Loading...
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_group->xg_gno, 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(to_perag(cur->bc_group), 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 ==
280 to_perag(cur->bc_group)->pagi_freecount);
281 }
282 }
283 return 0;
284}
285#else
286#define xfs_check_agi_freecount(cur) 0
287#endif
288
289/*
290 * Initialise a new set of inodes. When called without a transaction context
291 * (e.g. from recovery) we initiate a delayed write of the inode buffers rather
292 * than logging them (which in a transaction context puts them into the AIL
293 * for writeback rather than the xfsbufd queue).
294 */
295int
296xfs_ialloc_inode_init(
297 struct xfs_mount *mp,
298 struct xfs_trans *tp,
299 struct list_head *buffer_list,
300 int icount,
301 xfs_agnumber_t agno,
302 xfs_agblock_t agbno,
303 xfs_agblock_t length,
304 unsigned int gen)
305{
306 struct xfs_buf *fbuf;
307 struct xfs_dinode *free;
308 int nbufs;
309 int version;
310 int i, j;
311 xfs_daddr_t d;
312 xfs_ino_t ino = 0;
313 int error;
314
315 /*
316 * Loop over the new block(s), filling in the inodes. For small block
317 * sizes, manipulate the inodes in buffers which are multiples of the
318 * blocks size.
319 */
320 nbufs = length / M_IGEO(mp)->blocks_per_cluster;
321
322 /*
323 * Figure out what version number to use in the inodes we create. If
324 * the superblock version has caught up to the one that supports the new
325 * inode format, then use the new inode version. Otherwise use the old
326 * version so that old kernels will continue to be able to use the file
327 * system.
328 *
329 * For v3 inodes, we also need to write the inode number into the inode,
330 * so calculate the first inode number of the chunk here as
331 * XFS_AGB_TO_AGINO() only works within a filesystem block, not
332 * across multiple filesystem blocks (such as a cluster) and so cannot
333 * be used in the cluster buffer loop below.
334 *
335 * Further, because we are writing the inode directly into the buffer
336 * and calculating a CRC on the entire inode, we have ot log the entire
337 * inode so that the entire range the CRC covers is present in the log.
338 * That means for v3 inode we log the entire buffer rather than just the
339 * inode cores.
340 */
341 if (xfs_has_v3inodes(mp)) {
342 version = 3;
343 ino = XFS_AGINO_TO_INO(mp, agno, XFS_AGB_TO_AGINO(mp, agbno));
344
345 /*
346 * log the initialisation that is about to take place as an
347 * logical operation. This means the transaction does not
348 * need to log the physical changes to the inode buffers as log
349 * recovery will know what initialisation is actually needed.
350 * Hence we only need to log the buffers as "ordered" buffers so
351 * they track in the AIL as if they were physically logged.
352 */
353 if (tp)
354 xfs_icreate_log(tp, agno, agbno, icount,
355 mp->m_sb.sb_inodesize, length, gen);
356 } else
357 version = 2;
358
359 for (j = 0; j < nbufs; j++) {
360 /*
361 * Get the block.
362 */
363 d = XFS_AGB_TO_DADDR(mp, agno, agbno +
364 (j * M_IGEO(mp)->blocks_per_cluster));
365 error = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
366 mp->m_bsize * M_IGEO(mp)->blocks_per_cluster,
367 XBF_UNMAPPED, &fbuf);
368 if (error)
369 return error;
370
371 /* Initialize the inode buffers and log them appropriately. */
372 fbuf->b_ops = &xfs_inode_buf_ops;
373 xfs_buf_zero(fbuf, 0, BBTOB(fbuf->b_length));
374 for (i = 0; i < M_IGEO(mp)->inodes_per_cluster; i++) {
375 int ioffset = i << mp->m_sb.sb_inodelog;
376
377 free = xfs_make_iptr(mp, fbuf, i);
378 free->di_magic = cpu_to_be16(XFS_DINODE_MAGIC);
379 free->di_version = version;
380 free->di_gen = cpu_to_be32(gen);
381 free->di_next_unlinked = cpu_to_be32(NULLAGINO);
382
383 if (version == 3) {
384 free->di_ino = cpu_to_be64(ino);
385 ino++;
386 uuid_copy(&free->di_uuid,
387 &mp->m_sb.sb_meta_uuid);
388 xfs_dinode_calc_crc(mp, free);
389 } else if (tp) {
390 /* just log the inode core */
391 xfs_trans_log_buf(tp, fbuf, ioffset,
392 ioffset + XFS_DINODE_SIZE(mp) - 1);
393 }
394 }
395
396 if (tp) {
397 /*
398 * Mark the buffer as an inode allocation buffer so it
399 * sticks in AIL at the point of this allocation
400 * transaction. This ensures the they are on disk before
401 * the tail of the log can be moved past this
402 * transaction (i.e. by preventing relogging from moving
403 * it forward in the log).
404 */
405 xfs_trans_inode_alloc_buf(tp, fbuf);
406 if (version == 3) {
407 /*
408 * Mark the buffer as ordered so that they are
409 * not physically logged in the transaction but
410 * still tracked in the AIL as part of the
411 * transaction and pin the log appropriately.
412 */
413 xfs_trans_ordered_buf(tp, fbuf);
414 }
415 } else {
416 fbuf->b_flags |= XBF_DONE;
417 xfs_buf_delwri_queue(fbuf, buffer_list);
418 xfs_buf_relse(fbuf);
419 }
420 }
421 return 0;
422}
423
424/*
425 * Align startino and allocmask for a recently allocated sparse chunk such that
426 * they are fit for insertion (or merge) into the on-disk inode btrees.
427 *
428 * Background:
429 *
430 * When enabled, sparse inode support increases the inode alignment from cluster
431 * size to inode chunk size. This means that the minimum range between two
432 * non-adjacent inode records in the inobt is large enough for a full inode
433 * record. This allows for cluster sized, cluster aligned block allocation
434 * without need to worry about whether the resulting inode record overlaps with
435 * another record in the tree. Without this basic rule, we would have to deal
436 * with the consequences of overlap by potentially undoing recent allocations in
437 * the inode allocation codepath.
438 *
439 * Because of this alignment rule (which is enforced on mount), there are two
440 * inobt possibilities for newly allocated sparse chunks. One is that the
441 * aligned inode record for the chunk covers a range of inodes not already
442 * covered in the inobt (i.e., it is safe to insert a new sparse record). The
443 * other is that a record already exists at the aligned startino that considers
444 * the newly allocated range as sparse. In the latter case, record content is
445 * merged in hope that sparse inode chunks fill to full chunks over time.
446 */
447STATIC void
448xfs_align_sparse_ino(
449 struct xfs_mount *mp,
450 xfs_agino_t *startino,
451 uint16_t *allocmask)
452{
453 xfs_agblock_t agbno;
454 xfs_agblock_t mod;
455 int offset;
456
457 agbno = XFS_AGINO_TO_AGBNO(mp, *startino);
458 mod = agbno % mp->m_sb.sb_inoalignmt;
459 if (!mod)
460 return;
461
462 /* calculate the inode offset and align startino */
463 offset = XFS_AGB_TO_AGINO(mp, mod);
464 *startino -= offset;
465
466 /*
467 * Since startino has been aligned down, left shift allocmask such that
468 * it continues to represent the same physical inodes relative to the
469 * new startino.
470 */
471 *allocmask <<= offset / XFS_INODES_PER_HOLEMASK_BIT;
472}
473
474/*
475 * Determine whether the source inode record can merge into the target. Both
476 * records must be sparse, the inode ranges must match and there must be no
477 * allocation overlap between the records.
478 */
479STATIC bool
480__xfs_inobt_can_merge(
481 struct xfs_inobt_rec_incore *trec, /* tgt record */
482 struct xfs_inobt_rec_incore *srec) /* src record */
483{
484 uint64_t talloc;
485 uint64_t salloc;
486
487 /* records must cover the same inode range */
488 if (trec->ir_startino != srec->ir_startino)
489 return false;
490
491 /* both records must be sparse */
492 if (!xfs_inobt_issparse(trec->ir_holemask) ||
493 !xfs_inobt_issparse(srec->ir_holemask))
494 return false;
495
496 /* both records must track some inodes */
497 if (!trec->ir_count || !srec->ir_count)
498 return false;
499
500 /* can't exceed capacity of a full record */
501 if (trec->ir_count + srec->ir_count > XFS_INODES_PER_CHUNK)
502 return false;
503
504 /* verify there is no allocation overlap */
505 talloc = xfs_inobt_irec_to_allocmask(trec);
506 salloc = xfs_inobt_irec_to_allocmask(srec);
507 if (talloc & salloc)
508 return false;
509
510 return true;
511}
512
513/*
514 * Merge the source inode record into the target. The caller must call
515 * __xfs_inobt_can_merge() to ensure the merge is valid.
516 */
517STATIC void
518__xfs_inobt_rec_merge(
519 struct xfs_inobt_rec_incore *trec, /* target */
520 struct xfs_inobt_rec_incore *srec) /* src */
521{
522 ASSERT(trec->ir_startino == srec->ir_startino);
523
524 /* combine the counts */
525 trec->ir_count += srec->ir_count;
526 trec->ir_freecount += srec->ir_freecount;
527
528 /*
529 * Merge the holemask and free mask. For both fields, 0 bits refer to
530 * allocated inodes. We combine the allocated ranges with bitwise AND.
531 */
532 trec->ir_holemask &= srec->ir_holemask;
533 trec->ir_free &= srec->ir_free;
534}
535
536/*
537 * Insert a new sparse inode chunk into the associated inode allocation btree.
538 * The inode record for the sparse chunk is pre-aligned to a startino that
539 * should match any pre-existing sparse inode record in the tree. This allows
540 * sparse chunks to fill over time.
541 *
542 * If no preexisting record exists, the provided record is inserted.
543 * If there is a preexisting record, the provided record is merged with the
544 * existing record and updated in place. The merged record is returned in nrec.
545 *
546 * It is considered corruption if a merge is requested and not possible. Given
547 * the sparse inode alignment constraints, this should never happen.
548 */
549STATIC int
550xfs_inobt_insert_sprec(
551 struct xfs_perag *pag,
552 struct xfs_trans *tp,
553 struct xfs_buf *agbp,
554 struct xfs_inobt_rec_incore *nrec) /* in/out: new/merged rec. */
555{
556 struct xfs_mount *mp = pag_mount(pag);
557 struct xfs_btree_cur *cur;
558 int error;
559 int i;
560 struct xfs_inobt_rec_incore rec;
561
562 cur = xfs_inobt_init_cursor(pag, tp, agbp);
563
564 /* the new record is pre-aligned so we know where to look */
565 error = xfs_inobt_lookup(cur, nrec->ir_startino, XFS_LOOKUP_EQ, &i);
566 if (error)
567 goto error;
568 /* if nothing there, insert a new record and return */
569 if (i == 0) {
570 error = xfs_inobt_insert_rec(cur, nrec->ir_holemask,
571 nrec->ir_count, nrec->ir_freecount,
572 nrec->ir_free, &i);
573 if (error)
574 goto error;
575 if (XFS_IS_CORRUPT(mp, i != 1)) {
576 xfs_btree_mark_sick(cur);
577 error = -EFSCORRUPTED;
578 goto error;
579 }
580
581 goto out;
582 }
583
584 /*
585 * A record exists at this startino. Merge the records.
586 */
587 error = xfs_inobt_get_rec(cur, &rec, &i);
588 if (error)
589 goto error;
590 if (XFS_IS_CORRUPT(mp, i != 1)) {
591 xfs_btree_mark_sick(cur);
592 error = -EFSCORRUPTED;
593 goto error;
594 }
595 if (XFS_IS_CORRUPT(mp, rec.ir_startino != nrec->ir_startino)) {
596 xfs_btree_mark_sick(cur);
597 error = -EFSCORRUPTED;
598 goto error;
599 }
600
601 /*
602 * This should never fail. If we have coexisting records that
603 * cannot merge, something is seriously wrong.
604 */
605 if (XFS_IS_CORRUPT(mp, !__xfs_inobt_can_merge(nrec, &rec))) {
606 xfs_btree_mark_sick(cur);
607 error = -EFSCORRUPTED;
608 goto error;
609 }
610
611 trace_xfs_irec_merge_pre(pag, &rec, nrec);
612
613 /* merge to nrec to output the updated record */
614 __xfs_inobt_rec_merge(nrec, &rec);
615
616 trace_xfs_irec_merge_post(pag, nrec);
617
618 error = xfs_inobt_rec_check_count(mp, nrec);
619 if (error)
620 goto error;
621
622 error = xfs_inobt_update(cur, nrec);
623 if (error)
624 goto error;
625
626out:
627 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
628 return 0;
629error:
630 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
631 return error;
632}
633
634/*
635 * Insert a new sparse inode chunk into the free inode btree. The inode
636 * record for the sparse chunk is pre-aligned to a startino that should match
637 * any pre-existing sparse inode record in the tree. This allows sparse chunks
638 * to fill over time.
639 *
640 * The new record is always inserted, overwriting a pre-existing record if
641 * there is one.
642 */
643STATIC int
644xfs_finobt_insert_sprec(
645 struct xfs_perag *pag,
646 struct xfs_trans *tp,
647 struct xfs_buf *agbp,
648 struct xfs_inobt_rec_incore *nrec) /* in/out: new rec. */
649{
650 struct xfs_mount *mp = pag_mount(pag);
651 struct xfs_btree_cur *cur;
652 int error;
653 int i;
654
655 cur = xfs_finobt_init_cursor(pag, tp, agbp);
656
657 /* the new record is pre-aligned so we know where to look */
658 error = xfs_inobt_lookup(cur, nrec->ir_startino, XFS_LOOKUP_EQ, &i);
659 if (error)
660 goto error;
661 /* if nothing there, insert a new record and return */
662 if (i == 0) {
663 error = xfs_inobt_insert_rec(cur, nrec->ir_holemask,
664 nrec->ir_count, nrec->ir_freecount,
665 nrec->ir_free, &i);
666 if (error)
667 goto error;
668 if (XFS_IS_CORRUPT(mp, i != 1)) {
669 xfs_btree_mark_sick(cur);
670 error = -EFSCORRUPTED;
671 goto error;
672 }
673 } else {
674 error = xfs_inobt_update(cur, nrec);
675 if (error)
676 goto error;
677 }
678
679 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
680 return 0;
681error:
682 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
683 return error;
684}
685
686
687/*
688 * Allocate new inodes in the allocation group specified by agbp. Returns 0 if
689 * inodes were allocated in this AG; -EAGAIN if there was no space in this AG so
690 * the caller knows it can try another AG, a hard -ENOSPC when over the maximum
691 * inode count threshold, or the usual negative error code for other errors.
692 */
693STATIC int
694xfs_ialloc_ag_alloc(
695 struct xfs_perag *pag,
696 struct xfs_trans *tp,
697 struct xfs_buf *agbp)
698{
699 struct xfs_agi *agi;
700 struct xfs_alloc_arg args;
701 int error;
702 xfs_agino_t newino; /* new first inode's number */
703 xfs_agino_t newlen; /* new number of inodes */
704 int isaligned = 0; /* inode allocation at stripe */
705 /* unit boundary */
706 /* init. to full chunk */
707 struct xfs_inobt_rec_incore rec;
708 struct xfs_ino_geometry *igeo = M_IGEO(tp->t_mountp);
709 uint16_t allocmask = (uint16_t) -1;
710 int do_sparse = 0;
711
712 memset(&args, 0, sizeof(args));
713 args.tp = tp;
714 args.mp = tp->t_mountp;
715 args.fsbno = NULLFSBLOCK;
716 args.oinfo = XFS_RMAP_OINFO_INODES;
717 args.pag = pag;
718
719#ifdef DEBUG
720 /* randomly do sparse inode allocations */
721 if (xfs_has_sparseinodes(tp->t_mountp) &&
722 igeo->ialloc_min_blks < igeo->ialloc_blks)
723 do_sparse = get_random_u32_below(2);
724#endif
725
726 /*
727 * Locking will ensure that we don't have two callers in here
728 * at one time.
729 */
730 newlen = igeo->ialloc_inos;
731 if (igeo->maxicount &&
732 percpu_counter_read_positive(&args.mp->m_icount) + newlen >
733 igeo->maxicount)
734 return -ENOSPC;
735 args.minlen = args.maxlen = igeo->ialloc_blks;
736 /*
737 * First try to allocate inodes contiguous with the last-allocated
738 * chunk of inodes. If the filesystem is striped, this will fill
739 * an entire stripe unit with inodes.
740 */
741 agi = agbp->b_addr;
742 newino = be32_to_cpu(agi->agi_newino);
743 args.agbno = XFS_AGINO_TO_AGBNO(args.mp, newino) +
744 igeo->ialloc_blks;
745 if (do_sparse)
746 goto sparse_alloc;
747 if (likely(newino != NULLAGINO &&
748 (args.agbno < be32_to_cpu(agi->agi_length)))) {
749 args.prod = 1;
750
751 /*
752 * We need to take into account alignment here to ensure that
753 * we don't modify the free list if we fail to have an exact
754 * block. If we don't have an exact match, and every oher
755 * attempt allocation attempt fails, we'll end up cancelling
756 * a dirty transaction and shutting down.
757 *
758 * For an exact allocation, alignment must be 1,
759 * however we need to take cluster alignment into account when
760 * fixing up the freelist. Use the minalignslop field to
761 * indicate that extra blocks might be required for alignment,
762 * but not to use them in the actual exact allocation.
763 */
764 args.alignment = 1;
765 args.minalignslop = igeo->cluster_align - 1;
766
767 /* Allow space for the inode btree to split. */
768 args.minleft = igeo->inobt_maxlevels;
769 error = xfs_alloc_vextent_exact_bno(&args,
770 xfs_agbno_to_fsb(pag, args.agbno));
771 if (error)
772 return error;
773
774 /*
775 * This request might have dirtied the transaction if the AG can
776 * satisfy the request, but the exact block was not available.
777 * If the allocation did fail, subsequent requests will relax
778 * the exact agbno requirement and increase the alignment
779 * instead. It is critical that the total size of the request
780 * (len + alignment + slop) does not increase from this point
781 * on, so reset minalignslop to ensure it is not included in
782 * subsequent requests.
783 */
784 args.minalignslop = 0;
785 }
786
787 if (unlikely(args.fsbno == NULLFSBLOCK)) {
788 /*
789 * Set the alignment for the allocation.
790 * If stripe alignment is turned on then align at stripe unit
791 * boundary.
792 * If the cluster size is smaller than a filesystem block
793 * then we're doing I/O for inodes in filesystem block size
794 * pieces, so don't need alignment anyway.
795 */
796 isaligned = 0;
797 if (igeo->ialloc_align) {
798 ASSERT(!xfs_has_noalign(args.mp));
799 args.alignment = args.mp->m_dalign;
800 isaligned = 1;
801 } else
802 args.alignment = igeo->cluster_align;
803 /*
804 * Allocate a fixed-size extent of inodes.
805 */
806 args.prod = 1;
807 /*
808 * Allow space for the inode btree to split.
809 */
810 args.minleft = igeo->inobt_maxlevels;
811 error = xfs_alloc_vextent_near_bno(&args,
812 xfs_agbno_to_fsb(pag,
813 be32_to_cpu(agi->agi_root)));
814 if (error)
815 return error;
816 }
817
818 /*
819 * If stripe alignment is turned on, then try again with cluster
820 * alignment.
821 */
822 if (isaligned && args.fsbno == NULLFSBLOCK) {
823 args.alignment = igeo->cluster_align;
824 error = xfs_alloc_vextent_near_bno(&args,
825 xfs_agbno_to_fsb(pag,
826 be32_to_cpu(agi->agi_root)));
827 if (error)
828 return error;
829 }
830
831 /*
832 * Finally, try a sparse allocation if the filesystem supports it and
833 * the sparse allocation length is smaller than a full chunk.
834 */
835 if (xfs_has_sparseinodes(args.mp) &&
836 igeo->ialloc_min_blks < igeo->ialloc_blks &&
837 args.fsbno == NULLFSBLOCK) {
838sparse_alloc:
839 args.alignment = args.mp->m_sb.sb_spino_align;
840 args.prod = 1;
841
842 args.minlen = igeo->ialloc_min_blks;
843 args.maxlen = args.minlen;
844
845 /*
846 * The inode record will be aligned to full chunk size. We must
847 * prevent sparse allocation from AG boundaries that result in
848 * invalid inode records, such as records that start at agbno 0
849 * or extend beyond the AG.
850 *
851 * Set min agbno to the first aligned, non-zero agbno and max to
852 * the last aligned agbno that is at least one full chunk from
853 * the end of the AG.
854 */
855 args.min_agbno = args.mp->m_sb.sb_inoalignmt;
856 args.max_agbno = round_down(xfs_ag_block_count(args.mp,
857 pag_agno(pag)),
858 args.mp->m_sb.sb_inoalignmt) -
859 igeo->ialloc_blks;
860
861 error = xfs_alloc_vextent_near_bno(&args,
862 xfs_agbno_to_fsb(pag,
863 be32_to_cpu(agi->agi_root)));
864 if (error)
865 return error;
866
867 newlen = XFS_AGB_TO_AGINO(args.mp, args.len);
868 ASSERT(newlen <= XFS_INODES_PER_CHUNK);
869 allocmask = (1 << (newlen / XFS_INODES_PER_HOLEMASK_BIT)) - 1;
870 }
871
872 if (args.fsbno == NULLFSBLOCK)
873 return -EAGAIN;
874
875 ASSERT(args.len == args.minlen);
876
877 /*
878 * Stamp and write the inode buffers.
879 *
880 * Seed the new inode cluster with a random generation number. This
881 * prevents short-term reuse of generation numbers if a chunk is
882 * freed and then immediately reallocated. We use random numbers
883 * rather than a linear progression to prevent the next generation
884 * number from being easily guessable.
885 */
886 error = xfs_ialloc_inode_init(args.mp, tp, NULL, newlen, pag_agno(pag),
887 args.agbno, args.len, get_random_u32());
888
889 if (error)
890 return error;
891 /*
892 * Convert the results.
893 */
894 newino = XFS_AGB_TO_AGINO(args.mp, args.agbno);
895
896 if (xfs_inobt_issparse(~allocmask)) {
897 /*
898 * We've allocated a sparse chunk. Align the startino and mask.
899 */
900 xfs_align_sparse_ino(args.mp, &newino, &allocmask);
901
902 rec.ir_startino = newino;
903 rec.ir_holemask = ~allocmask;
904 rec.ir_count = newlen;
905 rec.ir_freecount = newlen;
906 rec.ir_free = XFS_INOBT_ALL_FREE;
907
908 /*
909 * Insert the sparse record into the inobt and allow for a merge
910 * if necessary. If a merge does occur, rec is updated to the
911 * merged record.
912 */
913 error = xfs_inobt_insert_sprec(pag, tp, agbp, &rec);
914 if (error == -EFSCORRUPTED) {
915 xfs_alert(args.mp,
916 "invalid sparse inode record: ino 0x%llx holemask 0x%x count %u",
917 xfs_agino_to_ino(pag, rec.ir_startino),
918 rec.ir_holemask, rec.ir_count);
919 xfs_force_shutdown(args.mp, SHUTDOWN_CORRUPT_INCORE);
920 }
921 if (error)
922 return error;
923
924 /*
925 * We can't merge the part we've just allocated as for the inobt
926 * due to finobt semantics. The original record may or may not
927 * exist independent of whether physical inodes exist in this
928 * sparse chunk.
929 *
930 * We must update the finobt record based on the inobt record.
931 * rec contains the fully merged and up to date inobt record
932 * from the previous call. Set merge false to replace any
933 * existing record with this one.
934 */
935 if (xfs_has_finobt(args.mp)) {
936 error = xfs_finobt_insert_sprec(pag, tp, agbp, &rec);
937 if (error)
938 return error;
939 }
940 } else {
941 /* full chunk - insert new records to both btrees */
942 error = xfs_inobt_insert(pag, tp, agbp, newino, newlen, false);
943 if (error)
944 return error;
945
946 if (xfs_has_finobt(args.mp)) {
947 error = xfs_inobt_insert(pag, tp, agbp, newino,
948 newlen, true);
949 if (error)
950 return error;
951 }
952 }
953
954 /*
955 * Update AGI counts and newino.
956 */
957 be32_add_cpu(&agi->agi_count, newlen);
958 be32_add_cpu(&agi->agi_freecount, newlen);
959 pag->pagi_freecount += newlen;
960 pag->pagi_count += newlen;
961 agi->agi_newino = cpu_to_be32(newino);
962
963 /*
964 * Log allocation group header fields
965 */
966 xfs_ialloc_log_agi(tp, agbp,
967 XFS_AGI_COUNT | XFS_AGI_FREECOUNT | XFS_AGI_NEWINO);
968 /*
969 * Modify/log superblock values for inode count and inode free count.
970 */
971 xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, (long)newlen);
972 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, (long)newlen);
973 return 0;
974}
975
976/*
977 * Try to retrieve the next record to the left/right from the current one.
978 */
979STATIC int
980xfs_ialloc_next_rec(
981 struct xfs_btree_cur *cur,
982 xfs_inobt_rec_incore_t *rec,
983 int *done,
984 int left)
985{
986 int error;
987 int i;
988
989 if (left)
990 error = xfs_btree_decrement(cur, 0, &i);
991 else
992 error = xfs_btree_increment(cur, 0, &i);
993
994 if (error)
995 return error;
996 *done = !i;
997 if (i) {
998 error = xfs_inobt_get_rec(cur, rec, &i);
999 if (error)
1000 return error;
1001 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1002 xfs_btree_mark_sick(cur);
1003 return -EFSCORRUPTED;
1004 }
1005 }
1006
1007 return 0;
1008}
1009
1010STATIC int
1011xfs_ialloc_get_rec(
1012 struct xfs_btree_cur *cur,
1013 xfs_agino_t agino,
1014 xfs_inobt_rec_incore_t *rec,
1015 int *done)
1016{
1017 int error;
1018 int i;
1019
1020 error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_EQ, &i);
1021 if (error)
1022 return error;
1023 *done = !i;
1024 if (i) {
1025 error = xfs_inobt_get_rec(cur, rec, &i);
1026 if (error)
1027 return error;
1028 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1029 xfs_btree_mark_sick(cur);
1030 return -EFSCORRUPTED;
1031 }
1032 }
1033
1034 return 0;
1035}
1036
1037/*
1038 * Return the offset of the first free inode in the record. If the inode chunk
1039 * is sparsely allocated, we convert the record holemask to inode granularity
1040 * and mask off the unallocated regions from the inode free mask.
