xref: /openbmc/linux/fs/jfs/jfs_dmap.c (revision d4c52c6a)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  *   Copyright (C) International Business Machines Corp., 2000-2004
4  *   Portions Copyright (C) Tino Reichardt, 2012
5  */
6 
7 #include <linux/fs.h>
8 #include <linux/slab.h>
9 #include "jfs_incore.h"
10 #include "jfs_superblock.h"
11 #include "jfs_dmap.h"
12 #include "jfs_imap.h"
13 #include "jfs_lock.h"
14 #include "jfs_metapage.h"
15 #include "jfs_debug.h"
16 #include "jfs_discard.h"
17 
18 /*
19  *	SERIALIZATION of the Block Allocation Map.
20  *
21  *	the working state of the block allocation map is accessed in
22  *	two directions:
23  *
24  *	1) allocation and free requests that start at the dmap
25  *	   level and move up through the dmap control pages (i.e.
26  *	   the vast majority of requests).
27  *
28  *	2) allocation requests that start at dmap control page
29  *	   level and work down towards the dmaps.
30  *
31  *	the serialization scheme used here is as follows.
32  *
33  *	requests which start at the bottom are serialized against each
34  *	other through buffers and each requests holds onto its buffers
35  *	as it works it way up from a single dmap to the required level
36  *	of dmap control page.
37  *	requests that start at the top are serialized against each other
38  *	and request that start from the bottom by the multiple read/single
39  *	write inode lock of the bmap inode. requests starting at the top
40  *	take this lock in write mode while request starting at the bottom
41  *	take the lock in read mode.  a single top-down request may proceed
42  *	exclusively while multiple bottoms-up requests may proceed
43  *	simultaneously (under the protection of busy buffers).
44  *
45  *	in addition to information found in dmaps and dmap control pages,
46  *	the working state of the block allocation map also includes read/
47  *	write information maintained in the bmap descriptor (i.e. total
48  *	free block count, allocation group level free block counts).
49  *	a single exclusive lock (BMAP_LOCK) is used to guard this information
50  *	in the face of multiple-bottoms up requests.
51  *	(lock ordering: IREAD_LOCK, BMAP_LOCK);
52  *
53  *	accesses to the persistent state of the block allocation map (limited
54  *	to the persistent bitmaps in dmaps) is guarded by (busy) buffers.
55  */
56 
57 #define BMAP_LOCK_INIT(bmp)	mutex_init(&bmp->db_bmaplock)
58 #define BMAP_LOCK(bmp)		mutex_lock(&bmp->db_bmaplock)
59 #define BMAP_UNLOCK(bmp)	mutex_unlock(&bmp->db_bmaplock)
60 
61 /*
62  * forward references
63  */
64 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
65 			int nblocks);
66 static void dbSplit(dmtree_t *tp, int leafno, int splitsz, int newval, bool is_ctl);
67 static int dbBackSplit(dmtree_t *tp, int leafno, bool is_ctl);
68 static int dbJoin(dmtree_t *tp, int leafno, int newval, bool is_ctl);
69 static void dbAdjTree(dmtree_t *tp, int leafno, int newval, bool is_ctl);
70 static int dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc,
71 		    int level);
72 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results);
73 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
74 		       int nblocks);
75 static int dbAllocNear(struct bmap * bmp, struct dmap * dp, s64 blkno,
76 		       int nblocks,
77 		       int l2nb, s64 * results);
78 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
79 		       int nblocks);
80 static int dbAllocDmapLev(struct bmap * bmp, struct dmap * dp, int nblocks,
81 			  int l2nb,
82 			  s64 * results);
83 static int dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb,
84 		     s64 * results);
85 static int dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno,
86 		      s64 * results);
87 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks);
88 static int dbFindBits(u32 word, int l2nb);
89 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno);
90 static int dbFindLeaf(dmtree_t *tp, int l2nb, int *leafidx, bool is_ctl);
91 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
92 		      int nblocks);
93 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
94 		      int nblocks);
95 static int dbMaxBud(u8 * cp);
96 static int blkstol2(s64 nb);
97 
98 static int cntlz(u32 value);
99 static int cnttz(u32 word);
100 
101 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
102 			 int nblocks);
103 static int dbInitDmap(struct dmap * dp, s64 blkno, int nblocks);
104 static int dbInitDmapTree(struct dmap * dp);
105 static int dbInitTree(struct dmaptree * dtp);
106 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i);
107 static int dbGetL2AGSize(s64 nblocks);
108 
109 /*
110  *	buddy table
111  *
112  * table used for determining buddy sizes within characters of
113  * dmap bitmap words.  the characters themselves serve as indexes
114  * into the table, with the table elements yielding the maximum
115  * binary buddy of free bits within the character.
116  */
117 static const s8 budtab[256] = {
118 	3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
119 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
120 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
121 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
122 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
123 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
124 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
125 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
126 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
127 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
128 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
129 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
130 	2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
131 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
132 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
133 	2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, -1
134 };
135 
136 /*
137  * NAME:	dbMount()
138  *
139  * FUNCTION:	initializate the block allocation map.
140  *
141  *		memory is allocated for the in-core bmap descriptor and
142  *		the in-core descriptor is initialized from disk.
143  *
144  * PARAMETERS:
145  *	ipbmap	- pointer to in-core inode for the block map.
146  *
147  * RETURN VALUES:
148  *	0	- success
149  *	-ENOMEM	- insufficient memory
150  *	-EIO	- i/o error
151  *	-EINVAL - wrong bmap data
152  */
153 int dbMount(struct inode *ipbmap)
154 {
155 	struct bmap *bmp;
156 	struct dbmap_disk *dbmp_le;
157 	struct metapage *mp;
158 	int i, err;
159 
160 	/*
161 	 * allocate/initialize the in-memory bmap descriptor
162 	 */
163 	/* allocate memory for the in-memory bmap descriptor */
164 	bmp = kmalloc(sizeof(struct bmap), GFP_KERNEL);
165 	if (bmp == NULL)
166 		return -ENOMEM;
167 
168 	/* read the on-disk bmap descriptor. */
169 	mp = read_metapage(ipbmap,
170 			   BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
171 			   PSIZE, 0);
172 	if (mp == NULL) {
173 		err = -EIO;
174 		goto err_kfree_bmp;
175 	}
176 
177 	/* copy the on-disk bmap descriptor to its in-memory version. */
178 	dbmp_le = (struct dbmap_disk *) mp->data;
179 	bmp->db_mapsize = le64_to_cpu(dbmp_le->dn_mapsize);
180 	bmp->db_nfree = le64_to_cpu(dbmp_le->dn_nfree);
181 
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 || bmp->db_numag > MAXAG) {
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 
241 err_release_metapage:
242 	release_metapage(mp);
243 err_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  */
265 int 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  */
287 int 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  */
355 int 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  */
448 int
449 dbUpdatePMap(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  */
628 int 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 */
679 unlock:
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
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  */
725 int 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 /*
903  * NAME:	dbReAlloc()
904  *
905  * FUNCTION:	attempt to extend a current allocation by a specified
906  *		number of blocks.
907  *
908  *		this routine attempts to satisfy the allocation request
909  *		by first trying to extend the existing allocation in
910  *		place by allocating the additional blocks as the blocks
911  *		immediately following the current allocation.  if these
912  *		blocks are not available, this routine will attempt to
913  *		allocate a new set of contiguous blocks large enough
914  *		to cover the existing allocation plus the additional
915  *		number of blocks required.
916  *
917  * PARAMETERS:
918  *	ip	    -  pointer to in-core inode requiring allocation.
919  *	blkno	    -  starting block of the current allocation.
920  *	nblocks	    -  number of contiguous blocks within the current
921  *		       allocation.
922  *	addnblocks  -  number of blocks to add to the allocation.
923  *	results	-      on successful return, set to the starting block number
924  *		       of the existing allocation if the existing allocation
925  *		       was extended in place or to a newly allocated contiguous
926  *		       range if the existing allocation could not be extended
927  *		       in place.
928  *
929  * RETURN VALUES:
930  *	0	- success
931  *	-ENOSPC	- insufficient disk resources
932  *	-EIO	- i/o error
933  */
934 int
935 dbReAlloc(struct inode *ip,
936 	  s64 blkno, s64 nblocks, s64 addnblocks, s64 * results)
937 {
938 	int rc;
939 
940 	/* try to extend the allocation in place.
941 	 */
942 	if ((rc = dbExtend(ip, blkno, nblocks, addnblocks)) == 0) {
943 		*results = blkno;
944 		return (0);
945 	} else {
946 		if (rc != -ENOSPC)
947 			return (rc);
948 	}
949 
950 	/* could not extend the allocation in place, so allocate a
951 	 * new set of blocks for the entire request (i.e. try to get
952 	 * a range of contiguous blocks large enough to cover the
953 	 * existing allocation plus the additional blocks.)
954 	 */
955 	return (dbAlloc
956 		(ip, blkno + nblocks - 1, addnblocks + nblocks, results));
957 }
958 
959 
960 /*
961  * NAME:	dbExtend()
962  *
963  * FUNCTION:	attempt to extend a current allocation by a specified
964  *		number of blocks.
965  *
966  *		this routine attempts to satisfy the allocation request
967  *		by first trying to extend the existing allocation in
968  *		place by allocating the additional blocks as the blocks
969  *		immediately following the current allocation.
970  *
971  * PARAMETERS:
972  *	ip	    -  pointer to in-core inode requiring allocation.
973  *	blkno	    -  starting block of the current allocation.
974  *	nblocks	    -  number of contiguous blocks within the current
975  *		       allocation.
976  *	addnblocks  -  number of blocks to add to the allocation.
977  *
978  * RETURN VALUES:
979  *	0	- success
980  *	-ENOSPC	- insufficient disk resources
981  *	-EIO	- i/o error
982  */
983 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks)
984 {
985 	struct jfs_sb_info *sbi = JFS_SBI(ip->i_sb);
986 	s64 lblkno, lastblkno, extblkno;
987 	uint rel_block;
988 	struct metapage *mp;
989 	struct dmap *dp;
990 	int rc;
991 	struct inode *ipbmap = sbi->ipbmap;
992 	struct bmap *bmp;
993 
994 	/*
995 	 * We don't want a non-aligned extent to cross a page boundary
996 	 */
997 	if (((rel_block = blkno & (sbi->nbperpage - 1))) &&
998 	    (rel_block + nblocks + addnblocks > sbi->nbperpage))
999 		return -ENOSPC;
1000 
1001 	/* get the last block of the current allocation */
1002 	lastblkno = blkno + nblocks - 1;
1003 
1004 	/* determine the block number of the block following
1005 	 * the existing allocation.
1006 	 */
1007 	extblkno = lastblkno + 1;
1008 
1009 	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
1010 
1011 	/* better be within the file system */
1012 	bmp = sbi->bmap;
1013 	if (lastblkno < 0 || lastblkno >= bmp->db_mapsize) {
1014 		IREAD_UNLOCK(ipbmap);
1015 		jfs_error(ip->i_sb, "the block is outside the filesystem\n");
1016 		return -EIO;
1017 	}
1018 
1019 	/* we'll attempt to extend the current allocation in place by
1020 	 * allocating the additional blocks as the blocks immediately
1021 	 * following the current allocation.  we only try to extend the
1022 	 * current allocation in place if the number of additional blocks
1023 	 * can fit into a dmap, the last block of the current allocation
1024 	 * is not the last block of the file system, and the start of the
1025 	 * inplace extension is not on an allocation group boundary.
1026 	 */
1027 	if (addnblocks > BPERDMAP || extblkno >= bmp->db_mapsize ||
1028 	    (extblkno & (bmp->db_agsize - 1)) == 0) {
1029 		IREAD_UNLOCK(ipbmap);
1030 		return -ENOSPC;
1031 	}
1032 
1033 	/* get the buffer for the dmap containing the first block
1034 	 * of the extension.
1035 	 */
1036 	lblkno = BLKTODMAP(extblkno, bmp->db_l2nbperpage);
1037 	mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
1038 	if (mp == NULL) {
1039 		IREAD_UNLOCK(ipbmap);
1040 		return -EIO;
1041 	}
1042 
1043 	dp = (struct dmap *) mp->data;
1044 
1045 	/* try to allocate the blocks immediately following the
1046 	 * current allocation.
1047 	 */
1048 	rc = dbAllocNext(bmp, dp, extblkno, (int) addnblocks);
1049 
1050 	IREAD_UNLOCK(ipbmap);
1051 
1052 	/* were we successful ? */
1053 	if (rc == 0)
1054 		write_metapage(mp);
1055 	else
1056 		/* we were not successful */
1057 		release_metapage(mp);
1058 
1059 	return (rc);
1060 }
1061 
1062 
1063 /*
1064  * NAME:	dbAllocNext()
1065  *
1066  * FUNCTION:	attempt to allocate the blocks of the specified block
1067  *		range within a dmap.
1068  *
1069  * PARAMETERS:
1070  *	bmp	-  pointer to bmap descriptor
1071  *	dp	-  pointer to dmap.
1072  *	blkno	-  starting block number of the range.
1073  *	nblocks	-  number of contiguous free blocks of the range.
1074  *
1075  * RETURN VALUES:
1076  *	0	- success
1077  *	-ENOSPC	- insufficient disk resources
1078  *	-EIO	- i/o error
1079  *
1080  * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1081  */
1082 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
1083 		       int nblocks)
1084 {
1085 	int dbitno, word, rembits, nb, nwords, wbitno, nw;
1086 	int l2size;
1087 	s8 *leaf;
1088 	u32 mask;
1089 
1090 	if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1091 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1092 		return -EIO;
1093 	}
1094 
1095 	/* pick up a pointer to the leaves of the dmap tree.
1096 	 */
1097 	leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1098 
1099 	/* determine the bit number and word within the dmap of the
1100 	 * starting block.
1101 	 */
1102 	dbitno = blkno & (BPERDMAP - 1);
1103 	word = dbitno >> L2DBWORD;
1104 
1105 	/* check if the specified block range is contained within
1106 	 * this dmap.
1107 	 */
1108 	if (dbitno + nblocks > BPERDMAP)
1109 		return -ENOSPC;
1110 
1111 	/* check if the starting leaf indicates that anything
1112 	 * is free.
1113 	 */
1114 	if (leaf[word] == NOFREE)
1115 		return -ENOSPC;
1116 
1117 	/* check the dmaps words corresponding to block range to see
1118 	 * if the block range is free.  not all bits of the first and
1119 	 * last words may be contained within the block range.  if this
1120 	 * is the case, we'll work against those words (i.e. partial first
1121 	 * and/or last) on an individual basis (a single pass) and examine
1122 	 * the actual bits to determine if they are free.  a single pass
1123 	 * will be used for all dmap words fully contained within the
1124 	 * specified range.  within this pass, the leaves of the dmap
1125 	 * tree will be examined to determine if the blocks are free. a
1126 	 * single leaf may describe the free space of multiple dmap
1127 	 * words, so we may visit only a subset of the actual leaves
1128 	 * corresponding to the dmap words of the block range.
1129 	 */
1130 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
1131 		/* determine the bit number within the word and
1132 		 * the number of bits within the word.
1133 		 */
1134 		wbitno = dbitno & (DBWORD - 1);
1135 		nb = min(rembits, DBWORD - wbitno);
1136 
1137 		/* check if only part of the word is to be examined.
1138 		 */
1139 		if (nb < DBWORD) {
1140 			/* check if the bits are free.
