xref: /linux/fs/jfs/jfs_dmap.c (revision 68550cbc6129159b7a6434796b721e8b66ee12f6)
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  */
64 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
65 			int nblocks);
66 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval);
67 static int dbBackSplit(dmtree_t * tp, int leafno);
68 static int dbJoin(dmtree_t * tp, int leafno, int newval);
69 static void dbAdjTree(dmtree_t * tp, int leafno, int newval);
70 static int dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc,
71 		    int level);
72 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results);
73 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
74 		       int nblocks);
75 static int dbAllocNear(struct bmap * bmp, struct dmap * dp, s64 blkno,
76 		       int nblocks,
77 		       int l2nb, s64 * results);
78 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
79 		       int nblocks);
80 static int dbAllocDmapLev(struct bmap * bmp, struct dmap * dp, int nblocks,
81 			  int l2nb,
82 			  s64 * results);
83 static int dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb,
84 		     s64 * results);
85 static int dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno,
86 		      s64 * results);
87 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks);
88 static int dbFindBits(u32 word, int l2nb);
89 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno);
90 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx);
91 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
92 		      int nblocks);
93 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
94 		      int nblocks);
95 static int dbMaxBud(u8 * cp);
96 static int blkstol2(s64 nb);
97 
98 static int cntlz(u32 value);
99 static int cnttz(u32 word);
100 
101 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
102 			 int nblocks);
103 static int dbInitDmap(struct dmap * dp, s64 blkno, int nblocks);
104 static int dbInitDmapTree(struct dmap * dp);
105 static int dbInitTree(struct dmaptree * dtp);
106 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i);
107 static int dbGetL2AGSize(s64 nblocks);
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  */
117 static const s8 budtab[256] = {
118 	3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
119 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
120 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
121 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
122 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
123 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
124 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
125 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
126 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
127 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
128 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
129 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
130 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
131 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
132 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
133 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, -1
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  */
152 int 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  */
229 int 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  */
250 int 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  */
318 int 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  */
410 int
411 dbUpdatePMap(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  */
590 int 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 */
641 unlock:
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
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  */
687 int 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  */
880 int 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  */
964 int
965 dbReAlloc(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  */
1013 static 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  */
1112 static 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  */
1243 static int
1244 dbAllocNear(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  */
1352 static int
1353 dbAllocAG(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  */
1551 static 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  */
1606 s64 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 = 1LL << 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  */
1714 static 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  */
1832 static int
1833 dbAllocCtl(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  */
1983 static int
1984 dbAllocDmapLev(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  */
2047 static 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  */
2102 static 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  */
2164 static 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  */
2311 static 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  */
2494 static int
2495 dbAdjCtl(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 				release_metapage(mp);
2554 				return rc;
2555 			}
2556 			oldval = dcp->stree[ti];
2557 		}
2558 		dbSplit((dmtree_t *) dcp, leafno, dcp->budmin, newval);
2559 	} else {
2560 		rc = dbJoin((dmtree_t *) dcp, leafno, newval);
2561 		if (rc) {
2562 			release_metapage(mp);
2563 			return rc;
2564 		}
2565 	}
2566 
2567 	/* check if the root of the current dmap control page changed due
2568 	 * to the update and if the current dmap control page is not at
2569 	 * the current top level (i.e. L0, L1, L2) of the map.  if so (i.e.
2570 	 * root changed and this is not the top level), call this routine
2571 	 * again (recursion) for the next higher level of the mapping to
2572 	 * reflect the change in root for the current dmap control page.
2573 	 */
2574 	if (dcp->stree[ROOT] != oldroot) {
2575 		/* are we below the top level of the map.  if so,
2576 		 * bubble the root up to the next higher level.
2577 		 */
2578 		if (level < bmp->db_maxlevel) {
2579 			/* bubble up the new root of this dmap control page to
2580 			 * the next level.
2581 			 */
2582 			if ((rc =
2583 			     dbAdjCtl(bmp, blkno, dcp->stree[ROOT], alloc,
2584 				      level + 1))) {
2585 				/* something went wrong in bubbling up the new
2586 				 * root value, so backout the changes to the
2587 				 * current dmap control page.
2588 				 */
2589 				if (alloc) {
2590 					dbJoin((dmtree_t *) dcp, leafno,
2591 					       oldval);
2592 				} else {
2593 					/* the dbJoin() above might have
2594 					 * caused a larger binary buddy system
2595 					 * to form and we may now be in the
2596 					 * middle of it.  if this is the case,
2597 					 * back split the buddies.
2598 					 */
2599 					if (dcp->stree[ti] == NOFREE)
2600 						dbBackSplit((dmtree_t *)
2601 							    dcp, leafno);
2602 					dbSplit((dmtree_t *) dcp, leafno,
2603 						dcp->budmin, oldval);
2604 				}
2605 
2606 				/* release the buffer and return the error.
2607 				 */
2608 				release_metapage(mp);
2609 				return (rc);
2610 			}
2611 		} else {
2612 			/* we're at the top level of the map. update
2613 			 * the bmap control page to reflect the size
2614 			 * of the maximum free buddy system.
2615 			 */
2616 			assert(level == bmp->db_maxlevel);
2617 			if (bmp->db_maxfreebud != oldroot) {
2618 				jfs_error(bmp->db_ipbmap->i_sb,
2619 					  "the maximum free buddy is not the old root\n");
2620 			}
2621 			bmp->db_maxfreebud = dcp->stree[ROOT];
2622 		}
2623 	}
2624 
2625 	/* write the buffer.
2626 	 */
2627 	write_metapage(mp);
2628 
2629 	return (0);
2630 }
2631 
2632 
2633 /*
2634  * NAME:	dbSplit()
2635  *
2636  * FUNCTION:	update the leaf of a dmtree with a new value, splitting
2637  *		the leaf from the binary buddy system of the dmtree's
2638  *		leaves, as required.
2639  *
2640  * PARAMETERS:
2641  *	tp	- pointer to the tree containing the leaf.
2642  *	leafno	- the number of the leaf to be updated.
2643  *	splitsz	- the size the binary buddy system starting at the leaf
2644  *		  must be split to, specified as the log2 number of blocks.
2645  *	newval	- the new value for the leaf.
2646  *
2647  * RETURN VALUES: none
2648  *
2649  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2650  */
2651 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval)
2652 {
2653 	int budsz;
2654 	int cursz;
2655 	s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2656 
2657 	/* check if the leaf needs to be split.
2658 	 */
2659 	if (leaf[leafno] > tp->dmt_budmin) {
2660 		/* the split occurs by cutting the buddy system in half
2661 		 * at the specified leaf until we reach the specified
2662 		 * size.  pick up the starting split size (current size
2663 		 * - 1 in l2) and the corresponding buddy size.
2664 		 */
2665 		cursz = leaf[leafno] - 1;
2666 		budsz = BUDSIZE(cursz, tp->dmt_budmin);
2667 
2668 		/* split until we reach the specified size.
2669 		 */
2670 		while (cursz >= splitsz) {
2671 			/* update the buddy's leaf with its new value.
2672 			 */
2673 			dbAdjTree(tp, leafno ^ budsz, cursz);
2674 
2675 			/* on to the next size and buddy.
2676 			 */
2677 			cursz -= 1;
2678 			budsz >>= 1;
2679 		}
2680 	}
2681 
2682 	/* adjust the dmap tree to reflect the specified leaf's new
2683 	 * value.
