]> git.proxmox.com Git - mirror_zfs.git/blob - include/sys/spa.h
Check ashift validity in 'zpool add'
[mirror_zfs.git] / include / sys / spa.h
1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
24 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
26 * Copyright 2013 Saso Kiselkov. All rights reserved.
27 */
28
29 #ifndef _SYS_SPA_H
30 #define _SYS_SPA_H
31
32 #include <sys/avl.h>
33 #include <sys/zfs_context.h>
34 #include <sys/kstat.h>
35 #include <sys/nvpair.h>
36 #include <sys/sysmacros.h>
37 #include <sys/types.h>
38 #include <sys/fs/zfs.h>
39 #include <sys/spa_checksum.h>
40 #include <sys/dmu.h>
41
42 #ifdef __cplusplus
43 extern "C" {
44 #endif
45
46 /*
47 * Forward references that lots of things need.
48 */
49 typedef struct spa spa_t;
50 typedef struct vdev vdev_t;
51 typedef struct metaslab metaslab_t;
52 typedef struct metaslab_group metaslab_group_t;
53 typedef struct metaslab_class metaslab_class_t;
54 typedef struct zio zio_t;
55 typedef struct zilog zilog_t;
56 typedef struct spa_aux_vdev spa_aux_vdev_t;
57 typedef struct ddt ddt_t;
58 typedef struct ddt_entry ddt_entry_t;
59 typedef struct zbookmark_phys zbookmark_phys_t;
60
61 struct dsl_pool;
62 struct dsl_dataset;
63
64 /*
65 * General-purpose 32-bit and 64-bit bitfield encodings.
66 */
67 #define BF32_DECODE(x, low, len) P2PHASE((x) >> (low), 1U << (len))
68 #define BF64_DECODE(x, low, len) P2PHASE((x) >> (low), 1ULL << (len))
69 #define BF32_ENCODE(x, low, len) (P2PHASE((x), 1U << (len)) << (low))
70 #define BF64_ENCODE(x, low, len) (P2PHASE((x), 1ULL << (len)) << (low))
71
72 #define BF32_GET(x, low, len) BF32_DECODE(x, low, len)
73 #define BF64_GET(x, low, len) BF64_DECODE(x, low, len)
74
75 #define BF32_SET(x, low, len, val) do { \
76 ASSERT3U(val, <, 1U << (len)); \
77 ASSERT3U(low + len, <=, 32); \
78 (x) ^= BF32_ENCODE((x >> low) ^ (val), low, len); \
79 _NOTE(CONSTCOND) } while (0)
80
81 #define BF64_SET(x, low, len, val) do { \
82 ASSERT3U(val, <, 1ULL << (len)); \
83 ASSERT3U(low + len, <=, 64); \
84 ((x) ^= BF64_ENCODE((x >> low) ^ (val), low, len)); \
85 _NOTE(CONSTCOND) } while (0)
86
87 #define BF32_GET_SB(x, low, len, shift, bias) \
88 ((BF32_GET(x, low, len) + (bias)) << (shift))
89 #define BF64_GET_SB(x, low, len, shift, bias) \
90 ((BF64_GET(x, low, len) + (bias)) << (shift))
91
92 #define BF32_SET_SB(x, low, len, shift, bias, val) do { \
93 ASSERT(IS_P2ALIGNED(val, 1U << shift)); \
94 ASSERT3S((val) >> (shift), >=, bias); \
95 BF32_SET(x, low, len, ((val) >> (shift)) - (bias)); \
96 _NOTE(CONSTCOND) } while (0)
97 #define BF64_SET_SB(x, low, len, shift, bias, val) do { \
98 ASSERT(IS_P2ALIGNED(val, 1ULL << shift)); \
99 ASSERT3S((val) >> (shift), >=, bias); \
100 BF64_SET(x, low, len, ((val) >> (shift)) - (bias)); \
101 _NOTE(CONSTCOND) } while (0)
102
103 /*
104 * We currently support block sizes from 512 bytes to 16MB.
105 * The benefits of larger blocks, and thus larger IO, need to be weighed
106 * against the cost of COWing a giant block to modify one byte, and the
107 * large latency of reading or writing a large block.
108 *
109 * Note that although blocks up to 16MB are supported, the recordsize
110 * property can not be set larger than zfs_max_recordsize (default 1MB).
111 * See the comment near zfs_max_recordsize in dsl_dataset.c for details.
112 *
113 * Note that although the LSIZE field of the blkptr_t can store sizes up
114 * to 32MB, the dnode's dn_datablkszsec can only store sizes up to
115 * 32MB - 512 bytes. Therefore, we limit SPA_MAXBLOCKSIZE to 16MB.
116 */
117 #define SPA_MINBLOCKSHIFT 9
118 #define SPA_OLD_MAXBLOCKSHIFT 17
119 #define SPA_MAXBLOCKSHIFT 24
120 #define SPA_MINBLOCKSIZE (1ULL << SPA_MINBLOCKSHIFT)
121 #define SPA_OLD_MAXBLOCKSIZE (1ULL << SPA_OLD_MAXBLOCKSHIFT)
122 #define SPA_MAXBLOCKSIZE (1ULL << SPA_MAXBLOCKSHIFT)
123
124 /*
125 * Alignment Shift (ashift) is an immutable, internal top-level vdev property
126 * which can only be set at vdev creation time. Physical writes are always done
127 * according to it, which makes 2^ashift the smallest possible IO on a vdev.
128 *
129 * We currently allow values ranging from 512 bytes (2^9 = 512) to 8 KiB
130 * (2^13 = 8,192).
131 */
132 #define ASHIFT_MIN 9
133 #define ASHIFT_MAX 13
134
135 /*
136 * Size of block to hold the configuration data (a packed nvlist)
137 */
138 #define SPA_CONFIG_BLOCKSIZE (1ULL << 14)
139
140 /*
141 * The DVA size encodings for LSIZE and PSIZE support blocks up to 32MB.
