xref: /openbmc/linux/fs/jfs/jfs_dmap.c (revision 1a2bd3bd)
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  *	-EINVAL - wrong bmap data
152  */
153 int dbMount(struct inode *ipbmap)
154 {
155 	struct bmap *bmp;
156 	struct dbmap_disk *dbmp_le;
157 	struct metapage *mp;
158 	int i, err;
159 
160 	/*
161 	 * allocate/initialize the in-memory bmap descriptor
162 	 */
163 	/* allocate memory for the in-memory bmap descriptor */
164 	bmp = kmalloc(sizeof(struct bmap), GFP_KERNEL);
165 	if (bmp == NULL)
166 		return -ENOMEM;
167 
168 	/* read the on-disk bmap descriptor. */
169 	mp = read_metapage(ipbmap,
170 			   BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
171 			   PSIZE, 0);
172 	if (mp == NULL) {
173 		err = -EIO;
174 		goto err_kfree_bmp;
175 	}
176 
177 	/* copy the on-disk bmap descriptor to its in-memory version. */
178 	dbmp_le = (struct dbmap_disk *) mp->data;
179 	bmp->db_mapsize = le64_to_cpu(dbmp_le->dn_mapsize);
180 	bmp->db_nfree = le64_to_cpu(dbmp_le->dn_nfree);
181 	bmp->db_l2nbperpage = le32_to_cpu(dbmp_le->dn_l2nbperpage);
182 	bmp->db_numag = le32_to_cpu(dbmp_le->dn_numag);
183 	if (!bmp->db_numag) {
184 		err = -EINVAL;
185 		goto err_release_metapage;
186 	}
187 
188 	bmp->db_maxlevel = le32_to_cpu(dbmp_le->dn_maxlevel);
189 	bmp->db_maxag = le32_to_cpu(dbmp_le->dn_maxag);
190 	bmp->db_agpref = le32_to_cpu(dbmp_le->dn_agpref);
191 	bmp->db_aglevel = le32_to_cpu(dbmp_le->dn_aglevel);
192 	bmp->db_agheight = le32_to_cpu(dbmp_le->dn_agheight);
193 	bmp->db_agwidth = le32_to_cpu(dbmp_le->dn_agwidth);
194 	bmp->db_agstart = le32_to_cpu(dbmp_le->dn_agstart);
195 	bmp->db_agl2size = le32_to_cpu(dbmp_le->dn_agl2size);
196 	if (bmp->db_agl2size > L2MAXL2SIZE - L2MAXAG ||
197 	    bmp->db_agl2size < 0) {
198 		err = -EINVAL;
199 		goto err_release_metapage;
200 	}
201 
202 	if (((bmp->db_mapsize - 1) >> bmp->db_agl2size) > MAXAG) {
203 		err = -EINVAL;
204 		goto err_release_metapage;
205 	}
206 
207 	for (i = 0; i < MAXAG; i++)
208 		bmp->db_agfree[i] = le64_to_cpu(dbmp_le->dn_agfree[i]);
209 	bmp->db_agsize = le64_to_cpu(dbmp_le->dn_agsize);
210 	bmp->db_maxfreebud = dbmp_le->dn_maxfreebud;
211 
212 	/* release the buffer. */
213 	release_metapage(mp);
214 
215 	/* bind the bmap inode and the bmap descriptor to each other. */
216 	bmp->db_ipbmap = ipbmap;
217 	JFS_SBI(ipbmap->i_sb)->bmap = bmp;
218 
219 	memset(bmp->db_active, 0, sizeof(bmp->db_active));
220 
221 	/*
222 	 * allocate/initialize the bmap lock
223 	 */
224 	BMAP_LOCK_INIT(bmp);
225 
226 	return (0);
227 
228 err_release_metapage:
229 	release_metapage(mp);
230 err_kfree_bmp:
231 	kfree(bmp);
232 	return err;
233 }
234 
235 
236 /*
237  * NAME:	dbUnmount()
238  *
239  * FUNCTION:	terminate the block allocation map in preparation for
240  *		file system unmount.
241  *
242  *		the in-core bmap descriptor is written to disk and
243  *		the memory for this descriptor is freed.
244  *
245  * PARAMETERS:
246  *	ipbmap	- pointer to in-core inode for the block map.
247  *
248  * RETURN VALUES:
249  *	0	- success
250  *	-EIO	- i/o error
251  */
252 int dbUnmount(struct inode *ipbmap, int mounterror)
253 {
254 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
255 
256 	if (!(mounterror || isReadOnly(ipbmap)))
257 		dbSync(ipbmap);
258 
259 	/*
260 	 * Invalidate the page cache buffers
261 	 */
262 	truncate_inode_pages(ipbmap->i_mapping, 0);
263 
264 	/* free the memory for the in-memory bmap. */
265 	kfree(bmp);
266 
267 	return (0);
268 }
269 
270 /*
271  *	dbSync()
272  */
273 int dbSync(struct inode *ipbmap)
274 {
275 	struct dbmap_disk *dbmp_le;
276 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
277 	struct metapage *mp;
278 	int i;
279 
280 	/*
281 	 * write bmap global control page
282 	 */
283 	/* get the buffer for the on-disk bmap descriptor. */
284 	mp = read_metapage(ipbmap,
285 			   BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
286 			   PSIZE, 0);
287 	if (mp == NULL) {
288 		jfs_err("dbSync: read_metapage failed!");
289 		return -EIO;
290 	}
291 	/* copy the in-memory version of the bmap to the on-disk version */
292 	dbmp_le = (struct dbmap_disk *) mp->data;
293 	dbmp_le->dn_mapsize = cpu_to_le64(bmp->db_mapsize);
294 	dbmp_le->dn_nfree = cpu_to_le64(bmp->db_nfree);
295 	dbmp_le->dn_l2nbperpage = cpu_to_le32(bmp->db_l2nbperpage);
296 	dbmp_le->dn_numag = cpu_to_le32(bmp->db_numag);
297 	dbmp_le->dn_maxlevel = cpu_to_le32(bmp->db_maxlevel);
298 	dbmp_le->dn_maxag = cpu_to_le32(bmp->db_maxag);
299 	dbmp_le->dn_agpref = cpu_to_le32(bmp->db_agpref);
300 	dbmp_le->dn_aglevel = cpu_to_le32(bmp->db_aglevel);
301 	dbmp_le->dn_agheight = cpu_to_le32(bmp->db_agheight);
302 	dbmp_le->dn_agwidth = cpu_to_le32(bmp->db_agwidth);
303 	dbmp_le->dn_agstart = cpu_to_le32(bmp->db_agstart);
304 	dbmp_le->dn_agl2size = cpu_to_le32(bmp->db_agl2size);
305 	for (i = 0; i < MAXAG; i++)
306 		dbmp_le->dn_agfree[i] = cpu_to_le64(bmp->db_agfree[i]);
307 	dbmp_le->dn_agsize = cpu_to_le64(bmp->db_agsize);
308 	dbmp_le->dn_maxfreebud = bmp->db_maxfreebud;
309 
310 	/* write the buffer */
311 	write_metapage(mp);
312 
313 	/*
314 	 * write out dirty pages of bmap
315 	 */
316 	filemap_write_and_wait(ipbmap->i_mapping);
317 
318 	diWriteSpecial(ipbmap, 0);
319 
320 	return (0);
321 }
322 
323 /*
324  * NAME:	dbFree()
325  *
326  * FUNCTION:	free the specified block range from the working block
327  *		allocation map.
328  *
329  *		the blocks will be free from the working map one dmap
330  *		at a time.
331  *
332  * PARAMETERS:
333  *	ip	- pointer to in-core inode;
334  *	blkno	- starting block number to be freed.
335  *	nblocks	- number of blocks to be freed.
336  *
337  * RETURN VALUES:
338  *	0	- success
339  *	-EIO	- i/o error
340  */
341 int dbFree(struct inode *ip, s64 blkno, s64 nblocks)
342 {
343 	struct metapage *mp;
344 	struct dmap *dp;
345 	int nb, rc;
346 	s64 lblkno, rem;
347 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
348 	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
349 	struct super_block *sb = ipbmap->i_sb;
350 
351 	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
352 
353 	/* block to be freed better be within the mapsize. */
354 	if (unlikely((blkno == 0) || (blkno + nblocks > bmp->db_mapsize))) {
355 		IREAD_UNLOCK(ipbmap);
356 		printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
357 		       (unsigned long long) blkno,
358 		       (unsigned long long) nblocks);
359 		jfs_error(ip->i_sb, "block to be freed is outside the map\n");
360 		return -EIO;
361 	}
362 
363 	/**
364 	 * TRIM the blocks, when mounted with discard option
365 	 */
366 	if (JFS_SBI(sb)->flag & JFS_DISCARD)
367 		if (JFS_SBI(sb)->minblks_trim <= nblocks)
368 			jfs_issue_discard(ipbmap, blkno, nblocks);
369 
370 	/*
371 	 * free the blocks a dmap at a time.
372 	 */
373 	mp = NULL;
374 	for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
375 		/* release previous dmap if any */
376 		if (mp) {
377 			write_metapage(mp);
378 		}
379 
380 		/* get the buffer for the current dmap. */
381 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
382 		mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
383 		if (mp == NULL) {
384 			IREAD_UNLOCK(ipbmap);
385 			return -EIO;
386 		}
387 		dp = (struct dmap *) mp->data;
388 
389 		/* determine the number of blocks to be freed from
390 		 * this dmap.
391 		 */
392 		nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
393 
394 		/* free the blocks. */
395 		if ((rc = dbFreeDmap(bmp, dp, blkno, nb))) {
396 			jfs_error(ip->i_sb, "error in block map\n");
397 			release_metapage(mp);
398 			IREAD_UNLOCK(ipbmap);
399 			return (rc);
400 		}
401 	}
402 
403 	/* write the last buffer. */
404 	if (mp)
405 		write_metapage(mp);
406 
407 	IREAD_UNLOCK(ipbmap);
408 
409 	return (0);
410 }
411 
412 
413 /*
414  * NAME:	dbUpdatePMap()
415  *
416  * FUNCTION:	update the allocation state (free or allocate) of the
417  *		specified block range in the persistent block allocation map.
418  *
419  *		the blocks will be updated in the persistent map one
420  *		dmap at a time.
421  *
422  * PARAMETERS:
423  *	ipbmap	- pointer to in-core inode for the block map.
424  *	free	- 'true' if block range is to be freed from the persistent
425  *		  map; 'false' if it is to be allocated.
426  *	blkno	- starting block number of the range.
427  *	nblocks	- number of contiguous blocks in the range.
428  *	tblk	- transaction block;
429  *
430  * RETURN VALUES:
431  *	0	- success
432  *	-EIO	- i/o error
433  */
434 int
435 dbUpdatePMap(struct inode *ipbmap,
436 	     int free, s64 blkno, s64 nblocks, struct tblock * tblk)
437 {
438 	int nblks, dbitno, wbitno, rbits;
439 	int word, nbits, nwords;
440 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
441 	s64 lblkno, rem, lastlblkno;
442 	u32 mask;
443 	struct dmap *dp;
444 	struct metapage *mp;
445 	struct jfs_log *log;
446 	int lsn, difft, diffp;
447 	unsigned long flags;
448 
449 	/* the blocks better be within the mapsize. */
450 	if (blkno + nblocks > bmp->db_mapsize) {
451 		printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
452 		       (unsigned long long) blkno,
453 		       (unsigned long long) nblocks);
454 		jfs_error(ipbmap->i_sb, "blocks are outside the map\n");
455 		return -EIO;
456 	}
457 
458 	/* compute delta of transaction lsn from log syncpt */
459 	lsn = tblk->lsn;
460 	log = (struct jfs_log *) JFS_SBI(tblk->sb)->log;
461 	logdiff(difft, lsn, log);
462 
463 	/*
464 	 * update the block state a dmap at a time.
465 	 */
466 	mp = NULL;
467 	lastlblkno = 0;
468 	for (rem = nblocks; rem > 0; rem -= nblks, blkno += nblks) {
469 		/* get the buffer for the current dmap. */
470 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
471 		if (lblkno != lastlblkno) {
472 			if (mp) {
473 				write_metapage(mp);
474 			}
475 
476 			mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE,
477 					   0);
478 			if (mp == NULL)
479 				return -EIO;
480 			metapage_wait_for_io(mp);
481 		}
482 		dp = (struct dmap *) mp->data;
483 
484 		/* determine the bit number and word within the dmap of
485 		 * the starting block.  also determine how many blocks
486 		 * are to be updated within this dmap.
487 		 */
488 		dbitno = blkno & (BPERDMAP - 1);
489 		word = dbitno >> L2DBWORD;
490 		nblks = min(rem, (s64)BPERDMAP - dbitno);
491 
492 		/* update the bits of the dmap words. the first and last
493 		 * words may only have a subset of their bits updated. if
494 		 * this is the case, we'll work against that word (i.e.
495 		 * partial first and/or last) only in a single pass.  a
496 		 * single pass will also be used to update all words that
497 		 * are to have all their bits updated.
498 		 */
499 		for (rbits = nblks; rbits > 0;
500 		     rbits -= nbits, dbitno += nbits) {
501 			/* determine the bit number within the word and
502 			 * the number of bits within the word.
503 			 */
504 			wbitno = dbitno & (DBWORD - 1);
505 			nbits = min(rbits, DBWORD - wbitno);
506 
507 			/* check if only part of the word is to be updated. */
508 			if (nbits < DBWORD) {
509 				/* update (free or allocate) the bits
510 				 * in this word.
511 				 */
512 				mask =
513 				    (ONES << (DBWORD - nbits) >> wbitno);
514 				if (free)
515 					dp->pmap[word] &=
516 					    cpu_to_le32(~mask);
517 				else
518 					dp->pmap[word] |=
519 					    cpu_to_le32(mask);
520 
521 				word += 1;
522 			} else {
523 				/* one or more words are to have all
524 				 * their bits updated.  determine how
525 				 * many words and how many bits.
526 				 */
527 				nwords = rbits >> L2DBWORD;
528 				nbits = nwords << L2DBWORD;
529 
530 				/* update (free or allocate) the bits
531 				 * in these words.
532 				 */
533 				if (free)
534 					memset(&dp->pmap[word], 0,
535 					       nwords * 4);
536 				else
537 					memset(&dp->pmap[word], (int) ONES,
538 					       nwords * 4);
539 
540 				word += nwords;
541 			}
542 		}
543 
544 		/*
545 		 * update dmap lsn
546 		 */
547 		if (lblkno == lastlblkno)
548 			continue;
549 
550 		lastlblkno = lblkno;
551 
552 		LOGSYNC_LOCK(log, flags);
553 		if (mp->lsn != 0) {
554 			/* inherit older/smaller lsn */
555 			logdiff(diffp, mp->lsn, log);
556 			if (difft < diffp) {
557 				mp->lsn = lsn;
558 
559 				/* move bp after tblock in logsync list */
560 				list_move(&mp->synclist, &tblk->synclist);
561 			}
562 
563 			/* inherit younger/larger clsn */
564 			logdiff(difft, tblk->clsn, log);
565 			logdiff(diffp, mp->clsn, log);
566 			if (difft > diffp)
567 				mp->clsn = tblk->clsn;
568 		} else {
569 			mp->log = log;
570 			mp->lsn = lsn;
571 
572 			/* insert bp after tblock in logsync list */
573 			log->count++;
574 			list_add(&mp->synclist, &tblk->synclist);
575 
576 			mp->clsn = tblk->clsn;
577 		}
578 		LOGSYNC_UNLOCK(log, flags);
579 	}
580 
581 	/* write the last buffer. */
582 	if (mp) {
583 		write_metapage(mp);
584 	}
585 
586 	return (0);
587 }
588 
589 
590 /*
591  * NAME:	dbNextAG()
592  *
593  * FUNCTION:	find the preferred allocation group for new allocations.
594  *
595  *		Within the allocation groups, we maintain a preferred
596  *		allocation group which consists of a group with at least
597  *		average free space.  It is the preferred group that we target
598  *		new inode allocation towards.  The tie-in between inode
599  *		allocation and block allocation occurs as we allocate the
600  *		first (data) block of an inode and specify the inode (block)
601  *		as the allocation hint for this block.
602  *
603  *		We try to avoid having more than one open file growing in
604  *		an allocation group, as this will lead to fragmentation.
605  *		This differs from the old OS/2 method of trying to keep
606  *		empty ags around for large allocations.
607  *
608  * PARAMETERS:
609  *	ipbmap	- pointer to in-core inode for the block map.
610  *
611  * RETURN VALUES:
612  *	the preferred allocation group number.
613  */
614 int dbNextAG(struct inode *ipbmap)
615 {
616 	s64 avgfree;
617 	int agpref;
618 	s64 hwm = 0;
619 	int i;
620 	int next_best = -1;
621 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
622 
623 	BMAP_LOCK(bmp);
624 
625 	/* determine the average number of free blocks within the ags. */
626 	avgfree = (u32)bmp->db_nfree / bmp->db_numag;
627 
628 	/*
629 	 * if the current preferred ag does not have an active allocator
630 	 * and has at least average freespace, return it
631 	 */
632 	agpref = bmp->db_agpref;
633 	if ((atomic_read(&bmp->db_active[agpref]) == 0) &&
634 	    (bmp->db_agfree[agpref] >= avgfree))
635 		goto unlock;
636 
637 	/* From the last preferred ag, find the next one with at least
638 	 * average free space.
