xref: /openbmc/linux/fs/xfs/libxfs/xfs_ialloc.c (revision 3dc4b6fb)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2000-2002,2005 Silicon Graphics, Inc.
4  * All Rights Reserved.
5  */
6 #include "xfs.h"
7 #include "xfs_fs.h"
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_bit.h"
13 #include "xfs_sb.h"
14 #include "xfs_mount.h"
15 #include "xfs_inode.h"
16 #include "xfs_btree.h"
17 #include "xfs_ialloc.h"
18 #include "xfs_ialloc_btree.h"
19 #include "xfs_alloc.h"
20 #include "xfs_errortag.h"
21 #include "xfs_error.h"
22 #include "xfs_bmap.h"
23 #include "xfs_trans.h"
24 #include "xfs_buf_item.h"
25 #include "xfs_icreate_item.h"
26 #include "xfs_icache.h"
27 #include "xfs_trace.h"
28 #include "xfs_log.h"
29 #include "xfs_rmap.h"
30 
31 /*
32  * Lookup a record by ino in the btree given by cur.
33  */
34 int					/* error */
35 xfs_inobt_lookup(
36 	struct xfs_btree_cur	*cur,	/* btree cursor */
37 	xfs_agino_t		ino,	/* starting inode of chunk */
38 	xfs_lookup_t		dir,	/* <=, >=, == */
39 	int			*stat)	/* success/failure */
40 {
41 	cur->bc_rec.i.ir_startino = ino;
42 	cur->bc_rec.i.ir_holemask = 0;
43 	cur->bc_rec.i.ir_count = 0;
44 	cur->bc_rec.i.ir_freecount = 0;
45 	cur->bc_rec.i.ir_free = 0;
46 	return xfs_btree_lookup(cur, dir, stat);
47 }
48 
49 /*
50  * Update the record referred to by cur to the value given.
51  * This either works (return 0) or gets an EFSCORRUPTED error.
52  */
53 STATIC int				/* error */
54 xfs_inobt_update(
55 	struct xfs_btree_cur	*cur,	/* btree cursor */
56 	xfs_inobt_rec_incore_t	*irec)	/* btree record */
57 {
58 	union xfs_btree_rec	rec;
59 
60 	rec.inobt.ir_startino = cpu_to_be32(irec->ir_startino);
61 	if (xfs_sb_version_hassparseinodes(&cur->bc_mp->m_sb)) {
62 		rec.inobt.ir_u.sp.ir_holemask = cpu_to_be16(irec->ir_holemask);
63 		rec.inobt.ir_u.sp.ir_count = irec->ir_count;
64 		rec.inobt.ir_u.sp.ir_freecount = irec->ir_freecount;
65 	} else {
66 		/* ir_holemask/ir_count not supported on-disk */
67 		rec.inobt.ir_u.f.ir_freecount = cpu_to_be32(irec->ir_freecount);
68 	}
69 	rec.inobt.ir_free = cpu_to_be64(irec->ir_free);
70 	return xfs_btree_update(cur, &rec);
71 }
72 
73 /* Convert on-disk btree record to incore inobt record. */
74 void
75 xfs_inobt_btrec_to_irec(
76 	struct xfs_mount		*mp,
77 	union xfs_btree_rec		*rec,
78 	struct xfs_inobt_rec_incore	*irec)
79 {
80 	irec->ir_startino = be32_to_cpu(rec->inobt.ir_startino);
81 	if (xfs_sb_version_hassparseinodes(&mp->m_sb)) {
82 		irec->ir_holemask = be16_to_cpu(rec->inobt.ir_u.sp.ir_holemask);
83 		irec->ir_count = rec->inobt.ir_u.sp.ir_count;
84 		irec->ir_freecount = rec->inobt.ir_u.sp.ir_freecount;
85 	} else {
86 		/*
87 		 * ir_holemask/ir_count not supported on-disk. Fill in hardcoded
88 		 * values for full inode chunks.
89 		 */
90 		irec->ir_holemask = XFS_INOBT_HOLEMASK_FULL;
91 		irec->ir_count = XFS_INODES_PER_CHUNK;
92 		irec->ir_freecount =
93 				be32_to_cpu(rec->inobt.ir_u.f.ir_freecount);
94 	}
95 	irec->ir_free = be64_to_cpu(rec->inobt.ir_free);
96 }
97 
98 /*
99  * Get the data from the pointed-to record.
100  */
101 int
102 xfs_inobt_get_rec(
103 	struct xfs_btree_cur		*cur,
104 	struct xfs_inobt_rec_incore	*irec,
105 	int				*stat)
106 {
107 	struct xfs_mount		*mp = cur->bc_mp;
108 	xfs_agnumber_t			agno = cur->bc_private.a.agno;
109 	union xfs_btree_rec		*rec;
110 	int				error;
111 	uint64_t			realfree;
112 
113 	error = xfs_btree_get_rec(cur, &rec, stat);
114 	if (error || *stat == 0)
115 		return error;
116 
117 	xfs_inobt_btrec_to_irec(mp, rec, irec);
118 
119 	if (!xfs_verify_agino(mp, agno, irec->ir_startino))
120 		goto out_bad_rec;
121 	if (irec->ir_count < XFS_INODES_PER_HOLEMASK_BIT ||
122 	    irec->ir_count > XFS_INODES_PER_CHUNK)
123 		goto out_bad_rec;
124 	if (irec->ir_freecount > XFS_INODES_PER_CHUNK)
125 		goto out_bad_rec;
126 
127 	/* if there are no holes, return the first available offset */
128 	if (!xfs_inobt_issparse(irec->ir_holemask))
129 		realfree = irec->ir_free;
130 	else
131 		realfree = irec->ir_free & xfs_inobt_irec_to_allocmask(irec);
132 	if (hweight64(realfree) != irec->ir_freecount)
133 		goto out_bad_rec;
134 
135 	return 0;
136 
137 out_bad_rec:
138 	xfs_warn(mp,
139 		"%s Inode BTree record corruption in AG %d detected!",
140 		cur->bc_btnum == XFS_BTNUM_INO ? "Used" : "Free", agno);
141 	xfs_warn(mp,
142 "start inode 0x%x, count 0x%x, free 0x%x freemask 0x%llx, holemask 0x%x",
143 		irec->ir_startino, irec->ir_count, irec->ir_freecount,
144 		irec->ir_free, irec->ir_holemask);
145 	return -EFSCORRUPTED;
146 }
147 
148 /*
149  * Insert a single inobt record. Cursor must already point to desired location.
150  */
151 int
152 xfs_inobt_insert_rec(
153 	struct xfs_btree_cur	*cur,
154 	uint16_t		holemask,
155 	uint8_t			count,
156 	int32_t			freecount,
157 	xfs_inofree_t		free,
158 	int			*stat)
159 {
160 	cur->bc_rec.i.ir_holemask = holemask;
161 	cur->bc_rec.i.ir_count = count;
162 	cur->bc_rec.i.ir_freecount = freecount;
163 	cur->bc_rec.i.ir_free = free;
164 	return xfs_btree_insert(cur, stat);
165 }
166 
167 /*
168  * Insert records describing a newly allocated inode chunk into the inobt.
169  */
170 STATIC int
171 xfs_inobt_insert(
172 	struct xfs_mount	*mp,
173 	struct xfs_trans	*tp,
174 	struct xfs_buf		*agbp,
175 	xfs_agino_t		newino,
176 	xfs_agino_t		newlen,
177 	xfs_btnum_t		btnum)
178 {
179 	struct xfs_btree_cur	*cur;
180 	struct xfs_agi		*agi = XFS_BUF_TO_AGI(agbp);
181 	xfs_agnumber_t		agno = be32_to_cpu(agi->agi_seqno);
182 	xfs_agino_t		thisino;
183 	int			i;
184 	int			error;
185 
186 	cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, btnum);
187 
188 	for (thisino = newino;
189 	     thisino < newino + newlen;
190 	     thisino += XFS_INODES_PER_CHUNK) {
191 		error = xfs_inobt_lookup(cur, thisino, XFS_LOOKUP_EQ, &i);
192 		if (error) {
193 			xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
194 			return error;
195 		}
196 		ASSERT(i == 0);
197 
198 		error = xfs_inobt_insert_rec(cur, XFS_INOBT_HOLEMASK_FULL,
199 					     XFS_INODES_PER_CHUNK,
200 					     XFS_INODES_PER_CHUNK,
201 					     XFS_INOBT_ALL_FREE, &i);
202 		if (error) {
203 			xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
204 			return error;
205 		}
206 		ASSERT(i == 1);
207 	}
208 
209 	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
210 
211 	return 0;
212 }
213 
214 /*
215  * Verify that the number of free inodes in the AGI is correct.
216  */
217 #ifdef DEBUG
218 STATIC int
219 xfs_check_agi_freecount(
220 	struct xfs_btree_cur	*cur,
221 	struct xfs_agi		*agi)
222 {
223 	if (cur->bc_nlevels == 1) {
224 		xfs_inobt_rec_incore_t rec;
225 		int		freecount = 0;
226 		int		error;
227 		int		i;
228 
229 		error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
230 		if (error)
231 			return error;
232 
233 		do {
234 			error = xfs_inobt_get_rec(cur, &rec, &i);
235 			if (error)
236 				return error;
237 
238 			if (i) {
239 				freecount += rec.ir_freecount;
240 				error = xfs_btree_increment(cur, 0, &i);
241 				if (error)
242 					return error;
243 			}
244 		} while (i == 1);
245 
246 		if (!XFS_FORCED_SHUTDOWN(cur->bc_mp))
247 			ASSERT(freecount == be32_to_cpu(agi->agi_freecount));
248 	}
249 	return 0;
250 }
251 #else
252 #define xfs_check_agi_freecount(cur, agi)	0
253 #endif
254 
255 /*
256  * Initialise a new set of inodes. When called without a transaction context
257  * (e.g. from recovery) we initiate a delayed write of the inode buffers rather
258  * than logging them (which in a transaction context puts them into the AIL
259  * for writeback rather than the xfsbufd queue).
260  */
261 int
262 xfs_ialloc_inode_init(
263 	struct xfs_mount	*mp,
264 	struct xfs_trans	*tp,
265 	struct list_head	*buffer_list,
266 	int			icount,
267 	xfs_agnumber_t		agno,
268 	xfs_agblock_t		agbno,
269 	xfs_agblock_t		length,
270 	unsigned int		gen)
271 {
272 	struct xfs_buf		*fbuf;
273 	struct xfs_dinode	*free;
274 	int			nbufs;
275 	int			version;
276 	int			i, j;
277 	xfs_daddr_t		d;
278 	xfs_ino_t		ino = 0;
279 
280 	/*
281 	 * Loop over the new block(s), filling in the inodes.  For small block
282 	 * sizes, manipulate the inodes in buffers  which are multiples of the
283 	 * blocks size.
284 	 */
285 	nbufs = length / M_IGEO(mp)->blocks_per_cluster;
286 
287 	/*
288 	 * Figure out what version number to use in the inodes we create.  If
289 	 * the superblock version has caught up to the one that supports the new
290 	 * inode format, then use the new inode version.  Otherwise use the old
291 	 * version so that old kernels will continue to be able to use the file
292 	 * system.
293 	 *
294 	 * For v3 inodes, we also need to write the inode number into the inode,
295 	 * so calculate the first inode number of the chunk here as
296 	 * XFS_AGB_TO_AGINO() only works within a filesystem block, not
297 	 * across multiple filesystem blocks (such as a cluster) and so cannot
298 	 * be used in the cluster buffer loop below.
299 	 *
300 	 * Further, because we are writing the inode directly into the buffer
301 	 * and calculating a CRC on the entire inode, we have ot log the entire
302 	 * inode so that the entire range the CRC covers is present in the log.
303 	 * That means for v3 inode we log the entire buffer rather than just the
304 	 * inode cores.
305 	 */
306 	if (xfs_sb_version_hascrc(&mp->m_sb)) {
307 		version = 3;
308 		ino = XFS_AGINO_TO_INO(mp, agno, XFS_AGB_TO_AGINO(mp, agbno));
309 
310 		/*
311 		 * log the initialisation that is about to take place as an
312 		 * logical operation. This means the transaction does not
313 		 * need to log the physical changes to the inode buffers as log
314 		 * recovery will know what initialisation is actually needed.
315 		 * Hence we only need to log the buffers as "ordered" buffers so
316 		 * they track in the AIL as if they were physically logged.
317 		 */
318 		if (tp)
319 			xfs_icreate_log(tp, agno, agbno, icount,
320 					mp->m_sb.sb_inodesize, length, gen);
321 	} else
322 		version = 2;
323 
324 	for (j = 0; j < nbufs; j++) {
325 		/*
326 		 * Get the block.
327 		 */
328 		d = XFS_AGB_TO_DADDR(mp, agno, agbno +
329 				(j * M_IGEO(mp)->blocks_per_cluster));
330 		fbuf = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
331 					 mp->m_bsize *
332 					 M_IGEO(mp)->blocks_per_cluster,
333 					 XBF_UNMAPPED);
334 		if (!fbuf)
335 			return -ENOMEM;
336 
337 		/* Initialize the inode buffers and log them appropriately. */
338 		fbuf->b_ops = &xfs_inode_buf_ops;
339 		xfs_buf_zero(fbuf, 0, BBTOB(fbuf->b_length));
340 		for (i = 0; i < M_IGEO(mp)->inodes_per_cluster; i++) {
341 			int	ioffset = i << mp->m_sb.sb_inodelog;
342 			uint	isize = xfs_dinode_size(version);
343 
344 			free = xfs_make_iptr(mp, fbuf, i);
345 			free->di_magic = cpu_to_be16(XFS_DINODE_MAGIC);
346 			free->di_version = version;
347 			free->di_gen = cpu_to_be32(gen);
348 			free->di_next_unlinked = cpu_to_be32(NULLAGINO);
349 
350 			if (version == 3) {
351 				free->di_ino = cpu_to_be64(ino);
352 				ino++;
353 				uuid_copy(&free->di_uuid,
354 					  &mp->m_sb.sb_meta_uuid);
355 				xfs_dinode_calc_crc(mp, free);
356 			} else if (tp) {
357 				/* just log the inode core */
358 				xfs_trans_log_buf(tp, fbuf, ioffset,
359 						  ioffset + isize - 1);
360 			}
361 		}
362 
363 		if (tp) {
364 			/*
365 			 * Mark the buffer as an inode allocation buffer so it
366 			 * sticks in AIL at the point of this allocation
367 			 * transaction. This ensures the they are on disk before
368 			 * the tail of the log can be moved past this
369 			 * transaction (i.e. by preventing relogging from moving
370 			 * it forward in the log).
