xref: /linux/fs/xfs/libxfs/xfs_ialloc.c (revision 0e9b70c1e3623fa110fb6be553e644524228ef60)
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_mount.h"
14 #include "xfs_inode.h"
15 #include "xfs_btree.h"
16 #include "xfs_ialloc.h"
17 #include "xfs_ialloc_btree.h"
18 #include "xfs_alloc.h"
19 #include "xfs_errortag.h"
20 #include "xfs_error.h"
21 #include "xfs_bmap.h"
22 #include "xfs_trans.h"
23 #include "xfs_buf_item.h"
24 #include "xfs_icreate_item.h"
25 #include "xfs_icache.h"
26 #include "xfs_trace.h"
27 #include "xfs_log.h"
28 #include "xfs_rmap.h"
29 #include "xfs_ag.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_has_sparseinodes(cur->bc_mp)) {
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 	const 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_has_sparseinodes(mp)) {
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 	union xfs_btree_rec		*rec;
109 	int				error;
110 	uint64_t			realfree;
111 
112 	error = xfs_btree_get_rec(cur, &rec, stat);
113 	if (error || *stat == 0)
114 		return error;
115 
116 	xfs_inobt_btrec_to_irec(mp, rec, irec);
117 
118 	if (!xfs_verify_agino(cur->bc_ag.pag, irec->ir_startino))
119 		goto out_bad_rec;
120 	if (irec->ir_count < XFS_INODES_PER_HOLEMASK_BIT ||
121 	    irec->ir_count > XFS_INODES_PER_CHUNK)
122 		goto out_bad_rec;
123 	if (irec->ir_freecount > XFS_INODES_PER_CHUNK)
124 		goto out_bad_rec;
125 
126 	/* if there are no holes, return the first available offset */
127 	if (!xfs_inobt_issparse(irec->ir_holemask))
128 		realfree = irec->ir_free;
129 	else
130 		realfree = irec->ir_free & xfs_inobt_irec_to_allocmask(irec);
131 	if (hweight64(realfree) != irec->ir_freecount)
132 		goto out_bad_rec;
133 
134 	return 0;
135 
136 out_bad_rec:
137 	xfs_warn(mp,
138 		"%s Inode BTree record corruption in AG %d detected!",
139 		cur->bc_btnum == XFS_BTNUM_INO ? "Used" : "Free",
140 		cur->bc_ag.pag->pag_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_perag	*pag,
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 	xfs_agino_t		thisino;
181 	int			i;
182 	int			error;
183 
184 	cur = xfs_inobt_init_cursor(pag, tp, agbp, btnum);
185 
186 	for (thisino = newino;
187 	     thisino < newino + newlen;
188 	     thisino += XFS_INODES_PER_CHUNK) {
189 		error = xfs_inobt_lookup(cur, thisino, XFS_LOOKUP_EQ, &i);
190 		if (error) {
191 			xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
192 			return error;
193 		}
194 		ASSERT(i == 0);
195 
196 		error = xfs_inobt_insert_rec(cur, XFS_INOBT_HOLEMASK_FULL,
197 					     XFS_INODES_PER_CHUNK,
198 					     XFS_INODES_PER_CHUNK,
199 					     XFS_INOBT_ALL_FREE, &i);
200 		if (error) {
201 			xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
202 			return error;
203 		}
204 		ASSERT(i == 1);
205 	}
206 
207 	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
208 
209 	return 0;
210 }
211 
212 /*
213  * Verify that the number of free inodes in the AGI is correct.
214  */
215 #ifdef DEBUG
216 static int
217 xfs_check_agi_freecount(
218 	struct xfs_btree_cur	*cur)
219 {
220 	if (cur->bc_nlevels == 1) {
221 		xfs_inobt_rec_incore_t rec;
222 		int		freecount = 0;
223 		int		error;
224 		int		i;
225 
226 		error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
227 		if (error)
228 			return error;
229 
230 		do {
231 			error = xfs_inobt_get_rec(cur, &rec, &i);
232 			if (error)
233 				return error;
234 
235 			if (i) {
236 				freecount += rec.ir_freecount;
237 				error = xfs_btree_increment(cur, 0, &i);
238 				if (error)
239 					return error;
240 			}
241 		} while (i == 1);
242 
243 		if (!xfs_is_shutdown(cur->bc_mp))
244 			ASSERT(freecount == cur->bc_ag.pag->pagi_freecount);
245 	}
246 	return 0;
247 }
248 #else
249 #define xfs_check_agi_freecount(cur)	0
250 #endif
251 
252 /*
253  * Initialise a new set of inodes. When called without a transaction context
254  * (e.g. from recovery) we initiate a delayed write of the inode buffers rather
255  * than logging them (which in a transaction context puts them into the AIL
256  * for writeback rather than the xfsbufd queue).
257  */
258 int
259 xfs_ialloc_inode_init(
260 	struct xfs_mount	*mp,
261 	struct xfs_trans	*tp,
262 	struct list_head	*buffer_list,
263 	int			icount,
264 	xfs_agnumber_t		agno,
265 	xfs_agblock_t		agbno,
266 	xfs_agblock_t		length,
267 	unsigned int		gen)
268 {
269 	struct xfs_buf		*fbuf;
270 	struct xfs_dinode	*free;
271 	int			nbufs;
272 	int			version;
273 	int			i, j;
274 	xfs_daddr_t		d;
275 	xfs_ino_t		ino = 0;
276 	int			error;
277 
278 	/*
279 	 * Loop over the new block(s), filling in the inodes.  For small block
280 	 * sizes, manipulate the inodes in buffers  which are multiples of the
281 	 * blocks size.
282 	 */
283 	nbufs = length / M_IGEO(mp)->blocks_per_cluster;
284 
285 	/*
286 	 * Figure out what version number to use in the inodes we create.  If
287 	 * the superblock version has caught up to the one that supports the new
288 	 * inode format, then use the new inode version.  Otherwise use the old
289 	 * version so that old kernels will continue to be able to use the file
290 	 * system.
291 	 *
292 	 * For v3 inodes, we also need to write the inode number into the inode,
293 	 * so calculate the first inode number of the chunk here as
294 	 * XFS_AGB_TO_AGINO() only works within a filesystem block, not
295 	 * across multiple filesystem blocks (such as a cluster) and so cannot
296 	 * be used in the cluster buffer loop below.
297 	 *
298 	 * Further, because we are writing the inode directly into the buffer
299 	 * and calculating a CRC on the entire inode, we have ot log the entire
300 	 * inode so that the entire range the CRC covers is present in the log.
301 	 * That means for v3 inode we log the entire buffer rather than just the
302 	 * inode cores.
303 	 */
304 	if (xfs_has_v3inodes(mp)) {
305 		version = 3;
306 		ino = XFS_AGINO_TO_INO(mp, agno, XFS_AGB_TO_AGINO(mp, agbno));
307 
308 		/*
309 		 * log the initialisation that is about to take place as an
310 		 * logical operation. This means the transaction does not
311 		 * need to log the physical changes to the inode buffers as log
312 		 * recovery will know what initialisation is actually needed.
313 		 * Hence we only need to log the buffers as "ordered" buffers so
314 		 * they track in the AIL as if they were physically logged.
315 		 */
316 		if (tp)
317 			xfs_icreate_log(tp, agno, agbno, icount,
318 					mp->m_sb.sb_inodesize, length, gen);
319 	} else
320 		version = 2;
321 
322 	for (j = 0; j < nbufs; j++) {
323 		/*
324 		 * Get the block.
325 		 */
326 		d = XFS_AGB_TO_DADDR(mp, agno, agbno +
327 				(j * M_IGEO(mp)->blocks_per_cluster));
328 		error = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
329 				mp->m_bsize * M_IGEO(mp)->blocks_per_cluster,
330 				XBF_UNMAPPED, &fbuf);
331 		if (error)
332 			return error;
333 
334 		/* Initialize the inode buffers and log them appropriately. */
335 		fbuf->b_ops = &xfs_inode_buf_ops;
336 		xfs_buf_zero(fbuf, 0, BBTOB(fbuf->b_length));
337 		for (i = 0; i < M_IGEO(mp)->inodes_per_cluster; i++) {
338 			int	ioffset = i << mp->m_sb.sb_inodelog;
339 
340 			free = xfs_make_iptr(mp, fbuf, i);
341 			free->di_magic = cpu_to_be16(XFS_DINODE_MAGIC);
342 			free->di_version = version;
343 			free->di_gen = cpu_to_be32(gen);
344 			free->di_next_unlinked = cpu_to_be32(NULLAGINO);
345 
346 			if (version == 3) {
347 				free->di_ino = cpu_to_be64(ino);
348 				ino++;
349 				uuid_copy(&free->di_uuid,
350 					  &mp->m_sb.sb_meta_uuid);
351 				xfs_dinode_calc_crc(mp, free);
352 			} else if (tp) {
353 				/* just log the inode core */
354 				xfs_trans_log_buf(tp, fbuf, ioffset,
355 					  ioffset + XFS_DINODE_SIZE(mp) - 1);
356 			}
357 		}
358 
359 		if (tp) {
360 			/*
361 			 * Mark the buffer as an inode allocation buffer so it
362 			 * sticks in AIL at the point of this allocation
363 			 * transaction. This ensures the they are on disk before
364 			 * the tail of the log can be moved past this
365 			 * transaction (i.e. by preventing relogging from moving
366 			 * it forward in the log).
367 			 */
368 			xfs_trans_inode_alloc_buf(tp, fbuf);
369 			if (version == 3) {
370 				/*
371 				 * Mark the buffer as ordered so that they are
372 				 * not physically logged in the transaction but
373 				 * still tracked in the AIL as part of the
374 				 * transaction and pin the log appropriately.
375 				 */
376 				xfs_trans_ordered_buf(tp, fbuf);
377 			}
378 		} else {
379 			fbuf->b_flags |= XBF_DONE;
380 			xfs_buf_delwri_queue(fbuf, buffer_list);
381 			xfs_buf_relse(fbuf);
382 		}
383 	}
384 	return 0;
385 }
386 
387 /*
388  * Align startino and allocmask for a recently allocated sparse chunk such that
389  * they are fit for insertion (or merge) into the on-disk inode btrees.
390  *
391  * Background:
392  *
393  * When enabled, sparse inode support increases the inode alignment from cluster
394  * size to inode chunk size. This means that the minimum range between two
395  * non-adjacent inode records in the inobt is large enough for a full inode
396  * record. This allows for cluster sized, cluster aligned block allocation
397  * without need to worry about whether the resulting inode record overlaps with
398  * another record in the tree. Without this basic rule, we would have to deal
399  * with the consequences of overlap by potentially undoing recent allocations in
400  * the inode allocation codepath.
401  *
402  * Because of this alignment rule (which is enforced on mount), there are two
403  * inobt possibilities for newly allocated sparse chunks. One is that the
404  * aligned inode record for the chunk covers a range of inodes not already
405  * covered in the inobt (i.e., it is safe to insert a new sparse record). The
406  * other is that a record already exists at the aligned startino that considers
407  * the newly allocated range as sparse. In the latter case, record content is
408  * merged in hope that sparse inode chunks fill to full chunks over time.
409  */
410 STATIC void
411 xfs_align_sparse_ino(
412 	struct xfs_mount		*mp,
413 	xfs_agino_t			*startino,
414 	uint16_t			*allocmask)
415 {
416 	xfs_agblock_t			agbno;
417 	xfs_agblock_t			mod;
418 	int				offset;
419 
420 	agbno = XFS_AGINO_TO_AGBNO(mp, *startino);
421 	mod = agbno % mp->m_sb.sb_inoalignmt;
422 	if (!mod)
423 		return;
424 
425 	/* calculate the inode offset and align startino */
426 	offset = XFS_AGB_TO_AGINO(mp, mod);
427 	*startino -= offset;
428 
429 	/*
430 	 * Since startino has been aligned down, left shift allocmask such that
431 	 * it continues to represent the same physical inodes relative to the
432 	 * new startino.
433 	 */
434 	*allocmask <<= offset / XFS_INODES_PER_HOLEMASK_BIT;
435 }
436 
437 /*
438  * Determine whether the source inode record can merge into the target. Both
439  * records must be sparse, the inode ranges must match and there must be no
440  * allocation overlap between the records.
441  */
442 STATIC bool
443 __xfs_inobt_can_merge(
444 	struct xfs_inobt_rec_incore	*trec,	/* tgt record */
445 	struct xfs_inobt_rec_incore	*srec)	/* src record */
446 {
447 	uint64_t			talloc;
448 	uint64_t			salloc;
449 
450 	/* records must cover the same inode range */
451 	if (trec->ir_startino != srec->ir_startino)
452 		return false;
453 
454 	/* both records must be sparse */
455 	if (!xfs_inobt_issparse(trec->ir_holemask) ||
456 	    !xfs_inobt_issparse(srec->ir_holemask))
457 		return false;
458 
459 	/* both records must track some inodes */
460 	if (!trec->ir_count || !srec->ir_count)
461 		return false;
462 
463 	/* can't exceed capacity of a full record */
464 	if (trec->ir_count + srec->ir_count > XFS_INODES_PER_CHUNK)
465 		return false;
466 
467 	/* verify there is no allocation overlap */
468 	talloc = xfs_inobt_irec_to_allocmask(trec);
469 	salloc = xfs_inobt_irec_to_allocmask(srec);
470 	if (talloc & salloc)
471 		return false;
472 
473 	return true;
474 }
475 
476 /*
477  * Merge the source inode record into the target. The caller must call
478  * __xfs_inobt_can_merge() to ensure the merge is valid.
479  */
480 STATIC void
481 __xfs_inobt_rec_merge(
482 	struct xfs_inobt_rec_incore	*trec,	/* target */
483 	struct xfs_inobt_rec_incore	*srec)	/* src */
484 {
485 	ASSERT(trec->ir_startino == srec->ir_startino);
486 
487 	/* combine the counts */
488 	trec->ir_count += srec->ir_count;
489 	trec->ir_freecount += srec->ir_freecount;
490 
491 	/*
492 	 * Merge the holemask and free mask. For both fields, 0 bits refer to
493 	 * allocated inodes. We combine the allocated ranges with bitwise AND.