639 	 */
640 	for (i = 0 ; i < bmp->db_numag; i++, agpref++) {
641 		if (agpref == bmp->db_numag)
642 			agpref = 0;
643 
644 		if (atomic_read(&bmp->db_active[agpref]))
645 			/* open file is currently growing in this ag */
646 			continue;
647 		if (bmp->db_agfree[agpref] >= avgfree) {
648 			/* Return this one */
649 			bmp->db_agpref = agpref;
650 			goto unlock;
651 		} else if (bmp->db_agfree[agpref] > hwm) {
652 			/* Less than avg. freespace, but best so far */
653 			hwm = bmp->db_agfree[agpref];
654 			next_best = agpref;
655 		}
656 	}
657 
658 	/*
659 	 * If no inactive ag was found with average freespace, use the
660 	 * next best
661 	 */
662 	if (next_best != -1)
663 		bmp->db_agpref = next_best;
664 	/* else leave db_agpref unchanged */
665 unlock:
666 	BMAP_UNLOCK(bmp);
667 
668 	/* return the preferred group.
669 	 */
670 	return (bmp->db_agpref);
671 }
672 
673 /*
674  * NAME:	dbAlloc()
675  *
676  * FUNCTION:	attempt to allocate a specified number of contiguous free
677  *		blocks from the working allocation block map.
678  *
679  *		the block allocation policy uses hints and a multi-step
680  *		approach.
681  *
682  *		for allocation requests smaller than the number of blocks
683  *		per dmap, we first try to allocate the new blocks
684  *		immediately following the hint.  if these blocks are not
685  *		available, we try to allocate blocks near the hint.  if
686  *		no blocks near the hint are available, we next try to
687  *		allocate within the same dmap as contains the hint.
688  *
689  *		if no blocks are available in the dmap or the allocation
690  *		request is larger than the dmap size, we try to allocate
691  *		within the same allocation group as contains the hint. if
692  *		this does not succeed, we finally try to allocate anywhere
693  *		within the aggregate.
694  *
695  *		we also try to allocate anywhere within the aggregate
696  *		for allocation requests larger than the allocation group
697  *		size or requests that specify no hint value.
698  *
699  * PARAMETERS:
700  *	ip	- pointer to in-core inode;
701  *	hint	- allocation hint.
702  *	nblocks	- number of contiguous blocks in the range.
703  *	results	- on successful return, set to the starting block number
704  *		  of the newly allocated contiguous range.
705  *
706  * RETURN VALUES:
707  *	0	- success
708  *	-ENOSPC	- insufficient disk resources
709  *	-EIO	- i/o error
710  */
711 int dbAlloc(struct inode *ip, s64 hint, s64 nblocks, s64 * results)
712 {
713 	int rc, agno;
714 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
715 	struct bmap *bmp;
716 	struct metapage *mp;
717 	s64 lblkno, blkno;
718 	struct dmap *dp;
719 	int l2nb;
720 	s64 mapSize;
721 	int writers;
722 
723 	/* assert that nblocks is valid */
724 	assert(nblocks > 0);
725 
726 	/* get the log2 number of blocks to be allocated.
727 	 * if the number of blocks is not a log2 multiple,
728 	 * it will be rounded up to the next log2 multiple.
729 	 */
730 	l2nb = BLKSTOL2(nblocks);
731 
732 	bmp = JFS_SBI(ip->i_sb)->bmap;
733 
734 	mapSize = bmp->db_mapsize;
735 
736 	/* the hint should be within the map */
737 	if (hint >= mapSize) {
738 		jfs_error(ip->i_sb, "the hint is outside the map\n");
739 		return -EIO;
740 	}
741 
742 	/* if the number of blocks to be allocated is greater than the
743 	 * allocation group size, try to allocate anywhere.
744 	 */
745 	if (l2nb > bmp->db_agl2size) {
746 		IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
747 
748 		rc = dbAllocAny(bmp, nblocks, l2nb, results);
749 
750 		goto write_unlock;
751 	}
752 
753 	/*
754 	 * If no hint, let dbNextAG recommend an allocation group
755 	 */
756 	if (hint == 0)
757 		goto pref_ag;
758 
759 	/* we would like to allocate close to the hint.  adjust the
760 	 * hint to the block following the hint since the allocators
761 	 * will start looking for free space starting at this point.
762 	 */
763 	blkno = hint + 1;
764 
765 	if (blkno >= bmp->db_mapsize)
766 		goto pref_ag;
767 
768 	agno = blkno >> bmp->db_agl2size;
769 
770 	/* check if blkno crosses over into a new allocation group.
771 	 * if so, check if we should allow allocations within this
772 	 * allocation group.
773 	 */
774 	if ((blkno & (bmp->db_agsize - 1)) == 0)
775 		/* check if the AG is currently being written to.
776 		 * if so, call dbNextAG() to find a non-busy
777 		 * AG with sufficient free space.
778 		 */
779 		if (atomic_read(&bmp->db_active[agno]))
780 			goto pref_ag;
781 
782 	/* check if the allocation request size can be satisfied from a
783 	 * single dmap.  if so, try to allocate from the dmap containing
784 	 * the hint using a tiered strategy.
785 	 */
786 	if (nblocks <= BPERDMAP) {
787 		IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
788 
789 		/* get the buffer for the dmap containing the hint.
790 		 */
791 		rc = -EIO;
792 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
793 		mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
794 		if (mp == NULL)
795 			goto read_unlock;
796 
797 		dp = (struct dmap *) mp->data;
798 
799 		/* first, try to satisfy the allocation request with the
800 		 * blocks beginning at the hint.
801 		 */
802 		if ((rc = dbAllocNext(bmp, dp, blkno, (int) nblocks))
803 		    != -ENOSPC) {
804 			if (rc == 0) {
805 				*results = blkno;
806 				mark_metapage_dirty(mp);
807 			}
808 
809 			release_metapage(mp);
810 			goto read_unlock;
811 		}
812 
813 		writers = atomic_read(&bmp->db_active[agno]);
814 		if ((writers > 1) ||
815 		    ((writers == 1) && (JFS_IP(ip)->active_ag != agno))) {
816 			/*
817 			 * Someone else is writing in this allocation
818 			 * group.  To avoid fragmenting, try another ag
819 			 */
820 			release_metapage(mp);
821 			IREAD_UNLOCK(ipbmap);
822 			goto pref_ag;
823 		}
824 
825 		/* next, try to satisfy the allocation request with blocks
826 		 * near the hint.
827 		 */
828 		if ((rc =
829 		     dbAllocNear(bmp, dp, blkno, (int) nblocks, l2nb, results))
830 		    != -ENOSPC) {
831 			if (rc == 0)
832 				mark_metapage_dirty(mp);
833 
834 			release_metapage(mp);
835 			goto read_unlock;
836 		}
837 
838 		/* try to satisfy the allocation request with blocks within
839 		 * the same dmap as the hint.
840 		 */
841 		if ((rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results))
842 		    != -ENOSPC) {
843 			if (rc == 0)
844 				mark_metapage_dirty(mp);
845 
846 			release_metapage(mp);
847 			goto read_unlock;
848 		}
849 
850 		release_metapage(mp);
851 		IREAD_UNLOCK(ipbmap);
852 	}
853 
854 	/* try to satisfy the allocation request with blocks within
855 	 * the same allocation group as the hint.
856 	 */
857 	IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
858 	if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) != -ENOSPC)
859 		goto write_unlock;
860 
861 	IWRITE_UNLOCK(ipbmap);
862 
863 
864       pref_ag:
865 	/*
866 	 * Let dbNextAG recommend a preferred allocation group
867 	 */
868 	agno = dbNextAG(ipbmap);
869 	IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
870 
871 	/* Try to allocate within this allocation group.  if that fails, try to
872 	 * allocate anywhere in the map.
873 	 */
874 	if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) == -ENOSPC)
875 		rc = dbAllocAny(bmp, nblocks, l2nb, results);
876 
877       write_unlock:
878 	IWRITE_UNLOCK(ipbmap);
879 
880 	return (rc);
881 
882       read_unlock:
883 	IREAD_UNLOCK(ipbmap);
884 
885 	return (rc);
886 }
887 
888 /*
889  * NAME:	dbReAlloc()
890  *
891  * FUNCTION:	attempt to extend a current allocation by a specified
892  *		number of blocks.
893  *
894  *		this routine attempts to satisfy the allocation request
895  *		by first trying to extend the existing allocation in
896  *		place by allocating the additional blocks as the blocks
897  *		immediately following the current allocation.  if these
898  *		blocks are not available, this routine will attempt to
899  *		allocate a new set of contiguous blocks large enough
900  *		to cover the existing allocation plus the additional
901  *		number of blocks required.
902  *
903  * PARAMETERS:
904  *	ip	    -  pointer to in-core inode requiring allocation.
905  *	blkno	    -  starting block of the current allocation.
906  *	nblocks	    -  number of contiguous blocks within the current
907  *		       allocation.
908  *	addnblocks  -  number of blocks to add to the allocation.
909  *	results	-      on successful return, set to the starting block number
910  *		       of the existing allocation if the existing allocation
911  *		       was extended in place or to a newly allocated contiguous
912  *		       range if the existing allocation could not be extended
913  *		       in place.
914  *
915  * RETURN VALUES:
916  *	0	- success
917  *	-ENOSPC	- insufficient disk resources
918  *	-EIO	- i/o error
919  */
920 int
921 dbReAlloc(struct inode *ip,
922 	  s64 blkno, s64 nblocks, s64 addnblocks, s64 * results)
923 {
924 	int rc;
925 
926 	/* try to extend the allocation in place.
927 	 */
928 	if ((rc = dbExtend(ip, blkno, nblocks, addnblocks)) == 0) {
929 		*results = blkno;
930 		return (0);
931 	} else {
932 		if (rc != -ENOSPC)
933 			return (rc);
934 	}
935 
936 	/* could not extend the allocation in place, so allocate a
937 	 * new set of blocks for the entire request (i.e. try to get
938 	 * a range of contiguous blocks large enough to cover the
939 	 * existing allocation plus the additional blocks.)
940 	 */
941 	return (dbAlloc
942 		(ip, blkno + nblocks - 1, addnblocks + nblocks, results));
943 }
944 
945 
946 /*
947  * NAME:	dbExtend()
948  *
949  * FUNCTION:	attempt to extend a current allocation by a specified
950  *		number of blocks.
951  *
952  *		this routine attempts to satisfy the allocation request
953  *		by first trying to extend the existing allocation in
954  *		place by allocating the additional blocks as the blocks
955  *		immediately following the current allocation.
956  *
957  * PARAMETERS:
958  *	ip	    -  pointer to in-core inode requiring allocation.
959  *	blkno	    -  starting block of the current allocation.
960  *	nblocks	    -  number of contiguous blocks within the current
961  *		       allocation.
962  *	addnblocks  -  number of blocks to add to the allocation.
963  *
964  * RETURN VALUES:
965  *	0	- success
966  *	-ENOSPC	- insufficient disk resources
967  *	-EIO	- i/o error
968  */
969 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks)
970 {
971 	struct jfs_sb_info *sbi = JFS_SBI(ip->i_sb);
972 	s64 lblkno, lastblkno, extblkno;
973 	uint rel_block;
974 	struct metapage *mp;
975 	struct dmap *dp;
976 	int rc;
977 	struct inode *ipbmap = sbi->ipbmap;
978 	struct bmap *bmp;
979 
980 	/*
981 	 * We don't want a non-aligned extent to cross a page boundary
982 	 */
983 	if (((rel_block = blkno & (sbi->nbperpage - 1))) &&
984 	    (rel_block + nblocks + addnblocks > sbi->nbperpage))
985 		return -ENOSPC;
986 
987 	/* get the last block of the current allocation */
988 	lastblkno = blkno + nblocks - 1;
989 
990 	/* determine the block number of the block following
991 	 * the existing allocation.
992 	 */
993 	extblkno = lastblkno + 1;
994 
995 	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
996 
997 	/* better be within the file system */
998 	bmp = sbi->bmap;
999 	if (lastblkno < 0 || lastblkno >= bmp->db_mapsize) {
1000 		IREAD_UNLOCK(ipbmap);
1001 		jfs_error(ip->i_sb, "the block is outside the filesystem\n");
1002 		return -EIO;
1003 	}
1004 
1005 	/* we'll attempt to extend the current allocation in place by
1006 	 * allocating the additional blocks as the blocks immediately
1007 	 * following the current allocation.  we only try to extend the
1008 	 * current allocation in place if the number of additional blocks
1009 	 * can fit into a dmap, the last block of the current allocation
1010 	 * is not the last block of the file system, and the start of the
1011 	 * inplace extension is not on an allocation group boundary.
1012 	 */
1013 	if (addnblocks > BPERDMAP || extblkno >= bmp->db_mapsize ||
1014 	    (extblkno & (bmp->db_agsize - 1)) == 0) {
1015 		IREAD_UNLOCK(ipbmap);
1016 		return -ENOSPC;
1017 	}
1018 
1019 	/* get the buffer for the dmap containing the first block
1020 	 * of the extension.
1021 	 */
1022 	lblkno = BLKTODMAP(extblkno, bmp->db_l2nbperpage);
1023 	mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
1024 	if (mp == NULL) {
1025 		IREAD_UNLOCK(ipbmap);
1026 		return -EIO;
1027 	}
1028 
1029 	dp = (struct dmap *) mp->data;
1030 
1031 	/* try to allocate the blocks immediately following the
1032 	 * current allocation.
1033 	 */
1034 	rc = dbAllocNext(bmp, dp, extblkno, (int) addnblocks);
1035 
1036 	IREAD_UNLOCK(ipbmap);
1037 
1038 	/* were we successful ? */
1039 	if (rc == 0)
1040 		write_metapage(mp);
1041 	else
1042 		/* we were not successful */
1043 		release_metapage(mp);
1044 
1045 	return (rc);
1046 }
1047 
1048 
1049 /*
1050  * NAME:	dbAllocNext()
1051  *
1052  * FUNCTION:	attempt to allocate the blocks of the specified block
1053  *		range within a dmap.
1054  *
1055  * PARAMETERS:
1056  *	bmp	-  pointer to bmap descriptor
1057  *	dp	-  pointer to dmap.
1058  *	blkno	-  starting block number of the range.
1059  *	nblocks	-  number of contiguous free blocks of the range.
1060  *
1061  * RETURN VALUES:
1062  *	0	- success
1063  *	-ENOSPC	- insufficient disk resources
1064  *	-EIO	- i/o error
1065  *
1066  * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1067  */
1068 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
1069 		       int nblocks)
1070 {
1071 	int dbitno, word, rembits, nb, nwords, wbitno, nw;
1072 	int l2size;
1073 	s8 *leaf;
1074 	u32 mask;
1075 
1076 	if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1077 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1078 		return -EIO;
1079 	}
1080 
1081 	/* pick up a pointer to the leaves of the dmap tree.
1082 	 */
1083 	leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1084 
1085 	/* determine the bit number and word within the dmap of the
1086 	 * starting block.
1087 	 */
1088 	dbitno = blkno & (BPERDMAP - 1);
1089 	word = dbitno >> L2DBWORD;
1090 
1091 	/* check if the specified block range is contained within
1092 	 * this dmap.
1093 	 */
1094 	if (dbitno + nblocks > BPERDMAP)
1095 		return -ENOSPC;
1096 
1097 	/* check if the starting leaf indicates that anything
1098 	 * is free.
1099 	 */
1100 	if (leaf[word] == NOFREE)
1101 		return -ENOSPC;
1102 
1103 	/* check the dmaps words corresponding to block range to see
1104 	 * if the block range is free.  not all bits of the first and
1105 	 * last words may be contained within the block range.  if this
1106 	 * is the case, we'll work against those words (i.e. partial first
1107 	 * and/or last) on an individual basis (a single pass) and examine
1108 	 * the actual bits to determine if they are free.  a single pass
1109 	 * will be used for all dmap words fully contained within the
1110 	 * specified range.  within this pass, the leaves of the dmap
1111 	 * tree will be examined to determine if the blocks are free. a
1112 	 * single leaf may describe the free space of multiple dmap
1113 	 * words, so we may visit only a subset of the actual leaves
1114 	 * corresponding to the dmap words of the block range.
1115 	 */
1116 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
1117 		/* determine the bit number within the word and
1118 		 * the number of bits within the word.
1119 		 */
1120 		wbitno = dbitno & (DBWORD - 1);
1121 		nb = min(rembits, DBWORD - wbitno);
1122 
1123 		/* check if only part of the word is to be examined.
1124 		 */
1125 		if (nb < DBWORD) {
1126 			/* check if the bits are free.
1127 			 */
1128 			mask = (ONES << (DBWORD - nb) >> wbitno);
1129 			if ((mask & ~le32_to_cpu(dp->wmap[word])) != mask)
1130 				return -ENOSPC;
1131 
1132 			word += 1;
1133 		} else {
1134 			/* one or more dmap words are fully contained
1135 			 * within the block range.  determine how many
1136 			 * words and how many bits.
1137 			 */
1138 			nwords = rembits >> L2DBWORD;
1139 			nb = nwords << L2DBWORD;
1140 
1141 			/* now examine the appropriate leaves to determine
1142 			 * if the blocks are free.
1143 			 */
1144 			while (nwords > 0) {
1145 				/* does the leaf describe any free space ?
1146 				 */
1147 				if (leaf[word] < BUDMIN)
1148 					return -ENOSPC;
1149 
1150 				/* determine the l2 number of bits provided
1151 				 * by this leaf.
1152 				 */
1153 				l2size =
1154 				    min_t(int, leaf[word], NLSTOL2BSZ(nwords));
1155 
1156 				/* determine how many words were handled.
1157 				 */
1158 				nw = BUDSIZE(l2size, BUDMIN);
1159 
1160 				nwords -= nw;
1161 				word += nw;
1162 			}
1163 		}
1164 	}
1165 
1166 	/* allocate the blocks.
1167 	 */
1168 	return (dbAllocDmap(bmp, dp, blkno, nblocks));
1169 }
1170 
1171 
1172 /*
1173  * NAME:	dbAllocNear()
1174  *
1175  * FUNCTION:	attempt to allocate a number of contiguous free blocks near
1176  *		a specified block (hint) within a dmap.
