xref: /kernel/linux/linux-5.10/fs/jfs/jfs_dmap.c (revision 8c2ecf20)
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 *   Copyright (C) International Business Machines Corp., 2000-2004
4 *   Portions Copyright (C) Tino Reichardt, 2012
5 */
6
7#include <linux/fs.h>
8#include <linux/slab.h>
9#include "jfs_incore.h"
10#include "jfs_superblock.h"
11#include "jfs_dmap.h"
12#include "jfs_imap.h"
13#include "jfs_lock.h"
14#include "jfs_metapage.h"
15#include "jfs_debug.h"
16#include "jfs_discard.h"
17
18/*
19 *	SERIALIZATION of the Block Allocation Map.
20 *
21 *	the working state of the block allocation map is accessed in
22 *	two directions:
23 *
24 *	1) allocation and free requests that start at the dmap
25 *	   level and move up through the dmap control pages (i.e.
26 *	   the vast majority of requests).
27 *
28 *	2) allocation requests that start at dmap control page
29 *	   level and work down towards the dmaps.
30 *
31 *	the serialization scheme used here is as follows.
32 *
33 *	requests which start at the bottom are serialized against each
34 *	other through buffers and each requests holds onto its buffers
35 *	as it works it way up from a single dmap to the required level
36 *	of dmap control page.
37 *	requests that start at the top are serialized against each other
38 *	and request that start from the bottom by the multiple read/single
39 *	write inode lock of the bmap inode. requests starting at the top
40 *	take this lock in write mode while request starting at the bottom
41 *	take the lock in read mode.  a single top-down request may proceed
42 *	exclusively while multiple bottoms-up requests may proceed
43 *	simultaneously (under the protection of busy buffers).
44 *
45 *	in addition to information found in dmaps and dmap control pages,
46 *	the working state of the block allocation map also includes read/
47 *	write information maintained in the bmap descriptor (i.e. total
48 *	free block count, allocation group level free block counts).
49 *	a single exclusive lock (BMAP_LOCK) is used to guard this information
50 *	in the face of multiple-bottoms up requests.
51 *	(lock ordering: IREAD_LOCK, BMAP_LOCK);
52 *
53 *	accesses to the persistent state of the block allocation map (limited
54 *	to the persistent bitmaps in dmaps) is guarded by (busy) buffers.
55 */
56
57#define BMAP_LOCK_INIT(bmp)	mutex_init(&bmp->db_bmaplock)
58#define BMAP_LOCK(bmp)		mutex_lock(&bmp->db_bmaplock)
59#define BMAP_UNLOCK(bmp)	mutex_unlock(&bmp->db_bmaplock)
60
61/*
62 * forward references
63 */
64static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
65			int nblocks);
66static void dbSplit(dmtree_t *tp, int leafno, int splitsz, int newval, bool is_ctl);
67static int dbBackSplit(dmtree_t *tp, int leafno, bool is_ctl);
68static int dbJoin(dmtree_t *tp, int leafno, int newval, bool is_ctl);
69static void dbAdjTree(dmtree_t *tp, int leafno, int newval, bool is_ctl);
70static int dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc,
71		    int level);
72static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results);
73static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
74		       int nblocks);
75static int dbAllocNear(struct bmap * bmp, struct dmap * dp, s64 blkno,
76		       int nblocks,
77		       int l2nb, s64 * results);
78static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
79		       int nblocks);
80static int dbAllocDmapLev(struct bmap * bmp, struct dmap * dp, int nblocks,
81			  int l2nb,
82			  s64 * results);
83static int dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb,
84		     s64 * results);
85static int dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno,
86		      s64 * results);
87static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks);
88static int dbFindBits(u32 word, int l2nb);
89static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno);
90static int dbFindLeaf(dmtree_t *tp, int l2nb, int *leafidx, bool is_ctl);
91static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
92		      int nblocks);
93static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
94		      int nblocks);
95static int dbMaxBud(u8 * cp);
96static int blkstol2(s64 nb);
97
98static int cntlz(u32 value);
99static int cnttz(u32 word);
100
101static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
102			 int nblocks);
103static int dbInitDmap(struct dmap * dp, s64 blkno, int nblocks);
104static int dbInitDmapTree(struct dmap * dp);
105static int dbInitTree(struct dmaptree * dtp);
106static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i);
107static int dbGetL2AGSize(s64 nblocks);
108
109/*
110 *	buddy table
111 *
112 * table used for determining buddy sizes within characters of
113 * dmap bitmap words.  the characters themselves serve as indexes
114 * into the table, with the table elements yielding the maximum
115 * binary buddy of free bits within the character.
116 */
117static const s8 budtab[256] = {
118	3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
119	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
120	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
121	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
122	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
123	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
124	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
125	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
126	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
127	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
128	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
129	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
130	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
131	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
132	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
133	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, -1
134};
135
136/*
137 * NAME:	dbMount()
138 *
139 * FUNCTION:	initializate the block allocation map.
140 *
141 *		memory is allocated for the in-core bmap descriptor and
142 *		the in-core descriptor is initialized from disk.
143 *
144 * PARAMETERS:
145 *	ipbmap	- pointer to in-core inode for the block map.
146 *
147 * RETURN VALUES:
148 *	0	- success
149 *	-ENOMEM	- insufficient memory
150 *	-EIO	- i/o error
151 *	-EINVAL - wrong bmap data
152 */
153int dbMount(struct inode *ipbmap)
154{
155	struct bmap *bmp;
156	struct dbmap_disk *dbmp_le;
157	struct metapage *mp;
158	int i, err;
159
160	/*
161	 * allocate/initialize the in-memory bmap descriptor
162	 */
163	/* allocate memory for the in-memory bmap descriptor */
164	bmp = kmalloc(sizeof(struct bmap), GFP_KERNEL);
165	if (bmp == NULL)
166		return -ENOMEM;
167
168	/* read the on-disk bmap descriptor. */
169	mp = read_metapage(ipbmap,
170			   BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
171			   PSIZE, 0);
172	if (mp == NULL) {
173		err = -EIO;
174		goto err_kfree_bmp;
175	}
176
177	/* copy the on-disk bmap descriptor to its in-memory version. */
178	dbmp_le = (struct dbmap_disk *) mp->data;
179	bmp->db_mapsize = le64_to_cpu(dbmp_le->dn_mapsize);
180	bmp->db_nfree = le64_to_cpu(dbmp_le->dn_nfree);
181
182	bmp->db_l2nbperpage = le32_to_cpu(dbmp_le->dn_l2nbperpage);
183	if (bmp->db_l2nbperpage > L2PSIZE - L2MINBLOCKSIZE ||
184		bmp->db_l2nbperpage < 0) {
185		err = -EINVAL;
186		goto err_release_metapage;
187	}
188
189	bmp->db_numag = le32_to_cpu(dbmp_le->dn_numag);
190	if (!bmp->db_numag) {
191		err = -EINVAL;
192		goto err_release_metapage;
193	}
194
195	bmp->db_maxlevel = le32_to_cpu(dbmp_le->dn_maxlevel);
196	bmp->db_maxag = le32_to_cpu(dbmp_le->dn_maxag);
197	bmp->db_agpref = le32_to_cpu(dbmp_le->dn_agpref);
198	if (bmp->db_maxag >= MAXAG || bmp->db_maxag < 0 ||
199		bmp->db_agpref >= MAXAG || bmp->db_agpref < 0) {
200		err = -EINVAL;
201		goto err_release_metapage;
202	}
203
204	bmp->db_aglevel = le32_to_cpu(dbmp_le->dn_aglevel);
205	bmp->db_agheight = le32_to_cpu(dbmp_le->dn_agheight);
206	bmp->db_agwidth = le32_to_cpu(dbmp_le->dn_agwidth);
207	bmp->db_agstart = le32_to_cpu(dbmp_le->dn_agstart);
208	bmp->db_agl2size = le32_to_cpu(dbmp_le->dn_agl2size);
209	if (bmp->db_agl2size > L2MAXL2SIZE - L2MAXAG ||
210	    bmp->db_agl2size < 0) {
211		err = -EINVAL;
212		goto err_release_metapage;
213	}
214
215	if (((bmp->db_mapsize - 1) >> bmp->db_agl2size) > MAXAG) {
216		err = -EINVAL;
217		goto err_release_metapage;
218	}
219
220	for (i = 0; i < MAXAG; i++)
221		bmp->db_agfree[i] = le64_to_cpu(dbmp_le->dn_agfree[i]);
222	bmp->db_agsize = le64_to_cpu(dbmp_le->dn_agsize);
223	bmp->db_maxfreebud = dbmp_le->dn_maxfreebud;
224
225	/* release the buffer. */
226	release_metapage(mp);
227
228	/* bind the bmap inode and the bmap descriptor to each other. */
229	bmp->db_ipbmap = ipbmap;
230	JFS_SBI(ipbmap->i_sb)->bmap = bmp;
231
232	memset(bmp->db_active, 0, sizeof(bmp->db_active));
233
234	/*
235	 * allocate/initialize the bmap lock
236	 */
237	BMAP_LOCK_INIT(bmp);
238
239	return (0);
240
241err_release_metapage:
242	release_metapage(mp);
243err_kfree_bmp:
244	kfree(bmp);
245	return err;
246}
247
248
249/*
250 * NAME:	dbUnmount()
251 *
252 * FUNCTION:	terminate the block allocation map in preparation for
253 *		file system unmount.
254 *
255 *		the in-core bmap descriptor is written to disk and
256 *		the memory for this descriptor is freed.
257 *
258 * PARAMETERS:
259 *	ipbmap	- pointer to in-core inode for the block map.
260 *
261 * RETURN VALUES:
262 *	0	- success
263 *	-EIO	- i/o error
264 */
265int dbUnmount(struct inode *ipbmap, int mounterror)
266{
267	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
268
269	if (!(mounterror || isReadOnly(ipbmap)))
270		dbSync(ipbmap);
271
272	/*
273	 * Invalidate the page cache buffers
274	 */
275	truncate_inode_pages(ipbmap->i_mapping, 0);
276
277	/* free the memory for the in-memory bmap. */
278	kfree(bmp);
279	JFS_SBI(ipbmap->i_sb)->bmap = NULL;
280
281	return (0);
282}
283
284/*
285 *	dbSync()
286 */
287int dbSync(struct inode *ipbmap)
288{
289	struct dbmap_disk *dbmp_le;
290	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
291	struct metapage *mp;
292	int i;
293
294	/*
295	 * write bmap global control page
296	 */
297	/* get the buffer for the on-disk bmap descriptor. */
298	mp = read_metapage(ipbmap,
299			   BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
300			   PSIZE, 0);
301	if (mp == NULL) {
302		jfs_err("dbSync: read_metapage failed!");
303		return -EIO;
304	}
305	/* copy the in-memory version of the bmap to the on-disk version */
306	dbmp_le = (struct dbmap_disk *) mp->data;
307	dbmp_le->dn_mapsize = cpu_to_le64(bmp->db_mapsize);
308	dbmp_le->dn_nfree = cpu_to_le64(bmp->db_nfree);
309	dbmp_le->dn_l2nbperpage = cpu_to_le32(bmp->db_l2nbperpage);
310	dbmp_le->dn_numag = cpu_to_le32(bmp->db_numag);
311	dbmp_le->dn_maxlevel = cpu_to_le32(bmp->db_maxlevel);
312	dbmp_le->dn_maxag = cpu_to_le32(bmp->db_maxag);
313	dbmp_le->dn_agpref = cpu_to_le32(bmp->db_agpref);
314	dbmp_le->dn_aglevel = cpu_to_le32(bmp->db_aglevel);
315	dbmp_le->dn_agheight = cpu_to_le32(bmp->db_agheight);
316	dbmp_le->dn_agwidth = cpu_to_le32(bmp->db_agwidth);
317	dbmp_le->dn_agstart = cpu_to_le32(bmp->db_agstart);
318	dbmp_le->dn_agl2size = cpu_to_le32(bmp->db_agl2size);
319	for (i = 0; i < MAXAG; i++)
320		dbmp_le->dn_agfree[i] = cpu_to_le64(bmp->db_agfree[i]);
321	dbmp_le->dn_agsize = cpu_to_le64(bmp->db_agsize);
322	dbmp_le->dn_maxfreebud = bmp->db_maxfreebud;
323
324	/* write the buffer */
325	write_metapage(mp);
326
327	/*
328	 * write out dirty pages of bmap
329	 */
330	filemap_write_and_wait(ipbmap->i_mapping);
331
332	diWriteSpecial(ipbmap, 0);
333
334	return (0);
335}
336
337/*
338 * NAME:	dbFree()
339 *
340 * FUNCTION:	free the specified block range from the working block
341 *		allocation map.
342 *
343 *		the blocks will be free from the working map one dmap
344 *		at a time.
345 *
346 * PARAMETERS:
347 *	ip	- pointer to in-core inode;
348 *	blkno	- starting block number to be freed.
349 *	nblocks	- number of blocks to be freed.
350 *
351 * RETURN VALUES:
352 *	0	- success
353 *	-EIO	- i/o error
354 */
355int dbFree(struct inode *ip, s64 blkno, s64 nblocks)
356{
357	struct metapage *mp;
358	struct dmap *dp;
359	int nb, rc;
360	s64 lblkno, rem;
361	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
362	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
363	struct super_block *sb = ipbmap->i_sb;
364
365	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
366
367	/* block to be freed better be within the mapsize. */
368	if (unlikely((blkno == 0) || (blkno + nblocks > bmp->db_mapsize))) {
369		IREAD_UNLOCK(ipbmap);
370		printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
371		       (unsigned long long) blkno,
372		       (unsigned long long) nblocks);
373		jfs_error(ip->i_sb, "block to be freed is outside the map\n");
374		return -EIO;
375	}
376
377	/**
378	 * TRIM the blocks, when mounted with discard option
379	 */
380	if (JFS_SBI(sb)->flag & JFS_DISCARD)
381		if (JFS_SBI(sb)->minblks_trim <= nblocks)
382			jfs_issue_discard(ipbmap, blkno, nblocks);
383
384	/*
385	 * free the blocks a dmap at a time.
386	 */
387	mp = NULL;
388	for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
389		/* release previous dmap if any */
390		if (mp) {
391			write_metapage(mp);
392		}
393
394		/* get the buffer for the current dmap. */
395		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
396		mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
397		if (mp == NULL) {
398			IREAD_UNLOCK(ipbmap);
399			return -EIO;
400		}
401		dp = (struct dmap *) mp->data;
402
403		/* determine the number of blocks to be freed from
404		 * this dmap.
405		 */
406		nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
407
408		/* free the blocks. */
409		if ((rc = dbFreeDmap(bmp, dp, blkno, nb))) {
410			jfs_error(ip->i_sb, "error in block map\n");
411			release_metapage(mp);
412			IREAD_UNLOCK(ipbmap);
413			return (rc);
414		}
415	}
416
417	/* write the last buffer. */
418	if (mp)
419		write_metapage(mp);
420
421	IREAD_UNLOCK(ipbmap);
422
423	return (0);
424}
425
426
427/*
428 * NAME:	dbUpdatePMap()
429 *
430 * FUNCTION:	update the allocation state (free or allocate) of the
431 *		specified block range in the persistent block allocation map.
432 *
433 *		the blocks will be updated in the persistent map one
434 *		dmap at a time.
435 *
436 * PARAMETERS:
437 *	ipbmap	- pointer to in-core inode for the block map.
438 *	free	- 'true' if block range is to be freed from the persistent
439 *		  map; 'false' if it is to be allocated.
440 *	blkno	- starting block number of the range.
441 *	nblocks	- number of contiguous blocks in the range.
442 *	tblk	- transaction block;
443 *
444 * RETURN VALUES:
445 *	0	- success
446 *	-EIO	- i/o error
447 */
448int
449dbUpdatePMap(struct inode *ipbmap,
450	     int free, s64 blkno, s64 nblocks, struct tblock * tblk)
451{
452	int nblks, dbitno, wbitno, rbits;
453	int word, nbits, nwords;
454	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
455	s64 lblkno, rem, lastlblkno;
456	u32 mask;
457	struct dmap *dp;
458	struct metapage *mp;
459	struct jfs_log *log;
460	int lsn, difft, diffp;
461	unsigned long flags;
462
463	/* the blocks better be within the mapsize. */
464	if (blkno + nblocks > bmp->db_mapsize) {
465		printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
466		       (unsigned long long) blkno,
467		       (unsigned long long) nblocks);
468		jfs_error(ipbmap->i_sb, "blocks are outside the map\n");
469		return -EIO;
470	}
471
472	/* compute delta of transaction lsn from log syncpt */
473	lsn = tblk->lsn;
474	log = (struct jfs_log *) JFS_SBI(tblk->sb)->log;
475	logdiff(difft, lsn, log);
476
477	/*
478	 * update the block state a dmap at a time.
479	 */
480	mp = NULL;
481	lastlblkno = 0;
482	for (rem = nblocks; rem > 0; rem -= nblks, blkno += nblks) {
483		/* get the buffer for the current dmap. */
484		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
485		if (lblkno != lastlblkno) {
486			if (mp) {
487				write_metapage(mp);
488			}
489
490			mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE,
491					   0);
492			if (mp == NULL)
493				return -EIO;
494			metapage_wait_for_io(mp);
495		}
496		dp = (struct dmap *) mp->data;
497
498		/* determine the bit number and word within the dmap of
499		 * the starting block.  also determine how many blocks
500		 * are to be updated within this dmap.
501		 */
502		dbitno = blkno & (BPERDMAP - 1);
503		word = dbitno >> L2DBWORD;
504		nblks = min(rem, (s64)BPERDMAP - dbitno);
505
506		/* update the bits of the dmap words. the first and last
507		 * words may only have a subset of their bits updated. if
508		 * this is the case, we'll work against that word (i.e.
509		 * partial first and/or last) only in a single pass.  a
510		 * single pass will also be used to update all words that
511		 * are to have all their bits updated.
