Loading...
1/*
2 * Copyright (c) 2000-2002,2005 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
17 */
18#include "xfs.h"
19#include "xfs_fs.h"
20#include "xfs_shared.h"
21#include "xfs_format.h"
22#include "xfs_log_format.h"
23#include "xfs_trans_resv.h"
24#include "xfs_bit.h"
25#include "xfs_sb.h"
26#include "xfs_mount.h"
27#include "xfs_defer.h"
28#include "xfs_inode.h"
29#include "xfs_btree.h"
30#include "xfs_ialloc.h"
31#include "xfs_ialloc_btree.h"
32#include "xfs_alloc.h"
33#include "xfs_rtalloc.h"
34#include "xfs_error.h"
35#include "xfs_bmap.h"
36#include "xfs_cksum.h"
37#include "xfs_trans.h"
38#include "xfs_buf_item.h"
39#include "xfs_icreate_item.h"
40#include "xfs_icache.h"
41#include "xfs_trace.h"
42#include "xfs_log.h"
43#include "xfs_rmap.h"
44
45
46/*
47 * Allocation group level functions.
48 */
49static inline int
50xfs_ialloc_cluster_alignment(
51 struct xfs_mount *mp)
52{
53 if (xfs_sb_version_hasalign(&mp->m_sb) &&
54 mp->m_sb.sb_inoalignmt >= xfs_icluster_size_fsb(mp))
55 return mp->m_sb.sb_inoalignmt;
56 return 1;
57}
58
59/*
60 * Lookup a record by ino in the btree given by cur.
61 */
62int /* error */
63xfs_inobt_lookup(
64 struct xfs_btree_cur *cur, /* btree cursor */
65 xfs_agino_t ino, /* starting inode of chunk */
66 xfs_lookup_t dir, /* <=, >=, == */
67 int *stat) /* success/failure */
68{
69 cur->bc_rec.i.ir_startino = ino;
70 cur->bc_rec.i.ir_holemask = 0;
71 cur->bc_rec.i.ir_count = 0;
72 cur->bc_rec.i.ir_freecount = 0;
73 cur->bc_rec.i.ir_free = 0;
74 return xfs_btree_lookup(cur, dir, stat);
75}
76
77/*
78 * Update the record referred to by cur to the value given.
79 * This either works (return 0) or gets an EFSCORRUPTED error.
80 */
81STATIC int /* error */
82xfs_inobt_update(
83 struct xfs_btree_cur *cur, /* btree cursor */
84 xfs_inobt_rec_incore_t *irec) /* btree record */
85{
86 union xfs_btree_rec rec;
87
88 rec.inobt.ir_startino = cpu_to_be32(irec->ir_startino);
89 if (xfs_sb_version_hassparseinodes(&cur->bc_mp->m_sb)) {
90 rec.inobt.ir_u.sp.ir_holemask = cpu_to_be16(irec->ir_holemask);
91 rec.inobt.ir_u.sp.ir_count = irec->ir_count;
92 rec.inobt.ir_u.sp.ir_freecount = irec->ir_freecount;
93 } else {
94 /* ir_holemask/ir_count not supported on-disk */
95 rec.inobt.ir_u.f.ir_freecount = cpu_to_be32(irec->ir_freecount);
96 }
97 rec.inobt.ir_free = cpu_to_be64(irec->ir_free);
98 return xfs_btree_update(cur, &rec);
99}
100
101/*
102 * Get the data from the pointed-to record.
103 */
104int /* error */
105xfs_inobt_get_rec(
106 struct xfs_btree_cur *cur, /* btree cursor */
107 xfs_inobt_rec_incore_t *irec, /* btree record */
108 int *stat) /* output: success/failure */
109{
110 union xfs_btree_rec *rec;
111 int error;
112
113 error = xfs_btree_get_rec(cur, &rec, stat);
114 if (error || *stat == 0)
115 return error;
116
117 irec->ir_startino = be32_to_cpu(rec->inobt.ir_startino);
118 if (xfs_sb_version_hassparseinodes(&cur->bc_mp->m_sb)) {
119 irec->ir_holemask = be16_to_cpu(rec->inobt.ir_u.sp.ir_holemask);
120 irec->ir_count = rec->inobt.ir_u.sp.ir_count;
121 irec->ir_freecount = rec->inobt.ir_u.sp.ir_freecount;
122 } else {
123 /*
124 * ir_holemask/ir_count not supported on-disk. Fill in hardcoded
125 * values for full inode chunks.
126 */
127 irec->ir_holemask = XFS_INOBT_HOLEMASK_FULL;
128 irec->ir_count = XFS_INODES_PER_CHUNK;
129 irec->ir_freecount =
130 be32_to_cpu(rec->inobt.ir_u.f.ir_freecount);
131 }
132 irec->ir_free = be64_to_cpu(rec->inobt.ir_free);
133
134 return 0;
135}
136
137/*
138 * Insert a single inobt record. Cursor must already point to desired location.
139 */
140STATIC int
141xfs_inobt_insert_rec(
142 struct xfs_btree_cur *cur,
143 __uint16_t holemask,
144 __uint8_t count,
145 __int32_t freecount,
146 xfs_inofree_t free,
147 int *stat)
148{
149 cur->bc_rec.i.ir_holemask = holemask;
150 cur->bc_rec.i.ir_count = count;
151 cur->bc_rec.i.ir_freecount = freecount;
152 cur->bc_rec.i.ir_free = free;
153 return xfs_btree_insert(cur, stat);
154}
155
156/*
157 * Insert records describing a newly allocated inode chunk into the inobt.
158 */
159STATIC int
160xfs_inobt_insert(
161 struct xfs_mount *mp,
162 struct xfs_trans *tp,
163 struct xfs_buf *agbp,
164 xfs_agino_t newino,
165 xfs_agino_t newlen,
166 xfs_btnum_t btnum)
167{
168 struct xfs_btree_cur *cur;
169 struct xfs_agi *agi = XFS_BUF_TO_AGI(agbp);
170 xfs_agnumber_t agno = be32_to_cpu(agi->agi_seqno);
171 xfs_agino_t thisino;
172 int i;
173 int error;
174
175 cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, btnum);
176
177 for (thisino = newino;
178 thisino < newino + newlen;
179 thisino += XFS_INODES_PER_CHUNK) {
180 error = xfs_inobt_lookup(cur, thisino, XFS_LOOKUP_EQ, &i);
181 if (error) {
182 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
183 return error;
184 }
185 ASSERT(i == 0);
186
187 error = xfs_inobt_insert_rec(cur, XFS_INOBT_HOLEMASK_FULL,
188 XFS_INODES_PER_CHUNK,
189 XFS_INODES_PER_CHUNK,
190 XFS_INOBT_ALL_FREE, &i);
191 if (error) {
192 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
193 return error;
194 }
195 ASSERT(i == 1);
196 }
197
198 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
199
200 return 0;
201}
202
203/*
204 * Verify that the number of free inodes in the AGI is correct.
205 */
206#ifdef DEBUG
207STATIC int
208xfs_check_agi_freecount(
209 struct xfs_btree_cur *cur,
210 struct xfs_agi *agi)
211{
212 if (cur->bc_nlevels == 1) {
213 xfs_inobt_rec_incore_t rec;
214 int freecount = 0;
215 int error;
216 int i;
217
218 error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
219 if (error)
220 return error;
221
222 do {
223 error = xfs_inobt_get_rec(cur, &rec, &i);
224 if (error)
225 return error;
226
227 if (i) {
228 freecount += rec.ir_freecount;
229 error = xfs_btree_increment(cur, 0, &i);
230 if (error)
231 return error;
232 }
233 } while (i == 1);
234
235 if (!XFS_FORCED_SHUTDOWN(cur->bc_mp))
236 ASSERT(freecount == be32_to_cpu(agi->agi_freecount));
237 }
238 return 0;
239}
240#else
241#define xfs_check_agi_freecount(cur, agi) 0
242#endif
243
244/*
245 * Initialise a new set of inodes. When called without a transaction context
246 * (e.g. from recovery) we initiate a delayed write of the inode buffers rather
247 * than logging them (which in a transaction context puts them into the AIL
248 * for writeback rather than the xfsbufd queue).
249 */
250int
251xfs_ialloc_inode_init(
252 struct xfs_mount *mp,
253 struct xfs_trans *tp,
254 struct list_head *buffer_list,
255 int icount,
256 xfs_agnumber_t agno,
257 xfs_agblock_t agbno,
258 xfs_agblock_t length,
259 unsigned int gen)
260{
261 struct xfs_buf *fbuf;
262 struct xfs_dinode *free;
263 int nbufs, blks_per_cluster, inodes_per_cluster;
264 int version;
265 int i, j;
266 xfs_daddr_t d;
267 xfs_ino_t ino = 0;
268
269 /*
270 * Loop over the new block(s), filling in the inodes. For small block
271 * sizes, manipulate the inodes in buffers which are multiples of the
272 * blocks size.
273 */
274 blks_per_cluster = xfs_icluster_size_fsb(mp);
275 inodes_per_cluster = blks_per_cluster << mp->m_sb.sb_inopblog;
276 nbufs = length / blks_per_cluster;
277
278 /*
279 * Figure out what version number to use in the inodes we create. If
280 * the superblock version has caught up to the one that supports the new
281 * inode format, then use the new inode version. Otherwise use the old
282 * version so that old kernels will continue to be able to use the file
283 * system.
284 *
285 * For v3 inodes, we also need to write the inode number into the inode,
286 * so calculate the first inode number of the chunk here as
287 * XFS_OFFBNO_TO_AGINO() only works within a filesystem block, not
288 * across multiple filesystem blocks (such as a cluster) and so cannot
289 * be used in the cluster buffer loop below.
290 *
291 * Further, because we are writing the inode directly into the buffer
292 * and calculating a CRC on the entire inode, we have ot log the entire
293 * inode so that the entire range the CRC covers is present in the log.
294 * That means for v3 inode we log the entire buffer rather than just the
295 * inode cores.
296 */
297 if (xfs_sb_version_hascrc(&mp->m_sb)) {
298 version = 3;
299 ino = XFS_AGINO_TO_INO(mp, agno,
300 XFS_OFFBNO_TO_AGINO(mp, agbno, 0));
301
302 /*
303 * log the initialisation that is about to take place as an
304 * logical operation. This means the transaction does not
305 * need to log the physical changes to the inode buffers as log
306 * recovery will know what initialisation is actually needed.
307 * Hence we only need to log the buffers as "ordered" buffers so
308 * they track in the AIL as if they were physically logged.
309 */
310 if (tp)
311 xfs_icreate_log(tp, agno, agbno, icount,
312 mp->m_sb.sb_inodesize, length, gen);
313 } else
314 version = 2;
315
316 for (j = 0; j < nbufs; j++) {
317 /*
318 * Get the block.
319 */
320 d = XFS_AGB_TO_DADDR(mp, agno, agbno + (j * blks_per_cluster));
321 fbuf = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
322 mp->m_bsize * blks_per_cluster,
323 XBF_UNMAPPED);
324 if (!fbuf)
325 return -ENOMEM;
326
327 /* Initialize the inode buffers and log them appropriately. */
328 fbuf->b_ops = &xfs_inode_buf_ops;
329 xfs_buf_zero(fbuf, 0, BBTOB(fbuf->b_length));
330 for (i = 0; i < inodes_per_cluster; i++) {
331 int ioffset = i << mp->m_sb.sb_inodelog;
332 uint isize = xfs_dinode_size(version);
333
334 free = xfs_make_iptr(mp, fbuf, i);
335 free->di_magic = cpu_to_be16(XFS_DINODE_MAGIC);
336 free->di_version = version;
337 free->di_gen = cpu_to_be32(gen);
338 free->di_next_unlinked = cpu_to_be32(NULLAGINO);
339
340 if (version == 3) {
341 free->di_ino = cpu_to_be64(ino);
342 ino++;
343 uuid_copy(&free->di_uuid,
344 &mp->m_sb.sb_meta_uuid);
345 xfs_dinode_calc_crc(mp, free);
346 } else if (tp) {
347 /* just log the inode core */
348 xfs_trans_log_buf(tp, fbuf, ioffset,
349 ioffset + isize - 1);
350 }
351 }
352
353 if (tp) {
354 /*
355 * Mark the buffer as an inode allocation buffer so it
356 * sticks in AIL at the point of this allocation
357 * transaction. This ensures the they are on disk before
358 * the tail of the log can be moved past this
359 * transaction (i.e. by preventing relogging from moving
360 * it forward in the log).
361 */
362 xfs_trans_inode_alloc_buf(tp, fbuf);
363 if (version == 3) {
364 /*
365 * Mark the buffer as ordered so that they are
366 * not physically logged in the transaction but
367 * still tracked in the AIL as part of the
368 * transaction and pin the log appropriately.
369 */
370 xfs_trans_ordered_buf(tp, fbuf);
371 xfs_trans_log_buf(tp, fbuf, 0,
372 BBTOB(fbuf->b_length) - 1);
373 }
374 } else {
375 fbuf->b_flags |= XBF_DONE;
376 xfs_buf_delwri_queue(fbuf, buffer_list);
377 xfs_buf_relse(fbuf);
378 }
379 }
380 return 0;
381}
382
383/*
384 * Align startino and allocmask for a recently allocated sparse chunk such that
385 * they are fit for insertion (or merge) into the on-disk inode btrees.
386 *
387 * Background:
388 *
389 * When enabled, sparse inode support increases the inode alignment from cluster
390 * size to inode chunk size. This means that the minimum range between two
391 * non-adjacent inode records in the inobt is large enough for a full inode
392 * record. This allows for cluster sized, cluster aligned block allocation
393 * without need to worry about whether the resulting inode record overlaps with
394 * another record in the tree. Without this basic rule, we would have to deal
395 * with the consequences of overlap by potentially undoing recent allocations in
396 * the inode allocation codepath.
397 *
398 * Because of this alignment rule (which is enforced on mount), there are two
399 * inobt possibilities for newly allocated sparse chunks. One is that the
400 * aligned inode record for the chunk covers a range of inodes not already
401 * covered in the inobt (i.e., it is safe to insert a new sparse record). The
402 * other is that a record already exists at the aligned startino that considers
403 * the newly allocated range as sparse. In the latter case, record content is
404 * merged in hope that sparse inode chunks fill to full chunks over time.
405 */
406STATIC void
407xfs_align_sparse_ino(
408 struct xfs_mount *mp,
409 xfs_agino_t *startino,
410 uint16_t *allocmask)
411{
412 xfs_agblock_t agbno;
413 xfs_agblock_t mod;
414 int offset;
415
416 agbno = XFS_AGINO_TO_AGBNO(mp, *startino);
417 mod = agbno % mp->m_sb.sb_inoalignmt;
418 if (!mod)
419 return;
420
421 /* calculate the inode offset and align startino */
422 offset = mod << mp->m_sb.sb_inopblog;
423 *startino -= offset;
424
425 /*
426 * Since startino has been aligned down, left shift allocmask such that
427 * it continues to represent the same physical inodes relative to the
428 * new startino.
429 */
430 *allocmask <<= offset / XFS_INODES_PER_HOLEMASK_BIT;
431}
432
433/*
434 * Determine whether the source inode record can merge into the target. Both
435 * records must be sparse, the inode ranges must match and there must be no
436 * allocation overlap between the records.
437 */
438STATIC bool
439__xfs_inobt_can_merge(
440 struct xfs_inobt_rec_incore *trec, /* tgt record */
441 struct xfs_inobt_rec_incore *srec) /* src record */
442{
443 uint64_t talloc;
444 uint64_t salloc;
445
446 /* records must cover the same inode range */
447 if (trec->ir_startino != srec->ir_startino)
448 return false;
449
450 /* both records must be sparse */
451 if (!xfs_inobt_issparse(trec->ir_holemask) ||
452 !xfs_inobt_issparse(srec->ir_holemask))
453 return false;
454
455 /* both records must track some inodes */
456 if (!trec->ir_count || !srec->ir_count)
457 return false;
458
459 /* can't exceed capacity of a full record */
460 if (trec->ir_count + srec->ir_count > XFS_INODES_PER_CHUNK)
461 return false;
462
463 /* verify there is no allocation overlap */
464 talloc = xfs_inobt_irec_to_allocmask(trec);
465 salloc = xfs_inobt_irec_to_allocmask(srec);
466 if (talloc & salloc)
467 return false;
468
469 return true;
470}
471
472/*
473 * Merge the source inode record into the target. The caller must call
474 * __xfs_inobt_can_merge() to ensure the merge is valid.
475 */
476STATIC void
477__xfs_inobt_rec_merge(
478 struct xfs_inobt_rec_incore *trec, /* target */
479 struct xfs_inobt_rec_incore *srec) /* src */
480{
481 ASSERT(trec->ir_startino == srec->ir_startino);
482
483 /* combine the counts */
484 trec->ir_count += srec->ir_count;
485 trec->ir_freecount += srec->ir_freecount;
486
487 /*
488 * Merge the holemask and free mask. For both fields, 0 bits refer to
489 * allocated inodes. We combine the allocated ranges with bitwise AND.
490 */
491 trec->ir_holemask &= srec->ir_holemask;
492 trec->ir_free &= srec->ir_free;
493}
494
495/*
496 * Insert a new sparse inode chunk into the associated inode btree. The inode
497 * record for the sparse chunk is pre-aligned to a startino that should match
498 * any pre-existing sparse inode record in the tree. This allows sparse chunks
499 * to fill over time.
500 *
501 * This function supports two modes of handling preexisting records depending on
502 * the merge flag. If merge is true, the provided record is merged with the
503 * existing record and updated in place. The merged record is returned in nrec.
504 * If merge is false, an existing record is replaced with the provided record.
505 * If no preexisting record exists, the provided record is always inserted.
506 *
507 * It is considered corruption if a merge is requested and not possible. Given
508 * the sparse inode alignment constraints, this should never happen.
509 */
510STATIC int
511xfs_inobt_insert_sprec(
512 struct xfs_mount *mp,
513 struct xfs_trans *tp,
514 struct xfs_buf *agbp,
515 int btnum,
516 struct xfs_inobt_rec_incore *nrec, /* in/out: new/merged rec. */
517 bool merge) /* merge or replace */
518{
519 struct xfs_btree_cur *cur;
520 struct xfs_agi *agi = XFS_BUF_TO_AGI(agbp);
521 xfs_agnumber_t agno = be32_to_cpu(agi->agi_seqno);
522 int error;
523 int i;
524 struct xfs_inobt_rec_incore rec;
525
526 cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, btnum);
527
528 /* the new record is pre-aligned so we know where to look */
529 error = xfs_inobt_lookup(cur, nrec->ir_startino, XFS_LOOKUP_EQ, &i);
530 if (error)
531 goto error;
532 /* if nothing there, insert a new record and return */
533 if (i == 0) {
534 error = xfs_inobt_insert_rec(cur, nrec->ir_holemask,
535 nrec->ir_count, nrec->ir_freecount,
536 nrec->ir_free, &i);
537 if (error)
538 goto error;
539 XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error);
540
541 goto out;
542 }
543
544 /*
545 * A record exists at this startino. Merge or replace the record
546 * depending on what we've been asked to do.
547 */
548 if (merge) {
549 error = xfs_inobt_get_rec(cur, &rec, &i);
550 if (error)
551 goto error;
552 XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error);
553 XFS_WANT_CORRUPTED_GOTO(mp,
554 rec.ir_startino == nrec->ir_startino,
555 error);
556
557 /*
558 * This should never fail. If we have coexisting records that
559 * cannot merge, something is seriously wrong.
560 */
561 XFS_WANT_CORRUPTED_GOTO(mp, __xfs_inobt_can_merge(nrec, &rec),
562 error);
563
564 trace_xfs_irec_merge_pre(mp, agno, rec.ir_startino,
565 rec.ir_holemask, nrec->ir_startino,
566 nrec->ir_holemask);
567
568 /* merge to nrec to output the updated record */
569 __xfs_inobt_rec_merge(nrec, &rec);
570
571 trace_xfs_irec_merge_post(mp, agno, nrec->ir_startino,
572 nrec->ir_holemask);
573
574 error = xfs_inobt_rec_check_count(mp, nrec);
575 if (error)
576 goto error;
577 }
578
579 error = xfs_inobt_update(cur, nrec);
580 if (error)
581 goto error;
582
583out:
584 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
585 return 0;
586error:
587 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
588 return error;
589}
590
591/*
592 * Allocate new inodes in the allocation group specified by agbp.
593 * Return 0 for success, else error code.
594 */
595STATIC int /* error code or 0 */
596xfs_ialloc_ag_alloc(
597 xfs_trans_t *tp, /* transaction pointer */
598 xfs_buf_t *agbp, /* alloc group buffer */
599 int *alloc)
600{
601 xfs_agi_t *agi; /* allocation group header */
602 xfs_alloc_arg_t args; /* allocation argument structure */
603 xfs_agnumber_t agno;
604 int error;
605 xfs_agino_t newino; /* new first inode's number */
606 xfs_agino_t newlen; /* new number of inodes */
607 int isaligned = 0; /* inode allocation at stripe unit */
608 /* boundary */
609 uint16_t allocmask = (uint16_t) -1; /* init. to full chunk */
610 struct xfs_inobt_rec_incore rec;
611 struct xfs_perag *pag;
612 int do_sparse = 0;
613
614 memset(&args, 0, sizeof(args));
615 args.tp = tp;
616 args.mp = tp->t_mountp;
617 args.fsbno = NULLFSBLOCK;
618 xfs_rmap_ag_owner(&args.oinfo, XFS_RMAP_OWN_INODES);
619
620#ifdef DEBUG
621 /* randomly do sparse inode allocations */
622 if (xfs_sb_version_hassparseinodes(&tp->t_mountp->m_sb) &&
623 args.mp->m_ialloc_min_blks < args.mp->m_ialloc_blks)
624 do_sparse = prandom_u32() & 1;
625#endif
626
627 /*
628 * Locking will ensure that we don't have two callers in here
629 * at one time.
630 */
631 newlen = args.mp->m_ialloc_inos;
632 if (args.mp->m_maxicount &&
633 percpu_counter_read_positive(&args.mp->m_icount) + newlen >
634 args.mp->m_maxicount)
635 return -ENOSPC;
636 args.minlen = args.maxlen = args.mp->m_ialloc_blks;
637 /*
638 * First try to allocate inodes contiguous with the last-allocated
639 * chunk of inodes. If the filesystem is striped, this will fill
640 * an entire stripe unit with inodes.
641 */
642 agi = XFS_BUF_TO_AGI(agbp);
643 newino = be32_to_cpu(agi->agi_newino);
644 agno = be32_to_cpu(agi->agi_seqno);
645 args.agbno = XFS_AGINO_TO_AGBNO(args.mp, newino) +
646 args.mp->m_ialloc_blks;
647 if (do_sparse)
648 goto sparse_alloc;
649 if (likely(newino != NULLAGINO &&
650 (args.agbno < be32_to_cpu(agi->agi_length)))) {
651 args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
652 args.type = XFS_ALLOCTYPE_THIS_BNO;
653 args.prod = 1;
654
655 /*
656 * We need to take into account alignment here to ensure that
657 * we don't modify the free list if we fail to have an exact
658 * block. If we don't have an exact match, and every oher
659 * attempt allocation attempt fails, we'll end up cancelling
660 * a dirty transaction and shutting down.
661 *
662 * For an exact allocation, alignment must be 1,
663 * however we need to take cluster alignment into account when
664 * fixing up the freelist. Use the minalignslop field to
665 * indicate that extra blocks might be required for alignment,
666 * but not to use them in the actual exact allocation.
667 */
668 args.alignment = 1;
669 args.minalignslop = xfs_ialloc_cluster_alignment(args.mp) - 1;
670
671 /* Allow space for the inode btree to split. */
672 args.minleft = args.mp->m_in_maxlevels - 1;
673 if ((error = xfs_alloc_vextent(&args)))
674 return error;
675
676 /*
677 * This request might have dirtied the transaction if the AG can
678 * satisfy the request, but the exact block was not available.
679 * If the allocation did fail, subsequent requests will relax
680 * the exact agbno requirement and increase the alignment
681 * instead. It is critical that the total size of the request
682 * (len + alignment + slop) does not increase from this point
683 * on, so reset minalignslop to ensure it is not included in
684 * subsequent requests.
685 */
686 args.minalignslop = 0;
687 }
688
689 if (unlikely(args.fsbno == NULLFSBLOCK)) {
690 /*
691 * Set the alignment for the allocation.
692 * If stripe alignment is turned on then align at stripe unit
693 * boundary.
694 * If the cluster size is smaller than a filesystem block
695 * then we're doing I/O for inodes in filesystem block size
696 * pieces, so don't need alignment anyway.
697 */
698 isaligned = 0;
699 if (args.mp->m_sinoalign) {
700 ASSERT(!(args.mp->m_flags & XFS_MOUNT_NOALIGN));
701 args.alignment = args.mp->m_dalign;
702 isaligned = 1;
703 } else
704 args.alignment = xfs_ialloc_cluster_alignment(args.mp);
705 /*
706 * Need to figure out where to allocate the inode blocks.
707 * Ideally they should be spaced out through the a.g.
708 * For now, just allocate blocks up front.
709 */
710 args.agbno = be32_to_cpu(agi->agi_root);
711 args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
712 /*
713 * Allocate a fixed-size extent of inodes.
714 */
715 args.type = XFS_ALLOCTYPE_NEAR_BNO;
716 args.prod = 1;
717 /*
718 * Allow space for the inode btree to split.
719 */
720 args.minleft = args.mp->m_in_maxlevels - 1;
721 if ((error = xfs_alloc_vextent(&args)))
722 return error;
723 }
724
725 /*
726 * If stripe alignment is turned on, then try again with cluster
727 * alignment.
