Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Copyright (C) International Business Machines Corp., 2000-2004
4 * Portions Copyright (C) Tino Reichardt, 2012
5 */
6
7#include <linux/fs.h>
8#include <linux/slab.h>
9#include "jfs_incore.h"
10#include "jfs_superblock.h"
11#include "jfs_dmap.h"
12#include "jfs_imap.h"
13#include "jfs_lock.h"
14#include "jfs_metapage.h"
15#include "jfs_debug.h"
16#include "jfs_discard.h"
17
18/*
19 * SERIALIZATION of the Block Allocation Map.
20 *
21 * the working state of the block allocation map is accessed in
22 * two directions:
23 *
24 * 1) allocation and free requests that start at the dmap
25 * level and move up through the dmap control pages (i.e.
26 * the vast majority of requests).
27 *
28 * 2) allocation requests that start at dmap control page
29 * level and work down towards the dmaps.
30 *
31 * the serialization scheme used here is as follows.
32 *
33 * requests which start at the bottom are serialized against each
34 * other through buffers and each requests holds onto its buffers
35 * as it works it way up from a single dmap to the required level
36 * of dmap control page.
37 * requests that start at the top are serialized against each other
38 * and request that start from the bottom by the multiple read/single
39 * write inode lock of the bmap inode. requests starting at the top
40 * take this lock in write mode while request starting at the bottom
41 * take the lock in read mode. a single top-down request may proceed
42 * exclusively while multiple bottoms-up requests may proceed
43 * simultaneously (under the protection of busy buffers).
44 *
45 * in addition to information found in dmaps and dmap control pages,
46 * the working state of the block allocation map also includes read/
47 * write information maintained in the bmap descriptor (i.e. total
48 * free block count, allocation group level free block counts).
49 * a single exclusive lock (BMAP_LOCK) is used to guard this information
50 * in the face of multiple-bottoms up requests.
51 * (lock ordering: IREAD_LOCK, BMAP_LOCK);
52 *
53 * accesses to the persistent state of the block allocation map (limited
54 * to the persistent bitmaps in dmaps) is guarded by (busy) buffers.
55 */
56
57#define BMAP_LOCK_INIT(bmp) mutex_init(&bmp->db_bmaplock)
58#define BMAP_LOCK(bmp) mutex_lock(&bmp->db_bmaplock)
59#define BMAP_UNLOCK(bmp) mutex_unlock(&bmp->db_bmaplock)
60
61/*
62 * forward references
63 */
64static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
65 int nblocks);
66static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval);
67static int dbBackSplit(dmtree_t * tp, int leafno);
68static int dbJoin(dmtree_t * tp, int leafno, int newval);
69static void dbAdjTree(dmtree_t * tp, int leafno, int newval);
70static int dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc,
71 int level);
72static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results);
73static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
74 int nblocks);
75static int dbAllocNear(struct bmap * bmp, struct dmap * dp, s64 blkno,
76 int nblocks,
77 int l2nb, s64 * results);
78static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
79 int nblocks);
80static int dbAllocDmapLev(struct bmap * bmp, struct dmap * dp, int nblocks,
81 int l2nb,
82 s64 * results);
83static int dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb,
84 s64 * results);
85static int dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno,
86 s64 * results);
87static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks);
88static int dbFindBits(u32 word, int l2nb);
89static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno);
90static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx);
91static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
92 int nblocks);
93static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
94 int nblocks);
95static int dbMaxBud(u8 * cp);
96static int blkstol2(s64 nb);
97
98static int cntlz(u32 value);
99static int cnttz(u32 word);
100
101static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
102 int nblocks);
103static int dbInitDmap(struct dmap * dp, s64 blkno, int nblocks);
104static int dbInitDmapTree(struct dmap * dp);
105static int dbInitTree(struct dmaptree * dtp);
106static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i);
107static int dbGetL2AGSize(s64 nblocks);
108
109/*
110 * buddy table
111 *
112 * table used for determining buddy sizes within characters of
113 * dmap bitmap words. the characters themselves serve as indexes
114 * into the table, with the table elements yielding the maximum
115 * binary buddy of free bits within the character.
116 */
117static const s8 budtab[256] = {
118 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
119 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
120 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
121 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
122 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
123 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
124 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
125 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
126 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
127 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
128 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
129 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
130 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
131 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
132 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
133 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, -1
134};
135
136/*
137 * NAME: dbMount()
138 *
139 * FUNCTION: initializate the block allocation map.
140 *
141 * memory is allocated for the in-core bmap descriptor and
142 * the in-core descriptor is initialized from disk.
143 *
144 * PARAMETERS:
145 * ipbmap - pointer to in-core inode for the block map.
146 *
147 * RETURN VALUES:
148 * 0 - success
149 * -ENOMEM - insufficient memory
150 * -EIO - i/o error
151 */
152int dbMount(struct inode *ipbmap)
153{
154 struct bmap *bmp;
155 struct dbmap_disk *dbmp_le;
156 struct metapage *mp;
157 int i;
158
159 /*
160 * allocate/initialize the in-memory bmap descriptor
161 */
162 /* allocate memory for the in-memory bmap descriptor */
163 bmp = kmalloc(sizeof(struct bmap), GFP_KERNEL);
164 if (bmp == NULL)
165 return -ENOMEM;
166
167 /* read the on-disk bmap descriptor. */
168 mp = read_metapage(ipbmap,
169 BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
170 PSIZE, 0);
171 if (mp == NULL) {
172 kfree(bmp);
173 return -EIO;
174 }
175
176 /* copy the on-disk bmap descriptor to its in-memory version. */
177 dbmp_le = (struct dbmap_disk *) mp->data;
178 bmp->db_mapsize = le64_to_cpu(dbmp_le->dn_mapsize);
179 bmp->db_nfree = le64_to_cpu(dbmp_le->dn_nfree);
180 bmp->db_l2nbperpage = le32_to_cpu(dbmp_le->dn_l2nbperpage);
181 bmp->db_numag = le32_to_cpu(dbmp_le->dn_numag);
182 bmp->db_maxlevel = le32_to_cpu(dbmp_le->dn_maxlevel);
183 bmp->db_maxag = le32_to_cpu(dbmp_le->dn_maxag);
184 bmp->db_agpref = le32_to_cpu(dbmp_le->dn_agpref);
185 bmp->db_aglevel = le32_to_cpu(dbmp_le->dn_aglevel);
186 bmp->db_agheight = le32_to_cpu(dbmp_le->dn_agheight);
187 bmp->db_agwidth = le32_to_cpu(dbmp_le->dn_agwidth);
188 bmp->db_agstart = le32_to_cpu(dbmp_le->dn_agstart);
189 bmp->db_agl2size = le32_to_cpu(dbmp_le->dn_agl2size);
190 for (i = 0; i < MAXAG; i++)
191 bmp->db_agfree[i] = le64_to_cpu(dbmp_le->dn_agfree[i]);
192 bmp->db_agsize = le64_to_cpu(dbmp_le->dn_agsize);
193 bmp->db_maxfreebud = dbmp_le->dn_maxfreebud;
194
195 /* release the buffer. */
196 release_metapage(mp);
197
198 /* bind the bmap inode and the bmap descriptor to each other. */
199 bmp->db_ipbmap = ipbmap;
200 JFS_SBI(ipbmap->i_sb)->bmap = bmp;
201
202 memset(bmp->db_active, 0, sizeof(bmp->db_active));
203
204 /*
205 * allocate/initialize the bmap lock
206 */
207 BMAP_LOCK_INIT(bmp);
208
209 return (0);
210}
211
212
213/*
214 * NAME: dbUnmount()
215 *
216 * FUNCTION: terminate the block allocation map in preparation for
217 * file system unmount.
218 *
219 * the in-core bmap descriptor is written to disk and
220 * the memory for this descriptor is freed.
221 *
222 * PARAMETERS:
223 * ipbmap - pointer to in-core inode for the block map.
224 *
225 * RETURN VALUES:
226 * 0 - success
227 * -EIO - i/o error
228 */
229int dbUnmount(struct inode *ipbmap, int mounterror)
230{
231 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
232
233 if (!(mounterror || isReadOnly(ipbmap)))
234 dbSync(ipbmap);
235
236 /*
237 * Invalidate the page cache buffers
238 */
239 truncate_inode_pages(ipbmap->i_mapping, 0);
240
241 /* free the memory for the in-memory bmap. */
242 kfree(bmp);
243
244 return (0);
245}
246
247/*
248 * dbSync()
249 */
250int dbSync(struct inode *ipbmap)
251{
252 struct dbmap_disk *dbmp_le;
253 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
254 struct metapage *mp;
255 int i;
256
257 /*
258 * write bmap global control page
259 */
260 /* get the buffer for the on-disk bmap descriptor. */
261 mp = read_metapage(ipbmap,
262 BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
263 PSIZE, 0);
264 if (mp == NULL) {
265 jfs_err("dbSync: read_metapage failed!");
266 return -EIO;
267 }
268 /* copy the in-memory version of the bmap to the on-disk version */
269 dbmp_le = (struct dbmap_disk *) mp->data;
270 dbmp_le->dn_mapsize = cpu_to_le64(bmp->db_mapsize);
271 dbmp_le->dn_nfree = cpu_to_le64(bmp->db_nfree);
272 dbmp_le->dn_l2nbperpage = cpu_to_le32(bmp->db_l2nbperpage);
273 dbmp_le->dn_numag = cpu_to_le32(bmp->db_numag);
274 dbmp_le->dn_maxlevel = cpu_to_le32(bmp->db_maxlevel);
275 dbmp_le->dn_maxag = cpu_to_le32(bmp->db_maxag);
276 dbmp_le->dn_agpref = cpu_to_le32(bmp->db_agpref);
277 dbmp_le->dn_aglevel = cpu_to_le32(bmp->db_aglevel);
278 dbmp_le->dn_agheight = cpu_to_le32(bmp->db_agheight);
279 dbmp_le->dn_agwidth = cpu_to_le32(bmp->db_agwidth);
280 dbmp_le->dn_agstart = cpu_to_le32(bmp->db_agstart);
281 dbmp_le->dn_agl2size = cpu_to_le32(bmp->db_agl2size);
282 for (i = 0; i < MAXAG; i++)
283 dbmp_le->dn_agfree[i] = cpu_to_le64(bmp->db_agfree[i]);
284 dbmp_le->dn_agsize = cpu_to_le64(bmp->db_agsize);
285 dbmp_le->dn_maxfreebud = bmp->db_maxfreebud;
286
287 /* write the buffer */
288 write_metapage(mp);
289
290 /*
291 * write out dirty pages of bmap
292 */
293 filemap_write_and_wait(ipbmap->i_mapping);
294
295 diWriteSpecial(ipbmap, 0);
296
297 return (0);
298}
299
300/*
301 * NAME: dbFree()
302 *
303 * FUNCTION: free the specified block range from the working block
304 * allocation map.
305 *
306 * the blocks will be free from the working map one dmap
307 * at a time.
308 *
309 * PARAMETERS:
310 * ip - pointer to in-core inode;
311 * blkno - starting block number to be freed.
312 * nblocks - number of blocks to be freed.
313 *
314 * RETURN VALUES:
315 * 0 - success
316 * -EIO - i/o error
317 */
318int dbFree(struct inode *ip, s64 blkno, s64 nblocks)
319{
320 struct metapage *mp;
321 struct dmap *dp;
322 int nb, rc;
323 s64 lblkno, rem;
324 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
325 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
326 struct super_block *sb = ipbmap->i_sb;
327
328 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
329
330 /* block to be freed better be within the mapsize. */
331 if (unlikely((blkno == 0) || (blkno + nblocks > bmp->db_mapsize))) {
332 IREAD_UNLOCK(ipbmap);
333 printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
334 (unsigned long long) blkno,
335 (unsigned long long) nblocks);
336 jfs_error(ip->i_sb, "block to be freed is outside the map\n");
337 return -EIO;
338 }
339
340 /**
341 * TRIM the blocks, when mounted with discard option
342 */
343 if (JFS_SBI(sb)->flag & JFS_DISCARD)
344 if (JFS_SBI(sb)->minblks_trim <= nblocks)
345 jfs_issue_discard(ipbmap, blkno, nblocks);
346
347 /*
348 * free the blocks a dmap at a time.
349 */
350 mp = NULL;
351 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
352 /* release previous dmap if any */
353 if (mp) {
354 write_metapage(mp);
355 }
356
357 /* get the buffer for the current dmap. */
358 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
359 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
360 if (mp == NULL) {
361 IREAD_UNLOCK(ipbmap);
362 return -EIO;
363 }
364 dp = (struct dmap *) mp->data;
365
366 /* determine the number of blocks to be freed from
367 * this dmap.
368 */
369 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
370
371 /* free the blocks. */
372 if ((rc = dbFreeDmap(bmp, dp, blkno, nb))) {
373 jfs_error(ip->i_sb, "error in block map\n");
374 release_metapage(mp);
375 IREAD_UNLOCK(ipbmap);
376 return (rc);
377 }
378 }
379
380 /* write the last buffer. */
381 write_metapage(mp);
382
383 IREAD_UNLOCK(ipbmap);
384
385 return (0);
386}
387
388
389/*
390 * NAME: dbUpdatePMap()
391 *
392 * FUNCTION: update the allocation state (free or allocate) of the
393 * specified block range in the persistent block allocation map.
394 *
395 * the blocks will be updated in the persistent map one
396 * dmap at a time.
397 *
398 * PARAMETERS:
399 * ipbmap - pointer to in-core inode for the block map.
400 * free - 'true' if block range is to be freed from the persistent
401 * map; 'false' if it is to be allocated.
402 * blkno - starting block number of the range.
403 * nblocks - number of contiguous blocks in the range.
404 * tblk - transaction block;
405 *
406 * RETURN VALUES:
407 * 0 - success
408 * -EIO - i/o error
409 */
410int
411dbUpdatePMap(struct inode *ipbmap,
412 int free, s64 blkno, s64 nblocks, struct tblock * tblk)
413{
414 int nblks, dbitno, wbitno, rbits;
415 int word, nbits, nwords;
416 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
417 s64 lblkno, rem, lastlblkno;
418 u32 mask;
419 struct dmap *dp;
420 struct metapage *mp;
421 struct jfs_log *log;
422 int lsn, difft, diffp;
423 unsigned long flags;
424
425 /* the blocks better be within the mapsize. */
426 if (blkno + nblocks > bmp->db_mapsize) {
427 printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
428 (unsigned long long) blkno,
429 (unsigned long long) nblocks);
430 jfs_error(ipbmap->i_sb, "blocks are outside the map\n");
431 return -EIO;
432 }
433
434 /* compute delta of transaction lsn from log syncpt */
435 lsn = tblk->lsn;
436 log = (struct jfs_log *) JFS_SBI(tblk->sb)->log;
437 logdiff(difft, lsn, log);
438
439 /*
440 * update the block state a dmap at a time.
441 */
442 mp = NULL;
443 lastlblkno = 0;
444 for (rem = nblocks; rem > 0; rem -= nblks, blkno += nblks) {
445 /* get the buffer for the current dmap. */
446 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
447 if (lblkno != lastlblkno) {
448 if (mp) {
449 write_metapage(mp);
450 }
451
452 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE,
453 0);
454 if (mp == NULL)
455 return -EIO;
456 metapage_wait_for_io(mp);
457 }
458 dp = (struct dmap *) mp->data;
459
460 /* determine the bit number and word within the dmap of
461 * the starting block. also determine how many blocks
462 * are to be updated within this dmap.
463 */
464 dbitno = blkno & (BPERDMAP - 1);
465 word = dbitno >> L2DBWORD;
466 nblks = min(rem, (s64)BPERDMAP - dbitno);
467
468 /* update the bits of the dmap words. the first and last
469 * words may only have a subset of their bits updated. if
470 * this is the case, we'll work against that word (i.e.
471 * partial first and/or last) only in a single pass. a
472 * single pass will also be used to update all words that
473 * are to have all their bits updated.
474 */
475 for (rbits = nblks; rbits > 0;
476 rbits -= nbits, dbitno += nbits) {
477 /* determine the bit number within the word and
478 * the number of bits within the word.
479 */
480 wbitno = dbitno & (DBWORD - 1);
481 nbits = min(rbits, DBWORD - wbitno);
482
483 /* check if only part of the word is to be updated. */
484 if (nbits < DBWORD) {
485 /* update (free or allocate) the bits
486 * in this word.
487 */
488 mask =
489 (ONES << (DBWORD - nbits) >> wbitno);
490 if (free)
491 dp->pmap[word] &=
492 cpu_to_le32(~mask);
493 else
494 dp->pmap[word] |=
495 cpu_to_le32(mask);
496
497 word += 1;
498 } else {
499 /* one or more words are to have all
500 * their bits updated. determine how
501 * many words and how many bits.
502 */
503 nwords = rbits >> L2DBWORD;
504 nbits = nwords << L2DBWORD;
505
506 /* update (free or allocate) the bits
507 * in these words.
508 */
509 if (free)
510 memset(&dp->pmap[word], 0,
511 nwords * 4);
512 else
513 memset(&dp->pmap[word], (int) ONES,
514 nwords * 4);
515
516 word += nwords;
517 }
518 }
519
520 /*
521 * update dmap lsn
522 */
523 if (lblkno == lastlblkno)
524 continue;
525
526 lastlblkno = lblkno;
527
528 LOGSYNC_LOCK(log, flags);
529 if (mp->lsn != 0) {
530 /* inherit older/smaller lsn */
531 logdiff(diffp, mp->lsn, log);
532 if (difft < diffp) {
533 mp->lsn = lsn;
534
535 /* move bp after tblock in logsync list */
536 list_move(&mp->synclist, &tblk->synclist);
537 }
538
539 /* inherit younger/larger clsn */
540 logdiff(difft, tblk->clsn, log);
541 logdiff(diffp, mp->clsn, log);
542 if (difft > diffp)
543 mp->clsn = tblk->clsn;
544 } else {
545 mp->log = log;
546 mp->lsn = lsn;
547
548 /* insert bp after tblock in logsync list */
549 log->count++;
550 list_add(&mp->synclist, &tblk->synclist);
551
552 mp->clsn = tblk->clsn;
553 }
554 LOGSYNC_UNLOCK(log, flags);
555 }
556
557 /* write the last buffer. */
558 if (mp) {
559 write_metapage(mp);
560 }
561
562 return (0);
563}
564
565
566/*
567 * NAME: dbNextAG()
568 *
569 * FUNCTION: find the preferred allocation group for new allocations.
570 *
571 * Within the allocation groups, we maintain a preferred
572 * allocation group which consists of a group with at least
573 * average free space. It is the preferred group that we target
574 * new inode allocation towards. The tie-in between inode
575 * allocation and block allocation occurs as we allocate the
576 * first (data) block of an inode and specify the inode (block)
577 * as the allocation hint for this block.
578 *
579 * We try to avoid having more than one open file growing in
580 * an allocation group, as this will lead to fragmentation.
581 * This differs from the old OS/2 method of trying to keep
582 * empty ags around for large allocations.
583 *
584 * PARAMETERS:
585 * ipbmap - pointer to in-core inode for the block map.
586 *
587 * RETURN VALUES:
588 * the preferred allocation group number.
589 */
590int dbNextAG(struct inode *ipbmap)
591{
592 s64 avgfree;
593 int agpref;
594 s64 hwm = 0;
595 int i;
596 int next_best = -1;
597 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
598
599 BMAP_LOCK(bmp);
600
601 /* determine the average number of free blocks within the ags. */
602 avgfree = (u32)bmp->db_nfree / bmp->db_numag;
603
604 /*
605 * if the current preferred ag does not have an active allocator
606 * and has at least average freespace, return it
607 */
608 agpref = bmp->db_agpref;
609 if ((atomic_read(&bmp->db_active[agpref]) == 0) &&
610 (bmp->db_agfree[agpref] >= avgfree))
611 goto unlock;
612
613 /* From the last preferred ag, find the next one with at least
614 * average free space.
615 */
616 for (i = 0 ; i < bmp->db_numag; i++, agpref++) {
617 if (agpref == bmp->db_numag)
618 agpref = 0;
619
620 if (atomic_read(&bmp->db_active[agpref]))
621 /* open file is currently growing in this ag */
622 continue;
623 if (bmp->db_agfree[agpref] >= avgfree) {
624 /* Return this one */
625 bmp->db_agpref = agpref;
626 goto unlock;
627 } else if (bmp->db_agfree[agpref] > hwm) {
628 /* Less than avg. freespace, but best so far */
629 hwm = bmp->db_agfree[agpref];
630 next_best = agpref;
631 }
632 }
633
634 /*
635 * If no inactive ag was found with average freespace, use the
636 * next best
637 */
638 if (next_best != -1)
639 bmp->db_agpref = next_best;
640 /* else leave db_agpref unchanged */
641unlock:
642 BMAP_UNLOCK(bmp);
643
644 /* return the preferred group.
645 */
646 return (bmp->db_agpref);
647}
648
649/*
650 * NAME: dbAlloc()
651 *
652 * FUNCTION: attempt to allocate a specified number of contiguous free
653 * blocks from the working allocation block map.
654 *
655 * the block allocation policy uses hints and a multi-step
656 * approach.
657 *
658 * for allocation requests smaller than the number of blocks
659 * per dmap, we first try to allocate the new blocks
660 * immediately following the hint. if these blocks are not
661 * available, we try to allocate blocks near the hint. if
662 * no blocks near the hint are available, we next try to
663 * allocate within the same dmap as contains the hint.
664 *
665 * if no blocks are available in the dmap or the allocation
666 * request is larger than the dmap size, we try to allocate
667 * within the same allocation group as contains the hint. if
668 * this does not succeed, we finally try to allocate anywhere
669 * within the aggregate.
670 *
671 * we also try to allocate anywhere within the aggregate for
672 * for allocation requests larger than the allocation group
673 * size or requests that specify no hint value.
674 *
675 * PARAMETERS:
676 * ip - pointer to in-core inode;
677 * hint - allocation hint.
678 * nblocks - number of contiguous blocks in the range.
679 * results - on successful return, set to the starting block number
680 * of the newly allocated contiguous range.
681 *
682 * RETURN VALUES:
683 * 0 - success
684 * -ENOSPC - insufficient disk resources
685 * -EIO - i/o error
686 */
687int dbAlloc(struct inode *ip, s64 hint, s64 nblocks, s64 * results)
688{
689 int rc, agno;
690 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
691 struct bmap *bmp;
692 struct metapage *mp;
693 s64 lblkno, blkno;
694 struct dmap *dp;
695 int l2nb;
696 s64 mapSize;
697 int writers;
698
699 /* assert that nblocks is valid */
700 assert(nblocks > 0);
701
702 /* get the log2 number of blocks to be allocated.
703 * if the number of blocks is not a log2 multiple,
704 * it will be rounded up to the next log2 multiple.
705 */
706 l2nb = BLKSTOL2(nblocks);
707
708 bmp = JFS_SBI(ip->i_sb)->bmap;
709
710 mapSize = bmp->db_mapsize;
711
712 /* the hint should be within the map */
713 if (hint >= mapSize) {
714 jfs_error(ip->i_sb, "the hint is outside the map\n");
715 return -EIO;
716 }
717
718 /* if the number of blocks to be allocated is greater than the
719 * allocation group size, try to allocate anywhere.
720 */
721 if (l2nb > bmp->db_agl2size) {
722 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
723
724 rc = dbAllocAny(bmp, nblocks, l2nb, results);
725
726 goto write_unlock;
727 }
728
729 /*
730 * If no hint, let dbNextAG recommend an allocation group
731 */
732 if (hint == 0)
733 goto pref_ag;
734
735 /* we would like to allocate close to the hint. adjust the
736 * hint to the block following the hint since the allocators
737 * will start looking for free space starting at this point.
738 */
739 blkno = hint + 1;
740
741 if (blkno >= bmp->db_mapsize)
742 goto pref_ag;
743
744 agno = blkno >> bmp->db_agl2size;
745
746 /* check if blkno crosses over into a new allocation group.
747 * if so, check if we should allow allocations within this
748 * allocation group.
749 */
750 if ((blkno & (bmp->db_agsize - 1)) == 0)
751 /* check if the AG is currently being written to.
752 * if so, call dbNextAG() to find a non-busy
753 * AG with sufficient free space.
754 */
755 if (atomic_read(&bmp->db_active[agno]))
756 goto pref_ag;
757
758 /* check if the allocation request size can be satisfied from a
759 * single dmap. if so, try to allocate from the dmap containing
760 * the hint using a tiered strategy.
761 */
762 if (nblocks <= BPERDMAP) {
763 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
764
765 /* get the buffer for the dmap containing the hint.
766 */
767 rc = -EIO;
768 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
769 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
770 if (mp == NULL)
771 goto read_unlock;
772
773 dp = (struct dmap *) mp->data;
774
775 /* first, try to satisfy the allocation request with the
776 * blocks beginning at the hint.
777 */
778 if ((rc = dbAllocNext(bmp, dp, blkno, (int) nblocks))
779 != -ENOSPC) {
780 if (rc == 0) {
781 *results = blkno;
782 mark_metapage_dirty(mp);
783 }
784
785 release_metapage(mp);
786 goto read_unlock;
787 }
788
789 writers = atomic_read(&bmp->db_active[agno]);
790 if ((writers > 1) ||
791 ((writers == 1) && (JFS_IP(ip)->active_ag != agno))) {
792 /*
793 * Someone else is writing in this allocation
794 * group. To avoid fragmenting, try another ag
795 */
796 release_metapage(mp);
797 IREAD_UNLOCK(ipbmap);
798 goto pref_ag;
799 }
800
801 /* next, try to satisfy the allocation request with blocks
802 * near the hint.
803 */
804 if ((rc =
805 dbAllocNear(bmp, dp, blkno, (int) nblocks, l2nb, results))
806 != -ENOSPC) {
807 if (rc == 0)
808 mark_metapage_dirty(mp);
809
810 release_metapage(mp);
811 goto read_unlock;
812 }
813
814 /* try to satisfy the allocation request with blocks within
815 * the same dmap as the hint.
816 */
817 if ((rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results))
818 != -ENOSPC) {
819 if (rc == 0)
820 mark_metapage_dirty(mp);
821
822 release_metapage(mp);
823 goto read_unlock;
824 }
825
826 release_metapage(mp);
827 IREAD_UNLOCK(ipbmap);
828 }
829
830 /* try to satisfy the allocation request with blocks within
831 * the same allocation group as the hint.
832 */
833 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
834 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) != -ENOSPC)
835 goto write_unlock;
836
837 IWRITE_UNLOCK(ipbmap);
838
839
840 pref_ag:
841 /*
842 * Let dbNextAG recommend a preferred allocation group
843 */
844 agno = dbNextAG(ipbmap);
845 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
846
847 /* Try to allocate within this allocation group. if that fails, try to
848 * allocate anywhere in the map.
849 */
850 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) == -ENOSPC)
851 rc = dbAllocAny(bmp, nblocks, l2nb, results);
852
853 write_unlock:
854 IWRITE_UNLOCK(ipbmap);
855
856 return (rc);
857
858 read_unlock:
859 IREAD_UNLOCK(ipbmap);
860
861 return (rc);
862}
863
864#ifdef _NOTYET
865/*
866 * NAME: dbAllocExact()
867 *
868 * FUNCTION: try to allocate the requested extent;
869 *
870 * PARAMETERS:
871 * ip - pointer to in-core inode;
872 * blkno - extent address;
873 * nblocks - extent length;
874 *
875 * RETURN VALUES:
876 * 0 - success
877 * -ENOSPC - insufficient disk resources
878 * -EIO - i/o error
879 */
880int dbAllocExact(struct inode *ip, s64 blkno, int nblocks)
881{
882 int rc;
883 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
884 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
885 struct dmap *dp;
886 s64 lblkno;
887 struct metapage *mp;
888
889 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
890
891 /*
892 * validate extent request:
893 *
894 * note: defragfs policy:
895 * max 64 blocks will be moved.
896 * allocation request size must be satisfied from a single dmap.
897 */
898 if (nblocks <= 0 || nblocks > BPERDMAP || blkno >= bmp->db_mapsize) {
899 IREAD_UNLOCK(ipbmap);
900 return -EINVAL;
901 }
902
903 if (nblocks > ((s64) 1 << bmp->db_maxfreebud)) {
904 /* the free space is no longer available */
905 IREAD_UNLOCK(ipbmap);
906 return -ENOSPC;
907 }
908
909 /* read in the dmap covering the extent */
910 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
911 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
912 if (mp == NULL) {
913 IREAD_UNLOCK(ipbmap);
914 return -EIO;
915 }
916 dp = (struct dmap *) mp->data;
917
918 /* try to allocate the requested extent */
919 rc = dbAllocNext(bmp, dp, blkno, nblocks);
920
921 IREAD_UNLOCK(ipbmap);
922
923 if (rc == 0)
924 mark_metapage_dirty(mp);
925
926 release_metapage(mp);
927
928 return (rc);
929}
930#endif /* _NOTYET */
931
932/*
933 * NAME: dbReAlloc()
934 *
935 * FUNCTION: attempt to extend a current allocation by a specified
936 * number of blocks.
937 *
938 * this routine attempts to satisfy the allocation request
939 * by first trying to extend the existing allocation in
940 * place by allocating the additional blocks as the blocks
941 * immediately following the current allocation. if these
942 * blocks are not available, this routine will attempt to
943 * allocate a new set of contiguous blocks large enough
944 * to cover the existing allocation plus the additional
945 * number of blocks required.
946 *
947 * PARAMETERS:
948 * ip - pointer to in-core inode requiring allocation.
949 * blkno - starting block of the current allocation.
950 * nblocks - number of contiguous blocks within the current
951 * allocation.
952 * addnblocks - number of blocks to add to the allocation.
953 * results - on successful return, set to the starting block number
954 * of the existing allocation if the existing allocation
955 * was extended in place or to a newly allocated contiguous
956 * range if the existing allocation could not be extended
957 * in place.
958 *
959 * RETURN VALUES:
960 * 0 - success
961 * -ENOSPC - insufficient disk resources
962 * -EIO - i/o error
963 */
964int
965dbReAlloc(struct inode *ip,
966 s64 blkno, s64 nblocks, s64 addnblocks, s64 * results)
967{
968 int rc;
969
970 /* try to extend the allocation in place.
971 */
972 if ((rc = dbExtend(ip, blkno, nblocks, addnblocks)) == 0) {
973 *results = blkno;
974 return (0);
975 } else {
976 if (rc != -ENOSPC)
977 return (rc);
978 }
979
980 /* could not extend the allocation in place, so allocate a
981 * new set of blocks for the entire request (i.e. try to get
982 * a range of contiguous blocks large enough to cover the
983 * existing allocation plus the additional blocks.)
984 */
985 return (dbAlloc
986 (ip, blkno + nblocks - 1, addnblocks + nblocks, results));
987}
988
989
990/*
991 * NAME: dbExtend()
992 *
993 * FUNCTION: attempt to extend a current allocation by a specified
994 * number of blocks.
995 *
996 * this routine attempts to satisfy the allocation request
997 * by first trying to extend the existing allocation in
998 * place by allocating the additional blocks as the blocks
999 * immediately following the current allocation.
1000 *
1001 * PARAMETERS:
1002 * ip - pointer to in-core inode requiring allocation.
1003 * blkno - starting block of the current allocation.
1004 * nblocks - number of contiguous blocks within the current
1005 * allocation.
1006 * addnblocks - number of blocks to add to the allocation.
1007 *
1008 * RETURN VALUES:
1009 * 0 - success
1010 * -ENOSPC - insufficient disk resources
1011 * -EIO - i/o error
1012 */
1013static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks)
1014{
1015 struct jfs_sb_info *sbi = JFS_SBI(ip->i_sb);
1016 s64 lblkno, lastblkno, extblkno;
1017 uint rel_block;
1018 struct metapage *mp;
1019 struct dmap *dp;
1020 int rc;
1021 struct inode *ipbmap = sbi->ipbmap;
1022 struct bmap *bmp;
1023
1024 /*
1025 * We don't want a non-aligned extent to cross a page boundary
1026 */
1027 if (((rel_block = blkno & (sbi->nbperpage - 1))) &&
1028 (rel_block + nblocks + addnblocks > sbi->nbperpage))
1029 return -ENOSPC;
1030
1031 /* get the last block of the current allocation */
1032 lastblkno = blkno + nblocks - 1;
1033
1034 /* determine the block number of the block following
1035 * the existing allocation.