1177  *
1178  *		starting with the dmap leaf that covers the hint, we'll
1179  *		check the next four contiguous leaves for sufficient free
1180  *		space.  if sufficient free space is found, we'll allocate
1181  *		the desired free space.
1182  *
1183  * PARAMETERS:
1184  *	bmp	-  pointer to bmap descriptor
1185  *	dp	-  pointer to dmap.
1186  *	blkno	-  block number to allocate near.
1187  *	nblocks	-  actual number of contiguous free blocks desired.
1188  *	l2nb	-  log2 number of contiguous free blocks desired.
1189  *	results	-  on successful return, set to the starting block number
1190  *		   of the newly allocated range.
1191  *
1192  * RETURN VALUES:
1193  *	0	- success
1194  *	-ENOSPC	- insufficient disk resources
1195  *	-EIO	- i/o error
1196  *
1197  * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1198  */
1199 static int
1200 dbAllocNear(struct bmap * bmp,
1201 	    struct dmap * dp, s64 blkno, int nblocks, int l2nb, s64 * results)
1202 {
1203 	int word, lword, rc;
1204 	s8 *leaf;
1205 
1206 	if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1207 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1208 		return -EIO;
1209 	}
1210 
1211 	leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1212 
1213 	/* determine the word within the dmap that holds the hint
1214 	 * (i.e. blkno).  also, determine the last word in the dmap
1215 	 * that we'll include in our examination.
1216 	 */
1217 	word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
1218 	lword = min(word + 4, LPERDMAP);
1219 
1220 	/* examine the leaves for sufficient free space.
1221 	 */
1222 	for (; word < lword; word++) {
1223 		/* does the leaf describe sufficient free space ?
1224 		 */
1225 		if (leaf[word] < l2nb)
1226 			continue;
1227 
1228 		/* determine the block number within the file system
1229 		 * of the first block described by this dmap word.
1230 		 */
1231 		blkno = le64_to_cpu(dp->start) + (word << L2DBWORD);
1232 
1233 		/* if not all bits of the dmap word are free, get the
1234 		 * starting bit number within the dmap word of the required
1235 		 * string of free bits and adjust the block number with the
1236 		 * value.
1237 		 */
1238 		if (leaf[word] < BUDMIN)
1239 			blkno +=
1240 			    dbFindBits(le32_to_cpu(dp->wmap[word]), l2nb);
1241 
1242 		/* allocate the blocks.
1243 		 */
1244 		if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1245 			*results = blkno;
1246 
1247 		return (rc);
1248 	}
1249 
1250 	return -ENOSPC;
1251 }
1252 
1253 
1254 /*
1255  * NAME:	dbAllocAG()
1256  *
1257  * FUNCTION:	attempt to allocate the specified number of contiguous
1258  *		free blocks within the specified allocation group.
1259  *
1260  *		unless the allocation group size is equal to the number
1261  *		of blocks per dmap, the dmap control pages will be used to
1262  *		find the required free space, if available.  we start the
1263  *		search at the highest dmap control page level which
1264  *		distinctly describes the allocation group's free space
1265  *		(i.e. the highest level at which the allocation group's
1266  *		free space is not mixed in with that of any other group).
1267  *		in addition, we start the search within this level at a
1268  *		height of the dmapctl dmtree at which the nodes distinctly
1269  *		describe the allocation group's free space.  at this height,
1270  *		the allocation group's free space may be represented by 1
1271  *		or two sub-trees, depending on the allocation group size.
1272  *		we search the top nodes of these subtrees left to right for
1273  *		sufficient free space.  if sufficient free space is found,
1274  *		the subtree is searched to find the leftmost leaf that
1275  *		has free space.  once we have made it to the leaf, we
1276  *		move the search to the next lower level dmap control page
1277  *		corresponding to this leaf.  we continue down the dmap control
1278  *		pages until we find the dmap that contains or starts the
1279  *		sufficient free space and we allocate at this dmap.
1280  *
1281  *		if the allocation group size is equal to the dmap size,
1282  *		we'll start at the dmap corresponding to the allocation
1283  *		group and attempt the allocation at this level.
1284  *
1285  *		the dmap control page search is also not performed if the
1286  *		allocation group is completely free and we go to the first
1287  *		dmap of the allocation group to do the allocation.  this is
1288  *		done because the allocation group may be part (not the first
1289  *		part) of a larger binary buddy system, causing the dmap
1290  *		control pages to indicate no free space (NOFREE) within
1291  *		the allocation group.
1292  *
1293  * PARAMETERS:
1294  *	bmp	-  pointer to bmap descriptor
1295  *	agno	- allocation group number.
1296  *	nblocks	-  actual number of contiguous free blocks desired.
1297  *	l2nb	-  log2 number of contiguous free blocks desired.
1298  *	results	-  on successful return, set to the starting block number
1299  *		   of the newly allocated range.
1300  *
1301  * RETURN VALUES:
1302  *	0	- success
1303  *	-ENOSPC	- insufficient disk resources
1304  *	-EIO	- i/o error
1305  *
1306  * note: IWRITE_LOCK(ipmap) held on entry/exit;
1307  */
1308 static int
1309 dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, s64 * results)
1310 {
1311 	struct metapage *mp;
1312 	struct dmapctl *dcp;
1313 	int rc, ti, i, k, m, n, agperlev;
1314 	s64 blkno, lblkno;
1315 	int budmin;
1316 
1317 	/* allocation request should not be for more than the
1318 	 * allocation group size.
1319 	 */
1320 	if (l2nb > bmp->db_agl2size) {
1321 		jfs_error(bmp->db_ipbmap->i_sb,
1322 			  "allocation request is larger than the allocation group size\n");
1323 		return -EIO;
1324 	}
1325 
1326 	/* determine the starting block number of the allocation
1327 	 * group.
1328 	 */
1329 	blkno = (s64) agno << bmp->db_agl2size;
1330 
1331 	/* check if the allocation group size is the minimum allocation
1332 	 * group size or if the allocation group is completely free. if
1333 	 * the allocation group size is the minimum size of BPERDMAP (i.e.
1334 	 * 1 dmap), there is no need to search the dmap control page (below)
1335 	 * that fully describes the allocation group since the allocation
1336 	 * group is already fully described by a dmap.  in this case, we
1337 	 * just call dbAllocCtl() to search the dmap tree and allocate the
1338 	 * required space if available.
1339 	 *
1340 	 * if the allocation group is completely free, dbAllocCtl() is
1341 	 * also called to allocate the required space.  this is done for
1342 	 * two reasons.  first, it makes no sense searching the dmap control
1343 	 * pages for free space when we know that free space exists.  second,
1344 	 * the dmap control pages may indicate that the allocation group
1345 	 * has no free space if the allocation group is part (not the first
1346 	 * part) of a larger binary buddy system.
1347 	 */
1348 	if (bmp->db_agsize == BPERDMAP
1349 	    || bmp->db_agfree[agno] == bmp->db_agsize) {
1350 		rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1351 		if ((rc == -ENOSPC) &&
1352 		    (bmp->db_agfree[agno] == bmp->db_agsize)) {
1353 			printk(KERN_ERR "blkno = %Lx, blocks = %Lx\n",
1354 			       (unsigned long long) blkno,
1355 			       (unsigned long long) nblocks);
1356 			jfs_error(bmp->db_ipbmap->i_sb,
1357 				  "dbAllocCtl failed in free AG\n");
1358 		}
1359 		return (rc);
1360 	}
1361 
1362 	/* the buffer for the dmap control page that fully describes the
1363 	 * allocation group.
1364 	 */
1365 	lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, bmp->db_aglevel);
1366 	mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1367 	if (mp == NULL)
1368 		return -EIO;
1369 	dcp = (struct dmapctl *) mp->data;
1370 	budmin = dcp->budmin;
1371 
1372 	if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1373 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
1374 		release_metapage(mp);
1375 		return -EIO;
1376 	}
1377 
1378 	/* search the subtree(s) of the dmap control page that describes
1379 	 * the allocation group, looking for sufficient free space.  to begin,
1380 	 * determine how many allocation groups are represented in a dmap
1381 	 * control page at the control page level (i.e. L0, L1, L2) that
1382 	 * fully describes an allocation group. next, determine the starting
1383 	 * tree index of this allocation group within the control page.
1384 	 */
1385 	agperlev =
1386 	    (1 << (L2LPERCTL - (bmp->db_agheight << 1))) / bmp->db_agwidth;
1387 	ti = bmp->db_agstart + bmp->db_agwidth * (agno & (agperlev - 1));
1388 
1389 	/* dmap control page trees fan-out by 4 and a single allocation
1390 	 * group may be described by 1 or 2 subtrees within the ag level
1391 	 * dmap control page, depending upon the ag size. examine the ag's
1392 	 * subtrees for sufficient free space, starting with the leftmost
1393 	 * subtree.
1394 	 */
1395 	for (i = 0; i < bmp->db_agwidth; i++, ti++) {
1396 		/* is there sufficient free space ?
1397 		 */
1398 		if (l2nb > dcp->stree[ti])
1399 			continue;
1400 
1401 		/* sufficient free space found in a subtree. now search down
1402 		 * the subtree to find the leftmost leaf that describes this
1403 		 * free space.
1404 		 */
1405 		for (k = bmp->db_agheight; k > 0; k--) {
1406 			for (n = 0, m = (ti << 2) + 1; n < 4; n++) {
1407 				if (l2nb <= dcp->stree[m + n]) {
1408 					ti = m + n;
1409 					break;
1410 				}
1411 			}
1412 			if (n == 4) {
1413 				jfs_error(bmp->db_ipbmap->i_sb,
1414 					  "failed descending stree\n");
1415 				release_metapage(mp);
1416 				return -EIO;
1417 			}
1418 		}
1419 
1420 		/* determine the block number within the file system
1421 		 * that corresponds to this leaf.
1422 		 */
1423 		if (bmp->db_aglevel == 2)
1424 			blkno = 0;
1425 		else if (bmp->db_aglevel == 1)
1426 			blkno &= ~(MAXL1SIZE - 1);
1427 		else		/* bmp->db_aglevel == 0 */
1428 			blkno &= ~(MAXL0SIZE - 1);
1429 
1430 		blkno +=
1431 		    ((s64) (ti - le32_to_cpu(dcp->leafidx))) << budmin;
1432 
1433 		/* release the buffer in preparation for going down
1434 		 * the next level of dmap control pages.
1435 		 */
1436 		release_metapage(mp);
1437 
1438 		/* check if we need to continue to search down the lower
1439 		 * level dmap control pages.  we need to if the number of
1440 		 * blocks required is less than maximum number of blocks
1441 		 * described at the next lower level.
1442 		 */
1443 		if (l2nb < budmin) {
1444 
1445 			/* search the lower level dmap control pages to get
1446 			 * the starting block number of the dmap that
1447 			 * contains or starts off the free space.
1448 			 */
1449 			if ((rc =
1450 			     dbFindCtl(bmp, l2nb, bmp->db_aglevel - 1,
1451 				       &blkno))) {
1452 				if (rc == -ENOSPC) {
1453 					jfs_error(bmp->db_ipbmap->i_sb,
1454 						  "control page inconsistent\n");
1455 					return -EIO;
1456 				}
1457 				return (rc);
1458 			}
1459 		}
1460 
1461 		/* allocate the blocks.
1462 		 */
1463 		rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1464 		if (rc == -ENOSPC) {
1465 			jfs_error(bmp->db_ipbmap->i_sb,
1466 				  "unable to allocate blocks\n");
1467 			rc = -EIO;
1468 		}
1469 		return (rc);
1470 	}
1471 
1472 	/* no space in the allocation group.  release the buffer and
1473 	 * return -ENOSPC.
1474 	 */
1475 	release_metapage(mp);
1476 
1477 	return -ENOSPC;
1478 }
1479 
1480 
1481 /*
1482  * NAME:	dbAllocAny()
1483  *
1484  * FUNCTION:	attempt to allocate the specified number of contiguous
1485  *		free blocks anywhere in the file system.
1486  *
1487  *		dbAllocAny() attempts to find the sufficient free space by
1488  *		searching down the dmap control pages, starting with the
1489  *		highest level (i.e. L0, L1, L2) control page.  if free space
1490  *		large enough to satisfy the desired free space is found, the
1491  *		desired free space is allocated.
1492  *
1493  * PARAMETERS:
1494  *	bmp	-  pointer to bmap descriptor
1495  *	nblocks	 -  actual number of contiguous free blocks desired.
1496  *	l2nb	 -  log2 number of contiguous free blocks desired.
1497  *	results	-  on successful return, set to the starting block number
1498  *		   of the newly allocated range.
1499  *
1500  * RETURN VALUES:
1501  *	0	- success
1502  *	-ENOSPC	- insufficient disk resources
1503  *	-EIO	- i/o error
1504  *
1505  * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1506  */
1507 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results)
1508 {
1509 	int rc;
1510 	s64 blkno = 0;
1511 
1512 	/* starting with the top level dmap control page, search
1513 	 * down the dmap control levels for sufficient free space.
1514 	 * if free space is found, dbFindCtl() returns the starting
1515 	 * block number of the dmap that contains or starts off the
1516 	 * range of free space.
1517 	 */
1518 	if ((rc = dbFindCtl(bmp, l2nb, bmp->db_maxlevel, &blkno)))
1519 		return (rc);
1520 
1521 	/* allocate the blocks.
1522 	 */
1523 	rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1524 	if (rc == -ENOSPC) {
1525 		jfs_error(bmp->db_ipbmap->i_sb, "unable to allocate blocks\n");
1526 		return -EIO;
1527 	}
1528 	return (rc);
1529 }
1530 
1531 
1532 /*
1533  * NAME:	dbDiscardAG()
1534  *
1535  * FUNCTION:	attempt to discard (TRIM) all free blocks of specific AG
1536  *
1537  *		algorithm:
1538  *		1) allocate blocks, as large as possible and save them
1539  *		   while holding IWRITE_LOCK on ipbmap
1540  *		2) trim all these saved block/length values
1541  *		3) mark the blocks free again
1542  *
1543  *		benefit:
1544  *		- we work only on one ag at some time, minimizing how long we
1545  *		  need to lock ipbmap
1546  *		- reading / writing the fs is possible most time, even on
1547  *		  trimming
1548  *
1549  *		downside:
1550  *		- we write two times to the dmapctl and dmap pages
1551  *		- but for me, this seems the best way, better ideas?
1552  *		/TR 2012
1553  *
1554  * PARAMETERS:
1555  *	ip	- pointer to in-core inode
1556  *	agno	- ag to trim
1557  *	minlen	- minimum value of contiguous blocks
1558  *
1559  * RETURN VALUES:
1560  *	s64	- actual number of blocks trimmed
1561  */
1562 s64 dbDiscardAG(struct inode *ip, int agno, s64 minlen)
1563 {
1564 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
1565 	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
1566 	s64 nblocks, blkno;
1567 	u64 trimmed = 0;
1568 	int rc, l2nb;
1569 	struct super_block *sb = ipbmap->i_sb;
1570 
1571 	struct range2trim {
1572 		u64 blkno;
1573 		u64 nblocks;
1574 	} *totrim, *tt;
1575 
1576 	/* max blkno / nblocks pairs to trim */
1577 	int count = 0, range_cnt;
1578 	u64 max_ranges;
1579 
1580 	/* prevent others from writing new stuff here, while trimming */
1581 	IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
1582 
1583 	nblocks = bmp->db_agfree[agno];
1584 	max_ranges = nblocks;
1585 	do_div(max_ranges, minlen);
1586 	range_cnt = min_t(u64, max_ranges + 1, 32 * 1024);
1587 	totrim = kmalloc_array(range_cnt, sizeof(struct range2trim), GFP_NOFS);
1588 	if (totrim == NULL) {
1589 		jfs_error(bmp->db_ipbmap->i_sb, "no memory for trim array\n");
1590 		IWRITE_UNLOCK(ipbmap);
1591 		return 0;
1592 	}
1593 
1594 	tt = totrim;
1595 	while (nblocks >= minlen) {
1596 		l2nb = BLKSTOL2(nblocks);
1597 
1598 		/* 0 = okay, -EIO = fatal, -ENOSPC -> try smaller block */
1599 		rc = dbAllocAG(bmp, agno, nblocks, l2nb, &blkno);
1600 		if (rc == 0) {
1601 			tt->blkno = blkno;
1602 			tt->nblocks = nblocks;
1603 			tt++; count++;
1604 
1605 			/* the whole ag is free, trim now */
1606 			if (bmp->db_agfree[agno] == 0)
1607 				break;
1608 
1609 			/* give a hint for the next while */
1610 			nblocks = bmp->db_agfree[agno];
1611 			continue;
1612 		} else if (rc == -ENOSPC) {
1613 			/* search for next smaller log2 block */
1614 			l2nb = BLKSTOL2(nblocks) - 1;
1615 			nblocks = 1LL << l2nb;
1616 		} else {
1617 			/* Trim any already allocated blocks */
1618 			jfs_error(bmp->db_ipbmap->i_sb, "-EIO\n");
1619 			break;
1620 		}
1621 
1622 		/* check, if our trim array is full */
1623 		if (unlikely(count >= range_cnt - 1))
1624 			break;
1625 	}
1626 	IWRITE_UNLOCK(ipbmap);
1627 
1628 	tt->nblocks = 0; /* mark the current end */
1629 	for (tt = totrim; tt->nblocks != 0; tt++) {
1630 		/* when mounted with online discard, dbFree() will
1631 		 * call jfs_issue_discard() itself */
1632 		if (!(JFS_SBI(sb)->flag & JFS_DISCARD))
1633 			jfs_issue_discard(ip, tt->blkno, tt->nblocks);
1634 		dbFree(ip, tt->blkno, tt->nblocks);
1635 		trimmed += tt->nblocks;
1636 	}
1637 	kfree(totrim);
1638 
1639 	return trimmed;
1640 }
1641 
1642 /*
1643  * NAME:	dbFindCtl()
1644  *
1645  * FUNCTION:	starting at a specified dmap control page level and block
1646  *		number, search down the dmap control levels for a range of
1647  *		contiguous free blocks large enough to satisfy an allocation
1648  *		request for the specified number of free blocks.