371 			 */
372 			xfs_trans_inode_alloc_buf(tp, fbuf);
373 			if (version == 3) {
374 				/*
375 				 * Mark the buffer as ordered so that they are
376 				 * not physically logged in the transaction but
377 				 * still tracked in the AIL as part of the
378 				 * transaction and pin the log appropriately.
379 				 */
380 				xfs_trans_ordered_buf(tp, fbuf);
381 			}
382 		} else {
383 			fbuf->b_flags |= XBF_DONE;
384 			xfs_buf_delwri_queue(fbuf, buffer_list);
385 			xfs_buf_relse(fbuf);
386 		}
387 	}
388 	return 0;
389 }
390 
391 /*
392  * Align startino and allocmask for a recently allocated sparse chunk such that
393  * they are fit for insertion (or merge) into the on-disk inode btrees.
394  *
395  * Background:
396  *
397  * When enabled, sparse inode support increases the inode alignment from cluster
398  * size to inode chunk size. This means that the minimum range between two
399  * non-adjacent inode records in the inobt is large enough for a full inode
400  * record. This allows for cluster sized, cluster aligned block allocation
401  * without need to worry about whether the resulting inode record overlaps with
402  * another record in the tree. Without this basic rule, we would have to deal
403  * with the consequences of overlap by potentially undoing recent allocations in
404  * the inode allocation codepath.
405  *
406  * Because of this alignment rule (which is enforced on mount), there are two
407  * inobt possibilities for newly allocated sparse chunks. One is that the
408  * aligned inode record for the chunk covers a range of inodes not already
409  * covered in the inobt (i.e., it is safe to insert a new sparse record). The
410  * other is that a record already exists at the aligned startino that considers
411  * the newly allocated range as sparse. In the latter case, record content is
412  * merged in hope that sparse inode chunks fill to full chunks over time.
413  */
414 STATIC void
415 xfs_align_sparse_ino(
416 	struct xfs_mount		*mp,
417 	xfs_agino_t			*startino,
418 	uint16_t			*allocmask)
419 {
420 	xfs_agblock_t			agbno;
421 	xfs_agblock_t			mod;
422 	int				offset;
423 
424 	agbno = XFS_AGINO_TO_AGBNO(mp, *startino);
425 	mod = agbno % mp->m_sb.sb_inoalignmt;
426 	if (!mod)
427 		return;
428 
429 	/* calculate the inode offset and align startino */
430 	offset = XFS_AGB_TO_AGINO(mp, mod);
431 	*startino -= offset;
432 
433 	/*
434 	 * Since startino has been aligned down, left shift allocmask such that
435 	 * it continues to represent the same physical inodes relative to the
436 	 * new startino.
437 	 */
438 	*allocmask <<= offset / XFS_INODES_PER_HOLEMASK_BIT;
439 }
440 
441 /*
442  * Determine whether the source inode record can merge into the target. Both
443  * records must be sparse, the inode ranges must match and there must be no
444  * allocation overlap between the records.
445  */
446 STATIC bool
447 __xfs_inobt_can_merge(
448 	struct xfs_inobt_rec_incore	*trec,	/* tgt record */
449 	struct xfs_inobt_rec_incore	*srec)	/* src record */
450 {
451 	uint64_t			talloc;
452 	uint64_t			salloc;
453 
454 	/* records must cover the same inode range */
455 	if (trec->ir_startino != srec->ir_startino)
456 		return false;
457 
458 	/* both records must be sparse */
459 	if (!xfs_inobt_issparse(trec->ir_holemask) ||
460 	    !xfs_inobt_issparse(srec->ir_holemask))
461 		return false;
462 
463 	/* both records must track some inodes */
464 	if (!trec->ir_count || !srec->ir_count)
465 		return false;
466 
467 	/* can't exceed capacity of a full record */
468 	if (trec->ir_count + srec->ir_count > XFS_INODES_PER_CHUNK)
469 		return false;
470 
471 	/* verify there is no allocation overlap */
472 	talloc = xfs_inobt_irec_to_allocmask(trec);
473 	salloc = xfs_inobt_irec_to_allocmask(srec);
474 	if (talloc & salloc)
475 		return false;
476 
477 	return true;
478 }
479 
480 /*
481  * Merge the source inode record into the target. The caller must call
482  * __xfs_inobt_can_merge() to ensure the merge is valid.
483  */
484 STATIC void
485 __xfs_inobt_rec_merge(
486 	struct xfs_inobt_rec_incore	*trec,	/* target */
487 	struct xfs_inobt_rec_incore	*srec)	/* src */
488 {
489 	ASSERT(trec->ir_startino == srec->ir_startino);
490 
491 	/* combine the counts */
492 	trec->ir_count += srec->ir_count;
493 	trec->ir_freecount += srec->ir_freecount;
494 
495 	/*
496 	 * Merge the holemask and free mask. For both fields, 0 bits refer to
497 	 * allocated inodes. We combine the allocated ranges with bitwise AND.
498 	 */
499 	trec->ir_holemask &= srec->ir_holemask;
500 	trec->ir_free &= srec->ir_free;
501 }
502 
503 /*
504  * Insert a new sparse inode chunk into the associated inode btree. The inode
505  * record for the sparse chunk is pre-aligned to a startino that should match
506  * any pre-existing sparse inode record in the tree. This allows sparse chunks
507  * to fill over time.
508  *
509  * This function supports two modes of handling preexisting records depending on
510  * the merge flag. If merge is true, the provided record is merged with the
511  * existing record and updated in place. The merged record is returned in nrec.
512  * If merge is false, an existing record is replaced with the provided record.
513  * If no preexisting record exists, the provided record is always inserted.
514  *
515  * It is considered corruption if a merge is requested and not possible. Given
516  * the sparse inode alignment constraints, this should never happen.
517  */
518 STATIC int
519 xfs_inobt_insert_sprec(
520 	struct xfs_mount		*mp,
521 	struct xfs_trans		*tp,
522 	struct xfs_buf			*agbp,
523 	int				btnum,
524 	struct xfs_inobt_rec_incore	*nrec,	/* in/out: new/merged rec. */
525 	bool				merge)	/* merge or replace */
526 {
527 	struct xfs_btree_cur		*cur;
528 	struct xfs_agi			*agi = XFS_BUF_TO_AGI(agbp);
529 	xfs_agnumber_t			agno = be32_to_cpu(agi->agi_seqno);
530 	int				error;
531 	int				i;
532 	struct xfs_inobt_rec_incore	rec;
533 
534 	cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, btnum);
535 
536 	/* the new record is pre-aligned so we know where to look */
537 	error = xfs_inobt_lookup(cur, nrec->ir_startino, XFS_LOOKUP_EQ, &i);
538 	if (error)
539 		goto error;
540 	/* if nothing there, insert a new record and return */
541 	if (i == 0) {
542 		error = xfs_inobt_insert_rec(cur, nrec->ir_holemask,
543 					     nrec->ir_count, nrec->ir_freecount,
544 					     nrec->ir_free, &i);
545 		if (error)
546 			goto error;
547 		XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error);
548 
549 		goto out;
550 	}
551 
552 	/*
553 	 * A record exists at this startino. Merge or replace the record
554 	 * depending on what we've been asked to do.
555 	 */
556 	if (merge) {
557 		error = xfs_inobt_get_rec(cur, &rec, &i);
558 		if (error)
559 			goto error;
560 		XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error);
561 		XFS_WANT_CORRUPTED_GOTO(mp,
562 					rec.ir_startino == nrec->ir_startino,
563 					error);
564 
565 		/*
566 		 * This should never fail. If we have coexisting records that
567 		 * cannot merge, something is seriously wrong.
568 		 */
569 		XFS_WANT_CORRUPTED_GOTO(mp, __xfs_inobt_can_merge(nrec, &rec),
570 					error);
571 
572 		trace_xfs_irec_merge_pre(mp, agno, rec.ir_startino,
573 					 rec.ir_holemask, nrec->ir_startino,
574 					 nrec->ir_holemask);
575 
576 		/* merge to nrec to output the updated record */
577 		__xfs_inobt_rec_merge(nrec, &rec);
578 
579 		trace_xfs_irec_merge_post(mp, agno, nrec->ir_startino,
580 					  nrec->ir_holemask);
581 
582 		error = xfs_inobt_rec_check_count(mp, nrec);
583 		if (error)
584 			goto error;
585 	}
586 
587 	error = xfs_inobt_update(cur, nrec);
588 	if (error)
589 		goto error;
590 
591 out:
592 	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
593 	return 0;
594 error:
595 	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
596 	return error;
597 }
598 
599 /*
600  * Allocate new inodes in the allocation group specified by agbp.
601  * Return 0 for success, else error code.
602  */
603 STATIC int
604 xfs_ialloc_ag_alloc(
605 	struct xfs_trans	*tp,
606 	struct xfs_buf		*agbp,
607 	int			*alloc)
608 {
609 	struct xfs_agi		*agi;
610 	struct xfs_alloc_arg	args;
611 	xfs_agnumber_t		agno;
612 	int			error;
613 	xfs_agino_t		newino;		/* new first inode's number */
614 	xfs_agino_t		newlen;		/* new number of inodes */
615 	int			isaligned = 0;	/* inode allocation at stripe */
616 						/* unit boundary */
617 	/* init. to full chunk */
618 	uint16_t		allocmask = (uint16_t) -1;
619 	struct xfs_inobt_rec_incore rec;
620 	struct xfs_perag	*pag;
621 	struct xfs_ino_geometry	*igeo = M_IGEO(tp->t_mountp);
622 	int			do_sparse = 0;
623 
624 	memset(&args, 0, sizeof(args));
625 	args.tp = tp;
626 	args.mp = tp->t_mountp;
627 	args.fsbno = NULLFSBLOCK;
628 	args.oinfo = XFS_RMAP_OINFO_INODES;
629 
630 #ifdef DEBUG
631 	/* randomly do sparse inode allocations */
632 	if (xfs_sb_version_hassparseinodes(&tp->t_mountp->m_sb) &&
633 	    igeo->ialloc_min_blks < igeo->ialloc_blks)
634 		do_sparse = prandom_u32() & 1;
635 #endif
636 
637 	/*
638 	 * Locking will ensure that we don't have two callers in here
639 	 * at one time.
640 	 */
641 	newlen = igeo->ialloc_inos;
642 	if (igeo->maxicount &&
643 	    percpu_counter_read_positive(&args.mp->m_icount) + newlen >
644 							igeo->maxicount)
645 		return -ENOSPC;
646 	args.minlen = args.maxlen = igeo->ialloc_blks;
647 	/*
648 	 * First try to allocate inodes contiguous with the last-allocated
649 	 * chunk of inodes.  If the filesystem is striped, this will fill
650 	 * an entire stripe unit with inodes.
651 	 */
652 	agi = XFS_BUF_TO_AGI(agbp);
653 	newino = be32_to_cpu(agi->agi_newino);
654 	agno = be32_to_cpu(agi->agi_seqno);
655 	args.agbno = XFS_AGINO_TO_AGBNO(args.mp, newino) +
656 		     igeo->ialloc_blks;
657 	if (do_sparse)
658 		goto sparse_alloc;
659 	if (likely(newino != NULLAGINO &&
660 		  (args.agbno < be32_to_cpu(agi->agi_length)))) {
661 		args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
662 		args.type = XFS_ALLOCTYPE_THIS_BNO;
663 		args.prod = 1;
664 
665 		/*
666 		 * We need to take into account alignment here to ensure that
667 		 * we don't modify the free list if we fail to have an exact
668 		 * block. If we don't have an exact match, and every oher
669 		 * attempt allocation attempt fails, we'll end up cancelling
670 		 * a dirty transaction and shutting down.
671 		 *
672 		 * For an exact allocation, alignment must be 1,
673 		 * however we need to take cluster alignment into account when
674 		 * fixing up the freelist. Use the minalignslop field to
675 		 * indicate that extra blocks might be required for alignment,
676 		 * but not to use them in the actual exact allocation.
677 		 */
678 		args.alignment = 1;
679 		args.minalignslop = igeo->cluster_align - 1;
680 
681 		/* Allow space for the inode btree to split. */
682 		args.minleft = igeo->inobt_maxlevels - 1;
683 		if ((error = xfs_alloc_vextent(&args)))
684 			return error;
685 
686 		/*
687 		 * This request might have dirtied the transaction if the AG can
688 		 * satisfy the request, but the exact block was not available.
689 		 * If the allocation did fail, subsequent requests will relax
690 		 * the exact agbno requirement and increase the alignment
691 		 * instead. It is critical that the total size of the request
692 		 * (len + alignment + slop) does not increase from this point
693 		 * on, so reset minalignslop to ensure it is not included in
694 		 * subsequent requests.
695 		 */
696 		args.minalignslop = 0;
697 	}
698 
699 	if (unlikely(args.fsbno == NULLFSBLOCK)) {
700 		/*
701 		 * Set the alignment for the allocation.
702 		 * If stripe alignment is turned on then align at stripe unit
703 		 * boundary.
704 		 * If the cluster size is smaller than a filesystem block
705 		 * then we're doing I/O for inodes in filesystem block size
706 		 * pieces, so don't need alignment anyway.
707 		 */
708 		isaligned = 0;
709 		if (igeo->ialloc_align) {
710 			ASSERT(!(args.mp->m_flags & XFS_MOUNT_NOALIGN));
711 			args.alignment = args.mp->m_dalign;
712 			isaligned = 1;
713 		} else
714 			args.alignment = igeo->cluster_align;
715 		/*
716 		 * Need to figure out where to allocate the inode blocks.
717 		 * Ideally they should be spaced out through the a.g.
718 		 * For now, just allocate blocks up front.
