Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2014 Red Hat, 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_mount.h"
13#include "xfs_trans.h"
14#include "xfs_alloc.h"
15#include "xfs_btree.h"
16#include "xfs_btree_staging.h"
17#include "xfs_rmap.h"
18#include "xfs_rmap_btree.h"
19#include "xfs_trace.h"
20#include "xfs_error.h"
21#include "xfs_extent_busy.h"
22#include "xfs_ag.h"
23#include "xfs_ag_resv.h"
24
25/*
26 * Reverse map btree.
27 *
28 * This is a per-ag tree used to track the owner(s) of a given extent. With
29 * reflink it is possible for there to be multiple owners, which is a departure
30 * from classic XFS. Owner records for data extents are inserted when the
31 * extent is mapped and removed when an extent is unmapped. Owner records for
32 * all other block types (i.e. metadata) are inserted when an extent is
33 * allocated and removed when an extent is freed. There can only be one owner
34 * of a metadata extent, usually an inode or some other metadata structure like
35 * an AG btree.
36 *
37 * The rmap btree is part of the free space management, so blocks for the tree
38 * are sourced from the agfl. Hence we need transaction reservation support for
39 * this tree so that the freelist is always large enough. This also impacts on
40 * the minimum space we need to leave free in the AG.
41 *
42 * The tree is ordered by [ag block, owner, offset]. This is a large key size,
43 * but it is the only way to enforce unique keys when a block can be owned by
44 * multiple files at any offset. There's no need to order/search by extent
45 * size for online updating/management of the tree. It is intended that most
46 * reverse lookups will be to find the owner(s) of a particular block, or to
47 * try to recover tree and file data from corrupt primary metadata.
48 */
49
50static struct xfs_btree_cur *
51xfs_rmapbt_dup_cursor(
52 struct xfs_btree_cur *cur)
53{
54 return xfs_rmapbt_init_cursor(cur->bc_mp, cur->bc_tp,
55 cur->bc_ag.agbp, cur->bc_ag.pag);
56}
57
58STATIC void
59xfs_rmapbt_set_root(
60 struct xfs_btree_cur *cur,
61 union xfs_btree_ptr *ptr,
62 int inc)
63{
64 struct xfs_buf *agbp = cur->bc_ag.agbp;
65 struct xfs_agf *agf = agbp->b_addr;
66 int btnum = cur->bc_btnum;
67
68 ASSERT(ptr->s != 0);
69
70 agf->agf_roots[btnum] = ptr->s;
71 be32_add_cpu(&agf->agf_levels[btnum], inc);
72 cur->bc_ag.pag->pagf_levels[btnum] += inc;
73
74 xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_ROOTS | XFS_AGF_LEVELS);
75}
76
77STATIC int
78xfs_rmapbt_alloc_block(
79 struct xfs_btree_cur *cur,
80 union xfs_btree_ptr *start,
81 union xfs_btree_ptr *new,
82 int *stat)
83{
84 struct xfs_buf *agbp = cur->bc_ag.agbp;
85 struct xfs_agf *agf = agbp->b_addr;
86 struct xfs_perag *pag = cur->bc_ag.pag;
87 int error;
88 xfs_agblock_t bno;
89
90 /* Allocate the new block from the freelist. If we can't, give up. */
91 error = xfs_alloc_get_freelist(cur->bc_tp, cur->bc_ag.agbp,
92 &bno, 1);
93 if (error)
94 return error;
95
96 trace_xfs_rmapbt_alloc_block(cur->bc_mp, pag->pag_agno, bno, 1);
97 if (bno == NULLAGBLOCK) {
98 *stat = 0;
99 return 0;
100 }
101
102 xfs_extent_busy_reuse(cur->bc_mp, pag, bno, 1, false);
103
104 new->s = cpu_to_be32(bno);
105 be32_add_cpu(&agf->agf_rmap_blocks, 1);
106 xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_RMAP_BLOCKS);
107
108 xfs_ag_resv_rmapbt_alloc(cur->bc_mp, pag->pag_agno);
109
110 *stat = 1;
111 return 0;
112}
113
114STATIC int
115xfs_rmapbt_free_block(
116 struct xfs_btree_cur *cur,
117 struct xfs_buf *bp)
118{
119 struct xfs_buf *agbp = cur->bc_ag.agbp;
120 struct xfs_agf *agf = agbp->b_addr;
121 struct xfs_perag *pag = cur->bc_ag.pag;
122 xfs_agblock_t bno;
123 int error;
124
125 bno = xfs_daddr_to_agbno(cur->bc_mp, XFS_BUF_ADDR(bp));
126 trace_xfs_rmapbt_free_block(cur->bc_mp, pag->pag_agno,
127 bno, 1);
128 be32_add_cpu(&agf->agf_rmap_blocks, -1);
129 xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_RMAP_BLOCKS);
130 error = xfs_alloc_put_freelist(cur->bc_tp, agbp, NULL, bno, 1);
131 if (error)
132 return error;
133
134 xfs_extent_busy_insert(cur->bc_tp, pag, bno, 1,
135 XFS_EXTENT_BUSY_SKIP_DISCARD);
136
137 xfs_ag_resv_free_extent(pag, XFS_AG_RESV_RMAPBT, NULL, 1);
138 return 0;
139}
140
141STATIC int
142xfs_rmapbt_get_minrecs(
143 struct xfs_btree_cur *cur,
144 int level)
145{
146 return cur->bc_mp->m_rmap_mnr[level != 0];
147}
148
149STATIC int
150xfs_rmapbt_get_maxrecs(
151 struct xfs_btree_cur *cur,
152 int level)
153{
154 return cur->bc_mp->m_rmap_mxr[level != 0];
155}
156
157STATIC void
158xfs_rmapbt_init_key_from_rec(
159 union xfs_btree_key *key,
160 union xfs_btree_rec *rec)
161{
162 key->rmap.rm_startblock = rec->rmap.rm_startblock;
163 key->rmap.rm_owner = rec->rmap.rm_owner;
164 key->rmap.rm_offset = rec->rmap.rm_offset;
165}
166
167/*
168 * The high key for a reverse mapping record can be computed by shifting
169 * the startblock and offset to the highest value that would still map
170 * to that record. In practice this means that we add blockcount-1 to
171 * the startblock for all records, and if the record is for a data/attr
172 * fork mapping, we add blockcount-1 to the offset too.
