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);
194 bmp->db_l2nbperpage = le32_to_cpu(dbmp_le->dn_l2nbperpage);
195 bmp->db_numag = le32_to_cpu(dbmp_le->dn_numag);
196 if (!bmp->db_numag) {
198 goto err_release_metapage;
201 bmp->db_maxlevel = le32_to_cpu(dbmp_le->dn_maxlevel);
202 bmp->db_maxag = le32_to_cpu(dbmp_le->dn_maxag);
203 bmp->db_agpref = le32_to_cpu(dbmp_le->dn_agpref);
204 bmp->db_aglevel = le32_to_cpu(dbmp_le->dn_aglevel);
205 bmp->db_agheight = le32_to_cpu(dbmp_le->dn_agheight);
206 bmp->db_agwidth = le32_to_cpu(dbmp_le->dn_agwidth);
207 bmp->db_agstart = le32_to_cpu(dbmp_le->dn_agstart);
208 bmp->db_agl2size = le32_to_cpu(dbmp_le->dn_agl2size);
209 if (bmp->db_agl2size > L2MAXL2SIZE - L2MAXAG ||
210 bmp->db_agl2size < 0) {
212 goto err_release_metapage;
215 if (((bmp->db_mapsize - 1) >> bmp->db_agl2size) > MAXAG) {
217 goto err_release_metapage;
220 for (i = 0; i < MAXAG; i++)
221 bmp->db_agfree[i] = le64_to_cpu(dbmp_le->dn_agfree[i]);
222 bmp->db_agsize = le64_to_cpu(dbmp_le->dn_agsize);
223 bmp->db_maxfreebud = dbmp_le->dn_maxfreebud;
225 /* release the buffer. */
226 release_metapage(mp);
228 /* bind the bmap inode and the bmap descriptor to each other. */
229 bmp->db_ipbmap = ipbmap;
230 JFS_SBI(ipbmap->i_sb)->bmap = bmp;
232 memset(bmp->db_active, 0, sizeof(bmp->db_active));
235 * allocate/initialize the bmap lock
241 err_release_metapage:
242 release_metapage(mp);
252 * FUNCTION: terminate the block allocation map in preparation for
253 * file system unmount.
255 * the in-core bmap descriptor is written to disk and
256 * the memory for this descriptor is freed.
259 * ipbmap - pointer to in-core inode for the block map.
265 int dbUnmount(struct inode *ipbmap, int mounterror)
267 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
269 if (!(mounterror || isReadOnly(ipbmap)))
273 * Invalidate the page cache buffers
275 truncate_inode_pages(ipbmap->i_mapping, 0);
277 /* free the memory for the in-memory bmap. */
286 int dbSync(struct inode *ipbmap)
288 struct dbmap_disk *dbmp_le;
289 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
294 * write bmap global control page
296 /* get the buffer for the on-disk bmap descriptor. */
297 mp = read_metapage(ipbmap,
298 BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
301 jfs_err("dbSync: read_metapage failed!");
304 /* copy the in-memory version of the bmap to the on-disk version */
305 dbmp_le = (struct dbmap_disk *) mp->data;
306 dbmp_le->dn_mapsize = cpu_to_le64(bmp->db_mapsize);
307 dbmp_le->dn_nfree = cpu_to_le64(bmp->db_nfree);
308 dbmp_le->dn_l2nbperpage = cpu_to_le32(bmp->db_l2nbperpage);
309 dbmp_le->dn_numag = cpu_to_le32(bmp->db_numag);
310 dbmp_le->dn_maxlevel = cpu_to_le32(bmp->db_maxlevel);
311 dbmp_le->dn_maxag = cpu_to_le32(bmp->db_maxag);
312 dbmp_le->dn_agpref = cpu_to_le32(bmp->db_agpref);
313 dbmp_le->dn_aglevel = cpu_to_le32(bmp->db_aglevel);
314 dbmp_le->dn_agheight = cpu_to_le32(bmp->db_agheight);
315 dbmp_le->dn_agwidth = cpu_to_le32(bmp->db_agwidth);
316 dbmp_le->dn_agstart = cpu_to_le32(bmp->db_agstart);
317 dbmp_le->dn_agl2size = cpu_to_le32(bmp->db_agl2size);
318 for (i = 0; i < MAXAG; i++)
319 dbmp_le->dn_agfree[i] = cpu_to_le64(bmp->db_agfree[i]);
320 dbmp_le->dn_agsize = cpu_to_le64(bmp->db_agsize);
321 dbmp_le->dn_maxfreebud = bmp->db_maxfreebud;
323 /* write the buffer */
327 * write out dirty pages of bmap
329 filemap_write_and_wait(ipbmap->i_mapping);
331 diWriteSpecial(ipbmap, 0);
339 * FUNCTION: free the specified block range from the working block
342 * the blocks will be free from the working map one dmap
346 * ip - pointer to in-core inode;
347 * blkno - starting block number to be freed.
348 * nblocks - number of blocks to be freed.
354 int dbFree(struct inode *ip, s64 blkno, s64 nblocks)
360 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
361 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
362 struct super_block *sb = ipbmap->i_sb;
364 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
366 /* block to be freed better be within the mapsize. */
367 if (unlikely((blkno == 0) || (blkno + nblocks > bmp->db_mapsize))) {
368 IREAD_UNLOCK(ipbmap);
369 printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
370 (unsigned long long) blkno,
371 (unsigned long long) nblocks);
372 jfs_error(ip->i_sb, "block to be freed is outside the map\n");
377 * TRIM the blocks, when mounted with discard option
379 if (JFS_SBI(sb)->flag & JFS_DISCARD)
380 if (JFS_SBI(sb)->minblks_trim <= nblocks)
381 jfs_issue_discard(ipbmap, blkno, nblocks);
384 * free the blocks a dmap at a time.
387 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
388 /* release previous dmap if any */
393 /* get the buffer for the current dmap. */
394 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
395 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
397 IREAD_UNLOCK(ipbmap);
400 dp = (struct dmap *) mp->data;
402 /* determine the number of blocks to be freed from
405 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
407 /* free the blocks. */
408 if ((rc = dbFreeDmap(bmp, dp, blkno, nb))) {
409 jfs_error(ip->i_sb, "error in block map\n");
410 release_metapage(mp);
411 IREAD_UNLOCK(ipbmap);
416 /* write the last buffer. */
420 IREAD_UNLOCK(ipbmap);
427 * NAME: dbUpdatePMap()
429 * FUNCTION: update the allocation state (free or allocate) of the
430 * specified block range in the persistent block allocation map.
432 * the blocks will be updated in the persistent map one
436 * ipbmap - pointer to in-core inode for the block map.
437 * free - 'true' if block range is to be freed from the persistent
438 * map; 'false' if it is to be allocated.
439 * blkno - starting block number of the range.
440 * nblocks - number of contiguous blocks in the range.
441 * tblk - transaction block;
448 dbUpdatePMap(struct inode *ipbmap,
449 int free, s64 blkno, s64 nblocks, struct tblock * tblk)
451 int nblks, dbitno, wbitno, rbits;
452 int word, nbits, nwords;
453 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
454 s64 lblkno, rem, lastlblkno;
459 int lsn, difft, diffp;
462 /* the blocks better be within the mapsize. */
463 if (blkno + nblocks > bmp->db_mapsize) {
464 printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
465 (unsigned long long) blkno,
466 (unsigned long long) nblocks);
467 jfs_error(ipbmap->i_sb, "blocks are outside the map\n");
471 /* compute delta of transaction lsn from log syncpt */
473 log = (struct jfs_log *) JFS_SBI(tblk->sb)->log;
474 logdiff(difft, lsn, log);
477 * update the block state a dmap at a time.
481 for (rem = nblocks; rem > 0; rem -= nblks, blkno += nblks) {
482 /* get the buffer for the current dmap. */
483 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
484 if (lblkno != lastlblkno) {
489 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE,
493 metapage_wait_for_io(mp);
495 dp = (struct dmap *) mp->data;
497 /* determine the bit number and word within the dmap of
498 * the starting block. also determine how many blocks
499 * are to be updated within this dmap.
501 dbitno = blkno & (BPERDMAP - 1);
502 word = dbitno >> L2DBWORD;
503 nblks = min(rem, (s64)BPERDMAP - dbitno);
505 /* update the bits of the dmap words. the first and last
506 * words may only have a subset of their bits updated. if
507 * this is the case, we'll work against that word (i.e.
508 * partial first and/or last) only in a single pass. a
509 * single pass will also be used to update all words that
510 * are to have all their bits updated.
512 for (rbits = nblks; rbits > 0;
513 rbits -= nbits, dbitno += nbits) {
514 /* determine the bit number within the word and
515 * the number of bits within the word.
517 wbitno = dbitno & (DBWORD - 1);
518 nbits = min(rbits, DBWORD - wbitno);
520 /* check if only part of the word is to be updated. */
521 if (nbits < DBWORD) {
522 /* update (free or allocate) the bits
526 (ONES << (DBWORD - nbits) >> wbitno);
536 /* one or more words are to have all
537 * their bits updated. determine how
538 * many words and how many bits.
540 nwords = rbits >> L2DBWORD;
541 nbits = nwords << L2DBWORD;
543 /* update (free or allocate) the bits
547 memset(&dp->pmap[word], 0,
550 memset(&dp->pmap[word], (int) ONES,
560 if (lblkno == lastlblkno)
565 LOGSYNC_LOCK(log, flags);
567 /* inherit older/smaller lsn */
568 logdiff(diffp, mp->lsn, log);
572 /* move bp after tblock in logsync list */
573 list_move(&mp->synclist, &tblk->synclist);
576 /* inherit younger/larger clsn */
577 logdiff(difft, tblk->clsn, log);
578 logdiff(diffp, mp->clsn, log);
580 mp->clsn = tblk->clsn;
585 /* insert bp after tblock in logsync list */
587 list_add(&mp->synclist, &tblk->synclist);
589 mp->clsn = tblk->clsn;
591 LOGSYNC_UNLOCK(log, flags);
594 /* write the last buffer. */
606 * FUNCTION: find the preferred allocation group for new allocations.
608 * Within the allocation groups, we maintain a preferred
609 * allocation group which consists of a group with at least
610 * average free space. It is the preferred group that we target
611 * new inode allocation towards. The tie-in between inode
612 * allocation and block allocation occurs as we allocate the
613 * first (data) block of an inode and specify the inode (block)
614 * as the allocation hint for this block.
616 * We try to avoid having more than one open file growing in
617 * an allocation group, as this will lead to fragmentation.
618 * This differs from the old OS/2 method of trying to keep
619 * empty ags around for large allocations.
622 * ipbmap - pointer to in-core inode for the block map.
625 * the preferred allocation group number.
627 int dbNextAG(struct inode *ipbmap)
634 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
638 /* determine the average number of free blocks within the ags. */
639 avgfree = (u32)bmp->db_nfree / bmp->db_numag;
642 * if the current preferred ag does not have an active allocator
643 * and has at least average freespace, return it
645 agpref = bmp->db_agpref;
646 if ((atomic_read(&bmp->db_active[agpref]) == 0) &&
647 (bmp->db_agfree[agpref] >= avgfree))
650 /* From the last preferred ag, find the next one with at least
651 * average free space.
653 for (i = 0 ; i < bmp->db_numag; i++, agpref++) {
654 if (agpref == bmp->db_numag)
657 if (atomic_read(&bmp->db_active[agpref]))
658 /* open file is currently growing in this ag */
660 if (bmp->db_agfree[agpref] >= avgfree) {
661 /* Return this one */
662 bmp->db_agpref = agpref;
664 } else if (bmp->db_agfree[agpref] > hwm) {
665 /* Less than avg. freespace, but best so far */
666 hwm = bmp->db_agfree[agpref];
672 * If no inactive ag was found with average freespace, use the
676 bmp->db_agpref = next_best;
677 /* else leave db_agpref unchanged */
681 /* return the preferred group.
683 return (bmp->db_agpref);
689 * FUNCTION: attempt to allocate a specified number of contiguous free
690 * blocks from the working allocation block map.
692 * the block allocation policy uses hints and a multi-step
695 * for allocation requests smaller than the number of blocks
696 * per dmap, we first try to allocate the new blocks
697 * immediately following the hint. if these blocks are not
698 * available, we try to allocate blocks near the hint. if
699 * no blocks near the hint are available, we next try to
700 * allocate within the same dmap as contains the hint.
702 * if no blocks are available in the dmap or the allocation
703 * request is larger than the dmap size, we try to allocate
704 * within the same allocation group as contains the hint. if
705 * this does not succeed, we finally try to allocate anywhere
706 * within the aggregate.
