xref: /openbmc/linux/fs/jfs/jfs_dmap.c (revision 293d5b43)
1 /*
2  *   Copyright (C) International Business Machines Corp., 2000-2004
3  *   Portions Copyright (C) Tino Reichardt, 2012
4  *
5  *   This program is free software;  you can redistribute it and/or modify
6  *   it under the terms of the GNU General Public License as published by
7  *   the Free Software Foundation; either version 2 of the License, or
8  *   (at your option) any later version.
9  *
10  *   This program is distributed in the hope that it will be useful,
11  *   but WITHOUT ANY WARRANTY;  without even the implied warranty of
12  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See
13  *   the GNU General Public License for more details.
14  *
15  *   You should have received a copy of the GNU General Public License
16  *   along with this program;  if not, write to the Free Software
17  *   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18  */
19 
20 #include <linux/fs.h>
21 #include <linux/slab.h>
22 #include "jfs_incore.h"
23 #include "jfs_superblock.h"
24 #include "jfs_dmap.h"
25 #include "jfs_imap.h"
26 #include "jfs_lock.h"
27 #include "jfs_metapage.h"
28 #include "jfs_debug.h"
29 #include "jfs_discard.h"
30 
31 /*
32  *	SERIALIZATION of the Block Allocation Map.
33  *
34  *	the working state of the block allocation map is accessed in
35  *	two directions:
36  *
37  *	1) allocation and free requests that start at the dmap
38  *	   level and move up through the dmap control pages (i.e.
39  *	   the vast majority of requests).
40  *
41  *	2) allocation requests that start at dmap control page
42  *	   level and work down towards the dmaps.
43  *
44  *	the serialization scheme used here is as follows.
45  *
46  *	requests which start at the bottom are serialized against each
47  *	other through buffers and each requests holds onto its buffers
48  *	as it works it way up from a single dmap to the required level
49  *	of dmap control page.
50  *	requests that start at the top are serialized against each other
51  *	and request that start from the bottom by the multiple read/single
52  *	write inode lock of the bmap inode. requests starting at the top
53  *	take this lock in write mode while request starting at the bottom
54  *	take the lock in read mode.  a single top-down request may proceed
55  *	exclusively while multiple bottoms-up requests may proceed
56  *	simultaneously (under the protection of busy buffers).
57  *
58  *	in addition to information found in dmaps and dmap control pages,
59  *	the working state of the block allocation map also includes read/
60  *	write information maintained in the bmap descriptor (i.e. total
61  *	free block count, allocation group level free block counts).
62  *	a single exclusive lock (BMAP_LOCK) is used to guard this information
63  *	in the face of multiple-bottoms up requests.
64  *	(lock ordering: IREAD_LOCK, BMAP_LOCK);
65  *
66  *	accesses to the persistent state of the block allocation map (limited
67  *	to the persistent bitmaps in dmaps) is guarded by (busy) buffers.
68  */
69 
70 #define BMAP_LOCK_INIT(bmp)	mutex_init(&bmp->db_bmaplock)
71 #define BMAP_LOCK(bmp)		mutex_lock(&bmp->db_bmaplock)
72 #define BMAP_UNLOCK(bmp)	mutex_unlock(&bmp->db_bmaplock)
73 
74 /*
75  * forward references
76  */
77 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
78 			int nblocks);
79 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval);
80 static int dbBackSplit(dmtree_t * tp, int leafno);
81 static int dbJoin(dmtree_t * tp, int leafno, int newval);
82 static void dbAdjTree(dmtree_t * tp, int leafno, int newval);
83 static int dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc,
84 		    int level);
85 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results);
86 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
87 		       int nblocks);
88 static int dbAllocNear(struct bmap * bmp, struct dmap * dp, s64 blkno,
89 		       int nblocks,
90 		       int l2nb, s64 * results);
91 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
92 		       int nblocks);
93 static int dbAllocDmapLev(struct bmap * bmp, struct dmap * dp, int nblocks,
94 			  int l2nb,
95 			  s64 * results);
96 static int dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb,
97 		     s64 * results);
98 static int dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno,
99 		      s64 * results);
100 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks);
101 static int dbFindBits(u32 word, int l2nb);
102 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno);
103 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx);
104 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
105 		      int nblocks);
106 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
107 		      int nblocks);
108 static int dbMaxBud(u8 * cp);
109 static int blkstol2(s64 nb);
110 
111 static int cntlz(u32 value);
112 static int cnttz(u32 word);
113 
114 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
115 			 int nblocks);
116 static int dbInitDmap(struct dmap * dp, s64 blkno, int nblocks);
117 static int dbInitDmapTree(struct dmap * dp);
118 static int dbInitTree(struct dmaptree * dtp);
119 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i);
120 static int dbGetL2AGSize(s64 nblocks);
121 
122 /*
123  *	buddy table
124  *
125  * table used for determining buddy sizes within characters of
126  * dmap bitmap words.  the characters themselves serve as indexes
127  * into the table, with the table elements yielding the maximum
128  * binary buddy of free bits within the character.
129  */
130 static const s8 budtab[256] = {
131 	3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
132 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
133 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
134 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
135 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
136 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
137 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
138 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
139 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
140 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
141 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
142 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
143 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
144 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
145 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
146 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, -1
147 };
148 
149 /*
150  * NAME:	dbMount()
151  *
152  * FUNCTION:	initializate the block allocation map.
153  *
154  *		memory is allocated for the in-core bmap descriptor and
155  *		the in-core descriptor is initialized from disk.
156  *
157  * PARAMETERS:
158  *	ipbmap	- pointer to in-core inode for the block map.
159  *
160  * RETURN VALUES:
161  *	0	- success
162  *	-ENOMEM	- insufficient memory
163  *	-EIO	- i/o error
164  */
165 int dbMount(struct inode *ipbmap)
166 {
167 	struct bmap *bmp;
168 	struct dbmap_disk *dbmp_le;
169 	struct metapage *mp;
170 	int i;
171 
172 	/*
173 	 * allocate/initialize the in-memory bmap descriptor
174 	 */
175 	/* allocate memory for the in-memory bmap descriptor */
176 	bmp = kmalloc(sizeof(struct bmap), GFP_KERNEL);
177 	if (bmp == NULL)
178 		return -ENOMEM;
179 
180 	/* read the on-disk bmap descriptor. */
181 	mp = read_metapage(ipbmap,
182 			   BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
183 			   PSIZE, 0);
184 	if (mp == NULL) {
185 		kfree(bmp);
186 		return -EIO;
187 	}
188 
189 	/* copy the on-disk bmap descriptor to its in-memory version. */
190 	dbmp_le = (struct dbmap_disk *) mp->data;
191 	bmp->db_mapsize = le64_to_cpu(dbmp_le->dn_mapsize);
192 	bmp->db_nfree = le64_to_cpu(dbmp_le->dn_nfree);
193 	bmp->db_l2nbperpage = le32_to_cpu(dbmp_le->dn_l2nbperpage);
194 	bmp->db_numag = le32_to_cpu(dbmp_le->dn_numag);
195 	bmp->db_maxlevel = le32_to_cpu(dbmp_le->dn_maxlevel);
196 	bmp->db_maxag = le32_to_cpu(dbmp_le->dn_maxag);
197 	bmp->db_agpref = le32_to_cpu(dbmp_le->dn_agpref);
198 	bmp->db_aglevel = le32_to_cpu(dbmp_le->dn_aglevel);
199 	bmp->db_agheight = le32_to_cpu(dbmp_le->dn_agheight);
200 	bmp->db_agwidth = le32_to_cpu(dbmp_le->dn_agwidth);
201 	bmp->db_agstart = le32_to_cpu(dbmp_le->dn_agstart);
202 	bmp->db_agl2size = le32_to_cpu(dbmp_le->dn_agl2size);
203 	for (i = 0; i < MAXAG; i++)
204 		bmp->db_agfree[i] = le64_to_cpu(dbmp_le->dn_agfree[i]);
205 	bmp->db_agsize = le64_to_cpu(dbmp_le->dn_agsize);
206 	bmp->db_maxfreebud = dbmp_le->dn_maxfreebud;
207 
208 	/* release the buffer. */
209 	release_metapage(mp);
210 
211 	/* bind the bmap inode and the bmap descriptor to each other. */
212 	bmp->db_ipbmap = ipbmap;
213 	JFS_SBI(ipbmap->i_sb)->bmap = bmp;
214 
215 	memset(bmp->db_active, 0, sizeof(bmp->db_active));
216 
217 	/*
218 	 * allocate/initialize the bmap lock
219 	 */
220 	BMAP_LOCK_INIT(bmp);
221 
222 	return (0);
223 }
224 
225 
226 /*
227  * NAME:	dbUnmount()
228  *
229  * FUNCTION:	terminate the block allocation map in preparation for
230  *		file system unmount.
231  *
232  *		the in-core bmap descriptor is written to disk and
233  *		the memory for this descriptor is freed.
234  *
235  * PARAMETERS:
236  *	ipbmap	- pointer to in-core inode for the block map.
237  *
238  * RETURN VALUES:
239  *	0	- success
240  *	-EIO	- i/o error
241  */
242 int dbUnmount(struct inode *ipbmap, int mounterror)
243 {
244 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
245 
246 	if (!(mounterror || isReadOnly(ipbmap)))
247 		dbSync(ipbmap);
248 
249 	/*
250 	 * Invalidate the page cache buffers
251 	 */
252 	truncate_inode_pages(ipbmap->i_mapping, 0);
253 
254 	/* free the memory for the in-memory bmap. */
255 	kfree(bmp);
256 
257 	return (0);
258 }
259 
260 /*
261  *	dbSync()
262  */
263 int dbSync(struct inode *ipbmap)
264 {
265 	struct dbmap_disk *dbmp_le;
266 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
267 	struct metapage *mp;
268 	int i;
269 
270 	/*
271 	 * write bmap global control page
272 	 */
273 	/* get the buffer for the on-disk bmap descriptor. */
274 	mp = read_metapage(ipbmap,
275 			   BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
276 			   PSIZE, 0);
277 	if (mp == NULL) {
278 		jfs_err("dbSync: read_metapage failed!");
279 		return -EIO;
280 	}
281 	/* copy the in-memory version of the bmap to the on-disk version */
282 	dbmp_le = (struct dbmap_disk *) mp->data;
283 	dbmp_le->dn_mapsize = cpu_to_le64(bmp->db_mapsize);
284 	dbmp_le->dn_nfree = cpu_to_le64(bmp->db_nfree);
285 	dbmp_le->dn_l2nbperpage = cpu_to_le32(bmp->db_l2nbperpage);
286 	dbmp_le->dn_numag = cpu_to_le32(bmp->db_numag);
287 	dbmp_le->dn_maxlevel = cpu_to_le32(bmp->db_maxlevel);
288 	dbmp_le->dn_maxag = cpu_to_le32(bmp->db_maxag);
289 	dbmp_le->dn_agpref = cpu_to_le32(bmp->db_agpref);
290 	dbmp_le->dn_aglevel = cpu_to_le32(bmp->db_aglevel);
291 	dbmp_le->dn_agheight = cpu_to_le32(bmp->db_agheight);
292 	dbmp_le->dn_agwidth = cpu_to_le32(bmp->db_agwidth);
293 	dbmp_le->dn_agstart = cpu_to_le32(bmp->db_agstart);
294 	dbmp_le->dn_agl2size = cpu_to_le32(bmp->db_agl2size);
295 	for (i = 0; i < MAXAG; i++)
296 		dbmp_le->dn_agfree[i] = cpu_to_le64(bmp->db_agfree[i]);
297 	dbmp_le->dn_agsize = cpu_to_le64(bmp->db_agsize);
298 	dbmp_le->dn_maxfreebud = bmp->db_maxfreebud;
299 
300 	/* write the buffer */
301 	write_metapage(mp);
302 
303 	/*
304 	 * write out dirty pages of bmap
305 	 */
306 	filemap_write_and_wait(ipbmap->i_mapping);
307 
308 	diWriteSpecial(ipbmap, 0);
309 
310 	return (0);
311 }
312 
313 /*
314  * NAME:	dbFree()
315  *
316  * FUNCTION:	free the specified block range from the working block
317  *		allocation map.
318  *
319  *		the blocks will be free from the working map one dmap
320  *		at a time.
321  *
322  * PARAMETERS:
323  *	ip	- pointer to in-core inode;
324  *	blkno	- starting block number to be freed.
325  *	nblocks	- number of blocks to be freed.
326  *
327  * RETURN VALUES:
328  *	0	- success
329  *	-EIO	- i/o error
330  */
331 int dbFree(struct inode *ip, s64 blkno, s64 nblocks)
332 {
333 	struct metapage *mp;
334 	struct dmap *dp;
335 	int nb, rc;
336 	s64 lblkno, rem;
337 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
338 	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
339 	struct super_block *sb = ipbmap->i_sb;
340 
341 	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
342 
343 	/* block to be freed better be within the mapsize. */
344 	if (unlikely((blkno == 0) || (blkno + nblocks > bmp->db_mapsize))) {
345 		IREAD_UNLOCK(ipbmap);
346 		printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
347 		       (unsigned long long) blkno,
348 		       (unsigned long long) nblocks);
349 		jfs_error(ip->i_sb, "block to be freed is outside the map\n");
350 		return -EIO;
351 	}
352 
353 	/**
354 	 * TRIM the blocks, when mounted with discard option
355 	 */
356 	if (JFS_SBI(sb)->flag & JFS_DISCARD)
357 		if (JFS_SBI(sb)->minblks_trim <= nblocks)
358 			jfs_issue_discard(ipbmap, blkno, nblocks);
359 
360 	/*
361 	 * free the blocks a dmap at a time.
362 	 */
363 	mp = NULL;
364 	for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
365 		/* release previous dmap if any */
366 		if (mp) {
367 			write_metapage(mp);
368 		}
369 
370 		/* get the buffer for the current dmap. */
371 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
372 		mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
373 		if (mp == NULL) {
374 			IREAD_UNLOCK(ipbmap);
375 			return -EIO;
376 		}
377 		dp = (struct dmap *) mp->data;
378 
379 		/* determine the number of blocks to be freed from
380 		 * this dmap.
381 		 */
382 		nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
383 
384 		/* free the blocks. */
385 		if ((rc = dbFreeDmap(bmp, dp, blkno, nb))) {
386 			jfs_error(ip->i_sb, "error in block map\n");
387 			release_metapage(mp);
388 			IREAD_UNLOCK(ipbmap);
389 			return (rc);
390 		}
391 	}
392 
393 	/* write the last buffer. */
394 	write_metapage(mp);
395 
396 	IREAD_UNLOCK(ipbmap);
397 
398 	return (0);
399 }
400 
401 
402 /*
403  * NAME:	dbUpdatePMap()
404  *
405  * FUNCTION:	update the allocation state (free or allocate) of the
406  *		specified block range in the persistent block allocation map.
407  *
408  *		the blocks will be updated in the persistent map one
409  *		dmap at a time.
410  *
411  * PARAMETERS:
412  *	ipbmap	- pointer to in-core inode for the block map.
413  *	free	- 'true' if block range is to be freed from the persistent
414  *		  map; 'false' if it is to be allocated.
415  *	blkno	- starting block number of the range.
416  *	nblocks	- number of contiguous blocks in the range.
417  *	tblk	- transaction block;
418  *
419  * RETURN VALUES:
420  *	0	- success
421  *	-EIO	- i/o error
422  */
423 int
424 dbUpdatePMap(struct inode *ipbmap,
425 	     int free, s64 blkno, s64 nblocks, struct tblock * tblk)
426 {
427 	int nblks, dbitno, wbitno, rbits;
428 	int word, nbits, nwords;
429 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
430 	s64 lblkno, rem, lastlblkno;
431 	u32 mask;
432 	struct dmap *dp;
433 	struct metapage *mp;
434 	struct jfs_log *log;
435 	int lsn, difft, diffp;
436 	unsigned long flags;
437 
438 	/* the blocks better be within the mapsize. */
439 	if (blkno + nblocks > bmp->db_mapsize) {
440 		printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
441 		       (unsigned long long) blkno,
442 		       (unsigned long long) nblocks);
443 		jfs_error(ipbmap->i_sb, "blocks are outside the map\n");
444 		return -EIO;
445 	}
446 
447 	/* compute delta of transaction lsn from log syncpt */
448 	lsn = tblk->lsn;
449 	log = (struct jfs_log *) JFS_SBI(tblk->sb)->log;
450 	logdiff(difft, lsn, log);
451 
452 	/*
453 	 * update the block state a dmap at a time.
454 	 */
455 	mp = NULL;
456 	lastlblkno = 0;
457 	for (rem = nblocks; rem > 0; rem -= nblks, blkno += nblks) {
458 		/* get the buffer for the current dmap. */
459 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
460 		if (lblkno != lastlblkno) {
461 			if (mp) {
462 				write_metapage(mp);
463 			}
464 
465 			mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE,
466 					   0);
467 			if (mp == NULL)
468 				return -EIO;
469 			metapage_wait_for_io(mp);
470 		}
471 		dp = (struct dmap *) mp->data;
472 
473 		/* determine the bit number and word within the dmap of
474 		 * the starting block.  also determine how many blocks
475 		 * are to be updated within this dmap.
476 		 */
477 		dbitno = blkno & (BPERDMAP - 1);
478 		word = dbitno >> L2DBWORD;
479 		nblks = min(rem, (s64)BPERDMAP - dbitno);
480 
481 		/* update the bits of the dmap words. the first and last
482 		 * words may only have a subset of their bits updated. if
483 		 * this is the case, we'll work against that word (i.e.
484 		 * partial first and/or last) only in a single pass.  a
485 		 * single pass will also be used to update all words that
486 		 * are to have all their bits updated.
487 		 */
488 		for (rbits = nblks; rbits > 0;
489 		     rbits -= nbits, dbitno += nbits) {
490 			/* determine the bit number within the word and
491 			 * the number of bits within the word.
492 			 */
493 			wbitno = dbitno & (DBWORD - 1);
494 			nbits = min(rbits, DBWORD - wbitno);
495 
496 			/* check if only part of the word is to be updated. */
497 			if (nbits < DBWORD) {
498 				/* update (free or allocate) the bits
499 				 * in this word.
500 				 */
501 				mask =
502 				    (ONES << (DBWORD - nbits) >> wbitno);
503 				if (free)
504 					dp->pmap[word] &=
505 					    cpu_to_le32(~mask);
506 				else
507 					dp->pmap[word] |=
508 					    cpu_to_le32(mask);
509 
510 				word += 1;
511 			} else {
512 				/* one or more words are to have all
513 				 * their bits updated.  determine how
514 				 * many words and how many bits.
