]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
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 | /* | |
428870ff | 22 | * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. |
be89734a | 23 | * Copyright (c) 2011, 2019 by Delphix. All rights reserved. |
a38718a6 | 24 | * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved. |
cc99f275 | 25 | * Copyright (c) 2017, Intel Corporation. |
34dc7c2f BB |
26 | */ |
27 | ||
f1512ee6 | 28 | #include <sys/sysmacros.h> |
34dc7c2f BB |
29 | #include <sys/zfs_context.h> |
30 | #include <sys/fm/fs/zfs.h> | |
31 | #include <sys/spa.h> | |
32 | #include <sys/txg.h> | |
33 | #include <sys/spa_impl.h> | |
34 | #include <sys/vdev_impl.h> | |
1b939560 | 35 | #include <sys/vdev_trim.h> |
34dc7c2f BB |
36 | #include <sys/zio_impl.h> |
37 | #include <sys/zio_compress.h> | |
38 | #include <sys/zio_checksum.h> | |
428870ff BB |
39 | #include <sys/dmu_objset.h> |
40 | #include <sys/arc.h> | |
41 | #include <sys/ddt.h> | |
9b67f605 | 42 | #include <sys/blkptr.h> |
b0bc7a84 | 43 | #include <sys/zfeature.h> |
d4a72f23 | 44 | #include <sys/dsl_scan.h> |
3dfb57a3 | 45 | #include <sys/metaslab_impl.h> |
193a37cb | 46 | #include <sys/time.h> |
d6662068 | 47 | #include <sys/trace_defs.h> |
a6255b7f | 48 | #include <sys/abd.h> |
b5256303 | 49 | #include <sys/dsl_crypt.h> |
492f64e9 | 50 | #include <sys/cityhash.h> |
34dc7c2f | 51 | |
34dc7c2f BB |
52 | /* |
53 | * ========================================================================== | |
54 | * I/O type descriptions | |
55 | * ========================================================================== | |
56 | */ | |
e8b96c60 | 57 | const char *zio_type_name[ZIO_TYPES] = { |
3dfb57a3 DB |
58 | /* |
59 | * Note: Linux kernel thread name length is limited | |
60 | * so these names will differ from upstream open zfs. | |
61 | */ | |
1b939560 | 62 | "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim" |
428870ff | 63 | }; |
34dc7c2f | 64 | |
27f2b90d | 65 | int zio_dva_throttle_enabled = B_TRUE; |
638dd5f4 | 66 | int zio_deadman_log_all = B_FALSE; |
3dfb57a3 | 67 | |
34dc7c2f BB |
68 | /* |
69 | * ========================================================================== | |
70 | * I/O kmem caches | |
71 | * ========================================================================== | |
72 | */ | |
73 | kmem_cache_t *zio_cache; | |
d164b209 | 74 | kmem_cache_t *zio_link_cache; |
34dc7c2f BB |
75 | kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; |
76 | kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
a6255b7f DQ |
77 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
78 | uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
79 | uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
80 | #endif | |
81 | ||
ad796b8a TH |
82 | /* Mark IOs as "slow" if they take longer than 30 seconds */ |
83 | int zio_slow_io_ms = (30 * MILLISEC); | |
34dc7c2f | 84 | |
fcff0f35 PD |
85 | #define BP_SPANB(indblkshift, level) \ |
86 | (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) | |
87 | #define COMPARE_META_LEVEL 0x80000000ul | |
55d85d5a GW |
88 | /* |
89 | * The following actions directly effect the spa's sync-to-convergence logic. | |
90 | * The values below define the sync pass when we start performing the action. | |
91 | * Care should be taken when changing these values as they directly impact | |
92 | * spa_sync() performance. Tuning these values may introduce subtle performance | |
93 | * pathologies and should only be done in the context of performance analysis. | |
94 | * These tunables will eventually be removed and replaced with #defines once | |
95 | * enough analysis has been done to determine optimal values. | |
96 | * | |
97 | * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that | |
98 | * regular blocks are not deferred. | |
be89734a MA |
99 | * |
100 | * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable | |
101 | * compression (including of metadata). In practice, we don't have this | |
102 | * many sync passes, so this has no effect. | |
103 | * | |
104 | * The original intent was that disabling compression would help the sync | |
105 | * passes to converge. However, in practice disabling compression increases | |
106 | * the average number of sync passes, because when we turn compression off, a | |
107 | * lot of block's size will change and thus we have to re-allocate (not | |
108 | * overwrite) them. It also increases the number of 128KB allocations (e.g. | |
109 | * for indirect blocks and spacemaps) because these will not be compressed. | |
110 | * The 128K allocations are especially detrimental to performance on highly | |
111 | * fragmented systems, which may have very few free segments of this size, | |
112 | * and may need to load new metaslabs to satisfy 128K allocations. | |
55d85d5a GW |
113 | */ |
114 | int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ | |
be89734a | 115 | int zfs_sync_pass_dont_compress = 8; /* don't compress starting in this pass */ |
55d85d5a GW |
116 | int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ |
117 | ||
34dc7c2f | 118 | /* |
b128c09f BB |
119 | * An allocating zio is one that either currently has the DVA allocate |
120 | * stage set or will have it later in its lifetime. | |
34dc7c2f | 121 | */ |
428870ff BB |
122 | #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) |
123 | ||
c409e464 | 124 | int zio_requeue_io_start_cut_in_line = 1; |
428870ff BB |
125 | |
126 | #ifdef ZFS_DEBUG | |
127 | int zio_buf_debug_limit = 16384; | |
128 | #else | |
129 | int zio_buf_debug_limit = 0; | |
130 | #endif | |
34dc7c2f | 131 | |
da6b4005 NB |
132 | static inline void __zio_execute(zio_t *zio); |
133 | ||
3dfb57a3 DB |
134 | static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t); |
135 | ||
34dc7c2f BB |
136 | void |
137 | zio_init(void) | |
138 | { | |
139 | size_t c; | |
140 | vmem_t *data_alloc_arena = NULL; | |
141 | ||
3941503c BB |
142 | zio_cache = kmem_cache_create("zio_cache", |
143 | sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); | |
d164b209 | 144 | zio_link_cache = kmem_cache_create("zio_link_cache", |
6795a698 | 145 | sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); |
34dc7c2f BB |
146 | |
147 | /* | |
148 | * For small buffers, we want a cache for each multiple of | |
f1512ee6 MA |
149 | * SPA_MINBLOCKSIZE. For larger buffers, we want a cache |
150 | * for each quarter-power of 2. | |
34dc7c2f BB |
151 | */ |
152 | for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { | |
153 | size_t size = (c + 1) << SPA_MINBLOCKSHIFT; | |
154 | size_t p2 = size; | |
155 | size_t align = 0; | |
6442f3cf | 156 | size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0; |
34dc7c2f | 157 | |
34328f3c | 158 | #if defined(_ILP32) && defined(_KERNEL) |
f1512ee6 MA |
159 | /* |
160 | * Cache size limited to 1M on 32-bit platforms until ARC | |
161 | * buffers no longer require virtual address space. | |
162 | */ | |
163 | if (size > zfs_max_recordsize) | |
164 | break; | |
165 | #endif | |
166 | ||
167 | while (!ISP2(p2)) | |
34dc7c2f BB |
168 | p2 &= p2 - 1; |
169 | ||
498877ba MA |
170 | #ifndef _KERNEL |
171 | /* | |
172 | * If we are using watchpoints, put each buffer on its own page, | |
173 | * to eliminate the performance overhead of trapping to the | |
174 | * kernel when modifying a non-watched buffer that shares the | |
175 | * page with a watched buffer. | |
176 | */ | |
177 | if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) | |
178 | continue; | |
fcf64f45 BB |
179 | /* |
180 | * Here's the problem - on 4K native devices in userland on | |
181 | * Linux using O_DIRECT, buffers must be 4K aligned or I/O | |
182 | * will fail with EINVAL, causing zdb (and others) to coredump. | |
183 | * Since userland probably doesn't need optimized buffer caches, | |
184 | * we just force 4K alignment on everything. | |
185 | */ | |
186 | align = 8 * SPA_MINBLOCKSIZE; | |
187 | #else | |
24fa2034 | 188 | if (size < PAGESIZE) { |
34dc7c2f | 189 | align = SPA_MINBLOCKSIZE; |
498877ba | 190 | } else if (IS_P2ALIGNED(size, p2 >> 2)) { |
24fa2034 | 191 | align = PAGESIZE; |
34dc7c2f | 192 | } |
fcf64f45 | 193 | #endif |
34dc7c2f BB |
194 | |
195 | if (align != 0) { | |
196 | char name[36]; | |
197 | (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); | |
198 | zio_buf_cache[c] = kmem_cache_create(name, size, | |
6442f3cf | 199 | align, NULL, NULL, NULL, NULL, NULL, cflags); |
34dc7c2f BB |
200 | |
201 | (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); | |
202 | zio_data_buf_cache[c] = kmem_cache_create(name, size, | |
ae6ba3db | 203 | align, NULL, NULL, NULL, NULL, |
6442f3cf | 204 | data_alloc_arena, cflags); |
34dc7c2f BB |
205 | } |
206 | } | |
207 | ||
208 | while (--c != 0) { | |
209 | ASSERT(zio_buf_cache[c] != NULL); | |
210 | if (zio_buf_cache[c - 1] == NULL) | |
211 | zio_buf_cache[c - 1] = zio_buf_cache[c]; | |
212 | ||
213 | ASSERT(zio_data_buf_cache[c] != NULL); | |
214 | if (zio_data_buf_cache[c - 1] == NULL) | |
215 | zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; | |
216 | } | |
217 | ||
34dc7c2f | 218 | zio_inject_init(); |
9759c60f ED |
219 | |
220 | lz4_init(); | |
34dc7c2f BB |
221 | } |
222 | ||
223 | void | |
224 | zio_fini(void) | |
225 | { | |
226 | size_t c; | |
227 | kmem_cache_t *last_cache = NULL; | |
228 | kmem_cache_t *last_data_cache = NULL; | |
229 | ||
230 | for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { | |
f1512ee6 MA |
231 | #ifdef _ILP32 |
232 | /* | |
233 | * Cache size limited to 1M on 32-bit platforms until ARC | |
234 | * buffers no longer require virtual address space. | |
235 | */ | |
236 | if (((c + 1) << SPA_MINBLOCKSHIFT) > zfs_max_recordsize) | |
237 | break; | |
a6255b7f DQ |
238 | #endif |
239 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) | |
240 | if (zio_buf_cache_allocs[c] != zio_buf_cache_frees[c]) | |
241 | (void) printf("zio_fini: [%d] %llu != %llu\n", | |
242 | (int)((c + 1) << SPA_MINBLOCKSHIFT), | |
243 | (long long unsigned)zio_buf_cache_allocs[c], | |
244 | (long long unsigned)zio_buf_cache_frees[c]); | |
f1512ee6 | 245 | #endif |
34dc7c2f BB |
246 | if (zio_buf_cache[c] != last_cache) { |
247 | last_cache = zio_buf_cache[c]; | |
248 | kmem_cache_destroy(zio_buf_cache[c]); | |
249 | } | |
250 | zio_buf_cache[c] = NULL; | |
251 | ||
252 | if (zio_data_buf_cache[c] != last_data_cache) { | |
253 | last_data_cache = zio_data_buf_cache[c]; | |
254 | kmem_cache_destroy(zio_data_buf_cache[c]); | |
255 | } | |
256 | zio_data_buf_cache[c] = NULL; | |
257 | } | |
258 | ||
d164b209 | 259 | kmem_cache_destroy(zio_link_cache); |
34dc7c2f BB |
260 | kmem_cache_destroy(zio_cache); |
261 | ||
262 | zio_inject_fini(); | |
9759c60f ED |
263 | |
264 | lz4_fini(); | |
34dc7c2f BB |
265 | } |
266 | ||
267 | /* | |
268 | * ========================================================================== | |
269 | * Allocate and free I/O buffers | |
270 | * ========================================================================== | |
271 | */ | |
272 | ||
273 | /* | |
274 | * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a | |
275 | * crashdump if the kernel panics, so use it judiciously. Obviously, it's | |
276 | * useful to inspect ZFS metadata, but if possible, we should avoid keeping | |
277 | * excess / transient data in-core during a crashdump. | |
278 | */ | |
279 | void * | |
280 | zio_buf_alloc(size_t size) | |
281 | { | |
282 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
283 | ||
63e3a861 | 284 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
a6255b7f DQ |
285 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
286 | atomic_add_64(&zio_buf_cache_allocs[c], 1); | |
287 | #endif | |
34dc7c2f | 288 | |
efcd79a8 | 289 | return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); |
34dc7c2f BB |
290 | } |
291 | ||
292 | /* | |
293 | * Use zio_data_buf_alloc to allocate data. The data will not appear in a | |
294 | * crashdump if the kernel panics. This exists so that we will limit the amount | |
295 | * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount | |
296 | * of kernel heap dumped to disk when the kernel panics) | |
297 | */ | |
298 | void * | |
299 | zio_data_buf_alloc(size_t size) | |
300 | { | |
301 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
302 | ||
63e3a861 | 303 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f | 304 | |
efcd79a8 | 305 | return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); |
34dc7c2f BB |
306 | } |
307 | ||
308 | void | |
309 | zio_buf_free(void *buf, size_t size) | |
310 | { | |
311 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
312 | ||
63e3a861 | 313 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
a6255b7f DQ |
314 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
315 | atomic_add_64(&zio_buf_cache_frees[c], 1); | |
316 | #endif | |
34dc7c2f BB |
317 | |
318 | kmem_cache_free(zio_buf_cache[c], buf); | |
319 | } | |
320 | ||
321 | void | |
322 | zio_data_buf_free(void *buf, size_t size) | |
323 | { | |
324 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
325 | ||
63e3a861 | 326 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f BB |
327 | |
328 | kmem_cache_free(zio_data_buf_cache[c], buf); | |
329 | } | |
330 | ||
84c07ada GN |
331 | static void |
332 | zio_abd_free(void *abd, size_t size) | |
333 | { | |
334 | abd_free((abd_t *)abd); | |
335 | } | |
336 | ||
34dc7c2f BB |
337 | /* |
338 | * ========================================================================== | |
339 | * Push and pop I/O transform buffers | |
340 | * ========================================================================== | |
341 | */ | |
d3c2ae1c | 342 | void |
a6255b7f | 343 | zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize, |
e9aa730c | 344 | zio_transform_func_t *transform) |
34dc7c2f | 345 | { |
79c76d5b | 346 | zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); |
34dc7c2f | 347 | |
a6255b7f | 348 | zt->zt_orig_abd = zio->io_abd; |
b128c09f | 349 | zt->zt_orig_size = zio->io_size; |
34dc7c2f | 350 | zt->zt_bufsize = bufsize; |
b128c09f | 351 | zt->zt_transform = transform; |
34dc7c2f BB |
352 | |
353 | zt->zt_next = zio->io_transform_stack; | |
354 | zio->io_transform_stack = zt; | |
355 | ||
a6255b7f | 356 | zio->io_abd = data; |
34dc7c2f BB |
357 | zio->io_size = size; |
358 | } | |
359 | ||
d3c2ae1c | 360 | void |
b128c09f | 361 | zio_pop_transforms(zio_t *zio) |
34dc7c2f | 362 | { |
b128c09f BB |
363 | zio_transform_t *zt; |
364 | ||
365 | while ((zt = zio->io_transform_stack) != NULL) { | |
366 | if (zt->zt_transform != NULL) | |
367 | zt->zt_transform(zio, | |
a6255b7f | 368 | zt->zt_orig_abd, zt->zt_orig_size); |
34dc7c2f | 369 | |
428870ff | 370 | if (zt->zt_bufsize != 0) |
a6255b7f | 371 | abd_free(zio->io_abd); |
34dc7c2f | 372 | |
a6255b7f | 373 | zio->io_abd = zt->zt_orig_abd; |
b128c09f BB |
374 | zio->io_size = zt->zt_orig_size; |
375 | zio->io_transform_stack = zt->zt_next; | |
34dc7c2f | 376 | |
b128c09f | 377 | kmem_free(zt, sizeof (zio_transform_t)); |
34dc7c2f BB |
378 | } |
379 | } | |
380 | ||
b128c09f BB |
381 | /* |
382 | * ========================================================================== | |
b5256303 | 383 | * I/O transform callbacks for subblocks, decompression, and decryption |
b128c09f BB |
384 | * ========================================================================== |
385 | */ | |
386 | static void | |
a6255b7f | 387 | zio_subblock(zio_t *zio, abd_t *data, uint64_t size) |
b128c09f BB |
388 | { |
389 | ASSERT(zio->io_size > size); | |
390 | ||
391 | if (zio->io_type == ZIO_TYPE_READ) | |
a6255b7f | 392 | abd_copy(data, zio->io_abd, size); |
b128c09f BB |
393 | } |
394 | ||
395 | static void | |
a6255b7f | 396 | zio_decompress(zio_t *zio, abd_t *data, uint64_t size) |
b128c09f | 397 | { |
a6255b7f DQ |
398 | if (zio->io_error == 0) { |
399 | void *tmp = abd_borrow_buf(data, size); | |
400 | int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), | |
401 | zio->io_abd, tmp, zio->io_size, size); | |
402 | abd_return_buf_copy(data, tmp, size); | |
403 | ||
c3bd3fb4 TC |
404 | if (zio_injection_enabled && ret == 0) |
405 | ret = zio_handle_fault_injection(zio, EINVAL); | |
406 | ||
a6255b7f DQ |
407 | if (ret != 0) |
408 | zio->io_error = SET_ERROR(EIO); | |
409 | } | |
b128c09f BB |
410 | } |
411 | ||
b5256303 TC |
412 | static void |
413 | zio_decrypt(zio_t *zio, abd_t *data, uint64_t size) | |
414 | { | |
415 | int ret; | |
416 | void *tmp; | |
417 | blkptr_t *bp = zio->io_bp; | |
ae76f45c TC |
418 | spa_t *spa = zio->io_spa; |
419 | uint64_t dsobj = zio->io_bookmark.zb_objset; | |
b5256303 TC |
420 | uint64_t lsize = BP_GET_LSIZE(bp); |
421 | dmu_object_type_t ot = BP_GET_TYPE(bp); | |
422 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
423 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
424 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
425 | boolean_t no_crypt = B_FALSE; | |
426 | ||
427 | ASSERT(BP_USES_CRYPT(bp)); | |
428 | ASSERT3U(size, !=, 0); | |
429 | ||
430 | if (zio->io_error != 0) | |
431 | return; | |
432 | ||
433 | /* | |
434 | * Verify the cksum of MACs stored in an indirect bp. It will always | |
435 | * be possible to verify this since it does not require an encryption | |
436 | * key. | |
437 | */ | |
438 | if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) { | |
439 | zio_crypt_decode_mac_bp(bp, mac); | |
440 | ||
441 | if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) { | |
442 | /* | |
443 | * We haven't decompressed the data yet, but | |
444 | * zio_crypt_do_indirect_mac_checksum() requires | |
445 | * decompressed data to be able to parse out the MACs | |
446 | * from the indirect block. We decompress it now and | |
447 | * throw away the result after we are finished. | |
448 | */ | |
449 | tmp = zio_buf_alloc(lsize); | |
450 | ret = zio_decompress_data(BP_GET_COMPRESS(bp), | |
451 | zio->io_abd, tmp, zio->io_size, lsize); | |
452 | if (ret != 0) { | |
453 | ret = SET_ERROR(EIO); | |
454 | goto error; | |
455 | } | |
456 | ret = zio_crypt_do_indirect_mac_checksum(B_FALSE, | |
457 | tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac); | |
458 | zio_buf_free(tmp, lsize); | |
459 | } else { | |
460 | ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE, | |
461 | zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac); | |
462 | } | |
463 | abd_copy(data, zio->io_abd, size); | |
464 | ||
be9a5c35 TC |
465 | if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) { |
466 | ret = zio_handle_decrypt_injection(spa, | |
467 | &zio->io_bookmark, ot, ECKSUM); | |
468 | } | |
b5256303 TC |
469 | if (ret != 0) |
470 | goto error; | |
471 | ||
472 | return; | |
473 | } | |
474 | ||
475 | /* | |
476 | * If this is an authenticated block, just check the MAC. It would be | |
477 | * nice to separate this out into its own flag, but for the moment | |
478 | * enum zio_flag is out of bits. | |
479 | */ | |
480 | if (BP_IS_AUTHENTICATED(bp)) { | |
481 | if (ot == DMU_OT_OBJSET) { | |
ae76f45c TC |
482 | ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, |
483 | dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp)); | |
b5256303 TC |
484 | } else { |
485 | zio_crypt_decode_mac_bp(bp, mac); | |
ae76f45c TC |
486 | ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, |
487 | zio->io_abd, size, mac); | |
be9a5c35 TC |
488 | if (zio_injection_enabled && ret == 0) { |
489 | ret = zio_handle_decrypt_injection(spa, | |
490 | &zio->io_bookmark, ot, ECKSUM); | |
491 | } | |
b5256303 TC |
492 | } |
493 | abd_copy(data, zio->io_abd, size); | |
494 | ||
495 | if (ret != 0) | |
496 | goto error; | |
497 | ||
498 | return; | |
499 | } | |
500 | ||
501 | zio_crypt_decode_params_bp(bp, salt, iv); | |
502 | ||
503 | if (ot == DMU_OT_INTENT_LOG) { | |
504 | tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t)); | |
505 | zio_crypt_decode_mac_zil(tmp, mac); | |
506 | abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t)); | |
507 | } else { | |
508 | zio_crypt_decode_mac_bp(bp, mac); | |
509 | } | |
510 | ||
be9a5c35 TC |
511 | ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp), |
512 | BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data, | |
513 | zio->io_abd, &no_crypt); | |
b5256303 TC |
514 | if (no_crypt) |
515 | abd_copy(data, zio->io_abd, size); | |
516 | ||
517 | if (ret != 0) | |
518 | goto error; | |
519 | ||
520 | return; | |
521 | ||
522 | error: | |
523 | /* assert that the key was found unless this was speculative */ | |
be9a5c35 | 524 | ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE)); |
b5256303 TC |
525 | |
526 | /* | |
527 | * If there was a decryption / authentication error return EIO as | |
528 | * the io_error. If this was not a speculative zio, create an ereport. | |
529 | */ | |
530 | if (ret == ECKSUM) { | |
a2c2ed1b | 531 | zio->io_error = SET_ERROR(EIO); |
b5256303 | 532 | if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) { |
be9a5c35 | 533 | spa_log_error(spa, &zio->io_bookmark); |
b5256303 | 534 | zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION, |
ae76f45c | 535 | spa, NULL, &zio->io_bookmark, zio, 0, 0); |
b5256303 TC |
536 | } |
537 | } else { | |
538 | zio->io_error = ret; | |
539 | } | |
540 | } | |
541 | ||
b128c09f BB |
542 | /* |
543 | * ========================================================================== | |
544 | * I/O parent/child relationships and pipeline interlocks | |
545 | * ========================================================================== | |
546 | */ | |
d164b209 | 547 | zio_t * |
3dfb57a3 | 548 | zio_walk_parents(zio_t *cio, zio_link_t **zl) |
d164b209 | 549 | { |
d164b209 | 550 | list_t *pl = &cio->io_parent_list; |
b128c09f | 551 | |
3dfb57a3 DB |
552 | *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl); |
553 | if (*zl == NULL) | |
d164b209 BB |
554 | return (NULL); |
555 | ||
3dfb57a3 DB |
556 | ASSERT((*zl)->zl_child == cio); |
557 | return ((*zl)->zl_parent); | |
d164b209 BB |
558 | } |
559 | ||
560 | zio_t * | |
3dfb57a3 | 561 | zio_walk_children(zio_t *pio, zio_link_t **zl) |
d164b209 | 562 | { |
d164b209 BB |
563 | list_t *cl = &pio->io_child_list; |
564 | ||
a8b2e306 TC |
565 | ASSERT(MUTEX_HELD(&pio->io_lock)); |
566 | ||
3dfb57a3 DB |
567 | *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl); |
568 | if (*zl == NULL) | |
d164b209 BB |
569 | return (NULL); |
570 | ||
3dfb57a3 DB |
571 | ASSERT((*zl)->zl_parent == pio); |
572 | return ((*zl)->zl_child); | |
d164b209 BB |
573 | } |
574 | ||
575 | zio_t * | |
576 | zio_unique_parent(zio_t *cio) | |
577 | { | |
3dfb57a3 DB |
578 | zio_link_t *zl = NULL; |
579 | zio_t *pio = zio_walk_parents(cio, &zl); | |
d164b209 | 580 | |
3dfb57a3 | 581 | VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL); |
d164b209 BB |
582 | return (pio); |
583 | } | |
584 | ||
585 | void | |
586 | zio_add_child(zio_t *pio, zio_t *cio) | |
b128c09f | 587 | { |
79c76d5b | 588 | zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); |
d164b209 BB |
589 | |
590 | /* | |
591 | * Logical I/Os can have logical, gang, or vdev children. | |
592 | * Gang I/Os can have gang or vdev children. | |
593 | * Vdev I/Os can only have vdev children. | |
594 | * The following ASSERT captures all of these constraints. | |
595 | */ | |
1ce23dca | 596 | ASSERT3S(cio->io_child_type, <=, pio->io_child_type); |
d164b209 BB |
597 | |
598 | zl->zl_parent = pio; | |
599 | zl->zl_child = cio; | |
600 | ||
b128c09f | 601 | mutex_enter(&pio->io_lock); |
a8b2e306 | 602 | mutex_enter(&cio->io_lock); |
d164b209 BB |
603 | |
604 | ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); | |
605 | ||
1c27024e | 606 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 BB |
607 | pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; |
608 | ||
609 | list_insert_head(&pio->io_child_list, zl); | |
610 | list_insert_head(&cio->io_parent_list, zl); | |
611 | ||
428870ff BB |
612 | pio->io_child_count++; |
613 | cio->io_parent_count++; | |
614 | ||
d164b209 | 615 | mutex_exit(&cio->io_lock); |
a8b2e306 | 616 | mutex_exit(&pio->io_lock); |
b128c09f BB |
617 | } |
618 | ||
34dc7c2f | 619 | static void |
d164b209 | 620 | zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) |
b128c09f | 621 | { |
d164b209 BB |
622 | ASSERT(zl->zl_parent == pio); |
623 | ASSERT(zl->zl_child == cio); | |
b128c09f BB |
624 | |
625 | mutex_enter(&pio->io_lock); | |
a8b2e306 | 626 | mutex_enter(&cio->io_lock); |
d164b209 BB |
627 | |
628 | list_remove(&pio->io_child_list, zl); | |
629 | list_remove(&cio->io_parent_list, zl); | |
630 | ||
428870ff BB |
631 | pio->io_child_count--; |
632 | cio->io_parent_count--; | |
633 | ||
d164b209 | 634 | mutex_exit(&cio->io_lock); |
a8b2e306 | 635 | mutex_exit(&pio->io_lock); |
d164b209 | 636 | kmem_cache_free(zio_link_cache, zl); |
b128c09f BB |
637 | } |
638 | ||
639 | static boolean_t | |
ddc751d5 | 640 | zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait) |
34dc7c2f | 641 | { |
b128c09f BB |
642 | boolean_t waiting = B_FALSE; |
643 | ||
644 | mutex_enter(&zio->io_lock); | |
645 | ASSERT(zio->io_stall == NULL); | |
ddc751d5 GW |
646 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) { |
647 | if (!(ZIO_CHILD_BIT_IS_SET(childbits, c))) | |
648 | continue; | |
649 | ||
650 | uint64_t *countp = &zio->io_children[c][wait]; | |
651 | if (*countp != 0) { | |
652 | zio->io_stage >>= 1; | |
653 | ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN); | |
654 | zio->io_stall = countp; | |
