2 * Copyright (c) 2014 Red Hat, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
18 #include "libxfs_priv.h"
20 #include "xfs_shared.h"
21 #include "xfs_format.h"
22 #include "xfs_log_format.h"
23 #include "xfs_trans_resv.h"
26 #include "xfs_mount.h"
27 #include "xfs_defer.h"
28 #include "xfs_inode.h"
29 #include "xfs_trans.h"
30 #include "xfs_alloc.h"
31 #include "xfs_btree.h"
33 #include "xfs_rmap_btree.h"
34 #include "xfs_trace.h"
35 #include "xfs_cksum.h"
40 * This is a per-ag tree used to track the owner(s) of a given extent. With
41 * reflink it is possible for there to be multiple owners, which is a departure
42 * from classic XFS. Owner records for data extents are inserted when the
43 * extent is mapped and removed when an extent is unmapped. Owner records for
44 * all other block types (i.e. metadata) are inserted when an extent is
45 * allocated and removed when an extent is freed. There can only be one owner
46 * of a metadata extent, usually an inode or some other metadata structure like
49 * The rmap btree is part of the free space management, so blocks for the tree
50 * are sourced from the agfl. Hence we need transaction reservation support for
51 * this tree so that the freelist is always large enough. This also impacts on
52 * the minimum space we need to leave free in the AG.
54 * The tree is ordered by [ag block, owner, offset]. This is a large key size,
55 * but it is the only way to enforce unique keys when a block can be owned by
56 * multiple files at any offset. There's no need to order/search by extent
57 * size for online updating/management of the tree. It is intended that most
58 * reverse lookups will be to find the owner(s) of a particular block, or to
59 * try to recover tree and file data from corrupt primary metadata.
62 static struct xfs_btree_cur
*
63 xfs_rmapbt_dup_cursor(
64 struct xfs_btree_cur
*cur
)
66 return xfs_rmapbt_init_cursor(cur
->bc_mp
, cur
->bc_tp
,
67 cur
->bc_private
.a
.agbp
, cur
->bc_private
.a
.agno
);
72 struct xfs_btree_cur
*cur
,
73 union xfs_btree_ptr
*ptr
,
76 struct xfs_buf
*agbp
= cur
->bc_private
.a
.agbp
;
77 struct xfs_agf
*agf
= XFS_BUF_TO_AGF(agbp
);
78 xfs_agnumber_t seqno
= be32_to_cpu(agf
->agf_seqno
);
79 int btnum
= cur
->bc_btnum
;
80 struct xfs_perag
*pag
= xfs_perag_get(cur
->bc_mp
, seqno
);
84 agf
->agf_roots
[btnum
] = ptr
->s
;
85 be32_add_cpu(&agf
->agf_levels
[btnum
], inc
);
86 pag
->pagf_levels
[btnum
] += inc
;
89 xfs_alloc_log_agf(cur
->bc_tp
, agbp
, XFS_AGF_ROOTS
| XFS_AGF_LEVELS
);
93 xfs_rmapbt_alloc_block(
94 struct xfs_btree_cur
*cur
,
95 union xfs_btree_ptr
*start
,
96 union xfs_btree_ptr
*new,
99 struct xfs_buf
*agbp
= cur
->bc_private
.a
.agbp
;
100 struct xfs_agf
*agf
= XFS_BUF_TO_AGF(agbp
);
104 XFS_BTREE_TRACE_CURSOR(cur
, XBT_ENTRY
);
106 /* Allocate the new block from the freelist. If we can't, give up. */
107 error
= xfs_alloc_get_freelist(cur
->bc_tp
, cur
->bc_private
.a
.agbp
,
110 XFS_BTREE_TRACE_CURSOR(cur
, XBT_ERROR
);
114 trace_xfs_rmapbt_alloc_block(cur
->bc_mp
, cur
->bc_private
.a
.agno
,
116 if (bno
== NULLAGBLOCK
) {
117 XFS_BTREE_TRACE_CURSOR(cur
, XBT_EXIT
);
122 xfs_extent_busy_reuse(cur
