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"
14 #include "xfs_mount.h"
15 #include "xfs_defer.h"
16 #include "xfs_inode.h"
17 #include "xfs_trans.h"
18 #include "xfs_alloc.h"
19 #include "xfs_btree.h"
21 #include "xfs_rmap_btree.h"
22 #include "xfs_trace.h"
23 #include "xfs_cksum.h"
24 #include "xfs_error.h"
25 #include "xfs_extent_busy.h"
26 #include "xfs_ag_resv.h"
31 * This is a per-ag tree used to track the owner(s) of a given extent. With
32 * reflink it is possible for there to be multiple owners, which is a departure
33 * from classic XFS. Owner records for data extents are inserted when the
34 * extent is mapped and removed when an extent is unmapped. Owner records for
35 * all other block types (i.e. metadata) are inserted when an extent is
36 * allocated and removed when an extent is freed. There can only be one owner
37 * of a metadata extent, usually an inode or some other metadata structure like
40 * The rmap btree is part of the free space management, so blocks for the tree
41 * are sourced from the agfl. Hence we need transaction reservation support for
42 * this tree so that the freelist is always large enough. This also impacts on
43 * the minimum space we need to leave free in the AG.
45 * The tree is ordered by [ag block, owner, offset]. This is a large key size,
46 * but it is the only way to enforce unique keys when a block can be owned by
47 * multiple files at any offset. There's no need to order/search by extent
48 * size for online updating/management of the tree. It is intended that most
49 * reverse lookups will be to find the owner(s) of a particular block, or to
50 * try to recover tree and file data from corrupt primary metadata.
53 static struct xfs_btree_cur *
54 xfs_rmapbt_dup_cursor(
55 struct xfs_btree_cur *cur)
57 return xfs_rmapbt_init_cursor(cur->bc_mp, cur->bc_tp,
58 cur->bc_private.a.agbp, cur->bc_private.a.agno);
63 struct xfs_btree_cur *cur,
64 union xfs_btree_ptr *ptr,
67 struct xfs_buf *agbp = cur->bc_private.a.agbp;
68 struct xfs_agf *agf = XFS_BUF_TO_AGF(agbp);
69 xfs_agnumber_t seqno = be32_to_cpu(agf->agf_seqno);
70 int btnum = cur->bc_btnum;
71 struct xfs_perag *pag = xfs_perag_get(cur->bc_mp, seqno);
75 agf->agf_roots[btnum] = ptr->s;
76 be32_add_cpu(&agf->agf_levels[btnum], inc);
77 pag->pagf_levels[btnum] += inc;
80 xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_ROOTS | XFS_AGF_LEVELS);
84 xfs_rmapbt_alloc_block(
85 struct xfs_btree_cur *cur,
86 union xfs_btree_ptr *start,
87 union xfs_btree_ptr *new,
90 struct xfs_buf *agbp = cur->bc_private.a.agbp;
91 struct xfs_agf *agf = XFS_BUF_TO_AGF(agbp);
95 /* Allocate the new block from the freelist. If we can't, give up. */
96 error = xfs_alloc_get_freelist(cur->bc_tp, cur->bc_private.a.agbp,
101 trace_xfs_rmapbt_alloc_block(cur->bc_mp, cur->bc_private.a.agno,
103 if (bno == NULLAGBLOCK) {
108 xfs_extent_busy_reuse(cur->bc_mp, cur->bc_private.a.agno, bno, 1,
111 xfs_trans_agbtree_delta(cur->bc_tp, 1);
112 new->s = cpu_to_be32(bno);
113 be32_add_cpu(&agf->agf_rmap_blocks, 1);
114 xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_RMAP_BLOCKS);
116 xfs_ag_resv_rmapbt_alloc(cur->bc_mp, cur->bc_private.a.agno);
123 xfs_rmapbt_free_block(
124 struct xfs_btree_cur *cur,
127 struct xfs_buf *agbp = cur->bc_private.a.agbp;
128 struct xfs_agf *agf = XFS_BUF_TO_AGF(agbp);
132 bno = xfs_daddr_to_agbno(cur->bc_mp, XFS_BUF_ADDR(bp));
133 trace_xfs_rmapbt_free_block(cur->bc_mp, cur->bc_private.a.agno,
135 be32_add_cpu(&agf->agf_rmap_blocks, -1);
136 xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_RMAP_BLOCKS);
137 error = xfs_alloc_put_freelist(cur->bc_tp, agbp, NULL, bno, 1);
141 xfs_extent_busy_insert(cur->bc_tp, be32_to_cpu(agf->agf_seqno), bno, 1,
142 XFS_EXTENT_BUSY_SKIP_DISCARD);
143 xfs_trans_agbtree_delta(cur->bc_tp, -1);
145 xfs_ag_resv_rmapbt_free(cur->bc_mp, cur->bc_private.a.agno);
151 xfs_rmapbt_get_minrecs(
152 struct xfs_btree_cur *cur,
155 return cur->bc_mp->m_rmap_mnr[level != 0];
159 xfs_rmapbt_get_maxrecs(
160 struct xfs_btree_cur *cur,
163 return cur->bc_mp->m_rmap_mxr[level != 0];
167 xfs_rmapbt_init_key_from_rec(
168 union xfs_btree_key *key,
169 union xfs_btree_rec *rec)
171 key->rmap.rm_startblock = rec->rmap.rm_startblock;
172 key->rmap.rm_owner = rec->rmap.rm_owner;
173 key->rmap.rm_offset = rec->rmap.rm_offset;
177 * The high key for a reverse mapping record can be computed by shifting
178 * the startblock and offset to the highest value that would still map
179 * to that record. In practice this means that we add blockcount-1 to
180 * the startblock for all records, and if the record is for a data/attr
181 * fork mapping, we add blockcount-1 to the offset too.