173 */
174STATIC void
175xfs_rmapbt_init_high_key_from_rec(
176 union xfs_btree_key *key,
177 union xfs_btree_rec *rec)
178{
179 uint64_t off;
180 int adj;
181
182 adj = be32_to_cpu(rec->rmap.rm_blockcount) - 1;
183
184 key->rmap.rm_startblock = rec->rmap.rm_startblock;
185 be32_add_cpu(&key->rmap.rm_startblock, adj);
186 key->rmap.rm_owner = rec->rmap.rm_owner;
187 key->rmap.rm_offset = rec->rmap.rm_offset;
188 if (XFS_RMAP_NON_INODE_OWNER(be64_to_cpu(rec->rmap.rm_owner)) ||
189 XFS_RMAP_IS_BMBT_BLOCK(be64_to_cpu(rec->rmap.rm_offset)))
190 return;
191 off = be64_to_cpu(key->rmap.rm_offset);
192 off = (XFS_RMAP_OFF(off) + adj) | (off & ~XFS_RMAP_OFF_MASK);
193 key->rmap.rm_offset = cpu_to_be64(off);
194}
195
196STATIC void
197xfs_rmapbt_init_rec_from_cur(
198 struct xfs_btree_cur *cur,
199 union xfs_btree_rec *rec)
200{
201 rec->rmap.rm_startblock = cpu_to_be32(cur->bc_rec.r.rm_startblock);
202 rec->rmap.rm_blockcount = cpu_to_be32(cur->bc_rec.r.rm_blockcount);
203 rec->rmap.rm_owner = cpu_to_be64(cur->bc_rec.r.rm_owner);
204 rec->rmap.rm_offset = cpu_to_be64(
205 xfs_rmap_irec_offset_pack(&cur->bc_rec.r));
206}
207
208STATIC void
209xfs_rmapbt_init_ptr_from_cur(
210 struct xfs_btree_cur *cur,
211 union xfs_btree_ptr *ptr)
212{
213 struct xfs_agf *agf = cur->bc_ag.agbp->b_addr;
214
215 ASSERT(cur->bc_ag.pag->pag_agno == be32_to_cpu(agf->agf_seqno));
216
217 ptr->s = agf->agf_roots[cur->bc_btnum];
218}
219
220STATIC int64_t
221xfs_rmapbt_key_diff(
222 struct xfs_btree_cur *cur,
223 union xfs_btree_key *key)
224{
225 struct xfs_rmap_irec *rec = &cur->bc_rec.r;
226 struct xfs_rmap_key *kp = &key->rmap;
227 __u64 x, y;
228 int64_t d;
229
230 d = (int64_t)be32_to_cpu(kp->rm_startblock) - rec->rm_startblock;
231 if (d)
232 return d;
233
234 x = be64_to_cpu(kp->rm_owner);
235 y = rec->rm_owner;
236 if (x > y)
237 return 1;
238 else if (y > x)
239 return -1;
240
241 x = XFS_RMAP_OFF(be64_to_cpu(kp->rm_offset));
242 y = rec->rm_offset;
243 if (x > y)
244 return 1;
245 else if (y > x)
246 return -1;
247 return 0;
248}
249
250STATIC int64_t
251xfs_rmapbt_diff_two_keys(
252 struct xfs_btree_cur *cur,
253 union xfs_btree_key *k1,
254 union xfs_btree_key *k2)
255{
256 struct xfs_rmap_key *kp1 = &k1->rmap;
257 struct xfs_rmap_key *kp2 = &k2->rmap;
258 int64_t d;
259 __u64 x, y;
260
261 d = (int64_t)be32_to_cpu(kp1->rm_startblock) -
262 be32_to_cpu(kp2->rm_startblock);
263 if (d)
264 return d;
265
266 x = be64_to_cpu(kp1->rm_owner);
267 y = be64_to_cpu(kp2->rm_owner);
268 if (x > y)
269 return 1;
270 else if (y > x)
271 return -1;
272
273 x = XFS_RMAP_OFF(be64_to_cpu(kp1->rm_offset));
274 y = XFS_RMAP_OFF(be64_to_cpu(kp2->rm_offset));
275 if (x > y)
276 return 1;
277 else if (y > x)
278 return -1;
279 return 0;
280}
281
282static xfs_failaddr_t
283xfs_rmapbt_verify(
284 struct xfs_buf *bp)
285{
286 struct xfs_mount *mp = bp->b_mount;
287 struct xfs_btree_block *block = XFS_BUF_TO_BLOCK(bp);
288 struct xfs_perag *pag = bp->b_pag;
289 xfs_failaddr_t fa;
290 unsigned int level;
291
292 /*
293 * magic number and level verification
294 *
295 * During growfs operations, we can't verify the exact level or owner as
296 * the perag is not fully initialised and hence not attached to the
297 * buffer. In this case, check against the maximum tree depth.
298 *
299 * Similarly, during log recovery we will have a perag structure
300 * attached, but the agf information will not yet have been initialised
301 * from the on disk AGF. Again, we can only check against maximum limits
302 * in this case.