1036 */
1037 extblkno = lastblkno + 1;
1038
1039 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
1040
1041 /* better be within the file system */
1042 bmp = sbi->bmap;
1043 if (lastblkno < 0 || lastblkno >= bmp->db_mapsize) {
1044 IREAD_UNLOCK(ipbmap);
1045 jfs_error(ip->i_sb, "the block is outside the filesystem\n");
1046 return -EIO;
1047 }
1048
1049 /* we'll attempt to extend the current allocation in place by
1050 * allocating the additional blocks as the blocks immediately
1051 * following the current allocation. we only try to extend the
1052 * current allocation in place if the number of additional blocks
1053 * can fit into a dmap, the last block of the current allocation
1054 * is not the last block of the file system, and the start of the
1055 * inplace extension is not on an allocation group boundary.
1056 */
1057 if (addnblocks > BPERDMAP || extblkno >= bmp->db_mapsize ||
1058 (extblkno & (bmp->db_agsize - 1)) == 0) {
1059 IREAD_UNLOCK(ipbmap);
1060 return -ENOSPC;
1061 }
1062
1063 /* get the buffer for the dmap containing the first block
1064 * of the extension.
1065 */
1066 lblkno = BLKTODMAP(extblkno, bmp->db_l2nbperpage);
1067 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
1068 if (mp == NULL) {
1069 IREAD_UNLOCK(ipbmap);
1070 return -EIO;
1071 }
1072
1073 dp = (struct dmap *) mp->data;
1074
1075 /* try to allocate the blocks immediately following the
1076 * current allocation.
1077 */
1078 rc = dbAllocNext(bmp, dp, extblkno, (int) addnblocks);
1079
1080 IREAD_UNLOCK(ipbmap);
1081
1082 /* were we successful ? */
1083 if (rc == 0)
1084 write_metapage(mp);
1085 else
1086 /* we were not successful */
1087 release_metapage(mp);
1088
1089 return (rc);
1090}
1091
1092
1093/*
1094 * NAME: dbAllocNext()
1095 *
1096 * FUNCTION: attempt to allocate the blocks of the specified block
1097 * range within a dmap.
1098 *
1099 * PARAMETERS:
1100 * bmp - pointer to bmap descriptor
1101 * dp - pointer to dmap.
1102 * blkno - starting block number of the range.
1103 * nblocks - number of contiguous free blocks of the range.
1104 *
1105 * RETURN VALUES:
1106 * 0 - success
1107 * -ENOSPC - insufficient disk resources
1108 * -EIO - i/o error
1109 *
1110 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1111 */
1112static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
1113 int nblocks)
1114{
1115 int dbitno, word, rembits, nb, nwords, wbitno, nw;
1116 int l2size;
1117 s8 *leaf;
1118 u32 mask;
1119
1120 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1121 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1122 return -EIO;
1123 }
1124
1125 /* pick up a pointer to the leaves of the dmap tree.
1126 */
1127 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1128
1129 /* determine the bit number and word within the dmap of the
1130 * starting block.
1131 */
1132 dbitno = blkno & (BPERDMAP - 1);
1133 word = dbitno >> L2DBWORD;
1134
1135 /* check if the specified block range is contained within
1136 * this dmap.
1137 */
1138 if (dbitno + nblocks > BPERDMAP)
1139 return -ENOSPC;
1140
1141 /* check if the starting leaf indicates that anything
1142 * is free.
1143 */
1144 if (leaf[word] == NOFREE)
1145 return -ENOSPC;
1146
1147 /* check the dmaps words corresponding to block range to see
1148 * if the block range is free. not all bits of the first and
1149 * last words may be contained within the block range. if this
1150 * is the case, we'll work against those words (i.e. partial first
1151 * and/or last) on an individual basis (a single pass) and examine
1152 * the actual bits to determine if they are free. a single pass
1153 * will be used for all dmap words fully contained within the
1154 * specified range. within this pass, the leaves of the dmap
1155 * tree will be examined to determine if the blocks are free. a
1156 * single leaf may describe the free space of multiple dmap
1157 * words, so we may visit only a subset of the actual leaves
1158 * corresponding to the dmap words of the block range.
1159 */
1160 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
1161 /* determine the bit number within the word and
1162 * the number of bits within the word.
1163 */
1164 wbitno = dbitno & (DBWORD - 1);
1165 nb = min(rembits, DBWORD - wbitno);
1166
1167 /* check if only part of the word is to be examined.
1168 */
1169 if (nb < DBWORD) {
1170 /* check if the bits are free.
1171 */
1172 mask = (ONES << (DBWORD - nb) >> wbitno);
1173 if ((mask & ~le32_to_cpu(dp->wmap[word])) != mask)
1174 return -ENOSPC;
1175
1176 word += 1;
1177 } else {
1178 /* one or more dmap words are fully contained
1179 * within the block range. determine how many
1180 * words and how many bits.
1181 */
1182 nwords = rembits >> L2DBWORD;
1183 nb = nwords << L2DBWORD;
1184
1185 /* now examine the appropriate leaves to determine
1186 * if the blocks are free.
1187 */
1188 while (nwords > 0) {
1189 /* does the leaf describe any free space ?
1190 */
1191 if (leaf[word] < BUDMIN)
1192 return -ENOSPC;
1193
1194 /* determine the l2 number of bits provided
1195 * by this leaf.
1196 */
1197 l2size =
1198 min_t(int, leaf[word], NLSTOL2BSZ(nwords));
1199
1200 /* determine how many words were handled.
1201 */
1202 nw = BUDSIZE(l2size, BUDMIN);
1203
1204 nwords -= nw;
1205 word += nw;
1206 }
1207 }
1208 }
1209
1210 /* allocate the blocks.
1211 */
1212 return (dbAllocDmap(bmp, dp, blkno, nblocks));
1213}
1214
1215
1216/*
1217 * NAME: dbAllocNear()
1218 *
1219 * FUNCTION: attempt to allocate a number of contiguous free blocks near
1220 * a specified block (hint) within a dmap.
1221 *
1222 * starting with the dmap leaf that covers the hint, we'll
1223 * check the next four contiguous leaves for sufficient free
1224 * space. if sufficient free space is found, we'll allocate
1225 * the desired free space.
1226 *
1227 * PARAMETERS:
1228 * bmp - pointer to bmap descriptor
1229 * dp - pointer to dmap.
1230 * blkno - block number to allocate near.
1231 * nblocks - actual number of contiguous free blocks desired.
1232 * l2nb - log2 number of contiguous free blocks desired.
1233 * results - on successful return, set to the starting block number
1234 * of the newly allocated range.
1235 *
1236 * RETURN VALUES:
1237 * 0 - success
1238 * -ENOSPC - insufficient disk resources
1239 * -EIO - i/o error
1240 *
1241 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1242 */
1243static int
1244dbAllocNear(struct bmap * bmp,
1245 struct dmap * dp, s64 blkno, int nblocks, int l2nb, s64 * results)
1246{
1247 int word, lword, rc;
1248 s8 *leaf;
1249
1250 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1251 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1252 return -EIO;
1253 }
1254
1255 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1256
1257 /* determine the word within the dmap that holds the hint
1258 * (i.e. blkno). also, determine the last word in the dmap
1259 * that we'll include in our examination.
1260 */
1261 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
1262 lword = min(word + 4, LPERDMAP);
1263
1264 /* examine the leaves for sufficient free space.
1265 */
1266 for (; word < lword; word++) {
1267 /* does the leaf describe sufficient free space ?
1268 */
1269 if (leaf[word] < l2nb)
1270 continue;
1271
1272 /* determine the block number within the file system
1273 * of the first block described by this dmap word.
1274 */
1275 blkno = le64_to_cpu(dp->start) + (word << L2DBWORD);
1276
1277 /* if not all bits of the dmap word are free, get the
1278 * starting bit number within the dmap word of the required
1279 * string of free bits and adjust the block number with the
1280 * value.
1281 */
1282 if (leaf[word] < BUDMIN)
1283 blkno +=
1284 dbFindBits(le32_to_cpu(dp->wmap[word]), l2nb);
1285
1286 /* allocate the blocks.
1287 */
1288 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1289 *results = blkno;
1290
1291 return (rc);
1292 }
1293
1294 return -ENOSPC;
1295}
1296
1297
1298/*
1299 * NAME: dbAllocAG()
1300 *
1301 * FUNCTION: attempt to allocate the specified number of contiguous
1302 * free blocks within the specified allocation group.
1303 *
1304 * unless the allocation group size is equal to the number
1305 * of blocks per dmap, the dmap control pages will be used to
1306 * find the required free space, if available. we start the
1307 * search at the highest dmap control page level which
1308 * distinctly describes the allocation group's free space
1309 * (i.e. the highest level at which the allocation group's
1310 * free space is not mixed in with that of any other group).
1311 * in addition, we start the search within this level at a
1312 * height of the dmapctl dmtree at which the nodes distinctly
1313 * describe the allocation group's free space. at this height,
1314 * the allocation group's free space may be represented by 1
1315 * or two sub-trees, depending on the allocation group size.
1316 * we search the top nodes of these subtrees left to right for
1317 * sufficient free space. if sufficient free space is found,
1318 * the subtree is searched to find the leftmost leaf that
1319 * has free space. once we have made it to the leaf, we
1320 * move the search to the next lower level dmap control page
1321 * corresponding to this leaf. we continue down the dmap control
1322 * pages until we find the dmap that contains or starts the
1323 * sufficient free space and we allocate at this dmap.
1324 *
1325 * if the allocation group size is equal to the dmap size,
1326 * we'll start at the dmap corresponding to the allocation
1327 * group and attempt the allocation at this level.
1328 *
1329 * the dmap control page search is also not performed if the
1330 * allocation group is completely free and we go to the first
1331 * dmap of the allocation group to do the allocation. this is
1332 * done because the allocation group may be part (not the first
1333 * part) of a larger binary buddy system, causing the dmap
1334 * control pages to indicate no free space (NOFREE) within
1335 * the allocation group.
1336 *
1337 * PARAMETERS:
1338 * bmp - pointer to bmap descriptor
1339 * agno - allocation group number.
1340 * nblocks - actual number of contiguous free blocks desired.
1341 * l2nb - log2 number of contiguous free blocks desired.
1342 * results - on successful return, set to the starting block number
1343 * of the newly allocated range.
1344 *
1345 * RETURN VALUES:
1346 * 0 - success
1347 * -ENOSPC - insufficient disk resources
1348 * -EIO - i/o error
1349 *
1350 * note: IWRITE_LOCK(ipmap) held on entry/exit;
1351 */
1352static int
1353dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, s64 * results)
1354{
1355 struct metapage *mp;
1356 struct dmapctl *dcp;
1357 int rc, ti, i, k, m, n, agperlev;
1358 s64 blkno, lblkno;
1359 int budmin;
1360
1361 /* allocation request should not be for more than the
1362 * allocation group size.
1363 */
1364 if (l2nb > bmp->db_agl2size) {
1365 jfs_error(bmp->db_ipbmap->i_sb,
1366 "allocation request is larger than the allocation group size\n");
1367 return -EIO;
1368 }
1369
1370 /* determine the starting block number of the allocation
1371 * group.
1372 */
1373 blkno = (s64) agno << bmp->db_agl2size;
1374
1375 /* check if the allocation group size is the minimum allocation
1376 * group size or if the allocation group is completely free. if
1377 * the allocation group size is the minimum size of BPERDMAP (i.e.
1378 * 1 dmap), there is no need to search the dmap control page (below)
1379 * that fully describes the allocation group since the allocation
1380 * group is already fully described by a dmap. in this case, we
1381 * just call dbAllocCtl() to search the dmap tree and allocate the
1382 * required space if available.
1383 *
1384 * if the allocation group is completely free, dbAllocCtl() is
1385 * also called to allocate the required space. this is done for
1386 * two reasons. first, it makes no sense searching the dmap control
1387 * pages for free space when we know that free space exists. second,
1388 * the dmap control pages may indicate that the allocation group
1389 * has no free space if the allocation group is part (not the first
1390 * part) of a larger binary buddy system.
1391 */
1392 if (bmp->db_agsize == BPERDMAP
1393 || bmp->db_agfree[agno] == bmp->db_agsize) {
1394 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1395 if ((rc == -ENOSPC) &&
1396 (bmp->db_agfree[agno] == bmp->db_agsize)) {
1397 printk(KERN_ERR "blkno = %Lx, blocks = %Lx\n",
1398 (unsigned long long) blkno,
1399 (unsigned long long) nblocks);
1400 jfs_error(bmp->db_ipbmap->i_sb,
1401 "dbAllocCtl failed in free AG\n");
1402 }
1403 return (rc);
1404 }
1405
1406 /* the buffer for the dmap control page that fully describes the
1407 * allocation group.
1408 */
1409 lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, bmp->db_aglevel);
1410 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1411 if (mp == NULL)
1412 return -EIO;
1413 dcp = (struct dmapctl *) mp->data;
1414 budmin = dcp->budmin;
1415
1416 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1417 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
1418 release_metapage(mp);
1419 return -EIO;
1420 }
1421
1422 /* search the subtree(s) of the dmap control page that describes
1423 * the allocation group, looking for sufficient free space. to begin,
1424 * determine how many allocation groups are represented in a dmap
1425 * control page at the control page level (i.e. L0, L1, L2) that
1426 * fully describes an allocation group. next, determine the starting
1427 * tree index of this allocation group within the control page.
1428 */
1429 agperlev =
1430 (1 << (L2LPERCTL - (bmp->db_agheight << 1))) / bmp->db_agwidth;
1431 ti = bmp->db_agstart + bmp->db_agwidth * (agno & (agperlev - 1));
1432
1433 /* dmap control page trees fan-out by 4 and a single allocation
1434 * group may be described by 1 or 2 subtrees within the ag level
1435 * dmap control page, depending upon the ag size. examine the ag's
1436 * subtrees for sufficient free space, starting with the leftmost
1437 * subtree.
1438 */
1439 for (i = 0; i < bmp->db_agwidth; i++, ti++) {
1440 /* is there sufficient free space ?
1441 */
1442 if (l2nb > dcp->stree[ti])
1443 continue;
1444
1445 /* sufficient free space found in a subtree. now search down
1446 * the subtree to find the leftmost leaf that describes this
1447 * free space.
1448 */
1449 for (k = bmp->db_agheight; k > 0; k--) {
1450 for (n = 0, m = (ti << 2) + 1; n < 4; n++) {
1451 if (l2nb <= dcp->stree[m + n]) {
1452 ti = m + n;
1453 break;
1454 }
1455 }
1456 if (n == 4) {
1457 jfs_error(bmp->db_ipbmap->i_sb,
1458 "failed descending stree\n");
1459 release_metapage(mp);
1460 return -EIO;
1461 }
1462 }
1463
1464 /* determine the block number within the file system
1465 * that corresponds to this leaf.
1466 */
1467 if (bmp->db_aglevel == 2)
1468 blkno = 0;
1469 else if (bmp->db_aglevel == 1)
1470 blkno &= ~(MAXL1SIZE - 1);
1471 else /* bmp->db_aglevel == 0 */
1472 blkno &= ~(MAXL0SIZE - 1);
1473
1474 blkno +=
1475 ((s64) (ti - le32_to_cpu(dcp->leafidx))) << budmin;
1476
1477 /* release the buffer in preparation for going down
1478 * the next level of dmap control pages.
1479 */
1480 release_metapage(mp);
1481
1482 /* check if we need to continue to search down the lower
1483 * level dmap control pages. we need to if the number of
1484 * blocks required is less than maximum number of blocks
1485 * described at the next lower level.
1486 */
1487 if (l2nb < budmin) {
1488
1489 /* search the lower level dmap control pages to get
1490 * the starting block number of the dmap that
1491 * contains or starts off the free space.
1492 */
1493 if ((rc =
1494 dbFindCtl(bmp, l2nb, bmp->db_aglevel - 1,
1495 &blkno))) {
1496 if (rc == -ENOSPC) {
1497 jfs_error(bmp->db_ipbmap->i_sb,
1498 "control page inconsistent\n");
1499 return -EIO;
1500 }
1501 return (rc);
1502 }
1503 }
1504
1505 /* allocate the blocks.
1506 */
1507 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1508 if (rc == -ENOSPC) {
1509 jfs_error(bmp->db_ipbmap->i_sb,
1510 "unable to allocate blocks\n");
1511 rc = -EIO;
1512 }
1513 return (rc);
1514 }
1515
1516 /* no space in the allocation group. release the buffer and
1517 * return -ENOSPC.
1518 */
1519 release_metapage(mp);
1520
1521 return -ENOSPC;
1522}
1523
1524
1525/*
1526 * NAME: dbAllocAny()
1527 *
1528 * FUNCTION: attempt to allocate the specified number of contiguous
1529 * free blocks anywhere in the file system.
1530 *
1531 * dbAllocAny() attempts to find the sufficient free space by
1532 * searching down the dmap control pages, starting with the
1533 * highest level (i.e. L0, L1, L2) control page. if free space
1534 * large enough to satisfy the desired free space is found, the
1535 * desired free space is allocated.
1536 *
1537 * PARAMETERS:
1538 * bmp - pointer to bmap descriptor
1539 * nblocks - actual number of contiguous free blocks desired.
1540 * l2nb - log2 number of contiguous free blocks desired.
1541 * results - on successful return, set to the starting block number
1542 * of the newly allocated range.
1543 *
1544 * RETURN VALUES:
1545 * 0 - success
1546 * -ENOSPC - insufficient disk resources
1547 * -EIO - i/o error
1548 *
1549 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1550 */
1551static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results)
1552{
1553 int rc;
1554 s64 blkno = 0;
1555
1556 /* starting with the top level dmap control page, search
1557 * down the dmap control levels for sufficient free space.
1558 * if free space is found, dbFindCtl() returns the starting
1559 * block number of the dmap that contains or starts off the
1560 * range of free space.
1561 */
1562 if ((rc = dbFindCtl(bmp, l2nb, bmp->db_maxlevel, &blkno)))
1563 return (rc);
1564
1565 /* allocate the blocks.
1566 */
1567 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1568 if (rc == -ENOSPC) {
1569 jfs_error(bmp->db_ipbmap->i_sb, "unable to allocate blocks\n");
1570 return -EIO;
1571 }
1572 return (rc);
1573}
1574
1575
1576/*
1577 * NAME: dbDiscardAG()
1578 *
1579 * FUNCTION: attempt to discard (TRIM) all free blocks of specific AG
1580 *
1581 * algorithm:
1582 * 1) allocate blocks, as large as possible and save them
1583 * while holding IWRITE_LOCK on ipbmap
1584 * 2) trim all these saved block/length values
1585 * 3) mark the blocks free again
1586 *
1587 * benefit:
1588 * - we work only on one ag at some time, minimizing how long we
1589 * need to lock ipbmap
1590 * - reading / writing the fs is possible most time, even on
1591 * trimming
1592 *
1593 * downside:
1594 * - we write two times to the dmapctl and dmap pages
1595 * - but for me, this seems the best way, better ideas?
1596 * /TR 2012
1597 *
1598 * PARAMETERS:
1599 * ip - pointer to in-core inode
1600 * agno - ag to trim
1601 * minlen - minimum value of contiguous blocks
1602 *
1603 * RETURN VALUES:
1604 * s64 - actual number of blocks trimmed
1605 */
1606s64 dbDiscardAG(struct inode *ip, int agno, s64 minlen)
1607{
1608 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
1609 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
1610 s64 nblocks, blkno;
1611 u64 trimmed = 0;
1612 int rc, l2nb;
1613 struct super_block *sb = ipbmap->i_sb;
1614
1615 struct range2trim {
1616 u64 blkno;
1617 u64 nblocks;
1618 } *totrim, *tt;
1619
1620 /* max blkno / nblocks pairs to trim */
1621 int count = 0, range_cnt;
1622 u64 max_ranges;
1623
1624 /* prevent others from writing new stuff here, while trimming */
1625 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
1626
1627 nblocks = bmp->db_agfree[agno];
1628 max_ranges = nblocks;
1629 do_div(max_ranges, minlen);
1630 range_cnt = min_t(u64, max_ranges + 1, 32 * 1024);
1631 totrim = kmalloc_array(range_cnt, sizeof(struct range2trim), GFP_NOFS);
1632 if (totrim == NULL) {
1633 jfs_error(bmp->db_ipbmap->i_sb, "no memory for trim array\n");
1634 IWRITE_UNLOCK(ipbmap);
1635 return 0;
1636 }
1637
1638 tt = totrim;
1639 while (nblocks >= minlen) {
1640 l2nb = BLKSTOL2(nblocks);
1641
1642 /* 0 = okay, -EIO = fatal, -ENOSPC -> try smaller block */
1643 rc = dbAllocAG(bmp, agno, nblocks, l2nb, &blkno);
1644 if (rc == 0) {
1645 tt->blkno = blkno;
1646 tt->nblocks = nblocks;
1647 tt++; count++;
1648
1649 /* the whole ag is free, trim now */
1650 if (bmp->db_agfree[agno] == 0)
1651 break;
1652
1653 /* give a hint for the next while */
1654 nblocks = bmp->db_agfree[agno];
1655 continue;
1656 } else if (rc == -ENOSPC) {
1657 /* search for next smaller log2 block */
1658 l2nb = BLKSTOL2(nblocks) - 1;
1659 nblocks = 1 << l2nb;
1660 } else {
1661 /* Trim any already allocated blocks */
1662 jfs_error(bmp->db_ipbmap->i_sb, "-EIO\n");
1663 break;
1664 }
1665
1666 /* check, if our trim array is full */
1667 if (unlikely(count >= range_cnt - 1))
1668 break;
1669 }
1670 IWRITE_UNLOCK(ipbmap);
1671
1672 tt->nblocks = 0; /* mark the current end */
1673 for (tt = totrim; tt->nblocks != 0; tt++) {
1674 /* when mounted with online discard, dbFree() will
1675 * call jfs_issue_discard() itself */
1676 if (!(JFS_SBI(sb)->flag & JFS_DISCARD))
1677 jfs_issue_discard(ip, tt->blkno, tt->nblocks);
1678 dbFree(ip, tt->blkno, tt->nblocks);
1679 trimmed += tt->nblocks;
1680 }
1681 kfree(totrim);
1682
1683 return trimmed;
1684}
1685
1686/*
1687 * NAME: dbFindCtl()
1688 *
1689 * FUNCTION: starting at a specified dmap control page level and block
1690 * number, search down the dmap control levels for a range of
1691 * contiguous free blocks large enough to satisfy an allocation
1692 * request for the specified number of free blocks.
1693 *
1694 * if sufficient contiguous free blocks are found, this routine
1695 * returns the starting block number within a dmap page that
1696 * contains or starts a range of contiqious free blocks that
1697 * is sufficient in size.
1698 *
1699 * PARAMETERS:
1700 * bmp - pointer to bmap descriptor
1701 * level - starting dmap control page level.
1702 * l2nb - log2 number of contiguous free blocks desired.
1703 * *blkno - on entry, starting block number for conducting the search.
1704 * on successful return, the first block within a dmap page
1705 * that contains or starts a range of contiguous free blocks.
1706 *
1707 * RETURN VALUES:
1708 * 0 - success
1709 * -ENOSPC - insufficient disk resources
1710 * -EIO - i/o error
1711 *
1712 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1713 */
1714static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno)
1715{
1716 int rc, leafidx, lev;
1717 s64 b, lblkno;
1718 struct dmapctl *dcp;
1719 int budmin;
1720 struct metapage *mp;
1721
1722 /* starting at the specified dmap control page level and block
1723 * number, search down the dmap control levels for the starting
1724 * block number of a dmap page that contains or starts off
1725 * sufficient free blocks.
1726 */
1727 for (lev = level, b = *blkno; lev >= 0; lev--) {
1728 /* get the buffer of the dmap control page for the block
1729 * number and level (i.e. L0, L1, L2).
1730 */
1731 lblkno = BLKTOCTL(b, bmp->db_l2nbperpage, lev);
1732 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1733 if (mp == NULL)
1734 return -EIO;
1735 dcp = (struct dmapctl *) mp->data;
1736 budmin = dcp->budmin;
1737
1738 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1739 jfs_error(bmp->db_ipbmap->i_sb,
1740 "Corrupt dmapctl page\n");
1741 release_metapage(mp);
1742 return -EIO;
1743 }
1744
1745 /* search the tree within the dmap control page for
1746 * sufficient free space. if sufficient free space is found,
1747 * dbFindLeaf() returns the index of the leaf at which
1748 * free space was found.
1749 */
1750 rc = dbFindLeaf((dmtree_t *) dcp, l2nb, &leafidx);
1751
1752 /* release the buffer.
1753 */
1754 release_metapage(mp);
1755
1756 /* space found ?
1757 */
1758 if (rc) {
1759 if (lev != level) {
1760 jfs_error(bmp->db_ipbmap->i_sb,
1761 "dmap inconsistent\n");
1762 return -EIO;
1763 }
1764 return -ENOSPC;
1765 }
1766
1767 /* adjust the block number to reflect the location within
1768 * the dmap control page (i.e. the leaf) at which free
1769 * space was found.
1770 */
1771 b += (((s64) leafidx) << budmin);
1772
1773 /* we stop the search at this dmap control page level if
1774 * the number of blocks required is greater than or equal
1775 * to the maximum number of blocks described at the next
1776 * (lower) level.
1777 */
1778 if (l2nb >= budmin)
1779 break;
1780 }
1781
1782 *blkno = b;
1783 return (0);
1784}
1785
1786
1787/*
1788 * NAME: dbAllocCtl()
1789 *
1790 * FUNCTION: attempt to allocate a specified number of contiguous
1791 * blocks starting within a specific dmap.
1792 *
1793 * this routine is called by higher level routines that search
1794 * the dmap control pages above the actual dmaps for contiguous
1795 * free space. the result of successful searches by these
1796 * routines are the starting block numbers within dmaps, with
1797 * the dmaps themselves containing the desired contiguous free
1798 * space or starting a contiguous free space of desired size
1799 * that is made up of the blocks of one or more dmaps. these
1800 * calls should not fail due to insufficent resources.
1801 *
1802 * this routine is called in some cases where it is not known
1803 * whether it will fail due to insufficient resources. more
1804 * specifically, this occurs when allocating from an allocation
1805 * group whose size is equal to the number of blocks per dmap.
1806 * in this case, the dmap control pages are not examined prior
1807 * to calling this routine (to save pathlength) and the call
1808 * might fail.
1809 *
1810 * for a request size that fits within a dmap, this routine relies
1811 * upon the dmap's dmtree to find the requested contiguous free
1812 * space. for request sizes that are larger than a dmap, the
1813 * requested free space will start at the first block of the
1814 * first dmap (i.e. blkno).
1815 *
1816 * PARAMETERS:
1817 * bmp - pointer to bmap descriptor
1818 * nblocks - actual number of contiguous free blocks to allocate.
1819 * l2nb - log2 number of contiguous free blocks to allocate.
1820 * blkno - starting block number of the dmap to start the allocation
1821 * from.
1822 * results - on successful return, set to the starting block number
1823 * of the newly allocated range.
1824 *
1825 * RETURN VALUES:
1826 * 0 - success
1827 * -ENOSPC - insufficient disk resources
1828 * -EIO - i/o error
1829 *
1830 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1831 */
1832static int
1833dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, s64 * results)
1834{
1835 int rc, nb;
1836 s64 b, lblkno, n;
1837 struct metapage *mp;
1838 struct dmap *dp;
1839
1840 /* check if the allocation request is confined to a single dmap.
1841 */
1842 if (l2nb <= L2BPERDMAP) {
1843 /* get the buffer for the dmap.
1844 */
1845 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
1846 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1847 if (mp == NULL)
1848 return -EIO;
1849 dp = (struct dmap *) mp->data;
1850
1851 /* try to allocate the blocks.
1852 */
1853 rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results);
1854 if (rc == 0)
1855 mark_metapage_dirty(mp);
1856
1857 release_metapage(mp);
1858
1859 return (rc);
1860 }
1861
1862 /* allocation request involving multiple dmaps. it must start on
1863 * a dmap boundary.
1864 */
1865 assert((blkno & (BPERDMAP - 1)) == 0);
1866
1867 /* allocate the blocks dmap by dmap.
1868 */
1869 for (n = nblocks, b = blkno; n > 0; n -= nb, b += nb) {
1870 /* get the buffer for the dmap.
1871 */
1872 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1873 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1874 if (mp == NULL) {
1875 rc = -EIO;
1876 goto backout;
1877 }
1878 dp = (struct dmap *) mp->data;
1879
1880 /* the dmap better be all free.
1881 */
1882 if (dp->tree.stree[ROOT] != L2BPERDMAP) {
1883 release_metapage(mp);
1884 jfs_error(bmp->db_ipbmap->i_sb,
1885 "the dmap is not all free\n");
1886 rc = -EIO;
1887 goto backout;
1888 }
1889
1890 /* determine how many blocks to allocate from this dmap.
1891 */
1892 nb = min_t(s64, n, BPERDMAP);
1893
1894 /* allocate the blocks from the dmap.
1895 */
1896 if ((rc = dbAllocDmap(bmp, dp, b, nb))) {
1897 release_metapage(mp);
1898 goto backout;
1899 }
1900
1901 /* write the buffer.
1902 */
1903 write_metapage(mp);
1904 }
1905
1906 /* set the results (starting block number) and return.
1907 */
1908 *results = blkno;
1909 return (0);
1910
1911 /* something failed in handling an allocation request involving
1912 * multiple dmaps. we'll try to clean up by backing out any
1913 * allocation that has already happened for this request. if
1914 * we fail in backing out the allocation, we'll mark the file
1915 * system to indicate that blocks have been leaked.
1916 */
1917 backout:
1918
1919 /* try to backout the allocations dmap by dmap.
1920 */
1921 for (n = nblocks - n, b = blkno; n > 0;
1922 n -= BPERDMAP, b += BPERDMAP) {
1923 /* get the buffer for this dmap.
1924 */
1925 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1926 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1927 if (mp == NULL) {
1928 /* could not back out. mark the file system
1929 * to indicate that we have leaked blocks.
1930 */
1931 jfs_error(bmp->db_ipbmap->i_sb,
1932 "I/O Error: Block Leakage\n");
1933 continue;
1934 }
1935 dp = (struct dmap *) mp->data;
1936
1937 /* free the blocks is this dmap.
1938 */
1939 if (dbFreeDmap(bmp, dp, b, BPERDMAP)) {
1940 /* could not back out. mark the file system
1941 * to indicate that we have leaked blocks.
1942 */
1943 release_metapage(mp);
1944 jfs_error(bmp->db_ipbmap->i_sb, "Block Leakage\n");
1945 continue;
1946 }
1947
1948 /* write the buffer.
1949 */
1950 write_metapage(mp);
1951 }
1952
1953 return (rc);
1954}
1955
1956
1957/*
1958 * NAME: dbAllocDmapLev()
1959 *
1960 * FUNCTION: attempt to allocate a specified number of contiguous blocks
1961 * from a specified dmap.
1962 *
1963 * this routine checks if the contiguous blocks are available.
1964 * if so, nblocks of blocks are allocated; otherwise, ENOSPC is
1965 * returned.
1966 *
1967 * PARAMETERS:
1968 * mp - pointer to bmap descriptor
1969 * dp - pointer to dmap to attempt to allocate blocks from.
1970 * l2nb - log2 number of contiguous block desired.
1971 * nblocks - actual number of contiguous block desired.
1972 * results - on successful return, set to the starting block number
1973 * of the newly allocated range.
1974 *
1975 * RETURN VALUES:
1976 * 0 - success
1977 * -ENOSPC - insufficient disk resources
1978 * -EIO - i/o error
1979 *
1980 * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or
1981 * IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit;
1982 */
1983static int
1984dbAllocDmapLev(struct bmap * bmp,
1985 struct dmap * dp, int nblocks, int l2nb, s64 * results)
1986{
1987 s64 blkno;
1988 int leafidx, rc;
1989
1990 /* can't be more than a dmaps worth of blocks */
1991 assert(l2nb <= L2BPERDMAP);
1992
1993 /* search the tree within the dmap page for sufficient
1994 * free space. if sufficient free space is found, dbFindLeaf()
1995 * returns the index of the leaf at which free space was found.
1996 */
1997 if (dbFindLeaf((dmtree_t *) & dp->tree, l2nb, &leafidx))
1998 return -ENOSPC;
1999
2000 /* determine the block number within the file system corresponding
2001 * to the leaf at which free space was found.
2002 */
2003 blkno = le64_to_cpu(dp->start) + (leafidx << L2DBWORD);
2004
2005 /* if not all bits of the dmap word are free, get the starting
2006 * bit number within the dmap word of the required string of free
2007 * bits and adjust the block number with this value.
2008 */
2009 if (dp->tree.stree[leafidx + LEAFIND] < BUDMIN)
2010 blkno += dbFindBits(le32_to_cpu(dp->wmap[leafidx]), l2nb);
2011
2012 /* allocate the blocks */
2013 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
2014 *results = blkno;
2015
2016 return (rc);
2017}
2018
2019
2020/*
2021 * NAME: dbAllocDmap()
2022 *
2023 * FUNCTION: adjust the disk allocation map to reflect the allocation
2024 * of a specified block range within a dmap.
