]>
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. |
10b3c7f5 MN |
26 | * Copyright (c) 2019, Klara Inc. |
27 | * Copyright (c) 2019, Allan Jude | |
f2286383 | 28 | * Copyright (c) 2021, Datto, Inc. |
34dc7c2f BB |
29 | */ |
30 | ||
f1512ee6 | 31 | #include <sys/sysmacros.h> |
34dc7c2f BB |
32 | #include <sys/zfs_context.h> |
33 | #include <sys/fm/fs/zfs.h> | |
34 | #include <sys/spa.h> | |
35 | #include <sys/txg.h> | |
36 | #include <sys/spa_impl.h> | |
37 | #include <sys/vdev_impl.h> | |
1b939560 | 38 | #include <sys/vdev_trim.h> |
34dc7c2f BB |
39 | #include <sys/zio_impl.h> |
40 | #include <sys/zio_compress.h> | |
41 | #include <sys/zio_checksum.h> | |
428870ff BB |
42 | #include <sys/dmu_objset.h> |
43 | #include <sys/arc.h> | |
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 | */ | |
1b939560 | 66 | "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "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 GA |
615 | spa_log_error(spa, &zio->io_bookmark, |
616 | &zio->io_bp->blk_birth); | |
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 |
ffd2e15d GW |
806 | * so. We only do this if the parent's zio type matches the |
807 | * child's type. Otherwise dispatch the parent zio in its | |
808 | * own taskq. | |
62840030 MA |
809 | * |
810 | * Having the caller execute the parent when possible reduces | |
811 | * locking on the zio taskq's, reduces context switch | |
812 | * overhead, and has no recursion penalty. Note that one | |
813 | * read from disk typically causes at least 3 zio's: a | |
814 | * zio_null(), the logical zio_read(), and then a physical | |
815 | * zio. When the physical ZIO completes, we are able to call | |
816 | * zio_done() on all 3 of these zio's from one invocation of | |
817 | * zio_execute() by returning the parent back to | |
818 | * zio_execute(). Since the parent isn't executed until this | |
819 | * thread returns back to zio_execute(), the caller should do | |
820 | * so promptly. | |
821 | * | |
822 | * In other cases, dispatching the parent prevents | |
823 | * overflowing the stack when we have deeply nested | |
824 | * parent-child relationships, as we do with the "mega zio" | |
825 | * of writes for spa_sync(), and the chain of ZIL blocks. | |
3dfb57a3 | 826 | */ |
ffd2e15d GW |
827 | if (next_to_executep != NULL && *next_to_executep == NULL && |
828 | pio->io_type == zio->io_type) { | |
62840030 MA |
829 | *next_to_executep = pio; |
830 | } else { | |
831 | zio_taskq_dispatch(pio, type, B_FALSE); | |
832 | } | |
b128c09f BB |
833 | } else { |
834 | mutex_exit(&pio->io_lock); | |
34dc7c2f BB |
835 | } |
836 | } | |
837 | ||
b128c09f BB |
838 | static void |
839 | zio_inherit_child_errors(zio_t *zio, enum zio_child c) | |
840 | { | |
841 | if (zio->io_child_error[c] != 0 && zio->io_error == 0) | |
842 | zio->io_error = zio->io_child_error[c]; | |
843 | } | |
844 | ||
3dfb57a3 | 845 | int |
64fc7762 | 846 | zio_bookmark_compare(const void *x1, const void *x2) |
3dfb57a3 DB |
847 | { |
848 | const zio_t *z1 = x1; | |
849 | const zio_t *z2 = x2; | |
3dfb57a3 | 850 | |
64fc7762 MA |
851 | if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset) |
852 | return (-1); | |
853 | if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset) | |
854 | return (1); | |
3dfb57a3 | 855 | |
64fc7762 MA |
856 | if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object) |
857 | return (-1); | |
858 | if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object) | |
859 | return (1); | |
3dfb57a3 | 860 | |
64fc7762 MA |
861 | if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level) |
862 | return (-1); | |
863 | if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level) | |
864 | return (1); | |
865 | ||
866 | if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid) | |
867 | return (-1); | |
868 | if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid) | |
869 | return (1); | |
870 | ||
871 | if (z1 < z2) | |
872 | return (-1); | |
873 | if (z1 > z2) | |
874 | return (1); | |
875 | ||
876 | return (0); | |
3dfb57a3 DB |
877 | } |
878 | ||
34dc7c2f BB |
879 | /* |
880 | * ========================================================================== | |
b128c09f | 881 | * Create the various types of I/O (read, write, free, etc) |
34dc7c2f BB |
882 | * ========================================================================== |
883 | */ | |
884 | static zio_t * | |
428870ff | 885 | zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
a6255b7f | 886 | abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done, |
2aa34383 | 887 | void *private, zio_type_t type, zio_priority_t priority, |
4938d01d | 888 | zio_flag_t flags, vdev_t *vd, uint64_t offset, |
2aa34383 DK |
889 | const zbookmark_phys_t *zb, enum zio_stage stage, |
890 | enum zio_stage pipeline) | |
34dc7c2f BB |
891 | { |
892 | zio_t *zio; | |
893 | ||
1b939560 | 894 | IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE); |
2aa34383 | 895 | ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); |
b128c09f BB |
896 | ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); |
897 | ||
898 | ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); | |
899 | ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); | |
900 | ASSERT(vd || stage == ZIO_STAGE_OPEN); | |
34dc7c2f | 901 | |
b5256303 | 902 | IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0); |
2aa34383 | 903 | |
79c76d5b | 904 | zio = kmem_cache_alloc(zio_cache, KM_SLEEP); |
861166b0 | 905 | memset(zio, 0, sizeof (zio_t)); |
3941503c | 906 | |
448d7aaa | 907 | mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL); |
3941503c BB |
908 | cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); |
909 | ||
910 | list_create(&zio->io_parent_list, sizeof (zio_link_t), | |
911 | offsetof(zio_link_t, zl_parent_node)); | |
912 | list_create(&zio->io_child_list, sizeof (zio_link_t), | |
913 | offsetof(zio_link_t, zl_child_node)); | |
4e21fd06 | 914 | metaslab_trace_init(&zio->io_alloc_list); |
d164b209 | 915 | |
b128c09f BB |
916 | if (vd != NULL) |
917 | zio->io_child_type = ZIO_CHILD_VDEV; | |
918 | else if (flags & ZIO_FLAG_GANG_CHILD) | |
919 | zio->io_child_type = ZIO_CHILD_GANG; | |
428870ff BB |
920 | else if (flags & ZIO_FLAG_DDT_CHILD) |
921 | zio->io_child_type = ZIO_CHILD_DDT; | |
b128c09f BB |
922 | else |
923 | zio->io_child_type = ZIO_CHILD_LOGICAL; | |
924 | ||
34dc7c2f | 925 | if (bp != NULL) { |
428870ff | 926 | if (type != ZIO_TYPE_WRITE || |
b4a08730 AM |
927 | zio->io_child_type == ZIO_CHILD_DDT) { |
928 | zio->io_bp_copy = *bp; | |
b128c09f | 929 | zio->io_bp = &zio->io_bp_copy; /* so caller can free */ |
b4a08730 AM |
930 | } else { |
931 | zio->io_bp = (blkptr_t *)bp; | |
932 | } | |
933 | zio->io_bp_orig = *bp; | |
9babb374 | 934 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) |
b128c09f | 935 | zio->io_logical = zio; |
9babb374 BB |
936 | if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) |
937 | pipeline |= ZIO_GANG_STAGES; | |
34dc7c2f | 938 | } |
b128c09f BB |
939 | |
940 | zio->io_spa = spa; | |
941 | zio->io_txg = txg; | |
34dc7c2f BB |
942 | zio->io_done = done; |
943 | zio->io_private = private; | |
944 | zio->io_type = type; | |
945 | zio->io_priority = priority; | |
b128c09f BB |
946 | zio->io_vd = vd; |
947 | zio->io_offset = offset; | |
a6255b7f | 948 | zio->io_orig_abd = zio->io_abd = data; |
2aa34383 DK |
949 | zio->io_orig_size = zio->io_size = psize; |
950 | zio->io_lsize = lsize; | |
b128c09f BB |
951 | zio->io_orig_flags = zio->io_flags = flags; |
952 | zio->io_orig_stage = zio->io_stage = stage; | |
953 | zio->io_orig_pipeline = zio->io_pipeline = pipeline; | |
3dfb57a3 | 954 | zio->io_pipeline_trace = ZIO_STAGE_OPEN; |
3bd4df38 | 955 | zio->io_allocator = ZIO_ALLOCATOR_NONE; |
34dc7c2f | 956 | |
3afdc97d AM |
957 | zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY) || |
958 | (pipeline & ZIO_STAGE_READY) == 0; | |
d164b209 BB |
959 | zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); |
960 | ||
b128c09f BB |
961 | if (zb != NULL) |
962 | zio->io_bookmark = *zb; | |
