]>
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. |
4f072827 | 23 | * Copyright (c) 2011, 2020 by Delphix. All rights reserved. |
a38718a6 | 24 | * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved. |
cc99f275 | 25 | * Copyright (c) 2017, Intel Corporation. |
10b3c7f5 MN |
26 | * Copyright (c) 2019, Klara Inc. |
27 | * Copyright (c) 2019, Allan Jude | |
f2286383 | 28 | * Copyright (c) 2021, Datto, Inc. |
34dc7c2f BB |
29 | */ |
30 | ||
f1512ee6 | 31 | #include <sys/sysmacros.h> |
34dc7c2f BB |
32 | #include <sys/zfs_context.h> |
33 | #include <sys/fm/fs/zfs.h> | |
34 | #include <sys/spa.h> | |
35 | #include <sys/txg.h> | |
36 | #include <sys/spa_impl.h> | |
37 | #include <sys/vdev_impl.h> | |
1b939560 | 38 | #include <sys/vdev_trim.h> |
34dc7c2f BB |
39 | #include <sys/zio_impl.h> |
40 | #include <sys/zio_compress.h> | |
41 | #include <sys/zio_checksum.h> | |
428870ff BB |
42 | #include <sys/dmu_objset.h> |
43 | #include <sys/arc.h> | |
44 | #include <sys/ddt.h> | |
9b67f605 | 45 | #include <sys/blkptr.h> |
b0bc7a84 | 46 | #include <sys/zfeature.h> |
d4a72f23 | 47 | #include <sys/dsl_scan.h> |
3dfb57a3 | 48 | #include <sys/metaslab_impl.h> |
193a37cb | 49 | #include <sys/time.h> |
e5d1c27e | 50 | #include <sys/trace_zfs.h> |
a6255b7f | 51 | #include <sys/abd.h> |
b5256303 | 52 | #include <sys/dsl_crypt.h> |
3f387973 | 53 | #include <cityhash.h> |
34dc7c2f | 54 | |
34dc7c2f BB |
55 | /* |
56 | * ========================================================================== | |
57 | * I/O type descriptions | |
58 | * ========================================================================== | |
59 | */ | |
18168da7 | 60 | const char *const zio_type_name[ZIO_TYPES] = { |
3dfb57a3 DB |
61 | /* |
62 | * Note: Linux kernel thread name length is limited | |
63 | * so these names will differ from upstream open zfs. | |
64 | */ | |
1b939560 | 65 | "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim" |
428870ff | 66 | }; |
34dc7c2f | 67 | |
27f2b90d | 68 | int zio_dva_throttle_enabled = B_TRUE; |
18168da7 | 69 | static int zio_deadman_log_all = B_FALSE; |
3dfb57a3 | 70 | |
34dc7c2f BB |
71 | /* |
72 | * ========================================================================== | |
73 | * I/O kmem caches | |
74 | * ========================================================================== | |
75 | */ | |
18168da7 AZ |
76 | static kmem_cache_t *zio_cache; |
77 | static kmem_cache_t *zio_link_cache; | |
34dc7c2f BB |
78 | kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; |
79 | kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
a6255b7f | 80 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
18168da7 AZ |
81 | static uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; |
82 | static uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
a6255b7f DQ |
83 | #endif |
84 | ||
ad796b8a | 85 | /* Mark IOs as "slow" if they take longer than 30 seconds */ |
18168da7 | 86 | static int zio_slow_io_ms = (30 * MILLISEC); |
34dc7c2f | 87 | |
fcff0f35 PD |
88 | #define BP_SPANB(indblkshift, level) \ |
89 | (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) | |
90 | #define COMPARE_META_LEVEL 0x80000000ul | |
55d85d5a GW |
91 | /* |
92 | * The following actions directly effect the spa's sync-to-convergence logic. | |
93 | * The values below define the sync pass when we start performing the action. | |
94 | * Care should be taken when changing these values as they directly impact | |
95 | * spa_sync() performance. Tuning these values may introduce subtle performance | |
96 | * pathologies and should only be done in the context of performance analysis. | |
97 | * These tunables will eventually be removed and replaced with #defines once | |
98 | * enough analysis has been done to determine optimal values. | |
99 | * | |
100 | * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that | |
101 | * regular blocks are not deferred. | |
be89734a MA |
102 | * |
103 | * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable | |
104 | * compression (including of metadata). In practice, we don't have this | |
105 | * many sync passes, so this has no effect. | |
106 | * | |
107 | * The original intent was that disabling compression would help the sync | |
108 | * passes to converge. However, in practice disabling compression increases | |
109 | * the average number of sync passes, because when we turn compression off, a | |
110 | * lot of block's size will change and thus we have to re-allocate (not | |
111 | * overwrite) them. It also increases the number of 128KB allocations (e.g. | |
112 | * for indirect blocks and spacemaps) because these will not be compressed. | |
113 | * The 128K allocations are especially detrimental to performance on highly | |
114 | * fragmented systems, which may have very few free segments of this size, | |
115 | * and may need to load new metaslabs to satisfy 128K allocations. | |
55d85d5a GW |
116 | */ |
117 | int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ | |
18168da7 AZ |
118 | static int zfs_sync_pass_dont_compress = 8; /* don't compress s. i. t. p. */ |
119 | static int zfs_sync_pass_rewrite = 2; /* rewrite new bps s. i. t. p. */ | |
55d85d5a | 120 | |
34dc7c2f | 121 | /* |
b128c09f BB |
122 | * An allocating zio is one that either currently has the DVA allocate |
123 | * stage set or will have it later in its lifetime. | |
34dc7c2f | 124 | */ |
428870ff BB |
125 | #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) |
126 | ||
3c502d3b MM |
127 | /* |
128 | * Enable smaller cores by excluding metadata | |
129 | * allocations as well. | |
130 | */ | |
131 | int zio_exclude_metadata = 0; | |
18168da7 | 132 | static int zio_requeue_io_start_cut_in_line = 1; |
428870ff BB |
133 | |
134 | #ifdef ZFS_DEBUG | |
18168da7 | 135 | static const int zio_buf_debug_limit = 16384; |
428870ff | 136 | #else |
18168da7 | 137 | static const int zio_buf_debug_limit = 0; |
428870ff | 138 | #endif |
34dc7c2f | 139 | |
da6b4005 NB |
140 | static inline void __zio_execute(zio_t *zio); |
141 | ||
3dfb57a3 DB |
142 | static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t); |
143 | ||
34dc7c2f BB |
144 | void |
145 | zio_init(void) | |
146 | { | |
147 | size_t c; | |
34dc7c2f | 148 | |
3941503c BB |
149 | zio_cache = kmem_cache_create("zio_cache", |
150 | sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); | |
d164b209 | 151 | zio_link_cache = kmem_cache_create("zio_link_cache", |
6795a698 | 152 | sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); |
34dc7c2f BB |
153 | |
154 | /* | |
155 | * For small buffers, we want a cache for each multiple of | |
f1512ee6 MA |
156 | * SPA_MINBLOCKSIZE. For larger buffers, we want a cache |
157 | * for each quarter-power of 2. | |
34dc7c2f BB |
158 | */ |
159 | for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { | |
160 | size_t size = (c + 1) << SPA_MINBLOCKSHIFT; | |
161 | size_t p2 = size; | |
162 | size_t align = 0; | |
3c502d3b MM |
163 | size_t data_cflags, cflags; |
164 | ||
165 | data_cflags = KMC_NODEBUG; | |
166 | cflags = (zio_exclude_metadata || size > zio_buf_debug_limit) ? | |
167 | KMC_NODEBUG : 0; | |
34dc7c2f | 168 | |
34328f3c | 169 | #if defined(_ILP32) && defined(_KERNEL) |
f1512ee6 MA |
170 | /* |
171 | * Cache size limited to 1M on 32-bit platforms until ARC | |
172 | * buffers no longer require virtual address space. | |
173 | */ | |
174 | if (size > zfs_max_recordsize) | |
175 | break; | |
176 | #endif | |
177 | ||
178 | while (!ISP2(p2)) | |
34dc7c2f BB |
179 | p2 &= p2 - 1; |
180 | ||
498877ba MA |
181 | #ifndef _KERNEL |
182 | /* | |
183 | * If we are using watchpoints, put each buffer on its own page, | |
184 | * to eliminate the performance overhead of trapping to the | |
185 | * kernel when modifying a non-watched buffer that shares the | |
186 | * page with a watched buffer. | |
187 | */ | |
188 | if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) | |
189 | continue; | |
fcf64f45 BB |
190 | /* |
191 | * Here's the problem - on 4K native devices in userland on | |
192 | * Linux using O_DIRECT, buffers must be 4K aligned or I/O | |
193 | * will fail with EINVAL, causing zdb (and others) to coredump. | |
194 | * Since userland probably doesn't need optimized buffer caches, | |
195 | * we just force 4K alignment on everything. | |
196 | */ | |
197 | align = 8 * SPA_MINBLOCKSIZE; | |
198 | #else | |
24fa2034 | 199 | if (size < PAGESIZE) { |
34dc7c2f | 200 | align = SPA_MINBLOCKSIZE; |
498877ba | 201 | } else if (IS_P2ALIGNED(size, p2 >> 2)) { |
24fa2034 | 202 | align = PAGESIZE; |
34dc7c2f | 203 | } |
fcf64f45 | 204 | #endif |
34dc7c2f BB |
205 | |
206 | if (align != 0) { | |
207 | char name[36]; | |
309c32c9 MG |
208 | if (cflags == data_cflags) { |
209 | /* | |
210 | * Resulting kmem caches would be identical. | |
211 | * Save memory by creating only one. | |
212 | */ | |
213 | (void) snprintf(name, sizeof (name), | |
214 | "zio_buf_comb_%lu", (ulong_t)size); | |
215 | zio_buf_cache[c] = kmem_cache_create(name, | |
216 | size, align, NULL, NULL, NULL, NULL, NULL, | |
217 | cflags); | |
218 | zio_data_buf_cache[c] = zio_buf_cache[c]; | |
219 | continue; | |
220 | } | |
c9e319fa JL |
221 | (void) snprintf(name, sizeof (name), "zio_buf_%lu", |
222 | (ulong_t)size); | |
34dc7c2f | 223 | zio_buf_cache[c] = kmem_cache_create(name, size, |
6442f3cf | 224 | align, NULL, NULL, NULL, NULL, NULL, cflags); |
34dc7c2f | 225 | |
c9e319fa JL |
226 | (void) snprintf(name, sizeof (name), "zio_data_buf_%lu", |
227 | (ulong_t)size); | |
34dc7c2f | 228 | zio_data_buf_cache[c] = kmem_cache_create(name, size, |
18ca574f | 229 | align, NULL, NULL, NULL, NULL, NULL, data_cflags); |
34dc7c2f BB |
230 | } |
231 | } | |
232 | ||
233 | while (--c != 0) { | |
234 | ASSERT(zio_buf_cache[c] != NULL); | |
235 | if (zio_buf_cache[c - 1] == NULL) | |
236 | zio_buf_cache[c - 1] = zio_buf_cache[c]; | |
237 | ||
238 | ASSERT(zio_data_buf_cache[c] != NULL); | |
239 | if (zio_data_buf_cache[c - 1] == NULL) | |
240 | zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; | |
241 | } | |
242 | ||
34dc7c2f | 243 | zio_inject_init(); |
9759c60f ED |
244 | |
245 | lz4_init(); | |
34dc7c2f BB |
246 | } |
247 | ||
248 | void | |
249 | zio_fini(void) | |
250 | { | |
309c32c9 | 251 | size_t n = SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; |
34dc7c2f | 252 | |
a6255b7f | 253 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
309c32c9 MG |
254 | for (size_t i = 0; i < n; i++) { |
255 | if (zio_buf_cache_allocs[i] != zio_buf_cache_frees[i]) | |
a6255b7f | 256 | (void) printf("zio_fini: [%d] %llu != %llu\n", |
309c32c9 MG |
257 | (int)((i + 1) << SPA_MINBLOCKSHIFT), |
258 | (long long unsigned)zio_buf_cache_allocs[i], | |
259 | (long long unsigned)zio_buf_cache_frees[i]); | |
260 | } | |
f1512ee6 | 261 | #endif |
309c32c9 MG |
262 | |
263 | /* | |
264 | * The same kmem cache can show up multiple times in both zio_buf_cache | |
265 | * and zio_data_buf_cache. Do a wasteful but trivially correct scan to | |
266 | * sort it out. | |
267 | */ | |
268 | for (size_t i = 0; i < n; i++) { | |
269 | kmem_cache_t *cache = zio_buf_cache[i]; | |
270 | if (cache == NULL) | |
271 | continue; | |
272 | for (size_t j = i; j < n; j++) { | |
273 | if (cache == zio_buf_cache[j]) | |
274 | zio_buf_cache[j] = NULL; | |
275 | if (cache == zio_data_buf_cache[j]) | |
276 | zio_data_buf_cache[j] = NULL; | |
34dc7c2f | 277 | } |
309c32c9 MG |
278 | kmem_cache_destroy(cache); |
279 | } | |
34dc7c2f | 280 | |
309c32c9 MG |
281 | for (size_t i = 0; i < n; i++) { |
282 | kmem_cache_t *cache = zio_data_buf_cache[i]; | |
283 | if (cache == NULL) | |
284 | continue; | |
285 | for (size_t j = i; j < n; j++) { | |
286 | if (cache == zio_data_buf_cache[j]) | |
287 | zio_data_buf_cache[j] = NULL; | |
34dc7c2f | 288 | } |
309c32c9 MG |
289 | kmem_cache_destroy(cache); |
290 | } | |
291 | ||
292 | for (size_t i = 0; i < n; i++) { | |
293 | VERIFY3P(zio_buf_cache[i], ==, NULL); | |
294 | VERIFY3P(zio_data_buf_cache[i], ==, NULL); | |
34dc7c2f BB |
295 | } |
296 | ||
d164b209 | 297 | kmem_cache_destroy(zio_link_cache); |
34dc7c2f BB |
298 | kmem_cache_destroy(zio_cache); |
299 | ||
300 | zio_inject_fini(); | |
9759c60f ED |
301 | |
302 | lz4_fini(); | |
34dc7c2f BB |
303 | } |
304 | ||
305 | /* | |
306 | * ========================================================================== | |
307 | * Allocate and free I/O buffers | |
308 | * ========================================================================== | |
309 | */ | |
310 | ||
311 | /* | |
312 | * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a | |
313 | * crashdump if the kernel panics, so use it judiciously. Obviously, it's | |
314 | * useful to inspect ZFS metadata, but if possible, we should avoid keeping | |
315 | * excess / transient data in-core during a crashdump. | |
316 | */ | |
317 | void * | |
318 | zio_buf_alloc(size_t size) | |
319 | { | |
320 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
321 | ||
63e3a861 | 322 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
a6255b7f DQ |
323 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
324 | atomic_add_64(&zio_buf_cache_allocs[c], 1); | |
325 | #endif | |
34dc7c2f | 326 | |
efcd79a8 | 327 | return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); |
34dc7c2f BB |
328 | } |
329 | ||
330 | /* | |
331 | * Use zio_data_buf_alloc to allocate data. The data will not appear in a | |
332 | * crashdump if the kernel panics. This exists so that we will limit the amount | |
333 | * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount | |
334 | * of kernel heap dumped to disk when the kernel panics) | |
335 | */ | |
336 | void * | |
337 | zio_data_buf_alloc(size_t size) | |
338 | { | |
339 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
340 | ||
63e3a861 | 341 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f | 342 | |
efcd79a8 | 343 | return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); |
34dc7c2f BB |
344 | } |
345 | ||
346 | void | |
347 | zio_buf_free(void *buf, size_t size) | |
348 | { | |
349 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
350 | ||
63e3a861 | 351 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
a6255b7f DQ |
352 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
353 | atomic_add_64(&zio_buf_cache_frees[c], 1); | |
354 | #endif | |
34dc7c2f BB |
355 | |
356 | kmem_cache_free(zio_buf_cache[c], buf); | |
357 | } | |
358 | ||
359 | void | |
360 | zio_data_buf_free(void *buf, size_t size) | |
361 | { | |
362 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
363 | ||
63e3a861 | 364 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f BB |
365 | |
366 | kmem_cache_free(zio_data_buf_cache[c], buf); | |
367 | } | |
368 | ||
84c07ada GN |
369 | static void |
370 | zio_abd_free(void *abd, size_t size) | |
371 | { | |
14e4e3cb | 372 | (void) size; |
84c07ada GN |
373 | abd_free((abd_t *)abd); |
374 | } | |
375 | ||
34dc7c2f BB |
376 | /* |
377 | * ========================================================================== | |
378 | * Push and pop I/O transform buffers | |
379 | * ========================================================================== | |
380 | */ | |
d3c2ae1c | 381 | void |
a6255b7f | 382 | zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize, |
e9aa730c | 383 | zio_transform_func_t *transform) |
34dc7c2f | 384 | { |
79c76d5b | 385 | zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); |
34dc7c2f | 386 | |
a6255b7f | 387 | zt->zt_orig_abd = zio->io_abd; |
b128c09f | 388 | zt->zt_orig_size = zio->io_size; |
34dc7c2f | 389 | zt->zt_bufsize = bufsize; |
b128c09f | 390 | zt->zt_transform = transform; |
34dc7c2f BB |
391 | |
392 | zt->zt_next = zio->io_transform_stack; | |
393 | zio->io_transform_stack = zt; | |
394 | ||
a6255b7f | 395 | zio->io_abd = data; |
34dc7c2f BB |
396 | zio->io_size = size; |
397 | } | |
398 | ||
d3c2ae1c | 399 | void |
b128c09f | 400 | zio_pop_transforms(zio_t *zio) |
34dc7c2f | 401 | { |
b128c09f BB |
402 | zio_transform_t *zt; |
403 | ||
404 | while ((zt = zio->io_transform_stack) != NULL) { | |
405 | if (zt->zt_transform != NULL) | |
406 | zt->zt_transform(zio, | |
a6255b7f | 407 | zt->zt_orig_abd, zt->zt_orig_size); |
34dc7c2f | 408 | |
428870ff | 409 | if (zt->zt_bufsize != 0) |
a6255b7f | 410 | abd_free(zio->io_abd); |
34dc7c2f | 411 | |
a6255b7f | 412 | zio->io_abd = zt->zt_orig_abd; |
b128c09f BB |
413 | zio->io_size = zt->zt_orig_size; |
414 | zio->io_transform_stack = zt->zt_next; | |
34dc7c2f | 415 | |
b128c09f | 416 | kmem_free(zt, sizeof (zio_transform_t)); |
34dc7c2f BB |
417 | } |
418 | } | |
419 | ||
b128c09f BB |
420 | /* |
421 | * ========================================================================== | |
b5256303 | 422 | * I/O transform callbacks for subblocks, decompression, and decryption |
b128c09f BB |
423 | * ========================================================================== |
424 | */ | |
425 | static void | |
a6255b7f | 426 | zio_subblock(zio_t *zio, abd_t *data, uint64_t size) |
b128c09f BB |
427 | { |
428 | ASSERT(zio->io_size > size); | |
429 | ||
430 | if (zio->io_type == ZIO_TYPE_READ) | |
a6255b7f | 431 | abd_copy(data, zio->io_abd, size); |
b128c09f BB |
432 | } |
433 | ||
434 | static void | |
a6255b7f | 435 | zio_decompress(zio_t *zio, abd_t *data, uint64_t size) |
b128c09f | 436 | { |
a6255b7f DQ |
437 | if (zio->io_error == 0) { |
438 | void *tmp = abd_borrow_buf(data, size); | |
439 | int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), | |
10b3c7f5 MN |
440 | zio->io_abd, tmp, zio->io_size, size, |
441 | &zio->io_prop.zp_complevel); | |
a6255b7f DQ |
442 | abd_return_buf_copy(data, tmp, size); |
443 | ||
c3bd3fb4 TC |
444 | if (zio_injection_enabled && ret == 0) |
445 | ret = zio_handle_fault_injection(zio, EINVAL); | |
446 | ||
a6255b7f DQ |
447 | if (ret != 0) |
448 | zio->io_error = SET_ERROR(EIO); | |
449 | } | |
b128c09f BB |
450 | } |
451 | ||
b5256303 TC |
452 | static void |
453 | zio_decrypt(zio_t *zio, abd_t *data, uint64_t size) | |
454 | { | |
455 | int ret; | |
456 | void *tmp; | |
457 | blkptr_t *bp = zio->io_bp; | |
ae76f45c TC |
458 | spa_t *spa = zio->io_spa; |
459 | uint64_t dsobj = zio->io_bookmark.zb_objset; | |
b5256303 TC |
460 | uint64_t lsize = BP_GET_LSIZE(bp); |
461 | dmu_object_type_t ot = BP_GET_TYPE(bp); | |
462 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
463 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
464 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
465 | boolean_t no_crypt = B_FALSE; | |
466 | ||
467 | ASSERT(BP_USES_CRYPT(bp)); | |
468 | ASSERT3U(size, !=, 0); | |
469 | ||
470 | if (zio->io_error != 0) | |
471 | return; | |
472 | ||
473 | /* | |
474 | * Verify the cksum of MACs stored in an indirect bp. It will always | |
475 | * be possible to verify this since it does not require an encryption | |
476 | * key. | |
477 | */ | |
478 | if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) { | |
479 | zio_crypt_decode_mac_bp(bp, mac); | |
480 | ||
481 | if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) { | |
482 | /* | |
483 | * We haven't decompressed the data yet, but | |
484 | * zio_crypt_do_indirect_mac_checksum() requires | |
485 | * decompressed data to be able to parse out the MACs | |
486 | * from the indirect block. We decompress it now and | |
487 | * throw away the result after we are finished. | |
488 | */ | |
489 | tmp = zio_buf_alloc(lsize); | |
490 | ret = zio_decompress_data(BP_GET_COMPRESS(bp), | |
10b3c7f5 MN |
491 | zio->io_abd, tmp, zio->io_size, lsize, |
492 | &zio->io_prop.zp_complevel); | |
b5256303 TC |
493 | if (ret != 0) { |
494 | ret = SET_ERROR(EIO); | |
495 | goto error; | |
496 | } | |
497 | ret = zio_crypt_do_indirect_mac_checksum(B_FALSE, | |
498 | tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac); | |
499 | zio_buf_free(tmp, lsize); | |
500 | } else { | |
501 | ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE, | |
502 | zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac); | |
503 | } | |
504 | abd_copy(data, zio->io_abd, size); | |
505 | ||
be9a5c35 TC |
506 | if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) { |
507 | ret = zio_handle_decrypt_injection(spa, | |
508 | &zio->io_bookmark, ot, ECKSUM); | |
509 | } | |
b5256303 TC |
510 | if (ret != 0) |
511 | goto error; | |
512 | ||
513 | return; | |
514 | } | |
515 | ||
516 | /* | |
517 | * If this is an authenticated block, just check the MAC. It would be | |
518 | * nice to separate this out into its own flag, but for the moment | |
519 | * enum zio_flag is out of bits. | |
520 | */ | |
521 | if (BP_IS_AUTHENTICATED(bp)) { | |
522 | if (ot == DMU_OT_OBJSET) { | |
ae76f45c TC |
523 | ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, |
524 | dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp)); | |
b5256303 TC |
525 | } else { |
526 | zio_crypt_decode_mac_bp(bp, mac); | |
ae76f45c TC |
527 | ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, |
528 | zio->io_abd, size, mac); | |
be9a5c35 TC |
529 | if (zio_injection_enabled && ret == 0) { |
530 | ret = zio_handle_decrypt_injection(spa, | |
531 | &zio->io_bookmark, ot, ECKSUM); | |
532 | } | |
b5256303 TC |
533 | } |
534 | abd_copy(data, zio->io_abd, size); | |
535 | ||
536 | if (ret != 0) | |
537 | goto error; | |
538 | ||
539 | return; | |
540 | } | |
541 | ||
542 | zio_crypt_decode_params_bp(bp, salt, iv); | |
543 | ||
544 | if (ot == DMU_OT_INTENT_LOG) { | |
545 | tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t)); | |
546 | zio_crypt_decode_mac_zil(tmp, mac); | |
547 | abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t)); | |
548 | } else { | |
549 | zio_crypt_decode_mac_bp(bp, mac); | |
550 | } | |
551 | ||
be9a5c35 TC |
552 | ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp), |
553 | BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data, | |
554 | zio->io_abd, &no_crypt); | |
b5256303 TC |
555 | if (no_crypt) |
556 | abd_copy(data, zio->io_abd, size); | |
557 | ||
558 | if (ret != 0) | |
559 | goto error; | |
560 | ||
561 | return; | |
562 | ||
563 | error: | |
564 | /* assert that the key was found unless this was speculative */ | |
be9a5c35 | 565 | ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE)); |
b5256303 TC |
566 | |
567 | /* | |
568 | * If there was a decryption / authentication error return EIO as | |
569 | * the io_error. If this was not a speculative zio, create an ereport. | |
570 | */ | |
571 | if (ret == ECKSUM) { | |
a2c2ed1b | 572 | zio->io_error = SET_ERROR(EIO); |
b5256303 | 573 | if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) { |
be9a5c35 | 574 | spa_log_error(spa, &zio->io_bookmark); |
1144586b | 575 | (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION, |
4f072827 | 576 | spa, NULL, &zio->io_bookmark, zio, 0); |
b5256303 TC |
577 | } |
578 | } else { | |
579 | zio->io_error = ret; | |
580 | } | |
581 | } | |
582 | ||
b128c09f BB |
583 | /* |
584 | * ========================================================================== | |
585 | * I/O parent/child relationships and pipeline interlocks | |
586 | * ========================================================================== | |
587 | */ | |
d164b209 | 588 | zio_t * |
3dfb57a3 | 589 | zio_walk_parents(zio_t *cio, zio_link_t **zl) |
d164b209 | 590 | { |
d164b209 | 591 | list_t *pl = &cio->io_parent_list; |
b128c09f | 592 | |
3dfb57a3 DB |
593 | *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl); |
594 | if (*zl == NULL) | |
d164b209 BB |
595 | return (NULL); |
596 | ||
3dfb57a3 DB |
597 | ASSERT((*zl)->zl_child == cio); |
598 | return ((*zl)->zl_parent); | |
d164b209 BB |
599 | } |
600 | ||
601 | zio_t * | |
3dfb57a3 | 602 | zio_walk_children(zio_t *pio, zio_link_t **zl) |
d164b209 | 603 | { |
d164b209 BB |
604 | list_t *cl = &pio->io_child_list; |
605 | ||
a8b2e306 TC |
606 | ASSERT(MUTEX_HELD(&pio->io_lock)); |
607 | ||
3dfb57a3 DB |
608 | *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl); |
609 | if (*zl == NULL) | |
d164b209 BB |
610 | return (NULL); |
611 | ||
3dfb57a3 DB |
612 | ASSERT((*zl)->zl_parent == pio); |
613 | return ((*zl)->zl_child); | |
d164b209 BB |
614 | } |
615 | ||
616 | zio_t * | |
617 | zio_unique_parent(zio_t *cio) | |
618 | { | |
3dfb57a3 DB |
619 | zio_link_t *zl = NULL; |
620 | zio_t *pio = zio_walk_parents(cio, &zl); | |
d164b209 | 621 | |
3dfb57a3 | 622 | VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL); |
d164b209 BB |
623 | return (pio); |
624 | } | |
625 | ||
626 | void | |
627 | zio_add_child(zio_t *pio, zio_t *cio) | |
b128c09f | 628 | { |
79c76d5b | 629 | zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); |
d164b209 BB |
630 | |
631 | /* | |
632 | * Logical I/Os can have logical, gang, or vdev children. | |
633 | * Gang I/Os can have gang or vdev children. | |
634 | * Vdev I/Os can only have vdev children. | |
635 | * The following ASSERT captures all of these constraints. | |
636 | */ | |
1ce23dca | 637 | ASSERT3S(cio->io_child_type, <=, pio->io_child_type); |
d164b209 BB |
638 | |
639 | zl->zl_parent = pio; | |
640 | zl->zl_child = cio; | |
641 | ||
b128c09f | 642 | mutex_enter(&pio->io_lock); |
a8b2e306 | 643 | mutex_enter(&cio->io_lock); |
d164b209 BB |
644 | |
645 | ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); | |
646 | ||
1c27024e | 647 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 BB |
648 | pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; |
649 | ||
650 | list_insert_head(&pio->io_child_list, zl); | |
651 | list_insert_head(&cio->io_parent_list, zl); | |
652 | ||
428870ff BB |
653 | pio->io_child_count++; |
654 | cio->io_parent_count++; | |
655 | ||
d164b209 | 656 | mutex_exit(&cio->io_lock); |
a8b2e306 | 657 | mutex_exit(&pio->io_lock); |
b128c09f BB |
658 | } |
659 | ||
34dc7c2f | 660 | static void |
d164b209 | 661 | zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) |
b128c09f | 662 | { |
d164b209 BB |
663 | ASSERT(zl->zl_parent == pio); |
664 | ASSERT(zl->zl_child == cio); | |
b128c09f BB |
665 | |
666 | mutex_enter(&pio->io_lock); | |
a8b2e306 | 667 | mutex_enter(&cio->io_lock); |
d164b209 BB |
668 | |
669 | list_remove(&pio->io_child_list, zl); | |
670 | list_remove(&cio->io_parent_list, zl); | |
671 | ||
428870ff BB |
672 | pio->io_child_count--; |
673 | cio->io_parent_count--; | |
674 | ||
d164b209 | 675 | mutex_exit(&cio->io_lock); |
a8b2e306 | 676 | mutex_exit(&pio->io_lock); |
d164b209 | 677 | kmem_cache_free(zio_link_cache, zl); |
b128c09f BB |
678 | } |
679 | ||
680 | static boolean_t | |
ddc751d5 | 681 | zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait) |
34dc7c2f | 682 | { |
b128c09f BB |
683 | boolean_t waiting = B_FALSE; |
684 | ||
685 | mutex_enter(&zio->io_lock); | |
686 | ASSERT(zio->io_stall == NULL); | |
ddc751d5 GW |
687 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) { |
688 | if (!(ZIO_CHILD_BIT_IS_SET(childbits, c))) | |
689 | continue; | |
690 | ||
691 | uint64_t *countp = &zio->io_children[c][wait]; | |