1649  *
1650  *		if sufficient contiguous free blocks are found, this routine
1651  *		returns the starting block number within a dmap page that
1652  *		contains or starts a range of contiqious free blocks that
1653  *		is sufficient in size.
1654  *
1655  * PARAMETERS:
1656  *	bmp	-  pointer to bmap descriptor
1657  *	level	-  starting dmap control page level.
1658  *	l2nb	-  log2 number of contiguous free blocks desired.
1659  *	*blkno	-  on entry, starting block number for conducting the search.
1660  *		   on successful return, the first block within a dmap page
1661  *		   that contains or starts a range of contiguous free blocks.
1662  *
1663  * RETURN VALUES:
1664  *	0	- success
1665  *	-ENOSPC	- insufficient disk resources
1666  *	-EIO	- i/o error
1667  *
1668  * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1669  */
1670 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno)
1671 {
1672 	int rc, leafidx, lev;
1673 	s64 b, lblkno;
1674 	struct dmapctl *dcp;
1675 	int budmin;
1676 	struct metapage *mp;
1677 
1678 	/* starting at the specified dmap control page level and block
1679 	 * number, search down the dmap control levels for the starting
1680 	 * block number of a dmap page that contains or starts off
1681 	 * sufficient free blocks.
1682 	 */
1683 	for (lev = level, b = *blkno; lev >= 0; lev--) {
1684 		/* get the buffer of the dmap control page for the block
1685 		 * number and level (i.e. L0, L1, L2).
1686 		 */
1687 		lblkno = BLKTOCTL(b, bmp->db_l2nbperpage, lev);
1688 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1689 		if (mp == NULL)
1690 			return -EIO;
1691 		dcp = (struct dmapctl *) mp->data;
1692 		budmin = dcp->budmin;
1693 
1694 		if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1695 			jfs_error(bmp->db_ipbmap->i_sb,
1696 				  "Corrupt dmapctl page\n");
1697 			release_metapage(mp);
1698 			return -EIO;
1699 		}
1700 
1701 		/* search the tree within the dmap control page for
1702 		 * sufficient free space.  if sufficient free space is found,
1703 		 * dbFindLeaf() returns the index of the leaf at which
1704 		 * free space was found.
1705 		 */
1706 		rc = dbFindLeaf((dmtree_t *) dcp, l2nb, &leafidx);
1707 
1708 		/* release the buffer.
1709 		 */
1710 		release_metapage(mp);
1711 
1712 		/* space found ?
1713 		 */
1714 		if (rc) {
1715 			if (lev != level) {
1716 				jfs_error(bmp->db_ipbmap->i_sb,
1717 					  "dmap inconsistent\n");
1718 				return -EIO;
1719 			}
1720 			return -ENOSPC;
1721 		}
1722 
1723 		/* adjust the block number to reflect the location within
1724 		 * the dmap control page (i.e. the leaf) at which free
1725 		 * space was found.
1726 		 */
1727 		b += (((s64) leafidx) << budmin);
1728 
1729 		/* we stop the search at this dmap control page level if
1730 		 * the number of blocks required is greater than or equal
1731 		 * to the maximum number of blocks described at the next
1732 		 * (lower) level.
1733 		 */
1734 		if (l2nb >= budmin)
1735 			break;
1736 	}
1737 
1738 	*blkno = b;
1739 	return (0);
1740 }
1741 
1742 
1743 /*
1744  * NAME:	dbAllocCtl()
1745  *
1746  * FUNCTION:	attempt to allocate a specified number of contiguous
1747  *		blocks starting within a specific dmap.
1748  *
1749  *		this routine is called by higher level routines that search
1750  *		the dmap control pages above the actual dmaps for contiguous
1751  *		free space.  the result of successful searches by these
1752  *		routines are the starting block numbers within dmaps, with
1753  *		the dmaps themselves containing the desired contiguous free
1754  *		space or starting a contiguous free space of desired size
1755  *		that is made up of the blocks of one or more dmaps. these
1756  *		calls should not fail due to insufficent resources.
1757  *
1758  *		this routine is called in some cases where it is not known
1759  *		whether it will fail due to insufficient resources.  more
1760  *		specifically, this occurs when allocating from an allocation
1761  *		group whose size is equal to the number of blocks per dmap.
1762  *		in this case, the dmap control pages are not examined prior
1763  *		to calling this routine (to save pathlength) and the call
1764  *		might fail.
1765  *
1766  *		for a request size that fits within a dmap, this routine relies
1767  *		upon the dmap's dmtree to find the requested contiguous free
1768  *		space.  for request sizes that are larger than a dmap, the
1769  *		requested free space will start at the first block of the
1770  *		first dmap (i.e. blkno).
1771  *
1772  * PARAMETERS:
1773  *	bmp	-  pointer to bmap descriptor
1774  *	nblocks	 -  actual number of contiguous free blocks to allocate.
1775  *	l2nb	 -  log2 number of contiguous free blocks to allocate.
1776  *	blkno	 -  starting block number of the dmap to start the allocation
1777  *		    from.
1778  *	results	-  on successful return, set to the starting block number
1779  *		   of the newly allocated range.
1780  *
1781  * RETURN VALUES:
1782  *	0	- success
1783  *	-ENOSPC	- insufficient disk resources
1784  *	-EIO	- i/o error
1785  *
1786  * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1787  */
1788 static int
1789 dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, s64 * results)
1790 {
1791 	int rc, nb;
1792 	s64 b, lblkno, n;
1793 	struct metapage *mp;
1794 	struct dmap *dp;
1795 
1796 	/* check if the allocation request is confined to a single dmap.
1797 	 */
1798 	if (l2nb <= L2BPERDMAP) {
1799 		/* get the buffer for the dmap.
1800 		 */
1801 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
1802 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1803 		if (mp == NULL)
1804 			return -EIO;
1805 		dp = (struct dmap *) mp->data;
1806 
1807 		/* try to allocate the blocks.
1808 		 */
1809 		rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results);
1810 		if (rc == 0)
1811 			mark_metapage_dirty(mp);
1812 
1813 		release_metapage(mp);
1814 
1815 		return (rc);
1816 	}
1817 
1818 	/* allocation request involving multiple dmaps. it must start on
1819 	 * a dmap boundary.
1820 	 */
1821 	assert((blkno & (BPERDMAP - 1)) == 0);
1822 
1823 	/* allocate the blocks dmap by dmap.
1824 	 */
1825 	for (n = nblocks, b = blkno; n > 0; n -= nb, b += nb) {
1826 		/* get the buffer for the dmap.
1827 		 */
1828 		lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1829 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1830 		if (mp == NULL) {
1831 			rc = -EIO;
1832 			goto backout;
1833 		}
1834 		dp = (struct dmap *) mp->data;
1835 
1836 		/* the dmap better be all free.
1837 		 */
1838 		if (dp->tree.stree[ROOT] != L2BPERDMAP) {
1839 			release_metapage(mp);
1840 			jfs_error(bmp->db_ipbmap->i_sb,
1841 				  "the dmap is not all free\n");
1842 			rc = -EIO;
1843 			goto backout;
1844 		}
1845 
1846 		/* determine how many blocks to allocate from this dmap.
1847 		 */
1848 		nb = min_t(s64, n, BPERDMAP);
1849 
1850 		/* allocate the blocks from the dmap.
1851 		 */
1852 		if ((rc = dbAllocDmap(bmp, dp, b, nb))) {
1853 			release_metapage(mp);
1854 			goto backout;
1855 		}
1856 
1857 		/* write the buffer.
1858 		 */
1859 		write_metapage(mp);
1860 	}
1861 
1862 	/* set the results (starting block number) and return.
1863 	 */
1864 	*results = blkno;
1865 	return (0);
1866 
1867 	/* something failed in handling an allocation request involving
1868 	 * multiple dmaps.  we'll try to clean up by backing out any
1869 	 * allocation that has already happened for this request.  if
1870 	 * we fail in backing out the allocation, we'll mark the file
1871 	 * system to indicate that blocks have been leaked.
1872 	 */
1873       backout:
1874 
1875 	/* try to backout the allocations dmap by dmap.
1876 	 */
1877 	for (n = nblocks - n, b = blkno; n > 0;
1878 	     n -= BPERDMAP, b += BPERDMAP) {
1879 		/* get the buffer for this dmap.
1880 		 */
1881 		lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1882 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1883 		if (mp == NULL) {
1884 			/* could not back out.  mark the file system
1885 			 * to indicate that we have leaked blocks.
1886 			 */
1887 			jfs_error(bmp->db_ipbmap->i_sb,
1888 				  "I/O Error: Block Leakage\n");
1889 			continue;
1890 		}
1891 		dp = (struct dmap *) mp->data;
1892 
1893 		/* free the blocks is this dmap.
1894 		 */
1895 		if (dbFreeDmap(bmp, dp, b, BPERDMAP)) {
1896 			/* could not back out.  mark the file system
1897 			 * to indicate that we have leaked blocks.
1898 			 */
1899 			release_metapage(mp);
1900 			jfs_error(bmp->db_ipbmap->i_sb, "Block Leakage\n");
1901 			continue;
1902 		}
1903 
1904 		/* write the buffer.
1905 		 */
1906 		write_metapage(mp);
1907 	}
1908 
1909 	return (rc);
1910 }
1911 
1912 
1913 /*
1914  * NAME:	dbAllocDmapLev()
1915  *
1916  * FUNCTION:	attempt to allocate a specified number of contiguous blocks
1917  *		from a specified dmap.
1918  *
1919  *		this routine checks if the contiguous blocks are available.
1920  *		if so, nblocks of blocks are allocated; otherwise, ENOSPC is
1921  *		returned.
1922  *
1923  * PARAMETERS:
1924  *	mp	-  pointer to bmap descriptor
1925  *	dp	-  pointer to dmap to attempt to allocate blocks from.
1926  *	l2nb	-  log2 number of contiguous block desired.
1927  *	nblocks	-  actual number of contiguous block desired.
1928  *	results	-  on successful return, set to the starting block number
1929  *		   of the newly allocated range.
1930  *
1931  * RETURN VALUES:
1932  *	0	- success
1933  *	-ENOSPC	- insufficient disk resources
1934  *	-EIO	- i/o error
1935  *
1936  * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or
1937  *	IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit;
1938  */
1939 static int
1940 dbAllocDmapLev(struct bmap * bmp,
1941 	       struct dmap * dp, int nblocks, int l2nb, s64 * results)
1942 {
1943 	s64 blkno;
1944 	int leafidx, rc;
1945 
1946 	/* can't be more than a dmaps worth of blocks */
1947 	assert(l2nb <= L2BPERDMAP);
1948 
1949 	/* search the tree within the dmap page for sufficient
1950 	 * free space.  if sufficient free space is found, dbFindLeaf()
1951 	 * returns the index of the leaf at which free space was found.
1952 	 */
1953 	if (dbFindLeaf((dmtree_t *) & dp->tree, l2nb, &leafidx))
1954 		return -ENOSPC;
1955 
1956 	/* determine the block number within the file system corresponding
1957 	 * to the leaf at which free space was found.
1958 	 */
1959 	blkno = le64_to_cpu(dp->start) + (leafidx << L2DBWORD);
1960 
1961 	/* if not all bits of the dmap word are free, get the starting
1962 	 * bit number within the dmap word of the required string of free
1963 	 * bits and adjust the block number with this value.
1964 	 */
1965 	if (dp->tree.stree[leafidx + LEAFIND] < BUDMIN)
1966 		blkno += dbFindBits(le32_to_cpu(dp->wmap[leafidx]), l2nb);
1967 
1968 	/* allocate the blocks */
1969 	if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1970 		*results = blkno;
1971 
1972 	return (rc);
1973 }
1974 
1975 
1976 /*
1977  * NAME:	dbAllocDmap()
1978  *
1979  * FUNCTION:	adjust the disk allocation map to reflect the allocation
1980  *		of a specified block range within a dmap.
1981  *
1982  *		this routine allocates the specified blocks from the dmap
1983  *		through a call to dbAllocBits(). if the allocation of the
1984  *		block range causes the maximum string of free blocks within
1985  *		the dmap to change (i.e. the value of the root of the dmap's
1986  *		dmtree), this routine will cause this change to be reflected
1987  *		up through the appropriate levels of the dmap control pages
1988  *		by a call to dbAdjCtl() for the L0 dmap control page that
1989  *		covers this dmap.
1990  *
1991  * PARAMETERS:
1992  *	bmp	-  pointer to bmap descriptor
1993  *	dp	-  pointer to dmap to allocate the block range from.
1994  *	blkno	-  starting block number of the block to be allocated.
1995  *	nblocks	-  number of blocks to be allocated.
1996  *
1997  * RETURN VALUES:
1998  *	0	- success
1999  *	-EIO	- i/o error
2000  *
2001  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2002  */
2003 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2004 		       int nblocks)
2005 {
2006 	s8 oldroot;
2007 	int rc;
2008 
2009 	/* save the current value of the root (i.e. maximum free string)
2010 	 * of the dmap tree.
2011 	 */
2012 	oldroot = dp->tree.stree[ROOT];
2013 
2014 	/* allocate the specified (blocks) bits */
2015 	dbAllocBits(bmp, dp, blkno, nblocks);
2016 
2017 	/* if the root has not changed, done. */
2018 	if (dp->tree.stree[ROOT] == oldroot)
2019 		return (0);
2020 
2021 	/* root changed. bubble the change up to the dmap control pages.
2022 	 * if the adjustment of the upper level control pages fails,
2023 	 * backout the bit allocation (thus making everything consistent).
2024 	 */
2025 	if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 1, 0)))
2026 		dbFreeBits(bmp, dp, blkno, nblocks);
2027 
2028 	return (rc);
2029 }
2030 
2031 
2032 /*
2033  * NAME:	dbFreeDmap()
2034  *
2035  * FUNCTION:	adjust the disk allocation map to reflect the allocation
2036  *		of a specified block range within a dmap.
2037  *
2038  *		this routine frees the specified blocks from the dmap through
2039  *		a call to dbFreeBits(). if the deallocation of the block range
2040  *		causes the maximum string of free blocks within the dmap to
2041  *		change (i.e. the value of the root of the dmap's dmtree), this
2042  *		routine will cause this change to be reflected up through the
2043  *		appropriate levels of the dmap control pages by a call to
2044  *		dbAdjCtl() for the L0 dmap control page that covers this dmap.
2045  *
2046  * PARAMETERS:
2047  *	bmp	-  pointer to bmap descriptor
2048  *	dp	-  pointer to dmap to free the block range from.
2049  *	blkno	-  starting block number of the block to be freed.
2050  *	nblocks	-  number of blocks to be freed.
2051  *
2052  * RETURN VALUES:
2053  *	0	- success
2054  *	-EIO	- i/o error
2055  *
2056  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2057  */
2058 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2059 		      int nblocks)
2060 {
2061 	s8 oldroot;
2062 	int rc = 0, word;
2063 
2064 	/* save the current value of the root (i.e. maximum free string)
2065 	 * of the dmap tree.
2066 	 */
2067 	oldroot = dp->tree.stree[ROOT];
2068 
2069 	/* free the specified (blocks) bits */
2070 	rc = dbFreeBits(bmp, dp, blkno, nblocks);
2071 
2072 	/* if error or the root has not changed, done. */
2073 	if (rc || (dp->tree.stree[ROOT] == oldroot))
2074 		return (rc);
2075 
2076 	/* root changed. bubble the change up to the dmap control pages.
2077 	 * if the adjustment of the upper level control pages fails,
2078 	 * backout the deallocation.
2079 	 */
2080 	if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 0, 0))) {
2081 		word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
2082 
2083 		/* as part of backing out the deallocation, we will have
2084 		 * to back split the dmap tree if the deallocation caused
2085 		 * the freed blocks to become part of a larger binary buddy
2086 		 * system.
2087 		 */
2088 		if (dp->tree.stree[word] == NOFREE)
2089 			dbBackSplit((dmtree_t *) & dp->tree, word);
2090 
2091 		dbAllocBits(bmp, dp, blkno, nblocks);
2092 	}
2093 
2094 	return (rc);
2095 }
2096 
2097 
2098 /*
2099  * NAME:	dbAllocBits()
2100  *
2101  * FUNCTION:	allocate a specified block range from a dmap.
2102  *
2103  *		this routine updates the dmap to reflect the working
2104  *		state allocation of the specified block range. it directly
2105  *		updates the bits of the working map and causes the adjustment
2106  *		of the binary buddy system described by the dmap's dmtree
2107  *		leaves to reflect the bits allocated.  it also causes the
2108  *		dmap's dmtree, as a whole, to reflect the allocated range.
2109  *
2110  * PARAMETERS:
2111  *	bmp	-  pointer to bmap descriptor
2112  *	dp	-  pointer to dmap to allocate bits from.
2113  *	blkno	-  starting block number of the bits to be allocated.
2114  *	nblocks	-  number of bits to be allocated.
2115  *
2116  * RETURN VALUES: none
2117  *
2118  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2119  */
2120 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2121 			int nblocks)
2122 {
2123 	int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2124 	dmtree_t *tp = (dmtree_t *) & dp->tree;
2125 	int size;
2126 	s8 *leaf;
2127 
2128 	/* pick up a pointer to the leaves of the dmap tree */
2129 	leaf = dp->tree.stree + LEAFIND;
2130 
2131 	/* determine the bit number and word within the dmap of the
2132 	 * starting block.
