xref: /openbmc/linux/fs/jfs/jfs_dmap.c (revision d2825fa9)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  *   Copyright (C) International Business Machines Corp., 2000-2004
4  *   Portions Copyright (C) Tino Reichardt, 2012
5  */
6 
7 #include <linux/fs.h>
8 #include <linux/slab.h>
9 #include "jfs_incore.h"
10 #include "jfs_superblock.h"
11 #include "jfs_dmap.h"
12 #include "jfs_imap.h"
13 #include "jfs_lock.h"
14 #include "jfs_metapage.h"
15 #include "jfs_debug.h"
16 #include "jfs_discard.h"
17 
18 /*
19  *	SERIALIZATION of the Block Allocation Map.
20  *
21  *	the working state of the block allocation map is accessed in
22  *	two directions:
23  *
24  *	1) allocation and free requests that start at the dmap
25  *	   level and move up through the dmap control pages (i.e.
26  *	   the vast majority of requests).
27  *
28  *	2) allocation requests that start at dmap control page
29  *	   level and work down towards the dmaps.
30  *
31  *	the serialization scheme used here is as follows.
32  *
33  *	requests which start at the bottom are serialized against each
34  *	other through buffers and each requests holds onto its buffers
35  *	as it works it way up from a single dmap to the required level
36  *	of dmap control page.
37  *	requests that start at the top are serialized against each other
38  *	and request that start from the bottom by the multiple read/single
39  *	write inode lock of the bmap inode. requests starting at the top
40  *	take this lock in write mode while request starting at the bottom
41  *	take the lock in read mode.  a single top-down request may proceed
42  *	exclusively while multiple bottoms-up requests may proceed
43  *	simultaneously (under the protection of busy buffers).
44  *
45  *	in addition to information found in dmaps and dmap control pages,
46  *	the working state of the block allocation map also includes read/
47  *	write information maintained in the bmap descriptor (i.e. total
48  *	free block count, allocation group level free block counts).
49  *	a single exclusive lock (BMAP_LOCK) is used to guard this information
50  *	in the face of multiple-bottoms up requests.
51  *	(lock ordering: IREAD_LOCK, BMAP_LOCK);
52  *
53  *	accesses to the persistent state of the block allocation map (limited
54  *	to the persistent bitmaps in dmaps) is guarded by (busy) buffers.
55  */
56 
57 #define BMAP_LOCK_INIT(bmp)	mutex_init(&bmp->db_bmaplock)
58 #define BMAP_LOCK(bmp)		mutex_lock(&bmp->db_bmaplock)
59 #define BMAP_UNLOCK(bmp)	mutex_unlock(&bmp->db_bmaplock)
60 
61 /*
62  * forward references
63  */
64 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
65 			int nblocks);
66 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval);
67 static int dbBackSplit(dmtree_t * tp, int leafno);
68 static int dbJoin(dmtree_t * tp, int leafno, int newval);
69 static void dbAdjTree(dmtree_t * tp, int leafno, int newval);
70 static int dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc,
71 		    int level);
72 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results);
73 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
74 		       int nblocks);
75 static int dbAllocNear(struct bmap * bmp, struct dmap * dp, s64 blkno,
76 		       int nblocks,
77 		       int l2nb, s64 * results);
78 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
79 		       int nblocks);
80 static int dbAllocDmapLev(struct bmap * bmp, struct dmap * dp, int nblocks,
81 			  int l2nb,
82 			  s64 * results);
83 static int dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb,
84 		     s64 * results);
85 static int dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno,
86 		      s64 * results);
87 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks);
88 static int dbFindBits(u32 word, int l2nb);
89 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno);
90 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx);
91 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
92 		      int nblocks);
93 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
94 		      int nblocks);
95 static int dbMaxBud(u8 * cp);
96 static int blkstol2(s64 nb);
97 
98 static int cntlz(u32 value);
99 static int cnttz(u32 word);
100 
101 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
102 			 int nblocks);
103 static int dbInitDmap(struct dmap * dp, s64 blkno, int nblocks);
104 static int dbInitDmapTree(struct dmap * dp);
105 static int dbInitTree(struct dmaptree * dtp);
106 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i);
107 static int dbGetL2AGSize(s64 nblocks);
108 
109 /*
110  *	buddy table
111  *
112  * table used for determining buddy sizes within characters of
113  * dmap bitmap words.  the characters themselves serve as indexes
114  * into the table, with the table elements yielding the maximum
115  * binary buddy of free bits within the character.
116  */
117 static const s8 budtab[256] = {
118 	3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
119 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
120 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
121 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
122 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
123 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
124 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
125 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
126 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
127 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
128 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
129 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
130 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
131 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
132 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
133 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, -1
134 };
135 
136 /*
137  * NAME:	dbMount()
138  *
139  * FUNCTION:	initializate the block allocation map.
140  *
141  *		memory is allocated for the in-core bmap descriptor and
142  *		the in-core descriptor is initialized from disk.
143  *
144  * PARAMETERS:
145  *	ipbmap	- pointer to in-core inode for the block map.
146  *
147  * RETURN VALUES:
148  *	0	- success
149  *	-ENOMEM	- insufficient memory
150  *	-EIO	- i/o error
151  *	-EINVAL - wrong bmap data
152  */
153 int dbMount(struct inode *ipbmap)
154 {
155 	struct bmap *bmp;
156 	struct dbmap_disk *dbmp_le;
157 	struct metapage *mp;
158 	int i;
159 
160 	/*
161 	 * allocate/initialize the in-memory bmap descriptor
162 	 */
163 	/* allocate memory for the in-memory bmap descriptor */
164 	bmp = kmalloc(sizeof(struct bmap), GFP_KERNEL);
165 	if (bmp == NULL)
166 		return -ENOMEM;
167 
168 	/* read the on-disk bmap descriptor. */
169 	mp = read_metapage(ipbmap,
170 			   BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
171 			   PSIZE, 0);
172 	if (mp == NULL) {
173 		kfree(bmp);
174 		return -EIO;
175 	}
176 
177 	/* copy the on-disk bmap descriptor to its in-memory version. */
178 	dbmp_le = (struct dbmap_disk *) mp->data;
179 	bmp->db_mapsize = le64_to_cpu(dbmp_le->dn_mapsize);
180 	bmp->db_nfree = le64_to_cpu(dbmp_le->dn_nfree);
181 	bmp->db_l2nbperpage = le32_to_cpu(dbmp_le->dn_l2nbperpage);
182 	bmp->db_numag = le32_to_cpu(dbmp_le->dn_numag);
183 	if (!bmp->db_numag) {
184 		release_metapage(mp);
185 		kfree(bmp);
186 		return -EINVAL;
187 	}
188 
189 	bmp->db_maxlevel = le32_to_cpu(dbmp_le->dn_maxlevel);
190 	bmp->db_maxag = le32_to_cpu(dbmp_le->dn_maxag);
191 	bmp->db_agpref = le32_to_cpu(dbmp_le->dn_agpref);
192 	bmp->db_aglevel = le32_to_cpu(dbmp_le->dn_aglevel);
193 	bmp->db_agheight = le32_to_cpu(dbmp_le->dn_agheight);
194 	bmp->db_agwidth = le32_to_cpu(dbmp_le->dn_agwidth);
195 	bmp->db_agstart = le32_to_cpu(dbmp_le->dn_agstart);
196 	bmp->db_agl2size = le32_to_cpu(dbmp_le->dn_agl2size);
197 	for (i = 0; i < MAXAG; i++)
198 		bmp->db_agfree[i] = le64_to_cpu(dbmp_le->dn_agfree[i]);
199 	bmp->db_agsize = le64_to_cpu(dbmp_le->dn_agsize);
200 	bmp->db_maxfreebud = dbmp_le->dn_maxfreebud;
201 
202 	/* release the buffer. */
203 	release_metapage(mp);
204 
205 	/* bind the bmap inode and the bmap descriptor to each other. */
206 	bmp->db_ipbmap = ipbmap;
207 	JFS_SBI(ipbmap->i_sb)->bmap = bmp;
208 
209 	memset(bmp->db_active, 0, sizeof(bmp->db_active));
210 
211 	/*
212 	 * allocate/initialize the bmap lock
213 	 */
214 	BMAP_LOCK_INIT(bmp);
215 
216 	return (0);
217 }
218 
219 
220 /*
221  * NAME:	dbUnmount()
222  *
223  * FUNCTION:	terminate the block allocation map in preparation for
224  *		file system unmount.
225  *
226  *		the in-core bmap descriptor is written to disk and
227  *		the memory for this descriptor is freed.
228  *
229  * PARAMETERS:
230  *	ipbmap	- pointer to in-core inode for the block map.
231  *
232  * RETURN VALUES:
233  *	0	- success
234  *	-EIO	- i/o error
235  */
236 int dbUnmount(struct inode *ipbmap, int mounterror)
237 {
238 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
239 
240 	if (!(mounterror || isReadOnly(ipbmap)))
241 		dbSync(ipbmap);
242 
243 	/*
244 	 * Invalidate the page cache buffers
245 	 */
246 	truncate_inode_pages(ipbmap->i_mapping, 0);
247 
248 	/* free the memory for the in-memory bmap. */
249 	kfree(bmp);
250 
251 	return (0);
252 }
253 
254 /*
255  *	dbSync()
256  */
257 int dbSync(struct inode *ipbmap)
258 {
259 	struct dbmap_disk *dbmp_le;
260 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
261 	struct metapage *mp;
262 	int i;
263 
264 	/*
265 	 * write bmap global control page
266 	 */
267 	/* get the buffer for the on-disk bmap descriptor. */
268 	mp = read_metapage(ipbmap,
269 			   BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
270 			   PSIZE, 0);
271 	if (mp == NULL) {
272 		jfs_err("dbSync: read_metapage failed!");
273 		return -EIO;
274 	}
275 	/* copy the in-memory version of the bmap to the on-disk version */
276 	dbmp_le = (struct dbmap_disk *) mp->data;
277 	dbmp_le->dn_mapsize = cpu_to_le64(bmp->db_mapsize);
278 	dbmp_le->dn_nfree = cpu_to_le64(bmp->db_nfree);
279 	dbmp_le->dn_l2nbperpage = cpu_to_le32(bmp->db_l2nbperpage);
280 	dbmp_le->dn_numag = cpu_to_le32(bmp->db_numag);
281 	dbmp_le->dn_maxlevel = cpu_to_le32(bmp->db_maxlevel);
282 	dbmp_le->dn_maxag = cpu_to_le32(bmp->db_maxag);
283 	dbmp_le->dn_agpref = cpu_to_le32(bmp->db_agpref);
284 	dbmp_le->dn_aglevel = cpu_to_le32(bmp->db_aglevel);
285 	dbmp_le->dn_agheight = cpu_to_le32(bmp->db_agheight);
286 	dbmp_le->dn_agwidth = cpu_to_le32(bmp->db_agwidth);
287 	dbmp_le->dn_agstart = cpu_to_le32(bmp->db_agstart);
288 	dbmp_le->dn_agl2size = cpu_to_le32(bmp->db_agl2size);
289 	for (i = 0; i < MAXAG; i++)
290 		dbmp_le->dn_agfree[i] = cpu_to_le64(bmp->db_agfree[i]);
291 	dbmp_le->dn_agsize = cpu_to_le64(bmp->db_agsize);
292 	dbmp_le->dn_maxfreebud = bmp->db_maxfreebud;
293 
294 	/* write the buffer */
295 	write_metapage(mp);
296 
297 	/*
298 	 * write out dirty pages of bmap
299 	 */
300 	filemap_write_and_wait(ipbmap->i_mapping);
301 
302 	diWriteSpecial(ipbmap, 0);
303 
304 	return (0);
305 }
306 
307 /*
308  * NAME:	dbFree()
309  *
310  * FUNCTION:	free the specified block range from the working block
311  *		allocation map.
312  *
313  *		the blocks will be free from the working map one dmap
314  *		at a time.
315  *
316  * PARAMETERS:
317  *	ip	- pointer to in-core inode;
318  *	blkno	- starting block number to be freed.
319  *	nblocks	- number of blocks to be freed.
320  *
321  * RETURN VALUES:
322  *	0	- success
323  *	-EIO	- i/o error
324  */
325 int dbFree(struct inode *ip, s64 blkno, s64 nblocks)
326 {
327 	struct metapage *mp;
328 	struct dmap *dp;
329 	int nb, rc;
330 	s64 lblkno, rem;
331 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
332 	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
333 	struct super_block *sb = ipbmap->i_sb;
334 
335 	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
336 
337 	/* block to be freed better be within the mapsize. */
338 	if (unlikely((blkno == 0) || (blkno + nblocks > bmp->db_mapsize))) {
339 		IREAD_UNLOCK(ipbmap);
340 		printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
341 		       (unsigned long long) blkno,
342 		       (unsigned long long) nblocks);
343 		jfs_error(ip->i_sb, "block to be freed is outside the map\n");
344 		return -EIO;
345 	}
346 
347 	/**
348 	 * TRIM the blocks, when mounted with discard option
349 	 */
350 	if (JFS_SBI(sb)->flag & JFS_DISCARD)
351 		if (JFS_SBI(sb)->minblks_trim <= nblocks)
352 			jfs_issue_discard(ipbmap, blkno, nblocks);
353 
354 	/*
355 	 * free the blocks a dmap at a time.
356 	 */
357 	mp = NULL;
358 	for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
359 		/* release previous dmap if any */
360 		if (mp) {
361 			write_metapage(mp);
362 		}
363 
364 		/* get the buffer for the current dmap. */
365 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
366 		mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
367 		if (mp == NULL) {
368 			IREAD_UNLOCK(ipbmap);
369 			return -EIO;
370 		}
371 		dp = (struct dmap *) mp->data;
372 
373 		/* determine the number of blocks to be freed from
374 		 * this dmap.
375 		 */
376 		nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
377 
378 		/* free the blocks. */
379 		if ((rc = dbFreeDmap(bmp, dp, blkno, nb))) {
380 			jfs_error(ip->i_sb, "error in block map\n");
381 			release_metapage(mp);
382 			IREAD_UNLOCK(ipbmap);
383 			return (rc);
384 		}
385 	}
386 
387 	/* write the last buffer. */
388 	write_metapage(mp);
389 
390 	IREAD_UNLOCK(ipbmap);
391 
392 	return (0);
393 }
394 
395 
396 /*
397  * NAME:	dbUpdatePMap()
398  *
399  * FUNCTION:	update the allocation state (free or allocate) of the
400  *		specified block range in the persistent block allocation map.
401  *
402  *		the blocks will be updated in the persistent map one
403  *		dmap at a time.
404  *
405  * PARAMETERS:
406  *	ipbmap	- pointer to in-core inode for the block map.
407  *	free	- 'true' if block range is to be freed from the persistent
408  *		  map; 'false' if it is to be allocated.
409  *	blkno	- starting block number of the range.
410  *	nblocks	- number of contiguous blocks in the range.
411  *	tblk	- transaction block;
412  *
413  * RETURN VALUES:
414  *	0	- success
415  *	-EIO	- i/o error
416  */
417 int
418 dbUpdatePMap(struct inode *ipbmap,
419 	     int free, s64 blkno, s64 nblocks, struct tblock * tblk)
420 {
421 	int nblks, dbitno, wbitno, rbits;
422 	int word, nbits, nwords;
423 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
424 	s64 lblkno, rem, lastlblkno;
425 	u32 mask;
426 	struct dmap *dp;
427 	struct metapage *mp;
428 	struct jfs_log *log;
429 	int lsn, difft, diffp;
430 	unsigned long flags;
431 
432 	/* the blocks better be within the mapsize. */
433 	if (blkno + nblocks > bmp->db_mapsize) {
434 		printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
435 		       (unsigned long long) blkno,
436 		       (unsigned long long) nblocks);
437 		jfs_error(ipbmap->i_sb, "blocks are outside the map\n");
438 		return -EIO;
439 	}
440 
441 	/* compute delta of transaction lsn from log syncpt */
442 	lsn = tblk->lsn;
443 	log = (struct jfs_log *) JFS_SBI(tblk->sb)->log;
444 	logdiff(difft, lsn, log);
445 
446 	/*
447 	 * update the block state a dmap at a time.
448 	 */
449 	mp = NULL;
450 	lastlblkno = 0;
451 	for (rem = nblocks; rem > 0; rem -= nblks, blkno += nblks) {
452 		/* get the buffer for the current dmap. */
453 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
454 		if (lblkno != lastlblkno) {
455 			if (mp) {
456 				write_metapage(mp);
457 			}
458 
459 			mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE,
460 					   0);
461 			if (mp == NULL)
462 				return -EIO;
463 			metapage_wait_for_io(mp);
464 		}
465 		dp = (struct dmap *) mp->data;
466 
467 		/* determine the bit number and word within the dmap of
468 		 * the starting block.  also determine how many blocks
469 		 * are to be updated within this dmap.
470 		 */
471 		dbitno = blkno & (BPERDMAP - 1);
472 		word = dbitno >> L2DBWORD;
473 		nblks = min(rem, (s64)BPERDMAP - dbitno);
474 
475 		/* update the bits of the dmap words. the first and last
476 		 * words may only have a subset of their bits updated. if
477 		 * this is the case, we'll work against that word (i.e.
478 		 * partial first and/or last) only in a single pass.  a
479 		 * single pass will also be used to update all words that
480 		 * are to have all their bits updated.
481 		 */
482 		for (rbits = nblks; rbits > 0;
483 		     rbits -= nbits, dbitno += nbits) {
484 			/* determine the bit number within the word and
485 			 * the number of bits within the word.
486 			 */
487 			wbitno = dbitno & (DBWORD - 1);
488 			nbits = min(rbits, DBWORD - wbitno);
489 
490 			/* check if only part of the word is to be updated. */
491 			if (nbits < DBWORD) {
492 				/* update (free or allocate) the bits
493 				 * in this word.
494 				 */
495 				mask =
496 				    (ONES << (DBWORD - nbits) >> wbitno);
497 				if (free)
498 					dp->pmap[word] &=
499 					    cpu_to_le32(~mask);
500 				else
501 					dp->pmap[word] |=
502 					    cpu_to_le32(mask);
503 
504 				word += 1;
505 			} else {
506 				/* one or more words are to have all
507 				 * their bits updated.  determine how
508 				 * many words and how many bits.
509 				 */
510 				nwords = rbits >> L2DBWORD;
511 				nbits = nwords << L2DBWORD;
512 
513 				/* update (free or allocate) the bits
514 				 * in these words.
