1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2014 Red Hat, Inc.
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
13 #include "xfs_mount.h"
14 #include "xfs_trans.h"
15 #include "xfs_alloc.h"
16 #include "xfs_btree.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_resv.h"
27 * This is a per-ag tree used to track the owner(s) of a given extent. With
28 * reflink it is possible for there to be multiple owners, which is a departure
29 * from classic XFS. Owner records for data extents are inserted when the
30 * extent is mapped and removed when an extent is unmapped. Owner records for
31 * all other block types (i.e. metadata) are inserted when an extent is
32 * allocated and removed when an extent is freed. There can only be one owner
33 * of a metadata extent, usually an inode or some other metadata structure like
36 * The rmap btree is part of the free space management, so blocks for the tree
37 * are sourced from the agfl. Hence we need transaction reservation support for
38 * this tree so that the freelist is always large enough. This also impacts on
39 * the minimum space we need to leave free in the AG.
41 * The tree is ordered by [ag block, owner, offset]. This is a large key size,
42 * but it is the only way to enforce unique keys when a block can be owned by
43 * multiple files at any offset. There's no need to order/search by extent
44 * size for online updating/management of the tree. It is intended that most
45 * reverse lookups will be to find the owner(s) of a particular block, or to
46 * try to recover tree and file data from corrupt primary metadata.
49 static struct xfs_btree_cur *
50 xfs_rmapbt_dup_cursor(
51 struct xfs_btree_cur *cur)
53 return xfs_rmapbt_init_cursor(cur->bc_mp, cur->bc_tp,
54 cur->bc_private.a.agbp, cur->bc_private.a.agno);
59 struct xfs_btree_cur *cur,
60 union xfs_btree_ptr *ptr,
63 struct xfs_buf *agbp = cur->bc_private.a.agbp;
64 struct xfs_agf *agf = XFS_BUF_TO_AGF(agbp);
65 xfs_agnumber_t seqno = be32_to_cpu(agf->agf_seqno);
66 int btnum = cur->bc_btnum;
67 struct xfs_perag *pag = xfs_perag_get(cur->bc_mp, seqno);
71 agf->agf_roots[btnum] = ptr->s;
72 be32_add_cpu(&agf->agf_levels[btnum], inc);
73 pag->pagf_levels[btnum] += inc;
76 xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_ROOTS | XFS_AGF_LEVELS);
80 xfs_rmapbt_alloc_block(
81 struct xfs_btree_cur *cur,
82 union xfs_btree_ptr *start,
83 union xfs_btree_ptr *new,
86 struct xfs_buf *agbp = cur->bc_private.a.agbp;
87 struct xfs_agf *agf = XFS_BUF_TO_AGF(agbp);
91 /* Allocate the new block from the freelist. If we can't, give up. */
92 error = xfs_alloc_get_freelist(cur->bc_tp, cur->bc_private.a.agbp,
97 trace_xfs_rmapbt_alloc_block(cur->bc_mp, cur->bc_private.a.agno,
99 if (bno == NULLAGBLOCK) {
104 xfs_extent_busy_reuse(cur->bc_mp, cur->bc_private.a.agno, bno, 1,
107 xfs_trans_agbtree_delta(cur->bc_tp, 1);
108 new->s = cpu_to_be32(bno);
109 be32_add_cpu(&agf->agf_rmap_blocks, 1);
110 xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_RMAP_BLOCKS);
112 xfs_ag_resv_rmapbt_alloc(cur->bc_mp, cur->bc_private.a.agno);
119 xfs_rmapbt_free_block(
120 struct xfs_btree_cur *cur,
123 struct xfs_buf *agbp = cur->bc_private.a.agbp;
124 struct xfs_agf *agf = XFS_BUF_TO_AGF(agbp);
128 bno = xfs_daddr_to_agbno(cur->bc_mp, XFS_BUF_ADDR(bp));
129 trace_xfs_rmapbt_free_block(cur->bc_mp, cur->bc_private.a.agno,
131 be32_add_cpu(&agf->agf_rmap_blocks, -1);
132 xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_RMAP_BLOCKS);
133 error = xfs_alloc_put_freelist(cur->bc_tp, agbp, NULL, bno, 1);
137 xfs_extent_busy_insert(cur->bc_tp, be32_to_cpu(agf->agf_seqno), bno, 1,
138 XFS_EXTENT_BUSY_SKIP_DISCARD);
139 xfs_trans_agbtree_delta(cur->bc_tp, -1);
141 xfs_ag_resv_rmapbt_free(cur->bc_mp, cur->bc_private.a.agno);
147 xfs_rmapbt_get_minrecs(
148 struct xfs_btree_cur *cur,
151 return cur->bc_mp->m_rmap_mnr[level != 0];
155 xfs_rmapbt_get_maxrecs(
156 struct xfs_btree_cur *cur,
159 return cur->bc_mp->m_rmap_mxr[level != 0];
163 xfs_rmapbt_init_key_from_rec(
164 union xfs_btree_key *key,
165 union xfs_btree_rec *rec)
167 key->rmap.rm_startblock = rec->rmap.rm_startblock;
168 key->rmap.rm_owner = rec->rmap.rm_owner;
169 key->rmap.rm_offset = rec->rmap.rm_offset;
173 * The high key for a reverse mapping record can be computed by shifting
174 * the startblock and offset to the highest value that would still map
175 * to that record. In practice this means that we add blockcount-1 to
176 * the startblock for all records, and if the record is for a data/attr
177 * fork mapping, we add blockcount-1 to the offset too.