2025 *
2026 * this routine allocates the specified blocks from the dmap
2027 * through a call to dbAllocBits(). if the allocation of the
2028 * block range causes the maximum string of free blocks within
2029 * the dmap to change (i.e. the value of the root of the dmap's
2030 * dmtree), this routine will cause this change to be reflected
2031 * up through the appropriate levels of the dmap control pages
2032 * by a call to dbAdjCtl() for the L0 dmap control page that
2033 * covers this dmap.
2034 *
2035 * PARAMETERS:
2036 * bmp - pointer to bmap descriptor
2037 * dp - pointer to dmap to allocate the block range from.
2038 * blkno - starting block number of the block to be allocated.
2039 * nblocks - number of blocks to be allocated.
2040 *
2041 * RETURN VALUES:
2042 * 0 - success
2043 * -EIO - i/o error
2044 *
2045 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2046 */
2047static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2048 int nblocks)
2049{
2050 s8 oldroot;
2051 int rc;
2052
2053 /* save the current value of the root (i.e. maximum free string)
2054 * of the dmap tree.
2055 */
2056 oldroot = dp->tree.stree[ROOT];
2057
2058 /* allocate the specified (blocks) bits */
2059 dbAllocBits(bmp, dp, blkno, nblocks);
2060
2061 /* if the root has not changed, done. */
2062 if (dp->tree.stree[ROOT] == oldroot)
2063 return (0);
2064
2065 /* root changed. bubble the change up to the dmap control pages.
2066 * if the adjustment of the upper level control pages fails,
2067 * backout the bit allocation (thus making everything consistent).
2068 */
2069 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 1, 0)))
2070 dbFreeBits(bmp, dp, blkno, nblocks);
2071
2072 return (rc);
2073}
2074
2075
2076/*
2077 * NAME: dbFreeDmap()
2078 *
2079 * FUNCTION: adjust the disk allocation map to reflect the allocation
2080 * of a specified block range within a dmap.
2081 *
2082 * this routine frees the specified blocks from the dmap through
2083 * a call to dbFreeBits(). if the deallocation of the block range
2084 * causes the maximum string of free blocks within the dmap to
2085 * change (i.e. the value of the root of the dmap's dmtree), this
2086 * routine will cause this change to be reflected up through the
2087 * appropriate levels of the dmap control pages by a call to
2088 * dbAdjCtl() for the L0 dmap control page that covers this dmap.
2089 *
2090 * PARAMETERS:
2091 * bmp - pointer to bmap descriptor
2092 * dp - pointer to dmap to free the block range from.
2093 * blkno - starting block number of the block to be freed.
2094 * nblocks - number of blocks to be freed.
2095 *
2096 * RETURN VALUES:
2097 * 0 - success
2098 * -EIO - i/o error
2099 *
2100 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2101 */
2102static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2103 int nblocks)
2104{
2105 s8 oldroot;
2106 int rc = 0, word;
2107
2108 /* save the current value of the root (i.e. maximum free string)
2109 * of the dmap tree.
2110 */
2111 oldroot = dp->tree.stree[ROOT];
2112
2113 /* free the specified (blocks) bits */
2114 rc = dbFreeBits(bmp, dp, blkno, nblocks);
2115
2116 /* if error or the root has not changed, done. */
2117 if (rc || (dp->tree.stree[ROOT] == oldroot))
2118 return (rc);
2119
2120 /* root changed. bubble the change up to the dmap control pages.
2121 * if the adjustment of the upper level control pages fails,
2122 * backout the deallocation.
2123 */
2124 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 0, 0))) {
2125 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
2126
2127 /* as part of backing out the deallocation, we will have
2128 * to back split the dmap tree if the deallocation caused
2129 * the freed blocks to become part of a larger binary buddy
2130 * system.
2131 */
2132 if (dp->tree.stree[word] == NOFREE)
2133 dbBackSplit((dmtree_t *) & dp->tree, word);
2134
2135 dbAllocBits(bmp, dp, blkno, nblocks);
2136 }
2137
2138 return (rc);
2139}
2140
2141
2142/*
2143 * NAME: dbAllocBits()
2144 *
2145 * FUNCTION: allocate a specified block range from a dmap.
2146 *
2147 * this routine updates the dmap to reflect the working
2148 * state allocation of the specified block range. it directly
2149 * updates the bits of the working map and causes the adjustment
2150 * of the binary buddy system described by the dmap's dmtree
2151 * leaves to reflect the bits allocated. it also causes the
2152 * dmap's dmtree, as a whole, to reflect the allocated range.
2153 *
2154 * PARAMETERS:
2155 * bmp - pointer to bmap descriptor
2156 * dp - pointer to dmap to allocate bits from.
2157 * blkno - starting block number of the bits to be allocated.
2158 * nblocks - number of bits to be allocated.
2159 *
2160 * RETURN VALUES: none
2161 *
2162 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2163 */
2164static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2165 int nblocks)
2166{
2167 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2168 dmtree_t *tp = (dmtree_t *) & dp->tree;
2169 int size;
2170 s8 *leaf;
2171
2172 /* pick up a pointer to the leaves of the dmap tree */
2173 leaf = dp->tree.stree + LEAFIND;
2174
2175 /* determine the bit number and word within the dmap of the
2176 * starting block.
2177 */
2178 dbitno = blkno & (BPERDMAP - 1);
2179 word = dbitno >> L2DBWORD;
2180
2181 /* block range better be within the dmap */
2182 assert(dbitno + nblocks <= BPERDMAP);
2183
2184 /* allocate the bits of the dmap's words corresponding to the block
2185 * range. not all bits of the first and last words may be contained
2186 * within the block range. if this is the case, we'll work against
2187 * those words (i.e. partial first and/or last) on an individual basis
2188 * (a single pass), allocating the bits of interest by hand and
2189 * updating the leaf corresponding to the dmap word. a single pass
2190 * will be used for all dmap words fully contained within the
2191 * specified range. within this pass, the bits of all fully contained
2192 * dmap words will be marked as free in a single shot and the leaves
2193 * will be updated. a single leaf may describe the free space of
2194 * multiple dmap words, so we may update only a subset of the actual
2195 * leaves corresponding to the dmap words of the block range.
2196 */
2197 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2198 /* determine the bit number within the word and
2199 * the number of bits within the word.
2200 */
2201 wbitno = dbitno & (DBWORD - 1);
2202 nb = min(rembits, DBWORD - wbitno);
2203
2204 /* check if only part of a word is to be allocated.
2205 */
2206 if (nb < DBWORD) {
2207 /* allocate (set to 1) the appropriate bits within
2208 * this dmap word.
2209 */
2210 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
2211 >> wbitno);
2212
2213 /* update the leaf for this dmap word. in addition
2214 * to setting the leaf value to the binary buddy max
2215 * of the updated dmap word, dbSplit() will split
2216 * the binary system of the leaves if need be.
2217 */
2218 dbSplit(tp, word, BUDMIN,
2219 dbMaxBud((u8 *) & dp->wmap[word]));
2220
2221 word += 1;
2222 } else {
2223 /* one or more dmap words are fully contained
2224 * within the block range. determine how many
2225 * words and allocate (set to 1) the bits of these
2226 * words.
2227 */
2228 nwords = rembits >> L2DBWORD;
2229 memset(&dp->wmap[word], (int) ONES, nwords * 4);
2230
2231 /* determine how many bits.
2232 */
2233 nb = nwords << L2DBWORD;
2234
2235 /* now update the appropriate leaves to reflect
2236 * the allocated words.
2237 */
2238 for (; nwords > 0; nwords -= nw) {
2239 if (leaf[word] < BUDMIN) {
2240 jfs_error(bmp->db_ipbmap->i_sb,
2241 "leaf page corrupt\n");
2242 break;
2243 }
2244
2245 /* determine what the leaf value should be
2246 * updated to as the minimum of the l2 number
2247 * of bits being allocated and the l2 number
2248 * of bits currently described by this leaf.
2249 */
2250 size = min_t(int, leaf[word],
2251 NLSTOL2BSZ(nwords));
2252
2253 /* update the leaf to reflect the allocation.
2254 * in addition to setting the leaf value to
2255 * NOFREE, dbSplit() will split the binary
2256 * system of the leaves to reflect the current
2257 * allocation (size).
2258 */
2259 dbSplit(tp, word, size, NOFREE);
2260
2261 /* get the number of dmap words handled */
2262 nw = BUDSIZE(size, BUDMIN);
2263 word += nw;
2264 }
2265 }
2266 }
2267
2268 /* update the free count for this dmap */
2269 le32_add_cpu(&dp->nfree, -nblocks);
2270
2271 BMAP_LOCK(bmp);
2272
2273 /* if this allocation group is completely free,
2274 * update the maximum allocation group number if this allocation
2275 * group is the new max.
2276 */
2277 agno = blkno >> bmp->db_agl2size;
2278 if (agno > bmp->db_maxag)
2279 bmp->db_maxag = agno;
2280
2281 /* update the free count for the allocation group and map */
2282 bmp->db_agfree[agno] -= nblocks;
2283 bmp->db_nfree -= nblocks;
2284
2285 BMAP_UNLOCK(bmp);
2286}
2287
2288
2289/*
2290 * NAME: dbFreeBits()
2291 *
2292 * FUNCTION: free a specified block range from a dmap.
2293 *
2294 * this routine updates the dmap to reflect the working
2295 * state allocation of the specified block range. it directly
2296 * updates the bits of the working map and causes the adjustment
2297 * of the binary buddy system described by the dmap's dmtree
2298 * leaves to reflect the bits freed. it also causes the dmap's
2299 * dmtree, as a whole, to reflect the deallocated range.
2300 *
2301 * PARAMETERS:
2302 * bmp - pointer to bmap descriptor
2303 * dp - pointer to dmap to free bits from.
2304 * blkno - starting block number of the bits to be freed.
2305 * nblocks - number of bits to be freed.
2306 *
2307 * RETURN VALUES: 0 for success
2308 *
2309 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2310 */
2311static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2312 int nblocks)
2313{
2314 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2315 dmtree_t *tp = (dmtree_t *) & dp->tree;
2316 int rc = 0;
2317 int size;
2318
2319 /* determine the bit number and word within the dmap of the
2320 * starting block.
2321 */
2322 dbitno = blkno & (BPERDMAP - 1);
2323 word = dbitno >> L2DBWORD;
2324
2325 /* block range better be within the dmap.
2326 */
2327 assert(dbitno + nblocks <= BPERDMAP);
2328
2329 /* free the bits of the dmaps words corresponding to the block range.
2330 * not all bits of the first and last words may be contained within
2331 * the block range. if this is the case, we'll work against those
2332 * words (i.e. partial first and/or last) on an individual basis
2333 * (a single pass), freeing the bits of interest by hand and updating
2334 * the leaf corresponding to the dmap word. a single pass will be used
2335 * for all dmap words fully contained within the specified range.
2336 * within this pass, the bits of all fully contained dmap words will
2337 * be marked as free in a single shot and the leaves will be updated. a
2338 * single leaf may describe the free space of multiple dmap words,
2339 * so we may update only a subset of the actual leaves corresponding
2340 * to the dmap words of the block range.
2341 *
2342 * dbJoin() is used to update leaf values and will join the binary
2343 * buddy system of the leaves if the new leaf values indicate this
2344 * should be done.
2345 */
2346 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2347 /* determine the bit number within the word and
2348 * the number of bits within the word.
2349 */
2350 wbitno = dbitno & (DBWORD - 1);
2351 nb = min(rembits, DBWORD - wbitno);
2352
2353 /* check if only part of a word is to be freed.
2354 */
2355 if (nb < DBWORD) {
2356 /* free (zero) the appropriate bits within this
2357 * dmap word.
2358 */
2359 dp->wmap[word] &=
2360 cpu_to_le32(~(ONES << (DBWORD - nb)
2361 >> wbitno));
2362
2363 /* update the leaf for this dmap word.
2364 */
2365 rc = dbJoin(tp, word,
2366 dbMaxBud((u8 *) & dp->wmap[word]));
2367 if (rc)
2368 return rc;
2369
2370 word += 1;
2371 } else {
2372 /* one or more dmap words are fully contained
2373 * within the block range. determine how many
2374 * words and free (zero) the bits of these words.
2375 */
2376 nwords = rembits >> L2DBWORD;
2377 memset(&dp->wmap[word], 0, nwords * 4);
2378
2379 /* determine how many bits.
2380 */
2381 nb = nwords << L2DBWORD;
2382
2383 /* now update the appropriate leaves to reflect
2384 * the freed words.
2385 */
2386 for (; nwords > 0; nwords -= nw) {
2387 /* determine what the leaf value should be
2388 * updated to as the minimum of the l2 number
2389 * of bits being freed and the l2 (max) number
2390 * of bits that can be described by this leaf.
2391 */
2392 size =
2393 min(LITOL2BSZ
2394 (word, L2LPERDMAP, BUDMIN),
2395 NLSTOL2BSZ(nwords));
2396
2397 /* update the leaf.
2398 */
2399 rc = dbJoin(tp, word, size);
2400 if (rc)
2401 return rc;
2402
2403 /* get the number of dmap words handled.
2404 */
2405 nw = BUDSIZE(size, BUDMIN);
2406 word += nw;
2407 }
2408 }
2409 }
2410
2411 /* update the free count for this dmap.
2412 */
2413 le32_add_cpu(&dp->nfree, nblocks);
2414
2415 BMAP_LOCK(bmp);
2416
2417 /* update the free count for the allocation group and
2418 * map.
2419 */
2420 agno = blkno >> bmp->db_agl2size;
2421 bmp->db_nfree += nblocks;
2422 bmp->db_agfree[agno] += nblocks;
2423
2424 /* check if this allocation group is not completely free and
2425 * if it is currently the maximum (rightmost) allocation group.
2426 * if so, establish the new maximum allocation group number by
2427 * searching left for the first allocation group with allocation.
2428 */
2429 if ((bmp->db_agfree[agno] == bmp->db_agsize && agno == bmp->db_maxag) ||
2430 (agno == bmp->db_numag - 1 &&
2431 bmp->db_agfree[agno] == (bmp-> db_mapsize & (BPERDMAP - 1)))) {
2432 while (bmp->db_maxag > 0) {
2433 bmp->db_maxag -= 1;
2434 if (bmp->db_agfree[bmp->db_maxag] !=
2435 bmp->db_agsize)
2436 break;
2437 }
2438
2439 /* re-establish the allocation group preference if the
2440 * current preference is right of the maximum allocation
2441 * group.
2442 */
2443 if (bmp->db_agpref > bmp->db_maxag)
2444 bmp->db_agpref = bmp->db_maxag;
2445 }
2446
2447 BMAP_UNLOCK(bmp);
2448
2449 return 0;
2450}
2451
2452
2453/*
2454 * NAME: dbAdjCtl()
2455 *
2456 * FUNCTION: adjust a dmap control page at a specified level to reflect
2457 * the change in a lower level dmap or dmap control page's
2458 * maximum string of free blocks (i.e. a change in the root
2459 * of the lower level object's dmtree) due to the allocation
2460 * or deallocation of a range of blocks with a single dmap.
2461 *
2462 * on entry, this routine is provided with the new value of
2463 * the lower level dmap or dmap control page root and the
2464 * starting block number of the block range whose allocation
2465 * or deallocation resulted in the root change. this range
2466 * is respresented by a single leaf of the current dmapctl
2467 * and the leaf will be updated with this value, possibly
2468 * causing a binary buddy system within the leaves to be
2469 * split or joined. the update may also cause the dmapctl's
2470 * dmtree to be updated.
2471 *
2472 * if the adjustment of the dmap control page, itself, causes its
2473 * root to change, this change will be bubbled up to the next dmap
2474 * control level by a recursive call to this routine, specifying
2475 * the new root value and the next dmap control page level to
2476 * be adjusted.
2477 * PARAMETERS:
2478 * bmp - pointer to bmap descriptor
2479 * blkno - the first block of a block range within a dmap. it is
2480 * the allocation or deallocation of this block range that
2481 * requires the dmap control page to be adjusted.
2482 * newval - the new value of the lower level dmap or dmap control
2483 * page root.
2484 * alloc - 'true' if adjustment is due to an allocation.
2485 * level - current level of dmap control page (i.e. L0, L1, L2) to
2486 * be adjusted.
2487 *
2488 * RETURN VALUES:
2489 * 0 - success
2490 * -EIO - i/o error
2491 *
2492 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2493 */
2494static int
2495dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, int level)
2496{
2497 struct metapage *mp;
2498 s8 oldroot;
2499 int oldval;
2500 s64 lblkno;
2501 struct dmapctl *dcp;
2502 int rc, leafno, ti;
2503
2504 /* get the buffer for the dmap control page for the specified
2505 * block number and control page level.
2506 */
2507 lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, level);
2508 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
2509 if (mp == NULL)
2510 return -EIO;
2511 dcp = (struct dmapctl *) mp->data;
2512
2513 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
2514 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
2515 release_metapage(mp);
2516 return -EIO;
2517 }
2518
2519 /* determine the leaf number corresponding to the block and
2520 * the index within the dmap control tree.
2521 */
2522 leafno = BLKTOCTLLEAF(blkno, dcp->budmin);
2523 ti = leafno + le32_to_cpu(dcp->leafidx);
2524
2525 /* save the current leaf value and the current root level (i.e.
2526 * maximum l2 free string described by this dmapctl).
2527 */
2528 oldval = dcp->stree[ti];
2529 oldroot = dcp->stree[ROOT];
2530
2531 /* check if this is a control page update for an allocation.
2532 * if so, update the leaf to reflect the new leaf value using
2533 * dbSplit(); otherwise (deallocation), use dbJoin() to update
2534 * the leaf with the new value. in addition to updating the
2535 * leaf, dbSplit() will also split the binary buddy system of
2536 * the leaves, if required, and bubble new values within the
2537 * dmapctl tree, if required. similarly, dbJoin() will join
2538 * the binary buddy system of leaves and bubble new values up
2539 * the dmapctl tree as required by the new leaf value.
2540 */
2541 if (alloc) {
2542 /* check if we are in the middle of a binary buddy
2543 * system. this happens when we are performing the
2544 * first allocation out of an allocation group that
2545 * is part (not the first part) of a larger binary
2546 * buddy system. if we are in the middle, back split
2547 * the system prior to calling dbSplit() which assumes
2548 * that it is at the front of a binary buddy system.
2549 */
2550 if (oldval == NOFREE) {
2551 rc = dbBackSplit((dmtree_t *) dcp, leafno);
2552 if (rc)
2553 return rc;
2554 oldval = dcp->stree[ti];
2555 }
2556 dbSplit((dmtree_t *) dcp, leafno, dcp->budmin, newval);
2557 } else {
2558 rc = dbJoin((dmtree_t *) dcp, leafno, newval);
2559 if (rc)
2560 return rc;
2561 }
2562
2563 /* check if the root of the current dmap control page changed due
2564 * to the update and if the current dmap control page is not at
2565 * the current top level (i.e. L0, L1, L2) of the map. if so (i.e.
2566 * root changed and this is not the top level), call this routine
2567 * again (recursion) for the next higher level of the mapping to
2568 * reflect the change in root for the current dmap control page.
2569 */
2570 if (dcp->stree[ROOT] != oldroot) {
2571 /* are we below the top level of the map. if so,
2572 * bubble the root up to the next higher level.
2573 */
2574 if (level < bmp->db_maxlevel) {
2575 /* bubble up the new root of this dmap control page to
2576 * the next level.
2577 */
2578 if ((rc =
2579 dbAdjCtl(bmp, blkno, dcp->stree[ROOT], alloc,
2580 level + 1))) {
2581 /* something went wrong in bubbling up the new
2582 * root value, so backout the changes to the
2583 * current dmap control page.
2584 */
2585 if (alloc) {
2586 dbJoin((dmtree_t *) dcp, leafno,
2587 oldval);
2588 } else {
2589 /* the dbJoin() above might have
2590 * caused a larger binary buddy system
2591 * to form and we may now be in the
2592 * middle of it. if this is the case,
2593 * back split the buddies.
2594 */
2595 if (dcp->stree[ti] == NOFREE)
2596 dbBackSplit((dmtree_t *)
2597 dcp, leafno);
2598 dbSplit((dmtree_t *) dcp, leafno,
2599 dcp->budmin, oldval);
2600 }
2601
2602 /* release the buffer and return the error.
2603 */
2604 release_metapage(mp);
2605 return (rc);
2606 }
2607 } else {
2608 /* we're at the top level of the map. update
2609 * the bmap control page to reflect the size
2610 * of the maximum free buddy system.
2611 */
2612 assert(level == bmp->db_maxlevel);
2613 if (bmp->db_maxfreebud != oldroot) {
2614 jfs_error(bmp->db_ipbmap->i_sb,
2615 "the maximum free buddy is not the old root\n");
2616 }
2617 bmp->db_maxfreebud = dcp->stree[ROOT];
2618 }
2619 }
2620
2621 /* write the buffer.
2622 */
2623 write_metapage(mp);
2624
2625 return (0);
2626}
2627
2628
2629/*
2630 * NAME: dbSplit()
2631 *
2632 * FUNCTION: update the leaf of a dmtree with a new value, splitting
2633 * the leaf from the binary buddy system of the dmtree's
2634 * leaves, as required.
2635 *
2636 * PARAMETERS:
2637 * tp - pointer to the tree containing the leaf.
2638 * leafno - the number of the leaf to be updated.
2639 * splitsz - the size the binary buddy system starting at the leaf
2640 * must be split to, specified as the log2 number of blocks.
2641 * newval - the new value for the leaf.
2642 *
2643 * RETURN VALUES: none
2644 *
2645 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2646 */
2647static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval)
2648{
2649 int budsz;
2650 int cursz;
2651 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2652
2653 /* check if the leaf needs to be split.
2654 */
2655 if (leaf[leafno] > tp->dmt_budmin) {
2656 /* the split occurs by cutting the buddy system in half
2657 * at the specified leaf until we reach the specified
2658 * size. pick up the starting split size (current size
2659 * - 1 in l2) and the corresponding buddy size.
2660 */
2661 cursz = leaf[leafno] - 1;
2662 budsz = BUDSIZE(cursz, tp->dmt_budmin);
2663
2664 /* split until we reach the specified size.
2665 */
2666 while (cursz >= splitsz) {
2667 /* update the buddy's leaf with its new value.
2668 */
2669 dbAdjTree(tp, leafno ^ budsz, cursz);
2670
2671 /* on to the next size and buddy.
2672 */
2673 cursz -= 1;
2674 budsz >>= 1;
2675 }
2676 }
2677
2678 /* adjust the dmap tree to reflect the specified leaf's new
2679 * value.
2680 */
2681 dbAdjTree(tp, leafno, newval);
2682}
2683
2684
2685/*
2686 * NAME: dbBackSplit()
2687 *
2688 * FUNCTION: back split the binary buddy system of dmtree leaves
2689 * that hold a specified leaf until the specified leaf
2690 * starts its own binary buddy system.
2691 *
2692 * the allocators typically perform allocations at the start
2693 * of binary buddy systems and dbSplit() is used to accomplish
2694 * any required splits. in some cases, however, allocation
2695 * may occur in the middle of a binary system and requires a
2696 * back split, with the split proceeding out from the middle of
2697 * the system (less efficient) rather than the start of the
2698 * system (more efficient). the cases in which a back split
2699 * is required are rare and are limited to the first allocation
2700 * within an allocation group which is a part (not first part)
2701 * of a larger binary buddy system and a few exception cases
2702 * in which a previous join operation must be backed out.
2703 *
2704 * PARAMETERS:
2705 * tp - pointer to the tree containing the leaf.
2706 * leafno - the number of the leaf to be updated.
2707 *
2708 * RETURN VALUES: none
2709 *
2710 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2711 */
2712static int dbBackSplit(dmtree_t * tp, int leafno)
2713{
2714 int budsz, bud, w, bsz, size;
2715 int cursz;
2716 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2717
2718 /* leaf should be part (not first part) of a binary
2719 * buddy system.
2720 */
2721 assert(leaf[leafno] == NOFREE);
2722
2723 /* the back split is accomplished by iteratively finding the leaf
2724 * that starts the buddy system that contains the specified leaf and
2725 * splitting that system in two. this iteration continues until
2726 * the specified leaf becomes the start of a buddy system.
2727 *
2728 * determine maximum possible l2 size for the specified leaf.
2729 */
2730 size =
2731 LITOL2BSZ(leafno, le32_to_cpu(tp->dmt_l2nleafs),
2732 tp->dmt_budmin);
2733
2734 /* determine the number of leaves covered by this size. this
2735 * is the buddy size that we will start with as we search for
2736 * the buddy system that contains the specified leaf.
2737 */
2738 budsz = BUDSIZE(size, tp->dmt_budmin);
2739
2740 /* back split.
2741 */
2742 while (leaf[leafno] == NOFREE) {
2743 /* find the leftmost buddy leaf.
2744 */
2745 for (w = leafno, bsz = budsz;; bsz <<= 1,
2746 w = (w < bud) ? w : bud) {
2747 if (bsz >= le32_to_cpu(tp->dmt_nleafs)) {
2748 jfs_err("JFS: block map error in dbBackSplit");
2749 return -EIO;
2750 }
2751
2752 /* determine the buddy.
2753 */
2754 bud = w ^ bsz;
2755
2756 /* check if this buddy is the start of the system.
2757 */
2758 if (leaf[bud] != NOFREE) {
2759 /* split the leaf at the start of the
2760 * system in two.
2761 */
2762 cursz = leaf[bud] - 1;
2763 dbSplit(tp, bud, cursz, cursz);
2764 break;
2765 }
2766 }
2767 }
2768
2769 if (leaf[leafno] != size) {
2770 jfs_err("JFS: wrong leaf value in dbBackSplit");
2771 return -EIO;
2772 }
2773 return 0;
2774}
2775
2776
2777/*
2778 * NAME: dbJoin()
2779 *
2780 * FUNCTION: update the leaf of a dmtree with a new value, joining
2781 * the leaf with other leaves of the dmtree into a multi-leaf
2782 * binary buddy system, as required.
2783 *
2784 * PARAMETERS:
2785 * tp - pointer to the tree containing the leaf.
2786 * leafno - the number of the leaf to be updated.
2787 * newval - the new value for the leaf.
2788 *
2789 * RETURN VALUES: none
2790 */
2791static int dbJoin(dmtree_t * tp, int leafno, int newval)
2792{
2793 int budsz, buddy;
2794 s8 *leaf;
2795
2796 /* can the new leaf value require a join with other leaves ?
2797 */
2798 if (newval >= tp->dmt_budmin) {
2799 /* pickup a pointer to the leaves of the tree.
2800 */
2801 leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2802
2803 /* try to join the specified leaf into a large binary
2804 * buddy system. the join proceeds by attempting to join
2805 * the specified leafno with its buddy (leaf) at new value.
2806 * if the join occurs, we attempt to join the left leaf
2807 * of the joined buddies with its buddy at new value + 1.
2808 * we continue to join until we find a buddy that cannot be
2809 * joined (does not have a value equal to the size of the
2810 * last join) or until all leaves have been joined into a
2811 * single system.
2812 *
2813 * get the buddy size (number of words covered) of
2814 * the new value.
2815 */
2816 budsz = BUDSIZE(newval, tp->dmt_budmin);
2817
2818 /* try to join.
2819 */
2820 while (budsz < le32_to_cpu(tp->dmt_nleafs)) {
2821 /* get the buddy leaf.
2822 */
2823 buddy = leafno ^ budsz;
2824
2825 /* if the leaf's new value is greater than its
2826 * buddy's value, we join no more.
2827 */
2828 if (newval > leaf[buddy])
2829 break;
2830
2831 /* It shouldn't be less */
2832 if (newval < leaf[buddy])
2833 return -EIO;
2834
2835 /* check which (leafno or buddy) is the left buddy.
2836 * the left buddy gets to claim the blocks resulting
2837 * from the join while the right gets to claim none.
2838 * the left buddy is also eligible to participate in
2839 * a join at the next higher level while the right
2840 * is not.
2841 *
2842 */
2843 if (leafno < buddy) {
2844 /* leafno is the left buddy.
2845 */
2846 dbAdjTree(tp, buddy, NOFREE);
2847 } else {
2848 /* buddy is the left buddy and becomes
2849 * leafno.
2850 */
2851 dbAdjTree(tp, leafno, NOFREE);
2852 leafno = buddy;
2853 }
2854
2855 /* on to try the next join.
2856 */
2857 newval += 1;
2858 budsz <<= 1;
2859 }
2860 }
2861
2862 /* update the leaf value.
2863 */
2864 dbAdjTree(tp, leafno, newval);
2865
2866 return 0;
2867}
2868
2869
2870/*
2871 * NAME: dbAdjTree()
2872 *
2873 * FUNCTION: update a leaf of a dmtree with a new value, adjusting
2874 * the dmtree, as required, to reflect the new leaf value.
2875 * the combination of any buddies must already be done before
2876 * this is called.
2877 *
2878 * PARAMETERS:
2879 * tp - pointer to the tree to be adjusted.
2880 * leafno - the number of the leaf to be updated.
2881 * newval - the new value for the leaf.
2882 *
2883 * RETURN VALUES: none
2884 */
2885static void dbAdjTree(dmtree_t * tp, int leafno, int newval)
2886{
2887 int lp, pp, k;
2888 int max;
2889
2890 /* pick up the index of the leaf for this leafno.
2891 */
2892 lp = leafno + le32_to_cpu(tp->dmt_leafidx);
2893
2894 /* is the current value the same as the old value ? if so,
2895 * there is nothing to do.
2896 */
2897 if (tp->dmt_stree[lp] == newval)
2898 return;
2899
2900 /* set the new value.
2901 */
2902 tp->dmt_stree[lp] = newval;
2903
2904 /* bubble the new value up the tree as required.
2905 */
2906 for (k = 0; k < le32_to_cpu(tp->dmt_height); k++) {
2907 /* get the index of the first leaf of the 4 leaf
2908 * group containing the specified leaf (leafno).
2909 */
2910 lp = ((lp - 1) & ~0x03) + 1;
2911
2912 /* get the index of the parent of this 4 leaf group.
2913 */
2914 pp = (lp - 1) >> 2;
2915
2916 /* determine the maximum of the 4 leaves.
2917 */
2918 max = TREEMAX(&tp->dmt_stree[lp]);
2919
2920 /* if the maximum of the 4 is the same as the
2921 * parent's value, we're done.
2922 */
2923 if (tp->dmt_stree[pp] == max)
2924 break;
2925
2926 /* parent gets new value.
2927 */
2928 tp->dmt_stree[pp] = max;
2929
2930 /* parent becomes leaf for next go-round.
2931 */
2932 lp = pp;
2933 }
2934}
2935
2936
2937/*
2938 * NAME: dbFindLeaf()
2939 *
2940 * FUNCTION: search a dmtree_t for sufficient free blocks, returning
2941 * the index of a leaf describing the free blocks if
2942 * sufficient free blocks are found.
2943 *
2944 * the search starts at the top of the dmtree_t tree and
2945 * proceeds down the tree to the leftmost leaf with sufficient
2946 * free space.
2947 *
2948 * PARAMETERS:
2949 * tp - pointer to the tree to be searched.
2950 * l2nb - log2 number of free blocks to search for.
2951 * leafidx - return pointer to be set to the index of the leaf
2952 * describing at least l2nb free blocks if sufficient
2953 * free blocks are found.
2954 *
2955 * RETURN VALUES:
2956 * 0 - success
2957 * -ENOSPC - insufficient free blocks.
2958 */
2959static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx)
2960{
2961 int ti, n = 0, k, x = 0;
2962
2963 /* first check the root of the tree to see if there is
2964 * sufficient free space.
2965 */
2966 if (l2nb > tp->dmt_stree[ROOT])
2967 return -ENOSPC;
2968
2969 /* sufficient free space available. now search down the tree
2970 * starting at the next level for the leftmost leaf that
2971 * describes sufficient free space.
2972 */
2973 for (k = le32_to_cpu(tp->dmt_height), ti = 1;
2974 k > 0; k--, ti = ((ti + n) << 2) + 1) {
2975 /* search the four nodes at this level, starting from
2976 * the left.
2977 */
2978 for (x = ti, n = 0; n < 4; n++) {
2979 /* sufficient free space found. move to the next
2980 * level (or quit if this is the last level).
2981 */
2982 if (l2nb <= tp->dmt_stree[x + n])
2983 break;
2984 }
2985
2986 /* better have found something since the higher
2987 * levels of the tree said it was here.
2988 */
2989 assert(n < 4);
2990 }
2991
2992 /* set the return to the leftmost leaf describing sufficient
2993 * free space.