515 				 */
516 				if (free)
517 					memset(&dp->pmap[word], 0,
518 					       nwords * 4);
519 				else
520 					memset(&dp->pmap[word], (int) ONES,
521 					       nwords * 4);
522 
523 				word += nwords;
524 			}
525 		}
526 
527 		/*
528 		 * update dmap lsn
529 		 */
530 		if (lblkno == lastlblkno)
531 			continue;
532 
533 		lastlblkno = lblkno;
534 
535 		LOGSYNC_LOCK(log, flags);
536 		if (mp->lsn != 0) {
537 			/* inherit older/smaller lsn */
538 			logdiff(diffp, mp->lsn, log);
539 			if (difft < diffp) {
540 				mp->lsn = lsn;
541 
542 				/* move bp after tblock in logsync list */
543 				list_move(&mp->synclist, &tblk->synclist);
544 			}
545 
546 			/* inherit younger/larger clsn */
547 			logdiff(difft, tblk->clsn, log);
548 			logdiff(diffp, mp->clsn, log);
549 			if (difft > diffp)
550 				mp->clsn = tblk->clsn;
551 		} else {
552 			mp->log = log;
553 			mp->lsn = lsn;
554 
555 			/* insert bp after tblock in logsync list */
556 			log->count++;
557 			list_add(&mp->synclist, &tblk->synclist);
558 
559 			mp->clsn = tblk->clsn;
560 		}
561 		LOGSYNC_UNLOCK(log, flags);
562 	}
563 
564 	/* write the last buffer. */
565 	if (mp) {
566 		write_metapage(mp);
567 	}
568 
569 	return (0);
570 }
571 
572 
573 /*
574  * NAME:	dbNextAG()
575  *
576  * FUNCTION:	find the preferred allocation group for new allocations.
577  *
578  *		Within the allocation groups, we maintain a preferred
579  *		allocation group which consists of a group with at least
580  *		average free space.  It is the preferred group that we target
581  *		new inode allocation towards.  The tie-in between inode
582  *		allocation and block allocation occurs as we allocate the
583  *		first (data) block of an inode and specify the inode (block)
584  *		as the allocation hint for this block.
585  *
586  *		We try to avoid having more than one open file growing in
587  *		an allocation group, as this will lead to fragmentation.
588  *		This differs from the old OS/2 method of trying to keep
589  *		empty ags around for large allocations.
590  *
591  * PARAMETERS:
592  *	ipbmap	- pointer to in-core inode for the block map.
593  *
594  * RETURN VALUES:
595  *	the preferred allocation group number.
596  */
597 int dbNextAG(struct inode *ipbmap)
598 {
599 	s64 avgfree;
600 	int agpref;
601 	s64 hwm = 0;
602 	int i;
603 	int next_best = -1;
604 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
605 
606 	BMAP_LOCK(bmp);
607 
608 	/* determine the average number of free blocks within the ags. */
609 	avgfree = (u32)bmp->db_nfree / bmp->db_numag;
610 
611 	/*
612 	 * if the current preferred ag does not have an active allocator
613 	 * and has at least average freespace, return it
614 	 */
615 	agpref = bmp->db_agpref;
616 	if ((atomic_read(&bmp->db_active[agpref]) == 0) &&
617 	    (bmp->db_agfree[agpref] >= avgfree))
618 		goto unlock;
619 
620 	/* From the last preferred ag, find the next one with at least
621 	 * average free space.
622 	 */
623 	for (i = 0 ; i < bmp->db_numag; i++, agpref++) {
624 		if (agpref == bmp->db_numag)
625 			agpref = 0;
626 
627 		if (atomic_read(&bmp->db_active[agpref]))
628 			/* open file is currently growing in this ag */
629 			continue;
630 		if (bmp->db_agfree[agpref] >= avgfree) {
631 			/* Return this one */
632 			bmp->db_agpref = agpref;
633 			goto unlock;
634 		} else if (bmp->db_agfree[agpref] > hwm) {
635 			/* Less than avg. freespace, but best so far */
636 			hwm = bmp->db_agfree[agpref];
637 			next_best = agpref;
638 		}
639 	}
640 
641 	/*
642 	 * If no inactive ag was found with average freespace, use the
643 	 * next best
644 	 */
645 	if (next_best != -1)
646 		bmp->db_agpref = next_best;
647 	/* else leave db_agpref unchanged */
648 unlock:
649 	BMAP_UNLOCK(bmp);
650 
651 	/* return the preferred group.
652 	 */
653 	return (bmp->db_agpref);
654 }
655 
656 /*
657  * NAME:	dbAlloc()
658  *
659  * FUNCTION:	attempt to allocate a specified number of contiguous free
660  *		blocks from the working allocation block map.
661  *
662  *		the block allocation policy uses hints and a multi-step
663  *		approach.
664  *
665  *		for allocation requests smaller than the number of blocks
666  *		per dmap, we first try to allocate the new blocks
667  *		immediately following the hint.  if these blocks are not
668  *		available, we try to allocate blocks near the hint.  if
669  *		no blocks near the hint are available, we next try to
670  *		allocate within the same dmap as contains the hint.
671  *
672  *		if no blocks are available in the dmap or the allocation
673  *		request is larger than the dmap size, we try to allocate
674  *		within the same allocation group as contains the hint. if
675  *		this does not succeed, we finally try to allocate anywhere
676  *		within the aggregate.
677  *
678  *		we also try to allocate anywhere within the aggregate
679  *		for allocation requests larger than the allocation group
680  *		size or requests that specify no hint value.
681  *
682  * PARAMETERS:
683  *	ip	- pointer to in-core inode;
684  *	hint	- allocation hint.
685  *	nblocks	- number of contiguous blocks in the range.
686  *	results	- on successful return, set to the starting block number
687  *		  of the newly allocated contiguous range.
688  *
689  * RETURN VALUES:
690  *	0	- success
691  *	-ENOSPC	- insufficient disk resources
692  *	-EIO	- i/o error
693  */
694 int dbAlloc(struct inode *ip, s64 hint, s64 nblocks, s64 * results)
695 {
696 	int rc, agno;
697 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
698 	struct bmap *bmp;
699 	struct metapage *mp;
700 	s64 lblkno, blkno;
701 	struct dmap *dp;
702 	int l2nb;
703 	s64 mapSize;
704 	int writers;
705 
706 	/* assert that nblocks is valid */
707 	assert(nblocks > 0);
708 
709 	/* get the log2 number of blocks to be allocated.
710 	 * if the number of blocks is not a log2 multiple,
711 	 * it will be rounded up to the next log2 multiple.
712 	 */
713 	l2nb = BLKSTOL2(nblocks);
714 
715 	bmp = JFS_SBI(ip->i_sb)->bmap;
716 
717 	mapSize = bmp->db_mapsize;
718 
719 	/* the hint should be within the map */
720 	if (hint >= mapSize) {
721 		jfs_error(ip->i_sb, "the hint is outside the map\n");
722 		return -EIO;
723 	}
724 
725 	/* if the number of blocks to be allocated is greater than the
726 	 * allocation group size, try to allocate anywhere.
727 	 */
728 	if (l2nb > bmp->db_agl2size) {
729 		IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
730 
731 		rc = dbAllocAny(bmp, nblocks, l2nb, results);
732 
733 		goto write_unlock;
734 	}
735 
736 	/*
737 	 * If no hint, let dbNextAG recommend an allocation group
738 	 */
739 	if (hint == 0)
740 		goto pref_ag;
741 
742 	/* we would like to allocate close to the hint.  adjust the
743 	 * hint to the block following the hint since the allocators
744 	 * will start looking for free space starting at this point.
745 	 */
746 	blkno = hint + 1;
747 
748 	if (blkno >= bmp->db_mapsize)
749 		goto pref_ag;
750 
751 	agno = blkno >> bmp->db_agl2size;
752 
753 	/* check if blkno crosses over into a new allocation group.
754 	 * if so, check if we should allow allocations within this
755 	 * allocation group.
756 	 */
757 	if ((blkno & (bmp->db_agsize - 1)) == 0)
758 		/* check if the AG is currently being written to.
759 		 * if so, call dbNextAG() to find a non-busy
760 		 * AG with sufficient free space.
761 		 */
762 		if (atomic_read(&bmp->db_active[agno]))
763 			goto pref_ag;
764 
765 	/* check if the allocation request size can be satisfied from a
766 	 * single dmap.  if so, try to allocate from the dmap containing
767 	 * the hint using a tiered strategy.
768 	 */
769 	if (nblocks <= BPERDMAP) {
770 		IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
771 
772 		/* get the buffer for the dmap containing the hint.
773 		 */
774 		rc = -EIO;
775 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
776 		mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
777 		if (mp == NULL)
778 			goto read_unlock;
779 
780 		dp = (struct dmap *) mp->data;
781 
782 		/* first, try to satisfy the allocation request with the
783 		 * blocks beginning at the hint.
784 		 */
785 		if ((rc = dbAllocNext(bmp, dp, blkno, (int) nblocks))
786 		    != -ENOSPC) {
787 			if (rc == 0) {
788 				*results = blkno;
789 				mark_metapage_dirty(mp);
790 			}
791 
792 			release_metapage(mp);
793 			goto read_unlock;
794 		}
795 
796 		writers = atomic_read(&bmp->db_active[agno]);
797 		if ((writers > 1) ||
798 		    ((writers == 1) && (JFS_IP(ip)->active_ag != agno))) {
799 			/*
800 			 * Someone else is writing in this allocation
801 			 * group.  To avoid fragmenting, try another ag
802 			 */
803 			release_metapage(mp);
804 			IREAD_UNLOCK(ipbmap);
805 			goto pref_ag;
806 		}
807 
808 		/* next, try to satisfy the allocation request with blocks
809 		 * near the hint.
810 		 */
811 		if ((rc =
812 		     dbAllocNear(bmp, dp, blkno, (int) nblocks, l2nb, results))
813 		    != -ENOSPC) {
814 			if (rc == 0)
815 				mark_metapage_dirty(mp);
816 
817 			release_metapage(mp);
818 			goto read_unlock;
819 		}
820 
821 		/* try to satisfy the allocation request with blocks within
822 		 * the same dmap as the hint.
823 		 */
824 		if ((rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results))
825 		    != -ENOSPC) {
826 			if (rc == 0)
827 				mark_metapage_dirty(mp);
828 
829 			release_metapage(mp);
830 			goto read_unlock;
831 		}
832 
833 		release_metapage(mp);
834 		IREAD_UNLOCK(ipbmap);
835 	}
836 
837 	/* try to satisfy the allocation request with blocks within
838 	 * the same allocation group as the hint.
839 	 */
840 	IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
841 	if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) != -ENOSPC)
842 		goto write_unlock;
843 
844 	IWRITE_UNLOCK(ipbmap);
845 
846 
847       pref_ag:
848 	/*
849 	 * Let dbNextAG recommend a preferred allocation group
850 	 */
851 	agno = dbNextAG(ipbmap);
852 	IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
853 
854 	/* Try to allocate within this allocation group.  if that fails, try to
855 	 * allocate anywhere in the map.
856 	 */
857 	if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) == -ENOSPC)
858 		rc = dbAllocAny(bmp, nblocks, l2nb, results);
859 
860       write_unlock:
861 	IWRITE_UNLOCK(ipbmap);
862 
863 	return (rc);
864 
865       read_unlock:
866 	IREAD_UNLOCK(ipbmap);
867 
868 	return (rc);
869 }
870 
871 #ifdef _NOTYET
872 /*
873  * NAME:	dbAllocExact()
874  *
875  * FUNCTION:	try to allocate the requested extent;
876  *
877  * PARAMETERS:
878  *	ip	- pointer to in-core inode;
879  *	blkno	- extent address;
880  *	nblocks	- extent length;
881  *
882  * RETURN VALUES:
883  *	0	- success
884  *	-ENOSPC	- insufficient disk resources
885  *	-EIO	- i/o error
886  */
887 int dbAllocExact(struct inode *ip, s64 blkno, int nblocks)
888 {
889 	int rc;
890 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
891 	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
892 	struct dmap *dp;
893 	s64 lblkno;
894 	struct metapage *mp;
895 
896 	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
897 
898 	/*
899 	 * validate extent request:
900 	 *
901 	 * note: defragfs policy:
902 	 *  max 64 blocks will be moved.
903 	 *  allocation request size must be satisfied from a single dmap.
904 	 */
905 	if (nblocks <= 0 || nblocks > BPERDMAP || blkno >= bmp->db_mapsize) {
906 		IREAD_UNLOCK(ipbmap);
907 		return -EINVAL;
908 	}
909 
910 	if (nblocks > ((s64) 1 << bmp->db_maxfreebud)) {
911 		/* the free space is no longer available */
912 		IREAD_UNLOCK(ipbmap);
913 		return -ENOSPC;
914 	}
915 
916 	/* read in the dmap covering the extent */
917 	lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
918 	mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
919 	if (mp == NULL) {
920 		IREAD_UNLOCK(ipbmap);
921 		return -EIO;
922 	}
923 	dp = (struct dmap *) mp->data;
924 
925 	/* try to allocate the requested extent */
926 	rc = dbAllocNext(bmp, dp, blkno, nblocks);
927 
928 	IREAD_UNLOCK(ipbmap);
929 
930 	if (rc == 0)
931 		mark_metapage_dirty(mp);
932 
933 	release_metapage(mp);
934 
935 	return (rc);
936 }
937 #endif /* _NOTYET */
938 
939 /*
940  * NAME:	dbReAlloc()
941  *
942  * FUNCTION:	attempt to extend a current allocation by a specified
943  *		number of blocks.
944  *
945  *		this routine attempts to satisfy the allocation request
946  *		by first trying to extend the existing allocation in
947  *		place by allocating the additional blocks as the blocks
948  *		immediately following the current allocation.  if these
949  *		blocks are not available, this routine will attempt to
950  *		allocate a new set of contiguous blocks large enough
951  *		to cover the existing allocation plus the additional
952  *		number of blocks required.
953  *
954  * PARAMETERS:
955  *	ip	    -  pointer to in-core inode requiring allocation.
956  *	blkno	    -  starting block of the current allocation.
957  *	nblocks	    -  number of contiguous blocks within the current
958  *		       allocation.
959  *	addnblocks  -  number of blocks to add to the allocation.
960  *	results	-      on successful return, set to the starting block number
961  *		       of the existing allocation if the existing allocation
962  *		       was extended in place or to a newly allocated contiguous
963  *		       range if the existing allocation could not be extended
964  *		       in place.
965  *
966  * RETURN VALUES:
967  *	0	- success
968  *	-ENOSPC	- insufficient disk resources
969  *	-EIO	- i/o error
970  */
971 int
972 dbReAlloc(struct inode *ip,
973 	  s64 blkno, s64 nblocks, s64 addnblocks, s64 * results)
974 {
975 	int rc;
976 
977 	/* try to extend the allocation in place.
978 	 */
979 	if ((rc = dbExtend(ip, blkno, nblocks, addnblocks)) == 0) {
980 		*results = blkno;
981 		return (0);
982 	} else {
983 		if (rc != -ENOSPC)
984 			return (rc);
985 	}
986 
987 	/* could not extend the allocation in place, so allocate a
988 	 * new set of blocks for the entire request (i.e. try to get
989 	 * a range of contiguous blocks large enough to cover the
990 	 * existing allocation plus the additional blocks.)
991 	 */
992 	return (dbAlloc
993 		(ip, blkno + nblocks - 1, addnblocks + nblocks, results));
994 }
995 
996 
997 /*
998  * NAME:	dbExtend()
999  *
1000  * FUNCTION:	attempt to extend a current allocation by a specified
1001  *		number of blocks.
1002  *
1003  *		this routine attempts to satisfy the allocation request
1004  *		by first trying to extend the existing allocation in
1005  *		place by allocating the additional blocks as the blocks
1006  *		immediately following the current allocation.
1007  *
1008  * PARAMETERS:
1009  *	ip	    -  pointer to in-core inode requiring allocation.
1010  *	blkno	    -  starting block of the current allocation.
1011  *	nblocks	    -  number of contiguous blocks within the current
1012  *		       allocation.
1013  *	addnblocks  -  number of blocks to add to the allocation.
1014  *
1015  * RETURN VALUES:
1016  *	0	- success
1017  *	-ENOSPC	- insufficient disk resources
1018  *	-EIO	- i/o error
1019  */
1020 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks)
1021 {
1022 	struct jfs_sb_info *sbi = JFS_SBI(ip->i_sb);
1023 	s64 lblkno, lastblkno, extblkno;
1024 	uint rel_block;
1025 	struct metapage *mp;
1026 	struct dmap *dp;
1027 	int rc;
1028 	struct inode *ipbmap = sbi->ipbmap;
1029 	struct bmap *bmp;
1030 
1031 	/*
1032 	 * We don't want a non-aligned extent to cross a page boundary
1033 	 */
1034 	if (((rel_block = blkno & (sbi->nbperpage - 1))) &&
1035 	    (rel_block + nblocks + addnblocks > sbi->nbperpage))
1036 		return -ENOSPC;
1037 
1038 	/* get the last block of the current allocation */
1039 	lastblkno = blkno + nblocks - 1;
1040 
1041 	/* determine the block number of the block following
1042 	 * the existing allocation.
1043 	 */
1044 	extblkno = lastblkno + 1;
1045 
1046 	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
1047 
1048 	/* better be within the file system */
1049 	bmp = sbi->bmap;
1050 	if (lastblkno < 0 || lastblkno >= bmp->db_mapsize) {
1051 		IREAD_UNLOCK(ipbmap);
1052 		jfs_error(ip->i_sb, "the block is outside the filesystem\n");
1053 		return -EIO;
1054 	}
1055 
1056 	/* we'll attempt to extend the current allocation in place by
1057 	 * allocating the additional blocks as the blocks immediately
1058 	 * following the current allocation.  we only try to extend the
1059 	 * current allocation in place if the number of additional blocks
1060 	 * can fit into a dmap, the last block of the current allocation
1061 	 * is not the last block of the file system, and the start of the
1062 	 * inplace extension is not on an allocation group boundary.
1063 	 */
1064 	if (addnblocks > BPERDMAP || extblkno >= bmp->db_mapsize ||
1065 	    (extblkno & (bmp->db_agsize - 1)) == 0) {
1066 		IREAD_UNLOCK(ipbmap);
1067 		return -ENOSPC;
1068 	}
1069 
1070 	/* get the buffer for the dmap containing the first block
1071 	 * of the extension.
