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