142 * The ASIZE encoding should be at least 64 times larger (6 more bits)
143 * to support up to 4-way RAID-Z mirror mode with worst-case gang block
144 * overhead, three DVAs per bp, plus one more bit in case we do anything
145 * else that expands the ASIZE.
146 */
147 #define SPA_LSIZEBITS 16 /* LSIZE up to 32M (2^16 * 512) */
148 #define SPA_PSIZEBITS 16 /* PSIZE up to 32M (2^16 * 512) */
149 #define SPA_ASIZEBITS 24 /* ASIZE up to 64 times larger */
150
151 #define SPA_COMPRESSBITS 7
152
153 /*
154 * All SPA data is represented by 128-bit data virtual addresses (DVAs).
155 * The members of the dva_t should be considered opaque outside the SPA.
156 */
157 typedef struct dva {
158 uint64_t dva_word[2];
159 } dva_t;
160
161
162 /*
163 * Some checksums/hashes need a 256-bit initialization salt. This salt is kept
164 * secret and is suitable for use in MAC algorithms as the key.
165 */
166 typedef struct zio_cksum_salt {
167 uint8_t zcs_bytes[32];
168 } zio_cksum_salt_t;
169
170 /*
171 * Each block is described by its DVAs, time of birth, checksum, etc.
172 * The word-by-word, bit-by-bit layout of the blkptr is as follows:
173 *
174 * 64 56 48 40 32 24 16 8 0
175 * +-------+-------+-------+-------+-------+-------+-------+-------+
176 * 0 | vdev1 | GRID | ASIZE |
177 * +-------+-------+-------+-------+-------+-------+-------+-------+
178 * 1 |G| offset1 |
179 * +-------+-------+-------+-------+-------+-------+-------+-------+
180 * 2 | vdev2 | GRID | ASIZE |
181 * +-------+-------+-------+-------+-------+-------+-------+-------+
182 * 3 |G| offset2 |
183 * +-------+-------+-------+-------+-------+-------+-------+-------+
184 * 4 | vdev3 | GRID | ASIZE |
185 * +-------+-------+-------+-------+-------+-------+-------+-------+
186 * 5 |G| offset3 |
187 * +-------+-------+-------+-------+-------+-------+-------+-------+
188 * 6 |BDX|lvl| type | cksum |E| comp| PSIZE | LSIZE |
189 * +-------+-------+-------+-------+-------+-------+-------+-------+
190 * 7 | padding |
191 * +-------+-------+-------+-------+-------+-------+-------+-------+
192 * 8 | padding |
193 * +-------+-------+-------+-------+-------+-------+-------+-------+
194 * 9 | physical birth txg |
195 * +-------+-------+-------+-------+-------+-------+-------+-------+
196 * a | logical birth txg |
197 * +-------+-------+-------+-------+-------+-------+-------+-------+
198 * b | fill count |
199 * +-------+-------+-------+-------+-------+-------+-------+-------+
200 * c | checksum[0] |
201 * +-------+-------+-------+-------+-------+-------+-------+-------+
202 * d | checksum[1] |
203 * +-------+-------+-------+-------+-------+-------+-------+-------+
204 * e | checksum[2] |
205 * +-------+-------+-------+-------+-------+-------+-------+-------+
206 * f | checksum[3] |
207 * +-------+-------+-------+-------+-------+-------+-------+-------+
208 *
209 * Legend:
210 *
211 * vdev virtual device ID
212 * offset offset into virtual device
213 * LSIZE logical size
214 * PSIZE physical size (after compression)
215 * ASIZE allocated size (including RAID-Z parity and gang block headers)
216 * GRID RAID-Z layout information (reserved for future use)
217 * cksum checksum function
218 * comp compression function
219 * G gang block indicator
220 * B byteorder (endianness)
221 * D dedup
222 * X encryption (on version 30, which is not supported)
223 * E blkptr_t contains embedded data (see below)
224 * lvl level of indirection
225 * type DMU object type
226 * phys birth txg of block allocation; zero if same as logical birth txg
227 * log. birth transaction group in which the block was logically born
228 * fill count number of non-zero blocks under this bp
229 * checksum[4] 256-bit checksum of the data this bp describes
230 */
231
232 /*
233 * "Embedded" blkptr_t's don't actually point to a block, instead they
234 * have a data payload embedded in the blkptr_t itself. See the comment
235 * in blkptr.c for more details.
236 *
237 * The blkptr_t is laid out as follows:
238 *
239 * 64 56 48 40 32 24 16 8 0
240 * +-------+-------+-------+-------+-------+-------+-------+-------+
241 * 0 | payload |
242 * 1 | payload |
243 * 2 | payload |
244 * 3 | payload |
245 * 4 | payload |
246 * 5 | payload |
247 * +-------+-------+-------+-------+-------+-------+-------+-------+
248 * 6 |BDX|lvl| type | etype |E| comp| PSIZE| LSIZE |
249 * +-------+-------+-------+-------+-------+-------+-------+-------+
250 * 7 | payload |
251 * 8 | payload |
252 * 9 | payload |
253 * +-------+-------+-------+-------+-------+-------+-------+-------+
254 * a | logical birth txg |
255 * +-------+-------+-------+-------+-------+-------+-------+-------+
256 * b | payload |
257 * c | payload |
258 * d | payload |
259 * e | payload |
260 * f | payload |
261 * +-------+-------+-------+-------+-------+-------+-------+-------+
262 *
263 * Legend:
264 *
265 * payload contains the embedded data
266 * B (byteorder) byteorder (endianness)
267 * D (dedup) padding (set to zero)
268 * X encryption (set to zero; see above)
269 * E (embedded) set to one
270 * lvl indirection level
271 * type DMU object type
272 * etype how to interpret embedded data (BP_EMBEDDED_TYPE_*)
273 * comp compression function of payload
274 * PSIZE size of payload after compression, in bytes
275 * LSIZE logical size of payload, in bytes
276 * note that 25 bits is enough to store the largest
277 * "normal" BP's LSIZE (2^16 * 2^9) in bytes
278 * log. birth transaction group in which the block was logically born
279 *
280 * Note that LSIZE and PSIZE are stored in bytes, whereas for non-embedded
281 * bp's they are stored in units of SPA_MINBLOCKSHIFT.
