xfs_ialloc.c revision ebd9027d
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 */
34int					/* error */
35xfs_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 */
53STATIC int				/* error */
54xfs_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. */
74void
75xfs_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 */
101int
102xfs_inobt_get_rec(
103	struct xfs_btree_cur		*cur,
104	struct xfs_inobt_rec_incore	*irec,
105	int				*stat)
106{
107	struct xfs_mount		*mp = cur->bc_mp;
108	xfs_agnumber_t			agno = cur->bc_ag.pag->pag_agno;
109	union xfs_btree_rec		*rec;
110	int				error;
111	uint64_t			realfree;
112
113	error = xfs_btree_get_rec(cur, &rec, stat);
114	if (error || *stat == 0)
115		return error;
116
117	xfs_inobt_btrec_to_irec(mp, rec, irec);
118
119	if (!xfs_verify_agino(mp, agno, irec->ir_startino))
120		goto out_bad_rec;
121	if (irec->ir_count < XFS_INODES_PER_HOLEMASK_BIT ||
122	    irec->ir_count > XFS_INODES_PER_CHUNK)
123		goto out_bad_rec;
124	if (irec->ir_freecount > XFS_INODES_PER_CHUNK)
125		goto out_bad_rec;
126
127	/* if there are no holes, return the first available offset */
128	if (!xfs_inobt_issparse(irec->ir_holemask))
129		realfree = irec->ir_free;
130	else
131		realfree = irec->ir_free & xfs_inobt_irec_to_allocmask(irec);
132	if (hweight64(realfree) != irec->ir_freecount)
133		goto out_bad_rec;
134
135	return 0;
136
137out_bad_rec:
138	xfs_warn(mp,
139		"%s Inode BTree record corruption in AG %d detected!",
140		cur->bc_btnum == XFS_BTNUM_INO ? "Used" : "Free", agno);
141	xfs_warn(mp,
142"start inode 0x%x, count 0x%x, free 0x%x freemask 0x%llx, holemask 0x%x",
143		irec->ir_startino, irec->ir_count, irec->ir_freecount,
144		irec->ir_free, irec->ir_holemask);
145	return -EFSCORRUPTED;
146}
147
148/*
149 * Insert a single inobt record. Cursor must already point to desired location.
150 */
151int
152xfs_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 */
170STATIC int
171xfs_inobt_insert(
172	struct xfs_mount	*mp,
173	struct xfs_trans	*tp,
174	struct xfs_buf		*agbp,
175	struct xfs_perag	*pag,
176	xfs_agino_t		newino,
177	xfs_agino_t		newlen,
178	xfs_btnum_t		btnum)
179{
180	struct xfs_btree_cur	*cur;
181	xfs_agino_t		thisino;
182	int			i;
183	int			error;
184
185	cur = xfs_inobt_init_cursor(mp, tp, agbp, pag, btnum);
186
187	for (thisino = newino;
188	     thisino < newino + newlen;
189	     thisino += XFS_INODES_PER_CHUNK) {
190		error = xfs_inobt_lookup(cur, thisino, XFS_LOOKUP_EQ, &i);
191		if (error) {
192			xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
193			return error;
194		}
195		ASSERT(i == 0);
196
197		error = xfs_inobt_insert_rec(cur, XFS_INOBT_HOLEMASK_FULL,
198					     XFS_INODES_PER_CHUNK,
199					     XFS_INODES_PER_CHUNK,
200					     XFS_INOBT_ALL_FREE, &i);
201		if (error) {
202			xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
203			return error;
204		}
205		ASSERT(i == 1);
206	}
207
208	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
209
210	return 0;
211}
212
213/*
214 * Verify that the number of free inodes in the AGI is correct.
215 */
216#ifdef DEBUG
217static int
218xfs_check_agi_freecount(
219	struct xfs_btree_cur	*cur)
220{
221	if (cur->bc_nlevels == 1) {
222		xfs_inobt_rec_incore_t rec;
223		int		freecount = 0;
224		int		error;
225		int		i;
226
227		error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
228		if (error)
229			return error;
230
231		do {
232			error = xfs_inobt_get_rec(cur, &rec, &i);
233			if (error)
234				return error;
235
236			if (i) {
237				freecount += rec.ir_freecount;
238				error = xfs_btree_increment(cur, 0, &i);
239				if (error)
240					return error;
241			}
242		} while (i == 1);
243
244		if (!xfs_is_shutdown(cur->bc_mp))
245			ASSERT(freecount == cur->bc_ag.pag->pagi_freecount);
246	}
247	return 0;
248}
249#else
250#define xfs_check_agi_freecount(cur)	0
251#endif
252
253/*
254 * Initialise a new set of inodes. When called without a transaction context
255 * (e.g. from recovery) we initiate a delayed write of the inode buffers rather
256 * than logging them (which in a transaction context puts them into the AIL
257 * for writeback rather than the xfsbufd queue).
258 */
259int
260xfs_ialloc_inode_init(
261	struct xfs_mount	*mp,
262	struct xfs_trans	*tp,
263	struct list_head	*buffer_list,
264	int			icount,
265	xfs_agnumber_t		agno,
266	xfs_agblock_t		agbno,
267	xfs_agblock_t		length,
268	unsigned int		gen)
269{
270	struct xfs_buf		*fbuf;
271	struct xfs_dinode	*free;
272	int			nbufs;
273	int			version;
274	int			i, j;
275	xfs_daddr_t		d;
276	xfs_ino_t		ino = 0;
277	int			error;
278
279	/*
280	 * Loop over the new block(s), filling in the inodes.  For small block
281	 * sizes, manipulate the inodes in buffers  which are multiples of the
282	 * blocks size.
283	 */
284	nbufs = length / M_IGEO(mp)->blocks_per_cluster;
285
286	/*
287	 * Figure out what version number to use in the inodes we create.  If
288	 * the superblock version has caught up to the one that supports the new
289	 * inode format, then use the new inode version.  Otherwise use the old
290	 * version so that old kernels will continue to be able to use the file
291	 * system.
292	 *
293	 * For v3 inodes, we also need to write the inode number into the inode,
294	 * so calculate the first inode number of the chunk here as
295	 * XFS_AGB_TO_AGINO() only works within a filesystem block, not
296	 * across multiple filesystem blocks (such as a cluster) and so cannot
297	 * be used in the cluster buffer loop below.
298	 *
299	 * Further, because we are writing the inode directly into the buffer
300	 * and calculating a CRC on the entire inode, we have ot log the entire
301	 * inode so that the entire range the CRC covers is present in the log.
302	 * That means for v3 inode we log the entire buffer rather than just the
303	 * inode cores.
304	 */
305	if (xfs_has_v3inodes(mp)) {
306		version = 3;
307		ino = XFS_AGINO_TO_INO(mp, agno, XFS_AGB_TO_AGINO(mp, agbno));
308
309		/*
310		 * log the initialisation that is about to take place as an
311		 * logical operation. This means the transaction does not
312		 * need to log the physical changes to the inode buffers as log
313		 * recovery will know what initialisation is actually needed.
314		 * Hence we only need to log the buffers as "ordered" buffers so
315		 * they track in the AIL as if they were physically logged.
316		 */
317		if (tp)
318			xfs_icreate_log(tp, agno, agbno, icount,
319					mp->m_sb.sb_inodesize, length, gen);
320	} else
321		version = 2;
322
323	for (j = 0; j < nbufs; j++) {
324		/*
325		 * Get the block.
326		 */
327		d = XFS_AGB_TO_DADDR(mp, agno, agbno +
328				(j * M_IGEO(mp)->blocks_per_cluster));
329		error = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
330				mp->m_bsize * M_IGEO(mp)->blocks_per_cluster,
331				XBF_UNMAPPED, &fbuf);
332		if (error)
333			return error;
334
335		/* Initialize the inode buffers and log them appropriately. */
336		fbuf->b_ops = &xfs_inode_buf_ops;
337		xfs_buf_zero(fbuf, 0, BBTOB(fbuf->b_length));
338		for (i = 0; i < M_IGEO(mp)->inodes_per_cluster; i++) {
339			int	ioffset = i << mp->m_sb.sb_inodelog;
340			uint	isize = XFS_DINODE_SIZE(&mp->m_sb);
341
342			free = xfs_make_iptr(mp, fbuf, i);
343			free->di_magic = cpu_to_be16(XFS_DINODE_MAGIC);
344			free->di_version = version;
345			free->di_gen = cpu_to_be32(gen);
346			free->di_next_unlinked = cpu_to_be32(NULLAGINO);
347
348			if (version == 3) {
349				free->di_ino = cpu_to_be64(ino);
350				ino++;
351				uuid_copy(&free->di_uuid,
352					  &mp->m_sb.sb_meta_uuid);
353				xfs_dinode_calc_crc(mp, free);
354			} else if (tp) {
355				/* just log the inode core */
356				xfs_trans_log_buf(tp, fbuf, ioffset,
357						  ioffset + isize - 1);
358			}
359		}
360
361		if (tp) {
362			/*
363			 * Mark the buffer as an inode allocation buffer so it
364			 * sticks in AIL at the point of this allocation
365			 * transaction. This ensures the they are on disk before
366			 * the tail of the log can be moved past this
367			 * transaction (i.e. by preventing relogging from moving
368			 * it forward in the log).
369			 */
370			xfs_trans_inode_alloc_buf(tp, fbuf);
371			if (version == 3) {
372				/*
373				 * Mark the buffer as ordered so that they are
374				 * not physically logged in the transaction but
375				 * still tracked in the AIL as part of the
376				 * transaction and pin the log appropriately.
377				 */
378				xfs_trans_ordered_buf(tp, fbuf);
379			}
380		} else {
381			fbuf->b_flags |= XBF_DONE;
382			xfs_buf_delwri_queue(fbuf, buffer_list);
383			xfs_buf_relse(fbuf);
384		}
385	}
386	return 0;
387}
388
389/*
390 * Align startino and allocmask for a recently allocated sparse chunk such that
391 * they are fit for insertion (or merge) into the on-disk inode btrees.
392 *
393 * Background:
394 *
395 * When enabled, sparse inode support increases the inode alignment from cluster
396 * size to inode chunk size. This means that the minimum range between two
397 * non-adjacent inode records in the inobt is large enough for a full inode
398 * record. This allows for cluster sized, cluster aligned block allocation
399 * without need to worry about whether the resulting inode record overlaps with
400 * another record in the tree. Without this basic rule, we would have to deal
401 * with the consequences of overlap by potentially undoing recent allocations in
402 * the inode allocation codepath.
403 *
404 * Because of this alignment rule (which is enforced on mount), there are two
405 * inobt possibilities for newly allocated sparse chunks. One is that the
406 * aligned inode record for the chunk covers a range of inodes not already
407 * covered in the inobt (i.e., it is safe to insert a new sparse record). The
408 * other is that a record already exists at the aligned startino that considers
409 * the newly allocated range as sparse. In the latter case, record content is
410 * merged in hope that sparse inode chunks fill to full chunks over time.
411 */
412STATIC void
413xfs_align_sparse_ino(
414	struct xfs_mount		*mp,
415	xfs_agino_t			*startino,
416	uint16_t			*allocmask)
417{
418	xfs_agblock_t			agbno;
419	xfs_agblock_t			mod;
420	int				offset;
421
422	agbno = XFS_AGINO_TO_AGBNO(mp, *startino);
423	mod = agbno % mp->m_sb.sb_inoalignmt;
424	if (!mod)
425		return;
426
427	/* calculate the inode offset and align startino */
428	offset = XFS_AGB_TO_AGINO(mp, mod);
429	*startino -= offset;
430
431	/*
432	 * Since startino has been aligned down, left shift allocmask such that
433	 * it continues to represent the same physical inodes relative to the
434	 * new startino.
435	 */
436	*allocmask <<= offset / XFS_INODES_PER_HOLEMASK_BIT;
437}
438
439/*
440 * Determine whether the source inode record can merge into the target. Both
441 * records must be sparse, the inode ranges must match and there must be no
442 * allocation overlap between the records.
443 */
444STATIC bool
445__xfs_inobt_can_merge(
446	struct xfs_inobt_rec_incore	*trec,	/* tgt record */
447	struct xfs_inobt_rec_incore	*srec)	/* src record */
448{
449	uint64_t			talloc;
450	uint64_t			salloc;
451
452	/* records must cover the same inode range */
453	if (trec->ir_startino != srec->ir_startino)
454		return false;
455
456	/* both records must be sparse */
457	if (!xfs_inobt_issparse(trec->ir_holemask) ||
458	    !xfs_inobt_issparse(srec->ir_holemask))
459		return false;
460
461	/* both records must track some inodes */
462	if (!trec->ir_count || !srec->ir_count)
463		return false;
464
465	/* can't exceed capacity of a full record */
466	if (trec->ir_count + srec->ir_count > XFS_INODES_PER_CHUNK)
467		return false;
468
469	/* verify there is no allocation overlap */
470	talloc = xfs_inobt_irec_to_allocmask(trec);
471	salloc = xfs_inobt_irec_to_allocmask(srec);
472	if (talloc & salloc)
473		return false;
474
475	return true;
476}
477
478/*
479 * Merge the source inode record into the target. The caller must call
480 * __xfs_inobt_can_merge() to ensure the merge is valid.