494 	 */
495 	trec->ir_holemask &= srec->ir_holemask;
496 	trec->ir_free &= srec->ir_free;
497 }
498 
499 /*
500  * Insert a new sparse inode chunk into the associated inode btree. The inode
501  * record for the sparse chunk is pre-aligned to a startino that should match
502  * any pre-existing sparse inode record in the tree. This allows sparse chunks
503  * to fill over time.
504  *
505  * This function supports two modes of handling preexisting records depending on
506  * the merge flag. If merge is true, the provided record is merged with the
507  * existing record and updated in place. The merged record is returned in nrec.
508  * If merge is false, an existing record is replaced with the provided record.
509  * If no preexisting record exists, the provided record is always inserted.
510  *
511  * It is considered corruption if a merge is requested and not possible. Given
512  * the sparse inode alignment constraints, this should never happen.
513  */
514 STATIC int
515 xfs_inobt_insert_sprec(
516 	struct xfs_perag		*pag,
517 	struct xfs_trans		*tp,
518 	struct xfs_buf			*agbp,
519 	int				btnum,
520 	struct xfs_inobt_rec_incore	*nrec,	/* in/out: new/merged rec. */
521 	bool				merge)	/* merge or replace */
522 {
523 	struct xfs_mount		*mp = pag->pag_mount;
524 	struct xfs_btree_cur		*cur;
525 	int				error;
526 	int				i;
527 	struct xfs_inobt_rec_incore	rec;
528 
529 	cur = xfs_inobt_init_cursor(pag, tp, agbp, btnum);
530 
531 	/* the new record is pre-aligned so we know where to look */
532 	error = xfs_inobt_lookup(cur, nrec->ir_startino, XFS_LOOKUP_EQ, &i);
533 	if (error)
534 		goto error;
535 	/* if nothing there, insert a new record and return */
536 	if (i == 0) {
537 		error = xfs_inobt_insert_rec(cur, nrec->ir_holemask,
538 					     nrec->ir_count, nrec->ir_freecount,
539 					     nrec->ir_free, &i);
540 		if (error)
541 			goto error;
542 		if (XFS_IS_CORRUPT(mp, i != 1)) {
543 			error = -EFSCORRUPTED;
544 			goto error;
545 		}
546 
547 		goto out;
548 	}
549 
550 	/*
551 	 * A record exists at this startino. Merge or replace the record
552 	 * depending on what we've been asked to do.
553 	 */
554 	if (merge) {
555 		error = xfs_inobt_get_rec(cur, &rec, &i);
556 		if (error)
557 			goto error;
558 		if (XFS_IS_CORRUPT(mp, i != 1)) {
559 			error = -EFSCORRUPTED;
560 			goto error;
561 		}
562 		if (XFS_IS_CORRUPT(mp, rec.ir_startino != nrec->ir_startino)) {
563 			error = -EFSCORRUPTED;
564 			goto error;
565 		}
566 
567 		/*
568 		 * This should never fail. If we have coexisting records that
569 		 * cannot merge, something is seriously wrong.
570 		 */
571 		if (XFS_IS_CORRUPT(mp, !__xfs_inobt_can_merge(nrec, &rec))) {
572 			error = -EFSCORRUPTED;
573 			goto error;
574 		}
575 
576 		trace_xfs_irec_merge_pre(mp, pag->pag_agno, rec.ir_startino,
577 					 rec.ir_holemask, nrec->ir_startino,
578 					 nrec->ir_holemask);
579 
580 		/* merge to nrec to output the updated record */
581 		__xfs_inobt_rec_merge(nrec, &rec);
582 
583 		trace_xfs_irec_merge_post(mp, pag->pag_agno, nrec->ir_startino,
584 					  nrec->ir_holemask);
585 
586 		error = xfs_inobt_rec_check_count(mp, nrec);
587 		if (error)
588 			goto error;
589 	}
590 
591 	error = xfs_inobt_update(cur, nrec);
592 	if (error)
593 		goto error;
594 
595 out:
596 	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
597 	return 0;
598 error:
599 	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
600 	return error;
601 }
602 
603 /*
604  * Allocate new inodes in the allocation group specified by agbp.  Returns 0 if
605  * inodes were allocated in this AG; -EAGAIN if there was no space in this AG so
606  * the caller knows it can try another AG, a hard -ENOSPC when over the maximum
607  * inode count threshold, or the usual negative error code for other errors.
608  */
609 STATIC int
610 xfs_ialloc_ag_alloc(
611 	struct xfs_perag	*pag,
612 	struct xfs_trans	*tp,
613 	struct xfs_buf		*agbp)
614 {
615 	struct xfs_agi		*agi;
616 	struct xfs_alloc_arg	args;
617 	int			error;
618 	xfs_agino_t		newino;		/* new first inode's number */
619 	xfs_agino_t		newlen;		/* new number of inodes */
620 	int			isaligned = 0;	/* inode allocation at stripe */
621 						/* unit boundary */
622 	/* init. to full chunk */
623 	struct xfs_inobt_rec_incore rec;
624 	struct xfs_ino_geometry	*igeo = M_IGEO(tp->t_mountp);
625 	uint16_t		allocmask = (uint16_t) -1;
626 	int			do_sparse = 0;
627 
628 	memset(&args, 0, sizeof(args));
629 	args.tp = tp;
630 	args.mp = tp->t_mountp;
631 	args.fsbno = NULLFSBLOCK;
632 	args.oinfo = XFS_RMAP_OINFO_INODES;
633 	args.pag = pag;
634 
635 #ifdef DEBUG
636 	/* randomly do sparse inode allocations */
637 	if (xfs_has_sparseinodes(tp->t_mountp) &&
638 	    igeo->ialloc_min_blks < igeo->ialloc_blks)
639 		do_sparse = get_random_u32_below(2);
640 #endif
641 
642 	/*
643 	 * Locking will ensure that we don't have two callers in here
644 	 * at one time.
645 	 */
646 	newlen = igeo->ialloc_inos;
647 	if (igeo->maxicount &&
648 	    percpu_counter_read_positive(&args.mp->m_icount) + newlen >
649 							igeo->maxicount)
650 		return -ENOSPC;
651 	args.minlen = args.maxlen = igeo->ialloc_blks;
652 	/*
653 	 * First try to allocate inodes contiguous with the last-allocated
654 	 * chunk of inodes.  If the filesystem is striped, this will fill
655 	 * an entire stripe unit with inodes.
656 	 */
657 	agi = agbp->b_addr;
658 	newino = be32_to_cpu(agi->agi_newino);
659 	args.agbno = XFS_AGINO_TO_AGBNO(args.mp, newino) +
660 		     igeo->ialloc_blks;
661 	if (do_sparse)
662 		goto sparse_alloc;
663 	if (likely(newino != NULLAGINO &&
664 		  (args.agbno < be32_to_cpu(agi->agi_length)))) {
665 		args.prod = 1;
666 
667 		/*
668 		 * We need to take into account alignment here to ensure that
669 		 * we don't modify the free list if we fail to have an exact
670 		 * block. If we don't have an exact match, and every oher
671 		 * attempt allocation attempt fails, we'll end up cancelling
672 		 * a dirty transaction and shutting down.
673 		 *
674 		 * For an exact allocation, alignment must be 1,
675 		 * however we need to take cluster alignment into account when
676 		 * fixing up the freelist. Use the minalignslop field to
677 		 * indicate that extra blocks might be required for alignment,
678 		 * but not to use them in the actual exact allocation.
679 		 */
680 		args.alignment = 1;
681 		args.minalignslop = igeo->cluster_align - 1;
682 
683 		/* Allow space for the inode btree to split. */
684 		args.minleft = igeo->inobt_maxlevels;
685 		error = xfs_alloc_vextent_exact_bno(&args,
686 				XFS_AGB_TO_FSB(args.mp, pag->pag_agno,
687 						args.agbno));
688 		if (error)
689 			return error;
690 
691 		/*
692 		 * This request might have dirtied the transaction if the AG can
693 		 * satisfy the request, but the exact block was not available.
694 		 * If the allocation did fail, subsequent requests will relax
695 		 * the exact agbno requirement and increase the alignment
696 		 * instead. It is critical that the total size of the request
697 		 * (len + alignment + slop) does not increase from this point
698 		 * on, so reset minalignslop to ensure it is not included in
699 		 * subsequent requests.
700 		 */
701 		args.minalignslop = 0;
702 	}
703 
704 	if (unlikely(args.fsbno == NULLFSBLOCK)) {
705 		/*
706 		 * Set the alignment for the allocation.
707 		 * If stripe alignment is turned on then align at stripe unit
708 		 * boundary.
709 		 * If the cluster size is smaller than a filesystem block
710 		 * then we're doing I/O for inodes in filesystem block size
711 		 * pieces, so don't need alignment anyway.
712 		 */
713 		isaligned = 0;
714 		if (igeo->ialloc_align) {
715 			ASSERT(!xfs_has_noalign(args.mp));
716 			args.alignment = args.mp->m_dalign;
717 			isaligned = 1;
718 		} else
719 			args.alignment = igeo->cluster_align;
720 		/*
721 		 * Allocate a fixed-size extent of inodes.
722 		 */
723 		args.prod = 1;
724 		/*
725 		 * Allow space for the inode btree to split.
726 		 */
727 		args.minleft = igeo->inobt_maxlevels;
728 		error = xfs_alloc_vextent_near_bno(&args,
729 				XFS_AGB_TO_FSB(args.mp, pag->pag_agno,
730 						be32_to_cpu(agi->agi_root)));
731 		if (error)
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.alignment = igeo->cluster_align;
741 		error = xfs_alloc_vextent_near_bno(&args,
742 				XFS_AGB_TO_FSB(args.mp, pag->pag_agno,
743 						be32_to_cpu(agi->agi_root)));
744 		if (error)
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_has_sparseinodes(args.mp) &&
753 	    igeo->ialloc_min_blks < igeo->ialloc_blks &&
754 	    args.fsbno == NULLFSBLOCK) {
755 sparse_alloc:
756 		args.alignment = args.mp->m_sb.sb_spino_align;
757 		args.prod = 1;
758 
759 		args.minlen = igeo->ialloc_min_blks;
760 		args.maxlen = args.minlen;
761 
762 		/*
763 		 * The inode record will be aligned to full chunk size. We must
764 		 * prevent sparse allocation from AG boundaries that result in
765 		 * invalid inode records, such as records that start at agbno 0
766 		 * or extend beyond the AG.
767 		 *
768 		 * Set min agbno to the first aligned, non-zero agbno and max to
769 		 * the last aligned agbno that is at least one full chunk from
770 		 * the end of the AG.
771 		 */
772 		args.min_agbno = args.mp->m_sb.sb_inoalignmt;
773 		args.max_agbno = round_down(args.mp->m_sb.sb_agblocks,
774 					    args.mp->m_sb.sb_inoalignmt) -
775 				 igeo->ialloc_blks;
776 
777 		error = xfs_alloc_vextent_near_bno(&args,
778 				XFS_AGB_TO_FSB(args.mp, pag->pag_agno,
779 						be32_to_cpu(agi->agi_root)));
780 		if (error)
781 			return error;
782 
783 		newlen = XFS_AGB_TO_AGINO(args.mp, args.len);
784 		ASSERT(newlen <= XFS_INODES_PER_CHUNK);
785 		allocmask = (1 << (newlen / XFS_INODES_PER_HOLEMASK_BIT)) - 1;
786 	}
787 
788 	if (args.fsbno == NULLFSBLOCK)
789 		return -EAGAIN;
790 
791 	ASSERT(args.len == args.minlen);
792 
793 	/*
794 	 * Stamp and write the inode buffers.
795 	 *
796 	 * Seed the new inode cluster with a random generation number. This
797 	 * prevents short-term reuse of generation numbers if a chunk is
798 	 * freed and then immediately reallocated. We use random numbers
799 	 * rather than a linear progression to prevent the next generation
800 	 * number from being easily guessable.
801 	 */
802 	error = xfs_ialloc_inode_init(args.mp, tp, NULL, newlen, pag->pag_agno,
803 			args.agbno, args.len, get_random_u32());
804 
805 	if (error)
806 		return error;
807 	/*
808 	 * Convert the results.
809 	 */
810 	newino = XFS_AGB_TO_AGINO(args.mp, args.agbno);
811 
812 	if (xfs_inobt_issparse(~allocmask)) {
813 		/*
814 		 * We've allocated a sparse chunk. Align the startino and mask.
815 		 */
816 		xfs_align_sparse_ino(args.mp, &newino, &allocmask);
817 
818 		rec.ir_startino = newino;
819 		rec.ir_holemask = ~allocmask;
820 		rec.ir_count = newlen;
821 		rec.ir_freecount = newlen;
822 		rec.ir_free = XFS_INOBT_ALL_FREE;
823 
824 		/*
825 		 * Insert the sparse record into the inobt and allow for a merge
826 		 * if necessary. If a merge does occur, rec is updated to the
827 		 * merged record.
828 		 */
829 		error = xfs_inobt_insert_sprec(pag, tp, agbp,
830 				XFS_BTNUM_INO, &rec, true);
831 		if (error == -EFSCORRUPTED) {
832 			xfs_alert(args.mp,
833 	"invalid sparse inode record: ino 0x%llx holemask 0x%x count %u",
834 				  XFS_AGINO_TO_INO(args.mp, pag->pag_agno,
835 						   rec.ir_startino),
836 				  rec.ir_holemask, rec.ir_count);
837 			xfs_force_shutdown(args.mp, SHUTDOWN_CORRUPT_INCORE);
838 		}
839 		if (error)
840 			return error;
841 
842 		/*
843 		 * We can't merge the part we've just allocated as for the inobt
844 		 * due to finobt semantics. The original record may or may not
845 		 * exist independent of whether physical inodes exist in this
846 		 * sparse chunk.