282 * Generally, the generic BP_GET_*() macros can be used on embedded BP's.
283 * The B, D, X, lvl, type, and comp fields are stored the same as with normal
284 * BP's so the BP_SET_* macros can be used with them. etype, PSIZE, LSIZE must
285 * be set with the BPE_SET_* macros. BP_SET_EMBEDDED() should be called before
286 * other macros, as they assert that they are only used on BP's of the correct
287 * "embedded-ness".
288 */
289
290 #define BPE_GET_ETYPE(bp) \
291 (ASSERT(BP_IS_EMBEDDED(bp)), \
292 BF64_GET((bp)->blk_prop, 40, 8))
293 #define BPE_SET_ETYPE(bp, t) do { \
294 ASSERT(BP_IS_EMBEDDED(bp)); \
295 BF64_SET((bp)->blk_prop, 40, 8, t); \
296 _NOTE(CONSTCOND) } while (0)
297
298 #define BPE_GET_LSIZE(bp) \
299 (ASSERT(BP_IS_EMBEDDED(bp)), \
300 BF64_GET_SB((bp)->blk_prop, 0, 25, 0, 1))
301 #define BPE_SET_LSIZE(bp, x) do { \
302 ASSERT(BP_IS_EMBEDDED(bp)); \
303 BF64_SET_SB((bp)->blk_prop, 0, 25, 0, 1, x); \
304 _NOTE(CONSTCOND) } while (0)
305
306 #define BPE_GET_PSIZE(bp) \
307 (ASSERT(BP_IS_EMBEDDED(bp)), \
308 BF64_GET_SB((bp)->blk_prop, 25, 7, 0, 1))
309 #define BPE_SET_PSIZE(bp, x) do { \
310 ASSERT(BP_IS_EMBEDDED(bp)); \
311 BF64_SET_SB((bp)->blk_prop, 25, 7, 0, 1, x); \
312 _NOTE(CONSTCOND) } while (0)
313
314 typedef enum bp_embedded_type {
315 BP_EMBEDDED_TYPE_DATA,
316 BP_EMBEDDED_TYPE_RESERVED, /* Reserved for an unintegrated feature. */
317 NUM_BP_EMBEDDED_TYPES = BP_EMBEDDED_TYPE_RESERVED
318 } bp_embedded_type_t;
319
320 #define BPE_NUM_WORDS 14
321 #define BPE_PAYLOAD_SIZE (BPE_NUM_WORDS * sizeof (uint64_t))
322 #define BPE_IS_PAYLOADWORD(bp, wp) \
323 ((wp) != &(bp)->blk_prop && (wp) != &(bp)->blk_birth)
324
325 #define SPA_BLKPTRSHIFT 7 /* blkptr_t is 128 bytes */
326 #define SPA_DVAS_PER_BP 3 /* Number of DVAs in a bp */
327
328 /*
329 * A block is a hole when it has either 1) never been written to, or
330 * 2) is zero-filled. In both cases, ZFS can return all zeroes for all reads
331 * without physically allocating disk space. Holes are represented in the
332 * blkptr_t structure by zeroed blk_dva. Correct checking for holes is
333 * done through the BP_IS_HOLE macro. For holes, the logical size, level,
334 * DMU object type, and birth times are all also stored for holes that
335 * were written to at some point (i.e. were punched after having been filled).
336 */
337 typedef struct blkptr {
338 dva_t blk_dva[SPA_DVAS_PER_BP]; /* Data Virtual Addresses */
339 uint64_t blk_prop; /* size, compression, type, etc */
340 uint64_t blk_pad[2]; /* Extra space for the future */
341 uint64_t blk_phys_birth; /* txg when block was allocated */
342 uint64_t blk_birth; /* transaction group at birth */
343 uint64_t blk_fill; /* fill count */
344 zio_cksum_t blk_cksum; /* 256-bit checksum */
345 } blkptr_t;
346
347 /*
348 * Macros to get and set fields in a bp or DVA.