2133 	 */
2134 	dbitno = blkno & (BPERDMAP - 1);
2135 	word = dbitno >> L2DBWORD;
2136 
2137 	/* block range better be within the dmap */
2138 	assert(dbitno + nblocks <= BPERDMAP);
2139 
2140 	/* allocate the bits of the dmap's words corresponding to the block
2141 	 * range. not all bits of the first and last words may be contained
2142 	 * within the block range.  if this is the case, we'll work against
2143 	 * those words (i.e. partial first and/or last) on an individual basis
2144 	 * (a single pass), allocating the bits of interest by hand and
2145 	 * updating the leaf corresponding to the dmap word. a single pass
2146 	 * will be used for all dmap words fully contained within the
2147 	 * specified range.  within this pass, the bits of all fully contained
2148 	 * dmap words will be marked as free in a single shot and the leaves
2149 	 * will be updated. a single leaf may describe the free space of
2150 	 * multiple dmap words, so we may update only a subset of the actual
2151 	 * leaves corresponding to the dmap words of the block range.
2152 	 */
2153 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2154 		/* determine the bit number within the word and
2155 		 * the number of bits within the word.
2156 		 */
2157 		wbitno = dbitno & (DBWORD - 1);
2158 		nb = min(rembits, DBWORD - wbitno);
2159 
2160 		/* check if only part of a word is to be allocated.
2161 		 */
2162 		if (nb < DBWORD) {
2163 			/* allocate (set to 1) the appropriate bits within
2164 			 * this dmap word.
2165 			 */
2166 			dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
2167 						      >> wbitno);
2168 
2169 			/* update the leaf for this dmap word. in addition
2170 			 * to setting the leaf value to the binary buddy max
2171 			 * of the updated dmap word, dbSplit() will split
2172 			 * the binary system of the leaves if need be.
2173 			 */
2174 			dbSplit(tp, word, BUDMIN,
2175 				dbMaxBud((u8 *) & dp->wmap[word]));
2176 
2177 			word += 1;
2178 		} else {
2179 			/* one or more dmap words are fully contained
2180 			 * within the block range.  determine how many
2181 			 * words and allocate (set to 1) the bits of these
2182 			 * words.
2183 			 */
2184 			nwords = rembits >> L2DBWORD;
2185 			memset(&dp->wmap[word], (int) ONES, nwords * 4);
2186 
2187 			/* determine how many bits.
2188 			 */
2189 			nb = nwords << L2DBWORD;
2190 
2191 			/* now update the appropriate leaves to reflect
2192 			 * the allocated words.
2193 			 */
2194 			for (; nwords > 0; nwords -= nw) {
2195 				if (leaf[word] < BUDMIN) {
2196 					jfs_error(bmp->db_ipbmap->i_sb,
2197 						  "leaf page corrupt\n");
2198 					break;
2199 				}
2200 
2201 				/* determine what the leaf value should be
2202 				 * updated to as the minimum of the l2 number
2203 				 * of bits being allocated and the l2 number
2204 				 * of bits currently described by this leaf.
2205 				 */
2206 				size = min_t(int, leaf[word],
2207 					     NLSTOL2BSZ(nwords));
2208 
2209 				/* update the leaf to reflect the allocation.
2210 				 * in addition to setting the leaf value to
2211 				 * NOFREE, dbSplit() will split the binary
2212 				 * system of the leaves to reflect the current
2213 				 * allocation (size).
2214 				 */
2215 				dbSplit(tp, word, size, NOFREE);
2216 
2217 				/* get the number of dmap words handled */
2218 				nw = BUDSIZE(size, BUDMIN);
2219 				word += nw;
2220 			}
2221 		}
2222 	}
2223 
2224 	/* update the free count for this dmap */
2225 	le32_add_cpu(&dp->nfree, -nblocks);
2226 
2227 	BMAP_LOCK(bmp);
2228 
2229 	/* if this allocation group is completely free,
2230 	 * update the maximum allocation group number if this allocation
2231 	 * group is the new max.
2232 	 */
2233 	agno = blkno >> bmp->db_agl2size;
2234 	if (agno > bmp->db_maxag)
2235 		bmp->db_maxag = agno;
2236 
2237 	/* update the free count for the allocation group and map */
2238 	bmp->db_agfree[agno] -= nblocks;
2239 	bmp->db_nfree -= nblocks;
2240 
2241 	BMAP_UNLOCK(bmp);
2242 }
2243 
2244 
2245 /*
2246  * NAME:	dbFreeBits()
2247  *
2248  * FUNCTION:	free a specified block range from a dmap.
2249  *
2250  *		this routine updates the dmap to reflect the working
2251  *		state allocation of the specified block range. it directly
2252  *		updates the bits of the working map and causes the adjustment
2253  *		of the binary buddy system described by the dmap's dmtree
2254  *		leaves to reflect the bits freed.  it also causes the dmap's
2255  *		dmtree, as a whole, to reflect the deallocated range.
2256  *
2257  * PARAMETERS:
2258  *	bmp	-  pointer to bmap descriptor
2259  *	dp	-  pointer to dmap to free bits from.
2260  *	blkno	-  starting block number of the bits to be freed.
2261  *	nblocks	-  number of bits to be freed.
2262  *
2263  * RETURN VALUES: 0 for success
2264  *
2265  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2266  */
2267 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2268 		       int nblocks)
2269 {
2270 	int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2271 	dmtree_t *tp = (dmtree_t *) & dp->tree;
2272 	int rc = 0;
2273 	int size;
2274 
2275 	/* determine the bit number and word within the dmap of the
2276 	 * starting block.
2277 	 */
2278 	dbitno = blkno & (BPERDMAP - 1);
2279 	word = dbitno >> L2DBWORD;
2280 
2281 	/* block range better be within the dmap.
2282 	 */
2283 	assert(dbitno + nblocks <= BPERDMAP);
2284 
2285 	/* free the bits of the dmaps words corresponding to the block range.
2286 	 * not all bits of the first and last words may be contained within
2287 	 * the block range.  if this is the case, we'll work against those
2288 	 * words (i.e. partial first and/or last) on an individual basis
2289 	 * (a single pass), freeing the bits of interest by hand and updating
2290 	 * the leaf corresponding to the dmap word. a single pass will be used
2291 	 * for all dmap words fully contained within the specified range.
2292 	 * within this pass, the bits of all fully contained dmap words will
2293 	 * be marked as free in a single shot and the leaves will be updated. a
2294 	 * single leaf may describe the free space of multiple dmap words,
2295 	 * so we may update only a subset of the actual leaves corresponding
2296 	 * to the dmap words of the block range.
2297 	 *
2298 	 * dbJoin() is used to update leaf values and will join the binary
2299 	 * buddy system of the leaves if the new leaf values indicate this
2300 	 * should be done.
2301 	 */
2302 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2303 		/* determine the bit number within the word and
2304 		 * the number of bits within the word.
2305 		 */
2306 		wbitno = dbitno & (DBWORD - 1);
2307 		nb = min(rembits, DBWORD - wbitno);
2308 
2309 		/* check if only part of a word is to be freed.
2310 		 */
2311 		if (nb < DBWORD) {
2312 			/* free (zero) the appropriate bits within this
2313 			 * dmap word.
2314 			 */
2315 			dp->wmap[word] &=
2316 			    cpu_to_le32(~(ONES << (DBWORD - nb)
2317 					  >> wbitno));
2318 
2319 			/* update the leaf for this dmap word.
2320 			 */
2321 			rc = dbJoin(tp, word,
2322 				    dbMaxBud((u8 *) & dp->wmap[word]));
2323 			if (rc)
2324 				return rc;
2325 
2326 			word += 1;
2327 		} else {
2328 			/* one or more dmap words are fully contained
2329 			 * within the block range.  determine how many
2330 			 * words and free (zero) the bits of these words.
2331 			 */
2332 			nwords = rembits >> L2DBWORD;
2333 			memset(&dp->wmap[word], 0, nwords * 4);
2334 
2335 			/* determine how many bits.
2336 			 */
2337 			nb = nwords << L2DBWORD;
2338 
2339 			/* now update the appropriate leaves to reflect
2340 			 * the freed words.
2341 			 */
2342 			for (; nwords > 0; nwords -= nw) {
2343 				/* determine what the leaf value should be
2344 				 * updated to as the minimum of the l2 number
2345 				 * of bits being freed and the l2 (max) number
2346 				 * of bits that can be described by this leaf.
2347 				 */
2348 				size =
2349 				    min(LITOL2BSZ
2350 					(word, L2LPERDMAP, BUDMIN),
2351 					NLSTOL2BSZ(nwords));
2352 
2353 				/* update the leaf.
2354 				 */
2355 				rc = dbJoin(tp, word, size);
2356 				if (rc)
2357 					return rc;
2358 
2359 				/* get the number of dmap words handled.
2360 				 */
2361 				nw = BUDSIZE(size, BUDMIN);
2362 				word += nw;
2363 			}
2364 		}
2365 	}
2366 
2367 	/* update the free count for this dmap.
2368 	 */
2369 	le32_add_cpu(&dp->nfree, nblocks);
2370 
2371 	BMAP_LOCK(bmp);
2372 
2373 	/* update the free count for the allocation group and
2374 	 * map.
2375 	 */
2376 	agno = blkno >> bmp->db_agl2size;
2377 	bmp->db_nfree += nblocks;
2378 	bmp->db_agfree[agno] += nblocks;
2379 
2380 	/* check if this allocation group is not completely free and
2381 	 * if it is currently the maximum (rightmost) allocation group.
2382 	 * if so, establish the new maximum allocation group number by
2383 	 * searching left for the first allocation group with allocation.
2384 	 */
2385 	if ((bmp->db_agfree[agno] == bmp->db_agsize && agno == bmp->db_maxag) ||
2386 	    (agno == bmp->db_numag - 1 &&
2387 	     bmp->db_agfree[agno] == (bmp-> db_mapsize & (BPERDMAP - 1)))) {
2388 		while (bmp->db_maxag > 0) {
2389 			bmp->db_maxag -= 1;
2390 			if (bmp->db_agfree[bmp->db_maxag] !=
2391 			    bmp->db_agsize)
2392 				break;
2393 		}
2394 
2395 		/* re-establish the allocation group preference if the
2396 		 * current preference is right of the maximum allocation
2397 		 * group.
2398 		 */
2399 		if (bmp->db_agpref > bmp->db_maxag)
2400 			bmp->db_agpref = bmp->db_maxag;
2401 	}
2402 
2403 	BMAP_UNLOCK(bmp);
2404 
2405 	return 0;
2406 }
2407 
2408 
2409 /*
2410  * NAME:	dbAdjCtl()
2411  *
2412  * FUNCTION:	adjust a dmap control page at a specified level to reflect
2413  *		the change in a lower level dmap or dmap control page's
2414  *		maximum string of free blocks (i.e. a change in the root
2415  *		of the lower level object's dmtree) due to the allocation
2416  *		or deallocation of a range of blocks with a single dmap.
2417  *
2418  *		on entry, this routine is provided with the new value of
2419  *		the lower level dmap or dmap control page root and the
2420  *		starting block number of the block range whose allocation
2421  *		or deallocation resulted in the root change.  this range
2422  *		is respresented by a single leaf of the current dmapctl
2423  *		and the leaf will be updated with this value, possibly
2424  *		causing a binary buddy system within the leaves to be
2425  *		split or joined.  the update may also cause the dmapctl's
2426  *		dmtree to be updated.
2427  *
2428  *		if the adjustment of the dmap control page, itself, causes its
2429  *		root to change, this change will be bubbled up to the next dmap
2430  *		control level by a recursive call to this routine, specifying
2431  *		the new root value and the next dmap control page level to
2432  *		be adjusted.
2433  * PARAMETERS:
2434  *	bmp	-  pointer to bmap descriptor
2435  *	blkno	-  the first block of a block range within a dmap.  it is
2436  *		   the allocation or deallocation of this block range that
2437  *		   requires the dmap control page to be adjusted.
2438  *	newval	-  the new value of the lower level dmap or dmap control
2439  *		   page root.
2440  *	alloc	-  'true' if adjustment is due to an allocation.
2441  *	level	-  current level of dmap control page (i.e. L0, L1, L2) to
2442  *		   be adjusted.
2443  *
2444  * RETURN VALUES:
2445  *	0	- success
2446  *	-EIO	- i/o error
2447  *
2448  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2449  */
2450 static int
2451 dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, int level)
2452 {
2453 	struct metapage *mp;
2454 	s8 oldroot;
2455 	int oldval;
2456 	s64 lblkno;
2457 	struct dmapctl *dcp;
2458 	int rc, leafno, ti;
2459 
2460 	/* get the buffer for the dmap control page for the specified
2461 	 * block number and control page level.
2462 	 */
2463 	lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, level);
2464 	mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
2465 	if (mp == NULL)
2466 		return -EIO;
2467 	dcp = (struct dmapctl *) mp->data;
2468 
2469 	if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
2470 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
2471 		release_metapage(mp);
2472 		return -EIO;
2473 	}
2474 
2475 	/* determine the leaf number corresponding to the block and
2476 	 * the index within the dmap control tree.
2477 	 */
2478 	leafno = BLKTOCTLLEAF(blkno, dcp->budmin);
2479 	ti = leafno + le32_to_cpu(dcp->leafidx);
2480 
2481 	/* save the current leaf value and the current root level (i.e.
2482 	 * maximum l2 free string described by this dmapctl).
2483 	 */
2484 	oldval = dcp->stree[ti];
2485 	oldroot = dcp->stree[ROOT];
2486 
2487 	/* check if this is a control page update for an allocation.
2488 	 * if so, update the leaf to reflect the new leaf value using
2489 	 * dbSplit(); otherwise (deallocation), use dbJoin() to update
2490 	 * the leaf with the new value.  in addition to updating the
2491 	 * leaf, dbSplit() will also split the binary buddy system of
2492 	 * the leaves, if required, and bubble new values within the
2493 	 * dmapctl tree, if required.  similarly, dbJoin() will join
2494 	 * the binary buddy system of leaves and bubble new values up
2495 	 * the dmapctl tree as required by the new leaf value.
2496 	 */
2497 	if (alloc) {
2498 		/* check if we are in the middle of a binary buddy
2499 		 * system.  this happens when we are performing the
2500 		 * first allocation out of an allocation group that
2501 		 * is part (not the first part) of a larger binary
2502 		 * buddy system.  if we are in the middle, back split
2503 		 * the system prior to calling dbSplit() which assumes
2504 		 * that it is at the front of a binary buddy system.
2505 		 */
2506 		if (oldval == NOFREE) {
2507 			rc = dbBackSplit((dmtree_t *) dcp, leafno);
2508 			if (rc) {
2509 				release_metapage(mp);
2510 				return rc;
2511 			}
2512 			oldval = dcp->stree[ti];
2513 		}
2514 		dbSplit((dmtree_t *) dcp, leafno, dcp->budmin, newval);
2515 	} else {
2516 		rc = dbJoin((dmtree_t *) dcp, leafno, newval);
2517 		if (rc) {
2518 			release_metapage(mp);
2519 			return rc;
2520 		}
2521 	}
2522 
2523 	/* check if the root of the current dmap control page changed due
2524 	 * to the update and if the current dmap control page is not at
2525 	 * the current top level (i.e. L0, L1, L2) of the map.  if so (i.e.
2526 	 * root changed and this is not the top level), call this routine
2527 	 * again (recursion) for the next higher level of the mapping to
2528 	 * reflect the change in root for the current dmap control page.
2529 	 */
2530 	if (dcp->stree[ROOT] != oldroot) {
2531 		/* are we below the top level of the map.  if so,
2532 		 * bubble the root up to the next higher level.
2533 		 */
2534 		if (level < bmp->db_maxlevel) {
2535 			/* bubble up the new root of this dmap control page to
2536 			 * the next level.
2537 			 */
2538 			if ((rc =
2539 			     dbAdjCtl(bmp, blkno, dcp->stree[ROOT], alloc,
2540 				      level + 1))) {
2541 				/* something went wrong in bubbling up the new
2542 				 * root value, so backout the changes to the
2543 				 * current dmap control page.
2544 				 */
2545 				if (alloc) {
2546 					dbJoin((dmtree_t *) dcp, leafno,
2547 					       oldval);
2548 				} else {
2549 					/* the dbJoin() above might have
2550 					 * caused a larger binary buddy system
2551 					 * to form and we may now be in the
2552 					 * middle of it.  if this is the case,
2553 					 * back split the buddies.
2554 					 */
2555 					if (dcp->stree[ti] == NOFREE)
2556 						dbBackSplit((dmtree_t *)
2557 							    dcp, leafno);
2558 					dbSplit((dmtree_t *) dcp, leafno,
2559 						dcp->budmin, oldval);
2560 				}
2561 
2562 				/* release the buffer and return the error.
2563 				 */
2564 				release_metapage(mp);
2565 				return (rc);
2566 			}
2567 		} else {
2568 			/* we're at the top level of the map. update
2569 			 * the bmap control page to reflect the size
2570 			 * of the maximum free buddy system.
2571 			 */
2572 			assert(level == bmp->db_maxlevel);
2573 			if (bmp->db_maxfreebud != oldroot) {
2574 				jfs_error(bmp->db_ipbmap->i_sb,
2575 					  "the maximum free buddy is not the old root\n");
2576 			}
2577 			bmp->db_maxfreebud = dcp->stree[ROOT];
2578 		}
2579 	}
2580 
2581 	/* write the buffer.