303 */
304 if (!xfs_verify_magic(bp, block->bb_magic))
305 return __this_address;
306
307 if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
308 return __this_address;
309 fa = xfs_btree_sblock_v5hdr_verify(bp);
310 if (fa)
311 return fa;
312
313 level = be16_to_cpu(block->bb_level);
314 if (pag && pag->pagf_init) {
315 if (level >= pag->pagf_levels[XFS_BTNUM_RMAPi])
316 return __this_address;
317 } else if (level >= mp->m_rmap_maxlevels)
318 return __this_address;
319
320 return xfs_btree_sblock_verify(bp, mp->m_rmap_mxr[level != 0]);
321}
322
323static void
324xfs_rmapbt_read_verify(
325 struct xfs_buf *bp)
326{
327 xfs_failaddr_t fa;
328
329 if (!xfs_btree_sblock_verify_crc(bp))
330 xfs_verifier_error(bp, -EFSBADCRC, __this_address);
331 else {
332 fa = xfs_rmapbt_verify(bp);
333 if (fa)
334 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
335 }
336
337 if (bp->b_error)
338 trace_xfs_btree_corrupt(bp, _RET_IP_);
339}
340
341static void
342xfs_rmapbt_write_verify(
343 struct xfs_buf *bp)
344{
345 xfs_failaddr_t fa;
346
347 fa = xfs_rmapbt_verify(bp);
348 if (fa) {
349 trace_xfs_btree_corrupt(bp, _RET_IP_);
350 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
351 return;
352 }
353 xfs_btree_sblock_calc_crc(bp);
354
355}
356
357const struct xfs_buf_ops xfs_rmapbt_buf_ops = {
358 .name = "xfs_rmapbt",
359 .magic = { 0, cpu_to_be32(XFS_RMAP_CRC_MAGIC) },
360 .verify_read = xfs_rmapbt_read_verify,
361 .verify_write = xfs_rmapbt_write_verify,
362 .verify_struct = xfs_rmapbt_verify,
363};
364
365STATIC int
366xfs_rmapbt_keys_inorder(
367 struct xfs_btree_cur *cur,
368 union xfs_btree_key *k1,
369 union xfs_btree_key *k2)
370{
371 uint32_t x;
372 uint32_t y;
373 uint64_t a;
374 uint64_t b;
375
376 x = be32_to_cpu(k1->rmap.rm_startblock);
377 y = be32_to_cpu(k2->rmap.rm_startblock);
378 if (x < y)
379 return 1;
380 else if (x > y)
381 return 0;
382 a = be64_to_cpu(k1->rmap.rm_owner);
383 b = be64_to_cpu(k2->rmap.rm_owner);
384 if (a < b)
385 return 1;
386 else if (a > b)
387 return 0;
388 a = XFS_RMAP_OFF(be64_to_cpu(k1->rmap.rm_offset));
389 b = XFS_RMAP_OFF(be64_to_cpu(k2->rmap.rm_offset));
390 if (a <= b)
391 return 1;
392 return 0;
393}
394
395STATIC int
396xfs_rmapbt_recs_inorder(
397 struct xfs_btree_cur *cur,
398 union xfs_btree_rec *r1,
399 union xfs_btree_rec *r2)
400{
401 uint32_t x;
402 uint32_t y;
403 uint64_t a;
404 uint64_t b;
405
406 x = be32_to_cpu(r1->rmap.rm_startblock);
407 y = be32_to_cpu(r2->rmap.rm_startblock);
408 if (x < y)
409 return 1;
410 else if (x > y)
411 return 0;
412 a = be64_to_cpu(r1->rmap.rm_owner);
413 b = be64_to_cpu(r2->rmap.rm_owner);
414 if (a < b)
415 return 1;
416 else if (a > b)
417 return 0;
418 a = XFS_RMAP_OFF(be64_to_cpu(r1->rmap.rm_offset));
419 b = XFS_RMAP_OFF(be64_to_cpu(r2->rmap.rm_offset));
420 if (a <= b)
421 return 1;
422 return 0;
423}
424
425static const struct xfs_btree_ops xfs_rmapbt_ops = {
426 .rec_len = sizeof(struct xfs_rmap_rec),
427 .key_len = 2 * sizeof(struct xfs_rmap_key),
428
429 .dup_cursor = xfs_rmapbt_dup_cursor,
430 .set_root = xfs_rmapbt_set_root,
431 .alloc_block = xfs_rmapbt_alloc_block,
432 .free_block = xfs_rmapbt_free_block,
433 .get_minrecs = xfs_rmapbt_get_minrecs,
434 .get_maxrecs = xfs_rmapbt_get_maxrecs,
435 .init_key_from_rec = xfs_rmapbt_init_key_from_rec,
436 .init_high_key_from_rec = xfs_rmapbt_init_high_key_from_rec,
437 .init_rec_from_cur = xfs_rmapbt_init_rec_from_cur,
438 .init_ptr_from_cur = xfs_rmapbt_init_ptr_from_cur,
439 .key_diff = xfs_rmapbt_key_diff,
440 .buf_ops = &xfs_rmapbt_buf_ops,
441 .diff_two_keys = xfs_rmapbt_diff_two_keys,
442 .keys_inorder = xfs_rmapbt_keys_inorder,
443 .recs_inorder = xfs_rmapbt_recs_inorder,
444};
445
446static struct xfs_btree_cur *
447xfs_rmapbt_init_common(
448 struct xfs_mount *mp,
449 struct xfs_trans *tp,
450 struct xfs_perag *pag)
451{
452 struct xfs_btree_cur *cur;
453
454 cur = kmem_cache_zalloc(xfs_btree_cur_zone, GFP_NOFS | __GFP_NOFAIL);
455 cur->bc_tp = tp;
456 cur->bc_mp = mp;
457 /* Overlapping btree; 2 keys per pointer. */
458 cur->bc_btnum = XFS_BTNUM_RMAP;
459 cur->bc_flags = XFS_BTREE_CRC_BLOCKS | XFS_BTREE_OVERLAPPING;
460 cur->bc_blocklog = mp->m_sb.sb_blocklog;
461 cur->bc_statoff = XFS_STATS_CALC_INDEX(xs_rmap_2);
462 cur->bc_ops = &xfs_rmapbt_ops;
463
464 /* take a reference for the cursor */
465 atomic_inc(&pag->pag_ref);
466 cur->bc_ag.pag = pag;
467
468 return cur;
469}
470
471/* Create a new reverse mapping btree cursor. */
472struct xfs_btree_cur *
473xfs_rmapbt_init_cursor(
474 struct xfs_mount *mp,
475 struct xfs_trans *tp,
476 struct xfs_buf *agbp,
477 struct xfs_perag *pag)
478{
479 struct xfs_agf *agf = agbp->b_addr;
480 struct xfs_btree_cur *cur;
481
482 cur = xfs_rmapbt_init_common(mp, tp, pag);
483 cur->bc_nlevels = be32_to_cpu(agf->agf_levels[XFS_BTNUM_RMAP]);
484 cur->bc_ag.agbp = agbp;
485 return cur;
486}
487
488/* Create a new reverse mapping btree cursor with a fake root for staging. */
489struct xfs_btree_cur *
490xfs_rmapbt_stage_cursor(
491 struct xfs_mount *mp,
492 struct xbtree_afakeroot *afake,
493 struct xfs_perag *pag)
494{
495 struct xfs_btree_cur *cur;
496
497 cur = xfs_rmapbt_init_common(mp, NULL, pag);
498 xfs_btree_stage_afakeroot(cur, afake);
499 return cur;
500}
501
502/*
503 * Install a new reverse mapping btree root. Caller is responsible for
504 * invalidating and freeing the old btree blocks.
505 */
506void
507xfs_rmapbt_commit_staged_btree(
508 struct xfs_btree_cur *cur,
509 struct xfs_trans *tp,
510 struct xfs_buf *agbp)
511{
512 struct xfs_agf *agf = agbp->b_addr;
513 struct xbtree_afakeroot *afake = cur->bc_ag.afake;
514
515 ASSERT(cur->bc_flags & XFS_BTREE_STAGING);
516
517 agf->agf_roots[cur->bc_btnum] = cpu_to_be32(afake->af_root);
518 agf->agf_levels[cur->bc_btnum] = cpu_to_be32(afake->af_levels);
519 agf->agf_rmap_blocks = cpu_to_be32(afake->af_blocks);
520 xfs_alloc_log_agf(tp, agbp, XFS_AGF_ROOTS | XFS_AGF_LEVELS |
521 XFS_AGF_RMAP_BLOCKS);
522 xfs_btree_commit_afakeroot(cur, tp, agbp, &xfs_rmapbt_ops);
523}
524
525/*
526 * Calculate number of records in an rmap btree block.