349 */
350 #define DVA_GET_ASIZE(dva) \
351 BF64_GET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, SPA_MINBLOCKSHIFT, 0)
352 #define DVA_SET_ASIZE(dva, x) \
353 BF64_SET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, \
354 SPA_MINBLOCKSHIFT, 0, x)
355
356 #define DVA_GET_GRID(dva) BF64_GET((dva)->dva_word[0], 24, 8)
357 #define DVA_SET_GRID(dva, x) BF64_SET((dva)->dva_word[0], 24, 8, x)
358
359 #define DVA_GET_VDEV(dva) BF64_GET((dva)->dva_word[0], 32, 32)
360 #define DVA_SET_VDEV(dva, x) BF64_SET((dva)->dva_word[0], 32, 32, x)
361
362 #define DVA_GET_OFFSET(dva) \
363 BF64_GET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0)
364 #define DVA_SET_OFFSET(dva, x) \
365 BF64_SET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0, x)
366
367 #define DVA_GET_GANG(dva) BF64_GET((dva)->dva_word[1], 63, 1)
368 #define DVA_SET_GANG(dva, x) BF64_SET((dva)->dva_word[1], 63, 1, x)
369
370 #define BP_GET_LSIZE(bp) \
371 (BP_IS_EMBEDDED(bp) ? \
372 (BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA ? BPE_GET_LSIZE(bp) : 0): \
373 BF64_GET_SB((bp)->blk_prop, 0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1))
374 #define BP_SET_LSIZE(bp, x) do { \
375 ASSERT(!BP_IS_EMBEDDED(bp)); \
376 BF64_SET_SB((bp)->blk_prop, \
377 0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \
378 _NOTE(CONSTCOND) } while (0)
379
380 #define BP_GET_PSIZE(bp) \
381 (BP_IS_EMBEDDED(bp) ? 0 : \
382 BF64_GET_SB((bp)->blk_prop, 16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1))
383 #define BP_SET_PSIZE(bp, x) do { \
384 ASSERT(!BP_IS_EMBEDDED(bp)); \
385 BF64_SET_SB((bp)->blk_prop, \
386 16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \
387 _NOTE(CONSTCOND) } while (0)
388
389 #define BP_GET_COMPRESS(bp) \
390 BF64_GET((bp)->blk_prop, 32, SPA_COMPRESSBITS)
391 #define BP_SET_COMPRESS(bp, x) \
392 BF64_SET((bp)->blk_prop, 32, SPA_COMPRESSBITS, x)
393
394 #define BP_IS_EMBEDDED(bp) BF64_GET((bp)->blk_prop, 39, 1)
395 #define BP_SET_EMBEDDED(bp, x) BF64_SET((bp)->blk_prop, 39, 1, x)
396
397 #define BP_GET_CHECKSUM(bp) \
398 (BP_IS_EMBEDDED(bp) ? ZIO_CHECKSUM_OFF : \
399 BF64_GET((bp)->blk_prop, 40, 8))
400 #define BP_SET_CHECKSUM(bp, x) do { \
401 ASSERT(!BP_IS_EMBEDDED(bp)); \
402 BF64_SET((bp)->blk_prop, 40, 8, x); \
403 _NOTE(CONSTCOND) } while (0)
404
405 #define BP_GET_TYPE(bp) BF64_GET((bp)->blk_prop, 48, 8)
406 #define BP_SET_TYPE(bp, x) BF64_SET((bp)->blk_prop, 48, 8, x)
407
408 #define BP_GET_LEVEL(bp) BF64_GET((bp)->blk_prop, 56, 5)
409 #define BP_SET_LEVEL(bp, x) BF64_SET((bp)->blk_prop, 56, 5, x)
410
411 #define BP_GET_DEDUP(bp) BF64_GET((bp)->blk_prop, 62, 1)
412 #define BP_SET_DEDUP(bp, x) BF64_SET((bp)->blk_prop, 62, 1, x)
413
414 #define BP_GET_BYTEORDER(bp) BF64_GET((bp)->blk_prop, 63, 1)
415 #define BP_SET_BYTEORDER(bp, x) BF64_SET((bp)->blk_prop, 63, 1, x)
416
417 #define BP_PHYSICAL_BIRTH(bp) \
418 (BP_IS_EMBEDDED(bp) ? 0 : \
419 (bp)->blk_phys_birth ? (bp)->blk_phys_birth : (bp)->blk_birth)
420
421 #define BP_SET_BIRTH(bp, logical, physical) \
422 { \
423 ASSERT(!BP_IS_EMBEDDED(bp)); \
424 (bp)->blk_birth = (logical); \
425 (bp)->blk_phys_birth = ((logical) == (physical) ? 0 : (physical)); \
426 }
427
428 #define BP_GET_FILL(bp) (BP_IS_EMBEDDED(bp) ? 1 : (bp)->blk_fill)
429
430 #define BP_IS_METADATA(bp) \
431 (BP_GET_LEVEL(bp) > 0 || DMU_OT_IS_METADATA(BP_GET_TYPE(bp)))
432
433 #define BP_GET_ASIZE(bp) \
434 (BP_IS_EMBEDDED(bp) ? 0 : \
435 DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \
436 DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \
437 DVA_GET_ASIZE(&(bp)->blk_dva[2]))
438
439 #define BP_GET_UCSIZE(bp) \
440 (BP_IS_METADATA(bp) ? BP_GET_PSIZE(bp) : BP_GET_LSIZE(bp))
441
442 #define BP_GET_NDVAS(bp) \
443 (BP_IS_EMBEDDED(bp) ? 0 : \
444 !!DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \
445 !!DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \
446 !!DVA_GET_ASIZE(&(bp)->blk_dva[2]))
447
448 #define BP_COUNT_GANG(bp) \
449 (BP_IS_EMBEDDED(bp) ? 0 : \
450 (DVA_GET_GANG(&(bp)->blk_dva[0]) + \
451 DVA_GET_GANG(&(bp)->blk_dva[1]) + \
452 DVA_GET_GANG(&(bp)->blk_dva[2])))
453
454 #define DVA_EQUAL(dva1, dva2) \
455 ((dva1)->dva_word[1] == (dva2)->dva_word[1] && \
456 (dva1)->dva_word[0] == (dva2)->dva_word[0])
457
458 #define BP_EQUAL(bp1, bp2) \
459 (BP_PHYSICAL_BIRTH(bp1) == BP_PHYSICAL_BIRTH(bp2) && \
460 (bp1)->blk_birth == (bp2)->blk_birth && \
461 DVA_EQUAL(&(bp1)->blk_dva[0], &(bp2)->blk_dva[0]) && \
462 DVA_EQUAL(&(bp1)->blk_dva[1], &(bp2)->blk_dva[1]) && \
463 DVA_EQUAL(&(bp1)->blk_dva[2], &(bp2)->blk_dva[2]))
464
465
466 #define DVA_IS_VALID(dva) (DVA_GET_ASIZE(dva) != 0)
467
468 #define BP_IDENTITY(bp) (ASSERT(!BP_IS_EMBEDDED(bp)), &(bp)->blk_dva[0])
469 #define BP_IS_GANG(bp) \
470 (BP_IS_EMBEDDED(bp) ? B_FALSE : DVA_GET_GANG(BP_IDENTITY(bp)))
471 #define DVA_IS_EMPTY(dva) ((dva)->dva_word[0] == 0ULL && \
472 (dva)->dva_word[1] == 0ULL)
473 #define BP_IS_HOLE(bp) \
474 (!BP_IS_EMBEDDED(bp) && DVA_IS_EMPTY(BP_IDENTITY(bp)))
475
476 /* BP_IS_RAIDZ(bp) assumes no block compression */
477 #define BP_IS_RAIDZ(bp) (DVA_GET_ASIZE(&(bp)->blk_dva[0]) > \
478 BP_GET_PSIZE(bp))
479
480 #define BP_ZERO(bp) \
481 { \
482 (bp)->blk_dva[0].dva_word[0] = 0; \
483 (bp)->blk_dva[0].dva_word[1] = 0; \
484 (bp)->blk_dva[1].dva_word[0] = 0; \
485 (bp)->blk_dva[1].dva_word[1] = 0; \
486 (bp)->blk_dva[2].dva_word[0] = 0; \
487 (bp)->blk_dva[2].dva_word[1] = 0; \
488 (bp)->blk_prop = 0; \
489 (bp)->blk_pad[0] = 0; \
490 (bp)->blk_pad[1] = 0; \
491 (bp)->blk_phys_birth = 0; \
492 (bp)->blk_birth = 0; \
493 (bp)->blk_fill = 0; \
494 ZIO_SET_CHECKSUM(&(bp)->blk_cksum, 0, 0, 0, 0); \
495 }
496
497 #ifdef _BIG_ENDIAN
498 #define ZFS_HOST_BYTEORDER (0ULL)
499 #else
500 #define ZFS_HOST_BYTEORDER (1ULL)
501 #endif
502
503 #define BP_SHOULD_BYTESWAP(bp) (BP_GET_BYTEORDER(bp) != ZFS_HOST_BYTEORDER)
504
505 #define BP_SPRINTF_LEN 320
506
507 /*
508 * This macro allows code sharing between zfs, libzpool, and mdb.