728 */
729 if (isaligned && args.fsbno == NULLFSBLOCK) {
730 args.type = XFS_ALLOCTYPE_NEAR_BNO;
731 args.agbno = be32_to_cpu(agi->agi_root);
732 args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
733 args.alignment = xfs_ialloc_cluster_alignment(args.mp);
734 if ((error = xfs_alloc_vextent(&args)))
735 return error;
736 }
737
738 /*
739 * Finally, try a sparse allocation if the filesystem supports it and
740 * the sparse allocation length is smaller than a full chunk.
741 */
742 if (xfs_sb_version_hassparseinodes(&args.mp->m_sb) &&
743 args.mp->m_ialloc_min_blks < args.mp->m_ialloc_blks &&
744 args.fsbno == NULLFSBLOCK) {
745sparse_alloc:
746 args.type = XFS_ALLOCTYPE_NEAR_BNO;
747 args.agbno = be32_to_cpu(agi->agi_root);
748 args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
749 args.alignment = args.mp->m_sb.sb_spino_align;
750 args.prod = 1;
751
752 args.minlen = args.mp->m_ialloc_min_blks;
753 args.maxlen = args.minlen;
754
755 /*
756 * The inode record will be aligned to full chunk size. We must
757 * prevent sparse allocation from AG boundaries that result in
758 * invalid inode records, such as records that start at agbno 0
759 * or extend beyond the AG.
760 *
761 * Set min agbno to the first aligned, non-zero agbno and max to
762 * the last aligned agbno that is at least one full chunk from
763 * the end of the AG.
764 */
765 args.min_agbno = args.mp->m_sb.sb_inoalignmt;
766 args.max_agbno = round_down(args.mp->m_sb.sb_agblocks,
767 args.mp->m_sb.sb_inoalignmt) -
768 args.mp->m_ialloc_blks;
769
770 error = xfs_alloc_vextent(&args);
771 if (error)
772 return error;
773
774 newlen = args.len << args.mp->m_sb.sb_inopblog;
775 ASSERT(newlen <= XFS_INODES_PER_CHUNK);
776 allocmask = (1 << (newlen / XFS_INODES_PER_HOLEMASK_BIT)) - 1;
777 }
778
779 if (args.fsbno == NULLFSBLOCK) {
780 *alloc = 0;
781 return 0;
782 }
783 ASSERT(args.len == args.minlen);
784
785 /*
786 * Stamp and write the inode buffers.
787 *
788 * Seed the new inode cluster with a random generation number. This
789 * prevents short-term reuse of generation numbers if a chunk is
790 * freed and then immediately reallocated. We use random numbers
791 * rather than a linear progression to prevent the next generation
792 * number from being easily guessable.
793 */
794 error = xfs_ialloc_inode_init(args.mp, tp, NULL, newlen, agno,
795 args.agbno, args.len, prandom_u32());
796
797 if (error)
798 return error;
799 /*
800 * Convert the results.
801 */
802 newino = XFS_OFFBNO_TO_AGINO(args.mp, args.agbno, 0);
803
804 if (xfs_inobt_issparse(~allocmask)) {
805 /*
806 * We've allocated a sparse chunk. Align the startino and mask.
807 */
808 xfs_align_sparse_ino(args.mp, &newino, &allocmask);
809
810 rec.ir_startino = newino;
811 rec.ir_holemask = ~allocmask;
812 rec.ir_count = newlen;
813 rec.ir_freecount = newlen;
814 rec.ir_free = XFS_INOBT_ALL_FREE;
815
816 /*
817 * Insert the sparse record into the inobt and allow for a merge
818 * if necessary. If a merge does occur, rec is updated to the
819 * merged record.
820 */
821 error = xfs_inobt_insert_sprec(args.mp, tp, agbp, XFS_BTNUM_INO,
822 &rec, true);
823 if (error == -EFSCORRUPTED) {
824 xfs_alert(args.mp,
825 "invalid sparse inode record: ino 0x%llx holemask 0x%x count %u",
826 XFS_AGINO_TO_INO(args.mp, agno,
827 rec.ir_startino),
828 rec.ir_holemask, rec.ir_count);
829 xfs_force_shutdown(args.mp, SHUTDOWN_CORRUPT_INCORE);
830 }
831 if (error)
832 return error;
833
834 /*
835 * We can't merge the part we've just allocated as for the inobt
836 * due to finobt semantics. The original record may or may not
837 * exist independent of whether physical inodes exist in this
838 * sparse chunk.
839 *
840 * We must update the finobt record based on the inobt record.
841 * rec contains the fully merged and up to date inobt record
842 * from the previous call. Set merge false to replace any
843 * existing record with this one.
844 */
845 if (xfs_sb_version_hasfinobt(&args.mp->m_sb)) {
846 error = xfs_inobt_insert_sprec(args.mp, tp, agbp,
847 XFS_BTNUM_FINO, &rec,
848 false);
849 if (error)
850 return error;
851 }
852 } else {
853 /* full chunk - insert new records to both btrees */
854 error = xfs_inobt_insert(args.mp, tp, agbp, newino, newlen,
855 XFS_BTNUM_INO);
856 if (error)
857 return error;
858
859 if (xfs_sb_version_hasfinobt(&args.mp->m_sb)) {
860 error = xfs_inobt_insert(args.mp, tp, agbp, newino,
861 newlen, XFS_BTNUM_FINO);
862 if (error)
863 return error;
864 }
865 }
866
867 /*
868 * Update AGI counts and newino.
869 */
870 be32_add_cpu(&agi->agi_count, newlen);
871 be32_add_cpu(&agi->agi_freecount, newlen);
872 pag = xfs_perag_get(args.mp, agno);
873 pag->pagi_freecount += newlen;
874 xfs_perag_put(pag);
875 agi->agi_newino = cpu_to_be32(newino);
876
877 /*
878 * Log allocation group header fields
879 */
880 xfs_ialloc_log_agi(tp, agbp,
881 XFS_AGI_COUNT | XFS_AGI_FREECOUNT | XFS_AGI_NEWINO);
882 /*
883 * Modify/log superblock values for inode count and inode free count.
884 */
885 xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, (long)newlen);
886 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, (long)newlen);
887 *alloc = 1;
888 return 0;
889}
890
891STATIC xfs_agnumber_t
892xfs_ialloc_next_ag(
893 xfs_mount_t *mp)
894{
895 xfs_agnumber_t agno;
896
897 spin_lock(&mp->m_agirotor_lock);
898 agno = mp->m_agirotor;
899 if (++mp->m_agirotor >= mp->m_maxagi)
900 mp->m_agirotor = 0;
901 spin_unlock(&mp->m_agirotor_lock);
902
903 return agno;
904}
905
906/*
907 * Select an allocation group to look for a free inode in, based on the parent
908 * inode and the mode. Return the allocation group buffer.
909 */
910STATIC xfs_agnumber_t
911xfs_ialloc_ag_select(
912 xfs_trans_t *tp, /* transaction pointer */
913 xfs_ino_t parent, /* parent directory inode number */
914 umode_t mode, /* bits set to indicate file type */
915 int okalloc) /* ok to allocate more space */
916{
917 xfs_agnumber_t agcount; /* number of ag's in the filesystem */
918 xfs_agnumber_t agno; /* current ag number */
919 int flags; /* alloc buffer locking flags */
920 xfs_extlen_t ineed; /* blocks needed for inode allocation */
921 xfs_extlen_t longest = 0; /* longest extent available */
922 xfs_mount_t *mp; /* mount point structure */
923 int needspace; /* file mode implies space allocated */
924 xfs_perag_t *pag; /* per allocation group data */
925 xfs_agnumber_t pagno; /* parent (starting) ag number */
926 int error;
927
928 /*
929 * Files of these types need at least one block if length > 0
930 * (and they won't fit in the inode, but that's hard to figure out).
931 */
932 needspace = S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode);
933 mp = tp->t_mountp;
934 agcount = mp->m_maxagi;
935 if (S_ISDIR(mode))
936 pagno = xfs_ialloc_next_ag(mp);
937 else {
938 pagno = XFS_INO_TO_AGNO(mp, parent);
939 if (pagno >= agcount)
940 pagno = 0;
941 }
942
943 ASSERT(pagno < agcount);
944
945 /*
946 * Loop through allocation groups, looking for one with a little
947 * free space in it. Note we don't look for free inodes, exactly.
948 * Instead, we include whether there is a need to allocate inodes
949 * to mean that blocks must be allocated for them,
950 * if none are currently free.
951 */
952 agno = pagno;
953 flags = XFS_ALLOC_FLAG_TRYLOCK;
954 for (;;) {
955 pag = xfs_perag_get(mp, agno);
956 if (!pag->pagi_inodeok) {
957 xfs_ialloc_next_ag(mp);
958 goto nextag;
959 }
960
961 if (!pag->pagi_init) {
962 error = xfs_ialloc_pagi_init(mp, tp, agno);
963 if (error)
964 goto nextag;
965 }
966
967 if (pag->pagi_freecount) {
968 xfs_perag_put(pag);
969 return agno;
970 }
971
972 if (!okalloc)
973 goto nextag;
974
975 if (!pag->pagf_init) {
976 error = xfs_alloc_pagf_init(mp, tp, agno, flags);
977 if (error)
978 goto nextag;
979 }
980
981 /*
982 * Check that there is enough free space for the file plus a
983 * chunk of inodes if we need to allocate some. If this is the
984 * first pass across the AGs, take into account the potential
985 * space needed for alignment of inode chunks when checking the
986 * longest contiguous free space in the AG - this prevents us
987 * from getting ENOSPC because we have free space larger than
988 * m_ialloc_blks but alignment constraints prevent us from using
989 * it.
990 *
991 * If we can't find an AG with space for full alignment slack to
992 * be taken into account, we must be near ENOSPC in all AGs.
993 * Hence we don't include alignment for the second pass and so
994 * if we fail allocation due to alignment issues then it is most
995 * likely a real ENOSPC condition.
996 */
997 ineed = mp->m_ialloc_min_blks;
998 if (flags && ineed > 1)
999 ineed += xfs_ialloc_cluster_alignment(mp);
1000 longest = pag->pagf_longest;
1001 if (!longest)
1002 longest = pag->pagf_flcount > 0;
1003
1004 if (pag->pagf_freeblks >= needspace + ineed &&
1005 longest >= ineed) {
1006 xfs_perag_put(pag);
1007 return agno;
1008 }
1009nextag:
1010 xfs_perag_put(pag);
1011 /*
1012 * No point in iterating over the rest, if we're shutting
1013 * down.
1014 */
1015 if (XFS_FORCED_SHUTDOWN(mp))
1016 return NULLAGNUMBER;
1017 agno++;
1018 if (agno >= agcount)
1019 agno = 0;
1020 if (agno == pagno) {
1021 if (flags == 0)
1022 return NULLAGNUMBER;
1023 flags = 0;
1024 }
1025 }
1026}
1027
1028/*
1029 * Try to retrieve the next record to the left/right from the current one.
1030 */
1031STATIC int
1032xfs_ialloc_next_rec(
1033 struct xfs_btree_cur *cur,
1034 xfs_inobt_rec_incore_t *rec,
1035 int *done,
1036 int left)
1037{
1038 int error;
1039 int i;
1040
1041 if (left)
1042 error = xfs_btree_decrement(cur, 0, &i);
1043 else
1044 error = xfs_btree_increment(cur, 0, &i);
1045
1046 if (error)
1047 return error;
1048 *done = !i;
1049 if (i) {
1050 error = xfs_inobt_get_rec(cur, rec, &i);
1051 if (error)
1052 return error;
1053 XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, i == 1);
1054 }
1055
1056 return 0;
1057}
1058
1059STATIC int
1060xfs_ialloc_get_rec(
1061 struct xfs_btree_cur *cur,
1062 xfs_agino_t agino,
1063 xfs_inobt_rec_incore_t *rec,
1064 int *done)
1065{
1066 int error;
1067 int i;
1068
1069 error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_EQ, &i);
1070 if (error)
1071 return error;
1072 *done = !i;
1073 if (i) {
1074 error = xfs_inobt_get_rec(cur, rec, &i);
1075 if (error)
1076 return error;
1077 XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, i == 1);
1078 }
1079
1080 return 0;
1081}
1082
1083/*
1084 * Return the offset of the first free inode in the record. If the inode chunk
1085 * is sparsely allocated, we convert the record holemask to inode granularity
1086 * and mask off the unallocated regions from the inode free mask.
1087 */
1088STATIC int
1089xfs_inobt_first_free_inode(
1090 struct xfs_inobt_rec_incore *rec)
1091{
1092 xfs_inofree_t realfree;
1093
1094 /* if there are no holes, return the first available offset */
1095 if (!xfs_inobt_issparse(rec->ir_holemask))
1096 return xfs_lowbit64(rec->ir_free);
1097
1098 realfree = xfs_inobt_irec_to_allocmask(rec);
1099 realfree &= rec->ir_free;
1100
1101 return xfs_lowbit64(realfree);
1102}
1103
1104/*
1105 * Allocate an inode using the inobt-only algorithm.
1106 */
1107STATIC int
1108xfs_dialloc_ag_inobt(
1109 struct xfs_trans *tp,
1110 struct xfs_buf *agbp,
1111 xfs_ino_t parent,
1112 xfs_ino_t *inop)
1113{
1114 struct xfs_mount *mp = tp->t_mountp;
1115 struct xfs_agi *agi = XFS_BUF_TO_AGI(agbp);
1116 xfs_agnumber_t agno = be32_to_cpu(agi->agi_seqno);
1117 xfs_agnumber_t pagno = XFS_INO_TO_AGNO(mp, parent);
1118 xfs_agino_t pagino = XFS_INO_TO_AGINO(mp, parent);
1119 struct xfs_perag *pag;
1120 struct xfs_btree_cur *cur, *tcur;
1121 struct xfs_inobt_rec_incore rec, trec;
1122 xfs_ino_t ino;
1123 int error;
1124 int offset;
1125 int i, j;
1126
1127 pag = xfs_perag_get(mp, agno);
1128
1129 ASSERT(pag->pagi_init);
1130 ASSERT(pag->pagi_inodeok);
1131 ASSERT(pag->pagi_freecount > 0);
1132
1133 restart_pagno:
1134 cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
1135 /*
1136 * If pagino is 0 (this is the root inode allocation) use newino.
1137 * This must work because we've just allocated some.
1138 */
1139 if (!pagino)
1140 pagino = be32_to_cpu(agi->agi_newino);
1141
1142 error = xfs_check_agi_freecount(cur, agi);
1143 if (error)
1144 goto error0;
1145
1146 /*
1147 * If in the same AG as the parent, try to get near the parent.
1148 */
1149 if (pagno == agno) {
1150 int doneleft; /* done, to the left */
1151 int doneright; /* done, to the right */
1152 int searchdistance = 10;
1153
1154 error = xfs_inobt_lookup(cur, pagino, XFS_LOOKUP_LE, &i);
1155 if (error)
1156 goto error0;
1157 XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error0);
1158
1159 error = xfs_inobt_get_rec(cur, &rec, &j);
1160 if (error)
1161 goto error0;
1162 XFS_WANT_CORRUPTED_GOTO(mp, j == 1, error0);
1163
1164 if (rec.ir_freecount > 0) {
1165 /*
1166 * Found a free inode in the same chunk
1167 * as the parent, done.
1168 */
1169 goto alloc_inode;
1170 }
1171
1172
1173 /*
1174 * In the same AG as parent, but parent's chunk is full.
1175 */
1176
1177 /* duplicate the cursor, search left & right simultaneously */
1178 error = xfs_btree_dup_cursor(cur, &tcur);
1179 if (error)
1180 goto error0;
1181
1182 /*
1183 * Skip to last blocks looked up if same parent inode.
1184 */
1185 if (pagino != NULLAGINO &&
1186 pag->pagl_pagino == pagino &&
1187 pag->pagl_leftrec != NULLAGINO &&
1188 pag->pagl_rightrec != NULLAGINO) {
1189 error = xfs_ialloc_get_rec(tcur, pag->pagl_leftrec,
1190 &trec, &doneleft);
1191 if (error)
1192 goto error1;
1193
1194 error = xfs_ialloc_get_rec(cur, pag->pagl_rightrec,
1195 &rec, &doneright);
1196 if (error)
1197 goto error1;
1198 } else {
1199 /* search left with tcur, back up 1 record */
1200 error = xfs_ialloc_next_rec(tcur, &trec, &doneleft, 1);
1201 if (error)
1202 goto error1;
1203
1204 /* search right with cur, go forward 1 record. */
1205 error = xfs_ialloc_next_rec(cur, &rec, &doneright, 0);
1206 if (error)
1207 goto error1;
1208 }
1209
1210 /*
1211 * Loop until we find an inode chunk with a free inode.
1212 */
1213 while (!doneleft || !doneright) {
1214 int useleft; /* using left inode chunk this time */
1215
1216 if (!--searchdistance) {
1217 /*
1218 * Not in range - save last search
1219 * location and allocate a new inode
1220 */
1221 xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1222 pag->pagl_leftrec = trec.ir_startino;
1223 pag->pagl_rightrec = rec.ir_startino;
1224 pag->pagl_pagino = pagino;
1225 goto newino;
1226 }
1227
1228 /* figure out the closer block if both are valid. */
1229 if (!doneleft && !doneright) {
1230 useleft = pagino -
1231 (trec.ir_startino + XFS_INODES_PER_CHUNK - 1) <
1232 rec.ir_startino - pagino;
1233 } else {
1234 useleft = !doneleft;
1235 }
1236
1237 /* free inodes to the left? */
1238 if (useleft && trec.ir_freecount) {
1239 rec = trec;
1240 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1241 cur = tcur;
1242
1243 pag->pagl_leftrec = trec.ir_startino;
1244 pag->pagl_rightrec = rec.ir_startino;
1245 pag->pagl_pagino = pagino;
1246 goto alloc_inode;
1247 }
1248
1249 /* free inodes to the right? */
1250 if (!useleft && rec.ir_freecount) {
1251 xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1252
1253 pag->pagl_leftrec = trec.ir_startino;
1254 pag->pagl_rightrec = rec.ir_startino;
1255 pag->pagl_pagino = pagino;
1256 goto alloc_inode;
1257 }
1258
1259 /* get next record to check */
1260 if (useleft) {
1261 error = xfs_ialloc_next_rec(tcur, &trec,
1262 &doneleft, 1);
1263 } else {
1264 error = xfs_ialloc_next_rec(cur, &rec,
1265 &doneright, 0);
1266 }
1267 if (error)
1268 goto error1;
1269 }
1270
1271 /*
1272 * We've reached the end of the btree. because
1273 * we are only searching a small chunk of the
1274 * btree each search, there is obviously free
1275 * inodes closer to the parent inode than we
1276 * are now. restart the search again.
1277 */
1278 pag->pagl_pagino = NULLAGINO;
1279 pag->pagl_leftrec = NULLAGINO;
1280 pag->pagl_rightrec = NULLAGINO;
1281 xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1282 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1283 goto restart_pagno;
1284 }
1285
1286 /*
1287 * In a different AG from the parent.
1288 * See if the most recently allocated block has any free.
1289 */
1290newino:
1291 if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1292 error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1293 XFS_LOOKUP_EQ, &i);
1294 if (error)
1295 goto error0;
1296
1297 if (i == 1) {
1298 error = xfs_inobt_get_rec(cur, &rec, &j);
1299 if (error)
1300 goto error0;
1301
1302 if (j == 1 && rec.ir_freecount > 0) {
1303 /*
1304 * The last chunk allocated in the group
1305 * still has a free inode.
1306 */
1307 goto alloc_inode;
1308 }
1309 }
1310 }
1311
1312 /*
1313 * None left in the last group, search the whole AG
1314 */
1315 error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1316 if (error)
1317 goto error0;
1318 XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error0);
1319
1320 for (;;) {
1321 error = xfs_inobt_get_rec(cur, &rec, &i);
1322 if (error)
1323 goto error0;
1324 XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error0);
1325 if (rec.ir_freecount > 0)
1326 break;
1327 error = xfs_btree_increment(cur, 0, &i);
1328 if (error)
1329 goto error0;
1330 XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error0);
1331 }
1332
1333alloc_inode:
1334 offset = xfs_inobt_first_free_inode(&rec);
1335 ASSERT(offset >= 0);
1336 ASSERT(offset < XFS_INODES_PER_CHUNK);
1337 ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1338 XFS_INODES_PER_CHUNK) == 0);
1339 ino = XFS_AGINO_TO_INO(mp, agno, rec.ir_startino + offset);
1340 rec.ir_free &= ~XFS_INOBT_MASK(offset);
1341 rec.ir_freecount--;
1342 error = xfs_inobt_update(cur, &rec);
1343 if (error)
1344 goto error0;
1345 be32_add_cpu(&agi->agi_freecount, -1);
1346 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1347 pag->pagi_freecount--;
1348
1349 error = xfs_check_agi_freecount(cur, agi);
1350 if (error)
1351 goto error0;
1352
1353 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1354 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1355 xfs_perag_put(pag);
1356 *inop = ino;
1357 return 0;
1358error1:
1359 xfs_btree_del_cursor(tcur, XFS_BTREE_ERROR);
1360error0:
1361 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1362 xfs_perag_put(pag);
1363 return error;
1364}
1365
1366/*
1367 * Use the free inode btree to allocate an inode based on distance from the
1368 * parent. Note that the provided cursor may be deleted and replaced.
1369 */
1370STATIC int
1371xfs_dialloc_ag_finobt_near(
1372 xfs_agino_t pagino,
1373 struct xfs_btree_cur **ocur,
1374 struct xfs_inobt_rec_incore *rec)
1375{
1376 struct xfs_btree_cur *lcur = *ocur; /* left search cursor */
1377 struct xfs_btree_cur *rcur; /* right search cursor */
1378 struct xfs_inobt_rec_incore rrec;
1379 int error;
1380 int i, j;
1381
1382 error = xfs_inobt_lookup(lcur, pagino, XFS_LOOKUP_LE, &i);
1383 if (error)
1384 return error;
1385
1386 if (i == 1) {
1387 error = xfs_inobt_get_rec(lcur, rec, &i);
1388 if (error)
1389 return error;
1390 XFS_WANT_CORRUPTED_RETURN(lcur->bc_mp, i == 1);
1391
1392 /*
1393 * See if we've landed in the parent inode record. The finobt
1394 * only tracks chunks with at least one free inode, so record
1395 * existence is enough.
1396 */
1397 if (pagino >= rec->ir_startino &&
1398 pagino < (rec->ir_startino + XFS_INODES_PER_CHUNK))
1399 return 0;
1400 }
1401
1402 error = xfs_btree_dup_cursor(lcur, &rcur);
1403 if (error)
1404 return error;
1405
1406 error = xfs_inobt_lookup(rcur, pagino, XFS_LOOKUP_GE, &j);
1407 if (error)
1408 goto error_rcur;
1409 if (j == 1) {
1410 error = xfs_inobt_get_rec(rcur, &rrec, &j);
1411 if (error)
1412 goto error_rcur;
1413 XFS_WANT_CORRUPTED_GOTO(lcur->bc_mp, j == 1, error_rcur);
1414 }
1415
1416 XFS_WANT_CORRUPTED_GOTO(lcur->bc_mp, i == 1 || j == 1, error_rcur);
1417 if (i == 1 && j == 1) {
1418 /*
1419 * Both the left and right records are valid. Choose the closer
1420 * inode chunk to the target.
1421 */
1422 if ((pagino - rec->ir_startino + XFS_INODES_PER_CHUNK - 1) >
1423 (rrec.ir_startino - pagino)) {
1424 *rec = rrec;
1425 xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1426 *ocur = rcur;
1427 } else {
1428 xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1429 }
1430 } else if (j == 1) {
1431 /* only the right record is valid */
1432 *rec = rrec;
1433 xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1434 *ocur = rcur;
1435 } else if (i == 1) {
1436 /* only the left record is valid */
1437 xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1438 }
1439
1440 return 0;
1441
1442error_rcur:
1443 xfs_btree_del_cursor(rcur, XFS_BTREE_ERROR);
1444 return error;
1445}
1446
1447/*
1448 * Use the free inode btree to find a free inode based on a newino hint. If
1449 * the hint is NULL, find the first free inode in the AG.
1450 */
1451STATIC int
1452xfs_dialloc_ag_finobt_newino(
1453 struct xfs_agi *agi,
1454 struct xfs_btree_cur *cur,
1455 struct xfs_inobt_rec_incore *rec)
1456{
1457 int error;
1458 int i;
1459
1460 if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1461 error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1462 XFS_LOOKUP_EQ, &i);
1463 if (error)
1464 return error;
1465 if (i == 1) {
1466 error = xfs_inobt_get_rec(cur, rec, &i);
1467 if (error)
1468 return error;
1469 XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, i == 1);
1470 return 0;
1471 }
1472 }
1473
1474 /*
1475 * Find the first inode available in the AG.
1476 */
1477 error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1478 if (error)
1479 return error;
1480 XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, i == 1);
1481
1482 error = xfs_inobt_get_rec(cur, rec, &i);
1483 if (error)
1484 return error;
1485 XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, i == 1);
1486
1487 return 0;
1488}
1489
1490/*
1491 * Update the inobt based on a modification made to the finobt. Also ensure that
1492 * the records from both trees are equivalent post-modification.