708 * we also try to allocate anywhere within the aggregate for
709 * for allocation requests larger than the allocation group
710 * size or requests that specify no hint value.
713 * ip - pointer to in-core inode;
714 * hint - allocation hint.
715 * nblocks - number of contiguous blocks in the range.
716 * results - on successful return, set to the starting block number
717 * of the newly allocated contiguous range.
721 * -ENOSPC - insufficient disk resources
724 int dbAlloc(struct inode *ip, s64 hint, s64 nblocks, s64 * results)
727 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
736 /* assert that nblocks is valid */
739 /* get the log2 number of blocks to be allocated.
740 * if the number of blocks is not a log2 multiple,
741 * it will be rounded up to the next log2 multiple.
743 l2nb = BLKSTOL2(nblocks);
745 bmp = JFS_SBI(ip->i_sb)->bmap;
747 mapSize = bmp->db_mapsize;
749 /* the hint should be within the map */
750 if (hint >= mapSize) {
751 jfs_error(ip->i_sb, "the hint is outside the map\n");
755 /* if the number of blocks to be allocated is greater than the
756 * allocation group size, try to allocate anywhere.
758 if (l2nb > bmp->db_agl2size) {
759 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
761 rc = dbAllocAny(bmp, nblocks, l2nb, results);
767 * If no hint, let dbNextAG recommend an allocation group
772 /* we would like to allocate close to the hint. adjust the
773 * hint to the block following the hint since the allocators
774 * will start looking for free space starting at this point.
778 if (blkno >= bmp->db_mapsize)
781 agno = blkno >> bmp->db_agl2size;
783 /* check if blkno crosses over into a new allocation group.
784 * if so, check if we should allow allocations within this
787 if ((blkno & (bmp->db_agsize - 1)) == 0)
788 /* check if the AG is currently being written to.
789 * if so, call dbNextAG() to find a non-busy
790 * AG with sufficient free space.
792 if (atomic_read(&bmp->db_active[agno]))
795 /* check if the allocation request size can be satisfied from a
796 * single dmap. if so, try to allocate from the dmap containing
797 * the hint using a tiered strategy.
799 if (nblocks <= BPERDMAP) {
800 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
802 /* get the buffer for the dmap containing the hint.
805 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
806 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
810 dp = (struct dmap *) mp->data;
812 /* first, try to satisfy the allocation request with the
813 * blocks beginning at the hint.
815 if ((rc = dbAllocNext(bmp, dp, blkno, (int) nblocks))
819 mark_metapage_dirty(mp);
822 release_metapage(mp);
826 writers = atomic_read(&bmp->db_active[agno]);
828 ((writers == 1) && (JFS_IP(ip)->active_ag != agno))) {
830 * Someone else is writing in this allocation
831 * group. To avoid fragmenting, try another ag
833 release_metapage(mp);
834 IREAD_UNLOCK(ipbmap);
838 /* next, try to satisfy the allocation request with blocks
842 dbAllocNear(bmp, dp, blkno, (int) nblocks, l2nb, results))
845 mark_metapage_dirty(mp);
847 release_metapage(mp);
851 /* try to satisfy the allocation request with blocks within
852 * the same dmap as the hint.
854 if ((rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results))
857 mark_metapage_dirty(mp);
859 release_metapage(mp);
863 release_metapage(mp);
864 IREAD_UNLOCK(ipbmap);
867 /* try to satisfy the allocation request with blocks within
868 * the same allocation group as the hint.
870 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
871 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) != -ENOSPC)
874 IWRITE_UNLOCK(ipbmap);
879 * Let dbNextAG recommend a preferred allocation group
881 agno = dbNextAG(ipbmap);
882 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
884 /* Try to allocate within this allocation group. if that fails, try to
885 * allocate anywhere in the map.
887 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) == -ENOSPC)
888 rc = dbAllocAny(bmp, nblocks, l2nb, results);
891 IWRITE_UNLOCK(ipbmap);
896 IREAD_UNLOCK(ipbmap);
903 * NAME: dbAllocExact()
905 * FUNCTION: try to allocate the requested extent;
908 * ip - pointer to in-core inode;
909 * blkno - extent address;
910 * nblocks - extent length;
914 * -ENOSPC - insufficient disk resources
917 int dbAllocExact(struct inode *ip, s64 blkno, int nblocks)
920 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
921 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
926 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
929 * validate extent request:
931 * note: defragfs policy:
932 * max 64 blocks will be moved.
933 * allocation request size must be satisfied from a single dmap.
935 if (nblocks <= 0 || nblocks > BPERDMAP || blkno >= bmp->db_mapsize) {
936 IREAD_UNLOCK(ipbmap);
940 if (nblocks > ((s64) 1 << bmp->db_maxfreebud)) {
941 /* the free space is no longer available */
942 IREAD_UNLOCK(ipbmap);
946 /* read in the dmap covering the extent */
947 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
948 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
950 IREAD_UNLOCK(ipbmap);
953 dp = (struct dmap *) mp->data;
955 /* try to allocate the requested extent */
956 rc = dbAllocNext(bmp, dp, blkno, nblocks);
958 IREAD_UNLOCK(ipbmap);
961 mark_metapage_dirty(mp);
963 release_metapage(mp);
972 * FUNCTION: attempt to extend a current allocation by a specified
975 * this routine attempts to satisfy the allocation request
976 * by first trying to extend the existing allocation in
977 * place by allocating the additional blocks as the blocks
978 * immediately following the current allocation. if these
979 * blocks are not available, this routine will attempt to
980 * allocate a new set of contiguous blocks large enough
981 * to cover the existing allocation plus the additional
982 * number of blocks required.
985 * ip - pointer to in-core inode requiring allocation.
986 * blkno - starting block of the current allocation.
987 * nblocks - number of contiguous blocks within the current
989 * addnblocks - number of blocks to add to the allocation.
990 * results - on successful return, set to the starting block number
991 * of the existing allocation if the existing allocation
992 * was extended in place or to a newly allocated contiguous
993 * range if the existing allocation could not be extended
998 * -ENOSPC - insufficient disk resources
1002 dbReAlloc(struct inode *ip,
1003 s64 blkno, s64 nblocks, s64 addnblocks, s64 * results)
1007 /* try to extend the allocation in place.
1009 if ((rc = dbExtend(ip, blkno, nblocks, addnblocks)) == 0) {
1017 /* could not extend the allocation in place, so allocate a
1018 * new set of blocks for the entire request (i.e. try to get
1019 * a range of contiguous blocks large enough to cover the
1020 * existing allocation plus the additional blocks.)
1023 (ip, blkno + nblocks - 1, addnblocks + nblocks, results));
1030 * FUNCTION: attempt to extend a current allocation by a specified
1033 * this routine attempts to satisfy the allocation request
1034 * by first trying to extend the existing allocation in
1035 * place by allocating the additional blocks as the blocks
1036 * immediately following the current allocation.
1039 * ip - pointer to in-core inode requiring allocation.
1040 * blkno - starting block of the current allocation.
1041 * nblocks - number of contiguous blocks within the current
1043 * addnblocks - number of blocks to add to the allocation.
1047 * -ENOSPC - insufficient disk resources
1050 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks)
1052 struct jfs_sb_info *sbi = JFS_SBI(ip->i_sb);
1053 s64 lblkno, lastblkno, extblkno;
1055 struct metapage *mp;
1058 struct inode *ipbmap = sbi->ipbmap;
1062 * We don't want a non-aligned extent to cross a page boundary
1064 if (((rel_block = blkno & (sbi->nbperpage - 1))) &&
1065 (rel_block + nblocks + addnblocks > sbi->nbperpage))
1068 /* get the last block of the current allocation */
1069 lastblkno = blkno + nblocks - 1;
1071 /* determine the block number of the block following
1072 * the existing allocation.
1074 extblkno = lastblkno + 1;
1076 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
1078 /* better be within the file system */
1080 if (lastblkno < 0 || lastblkno >= bmp->db_mapsize) {
1081 IREAD_UNLOCK(ipbmap);
1082 jfs_error(ip->i_sb, "the block is outside the filesystem\n");
1086 /* we'll attempt to extend the current allocation in place by
1087 * allocating the additional blocks as the blocks immediately
1088 * following the current allocation. we only try to extend the
1089 * current allocation in place if the number of additional blocks
1090 * can fit into a dmap, the last block of the current allocation
1091 * is not the last block of the file system, and the start of the
1092 * inplace extension is not on an allocation group boundary.
1094 if (addnblocks > BPERDMAP || extblkno >= bmp->db_mapsize ||
1095 (extblkno & (bmp->db_agsize - 1)) == 0) {
1096 IREAD_UNLOCK(ipbmap);
1100 /* get the buffer for the dmap containing the first block
1103 lblkno = BLKTODMAP(extblkno, bmp->db_l2nbperpage);
1104 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
1106 IREAD_UNLOCK(ipbmap);
1110 dp = (struct dmap *) mp->data;
1112 /* try to allocate the blocks immediately following the
1113 * current allocation.
1115 rc = dbAllocNext(bmp, dp, extblkno, (int) addnblocks);
1117 IREAD_UNLOCK(ipbmap);
1119 /* were we successful ? */
1123 /* we were not successful */
1124 release_metapage(mp);
1131 * NAME: dbAllocNext()
1133 * FUNCTION: attempt to allocate the blocks of the specified block
1134 * range within a dmap.
1137 * bmp - pointer to bmap descriptor
1138 * dp - pointer to dmap.
1139 * blkno - starting block number of the range.
1140 * nblocks - number of contiguous free blocks of the range.
1144 * -ENOSPC - insufficient disk resources
1147 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1149 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
1152 int dbitno, word, rembits, nb, nwords, wbitno, nw;
1157 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1158 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1162 /* pick up a pointer to the leaves of the dmap tree.
1164 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1166 /* determine the bit number and word within the dmap of the
1169 dbitno = blkno & (BPERDMAP - 1);
1170 word = dbitno >> L2DBWORD;
1172 /* check if the specified block range is contained within
1175 if (dbitno + nblocks > BPERDMAP)
1178 /* check if the starting leaf indicates that anything
1181 if (leaf[word] == NOFREE)
1184 /* check the dmaps words corresponding to block range to see
1185 * if the block range is free. not all bits of the first and
1186 * last words may be contained within the block range. if this
1187 * is the case, we'll work against those words (i.e. partial first
1188 * and/or last) on an individual basis (a single pass) and examine
1189 * the actual bits to determine if they are free. a single pass
1190 * will be used for all dmap words fully contained within the
1191 * specified range. within this pass, the leaves of the dmap
1192 * tree will be examined to determine if the blocks are free. a
1193 * single leaf may describe the free space of multiple dmap
1194 * words, so we may visit only a subset of the actual leaves
1195 * corresponding to the dmap words of the block range.
1197 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
1198 /* determine the bit number within the word and
1199 * the number of bits within the word.
1201 wbitno = dbitno & (DBWORD - 1);
1202 nb = min(rembits, DBWORD - wbitno);
1204 /* check if only part of the word is to be examined.
1207 /* check if the bits are free.
1209 mask = (ONES << (DBWORD - nb) >> wbitno);
1210 if ((mask & ~le32_to_cpu(dp->wmap[word])) != mask)
1215 /* one or more dmap words are fully contained
1216 * within the block range. determine how many
1217 * words and how many bits.
1219 nwords = rembits >> L2DBWORD;
1220 nb = nwords << L2DBWORD;
1222 /* now examine the appropriate leaves to determine
1223 * if the blocks are free.
1225 while (nwords > 0) {
1226 /* does the leaf describe any free space ?
1228 if (leaf[word] < BUDMIN)
1231 /* determine the l2 number of bits provided
1235 min_t(int, leaf[word], NLSTOL2BSZ(nwords));
1237 /* determine how many words were handled.
1239 nw = BUDSIZE(l2size, BUDMIN);
1247 /* allocate the blocks.
1249 return (dbAllocDmap(bmp, dp, blkno, nblocks));
1254 * NAME: dbAllocNear()
1256 * FUNCTION: attempt to allocate a number of contiguous free blocks near
1257 * a specified block (hint) within a dmap.
1259 * starting with the dmap leaf that covers the hint, we'll
1260 * check the next four contiguous leaves for sufficient free
1261 * space. if sufficient free space is found, we'll allocate
1262 * the desired free space.
1265 * bmp - pointer to bmap descriptor
1266 * dp - pointer to dmap.
1267 * blkno - block number to allocate near.
1268 * nblocks - actual number of contiguous free blocks desired.
1269 * l2nb - log2 number of contiguous free blocks desired.