481 */
482STATIC void
483__xfs_inobt_rec_merge(
484	struct xfs_inobt_rec_incore	*trec,	/* target */
485	struct xfs_inobt_rec_incore	*srec)	/* src */
486{
487	ASSERT(trec->ir_startino == srec->ir_startino);
488
489	/* combine the counts */
490	trec->ir_count += srec->ir_count;
491	trec->ir_freecount += srec->ir_freecount;
492
493	/*
494	 * Merge the holemask and free mask. For both fields, 0 bits refer to
495	 * allocated inodes. We combine the allocated ranges with bitwise AND.
496	 */
497	trec->ir_holemask &= srec->ir_holemask;
498	trec->ir_free &= srec->ir_free;
499}
500
501/*
502 * Insert a new sparse inode chunk into the associated inode btree. The inode
503 * record for the sparse chunk is pre-aligned to a startino that should match
504 * any pre-existing sparse inode record in the tree. This allows sparse chunks
505 * to fill over time.
506 *
507 * This function supports two modes of handling preexisting records depending on
508 * the merge flag. If merge is true, the provided record is merged with the
509 * existing record and updated in place. The merged record is returned in nrec.
510 * If merge is false, an existing record is replaced with the provided record.
511 * If no preexisting record exists, the provided record is always inserted.
512 *
513 * It is considered corruption if a merge is requested and not possible. Given
514 * the sparse inode alignment constraints, this should never happen.
515 */
516STATIC int
517xfs_inobt_insert_sprec(
518	struct xfs_mount		*mp,
519	struct xfs_trans		*tp,
520	struct xfs_buf			*agbp,
521	struct xfs_perag		*pag,
522	int				btnum,
523	struct xfs_inobt_rec_incore	*nrec,	/* in/out: new/merged rec. */
524	bool				merge)	/* merge or replace */
525{
526	struct xfs_btree_cur		*cur;
527	int				error;
528	int				i;
529	struct xfs_inobt_rec_incore	rec;
530
531	cur = xfs_inobt_init_cursor(mp, tp, agbp, pag, btnum);
532
533	/* the new record is pre-aligned so we know where to look */
534	error = xfs_inobt_lookup(cur, nrec->ir_startino, XFS_LOOKUP_EQ, &i);
535	if (error)
536		goto error;
537	/* if nothing there, insert a new record and return */
538	if (i == 0) {
539		error = xfs_inobt_insert_rec(cur, nrec->ir_holemask,
540					     nrec->ir_count, nrec->ir_freecount,
541					     nrec->ir_free, &i);
542		if (error)
543			goto error;
544		if (XFS_IS_CORRUPT(mp, i != 1)) {
545			error = -EFSCORRUPTED;
546			goto error;
547		}
548
549		goto out;
550	}
551
552	/*
553	 * A record exists at this startino. Merge or replace the record
554	 * depending on what we've been asked to do.
555	 */
556	if (merge) {
557		error = xfs_inobt_get_rec(cur, &rec, &i);
558		if (error)
559			goto error;
560		if (XFS_IS_CORRUPT(mp, i != 1)) {
561			error = -EFSCORRUPTED;
562			goto error;
563		}
564		if (XFS_IS_CORRUPT(mp, rec.ir_startino != nrec->ir_startino)) {
565			error = -EFSCORRUPTED;
566			goto error;
567		}
568
569		/*
570		 * This should never fail. If we have coexisting records that
571		 * cannot merge, something is seriously wrong.
572		 */
573		if (XFS_IS_CORRUPT(mp, !__xfs_inobt_can_merge(nrec, &rec))) {
574			error = -EFSCORRUPTED;
575			goto error;
576		}
577
578		trace_xfs_irec_merge_pre(mp, pag->pag_agno, rec.ir_startino,
579					 rec.ir_holemask, nrec->ir_startino,
580					 nrec->ir_holemask);
581
582		/* merge to nrec to output the updated record */
583		__xfs_inobt_rec_merge(nrec, &rec);
584
585		trace_xfs_irec_merge_post(mp, pag->pag_agno, nrec->ir_startino,
586					  nrec->ir_holemask);
587
588		error = xfs_inobt_rec_check_count(mp, nrec);
589		if (error)
590			goto error;
591	}
592
593	error = xfs_inobt_update(cur, nrec);
594	if (error)
595		goto error;
596
597out:
598	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
599	return 0;
600error:
601	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
602	return error;
603}
604
605/*
606 * Allocate new inodes in the allocation group specified by agbp.  Returns 0 if
607 * inodes were allocated in this AG; -EAGAIN if there was no space in this AG so
608 * the caller knows it can try another AG, a hard -ENOSPC when over the maximum
609 * inode count threshold, or the usual negative error code for other errors.
610 */
611STATIC int
612xfs_ialloc_ag_alloc(
613	struct xfs_trans	*tp,
614	struct xfs_buf		*agbp,
615	struct xfs_perag	*pag)
616{
617	struct xfs_agi		*agi;
618	struct xfs_alloc_arg	args;
619	int			error;
620	xfs_agino_t		newino;		/* new first inode's number */
621	xfs_agino_t		newlen;		/* new number of inodes */
622	int			isaligned = 0;	/* inode allocation at stripe */
623						/* unit boundary */
624	/* init. to full chunk */
625	struct xfs_inobt_rec_incore rec;
626	struct xfs_ino_geometry	*igeo = M_IGEO(tp->t_mountp);
627	uint16_t		allocmask = (uint16_t) -1;
628	int			do_sparse = 0;
629
630	memset(&args, 0, sizeof(args));
631	args.tp = tp;
632	args.mp = tp->t_mountp;
633	args.fsbno = NULLFSBLOCK;
634	args.oinfo = XFS_RMAP_OINFO_INODES;
635
636#ifdef DEBUG
637	/* randomly do sparse inode allocations */
638	if (xfs_has_sparseinodes(tp->t_mountp) &&
639	    igeo->ialloc_min_blks < igeo->ialloc_blks)
640		do_sparse = prandom_u32() & 1;
641#endif
642
643	/*
644	 * Locking will ensure that we don't have two callers in here
645	 * at one time.
646	 */
647	newlen = igeo->ialloc_inos;
648	if (igeo->maxicount &&
649	    percpu_counter_read_positive(&args.mp->m_icount) + newlen >
650							igeo->maxicount)
651		return -ENOSPC;
652	args.minlen = args.maxlen = igeo->ialloc_blks;
653	/*
654	 * First try to allocate inodes contiguous with the last-allocated
655	 * chunk of inodes.  If the filesystem is striped, this will fill
656	 * an entire stripe unit with inodes.
657	 */
658	agi = agbp->b_addr;
659	newino = be32_to_cpu(agi->agi_newino);
660	args.agbno = XFS_AGINO_TO_AGBNO(args.mp, newino) +
661		     igeo->ialloc_blks;
662	if (do_sparse)
663		goto sparse_alloc;
664	if (likely(newino != NULLAGINO &&
665		  (args.agbno < be32_to_cpu(agi->agi_length)))) {
666		args.fsbno = XFS_AGB_TO_FSB(args.mp, pag->pag_agno, args.agbno);
667		args.type = XFS_ALLOCTYPE_THIS_BNO;
668		args.prod = 1;
669
670		/*
671		 * We need to take into account alignment here to ensure that
672		 * we don't modify the free list if we fail to have an exact
673		 * block. If we don't have an exact match, and every oher
674		 * attempt allocation attempt fails, we'll end up cancelling
675		 * a dirty transaction and shutting down.
676		 *
677		 * For an exact allocation, alignment must be 1,
678		 * however we need to take cluster alignment into account when
679		 * fixing up the freelist. Use the minalignslop field to
680		 * indicate that extra blocks might be required for alignment,
681		 * but not to use them in the actual exact allocation.
682		 */
683		args.alignment = 1;
684		args.minalignslop = igeo->cluster_align - 1;
685
686		/* Allow space for the inode btree to split. */
687		args.minleft = igeo->inobt_maxlevels;
688		if ((error = xfs_alloc_vextent(&args)))
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		 * Need to figure out where to allocate the inode blocks.
722		 * Ideally they should be spaced out through the a.g.
723		 * For now, just allocate blocks up front.
724		 */
725		args.agbno = be32_to_cpu(agi->agi_root);
726		args.fsbno = XFS_AGB_TO_FSB(args.mp, pag->pag_agno, args.agbno);
727		/*
728		 * Allocate a fixed-size extent of inodes.
729		 */
730		args.type = XFS_ALLOCTYPE_NEAR_BNO;
731		args.prod = 1;
732		/*
733		 * Allow space for the inode btree to split.
734		 */
735		args.minleft = igeo->inobt_maxlevels;
736		if ((error = xfs_alloc_vextent(&args)))
737			return error;
738	}
739
740	/*
741	 * If stripe alignment is turned on, then try again with cluster
742	 * alignment.
743	 */
744	if (isaligned && args.fsbno == NULLFSBLOCK) {
745		args.type = XFS_ALLOCTYPE_NEAR_BNO;
746		args.agbno = be32_to_cpu(agi->agi_root);
747		args.fsbno = XFS_AGB_TO_FSB(args.mp, pag->pag_agno, args.agbno);
748		args.alignment = igeo->cluster_align;
749		if ((error = xfs_alloc_vextent(&args)))
750			return error;
751	}
752
753	/*
754	 * Finally, try a sparse allocation if the filesystem supports it and
755	 * the sparse allocation length is smaller than a full chunk.
756	 */
757	if (xfs_has_sparseinodes(args.mp) &&
758	    igeo->ialloc_min_blks < igeo->ialloc_blks &&
759	    args.fsbno == NULLFSBLOCK) {
760sparse_alloc:
761		args.type = XFS_ALLOCTYPE_NEAR_BNO;
762		args.agbno = be32_to_cpu(agi->agi_root);
763		args.fsbno = XFS_AGB_TO_FSB(args.mp, pag->pag_agno, args.agbno);
764		args.alignment = args.mp->m_sb.sb_spino_align;
765		args.prod = 1;
766
767		args.minlen = igeo->ialloc_min_blks;
768		args.maxlen = args.minlen;
769
770		/*
771		 * The inode record will be aligned to full chunk size. We must
772		 * prevent sparse allocation from AG boundaries that result in
773		 * invalid inode records, such as records that start at agbno 0
774		 * or extend beyond the AG.
775		 *
776		 * Set min agbno to the first aligned, non-zero agbno and max to
777		 * the last aligned agbno that is at least one full chunk from
778		 * the end of the AG.
779		 */
780		args.min_agbno = args.mp->m_sb.sb_inoalignmt;
781		args.max_agbno = round_down(args.mp->m_sb.sb_agblocks,
782					    args.mp->m_sb.sb_inoalignmt) -
783				 igeo->ialloc_blks;
784
785		error = xfs_alloc_vextent(&args);
786		if (error)
787			return error;
788
789		newlen = XFS_AGB_TO_AGINO(args.mp, args.len);
790		ASSERT(newlen <= XFS_INODES_PER_CHUNK);
791		allocmask = (1 << (newlen / XFS_INODES_PER_HOLEMASK_BIT)) - 1;
792	}
793
794	if (args.fsbno == NULLFSBLOCK)
795		return -EAGAIN;
796
797	ASSERT(args.len == args.minlen);
798
799	/*
800	 * Stamp and write the inode buffers.
801	 *
802	 * Seed the new inode cluster with a random generation number. This
803	 * prevents short-term reuse of generation numbers if a chunk is
804	 * freed and then immediately reallocated. We use random numbers
805	 * rather than a linear progression to prevent the next generation
806	 * number from being easily guessable.
807	 */
808	error = xfs_ialloc_inode_init(args.mp, tp, NULL, newlen, pag->pag_agno,
809			args.agbno, args.len, prandom_u32());
810
811	if (error)
812		return error;
813	/*
814	 * Convert the results.
815	 */
816	newino = XFS_AGB_TO_AGINO(args.mp, args.agbno);
817
818	if (xfs_inobt_issparse(~allocmask)) {
819		/*
820		 * We've allocated a sparse chunk. Align the startino and mask.
821		 */
822		xfs_align_sparse_ino(args.mp, &newino, &allocmask);
823
824		rec.ir_startino = newino;
825		rec.ir_holemask = ~allocmask;
826		rec.ir_count = newlen;
827		rec.ir_freecount = newlen;
828		rec.ir_free = XFS_INOBT_ALL_FREE;
829
830		/*
831		 * Insert the sparse record into the inobt and allow for a merge
832		 * if necessary. If a merge does occur, rec is updated to the
833		 * merged record.
834		 */
835		error = xfs_inobt_insert_sprec(args.mp, tp, agbp, pag,
836				XFS_BTNUM_INO, &rec, true);
837		if (error == -EFSCORRUPTED) {
838			xfs_alert(args.mp,
839	"invalid sparse inode record: ino 0x%llx holemask 0x%x count %u",
840				  XFS_AGINO_TO_INO(args.mp, pag->pag_agno,
841						   rec.ir_startino),
842				  rec.ir_holemask, rec.ir_count);
843			xfs_force_shutdown(args.mp, SHUTDOWN_CORRUPT_INCORE);
844		}
845		if (error)
846			return error;
847
848		/*
849		 * We can't merge the part we've just allocated as for the inobt
850		 * due to finobt semantics. The original record may or may not
851		 * exist independent of whether physical inodes exist in this
852		 * sparse chunk.