2684 	 */
2685 	dbAdjTree(tp, leafno, newval);
2686 }
2687 
2688 
2689 /*
2690  * NAME:	dbBackSplit()
2691  *
2692  * FUNCTION:	back split the binary buddy system of dmtree leaves
2693  *		that hold a specified leaf until the specified leaf
2694  *		starts its own binary buddy system.
2695  *
2696  *		the allocators typically perform allocations at the start
2697  *		of binary buddy systems and dbSplit() is used to accomplish
2698  *		any required splits.  in some cases, however, allocation
2699  *		may occur in the middle of a binary system and requires a
2700  *		back split, with the split proceeding out from the middle of
2701  *		the system (less efficient) rather than the start of the
2702  *		system (more efficient).  the cases in which a back split
2703  *		is required are rare and are limited to the first allocation
2704  *		within an allocation group which is a part (not first part)
2705  *		of a larger binary buddy system and a few exception cases
2706  *		in which a previous join operation must be backed out.
2707  *
2708  * PARAMETERS:
2709  *	tp	- pointer to the tree containing the leaf.
2710  *	leafno	- the number of the leaf to be updated.
2711  *
2712  * RETURN VALUES: none
2713  *
2714  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2715  */
2716 static int dbBackSplit(dmtree_t * tp, int leafno)
2717 {
2718 	int budsz, bud, w, bsz, size;
2719 	int cursz;
2720 	s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2721 
2722 	/* leaf should be part (not first part) of a binary
2723 	 * buddy system.
2724 	 */
2725 	assert(leaf[leafno] == NOFREE);
2726 
2727 	/* the back split is accomplished by iteratively finding the leaf
2728 	 * that starts the buddy system that contains the specified leaf and
2729 	 * splitting that system in two.  this iteration continues until
2730 	 * the specified leaf becomes the start of a buddy system.
2731 	 *
2732 	 * determine maximum possible l2 size for the specified leaf.
2733 	 */
2734 	size =
2735 	    LITOL2BSZ(leafno, le32_to_cpu(tp->dmt_l2nleafs),
2736 		      tp->dmt_budmin);
2737 
2738 	/* determine the number of leaves covered by this size.  this
2739 	 * is the buddy size that we will start with as we search for
2740 	 * the buddy system that contains the specified leaf.
2741 	 */
2742 	budsz = BUDSIZE(size, tp->dmt_budmin);
2743 
2744 	/* back split.
2745 	 */
2746 	while (leaf[leafno] == NOFREE) {
2747 		/* find the leftmost buddy leaf.
2748 		 */
2749 		for (w = leafno, bsz = budsz;; bsz <<= 1,
2750 		     w = (w < bud) ? w : bud) {
2751 			if (bsz >= le32_to_cpu(tp->dmt_nleafs)) {
2752 				jfs_err("JFS: block map error in dbBackSplit");
2753 				return -EIO;
2754 			}
2755 
2756 			/* determine the buddy.
2757 			 */
2758 			bud = w ^ bsz;
2759 
2760 			/* check if this buddy is the start of the system.
2761 			 */
2762 			if (leaf[bud] != NOFREE) {
2763 				/* split the leaf at the start of the
2764 				 * system in two.
2765 				 */
2766 				cursz = leaf[bud] - 1;
2767 				dbSplit(tp, bud, cursz, cursz);
2768 				break;
2769 			}
2770 		}
2771 	}
2772 
2773 	if (leaf[leafno] != size) {
2774 		jfs_err("JFS: wrong leaf value in dbBackSplit");
2775 		return -EIO;
2776 	}
2777 	return 0;
2778 }
2779 
2780 
2781 /*
2782  * NAME:	dbJoin()
2783  *
2784  * FUNCTION:	update the leaf of a dmtree with a new value, joining
2785  *		the leaf with other leaves of the dmtree into a multi-leaf
2786  *		binary buddy system, as required.
2787  *
2788  * PARAMETERS:
2789  *	tp	- pointer to the tree containing the leaf.
2790  *	leafno	- the number of the leaf to be updated.
2791  *	newval	- the new value for the leaf.
2792  *
2793  * RETURN VALUES: none
2794  */
2795 static int dbJoin(dmtree_t * tp, int leafno, int newval)
2796 {
2797 	int budsz, buddy;
2798 	s8 *leaf;
2799 
2800 	/* can the new leaf value require a join with other leaves ?
2801 	 */
2802 	if (newval >= tp->dmt_budmin) {
2803 		/* pickup a pointer to the leaves of the tree.
2804 		 */
2805 		leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2806 
2807 		/* try to join the specified leaf into a large binary
2808 		 * buddy system.  the join proceeds by attempting to join
2809 		 * the specified leafno with its buddy (leaf) at new value.
2810 		 * if the join occurs, we attempt to join the left leaf
2811 		 * of the joined buddies with its buddy at new value + 1.
2812 		 * we continue to join until we find a buddy that cannot be
2813 		 * joined (does not have a value equal to the size of the
2814 		 * last join) or until all leaves have been joined into a
2815 		 * single system.
2816 		 *
2817 		 * get the buddy size (number of words covered) of
2818 		 * the new value.
2819 		 */
2820 		budsz = BUDSIZE(newval, tp->dmt_budmin);
2821 
2822 		/* try to join.
2823 		 */
2824 		while (budsz < le32_to_cpu(tp->dmt_nleafs)) {
2825 			/* get the buddy leaf.
2826 			 */
2827 			buddy = leafno ^ budsz;
2828 
2829 			/* if the leaf's new value is greater than its
2830 			 * buddy's value, we join no more.
2831 			 */
2832 			if (newval > leaf[buddy])
2833 				break;
2834 
2835 			/* It shouldn't be less */
2836 			if (newval < leaf[buddy])
2837 				return -EIO;
2838 
2839 			/* check which (leafno or buddy) is the left buddy.
2840 			 * the left buddy gets to claim the blocks resulting
2841 			 * from the join while the right gets to claim none.
2842 			 * the left buddy is also eligible to participate in
2843 			 * a join at the next higher level while the right
2844 			 * is not.
2845 			 *
2846 			 */
2847 			if (leafno < buddy) {
2848 				/* leafno is the left buddy.
2849 				 */
2850 				dbAdjTree(tp, buddy, NOFREE);
2851 			} else {
2852 				/* buddy is the left buddy and becomes
2853 				 * leafno.
2854 				 */
2855 				dbAdjTree(tp, leafno, NOFREE);
2856 				leafno = buddy;
2857 			}
2858 
2859 			/* on to try the next join.
2860 			 */
2861 			newval += 1;
2862 			budsz <<= 1;
2863 		}
2864 	}
2865 
2866 	/* update the leaf value.
2867 	 */
2868 	dbAdjTree(tp, leafno, newval);
2869 
2870 	return 0;
2871 }
2872 
2873 
2874 /*
2875  * NAME:	dbAdjTree()
2876  *
2877  * FUNCTION:	update a leaf of a dmtree with a new value, adjusting
2878  *		the dmtree, as required, to reflect the new leaf value.
2879  *		the combination of any buddies must already be done before
2880  *		this is called.
2881  *
2882  * PARAMETERS:
2883  *	tp	- pointer to the tree to be adjusted.
2884  *	leafno	- the number of the leaf to be updated.
2885  *	newval	- the new value for the leaf.
2886  *
2887  * RETURN VALUES: none
2888  */
2889 static void dbAdjTree(dmtree_t * tp, int leafno, int newval)
2890 {
2891 	int lp, pp, k;
2892 	int max;
2893 
2894 	/* pick up the index of the leaf for this leafno.
