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"
13 #include "xfs_mount.h"
14 #include "xfs_trans.h"
15 #include "xfs_alloc.h"
16 #include "xfs_btree.h"
17 #include "xfs_btree_staging.h"
19 #include "xfs_rmap_btree.h"
20 #include "xfs_trace.h"
21 #include "xfs_ag_resv.h"
26 * This is a per-ag tree used to track the owner(s) of a given extent. With
27 * reflink it is possible for there to be multiple owners, which is a departure
28 * from classic XFS. Owner records for data extents are inserted when the
29 * extent is mapped and removed when an extent is unmapped. Owner records for
30 * all other block types (i.e. metadata) are inserted when an extent is
31 * allocated and removed when an extent is freed. There can only be one owner
32 * of a metadata extent, usually an inode or some other metadata structure like
35 * The rmap btree is part of the free space management, so blocks for the tree
36 * are sourced from the agfl. Hence we need transaction reservation support for
37 * this tree so that the freelist is always large enough. This also impacts on
38 * the minimum space we need to leave free in the AG.
40 * The tree is ordered by [ag block, owner, offset]. This is a large key size,
41 * but it is the only way to enforce unique keys when a block can be owned by
42 * multiple files at any offset. There's no need to order/search by extent
43 * size for online updating/management of the tree. It is intended that most
44 * reverse lookups will be to find the owner(s) of a particular block, or to
45 * try to recover tree and file data from corrupt primary metadata.
48 static struct xfs_btree_cur
*
49 xfs_rmapbt_dup_cursor(
50 struct xfs_btree_cur
*cur
)
52 return xfs_rmapbt_init_cursor(cur
->bc_mp
, cur
->bc_tp
,
53 cur
->bc_ag
.agbp
, cur
->bc_ag
.agno
);
58 struct xfs_btree_cur
*cur
,
59 union xfs_btree_ptr
*ptr
,
62 struct xfs_buf
*agbp
= cur
->bc_ag
.agbp
;
63 struct xfs_agf
*agf
= agbp
->b_addr
;
64 xfs_agnumber_t seqno
= be32_to_cpu(agf
->agf_seqno
);
65 int btnum
= cur
->bc_btnum
;
66 struct xfs_perag
*pag
= xfs_perag_get(cur
->bc_mp
, seqno
);
70 agf
->agf_roots
[btnum
] = ptr
->s
;
71 be32_add_cpu(&agf
->agf_levels
[btnum
], inc
);
72 pag
->pagf_levels
[btnum
] += inc
;
75 xfs_alloc_log_agf(cur
->bc_tp
, agbp
, XFS_AGF_ROOTS
| XFS_AGF_LEVELS
);
79 xfs_rmapbt_alloc_block(
80 struct xfs_btree_cur
*cur
,
81 union xfs_btree_ptr
*start
,
82 union xfs_btree_ptr
*new,
85 struct xfs_buf
*agbp
= cur
->bc_ag
.agbp
;
86 struct xfs_agf
*agf
= agbp
->b_addr
;
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
,
96 trace_xfs_rmapbt_alloc_block(cur
->bc_mp
, cur
->bc_ag
.agno
,
98 if (bno
== NULLAGBLOCK
) {
103 xfs_extent_busy_reuse(cur
->bc_mp
, cur
->bc_ag
.agno
, bno
, 1,
106 xfs_trans_agbtree_delta(cur
->bc_tp
, 1);
107 new->s
= cpu_to_be32(bno
);
108 be32_add_cpu(&agf
->agf_rmap_blocks
, 1);
109 xfs_alloc_log_agf(cur
->bc_tp