692 | if (*countp != 0) { | |
693 | zio->io_stage >>= 1; | |
694 | ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN); | |
695 | zio->io_stall = countp; | |
696 | waiting = B_TRUE; | |
697 | break; | |
698 | } | |
b128c09f BB |
699 | } |
700 | mutex_exit(&zio->io_lock); | |
b128c09f BB |
701 | return (waiting); |
702 | } | |
34dc7c2f | 703 | |
bf701a83 BB |
704 | __attribute__((always_inline)) |
705 | static inline void | |
62840030 MA |
706 | zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait, |
707 | zio_t **next_to_executep) | |
b128c09f BB |
708 | { |
709 | uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; | |
710 | int *errorp = &pio->io_child_error[zio->io_child_type]; | |
34dc7c2f | 711 | |
b128c09f BB |
712 | mutex_enter(&pio->io_lock); |
713 | if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) | |
714 | *errorp = zio_worst_error(*errorp, zio->io_error); | |
715 | pio->io_reexecute |= zio->io_reexecute; | |
716 | ASSERT3U(*countp, >, 0); | |
e8b96c60 MA |
717 | |
718 | (*countp)--; | |
719 | ||
720 | if (*countp == 0 && pio->io_stall == countp) { | |
3dfb57a3 DB |
721 | zio_taskq_type_t type = |
722 | pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE : | |
723 | ZIO_TASKQ_INTERRUPT; | |
b128c09f BB |
724 | pio->io_stall = NULL; |
725 | mutex_exit(&pio->io_lock); | |
62840030 | 726 | |
3dfb57a3 | 727 | /* |
62840030 MA |
728 | * If we can tell the caller to execute this parent next, do |
729 | * so. Otherwise dispatch the parent zio as its own task. | |
730 | * | |
731 | * Having the caller execute the parent when possible reduces | |
732 | * locking on the zio taskq's, reduces context switch | |
733 | * overhead, and has no recursion penalty. Note that one | |
734 | * read from disk typically causes at least 3 zio's: a | |
735 | * zio_null(), the logical zio_read(), and then a physical | |
736 | * zio. When the physical ZIO completes, we are able to call | |
737 | * zio_done() on all 3 of these zio's from one invocation of | |
738 | * zio_execute() by returning the parent back to | |
739 | * zio_execute(). Since the parent isn't executed until this | |
740 | * thread returns back to zio_execute(), the caller should do | |
741 | * so promptly. | |
742 | * | |
743 | * In other cases, dispatching the parent prevents | |
744 | * overflowing the stack when we have deeply nested | |
745 | * parent-child relationships, as we do with the "mega zio" | |
746 | * of writes for spa_sync(), and the chain of ZIL blocks. | |
3dfb57a3 | 747 | */ |
62840030 MA |
748 | if (next_to_executep != NULL && *next_to_executep == NULL) { |
749 | *next_to_executep = pio; | |
750 | } else { | |
751 | zio_taskq_dispatch(pio, type, B_FALSE); | |
752 | } | |
b128c09f BB |
753 | } else { |
754 | mutex_exit(&pio->io_lock); | |
34dc7c2f BB |
755 | } |
756 | } | |
757 | ||
b128c09f BB |
758 | static void |
759 | zio_inherit_child_errors(zio_t *zio, enum zio_child c) | |
760 | { | |
761 | if (zio->io_child_error[c] != 0 && zio->io_error == 0) | |
762 | zio->io_error = zio->io_child_error[c]; | |
763 | } | |
764 | ||
3dfb57a3 | 765 | int |
64fc7762 | 766 | zio_bookmark_compare(const void *x1, const void *x2) |
3dfb57a3 DB |
767 | { |
768 | const zio_t *z1 = x1; | |
769 | const zio_t *z2 = x2; | |
3dfb57a3 | 770 | |
64fc7762 MA |
771 | if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset) |
772 | return (-1); | |
773 | if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset) | |
774 | return (1); | |
3dfb57a3 | 775 | |
64fc7762 MA |
776 | if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object) |
777 | return (-1); | |
778 | if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object) | |
779 | return (1); | |
3dfb57a3 | 780 | |
64fc7762 MA |
781 | if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level) |
782 | return (-1); | |
783 | if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level) | |
784 | return (1); | |
785 | ||
786 | if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid) | |
787 | return (-1); | |
788 | if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid) | |
789 | return (1); | |
790 | ||
791 | if (z1 < z2) | |
792 | return (-1); | |
793 | if (z1 > z2) | |
794 | return (1); | |
795 | ||
796 | return (0); | |
3dfb57a3 DB |
797 | } |
798 | ||
34dc7c2f BB |
799 | /* |
800 | * ========================================================================== | |
b128c09f | 801 | * Create the various types of I/O (read, write, free, etc) |
34dc7c2f BB |
802 | * ========================================================================== |
803 | */ | |
804 | static zio_t * | |
428870ff | 805 | zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
a6255b7f | 806 | abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done, |
2aa34383 DK |
807 | void *private, zio_type_t type, zio_priority_t priority, |
808 | enum zio_flag flags, vdev_t *vd, uint64_t offset, | |
809 | const zbookmark_phys_t *zb, enum zio_stage stage, | |
810 | enum zio_stage pipeline) | |
34dc7c2f BB |
811 | { |
812 | zio_t *zio; | |
813 | ||
1b939560 | 814 | IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE); |
2aa34383 | 815 | ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); |
b128c09f BB |
816 | ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); |
817 | ||
818 | ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); | |
819 | ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); | |
820 | ASSERT(vd || stage == ZIO_STAGE_OPEN); | |
34dc7c2f | 821 | |
b5256303 | 822 | IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0); |
2aa34383 | 823 | |
79c76d5b | 824 | zio = kmem_cache_alloc(zio_cache, KM_SLEEP); |
3941503c BB |
825 | bzero(zio, sizeof (zio_t)); |
826 | ||
448d7aaa | 827 | mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL); |
3941503c BB |
828 | cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); |
829 | ||
830 | list_create(&zio->io_parent_list, sizeof (zio_link_t), | |
831 | offsetof(zio_link_t, zl_parent_node)); | |
832 | list_create(&zio->io_child_list, sizeof (zio_link_t), | |
833 | offsetof(zio_link_t, zl_child_node)); | |
4e21fd06 | 834 | metaslab_trace_init(&zio->io_alloc_list); |
d164b209 | 835 | |
b128c09f BB |
836 | if (vd != NULL) |
837 | zio->io_child_type = ZIO_CHILD_VDEV; | |
838 | else if (flags & ZIO_FLAG_GANG_CHILD) | |
839 | zio->io_child_type = ZIO_CHILD_GANG; | |
428870ff BB |
840 | else if (flags & ZIO_FLAG_DDT_CHILD) |
841 | zio->io_child_type = ZIO_CHILD_DDT; | |
b128c09f BB |
842 | else |
843 | zio->io_child_type = ZIO_CHILD_LOGICAL; | |
844 | ||
34dc7c2f | 845 | if (bp != NULL) { |
428870ff | 846 | zio->io_bp = (blkptr_t *)bp; |
34dc7c2f BB |
847 | zio->io_bp_copy = *bp; |
848 | zio->io_bp_orig = *bp; | |
428870ff BB |
849 | if (type != ZIO_TYPE_WRITE || |
850 | zio->io_child_type == ZIO_CHILD_DDT) | |
b128c09f | 851 | zio->io_bp = &zio->io_bp_copy; /* so caller can free */ |
9babb374 | 852 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) |
b128c09f | 853 | zio->io_logical = zio; |
9babb374 BB |
854 | if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) |
855 | pipeline |= ZIO_GANG_STAGES; | |
34dc7c2f | 856 | } |
b128c09f BB |
857 | |
858 | zio->io_spa = spa; | |
859 | zio->io_txg = txg; | |
34dc7c2f BB |
860 | zio->io_done = done; |
861 | zio->io_private = private; | |
862 | zio->io_type = type; | |
863 | zio->io_priority = priority; | |
b128c09f BB |
864 | zio->io_vd = vd; |
865 | zio->io_offset = offset; | |
a6255b7f | 866 | zio->io_orig_abd = zio->io_abd = data; |
2aa34383 DK |
867 | zio->io_orig_size = zio->io_size = psize; |
868 | zio->io_lsize = lsize; | |
b128c09f BB |
869 | zio->io_orig_flags = zio->io_flags = flags; |
870 | zio->io_orig_stage = zio->io_stage = stage; | |
871 | zio->io_orig_pipeline = zio->io_pipeline = pipeline; | |
3dfb57a3 | 872 | zio->io_pipeline_trace = ZIO_STAGE_OPEN; |
34dc7c2f | 873 | |
d164b209 BB |
874 | zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); |
875 | zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); | |
876 | ||
b128c09f BB |
877 | if (zb != NULL) |
878 | zio->io_bookmark = *zb; | |
879 | ||
880 | if (pio != NULL) { | |
1b50749c | 881 | zio->io_metaslab_class = pio->io_metaslab_class; |
b128c09f | 882 | if (zio->io_logical == NULL) |
34dc7c2f | 883 | zio->io_logical = pio->io_logical; |
9babb374 BB |
884 | if (zio->io_child_type == ZIO_CHILD_GANG) |
885 | zio->io_gang_leader = pio->io_gang_leader; | |
b128c09f | 886 | zio_add_child(pio, zio); |
34dc7c2f BB |
887 | } |
888 | ||
a38718a6 GA |
889 | taskq_init_ent(&zio->io_tqent); |
890 | ||
34dc7c2f BB |
891 | return (zio); |
892 | } | |
893 | ||
894 | static void | |
b128c09f | 895 | zio_destroy(zio_t *zio) |
34dc7c2f | 896 | { |
4e21fd06 | 897 | metaslab_trace_fini(&zio->io_alloc_list); |
3941503c BB |
898 | list_destroy(&zio->io_parent_list); |
899 | list_destroy(&zio->io_child_list); | |
900 | mutex_destroy(&zio->io_lock); | |
901 | cv_destroy(&zio->io_cv); | |
b128c09f | 902 | kmem_cache_free(zio_cache, zio); |
34dc7c2f BB |
903 | } |
904 | ||
905 | zio_t * | |
d164b209 | 906 | zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, |
428870ff | 907 | void *private, enum zio_flag flags) |
34dc7c2f BB |
908 | { |
909 | zio_t *zio; | |
910 | ||
2aa34383 | 911 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
d164b209 | 912 | ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
b128c09f | 913 | ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); |
34dc7c2f BB |
914 | |
915 | return (zio); | |
916 | } | |
917 | ||
918 | zio_t * | |
428870ff | 919 | zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) |
34dc7c2f | 920 | { |
d164b209 | 921 | return (zio_null(NULL, spa, NULL, done, private, flags)); |
34dc7c2f BB |
922 | } |
923 | ||
bc67cba7 PZ |
924 | static int |
925 | zfs_blkptr_verify_log(spa_t *spa, const blkptr_t *bp, | |
926 | enum blk_verify_flag blk_verify, const char *fmt, ...) | |
927 | { | |
928 | va_list adx; | |
929 | char buf[256]; | |
930 | ||
931 | va_start(adx, fmt); | |
932 | (void) vsnprintf(buf, sizeof (buf), fmt, adx); | |
933 | va_end(adx); | |
934 | ||
935 | switch (blk_verify) { | |
936 | case BLK_VERIFY_HALT: | |
f49db9b5 | 937 | dprintf_bp(bp, "blkptr at %p dprintf_bp():", bp); |
bc67cba7 PZ |
938 | zfs_panic_recover("%s: %s", spa_name(spa), buf); |
939 | break; | |
940 | case BLK_VERIFY_LOG: | |
941 | zfs_dbgmsg("%s: %s", spa_name(spa), buf); | |
942 | break; | |
943 | case BLK_VERIFY_ONLY: | |
944 | break; | |
945 | } | |
946 | ||
947 | return (1); | |
948 | } | |
949 | ||
950 | /* | |
951 | * Verify the block pointer fields contain reasonable values. This means | |
952 | * it only contains known object types, checksum/compression identifiers, | |
953 | * block sizes within the maximum allowed limits, valid DVAs, etc. | |
954 | * | |
955 | * If everything checks out B_TRUE is returned. The zfs_blkptr_verify | |
956 | * argument controls the behavior when an invalid field is detected. | |
957 | * | |
958 | * Modes for zfs_blkptr_verify: | |
959 | * 1) BLK_VERIFY_ONLY (evaluate the block) | |
960 | * 2) BLK_VERIFY_LOG (evaluate the block and log problems) | |
961 | * 3) BLK_VERIFY_HALT (call zfs_panic_recover on error) | |
962 | */ | |
963 | boolean_t | |
964 | zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp, boolean_t config_held, | |
965 | enum blk_verify_flag blk_verify) | |
63e3a861 | 966 | { |
bc67cba7 PZ |
967 | int errors = 0; |
968 | ||
63e3a861 | 969 | if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { |
bc67cba7 PZ |
970 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
971 | "blkptr at %p has invalid TYPE %llu", | |
63e3a861 MA |
972 | bp, (longlong_t)BP_GET_TYPE(bp)); |
973 | } | |
974 | if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS || | |
975 | BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) { | |
bc67cba7 PZ |
976 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
977 | "blkptr at %p has invalid CHECKSUM %llu", | |
63e3a861 MA |
978 | bp, (longlong_t)BP_GET_CHECKSUM(bp)); |
979 | } | |
980 | if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS || | |
981 | BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) { | |
bc67cba7 PZ |
982 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
983 | "blkptr at %p has invalid COMPRESS %llu", | |
63e3a861 MA |
984 | bp, (longlong_t)BP_GET_COMPRESS(bp)); |
985 | } | |
986 | if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
bc67cba7 PZ |
987 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
988 | "blkptr at %p has invalid LSIZE %llu", | |
63e3a861 MA |
989 | bp, (longlong_t)BP_GET_LSIZE(bp)); |
990 | } | |
991 | if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
bc67cba7 PZ |
992 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
993 | "blkptr at %p has invalid PSIZE %llu", | |
63e3a861 MA |
994 | bp, (longlong_t)BP_GET_PSIZE(bp)); |
995 | } | |
996 | ||
997 | if (BP_IS_EMBEDDED(bp)) { | |
746d4a45 | 998 | if (BPE_GET_ETYPE(bp) >= NUM_BP_EMBEDDED_TYPES) { |
bc67cba7 PZ |
999 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
1000 | "blkptr at %p has invalid ETYPE %llu", | |
63e3a861 MA |
1001 | bp, (longlong_t)BPE_GET_ETYPE(bp)); |
1002 | } | |
1003 | } | |
1004 | ||
6cb8e530 PZ |
1005 | /* |
1006 | * Do not verify individual DVAs if the config is not trusted. This | |
1007 | * will be done once the zio is executed in vdev_mirror_map_alloc. | |
1008 | */ | |
1009 | if (!spa->spa_trust_config) | |
b9ec4a15 | 1010 | return (errors == 0); |
6cb8e530 | 1011 | |
dc04a8c7 PD |
1012 | if (!config_held) |
1013 | spa_config_enter(spa, SCL_VDEV, bp, RW_READER); | |
1014 | else | |
1015 | ASSERT(spa_config_held(spa, SCL_VDEV, RW_WRITER)); | |
63e3a861 MA |
1016 | /* |
1017 | * Pool-specific checks. | |
1018 | * | |
1019 | * Note: it would be nice to verify that the blk_birth and | |
1020 | * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() | |
1021 | * allows the birth time of log blocks (and dmu_sync()-ed blocks | |
1022 | * that are in the log) to be arbitrarily large. | |
1023 | */ | |
1c27024e | 1024 | for (int i = 0; i < BP_GET_NDVAS(bp); i++) { |
2b56a634 MA |
1025 | const dva_t *dva = &bp->blk_dva[i]; |
1026 | uint64_t vdevid = DVA_GET_VDEV(dva); | |
1c27024e | 1027 | |
63e3a861 | 1028 | if (vdevid >= spa->spa_root_vdev->vdev_children) { |
bc67cba7 PZ |
1029 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
1030 | "blkptr at %p DVA %u has invalid VDEV %llu", | |
63e3a861 | 1031 | bp, i, (longlong_t)vdevid); |
ee3a23b8 | 1032 | continue; |
63e3a861 | 1033 | } |
1c27024e | 1034 | vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; |
63e3a861 | 1035 | if (vd == NULL) { |
bc67cba7 PZ |
1036 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
1037 | "blkptr at %p DVA %u has invalid VDEV %llu", | |
63e3a861 | 1038 | bp, i, (longlong_t)vdevid); |
ee3a23b8 | 1039 | continue; |
63e3a861 MA |
1040 | } |
1041 | if (vd->vdev_ops == &vdev_hole_ops) { | |
bc67cba7 PZ |
1042 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
1043 | "blkptr at %p DVA %u has hole VDEV %llu", | |
63e3a861 | 1044 | bp, i, (longlong_t)vdevid); |
ee3a23b8 | 1045 | continue; |
63e3a861 MA |
1046 | } |
1047 | if (vd->vdev_ops == &vdev_missing_ops) { | |
1048 | /* | |
1049 | * "missing" vdevs are valid during import, but we | |
1050 | * don't have their detailed info (e.g. asize), so | |
1051 | * we can't perform any more checks on them. | |
1052 | */ | |
1053 | continue; | |
1054 | } | |
2b56a634 MA |
1055 | uint64_t offset = DVA_GET_OFFSET(dva); |
1056 | uint64_t asize = DVA_GET_ASIZE(dva); | |
1057 | if (DVA_GET_GANG(dva)) | |
1058 | asize = vdev_gang_header_asize(vd); | |
63e3a861 | 1059 | if (offset + asize > vd->vdev_asize) { |
bc67cba7 PZ |
1060 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
1061 | "blkptr at %p DVA %u has invalid OFFSET %llu", | |
63e3a861 MA |
1062 | bp, i, (longlong_t)offset); |
1063 | } | |
1064 | } | |
f49db9b5 BB |
1065 | if (errors > 0) |
1066 | dprintf_bp(bp, "blkptr at %p dprintf_bp():", bp); | |
dc04a8c7 PD |
1067 | if (!config_held) |
1068 | spa_config_exit(spa, SCL_VDEV, bp); | |
bc67cba7 PZ |
1069 | |
1070 | return (errors == 0); | |
63e3a861 MA |
1071 | } |
1072 | ||
6cb8e530 PZ |
1073 | boolean_t |
1074 | zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp) | |
1075 | { | |
14e4e3cb | 1076 | (void) bp; |
6cb8e530 PZ |
1077 | uint64_t vdevid = DVA_GET_VDEV(dva); |
1078 | ||
1079 | if (vdevid >= spa->spa_root_vdev->vdev_children) | |
1080 | return (B_FALSE); | |
1081 | ||
1082 | vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; | |
1083 | if (vd == NULL) | |
1084 | return (B_FALSE); | |
1085 | ||
1086 | if (vd->vdev_ops == &vdev_hole_ops) | |
1087 | return (B_FALSE); | |
1088 | ||
1089 | if (vd->vdev_ops == &vdev_missing_ops) { | |
1090 | return (B_FALSE); | |
1091 | } | |
1092 | ||
1093 | uint64_t offset = DVA_GET_OFFSET(dva); | |
1094 | uint64_t asize = DVA_GET_ASIZE(dva); | |
1095 | ||
2b56a634 MA |
1096 | if (DVA_GET_GANG(dva)) |
1097 | asize = vdev_gang_header_asize(vd); | |
6cb8e530 PZ |
1098 | if (offset + asize > vd->vdev_asize) |
1099 | return (B_FALSE); | |
1100 | ||
1101 | return (B_TRUE); | |
1102 | } | |
1103 | ||
34dc7c2f | 1104 | zio_t * |
b128c09f | 1105 | zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, |
a6255b7f | 1106 | abd_t *data, uint64_t size, zio_done_func_t *done, void *private, |
5dbd68a3 | 1107 | zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) |
34dc7c2f BB |
1108 | { |
1109 | zio_t *zio; | |
1110 | ||
428870ff | 1111 | zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, |
2aa34383 | 1112 | data, size, size, done, private, |
b128c09f | 1113 | ZIO_TYPE_READ, priority, flags, NULL, 0, zb, |
428870ff BB |
1114 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
1115 | ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); | |
34dc7c2f | 1116 | |
b128c09f BB |
1117 | return (zio); |
1118 | } | |
34dc7c2f | 1119 | |
34dc7c2f | 1120 | zio_t * |
b128c09f | 1121 | zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, |
a6255b7f | 1122 | abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp, |
bc77ba73 PD |
1123 | zio_done_func_t *ready, zio_done_func_t *children_ready, |
1124 | zio_done_func_t *physdone, zio_done_func_t *done, | |
1125 | void *private, zio_priority_t priority, enum zio_flag flags, | |
1126 | const zbookmark_phys_t *zb) | |
34dc7c2f BB |
1127 | { |
1128 | zio_t *zio; | |
1129 | ||
b128c09f BB |
1130 | ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && |
1131 | zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && | |
1132 | zp->zp_compress >= ZIO_COMPRESS_OFF && | |
1133 | zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && | |
9ae529ec | 1134 | DMU_OT_IS_VALID(zp->zp_type) && |
b128c09f | 1135 | zp->zp_level < 32 && |
428870ff | 1136 | zp->zp_copies > 0 && |
03c6040b | 1137 | zp->zp_copies <= spa_max_replication(spa)); |
34dc7c2f | 1138 | |
2aa34383 | 1139 | zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private, |
b128c09f | 1140 | ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, |
428870ff BB |
1141 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
1142 | ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); | |
34dc7c2f BB |
1143 | |
1144 | zio->io_ready = ready; | |
bc77ba73 | 1145 | zio->io_children_ready = children_ready; |
e8b96c60 | 1146 | zio->io_physdone = physdone; |
b128c09f | 1147 | zio->io_prop = *zp; |
34dc7c2f | 1148 | |
9b67f605 MA |
1149 | /* |
1150 | * Data can be NULL if we are going to call zio_write_override() to | |
1151 | * provide the already-allocated BP. But we may need the data to | |
1152 | * verify a dedup hit (if requested). In this case, don't try to | |
b5256303 TC |
1153 | * dedup (just take the already-allocated BP verbatim). Encrypted |
1154 | * dedup blocks need data as well so we also disable dedup in this | |
1155 | * case. | |
9b67f605 | 1156 | */ |
b5256303 TC |
1157 | if (data == NULL && |
1158 | (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) { | |
9b67f605 MA |
1159 | zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; |
1160 | } | |
1161 | ||
34dc7c2f BB |
1162 | return (zio); |
1163 | } | |
1164 | ||
1165 | zio_t * | |
a6255b7f | 1166 | zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data, |
e8b96c60 | 1167 | uint64_t size, zio_done_func_t *done, void *private, |
5dbd68a3 | 1168 | zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) |
34dc7c2f BB |
1169 | { |
1170 | zio_t *zio; | |
1171 | ||
2aa34383 | 1172 | zio = zio_create(pio, spa, txg, bp, data, size, size, done, private, |
3dfb57a3 | 1173 | ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb, |
b128c09f | 1174 | ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); |
34dc7c2f BB |
1175 | |
1176 | return (zio); | |
1177 | } | |
1178 | ||
428870ff | 1179 | void |
03c6040b | 1180 | zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) |
428870ff BB |
1181 | { |
1182 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
1183 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1184 | ASSERT(zio->io_stage == ZIO_STAGE_OPEN); | |
1185 | ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); | |
1186 | ||
03c6040b GW |
1187 | /* |
1188 | * We must reset the io_prop to match the values that existed | |
1189 | * when the bp was first written by dmu_sync() keeping in mind | |
1190 | * that nopwrite and dedup are mutually exclusive. | |
1191 | */ | |
1192 | zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; | |
1193 | zio->io_prop.zp_nopwrite = nopwrite; | |
428870ff BB |
1194 | zio->io_prop.zp_copies = copies; |
1195 | zio->io_bp_override = bp; | |
1196 | } | |
1197 | ||
1198 | void | |
1199 | zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) | |
1200 | { | |
9b67f605 | 1201 | |
bc67cba7 | 1202 | (void) zfs_blkptr_verify(spa, bp, B_FALSE, BLK_VERIFY_HALT); |
a1d477c2 | 1203 | |
9b67f605 MA |
1204 | /* |
1205 | * The check for EMBEDDED is a performance optimization. We | |
1206 | * process the free here (by ignoring it) rather than | |
1207 | * putting it on the list and then processing it in zio_free_sync(). | |
1208 | */ | |
1209 | if (BP_IS_EMBEDDED(bp)) | |
1210 | return; | |
13fe0198 | 1211 | metaslab_check_free(spa, bp); |
2883cad5 MA |
1212 | |
1213 | /* | |
1214 | * Frees that are for the currently-syncing txg, are not going to be | |
1215 | * deferred, and which will not need to do a read (i.e. not GANG or | |
1216 | * DEDUP), can be processed immediately. Otherwise, put them on the | |
1217 | * in-memory list for later processing. | |
93e28d66 SD |
1218 | * |
1219 | * Note that we only defer frees after zfs_sync_pass_deferred_free | |
1220 | * when the log space map feature is disabled. [see relevant comment | |
1221 | * in spa_sync_iterate_to_convergence()] | |
2883cad5 | 1222 | */ |
93e28d66 SD |
1223 | if (BP_IS_GANG(bp) || |
1224 | BP_GET_DEDUP(bp) || | |
2883cad5 | 1225 | txg != spa->spa_syncing_txg || |
93e28d66 SD |
1226 | (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free && |
1227 | !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))) { | |
2883cad5 MA |
1228 | bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); |
1229 | } else { | |
9cdf7b1f | 1230 | VERIFY3P(zio_free_sync(NULL, spa, txg, bp, 0), ==, NULL); |
2883cad5 | 1231 | } |
428870ff BB |
1232 | } |
1233 | ||
9cdf7b1f MA |
1234 | /* |
1235 | * To improve performance, this function may return NULL if we were able | |
1236 | * to do the free immediately. This avoids the cost of creating a zio | |
1237 | * (and linking it to the parent, etc). | |
1238 | */ | |
34dc7c2f | 1239 | zio_t * |
428870ff BB |
1240 | zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
1241 | enum zio_flag flags) | |
34dc7c2f | 1242 | { |
428870ff BB |
1243 | ASSERT(!BP_IS_HOLE(bp)); |
1244 | ASSERT(spa_syncing_txg(spa) == txg); | |
34dc7c2f | 1245 | |
9b67f605 | 1246 | if (BP_IS_EMBEDDED(bp)) |
9cdf7b1f | 1247 | return (NULL); |
9b67f605 | 1248 | |
13fe0198 | 1249 | metaslab_check_free(spa, bp); |
8c841793 | 1250 | arc_freed(spa, bp); |
d4a72f23 | 1251 | dsl_scan_freed(spa, bp); |
13fe0198 | 1252 | |
9cdf7b1f MA |
1253 | if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) { |
1254 | /* | |
1255 | * GANG and DEDUP blocks can induce a read (for the gang block | |
1256 | * header, or the DDT), so issue them asynchronously so that | |
1257 | * this thread is not tied up. | |
1258 | */ | |
1259 | enum zio_stage stage = | |
1260 | ZIO_FREE_PIPELINE | ZIO_STAGE_ISSUE_ASYNC; | |
2883cad5 | 1261 | |
9cdf7b1f MA |
1262 | return (zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
1263 | BP_GET_PSIZE(bp), NULL, NULL, | |
1264 | ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, | |
1265 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage)); | |
1266 | } else { | |
1267 | metaslab_free(spa, bp, txg, B_FALSE); | |
1268 | return (NULL); | |
1269 | } | |
34dc7c2f BB |
1270 | } |
1271 | ||
1272 | zio_t * | |
428870ff BB |
1273 | zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
1274 | zio_done_func_t *done, void *private, enum zio_flag flags) | |
34dc7c2f BB |
1275 | { |
1276 | zio_t *zio; | |
1277 | ||
bc67cba7 PZ |
1278 | (void) zfs_blkptr_verify(spa, bp, flags & ZIO_FLAG_CONFIG_WRITER, |
1279 | BLK_VERIFY_HALT); | |
9b67f605 MA |
1280 | |
1281 | if (BP_IS_EMBEDDED(bp)) | |
1282 | return (zio_null(pio, spa, NULL, NULL, NULL, 0)); | |
1283 | ||
34dc7c2f BB |
1284 | /* |
1285 | * A claim is an allocation of a specific block. Claims are needed | |
1286 | * to support immediate writes in the intent log. The issue is that | |
1287 | * immediate writes contain committed data, but in a txg that was | |
1288 | * *not* committed. Upon opening the pool after an unclean shutdown, | |
1289 | * the intent log claims all blocks that contain immediate write data | |
1290 | * so that the SPA knows they're in use. | |
1291 | * | |
1292 | * All claims *must* be resolved in the first txg -- before the SPA | |
1293 | * starts allocating blocks -- so that nothing is allocated twice. | |
428870ff | 1294 | * If txg == 0 we just verify that the block is claimable. |
34dc7c2f | 1295 | */ |
d2734cce SD |
1296 | ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, |
1297 | spa_min_claim_txg(spa)); | |
1298 | ASSERT(txg == spa_min_claim_txg(spa) || txg == 0); | |
76d04993 | 1299 | ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(8) */ |
34dc7c2f | 1300 | |
b128c09f | 1301 | zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
2aa34383 DK |
1302 | BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, |
1303 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); | |
3dfb57a3 | 1304 | ASSERT0(zio->io_queued_timestamp); |
34dc7c2f BB |
1305 | |
1306 | return (zio); | |
1307 | } | |
1308 | ||
1309 | zio_t * | |
1310 | zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, | |
e8b96c60 | 1311 | zio_done_func_t *done, void *private, enum zio_flag flags) |
34dc7c2f BB |
1312 | { |
1313 | zio_t *zio; | |
1314 | int c; | |
1315 | ||
