]>
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 | |
1d3ba0bf | 9 | * or https://opensource.org/licenses/CDDL-1.0. |
34dc7c2f BB |
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. |
d7cf06a2 | 23 | * Copyright (c) 2011, 2022 by Delphix. All rights reserved. |
a38718a6 | 24 | * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved. |
cc99f275 | 25 | * Copyright (c) 2017, Intel Corporation. |
cac416f1 | 26 | * Copyright (c) 2019, 2023, 2024, Klara Inc. |
10b3c7f5 | 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> | |
67a1b037 | 44 | #include <sys/brt.h> |
428870ff | 45 | #include <sys/ddt.h> |
9b67f605 | 46 | #include <sys/blkptr.h> |
b0bc7a84 | 47 | #include <sys/zfeature.h> |
d4a72f23 | 48 | #include <sys/dsl_scan.h> |
3dfb57a3 | 49 | #include <sys/metaslab_impl.h> |
193a37cb | 50 | #include <sys/time.h> |
e5d1c27e | 51 | #include <sys/trace_zfs.h> |
a6255b7f | 52 | #include <sys/abd.h> |
b5256303 | 53 | #include <sys/dsl_crypt.h> |
3f387973 | 54 | #include <cityhash.h> |
34dc7c2f | 55 | |
34dc7c2f BB |
56 | /* |
57 | * ========================================================================== | |
58 | * I/O type descriptions | |
59 | * ========================================================================== | |
60 | */ | |
18168da7 | 61 | const char *const zio_type_name[ZIO_TYPES] = { |
3dfb57a3 DB |
62 | /* |
63 | * Note: Linux kernel thread name length is limited | |
64 | * so these names will differ from upstream open zfs. | |
65 | */ | |
d7605ae7 | 66 | "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_flush", "z_trim" |
428870ff | 67 | }; |
34dc7c2f | 68 | |
27f2b90d | 69 | int zio_dva_throttle_enabled = B_TRUE; |
18168da7 | 70 | static int zio_deadman_log_all = B_FALSE; |
3dfb57a3 | 71 | |
34dc7c2f BB |
72 | /* |
73 | * ========================================================================== | |
74 | * I/O kmem caches | |
75 | * ========================================================================== | |
76 | */ | |
18168da7 AZ |
77 | static kmem_cache_t *zio_cache; |
78 | static kmem_cache_t *zio_link_cache; | |
34dc7c2f BB |
79 | kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; |
80 | kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
a6255b7f | 81 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
18168da7 AZ |
82 | static uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; |
83 | static uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
a6255b7f DQ |
84 | #endif |
85 | ||
ad796b8a | 86 | /* Mark IOs as "slow" if they take longer than 30 seconds */ |
fdc2d303 | 87 | static uint_t zio_slow_io_ms = (30 * MILLISEC); |
34dc7c2f | 88 | |
fcff0f35 PD |
89 | #define BP_SPANB(indblkshift, level) \ |
90 | (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) | |
91 | #define COMPARE_META_LEVEL 0x80000000ul | |
55d85d5a GW |
92 | /* |
93 | * The following actions directly effect the spa's sync-to-convergence logic. | |
94 | * The values below define the sync pass when we start performing the action. | |
95 | * Care should be taken when changing these values as they directly impact | |
96 | * spa_sync() performance. Tuning these values may introduce subtle performance | |
97 | * pathologies and should only be done in the context of performance analysis. | |
98 | * These tunables will eventually be removed and replaced with #defines once | |
99 | * enough analysis has been done to determine optimal values. | |
100 | * | |
101 | * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that | |
102 | * regular blocks are not deferred. | |
be89734a MA |
103 | * |
104 | * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable | |
105 | * compression (including of metadata). In practice, we don't have this | |
106 | * many sync passes, so this has no effect. | |
107 | * | |
108 | * The original intent was that disabling compression would help the sync | |
109 | * passes to converge. However, in practice disabling compression increases | |
110 | * the average number of sync passes, because when we turn compression off, a | |
111 | * lot of block's size will change and thus we have to re-allocate (not | |
112 | * overwrite) them. It also increases the number of 128KB allocations (e.g. | |
113 | * for indirect blocks and spacemaps) because these will not be compressed. | |
114 | * The 128K allocations are especially detrimental to performance on highly | |
115 | * fragmented systems, which may have very few free segments of this size, | |
116 | * and may need to load new metaslabs to satisfy 128K allocations. | |
55d85d5a | 117 | */ |
fdc2d303 RY |
118 | |
119 | /* defer frees starting in this pass */ | |
120 | uint_t zfs_sync_pass_deferred_free = 2; | |
121 | ||
122 | /* don't compress starting in this pass */ | |
123 | static uint_t zfs_sync_pass_dont_compress = 8; | |
124 | ||
125 | /* rewrite new bps starting in this pass */ | |
126 | static uint_t zfs_sync_pass_rewrite = 2; | |
55d85d5a | 127 | |
34dc7c2f | 128 | /* |
b128c09f BB |
129 | * An allocating zio is one that either currently has the DVA allocate |
130 | * stage set or will have it later in its lifetime. | |
34dc7c2f | 131 | */ |
428870ff BB |
132 | #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) |
133 | ||
3c502d3b MM |
134 | /* |
135 | * Enable smaller cores by excluding metadata | |
136 | * allocations as well. | |
137 | */ | |
138 | int zio_exclude_metadata = 0; | |
18168da7 | 139 | static int zio_requeue_io_start_cut_in_line = 1; |
428870ff BB |
140 | |
141 | #ifdef ZFS_DEBUG | |
18168da7 | 142 | static const int zio_buf_debug_limit = 16384; |
428870ff | 143 | #else |
18168da7 | 144 | static const int zio_buf_debug_limit = 0; |
428870ff | 145 | #endif |
34dc7c2f | 146 | |
da6b4005 NB |
147 | static inline void __zio_execute(zio_t *zio); |
148 | ||
3dfb57a3 DB |
149 | static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t); |
150 | ||
34dc7c2f BB |
151 | void |
152 | zio_init(void) | |
153 | { | |
154 | size_t c; | |
34dc7c2f | 155 | |
3941503c BB |
156 | zio_cache = kmem_cache_create("zio_cache", |
157 | sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); | |
d164b209 | 158 | zio_link_cache = kmem_cache_create("zio_link_cache", |
6795a698 | 159 | sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); |
34dc7c2f | 160 | |
34dc7c2f BB |
161 | for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { |
162 | size_t size = (c + 1) << SPA_MINBLOCKSHIFT; | |
514d661c AM |
163 | size_t align, cflags, data_cflags; |
164 | char name[32]; | |
34dc7c2f | 165 | |
514d661c AM |
166 | /* |
167 | * Create cache for each half-power of 2 size, starting from | |
168 | * SPA_MINBLOCKSIZE. It should give us memory space efficiency | |
169 | * of ~7/8, sufficient for transient allocations mostly using | |
170 | * these caches. | |
171 | */ | |
172 | size_t p2 = size; | |
f1512ee6 | 173 | while (!ISP2(p2)) |
34dc7c2f | 174 | p2 &= p2 - 1; |
514d661c AM |
175 | if (!IS_P2ALIGNED(size, p2 / 2)) |
176 | continue; | |
34dc7c2f | 177 | |
498877ba MA |
178 | #ifndef _KERNEL |
179 | /* | |
180 | * If we are using watchpoints, put each buffer on its own page, | |
181 | * to eliminate the performance overhead of trapping to the | |
182 | * kernel when modifying a non-watched buffer that shares the | |
183 | * page with a watched buffer. | |
184 | */ | |
185 | if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) | |
186 | continue; | |
fcf64f45 | 187 | #endif |
34dc7c2f | 188 | |
514d661c AM |
189 | if (IS_P2ALIGNED(size, PAGESIZE)) |
190 | align = PAGESIZE; | |
191 | else | |
192 | align = 1 << (highbit64(size ^ (size - 1)) - 1); | |
34dc7c2f | 193 | |
514d661c AM |
194 | cflags = (zio_exclude_metadata || size > zio_buf_debug_limit) ? |
195 | KMC_NODEBUG : 0; | |
196 | data_cflags = KMC_NODEBUG; | |
197 | if (cflags == data_cflags) { | |
198 | /* | |
199 | * Resulting kmem caches would be identical. | |
200 | * Save memory by creating only one. | |
201 | */ | |
202 | (void) snprintf(name, sizeof (name), | |
203 | "zio_buf_comb_%lu", (ulong_t)size); | |
204 | zio_buf_cache[c] = kmem_cache_create(name, size, align, | |
205 | NULL, NULL, NULL, NULL, NULL, cflags); | |
206 | zio_data_buf_cache[c] = zio_buf_cache[c]; | |
207 | continue; | |
34dc7c2f | 208 | } |
514d661c AM |
209 | (void) snprintf(name, sizeof (name), "zio_buf_%lu", |
210 | (ulong_t)size); | |
211 | zio_buf_cache[c] = kmem_cache_create(name, size, align, | |
212 | NULL, NULL, NULL, NULL, NULL, cflags); | |
213 | ||
214 | (void) snprintf(name, sizeof (name), "zio_data_buf_%lu", | |
215 | (ulong_t)size); | |
216 | zio_data_buf_cache[c] = kmem_cache_create(name, size, align, | |
217 | NULL, NULL, NULL, NULL, NULL, data_cflags); | |
34dc7c2f BB |
218 | } |
219 | ||
220 | while (--c != 0) { | |
221 | ASSERT(zio_buf_cache[c] != NULL); | |
222 | if (zio_buf_cache[c - 1] == NULL) | |
223 | zio_buf_cache[c - 1] = zio_buf_cache[c]; | |
224 | ||
225 | ASSERT(zio_data_buf_cache[c] != NULL); | |
226 | if (zio_data_buf_cache[c - 1] == NULL) | |
227 | zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; | |
228 | } | |
229 | ||
34dc7c2f | 230 | zio_inject_init(); |
9759c60f ED |
231 | |
232 | lz4_init(); | |
34dc7c2f BB |
233 | } |
234 | ||
235 | void | |
236 | zio_fini(void) | |
237 | { | |
309c32c9 | 238 | size_t n = SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; |
34dc7c2f | 239 | |
a6255b7f | 240 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
309c32c9 MG |
241 | for (size_t i = 0; i < n; i++) { |
242 | if (zio_buf_cache_allocs[i] != zio_buf_cache_frees[i]) | |
a6255b7f | 243 | (void) printf("zio_fini: [%d] %llu != %llu\n", |
309c32c9 MG |
244 | (int)((i + 1) << SPA_MINBLOCKSHIFT), |
245 | (long long unsigned)zio_buf_cache_allocs[i], | |
246 | (long long unsigned)zio_buf_cache_frees[i]); | |
247 | } | |
f1512ee6 | 248 | #endif |
309c32c9 MG |
249 | |
250 | /* | |
251 | * The same kmem cache can show up multiple times in both zio_buf_cache | |
252 | * and zio_data_buf_cache. Do a wasteful but trivially correct scan to | |
253 | * sort it out. | |
254 | */ | |
255 | for (size_t i = 0; i < n; i++) { | |
256 | kmem_cache_t *cache = zio_buf_cache[i]; | |
257 | if (cache == NULL) | |
258 | continue; | |
259 | for (size_t j = i; j < n; j++) { | |
260 | if (cache == zio_buf_cache[j]) | |
261 | zio_buf_cache[j] = NULL; | |
262 | if (cache == zio_data_buf_cache[j]) | |
263 | zio_data_buf_cache[j] = NULL; | |
34dc7c2f | 264 | } |
309c32c9 MG |
265 | kmem_cache_destroy(cache); |
266 | } | |
34dc7c2f | 267 | |
309c32c9 MG |
268 | for (size_t i = 0; i < n; i++) { |
269 | kmem_cache_t *cache = zio_data_buf_cache[i]; | |
270 | if (cache == NULL) | |
271 | continue; | |
272 | for (size_t j = i; j < n; j++) { | |
273 | if (cache == zio_data_buf_cache[j]) | |
274 | zio_data_buf_cache[j] = NULL; | |
34dc7c2f | 275 | } |
309c32c9 MG |
276 | kmem_cache_destroy(cache); |
277 | } | |
278 | ||
279 | for (size_t i = 0; i < n; i++) { | |
280 | VERIFY3P(zio_buf_cache[i], ==, NULL); | |
281 | VERIFY3P(zio_data_buf_cache[i], ==, NULL); | |
34dc7c2f BB |
282 | } |
283 | ||
d164b209 | 284 | kmem_cache_destroy(zio_link_cache); |
34dc7c2f BB |
285 | kmem_cache_destroy(zio_cache); |
286 | ||
287 | zio_inject_fini(); | |
9759c60f ED |
288 | |
289 | lz4_fini(); | |
34dc7c2f BB |
290 | } |
291 | ||
292 | /* | |
293 | * ========================================================================== | |
294 | * Allocate and free I/O buffers | |
295 | * ========================================================================== | |
296 | */ | |
297 | ||
adcea23c AM |
298 | #ifdef ZFS_DEBUG |
299 | static const ulong_t zio_buf_canary = (ulong_t)0xdeadc0dedead210b; | |
300 | #endif | |
301 | ||
302 | /* | |
303 | * Use empty space after the buffer to detect overflows. | |
304 | * | |
305 | * Since zio_init() creates kmem caches only for certain set of buffer sizes, | |
306 | * allocations of different sizes may have some unused space after the data. | |
307 | * Filling part of that space with a known pattern on allocation and checking | |
308 | * it on free should allow us to detect some buffer overflows. | |
309 | */ | |
310 | static void | |
311 | zio_buf_put_canary(ulong_t *p, size_t size, kmem_cache_t **cache, size_t c) | |
312 | { | |
313 | #ifdef ZFS_DEBUG | |
314 | size_t off = P2ROUNDUP(size, sizeof (ulong_t)); | |
315 | ulong_t *canary = p + off / sizeof (ulong_t); | |
316 | size_t asize = (c + 1) << SPA_MINBLOCKSHIFT; | |
317 | if (c + 1 < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT && | |
318 | cache[c] == cache[c + 1]) | |
319 | asize = (c + 2) << SPA_MINBLOCKSHIFT; | |
320 | for (; off < asize; canary++, off += sizeof (ulong_t)) | |
321 | *canary = zio_buf_canary; | |
322 | #endif | |
323 | } | |
324 | ||
325 | static void | |
326 | zio_buf_check_canary(ulong_t *p, size_t size, kmem_cache_t **cache, size_t c) | |
327 | { | |
328 | #ifdef ZFS_DEBUG | |
329 | size_t off = P2ROUNDUP(size, sizeof (ulong_t)); | |
330 | ulong_t *canary = p + off / sizeof (ulong_t); | |
331 | size_t asize = (c + 1) << SPA_MINBLOCKSHIFT; | |
332 | if (c + 1 < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT && | |
333 | cache[c] == cache[c + 1]) | |
334 | asize = (c + 2) << SPA_MINBLOCKSHIFT; | |
335 | for (; off < asize; canary++, off += sizeof (ulong_t)) { | |
336 | if (unlikely(*canary != zio_buf_canary)) { | |
337 | PANIC("ZIO buffer overflow %p (%zu) + %zu %#lx != %#lx", | |
338 | p, size, (canary - p) * sizeof (ulong_t), | |
339 | *canary, zio_buf_canary); | |
340 | } | |
341 | } | |
342 | #endif | |
343 | } | |
344 | ||
34dc7c2f BB |
345 | /* |
346 | * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a | |
347 | * crashdump if the kernel panics, so use it judiciously. Obviously, it's | |
348 | * useful to inspect ZFS metadata, but if possible, we should avoid keeping | |
349 | * excess / transient data in-core during a crashdump. | |
350 | */ | |
351 | void * | |
352 | zio_buf_alloc(size_t size) | |
353 | { | |
354 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
355 | ||
63e3a861 | 356 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
a6255b7f DQ |
357 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
358 | atomic_add_64(&zio_buf_cache_allocs[c], 1); | |
359 | #endif | |
34dc7c2f | 360 | |
adcea23c AM |
361 | void *p = kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE); |
362 | zio_buf_put_canary(p, size, zio_buf_cache, c); | |
363 | return (p); | |
34dc7c2f BB |
364 | } |
365 | ||
366 | /* | |
367 | * Use zio_data_buf_alloc to allocate data. The data will not appear in a | |
368 | * crashdump if the kernel panics. This exists so that we will limit the amount | |
369 | * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount | |
370 | * of kernel heap dumped to disk when the kernel panics) | |
371 | */ | |
372 | void * | |
373 | zio_data_buf_alloc(size_t size) | |
374 | { | |
375 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
376 | ||
63e3a861 | 377 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f | 378 | |
adcea23c AM |
379 | void *p = kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE); |
380 | zio_buf_put_canary(p, size, zio_data_buf_cache, c); | |
381 | return (p); | |
34dc7c2f BB |
382 | } |
383 | ||
384 | void | |
385 | zio_buf_free(void *buf, size_t size) | |
386 | { | |
387 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
388 | ||
63e3a861 | 389 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
a6255b7f DQ |
390 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
391 | atomic_add_64(&zio_buf_cache_frees[c], 1); | |
392 | #endif | |
34dc7c2f | 393 | |
adcea23c | 394 | zio_buf_check_canary(buf, size, zio_buf_cache, c); |
34dc7c2f BB |
395 | kmem_cache_free(zio_buf_cache[c], buf); |
396 | } | |
397 | ||
398 | void | |
399 | zio_data_buf_free(void *buf, size_t size) | |
400 | { | |
401 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
402 | ||
63e3a861 | 403 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f | 404 | |
adcea23c | 405 | zio_buf_check_canary(buf, size, zio_data_buf_cache, c); |
34dc7c2f BB |
406 | kmem_cache_free(zio_data_buf_cache[c], buf); |
407 | } | |
408 | ||
84c07ada GN |
409 | static void |
410 | zio_abd_free(void *abd, size_t size) | |
411 | { | |
14e4e3cb | 412 | (void) size; |
84c07ada GN |
413 | abd_free((abd_t *)abd); |
414 | } | |
415 | ||
34dc7c2f BB |
416 | /* |
417 | * ========================================================================== | |
418 | * Push and pop I/O transform buffers | |
419 | * ========================================================================== | |
420 | */ | |
d3c2ae1c | 421 | void |
a6255b7f | 422 | zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize, |
e9aa730c | 423 | zio_transform_func_t *transform) |
34dc7c2f | 424 | { |
79c76d5b | 425 | zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); |
34dc7c2f | 426 | |
a6255b7f | 427 | zt->zt_orig_abd = zio->io_abd; |
b128c09f | 428 | zt->zt_orig_size = zio->io_size; |
34dc7c2f | 429 | zt->zt_bufsize = bufsize; |
b128c09f | 430 | zt->zt_transform = transform; |
34dc7c2f BB |
431 | |
432 | zt->zt_next = zio->io_transform_stack; | |
433 | zio->io_transform_stack = zt; | |
434 | ||
a6255b7f | 435 | zio->io_abd = data; |
34dc7c2f BB |
436 | zio->io_size = size; |
437 | } | |
438 | ||
d3c2ae1c | 439 | void |
b128c09f | 440 | zio_pop_transforms(zio_t *zio) |
34dc7c2f | 441 | { |
b128c09f BB |
442 | zio_transform_t *zt; |
443 | ||
444 | while ((zt = zio->io_transform_stack) != NULL) { | |
445 | if (zt->zt_transform != NULL) | |
446 | zt->zt_transform(zio, | |
a6255b7f | 447 | zt->zt_orig_abd, zt->zt_orig_size); |
34dc7c2f | 448 | |
428870ff | 449 | if (zt->zt_bufsize != 0) |
a6255b7f | 450 | abd_free(zio->io_abd); |
34dc7c2f | 451 | |
a6255b7f | 452 | zio->io_abd = zt->zt_orig_abd; |
b128c09f BB |
453 | zio->io_size = zt->zt_orig_size; |
454 | zio->io_transform_stack = zt->zt_next; | |
34dc7c2f | 455 | |
b128c09f | 456 | kmem_free(zt, sizeof (zio_transform_t)); |
34dc7c2f BB |
457 | } |
458 | } | |
459 | ||
b128c09f BB |
460 | /* |
461 | * ========================================================================== | |
b5256303 | 462 | * I/O transform callbacks for subblocks, decompression, and decryption |
b128c09f BB |
463 | * ========================================================================== |
464 | */ | |
465 | static void | |
a6255b7f | 466 | zio_subblock(zio_t *zio, abd_t *data, uint64_t size) |
b128c09f BB |
467 | { |
468 | ASSERT(zio->io_size > size); | |
469 | ||
470 | if (zio->io_type == ZIO_TYPE_READ) | |
a6255b7f | 471 | abd_copy(data, zio->io_abd, size); |
b128c09f BB |
472 | } |
473 | ||
474 | static void | |
a6255b7f | 475 | zio_decompress(zio_t *zio, abd_t *data, uint64_t size) |
b128c09f | 476 | { |
a6255b7f DQ |
477 | if (zio->io_error == 0) { |
478 | void *tmp = abd_borrow_buf(data, size); | |
479 | int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), | |
10b3c7f5 MN |
480 | zio->io_abd, tmp, zio->io_size, size, |
481 | &zio->io_prop.zp_complevel); | |
a6255b7f DQ |
482 | abd_return_buf_copy(data, tmp, size); |
483 | ||
c3bd3fb4 TC |
484 | if (zio_injection_enabled && ret == 0) |
485 | ret = zio_handle_fault_injection(zio, EINVAL); | |
486 | ||
a6255b7f DQ |
487 | if (ret != 0) |
488 | zio->io_error = SET_ERROR(EIO); | |
489 | } | |
b128c09f BB |
490 | } |
491 | ||
b5256303 TC |
492 | static void |
493 | zio_decrypt(zio_t *zio, abd_t *data, uint64_t size) | |
494 | { | |
495 | int ret; | |
496 | void *tmp; | |
497 | blkptr_t *bp = zio->io_bp; | |
ae76f45c TC |
498 | spa_t *spa = zio->io_spa; |
499 | uint64_t dsobj = zio->io_bookmark.zb_objset; | |
b5256303 TC |
500 | uint64_t lsize = BP_GET_LSIZE(bp); |
501 | dmu_object_type_t ot = BP_GET_TYPE(bp); | |
502 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
503 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
504 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
505 | boolean_t no_crypt = B_FALSE; | |
506 | ||
507 | ASSERT(BP_USES_CRYPT(bp)); | |
508 | ASSERT3U(size, !=, 0); | |
509 | ||
510 | if (zio->io_error != 0) | |
511 | return; | |
512 | ||
513 | /* | |
514 | * Verify the cksum of MACs stored in an indirect bp. It will always | |
515 | * be possible to verify this since it does not require an encryption | |
516 | * key. | |
517 | */ | |
518 | if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) { | |
519 | zio_crypt_decode_mac_bp(bp, mac); | |
520 | ||
521 | if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) { | |
522 | /* | |
523 | * We haven't decompressed the data yet, but | |
524 | * zio_crypt_do_indirect_mac_checksum() requires | |
525 | * decompressed data to be able to parse out the MACs | |
526 | * from the indirect block. We decompress it now and | |
527 | * throw away the result after we are finished. | |
528 | */ | |
529 | tmp = zio_buf_alloc(lsize); | |
530 | ret = zio_decompress_data(BP_GET_COMPRESS(bp), | |
10b3c7f5 MN |
531 | zio->io_abd, tmp, zio->io_size, lsize, |
532 | &zio->io_prop.zp_complevel); | |
b5256303 TC |
533 | if (ret != 0) { |
534 | ret = SET_ERROR(EIO); | |
535 | goto error; | |
536 | } | |
537 | ret = zio_crypt_do_indirect_mac_checksum(B_FALSE, | |
538 | tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac); | |
539 | zio_buf_free(tmp, lsize); | |
540 | } else { | |
541 | ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE, | |
542 | zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac); | |
543 | } | |
544 | abd_copy(data, zio->io_abd, size); | |
545 | ||
be9a5c35 TC |
546 | if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) { |
547 | ret = zio_handle_decrypt_injection(spa, | |
548 | &zio->io_bookmark, ot, ECKSUM); | |
549 | } | |
b5256303 TC |
550 | if (ret != 0) |
551 | goto error; | |
552 | ||
553 | return; | |
554 | } | |
555 | ||
556 | /* | |
557 | * If this is an authenticated block, just check the MAC. It would be | |
4938d01d RY |
558 | * nice to separate this out into its own flag, but when this was done, |
559 | * we had run out of bits in what is now zio_flag_t. Future cleanup | |
560 | * could make this a flag bit. | |
b5256303 TC |
561 | */ |
562 | if (BP_IS_AUTHENTICATED(bp)) { | |
563 | if (ot == DMU_OT_OBJSET) { | |
ae76f45c TC |
564 | ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, |
565 | dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp)); | |
b5256303 TC |
566 | } else { |
567 | zio_crypt_decode_mac_bp(bp, mac); | |
ae76f45c TC |
568 | ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, |
569 | zio->io_abd, size, mac); | |
be9a5c35 TC |
570 | if (zio_injection_enabled && ret == 0) { |
571 | ret = zio_handle_decrypt_injection(spa, | |
572 | &zio->io_bookmark, ot, ECKSUM); | |
573 | } | |
b5256303 TC |
574 | } |
575 | abd_copy(data, zio->io_abd, size); | |
576 | ||
577 | if (ret != 0) | |
578 | goto error; | |
579 | ||
580 | return; | |
581 | } | |
582 | ||
583 | zio_crypt_decode_params_bp(bp, salt, iv); | |
584 | ||
585 | if (ot == DMU_OT_INTENT_LOG) { | |
586 | tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t)); | |
587 | zio_crypt_decode_mac_zil(tmp, mac); | |
588 | abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t)); | |
589 | } else { | |
590 | zio_crypt_decode_mac_bp(bp, mac); | |
591 | } | |
592 | ||
be9a5c35 TC |
593 | ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp), |
594 | BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data, | |
595 | zio->io_abd, &no_crypt); | |
b5256303 TC |
596 | if (no_crypt) |
597 | abd_copy(data, zio->io_abd, size); | |
598 | ||
599 | if (ret != 0) | |
600 | goto error; | |
601 | ||
602 | return; | |
603 | ||
604 | error: | |
605 | /* assert that the key was found unless this was speculative */ | |
be9a5c35 | 606 | ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE)); |
b5256303 TC |
607 | |
608 | /* | |
609 | * If there was a decryption / authentication error return EIO as | |
610 | * the io_error. If this was not a speculative zio, create an ereport. | |
611 | */ | |
612 | if (ret == ECKSUM) { | |
a2c2ed1b | 613 | zio->io_error = SET_ERROR(EIO); |
b5256303 | 614 | if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) { |
431083f7 | 615 | spa_log_error(spa, &zio->io_bookmark, |
493fcce9 | 616 | BP_GET_LOGICAL_BIRTH(zio->io_bp)); |
1144586b | 617 | (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION, |
4f072827 | 618 | spa, NULL, &zio->io_bookmark, zio, 0); |
b5256303 TC |
619 | } |
620 | } else { | |
621 | zio->io_error = ret; | |
622 | } | |
623 | } | |
624 | ||
b128c09f BB |
625 | /* |
626 | * ========================================================================== | |
627 | * I/O parent/child relationships and pipeline interlocks | |
628 | * ========================================================================== | |
629 | */ | |
d164b209 | 630 | zio_t * |
3dfb57a3 | 631 | zio_walk_parents(zio_t *cio, zio_link_t **zl) |
d164b209 | 632 | { |
d164b209 | 633 | list_t *pl = &cio->io_parent_list; |
b128c09f | 634 | |
3dfb57a3 DB |
635 | *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl); |
636 | if (*zl == NULL) | |
d164b209 BB |
637 | return (NULL); |
638 | ||
3dfb57a3 DB |
639 | ASSERT((*zl)->zl_child == cio); |
640 | return ((*zl)->zl_parent); | |
d164b209 BB |
641 | } |
642 | ||
643 | zio_t * | |
3dfb57a3 | 644 | zio_walk_children(zio_t *pio, zio_link_t **zl) |
d164b209 | 645 | { |
d164b209 BB |
646 | list_t *cl = &pio->io_child_list; |
647 | ||
a8b2e306 TC |
648 | ASSERT(MUTEX_HELD(&pio->io_lock)); |
649 | ||
3dfb57a3 DB |
650 | *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl); |
651 | if (*zl == NULL) | |
d164b209 BB |
652 | return (NULL); |
653 | ||
3dfb57a3 DB |
654 | ASSERT((*zl)->zl_parent == pio); |
655 | return ((*zl)->zl_child); | |
d164b209 BB |
656 | } |
657 | ||
658 | zio_t * | |
659 | zio_unique_parent(zio_t *cio) | |
660 | { | |
3dfb57a3 DB |
661 | zio_link_t *zl = NULL; |
662 | zio_t *pio = zio_walk_parents(cio, &zl); | |
d164b209 | 663 | |
3dfb57a3 | 664 | VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL); |
d164b209 BB |
665 | return (pio); |
666 | } | |
667 | ||
668 | void | |
669 | zio_add_child(zio_t *pio, zio_t *cio) | |
b128c09f | 670 | { |
d164b209 BB |
671 | /* |
672 | * Logical I/Os can have logical, gang, or vdev children. | |
673 | * Gang I/Os can have gang or vdev children. | |
674 | * Vdev I/Os can only have vdev children. | |
675 | * The following ASSERT captures all of these constraints. | |
676 | */ | |
1ce23dca | 677 | ASSERT3S(cio->io_child_type, <=, pio->io_child_type); |
d164b209 | 678 | |
3afdc97d AM |
679 | /* Parent should not have READY stage if child doesn't have it. */ |
680 | IMPLY((cio->io_pipeline & ZIO_STAGE_READY) == 0 && | |
681 | (cio->io_child_type != ZIO_CHILD_VDEV), | |
682 | (pio->io_pipeline & ZIO_STAGE_READY) == 0); | |
683 | ||
b4a08730 | 684 | zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); |
d164b209 BB |
685 | zl->zl_parent = pio; |
686 | zl->zl_child = cio; | |
687 | ||
b128c09f | 688 | mutex_enter(&pio->io_lock); |
a8b2e306 | 689 | mutex_enter(&cio->io_lock); |
d164b209 BB |
690 | |
691 | ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); | |
692 | ||
b4a08730 | 693 | uint64_t *countp = pio->io_children[cio->io_child_type]; |
1c27024e | 694 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
b4a08730 | 695 | countp[w] += !cio->io_state[w]; |
d164b209 BB |
696 | |
697 | list_insert_head(&pio->io_child_list, zl); | |
698 | list_insert_head(&cio->io_parent_list, zl); | |
699 | ||
d164b209 | 700 | mutex_exit(&cio->io_lock); |
a8b2e306 | 701 | mutex_exit(&pio->io_lock); |
b128c09f BB |
702 | } |
703 | ||
b4a08730 AM |
704 | void |
705 | zio_add_child_first(zio_t *pio, zio_t *cio) | |
706 | { | |
707 | /* | |
708 | * Logical I/Os can have logical, gang, or vdev children. | |
709 | * Gang I/Os can have gang or vdev children. | |
710 | * Vdev I/Os can only have vdev children. | |
711 | * The following ASSERT captures all of these constraints. | |
712 | */ | |
713 | ASSERT3S(cio->io_child_type, <=, pio->io_child_type); | |
714 | ||
3afdc97d AM |
715 | /* Parent should not have READY stage if child doesn't have it. */ |
