Root/fs/jfs/jfs_dmap.c

1/*
2 * Copyright (C) International Business Machines Corp., 2000-2004
3 *
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See
12 * the GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
17 */
18
19#include <linux/fs.h>
20#include <linux/slab.h>
21#include "jfs_incore.h"
22#include "jfs_superblock.h"
23#include "jfs_dmap.h"
24#include "jfs_imap.h"
25#include "jfs_lock.h"
26#include "jfs_metapage.h"
27#include "jfs_debug.h"
28
29/*
30 * SERIALIZATION of the Block Allocation Map.
31 *
32 * the working state of the block allocation map is accessed in
33 * two directions:
34 *
35 * 1) allocation and free requests that start at the dmap
36 * level and move up through the dmap control pages (i.e.
37 * the vast majority of requests).
38 *
39 * 2) allocation requests that start at dmap control page
40 * level and work down towards the dmaps.
41 *
42 * the serialization scheme used here is as follows.
43 *
44 * requests which start at the bottom are serialized against each
45 * other through buffers and each requests holds onto its buffers
46 * as it works it way up from a single dmap to the required level
47 * of dmap control page.
48 * requests that start at the top are serialized against each other
49 * and request that start from the bottom by the multiple read/single
50 * write inode lock of the bmap inode. requests starting at the top
51 * take this lock in write mode while request starting at the bottom
52 * take the lock in read mode. a single top-down request may proceed
53 * exclusively while multiple bottoms-up requests may proceed
54 * simultaneously (under the protection of busy buffers).
55 *
56 * in addition to information found in dmaps and dmap control pages,
57 * the working state of the block allocation map also includes read/
58 * write information maintained in the bmap descriptor (i.e. total
59 * free block count, allocation group level free block counts).
60 * a single exclusive lock (BMAP_LOCK) is used to guard this information
61 * in the face of multiple-bottoms up requests.
62 * (lock ordering: IREAD_LOCK, BMAP_LOCK);
63 *
64 * accesses to the persistent state of the block allocation map (limited
65 * to the persistent bitmaps in dmaps) is guarded by (busy) buffers.
66 */
67
68#define BMAP_LOCK_INIT(bmp) mutex_init(&bmp->db_bmaplock)
69#define BMAP_LOCK(bmp) mutex_lock(&bmp->db_bmaplock)
70#define BMAP_UNLOCK(bmp) mutex_unlock(&bmp->db_bmaplock)
71
72/*
73 * forward references
74 */
75static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
76            int nblocks);
77static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval);
78static int dbBackSplit(dmtree_t * tp, int leafno);
79static int dbJoin(dmtree_t * tp, int leafno, int newval);
80static void dbAdjTree(dmtree_t * tp, int leafno, int newval);
81static int dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc,
82            int level);
83static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results);
84static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
85               int nblocks);
86static int dbAllocNear(struct bmap * bmp, struct dmap * dp, s64 blkno,
87               int nblocks,
88               int l2nb, s64 * results);
89static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
90               int nblocks);
91static int dbAllocDmapLev(struct bmap * bmp, struct dmap * dp, int nblocks,
92              int l2nb,
93              s64 * results);
94static int dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb,
95             s64 * results);
96static int dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno,
97              s64 * results);
98static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks);
99static int dbFindBits(u32 word, int l2nb);
100static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno);
101static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx);
102static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
103              int nblocks);
104static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
105              int nblocks);
106static int dbMaxBud(u8 * cp);
107s64 dbMapFileSizeToMapSize(struct inode *ipbmap);
108static int blkstol2(s64 nb);
109
110static int cntlz(u32 value);
111static int cnttz(u32 word);
112
113static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
114             int nblocks);
115static int dbInitDmap(struct dmap * dp, s64 blkno, int nblocks);
116static int dbInitDmapTree(struct dmap * dp);
117static int dbInitTree(struct dmaptree * dtp);
118static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i);
119static int dbGetL2AGSize(s64 nblocks);
120
121/*
122 * buddy table
123 *
124 * table used for determining buddy sizes within characters of
125 * dmap bitmap words. the characters themselves serve as indexes
126 * into the table, with the table elements yielding the maximum
127 * binary buddy of free bits within the character.
128 */
129static const s8 budtab[256] = {
130    3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
131    2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
132    2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
133    2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
134    2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
135    2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
136    2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
137    2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
138    2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
139    2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
140    2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
141    2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
142    2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
143    2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
144    2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
145    2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, -1
146};
147
148
149/*
150 * NAME: dbMount()
151 *
152 * FUNCTION: initializate the block allocation map.
153 *
154 * memory is allocated for the in-core bmap descriptor and
155 * the in-core descriptor is initialized from disk.
156 *
157 * PARAMETERS:
158 * ipbmap - pointer to in-core inode for the block map.
159 *
160 * RETURN VALUES:
161 * 0 - success
162 * -ENOMEM - insufficient memory
163 * -EIO - i/o error
164 */
165int dbMount(struct inode *ipbmap)
166{
167    struct bmap *bmp;
168    struct dbmap_disk *dbmp_le;
169    struct metapage *mp;
170    int i;
171
172    /*
173     * allocate/initialize the in-memory bmap descriptor
174     */
175    /* allocate memory for the in-memory bmap descriptor */
176    bmp = kmalloc(sizeof(struct bmap), GFP_KERNEL);
177    if (bmp == NULL)
178        return -ENOMEM;
179
180    /* read the on-disk bmap descriptor. */
181    mp = read_metapage(ipbmap,
182               BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
183               PSIZE, 0);
184    if (mp == NULL) {
185        kfree(bmp);
186        return -EIO;
187    }
188
189    /* copy the on-disk bmap descriptor to its in-memory version. */
190    dbmp_le = (struct dbmap_disk *) mp->data;
191    bmp->db_mapsize = le64_to_cpu(dbmp_le->dn_mapsize);
192    bmp->db_nfree = le64_to_cpu(dbmp_le->dn_nfree);
193    bmp->db_l2nbperpage = le32_to_cpu(dbmp_le->dn_l2nbperpage);
194    bmp->db_numag = le32_to_cpu(dbmp_le->dn_numag);
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    bmp->db_aglevel = le32_to_cpu(dbmp_le->dn_aglevel);
199    bmp->db_agheight = le32_to_cpu(dbmp_le->dn_agheight);
200    bmp->db_agwidth = le32_to_cpu(dbmp_le->dn_agwidth);
201    bmp->db_agstart = le32_to_cpu(dbmp_le->dn_agstart);
202    bmp->db_agl2size = le32_to_cpu(dbmp_le->dn_agl2size);
203    for (i = 0; i < MAXAG; i++)
204        bmp->db_agfree[i] = le64_to_cpu(dbmp_le->dn_agfree[i]);
205    bmp->db_agsize = le64_to_cpu(dbmp_le->dn_agsize);
206    bmp->db_maxfreebud = dbmp_le->dn_maxfreebud;
207
208    /* release the buffer. */
209    release_metapage(mp);
210
211    /* bind the bmap inode and the bmap descriptor to each other. */
212    bmp->db_ipbmap = ipbmap;
213    JFS_SBI(ipbmap->i_sb)->bmap = bmp;
214
215    memset(bmp->db_active, 0, sizeof(bmp->db_active));
216
217    /*
218     * allocate/initialize the bmap lock
219     */
220    BMAP_LOCK_INIT(bmp);
221
222    return (0);
223}
224
225
226/*
227 * NAME: dbUnmount()
228 *
229 * FUNCTION: terminate the block allocation map in preparation for
230 * file system unmount.
231 *
232 * the in-core bmap descriptor is written to disk and
233 * the memory for this descriptor is freed.
234 *
235 * PARAMETERS:
236 * ipbmap - pointer to in-core inode for the block map.
237 *
238 * RETURN VALUES:
239 * 0 - success
240 * -EIO - i/o error
241 */
242int dbUnmount(struct inode *ipbmap, int mounterror)
243{
244    struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
245
246    if (!(mounterror || isReadOnly(ipbmap)))
247        dbSync(ipbmap);
248
249    /*
250     * Invalidate the page cache buffers
251     */
252    truncate_inode_pages(ipbmap->i_mapping, 0);
253
254    /* free the memory for the in-memory bmap. */
255    kfree(bmp);
256
257    return (0);
258}
259
260/*
261 * dbSync()
262 */
263int dbSync(struct inode *ipbmap)
264{
265    struct dbmap_disk *dbmp_le;
266    struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
267    struct metapage *mp;
268    int i;
269
270    /*
271     * write bmap global control page
272     */
273    /* get the buffer for the on-disk bmap descriptor. */
274    mp = read_metapage(ipbmap,
275               BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
276               PSIZE, 0);
277    if (mp == NULL) {
278        jfs_err("dbSync: read_metapage failed!");
279        return -EIO;
280    }
281    /* copy the in-memory version of the bmap to the on-disk version */
282    dbmp_le = (struct dbmap_disk *) mp->data;
283    dbmp_le->dn_mapsize = cpu_to_le64(bmp->db_mapsize);
284    dbmp_le->dn_nfree = cpu_to_le64(bmp->db_nfree);
285    dbmp_le->dn_l2nbperpage = cpu_to_le32(bmp->db_l2nbperpage);
286    dbmp_le->dn_numag = cpu_to_le32(bmp->db_numag);
287    dbmp_le->dn_maxlevel = cpu_to_le32(bmp->db_maxlevel);
288    dbmp_le->dn_maxag = cpu_to_le32(bmp->db_maxag);
289    dbmp_le->dn_agpref = cpu_to_le32(bmp->db_agpref);
290    dbmp_le->dn_aglevel = cpu_to_le32(bmp->db_aglevel);
291    dbmp_le->dn_agheight = cpu_to_le32(bmp->db_agheight);
292    dbmp_le->dn_agwidth = cpu_to_le32(bmp->db_agwidth);
293    dbmp_le->dn_agstart = cpu_to_le32(bmp->db_agstart);
294    dbmp_le->dn_agl2size = cpu_to_le32(bmp->db_agl2size);
295    for (i = 0; i < MAXAG; i++)
296        dbmp_le->dn_agfree[i] = cpu_to_le64(bmp->db_agfree[i]);
297    dbmp_le->dn_agsize = cpu_to_le64(bmp->db_agsize);
298    dbmp_le->dn_maxfreebud = bmp->db_maxfreebud;
299
300    /* write the buffer */
301    write_metapage(mp);
302
303    /*
304     * write out dirty pages of bmap
305     */
306    filemap_write_and_wait(ipbmap->i_mapping);
307
308    diWriteSpecial(ipbmap, 0);
309
310    return (0);
311}
312
313
314/*
315 * NAME: dbFree()
316 *
317 * FUNCTION: free the specified block range from the working block
318 * allocation map.
319 *
320 * the blocks will be free from the working map one dmap
321 * at a time.
322 *
323 * PARAMETERS:
324 * ip - pointer to in-core inode;
325 * blkno - starting block number to be freed.
326 * nblocks - number of blocks to be freed.
327 *
328 * RETURN VALUES:
329 * 0 - success
330 * -EIO - i/o error
331 */
332int dbFree(struct inode *ip, s64 blkno, s64 nblocks)
333{
334    struct metapage *mp;
335    struct dmap *dp;
336    int nb, rc;
337    s64 lblkno, rem;
338    struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
339    struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
340
341    IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
342
343    /* block to be freed better be within the mapsize. */
344    if (unlikely((blkno == 0) || (blkno + nblocks > bmp->db_mapsize))) {
345        IREAD_UNLOCK(ipbmap);
346        printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
347               (unsigned long long) blkno,
348               (unsigned long long) nblocks);
349        jfs_error(ip->i_sb,
350              "dbFree: block to be freed is outside the map");
351        return -EIO;
352    }
353
354    /*
355     * free the blocks a dmap at a time.
356     */
357    mp = NULL;
358    for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
359        /* release previous dmap if any */
360        if (mp) {
361            write_metapage(mp);
362        }
363
364        /* get the buffer for the current dmap. */
365        lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
366        mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
367        if (mp == NULL) {
368            IREAD_UNLOCK(ipbmap);
369            return -EIO;
370        }
371        dp = (struct dmap *) mp->data;
372
373        /* determine the number of blocks to be freed from
374         * this dmap.
375         */
376        nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
377
378        /* free the blocks. */
379        if ((rc = dbFreeDmap(bmp, dp, blkno, nb))) {
380            jfs_error(ip->i_sb, "dbFree: error in block map\n");
381            release_metapage(mp);
382            IREAD_UNLOCK(ipbmap);
383            return (rc);
384        }
385    }
386
387    /* write the last buffer. */
388    write_metapage(mp);
389
390    IREAD_UNLOCK(ipbmap);
391
392    return (0);
393}
394
395
396/*
397 * NAME: dbUpdatePMap()
398 *
399 * FUNCTION: update the allocation state (free or allocate) of the
400 * specified block range in the persistent block allocation map.
401 *
402 * the blocks will be updated in the persistent map one
403 * dmap at a time.
404 *
405 * PARAMETERS:
406 * ipbmap - pointer to in-core inode for the block map.
407 * free - 'true' if block range is to be freed from the persistent
408 * map; 'false' if it is to be allocated.
409 * blkno - starting block number of the range.
410 * nblocks - number of contiguous blocks in the range.
411 * tblk - transaction block;
412 *
413 * RETURN VALUES:
414 * 0 - success
415 * -EIO - i/o error
416 */
417int
418dbUpdatePMap(struct inode *ipbmap,
419         int free, s64 blkno, s64 nblocks, struct tblock * tblk)
420{
421    int nblks, dbitno, wbitno, rbits;
422    int word, nbits, nwords;
423    struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
424    s64 lblkno, rem, lastlblkno;
425    u32 mask;
426    struct dmap *dp;
427    struct metapage *mp;
428    struct jfs_log *log;
429    int lsn, difft, diffp;
430    unsigned long flags;
431
432    /* the blocks better be within the mapsize. */
433    if (blkno + nblocks > bmp->db_mapsize) {
434        printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
435               (unsigned long long) blkno,
436               (unsigned long long) nblocks);
437        jfs_error(ipbmap->i_sb,
438              "dbUpdatePMap: blocks are outside the map");
439        return -EIO;
440    }
441
442    /* compute delta of transaction lsn from log syncpt */
443    lsn = tblk->lsn;
444    log = (struct jfs_log *) JFS_SBI(tblk->sb)->log;
445    logdiff(difft, lsn, log);
446
447    /*
448     * update the block state a dmap at a time.
449     */
450    mp = NULL;
451    lastlblkno = 0;
452    for (rem = nblocks; rem > 0; rem -= nblks, blkno += nblks) {
453        /* get the buffer for the current dmap. */
454        lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
455        if (lblkno != lastlblkno) {
456            if (mp) {
457                write_metapage(mp);
458            }
459
460            mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE,
461                       0);
462            if (mp == NULL)
463                return -EIO;
464            metapage_wait_for_io(mp);
465        }
466        dp = (struct dmap *) mp->data;
467
468        /* determine the bit number and word within the dmap of
469         * the starting block. also determine how many blocks
470         * are to be updated within this dmap.
471         */
472        dbitno = blkno & (BPERDMAP - 1);
473        word = dbitno >> L2DBWORD;
474        nblks = min(rem, (s64)BPERDMAP - dbitno);
475
476        /* update the bits of the dmap words. the first and last
477         * words may only have a subset of their bits updated. if
478         * this is the case, we'll work against that word (i.e.
479         * partial first and/or last) only in a single pass. a
480         * single pass will also be used to update all words that
481         * are to have all their bits updated.
482         */
483        for (rbits = nblks; rbits > 0;
484             rbits -= nbits, dbitno += nbits) {
485            /* determine the bit number within the word and
486             * the number of bits within the word.
487             */
488            wbitno = dbitno & (DBWORD - 1);
489            nbits = min(rbits, DBWORD - wbitno);
490
491            /* check if only part of the word is to be updated. */
492            if (nbits < DBWORD) {
493                /* update (free or allocate) the bits
494                 * in this word.
495                 */
496                mask =
497                    (ONES << (DBWORD - nbits) >> wbitno);
498                if (free)
499                    dp->pmap[word] &=
500                        cpu_to_le32(~mask);
501                else
502                    dp->pmap[word] |=
503                        cpu_to_le32(mask);
504
505                word += 1;
506            } else {
507                /* one or more words are to have all
508                 * their bits updated. determine how
509                 * many words and how many bits.
510                 */
511                nwords = rbits >> L2DBWORD;
512                nbits = nwords << L2DBWORD;
513
514                /* update (free or allocate) the bits
515                 * in these words.