515 				 */
516 				nwords = rbits >> L2DBWORD;
517 				nbits = nwords << L2DBWORD;
518 
519 				/* update (free or allocate) the bits
520 				 * in these words.
521 				 */
522 				if (free)
523 					memset(&dp->pmap[word], 0,
524 					       nwords * 4);
525 				else
526 					memset(&dp->pmap[word], (int) ONES,
527 					       nwords * 4);
528 
529 				word += nwords;
530 			}
531 		}
532 
533 		/*
534 		 * update dmap lsn
535 		 */
536 		if (lblkno == lastlblkno)
537 			continue;
538 
539 		lastlblkno = lblkno;
540 
541 		LOGSYNC_LOCK(log, flags);
542 		if (mp->lsn != 0) {
543 			/* inherit older/smaller lsn */
544 			logdiff(diffp, mp->lsn, log);
545 			if (difft < diffp) {
546 				mp->lsn = lsn;
547 
548 				/* move bp after tblock in logsync list */
549 				list_move(&mp->synclist, &tblk->synclist);
550 			}
551 
552 			/* inherit younger/larger clsn */
553 			logdiff(difft, tblk->clsn, log);
554 			logdiff(diffp, mp->clsn, log);
555 			if (difft > diffp)
556 				mp->clsn = tblk->clsn;
557 		} else {
558 			mp->log = log;
559 			mp->lsn = lsn;
560 
561 			/* insert bp after tblock in logsync list */
562 			log->count++;
563 			list_add(&mp->synclist, &tblk->synclist);
564 
565 			mp->clsn = tblk->clsn;
566 		}
567 		LOGSYNC_UNLOCK(log, flags);
568 	}
569 
570 	/* write the last buffer. */
571 	if (mp) {
572 		write_metapage(mp);
573 	}
574 
575 	return (0);
576 }
577 
578 
579 /*
580  * NAME:	dbNextAG()
581  *
582  * FUNCTION:	find the preferred allocation group for new allocations.
583  *
584  *		Within the allocation groups, we maintain a preferred
585  *		allocation group which consists of a group with at least
586  *		average free space.  It is the preferred group that we target
587  *		new inode allocation towards.  The tie-in between inode
588  *		allocation and block allocation occurs as we allocate the
589  *		first (data) block of an inode and specify the inode (block)
590  *		as the allocation hint for this block.
591  *
592  *		We try to avoid having more than one open file growing in
593  *		an allocation group, as this will lead to fragmentation.
594  *		This differs from the old OS/2 method of trying to keep
595  *		empty ags around for large allocations.
596  *
597  * PARAMETERS:
598  *	ipbmap	- pointer to in-core inode for the block map.
599  *
600  * RETURN VALUES:
601  *	the preferred allocation group number.
602  */
603 int dbNextAG(struct inode *ipbmap)
604 {
605 	s64 avgfree;
606 	int agpref;
607 	s64 hwm = 0;
608 	int i;
609 	int next_best = -1;
610 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
611 
612 	BMAP_LOCK(bmp);
613 
614 	/* determine the average number of free blocks within the ags. */
615 	avgfree = (u32)bmp->db_nfree / bmp->db_numag;
616 
617 	/*
618 	 * if the current preferred ag does not have an active allocator
619 	 * and has at least average freespace, return it
620 	 */
621 	agpref = bmp->db_agpref;
622 	if ((atomic_read(&bmp->db_active[agpref]) == 0) &&
623 	    (bmp->db_agfree[agpref] >= avgfree))
624 		goto unlock;
625 
626 	/* From the last preferred ag, find the next one with at least
627 	 * average free space.
628 	 */
629 	for (i = 0 ; i < bmp->db_numag; i++, agpref++) {
630 		if (agpref == bmp->db_numag)
631 			agpref = 0;
632 
633 		if (atomic_read(&bmp->db_active[agpref]))
634 			/* open file is currently growing in this ag */
635 			continue;
636 		if (bmp->db_agfree[agpref] >= avgfree) {
637 			/* Return this one */
638 			bmp->db_agpref = agpref;
639 			goto unlock;
640 		} else if (bmp->db_agfree[agpref] > hwm) {
641 			/* Less than avg. freespace, but best so far */
642 			hwm = bmp->db_agfree[agpref];
643 			next_best = agpref;
644 		}
645 	}
646 
647 	/*
648 	 * If no inactive ag was found with average freespace, use the
649 	 * next best
650 	 */
651 	if (next_best != -1)
652 		bmp->db_agpref = next_best;
653 	/* else leave db_agpref unchanged */
654 unlock:
655 	BMAP_UNLOCK(bmp);
656 
657 	/* return the preferred group.
658 	 */
659 	return (bmp->db_agpref);
660 }
661 
662 /*
663  * NAME:	dbAlloc()
664  *
665  * FUNCTION:	attempt to allocate a specified number of contiguous free
666  *		blocks from the working allocation block map.
667  *
668  *		the block allocation policy uses hints and a multi-step
669  *		approach.
670  *
671  *		for allocation requests smaller than the number of blocks
672  *		per dmap, we first try to allocate the new blocks
673  *		immediately following the hint.  if these blocks are not
674  *		available, we try to allocate blocks near the hint.  if
675  *		no blocks near the hint are available, we next try to
676  *		allocate within the same dmap as contains the hint.
677  *
678  *		if no blocks are available in the dmap or the allocation
679  *		request is larger than the dmap size, we try to allocate
680  *		within the same allocation group as contains the hint. if
681  *		this does not succeed, we finally try to allocate anywhere
682  *		within the aggregate.
683  *
684  *		we also try to allocate anywhere within the aggregate for
685  *		for allocation requests larger than the allocation group
686  *		size or requests that specify no hint value.
687  *
688  * PARAMETERS:
689  *	ip	- pointer to in-core inode;
690  *	hint	- allocation hint.
691  *	nblocks	- number of contiguous blocks in the range.
692  *	results	- on successful return, set to the starting block number
693  *		  of the newly allocated contiguous range.
694  *
695  * RETURN VALUES:
696  *	0	- success
697  *	-ENOSPC	- insufficient disk resources
698  *	-EIO	- i/o error
699  */
700 int dbAlloc(struct inode *ip, s64 hint, s64 nblocks, s64 * results)
701 {
702 	int rc, agno;
703 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
704 	struct bmap *bmp;
705 	struct metapage *mp;
706 	s64 lblkno, blkno;
707 	struct dmap *dp;
708 	int l2nb;
709 	s64 mapSize;
710 	int writers;
711 
712 	/* assert that nblocks is valid */
713 	assert(nblocks > 0);
714 
715 	/* get the log2 number of blocks to be allocated.
716 	 * if the number of blocks is not a log2 multiple,
717 	 * it will be rounded up to the next log2 multiple.
718 	 */
719 	l2nb = BLKSTOL2(nblocks);
720 
721 	bmp = JFS_SBI(ip->i_sb)->bmap;
722 
723 	mapSize = bmp->db_mapsize;
724 
725 	/* the hint should be within the map */
726 	if (hint >= mapSize) {
727 		jfs_error(ip->i_sb, "the hint is outside the map\n");
728 		return -EIO;
729 	}
730 
731 	/* if the number of blocks to be allocated is greater than the
732 	 * allocation group size, try to allocate anywhere.
733 	 */
734 	if (l2nb > bmp->db_agl2size) {
735 		IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
736 
737 		rc = dbAllocAny(bmp, nblocks, l2nb, results);
738 
739 		goto write_unlock;
740 	}
741 
742 	/*
743 	 * If no hint, let dbNextAG recommend an allocation group
744 	 */
745 	if (hint == 0)
746 		goto pref_ag;
747 
748 	/* we would like to allocate close to the hint.  adjust the
749 	 * hint to the block following the hint since the allocators
750 	 * will start looking for free space starting at this point.
751 	 */
752 	blkno = hint + 1;
753 
754 	if (blkno >= bmp->db_mapsize)
755 		goto pref_ag;
756 
757 	agno = blkno >> bmp->db_agl2size;
758 
759 	/* check if blkno crosses over into a new allocation group.
760 	 * if so, check if we should allow allocations within this
761 	 * allocation group.
762 	 */
763 	if ((blkno & (bmp->db_agsize - 1)) == 0)
764 		/* check if the AG is currently being written to.
765 		 * if so, call dbNextAG() to find a non-busy
766 		 * AG with sufficient free space.
767 		 */
768 		if (atomic_read(&bmp->db_active[agno]))
769 			goto pref_ag;
770 
771 	/* check if the allocation request size can be satisfied from a
772 	 * single dmap.  if so, try to allocate from the dmap containing
773 	 * the hint using a tiered strategy.
774 	 */
775 	if (nblocks <= BPERDMAP) {
776 		IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
777 
778 		/* get the buffer for the dmap containing the hint.
779 		 */
780 		rc = -EIO;
781 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
782 		mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
783 		if (mp == NULL)
784 			goto read_unlock;
785 
786 		dp = (struct dmap *) mp->data;
787 
788 		/* first, try to satisfy the allocation request with the
789 		 * blocks beginning at the hint.
790 		 */
791 		if ((rc = dbAllocNext(bmp, dp, blkno, (int) nblocks))
792 		    != -ENOSPC) {
793 			if (rc == 0) {
794 				*results = blkno;
795 				mark_metapage_dirty(mp);
796 			}
797 
798 			release_metapage(mp);
799 			goto read_unlock;
800 		}
801 
802 		writers = atomic_read(&bmp->db_active[agno]);
803 		if ((writers > 1) ||
804 		    ((writers == 1) && (JFS_IP(ip)->active_ag != agno))) {
805 			/*
806 			 * Someone else is writing in this allocation
807 			 * group.  To avoid fragmenting, try another ag
808 			 */
809 			release_metapage(mp);
810 			IREAD_UNLOCK(ipbmap);
811 			goto pref_ag;
812 		}
813 
814 		/* next, try to satisfy the allocation request with blocks
815 		 * near the hint.
816 		 */
817 		if ((rc =
818 		     dbAllocNear(bmp, dp, blkno, (int) nblocks, l2nb, results))
819 		    != -ENOSPC) {
820 			if (rc == 0)
821 				mark_metapage_dirty(mp);
822 
823 			release_metapage(mp);
824 			goto read_unlock;
825 		}
826 
827 		/* try to satisfy the allocation request with blocks within
828 		 * the same dmap as the hint.
829 		 */
830 		if ((rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results))
831 		    != -ENOSPC) {
832 			if (rc == 0)
833 				mark_metapage_dirty(mp);
834 
835 			release_metapage(mp);
836 			goto read_unlock;
837 		}
838 
839 		release_metapage(mp);
840 		IREAD_UNLOCK(ipbmap);
841 	}
842 
843 	/* try to satisfy the allocation request with blocks within
844 	 * the same allocation group as the hint.
845 	 */
846 	IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
847 	if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) != -ENOSPC)
848 		goto write_unlock;
849 
850 	IWRITE_UNLOCK(ipbmap);
851 
852 
853       pref_ag:
854 	/*
855 	 * Let dbNextAG recommend a preferred allocation group
856 	 */
857 	agno = dbNextAG(ipbmap);
858 	IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
859 
860 	/* Try to allocate within this allocation group.  if that fails, try to
861 	 * allocate anywhere in the map.
862 	 */
863 	if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) == -ENOSPC)
864 		rc = dbAllocAny(bmp, nblocks, l2nb, results);
865 
866       write_unlock:
867 	IWRITE_UNLOCK(ipbmap);
868 
869 	return (rc);
870 
871       read_unlock:
872 	IREAD_UNLOCK(ipbmap);
873 
874 	return (rc);
875 }
876 
877 #ifdef _NOTYET
878 /*
879  * NAME:	dbAllocExact()
880  *
881  * FUNCTION:	try to allocate the requested extent;
882  *
883  * PARAMETERS:
884  *	ip	- pointer to in-core inode;
885  *	blkno	- extent address;
886  *	nblocks	- extent length;
887  *
888  * RETURN VALUES:
889  *	0	- success
890  *	-ENOSPC	- insufficient disk resources
891  *	-EIO	- i/o error
892  */
893 int dbAllocExact(struct inode *ip, s64 blkno, int nblocks)
894 {
895 	int rc;
896 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
897 	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
898 	struct dmap *dp;
899 	s64 lblkno;
900 	struct metapage *mp;
901 
902 	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
903 
904 	/*
905 	 * validate extent request:
906 	 *
907 	 * note: defragfs policy:
908 	 *  max 64 blocks will be moved.
909 	 *  allocation request size must be satisfied from a single dmap.
910 	 */
911 	if (nblocks <= 0 || nblocks > BPERDMAP || blkno >= bmp->db_mapsize) {
912 		IREAD_UNLOCK(ipbmap);
913 		return -EINVAL;
914 	}
915 
916 	if (nblocks > ((s64) 1 << bmp->db_maxfreebud)) {
917 		/* the free space is no longer available */
918 		IREAD_UNLOCK(ipbmap);
919 		return -ENOSPC;
920 	}
921 
922 	/* read in the dmap covering the extent */
923 	lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
924 	mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
925 	if (mp == NULL) {
926 		IREAD_UNLOCK(ipbmap);
927 		return -EIO;
928 	}
929 	dp = (struct dmap *) mp->data;
930 
931 	/* try to allocate the requested extent */
932 	rc = dbAllocNext(bmp, dp, blkno, nblocks);
933 
934 	IREAD_UNLOCK(ipbmap);
935 
936 	if (rc == 0)
937 		mark_metapage_dirty(mp);
938 
939 	release_metapage(mp);
940 
941 	return (rc);
942 }
943 #endif /* _NOTYET */
944 
945 /*
946  * NAME:	dbReAlloc()
947  *
948  * FUNCTION:	attempt to extend a current allocation by a specified
949  *		number of blocks.
950  *
951  *		this routine attempts to satisfy the allocation request
952  *		by first trying to extend the existing allocation in
953  *		place by allocating the additional blocks as the blocks
954  *		immediately following the current allocation.  if these
955  *		blocks are not available, this routine will attempt to
956  *		allocate a new set of contiguous blocks large enough
957  *		to cover the existing allocation plus the additional
958  *		number of blocks required.
959  *
960  * PARAMETERS:
961  *	ip	    -  pointer to in-core inode requiring allocation.
962  *	blkno	    -  starting block of the current allocation.
963  *	nblocks	    -  number of contiguous blocks within the current
964  *		       allocation.
965  *	addnblocks  -  number of blocks to add to the allocation.
966  *	results	-      on successful return, set to the starting block number
967  *		       of the existing allocation if the existing allocation
968  *		       was extended in place or to a newly allocated contiguous
969  *		       range if the existing allocation could not be extended
970  *		       in place.
971  *
972  * RETURN VALUES:
973  *	0	- success
974  *	-ENOSPC	- insufficient disk resources
975  *	-EIO	- i/o error
976  */
977 int
978 dbReAlloc(struct inode *ip,
979 	  s64 blkno, s64 nblocks, s64 addnblocks, s64 * results)
980 {
981 	int rc;
982 
983 	/* try to extend the allocation in place.
984 	 */
985 	if ((rc = dbExtend(ip, blkno, nblocks, addnblocks)) == 0) {
986 		*results = blkno;
987 		return (0);
988 	} else {
989 		if (rc != -ENOSPC)
990 			return (rc);
991 	}
992 
993 	/* could not extend the allocation in place, so allocate a
994 	 * new set of blocks for the entire request (i.e. try to get
995 	 * a range of contiguous blocks large enough to cover the
996 	 * existing allocation plus the additional blocks.)
997 	 */
998 	return (dbAlloc
999 		(ip, blkno + nblocks - 1, addnblocks + nblocks, results));
1000 }
1001 
1002 
1003 /*
1004  * NAME:	dbExtend()
1005  *
1006  * FUNCTION:	attempt to extend a current allocation by a specified
1007  *		number of blocks.
1008  *
1009  *		this routine attempts to satisfy the allocation request
1010  *		by first trying to extend the existing allocation in
1011  *		place by allocating the additional blocks as the blocks
1012  *		immediately following the current allocation.
1013  *
1014  * PARAMETERS:
1015  *	ip	    -  pointer to in-core inode requiring allocation.
1016  *	blkno	    -  starting block of the current allocation.
1017  *	nblocks	    -  number of contiguous blocks within the current
1018  *		       allocation.
1019  *	addnblocks  -  number of blocks to add to the allocation.
1020  *
1021  * RETURN VALUES:
1022  *	0	- success
1023  *	-ENOSPC	- insufficient disk resources
1024  *	-EIO	- i/o error
1025  */
1026 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks)
1027 {
1028 	struct jfs_sb_info *sbi = JFS_SBI(ip->i_sb);
1029 	s64 lblkno, lastblkno, extblkno;
1030 	uint rel_block;
1031 	struct metapage *mp;
1032 	struct dmap *dp;
1033 	int rc;
1034 	struct inode *ipbmap = sbi->ipbmap;
1035 	struct bmap *bmp;
1036 
1037 	/*
1038 	 * We don't want a non-aligned extent to cross a page boundary
1039 	 */
1040 	if (((rel_block = blkno & (sbi->nbperpage - 1))) &&
1041 	    (rel_block + nblocks + addnblocks > sbi->nbperpage))
1042 		return -ENOSPC;
1043 
1044 	/* get the last block of the current allocation */
1045 	lastblkno = blkno + nblocks - 1;
1046 
1047 	/* determine the block number of the block following
1048 	 * the existing allocation.
1049 	 */
1050 	extblkno = lastblkno + 1;
1051 
1052 	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
1053 
1054 	/* better be within the file system */
1055 	bmp = sbi->bmap;
1056 	if (lastblkno < 0 || lastblkno >= bmp->db_mapsize) {
1057 		IREAD_UNLOCK(ipbmap);
1058 		jfs_error(ip->i_sb, "the block is outside the filesystem\n");
1059 		return -EIO;
1060 	}
1061 
1062 	/* we'll attempt to extend the current allocation in place by
1063 	 * allocating the additional blocks as the blocks immediately
1064 	 * following the current allocation.  we only try to extend the
1065 	 * current allocation in place if the number of additional blocks
1066 	 * can fit into a dmap, the last block of the current allocation
1067 	 * is not the last block of the file system, and the start of the
1068 	 * inplace extension is not on an allocation group boundary.
1069 	 */
1070 	if (addnblocks > BPERDMAP || extblkno >= bmp->db_mapsize ||
1071 	    (extblkno & (bmp->db_agsize - 1)) == 0) {
1072 		IREAD_UNLOCK(ipbmap);
1073 		return -ENOSPC;
1074 	}
1075 
1076 	/* get the buffer for the dmap containing the first block
1077 	 * of the extension.