1316 | if (vd->vdev_children == 0) { | |
2aa34383 | 1317 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
e8b96c60 | 1318 | ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
34dc7c2f BB |
1319 | ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); |
1320 | ||
34dc7c2f BB |
1321 | zio->io_cmd = cmd; |
1322 | } else { | |
d164b209 | 1323 | zio = zio_null(pio, spa, NULL, NULL, NULL, flags); |
34dc7c2f BB |
1324 | |
1325 | for (c = 0; c < vd->vdev_children; c++) | |
1326 | zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, | |
e8b96c60 | 1327 | done, private, flags)); |
34dc7c2f BB |
1328 | } |
1329 | ||
1330 | return (zio); | |
1331 | } | |
1332 | ||
1b939560 BB |
1333 | zio_t * |
1334 | zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
1335 | zio_done_func_t *done, void *private, zio_priority_t priority, | |
1336 | enum zio_flag flags, enum trim_flag trim_flags) | |
1337 | { | |
1338 | zio_t *zio; | |
1339 | ||
1340 | ASSERT0(vd->vdev_children); | |
1341 | ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift)); | |
1342 | ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift)); | |
1343 | ASSERT3U(size, !=, 0); | |
1344 | ||
1345 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done, | |
1346 | private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL, | |
1347 | vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE); | |
1348 | zio->io_trim_flags = trim_flags; | |
1349 | ||
1350 | return (zio); | |
1351 | } | |
1352 | ||
34dc7c2f BB |
1353 | zio_t * |
1354 | zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
a6255b7f | 1355 | abd_t *data, int checksum, zio_done_func_t *done, void *private, |
e8b96c60 | 1356 | zio_priority_t priority, enum zio_flag flags, boolean_t labels) |
34dc7c2f BB |
1357 | { |
1358 | zio_t *zio; | |
34dc7c2f | 1359 | |
b128c09f BB |
1360 | ASSERT(vd->vdev_children == 0); |
1361 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
1362 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
1363 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 1364 | |
2aa34383 DK |
1365 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
1366 | private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
1367 | offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); | |
34dc7c2f | 1368 | |
b128c09f | 1369 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f BB |
1370 | |
1371 | return (zio); | |
1372 | } | |
1373 | ||
1374 | zio_t * | |
1375 | zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
a6255b7f | 1376 | abd_t *data, int checksum, zio_done_func_t *done, void *private, |
e8b96c60 | 1377 | zio_priority_t priority, enum zio_flag flags, boolean_t labels) |
34dc7c2f | 1378 | { |
34dc7c2f | 1379 | zio_t *zio; |
34dc7c2f | 1380 | |
b128c09f BB |
1381 | ASSERT(vd->vdev_children == 0); |
1382 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
1383 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
1384 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 1385 | |
2aa34383 DK |
1386 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
1387 | private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
1388 | offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); | |
34dc7c2f | 1389 | |
b128c09f | 1390 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f | 1391 | |
3c67d83a | 1392 | if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { |
34dc7c2f | 1393 | /* |
428870ff | 1394 | * zec checksums are necessarily destructive -- they modify |
b128c09f | 1395 | * the end of the write buffer to hold the verifier/checksum. |
34dc7c2f | 1396 | * Therefore, we must make a local copy in case the data is |
b128c09f | 1397 | * being written to multiple places in parallel. |
34dc7c2f | 1398 | */ |
a6255b7f DQ |
1399 | abd_t *wbuf = abd_alloc_sametype(data, size); |
1400 | abd_copy(wbuf, data, size); | |
1401 | ||
b128c09f | 1402 | zio_push_transform(zio, wbuf, size, size, NULL); |
34dc7c2f BB |
1403 | } |
1404 | ||
1405 | return (zio); | |
1406 | } | |
1407 | ||
1408 | /* | |
b128c09f | 1409 | * Create a child I/O to do some work for us. |
34dc7c2f BB |
1410 | */ |
1411 | zio_t * | |
b128c09f | 1412 | zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, |
4ea3f864 GM |
1413 | abd_t *data, uint64_t size, int type, zio_priority_t priority, |
1414 | enum zio_flag flags, zio_done_func_t *done, void *private) | |
34dc7c2f | 1415 | { |
428870ff | 1416 | enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; |
b128c09f BB |
1417 | zio_t *zio; |
1418 | ||
a1d477c2 MA |
1419 | /* |
1420 | * vdev child I/Os do not propagate their error to the parent. | |
1421 | * Therefore, for correct operation the caller *must* check for | |
1422 | * and handle the error in the child i/o's done callback. | |
1423 | * The only exceptions are i/os that we don't care about | |
1424 | * (OPTIONAL or REPAIR). | |
1425 | */ | |
1426 | ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) || | |
1427 | done != NULL); | |
1428 | ||
34dc7c2f BB |
1429 | if (type == ZIO_TYPE_READ && bp != NULL) { |
1430 | /* | |
1431 | * If we have the bp, then the child should perform the | |
1432 | * checksum and the parent need not. This pushes error | |
1433 | * detection as close to the leaves as possible and | |
1434 | * eliminates redundant checksums in the interior nodes. | |
1435 | */ | |
428870ff BB |
1436 | pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; |
1437 | pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; | |
34dc7c2f BB |
1438 | } |
1439 | ||
a1d477c2 MA |
1440 | if (vd->vdev_ops->vdev_op_leaf) { |
1441 | ASSERT0(vd->vdev_children); | |
b128c09f | 1442 | offset += VDEV_LABEL_START_SIZE; |
a1d477c2 | 1443 | } |
b128c09f | 1444 | |
a1d477c2 | 1445 | flags |= ZIO_VDEV_CHILD_FLAGS(pio); |
428870ff BB |
1446 | |
1447 | /* | |
1448 | * If we've decided to do a repair, the write is not speculative -- | |
1449 | * even if the original read was. | |
1450 | */ | |
1451 | if (flags & ZIO_FLAG_IO_REPAIR) | |
1452 | flags &= ~ZIO_FLAG_SPECULATIVE; | |
1453 | ||
3dfb57a3 DB |
1454 | /* |
1455 | * If we're creating a child I/O that is not associated with a | |
1456 | * top-level vdev, then the child zio is not an allocating I/O. | |
1457 | * If this is a retried I/O then we ignore it since we will | |
1458 | * have already processed the original allocating I/O. | |
1459 | */ | |
1460 | if (flags & ZIO_FLAG_IO_ALLOCATING && | |
1461 | (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) { | |
cc99f275 DB |
1462 | ASSERT(pio->io_metaslab_class != NULL); |
1463 | ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled); | |
3dfb57a3 DB |
1464 | ASSERT(type == ZIO_TYPE_WRITE); |
1465 | ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE); | |
1466 | ASSERT(!(flags & ZIO_FLAG_IO_REPAIR)); | |
1467 | ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) || | |
1468 | pio->io_child_type == ZIO_CHILD_GANG); | |
1469 | ||
1470 | flags &= ~ZIO_FLAG_IO_ALLOCATING; | |
1471 | } | |
1472 | ||
1473 | ||
2aa34383 | 1474 | zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size, |
428870ff BB |
1475 | done, private, type, priority, flags, vd, offset, &pio->io_bookmark, |
1476 | ZIO_STAGE_VDEV_IO_START >> 1, pipeline); | |
3dfb57a3 | 1477 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); |
34dc7c2f | 1478 | |
e8b96c60 MA |
1479 | zio->io_physdone = pio->io_physdone; |
1480 | if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) | |
1481 | zio->io_logical->io_phys_children++; | |
1482 | ||
b128c09f | 1483 | return (zio); |
34dc7c2f BB |
1484 | } |
1485 | ||
b128c09f | 1486 | zio_t * |
a6255b7f | 1487 | zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size, |
9e052db4 | 1488 | zio_type_t type, zio_priority_t priority, enum zio_flag flags, |
e9aa730c | 1489 | zio_done_func_t *done, void *private) |
34dc7c2f | 1490 | { |
b128c09f | 1491 | zio_t *zio; |
34dc7c2f | 1492 | |
b128c09f | 1493 | ASSERT(vd->vdev_ops->vdev_op_leaf); |
34dc7c2f | 1494 | |
b128c09f | 1495 | zio = zio_create(NULL, vd->vdev_spa, 0, NULL, |
2aa34383 | 1496 | data, size, size, done, private, type, priority, |
e8b96c60 | 1497 | flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, |
b128c09f | 1498 | vd, offset, NULL, |
428870ff | 1499 | ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); |
34dc7c2f | 1500 | |
b128c09f | 1501 | return (zio); |
34dc7c2f BB |
1502 | } |
1503 | ||
1504 | void | |
b128c09f | 1505 | zio_flush(zio_t *zio, vdev_t *vd) |
34dc7c2f | 1506 | { |
b128c09f | 1507 | zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, |
e8b96c60 | 1508 | NULL, NULL, |
b128c09f | 1509 | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); |
34dc7c2f BB |
1510 | } |
1511 | ||
428870ff BB |
1512 | void |
1513 | zio_shrink(zio_t *zio, uint64_t size) | |
1514 | { | |
1ce23dca PS |
1515 | ASSERT3P(zio->io_executor, ==, NULL); |
1516 | ASSERT3U(zio->io_orig_size, ==, zio->io_size); | |
1517 | ASSERT3U(size, <=, zio->io_size); | |
428870ff BB |
1518 | |
1519 | /* | |
1520 | * We don't shrink for raidz because of problems with the | |
1521 | * reconstruction when reading back less than the block size. | |
1522 | * Note, BP_IS_RAIDZ() assumes no compression. | |
1523 | */ | |
1524 | ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); | |
2aa34383 DK |
1525 | if (!BP_IS_RAIDZ(zio->io_bp)) { |
1526 | /* we are not doing a raw write */ | |
1527 | ASSERT3U(zio->io_size, ==, zio->io_lsize); | |
1528 | zio->io_orig_size = zio->io_size = zio->io_lsize = size; | |
1529 | } | |
428870ff BB |
1530 | } |
1531 | ||
34dc7c2f BB |
1532 | /* |
1533 | * ========================================================================== | |
b128c09f | 1534 | * Prepare to read and write logical blocks |
34dc7c2f BB |
1535 | * ========================================================================== |
1536 | */ | |
b128c09f | 1537 | |
62840030 | 1538 | static zio_t * |
b128c09f | 1539 | zio_read_bp_init(zio_t *zio) |
34dc7c2f | 1540 | { |
b128c09f | 1541 | blkptr_t *bp = zio->io_bp; |
b5256303 TC |
1542 | uint64_t psize = |
1543 | BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); | |
34dc7c2f | 1544 | |
a1d477c2 MA |
1545 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
1546 | ||
fb5f0bc8 | 1547 | if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && |
9babb374 | 1548 | zio->io_child_type == ZIO_CHILD_LOGICAL && |
b5256303 | 1549 | !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { |
a6255b7f DQ |
1550 | zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), |
1551 | psize, psize, zio_decompress); | |
34dc7c2f | 1552 | } |
34dc7c2f | 1553 | |
b5256303 TC |
1554 | if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) || |
1555 | BP_HAS_INDIRECT_MAC_CKSUM(bp)) && | |
1556 | zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
1557 | zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), | |
1558 | psize, psize, zio_decrypt); | |
1559 | } | |
1560 | ||
9b67f605 | 1561 | if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { |
a6255b7f DQ |
1562 | int psize = BPE_GET_PSIZE(bp); |
1563 | void *data = abd_borrow_buf(zio->io_abd, psize); | |
1564 | ||
9b67f605 | 1565 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
a6255b7f DQ |
1566 | decode_embedded_bp_compressed(bp, data); |
1567 | abd_return_buf_copy(zio->io_abd, data, psize); | |
9b67f605 MA |
1568 | } else { |
1569 | ASSERT(!BP_IS_EMBEDDED(bp)); | |
a1d477c2 | 1570 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
9b67f605 MA |
1571 | } |
1572 | ||
9ae529ec | 1573 | if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) |
b128c09f BB |
1574 | zio->io_flags |= ZIO_FLAG_DONT_CACHE; |
1575 | ||
428870ff BB |
1576 | if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) |
1577 | zio->io_flags |= ZIO_FLAG_DONT_CACHE; | |
1578 | ||
1579 | if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) | |
1580 | zio->io_pipeline = ZIO_DDT_READ_PIPELINE; | |
1581 | ||
62840030 | 1582 | return (zio); |
34dc7c2f BB |
1583 | } |
1584 | ||
62840030 | 1585 | static zio_t * |
b128c09f | 1586 | zio_write_bp_init(zio_t *zio) |
34dc7c2f | 1587 | { |
b128c09f | 1588 | if (!IO_IS_ALLOCATING(zio)) |
62840030 | 1589 | return (zio); |
34dc7c2f | 1590 | |
428870ff BB |
1591 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1592 | ||
1593 | if (zio->io_bp_override) { | |
3dfb57a3 DB |
1594 | blkptr_t *bp = zio->io_bp; |
1595 | zio_prop_t *zp = &zio->io_prop; | |
1596 | ||
428870ff BB |
1597 | ASSERT(bp->blk_birth != zio->io_txg); |
1598 | ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); | |
1599 | ||
1600 | *bp = *zio->io_bp_override; | |
1601 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1602 | ||
9b67f605 | 1603 | if (BP_IS_EMBEDDED(bp)) |
62840030 | 1604 | return (zio); |
9b67f605 | 1605 | |
03c6040b GW |
1606 | /* |
1607 | * If we've been overridden and nopwrite is set then | |
1608 | * set the flag accordingly to indicate that a nopwrite | |
1609 | * has already occurred. | |
1610 | */ | |
1611 | if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { | |
1612 | ASSERT(!zp->zp_dedup); | |
3dfb57a3 | 1613 | ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum); |
03c6040b | 1614 | zio->io_flags |= ZIO_FLAG_NOPWRITE; |
62840030 | 1615 | return (zio); |
03c6040b GW |
1616 | } |
1617 | ||
1618 | ASSERT(!zp->zp_nopwrite); | |
1619 | ||
428870ff | 1620 | if (BP_IS_HOLE(bp) || !zp->zp_dedup) |
62840030 | 1621 | return (zio); |
428870ff | 1622 | |
3c67d83a TH |
1623 | ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & |
1624 | ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); | |
428870ff | 1625 | |
b5256303 TC |
1626 | if (BP_GET_CHECKSUM(bp) == zp->zp_checksum && |
1627 | !zp->zp_encrypt) { | |
428870ff BB |
1628 | BP_SET_DEDUP(bp, 1); |
1629 | zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; | |
62840030 | 1630 | return (zio); |
428870ff | 1631 | } |
3dfb57a3 DB |
1632 | |
1633 | /* | |
1634 | * We were unable to handle this as an override bp, treat | |
1635 | * it as a regular write I/O. | |
1636 | */ | |
5511754b | 1637 | zio->io_bp_override = NULL; |
3dfb57a3 DB |
1638 | *bp = zio->io_bp_orig; |
1639 | zio->io_pipeline = zio->io_orig_pipeline; | |
1640 | } | |
1641 | ||
62840030 | 1642 | return (zio); |
3dfb57a3 DB |
1643 | } |
1644 | ||
62840030 | 1645 | static zio_t * |
3dfb57a3 DB |
1646 | zio_write_compress(zio_t *zio) |
1647 | { | |
1648 | spa_t *spa = zio->io_spa; | |
1649 | zio_prop_t *zp = &zio->io_prop; | |
1650 | enum zio_compress compress = zp->zp_compress; | |
1651 | blkptr_t *bp = zio->io_bp; | |
1652 | uint64_t lsize = zio->io_lsize; | |
1653 | uint64_t psize = zio->io_size; | |
1654 | int pass = 1; | |
1655 | ||
3dfb57a3 DB |
1656 | /* |
1657 | * If our children haven't all reached the ready stage, | |
1658 | * wait for them and then repeat this pipeline stage. | |
1659 | */ | |
ddc751d5 GW |
1660 | if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT | |
1661 | ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) { | |
62840030 | 1662 | return (NULL); |
ddc751d5 | 1663 | } |
3dfb57a3 DB |
1664 | |
1665 | if (!IO_IS_ALLOCATING(zio)) | |
62840030 | 1666 | return (zio); |
3dfb57a3 DB |
1667 | |
1668 | if (zio->io_children_ready != NULL) { | |
1669 | /* | |
1670 | * Now that all our children are ready, run the callback | |
1671 | * associated with this zio in case it wants to modify the | |
1672 | * data to be written. | |
1673 | */ | |
1674 | ASSERT3U(zp->zp_level, >, 0); | |
1675 | zio->io_children_ready(zio); | |
428870ff | 1676 | } |
34dc7c2f | 1677 | |
3dfb57a3 DB |
1678 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1679 | ASSERT(zio->io_bp_override == NULL); | |
1680 | ||
b0bc7a84 | 1681 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { |
b128c09f BB |
1682 | /* |
1683 | * We're rewriting an existing block, which means we're | |
1684 | * working on behalf of spa_sync(). For spa_sync() to | |
1685 | * converge, it must eventually be the case that we don't | |
1686 | * have to allocate new blocks. But compression changes | |
1687 | * the blocksize, which forces a reallocate, and makes | |
1688 | * convergence take longer. Therefore, after the first | |
1689 | * few passes, stop compressing to ensure convergence. | |
1690 | */ | |
428870ff BB |
1691 | pass = spa_sync_pass(spa); |
1692 | ||
1693 | ASSERT(zio->io_txg == spa_syncing_txg(spa)); | |
1694 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1695 | ASSERT(!BP_GET_DEDUP(bp)); | |
34dc7c2f | 1696 | |
55d85d5a | 1697 | if (pass >= zfs_sync_pass_dont_compress) |
b128c09f | 1698 | compress = ZIO_COMPRESS_OFF; |
34dc7c2f | 1699 | |
b128c09f | 1700 | /* Make sure someone doesn't change their mind on overwrites */ |
9b67f605 | 1701 | ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), |
428870ff | 1702 | spa_max_replication(spa)) == BP_GET_NDVAS(bp)); |
b128c09f | 1703 | } |
34dc7c2f | 1704 | |
2aa34383 | 1705 | /* If it's a compressed write that is not raw, compress the buffer. */ |
b5256303 TC |
1706 | if (compress != ZIO_COMPRESS_OFF && |
1707 | !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { | |
428870ff | 1708 | void *cbuf = zio_buf_alloc(lsize); |
10b3c7f5 MN |
1709 | psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize, |
1710 | zp->zp_complevel); | |
1711 | if (psize == 0 || psize >= lsize) { | |
b128c09f | 1712 | compress = ZIO_COMPRESS_OFF; |
428870ff | 1713 | zio_buf_free(cbuf, lsize); |
b5256303 TC |
1714 | } else if (!zp->zp_dedup && !zp->zp_encrypt && |
1715 | psize <= BPE_PAYLOAD_SIZE && | |
9b67f605 MA |
1716 | zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && |
1717 | spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { | |
1718 | encode_embedded_bp_compressed(bp, | |
1719 | cbuf, compress, lsize, psize); | |
1720 | BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); | |
1721 | BP_SET_TYPE(bp, zio->io_prop.zp_type); | |
1722 | BP_SET_LEVEL(bp, zio->io_prop.zp_level); | |
1723 | zio_buf_free(cbuf, lsize); | |
1724 | bp->blk_birth = zio->io_txg; | |
1725 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1726 | ASSERT(spa_feature_is_active(spa, | |
1727 | SPA_FEATURE_EMBEDDED_DATA)); | |
62840030 | 1728 | return (zio); |
428870ff | 1729 | } else { |
9b67f605 | 1730 | /* |
b2255edc BB |
1731 | * Round compressed size up to the minimum allocation |
1732 | * size of the smallest-ashift device, and zero the | |
1733 | * tail. This ensures that the compressed size of the | |
1734 | * BP (and thus compressratio property) are correct, | |
c3520e7f MA |
1735 | * in that we charge for the padding used to fill out |
1736 | * the last sector. | |
9b67f605 | 1737 | */ |
b2255edc BB |
1738 | ASSERT3U(spa->spa_min_alloc, >=, SPA_MINBLOCKSHIFT); |
1739 | size_t rounded = (size_t)roundup(psize, | |
1740 | spa->spa_min_alloc); | |
c3520e7f | 1741 | if (rounded >= lsize) { |
9b67f605 MA |
1742 | compress = ZIO_COMPRESS_OFF; |
1743 | zio_buf_free(cbuf, lsize); | |
c3520e7f | 1744 | psize = lsize; |
9b67f605 | 1745 | } else { |
a6255b7f DQ |
1746 | abd_t *cdata = abd_get_from_buf(cbuf, lsize); |
1747 | abd_take_ownership_of_buf(cdata, B_TRUE); | |
1748 | abd_zero_off(cdata, psize, rounded - psize); | |
c3520e7f | 1749 | psize = rounded; |
a6255b7f | 1750 | zio_push_transform(zio, cdata, |
9b67f605 MA |
1751 | psize, lsize, NULL); |
1752 | } | |
b128c09f | 1753 | } |
3dfb57a3 DB |
1754 | |
1755 | /* | |
1756 | * We were unable to handle this as an override bp, treat | |
1757 | * it as a regular write I/O. | |
1758 | */ | |
1759 | zio->io_bp_override = NULL; | |
1760 | *bp = zio->io_bp_orig; | |
1761 | zio->io_pipeline = zio->io_orig_pipeline; | |
1762 | ||
b1d21733 TC |
1763 | } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 && |
1764 | zp->zp_type == DMU_OT_DNODE) { | |
1765 | /* | |
1766 | * The DMU actually relies on the zio layer's compression | |
1767 | * to free metadnode blocks that have had all contained | |
1768 | * dnodes freed. As a result, even when doing a raw | |
1769 | * receive, we must check whether the block can be compressed | |
1770 | * to a hole. | |
1771 | */ | |
1772 | psize = zio_compress_data(ZIO_COMPRESS_EMPTY, | |
10b3c7f5 MN |
1773 | zio->io_abd, NULL, lsize, zp->zp_complevel); |
1774 | if (psize == 0 || psize >= lsize) | |
b1d21733 | 1775 | compress = ZIO_COMPRESS_OFF; |
52a36bd4 GA |
1776 | } else if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS && |
1777 | !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) { | |
1778 | /* | |
1779 | * If we are raw receiving an encrypted dataset we should not | |
1780 | * take this codepath because it will change the on-disk block | |
1781 | * and decryption will fail. | |
1782 | */ | |
c634320e PD |
1783 | size_t rounded = MIN((size_t)roundup(psize, |
1784 | spa->spa_min_alloc), lsize); | |
1785 | ||
1786 | if (rounded != psize) { | |
1787 | abd_t *cdata = abd_alloc_linear(rounded, B_TRUE); | |
1788 | abd_zero_off(cdata, psize, rounded - psize); | |
1789 | abd_copy_off(cdata, zio->io_abd, 0, 0, psize); | |
1790 | psize = rounded; | |
1791 | zio_push_transform(zio, cdata, | |
1792 | psize, rounded, NULL); | |
1793 | } | |
2aa34383 DK |
1794 | } else { |
1795 | ASSERT3U(psize, !=, 0); | |
b128c09f | 1796 | } |
34dc7c2f | 1797 | |
b128c09f BB |
1798 | /* |
1799 | * The final pass of spa_sync() must be all rewrites, but the first | |
1800 | * few passes offer a trade-off: allocating blocks defers convergence, | |
1801 | * but newly allocated blocks are sequential, so they can be written | |
1802 | * to disk faster. Therefore, we allow the first few passes of | |
1803 | * spa_sync() to allocate new blocks, but force rewrites after that. | |
1804 | * There should only be a handful of blocks after pass 1 in any case. | |
1805 | */ | |
b0bc7a84 MG |
1806 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && |
1807 | BP_GET_PSIZE(bp) == psize && | |
55d85d5a | 1808 | pass >= zfs_sync_pass_rewrite) { |
cc99f275 | 1809 | VERIFY3U(psize, !=, 0); |
1c27024e | 1810 | enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; |
cc99f275 | 1811 | |
b128c09f BB |
1812 | zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; |
1813 | zio->io_flags |= ZIO_FLAG_IO_REWRITE; | |
1814 | } else { | |
1815 | BP_ZERO(bp); | |
1816 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
1817 | } | |
34dc7c2f | 1818 | |
428870ff | 1819 | if (psize == 0) { |
b0bc7a84 MG |
1820 | if (zio->io_bp_orig.blk_birth != 0 && |
1821 | spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { | |
1822 | BP_SET_LSIZE(bp, lsize); | |
1823 | BP_SET_TYPE(bp, zp->zp_type); | |
1824 | BP_SET_LEVEL(bp, zp->zp_level); | |
1825 | BP_SET_BIRTH(bp, zio->io_txg, 0); | |
1826 | } | |
b128c09f BB |
1827 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
1828 | } else { | |
1829 | ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); | |
1830 | BP_SET_LSIZE(bp, lsize); | |
b0bc7a84 MG |
1831 | BP_SET_TYPE(bp, zp->zp_type); |
1832 | BP_SET_LEVEL(bp, zp->zp_level); | |
428870ff | 1833 | BP_SET_PSIZE(bp, psize); |
b128c09f BB |
1834 | BP_SET_COMPRESS(bp, compress); |
1835 | BP_SET_CHECKSUM(bp, zp->zp_checksum); | |
428870ff | 1836 | BP_SET_DEDUP(bp, zp->zp_dedup); |
b128c09f | 1837 | BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); |
428870ff BB |
1838 | if (zp->zp_dedup) { |
1839 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1840 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
b5256303 TC |
1841 | ASSERT(!zp->zp_encrypt || |
1842 | DMU_OT_IS_ENCRYPTED(zp->zp_type)); | |
428870ff BB |
1843 | zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; |
1844 | } | |
03c6040b GW |
1845 | if (zp->zp_nopwrite) { |
1846 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1847 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
1848 | zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; | |
1849 | } | |
428870ff | 1850 | } |
62840030 | 1851 | return (zio); |
428870ff BB |
1852 | } |
1853 | ||
62840030 | 1854 | static zio_t * |
428870ff BB |
1855 | zio_free_bp_init(zio_t *zio) |
1856 | { | |
1857 | blkptr_t *bp = zio->io_bp; | |
1858 | ||
1859 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
1860 | if (BP_GET_DEDUP(bp)) | |
1861 | zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; | |
b128c09f | 1862 | } |
34dc7c2f | 1863 | |
a1d477c2 MA |
1864 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
1865 | ||
62840030 | 1866 | return (zio); |
34dc7c2f BB |
1867 | } |
1868 | ||
b128c09f BB |
1869 | /* |
1870 | * ========================================================================== | |
1871 | * Execute the I/O pipeline | |
1872 | * ========================================================================== | |
1873 | */ | |
1874 | ||
1875 | static void | |
7ef5e54e | 1876 | zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) |
34dc7c2f | 1877 | { |
428870ff | 1878 | spa_t *spa = zio->io_spa; |
b128c09f | 1879 | zio_type_t t = zio->io_type; |
a38718a6 | 1880 | int flags = (cutinline ? TQ_FRONT : 0); |
34dc7c2f BB |
1881 | |
1882 | /* | |
9babb374 BB |
1883 | * If we're a config writer or a probe, the normal issue and |
1884 | * interrupt threads may all be blocked waiting for the config lock. | |
1885 | * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. | |
34dc7c2f | 1886 | */ |
9babb374 | 1887 | if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) |
b128c09f | 1888 | t = ZIO_TYPE_NULL; |
34dc7c2f BB |
1889 | |
1890 | /* | |
b128c09f | 1891 | * A similar issue exists for the L2ARC write thread until L2ARC 2.0. |
34dc7c2f | 1892 | */ |
b128c09f BB |
1893 | if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) |
1894 | t = ZIO_TYPE_NULL; | |
34dc7c2f | 1895 | |
428870ff | 1896 | /* |
7ef5e54e AL |
1897 | * If this is a high priority I/O, then use the high priority taskq if |
1898 | * available. | |
428870ff | 1899 | */ |
18b14b17 GW |
1900 | if ((zio->io_priority == ZIO_PRIORITY_NOW || |
1901 | zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) && | |
7ef5e54e | 1902 | spa->spa_zio_taskq[t][q + 1].stqs_count != 0) |
428870ff BB |
1903 | q++; |
1904 | ||
1905 | ASSERT3U(q, <, ZIO_TASKQ_TYPES); | |
5cc556b4 | 1906 | |
a38718a6 GA |
1907 | /* |
1908 | * NB: We are assuming that the zio can only be dispatched | |
1909 | * to a single taskq at a time. It would be a grievous error | |
1910 | * to dispatch the zio to another taskq at the same time. | |
1911 | */ | |
1912 | ASSERT(taskq_empty_ent(&zio->io_tqent)); | |
23c13c7e AL |
1913 | spa_taskq_dispatch_ent(spa, t, q, zio_execute, zio, flags, |
1914 | &zio->io_tqent); | |
b128c09f | 1915 | } |
34dc7c2f | 1916 | |
b128c09f | 1917 | static boolean_t |
7ef5e54e | 1918 | zio_taskq_member(zio_t *zio, zio_taskq_type_t q) |
b128c09f | 1919 | { |
b128c09f | 1920 | spa_t *spa = zio->io_spa; |
34dc7c2f | 1921 | |
b3212d2f MA |
1922 | taskq_t *tq = taskq_of_curthread(); |
1923 | ||
1c27024e | 1924 | for (zio_type_t t = 0; t < ZIO_TYPES; t++) { |
7ef5e54e AL |
1925 | spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; |
1926 | uint_t i; | |
1927 | for (i = 0; i < tqs->stqs_count; i++) { | |