1041 */
1042STATIC int
1043xfs_inobt_first_free_inode(
1044 struct xfs_inobt_rec_incore *rec)
1045{
1046 xfs_inofree_t realfree;
1047
1048 /* if there are no holes, return the first available offset */
1049 if (!xfs_inobt_issparse(rec->ir_holemask))
1050 return xfs_lowbit64(rec->ir_free);
1051
1052 realfree = xfs_inobt_irec_to_allocmask(rec);
1053 realfree &= rec->ir_free;
1054
1055 return xfs_lowbit64(realfree);
1056}
1057
1058/*
1059 * If this AG has corrupt inodes, check if allocating this inode would fail
1060 * with corruption errors. Returns 0 if we're clear, or EAGAIN to try again
1061 * somewhere else.
1062 */
1063static int
1064xfs_dialloc_check_ino(
1065 struct xfs_perag *pag,
1066 struct xfs_trans *tp,
1067 xfs_ino_t ino)
1068{
1069 struct xfs_imap imap;
1070 struct xfs_buf *bp;
1071 int error;
1072
1073 error = xfs_imap(pag, tp, ino, &imap, 0);
1074 if (error)
1075 return -EAGAIN;
1076
1077 error = xfs_imap_to_bp(pag_mount(pag), tp, &imap, &bp);
1078 if (error)
1079 return -EAGAIN;
1080
1081 xfs_trans_brelse(tp, bp);
1082 return 0;
1083}
1084
1085/*
1086 * Allocate an inode using the inobt-only algorithm.
1087 */
1088STATIC int
1089xfs_dialloc_ag_inobt(
1090 struct xfs_perag *pag,
1091 struct xfs_trans *tp,
1092 struct xfs_buf *agbp,
1093 xfs_ino_t parent,
1094 xfs_ino_t *inop)
1095{
1096 struct xfs_mount *mp = tp->t_mountp;
1097 struct xfs_agi *agi = agbp->b_addr;
1098 xfs_agnumber_t pagno = XFS_INO_TO_AGNO(mp, parent);
1099 xfs_agino_t pagino = XFS_INO_TO_AGINO(mp, parent);
1100 struct xfs_btree_cur *cur, *tcur;
1101 struct xfs_inobt_rec_incore rec, trec;
1102 xfs_ino_t ino;
1103 int error;
1104 int offset;
1105 int i, j;
1106 int searchdistance = 10;
1107
1108 ASSERT(xfs_perag_initialised_agi(pag));
1109 ASSERT(xfs_perag_allows_inodes(pag));
1110 ASSERT(pag->pagi_freecount > 0);
1111
1112 restart_pagno:
1113 cur = xfs_inobt_init_cursor(pag, tp, agbp);
1114 /*
1115 * If pagino is 0 (this is the root inode allocation) use newino.
1116 * This must work because we've just allocated some.
1117 */
1118 if (!pagino)
1119 pagino = be32_to_cpu(agi->agi_newino);
1120
1121 error = xfs_check_agi_freecount(cur);
1122 if (error)
1123 goto error0;
1124
1125 /*
1126 * If in the same AG as the parent, try to get near the parent.
1127 */
1128 if (pagno == pag_agno(pag)) {
1129 int doneleft; /* done, to the left */
1130 int doneright; /* done, to the right */
1131
1132 error = xfs_inobt_lookup(cur, pagino, XFS_LOOKUP_LE, &i);
1133 if (error)
1134 goto error0;
1135 if (XFS_IS_CORRUPT(mp, i != 1)) {
1136 xfs_btree_mark_sick(cur);
1137 error = -EFSCORRUPTED;
1138 goto error0;
1139 }
1140
1141 error = xfs_inobt_get_rec(cur, &rec, &j);
1142 if (error)
1143 goto error0;
1144 if (XFS_IS_CORRUPT(mp, j != 1)) {
1145 xfs_btree_mark_sick(cur);
1146 error = -EFSCORRUPTED;
1147 goto error0;
1148 }
1149
1150 if (rec.ir_freecount > 0) {
1151 /*
1152 * Found a free inode in the same chunk
1153 * as the parent, done.
1154 */
1155 goto alloc_inode;
1156 }
1157
1158
1159 /*
1160 * In the same AG as parent, but parent's chunk is full.
1161 */
1162
1163 /* duplicate the cursor, search left & right simultaneously */
1164 error = xfs_btree_dup_cursor(cur, &tcur);
1165 if (error)
1166 goto error0;
1167
1168 /*
1169 * Skip to last blocks looked up if same parent inode.
1170 */
1171 if (pagino != NULLAGINO &&
1172 pag->pagl_pagino == pagino &&
1173 pag->pagl_leftrec != NULLAGINO &&
1174 pag->pagl_rightrec != NULLAGINO) {
1175 error = xfs_ialloc_get_rec(tcur, pag->pagl_leftrec,
1176 &trec, &doneleft);
1177 if (error)
1178 goto error1;
1179
1180 error = xfs_ialloc_get_rec(cur, pag->pagl_rightrec,
1181 &rec, &doneright);
1182 if (error)
1183 goto error1;
1184 } else {
1185 /* search left with tcur, back up 1 record */
1186 error = xfs_ialloc_next_rec(tcur, &trec, &doneleft, 1);
1187 if (error)
1188 goto error1;
1189
1190 /* search right with cur, go forward 1 record. */
1191 error = xfs_ialloc_next_rec(cur, &rec, &doneright, 0);
1192 if (error)
1193 goto error1;
1194 }
1195
1196 /*
1197 * Loop until we find an inode chunk with a free inode.
1198 */
1199 while (--searchdistance > 0 && (!doneleft || !doneright)) {
1200 int useleft; /* using left inode chunk this time */
1201
1202 /* figure out the closer block if both are valid. */
1203 if (!doneleft && !doneright) {
1204 useleft = pagino -
1205 (trec.ir_startino + XFS_INODES_PER_CHUNK - 1) <
1206 rec.ir_startino - pagino;
1207 } else {
1208 useleft = !doneleft;
1209 }
1210
1211 /* free inodes to the left? */
1212 if (useleft && trec.ir_freecount) {
1213 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1214 cur = tcur;
1215
1216 pag->pagl_leftrec = trec.ir_startino;
1217 pag->pagl_rightrec = rec.ir_startino;
1218 pag->pagl_pagino = pagino;
1219 rec = trec;
1220 goto alloc_inode;
1221 }
1222
1223 /* free inodes to the right? */
1224 if (!useleft && rec.ir_freecount) {
1225 xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1226
1227 pag->pagl_leftrec = trec.ir_startino;
1228 pag->pagl_rightrec = rec.ir_startino;
1229 pag->pagl_pagino = pagino;
1230 goto alloc_inode;
1231 }
1232
1233 /* get next record to check */
1234 if (useleft) {
1235 error = xfs_ialloc_next_rec(tcur, &trec,
1236 &doneleft, 1);
1237 } else {
1238 error = xfs_ialloc_next_rec(cur, &rec,
1239 &doneright, 0);
1240 }
1241 if (error)
1242 goto error1;
1243 }
1244
1245 if (searchdistance <= 0) {
1246 /*
1247 * Not in range - save last search
1248 * location and allocate a new inode
1249 */
1250 xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1251 pag->pagl_leftrec = trec.ir_startino;
1252 pag->pagl_rightrec = rec.ir_startino;
1253 pag->pagl_pagino = pagino;
1254
1255 } else {
1256 /*
1257 * We've reached the end of the btree. because
1258 * we are only searching a small chunk of the
1259 * btree each search, there is obviously free
1260 * inodes closer to the parent inode than we
1261 * are now. restart the search again.
1262 */
1263 pag->pagl_pagino = NULLAGINO;
1264 pag->pagl_leftrec = NULLAGINO;
1265 pag->pagl_rightrec = NULLAGINO;
1266 xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1267 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1268 goto restart_pagno;
1269 }
1270 }
1271
1272 /*
1273 * In a different AG from the parent.
1274 * See if the most recently allocated block has any free.
1275 */
1276 if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1277 error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1278 XFS_LOOKUP_EQ, &i);
1279 if (error)
1280 goto error0;
1281
1282 if (i == 1) {
1283 error = xfs_inobt_get_rec(cur, &rec, &j);
1284 if (error)
1285 goto error0;
1286
1287 if (j == 1 && rec.ir_freecount > 0) {
1288 /*
1289 * The last chunk allocated in the group
1290 * still has a free inode.
1291 */
1292 goto alloc_inode;
1293 }
1294 }
1295 }
1296
1297 /*
1298 * None left in the last group, search the whole AG
1299 */
1300 error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1301 if (error)
1302 goto error0;
1303 if (XFS_IS_CORRUPT(mp, i != 1)) {
1304 xfs_btree_mark_sick(cur);
1305 error = -EFSCORRUPTED;
1306 goto error0;
1307 }
1308
1309 for (;;) {
1310 error = xfs_inobt_get_rec(cur, &rec, &i);
1311 if (error)
1312 goto error0;
1313 if (XFS_IS_CORRUPT(mp, i != 1)) {
1314 xfs_btree_mark_sick(cur);
1315 error = -EFSCORRUPTED;
1316 goto error0;
1317 }
1318 if (rec.ir_freecount > 0)
1319 break;
1320 error = xfs_btree_increment(cur, 0, &i);
1321 if (error)
1322 goto error0;
1323 if (XFS_IS_CORRUPT(mp, i != 1)) {
1324 xfs_btree_mark_sick(cur);
1325 error = -EFSCORRUPTED;
1326 goto error0;
1327 }
1328 }
1329
1330alloc_inode:
1331 offset = xfs_inobt_first_free_inode(&rec);
1332 ASSERT(offset >= 0);
1333 ASSERT(offset < XFS_INODES_PER_CHUNK);
1334 ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1335 XFS_INODES_PER_CHUNK) == 0);
1336 ino = xfs_agino_to_ino(pag, rec.ir_startino + offset);
1337
1338 if (xfs_ag_has_sickness(pag, XFS_SICK_AG_INODES)) {
1339 error = xfs_dialloc_check_ino(pag, tp, ino);
1340 if (error)
1341 goto error0;
1342 }
1343
1344 rec.ir_free &= ~XFS_INOBT_MASK(offset);
1345 rec.ir_freecount--;
1346 error = xfs_inobt_update(cur, &rec);
1347 if (error)
1348 goto error0;
1349 be32_add_cpu(&agi->agi_freecount, -1);
1350 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1351 pag->pagi_freecount--;
1352
1353 error = xfs_check_agi_freecount(cur);
1354 if (error)
1355 goto error0;
1356
1357 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1358 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1359 *inop = ino;
1360 return 0;
1361error1:
1362 xfs_btree_del_cursor(tcur, XFS_BTREE_ERROR);
1363error0:
1364 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1365 return error;
1366}
1367
1368/*
1369 * Use the free inode btree to allocate an inode based on distance from the
1370 * parent. Note that the provided cursor may be deleted and replaced.
1371 */
1372STATIC int
1373xfs_dialloc_ag_finobt_near(
1374 xfs_agino_t pagino,
1375 struct xfs_btree_cur **ocur,
1376 struct xfs_inobt_rec_incore *rec)
1377{
1378 struct xfs_btree_cur *lcur = *ocur; /* left search cursor */
1379 struct xfs_btree_cur *rcur; /* right search cursor */
1380 struct xfs_inobt_rec_incore rrec;
1381 int error;
1382 int i, j;
1383
1384 error = xfs_inobt_lookup(lcur, pagino, XFS_LOOKUP_LE, &i);
1385 if (error)
1386 return error;
1387
1388 if (i == 1) {
1389 error = xfs_inobt_get_rec(lcur, rec, &i);
1390 if (error)
1391 return error;
1392 if (XFS_IS_CORRUPT(lcur->bc_mp, i != 1)) {
1393 xfs_btree_mark_sick(lcur);
1394 return -EFSCORRUPTED;
1395 }
1396
1397 /*
1398 * See if we've landed in the parent inode record. The finobt
1399 * only tracks chunks with at least one free inode, so record
1400 * existence is enough.
1401 */
1402 if (pagino >= rec->ir_startino &&
1403 pagino < (rec->ir_startino + XFS_INODES_PER_CHUNK))
1404 return 0;
1405 }
1406
1407 error = xfs_btree_dup_cursor(lcur, &rcur);
1408 if (error)
1409 return error;
1410
1411 error = xfs_inobt_lookup(rcur, pagino, XFS_LOOKUP_GE, &j);
1412 if (error)
1413 goto error_rcur;
1414 if (j == 1) {
1415 error = xfs_inobt_get_rec(rcur, &rrec, &j);
1416 if (error)
1417 goto error_rcur;
1418 if (XFS_IS_CORRUPT(lcur->bc_mp, j != 1)) {
1419 xfs_btree_mark_sick(lcur);
1420 error = -EFSCORRUPTED;
1421 goto error_rcur;
1422 }
1423 }
1424
1425 if (XFS_IS_CORRUPT(lcur->bc_mp, i != 1 && j != 1)) {
1426 xfs_btree_mark_sick(lcur);
1427 error = -EFSCORRUPTED;
1428 goto error_rcur;
1429 }
1430 if (i == 1 && j == 1) {
1431 /*
1432 * Both the left and right records are valid. Choose the closer
1433 * inode chunk to the target.
1434 */
1435 if ((pagino - rec->ir_startino + XFS_INODES_PER_CHUNK - 1) >
1436 (rrec.ir_startino - pagino)) {
1437 *rec = rrec;
1438 xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1439 *ocur = rcur;
1440 } else {
1441 xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1442 }
1443 } else if (j == 1) {
1444 /* only the right record is valid */
1445 *rec = rrec;
1446 xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1447 *ocur = rcur;
1448 } else if (i == 1) {
1449 /* only the left record is valid */
1450 xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1451 }
1452
1453 return 0;
1454
1455error_rcur:
1456 xfs_btree_del_cursor(rcur, XFS_BTREE_ERROR);
1457 return error;
1458}
1459
1460/*
1461 * Use the free inode btree to find a free inode based on a newino hint. If
1462 * the hint is NULL, find the first free inode in the AG.
1463 */
1464STATIC int
1465xfs_dialloc_ag_finobt_newino(
1466 struct xfs_agi *agi,
1467 struct xfs_btree_cur *cur,
1468 struct xfs_inobt_rec_incore *rec)
1469{
1470 int error;
1471 int i;
1472
1473 if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1474 error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1475 XFS_LOOKUP_EQ, &i);
1476 if (error)
1477 return error;
1478 if (i == 1) {
1479 error = xfs_inobt_get_rec(cur, rec, &i);
1480 if (error)
1481 return error;
1482 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1483 xfs_btree_mark_sick(cur);
1484 return -EFSCORRUPTED;
1485 }
1486 return 0;
1487 }
1488 }
1489
1490 /*
1491 * Find the first inode available in the AG.
1492 */
1493 error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1494 if (error)
1495 return error;
1496 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1497 xfs_btree_mark_sick(cur);
1498 return -EFSCORRUPTED;
1499 }
1500
1501 error = xfs_inobt_get_rec(cur, rec, &i);
1502 if (error)
1503 return error;
1504 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1505 xfs_btree_mark_sick(cur);
1506 return -EFSCORRUPTED;
1507 }
1508
1509 return 0;
1510}
1511
1512/*
1513 * Update the inobt based on a modification made to the finobt. Also ensure that
1514 * the records from both trees are equivalent post-modification.
1515 */
1516STATIC int
1517xfs_dialloc_ag_update_inobt(
1518 struct xfs_btree_cur *cur, /* inobt cursor */
1519 struct xfs_inobt_rec_incore *frec, /* finobt record */
1520 int offset) /* inode offset */
1521{
1522 struct xfs_inobt_rec_incore rec;
1523 int error;
1524 int i;
1525
1526 error = xfs_inobt_lookup(cur, frec->ir_startino, XFS_LOOKUP_EQ, &i);
1527 if (error)
1528 return error;
1529 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1530 xfs_btree_mark_sick(cur);
1531 return -EFSCORRUPTED;
1532 }
1533
1534 error = xfs_inobt_get_rec(cur, &rec, &i);
1535 if (error)
1536 return error;
1537 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1538 xfs_btree_mark_sick(cur);
1539 return -EFSCORRUPTED;
1540 }
1541 ASSERT((XFS_AGINO_TO_OFFSET(cur->bc_mp, rec.ir_startino) %
1542 XFS_INODES_PER_CHUNK) == 0);
1543
1544 rec.ir_free &= ~XFS_INOBT_MASK(offset);
1545 rec.ir_freecount--;
1546
1547 if (XFS_IS_CORRUPT(cur->bc_mp,
1548 rec.ir_free != frec->ir_free ||
1549 rec.ir_freecount != frec->ir_freecount)) {
1550 xfs_btree_mark_sick(cur);
1551 return -EFSCORRUPTED;
1552 }
1553
1554 return xfs_inobt_update(cur, &rec);
1555}
1556
1557/*
1558 * Allocate an inode using the free inode btree, if available. Otherwise, fall
1559 * back to the inobt search algorithm.
1560 *
1561 * The caller selected an AG for us, and made sure that free inodes are
1562 * available.
1563 */
1564static int
1565xfs_dialloc_ag(
1566 struct xfs_perag *pag,
1567 struct xfs_trans *tp,
1568 struct xfs_buf *agbp,
1569 xfs_ino_t parent,
1570 xfs_ino_t *inop)
1571{
1572 struct xfs_mount *mp = tp->t_mountp;
1573 struct xfs_agi *agi = agbp->b_addr;
1574 xfs_agnumber_t pagno = XFS_INO_TO_AGNO(mp, parent);
1575 xfs_agino_t pagino = XFS_INO_TO_AGINO(mp, parent);
1576 struct xfs_btree_cur *cur; /* finobt cursor */
1577 struct xfs_btree_cur *icur; /* inobt cursor */
1578 struct xfs_inobt_rec_incore rec;
1579 xfs_ino_t ino;
1580 int error;
1581 int offset;
1582 int i;
1583
1584 if (!xfs_has_finobt(mp))
1585 return xfs_dialloc_ag_inobt(pag, tp, agbp, parent, inop);
1586
1587 /*
1588 * If pagino is 0 (this is the root inode allocation) use newino.
1589 * This must work because we've just allocated some.
1590 */
1591 if (!pagino)
1592 pagino = be32_to_cpu(agi->agi_newino);
1593
1594 cur = xfs_finobt_init_cursor(pag, tp, agbp);
1595
1596 error = xfs_check_agi_freecount(cur);
1597 if (error)
1598 goto error_cur;
1599
1600 /*
1601 * The search algorithm depends on whether we're in the same AG as the
1602 * parent. If so, find the closest available inode to the parent. If
1603 * not, consider the agi hint or find the first free inode in the AG.
1604 */
1605 if (pag_agno(pag) == pagno)
1606 error = xfs_dialloc_ag_finobt_near(pagino, &cur, &rec);
1607 else
1608 error = xfs_dialloc_ag_finobt_newino(agi, cur, &rec);
1609 if (error)
1610 goto error_cur;
1611
1612 offset = xfs_inobt_first_free_inode(&rec);
1613 ASSERT(offset >= 0);
1614 ASSERT(offset < XFS_INODES_PER_CHUNK);
1615 ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1616 XFS_INODES_PER_CHUNK) == 0);
1617 ino = xfs_agino_to_ino(pag, rec.ir_startino + offset);
1618
1619 if (xfs_ag_has_sickness(pag, XFS_SICK_AG_INODES)) {
1620 error = xfs_dialloc_check_ino(pag, tp, ino);
1621 if (error)
1622 goto error_cur;
1623 }
1624
1625 /*
1626 * Modify or remove the finobt record.
1627 */
1628 rec.ir_free &= ~XFS_INOBT_MASK(offset);
1629 rec.ir_freecount--;
1630 if (rec.ir_freecount)
1631 error = xfs_inobt_update(cur, &rec);
1632 else
1633 error = xfs_btree_delete(cur, &i);
1634 if (error)
1635 goto error_cur;
1636
1637 /*
1638 * The finobt has now been updated appropriately. We haven't updated the
1639 * agi and superblock yet, so we can create an inobt cursor and validate
1640 * the original freecount. If all is well, make the equivalent update to
1641 * the inobt using the finobt record and offset information.
1642 */
1643 icur = xfs_inobt_init_cursor(pag, tp, agbp);
1644
1645 error = xfs_check_agi_freecount(icur);
1646 if (error)
1647 goto error_icur;
1648
1649 error = xfs_dialloc_ag_update_inobt(icur, &rec, offset);
1650 if (error)
1651 goto error_icur;
1652
1653 /*
1654 * Both trees have now been updated. We must update the perag and
1655 * superblock before we can check the freecount for each btree.
1656 */
1657 be32_add_cpu(&agi->agi_freecount, -1);
1658 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1659 pag->pagi_freecount--;
1660
1661 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1662
1663 error = xfs_check_agi_freecount(icur);
1664 if (error)
1665 goto error_icur;
1666 error = xfs_check_agi_freecount(cur);
1667 if (error)
1668 goto error_icur;
1669
1670 xfs_btree_del_cursor(icur, XFS_BTREE_NOERROR);
1671 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1672 *inop = ino;
1673 return 0;
1674
1675error_icur:
1676 xfs_btree_del_cursor(icur, XFS_BTREE_ERROR);
1677error_cur:
1678 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1679 return error;
1680}
1681
1682static int
1683xfs_dialloc_roll(
1684 struct xfs_trans **tpp,
1685 struct xfs_buf *agibp)
1686{
1687 struct xfs_trans *tp = *tpp;
1688 struct xfs_dquot_acct *dqinfo;
1689 int error;
1690
1691 /*
1692 * Hold to on to the agibp across the commit so no other allocation can
1693 * come in and take the free inodes we just allocated for our caller.
1694 */
1695 xfs_trans_bhold(tp, agibp);
1696
1697 /*
1698 * We want the quota changes to be associated with the next transaction,
1699 * NOT this one. So, detach the dqinfo from this and attach it to the
1700 * next transaction.
1701 */
1702 dqinfo = tp->t_dqinfo;
1703 tp->t_dqinfo = NULL;
1704
1705 error = xfs_trans_roll(&tp);
1706
1707 /* Re-attach the quota info that we detached from prev trx. */
1708 tp->t_dqinfo = dqinfo;
1709
1710 /*
1711 * Join the buffer even on commit error so that the buffer is released
1712 * when the caller cancels the transaction and doesn't have to handle
1713 * this error case specially.
1714 */
1715 xfs_trans_bjoin(tp, agibp);
1716 *tpp = tp;
1717 return error;
1718}
1719
1720static bool
1721xfs_dialloc_good_ag(
1722 struct xfs_perag *pag,
1723 struct xfs_trans *tp,
1724 umode_t mode,
1725 int flags,
1726 bool ok_alloc)
1727{
1728 struct xfs_mount *mp = tp->t_mountp;
1729 xfs_extlen_t ineed;
1730 xfs_extlen_t longest = 0;
1731 int needspace;
1732 int error;
1733
1734 if (!pag)
1735 return false;
1736 if (!xfs_perag_allows_inodes(pag))
1737 return false;
1738
1739 if (!xfs_perag_initialised_agi(pag)) {
1740 error = xfs_ialloc_read_agi(pag, tp, 0, NULL);
1741 if (error)
1742 return false;
1743 }
1744
1745 if (pag->pagi_freecount)
1746 return true;
1747 if (!ok_alloc)
1748 return false;
1749
1750 if (!xfs_perag_initialised_agf(pag)) {
1751 error = xfs_alloc_read_agf(pag, tp, flags, NULL);
1752 if (error)
1753 return false;
1754 }
1755
1756 /*
1757 * Check that there is enough free space for the file plus a chunk of
1758 * inodes if we need to allocate some. If this is the first pass across
1759 * the AGs, take into account the potential space needed for alignment
1760 * of inode chunks when checking the longest contiguous free space in
1761 * the AG - this prevents us from getting ENOSPC because we have free
1762 * space larger than ialloc_blks but alignment constraints prevent us
1763 * from using it.
1764 *
1765 * If we can't find an AG with space for full alignment slack to be
1766 * taken into account, we must be near ENOSPC in all AGs. Hence we
1767 * don't include alignment for the second pass and so if we fail
1768 * allocation due to alignment issues then it is most likely a real
1769 * ENOSPC condition.
1770 *
1771 * XXX(dgc): this calculation is now bogus thanks to the per-ag
1772 * reservations that xfs_alloc_fix_freelist() now does via
1773 * xfs_alloc_space_available(). When the AG fills up, pagf_freeblks will
1774 * be more than large enough for the check below to succeed, but
1775 * xfs_alloc_space_available() will fail because of the non-zero
1776 * metadata reservation and hence we won't actually be able to allocate
1777 * more inodes in this AG. We do soooo much unnecessary work near ENOSPC
1778 * because of this.
1779 */
1780 ineed = M_IGEO(mp)->ialloc_min_blks;
1781 if (flags && ineed > 1)
1782 ineed += M_IGEO(mp)->cluster_align;
1783 longest = pag->pagf_longest;
1784 if (!longest)
1785 longest = pag->pagf_flcount > 0;
1786 needspace = S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode);
1787
1788 if (pag->pagf_freeblks < needspace + ineed || longest < ineed)
1789 return false;
1790 return true;
1791}
1792
1793static int
1794xfs_dialloc_try_ag(
1795 struct xfs_perag *pag,
1796 struct xfs_trans **tpp,
1797 xfs_ino_t parent,
1798 xfs_ino_t *new_ino,
1799 bool ok_alloc)
1800{
1801 struct xfs_buf *agbp;
1802 xfs_ino_t ino;
1803 int error;
1804
1805 /*
1806 * Then read in the AGI buffer and recheck with the AGI buffer
1807 * lock held.
1808 */
1809 error = xfs_ialloc_read_agi(pag, *tpp, 0, &agbp);
1810 if (error)
1811 return error;
1812
1813 if (!pag->pagi_freecount) {
1814 if (!ok_alloc) {
1815 error = -EAGAIN;
1816 goto out_release;
1817 }
1818
1819 error = xfs_ialloc_ag_alloc(pag, *tpp, agbp);
1820 if (error < 0)
1821 goto out_release;
1822
1823 /*
1824 * We successfully allocated space for an inode cluster in this
1825 * AG. Roll the transaction so that we can allocate one of the
1826 * new inodes.