->bc_mp
, cur
->bc_private
.a
.agno
, bno
, 1,
125 xfs_trans_agbtree_delta(cur
->bc_tp
, 1);
126 new->s
= cpu_to_be32(bno
);
127 be32_add_cpu(&agf
->agf_rmap_blocks
, 1);
128 xfs_alloc_log_agf(cur
->bc_tp
, agbp
, XFS_AGF_RMAP_BLOCKS
);
130 XFS_BTREE_TRACE_CURSOR(cur
, XBT_EXIT
);
136 xfs_rmapbt_free_block(
137 struct xfs_btree_cur
*cur
,
140 struct xfs_buf
*agbp
= cur
->bc_private
.a
.agbp
;
141 struct xfs_agf
*agf
= XFS_BUF_TO_AGF(agbp
);
145 bno
= xfs_daddr_to_agbno(cur
->bc_mp
, XFS_BUF_ADDR(bp
));
146 trace_xfs_rmapbt_free_block(cur
->bc_mp
, cur
->bc_private
.a
.agno
,
148 be32_add_cpu(&agf
->agf_rmap_blocks
, -1);
149 xfs_alloc_log_agf(cur
->bc_tp
, agbp
, XFS_AGF_RMAP_BLOCKS
);
150 error
= xfs_alloc_put_freelist(cur
->bc_tp
, agbp
, NULL
, bno
, 1);
154 xfs_extent_busy_insert(cur
->bc_tp
, be32_to_cpu(agf
->agf_seqno
), bno
, 1,
155 XFS_EXTENT_BUSY_SKIP_DISCARD
);
156 xfs_trans_agbtree_delta(cur
->bc_tp
, -1);
162 xfs_rmapbt_get_minrecs(
163 struct xfs_btree_cur
*cur
,
166 return cur
->bc_mp
->m_rmap_mnr
[level
!= 0];
170 xfs_rmapbt_get_maxrecs(
171 struct xfs_btree_cur
*cur
,
174 return cur
->bc_mp
->m_rmap_mxr
[level
!= 0];
178 xfs_rmapbt_init_key_from_rec(
179 union xfs_btree_key
*key
,
180 union xfs_btree_rec
*rec
)
182 key
->rmap
.rm_startblock
= rec
->rmap
.rm_startblock
;
183 key
->rmap
.rm_owner
= rec
->rmap
.rm_owner
;
184 key
->rmap
.rm_offset
= rec
->rmap
.rm_offset
;
188 * The high key for a reverse mapping record can be computed by shifting
189 * the startblock and offset to the highest value that would still map
190 * to that record. In practice this means that we add blockcount-1 to
191 * the startblock for all records, and if the record is for a data/attr
192 * fork mapping, we add blockcount-1 to the offset too.
195 xfs_rmapbt_init_high_key_from_rec(
196 union xfs_btree_key
*key
,
197 union xfs_btree_rec
*rec
)
202 adj
= be32_to_cpu(rec
->rmap
.rm_blockcount
) - 1;
204 key
->rmap
.rm_startblock
= rec
->rmap
.rm_startblock
;
205 be32_add_cpu(&key
->rmap
.rm_startblock
, adj
);
206 key
->rmap
.rm_owner
= rec
->rmap
.rm_owner
;
207 key
->rmap
.rm_offset
= rec
->rmap
.rm_offset
;
208 if (XFS_RMAP_NON_INODE_OWNER(be64_to_cpu(rec
->rmap
.rm_owner
)) ||
209 XFS_RMAP_IS_BMBT_BLOCK(be64_to_cpu(rec
->rmap
.rm_offset
)))
211 off
= be64_to_cpu(key
->rmap
.rm_offset
);
212 off
= (XFS_RMAP_OFF(off
) + adj
) | (off
& ~XFS_RMAP_OFF_MASK
);
213 key
->rmap
.rm_offset
= cpu_to_be64(off
);
217 xfs_rmapbt_init_rec_from_cur(
218 struct xfs_btree_cur
*cur
,
219 union xfs_btree_rec
*rec
)
221 rec
->rmap
.rm_startblock
= cpu_to_be32(cur
->bc_rec
.r
.rm_startblock
);
222 rec
->rmap
.rm_blockcount
= cpu_to_be32(cur
->bc_rec
.r
.rm_blockcount
);
223 rec
->rmap
.rm_owner
= cpu_to_be64(cur
->bc_rec
.r
.rm_owner
);
224 rec
->rmap
.rm_offset
= cpu_to_be64(
225 xfs_rmap_irec_offset_pack(&cur
->bc_rec
.r
));
229 xfs_rmapbt_init_ptr_from_cur(
230 struct xfs_btree_cur
*cur
,
231 union xfs_btree_ptr
*ptr
)
233 struct xfs_agf
*agf
= XFS_BUF_TO_AGF(cur
->bc_private
.a
.agbp
);
235 ASSERT(cur
->bc_private
.a
.agno
== be32_to_cpu(agf
->agf_seqno
));
236 ASSERT(agf
->agf_roots
[cur
->bc_btnum
] != 0);
238 ptr
->s
= agf