b3212d2f | 1928 | if (tqs->stqs_taskq[i] == tq) |
7ef5e54e AL |
1929 | return (B_TRUE); |
1930 | } | |
1931 | } | |
34dc7c2f | 1932 | |
b128c09f BB |
1933 | return (B_FALSE); |
1934 | } | |
34dc7c2f | 1935 | |
62840030 | 1936 | static zio_t * |
b128c09f BB |
1937 | zio_issue_async(zio_t *zio) |
1938 | { | |
428870ff | 1939 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); |
b128c09f | 1940 | |
62840030 | 1941 | return (NULL); |
34dc7c2f BB |
1942 | } |
1943 | ||
b128c09f | 1944 | void |
23c13c7e | 1945 | zio_interrupt(void *zio) |
34dc7c2f | 1946 | { |
428870ff | 1947 | zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); |
b128c09f | 1948 | } |
34dc7c2f | 1949 | |
d3c1e45b MM |
1950 | void |
1951 | zio_delay_interrupt(zio_t *zio) | |
1952 | { | |
1953 | /* | |
1954 | * The timeout_generic() function isn't defined in userspace, so | |
1955 | * rather than trying to implement the function, the zio delay | |
1956 | * functionality has been disabled for userspace builds. | |
1957 | */ | |
1958 | ||
1959 | #ifdef _KERNEL | |
1960 | /* | |
1961 | * If io_target_timestamp is zero, then no delay has been registered | |
1962 | * for this IO, thus jump to the end of this function and "skip" the | |
1963 | * delay; issuing it directly to the zio layer. | |
1964 | */ | |
1965 | if (zio->io_target_timestamp != 0) { | |
1966 | hrtime_t now = gethrtime(); | |
1967 | ||
1968 | if (now >= zio->io_target_timestamp) { | |
1969 | /* | |
1970 | * This IO has already taken longer than the target | |
1971 | * delay to complete, so we don't want to delay it | |
1972 | * any longer; we "miss" the delay and issue it | |
1973 | * directly to the zio layer. This is likely due to | |
1974 | * the target latency being set to a value less than | |
1975 | * the underlying hardware can satisfy (e.g. delay | |
1976 | * set to 1ms, but the disks take 10ms to complete an | |
1977 | * IO request). | |
1978 | */ | |
1979 | ||
1980 | DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, | |
1981 | hrtime_t, now); | |
1982 | ||
1983 | zio_interrupt(zio); | |
1984 | } else { | |
1985 | taskqid_t tid; | |
1986 | hrtime_t diff = zio->io_target_timestamp - now; | |
1987 | clock_t expire_at_tick = ddi_get_lbolt() + | |
1988 | NSEC_TO_TICK(diff); | |
1989 | ||
1990 | DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, | |
1991 | hrtime_t, now, hrtime_t, diff); | |
1992 | ||
1993 | if (NSEC_TO_TICK(diff) == 0) { | |
1994 | /* Our delay is less than a jiffy - just spin */ | |
1995 | zfs_sleep_until(zio->io_target_timestamp); | |
1996 | zio_interrupt(zio); | |
1997 | } else { | |
1998 | /* | |
1999 | * Use taskq_dispatch_delay() in the place of | |
2000 | * OpenZFS's timeout_generic(). | |
2001 | */ | |
2002 | tid = taskq_dispatch_delay(system_taskq, | |
23c13c7e AL |
2003 | zio_interrupt, zio, TQ_NOSLEEP, |
2004 | expire_at_tick); | |
d3c1e45b MM |
2005 | if (tid == TASKQID_INVALID) { |
2006 | /* | |
2007 | * Couldn't allocate a task. Just | |
2008 | * finish the zio without a delay. | |
2009 | */ | |
2010 | zio_interrupt(zio); | |
2011 | } | |
2012 | } | |
2013 | } | |
2014 | return; | |
2015 | } | |
2016 | #endif | |
2017 | DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); | |
2018 | zio_interrupt(zio); | |
2019 | } | |
2020 | ||
8fb1ede1 | 2021 | static void |
638dd5f4 | 2022 | zio_deadman_impl(zio_t *pio, int ziodepth) |
8fb1ede1 BB |
2023 | { |
2024 | zio_t *cio, *cio_next; | |
2025 | zio_link_t *zl = NULL; | |
2026 | vdev_t *vd = pio->io_vd; | |
2027 | ||
638dd5f4 TC |
2028 | if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) { |
2029 | vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL; | |
8fb1ede1 BB |
2030 | zbookmark_phys_t *zb = &pio->io_bookmark; |
2031 | uint64_t delta = gethrtime() - pio->io_timestamp; | |
2032 | uint64_t failmode = spa_get_deadman_failmode(pio->io_spa); | |
2033 | ||
a887d653 | 2034 | zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu " |
8fb1ede1 | 2035 | "delta=%llu queued=%llu io=%llu " |
8e739b2c RE |
2036 | "path=%s " |
2037 | "last=%llu type=%d " | |
2038 | "priority=%d flags=0x%x stage=0x%x " | |
2039 | "pipeline=0x%x pipeline-trace=0x%x " | |
2040 | "objset=%llu object=%llu " | |
2041 | "level=%llu blkid=%llu " | |
2042 | "offset=%llu size=%llu " | |
2043 | "error=%d", | |
638dd5f4 | 2044 | ziodepth, pio, pio->io_timestamp, |
8e739b2c RE |
2045 | (u_longlong_t)delta, pio->io_delta, pio->io_delay, |
2046 | vd ? vd->vdev_path : "NULL", | |
2047 | vq ? vq->vq_io_complete_ts : 0, pio->io_type, | |
2048 | pio->io_priority, pio->io_flags, pio->io_stage, | |
2049 | pio->io_pipeline, pio->io_pipeline_trace, | |
2050 | (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object, | |
2051 | (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid, | |
2052 | (u_longlong_t)pio->io_offset, (u_longlong_t)pio->io_size, | |
2053 | pio->io_error); | |
1144586b | 2054 | (void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN, |
4f072827 | 2055 | pio->io_spa, vd, zb, pio, 0); |
8fb1ede1 BB |
2056 | |
2057 | if (failmode == ZIO_FAILURE_MODE_CONTINUE && | |
2058 | taskq_empty_ent(&pio->io_tqent)) { | |
2059 | zio_interrupt(pio); | |
2060 | } | |
2061 | } | |
2062 | ||
2063 | mutex_enter(&pio->io_lock); | |
2064 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { | |
2065 | cio_next = zio_walk_children(pio, &zl); | |
638dd5f4 | 2066 | zio_deadman_impl(cio, ziodepth + 1); |
8fb1ede1 BB |
2067 | } |
2068 | mutex_exit(&pio->io_lock); | |
2069 | } | |
2070 | ||
2071 | /* | |
2072 | * Log the critical information describing this zio and all of its children | |
2073 | * using the zfs_dbgmsg() interface then post deadman event for the ZED. | |
2074 | */ | |
2075 | void | |
2076 | zio_deadman(zio_t *pio, char *tag) | |
2077 | { | |
2078 | spa_t *spa = pio->io_spa; | |
2079 | char *name = spa_name(spa); | |
2080 | ||
2081 | if (!zfs_deadman_enabled || spa_suspended(spa)) | |
2082 | return; | |
2083 | ||
638dd5f4 | 2084 | zio_deadman_impl(pio, 0); |
8fb1ede1 BB |
2085 | |
2086 | switch (spa_get_deadman_failmode(spa)) { | |
2087 | case ZIO_FAILURE_MODE_WAIT: | |
2088 | zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name); | |
2089 | break; | |
2090 | ||
2091 | case ZIO_FAILURE_MODE_CONTINUE: | |
2092 | zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name); | |
2093 | break; | |
2094 | ||
2095 | case ZIO_FAILURE_MODE_PANIC: | |
2096 | fm_panic("%s determined I/O to pool '%s' is hung.", tag, name); | |
2097 | break; | |
2098 | } | |
2099 | } | |
2100 | ||
b128c09f BB |
2101 | /* |
2102 | * Execute the I/O pipeline until one of the following occurs: | |
2103 | * (1) the I/O completes; (2) the pipeline stalls waiting for | |
2104 | * dependent child I/Os; (3) the I/O issues, so we're waiting | |
2105 | * for an I/O completion interrupt; (4) the I/O is delegated by | |
2106 | * vdev-level caching or aggregation; (5) the I/O is deferred | |
2107 | * due to vdev-level queueing; (6) the I/O is handed off to | |
2108 | * another thread. In all cases, the pipeline stops whenever | |
8e07b99b | 2109 | * there's no CPU work; it never burns a thread in cv_wait_io(). |
b128c09f BB |
2110 | * |
2111 | * There's no locking on io_stage because there's no legitimate way | |
2112 | * for multiple threads to be attempting to process the same I/O. | |
2113 | */ | |
428870ff | 2114 | static zio_pipe_stage_t *zio_pipeline[]; |
34dc7c2f | 2115 | |
da6b4005 NB |
2116 | /* |
2117 | * zio_execute() is a wrapper around the static function | |
2118 | * __zio_execute() so that we can force __zio_execute() to be | |
2119 | * inlined. This reduces stack overhead which is important | |
2120 | * because __zio_execute() is called recursively in several zio | |
2121 | * code paths. zio_execute() itself cannot be inlined because | |
2122 | * it is externally visible. | |
2123 | */ | |
b128c09f | 2124 | void |
23c13c7e | 2125 | zio_execute(void *zio) |
da6b4005 | 2126 | { |
92119cc2 BB |
2127 | fstrans_cookie_t cookie; |
2128 | ||
2129 | cookie = spl_fstrans_mark(); | |
da6b4005 | 2130 | __zio_execute(zio); |
92119cc2 | 2131 | spl_fstrans_unmark(cookie); |
da6b4005 NB |
2132 | } |
2133 | ||
b58986ee BB |
2134 | /* |
2135 | * Used to determine if in the current context the stack is sized large | |
2136 | * enough to allow zio_execute() to be called recursively. A minimum | |
2137 | * stack size of 16K is required to avoid needing to re-dispatch the zio. | |
2138 | */ | |
65c7cc49 | 2139 | static boolean_t |
b58986ee BB |
2140 | zio_execute_stack_check(zio_t *zio) |
2141 | { | |
2142 | #if !defined(HAVE_LARGE_STACKS) | |
2143 | dsl_pool_t *dp = spa_get_dsl(zio->io_spa); | |
2144 | ||
2145 | /* Executing in txg_sync_thread() context. */ | |
2146 | if (dp && curthread == dp->dp_tx.tx_sync_thread) | |
2147 | return (B_TRUE); | |
2148 | ||
2149 | /* Pool initialization outside of zio_taskq context. */ | |
2150 | if (dp && spa_is_initializing(dp->dp_spa) && | |
2151 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) && | |
2152 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH)) | |
2153 | return (B_TRUE); | |
14e4e3cb AZ |
2154 | #else |
2155 | (void) zio; | |
b58986ee BB |
2156 | #endif /* HAVE_LARGE_STACKS */ |
2157 | ||
2158 | return (B_FALSE); | |
2159 | } | |
2160 | ||
da6b4005 NB |
2161 | __attribute__((always_inline)) |
2162 | static inline void | |
2163 | __zio_execute(zio_t *zio) | |
b128c09f | 2164 | { |
3dfb57a3 DB |
2165 | ASSERT3U(zio->io_queued_timestamp, >, 0); |
2166 | ||
b128c09f | 2167 | while (zio->io_stage < ZIO_STAGE_DONE) { |
428870ff BB |
2168 | enum zio_stage pipeline = zio->io_pipeline; |
2169 | enum zio_stage stage = zio->io_stage; | |
62840030 MA |
2170 | |
2171 | zio->io_executor = curthread; | |
34dc7c2f | 2172 | |
b128c09f | 2173 | ASSERT(!MUTEX_HELD(&zio->io_lock)); |
428870ff BB |
2174 | ASSERT(ISP2(stage)); |
2175 | ASSERT(zio->io_stall == NULL); | |
34dc7c2f | 2176 | |
428870ff BB |
2177 | do { |
2178 | stage <<= 1; | |
2179 | } while ((stage & pipeline) == 0); | |
b128c09f BB |
2180 | |
2181 | ASSERT(stage <= ZIO_STAGE_DONE); | |
34dc7c2f BB |
2182 | |
2183 | /* | |
b128c09f BB |
2184 | * If we are in interrupt context and this pipeline stage |
2185 | * will grab a config lock that is held across I/O, | |
428870ff BB |
2186 | * or may wait for an I/O that needs an interrupt thread |
2187 | * to complete, issue async to avoid deadlock. | |
2188 | * | |
2189 | * For VDEV_IO_START, we cut in line so that the io will | |
2190 | * be sent to disk promptly. | |
34dc7c2f | 2191 | */ |
91579709 BB |
2192 | if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && |
2193 | zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { | |
b58986ee BB |
2194 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? |
2195 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
91579709 BB |
2196 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
2197 | return; | |
2198 | } | |
2199 | ||
2200 | /* | |
b58986ee BB |
2201 | * If the current context doesn't have large enough stacks |
2202 | * the zio must be issued asynchronously to prevent overflow. | |
91579709 | 2203 | */ |
b58986ee BB |
2204 | if (zio_execute_stack_check(zio)) { |
2205 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? | |
2206 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
428870ff | 2207 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
b128c09f | 2208 | return; |
34dc7c2f BB |
2209 | } |
2210 | ||
b128c09f | 2211 | zio->io_stage = stage; |
3dfb57a3 | 2212 | zio->io_pipeline_trace |= zio->io_stage; |
34dc7c2f | 2213 | |
62840030 MA |
2214 | /* |
2215 | * The zio pipeline stage returns the next zio to execute | |
2216 | * (typically the same as this one), or NULL if we should | |
2217 | * stop. | |
2218 | */ | |
2219 | zio = zio_pipeline[highbit64(stage) - 1](zio); | |
34dc7c2f | 2220 | |
62840030 MA |
2221 | if (zio == NULL) |
2222 | return; | |
b128c09f | 2223 | } |
34dc7c2f BB |
2224 | } |
2225 | ||
da6b4005 | 2226 | |
b128c09f BB |
2227 | /* |
2228 | * ========================================================================== | |
2229 | * Initiate I/O, either sync or async | |
2230 | * ========================================================================== | |
2231 | */ | |
2232 | int | |
2233 | zio_wait(zio_t *zio) | |
34dc7c2f | 2234 | { |
9cdf7b1f MA |
2235 | /* |
2236 | * Some routines, like zio_free_sync(), may return a NULL zio | |
2237 | * to avoid the performance overhead of creating and then destroying | |
2238 | * an unneeded zio. For the callers' simplicity, we accept a NULL | |
2239 | * zio and ignore it. | |
2240 | */ | |
2241 | if (zio == NULL) | |
2242 | return (0); | |
2243 | ||
8fb1ede1 | 2244 | long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms); |
b128c09f | 2245 | int error; |
34dc7c2f | 2246 | |
1ce23dca PS |
2247 | ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN); |
2248 | ASSERT3P(zio->io_executor, ==, NULL); | |
34dc7c2f | 2249 | |
b128c09f | 2250 | zio->io_waiter = curthread; |
3dfb57a3 DB |
2251 | ASSERT0(zio->io_queued_timestamp); |
2252 | zio->io_queued_timestamp = gethrtime(); | |
34dc7c2f | 2253 | |
da6b4005 | 2254 | __zio_execute(zio); |
34dc7c2f | 2255 | |
b128c09f | 2256 | mutex_enter(&zio->io_lock); |
8fb1ede1 BB |
2257 | while (zio->io_executor != NULL) { |
2258 | error = cv_timedwait_io(&zio->io_cv, &zio->io_lock, | |
2259 | ddi_get_lbolt() + timeout); | |
2260 | ||
2261 | if (zfs_deadman_enabled && error == -1 && | |
2262 | gethrtime() - zio->io_queued_timestamp > | |
2263 | spa_deadman_ziotime(zio->io_spa)) { | |
2264 | mutex_exit(&zio->io_lock); | |
2265 | timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms); | |
2266 | zio_deadman(zio, FTAG); | |
2267 | mutex_enter(&zio->io_lock); | |
2268 | } | |
2269 | } | |
b128c09f | 2270 | mutex_exit(&zio->io_lock); |
34dc7c2f | 2271 | |
b128c09f BB |
2272 | error = zio->io_error; |
2273 | zio_destroy(zio); | |
34dc7c2f | 2274 | |
b128c09f BB |
2275 | return (error); |
2276 | } | |
34dc7c2f | 2277 | |
b128c09f BB |
2278 | void |
2279 | zio_nowait(zio_t *zio) | |
2280 | { | |
9cdf7b1f MA |
2281 | /* |
2282 | * See comment in zio_wait(). | |
2283 | */ | |
2284 | if (zio == NULL) | |
2285 | return; | |
2286 | ||
1ce23dca | 2287 | ASSERT3P(zio->io_executor, ==, NULL); |
34dc7c2f | 2288 | |
d164b209 BB |
2289 | if (zio->io_child_type == ZIO_CHILD_LOGICAL && |
2290 | zio_unique_parent(zio) == NULL) { | |
8878261f BB |
2291 | zio_t *pio; |
2292 | ||
34dc7c2f | 2293 | /* |
b128c09f | 2294 | * This is a logical async I/O with no parent to wait for it. |
9babb374 BB |
2295 | * We add it to the spa_async_root_zio "Godfather" I/O which |
2296 | * will ensure they complete prior to unloading the pool. | |
34dc7c2f | 2297 | */ |
b128c09f | 2298 | spa_t *spa = zio->io_spa; |
09eb36ce | 2299 | pio = spa->spa_async_zio_root[CPU_SEQID_UNSTABLE]; |
9babb374 | 2300 | |
8878261f | 2301 | zio_add_child(pio, zio); |
b128c09f | 2302 | } |
34dc7c2f | 2303 | |
3dfb57a3 DB |
2304 | ASSERT0(zio->io_queued_timestamp); |
2305 | zio->io_queued_timestamp = gethrtime(); | |
da6b4005 | 2306 | __zio_execute(zio); |
b128c09f | 2307 | } |
34dc7c2f | 2308 | |
b128c09f BB |
2309 | /* |
2310 | * ========================================================================== | |
1ce23dca | 2311 | * Reexecute, cancel, or suspend/resume failed I/O |
b128c09f BB |
2312 | * ========================================================================== |
2313 | */ | |
34dc7c2f | 2314 | |
b128c09f | 2315 | static void |
23c13c7e | 2316 | zio_reexecute(void *arg) |
b128c09f | 2317 | { |
23c13c7e | 2318 | zio_t *pio = arg; |
d164b209 BB |
2319 | zio_t *cio, *cio_next; |
2320 | ||
2321 | ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); | |
2322 | ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); | |
9babb374 BB |
2323 | ASSERT(pio->io_gang_leader == NULL); |
2324 | ASSERT(pio->io_gang_tree == NULL); | |
34dc7c2f | 2325 | |
b128c09f BB |
2326 | pio->io_flags = pio->io_orig_flags; |
2327 | pio->io_stage = pio->io_orig_stage; | |
2328 | pio->io_pipeline = pio->io_orig_pipeline; | |
2329 | pio->io_reexecute = 0; | |
03c6040b | 2330 | pio->io_flags |= ZIO_FLAG_REEXECUTED; |
3dfb57a3 | 2331 | pio->io_pipeline_trace = 0; |
b128c09f | 2332 | pio->io_error = 0; |
1c27024e | 2333 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 | 2334 | pio->io_state[w] = 0; |
1c27024e | 2335 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) |
b128c09f | 2336 | pio->io_child_error[c] = 0; |
34dc7c2f | 2337 | |
428870ff BB |
2338 | if (IO_IS_ALLOCATING(pio)) |
2339 | BP_ZERO(pio->io_bp); | |
34dc7c2f | 2340 | |
b128c09f BB |
2341 | /* |
2342 | * As we reexecute pio's children, new children could be created. | |
d164b209 | 2343 | * New children go to the head of pio's io_child_list, however, |
b128c09f | 2344 | * so we will (correctly) not reexecute them. The key is that |
d164b209 BB |
2345 | * the remainder of pio's io_child_list, from 'cio_next' onward, |
2346 | * cannot be affected by any side effects of reexecuting 'cio'. | |
b128c09f | 2347 | */ |
1c27024e | 2348 | zio_link_t *zl = NULL; |
a8b2e306 | 2349 | mutex_enter(&pio->io_lock); |
3dfb57a3 DB |
2350 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { |
2351 | cio_next = zio_walk_children(pio, &zl); | |
1c27024e | 2352 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 | 2353 | pio->io_children[cio->io_child_type][w]++; |
b128c09f | 2354 | mutex_exit(&pio->io_lock); |
d164b209 | 2355 | zio_reexecute(cio); |
a8b2e306 | 2356 | mutex_enter(&pio->io_lock); |
34dc7c2f | 2357 | } |
a8b2e306 | 2358 | mutex_exit(&pio->io_lock); |
34dc7c2f | 2359 | |
b128c09f BB |
2360 | /* |
2361 | * Now that all children have been reexecuted, execute the parent. | |
9babb374 | 2362 | * We don't reexecute "The Godfather" I/O here as it's the |
9e2c3bb4 | 2363 | * responsibility of the caller to wait on it. |
b128c09f | 2364 | */ |
3dfb57a3 DB |
2365 | if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) { |
2366 | pio->io_queued_timestamp = gethrtime(); | |
da6b4005 | 2367 | __zio_execute(pio); |
3dfb57a3 | 2368 | } |
34dc7c2f BB |
2369 | } |
2370 | ||
b128c09f | 2371 | void |
cec3a0a1 | 2372 | zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason) |
34dc7c2f | 2373 | { |
b128c09f BB |
2374 | if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) |
2375 | fm_panic("Pool '%s' has encountered an uncorrectable I/O " | |
2376 | "failure and the failure mode property for this pool " | |
2377 | "is set to panic.", spa_name(spa)); | |
34dc7c2f | 2378 | |
bf89c199 BB |
2379 | cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O " |
2380 | "failure and has been suspended.\n", spa_name(spa)); | |
2381 | ||
1144586b | 2382 | (void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, |
4f072827 | 2383 | NULL, NULL, 0); |
34dc7c2f | 2384 | |
b128c09f | 2385 | mutex_enter(&spa->spa_suspend_lock); |
34dc7c2f | 2386 | |
b128c09f | 2387 | if (spa->spa_suspend_zio_root == NULL) |
9babb374 BB |
2388 | spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, |
2389 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
2390 | ZIO_FLAG_GODFATHER); | |
34dc7c2f | 2391 | |
cec3a0a1 | 2392 | spa->spa_suspended = reason; |
34dc7c2f | 2393 | |
b128c09f | 2394 | if (zio != NULL) { |
9babb374 | 2395 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f BB |
2396 | ASSERT(zio != spa->spa_suspend_zio_root); |
2397 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
d164b209 | 2398 | ASSERT(zio_unique_parent(zio) == NULL); |
b128c09f BB |
2399 | ASSERT(zio->io_stage == ZIO_STAGE_DONE); |
2400 | zio_add_child(spa->spa_suspend_zio_root, zio); | |
2401 | } | |
34dc7c2f | 2402 | |
b128c09f BB |
2403 | mutex_exit(&spa->spa_suspend_lock); |
2404 | } | |
34dc7c2f | 2405 | |
9babb374 | 2406 | int |
b128c09f BB |
2407 | zio_resume(spa_t *spa) |
2408 | { | |
9babb374 | 2409 | zio_t *pio; |
34dc7c2f BB |
2410 | |
2411 | /* | |
b128c09f | 2412 | * Reexecute all previously suspended i/o. |
34dc7c2f | 2413 | */ |
b128c09f | 2414 | mutex_enter(&spa->spa_suspend_lock); |
cec3a0a1 | 2415 | spa->spa_suspended = ZIO_SUSPEND_NONE; |
b128c09f BB |
2416 | cv_broadcast(&spa->spa_suspend_cv); |
2417 | pio = spa->spa_suspend_zio_root; | |
2418 | spa->spa_suspend_zio_root = NULL; | |
2419 | mutex_exit(&spa->spa_suspend_lock); | |
2420 | ||
2421 | if (pio == NULL) | |
9babb374 | 2422 | return (0); |
34dc7c2f | 2423 | |
9babb374 BB |
2424 | zio_reexecute(pio); |
2425 | return (zio_wait(pio)); | |
b128c09f BB |
2426 | } |
2427 | ||
2428 | void | |
2429 | zio_resume_wait(spa_t *spa) | |
2430 | { | |
2431 | mutex_enter(&spa->spa_suspend_lock); | |
2432 | while (spa_suspended(spa)) | |
2433 | cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); | |
2434 | mutex_exit(&spa->spa_suspend_lock); | |
34dc7c2f BB |
2435 | } |
2436 | ||
2437 | /* | |
2438 | * ========================================================================== | |
b128c09f BB |
2439 | * Gang blocks. |
2440 | * | |
2441 | * A gang block is a collection of small blocks that looks to the DMU | |
2442 | * like one large block. When zio_dva_allocate() cannot find a block | |
2443 | * of the requested size, due to either severe fragmentation or the pool | |
2444 | * being nearly full, it calls zio_write_gang_block() to construct the | |
2445 | * block from smaller fragments. | |
2446 | * | |
2447 | * A gang block consists of a gang header (zio_gbh_phys_t) and up to | |
2448 | * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like | |
2449 | * an indirect block: it's an array of block pointers. It consumes | |
2450 | * only one sector and hence is allocatable regardless of fragmentation. | |
2451 | * The gang header's bps point to its gang members, which hold the data. | |
2452 | * | |
2453 | * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> | |
2454 | * as the verifier to ensure uniqueness of the SHA256 checksum. | |
2455 | * Critically, the gang block bp's blk_cksum is the checksum of the data, | |
2456 | * not the gang header. This ensures that data block signatures (needed for | |
2457 | * deduplication) are independent of how the block is physically stored. | |
2458 | * | |
2459 | * Gang blocks can be nested: a gang member may itself be a gang block. | |
2460 | * Thus every gang block is a tree in which root and all interior nodes are | |
2461 | * gang headers, and the leaves are normal blocks that contain user data. | |
2462 | * The root of the gang tree is called the gang leader. | |
2463 | * | |
2464 | * To perform any operation (read, rewrite, free, claim) on a gang block, | |
2465 | * zio_gang_assemble() first assembles the gang tree (minus data leaves) | |
2466 | * in the io_gang_tree field of the original logical i/o by recursively | |
2467 | * reading the gang leader and all gang headers below it. This yields | |
2468 | * an in-core tree containing the contents of every gang header and the | |
2469 | * bps for every constituent of the gang block. | |
2470 | * | |
2471 | * With the gang tree now assembled, zio_gang_issue() just walks the gang tree | |
2472 | * and invokes a callback on each bp. To free a gang block, zio_gang_issue() | |
2473 | * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. | |
2474 | * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). | |
2475 | * zio_read_gang() is a wrapper around zio_read() that omits reading gang | |
2476 | * headers, since we already have those in io_gang_tree. zio_rewrite_gang() | |
2477 | * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() | |
2478 | * of the gang header plus zio_checksum_compute() of the data to update the | |
2479 | * gang header's blk_cksum as described above. | |
2480 | * | |
2481 | * The two-phase assemble/issue model solves the problem of partial failure -- | |
2482 | * what if you'd freed part of a gang block but then couldn't read the | |
2483 | * gang header for another part? Assembling the entire gang tree first | |
2484 | * ensures that all the necessary gang header I/O has succeeded before | |
2485 | * starting the actual work of free, claim, or write. Once the gang tree | |
2486 | * is assembled, free and claim are in-memory operations that cannot fail. | |
2487 | * | |
2488 | * In the event that a gang write fails, zio_dva_unallocate() walks the | |
2489 | * gang tree to immediately free (i.e. insert back into the space map) | |
2490 | * everything we've allocated. This ensures that we don't get ENOSPC | |
2491 | * errors during repeated suspend/resume cycles due to a flaky device. | |
2492 | * | |
2493 | * Gang rewrites only happen during sync-to-convergence. If we can't assemble | |
2494 | * the gang tree, we won't modify the block, so we can safely defer the free | |
2495 | * (knowing that the block is still intact). If we *can* assemble the gang | |
2496 | * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free | |
2497 | * each constituent bp and we can allocate a new block on the next sync pass. | |
2498 | * | |
2499 | * In all cases, the gang tree allows complete recovery from partial failure. | |
34dc7c2f BB |
2500 | * ========================================================================== |
2501 | */ | |
b128c09f | 2502 | |
a6255b7f DQ |
2503 | static void |
2504 | zio_gang_issue_func_done(zio_t *zio) | |
2505 | { | |
e2af2acc | 2506 | abd_free(zio->io_abd); |
a6255b7f DQ |
2507 | } |
2508 | ||
b128c09f | 2509 | static zio_t * |
a6255b7f DQ |
2510 | zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, |
2511 | uint64_t offset) | |
34dc7c2f | 2512 | { |
b128c09f BB |
2513 | if (gn != NULL) |
2514 | return (pio); | |
34dc7c2f | 2515 | |
a6255b7f DQ |
2516 | return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset), |
2517 | BP_GET_PSIZE(bp), zio_gang_issue_func_done, | |
2518 | NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
b128c09f BB |
2519 | &pio->io_bookmark)); |
2520 | } | |
2521 | ||
a6255b7f DQ |
2522 | static zio_t * |
2523 | zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2524 | uint64_t offset) | |
b128c09f BB |
2525 | { |
2526 | zio_t *zio; | |
2527 | ||
2528 | if (gn != NULL) { | |
a6255b7f DQ |
2529 | abd_t *gbh_abd = |
2530 | abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
b128c09f | 2531 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
a6255b7f DQ |
2532 | gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL, |
2533 | pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
2534 | &pio->io_bookmark); | |
34dc7c2f | 2535 | /* |
b128c09f BB |
2536 | * As we rewrite each gang header, the pipeline will compute |
2537 | * a new gang block header checksum for it; but no one will | |
2538 | * compute a new data checksum, so we do that here. The one | |
2539 | * exception is the gang leader: the pipeline already computed | |
2540 | * its data checksum because that stage precedes gang assembly. | |
2541 | * (Presently, nothing actually uses interior data checksums; | |
2542 | * this is just good hygiene.) | |
34dc7c2f | 2543 | */ |
9babb374 | 2544 | if (gn != pio->io_gang_leader->io_gang_tree) { |