1827 */
1828 ASSERT(pag->pagi_freecount > 0);
1829 error = xfs_dialloc_roll(tpp, agbp);
1830 if (error)
1831 goto out_release;
1832 }
1833
1834 /* Allocate an inode in the found AG */
1835 error = xfs_dialloc_ag(pag, *tpp, agbp, parent, &ino);
1836 if (!error)
1837 *new_ino = ino;
1838 return error;
1839
1840out_release:
1841 xfs_trans_brelse(*tpp, agbp);
1842 return error;
1843}
1844
1845/*
1846 * Pick an AG for the new inode.
1847 *
1848 * Directories, symlinks, and regular files frequently allocate at least one
1849 * block, so factor that potential expansion when we examine whether an AG has
1850 * enough space for file creation. Try to keep metadata files all in the same
1851 * AG.
1852 */
1853static inline xfs_agnumber_t
1854xfs_dialloc_pick_ag(
1855 struct xfs_mount *mp,
1856 struct xfs_inode *dp,
1857 umode_t mode)
1858{
1859 xfs_agnumber_t start_agno;
1860
1861 if (!dp)
1862 return 0;
1863 if (xfs_is_metadir_inode(dp)) {
1864 if (mp->m_sb.sb_logstart)
1865 return XFS_FSB_TO_AGNO(mp, mp->m_sb.sb_logstart);
1866 return 0;
1867 }
1868
1869 if (S_ISDIR(mode))
1870 return (atomic_inc_return(&mp->m_agirotor) - 1) % mp->m_maxagi;
1871
1872 start_agno = XFS_INO_TO_AGNO(mp, dp->i_ino);
1873 if (start_agno >= mp->m_maxagi)
1874 start_agno = 0;
1875
1876 return start_agno;
1877}
1878
1879/*
1880 * Allocate an on-disk inode.
1881 *
1882 * Mode is used to tell whether the new inode is a directory and hence where to
1883 * locate it. The on-disk inode that is allocated will be returned in @new_ino
1884 * on success, otherwise an error will be set to indicate the failure (e.g.
1885 * -ENOSPC).
1886 */
1887int
1888xfs_dialloc(
1889 struct xfs_trans **tpp,
1890 const struct xfs_icreate_args *args,
1891 xfs_ino_t *new_ino)
1892{
1893 struct xfs_mount *mp = (*tpp)->t_mountp;
1894 struct xfs_perag *pag;
1895 struct xfs_ino_geometry *igeo = M_IGEO(mp);
1896 xfs_ino_t ino = NULLFSINO;
1897 xfs_ino_t parent = args->pip ? args->pip->i_ino : 0;
1898 xfs_agnumber_t agno;
1899 xfs_agnumber_t start_agno;
1900 umode_t mode = args->mode & S_IFMT;
1901 bool ok_alloc = true;
1902 bool low_space = false;
1903 int flags;
1904 int error = 0;
1905
1906 start_agno = xfs_dialloc_pick_ag(mp, args->pip, mode);
1907
1908 /*
1909 * If we have already hit the ceiling of inode blocks then clear
1910 * ok_alloc so we scan all available agi structures for a free
1911 * inode.
1912 *
1913 * Read rough value of mp->m_icount by percpu_counter_read_positive,
1914 * which will sacrifice the preciseness but improve the performance.
1915 */
1916 if (igeo->maxicount &&
1917 percpu_counter_read_positive(&mp->m_icount) + igeo->ialloc_inos
1918 > igeo->maxicount) {
1919 ok_alloc = false;
1920 }
1921
1922 /*
1923 * If we are near to ENOSPC, we want to prefer allocation from AGs that
1924 * have free inodes in them rather than use up free space allocating new
1925 * inode chunks. Hence we turn off allocation for the first non-blocking
1926 * pass through the AGs if we are near ENOSPC to consume free inodes
1927 * that we can immediately allocate, but then we allow allocation on the
1928 * second pass if we fail to find an AG with free inodes in it.
1929 */
1930 if (percpu_counter_read_positive(&mp->m_fdblocks) <
1931 mp->m_low_space[XFS_LOWSP_1_PCNT]) {
1932 ok_alloc = false;
1933 low_space = true;
1934 }
1935
1936 /*
1937 * Loop until we find an allocation group that either has free inodes
1938 * or in which we can allocate some inodes. Iterate through the
1939 * allocation groups upward, wrapping at the end.
1940 */
1941 flags = XFS_ALLOC_FLAG_TRYLOCK;
1942retry:
1943 for_each_perag_wrap_at(mp, start_agno, mp->m_maxagi, agno, pag) {
1944 if (xfs_dialloc_good_ag(pag, *tpp, mode, flags, ok_alloc)) {
1945 error = xfs_dialloc_try_ag(pag, tpp, parent,
1946 &ino, ok_alloc);
1947 if (error != -EAGAIN)
1948 break;
1949 error = 0;
1950 }
1951
1952 if (xfs_is_shutdown(mp)) {
1953 error = -EFSCORRUPTED;
1954 break;
1955 }
1956 }
1957 if (pag)
1958 xfs_perag_rele(pag);
1959 if (error)
1960 return error;
1961 if (ino == NULLFSINO) {
1962 if (flags) {
1963 flags = 0;
1964 if (low_space)
1965 ok_alloc = true;
1966 goto retry;
1967 }
1968 return -ENOSPC;
1969 }
1970
1971 /*
1972 * Protect against obviously corrupt allocation btree records. Later
1973 * xfs_iget checks will catch re-allocation of other active in-memory
1974 * and on-disk inodes. If we don't catch reallocating the parent inode
1975 * here we will deadlock in xfs_iget() so we have to do these checks
1976 * first.
1977 */
1978 if (ino == parent || !xfs_verify_dir_ino(mp, ino)) {
1979 xfs_alert(mp, "Allocated a known in-use inode 0x%llx!", ino);
1980 xfs_agno_mark_sick(mp, XFS_INO_TO_AGNO(mp, ino),
1981 XFS_SICK_AG_INOBT);
1982 return -EFSCORRUPTED;
1983 }
1984
1985 *new_ino = ino;
1986 return 0;
1987}
1988
1989/*
1990 * Free the blocks of an inode chunk. We must consider that the inode chunk
1991 * might be sparse and only free the regions that are allocated as part of the
1992 * chunk.
1993 */
1994static int
1995xfs_difree_inode_chunk(
1996 struct xfs_trans *tp,
1997 struct xfs_perag *pag,
1998 struct xfs_inobt_rec_incore *rec)
1999{
2000 struct xfs_mount *mp = tp->t_mountp;
2001 xfs_agblock_t sagbno = XFS_AGINO_TO_AGBNO(mp,
2002 rec->ir_startino);
2003 int startidx, endidx;
2004 int nextbit;
2005 xfs_agblock_t agbno;
2006 int contigblk;
2007 DECLARE_BITMAP(holemask, XFS_INOBT_HOLEMASK_BITS);
2008
2009 if (!xfs_inobt_issparse(rec->ir_holemask)) {
2010 /* not sparse, calculate extent info directly */
2011 return xfs_free_extent_later(tp, xfs_agbno_to_fsb(pag, sagbno),
2012 M_IGEO(mp)->ialloc_blks, &XFS_RMAP_OINFO_INODES,
2013 XFS_AG_RESV_NONE, 0);
2014 }
2015
2016 /* holemask is only 16-bits (fits in an unsigned long) */
2017 ASSERT(sizeof(rec->ir_holemask) <= sizeof(holemask[0]));
2018 holemask[0] = rec->ir_holemask;
2019
2020 /*
2021 * Find contiguous ranges of zeroes (i.e., allocated regions) in the
2022 * holemask and convert the start/end index of each range to an extent.
2023 * We start with the start and end index both pointing at the first 0 in
2024 * the mask.
2025 */
2026 startidx = endidx = find_first_zero_bit(holemask,
2027 XFS_INOBT_HOLEMASK_BITS);
2028 nextbit = startidx + 1;
2029 while (startidx < XFS_INOBT_HOLEMASK_BITS) {
2030 int error;
2031
2032 nextbit = find_next_zero_bit(holemask, XFS_INOBT_HOLEMASK_BITS,
2033 nextbit);
2034 /*
2035 * If the next zero bit is contiguous, update the end index of
2036 * the current range and continue.
2037 */
2038 if (nextbit != XFS_INOBT_HOLEMASK_BITS &&
2039 nextbit == endidx + 1) {
2040 endidx = nextbit;
2041 goto next;
2042 }
2043
2044 /*
2045 * nextbit is not contiguous with the current end index. Convert
2046 * the current start/end to an extent and add it to the free
2047 * list.
2048 */
2049 agbno = sagbno + (startidx * XFS_INODES_PER_HOLEMASK_BIT) /
2050 mp->m_sb.sb_inopblock;
2051 contigblk = ((endidx - startidx + 1) *
2052 XFS_INODES_PER_HOLEMASK_BIT) /
2053 mp->m_sb.sb_inopblock;
2054
2055 ASSERT(agbno % mp->m_sb.sb_spino_align == 0);
2056 ASSERT(contigblk % mp->m_sb.sb_spino_align == 0);
2057 error = xfs_free_extent_later(tp, xfs_agbno_to_fsb(pag, agbno),
2058 contigblk, &XFS_RMAP_OINFO_INODES,
2059 XFS_AG_RESV_NONE, 0);
2060 if (error)
2061 return error;
2062
2063 /* reset range to current bit and carry on... */
2064 startidx = endidx = nextbit;
2065
2066next:
2067 nextbit++;
2068 }
2069 return 0;
2070}
2071
2072STATIC int
2073xfs_difree_inobt(
2074 struct xfs_perag *pag,
2075 struct xfs_trans *tp,
2076 struct xfs_buf *agbp,
2077 xfs_agino_t agino,
2078 struct xfs_icluster *xic,
2079 struct xfs_inobt_rec_incore *orec)
2080{
2081 struct xfs_mount *mp = pag_mount(pag);
2082 struct xfs_agi *agi = agbp->b_addr;
2083 struct xfs_btree_cur *cur;
2084 struct xfs_inobt_rec_incore rec;
2085 int ilen;
2086 int error;
2087 int i;
2088 int off;
2089
2090 ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
2091 ASSERT(XFS_AGINO_TO_AGBNO(mp, agino) < be32_to_cpu(agi->agi_length));
2092
2093 /*
2094 * Initialize the cursor.
2095 */
2096 cur = xfs_inobt_init_cursor(pag, tp, agbp);
2097
2098 error = xfs_check_agi_freecount(cur);
2099 if (error)
2100 goto error0;
2101
2102 /*
2103 * Look for the entry describing this inode.
2104 */
2105 if ((error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i))) {
2106 xfs_warn(mp, "%s: xfs_inobt_lookup() returned error %d.",
2107 __func__, error);
2108 goto error0;
2109 }
2110 if (XFS_IS_CORRUPT(mp, i != 1)) {
2111 xfs_btree_mark_sick(cur);
2112 error = -EFSCORRUPTED;
2113 goto error0;
2114 }
2115 error = xfs_inobt_get_rec(cur, &rec, &i);
2116 if (error) {
2117 xfs_warn(mp, "%s: xfs_inobt_get_rec() returned error %d.",
2118 __func__, error);
2119 goto error0;
2120 }
2121 if (XFS_IS_CORRUPT(mp, i != 1)) {
2122 xfs_btree_mark_sick(cur);
2123 error = -EFSCORRUPTED;
2124 goto error0;
2125 }
2126 /*
2127 * Get the offset in the inode chunk.
2128 */
2129 off = agino - rec.ir_startino;
2130 ASSERT(off >= 0 && off < XFS_INODES_PER_CHUNK);
2131 ASSERT(!(rec.ir_free & XFS_INOBT_MASK(off)));
2132 /*
2133 * Mark the inode free & increment the count.
2134 */
2135 rec.ir_free |= XFS_INOBT_MASK(off);
2136 rec.ir_freecount++;
2137
2138 /*
2139 * When an inode chunk is free, it becomes eligible for removal. Don't
2140 * remove the chunk if the block size is large enough for multiple inode
2141 * chunks (that might not be free).
2142 */
2143 if (!xfs_has_ikeep(mp) && rec.ir_free == XFS_INOBT_ALL_FREE &&
2144 mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK) {
2145 xic->deleted = true;
2146 xic->first_ino = xfs_agino_to_ino(pag, rec.ir_startino);
2147 xic->alloc = xfs_inobt_irec_to_allocmask(&rec);
2148
2149 /*
2150 * Remove the inode cluster from the AGI B+Tree, adjust the
2151 * AGI and Superblock inode counts, and mark the disk space
2152 * to be freed when the transaction is committed.
2153 */
2154 ilen = rec.ir_freecount;
2155 be32_add_cpu(&agi->agi_count, -ilen);
2156 be32_add_cpu(&agi->agi_freecount, -(ilen - 1));
2157 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_COUNT | XFS_AGI_FREECOUNT);
2158 pag->pagi_freecount -= ilen - 1;
2159 pag->pagi_count -= ilen;
2160 xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, -ilen);
2161 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -(ilen - 1));
2162
2163 if ((error = xfs_btree_delete(cur, &i))) {
2164 xfs_warn(mp, "%s: xfs_btree_delete returned error %d.",
2165 __func__, error);
2166 goto error0;
2167 }
2168
2169 error = xfs_difree_inode_chunk(tp, pag, &rec);
2170 if (error)
2171 goto error0;
2172 } else {
2173 xic->deleted = false;
2174
2175 error = xfs_inobt_update(cur, &rec);
2176 if (error) {
2177 xfs_warn(mp, "%s: xfs_inobt_update returned error %d.",
2178 __func__, error);
2179 goto error0;
2180 }
2181
2182 /*
2183 * Change the inode free counts and log the ag/sb changes.
2184 */
2185 be32_add_cpu(&agi->agi_freecount, 1);
2186 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
2187 pag->pagi_freecount++;
2188 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, 1);
2189 }
2190
2191 error = xfs_check_agi_freecount(cur);
2192 if (error)
2193 goto error0;
2194
2195 *orec = rec;
2196 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2197 return 0;
2198
2199error0:
2200 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2201 return error;
2202}
2203
2204/*
2205 * Free an inode in the free inode btree.
2206 */
2207STATIC int
2208xfs_difree_finobt(
2209 struct xfs_perag *pag,
2210 struct xfs_trans *tp,
2211 struct xfs_buf *agbp,
2212 xfs_agino_t agino,
2213 struct xfs_inobt_rec_incore *ibtrec) /* inobt record */
2214{
2215 struct xfs_mount *mp = pag_mount(pag);
2216 struct xfs_btree_cur *cur;
2217 struct xfs_inobt_rec_incore rec;
2218 int offset = agino - ibtrec->ir_startino;
2219 int error;
2220 int i;
2221
2222 cur = xfs_finobt_init_cursor(pag, tp, agbp);
2223
2224 error = xfs_inobt_lookup(cur, ibtrec->ir_startino, XFS_LOOKUP_EQ, &i);
2225 if (error)
2226 goto error;
2227 if (i == 0) {
2228 /*
2229 * If the record does not exist in the finobt, we must have just
2230 * freed an inode in a previously fully allocated chunk. If not,
2231 * something is out of sync.
2232 */
2233 if (XFS_IS_CORRUPT(mp, ibtrec->ir_freecount != 1)) {
2234 xfs_btree_mark_sick(cur);
2235 error = -EFSCORRUPTED;
2236 goto error;
2237 }
2238
2239 error = xfs_inobt_insert_rec(cur, ibtrec->ir_holemask,
2240 ibtrec->ir_count,
2241 ibtrec->ir_freecount,
2242 ibtrec->ir_free, &i);
2243 if (error)
2244 goto error;
2245 ASSERT(i == 1);
2246
2247 goto out;
2248 }
2249
2250 /*
2251 * Read and update the existing record. We could just copy the ibtrec
2252 * across here, but that would defeat the purpose of having redundant
2253 * metadata. By making the modifications independently, we can catch
2254 * corruptions that we wouldn't see if we just copied from one record
2255 * to another.
2256 */
2257 error = xfs_inobt_get_rec(cur, &rec, &i);
2258 if (error)
2259 goto error;
2260 if (XFS_IS_CORRUPT(mp, i != 1)) {
2261 xfs_btree_mark_sick(cur);
2262 error = -EFSCORRUPTED;
2263 goto error;
2264 }
2265
2266 rec.ir_free |= XFS_INOBT_MASK(offset);
2267 rec.ir_freecount++;
2268
2269 if (XFS_IS_CORRUPT(mp,
2270 rec.ir_free != ibtrec->ir_free ||
2271 rec.ir_freecount != ibtrec->ir_freecount)) {
2272 xfs_btree_mark_sick(cur);
2273 error = -EFSCORRUPTED;
2274 goto error;
2275 }
2276
2277 /*
2278 * The content of inobt records should always match between the inobt
2279 * and finobt. The lifecycle of records in the finobt is different from
2280 * the inobt in that the finobt only tracks records with at least one
2281 * free inode. Hence, if all of the inodes are free and we aren't
2282 * keeping inode chunks permanently on disk, remove the record.
2283 * Otherwise, update the record with the new information.
2284 *
2285 * Note that we currently can't free chunks when the block size is large
2286 * enough for multiple chunks. Leave the finobt record to remain in sync
2287 * with the inobt.
2288 */
2289 if (!xfs_has_ikeep(mp) && rec.ir_free == XFS_INOBT_ALL_FREE &&
2290 mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK) {
2291 error = xfs_btree_delete(cur, &i);
2292 if (error)
2293 goto error;
2294 ASSERT(i == 1);
2295 } else {
2296 error = xfs_inobt_update(cur, &rec);
2297 if (error)
2298 goto error;
2299 }
2300
2301out:
2302 error = xfs_check_agi_freecount(cur);
2303 if (error)
2304 goto error;
2305
2306 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2307 return 0;
2308
2309error:
2310 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2311 return error;
2312}
2313
2314/*
2315 * Free disk inode. Carefully avoids touching the incore inode, all
2316 * manipulations incore are the caller's responsibility.
2317 * The on-disk inode is not changed by this operation, only the
2318 * btree (free inode mask) is changed.
2319 */
2320int
2321xfs_difree(
2322 struct xfs_trans *tp,
2323 struct xfs_perag *pag,
2324 xfs_ino_t inode,
2325 struct xfs_icluster *xic)
2326{
2327 /* REFERENCED */
2328 xfs_agblock_t agbno; /* block number containing inode */
2329 struct xfs_buf *agbp; /* buffer for allocation group header */
2330 xfs_agino_t agino; /* allocation group inode number */
2331 int error; /* error return value */
2332 struct xfs_mount *mp = tp->t_mountp;
2333 struct xfs_inobt_rec_incore rec;/* btree record */
2334
2335 /*
2336 * Break up inode number into its components.
2337 */
2338 if (pag_agno(pag) != XFS_INO_TO_AGNO(mp, inode)) {
2339 xfs_warn(mp, "%s: agno != pag_agno(pag) (%d != %d).",
2340 __func__, XFS_INO_TO_AGNO(mp, inode), pag_agno(pag));
2341 ASSERT(0);
2342 return -EINVAL;
2343 }
2344 agino = XFS_INO_TO_AGINO(mp, inode);
2345 if (inode != xfs_agino_to_ino(pag, agino)) {
2346 xfs_warn(mp, "%s: inode != xfs_agino_to_ino() (%llu != %llu).",
2347 __func__, (unsigned long long)inode,
2348 (unsigned long long)xfs_agino_to_ino(pag, agino));
2349 ASSERT(0);
2350 return -EINVAL;
2351 }
2352 agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2353 if (agbno >= xfs_ag_block_count(mp, pag_agno(pag))) {
2354 xfs_warn(mp, "%s: agbno >= xfs_ag_block_count (%d >= %d).",
2355 __func__, agbno, xfs_ag_block_count(mp, pag_agno(pag)));
2356 ASSERT(0);
2357 return -EINVAL;
2358 }
2359 /*
2360 * Get the allocation group header.
2361 */
2362 error = xfs_ialloc_read_agi(pag, tp, 0, &agbp);
2363 if (error) {
2364 xfs_warn(mp, "%s: xfs_ialloc_read_agi() returned error %d.",
2365 __func__, error);
2366 return error;
2367 }
2368
2369 /*
2370 * Fix up the inode allocation btree.
2371 */
2372 error = xfs_difree_inobt(pag, tp, agbp, agino, xic, &rec);
2373 if (error)
2374 goto error0;
2375
2376 /*
2377 * Fix up the free inode btree.
2378 */
2379 if (xfs_has_finobt(mp)) {
2380 error = xfs_difree_finobt(pag, tp, agbp, agino, &rec);
2381 if (error)
2382 goto error0;
2383 }
2384
2385 return 0;
2386
2387error0:
2388 return error;
2389}
2390
2391STATIC int
2392xfs_imap_lookup(
2393 struct xfs_perag *pag,
2394 struct xfs_trans *tp,
2395 xfs_agino_t agino,
2396 xfs_agblock_t agbno,
2397 xfs_agblock_t *chunk_agbno,
2398 xfs_agblock_t *offset_agbno,
2399 int flags)
2400{
2401 struct xfs_mount *mp = pag_mount(pag);
2402 struct xfs_inobt_rec_incore rec;
2403 struct xfs_btree_cur *cur;
2404 struct xfs_buf *agbp;
2405 int error;
2406 int i;
2407
2408 error = xfs_ialloc_read_agi(pag, tp, 0, &agbp);
2409 if (error) {
2410 xfs_alert(mp,
2411 "%s: xfs_ialloc_read_agi() returned error %d, agno %d",
2412 __func__, error, pag_agno(pag));
2413 return error;
2414 }
2415
2416 /*
2417 * Lookup the inode record for the given agino. If the record cannot be
2418 * found, then it's an invalid inode number and we should abort. Once
2419 * we have a record, we need to ensure it contains the inode number
2420 * we are looking up.
2421 */
2422 cur = xfs_inobt_init_cursor(pag, tp, agbp);
2423 error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i);
2424 if (!error) {
2425 if (i)
2426 error = xfs_inobt_get_rec(cur, &rec, &i);
2427 if (!error && i == 0)
2428 error = -EINVAL;
2429 }
2430
2431 xfs_trans_brelse(tp, agbp);
2432 xfs_btree_del_cursor(cur, error);
2433 if (error)
2434 return error;
2435
2436 /* check that the returned record contains the required inode */
2437 if (rec.ir_startino > agino ||
2438 rec.ir_startino + M_IGEO(mp)->ialloc_inos <= agino)
2439 return -EINVAL;
2440
2441 /* for untrusted inodes check it is allocated first */
2442 if ((flags & XFS_IGET_UNTRUSTED) &&
2443 (rec.ir_free & XFS_INOBT_MASK(agino - rec.ir_startino)))
2444 return -EINVAL;
2445
2446 *chunk_agbno = XFS_AGINO_TO_AGBNO(mp, rec.ir_startino);
2447 *offset_agbno = agbno - *chunk_agbno;
2448 return 0;
2449}
2450
2451/*
2452 * Return the location of the inode in imap, for mapping it into a buffer.
2453 */
2454int
2455xfs_imap(
2456 struct xfs_perag *pag,
2457 struct xfs_trans *tp,
2458 xfs_ino_t ino, /* inode to locate */
2459 struct xfs_imap *imap, /* location map structure */
2460 uint flags) /* flags for inode btree lookup */
2461{
2462 struct xfs_mount *mp = pag_mount(pag);
2463 xfs_agblock_t agbno; /* block number of inode in the alloc group */
2464 xfs_agino_t agino; /* inode number within alloc group */
2465 xfs_agblock_t chunk_agbno; /* first block in inode chunk */
2466 xfs_agblock_t cluster_agbno; /* first block in inode cluster */
2467 int error; /* error code */
2468 int offset; /* index of inode in its buffer */
2469 xfs_agblock_t offset_agbno; /* blks from chunk start to inode */
2470
2471 ASSERT(ino != NULLFSINO);
2472
2473 /*
2474 * Split up the inode number into its parts.
2475 */
2476 agino = XFS_INO_TO_AGINO(mp, ino);
2477 agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2478 if (agbno >= xfs_ag_block_count(mp, pag_agno(pag)) ||
2479 ino != xfs_agino_to_ino(pag, agino)) {
2480 error = -EINVAL;
2481#ifdef DEBUG
2482 /*
2483 * Don't output diagnostic information for untrusted inodes
2484 * as they can be invalid without implying corruption.
2485 */
2486 if (flags & XFS_IGET_UNTRUSTED)
2487 return error;
2488 if (agbno >= xfs_ag_block_count(mp, pag_agno(pag))) {
2489 xfs_alert(mp,
2490 "%s: agbno (0x%llx) >= mp->m_sb.sb_agblocks (0x%lx)",
2491 __func__, (unsigned long long)agbno,
2492 (unsigned long)xfs_ag_block_count(mp,
2493 pag_agno(pag)));
2494 }
2495 if (ino != xfs_agino_to_ino(pag, agino)) {
2496 xfs_alert(mp,
2497 "%s: ino (0x%llx) != xfs_agino_to_ino() (0x%llx)",
2498 __func__, ino,
2499 xfs_agino_to_ino(pag, agino));
2500 }
2501 xfs_stack_trace();
2502#endif /* DEBUG */
2503 return error;
2504 }
2505
2506 /*
2507 * For bulkstat and handle lookups, we have an untrusted inode number
2508 * that we have to verify is valid. We cannot do this just by reading
2509 * the inode buffer as it may have been unlinked and removed leaving
2510 * inodes in stale state on disk. Hence we have to do a btree lookup
2511 * in all cases where an untrusted inode number is passed.
2512 */
2513 if (flags & XFS_IGET_UNTRUSTED) {
2514 error = xfs_imap_lookup(pag, tp, agino, agbno,
2515 &chunk_agbno, &offset_agbno, flags);
2516 if (error)
2517 return error;
2518 goto out_map;
2519 }
2520
2521 /*
2522 * If the inode cluster size is the same as the blocksize or
2523 * smaller we get to the buffer by simple arithmetics.
2524 */
2525 if (M_IGEO(mp)->blocks_per_cluster == 1) {
2526 offset = XFS_INO_TO_OFFSET(mp, ino);
2527 ASSERT(offset < mp->m_sb.sb_inopblock);
2528
2529 imap->im_blkno = xfs_agbno_to_daddr(pag, agbno);
2530 imap->im_len = XFS_FSB_TO_BB(mp, 1);
2531 imap->im_boffset = (unsigned short)(offset <<
2532 mp->m_sb.sb_inodelog);
2533 return 0;
2534 }
2535
2536 /*
2537 * If the inode chunks are aligned then use simple maths to
2538 * find the location. Otherwise we have to do a btree
2539 * lookup to find the location.