516                 */
517                if (free)
518                    memset(&dp->pmap[word], 0,
519                           nwords * 4);
520                else
521                    memset(&dp->pmap[word], (int) ONES,
522                           nwords * 4);
523
524                word += nwords;
525            }
526        }
527
528        /*
529         * update dmap lsn
530         */
531        if (lblkno == lastlblkno)
532            continue;
533
534        lastlblkno = lblkno;
535
536        LOGSYNC_LOCK(log, flags);
537        if (mp->lsn != 0) {
538            /* inherit older/smaller lsn */
539            logdiff(diffp, mp->lsn, log);
540            if (difft < diffp) {
541                mp->lsn = lsn;
542
543                /* move bp after tblock in logsync list */
544                list_move(&mp->synclist, &tblk->synclist);
545            }
546
547            /* inherit younger/larger clsn */
548            logdiff(difft, tblk->clsn, log);
549            logdiff(diffp, mp->clsn, log);
550            if (difft > diffp)
551                mp->clsn = tblk->clsn;
552        } else {
553            mp->log = log;
554            mp->lsn = lsn;
555
556            /* insert bp after tblock in logsync list */
557            log->count++;
558            list_add(&mp->synclist, &tblk->synclist);
559
560            mp->clsn = tblk->clsn;
561        }
562        LOGSYNC_UNLOCK(log, flags);
563    }
564
565    /* write the last buffer. */
566    if (mp) {
567        write_metapage(mp);
568    }
569
570    return (0);
571}
572
573
574/*
575 * NAME: dbNextAG()
576 *
577 * FUNCTION: find the preferred allocation group for new allocations.
578 *
579 * Within the allocation groups, we maintain a preferred
580 * allocation group which consists of a group with at least
581 * average free space. It is the preferred group that we target
582 * new inode allocation towards. The tie-in between inode
583 * allocation and block allocation occurs as we allocate the
584 * first (data) block of an inode and specify the inode (block)
585 * as the allocation hint for this block.
586 *
587 * We try to avoid having more than one open file growing in
588 * an allocation group, as this will lead to fragmentation.
589 * This differs from the old OS/2 method of trying to keep
590 * empty ags around for large allocations.
591 *
592 * PARAMETERS:
593 * ipbmap - pointer to in-core inode for the block map.
594 *
595 * RETURN VALUES:
596 * the preferred allocation group number.
597 */
598int dbNextAG(struct inode *ipbmap)
599{
600    s64 avgfree;
601    int agpref;
602    s64 hwm = 0;
603    int i;
604    int next_best = -1;
605    struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
606
607    BMAP_LOCK(bmp);
608
609    /* determine the average number of free blocks within the ags. */
610    avgfree = (u32)bmp->db_nfree / bmp->db_numag;
611
612    /*
613     * if the current preferred ag does not have an active allocator
614     * and has at least average freespace, return it
615     */
616    agpref = bmp->db_agpref;
617    if ((atomic_read(&bmp->db_active[agpref]) == 0) &&
618        (bmp->db_agfree[agpref] >= avgfree))
619        goto unlock;
620
621    /* From the last preferred ag, find the next one with at least
622     * average free space.
623     */
624    for (i = 0 ; i < bmp->db_numag; i++, agpref++) {
625        if (agpref == bmp->db_numag)
626            agpref = 0;
627
628        if (atomic_read(&bmp->db_active[agpref]))
629            /* open file is currently growing in this ag */
630            continue;
631        if (bmp->db_agfree[agpref] >= avgfree) {
632            /* Return this one */
633            bmp->db_agpref = agpref;
634            goto unlock;
635        } else if (bmp->db_agfree[agpref] > hwm) {
636            /* Less than avg. freespace, but best so far */
637            hwm = bmp->db_agfree[agpref];
638            next_best = agpref;
639        }
640    }
641
642    /*
643     * If no inactive ag was found with average freespace, use the
644     * next best
645     */
646    if (next_best != -1)
647        bmp->db_agpref = next_best;
648    /* else leave db_agpref unchanged */
649unlock:
650    BMAP_UNLOCK(bmp);
651
652    /* return the preferred group.
653     */
654    return (bmp->db_agpref);
655}
656
657/*
658 * NAME: dbAlloc()
659 *
660 * FUNCTION: attempt to allocate a specified number of contiguous free
661 * blocks from the working allocation block map.
662 *
663 * the block allocation policy uses hints and a multi-step
664 * approach.
665 *
666 * for allocation requests smaller than the number of blocks
667 * per dmap, we first try to allocate the new blocks
668 * immediately following the hint. if these blocks are not
669 * available, we try to allocate blocks near the hint. if
670 * no blocks near the hint are available, we next try to
671 * allocate within the same dmap as contains the hint.
672 *
673 * if no blocks are available in the dmap or the allocation
674 * request is larger than the dmap size, we try to allocate
675 * within the same allocation group as contains the hint. if
676 * this does not succeed, we finally try to allocate anywhere
677 * within the aggregate.
678 *
679 * we also try to allocate anywhere within the aggregate for
680 * for allocation requests larger than the allocation group
681 * size or requests that specify no hint value.
682 *
683 * PARAMETERS:
684 * ip - pointer to in-core inode;
685 * hint - allocation hint.
686 * nblocks - number of contiguous blocks in the range.
687 * results - on successful return, set to the starting block number
688 * of the newly allocated contiguous range.
689 *
690 * RETURN VALUES:
691 * 0 - success
692 * -ENOSPC - insufficient disk resources
693 * -EIO - i/o error
694 */
695int dbAlloc(struct inode *ip, s64 hint, s64 nblocks, s64 * results)
696{
697    int rc, agno;
698    struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
699    struct bmap *bmp;
700    struct metapage *mp;
701    s64 lblkno, blkno;
702    struct dmap *dp;
703    int l2nb;
704    s64 mapSize;
705    int writers;
706
707    /* assert that nblocks is valid */
708    assert(nblocks > 0);
709
710    /* get the log2 number of blocks to be allocated.
711     * if the number of blocks is not a log2 multiple,
712     * it will be rounded up to the next log2 multiple.
713     */
714    l2nb = BLKSTOL2(nblocks);
715
716    bmp = JFS_SBI(ip->i_sb)->bmap;
717
718    mapSize = bmp->db_mapsize;
719
720    /* the hint should be within the map */
721    if (hint >= mapSize) {
722        jfs_error(ip->i_sb, "dbAlloc: the hint is outside the map");
723        return -EIO;
724    }
725
726    /* if the number of blocks to be allocated is greater than the
727     * allocation group size, try to allocate anywhere.
728     */
729    if (l2nb > bmp->db_agl2size) {
730        IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
731
732        rc = dbAllocAny(bmp, nblocks, l2nb, results);
733
734        goto write_unlock;
735    }
736
737    /*
738     * If no hint, let dbNextAG recommend an allocation group
739     */
740    if (hint == 0)
741        goto pref_ag;
742
743    /* we would like to allocate close to the hint. adjust the
744     * hint to the block following the hint since the allocators
745     * will start looking for free space starting at this point.
746     */
747    blkno = hint + 1;
748
749    if (blkno >= bmp->db_mapsize)
750        goto pref_ag;
751
752    agno = blkno >> bmp->db_agl2size;
753
754    /* check if blkno crosses over into a new allocation group.
755     * if so, check if we should allow allocations within this
756     * allocation group.
757     */
758    if ((blkno & (bmp->db_agsize - 1)) == 0)
759        /* check if the AG is currently being written to.
760         * if so, call dbNextAG() to find a non-busy
761         * AG with sufficient free space.
762         */
763        if (atomic_read(&bmp->db_active[agno]))
764            goto pref_ag;
765
766    /* check if the allocation request size can be satisfied from a
767     * single dmap. if so, try to allocate from the dmap containing
768     * the hint using a tiered strategy.
769     */
770    if (nblocks <= BPERDMAP) {
771        IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
772
773        /* get the buffer for the dmap containing the hint.
774         */
775        rc = -EIO;
776        lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
777        mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
778        if (mp == NULL)
779            goto read_unlock;
780
781        dp = (struct dmap *) mp->data;
782
783        /* first, try to satisfy the allocation request with the
784         * blocks beginning at the hint.
785         */
786        if ((rc = dbAllocNext(bmp, dp, blkno, (int) nblocks))
787            != -ENOSPC) {
788            if (rc == 0) {
789                *results = blkno;
790                mark_metapage_dirty(mp);
791            }
792
793            release_metapage(mp);
794            goto read_unlock;
795        }
796
797        writers = atomic_read(&bmp->db_active[agno]);
798        if ((writers > 1) ||
799            ((writers == 1) && (JFS_IP(ip)->active_ag != agno))) {
800            /*
801             * Someone else is writing in this allocation
802             * group. To avoid fragmenting, try another ag
803             */
804            release_metapage(mp);
805            IREAD_UNLOCK(ipbmap);
806            goto pref_ag;
807        }
808
809        /* next, try to satisfy the allocation request with blocks
810         * near the hint.
811         */
812        if ((rc =
813             dbAllocNear(bmp, dp, blkno, (int) nblocks, l2nb, results))
814            != -ENOSPC) {
815            if (rc == 0)
816                mark_metapage_dirty(mp);
817
818            release_metapage(mp);
819            goto read_unlock;
820        }
821
822        /* try to satisfy the allocation request with blocks within
823         * the same dmap as the hint.
824         */
825        if ((rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results))
826            != -ENOSPC) {
827            if (rc == 0)
828                mark_metapage_dirty(mp);
829
830            release_metapage(mp);
831            goto read_unlock;
832        }
833
834        release_metapage(mp);
835        IREAD_UNLOCK(ipbmap);
836    }
837
838    /* try to satisfy the allocation request with blocks within
839     * the same allocation group as the hint.
840     */
841    IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
842    if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) != -ENOSPC)
843        goto write_unlock;
844
845    IWRITE_UNLOCK(ipbmap);
846
847
848      pref_ag:
849    /*
850     * Let dbNextAG recommend a preferred allocation group
851     */
852    agno = dbNextAG(ipbmap);
853    IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
854
855    /* Try to allocate within this allocation group. if that fails, try to
856     * allocate anywhere in the map.
857     */
858    if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) == -ENOSPC)
859        rc = dbAllocAny(bmp, nblocks, l2nb, results);
860
861      write_unlock:
862    IWRITE_UNLOCK(ipbmap);
863
864    return (rc);
865
866      read_unlock:
867    IREAD_UNLOCK(ipbmap);
868
869    return (rc);
870}
871
872#ifdef _NOTYET
873/*
874 * NAME: dbAllocExact()
875 *
876 * FUNCTION: try to allocate the requested extent;
877 *
878 * PARAMETERS:
879 * ip - pointer to in-core inode;
880 * blkno - extent address;
881 * nblocks - extent length;
882 *
883 * RETURN VALUES:
884 * 0 - success
885 * -ENOSPC - insufficient disk resources
886 * -EIO - i/o error
887 */
888int dbAllocExact(struct inode *ip, s64 blkno, int nblocks)
889{
890    int rc;
891    struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
892    struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
893    struct dmap *dp;
894    s64 lblkno;
895    struct metapage *mp;
896
897    IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
898
899    /*
900     * validate extent request:
901     *
902     * note: defragfs policy:
903     * max 64 blocks will be moved.
904     * allocation request size must be satisfied from a single dmap.
905     */
906    if (nblocks <= 0 || nblocks > BPERDMAP || blkno >= bmp->db_mapsize) {
907        IREAD_UNLOCK(ipbmap);
908        return -EINVAL;
909    }
910
911    if (nblocks > ((s64) 1 << bmp->db_maxfreebud)) {
912        /* the free space is no longer available */
913        IREAD_UNLOCK(ipbmap);
914        return -ENOSPC;
915    }
916
917    /* read in the dmap covering the extent */
918    lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
919    mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
920    if (mp == NULL) {
921        IREAD_UNLOCK(ipbmap);
922        return -EIO;
923    }
924    dp = (struct dmap *) mp->data;
925
926    /* try to allocate the requested extent */
927    rc = dbAllocNext(bmp, dp, blkno, nblocks);
928
929    IREAD_UNLOCK(ipbmap);
930
931    if (rc == 0)
932        mark_metapage_dirty(mp);
933
934    release_metapage(mp);
935
936    return (rc);
937}
938#endif /* _NOTYET */
939
940/*
941 * NAME: dbReAlloc()
942 *
943 * FUNCTION: attempt to extend a current allocation by a specified
944 * number of blocks.
945 *
946 * this routine attempts to satisfy the allocation request
947 * by first trying to extend the existing allocation in
948 * place by allocating the additional blocks as the blocks
949 * immediately following the current allocation. if these
950 * blocks are not available, this routine will attempt to
951 * allocate a new set of contiguous blocks large enough
952 * to cover the existing allocation plus the additional
953 * number of blocks required.
954 *
955 * PARAMETERS:
956 * ip - pointer to in-core inode requiring allocation.
957 * blkno - starting block of the current allocation.
958 * nblocks - number of contiguous blocks within the current
959 * allocation.
960 * addnblocks - number of blocks to add to the allocation.
961 * results - on successful return, set to the starting block number
962 * of the existing allocation if the existing allocation
963 * was extended in place or to a newly allocated contiguous
964 * range if the existing allocation could not be extended
965 * in place.
966 *
967 * RETURN VALUES:
968 * 0 - success
969 * -ENOSPC - insufficient disk resources
970 * -EIO - i/o error
971 */
972int
973dbReAlloc(struct inode *ip,
974      s64 blkno, s64 nblocks, s64 addnblocks, s64 * results)
975{
976    int rc;
977
978    /* try to extend the allocation in place.
979     */
980    if ((rc = dbExtend(ip, blkno, nblocks, addnblocks)) == 0) {
981        *results = blkno;
982        return (0);
983    } else {
984        if (rc != -ENOSPC)
985            return (rc);
986    }
987
988    /* could not extend the allocation in place, so allocate a
989     * new set of blocks for the entire request (i.e. try to get
990     * a range of contiguous blocks large enough to cover the
991     * existing allocation plus the additional blocks.)
992     */
993    return (dbAlloc
994        (ip, blkno + nblocks - 1, addnblocks + nblocks, results));
995}
996
997
998/*
999 * NAME: dbExtend()
1000 *
1001 * FUNCTION: attempt to extend a current allocation by a specified
1002 * number of blocks.
1003 *
1004 * this routine attempts to satisfy the allocation request
1005 * by first trying to extend the existing allocation in
1006 * place by allocating the additional blocks as the blocks
1007 * immediately following the current allocation.
1008 *
1009 * PARAMETERS:
1010 * ip - pointer to in-core inode requiring allocation.
1011 * blkno - starting block of the current allocation.
1012 * nblocks - number of contiguous blocks within the current
1013 * allocation.
1014 * addnblocks - number of blocks to add to the allocation.
1015 *
1016 * RETURN VALUES:
1017 * 0 - success
1018 * -ENOSPC - insufficient disk resources
1019 * -EIO - i/o error
1020 */
1021static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks)
1022{
1023    struct jfs_sb_info *sbi = JFS_SBI(ip->i_sb);
1024    s64 lblkno, lastblkno, extblkno;
1025    uint rel_block;
1026    struct metapage *mp;
1027    struct dmap *dp;
1028    int rc;
1029    struct inode *ipbmap = sbi->ipbmap;
1030    struct bmap *bmp;
1031
1032    /*
1033     * We don't want a non-aligned extent to cross a page boundary
1034     */
1035    if (((rel_block = blkno & (sbi->nbperpage - 1))) &&
1036        (rel_block + nblocks + addnblocks > sbi->nbperpage))
1037        return -ENOSPC;
1038
1039    /* get the last block of the current allocation */
1040    lastblkno = blkno + nblocks - 1;
1041
1042    /* determine the block number of the block following
1043     * the existing allocation.
1044     */
1045    extblkno = lastblkno + 1;
1046
1047    IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
1048
1049    /* better be within the file system */
1050    bmp = sbi->bmap;
1051    if (lastblkno < 0 || lastblkno >= bmp->db_mapsize) {
1052        IREAD_UNLOCK(ipbmap);
1053        jfs_error(ip->i_sb,
1054              "dbExtend: the block is outside the filesystem");
1055        return -EIO;
1056    }
1057
1058    /* we'll attempt to extend the current allocation in place by
1059     * allocating the additional blocks as the blocks immediately
1060     * following the current allocation. we only try to extend the
1061     * current allocation in place if the number of additional blocks
1062     * can fit into a dmap, the last block of the current allocation
1063     * is not the last block of the file system, and the start of the
1064     * inplace extension is not on an allocation group boundary.