716 | IMPLY((cio->io_pipeline & ZIO_STAGE_READY) == 0 && | |
717 | (cio->io_child_type != ZIO_CHILD_VDEV), | |
718 | (pio->io_pipeline & ZIO_STAGE_READY) == 0); | |
719 | ||
b4a08730 AM |
720 | zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); |
721 | zl->zl_parent = pio; | |
722 | zl->zl_child = cio; | |
723 | ||
724 | ASSERT(list_is_empty(&cio->io_parent_list)); | |
725 | list_insert_head(&cio->io_parent_list, zl); | |
726 | ||
727 | mutex_enter(&pio->io_lock); | |
728 | ||
729 | ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); | |
730 | ||
731 | uint64_t *countp = pio->io_children[cio->io_child_type]; | |
732 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) | |
733 | countp[w] += !cio->io_state[w]; | |
734 | ||
735 | list_insert_head(&pio->io_child_list, zl); | |
736 | ||
737 | mutex_exit(&pio->io_lock); | |
738 | } | |
739 | ||
34dc7c2f | 740 | static void |
d164b209 | 741 | zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) |
b128c09f | 742 | { |
d164b209 BB |
743 | ASSERT(zl->zl_parent == pio); |
744 | ASSERT(zl->zl_child == cio); | |
b128c09f BB |
745 | |
746 | mutex_enter(&pio->io_lock); | |
a8b2e306 | 747 | mutex_enter(&cio->io_lock); |
d164b209 BB |
748 | |
749 | list_remove(&pio->io_child_list, zl); | |
750 | list_remove(&cio->io_parent_list, zl); | |
751 | ||
d164b209 | 752 | mutex_exit(&cio->io_lock); |
a8b2e306 | 753 | mutex_exit(&pio->io_lock); |
d164b209 | 754 | kmem_cache_free(zio_link_cache, zl); |
b128c09f BB |
755 | } |
756 | ||
757 | static boolean_t | |
ddc751d5 | 758 | zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait) |
34dc7c2f | 759 | { |
b128c09f BB |
760 | boolean_t waiting = B_FALSE; |
761 | ||
762 | mutex_enter(&zio->io_lock); | |
763 | ASSERT(zio->io_stall == NULL); | |
ddc751d5 GW |
764 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) { |
765 | if (!(ZIO_CHILD_BIT_IS_SET(childbits, c))) | |
766 | continue; | |
767 | ||
768 | uint64_t *countp = &zio->io_children[c][wait]; | |
769 | if (*countp != 0) { | |
770 | zio->io_stage >>= 1; | |
771 | ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN); | |
772 | zio->io_stall = countp; | |
773 | waiting = B_TRUE; | |
774 | break; | |
775 | } | |
b128c09f BB |
776 | } |
777 | mutex_exit(&zio->io_lock); | |
b128c09f BB |
778 | return (waiting); |
779 | } | |
34dc7c2f | 780 | |
bf701a83 BB |
781 | __attribute__((always_inline)) |
782 | static inline void | |
62840030 MA |
783 | zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait, |
784 | zio_t **next_to_executep) | |
b128c09f BB |
785 | { |
786 | uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; | |
787 | int *errorp = &pio->io_child_error[zio->io_child_type]; | |
34dc7c2f | 788 | |
b128c09f BB |
789 | mutex_enter(&pio->io_lock); |
790 | if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) | |
791 | *errorp = zio_worst_error(*errorp, zio->io_error); | |
792 | pio->io_reexecute |= zio->io_reexecute; | |
793 | ASSERT3U(*countp, >, 0); | |
e8b96c60 MA |
794 | |
795 | (*countp)--; | |
796 | ||
797 | if (*countp == 0 && pio->io_stall == countp) { | |
3dfb57a3 DB |
798 | zio_taskq_type_t type = |
799 | pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE : | |
800 | ZIO_TASKQ_INTERRUPT; | |
b128c09f BB |
801 | pio->io_stall = NULL; |
802 | mutex_exit(&pio->io_lock); | |
62840030 | 803 | |
3dfb57a3 | 804 | /* |
62840030 | 805 | * If we can tell the caller to execute this parent next, do |
a6edc0ad RN |
806 | * so. We do this if the parent's zio type matches the child's |
807 | * type, or if it's a zio_null() with no done callback, and so | |
808 | * has no actual work to do. Otherwise dispatch the parent zio | |
809 | * in its own taskq. | |
62840030 MA |
810 | * |
811 | * Having the caller execute the parent when possible reduces | |
812 | * locking on the zio taskq's, reduces context switch | |
813 | * overhead, and has no recursion penalty. Note that one | |
814 | * read from disk typically causes at least 3 zio's: a | |
815 | * zio_null(), the logical zio_read(), and then a physical | |
816 | * zio. When the physical ZIO completes, we are able to call | |
817 | * zio_done() on all 3 of these zio's from one invocation of | |
818 | * zio_execute() by returning the parent back to | |
819 | * zio_execute(). Since the parent isn't executed until this | |
820 | * thread returns back to zio_execute(), the caller should do | |
821 | * so promptly. | |
822 | * | |
823 | * In other cases, dispatching the parent prevents | |
824 | * overflowing the stack when we have deeply nested | |
825 | * parent-child relationships, as we do with the "mega zio" | |
826 | * of writes for spa_sync(), and the chain of ZIL blocks. | |
3dfb57a3 | 827 | */ |
ffd2e15d | 828 | if (next_to_executep != NULL && *next_to_executep == NULL && |
a6edc0ad RN |
829 | (pio->io_type == zio->io_type || |
830 | (pio->io_type == ZIO_TYPE_NULL && !pio->io_done))) { | |
62840030 MA |
831 | *next_to_executep = pio; |
832 | } else { | |
833 | zio_taskq_dispatch(pio, type, B_FALSE); | |
834 | } | |
b128c09f BB |
835 | } else { |
836 | mutex_exit(&pio->io_lock); | |
34dc7c2f BB |
837 | } |
838 | } | |
839 | ||
b128c09f BB |
840 | static void |
841 | zio_inherit_child_errors(zio_t *zio, enum zio_child c) | |
842 | { | |
843 | if (zio->io_child_error[c] != 0 && zio->io_error == 0) | |
844 | zio->io_error = zio->io_child_error[c]; | |
845 | } | |
846 | ||
3dfb57a3 | 847 | int |
64fc7762 | 848 | zio_bookmark_compare(const void *x1, const void *x2) |
3dfb57a3 DB |
849 | { |
850 | const zio_t *z1 = x1; | |
851 | const zio_t *z2 = x2; | |
3dfb57a3 | 852 | |
64fc7762 MA |
853 | if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset) |
854 | return (-1); | |
855 | if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset) | |
856 | return (1); | |
3dfb57a3 | 857 | |
64fc7762 MA |
858 | if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object) |
859 | return (-1); | |
860 | if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object) | |
861 | return (1); | |
3dfb57a3 | 862 | |
64fc7762 MA |
863 | if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level) |
864 | return (-1); | |
865 | if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level) | |
866 | return (1); | |
867 | ||
868 | if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid) | |
869 | return (-1); | |
870 | if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid) | |
871 | return (1); | |
872 | ||
873 | if (z1 < z2) | |
874 | return (-1); | |
875 | if (z1 > z2) | |
876 | return (1); | |
877 | ||
878 | return (0); | |
3dfb57a3 DB |
879 | } |
880 | ||
34dc7c2f BB |
881 | /* |
882 | * ========================================================================== | |
b128c09f | 883 | * Create the various types of I/O (read, write, free, etc) |
34dc7c2f BB |
884 | * ========================================================================== |
885 | */ | |
886 | static zio_t * | |
428870ff | 887 | zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
a6255b7f | 888 | abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done, |
2aa34383 | 889 | void *private, zio_type_t type, zio_priority_t priority, |
4938d01d | 890 | zio_flag_t flags, vdev_t *vd, uint64_t offset, |
2aa34383 DK |
891 | const zbookmark_phys_t *zb, enum zio_stage stage, |
892 | enum zio_stage pipeline) | |
34dc7c2f BB |
893 | { |
894 | zio_t *zio; | |
895 | ||
1b939560 | 896 | IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE); |
2aa34383 | 897 | ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); |
b128c09f BB |
898 | ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); |
899 | ||
900 | ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); | |
901 | ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); | |
902 | ASSERT(vd || stage == ZIO_STAGE_OPEN); | |
34dc7c2f | 903 | |
b5256303 | 904 | IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0); |
2aa34383 | 905 | |
79c76d5b | 906 | zio = kmem_cache_alloc(zio_cache, KM_SLEEP); |
861166b0 | 907 | memset(zio, 0, sizeof (zio_t)); |
3941503c | 908 | |
448d7aaa | 909 | mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL); |
3941503c BB |
910 | cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); |
911 | ||
912 | list_create(&zio->io_parent_list, sizeof (zio_link_t), | |
913 | offsetof(zio_link_t, zl_parent_node)); | |
914 | list_create(&zio->io_child_list, sizeof (zio_link_t), | |
915 | offsetof(zio_link_t, zl_child_node)); | |
4e21fd06 | 916 | metaslab_trace_init(&zio->io_alloc_list); |
d164b209 | 917 | |
b128c09f BB |
918 | if (vd != NULL) |
919 | zio->io_child_type = ZIO_CHILD_VDEV; | |
920 | else if (flags & ZIO_FLAG_GANG_CHILD) | |
921 | zio->io_child_type = ZIO_CHILD_GANG; | |
428870ff BB |
922 | else if (flags & ZIO_FLAG_DDT_CHILD) |
923 | zio->io_child_type = ZIO_CHILD_DDT; | |
b128c09f BB |
924 | else |
925 | zio->io_child_type = ZIO_CHILD_LOGICAL; | |
926 | ||
34dc7c2f | 927 | if (bp != NULL) { |
428870ff | 928 | if (type != ZIO_TYPE_WRITE || |
b4a08730 AM |
929 | zio->io_child_type == ZIO_CHILD_DDT) { |
930 | zio->io_bp_copy = *bp; | |
b128c09f | 931 | zio->io_bp = &zio->io_bp_copy; /* so caller can free */ |
b4a08730 AM |
932 | } else { |
933 | zio->io_bp = (blkptr_t *)bp; | |
934 | } | |
935 | zio->io_bp_orig = *bp; | |
9babb374 | 936 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) |
b128c09f | 937 | zio->io_logical = zio; |
9babb374 BB |
938 | if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) |
939 | pipeline |= ZIO_GANG_STAGES; | |
34dc7c2f | 940 | } |
b128c09f BB |
941 | |
942 | zio->io_spa = spa; | |
943 | zio->io_txg = txg; | |
34dc7c2f BB |
944 | zio->io_done = done; |
945 | zio->io_private = private; | |
946 | zio->io_type = type; | |
947 | zio->io_priority = priority; | |
b128c09f BB |
948 | zio->io_vd = vd; |
949 | zio->io_offset = offset; | |
a6255b7f | 950 | zio->io_orig_abd = zio->io_abd = data; |
2aa34383 DK |
951 | zio->io_orig_size = zio->io_size = psize; |
952 | zio->io_lsize = lsize; | |
b128c09f BB |
953 | zio->io_orig_flags = zio->io_flags = flags; |
954 | zio->io_orig_stage = zio->io_stage = stage; | |
955 | zio->io_orig_pipeline = zio->io_pipeline = pipeline; | |
3dfb57a3 | 956 | zio->io_pipeline_trace = ZIO_STAGE_OPEN; |
3bd4df38 | 957 | zio->io_allocator = ZIO_ALLOCATOR_NONE; |
34dc7c2f | 958 | |
3afdc97d AM |
959 | zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY) || |
960 | (pipeline & ZIO_STAGE_READY) == 0; | |
d164b209 BB |
961 | zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); |
962 | ||
b128c09f BB |
963 | if (zb != NULL) |
964 | zio->io_bookmark = *zb; | |
965 | ||
966 | if (pio != NULL) { | |
1b50749c | 967 | zio->io_metaslab_class = pio->io_metaslab_class; |
b128c09f | 968 | if (zio->io_logical == NULL) |
34dc7c2f | 969 | zio->io_logical = pio->io_logical; |
9babb374 BB |
970 | if (zio->io_child_type == ZIO_CHILD_GANG) |
971 | zio->io_gang_leader = pio->io_gang_leader; | |
b4a08730 | 972 | zio_add_child_first(pio, zio); |
34dc7c2f BB |
973 | } |
974 | ||
a38718a6 GA |
975 | taskq_init_ent(&zio->io_tqent); |
976 | ||
34dc7c2f BB |
977 | return (zio); |
978 | } | |
979 | ||
e8cf3a4f | 980 | void |
b128c09f | 981 | zio_destroy(zio_t *zio) |
34dc7c2f | 982 | { |
4e21fd06 | 983 | metaslab_trace_fini(&zio->io_alloc_list); |
3941503c BB |
984 | list_destroy(&zio->io_parent_list); |
985 | list_destroy(&zio->io_child_list); | |
986 | mutex_destroy(&zio->io_lock); | |
987 | cv_destroy(&zio->io_cv); | |
b128c09f | 988 | kmem_cache_free(zio_cache, zio); |
34dc7c2f BB |
989 | } |
990 | ||
3afdc97d AM |
991 | /* |
992 | * ZIO intended to be between others. Provides synchronization at READY | |
993 | * and DONE pipeline stages and calls the respective callbacks. | |
994 | */ | |
34dc7c2f | 995 | zio_t * |
d164b209 | 996 | zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, |
4938d01d | 997 | void *private, zio_flag_t flags) |
34dc7c2f BB |
998 | { |
999 | zio_t *zio; | |
1000 | ||
2aa34383 | 1001 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
d164b209 | 1002 | ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
b128c09f | 1003 | ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); |
34dc7c2f BB |
1004 | |
1005 | return (zio); | |
1006 | } | |
1007 | ||
3afdc97d AM |
1008 | /* |
1009 | * ZIO intended to be a root of a tree. Unlike null ZIO does not have a | |
1010 | * READY pipeline stage (is ready on creation), so it should not be used | |
1011 | * as child of any ZIO that may need waiting for grandchildren READY stage | |
1012 | * (any other ZIO type). | |
1013 | */ | |
34dc7c2f | 1014 | zio_t * |
4938d01d | 1015 | zio_root(spa_t *spa, zio_done_func_t *done, void *private, zio_flag_t flags) |
34dc7c2f | 1016 | { |
3afdc97d AM |
1017 | zio_t *zio; |
1018 | ||
1019 | zio = zio_create(NULL, spa, 0, NULL, NULL, 0, 0, done, private, | |
1020 | ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, NULL, 0, NULL, | |
1021 | ZIO_STAGE_OPEN, ZIO_ROOT_PIPELINE); | |
1022 | ||
1023 | return (zio); | |
34dc7c2f BB |
1024 | } |
1025 | ||
bc67cba7 PZ |
1026 | static int |
1027 | zfs_blkptr_verify_log(spa_t *spa, const blkptr_t *bp, | |
1028 | enum blk_verify_flag blk_verify, const char *fmt, ...) | |
1029 | { | |
1030 | va_list adx; | |
1031 | char buf[256]; | |
1032 | ||
1033 | va_start(adx, fmt); | |
1034 | (void) vsnprintf(buf, sizeof (buf), fmt, adx); | |
1035 | va_end(adx); | |
1036 | ||
3095ca91 MA |
1037 | zfs_dbgmsg("bad blkptr at %px: " |
1038 | "DVA[0]=%#llx/%#llx " | |
1039 | "DVA[1]=%#llx/%#llx " | |
1040 | "DVA[2]=%#llx/%#llx " | |
1041 | "prop=%#llx " | |
1042 | "pad=%#llx,%#llx " | |
1043 | "phys_birth=%#llx " | |
1044 | "birth=%#llx " | |
1045 | "fill=%#llx " | |
1046 | "cksum=%#llx/%#llx/%#llx/%#llx", | |
1047 | bp, | |
1048 | (long long)bp->blk_dva[0].dva_word[0], | |
1049 | (long long)bp->blk_dva[0].dva_word[1], | |
1050 | (long long)bp->blk_dva[1].dva_word[0], | |
1051 | (long long)bp->blk_dva[1].dva_word[1], | |
1052 | (long long)bp->blk_dva[2].dva_word[0], | |
1053 | (long long)bp->blk_dva[2].dva_word[1], | |
1054 | (long long)bp->blk_prop, | |
1055 | (long long)bp->blk_pad[0], | |
1056 | (long long)bp->blk_pad[1], | |
493fcce9 GW |
1057 | (long long)BP_GET_PHYSICAL_BIRTH(bp), |
1058 | (long long)BP_GET_LOGICAL_BIRTH(bp), | |
3095ca91 MA |
1059 | (long long)bp->blk_fill, |
1060 | (long long)bp->blk_cksum.zc_word[0], | |
1061 | (long long)bp->blk_cksum.zc_word[1], | |
1062 | (long long)bp->blk_cksum.zc_word[2], | |
1063 | (long long)bp->blk_cksum.zc_word[3]); | |
bc67cba7 PZ |
1064 | switch (blk_verify) { |
1065 | case BLK_VERIFY_HALT: | |
1066 | zfs_panic_recover("%s: %s", spa_name(spa), buf); | |
1067 | break; | |
1068 | case BLK_VERIFY_LOG: | |
1069 | zfs_dbgmsg("%s: %s", spa_name(spa), buf); | |
1070 | break; | |
1071 | case BLK_VERIFY_ONLY: | |
1072 | break; | |
1073 | } | |
1074 | ||
1075 | return (1); | |
1076 | } | |
1077 | ||
1078 | /* | |
1079 | * Verify the block pointer fields contain reasonable values. This means | |
1080 | * it only contains known object types, checksum/compression identifiers, | |
1081 | * block sizes within the maximum allowed limits, valid DVAs, etc. | |
1082 | * | |
1083 | * If everything checks out B_TRUE is returned. The zfs_blkptr_verify | |
1084 | * argument controls the behavior when an invalid field is detected. | |
1085 | * | |
3095ca91 MA |
1086 | * Values for blk_verify_flag: |
1087 | * BLK_VERIFY_ONLY: evaluate the block | |
1088 | * BLK_VERIFY_LOG: evaluate the block and log problems | |
1089 | * BLK_VERIFY_HALT: call zfs_panic_recover on error | |
1090 | * | |
1091 | * Values for blk_config_flag: | |
1092 | * BLK_CONFIG_HELD: caller holds SCL_VDEV for writer | |
1093 | * BLK_CONFIG_NEEDED: caller holds no config lock, SCL_VDEV will be | |
1094 | * obtained for reader | |
1095 | * BLK_CONFIG_SKIP: skip checks which require SCL_VDEV, for better | |
1096 | * performance | |
bc67cba7 PZ |
1097 | */ |
1098 | boolean_t | |
3095ca91 MA |
1099 | zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp, |
1100 | enum blk_config_flag blk_config, enum blk_verify_flag blk_verify) | |
63e3a861 | 1101 | { |
bc67cba7 PZ |
1102 | int errors = 0; |
1103 | ||
63e3a861 | 1104 | if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { |
bc67cba7 | 1105 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1106 | "blkptr at %px has invalid TYPE %llu", |
63e3a861 MA |
1107 | bp, (longlong_t)BP_GET_TYPE(bp)); |
1108 | } | |
2cd0f98f | 1109 | if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS) { |
bc67cba7 | 1110 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1111 | "blkptr at %px has invalid CHECKSUM %llu", |
63e3a861 MA |
1112 | bp, (longlong_t)BP_GET_CHECKSUM(bp)); |
1113 | } | |
2cd0f98f | 1114 | if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS) { |
bc67cba7 | 1115 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1116 | "blkptr at %px has invalid COMPRESS %llu", |
63e3a861 MA |
1117 | bp, (longlong_t)BP_GET_COMPRESS(bp)); |
1118 | } | |
1119 | if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
bc67cba7 | 1120 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1121 | "blkptr at %px has invalid LSIZE %llu", |
63e3a861 MA |
1122 | bp, (longlong_t)BP_GET_LSIZE(bp)); |
1123 | } | |
1124 | if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
bc67cba7 | 1125 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1126 | "blkptr at %px has invalid PSIZE %llu", |
63e3a861 MA |
1127 | bp, (longlong_t)BP_GET_PSIZE(bp)); |
1128 | } | |
1129 | ||
1130 | if (BP_IS_EMBEDDED(bp)) { | |
746d4a45 | 1131 | if (BPE_GET_ETYPE(bp) >= NUM_BP_EMBEDDED_TYPES) { |
bc67cba7 | 1132 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1133 | "blkptr at %px has invalid ETYPE %llu", |
63e3a861 MA |
1134 | bp, (longlong_t)BPE_GET_ETYPE(bp)); |
1135 | } | |
1136 | } | |
1137 | ||
6cb8e530 PZ |
1138 | /* |
1139 | * Do not verify individual DVAs if the config is not trusted. This | |
1140 | * will be done once the zio is executed in vdev_mirror_map_alloc. | |
1141 | */ | |
1142 | if (!spa->spa_trust_config) | |
b9ec4a15 | 1143 | return (errors == 0); |
6cb8e530 | 1144 | |
3095ca91 MA |
1145 | switch (blk_config) { |
1146 | case BLK_CONFIG_HELD: | |
dc04a8c7 | 1147 | ASSERT(spa_config_held(spa, SCL_VDEV, RW_WRITER)); |
3095ca91 MA |
1148 | break; |
1149 | case BLK_CONFIG_NEEDED: | |
1150 | spa_config_enter(spa, SCL_VDEV, bp, RW_READER); | |
1151 | break; | |
1152 | case BLK_CONFIG_SKIP: | |
1153 | return (errors == 0); | |
1154 | default: | |
1155 | panic("invalid blk_config %u", blk_config); | |
1156 | } | |
1157 | ||
63e3a861 MA |
1158 | /* |
1159 | * Pool-specific checks. | |
1160 | * | |
493fcce9 GW |
1161 | * Note: it would be nice to verify that the logical birth |
1162 | * and physical birth are not too large. However, | |
1163 | * spa_freeze() allows the birth time of log blocks (and | |
1164 | * dmu_sync()-ed blocks that are in the log) to be arbitrarily | |
1165 | * large. | |
63e3a861 | 1166 | */ |
1c27024e | 1167 | for (int i = 0; i < BP_GET_NDVAS(bp); i++) { |
2b56a634 MA |
1168 | const dva_t *dva = &bp->blk_dva[i]; |
1169 | uint64_t vdevid = DVA_GET_VDEV(dva); | |
1c27024e | 1170 | |
63e3a861 | 1171 | if (vdevid >= spa->spa_root_vdev->vdev_children) { |
bc67cba7 | 1172 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1173 | "blkptr at %px DVA %u has invalid VDEV %llu", |
63e3a861 | 1174 | bp, i, (longlong_t)vdevid); |
ee3a23b8 | 1175 | continue; |
63e3a861 | 1176 | } |
1c27024e | 1177 | vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; |
63e3a861 | 1178 | if (vd == NULL) { |
bc67cba7 | 1179 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1180 | "blkptr at %px DVA %u has invalid VDEV %llu", |
63e3a861 | 1181 | bp, i, (longlong_t)vdevid); |
ee3a23b8 | 1182 | continue; |
63e3a861 MA |
1183 | } |
1184 | if (vd->vdev_ops == &vdev_hole_ops) { | |
bc67cba7 | 1185 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1186 | "blkptr at %px DVA %u has hole VDEV %llu", |
63e3a861 | 1187 | bp, i, (longlong_t)vdevid); |
ee3a23b8 | 1188 | continue; |
63e3a861 MA |
1189 | } |
1190 | if (vd->vdev_ops == &vdev_missing_ops) { | |
1191 | /* | |
1192 | * "missing" vdevs are valid during import, but we | |
1193 | * don't have their detailed info (e.g. asize), so | |
1194 | * we can't perform any more checks on them. | |
1195 | */ | |
1196 | continue; | |
1197 | } | |
2b56a634 MA |
1198 | uint64_t offset = DVA_GET_OFFSET(dva); |
1199 | uint64_t asize = DVA_GET_ASIZE(dva); | |
1200 | if (DVA_GET_GANG(dva)) | |
1201 | asize = vdev_gang_header_asize(vd); | |
63e3a861 | 1202 | if (offset + asize > vd->vdev_asize) { |
bc67cba7 | 1203 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1204 | "blkptr at %px DVA %u has invalid OFFSET %llu", |
63e3a861 MA |
1205 | bp, i, (longlong_t)offset); |
1206 | } | |
1207 | } | |
3095ca91 | 1208 | if (blk_config == BLK_CONFIG_NEEDED) |
dc04a8c7 | 1209 | spa_config_exit(spa, SCL_VDEV, bp); |
bc67cba7 PZ |
1210 | |
1211 | return (errors == 0); | |
63e3a861 MA |
1212 | } |
1213 | ||
6cb8e530 PZ |
1214 | boolean_t |
1215 | zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp) | |
1216 | { | |
14e4e3cb | 1217 | (void) bp; |
6cb8e530 PZ |
1218 | uint64_t vdevid = DVA_GET_VDEV(dva); |
1219 | ||
1220 | if (vdevid >= spa->spa_root_vdev->vdev_children) | |
1221 | return (B_FALSE); | |
1222 | ||
1223 | vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; | |
1224 | if (vd == NULL) | |
1225 | return (B_FALSE); | |
1226 | ||
1227 | if (vd->vdev_ops == &vdev_hole_ops) | |
1228 | return (B_FALSE); | |
1229 | ||
1230 | if (vd->vdev_ops == &vdev_missing_ops) { | |
1231 | return (B_FALSE); | |
1232 | } | |
1233 | ||
1234 | uint64_t offset = DVA_GET_OFFSET(dva); | |
1235 | uint64_t asize = DVA_GET_ASIZE(dva); | |
1236 | ||
2b56a634 MA |
1237 | if (DVA_GET_GANG(dva)) |
1238 | asize = vdev_gang_header_asize(vd); | |
6cb8e530 PZ |
1239 | if (offset + asize > vd->vdev_asize) |
1240 | return (B_FALSE); | |
1241 | ||
1242 | return (B_TRUE); | |
1243 | } | |
1244 | ||
34dc7c2f | 1245 | zio_t * |
b128c09f | 1246 | zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, |
a6255b7f | 1247 | abd_t *data, uint64_t size, zio_done_func_t *done, void *private, |
4938d01d | 1248 | zio_priority_t priority, zio_flag_t flags, const zbookmark_phys_t *zb) |
34dc7c2f BB |
1249 | { |
1250 | zio_t *zio; | |
1251 | ||
493fcce9 | 1252 | zio = zio_create(pio, spa, BP_GET_BIRTH(bp), bp, |
2aa34383 | 1253 | data, size, size, done, private, |
b128c09f | 1254 | ZIO_TYPE_READ, priority, flags, NULL, 0, zb, |
428870ff BB |
1255 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
1256 | ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); | |
34dc7c2f | 1257 | |
b128c09f BB |
1258 | return (zio); |
1259 | } | |
34dc7c2f | 1260 | |
34dc7c2f | 1261 | zio_t * |
b128c09f | 1262 | zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, |
a6255b7f | 1263 | abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp, |
bc77ba73 | 1264 | zio_done_func_t *ready, zio_done_func_t *children_ready, |
ccec7fbe AM |
1265 | zio_done_func_t *done, void *private, zio_priority_t priority, |
1266 | zio_flag_t flags, const zbookmark_phys_t *zb) | |
34dc7c2f BB |
1267 | { |
1268 | zio_t *zio; | |
1269 | ||
b128c09f BB |
1270 | ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && |
1271 | zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && | |
1272 | zp->zp_compress >= ZIO_COMPRESS_OFF && | |
1273 | zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && | |
9ae529ec | 1274 | DMU_OT_IS_VALID(zp->zp_type) && |
b128c09f | 1275 | zp->zp_level < 32 && |
428870ff | 1276 | zp->zp_copies > 0 && |
03c6040b | 1277 | zp->zp_copies <= spa_max_replication(spa)); |
34dc7c2f | 1278 | |
2aa34383 | 1279 | zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private, |
b128c09f | 1280 | ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, |
428870ff BB |
1281 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
1282 | ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); | |
34dc7c2f BB |
1283 | |
1284 | zio->io_ready = ready; | |
bc77ba73 | 1285 | zio->io_children_ready = children_ready; |
b128c09f | 1286 | zio->io_prop = *zp; |
34dc7c2f | 1287 | |
9b67f605 MA |
1288 | /* |
1289 | * Data can be NULL if we are going to call zio_write_override() to | |
1290 | * provide the already-allocated BP. But we may need the data to | |
1291 | * verify a dedup hit (if requested). In this case, don't try to | |
b5256303 TC |
1292 | * dedup (just take the already-allocated BP verbatim). Encrypted |
1293 | * dedup blocks need data as well so we also disable dedup in this | |
1294 | * case. | |
9b67f605 | 1295 | */ |
b5256303 TC |
1296 | if (data == NULL && |
1297 | (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) { | |
9b67f605 MA |
1298 | zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; |
1299 | } | |
1300 | ||
34dc7c2f BB |
1301 | return (zio); |
1302 | } | |
1303 | ||
1304 | zio_t * | |
a6255b7f | 1305 | zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data, |
e8b96c60 | 1306 | uint64_t size, zio_done_func_t *done, void *private, |
4938d01d | 1307 | zio_priority_t priority, zio_flag_t flags, zbookmark_phys_t *zb) |
34dc7c2f BB |
1308 | { |
1309 | zio_t *zio; | |
1310 | ||
2aa34383 | 1311 | zio = zio_create(pio, spa, txg, bp, data, size, size, done, private, |
3dfb57a3 | 1312 | ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb, |
b128c09f | 1313 | ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); |
34dc7c2f BB |
1314 | |
1315 | return (zio); | |
1316 | } | |
1317 | ||
428870ff | 1318 | void |
67a1b037 PJD |
1319 | zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite, |
1320 | boolean_t brtwrite) | |
428870ff BB |
1321 | { |
1322 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
1323 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1324 | ASSERT(zio->io_stage == ZIO_STAGE_OPEN); | |
1325 | ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); | |
67a1b037 | 1326 | ASSERT(!brtwrite || !nopwrite); |
428870ff | 1327 | |
03c6040b GW |
1328 | /* |
1329 | * We must reset the io_prop to match the values that existed | |
1330 | * when the bp was first written by dmu_sync() keeping in mind | |
1331 | * that nopwrite and dedup are mutually exclusive. | |
1332 | */ | |
1333 | zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; | |
1334 | zio->io_prop.zp_nopwrite = nopwrite; | |
67a1b037 | 1335 | zio->io_prop.zp_brtwrite = brtwrite; |
428870ff BB |
1336 | zio->io_prop.zp_copies = copies; |
1337 | zio->io_bp_override = bp; | |
1338 | } | |
1339 | ||
1340 | void | |
1341 | zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) | |
1342 | { | |
9b67f605 | 1343 | |
3095ca91 | 1344 | (void) zfs_blkptr_verify(spa, bp, BLK_CONFIG_NEEDED, BLK_VERIFY_HALT); |
a1d477c2 | 1345 | |
9b67f605 MA |
1346 | /* |
1347 | * The check for EMBEDDED is a performance optimization. We | |
1348 | * process the free here (by ignoring it) rather than | |
1349 | * putting it on the list and then processing it in zio_free_sync(). | |
1350 | */ | |
1351 | if (BP_IS_EMBEDDED(bp)) | |
1352 | return; | |
2883cad5 MA |
1353 | |
1354 | /* | |
1355 | * Frees that are for the currently-syncing txg, are not going to be | |
1356 | * deferred, and which will not need to do a read (i.e. not GANG or | |
1357 | * DEDUP), can be processed immediately. Otherwise, put them on the | |
1358 | * in-memory list for later processing. | |
93e28d66 SD |
1359 | * |
1360 | * Note that we only defer frees after zfs_sync_pass_deferred_free | |
1361 | * when the log space map feature is disabled. [see relevant comment | |
1362 | * in spa_sync_iterate_to_convergence()] | |
2883cad5 | 1363 | */ |
93e28d66 SD |
1364 | if (BP_IS_GANG(bp) || |
1365 | BP_GET_DEDUP(bp) || | |
2883cad5 | 1366 | txg != spa->spa_syncing_txg || |
93e28d66 | 1367 | (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free && |
67a1b037 PJD |
1368 | !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) || |
1369 | brt_maybe_exists(spa, bp)) { | |
6694ca55 | 1370 | metaslab_check_free(spa, bp); |
2883cad5 MA |
1371 | bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); |