1072 	 */
1073 	lblkno = BLKTODMAP(extblkno, bmp->db_l2nbperpage);
1074 	mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
1075 	if (mp == NULL) {
1076 		IREAD_UNLOCK(ipbmap);
1077 		return -EIO;
1078 	}
1079 
1080 	dp = (struct dmap *) mp->data;
1081 
1082 	/* try to allocate the blocks immediately following the
1083 	 * current allocation.
1084 	 */
1085 	rc = dbAllocNext(bmp, dp, extblkno, (int) addnblocks);
1086 
1087 	IREAD_UNLOCK(ipbmap);
1088 
1089 	/* were we successful ? */
1090 	if (rc == 0)
1091 		write_metapage(mp);
1092 	else
1093 		/* we were not successful */
1094 		release_metapage(mp);
1095 
1096 	return (rc);
1097 }
1098 
1099 
1100 /*
1101  * NAME:	dbAllocNext()
1102  *
1103  * FUNCTION:	attempt to allocate the blocks of the specified block
1104  *		range within a dmap.
1105  *
1106  * PARAMETERS:
1107  *	bmp	-  pointer to bmap descriptor
1108  *	dp	-  pointer to dmap.
1109  *	blkno	-  starting block number of the range.
1110  *	nblocks	-  number of contiguous free blocks of the range.
1111  *
1112  * RETURN VALUES:
1113  *	0	- success
1114  *	-ENOSPC	- insufficient disk resources
1115  *	-EIO	- i/o error
1116  *
1117  * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1118  */
1119 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
1120 		       int nblocks)
1121 {
1122 	int dbitno, word, rembits, nb, nwords, wbitno, nw;
1123 	int l2size;
1124 	s8 *leaf;
1125 	u32 mask;
1126 
1127 	if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1128 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1129 		return -EIO;
1130 	}
1131 
1132 	/* pick up a pointer to the leaves of the dmap tree.
1133 	 */
1134 	leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1135 
1136 	/* determine the bit number and word within the dmap of the
1137 	 * starting block.
1138 	 */
1139 	dbitno = blkno & (BPERDMAP - 1);
1140 	word = dbitno >> L2DBWORD;
1141 
1142 	/* check if the specified block range is contained within
1143 	 * this dmap.
1144 	 */
1145 	if (dbitno + nblocks > BPERDMAP)
1146 		return -ENOSPC;
1147 
1148 	/* check if the starting leaf indicates that anything
1149 	 * is free.
1150 	 */
1151 	if (leaf[word] == NOFREE)
1152 		return -ENOSPC;
1153 
1154 	/* check the dmaps words corresponding to block range to see
1155 	 * if the block range is free.  not all bits of the first and
1156 	 * last words may be contained within the block range.  if this
1157 	 * is the case, we'll work against those words (i.e. partial first
1158 	 * and/or last) on an individual basis (a single pass) and examine
1159 	 * the actual bits to determine if they are free.  a single pass
1160 	 * will be used for all dmap words fully contained within the
1161 	 * specified range.  within this pass, the leaves of the dmap
1162 	 * tree will be examined to determine if the blocks are free. a
1163 	 * single leaf may describe the free space of multiple dmap
1164 	 * words, so we may visit only a subset of the actual leaves
1165 	 * corresponding to the dmap words of the block range.
1166 	 */
1167 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
1168 		/* determine the bit number within the word and
1169 		 * the number of bits within the word.
1170 		 */
1171 		wbitno = dbitno & (DBWORD - 1);
1172 		nb = min(rembits, DBWORD - wbitno);
1173 
1174 		/* check if only part of the word is to be examined.
1175 		 */
1176 		if (nb < DBWORD) {
1177 			/* check if the bits are free.
1178 			 */
1179 			mask = (ONES << (DBWORD - nb) >> wbitno);
1180 			if ((mask & ~le32_to_cpu(dp->wmap[word])) != mask)
1181 				return -ENOSPC;
1182 
1183 			word += 1;
1184 		} else {
1185 			/* one or more dmap words are fully contained
1186 			 * within the block range.  determine how many
1187 			 * words and how many bits.
1188 			 */
1189 			nwords = rembits >> L2DBWORD;
1190 			nb = nwords << L2DBWORD;
1191 
1192 			/* now examine the appropriate leaves to determine
1193 			 * if the blocks are free.
1194 			 */
1195 			while (nwords > 0) {
1196 				/* does the leaf describe any free space ?
1197 				 */
1198 				if (leaf[word] < BUDMIN)
1199 					return -ENOSPC;
1200 
1201 				/* determine the l2 number of bits provided
1202 				 * by this leaf.
1203 				 */
1204 				l2size =
1205 				    min_t(int, leaf[word], NLSTOL2BSZ(nwords));
1206 
1207 				/* determine how many words were handled.
1208 				 */
1209 				nw = BUDSIZE(l2size, BUDMIN);
1210 
1211 				nwords -= nw;
1212 				word += nw;
1213 			}
1214 		}
1215 	}
1216 
1217 	/* allocate the blocks.
1218 	 */
1219 	return (dbAllocDmap(bmp, dp, blkno, nblocks));
1220 }
1221 
1222 
1223 /*
1224  * NAME:	dbAllocNear()
1225  *
1226  * FUNCTION:	attempt to allocate a number of contiguous free blocks near
1227  *		a specified block (hint) within a dmap.
1228  *
1229  *		starting with the dmap leaf that covers the hint, we'll
1230  *		check the next four contiguous leaves for sufficient free
1231  *		space.  if sufficient free space is found, we'll allocate
1232  *		the desired free space.
1233  *
1234  * PARAMETERS:
1235  *	bmp	-  pointer to bmap descriptor
1236  *	dp	-  pointer to dmap.
1237  *	blkno	-  block number to allocate near.
1238  *	nblocks	-  actual number of contiguous free blocks desired.
1239  *	l2nb	-  log2 number of contiguous free blocks desired.
1240  *	results	-  on successful return, set to the starting block number
1241  *		   of the newly allocated range.
1242  *
1243  * RETURN VALUES:
1244  *	0	- success
1245  *	-ENOSPC	- insufficient disk resources
1246  *	-EIO	- i/o error
1247  *
1248  * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1249  */
1250 static int
1251 dbAllocNear(struct bmap * bmp,
1252 	    struct dmap * dp, s64 blkno, int nblocks, int l2nb, s64 * results)
1253 {
1254 	int word, lword, rc;
1255 	s8 *leaf;
1256 
1257 	if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1258 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1259 		return -EIO;
1260 	}
1261 
1262 	leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1263 
1264 	/* determine the word within the dmap that holds the hint
1265 	 * (i.e. blkno).  also, determine the last word in the dmap
1266 	 * that we'll include in our examination.
1267 	 */
1268 	word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
1269 	lword = min(word + 4, LPERDMAP);
1270 
1271 	/* examine the leaves for sufficient free space.
1272 	 */
1273 	for (; word < lword; word++) {
1274 		/* does the leaf describe sufficient free space ?
1275 		 */
1276 		if (leaf[word] < l2nb)
1277 			continue;
1278 
1279 		/* determine the block number within the file system
1280 		 * of the first block described by this dmap word.
1281 		 */
1282 		blkno = le64_to_cpu(dp->start) + (word << L2DBWORD);
1283 
1284 		/* if not all bits of the dmap word are free, get the
1285 		 * starting bit number within the dmap word of the required
1286 		 * string of free bits and adjust the block number with the
1287 		 * value.
1288 		 */
1289 		if (leaf[word] < BUDMIN)
1290 			blkno +=
1291 			    dbFindBits(le32_to_cpu(dp->wmap[word]), l2nb);
1292 
1293 		/* allocate the blocks.
1294 		 */
1295 		if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1296 			*results = blkno;
1297 
1298 		return (rc);
1299 	}
1300 
1301 	return -ENOSPC;
1302 }
1303 
1304 
1305 /*
1306  * NAME:	dbAllocAG()
1307  *
1308  * FUNCTION:	attempt to allocate the specified number of contiguous
1309  *		free blocks within the specified allocation group.
1310  *
1311  *		unless the allocation group size is equal to the number
1312  *		of blocks per dmap, the dmap control pages will be used to
1313  *		find the required free space, if available.  we start the
1314  *		search at the highest dmap control page level which
1315  *		distinctly describes the allocation group's free space
1316  *		(i.e. the highest level at which the allocation group's
1317  *		free space is not mixed in with that of any other group).
1318  *		in addition, we start the search within this level at a
1319  *		height of the dmapctl dmtree at which the nodes distinctly
1320  *		describe the allocation group's free space.  at this height,
1321  *		the allocation group's free space may be represented by 1
1322  *		or two sub-trees, depending on the allocation group size.
1323  *		we search the top nodes of these subtrees left to right for
1324  *		sufficient free space.  if sufficient free space is found,
1325  *		the subtree is searched to find the leftmost leaf that
1326  *		has free space.  once we have made it to the leaf, we
1327  *		move the search to the next lower level dmap control page
1328  *		corresponding to this leaf.  we continue down the dmap control
1329  *		pages until we find the dmap that contains or starts the
1330  *		sufficient free space and we allocate at this dmap.
1331  *
1332  *		if the allocation group size is equal to the dmap size,
1333  *		we'll start at the dmap corresponding to the allocation
1334  *		group and attempt the allocation at this level.
1335  *
1336  *		the dmap control page search is also not performed if the
1337  *		allocation group is completely free and we go to the first
1338  *		dmap of the allocation group to do the allocation.  this is
1339  *		done because the allocation group may be part (not the first
1340  *		part) of a larger binary buddy system, causing the dmap
1341  *		control pages to indicate no free space (NOFREE) within
1342  *		the allocation group.
1343  *
1344  * PARAMETERS:
1345  *	bmp	-  pointer to bmap descriptor
1346  *	agno	- allocation group number.
1347  *	nblocks	-  actual number of contiguous free blocks desired.
1348  *	l2nb	-  log2 number of contiguous free blocks desired.
1349  *	results	-  on successful return, set to the starting block number
1350  *		   of the newly allocated range.
1351  *
1352  * RETURN VALUES:
1353  *	0	- success
1354  *	-ENOSPC	- insufficient disk resources
1355  *	-EIO	- i/o error
1356  *
1357  * note: IWRITE_LOCK(ipmap) held on entry/exit;
1358  */
1359 static int
1360 dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, s64 * results)
1361 {
1362 	struct metapage *mp;
1363 	struct dmapctl *dcp;
1364 	int rc, ti, i, k, m, n, agperlev;
1365 	s64 blkno, lblkno;
1366 	int budmin;
1367 
1368 	/* allocation request should not be for more than the
1369 	 * allocation group size.
1370 	 */
1371 	if (l2nb > bmp->db_agl2size) {
1372 		jfs_error(bmp->db_ipbmap->i_sb,
1373 			  "allocation request is larger than the allocation group size\n");
1374 		return -EIO;
1375 	}
1376 
1377 	/* determine the starting block number of the allocation
1378 	 * group.
1379 	 */
1380 	blkno = (s64) agno << bmp->db_agl2size;
1381 
1382 	/* check if the allocation group size is the minimum allocation
1383 	 * group size or if the allocation group is completely free. if
1384 	 * the allocation group size is the minimum size of BPERDMAP (i.e.
1385 	 * 1 dmap), there is no need to search the dmap control page (below)
1386 	 * that fully describes the allocation group since the allocation
1387 	 * group is already fully described by a dmap.  in this case, we
1388 	 * just call dbAllocCtl() to search the dmap tree and allocate the
1389 	 * required space if available.
1390 	 *
1391 	 * if the allocation group is completely free, dbAllocCtl() is
1392 	 * also called to allocate the required space.  this is done for
1393 	 * two reasons.  first, it makes no sense searching the dmap control
1394 	 * pages for free space when we know that free space exists.  second,
1395 	 * the dmap control pages may indicate that the allocation group
1396 	 * has no free space if the allocation group is part (not the first
1397 	 * part) of a larger binary buddy system.
1398 	 */
1399 	if (bmp->db_agsize == BPERDMAP
1400 	    || bmp->db_agfree[agno] == bmp->db_agsize) {
1401 		rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1402 		if ((rc == -ENOSPC) &&
1403 		    (bmp->db_agfree[agno] == bmp->db_agsize)) {
1404 			printk(KERN_ERR "blkno = %Lx, blocks = %Lx\n",
1405 			       (unsigned long long) blkno,
1406 			       (unsigned long long) nblocks);
1407 			jfs_error(bmp->db_ipbmap->i_sb,
1408 				  "dbAllocCtl failed in free AG\n");
1409 		}
1410 		return (rc);
1411 	}
1412 
1413 	/* the buffer for the dmap control page that fully describes the
1414 	 * allocation group.
1415 	 */
1416 	lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, bmp->db_aglevel);
1417 	mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1418 	if (mp == NULL)
1419 		return -EIO;
1420 	dcp = (struct dmapctl *) mp->data;
1421 	budmin = dcp->budmin;
1422 
1423 	if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1424 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
1425 		release_metapage(mp);
1426 		return -EIO;
1427 	}
1428 
1429 	/* search the subtree(s) of the dmap control page that describes
1430 	 * the allocation group, looking for sufficient free space.  to begin,
1431 	 * determine how many allocation groups are represented in a dmap
1432 	 * control page at the control page level (i.e. L0, L1, L2) that
1433 	 * fully describes an allocation group. next, determine the starting
1434 	 * tree index of this allocation group within the control page.
1435 	 */
1436 	agperlev =
1437 	    (1 << (L2LPERCTL - (bmp->db_agheight << 1))) / bmp->db_agwidth;
1438 	ti = bmp->db_agstart + bmp->db_agwidth * (agno & (agperlev - 1));
1439 
1440 	/* dmap control page trees fan-out by 4 and a single allocation
1441 	 * group may be described by 1 or 2 subtrees within the ag level
1442 	 * dmap control page, depending upon the ag size. examine the ag's
1443 	 * subtrees for sufficient free space, starting with the leftmost
1444 	 * subtree.
1445 	 */
1446 	for (i = 0; i < bmp->db_agwidth; i++, ti++) {
1447 		/* is there sufficient free space ?
1448 		 */
1449 		if (l2nb > dcp->stree[ti])
1450 			continue;
1451 
1452 		/* sufficient free space found in a subtree. now search down
1453 		 * the subtree to find the leftmost leaf that describes this
1454 		 * free space.
1455 		 */
1456 		for (k = bmp->db_agheight; k > 0; k--) {
1457 			for (n = 0, m = (ti << 2) + 1; n < 4; n++) {
1458 				if (l2nb <= dcp->stree[m + n]) {
1459 					ti = m + n;
1460 					break;
1461 				}
1462 			}
1463 			if (n == 4) {
1464 				jfs_error(bmp->db_ipbmap->i_sb,
1465 					  "failed descending stree\n");
1466 				release_metapage(mp);
1467 				return -EIO;
1468 			}
1469 		}
1470 
1471 		/* determine the block number within the file system
1472 		 * that corresponds to this leaf.
1473 		 */
1474 		if (bmp->db_aglevel == 2)
1475 			blkno = 0;
1476 		else if (bmp->db_aglevel == 1)
1477 			blkno &= ~(MAXL1SIZE - 1);
1478 		else		/* bmp->db_aglevel == 0 */
1479 			blkno &= ~(MAXL0SIZE - 1);
1480 
1481 		blkno +=
1482 		    ((s64) (ti - le32_to_cpu(dcp->leafidx))) << budmin;
1483 
1484 		/* release the buffer in preparation for going down
1485 		 * the next level of dmap control pages.
1486 		 */
1487 		release_metapage(mp);
1488 
1489 		/* check if we need to continue to search down the lower
1490 		 * level dmap control pages.  we need to if the number of
1491 		 * blocks required is less than maximum number of blocks
1492 		 * described at the next lower level.
1493 		 */
1494 		if (l2nb < budmin) {
1495 
1496 			/* search the lower level dmap control pages to get
1497 			 * the starting block number of the dmap that
1498 			 * contains or starts off the free space.
1499 			 */
1500 			if ((rc =
1501 			     dbFindCtl(bmp, l2nb, bmp->db_aglevel - 1,
1502 				       &blkno))) {
1503 				if (rc == -ENOSPC) {
1504 					jfs_error(bmp->db_ipbmap->i_sb,
1505 						  "control page inconsistent\n");
1506 					return -EIO;
1507 				}
1508 				return (rc);
1509 			}
1510 		}
1511 
1512 		/* allocate the blocks.
1513 		 */
1514 		rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1515 		if (rc == -ENOSPC) {
1516 			jfs_error(bmp->db_ipbmap->i_sb,
1517 				  "unable to allocate blocks\n");
1518 			rc = -EIO;
1519 		}
1520 		return (rc);
1521 	}
1522 
1523 	/* no space in the allocation group.  release the buffer and
1524 	 * return -ENOSPC.
1525 	 */
1526 	release_metapage(mp);
1527 
1528 	return -ENOSPC;
1529 }
1530 
1531 
1532 /*
1533  * NAME:	dbAllocAny()
1534  *
1535  * FUNCTION:	attempt to allocate the specified number of contiguous
1536  *		free blocks anywhere in the file system.
1537  *
1538  *		dbAllocAny() attempts to find the sufficient free space by
1539  *		searching down the dmap control pages, starting with the
1540  *		highest level (i.e. L0, L1, L2) control page.  if free space
1541  *		large enough to satisfy the desired free space is found, the
1542  *		desired free space is allocated.
1543  *
1544  * PARAMETERS:
1545  *	bmp	-  pointer to bmap descriptor
1546  *	nblocks	 -  actual number of contiguous free blocks desired.
1547  *	l2nb	 -  log2 number of contiguous free blocks desired.
1548  *	results	-  on successful return, set to the starting block number
1549  *		   of the newly allocated range.
1550  *
1551  * RETURN VALUES:
1552  *	0	- success
1553  *	-ENOSPC	- insufficient disk resources
1554  *	-EIO	- i/o error
1555  *
1556  * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1557  */
1558 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results)
1559 {
1560 	int rc;
1561 	s64 blkno = 0;
1562 
1563 	/* starting with the top level dmap control page, search
1564 	 * down the dmap control levels for sufficient free space.
1565 	 * if free space is found, dbFindCtl() returns the starting
1566 	 * block number of the dmap that contains or starts off the
1567 	 * range of free space.
1568 	 */
1569 	if ((rc = dbFindCtl(bmp, l2nb, bmp->db_maxlevel, &blkno)))
1570 		return (rc);
1571 
1572 	/* allocate the blocks.