512		 */
513		for (rbits = nblks; rbits > 0;
514		     rbits -= nbits, dbitno += nbits) {
515			/* determine the bit number within the word and
516			 * the number of bits within the word.
517			 */
518			wbitno = dbitno & (DBWORD - 1);
519			nbits = min(rbits, DBWORD - wbitno);
520
521			/* check if only part of the word is to be updated. */
522			if (nbits < DBWORD) {
523				/* update (free or allocate) the bits
524				 * in this word.
525				 */
526				mask =
527				    (ONES << (DBWORD - nbits) >> wbitno);
528				if (free)
529					dp->pmap[word] &=
530					    cpu_to_le32(~mask);
531				else
532					dp->pmap[word] |=
533					    cpu_to_le32(mask);
534
535				word += 1;
536			} else {
537				/* one or more words are to have all
538				 * their bits updated.  determine how
539				 * many words and how many bits.
540				 */
541				nwords = rbits >> L2DBWORD;
542				nbits = nwords << L2DBWORD;
543
544				/* update (free or allocate) the bits
545				 * in these words.
546				 */
547				if (free)
548					memset(&dp->pmap[word], 0,
549					       nwords * 4);
550				else
551					memset(&dp->pmap[word], (int) ONES,
552					       nwords * 4);
553
554				word += nwords;
555			}
556		}
557
558		/*
559		 * update dmap lsn
560		 */
561		if (lblkno == lastlblkno)
562			continue;
563
564		lastlblkno = lblkno;
565
566		LOGSYNC_LOCK(log, flags);
567		if (mp->lsn != 0) {
568			/* inherit older/smaller lsn */
569			logdiff(diffp, mp->lsn, log);
570			if (difft < diffp) {
571				mp->lsn = lsn;
572
573				/* move bp after tblock in logsync list */
574				list_move(&mp->synclist, &tblk->synclist);
575			}
576
577			/* inherit younger/larger clsn */
578			logdiff(difft, tblk->clsn, log);
579			logdiff(diffp, mp->clsn, log);
580			if (difft > diffp)
581				mp->clsn = tblk->clsn;
582		} else {
583			mp->log = log;
584			mp->lsn = lsn;
585
586			/* insert bp after tblock in logsync list */
587			log->count++;
588			list_add(&mp->synclist, &tblk->synclist);
589
590			mp->clsn = tblk->clsn;
591		}
592		LOGSYNC_UNLOCK(log, flags);
593	}
594
595	/* write the last buffer. */
596	if (mp) {
597		write_metapage(mp);
598	}
599
600	return (0);
601}
602
603
604/*
605 * NAME:	dbNextAG()
606 *
607 * FUNCTION:	find the preferred allocation group for new allocations.
608 *
609 *		Within the allocation groups, we maintain a preferred
610 *		allocation group which consists of a group with at least
611 *		average free space.  It is the preferred group that we target
612 *		new inode allocation towards.  The tie-in between inode
613 *		allocation and block allocation occurs as we allocate the
614 *		first (data) block of an inode and specify the inode (block)
615 *		as the allocation hint for this block.
616 *
617 *		We try to avoid having more than one open file growing in
618 *		an allocation group, as this will lead to fragmentation.
619 *		This differs from the old OS/2 method of trying to keep
620 *		empty ags around for large allocations.
621 *
622 * PARAMETERS:
623 *	ipbmap	- pointer to in-core inode for the block map.
624 *
625 * RETURN VALUES:
626 *	the preferred allocation group number.
627 */
628int dbNextAG(struct inode *ipbmap)
629{
630	s64 avgfree;
631	int agpref;
632	s64 hwm = 0;
633	int i;
634	int next_best = -1;
635	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
636
637	BMAP_LOCK(bmp);
638
639	/* determine the average number of free blocks within the ags. */
640	avgfree = (u32)bmp->db_nfree / bmp->db_numag;
641
642	/*
643	 * if the current preferred ag does not have an active allocator
644	 * and has at least average freespace, return it
645	 */
646	agpref = bmp->db_agpref;
647	if ((atomic_read(&bmp->db_active[agpref]) == 0) &&
648	    (bmp->db_agfree[agpref] >= avgfree))
649		goto unlock;
650
651	/* From the last preferred ag, find the next one with at least
652	 * average free space.
653	 */
654	for (i = 0 ; i < bmp->db_numag; i++, agpref++) {
655		if (agpref == bmp->db_numag)
656			agpref = 0;
657
658		if (atomic_read(&bmp->db_active[agpref]))
659			/* open file is currently growing in this ag */
660			continue;
661		if (bmp->db_agfree[agpref] >= avgfree) {
662			/* Return this one */
663			bmp->db_agpref = agpref;
664			goto unlock;
665		} else if (bmp->db_agfree[agpref] > hwm) {
666			/* Less than avg. freespace, but best so far */
667			hwm = bmp->db_agfree[agpref];
668			next_best = agpref;
669		}
670	}
671
672	/*
673	 * If no inactive ag was found with average freespace, use the
674	 * next best
675	 */
676	if (next_best != -1)
677		bmp->db_agpref = next_best;
678	/* else leave db_agpref unchanged */
679unlock:
680	BMAP_UNLOCK(bmp);
681
682	/* return the preferred group.
683	 */
684	return (bmp->db_agpref);
685}
686
687/*
688 * NAME:	dbAlloc()
689 *
690 * FUNCTION:	attempt to allocate a specified number of contiguous free
691 *		blocks from the working allocation block map.
692 *
693 *		the block allocation policy uses hints and a multi-step
694 *		approach.
695 *
696 *		for allocation requests smaller than the number of blocks
697 *		per dmap, we first try to allocate the new blocks
698 *		immediately following the hint.  if these blocks are not
699 *		available, we try to allocate blocks near the hint.  if
700 *		no blocks near the hint are available, we next try to
701 *		allocate within the same dmap as contains the hint.
702 *
703 *		if no blocks are available in the dmap or the allocation
704 *		request is larger than the dmap size, we try to allocate
705 *		within the same allocation group as contains the hint. if
706 *		this does not succeed, we finally try to allocate anywhere
707 *		within the aggregate.
708 *
709 *		we also try to allocate anywhere within the aggregate for
710 *		for allocation requests larger than the allocation group
711 *		size or requests that specify no hint value.
712 *
713 * PARAMETERS:
714 *	ip	- pointer to in-core inode;
715 *	hint	- allocation hint.
716 *	nblocks	- number of contiguous blocks in the range.
717 *	results	- on successful return, set to the starting block number
718 *		  of the newly allocated contiguous range.
719 *
720 * RETURN VALUES:
721 *	0	- success
722 *	-ENOSPC	- insufficient disk resources
723 *	-EIO	- i/o error
724 */
725int dbAlloc(struct inode *ip, s64 hint, s64 nblocks, s64 * results)
726{
727	int rc, agno;
728	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
729	struct bmap *bmp;
730	struct metapage *mp;
731	s64 lblkno, blkno;
732	struct dmap *dp;
733	int l2nb;
734	s64 mapSize;
735	int writers;
736
737	/* assert that nblocks is valid */
738	assert(nblocks > 0);
739
740	/* get the log2 number of blocks to be allocated.
741	 * if the number of blocks is not a log2 multiple,
742	 * it will be rounded up to the next log2 multiple.
743	 */
744	l2nb = BLKSTOL2(nblocks);
745
746	bmp = JFS_SBI(ip->i_sb)->bmap;
747
748	mapSize = bmp->db_mapsize;
749
750	/* the hint should be within the map */
751	if (hint >= mapSize) {
752		jfs_error(ip->i_sb, "the hint is outside the map\n");
753		return -EIO;
754	}
755
756	/* if the number of blocks to be allocated is greater than the
757	 * allocation group size, try to allocate anywhere.
758	 */
759	if (l2nb > bmp->db_agl2size) {
760		IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
761
762		rc = dbAllocAny(bmp, nblocks, l2nb, results);
763
764		goto write_unlock;
765	}
766
767	/*
768	 * If no hint, let dbNextAG recommend an allocation group
769	 */
770	if (hint == 0)
771		goto pref_ag;
772
773	/* we would like to allocate close to the hint.  adjust the
774	 * hint to the block following the hint since the allocators
775	 * will start looking for free space starting at this point.
776	 */
777	blkno = hint + 1;
778
779	if (blkno >= bmp->db_mapsize)
780		goto pref_ag;
781
782	agno = blkno >> bmp->db_agl2size;
783
784	/* check if blkno crosses over into a new allocation group.
785	 * if so, check if we should allow allocations within this
786	 * allocation group.
787	 */
788	if ((blkno & (bmp->db_agsize - 1)) == 0)
789		/* check if the AG is currently being written to.
790		 * if so, call dbNextAG() to find a non-busy
791		 * AG with sufficient free space.
792		 */
793		if (atomic_read(&bmp->db_active[agno]))
794			goto pref_ag;
795
796	/* check if the allocation request size can be satisfied from a
797	 * single dmap.  if so, try to allocate from the dmap containing
798	 * the hint using a tiered strategy.
799	 */
800	if (nblocks <= BPERDMAP) {
801		IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
802
803		/* get the buffer for the dmap containing the hint.
804		 */
805		rc = -EIO;
806		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
807		mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
808		if (mp == NULL)
809			goto read_unlock;
810
811		dp = (struct dmap *) mp->data;
812
813		/* first, try to satisfy the allocation request with the
814		 * blocks beginning at the hint.
815		 */
816		if ((rc = dbAllocNext(bmp, dp, blkno, (int) nblocks))
817		    != -ENOSPC) {
818			if (rc == 0) {
819				*results = blkno;
820				mark_metapage_dirty(mp);
821			}
822
823			release_metapage(mp);
824			goto read_unlock;
825		}
826
827		writers = atomic_read(&bmp->db_active[agno]);
828		if ((writers > 1) ||
829		    ((writers == 1) && (JFS_IP(ip)->active_ag != agno))) {
830			/*
831			 * Someone else is writing in this allocation
832			 * group.  To avoid fragmenting, try another ag
833			 */
834			release_metapage(mp);
835			IREAD_UNLOCK(ipbmap);
836			goto pref_ag;
837		}
838
839		/* next, try to satisfy the allocation request with blocks
840		 * near the hint.
841		 */
842		if ((rc =
843		     dbAllocNear(bmp, dp, blkno, (int) nblocks, l2nb, results))
844		    != -ENOSPC) {
845			if (rc == 0)
846				mark_metapage_dirty(mp);
847
848			release_metapage(mp);
849			goto read_unlock;
850		}
851
852		/* try to satisfy the allocation request with blocks within
853		 * the same dmap as the hint.
854		 */
855		if ((rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results))
856		    != -ENOSPC) {
857			if (rc == 0)
858				mark_metapage_dirty(mp);
859
860			release_metapage(mp);
861			goto read_unlock;
862		}
863
864		release_metapage(mp);
865		IREAD_UNLOCK(ipbmap);
866	}
867
868	/* try to satisfy the allocation request with blocks within
869	 * the same allocation group as the hint.
870	 */
871	IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
872	if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) != -ENOSPC)
873		goto write_unlock;
874
875	IWRITE_UNLOCK(ipbmap);
876
877
878      pref_ag:
879	/*
880	 * Let dbNextAG recommend a preferred allocation group
881	 */
882	agno = dbNextAG(ipbmap);
883	IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
884
885	/* Try to allocate within this allocation group.  if that fails, try to
886	 * allocate anywhere in the map.
887	 */
888	if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) == -ENOSPC)
889		rc = dbAllocAny(bmp, nblocks, l2nb, results);
890
891      write_unlock:
892	IWRITE_UNLOCK(ipbmap);
893
894	return (rc);
895
896      read_unlock:
897	IREAD_UNLOCK(ipbmap);
898
899	return (rc);
900}
901
902#ifdef _NOTYET
903/*
904 * NAME:	dbAllocExact()
905 *
906 * FUNCTION:	try to allocate the requested extent;
907 *
908 * PARAMETERS:
909 *	ip	- pointer to in-core inode;
910 *	blkno	- extent address;
911 *	nblocks	- extent length;
912 *
913 * RETURN VALUES:
914 *	0	- success
915 *	-ENOSPC	- insufficient disk resources
916 *	-EIO	- i/o error
917 */
918int dbAllocExact(struct inode *ip, s64 blkno, int nblocks)
919{
920	int rc;
921	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
922	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
923	struct dmap *dp;
924	s64 lblkno;
925	struct metapage *mp;
926
927	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
928
929	/*
930	 * validate extent request:
931	 *
932	 * note: defragfs policy:
933	 *  max 64 blocks will be moved.
934	 *  allocation request size must be satisfied from a single dmap.
935	 */
936	if (nblocks <= 0 || nblocks > BPERDMAP || blkno >= bmp->db_mapsize) {
937		IREAD_UNLOCK(ipbmap);
938		return -EINVAL;
939	}
940
941	if (nblocks > ((s64) 1 << bmp->db_maxfreebud)) {
942		/* the free space is no longer available */
943		IREAD_UNLOCK(ipbmap);
944		return -ENOSPC;
945	}
946
947	/* read in the dmap covering the extent */
948	lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
949	mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
950	if (mp == NULL) {
951		IREAD_UNLOCK(ipbmap);
952		return -EIO;
953	}
954	dp = (struct dmap *) mp->data;
955
956	/* try to allocate the requested extent */
957	rc = dbAllocNext(bmp, dp, blkno, nblocks);
958
959	IREAD_UNLOCK(ipbmap);
960
961	if (rc == 0)
962		mark_metapage_dirty(mp);
963
964	release_metapage(mp);
965
966	return (rc);
967}
968#endif /* _NOTYET */
969
970/*
971 * NAME:	dbReAlloc()
972 *
973 * FUNCTION:	attempt to extend a current allocation by a specified
974 *		number of blocks.
975 *
976 *		this routine attempts to satisfy the allocation request
977 *		by first trying to extend the existing allocation in
978 *		place by allocating the additional blocks as the blocks
979 *		immediately following the current allocation.  if these
980 *		blocks are not available, this routine will attempt to
981 *		allocate a new set of contiguous blocks large enough
982 *		to cover the existing allocation plus the additional
983 *		number of blocks required.
984 *
985 * PARAMETERS:
986 *	ip	    -  pointer to in-core inode requiring allocation.
987 *	blkno	    -  starting block of the current allocation.
988 *	nblocks	    -  number of contiguous blocks within the current
989 *		       allocation.
990 *	addnblocks  -  number of blocks to add to the allocation.
991 *	results	-      on successful return, set to the starting block number
992 *		       of the existing allocation if the existing allocation
993 *		       was extended in place or to a newly allocated contiguous
994 *		       range if the existing allocation could not be extended
995 *		       in place.
996 *
997 * RETURN VALUES:
998 *	0	- success
999 *	-ENOSPC	- insufficient disk resources
1000 *	-EIO	- i/o error
1001 */
1002int
1003dbReAlloc(struct inode *ip,
1004	  s64 blkno, s64 nblocks, s64 addnblocks, s64 * results)
1005{
1006	int rc;
1007
1008	/* try to extend the allocation in place.
1009	 */
1010	if ((rc = dbExtend(ip, blkno, nblocks, addnblocks)) == 0) {
1011		*results = blkno;
1012		return (0);
1013	} else {
1014		if (rc != -ENOSPC)
1015			return (rc);
1016	}
1017
1018	/* could not extend the allocation in place, so allocate a
1019	 * new set of blocks for the entire request (i.e. try to get
1020	 * a range of contiguous blocks large enough to cover the
1021	 * existing allocation plus the additional blocks.)
1022	 */
1023	return (dbAlloc
1024		(ip, blkno + nblocks - 1, addnblocks + nblocks, results));
1025}
1026
1027
1028/*
1029 * NAME:	dbExtend()
1030 *
1031 * FUNCTION:	attempt to extend a current allocation by a specified
1032 *		number of blocks.
1033 *
1034 *		this routine attempts to satisfy the allocation request
1035 *		by first trying to extend the existing allocation in
1036 *		place by allocating the additional blocks as the blocks
1037 *		immediately following the current allocation.
1038 *
1039 * PARAMETERS:
1040 *	ip	    -  pointer to in-core inode requiring allocation.
1041 *	blkno	    -  starting block of the current allocation.
1042 *	nblocks	    -  number of contiguous blocks within the current
1043 *		       allocation.
1044 *	addnblocks  -  number of blocks to add to the allocation.
1045 *
1046 * RETURN VALUES:
1047 *	0	- success
1048 *	-ENOSPC	- insufficient disk resources
1049 *	-EIO	- i/o error
1050 */
1051static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks)
1052{
1053	struct jfs_sb_info *sbi = JFS_SBI(ip->i_sb);
1054	s64 lblkno, lastblkno, extblkno;
1055	uint rel_block;
1056	struct metapage *mp;
1057	struct dmap *dp;
1058	int rc;
1059	struct inode *ipbmap = sbi->ipbmap;
1060	struct bmap *bmp;
1061
1062	/*
1063	 * We don't want a non-aligned extent to cross a page boundary
1064	 */
1065	if (((rel_block = blkno & (sbi->nbperpage - 1))) &&
1066	    (rel_block + nblocks + addnblocks > sbi->nbperpage))
1067		return -ENOSPC;
1068
1069	/* get the last block of the current allocation */
1070	lastblkno = blkno + nblocks - 1;
1071
1072	/* determine the block number of the block following
1073	 * the existing allocation.
1074	 */
1075	extblkno = lastblkno + 1;
1076
1077	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
1078
1079	/* better be within the file system */
1080	bmp = sbi->bmap;
1081	if (lastblkno < 0 || lastblkno >= bmp->db_mapsize) {
1082		IREAD_UNLOCK(ipbmap);
1083		jfs_error(ip->i_sb, "the block is outside the filesystem\n");
1084		return -EIO;
1085	}
1086
1087	/* we'll attempt to extend the current allocation in place by
1088	 * allocating the additional blocks as the blocks immediately
1089	 * following the current allocation.  we only try to extend the
1090	 * current allocation in place if the number of additional blocks
1091	 * can fit into a dmap, the last block of the current allocation
1092	 * is not the last block of the file system, and the start of the
1093	 * inplace extension is not on an allocation group boundary.
