1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2014 Red Hat, Inc.
6 #include "libxfs_priv.h"
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_ag_resv.h"
29 * This is a per-ag tree used to track the owner(s) of a given extent. With
30 * reflink it is possible for there to be multiple owners, which is a departure
31 * from classic XFS. Owner records for data extents are inserted when the
32 * extent is mapped and removed when an extent is unmapped. Owner records for
33 * all other block types (i.e. metadata) are inserted when an extent is
34 * allocated and removed when an extent is freed. There can only be one owner
35 * of a metadata extent, usually an inode or some other metadata structure like
38 * The rmap btree is part of the free space management, so blocks for the tree
39 * are sourced from the agfl. Hence we need transaction reservation support for
40 * this tree so that the freelist is always large enough. This also impacts on
41 * the minimum space we need to leave free in the AG.
43 * The tree is ordered by [ag block, owner, offset]. This is a large key size,
44 * but it is the only way to enforce unique keys when a block can be owned by
45 * multiple files at any offset. There's no need to order/search by extent
46 * size for online updating/management of the tree. It is intended that most
47 * reverse lookups will be to find the owner(s) of a particular block, or to
48 * try to recover tree and file data from corrupt primary metadata.
51 static struct xfs_btree_cur
*
52 xfs_rmapbt_dup_cursor(
53 struct xfs_btree_cur
*cur
)
55 return xfs_rmapbt_init_cursor(cur
->bc_mp
, cur
->bc_tp
,
56 cur
->bc_private
.a
.agbp
, cur
->bc_private
.a
.agno
);
61 struct xfs_btree_cur
*cur
,
62 union xfs_btree_ptr
*ptr
,
65 struct xfs_buf
*agbp
= cur
->bc_private
.a
.agbp
;
66 struct xfs_agf
*agf
= XFS_BUF_TO_AGF(agbp
);
67 xfs_agnumber_t seqno
= be32_to_cpu(agf
->agf_seqno
);
68 int btnum
= cur
->bc_btnum
;
69 struct xfs_perag
*pag
= xfs_perag_get(cur
->bc_mp
, seqno
);
73 agf
->agf_roots
[btnum
] = ptr
->s
;
74 be32_add_cpu(&agf
->agf_levels
[btnum
], inc
);
75 pag
->pagf_levels
[btnum
] += inc
;
78 xfs_alloc_log_agf(cur
->bc_tp
, agbp
, XFS_AGF_ROOTS
| XFS_AGF_LEVELS
);
82 xfs_rmapbt_alloc_block(
83 struct xfs_btree_cur
*cur
,
84 union xfs_btree_ptr
*start
,
85 union xfs_btree_ptr
*new,
88 struct xfs_buf
*agbp
= cur
->bc_private
.a
.agbp
;
89 struct xfs_agf
*agf
= XFS_BUF_TO_AGF(agbp
);
93 /* Allocate the new block from the freelist. If we can't, give up. */
94 error
= xfs_alloc_get_freelist(cur
->bc_tp
, cur
->bc_private
.a
.agbp