2994 */
2995 *leafidx = x + n - le32_to_cpu(tp->dmt_leafidx);
2996
2997 return (0);
2998}
2999
3000
3001/*
3002 * NAME: dbFindBits()
3003 *
3004 * FUNCTION: find a specified number of binary buddy free bits within a
3005 * dmap bitmap word value.
3006 *
3007 * this routine searches the bitmap value for (1 << l2nb) free
3008 * bits at (1 << l2nb) alignments within the value.
3009 *
3010 * PARAMETERS:
3011 * word - dmap bitmap word value.
3012 * l2nb - number of free bits specified as a log2 number.
3013 *
3014 * RETURN VALUES:
3015 * starting bit number of free bits.
3016 */
3017static int dbFindBits(u32 word, int l2nb)
3018{
3019 int bitno, nb;
3020 u32 mask;
3021
3022 /* get the number of bits.
3023 */
3024 nb = 1 << l2nb;
3025 assert(nb <= DBWORD);
3026
3027 /* complement the word so we can use a mask (i.e. 0s represent
3028 * free bits) and compute the mask.
3029 */
3030 word = ~word;
3031 mask = ONES << (DBWORD - nb);
3032
3033 /* scan the word for nb free bits at nb alignments.
3034 */
3035 for (bitno = 0; mask != 0; bitno += nb, mask >>= nb) {
3036 if ((mask & word) == mask)
3037 break;
3038 }
3039
3040 ASSERT(bitno < 32);
3041
3042 /* return the bit number.
3043 */
3044 return (bitno);
3045}
3046
3047
3048/*
3049 * NAME: dbMaxBud(u8 *cp)
3050 *
3051 * FUNCTION: determine the largest binary buddy string of free
3052 * bits within 32-bits of the map.
3053 *
3054 * PARAMETERS:
3055 * cp - pointer to the 32-bit value.
3056 *
3057 * RETURN VALUES:
3058 * largest binary buddy of free bits within a dmap word.
3059 */
3060static int dbMaxBud(u8 * cp)
3061{
3062 signed char tmp1, tmp2;
3063
3064 /* check if the wmap word is all free. if so, the
3065 * free buddy size is BUDMIN.
3066 */
3067 if (*((uint *) cp) == 0)
3068 return (BUDMIN);
3069
3070 /* check if the wmap word is half free. if so, the
3071 * free buddy size is BUDMIN-1.
3072 */
3073 if (*((u16 *) cp) == 0 || *((u16 *) cp + 1) == 0)
3074 return (BUDMIN - 1);
3075
3076 /* not all free or half free. determine the free buddy
3077 * size thru table lookup using quarters of the wmap word.
3078 */
3079 tmp1 = max(budtab[cp[2]], budtab[cp[3]]);
3080 tmp2 = max(budtab[cp[0]], budtab[cp[1]]);
3081 return (max(tmp1, tmp2));
3082}
3083
3084
3085/*
3086 * NAME: cnttz(uint word)
3087 *
3088 * FUNCTION: determine the number of trailing zeros within a 32-bit
3089 * value.
3090 *
3091 * PARAMETERS:
3092 * value - 32-bit value to be examined.
3093 *
3094 * RETURN VALUES:
3095 * count of trailing zeros
3096 */
3097static int cnttz(u32 word)
3098{
3099 int n;
3100
3101 for (n = 0; n < 32; n++, word >>= 1) {
3102 if (word & 0x01)
3103 break;
3104 }
3105
3106 return (n);
3107}
3108
3109
3110/*
3111 * NAME: cntlz(u32 value)
3112 *
3113 * FUNCTION: determine the number of leading zeros within a 32-bit
3114 * value.
3115 *
3116 * PARAMETERS:
3117 * value - 32-bit value to be examined.
3118 *
3119 * RETURN VALUES:
3120 * count of leading zeros
3121 */
3122static int cntlz(u32 value)
3123{
3124 int n;
3125
3126 for (n = 0; n < 32; n++, value <<= 1) {
3127 if (value & HIGHORDER)
3128 break;
3129 }
3130 return (n);
3131}
3132
3133
3134/*
3135 * NAME: blkstol2(s64 nb)
3136 *
3137 * FUNCTION: convert a block count to its log2 value. if the block
3138 * count is not a l2 multiple, it is rounded up to the next
3139 * larger l2 multiple.
3140 *
3141 * PARAMETERS:
3142 * nb - number of blocks
3143 *
3144 * RETURN VALUES:
3145 * log2 number of blocks
3146 */
3147static int blkstol2(s64 nb)
3148{
3149 int l2nb;
3150 s64 mask; /* meant to be signed */
3151
3152 mask = (s64) 1 << (64 - 1);
3153
3154 /* count the leading bits.
3155 */
3156 for (l2nb = 0; l2nb < 64; l2nb++, mask >>= 1) {
3157 /* leading bit found.
3158 */
3159 if (nb & mask) {
3160 /* determine the l2 value.
3161 */
3162 l2nb = (64 - 1) - l2nb;
3163
3164 /* check if we need to round up.
3165 */
3166 if (~mask & nb)
3167 l2nb++;
3168
3169 return (l2nb);
3170 }
3171 }
3172 assert(0);
3173 return 0; /* fix compiler warning */
3174}
3175
3176
3177/*
3178 * NAME: dbAllocBottomUp()
3179 *
3180 * FUNCTION: alloc the specified block range from the working block
3181 * allocation map.
3182 *
3183 * the blocks will be alloc from the working map one dmap
3184 * at a time.
3185 *
3186 * PARAMETERS:
3187 * ip - pointer to in-core inode;
3188 * blkno - starting block number to be freed.
3189 * nblocks - number of blocks to be freed.
3190 *
3191 * RETURN VALUES:
3192 * 0 - success
3193 * -EIO - i/o error
3194 */
3195int dbAllocBottomUp(struct inode *ip, s64 blkno, s64 nblocks)
3196{
3197 struct metapage *mp;
3198 struct dmap *dp;
3199 int nb, rc;
3200 s64 lblkno, rem;
3201 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
3202 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
3203
3204 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
3205
3206 /* block to be allocated better be within the mapsize. */
3207 ASSERT(nblocks <= bmp->db_mapsize - blkno);
3208
3209 /*
3210 * allocate the blocks a dmap at a time.
3211 */
3212 mp = NULL;
3213 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
3214 /* release previous dmap if any */
3215 if (mp) {
3216 write_metapage(mp);
3217 }
3218
3219 /* get the buffer for the current dmap. */
3220 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
3221 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
3222 if (mp == NULL) {
3223 IREAD_UNLOCK(ipbmap);
3224 return -EIO;
3225 }
3226 dp = (struct dmap *) mp->data;
3227
3228 /* determine the number of blocks to be allocated from
3229 * this dmap.
3230 */
3231 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
3232
3233 /* allocate the blocks. */
3234 if ((rc = dbAllocDmapBU(bmp, dp, blkno, nb))) {
3235 release_metapage(mp);
3236 IREAD_UNLOCK(ipbmap);
3237 return (rc);
3238 }
3239 }
3240
3241 /* write the last buffer. */
3242 write_metapage(mp);
3243
3244 IREAD_UNLOCK(ipbmap);
3245
3246 return (0);
3247}
3248
3249
3250static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
3251 int nblocks)
3252{
3253 int rc;
3254 int dbitno, word, rembits, nb, nwords, wbitno, agno;
3255 s8 oldroot;
3256 struct dmaptree *tp = (struct dmaptree *) & dp->tree;
3257
3258 /* save the current value of the root (i.e. maximum free string)
3259 * of the dmap tree.
3260 */
3261 oldroot = tp->stree[ROOT];
3262
3263 /* determine the bit number and word within the dmap of the
3264 * starting block.
3265 */
3266 dbitno = blkno & (BPERDMAP - 1);
3267 word = dbitno >> L2DBWORD;
3268
3269 /* block range better be within the dmap */
3270 assert(dbitno + nblocks <= BPERDMAP);
3271
3272 /* allocate the bits of the dmap's words corresponding to the block
3273 * range. not all bits of the first and last words may be contained
3274 * within the block range. if this is the case, we'll work against
3275 * those words (i.e. partial first and/or last) on an individual basis
3276 * (a single pass), allocating the bits of interest by hand and
3277 * updating the leaf corresponding to the dmap word. a single pass
3278 * will be used for all dmap words fully contained within the
3279 * specified range. within this pass, the bits of all fully contained
3280 * dmap words will be marked as free in a single shot and the leaves
3281 * will be updated. a single leaf may describe the free space of
3282 * multiple dmap words, so we may update only a subset of the actual
3283 * leaves corresponding to the dmap words of the block range.
3284 */
3285 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
3286 /* determine the bit number within the word and
3287 * the number of bits within the word.
3288 */
3289 wbitno = dbitno & (DBWORD - 1);
3290 nb = min(rembits, DBWORD - wbitno);
3291
3292 /* check if only part of a word is to be allocated.
3293 */
3294 if (nb < DBWORD) {
3295 /* allocate (set to 1) the appropriate bits within
3296 * this dmap word.
3297 */
3298 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
3299 >> wbitno);
3300
3301 word++;
3302 } else {
3303 /* one or more dmap words are fully contained
3304 * within the block range. determine how many
3305 * words and allocate (set to 1) the bits of these
3306 * words.
3307 */
3308 nwords = rembits >> L2DBWORD;
3309 memset(&dp->wmap[word], (int) ONES, nwords * 4);
3310
3311 /* determine how many bits */
3312 nb = nwords << L2DBWORD;
3313 word += nwords;
3314 }
3315 }
3316
3317 /* update the free count for this dmap */
3318 le32_add_cpu(&dp->nfree, -nblocks);
3319
3320 /* reconstruct summary tree */
3321 dbInitDmapTree(dp);
3322
3323 BMAP_LOCK(bmp);
3324
3325 /* if this allocation group is completely free,
3326 * update the highest active allocation group number
3327 * if this allocation group is the new max.
3328 */
3329 agno = blkno >> bmp->db_agl2size;
3330 if (agno > bmp->db_maxag)
3331 bmp->db_maxag = agno;
3332
3333 /* update the free count for the allocation group and map */
3334 bmp->db_agfree[agno] -= nblocks;
3335 bmp->db_nfree -= nblocks;
3336
3337 BMAP_UNLOCK(bmp);
3338
3339 /* if the root has not changed, done. */
3340 if (tp->stree[ROOT] == oldroot)
3341 return (0);
3342
3343 /* root changed. bubble the change up to the dmap control pages.
3344 * if the adjustment of the upper level control pages fails,
3345 * backout the bit allocation (thus making everything consistent).
3346 */
3347 if ((rc = dbAdjCtl(bmp, blkno, tp->stree[ROOT], 1, 0)))
3348 dbFreeBits(bmp, dp, blkno, nblocks);
3349
3350 return (rc);
3351}
3352
3353
3354/*
3355 * NAME: dbExtendFS()
3356 *
3357 * FUNCTION: extend bmap from blkno for nblocks;
3358 * dbExtendFS() updates bmap ready for dbAllocBottomUp();
3359 *
3360 * L2
3361 * |
3362 * L1---------------------------------L1
3363 * | |
3364 * L0---------L0---------L0 L0---------L0---------L0
3365 * | | | | | |
3366 * d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,.,dm;
3367 * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm
3368 *
3369 * <---old---><----------------------------extend----------------------->
3370 */
3371int dbExtendFS(struct inode *ipbmap, s64 blkno, s64 nblocks)
3372{
3373 struct jfs_sb_info *sbi = JFS_SBI(ipbmap->i_sb);
3374 int nbperpage = sbi->nbperpage;
3375 int i, i0 = true, j, j0 = true, k, n;
3376 s64 newsize;
3377 s64 p;
3378 struct metapage *mp, *l2mp, *l1mp = NULL, *l0mp = NULL;
3379 struct dmapctl *l2dcp, *l1dcp, *l0dcp;
3380 struct dmap *dp;
3381 s8 *l0leaf, *l1leaf, *l2leaf;
3382 struct bmap *bmp = sbi->bmap;
3383 int agno, l2agsize, oldl2agsize;
3384 s64 ag_rem;
3385
3386 newsize = blkno + nblocks;
3387
3388 jfs_info("dbExtendFS: blkno:%Ld nblocks:%Ld newsize:%Ld",
3389 (long long) blkno, (long long) nblocks, (long long) newsize);
3390
3391 /*
3392 * initialize bmap control page.
3393 *
3394 * all the data in bmap control page should exclude
3395 * the mkfs hidden dmap page.
3396 */
3397
3398 /* update mapsize */
3399 bmp->db_mapsize = newsize;
3400 bmp->db_maxlevel = BMAPSZTOLEV(bmp->db_mapsize);
3401
3402 /* compute new AG size */
3403 l2agsize = dbGetL2AGSize(newsize);
3404 oldl2agsize = bmp->db_agl2size;
3405
3406 bmp->db_agl2size = l2agsize;
3407 bmp->db_agsize = 1 << l2agsize;
3408
3409 /* compute new number of AG */
3410 agno = bmp->db_numag;
3411 bmp->db_numag = newsize >> l2agsize;
3412 bmp->db_numag += ((u32) newsize % (u32) bmp->db_agsize) ? 1 : 0;
3413
3414 /*
3415 * reconfigure db_agfree[]
3416 * from old AG configuration to new AG configuration;
3417 *
3418 * coalesce contiguous k (newAGSize/oldAGSize) AGs;
3419 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn;
3420 * note: new AG size = old AG size * (2**x).
3421 */
3422 if (l2agsize == oldl2agsize)
3423 goto extend;
3424 k = 1 << (l2agsize - oldl2agsize);
3425 ag_rem = bmp->db_agfree[0]; /* save agfree[0] */
3426 for (i = 0, n = 0; i < agno; n++) {
3427 bmp->db_agfree[n] = 0; /* init collection point */
3428
3429 /* coalesce contiguous k AGs; */
3430 for (j = 0; j < k && i < agno; j++, i++) {
3431 /* merge AGi to AGn */
3432 bmp->db_agfree[n] += bmp->db_agfree[i];
3433 }
3434 }
3435 bmp->db_agfree[0] += ag_rem; /* restore agfree[0] */
3436
3437 for (; n < MAXAG; n++)
3438 bmp->db_agfree[n] = 0;
3439
3440 /*
3441 * update highest active ag number
3442 */
3443
3444 bmp->db_maxag = bmp->db_maxag / k;
3445
3446 /*
3447 * extend bmap
3448 *
3449 * update bit maps and corresponding level control pages;
3450 * global control page db_nfree, db_agfree[agno], db_maxfreebud;
3451 */
3452 extend:
3453 /* get L2 page */
3454 p = BMAPBLKNO + nbperpage; /* L2 page */
3455 l2mp = read_metapage(ipbmap, p, PSIZE, 0);
3456 if (!l2mp) {
3457 jfs_error(ipbmap->i_sb, "L2 page could not be read\n");
3458 return -EIO;
3459 }
3460 l2dcp = (struct dmapctl *) l2mp->data;
3461
3462 /* compute start L1 */
3463 k = blkno >> L2MAXL1SIZE;
3464 l2leaf = l2dcp->stree + CTLLEAFIND + k;
3465 p = BLKTOL1(blkno, sbi->l2nbperpage); /* L1 page */
3466
3467 /*
3468 * extend each L1 in L2
3469 */
3470 for (; k < LPERCTL; k++, p += nbperpage) {
3471 /* get L1 page */
3472 if (j0) {
3473 /* read in L1 page: (blkno & (MAXL1SIZE - 1)) */
3474 l1mp = read_metapage(ipbmap, p, PSIZE, 0);
3475 if (l1mp == NULL)
3476 goto errout;
3477 l1dcp = (struct dmapctl *) l1mp->data;
3478
3479 /* compute start L0 */
3480 j = (blkno & (MAXL1SIZE - 1)) >> L2MAXL0SIZE;
3481 l1leaf = l1dcp->stree + CTLLEAFIND + j;
3482 p = BLKTOL0(blkno, sbi->l2nbperpage);
3483 j0 = false;
3484 } else {
3485 /* assign/init L1 page */
3486 l1mp = get_metapage(ipbmap, p, PSIZE, 0);
3487 if (l1mp == NULL)
3488 goto errout;
3489
3490 l1dcp = (struct dmapctl *) l1mp->data;
3491
3492 /* compute start L0 */
3493 j = 0;
3494 l1leaf = l1dcp->stree + CTLLEAFIND;
3495 p += nbperpage; /* 1st L0 of L1.k */
3496 }
3497
3498 /*
3499 * extend each L0 in L1
3500 */
3501 for (; j < LPERCTL; j++) {
3502 /* get L0 page */
3503 if (i0) {
3504 /* read in L0 page: (blkno & (MAXL0SIZE - 1)) */
3505
3506 l0mp = read_metapage(ipbmap, p, PSIZE, 0);
3507 if (l0mp == NULL)
3508 goto errout;
3509 l0dcp = (struct dmapctl *) l0mp->data;
3510
3511 /* compute start dmap */
3512 i = (blkno & (MAXL0SIZE - 1)) >>
3513 L2BPERDMAP;
3514 l0leaf = l0dcp->stree + CTLLEAFIND + i;
3515 p = BLKTODMAP(blkno,
3516 sbi->l2nbperpage);
3517 i0 = false;
3518 } else {
3519 /* assign/init L0 page */
3520 l0mp = get_metapage(ipbmap, p, PSIZE, 0);
3521 if (l0mp == NULL)
3522 goto errout;
3523
3524 l0dcp = (struct dmapctl *) l0mp->data;
3525
3526 /* compute start dmap */
3527 i = 0;
3528 l0leaf = l0dcp->stree + CTLLEAFIND;
3529 p += nbperpage; /* 1st dmap of L0.j */
3530 }
3531
3532 /*
3533 * extend each dmap in L0
3534 */
3535 for (; i < LPERCTL; i++) {
3536 /*
3537 * reconstruct the dmap page, and
3538 * initialize corresponding parent L0 leaf
3539 */
3540 if ((n = blkno & (BPERDMAP - 1))) {
3541 /* read in dmap page: */
3542 mp = read_metapage(ipbmap, p,
3543 PSIZE, 0);
3544 if (mp == NULL)
3545 goto errout;
3546 n = min(nblocks, (s64)BPERDMAP - n);
3547 } else {
3548 /* assign/init dmap page */
3549 mp = read_metapage(ipbmap, p,
3550 PSIZE, 0);
3551 if (mp == NULL)
3552 goto errout;
3553
3554 n = min_t(s64, nblocks, BPERDMAP);
3555 }
3556
3557 dp = (struct dmap *) mp->data;
3558 *l0leaf = dbInitDmap(dp, blkno, n);
3559
3560 bmp->db_nfree += n;
3561 agno = le64_to_cpu(dp->start) >> l2agsize;
3562 bmp->db_agfree[agno] += n;
3563
3564 write_metapage(mp);
3565
3566 l0leaf++;
3567 p += nbperpage;
3568
3569 blkno += n;
3570 nblocks -= n;
3571 if (nblocks == 0)
3572 break;
3573 } /* for each dmap in a L0 */
3574
3575 /*
3576 * build current L0 page from its leaves, and
3577 * initialize corresponding parent L1 leaf
3578 */
3579 *l1leaf = dbInitDmapCtl(l0dcp, 0, ++i);
3580 write_metapage(l0mp);
3581 l0mp = NULL;
3582
3583 if (nblocks)
3584 l1leaf++; /* continue for next L0 */
3585 else {
3586 /* more than 1 L0 ? */
3587 if (j > 0)
3588 break; /* build L1 page */
3589 else {
3590 /* summarize in global bmap page */
3591 bmp->db_maxfreebud = *l1leaf;
3592 release_metapage(l1mp);
3593 release_metapage(l2mp);
3594 goto finalize;
3595 }
3596 }
3597 } /* for each L0 in a L1 */
3598
3599 /*
3600 * build current L1 page from its leaves, and
3601 * initialize corresponding parent L2 leaf
3602 */
3603 *l2leaf = dbInitDmapCtl(l1dcp, 1, ++j);
3604 write_metapage(l1mp);
3605 l1mp = NULL;
3606
3607 if (nblocks)
3608 l2leaf++; /* continue for next L1 */
3609 else {
3610 /* more than 1 L1 ? */
3611 if (k > 0)
3612 break; /* build L2 page */
3613 else {
3614 /* summarize in global bmap page */
3615 bmp->db_maxfreebud = *l2leaf;
3616 release_metapage(l2mp);
3617 goto finalize;
3618 }
3619 }
3620 } /* for each L1 in a L2 */
3621
3622 jfs_error(ipbmap->i_sb, "function has not returned as expected\n");
3623errout:
3624 if (l0mp)
3625 release_metapage(l0mp);
3626 if (l1mp)
3627 release_metapage(l1mp);
3628 release_metapage(l2mp);
3629 return -EIO;
3630
3631 /*
3632 * finalize bmap control page
3633 */
3634finalize:
3635
3636 return 0;
3637}
3638
3639
3640/*
3641 * dbFinalizeBmap()
3642 */
3643void dbFinalizeBmap(struct inode *ipbmap)
3644{
3645 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
3646 int actags, inactags, l2nl;
3647 s64 ag_rem, actfree, inactfree, avgfree;
3648 int i, n;
3649
3650 /*
3651 * finalize bmap control page
3652 */
3653//finalize:
3654 /*
3655 * compute db_agpref: preferred ag to allocate from
3656 * (the leftmost ag with average free space in it);
3657 */
3658//agpref:
3659 /* get the number of active ags and inacitve ags */
3660 actags = bmp->db_maxag + 1;
3661 inactags = bmp->db_numag - actags;
3662 ag_rem = bmp->db_mapsize & (bmp->db_agsize - 1); /* ??? */
3663
3664 /* determine how many blocks are in the inactive allocation
3665 * groups. in doing this, we must account for the fact that
3666 * the rightmost group might be a partial group (i.e. file
3667 * system size is not a multiple of the group size).
3668 */
3669 inactfree = (inactags && ag_rem) ?
3670 ((inactags - 1) << bmp->db_agl2size) + ag_rem
3671 : inactags << bmp->db_agl2size;
3672
3673 /* determine how many free blocks are in the active
3674 * allocation groups plus the average number of free blocks
3675 * within the active ags.
3676 */
3677 actfree = bmp->db_nfree - inactfree;
3678 avgfree = (u32) actfree / (u32) actags;
3679
3680 /* if the preferred allocation group has not average free space.
3681 * re-establish the preferred group as the leftmost
3682 * group with average free space.
3683 */
3684 if (bmp->db_agfree[bmp->db_agpref] < avgfree) {
3685 for (bmp->db_agpref = 0; bmp->db_agpref < actags;
3686 bmp->db_agpref++) {
3687 if (bmp->db_agfree[bmp->db_agpref] >= avgfree)
3688 break;
3689 }
3690 if (bmp->db_agpref >= bmp->db_numag) {
3691 jfs_error(ipbmap->i_sb,
3692 "cannot find ag with average freespace\n");
3693 }
3694 }
3695
3696 /*
3697 * compute db_aglevel, db_agheight, db_width, db_agstart:
3698 * an ag is covered in aglevel dmapctl summary tree,
3699 * at agheight level height (from leaf) with agwidth number of nodes
3700 * each, which starts at agstart index node of the smmary tree node
3701 * array;
3702 */
3703 bmp->db_aglevel = BMAPSZTOLEV(bmp->db_agsize);
3704 l2nl =
3705 bmp->db_agl2size - (L2BPERDMAP + bmp->db_aglevel * L2LPERCTL);
3706 bmp->db_agheight = l2nl >> 1;
3707 bmp->db_agwidth = 1 << (l2nl - (bmp->db_agheight << 1));
3708 for (i = 5 - bmp->db_agheight, bmp->db_agstart = 0, n = 1; i > 0;
3709 i--) {
3710 bmp->db_agstart += n;
3711 n <<= 2;
3712 }
3713
3714}
3715
3716
3717/*
3718 * NAME: dbInitDmap()/ujfs_idmap_page()
3719 *
3720 * FUNCTION: initialize working/persistent bitmap of the dmap page
3721 * for the specified number of blocks:
3722 *
3723 * at entry, the bitmaps had been initialized as free (ZEROS);
3724 * The number of blocks will only account for the actually
3725 * existing blocks. Blocks which don't actually exist in
3726 * the aggregate will be marked as allocated (ONES);
3727 *
3728 * PARAMETERS:
3729 * dp - pointer to page of map
3730 * nblocks - number of blocks this page
3731 *
3732 * RETURNS: NONE
3733 */
3734static int dbInitDmap(struct dmap * dp, s64 Blkno, int nblocks)
3735{
3736 int blkno, w, b, r, nw, nb, i;
3737
3738 /* starting block number within the dmap */
3739 blkno = Blkno & (BPERDMAP - 1);
3740
3741 if (blkno == 0) {
3742 dp->nblocks = dp->nfree = cpu_to_le32(nblocks);
3743 dp->start = cpu_to_le64(Blkno);
3744
3745 if (nblocks == BPERDMAP) {
3746 memset(&dp->wmap[0], 0, LPERDMAP * 4);
3747 memset(&dp->pmap[0], 0, LPERDMAP * 4);
3748 goto initTree;
3749 }
3750 } else {
3751 le32_add_cpu(&dp->nblocks, nblocks);
3752 le32_add_cpu(&dp->nfree, nblocks);
3753 }
3754
3755 /* word number containing start block number */
3756 w = blkno >> L2DBWORD;
3757
3758 /*
3759 * free the bits corresponding to the block range (ZEROS):
3760 * note: not all bits of the first and last words may be contained
3761 * within the block range.
3762 */
3763 for (r = nblocks; r > 0; r -= nb, blkno += nb) {
3764 /* number of bits preceding range to be freed in the word */
3765 b = blkno & (DBWORD - 1);
3766 /* number of bits to free in the word */
3767 nb = min(r, DBWORD - b);
3768
3769 /* is partial word to be freed ? */
3770 if (nb < DBWORD) {
3771 /* free (set to 0) from the bitmap word */
3772 dp->wmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3773 >> b));
3774 dp->pmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3775 >> b));
3776
3777 /* skip the word freed */
3778 w++;
3779 } else {
3780 /* free (set to 0) contiguous bitmap words */
3781 nw = r >> L2DBWORD;
3782 memset(&dp->wmap[w], 0, nw * 4);
3783 memset(&dp->pmap[w], 0, nw * 4);
3784
3785 /* skip the words freed */
3786 nb = nw << L2DBWORD;
3787 w += nw;
3788 }
3789 }
3790
3791 /*
3792 * mark bits following the range to be freed (non-existing
3793 * blocks) as allocated (ONES)
3794 */
3795
3796 if (blkno == BPERDMAP)
3797 goto initTree;
3798
3799 /* the first word beyond the end of existing blocks */
3800 w = blkno >> L2DBWORD;
3801
3802 /* does nblocks fall on a 32-bit boundary ? */
3803 b = blkno & (DBWORD - 1);
3804 if (b) {
3805 /* mark a partial word allocated */
3806 dp->wmap[w] = dp->pmap[w] = cpu_to_le32(ONES >> b);
3807 w++;
3808 }
3809
3810 /* set the rest of the words in the page to allocated (ONES) */
3811 for (i = w; i < LPERDMAP; i++)
3812 dp->pmap[i] = dp->wmap[i] = cpu_to_le32(ONES);
3813
3814 /*
3815 * init tree
3816 */
3817 initTree:
3818 return (dbInitDmapTree(dp));
3819}
3820
3821
3822/*
3823 * NAME: dbInitDmapTree()/ujfs_complete_dmap()
3824 *
3825 * FUNCTION: initialize summary tree of the specified dmap:
3826 *
3827 * at entry, bitmap of the dmap has been initialized;
3828 *
3829 * PARAMETERS:
3830 * dp - dmap to complete
3831 * blkno - starting block number for this dmap
3832 * treemax - will be filled in with max free for this dmap
3833 *
3834 * RETURNS: max free string at the root of the tree
3835 */
3836static int dbInitDmapTree(struct dmap * dp)
3837{
3838 struct dmaptree *tp;
3839 s8 *cp;
3840 int i;
3841
3842 /* init fixed info of tree */
3843 tp = &dp->tree;
3844 tp->nleafs = cpu_to_le32(LPERDMAP);
3845 tp->l2nleafs = cpu_to_le32(L2LPERDMAP);
3846 tp->leafidx = cpu_to_le32(LEAFIND);
3847 tp->height = cpu_to_le32(4);
3848 tp->budmin = BUDMIN;
3849
3850 /* init each leaf from corresponding wmap word:
3851 * note: leaf is set to NOFREE(-1) if all blocks of corresponding
3852 * bitmap word are allocated.
3853 */
3854 cp = tp->stree + le32_to_cpu(tp->leafidx);
3855 for (i = 0; i < LPERDMAP; i++)
3856 *cp++ = dbMaxBud((u8 *) & dp->wmap[i]);
3857
3858 /* build the dmap's binary buddy summary tree */
3859 return (dbInitTree(tp));
3860}
3861
3862
3863/*
3864 * NAME: dbInitTree()/ujfs_adjtree()
3865 *
3866 * FUNCTION: initialize binary buddy summary tree of a dmap or dmapctl.
3867 *
3868 * at entry, the leaves of the tree has been initialized
3869 * from corresponding bitmap word or root of summary tree
3870 * of the child control page;
3871 * configure binary buddy system at the leaf level, then
3872 * bubble up the values of the leaf nodes up the tree.
3873 *
3874 * PARAMETERS:
3875 * cp - Pointer to the root of the tree
3876 * l2leaves- Number of leaf nodes as a power of 2
3877 * l2min - Number of blocks that can be covered by a leaf
3878 * as a power of 2
3879 *
3880 * RETURNS: max free string at the root of the tree
3881 */
3882static int dbInitTree(struct dmaptree * dtp)
3883{
3884 int l2max, l2free, bsize, nextb, i;
3885 int child, parent, nparent;
3886 s8 *tp, *cp, *cp1;
3887
3888 tp = dtp->stree;
3889
3890 /* Determine the maximum free string possible for the leaves */
3891 l2max = le32_to_cpu(dtp->l2nleafs) + dtp->budmin;
3892
3893 /*
3894 * configure the leaf levevl into binary buddy system
3895 *
3896 * Try to combine buddies starting with a buddy size of 1
3897 * (i.e. two leaves). At a buddy size of 1 two buddy leaves
3898 * can be combined if both buddies have a maximum free of l2min;
3899 * the combination will result in the left-most buddy leaf having
3900 * a maximum free of l2min+1.
3901 * After processing all buddies for a given size, process buddies
3902 * at the next higher buddy size (i.e. current size * 2) and
3903 * the next maximum free (current free + 1).
3904 * This continues until the maximum possible buddy combination
3905 * yields maximum free.
3906 */
3907 for (l2free = dtp->budmin, bsize = 1; l2free < l2max;
3908 l2free++, bsize = nextb) {
3909 /* get next buddy size == current buddy pair size */
3910 nextb = bsize << 1;
3911
3912 /* scan each adjacent buddy pair at current buddy size */
3913 for (i = 0, cp = tp + le32_to_cpu(dtp->leafidx);
3914 i < le32_to_cpu(dtp->nleafs);
3915 i += nextb, cp += nextb) {
3916 /* coalesce if both adjacent buddies are max free */
3917 if (*cp == l2free && *(cp + bsize) == l2free) {
3918 *cp = l2free + 1; /* left take right */
3919 *(cp + bsize) = -1; /* right give left */
3920 }
3921 }
3922 }
3923
3924 /*
3925 * bubble summary information of leaves up the tree.
3926 *
3927 * Starting at the leaf node level, the four nodes described by
3928 * the higher level parent node are compared for a maximum free and
3929 * this maximum becomes the value of the parent node.
3930 * when all lower level nodes are processed in this fashion then
3931 * move up to the next level (parent becomes a lower level node) and
3932 * continue the process for that level.
3933 */
3934 for (child = le32_to_cpu(dtp->leafidx),
3935 nparent = le32_to_cpu(dtp->nleafs) >> 2;
3936 nparent > 0; nparent >>= 2, child = parent) {
3937 /* get index of 1st node of parent level */
3938 parent = (child - 1) >> 2;
3939
3940 /* set the value of the parent node as the maximum
3941 * of the four nodes of the current level.
3942 */
3943 for (i = 0, cp = tp + child, cp1 = tp + parent;
3944 i < nparent; i++, cp += 4, cp1++)
3945 *cp1 = TREEMAX(cp);
3946 }
3947
3948 return (*tp);
3949}
3950
3951
3952/*
3953 * dbInitDmapCtl()
3954 *
3955 * function: initialize dmapctl page
3956 */
3957static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i)
3958{ /* start leaf index not covered by range */
3959 s8 *cp;
3960
3961 dcp->nleafs = cpu_to_le32(LPERCTL);
3962 dcp->l2nleafs = cpu_to_le32(L2LPERCTL);
3963 dcp->leafidx = cpu_to_le32(CTLLEAFIND);
3964 dcp->height = cpu_to_le32(5);
3965 dcp->budmin = L2BPERDMAP + L2LPERCTL * level;
3966
3967 /*
3968 * initialize the leaves of current level that were not covered
3969 * by the specified input block range (i.e. the leaves have no
3970 * low level dmapctl or dmap).