527 */
528int
529xfs_rmapbt_maxrecs(
530 int blocklen,
531 int leaf)
532{
533 blocklen -= XFS_RMAP_BLOCK_LEN;
534
535 if (leaf)
536 return blocklen / sizeof(struct xfs_rmap_rec);
537 return blocklen /
538 (2 * sizeof(struct xfs_rmap_key) + sizeof(xfs_rmap_ptr_t));
539}
540
541/* Compute the maximum height of an rmap btree. */
542void
543xfs_rmapbt_compute_maxlevels(
544 struct xfs_mount *mp)
545{
546 /*
547 * On a non-reflink filesystem, the maximum number of rmap
548 * records is the number of blocks in the AG, hence the max
549 * rmapbt height is log_$maxrecs($agblocks). However, with
550 * reflink each AG block can have up to 2^32 (per the refcount
551 * record format) owners, which means that theoretically we
552 * could face up to 2^64 rmap records.
553 *
554 * That effectively means that the max rmapbt height must be
555 * XFS_BTREE_MAXLEVELS. "Fortunately" we'll run out of AG
556 * blocks to feed the rmapbt long before the rmapbt reaches
557 * maximum height. The reflink code uses ag_resv_critical to
558 * disallow reflinking when less than 10% of the per-AG metadata
559 * block reservation since the fallback is a regular file copy.
560 */
561 if (xfs_sb_version_hasreflink(&mp->m_sb))
562 mp->m_rmap_maxlevels = XFS_BTREE_MAXLEVELS;
563 else
564 mp->m_rmap_maxlevels = xfs_btree_compute_maxlevels(
565 mp->m_rmap_mnr, mp->m_sb.sb_agblocks);
566}
567
568/* Calculate the refcount btree size for some records. */
569xfs_extlen_t
570xfs_rmapbt_calc_size(
571 struct xfs_mount *mp,
572 unsigned long long len)
573{
574 return xfs_btree_calc_size(mp->m_rmap_mnr, len);
575}
576
577/*
578 * Calculate the maximum refcount btree size.
579 */
580xfs_extlen_t
581xfs_rmapbt_max_size(
582 struct xfs_mount *mp,
583 xfs_agblock_t agblocks)
584{
585 /* Bail out if we're uninitialized, which can happen in mkfs. */
586 if (mp->m_rmap_mxr[0] == 0)
587 return 0;
588
589 return xfs_rmapbt_calc_size(mp, agblocks);
590}
591
592/*
593 * Figure out how many blocks to reserve and how many are used by this btree.
594 */
595int
596xfs_rmapbt_calc_reserves(
597 struct xfs_mount *mp,
598 struct xfs_trans *tp,
599 struct xfs_perag *pag,
600 xfs_extlen_t *ask,
601 xfs_extlen_t *used)
602{
603 struct xfs_buf *agbp;
604 struct xfs_agf *agf;
605 xfs_agblock_t agblocks;
606 xfs_extlen_t tree_len;
607 int error;
608
609 if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
610 return 0;
611
612 error = xfs_alloc_read_agf(mp, tp, pag->pag_agno, 0, &agbp);
613 if (error)
614 return error;
615
616 agf = agbp->b_addr;
617 agblocks = be32_to_cpu(agf->agf_length);
618 tree_len = be32_to_cpu(agf->agf_rmap_blocks);
619 xfs_trans_brelse(tp, agbp);
620
621 /*
622 * The log is permanently allocated, so the space it occupies will
623 * never be available for the kinds of things that would require btree
624 * expansion. We therefore can pretend the space isn't there.
625 */
626 if (mp->m_sb.sb_logstart &&
627 XFS_FSB_TO_AGNO(mp, mp->m_sb.sb_logstart) == pag->pag_agno)
628 agblocks -= mp->m_sb.sb_logblocks;
629
630 /* Reserve 1% of the AG or enough for 1 block per record. */
631 *ask += max(agblocks / 100, xfs_rmapbt_max_size(mp, agblocks));
632 *used += tree_len;
633
634 return error;
635}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2014 Red Hat, 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_mount.h"
13#include "xfs_trans.h"
14#include "xfs_alloc.h"
15#include "xfs_btree.h"
16#include "xfs_btree_staging.h"
17#include "xfs_rmap.h"
18#include "xfs_rmap_btree.h"
19#include "xfs_trace.h"
20#include "xfs_error.h"
21#include "xfs_extent_busy.h"
22#include "xfs_ag.h"
23#include "xfs_ag_resv.h"
24
25static struct kmem_cache *xfs_rmapbt_cur_cache;
26
27/*
28 * Reverse map btree.
29 *
30 * This is a per-ag tree used to track the owner(s) of a given extent. With
31 * reflink it is possible for there to be multiple owners, which is a departure
32 * from classic XFS. Owner records for data extents are inserted when the
33 * extent is mapped and removed when an extent is unmapped. Owner records for
34 * all other block types (i.e. metadata) are inserted when an extent is
35 * allocated and removed when an extent is freed. There can only be one owner
36 * of a metadata extent, usually an inode or some other metadata structure like
37 * an AG btree.
38 *
39 * The rmap btree is part of the free space management, so blocks for the tree
40 * are sourced from the agfl. Hence we need transaction reservation support for
41 * this tree so that the freelist is always large enough. This also impacts on
42 * the minimum space we need to leave free in the AG.
43 *
44 * The tree is ordered by [ag block, owner, offset]. This is a large key size,
45 * but it is the only way to enforce unique keys when a block can be owned by
46 * multiple files at any offset. There's no need to order/search by extent
47 * size for online updating/management of the tree. It is intended that most
48 * reverse lookups will be to find the owner(s) of a particular block, or to
49 * try to recover tree and file data from corrupt primary metadata.