853		 *
854		 * We must update the finobt record based on the inobt record.
855		 * rec contains the fully merged and up to date inobt record
856		 * from the previous call. Set merge false to replace any
857		 * existing record with this one.
858		 */
859		if (xfs_has_finobt(args.mp)) {
860			error = xfs_inobt_insert_sprec(args.mp, tp, agbp, pag,
861				       XFS_BTNUM_FINO, &rec, false);
862			if (error)
863				return error;
864		}
865	} else {
866		/* full chunk - insert new records to both btrees */
867		error = xfs_inobt_insert(args.mp, tp, agbp, pag, newino, newlen,
868					 XFS_BTNUM_INO);
869		if (error)
870			return error;
871
872		if (xfs_has_finobt(args.mp)) {
873			error = xfs_inobt_insert(args.mp, tp, agbp, pag, newino,
874						 newlen, XFS_BTNUM_FINO);
875			if (error)
876				return error;
877		}
878	}
879
880	/*
881	 * Update AGI counts and newino.
882	 */
883	be32_add_cpu(&agi->agi_count, newlen);
884	be32_add_cpu(&agi->agi_freecount, newlen);
885	pag->pagi_freecount += newlen;
886	pag->pagi_count += newlen;
887	agi->agi_newino = cpu_to_be32(newino);
888
889	/*
890	 * Log allocation group header fields
891	 */
892	xfs_ialloc_log_agi(tp, agbp,
893		XFS_AGI_COUNT | XFS_AGI_FREECOUNT | XFS_AGI_NEWINO);
894	/*
895	 * Modify/log superblock values for inode count and inode free count.
896	 */
897	xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, (long)newlen);
898	xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, (long)newlen);
899	return 0;
900}
901
902/*
903 * Try to retrieve the next record to the left/right from the current one.
904 */
905STATIC int
906xfs_ialloc_next_rec(
907	struct xfs_btree_cur	*cur,
908	xfs_inobt_rec_incore_t	*rec,
909	int			*done,
910	int			left)
911{
912	int                     error;
913	int			i;
914
915	if (left)
916		error = xfs_btree_decrement(cur, 0, &i);
917	else
918		error = xfs_btree_increment(cur, 0, &i);
919
920	if (error)
921		return error;
922	*done = !i;
923	if (i) {
924		error = xfs_inobt_get_rec(cur, rec, &i);
925		if (error)
926			return error;
927		if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
928			return -EFSCORRUPTED;
929	}
930
931	return 0;
932}
933
934STATIC int
935xfs_ialloc_get_rec(
936	struct xfs_btree_cur	*cur,
937	xfs_agino_t		agino,
938	xfs_inobt_rec_incore_t	*rec,
939	int			*done)
940{
941	int                     error;
942	int			i;
943
944	error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_EQ, &i);
945	if (error)
946		return error;
947	*done = !i;
948	if (i) {
949		error = xfs_inobt_get_rec(cur, rec, &i);
950		if (error)
951			return error;
952		if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
953			return -EFSCORRUPTED;
954	}
955
956	return 0;
957}
958
959/*
960 * Return the offset of the first free inode in the record. If the inode chunk
961 * is sparsely allocated, we convert the record holemask to inode granularity
962 * and mask off the unallocated regions from the inode free mask.
963 */
964STATIC int
965xfs_inobt_first_free_inode(
966	struct xfs_inobt_rec_incore	*rec)
967{
968	xfs_inofree_t			realfree;
969
970	/* if there are no holes, return the first available offset */
971	if (!xfs_inobt_issparse(rec->ir_holemask))
972		return xfs_lowbit64(rec->ir_free);
973
974	realfree = xfs_inobt_irec_to_allocmask(rec);
975	realfree &= rec->ir_free;
976
977	return xfs_lowbit64(realfree);
978}
979
980/*
981 * Allocate an inode using the inobt-only algorithm.
982 */
983STATIC int
984xfs_dialloc_ag_inobt(
985	struct xfs_trans	*tp,
986	struct xfs_buf		*agbp,
987	struct xfs_perag	*pag,
988	xfs_ino_t		parent,
989	xfs_ino_t		*inop)
990{
991	struct xfs_mount	*mp = tp->t_mountp;
992	struct xfs_agi		*agi = agbp->b_addr;
993	xfs_agnumber_t		pagno = XFS_INO_TO_AGNO(mp, parent);
994	xfs_agino_t		pagino = XFS_INO_TO_AGINO(mp, parent);
995	struct xfs_btree_cur	*cur, *tcur;
996	struct xfs_inobt_rec_incore rec, trec;
997	xfs_ino_t		ino;
998	int			error;
999	int			offset;
1000	int			i, j;
1001	int			searchdistance = 10;
1002
1003	ASSERT(pag->pagi_init);
1004	ASSERT(pag->pagi_inodeok);
1005	ASSERT(pag->pagi_freecount > 0);
1006
1007 restart_pagno:
1008	cur = xfs_inobt_init_cursor(mp, tp, agbp, pag, XFS_BTNUM_INO);
1009	/*
1010	 * If pagino is 0 (this is the root inode allocation) use newino.
1011	 * This must work because we've just allocated some.
1012	 */
1013	if (!pagino)
1014		pagino = be32_to_cpu(agi->agi_newino);
1015
1016	error = xfs_check_agi_freecount(cur);
1017	if (error)
1018		goto error0;
1019
1020	/*
1021	 * If in the same AG as the parent, try to get near the parent.
1022	 */
1023	if (pagno == pag->pag_agno) {
1024		int		doneleft;	/* done, to the left */
1025		int		doneright;	/* done, to the right */
1026
1027		error = xfs_inobt_lookup(cur, pagino, XFS_LOOKUP_LE, &i);
1028		if (error)
1029			goto error0;
1030		if (XFS_IS_CORRUPT(mp, i != 1)) {
1031			error = -EFSCORRUPTED;
1032			goto error0;
1033		}
1034
1035		error = xfs_inobt_get_rec(cur, &rec, &j);
1036		if (error)
1037			goto error0;
1038		if (XFS_IS_CORRUPT(mp, j != 1)) {
1039			error = -EFSCORRUPTED;
1040			goto error0;
1041		}
1042
1043		if (rec.ir_freecount > 0) {
1044			/*
1045			 * Found a free inode in the same chunk
1046			 * as the parent, done.
1047			 */
1048			goto alloc_inode;
1049		}
1050
1051
1052		/*
1053		 * In the same AG as parent, but parent's chunk is full.
1054		 */
1055
1056		/* duplicate the cursor, search left & right simultaneously */
1057		error = xfs_btree_dup_cursor(cur, &tcur);
1058		if (error)
1059			goto error0;
1060
1061		/*
1062		 * Skip to last blocks looked up if same parent inode.
1063		 */
1064		if (pagino != NULLAGINO &&
1065		    pag->pagl_pagino == pagino &&
1066		    pag->pagl_leftrec != NULLAGINO &&
1067		    pag->pagl_rightrec != NULLAGINO) {
1068			error = xfs_ialloc_get_rec(tcur, pag->pagl_leftrec,
1069						   &trec, &doneleft);
1070			if (error)
1071				goto error1;
1072
1073			error = xfs_ialloc_get_rec(cur, pag->pagl_rightrec,
1074						   &rec, &doneright);
1075			if (error)
1076				goto error1;
1077		} else {
1078			/* search left with tcur, back up 1 record */
1079			error = xfs_ialloc_next_rec(tcur, &trec, &doneleft, 1);
1080			if (error)
1081				goto error1;
1082
1083			/* search right with cur, go forward 1 record. */
1084			error = xfs_ialloc_next_rec(cur, &rec, &doneright, 0);
1085			if (error)
1086				goto error1;
1087		}
1088
1089		/*
1090		 * Loop until we find an inode chunk with a free inode.
1091		 */
1092		while (--searchdistance > 0 && (!doneleft || !doneright)) {
1093			int	useleft;  /* using left inode chunk this time */
1094
1095			/* figure out the closer block if both are valid. */
1096			if (!doneleft && !doneright) {
1097				useleft = pagino -
1098				 (trec.ir_startino + XFS_INODES_PER_CHUNK - 1) <
1099				  rec.ir_startino - pagino;
1100			} else {
1101				useleft = !doneleft;
1102			}
1103
1104			/* free inodes to the left? */
1105			if (useleft && trec.ir_freecount) {
1106				xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1107				cur = tcur;
1108
1109				pag->pagl_leftrec = trec.ir_startino;
1110				pag->pagl_rightrec = rec.ir_startino;
1111				pag->pagl_pagino = pagino;
1112				rec = trec;
1113				goto alloc_inode;
1114			}
1115
1116			/* free inodes to the right? */
1117			if (!useleft && rec.ir_freecount) {
1118				xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1119
1120				pag->pagl_leftrec = trec.ir_startino;
1121				pag->pagl_rightrec = rec.ir_startino;
1122				pag->pagl_pagino = pagino;
1123				goto alloc_inode;
1124			}
1125
1126			/* get next record to check */
1127			if (useleft) {
1128				error = xfs_ialloc_next_rec(tcur, &trec,
1129								 &doneleft, 1);
1130			} else {
1131				error = xfs_ialloc_next_rec(cur, &rec,
1132								 &doneright, 0);
1133			}
1134			if (error)
1135				goto error1;
1136		}
1137
1138		if (searchdistance <= 0) {
1139			/*
1140			 * Not in range - save last search
1141			 * location and allocate a new inode
1142			 */
1143			xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1144			pag->pagl_leftrec = trec.ir_startino;
1145			pag->pagl_rightrec = rec.ir_startino;
1146			pag->pagl_pagino = pagino;
1147
1148		} else {
1149			/*
1150			 * We've reached the end of the btree. because
1151			 * we are only searching a small chunk of the
1152			 * btree each search, there is obviously free
1153			 * inodes closer to the parent inode than we
1154			 * are now. restart the search again.
1155			 */
1156			pag->pagl_pagino = NULLAGINO;
1157			pag->pagl_leftrec = NULLAGINO;
1158			pag->pagl_rightrec = NULLAGINO;
1159			xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1160			xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1161			goto restart_pagno;
1162		}
1163	}
1164
1165	/*
1166	 * In a different AG from the parent.
1167	 * See if the most recently allocated block has any free.
1168	 */
1169	if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1170		error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1171					 XFS_LOOKUP_EQ, &i);
1172		if (error)
1173			goto error0;
1174
1175		if (i == 1) {
1176			error = xfs_inobt_get_rec(cur, &rec, &j);
1177			if (error)
1178				goto error0;
1179
1180			if (j == 1 && rec.ir_freecount > 0) {
1181				/*
1182				 * The last chunk allocated in the group
1183				 * still has a free inode.
1184				 */
1185				goto alloc_inode;
1186			}
1187		}
1188	}
1189
1190	/*
1191	 * None left in the last group, search the whole AG
1192	 */
1193	error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1194	if (error)
1195		goto error0;
1196	if (XFS_IS_CORRUPT(mp, i != 1)) {
1197		error = -EFSCORRUPTED;
1198		goto error0;
1199	}
1200
1201	for (;;) {
1202		error = xfs_inobt_get_rec(cur, &rec, &i);
1203		if (error)
1204			goto error0;
1205		if (XFS_IS_CORRUPT(mp, i != 1)) {
1206			error = -EFSCORRUPTED;
1207			goto error0;
1208		}
1209		if (rec.ir_freecount > 0)
1210			break;
1211		error = xfs_btree_increment(cur, 0, &i);
1212		if (error)
1213			goto error0;
1214		if (XFS_IS_CORRUPT(mp, i != 1)) {
1215			error = -EFSCORRUPTED;
1216			goto error0;
1217		}
1218	}
1219
1220alloc_inode:
1221	offset = xfs_inobt_first_free_inode(&rec);
1222	ASSERT(offset >= 0);
1223	ASSERT(offset < XFS_INODES_PER_CHUNK);
1224	ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1225				   XFS_INODES_PER_CHUNK) == 0);
1226	ino = XFS_AGINO_TO_INO(mp, pag->pag_agno, rec.ir_startino + offset);
1227	rec.ir_free &= ~XFS_INOBT_MASK(offset);
1228	rec.ir_freecount--;
1229	error = xfs_inobt_update(cur, &rec);
1230	if (error)
1231		goto error0;
1232	be32_add_cpu(&agi->agi_freecount, -1);
1233	xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1234	pag->pagi_freecount--;
1235
1236	error = xfs_check_agi_freecount(cur);
1237	if (error)
1238		goto error0;
1239
1240	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1241	xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1242	*inop = ino;
1243	return 0;
1244error1:
1245	xfs_btree_del_cursor(tcur, XFS_BTREE_ERROR);
1246error0:
1247	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1248	return error;
1249}
1250
1251/*
1252 * Use the free inode btree to allocate an inode based on distance from the
1253 * parent. Note that the provided cursor may be deleted and replaced.
