1/* SPDX-License-Identifier: GPL-2.0 */
2#ifndef _BCACHEFS_EXTENTS_FORMAT_H
3#define _BCACHEFS_EXTENTS_FORMAT_H
4
5/*
6 * In extent bkeys, the value is a list of pointers (bch_extent_ptr), optionally
7 * preceded by checksum/compression information (bch_extent_crc32 or
8 * bch_extent_crc64).
9 *
10 * One major determining factor in the format of extents is how we handle and
11 * represent extents that have been partially overwritten and thus trimmed:
12 *
13 * If an extent is not checksummed or compressed, when the extent is trimmed we
14 * don't have to remember the extent we originally allocated and wrote: we can
15 * merely adjust ptr->offset to point to the start of the data that is currently
16 * live. The size field in struct bkey records the current (live) size of the
17 * extent, and is also used to mean "size of region on disk that we point to" in
18 * this case.
19 *
20 * Thus an extent that is not checksummed or compressed will consist only of a
21 * list of bch_extent_ptrs, with none of the fields in
22 * bch_extent_crc32/bch_extent_crc64.
23 *
24 * When an extent is checksummed or compressed, it's not possible to read only
25 * the data that is currently live: we have to read the entire extent that was
26 * originally written, and then return only the part of the extent that is
27 * currently live.
28 *
29 * Thus, in addition to the current size of the extent in struct bkey, we need
30 * to store the size of the originally allocated space - this is the
31 * compressed_size and uncompressed_size fields in bch_extent_crc32/64. Also,
32 * when the extent is trimmed, instead of modifying the offset field of the
33 * pointer, we keep a second smaller offset field - "offset into the original
34 * extent of the currently live region".
35 *
36 * The other major determining factor is replication and data migration:
37 *
38 * Each pointer may have its own bch_extent_crc32/64. When doing a replicated
39 * write, we will initially write all the replicas in the same format, with the
40 * same checksum type and compression format - however, when copygc runs later (or
41 * tiering/cache promotion, anything that moves data), it is not in general
42 * going to rewrite all the pointers at once - one of the replicas may be in a
43 * bucket on one device that has very little fragmentation while another lives
44 * in a bucket that has become heavily fragmented, and thus is being rewritten
45 * sooner than the rest.
46 *
47 * Thus it will only move a subset of the pointers (or in the case of
48 * tiering/cache promotion perhaps add a single pointer without dropping any
49 * current pointers), and if the extent has been partially overwritten it must
50 * write only the currently live portion (or copygc would not be able to reduce
51 * fragmentation!) - which necessitates a different bch_extent_crc format for
52 * the new pointer.
53 *
54 * But in the interests of space efficiency, we don't want to store one
55 * bch_extent_crc for each pointer if we don't have to.
56 *
57 * Thus, a bch_extent consists of bch_extent_crc32s, bch_extent_crc64s, and
58 * bch_extent_ptrs appended arbitrarily one after the other. We determine the
59 * type of a given entry with a scheme similar to utf8 (except we're encoding a
60 * type, not a size), encoding the type in the position of the first set bit:
61 *
62 * bch_extent_crc32 - 0b1
63 * bch_extent_ptr - 0b10
64 * bch_extent_crc64 - 0b100
65 *
66 * We do it this way because bch_extent_crc32 is _very_ constrained on bits (and
67 * bch_extent_crc64 is the least constrained).
68 *
69 * Then, each bch_extent_crc32/64 applies to the pointers that follow after it,
70 * until the next bch_extent_crc32/64.
71 *
72 * If there are no bch_extent_crcs preceding a bch_extent_ptr, then that pointer
73 * is neither checksummed nor compressed.
74 */
75
76#define BCH_EXTENT_ENTRY_TYPES() \
77 x(ptr, 0) \
78 x(crc32, 1) \
79 x(crc64, 2) \
80 x(crc128, 3) \
81 x(stripe_ptr, 4) \
82 x(rebalance, 5)
83#define BCH_EXTENT_ENTRY_MAX 6
84
85enum bch_extent_entry_type {
86#define x(f, n) BCH_EXTENT_ENTRY_##f = n,
87 BCH_EXTENT_ENTRY_TYPES()
88#undef x
89};
90
91/* Compressed/uncompressed size are stored biased by 1: */
92struct bch_extent_crc32 {
93#if defined(__LITTLE_ENDIAN_BITFIELD)
94 __u32 type:2,
95 _compressed_size:7,
96 _uncompressed_size:7,
97 offset:7,
98 _unused:1,
99 csum_type:4,
100 compression_type:4;
101 __u32 csum;
102#elif defined (__BIG_ENDIAN_BITFIELD)
103 __u32 csum;