50 */
51
52static struct xfs_btree_cur *
53xfs_rmapbt_dup_cursor(
54 struct xfs_btree_cur *cur)
55{
56 return xfs_rmapbt_init_cursor(cur->bc_mp, cur->bc_tp,
57 cur->bc_ag.agbp, cur->bc_ag.pag);
58}
59
60STATIC void
61xfs_rmapbt_set_root(
62 struct xfs_btree_cur *cur,
63 const union xfs_btree_ptr *ptr,
64 int inc)
65{
66 struct xfs_buf *agbp = cur->bc_ag.agbp;
67 struct xfs_agf *agf = agbp->b_addr;
68 int btnum = cur->bc_btnum;
69
70 ASSERT(ptr->s != 0);
71
72 agf->agf_roots[btnum] = ptr->s;
73 be32_add_cpu(&agf->agf_levels[btnum], inc);
74 cur->bc_ag.pag->pagf_levels[btnum] += inc;
75
76 xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_ROOTS | XFS_AGF_LEVELS);
77}
78
79STATIC int
80xfs_rmapbt_alloc_block(
81 struct xfs_btree_cur *cur,
82 const union xfs_btree_ptr *start,
83 union xfs_btree_ptr *new,
84 int *stat)
85{
86 struct xfs_buf *agbp = cur->bc_ag.agbp;
87 struct xfs_agf *agf = agbp->b_addr;
88 struct xfs_perag *pag = cur->bc_ag.pag;
89 int error;
90 xfs_agblock_t bno;
91
92 /* Allocate the new block from the freelist. If we can't, give up. */
93 error = xfs_alloc_get_freelist(pag, cur->bc_tp, cur->bc_ag.agbp,
94 &bno, 1);
95 if (error)
96 return error;
97
98 trace_xfs_rmapbt_alloc_block(cur->bc_mp, pag->pag_agno, bno, 1);
99 if (bno == NULLAGBLOCK) {
100 *stat = 0;
101 return 0;
102 }
103
104 xfs_extent_busy_reuse(cur->bc_mp, pag, bno, 1, false);
105
106 new->s = cpu_to_be32(bno);
107 be32_add_cpu(&agf->agf_rmap_blocks, 1);
108 xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_RMAP_BLOCKS);
109
110 xfs_ag_resv_rmapbt_alloc(cur->bc_mp, pag->pag_agno);
111
112 *stat = 1;
113 return 0;
114}
115
116STATIC int
117xfs_rmapbt_free_block(
118 struct xfs_btree_cur *cur,
119 struct xfs_buf *bp)
120{
121 struct xfs_buf *agbp = cur->bc_ag.agbp;
122 struct xfs_agf *agf = agbp->b_addr;
123 struct xfs_perag *pag = cur->bc_ag.pag;
124 xfs_agblock_t bno;
125 int error;
126
127 bno = xfs_daddr_to_agbno(cur->bc_mp, xfs_buf_daddr(bp));
128 trace_xfs_rmapbt_free_block(cur->bc_mp, pag->pag_agno,
129 bno, 1);
130 be32_add_cpu(&agf->agf_rmap_blocks, -1);
131 xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_RMAP_BLOCKS);
132 error = xfs_alloc_put_freelist(pag, cur->bc_tp, agbp, NULL, bno, 1);
133 if (error)
134 return error;
135
136 xfs_extent_busy_insert(cur->bc_tp, pag, bno, 1,
137 XFS_EXTENT_BUSY_SKIP_DISCARD);
138
139 xfs_ag_resv_free_extent(pag, XFS_AG_RESV_RMAPBT, NULL, 1);
140 return 0;
141}
142
143STATIC int
144xfs_rmapbt_get_minrecs(
145 struct xfs_btree_cur *cur,
146 int level)
147{
148 return cur->bc_mp->m_rmap_mnr[level != 0];
149}
150
151STATIC int
152xfs_rmapbt_get_maxrecs(
153 struct xfs_btree_cur *cur,
154 int level)
155{
156 return cur->bc_mp->m_rmap_mxr[level != 0];
157}
158
159/*
160 * Convert the ondisk record's offset field into the ondisk key's offset field.
161 * Fork and bmbt are significant parts of the rmap record key, but written
162 * status is merely a record attribute.
163 */
164static inline __be64 ondisk_rec_offset_to_key(const union xfs_btree_rec *rec)
165{
166 return rec->rmap.rm_offset & ~cpu_to_be64(XFS_RMAP_OFF_UNWRITTEN);
167}
168
169STATIC void
170xfs_rmapbt_init_key_from_rec(
171 union xfs_btree_key *key,
172 const union xfs_btree_rec *rec)
173{
174 key->rmap.rm_startblock = rec->rmap.rm_startblock;
175 key->rmap.rm_owner = rec->rmap.rm_owner;
176 key->rmap.rm_offset = ondisk_rec_offset_to_key(rec);
177}
178
179/*
180 * The high key for a reverse mapping record can be computed by shifting
181 * the startblock and offset to the highest value that would still map
182 * to that record. In practice this means that we add blockcount-1 to
183 * the startblock for all records, and if the record is for a data/attr
184 * fork mapping, we add blockcount-1 to the offset too.
185 */
186STATIC void
187xfs_rmapbt_init_high_key_from_rec(
188 union xfs_btree_key *key,
189 const union xfs_btree_rec *rec)
190{
191 uint64_t off;
192 int adj;
193
194 adj = be32_to_cpu(rec->rmap.rm_blockcount) - 1;
195
196 key->rmap.rm_startblock = rec->rmap.rm_startblock;
197 be32_add_cpu(&key->rmap.rm_startblock, adj);
198 key->rmap.rm_owner = rec->rmap.rm_owner;
199 key->rmap.rm_offset = ondisk_rec_offset_to_key(rec);
200 if (XFS_RMAP_NON_INODE_OWNER(be64_to_cpu(rec->rmap.rm_owner)) ||
201 XFS_RMAP_IS_BMBT_BLOCK(be64_to_cpu(rec->rmap.rm_offset)))
202 return;
203 off = be64_to_cpu(key->rmap.rm_offset);
204 off = (XFS_RMAP_OFF(off) + adj) | (off & ~XFS_RMAP_OFF_MASK);
205 key->rmap.rm_offset = cpu_to_be64(off);
206}
207
208STATIC void
209xfs_rmapbt_init_rec_from_cur(
210 struct xfs_btree_cur *cur,
211 union xfs_btree_rec *rec)
212{
213 rec->rmap.rm_startblock = cpu_to_be32(cur->bc_rec.r.rm_startblock);
214 rec->rmap.rm_blockcount = cpu_to_be32(cur->bc_rec.r.rm_blockcount);
215 rec->rmap.rm_owner = cpu_to_be64(cur->bc_rec.r.rm_owner);
216 rec->rmap.rm_offset = cpu_to_be64(
217 xfs_rmap_irec_offset_pack(&cur->bc_rec.r));
218}
219
220STATIC void
221xfs_rmapbt_init_ptr_from_cur(
222 struct xfs_btree_cur *cur,
223 union xfs_btree_ptr *ptr)
224{
225 struct xfs_agf *agf = cur->bc_ag.agbp->b_addr;
226
227 ASSERT(cur->bc_ag.pag->pag_agno == be32_to_cpu(agf->agf_seqno));
228
229 ptr->s = agf->agf_roots[cur->bc_btnum];
230}
231
232/*
233 * Mask the appropriate parts of the ondisk key field for a key comparison.