963 | ||
964 | if (pio != NULL) { | |
1b50749c | 965 | zio->io_metaslab_class = pio->io_metaslab_class; |
b128c09f | 966 | if (zio->io_logical == NULL) |
34dc7c2f | 967 | zio->io_logical = pio->io_logical; |
9babb374 BB |
968 | if (zio->io_child_type == ZIO_CHILD_GANG) |
969 | zio->io_gang_leader = pio->io_gang_leader; | |
b4a08730 | 970 | zio_add_child_first(pio, zio); |
34dc7c2f BB |
971 | } |
972 | ||
a38718a6 GA |
973 | taskq_init_ent(&zio->io_tqent); |
974 | ||
34dc7c2f BB |
975 | return (zio); |
976 | } | |
977 | ||
e8cf3a4f | 978 | void |
b128c09f | 979 | zio_destroy(zio_t *zio) |
34dc7c2f | 980 | { |
4e21fd06 | 981 | metaslab_trace_fini(&zio->io_alloc_list); |
3941503c BB |
982 | list_destroy(&zio->io_parent_list); |
983 | list_destroy(&zio->io_child_list); | |
984 | mutex_destroy(&zio->io_lock); | |
985 | cv_destroy(&zio->io_cv); | |
b128c09f | 986 | kmem_cache_free(zio_cache, zio); |
34dc7c2f BB |
987 | } |
988 | ||
3afdc97d AM |
989 | /* |
990 | * ZIO intended to be between others. Provides synchronization at READY | |
991 | * and DONE pipeline stages and calls the respective callbacks. | |
992 | */ | |
34dc7c2f | 993 | zio_t * |
d164b209 | 994 | zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, |
4938d01d | 995 | void *private, zio_flag_t flags) |
34dc7c2f BB |
996 | { |
997 | zio_t *zio; | |
998 | ||
2aa34383 | 999 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
d164b209 | 1000 | ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
b128c09f | 1001 | ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); |
34dc7c2f BB |
1002 | |
1003 | return (zio); | |
1004 | } | |
1005 | ||
3afdc97d AM |
1006 | /* |
1007 | * ZIO intended to be a root of a tree. Unlike null ZIO does not have a | |
1008 | * READY pipeline stage (is ready on creation), so it should not be used | |
1009 | * as child of any ZIO that may need waiting for grandchildren READY stage | |
1010 | * (any other ZIO type). | |
1011 | */ | |
34dc7c2f | 1012 | zio_t * |
4938d01d | 1013 | zio_root(spa_t *spa, zio_done_func_t *done, void *private, zio_flag_t flags) |
34dc7c2f | 1014 | { |
3afdc97d AM |
1015 | zio_t *zio; |
1016 | ||
1017 | zio = zio_create(NULL, spa, 0, NULL, NULL, 0, 0, done, private, | |
1018 | ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, NULL, 0, NULL, | |
1019 | ZIO_STAGE_OPEN, ZIO_ROOT_PIPELINE); | |
1020 | ||
1021 | return (zio); | |
34dc7c2f BB |
1022 | } |
1023 | ||
bc67cba7 PZ |
1024 | static int |
1025 | zfs_blkptr_verify_log(spa_t *spa, const blkptr_t *bp, | |
1026 | enum blk_verify_flag blk_verify, const char *fmt, ...) | |
1027 | { | |
1028 | va_list adx; | |
1029 | char buf[256]; | |
1030 | ||
1031 | va_start(adx, fmt); | |
1032 | (void) vsnprintf(buf, sizeof (buf), fmt, adx); | |
1033 | va_end(adx); | |
1034 | ||
3095ca91 MA |
1035 | zfs_dbgmsg("bad blkptr at %px: " |
1036 | "DVA[0]=%#llx/%#llx " | |
1037 | "DVA[1]=%#llx/%#llx " | |
1038 | "DVA[2]=%#llx/%#llx " | |
1039 | "prop=%#llx " | |
1040 | "pad=%#llx,%#llx " | |
1041 | "phys_birth=%#llx " | |
1042 | "birth=%#llx " | |
1043 | "fill=%#llx " | |
1044 | "cksum=%#llx/%#llx/%#llx/%#llx", | |
1045 | bp, | |
1046 | (long long)bp->blk_dva[0].dva_word[0], | |
1047 | (long long)bp->blk_dva[0].dva_word[1], | |
1048 | (long long)bp->blk_dva[1].dva_word[0], | |
1049 | (long long)bp->blk_dva[1].dva_word[1], | |
1050 | (long long)bp->blk_dva[2].dva_word[0], | |
1051 | (long long)bp->blk_dva[2].dva_word[1], | |
1052 | (long long)bp->blk_prop, | |
1053 | (long long)bp->blk_pad[0], | |
1054 | (long long)bp->blk_pad[1], | |
1055 | (long long)bp->blk_phys_birth, | |
1056 | (long long)bp->blk_birth, | |
1057 | (long long)bp->blk_fill, | |
1058 | (long long)bp->blk_cksum.zc_word[0], | |
1059 | (long long)bp->blk_cksum.zc_word[1], | |
1060 | (long long)bp->blk_cksum.zc_word[2], | |
1061 | (long long)bp->blk_cksum.zc_word[3]); | |
bc67cba7 PZ |
1062 | switch (blk_verify) { |
1063 | case BLK_VERIFY_HALT: | |
1064 | zfs_panic_recover("%s: %s", spa_name(spa), buf); | |
1065 | break; | |
1066 | case BLK_VERIFY_LOG: | |
1067 | zfs_dbgmsg("%s: %s", spa_name(spa), buf); | |
1068 | break; | |
1069 | case BLK_VERIFY_ONLY: | |
1070 | break; | |
1071 | } | |
1072 | ||
1073 | return (1); | |
1074 | } | |
1075 | ||
1076 | /* | |
1077 | * Verify the block pointer fields contain reasonable values. This means | |
1078 | * it only contains known object types, checksum/compression identifiers, | |
1079 | * block sizes within the maximum allowed limits, valid DVAs, etc. | |
1080 | * | |
1081 | * If everything checks out B_TRUE is returned. The zfs_blkptr_verify | |
1082 | * argument controls the behavior when an invalid field is detected. | |
1083 | * | |
3095ca91 MA |
1084 | * Values for blk_verify_flag: |
1085 | * BLK_VERIFY_ONLY: evaluate the block | |
1086 | * BLK_VERIFY_LOG: evaluate the block and log problems | |
1087 | * BLK_VERIFY_HALT: call zfs_panic_recover on error | |
1088 | * | |
1089 | * Values for blk_config_flag: | |
1090 | * BLK_CONFIG_HELD: caller holds SCL_VDEV for writer | |
1091 | * BLK_CONFIG_NEEDED: caller holds no config lock, SCL_VDEV will be | |
1092 | * obtained for reader | |
1093 | * BLK_CONFIG_SKIP: skip checks which require SCL_VDEV, for better | |
1094 | * performance | |
bc67cba7 PZ |
1095 | */ |
1096 | boolean_t | |
3095ca91 MA |
1097 | zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp, |
1098 | enum blk_config_flag blk_config, enum blk_verify_flag blk_verify) | |
63e3a861 | 1099 | { |
bc67cba7 PZ |
1100 | int errors = 0; |
1101 | ||
63e3a861 | 1102 | if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { |
bc67cba7 | 1103 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1104 | "blkptr at %px has invalid TYPE %llu", |
63e3a861 MA |
1105 | bp, (longlong_t)BP_GET_TYPE(bp)); |
1106 | } | |
2cd0f98f | 1107 | if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS) { |
bc67cba7 | 1108 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1109 | "blkptr at %px has invalid CHECKSUM %llu", |
63e3a861 MA |
1110 | bp, (longlong_t)BP_GET_CHECKSUM(bp)); |
1111 | } | |
2cd0f98f | 1112 | if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS) { |
bc67cba7 | 1113 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1114 | "blkptr at %px has invalid COMPRESS %llu", |
63e3a861 MA |
1115 | bp, (longlong_t)BP_GET_COMPRESS(bp)); |
1116 | } | |
1117 | if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
bc67cba7 | 1118 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1119 | "blkptr at %px has invalid LSIZE %llu", |
63e3a861 MA |
1120 | bp, (longlong_t)BP_GET_LSIZE(bp)); |
1121 | } | |
1122 | if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
bc67cba7 | 1123 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1124 | "blkptr at %px has invalid PSIZE %llu", |
63e3a861 MA |
1125 | bp, (longlong_t)BP_GET_PSIZE(bp)); |
1126 | } | |
1127 | ||
1128 | if (BP_IS_EMBEDDED(bp)) { | |
746d4a45 | 1129 | if (BPE_GET_ETYPE(bp) >= NUM_BP_EMBEDDED_TYPES) { |
bc67cba7 | 1130 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1131 | "blkptr at %px has invalid ETYPE %llu", |
63e3a861 MA |
1132 | bp, (longlong_t)BPE_GET_ETYPE(bp)); |
1133 | } | |
1134 | } | |
1135 | ||
6cb8e530 PZ |
1136 | /* |
1137 | * Do not verify individual DVAs if the config is not trusted. This | |
1138 | * will be done once the zio is executed in vdev_mirror_map_alloc. | |
1139 | */ | |
1140 | if (!spa->spa_trust_config) | |
b9ec4a15 | 1141 | return (errors == 0); |
6cb8e530 | 1142 | |
3095ca91 MA |
1143 | switch (blk_config) { |
1144 | case BLK_CONFIG_HELD: | |
dc04a8c7 | 1145 | ASSERT(spa_config_held(spa, SCL_VDEV, RW_WRITER)); |
3095ca91 MA |
1146 | break; |
1147 | case BLK_CONFIG_NEEDED: | |
1148 | spa_config_enter(spa, SCL_VDEV, bp, RW_READER); | |
1149 | break; | |
1150 | case BLK_CONFIG_SKIP: | |
1151 | return (errors == 0); | |
1152 | default: | |
1153 | panic("invalid blk_config %u", blk_config); | |
1154 | } | |
1155 | ||
63e3a861 MA |
1156 | /* |
1157 | * Pool-specific checks. | |
1158 | * | |
1159 | * Note: it would be nice to verify that the blk_birth and | |