3971 */
3972 cp = &dcp->stree[CTLLEAFIND + i];
3973 for (; i < LPERCTL; i++)
3974 *cp++ = NOFREE;
3975
3976 /* build the dmap's binary buddy summary tree */
3977 return (dbInitTree((struct dmaptree *) dcp));
3978}
3979
3980
3981/*
3982 * NAME: dbGetL2AGSize()/ujfs_getagl2size()
3983 *
3984 * FUNCTION: Determine log2(allocation group size) from aggregate size
3985 *
3986 * PARAMETERS:
3987 * nblocks - Number of blocks in aggregate
3988 *
3989 * RETURNS: log2(allocation group size) in aggregate blocks
3990 */
3991static int dbGetL2AGSize(s64 nblocks)
3992{
3993 s64 sz;
3994 s64 m;
3995 int l2sz;
3996
3997 if (nblocks < BPERDMAP * MAXAG)
3998 return (L2BPERDMAP);
3999
4000 /* round up aggregate size to power of 2 */
4001 m = ((u64) 1 << (64 - 1));
4002 for (l2sz = 64; l2sz >= 0; l2sz--, m >>= 1) {
4003 if (m & nblocks)
4004 break;
4005 }
4006
4007 sz = (s64) 1 << l2sz;
4008 if (sz < nblocks)
4009 l2sz += 1;
4010
4011 /* agsize = roundupSize/max_number_of_ag */
4012 return (l2sz - L2MAXAG);
4013}
4014
4015
4016/*
4017 * NAME: dbMapFileSizeToMapSize()
4018 *
4019 * FUNCTION: compute number of blocks the block allocation map file
4020 * can cover from the map file size;
4021 *
4022 * RETURNS: Number of blocks which can be covered by this block map file;
4023 */
4024
4025/*
4026 * maximum number of map pages at each level including control pages
4027 */
4028#define MAXL0PAGES (1 + LPERCTL)
4029#define MAXL1PAGES (1 + LPERCTL * MAXL0PAGES)
4030#define MAXL2PAGES (1 + LPERCTL * MAXL1PAGES)
4031
4032/*
4033 * convert number of map pages to the zero origin top dmapctl level
4034 */
4035#define BMAPPGTOLEV(npages) \
4036 (((npages) <= 3 + MAXL0PAGES) ? 0 : \
4037 ((npages) <= 2 + MAXL1PAGES) ? 1 : 2)
4038
4039s64 dbMapFileSizeToMapSize(struct inode * ipbmap)
4040{
4041 struct super_block *sb = ipbmap->i_sb;
4042 s64 nblocks;
4043 s64 npages, ndmaps;
4044 int level, i;
4045 int complete, factor;
4046
4047 nblocks = ipbmap->i_size >> JFS_SBI(sb)->l2bsize;
4048 npages = nblocks >> JFS_SBI(sb)->l2nbperpage;
4049 level = BMAPPGTOLEV(npages);
4050
4051 /* At each level, accumulate the number of dmap pages covered by
4052 * the number of full child levels below it;
4053 * repeat for the last incomplete child level.
4054 */
4055 ndmaps = 0;
4056 npages--; /* skip the first global control page */
4057 /* skip higher level control pages above top level covered by map */
4058 npages -= (2 - level);
4059 npages--; /* skip top level's control page */
4060 for (i = level; i >= 0; i--) {
4061 factor =
4062 (i == 2) ? MAXL1PAGES : ((i == 1) ? MAXL0PAGES : 1);
4063 complete = (u32) npages / factor;
4064 ndmaps += complete * ((i == 2) ? LPERCTL * LPERCTL :
4065 ((i == 1) ? LPERCTL : 1));
4066
4067 /* pages in last/incomplete child */
4068 npages = (u32) npages % factor;
4069 /* skip incomplete child's level control page */
4070 npages--;
4071 }
4072
4073 /* convert the number of dmaps into the number of blocks
4074 * which can be covered by the dmaps;
4075 */
4076 nblocks = ndmaps << L2BPERDMAP;
4077
4078 return (nblocks);
4079}
1/*
2 * Copyright (C) International Business Machines Corp., 2000-2004
3 * Portions Copyright (C) Tino Reichardt, 2012
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See
13 * the GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18 */
19
20#include <linux/fs.h>
21#include <linux/slab.h>
22#include "jfs_incore.h"
23#include "jfs_superblock.h"
24#include "jfs_dmap.h"
25#include "jfs_imap.h"
26#include "jfs_lock.h"
27#include "jfs_metapage.h"
28#include "jfs_debug.h"
29#include "jfs_discard.h"
30
31/*
32 * SERIALIZATION of the Block Allocation Map.
33 *
34 * the working state of the block allocation map is accessed in
35 * two directions:
36 *
37 * 1) allocation and free requests that start at the dmap
38 * level and move up through the dmap control pages (i.e.
39 * the vast majority of requests).
40 *
41 * 2) allocation requests that start at dmap control page
42 * level and work down towards the dmaps.
43 *
44 * the serialization scheme used here is as follows.
45 *
46 * requests which start at the bottom are serialized against each
47 * other through buffers and each requests holds onto its buffers
48 * as it works it way up from a single dmap to the required level
49 * of dmap control page.
50 * requests that start at the top are serialized against each other
51 * and request that start from the bottom by the multiple read/single
52 * write inode lock of the bmap inode. requests starting at the top
53 * take this lock in write mode while request starting at the bottom
54 * take the lock in read mode. a single top-down request may proceed
55 * exclusively while multiple bottoms-up requests may proceed
56 * simultaneously (under the protection of busy buffers).
57 *
58 * in addition to information found in dmaps and dmap control pages,
59 * the working state of the block allocation map also includes read/
60 * write information maintained in the bmap descriptor (i.e. total
61 * free block count, allocation group level free block counts).
62 * a single exclusive lock (BMAP_LOCK) is used to guard this information
63 * in the face of multiple-bottoms up requests.
64 * (lock ordering: IREAD_LOCK, BMAP_LOCK);
65 *
66 * accesses to the persistent state of the block allocation map (limited
67 * to the persistent bitmaps in dmaps) is guarded by (busy) buffers.
68 */
69
70#define BMAP_LOCK_INIT(bmp) mutex_init(&bmp->db_bmaplock)
71#define BMAP_LOCK(bmp) mutex_lock(&bmp->db_bmaplock)
72#define BMAP_UNLOCK(bmp) mutex_unlock(&bmp->db_bmaplock)
73
74/*
75 * forward references
76 */
77static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
78 int nblocks);
79static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval);
80static int dbBackSplit(dmtree_t * tp, int leafno);
81static int dbJoin(dmtree_t * tp, int leafno, int newval);
82static void dbAdjTree(dmtree_t * tp, int leafno, int newval);
83static int dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc,
84 int level);
85static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results);
86static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
87 int nblocks);
88static int dbAllocNear(struct bmap * bmp, struct dmap * dp, s64 blkno,
89 int nblocks,
90 int l2nb, s64 * results);
91static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
92 int nblocks);
93static int dbAllocDmapLev(struct bmap * bmp, struct dmap * dp, int nblocks,
94 int l2nb,
95 s64 * results);
96static int dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb,
97 s64 * results);
98static int dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno,
99 s64 * results);
100static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks);
101static int dbFindBits(u32 word, int l2nb);
102static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno);
103static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx);
104static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
105 int nblocks);
106static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
107 int nblocks);
108static int dbMaxBud(u8 * cp);
109static int blkstol2(s64 nb);
110
111static int cntlz(u32 value);
112static int cnttz(u32 word);
113
114static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
115 int nblocks);
116static int dbInitDmap(struct dmap * dp, s64 blkno, int nblocks);
117static int dbInitDmapTree(struct dmap * dp);
118static int dbInitTree(struct dmaptree * dtp);
119static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i);
120static int dbGetL2AGSize(s64 nblocks);
121
122/*
123 * buddy table
124 *
125 * table used for determining buddy sizes within characters of
126 * dmap bitmap words. the characters themselves serve as indexes
127 * into the table, with the table elements yielding the maximum
128 * binary buddy of free bits within the character.
129 */
130static const s8 budtab[256] = {
131 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
132 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
133 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
134 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
135 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
136 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
137 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
138 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
139 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
140 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
141 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
142 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
143 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
144 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
145 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
146 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, -1
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 * NAME: dbFree()
315 *
316 * FUNCTION: free the specified block range from the working block
317 * allocation map.
318 *
319 * the blocks will be free from the working map one dmap
320 * at a time.
321 *
322 * PARAMETERS:
323 * ip - pointer to in-core inode;
324 * blkno - starting block number to be freed.
325 * nblocks - number of blocks to be freed.
326 *
327 * RETURN VALUES:
328 * 0 - success
329 * -EIO - i/o error
330 */
331int dbFree(struct inode *ip, s64 blkno, s64 nblocks)
332{
333 struct metapage *mp;
334 struct dmap *dp;
335 int nb, rc;
336 s64 lblkno, rem;
337 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
338 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
339 struct super_block *sb = ipbmap->i_sb;
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, "block to be freed is outside the map\n");
350 return -EIO;
351 }
352
353 /**
354 * TRIM the blocks, when mounted with discard option
355 */
356 if (JFS_SBI(sb)->flag & JFS_DISCARD)
357 if (JFS_SBI(sb)->minblks_trim <= nblocks)
358 jfs_issue_discard(ipbmap, blkno, nblocks);
359
360 /*
361 * free the blocks a dmap at a time.
362 */
363 mp = NULL;
364 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
365 /* release previous dmap if any */
366 if (mp) {
367 write_metapage(mp);
368 }
369
370 /* get the buffer for the current dmap. */
371 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
372 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
373 if (mp == NULL) {
374 IREAD_UNLOCK(ipbmap);
375 return -EIO;
376 }
377 dp = (struct dmap *) mp->data;
378
379 /* determine the number of blocks to be freed from
380 * this dmap.
381 */
382 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
383
384 /* free the blocks. */
385 if ((rc = dbFreeDmap(bmp, dp, blkno, nb))) {
386 jfs_error(ip->i_sb, "error in block map\n");
387 release_metapage(mp);
388 IREAD_UNLOCK(ipbmap);
389 return (rc);
390 }
391 }
392
393 /* write the last buffer. */
394 write_metapage(mp);
395
396 IREAD_UNLOCK(ipbmap);
397
398 return (0);
399}
400
401
402/*
403 * NAME: dbUpdatePMap()
404 *
405 * FUNCTION: update the allocation state (free or allocate) of the
406 * specified block range in the persistent block allocation map.
407 *
408 * the blocks will be updated in the persistent map one
409 * dmap at a time.
410 *
411 * PARAMETERS:
412 * ipbmap - pointer to in-core inode for the block map.
413 * free - 'true' if block range is to be freed from the persistent
414 * map; 'false' if it is to be allocated.
415 * blkno - starting block number of the range.
416 * nblocks - number of contiguous blocks in the range.
417 * tblk - transaction block;
418 *
419 * RETURN VALUES:
420 * 0 - success
421 * -EIO - i/o error
422 */
423int
424dbUpdatePMap(struct inode *ipbmap,
425 int free, s64 blkno, s64 nblocks, struct tblock * tblk)
426{
427 int nblks, dbitno, wbitno, rbits;
428 int word, nbits, nwords;
429 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
430 s64 lblkno, rem, lastlblkno;
431 u32 mask;
432 struct dmap *dp;
433 struct metapage *mp;
434 struct jfs_log *log;
435 int lsn, difft, diffp;
436 unsigned long flags;
437
438 /* the blocks better be within the mapsize. */
439 if (blkno + nblocks > bmp->db_mapsize) {
440 printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
441 (unsigned long long) blkno,
442 (unsigned long long) nblocks);
443 jfs_error(ipbmap->i_sb, "blocks are outside the map\n");
444 return -EIO;
445 }
446
447 /* compute delta of transaction lsn from log syncpt */
448 lsn = tblk->lsn;
449 log = (struct jfs_log *) JFS_SBI(tblk->sb)->log;
450 logdiff(difft, lsn, log);
451
452 /*
453 * update the block state a dmap at a time.
454 */
455 mp = NULL;
456 lastlblkno = 0;
457 for (rem = nblocks; rem > 0; rem -= nblks, blkno += nblks) {
458 /* get the buffer for the current dmap. */
459 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
460 if (lblkno != lastlblkno) {
461 if (mp) {
462 write_metapage(mp);
463 }
464
465 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE,
466 0);
467 if (mp == NULL)
468 return -EIO;
469 metapage_wait_for_io(mp);
470 }
471 dp = (struct dmap *) mp->data;
472
473 /* determine the bit number and word within the dmap of
474 * the starting block. also determine how many blocks
475 * are to be updated within this dmap.
476 */
477 dbitno = blkno & (BPERDMAP - 1);
478 word = dbitno >> L2DBWORD;
479 nblks = min(rem, (s64)BPERDMAP - dbitno);
480
481 /* update the bits of the dmap words. the first and last
482 * words may only have a subset of their bits updated. if
483 * this is the case, we'll work against that word (i.e.
484 * partial first and/or last) only in a single pass. a
485 * single pass will also be used to update all words that
486 * are to have all their bits updated.
487 */
488 for (rbits = nblks; rbits > 0;
489 rbits -= nbits, dbitno += nbits) {
490 /* determine the bit number within the word and
491 * the number of bits within the word.
492 */
493 wbitno = dbitno & (DBWORD - 1);
494 nbits = min(rbits, DBWORD - wbitno);
495
496 /* check if only part of the word is to be updated. */
497 if (nbits < DBWORD) {
498 /* update (free or allocate) the bits
499 * in this word.
500 */
501 mask =
502 (ONES << (DBWORD - nbits) >> wbitno);
503 if (free)
504 dp->pmap[word] &=
505 cpu_to_le32(~mask);
506 else
507 dp->pmap[word] |=
508 cpu_to_le32(mask);
509
510 word += 1;
511 } else {
512 /* one or more words are to have all
513 * their bits updated. determine how
514 * many words and how many bits.
515 */
516 nwords = rbits >> L2DBWORD;
517 nbits = nwords << L2DBWORD;
518
519 /* update (free or allocate) the bits
520 * in these words.
521 */
522 if (free)
523 memset(&dp->pmap[word], 0,
524 nwords * 4);
525 else
526 memset(&dp->pmap[word], (int) ONES,
527 nwords * 4);
528
529 word += nwords;
530 }
531 }
532
533 /*
534 * update dmap lsn
535 */
536 if (lblkno == lastlblkno)
537 continue;
538
539 lastlblkno = lblkno;
540
541 LOGSYNC_LOCK(log, flags);
542 if (mp->lsn != 0) {
543 /* inherit older/smaller lsn */
544 logdiff(diffp, mp->lsn, log);
545 if (difft < diffp) {
546 mp->lsn = lsn;
547
548 /* move bp after tblock in logsync list */
549 list_move(&mp->synclist, &tblk->synclist);
550 }
551
552 /* inherit younger/larger clsn */
553 logdiff(difft, tblk->clsn, log);
554 logdiff(diffp, mp->clsn, log);
555 if (difft > diffp)
556 mp->clsn = tblk->clsn;
557 } else {
558 mp->log = log;
559 mp->lsn = lsn;
560
561 /* insert bp after tblock in logsync list */
562 log->count++;
563 list_add(&mp->synclist, &tblk->synclist);
564
565 mp->clsn = tblk->clsn;
566 }
567 LOGSYNC_UNLOCK(log, flags);
568 }
569
570 /* write the last buffer. */
571 if (mp) {
572 write_metapage(mp);
573 }
574
575 return (0);
576}
577
578
579/*
580 * NAME: dbNextAG()
581 *
582 * FUNCTION: find the preferred allocation group for new allocations.
583 *
584 * Within the allocation groups, we maintain a preferred
585 * allocation group which consists of a group with at least
586 * average free space. It is the preferred group that we target
587 * new inode allocation towards. The tie-in between inode
588 * allocation and block allocation occurs as we allocate the
589 * first (data) block of an inode and specify the inode (block)
590 * as the allocation hint for this block.
591 *
592 * We try to avoid having more than one open file growing in
593 * an allocation group, as this will lead to fragmentation.
594 * This differs from the old OS/2 method of trying to keep
595 * empty ags around for large allocations.
596 *
597 * PARAMETERS:
598 * ipbmap - pointer to in-core inode for the block map.
599 *
600 * RETURN VALUES:
601 * the preferred allocation group number.
602 */
603int dbNextAG(struct inode *ipbmap)
604{
605 s64 avgfree;
606 int agpref;
607 s64 hwm = 0;
608 int i;
609 int next_best = -1;
610 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
611
612 BMAP_LOCK(bmp);
613
614 /* determine the average number of free blocks within the ags. */
615 avgfree = (u32)bmp->db_nfree / bmp->db_numag;
616
617 /*
618 * if the current preferred ag does not have an active allocator
619 * and has at least average freespace, return it
620 */
621 agpref = bmp->db_agpref;
622 if ((atomic_read(&bmp->db_active[agpref]) == 0) &&
623 (bmp->db_agfree[agpref] >= avgfree))
624 goto unlock;
625
626 /* From the last preferred ag, find the next one with at least
627 * average free space.
628 */
629 for (i = 0 ; i < bmp->db_numag; i++, agpref++) {
630 if (agpref == bmp->db_numag)
631 agpref = 0;
632
633 if (atomic_read(&bmp->db_active[agpref]))
634 /* open file is currently growing in this ag */
635 continue;
636 if (bmp->db_agfree[agpref] >= avgfree) {
637 /* Return this one */
638 bmp->db_agpref = agpref;
639 goto unlock;
640 } else if (bmp->db_agfree[agpref] > hwm) {
641 /* Less than avg. freespace, but best so far */
642 hwm = bmp->db_agfree[agpref];
643 next_best = agpref;
644 }
645 }
646
647 /*
648 * If no inactive ag was found with average freespace, use the
649 * next best
650 */
651 if (next_best != -1)
652 bmp->db_agpref = next_best;
653 /* else leave db_agpref unchanged */
654unlock:
655 BMAP_UNLOCK(bmp);
656
657 /* return the preferred group.
658 */
659 return (bmp->db_agpref);
660}
661
662/*
663 * NAME: dbAlloc()
664 *
665 * FUNCTION: attempt to allocate a specified number of contiguous free
666 * blocks from the working allocation block map.
667 *
668 * the block allocation policy uses hints and a multi-step
669 * approach.
670 *
671 * for allocation requests smaller than the number of blocks
672 * per dmap, we first try to allocate the new blocks
673 * immediately following the hint. if these blocks are not
674 * available, we try to allocate blocks near the hint. if
675 * no blocks near the hint are available, we next try to
676 * allocate within the same dmap as contains the hint.
677 *
678 * if no blocks are available in the dmap or the allocation
679 * request is larger than the dmap size, we try to allocate
680 * within the same allocation group as contains the hint. if
681 * this does not succeed, we finally try to allocate anywhere
682 * within the aggregate.
683 *
684 * we also try to allocate anywhere within the aggregate for
685 * for allocation requests larger than the allocation group
686 * size or requests that specify no hint value.
687 *
688 * PARAMETERS:
689 * ip - pointer to in-core inode;
690 * hint - allocation hint.
691 * nblocks - number of contiguous blocks in the range.
692 * results - on successful return, set to the starting block number
693 * of the newly allocated contiguous range.
694 *
695 * RETURN VALUES:
696 * 0 - success
697 * -ENOSPC - insufficient disk resources
698 * -EIO - i/o error
699 */
700int dbAlloc(struct inode *ip, s64 hint, s64 nblocks, s64 * results)
701{
702 int rc, agno;
703 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
704 struct bmap *bmp;
705 struct metapage *mp;
706 s64 lblkno, blkno;
707 struct dmap *dp;
708 int l2nb;
709 s64 mapSize;
710 int writers;
711
712 /* assert that nblocks is valid */
713 assert(nblocks > 0);
714
715 /* get the log2 number of blocks to be allocated.
716 * if the number of blocks is not a log2 multiple,
717 * it will be rounded up to the next log2 multiple.
718 */
719 l2nb = BLKSTOL2(nblocks);
720
721 bmp = JFS_SBI(ip->i_sb)->bmap;
722
723 mapSize = bmp->db_mapsize;
724
725 /* the hint should be within the map */
726 if (hint >= mapSize) {
727 jfs_error(ip->i_sb, "the hint is outside the map\n");
728 return -EIO;
729 }
730
731 /* if the number of blocks to be allocated is greater than the
732 * allocation group size, try to allocate anywhere.
733 */
734 if (l2nb > bmp->db_agl2size) {
735 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
736
737 rc = dbAllocAny(bmp, nblocks, l2nb, results);
738
739 goto write_unlock;
740 }
741
742 /*
743 * If no hint, let dbNextAG recommend an allocation group
744 */
745 if (hint == 0)
746 goto pref_ag;
747
748 /* we would like to allocate close to the hint. adjust the
749 * hint to the block following the hint since the allocators
750 * will start looking for free space starting at this point.
751 */
752 blkno = hint + 1;
753
754 if (blkno >= bmp->db_mapsize)
755 goto pref_ag;
756
757 agno = blkno >> bmp->db_agl2size;
758
759 /* check if blkno crosses over into a new allocation group.
760 * if so, check if we should allow allocations within this
761 * allocation group.
762 */
763 if ((blkno & (bmp->db_agsize - 1)) == 0)
764 /* check if the AG is currently being written to.
765 * if so, call dbNextAG() to find a non-busy
766 * AG with sufficient free space.
767 */
768 if (atomic_read(&bmp->db_active[agno]))
769 goto pref_ag;
770
771 /* check if the allocation request size can be satisfied from a
772 * single dmap. if so, try to allocate from the dmap containing
773 * the hint using a tiered strategy.
774 */
775 if (nblocks <= BPERDMAP) {
776 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
777
778 /* get the buffer for the dmap containing the hint.
779 */
780 rc = -EIO;
781 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
782 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
783 if (mp == NULL)
784 goto read_unlock;
785
786 dp = (struct dmap *) mp->data;
787
788 /* first, try to satisfy the allocation request with the
789 * blocks beginning at the hint.
790 */
791 if ((rc = dbAllocNext(bmp, dp, blkno, (int) nblocks))
792 != -ENOSPC) {
793 if (rc == 0) {
794 *results = blkno;
795 mark_metapage_dirty(mp);
796 }
797
798 release_metapage(mp);
799 goto read_unlock;
800 }
801
802 writers = atomic_read(&bmp->db_active[agno]);
803 if ((writers > 1) ||
804 ((writers == 1) && (JFS_IP(ip)->active_ag != agno))) {
805 /*
806 * Someone else is writing in this allocation
807 * group. To avoid fragmenting, try another ag
808 */
809 release_metapage(mp);
810 IREAD_UNLOCK(ipbmap);
811 goto pref_ag;
812 }
813
814 /* next, try to satisfy the allocation request with blocks
815 * near the hint.
816 */
817 if ((rc =
818 dbAllocNear(bmp, dp, blkno, (int) nblocks, l2nb, results))
819 != -ENOSPC) {
820 if (rc == 0)
821 mark_metapage_dirty(mp);
822
823 release_metapage(mp);
824 goto read_unlock;
825 }
826
827 /* try to satisfy the allocation request with blocks within
828 * the same dmap as the hint.
829 */
830 if ((rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results))
831 != -ENOSPC) {
832 if (rc == 0)
833 mark_metapage_dirty(mp);
834
835 release_metapage(mp);
836 goto read_unlock;
837 }
838
839 release_metapage(mp);
840 IREAD_UNLOCK(ipbmap);
841 }
842
843 /* try to satisfy the allocation request with blocks within
844 * the same allocation group as the hint.
845 */
846 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
847 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) != -ENOSPC)
848 goto write_unlock;
849
850 IWRITE_UNLOCK(ipbmap);
851
852
853 pref_ag:
854 /*
855 * Let dbNextAG recommend a preferred allocation group
856 */
857 agno = dbNextAG(ipbmap);
858 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
859
860 /* Try to allocate within this allocation group. if that fails, try to
861 * allocate anywhere in the map.
862 */
863 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) == -ENOSPC)
864 rc = dbAllocAny(bmp, nblocks, l2nb, results);
865
866 write_unlock:
867 IWRITE_UNLOCK(ipbmap);
868
869 return (rc);
870
871 read_unlock:
872 IREAD_UNLOCK(ipbmap);
873
874 return (rc);
875}
876
877#ifdef _NOTYET
878/*
879 * NAME: dbAllocExact()
880 *
881 * FUNCTION: try to allocate the requested extent;
882 *
883 * PARAMETERS:
884 * ip - pointer to in-core inode;
885 * blkno - extent address;
886 * nblocks - extent length;
887 *
888 * RETURN VALUES:
889 * 0 - success
890 * -ENOSPC - insufficient disk resources
891 * -EIO - i/o error
892 */
893int dbAllocExact(struct inode *ip, s64 blkno, int nblocks)
894{
895 int rc;
896 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
897 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
898 struct dmap *dp;
899 s64 lblkno;
900 struct metapage *mp;
901
902 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
903
904 /*
905 * validate extent request:
906 *
907 * note: defragfs policy:
908 * max 64 blocks will be moved.
909 * allocation request size must be satisfied from a single dmap.
910 */
911 if (nblocks <= 0 || nblocks > BPERDMAP || blkno >= bmp->db_mapsize) {
912 IREAD_UNLOCK(ipbmap);
913 return -EINVAL;
914 }
915
916 if (nblocks > ((s64) 1 << bmp->db_maxfreebud)) {
917 /* the free space is no longer available */
918 IREAD_UNLOCK(ipbmap);
919 return -ENOSPC;
920 }
921
922 /* read in the dmap covering the extent */
923 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
924 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
925 if (mp == NULL) {
926 IREAD_UNLOCK(ipbmap);
927 return -EIO;
928 }
929 dp = (struct dmap *) mp->data;
930
931 /* try to allocate the requested extent */
932 rc = dbAllocNext(bmp, dp, blkno, nblocks);
933
934 IREAD_UNLOCK(ipbmap);
935
936 if (rc == 0)
937 mark_metapage_dirty(mp);
938
939 release_metapage(mp);
940
941 return (rc);
942}
943#endif /* _NOTYET */
944
945/*
946 * NAME: dbReAlloc()
947 *
948 * FUNCTION: attempt to extend a current allocation by a specified
949 * number of blocks.
950 *
951 * this routine attempts to satisfy the allocation request
952 * by first trying to extend the existing allocation in
953 * place by allocating the additional blocks as the blocks
954 * immediately following the current allocation. if these
955 * blocks are not available, this routine will attempt to
956 * allocate a new set of contiguous blocks large enough
957 * to cover the existing allocation plus the additional
958 * number of blocks required.
959 *
960 * PARAMETERS:
961 * ip - pointer to in-core inode requiring allocation.
962 * blkno - starting block of the current allocation.
963 * nblocks - number of contiguous blocks within the current
964 * allocation.
965 * addnblocks - number of blocks to add to the allocation.
966 * results - on successful return, set to the starting block number
967 * of the existing allocation if the existing allocation
968 * was extended in place or to a newly allocated contiguous
969 * range if the existing allocation could not be extended
970 * in place.
971 *
972 * RETURN VALUES:
973 * 0 - success
974 * -ENOSPC - insufficient disk resources
975 * -EIO - i/o error
976 */
977int
978dbReAlloc(struct inode *ip,
979 s64 blkno, s64 nblocks, s64 addnblocks, s64 * results)
980{
981 int rc;
982
983 /* try to extend the allocation in place.
984 */
985 if ((rc = dbExtend(ip, blkno, nblocks, addnblocks)) == 0) {
986 *results = blkno;
987 return (0);
988 } else {
989 if (rc != -ENOSPC)
990 return (rc);
991 }
992
993 /* could not extend the allocation in place, so allocate a
994 * new set of blocks for the entire request (i.e. try to get
995 * a range of contiguous blocks large enough to cover the
996 * existing allocation plus the additional blocks.)
997 */
998 return (dbAlloc
999 (ip, blkno + nblocks - 1, addnblocks + nblocks, results));
1000}
1001
1002
1003/*
1004 * NAME: dbExtend()
1005 *
1006 * FUNCTION: attempt to extend a current allocation by a specified
1007 * number of blocks.
1008 *
1009 * this routine attempts to satisfy the allocation request
1010 * by first trying to extend the existing allocation in
1011 * place by allocating the additional blocks as the blocks
1012 * immediately following the current allocation.
1013 *
1014 * PARAMETERS:
1015 * ip - pointer to in-core inode requiring allocation.
1016 * blkno - starting block of the current allocation.
1017 * nblocks - number of contiguous blocks within the current
1018 * allocation.
1019 * addnblocks - number of blocks to add to the allocation.
1020 *
1021 * RETURN VALUES:
1022 * 0 - success
1023 * -ENOSPC - insufficient disk resources
1024 * -EIO - i/o error
1025 */
1026static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks)
1027{
1028 struct jfs_sb_info *sbi = JFS_SBI(ip->i_sb);
1029 s64 lblkno, lastblkno, extblkno;
1030 uint rel_block;
1031 struct metapage *mp;
1032 struct dmap *dp;
1033 int rc;
1034 struct inode *ipbmap = sbi->ipbmap;
1035 struct bmap *bmp;
1036
1037 /*
1038 * We don't want a non-aligned extent to cross a page boundary
1039 */
1040 if (((rel_block = blkno & (sbi->nbperpage - 1))) &&
1041 (rel_block + nblocks + addnblocks > sbi->nbperpage))
1042 return -ENOSPC;
1043
1044 /* get the last block of the current allocation */
1045 lastblkno = blkno + nblocks - 1;
1046
1047 /* determine the block number of the block following
1048 * the existing allocation.
1049 */
1050 extblkno = lastblkno + 1;
1051
1052 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
1053
1054 /* better be within the file system */
1055 bmp = sbi->bmap;
1056 if (lastblkno < 0 || lastblkno >= bmp->db_mapsize) {
1057 IREAD_UNLOCK(ipbmap);
1058 jfs_error(ip->i_sb, "the block is outside the filesystem\n");
1059 return -EIO;
1060 }
1061
1062 /* we'll attempt to extend the current allocation in place by
1063 * allocating the additional blocks as the blocks immediately
1064 * following the current allocation. we only try to extend the
1065 * current allocation in place if the number of additional blocks
1066 * can fit into a dmap, the last block of the current allocation
1067 * is not the last block of the file system, and the start of the
1068 * inplace extension is not on an allocation group boundary.
1069 */
1070 if (addnblocks > BPERDMAP || extblkno >= bmp->db_mapsize ||
1071 (extblkno & (bmp->db_agsize - 1)) == 0) {
1072 IREAD_UNLOCK(ipbmap);
1073 return -ENOSPC;
1074 }
1075
1076 /* get the buffer for the dmap containing the first block
1077 * of the extension.
1078 */
1079 lblkno = BLKTODMAP(extblkno, bmp->db_l2nbperpage);
1080 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
1081 if (mp == NULL) {
1082 IREAD_UNLOCK(ipbmap);
1083 return -EIO;
1084 }
1085
1086 dp = (struct dmap *) mp->data;
1087
1088 /* try to allocate the blocks immediately following the
1089 * current allocation.
1090 */
1091 rc = dbAllocNext(bmp, dp, extblkno, (int) addnblocks);
1092
1093 IREAD_UNLOCK(ipbmap);
1094
1095 /* were we successful ? */
1096 if (rc == 0)
1097 write_metapage(mp);
1098 else
1099 /* we were not successful */
1100 release_metapage(mp);
1101
1102 return (rc);
1103}
1104
1105
1106/*
1107 * NAME: dbAllocNext()
1108 *
1109 * FUNCTION: attempt to allocate the blocks of the specified block
1110 * range within a dmap.
1111 *
1112 * PARAMETERS:
1113 * bmp - pointer to bmap descriptor
1114 * dp - pointer to dmap.
1115 * blkno - starting block number of the range.
1116 * nblocks - number of contiguous free blocks of the range.
1117 *
1118 * RETURN VALUES:
1119 * 0 - success
1120 * -ENOSPC - insufficient disk resources
1121 * -EIO - i/o error
1122 *
1123 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1124 */
1125static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
1126 int nblocks)
1127{
1128 int dbitno, word, rembits, nb, nwords, wbitno, nw;
1129 int l2size;
1130 s8 *leaf;
1131 u32 mask;
1132
1133 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1134 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1135 return -EIO;
1136 }
1137
1138 /* pick up a pointer to the leaves of the dmap tree.