234 * Fork and bmbt are significant parts of the rmap record key, but written
235 * status is merely a record attribute.
236 */
237static inline uint64_t offset_keymask(uint64_t offset)
238{
239 return offset & ~XFS_RMAP_OFF_UNWRITTEN;
240}
241
242STATIC int64_t
243xfs_rmapbt_key_diff(
244 struct xfs_btree_cur *cur,
245 const union xfs_btree_key *key)
246{
247 struct xfs_rmap_irec *rec = &cur->bc_rec.r;
248 const struct xfs_rmap_key *kp = &key->rmap;
249 __u64 x, y;
250 int64_t d;
251
252 d = (int64_t)be32_to_cpu(kp->rm_startblock) - rec->rm_startblock;
253 if (d)
254 return d;
255
256 x = be64_to_cpu(kp->rm_owner);
257 y = rec->rm_owner;
258 if (x > y)
259 return 1;
260 else if (y > x)
261 return -1;
262
263 x = offset_keymask(be64_to_cpu(kp->rm_offset));
264 y = offset_keymask(xfs_rmap_irec_offset_pack(rec));
265 if (x > y)
266 return 1;
267 else if (y > x)
268 return -1;
269 return 0;
270}
271
272STATIC int64_t
273xfs_rmapbt_diff_two_keys(
274 struct xfs_btree_cur *cur,
275 const union xfs_btree_key *k1,
276 const union xfs_btree_key *k2,
277 const union xfs_btree_key *mask)
278{
279 const struct xfs_rmap_key *kp1 = &k1->rmap;
280 const struct xfs_rmap_key *kp2 = &k2->rmap;
281 int64_t d;
282 __u64 x, y;
283
284 /* Doesn't make sense to mask off the physical space part */
285 ASSERT(!mask || mask->rmap.rm_startblock);
286
287 d = (int64_t)be32_to_cpu(kp1->rm_startblock) -
288 be32_to_cpu(kp2->rm_startblock);
289 if (d)
290 return d;
291
292 if (!mask || mask->rmap.rm_owner) {
293 x = be64_to_cpu(kp1->rm_owner);
294 y = be64_to_cpu(kp2->rm_owner);
295 if (x > y)
296 return 1;
297 else if (y > x)
298 return -1;
299 }
300
301 if (!mask || mask->rmap.rm_offset) {
302 /* Doesn't make sense to allow offset but not owner */
303 ASSERT(!mask || mask->rmap.rm_owner);
304
305 x = offset_keymask(be64_to_cpu(kp1->rm_offset));
306 y = offset_keymask(be64_to_cpu(kp2->rm_offset));
307 if (x > y)
308 return 1;
309 else if (y > x)
310 return -1;
311 }
312
313 return 0;
314}
315
316static xfs_failaddr_t
317xfs_rmapbt_verify(
318 struct xfs_buf *bp)
319{
320 struct xfs_mount *mp = bp->b_mount;
321 struct xfs_btree_block *block = XFS_BUF_TO_BLOCK(bp);
322 struct xfs_perag *pag = bp->b_pag;
323 xfs_failaddr_t fa;
324 unsigned int level;
325
326 /*
327 * magic number and level verification
328 *
329 * During growfs operations, we can't verify the exact level or owner as
330 * the perag is not fully initialised and hence not attached to the
331 * buffer. In this case, check against the maximum tree depth.
332 *
333 * Similarly, during log recovery we will have a perag structure
334 * attached, but the agf information will not yet have been initialised
335 * from the on disk AGF. Again, we can only check against maximum limits
336 * in this case.
337 */
338 if (!xfs_verify_magic(bp, block->bb_magic))
339 return __this_address;
340
341 if (!xfs_has_rmapbt(mp))
342 return __this_address;
343 fa = xfs_btree_sblock_v5hdr_verify(bp);
344 if (fa)
345 return fa;
346
347 level = be16_to_cpu(block->bb_level);
348 if (pag && xfs_perag_initialised_agf(pag)) {
349 if (level >= pag->pagf_levels[XFS_BTNUM_RMAPi])
350 return __this_address;
351 } else if (level >= mp->m_rmap_maxlevels)
352 return __this_address;
353
354 return xfs_btree_sblock_verify(bp, mp->m_rmap_mxr[level != 0]);
355}
356
357static void
358xfs_rmapbt_read_verify(
359 struct xfs_buf *bp)
360{
361 xfs_failaddr_t fa;
362
363 if (!xfs_btree_sblock_verify_crc(bp))
364 xfs_verifier_error(bp, -EFSBADCRC, __this_address);
365 else {
366 fa = xfs_rmapbt_verify(bp);
367 if (fa)
368 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
369 }
370
371 if (bp->b_error)
372 trace_xfs_btree_corrupt(bp, _RET_IP_);
373}
374
375static void
376xfs_rmapbt_write_verify(
377 struct xfs_buf *bp)
378{
379 xfs_failaddr_t fa;
380
381 fa = xfs_rmapbt_verify(bp);
382 if (fa) {
383 trace_xfs_btree_corrupt(bp, _RET_IP_);
384 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
385 return;
386 }
387 xfs_btree_sblock_calc_crc(bp);
388
389}
390
391const struct xfs_buf_ops xfs_rmapbt_buf_ops = {
392 .name = "xfs_rmapbt",
393 .magic = { 0, cpu_to_be32(XFS_RMAP_CRC_MAGIC) },
394 .verify_read = xfs_rmapbt_read_verify,
395 .verify_write = xfs_rmapbt_write_verify,
396 .verify_struct = xfs_rmapbt_verify,
397};
398
399STATIC int
400xfs_rmapbt_keys_inorder(
401 struct xfs_btree_cur *cur,
402 const union xfs_btree_key *k1,
403 const union xfs_btree_key *k2)
404{
405 uint32_t x;
406 uint32_t y;
407 uint64_t a;
408 uint64_t b;
409
410 x = be32_to_cpu(k1->rmap.rm_startblock);
411 y = be32_to_cpu(k2->rmap.rm_startblock);
412 if (x < y)
413 return 1;
414 else if (x > y)
415 return 0;
416 a = be64_to_cpu(k1->rmap.rm_owner);
417 b = be64_to_cpu(k2->rmap.rm_owner);
418 if (a < b)
419 return 1;
420 else if (a > b)
421 return 0;
422 a = offset_keymask(be64_to_cpu(k1->rmap.rm_offset));
423 b = offset_keymask(be64_to_cpu(k2->rmap.rm_offset));
424 if (a <= b)
425 return 1;
426 return 0;
427}
428
429STATIC int
430xfs_rmapbt_recs_inorder(
431 struct xfs_btree_cur *cur,
432 const union xfs_btree_rec *r1,
433 const union xfs_btree_rec *r2)
434{
435 uint32_t x;
436 uint32_t y;
437 uint64_t a;
438 uint64_t b;
439
440 x = be32_to_cpu(r1->rmap.rm_startblock);
441 y = be32_to_cpu(r2->rmap.rm_startblock);
442 if (x < y)
443 return 1;
444 else if (x > y)
445 return 0;
446 a = be64_to_cpu(r1->rmap.rm_owner);
447 b = be64_to_cpu(r2->rmap.rm_owner);
448 if (a < b)
449 return 1;
450 else if (a > b)
451 return 0;
452 a = offset_keymask(be64_to_cpu(r1->rmap.rm_offset));
453 b = offset_keymask(be64_to_cpu(r2->rmap.rm_offset));
454 if (a <= b)
455 return 1;
456 return 0;
457}
458
459STATIC enum xbtree_key_contig
460xfs_rmapbt_keys_contiguous(
461 struct xfs_btree_cur *cur,
462 const union xfs_btree_key *key1,
463 const union xfs_btree_key *key2,
464 const union xfs_btree_key *mask)
465{
466 ASSERT(!mask || mask->rmap.rm_startblock);
467
468 /*
469 * We only support checking contiguity of the physical space component.