2582 	 */
2583 	write_metapage(mp);
2584 
2585 	return (0);
2586 }
2587 
2588 
2589 /*
2590  * NAME:	dbSplit()
2591  *
2592  * FUNCTION:	update the leaf of a dmtree with a new value, splitting
2593  *		the leaf from the binary buddy system of the dmtree's
2594  *		leaves, as required.
2595  *
2596  * PARAMETERS:
2597  *	tp	- pointer to the tree containing the leaf.
2598  *	leafno	- the number of the leaf to be updated.
2599  *	splitsz	- the size the binary buddy system starting at the leaf
2600  *		  must be split to, specified as the log2 number of blocks.
2601  *	newval	- the new value for the leaf.
2602  *
2603  * RETURN VALUES: none
2604  *
2605  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2606  */
2607 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval)
2608 {
2609 	int budsz;
2610 	int cursz;
2611 	s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2612 
2613 	/* check if the leaf needs to be split.
2614 	 */
2615 	if (leaf[leafno] > tp->dmt_budmin) {
2616 		/* the split occurs by cutting the buddy system in half
2617 		 * at the specified leaf until we reach the specified
2618 		 * size.  pick up the starting split size (current size
2619 		 * - 1 in l2) and the corresponding buddy size.
2620 		 */
2621 		cursz = leaf[leafno] - 1;
2622 		budsz = BUDSIZE(cursz, tp->dmt_budmin);
2623 
2624 		/* split until we reach the specified size.
2625 		 */
2626 		while (cursz >= splitsz) {
2627 			/* update the buddy's leaf with its new value.
2628 			 */
2629 			dbAdjTree(tp, leafno ^ budsz, cursz);
2630 
2631 			/* on to the next size and buddy.
2632 			 */
2633 			cursz -= 1;
2634 			budsz >>= 1;
2635 		}
2636 	}
2637 
2638 	/* adjust the dmap tree to reflect the specified leaf's new
2639 	 * value.
2640 	 */
2641 	dbAdjTree(tp, leafno, newval);
2642 }
2643 
2644 
2645 /*
2646  * NAME:	dbBackSplit()
2647  *
2648  * FUNCTION:	back split the binary buddy system of dmtree leaves
2649  *		that hold a specified leaf until the specified leaf
2650  *		starts its own binary buddy system.
2651  *
2652  *		the allocators typically perform allocations at the start
2653  *		of binary buddy systems and dbSplit() is used to accomplish
2654  *		any required splits.  in some cases, however, allocation
2655  *		may occur in the middle of a binary system and requires a
2656  *		back split, with the split proceeding out from the middle of
2657  *		the system (less efficient) rather than the start of the
2658  *		system (more efficient).  the cases in which a back split
2659  *		is required are rare and are limited to the first allocation
2660  *		within an allocation group which is a part (not first part)
2661  *		of a larger binary buddy system and a few exception cases
2662  *		in which a previous join operation must be backed out.
2663  *
2664  * PARAMETERS:
2665  *	tp	- pointer to the tree containing the leaf.
2666  *	leafno	- the number of the leaf to be updated.
2667  *
2668  * RETURN VALUES: none
2669  *
2670  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2671  */
2672 static int dbBackSplit(dmtree_t * tp, int leafno)
2673 {
2674 	int budsz, bud, w, bsz, size;
2675 	int cursz;
2676 	s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2677 
2678 	/* leaf should be part (not first part) of a binary
2679 	 * buddy system.
2680 	 */
2681 	assert(leaf[leafno] == NOFREE);
2682 
2683 	/* the back split is accomplished by iteratively finding the leaf
2684 	 * that starts the buddy system that contains the specified leaf and
2685 	 * splitting that system in two.  this iteration continues until
2686 	 * the specified leaf becomes the start of a buddy system.
2687 	 *
2688 	 * determine maximum possible l2 size for the specified leaf.
2689 	 */
2690 	size =
2691 	    LITOL2BSZ(leafno, le32_to_cpu(tp->dmt_l2nleafs),
2692 		      tp->dmt_budmin);
2693 
2694 	/* determine the number of leaves covered by this size.  this
2695 	 * is the buddy size that we will start with as we search for
2696 	 * the buddy system that contains the specified leaf.
2697 	 */
2698 	budsz = BUDSIZE(size, tp->dmt_budmin);
2699 
2700 	/* back split.
2701 	 */
2702 	while (leaf[leafno] == NOFREE) {
2703 		/* find the leftmost buddy leaf.
2704 		 */
2705 		for (w = leafno, bsz = budsz;; bsz <<= 1,
2706 		     w = (w < bud) ? w : bud) {
2707 			if (bsz >= le32_to_cpu(tp->dmt_nleafs)) {
2708 				jfs_err("JFS: block map error in dbBackSplit");
2709 				return -EIO;
2710 			}
2711 
2712 			/* determine the buddy.
2713 			 */
2714 			bud = w ^ bsz;
2715 
2716 			/* check if this buddy is the start of the system.
2717 			 */
2718 			if (leaf[bud] != NOFREE) {
2719 				/* split the leaf at the start of the
2720 				 * system in two.
2721 				 */
2722 				cursz = leaf[bud] - 1;
2723 				dbSplit(tp, bud, cursz, cursz);
2724 				break;
2725 			}
2726 		}
2727 	}
2728 
2729 	if (leaf[leafno] != size) {
2730 		jfs_err("JFS: wrong leaf value in dbBackSplit");
2731 		return -EIO;
2732 	}
2733 	return 0;
2734 }
2735 
2736 
2737 /*
2738  * NAME:	dbJoin()
2739  *
2740  * FUNCTION:	update the leaf of a dmtree with a new value, joining
2741  *		the leaf with other leaves of the dmtree into a multi-leaf
2742  *		binary buddy system, as required.
2743  *
2744  * PARAMETERS:
2745  *	tp	- pointer to the tree containing the leaf.
2746  *	leafno	- the number of the leaf to be updated.
2747  *	newval	- the new value for the leaf.
2748  *
2749  * RETURN VALUES: none
2750  */
2751 static int dbJoin(dmtree_t * tp, int leafno, int newval)
2752 {
2753 	int budsz, buddy;
2754 	s8 *leaf;
2755 
2756 	/* can the new leaf value require a join with other leaves ?
2757 	 */
2758 	if (newval >= tp->dmt_budmin) {
2759 		/* pickup a pointer to the leaves of the tree.
2760 		 */
2761 		leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2762 
2763 		/* try to join the specified leaf into a large binary
2764 		 * buddy system.  the join proceeds by attempting to join
2765 		 * the specified leafno with its buddy (leaf) at new value.
2766 		 * if the join occurs, we attempt to join the left leaf
2767 		 * of the joined buddies with its buddy at new value + 1.
2768 		 * we continue to join until we find a buddy that cannot be
2769 		 * joined (does not have a value equal to the size of the
2770 		 * last join) or until all leaves have been joined into a
2771 		 * single system.
2772 		 *
2773 		 * get the buddy size (number of words covered) of
2774 		 * the new value.
2775 		 */
2776 		budsz = BUDSIZE(newval, tp->dmt_budmin);
2777 
2778 		/* try to join.
2779 		 */
2780 		while (budsz < le32_to_cpu(tp->dmt_nleafs)) {
2781 			/* get the buddy leaf.
2782 			 */
2783 			buddy = leafno ^ budsz;
2784 
2785 			/* if the leaf's new value is greater than its
2786 			 * buddy's value, we join no more.
2787 			 */
2788 			if (newval > leaf[buddy])
2789 				break;
2790 
2791 			/* It shouldn't be less */
2792 			if (newval < leaf[buddy])
2793 				return -EIO;
2794 
2795 			/* check which (leafno or buddy) is the left buddy.
2796 			 * the left buddy gets to claim the blocks resulting
2797 			 * from the join while the right gets to claim none.
2798 			 * the left buddy is also eligible to participate in
2799 			 * a join at the next higher level while the right
2800 			 * is not.
2801 			 *
2802 			 */
2803 			if (leafno < buddy) {
2804 				/* leafno is the left buddy.
2805 				 */
2806 				dbAdjTree(tp, buddy, NOFREE);
2807 			} else {
2808 				/* buddy is the left buddy and becomes
2809 				 * leafno.
2810 				 */
2811 				dbAdjTree(tp, leafno, NOFREE);
2812 				leafno = buddy;
2813 			}
2814 
2815 			/* on to try the next join.
2816 			 */
2817 			newval += 1;
2818 			budsz <<= 1;
2819 		}
2820 	}
2821 
2822 	/* update the leaf value.
2823 	 */
2824 	dbAdjTree(tp, leafno, newval);
2825 
2826 	return 0;
2827 }
2828 
2829 
2830 /*
2831  * NAME:	dbAdjTree()
2832  *
2833  * FUNCTION:	update a leaf of a dmtree with a new value, adjusting
2834  *		the dmtree, as required, to reflect the new leaf value.
2835  *		the combination of any buddies must already be done before
2836  *		this is called.
2837  *
2838  * PARAMETERS:
2839  *	tp	- pointer to the tree to be adjusted.
2840  *	leafno	- the number of the leaf to be updated.
2841  *	newval	- the new value for the leaf.
2842  *
2843  * RETURN VALUES: none
2844  */
2845 static void dbAdjTree(dmtree_t * tp, int leafno, int newval)
2846 {
2847 	int lp, pp, k;
2848 	int max;
2849 
2850 	/* pick up the index of the leaf for this leafno.
2851 	 */
2852 	lp = leafno + le32_to_cpu(tp->dmt_leafidx);
2853 
2854 	/* is the current value the same as the old value ?  if so,
2855 	 * there is nothing to do.
2856 	 */
2857 	if (tp->dmt_stree[lp] == newval)
2858 		return;
2859 
2860 	/* set the new value.
2861 	 */
2862 	tp->dmt_stree[lp] = newval;
2863 
2864 	/* bubble the new value up the tree as required.
2865 	 */
2866 	for (k = 0; k < le32_to_cpu(tp->dmt_height); k++) {
2867 		/* get the index of the first leaf of the 4 leaf
2868 		 * group containing the specified leaf (leafno).
2869 		 */
2870 		lp = ((lp - 1) & ~0x03) + 1;
2871 
2872 		/* get the index of the parent of this 4 leaf group.
2873 		 */
2874 		pp = (lp - 1) >> 2;
2875 
2876 		/* determine the maximum of the 4 leaves.
2877 		 */
2878 		max = TREEMAX(&tp->dmt_stree[lp]);
2879 
2880 		/* if the maximum of the 4 is the same as the
2881 		 * parent's value, we're done.
2882 		 */
2883 		if (tp->dmt_stree[pp] == max)
2884 			break;
2885 
2886 		/* parent gets new value.
2887 		 */
2888 		tp->dmt_stree[pp] = max;
2889 
2890 		/* parent becomes leaf for next go-round.
2891 		 */
2892 		lp = pp;
2893 	}
2894 }
2895 
2896 
2897 /*
2898  * NAME:	dbFindLeaf()
2899  *
2900  * FUNCTION:	search a dmtree_t for sufficient free blocks, returning
2901  *		the index of a leaf describing the free blocks if
2902  *		sufficient free blocks are found.
2903  *
2904  *		the search starts at the top of the dmtree_t tree and
2905  *		proceeds down the tree to the leftmost leaf with sufficient
2906  *		free space.
2907  *
2908  * PARAMETERS:
2909  *	tp	- pointer to the tree to be searched.
2910  *	l2nb	- log2 number of free blocks to search for.
2911  *	leafidx	- return pointer to be set to the index of the leaf
2912  *		  describing at least l2nb free blocks if sufficient
2913  *		  free blocks are found.
2914  *
2915  * RETURN VALUES:
2916  *	0	- success
2917  *	-ENOSPC	- insufficient free blocks.
2918  */
2919 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx)
2920 {
2921 	int ti, n = 0, k, x = 0;
2922 
2923 	/* first check the root of the tree to see if there is
2924 	 * sufficient free space.
2925 	 */
2926 	if (l2nb > tp->dmt_stree[ROOT])
2927 		return -ENOSPC;
2928 
2929 	/* sufficient free space available. now search down the tree
2930 	 * starting at the next level for the leftmost leaf that
2931 	 * describes sufficient free space.
2932 	 */
2933 	for (k = le32_to_cpu(tp->dmt_height), ti = 1;
2934 	     k > 0; k--, ti = ((ti + n) << 2) + 1) {
2935 		/* search the four nodes at this level, starting from
2936 		 * the left.
2937 		 */
2938 		for (x = ti, n = 0; n < 4; n++) {
2939 			/* sufficient free space found.  move to the next
2940 			 * level (or quit if this is the last level).
2941 			 */
2942 			if (l2nb <= tp->dmt_stree[x + n])
2943 				break;
2944 		}
2945 
2946 		/* better have found something since the higher
2947 		 * levels of the tree said it was here.
2948 		 */
2949 		assert(n < 4);
2950 	}
2951 
2952 	/* set the return to the leftmost leaf describing sufficient
2953 	 * free space.
2954 	 */
2955 	*leafidx = x + n - le32_to_cpu(tp->dmt_leafidx);
2956 
2957 	return (0);
2958 }
2959 
2960 
2961 /*
2962  * NAME:	dbFindBits()
2963  *
2964  * FUNCTION:	find a specified number of binary buddy free bits within a
2965  *		dmap bitmap word value.
2966  *
2967  *		this routine searches the bitmap value for (1 << l2nb) free
2968  *		bits at (1 << l2nb) alignments within the value.
2969  *
2970  * PARAMETERS:
2971  *	word	-  dmap bitmap word value.
2972  *	l2nb	-  number of free bits specified as a log2 number.
2973  *
2974  * RETURN VALUES:
2975  *	starting bit number of free bits.
2976  */
2977 static int dbFindBits(u32 word, int l2nb)
2978 {
2979 	int bitno, nb;
2980 	u32 mask;
2981 
2982 	/* get the number of bits.
2983 	 */
2984 	nb = 1 << l2nb;
2985 	assert(nb <= DBWORD);
2986 
2987 	/* complement the word so we can use a mask (i.e. 0s represent
2988 	 * free bits) and compute the mask.
2989 	 */
2990 	word = ~word;
2991 	mask = ONES << (DBWORD - nb);
2992 
2993 	/* scan the word for nb free bits at nb alignments.
2994 	 */
2995 	for (bitno = 0; mask != 0; bitno += nb, mask >>= nb) {
2996 		if ((mask & word) == mask)
2997 			break;
2998 	}
2999 
3000 	ASSERT(bitno < 32);
3001 
3002 	/* return the bit number.
3003 	 */
3004 	return (bitno);
3005 }
3006 
3007 
3008 /*
3009  * NAME:	dbMaxBud(u8 *cp)
3010  *
3011  * FUNCTION:	determine the largest binary buddy string of free
3012  *		bits within 32-bits of the map.
3013  *
3014  * PARAMETERS:
3015  *	cp	-  pointer to the 32-bit value.
3016  *
3017  * RETURN VALUES:
3018  *	largest binary buddy of free bits within a dmap word.
3019  */
3020 static int dbMaxBud(u8 * cp)
3021 {
3022 	signed char tmp1, tmp2;
3023 
3024 	/* check if the wmap word is all free. if so, the
3025 	 * free buddy size is BUDMIN.
3026 	 */
3027 	if (*((uint *) cp) == 0)
3028 		return (BUDMIN);
3029 
3030 	/* check if the wmap word is half free. if so, the
3031 	 * free buddy size is BUDMIN-1.
3032 	 */
3033 	if (*((u16 *) cp) == 0 || *((u16 *) cp + 1) == 0)
3034 		return (BUDMIN - 1);
3035 
3036 	/* not all free or half free. determine the free buddy
3037 	 * size thru table lookup using quarters of the wmap word.
3038 	 */
3039 	tmp1 = max(budtab[cp[2]], budtab[cp[3]]);
3040 	tmp2 = max(budtab[cp[0]], budtab[cp[1]]);
3041 	return (max(tmp1, tmp2));
3042 }
3043 
3044 
3045 /*
3046  * NAME:	cnttz(uint word)
3047  *
3048  * FUNCTION:	determine the number of trailing zeros within a 32-bit
3049  *		value.
3050  *
3051  * PARAMETERS:
3052  *	value	-  32-bit value to be examined.
3053  *
3054  * RETURN VALUES:
3055  *	count of trailing zeros
3056  */
3057 static int cnttz(u32 word)
3058 {
3059 	int n;
3060 
3061 	for (n = 0; n < 32; n++, word >>= 1) {
3062 		if (word & 0x01)
3063 			break;
3064 	}
3065 
3066 	return (n);
3067 }
3068 
3069 
3070 /*
3071  * NAME:	cntlz(u32 value)
3072  *
3073  * FUNCTION:	determine the number of leading zeros within a 32-bit
3074  *		value.
3075  *
3076  * PARAMETERS:
3077  *	value	-  32-bit value to be examined.
3078  *
3079  * RETURN VALUES:
3080  *	count of leading zeros
3081  */
3082 static int cntlz(u32 value)
3083 {
3084 	int n;
3085 
3086 	for (n = 0; n < 32; n++, value <<= 1) {
3087 		if (value & HIGHORDER)
3088 			break;
3089 	}
3090 	return (n);
3091 }
3092 
3093 
3094 /*
3095  * NAME:	blkstol2(s64 nb)
3096  *
3097  * FUNCTION:	convert a block count to its log2 value. if the block
3098  *		count is not a l2 multiple, it is rounded up to the next
3099  *		larger l2 multiple.
3100  *
3101  * PARAMETERS:
3102  *	nb	-  number of blocks
3103  *
3104  * RETURN VALUES:
3105  *	log2 number of blocks
3106  */
3107 static int blkstol2(s64 nb)
3108 {
3109 	int l2nb;
3110 	s64 mask;		/* meant to be signed */
3111 
3112 	mask = (s64) 1 << (64 - 1);
3113 
3114 	/* count the leading bits.