509 * 'func' is either snprintf() or mdb_snprintf().
510 * 'ws' (whitespace) can be ' ' for single-line format, '\n' for multi-line.
511 */
512 #define SNPRINTF_BLKPTR(func, ws, buf, size, bp, type, checksum, compress) \
513 { \
514 static const char *copyname[] = \
515 { "zero", "single", "double", "triple" }; \
516 int len = 0; \
517 int copies = 0; \
518 int d; \
519 \
520 if (bp == NULL) { \
521 len += func(buf + len, size - len, "<NULL>"); \
522 } else if (BP_IS_HOLE(bp)) { \
523 len += func(buf + len, size - len, \
524 "HOLE [L%llu %s] " \
525 "size=%llxL birth=%lluL", \
526 (u_longlong_t)BP_GET_LEVEL(bp), \
527 type, \
528 (u_longlong_t)BP_GET_LSIZE(bp), \
529 (u_longlong_t)bp->blk_birth); \
530 } else if (BP_IS_EMBEDDED(bp)) { \
531 len = func(buf + len, size - len, \
532 "EMBEDDED [L%llu %s] et=%u %s " \
533 "size=%llxL/%llxP birth=%lluL", \
534 (u_longlong_t)BP_GET_LEVEL(bp), \
535 type, \
536 (int)BPE_GET_ETYPE(bp), \
537 compress, \
538 (u_longlong_t)BPE_GET_LSIZE(bp), \
539 (u_longlong_t)BPE_GET_PSIZE(bp), \
540 (u_longlong_t)bp->blk_birth); \
541 } else { \
542 for (d = 0; d < BP_GET_NDVAS(bp); d++) { \
543 const dva_t *dva = &bp->blk_dva[d]; \
544 if (DVA_IS_VALID(dva)) \
545 copies++; \
546 len += func(buf + len, size - len, \
547 "DVA[%d]=<%llu:%llx:%llx>%c", d, \
548 (u_longlong_t)DVA_GET_VDEV(dva), \
549 (u_longlong_t)DVA_GET_OFFSET(dva), \
550 (u_longlong_t)DVA_GET_ASIZE(dva), \
551 ws); \
552 } \
553 if (BP_IS_GANG(bp) && \
554 DVA_GET_ASIZE(&bp->blk_dva[2]) <= \
555 DVA_GET_ASIZE(&bp->blk_dva[1]) / 2) \
556 copies--; \
557 len += func(buf + len, size - len, \
558 "[L%llu %s] %s %s %s %s %s %s%c" \
559 "size=%llxL/%llxP birth=%lluL/%lluP fill=%llu%c" \
560 "cksum=%llx:%llx:%llx:%llx", \
561 (u_longlong_t)BP_GET_LEVEL(bp), \
562 type, \
563 checksum, \
564 compress, \
565 BP_GET_BYTEORDER(bp) == 0 ? "BE" : "LE", \
566 BP_IS_GANG(bp) ? "gang" : "contiguous", \
567 BP_GET_DEDUP(bp) ? "dedup" : "unique", \
568 copyname[copies], \
569 ws, \
570 (u_longlong_t)BP_GET_LSIZE(bp), \
571 (u_longlong_t)BP_GET_PSIZE(bp), \
572 (u_longlong_t)bp->blk_birth, \
573 (u_longlong_t)BP_PHYSICAL_BIRTH(bp), \
574 (u_longlong_t)BP_GET_FILL(bp), \
575 ws, \
576 (u_longlong_t)bp->blk_cksum.zc_word[0], \
577 (u_longlong_t)bp->blk_cksum.zc_word[1], \
578 (u_longlong_t)bp->blk_cksum.zc_word[2], \
579 (u_longlong_t)bp->blk_cksum.zc_word[3]); \
580 } \
581 ASSERT(len < size); \
582 }
583
584 #define BP_GET_BUFC_TYPE(bp) \
585 (BP_IS_METADATA(bp) ? ARC_BUFC_METADATA : ARC_BUFC_DATA)
586
587 typedef enum spa_import_type {
588 SPA_IMPORT_EXISTING,
589 SPA_IMPORT_ASSEMBLE
590 } spa_import_type_t;
591
592 /* state manipulation functions */
593 extern int spa_open(const char *pool, spa_t **, void *tag);
594 extern int spa_open_rewind(const char *pool, spa_t **, void *tag,
595 nvlist_t *policy, nvlist_t **config);
596 extern int spa_get_stats(const char *pool, nvlist_t **config, char *altroot,
597 size_t buflen);
598 extern int spa_create(const char *pool, nvlist_t *config, nvlist_t *props,
599 nvlist_t *zplprops);
600 extern int spa_import(char *pool, nvlist_t *config, nvlist_t *props,
601 uint64_t flags);
602 extern nvlist_t *spa_tryimport(nvlist_t *tryconfig);
603 extern int spa_destroy(char *pool);
604 extern int spa_export(char *pool, nvlist_t **oldconfig, boolean_t force,
605 boolean_t hardforce);
606 extern int spa_reset(char *pool);
607 extern void spa_async_request(spa_t *spa, int flag);
608 extern void spa_async_unrequest(spa_t *spa, int flag);
609 extern void spa_async_suspend(spa_t *spa);
610 extern void spa_async_resume(spa_t *spa);
611 extern spa_t *spa_inject_addref(char *pool);
612 extern void spa_inject_delref(spa_t *spa);
613 extern void spa_scan_stat_init(spa_t *spa);
614 extern int spa_scan_get_stats(spa_t *spa, pool_scan_stat_t *ps);
615
616 #define SPA_ASYNC_CONFIG_UPDATE 0x01
617 #define SPA_ASYNC_REMOVE 0x02
618 #define SPA_ASYNC_PROBE 0x04
619 #define SPA_ASYNC_RESILVER_DONE 0x08
620 #define SPA_ASYNC_RESILVER 0x10
621 #define SPA_ASYNC_AUTOEXPAND 0x20
622 #define SPA_ASYNC_REMOVE_DONE 0x40
623 #define SPA_ASYNC_REMOVE_STOP 0x80
624
625 /*
626 * Controls the behavior of spa_vdev_remove().