1573 	 */
1574 	rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1575 	if (rc == -ENOSPC) {
1576 		jfs_error(bmp->db_ipbmap->i_sb, "unable to allocate blocks\n");
1577 		return -EIO;
1578 	}
1579 	return (rc);
1580 }
1581 
1582 
1583 /*
1584  * NAME:	dbDiscardAG()
1585  *
1586  * FUNCTION:	attempt to discard (TRIM) all free blocks of specific AG
1587  *
1588  *		algorithm:
1589  *		1) allocate blocks, as large as possible and save them
1590  *		   while holding IWRITE_LOCK on ipbmap
1591  *		2) trim all these saved block/length values
1592  *		3) mark the blocks free again
1593  *
1594  *		benefit:
1595  *		- we work only on one ag at some time, minimizing how long we
1596  *		  need to lock ipbmap
1597  *		- reading / writing the fs is possible most time, even on
1598  *		  trimming
1599  *
1600  *		downside:
1601  *		- we write two times to the dmapctl and dmap pages
1602  *		- but for me, this seems the best way, better ideas?
1603  *		/TR 2012
1604  *
1605  * PARAMETERS:
1606  *	ip	- pointer to in-core inode
1607  *	agno	- ag to trim
1608  *	minlen	- minimum value of contiguous blocks
1609  *
1610  * RETURN VALUES:
1611  *	s64	- actual number of blocks trimmed
1612  */
1613 s64 dbDiscardAG(struct inode *ip, int agno, s64 minlen)
1614 {
1615 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
1616 	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
1617 	s64 nblocks, blkno;
1618 	u64 trimmed = 0;
1619 	int rc, l2nb;
1620 	struct super_block *sb = ipbmap->i_sb;
1621 
1622 	struct range2trim {
1623 		u64 blkno;
1624 		u64 nblocks;
1625 	} *totrim, *tt;
1626 
1627 	/* max blkno / nblocks pairs to trim */
1628 	int count = 0, range_cnt;
1629 	u64 max_ranges;
1630 
1631 	/* prevent others from writing new stuff here, while trimming */
1632 	IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
1633 
1634 	nblocks = bmp->db_agfree[agno];
1635 	max_ranges = nblocks;
1636 	do_div(max_ranges, minlen);
1637 	range_cnt = min_t(u64, max_ranges + 1, 32 * 1024);
1638 	totrim = kmalloc_array(range_cnt, sizeof(struct range2trim), GFP_NOFS);
1639 	if (totrim == NULL) {
1640 		jfs_error(bmp->db_ipbmap->i_sb, "no memory for trim array\n");
1641 		IWRITE_UNLOCK(ipbmap);
1642 		return 0;
1643 	}
1644 
1645 	tt = totrim;
1646 	while (nblocks >= minlen) {
1647 		l2nb = BLKSTOL2(nblocks);
1648 
1649 		/* 0 = okay, -EIO = fatal, -ENOSPC -> try smaller block */
1650 		rc = dbAllocAG(bmp, agno, nblocks, l2nb, &blkno);
1651 		if (rc == 0) {
1652 			tt->blkno = blkno;
1653 			tt->nblocks = nblocks;
1654 			tt++; count++;
1655 
1656 			/* the whole ag is free, trim now */
1657 			if (bmp->db_agfree[agno] == 0)
1658 				break;
1659 
1660 			/* give a hint for the next while */
1661 			nblocks = bmp->db_agfree[agno];
1662 			continue;
1663 		} else if (rc == -ENOSPC) {
1664 			/* search for next smaller log2 block */
1665 			l2nb = BLKSTOL2(nblocks) - 1;
1666 			nblocks = 1LL << l2nb;
1667 		} else {
1668 			/* Trim any already allocated blocks */
1669 			jfs_error(bmp->db_ipbmap->i_sb, "-EIO\n");
1670 			break;
1671 		}
1672 
1673 		/* check, if our trim array is full */
1674 		if (unlikely(count >= range_cnt - 1))
1675 			break;
1676 	}
1677 	IWRITE_UNLOCK(ipbmap);
1678 
1679 	tt->nblocks = 0; /* mark the current end */
1680 	for (tt = totrim; tt->nblocks != 0; tt++) {
1681 		/* when mounted with online discard, dbFree() will
1682 		 * call jfs_issue_discard() itself */
1683 		if (!(JFS_SBI(sb)->flag & JFS_DISCARD))
1684 			jfs_issue_discard(ip, tt->blkno, tt->nblocks);
1685 		dbFree(ip, tt->blkno, tt->nblocks);
1686 		trimmed += tt->nblocks;
1687 	}
1688 	kfree(totrim);
1689 
1690 	return trimmed;
1691 }
1692 
1693 /*
1694  * NAME:	dbFindCtl()
1695  *
1696  * FUNCTION:	starting at a specified dmap control page level and block
1697  *		number, search down the dmap control levels for a range of
1698  *		contiguous free blocks large enough to satisfy an allocation
1699  *		request for the specified number of free blocks.
1700  *
1701  *		if sufficient contiguous free blocks are found, this routine
1702  *		returns the starting block number within a dmap page that
1703  *		contains or starts a range of contiqious free blocks that
1704  *		is sufficient in size.
1705  *
1706  * PARAMETERS:
1707  *	bmp	-  pointer to bmap descriptor
1708  *	level	-  starting dmap control page level.
1709  *	l2nb	-  log2 number of contiguous free blocks desired.
1710  *	*blkno	-  on entry, starting block number for conducting the search.
1711  *		   on successful return, the first block within a dmap page
1712  *		   that contains or starts a range of contiguous free blocks.
1713  *
1714  * RETURN VALUES:
1715  *	0	- success
1716  *	-ENOSPC	- insufficient disk resources
1717  *	-EIO	- i/o error
1718  *
1719  * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1720  */
1721 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno)
1722 {
1723 	int rc, leafidx, lev;
1724 	s64 b, lblkno;
1725 	struct dmapctl *dcp;
1726 	int budmin;
1727 	struct metapage *mp;
1728 
1729 	/* starting at the specified dmap control page level and block
1730 	 * number, search down the dmap control levels for the starting
1731 	 * block number of a dmap page that contains or starts off
1732 	 * sufficient free blocks.
1733 	 */
1734 	for (lev = level, b = *blkno; lev >= 0; lev--) {
1735 		/* get the buffer of the dmap control page for the block
1736 		 * number and level (i.e. L0, L1, L2).
1737 		 */
1738 		lblkno = BLKTOCTL(b, bmp->db_l2nbperpage, lev);
1739 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1740 		if (mp == NULL)
1741 			return -EIO;
1742 		dcp = (struct dmapctl *) mp->data;
1743 		budmin = dcp->budmin;
1744 
1745 		if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1746 			jfs_error(bmp->db_ipbmap->i_sb,
1747 				  "Corrupt dmapctl page\n");
1748 			release_metapage(mp);
1749 			return -EIO;
1750 		}
1751 
1752 		/* search the tree within the dmap control page for
1753 		 * sufficient free space.  if sufficient free space is found,
1754 		 * dbFindLeaf() returns the index of the leaf at which
1755 		 * free space was found.
1756 		 */
1757 		rc = dbFindLeaf((dmtree_t *) dcp, l2nb, &leafidx);
1758 
1759 		/* release the buffer.
1760 		 */
1761 		release_metapage(mp);
1762 
1763 		/* space found ?
1764 		 */
1765 		if (rc) {
1766 			if (lev != level) {
1767 				jfs_error(bmp->db_ipbmap->i_sb,
1768 					  "dmap inconsistent\n");
1769 				return -EIO;
1770 			}
1771 			return -ENOSPC;
1772 		}
1773 
1774 		/* adjust the block number to reflect the location within
1775 		 * the dmap control page (i.e. the leaf) at which free
1776 		 * space was found.
1777 		 */
1778 		b += (((s64) leafidx) << budmin);
1779 
1780 		/* we stop the search at this dmap control page level if
1781 		 * the number of blocks required is greater than or equal
1782 		 * to the maximum number of blocks described at the next
1783 		 * (lower) level.
1784 		 */
1785 		if (l2nb >= budmin)
1786 			break;
1787 	}
1788 
1789 	*blkno = b;
1790 	return (0);
1791 }
1792 
1793 
1794 /*
1795  * NAME:	dbAllocCtl()
1796  *
1797  * FUNCTION:	attempt to allocate a specified number of contiguous
1798  *		blocks starting within a specific dmap.
1799  *
1800  *		this routine is called by higher level routines that search
1801  *		the dmap control pages above the actual dmaps for contiguous
1802  *		free space.  the result of successful searches by these
1803  *		routines are the starting block numbers within dmaps, with
1804  *		the dmaps themselves containing the desired contiguous free
1805  *		space or starting a contiguous free space of desired size
1806  *		that is made up of the blocks of one or more dmaps. these
1807  *		calls should not fail due to insufficent resources.
1808  *
1809  *		this routine is called in some cases where it is not known
1810  *		whether it will fail due to insufficient resources.  more
1811  *		specifically, this occurs when allocating from an allocation
1812  *		group whose size is equal to the number of blocks per dmap.
1813  *		in this case, the dmap control pages are not examined prior
1814  *		to calling this routine (to save pathlength) and the call
1815  *		might fail.
1816  *
1817  *		for a request size that fits within a dmap, this routine relies
1818  *		upon the dmap's dmtree to find the requested contiguous free
1819  *		space.  for request sizes that are larger than a dmap, the
1820  *		requested free space will start at the first block of the
1821  *		first dmap (i.e. blkno).
1822  *
1823  * PARAMETERS:
1824  *	bmp	-  pointer to bmap descriptor
1825  *	nblocks	 -  actual number of contiguous free blocks to allocate.
1826  *	l2nb	 -  log2 number of contiguous free blocks to allocate.
1827  *	blkno	 -  starting block number of the dmap to start the allocation
1828  *		    from.
1829  *	results	-  on successful return, set to the starting block number
1830  *		   of the newly allocated range.
1831  *
1832  * RETURN VALUES:
1833  *	0	- success
1834  *	-ENOSPC	- insufficient disk resources
1835  *	-EIO	- i/o error
1836  *
1837  * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1838  */
1839 static int
1840 dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, s64 * results)
1841 {
1842 	int rc, nb;
1843 	s64 b, lblkno, n;
1844 	struct metapage *mp;
1845 	struct dmap *dp;
1846 
1847 	/* check if the allocation request is confined to a single dmap.
1848 	 */
1849 	if (l2nb <= L2BPERDMAP) {
1850 		/* get the buffer for the dmap.
1851 		 */
1852 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
1853 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1854 		if (mp == NULL)
1855 			return -EIO;
1856 		dp = (struct dmap *) mp->data;
1857 
1858 		/* try to allocate the blocks.
1859 		 */
1860 		rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results);
1861 		if (rc == 0)
1862 			mark_metapage_dirty(mp);
1863 
1864 		release_metapage(mp);
1865 
1866 		return (rc);
1867 	}
1868 
1869 	/* allocation request involving multiple dmaps. it must start on
1870 	 * a dmap boundary.
1871 	 */
1872 	assert((blkno & (BPERDMAP - 1)) == 0);
1873 
1874 	/* allocate the blocks dmap by dmap.
1875 	 */
1876 	for (n = nblocks, b = blkno; n > 0; n -= nb, b += nb) {
1877 		/* get the buffer for the dmap.
1878 		 */
1879 		lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1880 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1881 		if (mp == NULL) {
1882 			rc = -EIO;
1883 			goto backout;
1884 		}
1885 		dp = (struct dmap *) mp->data;
1886 
1887 		/* the dmap better be all free.
1888 		 */
1889 		if (dp->tree.stree[ROOT] != L2BPERDMAP) {
1890 			release_metapage(mp);
1891 			jfs_error(bmp->db_ipbmap->i_sb,
1892 				  "the dmap is not all free\n");
1893 			rc = -EIO;
1894 			goto backout;
1895 		}
1896 
1897 		/* determine how many blocks to allocate from this dmap.
1898 		 */
1899 		nb = min_t(s64, n, BPERDMAP);
1900 
1901 		/* allocate the blocks from the dmap.
1902 		 */
1903 		if ((rc = dbAllocDmap(bmp, dp, b, nb))) {
1904 			release_metapage(mp);
1905 			goto backout;
1906 		}
1907 
1908 		/* write the buffer.
1909 		 */
1910 		write_metapage(mp);
1911 	}
1912 
1913 	/* set the results (starting block number) and return.
1914 	 */
1915 	*results = blkno;
1916 	return (0);
1917 
1918 	/* something failed in handling an allocation request involving
1919 	 * multiple dmaps.  we'll try to clean up by backing out any
1920 	 * allocation that has already happened for this request.  if
1921 	 * we fail in backing out the allocation, we'll mark the file
1922 	 * system to indicate that blocks have been leaked.
1923 	 */
1924       backout:
1925 
1926 	/* try to backout the allocations dmap by dmap.
1927 	 */
1928 	for (n = nblocks - n, b = blkno; n > 0;
1929 	     n -= BPERDMAP, b += BPERDMAP) {
1930 		/* get the buffer for this dmap.
1931 		 */
1932 		lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1933 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1934 		if (mp == NULL) {
1935 			/* could not back out.  mark the file system
1936 			 * to indicate that we have leaked blocks.
1937 			 */
1938 			jfs_error(bmp->db_ipbmap->i_sb,
1939 				  "I/O Error: Block Leakage\n");
1940 			continue;
1941 		}
1942 		dp = (struct dmap *) mp->data;
1943 
1944 		/* free the blocks is this dmap.
1945 		 */
1946 		if (dbFreeDmap(bmp, dp, b, BPERDMAP)) {
1947 			/* could not back out.  mark the file system
1948 			 * to indicate that we have leaked blocks.
1949 			 */
1950 			release_metapage(mp);
1951 			jfs_error(bmp->db_ipbmap->i_sb, "Block Leakage\n");
1952 			continue;
1953 		}
1954 
1955 		/* write the buffer.
1956 		 */
1957 		write_metapage(mp);
1958 	}
1959 
1960 	return (rc);
1961 }
1962 
1963 
1964 /*
1965  * NAME:	dbAllocDmapLev()
1966  *
1967  * FUNCTION:	attempt to allocate a specified number of contiguous blocks
1968  *		from a specified dmap.
1969  *
1970  *		this routine checks if the contiguous blocks are available.
1971  *		if so, nblocks of blocks are allocated; otherwise, ENOSPC is
1972  *		returned.
1973  *
1974  * PARAMETERS:
1975  *	mp	-  pointer to bmap descriptor
1976  *	dp	-  pointer to dmap to attempt to allocate blocks from.
1977  *	l2nb	-  log2 number of contiguous block desired.
1978  *	nblocks	-  actual number of contiguous block desired.
1979  *	results	-  on successful return, set to the starting block number
1980  *		   of the newly allocated range.
1981  *
1982  * RETURN VALUES:
1983  *	0	- success
1984  *	-ENOSPC	- insufficient disk resources
1985  *	-EIO	- i/o error
1986  *
1987  * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or
1988  *	IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit;
1989  */
1990 static int
1991 dbAllocDmapLev(struct bmap * bmp,
1992 	       struct dmap * dp, int nblocks, int l2nb, s64 * results)
1993 {
1994 	s64 blkno;
1995 	int leafidx, rc;
1996 
1997 	/* can't be more than a dmaps worth of blocks */
1998 	assert(l2nb <= L2BPERDMAP);
1999 
2000 	/* search the tree within the dmap page for sufficient
2001 	 * free space.  if sufficient free space is found, dbFindLeaf()
2002 	 * returns the index of the leaf at which free space was found.
2003 	 */
2004 	if (dbFindLeaf((dmtree_t *) & dp->tree, l2nb, &leafidx))
2005 		return -ENOSPC;
2006 
2007 	/* determine the block number within the file system corresponding
2008 	 * to the leaf at which free space was found.
2009 	 */
2010 	blkno = le64_to_cpu(dp->start) + (leafidx << L2DBWORD);
2011 
2012 	/* if not all bits of the dmap word are free, get the starting
2013 	 * bit number within the dmap word of the required string of free
2014 	 * bits and adjust the block number with this value.
2015 	 */
2016 	if (dp->tree.stree[leafidx + LEAFIND] < BUDMIN)
2017 		blkno += dbFindBits(le32_to_cpu(dp->wmap[leafidx]), l2nb);
2018 
2019 	/* allocate the blocks */
2020 	if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
2021 		*results = blkno;
2022 
2023 	return (rc);
2024 }
2025 
2026 
2027 /*
2028  * NAME:	dbAllocDmap()
2029  *
2030  * FUNCTION:	adjust the disk allocation map to reflect the allocation
2031  *		of a specified block range within a dmap.
2032  *
2033  *		this routine allocates the specified blocks from the dmap
2034  *		through a call to dbAllocBits(). if the allocation of the
2035  *		block range causes the maximum string of free blocks within
2036  *		the dmap to change (i.e. the value of the root of the dmap's
2037  *		dmtree), this routine will cause this change to be reflected
2038  *		up through the appropriate levels of the dmap control pages
2039  *		by a call to dbAdjCtl() for the L0 dmap control page that
2040  *		covers this dmap.
2041  *
2042  * PARAMETERS:
2043  *	bmp	-  pointer to bmap descriptor
2044  *	dp	-  pointer to dmap to allocate the block range from.
2045  *	blkno	-  starting block number of the block to be allocated.
2046  *	nblocks	-  number of blocks to be allocated.
2047  *
2048  * RETURN VALUES:
2049  *	0	- success
2050  *	-EIO	- i/o error
2051  *
2052  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2053  */
2054 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2055 		       int nblocks)
2056 {
2057 	s8 oldroot;
2058 	int rc;
2059 
2060 	/* save the current value of the root (i.e. maximum free string)
2061 	 * of the dmap tree.
2062 	 */
2063 	oldroot = dp->tree.stree[ROOT];
2064 
2065 	/* allocate the specified (blocks) bits */
2066 	dbAllocBits(bmp, dp, blkno, nblocks);
2067 
2068 	/* if the root has not changed, done. */
2069 	if (dp->tree.stree[ROOT] == oldroot)
2070 		return (0);
2071 
2072 	/* root changed. bubble the change up to the dmap control pages.
2073 	 * if the adjustment of the upper level control pages fails,
2074 	 * backout the bit allocation (thus making everything consistent).
2075 	 */
2076 	if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 1, 0)))
2077 		dbFreeBits(bmp, dp, blkno, nblocks);
2078 
2079 	return (rc);
2080 }
2081 
2082 
2083 /*
2084  * NAME:	dbFreeDmap()
2085  *
2086  * FUNCTION:	adjust the disk allocation map to reflect the allocation
2087  *		of a specified block range within a dmap.