1065     */
1066    if (addnblocks > BPERDMAP || extblkno >= bmp->db_mapsize ||
1067        (extblkno & (bmp->db_agsize - 1)) == 0) {
1068        IREAD_UNLOCK(ipbmap);
1069        return -ENOSPC;
1070    }
1071
1072    /* get the buffer for the dmap containing the first block
1073     * of the extension.
1074     */
1075    lblkno = BLKTODMAP(extblkno, bmp->db_l2nbperpage);
1076    mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
1077    if (mp == NULL) {
1078        IREAD_UNLOCK(ipbmap);
1079        return -EIO;
1080    }
1081
1082    dp = (struct dmap *) mp->data;
1083
1084    /* try to allocate the blocks immediately following the
1085     * current allocation.
1086     */
1087    rc = dbAllocNext(bmp, dp, extblkno, (int) addnblocks);
1088
1089    IREAD_UNLOCK(ipbmap);
1090
1091    /* were we successful ? */
1092    if (rc == 0)
1093        write_metapage(mp);
1094    else
1095        /* we were not successful */
1096        release_metapage(mp);
1097
1098
1099    return (rc);
1100}
1101
1102
1103/*
1104 * NAME: dbAllocNext()
1105 *
1106 * FUNCTION: attempt to allocate the blocks of the specified block
1107 * range within a dmap.
1108 *
1109 * PARAMETERS:
1110 * bmp - pointer to bmap descriptor
1111 * dp - pointer to dmap.
1112 * blkno - starting block number of the range.
1113 * nblocks - number of contiguous free blocks of the range.
1114 *
1115 * RETURN VALUES:
1116 * 0 - success
1117 * -ENOSPC - insufficient disk resources
1118 * -EIO - i/o error
1119 *
1120 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1121 */
1122static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
1123               int nblocks)
1124{
1125    int dbitno, word, rembits, nb, nwords, wbitno, nw;
1126    int l2size;
1127    s8 *leaf;
1128    u32 mask;
1129
1130    if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1131        jfs_error(bmp->db_ipbmap->i_sb,
1132              "dbAllocNext: Corrupt dmap page");
1133        return -EIO;
1134    }
1135
1136    /* pick up a pointer to the leaves of the dmap tree.
1137     */
1138    leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1139
1140    /* determine the bit number and word within the dmap of the
1141     * starting block.
1142     */
1143    dbitno = blkno & (BPERDMAP - 1);
1144    word = dbitno >> L2DBWORD;
1145
1146    /* check if the specified block range is contained within
1147     * this dmap.
1148     */
1149    if (dbitno + nblocks > BPERDMAP)
1150        return -ENOSPC;
1151
1152    /* check if the starting leaf indicates that anything
1153     * is free.
1154     */
1155    if (leaf[word] == NOFREE)
1156        return -ENOSPC;
1157
1158    /* check the dmaps words corresponding to block range to see
1159     * if the block range is free. not all bits of the first and
1160     * last words may be contained within the block range. if this
1161     * is the case, we'll work against those words (i.e. partial first
1162     * and/or last) on an individual basis (a single pass) and examine
1163     * the actual bits to determine if they are free. a single pass
1164     * will be used for all dmap words fully contained within the
1165     * specified range. within this pass, the leaves of the dmap
1166     * tree will be examined to determine if the blocks are free. a
1167     * single leaf may describe the free space of multiple dmap
1168     * words, so we may visit only a subset of the actual leaves
1169     * corresponding to the dmap words of the block range.
1170     */
1171    for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
1172        /* determine the bit number within the word and
1173         * the number of bits within the word.
1174         */
1175        wbitno = dbitno & (DBWORD - 1);
1176        nb = min(rembits, DBWORD - wbitno);
1177
1178        /* check if only part of the word is to be examined.
1179         */
1180        if (nb < DBWORD) {
1181            /* check if the bits are free.
1182             */
1183            mask = (ONES << (DBWORD - nb) >> wbitno);
1184            if ((mask & ~le32_to_cpu(dp->wmap[word])) != mask)
1185                return -ENOSPC;
1186
1187            word += 1;
1188        } else {
1189            /* one or more dmap words are fully contained
1190             * within the block range. determine how many
1191             * words and how many bits.
1192             */
1193            nwords = rembits >> L2DBWORD;
1194            nb = nwords << L2DBWORD;
1195
1196            /* now examine the appropriate leaves to determine
1197             * if the blocks are free.
1198             */
1199            while (nwords > 0) {
1200                /* does the leaf describe any free space ?
1201                 */
1202                if (leaf[word] < BUDMIN)
1203                    return -ENOSPC;
1204
1205                /* determine the l2 number of bits provided
1206                 * by this leaf.
1207                 */
1208                l2size =
1209                    min((int)leaf[word], NLSTOL2BSZ(nwords));
1210
1211                /* determine how many words were handled.
1212                 */
1213                nw = BUDSIZE(l2size, BUDMIN);
1214
1215                nwords -= nw;
1216                word += nw;
1217            }
1218        }
1219    }
1220
1221    /* allocate the blocks.
1222     */
1223    return (dbAllocDmap(bmp, dp, blkno, nblocks));
1224}
1225
1226
1227/*
1228 * NAME: dbAllocNear()
1229 *
1230 * FUNCTION: attempt to allocate a number of contiguous free blocks near
1231 * a specified block (hint) within a dmap.
1232 *
1233 * starting with the dmap leaf that covers the hint, we'll
1234 * check the next four contiguous leaves for sufficient free
1235 * space. if sufficient free space is found, we'll allocate
1236 * the desired free space.
1237 *
1238 * PARAMETERS:
1239 * bmp - pointer to bmap descriptor
1240 * dp - pointer to dmap.
1241 * blkno - block number to allocate near.
1242 * nblocks - actual number of contiguous free blocks desired.
1243 * l2nb - log2 number of contiguous free blocks desired.
1244 * results - on successful return, set to the starting block number
1245 * of the newly allocated range.
1246 *
1247 * RETURN VALUES:
1248 * 0 - success
1249 * -ENOSPC - insufficient disk resources
1250 * -EIO - i/o error
1251 *
1252 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1253 */
1254static int
1255dbAllocNear(struct bmap * bmp,
1256        struct dmap * dp, s64 blkno, int nblocks, int l2nb, s64 * results)
1257{
1258    int word, lword, rc;
1259    s8 *leaf;
1260
1261    if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1262        jfs_error(bmp->db_ipbmap->i_sb,
1263              "dbAllocNear: Corrupt dmap page");
1264        return -EIO;
1265    }
1266
1267    leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1268
1269    /* determine the word within the dmap that holds the hint
1270     * (i.e. blkno). also, determine the last word in the dmap
1271     * that we'll include in our examination.
1272     */
1273    word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
1274    lword = min(word + 4, LPERDMAP);
1275
1276    /* examine the leaves for sufficient free space.
1277     */
1278    for (; word < lword; word++) {
1279        /* does the leaf describe sufficient free space ?
1280         */
1281        if (leaf[word] < l2nb)
1282            continue;
1283
1284        /* determine the block number within the file system
1285         * of the first block described by this dmap word.
1286         */
1287        blkno = le64_to_cpu(dp->start) + (word << L2DBWORD);
1288
1289        /* if not all bits of the dmap word are free, get the
1290         * starting bit number within the dmap word of the required
1291         * string of free bits and adjust the block number with the
1292         * value.
1293         */
1294        if (leaf[word] < BUDMIN)
1295            blkno +=
1296                dbFindBits(le32_to_cpu(dp->wmap[word]), l2nb);
1297
1298        /* allocate the blocks.
1299         */
1300        if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1301            *results = blkno;
1302
1303        return (rc);
1304    }
1305
1306    return -ENOSPC;
1307}
1308
1309
1310/*
1311 * NAME: dbAllocAG()
1312 *
1313 * FUNCTION: attempt to allocate the specified number of contiguous
1314 * free blocks within the specified allocation group.
1315 *
1316 * unless the allocation group size is equal to the number
1317 * of blocks per dmap, the dmap control pages will be used to
1318 * find the required free space, if available. we start the
1319 * search at the highest dmap control page level which
1320 * distinctly describes the allocation group's free space
1321 * (i.e. the highest level at which the allocation group's
1322 * free space is not mixed in with that of any other group).
1323 * in addition, we start the search within this level at a
1324 * height of the dmapctl dmtree at which the nodes distinctly
1325 * describe the allocation group's free space. at this height,
1326 * the allocation group's free space may be represented by 1
1327 * or two sub-trees, depending on the allocation group size.
1328 * we search the top nodes of these subtrees left to right for
1329 * sufficient free space. if sufficient free space is found,
1330 * the subtree is searched to find the leftmost leaf that
1331 * has free space. once we have made it to the leaf, we
1332 * move the search to the next lower level dmap control page
1333 * corresponding to this leaf. we continue down the dmap control
1334 * pages until we find the dmap that contains or starts the
1335 * sufficient free space and we allocate at this dmap.
1336 *
1337 * if the allocation group size is equal to the dmap size,
1338 * we'll start at the dmap corresponding to the allocation
1339 * group and attempt the allocation at this level.
1340 *
1341 * the dmap control page search is also not performed if the
1342 * allocation group is completely free and we go to the first
1343 * dmap of the allocation group to do the allocation. this is
1344 * done because the allocation group may be part (not the first
1345 * part) of a larger binary buddy system, causing the dmap
1346 * control pages to indicate no free space (NOFREE) within
1347 * the allocation group.
1348 *
1349 * PARAMETERS:
1350 * bmp - pointer to bmap descriptor
1351 * agno - allocation group number.
1352 * nblocks - actual number of contiguous free blocks desired.
1353 * l2nb - log2 number of contiguous free blocks desired.
1354 * results - on successful return, set to the starting block number
1355 * of the newly allocated range.
1356 *
1357 * RETURN VALUES:
1358 * 0 - success
1359 * -ENOSPC - insufficient disk resources
1360 * -EIO - i/o error
1361 *
1362 * note: IWRITE_LOCK(ipmap) held on entry/exit;
1363 */
1364static int
1365dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, s64 * results)
1366{
1367    struct metapage *mp;
1368    struct dmapctl *dcp;
1369    int rc, ti, i, k, m, n, agperlev;
1370    s64 blkno, lblkno;
1371    int budmin;
1372
1373    /* allocation request should not be for more than the
1374     * allocation group size.
1375     */
1376    if (l2nb > bmp->db_agl2size) {
1377        jfs_error(bmp->db_ipbmap->i_sb,
1378              "dbAllocAG: allocation request is larger than the "
1379              "allocation group size");
1380        return -EIO;
1381    }
1382
1383    /* determine the starting block number of the allocation
1384     * group.
1385     */
1386    blkno = (s64) agno << bmp->db_agl2size;
1387
1388    /* check if the allocation group size is the minimum allocation
1389     * group size or if the allocation group is completely free. if
1390     * the allocation group size is the minimum size of BPERDMAP (i.e.
1391     * 1 dmap), there is no need to search the dmap control page (below)
1392     * that fully describes the allocation group since the allocation
1393     * group is already fully described by a dmap. in this case, we
1394     * just call dbAllocCtl() to search the dmap tree and allocate the
1395     * required space if available.
1396     *
1397     * if the allocation group is completely free, dbAllocCtl() is
1398     * also called to allocate the required space. this is done for
1399     * two reasons. first, it makes no sense searching the dmap control
1400     * pages for free space when we know that free space exists. second,
1401     * the dmap control pages may indicate that the allocation group
1402     * has no free space if the allocation group is part (not the first
1403     * part) of a larger binary buddy system.
1404     */
1405    if (bmp->db_agsize == BPERDMAP
1406        || bmp->db_agfree[agno] == bmp->db_agsize) {
1407        rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1408        if ((rc == -ENOSPC) &&
1409            (bmp->db_agfree[agno] == bmp->db_agsize)) {
1410            printk(KERN_ERR "blkno = %Lx, blocks = %Lx\n",
1411                   (unsigned long long) blkno,
1412                   (unsigned long long) nblocks);
1413            jfs_error(bmp->db_ipbmap->i_sb,
1414                  "dbAllocAG: dbAllocCtl failed in free AG");
1415        }
1416        return (rc);
1417    }
1418
1419    /* the buffer for the dmap control page that fully describes the
1420     * allocation group.
1421     */
1422    lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, bmp->db_aglevel);
1423    mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1424    if (mp == NULL)
1425        return -EIO;
1426    dcp = (struct dmapctl *) mp->data;
1427    budmin = dcp->budmin;
1428
1429    if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1430        jfs_error(bmp->db_ipbmap->i_sb,
1431              "dbAllocAG: Corrupt dmapctl page");
1432        release_metapage(mp);
1433        return -EIO;
1434    }
1435
1436    /* search the subtree(s) of the dmap control page that describes
1437     * the allocation group, looking for sufficient free space. to begin,
1438     * determine how many allocation groups are represented in a dmap
1439     * control page at the control page level (i.e. L0, L1, L2) that
1440     * fully describes an allocation group. next, determine the starting
1441     * tree index of this allocation group within the control page.
1442     */
1443    agperlev =
1444        (1 << (L2LPERCTL - (bmp->db_agheight << 1))) / bmp->db_agwidth;
1445    ti = bmp->db_agstart + bmp->db_agwidth * (agno & (agperlev - 1));
1446
1447    /* dmap control page trees fan-out by 4 and a single allocation
1448     * group may be described by 1 or 2 subtrees within the ag level
1449     * dmap control page, depending upon the ag size. examine the ag's
1450     * subtrees for sufficient free space, starting with the leftmost
1451     * subtree.
1452     */
1453    for (i = 0; i < bmp->db_agwidth; i++, ti++) {
1454        /* is there sufficient free space ?
1455         */
1456        if (l2nb > dcp->stree[ti])
1457            continue;
1458
1459        /* sufficient free space found in a subtree. now search down
1460         * the subtree to find the leftmost leaf that describes this
1461         * free space.
1462         */
1463        for (k = bmp->db_agheight; k > 0; k--) {
1464            for (n = 0, m = (ti << 2) + 1; n < 4; n++) {
1465                if (l2nb <= dcp->stree[m + n]) {
1466                    ti = m + n;
1467                    break;
1468                }
1469            }
1470            if (n == 4) {
1471                jfs_error(bmp->db_ipbmap->i_sb,
1472                      "dbAllocAG: failed descending stree");
1473                release_metapage(mp);
1474                return -EIO;
1475            }
1476        }
1477
1478        /* determine the block number within the file system
1479         * that corresponds to this leaf.
1480         */
1481        if (bmp->db_aglevel == 2)
1482            blkno = 0;
1483        else if (bmp->db_aglevel == 1)
1484            blkno &= ~(MAXL1SIZE - 1);
1485        else /* bmp->db_aglevel == 0 */
1486            blkno &= ~(MAXL0SIZE - 1);
1487
1488        blkno +=
1489            ((s64) (ti - le32_to_cpu(dcp->leafidx))) << budmin;
1490
1491        /* release the buffer in preparation for going down
1492         * the next level of dmap control pages.
1493         */
1494        release_metapage(mp);
1495
1496        /* check if we need to continue to search down the lower
1497         * level dmap control pages. we need to if the number of
1498         * blocks required is less than maximum number of blocks
1499         * described at the next lower level.
1500         */
1501        if (l2nb < budmin) {
1502
1503            /* search the lower level dmap control pages to get
1504             * the starting block number of the dmap that
1505             * contains or starts off the free space.
1506             */
1507            if ((rc =
1508                 dbFindCtl(bmp, l2nb, bmp->db_aglevel - 1,
1509                       &blkno))) {
1510                if (rc == -ENOSPC) {
1511                    jfs_error(bmp->db_ipbmap->i_sb,
1512                          "dbAllocAG: control page "
1513                          "inconsistent");
1514                    return -EIO;
1515                }
1516                return (rc);
1517            }
1518        }
1519
1520        /* allocate the blocks.
1521         */
1522        rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1523        if (rc == -ENOSPC) {
1524            jfs_error(bmp->db_ipbmap->i_sb,
1525                  "dbAllocAG: unable to allocate blocks");
1526            rc = -EIO;
1527        }
1528        return (rc);
1529    }
1530
1531    /* no space in the allocation group. release the buffer and
1532     * return -ENOSPC.
1533     */
1534    release_metapage(mp);
1535
1536    return -ENOSPC;
1537}
1538
1539
1540/*
1541 * NAME: dbAllocAny()
1542 *
1543 * FUNCTION: attempt to allocate the specified number of contiguous
1544 * free blocks anywhere in the file system.
1545 *
1546 * dbAllocAny() attempts to find the sufficient free space by
1547 * searching down the dmap control pages, starting with the
1548 * highest level (i.e. L0, L1, L2) control page. if free space
1549 * large enough to satisfy the desired free space is found, the
1550 * desired free space is allocated.