2540 */
2541 if (M_IGEO(mp)->inoalign_mask) {
2542 offset_agbno = agbno & M_IGEO(mp)->inoalign_mask;
2543 chunk_agbno = agbno - offset_agbno;
2544 } else {
2545 error = xfs_imap_lookup(pag, tp, agino, agbno,
2546 &chunk_agbno, &offset_agbno, flags);
2547 if (error)
2548 return error;
2549 }
2550
2551out_map:
2552 ASSERT(agbno >= chunk_agbno);
2553 cluster_agbno = chunk_agbno +
2554 ((offset_agbno / M_IGEO(mp)->blocks_per_cluster) *
2555 M_IGEO(mp)->blocks_per_cluster);
2556 offset = ((agbno - cluster_agbno) * mp->m_sb.sb_inopblock) +
2557 XFS_INO_TO_OFFSET(mp, ino);
2558
2559 imap->im_blkno = xfs_agbno_to_daddr(pag, cluster_agbno);
2560 imap->im_len = XFS_FSB_TO_BB(mp, M_IGEO(mp)->blocks_per_cluster);
2561 imap->im_boffset = (unsigned short)(offset << mp->m_sb.sb_inodelog);
2562
2563 /*
2564 * If the inode number maps to a block outside the bounds
2565 * of the file system then return NULL rather than calling
2566 * read_buf and panicing when we get an error from the
2567 * driver.
2568 */
2569 if ((imap->im_blkno + imap->im_len) >
2570 XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks)) {
2571 xfs_alert(mp,
2572 "%s: (im_blkno (0x%llx) + im_len (0x%llx)) > sb_dblocks (0x%llx)",
2573 __func__, (unsigned long long) imap->im_blkno,
2574 (unsigned long long) imap->im_len,
2575 XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks));
2576 return -EINVAL;
2577 }
2578 return 0;
2579}
2580
2581/*
2582 * Log specified fields for the ag hdr (inode section). The growth of the agi
2583 * structure over time requires that we interpret the buffer as two logical
2584 * regions delineated by the end of the unlinked list. This is due to the size
2585 * of the hash table and its location in the middle of the agi.
2586 *
2587 * For example, a request to log a field before agi_unlinked and a field after
2588 * agi_unlinked could cause us to log the entire hash table and use an excessive
2589 * amount of log space. To avoid this behavior, log the region up through
2590 * agi_unlinked in one call and the region after agi_unlinked through the end of
2591 * the structure in another.
2592 */
2593void
2594xfs_ialloc_log_agi(
2595 struct xfs_trans *tp,
2596 struct xfs_buf *bp,
2597 uint32_t fields)
2598{
2599 int first; /* first byte number */
2600 int last; /* last byte number */
2601 static const short offsets[] = { /* field starting offsets */
2602 /* keep in sync with bit definitions */
2603 offsetof(xfs_agi_t, agi_magicnum),
2604 offsetof(xfs_agi_t, agi_versionnum),
2605 offsetof(xfs_agi_t, agi_seqno),
2606 offsetof(xfs_agi_t, agi_length),
2607 offsetof(xfs_agi_t, agi_count),
2608 offsetof(xfs_agi_t, agi_root),
2609 offsetof(xfs_agi_t, agi_level),
2610 offsetof(xfs_agi_t, agi_freecount),
2611 offsetof(xfs_agi_t, agi_newino),
2612 offsetof(xfs_agi_t, agi_dirino),
2613 offsetof(xfs_agi_t, agi_unlinked),
2614 offsetof(xfs_agi_t, agi_free_root),
2615 offsetof(xfs_agi_t, agi_free_level),
2616 offsetof(xfs_agi_t, agi_iblocks),
2617 sizeof(xfs_agi_t)
2618 };
2619#ifdef DEBUG
2620 struct xfs_agi *agi = bp->b_addr;
2621
2622 ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
2623#endif
2624
2625 /*
2626 * Compute byte offsets for the first and last fields in the first
2627 * region and log the agi buffer. This only logs up through
2628 * agi_unlinked.
2629 */
2630 if (fields & XFS_AGI_ALL_BITS_R1) {
2631 xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R1,
2632 &first, &last);
2633 xfs_trans_log_buf(tp, bp, first, last);
2634 }
2635
2636 /*
2637 * Mask off the bits in the first region and calculate the first and
2638 * last field offsets for any bits in the second region.
2639 */
2640 fields &= ~XFS_AGI_ALL_BITS_R1;
2641 if (fields) {
2642 xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R2,
2643 &first, &last);
2644 xfs_trans_log_buf(tp, bp, first, last);
2645 }
2646}
2647
2648static xfs_failaddr_t
2649xfs_agi_verify(
2650 struct xfs_buf *bp)
2651{
2652 struct xfs_mount *mp = bp->b_mount;
2653 struct xfs_agi *agi = bp->b_addr;
2654 xfs_failaddr_t fa;
2655 uint32_t agi_seqno = be32_to_cpu(agi->agi_seqno);
2656 uint32_t agi_length = be32_to_cpu(agi->agi_length);
2657 int i;
2658
2659 if (xfs_has_crc(mp)) {
2660 if (!uuid_equal(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid))
2661 return __this_address;
2662 if (!xfs_log_check_lsn(mp, be64_to_cpu(agi->agi_lsn)))
2663 return __this_address;
2664 }
2665
2666 /*
2667 * Validate the magic number of the agi block.
2668 */
2669 if (!xfs_verify_magic(bp, agi->agi_magicnum))
2670 return __this_address;
2671 if (!XFS_AGI_GOOD_VERSION(be32_to_cpu(agi->agi_versionnum)))
2672 return __this_address;
2673
2674 fa = xfs_validate_ag_length(bp, agi_seqno, agi_length);
2675 if (fa)
2676 return fa;
2677
2678 if (be32_to_cpu(agi->agi_level) < 1 ||
2679 be32_to_cpu(agi->agi_level) > M_IGEO(mp)->inobt_maxlevels)
2680 return __this_address;
2681
2682 if (xfs_has_finobt(mp) &&
2683 (be32_to_cpu(agi->agi_free_level) < 1 ||
2684 be32_to_cpu(agi->agi_free_level) > M_IGEO(mp)->inobt_maxlevels))
2685 return __this_address;
2686
2687 for (i = 0; i < XFS_AGI_UNLINKED_BUCKETS; i++) {
2688 if (agi->agi_unlinked[i] == cpu_to_be32(NULLAGINO))
2689 continue;
2690 if (!xfs_verify_ino(mp, be32_to_cpu(agi->agi_unlinked[i])))
2691 return __this_address;
2692 }
2693
2694 return NULL;
2695}
2696
2697static void
2698xfs_agi_read_verify(
2699 struct xfs_buf *bp)
2700{
2701 struct xfs_mount *mp = bp->b_mount;
2702 xfs_failaddr_t fa;
2703
2704 if (xfs_has_crc(mp) &&
2705 !xfs_buf_verify_cksum(bp, XFS_AGI_CRC_OFF))
2706 xfs_verifier_error(bp, -EFSBADCRC, __this_address);
2707 else {
2708 fa = xfs_agi_verify(bp);
2709 if (XFS_TEST_ERROR(fa, mp, XFS_ERRTAG_IALLOC_READ_AGI))
2710 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2711 }
2712}
2713
2714static void
2715xfs_agi_write_verify(
2716 struct xfs_buf *bp)
2717{
2718 struct xfs_mount *mp = bp->b_mount;
2719 struct xfs_buf_log_item *bip = bp->b_log_item;
2720 struct xfs_agi *agi = bp->b_addr;
2721 xfs_failaddr_t fa;
2722
2723 fa = xfs_agi_verify(bp);
2724 if (fa) {
2725 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2726 return;
2727 }
2728
2729 if (!xfs_has_crc(mp))
2730 return;
2731
2732 if (bip)
2733 agi->agi_lsn = cpu_to_be64(bip->bli_item.li_lsn);
2734 xfs_buf_update_cksum(bp, XFS_AGI_CRC_OFF);
2735}
2736
2737const struct xfs_buf_ops xfs_agi_buf_ops = {
2738 .name = "xfs_agi",
2739 .magic = { cpu_to_be32(XFS_AGI_MAGIC), cpu_to_be32(XFS_AGI_MAGIC) },
2740 .verify_read = xfs_agi_read_verify,
2741 .verify_write = xfs_agi_write_verify,
2742 .verify_struct = xfs_agi_verify,
2743};
2744
2745/*
2746 * Read in the allocation group header (inode allocation section)
2747 */
2748int
2749xfs_read_agi(
2750 struct xfs_perag *pag,
2751 struct xfs_trans *tp,
2752 xfs_buf_flags_t flags,
2753 struct xfs_buf **agibpp)
2754{
2755 struct xfs_mount *mp = pag_mount(pag);
2756 int error;
2757
2758 trace_xfs_read_agi(pag);
2759
2760 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
2761 XFS_AG_DADDR(mp, pag_agno(pag), XFS_AGI_DADDR(mp)),
2762 XFS_FSS_TO_BB(mp, 1), flags, agibpp, &xfs_agi_buf_ops);
2763 if (xfs_metadata_is_sick(error))
2764 xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
2765 if (error)
2766 return error;
2767 if (tp)
2768 xfs_trans_buf_set_type(tp, *agibpp, XFS_BLFT_AGI_BUF);
2769
2770 xfs_buf_set_ref(*agibpp, XFS_AGI_REF);
2771 return 0;
2772}
2773
2774/*
2775 * Read in the agi and initialise the per-ag data. If the caller supplies a
2776 * @agibpp, return the locked AGI buffer to them, otherwise release it.
2777 */
2778int
2779xfs_ialloc_read_agi(
2780 struct xfs_perag *pag,
2781 struct xfs_trans *tp,
2782 int flags,
2783 struct xfs_buf **agibpp)
2784{
2785 struct xfs_buf *agibp;
2786 struct xfs_agi *agi;
2787 int error;
2788
2789 trace_xfs_ialloc_read_agi(pag);
2790
2791 error = xfs_read_agi(pag, tp,
2792 (flags & XFS_IALLOC_FLAG_TRYLOCK) ? XBF_TRYLOCK : 0,
2793 &agibp);
2794 if (error)
2795 return error;
2796
2797 agi = agibp->b_addr;
2798 if (!xfs_perag_initialised_agi(pag)) {
2799 pag->pagi_freecount = be32_to_cpu(agi->agi_freecount);
2800 pag->pagi_count = be32_to_cpu(agi->agi_count);
2801 set_bit(XFS_AGSTATE_AGI_INIT, &pag->pag_opstate);
2802 }
2803
2804 /*
2805 * It's possible for these to be out of sync if
2806 * we are in the middle of a forced shutdown.
2807 */
2808 ASSERT(pag->pagi_freecount == be32_to_cpu(agi->agi_freecount) ||
2809 xfs_is_shutdown(pag_mount(pag)));
2810 if (agibpp)
2811 *agibpp = agibp;
2812 else
2813 xfs_trans_brelse(tp, agibp);
2814 return 0;
2815}
2816
2817/* How many inodes are backed by inode clusters ondisk? */
2818STATIC int
2819xfs_ialloc_count_ondisk(
2820 struct xfs_btree_cur *cur,
2821 xfs_agino_t low,
2822 xfs_agino_t high,
2823 unsigned int *allocated)
2824{
2825 struct xfs_inobt_rec_incore irec;
2826 unsigned int ret = 0;
2827 int has_record;
2828 int error;
2829
2830 error = xfs_inobt_lookup(cur, low, XFS_LOOKUP_LE, &has_record);
2831 if (error)
2832 return error;
2833
2834 while (has_record) {
2835 unsigned int i, hole_idx;
2836
2837 error = xfs_inobt_get_rec(cur, &irec, &has_record);
2838 if (error)
2839 return error;
2840 if (irec.ir_startino > high)
2841 break;
2842
2843 for (i = 0; i < XFS_INODES_PER_CHUNK; i++) {
2844 if (irec.ir_startino + i < low)
2845 continue;
2846 if (irec.ir_startino + i > high)
2847 break;
2848
2849 hole_idx = i / XFS_INODES_PER_HOLEMASK_BIT;
2850 if (!(irec.ir_holemask & (1U << hole_idx)))
2851 ret++;
2852 }
2853
2854 error = xfs_btree_increment(cur, 0, &has_record);
2855 if (error)
2856 return error;
2857 }
2858
2859 *allocated = ret;
2860 return 0;
2861}
2862
2863/* Is there an inode record covering a given extent? */
2864int
2865xfs_ialloc_has_inodes_at_extent(
2866 struct xfs_btree_cur *cur,
2867 xfs_agblock_t bno,
2868 xfs_extlen_t len,
2869 enum xbtree_recpacking *outcome)
2870{
2871 xfs_agino_t agino;
2872 xfs_agino_t last_agino;
2873 unsigned int allocated;
2874 int error;
2875
2876 agino = XFS_AGB_TO_AGINO(cur->bc_mp, bno);
2877 last_agino = XFS_AGB_TO_AGINO(cur->bc_mp, bno + len) - 1;
2878
2879 error = xfs_ialloc_count_ondisk(cur, agino, last_agino, &allocated);
2880 if (error)
2881 return error;
2882
2883 if (allocated == 0)
2884 *outcome = XBTREE_RECPACKING_EMPTY;
2885 else if (allocated == last_agino - agino + 1)
2886 *outcome = XBTREE_RECPACKING_FULL;
2887 else
2888 *outcome = XBTREE_RECPACKING_SPARSE;
2889 return 0;
2890}
2891
2892struct xfs_ialloc_count_inodes {
2893 xfs_agino_t count;
2894 xfs_agino_t freecount;
2895};
2896
2897/* Record inode counts across all inobt records. */
2898STATIC int
2899xfs_ialloc_count_inodes_rec(
2900 struct xfs_btree_cur *cur,
2901 const union xfs_btree_rec *rec,
2902 void *priv)
2903{
2904 struct xfs_inobt_rec_incore irec;
2905 struct xfs_ialloc_count_inodes *ci = priv;
2906 xfs_failaddr_t fa;
2907
2908 xfs_inobt_btrec_to_irec(cur->bc_mp, rec, &irec);
2909 fa = xfs_inobt_check_irec(to_perag(cur->bc_group), &irec);
2910 if (fa)
2911 return xfs_inobt_complain_bad_rec(cur, fa, &irec);
2912
2913 ci->count += irec.ir_count;
2914 ci->freecount += irec.ir_freecount;
2915
2916 return 0;
2917}
2918
2919/* Count allocated and free inodes under an inobt. */
2920int
2921xfs_ialloc_count_inodes(
2922 struct xfs_btree_cur *cur,
2923 xfs_agino_t *count,
2924 xfs_agino_t *freecount)
2925{
2926 struct xfs_ialloc_count_inodes ci = {0};
2927 int error;
2928
2929 ASSERT(xfs_btree_is_ino(cur->bc_ops));
2930 error = xfs_btree_query_all(cur, xfs_ialloc_count_inodes_rec, &ci);
2931 if (error)
2932 return error;
2933
2934 *count = ci.count;
2935 *freecount = ci.freecount;
2936 return 0;
2937}
2938
2939/*
2940 * Initialize inode-related geometry information.
2941 *
2942 * Compute the inode btree min and max levels and set maxicount.
2943 *
2944 * Set the inode cluster size. This may still be overridden by the file
2945 * system block size if it is larger than the chosen cluster size.
2946 *
2947 * For v5 filesystems, scale the cluster size with the inode size to keep a
2948 * constant ratio of inode per cluster buffer, but only if mkfs has set the
2949 * inode alignment value appropriately for larger cluster sizes.
2950 *
2951 * Then compute the inode cluster alignment information.
2952 */
2953void
2954xfs_ialloc_setup_geometry(
2955 struct xfs_mount *mp)
2956{
2957 struct xfs_sb *sbp = &mp->m_sb;
2958 struct xfs_ino_geometry *igeo = M_IGEO(mp);
2959 uint64_t icount;
2960 uint inodes;
2961
2962 igeo->new_diflags2 = 0;
2963 if (xfs_has_bigtime(mp))
2964 igeo->new_diflags2 |= XFS_DIFLAG2_BIGTIME;
2965 if (xfs_has_large_extent_counts(mp))
2966 igeo->new_diflags2 |= XFS_DIFLAG2_NREXT64;
2967
2968 /* Compute inode btree geometry. */
2969 igeo->agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
2970 igeo->inobt_mxr[0] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, true);
2971 igeo->inobt_mxr[1] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, false);
2972 igeo->inobt_mnr[0] = igeo->inobt_mxr[0] / 2;
2973 igeo->inobt_mnr[1] = igeo->inobt_mxr[1] / 2;
2974
2975 igeo->ialloc_inos = max_t(uint16_t, XFS_INODES_PER_CHUNK,
2976 sbp->sb_inopblock);
2977 igeo->ialloc_blks = igeo->ialloc_inos >> sbp->sb_inopblog;
2978
2979 if (sbp->sb_spino_align)
2980 igeo->ialloc_min_blks = sbp->sb_spino_align;
2981 else
2982 igeo->ialloc_min_blks = igeo->ialloc_blks;
2983
2984 /* Compute and fill in value of m_ino_geo.inobt_maxlevels. */
2985 inodes = (1LL << XFS_INO_AGINO_BITS(mp)) >> XFS_INODES_PER_CHUNK_LOG;
2986 igeo->inobt_maxlevels = xfs_btree_compute_maxlevels(igeo->inobt_mnr,
2987 inodes);
2988 ASSERT(igeo->inobt_maxlevels <= xfs_iallocbt_maxlevels_ondisk());
2989
2990 /*
2991 * Set the maximum inode count for this filesystem, being careful not
2992 * to use obviously garbage sb_inopblog/sb_inopblock values. Regular
2993 * users should never get here due to failing sb verification, but
2994 * certain users (xfs_db) need to be usable even with corrupt metadata.
2995 */
2996 if (sbp->sb_imax_pct && igeo->ialloc_blks) {
2997 /*
2998 * Make sure the maximum inode count is a multiple
2999 * of the units we allocate inodes in.
3000 */
3001 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
3002 do_div(icount, 100);
3003 do_div(icount, igeo->ialloc_blks);
3004 igeo->maxicount = XFS_FSB_TO_INO(mp,
3005 icount * igeo->ialloc_blks);
3006 } else {
3007 igeo->maxicount = 0;
3008 }
3009
3010 /*
3011 * Compute the desired size of an inode cluster buffer size, which
3012 * starts at 8K and (on v5 filesystems) scales up with larger inode
3013 * sizes.
3014 *
3015 * Preserve the desired inode cluster size because the sparse inodes
3016 * feature uses that desired size (not the actual size) to compute the
3017 * sparse inode alignment. The mount code validates this value, so we
3018 * cannot change the behavior.
3019 */
3020 igeo->inode_cluster_size_raw = XFS_INODE_BIG_CLUSTER_SIZE;
3021 if (xfs_has_v3inodes(mp)) {
3022 int new_size = igeo->inode_cluster_size_raw;
3023
3024 new_size *= mp->m_sb.sb_inodesize / XFS_DINODE_MIN_SIZE;
3025 if (mp->m_sb.sb_inoalignmt >= XFS_B_TO_FSBT(mp, new_size))
3026 igeo->inode_cluster_size_raw = new_size;
3027 }
3028
3029 /* Calculate inode cluster ratios. */
3030 if (igeo->inode_cluster_size_raw > mp->m_sb.sb_blocksize)
3031 igeo->blocks_per_cluster = XFS_B_TO_FSBT(mp,
3032 igeo->inode_cluster_size_raw);
3033 else
3034 igeo->blocks_per_cluster = 1;
3035 igeo->inode_cluster_size = XFS_FSB_TO_B(mp, igeo->blocks_per_cluster);
3036 igeo->inodes_per_cluster = XFS_FSB_TO_INO(mp, igeo->blocks_per_cluster);
3037
3038 /* Calculate inode cluster alignment. */
3039 if (xfs_has_align(mp) &&
3040 mp->m_sb.sb_inoalignmt >= igeo->blocks_per_cluster)
3041 igeo->cluster_align = mp->m_sb.sb_inoalignmt;
3042 else
3043 igeo->cluster_align = 1;
3044 igeo->inoalign_mask = igeo->cluster_align - 1;
3045 igeo->cluster_align_inodes = XFS_FSB_TO_INO(mp, igeo->cluster_align);
3046
3047 /*
3048 * If we are using stripe alignment, check whether
3049 * the stripe unit is a multiple of the inode alignment
3050 */
3051 if (mp->m_dalign && igeo->inoalign_mask &&
3052 !(mp->m_dalign & igeo->inoalign_mask))
3053 igeo->ialloc_align = mp->m_dalign;
3054 else
3055 igeo->ialloc_align = 0;
3056
3057 if (mp->m_sb.sb_blocksize > PAGE_SIZE)
3058 igeo->min_folio_order = mp->m_sb.sb_blocklog - PAGE_SHIFT;
3059 else
3060 igeo->min_folio_order = 0;
3061}
3062
3063/* Compute the location of the root directory inode that is laid out by mkfs. */
3064xfs_ino_t
3065xfs_ialloc_calc_rootino(
3066 struct xfs_mount *mp,
3067 int sunit)
3068{
3069 struct xfs_ino_geometry *igeo = M_IGEO(mp);
3070 xfs_agblock_t first_bno;
3071
3072 /*
3073 * Pre-calculate the geometry of AG 0. We know what it looks like
3074 * because libxfs knows how to create allocation groups now.
3075 *
3076 * first_bno is the first block in which mkfs could possibly have
3077 * allocated the root directory inode, once we factor in the metadata
3078 * that mkfs formats before it. Namely, the four AG headers...
3079 */
3080 first_bno = howmany(4 * mp->m_sb.sb_sectsize, mp->m_sb.sb_blocksize);
3081
3082 /* ...the two free space btree roots... */
3083 first_bno += 2;
3084
3085 /* ...the inode btree root... */
3086 first_bno += 1;
3087
3088 /* ...the initial AGFL... */
3089 first_bno += xfs_alloc_min_freelist(mp, NULL);
3090
3091 /* ...the free inode btree root... */
3092 if (xfs_has_finobt(mp))
3093 first_bno++;
3094
3095 /* ...the reverse mapping btree root... */
3096 if (xfs_has_rmapbt(mp))
3097 first_bno++;
3098
3099 /* ...the reference count btree... */
3100 if (xfs_has_reflink(mp))
3101 first_bno++;
3102
3103 /*
3104 * ...and the log, if it is allocated in the first allocation group.
3105 *
3106 * This can happen with filesystems that only have a single
3107 * allocation group, or very odd geometries created by old mkfs
3108 * versions on very small filesystems.
3109 */
3110 if (xfs_ag_contains_log(mp, 0))
3111 first_bno += mp->m_sb.sb_logblocks;
3112
3113 /*
3114 * Now round first_bno up to whatever allocation alignment is given
3115 * by the filesystem or was passed in.
3116 */
3117 if (xfs_has_dalign(mp) && igeo->ialloc_align > 0)
3118 first_bno = roundup(first_bno, sunit);
3119 else if (xfs_has_align(mp) &&
3120 mp->m_sb.sb_inoalignmt > 1)
3121 first_bno = roundup(first_bno, mp->m_sb.sb_inoalignmt);
3122
3123 return XFS_AGINO_TO_INO(mp, 0, XFS_AGB_TO_AGINO(mp, first_bno));
3124}
3125
3126/*
3127 * Ensure there are not sparse inode clusters that cross the new EOAG.
3128 *
3129 * This is a no-op for non-spinode filesystems since clusters are always fully
3130 * allocated and checking the bnobt suffices. However, a spinode filesystem
3131 * could have a record where the upper inodes are free blocks. If those blocks
3132 * were removed from the filesystem, the inode record would extend beyond EOAG,
3133 * which will be flagged as corruption.
3134 */
3135int
3136xfs_ialloc_check_shrink(
3137 struct xfs_perag *pag,
3138 struct xfs_trans *tp,
3139 struct xfs_buf *agibp,
3140 xfs_agblock_t new_length)
3141{
3142 struct xfs_inobt_rec_incore rec;
3143 struct xfs_btree_cur *cur;
3144 xfs_agino_t agino;
3145 int has;
3146 int error;
3147
3148 if (!xfs_has_sparseinodes(pag_mount(pag)))
3149 return 0;
3150
3151 cur = xfs_inobt_init_cursor(pag, tp, agibp);
3152
3153 /* Look up the inobt record that would correspond to the new EOFS. */
3154 agino = XFS_AGB_TO_AGINO(pag_mount(pag), new_length);
3155 error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &has);
3156 if (error || !has)
3157 goto out;
3158
3159 error = xfs_inobt_get_rec(cur, &rec, &has);
3160 if (error)
3161 goto out;
3162
3163 if (!has) {
3164 xfs_ag_mark_sick(pag, XFS_SICK_AG_INOBT);
3165 error = -EFSCORRUPTED;
3166 goto out;
3167 }
3168
3169 /* If the record covers inodes that would be beyond EOFS, bail out. */
3170 if (rec.ir_startino + XFS_INODES_PER_CHUNK > agino) {
3171 error = -ENOSPC;
3172 goto out;
3173 }
3174out:
3175 xfs_btree_del_cursor(cur, error);
3176 return error;
3177}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2002,2005 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
8#include "xfs_shared.h"
9#include "xfs_format.h"
10#include "xfs_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_bit.h"
13#include "xfs_sb.h"
14#include "xfs_mount.h"
15#include "xfs_inode.h"
16#include "xfs_btree.h"
17#include "xfs_ialloc.h"
18#include "xfs_ialloc_btree.h"
19#include "xfs_alloc.h"
20#include "xfs_errortag.h"
21#include "xfs_error.h"
22#include "xfs_bmap.h"
23#include "xfs_trans.h"
24#include "xfs_buf_item.h"
25#include "xfs_icreate_item.h"
26#include "xfs_icache.h"
27#include "xfs_trace.h"
28#include "xfs_log.h"
29#include "xfs_rmap.h"
30
31/*
32 * Lookup a record by ino in the btree given by cur.
33 */
34int /* error */
35xfs_inobt_lookup(
36 struct xfs_btree_cur *cur, /* btree cursor */
37 xfs_agino_t ino, /* starting inode of chunk */
38 xfs_lookup_t dir, /* <=, >=, == */
39 int *stat) /* success/failure */
40{
41 cur->bc_rec.i.ir_startino = ino;
42 cur->bc_rec.i.ir_holemask = 0;
43 cur->bc_rec.i.ir_count = 0;
44 cur->bc_rec.i.ir_freecount = 0;
45 cur->bc_rec.i.ir_free = 0;
46 return xfs_btree_lookup(cur, dir, stat);
47}
48
49/*
50 * Update the record referred to by cur to the value given.