1493 */
1494STATIC int
1495xfs_dialloc_ag_update_inobt(
1496 struct xfs_btree_cur *cur, /* inobt cursor */
1497 struct xfs_inobt_rec_incore *frec, /* finobt record */
1498 int offset) /* inode offset */
1499{
1500 struct xfs_inobt_rec_incore rec;
1501 int error;
1502 int i;
1503
1504 error = xfs_inobt_lookup(cur, frec->ir_startino, XFS_LOOKUP_EQ, &i);
1505 if (error)
1506 return error;
1507 XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, i == 1);
1508
1509 error = xfs_inobt_get_rec(cur, &rec, &i);
1510 if (error)
1511 return error;
1512 XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, i == 1);
1513 ASSERT((XFS_AGINO_TO_OFFSET(cur->bc_mp, rec.ir_startino) %
1514 XFS_INODES_PER_CHUNK) == 0);
1515
1516 rec.ir_free &= ~XFS_INOBT_MASK(offset);
1517 rec.ir_freecount--;
1518
1519 XFS_WANT_CORRUPTED_RETURN(cur->bc_mp, (rec.ir_free == frec->ir_free) &&
1520 (rec.ir_freecount == frec->ir_freecount));
1521
1522 return xfs_inobt_update(cur, &rec);
1523}
1524
1525/*
1526 * Allocate an inode using the free inode btree, if available. Otherwise, fall
1527 * back to the inobt search algorithm.
1528 *
1529 * The caller selected an AG for us, and made sure that free inodes are
1530 * available.
1531 */
1532STATIC int
1533xfs_dialloc_ag(
1534 struct xfs_trans *tp,
1535 struct xfs_buf *agbp,
1536 xfs_ino_t parent,
1537 xfs_ino_t *inop)
1538{
1539 struct xfs_mount *mp = tp->t_mountp;
1540 struct xfs_agi *agi = XFS_BUF_TO_AGI(agbp);
1541 xfs_agnumber_t agno = be32_to_cpu(agi->agi_seqno);
1542 xfs_agnumber_t pagno = XFS_INO_TO_AGNO(mp, parent);
1543 xfs_agino_t pagino = XFS_INO_TO_AGINO(mp, parent);
1544 struct xfs_perag *pag;
1545 struct xfs_btree_cur *cur; /* finobt cursor */
1546 struct xfs_btree_cur *icur; /* inobt cursor */
1547 struct xfs_inobt_rec_incore rec;
1548 xfs_ino_t ino;
1549 int error;
1550 int offset;
1551 int i;
1552
1553 if (!xfs_sb_version_hasfinobt(&mp->m_sb))
1554 return xfs_dialloc_ag_inobt(tp, agbp, parent, inop);
1555
1556 pag = xfs_perag_get(mp, agno);
1557
1558 /*
1559 * If pagino is 0 (this is the root inode allocation) use newino.
1560 * This must work because we've just allocated some.
1561 */
1562 if (!pagino)
1563 pagino = be32_to_cpu(agi->agi_newino);
1564
1565 cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_FINO);
1566
1567 error = xfs_check_agi_freecount(cur, agi);
1568 if (error)
1569 goto error_cur;
1570
1571 /*
1572 * The search algorithm depends on whether we're in the same AG as the
1573 * parent. If so, find the closest available inode to the parent. If
1574 * not, consider the agi hint or find the first free inode in the AG.
1575 */
1576 if (agno == pagno)
1577 error = xfs_dialloc_ag_finobt_near(pagino, &cur, &rec);
1578 else
1579 error = xfs_dialloc_ag_finobt_newino(agi, cur, &rec);
1580 if (error)
1581 goto error_cur;
1582
1583 offset = xfs_inobt_first_free_inode(&rec);
1584 ASSERT(offset >= 0);
1585 ASSERT(offset < XFS_INODES_PER_CHUNK);
1586 ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1587 XFS_INODES_PER_CHUNK) == 0);
1588 ino = XFS_AGINO_TO_INO(mp, agno, rec.ir_startino + offset);
1589
1590 /*
1591 * Modify or remove the finobt record.
1592 */
1593 rec.ir_free &= ~XFS_INOBT_MASK(offset);
1594 rec.ir_freecount--;
1595 if (rec.ir_freecount)
1596 error = xfs_inobt_update(cur, &rec);
1597 else
1598 error = xfs_btree_delete(cur, &i);
1599 if (error)
1600 goto error_cur;
1601
1602 /*
1603 * The finobt has now been updated appropriately. We haven't updated the
1604 * agi and superblock yet, so we can create an inobt cursor and validate
1605 * the original freecount. If all is well, make the equivalent update to
1606 * the inobt using the finobt record and offset information.
1607 */
1608 icur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
1609
1610 error = xfs_check_agi_freecount(icur, agi);
1611 if (error)
1612 goto error_icur;
1613
1614 error = xfs_dialloc_ag_update_inobt(icur, &rec, offset);
1615 if (error)
1616 goto error_icur;
1617
1618 /*
1619 * Both trees have now been updated. We must update the perag and
1620 * superblock before we can check the freecount for each btree.
1621 */
1622 be32_add_cpu(&agi->agi_freecount, -1);
1623 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1624 pag->pagi_freecount--;
1625
1626 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1627
1628 error = xfs_check_agi_freecount(icur, agi);
1629 if (error)
1630 goto error_icur;
1631 error = xfs_check_agi_freecount(cur, agi);
1632 if (error)
1633 goto error_icur;
1634
1635 xfs_btree_del_cursor(icur, XFS_BTREE_NOERROR);
1636 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1637 xfs_perag_put(pag);
1638 *inop = ino;
1639 return 0;
1640
1641error_icur:
1642 xfs_btree_del_cursor(icur, XFS_BTREE_ERROR);
1643error_cur:
1644 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1645 xfs_perag_put(pag);
1646 return error;
1647}
1648
1649/*
1650 * Allocate an inode on disk.
1651 *
1652 * Mode is used to tell whether the new inode will need space, and whether it
1653 * is a directory.
1654 *
1655 * This function is designed to be called twice if it has to do an allocation
1656 * to make more free inodes. On the first call, *IO_agbp should be set to NULL.
1657 * If an inode is available without having to performn an allocation, an inode
1658 * number is returned. In this case, *IO_agbp is set to NULL. If an allocation
1659 * needs to be done, xfs_dialloc returns the current AGI buffer in *IO_agbp.
1660 * The caller should then commit the current transaction, allocate a
1661 * new transaction, and call xfs_dialloc() again, passing in the previous value
1662 * of *IO_agbp. IO_agbp should be held across the transactions. Since the AGI
1663 * buffer is locked across the two calls, the second call is guaranteed to have
1664 * a free inode available.
1665 *
1666 * Once we successfully pick an inode its number is returned and the on-disk
1667 * data structures are updated. The inode itself is not read in, since doing so
1668 * would break ordering constraints with xfs_reclaim.
1669 */
1670int
1671xfs_dialloc(
1672 struct xfs_trans *tp,
1673 xfs_ino_t parent,
1674 umode_t mode,
1675 int okalloc,
1676 struct xfs_buf **IO_agbp,
1677 xfs_ino_t *inop)
1678{
1679 struct xfs_mount *mp = tp->t_mountp;
1680 struct xfs_buf *agbp;
1681 xfs_agnumber_t agno;
1682 int error;
1683 int ialloced;
1684 int noroom = 0;
1685 xfs_agnumber_t start_agno;
1686 struct xfs_perag *pag;
1687
1688 if (*IO_agbp) {
1689 /*
1690 * If the caller passes in a pointer to the AGI buffer,
1691 * continue where we left off before. In this case, we
1692 * know that the allocation group has free inodes.
1693 */
1694 agbp = *IO_agbp;
1695 goto out_alloc;
1696 }
1697
1698 /*
1699 * We do not have an agbp, so select an initial allocation
1700 * group for inode allocation.
1701 */
1702 start_agno = xfs_ialloc_ag_select(tp, parent, mode, okalloc);
1703 if (start_agno == NULLAGNUMBER) {
1704 *inop = NULLFSINO;
1705 return 0;
1706 }
1707
1708 /*
1709 * If we have already hit the ceiling of inode blocks then clear
1710 * okalloc so we scan all available agi structures for a free
1711 * inode.
1712 *
1713 * Read rough value of mp->m_icount by percpu_counter_read_positive,
1714 * which will sacrifice the preciseness but improve the performance.
1715 */
1716 if (mp->m_maxicount &&
1717 percpu_counter_read_positive(&mp->m_icount) + mp->m_ialloc_inos
1718 > mp->m_maxicount) {
1719 noroom = 1;
1720 okalloc = 0;
1721 }
1722
1723 /*
1724 * Loop until we find an allocation group that either has free inodes
1725 * or in which we can allocate some inodes. Iterate through the
1726 * allocation groups upward, wrapping at the end.
1727 */
1728 agno = start_agno;
1729 for (;;) {
1730 pag = xfs_perag_get(mp, agno);
1731 if (!pag->pagi_inodeok) {
1732 xfs_ialloc_next_ag(mp);
1733 goto nextag;
1734 }
1735
1736 if (!pag->pagi_init) {
1737 error = xfs_ialloc_pagi_init(mp, tp, agno);
1738 if (error)
1739 goto out_error;
1740 }
1741
1742 /*
1743 * Do a first racy fast path check if this AG is usable.
1744 */
1745 if (!pag->pagi_freecount && !okalloc)
1746 goto nextag;
1747
1748 /*
1749 * Then read in the AGI buffer and recheck with the AGI buffer
1750 * lock held.
1751 */
1752 error = xfs_ialloc_read_agi(mp, tp, agno, &agbp);
1753 if (error)
1754 goto out_error;
1755
1756 if (pag->pagi_freecount) {
1757 xfs_perag_put(pag);
1758 goto out_alloc;
1759 }
1760
1761 if (!okalloc)
1762 goto nextag_relse_buffer;
1763
1764
1765 error = xfs_ialloc_ag_alloc(tp, agbp, &ialloced);
1766 if (error) {
1767 xfs_trans_brelse(tp, agbp);
1768
1769 if (error != -ENOSPC)
1770 goto out_error;
1771
1772 xfs_perag_put(pag);
1773 *inop = NULLFSINO;
1774 return 0;
1775 }
1776
1777 if (ialloced) {
1778 /*
1779 * We successfully allocated some inodes, return
1780 * the current context to the caller so that it
1781 * can commit the current transaction and call
1782 * us again where we left off.
1783 */
1784 ASSERT(pag->pagi_freecount > 0);
1785 xfs_perag_put(pag);
1786
1787 *IO_agbp = agbp;
1788 *inop = NULLFSINO;
1789 return 0;
1790 }
1791
1792nextag_relse_buffer:
1793 xfs_trans_brelse(tp, agbp);
1794nextag:
1795 xfs_perag_put(pag);
1796 if (++agno == mp->m_sb.sb_agcount)
1797 agno = 0;
1798 if (agno == start_agno) {
1799 *inop = NULLFSINO;
1800 return noroom ? -ENOSPC : 0;
1801 }
1802 }
1803
1804out_alloc:
1805 *IO_agbp = NULL;
1806 return xfs_dialloc_ag(tp, agbp, parent, inop);
1807out_error:
1808 xfs_perag_put(pag);
1809 return error;
1810}
1811
1812/*
1813 * Free the blocks of an inode chunk. We must consider that the inode chunk
1814 * might be sparse and only free the regions that are allocated as part of the
1815 * chunk.
1816 */
1817STATIC void
1818xfs_difree_inode_chunk(
1819 struct xfs_mount *mp,
1820 xfs_agnumber_t agno,
1821 struct xfs_inobt_rec_incore *rec,
1822 struct xfs_defer_ops *dfops)
1823{
1824 xfs_agblock_t sagbno = XFS_AGINO_TO_AGBNO(mp, rec->ir_startino);
1825 int startidx, endidx;
1826 int nextbit;
1827 xfs_agblock_t agbno;
1828 int contigblk;
1829 struct xfs_owner_info oinfo;
1830 DECLARE_BITMAP(holemask, XFS_INOBT_HOLEMASK_BITS);
1831 xfs_rmap_ag_owner(&oinfo, XFS_RMAP_OWN_INODES);
1832
1833 if (!xfs_inobt_issparse(rec->ir_holemask)) {
1834 /* not sparse, calculate extent info directly */
1835 xfs_bmap_add_free(mp, dfops, XFS_AGB_TO_FSB(mp, agno, sagbno),
1836 mp->m_ialloc_blks, &oinfo);
1837 return;
1838 }
1839
1840 /* holemask is only 16-bits (fits in an unsigned long) */
1841 ASSERT(sizeof(rec->ir_holemask) <= sizeof(holemask[0]));
1842 holemask[0] = rec->ir_holemask;
1843
1844 /*
1845 * Find contiguous ranges of zeroes (i.e., allocated regions) in the
1846 * holemask and convert the start/end index of each range to an extent.
1847 * We start with the start and end index both pointing at the first 0 in
1848 * the mask.
1849 */
1850 startidx = endidx = find_first_zero_bit(holemask,
1851 XFS_INOBT_HOLEMASK_BITS);
1852 nextbit = startidx + 1;
1853 while (startidx < XFS_INOBT_HOLEMASK_BITS) {
1854 nextbit = find_next_zero_bit(holemask, XFS_INOBT_HOLEMASK_BITS,
1855 nextbit);
1856 /*
1857 * If the next zero bit is contiguous, update the end index of
1858 * the current range and continue.
1859 */
1860 if (nextbit != XFS_INOBT_HOLEMASK_BITS &&
1861 nextbit == endidx + 1) {
1862 endidx = nextbit;
1863 goto next;
1864 }
1865
1866 /*
1867 * nextbit is not contiguous with the current end index. Convert
1868 * the current start/end to an extent and add it to the free
1869 * list.
1870 */
1871 agbno = sagbno + (startidx * XFS_INODES_PER_HOLEMASK_BIT) /
1872 mp->m_sb.sb_inopblock;
1873 contigblk = ((endidx - startidx + 1) *
1874 XFS_INODES_PER_HOLEMASK_BIT) /
1875 mp->m_sb.sb_inopblock;
1876
1877 ASSERT(agbno % mp->m_sb.sb_spino_align == 0);
1878 ASSERT(contigblk % mp->m_sb.sb_spino_align == 0);
1879 xfs_bmap_add_free(mp, dfops, XFS_AGB_TO_FSB(mp, agno, agbno),
1880 contigblk, &oinfo);
1881
1882 /* reset range to current bit and carry on... */
1883 startidx = endidx = nextbit;
1884
1885next:
1886 nextbit++;
1887 }
1888}
1889
1890STATIC int
1891xfs_difree_inobt(
1892 struct xfs_mount *mp,
1893 struct xfs_trans *tp,
1894 struct xfs_buf *agbp,
1895 xfs_agino_t agino,
1896 struct xfs_defer_ops *dfops,
1897 struct xfs_icluster *xic,
1898 struct xfs_inobt_rec_incore *orec)
1899{
1900 struct xfs_agi *agi = XFS_BUF_TO_AGI(agbp);
1901 xfs_agnumber_t agno = be32_to_cpu(agi->agi_seqno);
1902 struct xfs_perag *pag;
1903 struct xfs_btree_cur *cur;
1904 struct xfs_inobt_rec_incore rec;
1905 int ilen;
1906 int error;
1907 int i;
1908 int off;
1909
1910 ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
1911 ASSERT(XFS_AGINO_TO_AGBNO(mp, agino) < be32_to_cpu(agi->agi_length));
1912
1913 /*
1914 * Initialize the cursor.
1915 */
1916 cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
1917
1918 error = xfs_check_agi_freecount(cur, agi);
1919 if (error)
1920 goto error0;
1921
1922 /*
1923 * Look for the entry describing this inode.
1924 */
1925 if ((error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i))) {
1926 xfs_warn(mp, "%s: xfs_inobt_lookup() returned error %d.",
1927 __func__, error);
1928 goto error0;
1929 }
1930 XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error0);
1931 error = xfs_inobt_get_rec(cur, &rec, &i);
1932 if (error) {
1933 xfs_warn(mp, "%s: xfs_inobt_get_rec() returned error %d.",
1934 __func__, error);
1935 goto error0;
1936 }
1937 XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error0);
1938 /*
1939 * Get the offset in the inode chunk.
1940 */
1941 off = agino - rec.ir_startino;
1942 ASSERT(off >= 0 && off < XFS_INODES_PER_CHUNK);
1943 ASSERT(!(rec.ir_free & XFS_INOBT_MASK(off)));
1944 /*
1945 * Mark the inode free & increment the count.
1946 */
1947 rec.ir_free |= XFS_INOBT_MASK(off);
1948 rec.ir_freecount++;
1949
1950 /*
1951 * When an inode chunk is free, it becomes eligible for removal. Don't
1952 * remove the chunk if the block size is large enough for multiple inode
1953 * chunks (that might not be free).
1954 */
1955 if (!(mp->m_flags & XFS_MOUNT_IKEEP) &&
1956 rec.ir_free == XFS_INOBT_ALL_FREE &&
1957 mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK) {
1958 xic->deleted = 1;
1959 xic->first_ino = XFS_AGINO_TO_INO(mp, agno, rec.ir_startino);
1960 xic->alloc = xfs_inobt_irec_to_allocmask(&rec);
1961
1962 /*
1963 * Remove the inode cluster from the AGI B+Tree, adjust the
1964 * AGI and Superblock inode counts, and mark the disk space
1965 * to be freed when the transaction is committed.
1966 */
1967 ilen = rec.ir_freecount;
1968 be32_add_cpu(&agi->agi_count, -ilen);
1969 be32_add_cpu(&agi->agi_freecount, -(ilen - 1));
1970 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_COUNT | XFS_AGI_FREECOUNT);
1971 pag = xfs_perag_get(mp, agno);
1972 pag->pagi_freecount -= ilen - 1;
1973 xfs_perag_put(pag);
1974 xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, -ilen);
1975 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -(ilen - 1));
1976
1977 if ((error = xfs_btree_delete(cur, &i))) {
1978 xfs_warn(mp, "%s: xfs_btree_delete returned error %d.",
1979 __func__, error);
1980 goto error0;
1981 }
1982
1983 xfs_difree_inode_chunk(mp, agno, &rec, dfops);
1984 } else {
1985 xic->deleted = 0;
1986
1987 error = xfs_inobt_update(cur, &rec);
1988 if (error) {
1989 xfs_warn(mp, "%s: xfs_inobt_update returned error %d.",
1990 __func__, error);
1991 goto error0;
1992 }
1993
1994 /*
1995 * Change the inode free counts and log the ag/sb changes.
1996 */
1997 be32_add_cpu(&agi->agi_freecount, 1);
1998 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1999 pag = xfs_perag_get(mp, agno);
2000 pag->pagi_freecount++;
2001 xfs_perag_put(pag);
2002 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, 1);
2003 }
2004
2005 error = xfs_check_agi_freecount(cur, agi);
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 xfs_agino_t agino,
2027 struct xfs_inobt_rec_incore *ibtrec) /* inobt record */
2028{
2029 struct xfs_agi *agi = XFS_BUF_TO_AGI(agbp);
2030 xfs_agnumber_t agno = be32_to_cpu(agi->agi_seqno);
2031 struct xfs_btree_cur *cur;
2032 struct xfs_inobt_rec_incore rec;
2033 int offset = agino - ibtrec->ir_startino;
2034 int error;
2035 int i;
2036
2037 cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_FINO);
2038
2039 error = xfs_inobt_lookup(cur, ibtrec->ir_startino, XFS_LOOKUP_EQ, &i);
2040 if (error)
2041 goto error;
2042 if (i == 0) {
2043 /*
2044 * If the record does not exist in the finobt, we must have just
2045 * freed an inode in a previously fully allocated chunk. If not,
2046 * something is out of sync.
2047 */
2048 XFS_WANT_CORRUPTED_GOTO(mp, ibtrec->ir_freecount == 1, error);
2049
2050 error = xfs_inobt_insert_rec(cur, ibtrec->ir_holemask,
2051 ibtrec->ir_count,
2052 ibtrec->ir_freecount,
2053 ibtrec->ir_free, &i);
2054 if (error)
2055 goto error;
2056 ASSERT(i == 1);
2057
2058 goto out;
2059 }
2060
2061 /*
2062 * Read and update the existing record. We could just copy the ibtrec
2063 * across here, but that would defeat the purpose of having redundant
2064 * metadata. By making the modifications independently, we can catch
2065 * corruptions that we wouldn't see if we just copied from one record
2066 * to another.
2067 */
2068 error = xfs_inobt_get_rec(cur, &rec, &i);
2069 if (error)
2070 goto error;
2071 XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error);
2072
2073 rec.ir_free |= XFS_INOBT_MASK(offset);
2074 rec.ir_freecount++;
2075
2076 XFS_WANT_CORRUPTED_GOTO(mp, (rec.ir_free == ibtrec->ir_free) &&
2077 (rec.ir_freecount == ibtrec->ir_freecount),
2078 error);
2079
2080 /*
2081 * The content of inobt records should always match between the inobt
2082 * and finobt. The lifecycle of records in the finobt is different from
2083 * the inobt in that the finobt only tracks records with at least one
2084 * free inode. Hence, if all of the inodes are free and we aren't
2085 * keeping inode chunks permanently on disk, remove the record.
2086 * Otherwise, update the record with the new information.
2087 *
2088 * Note that we currently can't free chunks when the block size is large
2089 * enough for multiple chunks. Leave the finobt record to remain in sync
2090 * with the inobt.
2091 */
2092 if (rec.ir_free == XFS_INOBT_ALL_FREE &&
2093 mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK &&
2094 !(mp->m_flags & XFS_MOUNT_IKEEP)) {
2095 error = xfs_btree_delete(cur, &i);
2096 if (error)
2097 goto error;
2098 ASSERT(i == 1);
2099 } else {
2100 error = xfs_inobt_update(cur, &rec);
2101 if (error)
2102 goto error;
2103 }
2104
2105out:
2106 error = xfs_check_agi_freecount(cur, agi);
2107 if (error)
2108 goto error;
2109
2110 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2111 return 0;
2112
2113error:
2114 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2115 return error;
2116}
2117
2118/*
2119 * Free disk inode. Carefully avoids touching the incore inode, all
2120 * manipulations incore are the caller's responsibility.
2121 * The on-disk inode is not changed by this operation, only the
2122 * btree (free inode mask) is changed.
2123 */
2124int
2125xfs_difree(
2126 struct xfs_trans *tp, /* transaction pointer */
2127 xfs_ino_t inode, /* inode to be freed */
2128 struct xfs_defer_ops *dfops, /* extents to free */
2129 struct xfs_icluster *xic) /* cluster info if deleted */
2130{
2131 /* REFERENCED */
2132 xfs_agblock_t agbno; /* block number containing inode */
2133 struct xfs_buf *agbp; /* buffer for allocation group header */
2134 xfs_agino_t agino; /* allocation group inode number */
2135 xfs_agnumber_t agno; /* allocation group number */
2136 int error; /* error return value */
2137 struct xfs_mount *mp; /* mount structure for filesystem */
2138 struct xfs_inobt_rec_incore rec;/* btree record */
2139
2140 mp = tp->t_mountp;
2141
2142 /*
2143 * Break up inode number into its components.
2144 */
2145 agno = XFS_INO_TO_AGNO(mp, inode);
2146 if (agno >= mp->m_sb.sb_agcount) {
2147 xfs_warn(mp, "%s: agno >= mp->m_sb.sb_agcount (%d >= %d).",
2148 __func__, agno, mp->m_sb.sb_agcount);
2149 ASSERT(0);
2150 return -EINVAL;
2151 }
2152 agino = XFS_INO_TO_AGINO(mp, inode);
2153 if (inode != XFS_AGINO_TO_INO(mp, agno, agino)) {
2154 xfs_warn(mp, "%s: inode != XFS_AGINO_TO_INO() (%llu != %llu).",
2155 __func__, (unsigned long long)inode,
2156 (unsigned long long)XFS_AGINO_TO_INO(mp, agno, agino));
2157 ASSERT(0);
2158 return -EINVAL;
2159 }
2160 agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2161 if (agbno >= mp->m_sb.sb_agblocks) {
2162 xfs_warn(mp, "%s: agbno >= mp->m_sb.sb_agblocks (%d >= %d).",
2163 __func__, agbno, mp->m_sb.sb_agblocks);
2164 ASSERT(0);
2165 return -EINVAL;
2166 }
2167 /*
2168 * Get the allocation group header.
2169 */
2170 error = xfs_ialloc_read_agi(mp, tp, agno, &agbp);
2171 if (error) {
2172 xfs_warn(mp, "%s: xfs_ialloc_read_agi() returned error %d.",
2173 __func__, error);
2174 return error;
2175 }
2176
2177 /*
2178 * Fix up the inode allocation btree.
2179 */
2180 error = xfs_difree_inobt(mp, tp, agbp, agino, dfops, xic, &rec);
2181 if (error)
2182 goto error0;
2183
2184 /*
2185 * Fix up the free inode btree.
2186 */
2187 if (xfs_sb_version_hasfinobt(&mp->m_sb)) {
2188 error = xfs_difree_finobt(mp, tp, agbp, agino, &rec);
2189 if (error)
2190 goto error0;
2191 }
2192
2193 return 0;
2194
2195error0:
2196 return error;
2197}
2198
2199STATIC int
2200xfs_imap_lookup(
2201 struct xfs_mount *mp,
2202 struct xfs_trans *tp,
2203 xfs_agnumber_t agno,
2204 xfs_agino_t agino,
2205 xfs_agblock_t agbno,
2206 xfs_agblock_t *chunk_agbno,
2207 xfs_agblock_t *offset_agbno,
2208 int flags)
2209{
2210 struct xfs_inobt_rec_incore rec;
2211 struct xfs_btree_cur *cur;
2212 struct xfs_buf *agbp;
2213 int error;
2214 int i;
2215
2216 error = xfs_ialloc_read_agi(mp, tp, agno, &agbp);
2217 if (error) {
2218 xfs_alert(mp,
2219 "%s: xfs_ialloc_read_agi() returned error %d, agno %d",
2220 __func__, error, agno);
2221 return error;
2222 }
2223
2224 /*
2225 * Lookup the inode record for the given agino. If the record cannot be
2226 * found, then it's an invalid inode number and we should abort. Once
2227 * we have a record, we need to ensure it contains the inode number
2228 * we are looking up.