719 		 */
720 		args.agbno = be32_to_cpu(agi->agi_root);
721 		args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
722 		/*
723 		 * Allocate a fixed-size extent of inodes.
724 		 */
725 		args.type = XFS_ALLOCTYPE_NEAR_BNO;
726 		args.prod = 1;
727 		/*
728 		 * Allow space for the inode btree to split.
729 		 */
730 		args.minleft = igeo->inobt_maxlevels - 1;
731 		if ((error = xfs_alloc_vextent(&args)))
732 			return error;
733 	}
734 
735 	/*
736 	 * If stripe alignment is turned on, then try again with cluster
737 	 * alignment.
738 	 */
739 	if (isaligned && args.fsbno == NULLFSBLOCK) {
740 		args.type = XFS_ALLOCTYPE_NEAR_BNO;
741 		args.agbno = be32_to_cpu(agi->agi_root);
742 		args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
743 		args.alignment = igeo->cluster_align;
744 		if ((error = xfs_alloc_vextent(&args)))
745 			return error;
746 	}
747 
748 	/*
749 	 * Finally, try a sparse allocation if the filesystem supports it and
750 	 * the sparse allocation length is smaller than a full chunk.
751 	 */
752 	if (xfs_sb_version_hassparseinodes(&args.mp->m_sb) &&
753 	    igeo->ialloc_min_blks < igeo->ialloc_blks &&
754 	    args.fsbno == NULLFSBLOCK) {
755 sparse_alloc:
756 		args.type = XFS_ALLOCTYPE_NEAR_BNO;
757 		args.agbno = be32_to_cpu(agi->agi_root);
758 		args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
759 		args.alignment = args.mp->m_sb.sb_spino_align;
760 		args.prod = 1;
761 
762 		args.minlen = igeo->ialloc_min_blks;
763 		args.maxlen = args.minlen;
764 
765 		/*
766 		 * The inode record will be aligned to full chunk size. We must
767 		 * prevent sparse allocation from AG boundaries that result in
768 		 * invalid inode records, such as records that start at agbno 0
769 		 * or extend beyond the AG.
770 		 *
771 		 * Set min agbno to the first aligned, non-zero agbno and max to
772 		 * the last aligned agbno that is at least one full chunk from
773 		 * the end of the AG.
774 		 */
775 		args.min_agbno = args.mp->m_sb.sb_inoalignmt;
776 		args.max_agbno = round_down(args.mp->m_sb.sb_agblocks,
777 					    args.mp->m_sb.sb_inoalignmt) -
778 				 igeo->ialloc_blks;
779 
780 		error = xfs_alloc_vextent(&args);
781 		if (error)
782 			return error;
783 
784 		newlen = XFS_AGB_TO_AGINO(args.mp, args.len);
785 		ASSERT(newlen <= XFS_INODES_PER_CHUNK);
786 		allocmask = (1 << (newlen / XFS_INODES_PER_HOLEMASK_BIT)) - 1;
787 	}
788 
789 	if (args.fsbno == NULLFSBLOCK) {
790 		*alloc = 0;
791 		return 0;
792 	}
793 	ASSERT(args.len == args.minlen);
794 
795 	/*
796 	 * Stamp and write the inode buffers.
797 	 *
798 	 * Seed the new inode cluster with a random generation number. This
799 	 * prevents short-term reuse of generation numbers if a chunk is
800 	 * freed and then immediately reallocated. We use random numbers
801 	 * rather than a linear progression to prevent the next generation
802 	 * number from being easily guessable.
803 	 */
804 	error = xfs_ialloc_inode_init(args.mp, tp, NULL, newlen, agno,
805 			args.agbno, args.len, prandom_u32());
806 
807 	if (error)
808 		return error;
809 	/*
810 	 * Convert the results.
811 	 */
812 	newino = XFS_AGB_TO_AGINO(args.mp, args.agbno);
813 
814 	if (xfs_inobt_issparse(~allocmask)) {
815 		/*
816 		 * We've allocated a sparse chunk. Align the startino and mask.
817 		 */
818 		xfs_align_sparse_ino(args.mp, &newino, &allocmask);
819 
820 		rec.ir_startino = newino;
821 		rec.ir_holemask = ~allocmask;
822 		rec.ir_count = newlen;
823 		rec.ir_freecount = newlen;
824 		rec.ir_free = XFS_INOBT_ALL_FREE;
825 
826 		/*
827 		 * Insert the sparse record into the inobt and allow for a merge
828 		 * if necessary. If a merge does occur, rec is updated to the
829 		 * merged record.
830 		 */
831 		error = xfs_inobt_insert_sprec(args.mp, tp, agbp, XFS_BTNUM_INO,
832 					       &rec, true);
833 		if (error == -EFSCORRUPTED) {
834 			xfs_alert(args.mp,
835 	"invalid sparse inode record: ino 0x%llx holemask 0x%x count %u",
836 				  XFS_AGINO_TO_INO(args.mp, agno,
837 						   rec.ir_startino),
838 				  rec.ir_holemask, rec.ir_count);
839 			xfs_force_shutdown(args.mp, SHUTDOWN_CORRUPT_INCORE);
840 		}
841 		if (error)
842 			return error;
843 
844 		/*
845 		 * We can't merge the part we've just allocated as for the inobt
846 		 * due to finobt semantics. The original record may or may not
847 		 * exist independent of whether physical inodes exist in this
848 		 * sparse chunk.
849 		 *
850 		 * We must update the finobt record based on the inobt record.
851 		 * rec contains the fully merged and up to date inobt record
852 		 * from the previous call. Set merge false to replace any
853 		 * existing record with this one.
854 		 */
855 		if (xfs_sb_version_hasfinobt(&args.mp->m_sb)) {
856 			error = xfs_inobt_insert_sprec(args.mp, tp, agbp,
857 						       XFS_BTNUM_FINO, &rec,
858 						       false);
859 			if (error)
860 				return error;
861 		}
862 	} else {
863 		/* full chunk - insert new records to both btrees */
864 		error = xfs_inobt_insert(args.mp, tp, agbp, newino, newlen,
865 					 XFS_BTNUM_INO);
866 		if (error)
867 			return error;
868 
869 		if (xfs_sb_version_hasfinobt(&args.mp->m_sb)) {
870 			error = xfs_inobt_insert(args.mp, tp, agbp, newino,
871 						 newlen, XFS_BTNUM_FINO);
872 			if (error)
873 				return error;
874 		}
875 	}
876 
877 	/*
878 	 * Update AGI counts and newino.
879 	 */
880 	be32_add_cpu(&agi->agi_count, newlen);
881 	be32_add_cpu(&agi->agi_freecount, newlen);
882 	pag = xfs_perag_get(args.mp, agno);
883 	pag->pagi_freecount += newlen;
884 	pag->pagi_count += newlen;
885 	xfs_perag_put(pag);
886 	agi->agi_newino = cpu_to_be32(newino);
887 
888 	/*
889 	 * Log allocation group header fields
890 	 */
891 	xfs_ialloc_log_agi(tp, agbp,
892 		XFS_AGI_COUNT | XFS_AGI_FREECOUNT | XFS_AGI_NEWINO);
893 	/*
894 	 * Modify/log superblock values for inode count and inode free count.
895 	 */
896 	xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, (long)newlen);
897 	xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, (long)newlen);
898 	*alloc = 1;
899 	return 0;
900 }
901 
902 STATIC xfs_agnumber_t
903 xfs_ialloc_next_ag(
904 	xfs_mount_t	*mp)
905 {
906 	xfs_agnumber_t	agno;
907 
908 	spin_lock(&mp->m_agirotor_lock);
909 	agno = mp->m_agirotor;
910 	if (++mp->m_agirotor >= mp->m_maxagi)
911 		mp->m_agirotor = 0;
912 	spin_unlock(&mp->m_agirotor_lock);
913 
914 	return agno;
915 }
916 
917 /*
918  * Select an allocation group to look for a free inode in, based on the parent
919  * inode and the mode.  Return the allocation group buffer.
920  */
921 STATIC xfs_agnumber_t
922 xfs_ialloc_ag_select(
923 	xfs_trans_t	*tp,		/* transaction pointer */
924 	xfs_ino_t	parent,		/* parent directory inode number */
925 	umode_t		mode)		/* bits set to indicate file type */
926 {
927 	xfs_agnumber_t	agcount;	/* number of ag's in the filesystem */
928 	xfs_agnumber_t	agno;		/* current ag number */
929 	int		flags;		/* alloc buffer locking flags */
930 	xfs_extlen_t	ineed;		/* blocks needed for inode allocation */
931 	xfs_extlen_t	longest = 0;	/* longest extent available */
932 	xfs_mount_t	*mp;		/* mount point structure */
933 	int		needspace;	/* file mode implies space allocated */
934 	xfs_perag_t	*pag;		/* per allocation group data */
935 	xfs_agnumber_t	pagno;		/* parent (starting) ag number */
936 	int		error;
937 
938 	/*
939 	 * Files of these types need at least one block if length > 0
940 	 * (and they won't fit in the inode, but that's hard to figure out).
941 	 */
942 	needspace = S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode);
943 	mp = tp->t_mountp;
944 	agcount = mp->m_maxagi;
945 	if (S_ISDIR(mode))
946 		pagno = xfs_ialloc_next_ag(mp);
947 	else {
948 		pagno = XFS_INO_TO_AGNO(mp, parent);
949 		if (pagno >= agcount)
950 			pagno = 0;
951 	}
952 
953 	ASSERT(pagno < agcount);
954 
955 	/*
956 	 * Loop through allocation groups, looking for one with a little
957 	 * free space in it.  Note we don't look for free inodes, exactly.
958 	 * Instead, we include whether there is a need to allocate inodes
959 	 * to mean that blocks must be allocated for them,
960 	 * if none are currently free.
961 	 */
962 	agno = pagno;
963 	flags = XFS_ALLOC_FLAG_TRYLOCK;
964 	for (;;) {
965 		pag = xfs_perag_get(mp, agno);
966 		if (!pag->pagi_inodeok) {
967 			xfs_ialloc_next_ag(mp);
968 			goto nextag;
969 		}
970 
971 		if (!pag->pagi_init) {
972 			error = xfs_ialloc_pagi_init(mp, tp, agno);
973 			if (error)
974 				goto nextag;
975 		}
976 
977 		if (pag->pagi_freecount) {
978 			xfs_perag_put(pag);
979 			return agno;
980 		}
981 
982 		if (!pag->pagf_init) {
983 			error = xfs_alloc_pagf_init(mp, tp, agno, flags);
984 			if (error)
985 				goto nextag;
986 		}
987 
988 		/*
989 		 * Check that there is enough free space for the file plus a
990 		 * chunk of inodes if we need to allocate some. If this is the
991 		 * first pass across the AGs, take into account the potential
992 		 * space needed for alignment of inode chunks when checking the
993 		 * longest contiguous free space in the AG - this prevents us
994 		 * from getting ENOSPC because we have free space larger than
995 		 * ialloc_blks but alignment constraints prevent us from using
996 		 * it.
997 		 *
998 		 * If we can't find an AG with space for full alignment slack to
999 		 * be taken into account, we must be near ENOSPC in all AGs.
1000 		 * Hence we don't include alignment for the second pass and so
1001 		 * if we fail allocation due to alignment issues then it is most
1002 		 * likely a real ENOSPC condition.
1003 		 */
1004 		ineed = M_IGEO(mp)->ialloc_min_blks;
1005 		if (flags && ineed > 1)
1006 			ineed += M_IGEO(mp)->cluster_align;
1007 		longest = pag->pagf_longest;
1008 		if (!longest)
1009 			longest = pag->pagf_flcount > 0;
1010 
1011 		if (pag->pagf_freeblks >= needspace + ineed &&
1012 		    longest >= ineed) {
1013 			xfs_perag_put(pag);
1014 			return agno;
1015 		}
1016 nextag:
1017 		xfs_perag_put(pag);
1018 		/*
1019 		 * No point in iterating over the rest, if we're shutting
1020 		 * down.
1021 		 */
1022 		if (XFS_FORCED_SHUTDOWN(mp))
1023 			return NULLAGNUMBER;
1024 		agno++;
1025 		if (agno >= agcount)
1026 			agno = 0;
1027 		if (agno == pagno) {
1028 			if (flags == 0)
1029 				return NULLAGNUMBER;
1030 			flags = 0;
1031 		}
1032 	}
1033 }
1034 
1035 /*
1036  * Try to retrieve the next record to the left/right from the current one.
1037  */
1038 STATIC int
1039 xfs_ialloc_next_rec(
1040 	struct xfs_btree_cur	*cur,
1041 	xfs_inobt_rec_incore_t	*rec,
1042 	int			*done,
1043 	int			left)
1044 {
1045 	int                     error;
1046 	int			i;
1047 
1048 	if (left)
1049 		error = xfs_btree_decrement(cur, 0, &i);
1050 	else
1051 		error = xfs_btree_increment(cur, 0, &i);
1052 
1053 	if (error)
1054 		return error;
1055 	*done = !i;
1056 	if (i) {
1057 		error = xfs_inobt_get_rec(cur, rec, &i);
1058 		if (error)
1059 			return error;
1060 		XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, i == 1);
1061 	}
1062 
1063 	return 0;
1064 }
1065 
1066 STATIC int
1067 xfs_ialloc_get_rec(
1068 	struct xfs_btree_cur	*cur,
1069 	xfs_agino_t		agino,
1070 	xfs_inobt_rec_incore_t	*rec,
1071 	int			*done)
1072 {
1073 	int                     error;
1074 	int			i;
1075 
1076 	error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_EQ, &i);
1077 	if (error)
1078 		return error;
1079 	*done = !i;
1080 	if (i) {
1081 		error = xfs_inobt_get_rec(cur, rec, &i);
1082 		if (error)
1083 			return error;
1084 		XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, i == 1);
1085 	}
1086 
1087 	return 0;
1088 }
1089 
1090 /*
1091  * Return the offset of the first free inode in the record. If the inode chunk
1092  * is sparsely allocated, we convert the record holemask to inode granularity
1093  * and mask off the unallocated regions from the inode free mask.