627 */
628 #define SPA_REMOVE_UNSPARE 0x01
629 #define SPA_REMOVE_DONE 0x02
630
631 /* device manipulation */
632 extern int spa_vdev_add(spa_t *spa, nvlist_t *nvroot);
633 extern int spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot,
634 int replacing);
635 extern int spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid,
636 int replace_done);
637 extern int spa_vdev_remove(spa_t *spa, uint64_t guid, boolean_t unspare);
638 extern boolean_t spa_vdev_remove_active(spa_t *spa);
639 extern int spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath);
640 extern int spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru);
641 extern int spa_vdev_split_mirror(spa_t *spa, char *newname, nvlist_t *config,
642 nvlist_t *props, boolean_t exp);
643
644 /* spare state (which is global across all pools) */
645 extern void spa_spare_add(vdev_t *vd);
646 extern void spa_spare_remove(vdev_t *vd);
647 extern boolean_t spa_spare_exists(uint64_t guid, uint64_t *pool, int *refcnt);
648 extern void spa_spare_activate(vdev_t *vd);
649
650 /* L2ARC state (which is global across all pools) */
651 extern void spa_l2cache_add(vdev_t *vd);
652 extern void spa_l2cache_remove(vdev_t *vd);
653 extern boolean_t spa_l2cache_exists(uint64_t guid, uint64_t *pool);
654 extern void spa_l2cache_activate(vdev_t *vd);
655 extern void spa_l2cache_drop(spa_t *spa);
656
657 /* scanning */
658 extern int spa_scan(spa_t *spa, pool_scan_func_t func);
659 extern int spa_scan_stop(spa_t *spa);
660
661 /* spa syncing */
662 extern void spa_sync(spa_t *spa, uint64_t txg); /* only for DMU use */
663 extern void spa_sync_allpools(void);
664
665 extern int zfs_sync_pass_deferred_free;
666
667 /* spa namespace global mutex */
668 extern kmutex_t spa_namespace_lock;
669
670 /*
671 * SPA configuration functions in spa_config.c
672 */
673
674 #define SPA_CONFIG_UPDATE_POOL 0
675 #define SPA_CONFIG_UPDATE_VDEVS 1
676
677 extern void spa_config_sync(spa_t *, boolean_t, boolean_t);
678 extern void spa_config_load(void);
679 extern nvlist_t *spa_all_configs(uint64_t *);
680 extern void spa_config_set(spa_t *spa, nvlist_t *config);
681 extern nvlist_t *spa_config_generate(spa_t *spa, vdev_t *vd, uint64_t txg,
682 int getstats);
683 extern void spa_config_update(spa_t *spa, int what);
684
685 /*
686 * Miscellaneous SPA routines in spa_misc.c
687 */
688
689 /* Namespace manipulation */
690 extern spa_t *spa_lookup(const char *name);
691 extern spa_t *spa_add(const char *name, nvlist_t *config, const char *altroot);
692 extern void spa_remove(spa_t *spa);
693 extern spa_t *spa_next(spa_t *prev);
694
695 /* Refcount functions */
696 extern void spa_open_ref(spa_t *spa, void *tag);
697 extern void spa_close(spa_t *spa, void *tag);
698 extern void spa_async_close(spa_t *spa, void *tag);
699 extern boolean_t spa_refcount_zero(spa_t *spa);
700
701 #define SCL_NONE 0x00
702 #define SCL_CONFIG 0x01
703 #define SCL_STATE 0x02
704 #define SCL_L2ARC 0x04 /* hack until L2ARC 2.0 */
705 #define SCL_ALLOC 0x08
706 #define SCL_ZIO 0x10
707 #define SCL_FREE 0x20
708 #define SCL_VDEV 0x40
709 #define SCL_LOCKS 7
710 #define SCL_ALL ((1 << SCL_LOCKS) - 1)
711 #define SCL_STATE_ALL (SCL_STATE | SCL_L2ARC | SCL_ZIO)
712
713 /* Historical pool statistics */
714 typedef struct spa_stats_history {
715 kmutex_t lock;
716 uint64_t count;
717 uint64_t size;
718 kstat_t *kstat;
719 void *private;
720 list_t list;
721 } spa_stats_history_t;
722
723 typedef struct spa_stats {
724 spa_stats_history_t read_history;
725 spa_stats_history_t txg_history;
726 spa_stats_history_t tx_assign_histogram;
727 spa_stats_history_t io_history;
728 } spa_stats_t;
729
730 typedef enum txg_state {
731 TXG_STATE_BIRTH = 0,
732 TXG_STATE_OPEN = 1,
733 TXG_STATE_QUIESCED = 2,
734 TXG_STATE_WAIT_FOR_SYNC = 3,
735 TXG_STATE_SYNCED = 4,
736 TXG_STATE_COMMITTED = 5,
737 } txg_state_t;
738
739 typedef struct txg_stat {
740 vdev_stat_t vs1;
741 vdev_stat_t vs2;
742 uint64_t txg;
743 uint64_t ndirty;
744 } txg_stat_t;
745
746 extern void spa_stats_init(spa_t *spa);
747 extern void spa_stats_destroy(spa_t *spa);
748 extern void spa_read_history_add(spa_t *spa, const zbookmark_phys_t *zb,
749 uint32_t aflags);