2229 */
2230 cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
2231 error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i);
2232 if (!error) {
2233 if (i)
2234 error = xfs_inobt_get_rec(cur, &rec, &i);
2235 if (!error && i == 0)
2236 error = -EINVAL;
2237 }
2238
2239 xfs_trans_brelse(tp, agbp);
2240 xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
2241 if (error)
2242 return error;
2243
2244 /* check that the returned record contains the required inode */
2245 if (rec.ir_startino > agino ||
2246 rec.ir_startino + mp->m_ialloc_inos <= agino)
2247 return -EINVAL;
2248
2249 /* for untrusted inodes check it is allocated first */
2250 if ((flags & XFS_IGET_UNTRUSTED) &&
2251 (rec.ir_free & XFS_INOBT_MASK(agino - rec.ir_startino)))
2252 return -EINVAL;
2253
2254 *chunk_agbno = XFS_AGINO_TO_AGBNO(mp, rec.ir_startino);
2255 *offset_agbno = agbno - *chunk_agbno;
2256 return 0;
2257}
2258
2259/*
2260 * Return the location of the inode in imap, for mapping it into a buffer.
2261 */
2262int
2263xfs_imap(
2264 xfs_mount_t *mp, /* file system mount structure */
2265 xfs_trans_t *tp, /* transaction pointer */
2266 xfs_ino_t ino, /* inode to locate */
2267 struct xfs_imap *imap, /* location map structure */
2268 uint flags) /* flags for inode btree lookup */
2269{
2270 xfs_agblock_t agbno; /* block number of inode in the alloc group */
2271 xfs_agino_t agino; /* inode number within alloc group */
2272 xfs_agnumber_t agno; /* allocation group number */
2273 int blks_per_cluster; /* num blocks per inode cluster */
2274 xfs_agblock_t chunk_agbno; /* first block in inode chunk */
2275 xfs_agblock_t cluster_agbno; /* first block in inode cluster */
2276 int error; /* error code */
2277 int offset; /* index of inode in its buffer */
2278 xfs_agblock_t offset_agbno; /* blks from chunk start to inode */
2279
2280 ASSERT(ino != NULLFSINO);
2281
2282 /*
2283 * Split up the inode number into its parts.
2284 */
2285 agno = XFS_INO_TO_AGNO(mp, ino);
2286 agino = XFS_INO_TO_AGINO(mp, ino);
2287 agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2288 if (agno >= mp->m_sb.sb_agcount || agbno >= mp->m_sb.sb_agblocks ||
2289 ino != XFS_AGINO_TO_INO(mp, agno, agino)) {
2290#ifdef DEBUG
2291 /*
2292 * Don't output diagnostic information for untrusted inodes
2293 * as they can be invalid without implying corruption.
2294 */
2295 if (flags & XFS_IGET_UNTRUSTED)
2296 return -EINVAL;
2297 if (agno >= mp->m_sb.sb_agcount) {
2298 xfs_alert(mp,
2299 "%s: agno (%d) >= mp->m_sb.sb_agcount (%d)",
2300 __func__, agno, mp->m_sb.sb_agcount);
2301 }
2302 if (agbno >= mp->m_sb.sb_agblocks) {
2303 xfs_alert(mp,
2304 "%s: agbno (0x%llx) >= mp->m_sb.sb_agblocks (0x%lx)",
2305 __func__, (unsigned long long)agbno,
2306 (unsigned long)mp->m_sb.sb_agblocks);
2307 }
2308 if (ino != XFS_AGINO_TO_INO(mp, agno, agino)) {
2309 xfs_alert(mp,
2310 "%s: ino (0x%llx) != XFS_AGINO_TO_INO() (0x%llx)",
2311 __func__, ino,
2312 XFS_AGINO_TO_INO(mp, agno, agino));
2313 }
2314 xfs_stack_trace();
2315#endif /* DEBUG */
2316 return -EINVAL;
2317 }
2318
2319 blks_per_cluster = xfs_icluster_size_fsb(mp);
2320
2321 /*
2322 * For bulkstat and handle lookups, we have an untrusted inode number
2323 * that we have to verify is valid. We cannot do this just by reading
2324 * the inode buffer as it may have been unlinked and removed leaving
2325 * inodes in stale state on disk. Hence we have to do a btree lookup
2326 * in all cases where an untrusted inode number is passed.
2327 */
2328 if (flags & XFS_IGET_UNTRUSTED) {
2329 error = xfs_imap_lookup(mp, tp, agno, agino, agbno,
2330 &chunk_agbno, &offset_agbno, flags);
2331 if (error)
2332 return error;
2333 goto out_map;
2334 }
2335
2336 /*
2337 * If the inode cluster size is the same as the blocksize or
2338 * smaller we get to the buffer by simple arithmetics.
2339 */
2340 if (blks_per_cluster == 1) {
2341 offset = XFS_INO_TO_OFFSET(mp, ino);
2342 ASSERT(offset < mp->m_sb.sb_inopblock);
2343
2344 imap->im_blkno = XFS_AGB_TO_DADDR(mp, agno, agbno);
2345 imap->im_len = XFS_FSB_TO_BB(mp, 1);
2346 imap->im_boffset = (unsigned short)(offset <<
2347 mp->m_sb.sb_inodelog);
2348 return 0;
2349 }
2350
2351 /*
2352 * If the inode chunks are aligned then use simple maths to
2353 * find the location. Otherwise we have to do a btree
2354 * lookup to find the location.
2355 */
2356 if (mp->m_inoalign_mask) {
2357 offset_agbno = agbno & mp->m_inoalign_mask;
2358 chunk_agbno = agbno - offset_agbno;
2359 } else {
2360 error = xfs_imap_lookup(mp, tp, agno, agino, agbno,
2361 &chunk_agbno, &offset_agbno, flags);
2362 if (error)
2363 return error;
2364 }
2365
2366out_map:
2367 ASSERT(agbno >= chunk_agbno);
2368 cluster_agbno = chunk_agbno +
2369 ((offset_agbno / blks_per_cluster) * blks_per_cluster);
2370 offset = ((agbno - cluster_agbno) * mp->m_sb.sb_inopblock) +
2371 XFS_INO_TO_OFFSET(mp, ino);
2372
2373 imap->im_blkno = XFS_AGB_TO_DADDR(mp, agno, cluster_agbno);
2374 imap->im_len = XFS_FSB_TO_BB(mp, blks_per_cluster);
2375 imap->im_boffset = (unsigned short)(offset << mp->m_sb.sb_inodelog);
2376
2377 /*
2378 * If the inode number maps to a block outside the bounds
2379 * of the file system then return NULL rather than calling
2380 * read_buf and panicing when we get an error from the
2381 * driver.
2382 */
2383 if ((imap->im_blkno + imap->im_len) >
2384 XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks)) {
2385 xfs_alert(mp,
2386 "%s: (im_blkno (0x%llx) + im_len (0x%llx)) > sb_dblocks (0x%llx)",
2387 __func__, (unsigned long long) imap->im_blkno,
2388 (unsigned long long) imap->im_len,
2389 XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks));
2390 return -EINVAL;
2391 }
2392 return 0;
2393}
2394
2395/*
2396 * Compute and fill in value of m_in_maxlevels.
2397 */
2398void
2399xfs_ialloc_compute_maxlevels(
2400 xfs_mount_t *mp) /* file system mount structure */
2401{
2402 uint inodes;
2403
2404 inodes = (1LL << XFS_INO_AGINO_BITS(mp)) >> XFS_INODES_PER_CHUNK_LOG;
2405 mp->m_in_maxlevels = xfs_btree_compute_maxlevels(mp, mp->m_inobt_mnr,
2406 inodes);
2407}
2408
2409/*
2410 * Log specified fields for the ag hdr (inode section). The growth of the agi
2411 * structure over time requires that we interpret the buffer as two logical
2412 * regions delineated by the end of the unlinked list. This is due to the size
2413 * of the hash table and its location in the middle of the agi.
2414 *
2415 * For example, a request to log a field before agi_unlinked and a field after
2416 * agi_unlinked could cause us to log the entire hash table and use an excessive
2417 * amount of log space. To avoid this behavior, log the region up through
2418 * agi_unlinked in one call and the region after agi_unlinked through the end of
2419 * the structure in another.
2420 */
2421void
2422xfs_ialloc_log_agi(
2423 xfs_trans_t *tp, /* transaction pointer */
2424 xfs_buf_t *bp, /* allocation group header buffer */
2425 int fields) /* bitmask of fields to log */
2426{
2427 int first; /* first byte number */
2428 int last; /* last byte number */
2429 static const short offsets[] = { /* field starting offsets */
2430 /* keep in sync with bit definitions */
2431 offsetof(xfs_agi_t, agi_magicnum),
2432 offsetof(xfs_agi_t, agi_versionnum),
2433 offsetof(xfs_agi_t, agi_seqno),
2434 offsetof(xfs_agi_t, agi_length),
2435 offsetof(xfs_agi_t, agi_count),
2436 offsetof(xfs_agi_t, agi_root),
2437 offsetof(xfs_agi_t, agi_level),
2438 offsetof(xfs_agi_t, agi_freecount),
2439 offsetof(xfs_agi_t, agi_newino),
2440 offsetof(xfs_agi_t, agi_dirino),
2441 offsetof(xfs_agi_t, agi_unlinked),
2442 offsetof(xfs_agi_t, agi_free_root),
2443 offsetof(xfs_agi_t, agi_free_level),
2444 sizeof(xfs_agi_t)
2445 };
2446#ifdef DEBUG
2447 xfs_agi_t *agi; /* allocation group header */
2448
2449 agi = XFS_BUF_TO_AGI(bp);
2450 ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
2451#endif
2452
2453 /*
2454 * Compute byte offsets for the first and last fields in the first
2455 * region and log the agi buffer. This only logs up through
2456 * agi_unlinked.
2457 */
2458 if (fields & XFS_AGI_ALL_BITS_R1) {
2459 xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R1,
2460 &first, &last);
2461 xfs_trans_log_buf(tp, bp, first, last);
2462 }
2463
2464 /*
2465 * Mask off the bits in the first region and calculate the first and
2466 * last field offsets for any bits in the second region.
2467 */
2468 fields &= ~XFS_AGI_ALL_BITS_R1;
2469 if (fields) {
2470 xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R2,
2471 &first, &last);
2472 xfs_trans_log_buf(tp, bp, first, last);
2473 }
2474}
2475
2476#ifdef DEBUG
2477STATIC void
2478xfs_check_agi_unlinked(
2479 struct xfs_agi *agi)
2480{
2481 int i;
2482
2483 for (i = 0; i < XFS_AGI_UNLINKED_BUCKETS; i++)
2484 ASSERT(agi->agi_unlinked[i]);
2485}
2486#else
2487#define xfs_check_agi_unlinked(agi)
2488#endif
2489
2490static bool
2491xfs_agi_verify(
2492 struct xfs_buf *bp)
2493{
2494 struct xfs_mount *mp = bp->b_target->bt_mount;
2495 struct xfs_agi *agi = XFS_BUF_TO_AGI(bp);
2496
2497 if (xfs_sb_version_hascrc(&mp->m_sb)) {
2498 if (!uuid_equal(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid))
2499 return false;
2500 if (!xfs_log_check_lsn(mp,
2501 be64_to_cpu(XFS_BUF_TO_AGI(bp)->agi_lsn)))
2502 return false;
2503 }
2504
2505 /*
2506 * Validate the magic number of the agi block.
2507 */
2508 if (agi->agi_magicnum != cpu_to_be32(XFS_AGI_MAGIC))
2509 return false;
2510 if (!XFS_AGI_GOOD_VERSION(be32_to_cpu(agi->agi_versionnum)))
2511 return false;
2512
2513 if (be32_to_cpu(agi->agi_level) < 1 ||
2514 be32_to_cpu(agi->agi_level) > XFS_BTREE_MAXLEVELS)
2515 return false;
2516
2517 if (xfs_sb_version_hasfinobt(&mp->m_sb) &&
2518 (be32_to_cpu(agi->agi_free_level) < 1 ||
2519 be32_to_cpu(agi->agi_free_level) > XFS_BTREE_MAXLEVELS))
2520 return false;
2521
2522 /*
2523 * during growfs operations, the perag is not fully initialised,
2524 * so we can't use it for any useful checking. growfs ensures we can't
2525 * use it by using uncached buffers that don't have the perag attached
2526 * so we can detect and avoid this problem.
2527 */
2528 if (bp->b_pag && be32_to_cpu(agi->agi_seqno) != bp->b_pag->pag_agno)
2529 return false;
2530
2531 xfs_check_agi_unlinked(agi);
2532 return true;
2533}
2534
2535static void
2536xfs_agi_read_verify(
2537 struct xfs_buf *bp)
2538{
2539 struct xfs_mount *mp = bp->b_target->bt_mount;
2540
2541 if (xfs_sb_version_hascrc(&mp->m_sb) &&
2542 !xfs_buf_verify_cksum(bp, XFS_AGI_CRC_OFF))
2543 xfs_buf_ioerror(bp, -EFSBADCRC);
2544 else if (XFS_TEST_ERROR(!xfs_agi_verify(bp), mp,
2545 XFS_ERRTAG_IALLOC_READ_AGI,
2546 XFS_RANDOM_IALLOC_READ_AGI))
2547 xfs_buf_ioerror(bp, -EFSCORRUPTED);
2548
2549 if (bp->b_error)
2550 xfs_verifier_error(bp);
2551}
2552
2553static void
2554xfs_agi_write_verify(
2555 struct xfs_buf *bp)
2556{
2557 struct xfs_mount *mp = bp->b_target->bt_mount;
2558 struct xfs_buf_log_item *bip = bp->b_fspriv;
2559
2560 if (!xfs_agi_verify(bp)) {
2561 xfs_buf_ioerror(bp, -EFSCORRUPTED);
2562 xfs_verifier_error(bp);
2563 return;
2564 }
2565
2566 if (!xfs_sb_version_hascrc(&mp->m_sb))
2567 return;
2568
2569 if (bip)
2570 XFS_BUF_TO_AGI(bp)->agi_lsn = cpu_to_be64(bip->bli_item.li_lsn);
2571 xfs_buf_update_cksum(bp, XFS_AGI_CRC_OFF);
2572}
2573
2574const struct xfs_buf_ops xfs_agi_buf_ops = {
2575 .name = "xfs_agi",
2576 .verify_read = xfs_agi_read_verify,
2577 .verify_write = xfs_agi_write_verify,
2578};
2579
2580/*
2581 * Read in the allocation group header (inode allocation section)
2582 */
2583int
2584xfs_read_agi(
2585 struct xfs_mount *mp, /* file system mount structure */
2586 struct xfs_trans *tp, /* transaction pointer */
2587 xfs_agnumber_t agno, /* allocation group number */
2588 struct xfs_buf **bpp) /* allocation group hdr buf */
2589{
2590 int error;
2591
2592 trace_xfs_read_agi(mp, agno);
2593
2594 ASSERT(agno != NULLAGNUMBER);
2595 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
2596 XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
2597 XFS_FSS_TO_BB(mp, 1), 0, bpp, &xfs_agi_buf_ops);
2598 if (error)
2599 return error;
2600 if (tp)
2601 xfs_trans_buf_set_type(tp, *bpp, XFS_BLFT_AGI_BUF);
2602
2603 xfs_buf_set_ref(*bpp, XFS_AGI_REF);
2604 return 0;
2605}
2606
2607int
2608xfs_ialloc_read_agi(
2609 struct xfs_mount *mp, /* file system mount structure */
2610 struct xfs_trans *tp, /* transaction pointer */
2611 xfs_agnumber_t agno, /* allocation group number */
2612 struct xfs_buf **bpp) /* allocation group hdr buf */
2613{
2614 struct xfs_agi *agi; /* allocation group header */
2615 struct xfs_perag *pag; /* per allocation group data */
2616 int error;
2617
2618 trace_xfs_ialloc_read_agi(mp, agno);
2619
2620 error = xfs_read_agi(mp, tp, agno, bpp);
2621 if (error)
2622 return error;
2623
2624 agi = XFS_BUF_TO_AGI(*bpp);
2625 pag = xfs_perag_get(mp, agno);
2626 if (!pag->pagi_init) {
2627 pag->pagi_freecount = be32_to_cpu(agi->agi_freecount);
2628 pag->pagi_count = be32_to_cpu(agi->agi_count);
2629 pag->pagi_init = 1;
2630 }
2631
2632 /*
2633 * It's possible for these to be out of sync if
2634 * we are in the middle of a forced shutdown.
2635 */
2636 ASSERT(pag->pagi_freecount == be32_to_cpu(agi->agi_freecount) ||
2637 XFS_FORCED_SHUTDOWN(mp));
2638 xfs_perag_put(pag);
2639 return 0;
2640}
2641
2642/*
2643 * Read in the agi to initialise the per-ag data in the mount structure
2644 */
2645int
2646xfs_ialloc_pagi_init(
2647 xfs_mount_t *mp, /* file system mount structure */
2648 xfs_trans_t *tp, /* transaction pointer */
2649 xfs_agnumber_t agno) /* allocation group number */
2650{
2651 xfs_buf_t *bp = NULL;
2652 int error;
2653
2654 error = xfs_ialloc_read_agi(mp, tp, agno, &bp);
2655 if (error)
2656 return error;
2657 if (bp)
2658 xfs_trans_brelse(tp, bp);
2659 return 0;
2660}
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#include "xfs_health.h"
31
32/*
33 * Lookup a record by ino in the btree given by cur.
34 */
35int /* error */
36xfs_inobt_lookup(
37 struct xfs_btree_cur *cur, /* btree cursor */
38 xfs_agino_t ino, /* starting inode of chunk */
39 xfs_lookup_t dir, /* <=, >=, == */
40 int *stat) /* success/failure */
41{
42 cur->bc_rec.i.ir_startino = ino;
43 cur->bc_rec.i.ir_holemask = 0;
44 cur->bc_rec.i.ir_count = 0;
45 cur->bc_rec.i.ir_freecount = 0;
46 cur->bc_rec.i.ir_free = 0;
47 return xfs_btree_lookup(cur, dir, stat);
48}
49
50/*
51 * Update the record referred to by cur to the value given.
52 * This either works (return 0) or gets an EFSCORRUPTED error.
53 */
54STATIC int /* error */
55xfs_inobt_update(
56 struct xfs_btree_cur *cur, /* btree cursor */
57 xfs_inobt_rec_incore_t *irec) /* btree record */
58{
59 union xfs_btree_rec rec;
60
61 rec.inobt.ir_startino = cpu_to_be32(irec->ir_startino);
62 if (xfs_has_sparseinodes(cur->bc_mp)) {
63 rec.inobt.ir_u.sp.ir_holemask = cpu_to_be16(irec->ir_holemask);
64 rec.inobt.ir_u.sp.ir_count = irec->ir_count;
65 rec.inobt.ir_u.sp.ir_freecount = irec->ir_freecount;
66 } else {
67 /* ir_holemask/ir_count not supported on-disk */
68 rec.inobt.ir_u.f.ir_freecount = cpu_to_be32(irec->ir_freecount);
69 }
70 rec.inobt.ir_free = cpu_to_be64(irec->ir_free);
71 return xfs_btree_update(cur, &rec);
72}
73
74/* Convert on-disk btree record to incore inobt record. */
75void
76xfs_inobt_btrec_to_irec(
77 struct xfs_mount *mp,
78 const union xfs_btree_rec *rec,
79 struct xfs_inobt_rec_incore *irec)
80{
81 irec->ir_startino = be32_to_cpu(rec->inobt.ir_startino);
82 if (xfs_has_sparseinodes(mp)) {
83 irec->ir_holemask = be16_to_cpu(rec->inobt.ir_u.sp.ir_holemask);
84 irec->ir_count = rec->inobt.ir_u.sp.ir_count;
85 irec->ir_freecount = rec->inobt.ir_u.sp.ir_freecount;
86 } else {
87 /*
88 * ir_holemask/ir_count not supported on-disk. Fill in hardcoded
89 * values for full inode chunks.
90 */
91 irec->ir_holemask = XFS_INOBT_HOLEMASK_FULL;
92 irec->ir_count = XFS_INODES_PER_CHUNK;
93 irec->ir_freecount =
94 be32_to_cpu(rec->inobt.ir_u.f.ir_freecount);
95 }
96 irec->ir_free = be64_to_cpu(rec->inobt.ir_free);
97}
98
99/* Compute the freecount of an incore inode record. */
100uint8_t
101xfs_inobt_rec_freecount(
102 const struct xfs_inobt_rec_incore *irec)
103{
104 uint64_t realfree = irec->ir_free;
105
106 if (xfs_inobt_issparse(irec->ir_holemask))
107 realfree &= xfs_inobt_irec_to_allocmask(irec);
108 return hweight64(realfree);
109}
110
111/* Simple checks for inode records. */
112xfs_failaddr_t
113xfs_inobt_check_irec(
114 struct xfs_perag *pag,
115 const struct xfs_inobt_rec_incore *irec)
116{
117 /* Record has to be properly aligned within the AG. */
118 if (!xfs_verify_agino(pag, irec->ir_startino))
119 return __this_address;
120 if (!xfs_verify_agino(pag,
121 irec->ir_startino + XFS_INODES_PER_CHUNK - 1))
122 return __this_address;
123 if (irec->ir_count < XFS_INODES_PER_HOLEMASK_BIT ||
124 irec->ir_count > XFS_INODES_PER_CHUNK)
125 return __this_address;
126 if (irec->ir_freecount > XFS_INODES_PER_CHUNK)
127 return __this_address;
128
129 if (xfs_inobt_rec_freecount(irec) != irec->ir_freecount)
130 return __this_address;
131
132 return NULL;
133}
134
135static inline int
136xfs_inobt_complain_bad_rec(
137 struct xfs_btree_cur *cur,
138 xfs_failaddr_t fa,
139 const struct xfs_inobt_rec_incore *irec)
140{
141 struct xfs_mount *mp = cur->bc_mp;
142
143 xfs_warn(mp,
144 "%sbt record corruption in AG %d detected at %pS!",
145 cur->bc_ops->name, cur->bc_ag.pag->pag_agno, fa);
146 xfs_warn(mp,
147"start inode 0x%x, count 0x%x, free 0x%x freemask 0x%llx, holemask 0x%x",
148 irec->ir_startino, irec->ir_count, irec->ir_freecount,
149 irec->ir_free, irec->ir_holemask);
150 xfs_btree_mark_sick(cur);
151 return -EFSCORRUPTED;
152}
153
154/*
155 * Get the data from the pointed-to record.
156 */
157int
158xfs_inobt_get_rec(
159 struct xfs_btree_cur *cur,
160 struct xfs_inobt_rec_incore *irec,
161 int *stat)
162{
163 struct xfs_mount *mp = cur->bc_mp;
164 union xfs_btree_rec *rec;
165 xfs_failaddr_t fa;
166 int error;
167
168 error = xfs_btree_get_rec(cur, &rec, stat);
169 if (error || *stat == 0)
170 return error;
171
172 xfs_inobt_btrec_to_irec(mp, rec, irec);
173 fa = xfs_inobt_check_irec(cur->bc_ag.pag, irec);
174 if (fa)
175 return xfs_inobt_complain_bad_rec(cur, fa, irec);
176
177 return 0;
178}
179
180/*
181 * Insert a single inobt record. Cursor must already point to desired location.
182 */
183int
184xfs_inobt_insert_rec(
185 struct xfs_btree_cur *cur,
186 uint16_t holemask,
187 uint8_t count,
188 int32_t freecount,
189 xfs_inofree_t free,
190 int *stat)
191{
192 cur->bc_rec.i.ir_holemask = holemask;
193 cur->bc_rec.i.ir_count = count;
194 cur->bc_rec.i.ir_freecount = freecount;
195 cur->bc_rec.i.ir_free = free;
196 return xfs_btree_insert(cur, stat);
197}
198
199/*
200 * Insert records describing a newly allocated inode chunk into the inobt.
201 */
202STATIC int
203xfs_inobt_insert(
204 struct xfs_perag *pag,
205 struct xfs_trans *tp,
206 struct xfs_buf *agbp,
207 xfs_agino_t newino,
208 xfs_agino_t newlen,
209 bool is_finobt)
210{
211 struct xfs_btree_cur *cur;
212 xfs_agino_t thisino;
213 int i;
214 int error;
215
216 if (is_finobt)
217 cur = xfs_finobt_init_cursor(pag, tp, agbp);
218 else
219 cur = xfs_inobt_init_cursor(pag, tp, agbp);
220
221 for (thisino = newino;
222 thisino < newino + newlen;
223 thisino += XFS_INODES_PER_CHUNK) {
224 error = xfs_inobt_lookup(cur, thisino, XFS_LOOKUP_EQ, &i);
225 if (error) {
226 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
227 return error;
228 }
229 ASSERT(i == 0);
230
231 error = xfs_inobt_insert_rec(cur, XFS_INOBT_HOLEMASK_FULL,
232 XFS_INODES_PER_CHUNK,
233 XFS_INODES_PER_CHUNK,
234 XFS_INOBT_ALL_FREE, &i);
235 if (error) {
236 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
237 return error;
238 }
239 ASSERT(i == 1);
240 }
241
242 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
243
244 return 0;
245}
246
247/*
248 * Verify that the number of free inodes in the AGI is correct.