1139 */
1140 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1141
1142 /* determine the bit number and word within the dmap of the
1143 * starting block.
1144 */
1145 dbitno = blkno & (BPERDMAP - 1);
1146 word = dbitno >> L2DBWORD;
1147
1148 /* check if the specified block range is contained within
1149 * this dmap.
1150 */
1151 if (dbitno + nblocks > BPERDMAP)
1152 return -ENOSPC;
1153
1154 /* check if the starting leaf indicates that anything
1155 * is free.
1156 */
1157 if (leaf[word] == NOFREE)
1158 return -ENOSPC;
1159
1160 /* check the dmaps words corresponding to block range to see
1161 * if the block range is free. not all bits of the first and
1162 * last words may be contained within the block range. if this
1163 * is the case, we'll work against those words (i.e. partial first
1164 * and/or last) on an individual basis (a single pass) and examine
1165 * the actual bits to determine if they are free. a single pass
1166 * will be used for all dmap words fully contained within the
1167 * specified range. within this pass, the leaves of the dmap
1168 * tree will be examined to determine if the blocks are free. a
1169 * single leaf may describe the free space of multiple dmap
1170 * words, so we may visit only a subset of the actual leaves
1171 * corresponding to the dmap words of the block range.
1172 */
1173 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
1174 /* determine the bit number within the word and
1175 * the number of bits within the word.
1176 */
1177 wbitno = dbitno & (DBWORD - 1);
1178 nb = min(rembits, DBWORD - wbitno);
1179
1180 /* check if only part of the word is to be examined.
1181 */
1182 if (nb < DBWORD) {
1183 /* check if the bits are free.
1184 */
1185 mask = (ONES << (DBWORD - nb) >> wbitno);
1186 if ((mask & ~le32_to_cpu(dp->wmap[word])) != mask)
1187 return -ENOSPC;
1188
1189 word += 1;
1190 } else {
1191 /* one or more dmap words are fully contained
1192 * within the block range. determine how many
1193 * words and how many bits.
1194 */
1195 nwords = rembits >> L2DBWORD;
1196 nb = nwords << L2DBWORD;
1197
1198 /* now examine the appropriate leaves to determine
1199 * if the blocks are free.
1200 */
1201 while (nwords > 0) {
1202 /* does the leaf describe any free space ?
1203 */
1204 if (leaf[word] < BUDMIN)
1205 return -ENOSPC;
1206
1207 /* determine the l2 number of bits provided
1208 * by this leaf.
1209 */
1210 l2size =
1211 min((int)leaf[word], NLSTOL2BSZ(nwords));
1212
1213 /* determine how many words were handled.
1214 */
1215 nw = BUDSIZE(l2size, BUDMIN);
1216
1217 nwords -= nw;
1218 word += nw;
1219 }
1220 }
1221 }
1222
1223 /* allocate the blocks.
1224 */
1225 return (dbAllocDmap(bmp, dp, blkno, nblocks));
1226}
1227
1228
1229/*
1230 * NAME: dbAllocNear()
1231 *
1232 * FUNCTION: attempt to allocate a number of contiguous free blocks near
1233 * a specified block (hint) within a dmap.
1234 *
1235 * starting with the dmap leaf that covers the hint, we'll
1236 * check the next four contiguous leaves for sufficient free
1237 * space. if sufficient free space is found, we'll allocate
1238 * the desired free space.
1239 *
1240 * PARAMETERS:
1241 * bmp - pointer to bmap descriptor
1242 * dp - pointer to dmap.
1243 * blkno - block number to allocate near.
1244 * nblocks - actual number of contiguous free blocks desired.
1245 * l2nb - log2 number of contiguous free blocks desired.
1246 * results - on successful return, set to the starting block number
1247 * of the newly allocated range.
1248 *
1249 * RETURN VALUES:
1250 * 0 - success
1251 * -ENOSPC - insufficient disk resources
1252 * -EIO - i/o error
1253 *
1254 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1255 */
1256static int
1257dbAllocNear(struct bmap * bmp,
1258 struct dmap * dp, s64 blkno, int nblocks, int l2nb, s64 * results)
1259{
1260 int word, lword, rc;
1261 s8 *leaf;
1262
1263 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1264 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1265 return -EIO;
1266 }
1267
1268 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1269
1270 /* determine the word within the dmap that holds the hint
1271 * (i.e. blkno). also, determine the last word in the dmap
1272 * that we'll include in our examination.
1273 */
1274 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
1275 lword = min(word + 4, LPERDMAP);
1276
1277 /* examine the leaves for sufficient free space.
1278 */
1279 for (; word < lword; word++) {
1280 /* does the leaf describe sufficient free space ?
1281 */
1282 if (leaf[word] < l2nb)
1283 continue;
1284
1285 /* determine the block number within the file system
1286 * of the first block described by this dmap word.
1287 */
1288 blkno = le64_to_cpu(dp->start) + (word << L2DBWORD);
1289
1290 /* if not all bits of the dmap word are free, get the
1291 * starting bit number within the dmap word of the required
1292 * string of free bits and adjust the block number with the
1293 * value.
1294 */
1295 if (leaf[word] < BUDMIN)
1296 blkno +=
1297 dbFindBits(le32_to_cpu(dp->wmap[word]), l2nb);
1298
1299 /* allocate the blocks.
1300 */
1301 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1302 *results = blkno;
1303
1304 return (rc);
1305 }
1306
1307 return -ENOSPC;
1308}
1309
1310
1311/*
1312 * NAME: dbAllocAG()
1313 *
1314 * FUNCTION: attempt to allocate the specified number of contiguous
1315 * free blocks within the specified allocation group.
1316 *
1317 * unless the allocation group size is equal to the number
1318 * of blocks per dmap, the dmap control pages will be used to
1319 * find the required free space, if available. we start the
1320 * search at the highest dmap control page level which
1321 * distinctly describes the allocation group's free space
1322 * (i.e. the highest level at which the allocation group's
1323 * free space is not mixed in with that of any other group).
1324 * in addition, we start the search within this level at a
1325 * height of the dmapctl dmtree at which the nodes distinctly
1326 * describe the allocation group's free space. at this height,
1327 * the allocation group's free space may be represented by 1
1328 * or two sub-trees, depending on the allocation group size.
1329 * we search the top nodes of these subtrees left to right for
1330 * sufficient free space. if sufficient free space is found,
1331 * the subtree is searched to find the leftmost leaf that
1332 * has free space. once we have made it to the leaf, we
1333 * move the search to the next lower level dmap control page
1334 * corresponding to this leaf. we continue down the dmap control
1335 * pages until we find the dmap that contains or starts the
1336 * sufficient free space and we allocate at this dmap.
1337 *
1338 * if the allocation group size is equal to the dmap size,
1339 * we'll start at the dmap corresponding to the allocation
1340 * group and attempt the allocation at this level.
1341 *
1342 * the dmap control page search is also not performed if the
1343 * allocation group is completely free and we go to the first
1344 * dmap of the allocation group to do the allocation. this is
1345 * done because the allocation group may be part (not the first
1346 * part) of a larger binary buddy system, causing the dmap
1347 * control pages to indicate no free space (NOFREE) within
1348 * the allocation group.
1349 *
1350 * PARAMETERS:
1351 * bmp - pointer to bmap descriptor
1352 * agno - allocation group number.
1353 * nblocks - actual number of contiguous free blocks desired.
1354 * l2nb - log2 number of contiguous free blocks desired.
1355 * results - on successful return, set to the starting block number
1356 * of the newly allocated range.
1357 *
1358 * RETURN VALUES:
1359 * 0 - success
1360 * -ENOSPC - insufficient disk resources
1361 * -EIO - i/o error
1362 *
1363 * note: IWRITE_LOCK(ipmap) held on entry/exit;
1364 */
1365static int
1366dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, s64 * results)
1367{
1368 struct metapage *mp;
1369 struct dmapctl *dcp;
1370 int rc, ti, i, k, m, n, agperlev;
1371 s64 blkno, lblkno;
1372 int budmin;
1373
1374 /* allocation request should not be for more than the
1375 * allocation group size.
1376 */
1377 if (l2nb > bmp->db_agl2size) {
1378 jfs_error(bmp->db_ipbmap->i_sb,
1379 "allocation request is larger than the allocation group size\n");
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 "dbAllocCtl failed in free AG\n");
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, "Corrupt dmapctl page\n");
1431 release_metapage(mp);
1432 return -EIO;
1433 }
1434
1435 /* search the subtree(s) of the dmap control page that describes
1436 * the allocation group, looking for sufficient free space. to begin,
1437 * determine how many allocation groups are represented in a dmap
1438 * control page at the control page level (i.e. L0, L1, L2) that
1439 * fully describes an allocation group. next, determine the starting
1440 * tree index of this allocation group within the control page.
1441 */
1442 agperlev =
1443 (1 << (L2LPERCTL - (bmp->db_agheight << 1))) / bmp->db_agwidth;
1444 ti = bmp->db_agstart + bmp->db_agwidth * (agno & (agperlev - 1));
1445
1446 /* dmap control page trees fan-out by 4 and a single allocation
1447 * group may be described by 1 or 2 subtrees within the ag level
1448 * dmap control page, depending upon the ag size. examine the ag's
1449 * subtrees for sufficient free space, starting with the leftmost
1450 * subtree.
1451 */
1452 for (i = 0; i < bmp->db_agwidth; i++, ti++) {
1453 /* is there sufficient free space ?
1454 */
1455 if (l2nb > dcp->stree[ti])
1456 continue;
1457
1458 /* sufficient free space found in a subtree. now search down
1459 * the subtree to find the leftmost leaf that describes this
1460 * free space.
1461 */
1462 for (k = bmp->db_agheight; k > 0; k--) {
1463 for (n = 0, m = (ti << 2) + 1; n < 4; n++) {
1464 if (l2nb <= dcp->stree[m + n]) {
1465 ti = m + n;
1466 break;
1467 }
1468 }
1469 if (n == 4) {
1470 jfs_error(bmp->db_ipbmap->i_sb,
1471 "failed descending stree\n");
1472 release_metapage(mp);
1473 return -EIO;
1474 }
1475 }
1476
1477 /* determine the block number within the file system
1478 * that corresponds to this leaf.
1479 */
1480 if (bmp->db_aglevel == 2)
1481 blkno = 0;
1482 else if (bmp->db_aglevel == 1)
1483 blkno &= ~(MAXL1SIZE - 1);
1484 else /* bmp->db_aglevel == 0 */
1485 blkno &= ~(MAXL0SIZE - 1);
1486
1487 blkno +=
1488 ((s64) (ti - le32_to_cpu(dcp->leafidx))) << budmin;
1489
1490 /* release the buffer in preparation for going down
1491 * the next level of dmap control pages.
1492 */
1493 release_metapage(mp);
1494
1495 /* check if we need to continue to search down the lower
1496 * level dmap control pages. we need to if the number of
1497 * blocks required is less than maximum number of blocks
1498 * described at the next lower level.
1499 */
1500 if (l2nb < budmin) {
1501
1502 /* search the lower level dmap control pages to get
1503 * the starting block number of the dmap that
1504 * contains or starts off the free space.
1505 */
1506 if ((rc =
1507 dbFindCtl(bmp, l2nb, bmp->db_aglevel - 1,
1508 &blkno))) {
1509 if (rc == -ENOSPC) {
1510 jfs_error(bmp->db_ipbmap->i_sb,
1511 "control page inconsistent\n");
1512 return -EIO;
1513 }
1514 return (rc);
1515 }
1516 }
1517
1518 /* allocate the blocks.
1519 */
1520 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1521 if (rc == -ENOSPC) {
1522 jfs_error(bmp->db_ipbmap->i_sb,
1523 "unable to allocate blocks\n");
1524 rc = -EIO;
1525 }
1526 return (rc);
1527 }
1528
1529 /* no space in the allocation group. release the buffer and
1530 * return -ENOSPC.
1531 */
1532 release_metapage(mp);
1533
1534 return -ENOSPC;
1535}
1536
1537
1538/*
1539 * NAME: dbAllocAny()
1540 *
1541 * FUNCTION: attempt to allocate the specified number of contiguous
1542 * free blocks anywhere in the file system.
1543 *
1544 * dbAllocAny() attempts to find the sufficient free space by
1545 * searching down the dmap control pages, starting with the
1546 * highest level (i.e. L0, L1, L2) control page. if free space
1547 * large enough to satisfy the desired free space is found, the
1548 * desired free space is allocated.
1549 *
1550 * PARAMETERS:
1551 * bmp - pointer to bmap descriptor
1552 * nblocks - actual number of contiguous free blocks desired.
1553 * l2nb - log2 number of contiguous free blocks desired.
1554 * results - on successful return, set to the starting block number
1555 * of the newly allocated range.
1556 *
1557 * RETURN VALUES:
1558 * 0 - success
1559 * -ENOSPC - insufficient disk resources
1560 * -EIO - i/o error
1561 *
1562 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1563 */
1564static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results)
1565{
1566 int rc;
1567 s64 blkno = 0;
1568
1569 /* starting with the top level dmap control page, search
1570 * down the dmap control levels for sufficient free space.
1571 * if free space is found, dbFindCtl() returns the starting
1572 * block number of the dmap that contains or starts off the
1573 * range of free space.
1574 */
1575 if ((rc = dbFindCtl(bmp, l2nb, bmp->db_maxlevel, &blkno)))
1576 return (rc);
1577
1578 /* allocate the blocks.
1579 */
1580 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1581 if (rc == -ENOSPC) {
1582 jfs_error(bmp->db_ipbmap->i_sb, "unable to allocate blocks\n");
1583 return -EIO;
1584 }
1585 return (rc);
1586}
1587
1588
1589/*
1590 * NAME: dbDiscardAG()
1591 *
1592 * FUNCTION: attempt to discard (TRIM) all free blocks of specific AG
1593 *
1594 * algorithm:
1595 * 1) allocate blocks, as large as possible and save them
1596 * while holding IWRITE_LOCK on ipbmap
1597 * 2) trim all these saved block/length values
1598 * 3) mark the blocks free again
1599 *
1600 * benefit:
1601 * - we work only on one ag at some time, minimizing how long we
1602 * need to lock ipbmap
1603 * - reading / writing the fs is possible most time, even on
1604 * trimming
1605 *
1606 * downside:
1607 * - we write two times to the dmapctl and dmap pages
1608 * - but for me, this seems the best way, better ideas?
1609 * /TR 2012
1610 *
1611 * PARAMETERS:
1612 * ip - pointer to in-core inode
1613 * agno - ag to trim
1614 * minlen - minimum value of contiguous blocks
1615 *
1616 * RETURN VALUES:
1617 * s64 - actual number of blocks trimmed
1618 */
1619s64 dbDiscardAG(struct inode *ip, int agno, s64 minlen)
1620{
1621 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
1622 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
1623 s64 nblocks, blkno;
1624 u64 trimmed = 0;
1625 int rc, l2nb;
1626 struct super_block *sb = ipbmap->i_sb;
1627
1628 struct range2trim {
1629 u64 blkno;
1630 u64 nblocks;
1631 } *totrim, *tt;
1632
1633 /* max blkno / nblocks pairs to trim */
1634 int count = 0, range_cnt;
1635 u64 max_ranges;
1636
1637 /* prevent others from writing new stuff here, while trimming */
1638 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
1639
1640 nblocks = bmp->db_agfree[agno];
1641 max_ranges = nblocks;
1642 do_div(max_ranges, minlen);
1643 range_cnt = min_t(u64, max_ranges + 1, 32 * 1024);
1644 totrim = kmalloc(sizeof(struct range2trim) * range_cnt, GFP_NOFS);
1645 if (totrim == NULL) {
1646 jfs_error(bmp->db_ipbmap->i_sb, "no memory for trim array\n");
1647 IWRITE_UNLOCK(ipbmap);
1648 return 0;
1649 }
1650
1651 tt = totrim;
1652 while (nblocks >= minlen) {
1653 l2nb = BLKSTOL2(nblocks);
1654
1655 /* 0 = okay, -EIO = fatal, -ENOSPC -> try smaller block */
1656 rc = dbAllocAG(bmp, agno, nblocks, l2nb, &blkno);
1657 if (rc == 0) {
1658 tt->blkno = blkno;
1659 tt->nblocks = nblocks;
1660 tt++; count++;
1661
1662 /* the whole ag is free, trim now */
1663 if (bmp->db_agfree[agno] == 0)
1664 break;
1665
1666 /* give a hint for the next while */
1667 nblocks = bmp->db_agfree[agno];
1668 continue;
1669 } else if (rc == -ENOSPC) {
1670 /* search for next smaller log2 block */
1671 l2nb = BLKSTOL2(nblocks) - 1;
1672 nblocks = 1 << l2nb;
1673 } else {
1674 /* Trim any already allocated blocks */
1675 jfs_error(bmp->db_ipbmap->i_sb, "-EIO\n");
1676 break;
1677 }
1678
1679 /* check, if our trim array is full */
1680 if (unlikely(count >= range_cnt - 1))
1681 break;
1682 }
1683 IWRITE_UNLOCK(ipbmap);
1684
1685 tt->nblocks = 0; /* mark the current end */
1686 for (tt = totrim; tt->nblocks != 0; tt++) {
1687 /* when mounted with online discard, dbFree() will
1688 * call jfs_issue_discard() itself */
1689 if (!(JFS_SBI(sb)->flag & JFS_DISCARD))
1690 jfs_issue_discard(ip, tt->blkno, tt->nblocks);
1691 dbFree(ip, tt->blkno, tt->nblocks);
1692 trimmed += tt->nblocks;
1693 }
1694 kfree(totrim);
1695
1696 return trimmed;
1697}
1698
1699/*
1700 * NAME: dbFindCtl()
1701 *
1702 * FUNCTION: starting at a specified dmap control page level and block
1703 * number, search down the dmap control levels for a range of
1704 * contiguous free blocks large enough to satisfy an allocation
1705 * request for the specified number of free blocks.
1706 *
1707 * if sufficient contiguous free blocks are found, this routine
1708 * returns the starting block number within a dmap page that
1709 * contains or starts a range of contiqious free blocks that
1710 * is sufficient in size.
1711 *
1712 * PARAMETERS:
1713 * bmp - pointer to bmap descriptor
1714 * level - starting dmap control page level.
1715 * l2nb - log2 number of contiguous free blocks desired.
1716 * *blkno - on entry, starting block number for conducting the search.
1717 * on successful return, the first block within a dmap page
1718 * that contains or starts a range of contiguous free blocks.
1719 *
1720 * RETURN VALUES:
1721 * 0 - success
1722 * -ENOSPC - insufficient disk resources
1723 * -EIO - i/o error
1724 *
1725 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1726 */
1727static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno)
1728{
1729 int rc, leafidx, lev;
1730 s64 b, lblkno;
1731 struct dmapctl *dcp;
1732 int budmin;
1733 struct metapage *mp;
1734
1735 /* starting at the specified dmap control page level and block
1736 * number, search down the dmap control levels for the starting
1737 * block number of a dmap page that contains or starts off
1738 * sufficient free blocks.
1739 */
1740 for (lev = level, b = *blkno; lev >= 0; lev--) {
1741 /* get the buffer of the dmap control page for the block
1742 * number and level (i.e. L0, L1, L2).
1743 */
1744 lblkno = BLKTOCTL(b, bmp->db_l2nbperpage, lev);
1745 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1746 if (mp == NULL)
1747 return -EIO;
1748 dcp = (struct dmapctl *) mp->data;
1749 budmin = dcp->budmin;
1750
1751 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1752 jfs_error(bmp->db_ipbmap->i_sb,
1753 "Corrupt dmapctl page\n");
1754 release_metapage(mp);
1755 return -EIO;
1756 }
1757
1758 /* search the tree within the dmap control page for
1759 * sufficient free space. if sufficient free space is found,
1760 * dbFindLeaf() returns the index of the leaf at which
1761 * free space was found.
1762 */
1763 rc = dbFindLeaf((dmtree_t *) dcp, l2nb, &leafidx);
1764
1765 /* release the buffer.
1766 */
1767 release_metapage(mp);
1768
1769 /* space found ?
1770 */
1771 if (rc) {
1772 if (lev != level) {
1773 jfs_error(bmp->db_ipbmap->i_sb,
1774 "dmap inconsistent\n");
1775 return -EIO;
1776 }
1777 return -ENOSPC;
1778 }
1779
1780 /* adjust the block number to reflect the location within
1781 * the dmap control page (i.e. the leaf) at which free
1782 * space was found.
1783 */
1784 b += (((s64) leafidx) << budmin);
1785
1786 /* we stop the search at this dmap control page level if
1787 * the number of blocks required is greater than or equal
1788 * to the maximum number of blocks described at the next
1789 * (lower) level.
1790 */
1791 if (l2nb >= budmin)
1792 break;
1793 }
1794
1795 *blkno = b;
1796 return (0);
1797}
1798
1799
1800/*
1801 * NAME: dbAllocCtl()
1802 *
1803 * FUNCTION: attempt to allocate a specified number of contiguous
1804 * blocks starting within a specific dmap.
1805 *
1806 * this routine is called by higher level routines that search
1807 * the dmap control pages above the actual dmaps for contiguous
1808 * free space. the result of successful searches by these
1809 * routines are the starting block numbers within dmaps, with
1810 * the dmaps themselves containing the desired contiguous free
1811 * space or starting a contiguous free space of desired size
1812 * that is made up of the blocks of one or more dmaps. these
1813 * calls should not fail due to insufficent resources.
1814 *
1815 * this routine is called in some cases where it is not known
1816 * whether it will fail due to insufficient resources. more
1817 * specifically, this occurs when allocating from an allocation
1818 * group whose size is equal to the number of blocks per dmap.
1819 * in this case, the dmap control pages are not examined prior
1820 * to calling this routine (to save pathlength) and the call
1821 * might fail.
1822 *
1823 * for a request size that fits within a dmap, this routine relies
1824 * upon the dmap's dmtree to find the requested contiguous free
1825 * space. for request sizes that are larger than a dmap, the
1826 * requested free space will start at the first block of the
1827 * first dmap (i.e. blkno).
1828 *
1829 * PARAMETERS:
1830 * bmp - pointer to bmap descriptor
1831 * nblocks - actual number of contiguous free blocks to allocate.
1832 * l2nb - log2 number of contiguous free blocks to allocate.
1833 * blkno - starting block number of the dmap to start the allocation
1834 * from.
1835 * results - on successful return, set to the starting block number
1836 * of the newly allocated range.
1837 *
1838 * RETURN VALUES:
1839 * 0 - success
1840 * -ENOSPC - insufficient disk resources
1841 * -EIO - i/o error
1842 *
1843 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1844 */
1845static int
1846dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, s64 * results)
1847{
1848 int rc, nb;
1849 s64 b, lblkno, n;
1850 struct metapage *mp;
1851 struct dmap *dp;
1852
1853 /* check if the allocation request is confined to a single dmap.
1854 */
1855 if (l2nb <= L2BPERDMAP) {
1856 /* get the buffer for the dmap.
1857 */
1858 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
1859 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1860 if (mp == NULL)
1861 return -EIO;
1862 dp = (struct dmap *) mp->data;
1863
1864 /* try to allocate the blocks.
1865 */
1866 rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results);
1867 if (rc == 0)
1868 mark_metapage_dirty(mp);
1869
1870 release_metapage(mp);
1871
1872 return (rc);
1873 }
1874
1875 /* allocation request involving multiple dmaps. it must start on
1876 * a dmap boundary.
1877 */
1878 assert((blkno & (BPERDMAP - 1)) == 0);
1879
1880 /* allocate the blocks dmap by dmap.
1881 */
1882 for (n = nblocks, b = blkno; n > 0; n -= nb, b += nb) {
1883 /* get the buffer for the dmap.
1884 */
1885 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1886 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1887 if (mp == NULL) {
1888 rc = -EIO;
1889 goto backout;
1890 }
1891 dp = (struct dmap *) mp->data;
1892
1893 /* the dmap better be all free.
1894 */
1895 if (dp->tree.stree[ROOT] != L2BPERDMAP) {
1896 release_metapage(mp);
1897 jfs_error(bmp->db_ipbmap->i_sb,
1898 "the dmap is not all free\n");
1899 rc = -EIO;
1900 goto backout;
1901 }
1902
1903 /* determine how many blocks to allocate from this dmap.
1904 */
1905 nb = min(n, (s64)BPERDMAP);
1906
1907 /* allocate the blocks from the dmap.
1908 */
1909 if ((rc = dbAllocDmap(bmp, dp, b, nb))) {
1910 release_metapage(mp);
1911 goto backout;
1912 }
1913
1914 /* write the buffer.
1915 */
1916 write_metapage(mp);
1917 }
1918
1919 /* set the results (starting block number) and return.
1920 */
1921 *results = blkno;
1922 return (0);
1923
1924 /* something failed in handling an allocation request involving
1925 * multiple dmaps. we'll try to clean up by backing out any
1926 * allocation that has already happened for this request. if
1927 * we fail in backing out the allocation, we'll mark the file
1928 * system to indicate that blocks have been leaked.
1929 */
1930 backout:
1931
1932 /* try to backout the allocations dmap by dmap.
1933 */
1934 for (n = nblocks - n, b = blkno; n > 0;
1935 n -= BPERDMAP, b += BPERDMAP) {
1936 /* get the buffer for this dmap.
1937 */
1938 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1939 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1940 if (mp == NULL) {
1941 /* could not back out. mark the file system
1942 * to indicate that we have leaked blocks.
1943 */
1944 jfs_error(bmp->db_ipbmap->i_sb,
1945 "I/O Error: Block Leakage\n");
1946 continue;
1947 }
1948 dp = (struct dmap *) mp->data;
1949
1950 /* free the blocks is this dmap.
1951 */
1952 if (dbFreeDmap(bmp, dp, b, BPERDMAP)) {
1953 /* could not back out. mark the file system
1954 * to indicate that we have leaked blocks.
1955 */
1956 release_metapage(mp);
1957 jfs_error(bmp->db_ipbmap->i_sb, "Block Leakage\n");
1958 continue;
1959 }
1960
1961 /* write the buffer.
1962 */
1963 write_metapage(mp);
1964 }
1965
1966 return (rc);
1967}
1968
1969
1970/*
1971 * NAME: dbAllocDmapLev()
1972 *
1973 * FUNCTION: attempt to allocate a specified number of contiguous blocks
1974 * from a specified dmap.
1975 *
1976 * this routine checks if the contiguous blocks are available.
1977 * if so, nblocks of blocks are allocated; otherwise, ENOSPC is
1978 * returned.
1979 *
1980 * PARAMETERS:
1981 * mp - pointer to bmap descriptor
1982 * dp - pointer to dmap to attempt to allocate blocks from.
1983 * l2nb - log2 number of contiguous block desired.
1984 * nblocks - actual number of contiguous block desired.
1985 * results - on successful return, set to the starting block number
1986 * of the newly allocated range.
1987 *
1988 * RETURN VALUES:
1989 * 0 - success
1990 * -ENOSPC - insufficient disk resources
1991 * -EIO - i/o error
1992 *
1993 * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or
1994 * IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit;
1995 */
1996static int
1997dbAllocDmapLev(struct bmap * bmp,
1998 struct dmap * dp, int nblocks, int l2nb, s64 * results)
1999{
2000 s64 blkno;
2001 int leafidx, rc;
2002
2003 /* can't be more than a dmaps worth of blocks */
2004 assert(l2nb <= L2BPERDMAP);
2005
2006 /* search the tree within the dmap page for sufficient
2007 * free space. if sufficient free space is found, dbFindLeaf()
2008 * returns the index of the leaf at which free space was found.
2009 */
2010 if (dbFindLeaf((dmtree_t *) & dp->tree, l2nb, &leafidx))
2011 return -ENOSPC;
2012
2013 /* determine the block number within the file system corresponding
2014 * to the leaf at which free space was found.
2015 */
2016 blkno = le64_to_cpu(dp->start) + (leafidx << L2DBWORD);
2017
2018 /* if not all bits of the dmap word are free, get the starting
2019 * bit number within the dmap word of the required string of free
2020 * bits and adjust the block number with this value.
2021 */
2022 if (dp->tree.stree[leafidx + LEAFIND] < BUDMIN)
2023 blkno += dbFindBits(le32_to_cpu(dp->wmap[leafidx]), l2nb);
2024
2025 /* allocate the blocks */
2026 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
2027 *results = blkno;
2028
2029 return (rc);
2030}
2031
2032
2033/*
2034 * NAME: dbAllocDmap()
2035 *
2036 * FUNCTION: adjust the disk allocation map to reflect the allocation
2037 * of a specified block range within a dmap.
2038 *
2039 * this routine allocates the specified blocks from the dmap
2040 * through a call to dbAllocBits(). if the allocation of the
2041 * block range causes the maximum string of free blocks within
2042 * the dmap to change (i.e. the value of the root of the dmap's
2043 * dmtree), this routine will cause this change to be reflected
2044 * up through the appropriate levels of the dmap control pages
2045 * by a call to dbAdjCtl() for the L0 dmap control page that
2046 * covers this dmap.
2047 *
2048 * PARAMETERS:
2049 * bmp - pointer to bmap descriptor
2050 * dp - pointer to dmap to allocate the block range from.
2051 * blkno - starting block number of the block to be allocated.
2052 * nblocks - number of blocks to be allocated.
2053 *
2054 * RETURN VALUES:
2055 * 0 - success
2056 * -EIO - i/o error
2057 *
2058 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2059 */
2060static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2061 int nblocks)
2062{
2063 s8 oldroot;
2064 int rc;
2065
2066 /* save the current value of the root (i.e. maximum free string)
2067 * of the dmap tree.
2068 */
2069 oldroot = dp->tree.stree[ROOT];
2070
2071 /* allocate the specified (blocks) bits */
2072 dbAllocBits(bmp, dp, blkno, nblocks);
2073
2074 /* if the root has not changed, done. */
2075 if (dp->tree.stree[ROOT] == oldroot)
2076 return (0);
2077
2078 /* root changed. bubble the change up to the dmap control pages.
2079 * if the adjustment of the upper level control pages fails,
2080 * backout the bit allocation (thus making everything consistent).
2081 */
2082 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 1, 0)))
2083 dbFreeBits(bmp, dp, blkno, nblocks);
2084
2085 return (rc);
2086}
2087
2088
2089/*
2090 * NAME: dbFreeDmap()
2091 *
2092 * FUNCTION: adjust the disk allocation map to reflect the allocation
2093 * of a specified block range within a dmap.
2094 *
2095 * this routine frees the specified blocks from the dmap through
2096 * a call to dbFreeBits(). if the deallocation of the block range
2097 * causes the maximum string of free blocks within the dmap to
2098 * change (i.e. the value of the root of the dmap's dmtree), this
2099 * routine will cause this change to be reflected up through the
2100 * appropriate levels of the dmap control pages by a call to
2101 * dbAdjCtl() for the L0 dmap control page that covers this dmap.
2102 *
2103 * PARAMETERS:
2104 * bmp - pointer to bmap descriptor
2105 * dp - pointer to dmap to free the block range from.
2106 * blkno - starting block number of the block to be freed.
2107 * nblocks - number of blocks to be freed.
2108 *
2109 * RETURN VALUES:
2110 * 0 - success
2111 * -EIO - i/o error
2112 *
2113 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2114 */
2115static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2116 int nblocks)
2117{
2118 s8 oldroot;
2119 int rc = 0, word;
2120
2121 /* save the current value of the root (i.e. maximum free string)
2122 * of the dmap tree.
2123 */
2124 oldroot = dp->tree.stree[ROOT];
2125
2126 /* free the specified (blocks) bits */
2127 rc = dbFreeBits(bmp, dp, blkno, nblocks);
2128
2129 /* if error or the root has not changed, done. */
2130 if (rc || (dp->tree.stree[ROOT] == oldroot))
2131 return (rc);
2132
2133 /* root changed. bubble the change up to the dmap control pages.
2134 * if the adjustment of the upper level control pages fails,
2135 * backout the deallocation.
2136 */
2137 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 0, 0))) {
2138 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
2139
2140 /* as part of backing out the deallocation, we will have
2141 * to back split the dmap tree if the deallocation caused
2142 * the freed blocks to become part of a larger binary buddy
2143 * system.
2144 */
2145 if (dp->tree.stree[word] == NOFREE)
2146 dbBackSplit((dmtree_t *) & dp->tree, word);
2147
2148 dbAllocBits(bmp, dp, blkno, nblocks);
2149 }
2150
2151 return (rc);
2152}
2153
2154
2155/*
2156 * NAME: dbAllocBits()
2157 *
2158 * FUNCTION: allocate a specified block range from a dmap.