1372 | } else { | |
9cdf7b1f | 1373 | VERIFY3P(zio_free_sync(NULL, spa, txg, bp, 0), ==, NULL); |
2883cad5 | 1374 | } |
428870ff BB |
1375 | } |
1376 | ||
9cdf7b1f MA |
1377 | /* |
1378 | * To improve performance, this function may return NULL if we were able | |
1379 | * to do the free immediately. This avoids the cost of creating a zio | |
1380 | * (and linking it to the parent, etc). | |
1381 | */ | |
34dc7c2f | 1382 | zio_t * |
428870ff | 1383 | zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
4938d01d | 1384 | zio_flag_t flags) |
34dc7c2f | 1385 | { |
428870ff BB |
1386 | ASSERT(!BP_IS_HOLE(bp)); |
1387 | ASSERT(spa_syncing_txg(spa) == txg); | |
34dc7c2f | 1388 | |
9b67f605 | 1389 | if (BP_IS_EMBEDDED(bp)) |
9cdf7b1f | 1390 | return (NULL); |
9b67f605 | 1391 | |
13fe0198 | 1392 | metaslab_check_free(spa, bp); |
8c841793 | 1393 | arc_freed(spa, bp); |
d4a72f23 | 1394 | dsl_scan_freed(spa, bp); |
13fe0198 | 1395 | |
67a1b037 PJD |
1396 | if (BP_IS_GANG(bp) || |
1397 | BP_GET_DEDUP(bp) || | |
1398 | brt_maybe_exists(spa, bp)) { | |
9cdf7b1f | 1399 | /* |
67a1b037 PJD |
1400 | * GANG, DEDUP and BRT blocks can induce a read (for the gang |
1401 | * block header, the DDT or the BRT), so issue them | |
1402 | * asynchronously so that this thread is not tied up. | |
9cdf7b1f MA |
1403 | */ |
1404 | enum zio_stage stage = | |
1405 | ZIO_FREE_PIPELINE | ZIO_STAGE_ISSUE_ASYNC; | |
2883cad5 | 1406 | |
9cdf7b1f MA |
1407 | return (zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
1408 | BP_GET_PSIZE(bp), NULL, NULL, | |
1409 | ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, | |
1410 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage)); | |
1411 | } else { | |
1412 | metaslab_free(spa, bp, txg, B_FALSE); | |
1413 | return (NULL); | |
1414 | } | |
34dc7c2f BB |
1415 | } |
1416 | ||
1417 | zio_t * | |
428870ff | 1418 | zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
4938d01d | 1419 | zio_done_func_t *done, void *private, zio_flag_t flags) |
34dc7c2f BB |
1420 | { |
1421 | zio_t *zio; | |
1422 | ||
3095ca91 MA |
1423 | (void) zfs_blkptr_verify(spa, bp, (flags & ZIO_FLAG_CONFIG_WRITER) ? |
1424 | BLK_CONFIG_HELD : BLK_CONFIG_NEEDED, BLK_VERIFY_HALT); | |
9b67f605 MA |
1425 | |
1426 | if (BP_IS_EMBEDDED(bp)) | |
1427 | return (zio_null(pio, spa, NULL, NULL, NULL, 0)); | |
1428 | ||
34dc7c2f BB |
1429 | /* |
1430 | * A claim is an allocation of a specific block. Claims are needed | |
1431 | * to support immediate writes in the intent log. The issue is that | |
1432 | * immediate writes contain committed data, but in a txg that was | |
1433 | * *not* committed. Upon opening the pool after an unclean shutdown, | |
1434 | * the intent log claims all blocks that contain immediate write data | |
1435 | * so that the SPA knows they're in use. | |
1436 | * | |
1437 | * All claims *must* be resolved in the first txg -- before the SPA | |
1438 | * starts allocating blocks -- so that nothing is allocated twice. | |
428870ff | 1439 | * If txg == 0 we just verify that the block is claimable. |
34dc7c2f | 1440 | */ |
493fcce9 | 1441 | ASSERT3U(BP_GET_LOGICAL_BIRTH(&spa->spa_uberblock.ub_rootbp), <, |
d2734cce SD |
1442 | spa_min_claim_txg(spa)); |
1443 | ASSERT(txg == spa_min_claim_txg(spa) || txg == 0); | |
76d04993 | 1444 | ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(8) */ |
34dc7c2f | 1445 | |
b128c09f | 1446 | zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
2aa34383 DK |
1447 | BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, |
1448 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); | |
3dfb57a3 | 1449 | ASSERT0(zio->io_queued_timestamp); |
34dc7c2f BB |
1450 | |
1451 | return (zio); | |
1452 | } | |
1453 | ||
1b939560 BB |
1454 | zio_t * |
1455 | zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
1456 | zio_done_func_t *done, void *private, zio_priority_t priority, | |
4938d01d | 1457 | zio_flag_t flags, enum trim_flag trim_flags) |
1b939560 BB |
1458 | { |
1459 | zio_t *zio; | |
1460 | ||
1461 | ASSERT0(vd->vdev_children); | |
1462 | ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift)); | |
1463 | ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift)); | |
1464 | ASSERT3U(size, !=, 0); | |
1465 | ||
1466 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done, | |
1467 | private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL, | |
1468 | vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE); | |
1469 | zio->io_trim_flags = trim_flags; | |
1470 | ||
1471 | return (zio); | |
1472 | } | |
1473 | ||
34dc7c2f BB |
1474 | zio_t * |
1475 | zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
a6255b7f | 1476 | abd_t *data, int checksum, zio_done_func_t *done, void *private, |
4938d01d | 1477 | zio_priority_t priority, zio_flag_t flags, boolean_t labels) |
34dc7c2f BB |
1478 | { |
1479 | zio_t *zio; | |
34dc7c2f | 1480 | |
b128c09f BB |
1481 | ASSERT(vd->vdev_children == 0); |
1482 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
1483 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
1484 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 1485 | |
2aa34383 DK |
1486 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
1487 | private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
1488 | offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); | |
34dc7c2f | 1489 | |
b128c09f | 1490 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f BB |
1491 | |
1492 | return (zio); | |
1493 | } | |
1494 | ||
1495 | zio_t * | |
1496 | zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
a6255b7f | 1497 | abd_t *data, int checksum, zio_done_func_t *done, void *private, |
4938d01d | 1498 | zio_priority_t priority, zio_flag_t flags, boolean_t labels) |
34dc7c2f | 1499 | { |
34dc7c2f | 1500 | zio_t *zio; |
34dc7c2f | 1501 | |
b128c09f BB |
1502 | ASSERT(vd->vdev_children == 0); |
1503 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
1504 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
1505 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 1506 | |
2aa34383 DK |
1507 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
1508 | private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
1509 | offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); | |
34dc7c2f | 1510 | |
b128c09f | 1511 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f | 1512 | |
3c67d83a | 1513 | if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { |
34dc7c2f | 1514 | /* |
428870ff | 1515 | * zec checksums are necessarily destructive -- they modify |
b128c09f | 1516 | * the end of the write buffer to hold the verifier/checksum. |
34dc7c2f | 1517 | * Therefore, we must make a local copy in case the data is |
b128c09f | 1518 | * being written to multiple places in parallel. |
34dc7c2f | 1519 | */ |
a6255b7f DQ |
1520 | abd_t *wbuf = abd_alloc_sametype(data, size); |
1521 | abd_copy(wbuf, data, size); | |
1522 | ||
b128c09f | 1523 | zio_push_transform(zio, wbuf, size, size, NULL); |
34dc7c2f BB |
1524 | } |
1525 | ||
1526 | return (zio); | |
1527 | } | |
1528 | ||
1529 | /* | |
b128c09f | 1530 | * Create a child I/O to do some work for us. |
34dc7c2f BB |
1531 | */ |
1532 | zio_t * | |
b128c09f | 1533 | zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, |
4ea3f864 | 1534 | abd_t *data, uint64_t size, int type, zio_priority_t priority, |
4938d01d | 1535 | zio_flag_t flags, zio_done_func_t *done, void *private) |
34dc7c2f | 1536 | { |
428870ff | 1537 | enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; |
b128c09f BB |
1538 | zio_t *zio; |
1539 | ||
a1d477c2 MA |
1540 | /* |
1541 | * vdev child I/Os do not propagate their error to the parent. | |
1542 | * Therefore, for correct operation the caller *must* check for | |
1543 | * and handle the error in the child i/o's done callback. | |
1544 | * The only exceptions are i/os that we don't care about | |
1545 | * (OPTIONAL or REPAIR). | |
1546 | */ | |
1547 | ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) || | |
1548 | done != NULL); | |
1549 | ||
34dc7c2f BB |
1550 | if (type == ZIO_TYPE_READ && bp != NULL) { |
1551 | /* | |
1552 | * If we have the bp, then the child should perform the | |
1553 | * checksum and the parent need not. This pushes error | |
1554 | * detection as close to the leaves as possible and | |
1555 | * eliminates redundant checksums in the interior nodes. | |
1556 | */ | |
428870ff BB |
1557 | pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; |
1558 | pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; | |
34dc7c2f BB |
1559 | } |
1560 | ||
a1d477c2 MA |
1561 | if (vd->vdev_ops->vdev_op_leaf) { |
1562 | ASSERT0(vd->vdev_children); | |
b128c09f | 1563 | offset += VDEV_LABEL_START_SIZE; |
a1d477c2 | 1564 | } |
b128c09f | 1565 | |
a1d477c2 | 1566 | flags |= ZIO_VDEV_CHILD_FLAGS(pio); |
428870ff BB |
1567 | |
1568 | /* | |
1569 | * If we've decided to do a repair, the write is not speculative -- | |
1570 | * even if the original read was. | |
1571 | */ | |
1572 | if (flags & ZIO_FLAG_IO_REPAIR) | |
1573 | flags &= ~ZIO_FLAG_SPECULATIVE; | |
1574 | ||
3dfb57a3 DB |
1575 | /* |
1576 | * If we're creating a child I/O that is not associated with a | |
1577 | * top-level vdev, then the child zio is not an allocating I/O. | |
1578 | * If this is a retried I/O then we ignore it since we will | |
1579 | * have already processed the original allocating I/O. | |
1580 | */ | |
1581 | if (flags & ZIO_FLAG_IO_ALLOCATING && | |
1582 | (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) { | |
cc99f275 DB |
1583 | ASSERT(pio->io_metaslab_class != NULL); |
1584 | ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled); | |
3dfb57a3 DB |
1585 | ASSERT(type == ZIO_TYPE_WRITE); |
1586 | ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE); | |
1587 | ASSERT(!(flags & ZIO_FLAG_IO_REPAIR)); | |
1588 | ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) || | |
1589 | pio->io_child_type == ZIO_CHILD_GANG); | |
1590 | ||
1591 | flags &= ~ZIO_FLAG_IO_ALLOCATING; | |
1592 | } | |
1593 | ||
2aa34383 | 1594 | zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size, |
428870ff BB |
1595 | done, private, type, priority, flags, vd, offset, &pio->io_bookmark, |
1596 | ZIO_STAGE_VDEV_IO_START >> 1, pipeline); | |
3dfb57a3 | 1597 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); |
34dc7c2f | 1598 | |
b128c09f | 1599 | return (zio); |
34dc7c2f BB |
1600 | } |
1601 | ||
b128c09f | 1602 | zio_t * |
a6255b7f | 1603 | zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size, |
4938d01d | 1604 | zio_type_t type, zio_priority_t priority, zio_flag_t flags, |
e9aa730c | 1605 | zio_done_func_t *done, void *private) |
34dc7c2f | 1606 | { |
b128c09f | 1607 | zio_t *zio; |
34dc7c2f | 1608 | |
b128c09f | 1609 | ASSERT(vd->vdev_ops->vdev_op_leaf); |
34dc7c2f | 1610 | |
b128c09f | 1611 | zio = zio_create(NULL, vd->vdev_spa, 0, NULL, |
2aa34383 | 1612 | data, size, size, done, private, type, priority, |
e8b96c60 | 1613 | flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, |
b128c09f | 1614 | vd, offset, NULL, |
428870ff | 1615 | ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); |
34dc7c2f | 1616 | |
b128c09f | 1617 | return (zio); |
34dc7c2f BB |
1618 | } |
1619 | ||
cac416f1 RN |
1620 | |
1621 | /* | |
1622 | * Send a flush command to the given vdev. Unlike most zio creation functions, | |
1623 | * the flush zios are issued immediately. You can wait on pio to pause until | |
1624 | * the flushes complete. | |
1625 | */ | |
34dc7c2f | 1626 | void |
5a3bffab | 1627 | zio_flush(zio_t *pio, vdev_t *vd) |
34dc7c2f | 1628 | { |
cac416f1 RN |
1629 | const zio_flag_t flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | |
1630 | ZIO_FLAG_DONT_RETRY; | |
1631 | ||
5a3bffab AM |
1632 | if (vd->vdev_nowritecache) |
1633 | return; | |
cac416f1 | 1634 | |
5a3bffab | 1635 | if (vd->vdev_children == 0) { |
c9c838aa | 1636 | zio_nowait(zio_create(pio, vd->vdev_spa, 0, NULL, NULL, 0, 0, |
d7605ae7 RN |
1637 | NULL, NULL, ZIO_TYPE_FLUSH, ZIO_PRIORITY_NOW, flags, vd, 0, |
1638 | NULL, ZIO_STAGE_OPEN, ZIO_FLUSH_PIPELINE)); | |
5a3bffab AM |
1639 | } else { |
1640 | for (uint64_t c = 0; c < vd->vdev_children; c++) | |
1641 | zio_flush(pio, vd->vdev_child[c]); | |
1642 | } | |
34dc7c2f BB |
1643 | } |
1644 | ||
428870ff BB |
1645 | void |
1646 | zio_shrink(zio_t *zio, uint64_t size) | |
1647 | { | |
1ce23dca PS |
1648 | ASSERT3P(zio->io_executor, ==, NULL); |
1649 | ASSERT3U(zio->io_orig_size, ==, zio->io_size); | |
1650 | ASSERT3U(size, <=, zio->io_size); | |
428870ff BB |
1651 | |
1652 | /* | |
1653 | * We don't shrink for raidz because of problems with the | |
1654 | * reconstruction when reading back less than the block size. | |
1655 | * Note, BP_IS_RAIDZ() assumes no compression. | |
1656 | */ | |
1657 | ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); | |
2aa34383 DK |
1658 | if (!BP_IS_RAIDZ(zio->io_bp)) { |
1659 | /* we are not doing a raw write */ | |
1660 | ASSERT3U(zio->io_size, ==, zio->io_lsize); | |
1661 | zio->io_orig_size = zio->io_size = zio->io_lsize = size; | |
1662 | } | |
428870ff BB |
1663 | } |
1664 | ||
d9bb583c AH |
1665 | /* |
1666 | * Round provided allocation size up to a value that can be allocated | |
1667 | * by at least some vdev(s) in the pool with minimum or no additional | |
1668 | * padding and without extra space usage on others | |
1669 | */ | |
1670 | static uint64_t | |
1671 | zio_roundup_alloc_size(spa_t *spa, uint64_t size) | |
1672 | { | |
1673 | if (size > spa->spa_min_alloc) | |
1674 | return (roundup(size, spa->spa_gcd_alloc)); | |
1675 | return (spa->spa_min_alloc); | |
1676 | } | |
1677 | ||
34dc7c2f BB |
1678 | /* |
1679 | * ========================================================================== | |
b128c09f | 1680 | * Prepare to read and write logical blocks |
34dc7c2f BB |
1681 | * ========================================================================== |
1682 | */ | |
b128c09f | 1683 | |
62840030 | 1684 | static zio_t * |
b128c09f | 1685 | zio_read_bp_init(zio_t *zio) |
34dc7c2f | 1686 | { |
b128c09f | 1687 | blkptr_t *bp = zio->io_bp; |
b5256303 TC |
1688 | uint64_t psize = |
1689 | BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); | |
34dc7c2f | 1690 | |
a1d477c2 MA |
1691 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
1692 | ||
fb5f0bc8 | 1693 | if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && |
9babb374 | 1694 | zio->io_child_type == ZIO_CHILD_LOGICAL && |
b5256303 | 1695 | !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { |
a6255b7f DQ |
1696 | zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), |
1697 | psize, psize, zio_decompress); | |
34dc7c2f | 1698 | } |
34dc7c2f | 1699 | |
b5256303 TC |
1700 | if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) || |
1701 | BP_HAS_INDIRECT_MAC_CKSUM(bp)) && | |
1702 | zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
1703 | zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), | |
1704 | psize, psize, zio_decrypt); | |
1705 | } | |
1706 | ||
9b67f605 | 1707 | if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { |
a6255b7f DQ |
1708 | int psize = BPE_GET_PSIZE(bp); |
1709 | void *data = abd_borrow_buf(zio->io_abd, psize); | |
1710 | ||
9b67f605 | 1711 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
a6255b7f DQ |
1712 | decode_embedded_bp_compressed(bp, data); |
1713 | abd_return_buf_copy(zio->io_abd, data, psize); | |
9b67f605 MA |
1714 | } else { |
1715 | ASSERT(!BP_IS_EMBEDDED(bp)); | |
1716 | } | |
1717 | ||
428870ff BB |
1718 | if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) |
1719 | zio->io_pipeline = ZIO_DDT_READ_PIPELINE; | |
1720 | ||
62840030 | 1721 | return (zio); |
34dc7c2f BB |
1722 | } |
1723 | ||
62840030 | 1724 | static zio_t * |
b128c09f | 1725 | zio_write_bp_init(zio_t *zio) |
34dc7c2f | 1726 | { |
b128c09f | 1727 | if (!IO_IS_ALLOCATING(zio)) |
62840030 | 1728 | return (zio); |
34dc7c2f | 1729 | |
428870ff BB |
1730 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1731 | ||
1732 | if (zio->io_bp_override) { | |
3dfb57a3 DB |
1733 | blkptr_t *bp = zio->io_bp; |
1734 | zio_prop_t *zp = &zio->io_prop; | |
1735 | ||
493fcce9 | 1736 | ASSERT(BP_GET_LOGICAL_BIRTH(bp) != zio->io_txg); |
428870ff BB |
1737 | |
1738 | *bp = *zio->io_bp_override; | |
1739 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1740 | ||
67a1b037 PJD |
1741 | if (zp->zp_brtwrite) |
1742 | return (zio); | |
1743 | ||
1744 | ASSERT(!BP_GET_DEDUP(zio->io_bp_override)); | |
1745 | ||
9b67f605 | 1746 | if (BP_IS_EMBEDDED(bp)) |
62840030 | 1747 | return (zio); |
9b67f605 | 1748 | |
03c6040b GW |
1749 | /* |
1750 | * If we've been overridden and nopwrite is set then | |
1751 | * set the flag accordingly to indicate that a nopwrite | |
1752 | * has already occurred. | |
1753 | */ | |
1754 | if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { | |
1755 | ASSERT(!zp->zp_dedup); | |
3dfb57a3 | 1756 | ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum); |
03c6040b | 1757 | zio->io_flags |= ZIO_FLAG_NOPWRITE; |
62840030 | 1758 | return (zio); |
03c6040b GW |
1759 | } |
1760 | ||
1761 | ASSERT(!zp->zp_nopwrite); | |
1762 | ||
428870ff | 1763 | if (BP_IS_HOLE(bp) || !zp->zp_dedup) |
62840030 | 1764 | return (zio); |
428870ff | 1765 | |
3c67d83a TH |
1766 | ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & |
1767 | ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); | |
428870ff | 1768 | |
b5256303 TC |
1769 | if (BP_GET_CHECKSUM(bp) == zp->zp_checksum && |
1770 | !zp->zp_encrypt) { | |
428870ff BB |
1771 | BP_SET_DEDUP(bp, 1); |
1772 | zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; | |
62840030 | 1773 | return (zio); |
428870ff | 1774 | } |
3dfb57a3 DB |
1775 | |
1776 | /* | |
1777 | * We were unable to handle this as an override bp, treat | |
1778 | * it as a regular write I/O. | |
1779 | */ | |
5511754b | 1780 | zio->io_bp_override = NULL; |
3dfb57a3 DB |
1781 | *bp = zio->io_bp_orig; |
1782 | zio->io_pipeline = zio->io_orig_pipeline; | |
1783 | } | |
1784 | ||
62840030 | 1785 | return (zio); |
3dfb57a3 DB |
1786 | } |
1787 | ||
62840030 | 1788 | static zio_t * |
3dfb57a3 DB |
1789 | zio_write_compress(zio_t *zio) |
1790 | { | |
1791 | spa_t *spa = zio->io_spa; | |
1792 | zio_prop_t *zp = &zio->io_prop; | |
1793 | enum zio_compress compress = zp->zp_compress; | |
1794 | blkptr_t *bp = zio->io_bp; | |
1795 | uint64_t lsize = zio->io_lsize; | |
1796 | uint64_t psize = zio->io_size; | |
fdc2d303 | 1797 | uint32_t pass = 1; |
3dfb57a3 | 1798 | |
3dfb57a3 DB |
1799 | /* |
1800 | * If our children haven't all reached the ready stage, | |
1801 | * wait for them and then repeat this pipeline stage. | |
1802 | */ | |
ddc751d5 GW |
1803 | if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT | |
1804 | ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) { | |
62840030 | 1805 | return (NULL); |
ddc751d5 | 1806 | } |
3dfb57a3 DB |
1807 | |
1808 | if (!IO_IS_ALLOCATING(zio)) | |
62840030 | 1809 | return (zio); |
3dfb57a3 DB |
1810 | |
1811 | if (zio->io_children_ready != NULL) { | |
1812 | /* | |
1813 | * Now that all our children are ready, run the callback | |
1814 | * associated with this zio in case it wants to modify the | |
1815 | * data to be written. | |
1816 | */ | |
1817 | ASSERT3U(zp->zp_level, >, 0); | |
1818 | zio->io_children_ready(zio); | |
428870ff | 1819 | } |
34dc7c2f | 1820 | |
3dfb57a3 DB |
1821 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1822 | ASSERT(zio->io_bp_override == NULL); | |
1823 | ||
493fcce9 | 1824 | if (!BP_IS_HOLE(bp) && BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg) { |
b128c09f BB |
1825 | /* |
1826 | * We're rewriting an existing block, which means we're | |
1827 | * working on behalf of spa_sync(). For spa_sync() to | |
1828 | * converge, it must eventually be the case that we don't | |
1829 | * have to allocate new blocks. But compression changes | |
1830 | * the blocksize, which forces a reallocate, and makes | |
1831 | * convergence take longer. Therefore, after the first | |
1832 | * few passes, stop compressing to ensure convergence. | |
1833 | */ | |
428870ff BB |
1834 | pass = spa_sync_pass(spa); |
1835 | ||
1836 | ASSERT(zio->io_txg == spa_syncing_txg(spa)); | |
1837 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1838 | ASSERT(!BP_GET_DEDUP(bp)); | |
34dc7c2f | 1839 | |
55d85d5a | 1840 | if (pass >= zfs_sync_pass_dont_compress) |
b128c09f | 1841 | compress = ZIO_COMPRESS_OFF; |
34dc7c2f | 1842 | |
b128c09f | 1843 | /* Make sure someone doesn't change their mind on overwrites */ |
ed39d668 SD |
1844 | ASSERT(BP_IS_EMBEDDED(bp) || BP_IS_GANG(bp) || |
1845 | MIN(zp->zp_copies, spa_max_replication(spa)) | |
1846 | == BP_GET_NDVAS(bp)); | |
b128c09f | 1847 | } |
34dc7c2f | 1848 | |
2aa34383 | 1849 | /* If it's a compressed write that is not raw, compress the buffer. */ |
b5256303 TC |
1850 | if (compress != ZIO_COMPRESS_OFF && |
1851 | !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { | |
bff26b02 RY |
1852 | void *cbuf = NULL; |
1853 | psize = zio_compress_data(compress, zio->io_abd, &cbuf, lsize, | |
10b3c7f5 | 1854 | zp->zp_complevel); |
bff26b02 | 1855 | if (psize == 0) { |
b128c09f | 1856 | compress = ZIO_COMPRESS_OFF; |
bff26b02 RY |
1857 | } else if (psize >= lsize) { |
1858 | compress = ZIO_COMPRESS_OFF; | |
1859 | if (cbuf != NULL) | |
1860 | zio_buf_free(cbuf, lsize); | |
b5256303 TC |
1861 | } else if (!zp->zp_dedup && !zp->zp_encrypt && |
1862 | psize <= BPE_PAYLOAD_SIZE && | |
9b67f605 MA |
1863 | zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && |
1864 | spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { | |
1865 | encode_embedded_bp_compressed(bp, | |
1866 | cbuf, compress, lsize, psize); | |
1867 | BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); | |
1868 | BP_SET_TYPE(bp, zio->io_prop.zp_type); | |
1869 | BP_SET_LEVEL(bp, zio->io_prop.zp_level); | |
1870 | zio_buf_free(cbuf, lsize); | |
493fcce9 | 1871 | BP_SET_LOGICAL_BIRTH(bp, zio->io_txg); |
9b67f605 MA |
1872 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
1873 | ASSERT(spa_feature_is_active(spa, | |
1874 | SPA_FEATURE_EMBEDDED_DATA)); | |
62840030 | 1875 | return (zio); |
428870ff | 1876 | } else { |
9b67f605 | 1877 | /* |
b2255edc BB |
1878 | * Round compressed size up to the minimum allocation |
1879 | * size of the smallest-ashift device, and zero the | |
1880 | * tail. This ensures that the compressed size of the | |
1881 | * BP (and thus compressratio property) are correct, | |
c3520e7f MA |
1882 | * in that we charge for the padding used to fill out |
1883 | * the last sector. | |
9b67f605 | 1884 | */ |
d9bb583c AH |
1885 | size_t rounded = (size_t)zio_roundup_alloc_size(spa, |
1886 | psize); | |
c3520e7f | 1887 | if (rounded >= lsize) { |
9b67f605 MA |
1888 | compress = ZIO_COMPRESS_OFF; |
1889 | zio_buf_free(cbuf, lsize); | |
c3520e7f | 1890 | psize = lsize; |
9b67f605 | 1891 | } else { |
a6255b7f DQ |
1892 | abd_t *cdata = abd_get_from_buf(cbuf, lsize); |
1893 | abd_take_ownership_of_buf(cdata, B_TRUE); | |
1894 | abd_zero_off(cdata, psize, rounded - psize); | |
c3520e7f | 1895 | psize = rounded; |
a6255b7f | 1896 | zio_push_transform(zio, cdata, |
9b67f605 MA |
1897 | psize, lsize, NULL); |
1898 | } | |
b128c09f | 1899 | } |
3dfb57a3 DB |
1900 | |
1901 | /* | |
1902 | * We were unable to handle this as an override bp, treat | |
1903 | * it as a regular write I/O. | |
1904 | */ | |
1905 | zio->io_bp_override = NULL; | |
1906 | *bp = zio->io_bp_orig; | |
1907 | zio->io_pipeline = zio->io_orig_pipeline; | |
1908 | ||
b1d21733 TC |
1909 | } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 && |
1910 | zp->zp_type == DMU_OT_DNODE) { | |
1911 | /* | |
1912 | * The DMU actually relies on the zio layer's compression | |
1913 | * to free metadnode blocks that have had all contained | |
1914 | * dnodes freed. As a result, even when doing a raw | |
1915 | * receive, we must check whether the block can be compressed | |
1916 | * to a hole. | |
1917 | */ | |
1918 | psize = zio_compress_data(ZIO_COMPRESS_EMPTY, | |
10b3c7f5 MN |
1919 | zio->io_abd, NULL, lsize, zp->zp_complevel); |
1920 | if (psize == 0 || psize >= lsize) | |
b1d21733 | 1921 | compress = ZIO_COMPRESS_OFF; |
52a36bd4 GA |
1922 | } else if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS && |
1923 | !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) { | |
1924 | /* | |
1925 | * If we are raw receiving an encrypted dataset we should not | |
1926 | * take this codepath because it will change the on-disk block | |
1927 | * and decryption will fail. | |
1928 | */ | |
d9bb583c AH |
1929 | size_t rounded = MIN((size_t)zio_roundup_alloc_size(spa, psize), |
1930 | lsize); | |
c634320e PD |
1931 | |
1932 | if (rounded != psize) { | |
1933 | abd_t *cdata = abd_alloc_linear(rounded, B_TRUE); | |
1934 | abd_zero_off(cdata, psize, rounded - psize); | |
1935 | abd_copy_off(cdata, zio->io_abd, 0, 0, psize); | |
1936 | psize = rounded; | |
1937 | zio_push_transform(zio, cdata, | |
1938 | psize, rounded, NULL); | |
1939 | } | |
2aa34383 DK |
1940 | } else { |
1941 | ASSERT3U(psize, !=, 0); | |
b128c09f | 1942 | } |
34dc7c2f | 1943 | |
b128c09f BB |
1944 | /* |
1945 | * The final pass of spa_sync() must be all rewrites, but the first | |
1946 | * few passes offer a trade-off: allocating blocks defers convergence, | |
1947 | * but newly allocated blocks are sequential, so they can be written | |
1948 | * to disk faster. Therefore, we allow the first few passes of | |
1949 | * spa_sync() to allocate new blocks, but force rewrites after that. | |
1950 | * There should only be a handful of blocks after pass 1 in any case. | |
1951 | */ | |
493fcce9 | 1952 | if (!BP_IS_HOLE(bp) && BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg && |
b0bc7a84 | 1953 | BP_GET_PSIZE(bp) == psize && |
55d85d5a | 1954 | pass >= zfs_sync_pass_rewrite) { |
cc99f275 | 1955 | VERIFY3U(psize, !=, 0); |
1c27024e | 1956 | enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; |
cc99f275 | 1957 | |
b128c09f BB |
1958 | zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; |
1959 | zio->io_flags |= ZIO_FLAG_IO_REWRITE; | |
1960 | } else { | |
1961 | BP_ZERO(bp); | |
1962 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
1963 | } | |
34dc7c2f | 1964 | |
428870ff | 1965 | if (psize == 0) { |
493fcce9 | 1966 | if (BP_GET_LOGICAL_BIRTH(&zio->io_bp_orig) != 0 && |
b0bc7a84 MG |
1967 | spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { |
1968 | BP_SET_LSIZE(bp, lsize); | |
1969 | BP_SET_TYPE(bp, zp->zp_type); | |
1970 | BP_SET_LEVEL(bp, zp->zp_level); | |
1971 | BP_SET_BIRTH(bp, zio->io_txg, 0); | |
1972 | } | |
b128c09f BB |
1973 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
1974 | } else { | |
1975 | ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); | |
1976 | BP_SET_LSIZE(bp, lsize); | |
b0bc7a84 MG |
1977 | BP_SET_TYPE(bp, zp->zp_type); |
1978 | BP_SET_LEVEL(bp, zp->zp_level); | |
428870ff | 1979 | BP_SET_PSIZE(bp, psize); |
b128c09f BB |
1980 | BP_SET_COMPRESS(bp, compress); |
1981 | BP_SET_CHECKSUM(bp, zp->zp_checksum); | |
428870ff | 1982 | BP_SET_DEDUP(bp, zp->zp_dedup); |
b128c09f | 1983 | BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); |
428870ff BB |
1984 | if (zp->zp_dedup) { |
1985 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1986 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
b5256303 TC |
1987 | ASSERT(!zp->zp_encrypt || |
1988 | DMU_OT_IS_ENCRYPTED(zp->zp_type)); | |
428870ff BB |
1989 | zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; |
1990 | } | |
03c6040b GW |
1991 | if (zp->zp_nopwrite) { |
1992 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1993 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
1994 | zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; | |
1995 | } | |
428870ff | 1996 | } |
62840030 | 1997 | return (zio); |
428870ff BB |
1998 | } |
1999 | ||
62840030 | 2000 | static zio_t * |
428870ff BB |
2001 | zio_free_bp_init(zio_t *zio) |
2002 | { | |
2003 | blkptr_t *bp = zio->io_bp; | |
2004 | ||
