1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright (C) International Business Machines Corp., 2000-2004
4 * Portions Copyright (C) Tino Reichardt, 2012
8 #include <linux/slab.h>
9 #include "jfs_incore.h"
10 #include "jfs_superblock.h"
14 #include "jfs_metapage.h"
15 #include "jfs_debug.h"
16 #include "jfs_discard.h"
19 * SERIALIZATION of the Block Allocation Map.
21 * the working state of the block allocation map is accessed in
24 * 1) allocation and free requests that start at the dmap
25 * level and move up through the dmap control pages (i.e.
26 * the vast majority of requests).
28 * 2) allocation requests that start at dmap control page
29 * level and work down towards the dmaps.
31 * the serialization scheme used here is as follows.
33 * requests which start at the bottom are serialized against each
34 * other through buffers and each requests holds onto its buffers
35 * as it works it way up from a single dmap to the required level
36 * of dmap control page.
37 * requests that start at the top are serialized against each other
38 * and request that start from the bottom by the multiple read/single
39 * write inode lock of the bmap inode. requests starting at the top
40 * take this lock in write mode while request starting at the bottom
41 * take the lock in read mode. a single top-down request may proceed
42 * exclusively while multiple bottoms-up requests may proceed
43 * simultaneously (under the protection of busy buffers).
45 * in addition to information found in dmaps and dmap control pages,
46 * the working state of the block allocation map also includes read/
47 * write information maintained in the bmap descriptor (i.e. total
48 * free block count, allocation group level free block counts).
49 * a single exclusive lock (BMAP_LOCK) is used to guard this information
50 * in the face of multiple-bottoms up requests.
51 * (lock ordering: IREAD_LOCK, BMAP_LOCK);
53 * accesses to the persistent state of the block allocation map (limited
54 * to the persistent bitmaps in dmaps) is guarded by (busy) buffers.
57 #define BMAP_LOCK_INIT(bmp) mutex_init(&bmp->db_bmaplock)
58 #define BMAP_LOCK(bmp) mutex_lock(&bmp->db_bmaplock)
59 #define BMAP_UNLOCK(bmp) mutex_unlock(&bmp->db_bmaplock)
64 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
66 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval);
67 static int dbBackSplit(dmtree_t * tp, int leafno);
68 static int dbJoin(dmtree_t * tp, int leafno, int newval);
69 static void dbAdjTree(dmtree_t * tp, int leafno, int newval);
70 static int dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc,
72 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results);
73 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
75 static int dbAllocNear(struct bmap * bmp, struct dmap * dp, s64 blkno,
77 int l2nb, s64 * results);
78 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
80 static int dbAllocDmapLev(struct bmap * bmp, struct dmap * dp, int nblocks,
83 static int dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb,
85 static int dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno,
87 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks);
88 static int dbFindBits(u32 word, int l2nb);
89 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno);
90 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx);
91 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
93 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
95 static int dbMaxBud(u8 * cp);
96 static int blkstol2(s64 nb);
98 static int cntlz(u32 value);
99 static int cnttz(u32 word);
101 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
103 static int dbInitDmap(struct dmap * dp, s64 blkno, int nblocks);
104 static int dbInitDmapTree(struct dmap * dp);
105 static int dbInitTree(struct dmaptree * dtp);
106 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i);
107 static int dbGetL2AGSize(s64 nblocks);
112 * table used for determining buddy sizes within characters of
113 * dmap bitmap words. the characters themselves serve as indexes
114 * into the table, with the table elements yielding the maximum
115 * binary buddy of free bits within the character.
117 static const s8 budtab[256] = {
118 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
119 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
120 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
121 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
122 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
123 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
124 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
125 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
126 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
127 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
128 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
129 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
130 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
131 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
132 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
133 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, -1
139 * FUNCTION: initializate the block allocation map.
141 * memory is allocated for the in-core bmap descriptor and
142 * the in-core descriptor is initialized from disk.
145 * ipbmap - pointer to in-core inode for the block map.
149 * -ENOMEM - insufficient memory
151 * -EINVAL - wrong bmap data
153 int dbMount(struct inode *ipbmap)
156 struct dbmap_disk *dbmp_le;
161 * allocate/initialize the in-memory bmap descriptor
163 /* allocate memory for the in-memory bmap descriptor */
164 bmp = kmalloc(sizeof(struct bmap), GFP_KERNEL);
168 /* read the on-disk bmap descriptor. */
169 mp = read_metapage(ipbmap,
170 BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
177 /* copy the on-disk bmap descriptor to its in-memory version. */
178 dbmp_le = (struct dbmap_disk *) mp->data;
179 bmp->db_mapsize = le64_to_cpu(dbmp_le->dn_mapsize);
180 bmp->db_nfree = le64_to_cpu(dbmp_le->dn_nfree);
182 bmp->db_l2nbperpage = le32_to_cpu(dbmp_le->dn_l2nbperpage);
183 if (bmp->db_l2nbperpage > L2PSIZE - L2MINBLOCKSIZE ||
184 bmp->db_l2nbperpage < 0) {
186 goto err_release_metapage;
189 bmp->db_numag = le32_to_cpu(dbmp_le->dn_numag);
190 if (!bmp->db_numag) {
192 goto err_release_metapage;
195 bmp->db_maxlevel = le32_to_cpu(dbmp_le->dn_maxlevel);
196 bmp->db_maxag = le32_to_cpu(dbmp_le->dn_maxag);
197 bmp->db_agpref = le32_to_cpu(dbmp_le->dn_agpref);
198 if (bmp->db_maxag >= MAXAG || bmp->db_maxag < 0 ||
199 bmp->db_agpref >= MAXAG || bmp->db_agpref < 0) {
201 goto err_release_metapage;
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. */
279 JFS_SBI(ipbmap->i_sb)->bmap = NULL;
287 int dbSync(struct inode *ipbmap)
289 struct dbmap_disk *dbmp_le;
290 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
295 * write bmap global control page
297 /* get the buffer for the on-disk bmap descriptor. */
298 mp = read_metapage(ipbmap,
299 BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
302 jfs_err("dbSync: read_metapage failed!");
305 /* copy the in-memory version of the bmap to the on-disk version */
306 dbmp_le = (struct dbmap_disk *) mp->data;
307 dbmp_le->dn_mapsize = cpu_to_le64(bmp->db_mapsize);
308 dbmp_le->dn_nfree = cpu_to_le64(bmp->db_nfree);
309 dbmp_le->dn_l2nbperpage = cpu_to_le32(bmp->db_l2nbperpage);
310 dbmp_le->dn_numag = cpu_to_le32(bmp->db_numag);
311 dbmp_le->dn_maxlevel = cpu_to_le32(bmp->db_maxlevel);
312 dbmp_le->dn_maxag = cpu_to_le32(bmp->db_maxag);
313 dbmp_le->dn_agpref = cpu_to_le32(bmp->db_agpref);
314 dbmp_le->dn_aglevel = cpu_to_le32(bmp->db_aglevel);
315 dbmp_le->dn_agheight = cpu_to_le32(bmp->db_agheight);
316 dbmp_le->dn_agwidth = cpu_to_le32(bmp->db_agwidth);
317 dbmp_le->dn_agstart = cpu_to_le32(bmp->db_agstart);
318 dbmp_le->dn_agl2size = cpu_to_le32(bmp->db_agl2size);
319 for (i = 0; i < MAXAG; i++)
320 dbmp_le->dn_agfree[i] = cpu_to_le64(bmp->db_agfree[i]);
321 dbmp_le->dn_agsize = cpu_to_le64(bmp->db_agsize);
322 dbmp_le->dn_maxfreebud = bmp->db_maxfreebud;
324 /* write the buffer */
328 * write out dirty pages of bmap
330 filemap_write_and_wait(ipbmap->i_mapping);
332 diWriteSpecial(ipbmap, 0);
340 * FUNCTION: free the specified block range from the working block
343 * the blocks will be free from the working map one dmap
347 * ip - pointer to in-core inode;
348 * blkno - starting block number to be freed.
349 * nblocks - number of blocks to be freed.
355 int dbFree(struct inode *ip, s64 blkno, s64 nblocks)
361 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
362 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
363 struct super_block *sb = ipbmap->i_sb;
365 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
367 /* block to be freed better be within the mapsize. */
368 if (unlikely((blkno == 0) || (blkno + nblocks > bmp->db_mapsize))) {
369 IREAD_UNLOCK(ipbmap);
370 printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
371 (unsigned long long) blkno,
372 (unsigned long long) nblocks);
373 jfs_error(ip->i_sb, "block to be freed is outside the map\n");
378 * TRIM the blocks, when mounted with discard option
380 if (JFS_SBI(sb)->flag & JFS_DISCARD)
381 if (JFS_SBI(sb)->minblks_trim <= nblocks)
382 jfs_issue_discard(ipbmap, blkno, nblocks);
385 * free the blocks a dmap at a time.
388 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
389 /* release previous dmap if any */
394 /* get the buffer for the current dmap. */
395 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
396 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
398 IREAD_UNLOCK(ipbmap);
401 dp = (struct dmap *) mp->data;
403 /* determine the number of blocks to be freed from
406 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
408 /* free the blocks. */
409 if ((rc = dbFreeDmap(bmp, dp, blkno, nb))) {
410 jfs_error(ip->i_sb, "error in block map\n");
411 release_metapage(mp);
412 IREAD_UNLOCK(ipbmap);
417 /* write the last buffer. */
421 IREAD_UNLOCK(ipbmap);
428 * NAME: dbUpdatePMap()
430 * FUNCTION: update the allocation state (free or allocate) of the
431 * specified block range in the persistent block allocation map.
433 * the blocks will be updated in the persistent map one
437 * ipbmap - pointer to in-core inode for the block map.
438 * free - 'true' if block range is to be freed from the persistent
439 * map; 'false' if it is to be allocated.
440 * blkno - starting block number of the range.
441 * nblocks - number of contiguous blocks in the range.
442 * tblk - transaction block;
449 dbUpdatePMap(struct inode *ipbmap,
450 int free, s64 blkno, s64 nblocks, struct tblock * tblk)
452 int nblks, dbitno, wbitno, rbits;
453 int word, nbits, nwords;
454 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
455 s64 lblkno, rem, lastlblkno;
460 int lsn, difft, diffp;
463 /* the blocks better be within the mapsize. */
464 if (blkno + nblocks > bmp->db_mapsize) {
465 printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
466 (unsigned long long) blkno,
467 (unsigned long long) nblocks);
468 jfs_error(ipbmap->i_sb, "blocks are outside the map\n");
472 /* compute delta of transaction lsn from log syncpt */
474 log = (struct jfs_log *) JFS_SBI(tblk->sb)->log;
475 logdiff(difft, lsn, log);
478 * update the block state a dmap at a time.
482 for (rem = nblocks; rem > 0; rem -= nblks, blkno += nblks) {
483 /* get the buffer for the current dmap. */
484 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
485 if (lblkno != lastlblkno) {
490 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE,
494 metapage_wait_for_io(mp);
496 dp = (struct dmap *) mp->data;
498 /* determine the bit number and word within the dmap of
499 * the starting block. also determine how many blocks
500 * are to be updated within this dmap.
502 dbitno = blkno & (BPERDMAP - 1);
503 word = dbitno >> L2DBWORD;
504 nblks = min(rem, (s64)BPERDMAP - dbitno);
506 /* update the bits of the dmap words. the first and last
507 * words may only have a subset of their bits updated. if
508 * this is the case, we'll work against that word (i.e.
509 * partial first and/or last) only in a single pass. a
510 * single pass will also be used to update all words that
511 * are to have all their bits updated.
513 for (rbits = nblks; rbits > 0;
514 rbits -= nbits, dbitno += nbits) {
515 /* determine the bit number within the word and
516 * the number of bits within the word.
518 wbitno = dbitno & (DBWORD - 1);
519 nbits = min(rbits, DBWORD - wbitno);
521 /* check if only part of the word is to be updated. */
522 if (nbits < DBWORD) {
523 /* update (free or allocate) the bits
527 (ONES << (DBWORD - nbits) >> wbitno);
537 /* one or more words are to have all
538 * their bits updated. determine how
539 * many words and how many bits.
541 nwords = rbits >> L2DBWORD;
542 nbits = nwords << L2DBWORD;
544 /* update (free or allocate) the bits
548 memset(&dp->pmap[word], 0,
551 memset(&dp->pmap[word], (int) ONES,
561 if (lblkno == lastlblkno)
566 LOGSYNC_LOCK(log, flags);
568 /* inherit older/smaller lsn */
569 logdiff(diffp, mp->lsn, log);
573 /* move bp after tblock in logsync list */
574 list_move(&mp->synclist, &tblk->synclist);
577 /* inherit younger/larger clsn */
578 logdiff(difft, tblk->clsn, log);
579 logdiff(diffp, mp->clsn, log);
581 mp->clsn = tblk->clsn;
586 /* insert bp after tblock in logsync list */
588 list_add(&mp->synclist, &tblk->synclist);
590 mp->clsn = tblk->clsn;
592 LOGSYNC_UNLOCK(log, flags);
595 /* write the last buffer. */
607 * FUNCTION: find the preferred allocation group for new allocations.