184 xfs_rmapbt_init_high_key_from_rec(
185 union xfs_btree_key *key,
186 union xfs_btree_rec *rec)
191 adj = be32_to_cpu(rec->rmap.rm_blockcount) - 1;
193 key->rmap.rm_startblock = rec->rmap.rm_startblock;
194 be32_add_cpu(&key->rmap.rm_startblock, adj);
195 key->rmap.rm_owner = rec->rmap.rm_owner;
196 key->rmap.rm_offset = rec->rmap.rm_offset;
197 if (XFS_RMAP_NON_INODE_OWNER(be64_to_cpu(rec->rmap.rm_owner)) ||
198 XFS_RMAP_IS_BMBT_BLOCK(be64_to_cpu(rec->rmap.rm_offset)))
200 off = be64_to_cpu(key->rmap.rm_offset);
201 off = (XFS_RMAP_OFF(off) + adj) | (off & ~XFS_RMAP_OFF_MASK);
202 key->rmap.rm_offset = cpu_to_be64(off);
206 xfs_rmapbt_init_rec_from_cur(
207 struct xfs_btree_cur *cur,
208 union xfs_btree_rec *rec)
210 rec->rmap.rm_startblock = cpu_to_be32(cur->bc_rec.r.rm_startblock);
211 rec->rmap.rm_blockcount = cpu_to_be32(cur->bc_rec.r.rm_blockcount);
212 rec->rmap.rm_owner = cpu_to_be64(cur->bc_rec.r.rm_owner);
213 rec->rmap.rm_offset = cpu_to_be64(
214 xfs_rmap_irec_offset_pack(&cur->bc_rec.r));
218 xfs_rmapbt_init_ptr_from_cur(
219 struct xfs_btree_cur *cur,
220 union xfs_btree_ptr *ptr)
222 struct xfs_agf *agf = XFS_BUF_TO_AGF(cur->bc_private.a.agbp);
224 ASSERT(cur->bc_private.a.agno == be32_to_cpu(agf->agf_seqno));
226 ptr->s = agf->agf_roots[cur->bc_btnum];
231 struct xfs_btree_cur *cur,
232 union xfs_btree_key *key)
234 struct xfs_rmap_irec *rec = &cur->bc_rec.r;
235 struct xfs_rmap_key *kp = &key->rmap;
239 d = (int64_t)be32_to_cpu(kp->rm_startblock) - rec->rm_startblock;
243 x = be64_to_cpu(kp->rm_owner);
250 x = XFS_RMAP_OFF(be64_to_cpu(kp->rm_offset));
260 xfs_rmapbt_diff_two_keys(
261 struct xfs_btree_cur *cur,
262 union xfs_btree_key *k1,
263 union xfs_btree_key *k2)
265 struct xfs_rmap_key *kp1 = &k1->rmap;
266 struct xfs_rmap_key *kp2 = &k2->rmap;
270 d = (int64_t)be32_to_cpu(kp1->rm_startblock) -
271 be32_to_cpu(kp2->rm_startblock);
275 x = be64_to_cpu(kp1->rm_owner);
276 y = be64_to_cpu(kp2->rm_owner);
282 x = XFS_RMAP_OFF(be64_to_cpu(kp1->rm_offset));
283 y = XFS_RMAP_OFF(be64_to_cpu(kp2->rm_offset));
291 static xfs_failaddr_t
295 struct xfs_mount *mp = bp->b_target->bt_mount;
296 struct xfs_btree_block *block = XFS_BUF_TO_BLOCK(bp);
297 struct xfs_perag *pag = bp->b_pag;
302 * magic number and level verification
304 * During growfs operations, we can't verify the exact level or owner as
305 * the perag is not fully initialised and hence not attached to the
306 * buffer. In this case, check against the maximum tree depth.