655 | waiting = B_TRUE; | |
656 | break; | |
657 | } | |
b128c09f BB |
658 | } |
659 | mutex_exit(&zio->io_lock); | |
b128c09f BB |
660 | return (waiting); |
661 | } | |
34dc7c2f | 662 | |
bf701a83 BB |
663 | __attribute__((always_inline)) |
664 | static inline void | |
62840030 MA |
665 | zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait, |
666 | zio_t **next_to_executep) | |
b128c09f BB |
667 | { |
668 | uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; | |
669 | int *errorp = &pio->io_child_error[zio->io_child_type]; | |
34dc7c2f | 670 | |
b128c09f BB |
671 | mutex_enter(&pio->io_lock); |
672 | if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) | |
673 | *errorp = zio_worst_error(*errorp, zio->io_error); | |
674 | pio->io_reexecute |= zio->io_reexecute; | |
675 | ASSERT3U(*countp, >, 0); | |
e8b96c60 MA |
676 | |
677 | (*countp)--; | |
678 | ||
679 | if (*countp == 0 && pio->io_stall == countp) { | |
3dfb57a3 DB |
680 | zio_taskq_type_t type = |
681 | pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE : | |
682 | ZIO_TASKQ_INTERRUPT; | |
b128c09f BB |
683 | pio->io_stall = NULL; |
684 | mutex_exit(&pio->io_lock); | |
62840030 | 685 | |
3dfb57a3 | 686 | /* |
62840030 MA |
687 | * If we can tell the caller to execute this parent next, do |
688 | * so. Otherwise dispatch the parent zio as its own task. | |
689 | * | |
690 | * Having the caller execute the parent when possible reduces | |
691 | * locking on the zio taskq's, reduces context switch | |
692 | * overhead, and has no recursion penalty. Note that one | |
693 | * read from disk typically causes at least 3 zio's: a | |
694 | * zio_null(), the logical zio_read(), and then a physical | |
695 | * zio. When the physical ZIO completes, we are able to call | |
696 | * zio_done() on all 3 of these zio's from one invocation of | |
697 | * zio_execute() by returning the parent back to | |
698 | * zio_execute(). Since the parent isn't executed until this | |
699 | * thread returns back to zio_execute(), the caller should do | |
700 | * so promptly. | |
701 | * | |
702 | * In other cases, dispatching the parent prevents | |
703 | * overflowing the stack when we have deeply nested | |
704 | * parent-child relationships, as we do with the "mega zio" | |
705 | * of writes for spa_sync(), and the chain of ZIL blocks. | |
3dfb57a3 | 706 | */ |
62840030 MA |
707 | if (next_to_executep != NULL && *next_to_executep == NULL) { |
708 | *next_to_executep = pio; | |
709 | } else { | |
710 | zio_taskq_dispatch(pio, type, B_FALSE); | |
711 | } | |
b128c09f BB |
712 | } else { |
713 | mutex_exit(&pio->io_lock); | |
34dc7c2f BB |
714 | } |
715 | } | |
716 | ||
b128c09f BB |
717 | static void |
718 | zio_inherit_child_errors(zio_t *zio, enum zio_child c) | |
719 | { | |
720 | if (zio->io_child_error[c] != 0 && zio->io_error == 0) | |
721 | zio->io_error = zio->io_child_error[c]; | |
722 | } | |
723 | ||
3dfb57a3 | 724 | int |
64fc7762 | 725 | zio_bookmark_compare(const void *x1, const void *x2) |
3dfb57a3 DB |
726 | { |
727 | const zio_t *z1 = x1; | |
728 | const zio_t *z2 = x2; | |
3dfb57a3 | 729 | |
64fc7762 MA |
730 | if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset) |
731 | return (-1); | |
732 | if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset) | |
733 | return (1); | |
3dfb57a3 | 734 | |
64fc7762 MA |
735 | if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object) |
736 | return (-1); | |
737 | if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object) | |
738 | return (1); | |
3dfb57a3 | 739 | |
64fc7762 MA |
740 | if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level) |
741 | return (-1); | |
742 | if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level) | |
743 | return (1); | |
744 | ||
745 | if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid) | |
746 | return (-1); | |
747 | if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid) | |
748 | return (1); | |
749 | ||
750 | if (z1 < z2) | |
751 | return (-1); | |
752 | if (z1 > z2) | |
753 | return (1); | |
754 | ||
755 | return (0); | |
3dfb57a3 DB |
756 | } |
757 | ||
34dc7c2f BB |
758 | /* |
759 | * ========================================================================== | |
b128c09f | 760 | * Create the various types of I/O (read, write, free, etc) |
34dc7c2f BB |
761 | * ========================================================================== |
762 | */ | |
763 | static zio_t * | |
428870ff | 764 | zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
a6255b7f | 765 | abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done, |
2aa34383 DK |
766 | void *private, zio_type_t type, zio_priority_t priority, |
767 | enum zio_flag flags, vdev_t *vd, uint64_t offset, | |
768 | const zbookmark_phys_t *zb, enum zio_stage stage, | |
769 | enum zio_stage pipeline) | |
34dc7c2f BB |
770 | { |
771 | zio_t *zio; | |
772 | ||
1b939560 | 773 | IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE); |
2aa34383 | 774 | ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); |
b128c09f BB |
775 | ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); |
776 | ||
777 | ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); | |
778 | ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); | |
779 | ASSERT(vd || stage == ZIO_STAGE_OPEN); | |
34dc7c2f | 780 | |
b5256303 | 781 | IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0); |
2aa34383 | 782 | |
79c76d5b | 783 | zio = kmem_cache_alloc(zio_cache, KM_SLEEP); |
3941503c BB |
784 | bzero(zio, sizeof (zio_t)); |
785 | ||
448d7aaa | 786 | mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL); |
3941503c BB |
787 | cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); |
788 | ||
789 | list_create(&zio->io_parent_list, sizeof (zio_link_t), | |
790 | offsetof(zio_link_t, zl_parent_node)); | |
791 | list_create(&zio->io_child_list, sizeof (zio_link_t), | |
792 | offsetof(zio_link_t, zl_child_node)); | |
4e21fd06 | 793 | metaslab_trace_init(&zio->io_alloc_list); |
d164b209 | 794 | |
b128c09f BB |
795 | if (vd != NULL) |
796 | zio->io_child_type = ZIO_CHILD_VDEV; | |
797 | else if (flags & ZIO_FLAG_GANG_CHILD) | |
798 | zio->io_child_type = ZIO_CHILD_GANG; | |
428870ff BB |
799 | else if (flags & ZIO_FLAG_DDT_CHILD) |
800 | zio->io_child_type = ZIO_CHILD_DDT; | |
b128c09f BB |
801 | else |
802 | zio->io_child_type = ZIO_CHILD_LOGICAL; | |
803 | ||
34dc7c2f | 804 | if (bp != NULL) { |
428870ff | 805 | zio->io_bp = (blkptr_t *)bp; |
34dc7c2f BB |
806 | zio->io_bp_copy = *bp; |
807 | zio->io_bp_orig = *bp; | |
428870ff BB |
808 | if (type != ZIO_TYPE_WRITE || |
809 | zio->io_child_type == ZIO_CHILD_DDT) | |
b128c09f | 810 | zio->io_bp = &zio->io_bp_copy; /* so caller can free */ |
9babb374 | 811 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) |
b128c09f | 812 | zio->io_logical = zio; |
9babb374 BB |
813 | if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) |
814 | pipeline |= ZIO_GANG_STAGES; | |
34dc7c2f | 815 | } |
b128c09f BB |
816 | |
817 | zio->io_spa = spa; | |
818 | zio->io_txg = txg; | |
34dc7c2f BB |
819 | zio->io_done = done; |
820 | zio->io_private = private; | |
821 | zio->io_type = type; | |
822 | zio->io_priority = priority; | |
b128c09f BB |
823 | zio->io_vd = vd; |
824 | zio->io_offset = offset; | |
a6255b7f | 825 | zio->io_orig_abd = zio->io_abd = data; |
2aa34383 DK |
826 | zio->io_orig_size = zio->io_size = psize; |
827 | zio->io_lsize = lsize; | |
b128c09f BB |
828 | zio->io_orig_flags = zio->io_flags = flags; |
829 | zio->io_orig_stage = zio->io_stage = stage; | |
830 | zio->io_orig_pipeline = zio->io_pipeline = pipeline; | |
3dfb57a3 | 831 | zio->io_pipeline_trace = ZIO_STAGE_OPEN; |
34dc7c2f | 832 | |
d164b209 BB |
833 | zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); |
834 | zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); | |
835 | ||
b128c09f BB |
836 | if (zb != NULL) |
837 | zio->io_bookmark = *zb; | |
838 | ||
839 | if (pio != NULL) { | |
cc99f275 DB |
840 | if (zio->io_metaslab_class == NULL) |
841 | zio->io_metaslab_class = pio->io_metaslab_class; | |
b128c09f | 842 | if (zio->io_logical == NULL) |
34dc7c2f | 843 | zio->io_logical = pio->io_logical; |
9babb374 BB |
844 | if (zio->io_child_type == ZIO_CHILD_GANG) |
845 | zio->io_gang_leader = pio->io_gang_leader; | |
b128c09f | 846 | zio_add_child(pio, zio); |
34dc7c2f BB |
847 | } |
848 | ||
a38718a6 GA |
849 | taskq_init_ent(&zio->io_tqent); |
850 | ||
34dc7c2f BB |
851 | return (zio); |
852 | } | |
853 | ||
854 | static void | |
b128c09f | 855 | zio_destroy(zio_t *zio) |
34dc7c2f | 856 | { |
4e21fd06 | 857 | metaslab_trace_fini(&zio->io_alloc_list); |
3941503c BB |
858 | list_destroy(&zio->io_parent_list); |
859 | list_destroy(&zio->io_child_list); | |
860 | mutex_destroy(&zio->io_lock); | |
861 | cv_destroy(&zio->io_cv); | |
b128c09f | 862 | kmem_cache_free(zio_cache, zio); |
34dc7c2f BB |
863 | } |
864 | ||
865 | zio_t * | |
d164b209 | 866 | zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, |
428870ff | 867 | void *private, enum zio_flag flags) |
34dc7c2f BB |
868 | { |
869 | zio_t *zio; | |
870 | ||
2aa34383 | 871 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
d164b209 | 872 | ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
b128c09f | 873 | ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); |
34dc7c2f BB |
874 | |
875 | return (zio); | |
876 | } | |
877 | ||
878 | zio_t * | |
428870ff | 879 | zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) |
34dc7c2f | 880 | { |
d164b209 | 881 | return (zio_null(NULL, spa, NULL, done, private, flags)); |
34dc7c2f BB |
882 | } |
883 | ||
dc04a8c7 PD |
884 | static void |
885 | zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp, boolean_t config_held) | |
63e3a861 | 886 | { |
63e3a861 MA |
887 | if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { |
888 | zfs_panic_recover("blkptr at %p has invalid TYPE %llu", | |
889 | bp, (longlong_t)BP_GET_TYPE(bp)); | |
890 | } | |
891 | if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS || | |
892 | BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) { | |
893 | zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu", | |
894 | bp, (longlong_t)BP_GET_CHECKSUM(bp)); | |
895 | } | |
896 | if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS || | |
897 | BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) { | |
898 | zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu", | |
899 | bp, (longlong_t)BP_GET_COMPRESS(bp)); | |
900 | } | |
901 | if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
902 | zfs_panic_recover("blkptr at %p has invalid LSIZE %llu", | |
903 | bp, (longlong_t)BP_GET_LSIZE(bp)); | |
904 | } | |
905 | if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
906 | zfs_panic_recover("blkptr at %p has invalid PSIZE %llu", | |
907 | bp, (longlong_t)BP_GET_PSIZE(bp)); | |
908 | } | |
909 | ||
910 | if (BP_IS_EMBEDDED(bp)) { | |
746d4a45 | 911 | if (BPE_GET_ETYPE(bp) >= NUM_BP_EMBEDDED_TYPES) { |
63e3a861 MA |
912 | zfs_panic_recover("blkptr at %p has invalid ETYPE %llu", |
913 | bp, (longlong_t)BPE_GET_ETYPE(bp)); | |
914 | } | |
915 | } | |
916 | ||
6cb8e530 PZ |
917 | /* |
918 | * Do not verify individual DVAs if the config is not trusted. This | |
919 | * will be done once the zio is executed in vdev_mirror_map_alloc. | |
920 | */ | |
921 | if (!spa->spa_trust_config) | |
922 | return; | |
923 | ||
dc04a8c7 PD |
924 | if (!config_held) |
925 | spa_config_enter(spa, SCL_VDEV, bp, RW_READER); | |
926 | else | |
927 | ASSERT(spa_config_held(spa, SCL_VDEV, RW_WRITER)); | |
63e3a861 MA |
928 | /* |
929 | * Pool-specific checks. | |
930 | * | |
931 | * Note: it would be nice to verify that the blk_birth and | |
932 | * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() | |
933 | * allows the birth time of log blocks (and dmu_sync()-ed blocks | |
934 | * that are in the log) to be arbitrarily large. | |
935 | */ | |
1c27024e | 936 | for (int i = 0; i < BP_GET_NDVAS(bp); i++) { |
63e3a861 | 937 | uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]); |
1c27024e | 938 | |
63e3a861 MA |
939 | if (vdevid >= spa->spa_root_vdev->vdev_children) { |
940 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
941 | "VDEV %llu", | |
942 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 943 | continue; |
63e3a861 | 944 | } |
1c27024e | 945 | vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; |
63e3a861 MA |
946 | if (vd == NULL) { |
947 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
948 | "VDEV %llu", | |
949 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 950 | continue; |
63e3a861 MA |
951 | } |
952 | if (vd->vdev_ops == &vdev_hole_ops) { | |
953 | zfs_panic_recover("blkptr at %p DVA %u has hole " | |
954 | "VDEV %llu", | |
955 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 956 | continue; |
63e3a861 MA |
957 | } |
958 | if (vd->vdev_ops == &vdev_missing_ops) { | |
959 | /* | |
960 | * "missing" vdevs are valid during import, but we | |
961 | * don't have their detailed info (e.g. asize), so | |
962 | * we can't perform any more checks on them. | |
963 | */ | |
964 | continue; | |
965 | } | |
1c27024e DB |
966 | uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]); |
967 | uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]); | |
63e3a861 MA |
968 | if (BP_IS_GANG(bp)) |
969 | asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
970 | if (offset + asize > vd->vdev_asize) { | |
971 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
972 | "OFFSET %llu", | |
973 | bp, i, (longlong_t)offset); | |
974 | } | |
975 | } | |
dc04a8c7 PD |
976 | if (!config_held) |
977 | spa_config_exit(spa, SCL_VDEV, bp); | |
63e3a861 MA |
978 | } |
979 | ||
6cb8e530 PZ |
980 | boolean_t |
981 | zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp) | |
982 | { | |
983 | uint64_t vdevid = DVA_GET_VDEV(dva); | |
984 | ||
985 | if (vdevid >= spa->spa_root_vdev->vdev_children) | |
986 | return (B_FALSE); | |
987 | ||
988 | vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; | |
989 | if (vd == NULL) | |
990 | return (B_FALSE); | |
991 | ||
992 | if (vd->vdev_ops == &vdev_hole_ops) | |
993 | return (B_FALSE); | |
994 | ||
995 | if (vd->vdev_ops == &vdev_missing_ops) { | |
996 | return (B_FALSE); | |
997 | } | |
998 | ||
999 | uint64_t offset = DVA_GET_OFFSET(dva); | |
1000 | uint64_t asize = DVA_GET_ASIZE(dva); | |
1001 | ||
1002 | if (BP_IS_GANG(bp)) | |
1003 | asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
1004 | if (offset + asize > vd->vdev_asize) | |
1005 | return (B_FALSE); | |
1006 | ||
1007 | return (B_TRUE); | |
1008 | } | |
1009 | ||
34dc7c2f | 1010 | zio_t * |
b128c09f | 1011 | zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, |
a6255b7f | 1012 | abd_t *data, uint64_t size, zio_done_func_t *done, void *private, |
5dbd68a3 | 1013 | zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) |
34dc7c2f BB |
1014 | { |
1015 | zio_t *zio; | |
1016 | ||
dc04a8c7 | 1017 | zfs_blkptr_verify(spa, bp, flags & ZIO_FLAG_CONFIG_WRITER); |
63e3a861 | 1018 | |
428870ff | 1019 | zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, |
2aa34383 | 1020 | data, size, size, done, private, |
b128c09f | 1021 | ZIO_TYPE_READ, priority, flags, NULL, 0, zb, |
428870ff BB |
1022 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
1023 | ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); | |
34dc7c2f | 1024 | |
b128c09f BB |
1025 | return (zio); |
1026 | } | |
34dc7c2f | 1027 | |
34dc7c2f | 1028 | zio_t * |
b128c09f | 1029 | zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, |
a6255b7f | 1030 | abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp, |
bc77ba73 PD |
1031 | zio_done_func_t *ready, zio_done_func_t *children_ready, |
1032 | zio_done_func_t *physdone, zio_done_func_t *done, | |
1033 | void *private, zio_priority_t priority, enum zio_flag flags, | |
1034 | const zbookmark_phys_t *zb) | |
34dc7c2f BB |
1035 | { |
1036 | zio_t *zio; | |
1037 | ||
b128c09f BB |
1038 | ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && |
1039 | zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && | |
1040 | zp->zp_compress >= ZIO_COMPRESS_OFF && | |
1041 | zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && | |
9ae529ec | 1042 | DMU_OT_IS_VALID(zp->zp_type) && |
b128c09f | 1043 | zp->zp_level < 32 && |
428870ff | 1044 | zp->zp_copies > 0 && |
03c6040b | 1045 | zp->zp_copies <= spa_max_replication(spa)); |
34dc7c2f | 1046 | |
2aa34383 | 1047 | zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private, |
b128c09f | 1048 | ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, |
428870ff BB |
1049 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
1050 | ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); | |
34dc7c2f BB |
1051 | |
1052 | zio->io_ready = ready; | |
bc77ba73 | 1053 | zio->io_children_ready = children_ready; |
e8b96c60 | 1054 | zio->io_physdone = physdone; |
b128c09f | 1055 | zio->io_prop = *zp; |
34dc7c2f | 1056 | |
9b67f605 MA |
1057 | /* |
1058 | * Data can be NULL if we are going to call zio_write_override() to | |
1059 | * provide the already-allocated BP. But we may need the data to | |
1060 | * verify a dedup hit (if requested). In this case, don't try to | |
b5256303 TC |
1061 | * dedup (just take the already-allocated BP verbatim). Encrypted |
1062 | * dedup blocks need data as well so we also disable dedup in this | |
1063 | * case. | |
9b67f605 | 1064 | */ |
b5256303 TC |
1065 | if (data == NULL && |
1066 | (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) { | |
9b67f605 MA |
1067 | zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; |
1068 | } | |
1069 | ||
34dc7c2f BB |
1070 | return (zio); |
1071 | } | |
1072 | ||
1073 | zio_t * | |
a6255b7f | 1074 | zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data, |
e8b96c60 | 1075 | uint64_t size, zio_done_func_t *done, void *private, |
5dbd68a3 | 1076 | zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) |
34dc7c2f BB |
1077 | { |
1078 | zio_t *zio; | |
1079 | ||
2aa34383 | 1080 | zio = zio_create(pio, spa, txg, bp, data, size, size, done, private, |
3dfb57a3 | 1081 | ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb, |
b128c09f | 1082 | ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); |
34dc7c2f BB |
1083 | |
1084 | return (zio); | |
1085 | } | |
1086 | ||
428870ff | 1087 | void |
03c6040b | 1088 | zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) |
428870ff BB |
1089 | { |
1090 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
1091 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1092 | ASSERT(zio->io_stage == ZIO_STAGE_OPEN); | |
1093 | ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); | |
1094 | ||
03c6040b GW |
1095 | /* |
1096 | * We must reset the io_prop to match the values that existed | |
1097 | * when the bp was first written by dmu_sync() keeping in mind | |
1098 | * that nopwrite and dedup are mutually exclusive. | |
1099 | */ | |
1100 | zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; | |
1101 | zio->io_prop.zp_nopwrite = nopwrite; | |
428870ff BB |
1102 | zio->io_prop.zp_copies = copies; |
1103 | zio->io_bp_override = bp; | |
1104 | } | |
1105 | ||
1106 | void | |
1107 | zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) | |
1108 | { | |
9b67f605 | 1109 | |
dc04a8c7 | 1110 | zfs_blkptr_verify(spa, bp, B_FALSE); |
a1d477c2 | 1111 | |
9b67f605 MA |
1112 | /* |
1113 | * The check for EMBEDDED is a performance optimization. We | |
1114 | * process the free here (by ignoring it) rather than | |
1115 | * putting it on the list and then processing it in zio_free_sync(). | |
1116 | */ | |
1117 | if (BP_IS_EMBEDDED(bp)) | |
1118 | return; | |
13fe0198 | 1119 | metaslab_check_free(spa, bp); |
2883cad5 MA |
1120 | |
1121 | /* | |
1122 | * Frees that are for the currently-syncing txg, are not going to be | |
1123 | * deferred, and which will not need to do a read (i.e. not GANG or | |
1124 | * DEDUP), can be processed immediately. Otherwise, put them on the | |
1125 | * in-memory list for later processing. | |
93e28d66 SD |
1126 | * |
1127 | * Note that we only defer frees after zfs_sync_pass_deferred_free | |
1128 | * when the log space map feature is disabled. [see relevant comment | |
1129 | * in spa_sync_iterate_to_convergence()] | |
2883cad5 | 1130 | */ |
93e28d66 SD |
1131 | if (BP_IS_GANG(bp) || |
1132 | BP_GET_DEDUP(bp) || | |
2883cad5 | 1133 | txg != spa->spa_syncing_txg || |
93e28d66 SD |
1134 | (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free && |
1135 | !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))) { | |
2883cad5 MA |
1136 | bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); |
1137 | } else { | |
1138 | VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0))); | |
1139 | } | |
428870ff BB |
1140 | } |
1141 | ||
34dc7c2f | 1142 | zio_t * |
428870ff BB |
1143 | zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
1144 | enum zio_flag flags) | |
34dc7c2f BB |
1145 | { |
1146 | zio_t *zio; | |
2883cad5 | 1147 | enum zio_stage stage = ZIO_FREE_PIPELINE; |
34dc7c2f | 1148 | |
428870ff BB |
1149 | ASSERT(!BP_IS_HOLE(bp)); |
1150 | ASSERT(spa_syncing_txg(spa) == txg); | |
34dc7c2f | 1151 | |
9b67f605 MA |
1152 | if (BP_IS_EMBEDDED(bp)) |
1153 | return (zio_null(pio, spa, NULL, NULL, NULL, 0)); | |
1154 | ||
13fe0198 | 1155 | metaslab_check_free(spa, bp); |
8c841793 | 1156 | arc_freed(spa, bp); |
d4a72f23 | 1157 | dsl_scan_freed(spa, bp); |
13fe0198 | 1158 | |
2883cad5 MA |
1159 | /* |
1160 | * GANG and DEDUP blocks can induce a read (for the gang block header, | |
1161 | * or the DDT), so issue them asynchronously so that this thread is | |
1162 | * not tied up. | |
1163 | */ | |
1164 | if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) | |
1165 | stage |= ZIO_STAGE_ISSUE_ASYNC; | |
1166 | ||
b128c09f | 1167 | zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
2aa34383 DK |
1168 | BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, |
1169 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage); | |
2883cad5 | 1170 | |
34dc7c2f BB |
1171 | return (zio); |
1172 | } | |
1173 | ||
1174 | zio_t * | |
428870ff BB |
1175 | zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
1176 | zio_done_func_t *done, void *private, enum zio_flag flags) | |
34dc7c2f BB |
1177 | { |
1178 | zio_t *zio; | |
1179 | ||
dc04a8c7 | 1180 | zfs_blkptr_verify(spa, bp, flags & ZIO_FLAG_CONFIG_WRITER); |
9b67f605 MA |
1181 | |
1182 | if (BP_IS_EMBEDDED(bp)) | |
1183 | return (zio_null(pio, spa, NULL, NULL, NULL, 0)); | |
1184 | ||
34dc7c2f BB |
1185 | /* |
1186 | * A claim is an allocation of a specific block. Claims are needed | |
1187 | * to support immediate writes in the intent log. The issue is that | |
1188 | * immediate writes contain committed data, but in a txg that was | |
1189 | * *not* committed. Upon opening the pool after an unclean shutdown, | |
1190 | * the intent log claims all blocks that contain immediate write data | |
1191 | * so that the SPA knows they're in use. | |
1192 | * | |
1193 | * All claims *must* be resolved in the first txg -- before the SPA | |
1194 | * starts allocating blocks -- so that nothing is allocated twice. | |
428870ff | 1195 | * If txg == 0 we just verify that the block is claimable. |
34dc7c2f | 1196 | */ |
d2734cce SD |
1197 | ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, |
1198 | spa_min_claim_txg(spa)); | |
1199 | ASSERT(txg == spa_min_claim_txg(spa) || txg == 0); | |
428870ff | 1200 | ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ |
34dc7c2f | 1201 | |
b128c09f | 1202 | zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
2aa34383 DK |
1203 | BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, |
1204 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); | |
3dfb57a3 | 1205 | ASSERT0(zio->io_queued_timestamp); |
34dc7c2f BB |
1206 | |
1207 | return (zio); | |
1208 | } | |
1209 | ||
1210 | zio_t * | |
1211 | zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, | |
e8b96c60 | 1212 | zio_done_func_t *done, void *private, enum zio_flag flags) |
34dc7c2f BB |
1213 | { |
1214 | zio_t *zio; | |
1215 | int c; | |
1216 | ||
1217 | if (vd->vdev_children == 0) { | |
2aa34383 | 1218 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
e8b96c60 | 1219 | ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
34dc7c2f BB |
1220 | ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); |
1221 | ||
34dc7c2f BB |
1222 | zio->io_cmd = cmd; |
1223 | } else { | |
d164b209 | 1224 | zio = zio_null(pio, spa, NULL, NULL, NULL, flags); |
34dc7c2f BB |
1225 | |
1226 | for (c = 0; c < vd->vdev_children; c++) | |
1227 | zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, | |
e8b96c60 | 1228 | done, private, flags)); |
34dc7c2f BB |
1229 | } |
1230 | ||
1231 | return (zio); | |
1232 | } | |
1233 | ||
1b939560 BB |
1234 | zio_t * |
1235 | zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
1236 | zio_done_func_t *done, void *private, zio_priority_t priority, | |
1237 | enum zio_flag flags, enum trim_flag trim_flags) | |
1238 | { | |
1239 | zio_t *zio; | |
1240 | ||
1241 | ASSERT0(vd->vdev_children); | |
1242 | ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift)); | |
1243 | ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift)); | |
1244 | ASSERT3U(size, !=, 0); | |
1245 | ||
1246 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done, | |
1247 | private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL, | |
1248 | vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE); | |