1551 *
1552 * PARAMETERS:
1553 * bmp - pointer to bmap descriptor
1554 * nblocks - actual number of contiguous free blocks desired.
1555 * l2nb - log2 number of contiguous free blocks desired.
1556 * results - on successful return, set to the starting block number
1557 * of the newly allocated range.
1558 *
1559 * RETURN VALUES:
1560 * 0 - success
1561 * -ENOSPC - insufficient disk resources
1562 * -EIO - i/o error
1563 *
1564 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1565 */
1566static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results)
1567{
1568    int rc;
1569    s64 blkno = 0;
1570
1571    /* starting with the top level dmap control page, search
1572     * down the dmap control levels for sufficient free space.
1573     * if free space is found, dbFindCtl() returns the starting
1574     * block number of the dmap that contains or starts off the
1575     * range of free space.
1576     */
1577    if ((rc = dbFindCtl(bmp, l2nb, bmp->db_maxlevel, &blkno)))
1578        return (rc);
1579
1580    /* allocate the blocks.
1581     */
1582    rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1583    if (rc == -ENOSPC) {
1584        jfs_error(bmp->db_ipbmap->i_sb,
1585              "dbAllocAny: unable to allocate blocks");
1586        return -EIO;
1587    }
1588    return (rc);
1589}
1590
1591
1592/*
1593 * NAME: dbFindCtl()
1594 *
1595 * FUNCTION: starting at a specified dmap control page level and block
1596 * number, search down the dmap control levels for a range of
1597 * contiguous free blocks large enough to satisfy an allocation
1598 * request for the specified number of free blocks.
1599 *
1600 * if sufficient contiguous free blocks are found, this routine
1601 * returns the starting block number within a dmap page that
1602 * contains or starts a range of contiqious free blocks that
1603 * is sufficient in size.
1604 *
1605 * PARAMETERS:
1606 * bmp - pointer to bmap descriptor
1607 * level - starting dmap control page level.
1608 * l2nb - log2 number of contiguous free blocks desired.
1609 * *blkno - on entry, starting block number for conducting the search.
1610 * on successful return, the first block within a dmap page
1611 * that contains or starts a range of contiguous free blocks.
1612 *
1613 * RETURN VALUES:
1614 * 0 - success
1615 * -ENOSPC - insufficient disk resources
1616 * -EIO - i/o error
1617 *
1618 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1619 */
1620static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno)
1621{
1622    int rc, leafidx, lev;
1623    s64 b, lblkno;
1624    struct dmapctl *dcp;
1625    int budmin;
1626    struct metapage *mp;
1627
1628    /* starting at the specified dmap control page level and block
1629     * number, search down the dmap control levels for the starting
1630     * block number of a dmap page that contains or starts off
1631     * sufficient free blocks.
1632     */
1633    for (lev = level, b = *blkno; lev >= 0; lev--) {
1634        /* get the buffer of the dmap control page for the block
1635         * number and level (i.e. L0, L1, L2).
1636         */
1637        lblkno = BLKTOCTL(b, bmp->db_l2nbperpage, lev);
1638        mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1639        if (mp == NULL)
1640            return -EIO;
1641        dcp = (struct dmapctl *) mp->data;
1642        budmin = dcp->budmin;
1643
1644        if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1645            jfs_error(bmp->db_ipbmap->i_sb,
1646                  "dbFindCtl: Corrupt dmapctl page");
1647            release_metapage(mp);
1648            return -EIO;
1649        }
1650
1651        /* search the tree within the dmap control page for
1652         * sufficient free space. if sufficient free space is found,
1653         * dbFindLeaf() returns the index of the leaf at which
1654         * free space was found.
1655         */
1656        rc = dbFindLeaf((dmtree_t *) dcp, l2nb, &leafidx);
1657
1658        /* release the buffer.
1659         */
1660        release_metapage(mp);
1661
1662        /* space found ?
1663         */
1664        if (rc) {
1665            if (lev != level) {
1666                jfs_error(bmp->db_ipbmap->i_sb,
1667                      "dbFindCtl: dmap inconsistent");
1668                return -EIO;
1669            }
1670            return -ENOSPC;
1671        }
1672
1673        /* adjust the block number to reflect the location within
1674         * the dmap control page (i.e. the leaf) at which free
1675         * space was found.
1676         */
1677        b += (((s64) leafidx) << budmin);
1678
1679        /* we stop the search at this dmap control page level if
1680         * the number of blocks required is greater than or equal
1681         * to the maximum number of blocks described at the next
1682         * (lower) level.
1683         */
1684        if (l2nb >= budmin)
1685            break;
1686    }
1687
1688    *blkno = b;
1689    return (0);
1690}
1691
1692
1693/*
1694 * NAME: dbAllocCtl()
1695 *
1696 * FUNCTION: attempt to allocate a specified number of contiguous
1697 * blocks starting within a specific dmap.
1698 *
1699 * this routine is called by higher level routines that search
1700 * the dmap control pages above the actual dmaps for contiguous
1701 * free space. the result of successful searches by these
1702 * routines are the starting block numbers within dmaps, with
1703 * the dmaps themselves containing the desired contiguous free
1704 * space or starting a contiguous free space of desired size
1705 * that is made up of the blocks of one or more dmaps. these
1706 * calls should not fail due to insufficent resources.
1707 *
1708 * this routine is called in some cases where it is not known
1709 * whether it will fail due to insufficient resources. more
1710 * specifically, this occurs when allocating from an allocation
1711 * group whose size is equal to the number of blocks per dmap.
1712 * in this case, the dmap control pages are not examined prior
1713 * to calling this routine (to save pathlength) and the call
1714 * might fail.
1715 *
1716 * for a request size that fits within a dmap, this routine relies
1717 * upon the dmap's dmtree to find the requested contiguous free
1718 * space. for request sizes that are larger than a dmap, the
1719 * requested free space will start at the first block of the
1720 * first dmap (i.e. blkno).
1721 *
1722 * PARAMETERS:
1723 * bmp - pointer to bmap descriptor
1724 * nblocks - actual number of contiguous free blocks to allocate.
1725 * l2nb - log2 number of contiguous free blocks to allocate.
1726 * blkno - starting block number of the dmap to start the allocation
1727 * from.
1728 * results - on successful return, set to the starting block number
1729 * of the newly allocated range.
1730 *
1731 * RETURN VALUES:
1732 * 0 - success
1733 * -ENOSPC - insufficient disk resources
1734 * -EIO - i/o error
1735 *
1736 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1737 */
1738static int
1739dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, s64 * results)
1740{
1741    int rc, nb;
1742    s64 b, lblkno, n;
1743    struct metapage *mp;
1744    struct dmap *dp;
1745
1746    /* check if the allocation request is confined to a single dmap.
1747     */
1748    if (l2nb <= L2BPERDMAP) {
1749        /* get the buffer for the dmap.
1750         */
1751        lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
1752        mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1753        if (mp == NULL)
1754            return -EIO;
1755        dp = (struct dmap *) mp->data;
1756
1757        /* try to allocate the blocks.
1758         */
1759        rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results);
1760        if (rc == 0)
1761            mark_metapage_dirty(mp);
1762
1763        release_metapage(mp);
1764
1765        return (rc);
1766    }
1767
1768    /* allocation request involving multiple dmaps. it must start on
1769     * a dmap boundary.
1770     */
1771    assert((blkno & (BPERDMAP - 1)) == 0);
1772
1773    /* allocate the blocks dmap by dmap.
1774     */
1775    for (n = nblocks, b = blkno; n > 0; n -= nb, b += nb) {
1776        /* get the buffer for the dmap.
1777         */
1778        lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1779        mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1780        if (mp == NULL) {
1781            rc = -EIO;
1782            goto backout;
1783        }
1784        dp = (struct dmap *) mp->data;
1785
1786        /* the dmap better be all free.
1787         */
1788        if (dp->tree.stree[ROOT] != L2BPERDMAP) {
1789            release_metapage(mp);
1790            jfs_error(bmp->db_ipbmap->i_sb,
1791                  "dbAllocCtl: the dmap is not all free");
1792            rc = -EIO;
1793            goto backout;
1794        }
1795
1796        /* determine how many blocks to allocate from this dmap.
1797         */
1798        nb = min(n, (s64)BPERDMAP);
1799
1800        /* allocate the blocks from the dmap.
1801         */
1802        if ((rc = dbAllocDmap(bmp, dp, b, nb))) {
1803            release_metapage(mp);
1804            goto backout;
1805        }
1806
1807        /* write the buffer.
1808         */
1809        write_metapage(mp);
1810    }
1811
1812    /* set the results (starting block number) and return.
1813     */
1814    *results = blkno;
1815    return (0);
1816
1817    /* something failed in handling an allocation request involving
1818     * multiple dmaps. we'll try to clean up by backing out any
1819     * allocation that has already happened for this request. if
1820     * we fail in backing out the allocation, we'll mark the file
1821     * system to indicate that blocks have been leaked.
1822     */
1823      backout:
1824
1825    /* try to backout the allocations dmap by dmap.
1826     */
1827    for (n = nblocks - n, b = blkno; n > 0;
1828         n -= BPERDMAP, b += BPERDMAP) {
1829        /* get the buffer for this dmap.
1830         */
1831        lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1832        mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1833        if (mp == NULL) {
1834            /* could not back out. mark the file system
1835             * to indicate that we have leaked blocks.
1836             */
1837            jfs_error(bmp->db_ipbmap->i_sb,
1838                  "dbAllocCtl: I/O Error: Block Leakage.");
1839            continue;
1840        }
1841        dp = (struct dmap *) mp->data;
1842
1843        /* free the blocks is this dmap.
1844         */
1845        if (dbFreeDmap(bmp, dp, b, BPERDMAP)) {
1846            /* could not back out. mark the file system
1847             * to indicate that we have leaked blocks.
1848             */
1849            release_metapage(mp);
1850            jfs_error(bmp->db_ipbmap->i_sb,
1851                  "dbAllocCtl: Block Leakage.");
1852            continue;
1853        }
1854
1855        /* write the buffer.
1856         */
1857        write_metapage(mp);
1858    }
1859
1860    return (rc);
1861}
1862
1863
1864/*
1865 * NAME: dbAllocDmapLev()
1866 *
1867 * FUNCTION: attempt to allocate a specified number of contiguous blocks
1868 * from a specified dmap.
1869 *
1870 * this routine checks if the contiguous blocks are available.
1871 * if so, nblocks of blocks are allocated; otherwise, ENOSPC is
1872 * returned.
1873 *
1874 * PARAMETERS:
1875 * mp - pointer to bmap descriptor
1876 * dp - pointer to dmap to attempt to allocate blocks from.
1877 * l2nb - log2 number of contiguous block desired.
1878 * nblocks - actual number of contiguous block desired.
1879 * results - on successful return, set to the starting block number
1880 * of the newly allocated range.
1881 *
1882 * RETURN VALUES:
1883 * 0 - success
1884 * -ENOSPC - insufficient disk resources
1885 * -EIO - i/o error
1886 *
1887 * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or
1888 * IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit;
1889 */
1890static int
1891dbAllocDmapLev(struct bmap * bmp,
1892           struct dmap * dp, int nblocks, int l2nb, s64 * results)
1893{
1894    s64 blkno;
1895    int leafidx, rc;
1896
1897    /* can't be more than a dmaps worth of blocks */
1898    assert(l2nb <= L2BPERDMAP);
1899
1900    /* search the tree within the dmap page for sufficient
1901     * free space. if sufficient free space is found, dbFindLeaf()
1902     * returns the index of the leaf at which free space was found.
1903     */
1904    if (dbFindLeaf((dmtree_t *) & dp->tree, l2nb, &leafidx))
1905        return -ENOSPC;
1906
1907    /* determine the block number within the file system corresponding
1908     * to the leaf at which free space was found.
1909     */
1910    blkno = le64_to_cpu(dp->start) + (leafidx << L2DBWORD);
1911
1912    /* if not all bits of the dmap word are free, get the starting
1913     * bit number within the dmap word of the required string of free
1914     * bits and adjust the block number with this value.
1915     */
1916    if (dp->tree.stree[leafidx + LEAFIND] < BUDMIN)
1917        blkno += dbFindBits(le32_to_cpu(dp->wmap[leafidx]), l2nb);
1918
1919    /* allocate the blocks */
1920    if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1921        *results = blkno;
1922
1923    return (rc);
1924}
1925
1926
1927/*
1928 * NAME: dbAllocDmap()
1929 *
1930 * FUNCTION: adjust the disk allocation map to reflect the allocation
1931 * of a specified block range within a dmap.
1932 *
1933 * this routine allocates the specified blocks from the dmap
1934 * through a call to dbAllocBits(). if the allocation of the
1935 * block range causes the maximum string of free blocks within
1936 * the dmap to change (i.e. the value of the root of the dmap's
1937 * dmtree), this routine will cause this change to be reflected
1938 * up through the appropriate levels of the dmap control pages
1939 * by a call to dbAdjCtl() for the L0 dmap control page that
1940 * covers this dmap.
1941 *
1942 * PARAMETERS:
1943 * bmp - pointer to bmap descriptor
1944 * dp - pointer to dmap to allocate the block range from.
1945 * blkno - starting block number of the block to be allocated.
1946 * nblocks - number of blocks to be allocated.
1947 *
1948 * RETURN VALUES:
1949 * 0 - success
1950 * -EIO - i/o error
1951 *
1952 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
1953 */
1954static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
1955               int nblocks)
1956{
1957    s8 oldroot;
1958    int rc;
1959
1960    /* save the current value of the root (i.e. maximum free string)
1961     * of the dmap tree.
1962     */
1963    oldroot = dp->tree.stree[ROOT];
1964
1965    /* allocate the specified (blocks) bits */
1966    dbAllocBits(bmp, dp, blkno, nblocks);
1967
1968    /* if the root has not changed, done. */
1969    if (dp->tree.stree[ROOT] == oldroot)
1970        return (0);
1971
1972    /* root changed. bubble the change up to the dmap control pages.
1973     * if the adjustment of the upper level control pages fails,
1974     * backout the bit allocation (thus making everything consistent).
1975     */
1976    if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 1, 0)))
1977        dbFreeBits(bmp, dp, blkno, nblocks);
1978
1979    return (rc);
1980}
1981
1982
1983/*
1984 * NAME: dbFreeDmap()
1985 *
1986 * FUNCTION: adjust the disk allocation map to reflect the allocation
1987 * of a specified block range within a dmap.
1988 *
1989 * this routine frees the specified blocks from the dmap through
1990 * a call to dbFreeBits(). if the deallocation of the block range
1991 * causes the maximum string of free blocks within the dmap to
1992 * change (i.e. the value of the root of the dmap's dmtree), this
1993 * routine will cause this change to be reflected up through the
1994 * appropriate levels of the dmap control pages by a call to
1995 * dbAdjCtl() for the L0 dmap control page that covers this dmap.
1996 *
1997 * PARAMETERS:
1998 * bmp - pointer to bmap descriptor
1999 * dp - pointer to dmap to free the block range from.
2000 * blkno - starting block number of the block to be freed.
2001 * nblocks - number of blocks to be freed.
2002 *
2003 * RETURN VALUES:
2004 * 0 - success
2005 * -EIO - i/o error
2006 *
2007 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2008 */
2009static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2010              int nblocks)
2011{
2012    s8 oldroot;
2013    int rc = 0, word;
2014
2015    /* save the current value of the root (i.e. maximum free string)
2016     * of the dmap tree.
2017     */
2018    oldroot = dp->tree.stree[ROOT];
2019
2020    /* free the specified (blocks) bits */
2021    rc = dbFreeBits(bmp, dp, blkno, nblocks);
2022
2023    /* if error or the root has not changed, done. */
2024    if (rc || (dp->tree.stree[ROOT] == oldroot))
2025        return (rc);
2026
2027    /* root changed. bubble the change up to the dmap control pages.
2028     * if the adjustment of the upper level control pages fails,
2029     * backout the deallocation.
2030     */
2031    if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 0, 0))) {
2032        word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
2033
2034        /* as part of backing out the deallocation, we will have
2035         * to back split the dmap tree if the deallocation caused
2036         * the freed blocks to become part of a larger binary buddy
2037         * system.
2038         */
2039        if (dp->tree.stree[word] == NOFREE)
2040            dbBackSplit((dmtree_t *) & dp->tree, word);
2041
2042        dbAllocBits(bmp, dp, blkno, nblocks);
2043    }
2044
2045    return (rc);
2046}
2047
2048
2049/*
2050 * NAME: dbAllocBits()
2051 *
2052 * FUNCTION: allocate a specified block range from a dmap.