1254 */
1255STATIC int
1256xfs_dialloc_ag_finobt_near(
1257	xfs_agino_t			pagino,
1258	struct xfs_btree_cur		**ocur,
1259	struct xfs_inobt_rec_incore	*rec)
1260{
1261	struct xfs_btree_cur		*lcur = *ocur;	/* left search cursor */
1262	struct xfs_btree_cur		*rcur;	/* right search cursor */
1263	struct xfs_inobt_rec_incore	rrec;
1264	int				error;
1265	int				i, j;
1266
1267	error = xfs_inobt_lookup(lcur, pagino, XFS_LOOKUP_LE, &i);
1268	if (error)
1269		return error;
1270
1271	if (i == 1) {
1272		error = xfs_inobt_get_rec(lcur, rec, &i);
1273		if (error)
1274			return error;
1275		if (XFS_IS_CORRUPT(lcur->bc_mp, i != 1))
1276			return -EFSCORRUPTED;
1277
1278		/*
1279		 * See if we've landed in the parent inode record. The finobt
1280		 * only tracks chunks with at least one free inode, so record
1281		 * existence is enough.
1282		 */
1283		if (pagino >= rec->ir_startino &&
1284		    pagino < (rec->ir_startino + XFS_INODES_PER_CHUNK))
1285			return 0;
1286	}
1287
1288	error = xfs_btree_dup_cursor(lcur, &rcur);
1289	if (error)
1290		return error;
1291
1292	error = xfs_inobt_lookup(rcur, pagino, XFS_LOOKUP_GE, &j);
1293	if (error)
1294		goto error_rcur;
1295	if (j == 1) {
1296		error = xfs_inobt_get_rec(rcur, &rrec, &j);
1297		if (error)
1298			goto error_rcur;
1299		if (XFS_IS_CORRUPT(lcur->bc_mp, j != 1)) {
1300			error = -EFSCORRUPTED;
1301			goto error_rcur;
1302		}
1303	}
1304
1305	if (XFS_IS_CORRUPT(lcur->bc_mp, i != 1 && j != 1)) {
1306		error = -EFSCORRUPTED;
1307		goto error_rcur;
1308	}
1309	if (i == 1 && j == 1) {
1310		/*
1311		 * Both the left and right records are valid. Choose the closer
1312		 * inode chunk to the target.
1313		 */
1314		if ((pagino - rec->ir_startino + XFS_INODES_PER_CHUNK - 1) >
1315		    (rrec.ir_startino - pagino)) {
1316			*rec = rrec;
1317			xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1318			*ocur = rcur;
1319		} else {
1320			xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1321		}
1322	} else if (j == 1) {
1323		/* only the right record is valid */
1324		*rec = rrec;
1325		xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1326		*ocur = rcur;
1327	} else if (i == 1) {
1328		/* only the left record is valid */
1329		xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1330	}
1331
1332	return 0;
1333
1334error_rcur:
1335	xfs_btree_del_cursor(rcur, XFS_BTREE_ERROR);
1336	return error;
1337}
1338
1339/*
1340 * Use the free inode btree to find a free inode based on a newino hint. If
1341 * the hint is NULL, find the first free inode in the AG.
1342 */
1343STATIC int
1344xfs_dialloc_ag_finobt_newino(
1345	struct xfs_agi			*agi,
1346	struct xfs_btree_cur		*cur,
1347	struct xfs_inobt_rec_incore	*rec)
1348{
1349	int error;
1350	int i;
1351
1352	if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1353		error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1354					 XFS_LOOKUP_EQ, &i);
1355		if (error)
1356			return error;
1357		if (i == 1) {
1358			error = xfs_inobt_get_rec(cur, rec, &i);
1359			if (error)
1360				return error;
1361			if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1362				return -EFSCORRUPTED;
1363			return 0;
1364		}
1365	}
1366
1367	/*
1368	 * Find the first inode available in the AG.
1369	 */
1370	error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1371	if (error)
1372		return error;
1373	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1374		return -EFSCORRUPTED;
1375
1376	error = xfs_inobt_get_rec(cur, rec, &i);
1377	if (error)
1378		return error;
1379	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1380		return -EFSCORRUPTED;
1381
1382	return 0;
1383}
1384
1385/*
1386 * Update the inobt based on a modification made to the finobt. Also ensure that
1387 * the records from both trees are equivalent post-modification.
1388 */
1389STATIC int
1390xfs_dialloc_ag_update_inobt(
1391	struct xfs_btree_cur		*cur,	/* inobt cursor */
1392	struct xfs_inobt_rec_incore	*frec,	/* finobt record */
1393	int				offset) /* inode offset */
1394{
1395	struct xfs_inobt_rec_incore	rec;
1396	int				error;
1397	int				i;
1398
1399	error = xfs_inobt_lookup(cur, frec->ir_startino, XFS_LOOKUP_EQ, &i);
1400	if (error)
1401		return error;
1402	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1403		return -EFSCORRUPTED;
1404
1405	error = xfs_inobt_get_rec(cur, &rec, &i);
1406	if (error)
1407		return error;
1408	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1409		return -EFSCORRUPTED;
1410	ASSERT((XFS_AGINO_TO_OFFSET(cur->bc_mp, rec.ir_startino) %
1411				   XFS_INODES_PER_CHUNK) == 0);
1412
1413	rec.ir_free &= ~XFS_INOBT_MASK(offset);
1414	rec.ir_freecount--;
1415
1416	if (XFS_IS_CORRUPT(cur->bc_mp,
1417			   rec.ir_free != frec->ir_free ||
1418			   rec.ir_freecount != frec->ir_freecount))
1419		return -EFSCORRUPTED;
1420
1421	return xfs_inobt_update(cur, &rec);
1422}
1423
1424/*
1425 * Allocate an inode using the free inode btree, if available. Otherwise, fall
1426 * back to the inobt search algorithm.
1427 *
1428 * The caller selected an AG for us, and made sure that free inodes are
1429 * available.
1430 */
1431static int
1432xfs_dialloc_ag(
1433	struct xfs_trans	*tp,
1434	struct xfs_buf		*agbp,
1435	struct xfs_perag	*pag,
1436	xfs_ino_t		parent,
1437	xfs_ino_t		*inop)
1438{
1439	struct xfs_mount		*mp = tp->t_mountp;
1440	struct xfs_agi			*agi = agbp->b_addr;
1441	xfs_agnumber_t			pagno = XFS_INO_TO_AGNO(mp, parent);
1442	xfs_agino_t			pagino = XFS_INO_TO_AGINO(mp, parent);
1443	struct xfs_btree_cur		*cur;	/* finobt cursor */
1444	struct xfs_btree_cur		*icur;	/* inobt cursor */
1445	struct xfs_inobt_rec_incore	rec;
1446	xfs_ino_t			ino;
1447	int				error;
1448	int				offset;
1449	int				i;
1450
1451	if (!xfs_has_finobt(mp))
1452		return xfs_dialloc_ag_inobt(tp, agbp, pag, parent, inop);
1453
1454	/*
1455	 * If pagino is 0 (this is the root inode allocation) use newino.
1456	 * This must work because we've just allocated some.
1457	 */
1458	if (!pagino)
1459		pagino = be32_to_cpu(agi->agi_newino);
1460
1461	cur = xfs_inobt_init_cursor(mp, tp, agbp, pag, XFS_BTNUM_FINO);
1462
1463	error = xfs_check_agi_freecount(cur);
1464	if (error)
1465		goto error_cur;
1466
1467	/*
1468	 * The search algorithm depends on whether we're in the same AG as the
1469	 * parent. If so, find the closest available inode to the parent. If
1470	 * not, consider the agi hint or find the first free inode in the AG.
1471	 */
1472	if (pag->pag_agno == pagno)
1473		error = xfs_dialloc_ag_finobt_near(pagino, &cur, &rec);
1474	else
1475		error = xfs_dialloc_ag_finobt_newino(agi, cur, &rec);
1476	if (error)
1477		goto error_cur;
1478
1479	offset = xfs_inobt_first_free_inode(&rec);
1480	ASSERT(offset >= 0);
1481	ASSERT(offset < XFS_INODES_PER_CHUNK);
1482	ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1483				   XFS_INODES_PER_CHUNK) == 0);
1484	ino = XFS_AGINO_TO_INO(mp, pag->pag_agno, rec.ir_startino + offset);
1485
1486	/*
1487	 * Modify or remove the finobt record.
1488	 */
1489	rec.ir_free &= ~XFS_INOBT_MASK(offset);
1490	rec.ir_freecount--;
1491	if (rec.ir_freecount)
1492		error = xfs_inobt_update(cur, &rec);
1493	else
1494		error = xfs_btree_delete(cur, &i);
1495	if (error)
1496		goto error_cur;
1497
1498	/*
1499	 * The finobt has now been updated appropriately. We haven't updated the
1500	 * agi and superblock yet, so we can create an inobt cursor and validate
1501	 * the original freecount. If all is well, make the equivalent update to
1502	 * the inobt using the finobt record and offset information.
1503	 */
1504	icur = xfs_inobt_init_cursor(mp, tp, agbp, pag, XFS_BTNUM_INO);
1505
1506	error = xfs_check_agi_freecount(icur);
1507	if (error)
1508		goto error_icur;
1509
1510	error = xfs_dialloc_ag_update_inobt(icur, &rec, offset);
1511	if (error)
1512		goto error_icur;
1513
1514	/*
1515	 * Both trees have now been updated. We must update the perag and
1516	 * superblock before we can check the freecount for each btree.
1517	 */
1518	be32_add_cpu(&agi->agi_freecount, -1);
1519	xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1520	pag->pagi_freecount--;
1521
1522	xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1523
1524	error = xfs_check_agi_freecount(icur);
1525	if (error)
1526		goto error_icur;
1527	error = xfs_check_agi_freecount(cur);
1528	if (error)
1529		goto error_icur;
1530
1531	xfs_btree_del_cursor(icur, XFS_BTREE_NOERROR);
1532	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1533	*inop = ino;
1534	return 0;
1535
1536error_icur:
1537	xfs_btree_del_cursor(icur, XFS_BTREE_ERROR);
1538error_cur:
1539	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1540	return error;
1541}
1542
1543static int
1544xfs_dialloc_roll(
1545	struct xfs_trans	**tpp,
1546	struct xfs_buf		*agibp)
1547{
1548	struct xfs_trans	*tp = *tpp;
1549	struct xfs_dquot_acct	*dqinfo;
1550	int			error;
1551
1552	/*
1553	 * Hold to on to the agibp across the commit so no other allocation can
1554	 * come in and take the free inodes we just allocated for our caller.
1555	 */
1556	xfs_trans_bhold(tp, agibp);
1557
1558	/*
1559	 * We want the quota changes to be associated with the next transaction,
1560	 * NOT this one. So, detach the dqinfo from this and attach it to the
1561	 * next transaction.
1562	 */
1563	dqinfo = tp->t_dqinfo;
1564	tp->t_dqinfo = NULL;
1565
1566	error = xfs_trans_roll(&tp);
1567
1568	/* Re-attach the quota info that we detached from prev trx. */
1569	tp->t_dqinfo = dqinfo;
1570
1571	/*
1572	 * Join the buffer even on commit error so that the buffer is released
1573	 * when the caller cancels the transaction and doesn't have to handle
1574	 * this error case specially.
1575	 */
1576	xfs_trans_bjoin(tp, agibp);
1577	*tpp = tp;
1578	return error;
1579}
1580
1581static xfs_agnumber_t
1582xfs_ialloc_next_ag(
1583	xfs_mount_t	*mp)
1584{
1585	xfs_agnumber_t	agno;
1586
1587	spin_lock(&mp->m_agirotor_lock);
1588	agno = mp->m_agirotor;
1589	if (++mp->m_agirotor >= mp->m_maxagi)
1590		mp->m_agirotor = 0;
1591	spin_unlock(&mp->m_agirotor_lock);
1592
1593	return agno;
1594}
1595
1596static bool
1597xfs_dialloc_good_ag(
1598	struct xfs_trans	*tp,
1599	struct xfs_perag	*pag,
1600	umode_t			mode,
1601	int			flags,
1602	bool			ok_alloc)
1603{
1604	struct xfs_mount	*mp = tp->t_mountp;
1605	xfs_extlen_t		ineed;
1606	xfs_extlen_t		longest = 0;
1607	int			needspace;
1608	int			error;
1609
1610	if (!pag->pagi_inodeok)
1611		return false;
1612
1613	if (!pag->pagi_init) {
1614		error = xfs_ialloc_pagi_init(mp, tp, pag->pag_agno);
1615		if (error)
1616			return false;
1617	}
1618
1619	if (pag->pagi_freecount)
1620		return true;
1621	if (!ok_alloc)
1622		return false;
1623
1624	if (!pag->pagf_init) {
1625		error = xfs_alloc_pagf_init(mp, tp, pag->pag_agno, flags);
1626		if (error)
1627			return false;
1628	}
1629
1630	/*
1631	 * Check that there is enough free space for the file plus a chunk of
1632	 * inodes if we need to allocate some. If this is the first pass across
1633	 * the AGs, take into account the potential space needed for alignment
1634	 * of inode chunks when checking the longest contiguous free space in
1635	 * the AG - this prevents us from getting ENOSPC because we have free
1636	 * space larger than ialloc_blks but alignment constraints prevent us
1637	 * from using it.
1638	 *
1639	 * If we can't find an AG with space for full alignment slack to be
1640	 * taken into account, we must be near ENOSPC in all AGs.  Hence we
1641	 * don't include alignment for the second pass and so if we fail
1642	 * allocation due to alignment issues then it is most likely a real
1643	 * ENOSPC condition.