2005 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
2006 | if (BP_GET_DEDUP(bp)) | |
2007 | zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; | |
b128c09f | 2008 | } |
34dc7c2f | 2009 | |
a1d477c2 MA |
2010 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
2011 | ||
62840030 | 2012 | return (zio); |
34dc7c2f BB |
2013 | } |
2014 | ||
b128c09f BB |
2015 | /* |
2016 | * ========================================================================== | |
2017 | * Execute the I/O pipeline | |
2018 | * ========================================================================== | |
2019 | */ | |
2020 | ||
2021 | static void | |
7ef5e54e | 2022 | zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) |
34dc7c2f | 2023 | { |
428870ff | 2024 | spa_t *spa = zio->io_spa; |
b128c09f | 2025 | zio_type_t t = zio->io_type; |
a38718a6 | 2026 | int flags = (cutinline ? TQ_FRONT : 0); |
34dc7c2f BB |
2027 | |
2028 | /* | |
9babb374 BB |
2029 | * If we're a config writer or a probe, the normal issue and |
2030 | * interrupt threads may all be blocked waiting for the config lock. | |
2031 | * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. | |
34dc7c2f | 2032 | */ |
9babb374 | 2033 | if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) |
b128c09f | 2034 | t = ZIO_TYPE_NULL; |
34dc7c2f BB |
2035 | |
2036 | /* | |
b128c09f | 2037 | * A similar issue exists for the L2ARC write thread until L2ARC 2.0. |
34dc7c2f | 2038 | */ |
b128c09f BB |
2039 | if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) |
2040 | t = ZIO_TYPE_NULL; | |
34dc7c2f | 2041 | |
428870ff | 2042 | /* |
7ef5e54e | 2043 | * If this is a high priority I/O, then use the high priority taskq if |
04bae5ec | 2044 | * available or cut the line otherwise. |
428870ff | 2045 | */ |
04bae5ec AM |
2046 | if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) { |
2047 | if (spa->spa_zio_taskq[t][q + 1].stqs_count != 0) | |
2048 | q++; | |
2049 | else | |
2050 | flags |= TQ_FRONT; | |
2051 | } | |
428870ff BB |
2052 | |
2053 | ASSERT3U(q, <, ZIO_TASKQ_TYPES); | |
5cc556b4 | 2054 | |
a38718a6 GA |
2055 | /* |
2056 | * NB: We are assuming that the zio can only be dispatched | |
2057 | * to a single taskq at a time. It would be a grievous error | |
2058 | * to dispatch the zio to another taskq at the same time. | |
2059 | */ | |
2060 | ASSERT(taskq_empty_ent(&zio->io_tqent)); | |
23c13c7e | 2061 | spa_taskq_dispatch_ent(spa, t, q, zio_execute, zio, flags, |
3bd4df38 | 2062 | &zio->io_tqent, zio); |
b128c09f | 2063 | } |
34dc7c2f | 2064 | |
b128c09f | 2065 | static boolean_t |
7ef5e54e | 2066 | zio_taskq_member(zio_t *zio, zio_taskq_type_t q) |
b128c09f | 2067 | { |
b128c09f | 2068 | spa_t *spa = zio->io_spa; |
34dc7c2f | 2069 | |
b3212d2f MA |
2070 | taskq_t *tq = taskq_of_curthread(); |
2071 | ||
1c27024e | 2072 | for (zio_type_t t = 0; t < ZIO_TYPES; t++) { |
7ef5e54e AL |
2073 | spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; |
2074 | uint_t i; | |
2075 | for (i = 0; i < tqs->stqs_count; i++) { | |
b3212d2f | 2076 | if (tqs->stqs_taskq[i] == tq) |
7ef5e54e AL |
2077 | return (B_TRUE); |
2078 | } | |
2079 | } | |
34dc7c2f | 2080 | |
b128c09f BB |
2081 | return (B_FALSE); |
2082 | } | |
34dc7c2f | 2083 | |
62840030 | 2084 | static zio_t * |
b128c09f BB |
2085 | zio_issue_async(zio_t *zio) |
2086 | { | |
3bd4df38 | 2087 | ASSERT((zio->io_type != ZIO_TYPE_WRITE) || ZIO_HAS_ALLOCATOR(zio)); |
428870ff | 2088 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); |
62840030 | 2089 | return (NULL); |
34dc7c2f BB |
2090 | } |
2091 | ||
b128c09f | 2092 | void |
23c13c7e | 2093 | zio_interrupt(void *zio) |
34dc7c2f | 2094 | { |
428870ff | 2095 | zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); |
b128c09f | 2096 | } |
34dc7c2f | 2097 | |
d3c1e45b MM |
2098 | void |
2099 | zio_delay_interrupt(zio_t *zio) | |
2100 | { | |
2101 | /* | |
2102 | * The timeout_generic() function isn't defined in userspace, so | |
2103 | * rather than trying to implement the function, the zio delay | |
2104 | * functionality has been disabled for userspace builds. | |
2105 | */ | |
2106 | ||
2107 | #ifdef _KERNEL | |
2108 | /* | |
2109 | * If io_target_timestamp is zero, then no delay has been registered | |
2110 | * for this IO, thus jump to the end of this function and "skip" the | |
2111 | * delay; issuing it directly to the zio layer. | |
2112 | */ | |
2113 | if (zio->io_target_timestamp != 0) { | |
2114 | hrtime_t now = gethrtime(); | |
2115 | ||
2116 | if (now >= zio->io_target_timestamp) { | |
2117 | /* | |
2118 | * This IO has already taken longer than the target | |
2119 | * delay to complete, so we don't want to delay it | |
2120 | * any longer; we "miss" the delay and issue it | |
2121 | * directly to the zio layer. This is likely due to | |
2122 | * the target latency being set to a value less than | |
2123 | * the underlying hardware can satisfy (e.g. delay | |
2124 | * set to 1ms, but the disks take 10ms to complete an | |
2125 | * IO request). | |
2126 | */ | |
2127 | ||
2128 | DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, | |
2129 | hrtime_t, now); | |
2130 | ||
2131 | zio_interrupt(zio); | |
2132 | } else { | |
2133 | taskqid_t tid; | |
2134 | hrtime_t diff = zio->io_target_timestamp - now; | |
2135 | clock_t expire_at_tick = ddi_get_lbolt() + | |
2136 | NSEC_TO_TICK(diff); | |
2137 | ||
2138 | DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, | |
2139 | hrtime_t, now, hrtime_t, diff); | |
2140 | ||
2141 | if (NSEC_TO_TICK(diff) == 0) { | |
2142 | /* Our delay is less than a jiffy - just spin */ | |
2143 | zfs_sleep_until(zio->io_target_timestamp); | |
2144 | zio_interrupt(zio); | |
2145 | } else { | |
2146 | /* | |
2147 | * Use taskq_dispatch_delay() in the place of | |
2148 | * OpenZFS's timeout_generic(). | |
2149 | */ | |
2150 | tid = taskq_dispatch_delay(system_taskq, | |
23c13c7e AL |
2151 | zio_interrupt, zio, TQ_NOSLEEP, |
2152 | expire_at_tick); | |
d3c1e45b MM |
2153 | if (tid == TASKQID_INVALID) { |
2154 | /* | |
2155 | * Couldn't allocate a task. Just | |
2156 | * finish the zio without a delay. | |
2157 | */ | |
2158 | zio_interrupt(zio); | |
2159 | } | |
2160 | } | |
2161 | } | |
2162 | return; | |
2163 | } | |
2164 | #endif | |
2165 | DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); | |
2166 | zio_interrupt(zio); | |
2167 | } | |
2168 | ||
8fb1ede1 | 2169 | static void |
638dd5f4 | 2170 | zio_deadman_impl(zio_t *pio, int ziodepth) |
8fb1ede1 BB |
2171 | { |
2172 | zio_t *cio, *cio_next; | |
2173 | zio_link_t *zl = NULL; | |
2174 | vdev_t *vd = pio->io_vd; | |
2175 | ||
638dd5f4 TC |
2176 | if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) { |
2177 | vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL; | |
8fb1ede1 BB |
2178 | zbookmark_phys_t *zb = &pio->io_bookmark; |
2179 | uint64_t delta = gethrtime() - pio->io_timestamp; | |
2180 | uint64_t failmode = spa_get_deadman_failmode(pio->io_spa); | |
2181 | ||
a887d653 | 2182 | zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu " |
8fb1ede1 | 2183 | "delta=%llu queued=%llu io=%llu " |
8e739b2c RE |
2184 | "path=%s " |
2185 | "last=%llu type=%d " | |
4938d01d | 2186 | "priority=%d flags=0x%llx stage=0x%x " |
8e739b2c RE |
2187 | "pipeline=0x%x pipeline-trace=0x%x " |
2188 | "objset=%llu object=%llu " | |
2189 | "level=%llu blkid=%llu " | |
2190 | "offset=%llu size=%llu " | |
2191 | "error=%d", | |
638dd5f4 | 2192 | ziodepth, pio, pio->io_timestamp, |
8e739b2c RE |
2193 | (u_longlong_t)delta, pio->io_delta, pio->io_delay, |
2194 | vd ? vd->vdev_path : "NULL", | |
2195 | vq ? vq->vq_io_complete_ts : 0, pio->io_type, | |
4938d01d RY |
2196 | pio->io_priority, (u_longlong_t)pio->io_flags, |
2197 | pio->io_stage, pio->io_pipeline, pio->io_pipeline_trace, | |
8e739b2c RE |
2198 | (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object, |
2199 | (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid, | |
2200 | (u_longlong_t)pio->io_offset, (u_longlong_t)pio->io_size, | |
2201 | pio->io_error); | |
1144586b | 2202 | (void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN, |
4f072827 | 2203 | pio->io_spa, vd, zb, pio, 0); |
8fb1ede1 BB |
2204 | |
2205 | if (failmode == ZIO_FAILURE_MODE_CONTINUE && | |
2206 | taskq_empty_ent(&pio->io_tqent)) { | |
2207 | zio_interrupt(pio); | |
2208 | } | |
2209 | } | |
2210 | ||
2211 | mutex_enter(&pio->io_lock); | |
2212 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { | |
2213 | cio_next = zio_walk_children(pio, &zl); | |
638dd5f4 | 2214 | zio_deadman_impl(cio, ziodepth + 1); |
8fb1ede1 BB |
2215 | } |
2216 | mutex_exit(&pio->io_lock); | |
2217 | } | |
2218 | ||
2219 | /* | |
2220 | * Log the critical information describing this zio and all of its children | |
2221 | * using the zfs_dbgmsg() interface then post deadman event for the ZED. | |
2222 | */ | |
2223 | void | |
dd66857d | 2224 | zio_deadman(zio_t *pio, const char *tag) |
8fb1ede1 BB |
2225 | { |
2226 | spa_t *spa = pio->io_spa; | |
2227 | char *name = spa_name(spa); | |
2228 | ||
2229 | if (!zfs_deadman_enabled || spa_suspended(spa)) | |
2230 | return; | |
2231 | ||
638dd5f4 | 2232 | zio_deadman_impl(pio, 0); |
8fb1ede1 BB |
2233 | |
2234 | switch (spa_get_deadman_failmode(spa)) { | |
2235 | case ZIO_FAILURE_MODE_WAIT: | |
2236 | zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name); | |
2237 | break; | |
2238 | ||
2239 | case ZIO_FAILURE_MODE_CONTINUE: | |
2240 | zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name); | |
2241 | break; | |
2242 | ||
2243 | case ZIO_FAILURE_MODE_PANIC: | |
2244 | fm_panic("%s determined I/O to pool '%s' is hung.", tag, name); | |
2245 | break; | |
2246 | } | |
2247 | } | |
2248 | ||
b128c09f BB |
2249 | /* |
2250 | * Execute the I/O pipeline until one of the following occurs: | |
2251 | * (1) the I/O completes; (2) the pipeline stalls waiting for | |
2252 | * dependent child I/Os; (3) the I/O issues, so we're waiting | |
2253 | * for an I/O completion interrupt; (4) the I/O is delegated by | |
2254 | * vdev-level caching or aggregation; (5) the I/O is deferred | |
2255 | * due to vdev-level queueing; (6) the I/O is handed off to | |
2256 | * another thread. In all cases, the pipeline stops whenever | |
8e07b99b | 2257 | * there's no CPU work; it never burns a thread in cv_wait_io(). |
b128c09f BB |
2258 | * |
2259 | * There's no locking on io_stage because there's no legitimate way | |
2260 | * for multiple threads to be attempting to process the same I/O. | |
2261 | */ | |
428870ff | 2262 | static zio_pipe_stage_t *zio_pipeline[]; |
34dc7c2f | 2263 | |
da6b4005 NB |
2264 | /* |
2265 | * zio_execute() is a wrapper around the static function | |
2266 | * __zio_execute() so that we can force __zio_execute() to be | |
2267 | * inlined. This reduces stack overhead which is important | |
2268 | * because __zio_execute() is called recursively in several zio | |
2269 | * code paths. zio_execute() itself cannot be inlined because | |
2270 | * it is externally visible. | |
2271 | */ | |
b128c09f | 2272 | void |
23c13c7e | 2273 | zio_execute(void *zio) |
da6b4005 | 2274 | { |
92119cc2 BB |
2275 | fstrans_cookie_t cookie; |
2276 | ||
2277 | cookie = spl_fstrans_mark(); | |
da6b4005 | 2278 | __zio_execute(zio); |
92119cc2 | 2279 | spl_fstrans_unmark(cookie); |
da6b4005 NB |
2280 | } |
2281 | ||
b58986ee BB |
2282 | /* |
2283 | * Used to determine if in the current context the stack is sized large | |
2284 | * enough to allow zio_execute() to be called recursively. A minimum | |
2285 | * stack size of 16K is required to avoid needing to re-dispatch the zio. | |
2286 | */ | |
65c7cc49 | 2287 | static boolean_t |
b58986ee BB |
2288 | zio_execute_stack_check(zio_t *zio) |
2289 | { | |
2290 | #if !defined(HAVE_LARGE_STACKS) | |
2291 | dsl_pool_t *dp = spa_get_dsl(zio->io_spa); | |
2292 | ||
2293 | /* Executing in txg_sync_thread() context. */ | |
2294 | if (dp && curthread == dp->dp_tx.tx_sync_thread) | |
2295 | return (B_TRUE); | |
2296 | ||
2297 | /* Pool initialization outside of zio_taskq context. */ | |
2298 | if (dp && spa_is_initializing(dp->dp_spa) && | |
2299 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) && | |
2300 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH)) | |
2301 | return (B_TRUE); | |
14e4e3cb AZ |
2302 | #else |
2303 | (void) zio; | |
b58986ee BB |
2304 | #endif /* HAVE_LARGE_STACKS */ |
2305 | ||
2306 | return (B_FALSE); | |
2307 | } | |
2308 | ||
da6b4005 NB |
2309 | __attribute__((always_inline)) |
2310 | static inline void | |
2311 | __zio_execute(zio_t *zio) | |
b128c09f | 2312 | { |
3dfb57a3 DB |
2313 | ASSERT3U(zio->io_queued_timestamp, >, 0); |
2314 | ||
b128c09f | 2315 | while (zio->io_stage < ZIO_STAGE_DONE) { |
428870ff BB |
2316 | enum zio_stage pipeline = zio->io_pipeline; |
2317 | enum zio_stage stage = zio->io_stage; | |
62840030 MA |
2318 | |
2319 | zio->io_executor = curthread; | |
34dc7c2f | 2320 | |
b128c09f | 2321 | ASSERT(!MUTEX_HELD(&zio->io_lock)); |
428870ff BB |
2322 | ASSERT(ISP2(stage)); |
2323 | ASSERT(zio->io_stall == NULL); | |
34dc7c2f | 2324 | |
428870ff BB |
2325 | do { |
2326 | stage <<= 1; | |
2327 | } while ((stage & pipeline) == 0); | |
b128c09f BB |
2328 | |
2329 | ASSERT(stage <= ZIO_STAGE_DONE); | |
34dc7c2f BB |
2330 | |
2331 | /* | |
b128c09f BB |
2332 | * If we are in interrupt context and this pipeline stage |
2333 | * will grab a config lock that is held across I/O, | |
428870ff BB |
2334 | * or may wait for an I/O that needs an interrupt thread |
2335 | * to complete, issue async to avoid deadlock. | |
2336 | * | |
2337 | * For VDEV_IO_START, we cut in line so that the io will | |
2338 | * be sent to disk promptly. | |
34dc7c2f | 2339 | */ |
91579709 BB |
2340 | if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && |
2341 | zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { | |
b58986ee BB |
2342 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? |
2343 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
91579709 BB |
2344 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
2345 | return; | |
2346 | } | |
2347 | ||
2348 | /* | |
b58986ee BB |
2349 | * If the current context doesn't have large enough stacks |
2350 | * the zio must be issued asynchronously to prevent overflow. | |
91579709 | 2351 | */ |
b58986ee BB |
2352 | if (zio_execute_stack_check(zio)) { |
2353 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? | |
2354 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
428870ff | 2355 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
b128c09f | 2356 | return; |
34dc7c2f BB |
2357 | } |
2358 | ||
b128c09f | 2359 | zio->io_stage = stage; |
3dfb57a3 | 2360 | zio->io_pipeline_trace |= zio->io_stage; |
34dc7c2f | 2361 | |
62840030 MA |
2362 | /* |
2363 | * The zio pipeline stage returns the next zio to execute | |
2364 | * (typically the same as this one), or NULL if we should | |
2365 | * stop. | |
2366 | */ | |
2367 | zio = zio_pipeline[highbit64(stage) - 1](zio); | |
34dc7c2f | 2368 | |
62840030 MA |
2369 | if (zio == NULL) |
2370 | return; | |
b128c09f | 2371 | } |
34dc7c2f BB |
2372 | } |
2373 | ||
da6b4005 | 2374 | |
b128c09f BB |
2375 | /* |
2376 | * ========================================================================== | |
2377 | * Initiate I/O, either sync or async | |
2378 | * ========================================================================== | |
2379 | */ | |
2380 | int | |
2381 | zio_wait(zio_t *zio) | |
34dc7c2f | 2382 | { |
9cdf7b1f MA |
2383 | /* |
2384 | * Some routines, like zio_free_sync(), may return a NULL zio | |
2385 | * to avoid the performance overhead of creating and then destroying | |
2386 | * an unneeded zio. For the callers' simplicity, we accept a NULL | |
2387 | * zio and ignore it. | |
2388 | */ | |
2389 | if (zio == NULL) | |
2390 | return (0); | |
2391 | ||
8fb1ede1 | 2392 | long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms); |
b128c09f | 2393 | int error; |
34dc7c2f | 2394 | |
1ce23dca PS |
2395 | ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN); |
2396 | ASSERT3P(zio->io_executor, ==, NULL); | |
34dc7c2f | 2397 | |
b128c09f | 2398 | zio->io_waiter = curthread; |
3dfb57a3 DB |
2399 | ASSERT0(zio->io_queued_timestamp); |
2400 | zio->io_queued_timestamp = gethrtime(); | |
34dc7c2f | 2401 | |
3bd4df38 EN |
2402 | if (zio->io_type == ZIO_TYPE_WRITE) { |
2403 | spa_select_allocator(zio); | |
2404 | } | |
da6b4005 | 2405 | __zio_execute(zio); |
34dc7c2f | 2406 | |
b128c09f | 2407 | mutex_enter(&zio->io_lock); |
8fb1ede1 BB |
2408 | while (zio->io_executor != NULL) { |
2409 | error = cv_timedwait_io(&zio->io_cv, &zio->io_lock, | |
2410 | ddi_get_lbolt() + timeout); | |
2411 | ||
2412 | if (zfs_deadman_enabled && error == -1 && | |
2413 | gethrtime() - zio->io_queued_timestamp > | |
2414 | spa_deadman_ziotime(zio->io_spa)) { | |
2415 | mutex_exit(&zio->io_lock); | |
2416 | timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms); | |
2417 | zio_deadman(zio, FTAG); | |
2418 | mutex_enter(&zio->io_lock); | |
2419 | } | |
2420 | } | |
b128c09f | 2421 | mutex_exit(&zio->io_lock); |
34dc7c2f | 2422 | |
b128c09f BB |
2423 | error = zio->io_error; |
2424 | zio_destroy(zio); | |
34dc7c2f | 2425 | |
b128c09f BB |
2426 | return (error); |
2427 | } | |
34dc7c2f | 2428 | |
b128c09f BB |
2429 | void |
2430 | zio_nowait(zio_t *zio) | |
2431 | { | |
9cdf7b1f MA |
2432 | /* |
2433 | * See comment in zio_wait(). | |
2434 | */ | |
2435 | if (zio == NULL) | |
2436 | return; | |
2437 | ||
1ce23dca | 2438 | ASSERT3P(zio->io_executor, ==, NULL); |
34dc7c2f | 2439 | |
d164b209 | 2440 | if (zio->io_child_type == ZIO_CHILD_LOGICAL && |
b035f2b2 | 2441 | list_is_empty(&zio->io_parent_list)) { |
8878261f BB |
2442 | zio_t *pio; |
2443 | ||
34dc7c2f | 2444 | /* |
b128c09f | 2445 | * This is a logical async I/O with no parent to wait for it. |
9babb374 BB |
2446 | * We add it to the spa_async_root_zio "Godfather" I/O which |
2447 | * will ensure they complete prior to unloading the pool. | |
34dc7c2f | 2448 | */ |
b128c09f | 2449 | spa_t *spa = zio->io_spa; |
09eb36ce | 2450 | pio = spa->spa_async_zio_root[CPU_SEQID_UNSTABLE]; |
9babb374 | 2451 | |
8878261f | 2452 | zio_add_child(pio, zio); |
b128c09f | 2453 | } |
34dc7c2f | 2454 | |
3dfb57a3 DB |
2455 | ASSERT0(zio->io_queued_timestamp); |
2456 | zio->io_queued_timestamp = gethrtime(); | |
3bd4df38 EN |
2457 | if (zio->io_type == ZIO_TYPE_WRITE) { |
2458 | spa_select_allocator(zio); | |
2459 | } | |
da6b4005 | 2460 | __zio_execute(zio); |
b128c09f | 2461 | } |
34dc7c2f | 2462 | |
b128c09f BB |
2463 | /* |
2464 | * ========================================================================== | |
1ce23dca | 2465 | * Reexecute, cancel, or suspend/resume failed I/O |
b128c09f BB |
2466 | * ========================================================================== |
2467 | */ | |
34dc7c2f | 2468 | |
b128c09f | 2469 | static void |
23c13c7e | 2470 | zio_reexecute(void *arg) |
b128c09f | 2471 | { |
23c13c7e | 2472 | zio_t *pio = arg; |
3afdc97d | 2473 | zio_t *cio, *cio_next, *gio; |
d164b209 BB |
2474 | |
2475 | ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); | |
2476 | ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); | |
9babb374 BB |
2477 | ASSERT(pio->io_gang_leader == NULL); |
2478 | ASSERT(pio->io_gang_tree == NULL); | |
34dc7c2f | 2479 | |
3afdc97d | 2480 | mutex_enter(&pio->io_lock); |
b128c09f BB |
2481 | pio->io_flags = pio->io_orig_flags; |
2482 | pio->io_stage = pio->io_orig_stage; | |
2483 | pio->io_pipeline = pio->io_orig_pipeline; | |
2484 | pio->io_reexecute = 0; | |
03c6040b | 2485 | pio->io_flags |= ZIO_FLAG_REEXECUTED; |
3dfb57a3 | 2486 | pio->io_pipeline_trace = 0; |
b128c09f | 2487 | pio->io_error = 0; |
3afdc97d AM |
2488 | pio->io_state[ZIO_WAIT_READY] = (pio->io_stage >= ZIO_STAGE_READY) || |
2489 | (pio->io_pipeline & ZIO_STAGE_READY) == 0; | |
2490 | pio->io_state[ZIO_WAIT_DONE] = (pio->io_stage >= ZIO_STAGE_DONE); | |
2491 | zio_link_t *zl = NULL; | |
2492 | while ((gio = zio_walk_parents(pio, &zl)) != NULL) { | |
2493 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) { | |
2494 | gio->io_children[pio->io_child_type][w] += | |
2495 | !pio->io_state[w]; | |
2496 | } | |
2497 | } | |
1c27024e | 2498 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) |
b128c09f | 2499 | pio->io_child_error[c] = 0; |
34dc7c2f | 2500 | |
428870ff BB |
2501 | if (IO_IS_ALLOCATING(pio)) |
2502 | BP_ZERO(pio->io_bp); | |
34dc7c2f | 2503 | |
b128c09f BB |
2504 | /* |
2505 | * As we reexecute pio's children, new children could be created. | |
d164b209 | 2506 | * New children go to the head of pio's io_child_list, however, |
b128c09f | 2507 | * so we will (correctly) not reexecute them. The key is that |
d164b209 BB |
2508 | * the remainder of pio's io_child_list, from 'cio_next' onward, |
2509 | * cannot be affected by any side effects of reexecuting 'cio'. | |
b128c09f | 2510 | */ |
3afdc97d | 2511 | zl = NULL; |
3dfb57a3 DB |
2512 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { |
2513 | cio_next = zio_walk_children(pio, &zl); | |
b128c09f | 2514 | mutex_exit(&pio->io_lock); |
d164b209 | 2515 | zio_reexecute(cio); |
a8b2e306 | 2516 | mutex_enter(&pio->io_lock); |
34dc7c2f | 2517 | } |
a8b2e306 | 2518 | mutex_exit(&pio->io_lock); |
34dc7c2f | 2519 | |
b128c09f BB |
2520 | /* |
2521 | * Now that all children have been reexecuted, execute the parent. | |
9babb374 | 2522 | * We don't reexecute "The Godfather" I/O here as it's the |
9e2c3bb4 | 2523 | * responsibility of the caller to wait on it. |
b128c09f | 2524 | */ |
3dfb57a3 DB |
2525 | if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) { |
2526 | pio->io_queued_timestamp = gethrtime(); | |
da6b4005 | 2527 | __zio_execute(pio); |
3dfb57a3 | 2528 | } |
34dc7c2f BB |
2529 | } |
2530 | ||
b128c09f | 2531 | void |
cec3a0a1 | 2532 | zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason) |
34dc7c2f | 2533 | { |
b128c09f BB |
2534 | if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) |
2535 | fm_panic("Pool '%s' has encountered an uncorrectable I/O " | |
2536 | "failure and the failure mode property for this pool " | |
2537 | "is set to panic.", spa_name(spa)); | |
34dc7c2f | 2538 | |
c3f2f1aa DB |
2539 | if (reason != ZIO_SUSPEND_MMP) { |
2540 | cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable " | |
2541 | "I/O failure and has been suspended.\n", spa_name(spa)); | |
2542 | } | |
bf89c199 | 2543 | |
1144586b | 2544 | (void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, |
4f072827 | 2545 | NULL, NULL, 0); |
34dc7c2f | 2546 | |
b128c09f | 2547 | mutex_enter(&spa->spa_suspend_lock); |
34dc7c2f | 2548 | |
b128c09f | 2549 | if (spa->spa_suspend_zio_root == NULL) |
9babb374 BB |
2550 | spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, |
2551 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
2552 | ZIO_FLAG_GODFATHER); | |
34dc7c2f | 2553 | |
cec3a0a1 | 2554 | spa->spa_suspended = reason; |
34dc7c2f | 2555 | |
b128c09f | 2556 | if (zio != NULL) { |
9babb374 | 2557 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f BB |
2558 | ASSERT(zio != spa->spa_suspend_zio_root); |
2559 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
d164b209 | 2560 | ASSERT(zio_unique_parent(zio) == NULL); |
b128c09f BB |
2561 | ASSERT(zio->io_stage == ZIO_STAGE_DONE); |
2562 | zio_add_child(spa->spa_suspend_zio_root, zio); | |
2563 | } | |
34dc7c2f | 2564 | |
b128c09f BB |
2565 | mutex_exit(&spa->spa_suspend_lock); |
2566 | } | |
34dc7c2f | 2567 | |
9babb374 | 2568 | int |
b128c09f BB |
2569 | zio_resume(spa_t *spa) |
2570 | { | |
9babb374 | 2571 | zio_t *pio; |
34dc7c2f BB |
2572 | |
2573 | /* | |
b128c09f | 2574 | * Reexecute all previously suspended i/o. |
34dc7c2f | 2575 | */ |
b128c09f | 2576 | mutex_enter(&spa->spa_suspend_lock); |
cec3a0a1 | 2577 | spa->spa_suspended = ZIO_SUSPEND_NONE; |
b128c09f BB |
2578 | cv_broadcast(&spa->spa_suspend_cv); |
2579 | pio = spa->spa_suspend_zio_root; | |
2580 | spa->spa_suspend_zio_root = NULL; | |
2581 | mutex_exit(&spa->spa_suspend_lock); | |
2582 | ||
2583 | if (pio == NULL) | |
9babb374 | 2584 | return (0); |
34dc7c2f | 2585 | |
9babb374 BB |
2586 | zio_reexecute(pio); |
2587 | return (zio_wait(pio)); | |
b128c09f BB |
2588 | } |
2589 | ||
2590 | void | |
2591 | zio_resume_wait(spa_t *spa) | |
2592 | { | |
2593 | mutex_enter(&spa->spa_suspend_lock); | |
2594 | while (spa_suspended(spa)) | |
2595 | cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); | |
2596 | mutex_exit(&spa->spa_suspend_lock); | |
34dc7c2f BB |
2597 | } |
2598 | ||
2599 | /* | |
2600 | * ========================================================================== | |
b128c09f BB |
2601 | * Gang blocks. |
2602 | * | |
2603 | * A gang block is a collection of small blocks that looks to the DMU | |
2604 | * like one large block. When zio_dva_allocate() cannot find a block | |
2605 | * of the requested size, due to either severe fragmentation or the pool | |
2606 | * being nearly full, it calls zio_write_gang_block() to construct the | |
2607 | * block from smaller fragments. | |
2608 | * | |
2609 | * A gang block consists of a gang header (zio_gbh_phys_t) and up to | |
2610 | * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like | |
2611 | * an indirect block: it's an array of block pointers. It consumes | |
2612 | * only one sector and hence is allocatable regardless of fragmentation. | |
2613 | * The gang header's bps point to its gang members, which hold the data. | |
2614 | * | |
2615 | * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> | |
2616 | * as the verifier to ensure uniqueness of the SHA256 checksum. | |
2617 | * Critically, the gang block bp's blk_cksum is the checksum of the data, | |
2618 | * not the gang header. This ensures that data block signatures (needed for | |
2619 | * deduplication) are independent of how the block is physically stored. | |
2620 | * | |
2621 | * Gang blocks can be nested: a gang member may itself be a gang block. | |
2622 | * Thus every gang block is a tree in which root and all interior nodes are | |
2623 | * gang headers, and the leaves are normal blocks that contain user data. | |
2624 | * The root of the gang tree is called the gang leader. | |
2625 | * | |
2626 | * To perform any operation (read, rewrite, free, claim) on a gang block, | |
2627 | * zio_gang_assemble() first assembles the gang tree (minus data leaves) | |
2628 | * in the io_gang_tree field of the original logical i/o by recursively | |
2629 | * reading the gang leader and all gang headers below it. This yields | |
2630 | * an in-core tree containing the contents of every gang header and the | |
2631 | * bps for every constituent of the gang block. | |
2632 | * | |
2633 | * With the gang tree now assembled, zio_gang_issue() just walks the gang tree | |
2634 | * and invokes a callback on each bp. To free a gang block, zio_gang_issue() | |
2635 | * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. | |
2636 | * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). | |
2637 | * zio_read_gang() is a wrapper around zio_read() that omits reading gang | |
2638 | * headers, since we already have those in io_gang_tree. zio_rewrite_gang() | |
2639 | * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() | |
2640 | * of the gang header plus zio_checksum_compute() of the data to update the | |
2641 | * gang header's blk_cksum as described above. | |
2642 | * | |
2643 | * The two-phase assemble/issue model solves the problem of partial failure -- | |
2644 | * what if you'd freed part of a gang block but then couldn't read the | |
2645 | * gang header for another part? Assembling the entire gang tree first | |
2646 | * ensures that all the necessary gang header I/O has succeeded before | |
2647 | * starting the actual work of free, claim, or write. Once the gang tree | |
2648 | * is assembled, free and claim are in-memory operations that cannot fail. | |
2649 | * | |
2650 | * In the event that a gang write fails, zio_dva_unallocate() walks the | |
2651 | * gang tree to immediately free (i.e. insert back into the space map) | |