, agbp
, XFS_AGF_RMAP_BLOCKS
);
111 xfs_ag_resv_rmapbt_alloc(cur
->bc_mp
, cur
->bc_ag
.agno
);
118 xfs_rmapbt_free_block(
119 struct xfs_btree_cur
*cur
,
122 struct xfs_buf
*agbp
= cur
->bc_ag
.agbp
;
123 struct xfs_agf
*agf
= agbp
->b_addr
;
127 bno
= xfs_daddr_to_agbno(cur
->bc_mp
, XFS_BUF_ADDR(bp
));
128 trace_xfs_rmapbt_free_block(cur
->bc_mp
, cur
->bc_ag
.agno
,
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(cur
->bc_tp
, agbp
, NULL
, bno
, 1);
136 xfs_extent_busy_insert(cur
->bc_tp
, be32_to_cpu(agf
->agf_seqno
), bno
, 1,
137 XFS_EXTENT_BUSY_SKIP_DISCARD
);
138 xfs_trans_agbtree_delta(cur
->bc_tp
, -1);
140 xfs_ag_resv_rmapbt_free(cur
->bc_mp
, cur
->bc_ag
.agno
);
146 xfs_rmapbt_get_minrecs(
147 struct xfs_btree_cur
*cur
,
150 return cur
->bc_mp
->m_rmap_mnr
[level
!= 0];
154 xfs_rmapbt_get_maxrecs(
155 struct xfs_btree_cur
*cur
,
158 return cur
->bc_mp
->m_rmap_mxr
[level
!= 0];
162 xfs_rmapbt_init_key_from_rec(
163 union xfs_btree_key
*key
,
164 union xfs_btree_rec
*rec
)
166 key
->rmap
.rm_startblock
= rec
->rmap
.rm_startblock
;
167 key
->rmap
.rm_owner
= rec
->rmap
.rm_owner
;
168 key
->rmap
.rm_offset
= rec
->rmap
.rm_offset
;
172 * The high key for a reverse mapping record can be computed by shifting
173 * the startblock and offset to the highest value that would still map
174 * to that record. In practice this means that we add blockcount-1 to
175 * the startblock for all records, and if the record is for a data/attr
176 * fork mapping, we add blockcount-1 to the offset too.
179 xfs_rmapbt_init_high_key_from_rec(
180 union xfs_btree_key
*key
,
181 union xfs_btree_rec
*rec
)
186 adj
= be32_to_cpu(rec
->rmap
.rm_blockcount
) - 1;
188 key
->rmap
.rm_startblock
= rec
->rmap
.rm_startblock
;
189 be32_add_cpu(&key
->rmap
.rm_startblock
, adj
);
190 key
->rmap
.rm_owner
= rec
->rmap
.rm_owner
;
191 key
->rmap
.rm_offset
= rec
->rmap
.rm_offset
;
192 if (XFS_RMAP_NON_INODE_OWNER(be64_to_cpu(rec
->rmap
.rm_owner
)) ||
193 XFS_RMAP_IS_BMBT_BLOCK(be64_to_cpu(rec
->rmap
.rm_offset
)))
195 off
= be64_to_cpu(key
->rmap
.rm_offset
);
196 off
= (XFS_RMAP_OFF(off
) + adj
) | (off
& ~XFS_RMAP_OFF_MASK
);
197 key
->rmap
.rm_offset
= cpu_to_be64(off
);
201 xfs_rmapbt_init_rec_from_cur(
202 struct xfs_btree_cur
*cur
,
203 union xfs_btree_rec
*rec
)
205 rec
->rmap
.rm_startblock
= cpu_to_be32(cur
->bc_rec
.r
.rm_startblock
);
206 rec
->rmap
.rm_blockcount
= cpu_to_be32(cur
->bc_rec
.r
.rm_blockcount
);
207 rec
->rmap
.rm_owner
= cpu_to_be64(cur
->bc_rec
.r
.rm_owner
);
208 rec
->rmap
.rm_offset
= cpu_to_be64(
209 xfs_rmap_irec_offset_pack(&cur
->bc_rec
.r
));
213 xfs_rmapbt_init_ptr_from_cur(
214 struct xfs_btree_cur
*cur
,
215 union xfs_btree_ptr
*ptr
)
217 struct xfs_agf
*agf
= cur
->bc_ag
.agbp
->b_addr
;
219 ASSERT(cur
->bc_ag
.agno
== be32_to_cpu(agf
->agf_seqno
));
221 ptr
->s
= agf
->agf_roots
[cur
->bc_btnum
];