847 		 *
848 		 * We must update the finobt record based on the inobt record.
849 		 * rec contains the fully merged and up to date inobt record
850 		 * from the previous call. Set merge false to replace any
851 		 * existing record with this one.
852 		 */
853 		if (xfs_has_finobt(args.mp)) {
854 			error = xfs_inobt_insert_sprec(pag, tp, agbp,
855 				       XFS_BTNUM_FINO, &rec, false);
856 			if (error)
857 				return error;
858 		}
859 	} else {
860 		/* full chunk - insert new records to both btrees */
861 		error = xfs_inobt_insert(pag, tp, agbp, newino, newlen,
862 					 XFS_BTNUM_INO);
863 		if (error)
864 			return error;
865 
866 		if (xfs_has_finobt(args.mp)) {
867 			error = xfs_inobt_insert(pag, tp, agbp, newino,
868 						 newlen, XFS_BTNUM_FINO);
869 			if (error)
870 				return error;
871 		}
872 	}
873 
874 	/*
875 	 * Update AGI counts and newino.
876 	 */
877 	be32_add_cpu(&agi->agi_count, newlen);
878 	be32_add_cpu(&agi->agi_freecount, newlen);
879 	pag->pagi_freecount += newlen;
880 	pag->pagi_count += newlen;
881 	agi->agi_newino = cpu_to_be32(newino);
882 
883 	/*
884 	 * Log allocation group header fields
885 	 */
886 	xfs_ialloc_log_agi(tp, agbp,
887 		XFS_AGI_COUNT | XFS_AGI_FREECOUNT | XFS_AGI_NEWINO);
888 	/*
889 	 * Modify/log superblock values for inode count and inode free count.
890 	 */
891 	xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, (long)newlen);
892 	xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, (long)newlen);
893 	return 0;
894 }
895 
896 /*
897  * Try to retrieve the next record to the left/right from the current one.
898  */
899 STATIC int
900 xfs_ialloc_next_rec(
901 	struct xfs_btree_cur	*cur,
902 	xfs_inobt_rec_incore_t	*rec,
903 	int			*done,
904 	int			left)
905 {
906 	int                     error;
907 	int			i;
908 
909 	if (left)
910 		error = xfs_btree_decrement(cur, 0, &i);
911 	else
912 		error = xfs_btree_increment(cur, 0, &i);
913 
914 	if (error)
915 		return error;
916 	*done = !i;
917 	if (i) {
918 		error = xfs_inobt_get_rec(cur, rec, &i);
919 		if (error)
920 			return error;
921 		if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
922 			return -EFSCORRUPTED;
923 	}
924 
925 	return 0;
926 }
927 
928 STATIC int
929 xfs_ialloc_get_rec(
930 	struct xfs_btree_cur	*cur,
931 	xfs_agino_t		agino,
932 	xfs_inobt_rec_incore_t	*rec,
933 	int			*done)
934 {
935 	int                     error;
936 	int			i;
937 
938 	error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_EQ, &i);
939 	if (error)
940 		return error;
941 	*done = !i;
942 	if (i) {
943 		error = xfs_inobt_get_rec(cur, rec, &i);
944 		if (error)
945 			return error;
946 		if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
947 			return -EFSCORRUPTED;
948 	}
949 
950 	return 0;
951 }
952 
953 /*
954  * Return the offset of the first free inode in the record. If the inode chunk
955  * is sparsely allocated, we convert the record holemask to inode granularity
956  * and mask off the unallocated regions from the inode free mask.
957  */
958 STATIC int
959 xfs_inobt_first_free_inode(
960 	struct xfs_inobt_rec_incore	*rec)
961 {
962 	xfs_inofree_t			realfree;
963 
964 	/* if there are no holes, return the first available offset */
965 	if (!xfs_inobt_issparse(rec->ir_holemask))
966 		return xfs_lowbit64(rec->ir_free);
967 
968 	realfree = xfs_inobt_irec_to_allocmask(rec);
969 	realfree &= rec->ir_free;
970 
971 	return xfs_lowbit64(realfree);
972 }
973 
974 /*
975  * Allocate an inode using the inobt-only algorithm.
976  */
977 STATIC int
978 xfs_dialloc_ag_inobt(
979 	struct xfs_perag	*pag,
980 	struct xfs_trans	*tp,
981 	struct xfs_buf		*agbp,
982 	xfs_ino_t		parent,
983 	xfs_ino_t		*inop)
984 {
985 	struct xfs_mount	*mp = tp->t_mountp;
986 	struct xfs_agi		*agi = agbp->b_addr;
987 	xfs_agnumber_t		pagno = XFS_INO_TO_AGNO(mp, parent);
988 	xfs_agino_t		pagino = XFS_INO_TO_AGINO(mp, parent);
989 	struct xfs_btree_cur	*cur, *tcur;
990 	struct xfs_inobt_rec_incore rec, trec;
991 	xfs_ino_t		ino;
992 	int			error;
993 	int			offset;
994 	int			i, j;
995 	int			searchdistance = 10;
996 
997 	ASSERT(xfs_perag_initialised_agi(pag));
998 	ASSERT(xfs_perag_allows_inodes(pag));
999 	ASSERT(pag->pagi_freecount > 0);
1000 
1001  restart_pagno:
1002 	cur = xfs_inobt_init_cursor(pag, tp, agbp, XFS_BTNUM_INO);
1003 	/*
1004 	 * If pagino is 0 (this is the root inode allocation) use newino.
1005 	 * This must work because we've just allocated some.
1006 	 */
1007 	if (!pagino)
1008 		pagino = be32_to_cpu(agi->agi_newino);
1009 
1010 	error = xfs_check_agi_freecount(cur);
1011 	if (error)
1012 		goto error0;
1013 
1014 	/*
1015 	 * If in the same AG as the parent, try to get near the parent.
1016 	 */
1017 	if (pagno == pag->pag_agno) {
1018 		int		doneleft;	/* done, to the left */
1019 		int		doneright;	/* done, to the right */
1020 
1021 		error = xfs_inobt_lookup(cur, pagino, XFS_LOOKUP_LE, &i);
1022 		if (error)
1023 			goto error0;
1024 		if (XFS_IS_CORRUPT(mp, i != 1)) {
1025 			error = -EFSCORRUPTED;
1026 			goto error0;
1027 		}
1028 
1029 		error = xfs_inobt_get_rec(cur, &rec, &j);
1030 		if (error)
1031 			goto error0;
1032 		if (XFS_IS_CORRUPT(mp, j != 1)) {
1033 			error = -EFSCORRUPTED;
1034 			goto error0;
1035 		}
1036 
1037 		if (rec.ir_freecount > 0) {
1038 			/*
1039 			 * Found a free inode in the same chunk
1040 			 * as the parent, done.
1041 			 */
1042 			goto alloc_inode;
1043 		}
1044 
1045 
1046 		/*
1047 		 * In the same AG as parent, but parent's chunk is full.
1048 		 */
1049 
1050 		/* duplicate the cursor, search left & right simultaneously */
1051 		error = xfs_btree_dup_cursor(cur, &tcur);
1052 		if (error)
1053 			goto error0;
1054 
1055 		/*
1056 		 * Skip to last blocks looked up if same parent inode.
1057 		 */
1058 		if (pagino != NULLAGINO &&
1059 		    pag->pagl_pagino == pagino &&
1060 		    pag->pagl_leftrec != NULLAGINO &&
1061 		    pag->pagl_rightrec != NULLAGINO) {
1062 			error = xfs_ialloc_get_rec(tcur, pag->pagl_leftrec,
1063 						   &trec, &doneleft);
1064 			if (error)
1065 				goto error1;
1066 
1067 			error = xfs_ialloc_get_rec(cur, pag->pagl_rightrec,
1068 						   &rec, &doneright);
1069 			if (error)
1070 				goto error1;
1071 		} else {
1072 			/* search left with tcur, back up 1 record */
1073 			error = xfs_ialloc_next_rec(tcur, &trec, &doneleft, 1);
1074 			if (error)
1075 				goto error1;
1076 
1077 			/* search right with cur, go forward 1 record. */
1078 			error = xfs_ialloc_next_rec(cur, &rec, &doneright, 0);
1079 			if (error)
1080 				goto error1;
1081 		}
1082 
1083 		/*
1084 		 * Loop until we find an inode chunk with a free inode.
1085 		 */
1086 		while (--searchdistance > 0 && (!doneleft || !doneright)) {
1087 			int	useleft;  /* using left inode chunk this time */
1088 
1089 			/* figure out the closer block if both are valid. */
1090 			if (!doneleft && !doneright) {
1091 				useleft = pagino -
1092 				 (trec.ir_startino + XFS_INODES_PER_CHUNK - 1) <
1093 				  rec.ir_startino - pagino;
1094 			} else {
1095 				useleft = !doneleft;
1096 			}
1097 
1098 			/* free inodes to the left? */
1099 			if (useleft && trec.ir_freecount) {
1100 				xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1101 				cur = tcur;
1102 
1103 				pag->pagl_leftrec = trec.ir_startino;
1104 				pag->pagl_rightrec = rec.ir_startino;
1105 				pag->pagl_pagino = pagino;
1106 				rec = trec;
1107 				goto alloc_inode;
1108 			}
1109 
1110 			/* free inodes to the right? */
1111 			if (!useleft && rec.ir_freecount) {
1112 				xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1113 
1114 				pag->pagl_leftrec = trec.ir_startino;
1115 				pag->pagl_rightrec = rec.ir_startino;
1116 				pag->pagl_pagino = pagino;
1117 				goto alloc_inode;
1118 			}
1119 
1120 			/* get next record to check */
1121 			if (useleft) {
1122 				error = xfs_ialloc_next_rec(tcur, &trec,
1123 								 &doneleft, 1);
1124 			} else {
1125 				error = xfs_ialloc_next_rec(cur, &rec,
1126 								 &doneright, 0);
1127 			}
1128 			if (error)
1129 				goto error1;
1130 		}
1131 
1132 		if (searchdistance <= 0) {
1133 			/*
1134 			 * Not in range - save last search
1135 			 * location and allocate a new inode
1136 			 */
1137 			xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1138 			pag->pagl_leftrec = trec.ir_startino;
1139 			pag->pagl_rightrec = rec.ir_startino;
1140 			pag->pagl_pagino = pagino;
1141 
1142 		} else {
1143 			/*
1144 			 * We've reached the end of the btree. because
1145 			 * we are only searching a small chunk of the
1146 			 * btree each search, there is obviously free
1147 			 * inodes closer to the parent inode than we
1148 			 * are now. restart the search again.
1149 			 */
1150 			pag->pagl_pagino = NULLAGINO;
1151 			pag->pagl_leftrec = NULLAGINO;
1152 			pag->pagl_rightrec = NULLAGINO;
1153 			xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1154 			xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1155 			goto restart_pagno;
1156 		}
1157 	}
1158 
1159 	/*
1160 	 * In a different AG from the parent.
1161 	 * See if the most recently allocated block has any free.
1162 	 */
1163 	if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1164 		error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1165 					 XFS_LOOKUP_EQ, &i);
1166 		if (error)
1167 			goto error0;
1168 
1169 		if (i == 1) {
1170 			error = xfs_inobt_get_rec(cur, &rec, &j);
1171 			if (error)
1172 				goto error0;
1173 
1174 			if (j == 1 && rec.ir_freecount > 0) {
1175 				/*
1176 				 * The last chunk allocated in the group
1177 				 * still has a free inode.
1178 				 */
1179 				goto alloc_inode;
1180 			}
1181 		}
1182 	}
1183 
1184 	/*
1185 	 * None left in the last group, search the whole AG
1186 	 */
1187 	error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1188 	if (error)
1189 		goto error0;
1190 	if (XFS_IS_CORRUPT(mp, i != 1)) {
1191 		error = -EFSCORRUPTED;
1192 		goto error0;
1193 	}
1194 
1195 	for (;;) {
1196 		error = xfs_inobt_get_rec(cur, &rec, &i);
1197 		if (error)
1198 			goto error0;
1199 		if (XFS_IS_CORRUPT(mp, i != 1)) {
1200 			error = -EFSCORRUPTED;
1201 			goto error0;
1202 		}
1203 		if (rec.ir_freecount > 0)
1204 			break;
1205 		error = xfs_btree_increment(cur, 0, &i);
1206 		if (error)
1207 			goto error0;
1208 		if (XFS_IS_CORRUPT(mp, i != 1)) {
1209 			error = -EFSCORRUPTED;
1210 			goto error0;
1211 		}
1212 	}
1213 
1214 alloc_inode:
1215 	offset = xfs_inobt_first_free_inode(&rec);
1216 	ASSERT(offset >= 0);
1217 	ASSERT(offset < XFS_INODES_PER_CHUNK);
1218 	ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1219 				   XFS_INODES_PER_CHUNK) == 0);
1220 	ino = XFS_AGINO_TO_INO(mp, pag->pag_agno, rec.ir_startino + offset);
1221 	rec.ir_free &= ~XFS_INOBT_MASK(offset);
1222 	rec.ir_freecount--;
1223 	error = xfs_inobt_update(cur, &rec);
1224 	if (error)
1225 		goto error0;
1226 	be32_add_cpu(&agi->agi_freecount, -1);
1227 	xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1228 	pag->pagi_freecount--;
1229 
1230 	error = xfs_check_agi_freecount(cur);
1231 	if (error)
1232 		goto error0;
1233 
1234 	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1235 	xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1236 	*inop = ino;
1237 	return 0;
1238 error1:
1239 	xfs_btree_del_cursor(tcur, XFS_BTREE_ERROR);
1240 error0:
1241 	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1242 	return error;
1243 }
1244 
1245 /*
1246  * Use the free inode btree to allocate an inode based on distance from the
1247  * parent. Note that the provided cursor may be deleted and replaced.