2159 *
2160 * this routine updates the dmap to reflect the working
2161 * state allocation of the specified block range. it directly
2162 * updates the bits of the working map and causes the adjustment
2163 * of the binary buddy system described by the dmap's dmtree
2164 * leaves to reflect the bits allocated. it also causes the
2165 * dmap's dmtree, as a whole, to reflect the allocated range.
2166 *
2167 * PARAMETERS:
2168 * bmp - pointer to bmap descriptor
2169 * dp - pointer to dmap to allocate bits from.
2170 * blkno - starting block number of the bits to be allocated.
2171 * nblocks - number of bits to be allocated.
2172 *
2173 * RETURN VALUES: none
2174 *
2175 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2176 */
2177static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2178 int nblocks)
2179{
2180 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2181 dmtree_t *tp = (dmtree_t *) & dp->tree;
2182 int size;
2183 s8 *leaf;
2184
2185 /* pick up a pointer to the leaves of the dmap tree */
2186 leaf = dp->tree.stree + LEAFIND;
2187
2188 /* determine the bit number and word within the dmap of the
2189 * starting block.
2190 */
2191 dbitno = blkno & (BPERDMAP - 1);
2192 word = dbitno >> L2DBWORD;
2193
2194 /* block range better be within the dmap */
2195 assert(dbitno + nblocks <= BPERDMAP);
2196
2197 /* allocate the bits of the dmap's words corresponding to the block
2198 * range. not all bits of the first and last words may be contained
2199 * within the block range. if this is the case, we'll work against
2200 * those words (i.e. partial first and/or last) on an individual basis
2201 * (a single pass), allocating the bits of interest by hand and
2202 * updating the leaf corresponding to the dmap word. a single pass
2203 * will be used for all dmap words fully contained within the
2204 * specified range. within this pass, the bits of all fully contained
2205 * dmap words will be marked as free in a single shot and the leaves
2206 * will be updated. a single leaf may describe the free space of
2207 * multiple dmap words, so we may update only a subset of the actual
2208 * leaves corresponding to the dmap words of the block range.
2209 */
2210 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2211 /* determine the bit number within the word and
2212 * the number of bits within the word.
2213 */
2214 wbitno = dbitno & (DBWORD - 1);
2215 nb = min(rembits, DBWORD - wbitno);
2216
2217 /* check if only part of a word is to be allocated.
2218 */
2219 if (nb < DBWORD) {
2220 /* allocate (set to 1) the appropriate bits within
2221 * this dmap word.
2222 */
2223 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
2224 >> wbitno);
2225
2226 /* update the leaf for this dmap word. in addition
2227 * to setting the leaf value to the binary buddy max
2228 * of the updated dmap word, dbSplit() will split
2229 * the binary system of the leaves if need be.
2230 */
2231 dbSplit(tp, word, BUDMIN,
2232 dbMaxBud((u8 *) & dp->wmap[word]));
2233
2234 word += 1;
2235 } else {
2236 /* one or more dmap words are fully contained
2237 * within the block range. determine how many
2238 * words and allocate (set to 1) the bits of these
2239 * words.
2240 */
2241 nwords = rembits >> L2DBWORD;
2242 memset(&dp->wmap[word], (int) ONES, nwords * 4);
2243
2244 /* determine how many bits.
2245 */
2246 nb = nwords << L2DBWORD;
2247
2248 /* now update the appropriate leaves to reflect
2249 * the allocated words.
2250 */
2251 for (; nwords > 0; nwords -= nw) {
2252 if (leaf[word] < BUDMIN) {
2253 jfs_error(bmp->db_ipbmap->i_sb,
2254 "leaf page corrupt\n");
2255 break;
2256 }
2257
2258 /* determine what the leaf value should be
2259 * updated to as the minimum of the l2 number
2260 * of bits being allocated and the l2 number
2261 * of bits currently described by this leaf.
2262 */
2263 size = min((int)leaf[word], NLSTOL2BSZ(nwords));
2264
2265 /* update the leaf to reflect the allocation.
2266 * in addition to setting the leaf value to
2267 * NOFREE, dbSplit() will split the binary
2268 * system of the leaves to reflect the current
2269 * allocation (size).
2270 */
2271 dbSplit(tp, word, size, NOFREE);
2272
2273 /* get the number of dmap words handled */
2274 nw = BUDSIZE(size, BUDMIN);
2275 word += nw;
2276 }
2277 }
2278 }
2279
2280 /* update the free count for this dmap */
2281 le32_add_cpu(&dp->nfree, -nblocks);
2282
2283 BMAP_LOCK(bmp);
2284
2285 /* if this allocation group is completely free,
2286 * update the maximum allocation group number if this allocation
2287 * group is the new max.
2288 */
2289 agno = blkno >> bmp->db_agl2size;
2290 if (agno > bmp->db_maxag)
2291 bmp->db_maxag = agno;
2292
2293 /* update the free count for the allocation group and map */
2294 bmp->db_agfree[agno] -= nblocks;
2295 bmp->db_nfree -= nblocks;
2296
2297 BMAP_UNLOCK(bmp);
2298}
2299
2300
2301/*
2302 * NAME: dbFreeBits()
2303 *
2304 * FUNCTION: free a specified block range from a dmap.
2305 *
2306 * this routine updates the dmap to reflect the working
2307 * state allocation of the specified block range. it directly
2308 * updates the bits of the working map and causes the adjustment
2309 * of the binary buddy system described by the dmap's dmtree
2310 * leaves to reflect the bits freed. it also causes the dmap's
2311 * dmtree, as a whole, to reflect the deallocated range.
2312 *
2313 * PARAMETERS:
2314 * bmp - pointer to bmap descriptor
2315 * dp - pointer to dmap to free bits from.
2316 * blkno - starting block number of the bits to be freed.
2317 * nblocks - number of bits to be freed.
2318 *
2319 * RETURN VALUES: 0 for success
2320 *
2321 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2322 */
2323static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2324 int nblocks)
2325{
2326 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2327 dmtree_t *tp = (dmtree_t *) & dp->tree;
2328 int rc = 0;
2329 int size;
2330
2331 /* determine the bit number and word within the dmap of the
2332 * starting block.
2333 */
2334 dbitno = blkno & (BPERDMAP - 1);
2335 word = dbitno >> L2DBWORD;
2336
2337 /* block range better be within the dmap.
2338 */
2339 assert(dbitno + nblocks <= BPERDMAP);
2340
2341 /* free the bits of the dmaps words corresponding to the block range.
2342 * not all bits of the first and last words may be contained within
2343 * the block range. if this is the case, we'll work against those
2344 * words (i.e. partial first and/or last) on an individual basis
2345 * (a single pass), freeing the bits of interest by hand and updating
2346 * the leaf corresponding to the dmap word. a single pass will be used
2347 * for all dmap words fully contained within the specified range.
2348 * within this pass, the bits of all fully contained dmap words will
2349 * be marked as free in a single shot and the leaves will be updated. a
2350 * single leaf may describe the free space of multiple dmap words,
2351 * so we may update only a subset of the actual leaves corresponding
2352 * to the dmap words of the block range.
2353 *
2354 * dbJoin() is used to update leaf values and will join the binary
2355 * buddy system of the leaves if the new leaf values indicate this
2356 * should be done.
2357 */
2358 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2359 /* determine the bit number within the word and
2360 * the number of bits within the word.
2361 */
2362 wbitno = dbitno & (DBWORD - 1);
2363 nb = min(rembits, DBWORD - wbitno);
2364
2365 /* check if only part of a word is to be freed.
2366 */
2367 if (nb < DBWORD) {
2368 /* free (zero) the appropriate bits within this
2369 * dmap word.
2370 */
2371 dp->wmap[word] &=
2372 cpu_to_le32(~(ONES << (DBWORD - nb)
2373 >> wbitno));
2374
2375 /* update the leaf for this dmap word.
2376 */
2377 rc = dbJoin(tp, word,
2378 dbMaxBud((u8 *) & dp->wmap[word]));
2379 if (rc)
2380 return rc;
2381
2382 word += 1;
2383 } else {
2384 /* one or more dmap words are fully contained
2385 * within the block range. determine how many
2386 * words and free (zero) the bits of these words.
2387 */
2388 nwords = rembits >> L2DBWORD;
2389 memset(&dp->wmap[word], 0, nwords * 4);
2390
2391 /* determine how many bits.
2392 */
2393 nb = nwords << L2DBWORD;
2394
2395 /* now update the appropriate leaves to reflect
2396 * the freed words.
2397 */
2398 for (; nwords > 0; nwords -= nw) {
2399 /* determine what the leaf value should be
2400 * updated to as the minimum of the l2 number
2401 * of bits being freed and the l2 (max) number
2402 * of bits that can be described by this leaf.
2403 */
2404 size =
2405 min(LITOL2BSZ
2406 (word, L2LPERDMAP, BUDMIN),
2407 NLSTOL2BSZ(nwords));
2408
2409 /* update the leaf.
2410 */
2411 rc = dbJoin(tp, word, size);
2412 if (rc)
2413 return rc;
2414
2415 /* get the number of dmap words handled.
2416 */
2417 nw = BUDSIZE(size, BUDMIN);
2418 word += nw;
2419 }
2420 }
2421 }
2422
2423 /* update the free count for this dmap.
2424 */
2425 le32_add_cpu(&dp->nfree, nblocks);
2426
2427 BMAP_LOCK(bmp);
2428
2429 /* update the free count for the allocation group and
2430 * map.
2431 */
2432 agno = blkno >> bmp->db_agl2size;
2433 bmp->db_nfree += nblocks;
2434 bmp->db_agfree[agno] += nblocks;
2435
2436 /* check if this allocation group is not completely free and
2437 * if it is currently the maximum (rightmost) allocation group.
2438 * if so, establish the new maximum allocation group number by
2439 * searching left for the first allocation group with allocation.
2440 */
2441 if ((bmp->db_agfree[agno] == bmp->db_agsize && agno == bmp->db_maxag) ||
2442 (agno == bmp->db_numag - 1 &&
2443 bmp->db_agfree[agno] == (bmp-> db_mapsize & (BPERDMAP - 1)))) {
2444 while (bmp->db_maxag > 0) {
2445 bmp->db_maxag -= 1;
2446 if (bmp->db_agfree[bmp->db_maxag] !=
2447 bmp->db_agsize)
2448 break;
2449 }
2450
2451 /* re-establish the allocation group preference if the
2452 * current preference is right of the maximum allocation
2453 * group.
2454 */
2455 if (bmp->db_agpref > bmp->db_maxag)
2456 bmp->db_agpref = bmp->db_maxag;
2457 }
2458
2459 BMAP_UNLOCK(bmp);
2460
2461 return 0;
2462}
2463
2464
2465/*
2466 * NAME: dbAdjCtl()
2467 *
2468 * FUNCTION: adjust a dmap control page at a specified level to reflect
2469 * the change in a lower level dmap or dmap control page's
2470 * maximum string of free blocks (i.e. a change in the root
2471 * of the lower level object's dmtree) due to the allocation
2472 * or deallocation of a range of blocks with a single dmap.
2473 *
2474 * on entry, this routine is provided with the new value of
2475 * the lower level dmap or dmap control page root and the
2476 * starting block number of the block range whose allocation
2477 * or deallocation resulted in the root change. this range
2478 * is respresented by a single leaf of the current dmapctl
2479 * and the leaf will be updated with this value, possibly
2480 * causing a binary buddy system within the leaves to be
2481 * split or joined. the update may also cause the dmapctl's
2482 * dmtree to be updated.
2483 *
2484 * if the adjustment of the dmap control page, itself, causes its
2485 * root to change, this change will be bubbled up to the next dmap
2486 * control level by a recursive call to this routine, specifying
2487 * the new root value and the next dmap control page level to
2488 * be adjusted.
2489 * PARAMETERS:
2490 * bmp - pointer to bmap descriptor
2491 * blkno - the first block of a block range within a dmap. it is
2492 * the allocation or deallocation of this block range that
2493 * requires the dmap control page to be adjusted.
2494 * newval - the new value of the lower level dmap or dmap control
2495 * page root.
2496 * alloc - 'true' if adjustment is due to an allocation.
2497 * level - current level of dmap control page (i.e. L0, L1, L2) to
2498 * be adjusted.
2499 *
2500 * RETURN VALUES:
2501 * 0 - success
2502 * -EIO - i/o error
2503 *
2504 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2505 */
2506static int
2507dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, int level)
2508{
2509 struct metapage *mp;
2510 s8 oldroot;
2511 int oldval;
2512 s64 lblkno;
2513 struct dmapctl *dcp;
2514 int rc, leafno, ti;
2515
2516 /* get the buffer for the dmap control page for the specified
2517 * block number and control page level.
2518 */
2519 lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, level);
2520 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
2521 if (mp == NULL)
2522 return -EIO;
2523 dcp = (struct dmapctl *) mp->data;
2524
2525 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
2526 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
2527 release_metapage(mp);
2528 return -EIO;
2529 }
2530
2531 /* determine the leaf number corresponding to the block and
2532 * the index within the dmap control tree.
2533 */
2534 leafno = BLKTOCTLLEAF(blkno, dcp->budmin);
2535 ti = leafno + le32_to_cpu(dcp->leafidx);
2536
2537 /* save the current leaf value and the current root level (i.e.
2538 * maximum l2 free string described by this dmapctl).
2539 */
2540 oldval = dcp->stree[ti];
2541 oldroot = dcp->stree[ROOT];
2542
2543 /* check if this is a control page update for an allocation.
2544 * if so, update the leaf to reflect the new leaf value using
2545 * dbSplit(); otherwise (deallocation), use dbJoin() to update
2546 * the leaf with the new value. in addition to updating the
2547 * leaf, dbSplit() will also split the binary buddy system of
2548 * the leaves, if required, and bubble new values within the
2549 * dmapctl tree, if required. similarly, dbJoin() will join
2550 * the binary buddy system of leaves and bubble new values up
2551 * the dmapctl tree as required by the new leaf value.
2552 */
2553 if (alloc) {
2554 /* check if we are in the middle of a binary buddy
2555 * system. this happens when we are performing the
2556 * first allocation out of an allocation group that
2557 * is part (not the first part) of a larger binary
2558 * buddy system. if we are in the middle, back split
2559 * the system prior to calling dbSplit() which assumes
2560 * that it is at the front of a binary buddy system.
2561 */
2562 if (oldval == NOFREE) {
2563 rc = dbBackSplit((dmtree_t *) dcp, leafno);
2564 if (rc)
2565 return rc;
2566 oldval = dcp->stree[ti];
2567 }
2568 dbSplit((dmtree_t *) dcp, leafno, dcp->budmin, newval);
2569 } else {
2570 rc = dbJoin((dmtree_t *) dcp, leafno, newval);
2571 if (rc)
2572 return rc;
2573 }
2574
2575 /* check if the root of the current dmap control page changed due
2576 * to the update and if the current dmap control page is not at
2577 * the current top level (i.e. L0, L1, L2) of the map. if so (i.e.
2578 * root changed and this is not the top level), call this routine
2579 * again (recursion) for the next higher level of the mapping to
2580 * reflect the change in root for the current dmap control page.
2581 */
2582 if (dcp->stree[ROOT] != oldroot) {
2583 /* are we below the top level of the map. if so,
2584 * bubble the root up to the next higher level.
2585 */
2586 if (level < bmp->db_maxlevel) {
2587 /* bubble up the new root of this dmap control page to
2588 * the next level.
2589 */
2590 if ((rc =
2591 dbAdjCtl(bmp, blkno, dcp->stree[ROOT], alloc,
2592 level + 1))) {
2593 /* something went wrong in bubbling up the new
2594 * root value, so backout the changes to the
2595 * current dmap control page.
2596 */
2597 if (alloc) {
2598 dbJoin((dmtree_t *) dcp, leafno,
2599 oldval);
2600 } else {
2601 /* the dbJoin() above might have
2602 * caused a larger binary buddy system
2603 * to form and we may now be in the
2604 * middle of it. if this is the case,
2605 * back split the buddies.
2606 */
2607 if (dcp->stree[ti] == NOFREE)
2608 dbBackSplit((dmtree_t *)
2609 dcp, leafno);
2610 dbSplit((dmtree_t *) dcp, leafno,
2611 dcp->budmin, oldval);
2612 }
2613
2614 /* release the buffer and return the error.
2615 */
2616 release_metapage(mp);
2617 return (rc);
2618 }
2619 } else {
2620 /* we're at the top level of the map. update
2621 * the bmap control page to reflect the size
2622 * of the maximum free buddy system.
2623 */
2624 assert(level == bmp->db_maxlevel);
2625 if (bmp->db_maxfreebud != oldroot) {
2626 jfs_error(bmp->db_ipbmap->i_sb,
2627 "the maximum free buddy is not the old root\n");
2628 }
2629 bmp->db_maxfreebud = dcp->stree[ROOT];
2630 }
2631 }
2632
2633 /* write the buffer.
2634 */
2635 write_metapage(mp);
2636
2637 return (0);
2638}
2639
2640
2641/*
2642 * NAME: dbSplit()
2643 *
2644 * FUNCTION: update the leaf of a dmtree with a new value, splitting
2645 * the leaf from the binary buddy system of the dmtree's
2646 * leaves, as required.
2647 *
2648 * PARAMETERS:
2649 * tp - pointer to the tree containing the leaf.
2650 * leafno - the number of the leaf to be updated.
2651 * splitsz - the size the binary buddy system starting at the leaf
2652 * must be split to, specified as the log2 number of blocks.
2653 * newval - the new value for the leaf.
2654 *
2655 * RETURN VALUES: none
2656 *
2657 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2658 */
2659static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval)
2660{
2661 int budsz;
2662 int cursz;
2663 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2664
2665 /* check if the leaf needs to be split.
2666 */
2667 if (leaf[leafno] > tp->dmt_budmin) {
2668 /* the split occurs by cutting the buddy system in half
2669 * at the specified leaf until we reach the specified
2670 * size. pick up the starting split size (current size
2671 * - 1 in l2) and the corresponding buddy size.
2672 */
2673 cursz = leaf[leafno] - 1;
2674 budsz = BUDSIZE(cursz, tp->dmt_budmin);
2675
2676 /* split until we reach the specified size.
2677 */
2678 while (cursz >= splitsz) {
2679 /* update the buddy's leaf with its new value.
2680 */
2681 dbAdjTree(tp, leafno ^ budsz, cursz);
2682
2683 /* on to the next size and buddy.
2684 */
2685 cursz -= 1;
2686 budsz >>= 1;
2687 }
2688 }
2689
2690 /* adjust the dmap tree to reflect the specified leaf's new
2691 * value.
2692 */
2693 dbAdjTree(tp, leafno, newval);
2694}
2695
2696
2697/*
2698 * NAME: dbBackSplit()
2699 *
2700 * FUNCTION: back split the binary buddy system of dmtree leaves
2701 * that hold a specified leaf until the specified leaf
2702 * starts its own binary buddy system.
2703 *
2704 * the allocators typically perform allocations at the start
2705 * of binary buddy systems and dbSplit() is used to accomplish
2706 * any required splits. in some cases, however, allocation
2707 * may occur in the middle of a binary system and requires a
2708 * back split, with the split proceeding out from the middle of
2709 * the system (less efficient) rather than the start of the
2710 * system (more efficient). the cases in which a back split
2711 * is required are rare and are limited to the first allocation
2712 * within an allocation group which is a part (not first part)
2713 * of a larger binary buddy system and a few exception cases
2714 * in which a previous join operation must be backed out.
2715 *
2716 * PARAMETERS:
2717 * tp - pointer to the tree containing the leaf.
2718 * leafno - the number of the leaf to be updated.
2719 *
2720 * RETURN VALUES: none
2721 *
2722 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2723 */
2724static int dbBackSplit(dmtree_t * tp, int leafno)
2725{
2726 int budsz, bud, w, bsz, size;
2727 int cursz;
2728 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2729
2730 /* leaf should be part (not first part) of a binary
2731 * buddy system.
2732 */
2733 assert(leaf[leafno] == NOFREE);
2734
2735 /* the back split is accomplished by iteratively finding the leaf
2736 * that starts the buddy system that contains the specified leaf and
2737 * splitting that system in two. this iteration continues until
2738 * the specified leaf becomes the start of a buddy system.
2739 *
2740 * determine maximum possible l2 size for the specified leaf.
2741 */
2742 size =
2743 LITOL2BSZ(leafno, le32_to_cpu(tp->dmt_l2nleafs),
2744 tp->dmt_budmin);
2745
2746 /* determine the number of leaves covered by this size. this
2747 * is the buddy size that we will start with as we search for
2748 * the buddy system that contains the specified leaf.
2749 */
2750 budsz = BUDSIZE(size, tp->dmt_budmin);
2751
2752 /* back split.
2753 */
2754 while (leaf[leafno] == NOFREE) {
2755 /* find the leftmost buddy leaf.
2756 */
2757 for (w = leafno, bsz = budsz;; bsz <<= 1,
2758 w = (w < bud) ? w : bud) {
2759 if (bsz >= le32_to_cpu(tp->dmt_nleafs)) {
2760 jfs_err("JFS: block map error in dbBackSplit");
2761 return -EIO;
2762 }
2763
2764 /* determine the buddy.
2765 */
2766 bud = w ^ bsz;
2767
2768 /* check if this buddy is the start of the system.
2769 */
2770 if (leaf[bud] != NOFREE) {
2771 /* split the leaf at the start of the
2772 * system in two.
2773 */
2774 cursz = leaf[bud] - 1;
2775 dbSplit(tp, bud, cursz, cursz);
2776 break;
2777 }
2778 }
2779 }
2780
2781 if (leaf[leafno] != size) {
2782 jfs_err("JFS: wrong leaf value in dbBackSplit");
2783 return -EIO;
2784 }
2785 return 0;
2786}
2787
2788
2789/*
2790 * NAME: dbJoin()
2791 *
2792 * FUNCTION: update the leaf of a dmtree with a new value, joining
2793 * the leaf with other leaves of the dmtree into a multi-leaf
2794 * binary buddy system, as required.
2795 *
2796 * PARAMETERS:
2797 * tp - pointer to the tree containing the leaf.
2798 * leafno - the number of the leaf to be updated.
2799 * newval - the new value for the leaf.
2800 *
2801 * RETURN VALUES: none
2802 */
2803static int dbJoin(dmtree_t * tp, int leafno, int newval)
2804{
2805 int budsz, buddy;
2806 s8 *leaf;
2807
2808 /* can the new leaf value require a join with other leaves ?
2809 */
2810 if (newval >= tp->dmt_budmin) {
2811 /* pickup a pointer to the leaves of the tree.
2812 */
2813 leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2814
2815 /* try to join the specified leaf into a large binary
2816 * buddy system. the join proceeds by attempting to join
2817 * the specified leafno with its buddy (leaf) at new value.
2818 * if the join occurs, we attempt to join the left leaf
2819 * of the joined buddies with its buddy at new value + 1.
2820 * we continue to join until we find a buddy that cannot be
2821 * joined (does not have a value equal to the size of the
2822 * last join) or until all leaves have been joined into a
2823 * single system.
2824 *
2825 * get the buddy size (number of words covered) of
2826 * the new value.
2827 */
2828 budsz = BUDSIZE(newval, tp->dmt_budmin);
2829
2830 /* try to join.
2831 */
2832 while (budsz < le32_to_cpu(tp->dmt_nleafs)) {
2833 /* get the buddy leaf.
2834 */
2835 buddy = leafno ^ budsz;
2836
2837 /* if the leaf's new value is greater than its
2838 * buddy's value, we join no more.
2839 */
2840 if (newval > leaf[buddy])
2841 break;
2842
2843 /* It shouldn't be less */
2844 if (newval < leaf[buddy])
2845 return -EIO;
2846
2847 /* check which (leafno or buddy) is the left buddy.
2848 * the left buddy gets to claim the blocks resulting
2849 * from the join while the right gets to claim none.
2850 * the left buddy is also eligible to participate in
2851 * a join at the next higher level while the right
2852 * is not.
2853 *
2854 */
2855 if (leafno < buddy) {
2856 /* leafno is the left buddy.
2857 */
2858 dbAdjTree(tp, buddy, NOFREE);
2859 } else {
2860 /* buddy is the left buddy and becomes
2861 * leafno.
2862 */
2863 dbAdjTree(tp, leafno, NOFREE);
2864 leafno = buddy;
2865 }
2866
2867 /* on to try the next join.
2868 */
2869 newval += 1;
2870 budsz <<= 1;
2871 }
2872 }
2873
2874 /* update the leaf value.
2875 */
2876 dbAdjTree(tp, leafno, newval);
2877
2878 return 0;
2879}
2880
2881
2882/*
2883 * NAME: dbAdjTree()
2884 *
2885 * FUNCTION: update a leaf of a dmtree with a new value, adjusting
2886 * the dmtree, as required, to reflect the new leaf value.
2887 * the combination of any buddies must already be done before
2888 * this is called.
2889 *
2890 * PARAMETERS:
2891 * tp - pointer to the tree to be adjusted.
2892 * leafno - the number of the leaf to be updated.
2893 * newval - the new value for the leaf.
2894 *
2895 * RETURN VALUES: none
2896 */
2897static void dbAdjTree(dmtree_t * tp, int leafno, int newval)
2898{
2899 int lp, pp, k;
2900 int max;
2901
2902 /* pick up the index of the leaf for this leafno.
2903 */
2904 lp = leafno + le32_to_cpu(tp->dmt_leafidx);
2905
2906 /* is the current value the same as the old value ? if so,
2907 * there is nothing to do.
2908 */
2909 if (tp->dmt_stree[lp] == newval)
2910 return;
2911
2912 /* set the new value.
2913 */
2914 tp->dmt_stree[lp] = newval;
2915
2916 /* bubble the new value up the tree as required.
2917 */
2918 for (k = 0; k < le32_to_cpu(tp->dmt_height); k++) {
2919 /* get the index of the first leaf of the 4 leaf
2920 * group containing the specified leaf (leafno).
2921 */
2922 lp = ((lp - 1) & ~0x03) + 1;
2923
2924 /* get the index of the parent of this 4 leaf group.
2925 */
2926 pp = (lp - 1) >> 2;
2927
2928 /* determine the maximum of the 4 leaves.
2929 */
2930 max = TREEMAX(&tp->dmt_stree[lp]);
2931
2932 /* if the maximum of the 4 is the same as the
2933 * parent's value, we're done.
2934 */
2935 if (tp->dmt_stree[pp] == max)
2936 break;
2937
2938 /* parent gets new value.
2939 */
2940 tp->dmt_stree[pp] = max;
2941
2942 /* parent becomes leaf for next go-round.
2943 */
2944 lp = pp;
2945 }
2946}
2947
2948
2949/*
2950 * NAME: dbFindLeaf()
2951 *
2952 * FUNCTION: search a dmtree_t for sufficient free blocks, returning
2953 * the index of a leaf describing the free blocks if
2954 * sufficient free blocks are found.
2955 *
2956 * the search starts at the top of the dmtree_t tree and
2957 * proceeds down the tree to the leftmost leaf with sufficient
2958 * free space.
2959 *
2960 * PARAMETERS:
2961 * tp - pointer to the tree to be searched.
2962 * l2nb - log2 number of free blocks to search for.
2963 * leafidx - return pointer to be set to the index of the leaf
2964 * describing at least l2nb free blocks if sufficient
2965 * free blocks are found.
2966 *
2967 * RETURN VALUES:
2968 * 0 - success
2969 * -ENOSPC - insufficient free blocks.
2970 */
2971static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx)
2972{
2973 int ti, n = 0, k, x = 0;
2974
2975 /* first check the root of the tree to see if there is
2976 * sufficient free space.
2977 */
2978 if (l2nb > tp->dmt_stree[ROOT])
2979 return -ENOSPC;
2980
2981 /* sufficient free space available. now search down the tree
2982 * starting at the next level for the leftmost leaf that
2983 * describes sufficient free space.
2984 */
2985 for (k = le32_to_cpu(tp->dmt_height), ti = 1;
2986 k > 0; k--, ti = ((ti + n) << 2) + 1) {
2987 /* search the four nodes at this level, starting from
2988 * the left.
2989 */
2990 for (x = ti, n = 0; n < 4; n++) {
2991 /* sufficient free space found. move to the next
2992 * level (or quit if this is the last level).
2993 */
2994 if (l2nb <= tp->dmt_stree[x + n])
2995 break;
2996 }
2997
2998 /* better have found something since the higher
2999 * levels of the tree said it was here.
3000 */
3001 assert(n < 4);
3002 }
3003
3004 /* set the return to the leftmost leaf describing sufficient
3005 * free space.
3006 */
3007 *leafidx = x + n - le32_to_cpu(tp->dmt_leafidx);
3008
3009 return (0);
3010}
3011
3012
3013/*
3014 * NAME: dbFindBits()
3015 *
3016 * FUNCTION: find a specified number of binary buddy free bits within a
3017 * dmap bitmap word value.
3018 *
3019 * this routine searches the bitmap value for (1 << l2nb) free
3020 * bits at (1 << l2nb) alignments within the value.
3021 *
3022 * PARAMETERS:
3023 * word - dmap bitmap word value.
3024 * l2nb - number of free bits specified as a log2 number.
3025 *
3026 * RETURN VALUES:
3027 * starting bit number of free bits.
3028 */
3029static int dbFindBits(u32 word, int l2nb)
3030{
3031 int bitno, nb;
3032 u32 mask;
3033
3034 /* get the number of bits.
3035 */
3036 nb = 1 << l2nb;
3037 assert(nb <= DBWORD);
3038
3039 /* complement the word so we can use a mask (i.e. 0s represent
3040 * free bits) and compute the mask.
3041 */
3042 word = ~word;
3043 mask = ONES << (DBWORD - nb);
3044
3045 /* scan the word for nb free bits at nb alignments.
3046 */
3047 for (bitno = 0; mask != 0; bitno += nb, mask >>= nb) {
3048 if ((mask & word) == mask)
3049 break;
3050 }
3051
3052 ASSERT(bitno < 32);
3053
3054 /* return the bit number.
3055 */
3056 return (bitno);
3057}
3058
3059
3060/*
3061 * NAME: dbMaxBud(u8 *cp)
3062 *
3063 * FUNCTION: determine the largest binary buddy string of free
3064 * bits within 32-bits of the map.
3065 *
3066 * PARAMETERS:
3067 * cp - pointer to the 32-bit value.
3068 *
3069 * RETURN VALUES:
3070 * largest binary buddy of free bits within a dmap word.
3071 */
3072static int dbMaxBud(u8 * cp)
3073{
3074 signed char tmp1, tmp2;
3075
3076 /* check if the wmap word is all free. if so, the
3077 * free buddy size is BUDMIN.
3078 */
3079 if (*((uint *) cp) == 0)
3080 return (BUDMIN);
3081
3082 /* check if the wmap word is half free. if so, the
3083 * free buddy size is BUDMIN-1.
3084 */
3085 if (*((u16 *) cp) == 0 || *((u16 *) cp + 1) == 0)
3086 return (BUDMIN - 1);
3087
3088 /* not all free or half free. determine the free buddy
3089 * size thru table lookup using quarters of the wmap word.
3090 */
3091 tmp1 = max(budtab[cp[2]], budtab[cp[3]]);
3092 tmp2 = max(budtab[cp[0]], budtab[cp[1]]);
3093 return (max(tmp1, tmp2));
3094}
3095
3096
3097/*
3098 * NAME: cnttz(uint word)
3099 *
3100 * FUNCTION: determine the number of trailing zeros within a 32-bit
3101 * value.
3102 *
3103 * PARAMETERS:
3104 * value - 32-bit value to be examined.
3105 *
3106 * RETURN VALUES:
3107 * count of trailing zeros
3108 */
3109static int cnttz(u32 word)
3110{
3111 int n;
3112
3113 for (n = 0; n < 32; n++, word >>= 1) {
3114 if (word & 0x01)
3115 break;
3116 }
3117
3118 return (n);
3119}
3120
3121
3122/*
3123 * NAME: cntlz(u32 value)
3124 *
3125 * FUNCTION: determine the number of leading zeros within a 32-bit
3126 * value.
3127 *
3128 * PARAMETERS:
3129 * value - 32-bit value to be examined.
3130 *
3131 * RETURN VALUES:
3132 * count of leading zeros
3133 */
3134static int cntlz(u32 value)
3135{
3136 int n;
3137
3138 for (n = 0; n < 32; n++, value <<= 1) {
3139 if (value & HIGHORDER)
3140 break;
3141 }
3142 return (n);
3143}
3144
3145
3146/*
3147 * NAME: blkstol2(s64 nb)
3148 *
3149 * FUNCTION: convert a block count to its log2 value. if the block
3150 * count is not a l2 multiple, it is rounded up to the next
3151 * larger l2 multiple.
3152 *
3153 * PARAMETERS:
3154 * nb - number of blocks
3155 *
3156 * RETURN VALUES:
3157 * log2 number of blocks
3158 */
3159static int blkstol2(s64 nb)
3160{
3161 int l2nb;
3162 s64 mask; /* meant to be signed */
3163
3164 mask = (s64) 1 << (64 - 1);
3165
3166 /* count the leading bits.