1160 | * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() | |
1161 | * allows the birth time of log blocks (and dmu_sync()-ed blocks | |
1162 | * that are in the log) to be arbitrarily large. | |
1163 | */ | |
1c27024e | 1164 | for (int i = 0; i < BP_GET_NDVAS(bp); i++) { |
2b56a634 MA |
1165 | const dva_t *dva = &bp->blk_dva[i]; |
1166 | uint64_t vdevid = DVA_GET_VDEV(dva); | |
1c27024e | 1167 | |
63e3a861 | 1168 | if (vdevid >= spa->spa_root_vdev->vdev_children) { |
bc67cba7 | 1169 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1170 | "blkptr at %px DVA %u has invalid VDEV %llu", |
63e3a861 | 1171 | bp, i, (longlong_t)vdevid); |
ee3a23b8 | 1172 | continue; |
63e3a861 | 1173 | } |
1c27024e | 1174 | vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; |
63e3a861 | 1175 | if (vd == NULL) { |
bc67cba7 | 1176 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1177 | "blkptr at %px DVA %u has invalid VDEV %llu", |
63e3a861 | 1178 | bp, i, (longlong_t)vdevid); |
ee3a23b8 | 1179 | continue; |
63e3a861 MA |
1180 | } |
1181 | if (vd->vdev_ops == &vdev_hole_ops) { | |
bc67cba7 | 1182 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1183 | "blkptr at %px DVA %u has hole VDEV %llu", |
63e3a861 | 1184 | bp, i, (longlong_t)vdevid); |
ee3a23b8 | 1185 | continue; |
63e3a861 MA |
1186 | } |
1187 | if (vd->vdev_ops == &vdev_missing_ops) { | |
1188 | /* | |
1189 | * "missing" vdevs are valid during import, but we | |
1190 | * don't have their detailed info (e.g. asize), so | |
1191 | * we can't perform any more checks on them. | |
1192 | */ | |
1193 | continue; | |
1194 | } | |
2b56a634 MA |
1195 | uint64_t offset = DVA_GET_OFFSET(dva); |
1196 | uint64_t asize = DVA_GET_ASIZE(dva); | |
1197 | if (DVA_GET_GANG(dva)) | |
1198 | asize = vdev_gang_header_asize(vd); | |
63e3a861 | 1199 | if (offset + asize > vd->vdev_asize) { |
bc67cba7 | 1200 | errors += zfs_blkptr_verify_log(spa, bp, blk_verify, |
3095ca91 | 1201 | "blkptr at %px DVA %u has invalid OFFSET %llu", |
63e3a861 MA |
1202 | bp, i, (longlong_t)offset); |
1203 | } | |
1204 | } | |
3095ca91 | 1205 | if (blk_config == BLK_CONFIG_NEEDED) |
dc04a8c7 | 1206 | spa_config_exit(spa, SCL_VDEV, bp); |
bc67cba7 PZ |
1207 | |
1208 | return (errors == 0); | |
63e3a861 MA |
1209 | } |
1210 | ||
6cb8e530 PZ |
1211 | boolean_t |
1212 | zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp) | |
1213 | { | |
14e4e3cb | 1214 | (void) bp; |
6cb8e530 PZ |
1215 | uint64_t vdevid = DVA_GET_VDEV(dva); |
1216 | ||
1217 | if (vdevid >= spa->spa_root_vdev->vdev_children) | |
1218 | return (B_FALSE); | |
1219 | ||
1220 | vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; | |
1221 | if (vd == NULL) | |
1222 | return (B_FALSE); | |
1223 | ||
1224 | if (vd->vdev_ops == &vdev_hole_ops) | |
1225 | return (B_FALSE); | |
1226 | ||
1227 | if (vd->vdev_ops == &vdev_missing_ops) { | |
1228 | return (B_FALSE); | |
1229 | } | |
1230 | ||
1231 | uint64_t offset = DVA_GET_OFFSET(dva); | |
1232 | uint64_t asize = DVA_GET_ASIZE(dva); | |
1233 | ||
2b56a634 MA |
1234 | if (DVA_GET_GANG(dva)) |
1235 | asize = vdev_gang_header_asize(vd); | |
6cb8e530 PZ |
1236 | if (offset + asize > vd->vdev_asize) |
1237 | return (B_FALSE); | |
1238 | ||
1239 | return (B_TRUE); | |
1240 | } | |
1241 | ||
34dc7c2f | 1242 | zio_t * |
b128c09f | 1243 | zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, |
a6255b7f | 1244 | abd_t *data, uint64_t size, zio_done_func_t *done, void *private, |
4938d01d | 1245 | zio_priority_t priority, zio_flag_t flags, const zbookmark_phys_t *zb) |
34dc7c2f BB |
1246 | { |
1247 | zio_t *zio; | |
1248 | ||
428870ff | 1249 | zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, |
2aa34383 | 1250 | data, size, size, done, private, |
b128c09f | 1251 | ZIO_TYPE_READ, priority, flags, NULL, 0, zb, |
428870ff BB |
1252 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
1253 | ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); | |
34dc7c2f | 1254 | |
b128c09f BB |
1255 | return (zio); |
1256 | } | |
34dc7c2f | 1257 | |
34dc7c2f | 1258 | zio_t * |
b128c09f | 1259 | zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, |
a6255b7f | 1260 | abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp, |
bc77ba73 | 1261 | zio_done_func_t *ready, zio_done_func_t *children_ready, |
ccec7fbe AM |
1262 | zio_done_func_t *done, void *private, zio_priority_t priority, |
1263 | zio_flag_t flags, const zbookmark_phys_t *zb) | |
34dc7c2f BB |
1264 | { |
1265 | zio_t *zio; | |
1266 | ||
b128c09f BB |
1267 | ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && |
1268 | zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && | |
1269 | zp->zp_compress >= ZIO_COMPRESS_OFF && | |
1270 | zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && | |
9ae529ec | 1271 | DMU_OT_IS_VALID(zp->zp_type) && |
b128c09f | 1272 | zp->zp_level < 32 && |
428870ff | 1273 | zp->zp_copies > 0 && |
03c6040b | 1274 | zp->zp_copies <= spa_max_replication(spa)); |
34dc7c2f | 1275 | |
2aa34383 | 1276 | zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private, |
b128c09f | 1277 | ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, |
428870ff BB |
1278 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
1279 | ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); | |
34dc7c2f BB |
1280 | |
1281 | zio->io_ready = ready; | |
bc77ba73 | 1282 | zio->io_children_ready = children_ready; |
b128c09f | 1283 | zio->io_prop = *zp; |
34dc7c2f | 1284 | |
9b67f605 MA |
1285 | /* |
1286 | * Data can be NULL if we are going to call zio_write_override() to | |
1287 | * provide the already-allocated BP. But we may need the data to | |
1288 | * verify a dedup hit (if requested). In this case, don't try to | |
b5256303 TC |
1289 | * dedup (just take the already-allocated BP verbatim). Encrypted |
1290 | * dedup blocks need data as well so we also disable dedup in this | |
1291 | * case. | |
9b67f605 | 1292 | */ |
b5256303 TC |
1293 | if (data == NULL && |
1294 | (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) { | |
9b67f605 MA |
1295 | zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; |
1296 | } | |
1297 | ||
34dc7c2f BB |
1298 | return (zio); |
1299 | } | |
1300 | ||
1301 | zio_t * | |
a6255b7f | 1302 | zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data, |
e8b96c60 | 1303 | uint64_t size, zio_done_func_t *done, void *private, |
4938d01d | 1304 | zio_priority_t priority, zio_flag_t flags, zbookmark_phys_t *zb) |
34dc7c2f BB |
1305 | { |
1306 | zio_t *zio; | |
1307 | ||
2aa34383 | 1308 | zio = zio_create(pio, spa, txg, bp, data, size, size, done, private, |
3dfb57a3 | 1309 | ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb, |
b128c09f | 1310 | ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); |
34dc7c2f BB |
1311 | |
1312 | return (zio); | |
1313 | } | |
1314 | ||
428870ff | 1315 | void |
67a1b037 PJD |
1316 | zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite, |
1317 | boolean_t brtwrite) | |
428870ff BB |
1318 | { |
1319 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
1320 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1321 | ASSERT(zio->io_stage == ZIO_STAGE_OPEN); | |
1322 | ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); | |
67a1b037 | 1323 | ASSERT(!brtwrite || !nopwrite); |
428870ff | 1324 | |
03c6040b GW |
1325 | /* |
1326 | * We must reset the io_prop to match the values that existed | |
1327 | * when the bp was first written by dmu_sync() keeping in mind | |
1328 | * that nopwrite and dedup are mutually exclusive. | |
1329 | */ | |
1330 | zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; | |
1331 | zio->io_prop.zp_nopwrite = nopwrite; | |
67a1b037 | 1332 | zio->io_prop.zp_brtwrite = brtwrite; |
428870ff BB |
1333 | zio->io_prop.zp_copies = copies; |
1334 | zio->io_bp_override = bp; | |
1335 | } | |
1336 | ||
1337 | void | |
1338 | zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) | |
1339 | { | |
9b67f605 | 1340 | |
3095ca91 | 1341 | (void) zfs_blkptr_verify(spa, bp, BLK_CONFIG_NEEDED, BLK_VERIFY_HALT); |
a1d477c2 | 1342 | |
