2 * Copyright (C) International Business Machines Corp., 2000-2004
3 * Portions Copyright (C) Tino Reichardt, 2012
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See
13 * the GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21 #include <linux/slab.h>
22 #include "jfs_incore.h"
23 #include "jfs_superblock.h"
27 #include "jfs_metapage.h"
28 #include "jfs_debug.h"
29 #include "jfs_discard.h"
32 * SERIALIZATION of the Block Allocation Map.
34 * the working state of the block allocation map is accessed in
37 * 1) allocation and free requests that start at the dmap
38 * level and move up through the dmap control pages (i.e.
39 * the vast majority of requests).
41 * 2) allocation requests that start at dmap control page
42 * level and work down towards the dmaps.
44 * the serialization scheme used here is as follows.
46 * requests which start at the bottom are serialized against each
47 * other through buffers and each requests holds onto its buffers
48 * as it works it way up from a single dmap to the required level
49 * of dmap control page.
50 * requests that start at the top are serialized against each other
51 * and request that start from the bottom by the multiple read/single
52 * write inode lock of the bmap inode. requests starting at the top
53 * take this lock in write mode while request starting at the bottom
54 * take the lock in read mode. a single top-down request may proceed
55 * exclusively while multiple bottoms-up requests may proceed
56 * simultaneously (under the protection of busy buffers).
58 * in addition to information found in dmaps and dmap control pages,
59 * the working state of the block allocation map also includes read/
60 * write information maintained in the bmap descriptor (i.e. total
61 * free block count, allocation group level free block counts).
62 * a single exclusive lock (BMAP_LOCK) is used to guard this information
63 * in the face of multiple-bottoms up requests.
64 * (lock ordering: IREAD_LOCK, BMAP_LOCK);
66 * accesses to the persistent state of the block allocation map (limited
67 * to the persistent bitmaps in dmaps) is guarded by (busy) buffers.
70 #define BMAP_LOCK_INIT(bmp) mutex_init(&bmp->db_bmaplock)
71 #define BMAP_LOCK(bmp) mutex_lock(&bmp->db_bmaplock)
72 #define BMAP_UNLOCK(bmp) mutex_unlock(&bmp->db_bmaplock)
77 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
79 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval);
80 static int dbBackSplit(dmtree_t * tp, int leafno);
81 static int dbJoin(dmtree_t * tp, int leafno, int newval);
82 static void dbAdjTree(dmtree_t * tp, int leafno, int newval);
83 static int dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc,
85 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results);
86 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
88 static int dbAllocNear(struct bmap * bmp, struct dmap * dp, s64 blkno,
90 int l2nb, s64 * results);
91 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
93 static int dbAllocDmapLev(struct bmap * bmp, struct dmap * dp, int nblocks,
96 static int dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb,
98 static int dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno,
100 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks);
101 static int dbFindBits(u32 word, int l2nb);
102 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno);
103 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx);
104 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
106 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
108 static int dbMaxBud(u8 * cp);
109 static int blkstol2(s64 nb);
111 static int cntlz(u32 value);
112 static int cnttz(u32 word);
114 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
116 static int dbInitDmap(struct dmap * dp, s64 blkno, int nblocks);
117 static int dbInitDmapTree(struct dmap * dp);
118 static int dbInitTree(struct dmaptree * dtp);
119 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i);
120 static int dbGetL2AGSize(s64 nblocks);
125 * table used for determining buddy sizes within characters of
126 * dmap bitmap words. the characters themselves serve as indexes
127 * into the table, with the table elements yielding the maximum
128 * binary buddy of free bits within the character.
130 static const s8 budtab[256] = {
131 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
132 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
133 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
134 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
135 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
136 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
137 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
138 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
139 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
140 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
141 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
142 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
143 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
144 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
145 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
146 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, -1
152 * FUNCTION: initializate the block allocation map.
154 * memory is allocated for the in-core bmap descriptor and
155 * the in-core descriptor is initialized from disk.
158 * ipbmap - pointer to in-core inode for the block map.
162 * -ENOMEM - insufficient memory
164 * -EINVAL - wrong bmap data
166 int dbMount(struct inode *ipbmap)
169 struct dbmap_disk *dbmp_le;
174 * allocate/initialize the in-memory bmap descriptor
176 /* allocate memory for the in-memory bmap descriptor */
177 bmp = kmalloc(sizeof(struct bmap), GFP_KERNEL);
181 /* read the on-disk bmap descriptor. */
182 mp = read_metapage(ipbmap,
183 BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
190 /* copy the on-disk bmap descriptor to its in-memory version. */
191 dbmp_le = (struct dbmap_disk *) mp->data;
192 bmp->db_mapsize = le64_to_cpu(dbmp_le->dn_mapsize);
193 bmp->db_nfree = le64_to_cpu(dbmp_le->dn_nfree);
195 bmp->db_l2nbperpage = le32_to_cpu(dbmp_le->dn_l2nbperpage);
196 if (bmp->db_l2nbperpage > L2PSIZE - L2MINBLOCKSIZE) {
198 goto err_release_metapage;
201 bmp->db_numag = le32_to_cpu(dbmp_le->dn_numag);
202 if (!bmp->db_numag) {
204 goto err_release_metapage;
207 bmp->db_maxlevel = le32_to_cpu(dbmp_le->dn_maxlevel);
208 bmp->db_maxag = le32_to_cpu(dbmp_le->dn_maxag);
209 bmp->db_agpref = le32_to_cpu(dbmp_le->dn_agpref);
210 bmp->db_aglevel = le32_to_cpu(dbmp_le->dn_aglevel);
211 bmp->db_agheight = le32_to_cpu(dbmp_le->dn_agheight);
212 bmp->db_agwidth = le32_to_cpu(dbmp_le->dn_agwidth);
213 bmp->db_agstart = le32_to_cpu(dbmp_le->dn_agstart);
214 bmp->db_agl2size = le32_to_cpu(dbmp_le->dn_agl2size);
215 if (bmp->db_agl2size > L2MAXL2SIZE - L2MAXAG ||
216 bmp->db_agl2size < 0) {
218 goto err_release_metapage;
221 if (((bmp->db_mapsize - 1) >> bmp->db_agl2size) > MAXAG) {
223 goto err_release_metapage;
226 for (i = 0; i < MAXAG; i++)
227 bmp->db_agfree[i] = le64_to_cpu(dbmp_le->dn_agfree[i]);
228 bmp->db_agsize = le64_to_cpu(dbmp_le->dn_agsize);
229 bmp->db_maxfreebud = dbmp_le->dn_maxfreebud;
231 /* release the buffer. */
232 release_metapage(mp);
234 /* bind the bmap inode and the bmap descriptor to each other. */
235 bmp->db_ipbmap = ipbmap;
236 JFS_SBI(ipbmap->i_sb)->bmap = bmp;
238 memset(bmp->db_active, 0, sizeof(bmp->db_active));
241 * allocate/initialize the bmap lock
247 err_release_metapage:
248 release_metapage(mp);
258 * FUNCTION: terminate the block allocation map in preparation for
259 * file system unmount.
261 * the in-core bmap descriptor is written to disk and
262 * the memory for this descriptor is freed.
265 * ipbmap - pointer to in-core inode for the block map.
271 int dbUnmount(struct inode *ipbmap, int mounterror)
273 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
275 if (!(mounterror || isReadOnly(ipbmap)))
279 * Invalidate the page cache buffers
281 truncate_inode_pages(ipbmap->i_mapping, 0);
283 /* free the memory for the in-memory bmap. */
285 JFS_SBI(ipbmap->i_sb)->bmap = NULL;
293 int dbSync(struct inode *ipbmap)
295 struct dbmap_disk *dbmp_le;
296 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
301 * write bmap global control page
303 /* get the buffer for the on-disk bmap descriptor. */
304 mp = read_metapage(ipbmap,
305 BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
308 jfs_err("dbSync: read_metapage failed!");
311 /* copy the in-memory version of the bmap to the on-disk version */
312 dbmp_le = (struct dbmap_disk *) mp->data;
313 dbmp_le->dn_mapsize = cpu_to_le64(bmp->db_mapsize);
314 dbmp_le->dn_nfree = cpu_to_le64(bmp->db_nfree);
315 dbmp_le->dn_l2nbperpage = cpu_to_le32(bmp->db_l2nbperpage);
316 dbmp_le->dn_numag = cpu_to_le32(bmp->db_numag);
317 dbmp_le->dn_maxlevel = cpu_to_le32(bmp->db_maxlevel);
318 dbmp_le->dn_maxag = cpu_to_le32(bmp->db_maxag);
319 dbmp_le->dn_agpref = cpu_to_le32(bmp->db_agpref);
320 dbmp_le->dn_aglevel = cpu_to_le32(bmp->db_aglevel);
321 dbmp_le->dn_agheight = cpu_to_le32(bmp->db_agheight);
322 dbmp_le->dn_agwidth = cpu_to_le32(bmp->db_agwidth);
323 dbmp_le->dn_agstart = cpu_to_le32(bmp->db_agstart);
324 dbmp_le->dn_agl2size = cpu_to_le32(bmp->db_agl2size);
325 for (i = 0; i < MAXAG; i++)
326 dbmp_le->dn_agfree[i] = cpu_to_le64(bmp->db_agfree[i]);
327 dbmp_le->dn_agsize = cpu_to_le64(bmp->db_agsize);
328 dbmp_le->dn_maxfreebud = bmp->db_maxfreebud;
330 /* write the buffer */
334 * write out dirty pages of bmap
336 filemap_write_and_wait(ipbmap->i_mapping);
338 diWriteSpecial(ipbmap, 0);
346 * FUNCTION: free the specified block range from the working block
349 * the blocks will be free from the working map one dmap
353 * ip - pointer to in-core inode;
354 * blkno - starting block number to be freed.
355 * nblocks - number of blocks to be freed.
361 int dbFree(struct inode *ip, s64 blkno, s64 nblocks)
367 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
368 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
369 struct super_block *sb = ipbmap->i_sb;
371 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
373 /* block to be freed better be within the mapsize. */
374 if (unlikely((blkno == 0) || (blkno + nblocks > bmp->db_mapsize))) {
375 IREAD_UNLOCK(ipbmap);
376 printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
377 (unsigned long long) blkno,
378 (unsigned long long) nblocks);
379 jfs_error(ip->i_sb, "block to be freed is outside the map\n");
384 * TRIM the blocks, when mounted with discard option
386 if (JFS_SBI(sb)->flag & JFS_DISCARD)
387 if (JFS_SBI(sb)->minblks_trim <= nblocks)
388 jfs_issue_discard(ipbmap, blkno, nblocks);
391 * free the blocks a dmap at a time.
394 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
395 /* release previous dmap if any */
400 /* get the buffer for the current dmap. */
401 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
402 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
404 IREAD_UNLOCK(ipbmap);
407 dp = (struct dmap *) mp->data;
409 /* determine the number of blocks to be freed from
412 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
414 /* free the blocks. */
415 if ((rc = dbFreeDmap(bmp, dp, blkno, nb))) {
416 jfs_error(ip->i_sb, "error in block map\n");
417 release_metapage(mp);
418 IREAD_UNLOCK(ipbmap);
423 /* write the last buffer. */
427 IREAD_UNLOCK(ipbmap);
434 * NAME: dbUpdatePMap()
436 * FUNCTION: update the allocation state (free or allocate) of the
437 * specified block range in the persistent block allocation map.
439 * the blocks will be updated in the persistent map one
443 * ipbmap - pointer to in-core inode for the block map.
444 * free - 'true' if block range is to be freed from the persistent
445 * map; 'false' if it is to be allocated.
446 * blkno - starting block number of the range.
447 * nblocks - number of contiguous blocks in the range.
448 * tblk - transaction block;
455 dbUpdatePMap(struct inode *ipbmap,
456 int free, s64 blkno, s64 nblocks, struct tblock * tblk)
458 int nblks, dbitno, wbitno, rbits;
459 int word, nbits, nwords;
460 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
461 s64 lblkno, rem, lastlblkno;
466 int lsn, difft, diffp;
469 /* the blocks better be within the mapsize. */
470 if (blkno + nblocks > bmp->db_mapsize) {
471 printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
472 (unsigned long long) blkno,
473 (unsigned long long) nblocks);
474 jfs_error(ipbmap->i_sb, "blocks are outside the map\n");
478 /* compute delta of transaction lsn from log syncpt */
480 log = (struct jfs_log *) JFS_SBI(tblk->sb)->log;
481 logdiff(difft, lsn, log);
484 * update the block state a dmap at a time.
488 for (rem = nblocks; rem > 0; rem -= nblks, blkno += nblks) {
489 /* get the buffer for the current dmap. */
490 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
491 if (lblkno != lastlblkno) {
496 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE,
500 metapage_wait_for_io(mp);
502 dp = (struct dmap *) mp->data;
504 /* determine the bit number and word within the dmap of
505 * the starting block. also determine how many blocks
506 * are to be updated within this dmap.
508 dbitno = blkno & (BPERDMAP - 1);
509 word = dbitno >> L2DBWORD;
510 nblks = min(rem, (s64)BPERDMAP - dbitno);
512 /* update the bits of the dmap words. the first and last
513 * words may only have a subset of their bits updated. if
514 * this is the case, we'll work against that word (i.e.