180 xfs_rmapbt_init_high_key_from_rec(
181 union xfs_btree_key *key,
182 union xfs_btree_rec *rec)
187 adj = be32_to_cpu(rec->rmap.rm_blockcount) - 1;
189 key->rmap.rm_startblock = rec->rmap.rm_startblock;
190 be32_add_cpu(&key->rmap.rm_startblock, adj);
191 key->rmap.rm_owner = rec->rmap.rm_owner;
192 key->rmap.rm_offset = rec->rmap.rm_offset;
193 if (XFS_RMAP_NON_INODE_OWNER(be64_to_cpu(rec->rmap.rm_owner)) ||
194 XFS_RMAP_IS_BMBT_BLOCK(be64_to_cpu(rec->rmap.rm_offset)))
196 off = be64_to_cpu(key->rmap.rm_offset);
197 off = (XFS_RMAP_OFF(off) + adj) | (off & ~XFS_RMAP_OFF_MASK);
198 key->rmap.rm_offset = cpu_to_be64(off);
202 xfs_rmapbt_init_rec_from_cur(
203 struct xfs_btree_cur *cur,
204 union xfs_btree_rec *rec)
206 rec->rmap.rm_startblock = cpu_to_be32(cur->bc_rec.r.rm_startblock);
207 rec->rmap.rm_blockcount = cpu_to_be32(cur->bc_rec.r.rm_blockcount);
208 rec->rmap.rm_owner = cpu_to_be64(cur->bc_rec.r.rm_owner);
209 rec->rmap.rm_offset = cpu_to_be64(
210 xfs_rmap_irec_offset_pack(&cur->bc_rec.r));
214 xfs_rmapbt_init_ptr_from_cur(
215 struct xfs_btree_cur *cur,
216 union xfs_btree_ptr *ptr)
218 struct xfs_agf *agf = XFS_BUF_TO_AGF(cur->bc_private.a.agbp);
220 ASSERT(cur->bc_private.a.agno == be32_to_cpu(agf->agf_seqno));
222 ptr->s = agf->agf_roots[cur->bc_btnum];
227 struct xfs_btree_cur *cur,
228 union xfs_btree_key *key)
230 struct xfs_rmap_irec *rec = &cur->bc_rec.r;
231 struct xfs_rmap_key *kp = &key->rmap;
235 d = (int64_t)be32_to_cpu(kp->rm_startblock) - rec->rm_startblock;
239 x = be64_to_cpu(kp->rm_owner);
246 x = XFS_RMAP_OFF(be64_to_cpu(kp->rm_offset));
256 xfs_rmapbt_diff_two_keys(
257 struct xfs_btree_cur *cur,
258 union xfs_btree_key *k1,
259 union xfs_btree_key *k2)
261 struct xfs_rmap_key *kp1 = &k1->rmap;
262 struct xfs_rmap_key *kp2 = &k2->rmap;
266 d = (int64_t)be32_to_cpu(kp1->rm_startblock) -
267 be32_to_cpu(kp2->rm_startblock);
271 x = be64_to_cpu(kp1->rm_owner);
272 y = be64_to_cpu(kp2->rm_owner);
278 x = XFS_RMAP_OFF(be64_to_cpu(kp1->rm_offset));
279 y = XFS_RMAP_OFF(be64_to_cpu(kp2->rm_offset));
287 static xfs_failaddr_t
291 struct xfs_mount *mp = bp->b_mount;
292 struct xfs_btree_block *block = XFS_BUF_TO_BLOCK(bp);
293 struct xfs_perag *pag = bp->b_pag;
298 * magic number and level verification
300 * During growfs operations, we can't verify the exact level or owner as
301 * the perag is not fully initialised and hence not attached to the
302 * buffer. In this case, check against the maximum tree depth.