51 * This either works (return 0) or gets an EFSCORRUPTED error.
52 */
53STATIC int /* error */
54xfs_inobt_update(
55 struct xfs_btree_cur *cur, /* btree cursor */
56 xfs_inobt_rec_incore_t *irec) /* btree record */
57{
58 union xfs_btree_rec rec;
59
60 rec.inobt.ir_startino = cpu_to_be32(irec->ir_startino);
61 if (xfs_sb_version_hassparseinodes(&cur->bc_mp->m_sb)) {
62 rec.inobt.ir_u.sp.ir_holemask = cpu_to_be16(irec->ir_holemask);
63 rec.inobt.ir_u.sp.ir_count = irec->ir_count;
64 rec.inobt.ir_u.sp.ir_freecount = irec->ir_freecount;
65 } else {
66 /* ir_holemask/ir_count not supported on-disk */
67 rec.inobt.ir_u.f.ir_freecount = cpu_to_be32(irec->ir_freecount);
68 }
69 rec.inobt.ir_free = cpu_to_be64(irec->ir_free);
70 return xfs_btree_update(cur, &rec);
71}
72
73/* Convert on-disk btree record to incore inobt record. */
74void
75xfs_inobt_btrec_to_irec(
76 struct xfs_mount *mp,
77 union xfs_btree_rec *rec,
78 struct xfs_inobt_rec_incore *irec)
79{
80 irec->ir_startino = be32_to_cpu(rec->inobt.ir_startino);
81 if (xfs_sb_version_hassparseinodes(&mp->m_sb)) {
82 irec->ir_holemask = be16_to_cpu(rec->inobt.ir_u.sp.ir_holemask);
83 irec->ir_count = rec->inobt.ir_u.sp.ir_count;
84 irec->ir_freecount = rec->inobt.ir_u.sp.ir_freecount;
85 } else {
86 /*
87 * ir_holemask/ir_count not supported on-disk. Fill in hardcoded
88 * values for full inode chunks.
89 */
90 irec->ir_holemask = XFS_INOBT_HOLEMASK_FULL;
91 irec->ir_count = XFS_INODES_PER_CHUNK;
92 irec->ir_freecount =
93 be32_to_cpu(rec->inobt.ir_u.f.ir_freecount);
94 }
95 irec->ir_free = be64_to_cpu(rec->inobt.ir_free);
96}
97
98/*
99 * Get the data from the pointed-to record.
100 */
101int
102xfs_inobt_get_rec(
103 struct xfs_btree_cur *cur,
104 struct xfs_inobt_rec_incore *irec,
105 int *stat)
106{
107 struct xfs_mount *mp = cur->bc_mp;
108 xfs_agnumber_t agno = cur->bc_ag.agno;
109 union xfs_btree_rec *rec;
110 int error;
111 uint64_t realfree;
112
113 error = xfs_btree_get_rec(cur, &rec, stat);
114 if (error || *stat == 0)
115 return error;
116
117 xfs_inobt_btrec_to_irec(mp, rec, irec);
118
119 if (!xfs_verify_agino(mp, agno, irec->ir_startino))
120 goto out_bad_rec;
121 if (irec->ir_count < XFS_INODES_PER_HOLEMASK_BIT ||
122 irec->ir_count > XFS_INODES_PER_CHUNK)
123 goto out_bad_rec;
124 if (irec->ir_freecount > XFS_INODES_PER_CHUNK)
125 goto out_bad_rec;
126
127 /* if there are no holes, return the first available offset */
128 if (!xfs_inobt_issparse(irec->ir_holemask))
129 realfree = irec->ir_free;
130 else
131 realfree = irec->ir_free & xfs_inobt_irec_to_allocmask(irec);
132 if (hweight64(realfree) != irec->ir_freecount)
133 goto out_bad_rec;
134
135 return 0;
136
137out_bad_rec:
138 xfs_warn(mp,
139 "%s Inode BTree record corruption in AG %d detected!",
140 cur->bc_btnum == XFS_BTNUM_INO ? "Used" : "Free", agno);
141 xfs_warn(mp,
142"start inode 0x%x, count 0x%x, free 0x%x freemask 0x%llx, holemask 0x%x",
143 irec->ir_startino, irec->ir_count, irec->ir_freecount,
144 irec->ir_free, irec->ir_holemask);
145 return -EFSCORRUPTED;
146}
147
148/*
149 * Insert a single inobt record. Cursor must already point to desired location.
150 */
151int
152xfs_inobt_insert_rec(
153 struct xfs_btree_cur *cur,
154 uint16_t holemask,
155 uint8_t count,
156 int32_t freecount,
157 xfs_inofree_t free,
158 int *stat)
159{
160 cur->bc_rec.i.ir_holemask = holemask;
161 cur->bc_rec.i.ir_count = count;
162 cur->bc_rec.i.ir_freecount = freecount;
163 cur->bc_rec.i.ir_free = free;
164 return xfs_btree_insert(cur, stat);
165}
166
167/*
168 * Insert records describing a newly allocated inode chunk into the inobt.
169 */
170STATIC int
171xfs_inobt_insert(
172 struct xfs_mount *mp,
173 struct xfs_trans *tp,
174 struct xfs_buf *agbp,
175 xfs_agino_t newino,
176 xfs_agino_t newlen,
177 xfs_btnum_t btnum)
178{
179 struct xfs_btree_cur *cur;
180 struct xfs_agi *agi = agbp->b_addr;
181 xfs_agnumber_t agno = be32_to_cpu(agi->agi_seqno);
182 xfs_agino_t thisino;
183 int i;
184 int error;
185
186 cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, btnum);
187
188 for (thisino = newino;
189 thisino < newino + newlen;
190 thisino += XFS_INODES_PER_CHUNK) {
191 error = xfs_inobt_lookup(cur, thisino, XFS_LOOKUP_EQ, &i);
192 if (error) {
193 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
194 return error;
195 }
196 ASSERT(i == 0);
197
198 error = xfs_inobt_insert_rec(cur, XFS_INOBT_HOLEMASK_FULL,
199 XFS_INODES_PER_CHUNK,
200 XFS_INODES_PER_CHUNK,
201 XFS_INOBT_ALL_FREE, &i);
202 if (error) {
203 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
204 return error;
205 }
206 ASSERT(i == 1);
207 }
208
209 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
210
211 return 0;
212}
213
214/*
215 * Verify that the number of free inodes in the AGI is correct.
216 */
217#ifdef DEBUG
218STATIC int
219xfs_check_agi_freecount(
220 struct xfs_btree_cur *cur,
221 struct xfs_agi *agi)
222{
223 if (cur->bc_nlevels == 1) {
224 xfs_inobt_rec_incore_t rec;
225 int freecount = 0;
226 int error;
227 int i;
228
229 error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
230 if (error)
231 return error;
232
233 do {
234 error = xfs_inobt_get_rec(cur, &rec, &i);
235 if (error)
236 return error;
237
238 if (i) {
239 freecount += rec.ir_freecount;
240 error = xfs_btree_increment(cur, 0, &i);
241 if (error)
242 return error;
243 }
244 } while (i == 1);
245
246 if (!XFS_FORCED_SHUTDOWN(cur->bc_mp))
247 ASSERT(freecount == be32_to_cpu(agi->agi_freecount));
248 }
249 return 0;
250}
251#else
252#define xfs_check_agi_freecount(cur, agi) 0
253#endif
254
255/*
256 * Initialise a new set of inodes. When called without a transaction context
257 * (e.g. from recovery) we initiate a delayed write of the inode buffers rather
258 * than logging them (which in a transaction context puts them into the AIL
259 * for writeback rather than the xfsbufd queue).
260 */
261int
262xfs_ialloc_inode_init(
263 struct xfs_mount *mp,
264 struct xfs_trans *tp,
265 struct list_head *buffer_list,
266 int icount,
267 xfs_agnumber_t agno,
268 xfs_agblock_t agbno,
269 xfs_agblock_t length,
270 unsigned int gen)
271{
272 struct xfs_buf *fbuf;
273 struct xfs_dinode *free;
274 int nbufs;
275 int version;
276 int i, j;
277 xfs_daddr_t d;
278 xfs_ino_t ino = 0;
279 int error;
280
281 /*
282 * Loop over the new block(s), filling in the inodes. For small block
283 * sizes, manipulate the inodes in buffers which are multiples of the
284 * blocks size.
285 */
286 nbufs = length / M_IGEO(mp)->blocks_per_cluster;
287
288 /*
289 * Figure out what version number to use in the inodes we create. If
290 * the superblock version has caught up to the one that supports the new
291 * inode format, then use the new inode version. Otherwise use the old
292 * version so that old kernels will continue to be able to use the file
293 * system.
294 *
295 * For v3 inodes, we also need to write the inode number into the inode,
296 * so calculate the first inode number of the chunk here as
297 * XFS_AGB_TO_AGINO() only works within a filesystem block, not
298 * across multiple filesystem blocks (such as a cluster) and so cannot
299 * be used in the cluster buffer loop below.
300 *
301 * Further, because we are writing the inode directly into the buffer
302 * and calculating a CRC on the entire inode, we have ot log the entire
303 * inode so that the entire range the CRC covers is present in the log.
304 * That means for v3 inode we log the entire buffer rather than just the
305 * inode cores.
306 */
307 if (xfs_sb_version_has_v3inode(&mp->m_sb)) {
308 version = 3;
309 ino = XFS_AGINO_TO_INO(mp, agno, XFS_AGB_TO_AGINO(mp, agbno));
310
311 /*
312 * log the initialisation that is about to take place as an
313 * logical operation. This means the transaction does not
314 * need to log the physical changes to the inode buffers as log
315 * recovery will know what initialisation is actually needed.
316 * Hence we only need to log the buffers as "ordered" buffers so
317 * they track in the AIL as if they were physically logged.
318 */
319 if (tp)
320 xfs_icreate_log(tp, agno, agbno, icount,
321 mp->m_sb.sb_inodesize, length, gen);
322 } else
323 version = 2;
324
325 for (j = 0; j < nbufs; j++) {
326 /*
327 * Get the block.
328 */
329 d = XFS_AGB_TO_DADDR(mp, agno, agbno +
330 (j * M_IGEO(mp)->blocks_per_cluster));
331 error = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
332 mp->m_bsize * M_IGEO(mp)->blocks_per_cluster,
333 XBF_UNMAPPED, &fbuf);
334 if (error)
335 return error;
336
337 /* Initialize the inode buffers and log them appropriately. */
338 fbuf->b_ops = &xfs_inode_buf_ops;
339 xfs_buf_zero(fbuf, 0, BBTOB(fbuf->b_length));
340 for (i = 0; i < M_IGEO(mp)->inodes_per_cluster; i++) {
341 int ioffset = i << mp->m_sb.sb_inodelog;
342 uint isize = XFS_DINODE_SIZE(&mp->m_sb);
343
344 free = xfs_make_iptr(mp, fbuf, i);
345 free->di_magic = cpu_to_be16(XFS_DINODE_MAGIC);
346 free->di_version = version;
347 free->di_gen = cpu_to_be32(gen);
348 free->di_next_unlinked = cpu_to_be32(NULLAGINO);
349
350 if (version == 3) {
351 free->di_ino = cpu_to_be64(ino);
352 ino++;
353 uuid_copy(&free->di_uuid,
354 &mp->m_sb.sb_meta_uuid);
355 xfs_dinode_calc_crc(mp, free);
356 } else if (tp) {
357 /* just log the inode core */
358 xfs_trans_log_buf(tp, fbuf, ioffset,
359 ioffset + isize - 1);
360 }
361 }
362
363 if (tp) {
364 /*
365 * Mark the buffer as an inode allocation buffer so it
366 * sticks in AIL at the point of this allocation
367 * transaction. This ensures the they are on disk before
368 * the tail of the log can be moved past this
369 * transaction (i.e. by preventing relogging from moving
370 * it forward in the log).
371 */
372 xfs_trans_inode_alloc_buf(tp, fbuf);
373 if (version == 3) {
374 /*
375 * Mark the buffer as ordered so that they are
376 * not physically logged in the transaction but
377 * still tracked in the AIL as part of the
378 * transaction and pin the log appropriately.
379 */
380 xfs_trans_ordered_buf(tp, fbuf);
381 }
382 } else {
383 fbuf->b_flags |= XBF_DONE;
384 xfs_buf_delwri_queue(fbuf, buffer_list);
385 xfs_buf_relse(fbuf);
386 }
387 }
388 return 0;
389}
390
391/*
392 * Align startino and allocmask for a recently allocated sparse chunk such that
393 * they are fit for insertion (or merge) into the on-disk inode btrees.
394 *
395 * Background:
396 *
397 * When enabled, sparse inode support increases the inode alignment from cluster
398 * size to inode chunk size. This means that the minimum range between two
399 * non-adjacent inode records in the inobt is large enough for a full inode
400 * record. This allows for cluster sized, cluster aligned block allocation
401 * without need to worry about whether the resulting inode record overlaps with
402 * another record in the tree. Without this basic rule, we would have to deal
403 * with the consequences of overlap by potentially undoing recent allocations in
404 * the inode allocation codepath.
405 *
406 * Because of this alignment rule (which is enforced on mount), there are two
407 * inobt possibilities for newly allocated sparse chunks. One is that the
408 * aligned inode record for the chunk covers a range of inodes not already
409 * covered in the inobt (i.e., it is safe to insert a new sparse record). The
410 * other is that a record already exists at the aligned startino that considers
411 * the newly allocated range as sparse. In the latter case, record content is
412 * merged in hope that sparse inode chunks fill to full chunks over time.
413 */
414STATIC void
415xfs_align_sparse_ino(
416 struct xfs_mount *mp,
417 xfs_agino_t *startino,
418 uint16_t *allocmask)
419{
420 xfs_agblock_t agbno;
421 xfs_agblock_t mod;
422 int offset;
423
424 agbno = XFS_AGINO_TO_AGBNO(mp, *startino);
425 mod = agbno % mp->m_sb.sb_inoalignmt;
426 if (!mod)
427 return;
428
429 /* calculate the inode offset and align startino */
430 offset = XFS_AGB_TO_AGINO(mp, mod);
431 *startino -= offset;
432
433 /*
434 * Since startino has been aligned down, left shift allocmask such that
435 * it continues to represent the same physical inodes relative to the
436 * new startino.
437 */
438 *allocmask <<= offset / XFS_INODES_PER_HOLEMASK_BIT;
439}
440
441/*
442 * Determine whether the source inode record can merge into the target. Both
443 * records must be sparse, the inode ranges must match and there must be no
444 * allocation overlap between the records.
445 */
446STATIC bool
447__xfs_inobt_can_merge(
448 struct xfs_inobt_rec_incore *trec, /* tgt record */
449 struct xfs_inobt_rec_incore *srec) /* src record */
450{
451 uint64_t talloc;
452 uint64_t salloc;
453
454 /* records must cover the same inode range */
455 if (trec->ir_startino != srec->ir_startino)
456 return false;
457
458 /* both records must be sparse */
459 if (!xfs_inobt_issparse(trec->ir_holemask) ||
460 !xfs_inobt_issparse(srec->ir_holemask))
461 return false;
462
463 /* both records must track some inodes */
464 if (!trec->ir_count || !srec->ir_count)
465 return false;
466
467 /* can't exceed capacity of a full record */
468 if (trec->ir_count + srec->ir_count > XFS_INODES_PER_CHUNK)
469 return false;
470
471 /* verify there is no allocation overlap */
472 talloc = xfs_inobt_irec_to_allocmask(trec);
473 salloc = xfs_inobt_irec_to_allocmask(srec);
474 if (talloc & salloc)
475 return false;
476
477 return true;
478}
479
480/*
481 * Merge the source inode record into the target. The caller must call
482 * __xfs_inobt_can_merge() to ensure the merge is valid.
483 */
484STATIC void
485__xfs_inobt_rec_merge(
486 struct xfs_inobt_rec_incore *trec, /* target */
487 struct xfs_inobt_rec_incore *srec) /* src */
488{
489 ASSERT(trec->ir_startino == srec->ir_startino);
490
491 /* combine the counts */
492 trec->ir_count += srec->ir_count;
493 trec->ir_freecount += srec->ir_freecount;
494
495 /*
496 * Merge the holemask and free mask. For both fields, 0 bits refer to
497 * allocated inodes. We combine the allocated ranges with bitwise AND.
498 */
499 trec->ir_holemask &= srec->ir_holemask;
500 trec->ir_free &= srec->ir_free;
501}
502
503/*
504 * Insert a new sparse inode chunk into the associated inode btree. The inode
505 * record for the sparse chunk is pre-aligned to a startino that should match
506 * any pre-existing sparse inode record in the tree. This allows sparse chunks
507 * to fill over time.
508 *
509 * This function supports two modes of handling preexisting records depending on
510 * the merge flag. If merge is true, the provided record is merged with the
511 * existing record and updated in place. The merged record is returned in nrec.
512 * If merge is false, an existing record is replaced with the provided record.
513 * If no preexisting record exists, the provided record is always inserted.
514 *
515 * It is considered corruption if a merge is requested and not possible. Given
516 * the sparse inode alignment constraints, this should never happen.
517 */
518STATIC int
519xfs_inobt_insert_sprec(
520 struct xfs_mount *mp,
521 struct xfs_trans *tp,
522 struct xfs_buf *agbp,
523 int btnum,
524 struct xfs_inobt_rec_incore *nrec, /* in/out: new/merged rec. */
525 bool merge) /* merge or replace */
526{
527 struct xfs_btree_cur *cur;
528 struct xfs_agi *agi = agbp->b_addr;
529 xfs_agnumber_t agno = be32_to_cpu(agi->agi_seqno);
530 int error;
531 int i;
532 struct xfs_inobt_rec_incore rec;
533
534 cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, btnum);
535
536 /* the new record is pre-aligned so we know where to look */
537 error = xfs_inobt_lookup(cur, nrec->ir_startino, XFS_LOOKUP_EQ, &i);
538 if (error)
539 goto error;
540 /* if nothing there, insert a new record and return */
541 if (i == 0) {
542 error = xfs_inobt_insert_rec(cur, nrec->ir_holemask,
543 nrec->ir_count, nrec->ir_freecount,
544 nrec->ir_free, &i);
545 if (error)
546 goto error;
547 if (XFS_IS_CORRUPT(mp, i != 1)) {
548 error = -EFSCORRUPTED;
549 goto error;
550 }
551
552 goto out;
553 }
554
555 /*
556 * A record exists at this startino. Merge or replace the record
557 * depending on what we've been asked to do.
558 */
559 if (merge) {
560 error = xfs_inobt_get_rec(cur, &rec, &i);
561 if (error)
562 goto error;
563 if (XFS_IS_CORRUPT(mp, i != 1)) {
564 error = -EFSCORRUPTED;
565 goto error;
566 }
567 if (XFS_IS_CORRUPT(mp, rec.ir_startino != nrec->ir_startino)) {
568 error = -EFSCORRUPTED;
569 goto error;
570 }
571
572 /*
573 * This should never fail. If we have coexisting records that
574 * cannot merge, something is seriously wrong.
575 */
576 if (XFS_IS_CORRUPT(mp, !__xfs_inobt_can_merge(nrec, &rec))) {
577 error = -EFSCORRUPTED;
578 goto error;
579 }
580
581 trace_xfs_irec_merge_pre(mp, agno, rec.ir_startino,
582 rec.ir_holemask, nrec->ir_startino,
583 nrec->ir_holemask);
584
585 /* merge to nrec to output the updated record */
586 __xfs_inobt_rec_merge(nrec, &rec);
587
588 trace_xfs_irec_merge_post(mp, agno, nrec->ir_startino,
589 nrec->ir_holemask);
590
591 error = xfs_inobt_rec_check_count(mp, nrec);
592 if (error)
593 goto error;
594 }
595
596 error = xfs_inobt_update(cur, nrec);
597 if (error)
598 goto error;
599
600out:
601 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
602 return 0;
603error:
604 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
605 return error;
606}
607
608/*
609 * Allocate new inodes in the allocation group specified by agbp.
610 * Return 0 for success, else error code.
611 */
612STATIC int
613xfs_ialloc_ag_alloc(
614 struct xfs_trans *tp,
615 struct xfs_buf *agbp,
616 int *alloc)
617{
618 struct xfs_agi *agi;
619 struct xfs_alloc_arg args;
620 xfs_agnumber_t agno;
621 int error;
622 xfs_agino_t newino; /* new first inode's number */
623 xfs_agino_t newlen; /* new number of inodes */
624 int isaligned = 0; /* inode allocation at stripe */
625 /* unit boundary */
626 /* init. to full chunk */
627 uint16_t allocmask = (uint16_t) -1;
628 struct xfs_inobt_rec_incore rec;
629 struct xfs_perag *pag;
630 struct xfs_ino_geometry *igeo = M_IGEO(tp->t_mountp);
631 int do_sparse = 0;
632
633 memset(&args, 0, sizeof(args));
634 args.tp = tp;
635 args.mp = tp->t_mountp;
636 args.fsbno = NULLFSBLOCK;
637 args.oinfo = XFS_RMAP_OINFO_INODES;
638
639#ifdef DEBUG
640 /* randomly do sparse inode allocations */
641 if (xfs_sb_version_hassparseinodes(&tp->t_mountp->m_sb) &&
642 igeo->ialloc_min_blks < igeo->ialloc_blks)
643 do_sparse = prandom_u32() & 1;
644#endif
645
646 /*
647 * Locking will ensure that we don't have two callers in here
648 * at one time.
649 */
650 newlen = igeo->ialloc_inos;
651 if (igeo->maxicount &&
652 percpu_counter_read_positive(&args.mp->m_icount) + newlen >
653 igeo->maxicount)
654 return -ENOSPC;
655 args.minlen = args.maxlen = igeo->ialloc_blks;
656 /*
657 * First try to allocate inodes contiguous with the last-allocated
658 * chunk of inodes. If the filesystem is striped, this will fill
659 * an entire stripe unit with inodes.
660 */
661 agi = agbp->b_addr;
662 newino = be32_to_cpu(agi->agi_newino);
663 agno = be32_to_cpu(agi->agi_seqno);
664 args.agbno = XFS_AGINO_TO_AGBNO(args.mp, newino) +
665 igeo->ialloc_blks;
666 if (do_sparse)
667 goto sparse_alloc;
668 if (likely(newino != NULLAGINO &&
669 (args.agbno < be32_to_cpu(agi->agi_length)))) {
670 args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
671 args.type = XFS_ALLOCTYPE_THIS_BNO;
672 args.prod = 1;
673
674 /*
675 * We need to take into account alignment here to ensure that
676 * we don't modify the free list if we fail to have an exact
677 * block. If we don't have an exact match, and every oher
678 * attempt allocation attempt fails, we'll end up cancelling
679 * a dirty transaction and shutting down.
680 *
681 * For an exact allocation, alignment must be 1,
682 * however we need to take cluster alignment into account when
683 * fixing up the freelist. Use the minalignslop field to
684 * indicate that extra blocks might be required for alignment,
685 * but not to use them in the actual exact allocation.
686 */
687 args.alignment = 1;
688 args.minalignslop = igeo->cluster_align - 1;
689
690 /* Allow space for the inode btree to split. */
691 args.minleft = igeo->inobt_maxlevels;
692 if ((error = xfs_alloc_vextent(&args)))
693 return error;
694
695 /*
696 * This request might have dirtied the transaction if the AG can
697 * satisfy the request, but the exact block was not available.
698 * If the allocation did fail, subsequent requests will relax
699 * the exact agbno requirement and increase the alignment
700 * instead. It is critical that the total size of the request
701 * (len + alignment + slop) does not increase from this point
702 * on, so reset minalignslop to ensure it is not included in
703 * subsequent requests.
704 */
705 args.minalignslop = 0;
706 }
707
708 if (unlikely(args.fsbno == NULLFSBLOCK)) {
709 /*
710 * Set the alignment for the allocation.
711 * If stripe alignment is turned on then align at stripe unit
712 * boundary.
713 * If the cluster size is smaller than a filesystem block
714 * then we're doing I/O for inodes in filesystem block size
715 * pieces, so don't need alignment anyway.
716 */
717 isaligned = 0;
718 if (igeo->ialloc_align) {
719 ASSERT(!(args.mp->m_flags & XFS_MOUNT_NOALIGN));
720 args.alignment = args.mp->m_dalign;
721 isaligned = 1;
722 } else
723 args.alignment = igeo->cluster_align;
724 /*
725 * Need to figure out where to allocate the inode blocks.
726 * Ideally they should be spaced out through the a.g.
727 * For now, just allocate blocks up front.
728 */
729 args.agbno = be32_to_cpu(agi->agi_root);
730 args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
731 /*
732 * Allocate a fixed-size extent of inodes.
733 */
734 args.type = XFS_ALLOCTYPE_NEAR_BNO;
735 args.prod = 1;
736 /*
737 * Allow space for the inode btree to split.
738 */
739 args.minleft = igeo->inobt_maxlevels;
740 if ((error = xfs_alloc_vextent(&args)))
741 return error;
742 }
743
744 /*
745 * If stripe alignment is turned on, then try again with cluster
746 * alignment.
747 */
748 if (isaligned && args.fsbno == NULLFSBLOCK) {
749 args.type = XFS_ALLOCTYPE_NEAR_BNO;
750 args.agbno = be32_to_cpu(agi->agi_root);
751 args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
752 args.alignment = igeo->cluster_align;
753 if ((error = xfs_alloc_vextent(&args)))
754 return error;
755 }
756
757 /*
758 * Finally, try a sparse allocation if the filesystem supports it and
759 * the sparse allocation length is smaller than a full chunk.
760 */
761 if (xfs_sb_version_hassparseinodes(&args.mp->m_sb) &&
762 igeo->ialloc_min_blks < igeo->ialloc_blks &&
763 args.fsbno == NULLFSBLOCK) {
764sparse_alloc:
765 args.type = XFS_ALLOCTYPE_NEAR_BNO;
766 args.agbno = be32_to_cpu(agi->agi_root);
767 args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
768 args.alignment = args.mp->m_sb.sb_spino_align;
769 args.prod = 1;
770
771 args.minlen = igeo->ialloc_min_blks;
772 args.maxlen = args.minlen;
773
774 /*
775 * The inode record will be aligned to full chunk size. We must
776 * prevent sparse allocation from AG boundaries that result in
777 * invalid inode records, such as records that start at agbno 0
778 * or extend beyond the AG.