1078 	 */
1079 	lblkno = BLKTODMAP(extblkno, bmp->db_l2nbperpage);
1080 	mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
1081 	if (mp == NULL) {
1082 		IREAD_UNLOCK(ipbmap);
1083 		return -EIO;
1084 	}
1085 
1086 	dp = (struct dmap *) mp->data;
1087 
1088 	/* try to allocate the blocks immediately following the
1089 	 * current allocation.
1090 	 */
1091 	rc = dbAllocNext(bmp, dp, extblkno, (int) addnblocks);
1092 
1093 	IREAD_UNLOCK(ipbmap);
1094 
1095 	/* were we successful ? */
1096 	if (rc == 0)
1097 		write_metapage(mp);
1098 	else
1099 		/* we were not successful */
1100 		release_metapage(mp);
1101 
1102 	return (rc);
1103 }
1104 
1105 
1106 /*
1107  * NAME:	dbAllocNext()
1108  *
1109  * FUNCTION:	attempt to allocate the blocks of the specified block
1110  *		range within a dmap.
1111  *
1112  * PARAMETERS:
1113  *	bmp	-  pointer to bmap descriptor
1114  *	dp	-  pointer to dmap.
1115  *	blkno	-  starting block number of the range.
1116  *	nblocks	-  number of contiguous free blocks of the range.
1117  *
1118  * RETURN VALUES:
1119  *	0	- success
1120  *	-ENOSPC	- insufficient disk resources
1121  *	-EIO	- i/o error
1122  *
1123  * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1124  */
1125 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
1126 		       int nblocks)
1127 {
1128 	int dbitno, word, rembits, nb, nwords, wbitno, nw;
1129 	int l2size;
1130 	s8 *leaf;
1131 	u32 mask;
1132 
1133 	if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1134 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1135 		return -EIO;
1136 	}
1137 
1138 	/* pick up a pointer to the leaves of the dmap tree.
1139 	 */
1140 	leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1141 
1142 	/* determine the bit number and word within the dmap of the
1143 	 * starting block.
1144 	 */
1145 	dbitno = blkno & (BPERDMAP - 1);
1146 	word = dbitno >> L2DBWORD;
1147 
1148 	/* check if the specified block range is contained within
1149 	 * this dmap.
1150 	 */
1151 	if (dbitno + nblocks > BPERDMAP)
1152 		return -ENOSPC;
1153 
1154 	/* check if the starting leaf indicates that anything
1155 	 * is free.
1156 	 */
1157 	if (leaf[word] == NOFREE)
1158 		return -ENOSPC;
1159 
1160 	/* check the dmaps words corresponding to block range to see
1161 	 * if the block range is free.  not all bits of the first and
1162 	 * last words may be contained within the block range.  if this
1163 	 * is the case, we'll work against those words (i.e. partial first
1164 	 * and/or last) on an individual basis (a single pass) and examine
1165 	 * the actual bits to determine if they are free.  a single pass
1166 	 * will be used for all dmap words fully contained within the
1167 	 * specified range.  within this pass, the leaves of the dmap
1168 	 * tree will be examined to determine if the blocks are free. a
1169 	 * single leaf may describe the free space of multiple dmap
1170 	 * words, so we may visit only a subset of the actual leaves
1171 	 * corresponding to the dmap words of the block range.
1172 	 */
1173 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
1174 		/* determine the bit number within the word and
1175 		 * the number of bits within the word.
1176 		 */
1177 		wbitno = dbitno & (DBWORD - 1);
1178 		nb = min(rembits, DBWORD - wbitno);
1179 
1180 		/* check if only part of the word is to be examined.
1181 		 */
1182 		if (nb < DBWORD) {
1183 			/* check if the bits are free.
1184 			 */
1185 			mask = (ONES << (DBWORD - nb) >> wbitno);
1186 			if ((mask & ~le32_to_cpu(dp->wmap[word])) != mask)
1187 				return -ENOSPC;
1188 
1189 			word += 1;
1190 		} else {
1191 			/* one or more dmap words are fully contained
1192 			 * within the block range.  determine how many
1193 			 * words and how many bits.
1194 			 */
1195 			nwords = rembits >> L2DBWORD;
1196 			nb = nwords << L2DBWORD;
1197 
1198 			/* now examine the appropriate leaves to determine
1199 			 * if the blocks are free.
1200 			 */
1201 			while (nwords > 0) {
1202 				/* does the leaf describe any free space ?
1203 				 */
1204 				if (leaf[word] < BUDMIN)
1205 					return -ENOSPC;
1206 
1207 				/* determine the l2 number of bits provided
1208 				 * by this leaf.
1209 				 */
1210 				l2size =
1211 				    min_t(int, leaf[word], NLSTOL2BSZ(nwords));
1212 
1213 				/* determine how many words were handled.
1214 				 */
1215 				nw = BUDSIZE(l2size, BUDMIN);
1216 
1217 				nwords -= nw;
1218 				word += nw;
1219 			}
1220 		}
1221 	}
1222 
1223 	/* allocate the blocks.
1224 	 */
1225 	return (dbAllocDmap(bmp, dp, blkno, nblocks));
1226 }
1227 
1228 
1229 /*
1230  * NAME:	dbAllocNear()
1231  *
1232  * FUNCTION:	attempt to allocate a number of contiguous free blocks near
1233  *		a specified block (hint) within a dmap.
1234  *
1235  *		starting with the dmap leaf that covers the hint, we'll
1236  *		check the next four contiguous leaves for sufficient free
1237  *		space.  if sufficient free space is found, we'll allocate
1238  *		the desired free space.
1239  *
1240  * PARAMETERS:
1241  *	bmp	-  pointer to bmap descriptor
1242  *	dp	-  pointer to dmap.
1243  *	blkno	-  block number to allocate near.
1244  *	nblocks	-  actual number of contiguous free blocks desired.
1245  *	l2nb	-  log2 number of contiguous free blocks desired.
1246  *	results	-  on successful return, set to the starting block number
1247  *		   of the newly allocated range.
1248  *
1249  * RETURN VALUES:
1250  *	0	- success
1251  *	-ENOSPC	- insufficient disk resources
1252  *	-EIO	- i/o error
1253  *
1254  * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1255  */
1256 static int
1257 dbAllocNear(struct bmap * bmp,
1258 	    struct dmap * dp, s64 blkno, int nblocks, int l2nb, s64 * results)
1259 {
1260 	int word, lword, rc;
1261 	s8 *leaf;
1262 
1263 	if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1264 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1265 		return -EIO;
1266 	}
1267 
1268 	leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1269 
1270 	/* determine the word within the dmap that holds the hint
1271 	 * (i.e. blkno).  also, determine the last word in the dmap
1272 	 * that we'll include in our examination.
1273 	 */
1274 	word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
1275 	lword = min(word + 4, LPERDMAP);
1276 
1277 	/* examine the leaves for sufficient free space.
1278 	 */
1279 	for (; word < lword; word++) {
1280 		/* does the leaf describe sufficient free space ?
1281 		 */
1282 		if (leaf[word] < l2nb)
1283 			continue;
1284 
1285 		/* determine the block number within the file system
1286 		 * of the first block described by this dmap word.
1287 		 */
1288 		blkno = le64_to_cpu(dp->start) + (word << L2DBWORD);
1289 
1290 		/* if not all bits of the dmap word are free, get the
1291 		 * starting bit number within the dmap word of the required
1292 		 * string of free bits and adjust the block number with the
1293 		 * value.
1294 		 */
1295 		if (leaf[word] < BUDMIN)
1296 			blkno +=
1297 			    dbFindBits(le32_to_cpu(dp->wmap[word]), l2nb);
1298 
1299 		/* allocate the blocks.
1300 		 */
1301 		if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1302 			*results = blkno;
1303 
1304 		return (rc);
1305 	}
1306 
1307 	return -ENOSPC;
1308 }
1309 
1310 
1311 /*
1312  * NAME:	dbAllocAG()
1313  *
1314  * FUNCTION:	attempt to allocate the specified number of contiguous
1315  *		free blocks within the specified allocation group.
1316  *
1317  *		unless the allocation group size is equal to the number
1318  *		of blocks per dmap, the dmap control pages will be used to
1319  *		find the required free space, if available.  we start the
1320  *		search at the highest dmap control page level which
1321  *		distinctly describes the allocation group's free space
1322  *		(i.e. the highest level at which the allocation group's
1323  *		free space is not mixed in with that of any other group).
1324  *		in addition, we start the search within this level at a
1325  *		height of the dmapctl dmtree at which the nodes distinctly
1326  *		describe the allocation group's free space.  at this height,
1327  *		the allocation group's free space may be represented by 1
1328  *		or two sub-trees, depending on the allocation group size.
1329  *		we search the top nodes of these subtrees left to right for
1330  *		sufficient free space.  if sufficient free space is found,
1331  *		the subtree is searched to find the leftmost leaf that
1332  *		has free space.  once we have made it to the leaf, we
1333  *		move the search to the next lower level dmap control page
1334  *		corresponding to this leaf.  we continue down the dmap control
1335  *		pages until we find the dmap that contains or starts the
1336  *		sufficient free space and we allocate at this dmap.
1337  *
1338  *		if the allocation group size is equal to the dmap size,
1339  *		we'll start at the dmap corresponding to the allocation
1340  *		group and attempt the allocation at this level.
1341  *
1342  *		the dmap control page search is also not performed if the
1343  *		allocation group is completely free and we go to the first
1344  *		dmap of the allocation group to do the allocation.  this is
1345  *		done because the allocation group may be part (not the first
1346  *		part) of a larger binary buddy system, causing the dmap
1347  *		control pages to indicate no free space (NOFREE) within
1348  *		the allocation group.
1349  *
1350  * PARAMETERS:
1351  *	bmp	-  pointer to bmap descriptor
1352  *	agno	- allocation group number.
1353  *	nblocks	-  actual number of contiguous free blocks desired.
1354  *	l2nb	-  log2 number of contiguous free blocks desired.
1355  *	results	-  on successful return, set to the starting block number
1356  *		   of the newly allocated range.
1357  *
1358  * RETURN VALUES:
1359  *	0	- success
1360  *	-ENOSPC	- insufficient disk resources
1361  *	-EIO	- i/o error
1362  *
1363  * note: IWRITE_LOCK(ipmap) held on entry/exit;
1364  */
1365 static int
1366 dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, s64 * results)
1367 {
1368 	struct metapage *mp;
1369 	struct dmapctl *dcp;
1370 	int rc, ti, i, k, m, n, agperlev;
1371 	s64 blkno, lblkno;
1372 	int budmin;
1373 
1374 	/* allocation request should not be for more than the
1375 	 * allocation group size.
1376 	 */
1377 	if (l2nb > bmp->db_agl2size) {
1378 		jfs_error(bmp->db_ipbmap->i_sb,
1379 			  "allocation request is larger than the allocation group size\n");
1380 		return -EIO;
1381 	}
1382 
1383 	/* determine the starting block number of the allocation
1384 	 * group.
1385 	 */
1386 	blkno = (s64) agno << bmp->db_agl2size;
1387 
1388 	/* check if the allocation group size is the minimum allocation
1389 	 * group size or if the allocation group is completely free. if
1390 	 * the allocation group size is the minimum size of BPERDMAP (i.e.
1391 	 * 1 dmap), there is no need to search the dmap control page (below)
1392 	 * that fully describes the allocation group since the allocation
1393 	 * group is already fully described by a dmap.  in this case, we
1394 	 * just call dbAllocCtl() to search the dmap tree and allocate the
1395 	 * required space if available.
1396 	 *
1397 	 * if the allocation group is completely free, dbAllocCtl() is
1398 	 * also called to allocate the required space.  this is done for
1399 	 * two reasons.  first, it makes no sense searching the dmap control
1400 	 * pages for free space when we know that free space exists.  second,
1401 	 * the dmap control pages may indicate that the allocation group
1402 	 * has no free space if the allocation group is part (not the first
1403 	 * part) of a larger binary buddy system.
1404 	 */
1405 	if (bmp->db_agsize == BPERDMAP
1406 	    || bmp->db_agfree[agno] == bmp->db_agsize) {
1407 		rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1408 		if ((rc == -ENOSPC) &&
1409 		    (bmp->db_agfree[agno] == bmp->db_agsize)) {
1410 			printk(KERN_ERR "blkno = %Lx, blocks = %Lx\n",
1411 			       (unsigned long long) blkno,
1412 			       (unsigned long long) nblocks);
1413 			jfs_error(bmp->db_ipbmap->i_sb,
1414 				  "dbAllocCtl failed in free AG\n");
1415 		}
1416 		return (rc);
1417 	}
1418 
1419 	/* the buffer for the dmap control page that fully describes the
1420 	 * allocation group.
1421 	 */
1422 	lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, bmp->db_aglevel);
1423 	mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1424 	if (mp == NULL)
1425 		return -EIO;
1426 	dcp = (struct dmapctl *) mp->data;
1427 	budmin = dcp->budmin;
1428 
1429 	if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1430 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
1431 		release_metapage(mp);
1432 		return -EIO;
1433 	}
1434 
1435 	/* search the subtree(s) of the dmap control page that describes
1436 	 * the allocation group, looking for sufficient free space.  to begin,
1437 	 * determine how many allocation groups are represented in a dmap
1438 	 * control page at the control page level (i.e. L0, L1, L2) that
1439 	 * fully describes an allocation group. next, determine the starting
1440 	 * tree index of this allocation group within the control page.
1441 	 */
1442 	agperlev =
1443 	    (1 << (L2LPERCTL - (bmp->db_agheight << 1))) / bmp->db_agwidth;
1444 	ti = bmp->db_agstart + bmp->db_agwidth * (agno & (agperlev - 1));
1445 
1446 	/* dmap control page trees fan-out by 4 and a single allocation
1447 	 * group may be described by 1 or 2 subtrees within the ag level
1448 	 * dmap control page, depending upon the ag size. examine the ag's
1449 	 * subtrees for sufficient free space, starting with the leftmost
1450 	 * subtree.
1451 	 */
1452 	for (i = 0; i < bmp->db_agwidth; i++, ti++) {
1453 		/* is there sufficient free space ?
1454 		 */
1455 		if (l2nb > dcp->stree[ti])
1456 			continue;
1457 
1458 		/* sufficient free space found in a subtree. now search down
1459 		 * the subtree to find the leftmost leaf that describes this
1460 		 * free space.
1461 		 */
1462 		for (k = bmp->db_agheight; k > 0; k--) {
1463 			for (n = 0, m = (ti << 2) + 1; n < 4; n++) {
1464 				if (l2nb <= dcp->stree[m + n]) {
1465 					ti = m + n;
1466 					break;
1467 				}
1468 			}
1469 			if (n == 4) {
1470 				jfs_error(bmp->db_ipbmap->i_sb,
1471 					  "failed descending stree\n");
1472 				release_metapage(mp);
1473 				return -EIO;
1474 			}
1475 		}
1476 
1477 		/* determine the block number within the file system
1478 		 * that corresponds to this leaf.
1479 		 */
1480 		if (bmp->db_aglevel == 2)
1481 			blkno = 0;
1482 		else if (bmp->db_aglevel == 1)
1483 			blkno &= ~(MAXL1SIZE - 1);
1484 		else		/* bmp->db_aglevel == 0 */
1485 			blkno &= ~(MAXL0SIZE - 1);
1486 
1487 		blkno +=
1488 		    ((s64) (ti - le32_to_cpu(dcp->leafidx))) << budmin;
1489 
1490 		/* release the buffer in preparation for going down
1491 		 * the next level of dmap control pages.
1492 		 */
1493 		release_metapage(mp);
1494 
1495 		/* check if we need to continue to search down the lower
1496 		 * level dmap control pages.  we need to if the number of
1497 		 * blocks required is less than maximum number of blocks
1498 		 * described at the next lower level.
1499 		 */
1500 		if (l2nb < budmin) {
1501 
1502 			/* search the lower level dmap control pages to get
1503 			 * the starting block number of the dmap that
1504 			 * contains or starts off the free space.
1505 			 */
1506 			if ((rc =
1507 			     dbFindCtl(bmp, l2nb, bmp->db_aglevel - 1,
1508 				       &blkno))) {
1509 				if (rc == -ENOSPC) {
1510 					jfs_error(bmp->db_ipbmap->i_sb,
1511 						  "control page inconsistent\n");
1512 					return -EIO;
1513 				}
1514 				return (rc);
1515 			}
1516 		}
1517 
1518 		/* allocate the blocks.
1519 		 */
1520 		rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1521 		if (rc == -ENOSPC) {
1522 			jfs_error(bmp->db_ipbmap->i_sb,
1523 				  "unable to allocate blocks\n");
1524 			rc = -EIO;
1525 		}
1526 		return (rc);
1527 	}
1528 
1529 	/* no space in the allocation group.  release the buffer and
1530 	 * return -ENOSPC.
1531 	 */
1532 	release_metapage(mp);
1533 
1534 	return -ENOSPC;
1535 }
1536 
1537 
1538 /*
1539  * NAME:	dbAllocAny()
1540  *
1541  * FUNCTION:	attempt to allocate the specified number of contiguous
1542  *		free blocks anywhere in the file system.
1543  *
1544  *		dbAllocAny() attempts to find the sufficient free space by
1545  *		searching down the dmap control pages, starting with the
1546  *		highest level (i.e. L0, L1, L2) control page.  if free space
1547  *		large enough to satisfy the desired free space is found, the
1548  *		desired free space is allocated.
1549  *
1550  * PARAMETERS:
1551  *	bmp	-  pointer to bmap descriptor
1552  *	nblocks	 -  actual number of contiguous free blocks desired.
1553  *	l2nb	 -  log2 number of contiguous free blocks desired.
1554  *	results	-  on successful return, set to the starting block number
1555  *		   of the newly allocated range.
1556  *
1557  * RETURN VALUES:
1558  *	0	- success
1559  *	-ENOSPC	- insufficient disk resources
1560  *	-EIO	- i/o error
1561  *
1562  * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1563  */
1564 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results)
1565 {
1566 	int rc;
1567 	s64 blkno = 0;
1568 
1569 	/* starting with the top level dmap control page, search
1570 	 * down the dmap control levels for sufficient free space.
1571 	 * if free space is found, dbFindCtl() returns the starting
1572 	 * block number of the dmap that contains or starts off the
1573 	 * range of free space.
1574 	 */
1575 	if ((rc = dbFindCtl(bmp, l2nb, bmp->db_maxlevel, &blkno)))
1576 		return (rc);
1577 
1578 	/* allocate the blocks.