750 extern void spa_txg_history_add(spa_t *spa, uint64_t txg, hrtime_t birth_time);
751 extern int spa_txg_history_set(spa_t *spa, uint64_t txg,
752 txg_state_t completed_state, hrtime_t completed_time);
753 extern txg_stat_t *spa_txg_history_init_io(spa_t *, uint64_t,
754 struct dsl_pool *);
755 extern void spa_txg_history_fini_io(spa_t *, txg_stat_t *);
756 extern void spa_tx_assign_add_nsecs(spa_t *spa, uint64_t nsecs);
757
758 /* Pool configuration locks */
759 extern int spa_config_tryenter(spa_t *spa, int locks, void *tag, krw_t rw);
760 extern void spa_config_enter(spa_t *spa, int locks, void *tag, krw_t rw);
761 extern void spa_config_exit(spa_t *spa, int locks, void *tag);
762 extern int spa_config_held(spa_t *spa, int locks, krw_t rw);
763
764 /* Pool vdev add/remove lock */
765 extern uint64_t spa_vdev_enter(spa_t *spa);
766 extern uint64_t spa_vdev_config_enter(spa_t *spa);
767 extern void spa_vdev_config_exit(spa_t *spa, vdev_t *vd, uint64_t txg,
768 int error, char *tag);
769 extern int spa_vdev_exit(spa_t *spa, vdev_t *vd, uint64_t txg, int error);
770
771 /* Pool vdev state change lock */
772 extern void spa_vdev_state_enter(spa_t *spa, int oplock);
773 extern int spa_vdev_state_exit(spa_t *spa, vdev_t *vd, int error);
774
775 /* Log state */
776 typedef enum spa_log_state {
777 SPA_LOG_UNKNOWN = 0, /* unknown log state */
778 SPA_LOG_MISSING, /* missing log(s) */
779 SPA_LOG_CLEAR, /* clear the log(s) */
780 SPA_LOG_GOOD, /* log(s) are good */
781 } spa_log_state_t;
782
783 extern spa_log_state_t spa_get_log_state(spa_t *spa);
784 extern void spa_set_log_state(spa_t *spa, spa_log_state_t state);
785 extern int spa_offline_log(spa_t *spa);
786
787 /* Log claim callback */
788 extern void spa_claim_notify(zio_t *zio);
789 extern void spa_deadman(void *);
790
791 /* Accessor functions */
792 extern boolean_t spa_shutting_down(spa_t *spa);
793 extern struct dsl_pool *spa_get_dsl(spa_t *spa);
794 extern boolean_t spa_is_initializing(spa_t *spa);
795 extern blkptr_t *spa_get_rootblkptr(spa_t *spa);
796 extern void spa_set_rootblkptr(spa_t *spa, const blkptr_t *bp);
797 extern void spa_altroot(spa_t *, char *, size_t);
798 extern int spa_sync_pass(spa_t *spa);
799 extern char *spa_name(spa_t *spa);
800 extern uint64_t spa_guid(spa_t *spa);
801 extern uint64_t spa_load_guid(spa_t *spa);
802 extern uint64_t spa_last_synced_txg(spa_t *spa);
803 extern uint64_t spa_first_txg(spa_t *spa);
804 extern uint64_t spa_syncing_txg(spa_t *spa);
805 extern uint64_t spa_version(spa_t *spa);
806 extern pool_state_t spa_state(spa_t *spa);
807 extern spa_load_state_t spa_load_state(spa_t *spa);
808 extern uint64_t spa_freeze_txg(spa_t *spa);
809 extern uint64_t spa_get_worst_case_asize(spa_t *spa, uint64_t lsize);
810 extern uint64_t spa_get_dspace(spa_t *spa);
811 extern uint64_t spa_get_slop_space(spa_t *spa);
812 extern void spa_update_dspace(spa_t *spa);
813 extern uint64_t spa_version(spa_t *spa);
814 extern boolean_t spa_deflate(spa_t *spa);
815 extern metaslab_class_t *spa_normal_class(spa_t *spa);
816 extern metaslab_class_t *spa_log_class(spa_t *spa);
817 extern void spa_evicting_os_register(spa_t *, objset_t *os);
818 extern void spa_evicting_os_deregister(spa_t *, objset_t *os);
819 extern void spa_evicting_os_wait(spa_t *spa);
820 extern int spa_max_replication(spa_t *spa);
821 extern int spa_prev_software_version(spa_t *spa);
822 extern uint8_t spa_get_failmode(spa_t *spa);
823 extern boolean_t spa_suspended(spa_t *spa);
824 extern uint64_t spa_bootfs(spa_t *spa);
825 extern uint64_t spa_delegation(spa_t *spa);
826 extern objset_t *spa_meta_objset(spa_t *spa);
827 extern uint64_t spa_deadman_synctime(spa_t *spa);
828
829 /* Miscellaneous support routines */
830 extern void spa_activate_mos_feature(spa_t *spa, const char *feature,
831 dmu_tx_t *tx);
832 extern void spa_deactivate_mos_feature(spa_t *spa, const char *feature);
833 extern int spa_rename(const char *oldname, const char *newname);
834 extern spa_t *spa_by_guid(uint64_t pool_guid, uint64_t device_guid);
835 extern boolean_t spa_guid_exists(uint64_t pool_guid, uint64_t device_guid);
836 extern char *spa_strdup(const char *);
837 extern void spa_strfree(char *);
838 extern uint64_t spa_get_random(uint64_t range);