3115 	 */
3116 	for (l2nb = 0; l2nb < 64; l2nb++, mask >>= 1) {
3117 		/* leading bit found.
3118 		 */
3119 		if (nb & mask) {
3120 			/* determine the l2 value.
3121 			 */
3122 			l2nb = (64 - 1) - l2nb;
3123 
3124 			/* check if we need to round up.
3125 			 */
3126 			if (~mask & nb)
3127 				l2nb++;
3128 
3129 			return (l2nb);
3130 		}
3131 	}
3132 	assert(0);
3133 	return 0;		/* fix compiler warning */
3134 }
3135 
3136 
3137 /*
3138  * NAME:	dbAllocBottomUp()
3139  *
3140  * FUNCTION:	alloc the specified block range from the working block
3141  *		allocation map.
3142  *
3143  *		the blocks will be alloc from the working map one dmap
3144  *		at a time.
3145  *
3146  * PARAMETERS:
3147  *	ip	-  pointer to in-core inode;
3148  *	blkno	-  starting block number to be freed.
3149  *	nblocks	-  number of blocks to be freed.
3150  *
3151  * RETURN VALUES:
3152  *	0	- success
3153  *	-EIO	- i/o error
3154  */
3155 int dbAllocBottomUp(struct inode *ip, s64 blkno, s64 nblocks)
3156 {
3157 	struct metapage *mp;
3158 	struct dmap *dp;
3159 	int nb, rc;
3160 	s64 lblkno, rem;
3161 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
3162 	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
3163 
3164 	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
3165 
3166 	/* block to be allocated better be within the mapsize. */
3167 	ASSERT(nblocks <= bmp->db_mapsize - blkno);
3168 
3169 	/*
3170 	 * allocate the blocks a dmap at a time.
3171 	 */
3172 	mp = NULL;
3173 	for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
3174 		/* release previous dmap if any */
3175 		if (mp) {
3176 			write_metapage(mp);
3177 		}
3178 
3179 		/* get the buffer for the current dmap. */
3180 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
3181 		mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
3182 		if (mp == NULL) {
3183 			IREAD_UNLOCK(ipbmap);
3184 			return -EIO;
3185 		}
3186 		dp = (struct dmap *) mp->data;
3187 
3188 		/* determine the number of blocks to be allocated from
3189 		 * this dmap.
3190 		 */
3191 		nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
3192 
3193 		/* allocate the blocks. */
3194 		if ((rc = dbAllocDmapBU(bmp, dp, blkno, nb))) {
3195 			release_metapage(mp);
3196 			IREAD_UNLOCK(ipbmap);
3197 			return (rc);
3198 		}
3199 	}
3200 
3201 	/* write the last buffer. */
3202 	write_metapage(mp);
3203 
3204 	IREAD_UNLOCK(ipbmap);
3205 
3206 	return (0);
3207 }
3208 
3209 
3210 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
3211 			 int nblocks)
3212 {
3213 	int rc;
3214 	int dbitno, word, rembits, nb, nwords, wbitno, agno;
3215 	s8 oldroot;
3216 	struct dmaptree *tp = (struct dmaptree *) & dp->tree;
3217 
3218 	/* save the current value of the root (i.e. maximum free string)
3219 	 * of the dmap tree.
3220 	 */
3221 	oldroot = tp->stree[ROOT];
3222 
3223 	/* determine the bit number and word within the dmap of the
3224 	 * starting block.
3225 	 */
3226 	dbitno = blkno & (BPERDMAP - 1);
3227 	word = dbitno >> L2DBWORD;
3228 
3229 	/* block range better be within the dmap */
3230 	assert(dbitno + nblocks <= BPERDMAP);
3231 
3232 	/* allocate the bits of the dmap's words corresponding to the block
3233 	 * range. not all bits of the first and last words may be contained
3234 	 * within the block range.  if this is the case, we'll work against
3235 	 * those words (i.e. partial first and/or last) on an individual basis
3236 	 * (a single pass), allocating the bits of interest by hand and
3237 	 * updating the leaf corresponding to the dmap word. a single pass
3238 	 * will be used for all dmap words fully contained within the
3239 	 * specified range.  within this pass, the bits of all fully contained
3240 	 * dmap words will be marked as free in a single shot and the leaves
3241 	 * will be updated. a single leaf may describe the free space of
3242 	 * multiple dmap words, so we may update only a subset of the actual
3243 	 * leaves corresponding to the dmap words of the block range.
3244 	 */
3245 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
3246 		/* determine the bit number within the word and
3247 		 * the number of bits within the word.
3248 		 */
3249 		wbitno = dbitno & (DBWORD - 1);
3250 		nb = min(rembits, DBWORD - wbitno);
3251 
3252 		/* check if only part of a word is to be allocated.
3253 		 */
3254 		if (nb < DBWORD) {
3255 			/* allocate (set to 1) the appropriate bits within
3256 			 * this dmap word.
3257 			 */
3258 			dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
3259 						      >> wbitno);
3260 
3261 			word++;
3262 		} else {
3263 			/* one or more dmap words are fully contained
3264 			 * within the block range.  determine how many
3265 			 * words and allocate (set to 1) the bits of these
3266 			 * words.
3267 			 */
3268 			nwords = rembits >> L2DBWORD;
3269 			memset(&dp->wmap[word], (int) ONES, nwords * 4);
3270 
3271 			/* determine how many bits */
3272 			nb = nwords << L2DBWORD;
3273 			word += nwords;
3274 		}
3275 	}
3276 
3277 	/* update the free count for this dmap */
3278 	le32_add_cpu(&dp->nfree, -nblocks);
3279 
3280 	/* reconstruct summary tree */
3281 	dbInitDmapTree(dp);
3282 
3283 	BMAP_LOCK(bmp);
3284 
3285 	/* if this allocation group is completely free,
3286 	 * update the highest active allocation group number
3287 	 * if this allocation group is the new max.
3288 	 */
3289 	agno = blkno >> bmp->db_agl2size;
3290 	if (agno > bmp->db_maxag)
3291 		bmp->db_maxag = agno;
3292 
3293 	/* update the free count for the allocation group and map */
3294 	bmp->db_agfree[agno] -= nblocks;
3295 	bmp->db_nfree -= nblocks;
3296 
3297 	BMAP_UNLOCK(bmp);
3298 
3299 	/* if the root has not changed, done. */
3300 	if (tp->stree[ROOT] == oldroot)
3301 		return (0);
3302 
3303 	/* root changed. bubble the change up to the dmap control pages.
3304 	 * if the adjustment of the upper level control pages fails,
3305 	 * backout the bit allocation (thus making everything consistent).
3306 	 */
3307 	if ((rc = dbAdjCtl(bmp, blkno, tp->stree[ROOT], 1, 0)))
3308 		dbFreeBits(bmp, dp, blkno, nblocks);
3309 
3310 	return (rc);
3311 }
3312 
3313 
3314 /*
3315  * NAME:	dbExtendFS()
3316  *
3317  * FUNCTION:	extend bmap from blkno for nblocks;
3318  *		dbExtendFS() updates bmap ready for dbAllocBottomUp();
3319  *
3320  * L2
3321  *  |
3322  *   L1---------------------------------L1
3323  *    |					 |
3324  *     L0---------L0---------L0		  L0---------L0---------L0
3325  *      |	   |	      |		   |	      |		 |
3326  *	 d0,...,dn  d0,...,dn  d0,...,dn    d0,...,dn  d0,...,dn  d0,.,dm;
3327  * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm
3328  *
3329  * <---old---><----------------------------extend----------------------->
3330  */
3331 int dbExtendFS(struct inode *ipbmap, s64 blkno,	s64 nblocks)
3332 {
3333 	struct jfs_sb_info *sbi = JFS_SBI(ipbmap->i_sb);
3334 	int nbperpage = sbi->nbperpage;
3335 	int i, i0 = true, j, j0 = true, k, n;
3336 	s64 newsize;
3337 	s64 p;
3338 	struct metapage *mp, *l2mp, *l1mp = NULL, *l0mp = NULL;
3339 	struct dmapctl *l2dcp, *l1dcp, *l0dcp;
3340 	struct dmap *dp;
3341 	s8 *l0leaf, *l1leaf, *l2leaf;
3342 	struct bmap *bmp = sbi->bmap;
3343 	int agno, l2agsize, oldl2agsize;
3344 	s64 ag_rem;
3345 
3346 	newsize = blkno + nblocks;
3347 
3348 	jfs_info("dbExtendFS: blkno:%Ld nblocks:%Ld newsize:%Ld",
3349 		 (long long) blkno, (long long) nblocks, (long long) newsize);
3350 
3351 	/*
3352 	 *	initialize bmap control page.
3353 	 *
3354 	 * all the data in bmap control page should exclude
3355 	 * the mkfs hidden dmap page.
3356 	 */
3357 
3358 	/* update mapsize */
3359 	bmp->db_mapsize = newsize;
3360 	bmp->db_maxlevel = BMAPSZTOLEV(bmp->db_mapsize);
3361 
3362 	/* compute new AG size */
3363 	l2agsize = dbGetL2AGSize(newsize);
3364 	oldl2agsize = bmp->db_agl2size;
3365 
3366 	bmp->db_agl2size = l2agsize;
3367 	bmp->db_agsize = 1 << l2agsize;
3368 
3369 	/* compute new number of AG */
3370 	agno = bmp->db_numag;
3371 	bmp->db_numag = newsize >> l2agsize;
3372 	bmp->db_numag += ((u32) newsize % (u32) bmp->db_agsize) ? 1 : 0;
3373 
3374 	/*
3375 	 *	reconfigure db_agfree[]
3376 	 * from old AG configuration to new AG configuration;
3377 	 *
3378 	 * coalesce contiguous k (newAGSize/oldAGSize) AGs;
3379 	 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn;
3380 	 * note: new AG size = old AG size * (2**x).
3381 	 */
3382 	if (l2agsize == oldl2agsize)
3383 		goto extend;
3384 	k = 1 << (l2agsize - oldl2agsize);
3385 	ag_rem = bmp->db_agfree[0];	/* save agfree[0] */
3386 	for (i = 0, n = 0; i < agno; n++) {
3387 		bmp->db_agfree[n] = 0;	/* init collection point */
3388 
3389 		/* coalesce contiguous k AGs; */
3390 		for (j = 0; j < k && i < agno; j++, i++) {
3391 			/* merge AGi to AGn */
3392 			bmp->db_agfree[n] += bmp->db_agfree[i];
3393 		}
3394 	}
3395 	bmp->db_agfree[0] += ag_rem;	/* restore agfree[0] */
3396 
3397 	for (; n < MAXAG; n++)
3398 		bmp->db_agfree[n] = 0;
3399 
3400 	/*
3401 	 * update highest active ag number
3402 	 */
3403 
3404 	bmp->db_maxag = bmp->db_maxag / k;
3405 
3406 	/*
3407 	 *	extend bmap
3408 	 *
3409 	 * update bit maps and corresponding level control pages;
3410 	 * global control page db_nfree, db_agfree[agno], db_maxfreebud;
3411 	 */
3412       extend:
3413 	/* get L2 page */
3414 	p = BMAPBLKNO + nbperpage;	/* L2 page */
3415 	l2mp = read_metapage(ipbmap, p, PSIZE, 0);
3416 	if (!l2mp) {
3417 		jfs_error(ipbmap->i_sb, "L2 page could not be read\n");
3418 		return -EIO;
3419 	}
3420 	l2dcp = (struct dmapctl *) l2mp->data;
3421 
3422 	/* compute start L1 */
3423 	k = blkno >> L2MAXL1SIZE;
3424 	l2leaf = l2dcp->stree + CTLLEAFIND + k;
3425 	p = BLKTOL1(blkno, sbi->l2nbperpage);	/* L1 page */
3426 
3427 	/*
3428 	 * extend each L1 in L2
3429 	 */
3430 	for (; k < LPERCTL; k++, p += nbperpage) {
3431 		/* get L1 page */
3432 		if (j0) {
3433 			/* read in L1 page: (blkno & (MAXL1SIZE - 1)) */
3434 			l1mp = read_metapage(ipbmap, p, PSIZE, 0);
3435 			if (l1mp == NULL)
3436 				goto errout;
3437 			l1dcp = (struct dmapctl *) l1mp->data;
3438 
3439 			/* compute start L0 */
3440 			j = (blkno & (MAXL1SIZE - 1)) >> L2MAXL0SIZE;
3441 			l1leaf = l1dcp->stree + CTLLEAFIND + j;
3442 			p = BLKTOL0(blkno, sbi->l2nbperpage);
3443 			j0 = false;
3444 		} else {
3445 			/* assign/init L1 page */
3446 			l1mp = get_metapage(ipbmap, p, PSIZE, 0);
3447 			if (l1mp == NULL)
3448 				goto errout;
3449 
3450 			l1dcp = (struct dmapctl *) l1mp->data;
3451 
3452 			/* compute start L0 */
3453 			j = 0;
3454 			l1leaf = l1dcp->stree + CTLLEAFIND;
3455 			p += nbperpage;	/* 1st L0 of L1.k */
3456 		}
3457 
3458 		/*
3459 		 * extend each L0 in L1
3460 		 */
3461 		for (; j < LPERCTL; j++) {
3462 			/* get L0 page */
3463 			if (i0) {
3464 				/* read in L0 page: (blkno & (MAXL0SIZE - 1)) */
3465 
3466 				l0mp = read_metapage(ipbmap, p, PSIZE, 0);
3467 				if (l0mp == NULL)
3468 					goto errout;
3469 				l0dcp = (struct dmapctl *) l0mp->data;
3470 
3471 				/* compute start dmap */
3472 				i = (blkno & (MAXL0SIZE - 1)) >>
3473 				    L2BPERDMAP;
3474 				l0leaf = l0dcp->stree + CTLLEAFIND + i;
3475 				p = BLKTODMAP(blkno,
3476 					      sbi->l2nbperpage);
3477 				i0 = false;
3478 			} else {
3479 				/* assign/init L0 page */
3480 				l0mp = get_metapage(ipbmap, p, PSIZE, 0);
3481 				if (l0mp == NULL)
3482 					goto errout;
3483 
3484 				l0dcp = (struct dmapctl *) l0mp->data;
3485 
3486 				/* compute start dmap */
3487 				i = 0;
3488 				l0leaf = l0dcp->stree + CTLLEAFIND;
3489 				p += nbperpage;	/* 1st dmap of L0.j */
3490 			}
3491 
3492 			/*
3493 			 * extend each dmap in L0
3494 			 */
3495 			for (; i < LPERCTL; i++) {
3496 				/*
3497 				 * reconstruct the dmap page, and
3498 				 * initialize corresponding parent L0 leaf
3499 				 */
3500 				if ((n = blkno & (BPERDMAP - 1))) {
3501 					/* read in dmap page: */
3502 					mp = read_metapage(ipbmap, p,
3503 							   PSIZE, 0);
3504 					if (mp == NULL)
3505 						goto errout;
3506 					n = min(nblocks, (s64)BPERDMAP - n);
3507 				} else {
3508 					/* assign/init dmap page */
3509 					mp = read_metapage(ipbmap, p,
3510 							   PSIZE, 0);
3511 					if (mp == NULL)
3512 						goto errout;
3513 
3514 					n = min_t(s64, nblocks, BPERDMAP);
3515 				}
3516 
3517 				dp = (struct dmap *) mp->data;
3518 				*l0leaf = dbInitDmap(dp, blkno, n);
3519 
3520 				bmp->db_nfree += n;
3521 				agno = le64_to_cpu(dp->start) >> l2agsize;
3522 				bmp->db_agfree[agno] += n;
3523 
3524 				write_metapage(mp);
3525 
3526 				l0leaf++;
3527 				p += nbperpage;
3528 
3529 				blkno += n;
3530 				nblocks -= n;
3531 				if (nblocks == 0)
3532 					break;
3533 			}	/* for each dmap in a L0 */
3534 
3535 			/*
3536 			 * build current L0 page from its leaves, and
3537 			 * initialize corresponding parent L1 leaf
3538 			 */
3539 			*l1leaf = dbInitDmapCtl(l0dcp, 0, ++i);
3540 			write_metapage(l0mp);
3541 			l0mp = NULL;
3542 
3543 			if (nblocks)
3544 				l1leaf++;	/* continue for next L0 */
3545 			else {
3546 				/* more than 1 L0 ? */
3547 				if (j > 0)
3548 					break;	/* build L1 page */
3549 				else {
3550 					/* summarize in global bmap page */
3551 					bmp->db_maxfreebud = *l1leaf;
3552 					release_metapage(l1mp);
3553 					release_metapage(l2mp);
3554 					goto finalize;
3555 				}
3556 			}
3557 		}		/* for each L0 in a L1 */
3558 
3559 		/*
3560 		 * build current L1 page from its leaves, and
3561 		 * initialize corresponding parent L2 leaf
3562 		 */
3563 		*l2leaf = dbInitDmapCtl(l1dcp, 1, ++j);
3564 		write_metapage(l1mp);
3565 		l1mp = NULL;
3566 
3567 		if (nblocks)
3568 			l2leaf++;	/* continue for next L1 */
3569 		else {
3570 			/* more than 1 L1 ? */
3571 			if (k > 0)
3572 				break;	/* build L2 page */
3573 			else {
3574 				/* summarize in global bmap page */
3575 				bmp->db_maxfreebud = *l2leaf;
3576 				release_metapage(l2mp);
3577 				goto finalize;
3578 			}
3579 		}
3580 	}			/* for each L1 in a L2 */
3581 
3582 	jfs_error(ipbmap->i_sb, "function has not returned as expected\n");
3583 errout:
3584 	if (l0mp)
3585 		release_metapage(l0mp);
3586 	if (l1mp)
3587 		release_metapage(l1mp);
3588 	release_metapage(l2mp);
3589 	return -EIO;
3590 
3591 	/*
3592 	 *	finalize bmap control page
3593 	 */
3594 finalize:
3595 
3596 	return 0;
3597 }
3598 
3599 
3600 /*
3601  *	dbFinalizeBmap()
3602  */
3603 void dbFinalizeBmap(struct inode *ipbmap)
3604 {
3605 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
3606 	int actags, inactags, l2nl;
3607 	s64 ag_rem, actfree, inactfree, avgfree;
3608 	int i, n;
3609 
3610 	/*
3611 	 *	finalize bmap control page
3612 	 */
3613 //finalize:
3614 	/*
3615 	 * compute db_agpref: preferred ag to allocate from
3616 	 * (the leftmost ag with average free space in it);
3617 	 */
3618 //agpref:
3619 	/* get the number of active ags and inactive ags */
3620 	actags = bmp->db_maxag + 1;
3621 	inactags = bmp->db_numag - actags;
3622 	ag_rem = bmp->db_mapsize & (bmp->db_agsize - 1);	/* ??? */
3623 
3624 	/* determine how many blocks are in the inactive allocation
3625 	 * groups. in doing this, we must account for the fact that
3626 	 * the rightmost group might be a partial group (i.e. file
3627 	 * system size is not a multiple of the group size).