2895 	 */
2896 	lp = leafno + le32_to_cpu(tp->dmt_leafidx);
2897 
2898 	/* is the current value the same as the old value ?  if so,
2899 	 * there is nothing to do.
2900 	 */
2901 	if (tp->dmt_stree[lp] == newval)
2902 		return;
2903 
2904 	/* set the new value.
2905 	 */
2906 	tp->dmt_stree[lp] = newval;
2907 
2908 	/* bubble the new value up the tree as required.
2909 	 */
2910 	for (k = 0; k < le32_to_cpu(tp->dmt_height); k++) {
2911 		/* get the index of the first leaf of the 4 leaf
2912 		 * group containing the specified leaf (leafno).
2913 		 */
2914 		lp = ((lp - 1) & ~0x03) + 1;
2915 
2916 		/* get the index of the parent of this 4 leaf group.
2917 		 */
2918 		pp = (lp - 1) >> 2;
2919 
2920 		/* determine the maximum of the 4 leaves.
2921 		 */
2922 		max = TREEMAX(&tp->dmt_stree[lp]);
2923 
2924 		/* if the maximum of the 4 is the same as the
2925 		 * parent's value, we're done.
2926 		 */
2927 		if (tp->dmt_stree[pp] == max)
2928 			break;
2929 
2930 		/* parent gets new value.
2931 		 */
2932 		tp->dmt_stree[pp] = max;
2933 
2934 		/* parent becomes leaf for next go-round.
2935 		 */
2936 		lp = pp;
2937 	}
2938 }
2939 
2940 
2941 /*
2942  * NAME:	dbFindLeaf()
2943  *
2944  * FUNCTION:	search a dmtree_t for sufficient free blocks, returning
2945  *		the index of a leaf describing the free blocks if
2946  *		sufficient free blocks are found.
2947  *
2948  *		the search starts at the top of the dmtree_t tree and
2949  *		proceeds down the tree to the leftmost leaf with sufficient
2950  *		free space.
2951  *
2952  * PARAMETERS:
2953  *	tp	- pointer to the tree to be searched.
2954  *	l2nb	- log2 number of free blocks to search for.
2955  *	leafidx	- return pointer to be set to the index of the leaf
2956  *		  describing at least l2nb free blocks if sufficient
2957  *		  free blocks are found.
2958  *
2959  * RETURN VALUES:
2960  *	0	- success
2961  *	-ENOSPC	- insufficient free blocks.
2962  */
2963 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx)
2964 {
2965 	int ti, n = 0, k, x = 0;
2966 
2967 	/* first check the root of the tree to see if there is
2968 	 * sufficient free space.
2969 	 */
2970 	if (l2nb > tp->dmt_stree[ROOT])
2971 		return -ENOSPC;
2972 
2973 	/* sufficient free space available. now search down the tree
2974 	 * starting at the next level for the leftmost leaf that
2975 	 * describes sufficient free space.
2976 	 */
2977 	for (k = le32_to_cpu(tp->dmt_height), ti = 1;
2978 	     k > 0; k--, ti = ((ti + n) << 2) + 1) {
2979 		/* search the four nodes at this level, starting from
2980 		 * the left.
2981 		 */
2982 		for (x = ti, n = 0; n < 4; n++) {
2983 			/* sufficient free space found.  move to the next
2984 			 * level (or quit if this is the last level).
2985 			 */
2986 			if (l2nb <= tp->dmt_stree[x + n])
2987 				break;
2988 		}
2989 
2990 		/* better have found something since the higher
2991 		 * levels of the tree said it was here.
2992 		 */
2993 		assert(n < 4);
2994 	}
2995 
2996 	/* set the return to the leftmost leaf describing sufficient
2997 	 * free space.
2998 	 */
2999 	*leafidx = x + n - le32_to_cpu(tp->dmt_leafidx);
3000 
3001 	return (0);
3002 }
3003 
3004 
3005 /*
3006  * NAME:	dbFindBits()
3007  *
3008  * FUNCTION:	find a specified number of binary buddy free bits within a
3009  *		dmap bitmap word value.
3010  *
3011  *		this routine searches the bitmap value for (1 << l2nb) free
3012  *		bits at (1 << l2nb) alignments within the value.
3013  *
3014  * PARAMETERS:
3015  *	word	-  dmap bitmap word value.
3016  *	l2nb	-  number of free bits specified as a log2 number.
3017  *
3018  * RETURN VALUES:
3019  *	starting bit number of free bits.
3020  */
3021 static int dbFindBits(u32 word, int l2nb)
3022 {
3023 	int bitno, nb;
3024 	u32 mask;
3025 
3026 	/* get the number of bits.
3027 	 */
3028 	nb = 1 << l2nb;
3029 	assert(nb <= DBWORD);
3030 
3031 	/* complement the word so we can use a mask (i.e. 0s represent
3032 	 * free bits) and compute the mask.
3033 	 */
3034 	word = ~word;
3035 	mask = ONES << (DBWORD - nb);
3036 
3037 	/* scan the word for nb free bits at nb alignments.
3038 	 */
3039 	for (bitno = 0; mask != 0; bitno += nb, mask >>= nb) {
3040 		if ((mask & word) == mask)
3041 			break;
3042 	}
3043 
3044 	ASSERT(bitno < 32);
3045 
3046 	/* return the bit number.
3047 	 */
3048 	return (bitno);
3049 }
3050 
3051 
3052 /*
3053  * NAME:	dbMaxBud(u8 *cp)
3054  *
3055  * FUNCTION:	determine the largest binary buddy string of free
3056  *		bits within 32-bits of the map.
3057  *
3058  * PARAMETERS:
3059  *	cp	-  pointer to the 32-bit value.
3060  *
3061  * RETURN VALUES:
3062  *	largest binary buddy of free bits within a dmap word.
3063  */
3064 static int dbMaxBud(u8 * cp)
3065 {
3066 	signed char tmp1, tmp2;
3067 
3068 	/* check if the wmap word is all free. if so, the
3069 	 * free buddy size is BUDMIN.
3070 	 */
3071 	if (*((uint *) cp) == 0)
3072 		return (BUDMIN);
3073 
3074 	/* check if the wmap word is half free. if so, the
3075 	 * free buddy size is BUDMIN-1.
3076 	 */
3077 	if (*((u16 *) cp) == 0 || *((u16 *) cp + 1) == 0)
3078 		return (BUDMIN - 1);
3079 
3080 	/* not all free or half free. determine the free buddy
3081 	 * size thru table lookup using quarters of the wmap word.
3082 	 */
3083 	tmp1 = max(budtab[cp[2]], budtab[cp[3]]);
3084 	tmp2 = max(budtab[cp[0]], budtab[cp[1]]);
3085 	return (max(tmp1, tmp2));
3086 }
3087 
3088 
3089 /*
3090  * NAME:	cnttz(uint word)
3091  *
3092  * FUNCTION:	determine the number of trailing zeros within a 32-bit
3093  *		value.
3094  *
3095  * PARAMETERS:
3096  *	value	-  32-bit value to be examined.
3097  *
3098  * RETURN VALUES:
3099  *	count of trailing zeros
3100  */
3101 static int cnttz(u32 word)
3102 {
3103 	int n;
3104 
3105 	for (n = 0; n < 32; n++, word >>= 1) {
3106 		if (word & 0x01)
3107 			break;
3108 	}
3109 
3110 	return (n);
3111 }
3112 
3113 
3114 /*
3115  * NAME:	cntlz(u32 value)
3116  *
3117  * FUNCTION:	determine the number of leading zeros within a 32-bit
3118  *		value.