,
99 trace_xfs_rmapbt_alloc_block(cur
->bc_mp
, cur
->bc_private
.a
.agno
,
101 if (bno
== NULLAGBLOCK
) {
106 xfs_extent_busy_reuse(cur
->bc_mp
, cur
->bc_private
.a
.agno
, bno
, 1,
109 xfs_trans_agbtree_delta(cur
->bc_tp
, 1);
110 new->s
= cpu_to_be32(bno
);
111 be32_add_cpu(&agf
->agf_rmap_blocks
, 1);
112 xfs_alloc_log_agf(cur
->bc_tp
, agbp
, XFS_AGF_RMAP_BLOCKS
);
114 xfs_ag_resv_rmapbt_alloc(cur
->bc_mp
, cur
->bc_private
.a
.agno
);
121 xfs_rmapbt_free_block(
122 struct xfs_btree_cur
*cur
,
125 struct xfs_buf
*agbp
= cur
->bc_private
.a
.agbp
;
126 struct xfs_agf
*agf
= XFS_BUF_TO_AGF(agbp
);
130 bno
= xfs_daddr_to_agbno(cur
->bc_mp
, XFS_BUF_ADDR(bp
));
131 trace_xfs_rmapbt_free_block(cur
->bc_mp
, cur
->bc_private
.a
.agno
,
133 be32_add_cpu(&agf
->agf_rmap_blocks
, -1);
134 xfs_alloc_log_agf(cur
->bc_tp
, agbp
, XFS_AGF_RMAP_BLOCKS
);
135 error
= xfs_alloc_put_freelist(cur
->bc_tp
, agbp
, NULL
, bno
, 1);
139 xfs_extent_busy_insert(cur
->bc_tp
, be32_to_cpu(agf
->agf_seqno
), bno
, 1,
140 XFS_EXTENT_BUSY_SKIP_DISCARD
);
141 xfs_trans_agbtree_delta(cur
->bc_tp
, -1);
143 xfs_ag_resv_rmapbt_free(cur
->bc_mp
, cur
->bc_private
.a
.agno
);
149 xfs_rmapbt_get_minrecs(
150 struct xfs_btree_cur
*cur
,
153 return cur
->bc_mp
->m_rmap_mnr
[level
!= 0];
157 xfs_rmapbt_get_maxrecs(
158 struct xfs_btree_cur
*cur
,
161 return cur
->bc_mp
->m_rmap_mxr
[level
!= 0];
165 xfs_rmapbt_init_key_from_rec(
166 union xfs_btree_key
*key
,
167 union xfs_btree_rec
*rec
)
169 key
->rmap
.rm_startblock
= rec
->rmap
.rm_startblock
;
170 key
->rmap
.rm_owner
= rec
->rmap
.rm_owner
;
171 key
->rmap
.rm_offset
= rec
->rmap
.rm_offset
;
175 * The high key for a reverse mapping record can be computed by shifting
176 * the startblock and offset to the highest value that would still map
177 * to that record. In practice this means that we add blockcount-1 to
178 * the startblock for all records, and if the record is for a data/attr
179 * fork mapping, we add blockcount-1 to the offset too.
182 xfs_rmapbt_init_high_key_from_rec(
183 union xfs_btree_key
*key
,
184 union xfs_btree_rec
*rec
)
189 adj
= be32_to_cpu(rec
->rmap
.rm_blockcount
) - 1;
191 key
->rmap
.rm_startblock
= rec
->rmap
.rm_startblock
;
192 be32_add_cpu(&key
->rmap
.rm_startblock
, adj
);
193 key
->rmap
.rm_owner
= rec
->rmap
.rm_owner
;
194 key
->rmap
.rm_offset
= rec
->rmap
.rm_offset
;
195 if (XFS_RMAP_NON_INODE_OWNER(be64_to_cpu(rec
->rmap
.rm_owner
)) ||
196 XFS_RMAP_IS_BMBT_BLOCK(be64_to_cpu(rec
->rmap
.rm_offset
)))
198 off
= be64_to_cpu(key
->rmap
.rm_offset
);
199 off
= (XFS_RMAP_OFF(off