a6255b7f DQ |
2545 | abd_t *buf = abd_get_offset(data, offset); |
2546 | ||
b128c09f | 2547 | zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), |
a6255b7f DQ |
2548 | buf, BP_GET_PSIZE(bp)); |
2549 | ||
e2af2acc | 2550 | abd_free(buf); |
b128c09f | 2551 | } |
428870ff BB |
2552 | /* |
2553 | * If we are here to damage data for testing purposes, | |
2554 | * leave the GBH alone so that we can detect the damage. | |
2555 | */ | |
2556 | if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) | |
2557 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; | |
34dc7c2f | 2558 | } else { |
b128c09f | 2559 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
a6255b7f DQ |
2560 | abd_get_offset(data, offset), BP_GET_PSIZE(bp), |
2561 | zio_gang_issue_func_done, NULL, pio->io_priority, | |
b128c09f | 2562 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); |
34dc7c2f BB |
2563 | } |
2564 | ||
b128c09f BB |
2565 | return (zio); |
2566 | } | |
34dc7c2f | 2567 | |
a6255b7f DQ |
2568 | static zio_t * |
2569 | zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2570 | uint64_t offset) | |
b128c09f | 2571 | { |
14e4e3cb AZ |
2572 | (void) gn, (void) data, (void) offset; |
2573 | ||
9cdf7b1f MA |
2574 | zio_t *zio = zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, |
2575 | ZIO_GANG_CHILD_FLAGS(pio)); | |
2576 | if (zio == NULL) { | |
2577 | zio = zio_null(pio, pio->io_spa, | |
2578 | NULL, NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)); | |
2579 | } | |
2580 | return (zio); | |
34dc7c2f BB |
2581 | } |
2582 | ||
a6255b7f DQ |
2583 | static zio_t * |
2584 | zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2585 | uint64_t offset) | |
34dc7c2f | 2586 | { |
14e4e3cb | 2587 | (void) gn, (void) data, (void) offset; |
b128c09f BB |
2588 | return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, |
2589 | NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); | |
2590 | } | |
2591 | ||
2592 | static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { | |
2593 | NULL, | |
2594 | zio_read_gang, | |
2595 | zio_rewrite_gang, | |
2596 | zio_free_gang, | |
2597 | zio_claim_gang, | |
2598 | NULL | |
2599 | }; | |
34dc7c2f | 2600 | |
b128c09f | 2601 | static void zio_gang_tree_assemble_done(zio_t *zio); |
34dc7c2f | 2602 | |
b128c09f BB |
2603 | static zio_gang_node_t * |
2604 | zio_gang_node_alloc(zio_gang_node_t **gnpp) | |
2605 | { | |
2606 | zio_gang_node_t *gn; | |
34dc7c2f | 2607 | |
b128c09f | 2608 | ASSERT(*gnpp == NULL); |
34dc7c2f | 2609 | |
79c76d5b | 2610 | gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); |
b128c09f BB |
2611 | gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); |
2612 | *gnpp = gn; | |
34dc7c2f | 2613 | |
b128c09f | 2614 | return (gn); |
34dc7c2f BB |
2615 | } |
2616 | ||
34dc7c2f | 2617 | static void |
b128c09f | 2618 | zio_gang_node_free(zio_gang_node_t **gnpp) |
34dc7c2f | 2619 | { |
b128c09f | 2620 | zio_gang_node_t *gn = *gnpp; |
34dc7c2f | 2621 | |
1c27024e | 2622 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f BB |
2623 | ASSERT(gn->gn_child[g] == NULL); |
2624 | ||
2625 | zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
2626 | kmem_free(gn, sizeof (*gn)); | |
2627 | *gnpp = NULL; | |
34dc7c2f BB |
2628 | } |
2629 | ||
b128c09f BB |
2630 | static void |
2631 | zio_gang_tree_free(zio_gang_node_t **gnpp) | |
34dc7c2f | 2632 | { |
b128c09f | 2633 | zio_gang_node_t *gn = *gnpp; |
34dc7c2f | 2634 | |
b128c09f BB |
2635 | if (gn == NULL) |
2636 | return; | |
34dc7c2f | 2637 | |
1c27024e | 2638 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f | 2639 | zio_gang_tree_free(&gn->gn_child[g]); |
34dc7c2f | 2640 | |
b128c09f | 2641 | zio_gang_node_free(gnpp); |
34dc7c2f BB |
2642 | } |
2643 | ||
b128c09f | 2644 | static void |
9babb374 | 2645 | zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) |
34dc7c2f | 2646 | { |
b128c09f | 2647 | zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); |
a6255b7f | 2648 | abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); |
b128c09f | 2649 | |
9babb374 | 2650 | ASSERT(gio->io_gang_leader == gio); |
b128c09f | 2651 | ASSERT(BP_IS_GANG(bp)); |
34dc7c2f | 2652 | |
a6255b7f DQ |
2653 | zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE, |
2654 | zio_gang_tree_assemble_done, gn, gio->io_priority, | |
2655 | ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); | |
b128c09f | 2656 | } |
34dc7c2f | 2657 | |
b128c09f BB |
2658 | static void |
2659 | zio_gang_tree_assemble_done(zio_t *zio) | |
2660 | { | |
9babb374 | 2661 | zio_t *gio = zio->io_gang_leader; |
b128c09f BB |
2662 | zio_gang_node_t *gn = zio->io_private; |
2663 | blkptr_t *bp = zio->io_bp; | |
34dc7c2f | 2664 | |
9babb374 | 2665 | ASSERT(gio == zio_unique_parent(zio)); |
428870ff | 2666 | ASSERT(zio->io_child_count == 0); |
34dc7c2f | 2667 | |
b128c09f BB |
2668 | if (zio->io_error) |
2669 | return; | |
34dc7c2f | 2670 | |
a6255b7f | 2671 | /* this ABD was created from a linear buf in zio_gang_tree_assemble */ |
b128c09f | 2672 | if (BP_SHOULD_BYTESWAP(bp)) |
a6255b7f | 2673 | byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size); |
34dc7c2f | 2674 | |
a6255b7f | 2675 | ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh); |
b128c09f | 2676 | ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); |
428870ff | 2677 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2678 | |
e2af2acc | 2679 | abd_free(zio->io_abd); |
a6255b7f | 2680 | |
1c27024e | 2681 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2682 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2683 | if (!BP_IS_GANG(gbp)) | |
2684 | continue; | |
9babb374 | 2685 | zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); |
b128c09f | 2686 | } |
34dc7c2f BB |
2687 | } |
2688 | ||
b128c09f | 2689 | static void |
a6255b7f DQ |
2690 | zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data, |
2691 | uint64_t offset) | |
34dc7c2f | 2692 | { |
9babb374 | 2693 | zio_t *gio = pio->io_gang_leader; |
b128c09f | 2694 | zio_t *zio; |
34dc7c2f | 2695 | |
b128c09f | 2696 | ASSERT(BP_IS_GANG(bp) == !!gn); |
9babb374 BB |
2697 | ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); |
2698 | ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); | |
34dc7c2f | 2699 | |
b128c09f BB |
2700 | /* |
2701 | * If you're a gang header, your data is in gn->gn_gbh. | |
2702 | * If you're a gang member, your data is in 'data' and gn == NULL. | |
2703 | */ | |
a6255b7f | 2704 | zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset); |
34dc7c2f | 2705 | |
b128c09f | 2706 | if (gn != NULL) { |
428870ff | 2707 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2708 | |
1c27024e | 2709 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2710 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2711 | if (BP_IS_HOLE(gbp)) | |
2712 | continue; | |
a6255b7f DQ |
2713 | zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data, |
2714 | offset); | |
2715 | offset += BP_GET_PSIZE(gbp); | |
b128c09f | 2716 | } |
34dc7c2f BB |
2717 | } |
2718 | ||
9babb374 | 2719 | if (gn == gio->io_gang_tree) |
a6255b7f | 2720 | ASSERT3U(gio->io_size, ==, offset); |
34dc7c2f | 2721 | |
b128c09f BB |
2722 | if (zio != pio) |
2723 | zio_nowait(zio); | |
34dc7c2f BB |
2724 | } |
2725 | ||
62840030 | 2726 | static zio_t * |
b128c09f | 2727 | zio_gang_assemble(zio_t *zio) |
34dc7c2f | 2728 | { |
b128c09f | 2729 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2730 | |
9babb374 BB |
2731 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); |
2732 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
2733 | ||
2734 | zio->io_gang_leader = zio; | |
34dc7c2f | 2735 | |
b128c09f | 2736 | zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); |
34dc7c2f | 2737 | |
62840030 | 2738 | return (zio); |
34dc7c2f BB |
2739 | } |
2740 | ||
62840030 | 2741 | static zio_t * |
b128c09f | 2742 | zio_gang_issue(zio_t *zio) |
34dc7c2f | 2743 | { |
b128c09f | 2744 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2745 | |
ddc751d5 | 2746 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) { |
62840030 | 2747 | return (NULL); |
ddc751d5 | 2748 | } |
34dc7c2f | 2749 | |
9babb374 BB |
2750 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); |
2751 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
34dc7c2f | 2752 | |
b128c09f | 2753 | if (zio->io_child_error[ZIO_CHILD_GANG] == 0) |
a6255b7f DQ |
2754 | zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd, |
2755 | 0); | |
b128c09f | 2756 | else |
9babb374 | 2757 | zio_gang_tree_free(&zio->io_gang_tree); |
34dc7c2f | 2758 | |
b128c09f | 2759 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f | 2760 | |
62840030 | 2761 | return (zio); |
34dc7c2f BB |
2762 | } |
2763 | ||
2764 | static void | |
b128c09f | 2765 | zio_write_gang_member_ready(zio_t *zio) |
34dc7c2f | 2766 | { |
d164b209 | 2767 | zio_t *pio = zio_unique_parent(zio); |
34dc7c2f BB |
2768 | dva_t *cdva = zio->io_bp->blk_dva; |
2769 | dva_t *pdva = pio->io_bp->blk_dva; | |
2770 | uint64_t asize; | |
2a8ba608 | 2771 | zio_t *gio __maybe_unused = zio->io_gang_leader; |
34dc7c2f | 2772 | |
b128c09f BB |
2773 | if (BP_IS_HOLE(zio->io_bp)) |
2774 | return; | |
2775 | ||
2776 | ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); | |
2777 | ||
2778 | ASSERT(zio->io_child_type == ZIO_CHILD_GANG); | |
428870ff BB |
2779 | ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); |
2780 | ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); | |
2781 | ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); | |
34dc7c2f | 2782 | ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); |
34dc7c2f BB |
2783 | |
2784 | mutex_enter(&pio->io_lock); | |
1c27024e | 2785 | for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { |
34dc7c2f BB |
2786 | ASSERT(DVA_GET_GANG(&pdva[d])); |
2787 | asize = DVA_GET_ASIZE(&pdva[d]); | |
2788 | asize += DVA_GET_ASIZE(&cdva[d]); | |
2789 | DVA_SET_ASIZE(&pdva[d], asize); | |
2790 | } | |
2791 | mutex_exit(&pio->io_lock); | |
2792 | } | |
2793 | ||
a6255b7f DQ |
2794 | static void |
2795 | zio_write_gang_done(zio_t *zio) | |
2796 | { | |
c955398b BL |
2797 | /* |
2798 | * The io_abd field will be NULL for a zio with no data. The io_flags | |
2799 | * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't | |
2800 | * check for it here as it is cleared in zio_ready. | |
2801 | */ | |
2802 | if (zio->io_abd != NULL) | |
e2af2acc | 2803 | abd_free(zio->io_abd); |
a6255b7f DQ |
2804 | } |
2805 | ||
62840030 | 2806 | static zio_t * |
aa755b35 | 2807 | zio_write_gang_block(zio_t *pio, metaslab_class_t *mc) |
34dc7c2f | 2808 | { |
b128c09f BB |
2809 | spa_t *spa = pio->io_spa; |
2810 | blkptr_t *bp = pio->io_bp; | |
9babb374 | 2811 | zio_t *gio = pio->io_gang_leader; |
b128c09f BB |
2812 | zio_t *zio; |
2813 | zio_gang_node_t *gn, **gnpp; | |
34dc7c2f | 2814 | zio_gbh_phys_t *gbh; |
a6255b7f | 2815 | abd_t *gbh_abd; |
b128c09f BB |
2816 | uint64_t txg = pio->io_txg; |
2817 | uint64_t resid = pio->io_size; | |
2818 | uint64_t lsize; | |
428870ff | 2819 | int copies = gio->io_prop.zp_copies; |
b5256303 | 2820 | int gbh_copies; |
b128c09f | 2821 | zio_prop_t zp; |
1c27024e | 2822 | int error; |
c955398b | 2823 | boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA); |
b5256303 TC |
2824 | |
2825 | /* | |
2826 | * encrypted blocks need DVA[2] free so encrypted gang headers can't | |
2827 | * have a third copy. | |
2828 | */ | |
2829 | gbh_copies = MIN(copies + 1, spa_max_replication(spa)); | |
2830 | if (gio->io_prop.zp_encrypt && gbh_copies >= SPA_DVAS_PER_BP) | |
2831 | gbh_copies = SPA_DVAS_PER_BP - 1; | |
2832 | ||
1c27024e | 2833 | int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER; |
3dfb57a3 DB |
2834 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
2835 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
c955398b | 2836 | ASSERT(has_data); |
3dfb57a3 DB |
2837 | |
2838 | flags |= METASLAB_ASYNC_ALLOC; | |
f8020c93 AM |
2839 | VERIFY(zfs_refcount_held(&mc->mc_allocator[pio->io_allocator]. |
2840 | mca_alloc_slots, pio)); | |
3dfb57a3 DB |
2841 | |
2842 | /* | |
2843 | * The logical zio has already placed a reservation for | |
2844 | * 'copies' allocation slots but gang blocks may require | |
2845 | * additional copies. These additional copies | |
2846 | * (i.e. gbh_copies - copies) are guaranteed to succeed | |
2847 | * since metaslab_class_throttle_reserve() always allows | |
2848 | * additional reservations for gang blocks. | |
2849 | */ | |
2850 | VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies, | |
492f64e9 | 2851 | pio->io_allocator, pio, flags)); |
3dfb57a3 DB |
2852 | } |
2853 | ||
2854 | error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE, | |
4e21fd06 | 2855 | bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, |
492f64e9 | 2856 | &pio->io_alloc_list, pio, pio->io_allocator); |
34dc7c2f | 2857 | if (error) { |
3dfb57a3 DB |
2858 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
2859 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
c955398b | 2860 | ASSERT(has_data); |
3dfb57a3 DB |
2861 | |
2862 | /* | |
2863 | * If we failed to allocate the gang block header then | |
2864 | * we remove any additional allocation reservations that | |
2865 | * we placed here. The original reservation will | |
2866 | * be removed when the logical I/O goes to the ready | |
2867 | * stage. | |
2868 | */ | |
2869 | metaslab_class_throttle_unreserve(mc, | |
492f64e9 | 2870 | gbh_copies - copies, pio->io_allocator, pio); |
3dfb57a3 DB |
2871 | } |
2872 | ||
b128c09f | 2873 | pio->io_error = error; |
62840030 | 2874 | return (pio); |
34dc7c2f BB |
2875 | } |
2876 | ||
9babb374 BB |
2877 | if (pio == gio) { |
2878 | gnpp = &gio->io_gang_tree; | |
b128c09f BB |
2879 | } else { |
2880 | gnpp = pio->io_private; | |
2881 | ASSERT(pio->io_ready == zio_write_gang_member_ready); | |
34dc7c2f BB |
2882 | } |
2883 | ||
b128c09f BB |
2884 | gn = zio_gang_node_alloc(gnpp); |
2885 | gbh = gn->gn_gbh; | |
2886 | bzero(gbh, SPA_GANGBLOCKSIZE); | |
a6255b7f | 2887 | gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE); |
34dc7c2f | 2888 | |
b128c09f BB |
2889 | /* |
2890 | * Create the gang header. | |
2891 | */ | |
a6255b7f DQ |
2892 | zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE, |
2893 | zio_write_gang_done, NULL, pio->io_priority, | |
2894 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); | |
34dc7c2f | 2895 | |
b128c09f BB |
2896 | /* |
2897 | * Create and nowait the gang children. | |
2898 | */ | |
1c27024e | 2899 | for (int g = 0; resid != 0; resid -= lsize, g++) { |
b128c09f BB |
2900 | lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), |
2901 | SPA_MINBLOCKSIZE); | |
2902 | ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); | |
2903 | ||
9babb374 | 2904 | zp.zp_checksum = gio->io_prop.zp_checksum; |
b128c09f | 2905 | zp.zp_compress = ZIO_COMPRESS_OFF; |
10b3c7f5 | 2906 | zp.zp_complevel = gio->io_prop.zp_complevel; |
b128c09f BB |
2907 | zp.zp_type = DMU_OT_NONE; |
2908 | zp.zp_level = 0; | |
428870ff | 2909 | zp.zp_copies = gio->io_prop.zp_copies; |
03c6040b GW |
2910 | zp.zp_dedup = B_FALSE; |
2911 | zp.zp_dedup_verify = B_FALSE; | |
2912 | zp.zp_nopwrite = B_FALSE; | |
4807c0ba TC |
2913 | zp.zp_encrypt = gio->io_prop.zp_encrypt; |
2914 | zp.zp_byteorder = gio->io_prop.zp_byteorder; | |
b5256303 TC |
2915 | bzero(zp.zp_salt, ZIO_DATA_SALT_LEN); |
2916 | bzero(zp.zp_iv, ZIO_DATA_IV_LEN); | |
2917 | bzero(zp.zp_mac, ZIO_DATA_MAC_LEN); | |
b128c09f | 2918 | |
1c27024e | 2919 | zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g], |
c955398b BL |
2920 | has_data ? abd_get_offset(pio->io_abd, pio->io_size - |
2921 | resid) : NULL, lsize, lsize, &zp, | |
2922 | zio_write_gang_member_ready, NULL, NULL, | |
a6255b7f | 2923 | zio_write_gang_done, &gn->gn_child[g], pio->io_priority, |
3dfb57a3 DB |
2924 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); |
2925 | ||
2926 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { | |
2927 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
c955398b | 2928 | ASSERT(has_data); |
3dfb57a3 DB |
2929 | |
2930 | /* | |
2931 | * Gang children won't throttle but we should | |
2932 | * account for their work, so reserve an allocation | |
2933 | * slot for them here. | |
2934 | */ | |
2935 | VERIFY(metaslab_class_throttle_reserve(mc, | |
492f64e9 | 2936 | zp.zp_copies, cio->io_allocator, cio, flags)); |
3dfb57a3 DB |
2937 | } |
2938 | zio_nowait(cio); | |
b128c09f | 2939 | } |
34dc7c2f BB |
2940 | |
2941 | /* | |
b128c09f | 2942 | * Set pio's pipeline to just wait for zio to finish. |
34dc7c2f | 2943 | */ |
b128c09f BB |
2944 | pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
2945 | ||
920dd524 ED |
2946 | /* |
2947 | * We didn't allocate this bp, so make sure it doesn't get unmarked. | |
2948 | */ | |
2949 | pio->io_flags &= ~ZIO_FLAG_FASTWRITE; | |
2950 | ||
b128c09f BB |
2951 | zio_nowait(zio); |
2952 | ||
62840030 | 2953 | return (pio); |
34dc7c2f BB |
2954 | } |
2955 | ||
03c6040b | 2956 | /* |
3c67d83a TH |
2957 | * The zio_nop_write stage in the pipeline determines if allocating a |
2958 | * new bp is necessary. The nopwrite feature can handle writes in | |
2959 | * either syncing or open context (i.e. zil writes) and as a result is | |
2960 | * mutually exclusive with dedup. | |
2961 | * | |
2962 | * By leveraging a cryptographically secure checksum, such as SHA256, we | |
2963 | * can compare the checksums of the new data and the old to determine if | |
2964 | * allocating a new block is required. Note that our requirements for | |
2965 | * cryptographic strength are fairly weak: there can't be any accidental | |
2966 | * hash collisions, but we don't need to be secure against intentional | |
2967 | * (malicious) collisions. To trigger a nopwrite, you have to be able | |
2968 | * to write the file to begin with, and triggering an incorrect (hash | |
2969 | * collision) nopwrite is no worse than simply writing to the file. | |
2970 | * That said, there are no known attacks against the checksum algorithms | |
2971 | * used for nopwrite, assuming that the salt and the checksums | |
2972 | * themselves remain secret. | |
03c6040b | 2973 | */ |
62840030 | 2974 | static zio_t * |
03c6040b GW |
2975 | zio_nop_write(zio_t *zio) |
2976 | { | |
2977 | blkptr_t *bp = zio->io_bp; | |
2978 | blkptr_t *bp_orig = &zio->io_bp_orig; | |
2979 | zio_prop_t *zp = &zio->io_prop; | |
2980 | ||
2981 | ASSERT(BP_GET_LEVEL(bp) == 0); | |
2982 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
2983 | ASSERT(zp->zp_nopwrite); | |
2984 | ASSERT(!zp->zp_dedup); | |
2985 | ASSERT(zio->io_bp_override == NULL); | |
2986 | ASSERT(IO_IS_ALLOCATING(zio)); | |
2987 | ||
2988 | /* | |
2989 | * Check to see if the original bp and the new bp have matching | |
2990 | * characteristics (i.e. same checksum, compression algorithms, etc). | |
2991 | * If they don't then just continue with the pipeline which will | |
2992 | * allocate a new bp. | |
2993 | */ | |
2994 | if (BP_IS_HOLE(bp_orig) || | |
3c67d83a TH |
2995 | !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & |
2996 | ZCHECKSUM_FLAG_NOPWRITE) || | |
b5256303 | 2997 | BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) || |
03c6040b GW |
2998 | BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || |
2999 | BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || | |
3000 | BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || | |
3001 | zp->zp_copies != BP_GET_NDVAS(bp_orig)) | |
62840030 | 3002 | return (zio); |
03c6040b GW |
3003 | |
3004 | /* | |
3005 | * If the checksums match then reset the pipeline so that we | |
3006 | * avoid allocating a new bp and issuing any I/O. | |
3007 | */ | |
3008 | if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { | |
3c67d83a TH |
3009 | ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & |
3010 | ZCHECKSUM_FLAG_NOPWRITE); | |
03c6040b GW |
3011 | ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); |
3012 | ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); | |
3013 | ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); | |
3014 | ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, | |
3015 | sizeof (uint64_t)) == 0); | |
3016 | ||
681a85cb GW |
3017 | /* |
3018 | * If we're overwriting a block that is currently on an | |
3019 | * indirect vdev, then ignore the nopwrite request and | |
3020 | * allow a new block to be allocated on a concrete vdev. | |
3021 | */ | |
3022 | spa_config_enter(zio->io_spa, SCL_VDEV, FTAG, RW_READER); | |
3023 | vdev_t *tvd = vdev_lookup_top(zio->io_spa, | |
3024 | DVA_GET_VDEV(&bp->blk_dva[0])); | |
3025 | if (tvd->vdev_ops == &vdev_indirect_ops) { | |
3026 | spa_config_exit(zio->io_spa, SCL_VDEV, FTAG); | |
3027 | return (zio); | |
3028 | } | |
3029 | spa_config_exit(zio->io_spa, SCL_VDEV, FTAG); | |
3030 | ||
03c6040b GW |
3031 | *bp = *bp_orig; |
3032 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
3033 | zio->io_flags |= ZIO_FLAG_NOPWRITE; | |
3034 | } | |
3035 | ||
62840030 | 3036 | return (zio); |
03c6040b GW |
3037 | } |
3038 | ||
34dc7c2f BB |
3039 | /* |
3040 | * ========================================================================== | |
428870ff | 3041 | * Dedup |
34dc7c2f BB |
3042 | * ========================================================================== |
3043 | */ | |
428870ff BB |
3044 | static void |
3045 | zio_ddt_child_read_done(zio_t *zio) | |
3046 | { | |
3047 | blkptr_t *bp = zio->io_bp; | |
3048 | ddt_entry_t *dde = zio->io_private; | |
3049 | ddt_phys_t *ddp; | |
3050 | zio_t *pio = zio_unique_parent(zio); | |
3051 | ||
3052 | mutex_enter(&pio->io_lock); | |
3053 | ddp = ddt_phys_select(dde, bp); | |
3054 | if (zio->io_error == 0) | |
3055 | ddt_phys_clear(ddp); /* this ddp doesn't need repair */ | |
a6255b7f DQ |
3056 | |
3057 | if (zio->io_error == 0 && dde->dde_repair_abd == NULL) | |
3058 | dde->dde_repair_abd = zio->io_abd; | |
428870ff | 3059 | else |
a6255b7f | 3060 | abd_free(zio->io_abd); |
428870ff BB |
3061 | mutex_exit(&pio->io_lock); |
3062 | } | |
3063 | ||
62840030 | 3064 | static zio_t * |
428870ff BB |
3065 | zio_ddt_read_start(zio_t *zio) |
3066 | { | |
3067 | blkptr_t *bp = zio->io_bp; | |
3068 | ||
3069 | ASSERT(BP_GET_DEDUP(bp)); | |
3070 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
3071 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
3072 | ||
3073 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
3074 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
3075 | ddt_entry_t *dde = ddt_repair_start(ddt, bp); | |
3076 | ddt_phys_t *ddp = dde->dde_phys; | |
3077 | ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); | |
3078 | blkptr_t blk; | |
3079 | ||
3080 | ASSERT(zio->io_vsd == NULL); | |
3081 | zio->io_vsd = dde; | |
3082 | ||
3083 | if (ddp_self == NULL) | |
62840030 | 3084 | return (zio); |
428870ff | 3085 | |
1c27024e | 3086 | for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { |
428870ff BB |
3087 | if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) |
3088 | continue; | |
3089 | ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, | |
3090 | &blk); | |
3091 | zio_nowait(zio_read(zio, zio->io_spa, &blk, | |
a6255b7f DQ |
3092 | abd_alloc_for_io(zio->io_size, B_TRUE), |
3093 | zio->io_size, zio_ddt_child_read_done, dde, | |
3094 | zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) | | |
3095 | ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark)); | |
428870ff | 3096 | } |
62840030 | 3097 | return (zio); |
428870ff BB |
3098 | } |
3099 | ||
3100 | zio_nowait(zio_read(zio, zio->io_spa, bp, | |
a6255b7f | 3101 | zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority, |
428870ff BB |
3102 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); |
3103 | ||
62840030 | 3104 | return (zio); |
428870ff BB |
3105 | } |
3106 | ||
62840030 | 3107 | static zio_t * |
428870ff BB |
3108 | zio_ddt_read_done(zio_t *zio) |
3109 | { | |
3110 | blkptr_t *bp = zio->io_bp; | |
3111 | ||
ddc751d5 | 3112 | if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) { |
62840030 | 3113 | return (NULL); |
ddc751d5 | 3114 | } |
428870ff BB |
3115 | |
3116 | ASSERT(BP_GET_DEDUP(bp)); | |
3117 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
3118 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
3119 | ||
3120 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
3121 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
3122 | ddt_entry_t *dde = zio->io_vsd; | |
3123 | if (ddt == NULL) { | |
3124 | ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); | |
62840030 | 3125 | return (zio); |
428870ff BB |
3126 | } |
3127 | if (dde == NULL) { | |
3128 | zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; | |
3129 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); | |
62840030 | 3130 | return (NULL); |
428870ff | 3131 | } |
a6255b7f DQ |
3132 | if (dde->dde_repair_abd != NULL) { |
3133 | abd_copy(zio->io_abd, dde->dde_repair_abd, | |
3134 | zio->io_size); | |
428870ff BB |
3135 | zio->io_child_error[ZIO_CHILD_DDT] = 0; |
3136 | } | |
3137 | ddt_repair_done(ddt, dde); | |
3138 | zio->io_vsd = NULL; | |
3139 | } | |
3140 | ||
3141 | ASSERT(zio->io_vsd == NULL); | |
3142 | ||
62840030 | 3143 | return (zio); |
428870ff BB |
3144 | } |
3145 | ||
3146 | static boolean_t | |
3147 | zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) | |
3148 | { | |
3149 | spa_t *spa = zio->io_spa; | |
c17bcf83 | 3150 | boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW); |
428870ff | 3151 | |
c17bcf83 | 3152 | ASSERT(!(zio->io_bp_override && do_raw)); |
2aa34383 | 3153 | |
428870ff BB |
3154 | /* |
3155 | * Note: we compare the original data, not the transformed data, | |
3156 | * because when zio->io_bp is an override bp, we will not have | |
3157 | * pushed the I/O transforms. That's an important optimization | |
3158 | * because otherwise we'd compress/encrypt all dmu_sync() data twice. | |
c17bcf83 | 3159 | * However, we should never get a raw, override zio so in these |
b5256303 | 3160 | * cases we can compare the io_abd directly. This is useful because |
c17bcf83 TC |
3161 | * it allows us to do dedup verification even if we don't have access |
3162 | * to the original data (for instance, if the encryption keys aren't | |
3163 | * loaded). | |
428870ff | 3164 | */ |
c17bcf83 | 3165 | |
1c27024e | 3166 | for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
3167 | zio_t *lio = dde->dde_lead_zio[p]; |
3168 | ||
c17bcf83 TC |
3169 | if (lio != NULL && do_raw) { |
3170 | return (lio->io_size != zio->io_size || | |
a6255b7f | 3171 | abd_cmp(zio->io_abd, lio->io_abd) != 0); |
c17bcf83 | 3172 | } else if (lio != NULL) { |
428870ff | 3173 | return (lio->io_orig_size != zio->io_orig_size || |
a6255b7f | 3174 | abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0); |
428870ff BB |
3175 | } |
3176 | } | |
3177 | ||
1c27024e | 3178 | for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
3179 | ddt_phys_t *ddp = &dde->dde_phys[p]; |