226 struct xfs_btree_cur
*cur
,
227 union xfs_btree_key
*key
)
229 struct xfs_rmap_irec
*rec
= &cur
->bc_rec
.r
;
230 struct xfs_rmap_key
*kp
= &key
->rmap
;
234 d
= (int64_t)be32_to_cpu(kp
->rm_startblock
) - rec
->rm_startblock
;
238 x
= be64_to_cpu(kp
->rm_owner
);
245 x
= XFS_RMAP_OFF(be64_to_cpu(kp
->rm_offset
));
255 xfs_rmapbt_diff_two_keys(
256 struct xfs_btree_cur
*cur
,
257 union xfs_btree_key
*k1
,
258 union xfs_btree_key
*k2
)
260 struct xfs_rmap_key
*kp1
= &k1
->rmap
;
261 struct xfs_rmap_key
*kp2
= &k2
->rmap
;
265 d
= (int64_t)be32_to_cpu(kp1
->rm_startblock
) -
266 be32_to_cpu(kp2
->rm_startblock
);
270 x
= be64_to_cpu(kp1
->rm_owner
);
271 y
= be64_to_cpu(kp2
->rm_owner
);
277 x
= XFS_RMAP_OFF(be64_to_cpu(kp1
->rm_offset
));
278 y
= XFS_RMAP_OFF(be64_to_cpu(kp2
->rm_offset
));
286 static xfs_failaddr_t
290 struct xfs_mount
*mp
= bp
->b_mount
;
291 struct xfs_btree_block
*block
= XFS_BUF_TO_BLOCK(bp
);
292 struct xfs_perag
*pag
= bp
->b_pag
;
297 * magic number and level verification
299 * During growfs operations, we can't verify the exact level or owner as
300 * the perag is not fully initialised and hence not attached to the
301 * buffer. In this case, check against the maximum tree depth.
303 * Similarly, during log recovery we will have a perag structure
304 * attached, but the agf information will not yet have been initialised
305 * from the on disk AGF. Again, we can only check against maximum limits
308 if (!xfs_verify_magic(bp
, block
->bb_magic
))
309 return __this_address
;
311 if (!xfs_sb_version_hasrmapbt(&mp
->m_sb
))
312 return __this_address
;
313 fa
= xfs_btree_sblock_v5hdr_verify(bp
);
317 level
= be16_to_cpu(block
->bb_level
);
318 if (pag
&& pag
->pagf_init
) {
319 if (level
>= pag
->pagf_levels
[XFS_BTNUM_RMAPi
])
320 return __this_address
;
321 } else if (level
>= mp
->m_rmap_maxlevels
)
322 return __this_address
;
324 return xfs_btree_sblock_verify(bp
, mp
->m_rmap_mxr
[level
!= 0]);
328 xfs_rmapbt_read_verify(
333 if (!xfs_btree_sblock_verify_crc(bp
))
334 xfs_verifier_error(bp
, -EFSBADCRC
, __this_address
);
336 fa
= xfs_rmapbt_verify(bp
);
338 xfs_verifier_error(bp
, -EFSCORRUPTED
, fa
);
342 trace_xfs_btree_corrupt(bp
, _RET_IP_
);
346 xfs_rmapbt_write_verify(
351 fa
= xfs_rmapbt_verify(bp
);
353 trace_xfs_btree_corrupt(bp
, _RET_IP_
);
354 xfs_verifier_error(bp
, -EFSCORRUPTED
, fa
);
357 xfs_btree_sblock_calc_crc(bp
);
361 const struct xfs_buf_ops xfs_rmapbt_buf_ops
= {
362 .name
= "xfs_rmapbt",
363 .magic
= { 0, cpu_to_be32(XFS_RMAP_CRC_MAGIC
) },
364 .verify_read
= xfs_rmapbt_read_verify
,
365 .verify_write
= xfs_rmapbt_write_verify
,
366 .verify_struct
= xfs_rmapbt_verify
,
370 xfs_rmapbt_keys_inorder(
371 struct xfs_btree_cur
*cur
,
372 union xfs_btree_key
*k1
,
373 union xfs_btree_key
*k2
)
380 x
= be32_to_cpu(k1
->rmap
.rm_startblock
);
381 y
= be32_to_cpu(k2
->rmap
.rm_startblock
);
386 a