2053 *
2054 * this routine updates the dmap to reflect the working
2055 * state allocation of the specified block range. it directly
2056 * updates the bits of the working map and causes the adjustment
2057 * of the binary buddy system described by the dmap's dmtree
2058 * leaves to reflect the bits allocated. it also causes the
2059 * dmap's dmtree, as a whole, to reflect the allocated range.
2060 *
2061 * PARAMETERS:
2062 * bmp - pointer to bmap descriptor
2063 * dp - pointer to dmap to allocate bits from.
2064 * blkno - starting block number of the bits to be allocated.
2065 * nblocks - number of bits to be allocated.
2066 *
2067 * RETURN VALUES: none
2068 *
2069 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2070 */
2071static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2072            int nblocks)
2073{
2074    int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2075    dmtree_t *tp = (dmtree_t *) & dp->tree;
2076    int size;
2077    s8 *leaf;
2078
2079    /* pick up a pointer to the leaves of the dmap tree */
2080    leaf = dp->tree.stree + LEAFIND;
2081
2082    /* determine the bit number and word within the dmap of the
2083     * starting block.
2084     */
2085    dbitno = blkno & (BPERDMAP - 1);
2086    word = dbitno >> L2DBWORD;
2087
2088    /* block range better be within the dmap */
2089    assert(dbitno + nblocks <= BPERDMAP);
2090
2091    /* allocate the bits of the dmap's words corresponding to the block
2092     * range. not all bits of the first and last words may be contained
2093     * within the block range. if this is the case, we'll work against
2094     * those words (i.e. partial first and/or last) on an individual basis
2095     * (a single pass), allocating the bits of interest by hand and
2096     * updating the leaf corresponding to the dmap word. a single pass
2097     * will be used for all dmap words fully contained within the
2098     * specified range. within this pass, the bits of all fully contained
2099     * dmap words will be marked as free in a single shot and the leaves
2100     * will be updated. a single leaf may describe the free space of
2101     * multiple dmap words, so we may update only a subset of the actual
2102     * leaves corresponding to the dmap words of the block range.
2103     */
2104    for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2105        /* determine the bit number within the word and
2106         * the number of bits within the word.
2107         */
2108        wbitno = dbitno & (DBWORD - 1);
2109        nb = min(rembits, DBWORD - wbitno);
2110
2111        /* check if only part of a word is to be allocated.
2112         */
2113        if (nb < DBWORD) {
2114            /* allocate (set to 1) the appropriate bits within
2115             * this dmap word.
2116             */
2117            dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
2118                              >> wbitno);
2119
2120            /* update the leaf for this dmap word. in addition
2121             * to setting the leaf value to the binary buddy max
2122             * of the updated dmap word, dbSplit() will split
2123             * the binary system of the leaves if need be.
2124             */
2125            dbSplit(tp, word, BUDMIN,
2126                dbMaxBud((u8 *) & dp->wmap[word]));
2127
2128            word += 1;
2129        } else {
2130            /* one or more dmap words are fully contained
2131             * within the block range. determine how many
2132             * words and allocate (set to 1) the bits of these
2133             * words.
2134             */
2135            nwords = rembits >> L2DBWORD;
2136            memset(&dp->wmap[word], (int) ONES, nwords * 4);
2137
2138            /* determine how many bits.
2139             */
2140            nb = nwords << L2DBWORD;
2141
2142            /* now update the appropriate leaves to reflect
2143             * the allocated words.
2144             */
2145            for (; nwords > 0; nwords -= nw) {
2146                if (leaf[word] < BUDMIN) {
2147                    jfs_error(bmp->db_ipbmap->i_sb,
2148                          "dbAllocBits: leaf page "
2149                          "corrupt");
2150                    break;
2151                }
2152
2153                /* determine what the leaf value should be
2154                 * updated to as the minimum of the l2 number
2155                 * of bits being allocated and the l2 number
2156                 * of bits currently described by this leaf.
2157                 */
2158                size = min((int)leaf[word], NLSTOL2BSZ(nwords));
2159
2160                /* update the leaf to reflect the allocation.
2161                 * in addition to setting the leaf value to
2162                 * NOFREE, dbSplit() will split the binary
2163                 * system of the leaves to reflect the current
2164                 * allocation (size).
2165                 */
2166                dbSplit(tp, word, size, NOFREE);
2167
2168                /* get the number of dmap words handled */
2169                nw = BUDSIZE(size, BUDMIN);
2170                word += nw;
2171            }
2172        }
2173    }
2174
2175    /* update the free count for this dmap */
2176    le32_add_cpu(&dp->nfree, -nblocks);
2177
2178    BMAP_LOCK(bmp);
2179
2180    /* if this allocation group is completely free,
2181     * update the maximum allocation group number if this allocation
2182     * group is the new max.
2183     */
2184    agno = blkno >> bmp->db_agl2size;
2185    if (agno > bmp->db_maxag)
2186        bmp->db_maxag = agno;
2187
2188    /* update the free count for the allocation group and map */
2189    bmp->db_agfree[agno] -= nblocks;
2190    bmp->db_nfree -= nblocks;
2191
2192    BMAP_UNLOCK(bmp);
2193}
2194
2195
2196/*
2197 * NAME: dbFreeBits()
2198 *
2199 * FUNCTION: free a specified block range from a dmap.
2200 *
2201 * this routine updates the dmap to reflect the working
2202 * state allocation of the specified block range. it directly
2203 * updates the bits of the working map and causes the adjustment
2204 * of the binary buddy system described by the dmap's dmtree
2205 * leaves to reflect the bits freed. it also causes the dmap's
2206 * dmtree, as a whole, to reflect the deallocated range.
2207 *
2208 * PARAMETERS:
2209 * bmp - pointer to bmap descriptor
2210 * dp - pointer to dmap to free bits from.
2211 * blkno - starting block number of the bits to be freed.
2212 * nblocks - number of bits to be freed.
2213 *
2214 * RETURN VALUES: 0 for success
2215 *
2216 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2217 */
2218static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2219               int nblocks)
2220{
2221    int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2222    dmtree_t *tp = (dmtree_t *) & dp->tree;
2223    int rc = 0;
2224    int size;
2225
2226    /* determine the bit number and word within the dmap of the
2227     * starting block.
2228     */
2229    dbitno = blkno & (BPERDMAP - 1);
2230    word = dbitno >> L2DBWORD;
2231
2232    /* block range better be within the dmap.
2233     */
2234    assert(dbitno + nblocks <= BPERDMAP);
2235
2236    /* free the bits of the dmaps words corresponding to the block range.
2237     * not all bits of the first and last words may be contained within
2238     * the block range. if this is the case, we'll work against those
2239     * words (i.e. partial first and/or last) on an individual basis
2240     * (a single pass), freeing the bits of interest by hand and updating
2241     * the leaf corresponding to the dmap word. a single pass will be used
2242     * for all dmap words fully contained within the specified range.
2243     * within this pass, the bits of all fully contained dmap words will
2244     * be marked as free in a single shot and the leaves will be updated. a
2245     * single leaf may describe the free space of multiple dmap words,
2246     * so we may update only a subset of the actual leaves corresponding
2247     * to the dmap words of the block range.
2248     *
2249     * dbJoin() is used to update leaf values and will join the binary
2250     * buddy system of the leaves if the new leaf values indicate this
2251     * should be done.
2252     */
2253    for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2254        /* determine the bit number within the word and
2255         * the number of bits within the word.
2256         */
2257        wbitno = dbitno & (DBWORD - 1);
2258        nb = min(rembits, DBWORD - wbitno);
2259
2260        /* check if only part of a word is to be freed.
2261         */
2262        if (nb < DBWORD) {
2263            /* free (zero) the appropriate bits within this
2264             * dmap word.
2265             */
2266            dp->wmap[word] &=
2267                cpu_to_le32(~(ONES << (DBWORD - nb)
2268                      >> wbitno));
2269
2270            /* update the leaf for this dmap word.
2271             */
2272            rc = dbJoin(tp, word,
2273                    dbMaxBud((u8 *) & dp->wmap[word]));
2274            if (rc)
2275                return rc;
2276
2277            word += 1;
2278        } else {
2279            /* one or more dmap words are fully contained
2280             * within the block range. determine how many
2281             * words and free (zero) the bits of these words.
2282             */
2283            nwords = rembits >> L2DBWORD;
2284            memset(&dp->wmap[word], 0, nwords * 4);
2285
2286            /* determine how many bits.
2287             */
2288            nb = nwords << L2DBWORD;
2289
2290            /* now update the appropriate leaves to reflect
2291             * the freed words.
2292             */
2293            for (; nwords > 0; nwords -= nw) {
2294                /* determine what the leaf value should be
2295                 * updated to as the minimum of the l2 number
2296                 * of bits being freed and the l2 (max) number
2297                 * of bits that can be described by this leaf.
2298                 */
2299                size =
2300                    min(LITOL2BSZ
2301                    (word, L2LPERDMAP, BUDMIN),
2302                    NLSTOL2BSZ(nwords));
2303
2304                /* update the leaf.
2305                 */
2306                rc = dbJoin(tp, word, size);
2307                if (rc)
2308                    return rc;
2309
2310                /* get the number of dmap words handled.
2311                 */
2312                nw = BUDSIZE(size, BUDMIN);
2313                word += nw;
2314            }
2315        }
2316    }
2317
2318    /* update the free count for this dmap.
2319     */
2320    le32_add_cpu(&dp->nfree, nblocks);
2321
2322    BMAP_LOCK(bmp);
2323
2324    /* update the free count for the allocation group and
2325     * map.
2326     */
2327    agno = blkno >> bmp->db_agl2size;
2328    bmp->db_nfree += nblocks;
2329    bmp->db_agfree[agno] += nblocks;
2330
2331    /* check if this allocation group is not completely free and
2332     * if it is currently the maximum (rightmost) allocation group.
2333     * if so, establish the new maximum allocation group number by
2334     * searching left for the first allocation group with allocation.
2335     */
2336    if ((bmp->db_agfree[agno] == bmp->db_agsize && agno == bmp->db_maxag) ||
2337        (agno == bmp->db_numag - 1 &&
2338         bmp->db_agfree[agno] == (bmp-> db_mapsize & (BPERDMAP - 1)))) {
2339        while (bmp->db_maxag > 0) {
2340            bmp->db_maxag -= 1;
2341            if (bmp->db_agfree[bmp->db_maxag] !=
2342                bmp->db_agsize)
2343                break;
2344        }
2345
2346        /* re-establish the allocation group preference if the
2347         * current preference is right of the maximum allocation
2348         * group.
2349         */
2350        if (bmp->db_agpref > bmp->db_maxag)
2351            bmp->db_agpref = bmp->db_maxag;
2352    }
2353
2354    BMAP_UNLOCK(bmp);
2355
2356    return 0;
2357}
2358
2359
2360/*
2361 * NAME: dbAdjCtl()
2362 *
2363 * FUNCTION: adjust a dmap control page at a specified level to reflect
2364 * the change in a lower level dmap or dmap control page's
2365 * maximum string of free blocks (i.e. a change in the root
2366 * of the lower level object's dmtree) due to the allocation
2367 * or deallocation of a range of blocks with a single dmap.
2368 *
2369 * on entry, this routine is provided with the new value of
2370 * the lower level dmap or dmap control page root and the
2371 * starting block number of the block range whose allocation
2372 * or deallocation resulted in the root change. this range
2373 * is respresented by a single leaf of the current dmapctl
2374 * and the leaf will be updated with this value, possibly
2375 * causing a binary buddy system within the leaves to be
2376 * split or joined. the update may also cause the dmapctl's
2377 * dmtree to be updated.
2378 *
2379 * if the adjustment of the dmap control page, itself, causes its
2380 * root to change, this change will be bubbled up to the next dmap
2381 * control level by a recursive call to this routine, specifying
2382 * the new root value and the next dmap control page level to
2383 * be adjusted.
2384 * PARAMETERS:
2385 * bmp - pointer to bmap descriptor
2386 * blkno - the first block of a block range within a dmap. it is
2387 * the allocation or deallocation of this block range that
2388 * requires the dmap control page to be adjusted.
2389 * newval - the new value of the lower level dmap or dmap control
2390 * page root.
2391 * alloc - 'true' if adjustment is due to an allocation.
2392 * level - current level of dmap control page (i.e. L0, L1, L2) to
2393 * be adjusted.
2394 *
2395 * RETURN VALUES:
2396 * 0 - success
2397 * -EIO - i/o error
2398 *
2399 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2400 */
2401static int
2402dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, int level)
2403{
2404    struct metapage *mp;
2405    s8 oldroot;
2406    int oldval;
2407    s64 lblkno;
2408    struct dmapctl *dcp;
2409    int rc, leafno, ti;
2410
2411    /* get the buffer for the dmap control page for the specified
2412     * block number and control page level.
2413     */
2414    lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, level);
2415    mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
2416    if (mp == NULL)
2417        return -EIO;
2418    dcp = (struct dmapctl *) mp->data;
2419
2420    if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
2421        jfs_error(bmp->db_ipbmap->i_sb,
2422              "dbAdjCtl: Corrupt dmapctl page");
2423        release_metapage(mp);
2424        return -EIO;
2425    }
2426
2427    /* determine the leaf number corresponding to the block and
2428     * the index within the dmap control tree.
2429     */
2430    leafno = BLKTOCTLLEAF(blkno, dcp->budmin);
2431    ti = leafno + le32_to_cpu(dcp->leafidx);
2432
2433    /* save the current leaf value and the current root level (i.e.
2434     * maximum l2 free string described by this dmapctl).
2435     */
2436    oldval = dcp->stree[ti];
2437    oldroot = dcp->stree[ROOT];
2438
2439    /* check if this is a control page update for an allocation.
2440     * if so, update the leaf to reflect the new leaf value using
2441     * dbSplit(); otherwise (deallocation), use dbJoin() to update
2442     * the leaf with the new value. in addition to updating the
2443     * leaf, dbSplit() will also split the binary buddy system of
2444     * the leaves, if required, and bubble new values within the
2445     * dmapctl tree, if required. similarly, dbJoin() will join
2446     * the binary buddy system of leaves and bubble new values up
2447     * the dmapctl tree as required by the new leaf value.
2448     */
2449    if (alloc) {
2450        /* check if we are in the middle of a binary buddy
2451         * system. this happens when we are performing the
2452         * first allocation out of an allocation group that
2453         * is part (not the first part) of a larger binary
2454         * buddy system. if we are in the middle, back split
2455         * the system prior to calling dbSplit() which assumes
2456         * that it is at the front of a binary buddy system.
2457         */
2458        if (oldval == NOFREE) {
2459            rc = dbBackSplit((dmtree_t *) dcp, leafno);
2460            if (rc)
2461                return rc;
2462            oldval = dcp->stree[ti];
2463        }
2464        dbSplit((dmtree_t *) dcp, leafno, dcp->budmin, newval);
2465    } else {
2466        rc = dbJoin((dmtree_t *) dcp, leafno, newval);
2467        if (rc)
2468            return rc;
2469    }
2470
2471    /* check if the root of the current dmap control page changed due
2472     * to the update and if the current dmap control page is not at
2473     * the current top level (i.e. L0, L1, L2) of the map. if so (i.e.
2474     * root changed and this is not the top level), call this routine
2475     * again (recursion) for the next higher level of the mapping to
2476     * reflect the change in root for the current dmap control page.
2477     */
2478    if (dcp->stree[ROOT] != oldroot) {
2479        /* are we below the top level of the map. if so,
2480         * bubble the root up to the next higher level.
2481         */
2482        if (level < bmp->db_maxlevel) {
2483            /* bubble up the new root of this dmap control page to
2484             * the next level.
2485             */
2486            if ((rc =
2487                 dbAdjCtl(bmp, blkno, dcp->stree[ROOT], alloc,
2488                      level + 1))) {
2489                /* something went wrong in bubbling up the new
2490                 * root value, so backout the changes to the
2491                 * current dmap control page.
2492                 */
2493                if (alloc) {
2494                    dbJoin((dmtree_t *) dcp, leafno,
2495                           oldval);
2496                } else {
2497                    /* the dbJoin() above might have
2498                     * caused a larger binary buddy system
2499                     * to form and we may now be in the
2500                     * middle of it. if this is the case,
2501                     * back split the buddies.
2502                     */
2503                    if (dcp->stree[ti] == NOFREE)
2504                        dbBackSplit((dmtree_t *)
2505                                dcp, leafno);
2506                    dbSplit((dmtree_t *) dcp, leafno,
2507                        dcp->budmin, oldval);
2508                }
2509
2510                /* release the buffer and return the error.
2511                 */
2512                release_metapage(mp);
2513                return (rc);
2514            }
2515        } else {
2516            /* we're at the top level of the map. update
2517             * the bmap control page to reflect the size
2518             * of the maximum free buddy system.