779 *
780 * Set min agbno to the first aligned, non-zero agbno and max to
781 * the last aligned agbno that is at least one full chunk from
782 * the end of the AG.
783 */
784 args.min_agbno = args.mp->m_sb.sb_inoalignmt;
785 args.max_agbno = round_down(args.mp->m_sb.sb_agblocks,
786 args.mp->m_sb.sb_inoalignmt) -
787 igeo->ialloc_blks;
788
789 error = xfs_alloc_vextent(&args);
790 if (error)
791 return error;
792
793 newlen = XFS_AGB_TO_AGINO(args.mp, args.len);
794 ASSERT(newlen <= XFS_INODES_PER_CHUNK);
795 allocmask = (1 << (newlen / XFS_INODES_PER_HOLEMASK_BIT)) - 1;
796 }
797
798 if (args.fsbno == NULLFSBLOCK) {
799 *alloc = 0;
800 return 0;
801 }
802 ASSERT(args.len == args.minlen);
803
804 /*
805 * Stamp and write the inode buffers.
806 *
807 * Seed the new inode cluster with a random generation number. This
808 * prevents short-term reuse of generation numbers if a chunk is
809 * freed and then immediately reallocated. We use random numbers
810 * rather than a linear progression to prevent the next generation
811 * number from being easily guessable.
812 */
813 error = xfs_ialloc_inode_init(args.mp, tp, NULL, newlen, agno,
814 args.agbno, args.len, prandom_u32());
815
816 if (error)
817 return error;
818 /*
819 * Convert the results.
820 */
821 newino = XFS_AGB_TO_AGINO(args.mp, args.agbno);
822
823 if (xfs_inobt_issparse(~allocmask)) {
824 /*
825 * We've allocated a sparse chunk. Align the startino and mask.
826 */
827 xfs_align_sparse_ino(args.mp, &newino, &allocmask);
828
829 rec.ir_startino = newino;
830 rec.ir_holemask = ~allocmask;
831 rec.ir_count = newlen;
832 rec.ir_freecount = newlen;
833 rec.ir_free = XFS_INOBT_ALL_FREE;
834
835 /*
836 * Insert the sparse record into the inobt and allow for a merge
837 * if necessary. If a merge does occur, rec is updated to the
838 * merged record.
839 */
840 error = xfs_inobt_insert_sprec(args.mp, tp, agbp, XFS_BTNUM_INO,
841 &rec, true);
842 if (error == -EFSCORRUPTED) {
843 xfs_alert(args.mp,
844 "invalid sparse inode record: ino 0x%llx holemask 0x%x count %u",
845 XFS_AGINO_TO_INO(args.mp, agno,
846 rec.ir_startino),
847 rec.ir_holemask, rec.ir_count);
848 xfs_force_shutdown(args.mp, SHUTDOWN_CORRUPT_INCORE);
849 }
850 if (error)
851 return error;
852
853 /*
854 * We can't merge the part we've just allocated as for the inobt
855 * due to finobt semantics. The original record may or may not
856 * exist independent of whether physical inodes exist in this
857 * sparse chunk.
858 *
859 * We must update the finobt record based on the inobt record.
860 * rec contains the fully merged and up to date inobt record
861 * from the previous call. Set merge false to replace any
862 * existing record with this one.
863 */
864 if (xfs_sb_version_hasfinobt(&args.mp->m_sb)) {
865 error = xfs_inobt_insert_sprec(args.mp, tp, agbp,
866 XFS_BTNUM_FINO, &rec,
867 false);
868 if (error)
869 return error;
870 }
871 } else {
872 /* full chunk - insert new records to both btrees */
873 error = xfs_inobt_insert(args.mp, tp, agbp, newino, newlen,
874 XFS_BTNUM_INO);
875 if (error)
876 return error;
877
878 if (xfs_sb_version_hasfinobt(&args.mp->m_sb)) {
879 error = xfs_inobt_insert(args.mp, tp, agbp, newino,
880 newlen, XFS_BTNUM_FINO);
881 if (error)
882 return error;
883 }
884 }
885
886 /*
887 * Update AGI counts and newino.
888 */
889 be32_add_cpu(&agi->agi_count, newlen);
890 be32_add_cpu(&agi->agi_freecount, newlen);
891 pag = agbp->b_pag;
892 pag->pagi_freecount += newlen;
893 pag->pagi_count += newlen;
894 agi->agi_newino = cpu_to_be32(newino);
895
896 /*
897 * Log allocation group header fields
898 */
899 xfs_ialloc_log_agi(tp, agbp,
900 XFS_AGI_COUNT | XFS_AGI_FREECOUNT | XFS_AGI_NEWINO);
901 /*
902 * Modify/log superblock values for inode count and inode free count.
903 */
904 xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, (long)newlen);
905 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, (long)newlen);
906 *alloc = 1;
907 return 0;
908}
909
910STATIC xfs_agnumber_t
911xfs_ialloc_next_ag(
912 xfs_mount_t *mp)
913{
914 xfs_agnumber_t agno;
915
916 spin_lock(&mp->m_agirotor_lock);
917 agno = mp->m_agirotor;
918 if (++mp->m_agirotor >= mp->m_maxagi)
919 mp->m_agirotor = 0;
920 spin_unlock(&mp->m_agirotor_lock);
921
922 return agno;
923}
924
925/*
926 * Select an allocation group to look for a free inode in, based on the parent
927 * inode and the mode. Return the allocation group buffer.
928 */
929STATIC xfs_agnumber_t
930xfs_ialloc_ag_select(
931 xfs_trans_t *tp, /* transaction pointer */
932 xfs_ino_t parent, /* parent directory inode number */
933 umode_t mode) /* bits set to indicate file type */
934{
935 xfs_agnumber_t agcount; /* number of ag's in the filesystem */
936 xfs_agnumber_t agno; /* current ag number */
937 int flags; /* alloc buffer locking flags */
938 xfs_extlen_t ineed; /* blocks needed for inode allocation */
939 xfs_extlen_t longest = 0; /* longest extent available */
940 xfs_mount_t *mp; /* mount point structure */
941 int needspace; /* file mode implies space allocated */
942 xfs_perag_t *pag; /* per allocation group data */
943 xfs_agnumber_t pagno; /* parent (starting) ag number */
944 int error;
945
946 /*
947 * Files of these types need at least one block if length > 0
948 * (and they won't fit in the inode, but that's hard to figure out).
949 */
950 needspace = S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode);
951 mp = tp->t_mountp;
952 agcount = mp->m_maxagi;
953 if (S_ISDIR(mode))
954 pagno = xfs_ialloc_next_ag(mp);
955 else {
956 pagno = XFS_INO_TO_AGNO(mp, parent);
957 if (pagno >= agcount)
958 pagno = 0;
959 }
960
961 ASSERT(pagno < agcount);
962
963 /*
964 * Loop through allocation groups, looking for one with a little
965 * free space in it. Note we don't look for free inodes, exactly.
966 * Instead, we include whether there is a need to allocate inodes
967 * to mean that blocks must be allocated for them,
968 * if none are currently free.
969 */
970 agno = pagno;
971 flags = XFS_ALLOC_FLAG_TRYLOCK;
972 for (;;) {
973 pag = xfs_perag_get(mp, agno);
974 if (!pag->pagi_inodeok) {
975 xfs_ialloc_next_ag(mp);
976 goto nextag;
977 }
978
979 if (!pag->pagi_init) {
980 error = xfs_ialloc_pagi_init(mp, tp, agno);
981 if (error)
982 goto nextag;
983 }
984
985 if (pag->pagi_freecount) {
986 xfs_perag_put(pag);
987 return agno;
988 }
989
990 if (!pag->pagf_init) {
991 error = xfs_alloc_pagf_init(mp, tp, agno, flags);
992 if (error)
993 goto nextag;
994 }
995
996 /*
997 * Check that there is enough free space for the file plus a
998 * chunk of inodes if we need to allocate some. If this is the
999 * first pass across the AGs, take into account the potential
1000 * space needed for alignment of inode chunks when checking the
1001 * longest contiguous free space in the AG - this prevents us
1002 * from getting ENOSPC because we have free space larger than
1003 * ialloc_blks but alignment constraints prevent us from using
1004 * it.
1005 *
1006 * If we can't find an AG with space for full alignment slack to
1007 * be taken into account, we must be near ENOSPC in all AGs.
1008 * Hence we don't include alignment for the second pass and so
1009 * if we fail allocation due to alignment issues then it is most
1010 * likely a real ENOSPC condition.
1011 */
1012 ineed = M_IGEO(mp)->ialloc_min_blks;
1013 if (flags && ineed > 1)
1014 ineed += M_IGEO(mp)->cluster_align;
1015 longest = pag->pagf_longest;
1016 if (!longest)
1017 longest = pag->pagf_flcount > 0;
1018
1019 if (pag->pagf_freeblks >= needspace + ineed &&
1020 longest >= ineed) {
1021 xfs_perag_put(pag);
1022 return agno;
1023 }
1024nextag:
1025 xfs_perag_put(pag);
1026 /*
1027 * No point in iterating over the rest, if we're shutting
1028 * down.
1029 */
1030 if (XFS_FORCED_SHUTDOWN(mp))
1031 return NULLAGNUMBER;
1032 agno++;
1033 if (agno >= agcount)
1034 agno = 0;
1035 if (agno == pagno) {
1036 if (flags == 0)
1037 return NULLAGNUMBER;
1038 flags = 0;
1039 }
1040 }
1041}
1042
1043/*
1044 * Try to retrieve the next record to the left/right from the current one.
1045 */
1046STATIC int
1047xfs_ialloc_next_rec(
1048 struct xfs_btree_cur *cur,
1049 xfs_inobt_rec_incore_t *rec,
1050 int *done,
1051 int left)
1052{
1053 int error;
1054 int i;
1055
1056 if (left)
1057 error = xfs_btree_decrement(cur, 0, &i);
1058 else
1059 error = xfs_btree_increment(cur, 0, &i);
1060
1061 if (error)
1062 return error;
1063 *done = !i;
1064 if (i) {
1065 error = xfs_inobt_get_rec(cur, rec, &i);
1066 if (error)
1067 return error;
1068 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1069 return -EFSCORRUPTED;
1070 }
1071
1072 return 0;
1073}
1074
1075STATIC int
1076xfs_ialloc_get_rec(
1077 struct xfs_btree_cur *cur,
1078 xfs_agino_t agino,
1079 xfs_inobt_rec_incore_t *rec,
1080 int *done)
1081{
1082 int error;
1083 int i;
1084
1085 error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_EQ, &i);
1086 if (error)
1087 return error;
1088 *done = !i;
1089 if (i) {
1090 error = xfs_inobt_get_rec(cur, rec, &i);
1091 if (error)
1092 return error;
1093 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1094 return -EFSCORRUPTED;
1095 }
1096
1097 return 0;
1098}
1099
1100/*
1101 * Return the offset of the first free inode in the record. If the inode chunk
1102 * is sparsely allocated, we convert the record holemask to inode granularity
1103 * and mask off the unallocated regions from the inode free mask.
1104 */
1105STATIC int
1106xfs_inobt_first_free_inode(
1107 struct xfs_inobt_rec_incore *rec)
1108{
1109 xfs_inofree_t realfree;
1110
1111 /* if there are no holes, return the first available offset */
1112 if (!xfs_inobt_issparse(rec->ir_holemask))
1113 return xfs_lowbit64(rec->ir_free);
1114
1115 realfree = xfs_inobt_irec_to_allocmask(rec);
1116 realfree &= rec->ir_free;
1117
1118 return xfs_lowbit64(realfree);
1119}
1120
1121/*
1122 * Allocate an inode using the inobt-only algorithm.
1123 */
1124STATIC int
1125xfs_dialloc_ag_inobt(
1126 struct xfs_trans *tp,
1127 struct xfs_buf *agbp,
1128 xfs_ino_t parent,
1129 xfs_ino_t *inop)
1130{
1131 struct xfs_mount *mp = tp->t_mountp;
1132 struct xfs_agi *agi = agbp->b_addr;
1133 xfs_agnumber_t agno = be32_to_cpu(agi->agi_seqno);
1134 xfs_agnumber_t pagno = XFS_INO_TO_AGNO(mp, parent);
1135 xfs_agino_t pagino = XFS_INO_TO_AGINO(mp, parent);
1136 struct xfs_perag *pag = agbp->b_pag;
1137 struct xfs_btree_cur *cur, *tcur;
1138 struct xfs_inobt_rec_incore rec, trec;
1139 xfs_ino_t ino;
1140 int error;
1141 int offset;
1142 int i, j;
1143 int searchdistance = 10;
1144
1145 ASSERT(pag->pagi_init);
1146 ASSERT(pag->pagi_inodeok);
1147 ASSERT(pag->pagi_freecount > 0);
1148
1149 restart_pagno:
1150 cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
1151 /*
1152 * If pagino is 0 (this is the root inode allocation) use newino.
1153 * This must work because we've just allocated some.
1154 */
1155 if (!pagino)
1156 pagino = be32_to_cpu(agi->agi_newino);
1157
1158 error = xfs_check_agi_freecount(cur, agi);
1159 if (error)
1160 goto error0;
1161
1162 /*
1163 * If in the same AG as the parent, try to get near the parent.
1164 */
1165 if (pagno == agno) {
1166 int doneleft; /* done, to the left */
1167 int doneright; /* done, to the right */
1168
1169 error = xfs_inobt_lookup(cur, pagino, XFS_LOOKUP_LE, &i);
1170 if (error)
1171 goto error0;
1172 if (XFS_IS_CORRUPT(mp, i != 1)) {
1173 error = -EFSCORRUPTED;
1174 goto error0;
1175 }
1176
1177 error = xfs_inobt_get_rec(cur, &rec, &j);
1178 if (error)
1179 goto error0;
1180 if (XFS_IS_CORRUPT(mp, j != 1)) {
1181 error = -EFSCORRUPTED;
1182 goto error0;
1183 }
1184
1185 if (rec.ir_freecount > 0) {
1186 /*
1187 * Found a free inode in the same chunk
1188 * as the parent, done.
1189 */
1190 goto alloc_inode;
1191 }
1192
1193
1194 /*
1195 * In the same AG as parent, but parent's chunk is full.
1196 */
1197
1198 /* duplicate the cursor, search left & right simultaneously */
1199 error = xfs_btree_dup_cursor(cur, &tcur);
1200 if (error)
1201 goto error0;
1202
1203 /*
1204 * Skip to last blocks looked up if same parent inode.
1205 */
1206 if (pagino != NULLAGINO &&
1207 pag->pagl_pagino == pagino &&
1208 pag->pagl_leftrec != NULLAGINO &&
1209 pag->pagl_rightrec != NULLAGINO) {
1210 error = xfs_ialloc_get_rec(tcur, pag->pagl_leftrec,
1211 &trec, &doneleft);
1212 if (error)
1213 goto error1;
1214
1215 error = xfs_ialloc_get_rec(cur, pag->pagl_rightrec,
1216 &rec, &doneright);
1217 if (error)
1218 goto error1;
1219 } else {
1220 /* search left with tcur, back up 1 record */
1221 error = xfs_ialloc_next_rec(tcur, &trec, &doneleft, 1);
1222 if (error)
1223 goto error1;
1224
1225 /* search right with cur, go forward 1 record. */
1226 error = xfs_ialloc_next_rec(cur, &rec, &doneright, 0);
1227 if (error)
1228 goto error1;
1229 }
1230
1231 /*
1232 * Loop until we find an inode chunk with a free inode.
1233 */
1234 while (--searchdistance > 0 && (!doneleft || !doneright)) {
1235 int useleft; /* using left inode chunk this time */
1236
1237 /* figure out the closer block if both are valid. */
1238 if (!doneleft && !doneright) {
1239 useleft = pagino -
1240 (trec.ir_startino + XFS_INODES_PER_CHUNK - 1) <
1241 rec.ir_startino - pagino;
1242 } else {
1243 useleft = !doneleft;
1244 }
1245
1246 /* free inodes to the left? */
1247 if (useleft && trec.ir_freecount) {
1248 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1249 cur = tcur;
1250
1251 pag->pagl_leftrec = trec.ir_startino;
1252 pag->pagl_rightrec = rec.ir_startino;
1253 pag->pagl_pagino = pagino;
1254 rec = trec;
1255 goto alloc_inode;
1256 }
1257
1258 /* free inodes to the right? */
1259 if (!useleft && rec.ir_freecount) {
1260 xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1261
1262 pag->pagl_leftrec = trec.ir_startino;
1263 pag->pagl_rightrec = rec.ir_startino;
1264 pag->pagl_pagino = pagino;
1265 goto alloc_inode;
1266 }
1267
1268 /* get next record to check */
1269 if (useleft) {
1270 error = xfs_ialloc_next_rec(tcur, &trec,
1271 &doneleft, 1);
1272 } else {
1273 error = xfs_ialloc_next_rec(cur, &rec,
1274 &doneright, 0);
1275 }
1276 if (error)
1277 goto error1;
1278 }
1279
1280 if (searchdistance <= 0) {
1281 /*
1282 * Not in range - save last search
1283 * location and allocate a new inode
1284 */
1285 xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1286 pag->pagl_leftrec = trec.ir_startino;
1287 pag->pagl_rightrec = rec.ir_startino;
1288 pag->pagl_pagino = pagino;
1289
1290 } else {
1291 /*
1292 * We've reached the end of the btree. because
1293 * we are only searching a small chunk of the
1294 * btree each search, there is obviously free
1295 * inodes closer to the parent inode than we
1296 * are now. restart the search again.
1297 */
1298 pag->pagl_pagino = NULLAGINO;
1299 pag->pagl_leftrec = NULLAGINO;
1300 pag->pagl_rightrec = NULLAGINO;
1301 xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1302 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1303 goto restart_pagno;
1304 }
1305 }
1306
1307 /*
1308 * In a different AG from the parent.
1309 * See if the most recently allocated block has any free.
1310 */
1311 if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1312 error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1313 XFS_LOOKUP_EQ, &i);
1314 if (error)
1315 goto error0;
1316
1317 if (i == 1) {
1318 error = xfs_inobt_get_rec(cur, &rec, &j);
1319 if (error)
1320 goto error0;
1321
1322 if (j == 1 && rec.ir_freecount > 0) {
1323 /*
1324 * The last chunk allocated in the group
1325 * still has a free inode.
1326 */
1327 goto alloc_inode;
1328 }
1329 }
1330 }
1331
1332 /*
1333 * None left in the last group, search the whole AG
1334 */
1335 error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1336 if (error)
1337 goto error0;
1338 if (XFS_IS_CORRUPT(mp, i != 1)) {
1339 error = -EFSCORRUPTED;
1340 goto error0;
1341 }
1342
1343 for (;;) {
1344 error = xfs_inobt_get_rec(cur, &rec, &i);
1345 if (error)
1346 goto error0;
1347 if (XFS_IS_CORRUPT(mp, i != 1)) {
1348 error = -EFSCORRUPTED;
1349 goto error0;
1350 }
1351 if (rec.ir_freecount > 0)
1352 break;
1353 error = xfs_btree_increment(cur, 0, &i);
1354 if (error)
1355 goto error0;
1356 if (XFS_IS_CORRUPT(mp, i != 1)) {
1357 error = -EFSCORRUPTED;
1358 goto error0;
1359 }
1360 }
1361
1362alloc_inode:
1363 offset = xfs_inobt_first_free_inode(&rec);
1364 ASSERT(offset >= 0);
1365 ASSERT(offset < XFS_INODES_PER_CHUNK);
1366 ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1367 XFS_INODES_PER_CHUNK) == 0);
1368 ino = XFS_AGINO_TO_INO(mp, agno, rec.ir_startino + offset);
1369 rec.ir_free &= ~XFS_INOBT_MASK(offset);
1370 rec.ir_freecount--;
1371 error = xfs_inobt_update(cur, &rec);
1372 if (error)
1373 goto error0;
1374 be32_add_cpu(&agi->agi_freecount, -1);
1375 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1376 pag->pagi_freecount--;
1377
1378 error = xfs_check_agi_freecount(cur, agi);
1379 if (error)
1380 goto error0;
1381
1382 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1383 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1384 *inop = ino;
1385 return 0;
1386error1:
1387 xfs_btree_del_cursor(tcur, XFS_BTREE_ERROR);
1388error0:
1389 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1390 return error;
1391}
1392
1393/*
1394 * Use the free inode btree to allocate an inode based on distance from the
1395 * parent. Note that the provided cursor may be deleted and replaced.
1396 */
1397STATIC int
1398xfs_dialloc_ag_finobt_near(
1399 xfs_agino_t pagino,
1400 struct xfs_btree_cur **ocur,
1401 struct xfs_inobt_rec_incore *rec)
1402{
1403 struct xfs_btree_cur *lcur = *ocur; /* left search cursor */
1404 struct xfs_btree_cur *rcur; /* right search cursor */
1405 struct xfs_inobt_rec_incore rrec;
1406 int error;
1407 int i, j;
1408
1409 error = xfs_inobt_lookup(lcur, pagino, XFS_LOOKUP_LE, &i);
1410 if (error)
1411 return error;
1412
1413 if (i == 1) {
1414 error = xfs_inobt_get_rec(lcur, rec, &i);
1415 if (error)
1416 return error;
1417 if (XFS_IS_CORRUPT(lcur->bc_mp, i != 1))
1418 return -EFSCORRUPTED;
1419
1420 /*
1421 * See if we've landed in the parent inode record. The finobt
1422 * only tracks chunks with at least one free inode, so record
1423 * existence is enough.
1424 */
1425 if (pagino >= rec->ir_startino &&
1426 pagino < (rec->ir_startino + XFS_INODES_PER_CHUNK))
1427 return 0;
1428 }
1429
1430 error = xfs_btree_dup_cursor(lcur, &rcur);
1431 if (error)
1432 return error;
1433
1434 error = xfs_inobt_lookup(rcur, pagino, XFS_LOOKUP_GE, &j);
1435 if (error)
1436 goto error_rcur;
1437 if (j == 1) {
1438 error = xfs_inobt_get_rec(rcur, &rrec, &j);
1439 if (error)
1440 goto error_rcur;
1441 if (XFS_IS_CORRUPT(lcur->bc_mp, j != 1)) {
1442 error = -EFSCORRUPTED;
1443 goto error_rcur;
1444 }
1445 }
1446
1447 if (XFS_IS_CORRUPT(lcur->bc_mp, i != 1 && j != 1)) {
1448 error = -EFSCORRUPTED;
1449 goto error_rcur;
1450 }
1451 if (i == 1 && j == 1) {
1452 /*
1453 * Both the left and right records are valid. Choose the closer
1454 * inode chunk to the target.
1455 */
1456 if ((pagino - rec->ir_startino + XFS_INODES_PER_CHUNK - 1) >
1457 (rrec.ir_startino - pagino)) {
1458 *rec = rrec;
1459 xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1460 *ocur = rcur;
1461 } else {
1462 xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1463 }
1464 } else if (j == 1) {
1465 /* only the right record is valid */
1466 *rec = rrec;
1467 xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1468 *ocur = rcur;
1469 } else if (i == 1) {
1470 /* only the left record is valid */
1471 xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1472 }
1473
1474 return 0;
1475
1476error_rcur:
1477 xfs_btree_del_cursor(rcur, XFS_BTREE_ERROR);
1478 return error;
1479}
1480
1481/*
1482 * Use the free inode btree to find a free inode based on a newino hint. If
1483 * the hint is NULL, find the first free inode in the AG.
1484 */
1485STATIC int
1486xfs_dialloc_ag_finobt_newino(
1487 struct xfs_agi *agi,
1488 struct xfs_btree_cur *cur,
1489 struct xfs_inobt_rec_incore *rec)
1490{
1491 int error;
1492 int i;
1493
1494 if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1495 error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1496 XFS_LOOKUP_EQ, &i);
1497 if (error)
1498 return error;
1499 if (i == 1) {
1500 error = xfs_inobt_get_rec(cur, rec, &i);
1501 if (error)
1502 return error;
1503 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1504 return -EFSCORRUPTED;
1505 return 0;
1506 }
1507 }
1508
1509 /*
1510 * Find the first inode available in the AG.
1511 */
1512 error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1513 if (error)
1514 return error;
1515 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1516 return -EFSCORRUPTED;
1517
1518 error = xfs_inobt_get_rec(cur, rec, &i);
1519 if (error)
1520 return error;
1521 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1522 return -EFSCORRUPTED;
1523
1524 return 0;
1525}
1526
1527/*
1528 * Update the inobt based on a modification made to the finobt. Also ensure that
1529 * the records from both trees are equivalent post-modification.
1530 */
1531STATIC int
1532xfs_dialloc_ag_update_inobt(
1533 struct xfs_btree_cur *cur, /* inobt cursor */
1534 struct xfs_inobt_rec_incore *frec, /* finobt record */
1535 int offset) /* inode offset */
1536{
1537 struct xfs_inobt_rec_incore rec;
1538 int error;
1539 int i;
1540
1541 error = xfs_inobt_lookup(cur, frec->ir_startino, XFS_LOOKUP_EQ, &i);
1542 if (error)
1543 return error;
1544 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1545 return -EFSCORRUPTED;
1546
1547 error = xfs_inobt_get_rec(cur, &rec, &i);
1548 if (error)
1549 return error;
1550 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
1551 return -EFSCORRUPTED;
1552 ASSERT((XFS_AGINO_TO_OFFSET(cur->bc_mp, rec.ir_startino) %
1553 XFS_INODES_PER_CHUNK) == 0);
1554
1555 rec.ir_free &= ~XFS_INOBT_MASK(offset);
1556 rec.ir_freecount--;
1557
1558 if (XFS_IS_CORRUPT(cur->bc_mp,
1559 rec.ir_free != frec->ir_free ||
1560 rec.ir_freecount != frec->ir_freecount))
1561 return -EFSCORRUPTED;
1562
1563 return xfs_inobt_update(cur, &rec);
1564}
1565
1566/*
1567 * Allocate an inode using the free inode btree, if available. Otherwise, fall
1568 * back to the inobt search algorithm.
1569 *
1570 * The caller selected an AG for us, and made sure that free inodes are
1571 * available.