1141 			 */
1142 			mask = (ONES << (DBWORD - nb) >> wbitno);
1143 			if ((mask & ~le32_to_cpu(dp->wmap[word])) != mask)
1144 				return -ENOSPC;
1145 
1146 			word += 1;
1147 		} else {
1148 			/* one or more dmap words are fully contained
1149 			 * within the block range.  determine how many
1150 			 * words and how many bits.
1151 			 */
1152 			nwords = rembits >> L2DBWORD;
1153 			nb = nwords << L2DBWORD;
1154 
1155 			/* now examine the appropriate leaves to determine
1156 			 * if the blocks are free.
1157 			 */
1158 			while (nwords > 0) {
1159 				/* does the leaf describe any free space ?
1160 				 */
1161 				if (leaf[word] < BUDMIN)
1162 					return -ENOSPC;
1163 
1164 				/* determine the l2 number of bits provided
1165 				 * by this leaf.
1166 				 */
1167 				l2size =
1168 				    min_t(int, leaf[word], NLSTOL2BSZ(nwords));
1169 
1170 				/* determine how many words were handled.
1171 				 */
1172 				nw = BUDSIZE(l2size, BUDMIN);
1173 
1174 				nwords -= nw;
1175 				word += nw;
1176 			}
1177 		}
1178 	}
1179 
1180 	/* allocate the blocks.
1181 	 */
1182 	return (dbAllocDmap(bmp, dp, blkno, nblocks));
1183 }
1184 
1185 
1186 /*
1187  * NAME:	dbAllocNear()
1188  *
1189  * FUNCTION:	attempt to allocate a number of contiguous free blocks near
1190  *		a specified block (hint) within a dmap.
1191  *
1192  *		starting with the dmap leaf that covers the hint, we'll
1193  *		check the next four contiguous leaves for sufficient free
1194  *		space.  if sufficient free space is found, we'll allocate
1195  *		the desired free space.
1196  *
1197  * PARAMETERS:
1198  *	bmp	-  pointer to bmap descriptor
1199  *	dp	-  pointer to dmap.
1200  *	blkno	-  block number to allocate near.
1201  *	nblocks	-  actual number of contiguous free blocks desired.
1202  *	l2nb	-  log2 number of contiguous free blocks desired.
1203  *	results	-  on successful return, set to the starting block number
1204  *		   of the newly allocated range.
1205  *
1206  * RETURN VALUES:
1207  *	0	- success
1208  *	-ENOSPC	- insufficient disk resources
1209  *	-EIO	- i/o error
1210  *
1211  * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1212  */
1213 static int
1214 dbAllocNear(struct bmap * bmp,
1215 	    struct dmap * dp, s64 blkno, int nblocks, int l2nb, s64 * results)
1216 {
1217 	int word, lword, rc;
1218 	s8 *leaf;
1219 
1220 	if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1221 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1222 		return -EIO;
1223 	}
1224 
1225 	leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1226 
1227 	/* determine the word within the dmap that holds the hint
1228 	 * (i.e. blkno).  also, determine the last word in the dmap
1229 	 * that we'll include in our examination.
1230 	 */
1231 	word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
1232 	lword = min(word + 4, LPERDMAP);
1233 
1234 	/* examine the leaves for sufficient free space.
1235 	 */
1236 	for (; word < lword; word++) {
1237 		/* does the leaf describe sufficient free space ?
1238 		 */
1239 		if (leaf[word] < l2nb)
1240 			continue;
1241 
1242 		/* determine the block number within the file system
1243 		 * of the first block described by this dmap word.
1244 		 */
1245 		blkno = le64_to_cpu(dp->start) + (word << L2DBWORD);
1246 
1247 		/* if not all bits of the dmap word are free, get the
1248 		 * starting bit number within the dmap word of the required
1249 		 * string of free bits and adjust the block number with the
1250 		 * value.
1251 		 */
1252 		if (leaf[word] < BUDMIN)
1253 			blkno +=
1254 			    dbFindBits(le32_to_cpu(dp->wmap[word]), l2nb);
1255 
1256 		/* allocate the blocks.
1257 		 */
1258 		if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1259 			*results = blkno;
1260 
1261 		return (rc);
1262 	}
1263 
1264 	return -ENOSPC;
1265 }
1266 
1267 
1268 /*
1269  * NAME:	dbAllocAG()
1270  *
1271  * FUNCTION:	attempt to allocate the specified number of contiguous
1272  *		free blocks within the specified allocation group.
1273  *
1274  *		unless the allocation group size is equal to the number
1275  *		of blocks per dmap, the dmap control pages will be used to
1276  *		find the required free space, if available.  we start the
1277  *		search at the highest dmap control page level which
1278  *		distinctly describes the allocation group's free space
1279  *		(i.e. the highest level at which the allocation group's
1280  *		free space is not mixed in with that of any other group).
1281  *		in addition, we start the search within this level at a
1282  *		height of the dmapctl dmtree at which the nodes distinctly
1283  *		describe the allocation group's free space.  at this height,
1284  *		the allocation group's free space may be represented by 1
1285  *		or two sub-trees, depending on the allocation group size.
1286  *		we search the top nodes of these subtrees left to right for
1287  *		sufficient free space.  if sufficient free space is found,
1288  *		the subtree is searched to find the leftmost leaf that
1289  *		has free space.  once we have made it to the leaf, we
1290  *		move the search to the next lower level dmap control page
1291  *		corresponding to this leaf.  we continue down the dmap control
1292  *		pages until we find the dmap that contains or starts the
1293  *		sufficient free space and we allocate at this dmap.
1294  *
1295  *		if the allocation group size is equal to the dmap size,
1296  *		we'll start at the dmap corresponding to the allocation
1297  *		group and attempt the allocation at this level.
1298  *
1299  *		the dmap control page search is also not performed if the
1300  *		allocation group is completely free and we go to the first
1301  *		dmap of the allocation group to do the allocation.  this is
1302  *		done because the allocation group may be part (not the first
1303  *		part) of a larger binary buddy system, causing the dmap
1304  *		control pages to indicate no free space (NOFREE) within
1305  *		the allocation group.
1306  *
1307  * PARAMETERS:
1308  *	bmp	-  pointer to bmap descriptor
1309  *	agno	- allocation group number.
1310  *	nblocks	-  actual number of contiguous free blocks desired.
1311  *	l2nb	-  log2 number of contiguous free blocks desired.
1312  *	results	-  on successful return, set to the starting block number
1313  *		   of the newly allocated range.
1314  *
1315  * RETURN VALUES:
1316  *	0	- success
1317  *	-ENOSPC	- insufficient disk resources
1318  *	-EIO	- i/o error
1319  *
1320  * note: IWRITE_LOCK(ipmap) held on entry/exit;
1321  */
1322 static int
1323 dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, s64 * results)
1324 {
1325 	struct metapage *mp;
1326 	struct dmapctl *dcp;
1327 	int rc, ti, i, k, m, n, agperlev;
1328 	s64 blkno, lblkno;
1329 	int budmin;
1330 
1331 	/* allocation request should not be for more than the
1332 	 * allocation group size.
1333 	 */
1334 	if (l2nb > bmp->db_agl2size) {
1335 		jfs_error(bmp->db_ipbmap->i_sb,
1336 			  "allocation request is larger than the allocation group size\n");
1337 		return -EIO;
1338 	}
1339 
1340 	/* determine the starting block number of the allocation
1341 	 * group.
1342 	 */
1343 	blkno = (s64) agno << bmp->db_agl2size;
1344 
1345 	/* check if the allocation group size is the minimum allocation
1346 	 * group size or if the allocation group is completely free. if
1347 	 * the allocation group size is the minimum size of BPERDMAP (i.e.
1348 	 * 1 dmap), there is no need to search the dmap control page (below)
1349 	 * that fully describes the allocation group since the allocation
1350 	 * group is already fully described by a dmap.  in this case, we
1351 	 * just call dbAllocCtl() to search the dmap tree and allocate the
1352 	 * required space if available.
1353 	 *
1354 	 * if the allocation group is completely free, dbAllocCtl() is
1355 	 * also called to allocate the required space.  this is done for
1356 	 * two reasons.  first, it makes no sense searching the dmap control
1357 	 * pages for free space when we know that free space exists.  second,
1358 	 * the dmap control pages may indicate that the allocation group
1359 	 * has no free space if the allocation group is part (not the first
1360 	 * part) of a larger binary buddy system.
1361 	 */
1362 	if (bmp->db_agsize == BPERDMAP
1363 	    || bmp->db_agfree[agno] == bmp->db_agsize) {
1364 		rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1365 		if ((rc == -ENOSPC) &&
1366 		    (bmp->db_agfree[agno] == bmp->db_agsize)) {
1367 			printk(KERN_ERR "blkno = %Lx, blocks = %Lx\n",
1368 			       (unsigned long long) blkno,
1369 			       (unsigned long long) nblocks);
1370 			jfs_error(bmp->db_ipbmap->i_sb,
1371 				  "dbAllocCtl failed in free AG\n");
1372 		}
1373 		return (rc);
1374 	}
1375 
1376 	/* the buffer for the dmap control page that fully describes the
1377 	 * allocation group.
1378 	 */
1379 	lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, bmp->db_aglevel);
1380 	mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1381 	if (mp == NULL)
1382 		return -EIO;
1383 	dcp = (struct dmapctl *) mp->data;
1384 	budmin = dcp->budmin;
1385 
1386 	if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1387 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
1388 		release_metapage(mp);
1389 		return -EIO;
1390 	}
1391 
1392 	/* search the subtree(s) of the dmap control page that describes
1393 	 * the allocation group, looking for sufficient free space.  to begin,
1394 	 * determine how many allocation groups are represented in a dmap
1395 	 * control page at the control page level (i.e. L0, L1, L2) that
1396 	 * fully describes an allocation group. next, determine the starting
1397 	 * tree index of this allocation group within the control page.
1398 	 */
1399 	agperlev =
1400 	    (1 << (L2LPERCTL - (bmp->db_agheight << 1))) / bmp->db_agwidth;
1401 	ti = bmp->db_agstart + bmp->db_agwidth * (agno & (agperlev - 1));
1402 
1403 	/* dmap control page trees fan-out by 4 and a single allocation
1404 	 * group may be described by 1 or 2 subtrees within the ag level
1405 	 * dmap control page, depending upon the ag size. examine the ag's
1406 	 * subtrees for sufficient free space, starting with the leftmost
1407 	 * subtree.
1408 	 */
1409 	for (i = 0; i < bmp->db_agwidth; i++, ti++) {
1410 		/* is there sufficient free space ?
1411 		 */
1412 		if (l2nb > dcp->stree[ti])
1413 			continue;
1414 
1415 		/* sufficient free space found in a subtree. now search down
1416 		 * the subtree to find the leftmost leaf that describes this
1417 		 * free space.
1418 		 */
1419 		for (k = bmp->db_agheight; k > 0; k--) {
1420 			for (n = 0, m = (ti << 2) + 1; n < 4; n++) {
1421 				if (l2nb <= dcp->stree[m + n]) {
1422 					ti = m + n;
1423 					break;
1424 				}
1425 			}
1426 			if (n == 4) {
1427 				jfs_error(bmp->db_ipbmap->i_sb,
1428 					  "failed descending stree\n");
1429 				release_metapage(mp);
1430 				return -EIO;
1431 			}
1432 		}
1433 
1434 		/* determine the block number within the file system
1435 		 * that corresponds to this leaf.
1436 		 */
1437 		if (bmp->db_aglevel == 2)
1438 			blkno = 0;
1439 		else if (bmp->db_aglevel == 1)
1440 			blkno &= ~(MAXL1SIZE - 1);
1441 		else		/* bmp->db_aglevel == 0 */
1442 			blkno &= ~(MAXL0SIZE - 1);
1443 
1444 		blkno +=
1445 		    ((s64) (ti - le32_to_cpu(dcp->leafidx))) << budmin;
1446 
1447 		/* release the buffer in preparation for going down
1448 		 * the next level of dmap control pages.
1449 		 */
1450 		release_metapage(mp);
1451 
1452 		/* check if we need to continue to search down the lower
1453 		 * level dmap control pages.  we need to if the number of
1454 		 * blocks required is less than maximum number of blocks
1455 		 * described at the next lower level.
1456 		 */
1457 		if (l2nb < budmin) {
1458 
1459 			/* search the lower level dmap control pages to get
1460 			 * the starting block number of the dmap that
1461 			 * contains or starts off the free space.
1462 			 */
1463 			if ((rc =
1464 			     dbFindCtl(bmp, l2nb, bmp->db_aglevel - 1,
1465 				       &blkno))) {
1466 				if (rc == -ENOSPC) {
1467 					jfs_error(bmp->db_ipbmap->i_sb,
1468 						  "control page inconsistent\n");
1469 					return -EIO;
1470 				}
1471 				return (rc);
1472 			}
1473 		}
1474 
1475 		/* allocate the blocks.
1476 		 */
1477 		rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1478 		if (rc == -ENOSPC) {
1479 			jfs_error(bmp->db_ipbmap->i_sb,
1480 				  "unable to allocate blocks\n");
1481 			rc = -EIO;
1482 		}
1483 		return (rc);
1484 	}
1485 
1486 	/* no space in the allocation group.  release the buffer and
1487 	 * return -ENOSPC.
1488 	 */
1489 	release_metapage(mp);
1490 
1491 	return -ENOSPC;
1492 }
1493 
1494 
1495 /*
1496  * NAME:	dbAllocAny()
1497  *
1498  * FUNCTION:	attempt to allocate the specified number of contiguous
1499  *		free blocks anywhere in the file system.
1500  *
1501  *		dbAllocAny() attempts to find the sufficient free space by
1502  *		searching down the dmap control pages, starting with the
1503  *		highest level (i.e. L0, L1, L2) control page.  if free space
1504  *		large enough to satisfy the desired free space is found, the
1505  *		desired free space is allocated.
1506  *
1507  * PARAMETERS:
1508  *	bmp	-  pointer to bmap descriptor
1509  *	nblocks	 -  actual number of contiguous free blocks desired.
1510  *	l2nb	 -  log2 number of contiguous free blocks desired.
1511  *	results	-  on successful return, set to the starting block number
1512  *		   of the newly allocated range.
1513  *
1514  * RETURN VALUES:
1515  *	0	- success
1516  *	-ENOSPC	- insufficient disk resources
1517  *	-EIO	- i/o error
1518  *
1519  * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1520  */
1521 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results)
1522 {
1523 	int rc;
1524 	s64 blkno = 0;
1525 
1526 	/* starting with the top level dmap control page, search
1527 	 * down the dmap control levels for sufficient free space.
1528 	 * if free space is found, dbFindCtl() returns the starting
1529 	 * block number of the dmap that contains or starts off the
1530 	 * range of free space.
1531 	 */
1532 	if ((rc = dbFindCtl(bmp, l2nb, bmp->db_maxlevel, &blkno)))
1533 		return (rc);
1534 
1535 	/* allocate the blocks.
1536 	 */
1537 	rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1538 	if (rc == -ENOSPC) {
1539 		jfs_error(bmp->db_ipbmap->i_sb, "unable to allocate blocks\n");
1540 		return -EIO;
1541 	}
1542 	return (rc);
1543 }
1544 
1545 
1546 /*
1547  * NAME:	dbDiscardAG()
1548  *
1549  * FUNCTION:	attempt to discard (TRIM) all free blocks of specific AG
1550  *
1551  *		algorithm:
1552  *		1) allocate blocks, as large as possible and save them
1553  *		   while holding IWRITE_LOCK on ipbmap
1554  *		2) trim all these saved block/length values
1555  *		3) mark the blocks free again
1556  *
1557  *		benefit:
1558  *		- we work only on one ag at some time, minimizing how long we
1559  *		  need to lock ipbmap
1560  *		- reading / writing the fs is possible most time, even on
1561  *		  trimming
1562  *
1563  *		downside:
1564  *		- we write two times to the dmapctl and dmap pages
1565  *		- but for me, this seems the best way, better ideas?