9b67f605 MA |
1343 | /* |
1344 | * The check for EMBEDDED is a performance optimization. We | |
1345 | * process the free here (by ignoring it) rather than | |
1346 | * putting it on the list and then processing it in zio_free_sync(). | |
1347 | */ | |
1348 | if (BP_IS_EMBEDDED(bp)) | |
1349 | return; | |
2883cad5 MA |
1350 | |
1351 | /* | |
1352 | * Frees that are for the currently-syncing txg, are not going to be | |
1353 | * deferred, and which will not need to do a read (i.e. not GANG or | |
1354 | * DEDUP), can be processed immediately. Otherwise, put them on the | |
1355 | * in-memory list for later processing. | |
93e28d66 SD |
1356 | * |
1357 | * Note that we only defer frees after zfs_sync_pass_deferred_free | |
1358 | * when the log space map feature is disabled. [see relevant comment | |
1359 | * in spa_sync_iterate_to_convergence()] | |
2883cad5 | 1360 | */ |
93e28d66 SD |
1361 | if (BP_IS_GANG(bp) || |
1362 | BP_GET_DEDUP(bp) || | |
2883cad5 | 1363 | txg != spa->spa_syncing_txg || |
93e28d66 | 1364 | (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free && |
67a1b037 PJD |
1365 | !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) || |
1366 | brt_maybe_exists(spa, bp)) { | |
6694ca55 | 1367 | metaslab_check_free(spa, bp); |
2883cad5 MA |
1368 | bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); |
1369 | } else { | |
9cdf7b1f | 1370 | VERIFY3P(zio_free_sync(NULL, spa, txg, bp, 0), ==, NULL); |
2883cad5 | 1371 | } |
428870ff BB |
1372 | } |
1373 | ||
9cdf7b1f MA |
1374 | /* |
1375 | * To improve performance, this function may return NULL if we were able | |
1376 | * to do the free immediately. This avoids the cost of creating a zio | |
1377 | * (and linking it to the parent, etc). | |
1378 | */ | |
34dc7c2f | 1379 | zio_t * |
428870ff | 1380 | zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
4938d01d | 1381 | zio_flag_t flags) |
34dc7c2f | 1382 | { |
428870ff BB |
1383 | ASSERT(!BP_IS_HOLE(bp)); |
1384 | ASSERT(spa_syncing_txg(spa) == txg); | |
34dc7c2f | 1385 | |
9b67f605 | 1386 | if (BP_IS_EMBEDDED(bp)) |
9cdf7b1f | 1387 | return (NULL); |
9b67f605 | 1388 | |
13fe0198 | 1389 | metaslab_check_free(spa, bp); |
8c841793 | 1390 | arc_freed(spa, bp); |
d4a72f23 | 1391 | dsl_scan_freed(spa, bp); |
13fe0198 | 1392 | |
67a1b037 PJD |
1393 | if (BP_IS_GANG(bp) || |
1394 | BP_GET_DEDUP(bp) || | |
1395 | brt_maybe_exists(spa, bp)) { | |
9cdf7b1f | 1396 | /* |
67a1b037 PJD |
1397 | * GANG, DEDUP and BRT blocks can induce a read (for the gang |
1398 | * block header, the DDT or the BRT), so issue them | |
1399 | * asynchronously so that this thread is not tied up. | |
9cdf7b1f MA |
1400 | */ |
1401 | enum zio_stage stage = | |
1402 | ZIO_FREE_PIPELINE | ZIO_STAGE_ISSUE_ASYNC; | |
2883cad5 | 1403 | |
9cdf7b1f MA |
1404 | return (zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
1405 | BP_GET_PSIZE(bp), NULL, NULL, | |
1406 | ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, | |
1407 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage)); | |
1408 | } else { | |
1409 | metaslab_free(spa, bp, txg, B_FALSE); | |
1410 | return (NULL); | |
1411 | } | |
34dc7c2f BB |
1412 | } |
1413 | ||
1414 | zio_t * | |
428870ff | 1415 | zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
4938d01d | 1416 | zio_done_func_t *done, void *private, zio_flag_t flags) |
34dc7c2f BB |
1417 | { |
1418 | zio_t *zio; | |
1419 | ||
3095ca91 MA |
1420 | (void) zfs_blkptr_verify(spa, bp, (flags & ZIO_FLAG_CONFIG_WRITER) ? |
1421 | BLK_CONFIG_HELD : BLK_CONFIG_NEEDED, BLK_VERIFY_HALT); | |
9b67f605 MA |
1422 | |
1423 | if (BP_IS_EMBEDDED(bp)) | |
1424 | return (zio_null(pio, spa, NULL, NULL, NULL, 0)); | |
1425 | ||
34dc7c2f BB |
1426 | /* |
1427 | * A claim is an allocation of a specific block. Claims are needed | |
1428 | * to support immediate writes in the intent log. The issue is that | |
1429 | * immediate writes contain committed data, but in a txg that was | |
1430 | * *not* committed. Upon opening the pool after an unclean shutdown, | |
1431 | * the intent log claims all blocks that contain immediate write data | |
1432 | * so that the SPA knows they're in use. | |
1433 | * | |
1434 | * All claims *must* be resolved in the first txg -- before the SPA | |
1435 | * starts allocating blocks -- so that nothing is allocated twice. | |
428870ff | 1436 | * If txg == 0 we just verify that the block is claimable. |
34dc7c2f | 1437 | */ |
d2734cce SD |
1438 | ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, |
1439 | spa_min_claim_txg(spa)); | |
1440 | ASSERT(txg == spa_min_claim_txg(spa) || txg == 0); | |
76d04993 | 1441 | ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(8) */ |
34dc7c2f | 1442 | |
b128c09f | 1443 | zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
2aa34383 DK |
1444 | BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, |
1445 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); | |
3dfb57a3 | 1446 | ASSERT0(zio->io_queued_timestamp); |
34dc7c2f BB |
1447 | |
1448 | return (zio); | |
1449 | } | |
1450 | ||
1451 | zio_t * | |
1452 | zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, | |
4938d01d | 1453 | zio_done_func_t *done, void *private, zio_flag_t flags) |
34dc7c2f | 1454 | { |
5a3bffab AM |
1455 | zio_t *zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
1456 | ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, | |
1457 | ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); | |
1458 | zio->io_cmd = cmd; | |
34dc7c2f BB |
1459 | return (zio); |
1460 | } | |
1461 | ||
1b939560 BB |
1462 | zio_t * |
1463 | zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
1464 | zio_done_func_t *done, void *private, zio_priority_t priority, | |
4938d01d | 1465 | zio_flag_t flags, enum trim_flag trim_flags) |
1b939560 BB |
1466 | { |
1467 | zio_t *zio; | |
1468 | ||
1469 | ASSERT0(vd->vdev_children); | |
1470 | ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift)); | |
1471 | ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift)); | |
1472 | ASSERT3U(size, !=, 0); | |
1473 | ||
1474 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done, | |
1475 | private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL, | |
1476 | vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE); | |
1477 | zio->io_trim_flags = trim_flags; | |
1478 | ||
1479 | return (zio); | |
1480 | } | |
1481 | ||
34dc7c2f BB |
1482 | zio_t * |
1483 | zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
a6255b7f | 1484 | abd_t *data, int checksum, zio_done_func_t *done, void *private, |
4938d01d | 1485 | zio_priority_t priority, zio_flag_t flags, boolean_t labels) |
34dc7c2f BB |
1486 | { |
1487 | zio_t *zio; | |
34dc7c2f | 1488 | |
b128c09f BB |
1489 | ASSERT(vd->vdev_children == 0); |
1490 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
1491 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
1492 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 1493 | |
2aa34383 DK |
1494 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
1495 | private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
1496 | offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); | |
34dc7c2f | 1497 | |
b128c09f | 1498 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f BB |
1499 | |
1500 | return (zio); | |
1501 | } | |
1502 | ||
1503 | zio_t * | |
1504 | zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
a6255b7f | 1505 | abd_t *data, int checksum, zio_done_func_t *done, void *private, |
4938d01d | 1506 | zio_priority_t priority, zio_flag_t flags, boolean_t labels) |
34dc7c2f | 1507 | { |
34dc7c2f | 1508 | zio_t *zio; |
34dc7c2f | 1509 | |
b128c09f BB |
1510 | ASSERT(vd->vdev_children == 0); |
1511 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
1512 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
1513 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 1514 | |
2aa34383 DK |
1515 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
1516 | private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
1517 | offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); | |