3119  *
3120  * PARAMETERS:
3121  *	value	-  32-bit value to be examined.
3122  *
3123  * RETURN VALUES:
3124  *	count of leading zeros
3125  */
3126 static int cntlz(u32 value)
3127 {
3128 	int n;
3129 
3130 	for (n = 0; n < 32; n++, value <<= 1) {
3131 		if (value & HIGHORDER)
3132 			break;
3133 	}
3134 	return (n);
3135 }
3136 
3137 
3138 /*
3139  * NAME:	blkstol2(s64 nb)
3140  *
3141  * FUNCTION:	convert a block count to its log2 value. if the block
3142  *		count is not a l2 multiple, it is rounded up to the next
3143  *		larger l2 multiple.
3144  *
3145  * PARAMETERS:
3146  *	nb	-  number of blocks
3147  *
3148  * RETURN VALUES:
3149  *	log2 number of blocks
3150  */
3151 static int blkstol2(s64 nb)
3152 {
3153 	int l2nb;
3154 	s64 mask;		/* meant to be signed */
3155 
3156 	mask = (s64) 1 << (64 - 1);
3157 
3158 	/* count the leading bits.
3159 	 */
3160 	for (l2nb = 0; l2nb < 64; l2nb++, mask >>= 1) {
3161 		/* leading bit found.
3162 		 */
3163 		if (nb & mask) {
3164 			/* determine the l2 value.
3165 			 */
3166 			l2nb = (64 - 1) - l2nb;
3167 
3168 			/* check if we need to round up.
3169 			 */
3170 			if (~mask & nb)
3171 				l2nb++;
3172 
3173 			return (l2nb);
3174 		}
3175 	}
3176 	assert(0);
3177 	return 0;		/* fix compiler warning */
3178 }
3179 
3180 
3181 /*
3182  * NAME:	dbAllocBottomUp()
3183  *
3184  * FUNCTION:	alloc the specified block range from the working block
3185  *		allocation map.
3186  *
3187  *		the blocks will be alloc from the working map one dmap
3188  *		at a time.
3189  *
3190  * PARAMETERS:
3191  *	ip	-  pointer to in-core inode;
3192  *	blkno	-  starting block number to be freed.
3193  *	nblocks	-  number of blocks to be freed.
3194  *
3195  * RETURN VALUES:
3196  *	0	- success
3197  *	-EIO	- i/o error
3198  */
3199 int dbAllocBottomUp(struct inode *ip, s64 blkno, s64 nblocks)
3200 {
3201 	struct metapage *mp;
3202 	struct dmap *dp;
3203 	int nb, rc;
3204 	s64 lblkno, rem;
3205 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
3206 	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
3207 
3208 	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
3209 
3210 	/* block to be allocated better be within the mapsize. */
3211 	ASSERT(nblocks <= bmp->db_mapsize - blkno);
3212 
3213 	/*
3214 	 * allocate the blocks a dmap at a time.
3215 	 */
3216 	mp = NULL;
3217 	for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
3218 		/* release previous dmap if any */
3219 		if (mp) {
3220 			write_metapage(mp);
3221 		}
3222 
3223 		/* get the buffer for the current dmap. */
3224 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
3225 		mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
3226 		if (mp == NULL) {
3227 			IREAD_UNLOCK(ipbmap);
3228 			return -EIO;
3229 		}
3230 		dp = (struct dmap *) mp->data;
3231 
3232 		/* determine the number of blocks to be allocated from
3233 		 * this dmap.
3234 		 */
3235 		nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
3236 
3237 		/* allocate the blocks. */
3238 		if ((rc = dbAllocDmapBU(bmp, dp, blkno, nb))) {
3239 			release_metapage(mp);
3240 			IREAD_UNLOCK(ipbmap);
3241 			return (rc);
3242 		}
3243 	}
3244 
3245 	/* write the last buffer. */
3246 	write_metapage(mp);
3247 
3248 	IREAD_UNLOCK(ipbmap);
3249 
3250 	return (0);
3251 }
3252 
3253 
3254 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
3255 			 int nblocks)
3256 {
3257 	int rc;
3258 	int dbitno, word, rembits, nb, nwords, wbitno, agno;
3259 	s8 oldroot;
3260 	struct dmaptree *tp = (struct dmaptree *) & dp->tree;
3261 
3262 	/* save the current value of the root (i.e. maximum free string)
3263 	 * of the dmap tree.
3264 	 */
3265 	oldroot = tp->stree[ROOT];
3266 
3267 	/* determine the bit number and word within the dmap of the
3268 	 * starting block.
3269 	 */
3270 	dbitno = blkno & (BPERDMAP - 1);
3271 	word = dbitno >> L2DBWORD;
3272 
3273 	/* block range better be within the dmap */
3274 	assert(dbitno + nblocks <= BPERDMAP);
3275 
3276 	/* allocate the bits of the dmap's words corresponding to the block
3277 	 * range. not all bits of the first and last words may be contained
3278 	 * within the block range.  if this is the case, we'll work against
3279 	 * those words (i.e. partial first and/or last) on an individual basis
3280 	 * (a single pass), allocating the bits of interest by hand and
3281 	 * updating the leaf corresponding to the dmap word. a single pass
3282 	 * will be used for all dmap words fully contained within the
3283 	 * specified range.  within this pass, the bits of all fully contained
3284 	 * dmap words will be marked as free in a single shot and the leaves
3285 	 * will be updated. a single leaf may describe the free space of
3286 	 * multiple dmap words, so we may update only a subset of the actual
3287 	 * leaves corresponding to the dmap words of the block range.
3288 	 */
3289 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
3290 		/* determine the bit number within the word and
3291 		 * the number of bits within the word.
3292 		 */
3293 		wbitno = dbitno & (DBWORD - 1);
3294 		nb = min(rembits, DBWORD - wbitno);
3295 
3296 		/* check if only part of a word is to be allocated.
3297 		 */
3298 		if (nb < DBWORD) {
3299 			/* allocate (set to 1) the appropriate bits within
3300 			 * this dmap word.
3301 			 */
3302 			dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
3303 						      >> wbitno);
3304 
3305 			word++;
3306 		} else {
3307 			/* one or more dmap words are fully contained
3308 			 * within the block range.  determine how many
3309 			 * words and allocate (set to 1) the bits of these
3310 			 * words.
3311 			 */
3312 			nwords = rembits >> L2DBWORD;
3313 			memset(&dp->wmap[word], (int) ONES, nwords * 4);
3314 
3315 			/* determine how many bits */
3316 			nb = nwords << L2DBWORD;
3317 			word += nwords;
3318 		}
3319 	}
3320 
3321 	/* update the free count for this dmap */
3322 	le32_add_cpu(&dp->nfree, -nblocks);
3323 
3324 	/* reconstruct summary tree */
3325 	dbInitDmapTree(dp);
3326 
3327 	BMAP_LOCK(bmp);
3328 
3329 	/* if this allocation group is completely free,
3330 	 * update the highest active allocation group number
3331 	 * if this allocation group is the new max.
3332 	 */
3333 	agno = blkno >> bmp->db_agl2size;
3334 	if (agno > bmp->db_maxag)
3335 		bmp->db_maxag = agno;
3336 
3337 	/* update the free count for the allocation group and map */
3338 	bmp->db_agfree[agno] -= nblocks;
3339 	bmp->db_nfree -= nblocks;
3340 
3341 	BMAP_UNLOCK(bmp);
3342 
3343 	/* if the root has not changed, done. */
3344 	if (tp->stree[ROOT] == oldroot)
3345 		return (0);
3346 
3347 	/* root changed. bubble the change up to the dmap control pages.