2519             */
2520            assert(level == bmp->db_maxlevel);
2521            if (bmp->db_maxfreebud != oldroot) {
2522                jfs_error(bmp->db_ipbmap->i_sb,
2523                      "dbAdjCtl: the maximum free buddy is "
2524                      "not the old root");
2525            }
2526            bmp->db_maxfreebud = dcp->stree[ROOT];
2527        }
2528    }
2529
2530    /* write the buffer.
2531     */
2532    write_metapage(mp);
2533
2534    return (0);
2535}
2536
2537
2538/*
2539 * NAME: dbSplit()
2540 *
2541 * FUNCTION: update the leaf of a dmtree with a new value, splitting
2542 * the leaf from the binary buddy system of the dmtree's
2543 * leaves, as required.
2544 *
2545 * PARAMETERS:
2546 * tp - pointer to the tree containing the leaf.
2547 * leafno - the number of the leaf to be updated.
2548 * splitsz - the size the binary buddy system starting at the leaf
2549 * must be split to, specified as the log2 number of blocks.
2550 * newval - the new value for the leaf.
2551 *
2552 * RETURN VALUES: none
2553 *
2554 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2555 */
2556static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval)
2557{
2558    int budsz;
2559    int cursz;
2560    s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2561
2562    /* check if the leaf needs to be split.
2563     */
2564    if (leaf[leafno] > tp->dmt_budmin) {
2565        /* the split occurs by cutting the buddy system in half
2566         * at the specified leaf until we reach the specified
2567         * size. pick up the starting split size (current size
2568         * - 1 in l2) and the corresponding buddy size.
2569         */
2570        cursz = leaf[leafno] - 1;
2571        budsz = BUDSIZE(cursz, tp->dmt_budmin);
2572
2573        /* split until we reach the specified size.
2574         */
2575        while (cursz >= splitsz) {
2576            /* update the buddy's leaf with its new value.
2577             */
2578            dbAdjTree(tp, leafno ^ budsz, cursz);
2579
2580            /* on to the next size and buddy.
2581             */
2582            cursz -= 1;
2583            budsz >>= 1;
2584        }
2585    }
2586
2587    /* adjust the dmap tree to reflect the specified leaf's new
2588     * value.
2589     */
2590    dbAdjTree(tp, leafno, newval);
2591}
2592
2593
2594/*
2595 * NAME: dbBackSplit()
2596 *
2597 * FUNCTION: back split the binary buddy system of dmtree leaves
2598 * that hold a specified leaf until the specified leaf
2599 * starts its own binary buddy system.
2600 *
2601 * the allocators typically perform allocations at the start
2602 * of binary buddy systems and dbSplit() is used to accomplish
2603 * any required splits. in some cases, however, allocation
2604 * may occur in the middle of a binary system and requires a
2605 * back split, with the split proceeding out from the middle of
2606 * the system (less efficient) rather than the start of the
2607 * system (more efficient). the cases in which a back split
2608 * is required are rare and are limited to the first allocation
2609 * within an allocation group which is a part (not first part)
2610 * of a larger binary buddy system and a few exception cases
2611 * in which a previous join operation must be backed out.
2612 *
2613 * PARAMETERS:
2614 * tp - pointer to the tree containing the leaf.
2615 * leafno - the number of the leaf to be updated.
2616 *
2617 * RETURN VALUES: none
2618 *
2619 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2620 */
2621static int dbBackSplit(dmtree_t * tp, int leafno)
2622{
2623    int budsz, bud, w, bsz, size;
2624    int cursz;
2625    s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2626
2627    /* leaf should be part (not first part) of a binary
2628     * buddy system.
2629     */
2630    assert(leaf[leafno] == NOFREE);
2631
2632    /* the back split is accomplished by iteratively finding the leaf
2633     * that starts the buddy system that contains the specified leaf and
2634     * splitting that system in two. this iteration continues until
2635     * the specified leaf becomes the start of a buddy system.
2636     *
2637     * determine maximum possible l2 size for the specified leaf.
2638     */
2639    size =
2640        LITOL2BSZ(leafno, le32_to_cpu(tp->dmt_l2nleafs),
2641              tp->dmt_budmin);
2642
2643    /* determine the number of leaves covered by this size. this
2644     * is the buddy size that we will start with as we search for
2645     * the buddy system that contains the specified leaf.
2646     */
2647    budsz = BUDSIZE(size, tp->dmt_budmin);
2648
2649    /* back split.
2650     */
2651    while (leaf[leafno] == NOFREE) {
2652        /* find the leftmost buddy leaf.
2653         */
2654        for (w = leafno, bsz = budsz;; bsz <<= 1,
2655             w = (w < bud) ? w : bud) {
2656            if (bsz >= le32_to_cpu(tp->dmt_nleafs)) {
2657                jfs_err("JFS: block map error in dbBackSplit");
2658                return -EIO;
2659            }
2660
2661            /* determine the buddy.
2662             */
2663            bud = w ^ bsz;
2664
2665            /* check if this buddy is the start of the system.
2666             */
2667            if (leaf[bud] != NOFREE) {
2668                /* split the leaf at the start of the
2669                 * system in two.
2670                 */
2671                cursz = leaf[bud] - 1;
2672                dbSplit(tp, bud, cursz, cursz);
2673                break;
2674            }
2675        }
2676    }
2677
2678    if (leaf[leafno] != size) {
2679        jfs_err("JFS: wrong leaf value in dbBackSplit");
2680        return -EIO;
2681    }
2682    return 0;
2683}
2684
2685
2686/*
2687 * NAME: dbJoin()
2688 *
2689 * FUNCTION: update the leaf of a dmtree with a new value, joining
2690 * the leaf with other leaves of the dmtree into a multi-leaf
2691 * binary buddy system, as required.
2692 *
2693 * PARAMETERS:
2694 * tp - pointer to the tree containing the leaf.
2695 * leafno - the number of the leaf to be updated.
2696 * newval - the new value for the leaf.
2697 *
2698 * RETURN VALUES: none
2699 */
2700static int dbJoin(dmtree_t * tp, int leafno, int newval)
2701{
2702    int budsz, buddy;
2703    s8 *leaf;
2704
2705    /* can the new leaf value require a join with other leaves ?
2706     */
2707    if (newval >= tp->dmt_budmin) {
2708        /* pickup a pointer to the leaves of the tree.
2709         */
2710        leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2711
2712        /* try to join the specified leaf into a large binary
2713         * buddy system. the join proceeds by attempting to join
2714         * the specified leafno with its buddy (leaf) at new value.
2715         * if the join occurs, we attempt to join the left leaf
2716         * of the joined buddies with its buddy at new value + 1.
2717         * we continue to join until we find a buddy that cannot be
2718         * joined (does not have a value equal to the size of the
2719         * last join) or until all leaves have been joined into a
2720         * single system.
2721         *
2722         * get the buddy size (number of words covered) of
2723         * the new value.
2724         */
2725        budsz = BUDSIZE(newval, tp->dmt_budmin);
2726
2727        /* try to join.
2728         */
2729        while (budsz < le32_to_cpu(tp->dmt_nleafs)) {
2730            /* get the buddy leaf.
2731             */
2732            buddy = leafno ^ budsz;
2733
2734            /* if the leaf's new value is greater than its
2735             * buddy's value, we join no more.
2736             */
2737            if (newval > leaf[buddy])
2738                break;
2739
2740            /* It shouldn't be less */
2741            if (newval < leaf[buddy])
2742                return -EIO;
2743
2744            /* check which (leafno or buddy) is the left buddy.
2745             * the left buddy gets to claim the blocks resulting
2746             * from the join while the right gets to claim none.
2747             * the left buddy is also eligible to participate in
2748             * a join at the next higher level while the right
2749             * is not.
2750             *
2751             */
2752            if (leafno < buddy) {
2753                /* leafno is the left buddy.
2754                 */
2755                dbAdjTree(tp, buddy, NOFREE);
2756            } else {
2757                /* buddy is the left buddy and becomes
2758                 * leafno.
2759                 */
2760                dbAdjTree(tp, leafno, NOFREE);
2761                leafno = buddy;
2762            }
2763
2764            /* on to try the next join.
2765             */
2766            newval += 1;
2767            budsz <<= 1;
2768        }
2769    }
2770
2771    /* update the leaf value.
2772     */
2773    dbAdjTree(tp, leafno, newval);
2774
2775    return 0;
2776}
2777
2778
2779/*
2780 * NAME: dbAdjTree()
2781 *
2782 * FUNCTION: update a leaf of a dmtree with a new value, adjusting
2783 * the dmtree, as required, to reflect the new leaf value.
2784 * the combination of any buddies must already be done before
2785 * this is called.
2786 *
2787 * PARAMETERS:
2788 * tp - pointer to the tree to be adjusted.
2789 * leafno - the number of the leaf to be updated.
2790 * newval - the new value for the leaf.
2791 *
2792 * RETURN VALUES: none
2793 */
2794static void dbAdjTree(dmtree_t * tp, int leafno, int newval)
2795{
2796    int lp, pp, k;
2797    int max;
2798
2799    /* pick up the index of the leaf for this leafno.
2800     */
2801    lp = leafno + le32_to_cpu(tp->dmt_leafidx);
2802
2803    /* is the current value the same as the old value ? if so,
2804     * there is nothing to do.
2805     */
2806    if (tp->dmt_stree[lp] == newval)
2807        return;
2808
2809    /* set the new value.
2810     */
2811    tp->dmt_stree[lp] = newval;
2812
2813    /* bubble the new value up the tree as required.
2814     */
2815    for (k = 0; k < le32_to_cpu(tp->dmt_height); k++) {
2816        /* get the index of the first leaf of the 4 leaf
2817         * group containing the specified leaf (leafno).
2818         */
2819        lp = ((lp - 1) & ~0x03) + 1;
2820
2821        /* get the index of the parent of this 4 leaf group.
2822         */
2823        pp = (lp - 1) >> 2;
2824
2825        /* determine the maximum of the 4 leaves.
2826         */
2827        max = TREEMAX(&tp->dmt_stree[lp]);
2828
2829        /* if the maximum of the 4 is the same as the
2830         * parent's value, we're done.
2831         */
2832        if (tp->dmt_stree[pp] == max)
2833            break;
2834
2835        /* parent gets new value.
2836         */
2837        tp->dmt_stree[pp] = max;
2838
2839        /* parent becomes leaf for next go-round.
2840         */
2841        lp = pp;
2842    }
2843}
2844
2845
2846/*
2847 * NAME: dbFindLeaf()
2848 *
2849 * FUNCTION: search a dmtree_t for sufficient free blocks, returning
2850 * the index of a leaf describing the free blocks if
2851 * sufficient free blocks are found.
2852 *
2853 * the search starts at the top of the dmtree_t tree and
2854 * proceeds down the tree to the leftmost leaf with sufficient
2855 * free space.
2856 *
2857 * PARAMETERS:
2858 * tp - pointer to the tree to be searched.
2859 * l2nb - log2 number of free blocks to search for.
2860 * leafidx - return pointer to be set to the index of the leaf
2861 * describing at least l2nb free blocks if sufficient
2862 * free blocks are found.
2863 *
2864 * RETURN VALUES:
2865 * 0 - success
2866 * -ENOSPC - insufficient free blocks.
2867 */
2868static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx)
2869{
2870    int ti, n = 0, k, x = 0;
2871
2872    /* first check the root of the tree to see if there is
2873     * sufficient free space.
2874     */
2875    if (l2nb > tp->dmt_stree[ROOT])
2876        return -ENOSPC;
2877
2878    /* sufficient free space available. now search down the tree
2879     * starting at the next level for the leftmost leaf that
2880     * describes sufficient free space.
2881     */
2882    for (k = le32_to_cpu(tp->dmt_height), ti = 1;
2883         k > 0; k--, ti = ((ti + n) << 2) + 1) {
2884        /* search the four nodes at this level, starting from
2885         * the left.
2886         */
2887        for (x = ti, n = 0; n < 4; n++) {
2888            /* sufficient free space found. move to the next
2889             * level (or quit if this is the last level).
2890             */
2891            if (l2nb <= tp->dmt_stree[x + n])
2892                break;
2893        }
2894
2895        /* better have found something since the higher
2896         * levels of the tree said it was here.
2897         */
2898        assert(n < 4);
2899    }
2900
2901    /* set the return to the leftmost leaf describing sufficient
2902     * free space.
2903     */
2904    *leafidx = x + n - le32_to_cpu(tp->dmt_leafidx);
2905
2906    return (0);
2907}
2908
2909
2910/*
2911 * NAME: dbFindBits()
2912 *
2913 * FUNCTION: find a specified number of binary buddy free bits within a
2914 * dmap bitmap word value.
2915 *
2916 * this routine searches the bitmap value for (1 << l2nb) free
2917 * bits at (1 << l2nb) alignments within the value.
2918 *
2919 * PARAMETERS:
2920 * word - dmap bitmap word value.
2921 * l2nb - number of free bits specified as a log2 number.
2922 *
2923 * RETURN VALUES:
2924 * starting bit number of free bits.
2925 */
2926static int dbFindBits(u32 word, int l2nb)
2927{
2928    int bitno, nb;
2929    u32 mask;
2930
2931    /* get the number of bits.
2932     */
2933    nb = 1 << l2nb;
2934    assert(nb <= DBWORD);
2935
2936    /* complement the word so we can use a mask (i.e. 0s represent
2937     * free bits) and compute the mask.
2938     */
2939    word = ~word;
2940    mask = ONES << (DBWORD - nb);
2941
2942    /* scan the word for nb free bits at nb alignments.
2943     */
2944    for (bitno = 0; mask != 0; bitno += nb, mask >>= nb) {
2945        if ((mask & word) == mask)
2946            break;
2947    }
2948
2949    ASSERT(bitno < 32);
2950
2951    /* return the bit number.
2952     */
2953    return (bitno);
2954}
2955
2956
2957/*
2958 * NAME: dbMaxBud(u8 *cp)
2959 *
2960 * FUNCTION: determine the largest binary buddy string of free
2961 * bits within 32-bits of the map.
2962 *
2963 * PARAMETERS:
2964 * cp - pointer to the 32-bit value.
2965 *
2966 * RETURN VALUES:
2967 * largest binary buddy of free bits within a dmap word.
2968 */
2969static int dbMaxBud(u8 * cp)
2970{
2971    signed char tmp1, tmp2;
2972
2973    /* check if the wmap word is all free. if so, the
2974     * free buddy size is BUDMIN.
2975     */
2976    if (*((uint *) cp) == 0)
2977        return (BUDMIN);
2978
2979    /* check if the wmap word is half free. if so, the
2980     * free buddy size is BUDMIN-1.
2981     */
2982    if (*((u16 *) cp) == 0 || *((u16 *) cp + 1) == 0)
2983        return (BUDMIN - 1);
2984
2985    /* not all free or half free. determine the free buddy
2986     * size thru table lookup using quarters of the wmap word.
2987     */
2988    tmp1 = max(budtab[cp[2]], budtab[cp[3]]);
2989    tmp2 = max(budtab[cp[0]], budtab[cp[1]]);
2990    return (max(tmp1, tmp2));
2991}
2992
2993
2994/*
2995 * NAME: cnttz(uint word)
2996 *
2997 * FUNCTION: determine the number of trailing zeros within a 32-bit
2998 * value.
2999 *
3000 * PARAMETERS:
3001 * value - 32-bit value to be examined.
3002 *
3003 * RETURN VALUES:
3004 * count of trailing zeros
3005 */
3006static int cnttz(u32 word)
3007{
3008    int n;
3009
3010    for (n = 0; n < 32; n++, word >>= 1) {
3011        if (word & 0x01)
3012            break;
3013    }
3014
3015    return (n);
3016}
3017
3018
3019/*
3020 * NAME: cntlz(u32 value)
3021 *
3022 * FUNCTION: determine the number of leading zeros within a 32-bit
3023 * value.
3024 *
3025 * PARAMETERS:
3026 * value - 32-bit value to be examined.
3027 *
3028 * RETURN VALUES:
3029 * count of leading zeros
3030 */
3031static int cntlz(u32 value)
3032{
3033    int n;
3034
3035    for (n = 0; n < 32; n++, value <<= 1) {
3036        if (value & HIGHORDER)
3037            break;
3038    }
3039    return (n);
3040}
3041
3042
3043/*
3044 * NAME: blkstol2(s64 nb)
3045 *
3046 * FUNCTION: convert a block count to its log2 value. if the block
3047 * count is not a l2 multiple, it is rounded up to the next
3048 * larger l2 multiple.
3049 *
3050 * PARAMETERS:
3051 * nb - number of blocks
3052 *
3053 * RETURN VALUES:
3054 * log2 number of blocks
3055 */
3056static int blkstol2(s64 nb)
3057{
3058    int l2nb;
3059    s64 mask; /* meant to be signed */
3060
3061    mask = (s64) 1 << (64 - 1);
3062
3063    /* count the leading bits.