1270 * results - on successful return, set to the starting block number
1271 * of the newly allocated range.
1275 * -ENOSPC - insufficient disk resources
1278 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1281 dbAllocNear(struct bmap * bmp,
1282 struct dmap * dp, s64 blkno, int nblocks, int l2nb, s64 * results)
1284 int word, lword, rc;
1287 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1288 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1292 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1294 /* determine the word within the dmap that holds the hint
1295 * (i.e. blkno). also, determine the last word in the dmap
1296 * that we'll include in our examination.
1298 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
1299 lword = min(word + 4, LPERDMAP);
1301 /* examine the leaves for sufficient free space.
1303 for (; word < lword; word++) {
1304 /* does the leaf describe sufficient free space ?
1306 if (leaf[word] < l2nb)
1309 /* determine the block number within the file system
1310 * of the first block described by this dmap word.
1312 blkno = le64_to_cpu(dp->start) + (word << L2DBWORD);
1314 /* if not all bits of the dmap word are free, get the
1315 * starting bit number within the dmap word of the required
1316 * string of free bits and adjust the block number with the
1319 if (leaf[word] < BUDMIN)
1321 dbFindBits(le32_to_cpu(dp->wmap[word]), l2nb);
1323 /* allocate the blocks.
1325 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1338 * FUNCTION: attempt to allocate the specified number of contiguous
1339 * free blocks within the specified allocation group.
1341 * unless the allocation group size is equal to the number
1342 * of blocks per dmap, the dmap control pages will be used to
1343 * find the required free space, if available. we start the
1344 * search at the highest dmap control page level which
1345 * distinctly describes the allocation group's free space
1346 * (i.e. the highest level at which the allocation group's
1347 * free space is not mixed in with that of any other group).
1348 * in addition, we start the search within this level at a
1349 * height of the dmapctl dmtree at which the nodes distinctly
1350 * describe the allocation group's free space. at this height,
1351 * the allocation group's free space may be represented by 1
1352 * or two sub-trees, depending on the allocation group size.
1353 * we search the top nodes of these subtrees left to right for
1354 * sufficient free space. if sufficient free space is found,
1355 * the subtree is searched to find the leftmost leaf that
1356 * has free space. once we have made it to the leaf, we
1357 * move the search to the next lower level dmap control page
1358 * corresponding to this leaf. we continue down the dmap control
1359 * pages until we find the dmap that contains or starts the
1360 * sufficient free space and we allocate at this dmap.
1362 * if the allocation group size is equal to the dmap size,
1363 * we'll start at the dmap corresponding to the allocation
1364 * group and attempt the allocation at this level.
1366 * the dmap control page search is also not performed if the
1367 * allocation group is completely free and we go to the first
1368 * dmap of the allocation group to do the allocation. this is
1369 * done because the allocation group may be part (not the first
1370 * part) of a larger binary buddy system, causing the dmap
1371 * control pages to indicate no free space (NOFREE) within
1372 * the allocation group.
1375 * bmp - pointer to bmap descriptor
1376 * agno - allocation group number.
1377 * nblocks - actual number of contiguous free blocks desired.
1378 * l2nb - log2 number of contiguous free blocks desired.
1379 * results - on successful return, set to the starting block number
1380 * of the newly allocated range.
1384 * -ENOSPC - insufficient disk resources
1387 * note: IWRITE_LOCK(ipmap) held on entry/exit;
1390 dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, s64 * results)
1392 struct metapage *mp;
1393 struct dmapctl *dcp;
1394 int rc, ti, i, k, m, n, agperlev;
1398 /* allocation request should not be for more than the
1399 * allocation group size.
1401 if (l2nb > bmp->db_agl2size) {
1402 jfs_error(bmp->db_ipbmap->i_sb,
1403 "allocation request is larger than the allocation group size\n");
1407 /* determine the starting block number of the allocation
1410 blkno = (s64) agno << bmp->db_agl2size;
1412 /* check if the allocation group size is the minimum allocation
1413 * group size or if the allocation group is completely free. if
1414 * the allocation group size is the minimum size of BPERDMAP (i.e.
1415 * 1 dmap), there is no need to search the dmap control page (below)
1416 * that fully describes the allocation group since the allocation
1417 * group is already fully described by a dmap. in this case, we
1418 * just call dbAllocCtl() to search the dmap tree and allocate the
1419 * required space if available.
1421 * if the allocation group is completely free, dbAllocCtl() is
1422 * also called to allocate the required space. this is done for
1423 * two reasons. first, it makes no sense searching the dmap control
1424 * pages for free space when we know that free space exists. second,
1425 * the dmap control pages may indicate that the allocation group
1426 * has no free space if the allocation group is part (not the first
1427 * part) of a larger binary buddy system.
1429 if (bmp->db_agsize == BPERDMAP
1430 || bmp->db_agfree[agno] == bmp->db_agsize) {
1431 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1432 if ((rc == -ENOSPC) &&
1433 (bmp->db_agfree[agno] == bmp->db_agsize)) {
1434 printk(KERN_ERR "blkno = %Lx, blocks = %Lx\n",
1435 (unsigned long long) blkno,
1436 (unsigned long long) nblocks);
1437 jfs_error(bmp->db_ipbmap->i_sb,
1438 "dbAllocCtl failed in free AG\n");
1443 /* the buffer for the dmap control page that fully describes the
1446 lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, bmp->db_aglevel);
1447 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1450 dcp = (struct dmapctl *) mp->data;
1451 budmin = dcp->budmin;
1453 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1454 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
1455 release_metapage(mp);
1459 /* search the subtree(s) of the dmap control page that describes
1460 * the allocation group, looking for sufficient free space. to begin,
1461 * determine how many allocation groups are represented in a dmap
1462 * control page at the control page level (i.e. L0, L1, L2) that
1463 * fully describes an allocation group. next, determine the starting
1464 * tree index of this allocation group within the control page.
1467 (1 << (L2LPERCTL - (bmp->db_agheight << 1))) / bmp->db_agwidth;
1468 ti = bmp->db_agstart + bmp->db_agwidth * (agno & (agperlev - 1));
1470 /* dmap control page trees fan-out by 4 and a single allocation
1471 * group may be described by 1 or 2 subtrees within the ag level
1472 * dmap control page, depending upon the ag size. examine the ag's
1473 * subtrees for sufficient free space, starting with the leftmost
1476 for (i = 0; i < bmp->db_agwidth; i++, ti++) {
1477 /* is there sufficient free space ?
1479 if (l2nb > dcp->stree[ti])
1482 /* sufficient free space found in a subtree. now search down
1483 * the subtree to find the leftmost leaf that describes this
1486 for (k = bmp->db_agheight; k > 0; k--) {
1487 for (n = 0, m = (ti << 2) + 1; n < 4; n++) {
1488 if (l2nb <= dcp->stree[m + n]) {
1494 jfs_error(bmp->db_ipbmap->i_sb,
1495 "failed descending stree\n");
1496 release_metapage(mp);
1501 /* determine the block number within the file system
1502 * that corresponds to this leaf.
1504 if (bmp->db_aglevel == 2)
1506 else if (bmp->db_aglevel == 1)
1507 blkno &= ~(MAXL1SIZE - 1);
1508 else /* bmp->db_aglevel == 0 */
1509 blkno &= ~(MAXL0SIZE - 1);
1512 ((s64) (ti - le32_to_cpu(dcp->leafidx))) << budmin;
1514 /* release the buffer in preparation for going down
1515 * the next level of dmap control pages.
1517 release_metapage(mp);
1519 /* check if we need to continue to search down the lower
1520 * level dmap control pages. we need to if the number of
1521 * blocks required is less than maximum number of blocks
1522 * described at the next lower level.
1524 if (l2nb < budmin) {
1526 /* search the lower level dmap control pages to get
1527 * the starting block number of the dmap that
1528 * contains or starts off the free space.
1531 dbFindCtl(bmp, l2nb, bmp->db_aglevel - 1,
1533 if (rc == -ENOSPC) {
1534 jfs_error(bmp->db_ipbmap->i_sb,
1535 "control page inconsistent\n");
1542 /* allocate the blocks.
1544 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1545 if (rc == -ENOSPC) {
1546 jfs_error(bmp->db_ipbmap->i_sb,
1547 "unable to allocate blocks\n");
1553 /* no space in the allocation group. release the buffer and
1556 release_metapage(mp);
1563 * NAME: dbAllocAny()
1565 * FUNCTION: attempt to allocate the specified number of contiguous
1566 * free blocks anywhere in the file system.
1568 * dbAllocAny() attempts to find the sufficient free space by
1569 * searching down the dmap control pages, starting with the
1570 * highest level (i.e. L0, L1, L2) control page. if free space
1571 * large enough to satisfy the desired free space is found, the
1572 * desired free space is allocated.
1575 * bmp - pointer to bmap descriptor
1576 * nblocks - actual number of contiguous free blocks desired.
1577 * l2nb - log2 number of contiguous free blocks desired.
1578 * results - on successful return, set to the starting block number
1579 * of the newly allocated range.
1583 * -ENOSPC - insufficient disk resources
1586 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1588 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results)
1593 /* starting with the top level dmap control page, search
1594 * down the dmap control levels for sufficient free space.
1595 * if free space is found, dbFindCtl() returns the starting
1596 * block number of the dmap that contains or starts off the
1597 * range of free space.
1599 if ((rc = dbFindCtl(bmp, l2nb, bmp->db_maxlevel, &blkno)))
1602 /* allocate the blocks.
1604 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1605 if (rc == -ENOSPC) {
1606 jfs_error(bmp->db_ipbmap->i_sb, "unable to allocate blocks\n");
1614 * NAME: dbDiscardAG()
1616 * FUNCTION: attempt to discard (TRIM) all free blocks of specific AG
1619 * 1) allocate blocks, as large as possible and save them
1620 * while holding IWRITE_LOCK on ipbmap
1621 * 2) trim all these saved block/length values
1622 * 3) mark the blocks free again
1625 * - we work only on one ag at some time, minimizing how long we
1626 * need to lock ipbmap
1627 * - reading / writing the fs is possible most time, even on
1631 * - we write two times to the dmapctl and dmap pages
1632 * - but for me, this seems the best way, better ideas?
1636 * ip - pointer to in-core inode
1638 * minlen - minimum value of contiguous blocks
1641 * s64 - actual number of blocks trimmed
1643 s64 dbDiscardAG(struct inode *ip, int agno, s64 minlen)
1645 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
1646 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
1650 struct super_block *sb = ipbmap->i_sb;
1657 /* max blkno / nblocks pairs to trim */
1658 int count = 0, range_cnt;
1661 /* prevent others from writing new stuff here, while trimming */
1662 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
1664 nblocks = bmp->db_agfree[agno];
1665 max_ranges = nblocks;
1666 do_div(max_ranges, minlen);
1667 range_cnt = min_t(u64, max_ranges + 1, 32 * 1024);
1668 totrim = kmalloc_array(range_cnt, sizeof(struct range2trim), GFP_NOFS);
1669 if (totrim == NULL) {
1670 jfs_error(bmp->db_ipbmap->i_sb, "no memory for trim array\n");
1671 IWRITE_UNLOCK(ipbmap);
1676 while (nblocks >= minlen) {
1677 l2nb = BLKSTOL2(nblocks);
1679 /* 0 = okay, -EIO = fatal, -ENOSPC -> try smaller block */
1680 rc = dbAllocAG(bmp, agno, nblocks, l2nb, &blkno);
1683 tt->nblocks = nblocks;
1686 /* the whole ag is free, trim now */
1687 if (bmp->db_agfree[agno] == 0)
1690 /* give a hint for the next while */
1691 nblocks = bmp->db_agfree[agno];
1693 } else if (rc == -ENOSPC) {
1694 /* search for next smaller log2 block */
1695 l2nb = BLKSTOL2(nblocks) - 1;
1696 nblocks = 1LL << l2nb;
1698 /* Trim any already allocated blocks */
1699 jfs_error(bmp->db_ipbmap->i_sb, "-EIO\n");
1703 /* check, if our trim array is full */
1704 if (unlikely(count >= range_cnt - 1))
1707 IWRITE_UNLOCK(ipbmap);
1709 tt->nblocks = 0; /* mark the current end */
1710 for (tt = totrim; tt->nblocks != 0; tt++) {
1711 /* when mounted with online discard, dbFree() will
1712 * call jfs_issue_discard() itself */
1713 if (!(JFS_SBI(sb)->flag & JFS_DISCARD))
1714 jfs_issue_discard(ip, tt->blkno, tt->nblocks);
1715 dbFree(ip, tt->blkno, tt->nblocks);
1716 trimmed += tt->nblocks;
1726 * FUNCTION: starting at a specified dmap control page level and block
1727 * number, search down the dmap control levels for a range of
1728 * contiguous free blocks large enough to satisfy an allocation
1729 * request for the specified number of free blocks.