249 */
250#ifdef DEBUG
251static int
252xfs_check_agi_freecount(
253 struct xfs_btree_cur *cur)
254{
255 if (cur->bc_nlevels == 1) {
256 xfs_inobt_rec_incore_t rec;
257 int freecount = 0;
258 int error;
259 int i;
260
261 error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
262 if (error)
263 return error;
264
265 do {
266 error = xfs_inobt_get_rec(cur, &rec, &i);
267 if (error)
268 return error;
269
270 if (i) {
271 freecount += rec.ir_freecount;
272 error = xfs_btree_increment(cur, 0, &i);
273 if (error)
274 return error;
275 }
276 } while (i == 1);
277
278 if (!xfs_is_shutdown(cur->bc_mp))
279 ASSERT(freecount == cur->bc_ag.pag->pagi_freecount);
280 }
281 return 0;
282}
283#else
284#define xfs_check_agi_freecount(cur) 0
285#endif
286
287/*
288 * Initialise a new set of inodes. When called without a transaction context
289 * (e.g. from recovery) we initiate a delayed write of the inode buffers rather
290 * than logging them (which in a transaction context puts them into the AIL
291 * for writeback rather than the xfsbufd queue).
292 */
293int
294xfs_ialloc_inode_init(
295 struct xfs_mount *mp,
296 struct xfs_trans *tp,
297 struct list_head *buffer_list,
298 int icount,
299 xfs_agnumber_t agno,
300 xfs_agblock_t agbno,
301 xfs_agblock_t length,
302 unsigned int gen)
303{
304 struct xfs_buf *fbuf;
305 struct xfs_dinode *free;
306 int nbufs;
307 int version;
308 int i, j;
309 xfs_daddr_t d;
310 xfs_ino_t ino = 0;
311 int error;
312
313 /*
314 * Loop over the new block(s), filling in the inodes. For small block
315 * sizes, manipulate the inodes in buffers which are multiples of the
316 * blocks size.
317 */
318 nbufs = length / M_IGEO(mp)->blocks_per_cluster;
319
320 /*
321 * Figure out what version number to use in the inodes we create. If
322 * the superblock version has caught up to the one that supports the new
323 * inode format, then use the new inode version. Otherwise use the old
324 * version so that old kernels will continue to be able to use the file
325 * system.
326 *
327 * For v3 inodes, we also need to write the inode number into the inode,
328 * so calculate the first inode number of the chunk here as
329 * XFS_AGB_TO_AGINO() only works within a filesystem block, not
330 * across multiple filesystem blocks (such as a cluster) and so cannot
331 * be used in the cluster buffer loop below.
332 *
333 * Further, because we are writing the inode directly into the buffer
334 * and calculating a CRC on the entire inode, we have ot log the entire
335 * inode so that the entire range the CRC covers is present in the log.
336 * That means for v3 inode we log the entire buffer rather than just the
337 * inode cores.
338 */
339 if (xfs_has_v3inodes(mp)) {
340 version = 3;
341 ino = XFS_AGINO_TO_INO(mp, agno, XFS_AGB_TO_AGINO(mp, agbno));
342
343 /*
344 * log the initialisation that is about to take place as an
345 * logical operation. This means the transaction does not
346 * need to log the physical changes to the inode buffers as log
347 * recovery will know what initialisation is actually needed.
348 * Hence we only need to log the buffers as "ordered" buffers so
349 * they track in the AIL as if they were physically logged.
350 */
351 if (tp)
352 xfs_icreate_log(tp, agno, agbno, icount,
353 mp->m_sb.sb_inodesize, length, gen);
354 } else
355 version = 2;
356
357 for (j = 0; j < nbufs; j++) {
358 /*
359 * Get the block.
360 */
361 d = XFS_AGB_TO_DADDR(mp, agno, agbno +
362 (j * M_IGEO(mp)->blocks_per_cluster));
363 error = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
364 mp->m_bsize * M_IGEO(mp)->blocks_per_cluster,
365 XBF_UNMAPPED, &fbuf);
366 if (error)
367 return error;
368
369 /* Initialize the inode buffers and log them appropriately. */
370 fbuf->b_ops = &xfs_inode_buf_ops;
371 xfs_buf_zero(fbuf, 0, BBTOB(fbuf->b_length));
372 for (i = 0; i < M_IGEO(mp)->inodes_per_cluster; i++) {
373 int ioffset = i << mp->m_sb.sb_inodelog;
374
375 free = xfs_make_iptr(mp, fbuf, i);
376 free->di_magic = cpu_to_be16(XFS_DINODE_MAGIC);
377 free->di_version = version;
378 free->di_gen = cpu_to_be32(gen);
379 free->di_next_unlinked = cpu_to_be32(NULLAGINO);
380
381 if (version == 3) {
382 free->di_ino = cpu_to_be64(ino);
383 ino++;
384 uuid_copy(&free->di_uuid,
385 &mp->m_sb.sb_meta_uuid);
386 xfs_dinode_calc_crc(mp, free);
387 } else if (tp) {
388 /* just log the inode core */
389 xfs_trans_log_buf(tp, fbuf, ioffset,
390 ioffset + XFS_DINODE_SIZE(mp) - 1);
391 }
392 }
393
394 if (tp) {
395 /*
396 * Mark the buffer as an inode allocation buffer so it
397 * sticks in AIL at the point of this allocation
398 * transaction. This ensures the they are on disk before
399 * the tail of the log can be moved past this
400 * transaction (i.e. by preventing relogging from moving
401 * it forward in the log).
402 */
403 xfs_trans_inode_alloc_buf(tp, fbuf);
404 if (version == 3) {
405 /*
406 * Mark the buffer as ordered so that they are
407 * not physically logged in the transaction but
408 * still tracked in the AIL as part of the
409 * transaction and pin the log appropriately.
410 */
411 xfs_trans_ordered_buf(tp, fbuf);
412 }
413 } else {
414 fbuf->b_flags |= XBF_DONE;
415 xfs_buf_delwri_queue(fbuf, buffer_list);
416 xfs_buf_relse(fbuf);
417 }
418 }
419 return 0;
420}
421
422/*
423 * Align startino and allocmask for a recently allocated sparse chunk such that
424 * they are fit for insertion (or merge) into the on-disk inode btrees.
425 *
426 * Background:
427 *
428 * When enabled, sparse inode support increases the inode alignment from cluster
429 * size to inode chunk size. This means that the minimum range between two
430 * non-adjacent inode records in the inobt is large enough for a full inode
431 * record. This allows for cluster sized, cluster aligned block allocation
432 * without need to worry about whether the resulting inode record overlaps with
433 * another record in the tree. Without this basic rule, we would have to deal
434 * with the consequences of overlap by potentially undoing recent allocations in
435 * the inode allocation codepath.
436 *
437 * Because of this alignment rule (which is enforced on mount), there are two
438 * inobt possibilities for newly allocated sparse chunks. One is that the
439 * aligned inode record for the chunk covers a range of inodes not already
440 * covered in the inobt (i.e., it is safe to insert a new sparse record). The
441 * other is that a record already exists at the aligned startino that considers
442 * the newly allocated range as sparse. In the latter case, record content is
443 * merged in hope that sparse inode chunks fill to full chunks over time.
444 */
445STATIC void
446xfs_align_sparse_ino(
447 struct xfs_mount *mp,
448 xfs_agino_t *startino,
449 uint16_t *allocmask)
450{
451 xfs_agblock_t agbno;
452 xfs_agblock_t mod;
453 int offset;
454
455 agbno = XFS_AGINO_TO_AGBNO(mp, *startino);
456 mod = agbno % mp->m_sb.sb_inoalignmt;
457 if (!mod)
458 return;
459
460 /* calculate the inode offset and align startino */
461 offset = XFS_AGB_TO_AGINO(mp, mod);
462 *startino -= offset;
463
464 /*
465 * Since startino has been aligned down, left shift allocmask such that
466 * it continues to represent the same physical inodes relative to the
467 * new startino.
468 */
469 *allocmask <<= offset / XFS_INODES_PER_HOLEMASK_BIT;
470}
471
472/*
473 * Determine whether the source inode record can merge into the target. Both
474 * records must be sparse, the inode ranges must match and there must be no
475 * allocation overlap between the records.
476 */
477STATIC bool
478__xfs_inobt_can_merge(
479 struct xfs_inobt_rec_incore *trec, /* tgt record */
480 struct xfs_inobt_rec_incore *srec) /* src record */
481{
482 uint64_t talloc;
483 uint64_t salloc;
484
485 /* records must cover the same inode range */
486 if (trec->ir_startino != srec->ir_startino)
487 return false;
488
489 /* both records must be sparse */
490 if (!xfs_inobt_issparse(trec->ir_holemask) ||
491 !xfs_inobt_issparse(srec->ir_holemask))
492 return false;
493
494 /* both records must track some inodes */
495 if (!trec->ir_count || !srec->ir_count)
496 return false;
497
498 /* can't exceed capacity of a full record */
499 if (trec->ir_count + srec->ir_count > XFS_INODES_PER_CHUNK)
500 return false;
501
502 /* verify there is no allocation overlap */
503 talloc = xfs_inobt_irec_to_allocmask(trec);
504 salloc = xfs_inobt_irec_to_allocmask(srec);
505 if (talloc & salloc)
506 return false;
507
508 return true;
509}
510
511/*
512 * Merge the source inode record into the target. The caller must call
513 * __xfs_inobt_can_merge() to ensure the merge is valid.
514 */
515STATIC void
516__xfs_inobt_rec_merge(
517 struct xfs_inobt_rec_incore *trec, /* target */
518 struct xfs_inobt_rec_incore *srec) /* src */
519{
520 ASSERT(trec->ir_startino == srec->ir_startino);
521
522 /* combine the counts */
523 trec->ir_count += srec->ir_count;
524 trec->ir_freecount += srec->ir_freecount;
525
526 /*
527 * Merge the holemask and free mask. For both fields, 0 bits refer to
528 * allocated inodes. We combine the allocated ranges with bitwise AND.
529 */
530 trec->ir_holemask &= srec->ir_holemask;
531 trec->ir_free &= srec->ir_free;
532}
533
534/*
535 * Insert a new sparse inode chunk into the associated inode allocation btree.
536 * The inode record for the sparse chunk is pre-aligned to a startino that
537 * should match any pre-existing sparse inode record in the tree. This allows
538 * sparse chunks to fill over time.
539 *
540 * If no preexisting record exists, the provided record is inserted.
541 * If there is a preexisting record, the provided record is merged with the
542 * existing record and updated in place. The merged record is returned in nrec.
543 *
544 * It is considered corruption if a merge is requested and not possible. Given
545 * the sparse inode alignment constraints, this should never happen.
546 */
547STATIC int
548xfs_inobt_insert_sprec(
549 struct xfs_perag *pag,
550 struct xfs_trans *tp,
551 struct xfs_buf *agbp,
552 struct xfs_inobt_rec_incore *nrec) /* in/out: new/merged rec. */
553{
554 struct xfs_mount *mp = pag->pag_mount;
555 struct xfs_btree_cur *cur;
556 int error;
557 int i;
558 struct xfs_inobt_rec_incore rec;
559
560 cur = xfs_inobt_init_cursor(pag, tp, agbp);
561
562 /* the new record is pre-aligned so we know where to look */
563 error = xfs_inobt_lookup(cur, nrec->ir_startino, XFS_LOOKUP_EQ, &i);
564 if (error)
565 goto error;
566 /* if nothing there, insert a new record and return */
567 if (i == 0) {
568 error = xfs_inobt_insert_rec(cur, nrec->ir_holemask,
569 nrec->ir_count, nrec->ir_freecount,
570 nrec->ir_free, &i);
571 if (error)
572 goto error;
573 if (XFS_IS_CORRUPT(mp, i != 1)) {
574 xfs_btree_mark_sick(cur);
575 error = -EFSCORRUPTED;
576 goto error;
577 }
578
579 goto out;
580 }
581
582 /*
583 * A record exists at this startino. Merge the records.
584 */
585 error = xfs_inobt_get_rec(cur, &rec, &i);
586 if (error)
587 goto error;
588 if (XFS_IS_CORRUPT(mp, i != 1)) {
589 xfs_btree_mark_sick(cur);
590 error = -EFSCORRUPTED;
591 goto error;
592 }
593 if (XFS_IS_CORRUPT(mp, rec.ir_startino != nrec->ir_startino)) {
594 xfs_btree_mark_sick(cur);
595 error = -EFSCORRUPTED;
596 goto error;
597 }
598
599 /*
600 * This should never fail. If we have coexisting records that
601 * cannot merge, something is seriously wrong.
602 */
603 if (XFS_IS_CORRUPT(mp, !__xfs_inobt_can_merge(nrec, &rec))) {
604 xfs_btree_mark_sick(cur);
605 error = -EFSCORRUPTED;
606 goto error;
607 }
608
609 trace_xfs_irec_merge_pre(mp, pag->pag_agno, rec.ir_startino,
610 rec.ir_holemask, nrec->ir_startino,
611 nrec->ir_holemask);
612
613 /* merge to nrec to output the updated record */
614 __xfs_inobt_rec_merge(nrec, &rec);
615
616 trace_xfs_irec_merge_post(mp, pag->pag_agno, nrec->ir_startino,
617 nrec->ir_holemask);
618
619 error = xfs_inobt_rec_check_count(mp, nrec);
620 if (error)
621 goto error;
622
623 error = xfs_inobt_update(cur, nrec);
624 if (error)
625 goto error;
626
627out:
628 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
629 return 0;
630error:
631 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
632 return error;
633}
634
635/*
636 * Insert a new sparse inode chunk into the free inode btree. The inode
637 * record for the sparse chunk is pre-aligned to a startino that should match
638 * any pre-existing sparse inode record in the tree. This allows sparse chunks
639 * to fill over time.
640 *
641 * The new record is always inserted, overwriting a pre-existing record if
642 * there is one.
643 */
644STATIC int
645xfs_finobt_insert_sprec(
646 struct xfs_perag *pag,
647 struct xfs_trans *tp,
648 struct xfs_buf *agbp,
649 struct xfs_inobt_rec_incore *nrec) /* in/out: new rec. */
650{
651 struct xfs_mount *mp = pag->pag_mount;
652 struct xfs_btree_cur *cur;
653 int error;
654 int i;
655
656 cur = xfs_finobt_init_cursor(pag, tp, agbp);
657
658 /* the new record is pre-aligned so we know where to look */
659 error = xfs_inobt_lookup(cur, nrec->ir_startino, XFS_LOOKUP_EQ, &i);
660 if (error)
661 goto error;
662 /* if nothing there, insert a new record and return */
663 if (i == 0) {
664 error = xfs_inobt_insert_rec(cur, nrec->ir_holemask,
665 nrec->ir_count, nrec->ir_freecount,
666 nrec->ir_free, &i);
667 if (error)
668 goto error;
669 if (XFS_IS_CORRUPT(mp, i != 1)) {
670 xfs_btree_mark_sick(cur);
671 error = -EFSCORRUPTED;
672 goto error;
673 }
674 } else {
675 error = xfs_inobt_update(cur, nrec);
676 if (error)
677 goto error;
678 }
679
680 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
681 return 0;
682error:
683 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
684 return error;
685}
686
687
688/*
689 * Allocate new inodes in the allocation group specified by agbp. Returns 0 if
690 * inodes were allocated in this AG; -EAGAIN if there was no space in this AG so
691 * the caller knows it can try another AG, a hard -ENOSPC when over the maximum
692 * inode count threshold, or the usual negative error code for other errors.
693 */
694STATIC int
695xfs_ialloc_ag_alloc(
696 struct xfs_perag *pag,
697 struct xfs_trans *tp,
698 struct xfs_buf *agbp)
699{
700 struct xfs_agi *agi;
701 struct xfs_alloc_arg args;
702 int error;
703 xfs_agino_t newino; /* new first inode's number */
704 xfs_agino_t newlen; /* new number of inodes */
705 int isaligned = 0; /* inode allocation at stripe */
706 /* unit boundary */
707 /* init. to full chunk */
708 struct xfs_inobt_rec_incore rec;
709 struct xfs_ino_geometry *igeo = M_IGEO(tp->t_mountp);
710 uint16_t allocmask = (uint16_t) -1;
711 int do_sparse = 0;
712
713 memset(&args, 0, sizeof(args));
714 args.tp = tp;
715 args.mp = tp->t_mountp;
716 args.fsbno = NULLFSBLOCK;
717 args.oinfo = XFS_RMAP_OINFO_INODES;
718 args.pag = pag;
719
720#ifdef DEBUG
721 /* randomly do sparse inode allocations */
722 if (xfs_has_sparseinodes(tp->t_mountp) &&
723 igeo->ialloc_min_blks < igeo->ialloc_blks)
724 do_sparse = get_random_u32_below(2);
725#endif
726
727 /*
728 * Locking will ensure that we don't have two callers in here
729 * at one time.
730 */
731 newlen = igeo->ialloc_inos;
732 if (igeo->maxicount &&
733 percpu_counter_read_positive(&args.mp->m_icount) + newlen >
734 igeo->maxicount)
735 return -ENOSPC;
736 args.minlen = args.maxlen = igeo->ialloc_blks;
737 /*
738 * First try to allocate inodes contiguous with the last-allocated
739 * chunk of inodes. If the filesystem is striped, this will fill
740 * an entire stripe unit with inodes.
741 */
742 agi = agbp->b_addr;
743 newino = be32_to_cpu(agi->agi_newino);
744 args.agbno = XFS_AGINO_TO_AGBNO(args.mp, newino) +
745 igeo->ialloc_blks;
746 if (do_sparse)
747 goto sparse_alloc;
748 if (likely(newino != NULLAGINO &&
749 (args.agbno < be32_to_cpu(agi->agi_length)))) {
750 args.prod = 1;
751
752 /*
753 * We need to take into account alignment here to ensure that
754 * we don't modify the free list if we fail to have an exact
755 * block. If we don't have an exact match, and every oher
756 * attempt allocation attempt fails, we'll end up cancelling
757 * a dirty transaction and shutting down.
758 *
759 * For an exact allocation, alignment must be 1,
760 * however we need to take cluster alignment into account when
761 * fixing up the freelist. Use the minalignslop field to
762 * indicate that extra blocks might be required for alignment,
763 * but not to use them in the actual exact allocation.
764 */
765 args.alignment = 1;
766 args.minalignslop = igeo->cluster_align - 1;
767
768 /* Allow space for the inode btree to split. */
769 args.minleft = igeo->inobt_maxlevels;
770 error = xfs_alloc_vextent_exact_bno(&args,
771 XFS_AGB_TO_FSB(args.mp, pag->pag_agno,
772 args.agbno));
773 if (error)
774 return error;
775
776 /*
777 * This request might have dirtied the transaction if the AG can
778 * satisfy the request, but the exact block was not available.
779 * If the allocation did fail, subsequent requests will relax
780 * the exact agbno requirement and increase the alignment
781 * instead. It is critical that the total size of the request
782 * (len + alignment + slop) does not increase from this point
783 * on, so reset minalignslop to ensure it is not included in
784 * subsequent requests.
785 */
786 args.minalignslop = 0;
787 }
788
789 if (unlikely(args.fsbno == NULLFSBLOCK)) {
790 /*
791 * Set the alignment for the allocation.
792 * If stripe alignment is turned on then align at stripe unit
793 * boundary.
794 * If the cluster size is smaller than a filesystem block
795 * then we're doing I/O for inodes in filesystem block size
796 * pieces, so don't need alignment anyway.
797 */
798 isaligned = 0;
799 if (igeo->ialloc_align) {
800 ASSERT(!xfs_has_noalign(args.mp));
801 args.alignment = args.mp->m_dalign;
802 isaligned = 1;
803 } else
804 args.alignment = igeo->cluster_align;
805 /*
806 * Allocate a fixed-size extent of inodes.
807 */
808 args.prod = 1;
809 /*
810 * Allow space for the inode btree to split.
811 */
812 args.minleft = igeo->inobt_maxlevels;
813 error = xfs_alloc_vextent_near_bno(&args,
814 XFS_AGB_TO_FSB(args.mp, pag->pag_agno,
815 be32_to_cpu(agi->agi_root)));
816 if (error)
817 return error;
818 }
819
820 /*
821 * If stripe alignment is turned on, then try again with cluster
822 * alignment.
823 */
824 if (isaligned && args.fsbno == NULLFSBLOCK) {
825 args.alignment = igeo->cluster_align;
826 error = xfs_alloc_vextent_near_bno(&args,
827 XFS_AGB_TO_FSB(args.mp, pag->pag_agno,
828 be32_to_cpu(agi->agi_root)));
829 if (error)
830 return error;
831 }
832
833 /*
834 * Finally, try a sparse allocation if the filesystem supports it and
835 * the sparse allocation length is smaller than a full chunk.
836 */
837 if (xfs_has_sparseinodes(args.mp) &&
838 igeo->ialloc_min_blks < igeo->ialloc_blks &&
839 args.fsbno == NULLFSBLOCK) {
840sparse_alloc:
841 args.alignment = args.mp->m_sb.sb_spino_align;
842 args.prod = 1;
843
844 args.minlen = igeo->ialloc_min_blks;
845 args.maxlen = args.minlen;
846
847 /*
848 * The inode record will be aligned to full chunk size. We must
849 * prevent sparse allocation from AG boundaries that result in
850 * invalid inode records, such as records that start at agbno 0
851 * or extend beyond the AG.
852 *
853 * Set min agbno to the first aligned, non-zero agbno and max to
854 * the last aligned agbno that is at least one full chunk from
855 * the end of the AG.
856 */
857 args.min_agbno = args.mp->m_sb.sb_inoalignmt;
858 args.max_agbno = round_down(args.mp->m_sb.sb_agblocks,
859 args.mp->m_sb.sb_inoalignmt) -
860 igeo->ialloc_blks;
861
862 error = xfs_alloc_vextent_near_bno(&args,
863 XFS_AGB_TO_FSB(args.mp, pag->pag_agno,
864 be32_to_cpu(agi->agi_root)));
865 if (error)
866 return error;
867
868 newlen = XFS_AGB_TO_AGINO(args.mp, args.len);
869 ASSERT(newlen <= XFS_INODES_PER_CHUNK);
870 allocmask = (1 << (newlen / XFS_INODES_PER_HOLEMASK_BIT)) - 1;
871 }
872
873 if (args.fsbno == NULLFSBLOCK)
874 return -EAGAIN;
875
876 ASSERT(args.len == args.minlen);
877
878 /*
879 * Stamp and write the inode buffers.
880 *
881 * Seed the new inode cluster with a random generation number. This
882 * prevents short-term reuse of generation numbers if a chunk is
883 * freed and then immediately reallocated. We use random numbers
884 * rather than a linear progression to prevent the next generation
885 * number from being easily guessable.
886 */
887 error = xfs_ialloc_inode_init(args.mp, tp, NULL, newlen, pag->pag_agno,
888 args.agbno, args.len, get_random_u32());
889
890 if (error)
891 return error;
892 /*
893 * Convert the results.
894 */
895 newino = XFS_AGB_TO_AGINO(args.mp, args.agbno);
896
897 if (xfs_inobt_issparse(~allocmask)) {
898 /*
899 * We've allocated a sparse chunk. Align the startino and mask.
900 */
901 xfs_align_sparse_ino(args.mp, &newino, &allocmask);
902
903 rec.ir_startino = newino;
904 rec.ir_holemask = ~allocmask;
905 rec.ir_count = newlen;
906 rec.ir_freecount = newlen;
907 rec.ir_free = XFS_INOBT_ALL_FREE;
908
909 /*
910 * Insert the sparse record into the inobt and allow for a merge
911 * if necessary. If a merge does occur, rec is updated to the
912 * merged record.
913 */
914 error = xfs_inobt_insert_sprec(pag, tp, agbp, &rec);
915 if (error == -EFSCORRUPTED) {
916 xfs_alert(args.mp,
917 "invalid sparse inode record: ino 0x%llx holemask 0x%x count %u",
918 XFS_AGINO_TO_INO(args.mp, pag->pag_agno,
919 rec.ir_startino),
920 rec.ir_holemask, rec.ir_count);
921 xfs_force_shutdown(args.mp, SHUTDOWN_CORRUPT_INCORE);
922 }
923 if (error)
924 return error;
925
926 /*
927 * We can't merge the part we've just allocated as for the inobt
928 * due to finobt semantics. The original record may or may not
929 * exist independent of whether physical inodes exist in this
930 * sparse chunk.
931 *
932 * We must update the finobt record based on the inobt record.
933 * rec contains the fully merged and up to date inobt record
934 * from the previous call. Set merge false to replace any
935 * existing record with this one.
936 */
937 if (xfs_has_finobt(args.mp)) {
938 error = xfs_finobt_insert_sprec(pag, tp, agbp, &rec);
939 if (error)
940 return error;
941 }
942 } else {
943 /* full chunk - insert new records to both btrees */
944 error = xfs_inobt_insert(pag, tp, agbp, newino, newlen, false);
945 if (error)
946 return error;
947
948 if (xfs_has_finobt(args.mp)) {
949 error = xfs_inobt_insert(pag, tp, agbp, newino,
950 newlen, true);
951 if (error)
952 return error;
953 }
954 }
955
956 /*
957 * Update AGI counts and newino.
958 */
959 be32_add_cpu(&agi->agi_count, newlen);
960 be32_add_cpu(&agi->agi_freecount, newlen);
961 pag->pagi_freecount += newlen;
962 pag->pagi_count += newlen;
963 agi->agi_newino = cpu_to_be32(newino);
964
965 /*
966 * Log allocation group header fields
967 */
968 xfs_ialloc_log_agi(tp, agbp,
969 XFS_AGI_COUNT | XFS_AGI_FREECOUNT | XFS_AGI_NEWINO);
970 /*
971 * Modify/log superblock values for inode count and inode free count.