470 * If any callers ever need more specificity than that, they'll have to
471 * implement it here.
472 */
473 ASSERT(!mask || (!mask->rmap.rm_owner && !mask->rmap.rm_offset));
474
475 return xbtree_key_contig(be32_to_cpu(key1->rmap.rm_startblock),
476 be32_to_cpu(key2->rmap.rm_startblock));
477}
478
479static const struct xfs_btree_ops xfs_rmapbt_ops = {
480 .rec_len = sizeof(struct xfs_rmap_rec),
481 .key_len = 2 * sizeof(struct xfs_rmap_key),
482
483 .dup_cursor = xfs_rmapbt_dup_cursor,
484 .set_root = xfs_rmapbt_set_root,
485 .alloc_block = xfs_rmapbt_alloc_block,
486 .free_block = xfs_rmapbt_free_block,
487 .get_minrecs = xfs_rmapbt_get_minrecs,
488 .get_maxrecs = xfs_rmapbt_get_maxrecs,
489 .init_key_from_rec = xfs_rmapbt_init_key_from_rec,
490 .init_high_key_from_rec = xfs_rmapbt_init_high_key_from_rec,
491 .init_rec_from_cur = xfs_rmapbt_init_rec_from_cur,
492 .init_ptr_from_cur = xfs_rmapbt_init_ptr_from_cur,
493 .key_diff = xfs_rmapbt_key_diff,
494 .buf_ops = &xfs_rmapbt_buf_ops,
495 .diff_two_keys = xfs_rmapbt_diff_two_keys,
496 .keys_inorder = xfs_rmapbt_keys_inorder,
497 .recs_inorder = xfs_rmapbt_recs_inorder,
498 .keys_contiguous = xfs_rmapbt_keys_contiguous,
499};
500
501static struct xfs_btree_cur *
502xfs_rmapbt_init_common(
503 struct xfs_mount *mp,
504 struct xfs_trans *tp,
505 struct xfs_perag *pag)
506{
507 struct xfs_btree_cur *cur;
508
509 /* Overlapping btree; 2 keys per pointer. */
510 cur = xfs_btree_alloc_cursor(mp, tp, XFS_BTNUM_RMAP,
511 mp->m_rmap_maxlevels, xfs_rmapbt_cur_cache);
512 cur->bc_flags = XFS_BTREE_CRC_BLOCKS | XFS_BTREE_OVERLAPPING;
513 cur->bc_statoff = XFS_STATS_CALC_INDEX(xs_rmap_2);
514 cur->bc_ops = &xfs_rmapbt_ops;
515
516 cur->bc_ag.pag = xfs_perag_hold(pag);
517 return cur;
518}
519
520/* Create a new reverse mapping btree cursor. */
521struct xfs_btree_cur *
522xfs_rmapbt_init_cursor(
523 struct xfs_mount *mp,
524 struct xfs_trans *tp,
525 struct xfs_buf *agbp,
526 struct xfs_perag *pag)
527{
528 struct xfs_agf *agf = agbp->b_addr;
529 struct xfs_btree_cur *cur;
530
531 cur = xfs_rmapbt_init_common(mp, tp, pag);
532 cur->bc_nlevels = be32_to_cpu(agf->agf_levels[XFS_BTNUM_RMAP]);
533 cur->bc_ag.agbp = agbp;
534 return cur;
535}
536
537/* Create a new reverse mapping btree cursor with a fake root for staging. */
538struct xfs_btree_cur *
539xfs_rmapbt_stage_cursor(
540 struct xfs_mount *mp,
541 struct xbtree_afakeroot *afake,
542 struct xfs_perag *pag)
543{
544 struct xfs_btree_cur *cur;
545
546 cur = xfs_rmapbt_init_common(mp, NULL, pag);
547 xfs_btree_stage_afakeroot(cur, afake);
548 return cur;
549}
550
551/*
552 * Install a new reverse mapping btree root. Caller is responsible for
553 * invalidating and freeing the old btree blocks.