1731 * if sufficient contiguous free blocks are found, this routine
1732 * returns the starting block number within a dmap page that
1733 * contains or starts a range of contiqious free blocks that
1734 * is sufficient in size.
1737 * bmp - pointer to bmap descriptor
1738 * level - starting dmap control page level.
1739 * l2nb - log2 number of contiguous free blocks desired.
1740 * *blkno - on entry, starting block number for conducting the search.
1741 * on successful return, the first block within a dmap page
1742 * that contains or starts a range of contiguous free blocks.
1746 * -ENOSPC - insufficient disk resources
1749 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1751 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno)
1753 int rc, leafidx, lev;
1755 struct dmapctl *dcp;
1757 struct metapage *mp;
1759 /* starting at the specified dmap control page level and block
1760 * number, search down the dmap control levels for the starting
1761 * block number of a dmap page that contains or starts off
1762 * sufficient free blocks.
1764 for (lev = level, b = *blkno; lev >= 0; lev--) {
1765 /* get the buffer of the dmap control page for the block
1766 * number and level (i.e. L0, L1, L2).
1768 lblkno = BLKTOCTL(b, bmp->db_l2nbperpage, lev);
1769 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1772 dcp = (struct dmapctl *) mp->data;
1773 budmin = dcp->budmin;
1775 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1776 jfs_error(bmp->db_ipbmap->i_sb,
1777 "Corrupt dmapctl page\n");
1778 release_metapage(mp);
1782 /* search the tree within the dmap control page for
1783 * sufficient free space. if sufficient free space is found,
1784 * dbFindLeaf() returns the index of the leaf at which
1785 * free space was found.
1787 rc = dbFindLeaf((dmtree_t *) dcp, l2nb, &leafidx);
1789 /* release the buffer.
1791 release_metapage(mp);
1797 jfs_error(bmp->db_ipbmap->i_sb,
1798 "dmap inconsistent\n");
1804 /* adjust the block number to reflect the location within
1805 * the dmap control page (i.e. the leaf) at which free
1808 b += (((s64) leafidx) << budmin);
1810 /* we stop the search at this dmap control page level if
1811 * the number of blocks required is greater than or equal
1812 * to the maximum number of blocks described at the next
1825 * NAME: dbAllocCtl()
1827 * FUNCTION: attempt to allocate a specified number of contiguous
1828 * blocks starting within a specific dmap.
1830 * this routine is called by higher level routines that search
1831 * the dmap control pages above the actual dmaps for contiguous
1832 * free space. the result of successful searches by these
1833 * routines are the starting block numbers within dmaps, with
1834 * the dmaps themselves containing the desired contiguous free
1835 * space or starting a contiguous free space of desired size
1836 * that is made up of the blocks of one or more dmaps. these
1837 * calls should not fail due to insufficent resources.
1839 * this routine is called in some cases where it is not known
1840 * whether it will fail due to insufficient resources. more
1841 * specifically, this occurs when allocating from an allocation
1842 * group whose size is equal to the number of blocks per dmap.
1843 * in this case, the dmap control pages are not examined prior
1844 * to calling this routine (to save pathlength) and the call
1847 * for a request size that fits within a dmap, this routine relies
1848 * upon the dmap's dmtree to find the requested contiguous free
1849 * space. for request sizes that are larger than a dmap, the
1850 * requested free space will start at the first block of the
1851 * first dmap (i.e. blkno).
1854 * bmp - pointer to bmap descriptor
1855 * nblocks - actual number of contiguous free blocks to allocate.
1856 * l2nb - log2 number of contiguous free blocks to allocate.
1857 * blkno - starting block number of the dmap to start the allocation
1859 * results - on successful return, set to the starting block number
1860 * of the newly allocated range.
1864 * -ENOSPC - insufficient disk resources
1867 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1870 dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, s64 * results)
1874 struct metapage *mp;
1877 /* check if the allocation request is confined to a single dmap.
1879 if (l2nb <= L2BPERDMAP) {
1880 /* get the buffer for the dmap.
1882 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
1883 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1886 dp = (struct dmap *) mp->data;
1888 /* try to allocate the blocks.
1890 rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results);
1892 mark_metapage_dirty(mp);
1894 release_metapage(mp);
1899 /* allocation request involving multiple dmaps. it must start on
1902 assert((blkno & (BPERDMAP - 1)) == 0);
1904 /* allocate the blocks dmap by dmap.
1906 for (n = nblocks, b = blkno; n > 0; n -= nb, b += nb) {
1907 /* get the buffer for the dmap.
1909 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1910 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1915 dp = (struct dmap *) mp->data;
1917 /* the dmap better be all free.
1919 if (dp->tree.stree[ROOT] != L2BPERDMAP) {
1920 release_metapage(mp);
1921 jfs_error(bmp->db_ipbmap->i_sb,
1922 "the dmap is not all free\n");
1927 /* determine how many blocks to allocate from this dmap.
1929 nb = min_t(s64, n, BPERDMAP);
1931 /* allocate the blocks from the dmap.
1933 if ((rc = dbAllocDmap(bmp, dp, b, nb))) {
1934 release_metapage(mp);
1938 /* write the buffer.
1943 /* set the results (starting block number) and return.
1948 /* something failed in handling an allocation request involving
1949 * multiple dmaps. we'll try to clean up by backing out any
1950 * allocation that has already happened for this request. if
1951 * we fail in backing out the allocation, we'll mark the file
1952 * system to indicate that blocks have been leaked.
1956 /* try to backout the allocations dmap by dmap.
1958 for (n = nblocks - n, b = blkno; n > 0;
1959 n -= BPERDMAP, b += BPERDMAP) {
1960 /* get the buffer for this dmap.
1962 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1963 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1965 /* could not back out. mark the file system
1966 * to indicate that we have leaked blocks.
1968 jfs_error(bmp->db_ipbmap->i_sb,
1969 "I/O Error: Block Leakage\n");
1972 dp = (struct dmap *) mp->data;
1974 /* free the blocks is this dmap.
1976 if (dbFreeDmap(bmp, dp, b, BPERDMAP)) {
1977 /* could not back out. mark the file system
1978 * to indicate that we have leaked blocks.
1980 release_metapage(mp);
1981 jfs_error(bmp->db_ipbmap->i_sb, "Block Leakage\n");
1985 /* write the buffer.
1995 * NAME: dbAllocDmapLev()
1997 * FUNCTION: attempt to allocate a specified number of contiguous blocks
1998 * from a specified dmap.
2000 * this routine checks if the contiguous blocks are available.
2001 * if so, nblocks of blocks are allocated; otherwise, ENOSPC is
2005 * mp - pointer to bmap descriptor
2006 * dp - pointer to dmap to attempt to allocate blocks from.
2007 * l2nb - log2 number of contiguous block desired.
2008 * nblocks - actual number of contiguous block desired.
2009 * results - on successful return, set to the starting block number
2010 * of the newly allocated range.
2014 * -ENOSPC - insufficient disk resources
2017 * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or
2018 * IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit;
2021 dbAllocDmapLev(struct bmap * bmp,
2022 struct dmap * dp, int nblocks, int l2nb, s64 * results)
2027 /* can't be more than a dmaps worth of blocks */
2028 assert(l2nb <= L2BPERDMAP);
2030 /* search the tree within the dmap page for sufficient
2031 * free space. if sufficient free space is found, dbFindLeaf()
2032 * returns the index of the leaf at which free space was found.
2034 if (dbFindLeaf((dmtree_t *) & dp->tree, l2nb, &leafidx))
2037 /* determine the block number within the file system corresponding
2038 * to the leaf at which free space was found.
2040 blkno = le64_to_cpu(dp->start) + (leafidx << L2DBWORD);
2042 /* if not all bits of the dmap word are free, get the starting
2043 * bit number within the dmap word of the required string of free
2044 * bits and adjust the block number with this value.
2046 if (dp->tree.stree[leafidx + LEAFIND] < BUDMIN)
2047 blkno += dbFindBits(le32_to_cpu(dp->wmap[leafidx]), l2nb);
2049 /* allocate the blocks */
2050 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
2058 * NAME: dbAllocDmap()
2060 * FUNCTION: adjust the disk allocation map to reflect the allocation
2061 * of a specified block range within a dmap.
2063 * this routine allocates the specified blocks from the dmap
2064 * through a call to dbAllocBits(). if the allocation of the
2065 * block range causes the maximum string of free blocks within
2066 * the dmap to change (i.e. the value of the root of the dmap's
2067 * dmtree), this routine will cause this change to be reflected
2068 * up through the appropriate levels of the dmap control pages
2069 * by a call to dbAdjCtl() for the L0 dmap control page that
2073 * bmp - pointer to bmap descriptor
2074 * dp - pointer to dmap to allocate the block range from.
2075 * blkno - starting block number of the block to be allocated.
2076 * nblocks - number of blocks to be allocated.
2082 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2084 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2090 /* save the current value of the root (i.e. maximum free string)
2093 oldroot = dp->tree.stree[ROOT];
2095 /* allocate the specified (blocks) bits */
2096 dbAllocBits(bmp, dp, blkno, nblocks);
2098 /* if the root has not changed, done. */
2099 if (dp->tree.stree[ROOT] == oldroot)
2102 /* root changed. bubble the change up to the dmap control pages.
2103 * if the adjustment of the upper level control pages fails,
2104 * backout the bit allocation (thus making everything consistent).
2106 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 1, 0)))
2107 dbFreeBits(bmp, dp, blkno, nblocks);
2114 * NAME: dbFreeDmap()
2116 * FUNCTION: adjust the disk allocation map to reflect the allocation
2117 * of a specified block range within a dmap.
2119 * this routine frees the specified blocks from the dmap through
2120 * a call to dbFreeBits(). if the deallocation of the block range
2121 * causes the maximum string of free blocks within the dmap to
2122 * change (i.e. the value of the root of the dmap's dmtree), this
2123 * routine will cause this change to be reflected up through the
2124 * appropriate levels of the dmap control pages by a call to
2125 * dbAdjCtl() for the L0 dmap control page that covers this dmap.
2128 * bmp - pointer to bmap descriptor
2129 * dp - pointer to dmap to free the block range from.
2130 * blkno - starting block number of the block to be freed.
2131 * nblocks - number of blocks to be freed.
2137 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2139 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2145 /* save the current value of the root (i.e. maximum free string)
2148 oldroot = dp->tree.stree[ROOT];
2150 /* free the specified (blocks) bits */
2151 rc = dbFreeBits(bmp, dp, blkno, nblocks);
2153 /* if error or the root has not changed, done. */
2154 if (rc || (dp->tree.stree[ROOT] == oldroot))
2157 /* root changed. bubble the change up to the dmap control pages.
2158 * if the adjustment of the upper level control pages fails,
2159 * backout the deallocation.
2161 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 0, 0))) {
2162 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
2164 /* as part of backing out the deallocation, we will have
2165 * to back split the dmap tree if the deallocation caused
2166 * the freed blocks to become part of a larger binary buddy
2169 if (dp->tree.stree[word] == NOFREE)
2170 dbBackSplit((dmtree_t *) & dp->tree, word);
2172 dbAllocBits(bmp, dp, blkno, nblocks);
2180 * NAME: dbAllocBits()
2182 * FUNCTION: allocate a specified block range from a dmap.
2184 * this routine updates the dmap to reflect the working
2185 * state allocation of the specified block range. it directly
2186 * updates the bits of the working map and causes the adjustment
2187 * of the binary buddy system described by the dmap's dmtree
2188 * leaves to reflect the bits allocated. it also causes the
2189 * dmap's dmtree, as a whole, to reflect the allocated range.
2192 * bmp - pointer to bmap descriptor
2193 * dp - pointer to dmap to allocate bits from.
2194 * blkno - starting block number of the bits to be allocated.
2195 * nblocks - number of bits to be allocated.