34dc7c2f | 1518 | |
b128c09f | 1519 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f | 1520 | |
3c67d83a | 1521 | if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { |
34dc7c2f | 1522 | /* |
428870ff | 1523 | * zec checksums are necessarily destructive -- they modify |
b128c09f | 1524 | * the end of the write buffer to hold the verifier/checksum. |
34dc7c2f | 1525 | * Therefore, we must make a local copy in case the data is |
b128c09f | 1526 | * being written to multiple places in parallel. |
34dc7c2f | 1527 | */ |
a6255b7f DQ |
1528 | abd_t *wbuf = abd_alloc_sametype(data, size); |
1529 | abd_copy(wbuf, data, size); | |
1530 | ||
b128c09f | 1531 | zio_push_transform(zio, wbuf, size, size, NULL); |
34dc7c2f BB |
1532 | } |
1533 | ||
1534 | return (zio); | |
1535 | } | |
1536 | ||
1537 | /* | |
b128c09f | 1538 | * Create a child I/O to do some work for us. |
34dc7c2f BB |
1539 | */ |
1540 | zio_t * | |
b128c09f | 1541 | zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, |
4ea3f864 | 1542 | abd_t *data, uint64_t size, int type, zio_priority_t priority, |
4938d01d | 1543 | zio_flag_t flags, zio_done_func_t *done, void *private) |
34dc7c2f | 1544 | { |
428870ff | 1545 | enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; |
b128c09f BB |
1546 | zio_t *zio; |
1547 | ||
a1d477c2 MA |
1548 | /* |
1549 | * vdev child I/Os do not propagate their error to the parent. | |
1550 | * Therefore, for correct operation the caller *must* check for | |
1551 | * and handle the error in the child i/o's done callback. | |
1552 | * The only exceptions are i/os that we don't care about | |
1553 | * (OPTIONAL or REPAIR). | |
1554 | */ | |
1555 | ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) || | |
1556 | done != NULL); | |
1557 | ||
34dc7c2f BB |
1558 | if (type == ZIO_TYPE_READ && bp != NULL) { |
1559 | /* | |
1560 | * If we have the bp, then the child should perform the | |
1561 | * checksum and the parent need not. This pushes error | |
1562 | * detection as close to the leaves as possible and | |
1563 | * eliminates redundant checksums in the interior nodes. | |
1564 | */ | |
428870ff BB |
1565 | pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; |
1566 | pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; | |
34dc7c2f BB |
1567 | } |
1568 | ||
a1d477c2 MA |
1569 | if (vd->vdev_ops->vdev_op_leaf) { |
1570 | ASSERT0(vd->vdev_children); | |
b128c09f | 1571 | offset += VDEV_LABEL_START_SIZE; |
a1d477c2 | 1572 | } |
b128c09f | 1573 | |
a1d477c2 | 1574 | flags |= ZIO_VDEV_CHILD_FLAGS(pio); |
428870ff BB |
1575 | |
1576 | /* | |
1577 | * If we've decided to do a repair, the write is not speculative -- | |
1578 | * even if the original read was. | |
1579 | */ | |
1580 | if (flags & ZIO_FLAG_IO_REPAIR) | |
1581 | flags &= ~ZIO_FLAG_SPECULATIVE; | |
1582 | ||
3dfb57a3 DB |
1583 | /* |
1584 | * If we're creating a child I/O that is not associated with a | |
1585 | * top-level vdev, then the child zio is not an allocating I/O. | |
1586 | * If this is a retried I/O then we ignore it since we will | |
1587 | * have already processed the original allocating I/O. | |
1588 | */ | |
1589 | if (flags & ZIO_FLAG_IO_ALLOCATING && | |
1590 | (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) { | |
cc99f275 DB |
1591 | ASSERT(pio->io_metaslab_class != NULL); |
1592 | ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled); | |
3dfb57a3 DB |
1593 | ASSERT(type == ZIO_TYPE_WRITE); |
1594 | ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE); | |
1595 | ASSERT(!(flags & ZIO_FLAG_IO_REPAIR)); | |
1596 | ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) || | |
1597 | pio->io_child_type == ZIO_CHILD_GANG); | |
1598 | ||
1599 | flags &= ~ZIO_FLAG_IO_ALLOCATING; | |
1600 | } | |
1601 | ||
2aa34383 | 1602 | zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size, |
428870ff BB |
1603 | done, private, type, priority, flags, vd, offset, &pio->io_bookmark, |
1604 | ZIO_STAGE_VDEV_IO_START >> 1, pipeline); | |
3dfb57a3 | 1605 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); |
34dc7c2f | 1606 | |
b128c09f | 1607 | return (zio); |
34dc7c2f BB |
1608 | } |
1609 | ||
b128c09f | 1610 | zio_t * |
a6255b7f | 1611 | zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size, |
4938d01d | 1612 | zio_type_t type, zio_priority_t priority, zio_flag_t flags, |
e9aa730c | 1613 | zio_done_func_t *done, void *private) |
34dc7c2f | 1614 | { |
b128c09f | 1615 | zio_t *zio; |
34dc7c2f | 1616 | |
b128c09f | 1617 | ASSERT(vd->vdev_ops->vdev_op_leaf); |
34dc7c2f | 1618 | |
b128c09f | 1619 | zio = zio_create(NULL, vd->vdev_spa, 0, NULL, |
2aa34383 | 1620 | data, size, size, done, private, type, priority, |
e8b96c60 | 1621 | flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, |
b128c09f | 1622 | vd, offset, NULL, |
428870ff | 1623 | ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); |
34dc7c2f | 1624 | |
b128c09f | 1625 | return (zio); |
34dc7c2f BB |
1626 | } |
1627 | ||
1628 | void | |
5a3bffab | 1629 | zio_flush(zio_t *pio, vdev_t *vd) |
34dc7c2f | 1630 | { |
5a3bffab AM |
1631 | if (vd->vdev_nowritecache) |
1632 | return; | |
1633 | if (vd->vdev_children == 0) { | |
1634 | zio_nowait(zio_ioctl(pio, vd->vdev_spa, vd, | |
1635 | DKIOCFLUSHWRITECACHE, NULL, NULL, ZIO_FLAG_CANFAIL | | |
1636 | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); | |
1637 | } else { | |
1638 | for (uint64_t c = 0; c < vd->vdev_children; c++) | |
1639 | zio_flush(pio, vd->vdev_child[c]); | |
1640 | } | |
34dc7c2f BB |
1641 | } |
1642 | ||
428870ff BB |
1643 | void |
1644 | zio_shrink(zio_t *zio, uint64_t size) | |
1645 | { | |
1ce23dca PS |
1646 | ASSERT3P(zio->io_executor, ==, NULL); |
1647 | ASSERT3U(zio->io_orig_size, ==, zio->io_size); | |
1648 | ASSERT3U(size, <=, zio->io_size); | |
428870ff BB |
1649 | |
1650 | /* | |
1651 | * We don't shrink for raidz because of problems with the | |
1652 | * reconstruction when reading back less than the block size. | |
1653 | * Note, BP_IS_RAIDZ() assumes no compression. | |
1654 | */ | |
1655 | ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); | |
2aa34383 DK |
1656 | if (!BP_IS_RAIDZ(zio->io_bp)) { |
1657 | /* we are not doing a raw write */ | |
1658 | ASSERT3U(zio->io_size, ==, zio->io_lsize); | |
1659 | zio->io_orig_size = zio->io_size = zio->io_lsize = size; | |
1660 | } | |
428870ff BB |
1661 | } |
1662 | ||
d9bb583c AH |
1663 | /* |
1664 | * Round provided allocation size up to a value that can be allocated | |
1665 | * by at least some vdev(s) in the pool with minimum or no additional | |
1666 | * padding and without extra space usage on others | |
1667 | */ | |
1668 | static uint64_t | |
1669 | zio_roundup_alloc_size(spa_t *spa, uint64_t size) | |
1670 | { | |
1671 | if (size > spa->spa_min_alloc) | |
1672 | return (roundup(size, spa->spa_gcd_alloc)); | |
1673 | return (spa->spa_min_alloc); | |
1674 | } | |
1675 | ||
34dc7c2f BB |
1676 | /* |
1677 | * ========================================================================== | |
b128c09f | 1678 | * Prepare to read and write logical blocks |
34dc7c2f BB |
1679 | * ========================================================================== |
1680 | */ | |
b128c09f | 1681 | |
62840030 | 1682 | static zio_t * |
b128c09f | 1683 | zio_read_bp_init(zio_t *zio) |
34dc7c2f | 1684 | { |
b128c09f | 1685 | blkptr_t *bp = zio->io_bp; |
b5256303 TC |
1686 | uint64_t psize = |
1687 | BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); | |
34dc7c2f | 1688 | |
a1d477c2 MA |
1689 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
1690 | ||
fb5f0bc8 | 1691 | if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && |
9babb374 | 1692 | zio->io_child_type == ZIO_CHILD_LOGICAL && |
b5256303 | 1693 | !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { |
a6255b7f DQ |
1694 | zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), |
1695 | psize, psize, zio_decompress); | |
34dc7c2f | 1696 | } |
34dc7c2f | 1697 | |
b5256303 TC |
1698 | if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) || |