1248  */
1249 STATIC int
1250 xfs_dialloc_ag_finobt_near(
1251 	xfs_agino_t			pagino,
1252 	struct xfs_btree_cur		**ocur,
1253 	struct xfs_inobt_rec_incore	*rec)
1254 {
1255 	struct xfs_btree_cur		*lcur = *ocur;	/* left search cursor */
1256 	struct xfs_btree_cur		*rcur;	/* right search cursor */
1257 	struct xfs_inobt_rec_incore	rrec;
1258 	int				error;
1259 	int				i, j;
1260 
1261 	error = xfs_inobt_lookup(lcur, pagino, XFS_LOOKUP_LE, &i);
1262 	if (error)
1263 		return error;
1264 
1265 	if (i == 1) {
1266 		error = xfs_inobt_get_rec(lcur, rec, &i);
1267 		if (error)
1268 			return error;
1269 		if (XFS_IS_CORRUPT(lcur->bc_mp, i != 1))
1270 			return -EFSCORRUPTED;
1271 
1272 		/*
1273 		 * See if we've landed in the parent inode record. The finobt
1274 		 * only tracks chunks with at least one free inode, so record
1275 		 * existence is enough.
1276 		 */
1277 		if (pagino >= rec->ir_startino &&
1278 		    pagino < (rec->ir_startino + XFS_INODES_PER_CHUNK))
1279 			return 0;
1280 	}
1281 
1282 	error = xfs_btree_dup_cursor(lcur, &rcur);
1283 	if (error)
1284 		return error;
1285 
1286 	error = xfs_inobt_lookup(rcur, pagino, XFS_LOOKUP_GE, &j);
1287 	if (error)
1288 		goto error_rcur;
1289 	if (j == 1) {
1290 		error = xfs_inobt_get_rec(rcur, &rrec, &j);
1291 		if (error)
1292 			goto error_rcur;
1293 		if (XFS_IS_CORRUPT(lcur->bc_mp, j != 1)) {
1294 			error = -EFSCORRUPTED;
1295 			goto error_rcur;
1296 		}
1297 	}
1298 
1299 	if (XFS_IS_CORRUPT(lcur->bc_mp, i != 1 && j != 1)) {
1300 		error = -EFSCORRUPTED;
1301 		goto error_rcur;
1302 	}
1303 	if (i == 1 && j == 1) {
1304 		/*
1305 		 * Both the left and right records are valid. Choose the closer
1306 		 * inode chunk to the target.
1307 		 */
1308 		if ((pagino - rec->ir_startino + XFS_INODES_PER_CHUNK - 1) >
1309 		    (rrec.ir_startino - pagino)) {
1310 			*rec = rrec;
1311 			xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1312 			*ocur = rcur;
1313 		} else {
1314 			xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1315 		}
1316 	} else if (j == 1) {
1317 		/* only the right record is valid */
1318 		*rec = rrec;
1319 		xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1320 		*ocur = rcur;
1321 	} else if (i == 1) {
1322 		/* only the left record is valid */
1323 		xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1324 	}
1325 
1326 	return 0;
1327 
1328 error_rcur:
1329 	xfs_btree_del_cursor(rcur, XFS_BTREE_ERROR);
1330 	return error;
1331 }
1332 
1333 /*
1334  * Use the free inode btree to find a free inode based on a newino hint. If
1335  * the hint is NULL, find the first free inode in the AG.
1336  */
1337 STATIC int
1338 xfs_dialloc_ag_finobt_newino(
1339 	struct xfs_agi			*agi,
1340 	struct xfs_btree_cur		*cur,
1341 	struct xfs_inobt_rec_incore	*rec)
1342 {
1343 	int error;
1344 	int i;
1345 
1346 	if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1347 		error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1348 					 XFS_LOOKUP_EQ, &i);
1349 		if (error)
1350 			return error;
1351 		if (i == 1) {
1352 			error = xfs_inobt_get_rec(cur, rec, &i);
1353 			if (error)
1354 				return error;
1355 			if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1356 				return -EFSCORRUPTED;
1357 			return 0;
1358 		}
1359 	}
1360 
1361 	/*
1362 	 * Find the first inode available in the AG.
1363 	 */
1364 	error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1365 	if (error)
1366 		return error;
1367 	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1368 		return -EFSCORRUPTED;
1369 
1370 	error = xfs_inobt_get_rec(cur, rec, &i);
1371 	if (error)
1372 		return error;
1373 	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1374 		return -EFSCORRUPTED;
1375 
1376 	return 0;
1377 }
1378 
1379 /*
1380  * Update the inobt based on a modification made to the finobt. Also ensure that
1381  * the records from both trees are equivalent post-modification.
1382  */
1383 STATIC int
1384 xfs_dialloc_ag_update_inobt(
1385 	struct xfs_btree_cur		*cur,	/* inobt cursor */
1386 	struct xfs_inobt_rec_incore	*frec,	/* finobt record */
1387 	int				offset) /* inode offset */
1388 {
1389 	struct xfs_inobt_rec_incore	rec;
1390 	int				error;
1391 	int				i;
1392 
1393 	error = xfs_inobt_lookup(cur, frec->ir_startino, XFS_LOOKUP_EQ, &i);
1394 	if (error)
1395 		return error;
1396 	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1397 		return -EFSCORRUPTED;
1398 
1399 	error = xfs_inobt_get_rec(cur, &rec, &i);
1400 	if (error)
1401 		return error;
1402 	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1403 		return -EFSCORRUPTED;
1404 	ASSERT((XFS_AGINO_TO_OFFSET(cur->bc_mp, rec.ir_startino) %
1405 				   XFS_INODES_PER_CHUNK) == 0);
1406 
1407 	rec.ir_free &= ~XFS_INOBT_MASK(offset);
1408 	rec.ir_freecount--;
1409 
1410 	if (XFS_IS_CORRUPT(cur->bc_mp,
1411 			   rec.ir_free != frec->ir_free ||
1412 			   rec.ir_freecount != frec->ir_freecount))
1413 		return -EFSCORRUPTED;
1414 
1415 	return xfs_inobt_update(cur, &rec);
1416 }
1417 
1418 /*
1419  * Allocate an inode using the free inode btree, if available. Otherwise, fall
1420  * back to the inobt search algorithm.
1421  *
1422  * The caller selected an AG for us, and made sure that free inodes are
1423  * available.
1424  */
1425 static int
1426 xfs_dialloc_ag(
1427 	struct xfs_perag	*pag,
1428 	struct xfs_trans	*tp,
1429 	struct xfs_buf		*agbp,
1430 	xfs_ino_t		parent,
1431 	xfs_ino_t		*inop)
1432 {
1433 	struct xfs_mount		*mp = tp->t_mountp;
1434 	struct xfs_agi			*agi = agbp->b_addr;
1435 	xfs_agnumber_t			pagno = XFS_INO_TO_AGNO(mp, parent);
1436 	xfs_agino_t			pagino = XFS_INO_TO_AGINO(mp, parent);
1437 	struct xfs_btree_cur		*cur;	/* finobt cursor */
1438 	struct xfs_btree_cur		*icur;	/* inobt cursor */
1439 	struct xfs_inobt_rec_incore	rec;
1440 	xfs_ino_t			ino;
1441 	int				error;
1442 	int				offset;
1443 	int				i;
1444 
1445 	if (!xfs_has_finobt(mp))
1446 		return xfs_dialloc_ag_inobt(pag, tp, agbp, parent, inop);
1447 
1448 	/*
1449 	 * If pagino is 0 (this is the root inode allocation) use newino.
1450 	 * This must work because we've just allocated some.
1451 	 */
1452 	if (!pagino)
1453 		pagino = be32_to_cpu(agi->agi_newino);
1454 
1455 	cur = xfs_inobt_init_cursor(pag, tp, agbp, XFS_BTNUM_FINO);
1456 
1457 	error = xfs_check_agi_freecount(cur);
1458 	if (error)
1459 		goto error_cur;
1460 
1461 	/*
1462 	 * The search algorithm depends on whether we're in the same AG as the
1463 	 * parent. If so, find the closest available inode to the parent. If
1464 	 * not, consider the agi hint or find the first free inode in the AG.
1465 	 */
1466 	if (pag->pag_agno == pagno)
1467 		error = xfs_dialloc_ag_finobt_near(pagino, &cur, &rec);
1468 	else
1469 		error = xfs_dialloc_ag_finobt_newino(agi, cur, &rec);
1470 	if (error)
1471 		goto error_cur;
1472 
1473 	offset = xfs_inobt_first_free_inode(&rec);
1474 	ASSERT(offset >= 0);
1475 	ASSERT(offset < XFS_INODES_PER_CHUNK);
1476 	ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1477 				   XFS_INODES_PER_CHUNK) == 0);
1478 	ino = XFS_AGINO_TO_INO(mp, pag->pag_agno, rec.ir_startino + offset);
1479 
1480 	/*
1481 	 * Modify or remove the finobt record.
1482 	 */
1483 	rec.ir_free &= ~XFS_INOBT_MASK(offset);
1484 	rec.ir_freecount--;
1485 	if (rec.ir_freecount)
1486 		error = xfs_inobt_update(cur, &rec);
1487 	else
1488 		error = xfs_btree_delete(cur, &i);
1489 	if (error)
1490 		goto error_cur;
1491 
1492 	/*
1493 	 * The finobt has now been updated appropriately. We haven't updated the
1494 	 * agi and superblock yet, so we can create an inobt cursor and validate
1495 	 * the original freecount. If all is well, make the equivalent update to
1496 	 * the inobt using the finobt record and offset information.
1497 	 */
1498 	icur = xfs_inobt_init_cursor(pag, tp, agbp, XFS_BTNUM_INO);
1499 
1500 	error = xfs_check_agi_freecount(icur);
1501 	if (error)
1502 		goto error_icur;
1503 
1504 	error = xfs_dialloc_ag_update_inobt(icur, &rec, offset);
1505 	if (error)
1506 		goto error_icur;
1507 
1508 	/*
1509 	 * Both trees have now been updated. We must update the perag and
1510 	 * superblock before we can check the freecount for each btree.
1511 	 */
1512 	be32_add_cpu(&agi->agi_freecount, -1);
1513 	xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1514 	pag->pagi_freecount--;
1515 
1516 	xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1517 
1518 	error = xfs_check_agi_freecount(icur);
1519 	if (error)
1520 		goto error_icur;
1521 	error = xfs_check_agi_freecount(cur);
1522 	if (error)
1523 		goto error_icur;
1524 
1525 	xfs_btree_del_cursor(icur, XFS_BTREE_NOERROR);
1526 	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1527 	*inop = ino;
1528 	return 0;
1529 
1530 error_icur:
1531 	xfs_btree_del_cursor(icur, XFS_BTREE_ERROR);
1532 error_cur:
1533 	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1534 	return error;
1535 }
1536 
1537 static int
1538 xfs_dialloc_roll(
1539 	struct xfs_trans	**tpp,
1540 	struct xfs_buf		*agibp)
1541 {
1542 	struct xfs_trans	*tp = *tpp;
1543 	struct xfs_dquot_acct	*dqinfo;
1544 	int			error;
1545 
1546 	/*
1547 	 * Hold to on to the agibp across the commit so no other allocation can
1548 	 * come in and take the free inodes we just allocated for our caller.
1549 	 */
1550 	xfs_trans_bhold(tp, agibp);
1551 
1552 	/*
1553 	 * We want the quota changes to be associated with the next transaction,
1554 	 * NOT this one. So, detach the dqinfo from this and attach it to the
1555 	 * next transaction.
1556 	 */
1557 	dqinfo = tp->t_dqinfo;
1558 	tp->t_dqinfo = NULL;
1559 
1560 	error = xfs_trans_roll(&tp);
1561 
1562 	/* Re-attach the quota info that we detached from prev trx. */
1563 	tp->t_dqinfo = dqinfo;
1564 
1565 	/*
1566 	 * Join the buffer even on commit error so that the buffer is released
1567 	 * when the caller cancels the transaction and doesn't have to handle
1568 	 * this error case specially.
1569 	 */
1570 	xfs_trans_bjoin(tp, agibp);
1571 	*tpp = tp;
1572 	return error;
1573 }
1574 
1575 static bool
1576 xfs_dialloc_good_ag(
1577 	struct xfs_perag	*pag,
1578 	struct xfs_trans	*tp,
1579 	umode_t			mode,
1580 	int			flags,
1581 	bool			ok_alloc)
1582 {
1583 	struct xfs_mount	*mp = tp->t_mountp;
1584 	xfs_extlen_t		ineed;
1585 	xfs_extlen_t		longest = 0;
1586 	int			needspace;
1587 	int			error;
1588 
1589 	if (!pag)
1590 		return false;
1591 	if (!xfs_perag_allows_inodes(pag))
1592 		return false;
1593 
1594 	if (!xfs_perag_initialised_agi(pag)) {
1595 		error = xfs_ialloc_read_agi(pag, tp, NULL);
1596 		if (error)
1597 			return false;
1598 	}
1599 
1600 	if (pag->pagi_freecount)
1601 		return true;
1602 	if (!ok_alloc)
1603 		return false;
1604 
1605 	if (!xfs_perag_initialised_agf(pag)) {
1606 		error = xfs_alloc_read_agf(pag, tp, flags, NULL);
1607 		if (error)
1608 			return false;
1609 	}
1610 
1611 	/*
1612 	 * Check that there is enough free space for the file plus a chunk of
1613 	 * inodes if we need to allocate some. If this is the first pass across
1614 	 * the AGs, take into account the potential space needed for alignment
1615 	 * of inode chunks when checking the longest contiguous free space in
1616 	 * the AG - this prevents us from getting ENOSPC because we have free
1617 	 * space larger than ialloc_blks but alignment constraints prevent us
1618 	 * from using it.
1619 	 *
1620 	 * If we can't find an AG with space for full alignment slack to be
1621 	 * taken into account, we must be near ENOSPC in all AGs.  Hence we
1622 	 * don't include alignment for the second pass and so if we fail
1623 	 * allocation due to alignment issues then it is most likely a real
1624 	 * ENOSPC condition.