2197 * RETURN VALUES: none
2199 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2201 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2204 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2205 dmtree_t *tp = (dmtree_t *) & dp->tree;
2209 /* pick up a pointer to the leaves of the dmap tree */
2210 leaf = dp->tree.stree + LEAFIND;
2212 /* determine the bit number and word within the dmap of the
2215 dbitno = blkno & (BPERDMAP - 1);
2216 word = dbitno >> L2DBWORD;
2218 /* block range better be within the dmap */
2219 assert(dbitno + nblocks <= BPERDMAP);
2221 /* allocate the bits of the dmap's words corresponding to the block
2222 * range. not all bits of the first and last words may be contained
2223 * within the block range. if this is the case, we'll work against
2224 * those words (i.e. partial first and/or last) on an individual basis
2225 * (a single pass), allocating the bits of interest by hand and
2226 * updating the leaf corresponding to the dmap word. a single pass
2227 * will be used for all dmap words fully contained within the
2228 * specified range. within this pass, the bits of all fully contained
2229 * dmap words will be marked as free in a single shot and the leaves
2230 * will be updated. a single leaf may describe the free space of
2231 * multiple dmap words, so we may update only a subset of the actual
2232 * leaves corresponding to the dmap words of the block range.
2234 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2235 /* determine the bit number within the word and
2236 * the number of bits within the word.
2238 wbitno = dbitno & (DBWORD - 1);
2239 nb = min(rembits, DBWORD - wbitno);
2241 /* check if only part of a word is to be allocated.
2244 /* allocate (set to 1) the appropriate bits within
2247 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
2250 /* update the leaf for this dmap word. in addition
2251 * to setting the leaf value to the binary buddy max
2252 * of the updated dmap word, dbSplit() will split
2253 * the binary system of the leaves if need be.
2255 dbSplit(tp, word, BUDMIN,
2256 dbMaxBud((u8 *) & dp->wmap[word]));
2260 /* one or more dmap words are fully contained
2261 * within the block range. determine how many
2262 * words and allocate (set to 1) the bits of these
2265 nwords = rembits >> L2DBWORD;
2266 memset(&dp->wmap[word], (int) ONES, nwords * 4);
2268 /* determine how many bits.
2270 nb = nwords << L2DBWORD;
2272 /* now update the appropriate leaves to reflect
2273 * the allocated words.
2275 for (; nwords > 0; nwords -= nw) {
2276 if (leaf[word] < BUDMIN) {
2277 jfs_error(bmp->db_ipbmap->i_sb,
2278 "leaf page corrupt\n");
2282 /* determine what the leaf value should be
2283 * updated to as the minimum of the l2 number
2284 * of bits being allocated and the l2 number
2285 * of bits currently described by this leaf.
2287 size = min_t(int, leaf[word],
2288 NLSTOL2BSZ(nwords));
2290 /* update the leaf to reflect the allocation.
2291 * in addition to setting the leaf value to
2292 * NOFREE, dbSplit() will split the binary
2293 * system of the leaves to reflect the current
2294 * allocation (size).
2296 dbSplit(tp, word, size, NOFREE);
2298 /* get the number of dmap words handled */
2299 nw = BUDSIZE(size, BUDMIN);
2305 /* update the free count for this dmap */
2306 le32_add_cpu(&dp->nfree, -nblocks);
2310 /* if this allocation group is completely free,
2311 * update the maximum allocation group number if this allocation
2312 * group is the new max.
2314 agno = blkno >> bmp->db_agl2size;
2315 if (agno > bmp->db_maxag)
2316 bmp->db_maxag = agno;
2318 /* update the free count for the allocation group and map */
2319 bmp->db_agfree[agno] -= nblocks;
2320 bmp->db_nfree -= nblocks;
2327 * NAME: dbFreeBits()
2329 * FUNCTION: free a specified block range from a dmap.
2331 * this routine updates the dmap to reflect the working
2332 * state allocation of the specified block range. it directly
2333 * updates the bits of the working map and causes the adjustment
2334 * of the binary buddy system described by the dmap's dmtree
2335 * leaves to reflect the bits freed. it also causes the dmap's
2336 * dmtree, as a whole, to reflect the deallocated range.
2339 * bmp - pointer to bmap descriptor
2340 * dp - pointer to dmap to free bits from.
2341 * blkno - starting block number of the bits to be freed.
2342 * nblocks - number of bits to be freed.
2344 * RETURN VALUES: 0 for success
2346 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2348 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2351 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2352 dmtree_t *tp = (dmtree_t *) & dp->tree;
2356 /* determine the bit number and word within the dmap of the
2359 dbitno = blkno & (BPERDMAP - 1);
2360 word = dbitno >> L2DBWORD;
2362 /* block range better be within the dmap.
2364 assert(dbitno + nblocks <= BPERDMAP);
2366 /* free the bits of the dmaps words corresponding to the block range.
2367 * not all bits of the first and last words may be contained within
2368 * the block range. if this is the case, we'll work against those
2369 * words (i.e. partial first and/or last) on an individual basis
2370 * (a single pass), freeing the bits of interest by hand and updating
2371 * the leaf corresponding to the dmap word. a single pass will be used
2372 * for all dmap words fully contained within the specified range.
2373 * within this pass, the bits of all fully contained dmap words will
2374 * be marked as free in a single shot and the leaves will be updated. a
2375 * single leaf may describe the free space of multiple dmap words,
2376 * so we may update only a subset of the actual leaves corresponding
2377 * to the dmap words of the block range.
2379 * dbJoin() is used to update leaf values and will join the binary
2380 * buddy system of the leaves if the new leaf values indicate this
2383 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2384 /* determine the bit number within the word and
2385 * the number of bits within the word.
2387 wbitno = dbitno & (DBWORD - 1);
2388 nb = min(rembits, DBWORD - wbitno);
2390 /* check if only part of a word is to be freed.
2393 /* free (zero) the appropriate bits within this
2397 cpu_to_le32(~(ONES << (DBWORD - nb)
2400 /* update the leaf for this dmap word.
2402 rc = dbJoin(tp, word,
2403 dbMaxBud((u8 *) & dp->wmap[word]));
2409 /* one or more dmap words are fully contained
2410 * within the block range. determine how many
2411 * words and free (zero) the bits of these words.
2413 nwords = rembits >> L2DBWORD;
2414 memset(&dp->wmap[word], 0, nwords * 4);
2416 /* determine how many bits.
2418 nb = nwords << L2DBWORD;
2420 /* now update the appropriate leaves to reflect
2423 for (; nwords > 0; nwords -= nw) {
2424 /* determine what the leaf value should be
2425 * updated to as the minimum of the l2 number
2426 * of bits being freed and the l2 (max) number
2427 * of bits that can be described by this leaf.
2431 (word, L2LPERDMAP, BUDMIN),
2432 NLSTOL2BSZ(nwords));
2436 rc = dbJoin(tp, word, size);
2440 /* get the number of dmap words handled.
2442 nw = BUDSIZE(size, BUDMIN);
2448 /* update the free count for this dmap.
2450 le32_add_cpu(&dp->nfree, nblocks);
2454 /* update the free count for the allocation group and
2457 agno = blkno >> bmp->db_agl2size;
2458 bmp->db_nfree += nblocks;
2459 bmp->db_agfree[agno] += nblocks;
2461 /* check if this allocation group is not completely free and
2462 * if it is currently the maximum (rightmost) allocation group.
2463 * if so, establish the new maximum allocation group number by
2464 * searching left for the first allocation group with allocation.
2466 if ((bmp->db_agfree[agno] == bmp->db_agsize && agno == bmp->db_maxag) ||
2467 (agno == bmp->db_numag - 1 &&
2468 bmp->db_agfree[agno] == (bmp-> db_mapsize & (BPERDMAP - 1)))) {
2469 while (bmp->db_maxag > 0) {
2471 if (bmp->db_agfree[bmp->db_maxag] !=
2476 /* re-establish the allocation group preference if the
2477 * current preference is right of the maximum allocation
2480 if (bmp->db_agpref > bmp->db_maxag)
2481 bmp->db_agpref = bmp->db_maxag;
2493 * FUNCTION: adjust a dmap control page at a specified level to reflect
2494 * the change in a lower level dmap or dmap control page's
2495 * maximum string of free blocks (i.e. a change in the root
2496 * of the lower level object's dmtree) due to the allocation
2497 * or deallocation of a range of blocks with a single dmap.
2499 * on entry, this routine is provided with the new value of
2500 * the lower level dmap or dmap control page root and the
2501 * starting block number of the block range whose allocation
2502 * or deallocation resulted in the root change. this range
2503 * is respresented by a single leaf of the current dmapctl
2504 * and the leaf will be updated with this value, possibly
2505 * causing a binary buddy system within the leaves to be
2506 * split or joined. the update may also cause the dmapctl's
2507 * dmtree to be updated.
2509 * if the adjustment of the dmap control page, itself, causes its
2510 * root to change, this change will be bubbled up to the next dmap
2511 * control level by a recursive call to this routine, specifying
2512 * the new root value and the next dmap control page level to
2515 * bmp - pointer to bmap descriptor
2516 * blkno - the first block of a block range within a dmap. it is
2517 * the allocation or deallocation of this block range that
2518 * requires the dmap control page to be adjusted.
2519 * newval - the new value of the lower level dmap or dmap control
2521 * alloc - 'true' if adjustment is due to an allocation.
2522 * level - current level of dmap control page (i.e. L0, L1, L2) to
2529 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2532 dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, int level)
2534 struct metapage *mp;
2538 struct dmapctl *dcp;
2541 /* get the buffer for the dmap control page for the specified
2542 * block number and control page level.
2544 lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, level);
2545 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
2548 dcp = (struct dmapctl *) mp->data;
2550 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
2551 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
2552 release_metapage(mp);
2556 /* determine the leaf number corresponding to the block and
2557 * the index within the dmap control tree.
2559 leafno = BLKTOCTLLEAF(blkno, dcp->budmin);
2560 ti = leafno + le32_to_cpu(dcp->leafidx);
2562 /* save the current leaf value and the current root level (i.e.
2563 * maximum l2 free string described by this dmapctl).
2565 oldval = dcp->stree[ti];
2566 oldroot = dcp->stree[ROOT];
2568 /* check if this is a control page update for an allocation.
2569 * if so, update the leaf to reflect the new leaf value using
2570 * dbSplit(); otherwise (deallocation), use dbJoin() to update
2571 * the leaf with the new value. in addition to updating the
2572 * leaf, dbSplit() will also split the binary buddy system of
2573 * the leaves, if required, and bubble new values within the
2574 * dmapctl tree, if required. similarly, dbJoin() will join
2575 * the binary buddy system of leaves and bubble new values up
2576 * the dmapctl tree as required by the new leaf value.
2579 /* check if we are in the middle of a binary buddy
2580 * system. this happens when we are performing the
2581 * first allocation out of an allocation group that
2582 * is part (not the first part) of a larger binary
2583 * buddy system. if we are in the middle, back split
2584 * the system prior to calling dbSplit() which assumes
2585 * that it is at the front of a binary buddy system.
2587 if (oldval == NOFREE) {
2588 rc = dbBackSplit((dmtree_t *) dcp, leafno);
2591 oldval = dcp->stree[ti];
2593 dbSplit((dmtree_t *) dcp, leafno, dcp->budmin, newval);
2595 rc = dbJoin((dmtree_t *) dcp, leafno, newval);
2600 /* check if the root of the current dmap control page changed due
2601 * to the update and if the current dmap control page is not at
2602 * the current top level (i.e. L0, L1, L2) of the map. if so (i.e.
2603 * root changed and this is not the top level), call this routine
2604 * again (recursion) for the next higher level of the mapping to
2605 * reflect the change in root for the current dmap control page.
2607 if (dcp->stree[ROOT] != oldroot) {
2608 /* are we below the top level of the map. if so,
2609 * bubble the root up to the next higher level.
2611 if (level < bmp->db_maxlevel) {
2612 /* bubble up the new root of this dmap control page to
2616 dbAdjCtl(bmp, blkno, dcp->stree[ROOT], alloc,
2618 /* something went wrong in bubbling up the new
2619 * root value, so backout the changes to the
2620 * current dmap control page.
2623 dbJoin((dmtree_t *) dcp, leafno,
2626 /* the dbJoin() above might have
2627 * caused a larger binary buddy system
2628 * to form and we may now be in the
2629 * middle of it. if this is the case,
2630 * back split the buddies.
2632 if (dcp->stree[ti] == NOFREE)
2633 dbBackSplit((dmtree_t *)
2635 dbSplit((dmtree_t *) dcp, leafno,
2636 dcp->budmin, oldval);
2639 /* release the buffer and return the error.
2641 release_metapage(mp);
2645 /* we're at the top level of the map. update
2646 * the bmap control page to reflect the size
2647 * of the maximum free buddy system.
2649 assert(level == bmp->db_maxlevel);
2650 if (bmp->db_maxfreebud != oldroot) {
2651 jfs_error(bmp->db_ipbmap->i_sb,
2652 "the maximum free buddy is not the old root\n");
2654 bmp->db_maxfreebud = dcp->stree[ROOT];
2658 /* write the buffer.