2652 | * everything we've allocated. This ensures that we don't get ENOSPC | |
2653 | * errors during repeated suspend/resume cycles due to a flaky device. | |
2654 | * | |
2655 | * Gang rewrites only happen during sync-to-convergence. If we can't assemble | |
2656 | * the gang tree, we won't modify the block, so we can safely defer the free | |
2657 | * (knowing that the block is still intact). If we *can* assemble the gang | |
2658 | * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free | |
2659 | * each constituent bp and we can allocate a new block on the next sync pass. | |
2660 | * | |
2661 | * In all cases, the gang tree allows complete recovery from partial failure. | |
34dc7c2f BB |
2662 | * ========================================================================== |
2663 | */ | |
b128c09f | 2664 | |
a6255b7f DQ |
2665 | static void |
2666 | zio_gang_issue_func_done(zio_t *zio) | |
2667 | { | |
e2af2acc | 2668 | abd_free(zio->io_abd); |
a6255b7f DQ |
2669 | } |
2670 | ||
b128c09f | 2671 | static zio_t * |
a6255b7f DQ |
2672 | zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, |
2673 | uint64_t offset) | |
34dc7c2f | 2674 | { |
b128c09f BB |
2675 | if (gn != NULL) |
2676 | return (pio); | |
34dc7c2f | 2677 | |
a6255b7f DQ |
2678 | return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset), |
2679 | BP_GET_PSIZE(bp), zio_gang_issue_func_done, | |
2680 | NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
b128c09f BB |
2681 | &pio->io_bookmark)); |
2682 | } | |
2683 | ||
a6255b7f DQ |
2684 | static zio_t * |
2685 | zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2686 | uint64_t offset) | |
b128c09f BB |
2687 | { |
2688 | zio_t *zio; | |
2689 | ||
2690 | if (gn != NULL) { | |
a6255b7f DQ |
2691 | abd_t *gbh_abd = |
2692 | abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
b128c09f | 2693 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
a6255b7f DQ |
2694 | gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL, |
2695 | pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
2696 | &pio->io_bookmark); | |
34dc7c2f | 2697 | /* |
b128c09f BB |
2698 | * As we rewrite each gang header, the pipeline will compute |
2699 | * a new gang block header checksum for it; but no one will | |
2700 | * compute a new data checksum, so we do that here. The one | |
2701 | * exception is the gang leader: the pipeline already computed | |
2702 | * its data checksum because that stage precedes gang assembly. | |
2703 | * (Presently, nothing actually uses interior data checksums; | |
2704 | * this is just good hygiene.) | |
34dc7c2f | 2705 | */ |
9babb374 | 2706 | if (gn != pio->io_gang_leader->io_gang_tree) { |
a6255b7f DQ |
2707 | abd_t *buf = abd_get_offset(data, offset); |
2708 | ||
b128c09f | 2709 | zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), |
a6255b7f DQ |
2710 | buf, BP_GET_PSIZE(bp)); |
2711 | ||
e2af2acc | 2712 | abd_free(buf); |
b128c09f | 2713 | } |
428870ff BB |
2714 | /* |
2715 | * If we are here to damage data for testing purposes, | |
2716 | * leave the GBH alone so that we can detect the damage. | |
2717 | */ | |
2718 | if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) | |
2719 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; | |
34dc7c2f | 2720 | } else { |
b128c09f | 2721 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
a6255b7f DQ |
2722 | abd_get_offset(data, offset), BP_GET_PSIZE(bp), |
2723 | zio_gang_issue_func_done, NULL, pio->io_priority, | |
b128c09f | 2724 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); |
34dc7c2f BB |
2725 | } |
2726 | ||
b128c09f BB |
2727 | return (zio); |
2728 | } | |
34dc7c2f | 2729 | |
a6255b7f DQ |
2730 | static zio_t * |
2731 | zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2732 | uint64_t offset) | |
b128c09f | 2733 | { |
14e4e3cb AZ |
2734 | (void) gn, (void) data, (void) offset; |
2735 | ||
9cdf7b1f MA |
2736 | zio_t *zio = zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, |
2737 | ZIO_GANG_CHILD_FLAGS(pio)); | |
2738 | if (zio == NULL) { | |
2739 | zio = zio_null(pio, pio->io_spa, | |
2740 | NULL, NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)); | |
2741 | } | |
2742 | return (zio); | |
34dc7c2f BB |
2743 | } |
2744 | ||
a6255b7f DQ |
2745 | static zio_t * |
2746 | zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2747 | uint64_t offset) | |
34dc7c2f | 2748 | { |
14e4e3cb | 2749 | (void) gn, (void) data, (void) offset; |
b128c09f BB |
2750 | return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, |
2751 | NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); | |
2752 | } | |
2753 | ||
2754 | static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { | |
2755 | NULL, | |
2756 | zio_read_gang, | |
2757 | zio_rewrite_gang, | |
2758 | zio_free_gang, | |
2759 | zio_claim_gang, | |
2760 | NULL | |
2761 | }; | |
34dc7c2f | 2762 | |
b128c09f | 2763 | static void zio_gang_tree_assemble_done(zio_t *zio); |
34dc7c2f | 2764 | |
b128c09f BB |
2765 | static zio_gang_node_t * |
2766 | zio_gang_node_alloc(zio_gang_node_t **gnpp) | |
2767 | { | |
2768 | zio_gang_node_t *gn; | |
34dc7c2f | 2769 | |
b128c09f | 2770 | ASSERT(*gnpp == NULL); |
34dc7c2f | 2771 | |
79c76d5b | 2772 | gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); |
b128c09f BB |
2773 | gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); |
2774 | *gnpp = gn; | |
34dc7c2f | 2775 | |
b128c09f | 2776 | return (gn); |
34dc7c2f BB |
2777 | } |
2778 | ||
34dc7c2f | 2779 | static void |
b128c09f | 2780 | zio_gang_node_free(zio_gang_node_t **gnpp) |
34dc7c2f | 2781 | { |
b128c09f | 2782 | zio_gang_node_t *gn = *gnpp; |
34dc7c2f | 2783 | |
1c27024e | 2784 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f BB |
2785 | ASSERT(gn->gn_child[g] == NULL); |
2786 | ||
2787 | zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
2788 | kmem_free(gn, sizeof (*gn)); | |
2789 | *gnpp = NULL; | |
34dc7c2f BB |
2790 | } |
2791 | ||
b128c09f BB |
2792 | static void |
2793 | zio_gang_tree_free(zio_gang_node_t **gnpp) | |
34dc7c2f | 2794 | { |
b128c09f | 2795 | zio_gang_node_t *gn = *gnpp; |
34dc7c2f | 2796 | |
b128c09f BB |
2797 | if (gn == NULL) |
2798 | return; | |
34dc7c2f | 2799 | |
1c27024e | 2800 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f | 2801 | zio_gang_tree_free(&gn->gn_child[g]); |
34dc7c2f | 2802 | |
b128c09f | 2803 | zio_gang_node_free(gnpp); |
34dc7c2f BB |
2804 | } |
2805 | ||
b128c09f | 2806 | static void |
9babb374 | 2807 | zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) |
34dc7c2f | 2808 | { |
b128c09f | 2809 | zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); |
a6255b7f | 2810 | abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); |
b128c09f | 2811 | |
9babb374 | 2812 | ASSERT(gio->io_gang_leader == gio); |
b128c09f | 2813 | ASSERT(BP_IS_GANG(bp)); |
34dc7c2f | 2814 | |
a6255b7f DQ |
2815 | zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE, |
2816 | zio_gang_tree_assemble_done, gn, gio->io_priority, | |
2817 | ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); | |
b128c09f | 2818 | } |
34dc7c2f | 2819 | |
b128c09f BB |
2820 | static void |
2821 | zio_gang_tree_assemble_done(zio_t *zio) | |
2822 | { | |
9babb374 | 2823 | zio_t *gio = zio->io_gang_leader; |
b128c09f BB |
2824 | zio_gang_node_t *gn = zio->io_private; |
2825 | blkptr_t *bp = zio->io_bp; | |
34dc7c2f | 2826 | |
9babb374 | 2827 | ASSERT(gio == zio_unique_parent(zio)); |
ccec7fbe | 2828 | ASSERT(list_is_empty(&zio->io_child_list)); |
34dc7c2f | 2829 | |
b128c09f BB |
2830 | if (zio->io_error) |
2831 | return; | |
34dc7c2f | 2832 | |
a6255b7f | 2833 | /* this ABD was created from a linear buf in zio_gang_tree_assemble */ |
b128c09f | 2834 | if (BP_SHOULD_BYTESWAP(bp)) |
a6255b7f | 2835 | byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size); |
34dc7c2f | 2836 | |
a6255b7f | 2837 | ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh); |
b128c09f | 2838 | ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); |
428870ff | 2839 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2840 | |
e2af2acc | 2841 | abd_free(zio->io_abd); |
a6255b7f | 2842 | |
1c27024e | 2843 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2844 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2845 | if (!BP_IS_GANG(gbp)) | |
2846 | continue; | |
9babb374 | 2847 | zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); |
b128c09f | 2848 | } |
34dc7c2f BB |
2849 | } |
2850 | ||
b128c09f | 2851 | static void |
a6255b7f DQ |
2852 | zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data, |
2853 | uint64_t offset) | |
34dc7c2f | 2854 | { |
9babb374 | 2855 | zio_t *gio = pio->io_gang_leader; |
b128c09f | 2856 | zio_t *zio; |
34dc7c2f | 2857 | |
b128c09f | 2858 | ASSERT(BP_IS_GANG(bp) == !!gn); |
9babb374 BB |
2859 | ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); |
2860 | ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); | |
34dc7c2f | 2861 | |
b128c09f BB |
2862 | /* |
2863 | * If you're a gang header, your data is in gn->gn_gbh. | |
2864 | * If you're a gang member, your data is in 'data' and gn == NULL. | |
2865 | */ | |
a6255b7f | 2866 | zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset); |
34dc7c2f | 2867 | |
b128c09f | 2868 | if (gn != NULL) { |
428870ff | 2869 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2870 | |
1c27024e | 2871 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2872 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2873 | if (BP_IS_HOLE(gbp)) | |
2874 | continue; | |
a6255b7f DQ |
2875 | zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data, |
2876 | offset); | |
2877 | offset += BP_GET_PSIZE(gbp); | |
b128c09f | 2878 | } |
34dc7c2f BB |
2879 | } |
2880 | ||
9babb374 | 2881 | if (gn == gio->io_gang_tree) |
a6255b7f | 2882 | ASSERT3U(gio->io_size, ==, offset); |
34dc7c2f | 2883 | |
b128c09f BB |
2884 | if (zio != pio) |
2885 | zio_nowait(zio); | |
34dc7c2f BB |
2886 | } |
2887 | ||
62840030 | 2888 | static zio_t * |
b128c09f | 2889 | zio_gang_assemble(zio_t *zio) |
34dc7c2f | 2890 | { |
b128c09f | 2891 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2892 | |
9babb374 BB |
2893 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); |
2894 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
2895 | ||
2896 | zio->io_gang_leader = zio; | |
34dc7c2f | 2897 | |
b128c09f | 2898 | zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); |
34dc7c2f | 2899 | |
62840030 | 2900 | return (zio); |
34dc7c2f BB |
2901 | } |
2902 | ||
62840030 | 2903 | static zio_t * |
b128c09f | 2904 | zio_gang_issue(zio_t *zio) |
34dc7c2f | 2905 | { |
b128c09f | 2906 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2907 | |
ddc751d5 | 2908 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) { |
62840030 | 2909 | return (NULL); |
ddc751d5 | 2910 | } |
34dc7c2f | 2911 | |
9babb374 BB |
2912 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); |
2913 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
34dc7c2f | 2914 | |
b128c09f | 2915 | if (zio->io_child_error[ZIO_CHILD_GANG] == 0) |
a6255b7f DQ |
2916 | zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd, |
2917 | 0); | |
b128c09f | 2918 | else |
9babb374 | 2919 | zio_gang_tree_free(&zio->io_gang_tree); |
34dc7c2f | 2920 | |
b128c09f | 2921 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f | 2922 | |
62840030 | 2923 | return (zio); |
34dc7c2f BB |
2924 | } |
2925 | ||
3bd4df38 EN |
2926 | static void |
2927 | zio_gang_inherit_allocator(zio_t *pio, zio_t *cio) | |
2928 | { | |
2929 | cio->io_allocator = pio->io_allocator; | |
3bd4df38 EN |
2930 | } |
2931 | ||
34dc7c2f | 2932 | static void |
b128c09f | 2933 | zio_write_gang_member_ready(zio_t *zio) |
34dc7c2f | 2934 | { |
d164b209 | 2935 | zio_t *pio = zio_unique_parent(zio); |
34dc7c2f BB |
2936 | dva_t *cdva = zio->io_bp->blk_dva; |
2937 | dva_t *pdva = pio->io_bp->blk_dva; | |
2938 | uint64_t asize; | |
2a8ba608 | 2939 | zio_t *gio __maybe_unused = zio->io_gang_leader; |
34dc7c2f | 2940 | |
b128c09f BB |
2941 | if (BP_IS_HOLE(zio->io_bp)) |
2942 | return; | |
2943 | ||
2944 | ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); | |
2945 | ||
2946 | ASSERT(zio->io_child_type == ZIO_CHILD_GANG); | |
428870ff BB |
2947 | ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); |
2948 | ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); | |
2949 | ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); | |
14872aaa | 2950 | VERIFY3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); |
34dc7c2f BB |
2951 | |
2952 | mutex_enter(&pio->io_lock); | |
1c27024e | 2953 | for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { |
34dc7c2f BB |
2954 | ASSERT(DVA_GET_GANG(&pdva[d])); |
2955 | asize = DVA_GET_ASIZE(&pdva[d]); | |
2956 | asize += DVA_GET_ASIZE(&cdva[d]); | |
2957 | DVA_SET_ASIZE(&pdva[d], asize); | |
2958 | } | |
2959 | mutex_exit(&pio->io_lock); | |
2960 | } | |
2961 | ||
a6255b7f DQ |
2962 | static void |
2963 | zio_write_gang_done(zio_t *zio) | |
2964 | { | |
c955398b BL |
2965 | /* |
2966 | * The io_abd field will be NULL for a zio with no data. The io_flags | |
2967 | * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't | |
2968 | * check for it here as it is cleared in zio_ready. | |
2969 | */ | |
2970 | if (zio->io_abd != NULL) | |
e2af2acc | 2971 | abd_free(zio->io_abd); |
a6255b7f DQ |
2972 | } |
2973 | ||
62840030 | 2974 | static zio_t * |
aa755b35 | 2975 | zio_write_gang_block(zio_t *pio, metaslab_class_t *mc) |
34dc7c2f | 2976 | { |
b128c09f BB |
2977 | spa_t *spa = pio->io_spa; |
2978 | blkptr_t *bp = pio->io_bp; | |
9babb374 | 2979 | zio_t *gio = pio->io_gang_leader; |
b128c09f BB |
2980 | zio_t *zio; |
2981 | zio_gang_node_t *gn, **gnpp; | |
34dc7c2f | 2982 | zio_gbh_phys_t *gbh; |
a6255b7f | 2983 | abd_t *gbh_abd; |
b128c09f BB |
2984 | uint64_t txg = pio->io_txg; |
2985 | uint64_t resid = pio->io_size; | |
2986 | uint64_t lsize; | |
428870ff | 2987 | int copies = gio->io_prop.zp_copies; |
b128c09f | 2988 | zio_prop_t zp; |
1c27024e | 2989 | int error; |
c955398b | 2990 | boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA); |
b5256303 TC |
2991 | |
2992 | /* | |
14872aaa MA |
2993 | * If one copy was requested, store 2 copies of the GBH, so that we |
2994 | * can still traverse all the data (e.g. to free or scrub) even if a | |
2995 | * block is damaged. Note that we can't store 3 copies of the GBH in | |
2996 | * all cases, e.g. with encryption, which uses DVA[2] for the IV+salt. | |
b5256303 | 2997 | */ |
14872aaa MA |
2998 | int gbh_copies = copies; |
2999 | if (gbh_copies == 1) { | |
3000 | gbh_copies = MIN(2, spa_max_replication(spa)); | |
3001 | } | |
b5256303 | 3002 | |
3bd4df38 | 3003 | ASSERT(ZIO_HAS_ALLOCATOR(pio)); |
1c27024e | 3004 | int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER; |
3dfb57a3 DB |
3005 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
3006 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
c955398b | 3007 | ASSERT(has_data); |
3dfb57a3 DB |
3008 | |
3009 | flags |= METASLAB_ASYNC_ALLOC; | |
f8020c93 AM |
3010 | VERIFY(zfs_refcount_held(&mc->mc_allocator[pio->io_allocator]. |
3011 | mca_alloc_slots, pio)); | |
3dfb57a3 DB |
3012 | |
3013 | /* | |
3014 | * The logical zio has already placed a reservation for | |
3015 | * 'copies' allocation slots but gang blocks may require | |
3016 | * additional copies. These additional copies | |
3017 | * (i.e. gbh_copies - copies) are guaranteed to succeed | |
3018 | * since metaslab_class_throttle_reserve() always allows | |
3019 | * additional reservations for gang blocks. | |
3020 | */ | |
3021 | VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies, | |
492f64e9 | 3022 | pio->io_allocator, pio, flags)); |
3dfb57a3 DB |
3023 | } |
3024 | ||
3025 | error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE, | |
4e21fd06 | 3026 | bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, |
492f64e9 | 3027 | &pio->io_alloc_list, pio, pio->io_allocator); |
34dc7c2f | 3028 | if (error) { |
3dfb57a3 DB |
3029 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
3030 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
c955398b | 3031 | ASSERT(has_data); |
3dfb57a3 DB |
3032 | |
3033 | /* | |
3034 | * If we failed to allocate the gang block header then | |
3035 | * we remove any additional allocation reservations that | |
3036 | * we placed here. The original reservation will | |
3037 | * be removed when the logical I/O goes to the ready | |
3038 | * stage. | |
3039 | */ | |
3040 | metaslab_class_throttle_unreserve(mc, | |
492f64e9 | 3041 | gbh_copies - copies, pio->io_allocator, pio); |
3dfb57a3 DB |
3042 | } |
3043 | ||
b128c09f | 3044 | pio->io_error = error; |
62840030 | 3045 | return (pio); |
34dc7c2f BB |
3046 | } |
3047 | ||
9babb374 BB |
3048 | if (pio == gio) { |
3049 | gnpp = &gio->io_gang_tree; | |
b128c09f BB |
3050 | } else { |
3051 | gnpp = pio->io_private; | |
3052 | ASSERT(pio->io_ready == zio_write_gang_member_ready); | |
34dc7c2f BB |
3053 | } |
3054 | ||
b128c09f BB |
3055 | gn = zio_gang_node_alloc(gnpp); |
3056 | gbh = gn->gn_gbh; | |
861166b0 | 3057 | memset(gbh, 0, SPA_GANGBLOCKSIZE); |
a6255b7f | 3058 | gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE); |
34dc7c2f | 3059 | |
b128c09f BB |
3060 | /* |
3061 | * Create the gang header. | |
3062 | */ | |
a6255b7f DQ |
3063 | zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE, |
3064 | zio_write_gang_done, NULL, pio->io_priority, | |
3065 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); | |
34dc7c2f | 3066 | |
3bd4df38 EN |
3067 | zio_gang_inherit_allocator(pio, zio); |
3068 | ||
b128c09f BB |
3069 | /* |
3070 | * Create and nowait the gang children. | |
3071 | */ | |
1c27024e | 3072 | for (int g = 0; resid != 0; resid -= lsize, g++) { |
b128c09f BB |
3073 | lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), |
3074 | SPA_MINBLOCKSIZE); | |
3075 | ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); | |
3076 | ||
9babb374 | 3077 | zp.zp_checksum = gio->io_prop.zp_checksum; |
b128c09f | 3078 | zp.zp_compress = ZIO_COMPRESS_OFF; |
10b3c7f5 | 3079 | zp.zp_complevel = gio->io_prop.zp_complevel; |
b128c09f BB |
3080 | zp.zp_type = DMU_OT_NONE; |
3081 | zp.zp_level = 0; | |
428870ff | 3082 | zp.zp_copies = gio->io_prop.zp_copies; |
03c6040b GW |
3083 | zp.zp_dedup = B_FALSE; |
3084 | zp.zp_dedup_verify = B_FALSE; | |
3085 | zp.zp_nopwrite = B_FALSE; | |
4807c0ba TC |
3086 | zp.zp_encrypt = gio->io_prop.zp_encrypt; |
3087 | zp.zp_byteorder = gio->io_prop.zp_byteorder; | |
861166b0 AZ |
3088 | memset(zp.zp_salt, 0, ZIO_DATA_SALT_LEN); |
3089 | memset(zp.zp_iv, 0, ZIO_DATA_IV_LEN); | |
3090 | memset(zp.zp_mac, 0, ZIO_DATA_MAC_LEN); | |
b128c09f | 3091 | |
1c27024e | 3092 | zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g], |
c955398b BL |
3093 | has_data ? abd_get_offset(pio->io_abd, pio->io_size - |
3094 | resid) : NULL, lsize, lsize, &zp, | |
ccec7fbe | 3095 | zio_write_gang_member_ready, NULL, |
a6255b7f | 3096 | zio_write_gang_done, &gn->gn_child[g], pio->io_priority, |
3dfb57a3 DB |
3097 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); |
3098 | ||
3bd4df38 EN |
3099 | zio_gang_inherit_allocator(zio, cio); |
3100 | ||
3dfb57a3 DB |
3101 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
3102 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
c955398b | 3103 | ASSERT(has_data); |
3dfb57a3 DB |
3104 | |
3105 | /* | |
3106 | * Gang children won't throttle but we should | |
3107 | * account for their work, so reserve an allocation | |
3108 | * slot for them here. | |
3109 | */ | |
3110 | VERIFY(metaslab_class_throttle_reserve(mc, | |
492f64e9 | 3111 | zp.zp_copies, cio->io_allocator, cio, flags)); |
3dfb57a3 DB |
3112 | } |
3113 | zio_nowait(cio); | |
b128c09f | 3114 | } |
34dc7c2f BB |
3115 | |
3116 | /* | |
b128c09f | 3117 | * Set pio's pipeline to just wait for zio to finish. |
34dc7c2f | 3118 | */ |
b128c09f BB |
3119 | pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
3120 | ||
3121 | zio_nowait(zio); | |
3122 | ||
62840030 | 3123 | return (pio); |
34dc7c2f BB |
3124 | } |
3125 | ||
03c6040b | 3126 | /* |
3c67d83a TH |
3127 | * The zio_nop_write stage in the pipeline determines if allocating a |
3128 | * new bp is necessary. The nopwrite feature can handle writes in | |
3129 | * either syncing or open context (i.e. zil writes) and as a result is | |
3130 | * mutually exclusive with dedup. | |
3131 | * | |
3132 | * By leveraging a cryptographically secure checksum, such as SHA256, we | |
3133 | * can compare the checksums of the new data and the old to determine if | |
3134 | * allocating a new block is required. Note that our requirements for | |
3135 | * cryptographic strength are fairly weak: there can't be any accidental | |
3136 | * hash collisions, but we don't need to be secure against intentional | |
3137 | * (malicious) collisions. To trigger a nopwrite, you have to be able | |
3138 | * to write the file to begin with, and triggering an incorrect (hash | |
3139 | * collision) nopwrite is no worse than simply writing to the file. | |
3140 | * That said, there are no known attacks against the checksum algorithms | |
3141 | * used for nopwrite, assuming that the salt and the checksums | |
3142 | * themselves remain secret. | |
03c6040b | 3143 | */ |
62840030 | 3144 | static zio_t * |
03c6040b GW |
3145 | zio_nop_write(zio_t *zio) |
3146 | { | |
3147 | blkptr_t *bp = zio->io_bp; | |
3148 | blkptr_t *bp_orig = &zio->io_bp_orig; | |
3149 | zio_prop_t *zp = &zio->io_prop; | |
3150 | ||
d7cf06a2 | 3151 | ASSERT(BP_IS_HOLE(bp)); |
03c6040b GW |
3152 | ASSERT(BP_GET_LEVEL(bp) == 0); |
3153 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
3154 | ASSERT(zp->zp_nopwrite); | |
3155 | ASSERT(!zp->zp_dedup); | |
3156 | ASSERT(zio->io_bp_override == NULL); | |
3157 | ASSERT(IO_IS_ALLOCATING(zio)); | |
3158 | ||
3159 | /* | |
3160 | * Check to see if the original bp and the new bp have matching | |
3161 | * characteristics (i.e. same checksum, compression algorithms, etc). | |
3162 | * If they don't then just continue with the pipeline which will | |
3163 | * allocate a new bp. | |
3164 | */ | |
3165 | if (BP_IS_HOLE(bp_orig) || | |
3c67d83a TH |
3166 | !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & |
3167 | ZCHECKSUM_FLAG_NOPWRITE) || | |
b5256303 | 3168 | BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) || |
03c6040b GW |
3169 | BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || |
3170 | BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || | |
3171 | BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || | |
3172 | zp->zp_copies != BP_GET_NDVAS(bp_orig)) | |
62840030 | 3173 | return (zio); |
03c6040b GW |
3174 | |
3175 | /* | |
3176 | * If the checksums match then reset the pipeline so that we | |
3177 | * avoid allocating a new bp and issuing any I/O. | |
3178 | */ | |
3179 | if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { | |
3c67d83a TH |
3180 | ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & |
3181 | ZCHECKSUM_FLAG_NOPWRITE); | |
03c6040b GW |
3182 | ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); |
3183 | ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); | |
3184 | ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); | |
d7cf06a2 | 3185 | ASSERT3U(bp->blk_prop, ==, bp_orig->blk_prop); |
03c6040b | 3186 | |
681a85cb GW |
3187 | /* |
3188 | * If we're overwriting a block that is currently on an | |
3189 | * indirect vdev, then ignore the nopwrite request and | |
3190 | * allow a new block to be allocated on a concrete vdev. | |
3191 | */ | |
3192 | spa_config_enter(zio->io_spa, SCL_VDEV, FTAG, RW_READER); | |
d7cf06a2 GW |
3193 | for (int d = 0; d < BP_GET_NDVAS(bp_orig); d++) { |
3194 | vdev_t *tvd = vdev_lookup_top(zio->io_spa, | |
3195 | DVA_GET_VDEV(&bp_orig->blk_dva[d])); | |
3196 | if (tvd->vdev_ops == &vdev_indirect_ops) { | |
3197 | spa_config_exit(zio->io_spa, SCL_VDEV, FTAG); | |
3198 | return (zio); | |
3199 | } | |
681a85cb GW |
3200 | } |
3201 | spa_config_exit(zio->io_spa, SCL_VDEV, FTAG); | |
3202 | ||
03c6040b GW |
3203 | *bp = *bp_orig; |
3204 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
3205 | zio->io_flags |= ZIO_FLAG_NOPWRITE; | |
3206 | } | |
3207 | ||
62840030 | 3208 | return (zio); |
03c6040b GW |
3209 | } |
3210 | ||
67a1b037 PJD |
3211 | /* |
3212 | * ========================================================================== | |
3213 | * Block Reference Table | |
3214 | * ========================================================================== | |
3215 | */ | |
3216 | static zio_t * | |
3217 | zio_brt_free(zio_t *zio) | |
3218 | { | |
3219 | blkptr_t *bp; | |
3220 | ||
3221 | bp = zio->io_bp; | |
3222 | ||
3223 | if (BP_GET_LEVEL(bp) > 0 || | |
3224 | BP_IS_METADATA(bp) || | |
3225 | !brt_maybe_exists(zio->io_spa, bp)) { | |
3226 | return (zio); | |
3227 | } | |
3228 | ||
3229 | if (!brt_entry_decref(zio->io_spa, bp)) { | |
3230 | /* | |
3231 | * This isn't the last reference, so we cannot free | |
3232 | * the data yet. | |
3233 | */ | |
3234 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
3235 | } | |
3236 | ||
3237 | return (zio); | |
3238 | } | |
3239 | ||
34dc7c2f BB |
3240 | /* |
3241 | * ========================================================================== | |
428870ff | 3242 | * Dedup |
34dc7c2f BB |
3243 | * ========================================================================== |
3244 | */ | |
428870ff BB |
3245 | static void |
3246 | zio_ddt_child_read_done(zio_t *zio) | |
3247 | { | |
3248 | blkptr_t *bp = zio->io_bp; | |
3249 | ddt_entry_t *dde = zio->io_private; | |
3250 | ddt_phys_t *ddp; | |
3251 | zio_t *pio = zio_unique_parent(zio); | |
3252 | ||
3253 | mutex_enter(&pio->io_lock); | |
3254 | ddp = ddt_phys_select(dde, bp); | |
3255 | if (zio->io_error == 0) | |
3256 | ddt_phys_clear(ddp); /* this ddp doesn't need repair */ | |
a6255b7f DQ |
3257 | |
3258 | if (zio->io_error == 0 && dde->dde_repair_abd == NULL) | |
3259 | dde->dde_repair_abd = zio->io_abd; | |
428870ff | 3260 | else |
a6255b7f | 3261 | abd_free(zio->io_abd); |
428870ff BB |
3262 | mutex_exit(&pio->io_lock); |
3263 | } | |
3264 | ||
62840030 | 3265 | static zio_t * |
428870ff BB |
3266 | zio_ddt_read_start(zio_t *zio) |
3267 | { | |
3268 | blkptr_t *bp = zio->io_bp; | |
3269 | ||
3270 | ASSERT(BP_GET_DEDUP(bp)); | |
3271 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
3272 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
3273 | ||
3274 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
3275 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
3276 | ddt_entry_t *dde = ddt_repair_start(ddt, bp); | |
3277 | ddt_phys_t *ddp = dde->dde_phys; | |
3278 | ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); | |
3279 | blkptr_t blk; | |
3280 | ||
3281 | ASSERT(zio->io_vsd == NULL); | |
3282 | zio->io_vsd = dde; | |
3283 | ||
3284 | if (ddp_self == NULL) | |
62840030 | 3285 | return (zio); |
428870ff | 3286 | |
1c27024e | 3287 | for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { |
428870ff BB |
3288 | if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) |
3289 | continue; | |
3290 | ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, | |
3291 | &blk); | |
3292 | zio_nowait(zio_read(zio, zio->io_spa, &blk, | |
a6255b7f DQ |
3293 | abd_alloc_for_io(zio->io_size, B_TRUE), |
3294 | zio->io_size, zio_ddt_child_read_done, dde, | |
3295 | zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) | | |
3296 | ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark)); | |
428870ff | 3297 | } |
62840030 | 3298 | return (zio); |
428870ff BB |
3299 | } |
3300 | ||
3301 | zio_nowait(zio_read(zio, zio->io_spa, bp, | |
a6255b7f | 3302 | zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority, |
428870ff BB |
3303 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); |
3304 | ||
62840030 | 3305 | return (zio); |
428870ff BB |
3306 | } |
3307 | ||
62840030 | 3308 | static zio_t * |
428870ff BB |
3309 | zio_ddt_read_done(zio_t *zio) |
3310 | { | |
3311 | blkptr_t *bp = zio->io_bp; | |