972 */
973 xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, (long)newlen);
974 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, (long)newlen);
975 return 0;
976}
977
978/*
979 * Try to retrieve the next record to the left/right from the current one.
980 */
981STATIC int
982xfs_ialloc_next_rec(
983 struct xfs_btree_cur *cur,
984 xfs_inobt_rec_incore_t *rec,
985 int *done,
986 int left)
987{
988 int error;
989 int i;
990
991 if (left)
992 error = xfs_btree_decrement(cur, 0, &i);
993 else
994 error = xfs_btree_increment(cur, 0, &i);
995
996 if (error)
997 return error;
998 *done = !i;
999 if (i) {
1000 error = xfs_inobt_get_rec(cur, rec, &i);
1001 if (error)
1002 return error;
1003 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1004 xfs_btree_mark_sick(cur);
1005 return -EFSCORRUPTED;
1006 }
1007 }
1008
1009 return 0;
1010}
1011
1012STATIC int
1013xfs_ialloc_get_rec(
1014 struct xfs_btree_cur *cur,
1015 xfs_agino_t agino,
1016 xfs_inobt_rec_incore_t *rec,
1017 int *done)
1018{
1019 int error;
1020 int i;
1021
1022 error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_EQ, &i);
1023 if (error)
1024 return error;
1025 *done = !i;
1026 if (i) {
1027 error = xfs_inobt_get_rec(cur, rec, &i);
1028 if (error)
1029 return error;
1030 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1031 xfs_btree_mark_sick(cur);
1032 return -EFSCORRUPTED;
1033 }
1034 }
1035
1036 return 0;
1037}
1038
1039/*
1040 * Return the offset of the first free inode in the record. If the inode chunk
1041 * is sparsely allocated, we convert the record holemask to inode granularity
1042 * and mask off the unallocated regions from the inode free mask.
1043 */
1044STATIC int
1045xfs_inobt_first_free_inode(
1046 struct xfs_inobt_rec_incore *rec)
1047{
1048 xfs_inofree_t realfree;
1049
1050 /* if there are no holes, return the first available offset */
1051 if (!xfs_inobt_issparse(rec->ir_holemask))
1052 return xfs_lowbit64(rec->ir_free);
1053
1054 realfree = xfs_inobt_irec_to_allocmask(rec);
1055 realfree &= rec->ir_free;
1056
1057 return xfs_lowbit64(realfree);
1058}
1059
1060/*
1061 * Allocate an inode using the inobt-only algorithm.
1062 */
1063STATIC int
1064xfs_dialloc_ag_inobt(
1065 struct xfs_perag *pag,
1066 struct xfs_trans *tp,
1067 struct xfs_buf *agbp,
1068 xfs_ino_t parent,
1069 xfs_ino_t *inop)
1070{
1071 struct xfs_mount *mp = tp->t_mountp;
1072 struct xfs_agi *agi = agbp->b_addr;
1073 xfs_agnumber_t pagno = XFS_INO_TO_AGNO(mp, parent);
1074 xfs_agino_t pagino = XFS_INO_TO_AGINO(mp, parent);
1075 struct xfs_btree_cur *cur, *tcur;
1076 struct xfs_inobt_rec_incore rec, trec;
1077 xfs_ino_t ino;
1078 int error;
1079 int offset;
1080 int i, j;
1081 int searchdistance = 10;
1082
1083 ASSERT(xfs_perag_initialised_agi(pag));
1084 ASSERT(xfs_perag_allows_inodes(pag));
1085 ASSERT(pag->pagi_freecount > 0);
1086
1087 restart_pagno:
1088 cur = xfs_inobt_init_cursor(pag, tp, agbp);
1089 /*
1090 * If pagino is 0 (this is the root inode allocation) use newino.
1091 * This must work because we've just allocated some.
1092 */
1093 if (!pagino)
1094 pagino = be32_to_cpu(agi->agi_newino);
1095
1096 error = xfs_check_agi_freecount(cur);
1097 if (error)
1098 goto error0;
1099
1100 /*
1101 * If in the same AG as the parent, try to get near the parent.
1102 */
1103 if (pagno == pag->pag_agno) {
1104 int doneleft; /* done, to the left */
1105 int doneright; /* done, to the right */
1106
1107 error = xfs_inobt_lookup(cur, pagino, XFS_LOOKUP_LE, &i);
1108 if (error)
1109 goto error0;
1110 if (XFS_IS_CORRUPT(mp, i != 1)) {
1111 xfs_btree_mark_sick(cur);
1112 error = -EFSCORRUPTED;
1113 goto error0;
1114 }
1115
1116 error = xfs_inobt_get_rec(cur, &rec, &j);
1117 if (error)
1118 goto error0;
1119 if (XFS_IS_CORRUPT(mp, j != 1)) {
1120 xfs_btree_mark_sick(cur);
1121 error = -EFSCORRUPTED;
1122 goto error0;
1123 }
1124
1125 if (rec.ir_freecount > 0) {
1126 /*
1127 * Found a free inode in the same chunk
1128 * as the parent, done.
1129 */
1130 goto alloc_inode;
1131 }
1132
1133
1134 /*
1135 * In the same AG as parent, but parent's chunk is full.
1136 */
1137
1138 /* duplicate the cursor, search left & right simultaneously */
1139 error = xfs_btree_dup_cursor(cur, &tcur);
1140 if (error)
1141 goto error0;
1142
1143 /*
1144 * Skip to last blocks looked up if same parent inode.
1145 */
1146 if (pagino != NULLAGINO &&
1147 pag->pagl_pagino == pagino &&
1148 pag->pagl_leftrec != NULLAGINO &&
1149 pag->pagl_rightrec != NULLAGINO) {
1150 error = xfs_ialloc_get_rec(tcur, pag->pagl_leftrec,
1151 &trec, &doneleft);
1152 if (error)
1153 goto error1;
1154
1155 error = xfs_ialloc_get_rec(cur, pag->pagl_rightrec,
1156 &rec, &doneright);
1157 if (error)
1158 goto error1;
1159 } else {
1160 /* search left with tcur, back up 1 record */
1161 error = xfs_ialloc_next_rec(tcur, &trec, &doneleft, 1);
1162 if (error)
1163 goto error1;
1164
1165 /* search right with cur, go forward 1 record. */
1166 error = xfs_ialloc_next_rec(cur, &rec, &doneright, 0);
1167 if (error)
1168 goto error1;
1169 }
1170
1171 /*
1172 * Loop until we find an inode chunk with a free inode.
1173 */
1174 while (--searchdistance > 0 && (!doneleft || !doneright)) {
1175 int useleft; /* using left inode chunk this time */
1176
1177 /* figure out the closer block if both are valid. */
1178 if (!doneleft && !doneright) {
1179 useleft = pagino -
1180 (trec.ir_startino + XFS_INODES_PER_CHUNK - 1) <
1181 rec.ir_startino - pagino;
1182 } else {
1183 useleft = !doneleft;
1184 }
1185
1186 /* free inodes to the left? */
1187 if (useleft && trec.ir_freecount) {
1188 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1189 cur = tcur;
1190
1191 pag->pagl_leftrec = trec.ir_startino;
1192 pag->pagl_rightrec = rec.ir_startino;
1193 pag->pagl_pagino = pagino;
1194 rec = trec;
1195 goto alloc_inode;
1196 }
1197
1198 /* free inodes to the right? */
1199 if (!useleft && rec.ir_freecount) {
1200 xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1201
1202 pag->pagl_leftrec = trec.ir_startino;
1203 pag->pagl_rightrec = rec.ir_startino;
1204 pag->pagl_pagino = pagino;
1205 goto alloc_inode;
1206 }
1207
1208 /* get next record to check */
1209 if (useleft) {
1210 error = xfs_ialloc_next_rec(tcur, &trec,
1211 &doneleft, 1);
1212 } else {
1213 error = xfs_ialloc_next_rec(cur, &rec,
1214 &doneright, 0);
1215 }
1216 if (error)
1217 goto error1;
1218 }
1219
1220 if (searchdistance <= 0) {
1221 /*
1222 * Not in range - save last search
1223 * location and allocate a new inode
1224 */
1225 xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1226 pag->pagl_leftrec = trec.ir_startino;
1227 pag->pagl_rightrec = rec.ir_startino;
1228 pag->pagl_pagino = pagino;
1229
1230 } else {
1231 /*
1232 * We've reached the end of the btree. because
1233 * we are only searching a small chunk of the
1234 * btree each search, there is obviously free
1235 * inodes closer to the parent inode than we
1236 * are now. restart the search again.
1237 */
1238 pag->pagl_pagino = NULLAGINO;
1239 pag->pagl_leftrec = NULLAGINO;
1240 pag->pagl_rightrec = NULLAGINO;
1241 xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1242 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1243 goto restart_pagno;
1244 }
1245 }
1246
1247 /*
1248 * In a different AG from the parent.
1249 * See if the most recently allocated block has any free.
1250 */
1251 if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1252 error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1253 XFS_LOOKUP_EQ, &i);
1254 if (error)
1255 goto error0;
1256
1257 if (i == 1) {
1258 error = xfs_inobt_get_rec(cur, &rec, &j);
1259 if (error)
1260 goto error0;
1261
1262 if (j == 1 && rec.ir_freecount > 0) {
1263 /*
1264 * The last chunk allocated in the group
1265 * still has a free inode.
1266 */
1267 goto alloc_inode;
1268 }
1269 }
1270 }
1271
1272 /*
1273 * None left in the last group, search the whole AG
1274 */
1275 error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1276 if (error)
1277 goto error0;
1278 if (XFS_IS_CORRUPT(mp, i != 1)) {
1279 xfs_btree_mark_sick(cur);
1280 error = -EFSCORRUPTED;
1281 goto error0;
1282 }
1283
1284 for (;;) {
1285 error = xfs_inobt_get_rec(cur, &rec, &i);
1286 if (error)
1287 goto error0;
1288 if (XFS_IS_CORRUPT(mp, i != 1)) {
1289 xfs_btree_mark_sick(cur);
1290 error = -EFSCORRUPTED;
1291 goto error0;
1292 }
1293 if (rec.ir_freecount > 0)
1294 break;
1295 error = xfs_btree_increment(cur, 0, &i);
1296 if (error)
1297 goto error0;
1298 if (XFS_IS_CORRUPT(mp, i != 1)) {
1299 xfs_btree_mark_sick(cur);
1300 error = -EFSCORRUPTED;
1301 goto error0;
1302 }
1303 }
1304
1305alloc_inode:
1306 offset = xfs_inobt_first_free_inode(&rec);
1307 ASSERT(offset >= 0);
1308 ASSERT(offset < XFS_INODES_PER_CHUNK);
1309 ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1310 XFS_INODES_PER_CHUNK) == 0);
1311 ino = XFS_AGINO_TO_INO(mp, pag->pag_agno, rec.ir_startino + offset);
1312 rec.ir_free &= ~XFS_INOBT_MASK(offset);
1313 rec.ir_freecount--;
1314 error = xfs_inobt_update(cur, &rec);
1315 if (error)
1316 goto error0;
1317 be32_add_cpu(&agi->agi_freecount, -1);
1318 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1319 pag->pagi_freecount--;
1320
1321 error = xfs_check_agi_freecount(cur);
1322 if (error)
1323 goto error0;
1324
1325 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1326 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1327 *inop = ino;
1328 return 0;
1329error1:
1330 xfs_btree_del_cursor(tcur, XFS_BTREE_ERROR);
1331error0:
1332 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1333 return error;
1334}
1335
1336/*
1337 * Use the free inode btree to allocate an inode based on distance from the
1338 * parent. Note that the provided cursor may be deleted and replaced.
1339 */
1340STATIC int
1341xfs_dialloc_ag_finobt_near(
1342 xfs_agino_t pagino,
1343 struct xfs_btree_cur **ocur,
1344 struct xfs_inobt_rec_incore *rec)
1345{
1346 struct xfs_btree_cur *lcur = *ocur; /* left search cursor */
1347 struct xfs_btree_cur *rcur; /* right search cursor */
1348 struct xfs_inobt_rec_incore rrec;
1349 int error;
1350 int i, j;
1351
1352 error = xfs_inobt_lookup(lcur, pagino, XFS_LOOKUP_LE, &i);
1353 if (error)
1354 return error;
1355
1356 if (i == 1) {
1357 error = xfs_inobt_get_rec(lcur, rec, &i);
1358 if (error)
1359 return error;
1360 if (XFS_IS_CORRUPT(lcur->bc_mp, i != 1)) {
1361 xfs_btree_mark_sick(lcur);
1362 return -EFSCORRUPTED;
1363 }
1364
1365 /*
1366 * See if we've landed in the parent inode record. The finobt
1367 * only tracks chunks with at least one free inode, so record
1368 * existence is enough.
1369 */
1370 if (pagino >= rec->ir_startino &&
1371 pagino < (rec->ir_startino + XFS_INODES_PER_CHUNK))
1372 return 0;
1373 }
1374
1375 error = xfs_btree_dup_cursor(lcur, &rcur);
1376 if (error)
1377 return error;
1378
1379 error = xfs_inobt_lookup(rcur, pagino, XFS_LOOKUP_GE, &j);
1380 if (error)
1381 goto error_rcur;
1382 if (j == 1) {
1383 error = xfs_inobt_get_rec(rcur, &rrec, &j);
1384 if (error)
1385 goto error_rcur;
1386 if (XFS_IS_CORRUPT(lcur->bc_mp, j != 1)) {
1387 xfs_btree_mark_sick(lcur);
1388 error = -EFSCORRUPTED;
1389 goto error_rcur;
1390 }
1391 }
1392
1393 if (XFS_IS_CORRUPT(lcur->bc_mp, i != 1 && j != 1)) {
1394 xfs_btree_mark_sick(lcur);
1395 error = -EFSCORRUPTED;
1396 goto error_rcur;
1397 }
1398 if (i == 1 && j == 1) {
1399 /*
1400 * Both the left and right records are valid. Choose the closer
1401 * inode chunk to the target.
1402 */
1403 if ((pagino - rec->ir_startino + XFS_INODES_PER_CHUNK - 1) >
1404 (rrec.ir_startino - pagino)) {
1405 *rec = rrec;
1406 xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1407 *ocur = rcur;
1408 } else {
1409 xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1410 }
1411 } else if (j == 1) {
1412 /* only the right record is valid */
1413 *rec = rrec;
1414 xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1415 *ocur = rcur;
1416 } else if (i == 1) {
1417 /* only the left record is valid */
1418 xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1419 }
1420
1421 return 0;
1422
1423error_rcur:
1424 xfs_btree_del_cursor(rcur, XFS_BTREE_ERROR);
1425 return error;
1426}
1427
1428/*
1429 * Use the free inode btree to find a free inode based on a newino hint. If
1430 * the hint is NULL, find the first free inode in the AG.
1431 */
1432STATIC int
1433xfs_dialloc_ag_finobt_newino(
1434 struct xfs_agi *agi,
1435 struct xfs_btree_cur *cur,
1436 struct xfs_inobt_rec_incore *rec)
1437{
1438 int error;
1439 int i;
1440
1441 if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1442 error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1443 XFS_LOOKUP_EQ, &i);
1444 if (error)
1445 return error;
1446 if (i == 1) {
1447 error = xfs_inobt_get_rec(cur, rec, &i);
1448 if (error)
1449 return error;
1450 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1451 xfs_btree_mark_sick(cur);
1452 return -EFSCORRUPTED;
1453 }
1454 return 0;
1455 }
1456 }
1457
1458 /*
1459 * Find the first inode available in the AG.
1460 */
1461 error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1462 if (error)
1463 return error;
1464 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1465 xfs_btree_mark_sick(cur);
1466 return -EFSCORRUPTED;
1467 }
1468
1469 error = xfs_inobt_get_rec(cur, rec, &i);
1470 if (error)
1471 return error;
1472 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1473 xfs_btree_mark_sick(cur);
1474 return -EFSCORRUPTED;
1475 }
1476
1477 return 0;
1478}
1479
1480/*
1481 * Update the inobt based on a modification made to the finobt. Also ensure that
1482 * the records from both trees are equivalent post-modification.
1483 */
1484STATIC int
1485xfs_dialloc_ag_update_inobt(
1486 struct xfs_btree_cur *cur, /* inobt cursor */
1487 struct xfs_inobt_rec_incore *frec, /* finobt record */
1488 int offset) /* inode offset */
1489{
1490 struct xfs_inobt_rec_incore rec;
1491 int error;
1492 int i;
1493
1494 error = xfs_inobt_lookup(cur, frec->ir_startino, XFS_LOOKUP_EQ, &i);
1495 if (error)
1496 return error;
1497 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1498 xfs_btree_mark_sick(cur);
1499 return -EFSCORRUPTED;
1500 }
1501
1502 error = xfs_inobt_get_rec(cur, &rec, &i);
1503 if (error)
1504 return error;
1505 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1506 xfs_btree_mark_sick(cur);
1507 return -EFSCORRUPTED;
1508 }
1509 ASSERT((XFS_AGINO_TO_OFFSET(cur->bc_mp, rec.ir_startino) %
1510 XFS_INODES_PER_CHUNK) == 0);
1511
1512 rec.ir_free &= ~XFS_INOBT_MASK(offset);
1513 rec.ir_freecount--;
1514
1515 if (XFS_IS_CORRUPT(cur->bc_mp,
1516 rec.ir_free != frec->ir_free ||
1517 rec.ir_freecount != frec->ir_freecount)) {
1518 xfs_btree_mark_sick(cur);
1519 return -EFSCORRUPTED;
1520 }
1521
1522 return xfs_inobt_update(cur, &rec);
1523}
1524
1525/*
1526 * Allocate an inode using the free inode btree, if available. Otherwise, fall
1527 * back to the inobt search algorithm.
1528 *
1529 * The caller selected an AG for us, and made sure that free inodes are
1530 * available.
1531 */
1532static int
1533xfs_dialloc_ag(
1534 struct xfs_perag *pag,
1535 struct xfs_trans *tp,
1536 struct xfs_buf *agbp,
1537 xfs_ino_t parent,
1538 xfs_ino_t *inop)
1539{
1540 struct xfs_mount *mp = tp->t_mountp;
1541 struct xfs_agi *agi = agbp->b_addr;
1542 xfs_agnumber_t pagno = XFS_INO_TO_AGNO(mp, parent);
1543 xfs_agino_t pagino = XFS_INO_TO_AGINO(mp, parent);
1544 struct xfs_btree_cur *cur; /* finobt cursor */
1545 struct xfs_btree_cur *icur; /* inobt cursor */
1546 struct xfs_inobt_rec_incore rec;
1547 xfs_ino_t ino;
1548 int error;
1549 int offset;
1550 int i;
1551
1552 if (!xfs_has_finobt(mp))
1553 return xfs_dialloc_ag_inobt(pag, tp, agbp, parent, inop);
1554
1555 /*
1556 * If pagino is 0 (this is the root inode allocation) use newino.
1557 * This must work because we've just allocated some.
1558 */
1559 if (!pagino)
1560 pagino = be32_to_cpu(agi->agi_newino);
1561
1562 cur = xfs_finobt_init_cursor(pag, tp, agbp);
1563
1564 error = xfs_check_agi_freecount(cur);
1565 if (error)
1566 goto error_cur;
1567
1568 /*
1569 * The search algorithm depends on whether we're in the same AG as the
1570 * parent. If so, find the closest available inode to the parent. If
1571 * not, consider the agi hint or find the first free inode in the AG.
1572 */
1573 if (pag->pag_agno == pagno)
1574 error = xfs_dialloc_ag_finobt_near(pagino, &cur, &rec);
1575 else
1576 error = xfs_dialloc_ag_finobt_newino(agi, cur, &rec);
1577 if (error)
1578 goto error_cur;
1579
1580 offset = xfs_inobt_first_free_inode(&rec);
1581 ASSERT(offset >= 0);
1582 ASSERT(offset < XFS_INODES_PER_CHUNK);
1583 ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1584 XFS_INODES_PER_CHUNK) == 0);
1585 ino = XFS_AGINO_TO_INO(mp, pag->pag_agno, rec.ir_startino + offset);
1586
1587 /*
1588 * Modify or remove the finobt record.
1589 */
1590 rec.ir_free &= ~XFS_INOBT_MASK(offset);
1591 rec.ir_freecount--;
1592 if (rec.ir_freecount)
1593 error = xfs_inobt_update(cur, &rec);
1594 else
1595 error = xfs_btree_delete(cur, &i);
1596 if (error)
1597 goto error_cur;
1598
1599 /*
1600 * The finobt has now been updated appropriately. We haven't updated the
1601 * agi and superblock yet, so we can create an inobt cursor and validate
1602 * the original freecount. If all is well, make the equivalent update to
1603 * the inobt using the finobt record and offset information.
1604 */
1605 icur = xfs_inobt_init_cursor(pag, tp, agbp);
1606
1607 error = xfs_check_agi_freecount(icur);
1608 if (error)
1609 goto error_icur;
1610
1611 error = xfs_dialloc_ag_update_inobt(icur, &rec, offset);
1612 if (error)
1613 goto error_icur;
1614
1615 /*
1616 * Both trees have now been updated. We must update the perag and
1617 * superblock before we can check the freecount for each btree.
1618 */
1619 be32_add_cpu(&agi->agi_freecount, -1);
1620 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1621 pag->pagi_freecount--;
1622
1623 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1624
1625 error = xfs_check_agi_freecount(icur);
1626 if (error)
1627 goto error_icur;
1628 error = xfs_check_agi_freecount(cur);
1629 if (error)
1630 goto error_icur;
1631
1632 xfs_btree_del_cursor(icur, XFS_BTREE_NOERROR);
1633 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1634 *inop = ino;
1635 return 0;
1636
1637error_icur:
1638 xfs_btree_del_cursor(icur, XFS_BTREE_ERROR);
1639error_cur:
1640 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1641 return error;
1642}
1643
1644static int
1645xfs_dialloc_roll(
1646 struct xfs_trans **tpp,
1647 struct xfs_buf *agibp)
1648{
1649 struct xfs_trans *tp = *tpp;
1650 struct xfs_dquot_acct *dqinfo;
1651 int error;
1652
1653 /*
1654 * Hold to on to the agibp across the commit so no other allocation can
1655 * come in and take the free inodes we just allocated for our caller.
1656 */
1657 xfs_trans_bhold(tp, agibp);
1658
1659 /*
1660 * We want the quota changes to be associated with the next transaction,
1661 * NOT this one. So, detach the dqinfo from this and attach it to the
1662 * next transaction.
1663 */
1664 dqinfo = tp->t_dqinfo;
1665 tp->t_dqinfo = NULL;
1666
1667 error = xfs_trans_roll(&tp);
1668
1669 /* Re-attach the quota info that we detached from prev trx. */
1670 tp->t_dqinfo = dqinfo;
1671
1672 /*
1673 * Join the buffer even on commit error so that the buffer is released
1674 * when the caller cancels the transaction and doesn't have to handle
1675 * this error case specially.
1676 */
1677 xfs_trans_bjoin(tp, agibp);
1678 *tpp = tp;
1679 return error;
1680}
1681
1682static bool
1683xfs_dialloc_good_ag(
1684 struct xfs_perag *pag,
1685 struct xfs_trans *tp,
1686 umode_t mode,
1687 int flags,
1688 bool ok_alloc)
1689{
1690 struct xfs_mount *mp = tp->t_mountp;
1691 xfs_extlen_t ineed;
1692 xfs_extlen_t longest = 0;
1693 int needspace;
1694 int error;
1695
1696 if (!pag)
1697 return false;
1698 if (!xfs_perag_allows_inodes(pag))
1699 return false;
1700
1701 if (!xfs_perag_initialised_agi(pag)) {
1702 error = xfs_ialloc_read_agi(pag, tp, NULL);
1703 if (error)
1704 return false;
1705 }
1706
1707 if (pag->pagi_freecount)
1708 return true;
1709 if (!ok_alloc)
1710 return false;
1711
1712 if (!xfs_perag_initialised_agf(pag)) {
1713 error = xfs_alloc_read_agf(pag, tp, flags, NULL);
1714 if (error)
1715 return false;
1716 }
1717
1718 /*
1719 * Check that there is enough free space for the file plus a chunk of
1720 * inodes if we need to allocate some. If this is the first pass across
1721 * the AGs, take into account the potential space needed for alignment
1722 * of inode chunks when checking the longest contiguous free space in
1723 * the AG - this prevents us from getting ENOSPC because we have free
1724 * space larger than ialloc_blks but alignment constraints prevent us
1725 * from using it.
1726 *
1727 * If we can't find an AG with space for full alignment slack to be
1728 * taken into account, we must be near ENOSPC in all AGs. Hence we
1729 * don't include alignment for the second pass and so if we fail
1730 * allocation due to alignment issues then it is most likely a real
1731 * ENOSPC condition.
1732 *
1733 * XXX(dgc): this calculation is now bogus thanks to the per-ag
1734 * reservations that xfs_alloc_fix_freelist() now does via
1735 * xfs_alloc_space_available(). When the AG fills up, pagf_freeblks will
1736 * be more than large enough for the check below to succeed, but
1737 * xfs_alloc_space_available() will fail because of the non-zero
1738 * metadata reservation and hence we won't actually be able to allocate
1739 * more inodes in this AG. We do soooo much unnecessary work near ENOSPC
1740 * because of this.