3167 */
3168 for (l2nb = 0; l2nb < 64; l2nb++, mask >>= 1) {
3169 /* leading bit found.
3170 */
3171 if (nb & mask) {
3172 /* determine the l2 value.
3173 */
3174 l2nb = (64 - 1) - l2nb;
3175
3176 /* check if we need to round up.
3177 */
3178 if (~mask & nb)
3179 l2nb++;
3180
3181 return (l2nb);
3182 }
3183 }
3184 assert(0);
3185 return 0; /* fix compiler warning */
3186}
3187
3188
3189/*
3190 * NAME: dbAllocBottomUp()
3191 *
3192 * FUNCTION: alloc the specified block range from the working block
3193 * allocation map.
3194 *
3195 * the blocks will be alloc from the working map one dmap
3196 * at a time.
3197 *
3198 * PARAMETERS:
3199 * ip - pointer to in-core inode;
3200 * blkno - starting block number to be freed.
3201 * nblocks - number of blocks to be freed.
3202 *
3203 * RETURN VALUES:
3204 * 0 - success
3205 * -EIO - i/o error
3206 */
3207int dbAllocBottomUp(struct inode *ip, s64 blkno, s64 nblocks)
3208{
3209 struct metapage *mp;
3210 struct dmap *dp;
3211 int nb, rc;
3212 s64 lblkno, rem;
3213 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
3214 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
3215
3216 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
3217
3218 /* block to be allocated better be within the mapsize. */
3219 ASSERT(nblocks <= bmp->db_mapsize - blkno);
3220
3221 /*
3222 * allocate the blocks a dmap at a time.
3223 */
3224 mp = NULL;
3225 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
3226 /* release previous dmap if any */
3227 if (mp) {
3228 write_metapage(mp);
3229 }
3230
3231 /* get the buffer for the current dmap. */
3232 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
3233 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
3234 if (mp == NULL) {
3235 IREAD_UNLOCK(ipbmap);
3236 return -EIO;
3237 }
3238 dp = (struct dmap *) mp->data;
3239
3240 /* determine the number of blocks to be allocated from
3241 * this dmap.
3242 */
3243 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
3244
3245 /* allocate the blocks. */
3246 if ((rc = dbAllocDmapBU(bmp, dp, blkno, nb))) {
3247 release_metapage(mp);
3248 IREAD_UNLOCK(ipbmap);
3249 return (rc);
3250 }
3251 }
3252
3253 /* write the last buffer. */
3254 write_metapage(mp);
3255
3256 IREAD_UNLOCK(ipbmap);
3257
3258 return (0);
3259}
3260
3261
3262static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
3263 int nblocks)
3264{
3265 int rc;
3266 int dbitno, word, rembits, nb, nwords, wbitno, agno;
3267 s8 oldroot;
3268 struct dmaptree *tp = (struct dmaptree *) & dp->tree;
3269
3270 /* save the current value of the root (i.e. maximum free string)
3271 * of the dmap tree.
3272 */
3273 oldroot = tp->stree[ROOT];
3274
3275 /* determine the bit number and word within the dmap of the
3276 * starting block.
3277 */
3278 dbitno = blkno & (BPERDMAP - 1);
3279 word = dbitno >> L2DBWORD;
3280
3281 /* block range better be within the dmap */
3282 assert(dbitno + nblocks <= BPERDMAP);
3283
3284 /* allocate the bits of the dmap's words corresponding to the block
3285 * range. not all bits of the first and last words may be contained
3286 * within the block range. if this is the case, we'll work against
3287 * those words (i.e. partial first and/or last) on an individual basis
3288 * (a single pass), allocating the bits of interest by hand and
3289 * updating the leaf corresponding to the dmap word. a single pass
3290 * will be used for all dmap words fully contained within the
3291 * specified range. within this pass, the bits of all fully contained
3292 * dmap words will be marked as free in a single shot and the leaves
3293 * will be updated. a single leaf may describe the free space of
3294 * multiple dmap words, so we may update only a subset of the actual
3295 * leaves corresponding to the dmap words of the block range.
3296 */
3297 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
3298 /* determine the bit number within the word and
3299 * the number of bits within the word.
3300 */
3301 wbitno = dbitno & (DBWORD - 1);
3302 nb = min(rembits, DBWORD - wbitno);
3303
3304 /* check if only part of a word is to be allocated.
3305 */
3306 if (nb < DBWORD) {
3307 /* allocate (set to 1) the appropriate bits within
3308 * this dmap word.
3309 */
3310 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
3311 >> wbitno);
3312
3313 word++;
3314 } else {
3315 /* one or more dmap words are fully contained
3316 * within the block range. determine how many
3317 * words and allocate (set to 1) the bits of these
3318 * words.
3319 */
3320 nwords = rembits >> L2DBWORD;
3321 memset(&dp->wmap[word], (int) ONES, nwords * 4);
3322
3323 /* determine how many bits */
3324 nb = nwords << L2DBWORD;
3325 word += nwords;
3326 }
3327 }
3328
3329 /* update the free count for this dmap */
3330 le32_add_cpu(&dp->nfree, -nblocks);
3331
3332 /* reconstruct summary tree */
3333 dbInitDmapTree(dp);
3334
3335 BMAP_LOCK(bmp);
3336
3337 /* if this allocation group is completely free,
3338 * update the highest active allocation group number
3339 * if this allocation group is the new max.
3340 */
3341 agno = blkno >> bmp->db_agl2size;
3342 if (agno > bmp->db_maxag)
3343 bmp->db_maxag = agno;
3344
3345 /* update the free count for the allocation group and map */
3346 bmp->db_agfree[agno] -= nblocks;
3347 bmp->db_nfree -= nblocks;
3348
3349 BMAP_UNLOCK(bmp);
3350
3351 /* if the root has not changed, done. */
3352 if (tp->stree[ROOT] == oldroot)
3353 return (0);
3354
3355 /* root changed. bubble the change up to the dmap control pages.
3356 * if the adjustment of the upper level control pages fails,
3357 * backout the bit allocation (thus making everything consistent).
3358 */
3359 if ((rc = dbAdjCtl(bmp, blkno, tp->stree[ROOT], 1, 0)))
3360 dbFreeBits(bmp, dp, blkno, nblocks);
3361
3362 return (rc);
3363}
3364
3365
3366/*
3367 * NAME: dbExtendFS()
3368 *
3369 * FUNCTION: extend bmap from blkno for nblocks;
3370 * dbExtendFS() updates bmap ready for dbAllocBottomUp();
3371 *
3372 * L2
3373 * |
3374 * L1---------------------------------L1
3375 * | |
3376 * L0---------L0---------L0 L0---------L0---------L0
3377 * | | | | | |
3378 * d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,.,dm;
3379 * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm
3380 *
3381 * <---old---><----------------------------extend----------------------->
3382 */
3383int dbExtendFS(struct inode *ipbmap, s64 blkno, s64 nblocks)
3384{
3385 struct jfs_sb_info *sbi = JFS_SBI(ipbmap->i_sb);
3386 int nbperpage = sbi->nbperpage;
3387 int i, i0 = true, j, j0 = true, k, n;
3388 s64 newsize;
3389 s64 p;
3390 struct metapage *mp, *l2mp, *l1mp = NULL, *l0mp = NULL;
3391 struct dmapctl *l2dcp, *l1dcp, *l0dcp;
3392 struct dmap *dp;
3393 s8 *l0leaf, *l1leaf, *l2leaf;
3394 struct bmap *bmp = sbi->bmap;
3395 int agno, l2agsize, oldl2agsize;
3396 s64 ag_rem;
3397
3398 newsize = blkno + nblocks;
3399
3400 jfs_info("dbExtendFS: blkno:%Ld nblocks:%Ld newsize:%Ld",
3401 (long long) blkno, (long long) nblocks, (long long) newsize);
3402
3403 /*
3404 * initialize bmap control page.
3405 *
3406 * all the data in bmap control page should exclude
3407 * the mkfs hidden dmap page.
3408 */
3409
3410 /* update mapsize */
3411 bmp->db_mapsize = newsize;
3412 bmp->db_maxlevel = BMAPSZTOLEV(bmp->db_mapsize);
3413
3414 /* compute new AG size */
3415 l2agsize = dbGetL2AGSize(newsize);
3416 oldl2agsize = bmp->db_agl2size;
3417
3418 bmp->db_agl2size = l2agsize;
3419 bmp->db_agsize = 1 << l2agsize;
3420
3421 /* compute new number of AG */
3422 agno = bmp->db_numag;
3423 bmp->db_numag = newsize >> l2agsize;
3424 bmp->db_numag += ((u32) newsize % (u32) bmp->db_agsize) ? 1 : 0;
3425
3426 /*
3427 * reconfigure db_agfree[]
3428 * from old AG configuration to new AG configuration;
3429 *
3430 * coalesce contiguous k (newAGSize/oldAGSize) AGs;
3431 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn;
3432 * note: new AG size = old AG size * (2**x).
3433 */
3434 if (l2agsize == oldl2agsize)
3435 goto extend;
3436 k = 1 << (l2agsize - oldl2agsize);
3437 ag_rem = bmp->db_agfree[0]; /* save agfree[0] */
3438 for (i = 0, n = 0; i < agno; n++) {
3439 bmp->db_agfree[n] = 0; /* init collection point */
3440
3441 /* coalesce contiguous k AGs; */
3442 for (j = 0; j < k && i < agno; j++, i++) {
3443 /* merge AGi to AGn */
3444 bmp->db_agfree[n] += bmp->db_agfree[i];
3445 }
3446 }
3447 bmp->db_agfree[0] += ag_rem; /* restore agfree[0] */
3448
3449 for (; n < MAXAG; n++)
3450 bmp->db_agfree[n] = 0;
3451
3452 /*
3453 * update highest active ag number
3454 */
3455
3456 bmp->db_maxag = bmp->db_maxag / k;
3457
3458 /*
3459 * extend bmap
3460 *
3461 * update bit maps and corresponding level control pages;
3462 * global control page db_nfree, db_agfree[agno], db_maxfreebud;
3463 */
3464 extend:
3465 /* get L2 page */
3466 p = BMAPBLKNO + nbperpage; /* L2 page */
3467 l2mp = read_metapage(ipbmap, p, PSIZE, 0);
3468 if (!l2mp) {
3469 jfs_error(ipbmap->i_sb, "L2 page could not be read\n");
3470 return -EIO;
3471 }
3472 l2dcp = (struct dmapctl *) l2mp->data;
3473
3474 /* compute start L1 */
3475 k = blkno >> L2MAXL1SIZE;
3476 l2leaf = l2dcp->stree + CTLLEAFIND + k;
3477 p = BLKTOL1(blkno, sbi->l2nbperpage); /* L1 page */
3478
3479 /*
3480 * extend each L1 in L2
3481 */
3482 for (; k < LPERCTL; k++, p += nbperpage) {
3483 /* get L1 page */
3484 if (j0) {
3485 /* read in L1 page: (blkno & (MAXL1SIZE - 1)) */
3486 l1mp = read_metapage(ipbmap, p, PSIZE, 0);
3487 if (l1mp == NULL)
3488 goto errout;
3489 l1dcp = (struct dmapctl *) l1mp->data;
3490
3491 /* compute start L0 */
3492 j = (blkno & (MAXL1SIZE - 1)) >> L2MAXL0SIZE;
3493 l1leaf = l1dcp->stree + CTLLEAFIND + j;
3494 p = BLKTOL0(blkno, sbi->l2nbperpage);
3495 j0 = false;
3496 } else {
3497 /* assign/init L1 page */
3498 l1mp = get_metapage(ipbmap, p, PSIZE, 0);
3499 if (l1mp == NULL)
3500 goto errout;
3501
3502 l1dcp = (struct dmapctl *) l1mp->data;
3503
3504 /* compute start L0 */
3505 j = 0;
3506 l1leaf = l1dcp->stree + CTLLEAFIND;
3507 p += nbperpage; /* 1st L0 of L1.k */
3508 }
3509
3510 /*
3511 * extend each L0 in L1
3512 */
3513 for (; j < LPERCTL; j++) {
3514 /* get L0 page */
3515 if (i0) {
3516 /* read in L0 page: (blkno & (MAXL0SIZE - 1)) */
3517
3518 l0mp = read_metapage(ipbmap, p, PSIZE, 0);
3519 if (l0mp == NULL)
3520 goto errout;
3521 l0dcp = (struct dmapctl *) l0mp->data;
3522
3523 /* compute start dmap */
3524 i = (blkno & (MAXL0SIZE - 1)) >>
3525 L2BPERDMAP;
3526 l0leaf = l0dcp->stree + CTLLEAFIND + i;
3527 p = BLKTODMAP(blkno,
3528 sbi->l2nbperpage);
3529 i0 = false;
3530 } else {
3531 /* assign/init L0 page */
3532 l0mp = get_metapage(ipbmap, p, PSIZE, 0);
3533 if (l0mp == NULL)
3534 goto errout;
3535
3536 l0dcp = (struct dmapctl *) l0mp->data;
3537
3538 /* compute start dmap */
3539 i = 0;
3540 l0leaf = l0dcp->stree + CTLLEAFIND;
3541 p += nbperpage; /* 1st dmap of L0.j */
3542 }
3543
3544 /*
3545 * extend each dmap in L0
3546 */
3547 for (; i < LPERCTL; i++) {
3548 /*
3549 * reconstruct the dmap page, and
3550 * initialize corresponding parent L0 leaf
3551 */
3552 if ((n = blkno & (BPERDMAP - 1))) {
3553 /* read in dmap page: */
3554 mp = read_metapage(ipbmap, p,
3555 PSIZE, 0);
3556 if (mp == NULL)
3557 goto errout;
3558 n = min(nblocks, (s64)BPERDMAP - n);
3559 } else {
3560 /* assign/init dmap page */
3561 mp = read_metapage(ipbmap, p,
3562 PSIZE, 0);
3563 if (mp == NULL)
3564 goto errout;
3565
3566 n = min(nblocks, (s64)BPERDMAP);
3567 }
3568
3569 dp = (struct dmap *) mp->data;
3570 *l0leaf = dbInitDmap(dp, blkno, n);
3571
3572 bmp->db_nfree += n;
3573 agno = le64_to_cpu(dp->start) >> l2agsize;
3574 bmp->db_agfree[agno] += n;
3575
3576 write_metapage(mp);
3577
3578 l0leaf++;
3579 p += nbperpage;
3580
3581 blkno += n;
3582 nblocks -= n;
3583 if (nblocks == 0)
3584 break;
3585 } /* for each dmap in a L0 */
3586
3587 /*
3588 * build current L0 page from its leaves, and
3589 * initialize corresponding parent L1 leaf
3590 */
3591 *l1leaf = dbInitDmapCtl(l0dcp, 0, ++i);
3592 write_metapage(l0mp);
3593 l0mp = NULL;
3594
3595 if (nblocks)
3596 l1leaf++; /* continue for next L0 */
3597 else {
3598 /* more than 1 L0 ? */
3599 if (j > 0)
3600 break; /* build L1 page */
3601 else {
3602 /* summarize in global bmap page */
3603 bmp->db_maxfreebud = *l1leaf;
3604 release_metapage(l1mp);
3605 release_metapage(l2mp);
3606 goto finalize;
3607 }
3608 }
3609 } /* for each L0 in a L1 */
3610
3611 /*
3612 * build current L1 page from its leaves, and
3613 * initialize corresponding parent L2 leaf
3614 */
3615 *l2leaf = dbInitDmapCtl(l1dcp, 1, ++j);
3616 write_metapage(l1mp);
3617 l1mp = NULL;
3618
3619 if (nblocks)
3620 l2leaf++; /* continue for next L1 */
3621 else {
3622 /* more than 1 L1 ? */
3623 if (k > 0)
3624 break; /* build L2 page */
3625 else {
3626 /* summarize in global bmap page */
3627 bmp->db_maxfreebud = *l2leaf;
3628 release_metapage(l2mp);
3629 goto finalize;
3630 }
3631 }
3632 } /* for each L1 in a L2 */
3633
3634 jfs_error(ipbmap->i_sb, "function has not returned as expected\n");
3635errout:
3636 if (l0mp)
3637 release_metapage(l0mp);
3638 if (l1mp)
3639 release_metapage(l1mp);
3640 release_metapage(l2mp);
3641 return -EIO;
3642
3643 /*
3644 * finalize bmap control page
3645 */
3646finalize:
3647
3648 return 0;
3649}
3650
3651
3652/*
3653 * dbFinalizeBmap()
3654 */
3655void dbFinalizeBmap(struct inode *ipbmap)
3656{
3657 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
3658 int actags, inactags, l2nl;
3659 s64 ag_rem, actfree, inactfree, avgfree;
3660 int i, n;
3661
3662 /*
3663 * finalize bmap control page
3664 */
3665//finalize:
3666 /*
3667 * compute db_agpref: preferred ag to allocate from
3668 * (the leftmost ag with average free space in it);
3669 */
3670//agpref:
3671 /* get the number of active ags and inacitve ags */
3672 actags = bmp->db_maxag + 1;
3673 inactags = bmp->db_numag - actags;
3674 ag_rem = bmp->db_mapsize & (bmp->db_agsize - 1); /* ??? */
3675
3676 /* determine how many blocks are in the inactive allocation
3677 * groups. in doing this, we must account for the fact that
3678 * the rightmost group might be a partial group (i.e. file
3679 * system size is not a multiple of the group size).
3680 */
3681 inactfree = (inactags && ag_rem) ?
3682 ((inactags - 1) << bmp->db_agl2size) + ag_rem
3683 : inactags << bmp->db_agl2size;
3684
3685 /* determine how many free blocks are in the active
3686 * allocation groups plus the average number of free blocks
3687 * within the active ags.
3688 */
3689 actfree = bmp->db_nfree - inactfree;
3690 avgfree = (u32) actfree / (u32) actags;
3691
3692 /* if the preferred allocation group has not average free space.
3693 * re-establish the preferred group as the leftmost
3694 * group with average free space.
3695 */
3696 if (bmp->db_agfree[bmp->db_agpref] < avgfree) {
3697 for (bmp->db_agpref = 0; bmp->db_agpref < actags;
3698 bmp->db_agpref++) {
3699 if (bmp->db_agfree[bmp->db_agpref] >= avgfree)
3700 break;
3701 }
3702 if (bmp->db_agpref >= bmp->db_numag) {
3703 jfs_error(ipbmap->i_sb,
3704 "cannot find ag with average freespace\n");
3705 }
3706 }
3707
3708 /*
3709 * compute db_aglevel, db_agheight, db_width, db_agstart:
3710 * an ag is covered in aglevel dmapctl summary tree,
3711 * at agheight level height (from leaf) with agwidth number of nodes
3712 * each, which starts at agstart index node of the smmary tree node
3713 * array;
3714 */
3715 bmp->db_aglevel = BMAPSZTOLEV(bmp->db_agsize);
3716 l2nl =
3717 bmp->db_agl2size - (L2BPERDMAP + bmp->db_aglevel * L2LPERCTL);
3718 bmp->db_agheight = l2nl >> 1;
3719 bmp->db_agwidth = 1 << (l2nl - (bmp->db_agheight << 1));
3720 for (i = 5 - bmp->db_agheight, bmp->db_agstart = 0, n = 1; i > 0;
3721 i--) {
3722 bmp->db_agstart += n;
3723 n <<= 2;
3724 }
3725
3726}
3727
3728
3729/*
3730 * NAME: dbInitDmap()/ujfs_idmap_page()
3731 *
3732 * FUNCTION: initialize working/persistent bitmap of the dmap page
3733 * for the specified number of blocks:
3734 *
3735 * at entry, the bitmaps had been initialized as free (ZEROS);
3736 * The number of blocks will only account for the actually
3737 * existing blocks. Blocks which don't actually exist in
3738 * the aggregate will be marked as allocated (ONES);
3739 *
3740 * PARAMETERS:
3741 * dp - pointer to page of map
3742 * nblocks - number of blocks this page
3743 *
3744 * RETURNS: NONE
3745 */
3746static int dbInitDmap(struct dmap * dp, s64 Blkno, int nblocks)
3747{
3748 int blkno, w, b, r, nw, nb, i;
3749
3750 /* starting block number within the dmap */
3751 blkno = Blkno & (BPERDMAP - 1);
3752
3753 if (blkno == 0) {
3754 dp->nblocks = dp->nfree = cpu_to_le32(nblocks);
3755 dp->start = cpu_to_le64(Blkno);
3756
3757 if (nblocks == BPERDMAP) {
3758 memset(&dp->wmap[0], 0, LPERDMAP * 4);
3759 memset(&dp->pmap[0], 0, LPERDMAP * 4);
3760 goto initTree;
3761 }
3762 } else {
3763 le32_add_cpu(&dp->nblocks, nblocks);
3764 le32_add_cpu(&dp->nfree, nblocks);
3765 }
3766
3767 /* word number containing start block number */
3768 w = blkno >> L2DBWORD;
3769
3770 /*
3771 * free the bits corresponding to the block range (ZEROS):
3772 * note: not all bits of the first and last words may be contained
3773 * within the block range.
3774 */
3775 for (r = nblocks; r > 0; r -= nb, blkno += nb) {
3776 /* number of bits preceding range to be freed in the word */
3777 b = blkno & (DBWORD - 1);
3778 /* number of bits to free in the word */
3779 nb = min(r, DBWORD - b);
3780
3781 /* is partial word to be freed ? */
3782 if (nb < DBWORD) {
3783 /* free (set to 0) from the bitmap word */
3784 dp->wmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3785 >> b));
3786 dp->pmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3787 >> b));
3788
3789 /* skip the word freed */
3790 w++;
3791 } else {
3792 /* free (set to 0) contiguous bitmap words */
3793 nw = r >> L2DBWORD;
3794 memset(&dp->wmap[w], 0, nw * 4);
3795 memset(&dp->pmap[w], 0, nw * 4);
3796
3797 /* skip the words freed */
3798 nb = nw << L2DBWORD;
3799 w += nw;
3800 }
3801 }
3802
3803 /*
3804 * mark bits following the range to be freed (non-existing
3805 * blocks) as allocated (ONES)
3806 */
3807
3808 if (blkno == BPERDMAP)
3809 goto initTree;
3810
3811 /* the first word beyond the end of existing blocks */
3812 w = blkno >> L2DBWORD;
3813
3814 /* does nblocks fall on a 32-bit boundary ? */
3815 b = blkno & (DBWORD - 1);
3816 if (b) {
3817 /* mark a partial word allocated */
3818 dp->wmap[w] = dp->pmap[w] = cpu_to_le32(ONES >> b);
3819 w++;
3820 }
3821
3822 /* set the rest of the words in the page to allocated (ONES) */
3823 for (i = w; i < LPERDMAP; i++)
3824 dp->pmap[i] = dp->wmap[i] = cpu_to_le32(ONES);
3825
3826 /*
3827 * init tree
3828 */
3829 initTree:
3830 return (dbInitDmapTree(dp));
3831}
3832
3833
3834/*
3835 * NAME: dbInitDmapTree()/ujfs_complete_dmap()
3836 *
3837 * FUNCTION: initialize summary tree of the specified dmap:
3838 *
3839 * at entry, bitmap of the dmap has been initialized;
3840 *
3841 * PARAMETERS:
3842 * dp - dmap to complete
3843 * blkno - starting block number for this dmap
3844 * treemax - will be filled in with max free for this dmap
3845 *
3846 * RETURNS: max free string at the root of the tree
3847 */
3848static int dbInitDmapTree(struct dmap * dp)
3849{
3850 struct dmaptree *tp;
3851 s8 *cp;
3852 int i;
3853
3854 /* init fixed info of tree */
3855 tp = &dp->tree;
3856 tp->nleafs = cpu_to_le32(LPERDMAP);
3857 tp->l2nleafs = cpu_to_le32(L2LPERDMAP);
3858 tp->leafidx = cpu_to_le32(LEAFIND);
3859 tp->height = cpu_to_le32(4);
3860 tp->budmin = BUDMIN;
3861
3862 /* init each leaf from corresponding wmap word:
3863 * note: leaf is set to NOFREE(-1) if all blocks of corresponding
3864 * bitmap word are allocated.
3865 */
3866 cp = tp->stree + le32_to_cpu(tp->leafidx);
3867 for (i = 0; i < LPERDMAP; i++)
3868 *cp++ = dbMaxBud((u8 *) & dp->wmap[i]);
3869
3870 /* build the dmap's binary buddy summary tree */
3871 return (dbInitTree(tp));
3872}
3873
3874
3875/*
3876 * NAME: dbInitTree()/ujfs_adjtree()
3877 *
3878 * FUNCTION: initialize binary buddy summary tree of a dmap or dmapctl.
3879 *
3880 * at entry, the leaves of the tree has been initialized
3881 * from corresponding bitmap word or root of summary tree
3882 * of the child control page;
3883 * configure binary buddy system at the leaf level, then
3884 * bubble up the values of the leaf nodes up the tree.
3885 *
3886 * PARAMETERS:
3887 * cp - Pointer to the root of the tree
3888 * l2leaves- Number of leaf nodes as a power of 2
3889 * l2min - Number of blocks that can be covered by a leaf
3890 * as a power of 2
3891 *
3892 * RETURNS: max free string at the root of the tree
3893 */
3894static int dbInitTree(struct dmaptree * dtp)
3895{
3896 int l2max, l2free, bsize, nextb, i;
3897 int child, parent, nparent;
3898 s8 *tp, *cp, *cp1;
3899
3900 tp = dtp->stree;
3901
3902 /* Determine the maximum free string possible for the leaves */
3903 l2max = le32_to_cpu(dtp->l2nleafs) + dtp->budmin;
3904
3905 /*
3906 * configure the leaf levevl into binary buddy system
3907 *
3908 * Try to combine buddies starting with a buddy size of 1
3909 * (i.e. two leaves). At a buddy size of 1 two buddy leaves
3910 * can be combined if both buddies have a maximum free of l2min;
3911 * the combination will result in the left-most buddy leaf having
3912 * a maximum free of l2min+1.
3913 * After processing all buddies for a given size, process buddies
3914 * at the next higher buddy size (i.e. current size * 2) and
3915 * the next maximum free (current free + 1).
3916 * This continues until the maximum possible buddy combination
3917 * yields maximum free.
3918 */
3919 for (l2free = dtp->budmin, bsize = 1; l2free < l2max;
3920 l2free++, bsize = nextb) {
3921 /* get next buddy size == current buddy pair size */
3922 nextb = bsize << 1;
3923
3924 /* scan each adjacent buddy pair at current buddy size */
3925 for (i = 0, cp = tp + le32_to_cpu(dtp->leafidx);
3926 i < le32_to_cpu(dtp->nleafs);
3927 i += nextb, cp += nextb) {
3928 /* coalesce if both adjacent buddies are max free */
3929 if (*cp == l2free && *(cp + bsize) == l2free) {
3930 *cp = l2free + 1; /* left take right */
3931 *(cp + bsize) = -1; /* right give left */
3932 }
3933 }
3934 }
3935
3936 /*
3937 * bubble summary information of leaves up the tree.
3938 *
3939 * Starting at the leaf node level, the four nodes described by
3940 * the higher level parent node are compared for a maximum free and
3941 * this maximum becomes the value of the parent node.
3942 * when all lower level nodes are processed in this fashion then
3943 * move up to the next level (parent becomes a lower level node) and
3944 * continue the process for that level.
3945 */
3946 for (child = le32_to_cpu(dtp->leafidx),
3947 nparent = le32_to_cpu(dtp->nleafs) >> 2;
3948 nparent > 0; nparent >>= 2, child = parent) {
3949 /* get index of 1st node of parent level */
3950 parent = (child - 1) >> 2;
3951
3952 /* set the value of the parent node as the maximum
3953 * of the four nodes of the current level.
3954 */
3955 for (i = 0, cp = tp + child, cp1 = tp + parent;
3956 i < nparent; i++, cp += 4, cp1++)
3957 *cp1 = TREEMAX(cp);
3958 }
3959
3960 return (*tp);
3961}
3962
3963
3964/*
3965 * dbInitDmapCtl()
3966 *
3967 * function: initialize dmapctl page
3968 */
3969static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i)
3970{ /* start leaf index not covered by range */
3971 s8 *cp;
3972
3973 dcp->nleafs = cpu_to_le32(LPERCTL);
3974 dcp->l2nleafs = cpu_to_le32(L2LPERCTL);
3975 dcp->leafidx = cpu_to_le32(CTLLEAFIND);
3976 dcp->height = cpu_to_le32(5);
3977 dcp->budmin = L2BPERDMAP + L2LPERCTL * level;
3978
3979 /*
3980 * initialize the leaves of current level that were not covered
3981 * by the specified input block range (i.e. the leaves have no
3982 * low level dmapctl or dmap).
3983 */
3984 cp = &dcp->stree[CTLLEAFIND + i];
3985 for (; i < LPERCTL; i++)
3986 *cp++ = NOFREE;
3987
3988 /* build the dmap's binary buddy summary tree */
3989 return (dbInitTree((struct dmaptree *) dcp));
3990}
3991
3992
3993/*
3994 * NAME: dbGetL2AGSize()/ujfs_getagl2size()
3995 *
3996 * FUNCTION: Determine log2(allocation group size) from aggregate size
3997 *
3998 * PARAMETERS:
3999 * nblocks - Number of blocks in aggregate
4000 *
4001 * RETURNS: log2(allocation group size) in aggregate blocks
4002 */
4003static int dbGetL2AGSize(s64 nblocks)
4004{
4005 s64 sz;
4006 s64 m;
4007 int l2sz;
4008
4009 if (nblocks < BPERDMAP * MAXAG)
4010 return (L2BPERDMAP);
4011
4012 /* round up aggregate size to power of 2 */
4013 m = ((u64) 1 << (64 - 1));
4014 for (l2sz = 64; l2sz >= 0; l2sz--, m >>= 1) {
4015 if (m & nblocks)
4016 break;
4017 }
4018
4019 sz = (s64) 1 << l2sz;
4020 if (sz < nblocks)
4021 l2sz += 1;
4022
4023 /* agsize = roundupSize/max_number_of_ag */
4024 return (l2sz - L2MAXAG);
4025}
4026
4027
4028/*
4029 * NAME: dbMapFileSizeToMapSize()
4030 *
4031 * FUNCTION: compute number of blocks the block allocation map file
4032 * can cover from the map file size;
4033 *
4034 * RETURNS: Number of blocks which can be covered by this block map file;
4035 */
4036
4037/*
4038 * maximum number of map pages at each level including control pages
4039 */
4040#define MAXL0PAGES (1 + LPERCTL)
4041#define MAXL1PAGES (1 + LPERCTL * MAXL0PAGES)
4042#define MAXL2PAGES (1 + LPERCTL * MAXL1PAGES)
4043
4044/*
4045 * convert number of map pages to the zero origin top dmapctl level
4046 */
4047#define BMAPPGTOLEV(npages) \
4048 (((npages) <= 3 + MAXL0PAGES) ? 0 : \
4049 ((npages) <= 2 + MAXL1PAGES) ? 1 : 2)
4050
4051s64 dbMapFileSizeToMapSize(struct inode * ipbmap)
4052{
4053 struct super_block *sb = ipbmap->i_sb;
4054 s64 nblocks;
4055 s64 npages, ndmaps;
4056 int level, i;
4057 int complete, factor;
4058
4059 nblocks = ipbmap->i_size >> JFS_SBI(sb)->l2bsize;
4060 npages = nblocks >> JFS_SBI(sb)->l2nbperpage;
4061 level = BMAPPGTOLEV(npages);
4062
4063 /* At each level, accumulate the number of dmap pages covered by
4064 * the number of full child levels below it;
4065 * repeat for the last incomplete child level.
4066 */
4067 ndmaps = 0;
4068 npages--; /* skip the first global control page */
4069 /* skip higher level control pages above top level covered by map */
4070 npages -= (2 - level);
4071 npages--; /* skip top level's control page */
4072 for (i = level; i >= 0; i--) {
4073 factor =
4074 (i == 2) ? MAXL1PAGES : ((i == 1) ? MAXL0PAGES : 1);
4075 complete = (u32) npages / factor;
4076 ndmaps += complete * ((i == 2) ? LPERCTL * LPERCTL :
4077 ((i == 1) ? LPERCTL : 1));
4078
4079 /* pages in last/incomplete child */
4080 npages = (u32) npages % factor;
4081 /* skip incomplete child's level control page */
4082 npages--;
4083 }
4084
4085 /* convert the number of dmaps into the number of blocks
4086 * which can be covered by the dmaps;
4087 */
4088 nblocks = ndmaps << L2BPERDMAP;
4089
4090 return (nblocks);
4091}