609 * Within the allocation groups, we maintain a preferred
610 * allocation group which consists of a group with at least
611 * average free space. It is the preferred group that we target
612 * new inode allocation towards. The tie-in between inode
613 * allocation and block allocation occurs as we allocate the
614 * first (data) block of an inode and specify the inode (block)
615 * as the allocation hint for this block.
617 * We try to avoid having more than one open file growing in
618 * an allocation group, as this will lead to fragmentation.
619 * This differs from the old OS/2 method of trying to keep
620 * empty ags around for large allocations.
623 * ipbmap - pointer to in-core inode for the block map.
626 * the preferred allocation group number.
628 int dbNextAG(struct inode *ipbmap)
635 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
639 /* determine the average number of free blocks within the ags. */
640 avgfree = (u32)bmp->db_nfree / bmp->db_numag;
643 * if the current preferred ag does not have an active allocator
644 * and has at least average freespace, return it
646 agpref = bmp->db_agpref;
647 if ((atomic_read(&bmp->db_active[agpref]) == 0) &&
648 (bmp->db_agfree[agpref] >= avgfree))
651 /* From the last preferred ag, find the next one with at least
652 * average free space.
654 for (i = 0 ; i < bmp->db_numag; i++, agpref++) {
655 if (agpref == bmp->db_numag)
658 if (atomic_read(&bmp->db_active[agpref]))
659 /* open file is currently growing in this ag */
661 if (bmp->db_agfree[agpref] >= avgfree) {
662 /* Return this one */
663 bmp->db_agpref = agpref;
665 } else if (bmp->db_agfree[agpref] > hwm) {
666 /* Less than avg. freespace, but best so far */
667 hwm = bmp->db_agfree[agpref];
673 * If no inactive ag was found with average freespace, use the
677 bmp->db_agpref = next_best;
678 /* else leave db_agpref unchanged */
682 /* return the preferred group.
684 return (bmp->db_agpref);
690 * FUNCTION: attempt to allocate a specified number of contiguous free
691 * blocks from the working allocation block map.
693 * the block allocation policy uses hints and a multi-step
696 * for allocation requests smaller than the number of blocks
697 * per dmap, we first try to allocate the new blocks
698 * immediately following the hint. if these blocks are not
699 * available, we try to allocate blocks near the hint. if
700 * no blocks near the hint are available, we next try to
701 * allocate within the same dmap as contains the hint.
703 * if no blocks are available in the dmap or the allocation
704 * request is larger than the dmap size, we try to allocate
705 * within the same allocation group as contains the hint. if
706 * this does not succeed, we finally try to allocate anywhere
707 * within the aggregate.
709 * we also try to allocate anywhere within the aggregate
710 * for allocation requests larger than the allocation group
711 * size or requests that specify no hint value.
714 * ip - pointer to in-core inode;
715 * hint - allocation hint.
716 * nblocks - number of contiguous blocks in the range.
717 * results - on successful return, set to the starting block number
718 * of the newly allocated contiguous range.
722 * -ENOSPC - insufficient disk resources
725 int dbAlloc(struct inode *ip, s64 hint, s64 nblocks, s64 * results)
728 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
737 /* assert that nblocks is valid */
740 /* get the log2 number of blocks to be allocated.
741 * if the number of blocks is not a log2 multiple,
742 * it will be rounded up to the next log2 multiple.
744 l2nb = BLKSTOL2(nblocks);
746 bmp = JFS_SBI(ip->i_sb)->bmap;
748 mapSize = bmp->db_mapsize;
750 /* the hint should be within the map */
751 if (hint >= mapSize) {
752 jfs_error(ip->i_sb, "the hint is outside the map\n");
756 /* if the number of blocks to be allocated is greater than the
757 * allocation group size, try to allocate anywhere.
759 if (l2nb > bmp->db_agl2size) {
760 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
762 rc = dbAllocAny(bmp, nblocks, l2nb, results);
768 * If no hint, let dbNextAG recommend an allocation group
773 /* we would like to allocate close to the hint. adjust the
774 * hint to the block following the hint since the allocators
775 * will start looking for free space starting at this point.
779 if (blkno >= bmp->db_mapsize)
782 agno = blkno >> bmp->db_agl2size;
784 /* check if blkno crosses over into a new allocation group.
785 * if so, check if we should allow allocations within this
788 if ((blkno & (bmp->db_agsize - 1)) == 0)
789 /* check if the AG is currently being written to.
790 * if so, call dbNextAG() to find a non-busy
791 * AG with sufficient free space.
793 if (atomic_read(&bmp->db_active[agno]))
796 /* check if the allocation request size can be satisfied from a
797 * single dmap. if so, try to allocate from the dmap containing
798 * the hint using a tiered strategy.
800 if (nblocks <= BPERDMAP) {
801 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
803 /* get the buffer for the dmap containing the hint.
806 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
807 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
811 dp = (struct dmap *) mp->data;
813 /* first, try to satisfy the allocation request with the
814 * blocks beginning at the hint.
816 if ((rc = dbAllocNext(bmp, dp, blkno, (int) nblocks))
820 mark_metapage_dirty(mp);
823 release_metapage(mp);
827 writers = atomic_read(&bmp->db_active[agno]);
829 ((writers == 1) && (JFS_IP(ip)->active_ag != agno))) {
831 * Someone else is writing in this allocation
832 * group. To avoid fragmenting, try another ag
834 release_metapage(mp);
835 IREAD_UNLOCK(ipbmap);
839 /* next, try to satisfy the allocation request with blocks
843 dbAllocNear(bmp, dp, blkno, (int) nblocks, l2nb, results))
846 mark_metapage_dirty(mp);
848 release_metapage(mp);
852 /* try to satisfy the allocation request with blocks within
853 * the same dmap as the hint.
855 if ((rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results))
858 mark_metapage_dirty(mp);
860 release_metapage(mp);
864 release_metapage(mp);
865 IREAD_UNLOCK(ipbmap);
868 /* try to satisfy the allocation request with blocks within
869 * the same allocation group as the hint.
871 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
872 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) != -ENOSPC)
875 IWRITE_UNLOCK(ipbmap);
880 * Let dbNextAG recommend a preferred allocation group
882 agno = dbNextAG(ipbmap);
883 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
885 /* Try to allocate within this allocation group. if that fails, try to
886 * allocate anywhere in the map.
888 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) == -ENOSPC)
889 rc = dbAllocAny(bmp, nblocks, l2nb, results);
892 IWRITE_UNLOCK(ipbmap);
897 IREAD_UNLOCK(ipbmap);
905 * FUNCTION: attempt to extend a current allocation by a specified
908 * this routine attempts to satisfy the allocation request
909 * by first trying to extend the existing allocation in
910 * place by allocating the additional blocks as the blocks
911 * immediately following the current allocation. if these
912 * blocks are not available, this routine will attempt to
913 * allocate a new set of contiguous blocks large enough
914 * to cover the existing allocation plus the additional
915 * number of blocks required.
918 * ip - pointer to in-core inode requiring allocation.
919 * blkno - starting block of the current allocation.
920 * nblocks - number of contiguous blocks within the current
922 * addnblocks - number of blocks to add to the allocation.
923 * results - on successful return, set to the starting block number
924 * of the existing allocation if the existing allocation
925 * was extended in place or to a newly allocated contiguous
926 * range if the existing allocation could not be extended
931 * -ENOSPC - insufficient disk resources
935 dbReAlloc(struct inode *ip,
936 s64 blkno, s64 nblocks, s64 addnblocks, s64 * results)
940 /* try to extend the allocation in place.
942 if ((rc = dbExtend(ip, blkno, nblocks, addnblocks)) == 0) {
950 /* could not extend the allocation in place, so allocate a
951 * new set of blocks for the entire request (i.e. try to get
952 * a range of contiguous blocks large enough to cover the
953 * existing allocation plus the additional blocks.)
956 (ip, blkno + nblocks - 1, addnblocks + nblocks, results));
963 * FUNCTION: attempt to extend a current allocation by a specified
966 * this routine attempts to satisfy the allocation request
967 * by first trying to extend the existing allocation in
968 * place by allocating the additional blocks as the blocks
969 * immediately following the current allocation.
972 * ip - pointer to in-core inode requiring allocation.
973 * blkno - starting block of the current allocation.
974 * nblocks - number of contiguous blocks within the current
976 * addnblocks - number of blocks to add to the allocation.
980 * -ENOSPC - insufficient disk resources
983 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks)
985 struct jfs_sb_info *sbi = JFS_SBI(ip->i_sb);
986 s64 lblkno, lastblkno, extblkno;
991 struct inode *ipbmap = sbi->ipbmap;
995 * We don't want a non-aligned extent to cross a page boundary
997 if (((rel_block = blkno & (sbi->nbperpage - 1))) &&
998 (rel_block + nblocks + addnblocks > sbi->nbperpage))
1001 /* get the last block of the current allocation */
1002 lastblkno = blkno + nblocks - 1;
1004 /* determine the block number of the block following
1005 * the existing allocation.
1007 extblkno = lastblkno + 1;
1009 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
1011 /* better be within the file system */
1013 if (lastblkno < 0 || lastblkno >= bmp->db_mapsize) {
1014 IREAD_UNLOCK(ipbmap);
1015 jfs_error(ip->i_sb, "the block is outside the filesystem\n");
1019 /* we'll attempt to extend the current allocation in place by
1020 * allocating the additional blocks as the blocks immediately
1021 * following the current allocation. we only try to extend the
1022 * current allocation in place if the number of additional blocks
1023 * can fit into a dmap, the last block of the current allocation
1024 * is not the last block of the file system, and the start of the
1025 * inplace extension is not on an allocation group boundary.
1027 if (addnblocks > BPERDMAP || extblkno >= bmp->db_mapsize ||
1028 (extblkno & (bmp->db_agsize - 1)) == 0) {
1029 IREAD_UNLOCK(ipbmap);
1033 /* get the buffer for the dmap containing the first block
1036 lblkno = BLKTODMAP(extblkno, bmp->db_l2nbperpage);
1037 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
1039 IREAD_UNLOCK(ipbmap);
1043 dp = (struct dmap *) mp->data;
1045 /* try to allocate the blocks immediately following the
1046 * current allocation.
1048 rc = dbAllocNext(bmp, dp, extblkno, (int) addnblocks);
1050 IREAD_UNLOCK(ipbmap);
1052 /* were we successful ? */
1056 /* we were not successful */
1057 release_metapage(mp);
1064 * NAME: dbAllocNext()
1066 * FUNCTION: attempt to allocate the blocks of the specified block
1067 * range within a dmap.
1070 * bmp - pointer to bmap descriptor
1071 * dp - pointer to dmap.
1072 * blkno - starting block number of the range.
1073 * nblocks - number of contiguous free blocks of the range.
1077 * -ENOSPC - insufficient disk resources
1080 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1082 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
1085 int dbitno, word, rembits, nb, nwords, wbitno, nw;
1090 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1091 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1095 /* pick up a pointer to the leaves of the dmap tree.
1097 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1099 /* determine the bit number and word within the dmap of the
1102 dbitno = blkno & (BPERDMAP - 1);
1103 word = dbitno >> L2DBWORD;
1105 /* check if the specified block range is contained within
1108 if (dbitno + nblocks > BPERDMAP)
1111 /* check if the starting leaf indicates that anything
1114 if (leaf[word] == NOFREE)
1117 /* check the dmaps words corresponding to block range to see
1118 * if the block range is free. not all bits of the first and
1119 * last words may be contained within the block range. if this
1120 * is the case, we'll work against those words (i.e. partial first
1121 * and/or last) on an individual basis (a single pass) and examine
1122 * the actual bits to determine if they are free. a single pass
1123 * will be used for all dmap words fully contained within the
1124 * specified range. within this pass, the leaves of the dmap
1125 * tree will be examined to determine if the blocks are free. a
1126 * single leaf may describe the free space of multiple dmap
1127 * words, so we may visit only a subset of the actual leaves
1128 * corresponding to the dmap words of the block range.
1130 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
1131 /* determine the bit number within the word and
1132 * the number of bits within the word.
1134 wbitno = dbitno & (DBWORD - 1);
1135 nb = min(rembits, DBWORD - wbitno);
1137 /* check if only part of the word is to be examined.
1140 /* check if the bits are free.
1142 mask = (ONES << (DBWORD - nb) >> wbitno);
1143 if ((mask & ~le32_to_cpu(dp->wmap[word])) != mask)
1148 /* one or more dmap words are fully contained
1149 * within the block range. determine how many
1150 * words and how many bits.