1566  *		/TR 2012
1567  *
1568  * PARAMETERS:
1569  *	ip	- pointer to in-core inode
1570  *	agno	- ag to trim
1571  *	minlen	- minimum value of contiguous blocks
1572  *
1573  * RETURN VALUES:
1574  *	s64	- actual number of blocks trimmed
1575  */
1576 s64 dbDiscardAG(struct inode *ip, int agno, s64 minlen)
1577 {
1578 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
1579 	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
1580 	s64 nblocks, blkno;
1581 	u64 trimmed = 0;
1582 	int rc, l2nb;
1583 	struct super_block *sb = ipbmap->i_sb;
1584 
1585 	struct range2trim {
1586 		u64 blkno;
1587 		u64 nblocks;
1588 	} *totrim, *tt;
1589 
1590 	/* max blkno / nblocks pairs to trim */
1591 	int count = 0, range_cnt;
1592 	u64 max_ranges;
1593 
1594 	/* prevent others from writing new stuff here, while trimming */
1595 	IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
1596 
1597 	nblocks = bmp->db_agfree[agno];
1598 	max_ranges = nblocks;
1599 	do_div(max_ranges, minlen);
1600 	range_cnt = min_t(u64, max_ranges + 1, 32 * 1024);
1601 	totrim = kmalloc_array(range_cnt, sizeof(struct range2trim), GFP_NOFS);
1602 	if (totrim == NULL) {
1603 		jfs_error(bmp->db_ipbmap->i_sb, "no memory for trim array\n");
1604 		IWRITE_UNLOCK(ipbmap);
1605 		return 0;
1606 	}
1607 
1608 	tt = totrim;
1609 	while (nblocks >= minlen) {
1610 		l2nb = BLKSTOL2(nblocks);
1611 
1612 		/* 0 = okay, -EIO = fatal, -ENOSPC -> try smaller block */
1613 		rc = dbAllocAG(bmp, agno, nblocks, l2nb, &blkno);
1614 		if (rc == 0) {
1615 			tt->blkno = blkno;
1616 			tt->nblocks = nblocks;
1617 			tt++; count++;
1618 
1619 			/* the whole ag is free, trim now */
1620 			if (bmp->db_agfree[agno] == 0)
1621 				break;
1622 
1623 			/* give a hint for the next while */
1624 			nblocks = bmp->db_agfree[agno];
1625 			continue;
1626 		} else if (rc == -ENOSPC) {
1627 			/* search for next smaller log2 block */
1628 			l2nb = BLKSTOL2(nblocks) - 1;
1629 			if (unlikely(l2nb < 0))
1630 				break;
1631 			nblocks = 1LL << l2nb;
1632 		} else {
1633 			/* Trim any already allocated blocks */
1634 			jfs_error(bmp->db_ipbmap->i_sb, "-EIO\n");
1635 			break;
1636 		}
1637 
1638 		/* check, if our trim array is full */
1639 		if (unlikely(count >= range_cnt - 1))
1640 			break;
1641 	}
1642 	IWRITE_UNLOCK(ipbmap);
1643 
1644 	tt->nblocks = 0; /* mark the current end */
1645 	for (tt = totrim; tt->nblocks != 0; tt++) {
1646 		/* when mounted with online discard, dbFree() will
1647 		 * call jfs_issue_discard() itself */
1648 		if (!(JFS_SBI(sb)->flag & JFS_DISCARD))
1649 			jfs_issue_discard(ip, tt->blkno, tt->nblocks);
1650 		dbFree(ip, tt->blkno, tt->nblocks);
1651 		trimmed += tt->nblocks;
1652 	}
1653 	kfree(totrim);
1654 
1655 	return trimmed;
1656 }
1657 
1658 /*
1659  * NAME:	dbFindCtl()
1660  *
1661  * FUNCTION:	starting at a specified dmap control page level and block
1662  *		number, search down the dmap control levels for a range of
1663  *		contiguous free blocks large enough to satisfy an allocation
1664  *		request for the specified number of free blocks.
1665  *
1666  *		if sufficient contiguous free blocks are found, this routine
1667  *		returns the starting block number within a dmap page that
1668  *		contains or starts a range of contiqious free blocks that
1669  *		is sufficient in size.
1670  *
1671  * PARAMETERS:
1672  *	bmp	-  pointer to bmap descriptor
1673  *	level	-  starting dmap control page level.
1674  *	l2nb	-  log2 number of contiguous free blocks desired.
1675  *	*blkno	-  on entry, starting block number for conducting the search.
1676  *		   on successful return, the first block within a dmap page
1677  *		   that contains or starts a range of contiguous free blocks.
1678  *
1679  * RETURN VALUES:
1680  *	0	- success
1681  *	-ENOSPC	- insufficient disk resources
1682  *	-EIO	- i/o error
1683  *
1684  * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1685  */
1686 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno)
1687 {
1688 	int rc, leafidx, lev;
1689 	s64 b, lblkno;
1690 	struct dmapctl *dcp;
1691 	int budmin;
1692 	struct metapage *mp;
1693 
1694 	/* starting at the specified dmap control page level and block
1695 	 * number, search down the dmap control levels for the starting
1696 	 * block number of a dmap page that contains or starts off
1697 	 * sufficient free blocks.
1698 	 */
1699 	for (lev = level, b = *blkno; lev >= 0; lev--) {
1700 		/* get the buffer of the dmap control page for the block
1701 		 * number and level (i.e. L0, L1, L2).
1702 		 */
1703 		lblkno = BLKTOCTL(b, bmp->db_l2nbperpage, lev);
1704 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1705 		if (mp == NULL)
1706 			return -EIO;
1707 		dcp = (struct dmapctl *) mp->data;
1708 		budmin = dcp->budmin;
1709 
1710 		if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1711 			jfs_error(bmp->db_ipbmap->i_sb,
1712 				  "Corrupt dmapctl page\n");
1713 			release_metapage(mp);
1714 			return -EIO;
1715 		}
1716 
1717 		/* search the tree within the dmap control page for
1718 		 * sufficient free space.  if sufficient free space is found,
1719 		 * dbFindLeaf() returns the index of the leaf at which
1720 		 * free space was found.
1721 		 */
1722 		rc = dbFindLeaf((dmtree_t *) dcp, l2nb, &leafidx, true);
1723 
1724 		/* release the buffer.
1725 		 */
1726 		release_metapage(mp);
1727 
1728 		/* space found ?
1729 		 */
1730 		if (rc) {
1731 			if (lev != level) {
1732 				jfs_error(bmp->db_ipbmap->i_sb,
1733 					  "dmap inconsistent\n");
1734 				return -EIO;
1735 			}
1736 			return -ENOSPC;
1737 		}
1738 
1739 		/* adjust the block number to reflect the location within
1740 		 * the dmap control page (i.e. the leaf) at which free
1741 		 * space was found.
1742 		 */
1743 		b += (((s64) leafidx) << budmin);
1744 
1745 		/* we stop the search at this dmap control page level if
1746 		 * the number of blocks required is greater than or equal
1747 		 * to the maximum number of blocks described at the next
1748 		 * (lower) level.
1749 		 */
1750 		if (l2nb >= budmin)
1751 			break;
1752 	}
1753 
1754 	*blkno = b;
1755 	return (0);
1756 }
1757 
1758 
1759 /*
1760  * NAME:	dbAllocCtl()
1761  *
1762  * FUNCTION:	attempt to allocate a specified number of contiguous
1763  *		blocks starting within a specific dmap.
1764  *
1765  *		this routine is called by higher level routines that search
1766  *		the dmap control pages above the actual dmaps for contiguous
1767  *		free space.  the result of successful searches by these
1768  *		routines are the starting block numbers within dmaps, with
1769  *		the dmaps themselves containing the desired contiguous free
1770  *		space or starting a contiguous free space of desired size
1771  *		that is made up of the blocks of one or more dmaps. these
1772  *		calls should not fail due to insufficent resources.
1773  *
1774  *		this routine is called in some cases where it is not known
1775  *		whether it will fail due to insufficient resources.  more
1776  *		specifically, this occurs when allocating from an allocation
1777  *		group whose size is equal to the number of blocks per dmap.
1778  *		in this case, the dmap control pages are not examined prior
1779  *		to calling this routine (to save pathlength) and the call
1780  *		might fail.
1781  *
1782  *		for a request size that fits within a dmap, this routine relies
1783  *		upon the dmap's dmtree to find the requested contiguous free
1784  *		space.  for request sizes that are larger than a dmap, the
1785  *		requested free space will start at the first block of the
1786  *		first dmap (i.e. blkno).
1787  *
1788  * PARAMETERS:
1789  *	bmp	-  pointer to bmap descriptor
1790  *	nblocks	 -  actual number of contiguous free blocks to allocate.
1791  *	l2nb	 -  log2 number of contiguous free blocks to allocate.
1792  *	blkno	 -  starting block number of the dmap to start the allocation
1793  *		    from.
1794  *	results	-  on successful return, set to the starting block number
1795  *		   of the newly allocated range.
1796  *
1797  * RETURN VALUES:
1798  *	0	- success
1799  *	-ENOSPC	- insufficient disk resources
1800  *	-EIO	- i/o error
1801  *
1802  * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1803  */
1804 static int
1805 dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, s64 * results)
1806 {
1807 	int rc, nb;
1808 	s64 b, lblkno, n;
1809 	struct metapage *mp;
1810 	struct dmap *dp;
1811 
1812 	/* check if the allocation request is confined to a single dmap.
1813 	 */
1814 	if (l2nb <= L2BPERDMAP) {
1815 		/* get the buffer for the dmap.
1816 		 */
1817 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
1818 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1819 		if (mp == NULL)
1820 			return -EIO;
1821 		dp = (struct dmap *) mp->data;
1822 
1823 		/* try to allocate the blocks.
1824 		 */
1825 		rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results);
1826 		if (rc == 0)
1827 			mark_metapage_dirty(mp);
1828 
1829 		release_metapage(mp);
1830 
1831 		return (rc);
1832 	}
1833 
1834 	/* allocation request involving multiple dmaps. it must start on
1835 	 * a dmap boundary.
1836 	 */
1837 	assert((blkno & (BPERDMAP - 1)) == 0);
1838 
1839 	/* allocate the blocks dmap by dmap.
1840 	 */
1841 	for (n = nblocks, b = blkno; n > 0; n -= nb, b += nb) {
1842 		/* get the buffer for the dmap.
1843 		 */
1844 		lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1845 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1846 		if (mp == NULL) {
1847 			rc = -EIO;
1848 			goto backout;
1849 		}
1850 		dp = (struct dmap *) mp->data;
1851 
1852 		/* the dmap better be all free.
1853 		 */
1854 		if (dp->tree.stree[ROOT] != L2BPERDMAP) {
1855 			release_metapage(mp);
1856 			jfs_error(bmp->db_ipbmap->i_sb,
1857 				  "the dmap is not all free\n");
1858 			rc = -EIO;
1859 			goto backout;
1860 		}
1861 
1862 		/* determine how many blocks to allocate from this dmap.
1863 		 */
1864 		nb = min_t(s64, n, BPERDMAP);
1865 
1866 		/* allocate the blocks from the dmap.
1867 		 */
1868 		if ((rc = dbAllocDmap(bmp, dp, b, nb))) {
1869 			release_metapage(mp);
1870 			goto backout;
1871 		}
1872 
1873 		/* write the buffer.
1874 		 */
1875 		write_metapage(mp);
1876 	}
1877 
1878 	/* set the results (starting block number) and return.
1879 	 */
1880 	*results = blkno;
1881 	return (0);
1882 
1883 	/* something failed in handling an allocation request involving
1884 	 * multiple dmaps.  we'll try to clean up by backing out any
1885 	 * allocation that has already happened for this request.  if
1886 	 * we fail in backing out the allocation, we'll mark the file
1887 	 * system to indicate that blocks have been leaked.
1888 	 */
1889       backout:
1890 
1891 	/* try to backout the allocations dmap by dmap.
1892 	 */
1893 	for (n = nblocks - n, b = blkno; n > 0;
1894 	     n -= BPERDMAP, b += BPERDMAP) {
1895 		/* get the buffer for this dmap.
1896 		 */
1897 		lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1898 		mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1899 		if (mp == NULL) {
1900 			/* could not back out.  mark the file system
1901 			 * to indicate that we have leaked blocks.
1902 			 */
1903 			jfs_error(bmp->db_ipbmap->i_sb,
1904 				  "I/O Error: Block Leakage\n");
1905 			continue;
1906 		}
1907 		dp = (struct dmap *) mp->data;
1908 
1909 		/* free the blocks is this dmap.
1910 		 */
1911 		if (dbFreeDmap(bmp, dp, b, BPERDMAP)) {
1912 			/* could not back out.  mark the file system
1913 			 * to indicate that we have leaked blocks.
1914 			 */
1915 			release_metapage(mp);
1916 			jfs_error(bmp->db_ipbmap->i_sb, "Block Leakage\n");
1917 			continue;
1918 		}
1919 
1920 		/* write the buffer.
1921 		 */
1922 		write_metapage(mp);
1923 	}
1924 
1925 	return (rc);
1926 }
1927 
1928 
1929 /*
1930  * NAME:	dbAllocDmapLev()
1931  *
1932  * FUNCTION:	attempt to allocate a specified number of contiguous blocks
1933  *		from a specified dmap.
1934  *
1935  *		this routine checks if the contiguous blocks are available.
1936  *		if so, nblocks of blocks are allocated; otherwise, ENOSPC is
1937  *		returned.
1938  *
1939  * PARAMETERS:
1940  *	mp	-  pointer to bmap descriptor
1941  *	dp	-  pointer to dmap to attempt to allocate blocks from.
1942  *	l2nb	-  log2 number of contiguous block desired.
1943  *	nblocks	-  actual number of contiguous block desired.
1944  *	results	-  on successful return, set to the starting block number
1945  *		   of the newly allocated range.
1946  *
1947  * RETURN VALUES:
1948  *	0	- success
1949  *	-ENOSPC	- insufficient disk resources
1950  *	-EIO	- i/o error
1951  *
1952  * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or
1953  *	IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit;
1954  */
1955 static int
1956 dbAllocDmapLev(struct bmap * bmp,
1957 	       struct dmap * dp, int nblocks, int l2nb, s64 * results)
1958 {
1959 	s64 blkno;
1960 	int leafidx, rc;
1961 
1962 	/* can't be more than a dmaps worth of blocks */
1963 	assert(l2nb <= L2BPERDMAP);
1964 
1965 	/* search the tree within the dmap page for sufficient
1966 	 * free space.  if sufficient free space is found, dbFindLeaf()
1967 	 * returns the index of the leaf at which free space was found.
1968 	 */
1969 	if (dbFindLeaf((dmtree_t *) &dp->tree, l2nb, &leafidx, false))
1970 		return -ENOSPC;
1971 
1972 	if (leafidx < 0)
1973 		return -EIO;
1974 
1975 	/* determine the block number within the file system corresponding
1976 	 * to the leaf at which free space was found.
1977 	 */
1978 	blkno = le64_to_cpu(dp->start) + (leafidx << L2DBWORD);
1979 
1980 	/* if not all bits of the dmap word are free, get the starting
1981 	 * bit number within the dmap word of the required string of free
1982 	 * bits and adjust the block number with this value.