1644	 *
1645	 * XXX(dgc): this calculation is now bogus thanks to the per-ag
1646	 * reservations that xfs_alloc_fix_freelist() now does via
1647	 * xfs_alloc_space_available(). When the AG fills up, pagf_freeblks will
1648	 * be more than large enough for the check below to succeed, but
1649	 * xfs_alloc_space_available() will fail because of the non-zero
1650	 * metadata reservation and hence we won't actually be able to allocate
1651	 * more inodes in this AG. We do soooo much unnecessary work near ENOSPC
1652	 * because of this.
1653	 */
1654	ineed = M_IGEO(mp)->ialloc_min_blks;
1655	if (flags && ineed > 1)
1656		ineed += M_IGEO(mp)->cluster_align;
1657	longest = pag->pagf_longest;
1658	if (!longest)
1659		longest = pag->pagf_flcount > 0;
1660	needspace = S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode);
1661
1662	if (pag->pagf_freeblks < needspace + ineed || longest < ineed)
1663		return false;
1664	return true;
1665}
1666
1667static int
1668xfs_dialloc_try_ag(
1669	struct xfs_trans	**tpp,
1670	struct xfs_perag	*pag,
1671	xfs_ino_t		parent,
1672	xfs_ino_t		*new_ino,
1673	bool			ok_alloc)
1674{
1675	struct xfs_buf		*agbp;
1676	xfs_ino_t		ino;
1677	int			error;
1678
1679	/*
1680	 * Then read in the AGI buffer and recheck with the AGI buffer
1681	 * lock held.
1682	 */
1683	error = xfs_ialloc_read_agi(pag->pag_mount, *tpp, pag->pag_agno, &agbp);
1684	if (error)
1685		return error;
1686
1687	if (!pag->pagi_freecount) {
1688		if (!ok_alloc) {
1689			error = -EAGAIN;
1690			goto out_release;
1691		}
1692
1693		error = xfs_ialloc_ag_alloc(*tpp, agbp, pag);
1694		if (error < 0)
1695			goto out_release;
1696
1697		/*
1698		 * We successfully allocated space for an inode cluster in this
1699		 * AG.  Roll the transaction so that we can allocate one of the
1700		 * new inodes.
1701		 */
1702		ASSERT(pag->pagi_freecount > 0);
1703		error = xfs_dialloc_roll(tpp, agbp);
1704		if (error)
1705			goto out_release;
1706	}
1707
1708	/* Allocate an inode in the found AG */
1709	error = xfs_dialloc_ag(*tpp, agbp, pag, parent, &ino);
1710	if (!error)
1711		*new_ino = ino;
1712	return error;
1713
1714out_release:
1715	xfs_trans_brelse(*tpp, agbp);
1716	return error;
1717}
1718
1719/*
1720 * Allocate an on-disk inode.
1721 *
1722 * Mode is used to tell whether the new inode is a directory and hence where to
1723 * locate it. The on-disk inode that is allocated will be returned in @new_ino
1724 * on success, otherwise an error will be set to indicate the failure (e.g.
1725 * -ENOSPC).
1726 */
1727int
1728xfs_dialloc(
1729	struct xfs_trans	**tpp,
1730	xfs_ino_t		parent,
1731	umode_t			mode,
1732	xfs_ino_t		*new_ino)
1733{
1734	struct xfs_mount	*mp = (*tpp)->t_mountp;
1735	xfs_agnumber_t		agno;
1736	int			error = 0;
1737	xfs_agnumber_t		start_agno;
1738	struct xfs_perag	*pag;
1739	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
1740	bool			ok_alloc = true;
1741	int			flags;
1742	xfs_ino_t		ino;
1743
1744	/*
1745	 * Directories, symlinks, and regular files frequently allocate at least
1746	 * one block, so factor that potential expansion when we examine whether
1747	 * an AG has enough space for file creation.
1748	 */
1749	if (S_ISDIR(mode))
1750		start_agno = xfs_ialloc_next_ag(mp);
1751	else {
1752		start_agno = XFS_INO_TO_AGNO(mp, parent);
1753		if (start_agno >= mp->m_maxagi)
1754			start_agno = 0;
1755	}
1756
1757	/*
1758	 * If we have already hit the ceiling of inode blocks then clear
1759	 * ok_alloc so we scan all available agi structures for a free
1760	 * inode.
1761	 *
1762	 * Read rough value of mp->m_icount by percpu_counter_read_positive,
1763	 * which will sacrifice the preciseness but improve the performance.
1764	 */
1765	if (igeo->maxicount &&
1766	    percpu_counter_read_positive(&mp->m_icount) + igeo->ialloc_inos
1767							> igeo->maxicount) {
1768		ok_alloc = false;
1769	}
1770
1771	/*
1772	 * Loop until we find an allocation group that either has free inodes
1773	 * or in which we can allocate some inodes.  Iterate through the
1774	 * allocation groups upward, wrapping at the end.
1775	 */
1776	agno = start_agno;
1777	flags = XFS_ALLOC_FLAG_TRYLOCK;
1778	for (;;) {
1779		pag = xfs_perag_get(mp, agno);
1780		if (xfs_dialloc_good_ag(*tpp, pag, mode, flags, ok_alloc)) {
1781			error = xfs_dialloc_try_ag(tpp, pag, parent,
1782					&ino, ok_alloc);
1783			if (error != -EAGAIN)
1784				break;
1785		}
1786
1787		if (xfs_is_shutdown(mp)) {
1788			error = -EFSCORRUPTED;
1789			break;
1790		}
1791		if (++agno == mp->m_maxagi)
1792			agno = 0;
1793		if (agno == start_agno) {
1794			if (!flags) {
1795				error = -ENOSPC;
1796				break;
1797			}
1798			flags = 0;
1799		}
1800		xfs_perag_put(pag);
1801	}
1802
1803	if (!error)
1804		*new_ino = ino;
1805	xfs_perag_put(pag);
1806	return error;
1807}
1808
1809/*
1810 * Free the blocks of an inode chunk. We must consider that the inode chunk
1811 * might be sparse and only free the regions that are allocated as part of the
1812 * chunk.
1813 */
1814STATIC void
1815xfs_difree_inode_chunk(
1816	struct xfs_trans		*tp,
1817	xfs_agnumber_t			agno,
1818	struct xfs_inobt_rec_incore	*rec)
1819{
1820	struct xfs_mount		*mp = tp->t_mountp;
1821	xfs_agblock_t			sagbno = XFS_AGINO_TO_AGBNO(mp,
1822							rec->ir_startino);
1823	int				startidx, endidx;
1824	int				nextbit;
1825	xfs_agblock_t			agbno;
1826	int				contigblk;
1827	DECLARE_BITMAP(holemask, XFS_INOBT_HOLEMASK_BITS);
1828
1829	if (!xfs_inobt_issparse(rec->ir_holemask)) {
1830		/* not sparse, calculate extent info directly */
1831		xfs_bmap_add_free(tp, XFS_AGB_TO_FSB(mp, agno, sagbno),
1832				  M_IGEO(mp)->ialloc_blks,
1833				  &XFS_RMAP_OINFO_INODES);
1834		return;
1835	}
1836
1837	/* holemask is only 16-bits (fits in an unsigned long) */
1838	ASSERT(sizeof(rec->ir_holemask) <= sizeof(holemask[0]));
1839	holemask[0] = rec->ir_holemask;
1840
1841	/*
1842	 * Find contiguous ranges of zeroes (i.e., allocated regions) in the
1843	 * holemask and convert the start/end index of each range to an extent.
1844	 * We start with the start and end index both pointing at the first 0 in
1845	 * the mask.
1846	 */
1847	startidx = endidx = find_first_zero_bit(holemask,
1848						XFS_INOBT_HOLEMASK_BITS);
1849	nextbit = startidx + 1;
1850	while (startidx < XFS_INOBT_HOLEMASK_BITS) {
1851		nextbit = find_next_zero_bit(holemask, XFS_INOBT_HOLEMASK_BITS,
1852					     nextbit);
1853		/*
1854		 * If the next zero bit is contiguous, update the end index of
1855		 * the current range and continue.
1856		 */
1857		if (nextbit != XFS_INOBT_HOLEMASK_BITS &&
1858		    nextbit == endidx + 1) {
1859			endidx = nextbit;
1860			goto next;
1861		}
1862
1863		/*
1864		 * nextbit is not contiguous with the current end index. Convert
1865		 * the current start/end to an extent and add it to the free
1866		 * list.
1867		 */
1868		agbno = sagbno + (startidx * XFS_INODES_PER_HOLEMASK_BIT) /
1869				  mp->m_sb.sb_inopblock;
1870		contigblk = ((endidx - startidx + 1) *
1871			     XFS_INODES_PER_HOLEMASK_BIT) /
1872			    mp->m_sb.sb_inopblock;
1873
1874		ASSERT(agbno % mp->m_sb.sb_spino_align == 0);
1875		ASSERT(contigblk % mp->m_sb.sb_spino_align == 0);
1876		xfs_bmap_add_free(tp, XFS_AGB_TO_FSB(mp, agno, agbno),
1877				  contigblk, &XFS_RMAP_OINFO_INODES);
1878
1879		/* reset range to current bit and carry on... */
1880		startidx = endidx = nextbit;
1881
1882next:
1883		nextbit++;
1884	}
1885}
1886
1887STATIC int
1888xfs_difree_inobt(
1889	struct xfs_mount		*mp,
1890	struct xfs_trans		*tp,
1891	struct xfs_buf			*agbp,
1892	struct xfs_perag		*pag,
1893	xfs_agino_t			agino,
1894	struct xfs_icluster		*xic,
1895	struct xfs_inobt_rec_incore	*orec)
1896{
1897	struct xfs_agi			*agi = agbp->b_addr;
1898	struct xfs_btree_cur		*cur;
1899	struct xfs_inobt_rec_incore	rec;
1900	int				ilen;
1901	int				error;
1902	int				i;
1903	int				off;
1904
1905	ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
1906	ASSERT(XFS_AGINO_TO_AGBNO(mp, agino) < be32_to_cpu(agi->agi_length));
1907
1908	/*
1909	 * Initialize the cursor.
1910	 */
1911	cur = xfs_inobt_init_cursor(mp, tp, agbp, pag, XFS_BTNUM_INO);
1912
1913	error = xfs_check_agi_freecount(cur);
1914	if (error)
1915		goto error0;
1916
1917	/*
1918	 * Look for the entry describing this inode.
1919	 */
1920	if ((error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i))) {
1921		xfs_warn(mp, "%s: xfs_inobt_lookup() returned error %d.",
1922			__func__, error);
1923		goto error0;
1924	}
1925	if (XFS_IS_CORRUPT(mp, i != 1)) {
1926		error = -EFSCORRUPTED;
1927		goto error0;
1928	}
1929	error = xfs_inobt_get_rec(cur, &rec, &i);
1930	if (error) {
1931		xfs_warn(mp, "%s: xfs_inobt_get_rec() returned error %d.",
1932			__func__, error);
1933		goto error0;
1934	}
1935	if (XFS_IS_CORRUPT(mp, i != 1)) {
1936		error = -EFSCORRUPTED;
1937		goto error0;
1938	}
1939	/*
1940	 * Get the offset in the inode chunk.
1941	 */
1942	off = agino - rec.ir_startino;
1943	ASSERT(off >= 0 && off < XFS_INODES_PER_CHUNK);
1944	ASSERT(!(rec.ir_free & XFS_INOBT_MASK(off)));
1945	/*
1946	 * Mark the inode free & increment the count.
1947	 */
1948	rec.ir_free |= XFS_INOBT_MASK(off);
1949	rec.ir_freecount++;
1950
1951	/*
1952	 * When an inode chunk is free, it becomes eligible for removal. Don't
1953	 * remove the chunk if the block size is large enough for multiple inode
1954	 * chunks (that might not be free).
1955	 */
1956	if (!xfs_has_ikeep(mp) && rec.ir_free == XFS_INOBT_ALL_FREE &&
1957	    mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK) {
1958		struct xfs_perag	*pag = agbp->b_pag;
1959
1960		xic->deleted = true;
1961		xic->first_ino = XFS_AGINO_TO_INO(mp, pag->pag_agno,
1962				rec.ir_startino);
1963		xic->alloc = xfs_inobt_irec_to_allocmask(&rec);
1964
1965		/*
1966		 * Remove the inode cluster from the AGI B+Tree, adjust the
1967		 * AGI and Superblock inode counts, and mark the disk space
1968		 * to be freed when the transaction is committed.
1969		 */
1970		ilen = rec.ir_freecount;
1971		be32_add_cpu(&agi->agi_count, -ilen);
1972		be32_add_cpu(&agi->agi_freecount, -(ilen - 1));
1973		xfs_ialloc_log_agi(tp, agbp, XFS_AGI_COUNT | XFS_AGI_FREECOUNT);
1974		pag->pagi_freecount -= ilen - 1;
1975		pag->pagi_count -= ilen;
1976		xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, -ilen);
1977		xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -(ilen - 1));
1978
1979		if ((error = xfs_btree_delete(cur, &i))) {
1980			xfs_warn(mp, "%s: xfs_btree_delete returned error %d.",
1981				__func__, error);
1982			goto error0;
1983		}
1984
1985		xfs_difree_inode_chunk(tp, pag->pag_agno, &rec);
1986	} else {
1987		xic->deleted = false;
1988
1989		error = xfs_inobt_update(cur, &rec);
1990		if (error) {
1991			xfs_warn(mp, "%s: xfs_inobt_update returned error %d.",
1992				__func__, error);
1993			goto error0;
1994		}
1995
1996		/*
1997		 * Change the inode free counts and log the ag/sb changes.