1094  */
1095 STATIC int
1096 xfs_inobt_first_free_inode(
1097 	struct xfs_inobt_rec_incore	*rec)
1098 {
1099 	xfs_inofree_t			realfree;
1100 
1101 	/* if there are no holes, return the first available offset */
1102 	if (!xfs_inobt_issparse(rec->ir_holemask))
1103 		return xfs_lowbit64(rec->ir_free);
1104 
1105 	realfree = xfs_inobt_irec_to_allocmask(rec);
1106 	realfree &= rec->ir_free;
1107 
1108 	return xfs_lowbit64(realfree);
1109 }
1110 
1111 /*
1112  * Allocate an inode using the inobt-only algorithm.
1113  */
1114 STATIC int
1115 xfs_dialloc_ag_inobt(
1116 	struct xfs_trans	*tp,
1117 	struct xfs_buf		*agbp,
1118 	xfs_ino_t		parent,
1119 	xfs_ino_t		*inop)
1120 {
1121 	struct xfs_mount	*mp = tp->t_mountp;
1122 	struct xfs_agi		*agi = XFS_BUF_TO_AGI(agbp);
1123 	xfs_agnumber_t		agno = be32_to_cpu(agi->agi_seqno);
1124 	xfs_agnumber_t		pagno = XFS_INO_TO_AGNO(mp, parent);
1125 	xfs_agino_t		pagino = XFS_INO_TO_AGINO(mp, parent);
1126 	struct xfs_perag	*pag;
1127 	struct xfs_btree_cur	*cur, *tcur;
1128 	struct xfs_inobt_rec_incore rec, trec;
1129 	xfs_ino_t		ino;
1130 	int			error;
1131 	int			offset;
1132 	int			i, j;
1133 	int			searchdistance = 10;
1134 
1135 	pag = xfs_perag_get(mp, agno);
1136 
1137 	ASSERT(pag->pagi_init);
1138 	ASSERT(pag->pagi_inodeok);
1139 	ASSERT(pag->pagi_freecount > 0);
1140 
1141  restart_pagno:
1142 	cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
1143 	/*
1144 	 * If pagino is 0 (this is the root inode allocation) use newino.
1145 	 * This must work because we've just allocated some.
1146 	 */
1147 	if (!pagino)
1148 		pagino = be32_to_cpu(agi->agi_newino);
1149 
1150 	error = xfs_check_agi_freecount(cur, agi);
1151 	if (error)
1152 		goto error0;
1153 
1154 	/*
1155 	 * If in the same AG as the parent, try to get near the parent.
1156 	 */
1157 	if (pagno == agno) {
1158 		int		doneleft;	/* done, to the left */
1159 		int		doneright;	/* done, to the right */
1160 
1161 		error = xfs_inobt_lookup(cur, pagino, XFS_LOOKUP_LE, &i);
1162 		if (error)
1163 			goto error0;
1164 		XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error0);
1165 
1166 		error = xfs_inobt_get_rec(cur, &rec, &j);
1167 		if (error)
1168 			goto error0;
1169 		XFS_WANT_CORRUPTED_GOTO(mp, j == 1, error0);
1170 
1171 		if (rec.ir_freecount > 0) {
1172 			/*
1173 			 * Found a free inode in the same chunk
1174 			 * as the parent, done.
1175 			 */
1176 			goto alloc_inode;
1177 		}
1178 
1179 
1180 		/*
1181 		 * In the same AG as parent, but parent's chunk is full.
1182 		 */
1183 
1184 		/* duplicate the cursor, search left & right simultaneously */
1185 		error = xfs_btree_dup_cursor(cur, &tcur);
1186 		if (error)
1187 			goto error0;
1188 
1189 		/*
1190 		 * Skip to last blocks looked up if same parent inode.
1191 		 */
1192 		if (pagino != NULLAGINO &&
1193 		    pag->pagl_pagino == pagino &&
1194 		    pag->pagl_leftrec != NULLAGINO &&
1195 		    pag->pagl_rightrec != NULLAGINO) {
1196 			error = xfs_ialloc_get_rec(tcur, pag->pagl_leftrec,
1197 						   &trec, &doneleft);
1198 			if (error)
1199 				goto error1;
1200 
1201 			error = xfs_ialloc_get_rec(cur, pag->pagl_rightrec,
1202 						   &rec, &doneright);
1203 			if (error)
1204 				goto error1;
1205 		} else {
1206 			/* search left with tcur, back up 1 record */
1207 			error = xfs_ialloc_next_rec(tcur, &trec, &doneleft, 1);
1208 			if (error)
1209 				goto error1;
1210 
1211 			/* search right with cur, go forward 1 record. */
1212 			error = xfs_ialloc_next_rec(cur, &rec, &doneright, 0);
1213 			if (error)
1214 				goto error1;
1215 		}
1216 
1217 		/*
1218 		 * Loop until we find an inode chunk with a free inode.
1219 		 */
1220 		while (--searchdistance > 0 && (!doneleft || !doneright)) {
1221 			int	useleft;  /* using left inode chunk this time */
1222 
1223 			/* figure out the closer block if both are valid. */
1224 			if (!doneleft && !doneright) {
1225 				useleft = pagino -
1226 				 (trec.ir_startino + XFS_INODES_PER_CHUNK - 1) <
1227 				  rec.ir_startino - pagino;
1228 			} else {
1229 				useleft = !doneleft;
1230 			}
1231 
1232 			/* free inodes to the left? */
1233 			if (useleft && trec.ir_freecount) {
1234 				xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1235 				cur = tcur;
1236 
1237 				pag->pagl_leftrec = trec.ir_startino;
1238 				pag->pagl_rightrec = rec.ir_startino;
1239 				pag->pagl_pagino = pagino;
1240 				rec = trec;
1241 				goto alloc_inode;
1242 			}
1243 
1244 			/* free inodes to the right? */
1245 			if (!useleft && rec.ir_freecount) {
1246 				xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1247 
1248 				pag->pagl_leftrec = trec.ir_startino;
1249 				pag->pagl_rightrec = rec.ir_startino;
1250 				pag->pagl_pagino = pagino;
1251 				goto alloc_inode;
1252 			}
1253 
1254 			/* get next record to check */
1255 			if (useleft) {
1256 				error = xfs_ialloc_next_rec(tcur, &trec,
1257 								 &doneleft, 1);
1258 			} else {
1259 				error = xfs_ialloc_next_rec(cur, &rec,
1260 								 &doneright, 0);
1261 			}
1262 			if (error)
1263 				goto error1;
1264 		}
1265 
1266 		if (searchdistance <= 0) {
1267 			/*
1268 			 * Not in range - save last search
1269 			 * location and allocate a new inode
1270 			 */
1271 			xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1272 			pag->pagl_leftrec = trec.ir_startino;
1273 			pag->pagl_rightrec = rec.ir_startino;
1274 			pag->pagl_pagino = pagino;
1275 
1276 		} else {
1277 			/*
1278 			 * We've reached the end of the btree. because
1279 			 * we are only searching a small chunk of the
1280 			 * btree each search, there is obviously free
1281 			 * inodes closer to the parent inode than we
1282 			 * are now. restart the search again.
1283 			 */
1284 			pag->pagl_pagino = NULLAGINO;
1285 			pag->pagl_leftrec = NULLAGINO;
1286 			pag->pagl_rightrec = NULLAGINO;
1287 			xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1288 			xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1289 			goto restart_pagno;
1290 		}
1291 	}
1292 
1293 	/*
1294 	 * In a different AG from the parent.
1295 	 * See if the most recently allocated block has any free.
1296 	 */
1297 	if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1298 		error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1299 					 XFS_LOOKUP_EQ, &i);
1300 		if (error)
1301 			goto error0;
1302 
1303 		if (i == 1) {
1304 			error = xfs_inobt_get_rec(cur, &rec, &j);
1305 			if (error)
1306 				goto error0;
1307 
1308 			if (j == 1 && rec.ir_freecount > 0) {
1309 				/*
1310 				 * The last chunk allocated in the group
1311 				 * still has a free inode.
1312 				 */
1313 				goto alloc_inode;
1314 			}
1315 		}
1316 	}
1317 
1318 	/*
1319 	 * None left in the last group, search the whole AG
1320 	 */
1321 	error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1322 	if (error)
1323 		goto error0;
1324 	XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error0);
1325 
1326 	for (;;) {
1327 		error = xfs_inobt_get_rec(cur, &rec, &i);
1328 		if (error)
1329 			goto error0;
1330 		XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error0);
1331 		if (rec.ir_freecount > 0)
1332 			break;
1333 		error = xfs_btree_increment(cur, 0, &i);
1334 		if (error)
1335 			goto error0;
1336 		XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error0);
1337 	}
1338 
1339 alloc_inode:
1340 	offset = xfs_inobt_first_free_inode(&rec);
1341 	ASSERT(offset >= 0);
1342 	ASSERT(offset < XFS_INODES_PER_CHUNK);
1343 	ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1344 				   XFS_INODES_PER_CHUNK) == 0);
1345 	ino = XFS_AGINO_TO_INO(mp, agno, rec.ir_startino + offset);
1346 	rec.ir_free &= ~XFS_INOBT_MASK(offset);
1347 	rec.ir_freecount--;
1348 	error = xfs_inobt_update(cur, &rec);
1349 	if (error)
1350 		goto error0;
1351 	be32_add_cpu(&agi->agi_freecount, -1);
1352 	xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1353 	pag->pagi_freecount--;
1354 
1355 	error = xfs_check_agi_freecount(cur, agi);
1356 	if (error)
1357 		goto error0;
1358 
1359 	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1360 	xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1361 	xfs_perag_put(pag);
1362 	*inop = ino;
1363 	return 0;
1364 error1:
1365 	xfs_btree_del_cursor(tcur, XFS_BTREE_ERROR);
1366 error0:
1367 	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1368 	xfs_perag_put(pag);
1369 	return error;
1370 }
1371 
1372 /*
1373  * Use the free inode btree to allocate an inode based on distance from the
1374  * parent. Note that the provided cursor may be deleted and replaced.
1375  */
1376 STATIC int
1377 xfs_dialloc_ag_finobt_near(
1378 	xfs_agino_t			pagino,
1379 	struct xfs_btree_cur		**ocur,
1380 	struct xfs_inobt_rec_incore	*rec)
1381 {
1382 	struct xfs_btree_cur		*lcur = *ocur;	/* left search cursor */
1383 	struct xfs_btree_cur		*rcur;	/* right search cursor */
1384 	struct xfs_inobt_rec_incore	rrec;
1385 	int				error;
1386 	int				i, j;
1387 
1388 	error = xfs_inobt_lookup(lcur, pagino, XFS_LOOKUP_LE, &i);
1389 	if (error)
1390 		return error;
1391 
1392 	if (i == 1) {
1393 		error = xfs_inobt_get_rec(lcur, rec, &i);
1394 		if (error)
1395 			return error;
1396 		XFS_WANT_CORRUPTED_RETURN(lcur->bc_mp, i == 1);
1397 
1398 		/*
1399 		 * See if we've landed in the parent inode record. The finobt
1400 		 * only tracks chunks with at least one free inode, so record
1401 		 * existence is enough.
1402 		 */
1403 		if (pagino >= rec->ir_startino &&
1404 		    pagino < (rec->ir_startino + XFS_INODES_PER_CHUNK))
1405 			return 0;
1406 	}
1407 
1408 	error = xfs_btree_dup_cursor(lcur, &rcur);
1409 	if (error)
1410 		return error;
1411 
1412 	error = xfs_inobt_lookup(rcur, pagino, XFS_LOOKUP_GE, &j);
1413 	if (error)
1414 		goto error_rcur;
1415 	if (j == 1) {
1416 		error = xfs_inobt_get_rec(rcur, &rrec, &j);
1417 		if (error)
1418 			goto error_rcur;
1419 		XFS_WANT_CORRUPTED_GOTO(lcur->bc_mp, j == 1, error_rcur);
1420 	}
1421 
1422 	XFS_WANT_CORRUPTED_GOTO(lcur->bc_mp, i == 1 || j == 1, error_rcur);
1423 	if (i == 1 && j == 1) {
1424 		/*
1425 		 * Both the left and right records are valid. Choose the closer
1426 		 * inode chunk to the target.
1427 		 */
1428 		if ((pagino - rec->ir_startino + XFS_INODES_PER_CHUNK - 1) >
1429 		    (rrec.ir_startino - pagino)) {
1430 			*rec = rrec;
1431 			xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1432 			*ocur = rcur;
1433 		} else {
1434 			xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1435 		}
1436 	} else if (j == 1) {
1437 		/* only the right record is valid */
1438 		*rec = rrec;
1439 		xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1440 		*ocur = rcur;
1441 	} else if (i == 1) {
1442 		/* only the left record is valid */
1443 		xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1444 	}
1445 
1446 	return 0;
1447 
1448 error_rcur:
1449 	xfs_btree_del_cursor(rcur, XFS_BTREE_ERROR);
1450 	return error;
1451 }
1452 
1453 /*
1454  * Use the free inode btree to find a free inode based on a newino hint. If
1455  * the hint is NULL, find the first free inode in the AG.
1456  */
1457 STATIC int
1458 xfs_dialloc_ag_finobt_newino(
1459 	struct xfs_agi			*agi,
1460 	struct xfs_btree_cur		*cur,
1461 	struct xfs_inobt_rec_incore	*rec)
1462 {
1463 	int error;
1464 	int i;
1465 
1466 	if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1467 		error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1468 					 XFS_LOOKUP_EQ, &i);
1469 		if (error)
1470 			return error;
1471 		if (i == 1) {
1472 			error = xfs_inobt_get_rec(cur, rec, &i);
1473 			if (error)
1474 				return error;
1475 			XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, i == 1);
1476 			return 0;
1477 		}
1478 	}
1479 
1480 	/*
1481 	 * Find the first inode available in the AG.
1482 	 */
1483 	error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1484 	if (error)
1485 		return error;
1486 	XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, i == 1);
1487 
1488 	error = xfs_inobt_get_rec(cur, rec, &i);
1489 	if (error)
1490 		return error;
1491 	XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, i == 1);
1492 
1493 	return 0;
1494 }
1495 
1496 /*
1497  * Update the inobt based on a modification made to the finobt. Also ensure that
1498  * the records from both trees are equivalent post-modification.