= be64_to_cpu(k1
->rmap
.rm_owner
);
387 b
= be64_to_cpu(k2
->rmap
.rm_owner
);
392 a
= XFS_RMAP_OFF(be64_to_cpu(k1
->rmap
.rm_offset
));
393 b
= XFS_RMAP_OFF(be64_to_cpu(k2
->rmap
.rm_offset
));
400 xfs_rmapbt_recs_inorder(
401 struct xfs_btree_cur
*cur
,
402 union xfs_btree_rec
*r1
,
403 union xfs_btree_rec
*r2
)
410 x
= be32_to_cpu(r1
->rmap
.rm_startblock
);
411 y
= be32_to_cpu(r2
->rmap
.rm_startblock
);
416 a
= be64_to_cpu(r1
->rmap
.rm_owner
);
417 b
= be64_to_cpu(r2
->rmap
.rm_owner
);
422 a
= XFS_RMAP_OFF(be64_to_cpu(r1
->rmap
.rm_offset
));
423 b
= XFS_RMAP_OFF(be64_to_cpu(r2
->rmap
.rm_offset
));
429 static const struct xfs_btree_ops xfs_rmapbt_ops
= {
430 .rec_len
= sizeof(struct xfs_rmap_rec
),
431 .key_len
= 2 * sizeof(struct xfs_rmap_key
),
433 .dup_cursor
= xfs_rmapbt_dup_cursor
,
434 .set_root
= xfs_rmapbt_set_root
,
435 .alloc_block
= xfs_rmapbt_alloc_block
,
436 .free_block
= xfs_rmapbt_free_block
,
437 .get_minrecs
= xfs_rmapbt_get_minrecs
,
438 .get_maxrecs
= xfs_rmapbt_get_maxrecs
,
439 .init_key_from_rec
= xfs_rmapbt_init_key_from_rec
,
440 .init_high_key_from_rec
= xfs_rmapbt_init_high_key_from_rec
,
441 .init_rec_from_cur
= xfs_rmapbt_init_rec_from_cur
,
442 .init_ptr_from_cur
= xfs_rmapbt_init_ptr_from_cur
,
443 .key_diff
= xfs_rmapbt_key_diff
,
444 .buf_ops
= &xfs_rmapbt_buf_ops
,
445 .diff_two_keys
= xfs_rmapbt_diff_two_keys
,
446 .keys_inorder
= xfs_rmapbt_keys_inorder
,
447 .recs_inorder
= xfs_rmapbt_recs_inorder
,
450 static struct xfs_btree_cur
*
451 xfs_rmapbt_init_common(
452 struct xfs_mount
*mp
,
453 struct xfs_trans
*tp
,
456 struct xfs_btree_cur
*cur
;
458 cur
= kmem_zone_zalloc(xfs_btree_cur_zone
, KM_NOFS
);
461 /* Overlapping btree; 2 keys per pointer. */
462 cur
->bc_btnum
= XFS_BTNUM_RMAP
;
463 cur
->bc_flags
= XFS_BTREE_CRC_BLOCKS
| XFS_BTREE_OVERLAPPING
;
464 cur
->bc_blocklog
= mp
->m_sb
.sb_blocklog
;
465 cur
->bc_statoff
= XFS_STATS_CALC_INDEX(xs_rmap_2
);
466 cur
->bc_ag
.agno
= agno
;
467 cur
->bc_ops
= &xfs_rmapbt_ops
;
472 /* Create a new reverse mapping btree cursor. */
473 struct xfs_btree_cur
*
474 xfs_rmapbt_init_cursor(
475 struct xfs_mount
*mp
,
476 struct xfs_trans
*tp
,
477 struct xfs_buf
*agbp
,
480 struct xfs_agf
*agf
= agbp
->b_addr
;
481 struct xfs_btree_cur
*cur
;
483 cur
= xfs_rmapbt_init_common(mp
, tp
, agno
);
484 cur
->bc_nlevels
= be32_to_cpu(agf
->agf_levels
[XFS_BTNUM_RMAP
]);
485 cur
->bc_ag
.agbp
= agbp
;
489 /* Create a new reverse mapping btree cursor with a fake root for staging. */
490 struct xfs_btree_cur
*
491 xfs_rmapbt_stage_cursor(
492 struct xfs_mount
*mp
,
493 struct xbtree_afakeroot
*afake
,
496 struct xfs_btree_cur
*cur
;
498 cur
= xfs_rmapbt_init_common(mp
, NULL
, agno
);
499 xfs_btree_stage_afakeroot(cur
, afake
);
504 * Install a new reverse mapping btree root. Caller is responsible for
505 * invalidating and freeing the old btree blocks.