1983 	 */
1984 	if (dp->tree.stree[leafidx + LEAFIND] < BUDMIN)
1985 		blkno += dbFindBits(le32_to_cpu(dp->wmap[leafidx]), l2nb);
1986 
1987 	/* allocate the blocks */
1988 	if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1989 		*results = blkno;
1990 
1991 	return (rc);
1992 }
1993 
1994 
1995 /*
1996  * NAME:	dbAllocDmap()
1997  *
1998  * FUNCTION:	adjust the disk allocation map to reflect the allocation
1999  *		of a specified block range within a dmap.
2000  *
2001  *		this routine allocates the specified blocks from the dmap
2002  *		through a call to dbAllocBits(). if the allocation of the
2003  *		block range causes the maximum string of free blocks within
2004  *		the dmap to change (i.e. the value of the root of the dmap's
2005  *		dmtree), this routine will cause this change to be reflected
2006  *		up through the appropriate levels of the dmap control pages
2007  *		by a call to dbAdjCtl() for the L0 dmap control page that
2008  *		covers this dmap.
2009  *
2010  * PARAMETERS:
2011  *	bmp	-  pointer to bmap descriptor
2012  *	dp	-  pointer to dmap to allocate the block range from.
2013  *	blkno	-  starting block number of the block to be allocated.
2014  *	nblocks	-  number of blocks to be allocated.
2015  *
2016  * RETURN VALUES:
2017  *	0	- success
2018  *	-EIO	- i/o error
2019  *
2020  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2021  */
2022 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2023 		       int nblocks)
2024 {
2025 	s8 oldroot;
2026 	int rc;
2027 
2028 	/* save the current value of the root (i.e. maximum free string)
2029 	 * of the dmap tree.
2030 	 */
2031 	oldroot = dp->tree.stree[ROOT];
2032 
2033 	/* allocate the specified (blocks) bits */
2034 	dbAllocBits(bmp, dp, blkno, nblocks);
2035 
2036 	/* if the root has not changed, done. */
2037 	if (dp->tree.stree[ROOT] == oldroot)
2038 		return (0);
2039 
2040 	/* root changed. bubble the change up to the dmap control pages.
2041 	 * if the adjustment of the upper level control pages fails,
2042 	 * backout the bit allocation (thus making everything consistent).
2043 	 */
2044 	if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 1, 0)))
2045 		dbFreeBits(bmp, dp, blkno, nblocks);
2046 
2047 	return (rc);
2048 }
2049 
2050 
2051 /*
2052  * NAME:	dbFreeDmap()
2053  *
2054  * FUNCTION:	adjust the disk allocation map to reflect the allocation
2055  *		of a specified block range within a dmap.
2056  *
2057  *		this routine frees the specified blocks from the dmap through
2058  *		a call to dbFreeBits(). if the deallocation of the block range
2059  *		causes the maximum string of free blocks within the dmap to
2060  *		change (i.e. the value of the root of the dmap's dmtree), this
2061  *		routine will cause this change to be reflected up through the
2062  *		appropriate levels of the dmap control pages by a call to
2063  *		dbAdjCtl() for the L0 dmap control page that covers this dmap.
2064  *
2065  * PARAMETERS:
2066  *	bmp	-  pointer to bmap descriptor
2067  *	dp	-  pointer to dmap to free the block range from.
2068  *	blkno	-  starting block number of the block to be freed.
2069  *	nblocks	-  number of blocks to be freed.
2070  *
2071  * RETURN VALUES:
2072  *	0	- success
2073  *	-EIO	- i/o error
2074  *
2075  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2076  */
2077 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2078 		      int nblocks)
2079 {
2080 	s8 oldroot;
2081 	int rc = 0, word;
2082 
2083 	/* save the current value of the root (i.e. maximum free string)
2084 	 * of the dmap tree.
2085 	 */
2086 	oldroot = dp->tree.stree[ROOT];
2087 
2088 	/* free the specified (blocks) bits */
2089 	rc = dbFreeBits(bmp, dp, blkno, nblocks);
2090 
2091 	/* if error or the root has not changed, done. */
2092 	if (rc || (dp->tree.stree[ROOT] == oldroot))
2093 		return (rc);
2094 
2095 	/* root changed. bubble the change up to the dmap control pages.
2096 	 * if the adjustment of the upper level control pages fails,
2097 	 * backout the deallocation.
2098 	 */
2099 	if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 0, 0))) {
2100 		word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
2101 
2102 		/* as part of backing out the deallocation, we will have
2103 		 * to back split the dmap tree if the deallocation caused
2104 		 * the freed blocks to become part of a larger binary buddy
2105 		 * system.
2106 		 */
2107 		if (dp->tree.stree[word] == NOFREE)
2108 			dbBackSplit((dmtree_t *)&dp->tree, word, false);
2109 
2110 		dbAllocBits(bmp, dp, blkno, nblocks);
2111 	}
2112 
2113 	return (rc);
2114 }
2115 
2116 
2117 /*
2118  * NAME:	dbAllocBits()
2119  *
2120  * FUNCTION:	allocate a specified block range from a dmap.
2121  *
2122  *		this routine updates the dmap to reflect the working
2123  *		state allocation of the specified block range. it directly
2124  *		updates the bits of the working map and causes the adjustment
2125  *		of the binary buddy system described by the dmap's dmtree
2126  *		leaves to reflect the bits allocated.  it also causes the
2127  *		dmap's dmtree, as a whole, to reflect the allocated range.
2128  *
2129  * PARAMETERS:
2130  *	bmp	-  pointer to bmap descriptor
2131  *	dp	-  pointer to dmap to allocate bits from.
2132  *	blkno	-  starting block number of the bits to be allocated.
2133  *	nblocks	-  number of bits to be allocated.
2134  *
2135  * RETURN VALUES: none
2136  *
2137  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2138  */
2139 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2140 			int nblocks)
2141 {
2142 	int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2143 	dmtree_t *tp = (dmtree_t *) & dp->tree;
2144 	int size;
2145 	s8 *leaf;
2146 
2147 	/* pick up a pointer to the leaves of the dmap tree */
2148 	leaf = dp->tree.stree + LEAFIND;
2149 
2150 	/* determine the bit number and word within the dmap of the
2151 	 * starting block.
2152 	 */
2153 	dbitno = blkno & (BPERDMAP - 1);
2154 	word = dbitno >> L2DBWORD;
2155 
2156 	/* block range better be within the dmap */
2157 	assert(dbitno + nblocks <= BPERDMAP);
2158 
2159 	/* allocate the bits of the dmap's words corresponding to the block
2160 	 * range. not all bits of the first and last words may be contained
2161 	 * within the block range.  if this is the case, we'll work against
2162 	 * those words (i.e. partial first and/or last) on an individual basis
2163 	 * (a single pass), allocating the bits of interest by hand and
2164 	 * updating the leaf corresponding to the dmap word. a single pass
2165 	 * will be used for all dmap words fully contained within the
2166 	 * specified range.  within this pass, the bits of all fully contained
2167 	 * dmap words will be marked as free in a single shot and the leaves
2168 	 * will be updated. a single leaf may describe the free space of
2169 	 * multiple dmap words, so we may update only a subset of the actual
2170 	 * leaves corresponding to the dmap words of the block range.
2171 	 */
2172 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2173 		/* determine the bit number within the word and
2174 		 * the number of bits within the word.
2175 		 */
2176 		wbitno = dbitno & (DBWORD - 1);
2177 		nb = min(rembits, DBWORD - wbitno);
2178 
2179 		/* check if only part of a word is to be allocated.
2180 		 */
2181 		if (nb < DBWORD) {
2182 			/* allocate (set to 1) the appropriate bits within
2183 			 * this dmap word.
2184 			 */
2185 			dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
2186 						      >> wbitno);
2187 
2188 			/* update the leaf for this dmap word. in addition
2189 			 * to setting the leaf value to the binary buddy max
2190 			 * of the updated dmap word, dbSplit() will split
2191 			 * the binary system of the leaves if need be.
2192 			 */
2193 			dbSplit(tp, word, BUDMIN,
2194 				dbMaxBud((u8 *)&dp->wmap[word]), false);
2195 
2196 			word += 1;
2197 		} else {
2198 			/* one or more dmap words are fully contained
2199 			 * within the block range.  determine how many
2200 			 * words and allocate (set to 1) the bits of these
2201 			 * words.
2202 			 */
2203 			nwords = rembits >> L2DBWORD;
2204 			memset(&dp->wmap[word], (int) ONES, nwords * 4);
2205 
2206 			/* determine how many bits.
2207 			 */
2208 			nb = nwords << L2DBWORD;
2209 
2210 			/* now update the appropriate leaves to reflect
2211 			 * the allocated words.
2212 			 */
2213 			for (; nwords > 0; nwords -= nw) {
2214 				if (leaf[word] < BUDMIN) {
2215 					jfs_error(bmp->db_ipbmap->i_sb,
2216 						  "leaf page corrupt\n");
2217 					break;
2218 				}
2219 
2220 				/* determine what the leaf value should be
2221 				 * updated to as the minimum of the l2 number
2222 				 * of bits being allocated and the l2 number
2223 				 * of bits currently described by this leaf.
2224 				 */
2225 				size = min_t(int, leaf[word],
2226 					     NLSTOL2BSZ(nwords));
2227 
2228 				/* update the leaf to reflect the allocation.
2229 				 * in addition to setting the leaf value to
2230 				 * NOFREE, dbSplit() will split the binary
2231 				 * system of the leaves to reflect the current
2232 				 * allocation (size).
2233 				 */
2234 				dbSplit(tp, word, size, NOFREE, false);
2235 
2236 				/* get the number of dmap words handled */
2237 				nw = BUDSIZE(size, BUDMIN);
2238 				word += nw;
2239 			}
2240 		}
2241 	}
2242 
2243 	/* update the free count for this dmap */
2244 	le32_add_cpu(&dp->nfree, -nblocks);
2245 
2246 	BMAP_LOCK(bmp);
2247 
2248 	/* if this allocation group is completely free,
2249 	 * update the maximum allocation group number if this allocation
2250 	 * group is the new max.
2251 	 */
2252 	agno = blkno >> bmp->db_agl2size;
2253 	if (agno > bmp->db_maxag)
2254 		bmp->db_maxag = agno;
2255 
2256 	/* update the free count for the allocation group and map */
2257 	bmp->db_agfree[agno] -= nblocks;
2258 	bmp->db_nfree -= nblocks;
2259 
2260 	BMAP_UNLOCK(bmp);
2261 }
2262 
2263 
2264 /*
2265  * NAME:	dbFreeBits()
2266  *
2267  * FUNCTION:	free a specified block range from a dmap.
2268  *
2269  *		this routine updates the dmap to reflect the working
2270  *		state allocation of the specified block range. it directly
2271  *		updates the bits of the working map and causes the adjustment
2272  *		of the binary buddy system described by the dmap's dmtree
2273  *		leaves to reflect the bits freed.  it also causes the dmap's
2274  *		dmtree, as a whole, to reflect the deallocated range.
2275  *
2276  * PARAMETERS:
2277  *	bmp	-  pointer to bmap descriptor
2278  *	dp	-  pointer to dmap to free bits from.
2279  *	blkno	-  starting block number of the bits to be freed.
2280  *	nblocks	-  number of bits to be freed.
2281  *
2282  * RETURN VALUES: 0 for success
2283  *
2284  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2285  */
2286 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2287 		       int nblocks)
2288 {
2289 	int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2290 	dmtree_t *tp = (dmtree_t *) & dp->tree;
2291 	int rc = 0;
2292 	int size;
2293 
2294 	/* determine the bit number and word within the dmap of the
2295 	 * starting block.
2296 	 */
2297 	dbitno = blkno & (BPERDMAP - 1);
2298 	word = dbitno >> L2DBWORD;
2299 
2300 	/* block range better be within the dmap.
2301 	 */
2302 	assert(dbitno + nblocks <= BPERDMAP);
2303 
2304 	/* free the bits of the dmaps words corresponding to the block range.
2305 	 * not all bits of the first and last words may be contained within
2306 	 * the block range.  if this is the case, we'll work against those
2307 	 * words (i.e. partial first and/or last) on an individual basis
2308 	 * (a single pass), freeing the bits of interest by hand and updating
2309 	 * the leaf corresponding to the dmap word. a single pass will be used
2310 	 * for all dmap words fully contained within the specified range.
2311 	 * within this pass, the bits of all fully contained dmap words will
2312 	 * be marked as free in a single shot and the leaves will be updated. a
2313 	 * single leaf may describe the free space of multiple dmap words,
2314 	 * so we may update only a subset of the actual leaves corresponding
2315 	 * to the dmap words of the block range.
2316 	 *
2317 	 * dbJoin() is used to update leaf values and will join the binary
2318 	 * buddy system of the leaves if the new leaf values indicate this
2319 	 * should be done.
2320 	 */
2321 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2322 		/* determine the bit number within the word and
2323 		 * the number of bits within the word.
2324 		 */
2325 		wbitno = dbitno & (DBWORD - 1);
2326 		nb = min(rembits, DBWORD - wbitno);
2327 
2328 		/* check if only part of a word is to be freed.
2329 		 */
2330 		if (nb < DBWORD) {
2331 			/* free (zero) the appropriate bits within this
2332 			 * dmap word.
2333 			 */
2334 			dp->wmap[word] &=
2335 			    cpu_to_le32(~(ONES << (DBWORD - nb)
2336 					  >> wbitno));
2337 
2338 			/* update the leaf for this dmap word.
2339 			 */
2340 			rc = dbJoin(tp, word,
2341 				    dbMaxBud((u8 *)&dp->wmap[word]), false);
2342 			if (rc)
2343 				return rc;
2344 
2345 			word += 1;
2346 		} else {
2347 			/* one or more dmap words are fully contained
2348 			 * within the block range.  determine how many
2349 			 * words and free (zero) the bits of these words.
2350 			 */
2351 			nwords = rembits >> L2DBWORD;
2352 			memset(&dp->wmap[word], 0, nwords * 4);
2353 
2354 			/* determine how many bits.
2355 			 */
2356 			nb = nwords << L2DBWORD;
2357 
2358 			/* now update the appropriate leaves to reflect
2359 			 * the freed words.
2360 			 */
2361 			for (; nwords > 0; nwords -= nw) {
2362 				/* determine what the leaf value should be
2363 				 * updated to as the minimum of the l2 number
2364 				 * of bits being freed and the l2 (max) number
2365 				 * of bits that can be described by this leaf.
2366 				 */
2367 				size =
2368 				    min(LITOL2BSZ
2369 					(word, L2LPERDMAP, BUDMIN),
2370 					NLSTOL2BSZ(nwords));
2371 
2372 				/* update the leaf.
2373 				 */
2374 				rc = dbJoin(tp, word, size, false);
2375 				if (rc)
2376 					return rc;
2377 
2378 				/* get the number of dmap words handled.
2379 				 */
2380 				nw = BUDSIZE(size, BUDMIN);
2381 				word += nw;
2382 			}
2383 		}
2384 	}
2385 
2386 	/* update the free count for this dmap.
2387 	 */
2388 	le32_add_cpu(&dp->nfree, nblocks);
2389 
2390 	BMAP_LOCK(bmp);
2391 
2392 	/* update the free count for the allocation group and
2393 	 * map.
2394 	 */
2395 	agno = blkno >> bmp->db_agl2size;
2396 	bmp->db_nfree += nblocks;
2397 	bmp->db_agfree[agno] += nblocks;
2398 
2399 	/* check if this allocation group is not completely free and
2400 	 * if it is currently the maximum (rightmost) allocation group.
2401 	 * if so, establish the new maximum allocation group number by
2402 	 * searching left for the first allocation group with allocation.