1094	 */
1095	if (addnblocks > BPERDMAP || extblkno >= bmp->db_mapsize ||
1096	    (extblkno & (bmp->db_agsize - 1)) == 0) {
1097		IREAD_UNLOCK(ipbmap);
1098		return -ENOSPC;
1099	}
1100
1101	/* get the buffer for the dmap containing the first block
1102	 * of the extension.
1103	 */
1104	lblkno = BLKTODMAP(extblkno, bmp->db_l2nbperpage);
1105	mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
1106	if (mp == NULL) {
1107		IREAD_UNLOCK(ipbmap);
1108		return -EIO;
1109	}
1110
1111	dp = (struct dmap *) mp->data;
1112
1113	/* try to allocate the blocks immediately following the
1114	 * current allocation.
1115	 */
1116	rc = dbAllocNext(bmp, dp, extblkno, (int) addnblocks);
1117
1118	IREAD_UNLOCK(ipbmap);
1119
1120	/* were we successful ? */
1121	if (rc == 0)
1122		write_metapage(mp);
1123	else
1124		/* we were not successful */
1125		release_metapage(mp);
1126
1127	return (rc);
1128}
1129
1130
1131/*
1132 * NAME:	dbAllocNext()
1133 *
1134 * FUNCTION:	attempt to allocate the blocks of the specified block
1135 *		range within a dmap.
1136 *
1137 * PARAMETERS:
1138 *	bmp	-  pointer to bmap descriptor
1139 *	dp	-  pointer to dmap.
1140 *	blkno	-  starting block number of the range.
1141 *	nblocks	-  number of contiguous free blocks of the range.
1142 *
1143 * RETURN VALUES:
1144 *	0	- success
1145 *	-ENOSPC	- insufficient disk resources
1146 *	-EIO	- i/o error
1147 *
1148 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1149 */
1150static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
1151		       int nblocks)
1152{
1153	int dbitno, word, rembits, nb, nwords, wbitno, nw;
1154	int l2size;
1155	s8 *leaf;
1156	u32 mask;
1157
1158	if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1159		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1160		return -EIO;
1161	}
1162
1163	/* pick up a pointer to the leaves of the dmap tree.
1164	 */
1165	leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1166
1167	/* determine the bit number and word within the dmap of the
1168	 * starting block.
1169	 */
1170	dbitno = blkno & (BPERDMAP - 1);
1171	word = dbitno >> L2DBWORD;
1172
1173	/* check if the specified block range is contained within
1174	 * this dmap.
1175	 */
1176	if (dbitno + nblocks > BPERDMAP)
1177		return -ENOSPC;
1178
1179	/* check if the starting leaf indicates that anything
1180	 * is free.
1181	 */
1182	if (leaf[word] == NOFREE)
1183		return -ENOSPC;
1184
1185	/* check the dmaps words corresponding to block range to see
1186	 * if the block range is free.  not all bits of the first and
1187	 * last words may be contained within the block range.  if this
1188	 * is the case, we'll work against those words (i.e. partial first
1189	 * and/or last) on an individual basis (a single pass) and examine
1190	 * the actual bits to determine if they are free.  a single pass
1191	 * will be used for all dmap words fully contained within the
1192	 * specified range.  within this pass, the leaves of the dmap
1193	 * tree will be examined to determine if the blocks are free. a
1194	 * single leaf may describe the free space of multiple dmap
1195	 * words, so we may visit only a subset of the actual leaves
1196	 * corresponding to the dmap words of the block range.
1197	 */
1198	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
1199		/* determine the bit number within the word and
1200		 * the number of bits within the word.
1201		 */
1202		wbitno = dbitno & (DBWORD - 1);
1203		nb = min(rembits, DBWORD - wbitno);
1204
1205		/* check if only part of the word is to be examined.
1206		 */
1207		if (nb < DBWORD) {
1208			/* check if the bits are free.
1209			 */
1210			mask = (ONES << (DBWORD - nb) >> wbitno);
1211			if ((mask & ~le32_to_cpu(dp->wmap[word])) != mask)
1212				return -ENOSPC;
1213
1214			word += 1;
1215		} else {
1216			/* one or more dmap words are fully contained
1217			 * within the block range.  determine how many
1218			 * words and how many bits.
1219			 */
1220			nwords = rembits >> L2DBWORD;
1221			nb = nwords << L2DBWORD;
1222
1223			/* now examine the appropriate leaves to determine
1224			 * if the blocks are free.
1225			 */
1226			while (nwords > 0) {
1227				/* does the leaf describe any free space ?
1228				 */
1229				if (leaf[word] < BUDMIN)
1230					return -ENOSPC;
1231
1232				/* determine the l2 number of bits provided
1233				 * by this leaf.
1234				 */
1235				l2size =
1236				    min_t(int, leaf[word], NLSTOL2BSZ(nwords));
1237
1238				/* determine how many words were handled.
1239				 */
1240				nw = BUDSIZE(l2size, BUDMIN);
1241
1242				nwords -= nw;
1243				word += nw;
1244			}
1245		}
1246	}
1247
1248	/* allocate the blocks.
1249	 */
1250	return (dbAllocDmap(bmp, dp, blkno, nblocks));
1251}
1252
1253
1254/*
1255 * NAME:	dbAllocNear()
1256 *
1257 * FUNCTION:	attempt to allocate a number of contiguous free blocks near
1258 *		a specified block (hint) within a dmap.
1259 *
1260 *		starting with the dmap leaf that covers the hint, we'll
1261 *		check the next four contiguous leaves for sufficient free
1262 *		space.  if sufficient free space is found, we'll allocate
1263 *		the desired free space.
1264 *
1265 * PARAMETERS:
1266 *	bmp	-  pointer to bmap descriptor
1267 *	dp	-  pointer to dmap.
1268 *	blkno	-  block number to allocate near.
1269 *	nblocks	-  actual number of contiguous free blocks desired.
1270 *	l2nb	-  log2 number of contiguous free blocks desired.
1271 *	results	-  on successful return, set to the starting block number
1272 *		   of the newly allocated range.
1273 *
1274 * RETURN VALUES:
1275 *	0	- success
1276 *	-ENOSPC	- insufficient disk resources
1277 *	-EIO	- i/o error
1278 *
1279 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1280 */
1281static int
1282dbAllocNear(struct bmap * bmp,
1283	    struct dmap * dp, s64 blkno, int nblocks, int l2nb, s64 * results)
1284{
1285	int word, lword, rc;
1286	s8 *leaf;
1287
1288	if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1289		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1290		return -EIO;
1291	}
1292
1293	leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1294
1295	/* determine the word within the dmap that holds the hint
1296	 * (i.e. blkno).  also, determine the last word in the dmap
1297	 * that we'll include in our examination.
1298	 */
1299	word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
1300	lword = min(word + 4, LPERDMAP);
1301
1302	/* examine the leaves for sufficient free space.
1303	 */
1304	for (; word < lword; word++) {
1305		/* does the leaf describe sufficient free space ?
1306		 */
1307		if (leaf[word] < l2nb)
1308			continue;
1309
1310		/* determine the block number within the file system
1311		 * of the first block described by this dmap word.
1312		 */
1313		blkno = le64_to_cpu(dp->start) + (word << L2DBWORD);
1314
1315		/* if not all bits of the dmap word are free, get the
1316		 * starting bit number within the dmap word of the required
1317		 * string of free bits and adjust the block number with the
1318		 * value.
1319		 */
1320		if (leaf[word] < BUDMIN)
1321			blkno +=
1322			    dbFindBits(le32_to_cpu(dp->wmap[word]), l2nb);
1323
1324		/* allocate the blocks.
1325		 */
1326		if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1327			*results = blkno;
1328
1329		return (rc);
1330	}
1331
1332	return -ENOSPC;
1333}
1334
1335
1336/*
1337 * NAME:	dbAllocAG()
1338 *
1339 * FUNCTION:	attempt to allocate the specified number of contiguous
1340 *		free blocks within the specified allocation group.
1341 *
1342 *		unless the allocation group size is equal to the number
1343 *		of blocks per dmap, the dmap control pages will be used to
1344 *		find the required free space, if available.  we start the
1345 *		search at the highest dmap control page level which
1346 *		distinctly describes the allocation group's free space
1347 *		(i.e. the highest level at which the allocation group's
1348 *		free space is not mixed in with that of any other group).
1349 *		in addition, we start the search within this level at a
1350 *		height of the dmapctl dmtree at which the nodes distinctly
1351 *		describe the allocation group's free space.  at this height,
1352 *		the allocation group's free space may be represented by 1
1353 *		or two sub-trees, depending on the allocation group size.
1354 *		we search the top nodes of these subtrees left to right for
1355 *		sufficient free space.  if sufficient free space is found,
1356 *		the subtree is searched to find the leftmost leaf that
1357 *		has free space.  once we have made it to the leaf, we
1358 *		move the search to the next lower level dmap control page
1359 *		corresponding to this leaf.  we continue down the dmap control
1360 *		pages until we find the dmap that contains or starts the
1361 *		sufficient free space and we allocate at this dmap.
1362 *
1363 *		if the allocation group size is equal to the dmap size,
1364 *		we'll start at the dmap corresponding to the allocation
1365 *		group and attempt the allocation at this level.
1366 *
1367 *		the dmap control page search is also not performed if the
1368 *		allocation group is completely free and we go to the first
1369 *		dmap of the allocation group to do the allocation.  this is
1370 *		done because the allocation group may be part (not the first
1371 *		part) of a larger binary buddy system, causing the dmap
1372 *		control pages to indicate no free space (NOFREE) within
1373 *		the allocation group.
1374 *
1375 * PARAMETERS:
1376 *	bmp	-  pointer to bmap descriptor
1377 *	agno	- allocation group number.
1378 *	nblocks	-  actual number of contiguous free blocks desired.
1379 *	l2nb	-  log2 number of contiguous free blocks desired.
1380 *	results	-  on successful return, set to the starting block number
1381 *		   of the newly allocated range.
1382 *
1383 * RETURN VALUES:
1384 *	0	- success
1385 *	-ENOSPC	- insufficient disk resources
1386 *	-EIO	- i/o error
1387 *
1388 * note: IWRITE_LOCK(ipmap) held on entry/exit;
1389 */
1390static int
1391dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, s64 * results)
1392{
1393	struct metapage *mp;
1394	struct dmapctl *dcp;
1395	int rc, ti, i, k, m, n, agperlev;
1396	s64 blkno, lblkno;
1397	int budmin;
1398
1399	/* allocation request should not be for more than the
1400	 * allocation group size.
1401	 */
1402	if (l2nb > bmp->db_agl2size) {
1403		jfs_error(bmp->db_ipbmap->i_sb,
1404			  "allocation request is larger than the allocation group size\n");
1405		return -EIO;
1406	}
1407
1408	/* determine the starting block number of the allocation
1409	 * group.
1410	 */
1411	blkno = (s64) agno << bmp->db_agl2size;
1412
1413	/* check if the allocation group size is the minimum allocation
1414	 * group size or if the allocation group is completely free. if
1415	 * the allocation group size is the minimum size of BPERDMAP (i.e.
1416	 * 1 dmap), there is no need to search the dmap control page (below)
1417	 * that fully describes the allocation group since the allocation
1418	 * group is already fully described by a dmap.  in this case, we
1419	 * just call dbAllocCtl() to search the dmap tree and allocate the
1420	 * required space if available.
1421	 *
1422	 * if the allocation group is completely free, dbAllocCtl() is
1423	 * also called to allocate the required space.  this is done for
1424	 * two reasons.  first, it makes no sense searching the dmap control
1425	 * pages for free space when we know that free space exists.  second,
1426	 * the dmap control pages may indicate that the allocation group
1427	 * has no free space if the allocation group is part (not the first
1428	 * part) of a larger binary buddy system.
1429	 */
1430	if (bmp->db_agsize == BPERDMAP
1431	    || bmp->db_agfree[agno] == bmp->db_agsize) {
1432		rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1433		if ((rc == -ENOSPC) &&
1434		    (bmp->db_agfree[agno] == bmp->db_agsize)) {
1435			printk(KERN_ERR "blkno = %Lx, blocks = %Lx\n",
1436			       (unsigned long long) blkno,
1437			       (unsigned long long) nblocks);
1438			jfs_error(bmp->db_ipbmap->i_sb,
1439				  "dbAllocCtl failed in free AG\n");
1440		}
1441		return (rc);
1442	}
1443
1444	/* the buffer for the dmap control page that fully describes the
1445	 * allocation group.
1446	 */
1447	lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, bmp->db_aglevel);
1448	mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1449	if (mp == NULL)
1450		return -EIO;
1451	dcp = (struct dmapctl *) mp->data;
1452	budmin = dcp->budmin;
1453
1454	if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1455		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
1456		release_metapage(mp);
1457		return -EIO;
1458	}
1459
1460	/* search the subtree(s) of the dmap control page that describes
1461	 * the allocation group, looking for sufficient free space.  to begin,
1462	 * determine how many allocation groups are represented in a dmap
1463	 * control page at the control page level (i.e. L0, L1, L2) that
1464	 * fully describes an allocation group. next, determine the starting
1465	 * tree index of this allocation group within the control page.
1466	 */
1467	agperlev =
1468	    (1 << (L2LPERCTL - (bmp->db_agheight << 1))) / bmp->db_agwidth;
1469	ti = bmp->db_agstart + bmp->db_agwidth * (agno & (agperlev - 1));
1470
1471	/* dmap control page trees fan-out by 4 and a single allocation
1472	 * group may be described by 1 or 2 subtrees within the ag level
1473	 * dmap control page, depending upon the ag size. examine the ag's
1474	 * subtrees for sufficient free space, starting with the leftmost
1475	 * subtree.
1476	 */
1477	for (i = 0; i < bmp->db_agwidth; i++, ti++) {
1478		/* is there sufficient free space ?
1479		 */
1480		if (l2nb > dcp->stree[ti])
1481			continue;
1482
1483		/* sufficient free space found in a subtree. now search down
1484		 * the subtree to find the leftmost leaf that describes this
1485		 * free space.
1486		 */
1487		for (k = bmp->db_agheight; k > 0; k--) {
1488			for (n = 0, m = (ti << 2) + 1; n < 4; n++) {
1489				if (l2nb <= dcp->stree[m + n]) {
1490					ti = m + n;
1491					break;
1492				}
1493			}
1494			if (n == 4) {
1495				jfs_error(bmp->db_ipbmap->i_sb,
1496					  "failed descending stree\n");
1497				release_metapage(mp);
1498				return -EIO;
1499			}
1500		}
1501
1502		/* determine the block number within the file system
1503		 * that corresponds to this leaf.
1504		 */
1505		if (bmp->db_aglevel == 2)
1506			blkno = 0;
1507		else if (bmp->db_aglevel == 1)
1508			blkno &= ~(MAXL1SIZE - 1);
1509		else		/* bmp->db_aglevel == 0 */
1510			blkno &= ~(MAXL0SIZE - 1);
1511
1512		blkno +=
1513		    ((s64) (ti - le32_to_cpu(dcp->leafidx))) << budmin;
1514
1515		/* release the buffer in preparation for going down
1516		 * the next level of dmap control pages.
1517		 */
1518		release_metapage(mp);
1519
1520		/* check if we need to continue to search down the lower
1521		 * level dmap control pages.  we need to if the number of
1522		 * blocks required is less than maximum number of blocks
1523		 * described at the next lower level.
1524		 */
1525		if (l2nb < budmin) {
1526
1527			/* search the lower level dmap control pages to get
1528			 * the starting block number of the dmap that
1529			 * contains or starts off the free space.
1530			 */
1531			if ((rc =
1532			     dbFindCtl(bmp, l2nb, bmp->db_aglevel - 1,
1533				       &blkno))) {
1534				if (rc == -ENOSPC) {
1535					jfs_error(bmp->db_ipbmap->i_sb,
1536						  "control page inconsistent\n");
1537					return -EIO;
1538				}
1539				return (rc);
1540			}
1541		}
1542
1543		/* allocate the blocks.
1544		 */
1545		rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1546		if (rc == -ENOSPC) {
1547			jfs_error(bmp->db_ipbmap->i_sb,
1548				  "unable to allocate blocks\n");
1549			rc = -EIO;
1550		}
1551		return (rc);
1552	}
1553
1554	/* no space in the allocation group.  release the buffer and
1555	 * return -ENOSPC.
1556	 */
1557	release_metapage(mp);
1558
1559	return -ENOSPC;
1560}
1561
1562
1563/*
1564 * NAME:	dbAllocAny()
1565 *
1566 * FUNCTION:	attempt to allocate the specified number of contiguous
1567 *		free blocks anywhere in the file system.
1568 *
1569 *		dbAllocAny() attempts to find the sufficient free space by
1570 *		searching down the dmap control pages, starting with the
1571 *		highest level (i.e. L0, L1, L2) control page.  if free space
1572 *		large enough to satisfy the desired free space is found, the
1573 *		desired free space is allocated.
1574 *
1575 * PARAMETERS:
1576 *	bmp	-  pointer to bmap descriptor
1577 *	nblocks	 -  actual number of contiguous free blocks desired.
1578 *	l2nb	 -  log2 number of contiguous free blocks desired.
1579 *	results	-  on successful return, set to the starting block number
1580 *		   of the newly allocated range.
1581 *
1582 * RETURN VALUES:
1583 *	0	- success
1584 *	-ENOSPC	- insufficient disk resources
1585 *	-EIO	- i/o error
1586 *
1587 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1588 */
1589static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results)
1590{
1591	int rc;
1592	s64 blkno = 0;
1593
1594	/* starting with the top level dmap control page, search
1595	 * down the dmap control levels for sufficient free space.