1579 	 */
1580 	rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1581 	if (rc == -ENOSPC) {
1582 		jfs_error(bmp->db_ipbmap->i_sb, "unable to allocate blocks\n");
1583 		return -EIO;
1584 	}
1585 	return (rc);
1586 }
1587 
1588 
1589 /*
1590  * NAME:	dbDiscardAG()
1591  *
1592  * FUNCTION:	attempt to discard (TRIM) all free blocks of specific AG
1593  *
1594  *		algorithm:
1595  *		1) allocate blocks, as large as possible and save them
1596  *		   while holding IWRITE_LOCK on ipbmap
1597  *		2) trim all these saved block/length values
1598  *		3) mark the blocks free again
1599  *
1600  *		benefit:
1601  *		- we work only on one ag at some time, minimizing how long we
1602  *		  need to lock ipbmap
1603  *		- reading / writing the fs is possible most time, even on
1604  *		  trimming
1605  *
1606  *		downside:
1607  *		- we write two times to the dmapctl and dmap pages
1608  *		- but for me, this seems the best way, better ideas?
1609  *		/TR 2012
1610  *
1611  * PARAMETERS:
1612  *	ip	- pointer to in-core inode
1613  *	agno	- ag to trim
1614  *	minlen	- minimum value of contiguous blocks
1615  *
1616  * RETURN VALUES:
1617  *	s64	- actual number of blocks trimmed
1618  */
1619 s64 dbDiscardAG(struct inode *ip, int agno, s64 minlen)
1620 {
1621 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
1622 	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
1623 	s64 nblocks, blkno;
1624 	u64 trimmed = 0;
1625 	int rc, l2nb;
1626 	struct super_block *sb = ipbmap->i_sb;
1627 
1628 	struct range2trim {
1629 		u64 blkno;
1630 		u64 nblocks;
1631 	} *totrim, *tt;
1632 
1633 	/* max blkno / nblocks pairs to trim */
1634 	int count = 0, range_cnt;
1635 	u64 max_ranges;
1636 
1637 	/* prevent others from writing new stuff here, while trimming */
1638 	IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
1639 
1640 	nblocks = bmp->db_agfree[agno];
1641 	max_ranges = nblocks;
1642 	do_div(max_ranges, minlen);
1643 	range_cnt = min_t(u64, max_ranges + 1, 32 * 1024);
1644 	totrim = kmalloc(sizeof(struct range2trim) * range_cnt, GFP_NOFS);
1645 	if (totrim == NULL) {
1646 		jfs_error(bmp->db_ipbmap->i_sb, "no memory for trim array\n");
1647 		IWRITE_UNLOCK(ipbmap);
1648 		return 0;
1649 	}
1650 
1651 	tt = totrim;
1652 	while (nblocks >= minlen) {
1653 		l2nb = BLKSTOL2(nblocks);
1654 
1655 		/* 0 = okay, -EIO = fatal, -ENOSPC -> try smaller block */
1656 		rc = dbAllocAG(bmp, agno, nblocks, l2nb, &blkno);
1657 		if (rc == 0) {
1658 			tt->blkno = blkno;
1659 			tt->nblocks = nblocks;
1660 			tt++; count++;
1661 
1662 			/* the whole ag is free, trim now */
1663 			if (bmp->db_agfree[agno] == 0)
1664 				break;
1665 
1666 			/* give a hint for the next while */
1667 			nblocks = bmp->db_agfree[agno];
1668 			continue;
1669 		} else if (rc == -ENOSPC) {
1670 			/* search for next smaller log2 block */
1671 			l2nb = BLKSTOL2(nblocks) - 1;
1672 			nblocks = 1 << l2nb;
1673 		} else {
1674 			/* Trim any already allocated blocks */
1675 			jfs_error(bmp->db_ipbmap->i_sb, "-EIO\n");
1676 			break;
1677 		}
1678 
1679 		/* check, if our trim array is full */
1680 		if (unlikely(count >= range_cnt - 1))
1681 			break;
1682 	}
1683 	IWRITE_UNLOCK(ipbmap);
1684 
1685 	tt->nblocks = 0; /* mark the current end */
1686 	for (tt = totrim; tt->nblocks != 0; tt++) {
1687 		/* when mounted with online discard, dbFree() will
1688 		 * call jfs_issue_discard() itself */
1689 		if (!(JFS_SBI(sb)->flag & JFS_DISCARD))
1690 			jfs_issue_discard(ip, tt->blkno, tt->nblocks);
1691 		dbFree(ip, tt->blkno, tt->nblocks);
1692 		trimmed += tt->nblocks;
1693 	}
1694 	kfree(totrim);
1695 
1696 	return trimmed;
1697 }
1698 
1699 /*
1700  * NAME:	dbFindCtl()
1701  *
1702  * FUNCTION:	starting at a specified dmap control page level and block
1703  *		number, search down the dmap control levels for a range of
1704  *		contiguous free blocks large enough to satisfy an allocation
1705  *		request for the specified number of free blocks.
1706  *
1707  *		if sufficient contiguous free blocks are found, this routine
1708  *		returns the starting block number within a dmap page that
1709  *		contains or starts a range of contiqious free blocks that
1710  *		is sufficient in size.
1711  *
1712  * PARAMETERS:
1713  *	bmp	-  pointer to bmap descriptor
1714  *	level	-  starting dmap control page level.
1715  *	l2nb	-  log2 number of contiguous free blocks desired.
1716  *	*blkno	-  on entry, starting block number for conducting the search.
1717  *		   on successful return, the first block within a dmap page
1718  *		   that contains or starts a range of contiguous free blocks.
1719  *
1720  * RETURN VALUES:
1721  *	0	- success
1722  *	-ENOSPC	- insufficient disk resources
1723  *	-EIO	- i/o error
1724  *
1725  * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1726  */
1727 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno)
1728 {
1729 	int rc, leafidx, lev;
1730 	s64 b, lblkno;
1731 	struct dmapctl *dcp;
1732 	int budmin;
1733 	struct metapage *mp;
1734 
1735 	/* starting at the specified dmap control page level and block
1736 	 * number, search down the dmap control levels for the starting
1737 	 * block number of a dmap page that contains or starts off
1738 	 * sufficient free blocks.
1739 	 */
1740 	for (lev = level, b = *blkno; lev >= 0; lev--) {
1741 		/* get the buffer of the dmap control page for the block
1742 		 * number and level (i.e. L0, L1, L2).
1743 		 */
1744 		lblkno = BLKTOCTL(b, bmp->db_l2nbperpage, lev);
1745 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1746 		if (mp == NULL)
1747 			return -EIO;
1748 		dcp = (struct dmapctl *) mp->data;
1749 		budmin = dcp->budmin;
1750 
1751 		if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1752 			jfs_error(bmp->db_ipbmap->i_sb,
1753 				  "Corrupt dmapctl page\n");
1754 			release_metapage(mp);
1755 			return -EIO;
1756 		}
1757 
1758 		/* search the tree within the dmap control page for
1759 		 * sufficient free space.  if sufficient free space is found,
1760 		 * dbFindLeaf() returns the index of the leaf at which
1761 		 * free space was found.
1762 		 */
1763 		rc = dbFindLeaf((dmtree_t *) dcp, l2nb, &leafidx);
1764 
1765 		/* release the buffer.
1766 		 */
1767 		release_metapage(mp);
1768 
1769 		/* space found ?
1770 		 */
1771 		if (rc) {
1772 			if (lev != level) {
1773 				jfs_error(bmp->db_ipbmap->i_sb,
1774 					  "dmap inconsistent\n");
1775 				return -EIO;
1776 			}
1777 			return -ENOSPC;
1778 		}
1779 
1780 		/* adjust the block number to reflect the location within
1781 		 * the dmap control page (i.e. the leaf) at which free
1782 		 * space was found.
1783 		 */
1784 		b += (((s64) leafidx) << budmin);
1785 
1786 		/* we stop the search at this dmap control page level if
1787 		 * the number of blocks required is greater than or equal
1788 		 * to the maximum number of blocks described at the next
1789 		 * (lower) level.
1790 		 */
1791 		if (l2nb >= budmin)
1792 			break;
1793 	}
1794 
1795 	*blkno = b;
1796 	return (0);
1797 }
1798 
1799 
1800 /*
1801  * NAME:	dbAllocCtl()
1802  *
1803  * FUNCTION:	attempt to allocate a specified number of contiguous
1804  *		blocks starting within a specific dmap.
1805  *
1806  *		this routine is called by higher level routines that search
1807  *		the dmap control pages above the actual dmaps for contiguous
1808  *		free space.  the result of successful searches by these
1809  *		routines are the starting block numbers within dmaps, with
1810  *		the dmaps themselves containing the desired contiguous free
1811  *		space or starting a contiguous free space of desired size
1812  *		that is made up of the blocks of one or more dmaps. these
1813  *		calls should not fail due to insufficent resources.
1814  *
1815  *		this routine is called in some cases where it is not known
1816  *		whether it will fail due to insufficient resources.  more
1817  *		specifically, this occurs when allocating from an allocation
1818  *		group whose size is equal to the number of blocks per dmap.
1819  *		in this case, the dmap control pages are not examined prior
1820  *		to calling this routine (to save pathlength) and the call
1821  *		might fail.
1822  *
1823  *		for a request size that fits within a dmap, this routine relies
1824  *		upon the dmap's dmtree to find the requested contiguous free
1825  *		space.  for request sizes that are larger than a dmap, the
1826  *		requested free space will start at the first block of the
1827  *		first dmap (i.e. blkno).
1828  *
1829  * PARAMETERS:
1830  *	bmp	-  pointer to bmap descriptor
1831  *	nblocks	 -  actual number of contiguous free blocks to allocate.
1832  *	l2nb	 -  log2 number of contiguous free blocks to allocate.
1833  *	blkno	 -  starting block number of the dmap to start the allocation
1834  *		    from.
1835  *	results	-  on successful return, set to the starting block number
1836  *		   of the newly allocated range.
1837  *
1838  * RETURN VALUES:
1839  *	0	- success
1840  *	-ENOSPC	- insufficient disk resources
1841  *	-EIO	- i/o error
1842  *
1843  * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1844  */
1845 static int
1846 dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, s64 * results)
1847 {
1848 	int rc, nb;
1849 	s64 b, lblkno, n;
1850 	struct metapage *mp;
1851 	struct dmap *dp;
1852 
1853 	/* check if the allocation request is confined to a single dmap.
1854 	 */
1855 	if (l2nb <= L2BPERDMAP) {
1856 		/* get the buffer for the dmap.
1857 		 */
1858 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
1859 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1860 		if (mp == NULL)
1861 			return -EIO;
1862 		dp = (struct dmap *) mp->data;
1863 
1864 		/* try to allocate the blocks.
1865 		 */
1866 		rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results);
1867 		if (rc == 0)
1868 			mark_metapage_dirty(mp);
1869 
1870 		release_metapage(mp);
1871 
1872 		return (rc);
1873 	}
1874 
1875 	/* allocation request involving multiple dmaps. it must start on
1876 	 * a dmap boundary.
1877 	 */
1878 	assert((blkno & (BPERDMAP - 1)) == 0);
1879 
1880 	/* allocate the blocks dmap by dmap.
1881 	 */
1882 	for (n = nblocks, b = blkno; n > 0; n -= nb, b += nb) {
1883 		/* get the buffer for the dmap.
1884 		 */
1885 		lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1886 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1887 		if (mp == NULL) {
1888 			rc = -EIO;
1889 			goto backout;
1890 		}
1891 		dp = (struct dmap *) mp->data;
1892 
1893 		/* the dmap better be all free.
1894 		 */
1895 		if (dp->tree.stree[ROOT] != L2BPERDMAP) {
1896 			release_metapage(mp);
1897 			jfs_error(bmp->db_ipbmap->i_sb,
1898 				  "the dmap is not all free\n");
1899 			rc = -EIO;
1900 			goto backout;
1901 		}
1902 
1903 		/* determine how many blocks to allocate from this dmap.
1904 		 */
1905 		nb = min_t(s64, n, BPERDMAP);
1906 
1907 		/* allocate the blocks from the dmap.
1908 		 */
1909 		if ((rc = dbAllocDmap(bmp, dp, b, nb))) {
1910 			release_metapage(mp);
1911 			goto backout;
1912 		}
1913 
1914 		/* write the buffer.
1915 		 */
1916 		write_metapage(mp);
1917 	}
1918 
1919 	/* set the results (starting block number) and return.
1920 	 */
1921 	*results = blkno;
1922 	return (0);
1923 
1924 	/* something failed in handling an allocation request involving
1925 	 * multiple dmaps.  we'll try to clean up by backing out any
1926 	 * allocation that has already happened for this request.  if
1927 	 * we fail in backing out the allocation, we'll mark the file
1928 	 * system to indicate that blocks have been leaked.
1929 	 */
1930       backout:
1931 
1932 	/* try to backout the allocations dmap by dmap.
1933 	 */
1934 	for (n = nblocks - n, b = blkno; n > 0;
1935 	     n -= BPERDMAP, b += BPERDMAP) {
1936 		/* get the buffer for this dmap.
1937 		 */
1938 		lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1939 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1940 		if (mp == NULL) {
1941 			/* could not back out.  mark the file system
1942 			 * to indicate that we have leaked blocks.
1943 			 */
1944 			jfs_error(bmp->db_ipbmap->i_sb,
1945 				  "I/O Error: Block Leakage\n");
1946 			continue;
1947 		}
1948 		dp = (struct dmap *) mp->data;
1949 
1950 		/* free the blocks is this dmap.
1951 		 */
1952 		if (dbFreeDmap(bmp, dp, b, BPERDMAP)) {
1953 			/* could not back out.  mark the file system
1954 			 * to indicate that we have leaked blocks.
1955 			 */
1956 			release_metapage(mp);
1957 			jfs_error(bmp->db_ipbmap->i_sb, "Block Leakage\n");
1958 			continue;
1959 		}
1960 
1961 		/* write the buffer.
1962 		 */
1963 		write_metapage(mp);
1964 	}
1965 
1966 	return (rc);
1967 }
1968 
1969 
1970 /*
1971  * NAME:	dbAllocDmapLev()
1972  *
1973  * FUNCTION:	attempt to allocate a specified number of contiguous blocks
1974  *		from a specified dmap.
1975  *
1976  *		this routine checks if the contiguous blocks are available.
1977  *		if so, nblocks of blocks are allocated; otherwise, ENOSPC is
1978  *		returned.
1979  *
1980  * PARAMETERS:
1981  *	mp	-  pointer to bmap descriptor
1982  *	dp	-  pointer to dmap to attempt to allocate blocks from.
1983  *	l2nb	-  log2 number of contiguous block desired.
1984  *	nblocks	-  actual number of contiguous block desired.
1985  *	results	-  on successful return, set to the starting block number
1986  *		   of the newly allocated range.
1987  *
1988  * RETURN VALUES:
1989  *	0	- success
1990  *	-ENOSPC	- insufficient disk resources
1991  *	-EIO	- i/o error
1992  *
1993  * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or
1994  *	IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit;
1995  */
1996 static int
1997 dbAllocDmapLev(struct bmap * bmp,
1998 	       struct dmap * dp, int nblocks, int l2nb, s64 * results)
1999 {
2000 	s64 blkno;
2001 	int leafidx, rc;
2002 
2003 	/* can't be more than a dmaps worth of blocks */
2004 	assert(l2nb <= L2BPERDMAP);
2005 
2006 	/* search the tree within the dmap page for sufficient
2007 	 * free space.  if sufficient free space is found, dbFindLeaf()
2008 	 * returns the index of the leaf at which free space was found.
2009 	 */
2010 	if (dbFindLeaf((dmtree_t *) & dp->tree, l2nb, &leafidx))
2011 		return -ENOSPC;
2012 
2013 	/* determine the block number within the file system corresponding
2014 	 * to the leaf at which free space was found.
2015 	 */
2016 	blkno = le64_to_cpu(dp->start) + (leafidx << L2DBWORD);
2017 
2018 	/* if not all bits of the dmap word are free, get the starting
2019 	 * bit number within the dmap word of the required string of free
2020 	 * bits and adjust the block number with this value.
2021 	 */
2022 	if (dp->tree.stree[leafidx + LEAFIND] < BUDMIN)
2023 		blkno += dbFindBits(le32_to_cpu(dp->wmap[leafidx]), l2nb);
2024 
2025 	/* allocate the blocks */
2026 	if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
2027 		*results = blkno;
2028 
2029 	return (rc);
2030 }
2031 
2032 
2033 /*
2034  * NAME:	dbAllocDmap()
2035  *
2036  * FUNCTION:	adjust the disk allocation map to reflect the allocation
2037  *		of a specified block range within a dmap.
2038  *
2039  *		this routine allocates the specified blocks from the dmap
2040  *		through a call to dbAllocBits(). if the allocation of the
2041  *		block range causes the maximum string of free blocks within
2042  *		the dmap to change (i.e. the value of the root of the dmap's
2043  *		dmtree), this routine will cause this change to be reflected
2044  *		up through the appropriate levels of the dmap control pages
2045  *		by a call to dbAdjCtl() for the L0 dmap control page that
2046  *		covers this dmap.
2047  *
2048  * PARAMETERS:
2049  *	bmp	-  pointer to bmap descriptor
2050  *	dp	-  pointer to dmap to allocate the block range from.
2051  *	blkno	-  starting block number of the block to be allocated.
2052  *	nblocks	-  number of blocks to be allocated.
2053  *
2054  * RETURN VALUES:
2055  *	0	- success
2056  *	-EIO	- i/o error
2057  *
2058  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2059  */
2060 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2061 		       int nblocks)
2062 {
2063 	s8 oldroot;
2064 	int rc;
2065 
2066 	/* save the current value of the root (i.e. maximum free string)
2067 	 * of the dmap tree.
2068 	 */
2069 	oldroot = dp->tree.stree[ROOT];
2070 
2071 	/* allocate the specified (blocks) bits */
2072 	dbAllocBits(bmp, dp, blkno, nblocks);
2073 
2074 	/* if the root has not changed, done. */
2075 	if (dp->tree.stree[ROOT] == oldroot)
2076 		return (0);
2077 
2078 	/* root changed. bubble the change up to the dmap control pages.
2079 	 * if the adjustment of the upper level control pages fails,
2080 	 * backout the bit allocation (thus making everything consistent).
2081 	 */
2082 	if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 1, 0)))
2083 		dbFreeBits(bmp, dp, blkno, nblocks);
2084 
2085 	return (rc);
2086 }
2087 
2088 
2089 /*
2090  * NAME:	dbFreeDmap()
2091  *
2092  * FUNCTION:	adjust the disk allocation map to reflect the allocation
2093  *		of a specified block range within a dmap.