1699 | BP_HAS_INDIRECT_MAC_CKSUM(bp)) && | |
1700 | zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
1701 | zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), | |
1702 | psize, psize, zio_decrypt); | |
1703 | } | |
1704 | ||
9b67f605 | 1705 | if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { |
a6255b7f DQ |
1706 | int psize = BPE_GET_PSIZE(bp); |
1707 | void *data = abd_borrow_buf(zio->io_abd, psize); | |
1708 | ||
9b67f605 | 1709 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
a6255b7f DQ |
1710 | decode_embedded_bp_compressed(bp, data); |
1711 | abd_return_buf_copy(zio->io_abd, data, psize); | |
9b67f605 MA |
1712 | } else { |
1713 | ASSERT(!BP_IS_EMBEDDED(bp)); | |
1714 | } | |
1715 | ||
428870ff BB |
1716 | if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) |
1717 | zio->io_pipeline = ZIO_DDT_READ_PIPELINE; | |
1718 | ||
62840030 | 1719 | return (zio); |
34dc7c2f BB |
1720 | } |
1721 | ||
62840030 | 1722 | static zio_t * |
b128c09f | 1723 | zio_write_bp_init(zio_t *zio) |
34dc7c2f | 1724 | { |
b128c09f | 1725 | if (!IO_IS_ALLOCATING(zio)) |
62840030 | 1726 | return (zio); |
34dc7c2f | 1727 | |
428870ff BB |
1728 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1729 | ||
1730 | if (zio->io_bp_override) { | |
3dfb57a3 DB |
1731 | blkptr_t *bp = zio->io_bp; |
1732 | zio_prop_t *zp = &zio->io_prop; | |
1733 | ||
428870ff | 1734 | ASSERT(bp->blk_birth != zio->io_txg); |
428870ff BB |
1735 | |
1736 | *bp = *zio->io_bp_override; | |
1737 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1738 | ||
67a1b037 PJD |
1739 | if (zp->zp_brtwrite) |
1740 | return (zio); | |
1741 | ||
1742 | ASSERT(!BP_GET_DEDUP(zio->io_bp_override)); | |
1743 | ||
9b67f605 | 1744 | if (BP_IS_EMBEDDED(bp)) |
62840030 | 1745 | return (zio); |
9b67f605 | 1746 | |
03c6040b GW |
1747 | /* |
1748 | * If we've been overridden and nopwrite is set then | |
1749 | * set the flag accordingly to indicate that a nopwrite | |
1750 | * has already occurred. | |
1751 | */ | |
1752 | if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { | |
1753 | ASSERT(!zp->zp_dedup); | |
3dfb57a3 | 1754 | ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum); |
03c6040b | 1755 | zio->io_flags |= ZIO_FLAG_NOPWRITE; |
62840030 | 1756 | return (zio); |
03c6040b GW |
1757 | } |
1758 | ||
1759 | ASSERT(!zp->zp_nopwrite); | |
1760 | ||
428870ff | 1761 | if (BP_IS_HOLE(bp) || !zp->zp_dedup) |
62840030 | 1762 | return (zio); |
428870ff | 1763 | |
3c67d83a TH |
1764 | ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & |
1765 | ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); | |
428870ff | 1766 | |
b5256303 TC |
1767 | if (BP_GET_CHECKSUM(bp) == zp->zp_checksum && |
1768 | !zp->zp_encrypt) { | |
428870ff BB |
1769 | BP_SET_DEDUP(bp, 1); |
1770 | zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; | |
62840030 | 1771 | return (zio); |
428870ff | 1772 | } |
3dfb57a3 DB |
1773 | |
1774 | /* | |
1775 | * We were unable to handle this as an override bp, treat | |
1776 | * it as a regular write I/O. | |
1777 | */ | |
5511754b | 1778 | zio->io_bp_override = NULL; |
3dfb57a3 DB |
1779 | *bp = zio->io_bp_orig; |
1780 | zio->io_pipeline = zio->io_orig_pipeline; | |
1781 | } | |
1782 | ||
62840030 | 1783 | return (zio); |
3dfb57a3 DB |
1784 | } |
1785 | ||
62840030 | 1786 | static zio_t * |
3dfb57a3 DB |
1787 | zio_write_compress(zio_t *zio) |
1788 | { | |
1789 | spa_t *spa = zio->io_spa; | |
1790 | zio_prop_t *zp = &zio->io_prop; | |
1791 | enum zio_compress compress = zp->zp_compress; | |
1792 | blkptr_t *bp = zio->io_bp; | |
1793 | uint64_t lsize = zio->io_lsize; | |
1794 | uint64_t psize = zio->io_size; | |
fdc2d303 | 1795 | uint32_t pass = 1; |
3dfb57a3 | 1796 | |
3dfb57a3 DB |
1797 | /* |
1798 | * If our children haven't all reached the ready stage, | |
1799 | * wait for them and then repeat this pipeline stage. | |
1800 | */ | |
ddc751d5 GW |
1801 | if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT | |
1802 | ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) { | |
62840030 | 1803 | return (NULL); |
ddc751d5 | 1804 | } |
3dfb57a3 DB |
1805 | |
1806 | if (!IO_IS_ALLOCATING(zio)) | |
62840030 | 1807 | return (zio); |
3dfb57a3 DB |
1808 | |
1809 | if (zio->io_children_ready != NULL) { | |
1810 | /* | |
1811 | * Now that all our children are ready, run the callback | |
1812 | * associated with this zio in case it wants to modify the | |
1813 | * data to be written. | |
1814 | */ | |
1815 | ASSERT3U(zp->zp_level, >, 0); | |
1816 | zio->io_children_ready(zio); | |
428870ff | 1817 | } |
34dc7c2f | 1818 | |
3dfb57a3 DB |
1819 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1820 | ASSERT(zio->io_bp_override == NULL); | |
1821 | ||
b0bc7a84 | 1822 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { |
b128c09f BB |
1823 | /* |
1824 | * We're rewriting an existing block, which means we're | |
1825 | * working on behalf of spa_sync(). For spa_sync() to | |
1826 | * converge, it must eventually be the case that we don't | |
1827 | * have to allocate new blocks. But compression changes | |
1828 | * the blocksize, which forces a reallocate, and makes | |
1829 | * convergence take longer. Therefore, after the first | |
1830 | * few passes, stop compressing to ensure convergence. | |
1831 | */ | |
428870ff BB |
1832 | pass = spa_sync_pass(spa); |
1833 | ||
1834 | ASSERT(zio->io_txg == spa_syncing_txg(spa)); | |
1835 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1836 | ASSERT(!BP_GET_DEDUP(bp)); | |
34dc7c2f | 1837 | |
55d85d5a | 1838 | if (pass >= zfs_sync_pass_dont_compress) |
b128c09f | 1839 | compress = ZIO_COMPRESS_OFF; |
34dc7c2f | 1840 | |
b128c09f | 1841 | /* Make sure someone doesn't change their mind on overwrites */ |
ed39d668 SD |
1842 | ASSERT(BP_IS_EMBEDDED(bp) || BP_IS_GANG(bp) || |
1843 | MIN(zp->zp_copies, spa_max_replication(spa)) | |
1844 | == BP_GET_NDVAS(bp)); | |
b128c09f | 1845 | } |
34dc7c2f | 1846 | |
2aa34383 | 1847 | /* If it's a compressed write that is not raw, compress the buffer. */ |
b5256303 TC |
1848 | if (compress != ZIO_COMPRESS_OFF && |
1849 | !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { | |
bff26b02 RY |
1850 | void *cbuf = NULL; |
1851 | psize = zio_compress_data(compress, zio->io_abd, &cbuf, lsize, | |
10b3c7f5 | 1852 | zp->zp_complevel); |
bff26b02 | 1853 | if (psize == 0) { |
b128c09f | 1854 | compress = ZIO_COMPRESS_OFF; |
bff26b02 RY |
1855 | } else if (psize >= lsize) { |
1856 | compress = ZIO_COMPRESS_OFF; | |
1857 | if (cbuf != NULL) | |
1858 | zio_buf_free(cbuf, lsize); | |
b5256303 TC |
1859 | } else if (!zp->zp_dedup && !zp->zp_encrypt && |
1860 | psize <= BPE_PAYLOAD_SIZE && | |
9b67f605 MA |
1861 | zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && |
1862 | spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { | |
1863 | encode_embedded_bp_compressed(bp, | |
1864 | cbuf, compress, lsize, psize); | |
1865 | BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); | |
1866 | BP_SET_TYPE(bp, zio->io_prop.zp_type); | |
1867 | BP_SET_LEVEL(bp, zio->io_prop.zp_level); | |
1868 | zio_buf_free(cbuf, lsize); | |
1869 | bp->blk_birth = zio->io_txg; | |
1870 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1871 | ASSERT(spa_feature_is_active(spa, | |
1872 | SPA_FEATURE_EMBEDDED_DATA)); | |
62840030 | 1873 | return (zio); |
428870ff | 1874 | } else { |
9b67f605 | 1875 | /* |
b2255edc BB |
1876 | * Round compressed size up to the minimum allocation |
1877 | * size of the smallest-ashift device, and zero the | |
1878 | * tail. This ensures that the compressed size of the | |
1879 | * BP (and thus compressratio property) are correct, | |
c3520e7f MA |
1880 | * in that we charge for the padding used to fill out |
1881 | * the last sector. | |
9b67f605 | 1882 | */ |
d9bb583c AH |
1883 | size_t rounded = (size_t)zio_roundup_alloc_size(spa, |
1884 | psize); | |
c3520e7f | 1885 | if (rounded >= lsize) { |
9b67f605 MA |
1886 | compress = ZIO_COMPRESS_OFF; |
1887 | zio_buf_free(cbuf, lsize); | |
c3520e7f | 1888 | psize = lsize; |
9b67f605 | 1889 | } else { |
a6255b7f DQ |