1596	 * if free space is found, dbFindCtl() returns the starting
1597	 * block number of the dmap that contains or starts off the
1598	 * range of free space.
1599	 */
1600	if ((rc = dbFindCtl(bmp, l2nb, bmp->db_maxlevel, &blkno)))
1601		return (rc);
1602
1603	/* allocate the blocks.
1604	 */
1605	rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1606	if (rc == -ENOSPC) {
1607		jfs_error(bmp->db_ipbmap->i_sb, "unable to allocate blocks\n");
1608		return -EIO;
1609	}
1610	return (rc);
1611}
1612
1613
1614/*
1615 * NAME:	dbDiscardAG()
1616 *
1617 * FUNCTION:	attempt to discard (TRIM) all free blocks of specific AG
1618 *
1619 *		algorithm:
1620 *		1) allocate blocks, as large as possible and save them
1621 *		   while holding IWRITE_LOCK on ipbmap
1622 *		2) trim all these saved block/length values
1623 *		3) mark the blocks free again
1624 *
1625 *		benefit:
1626 *		- we work only on one ag at some time, minimizing how long we
1627 *		  need to lock ipbmap
1628 *		- reading / writing the fs is possible most time, even on
1629 *		  trimming
1630 *
1631 *		downside:
1632 *		- we write two times to the dmapctl and dmap pages
1633 *		- but for me, this seems the best way, better ideas?
1634 *		/TR 2012
1635 *
1636 * PARAMETERS:
1637 *	ip	- pointer to in-core inode
1638 *	agno	- ag to trim
1639 *	minlen	- minimum value of contiguous blocks
1640 *
1641 * RETURN VALUES:
1642 *	s64	- actual number of blocks trimmed
1643 */
1644s64 dbDiscardAG(struct inode *ip, int agno, s64 minlen)
1645{
1646	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
1647	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
1648	s64 nblocks, blkno;
1649	u64 trimmed = 0;
1650	int rc, l2nb;
1651	struct super_block *sb = ipbmap->i_sb;
1652
1653	struct range2trim {
1654		u64 blkno;
1655		u64 nblocks;
1656	} *totrim, *tt;
1657
1658	/* max blkno / nblocks pairs to trim */
1659	int count = 0, range_cnt;
1660	u64 max_ranges;
1661
1662	/* prevent others from writing new stuff here, while trimming */
1663	IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
1664
1665	nblocks = bmp->db_agfree[agno];
1666	max_ranges = nblocks;
1667	do_div(max_ranges, minlen);
1668	range_cnt = min_t(u64, max_ranges + 1, 32 * 1024);
1669	totrim = kmalloc_array(range_cnt, sizeof(struct range2trim), GFP_NOFS);
1670	if (totrim == NULL) {
1671		jfs_error(bmp->db_ipbmap->i_sb, "no memory for trim array\n");
1672		IWRITE_UNLOCK(ipbmap);
1673		return 0;
1674	}
1675
1676	tt = totrim;
1677	while (nblocks >= minlen) {
1678		l2nb = BLKSTOL2(nblocks);
1679
1680		/* 0 = okay, -EIO = fatal, -ENOSPC -> try smaller block */
1681		rc = dbAllocAG(bmp, agno, nblocks, l2nb, &blkno);
1682		if (rc == 0) {
1683			tt->blkno = blkno;
1684			tt->nblocks = nblocks;
1685			tt++; count++;
1686
1687			/* the whole ag is free, trim now */
1688			if (bmp->db_agfree[agno] == 0)
1689				break;
1690
1691			/* give a hint for the next while */
1692			nblocks = bmp->db_agfree[agno];
1693			continue;
1694		} else if (rc == -ENOSPC) {
1695			/* search for next smaller log2 block */
1696			l2nb = BLKSTOL2(nblocks) - 1;
1697			nblocks = 1LL << l2nb;
1698		} else {
1699			/* Trim any already allocated blocks */
1700			jfs_error(bmp->db_ipbmap->i_sb, "-EIO\n");
1701			break;
1702		}
1703
1704		/* check, if our trim array is full */
1705		if (unlikely(count >= range_cnt - 1))
1706			break;
1707	}
1708	IWRITE_UNLOCK(ipbmap);
1709
1710	tt->nblocks = 0; /* mark the current end */
1711	for (tt = totrim; tt->nblocks != 0; tt++) {
1712		/* when mounted with online discard, dbFree() will
1713		 * call jfs_issue_discard() itself */
1714		if (!(JFS_SBI(sb)->flag & JFS_DISCARD))
1715			jfs_issue_discard(ip, tt->blkno, tt->nblocks);
1716		dbFree(ip, tt->blkno, tt->nblocks);
1717		trimmed += tt->nblocks;
1718	}
1719	kfree(totrim);
1720
1721	return trimmed;
1722}
1723
1724/*
1725 * NAME:	dbFindCtl()
1726 *
1727 * FUNCTION:	starting at a specified dmap control page level and block
1728 *		number, search down the dmap control levels for a range of
1729 *		contiguous free blocks large enough to satisfy an allocation
1730 *		request for the specified number of free blocks.
1731 *
1732 *		if sufficient contiguous free blocks are found, this routine
1733 *		returns the starting block number within a dmap page that
1734 *		contains or starts a range of contiqious free blocks that
1735 *		is sufficient in size.
1736 *
1737 * PARAMETERS:
1738 *	bmp	-  pointer to bmap descriptor
1739 *	level	-  starting dmap control page level.
1740 *	l2nb	-  log2 number of contiguous free blocks desired.
1741 *	*blkno	-  on entry, starting block number for conducting the search.
1742 *		   on successful return, the first block within a dmap page
1743 *		   that contains or starts a range of contiguous free blocks.
1744 *
1745 * RETURN VALUES:
1746 *	0	- success
1747 *	-ENOSPC	- insufficient disk resources
1748 *	-EIO	- i/o error
1749 *
1750 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1751 */
1752static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno)
1753{
1754	int rc, leafidx, lev;
1755	s64 b, lblkno;
1756	struct dmapctl *dcp;
1757	int budmin;
1758	struct metapage *mp;
1759
1760	/* starting at the specified dmap control page level and block
1761	 * number, search down the dmap control levels for the starting
1762	 * block number of a dmap page that contains or starts off
1763	 * sufficient free blocks.
1764	 */
1765	for (lev = level, b = *blkno; lev >= 0; lev--) {
1766		/* get the buffer of the dmap control page for the block
1767		 * number and level (i.e. L0, L1, L2).
1768		 */
1769		lblkno = BLKTOCTL(b, bmp->db_l2nbperpage, lev);
1770		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1771		if (mp == NULL)
1772			return -EIO;
1773		dcp = (struct dmapctl *) mp->data;
1774		budmin = dcp->budmin;
1775
1776		if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1777			jfs_error(bmp->db_ipbmap->i_sb,
1778				  "Corrupt dmapctl page\n");
1779			release_metapage(mp);
1780			return -EIO;
1781		}
1782
1783		/* search the tree within the dmap control page for
1784		 * sufficient free space.  if sufficient free space is found,
1785		 * dbFindLeaf() returns the index of the leaf at which
1786		 * free space was found.
1787		 */
1788		rc = dbFindLeaf((dmtree_t *) dcp, l2nb, &leafidx, true);
1789
1790		/* release the buffer.
1791		 */
1792		release_metapage(mp);
1793
1794		/* space found ?
1795		 */
1796		if (rc) {
1797			if (lev != level) {
1798				jfs_error(bmp->db_ipbmap->i_sb,
1799					  "dmap inconsistent\n");
1800				return -EIO;
1801			}
1802			return -ENOSPC;
1803		}
1804
1805		/* adjust the block number to reflect the location within
1806		 * the dmap control page (i.e. the leaf) at which free
1807		 * space was found.
1808		 */
1809		b += (((s64) leafidx) << budmin);
1810
1811		/* we stop the search at this dmap control page level if
1812		 * the number of blocks required is greater than or equal
1813		 * to the maximum number of blocks described at the next
1814		 * (lower) level.
1815		 */
1816		if (l2nb >= budmin)
1817			break;
1818	}
1819
1820	*blkno = b;
1821	return (0);
1822}
1823
1824
1825/*
1826 * NAME:	dbAllocCtl()
1827 *
1828 * FUNCTION:	attempt to allocate a specified number of contiguous
1829 *		blocks starting within a specific dmap.
1830 *
1831 *		this routine is called by higher level routines that search
1832 *		the dmap control pages above the actual dmaps for contiguous
1833 *		free space.  the result of successful searches by these
1834 *		routines are the starting block numbers within dmaps, with
1835 *		the dmaps themselves containing the desired contiguous free
1836 *		space or starting a contiguous free space of desired size
1837 *		that is made up of the blocks of one or more dmaps. these
1838 *		calls should not fail due to insufficent resources.
1839 *
1840 *		this routine is called in some cases where it is not known
1841 *		whether it will fail due to insufficient resources.  more
1842 *		specifically, this occurs when allocating from an allocation
1843 *		group whose size is equal to the number of blocks per dmap.
1844 *		in this case, the dmap control pages are not examined prior
1845 *		to calling this routine (to save pathlength) and the call
1846 *		might fail.
1847 *
1848 *		for a request size that fits within a dmap, this routine relies
1849 *		upon the dmap's dmtree to find the requested contiguous free
1850 *		space.  for request sizes that are larger than a dmap, the
1851 *		requested free space will start at the first block of the
1852 *		first dmap (i.e. blkno).
1853 *
1854 * PARAMETERS:
1855 *	bmp	-  pointer to bmap descriptor
1856 *	nblocks	 -  actual number of contiguous free blocks to allocate.
1857 *	l2nb	 -  log2 number of contiguous free blocks to allocate.
1858 *	blkno	 -  starting block number of the dmap to start the allocation
1859 *		    from.
1860 *	results	-  on successful return, set to the starting block number
1861 *		   of the newly allocated range.
1862 *
1863 * RETURN VALUES:
1864 *	0	- success
1865 *	-ENOSPC	- insufficient disk resources
1866 *	-EIO	- i/o error
1867 *
1868 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1869 */
1870static int
1871dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, s64 * results)
1872{
1873	int rc, nb;
1874	s64 b, lblkno, n;
1875	struct metapage *mp;
1876	struct dmap *dp;
1877
1878	/* check if the allocation request is confined to a single dmap.
1879	 */
1880	if (l2nb <= L2BPERDMAP) {
1881		/* get the buffer for the dmap.
1882		 */
1883		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
1884		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1885		if (mp == NULL)
1886			return -EIO;
1887		dp = (struct dmap *) mp->data;
1888
1889		/* try to allocate the blocks.
1890		 */
1891		rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results);
1892		if (rc == 0)
1893			mark_metapage_dirty(mp);
1894
1895		release_metapage(mp);
1896
1897		return (rc);
1898	}
1899
1900	/* allocation request involving multiple dmaps. it must start on
1901	 * a dmap boundary.
1902	 */
1903	assert((blkno & (BPERDMAP - 1)) == 0);
1904
1905	/* allocate the blocks dmap by dmap.
1906	 */
1907	for (n = nblocks, b = blkno; n > 0; n -= nb, b += nb) {
1908		/* get the buffer for the dmap.
1909		 */
1910		lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1911		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1912		if (mp == NULL) {
1913			rc = -EIO;
1914			goto backout;
1915		}
1916		dp = (struct dmap *) mp->data;
1917
1918		/* the dmap better be all free.
1919		 */
1920		if (dp->tree.stree[ROOT] != L2BPERDMAP) {
1921			release_metapage(mp);
1922			jfs_error(bmp->db_ipbmap->i_sb,
1923				  "the dmap is not all free\n");
1924			rc = -EIO;
1925			goto backout;
1926		}
1927
1928		/* determine how many blocks to allocate from this dmap.
1929		 */
1930		nb = min_t(s64, n, BPERDMAP);
1931
1932		/* allocate the blocks from the dmap.
1933		 */
1934		if ((rc = dbAllocDmap(bmp, dp, b, nb))) {
1935			release_metapage(mp);
1936			goto backout;
1937		}
1938
1939		/* write the buffer.
1940		 */
1941		write_metapage(mp);
1942	}
1943
1944	/* set the results (starting block number) and return.
1945	 */
1946	*results = blkno;
1947	return (0);
1948
1949	/* something failed in handling an allocation request involving
1950	 * multiple dmaps.  we'll try to clean up by backing out any
1951	 * allocation that has already happened for this request.  if
1952	 * we fail in backing out the allocation, we'll mark the file
1953	 * system to indicate that blocks have been leaked.
1954	 */
1955      backout:
1956
1957	/* try to backout the allocations dmap by dmap.
1958	 */
1959	for (n = nblocks - n, b = blkno; n > 0;
1960	     n -= BPERDMAP, b += BPERDMAP) {
1961		/* get the buffer for this dmap.
1962		 */
1963		lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1964		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1965		if (mp == NULL) {
1966			/* could not back out.  mark the file system
1967			 * to indicate that we have leaked blocks.
1968			 */
1969			jfs_error(bmp->db_ipbmap->i_sb,
1970				  "I/O Error: Block Leakage\n");
1971			continue;
1972		}
1973		dp = (struct dmap *) mp->data;
1974
1975		/* free the blocks is this dmap.
1976		 */
1977		if (dbFreeDmap(bmp, dp, b, BPERDMAP)) {
1978			/* could not back out.  mark the file system
1979			 * to indicate that we have leaked blocks.
1980			 */
1981			release_metapage(mp);
1982			jfs_error(bmp->db_ipbmap->i_sb, "Block Leakage\n");
1983			continue;
1984		}
1985
1986		/* write the buffer.
1987		 */
1988		write_metapage(mp);
1989	}
1990
1991	return (rc);
1992}
1993
1994
1995/*
1996 * NAME:	dbAllocDmapLev()
1997 *
1998 * FUNCTION:	attempt to allocate a specified number of contiguous blocks
1999 *		from a specified dmap.
2000 *
2001 *		this routine checks if the contiguous blocks are available.
2002 *		if so, nblocks of blocks are allocated; otherwise, ENOSPC is
2003 *		returned.
2004 *
2005 * PARAMETERS:
2006 *	mp	-  pointer to bmap descriptor
2007 *	dp	-  pointer to dmap to attempt to allocate blocks from.
2008 *	l2nb	-  log2 number of contiguous block desired.
2009 *	nblocks	-  actual number of contiguous block desired.
2010 *	results	-  on successful return, set to the starting block number
2011 *		   of the newly allocated range.
2012 *
2013 * RETURN VALUES:
2014 *	0	- success
2015 *	-ENOSPC	- insufficient disk resources
2016 *	-EIO	- i/o error
2017 *
2018 * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or
2019 *	IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit;
2020 */
2021static int
2022dbAllocDmapLev(struct bmap * bmp,
2023	       struct dmap * dp, int nblocks, int l2nb, s64 * results)
2024{
2025	s64 blkno;
2026	int leafidx, rc;
2027
2028	/* can't be more than a dmaps worth of blocks */
2029	assert(l2nb <= L2BPERDMAP);
2030
2031	/* search the tree within the dmap page for sufficient
2032	 * free space.  if sufficient free space is found, dbFindLeaf()
2033	 * returns the index of the leaf at which free space was found.
2034	 */
2035	if (dbFindLeaf((dmtree_t *) &dp->tree, l2nb, &leafidx, false))
2036		return -ENOSPC;
2037
2038	if (leafidx < 0)
2039		return -EIO;
2040
2041	/* determine the block number within the file system corresponding
2042	 * to the leaf at which free space was found.
2043	 */
2044	blkno = le64_to_cpu(dp->start) + (leafidx << L2DBWORD);
2045
2046	/* if not all bits of the dmap word are free, get the starting
2047	 * bit number within the dmap word of the required string of free
2048	 * bits and adjust the block number with this value.
2049	 */
2050	if (dp->tree.stree[leafidx + LEAFIND] < BUDMIN)
2051		blkno += dbFindBits(le32_to_cpu(dp->wmap[leafidx]), l2nb);
2052
2053	/* allocate the blocks */
2054	if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
2055		*results = blkno;
2056
2057	return (rc);
2058}
2059
2060
2061/*
2062 * NAME:	dbAllocDmap()
2063 *
2064 * FUNCTION:	adjust the disk allocation map to reflect the allocation
2065 *		of a specified block range within a dmap.
2066 *
2067 *		this routine allocates the specified blocks from the dmap
2068 *		through a call to dbAllocBits(). if the allocation of the
2069 *		block range causes the maximum string of free blocks within
2070 *		the dmap to change (i.e. the value of the root of the dmap's
2071 *		dmtree), this routine will cause this change to be reflected
2072 *		up through the appropriate levels of the dmap control pages
2073 *		by a call to dbAdjCtl() for the L0 dmap control page that
2074 *		covers this dmap.
2075 *
2076 * PARAMETERS:
2077 *	bmp	-  pointer to bmap descriptor
2078 *	dp	-  pointer to dmap to allocate the block range from.
2079 *	blkno	-  starting block number of the block to be allocated.
2080 *	nblocks	-  number of blocks to be allocated.
2081 *
2082 * RETURN VALUES:
2083 *	0	- success
2084 *	-EIO	- i/o error
2085 *
2086 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2087 */
2088static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2089		       int nblocks)
2090{
2091	s8 oldroot;
2092	int rc;
2093
2094	/* save the current value of the root (i.e. maximum free string)
2095	 * of the dmap tree.
2096	 */
2097	oldroot = dp->tree.stree[ROOT];
2098
2099	/* allocate the specified (blocks) bits */
2100	dbAllocBits(bmp, dp, blkno, nblocks);
2101
2102	/* if the root has not changed, done. */
2103	if (dp->tree.stree[ROOT] == oldroot)
2104		return (0);
2105
2106	/* root changed. bubble the change up to the dmap control pages.