304 * Similarly, during log recovery we will have a perag structure
305 * attached, but the agf information will not yet have been initialised
306 * from the on disk AGF. Again, we can only check against maximum limits
309 if (!xfs_verify_magic(bp, block->bb_magic))
310 return __this_address;
312 if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
313 return __this_address;
314 fa = xfs_btree_sblock_v5hdr_verify(bp);
318 level = be16_to_cpu(block->bb_level);
319 if (pag && pag->pagf_init) {
320 if (level >= pag->pagf_levels[XFS_BTNUM_RMAPi])
321 return __this_address;
322 } else if (level >= mp->m_rmap_maxlevels)
323 return __this_address;
325 return xfs_btree_sblock_verify(bp, mp->m_rmap_mxr[level != 0]);
329 xfs_rmapbt_read_verify(
334 if (!xfs_btree_sblock_verify_crc(bp))
335 xfs_verifier_error(bp, -EFSBADCRC, __this_address);
337 fa = xfs_rmapbt_verify(bp);
339 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
343 trace_xfs_btree_corrupt(bp, _RET_IP_);
347 xfs_rmapbt_write_verify(
352 fa = xfs_rmapbt_verify(bp);
354 trace_xfs_btree_corrupt(bp, _RET_IP_);
355 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
358 xfs_btree_sblock_calc_crc(bp);
362 const struct xfs_buf_ops xfs_rmapbt_buf_ops = {
363 .name = "xfs_rmapbt",
364 .magic = { 0, cpu_to_be32(XFS_RMAP_CRC_MAGIC) },
365 .verify_read = xfs_rmapbt_read_verify,
366 .verify_write = xfs_rmapbt_write_verify,
367 .verify_struct = xfs_rmapbt_verify,
371 xfs_rmapbt_keys_inorder(
372 struct xfs_btree_cur *cur,
373 union xfs_btree_key *k1,
374 union xfs_btree_key *k2)
381 x = be32_to_cpu(k1->rmap.rm_startblock);
382 y = be32_to_cpu(k2->rmap.rm_startblock);
387 a = be64_to_cpu(k1->rmap.rm_owner);
388 b = be64_to_cpu(k2->rmap.rm_owner);
393 a = XFS_RMAP_OFF(be64_to_cpu(k1->rmap.rm_offset));
394 b = XFS_RMAP_OFF(be64_to_cpu(k2->rmap.rm_offset));
401 xfs_rmapbt_recs_inorder(
402 struct xfs_btree_cur *cur,
403 union xfs_btree_rec *r1,
404 union xfs_btree_rec *r2)
411 x = be32_to_cpu(r1->rmap.rm_startblock);
412 y = be32_to_cpu(r2->rmap.rm_startblock);
417 a = be64_to_cpu(r1->rmap.rm_owner);
418 b = be64_to_cpu(r2->rmap.rm_owner);
423 a = XFS_RMAP_OFF(be64_to_cpu(r1->rmap.rm_offset));
424 b = XFS_RMAP_OFF(be64_to_cpu(r2->rmap.rm_offset));
430 static const struct xfs_btree_ops xfs_rmapbt_ops = {
431 .rec_len = sizeof(struct xfs_rmap_rec),
432 .key_len = 2 * sizeof(struct xfs_rmap_key),
434 .dup_cursor = xfs_rmapbt_dup_cursor,
435 .set_root = xfs_rmapbt_set_root,
436 .alloc_block = xfs_rmapbt_alloc_block,
437 .free_block = xfs_rmapbt_free_block,
438 .get_minrecs = xfs_rmapbt_get_minrecs,
439 .get_maxrecs = xfs_rmapbt_get_maxrecs,
440 .init_key_from_rec = xfs_rmapbt_init_key_from_rec,
441 .init_high_key_from_rec = xfs_rmapbt_init_high_key_from_rec,
442 .init_rec_from_cur = xfs_rmapbt_init_rec_from_cur,
443 .init_ptr_from_cur = xfs_rmapbt_init_ptr_from_cur,
444 .key_diff = xfs_rmapbt_key_diff,
445 .buf_ops = &xfs_rmapbt_buf_ops,
446 .diff_two_keys = xfs_rmapbt_diff_two_keys,
447 .keys_inorder = xfs_rmapbt_keys_inorder,
448 .recs_inorder = xfs_rmapbt_recs_inorder,
452 * Allocate a new allocation btree cursor.