508 xfs_rmapbt_commit_staged_btree(
509 struct xfs_btree_cur
*cur
,
510 struct xfs_trans
*tp
,
511 struct xfs_buf
*agbp
)
513 struct xfs_agf
*agf
= agbp
->b_addr
;
514 struct xbtree_afakeroot
*afake
= cur
->bc_ag
.afake
;
516 ASSERT(cur
->bc_flags
& XFS_BTREE_STAGING
);
518 agf
->agf_roots
[cur
->bc_btnum
] = cpu_to_be32(afake
->af_root
);
519 agf
->agf_levels
[cur
->bc_btnum
] = cpu_to_be32(afake
->af_levels
);
520 agf
->agf_rmap_blocks
= cpu_to_be32(afake
->af_blocks
);
521 xfs_alloc_log_agf(tp
, agbp
, XFS_AGF_ROOTS
| XFS_AGF_LEVELS
|
522 XFS_AGF_RMAP_BLOCKS
);
523 xfs_btree_commit_afakeroot(cur
, tp
, agbp
, &xfs_rmapbt_ops
);
527 * Calculate number of records in an rmap btree block.
534 blocklen
-= XFS_RMAP_BLOCK_LEN
;
537 return blocklen
/ sizeof(struct xfs_rmap_rec
);
539 (2 * sizeof(struct xfs_rmap_key
) + sizeof(xfs_rmap_ptr_t
));
542 /* Compute the maximum height of an rmap btree. */
544 xfs_rmapbt_compute_maxlevels(
545 struct xfs_mount
*mp
)
548 * On a non-reflink filesystem, the maximum number of rmap
549 * records is the number of blocks in the AG, hence the max
550 * rmapbt height is log_$maxrecs($agblocks). However, with
551 * reflink each AG block can have up to 2^32 (per the refcount
552 * record format) owners, which means that theoretically we
553 * could face up to 2^64 rmap records.
555 * That effectively means that the max rmapbt height must be
556 * XFS_BTREE_MAXLEVELS. "Fortunately" we'll run out of AG
557 * blocks to feed the rmapbt long before the rmapbt reaches
558 * maximum height. The reflink code uses ag_resv_critical to
559 * disallow reflinking when less than 10% of the per-AG metadata
560 * block reservation since the fallback is a regular file copy.
562 if (xfs_sb_version_hasreflink(&mp
->m_sb
))
563 mp
->m_rmap_maxlevels
= XFS_BTREE_MAXLEVELS
;
565 mp
->m_rmap_maxlevels
= xfs_btree_compute_maxlevels(
566 mp
->m_rmap_mnr
, mp
->m_sb
.sb_agblocks
);
569 /* Calculate the refcount btree size for some records. */
571 xfs_rmapbt_calc_size(
572 struct xfs_mount
*mp
,
573 unsigned long long len
)
575 return xfs_btree_calc_size(mp
->m_rmap_mnr
, len
);
579 * Calculate the maximum refcount btree size.
583 struct xfs_mount
*mp
,
584 xfs_agblock_t agblocks
)
586 /* Bail out if we're uninitialized, which can happen in mkfs. */
587 if (mp
->m_rmap_mxr
[0] == 0)
590 return xfs_rmapbt_calc_size(mp
, agblocks
);
594 * Figure out how many blocks to reserve and how many are used by this btree.
597 xfs_rmapbt_calc_reserves(
598 struct xfs_mount
*mp
,
599 struct xfs_trans
*tp
,
604 struct xfs_buf
*agbp
;
606 xfs_agblock_t agblocks
;
607 xfs_extlen_t tree_len
;
610 if (!xfs_sb_version_hasrmapbt(&mp
->m_sb
))
613 error
= xfs_alloc_read_agf(mp
, tp
, agno
, 0, &agbp
);
618 agblocks
= be32_to_cpu(agf
->agf_length
);
619 tree_len
= be32_to_cpu(agf
->agf_rmap_blocks
);
620 xfs_trans_brelse(tp
, agbp
);
623 * The log is permanently allocated, so the space it occupies will
624 * never be available for the kinds of things that would require btree
625 * expansion. We therefore can pretend the space isn't there.
627 if (mp
->m_sb
.sb_logstart
&&
628 XFS_FSB_TO_AGNO(mp
, mp
->m_sb
.sb_logstart
) == agno
)
629 agblocks
-= mp
->m_sb
.sb_logblocks
;
631 /* Reserve 1% of the AG or enough for 1 block per record. */
632 *ask
+= max(agblocks
/ 100, xfs_rmapbt_max_size(mp
, agblocks
));