2094  *
2095  *		this routine frees the specified blocks from the dmap through
2096  *		a call to dbFreeBits(). if the deallocation of the block range
2097  *		causes the maximum string of free blocks within the dmap to
2098  *		change (i.e. the value of the root of the dmap's dmtree), this
2099  *		routine will cause this change to be reflected up through the
2100  *		appropriate levels of the dmap control pages by a call to
2101  *		dbAdjCtl() for the L0 dmap control page that covers this dmap.
2102  *
2103  * PARAMETERS:
2104  *	bmp	-  pointer to bmap descriptor
2105  *	dp	-  pointer to dmap to free the block range from.
2106  *	blkno	-  starting block number of the block to be freed.
2107  *	nblocks	-  number of blocks to be freed.
2108  *
2109  * RETURN VALUES:
2110  *	0	- success
2111  *	-EIO	- i/o error
2112  *
2113  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2114  */
2115 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2116 		      int nblocks)
2117 {
2118 	s8 oldroot;
2119 	int rc = 0, word;
2120 
2121 	/* save the current value of the root (i.e. maximum free string)
2122 	 * of the dmap tree.
2123 	 */
2124 	oldroot = dp->tree.stree[ROOT];
2125 
2126 	/* free the specified (blocks) bits */
2127 	rc = dbFreeBits(bmp, dp, blkno, nblocks);
2128 
2129 	/* if error or the root has not changed, done. */
2130 	if (rc || (dp->tree.stree[ROOT] == oldroot))
2131 		return (rc);
2132 
2133 	/* root changed. bubble the change up to the dmap control pages.
2134 	 * if the adjustment of the upper level control pages fails,
2135 	 * backout the deallocation.
2136 	 */
2137 	if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 0, 0))) {
2138 		word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
2139 
2140 		/* as part of backing out the deallocation, we will have
2141 		 * to back split the dmap tree if the deallocation caused
2142 		 * the freed blocks to become part of a larger binary buddy
2143 		 * system.
2144 		 */
2145 		if (dp->tree.stree[word] == NOFREE)
2146 			dbBackSplit((dmtree_t *) & dp->tree, word);
2147 
2148 		dbAllocBits(bmp, dp, blkno, nblocks);
2149 	}
2150 
2151 	return (rc);
2152 }
2153 
2154 
2155 /*
2156  * NAME:	dbAllocBits()
2157  *
2158  * FUNCTION:	allocate a specified block range from a dmap.
2159  *
2160  *		this routine updates the dmap to reflect the working
2161  *		state allocation of the specified block range. it directly
2162  *		updates the bits of the working map and causes the adjustment
2163  *		of the binary buddy system described by the dmap's dmtree
2164  *		leaves to reflect the bits allocated.  it also causes the
2165  *		dmap's dmtree, as a whole, to reflect the allocated range.
2166  *
2167  * PARAMETERS:
2168  *	bmp	-  pointer to bmap descriptor
2169  *	dp	-  pointer to dmap to allocate bits from.
2170  *	blkno	-  starting block number of the bits to be allocated.
2171  *	nblocks	-  number of bits to be allocated.
2172  *
2173  * RETURN VALUES: none
2174  *
2175  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2176  */
2177 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2178 			int nblocks)
2179 {
2180 	int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2181 	dmtree_t *tp = (dmtree_t *) & dp->tree;
2182 	int size;
2183 	s8 *leaf;
2184 
2185 	/* pick up a pointer to the leaves of the dmap tree */
2186 	leaf = dp->tree.stree + LEAFIND;
2187 
2188 	/* determine the bit number and word within the dmap of the
2189 	 * starting block.
2190 	 */
2191 	dbitno = blkno & (BPERDMAP - 1);
2192 	word = dbitno >> L2DBWORD;
2193 
2194 	/* block range better be within the dmap */
2195 	assert(dbitno + nblocks <= BPERDMAP);
2196 
2197 	/* allocate the bits of the dmap's words corresponding to the block
2198 	 * range. not all bits of the first and last words may be contained
2199 	 * within the block range.  if this is the case, we'll work against
2200 	 * those words (i.e. partial first and/or last) on an individual basis
2201 	 * (a single pass), allocating the bits of interest by hand and
2202 	 * updating the leaf corresponding to the dmap word. a single pass
2203 	 * will be used for all dmap words fully contained within the
2204 	 * specified range.  within this pass, the bits of all fully contained
2205 	 * dmap words will be marked as free in a single shot and the leaves
2206 	 * will be updated. a single leaf may describe the free space of
2207 	 * multiple dmap words, so we may update only a subset of the actual
2208 	 * leaves corresponding to the dmap words of the block range.
2209 	 */
2210 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2211 		/* determine the bit number within the word and
2212 		 * the number of bits within the word.
2213 		 */
2214 		wbitno = dbitno & (DBWORD - 1);
2215 		nb = min(rembits, DBWORD - wbitno);
2216 
2217 		/* check if only part of a word is to be allocated.
2218 		 */
2219 		if (nb < DBWORD) {
2220 			/* allocate (set to 1) the appropriate bits within
2221 			 * this dmap word.
2222 			 */
2223 			dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
2224 						      >> wbitno);
2225 
2226 			/* update the leaf for this dmap word. in addition
2227 			 * to setting the leaf value to the binary buddy max
2228 			 * of the updated dmap word, dbSplit() will split
2229 			 * the binary system of the leaves if need be.
2230 			 */
2231 			dbSplit(tp, word, BUDMIN,
2232 				dbMaxBud((u8 *) & dp->wmap[word]));
2233 
2234 			word += 1;
2235 		} else {
2236 			/* one or more dmap words are fully contained
2237 			 * within the block range.  determine how many
2238 			 * words and allocate (set to 1) the bits of these
2239 			 * words.
2240 			 */
2241 			nwords = rembits >> L2DBWORD;
2242 			memset(&dp->wmap[word], (int) ONES, nwords * 4);
2243 
2244 			/* determine how many bits.
2245 			 */
2246 			nb = nwords << L2DBWORD;
2247 
2248 			/* now update the appropriate leaves to reflect
2249 			 * the allocated words.
2250 			 */
2251 			for (; nwords > 0; nwords -= nw) {
2252 				if (leaf[word] < BUDMIN) {
2253 					jfs_error(bmp->db_ipbmap->i_sb,
2254 						  "leaf page corrupt\n");
2255 					break;
2256 				}
2257 
2258 				/* determine what the leaf value should be
2259 				 * updated to as the minimum of the l2 number
2260 				 * of bits being allocated and the l2 number
2261 				 * of bits currently described by this leaf.
2262 				 */
2263 				size = min_t(int, leaf[word],
2264 					     NLSTOL2BSZ(nwords));
2265 
2266 				/* update the leaf to reflect the allocation.
2267 				 * in addition to setting the leaf value to
2268 				 * NOFREE, dbSplit() will split the binary
2269 				 * system of the leaves to reflect the current
2270 				 * allocation (size).
2271 				 */
2272 				dbSplit(tp, word, size, NOFREE);
2273 
2274 				/* get the number of dmap words handled */
2275 				nw = BUDSIZE(size, BUDMIN);
2276 				word += nw;
2277 			}
2278 		}
2279 	}
2280 
2281 	/* update the free count for this dmap */
2282 	le32_add_cpu(&dp->nfree, -nblocks);
2283 
2284 	BMAP_LOCK(bmp);
2285 
2286 	/* if this allocation group is completely free,
2287 	 * update the maximum allocation group number if this allocation
2288 	 * group is the new max.
2289 	 */
2290 	agno = blkno >> bmp->db_agl2size;
2291 	if (agno > bmp->db_maxag)
2292 		bmp->db_maxag = agno;
2293 
2294 	/* update the free count for the allocation group and map */
2295 	bmp->db_agfree[agno] -= nblocks;
2296 	bmp->db_nfree -= nblocks;
2297 
2298 	BMAP_UNLOCK(bmp);
2299 }
2300 
2301 
2302 /*
2303  * NAME:	dbFreeBits()
2304  *
2305  * FUNCTION:	free a specified block range from a dmap.
2306  *
2307  *		this routine updates the dmap to reflect the working
2308  *		state allocation of the specified block range. it directly
2309  *		updates the bits of the working map and causes the adjustment
2310  *		of the binary buddy system described by the dmap's dmtree
2311  *		leaves to reflect the bits freed.  it also causes the dmap's
2312  *		dmtree, as a whole, to reflect the deallocated range.
2313  *
2314  * PARAMETERS:
2315  *	bmp	-  pointer to bmap descriptor
2316  *	dp	-  pointer to dmap to free bits from.
2317  *	blkno	-  starting block number of the bits to be freed.
2318  *	nblocks	-  number of bits to be freed.
2319  *
2320  * RETURN VALUES: 0 for success
2321  *
2322  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2323  */
2324 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2325 		       int nblocks)
2326 {
2327 	int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2328 	dmtree_t *tp = (dmtree_t *) & dp->tree;
2329 	int rc = 0;
2330 	int size;
2331 
2332 	/* determine the bit number and word within the dmap of the
2333 	 * starting block.
2334 	 */
2335 	dbitno = blkno & (BPERDMAP - 1);
2336 	word = dbitno >> L2DBWORD;
2337 
2338 	/* block range better be within the dmap.
2339 	 */
2340 	assert(dbitno + nblocks <= BPERDMAP);
2341 
2342 	/* free the bits of the dmaps words corresponding to the block range.
2343 	 * not all bits of the first and last words may be contained within
2344 	 * the block range.  if this is the case, we'll work against those
2345 	 * words (i.e. partial first and/or last) on an individual basis
2346 	 * (a single pass), freeing the bits of interest by hand and updating
2347 	 * the leaf corresponding to the dmap word. a single pass will be used
2348 	 * for all dmap words fully contained within the specified range.
2349 	 * within this pass, the bits of all fully contained dmap words will
2350 	 * be marked as free in a single shot and the leaves will be updated. a
2351 	 * single leaf may describe the free space of multiple dmap words,
2352 	 * so we may update only a subset of the actual leaves corresponding
2353 	 * to the dmap words of the block range.
2354 	 *
2355 	 * dbJoin() is used to update leaf values and will join the binary
2356 	 * buddy system of the leaves if the new leaf values indicate this
2357 	 * should be done.
2358 	 */
2359 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2360 		/* determine the bit number within the word and
2361 		 * the number of bits within the word.
2362 		 */
2363 		wbitno = dbitno & (DBWORD - 1);
2364 		nb = min(rembits, DBWORD - wbitno);
2365 
2366 		/* check if only part of a word is to be freed.
2367 		 */
2368 		if (nb < DBWORD) {
2369 			/* free (zero) the appropriate bits within this
2370 			 * dmap word.
2371 			 */
2372 			dp->wmap[word] &=
2373 			    cpu_to_le32(~(ONES << (DBWORD - nb)
2374 					  >> wbitno));
2375 
2376 			/* update the leaf for this dmap word.
2377 			 */
2378 			rc = dbJoin(tp, word,
2379 				    dbMaxBud((u8 *) & dp->wmap[word]));
2380 			if (rc)
2381 				return rc;
2382 
2383 			word += 1;
2384 		} else {
2385 			/* one or more dmap words are fully contained
2386 			 * within the block range.  determine how many
2387 			 * words and free (zero) the bits of these words.
2388 			 */
2389 			nwords = rembits >> L2DBWORD;
2390 			memset(&dp->wmap[word], 0, nwords * 4);
2391 
2392 			/* determine how many bits.
2393 			 */
2394 			nb = nwords << L2DBWORD;
2395 
2396 			/* now update the appropriate leaves to reflect
2397 			 * the freed words.
2398 			 */
2399 			for (; nwords > 0; nwords -= nw) {
2400 				/* determine what the leaf value should be
2401 				 * updated to as the minimum of the l2 number
2402 				 * of bits being freed and the l2 (max) number
2403 				 * of bits that can be described by this leaf.
2404 				 */
2405 				size =
2406 				    min(LITOL2BSZ
2407 					(word, L2LPERDMAP, BUDMIN),
2408 					NLSTOL2BSZ(nwords));
2409 
2410 				/* update the leaf.
2411 				 */
2412 				rc = dbJoin(tp, word, size);
2413 				if (rc)
2414 					return rc;
2415 
2416 				/* get the number of dmap words handled.
2417 				 */
2418 				nw = BUDSIZE(size, BUDMIN);
2419 				word += nw;
2420 			}
2421 		}
2422 	}
2423 
2424 	/* update the free count for this dmap.
2425 	 */
2426 	le32_add_cpu(&dp->nfree, nblocks);
2427 
2428 	BMAP_LOCK(bmp);
2429 
2430 	/* update the free count for the allocation group and
2431 	 * map.
2432 	 */
2433 	agno = blkno >> bmp->db_agl2size;
2434 	bmp->db_nfree += nblocks;
2435 	bmp->db_agfree[agno] += nblocks;
2436 
2437 	/* check if this allocation group is not completely free and
2438 	 * if it is currently the maximum (rightmost) allocation group.
2439 	 * if so, establish the new maximum allocation group number by
2440 	 * searching left for the first allocation group with allocation.
2441 	 */
2442 	if ((bmp->db_agfree[agno] == bmp->db_agsize && agno == bmp->db_maxag) ||
2443 	    (agno == bmp->db_numag - 1 &&
2444 	     bmp->db_agfree[agno] == (bmp-> db_mapsize & (BPERDMAP - 1)))) {
2445 		while (bmp->db_maxag > 0) {
2446 			bmp->db_maxag -= 1;
2447 			if (bmp->db_agfree[bmp->db_maxag] !=
2448 			    bmp->db_agsize)
2449 				break;
2450 		}
2451 
2452 		/* re-establish the allocation group preference if the
2453 		 * current preference is right of the maximum allocation
2454 		 * group.
2455 		 */
2456 		if (bmp->db_agpref > bmp->db_maxag)
2457 			bmp->db_agpref = bmp->db_maxag;
2458 	}
2459 
2460 	BMAP_UNLOCK(bmp);
2461 
2462 	return 0;
2463 }
2464 
2465 
2466 /*
2467  * NAME:	dbAdjCtl()
2468  *
2469  * FUNCTION:	adjust a dmap control page at a specified level to reflect
2470  *		the change in a lower level dmap or dmap control page's
2471  *		maximum string of free blocks (i.e. a change in the root
2472  *		of the lower level object's dmtree) due to the allocation
2473  *		or deallocation of a range of blocks with a single dmap.
2474  *
2475  *		on entry, this routine is provided with the new value of
2476  *		the lower level dmap or dmap control page root and the
2477  *		starting block number of the block range whose allocation
2478  *		or deallocation resulted in the root change.  this range
2479  *		is respresented by a single leaf of the current dmapctl
2480  *		and the leaf will be updated with this value, possibly
2481  *		causing a binary buddy system within the leaves to be
2482  *		split or joined.  the update may also cause the dmapctl's
2483  *		dmtree to be updated.
2484  *
2485  *		if the adjustment of the dmap control page, itself, causes its
2486  *		root to change, this change will be bubbled up to the next dmap
2487  *		control level by a recursive call to this routine, specifying
2488  *		the new root value and the next dmap control page level to
2489  *		be adjusted.
2490  * PARAMETERS:
2491  *	bmp	-  pointer to bmap descriptor
2492  *	blkno	-  the first block of a block range within a dmap.  it is
2493  *		   the allocation or deallocation of this block range that
2494  *		   requires the dmap control page to be adjusted.
2495  *	newval	-  the new value of the lower level dmap or dmap control
2496  *		   page root.
2497  *	alloc	-  'true' if adjustment is due to an allocation.
2498  *	level	-  current level of dmap control page (i.e. L0, L1, L2) to
2499  *		   be adjusted.
2500  *
2501  * RETURN VALUES:
2502  *	0	- success
2503  *	-EIO	- i/o error
2504  *
2505  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2506  */
2507 static int
2508 dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, int level)
2509 {
2510 	struct metapage *mp;
2511 	s8 oldroot;
2512 	int oldval;
2513 	s64 lblkno;
2514 	struct dmapctl *dcp;
2515 	int rc, leafno, ti;
2516 
2517 	/* get the buffer for the dmap control page for the specified
2518 	 * block number and control page level.
2519 	 */
2520 	lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, level);
2521 	mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
2522 	if (mp == NULL)
2523 		return -EIO;
2524 	dcp = (struct dmapctl *) mp->data;
2525 
2526 	if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
2527 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
2528 		release_metapage(mp);
2529 		return -EIO;
2530 	}
2531 
2532 	/* determine the leaf number corresponding to the block and
2533 	 * the index within the dmap control tree.
2534 	 */
2535 	leafno = BLKTOCTLLEAF(blkno, dcp->budmin);
2536 	ti = leafno + le32_to_cpu(dcp->leafidx);
2537 
2538 	/* save the current leaf value and the current root level (i.e.
2539 	 * maximum l2 free string described by this dmapctl).
2540 	 */
2541 	oldval = dcp->stree[ti];
2542 	oldroot = dcp->stree[ROOT];
2543 
2544 	/* check if this is a control page update for an allocation.
2545 	 * if so, update the leaf to reflect the new leaf value using
2546 	 * dbSplit(); otherwise (deallocation), use dbJoin() to update
2547 	 * the leaf with the new value.  in addition to updating the
2548 	 * leaf, dbSplit() will also split the binary buddy system of
2549 	 * the leaves, if required, and bubble new values within the
2550 	 * dmapctl tree, if required.  similarly, dbJoin() will join
2551 	 * the binary buddy system of leaves and bubble new values up
2552 	 * the dmapctl tree as required by the new leaf value.
2553 	 */
2554 	if (alloc) {
2555 		/* check if we are in the middle of a binary buddy
2556 		 * system.  this happens when we are performing the
2557 		 * first allocation out of an allocation group that
2558 		 * is part (not the first part) of a larger binary
2559 		 * buddy system.  if we are in the middle, back split
2560 		 * the system prior to calling dbSplit() which assumes
2561 		 * that it is at the front of a binary buddy system.
2562 		 */
2563 		if (oldval == NOFREE) {
2564 			rc = dbBackSplit((dmtree_t *) dcp, leafno);
2565 			if (rc)
2566 				return rc;
2567 			oldval = dcp->stree[ti];
2568 		}
2569 		dbSplit((dmtree_t *) dcp, leafno, dcp->budmin, newval);
2570 	} else {
2571 		rc = dbJoin((dmtree_t *) dcp, leafno, newval);
2572 		if (rc)
2573 			return rc;
2574 	}
2575 
2576 	/* check if the root of the current dmap control page changed due
2577 	 * to the update and if the current dmap control page is not at
2578 	 * the current top level (i.e. L0, L1, L2) of the map.  if so (i.e.