1998		 */
1999		be32_add_cpu(&agi->agi_freecount, 1);
2000		xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
2001		pag->pagi_freecount++;
2002		xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, 1);
2003	}
2004
2005	error = xfs_check_agi_freecount(cur);
2006	if (error)
2007		goto error0;
2008
2009	*orec = rec;
2010	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2011	return 0;
2012
2013error0:
2014	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2015	return error;
2016}
2017
2018/*
2019 * Free an inode in the free inode btree.
2020 */
2021STATIC int
2022xfs_difree_finobt(
2023	struct xfs_mount		*mp,
2024	struct xfs_trans		*tp,
2025	struct xfs_buf			*agbp,
2026	struct xfs_perag		*pag,
2027	xfs_agino_t			agino,
2028	struct xfs_inobt_rec_incore	*ibtrec) /* inobt record */
2029{
2030	struct xfs_btree_cur		*cur;
2031	struct xfs_inobt_rec_incore	rec;
2032	int				offset = agino - ibtrec->ir_startino;
2033	int				error;
2034	int				i;
2035
2036	cur = xfs_inobt_init_cursor(mp, tp, agbp, pag, XFS_BTNUM_FINO);
2037
2038	error = xfs_inobt_lookup(cur, ibtrec->ir_startino, XFS_LOOKUP_EQ, &i);
2039	if (error)
2040		goto error;
2041	if (i == 0) {
2042		/*
2043		 * If the record does not exist in the finobt, we must have just
2044		 * freed an inode in a previously fully allocated chunk. If not,
2045		 * something is out of sync.
2046		 */
2047		if (XFS_IS_CORRUPT(mp, ibtrec->ir_freecount != 1)) {
2048			error = -EFSCORRUPTED;
2049			goto error;
2050		}
2051
2052		error = xfs_inobt_insert_rec(cur, ibtrec->ir_holemask,
2053					     ibtrec->ir_count,
2054					     ibtrec->ir_freecount,
2055					     ibtrec->ir_free, &i);
2056		if (error)
2057			goto error;
2058		ASSERT(i == 1);
2059
2060		goto out;
2061	}
2062
2063	/*
2064	 * Read and update the existing record. We could just copy the ibtrec
2065	 * across here, but that would defeat the purpose of having redundant
2066	 * metadata. By making the modifications independently, we can catch
2067	 * corruptions that we wouldn't see if we just copied from one record
2068	 * to another.
2069	 */
2070	error = xfs_inobt_get_rec(cur, &rec, &i);
2071	if (error)
2072		goto error;
2073	if (XFS_IS_CORRUPT(mp, i != 1)) {
2074		error = -EFSCORRUPTED;
2075		goto error;
2076	}
2077
2078	rec.ir_free |= XFS_INOBT_MASK(offset);
2079	rec.ir_freecount++;
2080
2081	if (XFS_IS_CORRUPT(mp,
2082			   rec.ir_free != ibtrec->ir_free ||
2083			   rec.ir_freecount != ibtrec->ir_freecount)) {
2084		error = -EFSCORRUPTED;
2085		goto error;
2086	}
2087
2088	/*
2089	 * The content of inobt records should always match between the inobt
2090	 * and finobt. The lifecycle of records in the finobt is different from
2091	 * the inobt in that the finobt only tracks records with at least one
2092	 * free inode. Hence, if all of the inodes are free and we aren't
2093	 * keeping inode chunks permanently on disk, remove the record.
2094	 * Otherwise, update the record with the new information.
2095	 *
2096	 * Note that we currently can't free chunks when the block size is large
2097	 * enough for multiple chunks. Leave the finobt record to remain in sync
2098	 * with the inobt.
2099	 */
2100	if (!xfs_has_ikeep(mp) && rec.ir_free == XFS_INOBT_ALL_FREE &&
2101	    mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK) {
2102		error = xfs_btree_delete(cur, &i);
2103		if (error)
2104			goto error;
2105		ASSERT(i == 1);
2106	} else {
2107		error = xfs_inobt_update(cur, &rec);
2108		if (error)
2109			goto error;
2110	}
2111
2112out:
2113	error = xfs_check_agi_freecount(cur);
2114	if (error)
2115		goto error;
2116
2117	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2118	return 0;
2119
2120error:
2121	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2122	return error;
2123}
2124
2125/*
2126 * Free disk inode.  Carefully avoids touching the incore inode, all
2127 * manipulations incore are the caller's responsibility.
2128 * The on-disk inode is not changed by this operation, only the
2129 * btree (free inode mask) is changed.
2130 */
2131int
2132xfs_difree(
2133	struct xfs_trans	*tp,
2134	struct xfs_perag	*pag,
2135	xfs_ino_t		inode,
2136	struct xfs_icluster	*xic)
2137{
2138	/* REFERENCED */
2139	xfs_agblock_t		agbno;	/* block number containing inode */
2140	struct xfs_buf		*agbp;	/* buffer for allocation group header */
2141	xfs_agino_t		agino;	/* allocation group inode number */
2142	int			error;	/* error return value */
2143	struct xfs_mount	*mp = tp->t_mountp;
2144	struct xfs_inobt_rec_incore rec;/* btree record */
2145
2146	/*
2147	 * Break up inode number into its components.
2148	 */
2149	if (pag->pag_agno != XFS_INO_TO_AGNO(mp, inode)) {
2150		xfs_warn(mp, "%s: agno != pag->pag_agno (%d != %d).",
2151			__func__, XFS_INO_TO_AGNO(mp, inode), pag->pag_agno);
2152		ASSERT(0);
2153		return -EINVAL;
2154	}
2155	agino = XFS_INO_TO_AGINO(mp, inode);
2156	if (inode != XFS_AGINO_TO_INO(mp, pag->pag_agno, agino))  {
2157		xfs_warn(mp, "%s: inode != XFS_AGINO_TO_INO() (%llu != %llu).",
2158			__func__, (unsigned long long)inode,
2159			(unsigned long long)XFS_AGINO_TO_INO(mp, pag->pag_agno, agino));
2160		ASSERT(0);
2161		return -EINVAL;
2162	}
2163	agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2164	if (agbno >= mp->m_sb.sb_agblocks)  {
2165		xfs_warn(mp, "%s: agbno >= mp->m_sb.sb_agblocks (%d >= %d).",
2166			__func__, agbno, mp->m_sb.sb_agblocks);
2167		ASSERT(0);
2168		return -EINVAL;
2169	}
2170	/*
2171	 * Get the allocation group header.
2172	 */
2173	error = xfs_ialloc_read_agi(mp, tp, pag->pag_agno, &agbp);
2174	if (error) {
2175		xfs_warn(mp, "%s: xfs_ialloc_read_agi() returned error %d.",
2176			__func__, error);
2177		return error;
2178	}
2179
2180	/*
2181	 * Fix up the inode allocation btree.
2182	 */
2183	error = xfs_difree_inobt(mp, tp, agbp, pag, agino, xic, &rec);
2184	if (error)
2185		goto error0;
2186
2187	/*
2188	 * Fix up the free inode btree.
2189	 */
2190	if (xfs_has_finobt(mp)) {
2191		error = xfs_difree_finobt(mp, tp, agbp, pag, agino, &rec);
2192		if (error)
2193			goto error0;
2194	}
2195
2196	return 0;
2197
2198error0:
2199	return error;
2200}
2201
2202STATIC int
2203xfs_imap_lookup(
2204	struct xfs_mount	*mp,
2205	struct xfs_trans	*tp,
2206	struct xfs_perag	*pag,
2207	xfs_agino_t		agino,
2208	xfs_agblock_t		agbno,
2209	xfs_agblock_t		*chunk_agbno,
2210	xfs_agblock_t		*offset_agbno,
2211	int			flags)
2212{
2213	struct xfs_inobt_rec_incore rec;
2214	struct xfs_btree_cur	*cur;
2215	struct xfs_buf		*agbp;
2216	int			error;
2217	int			i;
2218
2219	error = xfs_ialloc_read_agi(mp, tp, pag->pag_agno, &agbp);
2220	if (error) {
2221		xfs_alert(mp,
2222			"%s: xfs_ialloc_read_agi() returned error %d, agno %d",
2223			__func__, error, pag->pag_agno);
2224		return error;
2225	}
2226
2227	/*
2228	 * Lookup the inode record for the given agino. If the record cannot be
2229	 * found, then it's an invalid inode number and we should abort. Once
2230	 * we have a record, we need to ensure it contains the inode number
2231	 * we are looking up.
2232	 */
2233	cur = xfs_inobt_init_cursor(mp, tp, agbp, pag, XFS_BTNUM_INO);
2234	error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i);
2235	if (!error) {
2236		if (i)
2237			error = xfs_inobt_get_rec(cur, &rec, &i);
2238		if (!error && i == 0)
2239			error = -EINVAL;
2240	}
2241
2242	xfs_trans_brelse(tp, agbp);
2243	xfs_btree_del_cursor(cur, error);
2244	if (error)
2245		return error;
2246
2247	/* check that the returned record contains the required inode */
2248	if (rec.ir_startino > agino ||
2249	    rec.ir_startino + M_IGEO(mp)->ialloc_inos <= agino)
2250		return -EINVAL;
2251
2252	/* for untrusted inodes check it is allocated first */
2253	if ((flags & XFS_IGET_UNTRUSTED) &&
2254	    (rec.ir_free & XFS_INOBT_MASK(agino - rec.ir_startino)))
2255		return -EINVAL;
2256
2257	*chunk_agbno = XFS_AGINO_TO_AGBNO(mp, rec.ir_startino);
2258	*offset_agbno = agbno - *chunk_agbno;
2259	return 0;
2260}
2261
2262/*
2263 * Return the location of the inode in imap, for mapping it into a buffer.
2264 */
2265int
2266xfs_imap(
2267	struct xfs_mount	 *mp,	/* file system mount structure */
2268	struct xfs_trans	 *tp,	/* transaction pointer */
2269	xfs_ino_t		ino,	/* inode to locate */
2270	struct xfs_imap		*imap,	/* location map structure */
2271	uint			flags)	/* flags for inode btree lookup */
2272{
2273	xfs_agblock_t		agbno;	/* block number of inode in the alloc group */
2274	xfs_agino_t		agino;	/* inode number within alloc group */
2275	xfs_agblock_t		chunk_agbno;	/* first block in inode chunk */
2276	xfs_agblock_t		cluster_agbno;	/* first block in inode cluster */
2277	int			error;	/* error code */
2278	int			offset;	/* index of inode in its buffer */
2279	xfs_agblock_t		offset_agbno;	/* blks from chunk start to inode */
2280	struct xfs_perag	*pag;
2281
2282	ASSERT(ino != NULLFSINO);
2283
2284	/*
2285	 * Split up the inode number into its parts.
2286	 */
2287	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
2288	agino = XFS_INO_TO_AGINO(mp, ino);
2289	agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2290	if (!pag || agbno >= mp->m_sb.sb_agblocks ||
2291	    ino != XFS_AGINO_TO_INO(mp, pag->pag_agno, agino)) {
2292		error = -EINVAL;
2293#ifdef DEBUG
2294		/*
2295		 * Don't output diagnostic information for untrusted inodes
2296		 * as they can be invalid without implying corruption.
2297		 */
2298		if (flags & XFS_IGET_UNTRUSTED)
2299			goto out_drop;
2300		if (!pag) {
2301			xfs_alert(mp,
2302				"%s: agno (%d) >= mp->m_sb.sb_agcount (%d)",
2303				__func__, XFS_INO_TO_AGNO(mp, ino),
2304				mp->m_sb.sb_agcount);
2305		}
2306		if (agbno >= mp->m_sb.sb_agblocks) {
2307			xfs_alert(mp,
2308		"%s: agbno (0x%llx) >= mp->m_sb.sb_agblocks (0x%lx)",
2309				__func__, (unsigned long long)agbno,
2310				(unsigned long)mp->m_sb.sb_agblocks);
2311		}
2312		if (pag && ino != XFS_AGINO_TO_INO(mp, pag->pag_agno, agino)) {
2313			xfs_alert(mp,
2314		"%s: ino (0x%llx) != XFS_AGINO_TO_INO() (0x%llx)",
2315				__func__, ino,
2316				XFS_AGINO_TO_INO(mp, pag->pag_agno, agino));
2317		}
2318		xfs_stack_trace();
2319#endif /* DEBUG */
2320		goto out_drop;
2321	}
2322
2323	/*
2324	 * For bulkstat and handle lookups, we have an untrusted inode number
2325	 * that we have to verify is valid. We cannot do this just by reading
2326	 * the inode buffer as it may have been unlinked and removed leaving
2327	 * inodes in stale state on disk. Hence we have to do a btree lookup
2328	 * in all cases where an untrusted inode number is passed.