554 */
555void
556xfs_rmapbt_commit_staged_btree(
557 struct xfs_btree_cur *cur,
558 struct xfs_trans *tp,
559 struct xfs_buf *agbp)
560{
561 struct xfs_agf *agf = agbp->b_addr;
562 struct xbtree_afakeroot *afake = cur->bc_ag.afake;
563
564 ASSERT(cur->bc_flags & XFS_BTREE_STAGING);
565
566 agf->agf_roots[cur->bc_btnum] = cpu_to_be32(afake->af_root);
567 agf->agf_levels[cur->bc_btnum] = cpu_to_be32(afake->af_levels);
568 agf->agf_rmap_blocks = cpu_to_be32(afake->af_blocks);
569 xfs_alloc_log_agf(tp, agbp, XFS_AGF_ROOTS | XFS_AGF_LEVELS |
570 XFS_AGF_RMAP_BLOCKS);
571 xfs_btree_commit_afakeroot(cur, tp, agbp, &xfs_rmapbt_ops);
572}
573
574/* Calculate number of records in a reverse mapping btree block. */
575static inline unsigned int
576xfs_rmapbt_block_maxrecs(
577 unsigned int blocklen,
578 bool leaf)
579{
580 if (leaf)
581 return blocklen / sizeof(struct xfs_rmap_rec);
582 return blocklen /
583 (2 * sizeof(struct xfs_rmap_key) + sizeof(xfs_rmap_ptr_t));
584}
585
586/*
587 * Calculate number of records in an rmap btree block.
588 */
589int
590xfs_rmapbt_maxrecs(
591 int blocklen,
592 int leaf)
593{
594 blocklen -= XFS_RMAP_BLOCK_LEN;
595 return xfs_rmapbt_block_maxrecs(blocklen, leaf);
596}
597
598/* Compute the max possible height for reverse mapping btrees. */
599unsigned int
600xfs_rmapbt_maxlevels_ondisk(void)
601{
602 unsigned int minrecs[2];
603 unsigned int blocklen;
604
605 blocklen = XFS_MIN_CRC_BLOCKSIZE - XFS_BTREE_SBLOCK_CRC_LEN;
606
607 minrecs[0] = xfs_rmapbt_block_maxrecs(blocklen, true) / 2;
608 minrecs[1] = xfs_rmapbt_block_maxrecs(blocklen, false) / 2;
609
610 /*
611 * Compute the asymptotic maxlevels for an rmapbt on any reflink fs.
612 *
613 * On a reflink filesystem, each AG block can have up to 2^32 (per the
614 * refcount record format) owners, which means that theoretically we
615 * could face up to 2^64 rmap records. However, we're likely to run
616 * out of blocks in the AG long before that happens, which means that
617 * we must compute the max height based on what the btree will look
618 * like if it consumes almost all the blocks in the AG due to maximal
619 * sharing factor.
620 */
621 return xfs_btree_space_to_height(minrecs, XFS_MAX_CRC_AG_BLOCKS);
622}
623
624/* Compute the maximum height of an rmap btree. */
625void
626xfs_rmapbt_compute_maxlevels(
627 struct xfs_mount *mp)
628{
629 if (!xfs_has_rmapbt(mp)) {
630 mp->m_rmap_maxlevels = 0;
631 return;
632 }
633
634 if (xfs_has_reflink(mp)) {
635 /*
636 * Compute the asymptotic maxlevels for an rmap btree on a
637 * filesystem that supports reflink.
638 *
639 * On a reflink filesystem, each AG block can have up to 2^32
640 * (per the refcount record format) owners, which means that
641 * theoretically we could face up to 2^64 rmap records.
642 * However, we're likely to run out of blocks in the AG long
643 * before that happens, which means that we must compute the
644 * max height based on what the btree will look like if it
645 * consumes almost all the blocks in the AG due to maximal
646 * sharing factor.
647 */
648 mp->m_rmap_maxlevels = xfs_btree_space_to_height(mp->m_rmap_mnr,
649 mp->m_sb.sb_agblocks);
650 } else {
651 /*
652 * If there's no block sharing, compute the maximum rmapbt
653 * height assuming one rmap record per AG block.
654 */
655 mp->m_rmap_maxlevels = xfs_btree_compute_maxlevels(
656 mp->m_rmap_mnr, mp->m_sb.sb_agblocks);
657 }
658 ASSERT(mp->m_rmap_maxlevels <= xfs_rmapbt_maxlevels_ondisk());
659}
660
661/* Calculate the refcount btree size for some records. */
662xfs_extlen_t
663xfs_rmapbt_calc_size(
664 struct xfs_mount *mp,
665 unsigned long long len)
666{
667 return xfs_btree_calc_size(mp->m_rmap_mnr, len);
668}
669
670/*
671 * Calculate the maximum refcount btree size.
672 */
673xfs_extlen_t
674xfs_rmapbt_max_size(
675 struct xfs_mount *mp,
676 xfs_agblock_t agblocks)
677{
678 /* Bail out if we're uninitialized, which can happen in mkfs. */
679 if (mp->m_rmap_mxr[0] == 0)
680 return 0;
681
682 return xfs_rmapbt_calc_size(mp, agblocks);
683}
684
685/*
686 * Figure out how many blocks to reserve and how many are used by this btree.
687 */
688int
689xfs_rmapbt_calc_reserves(
690 struct xfs_mount *mp,
691 struct xfs_trans *tp,
692 struct xfs_perag *pag,
693 xfs_extlen_t *ask,
694 xfs_extlen_t *used)
695{
696 struct xfs_buf *agbp;
697 struct xfs_agf *agf;
698 xfs_agblock_t agblocks;
699 xfs_extlen_t tree_len;
700 int error;
701
702 if (!xfs_has_rmapbt(mp))
703 return 0;
704
705 error = xfs_alloc_read_agf(pag, tp, 0, &agbp);
706 if (error)
707 return error;
708
709 agf = agbp->b_addr;
710 agblocks = be32_to_cpu(agf->agf_length);
711 tree_len = be32_to_cpu(agf->agf_rmap_blocks);
712 xfs_trans_brelse(tp, agbp);
713
714 /*
715 * The log is permanently allocated, so the space it occupies will
716 * never be available for the kinds of things that would require btree
717 * expansion. We therefore can pretend the space isn't there.
718 */
719 if (xfs_ag_contains_log(mp, pag->pag_agno))
720 agblocks -= mp->m_sb.sb_logblocks;
721
722 /* Reserve 1% of the AG or enough for 1 block per record. */
723 *ask += max(agblocks / 100, xfs_rmapbt_max_size(mp, agblocks));
724 *used += tree_len;
725
726 return error;
727}
728
729int __init
730xfs_rmapbt_init_cur_cache(void)
731{
732 xfs_rmapbt_cur_cache = kmem_cache_create("xfs_rmapbt_cur",
733 xfs_btree_cur_sizeof(xfs_rmapbt_maxlevels_ondisk()),
734 0, 0, NULL);
735
736 if (!xfs_rmapbt_cur_cache)
737 return -ENOMEM;
738 return 0;
739}
740
741void
742xfs_rmapbt_destroy_cur_cache(void)
743{
744 kmem_cache_destroy(xfs_rmapbt_cur_cache);
745 xfs_rmapbt_cur_cache = NULL;
746}