) + adj
) | (off
& ~XFS_RMAP_OFF_MASK
);
200 key
->rmap
.rm_offset
= cpu_to_be64(off
);
204 xfs_rmapbt_init_rec_from_cur(
205 struct xfs_btree_cur
*cur
,
206 union xfs_btree_rec
*rec
)
208 rec
->rmap
.rm_startblock
= cpu_to_be32(cur
->bc_rec
.r
.rm_startblock
);
209 rec
->rmap
.rm_blockcount
= cpu_to_be32(cur
->bc_rec
.r
.rm_blockcount
);
210 rec
->rmap
.rm_owner
= cpu_to_be64(cur
->bc_rec
.r
.rm_owner
);
211 rec
->rmap
.rm_offset
= cpu_to_be64(
212 xfs_rmap_irec_offset_pack(&cur
->bc_rec
.r
));
216 xfs_rmapbt_init_ptr_from_cur(
217 struct xfs_btree_cur
*cur
,
218 union xfs_btree_ptr
*ptr
)
220 struct xfs_agf
*agf
= XFS_BUF_TO_AGF(cur
->bc_private
.a
.agbp
);
222 ASSERT(cur
->bc_private
.a
.agno
== be32_to_cpu(agf
->agf_seqno
));
224 ptr
->s
= agf
->agf_roots
[cur
->bc_btnum
];
229 struct xfs_btree_cur
*cur
,
230 union xfs_btree_key
*key
)
232 struct xfs_rmap_irec
*rec
= &cur
->bc_rec
.r
;
233 struct xfs_rmap_key
*kp
= &key
->rmap
;
237 d
= (int64_t)be32_to_cpu(kp
->rm_startblock
) - rec
->rm_startblock
;
241 x
= be64_to_cpu(kp
->rm_owner
);
248 x
= XFS_RMAP_OFF(be64_to_cpu(kp
->rm_offset
));
258 xfs_rmapbt_diff_two_keys(
259 struct xfs_btree_cur
*cur
,
260 union xfs_btree_key
*k1
,
261 union xfs_btree_key
*k2
)
263 struct xfs_rmap_key
*kp1
= &k1
->rmap
;
264 struct xfs_rmap_key
*kp2
= &k2
->rmap
;
268 d
= (int64_t)be32_to_cpu(kp1
->rm_startblock
) -
269 be32_to_cpu(kp2
->rm_startblock
);
273 x
= be64_to_cpu(kp1
->rm_owner
);
274 y
= be64_to_cpu(kp2
->rm_owner
);
280 x
= XFS_RMAP_OFF(be64_to_cpu(kp1
->rm_offset
));
281 y
= XFS_RMAP_OFF(be64_to_cpu(kp2
->rm_offset
));
289 static xfs_failaddr_t
293 struct xfs_mount
*mp
= bp
->b_target
->bt_mount
;
294 struct xfs_btree_block
*block
= XFS_BUF_TO_BLOCK(bp
);
295 struct xfs_perag
*pag
= bp
->b_pag
;
300 * magic number and level verification
302 * During growfs operations, we can't verify the exact level or owner as
303 * the perag is not fully initialised and hence not attached to the
304 * buffer. In this case, check against the maximum tree depth.
306 * Similarly, during log recovery we will have a perag structure
307 * attached, but the agf information will not yet have been initialised
308 * from the on disk AGF. Again, we can only check against maximum limits
311 if (block
->bb_magic
!= cpu_to_be32(XFS_RMAP_CRC_MAGIC
))
312 return __this_address
;
314 if (!xfs_sb_version_hasrmapbt(&mp
->m_sb
))
315 return __this_address
;
316 fa
= xfs_btree_sblock_v5hdr_verify(bp
);
320 level
= be16_to_cpu(block
->bb_level
);
321 if (pag
&& pag
->pagf_init
) {
322 if (level
>= pag
->pagf_levels
[XFS_BTNUM_RMAPi