3180 | ||
c17bcf83 TC |
3181 | if (ddp->ddp_phys_birth != 0 && do_raw) { |
3182 | blkptr_t blk = *zio->io_bp; | |
3183 | uint64_t psize; | |
a6255b7f | 3184 | abd_t *tmpabd; |
c17bcf83 TC |
3185 | int error; |
3186 | ||
3187 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
3188 | psize = BP_GET_PSIZE(&blk); | |
3189 | ||
3190 | if (psize != zio->io_size) | |
3191 | return (B_TRUE); | |
3192 | ||
3193 | ddt_exit(ddt); | |
3194 | ||
a6255b7f | 3195 | tmpabd = abd_alloc_for_io(psize, B_TRUE); |
c17bcf83 | 3196 | |
a6255b7f | 3197 | error = zio_wait(zio_read(NULL, spa, &blk, tmpabd, |
c17bcf83 TC |
3198 | psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ, |
3199 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
3200 | ZIO_FLAG_RAW, &zio->io_bookmark)); | |
3201 | ||
3202 | if (error == 0) { | |
a6255b7f | 3203 | if (abd_cmp(tmpabd, zio->io_abd) != 0) |
c17bcf83 TC |
3204 | error = SET_ERROR(ENOENT); |
3205 | } | |
3206 | ||
a6255b7f | 3207 | abd_free(tmpabd); |
c17bcf83 TC |
3208 | ddt_enter(ddt); |
3209 | return (error != 0); | |
3210 | } else if (ddp->ddp_phys_birth != 0) { | |
428870ff | 3211 | arc_buf_t *abuf = NULL; |
2a432414 | 3212 | arc_flags_t aflags = ARC_FLAG_WAIT; |
428870ff BB |
3213 | blkptr_t blk = *zio->io_bp; |
3214 | int error; | |
3215 | ||
3216 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
3217 | ||
c17bcf83 TC |
3218 | if (BP_GET_LSIZE(&blk) != zio->io_orig_size) |
3219 | return (B_TRUE); | |
3220 | ||
428870ff BB |
3221 | ddt_exit(ddt); |
3222 | ||
294f6806 | 3223 | error = arc_read(NULL, spa, &blk, |
428870ff BB |
3224 | arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, |
3225 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
3226 | &aflags, &zio->io_bookmark); | |
3227 | ||
3228 | if (error == 0) { | |
a6255b7f | 3229 | if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data, |
428870ff | 3230 | zio->io_orig_size) != 0) |
c17bcf83 | 3231 | error = SET_ERROR(ENOENT); |
d3c2ae1c | 3232 | arc_buf_destroy(abuf, &abuf); |
428870ff BB |
3233 | } |
3234 | ||
3235 | ddt_enter(ddt); | |
3236 | return (error != 0); | |
3237 | } | |
3238 | } | |
3239 | ||
3240 | return (B_FALSE); | |
3241 | } | |
3242 | ||
3243 | static void | |
3244 | zio_ddt_child_write_ready(zio_t *zio) | |
3245 | { | |
3246 | int p = zio->io_prop.zp_copies; | |
3247 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
3248 | ddt_entry_t *dde = zio->io_private; | |
3249 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
3250 | zio_t *pio; | |
3251 | ||
3252 | if (zio->io_error) | |
3253 | return; | |
3254 | ||
3255 | ddt_enter(ddt); | |
3256 | ||
3257 | ASSERT(dde->dde_lead_zio[p] == zio); | |
3258 | ||
3259 | ddt_phys_fill(ddp, zio->io_bp); | |
3260 | ||
1c27024e | 3261 | zio_link_t *zl = NULL; |
3dfb57a3 | 3262 | while ((pio = zio_walk_parents(zio, &zl)) != NULL) |
428870ff BB |
3263 | ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); |
3264 | ||
3265 | ddt_exit(ddt); | |
3266 | } | |
3267 | ||
3268 | static void | |
3269 | zio_ddt_child_write_done(zio_t *zio) | |
3270 | { | |
3271 | int p = zio->io_prop.zp_copies; | |
3272 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
3273 | ddt_entry_t *dde = zio->io_private; | |
3274 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
3275 | ||
3276 | ddt_enter(ddt); | |
3277 | ||
3278 | ASSERT(ddp->ddp_refcnt == 0); | |
3279 | ASSERT(dde->dde_lead_zio[p] == zio); | |
3280 | dde->dde_lead_zio[p] = NULL; | |
3281 | ||
3282 | if (zio->io_error == 0) { | |
3dfb57a3 DB |
3283 | zio_link_t *zl = NULL; |
3284 | while (zio_walk_parents(zio, &zl) != NULL) | |
428870ff BB |
3285 | ddt_phys_addref(ddp); |
3286 | } else { | |
3287 | ddt_phys_clear(ddp); | |
3288 | } | |
3289 | ||
3290 | ddt_exit(ddt); | |
3291 | } | |
3292 | ||
62840030 | 3293 | static zio_t * |
428870ff BB |
3294 | zio_ddt_write(zio_t *zio) |
3295 | { | |
3296 | spa_t *spa = zio->io_spa; | |
3297 | blkptr_t *bp = zio->io_bp; | |
3298 | uint64_t txg = zio->io_txg; | |
3299 | zio_prop_t *zp = &zio->io_prop; | |
3300 | int p = zp->zp_copies; | |
428870ff | 3301 | zio_t *cio = NULL; |
428870ff BB |
3302 | ddt_t *ddt = ddt_select(spa, bp); |
3303 | ddt_entry_t *dde; | |
3304 | ddt_phys_t *ddp; | |
3305 | ||
3306 | ASSERT(BP_GET_DEDUP(bp)); | |
3307 | ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); | |
3308 | ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); | |
c17bcf83 | 3309 | ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW))); |
428870ff BB |
3310 | |
3311 | ddt_enter(ddt); | |
3312 | dde = ddt_lookup(ddt, bp, B_TRUE); | |
3313 | ddp = &dde->dde_phys[p]; | |
3314 | ||
3315 | if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { | |
3316 | /* | |
3317 | * If we're using a weak checksum, upgrade to a strong checksum | |
3318 | * and try again. If we're already using a strong checksum, | |
3319 | * we can't resolve it, so just convert to an ordinary write. | |
3320 | * (And automatically e-mail a paper to Nature?) | |
3321 | */ | |
3c67d83a TH |
3322 | if (!(zio_checksum_table[zp->zp_checksum].ci_flags & |
3323 | ZCHECKSUM_FLAG_DEDUP)) { | |
428870ff BB |
3324 | zp->zp_checksum = spa_dedup_checksum(spa); |
3325 | zio_pop_transforms(zio); | |
3326 | zio->io_stage = ZIO_STAGE_OPEN; | |
3327 | BP_ZERO(bp); | |
3328 | } else { | |
03c6040b | 3329 | zp->zp_dedup = B_FALSE; |
accd6d9d | 3330 | BP_SET_DEDUP(bp, B_FALSE); |
428870ff | 3331 | } |
accd6d9d | 3332 | ASSERT(!BP_GET_DEDUP(bp)); |
428870ff BB |
3333 | zio->io_pipeline = ZIO_WRITE_PIPELINE; |
3334 | ddt_exit(ddt); | |
62840030 | 3335 | return (zio); |
428870ff BB |
3336 | } |
3337 | ||
428870ff BB |
3338 | if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { |
3339 | if (ddp->ddp_phys_birth != 0) | |
3340 | ddt_bp_fill(ddp, bp, txg); | |
3341 | if (dde->dde_lead_zio[p] != NULL) | |
3342 | zio_add_child(zio, dde->dde_lead_zio[p]); | |
3343 | else | |
3344 | ddt_phys_addref(ddp); | |
3345 | } else if (zio->io_bp_override) { | |
3346 | ASSERT(bp->blk_birth == txg); | |
3347 | ASSERT(BP_EQUAL(bp, zio->io_bp_override)); | |
3348 | ddt_phys_fill(ddp, bp); | |
3349 | ddt_phys_addref(ddp); | |
3350 | } else { | |
a6255b7f | 3351 | cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, |
2aa34383 | 3352 | zio->io_orig_size, zio->io_orig_size, zp, |
bc77ba73 | 3353 | zio_ddt_child_write_ready, NULL, NULL, |
428870ff BB |
3354 | zio_ddt_child_write_done, dde, zio->io_priority, |
3355 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); | |
3356 | ||
a6255b7f | 3357 | zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL); |
428870ff BB |
3358 | dde->dde_lead_zio[p] = cio; |
3359 | } | |
3360 | ||
3361 | ddt_exit(ddt); | |
3362 | ||
9cdf7b1f | 3363 | zio_nowait(cio); |
428870ff | 3364 | |
62840030 | 3365 | return (zio); |
428870ff BB |
3366 | } |
3367 | ||
3368 | ddt_entry_t *freedde; /* for debugging */ | |
b128c09f | 3369 | |
62840030 | 3370 | static zio_t * |
428870ff BB |
3371 | zio_ddt_free(zio_t *zio) |
3372 | { | |
3373 | spa_t *spa = zio->io_spa; | |
3374 | blkptr_t *bp = zio->io_bp; | |
3375 | ddt_t *ddt = ddt_select(spa, bp); | |
3376 | ddt_entry_t *dde; | |
3377 | ddt_phys_t *ddp; | |
3378 | ||
3379 | ASSERT(BP_GET_DEDUP(bp)); | |
3380 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
3381 | ||
3382 | ddt_enter(ddt); | |
3383 | freedde = dde = ddt_lookup(ddt, bp, B_TRUE); | |
5dc6af0e BB |
3384 | if (dde) { |
3385 | ddp = ddt_phys_select(dde, bp); | |
3386 | if (ddp) | |
3387 | ddt_phys_decref(ddp); | |
3388 | } | |
428870ff BB |
3389 | ddt_exit(ddt); |
3390 | ||
62840030 | 3391 | return (zio); |
428870ff BB |
3392 | } |
3393 | ||
3394 | /* | |
3395 | * ========================================================================== | |
3396 | * Allocate and free blocks | |
3397 | * ========================================================================== | |
3398 | */ | |
3dfb57a3 DB |
3399 | |
3400 | static zio_t * | |
492f64e9 | 3401 | zio_io_to_allocate(spa_t *spa, int allocator) |
3dfb57a3 DB |
3402 | { |
3403 | zio_t *zio; | |
3404 | ||
1b50749c | 3405 | ASSERT(MUTEX_HELD(&spa->spa_allocs[allocator].spaa_lock)); |
3dfb57a3 | 3406 | |
1b50749c | 3407 | zio = avl_first(&spa->spa_allocs[allocator].spaa_tree); |
3dfb57a3 DB |
3408 | if (zio == NULL) |
3409 | return (NULL); | |
3410 | ||
3411 | ASSERT(IO_IS_ALLOCATING(zio)); | |
3412 | ||
3413 | /* | |
3414 | * Try to place a reservation for this zio. If we're unable to | |
3415 | * reserve then we throttle. | |
3416 | */ | |
492f64e9 | 3417 | ASSERT3U(zio->io_allocator, ==, allocator); |
cc99f275 | 3418 | if (!metaslab_class_throttle_reserve(zio->io_metaslab_class, |
1b50749c | 3419 | zio->io_prop.zp_copies, allocator, zio, 0)) { |
3dfb57a3 DB |
3420 | return (NULL); |
3421 | } | |
3422 | ||
1b50749c | 3423 | avl_remove(&spa->spa_allocs[allocator].spaa_tree, zio); |
3dfb57a3 DB |
3424 | ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE); |
3425 | ||
3426 | return (zio); | |
3427 | } | |
3428 | ||
62840030 | 3429 | static zio_t * |
3dfb57a3 DB |
3430 | zio_dva_throttle(zio_t *zio) |
3431 | { | |
3432 | spa_t *spa = zio->io_spa; | |
3433 | zio_t *nio; | |
cc99f275 DB |
3434 | metaslab_class_t *mc; |
3435 | ||
3436 | /* locate an appropriate allocation class */ | |
3437 | mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type, | |
3438 | zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk); | |
3dfb57a3 DB |
3439 | |
3440 | if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE || | |
cc99f275 | 3441 | !mc->mc_alloc_throttle_enabled || |
3dfb57a3 DB |
3442 | zio->io_child_type == ZIO_CHILD_GANG || |
3443 | zio->io_flags & ZIO_FLAG_NODATA) { | |
62840030 | 3444 | return (zio); |
3dfb57a3 DB |
3445 | } |
3446 | ||
1b50749c | 3447 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); |
3dfb57a3 | 3448 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); |
3dfb57a3 DB |
3449 | ASSERT3U(zio->io_queued_timestamp, >, 0); |
3450 | ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE); | |
3451 | ||
492f64e9 PD |
3452 | zbookmark_phys_t *bm = &zio->io_bookmark; |
3453 | /* | |
3454 | * We want to try to use as many allocators as possible to help improve | |
3455 | * performance, but we also want logically adjacent IOs to be physically | |
3456 | * adjacent to improve sequential read performance. We chunk each object | |
3457 | * into 2^20 block regions, and then hash based on the objset, object, | |
3458 | * level, and region to accomplish both of these goals. | |
3459 | */ | |
1b50749c | 3460 | int allocator = (uint_t)cityhash4(bm->zb_objset, bm->zb_object, |
492f64e9 | 3461 | bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count; |
1b50749c | 3462 | zio->io_allocator = allocator; |
cc99f275 | 3463 | zio->io_metaslab_class = mc; |
1b50749c AM |
3464 | mutex_enter(&spa->spa_allocs[allocator].spaa_lock); |
3465 | avl_add(&spa->spa_allocs[allocator].spaa_tree, zio); | |
3466 | nio = zio_io_to_allocate(spa, allocator); | |
3467 | mutex_exit(&spa->spa_allocs[allocator].spaa_lock); | |
62840030 | 3468 | return (nio); |
3dfb57a3 DB |
3469 | } |
3470 | ||
cc99f275 | 3471 | static void |
492f64e9 | 3472 | zio_allocate_dispatch(spa_t *spa, int allocator) |
3dfb57a3 DB |
3473 | { |
3474 | zio_t *zio; | |
3475 | ||
1b50749c | 3476 | mutex_enter(&spa->spa_allocs[allocator].spaa_lock); |
492f64e9 | 3477 | zio = zio_io_to_allocate(spa, allocator); |
1b50749c | 3478 | mutex_exit(&spa->spa_allocs[allocator].spaa_lock); |
3dfb57a3 DB |
3479 | if (zio == NULL) |
3480 | return; | |
3481 | ||
3482 | ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE); | |
3483 | ASSERT0(zio->io_error); | |
3484 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE); | |
3485 | } | |
3486 | ||
62840030 | 3487 | static zio_t * |
34dc7c2f BB |
3488 | zio_dva_allocate(zio_t *zio) |
3489 | { | |
3490 | spa_t *spa = zio->io_spa; | |
cc99f275 | 3491 | metaslab_class_t *mc; |
34dc7c2f BB |
3492 | blkptr_t *bp = zio->io_bp; |
3493 | int error; | |
6d974228 | 3494 | int flags = 0; |
34dc7c2f | 3495 | |
9babb374 BB |
3496 | if (zio->io_gang_leader == NULL) { |
3497 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
3498 | zio->io_gang_leader = zio; | |
3499 | } | |
3500 | ||
34dc7c2f | 3501 | ASSERT(BP_IS_HOLE(bp)); |
c99c9001 | 3502 | ASSERT0(BP_GET_NDVAS(bp)); |
428870ff BB |
3503 | ASSERT3U(zio->io_prop.zp_copies, >, 0); |
3504 | ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); | |
34dc7c2f BB |
3505 | ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); |
3506 | ||
920dd524 | 3507 | flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0; |
3dfb57a3 DB |
3508 | if (zio->io_flags & ZIO_FLAG_NODATA) |
3509 | flags |= METASLAB_DONT_THROTTLE; | |
3510 | if (zio->io_flags & ZIO_FLAG_GANG_CHILD) | |
3511 | flags |= METASLAB_GANG_CHILD; | |
3512 | if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) | |
3513 | flags |= METASLAB_ASYNC_ALLOC; | |
3514 | ||
cc99f275 DB |
3515 | /* |
3516 | * if not already chosen, locate an appropriate allocation class | |
3517 | */ | |
3518 | mc = zio->io_metaslab_class; | |
3519 | if (mc == NULL) { | |
3520 | mc = spa_preferred_class(spa, zio->io_size, | |
3521 | zio->io_prop.zp_type, zio->io_prop.zp_level, | |
3522 | zio->io_prop.zp_zpl_smallblk); | |
3523 | zio->io_metaslab_class = mc; | |
3524 | } | |
3525 | ||
aa755b35 MA |
3526 | /* |
3527 | * Try allocating the block in the usual metaslab class. | |
3528 | * If that's full, allocate it in the normal class. | |
3529 | * If that's full, allocate as a gang block, | |
3530 | * and if all are full, the allocation fails (which shouldn't happen). | |
3531 | * | |
3532 | * Note that we do not fall back on embedded slog (ZIL) space, to | |
3533 | * preserve unfragmented slog space, which is critical for decent | |
3534 | * sync write performance. If a log allocation fails, we will fall | |
3535 | * back to spa_sync() which is abysmal for performance. | |
3536 | */ | |
b128c09f | 3537 | error = metaslab_alloc(spa, mc, zio->io_size, bp, |
4e21fd06 | 3538 | zio->io_prop.zp_copies, zio->io_txg, NULL, flags, |
492f64e9 | 3539 | &zio->io_alloc_list, zio, zio->io_allocator); |
34dc7c2f | 3540 | |
cc99f275 DB |
3541 | /* |
3542 | * Fallback to normal class when an alloc class is full | |
3543 | */ | |
3544 | if (error == ENOSPC && mc != spa_normal_class(spa)) { | |
3545 | /* | |
3546 | * If throttling, transfer reservation over to normal class. | |
3547 | * The io_allocator slot can remain the same even though we | |
3548 | * are switching classes. | |
3549 | */ | |
3550 | if (mc->mc_alloc_throttle_enabled && | |
3551 | (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) { | |
3552 | metaslab_class_throttle_unreserve(mc, | |
3553 | zio->io_prop.zp_copies, zio->io_allocator, zio); | |
3554 | zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING; | |
3555 | ||
aa755b35 MA |
3556 | VERIFY(metaslab_class_throttle_reserve( |
3557 | spa_normal_class(spa), | |
cc99f275 DB |
3558 | zio->io_prop.zp_copies, zio->io_allocator, zio, |
3559 | flags | METASLAB_MUST_RESERVE)); | |
cc99f275 | 3560 | } |
aa755b35 MA |
3561 | zio->io_metaslab_class = mc = spa_normal_class(spa); |
3562 | if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) { | |
3563 | zfs_dbgmsg("%s: metaslab allocation failure, " | |
3564 | "trying normal class: zio %px, size %llu, error %d", | |
8e739b2c RE |
3565 | spa_name(spa), zio, (u_longlong_t)zio->io_size, |
3566 | error); | |
aa755b35 | 3567 | } |
cc99f275 DB |
3568 | |
3569 | error = metaslab_alloc(spa, mc, zio->io_size, bp, | |
3570 | zio->io_prop.zp_copies, zio->io_txg, NULL, flags, | |
3571 | &zio->io_alloc_list, zio, zio->io_allocator); | |
3572 | } | |
3573 | ||
aa755b35 MA |
3574 | if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) { |
3575 | if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) { | |
3576 | zfs_dbgmsg("%s: metaslab allocation failure, " | |
3577 | "trying ganging: zio %px, size %llu, error %d", | |
8e739b2c RE |
3578 | spa_name(spa), zio, (u_longlong_t)zio->io_size, |
3579 | error); | |
aa755b35 MA |
3580 | } |
3581 | return (zio_write_gang_block(zio, mc)); | |
3582 | } | |
3dfb57a3 | 3583 | if (error != 0) { |
aa755b35 MA |
3584 | if (error != ENOSPC || |
3585 | (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC)) { | |
3586 | zfs_dbgmsg("%s: metaslab allocation failure: zio %px, " | |
3587 | "size %llu, error %d", | |
8e739b2c RE |
3588 | spa_name(spa), zio, (u_longlong_t)zio->io_size, |
3589 | error); | |
aa755b35 | 3590 | } |
34dc7c2f BB |
3591 | zio->io_error = error; |
3592 | } | |
3593 | ||
62840030 | 3594 | return (zio); |
34dc7c2f BB |
3595 | } |
3596 | ||
62840030 | 3597 | static zio_t * |
34dc7c2f BB |
3598 | zio_dva_free(zio_t *zio) |
3599 | { | |
b128c09f | 3600 | metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); |
34dc7c2f | 3601 | |
62840030 | 3602 | return (zio); |
34dc7c2f BB |
3603 | } |
3604 | ||
62840030 | 3605 | static zio_t * |
34dc7c2f BB |
3606 | zio_dva_claim(zio_t *zio) |
3607 | { | |
b128c09f BB |
3608 | int error; |
3609 | ||
3610 | error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); | |
3611 | if (error) | |
3612 | zio->io_error = error; | |
34dc7c2f | 3613 | |
62840030 | 3614 | return (zio); |
34dc7c2f BB |
3615 | } |
3616 | ||
b128c09f BB |
3617 | /* |
3618 | * Undo an allocation. This is used by zio_done() when an I/O fails | |
3619 | * and we want to give back the block we just allocated. | |
3620 | * This handles both normal blocks and gang blocks. | |
3621 | */ | |
3622 | static void | |
3623 | zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) | |
3624 | { | |
b128c09f | 3625 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); |
428870ff | 3626 | ASSERT(zio->io_bp_override == NULL); |
b128c09f BB |
3627 | |
3628 | if (!BP_IS_HOLE(bp)) | |
428870ff | 3629 | metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); |
b128c09f BB |
3630 | |
3631 | if (gn != NULL) { | |
1c27024e | 3632 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
3633 | zio_dva_unallocate(zio, gn->gn_child[g], |
3634 | &gn->gn_gbh->zg_blkptr[g]); | |
3635 | } | |
3636 | } | |
3637 | } | |
3638 | ||
3639 | /* | |
3640 | * Try to allocate an intent log block. Return 0 on success, errno on failure. | |
3641 | */ | |
3642 | int | |
b5256303 TC |
3643 | zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp, |
3644 | uint64_t size, boolean_t *slog) | |
b128c09f | 3645 | { |
428870ff | 3646 | int error = 1; |
4e21fd06 | 3647 | zio_alloc_list_t io_alloc_list; |
b128c09f | 3648 | |
428870ff BB |
3649 | ASSERT(txg > spa_syncing_txg(spa)); |
3650 | ||
4e21fd06 | 3651 | metaslab_trace_init(&io_alloc_list); |
cc99f275 DB |
3652 | |
3653 | /* | |
3654 | * Block pointer fields are useful to metaslabs for stats and debugging. | |
3655 | * Fill in the obvious ones before calling into metaslab_alloc(). | |
3656 | */ | |
3657 | BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); | |
3658 | BP_SET_PSIZE(new_bp, size); | |
3659 | BP_SET_LEVEL(new_bp, 0); | |
3660 | ||
492f64e9 PD |
3661 | /* |
3662 | * When allocating a zil block, we don't have information about | |
3663 | * the final destination of the block except the objset it's part | |
3664 | * of, so we just hash the objset ID to pick the allocator to get | |
3665 | * some parallelism. | |
3666 | */ | |
be5c6d96 | 3667 | int flags = METASLAB_FASTWRITE | METASLAB_ZIL; |
1b50749c AM |
3668 | int allocator = (uint_t)cityhash4(0, 0, 0, |
3669 | os->os_dsl_dataset->ds_object) % spa->spa_alloc_count; | |
aa755b35 MA |
3670 | error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1, |
3671 | txg, NULL, flags, &io_alloc_list, NULL, allocator); | |
3672 | *slog = (error == 0); | |
3673 | if (error != 0) { | |
3674 | error = metaslab_alloc(spa, spa_embedded_log_class(spa), size, | |
3675 | new_bp, 1, txg, NULL, flags, | |
3676 | &io_alloc_list, NULL, allocator); | |
3677 | } | |
3678 | if (error != 0) { | |
3679 | error = metaslab_alloc(spa, spa_normal_class(spa), size, | |
3680 | new_bp, 1, txg, NULL, flags, | |
3681 | &io_alloc_list, NULL, allocator); | |
ebf8e3a2 | 3682 | } |
4e21fd06 | 3683 | metaslab_trace_fini(&io_alloc_list); |
b128c09f BB |
3684 | |
3685 | if (error == 0) { | |
3686 | BP_SET_LSIZE(new_bp, size); | |
3687 | BP_SET_PSIZE(new_bp, size); | |
3688 | BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); | |
428870ff BB |
3689 | BP_SET_CHECKSUM(new_bp, |
3690 | spa_version(spa) >= SPA_VERSION_SLIM_ZIL | |
3691 | ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); | |
b128c09f BB |
3692 | BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); |
3693 | BP_SET_LEVEL(new_bp, 0); | |
428870ff | 3694 | BP_SET_DEDUP(new_bp, 0); |
b128c09f | 3695 | BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); |
b5256303 TC |
3696 | |
3697 | /* | |
3698 | * encrypted blocks will require an IV and salt. We generate | |
3699 | * these now since we will not be rewriting the bp at | |
3700 | * rewrite time. | |
3701 | */ | |
3702 | if (os->os_encrypted) { | |
3703 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
3704 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
3705 | ||
3706 | BP_SET_CRYPT(new_bp, B_TRUE); | |
3707 | VERIFY0(spa_crypt_get_salt(spa, | |
3708 | dmu_objset_id(os), salt)); | |
3709 | VERIFY0(zio_crypt_generate_iv(iv)); | |
3710 | ||
3711 | zio_crypt_encode_params_bp(new_bp, salt, iv); | |
3712 | } | |
1ce23dca PS |
3713 | } else { |
3714 | zfs_dbgmsg("%s: zil block allocation failure: " | |
8e739b2c RE |
3715 | "size %llu, error %d", spa_name(spa), (u_longlong_t)size, |
3716 | error); | |
b128c09f BB |
3717 | } |
3718 | ||
3719 | return (error); | |
3720 | } | |
3721 | ||
34dc7c2f BB |
3722 | /* |
3723 | * ========================================================================== | |
3724 | * Read and write to physical devices | |
3725 | * ========================================================================== | |
3726 | */ | |
98b25418 | 3727 | |
98b25418 GW |
3728 | /* |
3729 | * Issue an I/O to the underlying vdev. Typically the issue pipeline | |
3730 | * stops after this stage and will resume upon I/O completion. | |
3731 | * However, there are instances where the vdev layer may need to | |
3732 | * continue the pipeline when an I/O was not issued. Since the I/O | |
3733 | * that was sent to the vdev layer might be different than the one | |
3734 | * currently active in the pipeline (see vdev_queue_io()), we explicitly | |
3735 | * force the underlying vdev layers to call either zio_execute() or | |
3736 | * zio_interrupt() to ensure that the pipeline continues with the correct I/O. | |
3737 | */ | |
62840030 | 3738 | static zio_t * |
34dc7c2f BB |
3739 | zio_vdev_io_start(zio_t *zio) |
3740 | { | |
3741 | vdev_t *vd = zio->io_vd; | |
34dc7c2f BB |
3742 | uint64_t align; |
3743 | spa_t *spa = zio->io_spa; | |
3744 | ||
193a37cb TH |
3745 | zio->io_delay = 0; |
3746 | ||
b128c09f BB |
3747 | ASSERT(zio->io_error == 0); |
3748 | ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); | |
34dc7c2f | 3749 | |
b128c09f BB |
3750 | if (vd == NULL) { |
3751 | if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
3752 | spa_config_enter(spa, SCL_ZIO, zio, RW_READER); | |
34dc7c2f | 3753 | |
b128c09f BB |
3754 | /* |
3755 | * The mirror_ops handle multiple DVAs in a single BP. | |
3756 | */ | |
98b25418 | 3757 | vdev_mirror_ops.vdev_op_io_start(zio); |
62840030 | 3758 | return (NULL); |
34dc7c2f BB |
3759 | } |
3760 | ||
3dfb57a3 | 3761 | ASSERT3P(zio->io_logical, !=, zio); |
6cb8e530 PZ |
3762 | if (zio->io_type == ZIO_TYPE_WRITE) { |
3763 | ASSERT(spa->spa_trust_config); | |
3764 | ||
a1d477c2 MA |
3765 | /* |
3766 | * Note: the code can handle other kinds of writes, | |
3767 | * but we don't expect them. | |
3768 | */ | |
2a673e76 | 3769 | if (zio->io_vd->vdev_noalloc) { |
6cb8e530 PZ |
3770 | ASSERT(zio->io_flags & |
3771 | (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL | | |
3772 | ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE)); | |
3773 | } | |
a1d477c2 | 3774 | } |
3dfb57a3 | 3775 | |
b128c09f BB |
3776 | align = 1ULL << vd->vdev_top->vdev_ashift; |
3777 | ||
b02fe35d AR |
3778 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && |
3779 | P2PHASE(zio->io_size, align) != 0) { | |
3780 | /* Transform logical writes to be a full physical block size. */ | |
34dc7c2f | 3781 | uint64_t asize = P2ROUNDUP(zio->io_size, align); |
a6255b7f | 3782 | abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize); |
178e73b3 | 3783 | ASSERT(vd == vd->vdev_top); |
34dc7c2f | 3784 | if (zio->io_type == ZIO_TYPE_WRITE) { |
a6255b7f DQ |
3785 | abd_copy(abuf, zio->io_abd, zio->io_size); |
3786 | abd_zero_off(abuf, zio->io_size, asize - zio->io_size); | |
34dc7c2f | 3787 | } |
b128c09f | 3788 | zio_push_transform(zio, abuf, asize, asize, zio_subblock); |
34dc7c2f BB |
3789 | } |
3790 | ||
b02fe35d AR |
3791 | /* |
3792 | * If this is not a physical io, make sure that it is properly aligned | |
3793 | * before proceeding. | |
3794 | */ | |
3795 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { | |
3796 | ASSERT0(P2PHASE(zio->io_offset, align)); | |
3797 | ASSERT0(P2PHASE(zio->io_size, align)); | |
3798 | } else { | |
3799 | /* | |
3800 | * For physical writes, we allow 512b aligned writes and assume | |
3801 | * the device will perform a read-modify-write as necessary. | |
3802 | */ | |
3803 | ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); | |
3804 | ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); | |
3805 | } | |
3806 | ||
572e2857 | 3807 | VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); |
fb5f0bc8 BB |
3808 | |
3809 | /* | |
3810 | * If this is a repair I/O, and there's no self-healing involved -- | |
3811 | * that is, we're just resilvering what we expect to resilver -- | |
3812 | * then don't do the I/O unless zio's txg is actually in vd's DTL. | |
9e052db4 MA |
3813 | * This prevents spurious resilvering. |
3814 | * | |
3815 | * There are a few ways that we can end up creating these spurious | |
3816 | * resilver i/os: | |
3817 | * | |
3818 | * 1. A resilver i/o will be issued if any DVA in the BP has a | |
3819 | * dirty DTL. The mirror code will issue resilver writes to | |
3820 | * each DVA, including the one(s) that are not on vdevs with dirty | |
3821 | * DTLs. | |
3822 | * | |
3823 | * 2. With nested replication, which happens when we have a | |
3824 | * "replacing" or "spare" vdev that's a child of a mirror or raidz. | |
3825 | * For example, given mirror(replacing(A+B), C), it's likely that | |
3826 | * only A is out of date (it's the new device). In this case, we'll | |
3827 | * read from C, then use the data to resilver A+B -- but we don't | |
3828 | * actually want to resilver B, just A. The top-level mirror has no | |
3829 | * way to know this, so instead we just discard unnecessary repairs | |
3830 | * as we work our way down the vdev tree. | |
3831 | * | |
3832 | * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc. | |