1152 nwords = rembits >> L2DBWORD;
1153 nb = nwords << L2DBWORD;
1155 /* now examine the appropriate leaves to determine
1156 * if the blocks are free.
1158 while (nwords > 0) {
1159 /* does the leaf describe any free space ?
1161 if (leaf[word] < BUDMIN)
1164 /* determine the l2 number of bits provided
1168 min_t(int, leaf[word], NLSTOL2BSZ(nwords));
1170 /* determine how many words were handled.
1172 nw = BUDSIZE(l2size, BUDMIN);
1180 /* allocate the blocks.
1182 return (dbAllocDmap(bmp, dp, blkno, nblocks));
1187 * NAME: dbAllocNear()
1189 * FUNCTION: attempt to allocate a number of contiguous free blocks near
1190 * a specified block (hint) within a dmap.
1192 * starting with the dmap leaf that covers the hint, we'll
1193 * check the next four contiguous leaves for sufficient free
1194 * space. if sufficient free space is found, we'll allocate
1195 * the desired free space.
1198 * bmp - pointer to bmap descriptor
1199 * dp - pointer to dmap.
1200 * blkno - block number to allocate near.
1201 * nblocks - actual number of contiguous free blocks desired.
1202 * l2nb - log2 number of contiguous free blocks desired.
1203 * results - on successful return, set to the starting block number
1204 * of the newly allocated range.
1208 * -ENOSPC - insufficient disk resources
1211 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1214 dbAllocNear(struct bmap * bmp,
1215 struct dmap * dp, s64 blkno, int nblocks, int l2nb, s64 * results)
1217 int word, lword, rc;
1220 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1221 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1225 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1227 /* determine the word within the dmap that holds the hint
1228 * (i.e. blkno). also, determine the last word in the dmap
1229 * that we'll include in our examination.
1231 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
1232 lword = min(word + 4, LPERDMAP);
1234 /* examine the leaves for sufficient free space.
1236 for (; word < lword; word++) {
1237 /* does the leaf describe sufficient free space ?
1239 if (leaf[word] < l2nb)
1242 /* determine the block number within the file system
1243 * of the first block described by this dmap word.
1245 blkno = le64_to_cpu(dp->start) + (word << L2DBWORD);
1247 /* if not all bits of the dmap word are free, get the
1248 * starting bit number within the dmap word of the required
1249 * string of free bits and adjust the block number with the
1252 if (leaf[word] < BUDMIN)
1254 dbFindBits(le32_to_cpu(dp->wmap[word]), l2nb);
1256 /* allocate the blocks.
1258 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1271 * FUNCTION: attempt to allocate the specified number of contiguous
1272 * free blocks within the specified allocation group.
1274 * unless the allocation group size is equal to the number
1275 * of blocks per dmap, the dmap control pages will be used to
1276 * find the required free space, if available. we start the
1277 * search at the highest dmap control page level which
1278 * distinctly describes the allocation group's free space
1279 * (i.e. the highest level at which the allocation group's
1280 * free space is not mixed in with that of any other group).
1281 * in addition, we start the search within this level at a
1282 * height of the dmapctl dmtree at which the nodes distinctly
1283 * describe the allocation group's free space. at this height,
1284 * the allocation group's free space may be represented by 1
1285 * or two sub-trees, depending on the allocation group size.
1286 * we search the top nodes of these subtrees left to right for
1287 * sufficient free space. if sufficient free space is found,
1288 * the subtree is searched to find the leftmost leaf that
1289 * has free space. once we have made it to the leaf, we
1290 * move the search to the next lower level dmap control page
1291 * corresponding to this leaf. we continue down the dmap control
1292 * pages until we find the dmap that contains or starts the
1293 * sufficient free space and we allocate at this dmap.
1295 * if the allocation group size is equal to the dmap size,
1296 * we'll start at the dmap corresponding to the allocation
1297 * group and attempt the allocation at this level.
1299 * the dmap control page search is also not performed if the
1300 * allocation group is completely free and we go to the first
1301 * dmap of the allocation group to do the allocation. this is
1302 * done because the allocation group may be part (not the first
1303 * part) of a larger binary buddy system, causing the dmap
1304 * control pages to indicate no free space (NOFREE) within
1305 * the allocation group.
1308 * bmp - pointer to bmap descriptor
1309 * agno - allocation group number.
1310 * nblocks - actual number of contiguous free blocks desired.
1311 * l2nb - log2 number of contiguous free blocks desired.
1312 * results - on successful return, set to the starting block number
1313 * of the newly allocated range.
1317 * -ENOSPC - insufficient disk resources
1320 * note: IWRITE_LOCK(ipmap) held on entry/exit;
1323 dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, s64 * results)
1325 struct metapage *mp;
1326 struct dmapctl *dcp;
1327 int rc, ti, i, k, m, n, agperlev;
1331 /* allocation request should not be for more than the
1332 * allocation group size.
1334 if (l2nb > bmp->db_agl2size) {
1335 jfs_error(bmp->db_ipbmap->i_sb,
1336 "allocation request is larger than the allocation group size\n");
1340 /* determine the starting block number of the allocation
1343 blkno = (s64) agno << bmp->db_agl2size;
1345 /* check if the allocation group size is the minimum allocation
1346 * group size or if the allocation group is completely free. if
1347 * the allocation group size is the minimum size of BPERDMAP (i.e.
1348 * 1 dmap), there is no need to search the dmap control page (below)
1349 * that fully describes the allocation group since the allocation
1350 * group is already fully described by a dmap. in this case, we
1351 * just call dbAllocCtl() to search the dmap tree and allocate the
1352 * required space if available.
1354 * if the allocation group is completely free, dbAllocCtl() is
1355 * also called to allocate the required space. this is done for
1356 * two reasons. first, it makes no sense searching the dmap control
1357 * pages for free space when we know that free space exists. second,
1358 * the dmap control pages may indicate that the allocation group
1359 * has no free space if the allocation group is part (not the first
1360 * part) of a larger binary buddy system.
1362 if (bmp->db_agsize == BPERDMAP
1363 || bmp->db_agfree[agno] == bmp->db_agsize) {
1364 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1365 if ((rc == -ENOSPC) &&
1366 (bmp->db_agfree[agno] == bmp->db_agsize)) {
1367 printk(KERN_ERR "blkno = %Lx, blocks = %Lx\n",
1368 (unsigned long long) blkno,
1369 (unsigned long long) nblocks);
1370 jfs_error(bmp->db_ipbmap->i_sb,
1371 "dbAllocCtl failed in free AG\n");
1376 /* the buffer for the dmap control page that fully describes the
1379 lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, bmp->db_aglevel);
1380 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1383 dcp = (struct dmapctl *) mp->data;
1384 budmin = dcp->budmin;
1386 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1387 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
1388 release_metapage(mp);
1392 /* search the subtree(s) of the dmap control page that describes
1393 * the allocation group, looking for sufficient free space. to begin,
1394 * determine how many allocation groups are represented in a dmap
1395 * control page at the control page level (i.e. L0, L1, L2) that
1396 * fully describes an allocation group. next, determine the starting
1397 * tree index of this allocation group within the control page.
1400 (1 << (L2LPERCTL - (bmp->db_agheight << 1))) / bmp->db_agwidth;
1401 ti = bmp->db_agstart + bmp->db_agwidth * (agno & (agperlev - 1));
1403 /* dmap control page trees fan-out by 4 and a single allocation
1404 * group may be described by 1 or 2 subtrees within the ag level
1405 * dmap control page, depending upon the ag size. examine the ag's
1406 * subtrees for sufficient free space, starting with the leftmost
1409 for (i = 0; i < bmp->db_agwidth; i++, ti++) {
1410 /* is there sufficient free space ?
1412 if (l2nb > dcp->stree[ti])
1415 /* sufficient free space found in a subtree. now search down
1416 * the subtree to find the leftmost leaf that describes this
1419 for (k = bmp->db_agheight; k > 0; k--) {
1420 for (n = 0, m = (ti << 2) + 1; n < 4; n++) {
1421 if (l2nb <= dcp->stree[m + n]) {
1427 jfs_error(bmp->db_ipbmap->i_sb,
1428 "failed descending stree\n");
1429 release_metapage(mp);
1434 /* determine the block number within the file system
1435 * that corresponds to this leaf.
1437 if (bmp->db_aglevel == 2)
1439 else if (bmp->db_aglevel == 1)
1440 blkno &= ~(MAXL1SIZE - 1);
1441 else /* bmp->db_aglevel == 0 */
1442 blkno &= ~(MAXL0SIZE - 1);
1445 ((s64) (ti - le32_to_cpu(dcp->leafidx))) << budmin;
1447 /* release the buffer in preparation for going down
1448 * the next level of dmap control pages.
1450 release_metapage(mp);
1452 /* check if we need to continue to search down the lower
1453 * level dmap control pages. we need to if the number of
1454 * blocks required is less than maximum number of blocks
1455 * described at the next lower level.
1457 if (l2nb < budmin) {
1459 /* search the lower level dmap control pages to get
1460 * the starting block number of the dmap that
1461 * contains or starts off the free space.
1464 dbFindCtl(bmp, l2nb, bmp->db_aglevel - 1,
1466 if (rc == -ENOSPC) {
1467 jfs_error(bmp->db_ipbmap->i_sb,
1468 "control page inconsistent\n");
1475 /* allocate the blocks.
1477 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1478 if (rc == -ENOSPC) {
1479 jfs_error(bmp->db_ipbmap->i_sb,
1480 "unable to allocate blocks\n");
1486 /* no space in the allocation group. release the buffer and
1489 release_metapage(mp);
1496 * NAME: dbAllocAny()
1498 * FUNCTION: attempt to allocate the specified number of contiguous
1499 * free blocks anywhere in the file system.
1501 * dbAllocAny() attempts to find the sufficient free space by
1502 * searching down the dmap control pages, starting with the
1503 * highest level (i.e. L0, L1, L2) control page. if free space
1504 * large enough to satisfy the desired free space is found, the
1505 * desired free space is allocated.
1508 * bmp - pointer to bmap descriptor
1509 * nblocks - actual number of contiguous free blocks desired.
1510 * l2nb - log2 number of contiguous free blocks desired.
1511 * results - on successful return, set to the starting block number
1512 * of the newly allocated range.
1516 * -ENOSPC - insufficient disk resources
1519 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1521 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results)
1526 /* starting with the top level dmap control page, search
1527 * down the dmap control levels for sufficient free space.
1528 * if free space is found, dbFindCtl() returns the starting
1529 * block number of the dmap that contains or starts off the
1530 * range of free space.
1532 if ((rc = dbFindCtl(bmp, l2nb, bmp->db_maxlevel, &blkno)))
1535 /* allocate the blocks.
1537 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1538 if (rc == -ENOSPC) {
1539 jfs_error(bmp->db_ipbmap->i_sb, "unable to allocate blocks\n");
1547 * NAME: dbDiscardAG()
1549 * FUNCTION: attempt to discard (TRIM) all free blocks of specific AG
1552 * 1) allocate blocks, as large as possible and save them
1553 * while holding IWRITE_LOCK on ipbmap
1554 * 2) trim all these saved block/length values
1555 * 3) mark the blocks free again
1558 * - we work only on one ag at some time, minimizing how long we
1559 * need to lock ipbmap
1560 * - reading / writing the fs is possible most time, even on
1564 * - we write two times to the dmapctl and dmap pages
1565 * - but for me, this seems the best way, better ideas?