2579 	 * root changed and this is not the top level), call this routine
2580 	 * again (recursion) for the next higher level of the mapping to
2581 	 * reflect the change in root for the current dmap control page.
2582 	 */
2583 	if (dcp->stree[ROOT] != oldroot) {
2584 		/* are we below the top level of the map.  if so,
2585 		 * bubble the root up to the next higher level.
2586 		 */
2587 		if (level < bmp->db_maxlevel) {
2588 			/* bubble up the new root of this dmap control page to
2589 			 * the next level.
2590 			 */
2591 			if ((rc =
2592 			     dbAdjCtl(bmp, blkno, dcp->stree[ROOT], alloc,
2593 				      level + 1))) {
2594 				/* something went wrong in bubbling up the new
2595 				 * root value, so backout the changes to the
2596 				 * current dmap control page.
2597 				 */
2598 				if (alloc) {
2599 					dbJoin((dmtree_t *) dcp, leafno,
2600 					       oldval);
2601 				} else {
2602 					/* the dbJoin() above might have
2603 					 * caused a larger binary buddy system
2604 					 * to form and we may now be in the
2605 					 * middle of it.  if this is the case,
2606 					 * back split the buddies.
2607 					 */
2608 					if (dcp->stree[ti] == NOFREE)
2609 						dbBackSplit((dmtree_t *)
2610 							    dcp, leafno);
2611 					dbSplit((dmtree_t *) dcp, leafno,
2612 						dcp->budmin, oldval);
2613 				}
2614 
2615 				/* release the buffer and return the error.
2616 				 */
2617 				release_metapage(mp);
2618 				return (rc);
2619 			}
2620 		} else {
2621 			/* we're at the top level of the map. update
2622 			 * the bmap control page to reflect the size
2623 			 * of the maximum free buddy system.
2624 			 */
2625 			assert(level == bmp->db_maxlevel);
2626 			if (bmp->db_maxfreebud != oldroot) {
2627 				jfs_error(bmp->db_ipbmap->i_sb,
2628 					  "the maximum free buddy is not the old root\n");
2629 			}
2630 			bmp->db_maxfreebud = dcp->stree[ROOT];
2631 		}
2632 	}
2633 
2634 	/* write the buffer.
2635 	 */
2636 	write_metapage(mp);
2637 
2638 	return (0);
2639 }
2640 
2641 
2642 /*
2643  * NAME:	dbSplit()
2644  *
2645  * FUNCTION:	update the leaf of a dmtree with a new value, splitting
2646  *		the leaf from the binary buddy system of the dmtree's
2647  *		leaves, as required.
2648  *
2649  * PARAMETERS:
2650  *	tp	- pointer to the tree containing the leaf.
2651  *	leafno	- the number of the leaf to be updated.
2652  *	splitsz	- the size the binary buddy system starting at the leaf
2653  *		  must be split to, specified as the log2 number of blocks.
2654  *	newval	- the new value for the leaf.
2655  *
2656  * RETURN VALUES: none
2657  *
2658  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2659  */
2660 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval)
2661 {
2662 	int budsz;
2663 	int cursz;
2664 	s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2665 
2666 	/* check if the leaf needs to be split.
2667 	 */
2668 	if (leaf[leafno] > tp->dmt_budmin) {
2669 		/* the split occurs by cutting the buddy system in half
2670 		 * at the specified leaf until we reach the specified
2671 		 * size.  pick up the starting split size (current size
2672 		 * - 1 in l2) and the corresponding buddy size.
2673 		 */
2674 		cursz = leaf[leafno] - 1;
2675 		budsz = BUDSIZE(cursz, tp->dmt_budmin);
2676 
2677 		/* split until we reach the specified size.
2678 		 */
2679 		while (cursz >= splitsz) {
2680 			/* update the buddy's leaf with its new value.
2681 			 */
2682 			dbAdjTree(tp, leafno ^ budsz, cursz);
2683 
2684 			/* on to the next size and buddy.
2685 			 */
2686 			cursz -= 1;
2687 			budsz >>= 1;
2688 		}
2689 	}
2690 
2691 	/* adjust the dmap tree to reflect the specified leaf's new
2692 	 * value.
2693 	 */
2694 	dbAdjTree(tp, leafno, newval);
2695 }
2696 
2697 
2698 /*
2699  * NAME:	dbBackSplit()
2700  *
2701  * FUNCTION:	back split the binary buddy system of dmtree leaves
2702  *		that hold a specified leaf until the specified leaf
2703  *		starts its own binary buddy system.
2704  *
2705  *		the allocators typically perform allocations at the start
2706  *		of binary buddy systems and dbSplit() is used to accomplish
2707  *		any required splits.  in some cases, however, allocation
2708  *		may occur in the middle of a binary system and requires a
2709  *		back split, with the split proceeding out from the middle of
2710  *		the system (less efficient) rather than the start of the
2711  *		system (more efficient).  the cases in which a back split
2712  *		is required are rare and are limited to the first allocation
2713  *		within an allocation group which is a part (not first part)
2714  *		of a larger binary buddy system and a few exception cases
2715  *		in which a previous join operation must be backed out.
2716  *
2717  * PARAMETERS:
2718  *	tp	- pointer to the tree containing the leaf.
2719  *	leafno	- the number of the leaf to be updated.
2720  *
2721  * RETURN VALUES: none
2722  *
2723  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2724  */
2725 static int dbBackSplit(dmtree_t * tp, int leafno)
2726 {
2727 	int budsz, bud, w, bsz, size;
2728 	int cursz;
2729 	s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2730 
2731 	/* leaf should be part (not first part) of a binary
2732 	 * buddy system.
2733 	 */
2734 	assert(leaf[leafno] == NOFREE);
2735 
2736 	/* the back split is accomplished by iteratively finding the leaf
2737 	 * that starts the buddy system that contains the specified leaf and
2738 	 * splitting that system in two.  this iteration continues until
2739 	 * the specified leaf becomes the start of a buddy system.
2740 	 *
2741 	 * determine maximum possible l2 size for the specified leaf.
2742 	 */
2743 	size =
2744 	    LITOL2BSZ(leafno, le32_to_cpu(tp->dmt_l2nleafs),
2745 		      tp->dmt_budmin);
2746 
2747 	/* determine the number of leaves covered by this size.  this
2748 	 * is the buddy size that we will start with as we search for
2749 	 * the buddy system that contains the specified leaf.
2750 	 */
2751 	budsz = BUDSIZE(size, tp->dmt_budmin);
2752 
2753 	/* back split.
2754 	 */
2755 	while (leaf[leafno] == NOFREE) {
2756 		/* find the leftmost buddy leaf.
2757 		 */
2758 		for (w = leafno, bsz = budsz;; bsz <<= 1,
2759 		     w = (w < bud) ? w : bud) {
2760 			if (bsz >= le32_to_cpu(tp->dmt_nleafs)) {
2761 				jfs_err("JFS: block map error in dbBackSplit");
2762 				return -EIO;
2763 			}
2764 
2765 			/* determine the buddy.
2766 			 */
2767 			bud = w ^ bsz;
2768 
2769 			/* check if this buddy is the start of the system.
2770 			 */
2771 			if (leaf[bud] != NOFREE) {
2772 				/* split the leaf at the start of the
2773 				 * system in two.
2774 				 */
2775 				cursz = leaf[bud] - 1;
2776 				dbSplit(tp, bud, cursz, cursz);
2777 				break;
2778 			}
2779 		}
2780 	}
2781 
2782 	if (leaf[leafno] != size) {
2783 		jfs_err("JFS: wrong leaf value in dbBackSplit");
2784 		return -EIO;
2785 	}
2786 	return 0;
2787 }
2788 
2789 
2790 /*
2791  * NAME:	dbJoin()
2792  *
2793  * FUNCTION:	update the leaf of a dmtree with a new value, joining
2794  *		the leaf with other leaves of the dmtree into a multi-leaf
2795  *		binary buddy system, as required.
2796  *
2797  * PARAMETERS:
2798  *	tp	- pointer to the tree containing the leaf.
2799  *	leafno	- the number of the leaf to be updated.
2800  *	newval	- the new value for the leaf.
2801  *
2802  * RETURN VALUES: none
2803  */
2804 static int dbJoin(dmtree_t * tp, int leafno, int newval)
2805 {
2806 	int budsz, buddy;
2807 	s8 *leaf;
2808 
2809 	/* can the new leaf value require a join with other leaves ?
2810 	 */
2811 	if (newval >= tp->dmt_budmin) {
2812 		/* pickup a pointer to the leaves of the tree.
2813 		 */
2814 		leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2815 
2816 		/* try to join the specified leaf into a large binary
2817 		 * buddy system.  the join proceeds by attempting to join
2818 		 * the specified leafno with its buddy (leaf) at new value.
2819 		 * if the join occurs, we attempt to join the left leaf
2820 		 * of the joined buddies with its buddy at new value + 1.
2821 		 * we continue to join until we find a buddy that cannot be
2822 		 * joined (does not have a value equal to the size of the
2823 		 * last join) or until all leaves have been joined into a
2824 		 * single system.
2825 		 *
2826 		 * get the buddy size (number of words covered) of
2827 		 * the new value.
2828 		 */
2829 		budsz = BUDSIZE(newval, tp->dmt_budmin);
2830 
2831 		/* try to join.
2832 		 */
2833 		while (budsz < le32_to_cpu(tp->dmt_nleafs)) {
2834 			/* get the buddy leaf.
2835 			 */
2836 			buddy = leafno ^ budsz;
2837 
2838 			/* if the leaf's new value is greater than its
2839 			 * buddy's value, we join no more.
2840 			 */
2841 			if (newval > leaf[buddy])
2842 				break;
2843 
2844 			/* It shouldn't be less */
2845 			if (newval < leaf[buddy])
2846 				return -EIO;
2847 
2848 			/* check which (leafno or buddy) is the left buddy.
2849 			 * the left buddy gets to claim the blocks resulting
2850 			 * from the join while the right gets to claim none.
2851 			 * the left buddy is also eligible to participate in
2852 			 * a join at the next higher level while the right
2853 			 * is not.
2854 			 *
2855 			 */
2856 			if (leafno < buddy) {
2857 				/* leafno is the left buddy.
2858 				 */
2859 				dbAdjTree(tp, buddy, NOFREE);
2860 			} else {
2861 				/* buddy is the left buddy and becomes
2862 				 * leafno.
2863 				 */
2864 				dbAdjTree(tp, leafno, NOFREE);
2865 				leafno = buddy;
2866 			}
2867 
2868 			/* on to try the next join.
2869 			 */
2870 			newval += 1;
2871 			budsz <<= 1;
2872 		}
2873 	}
2874 
2875 	/* update the leaf value.
2876 	 */
2877 	dbAdjTree(tp, leafno, newval);
2878 
2879 	return 0;
2880 }
2881 
2882 
2883 /*
2884  * NAME:	dbAdjTree()
2885  *
2886  * FUNCTION:	update a leaf of a dmtree with a new value, adjusting
2887  *		the dmtree, as required, to reflect the new leaf value.
2888  *		the combination of any buddies must already be done before
2889  *		this is called.
2890  *
2891  * PARAMETERS:
2892  *	tp	- pointer to the tree to be adjusted.
2893  *	leafno	- the number of the leaf to be updated.
2894  *	newval	- the new value for the leaf.
2895  *
2896  * RETURN VALUES: none
2897  */
2898 static void dbAdjTree(dmtree_t * tp, int leafno, int newval)
2899 {
2900 	int lp, pp, k;
2901 	int max;
2902 
2903 	/* pick up the index of the leaf for this leafno.
2904 	 */
2905 	lp = leafno + le32_to_cpu(tp->dmt_leafidx);
2906 
2907 	/* is the current value the same as the old value ?  if so,
2908 	 * there is nothing to do.
2909 	 */
2910 	if (tp->dmt_stree[lp] == newval)
2911 		return;
2912 
2913 	/* set the new value.
2914 	 */
2915 	tp->dmt_stree[lp] = newval;
2916 
2917 	/* bubble the new value up the tree as required.
2918 	 */
2919 	for (k = 0; k < le32_to_cpu(tp->dmt_height); k++) {
2920 		/* get the index of the first leaf of the 4 leaf
2921 		 * group containing the specified leaf (leafno).
2922 		 */
2923 		lp = ((lp - 1) & ~0x03) + 1;
2924 
2925 		/* get the index of the parent of this 4 leaf group.
2926 		 */
2927 		pp = (lp - 1) >> 2;
2928 
2929 		/* determine the maximum of the 4 leaves.
2930 		 */
2931 		max = TREEMAX(&tp->dmt_stree[lp]);
2932 
2933 		/* if the maximum of the 4 is the same as the
2934 		 * parent's value, we're done.
2935 		 */
2936 		if (tp->dmt_stree[pp] == max)
2937 			break;
2938 
2939 		/* parent gets new value.
2940 		 */
2941 		tp->dmt_stree[pp] = max;
2942 
2943 		/* parent becomes leaf for next go-round.
2944 		 */
2945 		lp = pp;
2946 	}
2947 }
2948 
2949 
2950 /*
2951  * NAME:	dbFindLeaf()
2952  *
2953  * FUNCTION:	search a dmtree_t for sufficient free blocks, returning
2954  *		the index of a leaf describing the free blocks if
2955  *		sufficient free blocks are found.
2956  *
2957  *		the search starts at the top of the dmtree_t tree and
2958  *		proceeds down the tree to the leftmost leaf with sufficient
2959  *		free space.
2960  *
2961  * PARAMETERS:
2962  *	tp	- pointer to the tree to be searched.
2963  *	l2nb	- log2 number of free blocks to search for.
2964  *	leafidx	- return pointer to be set to the index of the leaf
2965  *		  describing at least l2nb free blocks if sufficient
2966  *		  free blocks are found.
2967  *
2968  * RETURN VALUES:
2969  *	0	- success
2970  *	-ENOSPC	- insufficient free blocks.
2971  */
2972 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx)
2973 {
2974 	int ti, n = 0, k, x = 0;
2975 
2976 	/* first check the root of the tree to see if there is
2977 	 * sufficient free space.
2978 	 */
2979 	if (l2nb > tp->dmt_stree[ROOT])
2980 		return -ENOSPC;
2981 
2982 	/* sufficient free space available. now search down the tree
2983 	 * starting at the next level for the leftmost leaf that
2984 	 * describes sufficient free space.
2985 	 */
2986 	for (k = le32_to_cpu(tp->dmt_height), ti = 1;
2987 	     k > 0; k--, ti = ((ti + n) << 2) + 1) {
2988 		/* search the four nodes at this level, starting from
2989 		 * the left.
2990 		 */
2991 		for (x = ti, n = 0; n < 4; n++) {
2992 			/* sufficient free space found.  move to the next
2993 			 * level (or quit if this is the last level).
2994 			 */
2995 			if (l2nb <= tp->dmt_stree[x + n])
2996 				break;
2997 		}
2998 
2999 		/* better have found something since the higher
3000 		 * levels of the tree said it was here.
3001 		 */
3002 		assert(n < 4);
3003 	}
3004 
3005 	/* set the return to the leftmost leaf describing sufficient
3006 	 * free space.
3007 	 */
3008 	*leafidx = x + n - le32_to_cpu(tp->dmt_leafidx);
3009 
3010 	return (0);
3011 }
3012 
3013 
3014 /*
3015  * NAME:	dbFindBits()
3016  *
3017  * FUNCTION:	find a specified number of binary buddy free bits within a
3018  *		dmap bitmap word value.
3019  *
3020  *		this routine searches the bitmap value for (1 << l2nb) free
3021  *		bits at (1 << l2nb) alignments within the value.
3022  *
3023  * PARAMETERS:
3024  *	word	-  dmap bitmap word value.
3025  *	l2nb	-  number of free bits specified as a log2 number.
3026  *
3027  * RETURN VALUES:
3028  *	starting bit number of free bits.
3029  */
3030 static int dbFindBits(u32 word, int l2nb)
3031 {
3032 	int bitno, nb;
3033 	u32 mask;
3034 
3035 	/* get the number of bits.
3036 	 */
3037 	nb = 1 << l2nb;
3038 	assert(nb <= DBWORD);
3039 
3040 	/* complement the word so we can use a mask (i.e. 0s represent
3041 	 * free bits) and compute the mask.
3042 	 */
3043 	word = ~word;
3044 	mask = ONES << (DBWORD - nb);
3045 
3046 	/* scan the word for nb free bits at nb alignments.
3047 	 */
3048 	for (bitno = 0; mask != 0; bitno += nb, mask >>= nb) {
3049 		if ((mask & word) == mask)
3050 			break;
3051 	}
3052 
3053 	ASSERT(bitno < 32);
3054 
3055 	/* return the bit number.
3056 	 */
3057 	return (bitno);
3058 }
3059 
3060 
3061 /*
3062  * NAME:	dbMaxBud(u8 *cp)
3063  *
3064  * FUNCTION:	determine the largest binary buddy string of free
3065  *		bits within 32-bits of the map.
3066  *
3067  * PARAMETERS:
3068  *	cp	-  pointer to the 32-bit value.
3069  *
3070  * RETURN VALUES:
3071  *	largest binary buddy of free bits within a dmap word.
3072  */
3073 static int dbMaxBud(u8 * cp)
3074 {
3075 	signed char tmp1, tmp2;
3076 
3077 	/* check if the wmap word is all free. if so, the
3078 	 * free buddy size is BUDMIN.
3079 	 */
3080 	if (*((uint *) cp) == 0)
3081 		return (BUDMIN);
3082 
3083 	/* check if the wmap word is half free. if so, the
3084 	 * free buddy size is BUDMIN-1.
3085 	 */
3086 	if (*((u16 *) cp) == 0 || *((u16 *) cp + 1) == 0)
3087 		return (BUDMIN - 1);
3088 
3089 	/* not all free or half free. determine the free buddy
3090 	 * size thru table lookup using quarters of the wmap word.
3091 	 */
3092 	tmp1 = max(budtab[cp[2]], budtab[cp[3]]);
3093 	tmp2 = max(budtab[cp[0]], budtab[cp[1]]);
3094 	return (max(tmp1, tmp2));
3095 }
3096 
3097 
3098 /*
3099  * NAME:	cnttz(uint word)
3100  *
3101  * FUNCTION:	determine the number of trailing zeros within a 32-bit
3102  *		value.
3103  *
3104  * PARAMETERS:
3105  *	value	-  32-bit value to be examined.