2107	 * if the adjustment of the upper level control pages fails,
2108	 * backout the bit allocation (thus making everything consistent).
2109	 */
2110	if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 1, 0)))
2111		dbFreeBits(bmp, dp, blkno, nblocks);
2112
2113	return (rc);
2114}
2115
2116
2117/*
2118 * NAME:	dbFreeDmap()
2119 *
2120 * FUNCTION:	adjust the disk allocation map to reflect the allocation
2121 *		of a specified block range within a dmap.
2122 *
2123 *		this routine frees the specified blocks from the dmap through
2124 *		a call to dbFreeBits(). if the deallocation of the block range
2125 *		causes the maximum string of free blocks within the dmap to
2126 *		change (i.e. the value of the root of the dmap's dmtree), this
2127 *		routine will cause this change to be reflected up through the
2128 *		appropriate levels of the dmap control pages by a call to
2129 *		dbAdjCtl() for the L0 dmap control page that covers this dmap.
2130 *
2131 * PARAMETERS:
2132 *	bmp	-  pointer to bmap descriptor
2133 *	dp	-  pointer to dmap to free the block range from.
2134 *	blkno	-  starting block number of the block to be freed.
2135 *	nblocks	-  number of blocks to be freed.
2136 *
2137 * RETURN VALUES:
2138 *	0	- success
2139 *	-EIO	- i/o error
2140 *
2141 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2142 */
2143static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2144		      int nblocks)
2145{
2146	s8 oldroot;
2147	int rc = 0, word;
2148
2149	/* save the current value of the root (i.e. maximum free string)
2150	 * of the dmap tree.
2151	 */
2152	oldroot = dp->tree.stree[ROOT];
2153
2154	/* free the specified (blocks) bits */
2155	rc = dbFreeBits(bmp, dp, blkno, nblocks);
2156
2157	/* if error or the root has not changed, done. */
2158	if (rc || (dp->tree.stree[ROOT] == oldroot))
2159		return (rc);
2160
2161	/* root changed. bubble the change up to the dmap control pages.
2162	 * if the adjustment of the upper level control pages fails,
2163	 * backout the deallocation.
2164	 */
2165	if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 0, 0))) {
2166		word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
2167
2168		/* as part of backing out the deallocation, we will have
2169		 * to back split the dmap tree if the deallocation caused
2170		 * the freed blocks to become part of a larger binary buddy
2171		 * system.
2172		 */
2173		if (dp->tree.stree[word] == NOFREE)
2174			dbBackSplit((dmtree_t *)&dp->tree, word, false);
2175
2176		dbAllocBits(bmp, dp, blkno, nblocks);
2177	}
2178
2179	return (rc);
2180}
2181
2182
2183/*
2184 * NAME:	dbAllocBits()
2185 *
2186 * FUNCTION:	allocate a specified block range from a dmap.
2187 *
2188 *		this routine updates the dmap to reflect the working
2189 *		state allocation of the specified block range. it directly
2190 *		updates the bits of the working map and causes the adjustment
2191 *		of the binary buddy system described by the dmap's dmtree
2192 *		leaves to reflect the bits allocated.  it also causes the
2193 *		dmap's dmtree, as a whole, to reflect the allocated range.
2194 *
2195 * PARAMETERS:
2196 *	bmp	-  pointer to bmap descriptor
2197 *	dp	-  pointer to dmap to allocate bits from.
2198 *	blkno	-  starting block number of the bits to be allocated.
2199 *	nblocks	-  number of bits to be allocated.
2200 *
2201 * RETURN VALUES: none
2202 *
2203 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2204 */
2205static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2206			int nblocks)
2207{
2208	int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2209	dmtree_t *tp = (dmtree_t *) & dp->tree;
2210	int size;
2211	s8 *leaf;
2212
2213	/* pick up a pointer to the leaves of the dmap tree */
2214	leaf = dp->tree.stree + LEAFIND;
2215
2216	/* determine the bit number and word within the dmap of the
2217	 * starting block.
2218	 */
2219	dbitno = blkno & (BPERDMAP - 1);
2220	word = dbitno >> L2DBWORD;
2221
2222	/* block range better be within the dmap */
2223	assert(dbitno + nblocks <= BPERDMAP);
2224
2225	/* allocate the bits of the dmap's words corresponding to the block
2226	 * range. not all bits of the first and last words may be contained
2227	 * within the block range.  if this is the case, we'll work against
2228	 * those words (i.e. partial first and/or last) on an individual basis
2229	 * (a single pass), allocating the bits of interest by hand and
2230	 * updating the leaf corresponding to the dmap word. a single pass
2231	 * will be used for all dmap words fully contained within the
2232	 * specified range.  within this pass, the bits of all fully contained
2233	 * dmap words will be marked as free in a single shot and the leaves
2234	 * will be updated. a single leaf may describe the free space of
2235	 * multiple dmap words, so we may update only a subset of the actual
2236	 * leaves corresponding to the dmap words of the block range.
2237	 */
2238	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2239		/* determine the bit number within the word and
2240		 * the number of bits within the word.
2241		 */
2242		wbitno = dbitno & (DBWORD - 1);
2243		nb = min(rembits, DBWORD - wbitno);
2244
2245		/* check if only part of a word is to be allocated.
2246		 */
2247		if (nb < DBWORD) {
2248			/* allocate (set to 1) the appropriate bits within
2249			 * this dmap word.
2250			 */
2251			dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
2252						      >> wbitno);
2253
2254			/* update the leaf for this dmap word. in addition
2255			 * to setting the leaf value to the binary buddy max
2256			 * of the updated dmap word, dbSplit() will split
2257			 * the binary system of the leaves if need be.
2258			 */
2259			dbSplit(tp, word, BUDMIN,
2260				dbMaxBud((u8 *)&dp->wmap[word]), false);
2261
2262			word += 1;
2263		} else {
2264			/* one or more dmap words are fully contained
2265			 * within the block range.  determine how many
2266			 * words and allocate (set to 1) the bits of these
2267			 * words.
2268			 */
2269			nwords = rembits >> L2DBWORD;
2270			memset(&dp->wmap[word], (int) ONES, nwords * 4);
2271
2272			/* determine how many bits.
2273			 */
2274			nb = nwords << L2DBWORD;
2275
2276			/* now update the appropriate leaves to reflect
2277			 * the allocated words.
2278			 */
2279			for (; nwords > 0; nwords -= nw) {
2280				if (leaf[word] < BUDMIN) {
2281					jfs_error(bmp->db_ipbmap->i_sb,
2282						  "leaf page corrupt\n");
2283					break;
2284				}
2285
2286				/* determine what the leaf value should be
2287				 * updated to as the minimum of the l2 number
2288				 * of bits being allocated and the l2 number
2289				 * of bits currently described by this leaf.
2290				 */
2291				size = min_t(int, leaf[word],
2292					     NLSTOL2BSZ(nwords));
2293
2294				/* update the leaf to reflect the allocation.
2295				 * in addition to setting the leaf value to
2296				 * NOFREE, dbSplit() will split the binary
2297				 * system of the leaves to reflect the current
2298				 * allocation (size).
2299				 */
2300				dbSplit(tp, word, size, NOFREE, false);
2301
2302				/* get the number of dmap words handled */
2303				nw = BUDSIZE(size, BUDMIN);
2304				word += nw;
2305			}
2306		}
2307	}
2308
2309	/* update the free count for this dmap */
2310	le32_add_cpu(&dp->nfree, -nblocks);
2311
2312	BMAP_LOCK(bmp);
2313
2314	/* if this allocation group is completely free,
2315	 * update the maximum allocation group number if this allocation
2316	 * group is the new max.
2317	 */
2318	agno = blkno >> bmp->db_agl2size;
2319	if (agno > bmp->db_maxag)
2320		bmp->db_maxag = agno;
2321
2322	/* update the free count for the allocation group and map */
2323	bmp->db_agfree[agno] -= nblocks;
2324	bmp->db_nfree -= nblocks;
2325
2326	BMAP_UNLOCK(bmp);
2327}
2328
2329
2330/*
2331 * NAME:	dbFreeBits()
2332 *
2333 * FUNCTION:	free a specified block range from a dmap.
2334 *
2335 *		this routine updates the dmap to reflect the working
2336 *		state allocation of the specified block range. it directly
2337 *		updates the bits of the working map and causes the adjustment
2338 *		of the binary buddy system described by the dmap's dmtree
2339 *		leaves to reflect the bits freed.  it also causes the dmap's
2340 *		dmtree, as a whole, to reflect the deallocated range.
2341 *
2342 * PARAMETERS:
2343 *	bmp	-  pointer to bmap descriptor
2344 *	dp	-  pointer to dmap to free bits from.
2345 *	blkno	-  starting block number of the bits to be freed.
2346 *	nblocks	-  number of bits to be freed.
2347 *
2348 * RETURN VALUES: 0 for success
2349 *
2350 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2351 */
2352static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2353		       int nblocks)
2354{
2355	int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2356	dmtree_t *tp = (dmtree_t *) & dp->tree;
2357	int rc = 0;
2358	int size;
2359
2360	/* determine the bit number and word within the dmap of the
2361	 * starting block.
2362	 */
2363	dbitno = blkno & (BPERDMAP - 1);
2364	word = dbitno >> L2DBWORD;
2365
2366	/* block range better be within the dmap.
2367	 */
2368	assert(dbitno + nblocks <= BPERDMAP);
2369
2370	/* free the bits of the dmaps words corresponding to the block range.
2371	 * not all bits of the first and last words may be contained within
2372	 * the block range.  if this is the case, we'll work against those
2373	 * words (i.e. partial first and/or last) on an individual basis
2374	 * (a single pass), freeing the bits of interest by hand and updating
2375	 * the leaf corresponding to the dmap word. a single pass will be used
2376	 * for all dmap words fully contained within the specified range.
2377	 * within this pass, the bits of all fully contained dmap words will
2378	 * be marked as free in a single shot and the leaves will be updated. a
2379	 * single leaf may describe the free space of multiple dmap words,
2380	 * so we may update only a subset of the actual leaves corresponding
2381	 * to the dmap words of the block range.
2382	 *
2383	 * dbJoin() is used to update leaf values and will join the binary
2384	 * buddy system of the leaves if the new leaf values indicate this
2385	 * should be done.
2386	 */
2387	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2388		/* determine the bit number within the word and
2389		 * the number of bits within the word.
2390		 */
2391		wbitno = dbitno & (DBWORD - 1);
2392		nb = min(rembits, DBWORD - wbitno);
2393
2394		/* check if only part of a word is to be freed.
2395		 */
2396		if (nb < DBWORD) {
2397			/* free (zero) the appropriate bits within this
2398			 * dmap word.
2399			 */
2400			dp->wmap[word] &=
2401			    cpu_to_le32(~(ONES << (DBWORD - nb)
2402					  >> wbitno));
2403
2404			/* update the leaf for this dmap word.
2405			 */
2406			rc = dbJoin(tp, word,
2407				    dbMaxBud((u8 *)&dp->wmap[word]), false);
2408			if (rc)
2409				return rc;
2410
2411			word += 1;
2412		} else {
2413			/* one or more dmap words are fully contained
2414			 * within the block range.  determine how many
2415			 * words and free (zero) the bits of these words.
2416			 */
2417			nwords = rembits >> L2DBWORD;
2418			memset(&dp->wmap[word], 0, nwords * 4);
2419
2420			/* determine how many bits.
2421			 */
2422			nb = nwords << L2DBWORD;
2423
2424			/* now update the appropriate leaves to reflect
2425			 * the freed words.
2426			 */
2427			for (; nwords > 0; nwords -= nw) {
2428				/* determine what the leaf value should be
2429				 * updated to as the minimum of the l2 number
2430				 * of bits being freed and the l2 (max) number
2431				 * of bits that can be described by this leaf.
2432				 */
2433				size =
2434				    min(LITOL2BSZ
2435					(word, L2LPERDMAP, BUDMIN),
2436					NLSTOL2BSZ(nwords));
2437
2438				/* update the leaf.
2439				 */
2440				rc = dbJoin(tp, word, size, false);
2441				if (rc)
2442					return rc;
2443
2444				/* get the number of dmap words handled.
2445				 */
2446				nw = BUDSIZE(size, BUDMIN);
2447				word += nw;
2448			}
2449		}
2450	}
2451
2452	/* update the free count for this dmap.
2453	 */
2454	le32_add_cpu(&dp->nfree, nblocks);
2455
2456	BMAP_LOCK(bmp);
2457
2458	/* update the free count for the allocation group and
2459	 * map.
2460	 */
2461	agno = blkno >> bmp->db_agl2size;
2462	bmp->db_nfree += nblocks;
2463	bmp->db_agfree[agno] += nblocks;
2464
2465	/* check if this allocation group is not completely free and
2466	 * if it is currently the maximum (rightmost) allocation group.
2467	 * if so, establish the new maximum allocation group number by
2468	 * searching left for the first allocation group with allocation.
2469	 */
2470	if ((bmp->db_agfree[agno] == bmp->db_agsize && agno == bmp->db_maxag) ||
2471	    (agno == bmp->db_numag - 1 &&
2472	     bmp->db_agfree[agno] == (bmp-> db_mapsize & (BPERDMAP - 1)))) {
2473		while (bmp->db_maxag > 0) {
2474			bmp->db_maxag -= 1;
2475			if (bmp->db_agfree[bmp->db_maxag] !=
2476			    bmp->db_agsize)
2477				break;
2478		}
2479
2480		/* re-establish the allocation group preference if the
2481		 * current preference is right of the maximum allocation
2482		 * group.
2483		 */
2484		if (bmp->db_agpref > bmp->db_maxag)
2485			bmp->db_agpref = bmp->db_maxag;
2486	}
2487
2488	BMAP_UNLOCK(bmp);
2489
2490	return 0;
2491}
2492
2493
2494/*
2495 * NAME:	dbAdjCtl()
2496 *
2497 * FUNCTION:	adjust a dmap control page at a specified level to reflect
2498 *		the change in a lower level dmap or dmap control page's
2499 *		maximum string of free blocks (i.e. a change in the root
2500 *		of the lower level object's dmtree) due to the allocation
2501 *		or deallocation of a range of blocks with a single dmap.
2502 *
2503 *		on entry, this routine is provided with the new value of
2504 *		the lower level dmap or dmap control page root and the
2505 *		starting block number of the block range whose allocation
2506 *		or deallocation resulted in the root change.  this range
2507 *		is respresented by a single leaf of the current dmapctl
2508 *		and the leaf will be updated with this value, possibly
2509 *		causing a binary buddy system within the leaves to be
2510 *		split or joined.  the update may also cause the dmapctl's
2511 *		dmtree to be updated.
2512 *
2513 *		if the adjustment of the dmap control page, itself, causes its
2514 *		root to change, this change will be bubbled up to the next dmap
2515 *		control level by a recursive call to this routine, specifying
2516 *		the new root value and the next dmap control page level to
2517 *		be adjusted.
2518 * PARAMETERS:
2519 *	bmp	-  pointer to bmap descriptor
2520 *	blkno	-  the first block of a block range within a dmap.  it is
2521 *		   the allocation or deallocation of this block range that
2522 *		   requires the dmap control page to be adjusted.
2523 *	newval	-  the new value of the lower level dmap or dmap control
2524 *		   page root.
2525 *	alloc	-  'true' if adjustment is due to an allocation.
2526 *	level	-  current level of dmap control page (i.e. L0, L1, L2) to
2527 *		   be adjusted.
2528 *
2529 * RETURN VALUES:
2530 *	0	- success
2531 *	-EIO	- i/o error
2532 *
2533 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2534 */
2535static int
2536dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, int level)
2537{
2538	struct metapage *mp;
2539	s8 oldroot;
2540	int oldval;
2541	s64 lblkno;
2542	struct dmapctl *dcp;
2543	int rc, leafno, ti;
2544
2545	/* get the buffer for the dmap control page for the specified
2546	 * block number and control page level.
2547	 */
2548	lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, level);
2549	mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
2550	if (mp == NULL)
2551		return -EIO;
2552	dcp = (struct dmapctl *) mp->data;
2553
2554	if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
2555		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
2556		release_metapage(mp);
2557		return -EIO;
2558	}
2559
2560	/* determine the leaf number corresponding to the block and
2561	 * the index within the dmap control tree.
2562	 */
2563	leafno = BLKTOCTLLEAF(blkno, dcp->budmin);
2564	ti = leafno + le32_to_cpu(dcp->leafidx);
2565
2566	/* save the current leaf value and the current root level (i.e.
2567	 * maximum l2 free string described by this dmapctl).
2568	 */
2569	oldval = dcp->stree[ti];
2570	oldroot = dcp->stree[ROOT];
2571
2572	/* check if this is a control page update for an allocation.
2573	 * if so, update the leaf to reflect the new leaf value using
2574	 * dbSplit(); otherwise (deallocation), use dbJoin() to update
2575	 * the leaf with the new value.  in addition to updating the
2576	 * leaf, dbSplit() will also split the binary buddy system of
2577	 * the leaves, if required, and bubble new values within the
2578	 * dmapctl tree, if required.  similarly, dbJoin() will join
2579	 * the binary buddy system of leaves and bubble new values up
2580	 * the dmapctl tree as required by the new leaf value.
2581	 */
2582	if (alloc) {
2583		/* check if we are in the middle of a binary buddy
2584		 * system.  this happens when we are performing the
2585		 * first allocation out of an allocation group that
2586		 * is part (not the first part) of a larger binary
2587		 * buddy system.  if we are in the middle, back split
2588		 * the system prior to calling dbSplit() which assumes
2589		 * that it is at the front of a binary buddy system.