2088  *
2089  *		this routine frees the specified blocks from the dmap through
2090  *		a call to dbFreeBits(). if the deallocation of the block range
2091  *		causes the maximum string of free blocks within the dmap to
2092  *		change (i.e. the value of the root of the dmap's dmtree), this
2093  *		routine will cause this change to be reflected up through the
2094  *		appropriate levels of the dmap control pages by a call to
2095  *		dbAdjCtl() for the L0 dmap control page that covers this dmap.
2096  *
2097  * PARAMETERS:
2098  *	bmp	-  pointer to bmap descriptor
2099  *	dp	-  pointer to dmap to free the block range from.
2100  *	blkno	-  starting block number of the block to be freed.
2101  *	nblocks	-  number of blocks to be freed.
2102  *
2103  * RETURN VALUES:
2104  *	0	- success
2105  *	-EIO	- i/o error
2106  *
2107  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2108  */
2109 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2110 		      int nblocks)
2111 {
2112 	s8 oldroot;
2113 	int rc = 0, word;
2114 
2115 	/* save the current value of the root (i.e. maximum free string)
2116 	 * of the dmap tree.
2117 	 */
2118 	oldroot = dp->tree.stree[ROOT];
2119 
2120 	/* free the specified (blocks) bits */
2121 	rc = dbFreeBits(bmp, dp, blkno, nblocks);
2122 
2123 	/* if error or the root has not changed, done. */
2124 	if (rc || (dp->tree.stree[ROOT] == oldroot))
2125 		return (rc);
2126 
2127 	/* root changed. bubble the change up to the dmap control pages.
2128 	 * if the adjustment of the upper level control pages fails,
2129 	 * backout the deallocation.
2130 	 */
2131 	if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 0, 0))) {
2132 		word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
2133 
2134 		/* as part of backing out the deallocation, we will have
2135 		 * to back split the dmap tree if the deallocation caused
2136 		 * the freed blocks to become part of a larger binary buddy
2137 		 * system.
2138 		 */
2139 		if (dp->tree.stree[word] == NOFREE)
2140 			dbBackSplit((dmtree_t *) & dp->tree, word);
2141 
2142 		dbAllocBits(bmp, dp, blkno, nblocks);
2143 	}
2144 
2145 	return (rc);
2146 }
2147 
2148 
2149 /*
2150  * NAME:	dbAllocBits()
2151  *
2152  * FUNCTION:	allocate a specified block range from a dmap.
2153  *
2154  *		this routine updates the dmap to reflect the working
2155  *		state allocation of the specified block range. it directly
2156  *		updates the bits of the working map and causes the adjustment
2157  *		of the binary buddy system described by the dmap's dmtree
2158  *		leaves to reflect the bits allocated.  it also causes the
2159  *		dmap's dmtree, as a whole, to reflect the allocated range.
2160  *
2161  * PARAMETERS:
2162  *	bmp	-  pointer to bmap descriptor
2163  *	dp	-  pointer to dmap to allocate bits from.
2164  *	blkno	-  starting block number of the bits to be allocated.
2165  *	nblocks	-  number of bits to be allocated.
2166  *
2167  * RETURN VALUES: none
2168  *
2169  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2170  */
2171 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2172 			int nblocks)
2173 {
2174 	int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2175 	dmtree_t *tp = (dmtree_t *) & dp->tree;
2176 	int size;
2177 	s8 *leaf;
2178 
2179 	/* pick up a pointer to the leaves of the dmap tree */
2180 	leaf = dp->tree.stree + LEAFIND;
2181 
2182 	/* determine the bit number and word within the dmap of the
2183 	 * starting block.
2184 	 */
2185 	dbitno = blkno & (BPERDMAP - 1);
2186 	word = dbitno >> L2DBWORD;
2187 
2188 	/* block range better be within the dmap */
2189 	assert(dbitno + nblocks <= BPERDMAP);
2190 
2191 	/* allocate the bits of the dmap's words corresponding to the block
2192 	 * range. not all bits of the first and last words may be contained
2193 	 * within the block range.  if this is the case, we'll work against
2194 	 * those words (i.e. partial first and/or last) on an individual basis
2195 	 * (a single pass), allocating the bits of interest by hand and
2196 	 * updating the leaf corresponding to the dmap word. a single pass
2197 	 * will be used for all dmap words fully contained within the
2198 	 * specified range.  within this pass, the bits of all fully contained
2199 	 * dmap words will be marked as free in a single shot and the leaves
2200 	 * will be updated. a single leaf may describe the free space of
2201 	 * multiple dmap words, so we may update only a subset of the actual
2202 	 * leaves corresponding to the dmap words of the block range.
2203 	 */
2204 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2205 		/* determine the bit number within the word and
2206 		 * the number of bits within the word.
2207 		 */
2208 		wbitno = dbitno & (DBWORD - 1);
2209 		nb = min(rembits, DBWORD - wbitno);
2210 
2211 		/* check if only part of a word is to be allocated.
2212 		 */
2213 		if (nb < DBWORD) {
2214 			/* allocate (set to 1) the appropriate bits within
2215 			 * this dmap word.
2216 			 */
2217 			dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
2218 						      >> wbitno);
2219 
2220 			/* update the leaf for this dmap word. in addition
2221 			 * to setting the leaf value to the binary buddy max
2222 			 * of the updated dmap word, dbSplit() will split
2223 			 * the binary system of the leaves if need be.
2224 			 */
2225 			dbSplit(tp, word, BUDMIN,
2226 				dbMaxBud((u8 *) & dp->wmap[word]));
2227 
2228 			word += 1;
2229 		} else {
2230 			/* one or more dmap words are fully contained
2231 			 * within the block range.  determine how many
2232 			 * words and allocate (set to 1) the bits of these
2233 			 * words.
2234 			 */
2235 			nwords = rembits >> L2DBWORD;
2236 			memset(&dp->wmap[word], (int) ONES, nwords * 4);
2237 
2238 			/* determine how many bits.
2239 			 */
2240 			nb = nwords << L2DBWORD;
2241 
2242 			/* now update the appropriate leaves to reflect
2243 			 * the allocated words.
2244 			 */
2245 			for (; nwords > 0; nwords -= nw) {
2246 				if (leaf[word] < BUDMIN) {
2247 					jfs_error(bmp->db_ipbmap->i_sb,
2248 						  "leaf page corrupt\n");
2249 					break;
2250 				}
2251 
2252 				/* determine what the leaf value should be
2253 				 * updated to as the minimum of the l2 number
2254 				 * of bits being allocated and the l2 number
2255 				 * of bits currently described by this leaf.
2256 				 */
2257 				size = min_t(int, leaf[word],
2258 					     NLSTOL2BSZ(nwords));
2259 
2260 				/* update the leaf to reflect the allocation.
2261 				 * in addition to setting the leaf value to
2262 				 * NOFREE, dbSplit() will split the binary
2263 				 * system of the leaves to reflect the current
2264 				 * allocation (size).
2265 				 */
2266 				dbSplit(tp, word, size, NOFREE);
2267 
2268 				/* get the number of dmap words handled */
2269 				nw = BUDSIZE(size, BUDMIN);
2270 				word += nw;
2271 			}
2272 		}
2273 	}
2274 
2275 	/* update the free count for this dmap */
2276 	le32_add_cpu(&dp->nfree, -nblocks);
2277 
2278 	BMAP_LOCK(bmp);
2279 
2280 	/* if this allocation group is completely free,
2281 	 * update the maximum allocation group number if this allocation
2282 	 * group is the new max.
2283 	 */
2284 	agno = blkno >> bmp->db_agl2size;
2285 	if (agno > bmp->db_maxag)
2286 		bmp->db_maxag = agno;
2287 
2288 	/* update the free count for the allocation group and map */
2289 	bmp->db_agfree[agno] -= nblocks;
2290 	bmp->db_nfree -= nblocks;
2291 
2292 	BMAP_UNLOCK(bmp);
2293 }
2294 
2295 
2296 /*
2297  * NAME:	dbFreeBits()
2298  *
2299  * FUNCTION:	free a specified block range from a dmap.
2300  *
2301  *		this routine updates the dmap to reflect the working
2302  *		state allocation of the specified block range. it directly
2303  *		updates the bits of the working map and causes the adjustment
2304  *		of the binary buddy system described by the dmap's dmtree
2305  *		leaves to reflect the bits freed.  it also causes the dmap's
2306  *		dmtree, as a whole, to reflect the deallocated range.
2307  *
2308  * PARAMETERS:
2309  *	bmp	-  pointer to bmap descriptor
2310  *	dp	-  pointer to dmap to free bits from.
2311  *	blkno	-  starting block number of the bits to be freed.
2312  *	nblocks	-  number of bits to be freed.
2313  *
2314  * RETURN VALUES: 0 for success
2315  *
2316  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2317  */
2318 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2319 		       int nblocks)
2320 {
2321 	int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2322 	dmtree_t *tp = (dmtree_t *) & dp->tree;
2323 	int rc = 0;
2324 	int size;
2325 
2326 	/* determine the bit number and word within the dmap of the
2327 	 * starting block.
2328 	 */
2329 	dbitno = blkno & (BPERDMAP - 1);
2330 	word = dbitno >> L2DBWORD;
2331 
2332 	/* block range better be within the dmap.
2333 	 */
2334 	assert(dbitno + nblocks <= BPERDMAP);
2335 
2336 	/* free the bits of the dmaps words corresponding to the block range.
2337 	 * not all bits of the first and last words may be contained within
2338 	 * the block range.  if this is the case, we'll work against those
2339 	 * words (i.e. partial first and/or last) on an individual basis
2340 	 * (a single pass), freeing the bits of interest by hand and updating
2341 	 * the leaf corresponding to the dmap word. a single pass will be used
2342 	 * for all dmap words fully contained within the specified range.
2343 	 * within this pass, the bits of all fully contained dmap words will
2344 	 * be marked as free in a single shot and the leaves will be updated. a
2345 	 * single leaf may describe the free space of multiple dmap words,
2346 	 * so we may update only a subset of the actual leaves corresponding
2347 	 * to the dmap words of the block range.
2348 	 *
2349 	 * dbJoin() is used to update leaf values and will join the binary
2350 	 * buddy system of the leaves if the new leaf values indicate this
2351 	 * should be done.
2352 	 */
2353 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2354 		/* determine the bit number within the word and
2355 		 * the number of bits within the word.
2356 		 */
2357 		wbitno = dbitno & (DBWORD - 1);
2358 		nb = min(rembits, DBWORD - wbitno);
2359 
2360 		/* check if only part of a word is to be freed.
2361 		 */
2362 		if (nb < DBWORD) {
2363 			/* free (zero) the appropriate bits within this
2364 			 * dmap word.
2365 			 */
2366 			dp->wmap[word] &=
2367 			    cpu_to_le32(~(ONES << (DBWORD - nb)
2368 					  >> wbitno));
2369 
2370 			/* update the leaf for this dmap word.
2371 			 */
2372 			rc = dbJoin(tp, word,
2373 				    dbMaxBud((u8 *) & dp->wmap[word]));
2374 			if (rc)
2375 				return rc;
2376 
2377 			word += 1;
2378 		} else {
2379 			/* one or more dmap words are fully contained
2380 			 * within the block range.  determine how many
2381 			 * words and free (zero) the bits of these words.
2382 			 */
2383 			nwords = rembits >> L2DBWORD;
2384 			memset(&dp->wmap[word], 0, nwords * 4);
2385 
2386 			/* determine how many bits.
2387 			 */
2388 			nb = nwords << L2DBWORD;
2389 
2390 			/* now update the appropriate leaves to reflect
2391 			 * the freed words.
2392 			 */
2393 			for (; nwords > 0; nwords -= nw) {
2394 				/* determine what the leaf value should be
2395 				 * updated to as the minimum of the l2 number
2396 				 * of bits being freed and the l2 (max) number
2397 				 * of bits that can be described by this leaf.
2398 				 */
2399 				size =
2400 				    min(LITOL2BSZ
2401 					(word, L2LPERDMAP, BUDMIN),
2402 					NLSTOL2BSZ(nwords));
2403 
2404 				/* update the leaf.
2405 				 */
2406 				rc = dbJoin(tp, word, size);
2407 				if (rc)
2408 					return rc;
2409 
2410 				/* get the number of dmap words handled.
2411 				 */
2412 				nw = BUDSIZE(size, BUDMIN);
2413 				word += nw;
2414 			}
2415 		}
2416 	}
2417 
2418 	/* update the free count for this dmap.
2419 	 */
2420 	le32_add_cpu(&dp->nfree, nblocks);
2421 
2422 	BMAP_LOCK(bmp);
2423 
2424 	/* update the free count for the allocation group and
2425 	 * map.
2426 	 */
2427 	agno = blkno >> bmp->db_agl2size;
2428 	bmp->db_nfree += nblocks;
2429 	bmp->db_agfree[agno] += nblocks;
2430 
2431 	/* check if this allocation group is not completely free and
2432 	 * if it is currently the maximum (rightmost) allocation group.
2433 	 * if so, establish the new maximum allocation group number by
2434 	 * searching left for the first allocation group with allocation.
2435 	 */
2436 	if ((bmp->db_agfree[agno] == bmp->db_agsize && agno == bmp->db_maxag) ||
2437 	    (agno == bmp->db_numag - 1 &&
2438 	     bmp->db_agfree[agno] == (bmp-> db_mapsize & (BPERDMAP - 1)))) {
2439 		while (bmp->db_maxag > 0) {
2440 			bmp->db_maxag -= 1;
2441 			if (bmp->db_agfree[bmp->db_maxag] !=
2442 			    bmp->db_agsize)
2443 				break;
2444 		}
2445 
2446 		/* re-establish the allocation group preference if the
2447 		 * current preference is right of the maximum allocation
2448 		 * group.
2449 		 */
2450 		if (bmp->db_agpref > bmp->db_maxag)
2451 			bmp->db_agpref = bmp->db_maxag;
2452 	}
2453 
2454 	BMAP_UNLOCK(bmp);
2455 
2456 	return 0;
2457 }
2458 
2459 
2460 /*
2461  * NAME:	dbAdjCtl()
2462  *
2463  * FUNCTION:	adjust a dmap control page at a specified level to reflect
2464  *		the change in a lower level dmap or dmap control page's
2465  *		maximum string of free blocks (i.e. a change in the root
2466  *		of the lower level object's dmtree) due to the allocation
2467  *		or deallocation of a range of blocks with a single dmap.
2468  *
2469  *		on entry, this routine is provided with the new value of
2470  *		the lower level dmap or dmap control page root and the
2471  *		starting block number of the block range whose allocation
2472  *		or deallocation resulted in the root change.  this range
2473  *		is respresented by a single leaf of the current dmapctl
2474  *		and the leaf will be updated with this value, possibly
2475  *		causing a binary buddy system within the leaves to be
2476  *		split or joined.  the update may also cause the dmapctl's
2477  *		dmtree to be updated.
2478  *
2479  *		if the adjustment of the dmap control page, itself, causes its
2480  *		root to change, this change will be bubbled up to the next dmap
2481  *		control level by a recursive call to this routine, specifying
2482  *		the new root value and the next dmap control page level to
2483  *		be adjusted.
2484  * PARAMETERS:
2485  *	bmp	-  pointer to bmap descriptor
2486  *	blkno	-  the first block of a block range within a dmap.  it is
2487  *		   the allocation or deallocation of this block range that
2488  *		   requires the dmap control page to be adjusted.
2489  *	newval	-  the new value of the lower level dmap or dmap control
2490  *		   page root.
2491  *	alloc	-  'true' if adjustment is due to an allocation.
2492  *	level	-  current level of dmap control page (i.e. L0, L1, L2) to
2493  *		   be adjusted.
2494  *
2495  * RETURN VALUES:
2496  *	0	- success
2497  *	-EIO	- i/o error
2498  *
2499  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2500  */
2501 static int
2502 dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, int level)
2503 {
2504 	struct metapage *mp;
2505 	s8 oldroot;
2506 	int oldval;
2507 	s64 lblkno;
2508 	struct dmapctl *dcp;
2509 	int rc, leafno, ti;
2510 
2511 	/* get the buffer for the dmap control page for the specified
2512 	 * block number and control page level.
2513 	 */
2514 	lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, level);
2515 	mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
2516 	if (mp == NULL)
2517 		return -EIO;
2518 	dcp = (struct dmapctl *) mp->data;
2519 
2520 	if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
2521 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
2522 		release_metapage(mp);
2523 		return -EIO;
2524 	}
2525 
2526 	/* determine the leaf number corresponding to the block and
2527 	 * the index within the dmap control tree.
2528 	 */
2529 	leafno = BLKTOCTLLEAF(blkno, dcp->budmin);
2530 	ti = leafno + le32_to_cpu(dcp->leafidx);
2531 
2532 	/* save the current leaf value and the current root level (i.e.
2533 	 * maximum l2 free string described by this dmapctl).
2534 	 */
2535 	oldval = dcp->stree[ti];
2536 	oldroot = dcp->stree[ROOT];
2537 
2538 	/* check if this is a control page update for an allocation.
2539 	 * if so, update the leaf to reflect the new leaf value using
2540 	 * dbSplit(); otherwise (deallocation), use dbJoin() to update
2541 	 * the leaf with the new value.  in addition to updating the
2542 	 * leaf, dbSplit() will also split the binary buddy system of
2543 	 * the leaves, if required, and bubble new values within the
2544 	 * dmapctl tree, if required.  similarly, dbJoin() will join
2545 	 * the binary buddy system of leaves and bubble new values up
2546 	 * the dmapctl tree as required by the new leaf value.
2547 	 */
2548 	if (alloc) {
2549 		/* check if we are in the middle of a binary buddy
2550 		 * system.  this happens when we are performing the
2551 		 * first allocation out of an allocation group that
2552 		 * is part (not the first part) of a larger binary
2553 		 * buddy system.  if we are in the middle, back split
2554 		 * the system prior to calling dbSplit() which assumes
2555 		 * that it is at the front of a binary buddy system.
2556 		 */
2557 		if (oldval == NOFREE) {
2558 			rc = dbBackSplit((dmtree_t *) dcp, leafno);
2559 			if (rc) {
2560 				release_metapage(mp);
2561 				return rc;
2562 			}
2563 			oldval = dcp->stree[ti];
2564 		}
2565 		dbSplit((dmtree_t *) dcp, leafno, dcp->budmin, newval);
2566 	} else {
2567 		rc = dbJoin((dmtree_t *) dcp, leafno, newval);
2568 		if (rc) {
2569 			release_metapage(mp);
2570 			return rc;
2571 		}
2572 	}
2573 
2574 	/* check if the root of the current dmap control page changed due
2575 	 * to the update and if the current dmap control page is not at
2576 	 * the current top level (i.e. L0, L1, L2) of the map.  if so (i.e.