1249 | zio->io_trim_flags = trim_flags; | |
1250 | ||
1251 | return (zio); | |
1252 | } | |
1253 | ||
34dc7c2f BB |
1254 | zio_t * |
1255 | zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
a6255b7f | 1256 | abd_t *data, int checksum, zio_done_func_t *done, void *private, |
e8b96c60 | 1257 | zio_priority_t priority, enum zio_flag flags, boolean_t labels) |
34dc7c2f BB |
1258 | { |
1259 | zio_t *zio; | |
34dc7c2f | 1260 | |
b128c09f BB |
1261 | ASSERT(vd->vdev_children == 0); |
1262 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
1263 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
1264 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 1265 | |
2aa34383 DK |
1266 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
1267 | private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
1268 | offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); | |
34dc7c2f | 1269 | |
b128c09f | 1270 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f BB |
1271 | |
1272 | return (zio); | |
1273 | } | |
1274 | ||
1275 | zio_t * | |
1276 | zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
a6255b7f | 1277 | abd_t *data, int checksum, zio_done_func_t *done, void *private, |
e8b96c60 | 1278 | zio_priority_t priority, enum zio_flag flags, boolean_t labels) |
34dc7c2f | 1279 | { |
34dc7c2f | 1280 | zio_t *zio; |
34dc7c2f | 1281 | |
b128c09f BB |
1282 | ASSERT(vd->vdev_children == 0); |
1283 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
1284 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
1285 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 1286 | |
2aa34383 DK |
1287 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
1288 | private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
1289 | offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); | |
34dc7c2f | 1290 | |
b128c09f | 1291 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f | 1292 | |
3c67d83a | 1293 | if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { |
34dc7c2f | 1294 | /* |
428870ff | 1295 | * zec checksums are necessarily destructive -- they modify |
b128c09f | 1296 | * the end of the write buffer to hold the verifier/checksum. |
34dc7c2f | 1297 | * Therefore, we must make a local copy in case the data is |
b128c09f | 1298 | * being written to multiple places in parallel. |
34dc7c2f | 1299 | */ |
a6255b7f DQ |
1300 | abd_t *wbuf = abd_alloc_sametype(data, size); |
1301 | abd_copy(wbuf, data, size); | |
1302 | ||
b128c09f | 1303 | zio_push_transform(zio, wbuf, size, size, NULL); |
34dc7c2f BB |
1304 | } |
1305 | ||
1306 | return (zio); | |
1307 | } | |
1308 | ||
1309 | /* | |
b128c09f | 1310 | * Create a child I/O to do some work for us. |
34dc7c2f BB |
1311 | */ |
1312 | zio_t * | |
b128c09f | 1313 | zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, |
4ea3f864 GM |
1314 | abd_t *data, uint64_t size, int type, zio_priority_t priority, |
1315 | enum zio_flag flags, zio_done_func_t *done, void *private) | |
34dc7c2f | 1316 | { |
428870ff | 1317 | enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; |
b128c09f BB |
1318 | zio_t *zio; |
1319 | ||
a1d477c2 MA |
1320 | /* |
1321 | * vdev child I/Os do not propagate their error to the parent. | |
1322 | * Therefore, for correct operation the caller *must* check for | |
1323 | * and handle the error in the child i/o's done callback. | |
1324 | * The only exceptions are i/os that we don't care about | |
1325 | * (OPTIONAL or REPAIR). | |
1326 | */ | |
1327 | ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) || | |
1328 | done != NULL); | |
1329 | ||
34dc7c2f BB |
1330 | if (type == ZIO_TYPE_READ && bp != NULL) { |
1331 | /* | |
1332 | * If we have the bp, then the child should perform the | |
1333 | * checksum and the parent need not. This pushes error | |
1334 | * detection as close to the leaves as possible and | |
1335 | * eliminates redundant checksums in the interior nodes. | |
1336 | */ | |
428870ff BB |
1337 | pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; |
1338 | pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; | |
34dc7c2f BB |
1339 | } |
1340 | ||
a1d477c2 MA |
1341 | if (vd->vdev_ops->vdev_op_leaf) { |
1342 | ASSERT0(vd->vdev_children); | |
b128c09f | 1343 | offset += VDEV_LABEL_START_SIZE; |
a1d477c2 | 1344 | } |
b128c09f | 1345 | |
a1d477c2 | 1346 | flags |= ZIO_VDEV_CHILD_FLAGS(pio); |
428870ff BB |
1347 | |
1348 | /* | |
1349 | * If we've decided to do a repair, the write is not speculative -- | |
1350 | * even if the original read was. | |
1351 | */ | |
1352 | if (flags & ZIO_FLAG_IO_REPAIR) | |
1353 | flags &= ~ZIO_FLAG_SPECULATIVE; | |
1354 | ||
3dfb57a3 DB |
1355 | /* |
1356 | * If we're creating a child I/O that is not associated with a | |
1357 | * top-level vdev, then the child zio is not an allocating I/O. | |
1358 | * If this is a retried I/O then we ignore it since we will | |
1359 | * have already processed the original allocating I/O. | |
1360 | */ | |
1361 | if (flags & ZIO_FLAG_IO_ALLOCATING && | |
1362 | (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) { | |
cc99f275 DB |
1363 | ASSERT(pio->io_metaslab_class != NULL); |
1364 | ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled); | |
3dfb57a3 DB |
1365 | ASSERT(type == ZIO_TYPE_WRITE); |
1366 | ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE); | |
1367 | ASSERT(!(flags & ZIO_FLAG_IO_REPAIR)); | |
1368 | ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) || | |
1369 | pio->io_child_type == ZIO_CHILD_GANG); | |
1370 | ||
1371 | flags &= ~ZIO_FLAG_IO_ALLOCATING; | |
1372 | } | |
1373 | ||
1374 | ||
2aa34383 | 1375 | zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size, |
428870ff BB |
1376 | done, private, type, priority, flags, vd, offset, &pio->io_bookmark, |
1377 | ZIO_STAGE_VDEV_IO_START >> 1, pipeline); | |
3dfb57a3 | 1378 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); |
34dc7c2f | 1379 | |
e8b96c60 MA |
1380 | zio->io_physdone = pio->io_physdone; |
1381 | if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) | |
1382 | zio->io_logical->io_phys_children++; | |
1383 | ||
b128c09f | 1384 | return (zio); |
34dc7c2f BB |
1385 | } |
1386 | ||
b128c09f | 1387 | zio_t * |
a6255b7f | 1388 | zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size, |
9e052db4 | 1389 | zio_type_t type, zio_priority_t priority, enum zio_flag flags, |
e9aa730c | 1390 | zio_done_func_t *done, void *private) |
34dc7c2f | 1391 | { |
b128c09f | 1392 | zio_t *zio; |
34dc7c2f | 1393 | |
b128c09f | 1394 | ASSERT(vd->vdev_ops->vdev_op_leaf); |
34dc7c2f | 1395 | |
b128c09f | 1396 | zio = zio_create(NULL, vd->vdev_spa, 0, NULL, |
2aa34383 | 1397 | data, size, size, done, private, type, priority, |
e8b96c60 | 1398 | flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, |
b128c09f | 1399 | vd, offset, NULL, |
428870ff | 1400 | ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); |
34dc7c2f | 1401 | |
b128c09f | 1402 | return (zio); |
34dc7c2f BB |
1403 | } |
1404 | ||
1405 | void | |
b128c09f | 1406 | zio_flush(zio_t *zio, vdev_t *vd) |
34dc7c2f | 1407 | { |
b128c09f | 1408 | zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, |
e8b96c60 | 1409 | NULL, NULL, |
b128c09f | 1410 | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); |
34dc7c2f BB |
1411 | } |
1412 | ||
428870ff BB |
1413 | void |
1414 | zio_shrink(zio_t *zio, uint64_t size) | |
1415 | { | |
1ce23dca PS |
1416 | ASSERT3P(zio->io_executor, ==, NULL); |
1417 | ASSERT3U(zio->io_orig_size, ==, zio->io_size); | |
1418 | ASSERT3U(size, <=, zio->io_size); | |
428870ff BB |
1419 | |
1420 | /* | |
1421 | * We don't shrink for raidz because of problems with the | |
1422 | * reconstruction when reading back less than the block size. | |
1423 | * Note, BP_IS_RAIDZ() assumes no compression. | |
1424 | */ | |
1425 | ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); | |
2aa34383 DK |
1426 | if (!BP_IS_RAIDZ(zio->io_bp)) { |
1427 | /* we are not doing a raw write */ | |
1428 | ASSERT3U(zio->io_size, ==, zio->io_lsize); | |
1429 | zio->io_orig_size = zio->io_size = zio->io_lsize = size; | |
1430 | } | |
428870ff BB |
1431 | } |
1432 | ||
34dc7c2f BB |
1433 | /* |
1434 | * ========================================================================== | |
b128c09f | 1435 | * Prepare to read and write logical blocks |
34dc7c2f BB |
1436 | * ========================================================================== |
1437 | */ | |
b128c09f | 1438 | |
62840030 | 1439 | static zio_t * |
b128c09f | 1440 | zio_read_bp_init(zio_t *zio) |
34dc7c2f | 1441 | { |
b128c09f | 1442 | blkptr_t *bp = zio->io_bp; |
b5256303 TC |
1443 | uint64_t psize = |
1444 | BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); | |
34dc7c2f | 1445 | |
a1d477c2 MA |
1446 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
1447 | ||
fb5f0bc8 | 1448 | if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && |
9babb374 | 1449 | zio->io_child_type == ZIO_CHILD_LOGICAL && |
b5256303 | 1450 | !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { |
a6255b7f DQ |
1451 | zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), |
1452 | psize, psize, zio_decompress); | |
34dc7c2f | 1453 | } |
34dc7c2f | 1454 | |
b5256303 TC |
1455 | if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) || |
1456 | BP_HAS_INDIRECT_MAC_CKSUM(bp)) && | |
1457 | zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
1458 | zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), | |
1459 | psize, psize, zio_decrypt); | |
1460 | } | |
1461 | ||
9b67f605 | 1462 | if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { |
a6255b7f DQ |
1463 | int psize = BPE_GET_PSIZE(bp); |
1464 | void *data = abd_borrow_buf(zio->io_abd, psize); | |
1465 | ||
9b67f605 | 1466 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
a6255b7f DQ |
1467 | decode_embedded_bp_compressed(bp, data); |
1468 | abd_return_buf_copy(zio->io_abd, data, psize); | |
9b67f605 MA |
1469 | } else { |
1470 | ASSERT(!BP_IS_EMBEDDED(bp)); | |
a1d477c2 | 1471 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
9b67f605 MA |
1472 | } |
1473 | ||
9ae529ec | 1474 | if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) |
b128c09f BB |
1475 | zio->io_flags |= ZIO_FLAG_DONT_CACHE; |
1476 | ||
428870ff BB |
1477 | if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) |
1478 | zio->io_flags |= ZIO_FLAG_DONT_CACHE; | |
1479 | ||
1480 | if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) | |
1481 | zio->io_pipeline = ZIO_DDT_READ_PIPELINE; | |
1482 | ||
62840030 | 1483 | return (zio); |
34dc7c2f BB |
1484 | } |
1485 | ||
62840030 | 1486 | static zio_t * |
b128c09f | 1487 | zio_write_bp_init(zio_t *zio) |
34dc7c2f | 1488 | { |
b128c09f | 1489 | if (!IO_IS_ALLOCATING(zio)) |
62840030 | 1490 | return (zio); |
34dc7c2f | 1491 | |
428870ff BB |
1492 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1493 | ||
1494 | if (zio->io_bp_override) { | |
3dfb57a3 DB |
1495 | blkptr_t *bp = zio->io_bp; |
1496 | zio_prop_t *zp = &zio->io_prop; | |
1497 | ||
428870ff BB |
1498 | ASSERT(bp->blk_birth != zio->io_txg); |
1499 | ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); | |
1500 | ||
1501 | *bp = *zio->io_bp_override; | |
1502 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1503 | ||
9b67f605 | 1504 | if (BP_IS_EMBEDDED(bp)) |
62840030 | 1505 | return (zio); |
9b67f605 | 1506 | |
03c6040b GW |
1507 | /* |
1508 | * If we've been overridden and nopwrite is set then | |
1509 | * set the flag accordingly to indicate that a nopwrite | |
1510 | * has already occurred. | |
1511 | */ | |
1512 | if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { | |
1513 | ASSERT(!zp->zp_dedup); | |
3dfb57a3 | 1514 | ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum); |
03c6040b | 1515 | zio->io_flags |= ZIO_FLAG_NOPWRITE; |
62840030 | 1516 | return (zio); |
03c6040b GW |
1517 | } |
1518 | ||
1519 | ASSERT(!zp->zp_nopwrite); | |
1520 | ||
428870ff | 1521 | if (BP_IS_HOLE(bp) || !zp->zp_dedup) |
62840030 | 1522 | return (zio); |
428870ff | 1523 | |
3c67d83a TH |
1524 | ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & |
1525 | ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); | |
428870ff | 1526 | |
b5256303 TC |
1527 | if (BP_GET_CHECKSUM(bp) == zp->zp_checksum && |
1528 | !zp->zp_encrypt) { | |
428870ff BB |
1529 | BP_SET_DEDUP(bp, 1); |
1530 | zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; | |
62840030 | 1531 | return (zio); |
428870ff | 1532 | } |
3dfb57a3 DB |
1533 | |
1534 | /* | |
1535 | * We were unable to handle this as an override bp, treat | |
1536 | * it as a regular write I/O. | |
1537 | */ | |
5511754b | 1538 | zio->io_bp_override = NULL; |
3dfb57a3 DB |
1539 | *bp = zio->io_bp_orig; |
1540 | zio->io_pipeline = zio->io_orig_pipeline; | |
1541 | } | |
1542 | ||
62840030 | 1543 | return (zio); |
3dfb57a3 DB |
1544 | } |
1545 | ||
62840030 | 1546 | static zio_t * |
3dfb57a3 DB |
1547 | zio_write_compress(zio_t *zio) |
1548 | { | |
1549 | spa_t *spa = zio->io_spa; | |
1550 | zio_prop_t *zp = &zio->io_prop; | |
1551 | enum zio_compress compress = zp->zp_compress; | |
1552 | blkptr_t *bp = zio->io_bp; | |
1553 | uint64_t lsize = zio->io_lsize; | |
1554 | uint64_t psize = zio->io_size; | |
1555 | int pass = 1; | |
1556 | ||
3dfb57a3 DB |
1557 | /* |
1558 | * If our children haven't all reached the ready stage, | |
1559 | * wait for them and then repeat this pipeline stage. | |
1560 | */ | |
ddc751d5 GW |
1561 | if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT | |
1562 | ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) { | |
62840030 | 1563 | return (NULL); |
ddc751d5 | 1564 | } |
3dfb57a3 DB |
1565 | |
1566 | if (!IO_IS_ALLOCATING(zio)) | |
62840030 | 1567 | return (zio); |
3dfb57a3 DB |
1568 | |
1569 | if (zio->io_children_ready != NULL) { | |
1570 | /* | |
1571 | * Now that all our children are ready, run the callback | |
1572 | * associated with this zio in case it wants to modify the | |
1573 | * data to be written. | |
1574 | */ | |
1575 | ASSERT3U(zp->zp_level, >, 0); | |
1576 | zio->io_children_ready(zio); | |
428870ff | 1577 | } |
34dc7c2f | 1578 | |
3dfb57a3 DB |
1579 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1580 | ASSERT(zio->io_bp_override == NULL); | |
1581 | ||
b0bc7a84 | 1582 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { |
b128c09f BB |
1583 | /* |
1584 | * We're rewriting an existing block, which means we're | |
1585 | * working on behalf of spa_sync(). For spa_sync() to | |
1586 | * converge, it must eventually be the case that we don't | |
1587 | * have to allocate new blocks. But compression changes | |
1588 | * the blocksize, which forces a reallocate, and makes | |
1589 | * convergence take longer. Therefore, after the first | |
1590 | * few passes, stop compressing to ensure convergence. | |
1591 | */ | |
428870ff BB |
1592 | pass = spa_sync_pass(spa); |
1593 | ||
1594 | ASSERT(zio->io_txg == spa_syncing_txg(spa)); | |
1595 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1596 | ASSERT(!BP_GET_DEDUP(bp)); | |
34dc7c2f | 1597 | |
55d85d5a | 1598 | if (pass >= zfs_sync_pass_dont_compress) |
b128c09f | 1599 | compress = ZIO_COMPRESS_OFF; |
34dc7c2f | 1600 | |
b128c09f | 1601 | /* Make sure someone doesn't change their mind on overwrites */ |
9b67f605 | 1602 | ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), |
428870ff | 1603 | spa_max_replication(spa)) == BP_GET_NDVAS(bp)); |
b128c09f | 1604 | } |
34dc7c2f | 1605 | |
2aa34383 | 1606 | /* If it's a compressed write that is not raw, compress the buffer. */ |
b5256303 TC |
1607 | if (compress != ZIO_COMPRESS_OFF && |
1608 | !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { | |
428870ff | 1609 | void *cbuf = zio_buf_alloc(lsize); |
a6255b7f | 1610 | psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize); |
428870ff | 1611 | if (psize == 0 || psize == lsize) { |
b128c09f | 1612 | compress = ZIO_COMPRESS_OFF; |
428870ff | 1613 | zio_buf_free(cbuf, lsize); |
b5256303 TC |
1614 | } else if (!zp->zp_dedup && !zp->zp_encrypt && |
1615 | psize <= BPE_PAYLOAD_SIZE && | |
9b67f605 MA |
1616 | zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && |
1617 | spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { | |
1618 | encode_embedded_bp_compressed(bp, | |
1619 | cbuf, compress, lsize, psize); | |
1620 | BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); | |
1621 | BP_SET_TYPE(bp, zio->io_prop.zp_type); | |
1622 | BP_SET_LEVEL(bp, zio->io_prop.zp_level); | |
1623 | zio_buf_free(cbuf, lsize); | |
1624 | bp->blk_birth = zio->io_txg; | |
1625 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1626 | ASSERT(spa_feature_is_active(spa, | |
1627 | SPA_FEATURE_EMBEDDED_DATA)); | |
62840030 | 1628 | return (zio); |
428870ff | 1629 | } else { |
9b67f605 | 1630 | /* |
c3520e7f MA |
1631 | * Round up compressed size up to the ashift |
1632 | * of the smallest-ashift device, and zero the tail. | |
1633 | * This ensures that the compressed size of the BP | |
1634 | * (and thus compressratio property) are correct, | |
1635 | * in that we charge for the padding used to fill out | |
1636 | * the last sector. | |
9b67f605 | 1637 | */ |
c3520e7f | 1638 | ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT); |
1c27024e | 1639 | size_t rounded = (size_t)P2ROUNDUP(psize, |
c3520e7f MA |
1640 | 1ULL << spa->spa_min_ashift); |
1641 | if (rounded >= lsize) { | |
9b67f605 MA |
1642 | compress = ZIO_COMPRESS_OFF; |
1643 | zio_buf_free(cbuf, lsize); | |
c3520e7f | 1644 | psize = lsize; |
9b67f605 | 1645 | } else { |
a6255b7f DQ |
1646 | abd_t *cdata = abd_get_from_buf(cbuf, lsize); |
1647 | abd_take_ownership_of_buf(cdata, B_TRUE); | |
1648 | abd_zero_off(cdata, psize, rounded - psize); | |
c3520e7f | 1649 | psize = rounded; |
a6255b7f | 1650 | zio_push_transform(zio, cdata, |
9b67f605 MA |
1651 | psize, lsize, NULL); |
1652 | } | |
b128c09f | 1653 | } |
3dfb57a3 DB |
1654 | |
1655 | /* | |
1656 | * We were unable to handle this as an override bp, treat | |
1657 | * it as a regular write I/O. | |
1658 | */ | |
1659 | zio->io_bp_override = NULL; | |
1660 | *bp = zio->io_bp_orig; | |
1661 | zio->io_pipeline = zio->io_orig_pipeline; | |
1662 | ||
b1d21733 TC |
1663 | } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 && |
1664 | zp->zp_type == DMU_OT_DNODE) { | |
1665 | /* | |
1666 | * The DMU actually relies on the zio layer's compression | |
1667 | * to free metadnode blocks that have had all contained | |
1668 | * dnodes freed. As a result, even when doing a raw | |
1669 | * receive, we must check whether the block can be compressed | |
1670 | * to a hole. | |
1671 | */ | |
1672 | psize = zio_compress_data(ZIO_COMPRESS_EMPTY, | |
1673 | zio->io_abd, NULL, lsize); | |
1674 | if (psize == 0) | |
1675 | compress = ZIO_COMPRESS_OFF; | |
2aa34383 DK |
1676 | } else { |
1677 | ASSERT3U(psize, !=, 0); | |
b128c09f | 1678 | } |
34dc7c2f | 1679 | |
b128c09f BB |
1680 | /* |
1681 | * The final pass of spa_sync() must be all rewrites, but the first | |
1682 | * few passes offer a trade-off: allocating blocks defers convergence, | |
1683 | * but newly allocated blocks are sequential, so they can be written | |
1684 | * to disk faster. Therefore, we allow the first few passes of | |
1685 | * spa_sync() to allocate new blocks, but force rewrites after that. | |
1686 | * There should only be a handful of blocks after pass 1 in any case. | |
1687 | */ | |
b0bc7a84 MG |
1688 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && |
1689 | BP_GET_PSIZE(bp) == psize && | |
55d85d5a | 1690 | pass >= zfs_sync_pass_rewrite) { |
cc99f275 | 1691 | VERIFY3U(psize, !=, 0); |
1c27024e | 1692 | enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; |
cc99f275 | 1693 | |
b128c09f BB |
1694 | zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; |
1695 | zio->io_flags |= ZIO_FLAG_IO_REWRITE; | |
1696 | } else { | |
1697 | BP_ZERO(bp); | |
1698 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
1699 | } | |
34dc7c2f | 1700 | |
428870ff | 1701 | if (psize == 0) { |
b0bc7a84 MG |
1702 | if (zio->io_bp_orig.blk_birth != 0 && |
1703 | spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { | |
1704 | BP_SET_LSIZE(bp, lsize); | |
1705 | BP_SET_TYPE(bp, zp->zp_type); | |
1706 | BP_SET_LEVEL(bp, zp->zp_level); | |
1707 | BP_SET_BIRTH(bp, zio->io_txg, 0); | |
1708 | } | |
b128c09f BB |
1709 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
1710 | } else { | |
1711 | ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); | |
1712 | BP_SET_LSIZE(bp, lsize); | |
b0bc7a84 MG |
1713 | BP_SET_TYPE(bp, zp->zp_type); |
1714 | BP_SET_LEVEL(bp, zp->zp_level); | |
428870ff | 1715 | BP_SET_PSIZE(bp, psize); |
b128c09f BB |
1716 | BP_SET_COMPRESS(bp, compress); |
1717 | BP_SET_CHECKSUM(bp, zp->zp_checksum); | |
428870ff | 1718 | BP_SET_DEDUP(bp, zp->zp_dedup); |
b128c09f | 1719 | BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); |
428870ff BB |
1720 | if (zp->zp_dedup) { |
1721 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1722 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
b5256303 TC |
1723 | ASSERT(!zp->zp_encrypt || |
1724 | DMU_OT_IS_ENCRYPTED(zp->zp_type)); | |
428870ff BB |
1725 | zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; |
1726 | } | |
03c6040b GW |
1727 | if (zp->zp_nopwrite) { |
1728 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1729 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
1730 | zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; | |
1731 | } | |
428870ff | 1732 | } |
62840030 | 1733 | return (zio); |
428870ff BB |
1734 | } |
1735 | ||
62840030 | 1736 | static zio_t * |
428870ff BB |
1737 | zio_free_bp_init(zio_t *zio) |
1738 | { | |
1739 | blkptr_t *bp = zio->io_bp; | |
1740 | ||
1741 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
1742 | if (BP_GET_DEDUP(bp)) | |
1743 | zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; | |
b128c09f | 1744 | } |
34dc7c2f | 1745 | |
a1d477c2 MA |
1746 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
1747 | ||
62840030 | 1748 | return (zio); |
34dc7c2f BB |
1749 | } |
1750 | ||
b128c09f BB |
1751 | /* |
1752 | * ========================================================================== | |
1753 | * Execute the I/O pipeline | |
1754 | * ========================================================================== | |
1755 | */ | |
1756 | ||
1757 | static void | |
7ef5e54e | 1758 | zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) |
34dc7c2f | 1759 | { |
428870ff | 1760 | spa_t *spa = zio->io_spa; |
b128c09f | 1761 | zio_type_t t = zio->io_type; |
a38718a6 | 1762 | int flags = (cutinline ? TQ_FRONT : 0); |
34dc7c2f BB |
1763 | |
1764 | /* | |
9babb374 BB |
1765 | * If we're a config writer or a probe, the normal issue and |
1766 | * interrupt threads may all be blocked waiting for the config lock. | |
1767 | * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. | |
34dc7c2f | 1768 | */ |
9babb374 | 1769 | if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) |
b128c09f | 1770 | t = ZIO_TYPE_NULL; |
34dc7c2f BB |
1771 | |
1772 | /* | |
b128c09f | 1773 | * A similar issue exists for the L2ARC write thread until L2ARC 2.0. |
34dc7c2f | 1774 | */ |
b128c09f BB |
1775 | if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) |
1776 | t = ZIO_TYPE_NULL; | |
34dc7c2f | 1777 | |
428870ff | 1778 | /* |
7ef5e54e AL |
1779 | * If this is a high priority I/O, then use the high priority taskq if |
1780 | * available. | |
428870ff | 1781 | */ |
18b14b17 GW |
1782 | if ((zio->io_priority == ZIO_PRIORITY_NOW || |
1783 | zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) && | |
7ef5e54e | 1784 | spa->spa_zio_taskq[t][q + 1].stqs_count != 0) |
428870ff BB |
1785 | q++; |
1786 | ||
1787 | ASSERT3U(q, <, ZIO_TASKQ_TYPES); | |
5cc556b4 | 1788 | |
a38718a6 GA |
1789 | /* |
1790 | * NB: We are assuming that the zio can only be dispatched | |
1791 | * to a single taskq at a time. It would be a grievous error | |
1792 | * to dispatch the zio to another taskq at the same time. | |
1793 | */ | |
1794 | ASSERT(taskq_empty_ent(&zio->io_tqent)); | |
7ef5e54e AL |
1795 | spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, |
1796 | flags, &zio->io_tqent); | |
b128c09f | 1797 | } |
34dc7c2f | 1798 | |
b128c09f | 1799 | static boolean_t |
7ef5e54e | 1800 | zio_taskq_member(zio_t *zio, zio_taskq_type_t q) |
b128c09f BB |
1801 | { |
1802 | kthread_t *executor = zio->io_executor; | |
1803 | spa_t *spa = zio->io_spa; | |
34dc7c2f | 1804 | |
1c27024e | 1805 | for (zio_type_t t = 0; t < ZIO_TYPES; t++) { |
7ef5e54e AL |
1806 | spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; |
1807 | uint_t i; | |
1808 | for (i = 0; i < tqs->stqs_count; i++) { | |
1809 | if (taskq_member(tqs->stqs_taskq[i], executor)) | |
1810 | return (B_TRUE); | |
1811 | } | |
1812 | } | |
34dc7c2f | 1813 | |
b128c09f BB |
1814 | return (B_FALSE); |
1815 | } | |
34dc7c2f | 1816 | |
62840030 | 1817 | static zio_t * |
b128c09f BB |
1818 | zio_issue_async(zio_t *zio) |
1819 | { | |
428870ff | 1820 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); |
b128c09f | 1821 | |
62840030 | 1822 | return (NULL); |
34dc7c2f BB |
1823 | } |
1824 | ||
b128c09f BB |
1825 | void |
1826 | zio_interrupt(zio_t *zio) | |
34dc7c2f | 1827 | { |
428870ff | 1828 | zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); |
b128c09f | 1829 | } |
34dc7c2f | 1830 | |
8fb1ede1 | 1831 | static void |
638dd5f4 | 1832 | zio_deadman_impl(zio_t *pio, int ziodepth) |
8fb1ede1 BB |
1833 | { |
1834 | zio_t *cio, *cio_next; | |