839 extern uint64_t spa_generate_guid(spa_t *spa);
840 extern void snprintf_blkptr(char *buf, size_t buflen, const blkptr_t *bp);
841 extern void spa_freeze(spa_t *spa);
842 extern int spa_change_guid(spa_t *spa);
843 extern void spa_upgrade(spa_t *spa, uint64_t version);
844 extern void spa_evict_all(void);
845 extern vdev_t *spa_lookup_by_guid(spa_t *spa, uint64_t guid,
846 boolean_t l2cache);
847 extern boolean_t spa_has_spare(spa_t *, uint64_t guid);
848 extern uint64_t dva_get_dsize_sync(spa_t *spa, const dva_t *dva);
849 extern uint64_t bp_get_dsize_sync(spa_t *spa, const blkptr_t *bp);
850 extern uint64_t bp_get_dsize(spa_t *spa, const blkptr_t *bp);
851 extern boolean_t spa_has_slogs(spa_t *spa);
852 extern boolean_t spa_is_root(spa_t *spa);
853 extern boolean_t spa_writeable(spa_t *spa);
854 extern boolean_t spa_has_pending_synctask(spa_t *spa);
855 extern int spa_maxblocksize(spa_t *spa);
856 extern int spa_maxdnodesize(spa_t *spa);
857 extern void zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp);
858
859 extern int spa_mode(spa_t *spa);
860 extern uint64_t strtonum(const char *str, char **nptr);
861
862 extern char *spa_his_ievent_table[];
863
864 extern void spa_history_create_obj(spa_t *spa, dmu_tx_t *tx);
865 extern int spa_history_get(spa_t *spa, uint64_t *offset, uint64_t *len_read,
866 char *his_buf);
867 extern int spa_history_log(spa_t *spa, const char *his_buf);
868 extern int spa_history_log_nvl(spa_t *spa, nvlist_t *nvl);
869 extern void spa_history_log_version(spa_t *spa, const char *operation);
870 extern void spa_history_log_internal(spa_t *spa, const char *operation,
871 dmu_tx_t *tx, const char *fmt, ...);
872 extern void spa_history_log_internal_ds(struct dsl_dataset *ds, const char *op,
873 dmu_tx_t *tx, const char *fmt, ...);
874 extern void spa_history_log_internal_dd(dsl_dir_t *dd, const char *operation,
875 dmu_tx_t *tx, const char *fmt, ...);
876
877 /* error handling */
878 struct zbookmark_phys;
879 extern void spa_log_error(spa_t *spa, zio_t *zio);
880 extern void zfs_ereport_post(const char *class, spa_t *spa, vdev_t *vd,
881 zio_t *zio, uint64_t stateoroffset, uint64_t length);
882 extern void zfs_post_remove(spa_t *spa, vdev_t *vd);
883 extern void zfs_post_state_change(spa_t *spa, vdev_t *vd, uint64_t laststate);
884 extern void zfs_post_autoreplace(spa_t *spa, vdev_t *vd);
885 extern void zfs_post_sysevent(spa_t *spa, vdev_t *vd, const char *name);
886 extern uint64_t spa_get_errlog_size(spa_t *spa);
887 extern int spa_get_errlog(spa_t *spa, void *uaddr, size_t *count);
888 extern void spa_errlog_rotate(spa_t *spa);
889 extern void spa_errlog_drain(spa_t *spa);
890 extern void spa_errlog_sync(spa_t *spa, uint64_t txg);
891 extern void spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub);
892
893 /* vdev cache */
894 extern void vdev_cache_stat_init(void);
895 extern void vdev_cache_stat_fini(void);
896
897 /* Initialization and termination */
898 extern void spa_init(int flags);
899 extern void spa_fini(void);
900 extern void spa_boot_init(void);
901
902 /* properties */
903 extern int spa_prop_set(spa_t *spa, nvlist_t *nvp);
904 extern int spa_prop_get(spa_t *spa, nvlist_t **nvp);
905 extern void spa_prop_clear_bootfs(spa_t *spa, uint64_t obj, dmu_tx_t *tx);
906 extern void spa_configfile_set(spa_t *, nvlist_t *, boolean_t);
907
908 /* asynchronous event notification */
909 extern void spa_event_notify(spa_t *spa, vdev_t *vdev, const char *name);
910
911 #ifdef ZFS_DEBUG
912 #define dprintf_bp(bp, fmt, ...) do { \
913 if (zfs_flags & ZFS_DEBUG_DPRINTF) { \
914 char *__blkbuf = kmem_alloc(BP_SPRINTF_LEN, KM_SLEEP); \
915 snprintf_blkptr(__blkbuf, BP_SPRINTF_LEN, (bp)); \
916 dprintf(fmt " %s\n", __VA_ARGS__, __blkbuf); \
917 kmem_free(__blkbuf, BP_SPRINTF_LEN); \
918 } \
919 _NOTE(CONSTCOND) } while (0)
920 #else
921 #define dprintf_bp(bp, fmt, ...)
922 #endif
923
924 extern boolean_t spa_debug_enabled(spa_t *spa);
925 #define spa_dbgmsg(spa, ...) \
926 { \
927 if (spa_debug_enabled(spa)) \
928 zfs_dbgmsg(__VA_ARGS__); \
929 }
930
931 extern int spa_mode_global; /* mode, e.g. FREAD | FWRITE */
932
933 #ifdef __cplusplus
934 }
935 #endif
936
937 #endif /* _SYS_SPA_H */