2403 	 */
2404 	if ((bmp->db_agfree[agno] == bmp->db_agsize && agno == bmp->db_maxag) ||
2405 	    (agno == bmp->db_numag - 1 &&
2406 	     bmp->db_agfree[agno] == (bmp-> db_mapsize & (BPERDMAP - 1)))) {
2407 		while (bmp->db_maxag > 0) {
2408 			bmp->db_maxag -= 1;
2409 			if (bmp->db_agfree[bmp->db_maxag] !=
2410 			    bmp->db_agsize)
2411 				break;
2412 		}
2413 
2414 		/* re-establish the allocation group preference if the
2415 		 * current preference is right of the maximum allocation
2416 		 * group.
2417 		 */
2418 		if (bmp->db_agpref > bmp->db_maxag)
2419 			bmp->db_agpref = bmp->db_maxag;
2420 	}
2421 
2422 	BMAP_UNLOCK(bmp);
2423 
2424 	return 0;
2425 }
2426 
2427 
2428 /*
2429  * NAME:	dbAdjCtl()
2430  *
2431  * FUNCTION:	adjust a dmap control page at a specified level to reflect
2432  *		the change in a lower level dmap or dmap control page's
2433  *		maximum string of free blocks (i.e. a change in the root
2434  *		of the lower level object's dmtree) due to the allocation
2435  *		or deallocation of a range of blocks with a single dmap.
2436  *
2437  *		on entry, this routine is provided with the new value of
2438  *		the lower level dmap or dmap control page root and the
2439  *		starting block number of the block range whose allocation
2440  *		or deallocation resulted in the root change.  this range
2441  *		is respresented by a single leaf of the current dmapctl
2442  *		and the leaf will be updated with this value, possibly
2443  *		causing a binary buddy system within the leaves to be
2444  *		split or joined.  the update may also cause the dmapctl's
2445  *		dmtree to be updated.
2446  *
2447  *		if the adjustment of the dmap control page, itself, causes its
2448  *		root to change, this change will be bubbled up to the next dmap
2449  *		control level by a recursive call to this routine, specifying
2450  *		the new root value and the next dmap control page level to
2451  *		be adjusted.
2452  * PARAMETERS:
2453  *	bmp	-  pointer to bmap descriptor
2454  *	blkno	-  the first block of a block range within a dmap.  it is
2455  *		   the allocation or deallocation of this block range that
2456  *		   requires the dmap control page to be adjusted.
2457  *	newval	-  the new value of the lower level dmap or dmap control
2458  *		   page root.
2459  *	alloc	-  'true' if adjustment is due to an allocation.
2460  *	level	-  current level of dmap control page (i.e. L0, L1, L2) to
2461  *		   be adjusted.
2462  *
2463  * RETURN VALUES:
2464  *	0	- success
2465  *	-EIO	- i/o error
2466  *
2467  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2468  */
2469 static int
2470 dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, int level)
2471 {
2472 	struct metapage *mp;
2473 	s8 oldroot;
2474 	int oldval;
2475 	s64 lblkno;
2476 	struct dmapctl *dcp;
2477 	int rc, leafno, ti;
2478 
2479 	/* get the buffer for the dmap control page for the specified
2480 	 * block number and control page level.
2481 	 */
2482 	lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, level);
2483 	mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
2484 	if (mp == NULL)
2485 		return -EIO;
2486 	dcp = (struct dmapctl *) mp->data;
2487 
2488 	if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
2489 		jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
2490 		release_metapage(mp);
2491 		return -EIO;
2492 	}
2493 
2494 	/* determine the leaf number corresponding to the block and
2495 	 * the index within the dmap control tree.
2496 	 */
2497 	leafno = BLKTOCTLLEAF(blkno, dcp->budmin);
2498 	ti = leafno + le32_to_cpu(dcp->leafidx);
2499 
2500 	/* save the current leaf value and the current root level (i.e.
2501 	 * maximum l2 free string described by this dmapctl).
2502 	 */
2503 	oldval = dcp->stree[ti];
2504 	oldroot = dcp->stree[ROOT];
2505 
2506 	/* check if this is a control page update for an allocation.
2507 	 * if so, update the leaf to reflect the new leaf value using
2508 	 * dbSplit(); otherwise (deallocation), use dbJoin() to update
2509 	 * the leaf with the new value.  in addition to updating the
2510 	 * leaf, dbSplit() will also split the binary buddy system of
2511 	 * the leaves, if required, and bubble new values within the
2512 	 * dmapctl tree, if required.  similarly, dbJoin() will join
2513 	 * the binary buddy system of leaves and bubble new values up
2514 	 * the dmapctl tree as required by the new leaf value.
2515 	 */
2516 	if (alloc) {
2517 		/* check if we are in the middle of a binary buddy
2518 		 * system.  this happens when we are performing the
2519 		 * first allocation out of an allocation group that
2520 		 * is part (not the first part) of a larger binary
2521 		 * buddy system.  if we are in the middle, back split
2522 		 * the system prior to calling dbSplit() which assumes
2523 		 * that it is at the front of a binary buddy system.
2524 		 */
2525 		if (oldval == NOFREE) {
2526 			rc = dbBackSplit((dmtree_t *)dcp, leafno, true);
2527 			if (rc) {
2528 				release_metapage(mp);
2529 				return rc;
2530 			}
2531 			oldval = dcp->stree[ti];
2532 		}
2533 		dbSplit((dmtree_t *) dcp, leafno, dcp->budmin, newval, true);
2534 	} else {
2535 		rc = dbJoin((dmtree_t *) dcp, leafno, newval, true);
2536 		if (rc) {
2537 			release_metapage(mp);
2538 			return rc;
2539 		}
2540 	}
2541 
2542 	/* check if the root of the current dmap control page changed due
2543 	 * to the update and if the current dmap control page is not at
2544 	 * the current top level (i.e. L0, L1, L2) of the map.  if so (i.e.
2545 	 * root changed and this is not the top level), call this routine
2546 	 * again (recursion) for the next higher level of the mapping to
2547 	 * reflect the change in root for the current dmap control page.
2548 	 */
2549 	if (dcp->stree[ROOT] != oldroot) {
2550 		/* are we below the top level of the map.  if so,
2551 		 * bubble the root up to the next higher level.
2552 		 */
2553 		if (level < bmp->db_maxlevel) {
2554 			/* bubble up the new root of this dmap control page to
2555 			 * the next level.
2556 			 */
2557 			if ((rc =
2558 			     dbAdjCtl(bmp, blkno, dcp->stree[ROOT], alloc,
2559 				      level + 1))) {
2560 				/* something went wrong in bubbling up the new
2561 				 * root value, so backout the changes to the
2562 				 * current dmap control page.
2563 				 */
2564 				if (alloc) {
2565 					dbJoin((dmtree_t *) dcp, leafno,
2566 					       oldval, true);
2567 				} else {
2568 					/* the dbJoin() above might have
2569 					 * caused a larger binary buddy system
2570 					 * to form and we may now be in the
2571 					 * middle of it.  if this is the case,
2572 					 * back split the buddies.
2573 					 */
2574 					if (dcp->stree[ti] == NOFREE)
2575 						dbBackSplit((dmtree_t *)
2576 							    dcp, leafno, true);
2577 					dbSplit((dmtree_t *) dcp, leafno,
2578 						dcp->budmin, oldval, true);
2579 				}
2580 
2581 				/* release the buffer and return the error.
2582 				 */
2583 				release_metapage(mp);
2584 				return (rc);
2585 			}
2586 		} else {
2587 			/* we're at the top level of the map. update
2588 			 * the bmap control page to reflect the size
2589 			 * of the maximum free buddy system.
2590 			 */
2591 			assert(level == bmp->db_maxlevel);
2592 			if (bmp->db_maxfreebud != oldroot) {
2593 				jfs_error(bmp->db_ipbmap->i_sb,
2594 					  "the maximum free buddy is not the old root\n");
2595 			}
2596 			bmp->db_maxfreebud = dcp->stree[ROOT];
2597 		}
2598 	}
2599 
2600 	/* write the buffer.
2601 	 */
2602 	write_metapage(mp);
2603 
2604 	return (0);
2605 }
2606 
2607 
2608 /*
2609  * NAME:	dbSplit()
2610  *
2611  * FUNCTION:	update the leaf of a dmtree with a new value, splitting
2612  *		the leaf from the binary buddy system of the dmtree's
2613  *		leaves, as required.
2614  *
2615  * PARAMETERS:
2616  *	tp	- pointer to the tree containing the leaf.
2617  *	leafno	- the number of the leaf to be updated.
2618  *	splitsz	- the size the binary buddy system starting at the leaf
2619  *		  must be split to, specified as the log2 number of blocks.
2620  *	newval	- the new value for the leaf.
2621  *
2622  * RETURN VALUES: none
2623  *
2624  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2625  */
2626 static void dbSplit(dmtree_t *tp, int leafno, int splitsz, int newval, bool is_ctl)
2627 {
2628 	int budsz;
2629 	int cursz;
2630 	s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2631 
2632 	/* check if the leaf needs to be split.
2633 	 */
2634 	if (leaf[leafno] > tp->dmt_budmin) {
2635 		/* the split occurs by cutting the buddy system in half
2636 		 * at the specified leaf until we reach the specified
2637 		 * size.  pick up the starting split size (current size
2638 		 * - 1 in l2) and the corresponding buddy size.
2639 		 */
2640 		cursz = leaf[leafno] - 1;
2641 		budsz = BUDSIZE(cursz, tp->dmt_budmin);
2642 
2643 		/* split until we reach the specified size.
2644 		 */
2645 		while (cursz >= splitsz) {
2646 			/* update the buddy's leaf with its new value.
2647 			 */
2648 			dbAdjTree(tp, leafno ^ budsz, cursz, is_ctl);
2649 
2650 			/* on to the next size and buddy.
2651 			 */
2652 			cursz -= 1;
2653 			budsz >>= 1;
2654 		}
2655 	}
2656 
2657 	/* adjust the dmap tree to reflect the specified leaf's new
2658 	 * value.
2659 	 */
2660 	dbAdjTree(tp, leafno, newval, is_ctl);
2661 }
2662 
2663 
2664 /*
2665  * NAME:	dbBackSplit()
2666  *
2667  * FUNCTION:	back split the binary buddy system of dmtree leaves
2668  *		that hold a specified leaf until the specified leaf
2669  *		starts its own binary buddy system.
2670  *
2671  *		the allocators typically perform allocations at the start
2672  *		of binary buddy systems and dbSplit() is used to accomplish
2673  *		any required splits.  in some cases, however, allocation
2674  *		may occur in the middle of a binary system and requires a
2675  *		back split, with the split proceeding out from the middle of
2676  *		the system (less efficient) rather than the start of the
2677  *		system (more efficient).  the cases in which a back split
2678  *		is required are rare and are limited to the first allocation
2679  *		within an allocation group which is a part (not first part)
2680  *		of a larger binary buddy system and a few exception cases
2681  *		in which a previous join operation must be backed out.
2682  *
2683  * PARAMETERS:
2684  *	tp	- pointer to the tree containing the leaf.
2685  *	leafno	- the number of the leaf to be updated.
2686  *
2687  * RETURN VALUES: none
2688  *
2689  * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2690  */
2691 static int dbBackSplit(dmtree_t *tp, int leafno, bool is_ctl)
2692 {
2693 	int budsz, bud, w, bsz, size;
2694 	int cursz;
2695 	s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2696 
2697 	/* leaf should be part (not first part) of a binary
2698 	 * buddy system.
2699 	 */
2700 	assert(leaf[leafno] == NOFREE);
2701 
2702 	/* the back split is accomplished by iteratively finding the leaf
2703 	 * that starts the buddy system that contains the specified leaf and
2704 	 * splitting that system in two.  this iteration continues until
2705 	 * the specified leaf becomes the start of a buddy system.
2706 	 *
2707 	 * determine maximum possible l2 size for the specified leaf.
2708 	 */
2709 	size =
2710 	    LITOL2BSZ(leafno, le32_to_cpu(tp->dmt_l2nleafs),
2711 		      tp->dmt_budmin);
2712 
2713 	/* determine the number of leaves covered by this size.  this
2714 	 * is the buddy size that we will start with as we search for
2715 	 * the buddy system that contains the specified leaf.
2716 	 */
2717 	budsz = BUDSIZE(size, tp->dmt_budmin);
2718 
2719 	/* back split.
2720 	 */
2721 	while (leaf[leafno] == NOFREE) {
2722 		/* find the leftmost buddy leaf.
2723 		 */
2724 		for (w = leafno, bsz = budsz;; bsz <<= 1,
2725 		     w = (w < bud) ? w : bud) {
2726 			if (bsz >= le32_to_cpu(tp->dmt_nleafs)) {
2727 				jfs_err("JFS: block map error in dbBackSplit");
2728 				return -EIO;
2729 			}
2730 
2731 			/* determine the buddy.
2732 			 */
2733 			bud = w ^ bsz;
2734 
2735 			/* check if this buddy is the start of the system.
2736 			 */
2737 			if (leaf[bud] != NOFREE) {
2738 				/* split the leaf at the start of the
2739 				 * system in two.
2740 				 */
2741 				cursz = leaf[bud] - 1;
2742 				dbSplit(tp, bud, cursz, cursz, is_ctl);
2743 				break;
2744 			}
2745 		}
2746 	}
2747 
2748 	if (leaf[leafno] != size) {
2749 		jfs_err("JFS: wrong leaf value in dbBackSplit");
2750 		return -EIO;
2751 	}
2752 	return 0;
2753 }
2754 
2755 
2756 /*
2757  * NAME:	dbJoin()
2758  *
2759  * FUNCTION:	update the leaf of a dmtree with a new value, joining
2760  *		the leaf with other leaves of the dmtree into a multi-leaf
2761  *		binary buddy system, as required.
2762  *
2763  * PARAMETERS:
2764  *	tp	- pointer to the tree containing the leaf.
2765  *	leafno	- the number of the leaf to be updated.
2766  *	newval	- the new value for the leaf.
2767  *
2768  * RETURN VALUES: none
2769  */
2770 static int dbJoin(dmtree_t *tp, int leafno, int newval, bool is_ctl)
2771 {
2772 	int budsz, buddy;
2773 	s8 *leaf;
2774 
2775 	/* can the new leaf value require a join with other leaves ?
2776 	 */
2777 	if (newval >= tp->dmt_budmin) {
2778 		/* pickup a pointer to the leaves of the tree.
2779 		 */
2780 		leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2781 
2782 		/* try to join the specified leaf into a large binary
2783 		 * buddy system.  the join proceeds by attempting to join
2784 		 * the specified leafno with its buddy (leaf) at new value.
2785 		 * if the join occurs, we attempt to join the left leaf
2786 		 * of the joined buddies with its buddy at new value + 1.
2787 		 * we continue to join until we find a buddy that cannot be
2788 		 * joined (does not have a value equal to the size of the
2789 		 * last join) or until all leaves have been joined into a
2790 		 * single system.
2791 		 *
2792 		 * get the buddy size (number of words covered) of
2793 		 * the new value.
2794 		 */
2795 		budsz = BUDSIZE(newval, tp->dmt_budmin);
2796 
2797 		/* try to join.
2798 		 */
2799 		while (budsz < le32_to_cpu(tp->dmt_nleafs)) {
2800 			/* get the buddy leaf.
2801 			 */
2802 			buddy = leafno ^ budsz;
2803 
2804 			/* if the leaf's new value is greater than its
2805 			 * buddy's value, we join no more.
2806 			 */
2807 			if (newval > leaf[buddy])
2808 				break;
2809 
2810 			/* It shouldn't be less */
2811 			if (newval < leaf[buddy])
2812 				return -EIO;
2813 
2814 			/* check which (leafno or buddy) is the left buddy.