1569 * ip - pointer to in-core inode
1571 * minlen - minimum value of contiguous blocks
1574 * s64 - actual number of blocks trimmed
1576 s64 dbDiscardAG(struct inode *ip, int agno, s64 minlen)
1578 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
1579 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
1583 struct super_block *sb = ipbmap->i_sb;
1590 /* max blkno / nblocks pairs to trim */
1591 int count = 0, range_cnt;
1594 /* prevent others from writing new stuff here, while trimming */
1595 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
1597 nblocks = bmp->db_agfree[agno];
1598 max_ranges = nblocks;
1599 do_div(max_ranges, minlen);
1600 range_cnt = min_t(u64, max_ranges + 1, 32 * 1024);
1601 totrim = kmalloc_array(range_cnt, sizeof(struct range2trim), GFP_NOFS);
1602 if (totrim == NULL) {
1603 jfs_error(bmp->db_ipbmap->i_sb, "no memory for trim array\n");
1604 IWRITE_UNLOCK(ipbmap);
1609 while (nblocks >= minlen) {
1610 l2nb = BLKSTOL2(nblocks);
1612 /* 0 = okay, -EIO = fatal, -ENOSPC -> try smaller block */
1613 rc = dbAllocAG(bmp, agno, nblocks, l2nb, &blkno);
1616 tt->nblocks = nblocks;
1619 /* the whole ag is free, trim now */
1620 if (bmp->db_agfree[agno] == 0)
1623 /* give a hint for the next while */
1624 nblocks = bmp->db_agfree[agno];
1626 } else if (rc == -ENOSPC) {
1627 /* search for next smaller log2 block */
1628 l2nb = BLKSTOL2(nblocks) - 1;
1629 nblocks = 1LL << l2nb;
1631 /* Trim any already allocated blocks */
1632 jfs_error(bmp->db_ipbmap->i_sb, "-EIO\n");
1636 /* check, if our trim array is full */
1637 if (unlikely(count >= range_cnt - 1))
1640 IWRITE_UNLOCK(ipbmap);
1642 tt->nblocks = 0; /* mark the current end */
1643 for (tt = totrim; tt->nblocks != 0; tt++) {
1644 /* when mounted with online discard, dbFree() will
1645 * call jfs_issue_discard() itself */
1646 if (!(JFS_SBI(sb)->flag & JFS_DISCARD))
1647 jfs_issue_discard(ip, tt->blkno, tt->nblocks);
1648 dbFree(ip, tt->blkno, tt->nblocks);
1649 trimmed += tt->nblocks;
1659 * FUNCTION: starting at a specified dmap control page level and block
1660 * number, search down the dmap control levels for a range of
1661 * contiguous free blocks large enough to satisfy an allocation
1662 * request for the specified number of free blocks.
1664 * if sufficient contiguous free blocks are found, this routine
1665 * returns the starting block number within a dmap page that
1666 * contains or starts a range of contiqious free blocks that
1667 * is sufficient in size.
1670 * bmp - pointer to bmap descriptor
1671 * level - starting dmap control page level.
1672 * l2nb - log2 number of contiguous free blocks desired.
1673 * *blkno - on entry, starting block number for conducting the search.
1674 * on successful return, the first block within a dmap page
1675 * that contains or starts a range of contiguous free blocks.
1679 * -ENOSPC - insufficient disk resources
1682 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1684 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno)
1686 int rc, leafidx, lev;
1688 struct dmapctl *dcp;
1690 struct metapage *mp;
1692 /* starting at the specified dmap control page level and block
1693 * number, search down the dmap control levels for the starting
1694 * block number of a dmap page that contains or starts off
1695 * sufficient free blocks.
1697 for (lev = level, b = *blkno; lev >= 0; lev--) {
1698 /* get the buffer of the dmap control page for the block
1699 * number and level (i.e. L0, L1, L2).
1701 lblkno = BLKTOCTL(b, bmp->db_l2nbperpage, lev);
1702 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1705 dcp = (struct dmapctl *) mp->data;
1706 budmin = dcp->budmin;
1708 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1709 jfs_error(bmp->db_ipbmap->i_sb,
1710 "Corrupt dmapctl page\n");
1711 release_metapage(mp);
1715 /* search the tree within the dmap control page for
1716 * sufficient free space. if sufficient free space is found,
1717 * dbFindLeaf() returns the index of the leaf at which
1718 * free space was found.
1720 rc = dbFindLeaf((dmtree_t *) dcp, l2nb, &leafidx);
1722 /* release the buffer.
1724 release_metapage(mp);
1730 jfs_error(bmp->db_ipbmap->i_sb,
1731 "dmap inconsistent\n");
1737 /* adjust the block number to reflect the location within
1738 * the dmap control page (i.e. the leaf) at which free
1741 b += (((s64) leafidx) << budmin);
1743 /* we stop the search at this dmap control page level if
1744 * the number of blocks required is greater than or equal
1745 * to the maximum number of blocks described at the next
1758 * NAME: dbAllocCtl()
1760 * FUNCTION: attempt to allocate a specified number of contiguous
1761 * blocks starting within a specific dmap.
1763 * this routine is called by higher level routines that search
1764 * the dmap control pages above the actual dmaps for contiguous
1765 * free space. the result of successful searches by these
1766 * routines are the starting block numbers within dmaps, with
1767 * the dmaps themselves containing the desired contiguous free
1768 * space or starting a contiguous free space of desired size
1769 * that is made up of the blocks of one or more dmaps. these
1770 * calls should not fail due to insufficent resources.
1772 * this routine is called in some cases where it is not known
1773 * whether it will fail due to insufficient resources. more
1774 * specifically, this occurs when allocating from an allocation
1775 * group whose size is equal to the number of blocks per dmap.
1776 * in this case, the dmap control pages are not examined prior
1777 * to calling this routine (to save pathlength) and the call
1780 * for a request size that fits within a dmap, this routine relies
1781 * upon the dmap's dmtree to find the requested contiguous free
1782 * space. for request sizes that are larger than a dmap, the
1783 * requested free space will start at the first block of the
1784 * first dmap (i.e. blkno).
1787 * bmp - pointer to bmap descriptor
1788 * nblocks - actual number of contiguous free blocks to allocate.
1789 * l2nb - log2 number of contiguous free blocks to allocate.
1790 * blkno - starting block number of the dmap to start the allocation
1792 * results - on successful return, set to the starting block number
1793 * of the newly allocated range.
1797 * -ENOSPC - insufficient disk resources
1800 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1803 dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, s64 * results)
1807 struct metapage *mp;
1810 /* check if the allocation request is confined to a single dmap.
1812 if (l2nb <= L2BPERDMAP) {
1813 /* get the buffer for the dmap.
1815 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
1816 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1819 dp = (struct dmap *) mp->data;
1821 /* try to allocate the blocks.
1823 rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results);
1825 mark_metapage_dirty(mp);
1827 release_metapage(mp);
1832 /* allocation request involving multiple dmaps. it must start on
1835 assert((blkno & (BPERDMAP - 1)) == 0);
1837 /* allocate the blocks dmap by dmap.
1839 for (n = nblocks, b = blkno; n > 0; n -= nb, b += nb) {
1840 /* get the buffer for the dmap.
1842 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1843 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1848 dp = (struct dmap *) mp->data;
1850 /* the dmap better be all free.
1852 if (dp->tree.stree[ROOT] != L2BPERDMAP) {
1853 release_metapage(mp);
1854 jfs_error(bmp->db_ipbmap->i_sb,
1855 "the dmap is not all free\n");
1860 /* determine how many blocks to allocate from this dmap.
1862 nb = min_t(s64, n, BPERDMAP);
1864 /* allocate the blocks from the dmap.
1866 if ((rc = dbAllocDmap(bmp, dp, b, nb))) {
1867 release_metapage(mp);
1871 /* write the buffer.
1876 /* set the results (starting block number) and return.
1881 /* something failed in handling an allocation request involving
1882 * multiple dmaps. we'll try to clean up by backing out any
1883 * allocation that has already happened for this request. if
1884 * we fail in backing out the allocation, we'll mark the file
1885 * system to indicate that blocks have been leaked.
1889 /* try to backout the allocations dmap by dmap.
1891 for (n = nblocks - n, b = blkno; n > 0;
1892 n -= BPERDMAP, b += BPERDMAP) {
1893 /* get the buffer for this dmap.
1895 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1896 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1898 /* could not back out. mark the file system
1899 * to indicate that we have leaked blocks.
1901 jfs_error(bmp->db_ipbmap->i_sb,
1902 "I/O Error: Block Leakage\n");
1905 dp = (struct dmap *) mp->data;
1907 /* free the blocks is this dmap.
1909 if (dbFreeDmap(bmp, dp, b, BPERDMAP)) {
1910 /* could not back out. mark the file system
1911 * to indicate that we have leaked blocks.
1913 release_metapage(mp);
1914 jfs_error(bmp->db_ipbmap->i_sb, "Block Leakage\n");
1918 /* write the buffer.
1928 * NAME: dbAllocDmapLev()
1930 * FUNCTION: attempt to allocate a specified number of contiguous blocks
1931 * from a specified dmap.
1933 * this routine checks if the contiguous blocks are available.
1934 * if so, nblocks of blocks are allocated; otherwise, ENOSPC is
1938 * mp - pointer to bmap descriptor
1939 * dp - pointer to dmap to attempt to allocate blocks from.
1940 * l2nb - log2 number of contiguous block desired.
1941 * nblocks - actual number of contiguous block desired.
1942 * results - on successful return, set to the starting block number
1943 * of the newly allocated range.
1947 * -ENOSPC - insufficient disk resources
1950 * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or
1951 * IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit;
1954 dbAllocDmapLev(struct bmap * bmp,
1955 struct dmap * dp, int nblocks, int l2nb, s64 * results)
1960 /* can't be more than a dmaps worth of blocks */
1961 assert(l2nb <= L2BPERDMAP);
1963 /* search the tree within the dmap page for sufficient
1964 * free space. if sufficient free space is found, dbFindLeaf()
1965 * returns the index of the leaf at which free space was found.
1967 if (dbFindLeaf((dmtree_t *) & dp->tree, l2nb, &leafidx))
1973 /* determine the block number within the file system corresponding
1974 * to the leaf at which free space was found.
1976 blkno = le64_to_cpu(dp->start) + (leafidx << L2DBWORD);
1978 /* if not all bits of the dmap word are free, get the starting
1979 * bit number within the dmap word of the required string of free
1980 * bits and adjust the block number with this value.
1982 if (dp->tree.stree[leafidx + LEAFIND] < BUDMIN)
1983 blkno += dbFindBits(le32_to_cpu(dp->wmap[leafidx]), l2nb);
1985 /* allocate the blocks */
1986 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1994 * NAME: dbAllocDmap()
1996 * FUNCTION: adjust the disk allocation map to reflect the allocation
1997 * of a specified block range within a dmap.
1999 * this routine allocates the specified blocks from the dmap
2000 * through a call to dbAllocBits(). if the allocation of the
2001 * block range causes the maximum string of free blocks within
2002 * the dmap to change (i.e. the value of the root of the dmap's
2003 * dmtree), this routine will cause this change to be reflected
2004 * up through the appropriate levels of the dmap control pages
2005 * by a call to dbAdjCtl() for the L0 dmap control page that
2009 * bmp - pointer to bmap descriptor
2010 * dp - pointer to dmap to allocate the block range from.
2011 * blkno - starting block number of the block to be allocated.
2012 * nblocks - number of blocks to be allocated.
2018 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2020 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2026 /* save the current value of the root (i.e. maximum free string)
2029 oldroot = dp->tree.stree[ROOT];
2031 /* allocate the specified (blocks) bits */
2032 dbAllocBits(bmp, dp, blkno, nblocks);
2034 /* if the root has not changed, done. */
2035 if (dp->tree.stree[ROOT] == oldroot)
2038 /* root changed. bubble the change up to the dmap control pages.
2039 * if the adjustment of the upper level control pages fails,
2040 * backout the bit allocation (thus making everything consistent).
2042 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 1, 0)))
2043 dbFreeBits(bmp, dp, blkno, nblocks);
2050 * NAME: dbFreeDmap()
2052 * FUNCTION: adjust the disk allocation map to reflect the allocation
2053 * of a specified block range within a dmap.
2055 * this routine frees the specified blocks from the dmap through
2056 * a call to dbFreeBits(). if the deallocation of the block range
2057 * causes the maximum string of free blocks within the dmap to
2058 * change (i.e. the value of the root of the dmap's dmtree), this
2059 * routine will cause this change to be reflected up through the
2060 * appropriate levels of the dmap control pages by a call to
2061 * dbAdjCtl() for the L0 dmap control page that covers this dmap.
2064 * bmp - pointer to bmap descriptor
2065 * dp - pointer to dmap to free the block range from.
2066 * blkno - starting block number of the block to be freed.
2067 * nblocks - number of blocks to be freed.
2073 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2075 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2081 /* save the current value of the root (i.e. maximum free string)
2084 oldroot = dp->tree.stree[ROOT];
2086 /* free the specified (blocks) bits */
2087 rc = dbFreeBits(bmp, dp, blkno, nblocks);
2089 /* if error or the root has not changed, done. */
2090 if (rc || (dp->tree.stree[ROOT] == oldroot))
2093 /* root changed. bubble the change up to the dmap control pages.
2094 * if the adjustment of the upper level control pages fails,
2095 * backout the deallocation.
2097 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 0, 0))) {
2098 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
2100 /* as part of backing out the deallocation, we will have
2101 * to back split the dmap tree if the deallocation caused
2102 * the freed blocks to become part of a larger binary buddy
2105 if (dp->tree.stree[word] == NOFREE)
2106 dbBackSplit((dmtree_t *) & dp->tree, word);
2108 dbAllocBits(bmp, dp, blkno, nblocks);
2116 * NAME: dbAllocBits()
2118 * FUNCTION: allocate a specified block range from a dmap.
2120 * this routine updates the dmap to reflect the working
2121 * state allocation of the specified block range. it directly
2122 * updates the bits of the working map and causes the adjustment
2123 * of the binary buddy system described by the dmap's dmtree
2124 * leaves to reflect the bits allocated. it also causes the
2125 * dmap's dmtree, as a whole, to reflect the allocated range.