1499  */
1500 STATIC int
1501 xfs_dialloc_ag_update_inobt(
1502 	struct xfs_btree_cur		*cur,	/* inobt cursor */
1503 	struct xfs_inobt_rec_incore	*frec,	/* finobt record */
1504 	int				offset) /* inode offset */
1505 {
1506 	struct xfs_inobt_rec_incore	rec;
1507 	int				error;
1508 	int				i;
1509 
1510 	error = xfs_inobt_lookup(cur, frec->ir_startino, XFS_LOOKUP_EQ, &i);
1511 	if (error)
1512 		return error;
1513 	XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, i == 1);
1514 
1515 	error = xfs_inobt_get_rec(cur, &rec, &i);
1516 	if (error)
1517 		return error;
1518 	XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, i == 1);
1519 	ASSERT((XFS_AGINO_TO_OFFSET(cur->bc_mp, rec.ir_startino) %
1520 				   XFS_INODES_PER_CHUNK) == 0);
1521 
1522 	rec.ir_free &= ~XFS_INOBT_MASK(offset);
1523 	rec.ir_freecount--;
1524 
1525 	XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, (rec.ir_free == frec->ir_free) &&
1526 				  (rec.ir_freecount == frec->ir_freecount));
1527 
1528 	return xfs_inobt_update(cur, &rec);
1529 }
1530 
1531 /*
1532  * Allocate an inode using the free inode btree, if available. Otherwise, fall
1533  * back to the inobt search algorithm.
1534  *
1535  * The caller selected an AG for us, and made sure that free inodes are
1536  * available.
1537  */
1538 STATIC int
1539 xfs_dialloc_ag(
1540 	struct xfs_trans	*tp,
1541 	struct xfs_buf		*agbp,
1542 	xfs_ino_t		parent,
1543 	xfs_ino_t		*inop)
1544 {
1545 	struct xfs_mount		*mp = tp->t_mountp;
1546 	struct xfs_agi			*agi = XFS_BUF_TO_AGI(agbp);
1547 	xfs_agnumber_t			agno = be32_to_cpu(agi->agi_seqno);
1548 	xfs_agnumber_t			pagno = XFS_INO_TO_AGNO(mp, parent);
1549 	xfs_agino_t			pagino = XFS_INO_TO_AGINO(mp, parent);
1550 	struct xfs_perag		*pag;
1551 	struct xfs_btree_cur		*cur;	/* finobt cursor */
1552 	struct xfs_btree_cur		*icur;	/* inobt cursor */
1553 	struct xfs_inobt_rec_incore	rec;
1554 	xfs_ino_t			ino;
1555 	int				error;
1556 	int				offset;
1557 	int				i;
1558 
1559 	if (!xfs_sb_version_hasfinobt(&mp->m_sb))
1560 		return xfs_dialloc_ag_inobt(tp, agbp, parent, inop);
1561 
1562 	pag = xfs_perag_get(mp, agno);
1563 
1564 	/*
1565 	 * If pagino is 0 (this is the root inode allocation) use newino.
1566 	 * This must work because we've just allocated some.
1567 	 */
1568 	if (!pagino)
1569 		pagino = be32_to_cpu(agi->agi_newino);
1570 
1571 	cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_FINO);
1572 
1573 	error = xfs_check_agi_freecount(cur, agi);
1574 	if (error)
1575 		goto error_cur;
1576 
1577 	/*
1578 	 * The search algorithm depends on whether we're in the same AG as the
1579 	 * parent. If so, find the closest available inode to the parent. If
1580 	 * not, consider the agi hint or find the first free inode in the AG.
1581 	 */
1582 	if (agno == pagno)
1583 		error = xfs_dialloc_ag_finobt_near(pagino, &cur, &rec);
1584 	else
1585 		error = xfs_dialloc_ag_finobt_newino(agi, cur, &rec);
1586 	if (error)
1587 		goto error_cur;
1588 
1589 	offset = xfs_inobt_first_free_inode(&rec);
1590 	ASSERT(offset >= 0);
1591 	ASSERT(offset < XFS_INODES_PER_CHUNK);
1592 	ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1593 				   XFS_INODES_PER_CHUNK) == 0);
1594 	ino = XFS_AGINO_TO_INO(mp, agno, rec.ir_startino + offset);
1595 
1596 	/*
1597 	 * Modify or remove the finobt record.
1598 	 */
1599 	rec.ir_free &= ~XFS_INOBT_MASK(offset);
1600 	rec.ir_freecount--;
1601 	if (rec.ir_freecount)
1602 		error = xfs_inobt_update(cur, &rec);
1603 	else
1604 		error = xfs_btree_delete(cur, &i);
1605 	if (error)
1606 		goto error_cur;
1607 
1608 	/*
1609 	 * The finobt has now been updated appropriately. We haven't updated the
1610 	 * agi and superblock yet, so we can create an inobt cursor and validate
1611 	 * the original freecount. If all is well, make the equivalent update to
1612 	 * the inobt using the finobt record and offset information.
1613 	 */
1614 	icur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
1615 
1616 	error = xfs_check_agi_freecount(icur, agi);
1617 	if (error)
1618 		goto error_icur;
1619 
1620 	error = xfs_dialloc_ag_update_inobt(icur, &rec, offset);
1621 	if (error)
1622 		goto error_icur;
1623 
1624 	/*
1625 	 * Both trees have now been updated. We must update the perag and
1626 	 * superblock before we can check the freecount for each btree.
1627 	 */
1628 	be32_add_cpu(&agi->agi_freecount, -1);
1629 	xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1630 	pag->pagi_freecount--;
1631 
1632 	xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1633 
1634 	error = xfs_check_agi_freecount(icur, agi);
1635 	if (error)
1636 		goto error_icur;
1637 	error = xfs_check_agi_freecount(cur, agi);
1638 	if (error)
1639 		goto error_icur;
1640 
1641 	xfs_btree_del_cursor(icur, XFS_BTREE_NOERROR);
1642 	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1643 	xfs_perag_put(pag);
1644 	*inop = ino;
1645 	return 0;
1646 
1647 error_icur:
1648 	xfs_btree_del_cursor(icur, XFS_BTREE_ERROR);
1649 error_cur:
1650 	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1651 	xfs_perag_put(pag);
1652 	return error;
1653 }
1654 
1655 /*
1656  * Allocate an inode on disk.
1657  *
1658  * Mode is used to tell whether the new inode will need space, and whether it
1659  * is a directory.
1660  *
1661  * This function is designed to be called twice if it has to do an allocation
1662  * to make more free inodes.  On the first call, *IO_agbp should be set to NULL.
1663  * If an inode is available without having to performn an allocation, an inode
1664  * number is returned.  In this case, *IO_agbp is set to NULL.  If an allocation
1665  * needs to be done, xfs_dialloc returns the current AGI buffer in *IO_agbp.
1666  * The caller should then commit the current transaction, allocate a
1667  * new transaction, and call xfs_dialloc() again, passing in the previous value
1668  * of *IO_agbp.  IO_agbp should be held across the transactions. Since the AGI
1669  * buffer is locked across the two calls, the second call is guaranteed to have
1670  * a free inode available.
1671  *
1672  * Once we successfully pick an inode its number is returned and the on-disk
1673  * data structures are updated.  The inode itself is not read in, since doing so
1674  * would break ordering constraints with xfs_reclaim.
1675  */
1676 int
1677 xfs_dialloc(
1678 	struct xfs_trans	*tp,
1679 	xfs_ino_t		parent,
1680 	umode_t			mode,
1681 	struct xfs_buf		**IO_agbp,
1682 	xfs_ino_t		*inop)
1683 {
1684 	struct xfs_mount	*mp = tp->t_mountp;
1685 	struct xfs_buf		*agbp;
1686 	xfs_agnumber_t		agno;
1687 	int			error;
1688 	int			ialloced;
1689 	int			noroom = 0;
1690 	xfs_agnumber_t		start_agno;
1691 	struct xfs_perag	*pag;
1692 	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
1693 	int			okalloc = 1;
1694 
1695 	if (*IO_agbp) {
1696 		/*
1697 		 * If the caller passes in a pointer to the AGI buffer,
1698 		 * continue where we left off before.  In this case, we
1699 		 * know that the allocation group has free inodes.
1700 		 */
1701 		agbp = *IO_agbp;
1702 		goto out_alloc;
1703 	}
1704 
1705 	/*
1706 	 * We do not have an agbp, so select an initial allocation
1707 	 * group for inode allocation.
1708 	 */
1709 	start_agno = xfs_ialloc_ag_select(tp, parent, mode);
1710 	if (start_agno == NULLAGNUMBER) {
1711 		*inop = NULLFSINO;
1712 		return 0;
1713 	}
1714 
1715 	/*
1716 	 * If we have already hit the ceiling of inode blocks then clear
1717 	 * okalloc so we scan all available agi structures for a free
1718 	 * inode.
1719 	 *
1720 	 * Read rough value of mp->m_icount by percpu_counter_read_positive,
1721 	 * which will sacrifice the preciseness but improve the performance.
1722 	 */
1723 	if (igeo->maxicount &&
1724 	    percpu_counter_read_positive(&mp->m_icount) + igeo->ialloc_inos
1725 							> igeo->maxicount) {
1726 		noroom = 1;
1727 		okalloc = 0;
1728 	}
1729 
1730 	/*
1731 	 * Loop until we find an allocation group that either has free inodes
1732 	 * or in which we can allocate some inodes.  Iterate through the
1733 	 * allocation groups upward, wrapping at the end.
1734 	 */
1735 	agno = start_agno;
1736 	for (;;) {
1737 		pag = xfs_perag_get(mp, agno);
1738 		if (!pag->pagi_inodeok) {
1739 			xfs_ialloc_next_ag(mp);
1740 			goto nextag;
1741 		}
1742 
1743 		if (!pag->pagi_init) {
1744 			error = xfs_ialloc_pagi_init(mp, tp, agno);
1745 			if (error)
1746 				goto out_error;
1747 		}
1748 
1749 		/*
1750 		 * Do a first racy fast path check if this AG is usable.
1751 		 */
1752 		if (!pag->pagi_freecount && !okalloc)
1753 			goto nextag;
1754 
1755 		/*
1756 		 * Then read in the AGI buffer and recheck with the AGI buffer
1757 		 * lock held.
1758 		 */
1759 		error = xfs_ialloc_read_agi(mp, tp, agno, &agbp);
1760 		if (error)
1761 			goto out_error;
1762 
1763 		if (pag->pagi_freecount) {
1764 			xfs_perag_put(pag);
1765 			goto out_alloc;
1766 		}
1767 
1768 		if (!okalloc)
1769 			goto nextag_relse_buffer;
1770 
1771 
1772 		error = xfs_ialloc_ag_alloc(tp, agbp, &ialloced);
1773 		if (error) {
1774 			xfs_trans_brelse(tp, agbp);
1775 
1776 			if (error != -ENOSPC)
1777 				goto out_error;
1778 
1779 			xfs_perag_put(pag);
1780 			*inop = NULLFSINO;
1781 			return 0;
1782 		}
1783 
1784 		if (ialloced) {
1785 			/*
1786 			 * We successfully allocated some inodes, return
1787 			 * the current context to the caller so that it
1788 			 * can commit the current transaction and call
1789 			 * us again where we left off.
1790 			 */
1791 			ASSERT(pag->pagi_freecount > 0);
1792 			xfs_perag_put(pag);
1793 
1794 			*IO_agbp = agbp;
1795 			*inop = NULLFSINO;
1796 			return 0;
1797 		}
1798 
1799 nextag_relse_buffer:
1800 		xfs_trans_brelse(tp, agbp);
1801 nextag:
1802 		xfs_perag_put(pag);
1803 		if (++agno == mp->m_sb.sb_agcount)
1804 			agno = 0;
1805 		if (agno == start_agno) {
1806 			*inop = NULLFSINO;
1807 			return noroom ? -ENOSPC : 0;
1808 		}
1809 	}
1810 
1811 out_alloc:
1812 	*IO_agbp = NULL;
1813 	return xfs_dialloc_ag(tp, agbp, parent, inop);
1814 out_error:
1815 	xfs_perag_put(pag);
1816 	return error;
1817 }
1818 
1819 /*
1820  * Free the blocks of an inode chunk. We must consider that the inode chunk
1821  * might be sparse and only free the regions that are allocated as part of the
1822  * chunk.
1823  */
1824 STATIC void
1825 xfs_difree_inode_chunk(
1826 	struct xfs_trans		*tp,
1827 	xfs_agnumber_t			agno,
1828 	struct xfs_inobt_rec_incore	*rec)
1829 {
1830 	struct xfs_mount		*mp = tp->t_mountp;
1831 	xfs_agblock_t			sagbno = XFS_AGINO_TO_AGBNO(mp,
1832 							rec->ir_startino);
1833 	int				startidx, endidx;
1834 	int				nextbit;
1835 	xfs_agblock_t			agbno;
1836 	int				contigblk;
1837 	DECLARE_BITMAP(holemask, XFS_INOBT_HOLEMASK_BITS);
1838 
1839 	if (!xfs_inobt_issparse(rec->ir_holemask)) {
1840 		/* not sparse, calculate extent info directly */
1841 		xfs_bmap_add_free(tp, XFS_AGB_TO_FSB(mp, agno, sagbno),
1842 				  M_IGEO(mp)->ialloc_blks,
1843 				  &XFS_RMAP_OINFO_INODES);
1844 		return;
1845 	}
1846 
1847 	/* holemask is only 16-bits (fits in an unsigned long) */
1848 	ASSERT(sizeof(rec->ir_holemask) <= sizeof(holemask[0]));
1849 	holemask[0] = rec->ir_holemask;
1850 
1851 	/*
1852 	 * Find contiguous ranges of zeroes (i.e., allocated regions) in the
1853 	 * holemask and convert the start/end index of each range to an extent.
1854 	 * We start with the start and end index both pointing at the first 0 in
1855 	 * the mask.
1856 	 */
1857 	startidx = endidx = find_first_zero_bit(holemask,
1858 						XFS_INOBT_HOLEMASK_BITS);
1859 	nextbit = startidx + 1;
1860 	while (startidx < XFS_INOBT_HOLEMASK_BITS) {
1861 		nextbit = find_next_zero_bit(holemask, XFS_INOBT_HOLEMASK_BITS,
1862 					     nextbit);
1863 		/*
1864 		 * If the next zero bit is contiguous, update the end index of
1865 		 * the current range and continue.