3064     */
3065    for (l2nb = 0; l2nb < 64; l2nb++, mask >>= 1) {
3066        /* leading bit found.
3067         */
3068        if (nb & mask) {
3069            /* determine the l2 value.
3070             */
3071            l2nb = (64 - 1) - l2nb;
3072
3073            /* check if we need to round up.
3074             */
3075            if (~mask & nb)
3076                l2nb++;
3077
3078            return (l2nb);
3079        }
3080    }
3081    assert(0);
3082    return 0; /* fix compiler warning */
3083}
3084
3085
3086/*
3087 * NAME: dbAllocBottomUp()
3088 *
3089 * FUNCTION: alloc the specified block range from the working block
3090 * allocation map.
3091 *
3092 * the blocks will be alloc from the working map one dmap
3093 * at a time.
3094 *
3095 * PARAMETERS:
3096 * ip - pointer to in-core inode;
3097 * blkno - starting block number to be freed.
3098 * nblocks - number of blocks to be freed.
3099 *
3100 * RETURN VALUES:
3101 * 0 - success
3102 * -EIO - i/o error
3103 */
3104int dbAllocBottomUp(struct inode *ip, s64 blkno, s64 nblocks)
3105{
3106    struct metapage *mp;
3107    struct dmap *dp;
3108    int nb, rc;
3109    s64 lblkno, rem;
3110    struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
3111    struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
3112
3113    IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
3114
3115    /* block to be allocated better be within the mapsize. */
3116    ASSERT(nblocks <= bmp->db_mapsize - blkno);
3117
3118    /*
3119     * allocate the blocks a dmap at a time.
3120     */
3121    mp = NULL;
3122    for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
3123        /* release previous dmap if any */
3124        if (mp) {
3125            write_metapage(mp);
3126        }
3127
3128        /* get the buffer for the current dmap. */
3129        lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
3130        mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
3131        if (mp == NULL) {
3132            IREAD_UNLOCK(ipbmap);
3133            return -EIO;
3134        }
3135        dp = (struct dmap *) mp->data;
3136
3137        /* determine the number of blocks to be allocated from
3138         * this dmap.
3139         */
3140        nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
3141
3142        /* allocate the blocks. */
3143        if ((rc = dbAllocDmapBU(bmp, dp, blkno, nb))) {
3144            release_metapage(mp);
3145            IREAD_UNLOCK(ipbmap);
3146            return (rc);
3147        }
3148    }
3149
3150    /* write the last buffer. */
3151    write_metapage(mp);
3152
3153    IREAD_UNLOCK(ipbmap);
3154
3155    return (0);
3156}
3157
3158
3159static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
3160             int nblocks)
3161{
3162    int rc;
3163    int dbitno, word, rembits, nb, nwords, wbitno, agno;
3164    s8 oldroot, *leaf;
3165    struct dmaptree *tp = (struct dmaptree *) & dp->tree;
3166
3167    /* save the current value of the root (i.e. maximum free string)
3168     * of the dmap tree.
3169     */
3170    oldroot = tp->stree[ROOT];
3171
3172    /* pick up a pointer to the leaves of the dmap tree */
3173    leaf = tp->stree + LEAFIND;
3174
3175    /* determine the bit number and word within the dmap of the
3176     * starting block.
3177     */
3178    dbitno = blkno & (BPERDMAP - 1);
3179    word = dbitno >> L2DBWORD;
3180
3181    /* block range better be within the dmap */
3182    assert(dbitno + nblocks <= BPERDMAP);
3183
3184    /* allocate the bits of the dmap's words corresponding to the block
3185     * range. not all bits of the first and last words may be contained
3186     * within the block range. if this is the case, we'll work against
3187     * those words (i.e. partial first and/or last) on an individual basis
3188     * (a single pass), allocating the bits of interest by hand and
3189     * updating the leaf corresponding to the dmap word. a single pass
3190     * will be used for all dmap words fully contained within the
3191     * specified range. within this pass, the bits of all fully contained
3192     * dmap words will be marked as free in a single shot and the leaves
3193     * will be updated. a single leaf may describe the free space of
3194     * multiple dmap words, so we may update only a subset of the actual
3195     * leaves corresponding to the dmap words of the block range.
3196     */
3197    for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
3198        /* determine the bit number within the word and
3199         * the number of bits within the word.
3200         */
3201        wbitno = dbitno & (DBWORD - 1);
3202        nb = min(rembits, DBWORD - wbitno);
3203
3204        /* check if only part of a word is to be allocated.
3205         */
3206        if (nb < DBWORD) {
3207            /* allocate (set to 1) the appropriate bits within
3208             * this dmap word.
3209             */
3210            dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
3211                              >> wbitno);
3212
3213            word++;
3214        } else {
3215            /* one or more dmap words are fully contained
3216             * within the block range. determine how many
3217             * words and allocate (set to 1) the bits of these
3218             * words.
3219             */
3220            nwords = rembits >> L2DBWORD;
3221            memset(&dp->wmap[word], (int) ONES, nwords * 4);
3222
3223            /* determine how many bits */
3224            nb = nwords << L2DBWORD;
3225            word += nwords;
3226        }
3227    }
3228
3229    /* update the free count for this dmap */
3230    le32_add_cpu(&dp->nfree, -nblocks);
3231
3232    /* reconstruct summary tree */
3233    dbInitDmapTree(dp);
3234
3235    BMAP_LOCK(bmp);
3236
3237    /* if this allocation group is completely free,
3238     * update the highest active allocation group number
3239     * if this allocation group is the new max.
3240     */
3241    agno = blkno >> bmp->db_agl2size;
3242    if (agno > bmp->db_maxag)
3243        bmp->db_maxag = agno;
3244
3245    /* update the free count for the allocation group and map */
3246    bmp->db_agfree[agno] -= nblocks;
3247    bmp->db_nfree -= nblocks;
3248
3249    BMAP_UNLOCK(bmp);
3250
3251    /* if the root has not changed, done. */
3252    if (tp->stree[ROOT] == oldroot)
3253        return (0);
3254
3255    /* root changed. bubble the change up to the dmap control pages.
3256     * if the adjustment of the upper level control pages fails,
3257     * backout the bit allocation (thus making everything consistent).
3258     */
3259    if ((rc = dbAdjCtl(bmp, blkno, tp->stree[ROOT], 1, 0)))
3260        dbFreeBits(bmp, dp, blkno, nblocks);
3261
3262    return (rc);
3263}
3264
3265
3266/*
3267 * NAME: dbExtendFS()
3268 *
3269 * FUNCTION: extend bmap from blkno for nblocks;
3270 * dbExtendFS() updates bmap ready for dbAllocBottomUp();
3271 *
3272 * L2
3273 * |
3274 * L1---------------------------------L1
3275 * | |
3276 * L0---------L0---------L0 L0---------L0---------L0
3277 * | | | | | |
3278 * d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,.,dm;
3279 * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm
3280 *
3281 * <---old---><----------------------------extend----------------------->
3282 */
3283int dbExtendFS(struct inode *ipbmap, s64 blkno, s64 nblocks)
3284{
3285    struct jfs_sb_info *sbi = JFS_SBI(ipbmap->i_sb);
3286    int nbperpage = sbi->nbperpage;
3287    int i, i0 = true, j, j0 = true, k, n;
3288    s64 newsize;
3289    s64 p;
3290    struct metapage *mp, *l2mp, *l1mp = NULL, *l0mp = NULL;
3291    struct dmapctl *l2dcp, *l1dcp, *l0dcp;
3292    struct dmap *dp;
3293    s8 *l0leaf, *l1leaf, *l2leaf;
3294    struct bmap *bmp = sbi->bmap;
3295    int agno, l2agsize, oldl2agsize;
3296    s64 ag_rem;
3297
3298    newsize = blkno + nblocks;
3299
3300    jfs_info("dbExtendFS: blkno:%Ld nblocks:%Ld newsize:%Ld",
3301         (long long) blkno, (long long) nblocks, (long long) newsize);
3302
3303    /*
3304     * initialize bmap control page.
3305     *
3306     * all the data in bmap control page should exclude
3307     * the mkfs hidden dmap page.
3308     */
3309
3310    /* update mapsize */
3311    bmp->db_mapsize = newsize;
3312    bmp->db_maxlevel = BMAPSZTOLEV(bmp->db_mapsize);
3313
3314    /* compute new AG size */
3315    l2agsize = dbGetL2AGSize(newsize);
3316    oldl2agsize = bmp->db_agl2size;
3317
3318    bmp->db_agl2size = l2agsize;
3319    bmp->db_agsize = 1 << l2agsize;
3320
3321    /* compute new number of AG */
3322    agno = bmp->db_numag;
3323    bmp->db_numag = newsize >> l2agsize;
3324    bmp->db_numag += ((u32) newsize % (u32) bmp->db_agsize) ? 1 : 0;
3325
3326    /*
3327     * reconfigure db_agfree[]
3328     * from old AG configuration to new AG configuration;
3329     *
3330     * coalesce contiguous k (newAGSize/oldAGSize) AGs;
3331     * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn;
3332     * note: new AG size = old AG size * (2**x).
3333     */
3334    if (l2agsize == oldl2agsize)
3335        goto extend;
3336    k = 1 << (l2agsize - oldl2agsize);
3337    ag_rem = bmp->db_agfree[0]; /* save agfree[0] */
3338    for (i = 0, n = 0; i < agno; n++) {
3339        bmp->db_agfree[n] = 0; /* init collection point */
3340
3341        /* coalesce contiguous k AGs; */
3342        for (j = 0; j < k && i < agno; j++, i++) {
3343            /* merge AGi to AGn */
3344            bmp->db_agfree[n] += bmp->db_agfree[i];
3345        }
3346    }
3347    bmp->db_agfree[0] += ag_rem; /* restore agfree[0] */
3348
3349    for (; n < MAXAG; n++)
3350        bmp->db_agfree[n] = 0;
3351
3352    /*
3353     * update highest active ag number
3354     */
3355
3356    bmp->db_maxag = bmp->db_maxag / k;
3357
3358    /*
3359     * extend bmap
3360     *
3361     * update bit maps and corresponding level control pages;
3362     * global control page db_nfree, db_agfree[agno], db_maxfreebud;
3363     */
3364      extend:
3365    /* get L2 page */
3366    p = BMAPBLKNO + nbperpage; /* L2 page */
3367    l2mp = read_metapage(ipbmap, p, PSIZE, 0);
3368    if (!l2mp) {
3369        jfs_error(ipbmap->i_sb, "dbExtendFS: L2 page could not be read");
3370        return -EIO;
3371    }
3372    l2dcp = (struct dmapctl *) l2mp->data;
3373
3374    /* compute start L1 */
3375    k = blkno >> L2MAXL1SIZE;
3376    l2leaf = l2dcp->stree + CTLLEAFIND + k;
3377    p = BLKTOL1(blkno, sbi->l2nbperpage); /* L1 page */
3378
3379    /*
3380     * extend each L1 in L2
3381     */
3382    for (; k < LPERCTL; k++, p += nbperpage) {
3383        /* get L1 page */
3384        if (j0) {
3385            /* read in L1 page: (blkno & (MAXL1SIZE - 1)) */
3386            l1mp = read_metapage(ipbmap, p, PSIZE, 0);
3387            if (l1mp == NULL)
3388                goto errout;
3389            l1dcp = (struct dmapctl *) l1mp->data;
3390
3391            /* compute start L0 */
3392            j = (blkno & (MAXL1SIZE - 1)) >> L2MAXL0SIZE;
3393            l1leaf = l1dcp->stree + CTLLEAFIND + j;
3394            p = BLKTOL0(blkno, sbi->l2nbperpage);
3395            j0 = false;
3396        } else {
3397            /* assign/init L1 page */
3398            l1mp = get_metapage(ipbmap, p, PSIZE, 0);
3399            if (l1mp == NULL)
3400                goto errout;
3401
3402            l1dcp = (struct dmapctl *) l1mp->data;
3403
3404            /* compute start L0 */
3405            j = 0;
3406            l1leaf = l1dcp->stree + CTLLEAFIND;
3407            p += nbperpage; /* 1st L0 of L1.k */
3408        }
3409
3410        /*
3411         * extend each L0 in L1
3412         */
3413        for (; j < LPERCTL; j++) {
3414            /* get L0 page */
3415            if (i0) {
3416                /* read in L0 page: (blkno & (MAXL0SIZE - 1)) */
3417
3418                l0mp = read_metapage(ipbmap, p, PSIZE, 0);
3419                if (l0mp == NULL)
3420                    goto errout;
3421                l0dcp = (struct dmapctl *) l0mp->data;
3422
3423                /* compute start dmap */
3424                i = (blkno & (MAXL0SIZE - 1)) >>
3425                    L2BPERDMAP;
3426                l0leaf = l0dcp->stree + CTLLEAFIND + i;
3427                p = BLKTODMAP(blkno,
3428                          sbi->l2nbperpage);
3429                i0 = false;
3430            } else {
3431                /* assign/init L0 page */
3432                l0mp = get_metapage(ipbmap, p, PSIZE, 0);
3433                if (l0mp == NULL)
3434                    goto errout;
3435
3436                l0dcp = (struct dmapctl *) l0mp->data;
3437
3438                /* compute start dmap */
3439                i = 0;
3440                l0leaf = l0dcp->stree + CTLLEAFIND;
3441                p += nbperpage; /* 1st dmap of L0.j */
3442            }
3443
3444            /*
3445             * extend each dmap in L0
3446             */
3447            for (; i < LPERCTL; i++) {
3448                /*
3449                 * reconstruct the dmap page, and
3450                 * initialize corresponding parent L0 leaf
3451                 */
3452                if ((n = blkno & (BPERDMAP - 1))) {
3453                    /* read in dmap page: */
3454                    mp = read_metapage(ipbmap, p,
3455                               PSIZE, 0);
3456                    if (mp == NULL)
3457                        goto errout;
3458                    n = min(nblocks, (s64)BPERDMAP - n);
3459                } else {
3460                    /* assign/init dmap page */
3461                    mp = read_metapage(ipbmap, p,
3462                               PSIZE, 0);
3463                    if (mp == NULL)
3464                        goto errout;
3465
3466                    n = min(nblocks, (s64)BPERDMAP);
3467                }
3468
3469                dp = (struct dmap *) mp->data;
3470                *l0leaf = dbInitDmap(dp, blkno, n);
3471
3472                bmp->db_nfree += n;
3473                agno = le64_to_cpu(dp->start) >> l2agsize;
3474                bmp->db_agfree[agno] += n;
3475
3476                write_metapage(mp);
3477
3478                l0leaf++;
3479                p += nbperpage;
3480
3481                blkno += n;
3482                nblocks -= n;
3483                if (nblocks == 0)
3484                    break;
3485            } /* for each dmap in a L0 */
3486
3487            /*
3488             * build current L0 page from its leaves, and
3489             * initialize corresponding parent L1 leaf
3490             */
3491            *l1leaf = dbInitDmapCtl(l0dcp, 0, ++i);
3492            write_metapage(l0mp);
3493            l0mp = NULL;
3494
3495            if (nblocks)
3496                l1leaf++; /* continue for next L0 */
3497            else {
3498                /* more than 1 L0 ? */
3499                if (j > 0)
3500                    break; /* build L1 page */
3501                else {
3502                    /* summarize in global bmap page */
3503                    bmp->db_maxfreebud = *l1leaf;
3504                    release_metapage(l1mp);
3505                    release_metapage(l2mp);
3506                    goto finalize;
3507                }
3508            }
3509        } /* for each L0 in a L1 */
3510
3511        /*
3512         * build current L1 page from its leaves, and
3513         * initialize corresponding parent L2 leaf
3514         */
3515        *l2leaf = dbInitDmapCtl(l1dcp, 1, ++j);
3516        write_metapage(l1mp);
3517        l1mp = NULL;
3518
3519        if (nblocks)
3520            l2leaf++; /* continue for next L1 */
3521        else {
3522            /* more than 1 L1 ? */
3523            if (k > 0)
3524                break; /* build L2 page */
3525            else {
3526                /* summarize in global bmap page */
3527                bmp->db_maxfreebud = *l2leaf;
3528                release_metapage(l2mp);
3529                goto finalize;
3530            }
3531        }
3532    } /* for each L1 in a L2 */
3533
3534    jfs_error(ipbmap->i_sb,
3535          "dbExtendFS: function has not returned as expected");
3536errout:
3537    if (l0mp)
3538        release_metapage(l0mp);
3539    if (l1mp)
3540        release_metapage(l1mp);
3541    release_metapage(l2mp);
3542    return -EIO;
3543
3544    /*
3545     * finalize bmap control page
3546     */
3547finalize:
3548
3549    return 0;
3550}
3551
3552
3553/*
3554 * dbFinalizeBmap()
3555 */
3556void dbFinalizeBmap(struct inode *ipbmap)
3557{
3558    struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
3559    int actags, inactags, l2nl;
3560    s64 ag_rem, actfree, inactfree, avgfree;
3561    int i, n;
3562
3563    /*
3564     * finalize bmap control page
3565     */
3566//finalize:
3567    /*
3568     * compute db_agpref: preferred ag to allocate from
3569     * (the leftmost ag with average free space in it);
3570     */
3571//agpref:
3572    /* get the number of active ags and inacitve ags */
3573    actags = bmp->db_maxag + 1;
3574    inactags = bmp->db_numag - actags;
3575    ag_rem = bmp->db_mapsize & (bmp->db_agsize - 1); /* ??? */
3576
3577    /* determine how many blocks are in the inactive allocation
3578     * groups. in doing this, we must account for the fact that
3579     * the rightmost group might be a partial group (i.e. file
3580     * system size is not a multiple of the group size).