454 struct xfs_btree_cur *
455 xfs_rmapbt_init_cursor(
456 struct xfs_mount *mp,
457 struct xfs_trans *tp,
458 struct xfs_buf *agbp,
461 struct xfs_agf *agf = XFS_BUF_TO_AGF(agbp);
462 struct xfs_btree_cur *cur;
464 cur = kmem_zone_zalloc(xfs_btree_cur_zone, KM_NOFS);
467 /* Overlapping btree; 2 keys per pointer. */
468 cur->bc_btnum = XFS_BTNUM_RMAP;
469 cur->bc_flags = XFS_BTREE_CRC_BLOCKS | XFS_BTREE_OVERLAPPING;
470 cur->bc_blocklog = mp->m_sb.sb_blocklog;
471 cur->bc_ops = &xfs_rmapbt_ops;
472 cur->bc_nlevels = be32_to_cpu(agf->agf_levels[XFS_BTNUM_RMAP]);
473 cur->bc_statoff = XFS_STATS_CALC_INDEX(xs_rmap_2);
475 cur->bc_private.a.agbp = agbp;
476 cur->bc_private.a.agno = agno;
482 * Calculate number of records in an rmap btree block.
489 blocklen -= XFS_RMAP_BLOCK_LEN;
492 return blocklen / sizeof(struct xfs_rmap_rec);
494 (2 * sizeof(struct xfs_rmap_key) + sizeof(xfs_rmap_ptr_t));
497 /* Compute the maximum height of an rmap btree. */
499 xfs_rmapbt_compute_maxlevels(
500 struct xfs_mount *mp)
503 * On a non-reflink filesystem, the maximum number of rmap
504 * records is the number of blocks in the AG, hence the max
505 * rmapbt height is log_$maxrecs($agblocks). However, with
506 * reflink each AG block can have up to 2^32 (per the refcount
507 * record format) owners, which means that theoretically we
508 * could face up to 2^64 rmap records.
510 * That effectively means that the max rmapbt height must be
511 * XFS_BTREE_MAXLEVELS. "Fortunately" we'll run out of AG
512 * blocks to feed the rmapbt long before the rmapbt reaches
513 * maximum height. The reflink code uses ag_resv_critical to
514 * disallow reflinking when less than 10% of the per-AG metadata
515 * block reservation since the fallback is a regular file copy.
517 if (xfs_sb_version_hasreflink(&mp->m_sb))
518 mp->m_rmap_maxlevels = XFS_BTREE_MAXLEVELS;
520 mp->m_rmap_maxlevels = xfs_btree_compute_maxlevels(
521 mp->m_rmap_mnr, mp->m_sb.sb_agblocks);
524 /* Calculate the refcount btree size for some records. */
526 xfs_rmapbt_calc_size(
527 struct xfs_mount *mp,
528 unsigned long long len)
530 return xfs_btree_calc_size(mp->m_rmap_mnr, len);
534 * Calculate the maximum refcount btree size.
538 struct xfs_mount *mp,
539 xfs_agblock_t agblocks)
541 /* Bail out if we're uninitialized, which can happen in mkfs. */
542 if (mp->m_rmap_mxr[0] == 0)
545 return xfs_rmapbt_calc_size(mp, agblocks);
549 * Figure out how many blocks to reserve and how many are used by this btree.
552 xfs_rmapbt_calc_reserves(
553 struct xfs_mount *mp,
554 struct xfs_trans *tp,
559 struct xfs_buf *agbp;
561 xfs_agblock_t agblocks;
562 xfs_extlen_t tree_len;
565 if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
568 error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
572 agf = XFS_BUF_TO_AGF(agbp);
573 agblocks = be32_to_cpu(agf->agf_length);
574 tree_len = be32_to_cpu(agf->agf_rmap_blocks);
575 xfs_trans_brelse(tp, agbp);
578 * The log is permanently allocated, so the space it occupies will
579 * never be available for the kinds of things that would require btree
580 * expansion. We therefore can pretend the space isn't there.
582 if (mp->m_sb.sb_logstart &&
583 XFS_FSB_TO_AGNO(mp, mp->m_sb.sb_logstart) == agno)
584 agblocks -= mp->m_sb.sb_logblocks;
586 /* Reserve 1% of the AG or enough for 1 block per record. */
587 *ask += max(agblocks / 100, xfs_rmapbt_max_size(mp, agblocks));