1890 | abd_t *cdata = abd_get_from_buf(cbuf, lsize); |
1891 | abd_take_ownership_of_buf(cdata, B_TRUE); | |
1892 | abd_zero_off(cdata, psize, rounded - psize); | |
c3520e7f | 1893 | psize = rounded; |
a6255b7f | 1894 | zio_push_transform(zio, cdata, |
9b67f605 MA |
1895 | psize, lsize, NULL); |
1896 | } | |
b128c09f | 1897 | } |
3dfb57a3 DB |
1898 | |
1899 | /* | |
1900 | * We were unable to handle this as an override bp, treat | |
1901 | * it as a regular write I/O. | |
1902 | */ | |
1903 | zio->io_bp_override = NULL; | |
1904 | *bp = zio->io_bp_orig; | |
1905 | zio->io_pipeline = zio->io_orig_pipeline; | |
1906 | ||
b1d21733 TC |
1907 | } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 && |
1908 | zp->zp_type == DMU_OT_DNODE) { | |
1909 | /* | |
1910 | * The DMU actually relies on the zio layer's compression | |
1911 | * to free metadnode blocks that have had all contained | |
1912 | * dnodes freed. As a result, even when doing a raw | |
1913 | * receive, we must check whether the block can be compressed | |
1914 | * to a hole. | |
1915 | */ | |
1916 | psize = zio_compress_data(ZIO_COMPRESS_EMPTY, | |
10b3c7f5 MN |
1917 | zio->io_abd, NULL, lsize, zp->zp_complevel); |
1918 | if (psize == 0 || psize >= lsize) | |
b1d21733 | 1919 | compress = ZIO_COMPRESS_OFF; |
52a36bd4 GA |
1920 | } else if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS && |
1921 | !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) { | |
1922 | /* | |
1923 | * If we are raw receiving an encrypted dataset we should not | |
1924 | * take this codepath because it will change the on-disk block | |
1925 | * and decryption will fail. | |
1926 | */ | |
d9bb583c AH |
1927 | size_t rounded = MIN((size_t)zio_roundup_alloc_size(spa, psize), |
1928 | lsize); | |
c634320e PD |
1929 | |
1930 | if (rounded != psize) { | |
1931 | abd_t *cdata = abd_alloc_linear(rounded, B_TRUE); | |
1932 | abd_zero_off(cdata, psize, rounded - psize); | |
1933 | abd_copy_off(cdata, zio->io_abd, 0, 0, psize); | |
1934 | psize = rounded; | |
1935 | zio_push_transform(zio, cdata, | |
1936 | psize, rounded, NULL); | |
1937 | } | |
2aa34383 DK |
1938 | } else { |
1939 | ASSERT3U(psize, !=, 0); | |
b128c09f | 1940 | } |
34dc7c2f | 1941 | |
b128c09f BB |
1942 | /* |
1943 | * The final pass of spa_sync() must be all rewrites, but the first | |
1944 | * few passes offer a trade-off: allocating blocks defers convergence, | |
1945 | * but newly allocated blocks are sequential, so they can be written | |
1946 | * to disk faster. Therefore, we allow the first few passes of | |
1947 | * spa_sync() to allocate new blocks, but force rewrites after that. | |
1948 | * There should only be a handful of blocks after pass 1 in any case. | |
1949 | */ | |
b0bc7a84 MG |
1950 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && |
1951 | BP_GET_PSIZE(bp) == psize && | |
55d85d5a | 1952 | pass >= zfs_sync_pass_rewrite) { |
cc99f275 | 1953 | VERIFY3U(psize, !=, 0); |
1c27024e | 1954 | enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; |
cc99f275 | 1955 | |
b128c09f BB |
1956 | zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; |
1957 | zio->io_flags |= ZIO_FLAG_IO_REWRITE; | |
1958 | } else { | |
1959 | BP_ZERO(bp); | |
1960 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
1961 | } | |
34dc7c2f | 1962 | |
428870ff | 1963 | if (psize == 0) { |
b0bc7a84 MG |
1964 | if (zio->io_bp_orig.blk_birth != 0 && |
1965 | spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { | |
1966 | BP_SET_LSIZE(bp, lsize); | |
1967 | BP_SET_TYPE(bp, zp->zp_type); | |
1968 | BP_SET_LEVEL(bp, zp->zp_level); | |
1969 | BP_SET_BIRTH(bp, zio->io_txg, 0); | |
1970 | } | |
b128c09f BB |
1971 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
1972 | } else { | |
1973 | ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); | |
1974 | BP_SET_LSIZE(bp, lsize); | |
b0bc7a84 MG |
1975 | BP_SET_TYPE(bp, zp->zp_type); |
1976 | BP_SET_LEVEL(bp, zp->zp_level); | |
428870ff | 1977 | BP_SET_PSIZE(bp, psize); |
b128c09f BB |
1978 | BP_SET_COMPRESS(bp, compress); |
1979 | BP_SET_CHECKSUM(bp, zp->zp_checksum); | |
428870ff | 1980 | BP_SET_DEDUP(bp, zp->zp_dedup); |
b128c09f | 1981 | BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); |
428870ff BB |
1982 | if (zp->zp_dedup) { |
1983 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1984 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
b5256303 TC |
1985 | ASSERT(!zp->zp_encrypt || |
1986 | DMU_OT_IS_ENCRYPTED(zp->zp_type)); | |
428870ff BB |
1987 | zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; |
1988 | } | |
03c6040b GW |
1989 | if (zp->zp_nopwrite) { |
1990 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1991 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
1992 | zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; | |
1993 | } | |
428870ff | 1994 | } |
62840030 | 1995 | return (zio); |
428870ff BB |
1996 | } |
1997 | ||
62840030 | 1998 | static zio_t * |
428870ff BB |
1999 | zio_free_bp_init(zio_t *zio) |
2000 | { | |
2001 | blkptr_t *bp = zio->io_bp; | |
2002 | ||
2003 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
2004 | if (BP_GET_DEDUP(bp)) | |
2005 | zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; | |
b128c09f | 2006 | } |
34dc7c2f | 2007 | |
a1d477c2 MA |
2008 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
2009 | ||
62840030 | 2010 | return (zio); |
34dc7c2f BB |
2011 | } |
2012 | ||
b128c09f BB |
2013 | /* |
2014 | * ========================================================================== | |
2015 | * Execute the I/O pipeline | |
2016 | * ========================================================================== | |
2017 | */ | |
2018 | ||
2019 | static void | |
7ef5e54e | 2020 | zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) |
34dc7c2f | 2021 | { |
428870ff | 2022 | spa_t *spa = zio->io_spa; |
b128c09f | 2023 | zio_type_t t = zio->io_type; |
a38718a6 | 2024 | int flags = (cutinline ? TQ_FRONT : 0); |
34dc7c2f BB |
2025 | |
2026 | /* | |
9babb374 BB |
2027 | * If we're a config writer or a probe, the normal issue and |
2028 | * interrupt threads may all be blocked waiting for the config lock. | |
2029 | * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. | |
34dc7c2f | 2030 | */ |
9babb374 | 2031 | if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) |
b128c09f | 2032 | t = ZIO_TYPE_NULL; |
34dc7c2f BB |
2033 | |
2034 | /* | |
b128c09f | 2035 | * A similar issue exists for the L2ARC write thread until L2ARC 2.0. |
34dc7c2f | 2036 | */ |
b128c09f BB |
2037 | if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) |
2038 | t = ZIO_TYPE_NULL; | |
34dc7c2f | 2039 | |
428870ff | 2040 | /* |
7ef5e54e AL |
2041 | * If this is a high priority I/O, then use the high priority taskq if |
2042 | * available. | |
428870ff | 2043 | */ |
18b14b17 GW |
2044 | if ((zio->io_priority == ZIO_PRIORITY_NOW || |
2045 | zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) && | |
7ef5e54e | 2046 | spa->spa_zio_taskq[t][q + 1].stqs_count != 0) |
428870ff BB |
2047 | q++; |
2048 | ||
2049 | ASSERT3U(q, <, ZIO_TASKQ_TYPES); | |
5cc556b4 | 2050 | |
a38718a6 GA |
2051 | /* |
2052 | * NB: We are assuming that the zio can only be dispatched | |
2053 | * to a single taskq at a time. It would be a grievous error | |
2054 | * to dispatch the zio to another taskq at the same time. | |
2055 | */ | |
2056 | ASSERT(taskq_empty_ent(&zio->io_tqent)); | |
23c13c7e | 2057 | spa_taskq_dispatch_ent(spa, t, q, zio_execute, zio, flags, |
3bd4df38 | 2058 | &zio->io_tqent, zio); |
b128c09f | 2059 | } |
34dc7c2f | 2060 | |
b128c09f | 2061 | static boolean_t |
7ef5e54e | 2062 | zio_taskq_member(zio_t *zio, zio_taskq_type_t q) |
b128c09f | 2063 | { |
b128c09f | 2064 | spa_t *spa = zio->io_spa; |
34dc7c2f | 2065 | |
b3212d2f MA |
2066 | taskq_t *tq = taskq_of_curthread(); |
2067 | ||
1c27024e | 2068 | for (zio_type_t t = 0; t < ZIO_TYPES; t++) { |
7ef5e54e AL |
2069 | spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; |
2070 | uint_t i; | |
2071 | for (i = 0; i < tqs->stqs_count; i++) { | |