2590		 */
2591		if (oldval == NOFREE) {
2592			rc = dbBackSplit((dmtree_t *)dcp, leafno, true);
2593			if (rc)
2594				return rc;
2595			oldval = dcp->stree[ti];
2596		}
2597		dbSplit((dmtree_t *) dcp, leafno, dcp->budmin, newval, true);
2598	} else {
2599		rc = dbJoin((dmtree_t *) dcp, leafno, newval, true);
2600		if (rc)
2601			return rc;
2602	}
2603
2604	/* check if the root of the current dmap control page changed due
2605	 * to the update and if the current dmap control page is not at
2606	 * the current top level (i.e. L0, L1, L2) of the map.  if so (i.e.
2607	 * root changed and this is not the top level), call this routine
2608	 * again (recursion) for the next higher level of the mapping to
2609	 * reflect the change in root for the current dmap control page.
2610	 */
2611	if (dcp->stree[ROOT] != oldroot) {
2612		/* are we below the top level of the map.  if so,
2613		 * bubble the root up to the next higher level.
2614		 */
2615		if (level < bmp->db_maxlevel) {
2616			/* bubble up the new root of this dmap control page to
2617			 * the next level.
2618			 */
2619			if ((rc =
2620			     dbAdjCtl(bmp, blkno, dcp->stree[ROOT], alloc,
2621				      level + 1))) {
2622				/* something went wrong in bubbling up the new
2623				 * root value, so backout the changes to the
2624				 * current dmap control page.
2625				 */
2626				if (alloc) {
2627					dbJoin((dmtree_t *) dcp, leafno,
2628					       oldval, true);
2629				} else {
2630					/* the dbJoin() above might have
2631					 * caused a larger binary buddy system
2632					 * to form and we may now be in the
2633					 * middle of it.  if this is the case,
2634					 * back split the buddies.
2635					 */
2636					if (dcp->stree[ti] == NOFREE)
2637						dbBackSplit((dmtree_t *)
2638							    dcp, leafno, true);
2639					dbSplit((dmtree_t *) dcp, leafno,
2640						dcp->budmin, oldval, true);
2641				}
2642
2643				/* release the buffer and return the error.
2644				 */
2645				release_metapage(mp);
2646				return (rc);
2647			}
2648		} else {
2649			/* we're at the top level of the map. update
2650			 * the bmap control page to reflect the size
2651			 * of the maximum free buddy system.
2652			 */
2653			assert(level == bmp->db_maxlevel);
2654			if (bmp->db_maxfreebud != oldroot) {
2655				jfs_error(bmp->db_ipbmap->i_sb,
2656					  "the maximum free buddy is not the old root\n");
2657			}
2658			bmp->db_maxfreebud = dcp->stree[ROOT];
2659		}
2660	}
2661
2662	/* write the buffer.
2663	 */
2664	write_metapage(mp);
2665
2666	return (0);
2667}
2668
2669
2670/*
2671 * NAME:	dbSplit()
2672 *
2673 * FUNCTION:	update the leaf of a dmtree with a new value, splitting
2674 *		the leaf from the binary buddy system of the dmtree's
2675 *		leaves, as required.
2676 *
2677 * PARAMETERS:
2678 *	tp	- pointer to the tree containing the leaf.
2679 *	leafno	- the number of the leaf to be updated.
2680 *	splitsz	- the size the binary buddy system starting at the leaf
2681 *		  must be split to, specified as the log2 number of blocks.
2682 *	newval	- the new value for the leaf.
2683 *
2684 * RETURN VALUES: none
2685 *
2686 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2687 */
2688static void dbSplit(dmtree_t *tp, int leafno, int splitsz, int newval, bool is_ctl)
2689{
2690	int budsz;
2691	int cursz;
2692	s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2693
2694	/* check if the leaf needs to be split.
2695	 */
2696	if (leaf[leafno] > tp->dmt_budmin) {
2697		/* the split occurs by cutting the buddy system in half
2698		 * at the specified leaf until we reach the specified
2699		 * size.  pick up the starting split size (current size
2700		 * - 1 in l2) and the corresponding buddy size.
2701		 */
2702		cursz = leaf[leafno] - 1;
2703		budsz = BUDSIZE(cursz, tp->dmt_budmin);
2704
2705		/* split until we reach the specified size.
2706		 */
2707		while (cursz >= splitsz) {
2708			/* update the buddy's leaf with its new value.
2709			 */
2710			dbAdjTree(tp, leafno ^ budsz, cursz, is_ctl);
2711
2712			/* on to the next size and buddy.
2713			 */
2714			cursz -= 1;
2715			budsz >>= 1;
2716		}
2717	}
2718
2719	/* adjust the dmap tree to reflect the specified leaf's new
2720	 * value.
2721	 */
2722	dbAdjTree(tp, leafno, newval, is_ctl);
2723}
2724
2725
2726/*
2727 * NAME:	dbBackSplit()
2728 *
2729 * FUNCTION:	back split the binary buddy system of dmtree leaves
2730 *		that hold a specified leaf until the specified leaf
2731 *		starts its own binary buddy system.
2732 *
2733 *		the allocators typically perform allocations at the start
2734 *		of binary buddy systems and dbSplit() is used to accomplish
2735 *		any required splits.  in some cases, however, allocation
2736 *		may occur in the middle of a binary system and requires a
2737 *		back split, with the split proceeding out from the middle of
2738 *		the system (less efficient) rather than the start of the
2739 *		system (more efficient).  the cases in which a back split
2740 *		is required are rare and are limited to the first allocation
2741 *		within an allocation group which is a part (not first part)
2742 *		of a larger binary buddy system and a few exception cases
2743 *		in which a previous join operation must be backed out.
2744 *
2745 * PARAMETERS:
2746 *	tp	- pointer to the tree containing the leaf.
2747 *	leafno	- the number of the leaf to be updated.
2748 *
2749 * RETURN VALUES: none
2750 *
2751 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2752 */
2753static int dbBackSplit(dmtree_t *tp, int leafno, bool is_ctl)
2754{
2755	int budsz, bud, w, bsz, size;
2756	int cursz;
2757	s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2758
2759	/* leaf should be part (not first part) of a binary
2760	 * buddy system.
2761	 */
2762	assert(leaf[leafno] == NOFREE);
2763
2764	/* the back split is accomplished by iteratively finding the leaf
2765	 * that starts the buddy system that contains the specified leaf and
2766	 * splitting that system in two.  this iteration continues until
2767	 * the specified leaf becomes the start of a buddy system.
2768	 *
2769	 * determine maximum possible l2 size for the specified leaf.
2770	 */
2771	size =
2772	    LITOL2BSZ(leafno, le32_to_cpu(tp->dmt_l2nleafs),
2773		      tp->dmt_budmin);
2774
2775	/* determine the number of leaves covered by this size.  this
2776	 * is the buddy size that we will start with as we search for
2777	 * the buddy system that contains the specified leaf.
2778	 */
2779	budsz = BUDSIZE(size, tp->dmt_budmin);
2780
2781	/* back split.
2782	 */
2783	while (leaf[leafno] == NOFREE) {
2784		/* find the leftmost buddy leaf.
2785		 */
2786		for (w = leafno, bsz = budsz;; bsz <<= 1,
2787		     w = (w < bud) ? w : bud) {
2788			if (bsz >= le32_to_cpu(tp->dmt_nleafs)) {
2789				jfs_err("JFS: block map error in dbBackSplit");
2790				return -EIO;
2791			}
2792
2793			/* determine the buddy.
2794			 */
2795			bud = w ^ bsz;
2796
2797			/* check if this buddy is the start of the system.
2798			 */
2799			if (leaf[bud] != NOFREE) {
2800				/* split the leaf at the start of the
2801				 * system in two.
2802				 */
2803				cursz = leaf[bud] - 1;
2804				dbSplit(tp, bud, cursz, cursz, is_ctl);
2805				break;
2806			}
2807		}
2808	}
2809
2810	if (leaf[leafno] != size) {
2811		jfs_err("JFS: wrong leaf value in dbBackSplit");
2812		return -EIO;
2813	}
2814	return 0;
2815}
2816
2817
2818/*
2819 * NAME:	dbJoin()
2820 *
2821 * FUNCTION:	update the leaf of a dmtree with a new value, joining
2822 *		the leaf with other leaves of the dmtree into a multi-leaf
2823 *		binary buddy system, as required.
2824 *
2825 * PARAMETERS:
2826 *	tp	- pointer to the tree containing the leaf.
2827 *	leafno	- the number of the leaf to be updated.
2828 *	newval	- the new value for the leaf.
2829 *
2830 * RETURN VALUES: none
2831 */
2832static int dbJoin(dmtree_t *tp, int leafno, int newval, bool is_ctl)
2833{
2834	int budsz, buddy;
2835	s8 *leaf;
2836
2837	/* can the new leaf value require a join with other leaves ?
2838	 */
2839	if (newval >= tp->dmt_budmin) {
2840		/* pickup a pointer to the leaves of the tree.
2841		 */
2842		leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2843
2844		/* try to join the specified leaf into a large binary
2845		 * buddy system.  the join proceeds by attempting to join
2846		 * the specified leafno with its buddy (leaf) at new value.
2847		 * if the join occurs, we attempt to join the left leaf
2848		 * of the joined buddies with its buddy at new value + 1.
2849		 * we continue to join until we find a buddy that cannot be
2850		 * joined (does not have a value equal to the size of the
2851		 * last join) or until all leaves have been joined into a
2852		 * single system.
2853		 *
2854		 * get the buddy size (number of words covered) of
2855		 * the new value.
2856		 */
2857		budsz = BUDSIZE(newval, tp->dmt_budmin);
2858
2859		/* try to join.
2860		 */
2861		while (budsz < le32_to_cpu(tp->dmt_nleafs)) {
2862			/* get the buddy leaf.
2863			 */
2864			buddy = leafno ^ budsz;
2865
2866			/* if the leaf's new value is greater than its
2867			 * buddy's value, we join no more.
2868			 */
2869			if (newval > leaf[buddy])
2870				break;
2871
2872			/* It shouldn't be less */
2873			if (newval < leaf[buddy])
2874				return -EIO;
2875
2876			/* check which (leafno or buddy) is the left buddy.
2877			 * the left buddy gets to claim the blocks resulting
2878			 * from the join while the right gets to claim none.
2879			 * the left buddy is also eligible to participate in
2880			 * a join at the next higher level while the right
2881			 * is not.
2882			 *
2883			 */
2884			if (leafno < buddy) {
2885				/* leafno is the left buddy.
2886				 */
2887				dbAdjTree(tp, buddy, NOFREE, is_ctl);
2888			} else {
2889				/* buddy is the left buddy and becomes
2890				 * leafno.
2891				 */
2892				dbAdjTree(tp, leafno, NOFREE, is_ctl);
2893				leafno = buddy;
2894			}
2895
2896			/* on to try the next join.
2897			 */
2898			newval += 1;
2899			budsz <<= 1;
2900		}
2901	}
2902
2903	/* update the leaf value.
2904	 */
2905	dbAdjTree(tp, leafno, newval, is_ctl);
2906
2907	return 0;
2908}
2909
2910
2911/*
2912 * NAME:	dbAdjTree()
2913 *
2914 * FUNCTION:	update a leaf of a dmtree with a new value, adjusting
2915 *		the dmtree, as required, to reflect the new leaf value.
2916 *		the combination of any buddies must already be done before
2917 *		this is called.
2918 *
2919 * PARAMETERS:
2920 *	tp	- pointer to the tree to be adjusted.
2921 *	leafno	- the number of the leaf to be updated.
2922 *	newval	- the new value for the leaf.
2923 *
2924 * RETURN VALUES: none
2925 */
2926static void dbAdjTree(dmtree_t *tp, int leafno, int newval, bool is_ctl)
2927{
2928	int lp, pp, k;
2929	int max, size;
2930
2931	size = is_ctl ? CTLTREESIZE : TREESIZE;
2932
2933	/* pick up the index of the leaf for this leafno.
2934	 */
2935	lp = leafno + le32_to_cpu(tp->dmt_leafidx);
2936
2937	if (WARN_ON_ONCE(lp >= size || lp < 0))
2938		return;
2939
2940	/* is the current value the same as the old value ?  if so,
2941	 * there is nothing to do.
2942	 */
2943	if (tp->dmt_stree[lp] == newval)
2944		return;
2945
2946	/* set the new value.
2947	 */
2948	tp->dmt_stree[lp] = newval;
2949
2950	/* bubble the new value up the tree as required.
2951	 */
2952	for (k = 0; k < le32_to_cpu(tp->dmt_height); k++) {
2953		/* get the index of the first leaf of the 4 leaf
2954		 * group containing the specified leaf (leafno).
2955		 */
2956		lp = ((lp - 1) & ~0x03) + 1;
2957
2958		/* get the index of the parent of this 4 leaf group.
2959		 */
2960		pp = (lp - 1) >> 2;
2961
2962		/* determine the maximum of the 4 leaves.
2963		 */
2964		max = TREEMAX(&tp->dmt_stree[lp]);
2965
2966		/* if the maximum of the 4 is the same as the
2967		 * parent's value, we're done.
2968		 */
2969		if (tp->dmt_stree[pp] == max)
2970			break;
2971
2972		/* parent gets new value.
2973		 */
2974		tp->dmt_stree[pp] = max;
2975
2976		/* parent becomes leaf for next go-round.
2977		 */
2978		lp = pp;
2979	}
2980}
2981
2982
2983/*
2984 * NAME:	dbFindLeaf()
2985 *
2986 * FUNCTION:	search a dmtree_t for sufficient free blocks, returning
2987 *		the index of a leaf describing the free blocks if
2988 *		sufficient free blocks are found.
2989 *
2990 *		the search starts at the top of the dmtree_t tree and
2991 *		proceeds down the tree to the leftmost leaf with sufficient
2992 *		free space.
2993 *
2994 * PARAMETERS:
2995 *	tp	- pointer to the tree to be searched.
2996 *	l2nb	- log2 number of free blocks to search for.
2997 *	leafidx	- return pointer to be set to the index of the leaf
2998 *		  describing at least l2nb free blocks if sufficient
2999 *		  free blocks are found.
3000 *	is_ctl	- determines if the tree is of type ctl
3001 *
3002 * RETURN VALUES:
3003 *	0	- success
3004 *	-ENOSPC	- insufficient free blocks.
3005 */
3006static int dbFindLeaf(dmtree_t *tp, int l2nb, int *leafidx, bool is_ctl)
3007{
3008	int ti, n = 0, k, x = 0;
3009	int max_size;
3010
3011	max_size = is_ctl ? CTLTREESIZE : TREESIZE;
3012
3013	/* first check the root of the tree to see if there is
3014	 * sufficient free space.
3015	 */
3016	if (l2nb > tp->dmt_stree[ROOT])
3017		return -ENOSPC;
3018
3019	/* sufficient free space available. now search down the tree
3020	 * starting at the next level for the leftmost leaf that
3021	 * describes sufficient free space.
3022	 */
3023	for (k = le32_to_cpu(tp->dmt_height), ti = 1;
3024	     k > 0; k--, ti = ((ti + n) << 2) + 1) {
3025		/* search the four nodes at this level, starting from
3026		 * the left.
3027		 */
3028		for (x = ti, n = 0; n < 4; n++) {
3029			/* sufficient free space found.  move to the next
3030			 * level (or quit if this is the last level).
3031			 */
3032			if (x + n > max_size)
3033				return -ENOSPC;
3034			if (l2nb <= tp->dmt_stree[x + n])
3035				break;
3036		}
3037
3038		/* better have found something since the higher
3039		 * levels of the tree said it was here.
3040		 */
3041		assert(n < 4);
3042	}
3043
3044	/* set the return to the leftmost leaf describing sufficient
3045	 * free space.
3046	 */
3047	*leafidx = x + n - le32_to_cpu(tp->dmt_leafidx);
3048
3049	return (0);
3050}
3051
3052
3053/*
3054 * NAME:	dbFindBits()
3055 *
3056 * FUNCTION:	find a specified number of binary buddy free bits within a
3057 *		dmap bitmap word value.
3058 *
3059 *		this routine searches the bitmap value for (1 << l2nb) free
3060 *		bits at (1 << l2nb) alignments within the value.
3061 *
3062 * PARAMETERS:
3063 *	word	-  dmap bitmap word value.
3064 *	l2nb	-  number of free bits specified as a log2 number.
3065 *
3066 * RETURN VALUES:
3067 *	starting bit number of free bits.
3068 */
3069static int dbFindBits(u32 word, int l2nb)
3070{
3071	int bitno, nb;
3072	u32 mask;
3073
3074	/* get the number of bits.
3075	 */
3076	nb = 1 << l2nb;
3077	assert(nb <= DBWORD);
3078
3079	/* complement the word so we can use a mask (i.e. 0s represent
3080	 * free bits) and compute the mask.
3081	 */
3082	word = ~word;
3083	mask = ONES << (DBWORD - nb);
3084
3085	/* scan the word for nb free bits at nb alignments.
3086	 */
3087	for (bitno = 0; mask != 0; bitno += nb, mask >>= nb) {
3088		if ((mask & word) == mask)
3089			break;
3090	}
3091
3092	ASSERT(bitno < 32);
3093
3094	/* return the bit number.
3095	 */
3096	return (bitno);
3097}
3098
3099
3100/*
3101 * NAME:	dbMaxBud(u8 *cp)
3102 *
3103 * FUNCTION:	determine the largest binary buddy string of free
3104 *		bits within 32-bits of the map.
3105 *
3106 * PARAMETERS:
3107 *	cp	-  pointer to the 32-bit value.
3108 *
3109 * RETURN VALUES:
3110 *	largest binary buddy of free bits within a dmap word.
3111 */
3112static int dbMaxBud(u8 * cp)
3113{
3114	signed char tmp1, tmp2;
3115
3116	/* check if the wmap word is all free. if so, the
3117	 * free buddy size is BUDMIN.
3118	 */
3119	if (*((uint *) cp) == 0)
3120		return (BUDMIN);
3121
3122	/* check if the wmap word is half free. if so, the
3123	 * free buddy size is BUDMIN-1.