515 * partial first and/or last) only in a single pass. a
516 * single pass will also be used to update all words that
517 * are to have all their bits updated.
519 for (rbits = nblks; rbits > 0;
520 rbits -= nbits, dbitno += nbits) {
521 /* determine the bit number within the word and
522 * the number of bits within the word.
524 wbitno = dbitno & (DBWORD - 1);
525 nbits = min(rbits, DBWORD - wbitno);
527 /* check if only part of the word is to be updated. */
528 if (nbits < DBWORD) {
529 /* update (free or allocate) the bits
533 (ONES << (DBWORD - nbits) >> wbitno);
543 /* one or more words are to have all
544 * their bits updated. determine how
545 * many words and how many bits.
547 nwords = rbits >> L2DBWORD;
548 nbits = nwords << L2DBWORD;
550 /* update (free or allocate) the bits
554 memset(&dp->pmap[word], 0,
557 memset(&dp->pmap[word], (int) ONES,
567 if (lblkno == lastlblkno)
572 LOGSYNC_LOCK(log, flags);
574 /* inherit older/smaller lsn */
575 logdiff(diffp, mp->lsn, log);
579 /* move bp after tblock in logsync list */
580 list_move(&mp->synclist, &tblk->synclist);
583 /* inherit younger/larger clsn */
584 logdiff(difft, tblk->clsn, log);
585 logdiff(diffp, mp->clsn, log);
587 mp->clsn = tblk->clsn;
592 /* insert bp after tblock in logsync list */
594 list_add(&mp->synclist, &tblk->synclist);
596 mp->clsn = tblk->clsn;
598 LOGSYNC_UNLOCK(log, flags);
601 /* write the last buffer. */
613 * FUNCTION: find the preferred allocation group for new allocations.
615 * Within the allocation groups, we maintain a preferred
616 * allocation group which consists of a group with at least
617 * average free space. It is the preferred group that we target
618 * new inode allocation towards. The tie-in between inode
619 * allocation and block allocation occurs as we allocate the
620 * first (data) block of an inode and specify the inode (block)
621 * as the allocation hint for this block.
623 * We try to avoid having more than one open file growing in
624 * an allocation group, as this will lead to fragmentation.
625 * This differs from the old OS/2 method of trying to keep
626 * empty ags around for large allocations.
629 * ipbmap - pointer to in-core inode for the block map.
632 * the preferred allocation group number.
634 int dbNextAG(struct inode *ipbmap)
641 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
645 /* determine the average number of free blocks within the ags. */
646 avgfree = (u32)bmp->db_nfree / bmp->db_numag;
649 * if the current preferred ag does not have an active allocator
650 * and has at least average freespace, return it
652 agpref = bmp->db_agpref;
653 if ((atomic_read(&bmp->db_active[agpref]) == 0) &&
654 (bmp->db_agfree[agpref] >= avgfree))
657 /* From the last preferred ag, find the next one with at least
658 * average free space.
660 for (i = 0 ; i < bmp->db_numag; i++, agpref++) {
661 if (agpref == bmp->db_numag)
664 if (atomic_read(&bmp->db_active[agpref]))
665 /* open file is currently growing in this ag */
667 if (bmp->db_agfree[agpref] >= avgfree) {
668 /* Return this one */
669 bmp->db_agpref = agpref;
671 } else if (bmp->db_agfree[agpref] > hwm) {
672 /* Less than avg. freespace, but best so far */
673 hwm = bmp->db_agfree[agpref];
679 * If no inactive ag was found with average freespace, use the
683 bmp->db_agpref = next_best;
684 /* else leave db_agpref unchanged */
688 /* return the preferred group.
690 return (bmp->db_agpref);
696 * FUNCTION: attempt to allocate a specified number of contiguous free
697 * blocks from the working allocation block map.
699 * the block allocation policy uses hints and a multi-step
702 * for allocation requests smaller than the number of blocks
703 * per dmap, we first try to allocate the new blocks
704 * immediately following the hint. if these blocks are not
705 * available, we try to allocate blocks near the hint. if
706 * no blocks near the hint are available, we next try to
707 * allocate within the same dmap as contains the hint.
709 * if no blocks are available in the dmap or the allocation
710 * request is larger than the dmap size, we try to allocate
711 * within the same allocation group as contains the hint. if
712 * this does not succeed, we finally try to allocate anywhere
713 * within the aggregate.
715 * we also try to allocate anywhere within the aggregate for
716 * for allocation requests larger than the allocation group
717 * size or requests that specify no hint value.
720 * ip - pointer to in-core inode;
721 * hint - allocation hint.
722 * nblocks - number of contiguous blocks in the range.
723 * results - on successful return, set to the starting block number
724 * of the newly allocated contiguous range.
728 * -ENOSPC - insufficient disk resources
731 int dbAlloc(struct inode *ip, s64 hint, s64 nblocks, s64 * results)
734 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
743 /* assert that nblocks is valid */
746 /* get the log2 number of blocks to be allocated.
747 * if the number of blocks is not a log2 multiple,
748 * it will be rounded up to the next log2 multiple.
750 l2nb = BLKSTOL2(nblocks);
752 bmp = JFS_SBI(ip->i_sb)->bmap;
754 mapSize = bmp->db_mapsize;
756 /* the hint should be within the map */
757 if (hint >= mapSize) {
758 jfs_error(ip->i_sb, "the hint is outside the map\n");
762 /* if the number of blocks to be allocated is greater than the
763 * allocation group size, try to allocate anywhere.
765 if (l2nb > bmp->db_agl2size) {
766 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
768 rc = dbAllocAny(bmp, nblocks, l2nb, results);
774 * If no hint, let dbNextAG recommend an allocation group
779 /* we would like to allocate close to the hint. adjust the
780 * hint to the block following the hint since the allocators
781 * will start looking for free space starting at this point.
785 if (blkno >= bmp->db_mapsize)
788 agno = blkno >> bmp->db_agl2size;
790 /* check if blkno crosses over into a new allocation group.
791 * if so, check if we should allow allocations within this
794 if ((blkno & (bmp->db_agsize - 1)) == 0)
795 /* check if the AG is currently being written to.
796 * if so, call dbNextAG() to find a non-busy
797 * AG with sufficient free space.
799 if (atomic_read(&bmp->db_active[agno]))
802 /* check if the allocation request size can be satisfied from a
803 * single dmap. if so, try to allocate from the dmap containing
804 * the hint using a tiered strategy.
806 if (nblocks <= BPERDMAP) {
807 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
809 /* get the buffer for the dmap containing the hint.
812 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
813 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
817 dp = (struct dmap *) mp->data;
819 /* first, try to satisfy the allocation request with the
820 * blocks beginning at the hint.
822 if ((rc = dbAllocNext(bmp, dp, blkno, (int) nblocks))
826 mark_metapage_dirty(mp);
829 release_metapage(mp);
833 writers = atomic_read(&bmp->db_active[agno]);
835 ((writers == 1) && (JFS_IP(ip)->active_ag != agno))) {
837 * Someone else is writing in this allocation
838 * group. To avoid fragmenting, try another ag
840 release_metapage(mp);
841 IREAD_UNLOCK(ipbmap);
845 /* next, try to satisfy the allocation request with blocks
849 dbAllocNear(bmp, dp, blkno, (int) nblocks, l2nb, results))
852 mark_metapage_dirty(mp);
854 release_metapage(mp);
858 /* try to satisfy the allocation request with blocks within
859 * the same dmap as the hint.
861 if ((rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results))
864 mark_metapage_dirty(mp);
866 release_metapage(mp);
870 release_metapage(mp);
871 IREAD_UNLOCK(ipbmap);
874 /* try to satisfy the allocation request with blocks within
875 * the same allocation group as the hint.
877 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
878 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) != -ENOSPC)
881 IWRITE_UNLOCK(ipbmap);
886 * Let dbNextAG recommend a preferred allocation group
888 agno = dbNextAG(ipbmap);
889 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
891 /* Try to allocate within this allocation group. if that fails, try to
892 * allocate anywhere in the map.
894 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) == -ENOSPC)
895 rc = dbAllocAny(bmp, nblocks, l2nb, results);
898 IWRITE_UNLOCK(ipbmap);
903 IREAD_UNLOCK(ipbmap);
910 * NAME: dbAllocExact()
912 * FUNCTION: try to allocate the requested extent;
915 * ip - pointer to in-core inode;
916 * blkno - extent address;
917 * nblocks - extent length;
921 * -ENOSPC - insufficient disk resources
924 int dbAllocExact(struct inode *ip, s64 blkno, int nblocks)
927 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
928 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
933 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
936 * validate extent request:
938 * note: defragfs policy:
939 * max 64 blocks will be moved.
940 * allocation request size must be satisfied from a single dmap.
942 if (nblocks <= 0 || nblocks > BPERDMAP || blkno >= bmp->db_mapsize) {
943 IREAD_UNLOCK(ipbmap);
947 if (nblocks > ((s64) 1 << bmp->db_maxfreebud)) {
948 /* the free space is no longer available */
949 IREAD_UNLOCK(ipbmap);
953 /* read in the dmap covering the extent */
954 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
955 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
957 IREAD_UNLOCK(ipbmap);
960 dp = (struct dmap *) mp->data;
962 /* try to allocate the requested extent */
963 rc = dbAllocNext(bmp, dp, blkno, nblocks);
965 IREAD_UNLOCK(ipbmap);
968 mark_metapage_dirty(mp);
970 release_metapage(mp);
979 * FUNCTION: attempt to extend a current allocation by a specified
982 * this routine attempts to satisfy the allocation request
983 * by first trying to extend the existing allocation in
984 * place by allocating the additional blocks as the blocks
985 * immediately following the current allocation. if these
986 * blocks are not available, this routine will attempt to
987 * allocate a new set of contiguous blocks large enough
988 * to cover the existing allocation plus the additional
989 * number of blocks required.
992 * ip - pointer to in-core inode requiring allocation.
993 * blkno - starting block of the current allocation.
994 * nblocks - number of contiguous blocks within the current
996 * addnblocks - number of blocks to add to the allocation.
997 * results - on successful return, set to the starting block number
998 * of the existing allocation if the existing allocation
999 * was extended in place or to a newly allocated contiguous
1000 * range if the existing allocation could not be extended
1005 * -ENOSPC - insufficient disk resources
1009 dbReAlloc(struct inode *ip,
1010 s64 blkno, s64 nblocks, s64 addnblocks, s64 * results)
1014 /* try to extend the allocation in place.
1016 if ((rc = dbExtend(ip, blkno, nblocks, addnblocks)) == 0) {
1024 /* could not extend the allocation in place, so allocate a
1025 * new set of blocks for the entire request (i.e. try to get
1026 * a range of contiguous blocks large enough to cover the
1027 * existing allocation plus the additional blocks.)
1030 (ip, blkno + nblocks - 1, addnblocks + nblocks, results));
1037 * FUNCTION: attempt to extend a current allocation by a specified
1040 * this routine attempts to satisfy the allocation request
1041 * by first trying to extend the existing allocation in
1042 * place by allocating the additional blocks as the blocks
1043 * immediately following the current allocation.
1046 * ip - pointer to in-core inode requiring allocation.
1047 * blkno - starting block of the current allocation.
1048 * nblocks - number of contiguous blocks within the current
1050 * addnblocks - number of blocks to add to the allocation.
1054 * -ENOSPC - insufficient disk resources
1057 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks)
1059 struct jfs_sb_info *sbi = JFS_SBI(ip->i_sb);
1060 s64 lblkno, lastblkno, extblkno;
1062 struct metapage *mp;
1065 struct inode *ipbmap = sbi->ipbmap;
1069 * We don't want a non-aligned extent to cross a page boundary
1071 if (((rel_block = blkno & (sbi->nbperpage - 1))) &&
1072 (rel_block + nblocks + addnblocks > sbi->nbperpage))
1075 /* get the last block of the current allocation */
1076 lastblkno = blkno + nblocks - 1;
1078 /* determine the block number of the block following
1079 * the existing allocation.
1081 extblkno = lastblkno + 1;
1083 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
1085 /* better be within the file system */
1087 if (lastblkno < 0 || lastblkno >= bmp->db_mapsize) {
1088 IREAD_UNLOCK(ipbmap);
1089 jfs_error(ip->i_sb, "the block is outside the filesystem\n");
1093 /* we'll attempt to extend the current allocation in place by
1094 * allocating the additional blocks as the blocks immediately
1095 * following the current allocation. we only try to extend the
1096 * current allocation in place if the number of additional blocks
1097 * can fit into a dmap, the last block of the current allocation
1098 * is not the last block of the file system, and the start of the
1099 * inplace extension is not on an allocation group boundary.
1101 if (addnblocks > BPERDMAP || extblkno >= bmp->db_mapsize ||
1102 (extblkno & (bmp->db_agsize - 1)) == 0) {
1103 IREAD_UNLOCK(ipbmap);
1107 /* get the buffer for the dmap containing the first block
1110 lblkno = BLKTODMAP(extblkno, bmp->db_l2nbperpage);
1111 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
1113 IREAD_UNLOCK(ipbmap);
1117 dp = (struct dmap *) mp->data;
1119 /* try to allocate the blocks immediately following the
1120 * current allocation.