2128 * bmp - pointer to bmap descriptor
2129 * dp - pointer to dmap to allocate bits from.
2130 * blkno - starting block number of the bits to be allocated.
2131 * nblocks - number of bits to be allocated.
2133 * RETURN VALUES: none
2135 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2137 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2140 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2141 dmtree_t *tp = (dmtree_t *) & dp->tree;
2145 /* pick up a pointer to the leaves of the dmap tree */
2146 leaf = dp->tree.stree + LEAFIND;
2148 /* determine the bit number and word within the dmap of the
2151 dbitno = blkno & (BPERDMAP - 1);
2152 word = dbitno >> L2DBWORD;
2154 /* block range better be within the dmap */
2155 assert(dbitno + nblocks <= BPERDMAP);
2157 /* allocate the bits of the dmap's words corresponding to the block
2158 * range. not all bits of the first and last words may be contained
2159 * within the block range. if this is the case, we'll work against
2160 * those words (i.e. partial first and/or last) on an individual basis
2161 * (a single pass), allocating the bits of interest by hand and
2162 * updating the leaf corresponding to the dmap word. a single pass
2163 * will be used for all dmap words fully contained within the
2164 * specified range. within this pass, the bits of all fully contained
2165 * dmap words will be marked as free in a single shot and the leaves
2166 * will be updated. a single leaf may describe the free space of
2167 * multiple dmap words, so we may update only a subset of the actual
2168 * leaves corresponding to the dmap words of the block range.
2170 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2171 /* determine the bit number within the word and
2172 * the number of bits within the word.
2174 wbitno = dbitno & (DBWORD - 1);
2175 nb = min(rembits, DBWORD - wbitno);
2177 /* check if only part of a word is to be allocated.
2180 /* allocate (set to 1) the appropriate bits within
2183 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
2186 /* update the leaf for this dmap word. in addition
2187 * to setting the leaf value to the binary buddy max
2188 * of the updated dmap word, dbSplit() will split
2189 * the binary system of the leaves if need be.
2191 dbSplit(tp, word, BUDMIN,
2192 dbMaxBud((u8 *) & dp->wmap[word]));
2196 /* one or more dmap words are fully contained
2197 * within the block range. determine how many
2198 * words and allocate (set to 1) the bits of these
2201 nwords = rembits >> L2DBWORD;
2202 memset(&dp->wmap[word], (int) ONES, nwords * 4);
2204 /* determine how many bits.
2206 nb = nwords << L2DBWORD;
2208 /* now update the appropriate leaves to reflect
2209 * the allocated words.
2211 for (; nwords > 0; nwords -= nw) {
2212 if (leaf[word] < BUDMIN) {
2213 jfs_error(bmp->db_ipbmap->i_sb,
2214 "leaf page corrupt\n");
2218 /* determine what the leaf value should be
2219 * updated to as the minimum of the l2 number
2220 * of bits being allocated and the l2 number
2221 * of bits currently described by this leaf.
2223 size = min_t(int, leaf[word],
2224 NLSTOL2BSZ(nwords));
2226 /* update the leaf to reflect the allocation.
2227 * in addition to setting the leaf value to
2228 * NOFREE, dbSplit() will split the binary
2229 * system of the leaves to reflect the current
2230 * allocation (size).
2232 dbSplit(tp, word, size, NOFREE);
2234 /* get the number of dmap words handled */
2235 nw = BUDSIZE(size, BUDMIN);
2241 /* update the free count for this dmap */
2242 le32_add_cpu(&dp->nfree, -nblocks);
2246 /* if this allocation group is completely free,
2247 * update the maximum allocation group number if this allocation
2248 * group is the new max.
2250 agno = blkno >> bmp->db_agl2size;
2251 if (agno > bmp->db_maxag)
2252 bmp->db_maxag = agno;
2254 /* update the free count for the allocation group and map */
2255 bmp->db_agfree[agno] -= nblocks;
2256 bmp->db_nfree -= nblocks;
2263 * NAME: dbFreeBits()
2265 * FUNCTION: free a specified block range from a dmap.
2267 * this routine updates the dmap to reflect the working
2268 * state allocation of the specified block range. it directly
2269 * updates the bits of the working map and causes the adjustment
2270 * of the binary buddy system described by the dmap's dmtree
2271 * leaves to reflect the bits freed. it also causes the dmap's
2272 * dmtree, as a whole, to reflect the deallocated range.
2275 * bmp - pointer to bmap descriptor
2276 * dp - pointer to dmap to free bits from.
2277 * blkno - starting block number of the bits to be freed.
2278 * nblocks - number of bits to be freed.
2280 * RETURN VALUES: 0 for success
2282 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2284 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2287 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2288 dmtree_t *tp = (dmtree_t *) & dp->tree;
2292 /* determine the bit number and word within the dmap of the
2295 dbitno = blkno & (BPERDMAP - 1);
2296 word = dbitno >> L2DBWORD;
2298 /* block range better be within the dmap.
2300 assert(dbitno + nblocks <= BPERDMAP);
2302 /* free the bits of the dmaps words corresponding to the block range.
2303 * not all bits of the first and last words may be contained within
2304 * the block range. if this is the case, we'll work against those
2305 * words (i.e. partial first and/or last) on an individual basis
2306 * (a single pass), freeing the bits of interest by hand and updating
2307 * the leaf corresponding to the dmap word. a single pass will be used
2308 * for all dmap words fully contained within the specified range.
2309 * within this pass, the bits of all fully contained dmap words will
2310 * be marked as free in a single shot and the leaves will be updated. a
2311 * single leaf may describe the free space of multiple dmap words,
2312 * so we may update only a subset of the actual leaves corresponding
2313 * to the dmap words of the block range.
2315 * dbJoin() is used to update leaf values and will join the binary
2316 * buddy system of the leaves if the new leaf values indicate this
2319 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2320 /* determine the bit number within the word and
2321 * the number of bits within the word.
2323 wbitno = dbitno & (DBWORD - 1);
2324 nb = min(rembits, DBWORD - wbitno);
2326 /* check if only part of a word is to be freed.
2329 /* free (zero) the appropriate bits within this
2333 cpu_to_le32(~(ONES << (DBWORD - nb)
2336 /* update the leaf for this dmap word.
2338 rc = dbJoin(tp, word,
2339 dbMaxBud((u8 *) & dp->wmap[word]));
2345 /* one or more dmap words are fully contained
2346 * within the block range. determine how many
2347 * words and free (zero) the bits of these words.
2349 nwords = rembits >> L2DBWORD;
2350 memset(&dp->wmap[word], 0, nwords * 4);
2352 /* determine how many bits.
2354 nb = nwords << L2DBWORD;
2356 /* now update the appropriate leaves to reflect
2359 for (; nwords > 0; nwords -= nw) {
2360 /* determine what the leaf value should be
2361 * updated to as the minimum of the l2 number
2362 * of bits being freed and the l2 (max) number
2363 * of bits that can be described by this leaf.
2367 (word, L2LPERDMAP, BUDMIN),
2368 NLSTOL2BSZ(nwords));
2372 rc = dbJoin(tp, word, size);
2376 /* get the number of dmap words handled.
2378 nw = BUDSIZE(size, BUDMIN);
2384 /* update the free count for this dmap.
2386 le32_add_cpu(&dp->nfree, nblocks);
2390 /* update the free count for the allocation group and
2393 agno = blkno >> bmp->db_agl2size;
2394 bmp->db_nfree += nblocks;
2395 bmp->db_agfree[agno] += nblocks;
2397 /* check if this allocation group is not completely free and
2398 * if it is currently the maximum (rightmost) allocation group.
2399 * if so, establish the new maximum allocation group number by
2400 * searching left for the first allocation group with allocation.
2402 if ((bmp->db_agfree[agno] == bmp->db_agsize && agno == bmp->db_maxag) ||
2403 (agno == bmp->db_numag - 1 &&
2404 bmp->db_agfree[agno] == (bmp-> db_mapsize & (BPERDMAP - 1)))) {
2405 while (bmp->db_maxag > 0) {
2407 if (bmp->db_agfree[bmp->db_maxag] !=
2412 /* re-establish the allocation group preference if the
2413 * current preference is right of the maximum allocation
2416 if (bmp->db_agpref > bmp->db_maxag)
2417 bmp->db_agpref = bmp->db_maxag;
2429 * FUNCTION: adjust a dmap control page at a specified level to reflect
2430 * the change in a lower level dmap or dmap control page's
2431 * maximum string of free blocks (i.e. a change in the root
2432 * of the lower level object's dmtree) due to the allocation
2433 * or deallocation of a range of blocks with a single dmap.
2435 * on entry, this routine is provided with the new value of
2436 * the lower level dmap or dmap control page root and the
2437 * starting block number of the block range whose allocation
2438 * or deallocation resulted in the root change. this range
2439 * is respresented by a single leaf of the current dmapctl
2440 * and the leaf will be updated with this value, possibly
2441 * causing a binary buddy system within the leaves to be
2442 * split or joined. the update may also cause the dmapctl's
2443 * dmtree to be updated.
2445 * if the adjustment of the dmap control page, itself, causes its
2446 * root to change, this change will be bubbled up to the next dmap
2447 * control level by a recursive call to this routine, specifying
2448 * the new root value and the next dmap control page level to
2451 * bmp - pointer to bmap descriptor
2452 * blkno - the first block of a block range within a dmap. it is
2453 * the allocation or deallocation of this block range that
2454 * requires the dmap control page to be adjusted.
2455 * newval - the new value of the lower level dmap or dmap control
2457 * alloc - 'true' if adjustment is due to an allocation.
2458 * level - current level of dmap control page (i.e. L0, L1, L2) to
2465 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2468 dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, int level)
2470 struct metapage *mp;
2474 struct dmapctl *dcp;
2477 /* get the buffer for the dmap control page for the specified
2478 * block number and control page level.
2480 lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, level);
2481 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
2484 dcp = (struct dmapctl *) mp->data;
2486 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
2487 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
2488 release_metapage(mp);
2492 /* determine the leaf number corresponding to the block and
2493 * the index within the dmap control tree.
2495 leafno = BLKTOCTLLEAF(blkno, dcp->budmin);
2496 ti = leafno + le32_to_cpu(dcp->leafidx);
2498 /* save the current leaf value and the current root level (i.e.
2499 * maximum l2 free string described by this dmapctl).
2501 oldval = dcp->stree[ti];
2502 oldroot = dcp->stree[ROOT];
2504 /* check if this is a control page update for an allocation.
2505 * if so, update the leaf to reflect the new leaf value using
2506 * dbSplit(); otherwise (deallocation), use dbJoin() to update
2507 * the leaf with the new value. in addition to updating the
2508 * leaf, dbSplit() will also split the binary buddy system of
2509 * the leaves, if required, and bubble new values within the
2510 * dmapctl tree, if required. similarly, dbJoin() will join
2511 * the binary buddy system of leaves and bubble new values up
2512 * the dmapctl tree as required by the new leaf value.
2515 /* check if we are in the middle of a binary buddy
2516 * system. this happens when we are performing the
2517 * first allocation out of an allocation group that
2518 * is part (not the first part) of a larger binary
2519 * buddy system. if we are in the middle, back split
2520 * the system prior to calling dbSplit() which assumes
2521 * that it is at the front of a binary buddy system.
2523 if (oldval == NOFREE) {
2524 rc = dbBackSplit((dmtree_t *) dcp, leafno);
2526 release_metapage(mp);
2529 oldval = dcp->stree[ti];
2531 dbSplit((dmtree_t *) dcp, leafno, dcp->budmin, newval);
2533 rc = dbJoin((dmtree_t *) dcp, leafno, newval);
2535 release_metapage(mp);
2540 /* check if the root of the current dmap control page changed due
2541 * to the update and if the current dmap control page is not at
2542 * the current top level (i.e. L0, L1, L2) of the map. if so (i.e.
2543 * root changed and this is not the top level), call this routine
2544 * again (recursion) for the next higher level of the mapping to
2545 * reflect the change in root for the current dmap control page.
2547 if (dcp->stree[ROOT] != oldroot) {
2548 /* are we below the top level of the map. if so,
2549 * bubble the root up to the next higher level.
2551 if (level < bmp->db_maxlevel) {
2552 /* bubble up the new root of this dmap control page to
2556 dbAdjCtl(bmp, blkno, dcp->stree[ROOT], alloc,
2558 /* something went wrong in bubbling up the new
2559 * root value, so backout the changes to the
2560 * current dmap control page.
2563 dbJoin((dmtree_t *) dcp, leafno,
2566 /* the dbJoin() above might have
2567 * caused a larger binary buddy system
2568 * to form and we may now be in the
2569 * middle of it. if this is the case,
2570 * back split the buddies.