308 * Similarly, during log recovery we will have a perag structure
309 * attached, but the agf information will not yet have been initialised
310 * from the on disk AGF. Again, we can only check against maximum limits
313 if (!xfs_verify_magic(bp, block->bb_magic))
314 return __this_address;
316 if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
317 return __this_address;
318 fa = xfs_btree_sblock_v5hdr_verify(bp);
322 level = be16_to_cpu(block->bb_level);
323 if (pag && pag->pagf_init) {
324 if (level >= pag->pagf_levels[XFS_BTNUM_RMAPi])
325 return __this_address;
326 } else if (level >= mp->m_rmap_maxlevels)
327 return __this_address;
329 return xfs_btree_sblock_verify(bp, mp->m_rmap_mxr[level != 0]);
333 xfs_rmapbt_read_verify(
338 if (!xfs_btree_sblock_verify_crc(bp))
339 xfs_verifier_error(bp, -EFSBADCRC, __this_address);
341 fa = xfs_rmapbt_verify(bp);
343 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
347 trace_xfs_btree_corrupt(bp, _RET_IP_);
351 xfs_rmapbt_write_verify(
356 fa = xfs_rmapbt_verify(bp);
358 trace_xfs_btree_corrupt(bp, _RET_IP_);
359 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
362 xfs_btree_sblock_calc_crc(bp);
366 const struct xfs_buf_ops xfs_rmapbt_buf_ops = {
367 .name = "xfs_rmapbt",
368 .magic = { 0, cpu_to_be32(XFS_RMAP_CRC_MAGIC) },
369 .verify_read = xfs_rmapbt_read_verify,
370 .verify_write = xfs_rmapbt_write_verify,
371 .verify_struct = xfs_rmapbt_verify,
375 xfs_rmapbt_keys_inorder(
376 struct xfs_btree_cur *cur,
377 union xfs_btree_key *k1,
378 union xfs_btree_key *k2)
385 x = be32_to_cpu(k1->rmap.rm_startblock);
386 y = be32_to_cpu(k2->rmap.rm_startblock);
391 a = be64_to_cpu(k1->rmap.rm_owner);
392 b = be64_to_cpu(k2->rmap.rm_owner);
397 a = XFS_RMAP_OFF(be64_to_cpu(k1->rmap.rm_offset));
398 b = XFS_RMAP_OFF(be64_to_cpu(k2->rmap.rm_offset));
405 xfs_rmapbt_recs_inorder(
406 struct xfs_btree_cur *cur,
407 union xfs_btree_rec *r1,
408 union xfs_btree_rec *r2)
415 x = be32_to_cpu(r1->rmap.rm_startblock);
416 y = be32_to_cpu(r2->rmap.rm_startblock);
421 a = be64_to_cpu(r1->rmap.rm_owner);
422 b = be64_to_cpu(r2->rmap.rm_owner);
427 a = XFS_RMAP_OFF(be64_to_cpu(r1->rmap.rm_offset));
428 b = XFS_RMAP_OFF(be64_to_cpu(r2->rmap.rm_offset));
434 static const struct xfs_btree_ops xfs_rmapbt_ops = {
435 .rec_len = sizeof(struct xfs_rmap_rec),
436 .key_len = 2 * sizeof(struct xfs_rmap_key),
438 .dup_cursor = xfs_rmapbt_dup_cursor,
439 .set_root = xfs_rmapbt_set_root,
440 .alloc_block = xfs_rmapbt_alloc_block,
441 .free_block = xfs_rmapbt_free_block,
442 .get_minrecs = xfs_rmapbt_get_minrecs,
443 .get_maxrecs = xfs_rmapbt_get_maxrecs,
444 .init_key_from_rec = xfs_rmapbt_init_key_from_rec,
445 .init_high_key_from_rec = xfs_rmapbt_init_high_key_from_rec,
446 .init_rec_from_cur = xfs_rmapbt_init_rec_from_cur,
447 .init_ptr_from_cur = xfs_rmapbt_init_ptr_from_cur,
448 .key_diff = xfs_rmapbt_key_diff,
449 .buf_ops = &xfs_rmapbt_buf_ops,
450 .diff_two_keys = xfs_rmapbt_diff_two_keys,
451 .keys_inorder = xfs_rmapbt_keys_inorder,
452 .recs_inorder = xfs_rmapbt_recs_inorder,
456 * Allocate a new allocation btree cursor.