1572 */
1573STATIC int
1574xfs_dialloc_ag(
1575 struct xfs_trans *tp,
1576 struct xfs_buf *agbp,
1577 xfs_ino_t parent,
1578 xfs_ino_t *inop)
1579{
1580 struct xfs_mount *mp = tp->t_mountp;
1581 struct xfs_agi *agi = agbp->b_addr;
1582 xfs_agnumber_t agno = be32_to_cpu(agi->agi_seqno);
1583 xfs_agnumber_t pagno = XFS_INO_TO_AGNO(mp, parent);
1584 xfs_agino_t pagino = XFS_INO_TO_AGINO(mp, parent);
1585 struct xfs_btree_cur *cur; /* finobt cursor */
1586 struct xfs_btree_cur *icur; /* inobt cursor */
1587 struct xfs_inobt_rec_incore rec;
1588 xfs_ino_t ino;
1589 int error;
1590 int offset;
1591 int i;
1592
1593 if (!xfs_sb_version_hasfinobt(&mp->m_sb))
1594 return xfs_dialloc_ag_inobt(tp, agbp, parent, inop);
1595
1596 /*
1597 * If pagino is 0 (this is the root inode allocation) use newino.
1598 * This must work because we've just allocated some.
1599 */
1600 if (!pagino)
1601 pagino = be32_to_cpu(agi->agi_newino);
1602
1603 cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_FINO);
1604
1605 error = xfs_check_agi_freecount(cur, agi);
1606 if (error)
1607 goto error_cur;
1608
1609 /*
1610 * The search algorithm depends on whether we're in the same AG as the
1611 * parent. If so, find the closest available inode to the parent. If
1612 * not, consider the agi hint or find the first free inode in the AG.
1613 */
1614 if (agno == pagno)
1615 error = xfs_dialloc_ag_finobt_near(pagino, &cur, &rec);
1616 else
1617 error = xfs_dialloc_ag_finobt_newino(agi, cur, &rec);
1618 if (error)
1619 goto error_cur;
1620
1621 offset = xfs_inobt_first_free_inode(&rec);
1622 ASSERT(offset >= 0);
1623 ASSERT(offset < XFS_INODES_PER_CHUNK);
1624 ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1625 XFS_INODES_PER_CHUNK) == 0);
1626 ino = XFS_AGINO_TO_INO(mp, agno, rec.ir_startino + offset);
1627
1628 /*
1629 * Modify or remove the finobt record.
1630 */
1631 rec.ir_free &= ~XFS_INOBT_MASK(offset);
1632 rec.ir_freecount--;
1633 if (rec.ir_freecount)
1634 error = xfs_inobt_update(cur, &rec);
1635 else
1636 error = xfs_btree_delete(cur, &i);
1637 if (error)
1638 goto error_cur;
1639
1640 /*
1641 * The finobt has now been updated appropriately. We haven't updated the
1642 * agi and superblock yet, so we can create an inobt cursor and validate
1643 * the original freecount. If all is well, make the equivalent update to
1644 * the inobt using the finobt record and offset information.
1645 */
1646 icur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
1647
1648 error = xfs_check_agi_freecount(icur, agi);
1649 if (error)
1650 goto error_icur;
1651
1652 error = xfs_dialloc_ag_update_inobt(icur, &rec, offset);
1653 if (error)
1654 goto error_icur;
1655
1656 /*
1657 * Both trees have now been updated. We must update the perag and
1658 * superblock before we can check the freecount for each btree.
1659 */
1660 be32_add_cpu(&agi->agi_freecount, -1);
1661 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1662 agbp->b_pag->pagi_freecount--;
1663
1664 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1665
1666 error = xfs_check_agi_freecount(icur, agi);
1667 if (error)
1668 goto error_icur;
1669 error = xfs_check_agi_freecount(cur, agi);
1670 if (error)
1671 goto error_icur;
1672
1673 xfs_btree_del_cursor(icur, XFS_BTREE_NOERROR);
1674 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1675 *inop = ino;
1676 return 0;
1677
1678error_icur:
1679 xfs_btree_del_cursor(icur, XFS_BTREE_ERROR);
1680error_cur:
1681 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1682 return error;
1683}
1684
1685/*
1686 * Allocate an inode on disk.
1687 *
1688 * Mode is used to tell whether the new inode will need space, and whether it
1689 * is a directory.
1690 *
1691 * This function is designed to be called twice if it has to do an allocation
1692 * to make more free inodes. On the first call, *IO_agbp should be set to NULL.
1693 * If an inode is available without having to performn an allocation, an inode
1694 * number is returned. In this case, *IO_agbp is set to NULL. If an allocation
1695 * needs to be done, xfs_dialloc returns the current AGI buffer in *IO_agbp.
1696 * The caller should then commit the current transaction, allocate a
1697 * new transaction, and call xfs_dialloc() again, passing in the previous value
1698 * of *IO_agbp. IO_agbp should be held across the transactions. Since the AGI
1699 * buffer is locked across the two calls, the second call is guaranteed to have
1700 * a free inode available.
1701 *
1702 * Once we successfully pick an inode its number is returned and the on-disk
1703 * data structures are updated. The inode itself is not read in, since doing so
1704 * would break ordering constraints with xfs_reclaim.
1705 */
1706int
1707xfs_dialloc(
1708 struct xfs_trans *tp,
1709 xfs_ino_t parent,
1710 umode_t mode,
1711 struct xfs_buf **IO_agbp,
1712 xfs_ino_t *inop)
1713{
1714 struct xfs_mount *mp = tp->t_mountp;
1715 struct xfs_buf *agbp;
1716 xfs_agnumber_t agno;
1717 int error;
1718 int ialloced;
1719 int noroom = 0;
1720 xfs_agnumber_t start_agno;
1721 struct xfs_perag *pag;
1722 struct xfs_ino_geometry *igeo = M_IGEO(mp);
1723 int okalloc = 1;
1724
1725 if (*IO_agbp) {
1726 /*
1727 * If the caller passes in a pointer to the AGI buffer,
1728 * continue where we left off before. In this case, we
1729 * know that the allocation group has free inodes.
1730 */
1731 agbp = *IO_agbp;
1732 goto out_alloc;
1733 }
1734
1735 /*
1736 * We do not have an agbp, so select an initial allocation
1737 * group for inode allocation.
1738 */
1739 start_agno = xfs_ialloc_ag_select(tp, parent, mode);
1740 if (start_agno == NULLAGNUMBER) {
1741 *inop = NULLFSINO;
1742 return 0;
1743 }
1744
1745 /*
1746 * If we have already hit the ceiling of inode blocks then clear
1747 * okalloc so we scan all available agi structures for a free
1748 * inode.
1749 *
1750 * Read rough value of mp->m_icount by percpu_counter_read_positive,
1751 * which will sacrifice the preciseness but improve the performance.
1752 */
1753 if (igeo->maxicount &&
1754 percpu_counter_read_positive(&mp->m_icount) + igeo->ialloc_inos
1755 > igeo->maxicount) {
1756 noroom = 1;
1757 okalloc = 0;
1758 }
1759
1760 /*
1761 * Loop until we find an allocation group that either has free inodes
1762 * or in which we can allocate some inodes. Iterate through the
1763 * allocation groups upward, wrapping at the end.
1764 */
1765 agno = start_agno;
1766 for (;;) {
1767 pag = xfs_perag_get(mp, agno);
1768 if (!pag->pagi_inodeok) {
1769 xfs_ialloc_next_ag(mp);
1770 goto nextag;
1771 }
1772
1773 if (!pag->pagi_init) {
1774 error = xfs_ialloc_pagi_init(mp, tp, agno);
1775 if (error)
1776 goto out_error;
1777 }
1778
1779 /*
1780 * Do a first racy fast path check if this AG is usable.
1781 */
1782 if (!pag->pagi_freecount && !okalloc)
1783 goto nextag;
1784
1785 /*
1786 * Then read in the AGI buffer and recheck with the AGI buffer
1787 * lock held.
1788 */
1789 error = xfs_ialloc_read_agi(mp, tp, agno, &agbp);
1790 if (error)
1791 goto out_error;
1792
1793 if (pag->pagi_freecount) {
1794 xfs_perag_put(pag);
1795 goto out_alloc;
1796 }
1797
1798 if (!okalloc)
1799 goto nextag_relse_buffer;
1800
1801
1802 error = xfs_ialloc_ag_alloc(tp, agbp, &ialloced);
1803 if (error) {
1804 xfs_trans_brelse(tp, agbp);
1805
1806 if (error != -ENOSPC)
1807 goto out_error;
1808
1809 xfs_perag_put(pag);
1810 *inop = NULLFSINO;
1811 return 0;
1812 }
1813
1814 if (ialloced) {
1815 /*
1816 * We successfully allocated some inodes, return
1817 * the current context to the caller so that it
1818 * can commit the current transaction and call
1819 * us again where we left off.
1820 */
1821 ASSERT(pag->pagi_freecount > 0);
1822 xfs_perag_put(pag);
1823
1824 *IO_agbp = agbp;
1825 *inop = NULLFSINO;
1826 return 0;
1827 }
1828
1829nextag_relse_buffer:
1830 xfs_trans_brelse(tp, agbp);
1831nextag:
1832 xfs_perag_put(pag);
1833 if (++agno == mp->m_sb.sb_agcount)
1834 agno = 0;
1835 if (agno == start_agno) {
1836 *inop = NULLFSINO;
1837 return noroom ? -ENOSPC : 0;
1838 }
1839 }
1840
1841out_alloc:
1842 *IO_agbp = NULL;
1843 return xfs_dialloc_ag(tp, agbp, parent, inop);
1844out_error:
1845 xfs_perag_put(pag);
1846 return error;
1847}
1848
1849/*
1850 * Free the blocks of an inode chunk. We must consider that the inode chunk
1851 * might be sparse and only free the regions that are allocated as part of the
1852 * chunk.
1853 */
1854STATIC void
1855xfs_difree_inode_chunk(
1856 struct xfs_trans *tp,
1857 xfs_agnumber_t agno,
1858 struct xfs_inobt_rec_incore *rec)
1859{
1860 struct xfs_mount *mp = tp->t_mountp;
1861 xfs_agblock_t sagbno = XFS_AGINO_TO_AGBNO(mp,
1862 rec->ir_startino);
1863 int startidx, endidx;
1864 int nextbit;
1865 xfs_agblock_t agbno;
1866 int contigblk;
1867 DECLARE_BITMAP(holemask, XFS_INOBT_HOLEMASK_BITS);
1868
1869 if (!xfs_inobt_issparse(rec->ir_holemask)) {
1870 /* not sparse, calculate extent info directly */
1871 xfs_bmap_add_free(tp, XFS_AGB_TO_FSB(mp, agno, sagbno),
1872 M_IGEO(mp)->ialloc_blks,
1873 &XFS_RMAP_OINFO_INODES);
1874 return;
1875 }
1876
1877 /* holemask is only 16-bits (fits in an unsigned long) */
1878 ASSERT(sizeof(rec->ir_holemask) <= sizeof(holemask[0]));
1879 holemask[0] = rec->ir_holemask;
1880
1881 /*
1882 * Find contiguous ranges of zeroes (i.e., allocated regions) in the
1883 * holemask and convert the start/end index of each range to an extent.
1884 * We start with the start and end index both pointing at the first 0 in
1885 * the mask.
1886 */
1887 startidx = endidx = find_first_zero_bit(holemask,
1888 XFS_INOBT_HOLEMASK_BITS);
1889 nextbit = startidx + 1;
1890 while (startidx < XFS_INOBT_HOLEMASK_BITS) {
1891 nextbit = find_next_zero_bit(holemask, XFS_INOBT_HOLEMASK_BITS,
1892 nextbit);
1893 /*
1894 * If the next zero bit is contiguous, update the end index of
1895 * the current range and continue.
1896 */
1897 if (nextbit != XFS_INOBT_HOLEMASK_BITS &&
1898 nextbit == endidx + 1) {
1899 endidx = nextbit;
1900 goto next;
1901 }
1902
1903 /*
1904 * nextbit is not contiguous with the current end index. Convert
1905 * the current start/end to an extent and add it to the free
1906 * list.
1907 */
1908 agbno = sagbno + (startidx * XFS_INODES_PER_HOLEMASK_BIT) /
1909 mp->m_sb.sb_inopblock;
1910 contigblk = ((endidx - startidx + 1) *
1911 XFS_INODES_PER_HOLEMASK_BIT) /
1912 mp->m_sb.sb_inopblock;
1913
1914 ASSERT(agbno % mp->m_sb.sb_spino_align == 0);
1915 ASSERT(contigblk % mp->m_sb.sb_spino_align == 0);
1916 xfs_bmap_add_free(tp, XFS_AGB_TO_FSB(mp, agno, agbno),
1917 contigblk, &XFS_RMAP_OINFO_INODES);
1918
1919 /* reset range to current bit and carry on... */
1920 startidx = endidx = nextbit;
1921
1922next:
1923 nextbit++;
1924 }
1925}
1926
1927STATIC int
1928xfs_difree_inobt(
1929 struct xfs_mount *mp,
1930 struct xfs_trans *tp,
1931 struct xfs_buf *agbp,
1932 xfs_agino_t agino,
1933 struct xfs_icluster *xic,
1934 struct xfs_inobt_rec_incore *orec)
1935{
1936 struct xfs_agi *agi = agbp->b_addr;
1937 xfs_agnumber_t agno = be32_to_cpu(agi->agi_seqno);
1938 struct xfs_btree_cur *cur;
1939 struct xfs_inobt_rec_incore rec;
1940 int ilen;
1941 int error;
1942 int i;
1943 int off;
1944
1945 ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
1946 ASSERT(XFS_AGINO_TO_AGBNO(mp, agino) < be32_to_cpu(agi->agi_length));
1947
1948 /*
1949 * Initialize the cursor.
1950 */
1951 cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
1952
1953 error = xfs_check_agi_freecount(cur, agi);
1954 if (error)
1955 goto error0;
1956
1957 /*
1958 * Look for the entry describing this inode.
1959 */
1960 if ((error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i))) {
1961 xfs_warn(mp, "%s: xfs_inobt_lookup() returned error %d.",
1962 __func__, error);
1963 goto error0;
1964 }
1965 if (XFS_IS_CORRUPT(mp, i != 1)) {
1966 error = -EFSCORRUPTED;
1967 goto error0;
1968 }
1969 error = xfs_inobt_get_rec(cur, &rec, &i);
1970 if (error) {
1971 xfs_warn(mp, "%s: xfs_inobt_get_rec() returned error %d.",
1972 __func__, error);
1973 goto error0;
1974 }
1975 if (XFS_IS_CORRUPT(mp, i != 1)) {
1976 error = -EFSCORRUPTED;
1977 goto error0;
1978 }
1979 /*
1980 * Get the offset in the inode chunk.
1981 */
1982 off = agino - rec.ir_startino;
1983 ASSERT(off >= 0 && off < XFS_INODES_PER_CHUNK);
1984 ASSERT(!(rec.ir_free & XFS_INOBT_MASK(off)));
1985 /*
1986 * Mark the inode free & increment the count.
1987 */
1988 rec.ir_free |= XFS_INOBT_MASK(off);
1989 rec.ir_freecount++;
1990
1991 /*
1992 * When an inode chunk is free, it becomes eligible for removal. Don't
1993 * remove the chunk if the block size is large enough for multiple inode
1994 * chunks (that might not be free).
1995 */
1996 if (!(mp->m_flags & XFS_MOUNT_IKEEP) &&
1997 rec.ir_free == XFS_INOBT_ALL_FREE &&
1998 mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK) {
1999 struct xfs_perag *pag = agbp->b_pag;
2000
2001 xic->deleted = true;
2002 xic->first_ino = XFS_AGINO_TO_INO(mp, agno, rec.ir_startino);
2003 xic->alloc = xfs_inobt_irec_to_allocmask(&rec);
2004
2005 /*
2006 * Remove the inode cluster from the AGI B+Tree, adjust the
2007 * AGI and Superblock inode counts, and mark the disk space
2008 * to be freed when the transaction is committed.
2009 */
2010 ilen = rec.ir_freecount;
2011 be32_add_cpu(&agi->agi_count, -ilen);
2012 be32_add_cpu(&agi->agi_freecount, -(ilen - 1));
2013 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_COUNT | XFS_AGI_FREECOUNT);
2014 pag->pagi_freecount -= ilen - 1;
2015 pag->pagi_count -= ilen;
2016 xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, -ilen);
2017 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -(ilen - 1));
2018
2019 if ((error = xfs_btree_delete(cur, &i))) {
2020 xfs_warn(mp, "%s: xfs_btree_delete returned error %d.",
2021 __func__, error);
2022 goto error0;
2023 }
2024
2025 xfs_difree_inode_chunk(tp, agno, &rec);
2026 } else {
2027 xic->deleted = false;
2028
2029 error = xfs_inobt_update(cur, &rec);
2030 if (error) {
2031 xfs_warn(mp, "%s: xfs_inobt_update returned error %d.",
2032 __func__, error);
2033 goto error0;
2034 }
2035
2036 /*
2037 * Change the inode free counts and log the ag/sb changes.
2038 */
2039 be32_add_cpu(&agi->agi_freecount, 1);
2040 xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
2041 agbp->b_pag->pagi_freecount++;
2042 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, 1);
2043 }
2044
2045 error = xfs_check_agi_freecount(cur, agi);
2046 if (error)
2047 goto error0;
2048
2049 *orec = rec;
2050 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2051 return 0;
2052
2053error0:
2054 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2055 return error;
2056}
2057
2058/*
2059 * Free an inode in the free inode btree.
2060 */
2061STATIC int
2062xfs_difree_finobt(
2063 struct xfs_mount *mp,
2064 struct xfs_trans *tp,
2065 struct xfs_buf *agbp,
2066 xfs_agino_t agino,
2067 struct xfs_inobt_rec_incore *ibtrec) /* inobt record */
2068{
2069 struct xfs_agi *agi = agbp->b_addr;
2070 xfs_agnumber_t agno = be32_to_cpu(agi->agi_seqno);
2071 struct xfs_btree_cur *cur;
2072 struct xfs_inobt_rec_incore rec;
2073 int offset = agino - ibtrec->ir_startino;
2074 int error;
2075 int i;
2076
2077 cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_FINO);
2078
2079 error = xfs_inobt_lookup(cur, ibtrec->ir_startino, XFS_LOOKUP_EQ, &i);
2080 if (error)
2081 goto error;
2082 if (i == 0) {
2083 /*
2084 * If the record does not exist in the finobt, we must have just
2085 * freed an inode in a previously fully allocated chunk. If not,
2086 * something is out of sync.
2087 */
2088 if (XFS_IS_CORRUPT(mp, ibtrec->ir_freecount != 1)) {
2089 error = -EFSCORRUPTED;
2090 goto error;
2091 }
2092
2093 error = xfs_inobt_insert_rec(cur, ibtrec->ir_holemask,
2094 ibtrec->ir_count,
2095 ibtrec->ir_freecount,
2096 ibtrec->ir_free, &i);
2097 if (error)
2098 goto error;
2099 ASSERT(i == 1);
2100
2101 goto out;
2102 }
2103
2104 /*
2105 * Read and update the existing record. We could just copy the ibtrec
2106 * across here, but that would defeat the purpose of having redundant
2107 * metadata. By making the modifications independently, we can catch
2108 * corruptions that we wouldn't see if we just copied from one record
2109 * to another.
2110 */
2111 error = xfs_inobt_get_rec(cur, &rec, &i);
2112 if (error)
2113 goto error;
2114 if (XFS_IS_CORRUPT(mp, i != 1)) {
2115 error = -EFSCORRUPTED;
2116 goto error;
2117 }
2118
2119 rec.ir_free |= XFS_INOBT_MASK(offset);
2120 rec.ir_freecount++;
2121
2122 if (XFS_IS_CORRUPT(mp,
2123 rec.ir_free != ibtrec->ir_free ||
2124 rec.ir_freecount != ibtrec->ir_freecount)) {
2125 error = -EFSCORRUPTED;
2126 goto error;
2127 }
2128
2129 /*
2130 * The content of inobt records should always match between the inobt
2131 * and finobt. The lifecycle of records in the finobt is different from
2132 * the inobt in that the finobt only tracks records with at least one
2133 * free inode. Hence, if all of the inodes are free and we aren't
2134 * keeping inode chunks permanently on disk, remove the record.
2135 * Otherwise, update the record with the new information.
2136 *
2137 * Note that we currently can't free chunks when the block size is large
2138 * enough for multiple chunks. Leave the finobt record to remain in sync
2139 * with the inobt.
2140 */
2141 if (rec.ir_free == XFS_INOBT_ALL_FREE &&
2142 mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK &&
2143 !(mp->m_flags & XFS_MOUNT_IKEEP)) {
2144 error = xfs_btree_delete(cur, &i);
2145 if (error)
2146 goto error;
2147 ASSERT(i == 1);
2148 } else {
2149 error = xfs_inobt_update(cur, &rec);
2150 if (error)
2151 goto error;
2152 }
2153
2154out:
2155 error = xfs_check_agi_freecount(cur, agi);
2156 if (error)
2157 goto error;
2158
2159 xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2160 return 0;
2161
2162error:
2163 xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2164 return error;
2165}
2166
2167/*
2168 * Free disk inode. Carefully avoids touching the incore inode, all
2169 * manipulations incore are the caller's responsibility.
2170 * The on-disk inode is not changed by this operation, only the
2171 * btree (free inode mask) is changed.
2172 */
2173int
2174xfs_difree(
2175 struct xfs_trans *tp, /* transaction pointer */
2176 xfs_ino_t inode, /* inode to be freed */
2177 struct xfs_icluster *xic) /* cluster info if deleted */
2178{
2179 /* REFERENCED */
2180 xfs_agblock_t agbno; /* block number containing inode */
2181 struct xfs_buf *agbp; /* buffer for allocation group header */
2182 xfs_agino_t agino; /* allocation group inode number */
2183 xfs_agnumber_t agno; /* allocation group number */
2184 int error; /* error return value */
2185 struct xfs_mount *mp; /* mount structure for filesystem */
2186 struct xfs_inobt_rec_incore rec;/* btree record */
2187
2188 mp = tp->t_mountp;
2189
2190 /*
2191 * Break up inode number into its components.
2192 */
2193 agno = XFS_INO_TO_AGNO(mp, inode);
2194 if (agno >= mp->m_sb.sb_agcount) {
2195 xfs_warn(mp, "%s: agno >= mp->m_sb.sb_agcount (%d >= %d).",
2196 __func__, agno, mp->m_sb.sb_agcount);
2197 ASSERT(0);
2198 return -EINVAL;
2199 }
2200 agino = XFS_INO_TO_AGINO(mp, inode);
2201 if (inode != XFS_AGINO_TO_INO(mp, agno, agino)) {
2202 xfs_warn(mp, "%s: inode != XFS_AGINO_TO_INO() (%llu != %llu).",
2203 __func__, (unsigned long long)inode,
2204 (unsigned long long)XFS_AGINO_TO_INO(mp, agno, agino));
2205 ASSERT(0);
2206 return -EINVAL;
2207 }
2208 agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2209 if (agbno >= mp->m_sb.sb_agblocks) {
2210 xfs_warn(mp, "%s: agbno >= mp->m_sb.sb_agblocks (%d >= %d).",
2211 __func__, agbno, mp->m_sb.sb_agblocks);
2212 ASSERT(0);
2213 return -EINVAL;
2214 }
2215 /*
2216 * Get the allocation group header.
2217 */
2218 error = xfs_ialloc_read_agi(mp, tp, agno, &agbp);
2219 if (error) {
2220 xfs_warn(mp, "%s: xfs_ialloc_read_agi() returned error %d.",
2221 __func__, error);
2222 return error;
2223 }
2224
2225 /*
2226 * Fix up the inode allocation btree.
2227 */
2228 error = xfs_difree_inobt(mp, tp, agbp, agino, xic, &rec);
2229 if (error)
2230 goto error0;
2231
2232 /*
2233 * Fix up the free inode btree.
2234 */
2235 if (xfs_sb_version_hasfinobt(&mp->m_sb)) {
2236 error = xfs_difree_finobt(mp, tp, agbp, agino, &rec);
2237 if (error)
2238 goto error0;
2239 }
2240
2241 return 0;
2242
2243error0:
2244 return error;
2245}
2246
2247STATIC int
2248xfs_imap_lookup(
2249 struct xfs_mount *mp,
2250 struct xfs_trans *tp,
2251 xfs_agnumber_t agno,
2252 xfs_agino_t agino,
2253 xfs_agblock_t agbno,
2254 xfs_agblock_t *chunk_agbno,
2255 xfs_agblock_t *offset_agbno,
2256 int flags)
2257{
2258 struct xfs_inobt_rec_incore rec;
2259 struct xfs_btree_cur *cur;
2260 struct xfs_buf *agbp;
2261 int error;
2262 int i;
2263
2264 error = xfs_ialloc_read_agi(mp, tp, agno, &agbp);
2265 if (error) {
2266 xfs_alert(mp,
2267 "%s: xfs_ialloc_read_agi() returned error %d, agno %d",
2268 __func__, error, agno);
2269 return error;
2270 }
2271
2272 /*
2273 * Lookup the inode record for the given agino. If the record cannot be
2274 * found, then it's an invalid inode number and we should abort. Once
2275 * we have a record, we need to ensure it contains the inode number
2276 * we are looking up.
2277 */
2278 cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
2279 error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i);
2280 if (!error) {
2281 if (i)
2282 error = xfs_inobt_get_rec(cur, &rec, &i);
2283 if (!error && i == 0)
2284 error = -EINVAL;
2285 }
2286
2287 xfs_trans_brelse(tp, agbp);
2288 xfs_btree_del_cursor(cur, error);
2289 if (error)
2290 return error;
2291
2292 /* check that the returned record contains the required inode */
2293 if (rec.ir_startino > agino ||
2294 rec.ir_startino + M_IGEO(mp)->ialloc_inos <= agino)
2295 return -EINVAL;
2296
2297 /* for untrusted inodes check it is allocated first */
2298 if ((flags & XFS_IGET_UNTRUSTED) &&
2299 (rec.ir_free & XFS_INOBT_MASK(agino - rec.ir_startino)))
2300 return -EINVAL;
2301
2302 *chunk_agbno = XFS_AGINO_TO_AGBNO(mp, rec.ir_startino);
2303 *offset_agbno = agbno - *chunk_agbno;
2304 return 0;
2305}
2306
2307/*
2308 * Return the location of the inode in imap, for mapping it into a buffer.