2572 if (dcp->stree[ti] == NOFREE)
2573 dbBackSplit((dmtree_t *)
2575 dbSplit((dmtree_t *) dcp, leafno,
2576 dcp->budmin, oldval);
2579 /* release the buffer and return the error.
2581 release_metapage(mp);
2585 /* we're at the top level of the map. update
2586 * the bmap control page to reflect the size
2587 * of the maximum free buddy system.
2589 assert(level == bmp->db_maxlevel);
2590 if (bmp->db_maxfreebud != oldroot) {
2591 jfs_error(bmp->db_ipbmap->i_sb,
2592 "the maximum free buddy is not the old root\n");
2594 bmp->db_maxfreebud = dcp->stree[ROOT];
2598 /* write the buffer.
2609 * FUNCTION: update the leaf of a dmtree with a new value, splitting
2610 * the leaf from the binary buddy system of the dmtree's
2611 * leaves, as required.
2614 * tp - pointer to the tree containing the leaf.
2615 * leafno - the number of the leaf to be updated.
2616 * splitsz - the size the binary buddy system starting at the leaf
2617 * must be split to, specified as the log2 number of blocks.
2618 * newval - the new value for the leaf.
2620 * RETURN VALUES: none
2622 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2624 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval)
2628 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2630 /* check if the leaf needs to be split.
2632 if (leaf[leafno] > tp->dmt_budmin) {
2633 /* the split occurs by cutting the buddy system in half
2634 * at the specified leaf until we reach the specified
2635 * size. pick up the starting split size (current size
2636 * - 1 in l2) and the corresponding buddy size.
2638 cursz = leaf[leafno] - 1;
2639 budsz = BUDSIZE(cursz, tp->dmt_budmin);
2641 /* split until we reach the specified size.
2643 while (cursz >= splitsz) {
2644 /* update the buddy's leaf with its new value.
2646 dbAdjTree(tp, leafno ^ budsz, cursz);
2648 /* on to the next size and buddy.
2655 /* adjust the dmap tree to reflect the specified leaf's new
2658 dbAdjTree(tp, leafno, newval);
2663 * NAME: dbBackSplit()
2665 * FUNCTION: back split the binary buddy system of dmtree leaves
2666 * that hold a specified leaf until the specified leaf
2667 * starts its own binary buddy system.
2669 * the allocators typically perform allocations at the start
2670 * of binary buddy systems and dbSplit() is used to accomplish
2671 * any required splits. in some cases, however, allocation
2672 * may occur in the middle of a binary system and requires a
2673 * back split, with the split proceeding out from the middle of
2674 * the system (less efficient) rather than the start of the
2675 * system (more efficient). the cases in which a back split
2676 * is required are rare and are limited to the first allocation
2677 * within an allocation group which is a part (not first part)
2678 * of a larger binary buddy system and a few exception cases
2679 * in which a previous join operation must be backed out.
2682 * tp - pointer to the tree containing the leaf.
2683 * leafno - the number of the leaf to be updated.
2685 * RETURN VALUES: none
2687 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2689 static int dbBackSplit(dmtree_t * tp, int leafno)
2691 int budsz, bud, w, bsz, size;
2693 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2695 /* leaf should be part (not first part) of a binary
2698 assert(leaf[leafno] == NOFREE);
2700 /* the back split is accomplished by iteratively finding the leaf
2701 * that starts the buddy system that contains the specified leaf and
2702 * splitting that system in two. this iteration continues until
2703 * the specified leaf becomes the start of a buddy system.
2705 * determine maximum possible l2 size for the specified leaf.
2708 LITOL2BSZ(leafno, le32_to_cpu(tp->dmt_l2nleafs),
2711 /* determine the number of leaves covered by this size. this
2712 * is the buddy size that we will start with as we search for
2713 * the buddy system that contains the specified leaf.
2715 budsz = BUDSIZE(size, tp->dmt_budmin);
2719 while (leaf[leafno] == NOFREE) {
2720 /* find the leftmost buddy leaf.
2722 for (w = leafno, bsz = budsz;; bsz <<= 1,
2723 w = (w < bud) ? w : bud) {
2724 if (bsz >= le32_to_cpu(tp->dmt_nleafs)) {
2725 jfs_err("JFS: block map error in dbBackSplit");
2729 /* determine the buddy.
2733 /* check if this buddy is the start of the system.
2735 if (leaf[bud] != NOFREE) {
2736 /* split the leaf at the start of the
2739 cursz = leaf[bud] - 1;
2740 dbSplit(tp, bud, cursz, cursz);
2746 if (leaf[leafno] != size) {
2747 jfs_err("JFS: wrong leaf value in dbBackSplit");
2757 * FUNCTION: update the leaf of a dmtree with a new value, joining
2758 * the leaf with other leaves of the dmtree into a multi-leaf
2759 * binary buddy system, as required.
2762 * tp - pointer to the tree containing the leaf.
2763 * leafno - the number of the leaf to be updated.
2764 * newval - the new value for the leaf.
2766 * RETURN VALUES: none
2768 static int dbJoin(dmtree_t * tp, int leafno, int newval)
2773 /* can the new leaf value require a join with other leaves ?
2775 if (newval >= tp->dmt_budmin) {
2776 /* pickup a pointer to the leaves of the tree.
2778 leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2780 /* try to join the specified leaf into a large binary
2781 * buddy system. the join proceeds by attempting to join
2782 * the specified leafno with its buddy (leaf) at new value.
2783 * if the join occurs, we attempt to join the left leaf
2784 * of the joined buddies with its buddy at new value + 1.
2785 * we continue to join until we find a buddy that cannot be
2786 * joined (does not have a value equal to the size of the
2787 * last join) or until all leaves have been joined into a
2790 * get the buddy size (number of words covered) of
2793 budsz = BUDSIZE(newval, tp->dmt_budmin);
2797 while (budsz < le32_to_cpu(tp->dmt_nleafs)) {
2798 /* get the buddy leaf.
2800 buddy = leafno ^ budsz;
2802 /* if the leaf's new value is greater than its
2803 * buddy's value, we join no more.
2805 if (newval > leaf[buddy])
2808 /* It shouldn't be less */
2809 if (newval < leaf[buddy])
2812 /* check which (leafno or buddy) is the left buddy.
2813 * the left buddy gets to claim the blocks resulting
2814 * from the join while the right gets to claim none.
2815 * the left buddy is also eligible to participate in
2816 * a join at the next higher level while the right
2820 if (leafno < buddy) {
2821 /* leafno is the left buddy.
2823 dbAdjTree(tp, buddy, NOFREE);
2825 /* buddy is the left buddy and becomes
2828 dbAdjTree(tp, leafno, NOFREE);
2832 /* on to try the next join.
2839 /* update the leaf value.
2841 dbAdjTree(tp, leafno, newval);
2850 * FUNCTION: update a leaf of a dmtree with a new value, adjusting
2851 * the dmtree, as required, to reflect the new leaf value.
2852 * the combination of any buddies must already be done before
2856 * tp - pointer to the tree to be adjusted.
2857 * leafno - the number of the leaf to be updated.
2858 * newval - the new value for the leaf.
2860 * RETURN VALUES: none
2862 static void dbAdjTree(dmtree_t * tp, int leafno, int newval)
2867 /* pick up the index of the leaf for this leafno.
2869 lp = leafno + le32_to_cpu(tp->dmt_leafidx);
2871 /* is the current value the same as the old value ? if so,
2872 * there is nothing to do.
2874 if (tp->dmt_stree[lp] == newval)
2877 /* set the new value.
2879 tp->dmt_stree[lp] = newval;
2881 /* bubble the new value up the tree as required.
2883 for (k = 0; k < le32_to_cpu(tp->dmt_height); k++) {
2884 /* get the index of the first leaf of the 4 leaf
2885 * group containing the specified leaf (leafno).
2887 lp = ((lp - 1) & ~0x03) + 1;
2889 /* get the index of the parent of this 4 leaf group.
2893 /* determine the maximum of the 4 leaves.
2895 max = TREEMAX(&tp->dmt_stree[lp]);
2897 /* if the maximum of the 4 is the same as the
2898 * parent's value, we're done.
2900 if (tp->dmt_stree[pp] == max)
2903 /* parent gets new value.
2905 tp->dmt_stree[pp] = max;
2907 /* parent becomes leaf for next go-round.
2915 * NAME: dbFindLeaf()
2917 * FUNCTION: search a dmtree_t for sufficient free blocks, returning
2918 * the index of a leaf describing the free blocks if
2919 * sufficient free blocks are found.
2921 * the search starts at the top of the dmtree_t tree and
2922 * proceeds down the tree to the leftmost leaf with sufficient
2926 * tp - pointer to the tree to be searched.
2927 * l2nb - log2 number of free blocks to search for.
2928 * leafidx - return pointer to be set to the index of the leaf
2929 * describing at least l2nb free blocks if sufficient
2930 * free blocks are found.
2934 * -ENOSPC - insufficient free blocks.
2936 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx)
2938 int ti, n = 0, k, x = 0;
2940 /* first check the root of the tree to see if there is
2941 * sufficient free space.
2943 if (l2nb > tp->dmt_stree[ROOT])
2946 /* sufficient free space available. now search down the tree
2947 * starting at the next level for the leftmost leaf that
2948 * describes sufficient free space.
2950 for (k = le32_to_cpu(tp->dmt_height), ti = 1;
2951 k > 0; k--, ti = ((ti + n) << 2) + 1) {
2952 /* search the four nodes at this level, starting from
2955 for (x = ti, n = 0; n < 4; n++) {
2956 /* sufficient free space found. move to the next
2957 * level (or quit if this is the last level).
2959 if (l2nb <= tp->dmt_stree[x + n])
2963 /* better have found something since the higher
2964 * levels of the tree said it was here.
2969 /* set the return to the leftmost leaf describing sufficient
2972 *leafidx = x + n - le32_to_cpu(tp->dmt_leafidx);
2979 * NAME: dbFindBits()
2981 * FUNCTION: find a specified number of binary buddy free bits within a
2982 * dmap bitmap word value.
2984 * this routine searches the bitmap value for (1 << l2nb) free
2985 * bits at (1 << l2nb) alignments within the value.
2988 * word - dmap bitmap word value.
2989 * l2nb - number of free bits specified as a log2 number.
2992 * starting bit number of free bits.
2994 static int dbFindBits(u32 word, int l2nb)
2999 /* get the number of bits.
3002 assert(nb <= DBWORD);
3004 /* complement the word so we can use a mask (i.e. 0s represent
3005 * free bits) and compute the mask.
3008 mask = ONES << (DBWORD - nb);
3010 /* scan the word for nb free bits at nb alignments.
3012 for (bitno = 0; mask != 0; bitno += nb, mask >>= nb) {
3013 if ((mask & word) == mask)
3019 /* return the bit number.
3026 * NAME: dbMaxBud(u8 *cp)
3028 * FUNCTION: determine the largest binary buddy string of free
3029 * bits within 32-bits of the map.
3032 * cp - pointer to the 32-bit value.
3035 * largest binary buddy of free bits within a dmap word.
3037 static int dbMaxBud(u8 * cp)
3039 signed char tmp1, tmp2;
3041 /* check if the wmap word is all free. if so, the
3042 * free buddy size is BUDMIN.
3044 if (*((uint *) cp) == 0)
3047 /* check if the wmap word is half free. if so, the
3048 * free buddy size is BUDMIN-1.
3050 if (*((u16 *) cp) == 0 || *((u16 *) cp + 1) == 0)
3051 return (BUDMIN - 1);
3053 /* not all free or half free. determine the free buddy
3054 * size thru table lookup using quarters of the wmap word.
3056 tmp1 = max(budtab[cp[2]], budtab[cp[3]]);
3057 tmp2 = max(budtab[cp[0]], budtab[cp[1]]);
3058 return (max(tmp1, tmp2));
3063 * NAME: cnttz(uint word)
3065 * FUNCTION: determine the number of trailing zeros within a 32-bit
3069 * value - 32-bit value to be examined.
3072 * count of trailing zeros
3074 static int cnttz(u32 word)
3078 for (n = 0; n < 32; n++, word >>= 1) {
3088 * NAME: cntlz(u32 value)
3090 * FUNCTION: determine the number of leading zeros within a 32-bit
3094 * value - 32-bit value to be examined.
3097 * count of leading zeros
3099 static int cntlz(u32 value)
3103 for (n = 0; n < 32; n++, value <<= 1) {
3104 if (value & HIGHORDER)
3112 * NAME: blkstol2(s64 nb)
3114 * FUNCTION: convert a block count to its log2 value. if the block
3115 * count is not a l2 multiple, it is rounded up to the next
3116 * larger l2 multiple.
3119 * nb - number of blocks
3122 * log2 number of blocks
3124 static int blkstol2(s64 nb)
3127 s64 mask; /* meant to be signed */
3129 mask = (s64) 1 << (64 - 1);
3131 /* count the leading bits.