3628 	 */
3629 	inactfree = (inactags && ag_rem) ?
3630 	    ((inactags - 1) << bmp->db_agl2size) + ag_rem
3631 	    : inactags << bmp->db_agl2size;
3632 
3633 	/* determine how many free blocks are in the active
3634 	 * allocation groups plus the average number of free blocks
3635 	 * within the active ags.
3636 	 */
3637 	actfree = bmp->db_nfree - inactfree;
3638 	avgfree = (u32) actfree / (u32) actags;
3639 
3640 	/* if the preferred allocation group has not average free space.
3641 	 * re-establish the preferred group as the leftmost
3642 	 * group with average free space.
3643 	 */
3644 	if (bmp->db_agfree[bmp->db_agpref] < avgfree) {
3645 		for (bmp->db_agpref = 0; bmp->db_agpref < actags;
3646 		     bmp->db_agpref++) {
3647 			if (bmp->db_agfree[bmp->db_agpref] >= avgfree)
3648 				break;
3649 		}
3650 		if (bmp->db_agpref >= bmp->db_numag) {
3651 			jfs_error(ipbmap->i_sb,
3652 				  "cannot find ag with average freespace\n");
3653 		}
3654 	}
3655 
3656 	/*
3657 	 * compute db_aglevel, db_agheight, db_width, db_agstart:
3658 	 * an ag is covered in aglevel dmapctl summary tree,
3659 	 * at agheight level height (from leaf) with agwidth number of nodes
3660 	 * each, which starts at agstart index node of the smmary tree node
3661 	 * array;
3662 	 */
3663 	bmp->db_aglevel = BMAPSZTOLEV(bmp->db_agsize);
3664 	l2nl =
3665 	    bmp->db_agl2size - (L2BPERDMAP + bmp->db_aglevel * L2LPERCTL);
3666 	bmp->db_agheight = l2nl >> 1;
3667 	bmp->db_agwidth = 1 << (l2nl - (bmp->db_agheight << 1));
3668 	for (i = 5 - bmp->db_agheight, bmp->db_agstart = 0, n = 1; i > 0;
3669 	     i--) {
3670 		bmp->db_agstart += n;
3671 		n <<= 2;
3672 	}
3673 
3674 }
3675 
3676 
3677 /*
3678  * NAME:	dbInitDmap()/ujfs_idmap_page()
3679  *
3680  * FUNCTION:	initialize working/persistent bitmap of the dmap page
3681  *		for the specified number of blocks:
3682  *
3683  *		at entry, the bitmaps had been initialized as free (ZEROS);
3684  *		The number of blocks will only account for the actually
3685  *		existing blocks. Blocks which don't actually exist in
3686  *		the aggregate will be marked as allocated (ONES);
3687  *
3688  * PARAMETERS:
3689  *	dp	- pointer to page of map
3690  *	nblocks	- number of blocks this page
3691  *
3692  * RETURNS: NONE
3693  */
3694 static int dbInitDmap(struct dmap * dp, s64 Blkno, int nblocks)
3695 {
3696 	int blkno, w, b, r, nw, nb, i;
3697 
3698 	/* starting block number within the dmap */
3699 	blkno = Blkno & (BPERDMAP - 1);
3700 
3701 	if (blkno == 0) {
3702 		dp->nblocks = dp->nfree = cpu_to_le32(nblocks);
3703 		dp->start = cpu_to_le64(Blkno);
3704 
3705 		if (nblocks == BPERDMAP) {
3706 			memset(&dp->wmap[0], 0, LPERDMAP * 4);
3707 			memset(&dp->pmap[0], 0, LPERDMAP * 4);
3708 			goto initTree;
3709 		}
3710 	} else {
3711 		le32_add_cpu(&dp->nblocks, nblocks);
3712 		le32_add_cpu(&dp->nfree, nblocks);
3713 	}
3714 
3715 	/* word number containing start block number */
3716 	w = blkno >> L2DBWORD;
3717 
3718 	/*
3719 	 * free the bits corresponding to the block range (ZEROS):
3720 	 * note: not all bits of the first and last words may be contained
3721 	 * within the block range.
3722 	 */
3723 	for (r = nblocks; r > 0; r -= nb, blkno += nb) {
3724 		/* number of bits preceding range to be freed in the word */
3725 		b = blkno & (DBWORD - 1);
3726 		/* number of bits to free in the word */
3727 		nb = min(r, DBWORD - b);
3728 
3729 		/* is partial word to be freed ? */
3730 		if (nb < DBWORD) {
3731 			/* free (set to 0) from the bitmap word */
3732 			dp->wmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3733 						     >> b));
3734 			dp->pmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3735 						     >> b));
3736 
3737 			/* skip the word freed */
3738 			w++;
3739 		} else {
3740 			/* free (set to 0) contiguous bitmap words */
3741 			nw = r >> L2DBWORD;
3742 			memset(&dp->wmap[w], 0, nw * 4);
3743 			memset(&dp->pmap[w], 0, nw * 4);
3744 
3745 			/* skip the words freed */
3746 			nb = nw << L2DBWORD;
3747 			w += nw;
3748 		}
3749 	}
3750 
3751 	/*
3752 	 * mark bits following the range to be freed (non-existing
3753 	 * blocks) as allocated (ONES)
3754 	 */
3755 
3756 	if (blkno == BPERDMAP)
3757 		goto initTree;
3758 
3759 	/* the first word beyond the end of existing blocks */
3760 	w = blkno >> L2DBWORD;
3761 
3762 	/* does nblocks fall on a 32-bit boundary ? */
3763 	b = blkno & (DBWORD - 1);
3764 	if (b) {
3765 		/* mark a partial word allocated */
3766 		dp->wmap[w] = dp->pmap[w] = cpu_to_le32(ONES >> b);
3767 		w++;
3768 	}
3769 
3770 	/* set the rest of the words in the page to allocated (ONES) */
3771 	for (i = w; i < LPERDMAP; i++)
3772 		dp->pmap[i] = dp->wmap[i] = cpu_to_le32(ONES);
3773 
3774 	/*
3775 	 * init tree
3776 	 */
3777       initTree:
3778 	return (dbInitDmapTree(dp));
3779 }
3780 
3781 
3782 /*
3783  * NAME:	dbInitDmapTree()/ujfs_complete_dmap()
3784  *
3785  * FUNCTION:	initialize summary tree of the specified dmap:
3786  *
3787  *		at entry, bitmap of the dmap has been initialized;
3788  *
3789  * PARAMETERS:
3790  *	dp	- dmap to complete
3791  *	blkno	- starting block number for this dmap
3792  *	treemax	- will be filled in with max free for this dmap
3793  *
3794  * RETURNS:	max free string at the root of the tree
3795  */
3796 static int dbInitDmapTree(struct dmap * dp)
3797 {
3798 	struct dmaptree *tp;
3799 	s8 *cp;
3800 	int i;
3801 
3802 	/* init fixed info of tree */
3803 	tp = &dp->tree;
3804 	tp->nleafs = cpu_to_le32(LPERDMAP);
3805 	tp->l2nleafs = cpu_to_le32(L2LPERDMAP);
3806 	tp->leafidx = cpu_to_le32(LEAFIND);
3807 	tp->height = cpu_to_le32(4);
3808 	tp->budmin = BUDMIN;
3809 
3810 	/* init each leaf from corresponding wmap word:
3811 	 * note: leaf is set to NOFREE(-1) if all blocks of corresponding
3812 	 * bitmap word are allocated.
3813 	 */
3814 	cp = tp->stree + le32_to_cpu(tp->leafidx);
3815 	for (i = 0; i < LPERDMAP; i++)
3816 		*cp++ = dbMaxBud((u8 *) & dp->wmap[i]);
3817 
3818 	/* build the dmap's binary buddy summary tree */
3819 	return (dbInitTree(tp));
3820 }
3821 
3822 
3823 /*
3824  * NAME:	dbInitTree()/ujfs_adjtree()
3825  *
3826  * FUNCTION:	initialize binary buddy summary tree of a dmap or dmapctl.
3827  *
3828  *		at entry, the leaves of the tree has been initialized
3829  *		from corresponding bitmap word or root of summary tree
3830  *		of the child control page;
3831  *		configure binary buddy system at the leaf level, then
3832  *		bubble up the values of the leaf nodes up the tree.
3833  *
3834  * PARAMETERS:
3835  *	cp	- Pointer to the root of the tree
3836  *	l2leaves- Number of leaf nodes as a power of 2
3837  *	l2min	- Number of blocks that can be covered by a leaf
3838  *		  as a power of 2
3839  *
3840  * RETURNS: max free string at the root of the tree
3841  */
3842 static int dbInitTree(struct dmaptree * dtp)
3843 {
3844 	int l2max, l2free, bsize, nextb, i;
3845 	int child, parent, nparent;
3846 	s8 *tp, *cp, *cp1;
3847 
3848 	tp = dtp->stree;
3849 
3850 	/* Determine the maximum free string possible for the leaves */
3851 	l2max = le32_to_cpu(dtp->l2nleafs) + dtp->budmin;
3852 
3853 	/*
3854 	 * configure the leaf levevl into binary buddy system
3855 	 *
3856 	 * Try to combine buddies starting with a buddy size of 1
3857 	 * (i.e. two leaves). At a buddy size of 1 two buddy leaves
3858 	 * can be combined if both buddies have a maximum free of l2min;
3859 	 * the combination will result in the left-most buddy leaf having
3860 	 * a maximum free of l2min+1.
3861 	 * After processing all buddies for a given size, process buddies
3862 	 * at the next higher buddy size (i.e. current size * 2) and
3863 	 * the next maximum free (current free + 1).
3864 	 * This continues until the maximum possible buddy combination
3865 	 * yields maximum free.
3866 	 */
3867 	for (l2free = dtp->budmin, bsize = 1; l2free < l2max;
3868 	     l2free++, bsize = nextb) {
3869 		/* get next buddy size == current buddy pair size */
3870 		nextb = bsize << 1;
3871 
3872 		/* scan each adjacent buddy pair at current buddy size */
3873 		for (i = 0, cp = tp + le32_to_cpu(dtp->leafidx);
3874 		     i < le32_to_cpu(dtp->nleafs);
3875 		     i += nextb, cp += nextb) {
3876 			/* coalesce if both adjacent buddies are max free */
3877 			if (*cp == l2free && *(cp + bsize) == l2free) {
3878 				*cp = l2free + 1;	/* left take right */
3879 				*(cp + bsize) = -1;	/* right give left */
3880 			}
3881 		}
3882 	}
3883 
3884 	/*
3885 	 * bubble summary information of leaves up the tree.
3886 	 *
3887 	 * Starting at the leaf node level, the four nodes described by
3888 	 * the higher level parent node are compared for a maximum free and
3889 	 * this maximum becomes the value of the parent node.
3890 	 * when all lower level nodes are processed in this fashion then
3891 	 * move up to the next level (parent becomes a lower level node) and
3892 	 * continue the process for that level.
3893 	 */
3894 	for (child = le32_to_cpu(dtp->leafidx),
3895 	     nparent = le32_to_cpu(dtp->nleafs) >> 2;
3896 	     nparent > 0; nparent >>= 2, child = parent) {
3897 		/* get index of 1st node of parent level */
3898 		parent = (child - 1) >> 2;
3899 
3900 		/* set the value of the parent node as the maximum
3901 		 * of the four nodes of the current level.
3902 		 */
3903 		for (i = 0, cp = tp + child, cp1 = tp + parent;
3904 		     i < nparent; i++, cp += 4, cp1++)
3905 			*cp1 = TREEMAX(cp);
3906 	}
3907 
3908 	return (*tp);
3909 }
3910 
3911 
3912 /*
3913  *	dbInitDmapCtl()
3914  *
3915  * function: initialize dmapctl page
3916  */
3917 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i)
3918 {				/* start leaf index not covered by range */
3919 	s8 *cp;
3920 
3921 	dcp->nleafs = cpu_to_le32(LPERCTL);
3922 	dcp->l2nleafs = cpu_to_le32(L2LPERCTL);
3923 	dcp->leafidx = cpu_to_le32(CTLLEAFIND);
3924 	dcp->height = cpu_to_le32(5);
3925 	dcp->budmin = L2BPERDMAP + L2LPERCTL * level;
3926 
3927 	/*
3928 	 * initialize the leaves of current level that were not covered
3929 	 * by the specified input block range (i.e. the leaves have no
3930 	 * low level dmapctl or dmap).
3931 	 */
3932 	cp = &dcp->stree[CTLLEAFIND + i];
3933 	for (; i < LPERCTL; i++)
3934 		*cp++ = NOFREE;
3935 
3936 	/* build the dmap's binary buddy summary tree */
3937 	return (dbInitTree((struct dmaptree *) dcp));
3938 }
3939 
3940 
3941 /*
3942  * NAME:	dbGetL2AGSize()/ujfs_getagl2size()
3943  *
3944  * FUNCTION:	Determine log2(allocation group size) from aggregate size
3945  *
3946  * PARAMETERS:
3947  *	nblocks	- Number of blocks in aggregate
3948  *
3949  * RETURNS: log2(allocation group size) in aggregate blocks
3950  */
3951 static int dbGetL2AGSize(s64 nblocks)
3952 {
3953 	s64 sz;
3954 	s64 m;
3955 	int l2sz;
3956 
3957 	if (nblocks < BPERDMAP * MAXAG)
3958 		return (L2BPERDMAP);
3959 
3960 	/* round up aggregate size to power of 2 */
3961 	m = ((u64) 1 << (64 - 1));
3962 	for (l2sz = 64; l2sz >= 0; l2sz--, m >>= 1) {
3963 		if (m & nblocks)
3964 			break;
3965 	}
3966 
3967 	sz = (s64) 1 << l2sz;
3968 	if (sz < nblocks)
3969 		l2sz += 1;
3970 
3971 	/* agsize = roundupSize/max_number_of_ag */
3972 	return (l2sz - L2MAXAG);
3973 }
3974 
3975 
3976 /*
3977  * NAME:	dbMapFileSizeToMapSize()
3978  *
3979  * FUNCTION:	compute number of blocks the block allocation map file
3980  *		can cover from the map file size;
3981  *
3982  * RETURNS:	Number of blocks which can be covered by this block map file;
3983  */
3984 
3985 /*
3986  * maximum number of map pages at each level including control pages
3987  */
3988 #define MAXL0PAGES	(1 + LPERCTL)
3989 #define MAXL1PAGES	(1 + LPERCTL * MAXL0PAGES)
3990 
3991 /*
3992  * convert number of map pages to the zero origin top dmapctl level
3993  */
3994 #define BMAPPGTOLEV(npages)	\
3995 	(((npages) <= 3 + MAXL0PAGES) ? 0 : \
3996 	 ((npages) <= 2 + MAXL1PAGES) ? 1 : 2)
3997 
3998 s64 dbMapFileSizeToMapSize(struct inode * ipbmap)
3999 {
4000 	struct super_block *sb = ipbmap->i_sb;
4001 	s64 nblocks;
4002 	s64 npages, ndmaps;
4003 	int level, i;
4004 	int complete, factor;
4005 
4006 	nblocks = ipbmap->i_size >> JFS_SBI(sb)->l2bsize;
4007 	npages = nblocks >> JFS_SBI(sb)->l2nbperpage;
4008 	level = BMAPPGTOLEV(npages);
4009 
4010 	/* At each level, accumulate the number of dmap pages covered by
4011 	 * the number of full child levels below it;
4012 	 * repeat for the last incomplete child level.
4013 	 */
4014 	ndmaps = 0;
4015 	npages--;		/* skip the first global control page */
4016 	/* skip higher level control pages above top level covered by map */
4017 	npages -= (2 - level);
4018 	npages--;		/* skip top level's control page */
4019 	for (i = level; i >= 0; i--) {
4020 		factor =
4021 		    (i == 2) ? MAXL1PAGES : ((i == 1) ? MAXL0PAGES : 1);
4022 		complete = (u32) npages / factor;
4023 		ndmaps += complete * ((i == 2) ? LPERCTL * LPERCTL :
4024 				      ((i == 1) ? LPERCTL : 1));
4025 
4026 		/* pages in last/incomplete child */
4027 		npages = (u32) npages % factor;
4028 		/* skip incomplete child's level control page */
4029 		npages--;
4030 	}
4031 
4032 	/* convert the number of dmaps into the number of blocks
4033 	 * which can be covered by the dmaps;
4034 	 */
4035 	nblocks = ndmaps << L2BPERDMAP;
4036 
4037 	return (nblocks);
4038 }
4039