1835 | zio_link_t *zl = NULL; | |
1836 | vdev_t *vd = pio->io_vd; | |
1837 | ||
638dd5f4 TC |
1838 | if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) { |
1839 | vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL; | |
8fb1ede1 BB |
1840 | zbookmark_phys_t *zb = &pio->io_bookmark; |
1841 | uint64_t delta = gethrtime() - pio->io_timestamp; | |
1842 | uint64_t failmode = spa_get_deadman_failmode(pio->io_spa); | |
1843 | ||
a887d653 | 1844 | zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu " |
8fb1ede1 BB |
1845 | "delta=%llu queued=%llu io=%llu " |
1846 | "path=%s last=%llu " | |
1847 | "type=%d priority=%d flags=0x%x " | |
1848 | "stage=0x%x pipeline=0x%x pipeline-trace=0x%x " | |
1849 | "objset=%llu object=%llu level=%llu blkid=%llu " | |
1850 | "offset=%llu size=%llu error=%d", | |
638dd5f4 | 1851 | ziodepth, pio, pio->io_timestamp, |
8fb1ede1 | 1852 | delta, pio->io_delta, pio->io_delay, |
638dd5f4 | 1853 | vd ? vd->vdev_path : "NULL", vq ? vq->vq_io_complete_ts : 0, |
8fb1ede1 | 1854 | pio->io_type, pio->io_priority, pio->io_flags, |
638dd5f4 | 1855 | pio->io_stage, pio->io_pipeline, pio->io_pipeline_trace, |
8fb1ede1 BB |
1856 | zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid, |
1857 | pio->io_offset, pio->io_size, pio->io_error); | |
1858 | zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN, | |
1859 | pio->io_spa, vd, zb, pio, 0, 0); | |
1860 | ||
1861 | if (failmode == ZIO_FAILURE_MODE_CONTINUE && | |
1862 | taskq_empty_ent(&pio->io_tqent)) { | |
1863 | zio_interrupt(pio); | |
1864 | } | |
1865 | } | |
1866 | ||
1867 | mutex_enter(&pio->io_lock); | |
1868 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { | |
1869 | cio_next = zio_walk_children(pio, &zl); | |
638dd5f4 | 1870 | zio_deadman_impl(cio, ziodepth + 1); |
8fb1ede1 BB |
1871 | } |
1872 | mutex_exit(&pio->io_lock); | |
1873 | } | |
1874 | ||
1875 | /* | |
1876 | * Log the critical information describing this zio and all of its children | |
1877 | * using the zfs_dbgmsg() interface then post deadman event for the ZED. | |
1878 | */ | |
1879 | void | |
1880 | zio_deadman(zio_t *pio, char *tag) | |
1881 | { | |
1882 | spa_t *spa = pio->io_spa; | |
1883 | char *name = spa_name(spa); | |
1884 | ||
1885 | if (!zfs_deadman_enabled || spa_suspended(spa)) | |
1886 | return; | |
1887 | ||
638dd5f4 | 1888 | zio_deadman_impl(pio, 0); |
8fb1ede1 BB |
1889 | |
1890 | switch (spa_get_deadman_failmode(spa)) { | |
1891 | case ZIO_FAILURE_MODE_WAIT: | |
1892 | zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name); | |
1893 | break; | |
1894 | ||
1895 | case ZIO_FAILURE_MODE_CONTINUE: | |
1896 | zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name); | |
1897 | break; | |
1898 | ||
1899 | case ZIO_FAILURE_MODE_PANIC: | |
1900 | fm_panic("%s determined I/O to pool '%s' is hung.", tag, name); | |
1901 | break; | |
1902 | } | |
1903 | } | |
1904 | ||
b128c09f BB |
1905 | /* |
1906 | * Execute the I/O pipeline until one of the following occurs: | |
1907 | * (1) the I/O completes; (2) the pipeline stalls waiting for | |
1908 | * dependent child I/Os; (3) the I/O issues, so we're waiting | |
1909 | * for an I/O completion interrupt; (4) the I/O is delegated by | |
1910 | * vdev-level caching or aggregation; (5) the I/O is deferred | |
1911 | * due to vdev-level queueing; (6) the I/O is handed off to | |
1912 | * another thread. In all cases, the pipeline stops whenever | |
8e07b99b | 1913 | * there's no CPU work; it never burns a thread in cv_wait_io(). |
b128c09f BB |
1914 | * |
1915 | * There's no locking on io_stage because there's no legitimate way | |
1916 | * for multiple threads to be attempting to process the same I/O. | |
1917 | */ | |
428870ff | 1918 | static zio_pipe_stage_t *zio_pipeline[]; |
34dc7c2f | 1919 | |
da6b4005 NB |
1920 | /* |
1921 | * zio_execute() is a wrapper around the static function | |
1922 | * __zio_execute() so that we can force __zio_execute() to be | |
1923 | * inlined. This reduces stack overhead which is important | |
1924 | * because __zio_execute() is called recursively in several zio | |
1925 | * code paths. zio_execute() itself cannot be inlined because | |
1926 | * it is externally visible. | |
1927 | */ | |
b128c09f BB |
1928 | void |
1929 | zio_execute(zio_t *zio) | |
da6b4005 | 1930 | { |
92119cc2 BB |
1931 | fstrans_cookie_t cookie; |
1932 | ||
1933 | cookie = spl_fstrans_mark(); | |
da6b4005 | 1934 | __zio_execute(zio); |
92119cc2 | 1935 | spl_fstrans_unmark(cookie); |
da6b4005 NB |
1936 | } |
1937 | ||
b58986ee BB |
1938 | /* |
1939 | * Used to determine if in the current context the stack is sized large | |
1940 | * enough to allow zio_execute() to be called recursively. A minimum | |
1941 | * stack size of 16K is required to avoid needing to re-dispatch the zio. | |
1942 | */ | |
1943 | boolean_t | |
1944 | zio_execute_stack_check(zio_t *zio) | |
1945 | { | |
1946 | #if !defined(HAVE_LARGE_STACKS) | |
1947 | dsl_pool_t *dp = spa_get_dsl(zio->io_spa); | |
1948 | ||
1949 | /* Executing in txg_sync_thread() context. */ | |
1950 | if (dp && curthread == dp->dp_tx.tx_sync_thread) | |
1951 | return (B_TRUE); | |
1952 | ||
1953 | /* Pool initialization outside of zio_taskq context. */ | |
1954 | if (dp && spa_is_initializing(dp->dp_spa) && | |
1955 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) && | |
1956 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH)) | |
1957 | return (B_TRUE); | |
1958 | #endif /* HAVE_LARGE_STACKS */ | |
1959 | ||
1960 | return (B_FALSE); | |
1961 | } | |
1962 | ||
da6b4005 NB |
1963 | __attribute__((always_inline)) |
1964 | static inline void | |
1965 | __zio_execute(zio_t *zio) | |
b128c09f | 1966 | { |
3dfb57a3 DB |
1967 | ASSERT3U(zio->io_queued_timestamp, >, 0); |
1968 | ||
b128c09f | 1969 | while (zio->io_stage < ZIO_STAGE_DONE) { |
428870ff BB |
1970 | enum zio_stage pipeline = zio->io_pipeline; |
1971 | enum zio_stage stage = zio->io_stage; | |
62840030 MA |
1972 | |
1973 | zio->io_executor = curthread; | |
34dc7c2f | 1974 | |
b128c09f | 1975 | ASSERT(!MUTEX_HELD(&zio->io_lock)); |
428870ff BB |
1976 | ASSERT(ISP2(stage)); |
1977 | ASSERT(zio->io_stall == NULL); | |
34dc7c2f | 1978 | |
428870ff BB |
1979 | do { |
1980 | stage <<= 1; | |
1981 | } while ((stage & pipeline) == 0); | |
b128c09f BB |
1982 | |
1983 | ASSERT(stage <= ZIO_STAGE_DONE); | |
34dc7c2f BB |
1984 | |
1985 | /* | |
b128c09f BB |
1986 | * If we are in interrupt context and this pipeline stage |
1987 | * will grab a config lock that is held across I/O, | |
428870ff BB |
1988 | * or may wait for an I/O that needs an interrupt thread |
1989 | * to complete, issue async to avoid deadlock. | |
1990 | * | |
1991 | * For VDEV_IO_START, we cut in line so that the io will | |
1992 | * be sent to disk promptly. | |
34dc7c2f | 1993 | */ |
91579709 BB |
1994 | if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && |
1995 | zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { | |
b58986ee BB |
1996 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? |
1997 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
91579709 BB |
1998 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
1999 | return; | |
2000 | } | |
2001 | ||
2002 | /* | |
b58986ee BB |
2003 | * If the current context doesn't have large enough stacks |
2004 | * the zio must be issued asynchronously to prevent overflow. | |
91579709 | 2005 | */ |
b58986ee BB |
2006 | if (zio_execute_stack_check(zio)) { |
2007 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? | |
2008 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
428870ff | 2009 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
b128c09f | 2010 | return; |
34dc7c2f BB |
2011 | } |
2012 | ||
b128c09f | 2013 | zio->io_stage = stage; |
3dfb57a3 | 2014 | zio->io_pipeline_trace |= zio->io_stage; |
34dc7c2f | 2015 | |
62840030 MA |
2016 | /* |
2017 | * The zio pipeline stage returns the next zio to execute | |
2018 | * (typically the same as this one), or NULL if we should | |
2019 | * stop. | |
2020 | */ | |
2021 | zio = zio_pipeline[highbit64(stage) - 1](zio); | |
34dc7c2f | 2022 | |
62840030 MA |
2023 | if (zio == NULL) |
2024 | return; | |
b128c09f | 2025 | } |
34dc7c2f BB |
2026 | } |
2027 | ||
da6b4005 | 2028 | |
b128c09f BB |
2029 | /* |
2030 | * ========================================================================== | |
2031 | * Initiate I/O, either sync or async | |
2032 | * ========================================================================== | |
2033 | */ | |
2034 | int | |
2035 | zio_wait(zio_t *zio) | |
34dc7c2f | 2036 | { |
8fb1ede1 | 2037 | long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms); |
b128c09f | 2038 | int error; |
34dc7c2f | 2039 | |
1ce23dca PS |
2040 | ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN); |
2041 | ASSERT3P(zio->io_executor, ==, NULL); | |
34dc7c2f | 2042 | |
b128c09f | 2043 | zio->io_waiter = curthread; |
3dfb57a3 DB |
2044 | ASSERT0(zio->io_queued_timestamp); |
2045 | zio->io_queued_timestamp = gethrtime(); | |
34dc7c2f | 2046 | |
da6b4005 | 2047 | __zio_execute(zio); |
34dc7c2f | 2048 | |
b128c09f | 2049 | mutex_enter(&zio->io_lock); |
8fb1ede1 BB |
2050 | while (zio->io_executor != NULL) { |
2051 | error = cv_timedwait_io(&zio->io_cv, &zio->io_lock, | |
2052 | ddi_get_lbolt() + timeout); | |
2053 | ||
2054 | if (zfs_deadman_enabled && error == -1 && | |
2055 | gethrtime() - zio->io_queued_timestamp > | |
2056 | spa_deadman_ziotime(zio->io_spa)) { | |
2057 | mutex_exit(&zio->io_lock); | |
2058 | timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms); | |
2059 | zio_deadman(zio, FTAG); | |
2060 | mutex_enter(&zio->io_lock); | |
2061 | } | |
2062 | } | |
b128c09f | 2063 | mutex_exit(&zio->io_lock); |
34dc7c2f | 2064 | |
b128c09f BB |
2065 | error = zio->io_error; |
2066 | zio_destroy(zio); | |
34dc7c2f | 2067 | |
b128c09f BB |
2068 | return (error); |
2069 | } | |
34dc7c2f | 2070 | |
b128c09f BB |
2071 | void |
2072 | zio_nowait(zio_t *zio) | |
2073 | { | |
1ce23dca | 2074 | ASSERT3P(zio->io_executor, ==, NULL); |
34dc7c2f | 2075 | |
d164b209 BB |
2076 | if (zio->io_child_type == ZIO_CHILD_LOGICAL && |
2077 | zio_unique_parent(zio) == NULL) { | |
8878261f BB |
2078 | zio_t *pio; |
2079 | ||
34dc7c2f | 2080 | /* |
b128c09f | 2081 | * This is a logical async I/O with no parent to wait for it. |
9babb374 BB |
2082 | * We add it to the spa_async_root_zio "Godfather" I/O which |
2083 | * will ensure they complete prior to unloading the pool. | |
34dc7c2f | 2084 | */ |
b128c09f | 2085 | spa_t *spa = zio->io_spa; |
8878261f BB |
2086 | kpreempt_disable(); |
2087 | pio = spa->spa_async_zio_root[CPU_SEQID]; | |
2088 | kpreempt_enable(); | |
9babb374 | 2089 | |
8878261f | 2090 | zio_add_child(pio, zio); |
b128c09f | 2091 | } |
34dc7c2f | 2092 | |
3dfb57a3 DB |
2093 | ASSERT0(zio->io_queued_timestamp); |
2094 | zio->io_queued_timestamp = gethrtime(); | |
da6b4005 | 2095 | __zio_execute(zio); |
b128c09f | 2096 | } |
34dc7c2f | 2097 | |
b128c09f BB |
2098 | /* |
2099 | * ========================================================================== | |
1ce23dca | 2100 | * Reexecute, cancel, or suspend/resume failed I/O |
b128c09f BB |
2101 | * ========================================================================== |
2102 | */ | |
34dc7c2f | 2103 | |
b128c09f BB |
2104 | static void |
2105 | zio_reexecute(zio_t *pio) | |
2106 | { | |
d164b209 BB |
2107 | zio_t *cio, *cio_next; |
2108 | ||
2109 | ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); | |
2110 | ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); | |
9babb374 BB |
2111 | ASSERT(pio->io_gang_leader == NULL); |
2112 | ASSERT(pio->io_gang_tree == NULL); | |
34dc7c2f | 2113 | |
b128c09f BB |
2114 | pio->io_flags = pio->io_orig_flags; |
2115 | pio->io_stage = pio->io_orig_stage; | |
2116 | pio->io_pipeline = pio->io_orig_pipeline; | |
2117 | pio->io_reexecute = 0; | |
03c6040b | 2118 | pio->io_flags |= ZIO_FLAG_REEXECUTED; |
3dfb57a3 | 2119 | pio->io_pipeline_trace = 0; |
b128c09f | 2120 | pio->io_error = 0; |
1c27024e | 2121 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 | 2122 | pio->io_state[w] = 0; |
1c27024e | 2123 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) |
b128c09f | 2124 | pio->io_child_error[c] = 0; |
34dc7c2f | 2125 | |
428870ff BB |
2126 | if (IO_IS_ALLOCATING(pio)) |
2127 | BP_ZERO(pio->io_bp); | |
34dc7c2f | 2128 | |
b128c09f BB |
2129 | /* |
2130 | * As we reexecute pio's children, new children could be created. | |
d164b209 | 2131 | * New children go to the head of pio's io_child_list, however, |
b128c09f | 2132 | * so we will (correctly) not reexecute them. The key is that |
d164b209 BB |
2133 | * the remainder of pio's io_child_list, from 'cio_next' onward, |
2134 | * cannot be affected by any side effects of reexecuting 'cio'. | |
b128c09f | 2135 | */ |
1c27024e | 2136 | zio_link_t *zl = NULL; |
a8b2e306 | 2137 | mutex_enter(&pio->io_lock); |
3dfb57a3 DB |
2138 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { |
2139 | cio_next = zio_walk_children(pio, &zl); | |
1c27024e | 2140 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 | 2141 | pio->io_children[cio->io_child_type][w]++; |
b128c09f | 2142 | mutex_exit(&pio->io_lock); |
d164b209 | 2143 | zio_reexecute(cio); |
a8b2e306 | 2144 | mutex_enter(&pio->io_lock); |
34dc7c2f | 2145 | } |
a8b2e306 | 2146 | mutex_exit(&pio->io_lock); |
34dc7c2f | 2147 | |
b128c09f BB |
2148 | /* |
2149 | * Now that all children have been reexecuted, execute the parent. | |
9babb374 | 2150 | * We don't reexecute "The Godfather" I/O here as it's the |
9e2c3bb4 | 2151 | * responsibility of the caller to wait on it. |
b128c09f | 2152 | */ |
3dfb57a3 DB |
2153 | if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) { |
2154 | pio->io_queued_timestamp = gethrtime(); | |
da6b4005 | 2155 | __zio_execute(pio); |
3dfb57a3 | 2156 | } |
34dc7c2f BB |
2157 | } |
2158 | ||
b128c09f | 2159 | void |
cec3a0a1 | 2160 | zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason) |
34dc7c2f | 2161 | { |
b128c09f BB |
2162 | if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) |
2163 | fm_panic("Pool '%s' has encountered an uncorrectable I/O " | |
2164 | "failure and the failure mode property for this pool " | |
2165 | "is set to panic.", spa_name(spa)); | |
34dc7c2f | 2166 | |
bf89c199 BB |
2167 | cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O " |
2168 | "failure and has been suspended.\n", spa_name(spa)); | |
2169 | ||
b5256303 TC |
2170 | zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, |
2171 | NULL, NULL, 0, 0); | |
34dc7c2f | 2172 | |
b128c09f | 2173 | mutex_enter(&spa->spa_suspend_lock); |
34dc7c2f | 2174 | |
b128c09f | 2175 | if (spa->spa_suspend_zio_root == NULL) |
9babb374 BB |
2176 | spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, |
2177 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
2178 | ZIO_FLAG_GODFATHER); | |
34dc7c2f | 2179 | |
cec3a0a1 | 2180 | spa->spa_suspended = reason; |
34dc7c2f | 2181 | |
b128c09f | 2182 | if (zio != NULL) { |
9babb374 | 2183 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f BB |
2184 | ASSERT(zio != spa->spa_suspend_zio_root); |
2185 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
d164b209 | 2186 | ASSERT(zio_unique_parent(zio) == NULL); |
b128c09f BB |
2187 | ASSERT(zio->io_stage == ZIO_STAGE_DONE); |
2188 | zio_add_child(spa->spa_suspend_zio_root, zio); | |
2189 | } | |
34dc7c2f | 2190 | |
b128c09f BB |
2191 | mutex_exit(&spa->spa_suspend_lock); |
2192 | } | |
34dc7c2f | 2193 | |
9babb374 | 2194 | int |
b128c09f BB |
2195 | zio_resume(spa_t *spa) |
2196 | { | |
9babb374 | 2197 | zio_t *pio; |
34dc7c2f BB |
2198 | |
2199 | /* | |
b128c09f | 2200 | * Reexecute all previously suspended i/o. |
34dc7c2f | 2201 | */ |
b128c09f | 2202 | mutex_enter(&spa->spa_suspend_lock); |
cec3a0a1 | 2203 | spa->spa_suspended = ZIO_SUSPEND_NONE; |
b128c09f BB |
2204 | cv_broadcast(&spa->spa_suspend_cv); |
2205 | pio = spa->spa_suspend_zio_root; | |
2206 | spa->spa_suspend_zio_root = NULL; | |
2207 | mutex_exit(&spa->spa_suspend_lock); | |
2208 | ||
2209 | if (pio == NULL) | |
9babb374 | 2210 | return (0); |
34dc7c2f | 2211 | |
9babb374 BB |
2212 | zio_reexecute(pio); |
2213 | return (zio_wait(pio)); | |
b128c09f BB |
2214 | } |
2215 | ||
2216 | void | |
2217 | zio_resume_wait(spa_t *spa) | |
2218 | { | |
2219 | mutex_enter(&spa->spa_suspend_lock); | |
2220 | while (spa_suspended(spa)) | |
2221 | cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); | |
2222 | mutex_exit(&spa->spa_suspend_lock); | |
34dc7c2f BB |
2223 | } |
2224 | ||
2225 | /* | |
2226 | * ========================================================================== | |
b128c09f BB |
2227 | * Gang blocks. |
2228 | * | |
2229 | * A gang block is a collection of small blocks that looks to the DMU | |
2230 | * like one large block. When zio_dva_allocate() cannot find a block | |
2231 | * of the requested size, due to either severe fragmentation or the pool | |
2232 | * being nearly full, it calls zio_write_gang_block() to construct the | |
2233 | * block from smaller fragments. | |
2234 | * | |
2235 | * A gang block consists of a gang header (zio_gbh_phys_t) and up to | |
2236 | * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like | |
2237 | * an indirect block: it's an array of block pointers. It consumes | |
2238 | * only one sector and hence is allocatable regardless of fragmentation. | |
2239 | * The gang header's bps point to its gang members, which hold the data. | |
2240 | * | |
2241 | * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> | |
2242 | * as the verifier to ensure uniqueness of the SHA256 checksum. | |
2243 | * Critically, the gang block bp's blk_cksum is the checksum of the data, | |
2244 | * not the gang header. This ensures that data block signatures (needed for | |
2245 | * deduplication) are independent of how the block is physically stored. | |
2246 | * | |
2247 | * Gang blocks can be nested: a gang member may itself be a gang block. | |
2248 | * Thus every gang block is a tree in which root and all interior nodes are | |
2249 | * gang headers, and the leaves are normal blocks that contain user data. | |
2250 | * The root of the gang tree is called the gang leader. | |
2251 | * | |
2252 | * To perform any operation (read, rewrite, free, claim) on a gang block, | |
2253 | * zio_gang_assemble() first assembles the gang tree (minus data leaves) | |
2254 | * in the io_gang_tree field of the original logical i/o by recursively | |
2255 | * reading the gang leader and all gang headers below it. This yields | |
2256 | * an in-core tree containing the contents of every gang header and the | |
2257 | * bps for every constituent of the gang block. | |
2258 | * | |
2259 | * With the gang tree now assembled, zio_gang_issue() just walks the gang tree | |
2260 | * and invokes a callback on each bp. To free a gang block, zio_gang_issue() | |
2261 | * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. | |
2262 | * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). | |
2263 | * zio_read_gang() is a wrapper around zio_read() that omits reading gang | |
2264 | * headers, since we already have those in io_gang_tree. zio_rewrite_gang() | |
2265 | * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() | |
2266 | * of the gang header plus zio_checksum_compute() of the data to update the | |
2267 | * gang header's blk_cksum as described above. | |
2268 | * | |
2269 | * The two-phase assemble/issue model solves the problem of partial failure -- | |
2270 | * what if you'd freed part of a gang block but then couldn't read the | |
2271 | * gang header for another part? Assembling the entire gang tree first | |
2272 | * ensures that all the necessary gang header I/O has succeeded before | |
2273 | * starting the actual work of free, claim, or write. Once the gang tree | |
2274 | * is assembled, free and claim are in-memory operations that cannot fail. | |
2275 | * | |
2276 | * In the event that a gang write fails, zio_dva_unallocate() walks the | |
2277 | * gang tree to immediately free (i.e. insert back into the space map) | |
2278 | * everything we've allocated. This ensures that we don't get ENOSPC | |
2279 | * errors during repeated suspend/resume cycles due to a flaky device. | |
2280 | * | |
2281 | * Gang rewrites only happen during sync-to-convergence. If we can't assemble | |
2282 | * the gang tree, we won't modify the block, so we can safely defer the free | |
2283 | * (knowing that the block is still intact). If we *can* assemble the gang | |
2284 | * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free | |
2285 | * each constituent bp and we can allocate a new block on the next sync pass. | |
2286 | * | |
2287 | * In all cases, the gang tree allows complete recovery from partial failure. | |
34dc7c2f BB |
2288 | * ========================================================================== |
2289 | */ | |
b128c09f | 2290 | |
a6255b7f DQ |
2291 | static void |
2292 | zio_gang_issue_func_done(zio_t *zio) | |
2293 | { | |
2294 | abd_put(zio->io_abd); | |
2295 | } | |
2296 | ||
b128c09f | 2297 | static zio_t * |
a6255b7f DQ |
2298 | zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, |
2299 | uint64_t offset) | |
34dc7c2f | 2300 | { |
b128c09f BB |
2301 | if (gn != NULL) |
2302 | return (pio); | |
34dc7c2f | 2303 | |
a6255b7f DQ |
2304 | return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset), |
2305 | BP_GET_PSIZE(bp), zio_gang_issue_func_done, | |
2306 | NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
b128c09f BB |
2307 | &pio->io_bookmark)); |
2308 | } | |
2309 | ||
a6255b7f DQ |
2310 | static zio_t * |
2311 | zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2312 | uint64_t offset) | |
b128c09f BB |
2313 | { |
2314 | zio_t *zio; | |
2315 | ||
2316 | if (gn != NULL) { | |
a6255b7f DQ |
2317 | abd_t *gbh_abd = |
2318 | abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
b128c09f | 2319 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
a6255b7f DQ |
2320 | gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL, |
2321 | pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
2322 | &pio->io_bookmark); | |
34dc7c2f | 2323 | /* |
b128c09f BB |
2324 | * As we rewrite each gang header, the pipeline will compute |
2325 | * a new gang block header checksum for it; but no one will | |
2326 | * compute a new data checksum, so we do that here. The one | |
2327 | * exception is the gang leader: the pipeline already computed | |
2328 | * its data checksum because that stage precedes gang assembly. | |
2329 | * (Presently, nothing actually uses interior data checksums; | |
2330 | * this is just good hygiene.) | |
34dc7c2f | 2331 | */ |
9babb374 | 2332 | if (gn != pio->io_gang_leader->io_gang_tree) { |
a6255b7f DQ |
2333 | abd_t *buf = abd_get_offset(data, offset); |
2334 | ||
b128c09f | 2335 | zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), |
a6255b7f DQ |
2336 | buf, BP_GET_PSIZE(bp)); |
2337 | ||
2338 | abd_put(buf); | |
b128c09f | 2339 | } |
428870ff BB |
2340 | /* |
2341 | * If we are here to damage data for testing purposes, | |
2342 | * leave the GBH alone so that we can detect the damage. | |
2343 | */ | |
2344 | if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) | |
2345 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; | |
34dc7c2f | 2346 | } else { |
b128c09f | 2347 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
a6255b7f DQ |
2348 | abd_get_offset(data, offset), BP_GET_PSIZE(bp), |
2349 | zio_gang_issue_func_done, NULL, pio->io_priority, | |
b128c09f | 2350 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); |
34dc7c2f BB |
2351 | } |
2352 | ||
b128c09f BB |
2353 | return (zio); |
2354 | } | |
34dc7c2f | 2355 | |
b128c09f | 2356 | /* ARGSUSED */ |
a6255b7f DQ |
2357 | static zio_t * |
2358 | zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2359 | uint64_t offset) | |
b128c09f | 2360 | { |
428870ff BB |
2361 | return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, |
2362 | ZIO_GANG_CHILD_FLAGS(pio))); | |
34dc7c2f BB |
2363 | } |
2364 | ||
b128c09f | 2365 | /* ARGSUSED */ |
a6255b7f DQ |
2366 | static zio_t * |
2367 | zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2368 | uint64_t offset) | |
34dc7c2f | 2369 | { |
b128c09f BB |
2370 | return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, |
2371 | NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); | |
2372 | } | |
2373 | ||
2374 | static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { | |
2375 | NULL, | |
2376 | zio_read_gang, | |
2377 | zio_rewrite_gang, | |
2378 | zio_free_gang, | |
2379 | zio_claim_gang, | |
2380 | NULL | |
2381 | }; | |
34dc7c2f | 2382 | |
b128c09f | 2383 | static void zio_gang_tree_assemble_done(zio_t *zio); |
34dc7c2f | 2384 | |
b128c09f BB |
2385 | static zio_gang_node_t * |
2386 | zio_gang_node_alloc(zio_gang_node_t **gnpp) | |
2387 | { | |
2388 | zio_gang_node_t *gn; | |
34dc7c2f | 2389 | |
b128c09f | 2390 | ASSERT(*gnpp == NULL); |
34dc7c2f | 2391 | |
79c76d5b | 2392 | gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); |
b128c09f BB |
2393 | gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); |
2394 | *gnpp = gn; | |
34dc7c2f | 2395 | |
b128c09f | 2396 | return (gn); |
34dc7c2f BB |
2397 | } |
2398 | ||
34dc7c2f | 2399 | static void |
b128c09f | 2400 | zio_gang_node_free(zio_gang_node_t **gnpp) |
34dc7c2f | 2401 | { |
b128c09f | 2402 | zio_gang_node_t *gn = *gnpp; |
34dc7c2f | 2403 | |
1c27024e | 2404 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f BB |