2669 * FUNCTION: update the leaf of a dmtree with a new value, splitting
2670 * the leaf from the binary buddy system of the dmtree's
2671 * leaves, as required.
2674 * tp - pointer to the tree containing the leaf.
2675 * leafno - the number of the leaf to be updated.
2676 * splitsz - the size the binary buddy system starting at the leaf
2677 * must be split to, specified as the log2 number of blocks.
2678 * newval - the new value for the leaf.
2680 * RETURN VALUES: none
2682 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2684 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval)
2688 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2690 /* check if the leaf needs to be split.
2692 if (leaf[leafno] > tp->dmt_budmin) {
2693 /* the split occurs by cutting the buddy system in half
2694 * at the specified leaf until we reach the specified
2695 * size. pick up the starting split size (current size
2696 * - 1 in l2) and the corresponding buddy size.
2698 cursz = leaf[leafno] - 1;
2699 budsz = BUDSIZE(cursz, tp->dmt_budmin);
2701 /* split until we reach the specified size.
2703 while (cursz >= splitsz) {
2704 /* update the buddy's leaf with its new value.
2706 dbAdjTree(tp, leafno ^ budsz, cursz);
2708 /* on to the next size and buddy.
2715 /* adjust the dmap tree to reflect the specified leaf's new
2718 dbAdjTree(tp, leafno, newval);
2723 * NAME: dbBackSplit()
2725 * FUNCTION: back split the binary buddy system of dmtree leaves
2726 * that hold a specified leaf until the specified leaf
2727 * starts its own binary buddy system.
2729 * the allocators typically perform allocations at the start
2730 * of binary buddy systems and dbSplit() is used to accomplish
2731 * any required splits. in some cases, however, allocation
2732 * may occur in the middle of a binary system and requires a
2733 * back split, with the split proceeding out from the middle of
2734 * the system (less efficient) rather than the start of the
2735 * system (more efficient). the cases in which a back split
2736 * is required are rare and are limited to the first allocation
2737 * within an allocation group which is a part (not first part)
2738 * of a larger binary buddy system and a few exception cases
2739 * in which a previous join operation must be backed out.
2742 * tp - pointer to the tree containing the leaf.
2743 * leafno - the number of the leaf to be updated.
2745 * RETURN VALUES: none
2747 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2749 static int dbBackSplit(dmtree_t * tp, int leafno)
2751 int budsz, bud, w, bsz, size;
2753 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2755 /* leaf should be part (not first part) of a binary
2758 assert(leaf[leafno] == NOFREE);
2760 /* the back split is accomplished by iteratively finding the leaf
2761 * that starts the buddy system that contains the specified leaf and
2762 * splitting that system in two. this iteration continues until
2763 * the specified leaf becomes the start of a buddy system.
2765 * determine maximum possible l2 size for the specified leaf.
2768 LITOL2BSZ(leafno, le32_to_cpu(tp->dmt_l2nleafs),
2771 /* determine the number of leaves covered by this size. this
2772 * is the buddy size that we will start with as we search for
2773 * the buddy system that contains the specified leaf.
2775 budsz = BUDSIZE(size, tp->dmt_budmin);
2779 while (leaf[leafno] == NOFREE) {
2780 /* find the leftmost buddy leaf.
2782 for (w = leafno, bsz = budsz;; bsz <<= 1,
2783 w = (w < bud) ? w : bud) {
2784 if (bsz >= le32_to_cpu(tp->dmt_nleafs)) {
2785 jfs_err("JFS: block map error in dbBackSplit");
2789 /* determine the buddy.
2793 /* check if this buddy is the start of the system.
2795 if (leaf[bud] != NOFREE) {
2796 /* split the leaf at the start of the
2799 cursz = leaf[bud] - 1;
2800 dbSplit(tp, bud, cursz, cursz);
2806 if (leaf[leafno] != size) {
2807 jfs_err("JFS: wrong leaf value in dbBackSplit");
2817 * FUNCTION: update the leaf of a dmtree with a new value, joining
2818 * the leaf with other leaves of the dmtree into a multi-leaf
2819 * binary buddy system, as required.
2822 * tp - pointer to the tree containing the leaf.
2823 * leafno - the number of the leaf to be updated.
2824 * newval - the new value for the leaf.
2826 * RETURN VALUES: none
2828 static int dbJoin(dmtree_t * tp, int leafno, int newval)
2833 /* can the new leaf value require a join with other leaves ?
2835 if (newval >= tp->dmt_budmin) {
2836 /* pickup a pointer to the leaves of the tree.
2838 leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2840 /* try to join the specified leaf into a large binary
2841 * buddy system. the join proceeds by attempting to join
2842 * the specified leafno with its buddy (leaf) at new value.
2843 * if the join occurs, we attempt to join the left leaf
2844 * of the joined buddies with its buddy at new value + 1.
2845 * we continue to join until we find a buddy that cannot be
2846 * joined (does not have a value equal to the size of the
2847 * last join) or until all leaves have been joined into a
2850 * get the buddy size (number of words covered) of
2853 budsz = BUDSIZE(newval, tp->dmt_budmin);
2857 while (budsz < le32_to_cpu(tp->dmt_nleafs)) {
2858 /* get the buddy leaf.
2860 buddy = leafno ^ budsz;
2862 /* if the leaf's new value is greater than its
2863 * buddy's value, we join no more.
2865 if (newval > leaf[buddy])
2868 /* It shouldn't be less */
2869 if (newval < leaf[buddy])
2872 /* check which (leafno or buddy) is the left buddy.
2873 * the left buddy gets to claim the blocks resulting
2874 * from the join while the right gets to claim none.
2875 * the left buddy is also eligible to participate in
2876 * a join at the next higher level while the right
2880 if (leafno < buddy) {
2881 /* leafno is the left buddy.
2883 dbAdjTree(tp, buddy, NOFREE);
2885 /* buddy is the left buddy and becomes
2888 dbAdjTree(tp, leafno, NOFREE);
2892 /* on to try the next join.
2899 /* update the leaf value.
2901 dbAdjTree(tp, leafno, newval);
2910 * FUNCTION: update a leaf of a dmtree with a new value, adjusting
2911 * the dmtree, as required, to reflect the new leaf value.
2912 * the combination of any buddies must already be done before
2916 * tp - pointer to the tree to be adjusted.
2917 * leafno - the number of the leaf to be updated.
2918 * newval - the new value for the leaf.
2920 * RETURN VALUES: none
2922 static void dbAdjTree(dmtree_t * tp, int leafno, int newval)
2927 /* pick up the index of the leaf for this leafno.
2929 lp = leafno + le32_to_cpu(tp->dmt_leafidx);
2931 /* is the current value the same as the old value ? if so,
2932 * there is nothing to do.
2934 if (tp->dmt_stree[lp] == newval)
2937 /* set the new value.
2939 tp->dmt_stree[lp] = newval;
2941 /* bubble the new value up the tree as required.
2943 for (k = 0; k < le32_to_cpu(tp->dmt_height); k++) {
2944 /* get the index of the first leaf of the 4 leaf
2945 * group containing the specified leaf (leafno).
2947 lp = ((lp - 1) & ~0x03) + 1;
2949 /* get the index of the parent of this 4 leaf group.
2953 /* determine the maximum of the 4 leaves.
2955 max = TREEMAX(&tp->dmt_stree[lp]);
2957 /* if the maximum of the 4 is the same as the
2958 * parent's value, we're done.
2960 if (tp->dmt_stree[pp] == max)
2963 /* parent gets new value.
2965 tp->dmt_stree[pp] = max;
2967 /* parent becomes leaf for next go-round.
2975 * NAME: dbFindLeaf()
2977 * FUNCTION: search a dmtree_t for sufficient free blocks, returning
2978 * the index of a leaf describing the free blocks if
2979 * sufficient free blocks are found.
2981 * the search starts at the top of the dmtree_t tree and
2982 * proceeds down the tree to the leftmost leaf with sufficient
2986 * tp - pointer to the tree to be searched.
2987 * l2nb - log2 number of free blocks to search for.
2988 * leafidx - return pointer to be set to the index of the leaf
2989 * describing at least l2nb free blocks if sufficient
2990 * free blocks are found.
2994 * -ENOSPC - insufficient free blocks.
2996 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx)
2998 int ti, n = 0, k, x = 0;
3000 /* first check the root of the tree to see if there is
3001 * sufficient free space.
3003 if (l2nb > tp->dmt_stree[ROOT])
3006 /* sufficient free space available. now search down the tree
3007 * starting at the next level for the leftmost leaf that
3008 * describes sufficient free space.
3010 for (k = le32_to_cpu(tp->dmt_height), ti = 1;
3011 k > 0; k--, ti = ((ti + n) << 2) + 1) {
3012 /* search the four nodes at this level, starting from
3015 for (x = ti, n = 0; n < 4; n++) {
3016 /* sufficient free space found. move to the next
3017 * level (or quit if this is the last level).
3019 if (l2nb <= tp->dmt_stree[x + n])
3023 /* better have found something since the higher
3024 * levels of the tree said it was here.
3029 /* set the return to the leftmost leaf describing sufficient
3032 *leafidx = x + n - le32_to_cpu(tp->dmt_leafidx);
3039 * NAME: dbFindBits()
3041 * FUNCTION: find a specified number of binary buddy free bits within a
3042 * dmap bitmap word value.
3044 * this routine searches the bitmap value for (1 << l2nb) free
3045 * bits at (1 << l2nb) alignments within the value.
3048 * word - dmap bitmap word value.
3049 * l2nb - number of free bits specified as a log2 number.
3052 * starting bit number of free bits.
3054 static int dbFindBits(u32 word, int l2nb)
3059 /* get the number of bits.
3062 assert(nb <= DBWORD);
3064 /* complement the word so we can use a mask (i.e. 0s represent
3065 * free bits) and compute the mask.
3068 mask = ONES << (DBWORD - nb);
3070 /* scan the word for nb free bits at nb alignments.
3072 for (bitno = 0; mask != 0; bitno += nb, mask >>= nb) {
3073 if ((mask & word) == mask)
3079 /* return the bit number.
3086 * NAME: dbMaxBud(u8 *cp)
3088 * FUNCTION: determine the largest binary buddy string of free
3089 * bits within 32-bits of the map.
3092 * cp - pointer to the 32-bit value.
3095 * largest binary buddy of free bits within a dmap word.
3097 static int dbMaxBud(u8 * cp)
3099 signed char tmp1, tmp2;
3101 /* check if the wmap word is all free. if so, the
3102 * free buddy size is BUDMIN.
3104 if (*((uint *) cp) == 0)
3107 /* check if the wmap word is half free. if so, the
3108 * free buddy size is BUDMIN-1.
3110 if (*((u16 *) cp) == 0 || *((u16 *) cp + 1) == 0)
3111 return (BUDMIN - 1);
3113 /* not all free or half free. determine the free buddy
3114 * size thru table lookup using quarters of the wmap word.
3116 tmp1 = max(budtab[cp[2]], budtab[cp[3]]);
3117 tmp2 = max(budtab[cp[0]], budtab[cp[1]]);
3118 return (max(tmp1, tmp2));
3123 * NAME: cnttz(uint word)
3125 * FUNCTION: determine the number of trailing zeros within a 32-bit
3129 * value - 32-bit value to be examined.
3132 * count of trailing zeros
3134 static int cnttz(u32 word)
3138 for (n = 0; n < 32; n++, word >>= 1) {
3148 * NAME: cntlz(u32 value)
3150 * FUNCTION: determine the number of leading zeros within a 32-bit
3154 * value - 32-bit value to be examined.
3157 * count of leading zeros
3159 static int cntlz(u32 value)
3163 for (n = 0; n < 32; n++, value <<= 1) {
3164 if (value & HIGHORDER)
3172 * NAME: blkstol2(s64 nb)
3174 * FUNCTION: convert a block count to its log2 value. if the block
3175 * count is not a l2 multiple, it is rounded up to the next
3176 * larger l2 multiple.
3179 * nb - number of blocks
3182 * log2 number of blocks
3184 static int blkstol2(s64 nb)
3187 s64 mask; /* meant to be signed */
3189 mask = (s64) 1 << (64 - 1);
3191 /* count the leading bits.
3193 for (l2nb = 0; l2nb < 64; l2nb++, mask >>= 1) {
3194 /* leading bit found.
3197 /* determine the l2 value.
3199 l2nb = (64 - 1) - l2nb;
3201 /* check if we need to round up.
3210 return 0; /* fix compiler warning */
3215 * NAME: dbAllocBottomUp()
3217 * FUNCTION: alloc the specified block range from the working block
3220 * the blocks will be alloc from the working map one dmap
3224 * ip - pointer to in-core inode;
3225 * blkno - starting block number to be freed.