3106  *
3107  * RETURN VALUES:
3108  *	count of trailing zeros
3109  */
3110 static int cnttz(u32 word)
3111 {
3112 	int n;
3113 
3114 	for (n = 0; n < 32; n++, word >>= 1) {
3115 		if (word & 0x01)
3116 			break;
3117 	}
3118 
3119 	return (n);
3120 }
3121 
3122 
3123 /*
3124  * NAME:	cntlz(u32 value)
3125  *
3126  * FUNCTION:	determine the number of leading zeros within a 32-bit
3127  *		value.
3128  *
3129  * PARAMETERS:
3130  *	value	-  32-bit value to be examined.
3131  *
3132  * RETURN VALUES:
3133  *	count of leading zeros
3134  */
3135 static int cntlz(u32 value)
3136 {
3137 	int n;
3138 
3139 	for (n = 0; n < 32; n++, value <<= 1) {
3140 		if (value & HIGHORDER)
3141 			break;
3142 	}
3143 	return (n);
3144 }
3145 
3146 
3147 /*
3148  * NAME:	blkstol2(s64 nb)
3149  *
3150  * FUNCTION:	convert a block count to its log2 value. if the block
3151  *		count is not a l2 multiple, it is rounded up to the next
3152  *		larger l2 multiple.
3153  *
3154  * PARAMETERS:
3155  *	nb	-  number of blocks
3156  *
3157  * RETURN VALUES:
3158  *	log2 number of blocks
3159  */
3160 static int blkstol2(s64 nb)
3161 {
3162 	int l2nb;
3163 	s64 mask;		/* meant to be signed */
3164 
3165 	mask = (s64) 1 << (64 - 1);
3166 
3167 	/* count the leading bits.
3168 	 */
3169 	for (l2nb = 0; l2nb < 64; l2nb++, mask >>= 1) {
3170 		/* leading bit found.
3171 		 */
3172 		if (nb & mask) {
3173 			/* determine the l2 value.
3174 			 */
3175 			l2nb = (64 - 1) - l2nb;
3176 
3177 			/* check if we need to round up.
3178 			 */
3179 			if (~mask & nb)
3180 				l2nb++;
3181 
3182 			return (l2nb);
3183 		}
3184 	}
3185 	assert(0);
3186 	return 0;		/* fix compiler warning */
3187 }
3188 
3189 
3190 /*
3191  * NAME:	dbAllocBottomUp()
3192  *
3193  * FUNCTION:	alloc the specified block range from the working block
3194  *		allocation map.
3195  *
3196  *		the blocks will be alloc from the working map one dmap
3197  *		at a time.
3198  *
3199  * PARAMETERS:
3200  *	ip	-  pointer to in-core inode;
3201  *	blkno	-  starting block number to be freed.
3202  *	nblocks	-  number of blocks to be freed.
3203  *
3204  * RETURN VALUES:
3205  *	0	- success
3206  *	-EIO	- i/o error
3207  */
3208 int dbAllocBottomUp(struct inode *ip, s64 blkno, s64 nblocks)
3209 {
3210 	struct metapage *mp;
3211 	struct dmap *dp;
3212 	int nb, rc;
3213 	s64 lblkno, rem;
3214 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
3215 	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
3216 
3217 	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
3218 
3219 	/* block to be allocated better be within the mapsize. */
3220 	ASSERT(nblocks <= bmp->db_mapsize - blkno);
3221 
3222 	/*
3223 	 * allocate the blocks a dmap at a time.
3224 	 */
3225 	mp = NULL;
3226 	for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
3227 		/* release previous dmap if any */
3228 		if (mp) {
3229 			write_metapage(mp);
3230 		}
3231 
3232 		/* get the buffer for the current dmap. */
3233 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
3234 		mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
3235 		if (mp == NULL) {
3236 			IREAD_UNLOCK(ipbmap);
3237 			return -EIO;
3238 		}
3239 		dp = (struct dmap *) mp->data;
3240 
3241 		/* determine the number of blocks to be allocated from
3242 		 * this dmap.
3243 		 */
3244 		nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
3245 
3246 		/* allocate the blocks. */
3247 		if ((rc = dbAllocDmapBU(bmp, dp, blkno, nb))) {
3248 			release_metapage(mp);
3249 			IREAD_UNLOCK(ipbmap);
3250 			return (rc);
3251 		}
3252 	}
3253 
3254 	/* write the last buffer. */
3255 	write_metapage(mp);
3256 
3257 	IREAD_UNLOCK(ipbmap);
3258 
3259 	return (0);
3260 }
3261 
3262 
3263 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
3264 			 int nblocks)
3265 {
3266 	int rc;
3267 	int dbitno, word, rembits, nb, nwords, wbitno, agno;
3268 	s8 oldroot;
3269 	struct dmaptree *tp = (struct dmaptree *) & dp->tree;
3270 
3271 	/* save the current value of the root (i.e. maximum free string)
3272 	 * of the dmap tree.
3273 	 */
3274 	oldroot = tp->stree[ROOT];
3275 
3276 	/* determine the bit number and word within the dmap of the
3277 	 * starting block.
3278 	 */
3279 	dbitno = blkno & (BPERDMAP - 1);
3280 	word = dbitno >> L2DBWORD;
3281 
3282 	/* block range better be within the dmap */
3283 	assert(dbitno + nblocks <= BPERDMAP);
3284 
3285 	/* allocate the bits of the dmap's words corresponding to the block
3286 	 * range. not all bits of the first and last words may be contained
3287 	 * within the block range.  if this is the case, we'll work against
3288 	 * those words (i.e. partial first and/or last) on an individual basis
3289 	 * (a single pass), allocating the bits of interest by hand and
3290 	 * updating the leaf corresponding to the dmap word. a single pass
3291 	 * will be used for all dmap words fully contained within the
3292 	 * specified range.  within this pass, the bits of all fully contained
3293 	 * dmap words will be marked as free in a single shot and the leaves
3294 	 * will be updated. a single leaf may describe the free space of
3295 	 * multiple dmap words, so we may update only a subset of the actual
3296 	 * leaves corresponding to the dmap words of the block range.
3297 	 */
3298 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
3299 		/* determine the bit number within the word and
3300 		 * the number of bits within the word.
3301 		 */
3302 		wbitno = dbitno & (DBWORD - 1);
3303 		nb = min(rembits, DBWORD - wbitno);
3304 
3305 		/* check if only part of a word is to be allocated.
3306 		 */
3307 		if (nb < DBWORD) {
3308 			/* allocate (set to 1) the appropriate bits within
3309 			 * this dmap word.
3310 			 */
3311 			dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
3312 						      >> wbitno);
3313 
3314 			word++;
3315 		} else {
3316 			/* one or more dmap words are fully contained
3317 			 * within the block range.  determine how many
3318 			 * words and allocate (set to 1) the bits of these
3319 			 * words.
3320 			 */
3321 			nwords = rembits >> L2DBWORD;
3322 			memset(&dp->wmap[word], (int) ONES, nwords * 4);
3323 
3324 			/* determine how many bits */
3325 			nb = nwords << L2DBWORD;
3326 			word += nwords;
3327 		}
3328 	}
3329 
3330 	/* update the free count for this dmap */
3331 	le32_add_cpu(&dp->nfree, -nblocks);
3332 
3333 	/* reconstruct summary tree */
3334 	dbInitDmapTree(dp);
3335 
3336 	BMAP_LOCK(bmp);
3337 
3338 	/* if this allocation group is completely free,
3339 	 * update the highest active allocation group number
3340 	 * if this allocation group is the new max.
3341 	 */
3342 	agno = blkno >> bmp->db_agl2size;
3343 	if (agno > bmp->db_maxag)
3344 		bmp->db_maxag = agno;
3345 
3346 	/* update the free count for the allocation group and map */
3347 	bmp->db_agfree[agno] -= nblocks;
3348 	bmp->db_nfree -= nblocks;
3349 
3350 	BMAP_UNLOCK(bmp);
3351 
3352 	/* if the root has not changed, done. */
3353 	if (tp->stree[ROOT] == oldroot)
3354 		return (0);
3355 
3356 	/* root changed. bubble the change up to the dmap control pages.
3357 	 * if the adjustment of the upper level control pages fails,
3358 	 * backout the bit allocation (thus making everything consistent).
3359 	 */
3360 	if ((rc = dbAdjCtl(bmp, blkno, tp->stree[ROOT], 1, 0)))
3361 		dbFreeBits(bmp, dp, blkno, nblocks);
3362 
3363 	return (rc);
3364 }
3365 
3366 
3367 /*
3368  * NAME:	dbExtendFS()
3369  *
3370  * FUNCTION:	extend bmap from blkno for nblocks;
3371  *		dbExtendFS() updates bmap ready for dbAllocBottomUp();
3372  *
3373  * L2
3374  *  |
3375  *   L1---------------------------------L1
3376  *    |					 |
3377  *     L0---------L0---------L0		  L0---------L0---------L0
3378  *      |	   |	      |		   |	      |		 |
3379  *	 d0,...,dn  d0,...,dn  d0,...,dn    d0,...,dn  d0,...,dn  d0,.,dm;
3380  * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm
3381  *
3382  * <---old---><----------------------------extend----------------------->
3383  */
3384 int dbExtendFS(struct inode *ipbmap, s64 blkno,	s64 nblocks)
3385 {
3386 	struct jfs_sb_info *sbi = JFS_SBI(ipbmap->i_sb);
3387 	int nbperpage = sbi->nbperpage;
3388 	int i, i0 = true, j, j0 = true, k, n;
3389 	s64 newsize;
3390 	s64 p;
3391 	struct metapage *mp, *l2mp, *l1mp = NULL, *l0mp = NULL;
3392 	struct dmapctl *l2dcp, *l1dcp, *l0dcp;
3393 	struct dmap *dp;
3394 	s8 *l0leaf, *l1leaf, *l2leaf;
3395 	struct bmap *bmp = sbi->bmap;
3396 	int agno, l2agsize, oldl2agsize;
3397 	s64 ag_rem;
3398 
3399 	newsize = blkno + nblocks;
3400 
3401 	jfs_info("dbExtendFS: blkno:%Ld nblocks:%Ld newsize:%Ld",
3402 		 (long long) blkno, (long long) nblocks, (long long) newsize);
3403 
3404 	/*
3405 	 *	initialize bmap control page.
3406 	 *
3407 	 * all the data in bmap control page should exclude
3408 	 * the mkfs hidden dmap page.
3409 	 */
3410 
3411 	/* update mapsize */
3412 	bmp->db_mapsize = newsize;
3413 	bmp->db_maxlevel = BMAPSZTOLEV(bmp->db_mapsize);
3414 
3415 	/* compute new AG size */
3416 	l2agsize = dbGetL2AGSize(newsize);
3417 	oldl2agsize = bmp->db_agl2size;
3418 
3419 	bmp->db_agl2size = l2agsize;
3420 	bmp->db_agsize = 1 << l2agsize;
3421 
3422 	/* compute new number of AG */
3423 	agno = bmp->db_numag;
3424 	bmp->db_numag = newsize >> l2agsize;
3425 	bmp->db_numag += ((u32) newsize % (u32) bmp->db_agsize) ? 1 : 0;
3426 
3427 	/*
3428 	 *	reconfigure db_agfree[]
3429 	 * from old AG configuration to new AG configuration;
3430 	 *
3431 	 * coalesce contiguous k (newAGSize/oldAGSize) AGs;
3432 	 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn;
3433 	 * note: new AG size = old AG size * (2**x).
3434 	 */
3435 	if (l2agsize == oldl2agsize)
3436 		goto extend;
3437 	k = 1 << (l2agsize - oldl2agsize);
3438 	ag_rem = bmp->db_agfree[0];	/* save agfree[0] */
3439 	for (i = 0, n = 0; i < agno; n++) {
3440 		bmp->db_agfree[n] = 0;	/* init collection point */
3441 
3442 		/* coalesce contiguous k AGs; */
3443 		for (j = 0; j < k && i < agno; j++, i++) {
3444 			/* merge AGi to AGn */
3445 			bmp->db_agfree[n] += bmp->db_agfree[i];
3446 		}
3447 	}
3448 	bmp->db_agfree[0] += ag_rem;	/* restore agfree[0] */
3449 
3450 	for (; n < MAXAG; n++)
3451 		bmp->db_agfree[n] = 0;
3452 
3453 	/*
3454 	 * update highest active ag number
3455 	 */
3456 
3457 	bmp->db_maxag = bmp->db_maxag / k;
3458 
3459 	/*
3460 	 *	extend bmap
3461 	 *
3462 	 * update bit maps and corresponding level control pages;
3463 	 * global control page db_nfree, db_agfree[agno], db_maxfreebud;
3464 	 */
3465       extend:
3466 	/* get L2 page */
3467 	p = BMAPBLKNO + nbperpage;	/* L2 page */
3468 	l2mp = read_metapage(ipbmap, p, PSIZE, 0);
3469 	if (!l2mp) {
3470 		jfs_error(ipbmap->i_sb, "L2 page could not be read\n");
3471 		return -EIO;
3472 	}
3473 	l2dcp = (struct dmapctl *) l2mp->data;
3474 
3475 	/* compute start L1 */
3476 	k = blkno >> L2MAXL1SIZE;
3477 	l2leaf = l2dcp->stree + CTLLEAFIND + k;
3478 	p = BLKTOL1(blkno, sbi->l2nbperpage);	/* L1 page */
3479 
3480 	/*
3481 	 * extend each L1 in L2
3482 	 */
3483 	for (; k < LPERCTL; k++, p += nbperpage) {
3484 		/* get L1 page */
3485 		if (j0) {
3486 			/* read in L1 page: (blkno & (MAXL1SIZE - 1)) */
3487 			l1mp = read_metapage(ipbmap, p, PSIZE, 0);
3488 			if (l1mp == NULL)
3489 				goto errout;
3490 			l1dcp = (struct dmapctl *) l1mp->data;
3491 
3492 			/* compute start L0 */
3493 			j = (blkno & (MAXL1SIZE - 1)) >> L2MAXL0SIZE;
3494 			l1leaf = l1dcp->stree + CTLLEAFIND + j;
3495 			p = BLKTOL0(blkno, sbi->l2nbperpage);
3496 			j0 = false;
3497 		} else {
3498 			/* assign/init L1 page */
3499 			l1mp = get_metapage(ipbmap, p, PSIZE, 0);
3500 			if (l1mp == NULL)
3501 				goto errout;
3502 
3503 			l1dcp = (struct dmapctl *) l1mp->data;
3504 
3505 			/* compute start L0 */
3506 			j = 0;
3507 			l1leaf = l1dcp->stree + CTLLEAFIND;
3508 			p += nbperpage;	/* 1st L0 of L1.k */
3509 		}
3510 
3511 		/*
3512 		 * extend each L0 in L1
3513 		 */
3514 		for (; j < LPERCTL; j++) {
3515 			/* get L0 page */
3516 			if (i0) {
3517 				/* read in L0 page: (blkno & (MAXL0SIZE - 1)) */
3518 
3519 				l0mp = read_metapage(ipbmap, p, PSIZE, 0);
3520 				if (l0mp == NULL)
3521 					goto errout;
3522 				l0dcp = (struct dmapctl *) l0mp->data;
3523 
3524 				/* compute start dmap */
3525 				i = (blkno & (MAXL0SIZE - 1)) >>
3526 				    L2BPERDMAP;
3527 				l0leaf = l0dcp->stree + CTLLEAFIND + i;
3528 				p = BLKTODMAP(blkno,
3529 					      sbi->l2nbperpage);
3530 				i0 = false;
3531 			} else {
3532 				/* assign/init L0 page */
3533 				l0mp = get_metapage(ipbmap, p, PSIZE, 0);
3534 				if (l0mp == NULL)
3535 					goto errout;
3536 
3537 				l0dcp = (struct dmapctl *) l0mp->data;
3538 
3539 				/* compute start dmap */
3540 				i = 0;
3541 				l0leaf = l0dcp->stree + CTLLEAFIND;
3542 				p += nbperpage;	/* 1st dmap of L0.j */
3543 			}
3544 
3545 			/*
3546 			 * extend each dmap in L0
3547 			 */
3548 			for (; i < LPERCTL; i++) {
3549 				/*
3550 				 * reconstruct the dmap page, and
3551 				 * initialize corresponding parent L0 leaf
3552 				 */
3553 				if ((n = blkno & (BPERDMAP - 1))) {
3554 					/* read in dmap page: */
3555 					mp = read_metapage(ipbmap, p,
3556 							   PSIZE, 0);
3557 					if (mp == NULL)
3558 						goto errout;
3559 					n = min(nblocks, (s64)BPERDMAP - n);
3560 				} else {
3561 					/* assign/init dmap page */
3562 					mp = read_metapage(ipbmap, p,
3563 							   PSIZE, 0);
3564 					if (mp == NULL)
3565 						goto errout;
3566 
3567 					n = min_t(s64, nblocks, BPERDMAP);
3568 				}
3569 
3570 				dp = (struct dmap *) mp->data;
3571 				*l0leaf = dbInitDmap(dp, blkno, n);
3572 
3573 				bmp->db_nfree += n;
3574 				agno = le64_to_cpu(dp->start) >> l2agsize;
3575 				bmp->db_agfree[agno] += n;
3576 
3577 				write_metapage(mp);
3578 
3579 				l0leaf++;
3580 				p += nbperpage;
3581 
3582 				blkno += n;
3583 				nblocks -= n;
3584 				if (nblocks == 0)
3585 					break;
3586 			}	/* for each dmap in a L0 */
3587 
3588 			/*
3589 			 * build current L0 page from its leaves, and
3590 			 * initialize corresponding parent L1 leaf
3591 			 */
3592 			*l1leaf = dbInitDmapCtl(l0dcp, 0, ++i);
3593 			write_metapage(l0mp);
3594 			l0mp = NULL;
3595 
3596 			if (nblocks)
3597 				l1leaf++;	/* continue for next L0 */
3598 			else {
3599 				/* more than 1 L0 ? */
3600 				if (j > 0)
3601 					break;	/* build L1 page */
3602 				else {
3603 					/* summarize in global bmap page */
3604 					bmp->db_maxfreebud = *l1leaf;
3605 					release_metapage(l1mp);
3606 					release_metapage(l2mp);
3607 					goto finalize;
3608 				}
3609 			}
3610 		}		/* for each L0 in a L1 */
3611 
3612 		/*
3613 		 * build current L1 page from its leaves, and
3614 		 * initialize corresponding parent L2 leaf
3615 		 */
3616 		*l2leaf = dbInitDmapCtl(l1dcp, 1, ++j);
3617 		write_metapage(l1mp);
3618 		l1mp = NULL;
3619 
3620 		if (nblocks)
3621 			l2leaf++;	/* continue for next L1 */
3622 		else {
3623 			/* more than 1 L1 ? */
3624 			if (k > 0)
3625 				break;	/* build L2 page */
3626 			else {
3627 				/* summarize in global bmap page */
3628 				bmp->db_maxfreebud = *l2leaf;
3629 				release_metapage(l2mp);
3630 				goto finalize;
3631 			}
3632 		}
3633 	}			/* for each L1 in a L2 */
3634 
3635 	jfs_error(ipbmap->i_sb, "function has not returned as expected\n");
3636 errout:
3637 	if (l0mp)
3638 		release_metapage(l0mp);
3639 	if (l1mp)
3640 		release_metapage(l1mp);
3641 	release_metapage(l2mp);
3642 	return -EIO;
3643 
3644 	/*
3645 	 *	finalize bmap control page
3646 	 */
3647 finalize:
3648 
3649 	return 0;
3650 }
3651 
3652 
3653 /*
3654  *	dbFinalizeBmap()
3655  */
3656 void dbFinalizeBmap(struct inode *ipbmap)
3657 {
3658 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
3659 	int actags, inactags, l2nl;
3660 	s64 ag_rem, actfree, inactfree, avgfree;
3661 	int i, n;
3662 
3663 	/*
3664 	 *	finalize bmap control page
3665 	 */
3666 //finalize:
3667 	/*
3668 	 * compute db_agpref: preferred ag to allocate from
3669 	 * (the leftmost ag with average free space in it);
3670 	 */
3671 //agpref:
3672 	/* get the number of active ags and inacitve ags */
3673 	actags = bmp->db_maxag + 1;
3674 	inactags = bmp->db_numag - actags;
3675 	ag_rem = bmp->db_mapsize & (bmp->db_agsize - 1);	/* ??? */
3676 
3677 	/* determine how many blocks are in the inactive allocation
3678 	 * groups. in doing this, we must account for the fact that
3679 	 * the rightmost group might be a partial group (i.e. file
3680 	 * system size is not a multiple of the group size).