])
323 return __this_address
;
324 } else if (level
>= mp
->m_rmap_maxlevels
)
325 return __this_address
;
327 return xfs_btree_sblock_verify(bp
, mp
->m_rmap_mxr
[level
!= 0]);
331 xfs_rmapbt_read_verify(
336 if (!xfs_btree_sblock_verify_crc(bp
))
337 xfs_verifier_error(bp
, -EFSBADCRC
, __this_address
);
339 fa
= xfs_rmapbt_verify(bp
);
341 xfs_verifier_error(bp
, -EFSCORRUPTED
, fa
);
345 trace_xfs_btree_corrupt(bp
, _RET_IP_
);
349 xfs_rmapbt_write_verify(
354 fa
= xfs_rmapbt_verify(bp
);
356 trace_xfs_btree_corrupt(bp
, _RET_IP_
);
357 xfs_verifier_error(bp
, -EFSCORRUPTED
, fa
);
360 xfs_btree_sblock_calc_crc(bp
);
364 const struct xfs_buf_ops xfs_rmapbt_buf_ops
= {
365 .name
= "xfs_rmapbt",
366 .verify_read
= xfs_rmapbt_read_verify
,
367 .verify_write
= xfs_rmapbt_write_verify
,
368 .verify_struct
= xfs_rmapbt_verify
,
372 xfs_rmapbt_keys_inorder(
373 struct xfs_btree_cur
*cur
,
374 union xfs_btree_key
*k1
,
375 union xfs_btree_key
*k2
)
382 x
= be32_to_cpu(k1
->rmap
.rm_startblock
);
383 y
= be32_to_cpu(k2
->rmap
.rm_startblock
);
388 a
= be64_to_cpu(k1
->rmap
.rm_owner
);
389 b
= be64_to_cpu(k2
->rmap
.rm_owner
);
394 a
= XFS_RMAP_OFF(be64_to_cpu(k1
->rmap
.rm_offset
));
395 b
= XFS_RMAP_OFF(be64_to_cpu(k2
->rmap
.rm_offset
));
402 xfs_rmapbt_recs_inorder(
403 struct xfs_btree_cur
*cur
,
404 union xfs_btree_rec
*r1
,
405 union xfs_btree_rec
*r2
)
412 x
= be32_to_cpu(r1
->rmap
.rm_startblock
);
413 y
= be32_to_cpu(r2
->rmap
.rm_startblock
);
418 a
= be64_to_cpu(r1
->rmap
.rm_owner
);
419 b
= be64_to_cpu(r2
->rmap
.rm_owner
);
424 a
= XFS_RMAP_OFF(be64_to_cpu(r1
->rmap
.rm_offset
));
425 b
= XFS_RMAP_OFF(be64_to_cpu(r2
->rmap
.rm_offset
));
431 static const struct xfs_btree_ops xfs_rmapbt_ops
= {
432 .rec_len
= sizeof(struct xfs_rmap_rec
),
433 .key_len
= 2 * sizeof(struct xfs_rmap_key
),
435 .dup_cursor
= xfs_rmapbt_dup_cursor
,
436 .set_root
= xfs_rmapbt_set_root
,
437 .alloc_block
= xfs_rmapbt_alloc_block
,
438 .free_block
= xfs_rmapbt_free_block
,
439 .get_minrecs
= xfs_rmapbt_get_minrecs
,
440 .get_maxrecs
= xfs_rmapbt_get_maxrecs
,
441 .init_key_from_rec
= xfs_rmapbt_init_key_from_rec
,
442 .init_high_key_from_rec
= xfs_rmapbt_init_high_key_from_rec
,
443 .init_rec_from_cur
= xfs_rmapbt_init_rec_from_cur
,
444 .init_ptr_from_cur
= xfs_rmapbt_init_ptr_from_cur
,
445 .key_diff
= xfs_rmapbt_key_diff
,
446 .buf_ops
= &xfs_rmapbt_buf_ops
,
447 .diff_two_keys
= xfs_rmapbt_diff_two_keys
,
448 .keys_inorder
= xfs_rmapbt_keys_inorder
,
449 .recs_inorder
= xfs_rmapbt_recs_inorder
,
453 * Allocate a new allocation btree cursor.