3226 * nblocks - number of blocks to be freed.
3232 int dbAllocBottomUp(struct inode *ip, s64 blkno, s64 nblocks)
3234 struct metapage *mp;
3238 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
3239 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
3241 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
3243 /* block to be allocated better be within the mapsize. */
3244 ASSERT(nblocks <= bmp->db_mapsize - blkno);
3247 * allocate the blocks a dmap at a time.
3250 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
3251 /* release previous dmap if any */
3256 /* get the buffer for the current dmap. */
3257 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
3258 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
3260 IREAD_UNLOCK(ipbmap);
3263 dp = (struct dmap *) mp->data;
3265 /* determine the number of blocks to be allocated from
3268 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
3270 /* allocate the blocks. */
3271 if ((rc = dbAllocDmapBU(bmp, dp, blkno, nb))) {
3272 release_metapage(mp);
3273 IREAD_UNLOCK(ipbmap);
3278 /* write the last buffer. */
3281 IREAD_UNLOCK(ipbmap);
3287 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
3291 int dbitno, word, rembits, nb, nwords, wbitno, agno;
3293 struct dmaptree *tp = (struct dmaptree *) & dp->tree;
3295 /* save the current value of the root (i.e. maximum free string)
3298 oldroot = tp->stree[ROOT];
3300 /* determine the bit number and word within the dmap of the
3303 dbitno = blkno & (BPERDMAP - 1);
3304 word = dbitno >> L2DBWORD;
3306 /* block range better be within the dmap */
3307 assert(dbitno + nblocks <= BPERDMAP);
3309 /* allocate the bits of the dmap's words corresponding to the block
3310 * range. not all bits of the first and last words may be contained
3311 * within the block range. if this is the case, we'll work against
3312 * those words (i.e. partial first and/or last) on an individual basis
3313 * (a single pass), allocating the bits of interest by hand and
3314 * updating the leaf corresponding to the dmap word. a single pass
3315 * will be used for all dmap words fully contained within the
3316 * specified range. within this pass, the bits of all fully contained
3317 * dmap words will be marked as free in a single shot and the leaves
3318 * will be updated. a single leaf may describe the free space of
3319 * multiple dmap words, so we may update only a subset of the actual
3320 * leaves corresponding to the dmap words of the block range.
3322 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
3323 /* determine the bit number within the word and
3324 * the number of bits within the word.
3326 wbitno = dbitno & (DBWORD - 1);
3327 nb = min(rembits, DBWORD - wbitno);
3329 /* check if only part of a word is to be allocated.
3332 /* allocate (set to 1) the appropriate bits within
3335 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
3340 /* one or more dmap words are fully contained
3341 * within the block range. determine how many
3342 * words and allocate (set to 1) the bits of these
3345 nwords = rembits >> L2DBWORD;
3346 memset(&dp->wmap[word], (int) ONES, nwords * 4);
3348 /* determine how many bits */
3349 nb = nwords << L2DBWORD;
3354 /* update the free count for this dmap */
3355 le32_add_cpu(&dp->nfree, -nblocks);
3357 /* reconstruct summary tree */
3362 /* if this allocation group is completely free,
3363 * update the highest active allocation group number
3364 * if this allocation group is the new max.
3366 agno = blkno >> bmp->db_agl2size;
3367 if (agno > bmp->db_maxag)
3368 bmp->db_maxag = agno;
3370 /* update the free count for the allocation group and map */
3371 bmp->db_agfree[agno] -= nblocks;
3372 bmp->db_nfree -= nblocks;
3376 /* if the root has not changed, done. */
3377 if (tp->stree[ROOT] == oldroot)
3380 /* root changed. bubble the change up to the dmap control pages.
3381 * if the adjustment of the upper level control pages fails,
3382 * backout the bit allocation (thus making everything consistent).
3384 if ((rc = dbAdjCtl(bmp, blkno, tp->stree[ROOT], 1, 0)))
3385 dbFreeBits(bmp, dp, blkno, nblocks);
3392 * NAME: dbExtendFS()
3394 * FUNCTION: extend bmap from blkno for nblocks;
3395 * dbExtendFS() updates bmap ready for dbAllocBottomUp();
3399 * L1---------------------------------L1
3401 * L0---------L0---------L0 L0---------L0---------L0
3403 * d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,.,dm;
3404 * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm
3406 * <---old---><----------------------------extend----------------------->
3408 int dbExtendFS(struct inode *ipbmap, s64 blkno, s64 nblocks)
3410 struct jfs_sb_info *sbi = JFS_SBI(ipbmap->i_sb);
3411 int nbperpage = sbi->nbperpage;
3412 int i, i0 = true, j, j0 = true, k, n;
3415 struct metapage *mp, *l2mp, *l1mp = NULL, *l0mp = NULL;
3416 struct dmapctl *l2dcp, *l1dcp, *l0dcp;
3418 s8 *l0leaf, *l1leaf, *l2leaf;
3419 struct bmap *bmp = sbi->bmap;
3420 int agno, l2agsize, oldl2agsize;
3423 newsize = blkno + nblocks;
3425 jfs_info("dbExtendFS: blkno:%Ld nblocks:%Ld newsize:%Ld",
3426 (long long) blkno, (long long) nblocks, (long long) newsize);
3429 * initialize bmap control page.
3431 * all the data in bmap control page should exclude
3432 * the mkfs hidden dmap page.
3435 /* update mapsize */
3436 bmp->db_mapsize = newsize;
3437 bmp->db_maxlevel = BMAPSZTOLEV(bmp->db_mapsize);
3439 /* compute new AG size */
3440 l2agsize = dbGetL2AGSize(newsize);
3441 oldl2agsize = bmp->db_agl2size;
3443 bmp->db_agl2size = l2agsize;
3444 bmp->db_agsize = 1 << l2agsize;
3446 /* compute new number of AG */
3447 agno = bmp->db_numag;
3448 bmp->db_numag = newsize >> l2agsize;
3449 bmp->db_numag += ((u32) newsize % (u32) bmp->db_agsize) ? 1 : 0;
3452 * reconfigure db_agfree[]
3453 * from old AG configuration to new AG configuration;
3455 * coalesce contiguous k (newAGSize/oldAGSize) AGs;
3456 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn;
3457 * note: new AG size = old AG size * (2**x).
3459 if (l2agsize == oldl2agsize)
3461 k = 1 << (l2agsize - oldl2agsize);
3462 ag_rem = bmp->db_agfree[0]; /* save agfree[0] */
3463 for (i = 0, n = 0; i < agno; n++) {
3464 bmp->db_agfree[n] = 0; /* init collection point */
3466 /* coalesce contiguous k AGs; */
3467 for (j = 0; j < k && i < agno; j++, i++) {
3468 /* merge AGi to AGn */
3469 bmp->db_agfree[n] += bmp->db_agfree[i];
3472 bmp->db_agfree[0] += ag_rem; /* restore agfree[0] */
3474 for (; n < MAXAG; n++)
3475 bmp->db_agfree[n] = 0;
3478 * update highest active ag number
3481 bmp->db_maxag = bmp->db_maxag / k;
3486 * update bit maps and corresponding level control pages;
3487 * global control page db_nfree, db_agfree[agno], db_maxfreebud;
3491 p = BMAPBLKNO + nbperpage; /* L2 page */
3492 l2mp = read_metapage(ipbmap, p, PSIZE, 0);
3494 jfs_error(ipbmap->i_sb, "L2 page could not be read\n");
3497 l2dcp = (struct dmapctl *) l2mp->data;
3499 /* compute start L1 */
3500 k = blkno >> L2MAXL1SIZE;
3501 l2leaf = l2dcp->stree + CTLLEAFIND + k;
3502 p = BLKTOL1(blkno, sbi->l2nbperpage); /* L1 page */
3505 * extend each L1 in L2
3507 for (; k < LPERCTL; k++, p += nbperpage) {
3510 /* read in L1 page: (blkno & (MAXL1SIZE - 1)) */
3511 l1mp = read_metapage(ipbmap, p, PSIZE, 0);
3514 l1dcp = (struct dmapctl *) l1mp->data;
3516 /* compute start L0 */
3517 j = (blkno & (MAXL1SIZE - 1)) >> L2MAXL0SIZE;
3518 l1leaf = l1dcp->stree + CTLLEAFIND + j;
3519 p = BLKTOL0(blkno, sbi->l2nbperpage);
3522 /* assign/init L1 page */
3523 l1mp = get_metapage(ipbmap, p, PSIZE, 0);
3527 l1dcp = (struct dmapctl *) l1mp->data;
3529 /* compute start L0 */
3531 l1leaf = l1dcp->stree + CTLLEAFIND;
3532 p += nbperpage; /* 1st L0 of L1.k */
3536 * extend each L0 in L1
3538 for (; j < LPERCTL; j++) {
3541 /* read in L0 page: (blkno & (MAXL0SIZE - 1)) */
3543 l0mp = read_metapage(ipbmap, p, PSIZE, 0);
3546 l0dcp = (struct dmapctl *) l0mp->data;
3548 /* compute start dmap */
3549 i = (blkno & (MAXL0SIZE - 1)) >>
3551 l0leaf = l0dcp->stree + CTLLEAFIND + i;
3552 p = BLKTODMAP(blkno,
3556 /* assign/init L0 page */
3557 l0mp = get_metapage(ipbmap, p, PSIZE, 0);
3561 l0dcp = (struct dmapctl *) l0mp->data;
3563 /* compute start dmap */
3565 l0leaf = l0dcp->stree + CTLLEAFIND;
3566 p += nbperpage; /* 1st dmap of L0.j */
3570 * extend each dmap in L0
3572 for (; i < LPERCTL; i++) {
3574 * reconstruct the dmap page, and
3575 * initialize corresponding parent L0 leaf
3577 if ((n = blkno & (BPERDMAP - 1))) {
3578 /* read in dmap page: */
3579 mp = read_metapage(ipbmap, p,
3583 n = min(nblocks, (s64)BPERDMAP - n);
3585 /* assign/init dmap page */
3586 mp = read_metapage(ipbmap, p,
3591 n = min_t(s64, nblocks, BPERDMAP);
3594 dp = (struct dmap *) mp->data;
3595 *l0leaf = dbInitDmap(dp, blkno, n);
3598 agno = le64_to_cpu(dp->start) >> l2agsize;
3599 bmp->db_agfree[agno] += n;
3610 } /* for each dmap in a L0 */
3613 * build current L0 page from its leaves, and
3614 * initialize corresponding parent L1 leaf
3616 *l1leaf = dbInitDmapCtl(l0dcp, 0, ++i);
3617 write_metapage(l0mp);
3621 l1leaf++; /* continue for next L0 */
3623 /* more than 1 L0 ? */
3625 break; /* build L1 page */
3627 /* summarize in global bmap page */
3628 bmp->db_maxfreebud = *l1leaf;
3629 release_metapage(l1mp);
3630 release_metapage(l2mp);
3634 } /* for each L0 in a L1 */
3637 * build current L1 page from its leaves, and
3638 * initialize corresponding parent L2 leaf
3640 *l2leaf = dbInitDmapCtl(l1dcp, 1, ++j);
3641 write_metapage(l1mp);
3645 l2leaf++; /* continue for next L1 */
3647 /* more than 1 L1 ? */
3649 break; /* build L2 page */
3651 /* summarize in global bmap page */
3652 bmp->db_maxfreebud = *l2leaf;
3653 release_metapage(l2mp);
3657 } /* for each L1 in a L2 */
3659 jfs_error(ipbmap->i_sb, "function has not returned as expected\n");
3662 release_metapage(l0mp);
3664 release_metapage(l1mp);
3665 release_metapage(l2mp);
3669 * finalize bmap control page
3680 void dbFinalizeBmap(struct inode *ipbmap)
3682 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
3683 int actags, inactags, l2nl;
3684 s64 ag_rem, actfree, inactfree, avgfree;
3688 * finalize bmap control page
3692 * compute db_agpref: preferred ag to allocate from
3693 * (the leftmost ag with average free space in it);
3696 /* get the number of active ags and inacitve ags */
3697 actags = bmp->db_maxag + 1;
3698 inactags = bmp->db_numag - actags;
3699 ag_rem = bmp->db_mapsize & (bmp->db_agsize - 1); /* ??? */
3701 /* determine how many blocks are in the inactive allocation
3702 * groups. in doing this, we must account for the fact that
3703 * the rightmost group might be a partial group (i.e. file
3704 * system size is not a multiple of the group size).
3706 inactfree = (inactags && ag_rem) ?