2577 	 * root changed and this is not the top level), call this routine
2578 	 * again (recursion) for the next higher level of the mapping to
2579 	 * reflect the change in root for the current dmap control page.
2580 	 */
2581 	if (dcp->stree[ROOT] != oldroot) {
2582 		/* are we below the top level of the map.  if so,
2583 		 * bubble the root up to the next higher level.
2584 		 */
2585 		if (level < bmp->db_maxlevel) {
2586 			/* bubble up the new root of this dmap control page to
2587 			 * the next level.
2588 			 */
2589 			if ((rc =
2590 			     dbAdjCtl(bmp, blkno, dcp->stree[ROOT], alloc,
2591 				      level + 1))) {
2592 				/* something went wrong in bubbling up the new
2593 				 * root value, so backout the changes to the
2594 				 * current dmap control page.
2595 				 */
2596 				if (alloc) {
2597 					dbJoin((dmtree_t *) dcp, leafno,
2598 					       oldval);
2599 				} else {
2600 					/* the dbJoin() above might have
2601 					 * caused a larger binary buddy system
2602 					 * to form and we may now be in the
2603 					 * middle of it.  if this is the case,
2604 					 * back split the buddies.
2605 					 */
2606 					if (dcp->stree[ti] == NOFREE)
2607 						dbBackSplit((dmtree_t *)
2608 							    dcp, leafno);
2609 					dbSplit((dmtree_t *) dcp, leafno,
2610 						dcp->budmin, oldval);
2611 				}
2612 
2613 				/* release the buffer and return the error.
2614 				 */
2615 				release_metapage(mp);
2616 				return (rc);
2617 			}
2618 		} else {
2619 			/* we're at the top level of the map. update
2620 			 * the bmap control page to reflect the size
2621 			 * of the maximum free buddy system.
2622 			 */
2623 			assert(level == bmp->db_maxlevel);
2624 			if (bmp->db_maxfreebud != oldroot) {
2625 				jfs_error(bmp->db_ipbmap->i_sb,
2626 					  "the maximum free buddy is not the old root\n");
2627 			}
2628 			bmp->db_maxfreebud = dcp->stree[ROOT];
2629 		}
2630 	}
2631 
2632 	/* write the buffer.
2633 	 */
2634 	write_metapage(mp);
2635 
2636 	return (0);
2637 }
2638 
2639 
2640 /*
2641  * NAME:	dbSplit()
2642  *
2643  * FUNCTION:	update the leaf of a dmtree with a new value, splitting
2644  *		the leaf from the binary buddy system of the dmtree's
2645  *		leaves, as required.
2646  *
2647  * PARAMETERS:
2648  *	tp	- pointer to the tree containing the leaf.
2649  *	leafno	- the number of the leaf to be updated.
2650  *	splitsz	- the size the binary buddy system starting at the leaf
2651  *		  must be split to, specified as the log2 number of blocks.
2652  *	newval	- the new value for the leaf.
2653  *
2654  * RETURN VALUES: none
2655  *
2656  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2657  */
2658 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval)
2659 {
2660 	int budsz;
2661 	int cursz;
2662 	s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2663 
2664 	/* check if the leaf needs to be split.
2665 	 */
2666 	if (leaf[leafno] > tp->dmt_budmin) {
2667 		/* the split occurs by cutting the buddy system in half
2668 		 * at the specified leaf until we reach the specified
2669 		 * size.  pick up the starting split size (current size
2670 		 * - 1 in l2) and the corresponding buddy size.
2671 		 */
2672 		cursz = leaf[leafno] - 1;
2673 		budsz = BUDSIZE(cursz, tp->dmt_budmin);
2674 
2675 		/* split until we reach the specified size.
2676 		 */
2677 		while (cursz >= splitsz) {
2678 			/* update the buddy's leaf with its new value.
2679 			 */
2680 			dbAdjTree(tp, leafno ^ budsz, cursz);
2681 
2682 			/* on to the next size and buddy.
2683 			 */
2684 			cursz -= 1;
2685 			budsz >>= 1;
2686 		}
2687 	}
2688 
2689 	/* adjust the dmap tree to reflect the specified leaf's new
2690 	 * value.
2691 	 */
2692 	dbAdjTree(tp, leafno, newval);
2693 }
2694 
2695 
2696 /*
2697  * NAME:	dbBackSplit()
2698  *
2699  * FUNCTION:	back split the binary buddy system of dmtree leaves
2700  *		that hold a specified leaf until the specified leaf
2701  *		starts its own binary buddy system.
2702  *
2703  *		the allocators typically perform allocations at the start
2704  *		of binary buddy systems and dbSplit() is used to accomplish
2705  *		any required splits.  in some cases, however, allocation
2706  *		may occur in the middle of a binary system and requires a
2707  *		back split, with the split proceeding out from the middle of
2708  *		the system (less efficient) rather than the start of the
2709  *		system (more efficient).  the cases in which a back split
2710  *		is required are rare and are limited to the first allocation
2711  *		within an allocation group which is a part (not first part)
2712  *		of a larger binary buddy system and a few exception cases
2713  *		in which a previous join operation must be backed out.
2714  *
2715  * PARAMETERS:
2716  *	tp	- pointer to the tree containing the leaf.
2717  *	leafno	- the number of the leaf to be updated.
2718  *
2719  * RETURN VALUES: none
2720  *
2721  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2722  */
2723 static int dbBackSplit(dmtree_t * tp, int leafno)
2724 {
2725 	int budsz, bud, w, bsz, size;
2726 	int cursz;
2727 	s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2728 
2729 	/* leaf should be part (not first part) of a binary
2730 	 * buddy system.
2731 	 */
2732 	assert(leaf[leafno] == NOFREE);
2733 
2734 	/* the back split is accomplished by iteratively finding the leaf
2735 	 * that starts the buddy system that contains the specified leaf and
2736 	 * splitting that system in two.  this iteration continues until
2737 	 * the specified leaf becomes the start of a buddy system.
2738 	 *
2739 	 * determine maximum possible l2 size for the specified leaf.
2740 	 */
2741 	size =
2742 	    LITOL2BSZ(leafno, le32_to_cpu(tp->dmt_l2nleafs),
2743 		      tp->dmt_budmin);
2744 
2745 	/* determine the number of leaves covered by this size.  this
2746 	 * is the buddy size that we will start with as we search for
2747 	 * the buddy system that contains the specified leaf.
2748 	 */
2749 	budsz = BUDSIZE(size, tp->dmt_budmin);
2750 
2751 	/* back split.
2752 	 */
2753 	while (leaf[leafno] == NOFREE) {
2754 		/* find the leftmost buddy leaf.
2755 		 */
2756 		for (w = leafno, bsz = budsz;; bsz <<= 1,
2757 		     w = (w < bud) ? w : bud) {
2758 			if (bsz >= le32_to_cpu(tp->dmt_nleafs)) {
2759 				jfs_err("JFS: block map error in dbBackSplit");
2760 				return -EIO;
2761 			}
2762 
2763 			/* determine the buddy.
2764 			 */
2765 			bud = w ^ bsz;
2766 
2767 			/* check if this buddy is the start of the system.
2768 			 */
2769 			if (leaf[bud] != NOFREE) {
2770 				/* split the leaf at the start of the
2771 				 * system in two.
2772 				 */
2773 				cursz = leaf[bud] - 1;
2774 				dbSplit(tp, bud, cursz, cursz);
2775 				break;
2776 			}
2777 		}
2778 	}
2779 
2780 	if (leaf[leafno] != size) {
2781 		jfs_err("JFS: wrong leaf value in dbBackSplit");
2782 		return -EIO;
2783 	}
2784 	return 0;
2785 }
2786 
2787 
2788 /*
2789  * NAME:	dbJoin()
2790  *
2791  * FUNCTION:	update the leaf of a dmtree with a new value, joining
2792  *		the leaf with other leaves of the dmtree into a multi-leaf
2793  *		binary buddy system, as required.
2794  *
2795  * PARAMETERS:
2796  *	tp	- pointer to the tree containing the leaf.
2797  *	leafno	- the number of the leaf to be updated.
2798  *	newval	- the new value for the leaf.
2799  *
2800  * RETURN VALUES: none
2801  */
2802 static int dbJoin(dmtree_t * tp, int leafno, int newval)
2803 {
2804 	int budsz, buddy;
2805 	s8 *leaf;
2806 
2807 	/* can the new leaf value require a join with other leaves ?
2808 	 */
2809 	if (newval >= tp->dmt_budmin) {
2810 		/* pickup a pointer to the leaves of the tree.
2811 		 */
2812 		leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2813 
2814 		/* try to join the specified leaf into a large binary
2815 		 * buddy system.  the join proceeds by attempting to join
2816 		 * the specified leafno with its buddy (leaf) at new value.
2817 		 * if the join occurs, we attempt to join the left leaf
2818 		 * of the joined buddies with its buddy at new value + 1.
2819 		 * we continue to join until we find a buddy that cannot be
2820 		 * joined (does not have a value equal to the size of the
2821 		 * last join) or until all leaves have been joined into a
2822 		 * single system.
2823 		 *
2824 		 * get the buddy size (number of words covered) of
2825 		 * the new value.
2826 		 */
2827 		budsz = BUDSIZE(newval, tp->dmt_budmin);
2828 
2829 		/* try to join.
2830 		 */
2831 		while (budsz < le32_to_cpu(tp->dmt_nleafs)) {
2832 			/* get the buddy leaf.
2833 			 */
2834 			buddy = leafno ^ budsz;
2835 
2836 			/* if the leaf's new value is greater than its
2837 			 * buddy's value, we join no more.
2838 			 */
2839 			if (newval > leaf[buddy])
2840 				break;
2841 
2842 			/* It shouldn't be less */
2843 			if (newval < leaf[buddy])
2844 				return -EIO;
2845 
2846 			/* check which (leafno or buddy) is the left buddy.
2847 			 * the left buddy gets to claim the blocks resulting
2848 			 * from the join while the right gets to claim none.
2849 			 * the left buddy is also eligible to participate in
2850 			 * a join at the next higher level while the right
2851 			 * is not.
2852 			 *
2853 			 */
2854 			if (leafno < buddy) {
2855 				/* leafno is the left buddy.
2856 				 */
2857 				dbAdjTree(tp, buddy, NOFREE);
2858 			} else {
2859 				/* buddy is the left buddy and becomes
2860 				 * leafno.
2861 				 */
2862 				dbAdjTree(tp, leafno, NOFREE);
2863 				leafno = buddy;
2864 			}
2865 
2866 			/* on to try the next join.
2867 			 */
2868 			newval += 1;
2869 			budsz <<= 1;
2870 		}
2871 	}
2872 
2873 	/* update the leaf value.
2874 	 */
2875 	dbAdjTree(tp, leafno, newval);
2876 
2877 	return 0;
2878 }
2879 
2880 
2881 /*
2882  * NAME:	dbAdjTree()
2883  *
2884  * FUNCTION:	update a leaf of a dmtree with a new value, adjusting
2885  *		the dmtree, as required, to reflect the new leaf value.
2886  *		the combination of any buddies must already be done before
2887  *		this is called.
2888  *
2889  * PARAMETERS:
2890  *	tp	- pointer to the tree to be adjusted.
2891  *	leafno	- the number of the leaf to be updated.
2892  *	newval	- the new value for the leaf.
2893  *
2894  * RETURN VALUES: none
2895  */
2896 static void dbAdjTree(dmtree_t * tp, int leafno, int newval)
2897 {
2898 	int lp, pp, k;
2899 	int max;
2900 
2901 	/* pick up the index of the leaf for this leafno.
2902 	 */
2903 	lp = leafno + le32_to_cpu(tp->dmt_leafidx);
2904 
2905 	/* is the current value the same as the old value ?  if so,
2906 	 * there is nothing to do.
2907 	 */
2908 	if (tp->dmt_stree[lp] == newval)
2909 		return;
2910 
2911 	/* set the new value.
2912 	 */
2913 	tp->dmt_stree[lp] = newval;
2914 
2915 	/* bubble the new value up the tree as required.
2916 	 */
2917 	for (k = 0; k < le32_to_cpu(tp->dmt_height); k++) {
2918 		/* get the index of the first leaf of the 4 leaf
2919 		 * group containing the specified leaf (leafno).
2920 		 */
2921 		lp = ((lp - 1) & ~0x03) + 1;
2922 
2923 		/* get the index of the parent of this 4 leaf group.
2924 		 */
2925 		pp = (lp - 1) >> 2;
2926 
2927 		/* determine the maximum of the 4 leaves.
2928 		 */
2929 		max = TREEMAX(&tp->dmt_stree[lp]);
2930 
2931 		/* if the maximum of the 4 is the same as the
2932 		 * parent's value, we're done.
2933 		 */
2934 		if (tp->dmt_stree[pp] == max)
2935 			break;
2936 
2937 		/* parent gets new value.
2938 		 */
2939 		tp->dmt_stree[pp] = max;
2940 
2941 		/* parent becomes leaf for next go-round.
2942 		 */
2943 		lp = pp;
2944 	}
2945 }
2946 
2947 
2948 /*
2949  * NAME:	dbFindLeaf()
2950  *
2951  * FUNCTION:	search a dmtree_t for sufficient free blocks, returning
2952  *		the index of a leaf describing the free blocks if
2953  *		sufficient free blocks are found.
2954  *
2955  *		the search starts at the top of the dmtree_t tree and
2956  *		proceeds down the tree to the leftmost leaf with sufficient
2957  *		free space.
2958  *
2959  * PARAMETERS:
2960  *	tp	- pointer to the tree to be searched.
2961  *	l2nb	- log2 number of free blocks to search for.
2962  *	leafidx	- return pointer to be set to the index of the leaf
2963  *		  describing at least l2nb free blocks if sufficient
2964  *		  free blocks are found.
2965  *
2966  * RETURN VALUES:
2967  *	0	- success
2968  *	-ENOSPC	- insufficient free blocks.
2969  */
2970 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx)
2971 {
2972 	int ti, n = 0, k, x = 0;
2973 
2974 	/* first check the root of the tree to see if there is
2975 	 * sufficient free space.
2976 	 */
2977 	if (l2nb > tp->dmt_stree[ROOT])
2978 		return -ENOSPC;
2979 
2980 	/* sufficient free space available. now search down the tree
2981 	 * starting at the next level for the leftmost leaf that
2982 	 * describes sufficient free space.
2983 	 */
2984 	for (k = le32_to_cpu(tp->dmt_height), ti = 1;
2985 	     k > 0; k--, ti = ((ti + n) << 2) + 1) {
2986 		/* search the four nodes at this level, starting from
2987 		 * the left.
2988 		 */
2989 		for (x = ti, n = 0; n < 4; n++) {
2990 			/* sufficient free space found.  move to the next
2991 			 * level (or quit if this is the last level).
2992 			 */
2993 			if (l2nb <= tp->dmt_stree[x + n])
2994 				break;
2995 		}
2996 
2997 		/* better have found something since the higher
2998 		 * levels of the tree said it was here.
2999 		 */
3000 		assert(n < 4);
3001 	}
3002 
3003 	/* set the return to the leftmost leaf describing sufficient
3004 	 * free space.
3005 	 */
3006 	*leafidx = x + n - le32_to_cpu(tp->dmt_leafidx);
3007 
3008 	return (0);
3009 }
3010 
3011 
3012 /*
3013  * NAME:	dbFindBits()
3014  *
3015  * FUNCTION:	find a specified number of binary buddy free bits within a
3016  *		dmap bitmap word value.
3017  *
3018  *		this routine searches the bitmap value for (1 << l2nb) free
3019  *		bits at (1 << l2nb) alignments within the value.
3020  *
3021  * PARAMETERS:
3022  *	word	-  dmap bitmap word value.
3023  *	l2nb	-  number of free bits specified as a log2 number.
3024  *
3025  * RETURN VALUES:
3026  *	starting bit number of free bits.
3027  */
3028 static int dbFindBits(u32 word, int l2nb)
3029 {
3030 	int bitno, nb;
3031 	u32 mask;
3032 
3033 	/* get the number of bits.
3034 	 */
3035 	nb = 1 << l2nb;
3036 	assert(nb <= DBWORD);
3037 
3038 	/* complement the word so we can use a mask (i.e. 0s represent
3039 	 * free bits) and compute the mask.
3040 	 */
3041 	word = ~word;
3042 	mask = ONES << (DBWORD - nb);
3043 
3044 	/* scan the word for nb free bits at nb alignments.
3045 	 */
3046 	for (bitno = 0; mask != 0; bitno += nb, mask >>= nb) {
3047 		if ((mask & word) == mask)
3048 			break;
3049 	}
3050 
3051 	ASSERT(bitno < 32);
3052 
3053 	/* return the bit number.
3054 	 */
3055 	return (bitno);
3056 }
3057 
3058 
3059 /*
3060  * NAME:	dbMaxBud(u8 *cp)
3061  *
3062  * FUNCTION:	determine the largest binary buddy string of free
3063  *		bits within 32-bits of the map.
3064  *
3065  * PARAMETERS:
3066  *	cp	-  pointer to the 32-bit value.
3067  *
3068  * RETURN VALUES:
3069  *	largest binary buddy of free bits within a dmap word.
3070  */
3071 static int dbMaxBud(u8 * cp)
3072 {
3073 	signed char tmp1, tmp2;
3074 
3075 	/* check if the wmap word is all free. if so, the
3076 	 * free buddy size is BUDMIN.
3077 	 */
3078 	if (*((uint *) cp) == 0)
3079 		return (BUDMIN);
3080 
3081 	/* check if the wmap word is half free. if so, the
3082 	 * free buddy size is BUDMIN-1.
3083 	 */
3084 	if (*((u16 *) cp) == 0 || *((u16 *) cp + 1) == 0)
3085 		return (BUDMIN - 1);
3086 
3087 	/* not all free or half free. determine the free buddy
3088 	 * size thru table lookup using quarters of the wmap word.
3089 	 */
3090 	tmp1 = max(budtab[cp[2]], budtab[cp[3]]);
3091 	tmp2 = max(budtab[cp[0]], budtab[cp[1]]);
3092 	return (max(tmp1, tmp2));
3093 }
3094 
3095 
3096 /*
3097  * NAME:	cnttz(uint word)
3098  *
3099  * FUNCTION:	determine the number of trailing zeros within a 32-bit
3100  *		value.
3101  *
3102  * PARAMETERS:
3103  *	value	-  32-bit value to be examined.