->agf_roots
[cur
->bc_btnum
];
243 struct xfs_btree_cur
*cur
,
244 union xfs_btree_key
*key
)
246 struct xfs_rmap_irec
*rec
= &cur
->bc_rec
.r
;
247 struct xfs_rmap_key
*kp
= &key
->rmap
;
251 d
= (__int64_t
)be32_to_cpu(kp
->rm_startblock
) - rec
->rm_startblock
;
255 x
= be64_to_cpu(kp
->rm_owner
);
262 x
= XFS_RMAP_OFF(be64_to_cpu(kp
->rm_offset
));
272 xfs_rmapbt_diff_two_keys(
273 struct xfs_btree_cur
*cur
,
274 union xfs_btree_key
*k1
,
275 union xfs_btree_key
*k2
)
277 struct xfs_rmap_key
*kp1
= &k1
->rmap
;
278 struct xfs_rmap_key
*kp2
= &k2
->rmap
;
282 d
= (__int64_t
)be32_to_cpu(kp1
->rm_startblock
) -
283 be32_to_cpu(kp2
->rm_startblock
);
287 x
= be64_to_cpu(kp1
->rm_owner
);
288 y
= be64_to_cpu(kp2
->rm_owner
);
294 x
= XFS_RMAP_OFF(be64_to_cpu(kp1
->rm_offset
));
295 y
= XFS_RMAP_OFF(be64_to_cpu(kp2
->rm_offset
));
307 struct xfs_mount
*mp
= bp
->b_target
->bt_mount
;
308 struct xfs_btree_block
*block
= XFS_BUF_TO_BLOCK(bp
);
309 struct xfs_perag
*pag
= bp
->b_pag
;
313 * magic number and level verification
315 * During growfs operations, we can't verify the exact level or owner as
316 * the perag is not fully initialised and hence not attached to the
317 * buffer. In this case, check against the maximum tree depth.
319 * Similarly, during log recovery we will have a perag structure
320 * attached, but the agf information will not yet have been initialised
321 * from the on disk AGF. Again, we can only check against maximum limits
324 if (block
->bb_magic
!= cpu_to_be32(XFS_RMAP_CRC_MAGIC
))
327 if (!xfs_sb_version_hasrmapbt(&mp
->m_sb
))
329 if (!xfs_btree_sblock_v5hdr_verify(bp
))
332 level
= be16_to_cpu(block
->bb_level
);
333 if (pag
&& pag
->pagf_init
) {
334 if (level
>= pag
->pagf_levels
[XFS_BTNUM_RMAPi
])
336 } else if (level
>= mp
->m_rmap_maxlevels
)
339 return xfs_btree_sblock_verify(bp
, mp
->m_rmap_mxr
[level
!= 0]);
343 xfs_rmapbt_read_verify(
346 if (!xfs_btree_sblock_verify_crc(bp
))
347 xfs_buf_ioerror(bp
, -EFSBADCRC
);
348 else if (!xfs_rmapbt_verify(bp
))
349 xfs_buf_ioerror(bp
, -EFSCORRUPTED
);
352 trace_xfs_btree_corrupt(bp
, _RET_IP_
);
353 xfs_verifier_error(bp
);
358 xfs_rmapbt_write_verify(
361 if (!xfs_rmapbt_verify(bp
)) {
362 trace_xfs_btree_corrupt(bp
, _RET_IP_
);
363 xfs_buf_ioerror(bp
, -EFSCORRUPTED
);
364 xfs_verifier_error(bp
);
367 xfs_btree_sblock_calc_crc(bp
);
371 const struct xfs_buf_ops xfs_rmapbt_buf_ops
= {
372 .name
= "xfs_rmapbt",
373 .verify_read
= xfs_rmapbt_read_verify
,
374 .verify_write
= xfs_rmapbt_write_verify
,
377 #if defined(DEBUG) || defined(XFS_WARN)
379 xfs_rmapbt_keys_inorder(
380 struct xfs_btree_cur
*cur
,
381 union xfs_btree_key
*k1
,
382 union xfs_btree_key
*k2
)
389 x
= be32_to_cpu(k1
->rmap
.rm_startblock
);
390 y
= be32_to_cpu(k2
->rmap
.rm_startblock
);
395 a
= be64_to_cpu(k1
->rmap
.rm_owner
);
396 b
= be64_to_cpu(k2
->rmap
.rm_owner
);
401 a
= XFS_RMAP_OFF(be64_to_cpu(k1
->rmap
.rm_offset
));
402 b
= XFS_RMAP_OFF(be64_to_cpu(k2
->rmap
.rm_offset
));
409 xfs_rmapbt_recs_inorder(
410 struct xfs_btree_cur
*cur
,
411 union xfs_btree_rec
*r1
,
412 union xfs_btree_rec
*r2
)
419 x
= be32_to_cpu(r1
->rmap
.rm_startblock
);
420 y