2329	 */
2330	if (flags & XFS_IGET_UNTRUSTED) {
2331		error = xfs_imap_lookup(mp, tp, pag, agino, agbno,
2332					&chunk_agbno, &offset_agbno, flags);
2333		if (error)
2334			goto out_drop;
2335		goto out_map;
2336	}
2337
2338	/*
2339	 * If the inode cluster size is the same as the blocksize or
2340	 * smaller we get to the buffer by simple arithmetics.
2341	 */
2342	if (M_IGEO(mp)->blocks_per_cluster == 1) {
2343		offset = XFS_INO_TO_OFFSET(mp, ino);
2344		ASSERT(offset < mp->m_sb.sb_inopblock);
2345
2346		imap->im_blkno = XFS_AGB_TO_DADDR(mp, pag->pag_agno, agbno);
2347		imap->im_len = XFS_FSB_TO_BB(mp, 1);
2348		imap->im_boffset = (unsigned short)(offset <<
2349							mp->m_sb.sb_inodelog);
2350		error = 0;
2351		goto out_drop;
2352	}
2353
2354	/*
2355	 * If the inode chunks are aligned then use simple maths to
2356	 * find the location. Otherwise we have to do a btree
2357	 * lookup to find the location.
2358	 */
2359	if (M_IGEO(mp)->inoalign_mask) {
2360		offset_agbno = agbno & M_IGEO(mp)->inoalign_mask;
2361		chunk_agbno = agbno - offset_agbno;
2362	} else {
2363		error = xfs_imap_lookup(mp, tp, pag, agino, agbno,
2364					&chunk_agbno, &offset_agbno, flags);
2365		if (error)
2366			goto out_drop;
2367	}
2368
2369out_map:
2370	ASSERT(agbno >= chunk_agbno);
2371	cluster_agbno = chunk_agbno +
2372		((offset_agbno / M_IGEO(mp)->blocks_per_cluster) *
2373		 M_IGEO(mp)->blocks_per_cluster);
2374	offset = ((agbno - cluster_agbno) * mp->m_sb.sb_inopblock) +
2375		XFS_INO_TO_OFFSET(mp, ino);
2376
2377	imap->im_blkno = XFS_AGB_TO_DADDR(mp, pag->pag_agno, cluster_agbno);
2378	imap->im_len = XFS_FSB_TO_BB(mp, M_IGEO(mp)->blocks_per_cluster);
2379	imap->im_boffset = (unsigned short)(offset << mp->m_sb.sb_inodelog);
2380
2381	/*
2382	 * If the inode number maps to a block outside the bounds
2383	 * of the file system then return NULL rather than calling
2384	 * read_buf and panicing when we get an error from the
2385	 * driver.
2386	 */
2387	if ((imap->im_blkno + imap->im_len) >
2388	    XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks)) {
2389		xfs_alert(mp,
2390	"%s: (im_blkno (0x%llx) + im_len (0x%llx)) > sb_dblocks (0x%llx)",
2391			__func__, (unsigned long long) imap->im_blkno,
2392			(unsigned long long) imap->im_len,
2393			XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks));
2394		error = -EINVAL;
2395		goto out_drop;
2396	}
2397	error = 0;
2398out_drop:
2399	if (pag)
2400		xfs_perag_put(pag);
2401	return error;
2402}
2403
2404/*
2405 * Log specified fields for the ag hdr (inode section). The growth of the agi
2406 * structure over time requires that we interpret the buffer as two logical
2407 * regions delineated by the end of the unlinked list. This is due to the size
2408 * of the hash table and its location in the middle of the agi.
2409 *
2410 * For example, a request to log a field before agi_unlinked and a field after
2411 * agi_unlinked could cause us to log the entire hash table and use an excessive
2412 * amount of log space. To avoid this behavior, log the region up through
2413 * agi_unlinked in one call and the region after agi_unlinked through the end of
2414 * the structure in another.
2415 */
2416void
2417xfs_ialloc_log_agi(
2418	xfs_trans_t	*tp,		/* transaction pointer */
2419	struct xfs_buf	*bp,		/* allocation group header buffer */
2420	int		fields)		/* bitmask of fields to log */
2421{
2422	int			first;		/* first byte number */
2423	int			last;		/* last byte number */
2424	static const short	offsets[] = {	/* field starting offsets */
2425					/* keep in sync with bit definitions */
2426		offsetof(xfs_agi_t, agi_magicnum),
2427		offsetof(xfs_agi_t, agi_versionnum),
2428		offsetof(xfs_agi_t, agi_seqno),
2429		offsetof(xfs_agi_t, agi_length),
2430		offsetof(xfs_agi_t, agi_count),
2431		offsetof(xfs_agi_t, agi_root),
2432		offsetof(xfs_agi_t, agi_level),
2433		offsetof(xfs_agi_t, agi_freecount),
2434		offsetof(xfs_agi_t, agi_newino),
2435		offsetof(xfs_agi_t, agi_dirino),
2436		offsetof(xfs_agi_t, agi_unlinked),
2437		offsetof(xfs_agi_t, agi_free_root),
2438		offsetof(xfs_agi_t, agi_free_level),
2439		offsetof(xfs_agi_t, agi_iblocks),
2440		sizeof(xfs_agi_t)
2441	};
2442#ifdef DEBUG
2443	struct xfs_agi		*agi = bp->b_addr;
2444
2445	ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
2446#endif
2447
2448	/*
2449	 * Compute byte offsets for the first and last fields in the first
2450	 * region and log the agi buffer. This only logs up through
2451	 * agi_unlinked.
2452	 */
2453	if (fields & XFS_AGI_ALL_BITS_R1) {
2454		xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R1,
2455				  &first, &last);
2456		xfs_trans_log_buf(tp, bp, first, last);
2457	}
2458
2459	/*
2460	 * Mask off the bits in the first region and calculate the first and
2461	 * last field offsets for any bits in the second region.
2462	 */
2463	fields &= ~XFS_AGI_ALL_BITS_R1;
2464	if (fields) {
2465		xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R2,
2466				  &first, &last);
2467		xfs_trans_log_buf(tp, bp, first, last);
2468	}
2469}
2470
2471static xfs_failaddr_t
2472xfs_agi_verify(
2473	struct xfs_buf	*bp)
2474{
2475	struct xfs_mount *mp = bp->b_mount;
2476	struct xfs_agi	*agi = bp->b_addr;
2477	int		i;
2478
2479	if (xfs_has_crc(mp)) {
2480		if (!uuid_equal(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid))
2481			return __this_address;
2482		if (!xfs_log_check_lsn(mp, be64_to_cpu(agi->agi_lsn)))
2483			return __this_address;
2484	}
2485
2486	/*
2487	 * Validate the magic number of the agi block.
2488	 */
2489	if (!xfs_verify_magic(bp, agi->agi_magicnum))
2490		return __this_address;
2491	if (!XFS_AGI_GOOD_VERSION(be32_to_cpu(agi->agi_versionnum)))
2492		return __this_address;
2493
2494	if (be32_to_cpu(agi->agi_level) < 1 ||
2495	    be32_to_cpu(agi->agi_level) > M_IGEO(mp)->inobt_maxlevels)
2496		return __this_address;
2497
2498	if (xfs_has_finobt(mp) &&
2499	    (be32_to_cpu(agi->agi_free_level) < 1 ||
2500	     be32_to_cpu(agi->agi_free_level) > M_IGEO(mp)->inobt_maxlevels))
2501		return __this_address;
2502
2503	/*
2504	 * during growfs operations, the perag is not fully initialised,
2505	 * so we can't use it for any useful checking. growfs ensures we can't
2506	 * use it by using uncached buffers that don't have the perag attached
2507	 * so we can detect and avoid this problem.
2508	 */
2509	if (bp->b_pag && be32_to_cpu(agi->agi_seqno) != bp->b_pag->pag_agno)
2510		return __this_address;
2511
2512	for (i = 0; i < XFS_AGI_UNLINKED_BUCKETS; i++) {
2513		if (agi->agi_unlinked[i] == cpu_to_be32(NULLAGINO))
2514			continue;
2515		if (!xfs_verify_ino(mp, be32_to_cpu(agi->agi_unlinked[i])))
2516			return __this_address;
2517	}
2518
2519	return NULL;
2520}
2521
2522static void
2523xfs_agi_read_verify(
2524	struct xfs_buf	*bp)
2525{
2526	struct xfs_mount *mp = bp->b_mount;
2527	xfs_failaddr_t	fa;
2528
2529	if (xfs_has_crc(mp) &&
2530	    !xfs_buf_verify_cksum(bp, XFS_AGI_CRC_OFF))
2531		xfs_verifier_error(bp, -EFSBADCRC, __this_address);
2532	else {
2533		fa = xfs_agi_verify(bp);
2534		if (XFS_TEST_ERROR(fa, mp, XFS_ERRTAG_IALLOC_READ_AGI))
2535			xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2536	}
2537}
2538
2539static void
2540xfs_agi_write_verify(
2541	struct xfs_buf	*bp)
2542{
2543	struct xfs_mount	*mp = bp->b_mount;
2544	struct xfs_buf_log_item	*bip = bp->b_log_item;
2545	struct xfs_agi		*agi = bp->b_addr;
2546	xfs_failaddr_t		fa;
2547
2548	fa = xfs_agi_verify(bp);
2549	if (fa) {
2550		xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2551		return;
2552	}
2553
2554	if (!xfs_has_crc(mp))
2555		return;
2556
2557	if (bip)
2558		agi->agi_lsn = cpu_to_be64(bip->bli_item.li_lsn);
2559	xfs_buf_update_cksum(bp, XFS_AGI_CRC_OFF);
2560}
2561
2562const struct xfs_buf_ops xfs_agi_buf_ops = {
2563	.name = "xfs_agi",
2564	.magic = { cpu_to_be32(XFS_AGI_MAGIC), cpu_to_be32(XFS_AGI_MAGIC) },
2565	.verify_read = xfs_agi_read_verify,
2566	.verify_write = xfs_agi_write_verify,
2567	.verify_struct = xfs_agi_verify,
2568};
2569
2570/*
2571 * Read in the allocation group header (inode allocation section)
2572 */
2573int
2574xfs_read_agi(
2575	struct xfs_mount	*mp,	/* file system mount structure */
2576	struct xfs_trans	*tp,	/* transaction pointer */
2577	xfs_agnumber_t		agno,	/* allocation group number */
2578	struct xfs_buf		**bpp)	/* allocation group hdr buf */
2579{
2580	int			error;
2581
2582	trace_xfs_read_agi(mp, agno);
2583
2584	ASSERT(agno != NULLAGNUMBER);
2585	error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
2586			XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
2587			XFS_FSS_TO_BB(mp, 1), 0, bpp, &xfs_agi_buf_ops);
2588	if (error)
2589		return error;
2590	if (tp)
2591		xfs_trans_buf_set_type(tp, *bpp, XFS_BLFT_AGI_BUF);
2592
2593	xfs_buf_set_ref(*bpp, XFS_AGI_REF);
2594	return 0;
2595}
2596
2597int
2598xfs_ialloc_read_agi(
2599	struct xfs_mount	*mp,	/* file system mount structure */
2600	struct xfs_trans	*tp,	/* transaction pointer */
2601	xfs_agnumber_t		agno,	/* allocation group number */
2602	struct xfs_buf		**bpp)	/* allocation group hdr buf */
2603{
2604	struct xfs_agi		*agi;	/* allocation group header */
2605	struct xfs_perag	*pag;	/* per allocation group data */
2606	int			error;
2607
2608	trace_xfs_ialloc_read_agi(mp, agno);
2609
2610	error = xfs_read_agi(mp, tp, agno, bpp);
2611	if (error)
2612		return error;
2613
2614	agi = (*bpp)->b_addr;
2615	pag = (*bpp)->b_pag;
2616	if (!pag->pagi_init) {
2617		pag->pagi_freecount = be32_to_cpu(agi->agi_freecount);
2618		pag->pagi_count = be32_to_cpu(agi->agi_count);
2619		pag->pagi_init = 1;
2620	}
2621
2622	/*
2623	 * It's possible for these to be out of sync if
2624	 * we are in the middle of a forced shutdown.