1866 		 */
1867 		if (nextbit != XFS_INOBT_HOLEMASK_BITS &&
1868 		    nextbit == endidx + 1) {
1869 			endidx = nextbit;
1870 			goto next;
1871 		}
1872 
1873 		/*
1874 		 * nextbit is not contiguous with the current end index. Convert
1875 		 * the current start/end to an extent and add it to the free
1876 		 * list.
1877 		 */
1878 		agbno = sagbno + (startidx * XFS_INODES_PER_HOLEMASK_BIT) /
1879 				  mp->m_sb.sb_inopblock;
1880 		contigblk = ((endidx - startidx + 1) *
1881 			     XFS_INODES_PER_HOLEMASK_BIT) /
1882 			    mp->m_sb.sb_inopblock;
1883 
1884 		ASSERT(agbno % mp->m_sb.sb_spino_align == 0);
1885 		ASSERT(contigblk % mp->m_sb.sb_spino_align == 0);
1886 		xfs_bmap_add_free(tp, XFS_AGB_TO_FSB(mp, agno, agbno),
1887 				  contigblk, &XFS_RMAP_OINFO_INODES);
1888 
1889 		/* reset range to current bit and carry on... */
1890 		startidx = endidx = nextbit;
1891 
1892 next:
1893 		nextbit++;
1894 	}
1895 }
1896 
1897 STATIC int
1898 xfs_difree_inobt(
1899 	struct xfs_mount		*mp,
1900 	struct xfs_trans		*tp,
1901 	struct xfs_buf			*agbp,
1902 	xfs_agino_t			agino,
1903 	struct xfs_icluster		*xic,
1904 	struct xfs_inobt_rec_incore	*orec)
1905 {
1906 	struct xfs_agi			*agi = XFS_BUF_TO_AGI(agbp);
1907 	xfs_agnumber_t			agno = be32_to_cpu(agi->agi_seqno);
1908 	struct xfs_perag		*pag;
1909 	struct xfs_btree_cur		*cur;
1910 	struct xfs_inobt_rec_incore	rec;
1911 	int				ilen;
1912 	int				error;
1913 	int				i;
1914 	int				off;
1915 
1916 	ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
1917 	ASSERT(XFS_AGINO_TO_AGBNO(mp, agino) < be32_to_cpu(agi->agi_length));
1918 
1919 	/*
1920 	 * Initialize the cursor.
1921 	 */
1922 	cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
1923 
1924 	error = xfs_check_agi_freecount(cur, agi);
1925 	if (error)
1926 		goto error0;
1927 
1928 	/*
1929 	 * Look for the entry describing this inode.
1930 	 */
1931 	if ((error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i))) {
1932 		xfs_warn(mp, "%s: xfs_inobt_lookup() returned error %d.",
1933 			__func__, error);
1934 		goto error0;
1935 	}
1936 	XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error0);
1937 	error = xfs_inobt_get_rec(cur, &rec, &i);
1938 	if (error) {
1939 		xfs_warn(mp, "%s: xfs_inobt_get_rec() returned error %d.",
1940 			__func__, error);
1941 		goto error0;
1942 	}
1943 	XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error0);
1944 	/*
1945 	 * Get the offset in the inode chunk.
1946 	 */
1947 	off = agino - rec.ir_startino;
1948 	ASSERT(off >= 0 && off < XFS_INODES_PER_CHUNK);
1949 	ASSERT(!(rec.ir_free & XFS_INOBT_MASK(off)));
1950 	/*
1951 	 * Mark the inode free & increment the count.
1952 	 */
1953 	rec.ir_free |= XFS_INOBT_MASK(off);
1954 	rec.ir_freecount++;
1955 
1956 	/*
1957 	 * When an inode chunk is free, it becomes eligible for removal. Don't
1958 	 * remove the chunk if the block size is large enough for multiple inode
1959 	 * chunks (that might not be free).
1960 	 */
1961 	if (!(mp->m_flags & XFS_MOUNT_IKEEP) &&
1962 	    rec.ir_free == XFS_INOBT_ALL_FREE &&
1963 	    mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK) {
1964 		xic->deleted = true;
1965 		xic->first_ino = XFS_AGINO_TO_INO(mp, agno, rec.ir_startino);
1966 		xic->alloc = xfs_inobt_irec_to_allocmask(&rec);
1967 
1968 		/*
1969 		 * Remove the inode cluster from the AGI B+Tree, adjust the
1970 		 * AGI and Superblock inode counts, and mark the disk space
1971 		 * to be freed when the transaction is committed.
1972 		 */
1973 		ilen = rec.ir_freecount;
1974 		be32_add_cpu(&agi->agi_count, -ilen);
1975 		be32_add_cpu(&agi->agi_freecount, -(ilen - 1));
1976 		xfs_ialloc_log_agi(tp, agbp, XFS_AGI_COUNT | XFS_AGI_FREECOUNT);
1977 		pag = xfs_perag_get(mp, agno);
1978 		pag->pagi_freecount -= ilen - 1;
1979 		pag->pagi_count -= ilen;
1980 		xfs_perag_put(pag);
1981 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, -ilen);
1982 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -(ilen - 1));
1983 
1984 		if ((error = xfs_btree_delete(cur, &i))) {
1985 			xfs_warn(mp, "%s: xfs_btree_delete returned error %d.",
1986 				__func__, error);
1987 			goto error0;
1988 		}
1989 
1990 		xfs_difree_inode_chunk(tp, agno, &rec);
1991 	} else {
1992 		xic->deleted = false;
1993 
1994 		error = xfs_inobt_update(cur, &rec);
1995 		if (error) {
1996 			xfs_warn(mp, "%s: xfs_inobt_update returned error %d.",
1997 				__func__, error);
1998 			goto error0;
1999 		}
2000 
2001 		/*
2002 		 * Change the inode free counts and log the ag/sb changes.
2003 		 */
2004 		be32_add_cpu(&agi->agi_freecount, 1);
2005 		xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
2006 		pag = xfs_perag_get(mp, agno);
2007 		pag->pagi_freecount++;
2008 		xfs_perag_put(pag);
2009 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, 1);
2010 	}
2011 
2012 	error = xfs_check_agi_freecount(cur, agi);
2013 	if (error)
2014 		goto error0;
2015 
2016 	*orec = rec;
2017 	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2018 	return 0;
2019 
2020 error0:
2021 	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2022 	return error;
2023 }
2024 
2025 /*
2026  * Free an inode in the free inode btree.
2027  */
2028 STATIC int
2029 xfs_difree_finobt(
2030 	struct xfs_mount		*mp,
2031 	struct xfs_trans		*tp,
2032 	struct xfs_buf			*agbp,
2033 	xfs_agino_t			agino,
2034 	struct xfs_inobt_rec_incore	*ibtrec) /* inobt record */
2035 {
2036 	struct xfs_agi			*agi = XFS_BUF_TO_AGI(agbp);
2037 	xfs_agnumber_t			agno = be32_to_cpu(agi->agi_seqno);
2038 	struct xfs_btree_cur		*cur;
2039 	struct xfs_inobt_rec_incore	rec;
2040 	int				offset = agino - ibtrec->ir_startino;
2041 	int				error;
2042 	int				i;
2043 
2044 	cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_FINO);
2045 
2046 	error = xfs_inobt_lookup(cur, ibtrec->ir_startino, XFS_LOOKUP_EQ, &i);
2047 	if (error)
2048 		goto error;
2049 	if (i == 0) {
2050 		/*
2051 		 * If the record does not exist in the finobt, we must have just
2052 		 * freed an inode in a previously fully allocated chunk. If not,
2053 		 * something is out of sync.
2054 		 */
2055 		XFS_WANT_CORRUPTED_GOTO(mp, ibtrec->ir_freecount == 1, error);
2056 
2057 		error = xfs_inobt_insert_rec(cur, ibtrec->ir_holemask,
2058 					     ibtrec->ir_count,
2059 					     ibtrec->ir_freecount,
2060 					     ibtrec->ir_free, &i);
2061 		if (error)
2062 			goto error;
2063 		ASSERT(i == 1);
2064 
2065 		goto out;
2066 	}
2067 
2068 	/*
2069 	 * Read and update the existing record. We could just copy the ibtrec
2070 	 * across here, but that would defeat the purpose of having redundant
2071 	 * metadata. By making the modifications independently, we can catch
2072 	 * corruptions that we wouldn't see if we just copied from one record
2073 	 * to another.
2074 	 */
2075 	error = xfs_inobt_get_rec(cur, &rec, &i);
2076 	if (error)
2077 		goto error;
2078 	XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error);
2079 
2080 	rec.ir_free |= XFS_INOBT_MASK(offset);
2081 	rec.ir_freecount++;
2082 
2083 	XFS_WANT_CORRUPTED_GOTO(mp, (rec.ir_free == ibtrec->ir_free) &&
2084 				(rec.ir_freecount == ibtrec->ir_freecount),
2085 				error);
2086 
2087 	/*
2088 	 * The content of inobt records should always match between the inobt
2089 	 * and finobt. The lifecycle of records in the finobt is different from
2090 	 * the inobt in that the finobt only tracks records with at least one
2091 	 * free inode. Hence, if all of the inodes are free and we aren't
2092 	 * keeping inode chunks permanently on disk, remove the record.
2093 	 * Otherwise, update the record with the new information.
2094 	 *
2095 	 * Note that we currently can't free chunks when the block size is large
2096 	 * enough for multiple chunks. Leave the finobt record to remain in sync
2097 	 * with the inobt.
2098 	 */
2099 	if (rec.ir_free == XFS_INOBT_ALL_FREE &&
2100 	    mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK &&
2101 	    !(mp->m_flags & XFS_MOUNT_IKEEP)) {
2102 		error = xfs_btree_delete(cur, &i);
2103 		if (error)
2104 			goto error;
2105 		ASSERT(i == 1);
2106 	} else {
2107 		error = xfs_inobt_update(cur, &rec);
2108 		if (error)
2109 			goto error;
2110 	}
2111 
2112 out:
2113 	error = xfs_check_agi_freecount(cur, agi);
2114 	if (error)
2115 		goto error;
2116 
2117 	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2118 	return 0;
2119 
2120 error:
2121 	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2122 	return error;
2123 }
2124 
2125 /*
2126  * Free disk inode.  Carefully avoids touching the incore inode, all
2127  * manipulations incore are the caller's responsibility.
2128  * The on-disk inode is not changed by this operation, only the
2129  * btree (free inode mask) is changed.
2130  */
2131 int
2132 xfs_difree(
2133 	struct xfs_trans	*tp,		/* transaction pointer */
2134 	xfs_ino_t		inode,		/* inode to be freed */
2135 	struct xfs_icluster	*xic)	/* cluster info if deleted */
2136 {
2137 	/* REFERENCED */
2138 	xfs_agblock_t		agbno;	/* block number containing inode */
2139 	struct xfs_buf		*agbp;	/* buffer for allocation group header */
2140 	xfs_agino_t		agino;	/* allocation group inode number */
2141 	xfs_agnumber_t		agno;	/* allocation group number */
2142 	int			error;	/* error return value */
2143 	struct xfs_mount	*mp;	/* mount structure for filesystem */
2144 	struct xfs_inobt_rec_incore rec;/* btree record */
2145 
2146 	mp = tp->t_mountp;
2147 
2148 	/*
2149 	 * Break up inode number into its components.
2150 	 */
2151 	agno = XFS_INO_TO_AGNO(mp, inode);
2152 	if (agno >= mp->m_sb.sb_agcount)  {
2153 		xfs_warn(mp, "%s: agno >= mp->m_sb.sb_agcount (%d >= %d).",
2154 			__func__, agno, mp->m_sb.sb_agcount);
2155 		ASSERT(0);
2156 		return -EINVAL;
2157 	}
2158 	agino = XFS_INO_TO_AGINO(mp, inode);
2159 	if (inode != XFS_AGINO_TO_INO(mp, agno, agino))  {
2160 		xfs_warn(mp, "%s: inode != XFS_AGINO_TO_INO() (%llu != %llu).",
2161 			__func__, (unsigned long long)inode,
2162 			(unsigned long long)XFS_AGINO_TO_INO(mp, agno, agino));
2163 		ASSERT(0);
2164 		return -EINVAL;
2165 	}
2166 	agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2167 	if (agbno >= mp->m_sb.sb_agblocks)  {
2168 		xfs_warn(mp, "%s: agbno >= mp->m_sb.sb_agblocks (%d >= %d).",
2169 			__func__, agbno, mp->m_sb.sb_agblocks);
2170 		ASSERT(0);
2171 		return -EINVAL;
2172 	}
2173 	/*
2174 	 * Get the allocation group header.
2175 	 */
2176 	error = xfs_ialloc_read_agi(mp, tp, agno, &agbp);
2177 	if (error) {
2178 		xfs_warn(mp, "%s: xfs_ialloc_read_agi() returned error %d.",
2179 			__func__, error);
2180 		return error;
2181 	}
2182 
2183 	/*
2184 	 * Fix up the inode allocation btree.
2185 	 */
2186 	error = xfs_difree_inobt(mp, tp, agbp, agino, xic, &rec);
2187 	if (error)
2188 		goto error0;
2189 
2190 	/*
2191 	 * Fix up the free inode btree.