1122 rc = dbAllocNext(bmp, dp, extblkno, (int) addnblocks);
1124 IREAD_UNLOCK(ipbmap);
1126 /* were we successful ? */
1130 /* we were not successful */
1131 release_metapage(mp);
1138 * NAME: dbAllocNext()
1140 * FUNCTION: attempt to allocate the blocks of the specified block
1141 * range within a dmap.
1144 * bmp - pointer to bmap descriptor
1145 * dp - pointer to dmap.
1146 * blkno - starting block number of the range.
1147 * nblocks - number of contiguous free blocks of the range.
1151 * -ENOSPC - insufficient disk resources
1154 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1156 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
1159 int dbitno, word, rembits, nb, nwords, wbitno, nw;
1164 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1165 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1169 /* pick up a pointer to the leaves of the dmap tree.
1171 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1173 /* determine the bit number and word within the dmap of the
1176 dbitno = blkno & (BPERDMAP - 1);
1177 word = dbitno >> L2DBWORD;
1179 /* check if the specified block range is contained within
1182 if (dbitno + nblocks > BPERDMAP)
1185 /* check if the starting leaf indicates that anything
1188 if (leaf[word] == NOFREE)
1191 /* check the dmaps words corresponding to block range to see
1192 * if the block range is free. not all bits of the first and
1193 * last words may be contained within the block range. if this
1194 * is the case, we'll work against those words (i.e. partial first
1195 * and/or last) on an individual basis (a single pass) and examine
1196 * the actual bits to determine if they are free. a single pass
1197 * will be used for all dmap words fully contained within the
1198 * specified range. within this pass, the leaves of the dmap
1199 * tree will be examined to determine if the blocks are free. a
1200 * single leaf may describe the free space of multiple dmap
1201 * words, so we may visit only a subset of the actual leaves
1202 * corresponding to the dmap words of the block range.
1204 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
1205 /* determine the bit number within the word and
1206 * the number of bits within the word.
1208 wbitno = dbitno & (DBWORD - 1);
1209 nb = min(rembits, DBWORD - wbitno);
1211 /* check if only part of the word is to be examined.
1214 /* check if the bits are free.
1216 mask = (ONES << (DBWORD - nb) >> wbitno);
1217 if ((mask & ~le32_to_cpu(dp->wmap[word])) != mask)
1222 /* one or more dmap words are fully contained
1223 * within the block range. determine how many
1224 * words and how many bits.
1226 nwords = rembits >> L2DBWORD;
1227 nb = nwords << L2DBWORD;
1229 /* now examine the appropriate leaves to determine
1230 * if the blocks are free.
1232 while (nwords > 0) {
1233 /* does the leaf describe any free space ?
1235 if (leaf[word] < BUDMIN)
1238 /* determine the l2 number of bits provided
1242 min_t(int, leaf[word], NLSTOL2BSZ(nwords));
1244 /* determine how many words were handled.
1246 nw = BUDSIZE(l2size, BUDMIN);
1254 /* allocate the blocks.
1256 return (dbAllocDmap(bmp, dp, blkno, nblocks));
1261 * NAME: dbAllocNear()
1263 * FUNCTION: attempt to allocate a number of contiguous free blocks near
1264 * a specified block (hint) within a dmap.
1266 * starting with the dmap leaf that covers the hint, we'll
1267 * check the next four contiguous leaves for sufficient free
1268 * space. if sufficient free space is found, we'll allocate
1269 * the desired free space.
1272 * bmp - pointer to bmap descriptor
1273 * dp - pointer to dmap.
1274 * blkno - block number to allocate near.
1275 * nblocks - actual number of contiguous free blocks desired.
1276 * l2nb - log2 number of contiguous free blocks desired.
1277 * results - on successful return, set to the starting block number
1278 * of the newly allocated range.
1282 * -ENOSPC - insufficient disk resources
1285 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1288 dbAllocNear(struct bmap * bmp,
1289 struct dmap * dp, s64 blkno, int nblocks, int l2nb, s64 * results)
1291 int word, lword, rc;
1294 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1295 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1299 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1301 /* determine the word within the dmap that holds the hint
1302 * (i.e. blkno). also, determine the last word in the dmap
1303 * that we'll include in our examination.
1305 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
1306 lword = min(word + 4, LPERDMAP);
1308 /* examine the leaves for sufficient free space.
1310 for (; word < lword; word++) {
1311 /* does the leaf describe sufficient free space ?
1313 if (leaf[word] < l2nb)
1316 /* determine the block number within the file system
1317 * of the first block described by this dmap word.
1319 blkno = le64_to_cpu(dp->start) + (word << L2DBWORD);
1321 /* if not all bits of the dmap word are free, get the
1322 * starting bit number within the dmap word of the required
1323 * string of free bits and adjust the block number with the
1326 if (leaf[word] < BUDMIN)
1328 dbFindBits(le32_to_cpu(dp->wmap[word]), l2nb);
1330 /* allocate the blocks.
1332 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1345 * FUNCTION: attempt to allocate the specified number of contiguous
1346 * free blocks within the specified allocation group.
1348 * unless the allocation group size is equal to the number
1349 * of blocks per dmap, the dmap control pages will be used to
1350 * find the required free space, if available. we start the
1351 * search at the highest dmap control page level which
1352 * distinctly describes the allocation group's free space
1353 * (i.e. the highest level at which the allocation group's
1354 * free space is not mixed in with that of any other group).
1355 * in addition, we start the search within this level at a
1356 * height of the dmapctl dmtree at which the nodes distinctly
1357 * describe the allocation group's free space. at this height,
1358 * the allocation group's free space may be represented by 1
1359 * or two sub-trees, depending on the allocation group size.
1360 * we search the top nodes of these subtrees left to right for
1361 * sufficient free space. if sufficient free space is found,
1362 * the subtree is searched to find the leftmost leaf that
1363 * has free space. once we have made it to the leaf, we
1364 * move the search to the next lower level dmap control page
1365 * corresponding to this leaf. we continue down the dmap control
1366 * pages until we find the dmap that contains or starts the
1367 * sufficient free space and we allocate at this dmap.
1369 * if the allocation group size is equal to the dmap size,
1370 * we'll start at the dmap corresponding to the allocation
1371 * group and attempt the allocation at this level.
1373 * the dmap control page search is also not performed if the
1374 * allocation group is completely free and we go to the first
1375 * dmap of the allocation group to do the allocation. this is
1376 * done because the allocation group may be part (not the first
1377 * part) of a larger binary buddy system, causing the dmap
1378 * control pages to indicate no free space (NOFREE) within
1379 * the allocation group.
1382 * bmp - pointer to bmap descriptor
1383 * agno - allocation group number.
1384 * nblocks - actual number of contiguous free blocks desired.
1385 * l2nb - log2 number of contiguous free blocks desired.
1386 * results - on successful return, set to the starting block number
1387 * of the newly allocated range.
1391 * -ENOSPC - insufficient disk resources
1394 * note: IWRITE_LOCK(ipmap) held on entry/exit;
1397 dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, s64 * results)
1399 struct metapage *mp;
1400 struct dmapctl *dcp;
1401 int rc, ti, i, k, m, n, agperlev;
1405 /* allocation request should not be for more than the
1406 * allocation group size.
1408 if (l2nb > bmp->db_agl2size) {
1409 jfs_error(bmp->db_ipbmap->i_sb,
1410 "allocation request is larger than the allocation group size\n");
1414 /* determine the starting block number of the allocation
1417 blkno = (s64) agno << bmp->db_agl2size;
1419 /* check if the allocation group size is the minimum allocation
1420 * group size or if the allocation group is completely free. if
1421 * the allocation group size is the minimum size of BPERDMAP (i.e.
1422 * 1 dmap), there is no need to search the dmap control page (below)
1423 * that fully describes the allocation group since the allocation
1424 * group is already fully described by a dmap. in this case, we
1425 * just call dbAllocCtl() to search the dmap tree and allocate the
1426 * required space if available.
1428 * if the allocation group is completely free, dbAllocCtl() is
1429 * also called to allocate the required space. this is done for
1430 * two reasons. first, it makes no sense searching the dmap control
1431 * pages for free space when we know that free space exists. second,
1432 * the dmap control pages may indicate that the allocation group
1433 * has no free space if the allocation group is part (not the first
1434 * part) of a larger binary buddy system.
1436 if (bmp->db_agsize == BPERDMAP
1437 || bmp->db_agfree[agno] == bmp->db_agsize) {
1438 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1439 if ((rc == -ENOSPC) &&
1440 (bmp->db_agfree[agno] == bmp->db_agsize)) {
1441 printk(KERN_ERR "blkno = %Lx, blocks = %Lx\n",
1442 (unsigned long long) blkno,
1443 (unsigned long long) nblocks);
1444 jfs_error(bmp->db_ipbmap->i_sb,
1445 "dbAllocCtl failed in free AG\n");
1450 /* the buffer for the dmap control page that fully describes the
1453 lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, bmp->db_aglevel);
1454 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1457 dcp = (struct dmapctl *) mp->data;
1458 budmin = dcp->budmin;
1460 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1461 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
1462 release_metapage(mp);
1466 /* search the subtree(s) of the dmap control page that describes
1467 * the allocation group, looking for sufficient free space. to begin,
1468 * determine how many allocation groups are represented in a dmap
1469 * control page at the control page level (i.e. L0, L1, L2) that
1470 * fully describes an allocation group. next, determine the starting
1471 * tree index of this allocation group within the control page.
1474 (1 << (L2LPERCTL - (bmp->db_agheight << 1))) / bmp->db_agwidth;
1475 ti = bmp->db_agstart + bmp->db_agwidth * (agno & (agperlev - 1));
1477 /* dmap control page trees fan-out by 4 and a single allocation
1478 * group may be described by 1 or 2 subtrees within the ag level
1479 * dmap control page, depending upon the ag size. examine the ag's
1480 * subtrees for sufficient free space, starting with the leftmost
1483 for (i = 0; i < bmp->db_agwidth; i++, ti++) {
1484 /* is there sufficient free space ?
1486 if (l2nb > dcp->stree[ti])
1489 /* sufficient free space found in a subtree. now search down
1490 * the subtree to find the leftmost leaf that describes this
1493 for (k = bmp->db_agheight; k > 0; k--) {
1494 for (n = 0, m = (ti << 2) + 1; n < 4; n++) {
1495 if (l2nb <= dcp->stree[m + n]) {
1501 jfs_error(bmp->db_ipbmap->i_sb,
1502 "failed descending stree\n");
1503 release_metapage(mp);
1508 /* determine the block number within the file system
1509 * that corresponds to this leaf.
1511 if (bmp->db_aglevel == 2)
1513 else if (bmp->db_aglevel == 1)
1514 blkno &= ~(MAXL1SIZE - 1);
1515 else /* bmp->db_aglevel == 0 */
1516 blkno &= ~(MAXL0SIZE - 1);
1519 ((s64) (ti - le32_to_cpu(dcp->leafidx))) << budmin;
1521 /* release the buffer in preparation for going down
1522 * the next level of dmap control pages.
1524 release_metapage(mp);
1526 /* check if we need to continue to search down the lower
1527 * level dmap control pages. we need to if the number of
1528 * blocks required is less than maximum number of blocks
1529 * described at the next lower level.
1531 if (l2nb < budmin) {
1533 /* search the lower level dmap control pages to get
1534 * the starting block number of the dmap that
1535 * contains or starts off the free space.
1538 dbFindCtl(bmp, l2nb, bmp->db_aglevel - 1,
1540 if (rc == -ENOSPC) {
1541 jfs_error(bmp->db_ipbmap->i_sb,
1542 "control page inconsistent\n");
1549 /* allocate the blocks.
1551 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1552 if (rc == -ENOSPC) {
1553 jfs_error(bmp->db_ipbmap->i_sb,
1554 "unable to allocate blocks\n");
1560 /* no space in the allocation group. release the buffer and
1563 release_metapage(mp);
1570 * NAME: dbAllocAny()
1572 * FUNCTION: attempt to allocate the specified number of contiguous
1573 * free blocks anywhere in the file system.
1575 * dbAllocAny() attempts to find the sufficient free space by
1576 * searching down the dmap control pages, starting with the
1577 * highest level (i.e. L0, L1, L2) control page. if free space
1578 * large enough to satisfy the desired free space is found, the
1579 * desired free space is allocated.
1582 * bmp - pointer to bmap descriptor
1583 * nblocks - actual number of contiguous free blocks desired.
1584 * l2nb - log2 number of contiguous free blocks desired.
1585 * results - on successful return, set to the starting block number
1586 * of the newly allocated range.
1590 * -ENOSPC - insufficient disk resources
1593 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1595 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results)
1600 /* starting with the top level dmap control page, search
1601 * down the dmap control levels for sufficient free space.
1602 * if free space is found, dbFindCtl() returns the starting
1603 * block number of the dmap that contains or starts off the
1604 * range of free space.
1606 if ((rc = dbFindCtl(bmp, l2nb, bmp->db_maxlevel, &blkno)))
1609 /* allocate the blocks.