1625 	 *
1626 	 * XXX(dgc): this calculation is now bogus thanks to the per-ag
1627 	 * reservations that xfs_alloc_fix_freelist() now does via
1628 	 * xfs_alloc_space_available(). When the AG fills up, pagf_freeblks will
1629 	 * be more than large enough for the check below to succeed, but
1630 	 * xfs_alloc_space_available() will fail because of the non-zero
1631 	 * metadata reservation and hence we won't actually be able to allocate
1632 	 * more inodes in this AG. We do soooo much unnecessary work near ENOSPC
1633 	 * because of this.
1634 	 */
1635 	ineed = M_IGEO(mp)->ialloc_min_blks;
1636 	if (flags && ineed > 1)
1637 		ineed += M_IGEO(mp)->cluster_align;
1638 	longest = pag->pagf_longest;
1639 	if (!longest)
1640 		longest = pag->pagf_flcount > 0;
1641 	needspace = S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode);
1642 
1643 	if (pag->pagf_freeblks < needspace + ineed || longest < ineed)
1644 		return false;
1645 	return true;
1646 }
1647 
1648 static int
1649 xfs_dialloc_try_ag(
1650 	struct xfs_perag	*pag,
1651 	struct xfs_trans	**tpp,
1652 	xfs_ino_t		parent,
1653 	xfs_ino_t		*new_ino,
1654 	bool			ok_alloc)
1655 {
1656 	struct xfs_buf		*agbp;
1657 	xfs_ino_t		ino;
1658 	int			error;
1659 
1660 	/*
1661 	 * Then read in the AGI buffer and recheck with the AGI buffer
1662 	 * lock held.
1663 	 */
1664 	error = xfs_ialloc_read_agi(pag, *tpp, &agbp);
1665 	if (error)
1666 		return error;
1667 
1668 	if (!pag->pagi_freecount) {
1669 		if (!ok_alloc) {
1670 			error = -EAGAIN;
1671 			goto out_release;
1672 		}
1673 
1674 		error = xfs_ialloc_ag_alloc(pag, *tpp, agbp);
1675 		if (error < 0)
1676 			goto out_release;
1677 
1678 		/*
1679 		 * We successfully allocated space for an inode cluster in this
1680 		 * AG.  Roll the transaction so that we can allocate one of the
1681 		 * new inodes.
1682 		 */
1683 		ASSERT(pag->pagi_freecount > 0);
1684 		error = xfs_dialloc_roll(tpp, agbp);
1685 		if (error)
1686 			goto out_release;
1687 	}
1688 
1689 	/* Allocate an inode in the found AG */
1690 	error = xfs_dialloc_ag(pag, *tpp, agbp, parent, &ino);
1691 	if (!error)
1692 		*new_ino = ino;
1693 	return error;
1694 
1695 out_release:
1696 	xfs_trans_brelse(*tpp, agbp);
1697 	return error;
1698 }
1699 
1700 /*
1701  * Allocate an on-disk inode.
1702  *
1703  * Mode is used to tell whether the new inode is a directory and hence where to
1704  * locate it. The on-disk inode that is allocated will be returned in @new_ino
1705  * on success, otherwise an error will be set to indicate the failure (e.g.
1706  * -ENOSPC).
1707  */
1708 int
1709 xfs_dialloc(
1710 	struct xfs_trans	**tpp,
1711 	xfs_ino_t		parent,
1712 	umode_t			mode,
1713 	xfs_ino_t		*new_ino)
1714 {
1715 	struct xfs_mount	*mp = (*tpp)->t_mountp;
1716 	xfs_agnumber_t		agno;
1717 	int			error = 0;
1718 	xfs_agnumber_t		start_agno;
1719 	struct xfs_perag	*pag;
1720 	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
1721 	bool			ok_alloc = true;
1722 	bool			low_space = false;
1723 	int			flags;
1724 	xfs_ino_t		ino = NULLFSINO;
1725 
1726 	/*
1727 	 * Directories, symlinks, and regular files frequently allocate at least
1728 	 * one block, so factor that potential expansion when we examine whether
1729 	 * an AG has enough space for file creation.
1730 	 */
1731 	if (S_ISDIR(mode))
1732 		start_agno = (atomic_inc_return(&mp->m_agirotor) - 1) %
1733 				mp->m_maxagi;
1734 	else {
1735 		start_agno = XFS_INO_TO_AGNO(mp, parent);
1736 		if (start_agno >= mp->m_maxagi)
1737 			start_agno = 0;
1738 	}
1739 
1740 	/*
1741 	 * If we have already hit the ceiling of inode blocks then clear
1742 	 * ok_alloc so we scan all available agi structures for a free
1743 	 * inode.
1744 	 *
1745 	 * Read rough value of mp->m_icount by percpu_counter_read_positive,
1746 	 * which will sacrifice the preciseness but improve the performance.
1747 	 */
1748 	if (igeo->maxicount &&
1749 	    percpu_counter_read_positive(&mp->m_icount) + igeo->ialloc_inos
1750 							> igeo->maxicount) {
1751 		ok_alloc = false;
1752 	}
1753 
1754 	/*
1755 	 * If we are near to ENOSPC, we want to prefer allocation from AGs that
1756 	 * have free inodes in them rather than use up free space allocating new
1757 	 * inode chunks. Hence we turn off allocation for the first non-blocking
1758 	 * pass through the AGs if we are near ENOSPC to consume free inodes
1759 	 * that we can immediately allocate, but then we allow allocation on the
1760 	 * second pass if we fail to find an AG with free inodes in it.
1761 	 */
1762 	if (percpu_counter_read_positive(&mp->m_fdblocks) <
1763 			mp->m_low_space[XFS_LOWSP_1_PCNT]) {
1764 		ok_alloc = false;
1765 		low_space = true;
1766 	}
1767 
1768 	/*
1769 	 * Loop until we find an allocation group that either has free inodes
1770 	 * or in which we can allocate some inodes.  Iterate through the
1771 	 * allocation groups upward, wrapping at the end.
1772 	 */
1773 	flags = XFS_ALLOC_FLAG_TRYLOCK;
1774 retry:
1775 	for_each_perag_wrap_at(mp, start_agno, mp->m_maxagi, agno, pag) {
1776 		if (xfs_dialloc_good_ag(pag, *tpp, mode, flags, ok_alloc)) {
1777 			error = xfs_dialloc_try_ag(pag, tpp, parent,
1778 					&ino, ok_alloc);
1779 			if (error != -EAGAIN)
1780 				break;
1781 			error = 0;
1782 		}
1783 
1784 		if (xfs_is_shutdown(mp)) {
1785 			error = -EFSCORRUPTED;
1786 			break;
1787 		}
1788 	}
1789 	if (pag)
1790 		xfs_perag_rele(pag);
1791 	if (error)
1792 		return error;
1793 	if (ino == NULLFSINO) {
1794 		if (flags) {
1795 			flags = 0;
1796 			if (low_space)
1797 				ok_alloc = true;
1798 			goto retry;
1799 		}
1800 		return -ENOSPC;
1801 	}
1802 	*new_ino = ino;
1803 	return 0;
1804 }
1805 
1806 /*
1807  * Free the blocks of an inode chunk. We must consider that the inode chunk
1808  * might be sparse and only free the regions that are allocated as part of the
1809  * chunk.
1810  */
1811 STATIC void
1812 xfs_difree_inode_chunk(
1813 	struct xfs_trans		*tp,
1814 	xfs_agnumber_t			agno,
1815 	struct xfs_inobt_rec_incore	*rec)
1816 {
1817 	struct xfs_mount		*mp = tp->t_mountp;
1818 	xfs_agblock_t			sagbno = XFS_AGINO_TO_AGBNO(mp,
1819 							rec->ir_startino);
1820 	int				startidx, endidx;
1821 	int				nextbit;
1822 	xfs_agblock_t			agbno;
1823 	int				contigblk;
1824 	DECLARE_BITMAP(holemask, XFS_INOBT_HOLEMASK_BITS);
1825 
1826 	if (!xfs_inobt_issparse(rec->ir_holemask)) {
1827 		/* not sparse, calculate extent info directly */
1828 		xfs_free_extent_later(tp, XFS_AGB_TO_FSB(mp, agno, sagbno),
1829 				  M_IGEO(mp)->ialloc_blks,
1830 				  &XFS_RMAP_OINFO_INODES);
1831 		return;
1832 	}
1833 
1834 	/* holemask is only 16-bits (fits in an unsigned long) */
1835 	ASSERT(sizeof(rec->ir_holemask) <= sizeof(holemask[0]));
1836 	holemask[0] = rec->ir_holemask;
1837 
1838 	/*
1839 	 * Find contiguous ranges of zeroes (i.e., allocated regions) in the
1840 	 * holemask and convert the start/end index of each range to an extent.
1841 	 * We start with the start and end index both pointing at the first 0 in
1842 	 * the mask.
1843 	 */
1844 	startidx = endidx = find_first_zero_bit(holemask,
1845 						XFS_INOBT_HOLEMASK_BITS);
1846 	nextbit = startidx + 1;
1847 	while (startidx < XFS_INOBT_HOLEMASK_BITS) {
1848 		nextbit = find_next_zero_bit(holemask, XFS_INOBT_HOLEMASK_BITS,
1849 					     nextbit);
1850 		/*
1851 		 * If the next zero bit is contiguous, update the end index of
1852 		 * the current range and continue.
1853 		 */
1854 		if (nextbit != XFS_INOBT_HOLEMASK_BITS &&
1855 		    nextbit == endidx + 1) {
1856 			endidx = nextbit;
1857 			goto next;
1858 		}
1859 
1860 		/*
1861 		 * nextbit is not contiguous with the current end index. Convert
1862 		 * the current start/end to an extent and add it to the free
1863 		 * list.
1864 		 */
1865 		agbno = sagbno + (startidx * XFS_INODES_PER_HOLEMASK_BIT) /
1866 				  mp->m_sb.sb_inopblock;
1867 		contigblk = ((endidx - startidx + 1) *
1868 			     XFS_INODES_PER_HOLEMASK_BIT) /
1869 			    mp->m_sb.sb_inopblock;
1870 
1871 		ASSERT(agbno % mp->m_sb.sb_spino_align == 0);
1872 		ASSERT(contigblk % mp->m_sb.sb_spino_align == 0);
1873 		xfs_free_extent_later(tp, XFS_AGB_TO_FSB(mp, agno, agbno),
1874 				  contigblk, &XFS_RMAP_OINFO_INODES);
1875 
1876 		/* reset range to current bit and carry on... */
1877 		startidx = endidx = nextbit;
1878 
1879 next:
1880 		nextbit++;
1881 	}
1882 }
1883 
1884 STATIC int
1885 xfs_difree_inobt(
1886 	struct xfs_perag		*pag,
1887 	struct xfs_trans		*tp,
1888 	struct xfs_buf			*agbp,
1889 	xfs_agino_t			agino,
1890 	struct xfs_icluster		*xic,
1891 	struct xfs_inobt_rec_incore	*orec)
1892 {
1893 	struct xfs_mount		*mp = pag->pag_mount;
1894 	struct xfs_agi			*agi = agbp->b_addr;
1895 	struct xfs_btree_cur		*cur;
1896 	struct xfs_inobt_rec_incore	rec;
1897 	int				ilen;
1898 	int				error;
1899 	int				i;
1900 	int				off;
1901 
1902 	ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
1903 	ASSERT(XFS_AGINO_TO_AGBNO(mp, agino) < be32_to_cpu(agi->agi_length));
1904 
1905 	/*
1906 	 * Initialize the cursor.
1907 	 */
1908 	cur = xfs_inobt_init_cursor(pag, tp, agbp, XFS_BTNUM_INO);
1909 
1910 	error = xfs_check_agi_freecount(cur);
1911 	if (error)
1912 		goto error0;
1913 
1914 	/*
1915 	 * Look for the entry describing this inode.
1916 	 */
1917 	if ((error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i))) {
1918 		xfs_warn(mp, "%s: xfs_inobt_lookup() returned error %d.",
1919 			__func__, error);
1920 		goto error0;
1921 	}
1922 	if (XFS_IS_CORRUPT(mp, i != 1)) {
1923 		error = -EFSCORRUPTED;
1924 		goto error0;
1925 	}
1926 	error = xfs_inobt_get_rec(cur, &rec, &i);
1927 	if (error) {
1928 		xfs_warn(mp, "%s: xfs_inobt_get_rec() returned error %d.",
1929 			__func__, error);
1930 		goto error0;
1931 	}
1932 	if (XFS_IS_CORRUPT(mp, i != 1)) {
1933 		error = -EFSCORRUPTED;
1934 		goto error0;
1935 	}
1936 	/*
1937 	 * Get the offset in the inode chunk.
1938 	 */
1939 	off = agino - rec.ir_startino;
1940 	ASSERT(off >= 0 && off < XFS_INODES_PER_CHUNK);
1941 	ASSERT(!(rec.ir_free & XFS_INOBT_MASK(off)));
1942 	/*
1943 	 * Mark the inode free & increment the count.
1944 	 */
1945 	rec.ir_free |= XFS_INOBT_MASK(off);
1946 	rec.ir_freecount++;
1947 
1948 	/*
1949 	 * When an inode chunk is free, it becomes eligible for removal. Don't
1950 	 * remove the chunk if the block size is large enough for multiple inode
1951 	 * chunks (that might not be free).
1952 	 */
1953 	if (!xfs_has_ikeep(mp) && rec.ir_free == XFS_INOBT_ALL_FREE &&
1954 	    mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK) {
1955 		struct xfs_perag	*pag = agbp->b_pag;
1956 
1957 		xic->deleted = true;
1958 		xic->first_ino = XFS_AGINO_TO_INO(mp, pag->pag_agno,
1959 				rec.ir_startino);
1960 		xic->alloc = xfs_inobt_irec_to_allocmask(&rec);
1961 
1962 		/*
1963 		 * Remove the inode cluster from the AGI B+Tree, adjust the
1964 		 * AGI and Superblock inode counts, and mark the disk space
1965 		 * to be freed when the transaction is committed.