2192 	 */
2193 	if (xfs_sb_version_hasfinobt(&mp->m_sb)) {
2194 		error = xfs_difree_finobt(mp, tp, agbp, agino, &rec);
2195 		if (error)
2196 			goto error0;
2197 	}
2198 
2199 	return 0;
2200 
2201 error0:
2202 	return error;
2203 }
2204 
2205 STATIC int
2206 xfs_imap_lookup(
2207 	struct xfs_mount	*mp,
2208 	struct xfs_trans	*tp,
2209 	xfs_agnumber_t		agno,
2210 	xfs_agino_t		agino,
2211 	xfs_agblock_t		agbno,
2212 	xfs_agblock_t		*chunk_agbno,
2213 	xfs_agblock_t		*offset_agbno,
2214 	int			flags)
2215 {
2216 	struct xfs_inobt_rec_incore rec;
2217 	struct xfs_btree_cur	*cur;
2218 	struct xfs_buf		*agbp;
2219 	int			error;
2220 	int			i;
2221 
2222 	error = xfs_ialloc_read_agi(mp, tp, agno, &agbp);
2223 	if (error) {
2224 		xfs_alert(mp,
2225 			"%s: xfs_ialloc_read_agi() returned error %d, agno %d",
2226 			__func__, error, agno);
2227 		return error;
2228 	}
2229 
2230 	/*
2231 	 * Lookup the inode record for the given agino. If the record cannot be
2232 	 * found, then it's an invalid inode number and we should abort. Once
2233 	 * we have a record, we need to ensure it contains the inode number
2234 	 * we are looking up.
2235 	 */
2236 	cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
2237 	error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i);
2238 	if (!error) {
2239 		if (i)
2240 			error = xfs_inobt_get_rec(cur, &rec, &i);
2241 		if (!error && i == 0)
2242 			error = -EINVAL;
2243 	}
2244 
2245 	xfs_trans_brelse(tp, agbp);
2246 	xfs_btree_del_cursor(cur, error);
2247 	if (error)
2248 		return error;
2249 
2250 	/* check that the returned record contains the required inode */
2251 	if (rec.ir_startino > agino ||
2252 	    rec.ir_startino + M_IGEO(mp)->ialloc_inos <= agino)
2253 		return -EINVAL;
2254 
2255 	/* for untrusted inodes check it is allocated first */
2256 	if ((flags & XFS_IGET_UNTRUSTED) &&
2257 	    (rec.ir_free & XFS_INOBT_MASK(agino - rec.ir_startino)))
2258 		return -EINVAL;
2259 
2260 	*chunk_agbno = XFS_AGINO_TO_AGBNO(mp, rec.ir_startino);
2261 	*offset_agbno = agbno - *chunk_agbno;
2262 	return 0;
2263 }
2264 
2265 /*
2266  * Return the location of the inode in imap, for mapping it into a buffer.
2267  */
2268 int
2269 xfs_imap(
2270 	xfs_mount_t	 *mp,	/* file system mount structure */
2271 	xfs_trans_t	 *tp,	/* transaction pointer */
2272 	xfs_ino_t	ino,	/* inode to locate */
2273 	struct xfs_imap	*imap,	/* location map structure */
2274 	uint		flags)	/* flags for inode btree lookup */
2275 {
2276 	xfs_agblock_t	agbno;	/* block number of inode in the alloc group */
2277 	xfs_agino_t	agino;	/* inode number within alloc group */
2278 	xfs_agnumber_t	agno;	/* allocation group number */
2279 	xfs_agblock_t	chunk_agbno;	/* first block in inode chunk */
2280 	xfs_agblock_t	cluster_agbno;	/* first block in inode cluster */
2281 	int		error;	/* error code */
2282 	int		offset;	/* index of inode in its buffer */
2283 	xfs_agblock_t	offset_agbno;	/* blks from chunk start to inode */
2284 
2285 	ASSERT(ino != NULLFSINO);
2286 
2287 	/*
2288 	 * Split up the inode number into its parts.
2289 	 */
2290 	agno = XFS_INO_TO_AGNO(mp, ino);
2291 	agino = XFS_INO_TO_AGINO(mp, ino);
2292 	agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2293 	if (agno >= mp->m_sb.sb_agcount || agbno >= mp->m_sb.sb_agblocks ||
2294 	    ino != XFS_AGINO_TO_INO(mp, agno, agino)) {
2295 #ifdef DEBUG
2296 		/*
2297 		 * Don't output diagnostic information for untrusted inodes
2298 		 * as they can be invalid without implying corruption.
2299 		 */
2300 		if (flags & XFS_IGET_UNTRUSTED)
2301 			return -EINVAL;
2302 		if (agno >= mp->m_sb.sb_agcount) {
2303 			xfs_alert(mp,
2304 				"%s: agno (%d) >= mp->m_sb.sb_agcount (%d)",
2305 				__func__, agno, mp->m_sb.sb_agcount);
2306 		}
2307 		if (agbno >= mp->m_sb.sb_agblocks) {
2308 			xfs_alert(mp,
2309 		"%s: agbno (0x%llx) >= mp->m_sb.sb_agblocks (0x%lx)",
2310 				__func__, (unsigned long long)agbno,
2311 				(unsigned long)mp->m_sb.sb_agblocks);
2312 		}
2313 		if (ino != XFS_AGINO_TO_INO(mp, agno, agino)) {
2314 			xfs_alert(mp,
2315 		"%s: ino (0x%llx) != XFS_AGINO_TO_INO() (0x%llx)",
2316 				__func__, ino,
2317 				XFS_AGINO_TO_INO(mp, agno, agino));
2318 		}
2319 		xfs_stack_trace();
2320 #endif /* DEBUG */
2321 		return -EINVAL;
2322 	}
2323 
2324 	/*
2325 	 * For bulkstat and handle lookups, we have an untrusted inode number
2326 	 * that we have to verify is valid. We cannot do this just by reading
2327 	 * the inode buffer as it may have been unlinked and removed leaving
2328 	 * inodes in stale state on disk. Hence we have to do a btree lookup
2329 	 * in all cases where an untrusted inode number is passed.
2330 	 */
2331 	if (flags & XFS_IGET_UNTRUSTED) {
2332 		error = xfs_imap_lookup(mp, tp, agno, agino, agbno,
2333 					&chunk_agbno, &offset_agbno, flags);
2334 		if (error)
2335 			return error;
2336 		goto out_map;
2337 	}
2338 
2339 	/*
2340 	 * If the inode cluster size is the same as the blocksize or
2341 	 * smaller we get to the buffer by simple arithmetics.
2342 	 */
2343 	if (M_IGEO(mp)->blocks_per_cluster == 1) {
2344 		offset = XFS_INO_TO_OFFSET(mp, ino);
2345 		ASSERT(offset < mp->m_sb.sb_inopblock);
2346 
2347 		imap->im_blkno = XFS_AGB_TO_DADDR(mp, agno, agbno);
2348 		imap->im_len = XFS_FSB_TO_BB(mp, 1);
2349 		imap->im_boffset = (unsigned short)(offset <<
2350 							mp->m_sb.sb_inodelog);
2351 		return 0;
2352 	}
2353 
2354 	/*
2355 	 * If the inode chunks are aligned then use simple maths to
2356 	 * find the location. Otherwise we have to do a btree
2357 	 * lookup to find the location.
2358 	 */
2359 	if (M_IGEO(mp)->inoalign_mask) {
2360 		offset_agbno = agbno & M_IGEO(mp)->inoalign_mask;
2361 		chunk_agbno = agbno - offset_agbno;
2362 	} else {
2363 		error = xfs_imap_lookup(mp, tp, agno, agino, agbno,
2364 					&chunk_agbno, &offset_agbno, flags);
2365 		if (error)
2366 			return error;
2367 	}
2368 
2369 out_map:
2370 	ASSERT(agbno >= chunk_agbno);
2371 	cluster_agbno = chunk_agbno +
2372 		((offset_agbno / M_IGEO(mp)->blocks_per_cluster) *
2373 		 M_IGEO(mp)->blocks_per_cluster);
2374 	offset = ((agbno - cluster_agbno) * mp->m_sb.sb_inopblock) +
2375 		XFS_INO_TO_OFFSET(mp, ino);
2376 
2377 	imap->im_blkno = XFS_AGB_TO_DADDR(mp, agno, cluster_agbno);
2378 	imap->im_len = XFS_FSB_TO_BB(mp, M_IGEO(mp)->blocks_per_cluster);
2379 	imap->im_boffset = (unsigned short)(offset << mp->m_sb.sb_inodelog);
2380 
2381 	/*
2382 	 * If the inode number maps to a block outside the bounds
2383 	 * of the file system then return NULL rather than calling
2384 	 * read_buf and panicing when we get an error from the
2385 	 * driver.
2386 	 */
2387 	if ((imap->im_blkno + imap->im_len) >
2388 	    XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks)) {
2389 		xfs_alert(mp,
2390 	"%s: (im_blkno (0x%llx) + im_len (0x%llx)) > sb_dblocks (0x%llx)",
2391 			__func__, (unsigned long long) imap->im_blkno,
2392 			(unsigned long long) imap->im_len,
2393 			XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks));
2394 		return -EINVAL;
2395 	}
2396 	return 0;
2397 }
2398 
2399 /*
2400  * Log specified fields for the ag hdr (inode section). The growth of the agi
2401  * structure over time requires that we interpret the buffer as two logical
2402  * regions delineated by the end of the unlinked list. This is due to the size
2403  * of the hash table and its location in the middle of the agi.
2404  *
2405  * For example, a request to log a field before agi_unlinked and a field after
2406  * agi_unlinked could cause us to log the entire hash table and use an excessive
2407  * amount of log space. To avoid this behavior, log the region up through
2408  * agi_unlinked in one call and the region after agi_unlinked through the end of
2409  * the structure in another.
2410  */
2411 void
2412 xfs_ialloc_log_agi(
2413 	xfs_trans_t	*tp,		/* transaction pointer */
2414 	xfs_buf_t	*bp,		/* allocation group header buffer */
2415 	int		fields)		/* bitmask of fields to log */
2416 {
2417 	int			first;		/* first byte number */
2418 	int			last;		/* last byte number */
2419 	static const short	offsets[] = {	/* field starting offsets */
2420 					/* keep in sync with bit definitions */
2421 		offsetof(xfs_agi_t, agi_magicnum),
2422 		offsetof(xfs_agi_t, agi_versionnum),
2423 		offsetof(xfs_agi_t, agi_seqno),
2424 		offsetof(xfs_agi_t, agi_length),
2425 		offsetof(xfs_agi_t, agi_count),
2426 		offsetof(xfs_agi_t, agi_root),
2427 		offsetof(xfs_agi_t, agi_level),
2428 		offsetof(xfs_agi_t, agi_freecount),
2429 		offsetof(xfs_agi_t, agi_newino),
2430 		offsetof(xfs_agi_t, agi_dirino),
2431 		offsetof(xfs_agi_t, agi_unlinked),
2432 		offsetof(xfs_agi_t, agi_free_root),
2433 		offsetof(xfs_agi_t, agi_free_level),
2434 		sizeof(xfs_agi_t)
2435 	};
2436 #ifdef DEBUG
2437 	xfs_agi_t		*agi;	/* allocation group header */
2438 
2439 	agi = XFS_BUF_TO_AGI(bp);
2440 	ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
2441 #endif
2442 
2443 	/*
2444 	 * Compute byte offsets for the first and last fields in the first
2445 	 * region and log the agi buffer. This only logs up through
2446 	 * agi_unlinked.
2447 	 */
2448 	if (fields & XFS_AGI_ALL_BITS_R1) {
2449 		xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R1,
2450 				  &first, &last);
2451 		xfs_trans_log_buf(tp, bp, first, last);
2452 	}
2453 
2454 	/*
2455 	 * Mask off the bits in the first region and calculate the first and
2456 	 * last field offsets for any bits in the second region.
2457 	 */
2458 	fields &= ~XFS_AGI_ALL_BITS_R1;
2459 	if (fields) {
2460 		xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R2,
2461 				  &first, &last);
2462 		xfs_trans_log_buf(tp, bp, first, last);
2463 	}
2464 }
2465 
2466 static xfs_failaddr_t
2467 xfs_agi_verify(
2468 	struct xfs_buf	*bp)
2469 {
2470 	struct xfs_mount *mp = bp->b_mount;
2471 	struct xfs_agi	*agi = XFS_BUF_TO_AGI(bp);
2472 	int		i;
2473 
2474 	if (xfs_sb_version_hascrc(&mp->m_sb)) {
2475 		if (!uuid_equal(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid))
2476 			return __this_address;
2477 		if (!xfs_log_check_lsn(mp,
2478 				be64_to_cpu(XFS_BUF_TO_AGI(bp)->agi_lsn)))
2479 			return __this_address;
2480 	}
2481 
2482 	/*
2483 	 * Validate the magic number of the agi block.
2484 	 */
2485 	if (!xfs_verify_magic(bp, agi->agi_magicnum))
2486 		return __this_address;
2487 	if (!XFS_AGI_GOOD_VERSION(be32_to_cpu(agi->agi_versionnum)))
2488 		return __this_address;
2489 
2490 	if (be32_to_cpu(agi->agi_level) < 1 ||
2491 	    be32_to_cpu(agi->agi_level) > XFS_BTREE_MAXLEVELS)
2492 		return __this_address;
2493 
2494 	if (xfs_sb_version_hasfinobt(&mp->m_sb) &&
2495 	    (be32_to_cpu(agi->agi_free_level) < 1 ||
2496 	     be32_to_cpu(agi->agi_free_level) > XFS_BTREE_MAXLEVELS))
2497 		return __this_address;
2498 
2499 	/*
2500 	 * during growfs operations, the perag is not fully initialised,
2501 	 * so we can't use it for any useful checking. growfs ensures we can't
2502 	 * use it by using uncached buffers that don't have the perag attached
2503 	 * so we can detect and avoid this problem.