3104  *
3105  * RETURN VALUES:
3106  *	count of trailing zeros
3107  */
3108 static int cnttz(u32 word)
3109 {
3110 	int n;
3111 
3112 	for (n = 0; n < 32; n++, word >>= 1) {
3113 		if (word & 0x01)
3114 			break;
3115 	}
3116 
3117 	return (n);
3118 }
3119 
3120 
3121 /*
3122  * NAME:	cntlz(u32 value)
3123  *
3124  * FUNCTION:	determine the number of leading zeros within a 32-bit
3125  *		value.
3126  *
3127  * PARAMETERS:
3128  *	value	-  32-bit value to be examined.
3129  *
3130  * RETURN VALUES:
3131  *	count of leading zeros
3132  */
3133 static int cntlz(u32 value)
3134 {
3135 	int n;
3136 
3137 	for (n = 0; n < 32; n++, value <<= 1) {
3138 		if (value & HIGHORDER)
3139 			break;
3140 	}
3141 	return (n);
3142 }
3143 
3144 
3145 /*
3146  * NAME:	blkstol2(s64 nb)
3147  *
3148  * FUNCTION:	convert a block count to its log2 value. if the block
3149  *		count is not a l2 multiple, it is rounded up to the next
3150  *		larger l2 multiple.
3151  *
3152  * PARAMETERS:
3153  *	nb	-  number of blocks
3154  *
3155  * RETURN VALUES:
3156  *	log2 number of blocks
3157  */
3158 static int blkstol2(s64 nb)
3159 {
3160 	int l2nb;
3161 	s64 mask;		/* meant to be signed */
3162 
3163 	mask = (s64) 1 << (64 - 1);
3164 
3165 	/* count the leading bits.
3166 	 */
3167 	for (l2nb = 0; l2nb < 64; l2nb++, mask >>= 1) {
3168 		/* leading bit found.
3169 		 */
3170 		if (nb & mask) {
3171 			/* determine the l2 value.
3172 			 */
3173 			l2nb = (64 - 1) - l2nb;
3174 
3175 			/* check if we need to round up.
3176 			 */
3177 			if (~mask & nb)
3178 				l2nb++;
3179 
3180 			return (l2nb);
3181 		}
3182 	}
3183 	assert(0);
3184 	return 0;		/* fix compiler warning */
3185 }
3186 
3187 
3188 /*
3189  * NAME:	dbAllocBottomUp()
3190  *
3191  * FUNCTION:	alloc the specified block range from the working block
3192  *		allocation map.
3193  *
3194  *		the blocks will be alloc from the working map one dmap
3195  *		at a time.
3196  *
3197  * PARAMETERS:
3198  *	ip	-  pointer to in-core inode;
3199  *	blkno	-  starting block number to be freed.
3200  *	nblocks	-  number of blocks to be freed.
3201  *
3202  * RETURN VALUES:
3203  *	0	- success
3204  *	-EIO	- i/o error
3205  */
3206 int dbAllocBottomUp(struct inode *ip, s64 blkno, s64 nblocks)
3207 {
3208 	struct metapage *mp;
3209 	struct dmap *dp;
3210 	int nb, rc;
3211 	s64 lblkno, rem;
3212 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
3213 	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
3214 
3215 	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
3216 
3217 	/* block to be allocated better be within the mapsize. */
3218 	ASSERT(nblocks <= bmp->db_mapsize - blkno);
3219 
3220 	/*
3221 	 * allocate the blocks a dmap at a time.
3222 	 */
3223 	mp = NULL;
3224 	for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
3225 		/* release previous dmap if any */
3226 		if (mp) {
3227 			write_metapage(mp);
3228 		}
3229 
3230 		/* get the buffer for the current dmap. */
3231 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
3232 		mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
3233 		if (mp == NULL) {
3234 			IREAD_UNLOCK(ipbmap);
3235 			return -EIO;
3236 		}
3237 		dp = (struct dmap *) mp->data;
3238 
3239 		/* determine the number of blocks to be allocated from
3240 		 * this dmap.
3241 		 */
3242 		nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
3243 
3244 		/* allocate the blocks. */
3245 		if ((rc = dbAllocDmapBU(bmp, dp, blkno, nb))) {
3246 			release_metapage(mp);
3247 			IREAD_UNLOCK(ipbmap);
3248 			return (rc);
3249 		}
3250 	}
3251 
3252 	/* write the last buffer. */
3253 	write_metapage(mp);
3254 
3255 	IREAD_UNLOCK(ipbmap);
3256 
3257 	return (0);
3258 }
3259 
3260 
3261 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
3262 			 int nblocks)
3263 {
3264 	int rc;
3265 	int dbitno, word, rembits, nb, nwords, wbitno, agno;
3266 	s8 oldroot;
3267 	struct dmaptree *tp = (struct dmaptree *) & dp->tree;
3268 
3269 	/* save the current value of the root (i.e. maximum free string)
3270 	 * of the dmap tree.
3271 	 */
3272 	oldroot = tp->stree[ROOT];
3273 
3274 	/* determine the bit number and word within the dmap of the
3275 	 * starting block.
3276 	 */
3277 	dbitno = blkno & (BPERDMAP - 1);
3278 	word = dbitno >> L2DBWORD;
3279 
3280 	/* block range better be within the dmap */
3281 	assert(dbitno + nblocks <= BPERDMAP);
3282 
3283 	/* allocate the bits of the dmap's words corresponding to the block
3284 	 * range. not all bits of the first and last words may be contained
3285 	 * within the block range.  if this is the case, we'll work against
3286 	 * those words (i.e. partial first and/or last) on an individual basis
3287 	 * (a single pass), allocating the bits of interest by hand and
3288 	 * updating the leaf corresponding to the dmap word. a single pass
3289 	 * will be used for all dmap words fully contained within the
3290 	 * specified range.  within this pass, the bits of all fully contained
3291 	 * dmap words will be marked as free in a single shot and the leaves
3292 	 * will be updated. a single leaf may describe the free space of
3293 	 * multiple dmap words, so we may update only a subset of the actual
3294 	 * leaves corresponding to the dmap words of the block range.
3295 	 */
3296 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
3297 		/* determine the bit number within the word and
3298 		 * the number of bits within the word.
3299 		 */
3300 		wbitno = dbitno & (DBWORD - 1);
3301 		nb = min(rembits, DBWORD - wbitno);
3302 
3303 		/* check if only part of a word is to be allocated.
3304 		 */
3305 		if (nb < DBWORD) {
3306 			/* allocate (set to 1) the appropriate bits within
3307 			 * this dmap word.
3308 			 */
3309 			dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
3310 						      >> wbitno);
3311 
3312 			word++;
3313 		} else {
3314 			/* one or more dmap words are fully contained
3315 			 * within the block range.  determine how many
3316 			 * words and allocate (set to 1) the bits of these
3317 			 * words.
3318 			 */
3319 			nwords = rembits >> L2DBWORD;
3320 			memset(&dp->wmap[word], (int) ONES, nwords * 4);
3321 
3322 			/* determine how many bits */
3323 			nb = nwords << L2DBWORD;
3324 			word += nwords;
3325 		}
3326 	}
3327 
3328 	/* update the free count for this dmap */
3329 	le32_add_cpu(&dp->nfree, -nblocks);
3330 
3331 	/* reconstruct summary tree */
3332 	dbInitDmapTree(dp);
3333 
3334 	BMAP_LOCK(bmp);
3335 
3336 	/* if this allocation group is completely free,
3337 	 * update the highest active allocation group number
3338 	 * if this allocation group is the new max.
3339 	 */
3340 	agno = blkno >> bmp->db_agl2size;
3341 	if (agno > bmp->db_maxag)
3342 		bmp->db_maxag = agno;
3343 
3344 	/* update the free count for the allocation group and map */
3345 	bmp->db_agfree[agno] -= nblocks;
3346 	bmp->db_nfree -= nblocks;
3347 
3348 	BMAP_UNLOCK(bmp);
3349 
3350 	/* if the root has not changed, done. */
3351 	if (tp->stree[ROOT] == oldroot)
3352 		return (0);
3353 
3354 	/* root changed. bubble the change up to the dmap control pages.
3355 	 * if the adjustment of the upper level control pages fails,
3356 	 * backout the bit allocation (thus making everything consistent).
3357 	 */
3358 	if ((rc = dbAdjCtl(bmp, blkno, tp->stree[ROOT], 1, 0)))
3359 		dbFreeBits(bmp, dp, blkno, nblocks);
3360 
3361 	return (rc);
3362 }
3363 
3364 
3365 /*
3366  * NAME:	dbExtendFS()
3367  *
3368  * FUNCTION:	extend bmap from blkno for nblocks;
3369  *		dbExtendFS() updates bmap ready for dbAllocBottomUp();
3370  *
3371  * L2
3372  *  |
3373  *   L1---------------------------------L1
3374  *    |					 |
3375  *     L0---------L0---------L0		  L0---------L0---------L0
3376  *      |	   |	      |		   |	      |		 |
3377  *	 d0,...,dn  d0,...,dn  d0,...,dn    d0,...,dn  d0,...,dn  d0,.,dm;
3378  * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm
3379  *
3380  * <---old---><----------------------------extend----------------------->
3381  */
3382 int dbExtendFS(struct inode *ipbmap, s64 blkno,	s64 nblocks)
3383 {
3384 	struct jfs_sb_info *sbi = JFS_SBI(ipbmap->i_sb);
3385 	int nbperpage = sbi->nbperpage;
3386 	int i, i0 = true, j, j0 = true, k, n;
3387 	s64 newsize;
3388 	s64 p;
3389 	struct metapage *mp, *l2mp, *l1mp = NULL, *l0mp = NULL;
3390 	struct dmapctl *l2dcp, *l1dcp, *l0dcp;
3391 	struct dmap *dp;
3392 	s8 *l0leaf, *l1leaf, *l2leaf;
3393 	struct bmap *bmp = sbi->bmap;
3394 	int agno, l2agsize, oldl2agsize;
3395 	s64 ag_rem;
3396 
3397 	newsize = blkno + nblocks;
3398 
3399 	jfs_info("dbExtendFS: blkno:%Ld nblocks:%Ld newsize:%Ld",
3400 		 (long long) blkno, (long long) nblocks, (long long) newsize);
3401 
3402 	/*
3403 	 *	initialize bmap control page.
3404 	 *
3405 	 * all the data in bmap control page should exclude
3406 	 * the mkfs hidden dmap page.
3407 	 */
3408 
3409 	/* update mapsize */
3410 	bmp->db_mapsize = newsize;
3411 	bmp->db_maxlevel = BMAPSZTOLEV(bmp->db_mapsize);
3412 
3413 	/* compute new AG size */
3414 	l2agsize = dbGetL2AGSize(newsize);
3415 	oldl2agsize = bmp->db_agl2size;
3416 
3417 	bmp->db_agl2size = l2agsize;
3418 	bmp->db_agsize = 1 << l2agsize;
3419 
3420 	/* compute new number of AG */
3421 	agno = bmp->db_numag;
3422 	bmp->db_numag = newsize >> l2agsize;
3423 	bmp->db_numag += ((u32) newsize % (u32) bmp->db_agsize) ? 1 : 0;
3424 
3425 	/*
3426 	 *	reconfigure db_agfree[]
3427 	 * from old AG configuration to new AG configuration;
3428 	 *
3429 	 * coalesce contiguous k (newAGSize/oldAGSize) AGs;
3430 	 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn;
3431 	 * note: new AG size = old AG size * (2**x).
3432 	 */
3433 	if (l2agsize == oldl2agsize)
3434 		goto extend;
3435 	k = 1 << (l2agsize - oldl2agsize);
3436 	ag_rem = bmp->db_agfree[0];	/* save agfree[0] */
3437 	for (i = 0, n = 0; i < agno; n++) {
3438 		bmp->db_agfree[n] = 0;	/* init collection point */
3439 
3440 		/* coalesce contiguous k AGs; */
3441 		for (j = 0; j < k && i < agno; j++, i++) {
3442 			/* merge AGi to AGn */
3443 			bmp->db_agfree[n] += bmp->db_agfree[i];
3444 		}
3445 	}
3446 	bmp->db_agfree[0] += ag_rem;	/* restore agfree[0] */
3447 
3448 	for (; n < MAXAG; n++)
3449 		bmp->db_agfree[n] = 0;
3450 
3451 	/*
3452 	 * update highest active ag number
3453 	 */
3454 
3455 	bmp->db_maxag = bmp->db_maxag / k;
3456 
3457 	/*
3458 	 *	extend bmap
3459 	 *
3460 	 * update bit maps and corresponding level control pages;
3461 	 * global control page db_nfree, db_agfree[agno], db_maxfreebud;
3462 	 */
3463       extend:
3464 	/* get L2 page */
3465 	p = BMAPBLKNO + nbperpage;	/* L2 page */
3466 	l2mp = read_metapage(ipbmap, p, PSIZE, 0);
3467 	if (!l2mp) {
3468 		jfs_error(ipbmap->i_sb, "L2 page could not be read\n");
3469 		return -EIO;
3470 	}
3471 	l2dcp = (struct dmapctl *) l2mp->data;
3472 
3473 	/* compute start L1 */
3474 	k = blkno >> L2MAXL1SIZE;
3475 	l2leaf = l2dcp->stree + CTLLEAFIND + k;
3476 	p = BLKTOL1(blkno, sbi->l2nbperpage);	/* L1 page */
3477 
3478 	/*
3479 	 * extend each L1 in L2
3480 	 */
3481 	for (; k < LPERCTL; k++, p += nbperpage) {
3482 		/* get L1 page */
3483 		if (j0) {
3484 			/* read in L1 page: (blkno & (MAXL1SIZE - 1)) */
3485 			l1mp = read_metapage(ipbmap, p, PSIZE, 0);
3486 			if (l1mp == NULL)
3487 				goto errout;
3488 			l1dcp = (struct dmapctl *) l1mp->data;
3489 
3490 			/* compute start L0 */
3491 			j = (blkno & (MAXL1SIZE - 1)) >> L2MAXL0SIZE;
3492 			l1leaf = l1dcp->stree + CTLLEAFIND + j;
3493 			p = BLKTOL0(blkno, sbi->l2nbperpage);
3494 			j0 = false;
3495 		} else {
3496 			/* assign/init L1 page */
3497 			l1mp = get_metapage(ipbmap, p, PSIZE, 0);
3498 			if (l1mp == NULL)
3499 				goto errout;
3500 
3501 			l1dcp = (struct dmapctl *) l1mp->data;
3502 
3503 			/* compute start L0 */
3504 			j = 0;
3505 			l1leaf = l1dcp->stree + CTLLEAFIND;
3506 			p += nbperpage;	/* 1st L0 of L1.k */
3507 		}
3508 
3509 		/*
3510 		 * extend each L0 in L1
3511 		 */
3512 		for (; j < LPERCTL; j++) {
3513 			/* get L0 page */
3514 			if (i0) {
3515 				/* read in L0 page: (blkno & (MAXL0SIZE - 1)) */
3516 
3517 				l0mp = read_metapage(ipbmap, p, PSIZE, 0);
3518 				if (l0mp == NULL)
3519 					goto errout;
3520 				l0dcp = (struct dmapctl *) l0mp->data;
3521 
3522 				/* compute start dmap */
3523 				i = (blkno & (MAXL0SIZE - 1)) >>
3524 				    L2BPERDMAP;
3525 				l0leaf = l0dcp->stree + CTLLEAFIND + i;
3526 				p = BLKTODMAP(blkno,
3527 					      sbi->l2nbperpage);
3528 				i0 = false;
3529 			} else {
3530 				/* assign/init L0 page */
3531 				l0mp = get_metapage(ipbmap, p, PSIZE, 0);
3532 				if (l0mp == NULL)
3533 					goto errout;
3534 
3535 				l0dcp = (struct dmapctl *) l0mp->data;
3536 
3537 				/* compute start dmap */
3538 				i = 0;
3539 				l0leaf = l0dcp->stree + CTLLEAFIND;
3540 				p += nbperpage;	/* 1st dmap of L0.j */
3541 			}
3542 
3543 			/*
3544 			 * extend each dmap in L0
3545 			 */
3546 			for (; i < LPERCTL; i++) {
3547 				/*
3548 				 * reconstruct the dmap page, and
3549 				 * initialize corresponding parent L0 leaf
3550 				 */
3551 				if ((n = blkno & (BPERDMAP - 1))) {
3552 					/* read in dmap page: */
3553 					mp = read_metapage(ipbmap, p,
3554 							   PSIZE, 0);
3555 					if (mp == NULL)
3556 						goto errout;
3557 					n = min(nblocks, (s64)BPERDMAP - n);
3558 				} else {
3559 					/* assign/init dmap page */
3560 					mp = read_metapage(ipbmap, p,
3561 							   PSIZE, 0);
3562 					if (mp == NULL)
3563 						goto errout;
3564 
3565 					n = min_t(s64, nblocks, BPERDMAP);
3566 				}
3567 
3568 				dp = (struct dmap *) mp->data;
3569 				*l0leaf = dbInitDmap(dp, blkno, n);
3570 
3571 				bmp->db_nfree += n;
3572 				agno = le64_to_cpu(dp->start) >> l2agsize;
3573 				bmp->db_agfree[agno] += n;
3574 
3575 				write_metapage(mp);
3576 
3577 				l0leaf++;
3578 				p += nbperpage;
3579 
3580 				blkno += n;
3581 				nblocks -= n;
3582 				if (nblocks == 0)
3583 					break;
3584 			}	/* for each dmap in a L0 */
3585 
3586 			/*
3587 			 * build current L0 page from its leaves, and
3588 			 * initialize corresponding parent L1 leaf
3589 			 */
3590 			*l1leaf = dbInitDmapCtl(l0dcp, 0, ++i);
3591 			write_metapage(l0mp);
3592 			l0mp = NULL;
3593 
3594 			if (nblocks)
3595 				l1leaf++;	/* continue for next L0 */
3596 			else {
3597 				/* more than 1 L0 ? */
3598 				if (j > 0)
3599 					break;	/* build L1 page */
3600 				else {
3601 					/* summarize in global bmap page */
3602 					bmp->db_maxfreebud = *l1leaf;
3603 					release_metapage(l1mp);
3604 					release_metapage(l2mp);
3605 					goto finalize;
3606 				}
3607 			}
3608 		}		/* for each L0 in a L1 */
3609 
3610 		/*
3611 		 * build current L1 page from its leaves, and
3612 		 * initialize corresponding parent L2 leaf
3613 		 */
3614 		*l2leaf = dbInitDmapCtl(l1dcp, 1, ++j);
3615 		write_metapage(l1mp);
3616 		l1mp = NULL;
3617 
3618 		if (nblocks)
3619 			l2leaf++;	/* continue for next L1 */
3620 		else {
3621 			/* more than 1 L1 ? */
3622 			if (k > 0)
3623 				break;	/* build L2 page */
3624 			else {
3625 				/* summarize in global bmap page */
3626 				bmp->db_maxfreebud = *l2leaf;
3627 				release_metapage(l2mp);
3628 				goto finalize;
3629 			}
3630 		}
3631 	}			/* for each L1 in a L2 */
3632 
3633 	jfs_error(ipbmap->i_sb, "function has not returned as expected\n");
3634 errout:
3635 	if (l0mp)
3636 		release_metapage(l0mp);
3637 	if (l1mp)
3638 		release_metapage(l1mp);
3639 	release_metapage(l2mp);
3640 	return -EIO;
3641 
3642 	/*
3643 	 *	finalize bmap control page
3644 	 */
3645 finalize:
3646 
3647 	return 0;
3648 }
3649 
3650 
3651 /*
3652  *	dbFinalizeBmap()
3653  */
3654 void dbFinalizeBmap(struct inode *ipbmap)
3655 {
3656 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
3657 	int actags, inactags, l2nl;
3658 	s64 ag_rem, actfree, inactfree, avgfree;
3659 	int i, n;
3660 
3661 	/*
3662 	 *	finalize bmap control page
3663 	 */
3664 //finalize:
3665 	/*
3666 	 * compute db_agpref: preferred ag to allocate from
3667 	 * (the leftmost ag with average free space in it);
3668 	 */
3669 //agpref:
3670 	/* get the number of active ags and inactive ags */
3671 	actags = bmp->db_maxag + 1;
3672 	inactags = bmp->db_numag - actags;
3673 	ag_rem = bmp->db_mapsize & (bmp->db_agsize - 1);	/* ??? */
3674 
3675 	/* determine how many blocks are in the inactive allocation
3676 	 * groups. in doing this, we must account for the fact that
3677 	 * the rightmost group might be a partial group (i.e. file
3678 	 * system size is not a multiple of the group size).