1611 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1612 if (rc == -ENOSPC) {
1613 jfs_error(bmp->db_ipbmap->i_sb, "unable to allocate blocks\n");
1621 * NAME: dbDiscardAG()
1623 * FUNCTION: attempt to discard (TRIM) all free blocks of specific AG
1626 * 1) allocate blocks, as large as possible and save them
1627 * while holding IWRITE_LOCK on ipbmap
1628 * 2) trim all these saved block/length values
1629 * 3) mark the blocks free again
1632 * - we work only on one ag at some time, minimizing how long we
1633 * need to lock ipbmap
1634 * - reading / writing the fs is possible most time, even on
1638 * - we write two times to the dmapctl and dmap pages
1639 * - but for me, this seems the best way, better ideas?
1643 * ip - pointer to in-core inode
1645 * minlen - minimum value of contiguous blocks
1648 * s64 - actual number of blocks trimmed
1650 s64 dbDiscardAG(struct inode *ip, int agno, s64 minlen)
1652 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
1653 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
1657 struct super_block *sb = ipbmap->i_sb;
1664 /* max blkno / nblocks pairs to trim */
1665 int count = 0, range_cnt;
1668 /* prevent others from writing new stuff here, while trimming */
1669 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
1671 nblocks = bmp->db_agfree[agno];
1672 max_ranges = nblocks;
1673 do_div(max_ranges, minlen);
1674 range_cnt = min_t(u64, max_ranges + 1, 32 * 1024);
1675 totrim = kmalloc(sizeof(struct range2trim) * range_cnt, GFP_NOFS);
1676 if (totrim == NULL) {
1677 jfs_error(bmp->db_ipbmap->i_sb, "no memory for trim array\n");
1678 IWRITE_UNLOCK(ipbmap);
1683 while (nblocks >= minlen) {
1684 l2nb = BLKSTOL2(nblocks);
1686 /* 0 = okay, -EIO = fatal, -ENOSPC -> try smaller block */
1687 rc = dbAllocAG(bmp, agno, nblocks, l2nb, &blkno);
1690 tt->nblocks = nblocks;
1693 /* the whole ag is free, trim now */
1694 if (bmp->db_agfree[agno] == 0)
1697 /* give a hint for the next while */
1698 nblocks = bmp->db_agfree[agno];
1700 } else if (rc == -ENOSPC) {
1701 /* search for next smaller log2 block */
1702 l2nb = BLKSTOL2(nblocks) - 1;
1703 nblocks = 1LL << l2nb;
1705 /* Trim any already allocated blocks */
1706 jfs_error(bmp->db_ipbmap->i_sb, "-EIO\n");
1710 /* check, if our trim array is full */
1711 if (unlikely(count >= range_cnt - 1))
1714 IWRITE_UNLOCK(ipbmap);
1716 tt->nblocks = 0; /* mark the current end */
1717 for (tt = totrim; tt->nblocks != 0; tt++) {
1718 /* when mounted with online discard, dbFree() will
1719 * call jfs_issue_discard() itself */
1720 if (!(JFS_SBI(sb)->flag & JFS_DISCARD))
1721 jfs_issue_discard(ip, tt->blkno, tt->nblocks);
1722 dbFree(ip, tt->blkno, tt->nblocks);
1723 trimmed += tt->nblocks;
1733 * FUNCTION: starting at a specified dmap control page level and block
1734 * number, search down the dmap control levels for a range of
1735 * contiguous free blocks large enough to satisfy an allocation
1736 * request for the specified number of free blocks.
1738 * if sufficient contiguous free blocks are found, this routine
1739 * returns the starting block number within a dmap page that
1740 * contains or starts a range of contiqious free blocks that
1741 * is sufficient in size.
1744 * bmp - pointer to bmap descriptor
1745 * level - starting dmap control page level.
1746 * l2nb - log2 number of contiguous free blocks desired.
1747 * *blkno - on entry, starting block number for conducting the search.
1748 * on successful return, the first block within a dmap page
1749 * that contains or starts a range of contiguous free blocks.
1753 * -ENOSPC - insufficient disk resources
1756 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1758 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno)
1760 int rc, leafidx, lev;
1762 struct dmapctl *dcp;
1764 struct metapage *mp;
1766 /* starting at the specified dmap control page level and block
1767 * number, search down the dmap control levels for the starting
1768 * block number of a dmap page that contains or starts off
1769 * sufficient free blocks.
1771 for (lev = level, b = *blkno; lev >= 0; lev--) {
1772 /* get the buffer of the dmap control page for the block
1773 * number and level (i.e. L0, L1, L2).
1775 lblkno = BLKTOCTL(b, bmp->db_l2nbperpage, lev);
1776 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1779 dcp = (struct dmapctl *) mp->data;
1780 budmin = dcp->budmin;
1782 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1783 jfs_error(bmp->db_ipbmap->i_sb,
1784 "Corrupt dmapctl page\n");
1785 release_metapage(mp);
1789 /* search the tree within the dmap control page for
1790 * sufficient free space. if sufficient free space is found,
1791 * dbFindLeaf() returns the index of the leaf at which
1792 * free space was found.
1794 rc = dbFindLeaf((dmtree_t *) dcp, l2nb, &leafidx);
1796 /* release the buffer.
1798 release_metapage(mp);
1804 jfs_error(bmp->db_ipbmap->i_sb,
1805 "dmap inconsistent\n");
1811 /* adjust the block number to reflect the location within
1812 * the dmap control page (i.e. the leaf) at which free
1815 b += (((s64) leafidx) << budmin);
1817 /* we stop the search at this dmap control page level if
1818 * the number of blocks required is greater than or equal
1819 * to the maximum number of blocks described at the next
1832 * NAME: dbAllocCtl()
1834 * FUNCTION: attempt to allocate a specified number of contiguous
1835 * blocks starting within a specific dmap.
1837 * this routine is called by higher level routines that search
1838 * the dmap control pages above the actual dmaps for contiguous
1839 * free space. the result of successful searches by these
1840 * routines are the starting block numbers within dmaps, with
1841 * the dmaps themselves containing the desired contiguous free
1842 * space or starting a contiguous free space of desired size
1843 * that is made up of the blocks of one or more dmaps. these
1844 * calls should not fail due to insufficent resources.
1846 * this routine is called in some cases where it is not known
1847 * whether it will fail due to insufficient resources. more
1848 * specifically, this occurs when allocating from an allocation
1849 * group whose size is equal to the number of blocks per dmap.
1850 * in this case, the dmap control pages are not examined prior
1851 * to calling this routine (to save pathlength) and the call
1854 * for a request size that fits within a dmap, this routine relies
1855 * upon the dmap's dmtree to find the requested contiguous free
1856 * space. for request sizes that are larger than a dmap, the
1857 * requested free space will start at the first block of the
1858 * first dmap (i.e. blkno).
1861 * bmp - pointer to bmap descriptor
1862 * nblocks - actual number of contiguous free blocks to allocate.
1863 * l2nb - log2 number of contiguous free blocks to allocate.
1864 * blkno - starting block number of the dmap to start the allocation
1866 * results - on successful return, set to the starting block number
1867 * of the newly allocated range.
1871 * -ENOSPC - insufficient disk resources
1874 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1877 dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, s64 * results)
1881 struct metapage *mp;
1884 /* check if the allocation request is confined to a single dmap.
1886 if (l2nb <= L2BPERDMAP) {
1887 /* get the buffer for the dmap.
1889 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
1890 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1893 dp = (struct dmap *) mp->data;
1895 /* try to allocate the blocks.
1897 rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results);
1899 mark_metapage_dirty(mp);
1901 release_metapage(mp);
1906 /* allocation request involving multiple dmaps. it must start on
1909 assert((blkno & (BPERDMAP - 1)) == 0);
1911 /* allocate the blocks dmap by dmap.
1913 for (n = nblocks, b = blkno; n > 0; n -= nb, b += nb) {
1914 /* get the buffer for the dmap.
1916 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1917 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1922 dp = (struct dmap *) mp->data;
1924 /* the dmap better be all free.
1926 if (dp->tree.stree[ROOT] != L2BPERDMAP) {
1927 release_metapage(mp);
1928 jfs_error(bmp->db_ipbmap->i_sb,
1929 "the dmap is not all free\n");
1934 /* determine how many blocks to allocate from this dmap.
1936 nb = min_t(s64, n, BPERDMAP);
1938 /* allocate the blocks from the dmap.
1940 if ((rc = dbAllocDmap(bmp, dp, b, nb))) {
1941 release_metapage(mp);
1945 /* write the buffer.
1950 /* set the results (starting block number) and return.
1955 /* something failed in handling an allocation request involving
1956 * multiple dmaps. we'll try to clean up by backing out any
1957 * allocation that has already happened for this request. if
1958 * we fail in backing out the allocation, we'll mark the file
1959 * system to indicate that blocks have been leaked.
1963 /* try to backout the allocations dmap by dmap.
1965 for (n = nblocks - n, b = blkno; n > 0;
1966 n -= BPERDMAP, b += BPERDMAP) {
1967 /* get the buffer for this dmap.
1969 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1970 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1972 /* could not back out. mark the file system
1973 * to indicate that we have leaked blocks.
1975 jfs_error(bmp->db_ipbmap->i_sb,
1976 "I/O Error: Block Leakage\n");
1979 dp = (struct dmap *) mp->data;
1981 /* free the blocks is this dmap.
1983 if (dbFreeDmap(bmp, dp, b, BPERDMAP)) {
1984 /* could not back out. mark the file system
1985 * to indicate that we have leaked blocks.
1987 release_metapage(mp);
1988 jfs_error(bmp->db_ipbmap->i_sb, "Block Leakage\n");
1992 /* write the buffer.
2002 * NAME: dbAllocDmapLev()
2004 * FUNCTION: attempt to allocate a specified number of contiguous blocks
2005 * from a specified dmap.
2007 * this routine checks if the contiguous blocks are available.
2008 * if so, nblocks of blocks are allocated; otherwise, ENOSPC is
2012 * mp - pointer to bmap descriptor
2013 * dp - pointer to dmap to attempt to allocate blocks from.
2014 * l2nb - log2 number of contiguous block desired.
2015 * nblocks - actual number of contiguous block desired.
2016 * results - on successful return, set to the starting block number
2017 * of the newly allocated range.
2021 * -ENOSPC - insufficient disk resources
2024 * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or
2025 * IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit;
2028 dbAllocDmapLev(struct bmap * bmp,
2029 struct dmap * dp, int nblocks, int l2nb, s64 * results)
2034 /* can't be more than a dmaps worth of blocks */
2035 assert(l2nb <= L2BPERDMAP);
2037 /* search the tree within the dmap page for sufficient
2038 * free space. if sufficient free space is found, dbFindLeaf()
2039 * returns the index of the leaf at which free space was found.
2041 if (dbFindLeaf((dmtree_t *) & dp->tree, l2nb, &leafidx))
2047 /* determine the block number within the file system corresponding
2048 * to the leaf at which free space was found.
2050 blkno = le64_to_cpu(dp->start) + (leafidx << L2DBWORD);
2052 /* if not all bits of the dmap word are free, get the starting
2053 * bit number within the dmap word of the required string of free
2054 * bits and adjust the block number with this value.
2056 if (dp->tree.stree[leafidx + LEAFIND] < BUDMIN)
2057 blkno += dbFindBits(le32_to_cpu(dp->wmap[leafidx]), l2nb);
2059 /* allocate the blocks */
2060 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
2068 * NAME: dbAllocDmap()
2070 * FUNCTION: adjust the disk allocation map to reflect the allocation
2071 * of a specified block range within a dmap.
2073 * this routine allocates the specified blocks from the dmap
2074 * through a call to dbAllocBits(). if the allocation of the
2075 * block range causes the maximum string of free blocks within
2076 * the dmap to change (i.e. the value of the root of the dmap's
2077 * dmtree), this routine will cause this change to be reflected
2078 * up through the appropriate levels of the dmap control pages
2079 * by a call to dbAdjCtl() for the L0 dmap control page that
2083 * bmp - pointer to bmap descriptor
2084 * dp - pointer to dmap to allocate the block range from.
2085 * blkno - starting block number of the block to be allocated.
2086 * nblocks - number of blocks to be allocated.
2092 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2094 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2100 /* save the current value of the root (i.e. maximum free string)
2103 oldroot = dp->tree.stree[ROOT];
2105 /* allocate the specified (blocks) bits */
2106 dbAllocBits(bmp, dp, blkno, nblocks);
2108 /* if the root has not changed, done. */
2109 if (dp->tree.stree[ROOT] == oldroot)
2112 /* root changed. bubble the change up to the dmap control pages.
2113 * if the adjustment of the upper level control pages fails,
2114 * backout the bit allocation (thus making everything consistent).
2116 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 1, 0)))
2117 dbFreeBits(bmp, dp, blkno, nblocks);
2124 * NAME: dbFreeDmap()
2126 * FUNCTION: adjust the disk allocation map to reflect the allocation
2127 * of a specified block range within a dmap.
2129 * this routine frees the specified blocks from the dmap through
2130 * a call to dbFreeBits(). if the deallocation of the block range
2131 * causes the maximum string of free blocks within the dmap to
2132 * change (i.e. the value of the root of the dmap's dmtree), this
2133 * routine will cause this change to be reflected up through the
2134 * appropriate levels of the dmap control pages by a call to
2135 * dbAdjCtl() for the L0 dmap control page that covers this dmap.