1966 		 */
1967 		ilen = rec.ir_freecount;
1968 		be32_add_cpu(&agi->agi_count, -ilen);
1969 		be32_add_cpu(&agi->agi_freecount, -(ilen - 1));
1970 		xfs_ialloc_log_agi(tp, agbp, XFS_AGI_COUNT | XFS_AGI_FREECOUNT);
1971 		pag->pagi_freecount -= ilen - 1;
1972 		pag->pagi_count -= ilen;
1973 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, -ilen);
1974 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -(ilen - 1));
1975 
1976 		if ((error = xfs_btree_delete(cur, &i))) {
1977 			xfs_warn(mp, "%s: xfs_btree_delete returned error %d.",
1978 				__func__, error);
1979 			goto error0;
1980 		}
1981 
1982 		xfs_difree_inode_chunk(tp, pag->pag_agno, &rec);
1983 	} else {
1984 		xic->deleted = false;
1985 
1986 		error = xfs_inobt_update(cur, &rec);
1987 		if (error) {
1988 			xfs_warn(mp, "%s: xfs_inobt_update returned error %d.",
1989 				__func__, error);
1990 			goto error0;
1991 		}
1992 
1993 		/*
1994 		 * Change the inode free counts and log the ag/sb changes.
1995 		 */
1996 		be32_add_cpu(&agi->agi_freecount, 1);
1997 		xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1998 		pag->pagi_freecount++;
1999 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, 1);
2000 	}
2001 
2002 	error = xfs_check_agi_freecount(cur);
2003 	if (error)
2004 		goto error0;
2005 
2006 	*orec = rec;
2007 	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2008 	return 0;
2009 
2010 error0:
2011 	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2012 	return error;
2013 }
2014 
2015 /*
2016  * Free an inode in the free inode btree.
2017  */
2018 STATIC int
2019 xfs_difree_finobt(
2020 	struct xfs_perag		*pag,
2021 	struct xfs_trans		*tp,
2022 	struct xfs_buf			*agbp,
2023 	xfs_agino_t			agino,
2024 	struct xfs_inobt_rec_incore	*ibtrec) /* inobt record */
2025 {
2026 	struct xfs_mount		*mp = pag->pag_mount;
2027 	struct xfs_btree_cur		*cur;
2028 	struct xfs_inobt_rec_incore	rec;
2029 	int				offset = agino - ibtrec->ir_startino;
2030 	int				error;
2031 	int				i;
2032 
2033 	cur = xfs_inobt_init_cursor(pag, tp, agbp, XFS_BTNUM_FINO);
2034 
2035 	error = xfs_inobt_lookup(cur, ibtrec->ir_startino, XFS_LOOKUP_EQ, &i);
2036 	if (error)
2037 		goto error;
2038 	if (i == 0) {
2039 		/*
2040 		 * If the record does not exist in the finobt, we must have just
2041 		 * freed an inode in a previously fully allocated chunk. If not,
2042 		 * something is out of sync.
2043 		 */
2044 		if (XFS_IS_CORRUPT(mp, ibtrec->ir_freecount != 1)) {
2045 			error = -EFSCORRUPTED;
2046 			goto error;
2047 		}
2048 
2049 		error = xfs_inobt_insert_rec(cur, ibtrec->ir_holemask,
2050 					     ibtrec->ir_count,
2051 					     ibtrec->ir_freecount,
2052 					     ibtrec->ir_free, &i);
2053 		if (error)
2054 			goto error;
2055 		ASSERT(i == 1);
2056 
2057 		goto out;
2058 	}
2059 
2060 	/*
2061 	 * Read and update the existing record. We could just copy the ibtrec
2062 	 * across here, but that would defeat the purpose of having redundant
2063 	 * metadata. By making the modifications independently, we can catch
2064 	 * corruptions that we wouldn't see if we just copied from one record
2065 	 * to another.
2066 	 */
2067 	error = xfs_inobt_get_rec(cur, &rec, &i);
2068 	if (error)
2069 		goto error;
2070 	if (XFS_IS_CORRUPT(mp, i != 1)) {
2071 		error = -EFSCORRUPTED;
2072 		goto error;
2073 	}
2074 
2075 	rec.ir_free |= XFS_INOBT_MASK(offset);
2076 	rec.ir_freecount++;
2077 
2078 	if (XFS_IS_CORRUPT(mp,
2079 			   rec.ir_free != ibtrec->ir_free ||
2080 			   rec.ir_freecount != ibtrec->ir_freecount)) {
2081 		error = -EFSCORRUPTED;
2082 		goto error;
2083 	}
2084 
2085 	/*
2086 	 * The content of inobt records should always match between the inobt
2087 	 * and finobt. The lifecycle of records in the finobt is different from
2088 	 * the inobt in that the finobt only tracks records with at least one
2089 	 * free inode. Hence, if all of the inodes are free and we aren't
2090 	 * keeping inode chunks permanently on disk, remove the record.
2091 	 * Otherwise, update the record with the new information.
2092 	 *
2093 	 * Note that we currently can't free chunks when the block size is large
2094 	 * enough for multiple chunks. Leave the finobt record to remain in sync
2095 	 * with the inobt.
2096 	 */
2097 	if (!xfs_has_ikeep(mp) && rec.ir_free == XFS_INOBT_ALL_FREE &&
2098 	    mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK) {
2099 		error = xfs_btree_delete(cur, &i);
2100 		if (error)
2101 			goto error;
2102 		ASSERT(i == 1);
2103 	} else {
2104 		error = xfs_inobt_update(cur, &rec);
2105 		if (error)
2106 			goto error;
2107 	}
2108 
2109 out:
2110 	error = xfs_check_agi_freecount(cur);
2111 	if (error)
2112 		goto error;
2113 
2114 	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2115 	return 0;
2116 
2117 error:
2118 	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2119 	return error;
2120 }
2121 
2122 /*
2123  * Free disk inode.  Carefully avoids touching the incore inode, all
2124  * manipulations incore are the caller's responsibility.
2125  * The on-disk inode is not changed by this operation, only the
2126  * btree (free inode mask) is changed.
2127  */
2128 int
2129 xfs_difree(
2130 	struct xfs_trans	*tp,
2131 	struct xfs_perag	*pag,
2132 	xfs_ino_t		inode,
2133 	struct xfs_icluster	*xic)
2134 {
2135 	/* REFERENCED */
2136 	xfs_agblock_t		agbno;	/* block number containing inode */
2137 	struct xfs_buf		*agbp;	/* buffer for allocation group header */
2138 	xfs_agino_t		agino;	/* allocation group inode number */
2139 	int			error;	/* error return value */
2140 	struct xfs_mount	*mp = tp->t_mountp;
2141 	struct xfs_inobt_rec_incore rec;/* btree record */
2142 
2143 	/*
2144 	 * Break up inode number into its components.
2145 	 */
2146 	if (pag->pag_agno != XFS_INO_TO_AGNO(mp, inode)) {
2147 		xfs_warn(mp, "%s: agno != pag->pag_agno (%d != %d).",
2148 			__func__, XFS_INO_TO_AGNO(mp, inode), pag->pag_agno);
2149 		ASSERT(0);
2150 		return -EINVAL;
2151 	}
2152 	agino = XFS_INO_TO_AGINO(mp, inode);
2153 	if (inode != XFS_AGINO_TO_INO(mp, pag->pag_agno, agino))  {
2154 		xfs_warn(mp, "%s: inode != XFS_AGINO_TO_INO() (%llu != %llu).",
2155 			__func__, (unsigned long long)inode,
2156 			(unsigned long long)XFS_AGINO_TO_INO(mp, pag->pag_agno, agino));
2157 		ASSERT(0);
2158 		return -EINVAL;
2159 	}
2160 	agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2161 	if (agbno >= mp->m_sb.sb_agblocks)  {
2162 		xfs_warn(mp, "%s: agbno >= mp->m_sb.sb_agblocks (%d >= %d).",
2163 			__func__, agbno, mp->m_sb.sb_agblocks);
2164 		ASSERT(0);
2165 		return -EINVAL;
2166 	}
2167 	/*
2168 	 * Get the allocation group header.
2169 	 */
2170 	error = xfs_ialloc_read_agi(pag, tp, &agbp);
2171 	if (error) {
2172 		xfs_warn(mp, "%s: xfs_ialloc_read_agi() returned error %d.",
2173 			__func__, error);
2174 		return error;
2175 	}
2176 
2177 	/*
2178 	 * Fix up the inode allocation btree.
2179 	 */
2180 	error = xfs_difree_inobt(pag, tp, agbp, agino, xic, &rec);
2181 	if (error)
2182 		goto error0;
2183 
2184 	/*
2185 	 * Fix up the free inode btree.
2186 	 */
2187 	if (xfs_has_finobt(mp)) {
2188 		error = xfs_difree_finobt(pag, tp, agbp, agino, &rec);
2189 		if (error)
2190 			goto error0;
2191 	}
2192 
2193 	return 0;
2194 
2195 error0:
2196 	return error;
2197 }
2198 
2199 STATIC int
2200 xfs_imap_lookup(
2201 	struct xfs_perag	*pag,
2202 	struct xfs_trans	*tp,
2203 	xfs_agino_t		agino,
2204 	xfs_agblock_t		agbno,
2205 	xfs_agblock_t		*chunk_agbno,
2206 	xfs_agblock_t		*offset_agbno,
2207 	int			flags)
2208 {
2209 	struct xfs_mount	*mp = pag->pag_mount;
2210 	struct xfs_inobt_rec_incore rec;
2211 	struct xfs_btree_cur	*cur;
2212 	struct xfs_buf		*agbp;
2213 	int			error;
2214 	int			i;
2215 
2216 	error = xfs_ialloc_read_agi(pag, tp, &agbp);
2217 	if (error) {
2218 		xfs_alert(mp,
2219 			"%s: xfs_ialloc_read_agi() returned error %d, agno %d",
2220 			__func__, error, pag->pag_agno);
2221 		return error;
2222 	}
2223 
2224 	/*
2225 	 * Lookup the inode record for the given agino. If the record cannot be
2226 	 * found, then it's an invalid inode number and we should abort. Once
2227 	 * we have a record, we need to ensure it contains the inode number
2228 	 * we are looking up.
2229 	 */
2230 	cur = xfs_inobt_init_cursor(pag, tp, agbp, XFS_BTNUM_INO);
2231 	error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i);
2232 	if (!error) {
2233 		if (i)
2234 			error = xfs_inobt_get_rec(cur, &rec, &i);
2235 		if (!error && i == 0)
2236 			error = -EINVAL;
2237 	}
2238 
2239 	xfs_trans_brelse(tp, agbp);
2240 	xfs_btree_del_cursor(cur, error);
2241 	if (error)
2242 		return error;
2243 
2244 	/* check that the returned record contains the required inode */
2245 	if (rec.ir_startino > agino ||
2246 	    rec.ir_startino + M_IGEO(mp)->ialloc_inos <= agino)
2247 		return -EINVAL;
2248 
2249 	/* for untrusted inodes check it is allocated first */
2250 	if ((flags & XFS_IGET_UNTRUSTED) &&
2251 	    (rec.ir_free & XFS_INOBT_MASK(agino - rec.ir_startino)))
2252 		return -EINVAL;
2253 
2254 	*chunk_agbno = XFS_AGINO_TO_AGBNO(mp, rec.ir_startino);
2255 	*offset_agbno = agbno - *chunk_agbno;
2256 	return 0;
2257 }
2258 
2259 /*
2260  * Return the location of the inode in imap, for mapping it into a buffer.
2261  */
2262 int
2263 xfs_imap(
2264 	struct xfs_perag	*pag,
2265 	struct xfs_trans	*tp,
2266 	xfs_ino_t		ino,	/* inode to locate */
2267 	struct xfs_imap		*imap,	/* location map structure */
2268 	uint			flags)	/* flags for inode btree lookup */
2269 {
2270 	struct xfs_mount	*mp = pag->pag_mount;
2271 	xfs_agblock_t		agbno;	/* block number of inode in the alloc group */
2272 	xfs_agino_t		agino;	/* inode number within alloc group */
2273 	xfs_agblock_t		chunk_agbno;	/* first block in inode chunk */
2274 	xfs_agblock_t		cluster_agbno;	/* first block in inode cluster */
2275 	int			error;	/* error code */
2276 	int			offset;	/* index of inode in its buffer */
2277 	xfs_agblock_t		offset_agbno;	/* blks from chunk start to inode */
2278 
2279 	ASSERT(ino != NULLFSINO);
2280 
2281 	/*
2282 	 * Split up the inode number into its parts.
2283 	 */
2284 	agino = XFS_INO_TO_AGINO(mp, ino);
2285 	agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2286 	if (agbno >= mp->m_sb.sb_agblocks ||
2287 	    ino != XFS_AGINO_TO_INO(mp, pag->pag_agno, agino)) {
2288 		error = -EINVAL;
2289 #ifdef DEBUG
2290 		/*
2291 		 * Don't output diagnostic information for untrusted inodes
2292 		 * as they can be invalid without implying corruption.
2293 		 */
2294 		if (flags & XFS_IGET_UNTRUSTED)
2295 			return error;
2296 		if (agbno >= mp->m_sb.sb_agblocks) {
2297 			xfs_alert(mp,
2298 		"%s: agbno (0x%llx) >= mp->m_sb.sb_agblocks (0x%lx)",
2299 				__func__, (unsigned long long)agbno,
2300 				(unsigned long)mp->m_sb.sb_agblocks);
2301 		}
2302 		if (ino != XFS_AGINO_TO_INO(mp, pag->pag_agno, agino)) {
2303 			xfs_alert(mp,
2304 		"%s: ino (0x%llx) != XFS_AGINO_TO_INO() (0x%llx)",
2305 				__func__, ino,
2306 				XFS_AGINO_TO_INO(mp, pag->pag_agno, agino));
2307 		}
2308 		xfs_stack_trace();
2309 #endif /* DEBUG */
2310 		return error;
2311 	}
2312 
2313 	/*
2314 	 * For bulkstat and handle lookups, we have an untrusted inode number
2315 	 * that we have to verify is valid. We cannot do this just by reading
2316 	 * the inode buffer as it may have been unlinked and removed leaving
2317 	 * inodes in stale state on disk. Hence we have to do a btree lookup
2318 	 * in all cases where an untrusted inode number is passed.