3681 	 */
3682 	inactfree = (inactags && ag_rem) ?
3683 	    ((inactags - 1) << bmp->db_agl2size) + ag_rem
3684 	    : inactags << bmp->db_agl2size;
3685 
3686 	/* determine how many free blocks are in the active
3687 	 * allocation groups plus the average number of free blocks
3688 	 * within the active ags.
3689 	 */
3690 	actfree = bmp->db_nfree - inactfree;
3691 	avgfree = (u32) actfree / (u32) actags;
3692 
3693 	/* if the preferred allocation group has not average free space.
3694 	 * re-establish the preferred group as the leftmost
3695 	 * group with average free space.
3696 	 */
3697 	if (bmp->db_agfree[bmp->db_agpref] < avgfree) {
3698 		for (bmp->db_agpref = 0; bmp->db_agpref < actags;
3699 		     bmp->db_agpref++) {
3700 			if (bmp->db_agfree[bmp->db_agpref] >= avgfree)
3701 				break;
3702 		}
3703 		if (bmp->db_agpref >= bmp->db_numag) {
3704 			jfs_error(ipbmap->i_sb,
3705 				  "cannot find ag with average freespace\n");
3706 		}
3707 	}
3708 
3709 	/*
3710 	 * compute db_aglevel, db_agheight, db_width, db_agstart:
3711 	 * an ag is covered in aglevel dmapctl summary tree,
3712 	 * at agheight level height (from leaf) with agwidth number of nodes
3713 	 * each, which starts at agstart index node of the smmary tree node
3714 	 * array;
3715 	 */
3716 	bmp->db_aglevel = BMAPSZTOLEV(bmp->db_agsize);
3717 	l2nl =
3718 	    bmp->db_agl2size - (L2BPERDMAP + bmp->db_aglevel * L2LPERCTL);
3719 	bmp->db_agheight = l2nl >> 1;
3720 	bmp->db_agwidth = 1 << (l2nl - (bmp->db_agheight << 1));
3721 	for (i = 5 - bmp->db_agheight, bmp->db_agstart = 0, n = 1; i > 0;
3722 	     i--) {
3723 		bmp->db_agstart += n;
3724 		n <<= 2;
3725 	}
3726 
3727 }
3728 
3729 
3730 /*
3731  * NAME:	dbInitDmap()/ujfs_idmap_page()
3732  *
3733  * FUNCTION:	initialize working/persistent bitmap of the dmap page
3734  *		for the specified number of blocks:
3735  *
3736  *		at entry, the bitmaps had been initialized as free (ZEROS);
3737  *		The number of blocks will only account for the actually
3738  *		existing blocks. Blocks which don't actually exist in
3739  *		the aggregate will be marked as allocated (ONES);
3740  *
3741  * PARAMETERS:
3742  *	dp	- pointer to page of map
3743  *	nblocks	- number of blocks this page
3744  *
3745  * RETURNS: NONE
3746  */
3747 static int dbInitDmap(struct dmap * dp, s64 Blkno, int nblocks)
3748 {
3749 	int blkno, w, b, r, nw, nb, i;
3750 
3751 	/* starting block number within the dmap */
3752 	blkno = Blkno & (BPERDMAP - 1);
3753 
3754 	if (blkno == 0) {
3755 		dp->nblocks = dp->nfree = cpu_to_le32(nblocks);
3756 		dp->start = cpu_to_le64(Blkno);
3757 
3758 		if (nblocks == BPERDMAP) {
3759 			memset(&dp->wmap[0], 0, LPERDMAP * 4);
3760 			memset(&dp->pmap[0], 0, LPERDMAP * 4);
3761 			goto initTree;
3762 		}
3763 	} else {
3764 		le32_add_cpu(&dp->nblocks, nblocks);
3765 		le32_add_cpu(&dp->nfree, nblocks);
3766 	}
3767 
3768 	/* word number containing start block number */
3769 	w = blkno >> L2DBWORD;
3770 
3771 	/*
3772 	 * free the bits corresponding to the block range (ZEROS):
3773 	 * note: not all bits of the first and last words may be contained
3774 	 * within the block range.
3775 	 */
3776 	for (r = nblocks; r > 0; r -= nb, blkno += nb) {
3777 		/* number of bits preceding range to be freed in the word */
3778 		b = blkno & (DBWORD - 1);
3779 		/* number of bits to free in the word */
3780 		nb = min(r, DBWORD - b);
3781 
3782 		/* is partial word to be freed ? */
3783 		if (nb < DBWORD) {
3784 			/* free (set to 0) from the bitmap word */
3785 			dp->wmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3786 						     >> b));
3787 			dp->pmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3788 						     >> b));
3789 
3790 			/* skip the word freed */
3791 			w++;
3792 		} else {
3793 			/* free (set to 0) contiguous bitmap words */
3794 			nw = r >> L2DBWORD;
3795 			memset(&dp->wmap[w], 0, nw * 4);
3796 			memset(&dp->pmap[w], 0, nw * 4);
3797 
3798 			/* skip the words freed */
3799 			nb = nw << L2DBWORD;
3800 			w += nw;
3801 		}
3802 	}
3803 
3804 	/*
3805 	 * mark bits following the range to be freed (non-existing
3806 	 * blocks) as allocated (ONES)
3807 	 */
3808 
3809 	if (blkno == BPERDMAP)
3810 		goto initTree;
3811 
3812 	/* the first word beyond the end of existing blocks */
3813 	w = blkno >> L2DBWORD;
3814 
3815 	/* does nblocks fall on a 32-bit boundary ? */
3816 	b = blkno & (DBWORD - 1);
3817 	if (b) {
3818 		/* mark a partial word allocated */
3819 		dp->wmap[w] = dp->pmap[w] = cpu_to_le32(ONES >> b);
3820 		w++;
3821 	}
3822 
3823 	/* set the rest of the words in the page to allocated (ONES) */
3824 	for (i = w; i < LPERDMAP; i++)
3825 		dp->pmap[i] = dp->wmap[i] = cpu_to_le32(ONES);
3826 
3827 	/*
3828 	 * init tree
3829 	 */
3830       initTree:
3831 	return (dbInitDmapTree(dp));
3832 }
3833 
3834 
3835 /*
3836  * NAME:	dbInitDmapTree()/ujfs_complete_dmap()
3837  *
3838  * FUNCTION:	initialize summary tree of the specified dmap:
3839  *
3840  *		at entry, bitmap of the dmap has been initialized;
3841  *
3842  * PARAMETERS:
3843  *	dp	- dmap to complete
3844  *	blkno	- starting block number for this dmap
3845  *	treemax	- will be filled in with max free for this dmap
3846  *
3847  * RETURNS:	max free string at the root of the tree
3848  */
3849 static int dbInitDmapTree(struct dmap * dp)
3850 {
3851 	struct dmaptree *tp;
3852 	s8 *cp;
3853 	int i;
3854 
3855 	/* init fixed info of tree */
3856 	tp = &dp->tree;
3857 	tp->nleafs = cpu_to_le32(LPERDMAP);
3858 	tp->l2nleafs = cpu_to_le32(L2LPERDMAP);
3859 	tp->leafidx = cpu_to_le32(LEAFIND);
3860 	tp->height = cpu_to_le32(4);
3861 	tp->budmin = BUDMIN;
3862 
3863 	/* init each leaf from corresponding wmap word:
3864 	 * note: leaf is set to NOFREE(-1) if all blocks of corresponding
3865 	 * bitmap word are allocated.
3866 	 */
3867 	cp = tp->stree + le32_to_cpu(tp->leafidx);
3868 	for (i = 0; i < LPERDMAP; i++)
3869 		*cp++ = dbMaxBud((u8 *) & dp->wmap[i]);
3870 
3871 	/* build the dmap's binary buddy summary tree */
3872 	return (dbInitTree(tp));
3873 }
3874 
3875 
3876 /*
3877  * NAME:	dbInitTree()/ujfs_adjtree()
3878  *
3879  * FUNCTION:	initialize binary buddy summary tree of a dmap or dmapctl.
3880  *
3881  *		at entry, the leaves of the tree has been initialized
3882  *		from corresponding bitmap word or root of summary tree
3883  *		of the child control page;
3884  *		configure binary buddy system at the leaf level, then
3885  *		bubble up the values of the leaf nodes up the tree.
3886  *
3887  * PARAMETERS:
3888  *	cp	- Pointer to the root of the tree
3889  *	l2leaves- Number of leaf nodes as a power of 2
3890  *	l2min	- Number of blocks that can be covered by a leaf
3891  *		  as a power of 2
3892  *
3893  * RETURNS: max free string at the root of the tree
3894  */
3895 static int dbInitTree(struct dmaptree * dtp)
3896 {
3897 	int l2max, l2free, bsize, nextb, i;
3898 	int child, parent, nparent;
3899 	s8 *tp, *cp, *cp1;
3900 
3901 	tp = dtp->stree;
3902 
3903 	/* Determine the maximum free string possible for the leaves */
3904 	l2max = le32_to_cpu(dtp->l2nleafs) + dtp->budmin;
3905 
3906 	/*
3907 	 * configure the leaf levevl into binary buddy system
3908 	 *
3909 	 * Try to combine buddies starting with a buddy size of 1
3910 	 * (i.e. two leaves). At a buddy size of 1 two buddy leaves
3911 	 * can be combined if both buddies have a maximum free of l2min;
3912 	 * the combination will result in the left-most buddy leaf having
3913 	 * a maximum free of l2min+1.
3914 	 * After processing all buddies for a given size, process buddies
3915 	 * at the next higher buddy size (i.e. current size * 2) and
3916 	 * the next maximum free (current free + 1).
3917 	 * This continues until the maximum possible buddy combination
3918 	 * yields maximum free.
3919 	 */
3920 	for (l2free = dtp->budmin, bsize = 1; l2free < l2max;
3921 	     l2free++, bsize = nextb) {
3922 		/* get next buddy size == current buddy pair size */
3923 		nextb = bsize << 1;
3924 
3925 		/* scan each adjacent buddy pair at current buddy size */
3926 		for (i = 0, cp = tp + le32_to_cpu(dtp->leafidx);
3927 		     i < le32_to_cpu(dtp->nleafs);
3928 		     i += nextb, cp += nextb) {
3929 			/* coalesce if both adjacent buddies are max free */
3930 			if (*cp == l2free && *(cp + bsize) == l2free) {
3931 				*cp = l2free + 1;	/* left take right */
3932 				*(cp + bsize) = -1;	/* right give left */
3933 			}
3934 		}
3935 	}
3936 
3937 	/*
3938 	 * bubble summary information of leaves up the tree.
3939 	 *
3940 	 * Starting at the leaf node level, the four nodes described by
3941 	 * the higher level parent node are compared for a maximum free and
3942 	 * this maximum becomes the value of the parent node.
3943 	 * when all lower level nodes are processed in this fashion then
3944 	 * move up to the next level (parent becomes a lower level node) and
3945 	 * continue the process for that level.
3946 	 */
3947 	for (child = le32_to_cpu(dtp->leafidx),
3948 	     nparent = le32_to_cpu(dtp->nleafs) >> 2;
3949 	     nparent > 0; nparent >>= 2, child = parent) {
3950 		/* get index of 1st node of parent level */
3951 		parent = (child - 1) >> 2;
3952 
3953 		/* set the value of the parent node as the maximum
3954 		 * of the four nodes of the current level.
3955 		 */
3956 		for (i = 0, cp = tp + child, cp1 = tp + parent;
3957 		     i < nparent; i++, cp += 4, cp1++)
3958 			*cp1 = TREEMAX(cp);
3959 	}
3960 
3961 	return (*tp);
3962 }
3963 
3964 
3965 /*
3966  *	dbInitDmapCtl()
3967  *
3968  * function: initialize dmapctl page
3969  */
3970 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i)
3971 {				/* start leaf index not covered by range */
3972 	s8 *cp;
3973 
3974 	dcp->nleafs = cpu_to_le32(LPERCTL);
3975 	dcp->l2nleafs = cpu_to_le32(L2LPERCTL);
3976 	dcp->leafidx = cpu_to_le32(CTLLEAFIND);
3977 	dcp->height = cpu_to_le32(5);
3978 	dcp->budmin = L2BPERDMAP + L2LPERCTL * level;
3979 
3980 	/*
3981 	 * initialize the leaves of current level that were not covered
3982 	 * by the specified input block range (i.e. the leaves have no
3983 	 * low level dmapctl or dmap).
3984 	 */
3985 	cp = &dcp->stree[CTLLEAFIND + i];
3986 	for (; i < LPERCTL; i++)
3987 		*cp++ = NOFREE;
3988 
3989 	/* build the dmap's binary buddy summary tree */
3990 	return (dbInitTree((struct dmaptree *) dcp));
3991 }
3992 
3993 
3994 /*
3995  * NAME:	dbGetL2AGSize()/ujfs_getagl2size()
3996  *
3997  * FUNCTION:	Determine log2(allocation group size) from aggregate size
3998  *
3999  * PARAMETERS:
4000  *	nblocks	- Number of blocks in aggregate
4001  *
4002  * RETURNS: log2(allocation group size) in aggregate blocks
4003  */
4004 static int dbGetL2AGSize(s64 nblocks)
4005 {
4006 	s64 sz;
4007 	s64 m;
4008 	int l2sz;
4009 
4010 	if (nblocks < BPERDMAP * MAXAG)
4011 		return (L2BPERDMAP);
4012 
4013 	/* round up aggregate size to power of 2 */
4014 	m = ((u64) 1 << (64 - 1));
4015 	for (l2sz = 64; l2sz >= 0; l2sz--, m >>= 1) {
4016 		if (m & nblocks)
4017 			break;
4018 	}
4019 
4020 	sz = (s64) 1 << l2sz;
4021 	if (sz < nblocks)
4022 		l2sz += 1;
4023 
4024 	/* agsize = roundupSize/max_number_of_ag */
4025 	return (l2sz - L2MAXAG);
4026 }
4027 
4028 
4029 /*
4030  * NAME:	dbMapFileSizeToMapSize()
4031  *
4032  * FUNCTION:	compute number of blocks the block allocation map file
4033  *		can cover from the map file size;
4034  *
4035  * RETURNS:	Number of blocks which can be covered by this block map file;
4036  */
4037 
4038 /*
4039  * maximum number of map pages at each level including control pages
4040  */
4041 #define MAXL0PAGES	(1 + LPERCTL)
4042 #define MAXL1PAGES	(1 + LPERCTL * MAXL0PAGES)
4043 #define MAXL2PAGES	(1 + LPERCTL * MAXL1PAGES)
4044 
4045 /*
4046  * convert number of map pages to the zero origin top dmapctl level
4047  */
4048 #define BMAPPGTOLEV(npages)	\
4049 	(((npages) <= 3 + MAXL0PAGES) ? 0 : \
4050 	 ((npages) <= 2 + MAXL1PAGES) ? 1 : 2)
4051 
4052 s64 dbMapFileSizeToMapSize(struct inode * ipbmap)
4053 {
4054 	struct super_block *sb = ipbmap->i_sb;
4055 	s64 nblocks;
4056 	s64 npages, ndmaps;
4057 	int level, i;
4058 	int complete, factor;
4059 
4060 	nblocks = ipbmap->i_size >> JFS_SBI(sb)->l2bsize;
4061 	npages = nblocks >> JFS_SBI(sb)->l2nbperpage;
4062 	level = BMAPPGTOLEV(npages);
4063 
4064 	/* At each level, accumulate the number of dmap pages covered by
4065 	 * the number of full child levels below it;
4066 	 * repeat for the last incomplete child level.
4067 	 */
4068 	ndmaps = 0;
4069 	npages--;		/* skip the first global control page */
4070 	/* skip higher level control pages above top level covered by map */
4071 	npages -= (2 - level);
4072 	npages--;		/* skip top level's control page */
4073 	for (i = level; i >= 0; i--) {
4074 		factor =
4075 		    (i == 2) ? MAXL1PAGES : ((i == 1) ? MAXL0PAGES : 1);
4076 		complete = (u32) npages / factor;
4077 		ndmaps += complete * ((i == 2) ? LPERCTL * LPERCTL :
4078 				      ((i == 1) ? LPERCTL : 1));
4079 
4080 		/* pages in last/incomplete child */
4081 		npages = (u32) npages % factor;
4082 		/* skip incomplete child's level control page */
4083 		npages--;
4084 	}
4085 
4086 	/* convert the number of dmaps into the number of blocks
4087 	 * which can be covered by the dmaps;
4088 	 */
4089 	nblocks = ndmaps << L2BPERDMAP;
4090 
4091 	return (nblocks);
4092 }
4093