2138 * bmp - pointer to bmap descriptor
2139 * dp - pointer to dmap to free the block range from.
2140 * blkno - starting block number of the block to be freed.
2141 * nblocks - number of blocks to be freed.
2147 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2149 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2155 /* save the current value of the root (i.e. maximum free string)
2158 oldroot = dp->tree.stree[ROOT];
2160 /* free the specified (blocks) bits */
2161 rc = dbFreeBits(bmp, dp, blkno, nblocks);
2163 /* if error or the root has not changed, done. */
2164 if (rc || (dp->tree.stree[ROOT] == oldroot))
2167 /* root changed. bubble the change up to the dmap control pages.
2168 * if the adjustment of the upper level control pages fails,
2169 * backout the deallocation.
2171 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 0, 0))) {
2172 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
2174 /* as part of backing out the deallocation, we will have
2175 * to back split the dmap tree if the deallocation caused
2176 * the freed blocks to become part of a larger binary buddy
2179 if (dp->tree.stree[word] == NOFREE)
2180 dbBackSplit((dmtree_t *) & dp->tree, word);
2182 dbAllocBits(bmp, dp, blkno, nblocks);
2190 * NAME: dbAllocBits()
2192 * FUNCTION: allocate a specified block range from a dmap.
2194 * this routine updates the dmap to reflect the working
2195 * state allocation of the specified block range. it directly
2196 * updates the bits of the working map and causes the adjustment
2197 * of the binary buddy system described by the dmap's dmtree
2198 * leaves to reflect the bits allocated. it also causes the
2199 * dmap's dmtree, as a whole, to reflect the allocated range.
2202 * bmp - pointer to bmap descriptor
2203 * dp - pointer to dmap to allocate bits from.
2204 * blkno - starting block number of the bits to be allocated.
2205 * nblocks - number of bits to be allocated.
2207 * RETURN VALUES: none
2209 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2211 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2214 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2215 dmtree_t *tp = (dmtree_t *) & dp->tree;
2219 /* pick up a pointer to the leaves of the dmap tree */
2220 leaf = dp->tree.stree + LEAFIND;
2222 /* determine the bit number and word within the dmap of the
2225 dbitno = blkno & (BPERDMAP - 1);
2226 word = dbitno >> L2DBWORD;
2228 /* block range better be within the dmap */
2229 assert(dbitno + nblocks <= BPERDMAP);
2231 /* allocate the bits of the dmap's words corresponding to the block
2232 * range. not all bits of the first and last words may be contained
2233 * within the block range. if this is the case, we'll work against
2234 * those words (i.e. partial first and/or last) on an individual basis
2235 * (a single pass), allocating the bits of interest by hand and
2236 * updating the leaf corresponding to the dmap word. a single pass
2237 * will be used for all dmap words fully contained within the
2238 * specified range. within this pass, the bits of all fully contained
2239 * dmap words will be marked as free in a single shot and the leaves
2240 * will be updated. a single leaf may describe the free space of
2241 * multiple dmap words, so we may update only a subset of the actual
2242 * leaves corresponding to the dmap words of the block range.
2244 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2245 /* determine the bit number within the word and
2246 * the number of bits within the word.
2248 wbitno = dbitno & (DBWORD - 1);
2249 nb = min(rembits, DBWORD - wbitno);
2251 /* check if only part of a word is to be allocated.
2254 /* allocate (set to 1) the appropriate bits within
2257 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
2260 /* update the leaf for this dmap word. in addition
2261 * to setting the leaf value to the binary buddy max
2262 * of the updated dmap word, dbSplit() will split
2263 * the binary system of the leaves if need be.
2265 dbSplit(tp, word, BUDMIN,
2266 dbMaxBud((u8 *) & dp->wmap[word]));
2270 /* one or more dmap words are fully contained
2271 * within the block range. determine how many
2272 * words and allocate (set to 1) the bits of these
2275 nwords = rembits >> L2DBWORD;
2276 memset(&dp->wmap[word], (int) ONES, nwords * 4);
2278 /* determine how many bits.
2280 nb = nwords << L2DBWORD;
2282 /* now update the appropriate leaves to reflect
2283 * the allocated words.
2285 for (; nwords > 0; nwords -= nw) {
2286 if (leaf[word] < BUDMIN) {
2287 jfs_error(bmp->db_ipbmap->i_sb,
2288 "leaf page corrupt\n");
2292 /* determine what the leaf value should be
2293 * updated to as the minimum of the l2 number
2294 * of bits being allocated and the l2 number
2295 * of bits currently described by this leaf.
2297 size = min_t(int, leaf[word],
2298 NLSTOL2BSZ(nwords));
2300 /* update the leaf to reflect the allocation.
2301 * in addition to setting the leaf value to
2302 * NOFREE, dbSplit() will split the binary
2303 * system of the leaves to reflect the current
2304 * allocation (size).
2306 dbSplit(tp, word, size, NOFREE);
2308 /* get the number of dmap words handled */
2309 nw = BUDSIZE(size, BUDMIN);
2315 /* update the free count for this dmap */
2316 le32_add_cpu(&dp->nfree, -nblocks);
2320 /* if this allocation group is completely free,
2321 * update the maximum allocation group number if this allocation
2322 * group is the new max.
2324 agno = blkno >> bmp->db_agl2size;
2325 if (agno > bmp->db_maxag)
2326 bmp->db_maxag = agno;
2328 /* update the free count for the allocation group and map */
2329 bmp->db_agfree[agno] -= nblocks;
2330 bmp->db_nfree -= nblocks;
2337 * NAME: dbFreeBits()
2339 * FUNCTION: free a specified block range from a dmap.
2341 * this routine updates the dmap to reflect the working
2342 * state allocation of the specified block range. it directly
2343 * updates the bits of the working map and causes the adjustment
2344 * of the binary buddy system described by the dmap's dmtree
2345 * leaves to reflect the bits freed. it also causes the dmap's
2346 * dmtree, as a whole, to reflect the deallocated range.
2349 * bmp - pointer to bmap descriptor
2350 * dp - pointer to dmap to free bits from.
2351 * blkno - starting block number of the bits to be freed.
2352 * nblocks - number of bits to be freed.
2354 * RETURN VALUES: 0 for success
2356 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2358 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2361 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2362 dmtree_t *tp = (dmtree_t *) & dp->tree;
2366 /* determine the bit number and word within the dmap of the
2369 dbitno = blkno & (BPERDMAP - 1);
2370 word = dbitno >> L2DBWORD;
2372 /* block range better be within the dmap.
2374 assert(dbitno + nblocks <= BPERDMAP);
2376 /* free the bits of the dmaps words corresponding to the block range.
2377 * not all bits of the first and last words may be contained within
2378 * the block range. if this is the case, we'll work against those
2379 * words (i.e. partial first and/or last) on an individual basis
2380 * (a single pass), freeing the bits of interest by hand and updating
2381 * the leaf corresponding to the dmap word. a single pass will be used
2382 * for all dmap words fully contained within the specified range.
2383 * within this pass, the bits of all fully contained dmap words will
2384 * be marked as free in a single shot and the leaves will be updated. a
2385 * single leaf may describe the free space of multiple dmap words,
2386 * so we may update only a subset of the actual leaves corresponding
2387 * to the dmap words of the block range.
2389 * dbJoin() is used to update leaf values and will join the binary
2390 * buddy system of the leaves if the new leaf values indicate this
2393 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2394 /* determine the bit number within the word and
2395 * the number of bits within the word.
2397 wbitno = dbitno & (DBWORD - 1);
2398 nb = min(rembits, DBWORD - wbitno);
2400 /* check if only part of a word is to be freed.
2403 /* free (zero) the appropriate bits within this
2407 cpu_to_le32(~(ONES << (DBWORD - nb)
2410 /* update the leaf for this dmap word.
2412 rc = dbJoin(tp, word,
2413 dbMaxBud((u8 *) & dp->wmap[word]));
2419 /* one or more dmap words are fully contained
2420 * within the block range. determine how many
2421 * words and free (zero) the bits of these words.
2423 nwords = rembits >> L2DBWORD;
2424 memset(&dp->wmap[word], 0, nwords * 4);
2426 /* determine how many bits.
2428 nb = nwords << L2DBWORD;
2430 /* now update the appropriate leaves to reflect
2433 for (; nwords > 0; nwords -= nw) {
2434 /* determine what the leaf value should be
2435 * updated to as the minimum of the l2 number
2436 * of bits being freed and the l2 (max) number
2437 * of bits that can be described by this leaf.
2441 (word, L2LPERDMAP, BUDMIN),
2442 NLSTOL2BSZ(nwords));
2446 rc = dbJoin(tp, word, size);
2450 /* get the number of dmap words handled.
2452 nw = BUDSIZE(size, BUDMIN);
2458 /* update the free count for this dmap.
2460 le32_add_cpu(&dp->nfree, nblocks);
2464 /* update the free count for the allocation group and
2467 agno = blkno >> bmp->db_agl2size;
2468 bmp->db_nfree += nblocks;
2469 bmp->db_agfree[agno] += nblocks;
2471 /* check if this allocation group is not completely free and
2472 * if it is currently the maximum (rightmost) allocation group.
2473 * if so, establish the new maximum allocation group number by
2474 * searching left for the first allocation group with allocation.
2476 if ((bmp->db_agfree[agno] == bmp->db_agsize && agno == bmp->db_maxag) ||
2477 (agno == bmp->db_numag - 1 &&
2478 bmp->db_agfree[agno] == (bmp-> db_mapsize & (BPERDMAP - 1)))) {
2479 while (bmp->db_maxag > 0) {
2481 if (bmp->db_agfree[bmp->db_maxag] !=
2486 /* re-establish the allocation group preference if the
2487 * current preference is right of the maximum allocation
2490 if (bmp->db_agpref > bmp->db_maxag)
2491 bmp->db_agpref = bmp->db_maxag;
2503 * FUNCTION: adjust a dmap control page at a specified level to reflect
2504 * the change in a lower level dmap or dmap control page's
2505 * maximum string of free blocks (i.e. a change in the root
2506 * of the lower level object's dmtree) due to the allocation
2507 * or deallocation of a range of blocks with a single dmap.
2509 * on entry, this routine is provided with the new value of
2510 * the lower level dmap or dmap control page root and the
2511 * starting block number of the block range whose allocation
2512 * or deallocation resulted in the root change. this range
2513 * is respresented by a single leaf of the current dmapctl
2514 * and the leaf will be updated with this value, possibly
2515 * causing a binary buddy system within the leaves to be
2516 * split or joined. the update may also cause the dmapctl's
2517 * dmtree to be updated.
2519 * if the adjustment of the dmap control page, itself, causes its
2520 * root to change, this change will be bubbled up to the next dmap
2521 * control level by a recursive call to this routine, specifying
2522 * the new root value and the next dmap control page level to
2525 * bmp - pointer to bmap descriptor
2526 * blkno - the first block of a block range within a dmap. it is
2527 * the allocation or deallocation of this block range that
2528 * requires the dmap control page to be adjusted.
2529 * newval - the new value of the lower level dmap or dmap control
2531 * alloc - 'true' if adjustment is due to an allocation.
2532 * level - current level of dmap control page (i.e. L0, L1, L2) to
2539 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2542 dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, int level)
2544 struct metapage *mp;
2548 struct dmapctl *dcp;
2551 /* get the buffer for the dmap control page for the specified
2552 * block number and control page level.
2554 lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, level);
2555 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
2558 dcp = (struct dmapctl *) mp->data;
2560 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
2561 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
2562 release_metapage(mp);
2566 /* determine the leaf number corresponding to the block and
2567 * the index within the dmap control tree.
2569 leafno = BLKTOCTLLEAF(blkno, dcp->budmin);
2570 ti = leafno + le32_to_cpu(dcp->leafidx);
2572 /* save the current leaf value and the current root level (i.e.
2573 * maximum l2 free string described by this dmapctl).
2575 oldval = dcp->stree[ti];
2576 oldroot = dcp->stree[ROOT];
2578 /* check if this is a control page update for an allocation.
2579 * if so, update the leaf to reflect the new leaf value using
2580 * dbSplit(); otherwise (deallocation), use dbJoin() to update
2581 * the leaf with the new value. in addition to updating the
2582 * leaf, dbSplit() will also split the binary buddy system of
2583 * the leaves, if required, and bubble new values within the
2584 * dmapctl tree, if required. similarly, dbJoin() will join
2585 * the binary buddy system of leaves and bubble new values up
2586 * the dmapctl tree as required by the new leaf value.
2589 /* check if we are in the middle of a binary buddy
2590 * system. this happens when we are performing the
2591 * first allocation out of an allocation group that
2592 * is part (not the first part) of a larger binary
2593 * buddy system. if we are in the middle, back split
2594 * the system prior to calling dbSplit() which assumes
2595 * that it is at the front of a binary buddy system.
2597 if (oldval == NOFREE) {
2598 rc = dbBackSplit((dmtree_t *) dcp, leafno);
2601 oldval = dcp->stree[ti];
2603 dbSplit((dmtree_t *) dcp, leafno, dcp->budmin, newval);
2605 rc = dbJoin((dmtree_t *) dcp, leafno, newval);
2610 /* check if the root of the current dmap control page changed due
2611 * to the update and if the current dmap control page is not at
2612 * the current top level (i.e. L0, L1, L2) of the map. if so (i.e.