458 struct xfs_btree_cur *
459 xfs_rmapbt_init_cursor(
460 struct xfs_mount *mp,
461 struct xfs_trans *tp,
462 struct xfs_buf *agbp,
465 struct xfs_agf *agf = XFS_BUF_TO_AGF(agbp);
466 struct xfs_btree_cur *cur;
468 cur = kmem_zone_zalloc(xfs_btree_cur_zone, KM_NOFS);
471 /* Overlapping btree; 2 keys per pointer. */
472 cur->bc_btnum = XFS_BTNUM_RMAP;
473 cur->bc_flags = XFS_BTREE_CRC_BLOCKS | XFS_BTREE_OVERLAPPING;
474 cur->bc_blocklog = mp->m_sb.sb_blocklog;
475 cur->bc_ops = &xfs_rmapbt_ops;
476 cur->bc_nlevels = be32_to_cpu(agf->agf_levels[XFS_BTNUM_RMAP]);
477 cur->bc_statoff = XFS_STATS_CALC_INDEX(xs_rmap_2);
479 cur->bc_private.a.agbp = agbp;
480 cur->bc_private.a.agno = agno;
486 * Calculate number of records in an rmap btree block.
493 blocklen -= XFS_RMAP_BLOCK_LEN;
496 return blocklen / sizeof(struct xfs_rmap_rec);
498 (2 * sizeof(struct xfs_rmap_key) + sizeof(xfs_rmap_ptr_t));
501 /* Compute the maximum height of an rmap btree. */
503 xfs_rmapbt_compute_maxlevels(
504 struct xfs_mount *mp)
507 * On a non-reflink filesystem, the maximum number of rmap
508 * records is the number of blocks in the AG, hence the max
509 * rmapbt height is log_$maxrecs($agblocks). However, with
510 * reflink each AG block can have up to 2^32 (per the refcount
511 * record format) owners, which means that theoretically we
512 * could face up to 2^64 rmap records.
514 * That effectively means that the max rmapbt height must be
515 * XFS_BTREE_MAXLEVELS. "Fortunately" we'll run out of AG
516 * blocks to feed the rmapbt long before the rmapbt reaches
517 * maximum height. The reflink code uses ag_resv_critical to
518 * disallow reflinking when less than 10% of the per-AG metadata
519 * block reservation since the fallback is a regular file copy.
521 if (xfs_sb_version_hasreflink(&mp->m_sb))
522 mp->m_rmap_maxlevels = XFS_BTREE_MAXLEVELS;
524 mp->m_rmap_maxlevels = xfs_btree_compute_maxlevels(
525 mp->m_rmap_mnr, mp->m_sb.sb_agblocks);
528 /* Calculate the refcount btree size for some records. */
530 xfs_rmapbt_calc_size(
531 struct xfs_mount *mp,
532 unsigned long long len)
534 return xfs_btree_calc_size(mp->m_rmap_mnr, len);
538 * Calculate the maximum refcount btree size.
542 struct xfs_mount *mp,
543 xfs_agblock_t agblocks)
545 /* Bail out if we're uninitialized, which can happen in mkfs. */
546 if (mp->m_rmap_mxr[0] == 0)
549 return xfs_rmapbt_calc_size(mp, agblocks);
553 * Figure out how many blocks to reserve and how many are used by this btree.
556 xfs_rmapbt_calc_reserves(
557 struct xfs_mount *mp,
558 struct xfs_trans *tp,
563 struct xfs_buf *agbp;
565 xfs_agblock_t agblocks;
566 xfs_extlen_t tree_len;
569 if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
572 error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
576 agf = XFS_BUF_TO_AGF(agbp);
577 agblocks = be32_to_cpu(agf->agf_length);
578 tree_len = be32_to_cpu(agf->agf_rmap_blocks);
579 xfs_trans_brelse(tp, agbp);
581 /* Reserve 1% of the AG or enough for 1 block per record. */
582 *ask += max(agblocks / 100, xfs_rmapbt_max_size(mp, agblocks));