3833 | * The same logic applies to any form of nested replication: ditto | |
3834 | * + mirror, RAID-Z + replacing, etc. | |
3835 | * | |
3836 | * However, indirect vdevs point off to other vdevs which may have | |
3837 | * DTL's, so we never bypass them. The child i/os on concrete vdevs | |
3838 | * will be properly bypassed instead. | |
b2255edc BB |
3839 | * |
3840 | * Leaf DTL_PARTIAL can be empty when a legitimate write comes from | |
3841 | * a dRAID spare vdev. For example, when a dRAID spare is first | |
3842 | * used, its spare blocks need to be written to but the leaf vdev's | |
3843 | * of such blocks can have empty DTL_PARTIAL. | |
3844 | * | |
3845 | * There seemed no clean way to allow such writes while bypassing | |
3846 | * spurious ones. At this point, just avoid all bypassing for dRAID | |
3847 | * for correctness. | |
fb5f0bc8 BB |
3848 | */ |
3849 | if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && | |
3850 | !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && | |
3851 | zio->io_txg != 0 && /* not a delegated i/o */ | |
9e052db4 | 3852 | vd->vdev_ops != &vdev_indirect_ops && |
b2255edc | 3853 | vd->vdev_top->vdev_ops != &vdev_draid_ops && |
fb5f0bc8 BB |
3854 | !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { |
3855 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
fb5f0bc8 | 3856 | zio_vdev_io_bypass(zio); |
62840030 | 3857 | return (zio); |
fb5f0bc8 | 3858 | } |
34dc7c2f | 3859 | |
b2255edc BB |
3860 | /* |
3861 | * Select the next best leaf I/O to process. Distributed spares are | |
3862 | * excluded since they dispatch the I/O directly to a leaf vdev after | |
3863 | * applying the dRAID mapping. | |
3864 | */ | |
3865 | if (vd->vdev_ops->vdev_op_leaf && | |
3866 | vd->vdev_ops != &vdev_draid_spare_ops && | |
3867 | (zio->io_type == ZIO_TYPE_READ || | |
3868 | zio->io_type == ZIO_TYPE_WRITE || | |
3869 | zio->io_type == ZIO_TYPE_TRIM)) { | |
b128c09f | 3870 | |
b0bc7a84 | 3871 | if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio)) |
62840030 | 3872 | return (zio); |
b128c09f BB |
3873 | |
3874 | if ((zio = vdev_queue_io(zio)) == NULL) | |
62840030 | 3875 | return (NULL); |
b128c09f BB |
3876 | |
3877 | if (!vdev_accessible(vd, zio)) { | |
2e528b49 | 3878 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f | 3879 | zio_interrupt(zio); |
62840030 | 3880 | return (NULL); |
b128c09f | 3881 | } |
67103816 | 3882 | zio->io_delay = gethrtime(); |
b128c09f BB |
3883 | } |
3884 | ||
98b25418 | 3885 | vd->vdev_ops->vdev_op_io_start(zio); |
62840030 | 3886 | return (NULL); |
34dc7c2f BB |
3887 | } |
3888 | ||
62840030 | 3889 | static zio_t * |
34dc7c2f BB |
3890 | zio_vdev_io_done(zio_t *zio) |
3891 | { | |
b128c09f BB |
3892 | vdev_t *vd = zio->io_vd; |
3893 | vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; | |
3894 | boolean_t unexpected_error = B_FALSE; | |
34dc7c2f | 3895 | |
ddc751d5 | 3896 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { |
62840030 | 3897 | return (NULL); |
ddc751d5 | 3898 | } |
34dc7c2f | 3899 | |
1b939560 BB |
3900 | ASSERT(zio->io_type == ZIO_TYPE_READ || |
3901 | zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM); | |
b128c09f | 3902 | |
193a37cb TH |
3903 | if (zio->io_delay) |
3904 | zio->io_delay = gethrtime() - zio->io_delay; | |
3905 | ||
b2255edc BB |
3906 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf && |
3907 | vd->vdev_ops != &vdev_draid_spare_ops) { | |
b128c09f BB |
3908 | vdev_queue_io_done(zio); |
3909 | ||
3910 | if (zio->io_type == ZIO_TYPE_WRITE) | |
3911 | vdev_cache_write(zio); | |
3912 | ||
3913 | if (zio_injection_enabled && zio->io_error == 0) | |
d977122d DB |
3914 | zio->io_error = zio_handle_device_injections(vd, zio, |
3915 | EIO, EILSEQ); | |
b128c09f BB |
3916 | |
3917 | if (zio_injection_enabled && zio->io_error == 0) | |
3918 | zio->io_error = zio_handle_label_injection(zio, EIO); | |
3919 | ||
1b939560 | 3920 | if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) { |
b128c09f | 3921 | if (!vdev_accessible(vd, zio)) { |
2e528b49 | 3922 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
3923 | } else { |
3924 | unexpected_error = B_TRUE; | |
3925 | } | |
3926 | } | |
3927 | } | |
3928 | ||
3929 | ops->vdev_op_io_done(zio); | |
34dc7c2f | 3930 | |
f43615d0 | 3931 | if (unexpected_error) |
d164b209 | 3932 | VERIFY(vdev_probe(vd, zio) == NULL); |
34dc7c2f | 3933 | |
62840030 | 3934 | return (zio); |
34dc7c2f BB |
3935 | } |
3936 | ||
a8b2e306 TC |
3937 | /* |
3938 | * This function is used to change the priority of an existing zio that is | |
3939 | * currently in-flight. This is used by the arc to upgrade priority in the | |
3940 | * event that a demand read is made for a block that is currently queued | |
3941 | * as a scrub or async read IO. Otherwise, the high priority read request | |
3942 | * would end up having to wait for the lower priority IO. | |
3943 | */ | |
3944 | void | |
3945 | zio_change_priority(zio_t *pio, zio_priority_t priority) | |
3946 | { | |
3947 | zio_t *cio, *cio_next; | |
3948 | zio_link_t *zl = NULL; | |
3949 | ||
3950 | ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
3951 | ||
3952 | if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) { | |
3953 | vdev_queue_change_io_priority(pio, priority); | |
3954 | } else { | |
3955 | pio->io_priority = priority; | |
3956 | } | |
3957 | ||
3958 | mutex_enter(&pio->io_lock); | |
3959 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { | |
3960 | cio_next = zio_walk_children(pio, &zl); | |
3961 | zio_change_priority(cio, priority); | |
3962 | } | |
3963 | mutex_exit(&pio->io_lock); | |
3964 | } | |
3965 | ||
428870ff BB |
3966 | /* |
3967 | * For non-raidz ZIOs, we can just copy aside the bad data read from the | |
3968 | * disk, and use that to finish the checksum ereport later. | |
3969 | */ | |
3970 | static void | |
3971 | zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, | |
84c07ada | 3972 | const abd_t *good_buf) |
428870ff BB |
3973 | { |
3974 | /* no processing needed */ | |
3975 | zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); | |
3976 | } | |
3977 | ||
428870ff | 3978 | void |
330c6c05 | 3979 | zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr) |
428870ff | 3980 | { |
84c07ada | 3981 | void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size); |
428870ff | 3982 | |
84c07ada | 3983 | abd_copy(abd, zio->io_abd, zio->io_size); |
428870ff BB |
3984 | |
3985 | zcr->zcr_cbinfo = zio->io_size; | |
84c07ada | 3986 | zcr->zcr_cbdata = abd; |
428870ff | 3987 | zcr->zcr_finish = zio_vsd_default_cksum_finish; |
84c07ada | 3988 | zcr->zcr_free = zio_abd_free; |
428870ff BB |
3989 | } |
3990 | ||
62840030 | 3991 | static zio_t * |
34dc7c2f BB |
3992 | zio_vdev_io_assess(zio_t *zio) |
3993 | { | |
3994 | vdev_t *vd = zio->io_vd; | |
b128c09f | 3995 | |
ddc751d5 | 3996 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { |
62840030 | 3997 | return (NULL); |
ddc751d5 | 3998 | } |
b128c09f BB |
3999 | |
4000 | if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
4001 | spa_config_exit(zio->io_spa, SCL_ZIO, zio); | |
4002 | ||
4003 | if (zio->io_vsd != NULL) { | |
428870ff | 4004 | zio->io_vsd_ops->vsd_free(zio); |
b128c09f | 4005 | zio->io_vsd = NULL; |
34dc7c2f BB |
4006 | } |
4007 | ||
b128c09f | 4008 | if (zio_injection_enabled && zio->io_error == 0) |
34dc7c2f BB |
4009 | zio->io_error = zio_handle_fault_injection(zio, EIO); |
4010 | ||
4011 | /* | |
4012 | * If the I/O failed, determine whether we should attempt to retry it. | |
428870ff BB |
4013 | * |
4014 | * On retry, we cut in line in the issue queue, since we don't want | |
4015 | * compression/checksumming/etc. work to prevent our (cheap) IO reissue. | |
34dc7c2f | 4016 | */ |
b128c09f BB |
4017 | if (zio->io_error && vd == NULL && |
4018 | !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { | |
4019 | ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ | |
4020 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ | |
34dc7c2f | 4021 | zio->io_error = 0; |
b128c09f BB |
4022 | zio->io_flags |= ZIO_FLAG_IO_RETRY | |
4023 | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; | |
428870ff BB |
4024 | zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; |
4025 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, | |
4026 | zio_requeue_io_start_cut_in_line); | |
62840030 | 4027 | return (NULL); |
34dc7c2f BB |
4028 | } |
4029 | ||
b128c09f BB |
4030 | /* |
4031 | * If we got an error on a leaf device, convert it to ENXIO | |
4032 | * if the device is not accessible at all. | |
4033 | */ | |
4034 | if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && | |
4035 | !vdev_accessible(vd, zio)) | |
2e528b49 | 4036 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
4037 | |
4038 | /* | |
4039 | * If we can't write to an interior vdev (mirror or RAID-Z), | |
4040 | * set vdev_cant_write so that we stop trying to allocate from it. | |
4041 | */ | |
4042 | if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && | |
13fe0198 | 4043 | vd != NULL && !vd->vdev_ops->vdev_op_leaf) { |
2b56a634 MA |
4044 | vdev_dbgmsg(vd, "zio_vdev_io_assess(zio=%px) setting " |
4045 | "cant_write=TRUE due to write failure with ENXIO", | |
4046 | zio); | |
b128c09f | 4047 | vd->vdev_cant_write = B_TRUE; |
13fe0198 | 4048 | } |
b128c09f | 4049 | |
298ec40b GM |
4050 | /* |
4051 | * If a cache flush returns ENOTSUP or ENOTTY, we know that no future | |
1b939560 BB |
4052 | * attempts will ever succeed. In this case we set a persistent |
4053 | * boolean flag so that we don't bother with it in the future. | |
298ec40b GM |
4054 | */ |
4055 | if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) && | |
4056 | zio->io_type == ZIO_TYPE_IOCTL && | |
4057 | zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL) | |
4058 | vd->vdev_nowritecache = B_TRUE; | |
4059 | ||
b128c09f BB |
4060 | if (zio->io_error) |
4061 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
4062 | ||
e8b96c60 MA |
4063 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf && |
4064 | zio->io_physdone != NULL) { | |
4065 | ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); | |
4066 | ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); | |
4067 | zio->io_physdone(zio->io_logical); | |
4068 | } | |
4069 | ||
62840030 | 4070 | return (zio); |
34dc7c2f BB |
4071 | } |
4072 | ||
4073 | void | |
4074 | zio_vdev_io_reissue(zio_t *zio) | |
4075 | { | |
4076 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
4077 | ASSERT(zio->io_error == 0); | |
4078 | ||
428870ff | 4079 | zio->io_stage >>= 1; |
34dc7c2f BB |
4080 | } |
4081 | ||
4082 | void | |
4083 | zio_vdev_io_redone(zio_t *zio) | |
4084 | { | |
4085 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); | |
4086 | ||
428870ff | 4087 | zio->io_stage >>= 1; |
34dc7c2f BB |
4088 | } |
4089 | ||
4090 | void | |
4091 | zio_vdev_io_bypass(zio_t *zio) | |
4092 | { | |
4093 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
4094 | ASSERT(zio->io_error == 0); | |
4095 | ||
4096 | zio->io_flags |= ZIO_FLAG_IO_BYPASS; | |
428870ff | 4097 | zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; |
34dc7c2f BB |
4098 | } |
4099 | ||
b5256303 TC |
4100 | /* |
4101 | * ========================================================================== | |
4102 | * Encrypt and store encryption parameters | |
4103 | * ========================================================================== | |
4104 | */ | |
4105 | ||
4106 | ||
4107 | /* | |
4108 | * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for | |
4109 | * managing the storage of encryption parameters and passing them to the | |
4110 | * lower-level encryption functions. | |
4111 | */ | |
62840030 | 4112 | static zio_t * |
b5256303 TC |
4113 | zio_encrypt(zio_t *zio) |
4114 | { | |
4115 | zio_prop_t *zp = &zio->io_prop; | |
4116 | spa_t *spa = zio->io_spa; | |
4117 | blkptr_t *bp = zio->io_bp; | |
4118 | uint64_t psize = BP_GET_PSIZE(bp); | |
ae76f45c | 4119 | uint64_t dsobj = zio->io_bookmark.zb_objset; |
b5256303 TC |
4120 | dmu_object_type_t ot = BP_GET_TYPE(bp); |
4121 | void *enc_buf = NULL; | |
4122 | abd_t *eabd = NULL; | |
4123 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
4124 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
4125 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
4126 | boolean_t no_crypt = B_FALSE; | |
4127 | ||
4128 | /* the root zio already encrypted the data */ | |
4129 | if (zio->io_child_type == ZIO_CHILD_GANG) | |
62840030 | 4130 | return (zio); |
b5256303 TC |
4131 | |
4132 | /* only ZIL blocks are re-encrypted on rewrite */ | |
4133 | if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG) | |
62840030 | 4134 | return (zio); |
b5256303 TC |
4135 | |
4136 | if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) { | |
4137 | BP_SET_CRYPT(bp, B_FALSE); | |
62840030 | 4138 | return (zio); |
b5256303 TC |
4139 | } |
4140 | ||
4141 | /* if we are doing raw encryption set the provided encryption params */ | |
4142 | if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) { | |
ae76f45c | 4143 | ASSERT0(BP_GET_LEVEL(bp)); |
b5256303 TC |
4144 | BP_SET_CRYPT(bp, B_TRUE); |
4145 | BP_SET_BYTEORDER(bp, zp->zp_byteorder); | |
4146 | if (ot != DMU_OT_OBJSET) | |
4147 | zio_crypt_encode_mac_bp(bp, zp->zp_mac); | |
ae76f45c TC |
4148 | |
4149 | /* dnode blocks must be written out in the provided byteorder */ | |
4150 | if (zp->zp_byteorder != ZFS_HOST_BYTEORDER && | |
4151 | ot == DMU_OT_DNODE) { | |
4152 | void *bswap_buf = zio_buf_alloc(psize); | |
4153 | abd_t *babd = abd_get_from_buf(bswap_buf, psize); | |
4154 | ||
4155 | ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); | |
4156 | abd_copy_to_buf(bswap_buf, zio->io_abd, psize); | |
4157 | dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf, | |
4158 | psize); | |
4159 | ||
4160 | abd_take_ownership_of_buf(babd, B_TRUE); | |
4161 | zio_push_transform(zio, babd, psize, psize, NULL); | |
4162 | } | |
4163 | ||
b5256303 TC |
4164 | if (DMU_OT_IS_ENCRYPTED(ot)) |
4165 | zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv); | |
62840030 | 4166 | return (zio); |
b5256303 TC |
4167 | } |
4168 | ||
4169 | /* indirect blocks only maintain a cksum of the lower level MACs */ | |
4170 | if (BP_GET_LEVEL(bp) > 0) { | |
4171 | BP_SET_CRYPT(bp, B_TRUE); | |
4172 | VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE, | |
4173 | zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp), | |
4174 | mac)); | |
4175 | zio_crypt_encode_mac_bp(bp, mac); | |
62840030 | 4176 | return (zio); |
b5256303 TC |
4177 | } |
4178 | ||
4179 | /* | |
4180 | * Objset blocks are a special case since they have 2 256-bit MACs | |
4181 | * embedded within them. | |
4182 | */ | |
4183 | if (ot == DMU_OT_OBJSET) { | |
4184 | ASSERT0(DMU_OT_IS_ENCRYPTED(ot)); | |
4185 | ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); | |
4186 | BP_SET_CRYPT(bp, B_TRUE); | |
ae76f45c TC |
4187 | VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj, |
4188 | zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp))); | |
62840030 | 4189 | return (zio); |
b5256303 TC |
4190 | } |
4191 | ||
4192 | /* unencrypted object types are only authenticated with a MAC */ | |
4193 | if (!DMU_OT_IS_ENCRYPTED(ot)) { | |
4194 | BP_SET_CRYPT(bp, B_TRUE); | |
ae76f45c TC |
4195 | VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj, |
4196 | zio->io_abd, psize, mac)); | |
b5256303 | 4197 | zio_crypt_encode_mac_bp(bp, mac); |
62840030 | 4198 | return (zio); |
b5256303 TC |
4199 | } |
4200 | ||
4201 | /* | |
4202 | * Later passes of sync-to-convergence may decide to rewrite data | |
4203 | * in place to avoid more disk reallocations. This presents a problem | |
d611989f | 4204 | * for encryption because this constitutes rewriting the new data with |
b5256303 TC |
4205 | * the same encryption key and IV. However, this only applies to blocks |
4206 | * in the MOS (particularly the spacemaps) and we do not encrypt the | |
4207 | * MOS. We assert that the zio is allocating or an intent log write | |
4208 | * to enforce this. | |
4209 | */ | |
4210 | ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG); | |
4211 | ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG); | |
4212 | ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION)); | |
4213 | ASSERT3U(psize, !=, 0); | |
4214 | ||
4215 | enc_buf = zio_buf_alloc(psize); | |
4216 | eabd = abd_get_from_buf(enc_buf, psize); | |
4217 | abd_take_ownership_of_buf(eabd, B_TRUE); | |
4218 | ||
4219 | /* | |
4220 | * For an explanation of what encryption parameters are stored | |
4221 | * where, see the block comment in zio_crypt.c. | |
4222 | */ | |
4223 | if (ot == DMU_OT_INTENT_LOG) { | |
4224 | zio_crypt_decode_params_bp(bp, salt, iv); | |
4225 | } else { | |
4226 | BP_SET_CRYPT(bp, B_TRUE); | |
4227 | } | |
4228 | ||
4229 | /* Perform the encryption. This should not fail */ | |
be9a5c35 TC |
4230 | VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark, |
4231 | BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), | |
4232 | salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt)); | |
b5256303 TC |
4233 | |
4234 | /* encode encryption metadata into the bp */ | |
4235 | if (ot == DMU_OT_INTENT_LOG) { | |
4236 | /* | |
4237 | * ZIL blocks store the MAC in the embedded checksum, so the | |
4238 | * transform must always be applied. | |
4239 | */ | |
4240 | zio_crypt_encode_mac_zil(enc_buf, mac); | |
4241 | zio_push_transform(zio, eabd, psize, psize, NULL); | |
4242 | } else { | |
4243 | BP_SET_CRYPT(bp, B_TRUE); | |
4244 | zio_crypt_encode_params_bp(bp, salt, iv); | |
4245 | zio_crypt_encode_mac_bp(bp, mac); | |
4246 | ||
4247 | if (no_crypt) { | |
4248 | ASSERT3U(ot, ==, DMU_OT_DNODE); | |
4249 | abd_free(eabd); | |
4250 | } else { | |
4251 | zio_push_transform(zio, eabd, psize, psize, NULL); | |
4252 | } | |
4253 | } | |
4254 | ||
62840030 | 4255 | return (zio); |
b5256303 TC |
4256 | } |
4257 | ||
34dc7c2f BB |
4258 | /* |
4259 | * ========================================================================== | |
4260 | * Generate and verify checksums | |
4261 | * ========================================================================== | |
4262 | */ | |
62840030 | 4263 | static zio_t * |
34dc7c2f BB |
4264 | zio_checksum_generate(zio_t *zio) |
4265 | { | |
34dc7c2f | 4266 | blkptr_t *bp = zio->io_bp; |
b128c09f | 4267 | enum zio_checksum checksum; |
34dc7c2f | 4268 | |
b128c09f BB |
4269 | if (bp == NULL) { |
4270 | /* | |
4271 | * This is zio_write_phys(). | |
4272 | * We're either generating a label checksum, or none at all. | |
4273 | */ | |
4274 | checksum = zio->io_prop.zp_checksum; | |
34dc7c2f | 4275 | |
b128c09f | 4276 | if (checksum == ZIO_CHECKSUM_OFF) |
62840030 | 4277 | return (zio); |
b128c09f BB |
4278 | |
4279 | ASSERT(checksum == ZIO_CHECKSUM_LABEL); | |
4280 | } else { | |
4281 | if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { | |
4282 | ASSERT(!IO_IS_ALLOCATING(zio)); | |
4283 | checksum = ZIO_CHECKSUM_GANG_HEADER; | |
4284 | } else { | |
4285 | checksum = BP_GET_CHECKSUM(bp); | |
4286 | } | |
4287 | } | |
34dc7c2f | 4288 | |
a6255b7f | 4289 | zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size); |
34dc7c2f | 4290 | |
62840030 | 4291 | return (zio); |
34dc7c2f BB |
4292 | } |
4293 | ||
62840030 | 4294 | static zio_t * |
b128c09f | 4295 | zio_checksum_verify(zio_t *zio) |
34dc7c2f | 4296 | { |
428870ff | 4297 | zio_bad_cksum_t info; |
b128c09f BB |
4298 | blkptr_t *bp = zio->io_bp; |
4299 | int error; | |
34dc7c2f | 4300 | |
428870ff BB |
4301 | ASSERT(zio->io_vd != NULL); |
4302 | ||
b128c09f BB |
4303 | if (bp == NULL) { |
4304 | /* | |
4305 | * This is zio_read_phys(). | |
4306 | * We're either verifying a label checksum, or nothing at all. | |
4307 | */ | |
4308 | if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) | |
62840030 | 4309 | return (zio); |
34dc7c2f | 4310 | |
b2255edc | 4311 | ASSERT3U(zio->io_prop.zp_checksum, ==, ZIO_CHECKSUM_LABEL); |
b128c09f | 4312 | } |
34dc7c2f | 4313 | |
428870ff | 4314 | if ((error = zio_checksum_error(zio, &info)) != 0) { |
b128c09f | 4315 | zio->io_error = error; |
7a3066ff MA |
4316 | if (error == ECKSUM && |
4317 | !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { | |
03e02e5b | 4318 | (void) zfs_ereport_start_checksum(zio->io_spa, |
b5256303 | 4319 | zio->io_vd, &zio->io_bookmark, zio, |
330c6c05 | 4320 | zio->io_offset, zio->io_size, &info); |
03e02e5b DB |
4321 | mutex_enter(&zio->io_vd->vdev_stat_lock); |
4322 | zio->io_vd->vdev_stat.vs_checksum_errors++; | |
4323 | mutex_exit(&zio->io_vd->vdev_stat_lock); | |
b128c09f | 4324 | } |
34dc7c2f BB |
4325 | } |
4326 | ||
62840030 | 4327 | return (zio); |
34dc7c2f BB |
4328 | } |
4329 | ||
4330 | /* | |
4331 | * Called by RAID-Z to ensure we don't compute the checksum twice. | |
4332 | */ | |
4333 | void | |
4334 | zio_checksum_verified(zio_t *zio) | |
4335 | { | |
428870ff | 4336 | zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; |
34dc7c2f BB |
4337 | } |
4338 | ||
4339 | /* | |
b128c09f BB |
4340 | * ========================================================================== |
4341 | * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. | |
9b67f605 | 4342 | * An error of 0 indicates success. ENXIO indicates whole-device failure, |
d611989f | 4343 | * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO |
b128c09f BB |
4344 | * indicate errors that are specific to one I/O, and most likely permanent. |
4345 | * Any other error is presumed to be worse because we weren't expecting it. | |
4346 | * ========================================================================== | |
34dc7c2f | 4347 | */ |
b128c09f BB |
4348 | int |
4349 | zio_worst_error(int e1, int e2) | |
34dc7c2f | 4350 | { |
b128c09f BB |
4351 | static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; |
4352 | int r1, r2; | |
4353 | ||
4354 | for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) | |
4355 | if (e1 == zio_error_rank[r1]) | |
4356 | break; | |
34dc7c2f | 4357 | |
b128c09f BB |
4358 | for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) |
4359 | if (e2 == zio_error_rank[r2]) | |
4360 | break; | |
4361 | ||
4362 | return (r1 > r2 ? e1 : e2); | |
34dc7c2f BB |
4363 | } |
4364 | ||
4365 | /* | |
4366 | * ========================================================================== | |
b128c09f | 4367 | * I/O completion |
34dc7c2f BB |
4368 | * ========================================================================== |
4369 | */ | |
62840030 | 4370 | static zio_t * |
b128c09f | 4371 | zio_ready(zio_t *zio) |
34dc7c2f | 4372 | { |
b128c09f | 4373 | blkptr_t *bp = zio->io_bp; |
d164b209 | 4374 | zio_t *pio, *pio_next; |
3dfb57a3 | 4375 | zio_link_t *zl = NULL; |
34dc7c2f | 4376 | |
ddc751d5 GW |
4377 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT, |
4378 | ZIO_WAIT_READY)) { | |
62840030 | 4379 | return (NULL); |
ddc751d5 | 4380 | } |
34dc7c2f | 4381 | |
9babb374 | 4382 | if (zio->io_ready) { |
b128c09f | 4383 | ASSERT(IO_IS_ALLOCATING(zio)); |
03c6040b GW |
4384 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || |
4385 | (zio->io_flags & ZIO_FLAG_NOPWRITE)); | |
b128c09f | 4386 | ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); |
34dc7c2f | 4387 | |
b128c09f BB |
4388 | zio->io_ready(zio); |
4389 | } | |
34dc7c2f | 4390 | |
b128c09f BB |
4391 | if (bp != NULL && bp != &zio->io_bp_copy) |
4392 | zio->io_bp_copy = *bp; | |
34dc7c2f | 4393 | |
3dfb57a3 | 4394 | if (zio->io_error != 0) { |
b128c09f | 4395 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f | 4396 | |
3dfb57a3 DB |
4397 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
4398 | ASSERT(IO_IS_ALLOCATING(zio)); | |
4399 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
cc99f275 DB |
4400 | ASSERT(zio->io_metaslab_class != NULL); |
4401 | ||
3dfb57a3 DB |
4402 | /* |
4403 | * We were unable to allocate anything, unreserve and | |
4404 | * issue the next I/O to allocate. | |
4405 | */ | |
4406 | metaslab_class_throttle_unreserve( | |
cc99f275 DB |
4407 | zio->io_metaslab_class, zio->io_prop.zp_copies, |
4408 | zio->io_allocator, zio); | |
492f64e9 | 4409 | zio_allocate_dispatch(zio->io_spa, zio->io_allocator); |
3dfb57a3 DB |
4410 | } |
4411 | } | |
4412 | ||
d164b209 BB |
4413 | mutex_enter(&zio->io_lock); |
4414 | zio->io_state[ZIO_WAIT_READY] = 1; | |
3dfb57a3 | 4415 | pio = zio_walk_parents(zio, &zl); |
d164b209 BB |
4416 | mutex_exit(&zio->io_lock); |
4417 | ||
4418 | /* | |
4419 | * As we notify zio's parents, new parents could be added. | |
4420 | * New parents go to the head of zio's io_parent_list, however, | |
4421 | * so we will (correctly) not notify them. The remainder of zio's | |
4422 | * io_parent_list, from 'pio_next' onward, cannot change because | |
4423 | * all parents must wait for us to be done before they can be done. | |
4424 | */ | |
4425 | for (; pio != NULL; pio = pio_next) { | |
3dfb57a3 | 4426 | pio_next = zio_walk_parents(zio, &zl); |
62840030 | 4427 | zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL); |
d164b209 | 4428 | } |
34dc7c2f | 4429 | |
428870ff BB |
4430 | if (zio->io_flags & ZIO_FLAG_NODATA) { |
4431 | if (BP_IS_GANG(bp)) { | |
4432 | zio->io_flags &= ~ZIO_FLAG_NODATA; | |
4433 | } else { | |
a6255b7f | 4434 | ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE); |
428870ff BB |
4435 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; |
4436 | } | |
4437 | } | |
4438 | ||
4439 | if (zio_injection_enabled && | |
4440 | zio->io_spa->spa_syncing_txg == zio->io_txg) | |
4441 | zio_handle_ignored_writes(zio); | |
4442 | ||
62840030 | 4443 | return (zio); |
34dc7c2f BB |
4444 | } |
4445 | ||
3dfb57a3 DB |
4446 | /* |
4447 | * Update the allocation throttle accounting. | |
4448 | */ | |
4449 | static void | |
4450 | zio_dva_throttle_done(zio_t *zio) | |
4451 | { | |
2a8ba608 | 4452 | zio_t *lio __maybe_unused = zio->io_logical; |
3dfb57a3 DB |
4453 | zio_t *pio = zio_unique_parent(zio); |
4454 | vdev_t *vd = zio->io_vd; | |
4455 | int flags = METASLAB_ASYNC_ALLOC; | |
4456 | ||
4457 | ASSERT3P(zio->io_bp, !=, NULL); | |
4458 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); | |
4459 | ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE); | |
4460 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); | |
4461 | ASSERT(vd != NULL); | |
4462 | ASSERT3P(vd, ==, vd->vdev_top); | |
21df134f SB |
4463 | ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY)); |
4464 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); | |
3dfb57a3 DB |
4465 | ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING); |
4466 | ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
4467 | ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA)); | |
4468 | ||
4469 | /* | |
4470 | * Parents of gang children can have two flavors -- ones that | |
4471 | * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set) | |
4472 | * and ones that allocated the constituent blocks. The allocation | |
4473 | * throttle needs to know the allocating parent zio so we must find | |