3133 for (l2nb = 0; l2nb < 64; l2nb++, mask >>= 1) {
3134 /* leading bit found.
3137 /* determine the l2 value.
3139 l2nb = (64 - 1) - l2nb;
3141 /* check if we need to round up.
3150 return 0; /* fix compiler warning */
3155 * NAME: dbAllocBottomUp()
3157 * FUNCTION: alloc the specified block range from the working block
3160 * the blocks will be alloc from the working map one dmap
3164 * ip - pointer to in-core inode;
3165 * blkno - starting block number to be freed.
3166 * nblocks - number of blocks to be freed.
3172 int dbAllocBottomUp(struct inode *ip, s64 blkno, s64 nblocks)
3174 struct metapage *mp;
3178 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
3179 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
3181 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
3183 /* block to be allocated better be within the mapsize. */
3184 ASSERT(nblocks <= bmp->db_mapsize - blkno);
3187 * allocate the blocks a dmap at a time.
3190 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
3191 /* release previous dmap if any */
3196 /* get the buffer for the current dmap. */
3197 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
3198 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
3200 IREAD_UNLOCK(ipbmap);
3203 dp = (struct dmap *) mp->data;
3205 /* determine the number of blocks to be allocated from
3208 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
3210 /* allocate the blocks. */
3211 if ((rc = dbAllocDmapBU(bmp, dp, blkno, nb))) {
3212 release_metapage(mp);
3213 IREAD_UNLOCK(ipbmap);
3218 /* write the last buffer. */
3221 IREAD_UNLOCK(ipbmap);
3227 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
3231 int dbitno, word, rembits, nb, nwords, wbitno, agno;
3233 struct dmaptree *tp = (struct dmaptree *) & dp->tree;
3235 /* save the current value of the root (i.e. maximum free string)
3238 oldroot = tp->stree[ROOT];
3240 /* determine the bit number and word within the dmap of the
3243 dbitno = blkno & (BPERDMAP - 1);
3244 word = dbitno >> L2DBWORD;
3246 /* block range better be within the dmap */
3247 assert(dbitno + nblocks <= BPERDMAP);
3249 /* allocate the bits of the dmap's words corresponding to the block
3250 * range. not all bits of the first and last words may be contained
3251 * within the block range. if this is the case, we'll work against
3252 * those words (i.e. partial first and/or last) on an individual basis
3253 * (a single pass), allocating the bits of interest by hand and
3254 * updating the leaf corresponding to the dmap word. a single pass
3255 * will be used for all dmap words fully contained within the
3256 * specified range. within this pass, the bits of all fully contained
3257 * dmap words will be marked as free in a single shot and the leaves
3258 * will be updated. a single leaf may describe the free space of
3259 * multiple dmap words, so we may update only a subset of the actual
3260 * leaves corresponding to the dmap words of the block range.
3262 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
3263 /* determine the bit number within the word and
3264 * the number of bits within the word.
3266 wbitno = dbitno & (DBWORD - 1);
3267 nb = min(rembits, DBWORD - wbitno);
3269 /* check if only part of a word is to be allocated.
3272 /* allocate (set to 1) the appropriate bits within
3275 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
3280 /* one or more dmap words are fully contained
3281 * within the block range. determine how many
3282 * words and allocate (set to 1) the bits of these
3285 nwords = rembits >> L2DBWORD;
3286 memset(&dp->wmap[word], (int) ONES, nwords * 4);
3288 /* determine how many bits */
3289 nb = nwords << L2DBWORD;
3294 /* update the free count for this dmap */
3295 le32_add_cpu(&dp->nfree, -nblocks);
3297 /* reconstruct summary tree */
3302 /* if this allocation group is completely free,
3303 * update the highest active allocation group number
3304 * if this allocation group is the new max.
3306 agno = blkno >> bmp->db_agl2size;
3307 if (agno > bmp->db_maxag)
3308 bmp->db_maxag = agno;
3310 /* update the free count for the allocation group and map */
3311 bmp->db_agfree[agno] -= nblocks;
3312 bmp->db_nfree -= nblocks;
3316 /* if the root has not changed, done. */
3317 if (tp->stree[ROOT] == oldroot)
3320 /* root changed. bubble the change up to the dmap control pages.
3321 * if the adjustment of the upper level control pages fails,
3322 * backout the bit allocation (thus making everything consistent).
3324 if ((rc = dbAdjCtl(bmp, blkno, tp->stree[ROOT], 1, 0)))
3325 dbFreeBits(bmp, dp, blkno, nblocks);
3332 * NAME: dbExtendFS()
3334 * FUNCTION: extend bmap from blkno for nblocks;
3335 * dbExtendFS() updates bmap ready for dbAllocBottomUp();
3339 * L1---------------------------------L1
3341 * L0---------L0---------L0 L0---------L0---------L0
3343 * d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,.,dm;
3344 * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm
3346 * <---old---><----------------------------extend----------------------->
3348 int dbExtendFS(struct inode *ipbmap, s64 blkno, s64 nblocks)
3350 struct jfs_sb_info *sbi = JFS_SBI(ipbmap->i_sb);
3351 int nbperpage = sbi->nbperpage;
3352 int i, i0 = true, j, j0 = true, k, n;
3355 struct metapage *mp, *l2mp, *l1mp = NULL, *l0mp = NULL;
3356 struct dmapctl *l2dcp, *l1dcp, *l0dcp;
3358 s8 *l0leaf, *l1leaf, *l2leaf;
3359 struct bmap *bmp = sbi->bmap;
3360 int agno, l2agsize, oldl2agsize;
3363 newsize = blkno + nblocks;
3365 jfs_info("dbExtendFS: blkno:%Ld nblocks:%Ld newsize:%Ld",
3366 (long long) blkno, (long long) nblocks, (long long) newsize);
3369 * initialize bmap control page.
3371 * all the data in bmap control page should exclude
3372 * the mkfs hidden dmap page.
3375 /* update mapsize */
3376 bmp->db_mapsize = newsize;
3377 bmp->db_maxlevel = BMAPSZTOLEV(bmp->db_mapsize);
3379 /* compute new AG size */
3380 l2agsize = dbGetL2AGSize(newsize);
3381 oldl2agsize = bmp->db_agl2size;
3383 bmp->db_agl2size = l2agsize;
3384 bmp->db_agsize = 1 << l2agsize;
3386 /* compute new number of AG */
3387 agno = bmp->db_numag;
3388 bmp->db_numag = newsize >> l2agsize;
3389 bmp->db_numag += ((u32) newsize % (u32) bmp->db_agsize) ? 1 : 0;
3392 * reconfigure db_agfree[]
3393 * from old AG configuration to new AG configuration;
3395 * coalesce contiguous k (newAGSize/oldAGSize) AGs;
3396 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn;
3397 * note: new AG size = old AG size * (2**x).
3399 if (l2agsize == oldl2agsize)
3401 k = 1 << (l2agsize - oldl2agsize);
3402 ag_rem = bmp->db_agfree[0]; /* save agfree[0] */
3403 for (i = 0, n = 0; i < agno; n++) {
3404 bmp->db_agfree[n] = 0; /* init collection point */
3406 /* coalesce contiguous k AGs; */
3407 for (j = 0; j < k && i < agno; j++, i++) {
3408 /* merge AGi to AGn */
3409 bmp->db_agfree[n] += bmp->db_agfree[i];
3412 bmp->db_agfree[0] += ag_rem; /* restore agfree[0] */
3414 for (; n < MAXAG; n++)
3415 bmp->db_agfree[n] = 0;
3418 * update highest active ag number
3421 bmp->db_maxag = bmp->db_maxag / k;
3426 * update bit maps and corresponding level control pages;
3427 * global control page db_nfree, db_agfree[agno], db_maxfreebud;
3431 p = BMAPBLKNO + nbperpage; /* L2 page */
3432 l2mp = read_metapage(ipbmap, p, PSIZE, 0);
3434 jfs_error(ipbmap->i_sb, "L2 page could not be read\n");
3437 l2dcp = (struct dmapctl *) l2mp->data;
3439 /* compute start L1 */
3440 k = blkno >> L2MAXL1SIZE;
3441 l2leaf = l2dcp->stree + CTLLEAFIND + k;
3442 p = BLKTOL1(blkno, sbi->l2nbperpage); /* L1 page */
3445 * extend each L1 in L2
3447 for (; k < LPERCTL; k++, p += nbperpage) {
3450 /* read in L1 page: (blkno & (MAXL1SIZE - 1)) */
3451 l1mp = read_metapage(ipbmap, p, PSIZE, 0);
3454 l1dcp = (struct dmapctl *) l1mp->data;
3456 /* compute start L0 */
3457 j = (blkno & (MAXL1SIZE - 1)) >> L2MAXL0SIZE;
3458 l1leaf = l1dcp->stree + CTLLEAFIND + j;
3459 p = BLKTOL0(blkno, sbi->l2nbperpage);
3462 /* assign/init L1 page */
3463 l1mp = get_metapage(ipbmap, p, PSIZE, 0);
3467 l1dcp = (struct dmapctl *) l1mp->data;
3469 /* compute start L0 */
3471 l1leaf = l1dcp->stree + CTLLEAFIND;
3472 p += nbperpage; /* 1st L0 of L1.k */
3476 * extend each L0 in L1
3478 for (; j < LPERCTL; j++) {
3481 /* read in L0 page: (blkno & (MAXL0SIZE - 1)) */
3483 l0mp = read_metapage(ipbmap, p, PSIZE, 0);
3486 l0dcp = (struct dmapctl *) l0mp->data;
3488 /* compute start dmap */
3489 i = (blkno & (MAXL0SIZE - 1)) >>
3491 l0leaf = l0dcp->stree + CTLLEAFIND + i;
3492 p = BLKTODMAP(blkno,
3496 /* assign/init L0 page */
3497 l0mp = get_metapage(ipbmap, p, PSIZE, 0);
3501 l0dcp = (struct dmapctl *) l0mp->data;
3503 /* compute start dmap */
3505 l0leaf = l0dcp->stree + CTLLEAFIND;
3506 p += nbperpage; /* 1st dmap of L0.j */
3510 * extend each dmap in L0
3512 for (; i < LPERCTL; i++) {
3514 * reconstruct the dmap page, and
3515 * initialize corresponding parent L0 leaf
3517 if ((n = blkno & (BPERDMAP - 1))) {
3518 /* read in dmap page: */
3519 mp = read_metapage(ipbmap, p,
3523 n = min(nblocks, (s64)BPERDMAP - n);
3525 /* assign/init dmap page */
3526 mp = read_metapage(ipbmap, p,
3531 n = min_t(s64, nblocks, BPERDMAP);
3534 dp = (struct dmap *) mp->data;
3535 *l0leaf = dbInitDmap(dp, blkno, n);
3538 agno = le64_to_cpu(dp->start) >> l2agsize;
3539 bmp->db_agfree[agno] += n;
3550 } /* for each dmap in a L0 */
3553 * build current L0 page from its leaves, and
3554 * initialize corresponding parent L1 leaf
3556 *l1leaf = dbInitDmapCtl(l0dcp, 0, ++i);
3557 write_metapage(l0mp);
3561 l1leaf++; /* continue for next L0 */
3563 /* more than 1 L0 ? */
3565 break; /* build L1 page */
3567 /* summarize in global bmap page */
3568 bmp->db_maxfreebud = *l1leaf;
3569 release_metapage(l1mp);
3570 release_metapage(l2mp);
3574 } /* for each L0 in a L1 */
3577 * build current L1 page from its leaves, and
3578 * initialize corresponding parent L2 leaf
3580 *l2leaf = dbInitDmapCtl(l1dcp, 1, ++j);
3581 write_metapage(l1mp);
3585 l2leaf++; /* continue for next L1 */
3587 /* more than 1 L1 ? */
3589 break; /* build L2 page */
3591 /* summarize in global bmap page */
3592 bmp->db_maxfreebud = *l2leaf;
3593 release_metapage(l2mp);
3597 } /* for each L1 in a L2 */
3599 jfs_error(ipbmap->i_sb, "function has not returned as expected\n");
3602 release_metapage(l0mp);
3604 release_metapage(l1mp);
3605 release_metapage(l2mp);
3609 * finalize bmap control page
3620 void dbFinalizeBmap(struct inode *ipbmap)
3622 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
3623 int actags, inactags, l2nl;
3624 s64 ag_rem, actfree, inactfree, avgfree;
3628 * finalize bmap control page
3632 * compute db_agpref: preferred ag to allocate from
3633 * (the leftmost ag with average free space in it);
3636 /* get the number of active ags and inactive ags */
3637 actags = bmp->db_maxag + 1;
3638 inactags = bmp->db_numag - actags;
3639 ag_rem = bmp->db_mapsize & (bmp->db_agsize - 1); /* ??? */
3641 /* determine how many blocks are in the inactive allocation
3642 * groups. in doing this, we must account for the fact that
3643 * the rightmost group might be a partial group (i.e. file
3644 * system size is not a multiple of the group size).