1741 */
1742 ineed = M_IGEO(mp)->ialloc_min_blks;
1743 if (flags && ineed > 1)
1744 ineed += M_IGEO(mp)->cluster_align;
1745 longest = pag->pagf_longest;
1746 if (!longest)
1747 longest = pag->pagf_flcount > 0;
1748 needspace = S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode);
1749
1750 if (pag->pagf_freeblks < needspace + ineed || longest < ineed)
1751 return false;
1752 return true;
1753}
1754
1755static int
1756xfs_dialloc_try_ag(
1757 struct xfs_perag *pag,
1758 struct xfs_trans **tpp,
1759 xfs_ino_t parent,
1760 xfs_ino_t *new_ino,
1761 bool ok_alloc)
1762{
1763 struct xfs_buf *agbp;
1764 xfs_ino_t ino;
1765 int error;
1766
1767 /*
1768 * Then read in the AGI buffer and recheck with the AGI buffer
1769 * lock held.
1770 */
1771 error = xfs_ialloc_read_agi(pag, *tpp, &agbp);
1772 if (error)
1773 return error;
1774
1775 if (!pag->pagi_freecount) {
1776 if (!ok_alloc) {
1777 error = -EAGAIN;
1778 goto out_release;
1779 }
1780
1781 error = xfs_ialloc_ag_alloc(pag, *tpp, agbp);
1782 if (error < 0)
1783 goto out_release;
1784
1785 /*
1786 * We successfully allocated space for an inode cluster in this
1787 * AG. Roll the transaction so that we can allocate one of the
1788 * new inodes.
1789 */
1790 ASSERT(pag->pagi_freecount > 0);
1791 error = xfs_dialloc_roll(tpp, agbp);
1792 if (error)
1793 goto out_release;
1794 }
1795
1796 /* Allocate an inode in the found AG */
1797 error = xfs_dialloc_ag(pag, *tpp, agbp, parent, &ino);
1798 if (!error)
1799 *new_ino = ino;
1800 return error;
1801
1802out_release:
1803 xfs_trans_brelse(*tpp, agbp);
1804 return error;
1805}
1806
1807/*
1808 * Allocate an on-disk inode.
1809 *
1810 * Mode is used to tell whether the new inode is a directory and hence where to
1811 * locate it. The on-disk inode that is allocated will be returned in @new_ino
1812 * on success, otherwise an error will be set to indicate the failure (e.g.
1813 * -ENOSPC).
1814 */
1815int
1816xfs_dialloc(
1817 struct xfs_trans **tpp,
1818 xfs_ino_t parent,
1819 umode_t mode,
1820 xfs_ino_t *new_ino)
1821{
1822 struct xfs_mount *mp = (*tpp)->t_mountp;
1823 xfs_agnumber_t agno;
1824 int error = 0;
1825 xfs_agnumber_t start_agno;
1826 struct xfs_perag *pag;
1827 struct xfs_ino_geometry *igeo = M_IGEO(mp);
1828 bool ok_alloc = true;
1829 bool low_space = false;
1830 int flags;
1831 xfs_ino_t ino = NULLFSINO;
1832
1833 /*
1834 * Directories, symlinks, and regular files frequently allocate at least
1835 * one block, so factor that potential expansion when we examine whether
1836 * an AG has enough space for file creation.
1837 */
1838 if (S_ISDIR(mode))
1839 start_agno = (atomic_inc_return(&mp->m_agirotor) - 1) %
1840 mp->m_maxagi;
1841 else {
1842 start_agno = XFS_INO_TO_AGNO(mp, parent);
1843 if (start_agno >= mp->m_maxagi)
1844 start_agno = 0;
1845 }
1846
1847 /*
1848 * If we have already hit the ceiling of inode blocks then clear
1849 * ok_alloc so we scan all available agi structures for a free
1850 * inode.
1851 *
1852 * Read rough value of mp->m_icount by percpu_counter_read_positive,
1853 * which will sacrifice the preciseness but improve the performance.
1854 */
1855 if (igeo->maxicount &&
1856 percpu_counter_read_positive(&mp->m_icount) + igeo->ialloc_inos
1857 > igeo->maxicount) {
1858 ok_alloc = false;
1859 }
1860
1861 /*
1862 * If we are near to ENOSPC, we want to prefer allocation from AGs that
1863 * have free inodes in them rather than use up free space allocating new
1864 * inode chunks. Hence we turn off allocation for the first non-blocking
1865 * pass through the AGs if we are near ENOSPC to consume free inodes
1866 * that we can immediately allocate, but then we allow allocation on the
1867 * second pass if we fail to find an AG with free inodes in it.
1868 */
1869 if (percpu_counter_read_positive(&mp->m_fdblocks) <
1870 mp->m_low_space[XFS_LOWSP_1_PCNT]) {
1871 ok_alloc = false;
1872 low_space = true;
1873 }
1874
1875 /*
1876 * Loop until we find an allocation group that either has free inodes
1877 * or in which we can allocate some inodes. Iterate through the
1878 * allocation groups upward, wrapping at the end.
1879 */
1880 flags = XFS_ALLOC_FLAG_TRYLOCK;
1881retry:
1882 for_each_perag_wrap_at(mp, start_agno, mp->m_maxagi, agno, pag) {
1883 if (xfs_dialloc_good_ag(pag, *tpp, mode, flags, ok_alloc)) {
1884 error = xfs_dialloc_try_ag(pag, tpp, parent,
1885 &ino, ok_alloc);
1886 if (error != -EAGAIN)
1887 break;
1888 error = 0;
1889 }
1890
1891 if (xfs_is_shutdown(mp)) {
1892 error = -EFSCORRUPTED;
1893 break;
1894 }
1895 }
1896 if (pag)
1897 xfs_perag_rele(pag);
1898 if (error)
1899 return error;
1900 if (ino == NULLFSINO) {
1901 if (flags) {
1902 flags = 0;
1903 if (low_space)
1904 ok_alloc = true;
1905 goto retry;
1906 }
1907 return -ENOSPC;
1908 }
1909 *new_ino = ino;
1910 return 0;
1911}
1912
1913/*
1914 * Free the blocks of an inode chunk. We must consider that the inode chunk
1915 * might be sparse and only free the regions that are allocated as part of the
1916 * chunk.
1917 */
1918static int
1919xfs_difree_inode_chunk(
1920 struct xfs_trans *tp,
1921 xfs_agnumber_t agno,
1922 struct xfs_inobt_rec_incore *rec)
1923{
1924 struct xfs_mount *mp = tp->t_mountp;
1925 xfs_agblock_t sagbno = XFS_AGINO_TO_AGBNO(mp,
1926 rec->ir_startino);
1927 int startidx, endidx;
1928 int nextbit;
1929 xfs_agblock_t agbno;
1930 int contigblk;
1931 DECLARE_BITMAP(holemask, XFS_INOBT_HOLEMASK_BITS);
1932
1933 if (!xfs_inobt_issparse(rec->ir_holemask)) {
1934 /* not sparse, calculate extent info directly */
1935 return xfs_free_extent_later(tp,
1936 XFS_AGB_TO_FSB(mp, agno, sagbno),
1937 M_IGEO(mp)->ialloc_blks, &XFS_RMAP_OINFO_INODES,
1938 XFS_AG_RESV_NONE, false);
1939 }
1940
1941 /* holemask is only 16-bits (fits in an unsigned long) */
1942 ASSERT(sizeof(rec->ir_holemask) <= sizeof(holemask[0]));
1943 holemask[0] = rec->ir_holemask;
1944
1945 /*
1946 * Find contiguous ranges of zeroes (i.e., allocated regions) in the
1947 * holemask and convert the start/end index of each range to an extent.
1948 * We start with the start and end index both pointing at the first 0 in
1949 * the mask.
1950 */
1951 startidx = endidx = find_first_zero_bit(holemask,
1952 XFS_INOBT_HOLEMASK_BITS);
1953 nextbit = startidx + 1;
1954 while (startidx < XFS_INOBT_HOLEMASK_BITS) {
1955 int error;
1956
1957 nextbit = find_next_zero_bit(holemask, XFS_INOBT_HOLEMASK_BITS,
1958 nextbit);
1959 /*
1960 * If the next zero bit is contiguous, update the end index of
1961 * the current range and continue.
1962 */
1963 if (nextbit != XFS_INOBT_HOLEMASK_BITS &&
1964 nextbit == endidx + 1) {
1965 endidx = nextbit;
1966 goto next;
1967 }
1968
1969 /*
1970 * nextbit is not contiguous with the current end index. Convert
1971 * the current start/end to an extent and add it to the free
1972 * list.
1973 */
1974 agbno = sagbno + (startidx * XFS_INODES_PER_HOLEMASK_BIT) /
1975 mp->m_sb.sb_inopblock;
1976 contigblk = ((endidx - startidx + 1) *
1977 XFS_INODES_PER_HOLEMASK_BIT) /
1978 mp->m_sb.sb_inopblock;
1979
1980 ASSERT(agbno % mp->m_sb.sb_spino_align == 0);
1981 ASSERT(contigblk % mp->m_sb.sb_spino_align == 0);
1982 error = xfs_free_extent_later(tp,
1983 XFS_AGB_TO_FSB(mp, agno, agbno), contigblk,
1984 &XFS_RMAP_OINFO_INODES, XFS_AG_RESV_NONE,
1985 false);
1986 if (error)
1987 return error;
1988
1989 /* reset range to current bit and carry on... */
1990 startidx = endidx = nextbit;
1991
1992next:
1993 nextbit++;
1994 }
1995 return 0;
1996}
1997
1998STATIC int
1999xfs_difree_inobt(
2000 struct xfs_perag *pag,
2001 struct xfs_trans *tp,
2002 struct xfs_buf *agbp,
2003 xfs_agino_t agino,
2004 struct xfs_icluster *xic,
2005 struct xfs_inobt_rec_incore *orec)
2006{
2007 struct xfs_mount *mp = pag->pag_mount;
2008 struct xfs_agi *agi = agbp->b_addr;
2009 struct xfs_btree_cur *cur;
2010 struct xfs_inobt_rec_incore rec;
2011 int ilen;
2012 int error;
2013 int i;
2014 int off;
2015
2016 ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
2017 ASSERT(XFS_AGINO_TO_AGBNO(mp, agino) < be32_to_cpu(agi->agi_length));
2018
2019 /*
2020 * Initialize the cursor.
2021 */
2022 cur = xfs_inobt_init_cursor(pag, tp, agbp);
2023
2024 error = xfs_check_agi_freecount(cur);
2025 if (error)
2026 goto error0;
2027
2028 /*
2029 * Look for the entry describing this inode.
2030 */
2031 if ((error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i))) {
2032 xfs_warn(mp, "%s: xfs_inobt_lookup() returned error %d.",
2033 __func__, error);
2034 goto error0;
2035 }
2036 if (XFS_IS_CORRUPT(mp, i != 1)) {
2037 xfs_btree_mark_sick(cur);
2038 error = -EFSCORRUPTED;
2039 goto error0;
2040 }
2041 error = xfs_inobt_get_rec(cur, &rec, &i);
2042 if (error) {
2043 xfs_warn(mp, "%s: xfs_inobt_get_rec() returned error %d.",
2044 __func__, error);
2045 goto error0;
2046 }
2047 if (XFS_IS_CORRUPT(mp, i != 1)) {
2048 xfs_btree_mark_sick(cur);
2049 error = -EFSCORRUPTED;
2050 goto error0;
2051 }
2052 /*
2053 * Get the offset in the inode chunk.
2054 */
2055 off = agino - rec.ir_startino;
2056 ASSERT(off >= 0 && off < XFS_INODES_PER_CHUNK);
2057 ASSERT(!(rec.ir_free & XFS_INOBT_MASK(off)));
2058 /*
2059 * Mark the inode free & increment the count.
2060 */
2061 rec.ir_free |= XFS_INOBT_MASK(off);
2062 rec.ir_freecount++;
2063
2064 /*
2065 * When an inode chunk is free, it becomes eligible for removal. Don't
2066 * remove the chunk if the block size is large enough for multiple inode
2067 * chunks (that might not be free).
2068 */
2069 if (!xfs_has_ikeep(mp) && rec.ir_free == XFS_INOBT_ALL_FREE &&
2070 mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK) {
2071 xic->deleted = true;
2072 xic->first_ino = XFS_AGINO_TO_INO(mp, pag->pag_agno,
2073 rec.ir_startino);
2074 xic->alloc = xfs_inobt_irec_to_allocmask(&rec);
2075
2076 /*
2077 * Remove the inode cluster from the AGI B+Tree, adjust the
2078 * AGI and Superblock inode counts, and mark the disk space
2079 * to be freed when the transaction is committed.
2080 */
2081 ilen = rec.ir_freecount;
2082 be32_add_cpu(&agi->agi_count, -ilen);
2083 be32_add_cpu(&agi->agi_freecount, -(ilen - 1));
2084 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_COUNT | XFS_AGI_FREECOUNT);
2085 pag->pagi_freecount -= ilen - 1;
2086 pag->pagi_count -= ilen;
2087 xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, -ilen);
2088 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -(ilen - 1));
2089
2090 if ((error = xfs_btree_delete(cur, &i))) {
2091 xfs_warn(mp, "%s: xfs_btree_delete returned error %d.",
2092 __func__, error);
2093 goto error0;
2094 }
2095
2096 error = xfs_difree_inode_chunk(tp, pag->pag_agno, &rec);
2097 if (error)
2098 goto error0;
2099 } else {
2100 xic->deleted = false;
2101
2102 error = xfs_inobt_update(cur, &rec);
2103 if (error) {
2104 xfs_warn(mp, "%s: xfs_inobt_update returned error %d.",
2105 __func__, error);
2106 goto error0;
2107 }
2108
2109 /*
2110 * Change the inode free counts and log the ag/sb changes.
2111 */
2112 be32_add_cpu(&agi->agi_freecount, 1);
2113 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
2114 pag->pagi_freecount++;
2115 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, 1);
2116 }
2117
2118 error = xfs_check_agi_freecount(cur);
2119 if (error)
2120 goto error0;
2121
2122 *orec = rec;
2123 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2124 return 0;
2125
2126error0:
2127 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2128 return error;
2129}
2130
2131/*
2132 * Free an inode in the free inode btree.
2133 */
2134STATIC int
2135xfs_difree_finobt(
2136 struct xfs_perag *pag,
2137 struct xfs_trans *tp,
2138 struct xfs_buf *agbp,
2139 xfs_agino_t agino,
2140 struct xfs_inobt_rec_incore *ibtrec) /* inobt record */
2141{
2142 struct xfs_mount *mp = pag->pag_mount;
2143 struct xfs_btree_cur *cur;
2144 struct xfs_inobt_rec_incore rec;
2145 int offset = agino - ibtrec->ir_startino;
2146 int error;
2147 int i;
2148
2149 cur = xfs_finobt_init_cursor(pag, tp, agbp);
2150
2151 error = xfs_inobt_lookup(cur, ibtrec->ir_startino, XFS_LOOKUP_EQ, &i);
2152 if (error)
2153 goto error;
2154 if (i == 0) {
2155 /*
2156 * If the record does not exist in the finobt, we must have just
2157 * freed an inode in a previously fully allocated chunk. If not,
2158 * something is out of sync.
2159 */
2160 if (XFS_IS_CORRUPT(mp, ibtrec->ir_freecount != 1)) {
2161 xfs_btree_mark_sick(cur);
2162 error = -EFSCORRUPTED;
2163 goto error;
2164 }
2165
2166 error = xfs_inobt_insert_rec(cur, ibtrec->ir_holemask,
2167 ibtrec->ir_count,
2168 ibtrec->ir_freecount,
2169 ibtrec->ir_free, &i);
2170 if (error)
2171 goto error;
2172 ASSERT(i == 1);
2173
2174 goto out;
2175 }
2176
2177 /*
2178 * Read and update the existing record. We could just copy the ibtrec
2179 * across here, but that would defeat the purpose of having redundant
2180 * metadata. By making the modifications independently, we can catch
2181 * corruptions that we wouldn't see if we just copied from one record
2182 * to another.
2183 */
2184 error = xfs_inobt_get_rec(cur, &rec, &i);
2185 if (error)
2186 goto error;
2187 if (XFS_IS_CORRUPT(mp, i != 1)) {
2188 xfs_btree_mark_sick(cur);
2189 error = -EFSCORRUPTED;
2190 goto error;
2191 }
2192
2193 rec.ir_free |= XFS_INOBT_MASK(offset);
2194 rec.ir_freecount++;
2195
2196 if (XFS_IS_CORRUPT(mp,
2197 rec.ir_free != ibtrec->ir_free ||
2198 rec.ir_freecount != ibtrec->ir_freecount)) {
2199 xfs_btree_mark_sick(cur);
2200 error = -EFSCORRUPTED;
2201 goto error;
2202 }
2203
2204 /*
2205 * The content of inobt records should always match between the inobt
2206 * and finobt. The lifecycle of records in the finobt is different from
2207 * the inobt in that the finobt only tracks records with at least one
2208 * free inode. Hence, if all of the inodes are free and we aren't
2209 * keeping inode chunks permanently on disk, remove the record.
2210 * Otherwise, update the record with the new information.
2211 *
2212 * Note that we currently can't free chunks when the block size is large
2213 * enough for multiple chunks. Leave the finobt record to remain in sync
2214 * with the inobt.
2215 */
2216 if (!xfs_has_ikeep(mp) && rec.ir_free == XFS_INOBT_ALL_FREE &&
2217 mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK) {
2218 error = xfs_btree_delete(cur, &i);
2219 if (error)
2220 goto error;
2221 ASSERT(i == 1);
2222 } else {
2223 error = xfs_inobt_update(cur, &rec);
2224 if (error)
2225 goto error;
2226 }
2227
2228out:
2229 error = xfs_check_agi_freecount(cur);
2230 if (error)
2231 goto error;
2232
2233 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2234 return 0;
2235
2236error:
2237 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2238 return error;
2239}
2240
2241/*
2242 * Free disk inode. Carefully avoids touching the incore inode, all
2243 * manipulations incore are the caller's responsibility.
2244 * The on-disk inode is not changed by this operation, only the
2245 * btree (free inode mask) is changed.
2246 */
2247int
2248xfs_difree(
2249 struct xfs_trans *tp,
2250 struct xfs_perag *pag,
2251 xfs_ino_t inode,
2252 struct xfs_icluster *xic)
2253{
2254 /* REFERENCED */
2255 xfs_agblock_t agbno; /* block number containing inode */
2256 struct xfs_buf *agbp; /* buffer for allocation group header */
2257 xfs_agino_t agino; /* allocation group inode number */
2258 int error; /* error return value */
2259 struct xfs_mount *mp = tp->t_mountp;
2260 struct xfs_inobt_rec_incore rec;/* btree record */
2261
2262 /*
2263 * Break up inode number into its components.
2264 */
2265 if (pag->pag_agno != XFS_INO_TO_AGNO(mp, inode)) {
2266 xfs_warn(mp, "%s: agno != pag->pag_agno (%d != %d).",
2267 __func__, XFS_INO_TO_AGNO(mp, inode), pag->pag_agno);
2268 ASSERT(0);
2269 return -EINVAL;
2270 }
2271 agino = XFS_INO_TO_AGINO(mp, inode);
2272 if (inode != XFS_AGINO_TO_INO(mp, pag->pag_agno, agino)) {
2273 xfs_warn(mp, "%s: inode != XFS_AGINO_TO_INO() (%llu != %llu).",
2274 __func__, (unsigned long long)inode,
2275 (unsigned long long)XFS_AGINO_TO_INO(mp, pag->pag_agno, agino));
2276 ASSERT(0);
2277 return -EINVAL;
2278 }
2279 agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2280 if (agbno >= mp->m_sb.sb_agblocks) {
2281 xfs_warn(mp, "%s: agbno >= mp->m_sb.sb_agblocks (%d >= %d).",
2282 __func__, agbno, mp->m_sb.sb_agblocks);
2283 ASSERT(0);
2284 return -EINVAL;
2285 }
2286 /*
2287 * Get the allocation group header.
2288 */
2289 error = xfs_ialloc_read_agi(pag, tp, &agbp);
2290 if (error) {
2291 xfs_warn(mp, "%s: xfs_ialloc_read_agi() returned error %d.",
2292 __func__, error);
2293 return error;
2294 }
2295
2296 /*
2297 * Fix up the inode allocation btree.
2298 */
2299 error = xfs_difree_inobt(pag, tp, agbp, agino, xic, &rec);
2300 if (error)
2301 goto error0;
2302
2303 /*
2304 * Fix up the free inode btree.
2305 */
2306 if (xfs_has_finobt(mp)) {
2307 error = xfs_difree_finobt(pag, tp, agbp, agino, &rec);
2308 if (error)
2309 goto error0;
2310 }
2311
2312 return 0;
2313
2314error0:
2315 return error;
2316}
2317
2318STATIC int
2319xfs_imap_lookup(
2320 struct xfs_perag *pag,
2321 struct xfs_trans *tp,
2322 xfs_agino_t agino,
2323 xfs_agblock_t agbno,
2324 xfs_agblock_t *chunk_agbno,
2325 xfs_agblock_t *offset_agbno,
2326 int flags)
2327{
2328 struct xfs_mount *mp = pag->pag_mount;
2329 struct xfs_inobt_rec_incore rec;
2330 struct xfs_btree_cur *cur;
2331 struct xfs_buf *agbp;
2332 int error;
2333 int i;
2334
2335 error = xfs_ialloc_read_agi(pag, tp, &agbp);
2336 if (error) {
2337 xfs_alert(mp,
2338 "%s: xfs_ialloc_read_agi() returned error %d, agno %d",
2339 __func__, error, pag->pag_agno);
2340 return error;
2341 }
2342
2343 /*
2344 * Lookup the inode record for the given agino. If the record cannot be
2345 * found, then it's an invalid inode number and we should abort. Once
2346 * we have a record, we need to ensure it contains the inode number
2347 * we are looking up.
2348 */
2349 cur = xfs_inobt_init_cursor(pag, tp, agbp);
2350 error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i);
2351 if (!error) {
2352 if (i)
2353 error = xfs_inobt_get_rec(cur, &rec, &i);
2354 if (!error && i == 0)
2355 error = -EINVAL;
2356 }
2357
2358 xfs_trans_brelse(tp, agbp);
2359 xfs_btree_del_cursor(cur, error);
2360 if (error)
2361 return error;
2362
2363 /* check that the returned record contains the required inode */
2364 if (rec.ir_startino > agino ||
2365 rec.ir_startino + M_IGEO(mp)->ialloc_inos <= agino)
2366 return -EINVAL;
2367
2368 /* for untrusted inodes check it is allocated first */
2369 if ((flags & XFS_IGET_UNTRUSTED) &&
2370 (rec.ir_free & XFS_INOBT_MASK(agino - rec.ir_startino)))
2371 return -EINVAL;
2372
2373 *chunk_agbno = XFS_AGINO_TO_AGBNO(mp, rec.ir_startino);
2374 *offset_agbno = agbno - *chunk_agbno;
2375 return 0;
2376}
2377
2378/*
2379 * Return the location of the inode in imap, for mapping it into a buffer.
2380 */
2381int
2382xfs_imap(
2383 struct xfs_perag *pag,
2384 struct xfs_trans *tp,
2385 xfs_ino_t ino, /* inode to locate */
2386 struct xfs_imap *imap, /* location map structure */
2387 uint flags) /* flags for inode btree lookup */
2388{
2389 struct xfs_mount *mp = pag->pag_mount;
2390 xfs_agblock_t agbno; /* block number of inode in the alloc group */
2391 xfs_agino_t agino; /* inode number within alloc group */
2392 xfs_agblock_t chunk_agbno; /* first block in inode chunk */
2393 xfs_agblock_t cluster_agbno; /* first block in inode cluster */
2394 int error; /* error code */
2395 int offset; /* index of inode in its buffer */
2396 xfs_agblock_t offset_agbno; /* blks from chunk start to inode */
2397
2398 ASSERT(ino != NULLFSINO);
2399
2400 /*
2401 * Split up the inode number into its parts.
2402 */
2403 agino = XFS_INO_TO_AGINO(mp, ino);
2404 agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2405 if (agbno >= mp->m_sb.sb_agblocks ||
2406 ino != XFS_AGINO_TO_INO(mp, pag->pag_agno, agino)) {
2407 error = -EINVAL;
2408#ifdef DEBUG
2409 /*
2410 * Don't output diagnostic information for untrusted inodes
2411 * as they can be invalid without implying corruption.
2412 */
2413 if (flags & XFS_IGET_UNTRUSTED)
2414 return error;
2415 if (agbno >= mp->m_sb.sb_agblocks) {
2416 xfs_alert(mp,
2417 "%s: agbno (0x%llx) >= mp->m_sb.sb_agblocks (0x%lx)",
2418 __func__, (unsigned long long)agbno,
2419 (unsigned long)mp->m_sb.sb_agblocks);
2420 }
2421 if (ino != XFS_AGINO_TO_INO(mp, pag->pag_agno, agino)) {
2422 xfs_alert(mp,
2423 "%s: ino (0x%llx) != XFS_AGINO_TO_INO() (0x%llx)",
2424 __func__, ino,
2425 XFS_AGINO_TO_INO(mp, pag->pag_agno, agino));
2426 }
2427 xfs_stack_trace();
2428#endif /* DEBUG */
2429 return error;
2430 }
2431
2432 /*
2433 * For bulkstat and handle lookups, we have an untrusted inode number
2434 * that we have to verify is valid. We cannot do this just by reading
2435 * the inode buffer as it may have been unlinked and removed leaving
2436 * inodes in stale state on disk. Hence we have to do a btree lookup
2437 * in all cases where an untrusted inode number is passed.
2438 */
2439 if (flags & XFS_IGET_UNTRUSTED) {
2440 error = xfs_imap_lookup(pag, tp, agino, agbno,
2441 &chunk_agbno, &offset_agbno, flags);
2442 if (error)
2443 return error;
2444 goto out_map;
2445 }
2446
2447 /*
2448 * If the inode cluster size is the same as the blocksize or
2449 * smaller we get to the buffer by simple arithmetics.