2613 * root changed and this is not the top level), call this routine
2614 * again (recursion) for the next higher level of the mapping to
2615 * reflect the change in root for the current dmap control page.
2617 if (dcp->stree[ROOT] != oldroot) {
2618 /* are we below the top level of the map. if so,
2619 * bubble the root up to the next higher level.
2621 if (level < bmp->db_maxlevel) {
2622 /* bubble up the new root of this dmap control page to
2626 dbAdjCtl(bmp, blkno, dcp->stree[ROOT], alloc,
2628 /* something went wrong in bubbling up the new
2629 * root value, so backout the changes to the
2630 * current dmap control page.
2633 dbJoin((dmtree_t *) dcp, leafno,
2636 /* the dbJoin() above might have
2637 * caused a larger binary buddy system
2638 * to form and we may now be in the
2639 * middle of it. if this is the case,
2640 * back split the buddies.
2642 if (dcp->stree[ti] == NOFREE)
2643 dbBackSplit((dmtree_t *)
2645 dbSplit((dmtree_t *) dcp, leafno,
2646 dcp->budmin, oldval);
2649 /* release the buffer and return the error.
2651 release_metapage(mp);
2655 /* we're at the top level of the map. update
2656 * the bmap control page to reflect the size
2657 * of the maximum free buddy system.
2659 assert(level == bmp->db_maxlevel);
2660 if (bmp->db_maxfreebud != oldroot) {
2661 jfs_error(bmp->db_ipbmap->i_sb,
2662 "the maximum free buddy is not the old root\n");
2664 bmp->db_maxfreebud = dcp->stree[ROOT];
2668 /* write the buffer.
2679 * FUNCTION: update the leaf of a dmtree with a new value, splitting
2680 * the leaf from the binary buddy system of the dmtree's
2681 * leaves, as required.
2684 * tp - pointer to the tree containing the leaf.
2685 * leafno - the number of the leaf to be updated.
2686 * splitsz - the size the binary buddy system starting at the leaf
2687 * must be split to, specified as the log2 number of blocks.
2688 * newval - the new value for the leaf.
2690 * RETURN VALUES: none
2692 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2694 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval)
2698 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2700 /* check if the leaf needs to be split.
2702 if (leaf[leafno] > tp->dmt_budmin) {
2703 /* the split occurs by cutting the buddy system in half
2704 * at the specified leaf until we reach the specified
2705 * size. pick up the starting split size (current size
2706 * - 1 in l2) and the corresponding buddy size.
2708 cursz = leaf[leafno] - 1;
2709 budsz = BUDSIZE(cursz, tp->dmt_budmin);
2711 /* split until we reach the specified size.
2713 while (cursz >= splitsz) {
2714 /* update the buddy's leaf with its new value.
2716 dbAdjTree(tp, leafno ^ budsz, cursz);
2718 /* on to the next size and buddy.
2725 /* adjust the dmap tree to reflect the specified leaf's new
2728 dbAdjTree(tp, leafno, newval);
2733 * NAME: dbBackSplit()
2735 * FUNCTION: back split the binary buddy system of dmtree leaves
2736 * that hold a specified leaf until the specified leaf
2737 * starts its own binary buddy system.
2739 * the allocators typically perform allocations at the start
2740 * of binary buddy systems and dbSplit() is used to accomplish
2741 * any required splits. in some cases, however, allocation
2742 * may occur in the middle of a binary system and requires a
2743 * back split, with the split proceeding out from the middle of
2744 * the system (less efficient) rather than the start of the
2745 * system (more efficient). the cases in which a back split
2746 * is required are rare and are limited to the first allocation
2747 * within an allocation group which is a part (not first part)
2748 * of a larger binary buddy system and a few exception cases
2749 * in which a previous join operation must be backed out.
2752 * tp - pointer to the tree containing the leaf.
2753 * leafno - the number of the leaf to be updated.
2755 * RETURN VALUES: none
2757 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2759 static int dbBackSplit(dmtree_t * tp, int leafno)
2761 int budsz, bud, w, bsz, size;
2763 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2765 /* leaf should be part (not first part) of a binary
2768 assert(leaf[leafno] == NOFREE);
2770 /* the back split is accomplished by iteratively finding the leaf
2771 * that starts the buddy system that contains the specified leaf and
2772 * splitting that system in two. this iteration continues until
2773 * the specified leaf becomes the start of a buddy system.
2775 * determine maximum possible l2 size for the specified leaf.
2778 LITOL2BSZ(leafno, le32_to_cpu(tp->dmt_l2nleafs),
2781 /* determine the number of leaves covered by this size. this
2782 * is the buddy size that we will start with as we search for
2783 * the buddy system that contains the specified leaf.
2785 budsz = BUDSIZE(size, tp->dmt_budmin);
2789 while (leaf[leafno] == NOFREE) {
2790 /* find the leftmost buddy leaf.
2792 for (w = leafno, bsz = budsz;; bsz <<= 1,
2793 w = (w < bud) ? w : bud) {
2794 if (bsz >= le32_to_cpu(tp->dmt_nleafs)) {
2795 jfs_err("JFS: block map error in dbBackSplit");
2799 /* determine the buddy.
2803 /* check if this buddy is the start of the system.
2805 if (leaf[bud] != NOFREE) {
2806 /* split the leaf at the start of the
2809 cursz = leaf[bud] - 1;
2810 dbSplit(tp, bud, cursz, cursz);
2816 if (leaf[leafno] != size) {
2817 jfs_err("JFS: wrong leaf value in dbBackSplit");
2827 * FUNCTION: update the leaf of a dmtree with a new value, joining
2828 * the leaf with other leaves of the dmtree into a multi-leaf
2829 * binary buddy system, as required.
2832 * tp - pointer to the tree containing the leaf.
2833 * leafno - the number of the leaf to be updated.
2834 * newval - the new value for the leaf.
2836 * RETURN VALUES: none
2838 static int dbJoin(dmtree_t * tp, int leafno, int newval)
2843 /* can the new leaf value require a join with other leaves ?
2845 if (newval >= tp->dmt_budmin) {
2846 /* pickup a pointer to the leaves of the tree.
2848 leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2850 /* try to join the specified leaf into a large binary
2851 * buddy system. the join proceeds by attempting to join
2852 * the specified leafno with its buddy (leaf) at new value.
2853 * if the join occurs, we attempt to join the left leaf
2854 * of the joined buddies with its buddy at new value + 1.
2855 * we continue to join until we find a buddy that cannot be
2856 * joined (does not have a value equal to the size of the
2857 * last join) or until all leaves have been joined into a
2860 * get the buddy size (number of words covered) of
2863 budsz = BUDSIZE(newval, tp->dmt_budmin);
2867 while (budsz < le32_to_cpu(tp->dmt_nleafs)) {
2868 /* get the buddy leaf.
2870 buddy = leafno ^ budsz;
2872 /* if the leaf's new value is greater than its
2873 * buddy's value, we join no more.
2875 if (newval > leaf[buddy])
2878 /* It shouldn't be less */
2879 if (newval < leaf[buddy])
2882 /* check which (leafno or buddy) is the left buddy.
2883 * the left buddy gets to claim the blocks resulting
2884 * from the join while the right gets to claim none.
2885 * the left buddy is also eligible to participate in
2886 * a join at the next higher level while the right
2890 if (leafno < buddy) {
2891 /* leafno is the left buddy.
2893 dbAdjTree(tp, buddy, NOFREE);
2895 /* buddy is the left buddy and becomes
2898 dbAdjTree(tp, leafno, NOFREE);
2902 /* on to try the next join.
2909 /* update the leaf value.
2911 dbAdjTree(tp, leafno, newval);
2920 * FUNCTION: update a leaf of a dmtree with a new value, adjusting
2921 * the dmtree, as required, to reflect the new leaf value.
2922 * the combination of any buddies must already be done before
2926 * tp - pointer to the tree to be adjusted.
2927 * leafno - the number of the leaf to be updated.
2928 * newval - the new value for the leaf.
2930 * RETURN VALUES: none
2932 static void dbAdjTree(dmtree_t * tp, int leafno, int newval)
2937 /* pick up the index of the leaf for this leafno.
2939 lp = leafno + le32_to_cpu(tp->dmt_leafidx);
2941 /* is the current value the same as the old value ? if so,
2942 * there is nothing to do.
2944 if (tp->dmt_stree[lp] == newval)
2947 /* set the new value.
2949 tp->dmt_stree[lp] = newval;
2951 /* bubble the new value up the tree as required.
2953 for (k = 0; k < le32_to_cpu(tp->dmt_height); k++) {
2954 /* get the index of the first leaf of the 4 leaf
2955 * group containing the specified leaf (leafno).
2957 lp = ((lp - 1) & ~0x03) + 1;
2959 /* get the index of the parent of this 4 leaf group.
2963 /* determine the maximum of the 4 leaves.
2965 max = TREEMAX(&tp->dmt_stree[lp]);
2967 /* if the maximum of the 4 is the same as the
2968 * parent's value, we're done.
2970 if (tp->dmt_stree[pp] == max)
2973 /* parent gets new value.
2975 tp->dmt_stree[pp] = max;
2977 /* parent becomes leaf for next go-round.
2985 * NAME: dbFindLeaf()
2987 * FUNCTION: search a dmtree_t for sufficient free blocks, returning
2988 * the index of a leaf describing the free blocks if
2989 * sufficient free blocks are found.
2991 * the search starts at the top of the dmtree_t tree and
2992 * proceeds down the tree to the leftmost leaf with sufficient
2996 * tp - pointer to the tree to be searched.
2997 * l2nb - log2 number of free blocks to search for.
2998 * leafidx - return pointer to be set to the index of the leaf
2999 * describing at least l2nb free blocks if sufficient
3000 * free blocks are found.
3004 * -ENOSPC - insufficient free blocks.
3006 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx)
3008 int ti, n = 0, k, x = 0;
3010 /* first check the root of the tree to see if there is
3011 * sufficient free space.
3013 if (l2nb > tp->dmt_stree[ROOT])
3016 /* sufficient free space available. now search down the tree
3017 * starting at the next level for the leftmost leaf that
3018 * describes sufficient free space.
3020 for (k = le32_to_cpu(tp->dmt_height), ti = 1;
3021 k > 0; k--, ti = ((ti + n) << 2) + 1) {
3022 /* search the four nodes at this level, starting from
3025 for (x = ti, n = 0; n < 4; n++) {
3026 /* sufficient free space found. move to the next
3027 * level (or quit if this is the last level).
3029 if (l2nb <= tp->dmt_stree[x + n])
3033 /* better have found something since the higher
3034 * levels of the tree said it was here.
3039 /* set the return to the leftmost leaf describing sufficient
3042 *leafidx = x + n - le32_to_cpu(tp->dmt_leafidx);
3049 * NAME: dbFindBits()
3051 * FUNCTION: find a specified number of binary buddy free bits within a
3052 * dmap bitmap word value.
3054 * this routine searches the bitmap value for (1 << l2nb) free
3055 * bits at (1 << l2nb) alignments within the value.
3058 * word - dmap bitmap word value.
3059 * l2nb - number of free bits specified as a log2 number.
3062 * starting bit number of free bits.
3064 static int dbFindBits(u32 word, int l2nb)
3069 /* get the number of bits.
3072 assert(nb <= DBWORD);
3074 /* complement the word so we can use a mask (i.e. 0s represent
3075 * free bits) and compute the mask.
3078 mask = ONES << (DBWORD - nb);
3080 /* scan the word for nb free bits at nb alignments.
3082 for (bitno = 0; mask != 0; bitno += nb, mask >>= nb) {
3083 if ((mask & word) == mask)
3089 /* return the bit number.
3096 * NAME: dbMaxBud(u8 *cp)
3098 * FUNCTION: determine the largest binary buddy string of free
3099 * bits within 32-bits of the map.
3102 * cp - pointer to the 32-bit value.
3105 * largest binary buddy of free bits within a dmap word.
3107 static int dbMaxBud(u8 * cp)
3109 signed char tmp1, tmp2;
3111 /* check if the wmap word is all free. if so, the
3112 * free buddy size is BUDMIN.
3114 if (*((uint *) cp) == 0)
3117 /* check if the wmap word is half free. if so, the
3118 * free buddy size is BUDMIN-1.
3120 if (*((u16 *) cp) == 0 || *((u16 *) cp + 1) == 0)
3121 return (BUDMIN - 1);
3123 /* not all free or half free. determine the free buddy
3124 * size thru table lookup using quarters of the wmap word.
3126 tmp1 = max(budtab[cp[2]], budtab[cp[3]]);
3127 tmp2 = max(budtab[cp[0]], budtab[cp[1]]);
3128 return (max(tmp1, tmp2));
3133 * NAME: cnttz(uint word)
3135 * FUNCTION: determine the number of trailing zeros within a 32-bit
3139 * value - 32-bit value to be examined.
3142 * count of trailing zeros
3144 static int cnttz(u32 word)
3148 for (n = 0; n < 32; n++, word >>= 1) {
3158 * NAME: cntlz(u32 value)
3160 * FUNCTION: determine the number of leading zeros within a 32-bit
3164 * value - 32-bit value to be examined.