2405 | ASSERT(gn->gn_child[g] == NULL); |
2406 | ||
2407 | zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
2408 | kmem_free(gn, sizeof (*gn)); | |
2409 | *gnpp = NULL; | |
34dc7c2f BB |
2410 | } |
2411 | ||
b128c09f BB |
2412 | static void |
2413 | zio_gang_tree_free(zio_gang_node_t **gnpp) | |
34dc7c2f | 2414 | { |
b128c09f | 2415 | zio_gang_node_t *gn = *gnpp; |
34dc7c2f | 2416 | |
b128c09f BB |
2417 | if (gn == NULL) |
2418 | return; | |
34dc7c2f | 2419 | |
1c27024e | 2420 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f | 2421 | zio_gang_tree_free(&gn->gn_child[g]); |
34dc7c2f | 2422 | |
b128c09f | 2423 | zio_gang_node_free(gnpp); |
34dc7c2f BB |
2424 | } |
2425 | ||
b128c09f | 2426 | static void |
9babb374 | 2427 | zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) |
34dc7c2f | 2428 | { |
b128c09f | 2429 | zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); |
a6255b7f | 2430 | abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); |
b128c09f | 2431 | |
9babb374 | 2432 | ASSERT(gio->io_gang_leader == gio); |
b128c09f | 2433 | ASSERT(BP_IS_GANG(bp)); |
34dc7c2f | 2434 | |
a6255b7f DQ |
2435 | zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE, |
2436 | zio_gang_tree_assemble_done, gn, gio->io_priority, | |
2437 | ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); | |
b128c09f | 2438 | } |
34dc7c2f | 2439 | |
b128c09f BB |
2440 | static void |
2441 | zio_gang_tree_assemble_done(zio_t *zio) | |
2442 | { | |
9babb374 | 2443 | zio_t *gio = zio->io_gang_leader; |
b128c09f BB |
2444 | zio_gang_node_t *gn = zio->io_private; |
2445 | blkptr_t *bp = zio->io_bp; | |
34dc7c2f | 2446 | |
9babb374 | 2447 | ASSERT(gio == zio_unique_parent(zio)); |
428870ff | 2448 | ASSERT(zio->io_child_count == 0); |
34dc7c2f | 2449 | |
b128c09f BB |
2450 | if (zio->io_error) |
2451 | return; | |
34dc7c2f | 2452 | |
a6255b7f | 2453 | /* this ABD was created from a linear buf in zio_gang_tree_assemble */ |
b128c09f | 2454 | if (BP_SHOULD_BYTESWAP(bp)) |
a6255b7f | 2455 | byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size); |
34dc7c2f | 2456 | |
a6255b7f | 2457 | ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh); |
b128c09f | 2458 | ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); |
428870ff | 2459 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2460 | |
a6255b7f DQ |
2461 | abd_put(zio->io_abd); |
2462 | ||
1c27024e | 2463 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2464 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2465 | if (!BP_IS_GANG(gbp)) | |
2466 | continue; | |
9babb374 | 2467 | zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); |
b128c09f | 2468 | } |
34dc7c2f BB |
2469 | } |
2470 | ||
b128c09f | 2471 | static void |
a6255b7f DQ |
2472 | zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data, |
2473 | uint64_t offset) | |
34dc7c2f | 2474 | { |
9babb374 | 2475 | zio_t *gio = pio->io_gang_leader; |
b128c09f | 2476 | zio_t *zio; |
34dc7c2f | 2477 | |
b128c09f | 2478 | ASSERT(BP_IS_GANG(bp) == !!gn); |
9babb374 BB |
2479 | ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); |
2480 | ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); | |
34dc7c2f | 2481 | |
b128c09f BB |
2482 | /* |
2483 | * If you're a gang header, your data is in gn->gn_gbh. | |
2484 | * If you're a gang member, your data is in 'data' and gn == NULL. | |
2485 | */ | |
a6255b7f | 2486 | zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset); |
34dc7c2f | 2487 | |
b128c09f | 2488 | if (gn != NULL) { |
428870ff | 2489 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2490 | |
1c27024e | 2491 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2492 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2493 | if (BP_IS_HOLE(gbp)) | |
2494 | continue; | |
a6255b7f DQ |
2495 | zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data, |
2496 | offset); | |
2497 | offset += BP_GET_PSIZE(gbp); | |
b128c09f | 2498 | } |
34dc7c2f BB |
2499 | } |
2500 | ||
9babb374 | 2501 | if (gn == gio->io_gang_tree) |
a6255b7f | 2502 | ASSERT3U(gio->io_size, ==, offset); |
34dc7c2f | 2503 | |
b128c09f BB |
2504 | if (zio != pio) |
2505 | zio_nowait(zio); | |
34dc7c2f BB |
2506 | } |
2507 | ||
62840030 | 2508 | static zio_t * |
b128c09f | 2509 | zio_gang_assemble(zio_t *zio) |
34dc7c2f | 2510 | { |
b128c09f | 2511 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2512 | |
9babb374 BB |
2513 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); |
2514 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
2515 | ||
2516 | zio->io_gang_leader = zio; | |
34dc7c2f | 2517 | |
b128c09f | 2518 | zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); |
34dc7c2f | 2519 | |
62840030 | 2520 | return (zio); |
34dc7c2f BB |
2521 | } |
2522 | ||
62840030 | 2523 | static zio_t * |
b128c09f | 2524 | zio_gang_issue(zio_t *zio) |
34dc7c2f | 2525 | { |
b128c09f | 2526 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2527 | |
ddc751d5 | 2528 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) { |
62840030 | 2529 | return (NULL); |
ddc751d5 | 2530 | } |
34dc7c2f | 2531 | |
9babb374 BB |
2532 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); |
2533 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
34dc7c2f | 2534 | |
b128c09f | 2535 | if (zio->io_child_error[ZIO_CHILD_GANG] == 0) |
a6255b7f DQ |
2536 | zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd, |
2537 | 0); | |
b128c09f | 2538 | else |
9babb374 | 2539 | zio_gang_tree_free(&zio->io_gang_tree); |
34dc7c2f | 2540 | |
b128c09f | 2541 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f | 2542 | |
62840030 | 2543 | return (zio); |
34dc7c2f BB |
2544 | } |
2545 | ||
2546 | static void | |
b128c09f | 2547 | zio_write_gang_member_ready(zio_t *zio) |
34dc7c2f | 2548 | { |
d164b209 | 2549 | zio_t *pio = zio_unique_parent(zio); |
34dc7c2f BB |
2550 | dva_t *cdva = zio->io_bp->blk_dva; |
2551 | dva_t *pdva = pio->io_bp->blk_dva; | |
2552 | uint64_t asize; | |
d1d7e268 | 2553 | ASSERTV(zio_t *gio = zio->io_gang_leader); |
34dc7c2f | 2554 | |
b128c09f BB |
2555 | if (BP_IS_HOLE(zio->io_bp)) |
2556 | return; | |
2557 | ||
2558 | ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); | |
2559 | ||
2560 | ASSERT(zio->io_child_type == ZIO_CHILD_GANG); | |
428870ff BB |
2561 | ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); |
2562 | ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); | |
2563 | ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); | |
34dc7c2f | 2564 | ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); |
34dc7c2f BB |
2565 | |
2566 | mutex_enter(&pio->io_lock); | |
1c27024e | 2567 | for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { |
34dc7c2f BB |
2568 | ASSERT(DVA_GET_GANG(&pdva[d])); |
2569 | asize = DVA_GET_ASIZE(&pdva[d]); | |
2570 | asize += DVA_GET_ASIZE(&cdva[d]); | |
2571 | DVA_SET_ASIZE(&pdva[d], asize); | |
2572 | } | |
2573 | mutex_exit(&pio->io_lock); | |
2574 | } | |
2575 | ||
a6255b7f DQ |
2576 | static void |
2577 | zio_write_gang_done(zio_t *zio) | |
2578 | { | |
c955398b BL |
2579 | /* |
2580 | * The io_abd field will be NULL for a zio with no data. The io_flags | |
2581 | * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't | |
2582 | * check for it here as it is cleared in zio_ready. | |
2583 | */ | |
2584 | if (zio->io_abd != NULL) | |
2585 | abd_put(zio->io_abd); | |
a6255b7f DQ |
2586 | } |
2587 | ||
62840030 | 2588 | static zio_t * |
b128c09f | 2589 | zio_write_gang_block(zio_t *pio) |
34dc7c2f | 2590 | { |
b128c09f | 2591 | spa_t *spa = pio->io_spa; |
3dfb57a3 | 2592 | metaslab_class_t *mc = spa_normal_class(spa); |
b128c09f | 2593 | blkptr_t *bp = pio->io_bp; |
9babb374 | 2594 | zio_t *gio = pio->io_gang_leader; |
b128c09f BB |
2595 | zio_t *zio; |
2596 | zio_gang_node_t *gn, **gnpp; | |
34dc7c2f | 2597 | zio_gbh_phys_t *gbh; |
a6255b7f | 2598 | abd_t *gbh_abd; |
b128c09f BB |
2599 | uint64_t txg = pio->io_txg; |
2600 | uint64_t resid = pio->io_size; | |
2601 | uint64_t lsize; | |
428870ff | 2602 | int copies = gio->io_prop.zp_copies; |
b5256303 | 2603 | int gbh_copies; |
b128c09f | 2604 | zio_prop_t zp; |
1c27024e | 2605 | int error; |
c955398b | 2606 | boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA); |
b5256303 TC |
2607 | |
2608 | /* | |
2609 | * encrypted blocks need DVA[2] free so encrypted gang headers can't | |
2610 | * have a third copy. | |
2611 | */ | |
2612 | gbh_copies = MIN(copies + 1, spa_max_replication(spa)); | |
2613 | if (gio->io_prop.zp_encrypt && gbh_copies >= SPA_DVAS_PER_BP) | |
2614 | gbh_copies = SPA_DVAS_PER_BP - 1; | |
2615 | ||
1c27024e | 2616 | int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER; |
3dfb57a3 DB |
2617 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
2618 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
c955398b | 2619 | ASSERT(has_data); |
3dfb57a3 DB |
2620 | |
2621 | flags |= METASLAB_ASYNC_ALLOC; | |
424fd7c3 | 2622 | VERIFY(zfs_refcount_held(&mc->mc_alloc_slots[pio->io_allocator], |
492f64e9 | 2623 | pio)); |
3dfb57a3 DB |
2624 | |
2625 | /* | |
2626 | * The logical zio has already placed a reservation for | |
2627 | * 'copies' allocation slots but gang blocks may require | |
2628 | * additional copies. These additional copies | |
2629 | * (i.e. gbh_copies - copies) are guaranteed to succeed | |
2630 | * since metaslab_class_throttle_reserve() always allows | |
2631 | * additional reservations for gang blocks. | |
2632 | */ | |
2633 | VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies, | |
492f64e9 | 2634 | pio->io_allocator, pio, flags)); |
3dfb57a3 DB |
2635 | } |
2636 | ||
2637 | error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE, | |
4e21fd06 | 2638 | bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, |
492f64e9 | 2639 | &pio->io_alloc_list, pio, pio->io_allocator); |
34dc7c2f | 2640 | if (error) { |
3dfb57a3 DB |
2641 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
2642 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
c955398b | 2643 | ASSERT(has_data); |
3dfb57a3 DB |
2644 | |
2645 | /* | |
2646 | * If we failed to allocate the gang block header then | |
2647 | * we remove any additional allocation reservations that | |
2648 | * we placed here. The original reservation will | |
2649 | * be removed when the logical I/O goes to the ready | |
2650 | * stage. | |
2651 | */ | |
2652 | metaslab_class_throttle_unreserve(mc, | |
492f64e9 | 2653 | gbh_copies - copies, pio->io_allocator, pio); |
3dfb57a3 DB |
2654 | } |
2655 | ||
b128c09f | 2656 | pio->io_error = error; |
62840030 | 2657 | return (pio); |
34dc7c2f BB |
2658 | } |
2659 | ||
9babb374 BB |
2660 | if (pio == gio) { |
2661 | gnpp = &gio->io_gang_tree; | |
b128c09f BB |
2662 | } else { |
2663 | gnpp = pio->io_private; | |
2664 | ASSERT(pio->io_ready == zio_write_gang_member_ready); | |
34dc7c2f BB |
2665 | } |
2666 | ||
b128c09f BB |
2667 | gn = zio_gang_node_alloc(gnpp); |
2668 | gbh = gn->gn_gbh; | |
2669 | bzero(gbh, SPA_GANGBLOCKSIZE); | |
a6255b7f | 2670 | gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE); |
34dc7c2f | 2671 | |
b128c09f BB |
2672 | /* |
2673 | * Create the gang header. | |
2674 | */ | |
a6255b7f DQ |
2675 | zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE, |
2676 | zio_write_gang_done, NULL, pio->io_priority, | |
2677 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); | |
34dc7c2f | 2678 | |
b128c09f BB |
2679 | /* |
2680 | * Create and nowait the gang children. | |
2681 | */ | |
1c27024e | 2682 | for (int g = 0; resid != 0; resid -= lsize, g++) { |
b128c09f BB |
2683 | lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), |
2684 | SPA_MINBLOCKSIZE); | |
2685 | ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); | |
2686 | ||
9babb374 | 2687 | zp.zp_checksum = gio->io_prop.zp_checksum; |
b128c09f BB |
2688 | zp.zp_compress = ZIO_COMPRESS_OFF; |
2689 | zp.zp_type = DMU_OT_NONE; | |
2690 | zp.zp_level = 0; | |
428870ff | 2691 | zp.zp_copies = gio->io_prop.zp_copies; |
03c6040b GW |
2692 | zp.zp_dedup = B_FALSE; |
2693 | zp.zp_dedup_verify = B_FALSE; | |
2694 | zp.zp_nopwrite = B_FALSE; | |
4807c0ba TC |
2695 | zp.zp_encrypt = gio->io_prop.zp_encrypt; |
2696 | zp.zp_byteorder = gio->io_prop.zp_byteorder; | |
b5256303 TC |
2697 | bzero(zp.zp_salt, ZIO_DATA_SALT_LEN); |
2698 | bzero(zp.zp_iv, ZIO_DATA_IV_LEN); | |
2699 | bzero(zp.zp_mac, ZIO_DATA_MAC_LEN); | |
b128c09f | 2700 | |
1c27024e | 2701 | zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g], |
c955398b BL |
2702 | has_data ? abd_get_offset(pio->io_abd, pio->io_size - |
2703 | resid) : NULL, lsize, lsize, &zp, | |
2704 | zio_write_gang_member_ready, NULL, NULL, | |
a6255b7f | 2705 | zio_write_gang_done, &gn->gn_child[g], pio->io_priority, |
3dfb57a3 DB |
2706 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); |
2707 | ||
2708 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { | |
2709 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
c955398b | 2710 | ASSERT(has_data); |
3dfb57a3 DB |
2711 | |
2712 | /* | |
2713 | * Gang children won't throttle but we should | |
2714 | * account for their work, so reserve an allocation | |
2715 | * slot for them here. | |
2716 | */ | |
2717 | VERIFY(metaslab_class_throttle_reserve(mc, | |
492f64e9 | 2718 | zp.zp_copies, cio->io_allocator, cio, flags)); |
3dfb57a3 DB |
2719 | } |
2720 | zio_nowait(cio); | |
b128c09f | 2721 | } |
34dc7c2f BB |
2722 | |
2723 | /* | |
b128c09f | 2724 | * Set pio's pipeline to just wait for zio to finish. |
34dc7c2f | 2725 | */ |
b128c09f BB |
2726 | pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
2727 | ||
920dd524 ED |
2728 | /* |
2729 | * We didn't allocate this bp, so make sure it doesn't get unmarked. | |
2730 | */ | |
2731 | pio->io_flags &= ~ZIO_FLAG_FASTWRITE; | |
2732 | ||
b128c09f BB |
2733 | zio_nowait(zio); |
2734 | ||
62840030 | 2735 | return (pio); |
34dc7c2f BB |
2736 | } |
2737 | ||
03c6040b | 2738 | /* |
3c67d83a TH |
2739 | * The zio_nop_write stage in the pipeline determines if allocating a |
2740 | * new bp is necessary. The nopwrite feature can handle writes in | |
2741 | * either syncing or open context (i.e. zil writes) and as a result is | |
2742 | * mutually exclusive with dedup. | |
2743 | * | |
2744 | * By leveraging a cryptographically secure checksum, such as SHA256, we | |
2745 | * can compare the checksums of the new data and the old to determine if | |
2746 | * allocating a new block is required. Note that our requirements for | |
2747 | * cryptographic strength are fairly weak: there can't be any accidental | |
2748 | * hash collisions, but we don't need to be secure against intentional | |
2749 | * (malicious) collisions. To trigger a nopwrite, you have to be able | |
2750 | * to write the file to begin with, and triggering an incorrect (hash | |
2751 | * collision) nopwrite is no worse than simply writing to the file. | |
2752 | * That said, there are no known attacks against the checksum algorithms | |
2753 | * used for nopwrite, assuming that the salt and the checksums | |
2754 | * themselves remain secret. | |
03c6040b | 2755 | */ |
62840030 | 2756 | static zio_t * |
03c6040b GW |
2757 | zio_nop_write(zio_t *zio) |
2758 | { | |
2759 | blkptr_t *bp = zio->io_bp; | |
2760 | blkptr_t *bp_orig = &zio->io_bp_orig; | |
2761 | zio_prop_t *zp = &zio->io_prop; | |
2762 | ||
2763 | ASSERT(BP_GET_LEVEL(bp) == 0); | |
2764 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
2765 | ASSERT(zp->zp_nopwrite); | |
2766 | ASSERT(!zp->zp_dedup); | |
2767 | ASSERT(zio->io_bp_override == NULL); | |
2768 | ASSERT(IO_IS_ALLOCATING(zio)); | |
2769 | ||
2770 | /* | |
2771 | * Check to see if the original bp and the new bp have matching | |
2772 | * characteristics (i.e. same checksum, compression algorithms, etc). | |
2773 | * If they don't then just continue with the pipeline which will | |
2774 | * allocate a new bp. | |
2775 | */ | |
2776 | if (BP_IS_HOLE(bp_orig) || | |
3c67d83a TH |
2777 | !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & |
2778 | ZCHECKSUM_FLAG_NOPWRITE) || | |
b5256303 | 2779 | BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) || |
03c6040b GW |
2780 | BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || |
2781 | BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || | |
2782 | BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || | |
2783 | zp->zp_copies != BP_GET_NDVAS(bp_orig)) | |
62840030 | 2784 | return (zio); |
03c6040b GW |
2785 | |
2786 | /* | |
2787 | * If the checksums match then reset the pipeline so that we | |
2788 | * avoid allocating a new bp and issuing any I/O. | |
2789 | */ | |
2790 | if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { | |
3c67d83a TH |
2791 | ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & |
2792 | ZCHECKSUM_FLAG_NOPWRITE); | |
03c6040b GW |
2793 | ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); |
2794 | ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); | |
2795 | ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); | |
2796 | ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, | |
2797 | sizeof (uint64_t)) == 0); | |
2798 | ||
681a85cb GW |
2799 | /* |
2800 | * If we're overwriting a block that is currently on an | |
2801 | * indirect vdev, then ignore the nopwrite request and | |
2802 | * allow a new block to be allocated on a concrete vdev. | |
2803 | */ | |
2804 | spa_config_enter(zio->io_spa, SCL_VDEV, FTAG, RW_READER); | |
2805 | vdev_t *tvd = vdev_lookup_top(zio->io_spa, | |
2806 | DVA_GET_VDEV(&bp->blk_dva[0])); | |
2807 | if (tvd->vdev_ops == &vdev_indirect_ops) { | |
2808 | spa_config_exit(zio->io_spa, SCL_VDEV, FTAG); | |
2809 | return (zio); | |
2810 | } | |
2811 | spa_config_exit(zio->io_spa, SCL_VDEV, FTAG); | |
2812 | ||
03c6040b GW |
2813 | *bp = *bp_orig; |
2814 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
2815 | zio->io_flags |= ZIO_FLAG_NOPWRITE; | |
2816 | } | |
2817 | ||
62840030 | 2818 | return (zio); |
03c6040b GW |
2819 | } |
2820 | ||
34dc7c2f BB |
2821 | /* |
2822 | * ========================================================================== | |
428870ff | 2823 | * Dedup |
34dc7c2f BB |
2824 | * ========================================================================== |
2825 | */ | |
428870ff BB |
2826 | static void |
2827 | zio_ddt_child_read_done(zio_t *zio) | |
2828 | { | |
2829 | blkptr_t *bp = zio->io_bp; | |
2830 | ddt_entry_t *dde = zio->io_private; | |
2831 | ddt_phys_t *ddp; | |
2832 | zio_t *pio = zio_unique_parent(zio); | |
2833 | ||
2834 | mutex_enter(&pio->io_lock); | |
2835 | ddp = ddt_phys_select(dde, bp); | |
2836 | if (zio->io_error == 0) | |
2837 | ddt_phys_clear(ddp); /* this ddp doesn't need repair */ | |
a6255b7f DQ |
2838 | |
2839 | if (zio->io_error == 0 && dde->dde_repair_abd == NULL) | |
2840 | dde->dde_repair_abd = zio->io_abd; | |
428870ff | 2841 | else |
a6255b7f | 2842 | abd_free(zio->io_abd); |
428870ff BB |
2843 | mutex_exit(&pio->io_lock); |
2844 | } | |
2845 | ||
62840030 | 2846 | static zio_t * |
428870ff BB |
2847 | zio_ddt_read_start(zio_t *zio) |
2848 | { | |
2849 | blkptr_t *bp = zio->io_bp; | |
2850 | ||
2851 | ASSERT(BP_GET_DEDUP(bp)); | |
2852 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
2853 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
2854 | ||
2855 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
2856 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
2857 | ddt_entry_t *dde = ddt_repair_start(ddt, bp); | |
2858 | ddt_phys_t *ddp = dde->dde_phys; | |
2859 | ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); | |
2860 | blkptr_t blk; | |
2861 | ||
2862 | ASSERT(zio->io_vsd == NULL); | |
2863 | zio->io_vsd = dde; | |
2864 | ||
2865 | if (ddp_self == NULL) | |
62840030 | 2866 | return (zio); |
428870ff | 2867 | |
1c27024e | 2868 | for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { |
428870ff BB |
2869 | if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) |
2870 | continue; | |
2871 | ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, | |
2872 | &blk); | |
2873 | zio_nowait(zio_read(zio, zio->io_spa, &blk, | |
a6255b7f DQ |
2874 | abd_alloc_for_io(zio->io_size, B_TRUE), |
2875 | zio->io_size, zio_ddt_child_read_done, dde, | |
2876 | zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) | | |
2877 | ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark)); | |
428870ff | 2878 | } |
62840030 | 2879 | return (zio); |
428870ff BB |
2880 | } |
2881 | ||
2882 | zio_nowait(zio_read(zio, zio->io_spa, bp, | |
a6255b7f | 2883 | zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority, |
428870ff BB |
2884 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); |
2885 | ||
62840030 | 2886 | return (zio); |
428870ff BB |
2887 | } |
2888 | ||
62840030 | 2889 | static zio_t * |
428870ff BB |
2890 | zio_ddt_read_done(zio_t *zio) |
2891 | { | |
2892 | blkptr_t *bp = zio->io_bp; | |
2893 | ||
ddc751d5 | 2894 | if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) { |
62840030 | 2895 | return (NULL); |
ddc751d5 | 2896 | } |
428870ff BB |
2897 | |
2898 | ASSERT(BP_GET_DEDUP(bp)); | |
2899 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
2900 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
2901 | ||
2902 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
2903 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
2904 | ddt_entry_t *dde = zio->io_vsd; | |
2905 | if (ddt == NULL) { | |
2906 | ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); | |
62840030 | 2907 | return (zio); |
428870ff BB |
2908 | } |
2909 | if (dde == NULL) { | |
2910 | zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; | |
2911 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); | |
62840030 | 2912 | return (NULL); |
428870ff | 2913 | } |
a6255b7f DQ |
2914 | if (dde->dde_repair_abd != NULL) { |
2915 | abd_copy(zio->io_abd, dde->dde_repair_abd, | |
2916 | zio->io_size); | |
428870ff BB |
2917 | zio->io_child_error[ZIO_CHILD_DDT] = 0; |
2918 | } | |
2919 | ddt_repair_done(ddt, dde); | |
2920 | zio->io_vsd = NULL; | |
2921 | } | |
2922 | ||
2923 | ASSERT(zio->io_vsd == NULL); | |
2924 | ||
62840030 | 2925 | return (zio); |
428870ff BB |
2926 | } |
2927 | ||
2928 | static boolean_t | |
2929 | zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) | |
2930 | { | |
2931 | spa_t *spa = zio->io_spa; | |
c17bcf83 | 2932 | boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW); |
428870ff | 2933 | |
c17bcf83 | 2934 | ASSERT(!(zio->io_bp_override && do_raw)); |
2aa34383 | 2935 | |
428870ff BB |
2936 | /* |
2937 | * Note: we compare the original data, not the transformed data, | |
2938 | * because when zio->io_bp is an override bp, we will not have | |
2939 | * pushed the I/O transforms. That's an important optimization | |
2940 | * because otherwise we'd compress/encrypt all dmu_sync() data twice. | |
c17bcf83 | 2941 | * However, we should never get a raw, override zio so in these |
b5256303 | 2942 | * cases we can compare the io_abd directly. This is useful because |
c17bcf83 TC |
2943 | * it allows us to do dedup verification even if we don't have access |
2944 | * to the original data (for instance, if the encryption keys aren't | |
2945 | * loaded). | |
428870ff | 2946 | */ |
c17bcf83 | 2947 | |
1c27024e | 2948 | for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
2949 | zio_t *lio = dde->dde_lead_zio[p]; |
2950 | ||
c17bcf83 TC |
2951 | if (lio != NULL && do_raw) { |
2952 | return (lio->io_size != zio->io_size || | |
a6255b7f | 2953 | abd_cmp(zio->io_abd, lio->io_abd) != 0); |
c17bcf83 | 2954 | } else if (lio != NULL) { |
428870ff | 2955 | return (lio->io_orig_size != zio->io_orig_size || |
a6255b7f | 2956 | abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0); |
428870ff BB |
2957 | } |
2958 | } | |
2959 | ||
1c27024e | 2960 | for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
2961 | ddt_phys_t *ddp = &dde->dde_phys[p]; |
2962 | ||
c17bcf83 TC |
2963 | if (ddp->ddp_phys_birth != 0 && do_raw) { |
2964 | blkptr_t blk = *zio->io_bp; | |
2965 | uint64_t psize; | |
a6255b7f | 2966 | abd_t *tmpabd; |
c17bcf83 TC |
2967 | int error; |
2968 | ||
2969 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
2970 | psize = BP_GET_PSIZE(&blk); | |
2971 | ||
2972 | if (psize != zio->io_size) | |
2973 | return (B_TRUE); | |
2974 | ||
2975 | ddt_exit(ddt); | |
2976 | ||
a6255b7f | 2977 | tmpabd = abd_alloc_for_io(psize, B_TRUE); |
c17bcf83 | 2978 | |
a6255b7f | 2979 | error = zio_wait(zio_read(NULL, spa, &blk, tmpabd, |
c17bcf83 TC |
2980 | psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ, |
2981 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
2982 | ZIO_FLAG_RAW, &zio->io_bookmark)); | |
2983 | ||
2984 | if (error == 0) { | |
a6255b7f | 2985 | if (abd_cmp(tmpabd, zio->io_abd) != 0) |
c17bcf83 TC |
2986 | error = SET_ERROR(ENOENT); |
2987 | } | |
2988 | ||
a6255b7f | 2989 | abd_free(tmpabd); |
c17bcf83 TC |
2990 | ddt_enter(ddt); |
2991 | return (error != 0); | |
2992 | } else if (ddp->ddp_phys_birth != 0) { | |
428870ff | 2993 | arc_buf_t *abuf = NULL; |
2a432414 | 2994 | arc_flags_t aflags = ARC_FLAG_WAIT; |
428870ff BB |
2995 | blkptr_t blk = *zio->io_bp; |
2996 | int error; | |
2997 | ||
2998 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
2999 | ||
c17bcf83 TC |
3000 | if (BP_GET_LSIZE(&blk) != zio->io_orig_size) |
3001 | return (B_TRUE); | |
3002 | ||
428870ff BB |
3003 | ddt_exit(ddt); |
3004 | ||
294f6806 | 3005 | error = arc_read(NULL, spa, &blk, |
428870ff BB |
3006 | arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, |