3348 	 * if the adjustment of the upper level control pages fails,
3349 	 * backout the bit allocation (thus making everything consistent).
3350 	 */
3351 	if ((rc = dbAdjCtl(bmp, blkno, tp->stree[ROOT], 1, 0)))
3352 		dbFreeBits(bmp, dp, blkno, nblocks);
3353 
3354 	return (rc);
3355 }
3356 
3357 
3358 /*
3359  * NAME:	dbExtendFS()
3360  *
3361  * FUNCTION:	extend bmap from blkno for nblocks;
3362  *		dbExtendFS() updates bmap ready for dbAllocBottomUp();
3363  *
3364  * L2
3365  *  |
3366  *   L1---------------------------------L1
3367  *    |					 |
3368  *     L0---------L0---------L0		  L0---------L0---------L0
3369  *      |	   |	      |		   |	      |		 |
3370  *	 d0,...,dn  d0,...,dn  d0,...,dn    d0,...,dn  d0,...,dn  d0,.,dm;
3371  * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm
3372  *
3373  * <---old---><----------------------------extend----------------------->
3374  */
3375 int dbExtendFS(struct inode *ipbmap, s64 blkno,	s64 nblocks)
3376 {
3377 	struct jfs_sb_info *sbi = JFS_SBI(ipbmap->i_sb);
3378 	int nbperpage = sbi->nbperpage;
3379 	int i, i0 = true, j, j0 = true, k, n;
3380 	s64 newsize;
3381 	s64 p;
3382 	struct metapage *mp, *l2mp, *l1mp = NULL, *l0mp = NULL;
3383 	struct dmapctl *l2dcp, *l1dcp, *l0dcp;
3384 	struct dmap *dp;
3385 	s8 *l0leaf, *l1leaf, *l2leaf;
3386 	struct bmap *bmp = sbi->bmap;
3387 	int agno, l2agsize, oldl2agsize;
3388 	s64 ag_rem;
3389 
3390 	newsize = blkno + nblocks;
3391 
3392 	jfs_info("dbExtendFS: blkno:%Ld nblocks:%Ld newsize:%Ld",
3393 		 (long long) blkno, (long long) nblocks, (long long) newsize);
3394 
3395 	/*
3396 	 *	initialize bmap control page.
3397 	 *
3398 	 * all the data in bmap control page should exclude
3399 	 * the mkfs hidden dmap page.
3400 	 */
3401 
3402 	/* update mapsize */
3403 	bmp->db_mapsize = newsize;
3404 	bmp->db_maxlevel = BMAPSZTOLEV(bmp->db_mapsize);
3405 
3406 	/* compute new AG size */
3407 	l2agsize = dbGetL2AGSize(newsize);
3408 	oldl2agsize = bmp->db_agl2size;
3409 
3410 	bmp->db_agl2size = l2agsize;
3411 	bmp->db_agsize = 1 << l2agsize;
3412 
3413 	/* compute new number of AG */
3414 	agno = bmp->db_numag;
3415 	bmp->db_numag = newsize >> l2agsize;
3416 	bmp->db_numag += ((u32) newsize % (u32) bmp->db_agsize) ? 1 : 0;
3417 
3418 	/*
3419 	 *	reconfigure db_agfree[]
3420 	 * from old AG configuration to new AG configuration;
3421 	 *
3422 	 * coalesce contiguous k (newAGSize/oldAGSize) AGs;
3423 	 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn;
3424 	 * note: new AG size = old AG size * (2**x).
3425 	 */
3426 	if (l2agsize == oldl2agsize)
3427 		goto extend;
3428 	k = 1 << (l2agsize - oldl2agsize);
3429 	ag_rem = bmp->db_agfree[0];	/* save agfree[0] */
3430 	for (i = 0, n = 0; i < agno; n++) {
3431 		bmp->db_agfree[n] = 0;	/* init collection point */
3432 
3433 		/* coalesce contiguous k AGs; */
3434 		for (j = 0; j < k && i < agno; j++, i++) {
3435 			/* merge AGi to AGn */
3436 			bmp->db_agfree[n] += bmp->db_agfree[i];
3437 		}
3438 	}
3439 	bmp->db_agfree[0] += ag_rem;	/* restore agfree[0] */
3440 
3441 	for (; n < MAXAG; n++)
3442 		bmp->db_agfree[n] = 0;
3443 
3444 	/*
3445 	 * update highest active ag number
3446 	 */
3447 
3448 	bmp->db_maxag = bmp->db_maxag / k;
3449 
3450 	/*
3451 	 *	extend bmap
3452 	 *
3453 	 * update bit maps and corresponding level control pages;
3454 	 * global control page db_nfree, db_agfree[agno], db_maxfreebud;
3455 	 */
3456       extend:
3457 	/* get L2 page */
3458 	p = BMAPBLKNO + nbperpage;	/* L2 page */
3459 	l2mp = read_metapage(ipbmap, p, PSIZE, 0);
3460 	if (!l2mp) {
3461 		jfs_error(ipbmap->i_sb, "L2 page could not be read\n");
3462 		return -EIO;
3463 	}
3464 	l2dcp = (struct dmapctl *) l2mp->data;
3465 
3466 	/* compute start L1 */
3467 	k = blkno >> L2MAXL1SIZE;
3468 	l2leaf = l2dcp->stree + CTLLEAFIND + k;
3469 	p = BLKTOL1(blkno, sbi->l2nbperpage);	/* L1 page */
3470 
3471 	/*
3472 	 * extend each L1 in L2
3473 	 */
3474 	for (; k < LPERCTL; k++, p += nbperpage) {
3475 		/* get L1 page */
3476 		if (j0) {
3477 			/* read in L1 page: (blkno & (MAXL1SIZE - 1)) */
3478 			l1mp = read_metapage(ipbmap, p, PSIZE, 0);
3479 			if (l1mp == NULL)
3480 				goto errout;
3481 			l1dcp = (struct dmapctl *) l1mp->data;
3482 
3483 			/* compute start L0 */
3484 			j = (blkno & (MAXL1SIZE - 1)) >> L2MAXL0SIZE;
3485 			l1leaf = l1dcp->stree + CTLLEAFIND + j;
3486 			p = BLKTOL0(blkno, sbi->l2nbperpage);
3487 			j0 = false;
3488 		} else {
3489 			/* assign/init L1 page */
3490 			l1mp = get_metapage(ipbmap, p, PSIZE, 0);
3491 			if (l1mp == NULL)
3492 				goto errout;
3493 
3494 			l1dcp = (struct dmapctl *) l1mp->data;
3495 
3496 			/* compute start L0 */
3497 			j = 0;
3498 			l1leaf = l1dcp->stree + CTLLEAFIND;
3499 			p += nbperpage;	/* 1st L0 of L1.k */
3500 		}
3501 
3502 		/*
3503 		 * extend each L0 in L1
3504 		 */
3505 		for (; j < LPERCTL; j++) {
3506 			/* get L0 page */
3507 			if (i0) {
3508 				/* read in L0 page: (blkno & (MAXL0SIZE - 1)) */
3509 