104 __u32 compression_type:4,
105 csum_type:4,
106 _unused:1,
107 offset:7,
108 _uncompressed_size:7,
109 _compressed_size:7,
110 type:2;
111#endif
112} __packed __aligned(8);
113
114#define CRC32_SIZE_MAX (1U << 7)
115#define CRC32_NONCE_MAX 0
116
117struct bch_extent_crc64 {
118#if defined(__LITTLE_ENDIAN_BITFIELD)
119 __u64 type:3,
120 _compressed_size:9,
121 _uncompressed_size:9,
122 offset:9,
123 nonce:10,
124 csum_type:4,
125 compression_type:4,
126 csum_hi:16;
127#elif defined (__BIG_ENDIAN_BITFIELD)
128 __u64 csum_hi:16,
129 compression_type:4,
130 csum_type:4,
131 nonce:10,
132 offset:9,
133 _uncompressed_size:9,
134 _compressed_size:9,
135 type:3;
136#endif
137 __u64 csum_lo;
138} __packed __aligned(8);
139
140#define CRC64_SIZE_MAX (1U << 9)
141#define CRC64_NONCE_MAX ((1U << 10) - 1)
142
143struct bch_extent_crc128 {
144#if defined(__LITTLE_ENDIAN_BITFIELD)
145 __u64 type:4,
146 _compressed_size:13,
147 _uncompressed_size:13,
148 offset:13,
149 nonce:13,
150 csum_type:4,
151 compression_type:4;
152#elif defined (__BIG_ENDIAN_BITFIELD)
153 __u64 compression_type:4,
154 csum_type:4,
155 nonce:13,
156 offset:13,
157 _uncompressed_size:13,
158 _compressed_size:13,
159 type:4;
160#endif
161 struct bch_csum csum;
162} __packed __aligned(8);
163
164#define CRC128_SIZE_MAX (1U << 13)
165#define CRC128_NONCE_MAX ((1U << 13) - 1)
166
167/*
168 * @reservation - pointer hasn't been written to, just reserved
169 */
170struct bch_extent_ptr {
171#if defined(__LITTLE_ENDIAN_BITFIELD)
172 __u64 type:1,
173 cached:1,
174 unused:1,
175 unwritten:1,
176 offset:44, /* 8 petabytes */
177 dev:8,
178 gen:8;
179#elif defined (__BIG_ENDIAN_BITFIELD)
180 __u64 gen:8,
181 dev:8,
182 offset:44,
183 unwritten:1,
184 unused:1,
185 cached:1,
186 type:1;
187#endif
188} __packed __aligned(8);
189
190struct bch_extent_stripe_ptr {
191#if defined(__LITTLE_ENDIAN_BITFIELD)
192 __u64 type:5,
193 block:8,
194 redundancy:4,
195 idx:47;
196#elif defined (__BIG_ENDIAN_BITFIELD)
197 __u64 idx:47,
198 redundancy:4,
199 block:8,
200 type:5;
201#endif
202};
203
204struct bch_extent_rebalance {
205#if defined(__LITTLE_ENDIAN_BITFIELD)
206 __u64 type:6,
207 unused:34,
208 compression:8, /* enum bch_compression_opt */
209 target:16;
210#elif defined (__BIG_ENDIAN_BITFIELD)
211 __u64 target:16,
212 compression:8,
213 unused:34,
214 type:6;
215#endif
216};
217
218union bch_extent_entry {
219#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ || __BITS_PER_LONG == 64
220 unsigned long type;
221#elif __BITS_PER_LONG == 32
222 struct {
223 unsigned long pad;
224 unsigned long type;
225 };
226#else
227#error edit for your odd byteorder.
228#endif
229
230#define x(f, n) struct bch_extent_##f f;
231 BCH_EXTENT_ENTRY_TYPES()
232#undef x
233};
234
235struct bch_btree_ptr {
236 struct bch_val v;
237
238 __u64 _data[0];
239 struct bch_extent_ptr start[];
240} __packed __aligned(8);
241
242struct bch_btree_ptr_v2 {
243 struct bch_val v;
244
245 __u64 mem_ptr;
246 __le64 seq;
247 __le16 sectors_written;
248 __le16 flags;
249 struct bpos min_key;
250 __u64 _data[0];
251 struct bch_extent_ptr start[];
252} __packed __aligned(8);
253
254LE16_BITMASK(BTREE_PTR_RANGE_UPDATED, struct bch_btree_ptr_v2, flags, 0, 1);
255
256struct bch_extent {
257 struct bch_val v;
258
259 __u64 _data[0];
260 union bch_extent_entry start[];
261} __packed __aligned(8);
262
263/* Maximum size (in u64s) a single pointer could be: */
264#define BKEY_EXTENT_PTR_U64s_MAX\
265 ((sizeof(struct bch_extent_crc128) + \
266 sizeof(struct bch_extent_ptr)) / sizeof(__u64))
267
268/* Maximum possible size of an entire extent value: */
269#define BKEY_EXTENT_VAL_U64s_MAX \
270 (1 + BKEY_EXTENT_PTR_U64s_MAX * (BCH_REPLICAS_MAX + 1))
271
272/* * Maximum possible size of an entire extent, key + value: */
273#define BKEY_EXTENT_U64s_MAX (BKEY_U64s + BKEY_EXTENT_VAL_U64s_MAX)
274
275/* Btree pointers don't carry around checksums: */
276#define BKEY_BTREE_PTR_VAL_U64s_MAX \
277 ((sizeof(struct bch_btree_ptr_v2) + \
278 sizeof(struct bch_extent_ptr) * BCH_REPLICAS_MAX) / sizeof(__u64))
279#define BKEY_BTREE_PTR_U64s_MAX \
280 (BKEY_U64s + BKEY_BTREE_PTR_VAL_U64s_MAX)
281
282struct bch_reservation {
283 struct bch_val v;
284
285 __le32 generation;
286 __u8 nr_replicas;
287 __u8 pad[3];
288} __packed __aligned(8);
289
290struct bch_inline_data {
291 struct bch_val v;
292 u8 data[];
293};
294
295#endif /* _BCACHEFS_EXTENTS_FORMAT_H */
296

source code of linux/fs/bcachefs/extents_format.h