2319 	 */
2320 	if (flags & XFS_IGET_UNTRUSTED) {
2321 		error = xfs_imap_lookup(pag, tp, agino, agbno,
2322 					&chunk_agbno, &offset_agbno, flags);
2323 		if (error)
2324 			return error;
2325 		goto out_map;
2326 	}
2327 
2328 	/*
2329 	 * If the inode cluster size is the same as the blocksize or
2330 	 * smaller we get to the buffer by simple arithmetics.
2331 	 */
2332 	if (M_IGEO(mp)->blocks_per_cluster == 1) {
2333 		offset = XFS_INO_TO_OFFSET(mp, ino);
2334 		ASSERT(offset < mp->m_sb.sb_inopblock);
2335 
2336 		imap->im_blkno = XFS_AGB_TO_DADDR(mp, pag->pag_agno, agbno);
2337 		imap->im_len = XFS_FSB_TO_BB(mp, 1);
2338 		imap->im_boffset = (unsigned short)(offset <<
2339 							mp->m_sb.sb_inodelog);
2340 		return 0;
2341 	}
2342 
2343 	/*
2344 	 * If the inode chunks are aligned then use simple maths to
2345 	 * find the location. Otherwise we have to do a btree
2346 	 * lookup to find the location.
2347 	 */
2348 	if (M_IGEO(mp)->inoalign_mask) {
2349 		offset_agbno = agbno & M_IGEO(mp)->inoalign_mask;
2350 		chunk_agbno = agbno - offset_agbno;
2351 	} else {
2352 		error = xfs_imap_lookup(pag, tp, agino, agbno,
2353 					&chunk_agbno, &offset_agbno, flags);
2354 		if (error)
2355 			return error;
2356 	}
2357 
2358 out_map:
2359 	ASSERT(agbno >= chunk_agbno);
2360 	cluster_agbno = chunk_agbno +
2361 		((offset_agbno / M_IGEO(mp)->blocks_per_cluster) *
2362 		 M_IGEO(mp)->blocks_per_cluster);
2363 	offset = ((agbno - cluster_agbno) * mp->m_sb.sb_inopblock) +
2364 		XFS_INO_TO_OFFSET(mp, ino);
2365 
2366 	imap->im_blkno = XFS_AGB_TO_DADDR(mp, pag->pag_agno, cluster_agbno);
2367 	imap->im_len = XFS_FSB_TO_BB(mp, M_IGEO(mp)->blocks_per_cluster);
2368 	imap->im_boffset = (unsigned short)(offset << mp->m_sb.sb_inodelog);
2369 
2370 	/*
2371 	 * If the inode number maps to a block outside the bounds
2372 	 * of the file system then return NULL rather than calling
2373 	 * read_buf and panicing when we get an error from the
2374 	 * driver.
2375 	 */
2376 	if ((imap->im_blkno + imap->im_len) >
2377 	    XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks)) {
2378 		xfs_alert(mp,
2379 	"%s: (im_blkno (0x%llx) + im_len (0x%llx)) > sb_dblocks (0x%llx)",
2380 			__func__, (unsigned long long) imap->im_blkno,
2381 			(unsigned long long) imap->im_len,
2382 			XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks));
2383 		return -EINVAL;
2384 	}
2385 	return 0;
2386 }
2387 
2388 /*
2389  * Log specified fields for the ag hdr (inode section). The growth of the agi
2390  * structure over time requires that we interpret the buffer as two logical
2391  * regions delineated by the end of the unlinked list. This is due to the size
2392  * of the hash table and its location in the middle of the agi.
2393  *
2394  * For example, a request to log a field before agi_unlinked and a field after
2395  * agi_unlinked could cause us to log the entire hash table and use an excessive
2396  * amount of log space. To avoid this behavior, log the region up through
2397  * agi_unlinked in one call and the region after agi_unlinked through the end of
2398  * the structure in another.
2399  */
2400 void
2401 xfs_ialloc_log_agi(
2402 	struct xfs_trans	*tp,
2403 	struct xfs_buf		*bp,
2404 	uint32_t		fields)
2405 {
2406 	int			first;		/* first byte number */
2407 	int			last;		/* last byte number */
2408 	static const short	offsets[] = {	/* field starting offsets */
2409 					/* keep in sync with bit definitions */
2410 		offsetof(xfs_agi_t, agi_magicnum),
2411 		offsetof(xfs_agi_t, agi_versionnum),
2412 		offsetof(xfs_agi_t, agi_seqno),
2413 		offsetof(xfs_agi_t, agi_length),
2414 		offsetof(xfs_agi_t, agi_count),
2415 		offsetof(xfs_agi_t, agi_root),
2416 		offsetof(xfs_agi_t, agi_level),
2417 		offsetof(xfs_agi_t, agi_freecount),
2418 		offsetof(xfs_agi_t, agi_newino),
2419 		offsetof(xfs_agi_t, agi_dirino),
2420 		offsetof(xfs_agi_t, agi_unlinked),
2421 		offsetof(xfs_agi_t, agi_free_root),
2422 		offsetof(xfs_agi_t, agi_free_level),
2423 		offsetof(xfs_agi_t, agi_iblocks),
2424 		sizeof(xfs_agi_t)
2425 	};
2426 #ifdef DEBUG
2427 	struct xfs_agi		*agi = bp->b_addr;
2428 
2429 	ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
2430 #endif
2431 
2432 	/*
2433 	 * Compute byte offsets for the first and last fields in the first
2434 	 * region and log the agi buffer. This only logs up through
2435 	 * agi_unlinked.
2436 	 */
2437 	if (fields & XFS_AGI_ALL_BITS_R1) {
2438 		xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R1,
2439 				  &first, &last);
2440 		xfs_trans_log_buf(tp, bp, first, last);
2441 	}
2442 
2443 	/*
2444 	 * Mask off the bits in the first region and calculate the first and
2445 	 * last field offsets for any bits in the second region.
2446 	 */
2447 	fields &= ~XFS_AGI_ALL_BITS_R1;
2448 	if (fields) {
2449 		xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R2,
2450 				  &first, &last);
2451 		xfs_trans_log_buf(tp, bp, first, last);
2452 	}
2453 }
2454 
2455 static xfs_failaddr_t
2456 xfs_agi_verify(
2457 	struct xfs_buf	*bp)
2458 {
2459 	struct xfs_mount *mp = bp->b_mount;
2460 	struct xfs_agi	*agi = bp->b_addr;
2461 	int		i;
2462 
2463 	if (xfs_has_crc(mp)) {
2464 		if (!uuid_equal(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid))
2465 			return __this_address;
2466 		if (!xfs_log_check_lsn(mp, be64_to_cpu(agi->agi_lsn)))
2467 			return __this_address;
2468 	}
2469 
2470 	/*
2471 	 * Validate the magic number of the agi block.
2472 	 */
2473 	if (!xfs_verify_magic(bp, agi->agi_magicnum))
2474 		return __this_address;
2475 	if (!XFS_AGI_GOOD_VERSION(be32_to_cpu(agi->agi_versionnum)))
2476 		return __this_address;
2477 
2478 	if (be32_to_cpu(agi->agi_level) < 1 ||
2479 	    be32_to_cpu(agi->agi_level) > M_IGEO(mp)->inobt_maxlevels)
2480 		return __this_address;
2481 
2482 	if (xfs_has_finobt(mp) &&
2483 	    (be32_to_cpu(agi->agi_free_level) < 1 ||
2484 	     be32_to_cpu(agi->agi_free_level) > M_IGEO(mp)->inobt_maxlevels))
2485 		return __this_address;
2486 
2487 	/*
2488 	 * during growfs operations, the perag is not fully initialised,
2489 	 * so we can't use it for any useful checking. growfs ensures we can't
2490 	 * use it by using uncached buffers that don't have the perag attached
2491 	 * so we can detect and avoid this problem.
2492 	 */
2493 	if (bp->b_pag && be32_to_cpu(agi->agi_seqno) != bp->b_pag->pag_agno)
2494 		return __this_address;
2495 
2496 	for (i = 0; i < XFS_AGI_UNLINKED_BUCKETS; i++) {
2497 		if (agi->agi_unlinked[i] == cpu_to_be32(NULLAGINO))
2498 			continue;
2499 		if (!xfs_verify_ino(mp, be32_to_cpu(agi->agi_unlinked[i])))
2500 			return __this_address;
2501 	}
2502 
2503 	return NULL;
2504 }
2505 
2506 static void
2507 xfs_agi_read_verify(
2508 	struct xfs_buf	*bp)
2509 {
2510 	struct xfs_mount *mp = bp->b_mount;
2511 	xfs_failaddr_t	fa;
2512 
2513 	if (xfs_has_crc(mp) &&
2514 	    !xfs_buf_verify_cksum(bp, XFS_AGI_CRC_OFF))
2515 		xfs_verifier_error(bp, -EFSBADCRC, __this_address);
2516 	else {
2517 		fa = xfs_agi_verify(bp);
2518 		if (XFS_TEST_ERROR(fa, mp, XFS_ERRTAG_IALLOC_READ_AGI))
2519 			xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2520 	}
2521 }
2522 
2523 static void
2524 xfs_agi_write_verify(
2525 	struct xfs_buf	*bp)
2526 {
2527 	struct xfs_mount	*mp = bp->b_mount;
2528 	struct xfs_buf_log_item	*bip = bp->b_log_item;
2529 	struct xfs_agi		*agi = bp->b_addr;
2530 	xfs_failaddr_t		fa;
2531 
2532 	fa = xfs_agi_verify(bp);
2533 	if (fa) {
2534 		xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2535 		return;
2536 	}
2537 
2538 	if (!xfs_has_crc(mp))
2539 		return;
2540 
2541 	if (bip)
2542 		agi->agi_lsn = cpu_to_be64(bip->bli_item.li_lsn);
2543 	xfs_buf_update_cksum(bp, XFS_AGI_CRC_OFF);
2544 }
2545 
2546 const struct xfs_buf_ops xfs_agi_buf_ops = {
2547 	.name = "xfs_agi",
2548 	.magic = { cpu_to_be32(XFS_AGI_MAGIC), cpu_to_be32(XFS_AGI_MAGIC) },
2549 	.verify_read = xfs_agi_read_verify,
2550 	.verify_write = xfs_agi_write_verify,
2551 	.verify_struct = xfs_agi_verify,
2552 };
2553 
2554 /*
2555  * Read in the allocation group header (inode allocation section)
2556  */
2557 int
2558 xfs_read_agi(
2559 	struct xfs_perag	*pag,
2560 	struct xfs_trans	*tp,
2561 	struct xfs_buf		**agibpp)
2562 {
2563 	struct xfs_mount	*mp = pag->pag_mount;
2564 	int			error;
2565 
2566 	trace_xfs_read_agi(pag->pag_mount, pag->pag_agno);
2567 
2568 	error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
2569 			XFS_AG_DADDR(mp, pag->pag_agno, XFS_AGI_DADDR(mp)),
2570 			XFS_FSS_TO_BB(mp, 1), 0, agibpp, &xfs_agi_buf_ops);
2571 	if (error)
2572 		return error;
2573 	if (tp)
2574 		xfs_trans_buf_set_type(tp, *agibpp, XFS_BLFT_AGI_BUF);
2575 
2576 	xfs_buf_set_ref(*agibpp, XFS_AGI_REF);
2577 	return 0;
2578 }
2579 
2580 /*
2581  * Read in the agi and initialise the per-ag data. If the caller supplies a
2582  * @agibpp, return the locked AGI buffer to them, otherwise release it.
2583  */
2584 int
2585 xfs_ialloc_read_agi(
2586 	struct xfs_perag	*pag,
2587 	struct xfs_trans	*tp,
2588 	struct xfs_buf		**agibpp)
2589 {
2590 	struct xfs_buf		*agibp;
2591 	struct xfs_agi		*agi;
2592 	int			error;
2593 
2594 	trace_xfs_ialloc_read_agi(pag->pag_mount, pag->pag_agno);
2595 
2596 	error = xfs_read_agi(pag, tp, &agibp);
2597 	if (error)
2598 		return error;
2599 
2600 	agi = agibp->b_addr;
2601 	if (!xfs_perag_initialised_agi(pag)) {
2602 		pag->pagi_freecount = be32_to_cpu(agi->agi_freecount);
2603 		pag->pagi_count = be32_to_cpu(agi->agi_count);
2604 		set_bit(XFS_AGSTATE_AGI_INIT, &pag->pag_opstate);
2605 	}
2606 
2607 	/*
2608 	 * It's possible for these to be out of sync if
2609 	 * we are in the middle of a forced shutdown.