455 struct xfs_btree_cur
*
456 xfs_rmapbt_init_cursor(
457 struct xfs_mount
*mp
,
458 struct xfs_trans
*tp
,
459 struct xfs_buf
*agbp
,
462 struct xfs_agf
*agf
= XFS_BUF_TO_AGF(agbp
);
463 struct xfs_btree_cur
*cur
;
465 cur
= kmem_zone_zalloc(xfs_btree_cur_zone
, KM_NOFS
);
468 /* Overlapping btree; 2 keys per pointer. */
469 cur
->bc_btnum
= XFS_BTNUM_RMAP
;
470 cur
->bc_flags
= XFS_BTREE_CRC_BLOCKS
| XFS_BTREE_OVERLAPPING
;
471 cur
->bc_blocklog
= mp
->m_sb
.sb_blocklog
;
472 cur
->bc_ops
= &xfs_rmapbt_ops
;
473 cur
->bc_nlevels
= be32_to_cpu(agf
->agf_levels
[XFS_BTNUM_RMAP
]);
474 cur
->bc_statoff
= XFS_STATS_CALC_INDEX(xs_rmap_2
);
476 cur
->bc_private
.a
.agbp
= agbp
;
477 cur
->bc_private
.a
.agno
= agno
;
483 * Calculate number of records in an rmap btree block.
490 blocklen
-= XFS_RMAP_BLOCK_LEN
;
493 return blocklen
/ sizeof(struct xfs_rmap_rec
);
495 (2 * sizeof(struct xfs_rmap_key
) + sizeof(xfs_rmap_ptr_t
));
498 /* Compute the maximum height of an rmap btree. */
500 xfs_rmapbt_compute_maxlevels(
501 struct xfs_mount
*mp
)
504 * On a non-reflink filesystem, the maximum number of rmap
505 * records is the number of blocks in the AG, hence the max
506 * rmapbt height is log_$maxrecs($agblocks). However, with
507 * reflink each AG block can have up to 2^32 (per the refcount
508 * record format) owners, which means that theoretically we
509 * could face up to 2^64 rmap records.
511 * That effectively means that the max rmapbt height must be
512 * XFS_BTREE_MAXLEVELS. "Fortunately" we'll run out of AG
513 * blocks to feed the rmapbt long before the rmapbt reaches
514 * maximum height. The reflink code uses ag_resv_critical to
515 * disallow reflinking when less than 10% of the per-AG metadata
516 * block reservation since the fallback is a regular file copy.
518 if (xfs_sb_version_hasreflink(&mp
->m_sb
))
519 mp
->m_rmap_maxlevels
= XFS_BTREE_MAXLEVELS
;
521 mp
->m_rmap_maxlevels
= xfs_btree_compute_maxlevels(
522 mp
->m_rmap_mnr
, mp
->m_sb
.sb_agblocks
);
525 /* Calculate the refcount btree size for some records. */
527 xfs_rmapbt_calc_size(
528 struct xfs_mount
*mp
,
529 unsigned long long len
)
531 return xfs_btree_calc_size(mp
->m_rmap_mnr
, len
);
535 * Calculate the maximum refcount btree size.
539 struct xfs_mount
*mp
,
540 xfs_agblock_t agblocks
)
542 /* Bail out if we're uninitialized, which can happen in mkfs. */
543 if (mp
->m_rmap_mxr
[0] == 0)
546 return xfs_rmapbt_calc_size(mp
, agblocks
);
550 * Figure out how many blocks to reserve and how many are used by this btree.
553 xfs_rmapbt_calc_reserves(
554 struct xfs_mount
*mp
,
555 struct xfs_trans
*tp
,
560 struct xfs_buf
*agbp
;
562 xfs_agblock_t agblocks
;
563 xfs_extlen_t tree_len
;
566 if (!xfs_sb_version_hasrmapbt(&mp
->m_sb
))
569 error
= xfs_alloc_read_agf(mp
, tp
, agno
, 0, &agbp
);
573 agf
= XFS_BUF_TO_AGF(agbp
);
574 agblocks
= be32_to_cpu(agf
->agf_length
);
575 tree_len
= be32_to_cpu(agf
->agf_rmap_blocks
);
576 xfs_trans_brelse(tp
, agbp
);
578 /* Reserve 1% of the AG or enough for 1 block per record. */
579 *ask
+= max(agblocks
/ 100, xfs_rmapbt_max_size(mp
, agblocks
));