3312 | ||
ddc751d5 | 3313 | if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) { |
62840030 | 3314 | return (NULL); |
ddc751d5 | 3315 | } |
428870ff BB |
3316 | |
3317 | ASSERT(BP_GET_DEDUP(bp)); | |
3318 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
3319 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
3320 | ||
3321 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
3322 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
3323 | ddt_entry_t *dde = zio->io_vsd; | |
3324 | if (ddt == NULL) { | |
3325 | ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); | |
62840030 | 3326 | return (zio); |
428870ff BB |
3327 | } |
3328 | if (dde == NULL) { | |
3329 | zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; | |
3330 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); | |
62840030 | 3331 | return (NULL); |
428870ff | 3332 | } |
a6255b7f DQ |
3333 | if (dde->dde_repair_abd != NULL) { |
3334 | abd_copy(zio->io_abd, dde->dde_repair_abd, | |
3335 | zio->io_size); | |
428870ff BB |
3336 | zio->io_child_error[ZIO_CHILD_DDT] = 0; |
3337 | } | |
3338 | ddt_repair_done(ddt, dde); | |
3339 | zio->io_vsd = NULL; | |
3340 | } | |
3341 | ||
3342 | ASSERT(zio->io_vsd == NULL); | |
3343 | ||
62840030 | 3344 | return (zio); |
428870ff BB |
3345 | } |
3346 | ||
3347 | static boolean_t | |
3348 | zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) | |
3349 | { | |
3350 | spa_t *spa = zio->io_spa; | |
c17bcf83 | 3351 | boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW); |
428870ff | 3352 | |
c17bcf83 | 3353 | ASSERT(!(zio->io_bp_override && do_raw)); |
2aa34383 | 3354 | |
428870ff BB |
3355 | /* |
3356 | * Note: we compare the original data, not the transformed data, | |
3357 | * because when zio->io_bp is an override bp, we will not have | |
3358 | * pushed the I/O transforms. That's an important optimization | |
3359 | * because otherwise we'd compress/encrypt all dmu_sync() data twice. | |
c17bcf83 | 3360 | * However, we should never get a raw, override zio so in these |
b5256303 | 3361 | * cases we can compare the io_abd directly. This is useful because |
c17bcf83 TC |
3362 | * it allows us to do dedup verification even if we don't have access |
3363 | * to the original data (for instance, if the encryption keys aren't | |
3364 | * loaded). | |
428870ff | 3365 | */ |
c17bcf83 | 3366 | |
1c27024e | 3367 | for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
3368 | zio_t *lio = dde->dde_lead_zio[p]; |
3369 | ||
c17bcf83 TC |
3370 | if (lio != NULL && do_raw) { |
3371 | return (lio->io_size != zio->io_size || | |
a6255b7f | 3372 | abd_cmp(zio->io_abd, lio->io_abd) != 0); |
c17bcf83 | 3373 | } else if (lio != NULL) { |
428870ff | 3374 | return (lio->io_orig_size != zio->io_orig_size || |
a6255b7f | 3375 | abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0); |
428870ff BB |
3376 | } |
3377 | } | |
3378 | ||
1c27024e | 3379 | for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
3380 | ddt_phys_t *ddp = &dde->dde_phys[p]; |
3381 | ||
c17bcf83 TC |
3382 | if (ddp->ddp_phys_birth != 0 && do_raw) { |
3383 | blkptr_t blk = *zio->io_bp; | |
3384 | uint64_t psize; | |
a6255b7f | 3385 | abd_t *tmpabd; |
c17bcf83 TC |
3386 | int error; |
3387 | ||
3388 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
3389 | psize = BP_GET_PSIZE(&blk); | |
3390 | ||
3391 | if (psize != zio->io_size) | |
3392 | return (B_TRUE); | |
3393 | ||
3394 | ddt_exit(ddt); | |
3395 | ||
a6255b7f | 3396 | tmpabd = abd_alloc_for_io(psize, B_TRUE); |
c17bcf83 | 3397 | |
a6255b7f | 3398 | error = zio_wait(zio_read(NULL, spa, &blk, tmpabd, |
c17bcf83 TC |
3399 | psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ, |
3400 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
3401 | ZIO_FLAG_RAW, &zio->io_bookmark)); | |
3402 | ||
3403 | if (error == 0) { | |
a6255b7f | 3404 | if (abd_cmp(tmpabd, zio->io_abd) != 0) |
c17bcf83 TC |
3405 | error = SET_ERROR(ENOENT); |
3406 | } | |
3407 | ||
a6255b7f | 3408 | abd_free(tmpabd); |
c17bcf83 TC |
3409 | ddt_enter(ddt); |
3410 | return (error != 0); | |
3411 | } else if (ddp->ddp_phys_birth != 0) { | |
428870ff | 3412 | arc_buf_t *abuf = NULL; |
2a432414 | 3413 | arc_flags_t aflags = ARC_FLAG_WAIT; |
428870ff BB |
3414 | blkptr_t blk = *zio->io_bp; |
3415 | int error; | |
3416 | ||
3417 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
3418 | ||
c17bcf83 TC |
3419 | if (BP_GET_LSIZE(&blk) != zio->io_orig_size) |
3420 | return (B_TRUE); | |
3421 | ||
428870ff BB |
3422 | ddt_exit(ddt); |
3423 | ||
294f6806 | 3424 | error = arc_read(NULL, spa, &blk, |
428870ff BB |
3425 | arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, |
3426 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
3427 | &aflags, &zio->io_bookmark); | |
3428 | ||
3429 | if (error == 0) { | |
a6255b7f | 3430 | if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data, |
428870ff | 3431 | zio->io_orig_size) != 0) |
c17bcf83 | 3432 | error = SET_ERROR(ENOENT); |
d3c2ae1c | 3433 | arc_buf_destroy(abuf, &abuf); |
428870ff BB |
3434 | } |
3435 | ||
3436 | ddt_enter(ddt); | |
3437 | return (error != 0); | |
3438 | } | |
3439 | } | |
3440 | ||
3441 | return (B_FALSE); | |
3442 | } | |
3443 | ||
3444 | static void | |
3445 | zio_ddt_child_write_ready(zio_t *zio) | |
3446 | { | |
3447 | int p = zio->io_prop.zp_copies; | |
3448 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
3449 | ddt_entry_t *dde = zio->io_private; | |
3450 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
3451 | zio_t *pio; | |
3452 | ||
3453 | if (zio->io_error) | |
3454 | return; | |
3455 | ||
3456 | ddt_enter(ddt); | |
3457 | ||
3458 | ASSERT(dde->dde_lead_zio[p] == zio); | |
3459 | ||
3460 | ddt_phys_fill(ddp, zio->io_bp); | |
3461 | ||
1c27024e | 3462 | zio_link_t *zl = NULL; |
3dfb57a3 | 3463 | while ((pio = zio_walk_parents(zio, &zl)) != NULL) |
428870ff BB |
3464 | ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); |
3465 | ||
3466 | ddt_exit(ddt); | |
3467 | } | |
3468 | ||
3469 | static void | |
3470 | zio_ddt_child_write_done(zio_t *zio) | |
3471 | { | |
3472 | int p = zio->io_prop.zp_copies; | |
3473 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
3474 | ddt_entry_t *dde = zio->io_private; | |
3475 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
3476 | ||
3477 | ddt_enter(ddt); | |
3478 | ||
3479 | ASSERT(ddp->ddp_refcnt == 0); | |
3480 | ASSERT(dde->dde_lead_zio[p] == zio); | |
3481 | dde->dde_lead_zio[p] = NULL; | |
3482 | ||
3483 | if (zio->io_error == 0) { | |
3dfb57a3 DB |
3484 | zio_link_t *zl = NULL; |
3485 | while (zio_walk_parents(zio, &zl) != NULL) | |
428870ff BB |
3486 | ddt_phys_addref(ddp); |
3487 | } else { | |
3488 | ddt_phys_clear(ddp); | |
3489 | } | |
3490 | ||
3491 | ddt_exit(ddt); | |
3492 | } | |
3493 | ||
62840030 | 3494 | static zio_t * |
428870ff BB |
3495 | zio_ddt_write(zio_t *zio) |
3496 | { | |
3497 | spa_t *spa = zio->io_spa; | |
3498 | blkptr_t *bp = zio->io_bp; | |
3499 | uint64_t txg = zio->io_txg; | |
3500 | zio_prop_t *zp = &zio->io_prop; | |
3501 | int p = zp->zp_copies; | |
428870ff | 3502 | zio_t *cio = NULL; |
428870ff BB |
3503 | ddt_t *ddt = ddt_select(spa, bp); |
3504 | ddt_entry_t *dde; | |
3505 | ddt_phys_t *ddp; | |
3506 | ||
3507 | ASSERT(BP_GET_DEDUP(bp)); | |
3508 | ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); | |
3509 | ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); | |
c17bcf83 | 3510 | ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW))); |
428870ff BB |
3511 | |
3512 | ddt_enter(ddt); | |
3513 | dde = ddt_lookup(ddt, bp, B_TRUE); | |
3514 | ddp = &dde->dde_phys[p]; | |
3515 | ||
3516 | if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { | |
3517 | /* | |
3518 | * If we're using a weak checksum, upgrade to a strong checksum | |
3519 | * and try again. If we're already using a strong checksum, | |
3520 | * we can't resolve it, so just convert to an ordinary write. | |
3521 | * (And automatically e-mail a paper to Nature?) | |
3522 | */ | |
3c67d83a TH |
3523 | if (!(zio_checksum_table[zp->zp_checksum].ci_flags & |
3524 | ZCHECKSUM_FLAG_DEDUP)) { | |
428870ff BB |
3525 | zp->zp_checksum = spa_dedup_checksum(spa); |
3526 | zio_pop_transforms(zio); | |
3527 | zio->io_stage = ZIO_STAGE_OPEN; | |
3528 | BP_ZERO(bp); | |
3529 | } else { | |
03c6040b | 3530 | zp->zp_dedup = B_FALSE; |
accd6d9d | 3531 | BP_SET_DEDUP(bp, B_FALSE); |
428870ff | 3532 | } |
accd6d9d | 3533 | ASSERT(!BP_GET_DEDUP(bp)); |
428870ff BB |
3534 | zio->io_pipeline = ZIO_WRITE_PIPELINE; |
3535 | ddt_exit(ddt); | |
62840030 | 3536 | return (zio); |
428870ff BB |
3537 | } |
3538 | ||
428870ff BB |
3539 | if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { |
3540 | if (ddp->ddp_phys_birth != 0) | |
3541 | ddt_bp_fill(ddp, bp, txg); | |
3542 | if (dde->dde_lead_zio[p] != NULL) | |
3543 | zio_add_child(zio, dde->dde_lead_zio[p]); | |
3544 | else | |
3545 | ddt_phys_addref(ddp); | |
3546 | } else if (zio->io_bp_override) { | |
493fcce9 | 3547 | ASSERT(BP_GET_LOGICAL_BIRTH(bp) == txg); |
428870ff BB |
3548 | ASSERT(BP_EQUAL(bp, zio->io_bp_override)); |
3549 | ddt_phys_fill(ddp, bp); | |
3550 | ddt_phys_addref(ddp); | |
3551 | } else { | |
a6255b7f | 3552 | cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, |
2aa34383 | 3553 | zio->io_orig_size, zio->io_orig_size, zp, |
ccec7fbe | 3554 | zio_ddt_child_write_ready, NULL, |
428870ff BB |
3555 | zio_ddt_child_write_done, dde, zio->io_priority, |
3556 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); | |
3557 | ||
a6255b7f | 3558 | zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL); |
428870ff BB |
3559 | dde->dde_lead_zio[p] = cio; |
3560 | } | |
3561 | ||
3562 | ddt_exit(ddt); | |
3563 | ||
9cdf7b1f | 3564 | zio_nowait(cio); |
428870ff | 3565 | |
62840030 | 3566 | return (zio); |
428870ff BB |
3567 | } |
3568 | ||
27218a32 | 3569 | static ddt_entry_t *freedde; /* for debugging */ |
b128c09f | 3570 | |
62840030 | 3571 | static zio_t * |
428870ff BB |
3572 | zio_ddt_free(zio_t *zio) |
3573 | { | |
3574 | spa_t *spa = zio->io_spa; | |
3575 | blkptr_t *bp = zio->io_bp; | |
3576 | ddt_t *ddt = ddt_select(spa, bp); | |
3577 | ddt_entry_t *dde; | |
3578 | ddt_phys_t *ddp; | |
3579 | ||
3580 | ASSERT(BP_GET_DEDUP(bp)); | |
3581 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
3582 | ||
3583 | ddt_enter(ddt); | |
3584 | freedde = dde = ddt_lookup(ddt, bp, B_TRUE); | |
5dc6af0e BB |
3585 | if (dde) { |
3586 | ddp = ddt_phys_select(dde, bp); | |
3587 | if (ddp) | |
3588 | ddt_phys_decref(ddp); | |
3589 | } | |
428870ff BB |
3590 | ddt_exit(ddt); |
3591 | ||
62840030 | 3592 | return (zio); |
428870ff BB |
3593 | } |
3594 | ||
3595 | /* | |
3596 | * ========================================================================== | |
3597 | * Allocate and free blocks | |
3598 | * ========================================================================== | |
3599 | */ | |
3dfb57a3 DB |
3600 | |
3601 | static zio_t * | |
492f64e9 | 3602 | zio_io_to_allocate(spa_t *spa, int allocator) |
3dfb57a3 DB |
3603 | { |
3604 | zio_t *zio; | |
3605 | ||
1b50749c | 3606 | ASSERT(MUTEX_HELD(&spa->spa_allocs[allocator].spaa_lock)); |
3dfb57a3 | 3607 | |
1b50749c | 3608 | zio = avl_first(&spa->spa_allocs[allocator].spaa_tree); |
3dfb57a3 DB |
3609 | if (zio == NULL) |
3610 | return (NULL); | |
3611 | ||
3612 | ASSERT(IO_IS_ALLOCATING(zio)); | |
3bd4df38 | 3613 | ASSERT(ZIO_HAS_ALLOCATOR(zio)); |
3dfb57a3 DB |
3614 | |
3615 | /* | |
3616 | * Try to place a reservation for this zio. If we're unable to | |
3617 | * reserve then we throttle. | |
3618 | */ | |
492f64e9 | 3619 | ASSERT3U(zio->io_allocator, ==, allocator); |
cc99f275 | 3620 | if (!metaslab_class_throttle_reserve(zio->io_metaslab_class, |
1b50749c | 3621 | zio->io_prop.zp_copies, allocator, zio, 0)) { |
3dfb57a3 DB |
3622 | return (NULL); |
3623 | } | |
3624 | ||
1b50749c | 3625 | avl_remove(&spa->spa_allocs[allocator].spaa_tree, zio); |
3dfb57a3 DB |
3626 | ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE); |
3627 | ||
3628 | return (zio); | |
3629 | } | |
3630 | ||
62840030 | 3631 | static zio_t * |
3dfb57a3 DB |
3632 | zio_dva_throttle(zio_t *zio) |
3633 | { | |
3634 | spa_t *spa = zio->io_spa; | |
3635 | zio_t *nio; | |
cc99f275 DB |
3636 | metaslab_class_t *mc; |
3637 | ||
3638 | /* locate an appropriate allocation class */ | |
3639 | mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type, | |
3640 | zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk); | |
3dfb57a3 DB |
3641 | |
3642 | if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE || | |
cc99f275 | 3643 | !mc->mc_alloc_throttle_enabled || |
3dfb57a3 DB |
3644 | zio->io_child_type == ZIO_CHILD_GANG || |
3645 | zio->io_flags & ZIO_FLAG_NODATA) { | |
62840030 | 3646 | return (zio); |
3dfb57a3 DB |
3647 | } |
3648 | ||
1b50749c | 3649 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); |
3bd4df38 | 3650 | ASSERT(ZIO_HAS_ALLOCATOR(zio)); |
3dfb57a3 | 3651 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); |
3dfb57a3 DB |
3652 | ASSERT3U(zio->io_queued_timestamp, >, 0); |
3653 | ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE); | |
3654 | ||
3bd4df38 | 3655 | int allocator = zio->io_allocator; |
cc99f275 | 3656 | zio->io_metaslab_class = mc; |
1b50749c AM |
3657 | mutex_enter(&spa->spa_allocs[allocator].spaa_lock); |
3658 | avl_add(&spa->spa_allocs[allocator].spaa_tree, zio); | |
3659 | nio = zio_io_to_allocate(spa, allocator); | |
3660 | mutex_exit(&spa->spa_allocs[allocator].spaa_lock); | |
62840030 | 3661 | return (nio); |
3dfb57a3 DB |
3662 | } |
3663 | ||
cc99f275 | 3664 | static void |
492f64e9 | 3665 | zio_allocate_dispatch(spa_t *spa, int allocator) |
3dfb57a3 DB |
3666 | { |
3667 | zio_t *zio; | |
3668 | ||
1b50749c | 3669 | mutex_enter(&spa->spa_allocs[allocator].spaa_lock); |
492f64e9 | 3670 | zio = zio_io_to_allocate(spa, allocator); |
1b50749c | 3671 | mutex_exit(&spa->spa_allocs[allocator].spaa_lock); |
3dfb57a3 DB |
3672 | if (zio == NULL) |
3673 | return; | |
3674 | ||
3675 | ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE); | |
3676 | ASSERT0(zio->io_error); | |
3677 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE); | |
3678 | } | |
3679 | ||
62840030 | 3680 | static zio_t * |
34dc7c2f BB |
3681 | zio_dva_allocate(zio_t *zio) |
3682 | { | |
3683 | spa_t *spa = zio->io_spa; | |
cc99f275 | 3684 | metaslab_class_t *mc; |
34dc7c2f BB |
3685 | blkptr_t *bp = zio->io_bp; |
3686 | int error; | |
6d974228 | 3687 | int flags = 0; |
34dc7c2f | 3688 | |
9babb374 BB |
3689 | if (zio->io_gang_leader == NULL) { |
3690 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
3691 | zio->io_gang_leader = zio; | |
3692 | } | |
3693 | ||
34dc7c2f | 3694 | ASSERT(BP_IS_HOLE(bp)); |
c99c9001 | 3695 | ASSERT0(BP_GET_NDVAS(bp)); |
428870ff BB |
3696 | ASSERT3U(zio->io_prop.zp_copies, >, 0); |
3697 | ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); | |
34dc7c2f BB |
3698 | ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); |
3699 | ||
3dfb57a3 DB |
3700 | if (zio->io_flags & ZIO_FLAG_NODATA) |
3701 | flags |= METASLAB_DONT_THROTTLE; | |
3702 | if (zio->io_flags & ZIO_FLAG_GANG_CHILD) | |
3703 | flags |= METASLAB_GANG_CHILD; | |
3704 | if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) | |
3705 | flags |= METASLAB_ASYNC_ALLOC; | |
3706 | ||
cc99f275 DB |
3707 | /* |
3708 | * if not already chosen, locate an appropriate allocation class | |
3709 | */ | |
3710 | mc = zio->io_metaslab_class; | |
3711 | if (mc == NULL) { | |
3712 | mc = spa_preferred_class(spa, zio->io_size, | |
3713 | zio->io_prop.zp_type, zio->io_prop.zp_level, | |
3714 | zio->io_prop.zp_zpl_smallblk); | |
3715 | zio->io_metaslab_class = mc; | |
3716 | } | |
3717 | ||
aa755b35 MA |
3718 | /* |
3719 | * Try allocating the block in the usual metaslab class. | |
3720 | * If that's full, allocate it in the normal class. | |
3721 | * If that's full, allocate as a gang block, | |
3722 | * and if all are full, the allocation fails (which shouldn't happen). | |
3723 | * | |
3724 | * Note that we do not fall back on embedded slog (ZIL) space, to | |
3725 | * preserve unfragmented slog space, which is critical for decent | |
3726 | * sync write performance. If a log allocation fails, we will fall | |
3727 | * back to spa_sync() which is abysmal for performance. | |
3728 | */ | |
3bd4df38 | 3729 | ASSERT(ZIO_HAS_ALLOCATOR(zio)); |
b128c09f | 3730 | error = metaslab_alloc(spa, mc, zio->io_size, bp, |
4e21fd06 | 3731 | zio->io_prop.zp_copies, zio->io_txg, NULL, flags, |
492f64e9 | 3732 | &zio->io_alloc_list, zio, zio->io_allocator); |
34dc7c2f | 3733 | |
cc99f275 DB |
3734 | /* |
3735 | * Fallback to normal class when an alloc class is full | |
3736 | */ | |
3737 | if (error == ENOSPC && mc != spa_normal_class(spa)) { | |
3738 | /* | |
3739 | * If throttling, transfer reservation over to normal class. | |
3740 | * The io_allocator slot can remain the same even though we | |
3741 | * are switching classes. | |
3742 | */ | |
3743 | if (mc->mc_alloc_throttle_enabled && | |
3744 | (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) { | |
3745 | metaslab_class_throttle_unreserve(mc, | |
3746 | zio->io_prop.zp_copies, zio->io_allocator, zio); | |
3747 | zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING; | |
3748 | ||
aa755b35 MA |
3749 | VERIFY(metaslab_class_throttle_reserve( |
3750 | spa_normal_class(spa), | |
cc99f275 DB |
3751 | zio->io_prop.zp_copies, zio->io_allocator, zio, |
3752 | flags | METASLAB_MUST_RESERVE)); | |
cc99f275 | 3753 | } |
aa755b35 MA |
3754 | zio->io_metaslab_class = mc = spa_normal_class(spa); |
3755 | if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) { | |
3756 | zfs_dbgmsg("%s: metaslab allocation failure, " | |
3757 | "trying normal class: zio %px, size %llu, error %d", | |
8e739b2c RE |
3758 | spa_name(spa), zio, (u_longlong_t)zio->io_size, |
3759 | error); | |
aa755b35 | 3760 | } |
cc99f275 DB |
3761 | |
3762 | error = metaslab_alloc(spa, mc, zio->io_size, bp, | |
3763 | zio->io_prop.zp_copies, zio->io_txg, NULL, flags, | |
3764 | &zio->io_alloc_list, zio, zio->io_allocator); | |
3765 | } | |
3766 | ||
aa755b35 MA |
3767 | if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) { |
3768 | if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) { | |
3769 | zfs_dbgmsg("%s: metaslab allocation failure, " | |
3770 | "trying ganging: zio %px, size %llu, error %d", | |
8e739b2c RE |
3771 | spa_name(spa), zio, (u_longlong_t)zio->io_size, |
3772 | error); | |
aa755b35 MA |
3773 | } |
3774 | return (zio_write_gang_block(zio, mc)); | |
3775 | } | |
3dfb57a3 | 3776 | if (error != 0) { |
aa755b35 MA |
3777 | if (error != ENOSPC || |
3778 | (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC)) { | |
3779 | zfs_dbgmsg("%s: metaslab allocation failure: zio %px, " | |
3780 | "size %llu, error %d", | |
8e739b2c RE |
3781 | spa_name(spa), zio, (u_longlong_t)zio->io_size, |
3782 | error); | |
aa755b35 | 3783 | } |
34dc7c2f BB |
3784 | zio->io_error = error; |
3785 | } | |
3786 | ||
62840030 | 3787 | return (zio); |
34dc7c2f BB |
3788 | } |
3789 | ||
62840030 | 3790 | static zio_t * |
34dc7c2f BB |
3791 | zio_dva_free(zio_t *zio) |
3792 | { | |
b128c09f | 3793 | metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); |
34dc7c2f | 3794 | |
62840030 | 3795 | return (zio); |
34dc7c2f BB |
3796 | } |
3797 | ||
62840030 | 3798 | static zio_t * |
34dc7c2f BB |
3799 | zio_dva_claim(zio_t *zio) |
3800 | { | |
b128c09f BB |
3801 | int error; |
3802 | ||
3803 | error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); | |
3804 | if (error) | |
3805 | zio->io_error = error; | |
34dc7c2f | 3806 | |
62840030 | 3807 | return (zio); |
34dc7c2f BB |
3808 | } |
3809 | ||
b128c09f BB |
3810 | /* |
3811 | * Undo an allocation. This is used by zio_done() when an I/O fails | |
3812 | * and we want to give back the block we just allocated. | |
3813 | * This handles both normal blocks and gang blocks. | |
3814 | */ | |
3815 | static void | |
3816 | zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) | |
3817 | { | |
493fcce9 | 3818 | ASSERT(BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg || BP_IS_HOLE(bp)); |
428870ff | 3819 | ASSERT(zio->io_bp_override == NULL); |
b128c09f | 3820 | |
493fcce9 GW |
3821 | if (!BP_IS_HOLE(bp)) { |
3822 | metaslab_free(zio->io_spa, bp, BP_GET_LOGICAL_BIRTH(bp), | |
3823 | B_TRUE); | |
3824 | } | |
b128c09f BB |
3825 | |
3826 | if (gn != NULL) { | |
1c27024e | 3827 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
3828 | zio_dva_unallocate(zio, gn->gn_child[g], |
3829 | &gn->gn_gbh->zg_blkptr[g]); | |
3830 | } | |
3831 | } | |
3832 | } | |
3833 | ||
3834 | /* | |
3835 | * Try to allocate an intent log block. Return 0 on success, errno on failure. | |
3836 | */ | |
3837 | int | |
b5256303 TC |
3838 | zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp, |
3839 | uint64_t size, boolean_t *slog) | |
b128c09f | 3840 | { |
428870ff | 3841 | int error = 1; |
4e21fd06 | 3842 | zio_alloc_list_t io_alloc_list; |
b128c09f | 3843 | |
428870ff BB |
3844 | ASSERT(txg > spa_syncing_txg(spa)); |
3845 | ||
4e21fd06 | 3846 | metaslab_trace_init(&io_alloc_list); |
cc99f275 DB |
3847 | |
3848 | /* | |
3849 | * Block pointer fields are useful to metaslabs for stats and debugging. | |
3850 | * Fill in the obvious ones before calling into metaslab_alloc(). | |
3851 | */ | |
3852 | BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); | |
3853 | BP_SET_PSIZE(new_bp, size); | |
3854 | BP_SET_LEVEL(new_bp, 0); | |
3855 | ||
492f64e9 PD |
3856 | /* |
3857 | * When allocating a zil block, we don't have information about | |
3858 | * the final destination of the block except the objset it's part | |
3859 | * of, so we just hash the objset ID to pick the allocator to get | |
3860 | * some parallelism. | |
3861 | */ | |
b22bab25 | 3862 | int flags = METASLAB_ZIL; |
1b50749c AM |
3863 | int allocator = (uint_t)cityhash4(0, 0, 0, |
3864 | os->os_dsl_dataset->ds_object) % spa->spa_alloc_count; | |
aa755b35 MA |
3865 | error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1, |
3866 | txg, NULL, flags, &io_alloc_list, NULL, allocator); | |
3867 | *slog = (error == 0); | |
3868 | if (error != 0) { | |
3869 | error = metaslab_alloc(spa, spa_embedded_log_class(spa), size, | |
3870 | new_bp, 1, txg, NULL, flags, | |
3871 | &io_alloc_list, NULL, allocator); | |
3872 | } | |
3873 | if (error != 0) { | |
3874 | error = metaslab_alloc(spa, spa_normal_class(spa), size, | |
3875 | new_bp, 1, txg, NULL, flags, | |
3876 | &io_alloc_list, NULL, allocator); | |
ebf8e3a2 | 3877 | } |
4e21fd06 | 3878 | metaslab_trace_fini(&io_alloc_list); |
b128c09f BB |
3879 | |
3880 | if (error == 0) { | |
3881 | BP_SET_LSIZE(new_bp, size); | |
3882 | BP_SET_PSIZE(new_bp, size); | |
3883 | BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); | |
428870ff BB |
3884 | BP_SET_CHECKSUM(new_bp, |
3885 | spa_version(spa) >= SPA_VERSION_SLIM_ZIL | |
3886 | ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); | |
b128c09f BB |
3887 | BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); |
3888 | BP_SET_LEVEL(new_bp, 0); | |
428870ff | 3889 | BP_SET_DEDUP(new_bp, 0); |
b128c09f | 3890 | BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); |
b5256303 TC |
3891 | |
3892 | /* | |
3893 | * encrypted blocks will require an IV and salt. We generate | |
3894 | * these now since we will not be rewriting the bp at | |
3895 | * rewrite time. | |
3896 | */ | |
3897 | if (os->os_encrypted) { | |
3898 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
3899 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
3900 | ||
3901 | BP_SET_CRYPT(new_bp, B_TRUE); | |
3902 | VERIFY0(spa_crypt_get_salt(spa, | |
3903 | dmu_objset_id(os), salt)); | |
3904 | VERIFY0(zio_crypt_generate_iv(iv)); | |
3905 | ||
3906 | zio_crypt_encode_params_bp(new_bp, salt, iv); | |
3907 | } | |
1ce23dca PS |
3908 | } else { |
3909 | zfs_dbgmsg("%s: zil block allocation failure: " | |
8e739b2c RE |
3910 | "size %llu, error %d", spa_name(spa), (u_longlong_t)size, |
3911 | error); | |
b128c09f BB |
3912 | } |
3913 | ||
3914 | return (error); | |
3915 | } | |
3916 | ||
34dc7c2f BB |
3917 | /* |
3918 | * ========================================================================== | |
3919 | * Read and write to physical devices | |
3920 | * ========================================================================== | |
3921 | */ | |
98b25418 | 3922 | |
98b25418 GW |
3923 | /* |
3924 | * Issue an I/O to the underlying vdev. Typically the issue pipeline | |
3925 | * stops after this stage and will resume upon I/O completion. | |
3926 | * However, there are instances where the vdev layer may need to | |
3927 | * continue the pipeline when an I/O was not issued. Since the I/O | |
3928 | * that was sent to the vdev layer might be different than the one | |
3929 | * currently active in the pipeline (see vdev_queue_io()), we explicitly | |
3930 | * force the underlying vdev layers to call either zio_execute() or | |
3931 | * zio_interrupt() to ensure that the pipeline continues with the correct I/O. | |
3932 | */ | |
62840030 | 3933 | static zio_t * |
34dc7c2f BB |
3934 | zio_vdev_io_start(zio_t *zio) |
3935 | { | |
3936 | vdev_t *vd = zio->io_vd; | |
34dc7c2f BB |
3937 | uint64_t align; |
3938 | spa_t *spa = zio->io_spa; | |
3939 | ||
193a37cb TH |
3940 | zio->io_delay = 0; |
3941 | ||
b128c09f BB |
3942 | ASSERT(zio->io_error == 0); |
3943 | ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); | |
34dc7c2f | 3944 | |
b128c09f BB |
3945 | if (vd == NULL) { |
3946 | if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
3947 | spa_config_enter(spa, SCL_ZIO, zio, RW_READER); | |
34dc7c2f | 3948 | |
b128c09f BB |
3949 | /* |
3950 | * The mirror_ops handle multiple DVAs in a single BP. | |
3951 | */ | |
98b25418 | 3952 | vdev_mirror_ops.vdev_op_io_start(zio); |
62840030 | 3953 | return (NULL); |
34dc7c2f BB |
3954 | } |
3955 | ||
3dfb57a3 | 3956 | ASSERT3P(zio->io_logical, !=, zio); |
6cb8e530 PZ |
3957 | if (zio->io_type == ZIO_TYPE_WRITE) { |
3958 | ASSERT(spa->spa_trust_config); | |
3959 | ||
a1d477c2 MA |
3960 | /* |
3961 | * Note: the code can handle other kinds of writes, | |
3962 | * but we don't expect them. | |
3963 | */ | |
2a673e76 | 3964 | if (zio->io_vd->vdev_noalloc) { |
6cb8e530 PZ |
3965 | ASSERT(zio->io_flags & |
3966 | (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL | | |
3967 | ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE)); | |
3968 | } | |
a1d477c2 | 3969 | } |
3dfb57a3 | 3970 | |
b128c09f BB |
3971 | align = 1ULL << vd->vdev_top->vdev_ashift; |
3972 | ||
b02fe35d AR |
3973 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && |
3974 | P2PHASE(zio->io_size, align) != 0) { | |
3975 | /* Transform logical writes to be a full physical block size. */ | |
34dc7c2f | 3976 | uint64_t asize = P2ROUNDUP(zio->io_size, align); |
a6255b7f | 3977 | abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize); |
178e73b3 | 3978 | ASSERT(vd == vd->vdev_top); |
34dc7c2f | 3979 | if (zio->io_type == ZIO_TYPE_WRITE) { |
a6255b7f DQ |
3980 | abd_copy(abuf, zio->io_abd, zio->io_size); |
3981 | abd_zero_off(abuf, zio->io_size, asize - zio->io_size); | |
34dc7c2f | 3982 | } |
b128c09f | 3983 | zio_push_transform(zio, abuf, asize, asize, zio_subblock); |
34dc7c2f BB |
3984 | } |
3985 | ||
b02fe35d AR |
3986 | /* |
3987 | * If this is not a physical io, make sure that it is properly aligned | |
3988 | * before proceeding. | |
3989 | */ | |
3990 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { | |
3991 | ASSERT0(P2PHASE(zio->io_offset, align)); | |
3992 | ASSERT0(P2PHASE(zio->io_size, align)); | |
3993 | } else { | |
3994 | /* | |
3995 | * For physical writes, we allow 512b aligned writes and assume | |
3996 | * the device will perform a read-modify-write as necessary. | |
3997 | */ | |
3998 | ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); | |
3999 | ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); | |
4000 | } | |
4001 | ||
572e2857 | 4002 | VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); |