3007 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
3008 | &aflags, &zio->io_bookmark); | |
3009 | ||
3010 | if (error == 0) { | |
a6255b7f | 3011 | if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data, |
428870ff | 3012 | zio->io_orig_size) != 0) |
c17bcf83 | 3013 | error = SET_ERROR(ENOENT); |
d3c2ae1c | 3014 | arc_buf_destroy(abuf, &abuf); |
428870ff BB |
3015 | } |
3016 | ||
3017 | ddt_enter(ddt); | |
3018 | return (error != 0); | |
3019 | } | |
3020 | } | |
3021 | ||
3022 | return (B_FALSE); | |
3023 | } | |
3024 | ||
3025 | static void | |
3026 | zio_ddt_child_write_ready(zio_t *zio) | |
3027 | { | |
3028 | int p = zio->io_prop.zp_copies; | |
3029 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
3030 | ddt_entry_t *dde = zio->io_private; | |
3031 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
3032 | zio_t *pio; | |
3033 | ||
3034 | if (zio->io_error) | |
3035 | return; | |
3036 | ||
3037 | ddt_enter(ddt); | |
3038 | ||
3039 | ASSERT(dde->dde_lead_zio[p] == zio); | |
3040 | ||
3041 | ddt_phys_fill(ddp, zio->io_bp); | |
3042 | ||
1c27024e | 3043 | zio_link_t *zl = NULL; |
3dfb57a3 | 3044 | while ((pio = zio_walk_parents(zio, &zl)) != NULL) |
428870ff BB |
3045 | ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); |
3046 | ||
3047 | ddt_exit(ddt); | |
3048 | } | |
3049 | ||
3050 | static void | |
3051 | zio_ddt_child_write_done(zio_t *zio) | |
3052 | { | |
3053 | int p = zio->io_prop.zp_copies; | |
3054 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
3055 | ddt_entry_t *dde = zio->io_private; | |
3056 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
3057 | ||
3058 | ddt_enter(ddt); | |
3059 | ||
3060 | ASSERT(ddp->ddp_refcnt == 0); | |
3061 | ASSERT(dde->dde_lead_zio[p] == zio); | |
3062 | dde->dde_lead_zio[p] = NULL; | |
3063 | ||
3064 | if (zio->io_error == 0) { | |
3dfb57a3 DB |
3065 | zio_link_t *zl = NULL; |
3066 | while (zio_walk_parents(zio, &zl) != NULL) | |
428870ff BB |
3067 | ddt_phys_addref(ddp); |
3068 | } else { | |
3069 | ddt_phys_clear(ddp); | |
3070 | } | |
3071 | ||
3072 | ddt_exit(ddt); | |
3073 | } | |
3074 | ||
62840030 | 3075 | static zio_t * |
428870ff BB |
3076 | zio_ddt_write(zio_t *zio) |
3077 | { | |
3078 | spa_t *spa = zio->io_spa; | |
3079 | blkptr_t *bp = zio->io_bp; | |
3080 | uint64_t txg = zio->io_txg; | |
3081 | zio_prop_t *zp = &zio->io_prop; | |
3082 | int p = zp->zp_copies; | |
428870ff | 3083 | zio_t *cio = NULL; |
428870ff BB |
3084 | ddt_t *ddt = ddt_select(spa, bp); |
3085 | ddt_entry_t *dde; | |
3086 | ddt_phys_t *ddp; | |
3087 | ||
3088 | ASSERT(BP_GET_DEDUP(bp)); | |
3089 | ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); | |
3090 | ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); | |
c17bcf83 | 3091 | ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW))); |
428870ff BB |
3092 | |
3093 | ddt_enter(ddt); | |
3094 | dde = ddt_lookup(ddt, bp, B_TRUE); | |
3095 | ddp = &dde->dde_phys[p]; | |
3096 | ||
3097 | if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { | |
3098 | /* | |
3099 | * If we're using a weak checksum, upgrade to a strong checksum | |
3100 | * and try again. If we're already using a strong checksum, | |
3101 | * we can't resolve it, so just convert to an ordinary write. | |
3102 | * (And automatically e-mail a paper to Nature?) | |
3103 | */ | |
3c67d83a TH |
3104 | if (!(zio_checksum_table[zp->zp_checksum].ci_flags & |
3105 | ZCHECKSUM_FLAG_DEDUP)) { | |
428870ff BB |
3106 | zp->zp_checksum = spa_dedup_checksum(spa); |
3107 | zio_pop_transforms(zio); | |
3108 | zio->io_stage = ZIO_STAGE_OPEN; | |
3109 | BP_ZERO(bp); | |
3110 | } else { | |
03c6040b | 3111 | zp->zp_dedup = B_FALSE; |
accd6d9d | 3112 | BP_SET_DEDUP(bp, B_FALSE); |
428870ff | 3113 | } |
accd6d9d | 3114 | ASSERT(!BP_GET_DEDUP(bp)); |
428870ff BB |
3115 | zio->io_pipeline = ZIO_WRITE_PIPELINE; |
3116 | ddt_exit(ddt); | |
62840030 | 3117 | return (zio); |
428870ff BB |
3118 | } |
3119 | ||
428870ff BB |
3120 | if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { |
3121 | if (ddp->ddp_phys_birth != 0) | |
3122 | ddt_bp_fill(ddp, bp, txg); | |
3123 | if (dde->dde_lead_zio[p] != NULL) | |
3124 | zio_add_child(zio, dde->dde_lead_zio[p]); | |
3125 | else | |
3126 | ddt_phys_addref(ddp); | |
3127 | } else if (zio->io_bp_override) { | |
3128 | ASSERT(bp->blk_birth == txg); | |
3129 | ASSERT(BP_EQUAL(bp, zio->io_bp_override)); | |
3130 | ddt_phys_fill(ddp, bp); | |
3131 | ddt_phys_addref(ddp); | |
3132 | } else { | |
a6255b7f | 3133 | cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, |
2aa34383 | 3134 | zio->io_orig_size, zio->io_orig_size, zp, |
bc77ba73 | 3135 | zio_ddt_child_write_ready, NULL, NULL, |
428870ff BB |
3136 | zio_ddt_child_write_done, dde, zio->io_priority, |
3137 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); | |
3138 | ||
a6255b7f | 3139 | zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL); |
428870ff BB |
3140 | dde->dde_lead_zio[p] = cio; |
3141 | } | |
3142 | ||
3143 | ddt_exit(ddt); | |
3144 | ||
3145 | if (cio) | |
3146 | zio_nowait(cio); | |
428870ff | 3147 | |
62840030 | 3148 | return (zio); |
428870ff BB |
3149 | } |
3150 | ||
3151 | ddt_entry_t *freedde; /* for debugging */ | |
b128c09f | 3152 | |
62840030 | 3153 | static zio_t * |
428870ff BB |
3154 | zio_ddt_free(zio_t *zio) |
3155 | { | |
3156 | spa_t *spa = zio->io_spa; | |
3157 | blkptr_t *bp = zio->io_bp; | |
3158 | ddt_t *ddt = ddt_select(spa, bp); | |
3159 | ddt_entry_t *dde; | |
3160 | ddt_phys_t *ddp; | |
3161 | ||
3162 | ASSERT(BP_GET_DEDUP(bp)); | |
3163 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
3164 | ||
3165 | ddt_enter(ddt); | |
3166 | freedde = dde = ddt_lookup(ddt, bp, B_TRUE); | |
5dc6af0e BB |
3167 | if (dde) { |
3168 | ddp = ddt_phys_select(dde, bp); | |
3169 | if (ddp) | |
3170 | ddt_phys_decref(ddp); | |
3171 | } | |
428870ff BB |
3172 | ddt_exit(ddt); |
3173 | ||
62840030 | 3174 | return (zio); |
428870ff BB |
3175 | } |
3176 | ||
3177 | /* | |
3178 | * ========================================================================== | |
3179 | * Allocate and free blocks | |
3180 | * ========================================================================== | |
3181 | */ | |
3dfb57a3 DB |
3182 | |
3183 | static zio_t * | |
492f64e9 | 3184 | zio_io_to_allocate(spa_t *spa, int allocator) |
3dfb57a3 DB |
3185 | { |
3186 | zio_t *zio; | |
3187 | ||
492f64e9 | 3188 | ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator])); |
3dfb57a3 | 3189 | |
492f64e9 | 3190 | zio = avl_first(&spa->spa_alloc_trees[allocator]); |
3dfb57a3 DB |
3191 | if (zio == NULL) |
3192 | return (NULL); | |
3193 | ||
3194 | ASSERT(IO_IS_ALLOCATING(zio)); | |
3195 | ||
3196 | /* | |
3197 | * Try to place a reservation for this zio. If we're unable to | |
3198 | * reserve then we throttle. | |
3199 | */ | |
492f64e9 | 3200 | ASSERT3U(zio->io_allocator, ==, allocator); |
cc99f275 | 3201 | if (!metaslab_class_throttle_reserve(zio->io_metaslab_class, |
492f64e9 | 3202 | zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) { |
3dfb57a3 DB |
3203 | return (NULL); |
3204 | } | |
3205 | ||
492f64e9 | 3206 | avl_remove(&spa->spa_alloc_trees[allocator], zio); |
3dfb57a3 DB |
3207 | ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE); |
3208 | ||
3209 | return (zio); | |
3210 | } | |
3211 | ||
62840030 | 3212 | static zio_t * |
3dfb57a3 DB |
3213 | zio_dva_throttle(zio_t *zio) |
3214 | { | |
3215 | spa_t *spa = zio->io_spa; | |
3216 | zio_t *nio; | |
cc99f275 DB |
3217 | metaslab_class_t *mc; |
3218 | ||
3219 | /* locate an appropriate allocation class */ | |
3220 | mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type, | |
3221 | zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk); | |
3dfb57a3 DB |
3222 | |
3223 | if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE || | |
cc99f275 | 3224 | !mc->mc_alloc_throttle_enabled || |
3dfb57a3 DB |
3225 | zio->io_child_type == ZIO_CHILD_GANG || |
3226 | zio->io_flags & ZIO_FLAG_NODATA) { | |
62840030 | 3227 | return (zio); |
3dfb57a3 DB |
3228 | } |
3229 | ||
3230 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
3231 | ||
3232 | ASSERT3U(zio->io_queued_timestamp, >, 0); | |
3233 | ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE); | |
3234 | ||
492f64e9 PD |
3235 | zbookmark_phys_t *bm = &zio->io_bookmark; |
3236 | /* | |
3237 | * We want to try to use as many allocators as possible to help improve | |
3238 | * performance, but we also want logically adjacent IOs to be physically | |
3239 | * adjacent to improve sequential read performance. We chunk each object | |
3240 | * into 2^20 block regions, and then hash based on the objset, object, | |
3241 | * level, and region to accomplish both of these goals. | |
3242 | */ | |
3243 | zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object, | |
3244 | bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count; | |
3245 | mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]); | |
3dfb57a3 | 3246 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); |
cc99f275 | 3247 | zio->io_metaslab_class = mc; |
492f64e9 | 3248 | avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio); |
cc99f275 | 3249 | nio = zio_io_to_allocate(spa, zio->io_allocator); |
492f64e9 | 3250 | mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]); |
62840030 | 3251 | return (nio); |
3dfb57a3 DB |
3252 | } |
3253 | ||
cc99f275 | 3254 | static void |
492f64e9 | 3255 | zio_allocate_dispatch(spa_t *spa, int allocator) |
3dfb57a3 DB |
3256 | { |
3257 | zio_t *zio; | |
3258 | ||
492f64e9 PD |
3259 | mutex_enter(&spa->spa_alloc_locks[allocator]); |
3260 | zio = zio_io_to_allocate(spa, allocator); | |
3261 | mutex_exit(&spa->spa_alloc_locks[allocator]); | |
3dfb57a3 DB |
3262 | if (zio == NULL) |
3263 | return; | |
3264 | ||
3265 | ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE); | |
3266 | ASSERT0(zio->io_error); | |
3267 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE); | |
3268 | } | |
3269 | ||
62840030 | 3270 | static zio_t * |
34dc7c2f BB |
3271 | zio_dva_allocate(zio_t *zio) |
3272 | { | |
3273 | spa_t *spa = zio->io_spa; | |
cc99f275 | 3274 | metaslab_class_t *mc; |
34dc7c2f BB |
3275 | blkptr_t *bp = zio->io_bp; |
3276 | int error; | |
6d974228 | 3277 | int flags = 0; |
34dc7c2f | 3278 | |
9babb374 BB |
3279 | if (zio->io_gang_leader == NULL) { |
3280 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
3281 | zio->io_gang_leader = zio; | |
3282 | } | |
3283 | ||
34dc7c2f | 3284 | ASSERT(BP_IS_HOLE(bp)); |
c99c9001 | 3285 | ASSERT0(BP_GET_NDVAS(bp)); |
428870ff BB |
3286 | ASSERT3U(zio->io_prop.zp_copies, >, 0); |
3287 | ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); | |
34dc7c2f BB |
3288 | ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); |
3289 | ||
920dd524 | 3290 | flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0; |
3dfb57a3 DB |
3291 | if (zio->io_flags & ZIO_FLAG_NODATA) |
3292 | flags |= METASLAB_DONT_THROTTLE; | |
3293 | if (zio->io_flags & ZIO_FLAG_GANG_CHILD) | |
3294 | flags |= METASLAB_GANG_CHILD; | |
3295 | if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) | |
3296 | flags |= METASLAB_ASYNC_ALLOC; | |
3297 | ||
cc99f275 DB |
3298 | /* |
3299 | * if not already chosen, locate an appropriate allocation class | |
3300 | */ | |
3301 | mc = zio->io_metaslab_class; | |
3302 | if (mc == NULL) { | |
3303 | mc = spa_preferred_class(spa, zio->io_size, | |
3304 | zio->io_prop.zp_type, zio->io_prop.zp_level, | |
3305 | zio->io_prop.zp_zpl_smallblk); | |
3306 | zio->io_metaslab_class = mc; | |
3307 | } | |
3308 | ||
b128c09f | 3309 | error = metaslab_alloc(spa, mc, zio->io_size, bp, |
4e21fd06 | 3310 | zio->io_prop.zp_copies, zio->io_txg, NULL, flags, |
492f64e9 | 3311 | &zio->io_alloc_list, zio, zio->io_allocator); |
34dc7c2f | 3312 | |
cc99f275 DB |
3313 | /* |
3314 | * Fallback to normal class when an alloc class is full | |
3315 | */ | |
3316 | if (error == ENOSPC && mc != spa_normal_class(spa)) { | |
3317 | /* | |
3318 | * If throttling, transfer reservation over to normal class. | |
3319 | * The io_allocator slot can remain the same even though we | |
3320 | * are switching classes. | |
3321 | */ | |
3322 | if (mc->mc_alloc_throttle_enabled && | |
3323 | (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) { | |
3324 | metaslab_class_throttle_unreserve(mc, | |
3325 | zio->io_prop.zp_copies, zio->io_allocator, zio); | |
3326 | zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING; | |
3327 | ||
3328 | mc = spa_normal_class(spa); | |
3329 | VERIFY(metaslab_class_throttle_reserve(mc, | |
3330 | zio->io_prop.zp_copies, zio->io_allocator, zio, | |
3331 | flags | METASLAB_MUST_RESERVE)); | |
3332 | } else { | |
3333 | mc = spa_normal_class(spa); | |
3334 | } | |
3335 | zio->io_metaslab_class = mc; | |
3336 | ||
3337 | error = metaslab_alloc(spa, mc, zio->io_size, bp, | |
3338 | zio->io_prop.zp_copies, zio->io_txg, NULL, flags, | |
3339 | &zio->io_alloc_list, zio, zio->io_allocator); | |
3340 | } | |
3341 | ||
3dfb57a3 | 3342 | if (error != 0) { |
a887d653 | 3343 | zfs_dbgmsg("%s: metaslab allocation failure: zio %px, " |
6d974228 GW |
3344 | "size %llu, error %d", spa_name(spa), zio, zio->io_size, |
3345 | error); | |
b128c09f BB |
3346 | if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) |
3347 | return (zio_write_gang_block(zio)); | |
34dc7c2f BB |
3348 | zio->io_error = error; |
3349 | } | |
3350 | ||
62840030 | 3351 | return (zio); |
34dc7c2f BB |
3352 | } |
3353 | ||
62840030 | 3354 | static zio_t * |
34dc7c2f BB |
3355 | zio_dva_free(zio_t *zio) |
3356 | { | |
b128c09f | 3357 | metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); |
34dc7c2f | 3358 | |
62840030 | 3359 | return (zio); |
34dc7c2f BB |
3360 | } |
3361 | ||
62840030 | 3362 | static zio_t * |
34dc7c2f BB |
3363 | zio_dva_claim(zio_t *zio) |
3364 | { | |
b128c09f BB |
3365 | int error; |
3366 | ||
3367 | error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); | |
3368 | if (error) | |
3369 | zio->io_error = error; | |
34dc7c2f | 3370 | |
62840030 | 3371 | return (zio); |
34dc7c2f BB |
3372 | } |
3373 | ||
b128c09f BB |
3374 | /* |
3375 | * Undo an allocation. This is used by zio_done() when an I/O fails | |
3376 | * and we want to give back the block we just allocated. | |
3377 | * This handles both normal blocks and gang blocks. | |
3378 | */ | |
3379 | static void | |
3380 | zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) | |
3381 | { | |
b128c09f | 3382 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); |
428870ff | 3383 | ASSERT(zio->io_bp_override == NULL); |
b128c09f BB |
3384 | |
3385 | if (!BP_IS_HOLE(bp)) | |
428870ff | 3386 | metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); |
b128c09f BB |
3387 | |
3388 | if (gn != NULL) { | |
1c27024e | 3389 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
3390 | zio_dva_unallocate(zio, gn->gn_child[g], |
3391 | &gn->gn_gbh->zg_blkptr[g]); | |
3392 | } | |
3393 | } | |
3394 | } | |
3395 | ||
3396 | /* | |
3397 | * Try to allocate an intent log block. Return 0 on success, errno on failure. | |
3398 | */ | |
3399 | int | |
b5256303 TC |
3400 | zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp, |
3401 | uint64_t size, boolean_t *slog) | |
b128c09f | 3402 | { |
428870ff | 3403 | int error = 1; |
4e21fd06 | 3404 | zio_alloc_list_t io_alloc_list; |
b128c09f | 3405 | |
428870ff BB |
3406 | ASSERT(txg > spa_syncing_txg(spa)); |
3407 | ||
4e21fd06 | 3408 | metaslab_trace_init(&io_alloc_list); |
cc99f275 DB |
3409 | |
3410 | /* | |
3411 | * Block pointer fields are useful to metaslabs for stats and debugging. | |
3412 | * Fill in the obvious ones before calling into metaslab_alloc(). | |
3413 | */ | |
3414 | BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); | |
3415 | BP_SET_PSIZE(new_bp, size); | |
3416 | BP_SET_LEVEL(new_bp, 0); | |
3417 | ||
492f64e9 PD |
3418 | /* |
3419 | * When allocating a zil block, we don't have information about | |
3420 | * the final destination of the block except the objset it's part | |
3421 | * of, so we just hash the objset ID to pick the allocator to get | |
3422 | * some parallelism. | |
3423 | */ | |
1b7c1e5c | 3424 | error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1, |
492f64e9 PD |
3425 | txg, NULL, METASLAB_FASTWRITE, &io_alloc_list, NULL, |
3426 | cityhash4(0, 0, 0, os->os_dsl_dataset->ds_object) % | |
3427 | spa->spa_alloc_count); | |
1b7c1e5c GDN |
3428 | if (error == 0) { |
3429 | *slog = TRUE; | |
3430 | } else { | |
428870ff | 3431 | error = metaslab_alloc(spa, spa_normal_class(spa), size, |
4e21fd06 | 3432 | new_bp, 1, txg, NULL, METASLAB_FASTWRITE, |
492f64e9 PD |
3433 | &io_alloc_list, NULL, cityhash4(0, 0, 0, |
3434 | os->os_dsl_dataset->ds_object) % spa->spa_alloc_count); | |
1b7c1e5c GDN |
3435 | if (error == 0) |
3436 | *slog = FALSE; | |
ebf8e3a2 | 3437 | } |
4e21fd06 | 3438 | metaslab_trace_fini(&io_alloc_list); |
b128c09f BB |
3439 | |
3440 | if (error == 0) { | |
3441 | BP_SET_LSIZE(new_bp, size); | |
3442 | BP_SET_PSIZE(new_bp, size); | |
3443 | BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); | |
428870ff BB |
3444 | BP_SET_CHECKSUM(new_bp, |
3445 | spa_version(spa) >= SPA_VERSION_SLIM_ZIL | |
3446 | ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); | |
b128c09f BB |
3447 | BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); |
3448 | BP_SET_LEVEL(new_bp, 0); | |
428870ff | 3449 | BP_SET_DEDUP(new_bp, 0); |
b128c09f | 3450 | BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); |
b5256303 TC |
3451 | |
3452 | /* | |
3453 | * encrypted blocks will require an IV and salt. We generate | |
3454 | * these now since we will not be rewriting the bp at | |
3455 | * rewrite time. | |
3456 | */ | |
3457 | if (os->os_encrypted) { | |
3458 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
3459 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
3460 | ||
3461 | BP_SET_CRYPT(new_bp, B_TRUE); | |
3462 | VERIFY0(spa_crypt_get_salt(spa, | |
3463 | dmu_objset_id(os), salt)); | |
3464 | VERIFY0(zio_crypt_generate_iv(iv)); | |
3465 | ||
3466 | zio_crypt_encode_params_bp(new_bp, salt, iv); | |
3467 | } | |
1ce23dca PS |
3468 | } else { |
3469 | zfs_dbgmsg("%s: zil block allocation failure: " | |
3470 | "size %llu, error %d", spa_name(spa), size, error); | |
b128c09f BB |
3471 | } |
3472 | ||
3473 | return (error); | |
3474 | } | |
3475 | ||
34dc7c2f BB |
3476 | /* |
3477 | * ========================================================================== | |
3478 | * Read and write to physical devices | |
3479 | * ========================================================================== | |
3480 | */ | |
98b25418 | 3481 | |
98b25418 GW |
3482 | /* |
3483 | * Issue an I/O to the underlying vdev. Typically the issue pipeline | |
3484 | * stops after this stage and will resume upon I/O completion. | |
3485 | * However, there are instances where the vdev layer may need to | |
3486 | * continue the pipeline when an I/O was not issued. Since the I/O | |
3487 | * that was sent to the vdev layer might be different than the one | |
3488 | * currently active in the pipeline (see vdev_queue_io()), we explicitly | |
3489 | * force the underlying vdev layers to call either zio_execute() or | |
3490 | * zio_interrupt() to ensure that the pipeline continues with the correct I/O. | |
3491 | */ | |
62840030 | 3492 | static zio_t * |
34dc7c2f BB |
3493 | zio_vdev_io_start(zio_t *zio) |
3494 | { | |
3495 | vdev_t *vd = zio->io_vd; | |
34dc7c2f BB |
3496 | uint64_t align; |
3497 | spa_t *spa = zio->io_spa; | |
3498 | ||
193a37cb TH |
3499 | zio->io_delay = 0; |
3500 | ||
b128c09f BB |
3501 | ASSERT(zio->io_error == 0); |
3502 | ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); | |
34dc7c2f | 3503 | |
b128c09f BB |
3504 | if (vd == NULL) { |
3505 | if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
3506 | spa_config_enter(spa, SCL_ZIO, zio, RW_READER); | |
34dc7c2f | 3507 | |
b128c09f BB |
3508 | /* |
3509 | * The mirror_ops handle multiple DVAs in a single BP. | |
3510 | */ | |
98b25418 | 3511 | vdev_mirror_ops.vdev_op_io_start(zio); |
62840030 | 3512 | return (NULL); |
34dc7c2f BB |
3513 | } |
3514 | ||
3dfb57a3 | 3515 | ASSERT3P(zio->io_logical, !=, zio); |
6cb8e530 PZ |
3516 | if (zio->io_type == ZIO_TYPE_WRITE) { |
3517 | ASSERT(spa->spa_trust_config); | |
3518 | ||
a1d477c2 MA |
3519 | /* |
3520 | * Note: the code can handle other kinds of writes, | |
3521 | * but we don't expect them. | |
3522 | */ | |
6cb8e530 PZ |
3523 | if (zio->io_vd->vdev_removing) { |
3524 | ASSERT(zio->io_flags & | |
3525 | (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL | | |
3526 | ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE)); | |
3527 | } | |
a1d477c2 | 3528 | } |
3dfb57a3 | 3529 | |
b128c09f BB |
3530 | align = 1ULL << vd->vdev_top->vdev_ashift; |
3531 | ||
b02fe35d AR |
3532 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && |
3533 | P2PHASE(zio->io_size, align) != 0) { | |
3534 | /* Transform logical writes to be a full physical block size. */ | |
34dc7c2f | 3535 | uint64_t asize = P2ROUNDUP(zio->io_size, align); |
a6255b7f | 3536 | abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize); |
178e73b3 | 3537 | ASSERT(vd == vd->vdev_top); |
34dc7c2f | 3538 | if (zio->io_type == ZIO_TYPE_WRITE) { |
a6255b7f DQ |
3539 | abd_copy(abuf, zio->io_abd, zio->io_size); |
3540 | abd_zero_off(abuf, zio->io_size, asize - zio->io_size); | |
34dc7c2f | 3541 | } |
b128c09f | 3542 | zio_push_transform(zio, abuf, asize, asize, zio_subblock); |
34dc7c2f BB |
3543 | } |
3544 | ||
b02fe35d AR |
3545 | /* |
3546 | * If this is not a physical io, make sure that it is properly aligned | |
3547 | * before proceeding. | |
3548 | */ | |
3549 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { | |
3550 | ASSERT0(P2PHASE(zio->io_offset, align)); | |
3551 | ASSERT0(P2PHASE(zio->io_size, align)); | |
3552 | } else { | |
3553 | /* | |
3554 | * For physical writes, we allow 512b aligned writes and assume | |
3555 | * the device will perform a read-modify-write as necessary. | |
3556 | */ | |
3557 | ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); | |
3558 | ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); | |
3559 | } | |
3560 | ||
572e2857 | 3561 | VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); |
fb5f0bc8 BB |
3562 | |
3563 | /* | |
3564 | * If this is a repair I/O, and there's no self-healing involved -- | |
3565 | * that is, we're just resilvering what we expect to resilver -- | |
3566 | * then don't do the I/O unless zio's txg is actually in vd's DTL. | |
9e052db4 MA |
3567 | * This prevents spurious resilvering. |
3568 | * | |
3569 | * There are a few ways that we can end up creating these spurious | |
3570 | * resilver i/os: | |
3571 | * | |
3572 | * 1. A resilver i/o will be issued if any DVA in the BP has a | |
3573 | * dirty DTL. The mirror code will issue resilver writes to | |
3574 | * each DVA, including the one(s) that are not on vdevs with dirty | |
3575 | * DTLs. | |
3576 | * | |
3577 | * 2. With nested replication, which happens when we have a | |
3578 | * "replacing" or "spare" vdev that's a child of a mirror or raidz. | |
3579 | * For example, given mirror(replacing(A+B), C), it's likely that | |
3580 | * only A is out of date (it's the new device). In this case, we'll | |
3581 | * read from C, then use the data to resilver A+B -- but we don't | |
3582 | * actually want to resilver B, just A. The top-level mirror has no | |
3583 | * way to know this, so instead we just discard unnecessary repairs | |
3584 | * as we work our way down the vdev tree. | |
3585 | * | |
3586 | * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc. | |
3587 | * The same logic applies to any form of nested replication: ditto | |
3588 | * + mirror, RAID-Z + replacing, etc. | |
3589 | * | |
3590 | * However, indirect vdevs point off to other vdevs which may have | |
3591 | * DTL's, so we never bypass them. The child i/os on concrete vdevs | |
3592 | * will be properly bypassed instead. | |
fb5f0bc8 BB |
3593 | */ |
3594 | if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && | |
3595 | !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && | |
3596 | zio->io_txg != 0 && /* not a delegated i/o */ | |
9e052db4 | 3597 | vd->vdev_ops != &vdev_indirect_ops && |
fb5f0bc8 BB |
3598 | !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { |
3599 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
fb5f0bc8 | 3600 | zio_vdev_io_bypass(zio); |
62840030 | 3601 | return (zio); |
fb5f0bc8 | 3602 | } |
34dc7c2f | 3603 | |
1b939560 BB |
3604 | if (vd->vdev_ops->vdev_op_leaf && (zio->io_type == ZIO_TYPE_READ || |
3605 | zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM)) { | |
b128c09f | 3606 | |
b0bc7a84 | 3607 | if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio)) |
62840030 | 3608 | return (zio); |
b128c09f BB |
3609 | |
3610 | if ((zio = vdev_queue_io(zio)) == NULL) | |
62840030 | 3611 | return (NULL); |
b128c09f BB |
3612 | |
3613 | if (!vdev_accessible(vd, zio)) { | |
2e528b49 | 3614 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f | 3615 | zio_interrupt(zio); |
62840030 | 3616 | return (NULL); |
b128c09f | 3617 | } |
67103816 | 3618 | zio->io_delay = gethrtime(); |
b128c09f BB |
3619 | } |
3620 | ||
98b25418 | 3621 | vd->vdev_ops->vdev_op_io_start(zio); |
62840030 | 3622 | return (NULL); |
34dc7c2f BB |
3623 | } |
3624 | ||
62840030 | 3625 | static zio_t * |
34dc7c2f BB |
3626 | zio_vdev_io_done(zio_t *zio) |