3510 				l0mp = read_metapage(ipbmap, p, PSIZE, 0);
3511 				if (l0mp == NULL)
3512 					goto errout;
3513 				l0dcp = (struct dmapctl *) l0mp->data;
3514 
3515 				/* compute start dmap */
3516 				i = (blkno & (MAXL0SIZE - 1)) >>
3517 				    L2BPERDMAP;
3518 				l0leaf = l0dcp->stree + CTLLEAFIND + i;
3519 				p = BLKTODMAP(blkno,
3520 					      sbi->l2nbperpage);
3521 				i0 = false;
3522 			} else {
3523 				/* assign/init L0 page */
3524 				l0mp = get_metapage(ipbmap, p, PSIZE, 0);
3525 				if (l0mp == NULL)
3526 					goto errout;
3527 
3528 				l0dcp = (struct dmapctl *) l0mp->data;
3529 
3530 				/* compute start dmap */
3531 				i = 0;
3532 				l0leaf = l0dcp->stree + CTLLEAFIND;
3533 				p += nbperpage;	/* 1st dmap of L0.j */
3534 			}
3535 
3536 			/*
3537 			 * extend each dmap in L0
3538 			 */
3539 			for (; i < LPERCTL; i++) {
3540 				/*
3541 				 * reconstruct the dmap page, and
3542 				 * initialize corresponding parent L0 leaf
3543 				 */
3544 				if ((n = blkno & (BPERDMAP - 1))) {
3545 					/* read in dmap page: */
3546 					mp = read_metapage(ipbmap, p,
3547 							   PSIZE, 0);
3548 					if (mp == NULL)
3549 						goto errout;
3550 					n = min(nblocks, (s64)BPERDMAP - n);
3551 				} else {
3552 					/* assign/init dmap page */
3553 					mp = read_metapage(ipbmap, p,
3554 							   PSIZE, 0);
3555 					if (mp == NULL)
3556 						goto errout;
3557 
3558 					n = min_t(s64, nblocks, BPERDMAP);
3559 				}
3560 
3561 				dp = (struct dmap *) mp->data;
3562 				*l0leaf = dbInitDmap(dp, blkno, n);
3563 
3564 				bmp->db_nfree += n;
3565 				agno = le64_to_cpu(dp->start) >> l2agsize;
3566 				bmp->db_agfree[agno] += n;
3567 
3568 				write_metapage(mp);
3569 
3570 				l0leaf++;
3571 				p += nbperpage;
3572 
3573 				blkno += n;
3574 				nblocks -= n;
3575 				if (nblocks == 0)
3576 					break;
3577 			}	/* for each dmap in a L0 */
3578 
3579 			/*
3580 			 * build current L0 page from its leaves, and
3581 			 * initialize corresponding parent L1 leaf
3582 			 */
3583 			*l1leaf = dbInitDmapCtl(l0dcp, 0, ++i);
3584 			write_metapage(l0mp);
3585 			l0mp = NULL;
3586 
3587 			if (nblocks)
3588 				l1leaf++;	/* continue for next L0 */
3589 			else {
3590 				/* more than 1 L0 ? */
3591 				if (j > 0)
3592 					break;	/* build L1 page */
3593 				else {
3594 					/* summarize in global bmap page */
3595 					bmp->db_maxfreebud = *l1leaf;
3596 					release_metapage(l1mp);
3597 					release_metapage(l2mp);
3598 					goto finalize;
3599 				}
3600 			}
3601 		}		/* for each L0 in a L1 */
3602 
3603 		/*
3604 		 * build current L1 page from its leaves, and
3605 		 * initialize corresponding parent L2 leaf
3606 		 */
3607 		*l2leaf = dbInitDmapCtl(l1dcp, 1, ++j);
3608 		write_metapage(l1mp);
3609 		l1mp = NULL;
3610 
3611 		if (nblocks)
3612 			l2leaf++;	/* continue for next L1 */
3613 		else {
3614 			/* more than 1 L1 ? */
3615 			if (k > 0)
3616 				break;	/* build L2 page */
3617 			else {
3618 				/* summarize in global bmap page */
3619 				bmp->db_maxfreebud = *l2leaf;
3620 				release_metapage(l2mp);
3621 				goto finalize;
3622 			}
3623 		}
3624 	}			/* for each L1 in a L2 */
3625 
3626 	jfs_error(ipbmap->i_sb, "function has not returned as expected\n");
3627 errout:
3628 	if (l0mp)
3629 		release_metapage(l0mp);
3630 	if (l1mp)
3631 		release_metapage(l1mp);
3632 	release_metapage(l2mp);
3633 	return -EIO;
3634 
3635 	/*
3636 	 *	finalize bmap control page
3637 	 */
3638 finalize:
3639 
3640 	return 0;
3641 }
3642 
3643 
3644 /*
3645  *	dbFinalizeBmap()
3646  */
3647 void dbFinalizeBmap(struct inode *ipbmap)
3648 {
3649 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
3650 	int actags, inactags, l2nl;
3651 	s64 ag_rem, actfree, inactfree, avgfree;
3652 	int i, n;
3653 
3654 	/*
3655 	 *	finalize bmap control page
3656 	 */
3657 //finalize:
3658 	/*
3659 	 * compute db_agpref: preferred ag to allocate from
3660 	 * (the leftmost ag with average free space in it);
3661 	 */
3662 //agpref:
3663 	/* get the number of active ags and inactive ags */
3664 	actags = bmp->db_maxag + 1;
3665 	inactags = bmp->db_numag - actags;
3666 	ag_rem = bmp->db_mapsize & (bmp->db_agsize - 1);	/* ??? */
3667 
3668 	/* determine how many blocks are in the inactive allocation
3669 	 * groups. in doing this, we must account for the fact that
3670 	 * the rightmost group might be a partial group (i.e. file
3671 	 * system size is not a multiple of the group size).
3672 	 */
3673 	inactfree = (inactags && ag_rem) ?
3674 	    ((inactags - 1) << bmp->db_agl2size) + ag_rem
3675 	    : inactags << bmp->db_agl2size;
3676 
3677 	/* determine how many free blocks are in the active
3678 	 * allocation groups plus the average number of free blocks
3679 	 * within the active ags.
3680 	 */
3681 	actfree = bmp->db_nfree - inactfree;
3682 	avgfree = (u32) actfree / (u32) actags;
3683 
3684 	/* if the preferred allocation group has not average free space.
3685 	 * re-establish the preferred group as the leftmost
3686 	 * group with average free space.