2504 	 */
2505 	if (bp->b_pag && be32_to_cpu(agi->agi_seqno) != bp->b_pag->pag_agno)
2506 		return __this_address;
2507 
2508 	for (i = 0; i < XFS_AGI_UNLINKED_BUCKETS; i++) {
2509 		if (agi->agi_unlinked[i] == cpu_to_be32(NULLAGINO))
2510 			continue;
2511 		if (!xfs_verify_ino(mp, be32_to_cpu(agi->agi_unlinked[i])))
2512 			return __this_address;
2513 	}
2514 
2515 	return NULL;
2516 }
2517 
2518 static void
2519 xfs_agi_read_verify(
2520 	struct xfs_buf	*bp)
2521 {
2522 	struct xfs_mount *mp = bp->b_mount;
2523 	xfs_failaddr_t	fa;
2524 
2525 	if (xfs_sb_version_hascrc(&mp->m_sb) &&
2526 	    !xfs_buf_verify_cksum(bp, XFS_AGI_CRC_OFF))
2527 		xfs_verifier_error(bp, -EFSBADCRC, __this_address);
2528 	else {
2529 		fa = xfs_agi_verify(bp);
2530 		if (XFS_TEST_ERROR(fa, mp, XFS_ERRTAG_IALLOC_READ_AGI))
2531 			xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2532 	}
2533 }
2534 
2535 static void
2536 xfs_agi_write_verify(
2537 	struct xfs_buf	*bp)
2538 {
2539 	struct xfs_mount	*mp = bp->b_mount;
2540 	struct xfs_buf_log_item	*bip = bp->b_log_item;
2541 	xfs_failaddr_t		fa;
2542 
2543 	fa = xfs_agi_verify(bp);
2544 	if (fa) {
2545 		xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2546 		return;
2547 	}
2548 
2549 	if (!xfs_sb_version_hascrc(&mp->m_sb))
2550 		return;
2551 
2552 	if (bip)
2553 		XFS_BUF_TO_AGI(bp)->agi_lsn = cpu_to_be64(bip->bli_item.li_lsn);
2554 	xfs_buf_update_cksum(bp, XFS_AGI_CRC_OFF);
2555 }
2556 
2557 const struct xfs_buf_ops xfs_agi_buf_ops = {
2558 	.name = "xfs_agi",
2559 	.magic = { cpu_to_be32(XFS_AGI_MAGIC), cpu_to_be32(XFS_AGI_MAGIC) },
2560 	.verify_read = xfs_agi_read_verify,
2561 	.verify_write = xfs_agi_write_verify,
2562 	.verify_struct = xfs_agi_verify,
2563 };
2564 
2565 /*
2566  * Read in the allocation group header (inode allocation section)
2567  */
2568 int
2569 xfs_read_agi(
2570 	struct xfs_mount	*mp,	/* file system mount structure */
2571 	struct xfs_trans	*tp,	/* transaction pointer */
2572 	xfs_agnumber_t		agno,	/* allocation group number */
2573 	struct xfs_buf		**bpp)	/* allocation group hdr buf */
2574 {
2575 	int			error;
2576 
2577 	trace_xfs_read_agi(mp, agno);
2578 
2579 	ASSERT(agno != NULLAGNUMBER);
2580 	error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
2581 			XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
2582 			XFS_FSS_TO_BB(mp, 1), 0, bpp, &xfs_agi_buf_ops);
2583 	if (error)
2584 		return error;
2585 	if (tp)
2586 		xfs_trans_buf_set_type(tp, *bpp, XFS_BLFT_AGI_BUF);
2587 
2588 	xfs_buf_set_ref(*bpp, XFS_AGI_REF);
2589 	return 0;
2590 }
2591 
2592 int
2593 xfs_ialloc_read_agi(
2594 	struct xfs_mount	*mp,	/* file system mount structure */
2595 	struct xfs_trans	*tp,	/* transaction pointer */
2596 	xfs_agnumber_t		agno,	/* allocation group number */
2597 	struct xfs_buf		**bpp)	/* allocation group hdr buf */
2598 {
2599 	struct xfs_agi		*agi;	/* allocation group header */
2600 	struct xfs_perag	*pag;	/* per allocation group data */
2601 	int			error;
2602 
2603 	trace_xfs_ialloc_read_agi(mp, agno);
2604 
2605 	error = xfs_read_agi(mp, tp, agno, bpp);
2606 	if (error)
2607 		return error;
2608 
2609 	agi = XFS_BUF_TO_AGI(*bpp);
2610 	pag = xfs_perag_get(mp, agno);
2611 	if (!pag->pagi_init) {
2612 		pag->pagi_freecount = be32_to_cpu(agi->agi_freecount);
2613 		pag->pagi_count = be32_to_cpu(agi->agi_count);
2614 		pag->pagi_init = 1;
2615 	}
2616 
2617 	/*
2618 	 * It's possible for these to be out of sync if
2619 	 * we are in the middle of a forced shutdown.
2620 	 */
2621 	ASSERT(pag->pagi_freecount == be32_to_cpu(agi->agi_freecount) ||
2622 		XFS_FORCED_SHUTDOWN(mp));
2623 	xfs_perag_put(pag);
2624 	return 0;
2625 }
2626 
2627 /*
2628  * Read in the agi to initialise the per-ag data in the mount structure
2629  */
2630 int
2631 xfs_ialloc_pagi_init(
2632 	xfs_mount_t	*mp,		/* file system mount structure */
2633 	xfs_trans_t	*tp,		/* transaction pointer */
2634 	xfs_agnumber_t	agno)		/* allocation group number */
2635 {
2636 	xfs_buf_t	*bp = NULL;
2637 	int		error;
2638 
2639 	error = xfs_ialloc_read_agi(mp, tp, agno, &bp);
2640 	if (error)
2641 		return error;
2642 	if (bp)
2643 		xfs_trans_brelse(tp, bp);
2644 	return 0;
2645 }
2646 
2647 /* Is there an inode record covering a given range of inode numbers? */
2648 int
2649 xfs_ialloc_has_inode_record(
2650 	struct xfs_btree_cur	*cur,
2651 	xfs_agino_t		low,
2652 	xfs_agino_t		high,
2653 	bool			*exists)
2654 {
2655 	struct xfs_inobt_rec_incore	irec;
2656 	xfs_agino_t		agino;
2657 	uint16_t		holemask;
2658 	int			has_record;
2659 	int			i;
2660 	int			error;
2661 
2662 	*exists = false;
2663 	error = xfs_inobt_lookup(cur, low, XFS_LOOKUP_LE, &has_record);
2664 	while (error == 0 && has_record) {
2665 		error = xfs_inobt_get_rec(cur, &irec, &has_record);
2666 		if (error || irec.ir_startino > high)
2667 			break;
2668 
2669 		agino = irec.ir_startino;
2670 		holemask = irec.ir_holemask;
2671 		for (i = 0; i < XFS_INOBT_HOLEMASK_BITS; holemask >>= 1,
2672 				i++, agino += XFS_INODES_PER_HOLEMASK_BIT) {
2673 			if (holemask & 1)
2674 				continue;
2675 			if (agino + XFS_INODES_PER_HOLEMASK_BIT > low &&
2676 					agino <= high) {
2677 				*exists = true;
2678 				return 0;
2679 			}
2680 		}
2681 
2682 		error = xfs_btree_increment(cur, 0, &has_record);
2683 	}
2684 	return error;
2685 }
2686 
2687 /* Is there an inode record covering a given extent? */
2688 int
2689 xfs_ialloc_has_inodes_at_extent(
2690 	struct xfs_btree_cur	*cur,
2691 	xfs_agblock_t		bno,
2692 	xfs_extlen_t		len,
2693 	bool			*exists)
2694 {
2695 	xfs_agino_t		low;
2696 	xfs_agino_t		high;
2697 
2698 	low = XFS_AGB_TO_AGINO(cur->bc_mp, bno);
2699 	high = XFS_AGB_TO_AGINO(cur->bc_mp, bno + len) - 1;
2700 
2701 	return xfs_ialloc_has_inode_record(cur, low, high, exists);
2702 }
2703 
2704 struct xfs_ialloc_count_inodes {
2705 	xfs_agino_t			count;
2706 	xfs_agino_t			freecount;
2707 };
2708 
2709 /* Record inode counts across all inobt records. */
2710 STATIC int
2711 xfs_ialloc_count_inodes_rec(
2712 	struct xfs_btree_cur		*cur,
2713 	union xfs_btree_rec		*rec,
2714 	void				*priv)
2715 {
2716 	struct xfs_inobt_rec_incore	irec;
2717 	struct xfs_ialloc_count_inodes	*ci = priv;
2718 
2719 	xfs_inobt_btrec_to_irec(cur->bc_mp, rec, &irec);
2720 	ci->count += irec.ir_count;
2721 	ci->freecount += irec.ir_freecount;
2722 
2723 	return 0;
2724 }
2725 
2726 /* Count allocated and free inodes under an inobt. */
2727 int
2728 xfs_ialloc_count_inodes(
2729 	struct xfs_btree_cur		*cur,
2730 	xfs_agino_t			*count,
2731 	xfs_agino_t			*freecount)
2732 {
2733 	struct xfs_ialloc_count_inodes	ci = {0};
2734 	int				error;
2735 
2736 	ASSERT(cur->bc_btnum == XFS_BTNUM_INO);
2737 	error = xfs_btree_query_all(cur, xfs_ialloc_count_inodes_rec, &ci);
2738 	if (error)
2739 		return error;
2740 
2741 	*count = ci.count;
2742 	*freecount = ci.freecount;
2743 	return 0;
2744 }
2745 
2746 /*
2747  * Initialize inode-related geometry information.
2748  *
2749  * Compute the inode btree min and max levels and set maxicount.
2750  *
2751  * Set the inode cluster size.  This may still be overridden by the file
2752  * system block size if it is larger than the chosen cluster size.
2753  *
2754  * For v5 filesystems, scale the cluster size with the inode size to keep a
2755  * constant ratio of inode per cluster buffer, but only if mkfs has set the
2756  * inode alignment value appropriately for larger cluster sizes.
2757  *
2758  * Then compute the inode cluster alignment information.
2759  */
2760 void
2761 xfs_ialloc_setup_geometry(
2762 	struct xfs_mount	*mp)
2763 {
2764 	struct xfs_sb		*sbp = &mp->m_sb;
2765 	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
2766 	uint64_t		icount;
2767 	uint			inodes;
2768 
2769 	/* Compute inode btree geometry. */
2770 	igeo->agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
2771 	igeo->inobt_mxr[0] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, 1);
2772 	igeo->inobt_mxr[1] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, 0);
2773 	igeo->inobt_mnr[0] = igeo->inobt_mxr[0] / 2;
2774 	igeo->inobt_mnr[1] = igeo->inobt_mxr[1] / 2;
2775 
2776 	igeo->ialloc_inos = max_t(uint16_t, XFS_INODES_PER_CHUNK,
2777 			sbp->sb_inopblock);
2778 	igeo->ialloc_blks = igeo->ialloc_inos >> sbp->sb_inopblog;
2779 
2780 	if (sbp->sb_spino_align)
2781 		igeo->ialloc_min_blks = sbp->sb_spino_align;
2782 	else
2783 		igeo->ialloc_min_blks = igeo->ialloc_blks;
2784 
2785 	/* Compute and fill in value of m_ino_geo.inobt_maxlevels. */
2786 	inodes = (1LL << XFS_INO_AGINO_BITS(mp)) >> XFS_INODES_PER_CHUNK_LOG;
2787 	igeo->inobt_maxlevels = xfs_btree_compute_maxlevels(igeo->inobt_mnr,
2788 			inodes);
2789 
2790 	/*
2791 	 * Set the maximum inode count for this filesystem, being careful not
2792 	 * to use obviously garbage sb_inopblog/sb_inopblock values.  Regular
2793 	 * users should never get here due to failing sb verification, but
2794 	 * certain users (xfs_db) need to be usable even with corrupt metadata.
2795 	 */
2796 	if (sbp->sb_imax_pct && igeo->ialloc_blks) {
2797 		/*
2798 		 * Make sure the maximum inode count is a multiple
2799 		 * of the units we allocate inodes in.
2800 		 */
2801 		icount = sbp->sb_dblocks * sbp->sb_imax_pct;
2802 		do_div(icount, 100);
2803 		do_div(icount, igeo->ialloc_blks);
2804 		igeo->maxicount = XFS_FSB_TO_INO(mp,
2805 				icount * igeo->ialloc_blks);
2806 	} else {
2807 		igeo->maxicount = 0;
2808 	}
2809 
2810 	/*
2811 	 * Compute the desired size of an inode cluster buffer size, which
2812 	 * starts at 8K and (on v5 filesystems) scales up with larger inode
2813 	 * sizes.
2814 	 *
2815 	 * Preserve the desired inode cluster size because the sparse inodes
2816 	 * feature uses that desired size (not the actual size) to compute the
2817 	 * sparse inode alignment.  The mount code validates this value, so we
2818 	 * cannot change the behavior.
2819 	 */
2820 	igeo->inode_cluster_size_raw = XFS_INODE_BIG_CLUSTER_SIZE;
2821 	if (xfs_sb_version_hascrc(&mp->m_sb)) {
2822 		int	new_size = igeo->inode_cluster_size_raw;
2823 
2824 		new_size *= mp->m_sb.sb_inodesize / XFS_DINODE_MIN_SIZE;
2825 		if (mp->m_sb.sb_inoalignmt >= XFS_B_TO_FSBT(mp, new_size))
2826 			igeo->inode_cluster_size_raw = new_size;
2827 	}
2828 
2829 	/* Calculate inode cluster ratios. */
2830 	if (igeo->inode_cluster_size_raw > mp->m_sb.sb_blocksize)
2831 		igeo->blocks_per_cluster = XFS_B_TO_FSBT(mp,
2832 				igeo->inode_cluster_size_raw);
2833 	else
2834 		igeo->blocks_per_cluster = 1;
2835 	igeo->inode_cluster_size = XFS_FSB_TO_B(mp, igeo->blocks_per_cluster);
2836 	igeo->inodes_per_cluster = XFS_FSB_TO_INO(mp, igeo->blocks_per_cluster);
2837 
2838 	/* Calculate inode cluster alignment. */
2839 	if (xfs_sb_version_hasalign(&mp->m_sb) &&
2840 	    mp->m_sb.sb_inoalignmt >= igeo->blocks_per_cluster)
2841 		igeo->cluster_align = mp->m_sb.sb_inoalignmt;
2842 	else
2843 		igeo->cluster_align = 1;
2844 	igeo->inoalign_mask = igeo->cluster_align - 1;
2845 	igeo->cluster_align_inodes = XFS_FSB_TO_INO(mp, igeo->cluster_align);
2846 
2847 	/*
2848 	 * If we are using stripe alignment, check whether
2849 	 * the stripe unit is a multiple of the inode alignment
2850 	 */
2851 	if (mp->m_dalign && igeo->inoalign_mask &&
2852 	    !(mp->m_dalign & igeo->inoalign_mask))
2853 		igeo->ialloc_align = mp->m_dalign;
2854 	else
2855 		igeo->ialloc_align = 0;
2856 }
2857