3124	 */
3125	if (*((u16 *) cp) == 0 || *((u16 *) cp + 1) == 0)
3126		return (BUDMIN - 1);
3127
3128	/* not all free or half free. determine the free buddy
3129	 * size thru table lookup using quarters of the wmap word.
3130	 */
3131	tmp1 = max(budtab[cp[2]], budtab[cp[3]]);
3132	tmp2 = max(budtab[cp[0]], budtab[cp[1]]);
3133	return (max(tmp1, tmp2));
3134}
3135
3136
3137/*
3138 * NAME:	cnttz(uint word)
3139 *
3140 * FUNCTION:	determine the number of trailing zeros within a 32-bit
3141 *		value.
3142 *
3143 * PARAMETERS:
3144 *	value	-  32-bit value to be examined.
3145 *
3146 * RETURN VALUES:
3147 *	count of trailing zeros
3148 */
3149static int cnttz(u32 word)
3150{
3151	int n;
3152
3153	for (n = 0; n < 32; n++, word >>= 1) {
3154		if (word & 0x01)
3155			break;
3156	}
3157
3158	return (n);
3159}
3160
3161
3162/*
3163 * NAME:	cntlz(u32 value)
3164 *
3165 * FUNCTION:	determine the number of leading zeros within a 32-bit
3166 *		value.
3167 *
3168 * PARAMETERS:
3169 *	value	-  32-bit value to be examined.
3170 *
3171 * RETURN VALUES:
3172 *	count of leading zeros
3173 */
3174static int cntlz(u32 value)
3175{
3176	int n;
3177
3178	for (n = 0; n < 32; n++, value <<= 1) {
3179		if (value & HIGHORDER)
3180			break;
3181	}
3182	return (n);
3183}
3184
3185
3186/*
3187 * NAME:	blkstol2(s64 nb)
3188 *
3189 * FUNCTION:	convert a block count to its log2 value. if the block
3190 *		count is not a l2 multiple, it is rounded up to the next
3191 *		larger l2 multiple.
3192 *
3193 * PARAMETERS:
3194 *	nb	-  number of blocks
3195 *
3196 * RETURN VALUES:
3197 *	log2 number of blocks
3198 */
3199static int blkstol2(s64 nb)
3200{
3201	int l2nb;
3202	s64 mask;		/* meant to be signed */
3203
3204	mask = (s64) 1 << (64 - 1);
3205
3206	/* count the leading bits.
3207	 */
3208	for (l2nb = 0; l2nb < 64; l2nb++, mask >>= 1) {
3209		/* leading bit found.
3210		 */
3211		if (nb & mask) {
3212			/* determine the l2 value.
3213			 */
3214			l2nb = (64 - 1) - l2nb;
3215
3216			/* check if we need to round up.
3217			 */
3218			if (~mask & nb)
3219				l2nb++;
3220
3221			return (l2nb);
3222		}
3223	}
3224	assert(0);
3225	return 0;		/* fix compiler warning */
3226}
3227
3228
3229/*
3230 * NAME:	dbAllocBottomUp()
3231 *
3232 * FUNCTION:	alloc the specified block range from the working block
3233 *		allocation map.
3234 *
3235 *		the blocks will be alloc from the working map one dmap
3236 *		at a time.
3237 *
3238 * PARAMETERS:
3239 *	ip	-  pointer to in-core inode;
3240 *	blkno	-  starting block number to be freed.
3241 *	nblocks	-  number of blocks to be freed.
3242 *
3243 * RETURN VALUES:
3244 *	0	- success
3245 *	-EIO	- i/o error
3246 */
3247int dbAllocBottomUp(struct inode *ip, s64 blkno, s64 nblocks)
3248{
3249	struct metapage *mp;
3250	struct dmap *dp;
3251	int nb, rc;
3252	s64 lblkno, rem;
3253	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
3254	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
3255
3256	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
3257
3258	/* block to be allocated better be within the mapsize. */
3259	ASSERT(nblocks <= bmp->db_mapsize - blkno);
3260
3261	/*
3262	 * allocate the blocks a dmap at a time.
3263	 */
3264	mp = NULL;
3265	for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
3266		/* release previous dmap if any */
3267		if (mp) {
3268			write_metapage(mp);
3269		}
3270
3271		/* get the buffer for the current dmap. */
3272		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
3273		mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
3274		if (mp == NULL) {
3275			IREAD_UNLOCK(ipbmap);
3276			return -EIO;
3277		}
3278		dp = (struct dmap *) mp->data;
3279
3280		/* determine the number of blocks to be allocated from
3281		 * this dmap.
3282		 */
3283		nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
3284
3285		/* allocate the blocks. */
3286		if ((rc = dbAllocDmapBU(bmp, dp, blkno, nb))) {
3287			release_metapage(mp);
3288			IREAD_UNLOCK(ipbmap);
3289			return (rc);
3290		}
3291	}
3292
3293	/* write the last buffer. */
3294	write_metapage(mp);
3295
3296	IREAD_UNLOCK(ipbmap);
3297
3298	return (0);
3299}
3300
3301
3302static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
3303			 int nblocks)
3304{
3305	int rc;
3306	int dbitno, word, rembits, nb, nwords, wbitno, agno;
3307	s8 oldroot;
3308	struct dmaptree *tp = (struct dmaptree *) & dp->tree;
3309
3310	/* save the current value of the root (i.e. maximum free string)
3311	 * of the dmap tree.
3312	 */
3313	oldroot = tp->stree[ROOT];
3314
3315	/* determine the bit number and word within the dmap of the
3316	 * starting block.
3317	 */
3318	dbitno = blkno & (BPERDMAP - 1);
3319	word = dbitno >> L2DBWORD;
3320
3321	/* block range better be within the dmap */
3322	assert(dbitno + nblocks <= BPERDMAP);
3323
3324	/* allocate the bits of the dmap's words corresponding to the block
3325	 * range. not all bits of the first and last words may be contained
3326	 * within the block range.  if this is the case, we'll work against
3327	 * those words (i.e. partial first and/or last) on an individual basis
3328	 * (a single pass), allocating the bits of interest by hand and
3329	 * updating the leaf corresponding to the dmap word. a single pass
3330	 * will be used for all dmap words fully contained within the
3331	 * specified range.  within this pass, the bits of all fully contained
3332	 * dmap words will be marked as free in a single shot and the leaves
3333	 * will be updated. a single leaf may describe the free space of
3334	 * multiple dmap words, so we may update only a subset of the actual
3335	 * leaves corresponding to the dmap words of the block range.
3336	 */
3337	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
3338		/* determine the bit number within the word and
3339		 * the number of bits within the word.
3340		 */
3341		wbitno = dbitno & (DBWORD - 1);
3342		nb = min(rembits, DBWORD - wbitno);
3343
3344		/* check if only part of a word is to be allocated.
3345		 */
3346		if (nb < DBWORD) {
3347			/* allocate (set to 1) the appropriate bits within
3348			 * this dmap word.
3349			 */
3350			dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
3351						      >> wbitno);
3352
3353			word++;
3354		} else {
3355			/* one or more dmap words are fully contained
3356			 * within the block range.  determine how many
3357			 * words and allocate (set to 1) the bits of these
3358			 * words.
3359			 */
3360			nwords = rembits >> L2DBWORD;
3361			memset(&dp->wmap[word], (int) ONES, nwords * 4);
3362
3363			/* determine how many bits */
3364			nb = nwords << L2DBWORD;
3365			word += nwords;
3366		}
3367	}
3368
3369	/* update the free count for this dmap */
3370	le32_add_cpu(&dp->nfree, -nblocks);
3371
3372	/* reconstruct summary tree */
3373	dbInitDmapTree(dp);
3374
3375	BMAP_LOCK(bmp);
3376
3377	/* if this allocation group is completely free,
3378	 * update the highest active allocation group number
3379	 * if this allocation group is the new max.
3380	 */
3381	agno = blkno >> bmp->db_agl2size;
3382	if (agno > bmp->db_maxag)
3383		bmp->db_maxag = agno;
3384
3385	/* update the free count for the allocation group and map */
3386	bmp->db_agfree[agno] -= nblocks;
3387	bmp->db_nfree -= nblocks;
3388
3389	BMAP_UNLOCK(bmp);
3390
3391	/* if the root has not changed, done. */
3392	if (tp->stree[ROOT] == oldroot)
3393		return (0);
3394
3395	/* root changed. bubble the change up to the dmap control pages.
3396	 * if the adjustment of the upper level control pages fails,
3397	 * backout the bit allocation (thus making everything consistent).
3398	 */
3399	if ((rc = dbAdjCtl(bmp, blkno, tp->stree[ROOT], 1, 0)))
3400		dbFreeBits(bmp, dp, blkno, nblocks);
3401
3402	return (rc);
3403}
3404
3405
3406/*
3407 * NAME:	dbExtendFS()
3408 *
3409 * FUNCTION:	extend bmap from blkno for nblocks;
3410 *		dbExtendFS() updates bmap ready for dbAllocBottomUp();
3411 *
3412 * L2
3413 *  |
3414 *   L1---------------------------------L1
3415 *    |					 |
3416 *     L0---------L0---------L0		  L0---------L0---------L0
3417 *      |	   |	      |		   |	      |		 |
3418 *	 d0,...,dn  d0,...,dn  d0,...,dn    d0,...,dn  d0,...,dn  d0,.,dm;
3419 * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm
3420 *
3421 * <---old---><----------------------------extend----------------------->
3422 */
3423int dbExtendFS(struct inode *ipbmap, s64 blkno,	s64 nblocks)
3424{
3425	struct jfs_sb_info *sbi = JFS_SBI(ipbmap->i_sb);
3426	int nbperpage = sbi->nbperpage;
3427	int i, i0 = true, j, j0 = true, k, n;
3428	s64 newsize;
3429	s64 p;
3430	struct metapage *mp, *l2mp, *l1mp = NULL, *l0mp = NULL;
3431	struct dmapctl *l2dcp, *l1dcp, *l0dcp;
3432	struct dmap *dp;
3433	s8 *l0leaf, *l1leaf, *l2leaf;
3434	struct bmap *bmp = sbi->bmap;
3435	int agno, l2agsize, oldl2agsize;
3436	s64 ag_rem;
3437
3438	newsize = blkno + nblocks;
3439
3440	jfs_info("dbExtendFS: blkno:%Ld nblocks:%Ld newsize:%Ld",
3441		 (long long) blkno, (long long) nblocks, (long long) newsize);
3442
3443	/*
3444	 *	initialize bmap control page.
3445	 *
3446	 * all the data in bmap control page should exclude
3447	 * the mkfs hidden dmap page.
3448	 */
3449
3450	/* update mapsize */
3451	bmp->db_mapsize = newsize;
3452	bmp->db_maxlevel = BMAPSZTOLEV(bmp->db_mapsize);
3453
3454	/* compute new AG size */
3455	l2agsize = dbGetL2AGSize(newsize);
3456	oldl2agsize = bmp->db_agl2size;
3457
3458	bmp->db_agl2size = l2agsize;
3459	bmp->db_agsize = 1 << l2agsize;
3460
3461	/* compute new number of AG */
3462	agno = bmp->db_numag;
3463	bmp->db_numag = newsize >> l2agsize;
3464	bmp->db_numag += ((u32) newsize % (u32) bmp->db_agsize) ? 1 : 0;
3465
3466	/*
3467	 *	reconfigure db_agfree[]
3468	 * from old AG configuration to new AG configuration;
3469	 *
3470	 * coalesce contiguous k (newAGSize/oldAGSize) AGs;
3471	 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn;
3472	 * note: new AG size = old AG size * (2**x).
3473	 */
3474	if (l2agsize == oldl2agsize)
3475		goto extend;
3476	k = 1 << (l2agsize - oldl2agsize);
3477	ag_rem = bmp->db_agfree[0];	/* save agfree[0] */
3478	for (i = 0, n = 0; i < agno; n++) {
3479		bmp->db_agfree[n] = 0;	/* init collection point */
3480
3481		/* coalesce contiguous k AGs; */
3482		for (j = 0; j < k && i < agno; j++, i++) {
3483			/* merge AGi to AGn */
3484			bmp->db_agfree[n] += bmp->db_agfree[i];
3485		}
3486	}
3487	bmp->db_agfree[0] += ag_rem;	/* restore agfree[0] */
3488
3489	for (; n < MAXAG; n++)
3490		bmp->db_agfree[n] = 0;
3491
3492	/*
3493	 * update highest active ag number
3494	 */
3495
3496	bmp->db_maxag = bmp->db_maxag / k;
3497
3498	/*
3499	 *	extend bmap
3500	 *
3501	 * update bit maps and corresponding level control pages;
3502	 * global control page db_nfree, db_agfree[agno], db_maxfreebud;
3503	 */
3504      extend:
3505	/* get L2 page */
3506	p = BMAPBLKNO + nbperpage;	/* L2 page */
3507	l2mp = read_metapage(ipbmap, p, PSIZE, 0);
3508	if (!l2mp) {
3509		jfs_error(ipbmap->i_sb, "L2 page could not be read\n");
3510		return -EIO;
3511	}
3512	l2dcp = (struct dmapctl *) l2mp->data;
3513
3514	/* compute start L1 */
3515	k = blkno >> L2MAXL1SIZE;
3516	l2leaf = l2dcp->stree + CTLLEAFIND + k;
3517	p = BLKTOL1(blkno, sbi->l2nbperpage);	/* L1 page */
3518
3519	/*
3520	 * extend each L1 in L2
3521	 */
3522	for (; k < LPERCTL; k++, p += nbperpage) {
3523		/* get L1 page */
3524		if (j0) {
3525			/* read in L1 page: (blkno & (MAXL1SIZE - 1)) */
3526			l1mp = read_metapage(ipbmap, p, PSIZE, 0);
3527			if (l1mp == NULL)
3528				goto errout;
3529			l1dcp = (struct dmapctl *) l1mp->data;
3530
3531			/* compute start L0 */
3532			j = (blkno & (MAXL1SIZE - 1)) >> L2MAXL0SIZE;
3533			l1leaf = l1dcp->stree + CTLLEAFIND + j;
3534			p = BLKTOL0(blkno, sbi->l2nbperpage);
3535			j0 = false;
3536		} else {
3537			/* assign/init L1 page */
3538			l1mp = get_metapage(ipbmap, p, PSIZE, 0);
3539			if (l1mp == NULL)
3540				goto errout;
3541
3542			l1dcp = (struct dmapctl *) l1mp->data;
3543
3544			/* compute start L0 */
3545			j = 0;
3546			l1leaf = l1dcp->stree + CTLLEAFIND;
3547			p += nbperpage;	/* 1st L0 of L1.k */
3548		}
3549
3550		/*
3551		 * extend each L0 in L1
3552		 */
3553		for (; j < LPERCTL; j++) {
3554			/* get L0 page */
3555			if (i0) {
3556				/* read in L0 page: (blkno & (MAXL0SIZE - 1)) */
3557
3558				l0mp = read_metapage(ipbmap, p, PSIZE, 0);
3559				if (l0mp == NULL)
3560					goto errout;
3561				l0dcp = (struct dmapctl *) l0mp->data;
3562
3563				/* compute start dmap */
3564				i = (blkno & (MAXL0SIZE - 1)) >>
3565				    L2BPERDMAP;
3566				l0leaf = l0dcp->stree + CTLLEAFIND + i;
3567				p = BLKTODMAP(blkno,
3568					      sbi->l2nbperpage);
3569				i0 = false;
3570			} else {
3571				/* assign/init L0 page */
3572				l0mp = get_metapage(ipbmap, p, PSIZE, 0);
3573				if (l0mp == NULL)
3574					goto errout;
3575
3576				l0dcp = (struct dmapctl *) l0mp->data;
3577
3578				/* compute start dmap */
3579				i = 0;
3580				l0leaf = l0dcp->stree + CTLLEAFIND;
3581				p += nbperpage;	/* 1st dmap of L0.j */
3582			}
3583
3584			/*
3585			 * extend each dmap in L0
3586			 */
3587			for (; i < LPERCTL; i++) {
3588				/*
3589				 * reconstruct the dmap page, and
3590				 * initialize corresponding parent L0 leaf
3591				 */
3592				if ((n = blkno & (BPERDMAP - 1))) {
3593					/* read in dmap page: */
3594					mp = read_metapage(ipbmap, p,
3595							   PSIZE, 0);
3596					if (mp == NULL)
3597						goto errout;
3598					n = min(nblocks, (s64)BPERDMAP - n);
3599				} else {
3600					/* assign/init dmap page */
3601					mp = read_metapage(ipbmap, p,
3602							   PSIZE, 0);
3603					if (mp == NULL)
3604						goto errout;
3605
3606					n = min_t(s64, nblocks, BPERDMAP);
3607				}
3608
3609				dp = (struct dmap *) mp->data;
3610				*l0leaf = dbInitDmap(dp, blkno, n);
3611
3612				bmp->db_nfree += n;
3613				agno = le64_to_cpu(dp->start) >> l2agsize;
3614				bmp->db_agfree[agno] += n;
3615
3616				write_metapage(mp);
3617
3618				l0leaf++;
3619				p += nbperpage;
3620
3621				blkno += n;
3622				nblocks -= n;
3623				if (nblocks == 0)
3624					break;
3625			}	/* for each dmap in a L0 */
3626
3627			/*
3628			 * build current L0 page from its leaves, and
3629			 * initialize corresponding parent L1 leaf
3630			 */
3631			*l1leaf = dbInitDmapCtl(l0dcp, 0, ++i);
3632			write_metapage(l0mp);
3633			l0mp = NULL;
3634
3635			if (nblocks)
3636				l1leaf++;	/* continue for next L0 */
3637			else {
3638				/* more than 1 L0 ? */
3639				if (j > 0)
3640					break;	/* build L1 page */
3641				else {
3642					/* summarize in global bmap page */
3643					bmp->db_maxfreebud = *l1leaf;
3644					release_metapage(l1mp);
3645					release_metapage(l2mp);
3646					goto finalize;
3647				}
3648			}
3649		}		/* for each L0 in a L1 */
3650
3651		/*
3652		 * build current L1 page from its leaves, and
3653		 * initialize corresponding parent L2 leaf
3654		 */
3655		*l2leaf = dbInitDmapCtl(l1dcp, 1, ++j);
3656		write_metapage(l1mp);
3657		l1mp = NULL;
3658
3659		if (nblocks)
3660			l2leaf++;	/* continue for next L1 */
3661		else {
3662			/* more than 1 L1 ? */
3663			if (k > 0)
3664				break;	/* build L2 page */
3665			else {
3666				/* summarize in global bmap page */
3667				bmp->db_maxfreebud = *l2leaf;
3668				release_metapage(l2mp);
3669				goto finalize;
3670			}
3671		}
3672	}			/* for each L1 in a L2 */
3673
3674	jfs_error(ipbmap->i_sb, "function has not returned as expected\n");
3675errout:
3676	if (l0mp)
3677		release_metapage(l0mp);
3678	if (l1mp)
3679		release_metapage(l1mp);
3680	release_metapage(l2mp);
3681	return -EIO;
3682
3683	/*
3684	 *	finalize bmap control page
3685	 */
3686finalize:
3687
3688	return 0;
3689}
3690
3691
3692/*
3693 *	dbFinalizeBmap()
3694 */
3695void dbFinalizeBmap(struct inode *ipbmap)
3696{
3697	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
3698	int actags, inactags, l2nl;
3699	s64 ag_rem, actfree, inactfree, avgfree;
3700	int i, n;
3701
3702	/*
3703	 *	finalize bmap control page
3704	 */
3705//finalize:
3706	/*
3707	 * compute db_agpref: preferred ag to allocate from
3708	 * (the leftmost ag with average free space in it);
3709	 */
3710//agpref:
3711	/* get the number of active ags and inacitve ags */
3712	actags = bmp->db_maxag + 1;
3713	inactags = bmp->db_numag - actags;
3714	ag_rem = bmp->db_mapsize & (bmp->db_agsize - 1);	/* ??? */
3715
3716	/* determine how many blocks are in the inactive allocation
3717	 * groups. in doing this, we must account for the fact that
3718	 * the rightmost group might be a partial group (i.e. file
3719	 * system size is not a multiple of the group size).