2610 	 */
2611 	ASSERT(pag->pagi_freecount == be32_to_cpu(agi->agi_freecount) ||
2612 		xfs_is_shutdown(pag->pag_mount));
2613 	if (agibpp)
2614 		*agibpp = agibp;
2615 	else
2616 		xfs_trans_brelse(tp, agibp);
2617 	return 0;
2618 }
2619 
2620 /* Is there an inode record covering a given range of inode numbers? */
2621 int
2622 xfs_ialloc_has_inode_record(
2623 	struct xfs_btree_cur	*cur,
2624 	xfs_agino_t		low,
2625 	xfs_agino_t		high,
2626 	bool			*exists)
2627 {
2628 	struct xfs_inobt_rec_incore	irec;
2629 	xfs_agino_t		agino;
2630 	uint16_t		holemask;
2631 	int			has_record;
2632 	int			i;
2633 	int			error;
2634 
2635 	*exists = false;
2636 	error = xfs_inobt_lookup(cur, low, XFS_LOOKUP_LE, &has_record);
2637 	while (error == 0 && has_record) {
2638 		error = xfs_inobt_get_rec(cur, &irec, &has_record);
2639 		if (error || irec.ir_startino > high)
2640 			break;
2641 
2642 		agino = irec.ir_startino;
2643 		holemask = irec.ir_holemask;
2644 		for (i = 0; i < XFS_INOBT_HOLEMASK_BITS; holemask >>= 1,
2645 				i++, agino += XFS_INODES_PER_HOLEMASK_BIT) {
2646 			if (holemask & 1)
2647 				continue;
2648 			if (agino + XFS_INODES_PER_HOLEMASK_BIT > low &&
2649 					agino <= high) {
2650 				*exists = true;
2651 				return 0;
2652 			}
2653 		}
2654 
2655 		error = xfs_btree_increment(cur, 0, &has_record);
2656 	}
2657 	return error;
2658 }
2659 
2660 /* Is there an inode record covering a given extent? */
2661 int
2662 xfs_ialloc_has_inodes_at_extent(
2663 	struct xfs_btree_cur	*cur,
2664 	xfs_agblock_t		bno,
2665 	xfs_extlen_t		len,
2666 	bool			*exists)
2667 {
2668 	xfs_agino_t		low;
2669 	xfs_agino_t		high;
2670 
2671 	low = XFS_AGB_TO_AGINO(cur->bc_mp, bno);
2672 	high = XFS_AGB_TO_AGINO(cur->bc_mp, bno + len) - 1;
2673 
2674 	return xfs_ialloc_has_inode_record(cur, low, high, exists);
2675 }
2676 
2677 struct xfs_ialloc_count_inodes {
2678 	xfs_agino_t			count;
2679 	xfs_agino_t			freecount;
2680 };
2681 
2682 /* Record inode counts across all inobt records. */
2683 STATIC int
2684 xfs_ialloc_count_inodes_rec(
2685 	struct xfs_btree_cur		*cur,
2686 	const union xfs_btree_rec	*rec,
2687 	void				*priv)
2688 {
2689 	struct xfs_inobt_rec_incore	irec;
2690 	struct xfs_ialloc_count_inodes	*ci = priv;
2691 
2692 	xfs_inobt_btrec_to_irec(cur->bc_mp, rec, &irec);
2693 	ci->count += irec.ir_count;
2694 	ci->freecount += irec.ir_freecount;
2695 
2696 	return 0;
2697 }
2698 
2699 /* Count allocated and free inodes under an inobt. */
2700 int
2701 xfs_ialloc_count_inodes(
2702 	struct xfs_btree_cur		*cur,
2703 	xfs_agino_t			*count,
2704 	xfs_agino_t			*freecount)
2705 {
2706 	struct xfs_ialloc_count_inodes	ci = {0};
2707 	int				error;
2708 
2709 	ASSERT(cur->bc_btnum == XFS_BTNUM_INO);
2710 	error = xfs_btree_query_all(cur, xfs_ialloc_count_inodes_rec, &ci);
2711 	if (error)
2712 		return error;
2713 
2714 	*count = ci.count;
2715 	*freecount = ci.freecount;
2716 	return 0;
2717 }
2718 
2719 /*
2720  * Initialize inode-related geometry information.
2721  *
2722  * Compute the inode btree min and max levels and set maxicount.
2723  *
2724  * Set the inode cluster size.  This may still be overridden by the file
2725  * system block size if it is larger than the chosen cluster size.
2726  *
2727  * For v5 filesystems, scale the cluster size with the inode size to keep a
2728  * constant ratio of inode per cluster buffer, but only if mkfs has set the
2729  * inode alignment value appropriately for larger cluster sizes.
2730  *
2731  * Then compute the inode cluster alignment information.
2732  */
2733 void
2734 xfs_ialloc_setup_geometry(
2735 	struct xfs_mount	*mp)
2736 {
2737 	struct xfs_sb		*sbp = &mp->m_sb;
2738 	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
2739 	uint64_t		icount;
2740 	uint			inodes;
2741 
2742 	igeo->new_diflags2 = 0;
2743 	if (xfs_has_bigtime(mp))
2744 		igeo->new_diflags2 |= XFS_DIFLAG2_BIGTIME;
2745 	if (xfs_has_large_extent_counts(mp))
2746 		igeo->new_diflags2 |= XFS_DIFLAG2_NREXT64;
2747 
2748 	/* Compute inode btree geometry. */
2749 	igeo->agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
2750 	igeo->inobt_mxr[0] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, 1);
2751 	igeo->inobt_mxr[1] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, 0);
2752 	igeo->inobt_mnr[0] = igeo->inobt_mxr[0] / 2;
2753 	igeo->inobt_mnr[1] = igeo->inobt_mxr[1] / 2;
2754 
2755 	igeo->ialloc_inos = max_t(uint16_t, XFS_INODES_PER_CHUNK,
2756 			sbp->sb_inopblock);
2757 	igeo->ialloc_blks = igeo->ialloc_inos >> sbp->sb_inopblog;
2758 
2759 	if (sbp->sb_spino_align)
2760 		igeo->ialloc_min_blks = sbp->sb_spino_align;
2761 	else
2762 		igeo->ialloc_min_blks = igeo->ialloc_blks;
2763 
2764 	/* Compute and fill in value of m_ino_geo.inobt_maxlevels. */
2765 	inodes = (1LL << XFS_INO_AGINO_BITS(mp)) >> XFS_INODES_PER_CHUNK_LOG;
2766 	igeo->inobt_maxlevels = xfs_btree_compute_maxlevels(igeo->inobt_mnr,
2767 			inodes);
2768 	ASSERT(igeo->inobt_maxlevels <= xfs_iallocbt_maxlevels_ondisk());
2769 
2770 	/*
2771 	 * Set the maximum inode count for this filesystem, being careful not
2772 	 * to use obviously garbage sb_inopblog/sb_inopblock values.  Regular
2773 	 * users should never get here due to failing sb verification, but
2774 	 * certain users (xfs_db) need to be usable even with corrupt metadata.
2775 	 */
2776 	if (sbp->sb_imax_pct && igeo->ialloc_blks) {
2777 		/*
2778 		 * Make sure the maximum inode count is a multiple
2779 		 * of the units we allocate inodes in.
2780 		 */
2781 		icount = sbp->sb_dblocks * sbp->sb_imax_pct;
2782 		do_div(icount, 100);
2783 		do_div(icount, igeo->ialloc_blks);
2784 		igeo->maxicount = XFS_FSB_TO_INO(mp,
2785 				icount * igeo->ialloc_blks);
2786 	} else {
2787 		igeo->maxicount = 0;
2788 	}
2789 
2790 	/*
2791 	 * Compute the desired size of an inode cluster buffer size, which
2792 	 * starts at 8K and (on v5 filesystems) scales up with larger inode
2793 	 * sizes.
2794 	 *
2795 	 * Preserve the desired inode cluster size because the sparse inodes
2796 	 * feature uses that desired size (not the actual size) to compute the
2797 	 * sparse inode alignment.  The mount code validates this value, so we
2798 	 * cannot change the behavior.
2799 	 */
2800 	igeo->inode_cluster_size_raw = XFS_INODE_BIG_CLUSTER_SIZE;
2801 	if (xfs_has_v3inodes(mp)) {
2802 		int	new_size = igeo->inode_cluster_size_raw;
2803 
2804 		new_size *= mp->m_sb.sb_inodesize / XFS_DINODE_MIN_SIZE;
2805 		if (mp->m_sb.sb_inoalignmt >= XFS_B_TO_FSBT(mp, new_size))
2806 			igeo->inode_cluster_size_raw = new_size;
2807 	}
2808 
2809 	/* Calculate inode cluster ratios. */
2810 	if (igeo->inode_cluster_size_raw > mp->m_sb.sb_blocksize)
2811 		igeo->blocks_per_cluster = XFS_B_TO_FSBT(mp,
2812 				igeo->inode_cluster_size_raw);
2813 	else
2814 		igeo->blocks_per_cluster = 1;
2815 	igeo->inode_cluster_size = XFS_FSB_TO_B(mp, igeo->blocks_per_cluster);
2816 	igeo->inodes_per_cluster = XFS_FSB_TO_INO(mp, igeo->blocks_per_cluster);
2817 
2818 	/* Calculate inode cluster alignment. */
2819 	if (xfs_has_align(mp) &&
2820 	    mp->m_sb.sb_inoalignmt >= igeo->blocks_per_cluster)
2821 		igeo->cluster_align = mp->m_sb.sb_inoalignmt;
2822 	else
2823 		igeo->cluster_align = 1;
2824 	igeo->inoalign_mask = igeo->cluster_align - 1;
2825 	igeo->cluster_align_inodes = XFS_FSB_TO_INO(mp, igeo->cluster_align);
2826 
2827 	/*
2828 	 * If we are using stripe alignment, check whether
2829 	 * the stripe unit is a multiple of the inode alignment
2830 	 */
2831 	if (mp->m_dalign && igeo->inoalign_mask &&
2832 	    !(mp->m_dalign & igeo->inoalign_mask))
2833 		igeo->ialloc_align = mp->m_dalign;
2834 	else
2835 		igeo->ialloc_align = 0;
2836 }
2837 
2838 /* Compute the location of the root directory inode that is laid out by mkfs. */
2839 xfs_ino_t
2840 xfs_ialloc_calc_rootino(
2841 	struct xfs_mount	*mp,
2842 	int			sunit)
2843 {
2844 	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
2845 	xfs_agblock_t		first_bno;
2846 
2847 	/*
2848 	 * Pre-calculate the geometry of AG 0.  We know what it looks like
2849 	 * because libxfs knows how to create allocation groups now.
2850 	 *
2851 	 * first_bno is the first block in which mkfs could possibly have
2852 	 * allocated the root directory inode, once we factor in the metadata
2853 	 * that mkfs formats before it.  Namely, the four AG headers...
2854 	 */
2855 	first_bno = howmany(4 * mp->m_sb.sb_sectsize, mp->m_sb.sb_blocksize);
2856 
2857 	/* ...the two free space btree roots... */
2858 	first_bno += 2;
2859 
2860 	/* ...the inode btree root... */
2861 	first_bno += 1;
2862 
2863 	/* ...the initial AGFL... */
2864 	first_bno += xfs_alloc_min_freelist(mp, NULL);
2865 
2866 	/* ...the free inode btree root... */
2867 	if (xfs_has_finobt(mp))
2868 		first_bno++;
2869 
2870 	/* ...the reverse mapping btree root... */
2871 	if (xfs_has_rmapbt(mp))
2872 		first_bno++;
2873 
2874 	/* ...the reference count btree... */
2875 	if (xfs_has_reflink(mp))
2876 		first_bno++;
2877 
2878 	/*
2879 	 * ...and the log, if it is allocated in the first allocation group.
2880 	 *
2881 	 * This can happen with filesystems that only have a single
2882 	 * allocation group, or very odd geometries created by old mkfs
2883 	 * versions on very small filesystems.
2884 	 */
2885 	if (xfs_ag_contains_log(mp, 0))
2886 		 first_bno += mp->m_sb.sb_logblocks;
2887 
2888 	/*
2889 	 * Now round first_bno up to whatever allocation alignment is given
2890 	 * by the filesystem or was passed in.
2891 	 */
2892 	if (xfs_has_dalign(mp) && igeo->ialloc_align > 0)
2893 		first_bno = roundup(first_bno, sunit);
2894 	else if (xfs_has_align(mp) &&
2895 			mp->m_sb.sb_inoalignmt > 1)
2896 		first_bno = roundup(first_bno, mp->m_sb.sb_inoalignmt);
2897 
2898 	return XFS_AGINO_TO_INO(mp, 0, XFS_AGB_TO_AGINO(mp, first_bno));
2899 }
2900 
2901 /*
2902  * Ensure there are not sparse inode clusters that cross the new EOAG.
2903  *
2904  * This is a no-op for non-spinode filesystems since clusters are always fully
2905  * allocated and checking the bnobt suffices.  However, a spinode filesystem
2906  * could have a record where the upper inodes are free blocks.  If those blocks
2907  * were removed from the filesystem, the inode record would extend beyond EOAG,
2908  * which will be flagged as corruption.
2909  */
2910 int
2911 xfs_ialloc_check_shrink(
2912 	struct xfs_perag	*pag,
2913 	struct xfs_trans	*tp,
2914 	struct xfs_buf		*agibp,
2915 	xfs_agblock_t		new_length)
2916 {
2917 	struct xfs_inobt_rec_incore rec;
2918 	struct xfs_btree_cur	*cur;
2919 	xfs_agino_t		agino;
2920 	int			has;
2921 	int			error;
2922 
2923 	if (!xfs_has_sparseinodes(pag->pag_mount))
2924 		return 0;
2925 
2926 	cur = xfs_inobt_init_cursor(pag, tp, agibp, XFS_BTNUM_INO);
2927 
2928 	/* Look up the inobt record that would correspond to the new EOFS. */
2929 	agino = XFS_AGB_TO_AGINO(pag->pag_mount, new_length);
2930 	error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &has);
2931 	if (error || !has)
2932 		goto out;
2933 
2934 	error = xfs_inobt_get_rec(cur, &rec, &has);
2935 	if (error)
2936 		goto out;
2937 
2938 	if (!has) {
2939 		error = -EFSCORRUPTED;
2940 		goto out;
2941 	}
2942 
2943 	/* If the record covers inodes that would be beyond EOFS, bail out. */
2944 	if (rec.ir_startino + XFS_INODES_PER_CHUNK > agino) {
2945 		error = -ENOSPC;
2946 		goto out;
2947 	}
2948 out:
2949 	xfs_btree_del_cursor(cur, error);
2950 	return error;
2951 }
2952