3167 * count of leading zeros
3169 static int cntlz(u32 value)
3173 for (n = 0; n < 32; n++, value <<= 1) {
3174 if (value & HIGHORDER)
3182 * NAME: blkstol2(s64 nb)
3184 * FUNCTION: convert a block count to its log2 value. if the block
3185 * count is not a l2 multiple, it is rounded up to the next
3186 * larger l2 multiple.
3189 * nb - number of blocks
3192 * log2 number of blocks
3194 static int blkstol2(s64 nb)
3197 s64 mask; /* meant to be signed */
3199 mask = (s64) 1 << (64 - 1);
3201 /* count the leading bits.
3203 for (l2nb = 0; l2nb < 64; l2nb++, mask >>= 1) {
3204 /* leading bit found.
3207 /* determine the l2 value.
3209 l2nb = (64 - 1) - l2nb;
3211 /* check if we need to round up.
3220 return 0; /* fix compiler warning */
3225 * NAME: dbAllocBottomUp()
3227 * FUNCTION: alloc the specified block range from the working block
3230 * the blocks will be alloc from the working map one dmap
3234 * ip - pointer to in-core inode;
3235 * blkno - starting block number to be freed.
3236 * nblocks - number of blocks to be freed.
3242 int dbAllocBottomUp(struct inode *ip, s64 blkno, s64 nblocks)
3244 struct metapage *mp;
3248 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
3249 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
3251 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
3253 /* block to be allocated better be within the mapsize. */
3254 ASSERT(nblocks <= bmp->db_mapsize - blkno);
3257 * allocate the blocks a dmap at a time.
3260 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
3261 /* release previous dmap if any */
3266 /* get the buffer for the current dmap. */
3267 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
3268 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
3270 IREAD_UNLOCK(ipbmap);
3273 dp = (struct dmap *) mp->data;
3275 /* determine the number of blocks to be allocated from
3278 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
3280 /* allocate the blocks. */
3281 if ((rc = dbAllocDmapBU(bmp, dp, blkno, nb))) {
3282 release_metapage(mp);
3283 IREAD_UNLOCK(ipbmap);
3288 /* write the last buffer. */
3291 IREAD_UNLOCK(ipbmap);
3297 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
3301 int dbitno, word, rembits, nb, nwords, wbitno, agno;
3303 struct dmaptree *tp = (struct dmaptree *) & dp->tree;
3305 /* save the current value of the root (i.e. maximum free string)
3308 oldroot = tp->stree[ROOT];
3310 /* determine the bit number and word within the dmap of the
3313 dbitno = blkno & (BPERDMAP - 1);
3314 word = dbitno >> L2DBWORD;
3316 /* block range better be within the dmap */
3317 assert(dbitno + nblocks <= BPERDMAP);
3319 /* allocate the bits of the dmap's words corresponding to the block
3320 * range. not all bits of the first and last words may be contained
3321 * within the block range. if this is the case, we'll work against
3322 * those words (i.e. partial first and/or last) on an individual basis
3323 * (a single pass), allocating the bits of interest by hand and
3324 * updating the leaf corresponding to the dmap word. a single pass
3325 * will be used for all dmap words fully contained within the
3326 * specified range. within this pass, the bits of all fully contained
3327 * dmap words will be marked as free in a single shot and the leaves
3328 * will be updated. a single leaf may describe the free space of
3329 * multiple dmap words, so we may update only a subset of the actual
3330 * leaves corresponding to the dmap words of the block range.
3332 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
3333 /* determine the bit number within the word and
3334 * the number of bits within the word.
3336 wbitno = dbitno & (DBWORD - 1);
3337 nb = min(rembits, DBWORD - wbitno);
3339 /* check if only part of a word is to be allocated.
3342 /* allocate (set to 1) the appropriate bits within
3345 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
3350 /* one or more dmap words are fully contained
3351 * within the block range. determine how many
3352 * words and allocate (set to 1) the bits of these
3355 nwords = rembits >> L2DBWORD;
3356 memset(&dp->wmap[word], (int) ONES, nwords * 4);
3358 /* determine how many bits */
3359 nb = nwords << L2DBWORD;
3364 /* update the free count for this dmap */
3365 le32_add_cpu(&dp->nfree, -nblocks);
3367 /* reconstruct summary tree */
3372 /* if this allocation group is completely free,
3373 * update the highest active allocation group number
3374 * if this allocation group is the new max.
3376 agno = blkno >> bmp->db_agl2size;
3377 if (agno > bmp->db_maxag)
3378 bmp->db_maxag = agno;
3380 /* update the free count for the allocation group and map */
3381 bmp->db_agfree[agno] -= nblocks;
3382 bmp->db_nfree -= nblocks;
3386 /* if the root has not changed, done. */
3387 if (tp->stree[ROOT] == oldroot)
3390 /* root changed. bubble the change up to the dmap control pages.
3391 * if the adjustment of the upper level control pages fails,
3392 * backout the bit allocation (thus making everything consistent).
3394 if ((rc = dbAdjCtl(bmp, blkno, tp->stree[ROOT], 1, 0)))
3395 dbFreeBits(bmp, dp, blkno, nblocks);
3402 * NAME: dbExtendFS()
3404 * FUNCTION: extend bmap from blkno for nblocks;
3405 * dbExtendFS() updates bmap ready for dbAllocBottomUp();
3409 * L1---------------------------------L1
3411 * L0---------L0---------L0 L0---------L0---------L0
3413 * d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,.,dm;
3414 * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm
3416 * <---old---><----------------------------extend----------------------->
3418 int dbExtendFS(struct inode *ipbmap, s64 blkno, s64 nblocks)
3420 struct jfs_sb_info *sbi = JFS_SBI(ipbmap->i_sb);
3421 int nbperpage = sbi->nbperpage;
3422 int i, i0 = true, j, j0 = true, k, n;
3425 struct metapage *mp, *l2mp, *l1mp = NULL, *l0mp = NULL;
3426 struct dmapctl *l2dcp, *l1dcp, *l0dcp;
3428 s8 *l0leaf, *l1leaf, *l2leaf;
3429 struct bmap *bmp = sbi->bmap;
3430 int agno, l2agsize, oldl2agsize;
3433 newsize = blkno + nblocks;
3435 jfs_info("dbExtendFS: blkno:%Ld nblocks:%Ld newsize:%Ld",
3436 (long long) blkno, (long long) nblocks, (long long) newsize);
3439 * initialize bmap control page.
3441 * all the data in bmap control page should exclude
3442 * the mkfs hidden dmap page.
3445 /* update mapsize */
3446 bmp->db_mapsize = newsize;
3447 bmp->db_maxlevel = BMAPSZTOLEV(bmp->db_mapsize);
3449 /* compute new AG size */
3450 l2agsize = dbGetL2AGSize(newsize);
3451 oldl2agsize = bmp->db_agl2size;
3453 bmp->db_agl2size = l2agsize;
3454 bmp->db_agsize = 1 << l2agsize;
3456 /* compute new number of AG */
3457 agno = bmp->db_numag;
3458 bmp->db_numag = newsize >> l2agsize;
3459 bmp->db_numag += ((u32) newsize % (u32) bmp->db_agsize) ? 1 : 0;
3462 * reconfigure db_agfree[]
3463 * from old AG configuration to new AG configuration;
3465 * coalesce contiguous k (newAGSize/oldAGSize) AGs;
3466 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn;
3467 * note: new AG size = old AG size * (2**x).
3469 if (l2agsize == oldl2agsize)
3471 k = 1 << (l2agsize - oldl2agsize);
3472 ag_rem = bmp->db_agfree[0]; /* save agfree[0] */
3473 for (i = 0, n = 0; i < agno; n++) {
3474 bmp->db_agfree[n] = 0; /* init collection point */
3476 /* coalesce contiguous k AGs; */
3477 for (j = 0; j < k && i < agno; j++, i++) {
3478 /* merge AGi to AGn */
3479 bmp->db_agfree[n] += bmp->db_agfree[i];
3482 bmp->db_agfree[0] += ag_rem; /* restore agfree[0] */
3484 for (; n < MAXAG; n++)
3485 bmp->db_agfree[n] = 0;
3488 * update highest active ag number
3491 bmp->db_maxag = bmp->db_maxag / k;
3496 * update bit maps and corresponding level control pages;
3497 * global control page db_nfree, db_agfree[agno], db_maxfreebud;
3501 p = BMAPBLKNO + nbperpage; /* L2 page */
3502 l2mp = read_metapage(ipbmap, p, PSIZE, 0);
3504 jfs_error(ipbmap->i_sb, "L2 page could not be read\n");
3507 l2dcp = (struct dmapctl *) l2mp->data;
3509 /* compute start L1 */
3510 k = blkno >> L2MAXL1SIZE;
3511 l2leaf = l2dcp->stree + CTLLEAFIND + k;
3512 p = BLKTOL1(blkno, sbi->l2nbperpage); /* L1 page */
3515 * extend each L1 in L2
3517 for (; k < LPERCTL; k++, p += nbperpage) {
3520 /* read in L1 page: (blkno & (MAXL1SIZE - 1)) */
3521 l1mp = read_metapage(ipbmap, p, PSIZE, 0);
3524 l1dcp = (struct dmapctl *) l1mp->data;
3526 /* compute start L0 */
3527 j = (blkno & (MAXL1SIZE - 1)) >> L2MAXL0SIZE;
3528 l1leaf = l1dcp->stree + CTLLEAFIND + j;
3529 p = BLKTOL0(blkno, sbi->l2nbperpage);
3532 /* assign/init L1 page */
3533 l1mp = get_metapage(ipbmap, p, PSIZE, 0);
3537 l1dcp = (struct dmapctl *) l1mp->data;
3539 /* compute start L0 */
3541 l1leaf = l1dcp->stree + CTLLEAFIND;
3542 p += nbperpage; /* 1st L0 of L1.k */
3546 * extend each L0 in L1
3548 for (; j < LPERCTL; j++) {
3551 /* read in L0 page: (blkno & (MAXL0SIZE - 1)) */
3553 l0mp = read_metapage(ipbmap, p, PSIZE, 0);
3556 l0dcp = (struct dmapctl *) l0mp->data;
3558 /* compute start dmap */
3559 i = (blkno & (MAXL0SIZE - 1)) >>
3561 l0leaf = l0dcp->stree + CTLLEAFIND + i;
3562 p = BLKTODMAP(blkno,
3566 /* assign/init L0 page */
3567 l0mp = get_metapage(ipbmap, p, PSIZE, 0);
3571 l0dcp = (struct dmapctl *) l0mp->data;
3573 /* compute start dmap */
3575 l0leaf = l0dcp->stree + CTLLEAFIND;
3576 p += nbperpage; /* 1st dmap of L0.j */
3580 * extend each dmap in L0
3582 for (; i < LPERCTL; i++) {
3584 * reconstruct the dmap page, and
3585 * initialize corresponding parent L0 leaf
3587 if ((n = blkno & (BPERDMAP - 1))) {
3588 /* read in dmap page: */
3589 mp = read_metapage(ipbmap, p,
3593 n = min(nblocks, (s64)BPERDMAP - n);
3595 /* assign/init dmap page */
3596 mp = read_metapage(ipbmap, p,
3601 n = min_t(s64, nblocks, BPERDMAP);
3604 dp = (struct dmap *) mp->data;
3605 *l0leaf = dbInitDmap(dp, blkno, n);
3608 agno = le64_to_cpu(dp->start) >> l2agsize;
3609 bmp->db_agfree[agno] += n;
3620 } /* for each dmap in a L0 */
3623 * build current L0 page from its leaves, and
3624 * initialize corresponding parent L1 leaf
3626 *l1leaf = dbInitDmapCtl(l0dcp, 0, ++i);
3627 write_metapage(l0mp);
3631 l1leaf++; /* continue for next L0 */
3633 /* more than 1 L0 ? */
3635 break; /* build L1 page */
3637 /* summarize in global bmap page */
3638 bmp->db_maxfreebud = *l1leaf;
3639 release_metapage(l1mp);
3640 release_metapage(l2mp);
3644 } /* for each L0 in a L1 */
3647 * build current L1 page from its leaves, and
3648 * initialize corresponding parent L2 leaf
3650 *l2leaf = dbInitDmapCtl(l1dcp, 1, ++j);
3651 write_metapage(l1mp);
3655 l2leaf++; /* continue for next L1 */
3657 /* more than 1 L1 ? */
3659 break; /* build L2 page */
3661 /* summarize in global bmap page */
3662 bmp->db_maxfreebud = *l2leaf;
3663 release_metapage(l2mp);
3667 } /* for each L1 in a L2 */
3669 jfs_error(ipbmap->i_sb, "function has not returned as expected\n");
3672 release_metapage(l0mp);
3674 release_metapage(l1mp);
3675 release_metapage(l2mp);
3679 * finalize bmap control page
3690 void dbFinalizeBmap(struct inode *ipbmap)
3692 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
3693 int actags, inactags, l2nl;
3694 s64 ag_rem, actfree, inactfree, avgfree;
3698 * finalize bmap control page
3702 * compute db_agpref: preferred ag to allocate from
3703 * (the leftmost ag with average free space in it);
3706 /* get the number of active ags and inacitve ags */
3707 actags = bmp->db_maxag + 1;
3708 inactags = bmp->db_numag - actags;
3709 ag_rem = bmp->db_mapsize & (bmp->db_agsize - 1); /* ??? */
3711 /* determine how many blocks are in the inactive allocation
3712 * groups. in doing this, we must account for the fact that
3713 * the rightmost group might be a partial group (i.e. file
3714 * system size is not a multiple of the group size).