3627 | { | |
b128c09f BB |
3628 | vdev_t *vd = zio->io_vd; |
3629 | vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; | |
3630 | boolean_t unexpected_error = B_FALSE; | |
34dc7c2f | 3631 | |
ddc751d5 | 3632 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { |
62840030 | 3633 | return (NULL); |
ddc751d5 | 3634 | } |
34dc7c2f | 3635 | |
1b939560 BB |
3636 | ASSERT(zio->io_type == ZIO_TYPE_READ || |
3637 | zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM); | |
b128c09f | 3638 | |
193a37cb TH |
3639 | if (zio->io_delay) |
3640 | zio->io_delay = gethrtime() - zio->io_delay; | |
3641 | ||
b128c09f BB |
3642 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { |
3643 | ||
3644 | vdev_queue_io_done(zio); | |
3645 | ||
3646 | if (zio->io_type == ZIO_TYPE_WRITE) | |
3647 | vdev_cache_write(zio); | |
3648 | ||
3649 | if (zio_injection_enabled && zio->io_error == 0) | |
d977122d DB |
3650 | zio->io_error = zio_handle_device_injections(vd, zio, |
3651 | EIO, EILSEQ); | |
b128c09f BB |
3652 | |
3653 | if (zio_injection_enabled && zio->io_error == 0) | |
3654 | zio->io_error = zio_handle_label_injection(zio, EIO); | |
3655 | ||
1b939560 | 3656 | if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) { |
b128c09f | 3657 | if (!vdev_accessible(vd, zio)) { |
2e528b49 | 3658 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
3659 | } else { |
3660 | unexpected_error = B_TRUE; | |
3661 | } | |
3662 | } | |
3663 | } | |
3664 | ||
3665 | ops->vdev_op_io_done(zio); | |
34dc7c2f | 3666 | |
f43615d0 | 3667 | if (unexpected_error) |
d164b209 | 3668 | VERIFY(vdev_probe(vd, zio) == NULL); |
34dc7c2f | 3669 | |
62840030 | 3670 | return (zio); |
34dc7c2f BB |
3671 | } |
3672 | ||
a8b2e306 TC |
3673 | /* |
3674 | * This function is used to change the priority of an existing zio that is | |
3675 | * currently in-flight. This is used by the arc to upgrade priority in the | |
3676 | * event that a demand read is made for a block that is currently queued | |
3677 | * as a scrub or async read IO. Otherwise, the high priority read request | |
3678 | * would end up having to wait for the lower priority IO. | |
3679 | */ | |
3680 | void | |
3681 | zio_change_priority(zio_t *pio, zio_priority_t priority) | |
3682 | { | |
3683 | zio_t *cio, *cio_next; | |
3684 | zio_link_t *zl = NULL; | |
3685 | ||
3686 | ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
3687 | ||
3688 | if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) { | |
3689 | vdev_queue_change_io_priority(pio, priority); | |
3690 | } else { | |
3691 | pio->io_priority = priority; | |
3692 | } | |
3693 | ||
3694 | mutex_enter(&pio->io_lock); | |
3695 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { | |
3696 | cio_next = zio_walk_children(pio, &zl); | |
3697 | zio_change_priority(cio, priority); | |
3698 | } | |
3699 | mutex_exit(&pio->io_lock); | |
3700 | } | |
3701 | ||
428870ff BB |
3702 | /* |
3703 | * For non-raidz ZIOs, we can just copy aside the bad data read from the | |
3704 | * disk, and use that to finish the checksum ereport later. | |
3705 | */ | |
3706 | static void | |
3707 | zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, | |
84c07ada | 3708 | const abd_t *good_buf) |
428870ff BB |
3709 | { |
3710 | /* no processing needed */ | |
3711 | zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); | |
3712 | } | |
3713 | ||
3714 | /*ARGSUSED*/ | |
3715 | void | |
3716 | zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) | |
3717 | { | |
84c07ada | 3718 | void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size); |
428870ff | 3719 | |
84c07ada | 3720 | abd_copy(abd, zio->io_abd, zio->io_size); |
428870ff BB |
3721 | |
3722 | zcr->zcr_cbinfo = zio->io_size; | |
84c07ada | 3723 | zcr->zcr_cbdata = abd; |
428870ff | 3724 | zcr->zcr_finish = zio_vsd_default_cksum_finish; |
84c07ada | 3725 | zcr->zcr_free = zio_abd_free; |
428870ff BB |
3726 | } |
3727 | ||
62840030 | 3728 | static zio_t * |
34dc7c2f BB |
3729 | zio_vdev_io_assess(zio_t *zio) |
3730 | { | |
3731 | vdev_t *vd = zio->io_vd; | |
b128c09f | 3732 | |
ddc751d5 | 3733 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { |
62840030 | 3734 | return (NULL); |
ddc751d5 | 3735 | } |
b128c09f BB |
3736 | |
3737 | if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
3738 | spa_config_exit(zio->io_spa, SCL_ZIO, zio); | |
3739 | ||
3740 | if (zio->io_vsd != NULL) { | |
428870ff | 3741 | zio->io_vsd_ops->vsd_free(zio); |
b128c09f | 3742 | zio->io_vsd = NULL; |
34dc7c2f BB |
3743 | } |
3744 | ||
b128c09f | 3745 | if (zio_injection_enabled && zio->io_error == 0) |
34dc7c2f BB |
3746 | zio->io_error = zio_handle_fault_injection(zio, EIO); |
3747 | ||
3748 | /* | |
3749 | * If the I/O failed, determine whether we should attempt to retry it. | |
428870ff BB |
3750 | * |
3751 | * On retry, we cut in line in the issue queue, since we don't want | |
3752 | * compression/checksumming/etc. work to prevent our (cheap) IO reissue. | |
34dc7c2f | 3753 | */ |
b128c09f BB |
3754 | if (zio->io_error && vd == NULL && |
3755 | !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { | |
3756 | ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ | |
3757 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ | |
34dc7c2f | 3758 | zio->io_error = 0; |
b128c09f BB |
3759 | zio->io_flags |= ZIO_FLAG_IO_RETRY | |
3760 | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; | |
428870ff BB |
3761 | zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; |
3762 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, | |
3763 | zio_requeue_io_start_cut_in_line); | |
62840030 | 3764 | return (NULL); |
34dc7c2f BB |
3765 | } |
3766 | ||
b128c09f BB |
3767 | /* |
3768 | * If we got an error on a leaf device, convert it to ENXIO | |
3769 | * if the device is not accessible at all. | |
3770 | */ | |
3771 | if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && | |
3772 | !vdev_accessible(vd, zio)) | |
2e528b49 | 3773 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
3774 | |
3775 | /* | |
3776 | * If we can't write to an interior vdev (mirror or RAID-Z), | |
3777 | * set vdev_cant_write so that we stop trying to allocate from it. | |
3778 | */ | |
3779 | if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && | |
13fe0198 | 3780 | vd != NULL && !vd->vdev_ops->vdev_op_leaf) { |
b128c09f | 3781 | vd->vdev_cant_write = B_TRUE; |
13fe0198 | 3782 | } |
b128c09f | 3783 | |
298ec40b GM |
3784 | /* |
3785 | * If a cache flush returns ENOTSUP or ENOTTY, we know that no future | |
1b939560 BB |
3786 | * attempts will ever succeed. In this case we set a persistent |
3787 | * boolean flag so that we don't bother with it in the future. | |
298ec40b GM |
3788 | */ |
3789 | if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) && | |
3790 | zio->io_type == ZIO_TYPE_IOCTL && | |
3791 | zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL) | |
3792 | vd->vdev_nowritecache = B_TRUE; | |
3793 | ||
b128c09f BB |
3794 | if (zio->io_error) |
3795 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
3796 | ||
e8b96c60 MA |
3797 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf && |
3798 | zio->io_physdone != NULL) { | |
3799 | ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); | |
3800 | ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); | |
3801 | zio->io_physdone(zio->io_logical); | |
3802 | } | |
3803 | ||
62840030 | 3804 | return (zio); |
34dc7c2f BB |
3805 | } |
3806 | ||
3807 | void | |
3808 | zio_vdev_io_reissue(zio_t *zio) | |
3809 | { | |
3810 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
3811 | ASSERT(zio->io_error == 0); | |
3812 | ||
428870ff | 3813 | zio->io_stage >>= 1; |
34dc7c2f BB |
3814 | } |
3815 | ||
3816 | void | |
3817 | zio_vdev_io_redone(zio_t *zio) | |
3818 | { | |
3819 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); | |
3820 | ||
428870ff | 3821 | zio->io_stage >>= 1; |
34dc7c2f BB |
3822 | } |
3823 | ||
3824 | void | |
3825 | zio_vdev_io_bypass(zio_t *zio) | |
3826 | { | |
3827 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
3828 | ASSERT(zio->io_error == 0); | |
3829 | ||
3830 | zio->io_flags |= ZIO_FLAG_IO_BYPASS; | |
428870ff | 3831 | zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; |
34dc7c2f BB |
3832 | } |
3833 | ||
b5256303 TC |
3834 | /* |
3835 | * ========================================================================== | |
3836 | * Encrypt and store encryption parameters | |
3837 | * ========================================================================== | |
3838 | */ | |
3839 | ||
3840 | ||
3841 | /* | |
3842 | * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for | |
3843 | * managing the storage of encryption parameters and passing them to the | |
3844 | * lower-level encryption functions. | |
3845 | */ | |
62840030 | 3846 | static zio_t * |
b5256303 TC |
3847 | zio_encrypt(zio_t *zio) |
3848 | { | |
3849 | zio_prop_t *zp = &zio->io_prop; | |
3850 | spa_t *spa = zio->io_spa; | |
3851 | blkptr_t *bp = zio->io_bp; | |
3852 | uint64_t psize = BP_GET_PSIZE(bp); | |
ae76f45c | 3853 | uint64_t dsobj = zio->io_bookmark.zb_objset; |
b5256303 TC |
3854 | dmu_object_type_t ot = BP_GET_TYPE(bp); |
3855 | void *enc_buf = NULL; | |
3856 | abd_t *eabd = NULL; | |
3857 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
3858 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
3859 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
3860 | boolean_t no_crypt = B_FALSE; | |
3861 | ||
3862 | /* the root zio already encrypted the data */ | |
3863 | if (zio->io_child_type == ZIO_CHILD_GANG) | |
62840030 | 3864 | return (zio); |
b5256303 TC |
3865 | |
3866 | /* only ZIL blocks are re-encrypted on rewrite */ | |
3867 | if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG) | |
62840030 | 3868 | return (zio); |
b5256303 TC |
3869 | |
3870 | if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) { | |
3871 | BP_SET_CRYPT(bp, B_FALSE); | |
62840030 | 3872 | return (zio); |
b5256303 TC |
3873 | } |
3874 | ||
3875 | /* if we are doing raw encryption set the provided encryption params */ | |
3876 | if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) { | |
ae76f45c | 3877 | ASSERT0(BP_GET_LEVEL(bp)); |
b5256303 TC |
3878 | BP_SET_CRYPT(bp, B_TRUE); |
3879 | BP_SET_BYTEORDER(bp, zp->zp_byteorder); | |
3880 | if (ot != DMU_OT_OBJSET) | |
3881 | zio_crypt_encode_mac_bp(bp, zp->zp_mac); | |
ae76f45c TC |
3882 | |
3883 | /* dnode blocks must be written out in the provided byteorder */ | |
3884 | if (zp->zp_byteorder != ZFS_HOST_BYTEORDER && | |
3885 | ot == DMU_OT_DNODE) { | |
3886 | void *bswap_buf = zio_buf_alloc(psize); | |
3887 | abd_t *babd = abd_get_from_buf(bswap_buf, psize); | |
3888 | ||
3889 | ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); | |
3890 | abd_copy_to_buf(bswap_buf, zio->io_abd, psize); | |
3891 | dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf, | |
3892 | psize); | |
3893 | ||
3894 | abd_take_ownership_of_buf(babd, B_TRUE); | |
3895 | zio_push_transform(zio, babd, psize, psize, NULL); | |
3896 | } | |
3897 | ||
b5256303 TC |
3898 | if (DMU_OT_IS_ENCRYPTED(ot)) |
3899 | zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv); | |
62840030 | 3900 | return (zio); |
b5256303 TC |
3901 | } |
3902 | ||
3903 | /* indirect blocks only maintain a cksum of the lower level MACs */ | |
3904 | if (BP_GET_LEVEL(bp) > 0) { | |
3905 | BP_SET_CRYPT(bp, B_TRUE); | |
3906 | VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE, | |
3907 | zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp), | |
3908 | mac)); | |
3909 | zio_crypt_encode_mac_bp(bp, mac); | |
62840030 | 3910 | return (zio); |
b5256303 TC |
3911 | } |
3912 | ||
3913 | /* | |
3914 | * Objset blocks are a special case since they have 2 256-bit MACs | |
3915 | * embedded within them. | |
3916 | */ | |
3917 | if (ot == DMU_OT_OBJSET) { | |
3918 | ASSERT0(DMU_OT_IS_ENCRYPTED(ot)); | |
3919 | ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); | |
3920 | BP_SET_CRYPT(bp, B_TRUE); | |
ae76f45c TC |
3921 | VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj, |
3922 | zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp))); | |
62840030 | 3923 | return (zio); |
b5256303 TC |
3924 | } |
3925 | ||
3926 | /* unencrypted object types are only authenticated with a MAC */ | |
3927 | if (!DMU_OT_IS_ENCRYPTED(ot)) { | |
3928 | BP_SET_CRYPT(bp, B_TRUE); | |
ae76f45c TC |
3929 | VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj, |
3930 | zio->io_abd, psize, mac)); | |
b5256303 | 3931 | zio_crypt_encode_mac_bp(bp, mac); |
62840030 | 3932 | return (zio); |
b5256303 TC |
3933 | } |
3934 | ||
3935 | /* | |
3936 | * Later passes of sync-to-convergence may decide to rewrite data | |
3937 | * in place to avoid more disk reallocations. This presents a problem | |
d611989f | 3938 | * for encryption because this constitutes rewriting the new data with |
b5256303 TC |
3939 | * the same encryption key and IV. However, this only applies to blocks |
3940 | * in the MOS (particularly the spacemaps) and we do not encrypt the | |
3941 | * MOS. We assert that the zio is allocating or an intent log write | |
3942 | * to enforce this. | |
3943 | */ | |
3944 | ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG); | |
3945 | ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG); | |
3946 | ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION)); | |
3947 | ASSERT3U(psize, !=, 0); | |
3948 | ||
3949 | enc_buf = zio_buf_alloc(psize); | |
3950 | eabd = abd_get_from_buf(enc_buf, psize); | |
3951 | abd_take_ownership_of_buf(eabd, B_TRUE); | |
3952 | ||
3953 | /* | |
3954 | * For an explanation of what encryption parameters are stored | |
3955 | * where, see the block comment in zio_crypt.c. | |
3956 | */ | |
3957 | if (ot == DMU_OT_INTENT_LOG) { | |
3958 | zio_crypt_decode_params_bp(bp, salt, iv); | |
3959 | } else { | |
3960 | BP_SET_CRYPT(bp, B_TRUE); | |
3961 | } | |
3962 | ||
3963 | /* Perform the encryption. This should not fail */ | |
be9a5c35 TC |
3964 | VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark, |
3965 | BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), | |
3966 | salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt)); | |
b5256303 TC |
3967 | |
3968 | /* encode encryption metadata into the bp */ | |
3969 | if (ot == DMU_OT_INTENT_LOG) { | |
3970 | /* | |
3971 | * ZIL blocks store the MAC in the embedded checksum, so the | |
3972 | * transform must always be applied. | |
3973 | */ | |
3974 | zio_crypt_encode_mac_zil(enc_buf, mac); | |
3975 | zio_push_transform(zio, eabd, psize, psize, NULL); | |
3976 | } else { | |
3977 | BP_SET_CRYPT(bp, B_TRUE); | |
3978 | zio_crypt_encode_params_bp(bp, salt, iv); | |
3979 | zio_crypt_encode_mac_bp(bp, mac); | |
3980 | ||
3981 | if (no_crypt) { | |
3982 | ASSERT3U(ot, ==, DMU_OT_DNODE); | |
3983 | abd_free(eabd); | |
3984 | } else { | |
3985 | zio_push_transform(zio, eabd, psize, psize, NULL); | |
3986 | } | |
3987 | } | |
3988 | ||
62840030 | 3989 | return (zio); |
b5256303 TC |
3990 | } |
3991 | ||
34dc7c2f BB |
3992 | /* |
3993 | * ========================================================================== | |
3994 | * Generate and verify checksums | |
3995 | * ========================================================================== | |
3996 | */ | |
62840030 | 3997 | static zio_t * |
34dc7c2f BB |
3998 | zio_checksum_generate(zio_t *zio) |
3999 | { | |
34dc7c2f | 4000 | blkptr_t *bp = zio->io_bp; |
b128c09f | 4001 | enum zio_checksum checksum; |
34dc7c2f | 4002 | |
b128c09f BB |
4003 | if (bp == NULL) { |
4004 | /* | |
4005 | * This is zio_write_phys(). | |
4006 | * We're either generating a label checksum, or none at all. | |
4007 | */ | |
4008 | checksum = zio->io_prop.zp_checksum; | |
34dc7c2f | 4009 | |
b128c09f | 4010 | if (checksum == ZIO_CHECKSUM_OFF) |
62840030 | 4011 | return (zio); |
b128c09f BB |
4012 | |
4013 | ASSERT(checksum == ZIO_CHECKSUM_LABEL); | |
4014 | } else { | |
4015 | if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { | |
4016 | ASSERT(!IO_IS_ALLOCATING(zio)); | |
4017 | checksum = ZIO_CHECKSUM_GANG_HEADER; | |
4018 | } else { | |
4019 | checksum = BP_GET_CHECKSUM(bp); | |
4020 | } | |
4021 | } | |
34dc7c2f | 4022 | |
a6255b7f | 4023 | zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size); |
34dc7c2f | 4024 | |
62840030 | 4025 | return (zio); |
34dc7c2f BB |
4026 | } |
4027 | ||
62840030 | 4028 | static zio_t * |
b128c09f | 4029 | zio_checksum_verify(zio_t *zio) |
34dc7c2f | 4030 | { |
428870ff | 4031 | zio_bad_cksum_t info; |
b128c09f BB |
4032 | blkptr_t *bp = zio->io_bp; |
4033 | int error; | |
34dc7c2f | 4034 | |
428870ff BB |
4035 | ASSERT(zio->io_vd != NULL); |
4036 | ||
b128c09f BB |
4037 | if (bp == NULL) { |
4038 | /* | |
4039 | * This is zio_read_phys(). | |
4040 | * We're either verifying a label checksum, or nothing at all. | |
4041 | */ | |
4042 | if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) | |
62840030 | 4043 | return (zio); |
34dc7c2f | 4044 | |
b128c09f BB |
4045 | ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); |
4046 | } | |
34dc7c2f | 4047 | |
428870ff | 4048 | if ((error = zio_checksum_error(zio, &info)) != 0) { |
b128c09f | 4049 | zio->io_error = error; |
7a3066ff MA |
4050 | if (error == ECKSUM && |
4051 | !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { | |
2bbec1c9 TH |
4052 | mutex_enter(&zio->io_vd->vdev_stat_lock); |
4053 | zio->io_vd->vdev_stat.vs_checksum_errors++; | |
4054 | mutex_exit(&zio->io_vd->vdev_stat_lock); | |
4055 | ||
428870ff | 4056 | zfs_ereport_start_checksum(zio->io_spa, |
b5256303 TC |
4057 | zio->io_vd, &zio->io_bookmark, zio, |
4058 | zio->io_offset, zio->io_size, NULL, &info); | |
b128c09f | 4059 | } |
34dc7c2f BB |
4060 | } |
4061 | ||
62840030 | 4062 | return (zio); |
34dc7c2f BB |
4063 | } |
4064 | ||
4065 | /* | |
4066 | * Called by RAID-Z to ensure we don't compute the checksum twice. | |
4067 | */ | |
4068 | void | |
4069 | zio_checksum_verified(zio_t *zio) | |
4070 | { | |
428870ff | 4071 | zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; |
34dc7c2f BB |
4072 | } |
4073 | ||
4074 | /* | |
b128c09f BB |
4075 | * ========================================================================== |
4076 | * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. | |
9b67f605 | 4077 | * An error of 0 indicates success. ENXIO indicates whole-device failure, |
d611989f | 4078 | * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO |
b128c09f BB |
4079 | * indicate errors that are specific to one I/O, and most likely permanent. |
4080 | * Any other error is presumed to be worse because we weren't expecting it. | |
4081 | * ========================================================================== | |
34dc7c2f | 4082 | */ |
b128c09f BB |
4083 | int |
4084 | zio_worst_error(int e1, int e2) | |
34dc7c2f | 4085 | { |
b128c09f BB |
4086 | static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; |
4087 | int r1, r2; | |
4088 | ||
4089 | for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) | |
4090 | if (e1 == zio_error_rank[r1]) | |
4091 | break; | |
34dc7c2f | 4092 | |
b128c09f BB |
4093 | for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) |
4094 | if (e2 == zio_error_rank[r2]) | |
4095 | break; | |
4096 | ||
4097 | return (r1 > r2 ? e1 : e2); | |
34dc7c2f BB |
4098 | } |
4099 | ||
4100 | /* | |
4101 | * ========================================================================== | |
b128c09f | 4102 | * I/O completion |
34dc7c2f BB |
4103 | * ========================================================================== |
4104 | */ | |
62840030 | 4105 | static zio_t * |
b128c09f | 4106 | zio_ready(zio_t *zio) |
34dc7c2f | 4107 | { |
b128c09f | 4108 | blkptr_t *bp = zio->io_bp; |
d164b209 | 4109 | zio_t *pio, *pio_next; |
3dfb57a3 | 4110 | zio_link_t *zl = NULL; |
34dc7c2f | 4111 | |
ddc751d5 GW |
4112 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT, |
4113 | ZIO_WAIT_READY)) { | |
62840030 | 4114 | return (NULL); |
ddc751d5 | 4115 | } |
34dc7c2f | 4116 | |
9babb374 | 4117 | if (zio->io_ready) { |
b128c09f | 4118 | ASSERT(IO_IS_ALLOCATING(zio)); |
03c6040b GW |
4119 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || |
4120 | (zio->io_flags & ZIO_FLAG_NOPWRITE)); | |
b128c09f | 4121 | ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); |
34dc7c2f | 4122 | |
b128c09f BB |
4123 | zio->io_ready(zio); |
4124 | } | |
34dc7c2f | 4125 | |
b128c09f BB |
4126 | if (bp != NULL && bp != &zio->io_bp_copy) |
4127 | zio->io_bp_copy = *bp; | |
34dc7c2f | 4128 | |
3dfb57a3 | 4129 | if (zio->io_error != 0) { |
b128c09f | 4130 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f | 4131 | |
3dfb57a3 DB |
4132 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
4133 | ASSERT(IO_IS_ALLOCATING(zio)); | |
4134 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
cc99f275 DB |
4135 | ASSERT(zio->io_metaslab_class != NULL); |
4136 | ||
3dfb57a3 DB |
4137 | /* |
4138 | * We were unable to allocate anything, unreserve and | |
4139 | * issue the next I/O to allocate. | |
4140 | */ | |
4141 | metaslab_class_throttle_unreserve( | |
cc99f275 DB |
4142 | zio->io_metaslab_class, zio->io_prop.zp_copies, |
4143 | zio->io_allocator, zio); | |
492f64e9 | 4144 | zio_allocate_dispatch(zio->io_spa, zio->io_allocator); |
3dfb57a3 DB |
4145 | } |
4146 | } | |
4147 | ||
d164b209 BB |
4148 | mutex_enter(&zio->io_lock); |
4149 | zio->io_state[ZIO_WAIT_READY] = 1; | |
3dfb57a3 | 4150 | pio = zio_walk_parents(zio, &zl); |
d164b209 BB |
4151 | mutex_exit(&zio->io_lock); |
4152 | ||
4153 | /* | |
4154 | * As we notify zio's parents, new parents could be added. | |
4155 | * New parents go to the head of zio's io_parent_list, however, | |
4156 | * so we will (correctly) not notify them. The remainder of zio's | |
4157 | * io_parent_list, from 'pio_next' onward, cannot change because | |
4158 | * all parents must wait for us to be done before they can be done. | |
4159 | */ | |
4160 | for (; pio != NULL; pio = pio_next) { | |
3dfb57a3 | 4161 | pio_next = zio_walk_parents(zio, &zl); |
62840030 | 4162 | zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL); |
d164b209 | 4163 | } |
34dc7c2f | 4164 | |
428870ff BB |
4165 | if (zio->io_flags & ZIO_FLAG_NODATA) { |
4166 | if (BP_IS_GANG(bp)) { | |
4167 | zio->io_flags &= ~ZIO_FLAG_NODATA; | |
4168 | } else { | |
a6255b7f | 4169 | ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE); |
428870ff BB |
4170 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; |
4171 | } | |
4172 | } | |
4173 | ||
4174 | if (zio_injection_enabled && | |
4175 | zio->io_spa->spa_syncing_txg == zio->io_txg) | |
4176 | zio_handle_ignored_writes(zio); | |
4177 | ||
62840030 | 4178 | return (zio); |
34dc7c2f BB |
4179 | } |
4180 | ||
3dfb57a3 DB |
4181 | /* |
4182 | * Update the allocation throttle accounting. | |
4183 | */ | |
4184 | static void | |
4185 | zio_dva_throttle_done(zio_t *zio) | |
4186 | { | |
1c27024e | 4187 | ASSERTV(zio_t *lio = zio->io_logical); |
3dfb57a3 DB |
4188 | zio_t *pio = zio_unique_parent(zio); |
4189 | vdev_t *vd = zio->io_vd; | |
4190 | int flags = METASLAB_ASYNC_ALLOC; | |
4191 | ||
4192 | ASSERT3P(zio->io_bp, !=, NULL); | |
4193 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); | |
4194 | ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE); | |
4195 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); | |
4196 | ASSERT(vd != NULL); | |
4197 | ASSERT3P(vd, ==, vd->vdev_top); | |
21df134f SB |
4198 | ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY)); |
4199 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); | |
3dfb57a3 DB |
4200 | ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING); |
4201 | ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
4202 | ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA)); | |
4203 | ||
4204 | /* | |
4205 | * Parents of gang children can have two flavors -- ones that | |
4206 | * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set) | |
4207 | * and ones that allocated the constituent blocks. The allocation | |
4208 | * throttle needs to know the allocating parent zio so we must find | |
4209 | * it here. | |
4210 | */ | |
4211 | if (pio->io_child_type == ZIO_CHILD_GANG) { | |
4212 | /* | |
4213 | * If our parent is a rewrite gang child then our grandparent | |
4214 | * would have been the one that performed the allocation. | |
4215 | */ | |
4216 | if (pio->io_flags & ZIO_FLAG_IO_REWRITE) | |
4217 | pio = zio_unique_parent(pio); | |
4218 | flags |= METASLAB_GANG_CHILD; | |
4219 | } | |
4220 | ||
4221 | ASSERT(IO_IS_ALLOCATING(pio)); | |
4222 | ASSERT3P(zio, !=, zio->io_logical); | |
4223 | ASSERT(zio->io_logical != NULL); | |
4224 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); | |
4225 | ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE); | |
cc99f275 | 4226 | ASSERT(zio->io_metaslab_class != NULL); |
3dfb57a3 DB |
4227 | |
4228 | mutex_enter(&pio->io_lock); | |
492f64e9 PD |
4229 | metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags, |
4230 | pio->io_allocator, B_TRUE); | |
3dfb57a3 DB |
4231 | mutex_exit(&pio->io_lock); |
4232 | ||
cc99f275 DB |
4233 | metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1, |
4234 | pio->io_allocator, pio); | |
3dfb57a3 DB |
4235 | |
4236 | /* | |
4237 | * Call into the pipeline to see if there is more work that | |
4238 | * needs to be done. If there is work to be done it will be | |
4239 | * dispatched to another taskq thread. | |
4240 | */ | |
492f64e9 | 4241 | zio_allocate_dispatch(zio->io_spa, pio->io_allocator); |