= be32_to_cpu(r2
->rmap
.rm_startblock
);
425 a
= be64_to_cpu(r1
->rmap
.rm_owner
);
426 b
= be64_to_cpu(r2
->rmap
.rm_owner
);
431 a
= XFS_RMAP_OFF(be64_to_cpu(r1
->rmap
.rm_offset
));
432 b
= XFS_RMAP_OFF(be64_to_cpu(r2
->rmap
.rm_offset
));
439 static const struct xfs_btree_ops xfs_rmapbt_ops
= {
440 .rec_len
= sizeof(struct xfs_rmap_rec
),
441 .key_len
= 2 * sizeof(struct xfs_rmap_key
),
443 .dup_cursor
= xfs_rmapbt_dup_cursor
,
444 .set_root
= xfs_rmapbt_set_root
,
445 .alloc_block
= xfs_rmapbt_alloc_block
,
446 .free_block
= xfs_rmapbt_free_block
,
447 .get_minrecs
= xfs_rmapbt_get_minrecs
,
448 .get_maxrecs
= xfs_rmapbt_get_maxrecs
,
449 .init_key_from_rec
= xfs_rmapbt_init_key_from_rec
,
450 .init_high_key_from_rec
= xfs_rmapbt_init_high_key_from_rec
,
451 .init_rec_from_cur
= xfs_rmapbt_init_rec_from_cur
,
452 .init_ptr_from_cur
= xfs_rmapbt_init_ptr_from_cur
,
453 .key_diff
= xfs_rmapbt_key_diff
,
454 .buf_ops
= &xfs_rmapbt_buf_ops
,
455 .diff_two_keys
= xfs_rmapbt_diff_two_keys
,
456 #if defined(DEBUG) || defined(XFS_WARN)
457 .keys_inorder
= xfs_rmapbt_keys_inorder
,
458 .recs_inorder
= xfs_rmapbt_recs_inorder
,
463 * Allocate a new allocation btree cursor.
465 struct xfs_btree_cur
*
466 xfs_rmapbt_init_cursor(
467 struct xfs_mount
*mp
,
468 struct xfs_trans
*tp
,
469 struct xfs_buf
*agbp
,
472 struct xfs_agf
*agf
= XFS_BUF_TO_AGF(agbp
);
473 struct xfs_btree_cur
*cur
;
475 cur
= kmem_zone_zalloc(xfs_btree_cur_zone
, KM_NOFS
);
478 /* Overlapping btree; 2 keys per pointer. */
479 cur
->bc_btnum
= XFS_BTNUM_RMAP
;
480 cur
->bc_flags
= XFS_BTREE_CRC_BLOCKS
| XFS_BTREE_OVERLAPPING
;
481 cur
->bc_blocklog
= mp
->m_sb
.sb_blocklog
;
482 cur
->bc_ops
= &xfs_rmapbt_ops
;
483 cur
->bc_nlevels
= be32_to_cpu(agf
->agf_levels
[XFS_BTNUM_RMAP
]);
485 cur
->bc_private
.a
.agbp
= agbp
;
486 cur
->bc_private
.a
.agno
= agno
;
492 * Calculate number of records in an rmap btree block.
496 struct xfs_mount
*mp
,
500 blocklen
-= XFS_RMAP_BLOCK_LEN
;
503 return blocklen
/ sizeof(struct xfs_rmap_rec
);
505 (2 * sizeof(struct xfs_rmap_key
) + sizeof(xfs_rmap_ptr_t
));
508 /* Compute the maximum height of an rmap btree. */
510 xfs_rmapbt_compute_maxlevels(
511 struct xfs_mount
*mp
)
514 * On a non-reflink filesystem, the maximum number of rmap
515 * records is the number of blocks in the AG, hence the max
516 * rmapbt height is log_$maxrecs($agblocks). However, with
517 * reflink each AG block can have up to 2^32 (per the refcount
518 * record format) owners, which means that theoretically we
519 * could face up to 2^64 rmap records.
521 * That effectively means that the max rmapbt height must be
522 * XFS_BTREE_MAXLEVELS. "Fortunately" we'll run out of AG
523 * blocks to feed the rmapbt long before the rmapbt reaches
524 * maximum height. The reflink code uses ag_resv_critical to
525 * disallow reflinking when less than 10% of the per-AG metadata
526 * block reservation since the fallback is a regular file copy.
528 if (xfs_sb_version_hasreflink(&mp
->m_sb
))
529 mp
->m_rmap_maxlevels
= XFS_BTREE_MAXLEVELS
;
531 mp
->m_rmap_maxlevels
= xfs_btree_compute_maxlevels(mp
,
532 mp
->m_rmap_mnr
, mp
->m_sb
.sb_agblocks
);