3707 ((inactags - 1) << bmp->db_agl2size) + ag_rem
3708 : inactags << bmp->db_agl2size;
3710 /* determine how many free blocks are in the active
3711 * allocation groups plus the average number of free blocks
3712 * within the active ags.
3714 actfree = bmp->db_nfree - inactfree;
3715 avgfree = (u32) actfree / (u32) actags;
3717 /* if the preferred allocation group has not average free space.
3718 * re-establish the preferred group as the leftmost
3719 * group with average free space.
3721 if (bmp->db_agfree[bmp->db_agpref] < avgfree) {
3722 for (bmp->db_agpref = 0; bmp->db_agpref < actags;
3724 if (bmp->db_agfree[bmp->db_agpref] >= avgfree)
3727 if (bmp->db_agpref >= bmp->db_numag) {
3728 jfs_error(ipbmap->i_sb,
3729 "cannot find ag with average freespace\n");
3734 * compute db_aglevel, db_agheight, db_width, db_agstart:
3735 * an ag is covered in aglevel dmapctl summary tree,
3736 * at agheight level height (from leaf) with agwidth number of nodes
3737 * each, which starts at agstart index node of the smmary tree node
3740 bmp->db_aglevel = BMAPSZTOLEV(bmp->db_agsize);
3742 bmp->db_agl2size - (L2BPERDMAP + bmp->db_aglevel * L2LPERCTL);
3743 bmp->db_agheight = l2nl >> 1;
3744 bmp->db_agwidth = 1 << (l2nl - (bmp->db_agheight << 1));
3745 for (i = 5 - bmp->db_agheight, bmp->db_agstart = 0, n = 1; i > 0;
3747 bmp->db_agstart += n;
3755 * NAME: dbInitDmap()/ujfs_idmap_page()
3757 * FUNCTION: initialize working/persistent bitmap of the dmap page
3758 * for the specified number of blocks:
3760 * at entry, the bitmaps had been initialized as free (ZEROS);
3761 * The number of blocks will only account for the actually
3762 * existing blocks. Blocks which don't actually exist in
3763 * the aggregate will be marked as allocated (ONES);
3766 * dp - pointer to page of map
3767 * nblocks - number of blocks this page
3771 static int dbInitDmap(struct dmap * dp, s64 Blkno, int nblocks)
3773 int blkno, w, b, r, nw, nb, i;
3775 /* starting block number within the dmap */
3776 blkno = Blkno & (BPERDMAP - 1);
3779 dp->nblocks = dp->nfree = cpu_to_le32(nblocks);
3780 dp->start = cpu_to_le64(Blkno);
3782 if (nblocks == BPERDMAP) {
3783 memset(&dp->wmap[0], 0, LPERDMAP * 4);
3784 memset(&dp->pmap[0], 0, LPERDMAP * 4);
3788 le32_add_cpu(&dp->nblocks, nblocks);
3789 le32_add_cpu(&dp->nfree, nblocks);
3792 /* word number containing start block number */
3793 w = blkno >> L2DBWORD;
3796 * free the bits corresponding to the block range (ZEROS):
3797 * note: not all bits of the first and last words may be contained
3798 * within the block range.
3800 for (r = nblocks; r > 0; r -= nb, blkno += nb) {
3801 /* number of bits preceding range to be freed in the word */
3802 b = blkno & (DBWORD - 1);
3803 /* number of bits to free in the word */
3804 nb = min(r, DBWORD - b);
3806 /* is partial word to be freed ? */
3808 /* free (set to 0) from the bitmap word */
3809 dp->wmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3811 dp->pmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3814 /* skip the word freed */
3817 /* free (set to 0) contiguous bitmap words */
3819 memset(&dp->wmap[w], 0, nw * 4);
3820 memset(&dp->pmap[w], 0, nw * 4);
3822 /* skip the words freed */
3823 nb = nw << L2DBWORD;
3829 * mark bits following the range to be freed (non-existing
3830 * blocks) as allocated (ONES)
3833 if (blkno == BPERDMAP)
3836 /* the first word beyond the end of existing blocks */
3837 w = blkno >> L2DBWORD;
3839 /* does nblocks fall on a 32-bit boundary ? */
3840 b = blkno & (DBWORD - 1);
3842 /* mark a partial word allocated */
3843 dp->wmap[w] = dp->pmap[w] = cpu_to_le32(ONES >> b);
3847 /* set the rest of the words in the page to allocated (ONES) */
3848 for (i = w; i < LPERDMAP; i++)
3849 dp->pmap[i] = dp->wmap[i] = cpu_to_le32(ONES);
3855 return (dbInitDmapTree(dp));
3860 * NAME: dbInitDmapTree()/ujfs_complete_dmap()
3862 * FUNCTION: initialize summary tree of the specified dmap:
3864 * at entry, bitmap of the dmap has been initialized;
3867 * dp - dmap to complete
3868 * blkno - starting block number for this dmap
3869 * treemax - will be filled in with max free for this dmap
3871 * RETURNS: max free string at the root of the tree
3873 static int dbInitDmapTree(struct dmap * dp)
3875 struct dmaptree *tp;
3879 /* init fixed info of tree */
3881 tp->nleafs = cpu_to_le32(LPERDMAP);
3882 tp->l2nleafs = cpu_to_le32(L2LPERDMAP);
3883 tp->leafidx = cpu_to_le32(LEAFIND);
3884 tp->height = cpu_to_le32(4);
3885 tp->budmin = BUDMIN;
3887 /* init each leaf from corresponding wmap word:
3888 * note: leaf is set to NOFREE(-1) if all blocks of corresponding
3889 * bitmap word are allocated.
3891 cp = tp->stree + le32_to_cpu(tp->leafidx);
3892 for (i = 0; i < LPERDMAP; i++)
3893 *cp++ = dbMaxBud((u8 *) & dp->wmap[i]);
3895 /* build the dmap's binary buddy summary tree */
3896 return (dbInitTree(tp));
3901 * NAME: dbInitTree()/ujfs_adjtree()
3903 * FUNCTION: initialize binary buddy summary tree of a dmap or dmapctl.
3905 * at entry, the leaves of the tree has been initialized
3906 * from corresponding bitmap word or root of summary tree
3907 * of the child control page;
3908 * configure binary buddy system at the leaf level, then
3909 * bubble up the values of the leaf nodes up the tree.
3912 * cp - Pointer to the root of the tree
3913 * l2leaves- Number of leaf nodes as a power of 2
3914 * l2min - Number of blocks that can be covered by a leaf
3917 * RETURNS: max free string at the root of the tree
3919 static int dbInitTree(struct dmaptree * dtp)
3921 int l2max, l2free, bsize, nextb, i;
3922 int child, parent, nparent;
3927 /* Determine the maximum free string possible for the leaves */
3928 l2max = le32_to_cpu(dtp->l2nleafs) + dtp->budmin;
3931 * configure the leaf levevl into binary buddy system
3933 * Try to combine buddies starting with a buddy size of 1
3934 * (i.e. two leaves). At a buddy size of 1 two buddy leaves
3935 * can be combined if both buddies have a maximum free of l2min;
3936 * the combination will result in the left-most buddy leaf having
3937 * a maximum free of l2min+1.
3938 * After processing all buddies for a given size, process buddies
3939 * at the next higher buddy size (i.e. current size * 2) and
3940 * the next maximum free (current free + 1).
3941 * This continues until the maximum possible buddy combination
3942 * yields maximum free.
3944 for (l2free = dtp->budmin, bsize = 1; l2free < l2max;
3945 l2free++, bsize = nextb) {
3946 /* get next buddy size == current buddy pair size */
3949 /* scan each adjacent buddy pair at current buddy size */
3950 for (i = 0, cp = tp + le32_to_cpu(dtp->leafidx);
3951 i < le32_to_cpu(dtp->nleafs);
3952 i += nextb, cp += nextb) {
3953 /* coalesce if both adjacent buddies are max free */
3954 if (*cp == l2free && *(cp + bsize) == l2free) {
3955 *cp = l2free + 1; /* left take right */
3956 *(cp + bsize) = -1; /* right give left */
3962 * bubble summary information of leaves up the tree.
3964 * Starting at the leaf node level, the four nodes described by
3965 * the higher level parent node are compared for a maximum free and
3966 * this maximum becomes the value of the parent node.
3967 * when all lower level nodes are processed in this fashion then
3968 * move up to the next level (parent becomes a lower level node) and
3969 * continue the process for that level.
3971 for (child = le32_to_cpu(dtp->leafidx),
3972 nparent = le32_to_cpu(dtp->nleafs) >> 2;
3973 nparent > 0; nparent >>= 2, child = parent) {
3974 /* get index of 1st node of parent level */
3975 parent = (child - 1) >> 2;
3977 /* set the value of the parent node as the maximum
3978 * of the four nodes of the current level.
3980 for (i = 0, cp = tp + child, cp1 = tp + parent;
3981 i < nparent; i++, cp += 4, cp1++)
3992 * function: initialize dmapctl page
3994 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i)
3995 { /* start leaf index not covered by range */
3998 dcp->nleafs = cpu_to_le32(LPERCTL);
3999 dcp->l2nleafs = cpu_to_le32(L2LPERCTL);
4000 dcp->leafidx = cpu_to_le32(CTLLEAFIND);
4001 dcp->height = cpu_to_le32(5);
4002 dcp->budmin = L2BPERDMAP + L2LPERCTL * level;
4005 * initialize the leaves of current level that were not covered
4006 * by the specified input block range (i.e. the leaves have no
4007 * low level dmapctl or dmap).
4009 cp = &dcp->stree[CTLLEAFIND + i];
4010 for (; i < LPERCTL; i++)
4013 /* build the dmap's binary buddy summary tree */
4014 return (dbInitTree((struct dmaptree *) dcp));
4019 * NAME: dbGetL2AGSize()/ujfs_getagl2size()
4021 * FUNCTION: Determine log2(allocation group size) from aggregate size
4024 * nblocks - Number of blocks in aggregate
4026 * RETURNS: log2(allocation group size) in aggregate blocks
4028 static int dbGetL2AGSize(s64 nblocks)
4034 if (nblocks < BPERDMAP * MAXAG)
4035 return (L2BPERDMAP);
4037 /* round up aggregate size to power of 2 */
4038 m = ((u64) 1 << (64 - 1));
4039 for (l2sz = 64; l2sz >= 0; l2sz--, m >>= 1) {
4044 sz = (s64) 1 << l2sz;
4048 /* agsize = roundupSize/max_number_of_ag */
4049 return (l2sz - L2MAXAG);
4054 * NAME: dbMapFileSizeToMapSize()
4056 * FUNCTION: compute number of blocks the block allocation map file
4057 * can cover from the map file size;
4059 * RETURNS: Number of blocks which can be covered by this block map file;
4063 * maximum number of map pages at each level including control pages
4065 #define MAXL0PAGES (1 + LPERCTL)
4066 #define MAXL1PAGES (1 + LPERCTL * MAXL0PAGES)
4067 #define MAXL2PAGES (1 + LPERCTL * MAXL1PAGES)
4070 * convert number of map pages to the zero origin top dmapctl level
4072 #define BMAPPGTOLEV(npages) \
4073 (((npages) <= 3 + MAXL0PAGES) ? 0 : \
4074 ((npages) <= 2 + MAXL1PAGES) ? 1 : 2)
4076 s64 dbMapFileSizeToMapSize(struct inode * ipbmap)
4078 struct super_block *sb = ipbmap->i_sb;
4082 int complete, factor;
4084 nblocks = ipbmap->i_size >> JFS_SBI(sb)->l2bsize;
4085 npages = nblocks >> JFS_SBI(sb)->l2nbperpage;
4086 level = BMAPPGTOLEV(npages);
4088 /* At each level, accumulate the number of dmap pages covered by
4089 * the number of full child levels below it;
4090 * repeat for the last incomplete child level.
4093 npages--; /* skip the first global control page */
4094 /* skip higher level control pages above top level covered by map */
4095 npages -= (2 - level);
4096 npages--; /* skip top level's control page */
4097 for (i = level; i >= 0; i--) {
4099 (i == 2) ? MAXL1PAGES : ((i == 1) ? MAXL0PAGES : 1);
4100 complete = (u32) npages / factor;
4101 ndmaps += complete * ((i == 2) ? LPERCTL * LPERCTL :
4102 ((i == 1) ? LPERCTL : 1));
4104 /* pages in last/incomplete child */
4105 npages = (u32) npages % factor;
4106 /* skip incomplete child's level control page */
4110 /* convert the number of dmaps into the number of blocks
4111 * which can be covered by the dmaps;
4113 nblocks = ndmaps << L2BPERDMAP;