3dfb57a3 DB |
4242 | } |
4243 | ||
62840030 | 4244 | static zio_t * |
b128c09f | 4245 | zio_done(zio_t *zio) |
34dc7c2f | 4246 | { |
3dfb57a3 DB |
4247 | /* |
4248 | * Always attempt to keep stack usage minimal here since | |
d611989f | 4249 | * we can be called recursively up to 19 levels deep. |
3dfb57a3 | 4250 | */ |
84c07ada | 4251 | const uint64_t psize = zio->io_size; |
d164b209 | 4252 | zio_t *pio, *pio_next; |
3dfb57a3 | 4253 | zio_link_t *zl = NULL; |
34dc7c2f | 4254 | |
b128c09f | 4255 | /* |
9babb374 | 4256 | * If our children haven't all completed, |
b128c09f BB |
4257 | * wait for them and then repeat this pipeline stage. |
4258 | */ | |
ddc751d5 | 4259 | if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) { |
62840030 | 4260 | return (NULL); |
ddc751d5 | 4261 | } |
34dc7c2f | 4262 | |
3dfb57a3 DB |
4263 | /* |
4264 | * If the allocation throttle is enabled, then update the accounting. | |
4265 | * We only track child I/Os that are part of an allocating async | |
4266 | * write. We must do this since the allocation is performed | |
4267 | * by the logical I/O but the actual write is done by child I/Os. | |
4268 | */ | |
4269 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING && | |
4270 | zio->io_child_type == ZIO_CHILD_VDEV) { | |
cc99f275 DB |
4271 | ASSERT(zio->io_metaslab_class != NULL); |
4272 | ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled); | |
3dfb57a3 DB |
4273 | zio_dva_throttle_done(zio); |
4274 | } | |
4275 | ||
4276 | /* | |
4277 | * If the allocation throttle is enabled, verify that | |
4278 | * we have decremented the refcounts for every I/O that was throttled. | |
4279 | */ | |
4280 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { | |
4281 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
4282 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
4283 | ASSERT(zio->io_bp != NULL); | |
cc99f275 | 4284 | |
492f64e9 PD |
4285 | metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio, |
4286 | zio->io_allocator); | |
424fd7c3 | 4287 | VERIFY(zfs_refcount_not_held( |
cc99f275 | 4288 | &zio->io_metaslab_class->mc_alloc_slots[zio->io_allocator], |
492f64e9 | 4289 | zio)); |
3dfb57a3 DB |
4290 | } |
4291 | ||
4292 | ||
1c27024e DB |
4293 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) |
4294 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) | |
b128c09f BB |
4295 | ASSERT(zio->io_children[c][w] == 0); |
4296 | ||
9b67f605 | 4297 | if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) { |
c776b317 BB |
4298 | ASSERT(zio->io_bp->blk_pad[0] == 0); |
4299 | ASSERT(zio->io_bp->blk_pad[1] == 0); | |
d1d7e268 MK |
4300 | ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, |
4301 | sizeof (blkptr_t)) == 0 || | |
c776b317 BB |
4302 | (zio->io_bp == zio_unique_parent(zio)->io_bp)); |
4303 | if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) && | |
428870ff | 4304 | zio->io_bp_override == NULL && |
b128c09f | 4305 | !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { |
d1d7e268 MK |
4306 | ASSERT3U(zio->io_prop.zp_copies, <=, |
4307 | BP_GET_NDVAS(zio->io_bp)); | |
c776b317 | 4308 | ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 || |
d1d7e268 MK |
4309 | (BP_COUNT_GANG(zio->io_bp) == |
4310 | BP_GET_NDVAS(zio->io_bp))); | |
b128c09f | 4311 | } |
03c6040b GW |
4312 | if (zio->io_flags & ZIO_FLAG_NOPWRITE) |
4313 | VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig)); | |
b128c09f BB |
4314 | } |
4315 | ||
4316 | /* | |
428870ff | 4317 | * If there were child vdev/gang/ddt errors, they apply to us now. |
b128c09f BB |
4318 | */ |
4319 | zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); | |
4320 | zio_inherit_child_errors(zio, ZIO_CHILD_GANG); | |
428870ff BB |
4321 | zio_inherit_child_errors(zio, ZIO_CHILD_DDT); |
4322 | ||
4323 | /* | |
4324 | * If the I/O on the transformed data was successful, generate any | |
4325 | * checksum reports now while we still have the transformed data. | |
4326 | */ | |
4327 | if (zio->io_error == 0) { | |
4328 | while (zio->io_cksum_report != NULL) { | |
4329 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
4330 | uint64_t align = zcr->zcr_align; | |
a6255b7f | 4331 | uint64_t asize = P2ROUNDUP(psize, align); |
a6255b7f DQ |
4332 | abd_t *adata = zio->io_abd; |
4333 | ||
4334 | if (asize != psize) { | |
84c07ada | 4335 | adata = abd_alloc(asize, B_TRUE); |
a6255b7f DQ |
4336 | abd_copy(adata, zio->io_abd, psize); |
4337 | abd_zero_off(adata, psize, asize - psize); | |
428870ff BB |
4338 | } |
4339 | ||
4340 | zio->io_cksum_report = zcr->zcr_next; | |
4341 | zcr->zcr_next = NULL; | |
84c07ada | 4342 | zcr->zcr_finish(zcr, adata); |
428870ff BB |
4343 | zfs_ereport_free_checksum(zcr); |
4344 | ||
a6255b7f DQ |
4345 | if (asize != psize) |
4346 | abd_free(adata); | |
428870ff BB |
4347 | } |
4348 | } | |
b128c09f BB |
4349 | |
4350 | zio_pop_transforms(zio); /* note: may set zio->io_error */ | |
4351 | ||
a6255b7f | 4352 | vdev_stat_update(zio, psize); |
b128c09f | 4353 | |
a69052be | 4354 | /* |
cc92e9d0 | 4355 | * If this I/O is attached to a particular vdev is slow, exceeding |
72f53c56 MJ |
4356 | * 30 seconds to complete, post an error described the I/O delay. |
4357 | * We ignore these errors if the device is currently unavailable. | |
a69052be | 4358 | */ |
ad796b8a TH |
4359 | if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) { |
4360 | if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) { | |
4361 | /* | |
4362 | * We want to only increment our slow IO counters if | |
4363 | * the IO is valid (i.e. not if the drive is removed). | |
4364 | * | |
4365 | * zfs_ereport_post() will also do these checks, but | |
4366 | * it can also ratelimit and have other failures, so we | |
4367 | * need to increment the slow_io counters independent | |
4368 | * of it. | |
4369 | */ | |
4370 | if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY, | |
4371 | zio->io_spa, zio->io_vd, zio)) { | |
4372 | mutex_enter(&zio->io_vd->vdev_stat_lock); | |
4373 | zio->io_vd->vdev_stat.vs_slow_ios++; | |
4374 | mutex_exit(&zio->io_vd->vdev_stat_lock); | |
4375 | ||
4376 | zfs_ereport_post(FM_EREPORT_ZFS_DELAY, | |
4377 | zio->io_spa, zio->io_vd, &zio->io_bookmark, | |
4378 | zio, 0, 0); | |
4379 | } | |
4380 | } | |
72f53c56 | 4381 | } |
a69052be | 4382 | |
b128c09f BB |
4383 | if (zio->io_error) { |
4384 | /* | |
4385 | * If this I/O is attached to a particular vdev, | |
4386 | * generate an error message describing the I/O failure | |
4387 | * at the block level. We ignore these errors if the | |
4388 | * device is currently unavailable. | |
4389 | */ | |
c776b317 | 4390 | if (zio->io_error != ECKSUM && zio->io_vd != NULL && |
2bbec1c9 TH |
4391 | !vdev_is_dead(zio->io_vd)) { |
4392 | mutex_enter(&zio->io_vd->vdev_stat_lock); | |
4393 | if (zio->io_type == ZIO_TYPE_READ) { | |
4394 | zio->io_vd->vdev_stat.vs_read_errors++; | |
4395 | } else if (zio->io_type == ZIO_TYPE_WRITE) { | |
4396 | zio->io_vd->vdev_stat.vs_write_errors++; | |
4397 | } | |
4398 | mutex_exit(&zio->io_vd->vdev_stat_lock); | |
4399 | ||
c776b317 | 4400 | zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa, |
b5256303 | 4401 | zio->io_vd, &zio->io_bookmark, zio, 0, 0); |
2bbec1c9 | 4402 | } |
34dc7c2f | 4403 | |
428870ff BB |
4404 | if ((zio->io_error == EIO || !(zio->io_flags & |
4405 | (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && | |
c776b317 | 4406 | zio == zio->io_logical) { |
b128c09f BB |
4407 | /* |
4408 | * For logical I/O requests, tell the SPA to log the | |
4409 | * error and generate a logical data ereport. | |
4410 | */ | |
b5256303 | 4411 | spa_log_error(zio->io_spa, &zio->io_bookmark); |
d1d7e268 | 4412 | zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, |
b5256303 | 4413 | NULL, &zio->io_bookmark, zio, 0, 0); |
b128c09f BB |
4414 | } |
4415 | } | |
34dc7c2f | 4416 | |
c776b317 | 4417 | if (zio->io_error && zio == zio->io_logical) { |
b128c09f BB |
4418 | /* |
4419 | * Determine whether zio should be reexecuted. This will | |
4420 | * propagate all the way to the root via zio_notify_parent(). | |
4421 | */ | |
c776b317 | 4422 | ASSERT(zio->io_vd == NULL && zio->io_bp != NULL); |
428870ff | 4423 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); |
b128c09f | 4424 | |
428870ff BB |
4425 | if (IO_IS_ALLOCATING(zio) && |
4426 | !(zio->io_flags & ZIO_FLAG_CANFAIL)) { | |
b128c09f BB |
4427 | if (zio->io_error != ENOSPC) |
4428 | zio->io_reexecute |= ZIO_REEXECUTE_NOW; | |
4429 | else | |
4430 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff | 4431 | } |
b128c09f BB |
4432 | |
4433 | if ((zio->io_type == ZIO_TYPE_READ || | |
4434 | zio->io_type == ZIO_TYPE_FREE) && | |
572e2857 | 4435 | !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && |
b128c09f | 4436 | zio->io_error == ENXIO && |
c776b317 BB |
4437 | spa_load_state(zio->io_spa) == SPA_LOAD_NONE && |
4438 | spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE) | |
b128c09f BB |
4439 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; |
4440 | ||
4441 | if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) | |
4442 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff BB |
4443 | |
4444 | /* | |
4445 | * Here is a possibly good place to attempt to do | |
4446 | * either combinatorial reconstruction or error correction | |
4447 | * based on checksums. It also might be a good place | |
4448 | * to send out preliminary ereports before we suspend | |
4449 | * processing. | |
4450 | */ | |
34dc7c2f BB |
4451 | } |
4452 | ||
4453 | /* | |
b128c09f BB |
4454 | * If there were logical child errors, they apply to us now. |
4455 | * We defer this until now to avoid conflating logical child | |
4456 | * errors with errors that happened to the zio itself when | |
4457 | * updating vdev stats and reporting FMA events above. | |
34dc7c2f | 4458 | */ |
b128c09f | 4459 | zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); |
34dc7c2f | 4460 | |
428870ff BB |
4461 | if ((zio->io_error || zio->io_reexecute) && |
4462 | IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && | |
03c6040b | 4463 | !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) |
c776b317 | 4464 | zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp); |
9babb374 BB |
4465 | |
4466 | zio_gang_tree_free(&zio->io_gang_tree); | |
4467 | ||
4468 | /* | |
4469 | * Godfather I/Os should never suspend. | |
4470 | */ | |
4471 | if ((zio->io_flags & ZIO_FLAG_GODFATHER) && | |
4472 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) | |
a32494d2 | 4473 | zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND; |
9babb374 | 4474 | |
b128c09f BB |
4475 | if (zio->io_reexecute) { |
4476 | /* | |
4477 | * This is a logical I/O that wants to reexecute. | |
4478 | * | |
4479 | * Reexecute is top-down. When an i/o fails, if it's not | |
4480 | * the root, it simply notifies its parent and sticks around. | |
4481 | * The parent, seeing that it still has children in zio_done(), | |
4482 | * does the same. This percolates all the way up to the root. | |
4483 | * The root i/o will reexecute or suspend the entire tree. | |
4484 | * | |
4485 | * This approach ensures that zio_reexecute() honors | |
4486 | * all the original i/o dependency relationships, e.g. | |
4487 | * parents not executing until children are ready. | |
4488 | */ | |
4489 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
34dc7c2f | 4490 | |
9babb374 | 4491 | zio->io_gang_leader = NULL; |
b128c09f | 4492 | |
d164b209 BB |
4493 | mutex_enter(&zio->io_lock); |
4494 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
4495 | mutex_exit(&zio->io_lock); | |
4496 | ||
9babb374 BB |
4497 | /* |
4498 | * "The Godfather" I/O monitors its children but is | |
4499 | * not a true parent to them. It will track them through | |
4500 | * the pipeline but severs its ties whenever they get into | |
4501 | * trouble (e.g. suspended). This allows "The Godfather" | |
4502 | * I/O to return status without blocking. | |
4503 | */ | |
3dfb57a3 DB |
4504 | zl = NULL; |
4505 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; | |
4506 | pio = pio_next) { | |
4507 | zio_link_t *remove_zl = zl; | |
4508 | pio_next = zio_walk_parents(zio, &zl); | |
9babb374 BB |
4509 | |
4510 | if ((pio->io_flags & ZIO_FLAG_GODFATHER) && | |
4511 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { | |
3dfb57a3 | 4512 | zio_remove_child(pio, zio, remove_zl); |
62840030 MA |
4513 | /* |
4514 | * This is a rare code path, so we don't | |
4515 | * bother with "next_to_execute". | |
4516 | */ | |
4517 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, | |
4518 | NULL); | |
9babb374 BB |
4519 | } |
4520 | } | |
4521 | ||
d164b209 | 4522 | if ((pio = zio_unique_parent(zio)) != NULL) { |
b128c09f BB |
4523 | /* |
4524 | * We're not a root i/o, so there's nothing to do | |
4525 | * but notify our parent. Don't propagate errors | |
4526 | * upward since we haven't permanently failed yet. | |
4527 | */ | |
9babb374 | 4528 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f | 4529 | zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; |
62840030 MA |
4530 | /* |
4531 | * This is a rare code path, so we don't bother with | |
4532 | * "next_to_execute". | |
4533 | */ | |
4534 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL); | |
b128c09f BB |
4535 | } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { |
4536 | /* | |
4537 | * We'd fail again if we reexecuted now, so suspend | |
4538 | * until conditions improve (e.g. device comes online). | |
4539 | */ | |
cec3a0a1 | 4540 | zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR); |
b128c09f BB |
4541 | } else { |
4542 | /* | |
4543 | * Reexecution is potentially a huge amount of work. | |
4544 | * Hand it off to the otherwise-unused claim taskq. | |
4545 | */ | |
a38718a6 | 4546 | ASSERT(taskq_empty_ent(&zio->io_tqent)); |
7ef5e54e AL |
4547 | spa_taskq_dispatch_ent(zio->io_spa, |
4548 | ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE, | |
a38718a6 GA |
4549 | (task_func_t *)zio_reexecute, zio, 0, |
4550 | &zio->io_tqent); | |
b128c09f | 4551 | } |
62840030 | 4552 | return (NULL); |
34dc7c2f BB |
4553 | } |
4554 | ||
428870ff | 4555 | ASSERT(zio->io_child_count == 0); |
b128c09f BB |
4556 | ASSERT(zio->io_reexecute == 0); |
4557 | ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); | |
34dc7c2f | 4558 | |
428870ff BB |
4559 | /* |
4560 | * Report any checksum errors, since the I/O is complete. | |
4561 | */ | |
4562 | while (zio->io_cksum_report != NULL) { | |
4563 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
4564 | zio->io_cksum_report = zcr->zcr_next; | |
4565 | zcr->zcr_next = NULL; | |
4566 | zcr->zcr_finish(zcr, NULL); | |
4567 | zfs_ereport_free_checksum(zcr); | |
4568 | } | |
4569 | ||
920dd524 | 4570 | if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp && |
9b67f605 MA |
4571 | !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) && |
4572 | !(zio->io_flags & ZIO_FLAG_NOPWRITE)) { | |
920dd524 ED |
4573 | metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp); |
4574 | } | |
4575 | ||
d164b209 BB |
4576 | /* |
4577 | * It is the responsibility of the done callback to ensure that this | |
4578 | * particular zio is no longer discoverable for adoption, and as | |
4579 | * such, cannot acquire any new parents. | |
4580 | */ | |
b128c09f BB |
4581 | if (zio->io_done) |
4582 | zio->io_done(zio); | |
34dc7c2f | 4583 | |
d164b209 BB |
4584 | mutex_enter(&zio->io_lock); |
4585 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
4586 | mutex_exit(&zio->io_lock); | |
34dc7c2f | 4587 | |
62840030 MA |
4588 | /* |
4589 | * We are done executing this zio. We may want to execute a parent | |
4590 | * next. See the comment in zio_notify_parent(). | |
4591 | */ | |
4592 | zio_t *next_to_execute = NULL; | |
3dfb57a3 DB |
4593 | zl = NULL; |
4594 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) { | |
4595 | zio_link_t *remove_zl = zl; | |
4596 | pio_next = zio_walk_parents(zio, &zl); | |
4597 | zio_remove_child(pio, zio, remove_zl); | |
62840030 | 4598 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute); |
b128c09f | 4599 | } |
34dc7c2f | 4600 | |
b128c09f BB |
4601 | if (zio->io_waiter != NULL) { |
4602 | mutex_enter(&zio->io_lock); | |
4603 | zio->io_executor = NULL; | |
4604 | cv_broadcast(&zio->io_cv); | |
4605 | mutex_exit(&zio->io_lock); | |
4606 | } else { | |
4607 | zio_destroy(zio); | |
4608 | } | |
34dc7c2f | 4609 | |
62840030 | 4610 | return (next_to_execute); |
34dc7c2f BB |
4611 | } |
4612 | ||
4613 | /* | |
b128c09f BB |
4614 | * ========================================================================== |
4615 | * I/O pipeline definition | |
4616 | * ========================================================================== | |
34dc7c2f | 4617 | */ |
428870ff | 4618 | static zio_pipe_stage_t *zio_pipeline[] = { |
b128c09f | 4619 | NULL, |
b128c09f | 4620 | zio_read_bp_init, |
3dfb57a3 | 4621 | zio_write_bp_init, |
428870ff BB |
4622 | zio_free_bp_init, |
4623 | zio_issue_async, | |
3dfb57a3 | 4624 | zio_write_compress, |
b5256303 | 4625 | zio_encrypt, |
b128c09f | 4626 | zio_checksum_generate, |
03c6040b | 4627 | zio_nop_write, |
428870ff BB |
4628 | zio_ddt_read_start, |
4629 | zio_ddt_read_done, | |
4630 | zio_ddt_write, | |
4631 | zio_ddt_free, | |
b128c09f BB |
4632 | zio_gang_assemble, |
4633 | zio_gang_issue, | |
3dfb57a3 | 4634 | zio_dva_throttle, |
b128c09f BB |
4635 | zio_dva_allocate, |
4636 | zio_dva_free, | |
4637 | zio_dva_claim, | |
4638 | zio_ready, | |
4639 | zio_vdev_io_start, | |
4640 | zio_vdev_io_done, | |
4641 | zio_vdev_io_assess, | |
4642 | zio_checksum_verify, | |
4643 | zio_done | |
4644 | }; | |
c28b2279 | 4645 | |
9ae529ec | 4646 | |
9ae529ec | 4647 | |
9ae529ec | 4648 | |
fcff0f35 PD |
4649 | /* |
4650 | * Compare two zbookmark_phys_t's to see which we would reach first in a | |
4651 | * pre-order traversal of the object tree. | |
4652 | * | |
4653 | * This is simple in every case aside from the meta-dnode object. For all other | |
4654 | * objects, we traverse them in order (object 1 before object 2, and so on). | |
4655 | * However, all of these objects are traversed while traversing object 0, since | |
4656 | * the data it points to is the list of objects. Thus, we need to convert to a | |
4657 | * canonical representation so we can compare meta-dnode bookmarks to | |
4658 | * non-meta-dnode bookmarks. | |
4659 | * | |
4660 | * We do this by calculating "equivalents" for each field of the zbookmark. | |
4661 | * zbookmarks outside of the meta-dnode use their own object and level, and | |
4662 | * calculate the level 0 equivalent (the first L0 blkid that is contained in the | |
4663 | * blocks this bookmark refers to) by multiplying their blkid by their span | |
4664 | * (the number of L0 blocks contained within one block at their level). | |
4665 | * zbookmarks inside the meta-dnode calculate their object equivalent | |
4666 | * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use | |
4667 | * level + 1<<31 (any value larger than a level could ever be) for their level. | |
4668 | * This causes them to always compare before a bookmark in their object | |
4669 | * equivalent, compare appropriately to bookmarks in other objects, and to | |
4670 | * compare appropriately to other bookmarks in the meta-dnode. | |
4671 | */ | |
4672 | int | |
4673 | zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, | |
4674 | const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) | |
4675 | { | |
4676 | /* | |
4677 | * These variables represent the "equivalent" values for the zbookmark, | |
4678 | * after converting zbookmarks inside the meta dnode to their | |
4679 | * normal-object equivalents. | |
4680 | */ | |
4681 | uint64_t zb1obj, zb2obj; | |
4682 | uint64_t zb1L0, zb2L0; | |
4683 | uint64_t zb1level, zb2level; | |
4684 | ||
4685 | if (zb1->zb_object == zb2->zb_object && | |
4686 | zb1->zb_level == zb2->zb_level && | |
4687 | zb1->zb_blkid == zb2->zb_blkid) | |
4688 | return (0); | |
9ae529ec | 4689 | |
30af21b0 PD |
4690 | IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT); |
4691 | IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT); | |
4692 | ||
fcff0f35 PD |
4693 | /* |
4694 | * BP_SPANB calculates the span in blocks. | |
4695 | */ | |
4696 | zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); | |
4697 | zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); | |
9ae529ec CS |
4698 | |
4699 | if (zb1->zb_object == DMU_META_DNODE_OBJECT) { | |
fcff0f35 PD |
4700 | zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); |
4701 | zb1L0 = 0; | |
4702 | zb1level = zb1->zb_level + COMPARE_META_LEVEL; | |
4703 | } else { | |
4704 | zb1obj = zb1->zb_object; | |
4705 | zb1level = zb1->zb_level; | |
9ae529ec CS |
4706 | } |
4707 | ||
fcff0f35 PD |
4708 | if (zb2->zb_object == DMU_META_DNODE_OBJECT) { |
4709 | zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); | |
4710 | zb2L0 = 0; | |
4711 | zb2level = zb2->zb_level + COMPARE_META_LEVEL; | |
4712 | } else { | |
4713 | zb2obj = zb2->zb_object; | |
4714 | zb2level = zb2->zb_level; | |
4715 | } | |
4716 | ||
4717 | /* Now that we have a canonical representation, do the comparison. */ | |
4718 | if (zb1obj != zb2obj) | |
4719 | return (zb1obj < zb2obj ? -1 : 1); | |
4720 | else if (zb1L0 != zb2L0) | |
4721 | return (zb1L0 < zb2L0 ? -1 : 1); | |
4722 | else if (zb1level != zb2level) | |
4723 | return (zb1level > zb2level ? -1 : 1); | |
4724 | /* | |
4725 | * This can (theoretically) happen if the bookmarks have the same object | |
4726 | * and level, but different blkids, if the block sizes are not the same. | |
4727 | * There is presently no way to change the indirect block sizes | |
4728 | */ | |
4729 | return (0); | |
4730 | } | |
4731 | ||
4732 | /* | |
4733 | * This function checks the following: given that last_block is the place that | |
4734 | * our traversal stopped last time, does that guarantee that we've visited | |
4735 | * every node under subtree_root? Therefore, we can't just use the raw output | |
4736 | * of zbookmark_compare. We have to pass in a modified version of | |
4737 | * subtree_root; by incrementing the block id, and then checking whether | |
4738 | * last_block is before or equal to that, we can tell whether or not having | |
4739 | * visited last_block implies that all of subtree_root's children have been | |
4740 | * visited. | |
4741 | */ | |
4742 | boolean_t | |
4743 | zbookmark_subtree_completed(const dnode_phys_t *dnp, | |
4744 | const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) | |
4745 | { | |
4746 | zbookmark_phys_t mod_zb = *subtree_root; | |
4747 | mod_zb.zb_blkid++; | |
4748 | ASSERT(last_block->zb_level == 0); | |
4749 | ||
4750 | /* The objset_phys_t isn't before anything. */ | |
4751 | if (dnp == NULL) | |
9ae529ec | 4752 | return (B_FALSE); |
fcff0f35 PD |
4753 | |
4754 | /* | |
4755 | * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the | |
4756 | * data block size in sectors, because that variable is only used if | |
4757 | * the bookmark refers to a block in the meta-dnode. Since we don't | |
4758 | * know without examining it what object it refers to, and there's no | |
4759 | * harm in passing in this value in other cases, we always pass it in. | |
4760 | * | |
4761 | * We pass in 0 for the indirect block size shift because zb2 must be | |
4762 | * level 0. The indirect block size is only used to calculate the span | |
4763 | * of the bookmark, but since the bookmark must be level 0, the span is | |
4764 | * always 1, so the math works out. | |
4765 | * | |
4766 | * If you make changes to how the zbookmark_compare code works, be sure | |
4767 | * to make sure that this code still works afterwards. | |
4768 | */ | |
4769 | return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, | |
4770 | 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, | |
4771 | last_block) <= 0); | |
9ae529ec CS |
4772 | } |
4773 | ||
c28b2279 | 4774 | EXPORT_SYMBOL(zio_type_name); |
81971b13 BB |
4775 | EXPORT_SYMBOL(zio_buf_alloc); |
4776 | EXPORT_SYMBOL(zio_data_buf_alloc); | |
4777 | EXPORT_SYMBOL(zio_buf_free); | |
4778 | EXPORT_SYMBOL(zio_data_buf_free); | |
c28b2279 | 4779 | |
03fdcb9a MM |
4780 | /* BEGIN CSTYLED */ |
4781 | ZFS_MODULE_PARAM(zfs_zio, zio_, slow_io_ms, INT, ZMOD_RW, | |
ad796b8a | 4782 | "Max I/O completion time (milliseconds) before marking it as slow"); |
c409e464 | 4783 | |
03fdcb9a MM |
4784 | ZFS_MODULE_PARAM(zfs_zio, zio_, requeue_io_start_cut_in_line, INT, ZMOD_RW, |
4785 | "Prioritize requeued I/O"); | |
29dee3ee | 4786 | |
03fdcb9a | 4787 | ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_deferred_free, INT, ZMOD_RW, |
d1d7e268 | 4788 | "Defer frees starting in this pass"); |
29dee3ee | 4789 | |
03fdcb9a | 4790 | ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_dont_compress, INT, ZMOD_RW, |
d1d7e268 | 4791 | "Don't compress starting in this pass"); |
29dee3ee | 4792 | |
03fdcb9a | 4793 | ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_rewrite, INT, ZMOD_RW, |
d1d7e268 | 4794 | "Rewrite new bps starting in this pass"); |
3dfb57a3 | 4795 | |
03fdcb9a | 4796 | ZFS_MODULE_PARAM(zfs_zio, zio_, dva_throttle_enabled, INT, ZMOD_RW, |
3dfb57a3 | 4797 | "Throttle block allocations in the ZIO pipeline"); |
638dd5f4 | 4798 | |
03fdcb9a | 4799 | ZFS_MODULE_PARAM(zfs_zio, zio_, deadman_log_all, INT, ZMOD_RW, |
638dd5f4 | 4800 | "Log all slow ZIOs, not just those with vdevs"); |
03fdcb9a | 4801 | /* END CSTYLED */ |