3687 	 */
3688 	if (bmp->db_agfree[bmp->db_agpref] < avgfree) {
3689 		for (bmp->db_agpref = 0; bmp->db_agpref < actags;
3690 		     bmp->db_agpref++) {
3691 			if (bmp->db_agfree[bmp->db_agpref] >= avgfree)
3692 				break;
3693 		}
3694 		if (bmp->db_agpref >= bmp->db_numag) {
3695 			jfs_error(ipbmap->i_sb,
3696 				  "cannot find ag with average freespace\n");
3697 		}
3698 	}
3699 
3700 	/*
3701 	 * compute db_aglevel, db_agheight, db_width, db_agstart:
3702 	 * an ag is covered in aglevel dmapctl summary tree,
3703 	 * at agheight level height (from leaf) with agwidth number of nodes
3704 	 * each, which starts at agstart index node of the smmary tree node
3705 	 * array;
3706 	 */
3707 	bmp->db_aglevel = BMAPSZTOLEV(bmp->db_agsize);
3708 	l2nl =
3709 	    bmp->db_agl2size - (L2BPERDMAP + bmp->db_aglevel * L2LPERCTL);
3710 	bmp->db_agheight = l2nl >> 1;
3711 	bmp->db_agwidth = 1 << (l2nl - (bmp->db_agheight << 1));
3712 	for (i = 5 - bmp->db_agheight, bmp->db_agstart = 0, n = 1; i > 0;
3713 	     i--) {
3714 		bmp->db_agstart += n;
3715 		n <<= 2;
3716 	}
3717 
3718 }
3719 
3720 
3721 /*
3722  * NAME:	dbInitDmap()/ujfs_idmap_page()
3723  *
3724  * FUNCTION:	initialize working/persistent bitmap of the dmap page
3725  *		for the specified number of blocks:
3726  *
3727  *		at entry, the bitmaps had been initialized as free (ZEROS);
3728  *		The number of blocks will only account for the actually
3729  *		existing blocks. Blocks which don't actually exist in
3730  *		the aggregate will be marked as allocated (ONES);
3731  *
3732  * PARAMETERS:
3733  *	dp	- pointer to page of map
3734  *	nblocks	- number of blocks this page
3735  *
3736  * RETURNS: NONE
3737  */
3738 static int dbInitDmap(struct dmap * dp, s64 Blkno, int nblocks)
3739 {
3740 	int blkno, w, b, r, nw, nb, i;
3741 
3742 	/* starting block number within the dmap */
3743 	blkno = Blkno & (BPERDMAP - 1);
3744 
3745 	if (blkno == 0) {
3746 		dp->nblocks = dp->nfree = cpu_to_le32(nblocks);
3747 		dp->start = cpu_to_le64(Blkno);
3748 
3749 		if (nblocks == BPERDMAP) {
3750 			memset(&dp->wmap[0], 0, LPERDMAP * 4);
3751 			memset(&dp->pmap[0], 0, LPERDMAP * 4);
3752 			goto initTree;
3753 		}
3754 	} else {
3755 		le32_add_cpu(&dp->nblocks, nblocks);
3756 		le32_add_cpu(&dp->nfree, nblocks);
3757 	}
3758 
3759 	/* word number containing start block number */
3760 	w = blkno >> L2DBWORD;
3761 
3762 	/*
3763 	 * free the bits corresponding to the block range (ZEROS):
3764 	 * note: not all bits of the first and last words may be contained
3765 	 * within the block range.
3766 	 */
3767 	for (r = nblocks; r > 0; r -= nb, blkno += nb) {
3768 		/* number of bits preceding range to be freed in the word */
3769 		b = blkno & (DBWORD - 1);
3770 		/* number of bits to free in the word */
3771 		nb = min(r, DBWORD - b);
3772 
3773 		/* is partial word to be freed ? */
3774 		if (nb < DBWORD) {
3775 			/* free (set to 0) from the bitmap word */
3776 			dp->wmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3777 						     >> b));
3778 			dp->pmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3779 						     >> b));
3780 
3781 			/* skip the word freed */
3782 			w++;
3783 		} else {
3784 			/* free (set to 0) contiguous bitmap words */
3785 			nw = r >> L2DBWORD;
3786 			memset(&dp->wmap[w], 0, nw * 4);
3787 			memset(&dp->pmap[w], 0, nw * 4);
3788 
3789 			/* skip the words freed */
3790 			nb = nw << L2DBWORD;
3791 			w += nw;
3792 		}
3793 	}
3794 
3795 	/*
3796 	 * mark bits following the range to be freed (non-existing
3797 	 * blocks) as allocated (ONES)
3798 	 */
3799 
3800 	if (blkno == BPERDMAP)
3801 		goto initTree;
3802 
3803 	/* the first word beyond the end of existing blocks */
3804 	w = blkno >> L2DBWORD;
3805 
3806 	/* does nblocks fall on a 32-bit boundary ? */
3807 	b = blkno & (DBWORD - 1);
3808 	if (b) {
3809 		/* mark a partial word allocated */
3810 		dp->wmap[w] = dp->pmap[w] = cpu_to_le32(ONES >> b);
3811 		w++;
3812 	}
3813 
3814 	/* set the rest of the words in the page to allocated (ONES) */
3815 	for (i = w; i < LPERDMAP; i++)
3816 		dp->pmap[i] = dp->wmap[i] = cpu_to_le32(ONES);
3817 
3818 	/*
3819 	 * init tree
3820 	 */
3821       initTree:
3822 	return (dbInitDmapTree(dp));
3823 }
3824 
3825 
3826 /*
3827  * NAME:	dbInitDmapTree()/ujfs_complete_dmap()
3828  *
3829  * FUNCTION:	initialize summary tree of the specified dmap:
3830  *
3831  *		at entry, bitmap of the dmap has been initialized;
3832  *
3833  * PARAMETERS:
3834  *	dp	- dmap to complete
3835  *	blkno	- starting block number for this dmap
3836  *	treemax	- will be filled in with max free for this dmap
3837  *
3838  * RETURNS:	max free string at the root of the tree
3839  */
3840 static int dbInitDmapTree(struct dmap * dp)
3841 {
3842 	struct dmaptree *tp;
3843 	s8 *cp;
3844 	int i;
3845 
3846 	/* init fixed info of tree */
3847 	tp = &dp->tree;
3848 	tp->nleafs = cpu_to_le32(LPERDMAP);
3849 	tp->l2nleafs = cpu_to_le32(L2LPERDMAP);
3850 	tp->leafidx = cpu_to_le32(LEAFIND);
3851 	tp->height = cpu_to_le32(4);
3852 	tp->budmin = BUDMIN;
3853 
3854 	/* init each leaf from corresponding wmap word:
3855 	 * note: leaf is set to NOFREE(-1) if all blocks of corresponding
3856 	 * bitmap word are allocated.
3857 	 */
3858 	cp = tp->stree + le32_to_cpu(tp->leafidx);
3859 	for (i = 0; i < LPERDMAP; i++)
3860 		*cp++ = dbMaxBud((u8 *) & dp->wmap[i]);
3861 
3862 	/* build the dmap's binary buddy summary tree */
3863 	return (dbInitTree(tp));
3864 }
3865 
3866 
3867 /*
3868  * NAME:	dbInitTree()/ujfs_adjtree()
3869  *
3870  * FUNCTION:	initialize binary buddy summary tree of a dmap or dmapctl.
3871  *
3872  *		at entry, the leaves of the tree has been initialized
3873  *		from corresponding bitmap word or root of summary tree
3874  *		of the child control page;
3875  *		configure binary buddy system at the leaf level, then
3876  *		bubble up the values of the leaf nodes up the tree.
3877  *
3878  * PARAMETERS:
3879  *	cp	- Pointer to the root of the tree
3880  *	l2leaves- Number of leaf nodes as a power of 2
3881  *	l2min	- Number of blocks that can be covered by a leaf
3882  *		  as a power of 2
3883  *
3884  * RETURNS: max free string at the root of the tree
3885  */
3886 static int dbInitTree(struct dmaptree * dtp)
3887 {
3888 	int l2max, l2free, bsize, nextb, i;
3889 	int child, parent, nparent;
3890 	s8 *tp, *cp, *cp1;
3891 
3892 	tp = dtp->stree;
3893 
3894 	/* Determine the maximum free string possible for the leaves */
3895 	l2max = le32_to_cpu(dtp->l2nleafs) + dtp->budmin;
3896 
3897 	/*
3898 	 * configure the leaf levevl into binary buddy system
3899 	 *
3900 	 * Try to combine buddies starting with a buddy size of 1
3901 	 * (i.e. two leaves). At a buddy size of 1 two buddy leaves
3902 	 * can be combined if both buddies have a maximum free of l2min;
3903 	 * the combination will result in the left-most buddy leaf having
3904 	 * a maximum free of l2min+1.