fb5f0bc8 BB |
4003 | |
4004 | /* | |
4005 | * If this is a repair I/O, and there's no self-healing involved -- | |
4006 | * that is, we're just resilvering what we expect to resilver -- | |
4007 | * then don't do the I/O unless zio's txg is actually in vd's DTL. | |
9e052db4 MA |
4008 | * This prevents spurious resilvering. |
4009 | * | |
4010 | * There are a few ways that we can end up creating these spurious | |
4011 | * resilver i/os: | |
4012 | * | |
4013 | * 1. A resilver i/o will be issued if any DVA in the BP has a | |
4014 | * dirty DTL. The mirror code will issue resilver writes to | |
4015 | * each DVA, including the one(s) that are not on vdevs with dirty | |
4016 | * DTLs. | |
4017 | * | |
4018 | * 2. With nested replication, which happens when we have a | |
4019 | * "replacing" or "spare" vdev that's a child of a mirror or raidz. | |
4020 | * For example, given mirror(replacing(A+B), C), it's likely that | |
4021 | * only A is out of date (it's the new device). In this case, we'll | |
4022 | * read from C, then use the data to resilver A+B -- but we don't | |
4023 | * actually want to resilver B, just A. The top-level mirror has no | |
4024 | * way to know this, so instead we just discard unnecessary repairs | |
4025 | * as we work our way down the vdev tree. | |
4026 | * | |
4027 | * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc. | |
4028 | * The same logic applies to any form of nested replication: ditto | |
4029 | * + mirror, RAID-Z + replacing, etc. | |
4030 | * | |
4031 | * However, indirect vdevs point off to other vdevs which may have | |
4032 | * DTL's, so we never bypass them. The child i/os on concrete vdevs | |
4033 | * will be properly bypassed instead. | |
b2255edc BB |
4034 | * |
4035 | * Leaf DTL_PARTIAL can be empty when a legitimate write comes from | |
4036 | * a dRAID spare vdev. For example, when a dRAID spare is first | |
4037 | * used, its spare blocks need to be written to but the leaf vdev's | |
4038 | * of such blocks can have empty DTL_PARTIAL. | |
4039 | * | |
4040 | * There seemed no clean way to allow such writes while bypassing | |
4041 | * spurious ones. At this point, just avoid all bypassing for dRAID | |
4042 | * for correctness. | |
fb5f0bc8 BB |
4043 | */ |
4044 | if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && | |
4045 | !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && | |
4046 | zio->io_txg != 0 && /* not a delegated i/o */ | |
9e052db4 | 4047 | vd->vdev_ops != &vdev_indirect_ops && |
b2255edc | 4048 | vd->vdev_top->vdev_ops != &vdev_draid_ops && |
fb5f0bc8 BB |
4049 | !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { |
4050 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
fb5f0bc8 | 4051 | zio_vdev_io_bypass(zio); |
62840030 | 4052 | return (zio); |
fb5f0bc8 | 4053 | } |
34dc7c2f | 4054 | |
b2255edc BB |
4055 | /* |
4056 | * Select the next best leaf I/O to process. Distributed spares are | |
4057 | * excluded since they dispatch the I/O directly to a leaf vdev after | |
4058 | * applying the dRAID mapping. | |
4059 | */ | |
4060 | if (vd->vdev_ops->vdev_op_leaf && | |
4061 | vd->vdev_ops != &vdev_draid_spare_ops && | |
4062 | (zio->io_type == ZIO_TYPE_READ || | |
4063 | zio->io_type == ZIO_TYPE_WRITE || | |
4064 | zio->io_type == ZIO_TYPE_TRIM)) { | |
b128c09f | 4065 | |
4725e543 RN |
4066 | if (zio_handle_device_injection(vd, zio, ENOSYS) != 0) { |
4067 | /* | |
4068 | * "no-op" injections return success, but do no actual | |
4069 | * work. Just skip the remaining vdev stages. | |
4070 | */ | |
4071 | zio_vdev_io_bypass(zio); | |
4072 | zio_interrupt(zio); | |
4073 | return (NULL); | |
4074 | } | |
4075 | ||
b128c09f | 4076 | if ((zio = vdev_queue_io(zio)) == NULL) |
62840030 | 4077 | return (NULL); |
b128c09f BB |
4078 | |
4079 | if (!vdev_accessible(vd, zio)) { | |
2e528b49 | 4080 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f | 4081 | zio_interrupt(zio); |
62840030 | 4082 | return (NULL); |
b128c09f | 4083 | } |
67103816 | 4084 | zio->io_delay = gethrtime(); |
b128c09f BB |
4085 | } |
4086 | ||
98b25418 | 4087 | vd->vdev_ops->vdev_op_io_start(zio); |
62840030 | 4088 | return (NULL); |
34dc7c2f BB |
4089 | } |
4090 | ||
62840030 | 4091 | static zio_t * |
34dc7c2f BB |
4092 | zio_vdev_io_done(zio_t *zio) |
4093 | { | |
b128c09f BB |
4094 | vdev_t *vd = zio->io_vd; |
4095 | vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; | |
4096 | boolean_t unexpected_error = B_FALSE; | |
34dc7c2f | 4097 | |
ddc751d5 | 4098 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { |
62840030 | 4099 | return (NULL); |
ddc751d5 | 4100 | } |
34dc7c2f | 4101 | |
1b939560 | 4102 | ASSERT(zio->io_type == ZIO_TYPE_READ || |
76d1dde9 | 4103 | zio->io_type == ZIO_TYPE_WRITE || |
d7605ae7 | 4104 | zio->io_type == ZIO_TYPE_FLUSH || |
76d1dde9 | 4105 | zio->io_type == ZIO_TYPE_TRIM); |
b128c09f | 4106 | |
193a37cb TH |
4107 | if (zio->io_delay) |
4108 | zio->io_delay = gethrtime() - zio->io_delay; | |
4109 | ||
b2255edc BB |
4110 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf && |
4111 | vd->vdev_ops != &vdev_draid_spare_ops) { | |
d7605ae7 | 4112 | if (zio->io_type != ZIO_TYPE_FLUSH) |
76d1dde9 | 4113 | vdev_queue_io_done(zio); |
b128c09f | 4114 | |
b128c09f | 4115 | if (zio_injection_enabled && zio->io_error == 0) |
d977122d DB |
4116 | zio->io_error = zio_handle_device_injections(vd, zio, |
4117 | EIO, EILSEQ); | |
b128c09f BB |
4118 | |
4119 | if (zio_injection_enabled && zio->io_error == 0) | |
4120 | zio->io_error = zio_handle_label_injection(zio, EIO); | |
4121 | ||
9f83eec0 AM |
4122 | if (zio->io_error && zio->io_type != ZIO_TYPE_FLUSH && |
4123 | zio->io_type != ZIO_TYPE_TRIM) { | |
b128c09f | 4124 | if (!vdev_accessible(vd, zio)) { |
2e528b49 | 4125 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
4126 | } else { |
4127 | unexpected_error = B_TRUE; | |
4128 | } | |
4129 | } | |
4130 | } | |
4131 | ||
4132 | ops->vdev_op_io_done(zio); | |
34dc7c2f | 4133 | |
55c12724 | 4134 | if (unexpected_error && vd->vdev_remove_wanted == B_FALSE) |
d164b209 | 4135 | VERIFY(vdev_probe(vd, zio) == NULL); |
34dc7c2f | 4136 | |
62840030 | 4137 | return (zio); |
34dc7c2f BB |
4138 | } |
4139 | ||
a8b2e306 TC |
4140 | /* |
4141 | * This function is used to change the priority of an existing zio that is | |
4142 | * currently in-flight. This is used by the arc to upgrade priority in the | |
4143 | * event that a demand read is made for a block that is currently queued | |
4144 | * as a scrub or async read IO. Otherwise, the high priority read request | |
4145 | * would end up having to wait for the lower priority IO. | |
4146 | */ | |
4147 | void | |
4148 | zio_change_priority(zio_t *pio, zio_priority_t priority) | |
4149 | { | |
4150 | zio_t *cio, *cio_next; | |
4151 | zio_link_t *zl = NULL; | |
4152 | ||
4153 | ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
4154 | ||
4155 | if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) { | |
4156 | vdev_queue_change_io_priority(pio, priority); | |
4157 | } else { | |
4158 | pio->io_priority = priority; | |
4159 | } | |
4160 | ||
4161 | mutex_enter(&pio->io_lock); | |
4162 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { | |
4163 | cio_next = zio_walk_children(pio, &zl); | |
4164 | zio_change_priority(cio, priority); | |
4165 | } | |
4166 | mutex_exit(&pio->io_lock); | |
4167 | } | |
4168 | ||
428870ff BB |
4169 | /* |
4170 | * For non-raidz ZIOs, we can just copy aside the bad data read from the | |
4171 | * disk, and use that to finish the checksum ereport later. | |
4172 | */ | |
4173 | static void | |
4174 | zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, | |
84c07ada | 4175 | const abd_t *good_buf) |
428870ff BB |
4176 | { |
4177 | /* no processing needed */ | |
4178 | zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); | |
4179 | } | |
4180 | ||
428870ff | 4181 | void |
330c6c05 | 4182 | zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr) |
428870ff | 4183 | { |
84c07ada | 4184 | void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size); |
428870ff | 4185 | |
84c07ada | 4186 | abd_copy(abd, zio->io_abd, zio->io_size); |
428870ff BB |
4187 | |
4188 | zcr->zcr_cbinfo = zio->io_size; | |
84c07ada | 4189 | zcr->zcr_cbdata = abd; |
428870ff | 4190 | zcr->zcr_finish = zio_vsd_default_cksum_finish; |
84c07ada | 4191 | zcr->zcr_free = zio_abd_free; |
428870ff BB |
4192 | } |
4193 | ||
62840030 | 4194 | static zio_t * |
34dc7c2f BB |
4195 | zio_vdev_io_assess(zio_t *zio) |
4196 | { | |
4197 | vdev_t *vd = zio->io_vd; | |
b128c09f | 4198 | |
ddc751d5 | 4199 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { |
62840030 | 4200 | return (NULL); |
ddc751d5 | 4201 | } |
b128c09f BB |
4202 | |
4203 | if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
4204 | spa_config_exit(zio->io_spa, SCL_ZIO, zio); | |
4205 | ||
4206 | if (zio->io_vsd != NULL) { | |
428870ff | 4207 | zio->io_vsd_ops->vsd_free(zio); |
b128c09f | 4208 | zio->io_vsd = NULL; |
34dc7c2f BB |
4209 | } |
4210 | ||
b128c09f | 4211 | if (zio_injection_enabled && zio->io_error == 0) |
34dc7c2f BB |
4212 | zio->io_error = zio_handle_fault_injection(zio, EIO); |
4213 | ||
4214 | /* | |
4215 | * If the I/O failed, determine whether we should attempt to retry it. | |
428870ff BB |
4216 | * |
4217 | * On retry, we cut in line in the issue queue, since we don't want | |
4218 | * compression/checksumming/etc. work to prevent our (cheap) IO reissue. | |
34dc7c2f | 4219 | */ |
b128c09f BB |
4220 | if (zio->io_error && vd == NULL && |
4221 | !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { | |
4222 | ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ | |
4223 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ | |
34dc7c2f | 4224 | zio->io_error = 0; |
70ea484e | 4225 | zio->io_flags |= ZIO_FLAG_IO_RETRY | ZIO_FLAG_DONT_AGGREGATE; |
428870ff BB |
4226 | zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; |
4227 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, | |
4228 | zio_requeue_io_start_cut_in_line); | |
62840030 | 4229 | return (NULL); |
34dc7c2f BB |
4230 | } |
4231 | ||
b128c09f BB |
4232 | /* |
4233 | * If we got an error on a leaf device, convert it to ENXIO | |
4234 | * if the device is not accessible at all. | |
4235 | */ | |
4236 | if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && | |
4237 | !vdev_accessible(vd, zio)) | |
2e528b49 | 4238 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
4239 | |
4240 | /* | |
4241 | * If we can't write to an interior vdev (mirror or RAID-Z), | |
4242 | * set vdev_cant_write so that we stop trying to allocate from it. | |
4243 | */ | |
4244 | if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && | |
13fe0198 | 4245 | vd != NULL && !vd->vdev_ops->vdev_op_leaf) { |
2b56a634 MA |
4246 | vdev_dbgmsg(vd, "zio_vdev_io_assess(zio=%px) setting " |
4247 | "cant_write=TRUE due to write failure with ENXIO", | |
4248 | zio); | |
b128c09f | 4249 | vd->vdev_cant_write = B_TRUE; |
13fe0198 | 4250 | } |
b128c09f | 4251 | |
298ec40b GM |
4252 | /* |
4253 | * If a cache flush returns ENOTSUP or ENOTTY, we know that no future | |
1b939560 BB |
4254 | * attempts will ever succeed. In this case we set a persistent |
4255 | * boolean flag so that we don't bother with it in the future. | |
298ec40b GM |
4256 | */ |
4257 | if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) && | |
d7605ae7 | 4258 | zio->io_type == ZIO_TYPE_FLUSH && vd != NULL) |
298ec40b GM |
4259 | vd->vdev_nowritecache = B_TRUE; |
4260 | ||
b128c09f BB |
4261 | if (zio->io_error) |
4262 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
4263 | ||
62840030 | 4264 | return (zio); |
34dc7c2f BB |
4265 | } |
4266 | ||
4267 | void | |
4268 | zio_vdev_io_reissue(zio_t *zio) | |
4269 | { | |
4270 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
4271 | ASSERT(zio->io_error == 0); | |
4272 | ||
428870ff | 4273 | zio->io_stage >>= 1; |
34dc7c2f BB |
4274 | } |
4275 | ||
4276 | void | |
4277 | zio_vdev_io_redone(zio_t *zio) | |
4278 | { | |
4279 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); | |
4280 | ||
428870ff | 4281 | zio->io_stage >>= 1; |
34dc7c2f BB |
4282 | } |
4283 | ||
4284 | void | |
4285 | zio_vdev_io_bypass(zio_t *zio) | |
4286 | { | |
4287 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
4288 | ASSERT(zio->io_error == 0); | |
4289 | ||
4290 | zio->io_flags |= ZIO_FLAG_IO_BYPASS; | |
428870ff | 4291 | zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; |
34dc7c2f BB |
4292 | } |
4293 | ||
b5256303 TC |
4294 | /* |
4295 | * ========================================================================== | |
4296 | * Encrypt and store encryption parameters | |
4297 | * ========================================================================== | |
4298 | */ | |
4299 | ||
4300 | ||
4301 | /* | |
4302 | * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for | |
4303 | * managing the storage of encryption parameters and passing them to the | |
4304 | * lower-level encryption functions. | |
4305 | */ | |
62840030 | 4306 | static zio_t * |
b5256303 TC |
4307 | zio_encrypt(zio_t *zio) |
4308 | { | |
4309 | zio_prop_t *zp = &zio->io_prop; | |
4310 | spa_t *spa = zio->io_spa; | |
4311 | blkptr_t *bp = zio->io_bp; | |
4312 | uint64_t psize = BP_GET_PSIZE(bp); | |
ae76f45c | 4313 | uint64_t dsobj = zio->io_bookmark.zb_objset; |
b5256303 TC |
4314 | dmu_object_type_t ot = BP_GET_TYPE(bp); |
4315 | void *enc_buf = NULL; | |
4316 | abd_t *eabd = NULL; | |
4317 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
4318 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
4319 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
4320 | boolean_t no_crypt = B_FALSE; | |
4321 | ||
4322 | /* the root zio already encrypted the data */ | |
4323 | if (zio->io_child_type == ZIO_CHILD_GANG) | |
62840030 | 4324 | return (zio); |
b5256303 TC |
4325 | |
4326 | /* only ZIL blocks are re-encrypted on rewrite */ | |
4327 | if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG) | |
62840030 | 4328 | return (zio); |
b5256303 TC |
4329 | |
4330 | if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) { | |
4331 | BP_SET_CRYPT(bp, B_FALSE); | |
62840030 | 4332 | return (zio); |
b5256303 TC |
4333 | } |
4334 | ||
4335 | /* if we are doing raw encryption set the provided encryption params */ | |
4336 | if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) { | |
ae76f45c | 4337 | ASSERT0(BP_GET_LEVEL(bp)); |
b5256303 TC |
4338 | BP_SET_CRYPT(bp, B_TRUE); |
4339 | BP_SET_BYTEORDER(bp, zp->zp_byteorder); | |
4340 | if (ot != DMU_OT_OBJSET) | |
4341 | zio_crypt_encode_mac_bp(bp, zp->zp_mac); | |
ae76f45c TC |
4342 | |
4343 | /* dnode blocks must be written out in the provided byteorder */ | |
4344 | if (zp->zp_byteorder != ZFS_HOST_BYTEORDER && | |
4345 | ot == DMU_OT_DNODE) { | |
4346 | void *bswap_buf = zio_buf_alloc(psize); | |
4347 | abd_t *babd = abd_get_from_buf(bswap_buf, psize); | |
4348 | ||
4349 | ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); | |
4350 | abd_copy_to_buf(bswap_buf, zio->io_abd, psize); | |
4351 | dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf, | |
4352 | psize); | |
4353 | ||
4354 | abd_take_ownership_of_buf(babd, B_TRUE); | |
4355 | zio_push_transform(zio, babd, psize, psize, NULL); | |
4356 | } | |
4357 | ||
b5256303 TC |
4358 | if (DMU_OT_IS_ENCRYPTED(ot)) |
4359 | zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv); | |
62840030 | 4360 | return (zio); |
b5256303 TC |
4361 | } |
4362 | ||
4363 | /* indirect blocks only maintain a cksum of the lower level MACs */ | |
4364 | if (BP_GET_LEVEL(bp) > 0) { | |
4365 | BP_SET_CRYPT(bp, B_TRUE); | |
4366 | VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE, | |
4367 | zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp), | |
4368 | mac)); | |
4369 | zio_crypt_encode_mac_bp(bp, mac); | |
62840030 | 4370 | return (zio); |
b5256303 TC |
4371 | } |
4372 | ||
4373 | /* | |
4374 | * Objset blocks are a special case since they have 2 256-bit MACs | |
4375 | * embedded within them. | |
4376 | */ | |
4377 | if (ot == DMU_OT_OBJSET) { | |
4378 | ASSERT0(DMU_OT_IS_ENCRYPTED(ot)); | |
4379 | ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); | |
4380 | BP_SET_CRYPT(bp, B_TRUE); | |
ae76f45c TC |
4381 | VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj, |
4382 | zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp))); | |
62840030 | 4383 | return (zio); |
b5256303 TC |
4384 | } |
4385 | ||
4386 | /* unencrypted object types are only authenticated with a MAC */ | |
4387 | if (!DMU_OT_IS_ENCRYPTED(ot)) { | |
4388 | BP_SET_CRYPT(bp, B_TRUE); | |
ae76f45c TC |
4389 | VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj, |
4390 | zio->io_abd, psize, mac)); | |
b5256303 | 4391 | zio_crypt_encode_mac_bp(bp, mac); |
62840030 | 4392 | return (zio); |
b5256303 TC |
4393 | } |
4394 | ||
4395 | /* | |
4396 | * Later passes of sync-to-convergence may decide to rewrite data | |
4397 | * in place to avoid more disk reallocations. This presents a problem | |
d611989f | 4398 | * for encryption because this constitutes rewriting the new data with |
b5256303 TC |
4399 | * the same encryption key and IV. However, this only applies to blocks |
4400 | * in the MOS (particularly the spacemaps) and we do not encrypt the | |
4401 | * MOS. We assert that the zio is allocating or an intent log write | |
4402 | * to enforce this. | |
4403 | */ | |
4404 | ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG); | |
4405 | ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG); | |
4406 | ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION)); | |
4407 | ASSERT3U(psize, !=, 0); | |
4408 | ||
4409 | enc_buf = zio_buf_alloc(psize); | |
4410 | eabd = abd_get_from_buf(enc_buf, psize); | |
4411 | abd_take_ownership_of_buf(eabd, B_TRUE); | |
4412 | ||
4413 | /* | |
4414 | * For an explanation of what encryption parameters are stored | |
4415 | * where, see the block comment in zio_crypt.c. | |
4416 | */ | |
4417 | if (ot == DMU_OT_INTENT_LOG) { | |
4418 | zio_crypt_decode_params_bp(bp, salt, iv); | |
4419 | } else { | |
4420 | BP_SET_CRYPT(bp, B_TRUE); | |
4421 | } | |
4422 | ||
4423 | /* Perform the encryption. This should not fail */ | |
be9a5c35 TC |
4424 | VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark, |
4425 | BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), | |
4426 | salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt)); | |
b5256303 TC |
4427 | |
4428 | /* encode encryption metadata into the bp */ | |
4429 | if (ot == DMU_OT_INTENT_LOG) { | |
4430 | /* | |
4431 | * ZIL blocks store the MAC in the embedded checksum, so the | |
4432 | * transform must always be applied. | |
4433 | */ | |
4434 | zio_crypt_encode_mac_zil(enc_buf, mac); | |
4435 | zio_push_transform(zio, eabd, psize, psize, NULL); | |
4436 | } else { | |
4437 | BP_SET_CRYPT(bp, B_TRUE); | |
4438 | zio_crypt_encode_params_bp(bp, salt, iv); | |
4439 | zio_crypt_encode_mac_bp(bp, mac); | |
4440 | ||
4441 | if (no_crypt) { | |
4442 | ASSERT3U(ot, ==, DMU_OT_DNODE); | |
4443 | abd_free(eabd); | |
4444 | } else { | |
4445 | zio_push_transform(zio, eabd, psize, psize, NULL); | |
4446 | } | |
4447 | } | |
4448 | ||
62840030 | 4449 | return (zio); |
b5256303 TC |
4450 | } |
4451 | ||
34dc7c2f BB |
4452 | /* |
4453 | * ========================================================================== | |
4454 | * Generate and verify checksums | |
4455 | * ========================================================================== | |
4456 | */ | |
62840030 | 4457 | static zio_t * |
34dc7c2f BB |
4458 | zio_checksum_generate(zio_t *zio) |
4459 | { | |
34dc7c2f | 4460 | blkptr_t *bp = zio->io_bp; |
b128c09f | 4461 | enum zio_checksum checksum; |
34dc7c2f | 4462 | |
b128c09f BB |
4463 | if (bp == NULL) { |
4464 | /* | |
4465 | * This is zio_write_phys(). | |
4466 | * We're either generating a label checksum, or none at all. | |
4467 | */ | |
4468 | checksum = zio->io_prop.zp_checksum; | |
34dc7c2f | 4469 | |
b128c09f | 4470 | if (checksum == ZIO_CHECKSUM_OFF) |
62840030 | 4471 | return (zio); |
b128c09f BB |
4472 | |
4473 | ASSERT(checksum == ZIO_CHECKSUM_LABEL); | |
4474 | } else { | |
4475 | if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { | |
4476 | ASSERT(!IO_IS_ALLOCATING(zio)); | |
4477 | checksum = ZIO_CHECKSUM_GANG_HEADER; | |
4478 | } else { | |
4479 | checksum = BP_GET_CHECKSUM(bp); | |
4480 | } | |
4481 | } | |
34dc7c2f | 4482 | |
a6255b7f | 4483 | zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size); |
34dc7c2f | 4484 | |
62840030 | 4485 | return (zio); |
34dc7c2f BB |
4486 | } |
4487 | ||
62840030 | 4488 | static zio_t * |
b128c09f | 4489 | zio_checksum_verify(zio_t *zio) |
34dc7c2f | 4490 | { |
428870ff | 4491 | zio_bad_cksum_t info; |
b128c09f BB |
4492 | blkptr_t *bp = zio->io_bp; |
4493 | int error; | |
34dc7c2f | 4494 | |
428870ff BB |
4495 | ASSERT(zio->io_vd != NULL); |
4496 | ||
b128c09f BB |
4497 | if (bp == NULL) { |
4498 | /* | |
4499 | * This is zio_read_phys(). | |
4500 | * We're either verifying a label checksum, or nothing at all. | |
4501 | */ | |
4502 | if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) | |
62840030 | 4503 | return (zio); |
34dc7c2f | 4504 | |
b2255edc | 4505 | ASSERT3U(zio->io_prop.zp_checksum, ==, ZIO_CHECKSUM_LABEL); |
b128c09f | 4506 | } |
34dc7c2f | 4507 | |
428870ff | 4508 | if ((error = zio_checksum_error(zio, &info)) != 0) { |
b128c09f | 4509 | zio->io_error = error; |
7a3066ff MA |
4510 | if (error == ECKSUM && |
4511 | !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { | |
03e02e5b DB |
4512 | mutex_enter(&zio->io_vd->vdev_stat_lock); |
4513 | zio->io_vd->vdev_stat.vs_checksum_errors++; | |
4514 | mutex_exit(&zio->io_vd->vdev_stat_lock); | |
7a75f74c RW |
4515 | (void) zfs_ereport_start_checksum(zio->io_spa, |
4516 | zio->io_vd, &zio->io_bookmark, zio, | |
4517 | zio->io_offset, zio->io_size, &info); | |
b128c09f | 4518 | } |
34dc7c2f BB |
4519 | } |
4520 | ||
62840030 | 4521 | return (zio); |
34dc7c2f BB |
4522 | } |
4523 | ||
4524 | /* | |
4525 | * Called by RAID-Z to ensure we don't compute the checksum twice. | |
4526 | */ | |
4527 | void | |
4528 | zio_checksum_verified(zio_t *zio) | |
4529 | { | |
428870ff | 4530 | zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; |
34dc7c2f BB |
4531 | } |
4532 | ||
4533 | /* | |
b128c09f BB |
4534 | * ========================================================================== |
4535 | * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. | |
9b67f605 | 4536 | * An error of 0 indicates success. ENXIO indicates whole-device failure, |
d611989f | 4537 | * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO |
b128c09f BB |
4538 | * indicate errors that are specific to one I/O, and most likely permanent. |
4539 | * Any other error is presumed to be worse because we weren't expecting it. | |
4540 | * ========================================================================== | |
34dc7c2f | 4541 | */ |
b128c09f BB |
4542 | int |
4543 | zio_worst_error(int e1, int e2) | |
34dc7c2f | 4544 | { |
b128c09f BB |
4545 | static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; |
4546 | int r1, r2; | |
4547 | ||
4548 | for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) | |
4549 | if (e1 == zio_error_rank[r1]) | |
4550 | break; | |
34dc7c2f | 4551 | |
b128c09f BB |
4552 | for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) |
4553 | if (e2 == zio_error_rank[r2]) | |
4554 | break; | |
4555 | ||
4556 | return (r1 > r2 ? e1 : e2); | |
34dc7c2f BB |
4557 | } |
4558 | ||
4559 | /* | |
4560 | * ========================================================================== | |
b128c09f | 4561 | * I/O completion |
34dc7c2f BB |
4562 | * ========================================================================== |
4563 | */ | |
62840030 | 4564 | static zio_t * |
b128c09f | 4565 | zio_ready(zio_t *zio) |
34dc7c2f | 4566 | { |
b128c09f | 4567 | blkptr_t *bp = zio->io_bp; |
d164b209 | 4568 | zio_t *pio, *pio_next; |
3dfb57a3 | 4569 | zio_link_t *zl = NULL; |
34dc7c2f | 4570 | |
eda3fcd5 AM |
4571 | if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT | |
4572 | ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT, ZIO_WAIT_READY)) { | |
62840030 | 4573 | return (NULL); |
ddc751d5 | 4574 | } |
34dc7c2f | 4575 | |
9babb374 | 4576 | if (zio->io_ready) { |
b128c09f | 4577 | ASSERT(IO_IS_ALLOCATING(zio)); |
493fcce9 GW |
4578 | ASSERT(BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg || |
4579 | BP_IS_HOLE(bp) || (zio->io_flags & ZIO_FLAG_NOPWRITE)); | |
b128c09f | 4580 | ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); |
34dc7c2f | 4581 | |
b128c09f BB |
4582 | zio->io_ready(zio); |
4583 | } | |
34dc7c2f | 4584 | |
b4a08730 | 4585 | #ifdef ZFS_DEBUG |
b128c09f BB |
4586 | if (bp != NULL && bp != &zio->io_bp_copy) |
4587 | zio->io_bp_copy = *bp; | |
b4a08730 | 4588 | #endif |
34dc7c2f | 4589 | |
3dfb57a3 | 4590 | if (zio->io_error != 0) { |
b128c09f | 4591 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f | 4592 | |
3dfb57a3 DB |
4593 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
4594 | ASSERT(IO_IS_ALLOCATING(zio)); | |
4595 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
cc99f275 | 4596 | ASSERT(zio->io_metaslab_class != NULL); |
3bd4df38 | 4597 | ASSERT(ZIO_HAS_ALLOCATOR(zio)); |
cc99f275 | 4598 | |
3dfb57a3 DB |
4599 | /* |
4600 | * We were unable to allocate anything, unreserve and | |
4601 | * issue the next I/O to allocate. | |
4602 | */ | |
4603 | metaslab_class_throttle_unreserve( | |
cc99f275 DB |
4604 | zio->io_metaslab_class, zio->io_prop.zp_copies, |
4605 | zio->io_allocator, zio); | |
492f64e9 | 4606 | zio_allocate_dispatch(zio->io_spa, zio->io_allocator); |
3dfb57a3 DB |
4607 | } |
4608 | } | |
4609 | ||
d164b209 BB |
4610 | mutex_enter(&zio->io_lock); |
4611 | zio->io_state[ZIO_WAIT_READY] = 1; | |
3dfb57a3 | 4612 | pio = zio_walk_parents(zio, &zl); |
d164b209 BB |
4613 | mutex_exit(&zio->io_lock); |
4614 | ||
4615 | /* | |
4616 | * As we notify zio's parents, new parents could be added. | |
4617 | * New parents go to the head of zio's io_parent_list, however, | |
4618 | * so we will (correctly) not notify them. The remainder of zio's | |
4619 | * io_parent_list, from 'pio_next' onward, cannot change because | |
4620 | * all parents must wait for us to be done before they can be done. | |
4621 | */ | |
4622 | for (; pio != NULL; pio = pio_next) { | |
3dfb57a3 | 4623 | pio_next = zio_walk_parents(zio, &zl); |
62840030 | 4624 | zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL); |
d164b209 | 4625 | } |
34dc7c2f | 4626 | |
428870ff | 4627 | if (zio->io_flags & ZIO_FLAG_NODATA) { |
7cb67d62 | 4628 | if (bp != NULL && BP_IS_GANG(bp)) { |
428870ff BB |
4629 | zio->io_flags &= ~ZIO_FLAG_NODATA; |
4630 | } else { | |
a6255b7f | 4631 | ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE); |
428870ff BB |
4632 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; |
4633 | } | |
4634 | } | |
4635 | ||
4636 | if (zio_injection_enabled && | |
4637 | zio->io_spa->spa_syncing_txg == zio->io_txg) | |
4638 | zio_handle_ignored_writes(zio); | |
4639 | ||
62840030 | 4640 | return (zio); |
34dc7c2f BB |
4641 | } |
4642 | ||
3dfb57a3 DB |
4643 | /* |
4644 | * Update the allocation throttle accounting. | |
4645 | */ | |
4646 | static void | |
4647 | zio_dva_throttle_done(zio_t *zio) | |
4648 | { | |
2a8ba608 | 4649 | zio_t *lio __maybe_unused = zio->io_logical; |
3dfb57a3 DB |
4650 | zio_t *pio = zio_unique_parent(zio); |
4651 | vdev_t *vd = zio->io_vd; | |
4652 | int flags = METASLAB_ASYNC_ALLOC; | |
4653 | ||
4654 | ASSERT3P(zio->io_bp, !=, NULL); | |
4655 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); | |
4656 | ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE); | |
4657 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); | |
4658 | ASSERT(vd != NULL); | |
4659 | ASSERT3P(vd, ==, vd->vdev_top); | |
21df134f SB |
4660 | ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY)); |
4661 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); | |