3720	 */
3721	inactfree = (inactags && ag_rem) ?
3722	    ((inactags - 1) << bmp->db_agl2size) + ag_rem
3723	    : inactags << bmp->db_agl2size;
3724
3725	/* determine how many free blocks are in the active
3726	 * allocation groups plus the average number of free blocks
3727	 * within the active ags.
3728	 */
3729	actfree = bmp->db_nfree - inactfree;
3730	avgfree = (u32) actfree / (u32) actags;
3731
3732	/* if the preferred allocation group has not average free space.
3733	 * re-establish the preferred group as the leftmost
3734	 * group with average free space.
3735	 */
3736	if (bmp->db_agfree[bmp->db_agpref] < avgfree) {
3737		for (bmp->db_agpref = 0; bmp->db_agpref < actags;
3738		     bmp->db_agpref++) {
3739			if (bmp->db_agfree[bmp->db_agpref] >= avgfree)
3740				break;
3741		}
3742		if (bmp->db_agpref >= bmp->db_numag) {
3743			jfs_error(ipbmap->i_sb,
3744				  "cannot find ag with average freespace\n");
3745		}
3746	}
3747
3748	/*
3749	 * compute db_aglevel, db_agheight, db_width, db_agstart:
3750	 * an ag is covered in aglevel dmapctl summary tree,
3751	 * at agheight level height (from leaf) with agwidth number of nodes
3752	 * each, which starts at agstart index node of the smmary tree node
3753	 * array;
3754	 */
3755	bmp->db_aglevel = BMAPSZTOLEV(bmp->db_agsize);
3756	l2nl =
3757	    bmp->db_agl2size - (L2BPERDMAP + bmp->db_aglevel * L2LPERCTL);
3758	bmp->db_agheight = l2nl >> 1;
3759	bmp->db_agwidth = 1 << (l2nl - (bmp->db_agheight << 1));
3760	for (i = 5 - bmp->db_agheight, bmp->db_agstart = 0, n = 1; i > 0;
3761	     i--) {
3762		bmp->db_agstart += n;
3763		n <<= 2;
3764	}
3765
3766}
3767
3768
3769/*
3770 * NAME:	dbInitDmap()/ujfs_idmap_page()
3771 *
3772 * FUNCTION:	initialize working/persistent bitmap of the dmap page
3773 *		for the specified number of blocks:
3774 *
3775 *		at entry, the bitmaps had been initialized as free (ZEROS);
3776 *		The number of blocks will only account for the actually
3777 *		existing blocks. Blocks which don't actually exist in
3778 *		the aggregate will be marked as allocated (ONES);
3779 *
3780 * PARAMETERS:
3781 *	dp	- pointer to page of map
3782 *	nblocks	- number of blocks this page
3783 *
3784 * RETURNS: NONE
3785 */
3786static int dbInitDmap(struct dmap * dp, s64 Blkno, int nblocks)
3787{
3788	int blkno, w, b, r, nw, nb, i;
3789
3790	/* starting block number within the dmap */
3791	blkno = Blkno & (BPERDMAP - 1);
3792
3793	if (blkno == 0) {
3794		dp->nblocks = dp->nfree = cpu_to_le32(nblocks);
3795		dp->start = cpu_to_le64(Blkno);
3796
3797		if (nblocks == BPERDMAP) {
3798			memset(&dp->wmap[0], 0, LPERDMAP * 4);
3799			memset(&dp->pmap[0], 0, LPERDMAP * 4);
3800			goto initTree;
3801		}
3802	} else {
3803		le32_add_cpu(&dp->nblocks, nblocks);
3804		le32_add_cpu(&dp->nfree, nblocks);
3805	}
3806
3807	/* word number containing start block number */
3808	w = blkno >> L2DBWORD;
3809
3810	/*
3811	 * free the bits corresponding to the block range (ZEROS):
3812	 * note: not all bits of the first and last words may be contained
3813	 * within the block range.
3814	 */
3815	for (r = nblocks; r > 0; r -= nb, blkno += nb) {
3816		/* number of bits preceding range to be freed in the word */
3817		b = blkno & (DBWORD - 1);
3818		/* number of bits to free in the word */
3819		nb = min(r, DBWORD - b);
3820
3821		/* is partial word to be freed ? */
3822		if (nb < DBWORD) {
3823			/* free (set to 0) from the bitmap word */
3824			dp->wmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3825						     >> b));
3826			dp->pmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3827						     >> b));
3828
3829			/* skip the word freed */
3830			w++;
3831		} else {
3832			/* free (set to 0) contiguous bitmap words */
3833			nw = r >> L2DBWORD;
3834			memset(&dp->wmap[w], 0, nw * 4);
3835			memset(&dp->pmap[w], 0, nw * 4);
3836
3837			/* skip the words freed */
3838			nb = nw << L2DBWORD;
3839			w += nw;
3840		}
3841	}
3842
3843	/*
3844	 * mark bits following the range to be freed (non-existing
3845	 * blocks) as allocated (ONES)
3846	 */
3847
3848	if (blkno == BPERDMAP)
3849		goto initTree;
3850
3851	/* the first word beyond the end of existing blocks */
3852	w = blkno >> L2DBWORD;
3853
3854	/* does nblocks fall on a 32-bit boundary ? */
3855	b = blkno & (DBWORD - 1);
3856	if (b) {
3857		/* mark a partial word allocated */
3858		dp->wmap[w] = dp->pmap[w] = cpu_to_le32(ONES >> b);
3859		w++;
3860	}
3861
3862	/* set the rest of the words in the page to allocated (ONES) */
3863	for (i = w; i < LPERDMAP; i++)
3864		dp->pmap[i] = dp->wmap[i] = cpu_to_le32(ONES);
3865
3866	/*
3867	 * init tree
3868	 */
3869      initTree:
3870	return (dbInitDmapTree(dp));
3871}
3872
3873
3874/*
3875 * NAME:	dbInitDmapTree()/ujfs_complete_dmap()
3876 *
3877 * FUNCTION:	initialize summary tree of the specified dmap:
3878 *
3879 *		at entry, bitmap of the dmap has been initialized;
3880 *
3881 * PARAMETERS:
3882 *	dp	- dmap to complete
3883 *	blkno	- starting block number for this dmap
3884 *	treemax	- will be filled in with max free for this dmap
3885 *
3886 * RETURNS:	max free string at the root of the tree
3887 */
3888static int dbInitDmapTree(struct dmap * dp)
3889{
3890	struct dmaptree *tp;
3891	s8 *cp;
3892	int i;
3893
3894	/* init fixed info of tree */
3895	tp = &dp->tree;
3896	tp->nleafs = cpu_to_le32(LPERDMAP);
3897	tp->l2nleafs = cpu_to_le32(L2LPERDMAP);
3898	tp->leafidx = cpu_to_le32(LEAFIND);
3899	tp->height = cpu_to_le32(4);
3900	tp->budmin = BUDMIN;
3901
3902	/* init each leaf from corresponding wmap word:
3903	 * note: leaf is set to NOFREE(-1) if all blocks of corresponding
3904	 * bitmap word are allocated.
3905	 */
3906	cp = tp->stree + le32_to_cpu(tp->leafidx);
3907	for (i = 0; i < LPERDMAP; i++)
3908		*cp++ = dbMaxBud((u8 *) & dp->wmap[i]);
3909
3910	/* build the dmap's binary buddy summary tree */
3911	return (dbInitTree(tp));
3912}
3913
3914
3915/*
3916 * NAME:	dbInitTree()/ujfs_adjtree()
3917 *
3918 * FUNCTION:	initialize binary buddy summary tree of a dmap or dmapctl.
3919 *
3920 *		at entry, the leaves of the tree has been initialized
3921 *		from corresponding bitmap word or root of summary tree
3922 *		of the child control page;
3923 *		configure binary buddy system at the leaf level, then
3924 *		bubble up the values of the leaf nodes up the tree.
3925 *
3926 * PARAMETERS:
3927 *	cp	- Pointer to the root of the tree
3928 *	l2leaves- Number of leaf nodes as a power of 2
3929 *	l2min	- Number of blocks that can be covered by a leaf
3930 *		  as a power of 2
3931 *
3932 * RETURNS: max free string at the root of the tree
3933 */
3934static int dbInitTree(struct dmaptree * dtp)
3935{
3936	int l2max, l2free, bsize, nextb, i;
3937	int child, parent, nparent;
3938	s8 *tp, *cp, *cp1;
3939
3940	tp = dtp->stree;
3941
3942	/* Determine the maximum free string possible for the leaves */
3943	l2max = le32_to_cpu(dtp->l2nleafs) + dtp->budmin;
3944
3945	/*
3946	 * configure the leaf levevl into binary buddy system
3947	 *
3948	 * Try to combine buddies starting with a buddy size of 1
3949	 * (i.e. two leaves). At a buddy size of 1 two buddy leaves
3950	 * can be combined if both buddies have a maximum free of l2min;
3951	 * the combination will result in the left-most buddy leaf having
3952	 * a maximum free of l2min+1.
3953	 * After processing all buddies for a given size, process buddies
3954	 * at the next higher buddy size (i.e. current size * 2) and
3955	 * the next maximum free (current free + 1).
3956	 * This continues until the maximum possible buddy combination
3957	 * yields maximum free.
3958	 */
3959	for (l2free = dtp->budmin, bsize = 1; l2free < l2max;
3960	     l2free++, bsize = nextb) {
3961		/* get next buddy size == current buddy pair size */
3962		nextb = bsize << 1;
3963
3964		/* scan each adjacent buddy pair at current buddy size */
3965		for (i = 0, cp = tp + le32_to_cpu(dtp->leafidx);
3966		     i < le32_to_cpu(dtp->nleafs);
3967		     i += nextb, cp += nextb) {
3968			/* coalesce if both adjacent buddies are max free */
3969			if (*cp == l2free && *(cp + bsize) == l2free) {
3970				*cp = l2free + 1;	/* left take right */
3971				*(cp + bsize) = -1;	/* right give left */
3972			}
3973		}
3974	}
3975
3976	/*
3977	 * bubble summary information of leaves up the tree.
3978	 *
3979	 * Starting at the leaf node level, the four nodes described by
3980	 * the higher level parent node are compared for a maximum free and
3981	 * this maximum becomes the value of the parent node.
3982	 * when all lower level nodes are processed in this fashion then
3983	 * move up to the next level (parent becomes a lower level node) and
3984	 * continue the process for that level.
3985	 */
3986	for (child = le32_to_cpu(dtp->leafidx),
3987	     nparent = le32_to_cpu(dtp->nleafs) >> 2;
3988	     nparent > 0; nparent >>= 2, child = parent) {
3989		/* get index of 1st node of parent level */
3990		parent = (child - 1) >> 2;
3991
3992		/* set the value of the parent node as the maximum
3993		 * of the four nodes of the current level.
3994		 */
3995		for (i = 0, cp = tp + child, cp1 = tp + parent;
3996		     i < nparent; i++, cp += 4, cp1++)
3997			*cp1 = TREEMAX(cp);
3998	}
3999
4000	return (*tp);
4001}
4002
4003
4004/*
4005 *	dbInitDmapCtl()
4006 *
4007 * function: initialize dmapctl page
4008 */
4009static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i)
4010{				/* start leaf index not covered by range */
4011	s8 *cp;
4012
4013	dcp->nleafs = cpu_to_le32(LPERCTL);
4014	dcp->l2nleafs = cpu_to_le32(L2LPERCTL);
4015	dcp->leafidx = cpu_to_le32(CTLLEAFIND);
4016	dcp->height = cpu_to_le32(5);
4017	dcp->budmin = L2BPERDMAP + L2LPERCTL * level;
4018
4019	/*
4020	 * initialize the leaves of current level that were not covered
4021	 * by the specified input block range (i.e. the leaves have no
4022	 * low level dmapctl or dmap).
4023	 */
4024	cp = &dcp->stree[CTLLEAFIND + i];
4025	for (; i < LPERCTL; i++)
4026		*cp++ = NOFREE;
4027
4028	/* build the dmap's binary buddy summary tree */
4029	return (dbInitTree((struct dmaptree *) dcp));
4030}
4031
4032
4033/*
4034 * NAME:	dbGetL2AGSize()/ujfs_getagl2size()
4035 *
4036 * FUNCTION:	Determine log2(allocation group size) from aggregate size
4037 *
4038 * PARAMETERS:
4039 *	nblocks	- Number of blocks in aggregate
4040 *
4041 * RETURNS: log2(allocation group size) in aggregate blocks
4042 */
4043static int dbGetL2AGSize(s64 nblocks)
4044{
4045	s64 sz;
4046	s64 m;
4047	int l2sz;
4048
4049	if (nblocks < BPERDMAP * MAXAG)
4050		return (L2BPERDMAP);
4051
4052	/* round up aggregate size to power of 2 */
4053	m = ((u64) 1 << (64 - 1));
4054	for (l2sz = 64; l2sz >= 0; l2sz--, m >>= 1) {
4055		if (m & nblocks)
4056			break;
4057	}
4058
4059	sz = (s64) 1 << l2sz;
4060	if (sz < nblocks)
4061		l2sz += 1;
4062
4063	/* agsize = roundupSize/max_number_of_ag */
4064	return (l2sz - L2MAXAG);
4065}
4066
4067
4068/*
4069 * NAME:	dbMapFileSizeToMapSize()
4070 *
4071 * FUNCTION:	compute number of blocks the block allocation map file
4072 *		can cover from the map file size;
4073 *
4074 * RETURNS:	Number of blocks which can be covered by this block map file;
4075 */
4076
4077/*
4078 * maximum number of map pages at each level including control pages
4079 */
4080#define MAXL0PAGES	(1 + LPERCTL)
4081#define MAXL1PAGES	(1 + LPERCTL * MAXL0PAGES)
4082
4083/*
4084 * convert number of map pages to the zero origin top dmapctl level
4085 */
4086#define BMAPPGTOLEV(npages)	\
4087	(((npages) <= 3 + MAXL0PAGES) ? 0 : \
4088	 ((npages) <= 2 + MAXL1PAGES) ? 1 : 2)
4089
4090s64 dbMapFileSizeToMapSize(struct inode * ipbmap)
4091{
4092	struct super_block *sb = ipbmap->i_sb;
4093	s64 nblocks;
4094	s64 npages, ndmaps;
4095	int level, i;
4096	int complete, factor;
4097
4098	nblocks = ipbmap->i_size >> JFS_SBI(sb)->l2bsize;
4099	npages = nblocks >> JFS_SBI(sb)->l2nbperpage;
4100	level = BMAPPGTOLEV(npages);
4101
4102	/* At each level, accumulate the number of dmap pages covered by
4103	 * the number of full child levels below it;
4104	 * repeat for the last incomplete child level.
4105	 */
4106	ndmaps = 0;
4107	npages--;		/* skip the first global control page */
4108	/* skip higher level control pages above top level covered by map */
4109	npages -= (2 - level);
4110	npages--;		/* skip top level's control page */
4111	for (i = level; i >= 0; i--) {
4112		factor =
4113		    (i == 2) ? MAXL1PAGES : ((i == 1) ? MAXL0PAGES : 1);
4114		complete = (u32) npages / factor;
4115		ndmaps += complete * ((i == 2) ? LPERCTL * LPERCTL :
4116				      ((i == 1) ? LPERCTL : 1));
4117
4118		/* pages in last/incomplete child */
4119		npages = (u32) npages % factor;
4120		/* skip incomplete child's level control page */
4121		npages--;
4122	}
4123
4124	/* convert the number of dmaps into the number of blocks
4125	 * which can be covered by the dmaps;
4126	 */
4127	nblocks = ndmaps << L2BPERDMAP;
4128
4129	return (nblocks);
4130}
4131