3581     */
3582    inactfree = (inactags && ag_rem) ?
3583        ((inactags - 1) << bmp->db_agl2size) + ag_rem
3584        : inactags << bmp->db_agl2size;
3585
3586    /* determine how many free blocks are in the active
3587     * allocation groups plus the average number of free blocks
3588     * within the active ags.
3589     */
3590    actfree = bmp->db_nfree - inactfree;
3591    avgfree = (u32) actfree / (u32) actags;
3592
3593    /* if the preferred allocation group has not average free space.
3594     * re-establish the preferred group as the leftmost
3595     * group with average free space.
3596     */
3597    if (bmp->db_agfree[bmp->db_agpref] < avgfree) {
3598        for (bmp->db_agpref = 0; bmp->db_agpref < actags;
3599             bmp->db_agpref++) {
3600            if (bmp->db_agfree[bmp->db_agpref] >= avgfree)
3601                break;
3602        }
3603        if (bmp->db_agpref >= bmp->db_numag) {
3604            jfs_error(ipbmap->i_sb,
3605                  "cannot find ag with average freespace");
3606        }
3607    }
3608
3609    /*
3610     * compute db_aglevel, db_agheight, db_width, db_agstart:
3611     * an ag is covered in aglevel dmapctl summary tree,
3612     * at agheight level height (from leaf) with agwidth number of nodes
3613     * each, which starts at agstart index node of the smmary tree node
3614     * array;
3615     */
3616    bmp->db_aglevel = BMAPSZTOLEV(bmp->db_agsize);
3617    l2nl =
3618        bmp->db_agl2size - (L2BPERDMAP + bmp->db_aglevel * L2LPERCTL);
3619    bmp->db_agheight = l2nl >> 1;
3620    bmp->db_agwidth = 1 << (l2nl - (bmp->db_agheight << 1));
3621    for (i = 5 - bmp->db_agheight, bmp->db_agstart = 0, n = 1; i > 0;
3622         i--) {
3623        bmp->db_agstart += n;
3624        n <<= 2;
3625    }
3626
3627}
3628
3629
3630/*
3631 * NAME: dbInitDmap()/ujfs_idmap_page()
3632 *
3633 * FUNCTION: initialize working/persistent bitmap of the dmap page
3634 * for the specified number of blocks:
3635 *
3636 * at entry, the bitmaps had been initialized as free (ZEROS);
3637 * The number of blocks will only account for the actually
3638 * existing blocks. Blocks which don't actually exist in
3639 * the aggregate will be marked as allocated (ONES);
3640 *
3641 * PARAMETERS:
3642 * dp - pointer to page of map
3643 * nblocks - number of blocks this page
3644 *
3645 * RETURNS: NONE
3646 */
3647static int dbInitDmap(struct dmap * dp, s64 Blkno, int nblocks)
3648{
3649    int blkno, w, b, r, nw, nb, i;
3650
3651    /* starting block number within the dmap */
3652    blkno = Blkno & (BPERDMAP - 1);
3653
3654    if (blkno == 0) {
3655        dp->nblocks = dp->nfree = cpu_to_le32(nblocks);
3656        dp->start = cpu_to_le64(Blkno);
3657
3658        if (nblocks == BPERDMAP) {
3659            memset(&dp->wmap[0], 0, LPERDMAP * 4);
3660            memset(&dp->pmap[0], 0, LPERDMAP * 4);
3661            goto initTree;
3662        }
3663    } else {
3664        le32_add_cpu(&dp->nblocks, nblocks);
3665        le32_add_cpu(&dp->nfree, nblocks);
3666    }
3667
3668    /* word number containing start block number */
3669    w = blkno >> L2DBWORD;
3670
3671    /*
3672     * free the bits corresponding to the block range (ZEROS):
3673     * note: not all bits of the first and last words may be contained
3674     * within the block range.
3675     */
3676    for (r = nblocks; r > 0; r -= nb, blkno += nb) {
3677        /* number of bits preceding range to be freed in the word */
3678        b = blkno & (DBWORD - 1);
3679        /* number of bits to free in the word */
3680        nb = min(r, DBWORD - b);
3681
3682        /* is partial word to be freed ? */
3683        if (nb < DBWORD) {
3684            /* free (set to 0) from the bitmap word */
3685            dp->wmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3686                             >> b));
3687            dp->pmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3688                             >> b));
3689
3690            /* skip the word freed */
3691            w++;
3692        } else {
3693            /* free (set to 0) contiguous bitmap words */
3694            nw = r >> L2DBWORD;
3695            memset(&dp->wmap[w], 0, nw * 4);
3696            memset(&dp->pmap[w], 0, nw * 4);
3697
3698            /* skip the words freed */
3699            nb = nw << L2DBWORD;
3700            w += nw;
3701        }
3702    }
3703
3704    /*
3705     * mark bits following the range to be freed (non-existing
3706     * blocks) as allocated (ONES)
3707     */
3708
3709    if (blkno == BPERDMAP)
3710        goto initTree;
3711
3712    /* the first word beyond the end of existing blocks */
3713    w = blkno >> L2DBWORD;
3714
3715    /* does nblocks fall on a 32-bit boundary ? */
3716    b = blkno & (DBWORD - 1);
3717    if (b) {
3718        /* mark a partial word allocated */
3719        dp->wmap[w] = dp->pmap[w] = cpu_to_le32(ONES >> b);
3720        w++;
3721    }
3722
3723    /* set the rest of the words in the page to allocated (ONES) */
3724    for (i = w; i < LPERDMAP; i++)
3725        dp->pmap[i] = dp->wmap[i] = cpu_to_le32(ONES);
3726
3727    /*
3728     * init tree
3729     */
3730      initTree:
3731    return (dbInitDmapTree(dp));
3732}
3733
3734
3735/*
3736 * NAME: dbInitDmapTree()/ujfs_complete_dmap()
3737 *
3738 * FUNCTION: initialize summary tree of the specified dmap:
3739 *
3740 * at entry, bitmap of the dmap has been initialized;
3741 *
3742 * PARAMETERS:
3743 * dp - dmap to complete
3744 * blkno - starting block number for this dmap
3745 * treemax - will be filled in with max free for this dmap
3746 *
3747 * RETURNS: max free string at the root of the tree
3748 */
3749static int dbInitDmapTree(struct dmap * dp)
3750{
3751    struct dmaptree *tp;
3752    s8 *cp;
3753    int i;
3754
3755    /* init fixed info of tree */
3756    tp = &dp->tree;
3757    tp->nleafs = cpu_to_le32(LPERDMAP);
3758    tp->l2nleafs = cpu_to_le32(L2LPERDMAP);
3759    tp->leafidx = cpu_to_le32(LEAFIND);
3760    tp->height = cpu_to_le32(4);
3761    tp->budmin = BUDMIN;
3762
3763    /* init each leaf from corresponding wmap word:
3764     * note: leaf is set to NOFREE(-1) if all blocks of corresponding
3765     * bitmap word are allocated.
3766     */
3767    cp = tp->stree + le32_to_cpu(tp->leafidx);
3768    for (i = 0; i < LPERDMAP; i++)
3769        *cp++ = dbMaxBud((u8 *) & dp->wmap[i]);
3770
3771    /* build the dmap's binary buddy summary tree */
3772    return (dbInitTree(tp));
3773}
3774
3775
3776/*
3777 * NAME: dbInitTree()/ujfs_adjtree()
3778 *
3779 * FUNCTION: initialize binary buddy summary tree of a dmap or dmapctl.
3780 *
3781 * at entry, the leaves of the tree has been initialized
3782 * from corresponding bitmap word or root of summary tree
3783 * of the child control page;
3784 * configure binary buddy system at the leaf level, then
3785 * bubble up the values of the leaf nodes up the tree.
3786 *
3787 * PARAMETERS:
3788 * cp - Pointer to the root of the tree
3789 * l2leaves- Number of leaf nodes as a power of 2
3790 * l2min - Number of blocks that can be covered by a leaf
3791 * as a power of 2
3792 *
3793 * RETURNS: max free string at the root of the tree
3794 */
3795static int dbInitTree(struct dmaptree * dtp)
3796{
3797    int l2max, l2free, bsize, nextb, i;
3798    int child, parent, nparent;
3799    s8 *tp, *cp, *cp1;
3800
3801    tp = dtp->stree;
3802
3803    /* Determine the maximum free string possible for the leaves */
3804    l2max = le32_to_cpu(dtp->l2nleafs) + dtp->budmin;
3805
3806    /*
3807     * configure the leaf levevl into binary buddy system
3808     *
3809     * Try to combine buddies starting with a buddy size of 1
3810     * (i.e. two leaves). At a buddy size of 1 two buddy leaves
3811     * can be combined if both buddies have a maximum free of l2min;
3812     * the combination will result in the left-most buddy leaf having
3813     * a maximum free of l2min+1.
3814     * After processing all buddies for a given size, process buddies
3815     * at the next higher buddy size (i.e. current size * 2) and
3816     * the next maximum free (current free + 1).
3817     * This continues until the maximum possible buddy combination
3818     * yields maximum free.
3819     */
3820    for (l2free = dtp->budmin, bsize = 1; l2free < l2max;
3821         l2free++, bsize = nextb) {
3822        /* get next buddy size == current buddy pair size */
3823        nextb = bsize << 1;
3824
3825        /* scan each adjacent buddy pair at current buddy size */
3826        for (i = 0, cp = tp + le32_to_cpu(dtp->leafidx);
3827             i < le32_to_cpu(dtp->nleafs);
3828             i += nextb, cp += nextb) {
3829            /* coalesce if both adjacent buddies are max free */
3830            if (*cp == l2free && *(cp + bsize) == l2free) {
3831                *cp = l2free + 1; /* left take right */
3832                *(cp + bsize) = -1; /* right give left */
3833            }
3834        }
3835    }
3836
3837    /*
3838     * bubble summary information of leaves up the tree.
3839     *
3840     * Starting at the leaf node level, the four nodes described by
3841     * the higher level parent node are compared for a maximum free and
3842     * this maximum becomes the value of the parent node.
3843     * when all lower level nodes are processed in this fashion then
3844     * move up to the next level (parent becomes a lower level node) and
3845     * continue the process for that level.
3846     */
3847    for (child = le32_to_cpu(dtp->leafidx),
3848         nparent = le32_to_cpu(dtp->nleafs) >> 2;
3849         nparent > 0; nparent >>= 2, child = parent) {
3850        /* get index of 1st node of parent level */
3851        parent = (child - 1) >> 2;
3852
3853        /* set the value of the parent node as the maximum
3854         * of the four nodes of the current level.
3855         */
3856        for (i = 0, cp = tp + child, cp1 = tp + parent;
3857             i < nparent; i++, cp += 4, cp1++)
3858            *cp1 = TREEMAX(cp);
3859    }
3860
3861    return (*tp);
3862}
3863
3864
3865/*
3866 * dbInitDmapCtl()
3867 *
3868 * function: initialize dmapctl page
3869 */
3870static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i)
3871{ /* start leaf index not covered by range */
3872    s8 *cp;
3873
3874    dcp->nleafs = cpu_to_le32(LPERCTL);
3875    dcp->l2nleafs = cpu_to_le32(L2LPERCTL);
3876    dcp->leafidx = cpu_to_le32(CTLLEAFIND);
3877    dcp->height = cpu_to_le32(5);
3878    dcp->budmin = L2BPERDMAP + L2LPERCTL * level;
3879
3880    /*
3881     * initialize the leaves of current level that were not covered
3882     * by the specified input block range (i.e. the leaves have no
3883     * low level dmapctl or dmap).
3884     */
3885    cp = &dcp->stree[CTLLEAFIND + i];
3886    for (; i < LPERCTL; i++)
3887        *cp++ = NOFREE;
3888
3889    /* build the dmap's binary buddy summary tree */
3890    return (dbInitTree((struct dmaptree *) dcp));
3891}
3892
3893
3894/*
3895 * NAME: dbGetL2AGSize()/ujfs_getagl2size()
3896 *
3897 * FUNCTION: Determine log2(allocation group size) from aggregate size
3898 *
3899 * PARAMETERS:
3900 * nblocks - Number of blocks in aggregate
3901 *
3902 * RETURNS: log2(allocation group size) in aggregate blocks
3903 */
3904static int dbGetL2AGSize(s64 nblocks)
3905{
3906    s64 sz;
3907    s64 m;
3908    int l2sz;
3909
3910    if (nblocks < BPERDMAP * MAXAG)
3911        return (L2BPERDMAP);
3912
3913    /* round up aggregate size to power of 2 */
3914    m = ((u64) 1 << (64 - 1));
3915    for (l2sz = 64; l2sz >= 0; l2sz--, m >>= 1) {
3916        if (m & nblocks)
3917            break;
3918    }
3919
3920    sz = (s64) 1 << l2sz;
3921    if (sz < nblocks)
3922        l2sz += 1;
3923
3924    /* agsize = roundupSize/max_number_of_ag */
3925    return (l2sz - L2MAXAG);
3926}
3927
3928
3929/*
3930 * NAME: dbMapFileSizeToMapSize()
3931 *
3932 * FUNCTION: compute number of blocks the block allocation map file
3933 * can cover from the map file size;
3934 *
3935 * RETURNS: Number of blocks which can be covered by this block map file;
3936 */
3937
3938/*
3939 * maximum number of map pages at each level including control pages
3940 */
3941#define MAXL0PAGES (1 + LPERCTL)
3942#define MAXL1PAGES (1 + LPERCTL * MAXL0PAGES)
3943#define MAXL2PAGES (1 + LPERCTL * MAXL1PAGES)
3944
3945/*
3946 * convert number of map pages to the zero origin top dmapctl level
3947 */
3948#define BMAPPGTOLEV(npages) \
3949    (((npages) <= 3 + MAXL0PAGES) ? 0 : \
3950     ((npages) <= 2 + MAXL1PAGES) ? 1 : 2)
3951
3952s64 dbMapFileSizeToMapSize(struct inode * ipbmap)
3953{
3954    struct super_block *sb = ipbmap->i_sb;
3955    s64 nblocks;
3956    s64 npages, ndmaps;
3957    int level, i;
3958    int complete, factor;
3959
3960    nblocks = ipbmap->i_size >> JFS_SBI(sb)->l2bsize;
3961    npages = nblocks >> JFS_SBI(sb)->l2nbperpage;
3962    level = BMAPPGTOLEV(npages);
3963
3964    /* At each level, accumulate the number of dmap pages covered by
3965     * the number of full child levels below it;
3966     * repeat for the last incomplete child level.
3967     */
3968    ndmaps = 0;
3969    npages--; /* skip the first global control page */
3970    /* skip higher level control pages above top level covered by map */
3971    npages -= (2 - level);
3972    npages--; /* skip top level's control page */
3973    for (i = level; i >= 0; i--) {
3974        factor =
3975            (i == 2) ? MAXL1PAGES : ((i == 1) ? MAXL0PAGES : 1);
3976        complete = (u32) npages / factor;
3977        ndmaps += complete * ((i == 2) ? LPERCTL * LPERCTL :
3978                      ((i == 1) ? LPERCTL : 1));
3979
3980        /* pages in last/incomplete child */
3981        npages = (u32) npages % factor;
3982        /* skip incomplete child's level control page */
3983        npages--;
3984    }
3985
3986    /* convert the number of dmaps into the number of blocks
3987     * which can be covered by the dmaps;
3988     */
3989    nblocks = ndmaps << L2BPERDMAP;
3990
3991    return (nblocks);
3992}
3993

Archive Download this file



interactive