3905 	 * After processing all buddies for a given size, process buddies
3906 	 * at the next higher buddy size (i.e. current size * 2) and
3907 	 * the next maximum free (current free + 1).
3908 	 * This continues until the maximum possible buddy combination
3909 	 * yields maximum free.
3910 	 */
3911 	for (l2free = dtp->budmin, bsize = 1; l2free < l2max;
3912 	     l2free++, bsize = nextb) {
3913 		/* get next buddy size == current buddy pair size */
3914 		nextb = bsize << 1;
3915 
3916 		/* scan each adjacent buddy pair at current buddy size */
3917 		for (i = 0, cp = tp + le32_to_cpu(dtp->leafidx);
3918 		     i < le32_to_cpu(dtp->nleafs);
3919 		     i += nextb, cp += nextb) {
3920 			/* coalesce if both adjacent buddies are max free */
3921 			if (*cp == l2free && *(cp + bsize) == l2free) {
3922 				*cp = l2free + 1;	/* left take right */
3923 				*(cp + bsize) = -1;	/* right give left */
3924 			}
3925 		}
3926 	}
3927 
3928 	/*
3929 	 * bubble summary information of leaves up the tree.
3930 	 *
3931 	 * Starting at the leaf node level, the four nodes described by
3932 	 * the higher level parent node are compared for a maximum free and
3933 	 * this maximum becomes the value of the parent node.
3934 	 * when all lower level nodes are processed in this fashion then
3935 	 * move up to the next level (parent becomes a lower level node) and
3936 	 * continue the process for that level.
3937 	 */
3938 	for (child = le32_to_cpu(dtp->leafidx),
3939 	     nparent = le32_to_cpu(dtp->nleafs) >> 2;
3940 	     nparent > 0; nparent >>= 2, child = parent) {
3941 		/* get index of 1st node of parent level */
3942 		parent = (child - 1) >> 2;
3943 
3944 		/* set the value of the parent node as the maximum
3945 		 * of the four nodes of the current level.
3946 		 */
3947 		for (i = 0, cp = tp + child, cp1 = tp + parent;
3948 		     i < nparent; i++, cp += 4, cp1++)
3949 			*cp1 = TREEMAX(cp);
3950 	}
3951 
3952 	return (*tp);
3953 }
3954 
3955 
3956 /*
3957  *	dbInitDmapCtl()
3958  *
3959  * function: initialize dmapctl page
3960  */
3961 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i)
3962 {				/* start leaf index not covered by range */
3963 	s8 *cp;
3964 
3965 	dcp->nleafs = cpu_to_le32(LPERCTL);
3966 	dcp->l2nleafs = cpu_to_le32(L2LPERCTL);
3967 	dcp->leafidx = cpu_to_le32(CTLLEAFIND);
3968 	dcp->height = cpu_to_le32(5);
3969 	dcp->budmin = L2BPERDMAP + L2LPERCTL * level;
3970 
3971 	/*
3972 	 * initialize the leaves of current level that were not covered
3973 	 * by the specified input block range (i.e. the leaves have no
3974 	 * low level dmapctl or dmap).
3975 	 */
3976 	cp = &dcp->stree[CTLLEAFIND + i];
3977 	for (; i < LPERCTL; i++)
3978 		*cp++ = NOFREE;
3979 
3980 	/* build the dmap's binary buddy summary tree */
3981 	return (dbInitTree((struct dmaptree *) dcp));
3982 }
3983 
3984 
3985 /*
3986  * NAME:	dbGetL2AGSize()/ujfs_getagl2size()
3987  *
3988  * FUNCTION:	Determine log2(allocation group size) from aggregate size
3989  *
3990  * PARAMETERS:
3991  *	nblocks	- Number of blocks in aggregate
3992  *
3993  * RETURNS: log2(allocation group size) in aggregate blocks
3994  */
3995 static int dbGetL2AGSize(s64 nblocks)
3996 {
3997 	s64 sz;
3998 	s64 m;
3999 	int l2sz;
4000 
4001 	if (nblocks < BPERDMAP * MAXAG)
4002 		return (L2BPERDMAP);
4003 
4004 	/* round up aggregate size to power of 2 */
4005 	m = ((u64) 1 << (64 - 1));
4006 	for (l2sz = 64; l2sz >= 0; l2sz--, m >>= 1) {
4007 		if (m & nblocks)
4008 			break;
4009 	}
4010 
4011 	sz = (s64) 1 << l2sz;
4012 	if (sz < nblocks)
4013 		l2sz += 1;
4014 
4015 	/* agsize = roundupSize/max_number_of_ag */
4016 	return (l2sz - L2MAXAG);
4017 }
4018 
4019 
4020 /*
4021  * NAME:	dbMapFileSizeToMapSize()
4022  *
4023  * FUNCTION:	compute number of blocks the block allocation map file
4024  *		can cover from the map file size;
4025  *
4026  * RETURNS:	Number of blocks which can be covered by this block map file;
4027  */
4028 
4029 /*
4030  * maximum number of map pages at each level including control pages
4031  */
4032 #define MAXL0PAGES	(1 + LPERCTL)
4033 #define MAXL1PAGES	(1 + LPERCTL * MAXL0PAGES)
4034 
4035 /*
4036  * convert number of map pages to the zero origin top dmapctl level
4037  */
4038 #define BMAPPGTOLEV(npages)	\
4039 	(((npages) <= 3 + MAXL0PAGES) ? 0 : \
4040 	 ((npages) <= 2 + MAXL1PAGES) ? 1 : 2)
4041 
4042 s64 dbMapFileSizeToMapSize(struct inode * ipbmap)
4043 {
4044 	struct super_block *sb = ipbmap->i_sb;
4045 	s64 nblocks;
4046 	s64 npages, ndmaps;
4047 	int level, i;
4048 	int complete, factor;
4049 
4050 	nblocks = ipbmap->i_size >> JFS_SBI(sb)->l2bsize;
4051 	npages = nblocks >> JFS_SBI(sb)->l2nbperpage;
4052 	level = BMAPPGTOLEV(npages);
4053 
4054 	/* At each level, accumulate the number of dmap pages covered by
4055 	 * the number of full child levels below it;
4056 	 * repeat for the last incomplete child level.
4057 	 */
4058 	ndmaps = 0;
4059 	npages--;		/* skip the first global control page */
4060 	/* skip higher level control pages above top level covered by map */
4061 	npages -= (2 - level);
4062 	npages--;		/* skip top level's control page */
4063 	for (i = level; i >= 0; i--) {
4064 		factor =
4065 		    (i == 2) ? MAXL1PAGES : ((i == 1) ? MAXL0PAGES : 1);
4066 		complete = (u32) npages / factor;
4067 		ndmaps += complete * ((i == 2) ? LPERCTL * LPERCTL :
4068 				      ((i == 1) ? LPERCTL : 1));
4069 
4070 		/* pages in last/incomplete child */
4071 		npages = (u32) npages % factor;
4072 		/* skip incomplete child's level control page */
4073 		npages--;
4074 	}
4075 
4076 	/* convert the number of dmaps into the number of blocks
4077 	 * which can be covered by the dmaps;
4078 	 */
4079 	nblocks = ndmaps << L2BPERDMAP;
4080 
4081 	return (nblocks);
4082 }
4083