2815 			 * the left buddy gets to claim the blocks resulting
2816 			 * from the join while the right gets to claim none.
2817 			 * the left buddy is also eligible to participate in
2818 			 * a join at the next higher level while the right
2819 			 * is not.
2820 			 *
2821 			 */
2822 			if (leafno < buddy) {
2823 				/* leafno is the left buddy.
2824 				 */
2825 				dbAdjTree(tp, buddy, NOFREE, is_ctl);
2826 			} else {
2827 				/* buddy is the left buddy and becomes
2828 				 * leafno.
2829 				 */
2830 				dbAdjTree(tp, leafno, NOFREE, is_ctl);
2831 				leafno = buddy;
2832 			}
2833 
2834 			/* on to try the next join.
2835 			 */
2836 			newval += 1;
2837 			budsz <<= 1;
2838 		}
2839 	}
2840 
2841 	/* update the leaf value.
2842 	 */
2843 	dbAdjTree(tp, leafno, newval, is_ctl);
2844 
2845 	return 0;
2846 }
2847 
2848 
2849 /*
2850  * NAME:	dbAdjTree()
2851  *
2852  * FUNCTION:	update a leaf of a dmtree with a new value, adjusting
2853  *		the dmtree, as required, to reflect the new leaf value.
2854  *		the combination of any buddies must already be done before
2855  *		this is called.
2856  *
2857  * PARAMETERS:
2858  *	tp	- pointer to the tree to be adjusted.
2859  *	leafno	- the number of the leaf to be updated.
2860  *	newval	- the new value for the leaf.
2861  *
2862  * RETURN VALUES: none
2863  */
2864 static void dbAdjTree(dmtree_t *tp, int leafno, int newval, bool is_ctl)
2865 {
2866 	int lp, pp, k;
2867 	int max, size;
2868 
2869 	size = is_ctl ? CTLTREESIZE : TREESIZE;
2870 
2871 	/* pick up the index of the leaf for this leafno.
2872 	 */
2873 	lp = leafno + le32_to_cpu(tp->dmt_leafidx);
2874 
2875 	if (WARN_ON_ONCE(lp >= size || lp < 0))
2876 		return;
2877 
2878 	/* is the current value the same as the old value ?  if so,
2879 	 * there is nothing to do.
2880 	 */
2881 	if (tp->dmt_stree[lp] == newval)
2882 		return;
2883 
2884 	/* set the new value.
2885 	 */
2886 	tp->dmt_stree[lp] = newval;
2887 
2888 	/* bubble the new value up the tree as required.
2889 	 */
2890 	for (k = 0; k < le32_to_cpu(tp->dmt_height); k++) {
2891 		/* get the index of the first leaf of the 4 leaf
2892 		 * group containing the specified leaf (leafno).
2893 		 */
2894 		lp = ((lp - 1) & ~0x03) + 1;
2895 
2896 		/* get the index of the parent of this 4 leaf group.
2897 		 */
2898 		pp = (lp - 1) >> 2;
2899 
2900 		/* determine the maximum of the 4 leaves.
2901 		 */
2902 		max = TREEMAX(&tp->dmt_stree[lp]);
2903 
2904 		/* if the maximum of the 4 is the same as the
2905 		 * parent's value, we're done.
2906 		 */
2907 		if (tp->dmt_stree[pp] == max)
2908 			break;
2909 
2910 		/* parent gets new value.
2911 		 */
2912 		tp->dmt_stree[pp] = max;
2913 
2914 		/* parent becomes leaf for next go-round.
2915 		 */
2916 		lp = pp;
2917 	}
2918 }
2919 
2920 
2921 /*
2922  * NAME:	dbFindLeaf()
2923  *
2924  * FUNCTION:	search a dmtree_t for sufficient free blocks, returning
2925  *		the index of a leaf describing the free blocks if
2926  *		sufficient free blocks are found.
2927  *
2928  *		the search starts at the top of the dmtree_t tree and
2929  *		proceeds down the tree to the leftmost leaf with sufficient
2930  *		free space.
2931  *
2932  * PARAMETERS:
2933  *	tp	- pointer to the tree to be searched.
2934  *	l2nb	- log2 number of free blocks to search for.
2935  *	leafidx	- return pointer to be set to the index of the leaf
2936  *		  describing at least l2nb free blocks if sufficient
2937  *		  free blocks are found.
2938  *	is_ctl	- determines if the tree is of type ctl
2939  *
2940  * RETURN VALUES:
2941  *	0	- success
2942  *	-ENOSPC	- insufficient free blocks.
2943  */
2944 static int dbFindLeaf(dmtree_t *tp, int l2nb, int *leafidx, bool is_ctl)
2945 {
2946 	int ti, n = 0, k, x = 0;
2947 	int max_size, max_idx;
2948 
2949 	max_size = is_ctl ? CTLTREESIZE : TREESIZE;
2950 	max_idx = is_ctl ? LPERCTL : LPERDMAP;
2951 
2952 	/* first check the root of the tree to see if there is
2953 	 * sufficient free space.
2954 	 */
2955 	if (l2nb > tp->dmt_stree[ROOT])
2956 		return -ENOSPC;
2957 
2958 	/* sufficient free space available. now search down the tree
2959 	 * starting at the next level for the leftmost leaf that
2960 	 * describes sufficient free space.
2961 	 */
2962 	for (k = le32_to_cpu(tp->dmt_height), ti = 1;
2963 	     k > 0; k--, ti = ((ti + n) << 2) + 1) {
2964 		/* search the four nodes at this level, starting from
2965 		 * the left.
2966 		 */
2967 		for (x = ti, n = 0; n < 4; n++) {
2968 			/* sufficient free space found.  move to the next
2969 			 * level (or quit if this is the last level).
2970 			 */
2971 			if (x + n > max_size)
2972 				return -ENOSPC;
2973 			if (l2nb <= tp->dmt_stree[x + n])
2974 				break;
2975 		}
2976 
2977 		/* better have found something since the higher
2978 		 * levels of the tree said it was here.
2979 		 */
2980 		assert(n < 4);
2981 	}
2982 	if (le32_to_cpu(tp->dmt_leafidx) >= max_idx)
2983 		return -ENOSPC;
2984 
2985 	/* set the return to the leftmost leaf describing sufficient
2986 	 * free space.
2987 	 */
2988 	*leafidx = x + n - le32_to_cpu(tp->dmt_leafidx);
2989 
2990 	return (0);
2991 }
2992 
2993 
2994 /*
2995  * NAME:	dbFindBits()
2996  *
2997  * FUNCTION:	find a specified number of binary buddy free bits within a
2998  *		dmap bitmap word value.
2999  *
3000  *		this routine searches the bitmap value for (1 << l2nb) free
3001  *		bits at (1 << l2nb) alignments within the value.
3002  *
3003  * PARAMETERS:
3004  *	word	-  dmap bitmap word value.
3005  *	l2nb	-  number of free bits specified as a log2 number.
3006  *
3007  * RETURN VALUES:
3008  *	starting bit number of free bits.
3009  */
3010 static int dbFindBits(u32 word, int l2nb)
3011 {
3012 	int bitno, nb;
3013 	u32 mask;
3014 
3015 	/* get the number of bits.
3016 	 */
3017 	nb = 1 << l2nb;
3018 	assert(nb <= DBWORD);
3019 
3020 	/* complement the word so we can use a mask (i.e. 0s represent
3021 	 * free bits) and compute the mask.
3022 	 */
3023 	word = ~word;
3024 	mask = ONES << (DBWORD - nb);
3025 
3026 	/* scan the word for nb free bits at nb alignments.
3027 	 */
3028 	for (bitno = 0; mask != 0; bitno += nb, mask = (mask >> nb)) {
3029 		if ((mask & word) == mask)
3030 			break;
3031 	}
3032 
3033 	ASSERT(bitno < 32);
3034 
3035 	/* return the bit number.
3036 	 */
3037 	return (bitno);
3038 }
3039 
3040 
3041 /*
3042  * NAME:	dbMaxBud(u8 *cp)
3043  *
3044  * FUNCTION:	determine the largest binary buddy string of free
3045  *		bits within 32-bits of the map.
3046  *
3047  * PARAMETERS:
3048  *	cp	-  pointer to the 32-bit value.
3049  *
3050  * RETURN VALUES:
3051  *	largest binary buddy of free bits within a dmap word.
3052  */
3053 static int dbMaxBud(u8 * cp)
3054 {
3055 	signed char tmp1, tmp2;
3056 
3057 	/* check if the wmap word is all free. if so, the
3058 	 * free buddy size is BUDMIN.
3059 	 */
3060 	if (*((uint *) cp) == 0)
3061 		return (BUDMIN);
3062 
3063 	/* check if the wmap word is half free. if so, the
3064 	 * free buddy size is BUDMIN-1.
3065 	 */
3066 	if (*((u16 *) cp) == 0 || *((u16 *) cp + 1) == 0)
3067 		return (BUDMIN - 1);
3068 
3069 	/* not all free or half free. determine the free buddy
3070 	 * size thru table lookup using quarters of the wmap word.
3071 	 */
3072 	tmp1 = max(budtab[cp[2]], budtab[cp[3]]);
3073 	tmp2 = max(budtab[cp[0]], budtab[cp[1]]);
3074 	return (max(tmp1, tmp2));
3075 }
3076 
3077 
3078 /*
3079  * NAME:	cnttz(uint word)
3080  *
3081  * FUNCTION:	determine the number of trailing zeros within a 32-bit
3082  *		value.
3083  *
3084  * PARAMETERS:
3085  *	value	-  32-bit value to be examined.
3086  *
3087  * RETURN VALUES:
3088  *	count of trailing zeros
3089  */
3090 static int cnttz(u32 word)
3091 {
3092 	int n;
3093 
3094 	for (n = 0; n < 32; n++, word >>= 1) {
3095 		if (word & 0x01)
3096 			break;
3097 	}
3098 
3099 	return (n);
3100 }
3101 
3102 
3103 /*
3104  * NAME:	cntlz(u32 value)
3105  *
3106  * FUNCTION:	determine the number of leading zeros within a 32-bit
3107  *		value.
3108  *
3109  * PARAMETERS:
3110  *	value	-  32-bit value to be examined.
3111  *
3112  * RETURN VALUES:
3113  *	count of leading zeros
3114  */
3115 static int cntlz(u32 value)
3116 {
3117 	int n;
3118 
3119 	for (n = 0; n < 32; n++, value <<= 1) {
3120 		if (value & HIGHORDER)
3121 			break;
3122 	}
3123 	return (n);
3124 }
3125 
3126 
3127 /*
3128  * NAME:	blkstol2(s64 nb)
3129  *
3130  * FUNCTION:	convert a block count to its log2 value. if the block
3131  *		count is not a l2 multiple, it is rounded up to the next
3132  *		larger l2 multiple.
3133  *
3134  * PARAMETERS:
3135  *	nb	-  number of blocks
3136  *
3137  * RETURN VALUES:
3138  *	log2 number of blocks
3139  */
3140 static int blkstol2(s64 nb)
3141 {
3142 	int l2nb;
3143 	s64 mask;		/* meant to be signed */
3144 
3145 	mask = (s64) 1 << (64 - 1);
3146 
3147 	/* count the leading bits.
3148 	 */
3149 	for (l2nb = 0; l2nb < 64; l2nb++, mask >>= 1) {
3150 		/* leading bit found.
3151 		 */
3152 		if (nb & mask) {
3153 			/* determine the l2 value.
3154 			 */
3155 			l2nb = (64 - 1) - l2nb;
3156 
3157 			/* check if we need to round up.
3158 			 */
3159 			if (~mask & nb)
3160 				l2nb++;
3161 
3162 			return (l2nb);
3163 		}
3164 	}
3165 	assert(0);
3166 	return 0;		/* fix compiler warning */
3167 }
3168 
3169 
3170 /*
3171  * NAME:	dbAllocBottomUp()
3172  *
3173  * FUNCTION:	alloc the specified block range from the working block
3174  *		allocation map.
3175  *
3176  *		the blocks will be alloc from the working map one dmap
3177  *		at a time.
3178  *
3179  * PARAMETERS:
3180  *	ip	-  pointer to in-core inode;
3181  *	blkno	-  starting block number to be freed.
3182  *	nblocks	-  number of blocks to be freed.
3183  *
3184  * RETURN VALUES:
3185  *	0	- success
3186  *	-EIO	- i/o error
3187  */
3188 int dbAllocBottomUp(struct inode *ip, s64 blkno, s64 nblocks)
3189 {
3190 	struct metapage *mp;
3191 	struct dmap *dp;
3192 	int nb, rc;
3193 	s64 lblkno, rem;
3194 	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
3195 	struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
3196 
3197 	IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
3198 
3199 	/* block to be allocated better be within the mapsize. */
3200 	ASSERT(nblocks <= bmp->db_mapsize - blkno);
3201 
3202 	/*
3203 	 * allocate the blocks a dmap at a time.
3204 	 */
3205 	mp = NULL;
3206 	for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
3207 		/* release previous dmap if any */
3208 		if (mp) {
3209 			write_metapage(mp);
3210 		}
3211 
3212 		/* get the buffer for the current dmap. */
3213 		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
3214 		mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
3215 		if (mp == NULL) {
3216 			IREAD_UNLOCK(ipbmap);
3217 			return -EIO;
3218 		}
3219 		dp = (struct dmap *) mp->data;
3220 
3221 		/* determine the number of blocks to be allocated from
3222 		 * this dmap.
3223 		 */
3224 		nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
3225 
3226 		/* allocate the blocks. */
3227 		if ((rc = dbAllocDmapBU(bmp, dp, blkno, nb))) {
3228 			release_metapage(mp);
3229 			IREAD_UNLOCK(ipbmap);
3230 			return (rc);
3231 		}
3232 	}
3233 
3234 	/* write the last buffer. */
3235 	write_metapage(mp);
3236 
3237 	IREAD_UNLOCK(ipbmap);
3238 
3239 	return (0);
3240 }
3241 
3242 
3243 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
3244 			 int nblocks)
3245 {
3246 	int rc;
3247 	int dbitno, word, rembits, nb, nwords, wbitno, agno;
3248 	s8 oldroot;
3249 	struct dmaptree *tp = (struct dmaptree *) & dp->tree;
3250 
3251 	/* save the current value of the root (i.e. maximum free string)
3252 	 * of the dmap tree.
3253 	 */
3254 	oldroot = tp->stree[ROOT];
3255 
3256 	/* determine the bit number and word within the dmap of the
3257 	 * starting block.
3258 	 */
3259 	dbitno = blkno & (BPERDMAP - 1);
3260 	word = dbitno >> L2DBWORD;
3261 
3262 	/* block range better be within the dmap */
3263 	assert(dbitno + nblocks <= BPERDMAP);
3264 
3265 	/* allocate the bits of the dmap's words corresponding to the block
3266 	 * range. not all bits of the first and last words may be contained
3267 	 * within the block range.  if this is the case, we'll work against
3268 	 * those words (i.e. partial first and/or last) on an individual basis
3269 	 * (a single pass), allocating the bits of interest by hand and
3270 	 * updating the leaf corresponding to the dmap word. a single pass
3271 	 * will be used for all dmap words fully contained within the
3272 	 * specified range.  within this pass, the bits of all fully contained
3273 	 * dmap words will be marked as free in a single shot and the leaves
3274 	 * will be updated. a single leaf may describe the free space of
3275 	 * multiple dmap words, so we may update only a subset of the actual
3276 	 * leaves corresponding to the dmap words of the block range.