3646 inactfree = (inactags && ag_rem) ?
3647 ((inactags - 1) << bmp->db_agl2size) + ag_rem
3648 : inactags << bmp->db_agl2size;
3650 /* determine how many free blocks are in the active
3651 * allocation groups plus the average number of free blocks
3652 * within the active ags.
3654 actfree = bmp->db_nfree - inactfree;
3655 avgfree = (u32) actfree / (u32) actags;
3657 /* if the preferred allocation group has not average free space.
3658 * re-establish the preferred group as the leftmost
3659 * group with average free space.
3661 if (bmp->db_agfree[bmp->db_agpref] < avgfree) {
3662 for (bmp->db_agpref = 0; bmp->db_agpref < actags;
3664 if (bmp->db_agfree[bmp->db_agpref] >= avgfree)
3667 if (bmp->db_agpref >= bmp->db_numag) {
3668 jfs_error(ipbmap->i_sb,
3669 "cannot find ag with average freespace\n");
3674 * compute db_aglevel, db_agheight, db_width, db_agstart:
3675 * an ag is covered in aglevel dmapctl summary tree,
3676 * at agheight level height (from leaf) with agwidth number of nodes
3677 * each, which starts at agstart index node of the smmary tree node
3680 bmp->db_aglevel = BMAPSZTOLEV(bmp->db_agsize);
3682 bmp->db_agl2size - (L2BPERDMAP + bmp->db_aglevel * L2LPERCTL);
3683 bmp->db_agheight = l2nl >> 1;
3684 bmp->db_agwidth = 1 << (l2nl - (bmp->db_agheight << 1));
3685 for (i = 5 - bmp->db_agheight, bmp->db_agstart = 0, n = 1; i > 0;
3687 bmp->db_agstart += n;
3695 * NAME: dbInitDmap()/ujfs_idmap_page()
3697 * FUNCTION: initialize working/persistent bitmap of the dmap page
3698 * for the specified number of blocks:
3700 * at entry, the bitmaps had been initialized as free (ZEROS);
3701 * The number of blocks will only account for the actually
3702 * existing blocks. Blocks which don't actually exist in
3703 * the aggregate will be marked as allocated (ONES);
3706 * dp - pointer to page of map
3707 * nblocks - number of blocks this page
3711 static int dbInitDmap(struct dmap * dp, s64 Blkno, int nblocks)
3713 int blkno, w, b, r, nw, nb, i;
3715 /* starting block number within the dmap */
3716 blkno = Blkno & (BPERDMAP - 1);
3719 dp->nblocks = dp->nfree = cpu_to_le32(nblocks);
3720 dp->start = cpu_to_le64(Blkno);
3722 if (nblocks == BPERDMAP) {
3723 memset(&dp->wmap[0], 0, LPERDMAP * 4);
3724 memset(&dp->pmap[0], 0, LPERDMAP * 4);
3728 le32_add_cpu(&dp->nblocks, nblocks);
3729 le32_add_cpu(&dp->nfree, nblocks);
3732 /* word number containing start block number */
3733 w = blkno >> L2DBWORD;
3736 * free the bits corresponding to the block range (ZEROS):
3737 * note: not all bits of the first and last words may be contained
3738 * within the block range.
3740 for (r = nblocks; r > 0; r -= nb, blkno += nb) {
3741 /* number of bits preceding range to be freed in the word */
3742 b = blkno & (DBWORD - 1);
3743 /* number of bits to free in the word */
3744 nb = min(r, DBWORD - b);
3746 /* is partial word to be freed ? */
3748 /* free (set to 0) from the bitmap word */
3749 dp->wmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3751 dp->pmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3754 /* skip the word freed */
3757 /* free (set to 0) contiguous bitmap words */
3759 memset(&dp->wmap[w], 0, nw * 4);
3760 memset(&dp->pmap[w], 0, nw * 4);
3762 /* skip the words freed */
3763 nb = nw << L2DBWORD;
3769 * mark bits following the range to be freed (non-existing
3770 * blocks) as allocated (ONES)
3773 if (blkno == BPERDMAP)
3776 /* the first word beyond the end of existing blocks */
3777 w = blkno >> L2DBWORD;
3779 /* does nblocks fall on a 32-bit boundary ? */
3780 b = blkno & (DBWORD - 1);
3782 /* mark a partial word allocated */
3783 dp->wmap[w] = dp->pmap[w] = cpu_to_le32(ONES >> b);
3787 /* set the rest of the words in the page to allocated (ONES) */
3788 for (i = w; i < LPERDMAP; i++)
3789 dp->pmap[i] = dp->wmap[i] = cpu_to_le32(ONES);
3795 return (dbInitDmapTree(dp));
3800 * NAME: dbInitDmapTree()/ujfs_complete_dmap()
3802 * FUNCTION: initialize summary tree of the specified dmap:
3804 * at entry, bitmap of the dmap has been initialized;
3807 * dp - dmap to complete
3808 * blkno - starting block number for this dmap
3809 * treemax - will be filled in with max free for this dmap
3811 * RETURNS: max free string at the root of the tree
3813 static int dbInitDmapTree(struct dmap * dp)
3815 struct dmaptree *tp;
3819 /* init fixed info of tree */
3821 tp->nleafs = cpu_to_le32(LPERDMAP);
3822 tp->l2nleafs = cpu_to_le32(L2LPERDMAP);
3823 tp->leafidx = cpu_to_le32(LEAFIND);
3824 tp->height = cpu_to_le32(4);
3825 tp->budmin = BUDMIN;
3827 /* init each leaf from corresponding wmap word:
3828 * note: leaf is set to NOFREE(-1) if all blocks of corresponding
3829 * bitmap word are allocated.
3831 cp = tp->stree + le32_to_cpu(tp->leafidx);
3832 for (i = 0; i < LPERDMAP; i++)
3833 *cp++ = dbMaxBud((u8 *) & dp->wmap[i]);
3835 /* build the dmap's binary buddy summary tree */
3836 return (dbInitTree(tp));
3841 * NAME: dbInitTree()/ujfs_adjtree()
3843 * FUNCTION: initialize binary buddy summary tree of a dmap or dmapctl.
3845 * at entry, the leaves of the tree has been initialized
3846 * from corresponding bitmap word or root of summary tree
3847 * of the child control page;
3848 * configure binary buddy system at the leaf level, then
3849 * bubble up the values of the leaf nodes up the tree.
3852 * cp - Pointer to the root of the tree
3853 * l2leaves- Number of leaf nodes as a power of 2
3854 * l2min - Number of blocks that can be covered by a leaf
3857 * RETURNS: max free string at the root of the tree
3859 static int dbInitTree(struct dmaptree * dtp)
3861 int l2max, l2free, bsize, nextb, i;
3862 int child, parent, nparent;
3867 /* Determine the maximum free string possible for the leaves */
3868 l2max = le32_to_cpu(dtp->l2nleafs) + dtp->budmin;
3871 * configure the leaf level into binary buddy system
3873 * Try to combine buddies starting with a buddy size of 1
3874 * (i.e. two leaves). At a buddy size of 1 two buddy leaves
3875 * can be combined if both buddies have a maximum free of l2min;
3876 * the combination will result in the left-most buddy leaf having
3877 * a maximum free of l2min+1.
3878 * After processing all buddies for a given size, process buddies
3879 * at the next higher buddy size (i.e. current size * 2) and
3880 * the next maximum free (current free + 1).
3881 * This continues until the maximum possible buddy combination
3882 * yields maximum free.
3884 for (l2free = dtp->budmin, bsize = 1; l2free < l2max;
3885 l2free++, bsize = nextb) {
3886 /* get next buddy size == current buddy pair size */
3889 /* scan each adjacent buddy pair at current buddy size */
3890 for (i = 0, cp = tp + le32_to_cpu(dtp->leafidx);
3891 i < le32_to_cpu(dtp->nleafs);
3892 i += nextb, cp += nextb) {
3893 /* coalesce if both adjacent buddies are max free */
3894 if (*cp == l2free && *(cp + bsize) == l2free) {
3895 *cp = l2free + 1; /* left take right */
3896 *(cp + bsize) = -1; /* right give left */
3902 * bubble summary information of leaves up the tree.
3904 * Starting at the leaf node level, the four nodes described by
3905 * the higher level parent node are compared for a maximum free and
3906 * this maximum becomes the value of the parent node.
3907 * when all lower level nodes are processed in this fashion then
3908 * move up to the next level (parent becomes a lower level node) and
3909 * continue the process for that level.
3911 for (child = le32_to_cpu(dtp->leafidx),
3912 nparent = le32_to_cpu(dtp->nleafs) >> 2;
3913 nparent > 0; nparent >>= 2, child = parent) {
3914 /* get index of 1st node of parent level */
3915 parent = (child - 1) >> 2;
3917 /* set the value of the parent node as the maximum
3918 * of the four nodes of the current level.
3920 for (i = 0, cp = tp + child, cp1 = tp + parent;
3921 i < nparent; i++, cp += 4, cp1++)
3932 * function: initialize dmapctl page
3934 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i)
3935 { /* start leaf index not covered by range */
3938 dcp->nleafs = cpu_to_le32(LPERCTL);
3939 dcp->l2nleafs = cpu_to_le32(L2LPERCTL);
3940 dcp->leafidx = cpu_to_le32(CTLLEAFIND);
3941 dcp->height = cpu_to_le32(5);
3942 dcp->budmin = L2BPERDMAP + L2LPERCTL * level;
3945 * initialize the leaves of current level that were not covered
3946 * by the specified input block range (i.e. the leaves have no
3947 * low level dmapctl or dmap).
3949 cp = &dcp->stree[CTLLEAFIND + i];
3950 for (; i < LPERCTL; i++)
3953 /* build the dmap's binary buddy summary tree */
3954 return (dbInitTree((struct dmaptree *) dcp));
3959 * NAME: dbGetL2AGSize()/ujfs_getagl2size()
3961 * FUNCTION: Determine log2(allocation group size) from aggregate size
3964 * nblocks - Number of blocks in aggregate
3966 * RETURNS: log2(allocation group size) in aggregate blocks
3968 static int dbGetL2AGSize(s64 nblocks)
3974 if (nblocks < BPERDMAP * MAXAG)
3975 return (L2BPERDMAP);
3977 /* round up aggregate size to power of 2 */
3978 m = ((u64) 1 << (64 - 1));
3979 for (l2sz = 64; l2sz >= 0; l2sz--, m >>= 1) {
3984 sz = (s64) 1 << l2sz;
3988 /* agsize = roundupSize/max_number_of_ag */
3989 return (l2sz - L2MAXAG);
3994 * NAME: dbMapFileSizeToMapSize()
3996 * FUNCTION: compute number of blocks the block allocation map file
3997 * can cover from the map file size;
3999 * RETURNS: Number of blocks which can be covered by this block map file;
4003 * maximum number of map pages at each level including control pages
4005 #define MAXL0PAGES (1 + LPERCTL)
4006 #define MAXL1PAGES (1 + LPERCTL * MAXL0PAGES)
4009 * convert number of map pages to the zero origin top dmapctl level
4011 #define BMAPPGTOLEV(npages) \
4012 (((npages) <= 3 + MAXL0PAGES) ? 0 : \
4013 ((npages) <= 2 + MAXL1PAGES) ? 1 : 2)
4015 s64 dbMapFileSizeToMapSize(struct inode * ipbmap)
4017 struct super_block *sb = ipbmap->i_sb;
4021 int complete, factor;
4023 nblocks = ipbmap->i_size >> JFS_SBI(sb)->l2bsize;
4024 npages = nblocks >> JFS_SBI(sb)->l2nbperpage;
4025 level = BMAPPGTOLEV(npages);
4027 /* At each level, accumulate the number of dmap pages covered by
4028 * the number of full child levels below it;
4029 * repeat for the last incomplete child level.
4032 npages--; /* skip the first global control page */
4033 /* skip higher level control pages above top level covered by map */
4034 npages -= (2 - level);
4035 npages--; /* skip top level's control page */
4036 for (i = level; i >= 0; i--) {
4038 (i == 2) ? MAXL1PAGES : ((i == 1) ? MAXL0PAGES : 1);
4039 complete = (u32) npages / factor;
4040 ndmaps += complete * ((i == 2) ? LPERCTL * LPERCTL :
4041 ((i == 1) ? LPERCTL : 1));
4043 /* pages in last/incomplete child */
4044 npages = (u32) npages % factor;
4045 /* skip incomplete child's level control page */
4049 /* convert the number of dmaps into the number of blocks
4050 * which can be covered by the dmaps;
4052 nblocks = ndmaps << L2BPERDMAP;