3dfb57a3 DB |
4662 | ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING); |
4663 | ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
4664 | ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA)); | |
4665 | ||
4666 | /* | |
4667 | * Parents of gang children can have two flavors -- ones that | |
4668 | * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set) | |
4669 | * and ones that allocated the constituent blocks. The allocation | |
4670 | * throttle needs to know the allocating parent zio so we must find | |
4671 | * it here. | |
4672 | */ | |
4673 | if (pio->io_child_type == ZIO_CHILD_GANG) { | |
4674 | /* | |
4675 | * If our parent is a rewrite gang child then our grandparent | |
4676 | * would have been the one that performed the allocation. | |
4677 | */ | |
4678 | if (pio->io_flags & ZIO_FLAG_IO_REWRITE) | |
4679 | pio = zio_unique_parent(pio); | |
4680 | flags |= METASLAB_GANG_CHILD; | |
4681 | } | |
4682 | ||
4683 | ASSERT(IO_IS_ALLOCATING(pio)); | |
3bd4df38 | 4684 | ASSERT(ZIO_HAS_ALLOCATOR(pio)); |
3dfb57a3 DB |
4685 | ASSERT3P(zio, !=, zio->io_logical); |
4686 | ASSERT(zio->io_logical != NULL); | |
4687 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); | |
4688 | ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE); | |
cc99f275 | 4689 | ASSERT(zio->io_metaslab_class != NULL); |
3dfb57a3 DB |
4690 | |
4691 | mutex_enter(&pio->io_lock); | |
492f64e9 PD |
4692 | metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags, |
4693 | pio->io_allocator, B_TRUE); | |
3dfb57a3 DB |
4694 | mutex_exit(&pio->io_lock); |
4695 | ||
cc99f275 DB |
4696 | metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1, |
4697 | pio->io_allocator, pio); | |
3dfb57a3 DB |
4698 | |
4699 | /* | |
4700 | * Call into the pipeline to see if there is more work that | |
4701 | * needs to be done. If there is work to be done it will be | |
4702 | * dispatched to another taskq thread. | |
4703 | */ | |
492f64e9 | 4704 | zio_allocate_dispatch(zio->io_spa, pio->io_allocator); |
3dfb57a3 DB |
4705 | } |
4706 | ||
62840030 | 4707 | static zio_t * |
b128c09f | 4708 | zio_done(zio_t *zio) |
34dc7c2f | 4709 | { |
3dfb57a3 DB |
4710 | /* |
4711 | * Always attempt to keep stack usage minimal here since | |
d611989f | 4712 | * we can be called recursively up to 19 levels deep. |
3dfb57a3 | 4713 | */ |
84c07ada | 4714 | const uint64_t psize = zio->io_size; |
d164b209 | 4715 | zio_t *pio, *pio_next; |
3dfb57a3 | 4716 | zio_link_t *zl = NULL; |
34dc7c2f | 4717 | |
b128c09f | 4718 | /* |
9babb374 | 4719 | * If our children haven't all completed, |
b128c09f BB |
4720 | * wait for them and then repeat this pipeline stage. |
4721 | */ | |
ddc751d5 | 4722 | if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) { |
62840030 | 4723 | return (NULL); |
ddc751d5 | 4724 | } |
34dc7c2f | 4725 | |
3dfb57a3 DB |
4726 | /* |
4727 | * If the allocation throttle is enabled, then update the accounting. | |
4728 | * We only track child I/Os that are part of an allocating async | |
4729 | * write. We must do this since the allocation is performed | |
4730 | * by the logical I/O but the actual write is done by child I/Os. | |
4731 | */ | |
4732 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING && | |
4733 | zio->io_child_type == ZIO_CHILD_VDEV) { | |
cc99f275 DB |
4734 | ASSERT(zio->io_metaslab_class != NULL); |
4735 | ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled); | |
3dfb57a3 DB |
4736 | zio_dva_throttle_done(zio); |
4737 | } | |
4738 | ||
4739 | /* | |
4740 | * If the allocation throttle is enabled, verify that | |
4741 | * we have decremented the refcounts for every I/O that was throttled. | |
4742 | */ | |
4743 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { | |
4744 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
4745 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
4746 | ASSERT(zio->io_bp != NULL); | |
3bd4df38 | 4747 | ASSERT(ZIO_HAS_ALLOCATOR(zio)); |
cc99f275 | 4748 | |
492f64e9 PD |
4749 | metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio, |
4750 | zio->io_allocator); | |
f8020c93 AM |
4751 | VERIFY(zfs_refcount_not_held(&zio->io_metaslab_class-> |
4752 | mc_allocator[zio->io_allocator].mca_alloc_slots, zio)); | |
3dfb57a3 DB |
4753 | } |
4754 | ||
4755 | ||
1c27024e DB |
4756 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) |
4757 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) | |
b128c09f BB |
4758 | ASSERT(zio->io_children[c][w] == 0); |
4759 | ||
9b67f605 | 4760 | if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) { |
c776b317 BB |
4761 | ASSERT(zio->io_bp->blk_pad[0] == 0); |
4762 | ASSERT(zio->io_bp->blk_pad[1] == 0); | |
861166b0 | 4763 | ASSERT(memcmp(zio->io_bp, &zio->io_bp_copy, |
d1d7e268 | 4764 | sizeof (blkptr_t)) == 0 || |
c776b317 BB |
4765 | (zio->io_bp == zio_unique_parent(zio)->io_bp)); |
4766 | if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) && | |
428870ff | 4767 | zio->io_bp_override == NULL && |
b128c09f | 4768 | !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { |
d1d7e268 MK |
4769 | ASSERT3U(zio->io_prop.zp_copies, <=, |
4770 | BP_GET_NDVAS(zio->io_bp)); | |
c776b317 | 4771 | ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 || |
d1d7e268 MK |
4772 | (BP_COUNT_GANG(zio->io_bp) == |
4773 | BP_GET_NDVAS(zio->io_bp))); | |
b128c09f | 4774 | } |
03c6040b GW |
4775 | if (zio->io_flags & ZIO_FLAG_NOPWRITE) |
4776 | VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig)); | |
b128c09f BB |
4777 | } |
4778 | ||
4779 | /* | |
428870ff | 4780 | * If there were child vdev/gang/ddt errors, they apply to us now. |
b128c09f BB |
4781 | */ |
4782 | zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); | |
4783 | zio_inherit_child_errors(zio, ZIO_CHILD_GANG); | |
428870ff BB |
4784 | zio_inherit_child_errors(zio, ZIO_CHILD_DDT); |
4785 | ||
4786 | /* | |
4787 | * If the I/O on the transformed data was successful, generate any | |
4788 | * checksum reports now while we still have the transformed data. | |
4789 | */ | |
4790 | if (zio->io_error == 0) { | |
4791 | while (zio->io_cksum_report != NULL) { | |
4792 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
4793 | uint64_t align = zcr->zcr_align; | |
a6255b7f | 4794 | uint64_t asize = P2ROUNDUP(psize, align); |
a6255b7f DQ |
4795 | abd_t *adata = zio->io_abd; |
4796 | ||
f2286383 | 4797 | if (adata != NULL && asize != psize) { |
84c07ada | 4798 | adata = abd_alloc(asize, B_TRUE); |
a6255b7f DQ |
4799 | abd_copy(adata, zio->io_abd, psize); |
4800 | abd_zero_off(adata, psize, asize - psize); | |
428870ff BB |
4801 | } |
4802 | ||
4803 | zio->io_cksum_report = zcr->zcr_next; | |
4804 | zcr->zcr_next = NULL; | |
84c07ada | 4805 | zcr->zcr_finish(zcr, adata); |
428870ff BB |
4806 | zfs_ereport_free_checksum(zcr); |
4807 | ||
f2286383 | 4808 | if (adata != NULL && asize != psize) |
a6255b7f | 4809 | abd_free(adata); |
428870ff BB |
4810 | } |
4811 | } | |
b128c09f BB |
4812 | |
4813 | zio_pop_transforms(zio); /* note: may set zio->io_error */ | |
4814 | ||
a6255b7f | 4815 | vdev_stat_update(zio, psize); |
b128c09f | 4816 | |
a69052be | 4817 | /* |
cc92e9d0 | 4818 | * If this I/O is attached to a particular vdev is slow, exceeding |
72f53c56 MJ |
4819 | * 30 seconds to complete, post an error described the I/O delay. |
4820 | * We ignore these errors if the device is currently unavailable. | |
a69052be | 4821 | */ |
ad796b8a TH |
4822 | if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) { |
4823 | if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) { | |
4824 | /* | |
4825 | * We want to only increment our slow IO counters if | |
4826 | * the IO is valid (i.e. not if the drive is removed). | |
4827 | * | |
4828 | * zfs_ereport_post() will also do these checks, but | |
4829 | * it can also ratelimit and have other failures, so we | |
4830 | * need to increment the slow_io counters independent | |
4831 | * of it. | |
4832 | */ | |
4833 | if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY, | |
4834 | zio->io_spa, zio->io_vd, zio)) { | |
4835 | mutex_enter(&zio->io_vd->vdev_stat_lock); | |
4836 | zio->io_vd->vdev_stat.vs_slow_ios++; | |
4837 | mutex_exit(&zio->io_vd->vdev_stat_lock); | |
4838 | ||
1144586b | 4839 | (void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY, |
ad796b8a | 4840 | zio->io_spa, zio->io_vd, &zio->io_bookmark, |
4f072827 | 4841 | zio, 0); |
ad796b8a TH |
4842 | } |
4843 | } | |
72f53c56 | 4844 | } |
a69052be | 4845 | |
b128c09f BB |
4846 | if (zio->io_error) { |
4847 | /* | |
4848 | * If this I/O is attached to a particular vdev, | |
4849 | * generate an error message describing the I/O failure | |
4850 | * at the block level. We ignore these errors if the | |
4851 | * device is currently unavailable. | |
4852 | */ | |
c776b317 | 4853 | if (zio->io_error != ECKSUM && zio->io_vd != NULL && |
2bbec1c9 | 4854 | !vdev_is_dead(zio->io_vd)) { |
4f072827 DB |
4855 | int ret = zfs_ereport_post(FM_EREPORT_ZFS_IO, |
4856 | zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0); | |
4857 | if (ret != EALREADY) { | |
4858 | mutex_enter(&zio->io_vd->vdev_stat_lock); | |
4859 | if (zio->io_type == ZIO_TYPE_READ) | |
4860 | zio->io_vd->vdev_stat.vs_read_errors++; | |
4861 | else if (zio->io_type == ZIO_TYPE_WRITE) | |
4862 | zio->io_vd->vdev_stat.vs_write_errors++; | |
4863 | mutex_exit(&zio->io_vd->vdev_stat_lock); | |
2bbec1c9 | 4864 | } |
2bbec1c9 | 4865 | } |
34dc7c2f | 4866 | |
428870ff BB |
4867 | if ((zio->io_error == EIO || !(zio->io_flags & |
4868 | (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && | |
c776b317 | 4869 | zio == zio->io_logical) { |
b128c09f BB |
4870 | /* |
4871 | * For logical I/O requests, tell the SPA to log the | |
4872 | * error and generate a logical data ereport. | |
4873 | */ | |
431083f7 | 4874 | spa_log_error(zio->io_spa, &zio->io_bookmark, |
493fcce9 | 4875 | BP_GET_LOGICAL_BIRTH(zio->io_bp)); |
1144586b | 4876 | (void) zfs_ereport_post(FM_EREPORT_ZFS_DATA, |
4f072827 | 4877 | zio->io_spa, NULL, &zio->io_bookmark, zio, 0); |
b128c09f BB |
4878 | } |
4879 | } | |
34dc7c2f | 4880 | |
c776b317 | 4881 | if (zio->io_error && zio == zio->io_logical) { |
b128c09f BB |
4882 | /* |
4883 | * Determine whether zio should be reexecuted. This will | |
4884 | * propagate all the way to the root via zio_notify_parent(). | |
4885 | */ | |
c776b317 | 4886 | ASSERT(zio->io_vd == NULL && zio->io_bp != NULL); |
428870ff | 4887 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); |
b128c09f | 4888 | |
428870ff BB |
4889 | if (IO_IS_ALLOCATING(zio) && |
4890 | !(zio->io_flags & ZIO_FLAG_CANFAIL)) { | |
b128c09f BB |
4891 | if (zio->io_error != ENOSPC) |
4892 | zio->io_reexecute |= ZIO_REEXECUTE_NOW; | |
4893 | else | |
4894 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff | 4895 | } |
b128c09f BB |
4896 | |
4897 | if ((zio->io_type == ZIO_TYPE_READ || | |
4898 | zio->io_type == ZIO_TYPE_FREE) && | |
572e2857 | 4899 | !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && |
b128c09f | 4900 | zio->io_error == ENXIO && |
c776b317 BB |
4901 | spa_load_state(zio->io_spa) == SPA_LOAD_NONE && |
4902 | spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE) | |
b128c09f BB |
4903 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; |
4904 | ||
4905 | if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) | |
4906 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff BB |
4907 | |
4908 | /* | |
4909 | * Here is a possibly good place to attempt to do | |
4910 | * either combinatorial reconstruction or error correction | |
4911 | * based on checksums. It also might be a good place | |
4912 | * to send out preliminary ereports before we suspend | |
4913 | * processing. | |
4914 | */ | |
34dc7c2f BB |
4915 | } |
4916 | ||
4917 | /* | |
b128c09f BB |
4918 | * If there were logical child errors, they apply to us now. |
4919 | * We defer this until now to avoid conflating logical child | |
4920 | * errors with errors that happened to the zio itself when | |
4921 | * updating vdev stats and reporting FMA events above. | |
34dc7c2f | 4922 | */ |
b128c09f | 4923 | zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); |
34dc7c2f | 4924 | |
428870ff BB |
4925 | if ((zio->io_error || zio->io_reexecute) && |
4926 | IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && | |
03c6040b | 4927 | !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) |
c776b317 | 4928 | zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp); |
9babb374 BB |
4929 | |
4930 | zio_gang_tree_free(&zio->io_gang_tree); | |
4931 | ||
4932 | /* | |
4933 | * Godfather I/Os should never suspend. | |
4934 | */ | |
4935 | if ((zio->io_flags & ZIO_FLAG_GODFATHER) && | |
4936 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) | |
a32494d2 | 4937 | zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND; |
9babb374 | 4938 | |
b128c09f BB |
4939 | if (zio->io_reexecute) { |
4940 | /* | |
4941 | * This is a logical I/O that wants to reexecute. | |
4942 | * | |
4943 | * Reexecute is top-down. When an i/o fails, if it's not | |
4944 | * the root, it simply notifies its parent and sticks around. | |
4945 | * The parent, seeing that it still has children in zio_done(), | |
4946 | * does the same. This percolates all the way up to the root. | |
4947 | * The root i/o will reexecute or suspend the entire tree. | |
4948 | * | |
4949 | * This approach ensures that zio_reexecute() honors | |
4950 | * all the original i/o dependency relationships, e.g. | |
4951 | * parents not executing until children are ready. | |
4952 | */ | |
4953 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
34dc7c2f | 4954 | |
9babb374 | 4955 | zio->io_gang_leader = NULL; |
b128c09f | 4956 | |
d164b209 BB |
4957 | mutex_enter(&zio->io_lock); |
4958 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
4959 | mutex_exit(&zio->io_lock); | |
4960 | ||
9babb374 BB |
4961 | /* |
4962 | * "The Godfather" I/O monitors its children but is | |
4963 | * not a true parent to them. It will track them through | |
4964 | * the pipeline but severs its ties whenever they get into | |
4965 | * trouble (e.g. suspended). This allows "The Godfather" | |
4966 | * I/O to return status without blocking. | |
4967 | */ | |
3dfb57a3 DB |
4968 | zl = NULL; |
4969 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; | |
4970 | pio = pio_next) { | |
4971 | zio_link_t *remove_zl = zl; | |
4972 | pio_next = zio_walk_parents(zio, &zl); | |
9babb374 BB |
4973 | |
4974 | if ((pio->io_flags & ZIO_FLAG_GODFATHER) && | |
4975 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { | |
3dfb57a3 | 4976 | zio_remove_child(pio, zio, remove_zl); |
62840030 MA |
4977 | /* |
4978 | * This is a rare code path, so we don't | |
4979 | * bother with "next_to_execute". | |
4980 | */ | |
4981 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, | |
4982 | NULL); | |
9babb374 BB |
4983 | } |
4984 | } | |
4985 | ||
d164b209 | 4986 | if ((pio = zio_unique_parent(zio)) != NULL) { |
b128c09f BB |
4987 | /* |
4988 | * We're not a root i/o, so there's nothing to do | |
4989 | * but notify our parent. Don't propagate errors | |
4990 | * upward since we haven't permanently failed yet. | |
4991 | */ | |
9babb374 | 4992 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f | 4993 | zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; |
62840030 MA |
4994 | /* |
4995 | * This is a rare code path, so we don't bother with | |
4996 | * "next_to_execute". | |
4997 | */ | |
4998 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL); | |
b128c09f BB |
4999 | } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { |
5000 | /* | |
5001 | * We'd fail again if we reexecuted now, so suspend | |
5002 | * until conditions improve (e.g. device comes online). | |
5003 | */ | |
cec3a0a1 | 5004 | zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR); |
b128c09f BB |
5005 | } else { |
5006 | /* | |
5007 | * Reexecution is potentially a huge amount of work. | |
5008 | * Hand it off to the otherwise-unused claim taskq. | |
5009 | */ | |
a38718a6 | 5010 | ASSERT(taskq_empty_ent(&zio->io_tqent)); |
7ef5e54e AL |
5011 | spa_taskq_dispatch_ent(zio->io_spa, |
5012 | ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE, | |
3bd4df38 | 5013 | zio_reexecute, zio, 0, &zio->io_tqent, NULL); |
b128c09f | 5014 | } |
62840030 | 5015 | return (NULL); |
34dc7c2f BB |
5016 | } |
5017 | ||
ccec7fbe | 5018 | ASSERT(list_is_empty(&zio->io_child_list)); |
b128c09f BB |
5019 | ASSERT(zio->io_reexecute == 0); |
5020 | ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); | |
34dc7c2f | 5021 | |
428870ff BB |
5022 | /* |
5023 | * Report any checksum errors, since the I/O is complete. | |
5024 | */ | |
5025 | while (zio->io_cksum_report != NULL) { | |
5026 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
5027 | zio->io_cksum_report = zcr->zcr_next; | |
5028 | zcr->zcr_next = NULL; | |
5029 | zcr->zcr_finish(zcr, NULL); | |
5030 | zfs_ereport_free_checksum(zcr); | |
5031 | } | |
5032 | ||
d164b209 BB |
5033 | /* |
5034 | * It is the responsibility of the done callback to ensure that this | |
5035 | * particular zio is no longer discoverable for adoption, and as | |
5036 | * such, cannot acquire any new parents. | |
5037 | */ | |
b128c09f BB |
5038 | if (zio->io_done) |
5039 | zio->io_done(zio); | |
34dc7c2f | 5040 | |
d164b209 BB |
5041 | mutex_enter(&zio->io_lock); |
5042 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
5043 | mutex_exit(&zio->io_lock); | |
34dc7c2f | 5044 | |
62840030 MA |
5045 | /* |
5046 | * We are done executing this zio. We may want to execute a parent | |
5047 | * next. See the comment in zio_notify_parent(). | |
5048 | */ | |
5049 | zio_t *next_to_execute = NULL; | |
3dfb57a3 DB |
5050 | zl = NULL; |
5051 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) { | |
5052 | zio_link_t *remove_zl = zl; | |
5053 | pio_next = zio_walk_parents(zio, &zl); | |
5054 | zio_remove_child(pio, zio, remove_zl); | |
62840030 | 5055 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute); |
b128c09f | 5056 | } |
34dc7c2f | 5057 | |
b128c09f BB |
5058 | if (zio->io_waiter != NULL) { |
5059 | mutex_enter(&zio->io_lock); | |
5060 | zio->io_executor = NULL; | |
5061 | cv_broadcast(&zio->io_cv); | |
5062 | mutex_exit(&zio->io_lock); | |
5063 | } else { | |
5064 | zio_destroy(zio); | |
5065 | } | |
34dc7c2f | 5066 | |
62840030 | 5067 | return (next_to_execute); |
34dc7c2f BB |
5068 | } |
5069 | ||
5070 | /* | |
b128c09f BB |
5071 | * ========================================================================== |
5072 | * I/O pipeline definition | |
5073 | * ========================================================================== | |
34dc7c2f | 5074 | */ |
428870ff | 5075 | static zio_pipe_stage_t *zio_pipeline[] = { |
b128c09f | 5076 | NULL, |
b128c09f | 5077 | zio_read_bp_init, |
3dfb57a3 | 5078 | zio_write_bp_init, |
428870ff BB |
5079 | zio_free_bp_init, |
5080 | zio_issue_async, | |
3dfb57a3 | 5081 | zio_write_compress, |
b5256303 | 5082 | zio_encrypt, |
b128c09f | 5083 | zio_checksum_generate, |
03c6040b | 5084 | zio_nop_write, |
67a1b037 | 5085 | zio_brt_free, |
428870ff BB |
5086 | zio_ddt_read_start, |
5087 | zio_ddt_read_done, | |
5088 | zio_ddt_write, | |
5089 | zio_ddt_free, | |
b128c09f BB |
5090 | zio_gang_assemble, |
5091 | zio_gang_issue, | |
3dfb57a3 | 5092 | zio_dva_throttle, |
b128c09f BB |
5093 | zio_dva_allocate, |
5094 | zio_dva_free, | |
5095 | zio_dva_claim, | |
5096 | zio_ready, | |
5097 | zio_vdev_io_start, | |
5098 | zio_vdev_io_done, | |
5099 | zio_vdev_io_assess, | |
5100 | zio_checksum_verify, | |
5101 | zio_done | |
5102 | }; | |
c28b2279 | 5103 | |
9ae529ec | 5104 | |
9ae529ec | 5105 | |
9ae529ec | 5106 | |
fcff0f35 PD |
5107 | /* |
5108 | * Compare two zbookmark_phys_t's to see which we would reach first in a | |
5109 | * pre-order traversal of the object tree. | |
5110 | * | |
5111 | * This is simple in every case aside from the meta-dnode object. For all other | |
5112 | * objects, we traverse them in order (object 1 before object 2, and so on). | |
5113 | * However, all of these objects are traversed while traversing object 0, since | |
5114 | * the data it points to is the list of objects. Thus, we need to convert to a | |
5115 | * canonical representation so we can compare meta-dnode bookmarks to | |
5116 | * non-meta-dnode bookmarks. | |
5117 | * | |
5118 | * We do this by calculating "equivalents" for each field of the zbookmark. | |
5119 | * zbookmarks outside of the meta-dnode use their own object and level, and | |
5120 | * calculate the level 0 equivalent (the first L0 blkid that is contained in the | |
5121 | * blocks this bookmark refers to) by multiplying their blkid by their span | |
5122 | * (the number of L0 blocks contained within one block at their level). | |
5123 | * zbookmarks inside the meta-dnode calculate their object equivalent | |
5124 | * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use | |
5125 | * level + 1<<31 (any value larger than a level could ever be) for their level. | |
5126 | * This causes them to always compare before a bookmark in their object | |
5127 | * equivalent, compare appropriately to bookmarks in other objects, and to | |
5128 | * compare appropriately to other bookmarks in the meta-dnode. | |
5129 | */ | |
5130 | int | |
5131 | zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, | |
5132 | const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) | |
5133 | { | |
5134 | /* | |
5135 | * These variables represent the "equivalent" values for the zbookmark, | |
5136 | * after converting zbookmarks inside the meta dnode to their | |
5137 | * normal-object equivalents. | |
5138 | */ | |
5139 | uint64_t zb1obj, zb2obj; | |
5140 | uint64_t zb1L0, zb2L0; | |
5141 | uint64_t zb1level, zb2level; | |
5142 | ||
5143 | if (zb1->zb_object == zb2->zb_object && | |
5144 | zb1->zb_level == zb2->zb_level && | |
5145 | zb1->zb_blkid == zb2->zb_blkid) | |
5146 | return (0); | |
9ae529ec | 5147 | |
30af21b0 PD |
5148 | IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT); |
5149 | IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT); | |
5150 | ||
fcff0f35 PD |
5151 | /* |
5152 | * BP_SPANB calculates the span in blocks. | |
5153 | */ | |
5154 | zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); | |
5155 | zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); | |
9ae529ec CS |
5156 | |
5157 | if (zb1->zb_object == DMU_META_DNODE_OBJECT) { | |
fcff0f35 PD |
5158 | zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); |
5159 | zb1L0 = 0; | |
5160 | zb1level = zb1->zb_level + COMPARE_META_LEVEL; | |
5161 | } else { | |
5162 | zb1obj = zb1->zb_object; | |
5163 | zb1level = zb1->zb_level; | |
9ae529ec CS |
5164 | } |
5165 | ||
fcff0f35 PD |
5166 | if (zb2->zb_object == DMU_META_DNODE_OBJECT) { |
5167 | zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); | |
5168 | zb2L0 = 0; | |
5169 | zb2level = zb2->zb_level + COMPARE_META_LEVEL; | |
5170 | } else { | |
5171 | zb2obj = zb2->zb_object; | |
5172 | zb2level = zb2->zb_level; | |
5173 | } | |
5174 | ||
5175 | /* Now that we have a canonical representation, do the comparison. */ | |
5176 | if (zb1obj != zb2obj) | |
5177 | return (zb1obj < zb2obj ? -1 : 1); | |
5178 | else if (zb1L0 != zb2L0) | |
5179 | return (zb1L0 < zb2L0 ? -1 : 1); | |
5180 | else if (zb1level != zb2level) | |
5181 | return (zb1level > zb2level ? -1 : 1); | |
5182 | /* | |
5183 | * This can (theoretically) happen if the bookmarks have the same object | |
5184 | * and level, but different blkids, if the block sizes are not the same. | |
5185 | * There is presently no way to change the indirect block sizes | |
5186 | */ | |
5187 | return (0); | |
5188 | } | |
5189 | ||
5190 | /* | |
5191 | * This function checks the following: given that last_block is the place that | |
5192 | * our traversal stopped last time, does that guarantee that we've visited | |
5193 | * every node under subtree_root? Therefore, we can't just use the raw output | |
5194 | * of zbookmark_compare. We have to pass in a modified version of | |
5195 | * subtree_root; by incrementing the block id, and then checking whether | |
5196 | * last_block is before or equal to that, we can tell whether or not having | |
5197 | * visited last_block implies that all of subtree_root's children have been | |
5198 | * visited. | |
5199 | */ | |
5200 | boolean_t | |
5201 | zbookmark_subtree_completed(const dnode_phys_t *dnp, | |
5202 | const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) | |
5203 | { | |
5204 | zbookmark_phys_t mod_zb = *subtree_root; | |
5205 | mod_zb.zb_blkid++; | |
33dba8c7 | 5206 | ASSERT0(last_block->zb_level); |
fcff0f35 PD |
5207 | |
5208 | /* The objset_phys_t isn't before anything. */ | |
5209 | if (dnp == NULL) | |
9ae529ec | 5210 | return (B_FALSE); |
fcff0f35 PD |
5211 | |
5212 | /* | |
5213 | * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the | |
5214 | * data block size in sectors, because that variable is only used if | |
5215 | * the bookmark refers to a block in the meta-dnode. Since we don't | |
5216 | * know without examining it what object it refers to, and there's no | |
5217 | * harm in passing in this value in other cases, we always pass it in. | |
5218 | * | |
5219 | * We pass in 0 for the indirect block size shift because zb2 must be | |
5220 | * level 0. The indirect block size is only used to calculate the span | |
5221 | * of the bookmark, but since the bookmark must be level 0, the span is | |
5222 | * always 1, so the math works out. | |
5223 | * | |
5224 | * If you make changes to how the zbookmark_compare code works, be sure | |
5225 | * to make sure that this code still works afterwards. | |
5226 | */ | |
5227 | return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, | |
5228 | 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, | |
5229 | last_block) <= 0); | |
9ae529ec CS |
5230 | } |
5231 | ||
33dba8c7 AM |
5232 | /* |
5233 | * This function is similar to zbookmark_subtree_completed(), but returns true | |
5234 | * if subtree_root is equal or ahead of last_block, i.e. still to be done. | |
5235 | */ | |
5236 | boolean_t | |
5237 | zbookmark_subtree_tbd(const dnode_phys_t *dnp, | |
5238 | const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) | |
5239 | { | |
5240 | ASSERT0(last_block->zb_level); | |
5241 | if (dnp == NULL) | |
5242 | return (B_FALSE); | |
5243 | return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, | |
5244 | 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, subtree_root, | |
5245 | last_block) >= 0); | |
5246 | } | |
5247 | ||
c28b2279 | 5248 | EXPORT_SYMBOL(zio_type_name); |
81971b13 BB |
5249 | EXPORT_SYMBOL(zio_buf_alloc); |
5250 | EXPORT_SYMBOL(zio_data_buf_alloc); | |
5251 | EXPORT_SYMBOL(zio_buf_free); | |
5252 | EXPORT_SYMBOL(zio_data_buf_free); | |
c28b2279 | 5253 | |
03fdcb9a | 5254 | ZFS_MODULE_PARAM(zfs_zio, zio_, slow_io_ms, INT, ZMOD_RW, |
ad796b8a | 5255 | "Max I/O completion time (milliseconds) before marking it as slow"); |
c409e464 | 5256 | |
03fdcb9a MM |
5257 | ZFS_MODULE_PARAM(zfs_zio, zio_, requeue_io_start_cut_in_line, INT, ZMOD_RW, |
5258 | "Prioritize requeued I/O"); | |
29dee3ee | 5259 | |
fdc2d303 | 5260 | ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_deferred_free, UINT, ZMOD_RW, |
d1d7e268 | 5261 | "Defer frees starting in this pass"); |
29dee3ee | 5262 | |
fdc2d303 | 5263 | ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_dont_compress, UINT, ZMOD_RW, |
d1d7e268 | 5264 | "Don't compress starting in this pass"); |
29dee3ee | 5265 | |
fdc2d303 | 5266 | ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_rewrite, UINT, ZMOD_RW, |
d1d7e268 | 5267 | "Rewrite new bps starting in this pass"); |
3dfb57a3 | 5268 | |
03fdcb9a | 5269 | ZFS_MODULE_PARAM(zfs_zio, zio_, dva_throttle_enabled, INT, ZMOD_RW, |
3dfb57a3 | 5270 | "Throttle block allocations in the ZIO pipeline"); |
638dd5f4 | 5271 | |
03fdcb9a | 5272 | ZFS_MODULE_PARAM(zfs_zio, zio_, deadman_log_all, INT, ZMOD_RW, |
638dd5f4 | 5273 | "Log all slow ZIOs, not just those with vdevs"); |