3277 	 */
3278 	for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
3279 		/* determine the bit number within the word and
3280 		 * the number of bits within the word.
3281 		 */
3282 		wbitno = dbitno & (DBWORD - 1);
3283 		nb = min(rembits, DBWORD - wbitno);
3284 
3285 		/* check if only part of a word is to be allocated.
3286 		 */
3287 		if (nb < DBWORD) {
3288 			/* allocate (set to 1) the appropriate bits within
3289 			 * this dmap word.
3290 			 */
3291 			dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
3292 						      >> wbitno);
3293 
3294 			word++;
3295 		} else {
3296 			/* one or more dmap words are fully contained
3297 			 * within the block range.  determine how many
3298 			 * words and allocate (set to 1) the bits of these
3299 			 * words.
3300 			 */
3301 			nwords = rembits >> L2DBWORD;
3302 			memset(&dp->wmap[word], (int) ONES, nwords * 4);
3303 
3304 			/* determine how many bits */
3305 			nb = nwords << L2DBWORD;
3306 			word += nwords;
3307 		}
3308 	}
3309 
3310 	/* update the free count for this dmap */
3311 	le32_add_cpu(&dp->nfree, -nblocks);
3312 
3313 	/* reconstruct summary tree */
3314 	dbInitDmapTree(dp);
3315 
3316 	BMAP_LOCK(bmp);
3317 
3318 	/* if this allocation group is completely free,
3319 	 * update the highest active allocation group number
3320 	 * if this allocation group is the new max.
3321 	 */
3322 	agno = blkno >> bmp->db_agl2size;
3323 	if (agno > bmp->db_maxag)
3324 		bmp->db_maxag = agno;
3325 
3326 	/* update the free count for the allocation group and map */
3327 	bmp->db_agfree[agno] -= nblocks;
3328 	bmp->db_nfree -= nblocks;
3329 
3330 	BMAP_UNLOCK(bmp);
3331 
3332 	/* if the root has not changed, done. */
3333 	if (tp->stree[ROOT] == oldroot)
3334 		return (0);
3335 
3336 	/* root changed. bubble the change up to the dmap control pages.
3337 	 * if the adjustment of the upper level control pages fails,
3338 	 * backout the bit allocation (thus making everything consistent).
3339 	 */
3340 	if ((rc = dbAdjCtl(bmp, blkno, tp->stree[ROOT], 1, 0)))
3341 		dbFreeBits(bmp, dp, blkno, nblocks);
3342 
3343 	return (rc);
3344 }
3345 
3346 
3347 /*
3348  * NAME:	dbExtendFS()
3349  *
3350  * FUNCTION:	extend bmap from blkno for nblocks;
3351  *		dbExtendFS() updates bmap ready for dbAllocBottomUp();
3352  *
3353  * L2
3354  *  |
3355  *   L1---------------------------------L1
3356  *    |					 |
3357  *     L0---------L0---------L0		  L0---------L0---------L0
3358  *      |	   |	      |		   |	      |		 |
3359  *	 d0,...,dn  d0,...,dn  d0,...,dn    d0,...,dn  d0,...,dn  d0,.,dm;
3360  * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm
3361  *
3362  * <---old---><----------------------------extend----------------------->
3363  */
3364 int dbExtendFS(struct inode *ipbmap, s64 blkno,	s64 nblocks)
3365 {
3366 	struct jfs_sb_info *sbi = JFS_SBI(ipbmap->i_sb);
3367 	int nbperpage = sbi->nbperpage;
3368 	int i, i0 = true, j, j0 = true, k, n;
3369 	s64 newsize;
3370 	s64 p;
3371 	struct metapage *mp, *l2mp, *l1mp = NULL, *l0mp = NULL;
3372 	struct dmapctl *l2dcp, *l1dcp, *l0dcp;
3373 	struct dmap *dp;
3374 	s8 *l0leaf, *l1leaf, *l2leaf;
3375 	struct bmap *bmp = sbi->bmap;
3376 	int agno, l2agsize, oldl2agsize;
3377 	s64 ag_rem;
3378 
3379 	newsize = blkno + nblocks;
3380 
3381 	jfs_info("dbExtendFS: blkno:%Ld nblocks:%Ld newsize:%Ld",
3382 		 (long long) blkno, (long long) nblocks, (long long) newsize);
3383 
3384 	/*
3385 	 *	initialize bmap control page.
3386 	 *
3387 	 * all the data in bmap control page should exclude
3388 	 * the mkfs hidden dmap page.
3389 	 */
3390 
3391 	/* update mapsize */
3392 	bmp->db_mapsize = newsize;
3393 	bmp->db_maxlevel = BMAPSZTOLEV(bmp->db_mapsize);
3394 
3395 	/* compute new AG size */
3396 	l2agsize = dbGetL2AGSize(newsize);
3397 	oldl2agsize = bmp->db_agl2size;
3398 
3399 	bmp->db_agl2size = l2agsize;
3400 	bmp->db_agsize = 1 << l2agsize;
3401 
3402 	/* compute new number of AG */
3403 	agno = bmp->db_numag;
3404 	bmp->db_numag = newsize >> l2agsize;
3405 	bmp->db_numag += ((u32) newsize % (u32) bmp->db_agsize) ? 1 : 0;
3406 
3407 	/*
3408 	 *	reconfigure db_agfree[]
3409 	 * from old AG configuration to new AG configuration;
3410 	 *
3411 	 * coalesce contiguous k (newAGSize/oldAGSize) AGs;
3412 	 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn;
3413 	 * note: new AG size = old AG size * (2**x).
3414 	 */
3415 	if (l2agsize == oldl2agsize)
3416 		goto extend;
3417 	k = 1 << (l2agsize - oldl2agsize);
3418 	ag_rem = bmp->db_agfree[0];	/* save agfree[0] */
3419 	for (i = 0, n = 0; i < agno; n++) {
3420 		bmp->db_agfree[n] = 0;	/* init collection point */
3421 
3422 		/* coalesce contiguous k AGs; */
3423 		for (j = 0; j < k && i < agno; j++, i++) {
3424 			/* merge AGi to AGn */
3425 			bmp->db_agfree[n] += bmp->db_agfree[i];
3426 		}
3427 	}
3428 	bmp->db_agfree[0] += ag_rem;	/* restore agfree[0] */
3429 
3430 	for (; n < MAXAG; n++)
3431 		bmp->db_agfree[n] = 0;
3432 
3433 	/*
3434 	 * update highest active ag number
3435 	 */
3436 
3437 	bmp->db_maxag = bmp->db_maxag / k;
3438 
3439 	/*
3440 	 *	extend bmap
3441 	 *
3442 	 * update bit maps and corresponding level control pages;
3443 	 * global control page db_nfree, db_agfree[agno], db_maxfreebud;
3444 	 */
3445       extend:
3446 	/* get L2 page */
3447 	p = BMAPBLKNO + nbperpage;	/* L2 page */
3448 	l2mp = read_metapage(ipbmap, p, PSIZE, 0);
3449 	if (!l2mp) {
3450 		jfs_error(ipbmap->i_sb, "L2 page could not be read\n");
3451 		return -EIO;
3452 	}
3453 	l2dcp = (struct dmapctl *) l2mp->data;
3454 
3455 	/* compute start L1 */
3456 	k = blkno >> L2MAXL1SIZE;
3457 	l2leaf = l2dcp->stree + CTLLEAFIND + k;
3458 	p = BLKTOL1(blkno, sbi->l2nbperpage);	/* L1 page */
3459 
3460 	/*
3461 	 * extend each L1 in L2
3462 	 */
3463 	for (; k < LPERCTL; k++, p += nbperpage) {
3464 		/* get L1 page */
3465 		if (j0) {
3466 			/* read in L1 page: (blkno & (MAXL1SIZE - 1)) */
3467 			l1mp = read_metapage(ipbmap, p, PSIZE, 0);
3468 			if (l1mp == NULL)
3469 				goto errout;
3470 			l1dcp = (struct dmapctl *) l1mp->data;
3471 
3472 			/* compute start L0 */
3473 			j = (blkno & (MAXL1SIZE - 1)) >> L2MAXL0SIZE;
3474 			l1leaf = l1dcp->stree + CTLLEAFIND + j;
3475 			p = BLKTOL0(blkno, sbi->l2nbperpage);
3476 			j0 = false;
3477 		} else {
3478 			/* assign/init L1 page */
3479 			l1mp = get_metapage(ipbmap, p, PSIZE, 0);
3480 			if (l1mp == NULL)
3481 				goto errout;
3482 
3483 			l1dcp = (struct dmapctl *) l1mp->data;
3484 
3485 			/* compute start L0 */
3486 			j = 0;
3487 			l1leaf = l1dcp->stree + CTLLEAFIND;
3488 			p += nbperpage;	/* 1st L0 of L1.k */
3489 		}
3490 
3491 		/*
3492 		 * extend each L0 in L1
3493 		 */
3494 		for (; j < LPERCTL; j++) {
3495 			/* get L0 page */
3496 			if (i0) {
3497 				/* read in L0 page: (blkno & (MAXL0SIZE - 1)) */
3498 
3499 				l0mp = read_metapage(ipbmap, p, PSIZE, 0);
3500 				if (l0mp == NULL)
3501 					goto errout;
3502 				l0dcp = (struct dmapctl *) l0mp->data;
3503 
3504 				/* compute start dmap */
3505 				i = (blkno & (MAXL0SIZE - 1)) >>
3506 				    L2BPERDMAP;
3507 				l0leaf = l0dcp->stree + CTLLEAFIND + i;
3508 				p = BLKTODMAP(blkno,
3509 					      sbi->l2nbperpage);
3510 				i0 = false;
3511 			} else {
3512 				/* assign/init L0 page */
3513 				l0mp = get_metapage(ipbmap, p, PSIZE, 0);
3514 				if (l0mp == NULL)
3515 					goto errout;
3516 
3517 				l0dcp = (struct dmapctl *) l0mp->data;
3518 
3519 				/* compute start dmap */
3520 				i = 0;
3521 				l0leaf = l0dcp->stree + CTLLEAFIND;
3522 				p += nbperpage;	/* 1st dmap of L0.j */
3523 			}
3524 
3525 			/*
3526 			 * extend each dmap in L0
3527 			 */
3528 			for (; i < LPERCTL; i++) {
3529 				/*
3530 				 * reconstruct the dmap page, and
3531 				 * initialize corresponding parent L0 leaf
3532 				 */
3533 				if ((n = blkno & (BPERDMAP - 1))) {
3534 					/* read in dmap page: */
3535 					mp = read_metapage(ipbmap, p,
3536 							   PSIZE, 0);
3537 					if (mp == NULL)
3538 						goto errout;
3539 					n = min(nblocks, (s64)BPERDMAP - n);
3540 				} else {
3541 					/* assign/init dmap page */
3542 					mp = read_metapage(ipbmap, p,
3543 							   PSIZE, 0);
3544 					if (mp == NULL)
3545 						goto errout;
3546 
3547 					n = min_t(s64, nblocks, BPERDMAP);
3548 				}
3549 
3550 				dp = (struct dmap *) mp->data;
3551 				*l0leaf = dbInitDmap(dp, blkno, n);
3552 
3553 				bmp->db_nfree += n;
3554 				agno = le64_to_cpu(dp->start) >> l2agsize;
3555 				bmp->db_agfree[agno] += n;
3556 
3557 				write_metapage(mp);
3558 
3559 				l0leaf++;
3560 				p += nbperpage;
3561 
3562 				blkno += n;
3563 				nblocks -= n;
3564 				if (nblocks == 0)
3565 					break;
3566 			}	/* for each dmap in a L0 */
3567 
3568 			/*
3569 			 * build current L0 page from its leaves, and
3570 			 * initialize corresponding parent L1 leaf
3571 			 */
3572 			*l1leaf = dbInitDmapCtl(l0dcp, 0, ++i);
3573 			write_metapage(l0mp);
3574 			l0mp = NULL;
3575 
3576 			if (nblocks)
3577 				l1leaf++;	/* continue for next L0 */
3578 			else {
3579 				/* more than 1 L0 ? */
3580 				if (j > 0)
3581 					break;	/* build L1 page */
3582 				else {
3583 					/* summarize in global bmap page */
3584 					bmp->db_maxfreebud = *l1leaf;
3585 					release_metapage(l1mp);
3586 					release_metapage(l2mp);
3587 					goto finalize;
3588 				}
3589 			}
3590 		}		/* for each L0 in a L1 */
3591 
3592 		/*
3593 		 * build current L1 page from its leaves, and
3594 		 * initialize corresponding parent L2 leaf
3595 		 */
3596 		*l2leaf = dbInitDmapCtl(l1dcp, 1, ++j);
3597 		write_metapage(l1mp);
3598 		l1mp = NULL;
3599 
3600 		if (nblocks)
3601 			l2leaf++;	/* continue for next L1 */
3602 		else {
3603 			/* more than 1 L1 ? */
3604 			if (k > 0)
3605 				break;	/* build L2 page */
3606 			else {
3607 				/* summarize in global bmap page */
3608 				bmp->db_maxfreebud = *l2leaf;
3609 				release_metapage(l2mp);
3610 				goto finalize;
3611 			}
3612 		}
3613 	}			/* for each L1 in a L2 */
3614 
3615 	jfs_error(ipbmap->i_sb, "function has not returned as expected\n");
3616 errout:
3617 	if (l0mp)
3618 		release_metapage(l0mp);
3619 	if (l1mp)
3620 		release_metapage(l1mp);
3621 	release_metapage(l2mp);
3622 	return -EIO;
3623 
3624 	/*
3625 	 *	finalize bmap control page
3626 	 */
3627 finalize:
3628 
3629 	return 0;
3630 }
3631 
3632 
3633 /*
3634  *	dbFinalizeBmap()
3635  */
3636 void dbFinalizeBmap(struct inode *ipbmap)
3637 {
3638 	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
3639 	int actags, inactags, l2nl;
3640 	s64 ag_rem, actfree, inactfree, avgfree;
3641 	int i, n;
3642 
3643 	/*
3644 	 *	finalize bmap control page
3645 	 */
3646 //finalize:
3647 	/*
3648 	 * compute db_agpref: preferred ag to allocate from
3649 	 * (the leftmost ag with average free space in it);
3650 	 */
3651 //agpref:
3652 	/* get the number of active ags and inactive ags */
3653 	actags = bmp->db_maxag + 1;
3654 	inactags = bmp->db_numag - actags;
3655 	ag_rem = bmp->db_mapsize & (bmp->db_agsize - 1);	/* ??? */
3656 
3657 	/* determine how many blocks are in the inactive allocation
3658 	 * groups. in doing this, we must account for the fact that
3659 	 * the rightmost group might be a partial group (i.e. file
3660 	 * system size is not a multiple of the group size).
3661 	 */
3662 	inactfree = (inactags && ag_rem) ?
3663 	    ((inactags - 1) << bmp->db_agl2size) + ag_rem
3664 	    : inactags << bmp->db_agl2size;
3665 
3666 	/* determine how many free blocks are in the active
3667 	 * allocation groups plus the average number of free blocks
3668 	 * within the active ags.
3669 	 */
3670 	actfree = bmp->db_nfree - inactfree;
3671 	avgfree = (u32) actfree / (u32) actags;
3672 
3673 	/* if the preferred allocation group has not average free space.
3674 	 * re-establish the preferred group as the leftmost
3675 	 * group with average free space.