2309 */
2310int
2311xfs_imap(
2312 xfs_mount_t *mp, /* file system mount structure */
2313 xfs_trans_t *tp, /* transaction pointer */
2314 xfs_ino_t ino, /* inode to locate */
2315 struct xfs_imap *imap, /* location map structure */
2316 uint flags) /* flags for inode btree lookup */
2317{
2318 xfs_agblock_t agbno; /* block number of inode in the alloc group */
2319 xfs_agino_t agino; /* inode number within alloc group */
2320 xfs_agnumber_t agno; /* allocation group number */
2321 xfs_agblock_t chunk_agbno; /* first block in inode chunk */
2322 xfs_agblock_t cluster_agbno; /* first block in inode cluster */
2323 int error; /* error code */
2324 int offset; /* index of inode in its buffer */
2325 xfs_agblock_t offset_agbno; /* blks from chunk start to inode */
2326
2327 ASSERT(ino != NULLFSINO);
2328
2329 /*
2330 * Split up the inode number into its parts.
2331 */
2332 agno = XFS_INO_TO_AGNO(mp, ino);
2333 agino = XFS_INO_TO_AGINO(mp, ino);
2334 agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2335 if (agno >= mp->m_sb.sb_agcount || agbno >= mp->m_sb.sb_agblocks ||
2336 ino != XFS_AGINO_TO_INO(mp, agno, agino)) {
2337#ifdef DEBUG
2338 /*
2339 * Don't output diagnostic information for untrusted inodes
2340 * as they can be invalid without implying corruption.
2341 */
2342 if (flags & XFS_IGET_UNTRUSTED)
2343 return -EINVAL;
2344 if (agno >= mp->m_sb.sb_agcount) {
2345 xfs_alert(mp,
2346 "%s: agno (%d) >= mp->m_sb.sb_agcount (%d)",
2347 __func__, agno, mp->m_sb.sb_agcount);
2348 }
2349 if (agbno >= mp->m_sb.sb_agblocks) {
2350 xfs_alert(mp,
2351 "%s: agbno (0x%llx) >= mp->m_sb.sb_agblocks (0x%lx)",
2352 __func__, (unsigned long long)agbno,
2353 (unsigned long)mp->m_sb.sb_agblocks);
2354 }
2355 if (ino != XFS_AGINO_TO_INO(mp, agno, agino)) {
2356 xfs_alert(mp,
2357 "%s: ino (0x%llx) != XFS_AGINO_TO_INO() (0x%llx)",
2358 __func__, ino,
2359 XFS_AGINO_TO_INO(mp, agno, agino));
2360 }
2361 xfs_stack_trace();
2362#endif /* DEBUG */
2363 return -EINVAL;
2364 }
2365
2366 /*
2367 * For bulkstat and handle lookups, we have an untrusted inode number
2368 * that we have to verify is valid. We cannot do this just by reading
2369 * the inode buffer as it may have been unlinked and removed leaving
2370 * inodes in stale state on disk. Hence we have to do a btree lookup
2371 * in all cases where an untrusted inode number is passed.
2372 */
2373 if (flags & XFS_IGET_UNTRUSTED) {
2374 error = xfs_imap_lookup(mp, tp, agno, agino, agbno,
2375 &chunk_agbno, &offset_agbno, flags);
2376 if (error)
2377 return error;
2378 goto out_map;
2379 }
2380
2381 /*
2382 * If the inode cluster size is the same as the blocksize or
2383 * smaller we get to the buffer by simple arithmetics.
2384 */
2385 if (M_IGEO(mp)->blocks_per_cluster == 1) {
2386 offset = XFS_INO_TO_OFFSET(mp, ino);
2387 ASSERT(offset < mp->m_sb.sb_inopblock);
2388
2389 imap->im_blkno = XFS_AGB_TO_DADDR(mp, agno, agbno);
2390 imap->im_len = XFS_FSB_TO_BB(mp, 1);
2391 imap->im_boffset = (unsigned short)(offset <<
2392 mp->m_sb.sb_inodelog);
2393 return 0;
2394 }
2395
2396 /*
2397 * If the inode chunks are aligned then use simple maths to
2398 * find the location. Otherwise we have to do a btree
2399 * lookup to find the location.
2400 */
2401 if (M_IGEO(mp)->inoalign_mask) {
2402 offset_agbno = agbno & M_IGEO(mp)->inoalign_mask;
2403 chunk_agbno = agbno - offset_agbno;
2404 } else {
2405 error = xfs_imap_lookup(mp, tp, agno, agino, agbno,
2406 &chunk_agbno, &offset_agbno, flags);
2407 if (error)
2408 return error;
2409 }
2410
2411out_map:
2412 ASSERT(agbno >= chunk_agbno);
2413 cluster_agbno = chunk_agbno +
2414 ((offset_agbno / M_IGEO(mp)->blocks_per_cluster) *
2415 M_IGEO(mp)->blocks_per_cluster);
2416 offset = ((agbno - cluster_agbno) * mp->m_sb.sb_inopblock) +
2417 XFS_INO_TO_OFFSET(mp, ino);
2418
2419 imap->im_blkno = XFS_AGB_TO_DADDR(mp, agno, cluster_agbno);
2420 imap->im_len = XFS_FSB_TO_BB(mp, M_IGEO(mp)->blocks_per_cluster);
2421 imap->im_boffset = (unsigned short)(offset << mp->m_sb.sb_inodelog);
2422
2423 /*
2424 * If the inode number maps to a block outside the bounds
2425 * of the file system then return NULL rather than calling
2426 * read_buf and panicing when we get an error from the
2427 * driver.
2428 */
2429 if ((imap->im_blkno + imap->im_len) >
2430 XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks)) {
2431 xfs_alert(mp,
2432 "%s: (im_blkno (0x%llx) + im_len (0x%llx)) > sb_dblocks (0x%llx)",
2433 __func__, (unsigned long long) imap->im_blkno,
2434 (unsigned long long) imap->im_len,
2435 XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks));
2436 return -EINVAL;
2437 }
2438 return 0;
2439}
2440
2441/*
2442 * Log specified fields for the ag hdr (inode section). The growth of the agi
2443 * structure over time requires that we interpret the buffer as two logical
2444 * regions delineated by the end of the unlinked list. This is due to the size
2445 * of the hash table and its location in the middle of the agi.
2446 *
2447 * For example, a request to log a field before agi_unlinked and a field after
2448 * agi_unlinked could cause us to log the entire hash table and use an excessive
2449 * amount of log space. To avoid this behavior, log the region up through
2450 * agi_unlinked in one call and the region after agi_unlinked through the end of
2451 * the structure in another.
2452 */
2453void
2454xfs_ialloc_log_agi(
2455 xfs_trans_t *tp, /* transaction pointer */
2456 xfs_buf_t *bp, /* allocation group header buffer */
2457 int fields) /* bitmask of fields to log */
2458{
2459 int first; /* first byte number */
2460 int last; /* last byte number */
2461 static const short offsets[] = { /* field starting offsets */
2462 /* keep in sync with bit definitions */
2463 offsetof(xfs_agi_t, agi_magicnum),
2464 offsetof(xfs_agi_t, agi_versionnum),
2465 offsetof(xfs_agi_t, agi_seqno),
2466 offsetof(xfs_agi_t, agi_length),
2467 offsetof(xfs_agi_t, agi_count),
2468 offsetof(xfs_agi_t, agi_root),
2469 offsetof(xfs_agi_t, agi_level),
2470 offsetof(xfs_agi_t, agi_freecount),
2471 offsetof(xfs_agi_t, agi_newino),
2472 offsetof(xfs_agi_t, agi_dirino),
2473 offsetof(xfs_agi_t, agi_unlinked),
2474 offsetof(xfs_agi_t, agi_free_root),
2475 offsetof(xfs_agi_t, agi_free_level),
2476 sizeof(xfs_agi_t)
2477 };
2478#ifdef DEBUG
2479 struct xfs_agi *agi = bp->b_addr;
2480
2481 ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
2482#endif
2483
2484 /*
2485 * Compute byte offsets for the first and last fields in the first
2486 * region and log the agi buffer. This only logs up through
2487 * agi_unlinked.
2488 */
2489 if (fields & XFS_AGI_ALL_BITS_R1) {
2490 xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R1,
2491 &first, &last);
2492 xfs_trans_log_buf(tp, bp, first, last);
2493 }
2494
2495 /*
2496 * Mask off the bits in the first region and calculate the first and
2497 * last field offsets for any bits in the second region.
2498 */
2499 fields &= ~XFS_AGI_ALL_BITS_R1;
2500 if (fields) {
2501 xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R2,
2502 &first, &last);
2503 xfs_trans_log_buf(tp, bp, first, last);
2504 }
2505}
2506
2507static xfs_failaddr_t
2508xfs_agi_verify(
2509 struct xfs_buf *bp)
2510{
2511 struct xfs_mount *mp = bp->b_mount;
2512 struct xfs_agi *agi = bp->b_addr;
2513 int i;
2514
2515 if (xfs_sb_version_hascrc(&mp->m_sb)) {
2516 if (!uuid_equal(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid))
2517 return __this_address;
2518 if (!xfs_log_check_lsn(mp, be64_to_cpu(agi->agi_lsn)))
2519 return __this_address;
2520 }
2521
2522 /*
2523 * Validate the magic number of the agi block.
2524 */
2525 if (!xfs_verify_magic(bp, agi->agi_magicnum))
2526 return __this_address;
2527 if (!XFS_AGI_GOOD_VERSION(be32_to_cpu(agi->agi_versionnum)))
2528 return __this_address;
2529
2530 if (be32_to_cpu(agi->agi_level) < 1 ||
2531 be32_to_cpu(agi->agi_level) > XFS_BTREE_MAXLEVELS)
2532 return __this_address;
2533
2534 if (xfs_sb_version_hasfinobt(&mp->m_sb) &&
2535 (be32_to_cpu(agi->agi_free_level) < 1 ||
2536 be32_to_cpu(agi->agi_free_level) > XFS_BTREE_MAXLEVELS))
2537 return __this_address;
2538
2539 /*
2540 * during growfs operations, the perag is not fully initialised,
2541 * so we can't use it for any useful checking. growfs ensures we can't
2542 * use it by using uncached buffers that don't have the perag attached
2543 * so we can detect and avoid this problem.
2544 */
2545 if (bp->b_pag && be32_to_cpu(agi->agi_seqno) != bp->b_pag->pag_agno)
2546 return __this_address;
2547
2548 for (i = 0; i < XFS_AGI_UNLINKED_BUCKETS; i++) {
2549 if (agi->agi_unlinked[i] == cpu_to_be32(NULLAGINO))
2550 continue;
2551 if (!xfs_verify_ino(mp, be32_to_cpu(agi->agi_unlinked[i])))
2552 return __this_address;
2553 }
2554
2555 return NULL;
2556}
2557
2558static void
2559xfs_agi_read_verify(
2560 struct xfs_buf *bp)
2561{
2562 struct xfs_mount *mp = bp->b_mount;
2563 xfs_failaddr_t fa;
2564
2565 if (xfs_sb_version_hascrc(&mp->m_sb) &&
2566 !xfs_buf_verify_cksum(bp, XFS_AGI_CRC_OFF))
2567 xfs_verifier_error(bp, -EFSBADCRC, __this_address);
2568 else {
2569 fa = xfs_agi_verify(bp);
2570 if (XFS_TEST_ERROR(fa, mp, XFS_ERRTAG_IALLOC_READ_AGI))
2571 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2572 }
2573}
2574
2575static void
2576xfs_agi_write_verify(
2577 struct xfs_buf *bp)
2578{
2579 struct xfs_mount *mp = bp->b_mount;
2580 struct xfs_buf_log_item *bip = bp->b_log_item;
2581 struct xfs_agi *agi = bp->b_addr;
2582 xfs_failaddr_t fa;
2583
2584 fa = xfs_agi_verify(bp);
2585 if (fa) {
2586 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2587 return;
2588 }
2589
2590 if (!xfs_sb_version_hascrc(&mp->m_sb))
2591 return;
2592
2593 if (bip)
2594 agi->agi_lsn = cpu_to_be64(bip->bli_item.li_lsn);
2595 xfs_buf_update_cksum(bp, XFS_AGI_CRC_OFF);
2596}
2597
2598const struct xfs_buf_ops xfs_agi_buf_ops = {
2599 .name = "xfs_agi",
2600 .magic = { cpu_to_be32(XFS_AGI_MAGIC), cpu_to_be32(XFS_AGI_MAGIC) },
2601 .verify_read = xfs_agi_read_verify,
2602 .verify_write = xfs_agi_write_verify,
2603 .verify_struct = xfs_agi_verify,
2604};
2605
2606/*
2607 * Read in the allocation group header (inode allocation section)
2608 */
2609int
2610xfs_read_agi(
2611 struct xfs_mount *mp, /* file system mount structure */
2612 struct xfs_trans *tp, /* transaction pointer */
2613 xfs_agnumber_t agno, /* allocation group number */
2614 struct xfs_buf **bpp) /* allocation group hdr buf */
2615{
2616 int error;
2617
2618 trace_xfs_read_agi(mp, agno);
2619
2620 ASSERT(agno != NULLAGNUMBER);
2621 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
2622 XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
2623 XFS_FSS_TO_BB(mp, 1), 0, bpp, &xfs_agi_buf_ops);
2624 if (error)
2625 return error;
2626 if (tp)
2627 xfs_trans_buf_set_type(tp, *bpp, XFS_BLFT_AGI_BUF);
2628
2629 xfs_buf_set_ref(*bpp, XFS_AGI_REF);
2630 return 0;
2631}
2632
2633int
2634xfs_ialloc_read_agi(
2635 struct xfs_mount *mp, /* file system mount structure */
2636 struct xfs_trans *tp, /* transaction pointer */
2637 xfs_agnumber_t agno, /* allocation group number */
2638 struct xfs_buf **bpp) /* allocation group hdr buf */
2639{
2640 struct xfs_agi *agi; /* allocation group header */
2641 struct xfs_perag *pag; /* per allocation group data */
2642 int error;
2643
2644 trace_xfs_ialloc_read_agi(mp, agno);
2645
2646 error = xfs_read_agi(mp, tp, agno, bpp);
2647 if (error)
2648 return error;
2649
2650 agi = (*bpp)->b_addr;
2651 pag = (*bpp)->b_pag;
2652 if (!pag->pagi_init) {
2653 pag->pagi_freecount = be32_to_cpu(agi->agi_freecount);
2654 pag->pagi_count = be32_to_cpu(agi->agi_count);
2655 pag->pagi_init = 1;
2656 }
2657
2658 /*
2659 * It's possible for these to be out of sync if
2660 * we are in the middle of a forced shutdown.
2661 */
2662 ASSERT(pag->pagi_freecount == be32_to_cpu(agi->agi_freecount) ||
2663 XFS_FORCED_SHUTDOWN(mp));
2664 return 0;
2665}
2666
2667/*
2668 * Read in the agi to initialise the per-ag data in the mount structure
2669 */
2670int
2671xfs_ialloc_pagi_init(
2672 xfs_mount_t *mp, /* file system mount structure */
2673 xfs_trans_t *tp, /* transaction pointer */
2674 xfs_agnumber_t agno) /* allocation group number */
2675{
2676 xfs_buf_t *bp = NULL;
2677 int error;
2678
2679 error = xfs_ialloc_read_agi(mp, tp, agno, &bp);
2680 if (error)
2681 return error;
2682 if (bp)
2683 xfs_trans_brelse(tp, bp);
2684 return 0;
2685}
2686
2687/* Is there an inode record covering a given range of inode numbers? */
2688int
2689xfs_ialloc_has_inode_record(
2690 struct xfs_btree_cur *cur,
2691 xfs_agino_t low,
2692 xfs_agino_t high,
2693 bool *exists)
2694{
2695 struct xfs_inobt_rec_incore irec;
2696 xfs_agino_t agino;
2697 uint16_t holemask;
2698 int has_record;
2699 int i;
2700 int error;
2701
2702 *exists = false;
2703 error = xfs_inobt_lookup(cur, low, XFS_LOOKUP_LE, &has_record);
2704 while (error == 0 && has_record) {
2705 error = xfs_inobt_get_rec(cur, &irec, &has_record);
2706 if (error || irec.ir_startino > high)
2707 break;
2708
2709 agino = irec.ir_startino;
2710 holemask = irec.ir_holemask;
2711 for (i = 0; i < XFS_INOBT_HOLEMASK_BITS; holemask >>= 1,
2712 i++, agino += XFS_INODES_PER_HOLEMASK_BIT) {
2713 if (holemask & 1)
2714 continue;
2715 if (agino + XFS_INODES_PER_HOLEMASK_BIT > low &&
2716 agino <= high) {
2717 *exists = true;
2718 return 0;
2719 }
2720 }
2721
2722 error = xfs_btree_increment(cur, 0, &has_record);
2723 }
2724 return error;
2725}
2726
2727/* Is there an inode record covering a given extent? */
2728int
2729xfs_ialloc_has_inodes_at_extent(
2730 struct xfs_btree_cur *cur,
2731 xfs_agblock_t bno,
2732 xfs_extlen_t len,
2733 bool *exists)
2734{
2735 xfs_agino_t low;
2736 xfs_agino_t high;
2737
2738 low = XFS_AGB_TO_AGINO(cur->bc_mp, bno);
2739 high = XFS_AGB_TO_AGINO(cur->bc_mp, bno + len) - 1;
2740
2741 return xfs_ialloc_has_inode_record(cur, low, high, exists);
2742}
2743
2744struct xfs_ialloc_count_inodes {
2745 xfs_agino_t count;
2746 xfs_agino_t freecount;
2747};
2748
2749/* Record inode counts across all inobt records. */
2750STATIC int
2751xfs_ialloc_count_inodes_rec(
2752 struct xfs_btree_cur *cur,
2753 union xfs_btree_rec *rec,
2754 void *priv)
2755{
2756 struct xfs_inobt_rec_incore irec;
2757 struct xfs_ialloc_count_inodes *ci = priv;
2758
2759 xfs_inobt_btrec_to_irec(cur->bc_mp, rec, &irec);
2760 ci->count += irec.ir_count;
2761 ci->freecount += irec.ir_freecount;
2762
2763 return 0;
2764}
2765
2766/* Count allocated and free inodes under an inobt. */
2767int
2768xfs_ialloc_count_inodes(
2769 struct xfs_btree_cur *cur,
2770 xfs_agino_t *count,
2771 xfs_agino_t *freecount)
2772{
2773 struct xfs_ialloc_count_inodes ci = {0};
2774 int error;
2775
2776 ASSERT(cur->bc_btnum == XFS_BTNUM_INO);
2777 error = xfs_btree_query_all(cur, xfs_ialloc_count_inodes_rec, &ci);
2778 if (error)
2779 return error;
2780
2781 *count = ci.count;
2782 *freecount = ci.freecount;
2783 return 0;
2784}
2785
2786/*
2787 * Initialize inode-related geometry information.
2788 *
2789 * Compute the inode btree min and max levels and set maxicount.
2790 *
2791 * Set the inode cluster size. This may still be overridden by the file
2792 * system block size if it is larger than the chosen cluster size.
2793 *
2794 * For v5 filesystems, scale the cluster size with the inode size to keep a
2795 * constant ratio of inode per cluster buffer, but only if mkfs has set the
2796 * inode alignment value appropriately for larger cluster sizes.
2797 *
2798 * Then compute the inode cluster alignment information.
2799 */
2800void
2801xfs_ialloc_setup_geometry(
2802 struct xfs_mount *mp)
2803{
2804 struct xfs_sb *sbp = &mp->m_sb;
2805 struct xfs_ino_geometry *igeo = M_IGEO(mp);
2806 uint64_t icount;
2807 uint inodes;
2808
2809 /* Compute inode btree geometry. */
2810 igeo->agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
2811 igeo->inobt_mxr[0] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, 1);
2812 igeo->inobt_mxr[1] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, 0);
2813 igeo->inobt_mnr[0] = igeo->inobt_mxr[0] / 2;
2814 igeo->inobt_mnr[1] = igeo->inobt_mxr[1] / 2;
2815
2816 igeo->ialloc_inos = max_t(uint16_t, XFS_INODES_PER_CHUNK,
2817 sbp->sb_inopblock);
2818 igeo->ialloc_blks = igeo->ialloc_inos >> sbp->sb_inopblog;
2819
2820 if (sbp->sb_spino_align)
2821 igeo->ialloc_min_blks = sbp->sb_spino_align;
2822 else
2823 igeo->ialloc_min_blks = igeo->ialloc_blks;
2824
2825 /* Compute and fill in value of m_ino_geo.inobt_maxlevels. */
2826 inodes = (1LL << XFS_INO_AGINO_BITS(mp)) >> XFS_INODES_PER_CHUNK_LOG;
2827 igeo->inobt_maxlevels = xfs_btree_compute_maxlevels(igeo->inobt_mnr,
2828 inodes);
2829
2830 /*
2831 * Set the maximum inode count for this filesystem, being careful not
2832 * to use obviously garbage sb_inopblog/sb_inopblock values. Regular
2833 * users should never get here due to failing sb verification, but
2834 * certain users (xfs_db) need to be usable even with corrupt metadata.
2835 */
2836 if (sbp->sb_imax_pct && igeo->ialloc_blks) {
2837 /*
2838 * Make sure the maximum inode count is a multiple
2839 * of the units we allocate inodes in.
2840 */
2841 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
2842 do_div(icount, 100);
2843 do_div(icount, igeo->ialloc_blks);
2844 igeo->maxicount = XFS_FSB_TO_INO(mp,
2845 icount * igeo->ialloc_blks);
2846 } else {
2847 igeo->maxicount = 0;
2848 }
2849
2850 /*
2851 * Compute the desired size of an inode cluster buffer size, which
2852 * starts at 8K and (on v5 filesystems) scales up with larger inode
2853 * sizes.
2854 *
2855 * Preserve the desired inode cluster size because the sparse inodes
2856 * feature uses that desired size (not the actual size) to compute the
2857 * sparse inode alignment. The mount code validates this value, so we
2858 * cannot change the behavior.
2859 */
2860 igeo->inode_cluster_size_raw = XFS_INODE_BIG_CLUSTER_SIZE;
2861 if (xfs_sb_version_has_v3inode(&mp->m_sb)) {
2862 int new_size = igeo->inode_cluster_size_raw;
2863
2864 new_size *= mp->m_sb.sb_inodesize / XFS_DINODE_MIN_SIZE;
2865 if (mp->m_sb.sb_inoalignmt >= XFS_B_TO_FSBT(mp, new_size))
2866 igeo->inode_cluster_size_raw = new_size;
2867 }
2868
2869 /* Calculate inode cluster ratios. */
2870 if (igeo->inode_cluster_size_raw > mp->m_sb.sb_blocksize)
2871 igeo->blocks_per_cluster = XFS_B_TO_FSBT(mp,
2872 igeo->inode_cluster_size_raw);
2873 else
2874 igeo->blocks_per_cluster = 1;
2875 igeo->inode_cluster_size = XFS_FSB_TO_B(mp, igeo->blocks_per_cluster);
2876 igeo->inodes_per_cluster = XFS_FSB_TO_INO(mp, igeo->blocks_per_cluster);
2877
2878 /* Calculate inode cluster alignment. */
2879 if (xfs_sb_version_hasalign(&mp->m_sb) &&
2880 mp->m_sb.sb_inoalignmt >= igeo->blocks_per_cluster)
2881 igeo->cluster_align = mp->m_sb.sb_inoalignmt;
2882 else
2883 igeo->cluster_align = 1;
2884 igeo->inoalign_mask = igeo->cluster_align - 1;
2885 igeo->cluster_align_inodes = XFS_FSB_TO_INO(mp, igeo->cluster_align);
2886
2887 /*
2888 * If we are using stripe alignment, check whether
2889 * the stripe unit is a multiple of the inode alignment
2890 */
2891 if (mp->m_dalign && igeo->inoalign_mask &&
2892 !(mp->m_dalign & igeo->inoalign_mask))
2893 igeo->ialloc_align = mp->m_dalign;
2894 else
2895 igeo->ialloc_align = 0;
2896}
2897
2898/* Compute the location of the root directory inode that is laid out by mkfs. */
2899xfs_ino_t
2900xfs_ialloc_calc_rootino(
2901 struct xfs_mount *mp,
2902 int sunit)
2903{
2904 struct xfs_ino_geometry *igeo = M_IGEO(mp);
2905 xfs_agblock_t first_bno;
2906
2907 /*
2908 * Pre-calculate the geometry of AG 0. We know what it looks like
2909 * because libxfs knows how to create allocation groups now.
2910 *
2911 * first_bno is the first block in which mkfs could possibly have
2912 * allocated the root directory inode, once we factor in the metadata
2913 * that mkfs formats before it. Namely, the four AG headers...
2914 */
2915 first_bno = howmany(4 * mp->m_sb.sb_sectsize, mp->m_sb.sb_blocksize);
2916
2917 /* ...the two free space btree roots... */
2918 first_bno += 2;
2919
2920 /* ...the inode btree root... */
2921 first_bno += 1;
2922
2923 /* ...the initial AGFL... */
2924 first_bno += xfs_alloc_min_freelist(mp, NULL);
2925
2926 /* ...the free inode btree root... */
2927 if (xfs_sb_version_hasfinobt(&mp->m_sb))
2928 first_bno++;
2929
2930 /* ...the reverse mapping btree root... */
2931 if (xfs_sb_version_hasrmapbt(&mp->m_sb))
2932 first_bno++;
2933
2934 /* ...the reference count btree... */
2935 if (xfs_sb_version_hasreflink(&mp->m_sb))
2936 first_bno++;
2937
2938 /*
2939 * ...and the log, if it is allocated in the first allocation group.
2940 *
2941 * This can happen with filesystems that only have a single
2942 * allocation group, or very odd geometries created by old mkfs
2943 * versions on very small filesystems.
2944 */
2945 if (mp->m_sb.sb_logstart &&
2946 XFS_FSB_TO_AGNO(mp, mp->m_sb.sb_logstart) == 0)
2947 first_bno += mp->m_sb.sb_logblocks;
2948
2949 /*
2950 * Now round first_bno up to whatever allocation alignment is given
2951 * by the filesystem or was passed in.
2952 */
2953 if (xfs_sb_version_hasdalign(&mp->m_sb) && igeo->ialloc_align > 0)
2954 first_bno = roundup(first_bno, sunit);
2955 else if (xfs_sb_version_hasalign(&mp->m_sb) &&
2956 mp->m_sb.sb_inoalignmt > 1)
2957 first_bno = roundup(first_bno, mp->m_sb.sb_inoalignmt);
2958
2959 return XFS_AGINO_TO_INO(mp, 0, XFS_AGB_TO_AGINO(mp, first_bno));
2960}