2450 */
2451 if (M_IGEO(mp)->blocks_per_cluster == 1) {
2452 offset = XFS_INO_TO_OFFSET(mp, ino);
2453 ASSERT(offset < mp->m_sb.sb_inopblock);
2454
2455 imap->im_blkno = XFS_AGB_TO_DADDR(mp, pag->pag_agno, agbno);
2456 imap->im_len = XFS_FSB_TO_BB(mp, 1);
2457 imap->im_boffset = (unsigned short)(offset <<
2458 mp->m_sb.sb_inodelog);
2459 return 0;
2460 }
2461
2462 /*
2463 * If the inode chunks are aligned then use simple maths to
2464 * find the location. Otherwise we have to do a btree
2465 * lookup to find the location.
2466 */
2467 if (M_IGEO(mp)->inoalign_mask) {
2468 offset_agbno = agbno & M_IGEO(mp)->inoalign_mask;
2469 chunk_agbno = agbno - offset_agbno;
2470 } else {
2471 error = xfs_imap_lookup(pag, tp, agino, agbno,
2472 &chunk_agbno, &offset_agbno, flags);
2473 if (error)
2474 return error;
2475 }
2476
2477out_map:
2478 ASSERT(agbno >= chunk_agbno);
2479 cluster_agbno = chunk_agbno +
2480 ((offset_agbno / M_IGEO(mp)->blocks_per_cluster) *
2481 M_IGEO(mp)->blocks_per_cluster);
2482 offset = ((agbno - cluster_agbno) * mp->m_sb.sb_inopblock) +
2483 XFS_INO_TO_OFFSET(mp, ino);
2484
2485 imap->im_blkno = XFS_AGB_TO_DADDR(mp, pag->pag_agno, cluster_agbno);
2486 imap->im_len = XFS_FSB_TO_BB(mp, M_IGEO(mp)->blocks_per_cluster);
2487 imap->im_boffset = (unsigned short)(offset << mp->m_sb.sb_inodelog);
2488
2489 /*
2490 * If the inode number maps to a block outside the bounds
2491 * of the file system then return NULL rather than calling
2492 * read_buf and panicing when we get an error from the
2493 * driver.
2494 */
2495 if ((imap->im_blkno + imap->im_len) >
2496 XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks)) {
2497 xfs_alert(mp,
2498 "%s: (im_blkno (0x%llx) + im_len (0x%llx)) > sb_dblocks (0x%llx)",
2499 __func__, (unsigned long long) imap->im_blkno,
2500 (unsigned long long) imap->im_len,
2501 XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks));
2502 return -EINVAL;
2503 }
2504 return 0;
2505}
2506
2507/*
2508 * Log specified fields for the ag hdr (inode section). The growth of the agi
2509 * structure over time requires that we interpret the buffer as two logical
2510 * regions delineated by the end of the unlinked list. This is due to the size
2511 * of the hash table and its location in the middle of the agi.
2512 *
2513 * For example, a request to log a field before agi_unlinked and a field after
2514 * agi_unlinked could cause us to log the entire hash table and use an excessive
2515 * amount of log space. To avoid this behavior, log the region up through
2516 * agi_unlinked in one call and the region after agi_unlinked through the end of
2517 * the structure in another.
2518 */
2519void
2520xfs_ialloc_log_agi(
2521 struct xfs_trans *tp,
2522 struct xfs_buf *bp,
2523 uint32_t fields)
2524{
2525 int first; /* first byte number */
2526 int last; /* last byte number */
2527 static const short offsets[] = { /* field starting offsets */
2528 /* keep in sync with bit definitions */
2529 offsetof(xfs_agi_t, agi_magicnum),
2530 offsetof(xfs_agi_t, agi_versionnum),
2531 offsetof(xfs_agi_t, agi_seqno),
2532 offsetof(xfs_agi_t, agi_length),
2533 offsetof(xfs_agi_t, agi_count),
2534 offsetof(xfs_agi_t, agi_root),
2535 offsetof(xfs_agi_t, agi_level),
2536 offsetof(xfs_agi_t, agi_freecount),
2537 offsetof(xfs_agi_t, agi_newino),
2538 offsetof(xfs_agi_t, agi_dirino),
2539 offsetof(xfs_agi_t, agi_unlinked),
2540 offsetof(xfs_agi_t, agi_free_root),
2541 offsetof(xfs_agi_t, agi_free_level),
2542 offsetof(xfs_agi_t, agi_iblocks),
2543 sizeof(xfs_agi_t)
2544 };
2545#ifdef DEBUG
2546 struct xfs_agi *agi = bp->b_addr;
2547
2548 ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
2549#endif
2550
2551 /*
2552 * Compute byte offsets for the first and last fields in the first
2553 * region and log the agi buffer. This only logs up through
2554 * agi_unlinked.
2555 */
2556 if (fields & XFS_AGI_ALL_BITS_R1) {
2557 xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R1,
2558 &first, &last);
2559 xfs_trans_log_buf(tp, bp, first, last);
2560 }
2561
2562 /*
2563 * Mask off the bits in the first region and calculate the first and
2564 * last field offsets for any bits in the second region.
2565 */
2566 fields &= ~XFS_AGI_ALL_BITS_R1;
2567 if (fields) {
2568 xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R2,
2569 &first, &last);
2570 xfs_trans_log_buf(tp, bp, first, last);
2571 }
2572}
2573
2574static xfs_failaddr_t
2575xfs_agi_verify(
2576 struct xfs_buf *bp)
2577{
2578 struct xfs_mount *mp = bp->b_mount;
2579 struct xfs_agi *agi = bp->b_addr;
2580 xfs_failaddr_t fa;
2581 uint32_t agi_seqno = be32_to_cpu(agi->agi_seqno);
2582 uint32_t agi_length = be32_to_cpu(agi->agi_length);
2583 int i;
2584
2585 if (xfs_has_crc(mp)) {
2586 if (!uuid_equal(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid))
2587 return __this_address;
2588 if (!xfs_log_check_lsn(mp, be64_to_cpu(agi->agi_lsn)))
2589 return __this_address;
2590 }
2591
2592 /*
2593 * Validate the magic number of the agi block.
2594 */
2595 if (!xfs_verify_magic(bp, agi->agi_magicnum))
2596 return __this_address;
2597 if (!XFS_AGI_GOOD_VERSION(be32_to_cpu(agi->agi_versionnum)))
2598 return __this_address;
2599
2600 fa = xfs_validate_ag_length(bp, agi_seqno, agi_length);
2601 if (fa)
2602 return fa;
2603
2604 if (be32_to_cpu(agi->agi_level) < 1 ||
2605 be32_to_cpu(agi->agi_level) > M_IGEO(mp)->inobt_maxlevels)
2606 return __this_address;
2607
2608 if (xfs_has_finobt(mp) &&
2609 (be32_to_cpu(agi->agi_free_level) < 1 ||
2610 be32_to_cpu(agi->agi_free_level) > M_IGEO(mp)->inobt_maxlevels))
2611 return __this_address;
2612
2613 for (i = 0; i < XFS_AGI_UNLINKED_BUCKETS; i++) {
2614 if (agi->agi_unlinked[i] == cpu_to_be32(NULLAGINO))
2615 continue;
2616 if (!xfs_verify_ino(mp, be32_to_cpu(agi->agi_unlinked[i])))
2617 return __this_address;
2618 }
2619
2620 return NULL;
2621}
2622
2623static void
2624xfs_agi_read_verify(
2625 struct xfs_buf *bp)
2626{
2627 struct xfs_mount *mp = bp->b_mount;
2628 xfs_failaddr_t fa;
2629
2630 if (xfs_has_crc(mp) &&
2631 !xfs_buf_verify_cksum(bp, XFS_AGI_CRC_OFF))
2632 xfs_verifier_error(bp, -EFSBADCRC, __this_address);
2633 else {
2634 fa = xfs_agi_verify(bp);
2635 if (XFS_TEST_ERROR(fa, mp, XFS_ERRTAG_IALLOC_READ_AGI))
2636 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2637 }
2638}
2639
2640static void
2641xfs_agi_write_verify(
2642 struct xfs_buf *bp)
2643{
2644 struct xfs_mount *mp = bp->b_mount;
2645 struct xfs_buf_log_item *bip = bp->b_log_item;
2646 struct xfs_agi *agi = bp->b_addr;
2647 xfs_failaddr_t fa;
2648
2649 fa = xfs_agi_verify(bp);
2650 if (fa) {
2651 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2652 return;
2653 }
2654
2655 if (!xfs_has_crc(mp))
2656 return;
2657
2658 if (bip)
2659 agi->agi_lsn = cpu_to_be64(bip->bli_item.li_lsn);
2660 xfs_buf_update_cksum(bp, XFS_AGI_CRC_OFF);
2661}
2662
2663const struct xfs_buf_ops xfs_agi_buf_ops = {
2664 .name = "xfs_agi",
2665 .magic = { cpu_to_be32(XFS_AGI_MAGIC), cpu_to_be32(XFS_AGI_MAGIC) },
2666 .verify_read = xfs_agi_read_verify,
2667 .verify_write = xfs_agi_write_verify,
2668 .verify_struct = xfs_agi_verify,
2669};
2670
2671/*
2672 * Read in the allocation group header (inode allocation section)
2673 */
2674int
2675xfs_read_agi(
2676 struct xfs_perag *pag,
2677 struct xfs_trans *tp,
2678 struct xfs_buf **agibpp)
2679{
2680 struct xfs_mount *mp = pag->pag_mount;
2681 int error;
2682
2683 trace_xfs_read_agi(pag->pag_mount, pag->pag_agno);
2684
2685 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
2686 XFS_AG_DADDR(mp, pag->pag_agno, XFS_AGI_DADDR(mp)),
2687 XFS_FSS_TO_BB(mp, 1), 0, agibpp, &xfs_agi_buf_ops);
2688 if (xfs_metadata_is_sick(error))
2689 xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
2690 if (error)
2691 return error;
2692 if (tp)
2693 xfs_trans_buf_set_type(tp, *agibpp, XFS_BLFT_AGI_BUF);
2694
2695 xfs_buf_set_ref(*agibpp, XFS_AGI_REF);
2696 return 0;
2697}
2698
2699/*
2700 * Read in the agi and initialise the per-ag data. If the caller supplies a
2701 * @agibpp, return the locked AGI buffer to them, otherwise release it.
2702 */
2703int
2704xfs_ialloc_read_agi(
2705 struct xfs_perag *pag,
2706 struct xfs_trans *tp,
2707 struct xfs_buf **agibpp)
2708{
2709 struct xfs_buf *agibp;
2710 struct xfs_agi *agi;
2711 int error;
2712
2713 trace_xfs_ialloc_read_agi(pag->pag_mount, pag->pag_agno);
2714
2715 error = xfs_read_agi(pag, tp, &agibp);
2716 if (error)
2717 return error;
2718
2719 agi = agibp->b_addr;
2720 if (!xfs_perag_initialised_agi(pag)) {
2721 pag->pagi_freecount = be32_to_cpu(agi->agi_freecount);
2722 pag->pagi_count = be32_to_cpu(agi->agi_count);
2723 set_bit(XFS_AGSTATE_AGI_INIT, &pag->pag_opstate);
2724 }
2725
2726 /*
2727 * It's possible for these to be out of sync if
2728 * we are in the middle of a forced shutdown.
2729 */
2730 ASSERT(pag->pagi_freecount == be32_to_cpu(agi->agi_freecount) ||
2731 xfs_is_shutdown(pag->pag_mount));
2732 if (agibpp)
2733 *agibpp = agibp;
2734 else
2735 xfs_trans_brelse(tp, agibp);
2736 return 0;
2737}
2738
2739/* How many inodes are backed by inode clusters ondisk? */
2740STATIC int
2741xfs_ialloc_count_ondisk(
2742 struct xfs_btree_cur *cur,
2743 xfs_agino_t low,
2744 xfs_agino_t high,
2745 unsigned int *allocated)
2746{
2747 struct xfs_inobt_rec_incore irec;
2748 unsigned int ret = 0;
2749 int has_record;
2750 int error;
2751
2752 error = xfs_inobt_lookup(cur, low, XFS_LOOKUP_LE, &has_record);
2753 if (error)
2754 return error;
2755
2756 while (has_record) {
2757 unsigned int i, hole_idx;
2758
2759 error = xfs_inobt_get_rec(cur, &irec, &has_record);
2760 if (error)
2761 return error;
2762 if (irec.ir_startino > high)
2763 break;
2764
2765 for (i = 0; i < XFS_INODES_PER_CHUNK; i++) {
2766 if (irec.ir_startino + i < low)
2767 continue;
2768 if (irec.ir_startino + i > high)
2769 break;
2770
2771 hole_idx = i / XFS_INODES_PER_HOLEMASK_BIT;
2772 if (!(irec.ir_holemask & (1U << hole_idx)))
2773 ret++;
2774 }
2775
2776 error = xfs_btree_increment(cur, 0, &has_record);
2777 if (error)
2778 return error;
2779 }
2780
2781 *allocated = ret;
2782 return 0;
2783}
2784
2785/* Is there an inode record covering a given extent? */
2786int
2787xfs_ialloc_has_inodes_at_extent(
2788 struct xfs_btree_cur *cur,
2789 xfs_agblock_t bno,
2790 xfs_extlen_t len,
2791 enum xbtree_recpacking *outcome)
2792{
2793 xfs_agino_t agino;
2794 xfs_agino_t last_agino;
2795 unsigned int allocated;
2796 int error;
2797
2798 agino = XFS_AGB_TO_AGINO(cur->bc_mp, bno);
2799 last_agino = XFS_AGB_TO_AGINO(cur->bc_mp, bno + len) - 1;
2800
2801 error = xfs_ialloc_count_ondisk(cur, agino, last_agino, &allocated);
2802 if (error)
2803 return error;
2804
2805 if (allocated == 0)
2806 *outcome = XBTREE_RECPACKING_EMPTY;
2807 else if (allocated == last_agino - agino + 1)
2808 *outcome = XBTREE_RECPACKING_FULL;
2809 else
2810 *outcome = XBTREE_RECPACKING_SPARSE;
2811 return 0;
2812}
2813
2814struct xfs_ialloc_count_inodes {
2815 xfs_agino_t count;
2816 xfs_agino_t freecount;
2817};
2818
2819/* Record inode counts across all inobt records. */
2820STATIC int
2821xfs_ialloc_count_inodes_rec(
2822 struct xfs_btree_cur *cur,
2823 const union xfs_btree_rec *rec,
2824 void *priv)
2825{
2826 struct xfs_inobt_rec_incore irec;
2827 struct xfs_ialloc_count_inodes *ci = priv;
2828 xfs_failaddr_t fa;
2829
2830 xfs_inobt_btrec_to_irec(cur->bc_mp, rec, &irec);
2831 fa = xfs_inobt_check_irec(cur->bc_ag.pag, &irec);
2832 if (fa)
2833 return xfs_inobt_complain_bad_rec(cur, fa, &irec);
2834
2835 ci->count += irec.ir_count;
2836 ci->freecount += irec.ir_freecount;
2837
2838 return 0;
2839}
2840
2841/* Count allocated and free inodes under an inobt. */
2842int
2843xfs_ialloc_count_inodes(
2844 struct xfs_btree_cur *cur,
2845 xfs_agino_t *count,
2846 xfs_agino_t *freecount)
2847{
2848 struct xfs_ialloc_count_inodes ci = {0};
2849 int error;
2850
2851 ASSERT(xfs_btree_is_ino(cur->bc_ops));
2852 error = xfs_btree_query_all(cur, xfs_ialloc_count_inodes_rec, &ci);
2853 if (error)
2854 return error;
2855
2856 *count = ci.count;
2857 *freecount = ci.freecount;
2858 return 0;
2859}
2860
2861/*
2862 * Initialize inode-related geometry information.
2863 *
2864 * Compute the inode btree min and max levels and set maxicount.
2865 *
2866 * Set the inode cluster size. This may still be overridden by the file
2867 * system block size if it is larger than the chosen cluster size.
2868 *
2869 * For v5 filesystems, scale the cluster size with the inode size to keep a
2870 * constant ratio of inode per cluster buffer, but only if mkfs has set the
2871 * inode alignment value appropriately for larger cluster sizes.
2872 *
2873 * Then compute the inode cluster alignment information.
2874 */
2875void
2876xfs_ialloc_setup_geometry(
2877 struct xfs_mount *mp)
2878{
2879 struct xfs_sb *sbp = &mp->m_sb;
2880 struct xfs_ino_geometry *igeo = M_IGEO(mp);
2881 uint64_t icount;
2882 uint inodes;
2883
2884 igeo->new_diflags2 = 0;
2885 if (xfs_has_bigtime(mp))
2886 igeo->new_diflags2 |= XFS_DIFLAG2_BIGTIME;
2887 if (xfs_has_large_extent_counts(mp))
2888 igeo->new_diflags2 |= XFS_DIFLAG2_NREXT64;
2889
2890 /* Compute inode btree geometry. */
2891 igeo->agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
2892 igeo->inobt_mxr[0] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, 1);
2893 igeo->inobt_mxr[1] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, 0);
2894 igeo->inobt_mnr[0] = igeo->inobt_mxr[0] / 2;
2895 igeo->inobt_mnr[1] = igeo->inobt_mxr[1] / 2;
2896
2897 igeo->ialloc_inos = max_t(uint16_t, XFS_INODES_PER_CHUNK,
2898 sbp->sb_inopblock);
2899 igeo->ialloc_blks = igeo->ialloc_inos >> sbp->sb_inopblog;
2900
2901 if (sbp->sb_spino_align)
2902 igeo->ialloc_min_blks = sbp->sb_spino_align;
2903 else
2904 igeo->ialloc_min_blks = igeo->ialloc_blks;
2905
2906 /* Compute and fill in value of m_ino_geo.inobt_maxlevels. */
2907 inodes = (1LL << XFS_INO_AGINO_BITS(mp)) >> XFS_INODES_PER_CHUNK_LOG;
2908 igeo->inobt_maxlevels = xfs_btree_compute_maxlevels(igeo->inobt_mnr,
2909 inodes);
2910 ASSERT(igeo->inobt_maxlevels <= xfs_iallocbt_maxlevels_ondisk());
2911
2912 /*
2913 * Set the maximum inode count for this filesystem, being careful not
2914 * to use obviously garbage sb_inopblog/sb_inopblock values. Regular
2915 * users should never get here due to failing sb verification, but
2916 * certain users (xfs_db) need to be usable even with corrupt metadata.
2917 */
2918 if (sbp->sb_imax_pct && igeo->ialloc_blks) {
2919 /*
2920 * Make sure the maximum inode count is a multiple
2921 * of the units we allocate inodes in.
2922 */
2923 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
2924 do_div(icount, 100);
2925 do_div(icount, igeo->ialloc_blks);
2926 igeo->maxicount = XFS_FSB_TO_INO(mp,
2927 icount * igeo->ialloc_blks);
2928 } else {
2929 igeo->maxicount = 0;
2930 }
2931
2932 /*
2933 * Compute the desired size of an inode cluster buffer size, which
2934 * starts at 8K and (on v5 filesystems) scales up with larger inode
2935 * sizes.
2936 *
2937 * Preserve the desired inode cluster size because the sparse inodes
2938 * feature uses that desired size (not the actual size) to compute the
2939 * sparse inode alignment. The mount code validates this value, so we
2940 * cannot change the behavior.
2941 */
2942 igeo->inode_cluster_size_raw = XFS_INODE_BIG_CLUSTER_SIZE;
2943 if (xfs_has_v3inodes(mp)) {
2944 int new_size = igeo->inode_cluster_size_raw;
2945
2946 new_size *= mp->m_sb.sb_inodesize / XFS_DINODE_MIN_SIZE;
2947 if (mp->m_sb.sb_inoalignmt >= XFS_B_TO_FSBT(mp, new_size))
2948 igeo->inode_cluster_size_raw = new_size;
2949 }
2950
2951 /* Calculate inode cluster ratios. */
2952 if (igeo->inode_cluster_size_raw > mp->m_sb.sb_blocksize)
2953 igeo->blocks_per_cluster = XFS_B_TO_FSBT(mp,
2954 igeo->inode_cluster_size_raw);
2955 else
2956 igeo->blocks_per_cluster = 1;
2957 igeo->inode_cluster_size = XFS_FSB_TO_B(mp, igeo->blocks_per_cluster);
2958 igeo->inodes_per_cluster = XFS_FSB_TO_INO(mp, igeo->blocks_per_cluster);
2959
2960 /* Calculate inode cluster alignment. */
2961 if (xfs_has_align(mp) &&
2962 mp->m_sb.sb_inoalignmt >= igeo->blocks_per_cluster)
2963 igeo->cluster_align = mp->m_sb.sb_inoalignmt;
2964 else
2965 igeo->cluster_align = 1;
2966 igeo->inoalign_mask = igeo->cluster_align - 1;
2967 igeo->cluster_align_inodes = XFS_FSB_TO_INO(mp, igeo->cluster_align);
2968
2969 /*
2970 * If we are using stripe alignment, check whether
2971 * the stripe unit is a multiple of the inode alignment
2972 */
2973 if (mp->m_dalign && igeo->inoalign_mask &&
2974 !(mp->m_dalign & igeo->inoalign_mask))
2975 igeo->ialloc_align = mp->m_dalign;
2976 else
2977 igeo->ialloc_align = 0;
2978}
2979
2980/* Compute the location of the root directory inode that is laid out by mkfs. */
2981xfs_ino_t
2982xfs_ialloc_calc_rootino(
2983 struct xfs_mount *mp,
2984 int sunit)
2985{
2986 struct xfs_ino_geometry *igeo = M_IGEO(mp);
2987 xfs_agblock_t first_bno;
2988
2989 /*
2990 * Pre-calculate the geometry of AG 0. We know what it looks like
2991 * because libxfs knows how to create allocation groups now.
2992 *
2993 * first_bno is the first block in which mkfs could possibly have
2994 * allocated the root directory inode, once we factor in the metadata
2995 * that mkfs formats before it. Namely, the four AG headers...
2996 */
2997 first_bno = howmany(4 * mp->m_sb.sb_sectsize, mp->m_sb.sb_blocksize);
2998
2999 /* ...the two free space btree roots... */
3000 first_bno += 2;
3001
3002 /* ...the inode btree root... */
3003 first_bno += 1;
3004
3005 /* ...the initial AGFL... */
3006 first_bno += xfs_alloc_min_freelist(mp, NULL);
3007
3008 /* ...the free inode btree root... */
3009 if (xfs_has_finobt(mp))
3010 first_bno++;
3011
3012 /* ...the reverse mapping btree root... */
3013 if (xfs_has_rmapbt(mp))
3014 first_bno++;
3015
3016 /* ...the reference count btree... */
3017 if (xfs_has_reflink(mp))
3018 first_bno++;
3019
3020 /*
3021 * ...and the log, if it is allocated in the first allocation group.
3022 *
3023 * This can happen with filesystems that only have a single
3024 * allocation group, or very odd geometries created by old mkfs
3025 * versions on very small filesystems.
3026 */
3027 if (xfs_ag_contains_log(mp, 0))
3028 first_bno += mp->m_sb.sb_logblocks;
3029
3030 /*
3031 * Now round first_bno up to whatever allocation alignment is given
3032 * by the filesystem or was passed in.
3033 */
3034 if (xfs_has_dalign(mp) && igeo->ialloc_align > 0)
3035 first_bno = roundup(first_bno, sunit);
3036 else if (xfs_has_align(mp) &&
3037 mp->m_sb.sb_inoalignmt > 1)
3038 first_bno = roundup(first_bno, mp->m_sb.sb_inoalignmt);
3039
3040 return XFS_AGINO_TO_INO(mp, 0, XFS_AGB_TO_AGINO(mp, first_bno));
3041}
3042
3043/*
3044 * Ensure there are not sparse inode clusters that cross the new EOAG.
3045 *
3046 * This is a no-op for non-spinode filesystems since clusters are always fully
3047 * allocated and checking the bnobt suffices. However, a spinode filesystem
3048 * could have a record where the upper inodes are free blocks. If those blocks
3049 * were removed from the filesystem, the inode record would extend beyond EOAG,
3050 * which will be flagged as corruption.
3051 */
3052int
3053xfs_ialloc_check_shrink(
3054 struct xfs_perag *pag,
3055 struct xfs_trans *tp,
3056 struct xfs_buf *agibp,
3057 xfs_agblock_t new_length)
3058{
3059 struct xfs_inobt_rec_incore rec;
3060 struct xfs_btree_cur *cur;
3061 xfs_agino_t agino;
3062 int has;
3063 int error;
3064
3065 if (!xfs_has_sparseinodes(pag->pag_mount))
3066 return 0;
3067
3068 cur = xfs_inobt_init_cursor(pag, tp, agibp);
3069
3070 /* Look up the inobt record that would correspond to the new EOFS. */
3071 agino = XFS_AGB_TO_AGINO(pag->pag_mount, new_length);
3072 error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &has);
3073 if (error || !has)
3074 goto out;
3075
3076 error = xfs_inobt_get_rec(cur, &rec, &has);
3077 if (error)
3078 goto out;
3079
3080 if (!has) {
3081 xfs_ag_mark_sick(pag, XFS_SICK_AG_INOBT);
3082 error = -EFSCORRUPTED;
3083 goto out;
3084 }
3085
3086 /* If the record covers inodes that would be beyond EOFS, bail out. */
3087 if (rec.ir_startino + XFS_INODES_PER_CHUNK > agino) {
3088 error = -ENOSPC;
3089 goto out;
3090 }
3091out:
3092 xfs_btree_del_cursor(cur, error);
3093 return error;
3094}