3679 	 */
3680 	inactfree = (inactags && ag_rem) ?
3681 	    ((inactags - 1) << bmp->db_agl2size) + ag_rem
3682 	    : inactags << bmp->db_agl2size;
3683 
3684 	/* determine how many free blocks are in the active
3685 	 * allocation groups plus the average number of free blocks
3686 	 * within the active ags.
3687 	 */
3688 	actfree = bmp->db_nfree - inactfree;
3689 	avgfree = (u32) actfree / (u32) actags;
3690 
3691 	/* if the preferred allocation group has not average free space.
3692 	 * re-establish the preferred group as the leftmost
3693 	 * group with average free space.
3694 	 */
3695 	if (bmp->db_agfree[bmp->db_agpref] < avgfree) {
3696 		for (bmp->db_agpref = 0; bmp->db_agpref < actags;
3697 		     bmp->db_agpref++) {
3698 			if (bmp->db_agfree[bmp->db_agpref] >= avgfree)
3699 				break;
3700 		}
3701 		if (bmp->db_agpref >= bmp->db_numag) {
3702 			jfs_error(ipbmap->i_sb,
3703 				  "cannot find ag with average freespace\n");
3704 		}
3705 	}
3706 
3707 	/*
3708 	 * compute db_aglevel, db_agheight, db_width, db_agstart:
3709 	 * an ag is covered in aglevel dmapctl summary tree,
3710 	 * at agheight level height (from leaf) with agwidth number of nodes
3711 	 * each, which starts at agstart index node of the smmary tree node
3712 	 * array;
3713 	 */
3714 	bmp->db_aglevel = BMAPSZTOLEV(bmp->db_agsize);
3715 	l2nl =
3716 	    bmp->db_agl2size - (L2BPERDMAP + bmp->db_aglevel * L2LPERCTL);
3717 	bmp->db_agheight = l2nl >> 1;
3718 	bmp->db_agwidth = 1 << (l2nl - (bmp->db_agheight << 1));
3719 	for (i = 5 - bmp->db_agheight, bmp->db_agstart = 0, n = 1; i > 0;
3720 	     i--) {
3721 		bmp->db_agstart += n;
3722 		n <<= 2;
3723 	}
3724 
3725 }
3726 
3727 
3728 /*
3729  * NAME:	dbInitDmap()/ujfs_idmap_page()
3730  *
3731  * FUNCTION:	initialize working/persistent bitmap of the dmap page
3732  *		for the specified number of blocks:
3733  *
3734  *		at entry, the bitmaps had been initialized as free (ZEROS);
3735  *		The number of blocks will only account for the actually
3736  *		existing blocks. Blocks which don't actually exist in
3737  *		the aggregate will be marked as allocated (ONES);
3738  *
3739  * PARAMETERS:
3740  *	dp	- pointer to page of map
3741  *	nblocks	- number of blocks this page
3742  *
3743  * RETURNS: NONE
3744  */
3745 static int dbInitDmap(struct dmap * dp, s64 Blkno, int nblocks)
3746 {
3747 	int blkno, w, b, r, nw, nb, i;
3748 
3749 	/* starting block number within the dmap */
3750 	blkno = Blkno & (BPERDMAP - 1);
3751 
3752 	if (blkno == 0) {
3753 		dp->nblocks = dp->nfree = cpu_to_le32(nblocks);
3754 		dp->start = cpu_to_le64(Blkno);
3755 
3756 		if (nblocks == BPERDMAP) {
3757 			memset(&dp->wmap[0], 0, LPERDMAP * 4);
3758 			memset(&dp->pmap[0], 0, LPERDMAP * 4);
3759 			goto initTree;
3760 		}
3761 	} else {
3762 		le32_add_cpu(&dp->nblocks, nblocks);
3763 		le32_add_cpu(&dp->nfree, nblocks);
3764 	}
3765 
3766 	/* word number containing start block number */
3767 	w = blkno >> L2DBWORD;
3768 
3769 	/*
3770 	 * free the bits corresponding to the block range (ZEROS):
3771 	 * note: not all bits of the first and last words may be contained
3772 	 * within the block range.
3773 	 */
3774 	for (r = nblocks; r > 0; r -= nb, blkno += nb) {
3775 		/* number of bits preceding range to be freed in the word */
3776 		b = blkno & (DBWORD - 1);
3777 		/* number of bits to free in the word */
3778 		nb = min(r, DBWORD - b);
3779 
3780 		/* is partial word to be freed ? */
3781 		if (nb < DBWORD) {
3782 			/* free (set to 0) from the bitmap word */
3783 			dp->wmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3784 						     >> b));
3785 			dp->pmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3786 						     >> b));
3787 
3788 			/* skip the word freed */
3789 			w++;
3790 		} else {
3791 			/* free (set to 0) contiguous bitmap words */
3792 			nw = r >> L2DBWORD;
3793 			memset(&dp->wmap[w], 0, nw * 4);
3794 			memset(&dp->pmap[w], 0, nw * 4);
3795 
3796 			/* skip the words freed */
3797 			nb = nw << L2DBWORD;
3798 			w += nw;
3799 		}
3800 	}
3801 
3802 	/*
3803 	 * mark bits following the range to be freed (non-existing
3804 	 * blocks) as allocated (ONES)
3805 	 */
3806 
3807 	if (blkno == BPERDMAP)
3808 		goto initTree;
3809 
3810 	/* the first word beyond the end of existing blocks */
3811 	w = blkno >> L2DBWORD;
3812 
3813 	/* does nblocks fall on a 32-bit boundary ? */
3814 	b = blkno & (DBWORD - 1);
3815 	if (b) {
3816 		/* mark a partial word allocated */
3817 		dp->wmap[w] = dp->pmap[w] = cpu_to_le32(ONES >> b);
3818 		w++;
3819 	}
3820 
3821 	/* set the rest of the words in the page to allocated (ONES) */
3822 	for (i = w; i < LPERDMAP; i++)
3823 		dp->pmap[i] = dp->wmap[i] = cpu_to_le32(ONES);
3824 
3825 	/*
3826 	 * init tree
3827 	 */
3828       initTree:
3829 	return (dbInitDmapTree(dp));
3830 }
3831 
3832 
3833 /*
3834  * NAME:	dbInitDmapTree()/ujfs_complete_dmap()
3835  *
3836  * FUNCTION:	initialize summary tree of the specified dmap:
3837  *
3838  *		at entry, bitmap of the dmap has been initialized;
3839  *
3840  * PARAMETERS:
3841  *	dp	- dmap to complete
3842  *	blkno	- starting block number for this dmap
3843  *	treemax	- will be filled in with max free for this dmap
3844  *
3845  * RETURNS:	max free string at the root of the tree
3846  */
3847 static int dbInitDmapTree(struct dmap * dp)
3848 {
3849 	struct dmaptree *tp;
3850 	s8 *cp;
3851 	int i;
3852 
3853 	/* init fixed info of tree */
3854 	tp = &dp->tree;
3855 	tp->nleafs = cpu_to_le32(LPERDMAP);
3856 	tp->l2nleafs = cpu_to_le32(L2LPERDMAP);
3857 	tp->leafidx = cpu_to_le32(LEAFIND);
3858 	tp->height = cpu_to_le32(4);
3859 	tp->budmin = BUDMIN;
3860 
3861 	/* init each leaf from corresponding wmap word:
3862 	 * note: leaf is set to NOFREE(-1) if all blocks of corresponding
3863 	 * bitmap word are allocated.
3864 	 */
3865 	cp = tp->stree + le32_to_cpu(tp->leafidx);
3866 	for (i = 0; i < LPERDMAP; i++)
3867 		*cp++ = dbMaxBud((u8 *) & dp->wmap[i]);
3868 
3869 	/* build the dmap's binary buddy summary tree */
3870 	return (dbInitTree(tp));
3871 }
3872 
3873 
3874 /*
3875  * NAME:	dbInitTree()/ujfs_adjtree()
3876  *
3877  * FUNCTION:	initialize binary buddy summary tree of a dmap or dmapctl.
3878  *
3879  *		at entry, the leaves of the tree has been initialized
3880  *		from corresponding bitmap word or root of summary tree
3881  *		of the child control page;
3882  *		configure binary buddy system at the leaf level, then
3883  *		bubble up the values of the leaf nodes up the tree.
3884  *
3885  * PARAMETERS:
3886  *	cp	- Pointer to the root of the tree
3887  *	l2leaves- Number of leaf nodes as a power of 2
3888  *	l2min	- Number of blocks that can be covered by a leaf
3889  *		  as a power of 2
3890  *
3891  * RETURNS: max free string at the root of the tree
3892  */
3893 static int dbInitTree(struct dmaptree * dtp)
3894 {
3895 	int l2max, l2free, bsize, nextb, i;
3896 	int child, parent, nparent;
3897 	s8 *tp, *cp, *cp1;
3898 
3899 	tp = dtp->stree;
3900 
3901 	/* Determine the maximum free string possible for the leaves */
3902 	l2max = le32_to_cpu(dtp->l2nleafs) + dtp->budmin;
3903 
3904 	/*
3905 	 * configure the leaf levevl into binary buddy system
3906 	 *
3907 	 * Try to combine buddies starting with a buddy size of 1
3908 	 * (i.e. two leaves). At a buddy size of 1 two buddy leaves
3909 	 * can be combined if both buddies have a maximum free of l2min;
3910 	 * the combination will result in the left-most buddy leaf having
3911 	 * a maximum free of l2min+1.
3912 	 * After processing all buddies for a given size, process buddies
3913 	 * at the next higher buddy size (i.e. current size * 2) and
3914 	 * the next maximum free (current free + 1).
3915 	 * This continues until the maximum possible buddy combination
3916 	 * yields maximum free.
3917 	 */
3918 	for (l2free = dtp->budmin, bsize = 1; l2free < l2max;
3919 	     l2free++, bsize = nextb) {
3920 		/* get next buddy size == current buddy pair size */
3921 		nextb = bsize << 1;
3922 
3923 		/* scan each adjacent buddy pair at current buddy size */
3924 		for (i = 0, cp = tp + le32_to_cpu(dtp->leafidx);
3925 		     i < le32_to_cpu(dtp->nleafs);
3926 		     i += nextb, cp += nextb) {
3927 			/* coalesce if both adjacent buddies are max free */
3928 			if (*cp == l2free && *(cp + bsize) == l2free) {
3929 				*cp = l2free + 1;	/* left take right */
3930 				*(cp + bsize) = -1;	/* right give left */
3931 			}
3932 		}
3933 	}
3934 
3935 	/*
3936 	 * bubble summary information of leaves up the tree.
3937 	 *
3938 	 * Starting at the leaf node level, the four nodes described by
3939 	 * the higher level parent node are compared for a maximum free and
3940 	 * this maximum becomes the value of the parent node.
3941 	 * when all lower level nodes are processed in this fashion then
3942 	 * move up to the next level (parent becomes a lower level node) and
3943 	 * continue the process for that level.
3944 	 */
3945 	for (child = le32_to_cpu(dtp->leafidx),
3946 	     nparent = le32_to_cpu(dtp->nleafs) >> 2;
3947 	     nparent > 0; nparent >>= 2, child = parent) {
3948 		/* get index of 1st node of parent level */
3949 		parent = (child - 1) >> 2;
3950 
3951 		/* set the value of the parent node as the maximum
3952 		 * of the four nodes of the current level.
3953 		 */
3954 		for (i = 0, cp = tp + child, cp1 = tp + parent;
3955 		     i < nparent; i++, cp += 4, cp1++)
3956 			*cp1 = TREEMAX(cp);
3957 	}
3958 
3959 	return (*tp);
3960 }
3961 
3962 
3963 /*
3964  *	dbInitDmapCtl()
3965  *
3966  * function: initialize dmapctl page
3967  */
3968 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i)
3969 {				/* start leaf index not covered by range */
3970 	s8 *cp;
3971 
3972 	dcp->nleafs = cpu_to_le32(LPERCTL);
3973 	dcp->l2nleafs = cpu_to_le32(L2LPERCTL);
3974 	dcp->leafidx = cpu_to_le32(CTLLEAFIND);
3975 	dcp->height = cpu_to_le32(5);
3976 	dcp->budmin = L2BPERDMAP + L2LPERCTL * level;
3977 
3978 	/*
3979 	 * initialize the leaves of current level that were not covered
3980 	 * by the specified input block range (i.e. the leaves have no
3981 	 * low level dmapctl or dmap).
3982 	 */
3983 	cp = &dcp->stree[CTLLEAFIND + i];
3984 	for (; i < LPERCTL; i++)
3985 		*cp++ = NOFREE;
3986 
3987 	/* build the dmap's binary buddy summary tree */
3988 	return (dbInitTree((struct dmaptree *) dcp));
3989 }
3990 
3991 
3992 /*
3993  * NAME:	dbGetL2AGSize()/ujfs_getagl2size()
3994  *
3995  * FUNCTION:	Determine log2(allocation group size) from aggregate size
3996  *
3997  * PARAMETERS:
3998  *	nblocks	- Number of blocks in aggregate
3999  *
4000  * RETURNS: log2(allocation group size) in aggregate blocks
4001  */
4002 static int dbGetL2AGSize(s64 nblocks)
4003 {
4004 	s64 sz;
4005 	s64 m;
4006 	int l2sz;
4007 
4008 	if (nblocks < BPERDMAP * MAXAG)
4009 		return (L2BPERDMAP);
4010 
4011 	/* round up aggregate size to power of 2 */
4012 	m = ((u64) 1 << (64 - 1));
4013 	for (l2sz = 64; l2sz >= 0; l2sz--, m >>= 1) {
4014 		if (m & nblocks)
4015 			break;
4016 	}
4017 
4018 	sz = (s64) 1 << l2sz;
4019 	if (sz < nblocks)
4020 		l2sz += 1;
4021 
4022 	/* agsize = roundupSize/max_number_of_ag */
4023 	return (l2sz - L2MAXAG);
4024 }
4025 
4026 
4027 /*
4028  * NAME:	dbMapFileSizeToMapSize()
4029  *
4030  * FUNCTION:	compute number of blocks the block allocation map file
4031  *		can cover from the map file size;
4032  *
4033  * RETURNS:	Number of blocks which can be covered by this block map file;
4034  */
4035 
4036 /*
4037  * maximum number of map pages at each level including control pages
4038  */
4039 #define MAXL0PAGES	(1 + LPERCTL)
4040 #define MAXL1PAGES	(1 + LPERCTL * MAXL0PAGES)
4041 
4042 /*
4043  * convert number of map pages to the zero origin top dmapctl level
4044  */
4045 #define BMAPPGTOLEV(npages)	\
4046 	(((npages) <= 3 + MAXL0PAGES) ? 0 : \
4047 	 ((npages) <= 2 + MAXL1PAGES) ? 1 : 2)
4048 
4049 s64 dbMapFileSizeToMapSize(struct inode * ipbmap)
4050 {
4051 	struct super_block *sb = ipbmap->i_sb;
4052 	s64 nblocks;
4053 	s64 npages, ndmaps;
4054 	int level, i;
4055 	int complete, factor;
4056 
4057 	nblocks = ipbmap->i_size >> JFS_SBI(sb)->l2bsize;
4058 	npages = nblocks >> JFS_SBI(sb)->l2nbperpage;
4059 	level = BMAPPGTOLEV(npages);
4060 
4061 	/* At each level, accumulate the number of dmap pages covered by
4062 	 * the number of full child levels below it;
4063 	 * repeat for the last incomplete child level.
4064 	 */
4065 	ndmaps = 0;
4066 	npages--;		/* skip the first global control page */
4067 	/* skip higher level control pages above top level covered by map */
4068 	npages -= (2 - level);
4069 	npages--;		/* skip top level's control page */
4070 	for (i = level; i >= 0; i--) {
4071 		factor =
4072 		    (i == 2) ? MAXL1PAGES : ((i == 1) ? MAXL0PAGES : 1);
4073 		complete = (u32) npages / factor;
4074 		ndmaps += complete * ((i == 2) ? LPERCTL * LPERCTL :
4075 				      ((i == 1) ? LPERCTL : 1));
4076 
4077 		/* pages in last/incomplete child */
4078 		npages = (u32) npages % factor;
4079 		/* skip incomplete child's level control page */
4080 		npages--;
4081 	}
4082 
4083 	/* convert the number of dmaps into the number of blocks
4084 	 * which can be covered by the dmaps;
4085 	 */
4086 	nblocks = ndmaps << L2BPERDMAP;
4087 
4088 	return (nblocks);
4089 }
4090