3676 	 */
3677 	if (bmp->db_agfree[bmp->db_agpref] < avgfree) {
3678 		for (bmp->db_agpref = 0; bmp->db_agpref < actags;
3679 		     bmp->db_agpref++) {
3680 			if (bmp->db_agfree[bmp->db_agpref] >= avgfree)
3681 				break;
3682 		}
3683 		if (bmp->db_agpref >= bmp->db_numag) {
3684 			jfs_error(ipbmap->i_sb,
3685 				  "cannot find ag with average freespace\n");
3686 		}
3687 	}
3688 
3689 	/*
3690 	 * compute db_aglevel, db_agheight, db_width, db_agstart:
3691 	 * an ag is covered in aglevel dmapctl summary tree,
3692 	 * at agheight level height (from leaf) with agwidth number of nodes
3693 	 * each, which starts at agstart index node of the smmary tree node
3694 	 * array;
3695 	 */
3696 	bmp->db_aglevel = BMAPSZTOLEV(bmp->db_agsize);
3697 	l2nl =
3698 	    bmp->db_agl2size - (L2BPERDMAP + bmp->db_aglevel * L2LPERCTL);
3699 	bmp->db_agheight = l2nl >> 1;
3700 	bmp->db_agwidth = 1 << (l2nl - (bmp->db_agheight << 1));
3701 	for (i = 5 - bmp->db_agheight, bmp->db_agstart = 0, n = 1; i > 0;
3702 	     i--) {
3703 		bmp->db_agstart += n;
3704 		n <<= 2;
3705 	}
3706 
3707 }
3708 
3709 
3710 /*
3711  * NAME:	dbInitDmap()/ujfs_idmap_page()
3712  *
3713  * FUNCTION:	initialize working/persistent bitmap of the dmap page
3714  *		for the specified number of blocks:
3715  *
3716  *		at entry, the bitmaps had been initialized as free (ZEROS);
3717  *		The number of blocks will only account for the actually
3718  *		existing blocks. Blocks which don't actually exist in
3719  *		the aggregate will be marked as allocated (ONES);
3720  *
3721  * PARAMETERS:
3722  *	dp	- pointer to page of map
3723  *	nblocks	- number of blocks this page
3724  *
3725  * RETURNS: NONE
3726  */
3727 static int dbInitDmap(struct dmap * dp, s64 Blkno, int nblocks)
3728 {
3729 	int blkno, w, b, r, nw, nb, i;
3730 
3731 	/* starting block number within the dmap */
3732 	blkno = Blkno & (BPERDMAP - 1);
3733 
3734 	if (blkno == 0) {
3735 		dp->nblocks = dp->nfree = cpu_to_le32(nblocks);
3736 		dp->start = cpu_to_le64(Blkno);
3737 
3738 		if (nblocks == BPERDMAP) {
3739 			memset(&dp->wmap[0], 0, LPERDMAP * 4);
3740 			memset(&dp->pmap[0], 0, LPERDMAP * 4);
3741 			goto initTree;
3742 		}
3743 	} else {
3744 		le32_add_cpu(&dp->nblocks, nblocks);
3745 		le32_add_cpu(&dp->nfree, nblocks);
3746 	}
3747 
3748 	/* word number containing start block number */
3749 	w = blkno >> L2DBWORD;
3750 
3751 	/*
3752 	 * free the bits corresponding to the block range (ZEROS):
3753 	 * note: not all bits of the first and last words may be contained
3754 	 * within the block range.
3755 	 */
3756 	for (r = nblocks; r > 0; r -= nb, blkno += nb) {
3757 		/* number of bits preceding range to be freed in the word */
3758 		b = blkno & (DBWORD - 1);
3759 		/* number of bits to free in the word */
3760 		nb = min(r, DBWORD - b);
3761 
3762 		/* is partial word to be freed ? */
3763 		if (nb < DBWORD) {
3764 			/* free (set to 0) from the bitmap word */
3765 			dp->wmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3766 						     >> b));
3767 			dp->pmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3768 						     >> b));
3769 
3770 			/* skip the word freed */
3771 			w++;
3772 		} else {
3773 			/* free (set to 0) contiguous bitmap words */
3774 			nw = r >> L2DBWORD;
3775 			memset(&dp->wmap[w], 0, nw * 4);
3776 			memset(&dp->pmap[w], 0, nw * 4);
3777 
3778 			/* skip the words freed */
3779 			nb = nw << L2DBWORD;
3780 			w += nw;
3781 		}
3782 	}
3783 
3784 	/*
3785 	 * mark bits following the range to be freed (non-existing
3786 	 * blocks) as allocated (ONES)
3787 	 */
3788 
3789 	if (blkno == BPERDMAP)
3790 		goto initTree;
3791 
3792 	/* the first word beyond the end of existing blocks */
3793 	w = blkno >> L2DBWORD;
3794 
3795 	/* does nblocks fall on a 32-bit boundary ? */
3796 	b = blkno & (DBWORD - 1);
3797 	if (b) {
3798 		/* mark a partial word allocated */
3799 		dp->wmap[w] = dp->pmap[w] = cpu_to_le32(ONES >> b);
3800 		w++;
3801 	}
3802 
3803 	/* set the rest of the words in the page to allocated (ONES) */
3804 	for (i = w; i < LPERDMAP; i++)
3805 		dp->pmap[i] = dp->wmap[i] = cpu_to_le32(ONES);
3806 
3807 	/*
3808 	 * init tree
3809 	 */
3810       initTree:
3811 	return (dbInitDmapTree(dp));
3812 }
3813 
3814 
3815 /*
3816  * NAME:	dbInitDmapTree()/ujfs_complete_dmap()
3817  *
3818  * FUNCTION:	initialize summary tree of the specified dmap:
3819  *
3820  *		at entry, bitmap of the dmap has been initialized;
3821  *
3822  * PARAMETERS:
3823  *	dp	- dmap to complete
3824  *	blkno	- starting block number for this dmap
3825  *	treemax	- will be filled in with max free for this dmap
3826  *
3827  * RETURNS:	max free string at the root of the tree
3828  */
3829 static int dbInitDmapTree(struct dmap * dp)
3830 {
3831 	struct dmaptree *tp;
3832 	s8 *cp;
3833 	int i;
3834 
3835 	/* init fixed info of tree */
3836 	tp = &dp->tree;
3837 	tp->nleafs = cpu_to_le32(LPERDMAP);
3838 	tp->l2nleafs = cpu_to_le32(L2LPERDMAP);
3839 	tp->leafidx = cpu_to_le32(LEAFIND);
3840 	tp->height = cpu_to_le32(4);
3841 	tp->budmin = BUDMIN;
3842 
3843 	/* init each leaf from corresponding wmap word:
3844 	 * note: leaf is set to NOFREE(-1) if all blocks of corresponding
3845 	 * bitmap word are allocated.
3846 	 */
3847 	cp = tp->stree + le32_to_cpu(tp->leafidx);
3848 	for (i = 0; i < LPERDMAP; i++)
3849 		*cp++ = dbMaxBud((u8 *) & dp->wmap[i]);
3850 
3851 	/* build the dmap's binary buddy summary tree */
3852 	return (dbInitTree(tp));
3853 }
3854 
3855 
3856 /*
3857  * NAME:	dbInitTree()/ujfs_adjtree()
3858  *
3859  * FUNCTION:	initialize binary buddy summary tree of a dmap or dmapctl.
3860  *
3861  *		at entry, the leaves of the tree has been initialized
3862  *		from corresponding bitmap word or root of summary tree
3863  *		of the child control page;
3864  *		configure binary buddy system at the leaf level, then
3865  *		bubble up the values of the leaf nodes up the tree.
3866  *
3867  * PARAMETERS:
3868  *	cp	- Pointer to the root of the tree
3869  *	l2leaves- Number of leaf nodes as a power of 2
3870  *	l2min	- Number of blocks that can be covered by a leaf
3871  *		  as a power of 2
3872  *
3873  * RETURNS: max free string at the root of the tree
3874  */
3875 static int dbInitTree(struct dmaptree * dtp)
3876 {
3877 	int l2max, l2free, bsize, nextb, i;
3878 	int child, parent, nparent;
3879 	s8 *tp, *cp, *cp1;
3880 
3881 	tp = dtp->stree;
3882 
3883 	/* Determine the maximum free string possible for the leaves */
3884 	l2max = le32_to_cpu(dtp->l2nleafs) + dtp->budmin;
3885 
3886 	/*
3887 	 * configure the leaf level into binary buddy system
3888 	 *
3889 	 * Try to combine buddies starting with a buddy size of 1
3890 	 * (i.e. two leaves). At a buddy size of 1 two buddy leaves
3891 	 * can be combined if both buddies have a maximum free of l2min;
3892 	 * the combination will result in the left-most buddy leaf having
3893 	 * a maximum free of l2min+1.
3894 	 * After processing all buddies for a given size, process buddies
3895 	 * at the next higher buddy size (i.e. current size * 2) and
3896 	 * the next maximum free (current free + 1).
3897 	 * This continues until the maximum possible buddy combination
3898 	 * yields maximum free.
3899 	 */
3900 	for (l2free = dtp->budmin, bsize = 1; l2free < l2max;
3901 	     l2free++, bsize = nextb) {
3902 		/* get next buddy size == current buddy pair size */
3903 		nextb = bsize << 1;
3904 
3905 		/* scan each adjacent buddy pair at current buddy size */
3906 		for (i = 0, cp = tp + le32_to_cpu(dtp->leafidx);
3907 		     i < le32_to_cpu(dtp->nleafs);
3908 		     i += nextb, cp += nextb) {
3909 			/* coalesce if both adjacent buddies are max free */
3910 			if (*cp == l2free && *(cp + bsize) == l2free) {
3911 				*cp = l2free + 1;	/* left take right */
3912 				*(cp + bsize) = -1;	/* right give left */
3913 			}
3914 		}
3915 	}
3916 
3917 	/*
3918 	 * bubble summary information of leaves up the tree.
3919 	 *
3920 	 * Starting at the leaf node level, the four nodes described by
3921 	 * the higher level parent node are compared for a maximum free and
3922 	 * this maximum becomes the value of the parent node.
3923 	 * when all lower level nodes are processed in this fashion then
3924 	 * move up to the next level (parent becomes a lower level node) and
3925 	 * continue the process for that level.
3926 	 */
3927 	for (child = le32_to_cpu(dtp->leafidx),
3928 	     nparent = le32_to_cpu(dtp->nleafs) >> 2;
3929 	     nparent > 0; nparent >>= 2, child = parent) {
3930 		/* get index of 1st node of parent level */
3931 		parent = (child - 1) >> 2;
3932 
3933 		/* set the value of the parent node as the maximum
3934 		 * of the four nodes of the current level.
3935 		 */
3936 		for (i = 0, cp = tp + child, cp1 = tp + parent;
3937 		     i < nparent; i++, cp += 4, cp1++)
3938 			*cp1 = TREEMAX(cp);
3939 	}
3940 
3941 	return (*tp);
3942 }
3943 
3944 
3945 /*
3946  *	dbInitDmapCtl()
3947  *
3948  * function: initialize dmapctl page
3949  */
3950 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i)
3951 {				/* start leaf index not covered by range */
3952 	s8 *cp;
3953 
3954 	dcp->nleafs = cpu_to_le32(LPERCTL);
3955 	dcp->l2nleafs = cpu_to_le32(L2LPERCTL);
3956 	dcp->leafidx = cpu_to_le32(CTLLEAFIND);
3957 	dcp->height = cpu_to_le32(5);
3958 	dcp->budmin = L2BPERDMAP + L2LPERCTL * level;
3959 
3960 	/*
3961 	 * initialize the leaves of current level that were not covered
3962 	 * by the specified input block range (i.e. the leaves have no
3963 	 * low level dmapctl or dmap).
3964 	 */
3965 	cp = &dcp->stree[CTLLEAFIND + i];
3966 	for (; i < LPERCTL; i++)
3967 		*cp++ = NOFREE;
3968 
3969 	/* build the dmap's binary buddy summary tree */
3970 	return (dbInitTree((struct dmaptree *) dcp));
3971 }
3972 
3973 
3974 /*
3975  * NAME:	dbGetL2AGSize()/ujfs_getagl2size()
3976  *
3977  * FUNCTION:	Determine log2(allocation group size) from aggregate size
3978  *
3979  * PARAMETERS:
3980  *	nblocks	- Number of blocks in aggregate
3981  *
3982  * RETURNS: log2(allocation group size) in aggregate blocks
3983  */
3984 static int dbGetL2AGSize(s64 nblocks)
3985 {
3986 	s64 sz;
3987 	s64 m;
3988 	int l2sz;
3989 
3990 	if (nblocks < BPERDMAP * MAXAG)
3991 		return (L2BPERDMAP);
3992 
3993 	/* round up aggregate size to power of 2 */
3994 	m = ((u64) 1 << (64 - 1));
3995 	for (l2sz = 64; l2sz >= 0; l2sz--, m >>= 1) {
3996 		if (m & nblocks)
3997 			break;
3998 	}
3999 
4000 	sz = (s64) 1 << l2sz;
4001 	if (sz < nblocks)
4002 		l2sz += 1;
4003 
4004 	/* agsize = roundupSize/max_number_of_ag */
4005 	return (l2sz - L2MAXAG);
4006 }
4007 
4008 
4009 /*
4010  * NAME:	dbMapFileSizeToMapSize()
4011  *
4012  * FUNCTION:	compute number of blocks the block allocation map file
4013  *		can cover from the map file size;
4014  *
4015  * RETURNS:	Number of blocks which can be covered by this block map file;
4016  */
4017 
4018 /*
4019  * maximum number of map pages at each level including control pages
4020  */
4021 #define MAXL0PAGES	(1 + LPERCTL)
4022 #define MAXL1PAGES	(1 + LPERCTL * MAXL0PAGES)
4023 
4024 /*
4025  * convert number of map pages to the zero origin top dmapctl level
4026  */
4027 #define BMAPPGTOLEV(npages)	\
4028 	(((npages) <= 3 + MAXL0PAGES) ? 0 : \
4029 	 ((npages) <= 2 + MAXL1PAGES) ? 1 : 2)
4030 
4031 s64 dbMapFileSizeToMapSize(struct inode * ipbmap)
4032 {
4033 	struct super_block *sb = ipbmap->i_sb;
4034 	s64 nblocks;
4035 	s64 npages, ndmaps;
4036 	int level, i;
4037 	int complete, factor;
4038 
4039 	nblocks = ipbmap->i_size >> JFS_SBI(sb)->l2bsize;
4040 	npages = nblocks >> JFS_SBI(sb)->l2nbperpage;
4041 	level = BMAPPGTOLEV(npages);
4042 
4043 	/* At each level, accumulate the number of dmap pages covered by
4044 	 * the number of full child levels below it;
4045 	 * repeat for the last incomplete child level.
4046 	 */
4047 	ndmaps = 0;
4048 	npages--;		/* skip the first global control page */
4049 	/* skip higher level control pages above top level covered by map */
4050 	npages -= (2 - level);
4051 	npages--;		/* skip top level's control page */
4052 	for (i = level; i >= 0; i--) {
4053 		factor =
4054 		    (i == 2) ? MAXL1PAGES : ((i == 1) ? MAXL0PAGES : 1);
4055 		complete = (u32) npages / factor;
4056 		ndmaps += complete * ((i == 2) ? LPERCTL * LPERCTL :
4057 				      ((i == 1) ? LPERCTL : 1));
4058 
4059 		/* pages in last/incomplete child */
4060 		npages = (u32) npages % factor;
4061 		/* skip incomplete child's level control page */
4062 		npages--;
4063 	}
4064 
4065 	/* convert the number of dmaps into the number of blocks
4066 	 * which can be covered by the dmaps;
4067 	 */
4068 	nblocks = ndmaps << L2BPERDMAP;
4069 
4070 	return (nblocks);
4071 }
4072