2625	 */
2626	ASSERT(pag->pagi_freecount == be32_to_cpu(agi->agi_freecount) ||
2627		xfs_is_shutdown(mp));
2628	return 0;
2629}
2630
2631/*
2632 * Read in the agi to initialise the per-ag data in the mount structure
2633 */
2634int
2635xfs_ialloc_pagi_init(
2636	xfs_mount_t	*mp,		/* file system mount structure */
2637	xfs_trans_t	*tp,		/* transaction pointer */
2638	xfs_agnumber_t	agno)		/* allocation group number */
2639{
2640	struct xfs_buf	*bp = NULL;
2641	int		error;
2642
2643	error = xfs_ialloc_read_agi(mp, tp, agno, &bp);
2644	if (error)
2645		return error;
2646	if (bp)
2647		xfs_trans_brelse(tp, bp);
2648	return 0;
2649}
2650
2651/* Is there an inode record covering a given range of inode numbers? */
2652int
2653xfs_ialloc_has_inode_record(
2654	struct xfs_btree_cur	*cur,
2655	xfs_agino_t		low,
2656	xfs_agino_t		high,
2657	bool			*exists)
2658{
2659	struct xfs_inobt_rec_incore	irec;
2660	xfs_agino_t		agino;
2661	uint16_t		holemask;
2662	int			has_record;
2663	int			i;
2664	int			error;
2665
2666	*exists = false;
2667	error = xfs_inobt_lookup(cur, low, XFS_LOOKUP_LE, &has_record);
2668	while (error == 0 && has_record) {
2669		error = xfs_inobt_get_rec(cur, &irec, &has_record);
2670		if (error || irec.ir_startino > high)
2671			break;
2672
2673		agino = irec.ir_startino;
2674		holemask = irec.ir_holemask;
2675		for (i = 0; i < XFS_INOBT_HOLEMASK_BITS; holemask >>= 1,
2676				i++, agino += XFS_INODES_PER_HOLEMASK_BIT) {
2677			if (holemask & 1)
2678				continue;
2679			if (agino + XFS_INODES_PER_HOLEMASK_BIT > low &&
2680					agino <= high) {
2681				*exists = true;
2682				return 0;
2683			}
2684		}
2685
2686		error = xfs_btree_increment(cur, 0, &has_record);
2687	}
2688	return error;
2689}
2690
2691/* Is there an inode record covering a given extent? */
2692int
2693xfs_ialloc_has_inodes_at_extent(
2694	struct xfs_btree_cur	*cur,
2695	xfs_agblock_t		bno,
2696	xfs_extlen_t		len,
2697	bool			*exists)
2698{
2699	xfs_agino_t		low;
2700	xfs_agino_t		high;
2701
2702	low = XFS_AGB_TO_AGINO(cur->bc_mp, bno);
2703	high = XFS_AGB_TO_AGINO(cur->bc_mp, bno + len) - 1;
2704
2705	return xfs_ialloc_has_inode_record(cur, low, high, exists);
2706}
2707
2708struct xfs_ialloc_count_inodes {
2709	xfs_agino_t			count;
2710	xfs_agino_t			freecount;
2711};
2712
2713/* Record inode counts across all inobt records. */
2714STATIC int
2715xfs_ialloc_count_inodes_rec(
2716	struct xfs_btree_cur		*cur,
2717	const union xfs_btree_rec	*rec,
2718	void				*priv)
2719{
2720	struct xfs_inobt_rec_incore	irec;
2721	struct xfs_ialloc_count_inodes	*ci = priv;
2722
2723	xfs_inobt_btrec_to_irec(cur->bc_mp, rec, &irec);
2724	ci->count += irec.ir_count;
2725	ci->freecount += irec.ir_freecount;
2726
2727	return 0;
2728}
2729
2730/* Count allocated and free inodes under an inobt. */
2731int
2732xfs_ialloc_count_inodes(
2733	struct xfs_btree_cur		*cur,
2734	xfs_agino_t			*count,
2735	xfs_agino_t			*freecount)
2736{
2737	struct xfs_ialloc_count_inodes	ci = {0};
2738	int				error;
2739
2740	ASSERT(cur->bc_btnum == XFS_BTNUM_INO);
2741	error = xfs_btree_query_all(cur, xfs_ialloc_count_inodes_rec, &ci);
2742	if (error)
2743		return error;
2744
2745	*count = ci.count;
2746	*freecount = ci.freecount;
2747	return 0;
2748}
2749
2750/*
2751 * Initialize inode-related geometry information.
2752 *
2753 * Compute the inode btree min and max levels and set maxicount.
2754 *
2755 * Set the inode cluster size.  This may still be overridden by the file
2756 * system block size if it is larger than the chosen cluster size.
2757 *
2758 * For v5 filesystems, scale the cluster size with the inode size to keep a
2759 * constant ratio of inode per cluster buffer, but only if mkfs has set the
2760 * inode alignment value appropriately for larger cluster sizes.
2761 *
2762 * Then compute the inode cluster alignment information.
2763 */
2764void
2765xfs_ialloc_setup_geometry(
2766	struct xfs_mount	*mp)
2767{
2768	struct xfs_sb		*sbp = &mp->m_sb;
2769	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
2770	uint64_t		icount;
2771	uint			inodes;
2772
2773	igeo->new_diflags2 = 0;
2774	if (xfs_has_bigtime(mp))
2775		igeo->new_diflags2 |= XFS_DIFLAG2_BIGTIME;
2776
2777	/* Compute inode btree geometry. */
2778	igeo->agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
2779	igeo->inobt_mxr[0] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, 1);
2780	igeo->inobt_mxr[1] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, 0);
2781	igeo->inobt_mnr[0] = igeo->inobt_mxr[0] / 2;
2782	igeo->inobt_mnr[1] = igeo->inobt_mxr[1] / 2;
2783
2784	igeo->ialloc_inos = max_t(uint16_t, XFS_INODES_PER_CHUNK,
2785			sbp->sb_inopblock);
2786	igeo->ialloc_blks = igeo->ialloc_inos >> sbp->sb_inopblog;
2787
2788	if (sbp->sb_spino_align)
2789		igeo->ialloc_min_blks = sbp->sb_spino_align;
2790	else
2791		igeo->ialloc_min_blks = igeo->ialloc_blks;
2792
2793	/* Compute and fill in value of m_ino_geo.inobt_maxlevels. */
2794	inodes = (1LL << XFS_INO_AGINO_BITS(mp)) >> XFS_INODES_PER_CHUNK_LOG;
2795	igeo->inobt_maxlevels = xfs_btree_compute_maxlevels(igeo->inobt_mnr,
2796			inodes);
2797
2798	/*
2799	 * Set the maximum inode count for this filesystem, being careful not
2800	 * to use obviously garbage sb_inopblog/sb_inopblock values.  Regular
2801	 * users should never get here due to failing sb verification, but
2802	 * certain users (xfs_db) need to be usable even with corrupt metadata.
2803	 */
2804	if (sbp->sb_imax_pct && igeo->ialloc_blks) {
2805		/*
2806		 * Make sure the maximum inode count is a multiple
2807		 * of the units we allocate inodes in.
2808		 */
2809		icount = sbp->sb_dblocks * sbp->sb_imax_pct;
2810		do_div(icount, 100);
2811		do_div(icount, igeo->ialloc_blks);
2812		igeo->maxicount = XFS_FSB_TO_INO(mp,
2813				icount * igeo->ialloc_blks);
2814	} else {
2815		igeo->maxicount = 0;
2816	}
2817
2818	/*
2819	 * Compute the desired size of an inode cluster buffer size, which
2820	 * starts at 8K and (on v5 filesystems) scales up with larger inode
2821	 * sizes.
2822	 *
2823	 * Preserve the desired inode cluster size because the sparse inodes
2824	 * feature uses that desired size (not the actual size) to compute the
2825	 * sparse inode alignment.  The mount code validates this value, so we
2826	 * cannot change the behavior.
2827	 */
2828	igeo->inode_cluster_size_raw = XFS_INODE_BIG_CLUSTER_SIZE;
2829	if (xfs_has_v3inodes(mp)) {
2830		int	new_size = igeo->inode_cluster_size_raw;
2831
2832		new_size *= mp->m_sb.sb_inodesize / XFS_DINODE_MIN_SIZE;
2833		if (mp->m_sb.sb_inoalignmt >= XFS_B_TO_FSBT(mp, new_size))
2834			igeo->inode_cluster_size_raw = new_size;
2835	}
2836
2837	/* Calculate inode cluster ratios. */
2838	if (igeo->inode_cluster_size_raw > mp->m_sb.sb_blocksize)
2839		igeo->blocks_per_cluster = XFS_B_TO_FSBT(mp,
2840				igeo->inode_cluster_size_raw);
2841	else
2842		igeo->blocks_per_cluster = 1;
2843	igeo->inode_cluster_size = XFS_FSB_TO_B(mp, igeo->blocks_per_cluster);
2844	igeo->inodes_per_cluster = XFS_FSB_TO_INO(mp, igeo->blocks_per_cluster);
2845
2846	/* Calculate inode cluster alignment. */
2847	if (xfs_has_align(mp) &&
2848	    mp->m_sb.sb_inoalignmt >= igeo->blocks_per_cluster)
2849		igeo->cluster_align = mp->m_sb.sb_inoalignmt;
2850	else
2851		igeo->cluster_align = 1;
2852	igeo->inoalign_mask = igeo->cluster_align - 1;
2853	igeo->cluster_align_inodes = XFS_FSB_TO_INO(mp, igeo->cluster_align);
2854
2855	/*
2856	 * If we are using stripe alignment, check whether
2857	 * the stripe unit is a multiple of the inode alignment
2858	 */
2859	if (mp->m_dalign && igeo->inoalign_mask &&
2860	    !(mp->m_dalign & igeo->inoalign_mask))
2861		igeo->ialloc_align = mp->m_dalign;
2862	else
2863		igeo->ialloc_align = 0;
2864}
2865
2866/* Compute the location of the root directory inode that is laid out by mkfs. */
2867xfs_ino_t
2868xfs_ialloc_calc_rootino(
2869	struct xfs_mount	*mp,
2870	int			sunit)
2871{
2872	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
2873	xfs_agblock_t		first_bno;
2874
2875	/*
2876	 * Pre-calculate the geometry of AG 0.  We know what it looks like
2877	 * because libxfs knows how to create allocation groups now.
2878	 *
2879	 * first_bno is the first block in which mkfs could possibly have
2880	 * allocated the root directory inode, once we factor in the metadata
2881	 * that mkfs formats before it.  Namely, the four AG headers...
2882	 */
2883	first_bno = howmany(4 * mp->m_sb.sb_sectsize, mp->m_sb.sb_blocksize);
2884
2885	/* ...the two free space btree roots... */
2886	first_bno += 2;
2887
2888	/* ...the inode btree root... */
2889	first_bno += 1;
2890
2891	/* ...the initial AGFL... */
2892	first_bno += xfs_alloc_min_freelist(mp, NULL);
2893
2894	/* ...the free inode btree root... */
2895	if (xfs_has_finobt(mp))
2896		first_bno++;
2897
2898	/* ...the reverse mapping btree root... */
2899	if (xfs_has_rmapbt(mp))
2900		first_bno++;
2901
2902	/* ...the reference count btree... */
2903	if (xfs_has_reflink(mp))
2904		first_bno++;
2905
2906	/*
2907	 * ...and the log, if it is allocated in the first allocation group.
2908	 *
2909	 * This can happen with filesystems that only have a single
2910	 * allocation group, or very odd geometries created by old mkfs
2911	 * versions on very small filesystems.
2912	 */
2913	if (mp->m_sb.sb_logstart &&
2914	    XFS_FSB_TO_AGNO(mp, mp->m_sb.sb_logstart) == 0)
2915		 first_bno += mp->m_sb.sb_logblocks;
2916
2917	/*
2918	 * Now round first_bno up to whatever allocation alignment is given
2919	 * by the filesystem or was passed in.
2920	 */
2921	if (xfs_has_dalign(mp) && igeo->ialloc_align > 0)
2922		first_bno = roundup(first_bno, sunit);
2923	else if (xfs_has_align(mp) &&
2924			mp->m_sb.sb_inoalignmt > 1)
2925		first_bno = roundup(first_bno, mp->m_sb.sb_inoalignmt);
2926
2927	return XFS_AGINO_TO_INO(mp, 0, XFS_AGB_TO_AGINO(mp, first_bno));
2928}
2929
2930/*
2931 * Ensure there are not sparse inode clusters that cross the new EOAG.
2932 *
2933 * This is a no-op for non-spinode filesystems since clusters are always fully
2934 * allocated and checking the bnobt suffices.  However, a spinode filesystem
2935 * could have a record where the upper inodes are free blocks.  If those blocks
2936 * were removed from the filesystem, the inode record would extend beyond EOAG,
2937 * which will be flagged as corruption.
2938 */
2939int
2940xfs_ialloc_check_shrink(
2941	struct xfs_trans	*tp,
2942	xfs_agnumber_t		agno,
2943	struct xfs_buf		*agibp,
2944	xfs_agblock_t		new_length)
2945{
2946	struct xfs_inobt_rec_incore rec;
2947	struct xfs_btree_cur	*cur;
2948	struct xfs_mount	*mp = tp->t_mountp;
2949	struct xfs_perag	*pag;
2950	xfs_agino_t		agino = XFS_AGB_TO_AGINO(mp, new_length);
2951	int			has;
2952	int			error;
2953
2954	if (!xfs_has_sparseinodes(mp))
2955		return 0;
2956
2957	pag = xfs_perag_get(mp, agno);
2958	cur = xfs_inobt_init_cursor(mp, tp, agibp, pag, XFS_BTNUM_INO);
2959
2960	/* Look up the inobt record that would correspond to the new EOFS. */
2961	error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &has);
2962	if (error || !has)
2963		goto out;
2964
2965	error = xfs_inobt_get_rec(cur, &rec, &has);
2966	if (error)
2967		goto out;
2968
2969	if (!has) {
2970		error = -EFSCORRUPTED;
2971		goto out;
2972	}
2973
2974	/* If the record covers inodes that would be beyond EOFS, bail out. */
2975	if (rec.ir_startino + XFS_INODES_PER_CHUNK > agino) {
2976		error = -ENOSPC;
2977		goto out;
2978	}
2979out:
2980	xfs_btree_del_cursor(cur, error);
2981	xfs_perag_put(pag);
2982	return error;
2983}
2984