3716 inactfree = (inactags && ag_rem) ?
3717 ((inactags - 1) << bmp->db_agl2size) + ag_rem
3718 : inactags << bmp->db_agl2size;
3720 /* determine how many free blocks are in the active
3721 * allocation groups plus the average number of free blocks
3722 * within the active ags.
3724 actfree = bmp->db_nfree - inactfree;
3725 avgfree = (u32) actfree / (u32) actags;
3727 /* if the preferred allocation group has not average free space.
3728 * re-establish the preferred group as the leftmost
3729 * group with average free space.
3731 if (bmp->db_agfree[bmp->db_agpref] < avgfree) {
3732 for (bmp->db_agpref = 0; bmp->db_agpref < actags;
3734 if (bmp->db_agfree[bmp->db_agpref] >= avgfree)
3737 if (bmp->db_agpref >= bmp->db_numag) {
3738 jfs_error(ipbmap->i_sb,
3739 "cannot find ag with average freespace\n");
3744 * compute db_aglevel, db_agheight, db_width, db_agstart:
3745 * an ag is covered in aglevel dmapctl summary tree,
3746 * at agheight level height (from leaf) with agwidth number of nodes
3747 * each, which starts at agstart index node of the smmary tree node
3750 bmp->db_aglevel = BMAPSZTOLEV(bmp->db_agsize);
3752 bmp->db_agl2size - (L2BPERDMAP + bmp->db_aglevel * L2LPERCTL);
3753 bmp->db_agheight = l2nl >> 1;
3754 bmp->db_agwidth = 1 << (l2nl - (bmp->db_agheight << 1));
3755 for (i = 5 - bmp->db_agheight, bmp->db_agstart = 0, n = 1; i > 0;
3757 bmp->db_agstart += n;
3765 * NAME: dbInitDmap()/ujfs_idmap_page()
3767 * FUNCTION: initialize working/persistent bitmap of the dmap page
3768 * for the specified number of blocks:
3770 * at entry, the bitmaps had been initialized as free (ZEROS);
3771 * The number of blocks will only account for the actually
3772 * existing blocks. Blocks which don't actually exist in
3773 * the aggregate will be marked as allocated (ONES);
3776 * dp - pointer to page of map
3777 * nblocks - number of blocks this page
3781 static int dbInitDmap(struct dmap * dp, s64 Blkno, int nblocks)
3783 int blkno, w, b, r, nw, nb, i;
3785 /* starting block number within the dmap */
3786 blkno = Blkno & (BPERDMAP - 1);
3789 dp->nblocks = dp->nfree = cpu_to_le32(nblocks);
3790 dp->start = cpu_to_le64(Blkno);
3792 if (nblocks == BPERDMAP) {
3793 memset(&dp->wmap[0], 0, LPERDMAP * 4);
3794 memset(&dp->pmap[0], 0, LPERDMAP * 4);
3798 le32_add_cpu(&dp->nblocks, nblocks);
3799 le32_add_cpu(&dp->nfree, nblocks);
3802 /* word number containing start block number */
3803 w = blkno >> L2DBWORD;
3806 * free the bits corresponding to the block range (ZEROS):
3807 * note: not all bits of the first and last words may be contained
3808 * within the block range.
3810 for (r = nblocks; r > 0; r -= nb, blkno += nb) {
3811 /* number of bits preceding range to be freed in the word */
3812 b = blkno & (DBWORD - 1);
3813 /* number of bits to free in the word */
3814 nb = min(r, DBWORD - b);
3816 /* is partial word to be freed ? */
3818 /* free (set to 0) from the bitmap word */
3819 dp->wmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3821 dp->pmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3824 /* skip the word freed */
3827 /* free (set to 0) contiguous bitmap words */
3829 memset(&dp->wmap[w], 0, nw * 4);
3830 memset(&dp->pmap[w], 0, nw * 4);
3832 /* skip the words freed */
3833 nb = nw << L2DBWORD;
3839 * mark bits following the range to be freed (non-existing
3840 * blocks) as allocated (ONES)
3843 if (blkno == BPERDMAP)
3846 /* the first word beyond the end of existing blocks */
3847 w = blkno >> L2DBWORD;
3849 /* does nblocks fall on a 32-bit boundary ? */
3850 b = blkno & (DBWORD - 1);
3852 /* mark a partial word allocated */
3853 dp->wmap[w] = dp->pmap[w] = cpu_to_le32(ONES >> b);
3857 /* set the rest of the words in the page to allocated (ONES) */
3858 for (i = w; i < LPERDMAP; i++)
3859 dp->pmap[i] = dp->wmap[i] = cpu_to_le32(ONES);
3865 return (dbInitDmapTree(dp));
3870 * NAME: dbInitDmapTree()/ujfs_complete_dmap()
3872 * FUNCTION: initialize summary tree of the specified dmap:
3874 * at entry, bitmap of the dmap has been initialized;
3877 * dp - dmap to complete
3878 * blkno - starting block number for this dmap
3879 * treemax - will be filled in with max free for this dmap
3881 * RETURNS: max free string at the root of the tree
3883 static int dbInitDmapTree(struct dmap * dp)
3885 struct dmaptree *tp;
3889 /* init fixed info of tree */
3891 tp->nleafs = cpu_to_le32(LPERDMAP);
3892 tp->l2nleafs = cpu_to_le32(L2LPERDMAP);
3893 tp->leafidx = cpu_to_le32(LEAFIND);
3894 tp->height = cpu_to_le32(4);
3895 tp->budmin = BUDMIN;
3897 /* init each leaf from corresponding wmap word:
3898 * note: leaf is set to NOFREE(-1) if all blocks of corresponding
3899 * bitmap word are allocated.
3901 cp = tp->stree + le32_to_cpu(tp->leafidx);
3902 for (i = 0; i < LPERDMAP; i++)
3903 *cp++ = dbMaxBud((u8 *) & dp->wmap[i]);
3905 /* build the dmap's binary buddy summary tree */
3906 return (dbInitTree(tp));
3911 * NAME: dbInitTree()/ujfs_adjtree()
3913 * FUNCTION: initialize binary buddy summary tree of a dmap or dmapctl.
3915 * at entry, the leaves of the tree has been initialized
3916 * from corresponding bitmap word or root of summary tree
3917 * of the child control page;
3918 * configure binary buddy system at the leaf level, then
3919 * bubble up the values of the leaf nodes up the tree.
3922 * cp - Pointer to the root of the tree
3923 * l2leaves- Number of leaf nodes as a power of 2
3924 * l2min - Number of blocks that can be covered by a leaf
3927 * RETURNS: max free string at the root of the tree
3929 static int dbInitTree(struct dmaptree * dtp)
3931 int l2max, l2free, bsize, nextb, i;
3932 int child, parent, nparent;
3937 /* Determine the maximum free string possible for the leaves */
3938 l2max = le32_to_cpu(dtp->l2nleafs) + dtp->budmin;
3941 * configure the leaf levevl into binary buddy system
3943 * Try to combine buddies starting with a buddy size of 1
3944 * (i.e. two leaves). At a buddy size of 1 two buddy leaves
3945 * can be combined if both buddies have a maximum free of l2min;
3946 * the combination will result in the left-most buddy leaf having
3947 * a maximum free of l2min+1.
3948 * After processing all buddies for a given size, process buddies
3949 * at the next higher buddy size (i.e. current size * 2) and
3950 * the next maximum free (current free + 1).
3951 * This continues until the maximum possible buddy combination
3952 * yields maximum free.
3954 for (l2free = dtp->budmin, bsize = 1; l2free < l2max;
3955 l2free++, bsize = nextb) {
3956 /* get next buddy size == current buddy pair size */
3959 /* scan each adjacent buddy pair at current buddy size */
3960 for (i = 0, cp = tp + le32_to_cpu(dtp->leafidx);
3961 i < le32_to_cpu(dtp->nleafs);
3962 i += nextb, cp += nextb) {
3963 /* coalesce if both adjacent buddies are max free */
3964 if (*cp == l2free && *(cp + bsize) == l2free) {
3965 *cp = l2free + 1; /* left take right */
3966 *(cp + bsize) = -1; /* right give left */
3972 * bubble summary information of leaves up the tree.
3974 * Starting at the leaf node level, the four nodes described by
3975 * the higher level parent node are compared for a maximum free and
3976 * this maximum becomes the value of the parent node.
3977 * when all lower level nodes are processed in this fashion then
3978 * move up to the next level (parent becomes a lower level node) and
3979 * continue the process for that level.
3981 for (child = le32_to_cpu(dtp->leafidx),
3982 nparent = le32_to_cpu(dtp->nleafs) >> 2;
3983 nparent > 0; nparent >>= 2, child = parent) {
3984 /* get index of 1st node of parent level */
3985 parent = (child - 1) >> 2;
3987 /* set the value of the parent node as the maximum
3988 * of the four nodes of the current level.
3990 for (i = 0, cp = tp + child, cp1 = tp + parent;
3991 i < nparent; i++, cp += 4, cp1++)
4002 * function: initialize dmapctl page
4004 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i)
4005 { /* start leaf index not covered by range */
4008 dcp->nleafs = cpu_to_le32(LPERCTL);
4009 dcp->l2nleafs = cpu_to_le32(L2LPERCTL);
4010 dcp->leafidx = cpu_to_le32(CTLLEAFIND);
4011 dcp->height = cpu_to_le32(5);
4012 dcp->budmin = L2BPERDMAP + L2LPERCTL * level;
4015 * initialize the leaves of current level that were not covered
4016 * by the specified input block range (i.e. the leaves have no
4017 * low level dmapctl or dmap).
4019 cp = &dcp->stree[CTLLEAFIND + i];
4020 for (; i < LPERCTL; i++)
4023 /* build the dmap's binary buddy summary tree */
4024 return (dbInitTree((struct dmaptree *) dcp));
4029 * NAME: dbGetL2AGSize()/ujfs_getagl2size()
4031 * FUNCTION: Determine log2(allocation group size) from aggregate size
4034 * nblocks - Number of blocks in aggregate
4036 * RETURNS: log2(allocation group size) in aggregate blocks
4038 static int dbGetL2AGSize(s64 nblocks)
4044 if (nblocks < BPERDMAP * MAXAG)
4045 return (L2BPERDMAP);
4047 /* round up aggregate size to power of 2 */
4048 m = ((u64) 1 << (64 - 1));
4049 for (l2sz = 64; l2sz >= 0; l2sz--, m >>= 1) {
4054 sz = (s64) 1 << l2sz;
4058 /* agsize = roundupSize/max_number_of_ag */
4059 return (l2sz - L2MAXAG);
4064 * NAME: dbMapFileSizeToMapSize()
4066 * FUNCTION: compute number of blocks the block allocation map file
4067 * can cover from the map file size;
4069 * RETURNS: Number of blocks which can be covered by this block map file;
4073 * maximum number of map pages at each level including control pages
4075 #define MAXL0PAGES (1 + LPERCTL)
4076 #define MAXL1PAGES (1 + LPERCTL * MAXL0PAGES)
4077 #define MAXL2PAGES (1 + LPERCTL * MAXL1PAGES)
4080 * convert number of map pages to the zero origin top dmapctl level
4082 #define BMAPPGTOLEV(npages) \
4083 (((npages) <= 3 + MAXL0PAGES) ? 0 : \
4084 ((npages) <= 2 + MAXL1PAGES) ? 1 : 2)
4086 s64 dbMapFileSizeToMapSize(struct inode * ipbmap)
4088 struct super_block *sb = ipbmap->i_sb;
4092 int complete, factor;
4094 nblocks = ipbmap->i_size >> JFS_SBI(sb)->l2bsize;
4095 npages = nblocks >> JFS_SBI(sb)->l2nbperpage;
4096 level = BMAPPGTOLEV(npages);
4098 /* At each level, accumulate the number of dmap pages covered by
4099 * the number of full child levels below it;
4100 * repeat for the last incomplete child level.
4103 npages--; /* skip the first global control page */
4104 /* skip higher level control pages above top level covered by map */
4105 npages -= (2 - level);
4106 npages--; /* skip top level's control page */
4107 for (i = level; i >= 0; i--) {
4109 (i == 2) ? MAXL1PAGES : ((i == 1) ? MAXL0PAGES : 1);
4110 complete = (u32) npages / factor;
4111 ndmaps += complete * ((i == 2) ? LPERCTL * LPERCTL :
4112 ((i == 1) ? LPERCTL : 1));
4114 /* pages in last/incomplete child */
4115 npages = (u32) npages % factor;
4116 /* skip incomplete child's level control page */
4120 /* convert the number of dmaps into the number of blocks
4121 * which can be covered by the dmaps;
4123 nblocks = ndmaps << L2BPERDMAP;