4474 | * it here. | |
4475 | */ | |
4476 | if (pio->io_child_type == ZIO_CHILD_GANG) { | |
4477 | /* | |
4478 | * If our parent is a rewrite gang child then our grandparent | |
4479 | * would have been the one that performed the allocation. | |
4480 | */ | |
4481 | if (pio->io_flags & ZIO_FLAG_IO_REWRITE) | |
4482 | pio = zio_unique_parent(pio); | |
4483 | flags |= METASLAB_GANG_CHILD; | |
4484 | } | |
4485 | ||
4486 | ASSERT(IO_IS_ALLOCATING(pio)); | |
4487 | ASSERT3P(zio, !=, zio->io_logical); | |
4488 | ASSERT(zio->io_logical != NULL); | |
4489 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); | |
4490 | ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE); | |
cc99f275 | 4491 | ASSERT(zio->io_metaslab_class != NULL); |
3dfb57a3 DB |
4492 | |
4493 | mutex_enter(&pio->io_lock); | |
492f64e9 PD |
4494 | metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags, |
4495 | pio->io_allocator, B_TRUE); | |
3dfb57a3 DB |
4496 | mutex_exit(&pio->io_lock); |
4497 | ||
cc99f275 DB |
4498 | metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1, |
4499 | pio->io_allocator, pio); | |
3dfb57a3 DB |
4500 | |
4501 | /* | |
4502 | * Call into the pipeline to see if there is more work that | |
4503 | * needs to be done. If there is work to be done it will be | |
4504 | * dispatched to another taskq thread. | |
4505 | */ | |
492f64e9 | 4506 | zio_allocate_dispatch(zio->io_spa, pio->io_allocator); |
3dfb57a3 DB |
4507 | } |
4508 | ||
62840030 | 4509 | static zio_t * |
b128c09f | 4510 | zio_done(zio_t *zio) |
34dc7c2f | 4511 | { |
3dfb57a3 DB |
4512 | /* |
4513 | * Always attempt to keep stack usage minimal here since | |
d611989f | 4514 | * we can be called recursively up to 19 levels deep. |
3dfb57a3 | 4515 | */ |
84c07ada | 4516 | const uint64_t psize = zio->io_size; |
d164b209 | 4517 | zio_t *pio, *pio_next; |
3dfb57a3 | 4518 | zio_link_t *zl = NULL; |
34dc7c2f | 4519 | |
b128c09f | 4520 | /* |
9babb374 | 4521 | * If our children haven't all completed, |
b128c09f BB |
4522 | * wait for them and then repeat this pipeline stage. |
4523 | */ | |
ddc751d5 | 4524 | if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) { |
62840030 | 4525 | return (NULL); |
ddc751d5 | 4526 | } |
34dc7c2f | 4527 | |
3dfb57a3 DB |
4528 | /* |
4529 | * If the allocation throttle is enabled, then update the accounting. | |
4530 | * We only track child I/Os that are part of an allocating async | |
4531 | * write. We must do this since the allocation is performed | |
4532 | * by the logical I/O but the actual write is done by child I/Os. | |
4533 | */ | |
4534 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING && | |
4535 | zio->io_child_type == ZIO_CHILD_VDEV) { | |
cc99f275 DB |
4536 | ASSERT(zio->io_metaslab_class != NULL); |
4537 | ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled); | |
3dfb57a3 DB |
4538 | zio_dva_throttle_done(zio); |
4539 | } | |
4540 | ||
4541 | /* | |
4542 | * If the allocation throttle is enabled, verify that | |
4543 | * we have decremented the refcounts for every I/O that was throttled. | |
4544 | */ | |
4545 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { | |
4546 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
4547 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
4548 | ASSERT(zio->io_bp != NULL); | |
cc99f275 | 4549 | |
492f64e9 PD |
4550 | metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio, |
4551 | zio->io_allocator); | |
f8020c93 AM |
4552 | VERIFY(zfs_refcount_not_held(&zio->io_metaslab_class-> |
4553 | mc_allocator[zio->io_allocator].mca_alloc_slots, zio)); | |
3dfb57a3 DB |
4554 | } |
4555 | ||
4556 | ||
1c27024e DB |
4557 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) |
4558 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) | |
b128c09f BB |
4559 | ASSERT(zio->io_children[c][w] == 0); |
4560 | ||
9b67f605 | 4561 | if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) { |
c776b317 BB |
4562 | ASSERT(zio->io_bp->blk_pad[0] == 0); |
4563 | ASSERT(zio->io_bp->blk_pad[1] == 0); | |
d1d7e268 MK |
4564 | ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, |
4565 | sizeof (blkptr_t)) == 0 || | |
c776b317 BB |
4566 | (zio->io_bp == zio_unique_parent(zio)->io_bp)); |
4567 | if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) && | |
428870ff | 4568 | zio->io_bp_override == NULL && |
b128c09f | 4569 | !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { |
d1d7e268 MK |
4570 | ASSERT3U(zio->io_prop.zp_copies, <=, |
4571 | BP_GET_NDVAS(zio->io_bp)); | |
c776b317 | 4572 | ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 || |
d1d7e268 MK |
4573 | (BP_COUNT_GANG(zio->io_bp) == |
4574 | BP_GET_NDVAS(zio->io_bp))); | |
b128c09f | 4575 | } |
03c6040b GW |
4576 | if (zio->io_flags & ZIO_FLAG_NOPWRITE) |
4577 | VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig)); | |
b128c09f BB |
4578 | } |
4579 | ||
4580 | /* | |
428870ff | 4581 | * If there were child vdev/gang/ddt errors, they apply to us now. |
b128c09f BB |
4582 | */ |
4583 | zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); | |
4584 | zio_inherit_child_errors(zio, ZIO_CHILD_GANG); | |
428870ff BB |
4585 | zio_inherit_child_errors(zio, ZIO_CHILD_DDT); |
4586 | ||
4587 | /* | |
4588 | * If the I/O on the transformed data was successful, generate any | |
4589 | * checksum reports now while we still have the transformed data. | |
4590 | */ | |
4591 | if (zio->io_error == 0) { | |
4592 | while (zio->io_cksum_report != NULL) { | |
4593 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
4594 | uint64_t align = zcr->zcr_align; | |
a6255b7f | 4595 | uint64_t asize = P2ROUNDUP(psize, align); |
a6255b7f DQ |
4596 | abd_t *adata = zio->io_abd; |
4597 | ||
f2286383 | 4598 | if (adata != NULL && asize != psize) { |
84c07ada | 4599 | adata = abd_alloc(asize, B_TRUE); |
a6255b7f DQ |
4600 | abd_copy(adata, zio->io_abd, psize); |
4601 | abd_zero_off(adata, psize, asize - psize); | |
428870ff BB |
4602 | } |
4603 | ||
4604 | zio->io_cksum_report = zcr->zcr_next; | |
4605 | zcr->zcr_next = NULL; | |
84c07ada | 4606 | zcr->zcr_finish(zcr, adata); |
428870ff BB |
4607 | zfs_ereport_free_checksum(zcr); |
4608 | ||
f2286383 | 4609 | if (adata != NULL && asize != psize) |
a6255b7f | 4610 | abd_free(adata); |
428870ff BB |
4611 | } |
4612 | } | |
b128c09f BB |
4613 | |
4614 | zio_pop_transforms(zio); /* note: may set zio->io_error */ | |
4615 | ||
a6255b7f | 4616 | vdev_stat_update(zio, psize); |
b128c09f | 4617 | |
a69052be | 4618 | /* |
cc92e9d0 | 4619 | * If this I/O is attached to a particular vdev is slow, exceeding |
72f53c56 MJ |
4620 | * 30 seconds to complete, post an error described the I/O delay. |
4621 | * We ignore these errors if the device is currently unavailable. | |
a69052be | 4622 | */ |
ad796b8a TH |
4623 | if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) { |
4624 | if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) { | |
4625 | /* | |
4626 | * We want to only increment our slow IO counters if | |
4627 | * the IO is valid (i.e. not if the drive is removed). | |
4628 | * | |
4629 | * zfs_ereport_post() will also do these checks, but | |
4630 | * it can also ratelimit and have other failures, so we | |
4631 | * need to increment the slow_io counters independent | |
4632 | * of it. | |
4633 | */ | |
4634 | if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY, | |
4635 | zio->io_spa, zio->io_vd, zio)) { | |
4636 | mutex_enter(&zio->io_vd->vdev_stat_lock); | |
4637 | zio->io_vd->vdev_stat.vs_slow_ios++; | |
4638 | mutex_exit(&zio->io_vd->vdev_stat_lock); | |
4639 | ||
1144586b | 4640 | (void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY, |
ad796b8a | 4641 | zio->io_spa, zio->io_vd, &zio->io_bookmark, |
4f072827 | 4642 | zio, 0); |
ad796b8a TH |
4643 | } |
4644 | } | |
72f53c56 | 4645 | } |
a69052be | 4646 | |
b128c09f BB |
4647 | if (zio->io_error) { |
4648 | /* | |
4649 | * If this I/O is attached to a particular vdev, | |
4650 | * generate an error message describing the I/O failure | |
4651 | * at the block level. We ignore these errors if the | |
4652 | * device is currently unavailable. | |
4653 | */ | |
c776b317 | 4654 | if (zio->io_error != ECKSUM && zio->io_vd != NULL && |
2bbec1c9 | 4655 | !vdev_is_dead(zio->io_vd)) { |
4f072827 DB |
4656 | int ret = zfs_ereport_post(FM_EREPORT_ZFS_IO, |
4657 | zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0); | |
4658 | if (ret != EALREADY) { | |
4659 | mutex_enter(&zio->io_vd->vdev_stat_lock); | |
4660 | if (zio->io_type == ZIO_TYPE_READ) | |
4661 | zio->io_vd->vdev_stat.vs_read_errors++; | |
4662 | else if (zio->io_type == ZIO_TYPE_WRITE) | |
4663 | zio->io_vd->vdev_stat.vs_write_errors++; | |
4664 | mutex_exit(&zio->io_vd->vdev_stat_lock); | |
2bbec1c9 | 4665 | } |
2bbec1c9 | 4666 | } |
34dc7c2f | 4667 | |
428870ff BB |
4668 | if ((zio->io_error == EIO || !(zio->io_flags & |
4669 | (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && | |
c776b317 | 4670 | zio == zio->io_logical) { |
b128c09f BB |
4671 | /* |
4672 | * For logical I/O requests, tell the SPA to log the | |
4673 | * error and generate a logical data ereport. | |
4674 | */ | |
b5256303 | 4675 | spa_log_error(zio->io_spa, &zio->io_bookmark); |
1144586b | 4676 | (void) zfs_ereport_post(FM_EREPORT_ZFS_DATA, |
4f072827 | 4677 | zio->io_spa, NULL, &zio->io_bookmark, zio, 0); |
b128c09f BB |
4678 | } |
4679 | } | |
34dc7c2f | 4680 | |
c776b317 | 4681 | if (zio->io_error && zio == zio->io_logical) { |
b128c09f BB |
4682 | /* |
4683 | * Determine whether zio should be reexecuted. This will | |
4684 | * propagate all the way to the root via zio_notify_parent(). | |
4685 | */ | |
c776b317 | 4686 | ASSERT(zio->io_vd == NULL && zio->io_bp != NULL); |
428870ff | 4687 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); |
b128c09f | 4688 | |
428870ff BB |
4689 | if (IO_IS_ALLOCATING(zio) && |
4690 | !(zio->io_flags & ZIO_FLAG_CANFAIL)) { | |
b128c09f BB |
4691 | if (zio->io_error != ENOSPC) |
4692 | zio->io_reexecute |= ZIO_REEXECUTE_NOW; | |
4693 | else | |
4694 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff | 4695 | } |
b128c09f BB |
4696 | |
4697 | if ((zio->io_type == ZIO_TYPE_READ || | |
4698 | zio->io_type == ZIO_TYPE_FREE) && | |
572e2857 | 4699 | !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && |
b128c09f | 4700 | zio->io_error == ENXIO && |
c776b317 BB |
4701 | spa_load_state(zio->io_spa) == SPA_LOAD_NONE && |
4702 | spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE) | |
b128c09f BB |
4703 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; |
4704 | ||
4705 | if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) | |
4706 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff BB |
4707 | |
4708 | /* | |
4709 | * Here is a possibly good place to attempt to do | |
4710 | * either combinatorial reconstruction or error correction | |
4711 | * based on checksums. It also might be a good place | |
4712 | * to send out preliminary ereports before we suspend | |
4713 | * processing. | |
4714 | */ | |
34dc7c2f BB |
4715 | } |
4716 | ||
4717 | /* | |
b128c09f BB |
4718 | * If there were logical child errors, they apply to us now. |
4719 | * We defer this until now to avoid conflating logical child | |
4720 | * errors with errors that happened to the zio itself when | |
4721 | * updating vdev stats and reporting FMA events above. | |
34dc7c2f | 4722 | */ |
b128c09f | 4723 | zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); |
34dc7c2f | 4724 | |
428870ff BB |
4725 | if ((zio->io_error || zio->io_reexecute) && |
4726 | IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && | |
03c6040b | 4727 | !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) |
c776b317 | 4728 | zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp); |
9babb374 BB |
4729 | |
4730 | zio_gang_tree_free(&zio->io_gang_tree); | |
4731 | ||
4732 | /* | |
4733 | * Godfather I/Os should never suspend. | |
4734 | */ | |
4735 | if ((zio->io_flags & ZIO_FLAG_GODFATHER) && | |
4736 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) | |
a32494d2 | 4737 | zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND; |
9babb374 | 4738 | |
b128c09f BB |
4739 | if (zio->io_reexecute) { |
4740 | /* | |
4741 | * This is a logical I/O that wants to reexecute. | |
4742 | * | |
4743 | * Reexecute is top-down. When an i/o fails, if it's not | |
4744 | * the root, it simply notifies its parent and sticks around. | |
4745 | * The parent, seeing that it still has children in zio_done(), | |
4746 | * does the same. This percolates all the way up to the root. | |
4747 | * The root i/o will reexecute or suspend the entire tree. | |
4748 | * | |
4749 | * This approach ensures that zio_reexecute() honors | |
4750 | * all the original i/o dependency relationships, e.g. | |
4751 | * parents not executing until children are ready. | |
4752 | */ | |
4753 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
34dc7c2f | 4754 | |
9babb374 | 4755 | zio->io_gang_leader = NULL; |
b128c09f | 4756 | |
d164b209 BB |
4757 | mutex_enter(&zio->io_lock); |
4758 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
4759 | mutex_exit(&zio->io_lock); | |
4760 | ||
9babb374 BB |
4761 | /* |
4762 | * "The Godfather" I/O monitors its children but is | |
4763 | * not a true parent to them. It will track them through | |
4764 | * the pipeline but severs its ties whenever they get into | |
4765 | * trouble (e.g. suspended). This allows "The Godfather" | |
4766 | * I/O to return status without blocking. | |
4767 | */ | |
3dfb57a3 DB |
4768 | zl = NULL; |
4769 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; | |
4770 | pio = pio_next) { | |
4771 | zio_link_t *remove_zl = zl; | |
4772 | pio_next = zio_walk_parents(zio, &zl); | |
9babb374 BB |
4773 | |
4774 | if ((pio->io_flags & ZIO_FLAG_GODFATHER) && | |
4775 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { | |
3dfb57a3 | 4776 | zio_remove_child(pio, zio, remove_zl); |
62840030 MA |
4777 | /* |
4778 | * This is a rare code path, so we don't | |
4779 | * bother with "next_to_execute". | |
4780 | */ | |
4781 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, | |
4782 | NULL); | |
9babb374 BB |
4783 | } |
4784 | } | |
4785 | ||
d164b209 | 4786 | if ((pio = zio_unique_parent(zio)) != NULL) { |
b128c09f BB |
4787 | /* |
4788 | * We're not a root i/o, so there's nothing to do | |
4789 | * but notify our parent. Don't propagate errors | |
4790 | * upward since we haven't permanently failed yet. | |
4791 | */ | |
9babb374 | 4792 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f | 4793 | zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; |
62840030 MA |
4794 | /* |
4795 | * This is a rare code path, so we don't bother with | |
4796 | * "next_to_execute". | |
4797 | */ | |
4798 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL); | |
b128c09f BB |
4799 | } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { |
4800 | /* | |
4801 | * We'd fail again if we reexecuted now, so suspend | |
4802 | * until conditions improve (e.g. device comes online). | |
4803 | */ | |
cec3a0a1 | 4804 | zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR); |
b128c09f BB |
4805 | } else { |
4806 | /* | |
4807 | * Reexecution is potentially a huge amount of work. | |
4808 | * Hand it off to the otherwise-unused claim taskq. | |
4809 | */ | |
a38718a6 | 4810 | ASSERT(taskq_empty_ent(&zio->io_tqent)); |
7ef5e54e AL |
4811 | spa_taskq_dispatch_ent(zio->io_spa, |
4812 | ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE, | |
23c13c7e | 4813 | zio_reexecute, zio, 0, &zio->io_tqent); |
b128c09f | 4814 | } |
62840030 | 4815 | return (NULL); |
34dc7c2f BB |
4816 | } |
4817 | ||
428870ff | 4818 | ASSERT(zio->io_child_count == 0); |
b128c09f BB |
4819 | ASSERT(zio->io_reexecute == 0); |
4820 | ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); | |
34dc7c2f | 4821 | |
428870ff BB |
4822 | /* |
4823 | * Report any checksum errors, since the I/O is complete. | |
4824 | */ | |
4825 | while (zio->io_cksum_report != NULL) { | |
4826 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
4827 | zio->io_cksum_report = zcr->zcr_next; | |
4828 | zcr->zcr_next = NULL; | |
4829 | zcr->zcr_finish(zcr, NULL); | |
4830 | zfs_ereport_free_checksum(zcr); | |
4831 | } | |
4832 | ||
920dd524 | 4833 | if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp && |
9b67f605 MA |
4834 | !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) && |
4835 | !(zio->io_flags & ZIO_FLAG_NOPWRITE)) { | |
920dd524 ED |
4836 | metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp); |
4837 | } | |
4838 | ||
d164b209 BB |
4839 | /* |
4840 | * It is the responsibility of the done callback to ensure that this | |
4841 | * particular zio is no longer discoverable for adoption, and as | |
4842 | * such, cannot acquire any new parents. | |
4843 | */ | |
b128c09f BB |
4844 | if (zio->io_done) |
4845 | zio->io_done(zio); | |
34dc7c2f | 4846 | |
d164b209 BB |
4847 | mutex_enter(&zio->io_lock); |
4848 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
4849 | mutex_exit(&zio->io_lock); | |
34dc7c2f | 4850 | |
62840030 MA |
4851 | /* |
4852 | * We are done executing this zio. We may want to execute a parent | |
4853 | * next. See the comment in zio_notify_parent(). | |
4854 | */ | |
4855 | zio_t *next_to_execute = NULL; | |
3dfb57a3 DB |
4856 | zl = NULL; |
4857 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) { | |
4858 | zio_link_t *remove_zl = zl; | |
4859 | pio_next = zio_walk_parents(zio, &zl); | |
4860 | zio_remove_child(pio, zio, remove_zl); | |
62840030 | 4861 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute); |
b128c09f | 4862 | } |
34dc7c2f | 4863 | |
b128c09f BB |
4864 | if (zio->io_waiter != NULL) { |
4865 | mutex_enter(&zio->io_lock); | |
4866 | zio->io_executor = NULL; | |
4867 | cv_broadcast(&zio->io_cv); | |
4868 | mutex_exit(&zio->io_lock); | |
4869 | } else { | |
4870 | zio_destroy(zio); | |
4871 | } | |
34dc7c2f | 4872 | |
62840030 | 4873 | return (next_to_execute); |
34dc7c2f BB |
4874 | } |
4875 | ||
4876 | /* | |
b128c09f BB |
4877 | * ========================================================================== |
4878 | * I/O pipeline definition | |
4879 | * ========================================================================== | |
34dc7c2f | 4880 | */ |
428870ff | 4881 | static zio_pipe_stage_t *zio_pipeline[] = { |
b128c09f | 4882 | NULL, |
b128c09f | 4883 | zio_read_bp_init, |
3dfb57a3 | 4884 | zio_write_bp_init, |
428870ff BB |
4885 | zio_free_bp_init, |
4886 | zio_issue_async, | |
3dfb57a3 | 4887 | zio_write_compress, |
b5256303 | 4888 | zio_encrypt, |
b128c09f | 4889 | zio_checksum_generate, |
03c6040b | 4890 | zio_nop_write, |
428870ff BB |
4891 | zio_ddt_read_start, |
4892 | zio_ddt_read_done, | |
4893 | zio_ddt_write, | |
4894 | zio_ddt_free, | |
b128c09f BB |
4895 | zio_gang_assemble, |
4896 | zio_gang_issue, | |
3dfb57a3 | 4897 | zio_dva_throttle, |
b128c09f BB |
4898 | zio_dva_allocate, |
4899 | zio_dva_free, | |
4900 | zio_dva_claim, | |
4901 | zio_ready, | |
4902 | zio_vdev_io_start, | |
4903 | zio_vdev_io_done, | |
4904 | zio_vdev_io_assess, | |
4905 | zio_checksum_verify, | |
4906 | zio_done | |
4907 | }; | |
c28b2279 | 4908 | |
9ae529ec | 4909 | |
9ae529ec | 4910 | |
9ae529ec | 4911 | |
fcff0f35 PD |
4912 | /* |
4913 | * Compare two zbookmark_phys_t's to see which we would reach first in a | |
4914 | * pre-order traversal of the object tree. | |
4915 | * | |
4916 | * This is simple in every case aside from the meta-dnode object. For all other | |
4917 | * objects, we traverse them in order (object 1 before object 2, and so on). | |
4918 | * However, all of these objects are traversed while traversing object 0, since | |
4919 | * the data it points to is the list of objects. Thus, we need to convert to a | |
4920 | * canonical representation so we can compare meta-dnode bookmarks to | |
4921 | * non-meta-dnode bookmarks. | |
4922 | * | |
4923 | * We do this by calculating "equivalents" for each field of the zbookmark. | |
4924 | * zbookmarks outside of the meta-dnode use their own object and level, and | |
4925 | * calculate the level 0 equivalent (the first L0 blkid that is contained in the | |
4926 | * blocks this bookmark refers to) by multiplying their blkid by their span | |
4927 | * (the number of L0 blocks contained within one block at their level). | |
4928 | * zbookmarks inside the meta-dnode calculate their object equivalent | |
4929 | * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use | |
4930 | * level + 1<<31 (any value larger than a level could ever be) for their level. | |
4931 | * This causes them to always compare before a bookmark in their object | |
4932 | * equivalent, compare appropriately to bookmarks in other objects, and to | |
4933 | * compare appropriately to other bookmarks in the meta-dnode. | |
4934 | */ | |
4935 | int | |
4936 | zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, | |
4937 | const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) | |
4938 | { | |
4939 | /* | |
4940 | * These variables represent the "equivalent" values for the zbookmark, | |
4941 | * after converting zbookmarks inside the meta dnode to their | |
4942 | * normal-object equivalents. | |
4943 | */ | |
4944 | uint64_t zb1obj, zb2obj; | |
4945 | uint64_t zb1L0, zb2L0; | |
4946 | uint64_t zb1level, zb2level; | |
4947 | ||
4948 | if (zb1->zb_object == zb2->zb_object && | |
4949 | zb1->zb_level == zb2->zb_level && | |
4950 | zb1->zb_blkid == zb2->zb_blkid) | |
4951 | return (0); | |
9ae529ec | 4952 | |
30af21b0 PD |
4953 | IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT); |
4954 | IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT); | |
4955 | ||
fcff0f35 PD |
4956 | /* |
4957 | * BP_SPANB calculates the span in blocks. | |
4958 | */ | |
4959 | zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); | |
4960 | zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); | |
9ae529ec CS |
4961 | |
4962 | if (zb1->zb_object == DMU_META_DNODE_OBJECT) { | |
fcff0f35 PD |
4963 | zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); |
4964 | zb1L0 = 0; | |
4965 | zb1level = zb1->zb_level + COMPARE_META_LEVEL; | |
4966 | } else { | |
4967 | zb1obj = zb1->zb_object; | |
4968 | zb1level = zb1->zb_level; | |
9ae529ec CS |
4969 | } |
4970 | ||
fcff0f35 PD |
4971 | if (zb2->zb_object == DMU_META_DNODE_OBJECT) { |
4972 | zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); | |
4973 | zb2L0 = 0; | |
4974 | zb2level = zb2->zb_level + COMPARE_META_LEVEL; | |
4975 | } else { | |
4976 | zb2obj = zb2->zb_object; | |
4977 | zb2level = zb2->zb_level; | |
4978 | } | |
4979 | ||
4980 | /* Now that we have a canonical representation, do the comparison. */ | |
4981 | if (zb1obj != zb2obj) | |
4982 | return (zb1obj < zb2obj ? -1 : 1); | |
4983 | else if (zb1L0 != zb2L0) | |
4984 | return (zb1L0 < zb2L0 ? -1 : 1); | |
4985 | else if (zb1level != zb2level) | |
4986 | return (zb1level > zb2level ? -1 : 1); | |
4987 | /* | |
4988 | * This can (theoretically) happen if the bookmarks have the same object | |
4989 | * and level, but different blkids, if the block sizes are not the same. | |
4990 | * There is presently no way to change the indirect block sizes | |
4991 | */ | |
4992 | return (0); | |
4993 | } | |
4994 | ||
4995 | /* | |
4996 | * This function checks the following: given that last_block is the place that | |
4997 | * our traversal stopped last time, does that guarantee that we've visited | |
4998 | * every node under subtree_root? Therefore, we can't just use the raw output | |
4999 | * of zbookmark_compare. We have to pass in a modified version of | |
5000 | * subtree_root; by incrementing the block id, and then checking whether | |
5001 | * last_block is before or equal to that, we can tell whether or not having | |
5002 | * visited last_block implies that all of subtree_root's children have been | |
5003 | * visited. | |
5004 | */ | |
5005 | boolean_t | |
5006 | zbookmark_subtree_completed(const dnode_phys_t *dnp, | |
5007 | const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) | |
5008 | { | |
5009 | zbookmark_phys_t mod_zb = *subtree_root; | |
5010 | mod_zb.zb_blkid++; | |
5011 | ASSERT(last_block->zb_level == 0); | |
5012 | ||
5013 | /* The objset_phys_t isn't before anything. */ | |
5014 | if (dnp == NULL) | |
9ae529ec | 5015 | return (B_FALSE); |
fcff0f35 PD |
5016 | |
5017 | /* | |
5018 | * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the | |
5019 | * data block size in sectors, because that variable is only used if | |
5020 | * the bookmark refers to a block in the meta-dnode. Since we don't | |
5021 | * know without examining it what object it refers to, and there's no | |
5022 | * harm in passing in this value in other cases, we always pass it in. | |
5023 | * | |
5024 | * We pass in 0 for the indirect block size shift because zb2 must be | |
5025 | * level 0. The indirect block size is only used to calculate the span | |
5026 | * of the bookmark, but since the bookmark must be level 0, the span is | |
5027 | * always 1, so the math works out. | |
5028 | * | |
5029 | * If you make changes to how the zbookmark_compare code works, be sure | |
5030 | * to make sure that this code still works afterwards. | |
5031 | */ | |
5032 | return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, | |
5033 | 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, | |
5034 | last_block) <= 0); | |
9ae529ec CS |
5035 | } |
5036 | ||
c28b2279 | 5037 | EXPORT_SYMBOL(zio_type_name); |
81971b13 BB |
5038 | EXPORT_SYMBOL(zio_buf_alloc); |
5039 | EXPORT_SYMBOL(zio_data_buf_alloc); | |
5040 | EXPORT_SYMBOL(zio_buf_free); | |
5041 | EXPORT_SYMBOL(zio_data_buf_free); | |
c28b2279 | 5042 | |
03fdcb9a | 5043 | ZFS_MODULE_PARAM(zfs_zio, zio_, slow_io_ms, INT, ZMOD_RW, |
ad796b8a | 5044 | "Max I/O completion time (milliseconds) before marking it as slow"); |
c409e464 | 5045 | |
03fdcb9a MM |
5046 | ZFS_MODULE_PARAM(zfs_zio, zio_, requeue_io_start_cut_in_line, INT, ZMOD_RW, |
5047 | "Prioritize requeued I/O"); | |
29dee3ee | 5048 | |
03fdcb9a | 5049 | ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_deferred_free, INT, ZMOD_RW, |
d1d7e268 | 5050 | "Defer frees starting in this pass"); |
29dee3ee | 5051 | |
03fdcb9a | 5052 | ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_dont_compress, INT, ZMOD_RW, |
d1d7e268 | 5053 | "Don't compress starting in this pass"); |
29dee3ee | 5054 | |
03fdcb9a | 5055 | ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_rewrite, INT, ZMOD_RW, |
d1d7e268 | 5056 | "Rewrite new bps starting in this pass"); |
3dfb57a3 | 5057 | |
03fdcb9a | 5058 | ZFS_MODULE_PARAM(zfs_zio, zio_, dva_throttle_enabled, INT, ZMOD_RW, |
3dfb57a3 | 5059 | "Throttle block allocations in the ZIO pipeline"); |
638dd5f4 | 5060 | |
03fdcb9a | 5061 | ZFS_MODULE_PARAM(zfs_zio, zio_, deadman_log_all, INT, ZMOD_RW, |
638dd5f4 | 5062 | "Log all slow ZIOs, not just those with vdevs"); |