b3212d2f | 2072 | if (tqs->stqs_taskq[i] == tq) |
7ef5e54e AL |
2073 | return (B_TRUE); |
2074 | } | |
2075 | } | |
34dc7c2f | 2076 | |
b128c09f BB |
2077 | return (B_FALSE); |
2078 | } | |
34dc7c2f | 2079 | |
62840030 | 2080 | static zio_t * |
b128c09f BB |
2081 | zio_issue_async(zio_t *zio) |
2082 | { | |
3bd4df38 | 2083 | ASSERT((zio->io_type != ZIO_TYPE_WRITE) || ZIO_HAS_ALLOCATOR(zio)); |
428870ff | 2084 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); |
62840030 | 2085 | return (NULL); |
34dc7c2f BB |
2086 | } |
2087 | ||
b128c09f | 2088 | void |
23c13c7e | 2089 | zio_interrupt(void *zio) |
34dc7c2f | 2090 | { |
428870ff | 2091 | zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); |
b128c09f | 2092 | } |
34dc7c2f | 2093 | |
d3c1e45b MM |
2094 | void |
2095 | zio_delay_interrupt(zio_t *zio) | |
2096 | { | |
2097 | /* | |
2098 | * The timeout_generic() function isn't defined in userspace, so | |
2099 | * rather than trying to implement the function, the zio delay | |
2100 | * functionality has been disabled for userspace builds. | |
2101 | */ | |
2102 | ||
2103 | #ifdef _KERNEL | |
2104 | /* | |
2105 | * If io_target_timestamp is zero, then no delay has been registered | |
2106 | * for this IO, thus jump to the end of this function and "skip" the | |
2107 | * delay; issuing it directly to the zio layer. | |
2108 | */ | |
2109 | if (zio->io_target_timestamp != 0) { | |
2110 | hrtime_t now = gethrtime(); | |
2111 | ||
2112 | if (now >= zio->io_target_timestamp) { | |
2113 | /* | |
2114 | * This IO has already taken longer than the target | |
2115 | * delay to complete, so we don't want to delay it | |
2116 | * any longer; we "miss" the delay and issue it | |
2117 | * directly to the zio layer. This is likely due to | |
2118 | * the target latency being set to a value less than | |
2119 | * the underlying hardware can satisfy (e.g. delay | |
2120 | * set to 1ms, but the disks take 10ms to complete an | |
2121 | * IO request). | |
2122 | */ | |
2123 | ||
2124 | DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, | |
2125 | hrtime_t, now); | |
2126 | ||
2127 | zio_interrupt(zio); | |
2128 | } else { | |
2129 | taskqid_t tid; | |
2130 | hrtime_t diff = zio->io_target_timestamp - now; | |
2131 | clock_t expire_at_tick = ddi_get_lbolt() + | |
2132 | NSEC_TO_TICK(diff); | |
2133 | ||
2134 | DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, | |
2135 | hrtime_t, now, hrtime_t, diff); | |
2136 | ||
2137 | if (NSEC_TO_TICK(diff) == 0) { | |
2138 | /* Our delay is less than a jiffy - just spin */ | |
2139 | zfs_sleep_until(zio->io_target_timestamp); | |
2140 | zio_interrupt(zio); | |
2141 | } else { | |
2142 | /* | |
2143 | * Use taskq_dispatch_delay() in the place of | |
2144 | * OpenZFS's timeout_generic(). | |
2145 | */ | |
2146 | tid = taskq_dispatch_delay(system_taskq, | |
23c13c7e AL |
2147 | zio_interrupt, zio, TQ_NOSLEEP, |
2148 | expire_at_tick); | |
d3c1e45b MM |
2149 | if (tid == TASKQID_INVALID) { |
2150 | /* | |
2151 | * Couldn't allocate a task. Just | |
2152 | * finish the zio without a delay. | |
2153 | */ | |
2154 | zio_interrupt(zio); | |
2155 | } | |
2156 | } | |
2157 | } | |
2158 | return; | |
2159 | } | |
2160 | #endif | |
2161 | DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); | |
2162 | zio_interrupt(zio); | |
2163 | } | |
2164 | ||
8fb1ede1 | 2165 | static void |
638dd5f4 | 2166 | zio_deadman_impl(zio_t *pio, int ziodepth) |
8fb1ede1 BB |
2167 | { |
2168 | zio_t *cio, *cio_next; | |
2169 | zio_link_t *zl = NULL; | |
2170 | vdev_t *vd = pio->io_vd; | |
2171 | ||
638dd5f4 TC |
2172 | if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) { |
2173 | vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL; | |
8fb1ede1 BB |
2174 | zbookmark_phys_t *zb = &pio->io_bookmark; |
2175 | uint64_t delta = gethrtime() - pio->io_timestamp; | |
2176 | uint64_t failmode = spa_get_deadman_failmode(pio->io_spa); | |
2177 | ||
a887d653 | 2178 | zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu " |
8fb1ede1 | 2179 | "delta=%llu queued=%llu io=%llu " |
8e739b2c RE |
2180 | "path=%s " |
2181 | "last=%llu type=%d " | |
4938d01d | 2182 | "priority=%d flags=0x%llx stage=0x%x " |
8e739b2c RE |
2183 | "pipeline=0x%x pipeline-trace=0x%x " |
2184 | "objset=%llu object=%llu " | |
2185 | "level=%llu blkid=%llu " | |
2186 | "offset=%llu size=%llu " | |
2187 | "error=%d", | |
638dd5f4 | 2188 | ziodepth, pio, pio->io_timestamp, |
8e739b2c RE |
2189 | (u_longlong_t)delta, pio->io_delta, pio->io_delay, |
2190 | vd ? vd->vdev_path : "NULL", | |
2191 | vq ? vq->vq_io_complete_ts : 0, pio->io_type, | |
4938d01d RY |
2192 | pio->io_priority, (u_longlong_t)pio->io_flags, |
2193 | pio->io_stage, pio->io_pipeline, pio->io_pipeline_trace, | |
8e739b2c RE |
2194 | (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object, |
2195 | (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid, | |
2196 | (u_longlong_t)pio->io_offset, (u_longlong_t)pio->io_size, | |
2197 | pio->io_error); | |
1144586b | 2198 | (void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN, |
4f072827 | 2199 | pio->io_spa, vd, zb, pio, 0); |
8fb1ede1 BB |
2200 | |
2201 | if (failmode == ZIO_FAILURE_MODE_CONTINUE && | |
2202 | taskq_empty_ent(&pio->io_tqent)) { | |
2203 | zio_interrupt(pio); | |
2204 | } | |
2205 | } | |
2206 | ||
2207 | mutex_enter(&pio->io_lock); | |
2208 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { | |
2209 | cio_next = zio_walk_children(pio, &zl); | |
638dd5f4 | 2210 | zio_deadman_impl(cio, ziodepth + 1); |
8fb1ede1 BB |
2211 | } |
2212 | mutex_exit(&pio->io_lock); | |
2213 | } | |
2214 | ||
2215 | /* | |
2216 | * Log the critical information describing this zio and all of its children | |
2217 | * using the zfs_dbgmsg() interface then post deadman event for the ZED. | |
2218 | */ | |
2219 | void | |
dd66857d | 2220 | zio_deadman(zio_t *pio, const char *tag) |
8fb1ede1 BB |
2221 | { |
2222 | spa_t *spa = pio->io_spa; | |
2223 | char *name = spa_name(spa); | |
2224 | ||
2225 | if (!zfs_deadman_enabled || spa_suspended(spa)) | |
2226 | return; | |
2227 | ||
638dd5f4 | 2228 | zio_deadman_impl(pio, 0); |
8fb1ede1 BB |
2229 | |
2230 | switch (spa_get_deadman_failmode(spa)) { | |
2231 | case ZIO_FAILURE_MODE_WAIT: | |
2232 | zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name); | |
2233 | break; | |
2234 | ||
2235 | case ZIO_FAILURE_MODE_CONTINUE: | |
2236 | zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name); | |
2237 | break; | |
2238 | ||
2239 | case ZIO_FAILURE_MODE_PANIC: | |
2240 | fm_panic("%s determined I/O to pool '%s' is hung.", tag, name); | |
2241 | break; | |
2242 | } | |
2243 | } | |
2244 | ||
b128c09f BB |
2245 | /* |
2246 | * Execute the I/O pipeline until one of the following occurs: | |
2247 | * (1) the I/O completes; (2) the pipeline stalls waiting for | |
2248 | * dependent child I/Os; (3) the I/O issues, so we're waiting | |
2249 | * for an I/O completion interrupt; (4) the I/O is delegated by | |
2250 | * vdev-level caching or aggregation; (5) the I/O is deferred | |
2251 | * due to vdev-level queueing; (6) the I/O is handed off to | |
2252 | * another thread. In all cases, the pipeline stops whenever | |
8e07b99b | 2253 | * there's no CPU work; it never burns a thread in cv_wait_io(). |
b128c09f BB |
2254 | * |
2255 | * There's no locking on io_stage because there's no legitimate way | |
2256 | * for multiple threads to be attempting to process the same I/O. | |
2257 | */ | |
428870ff | 2258 | static zio_pipe_stage_t *zio_pipeline[]; |
34dc7c2f | 2259 | |
da6b4005 NB |
2260 | /* |
2261 | * zio_execute() is a wrapper around the static function | |
2262 | * __zio_execute() so that we can force __zio_execute() to be | |
2263 | * inlined. This reduces stack overhead which is important | |
2264 | * because __zio_execute() is called recursively in several zio | |
2265 | * code paths. zio_execute() itself cannot be inlined because | |
2266 | * it is externally visible. | |
2267 | */ | |
b128c09f | 2268 | void |
23c13c7e | 2269 | zio_execute(void *zio) |
da6b4005 | 2270 | { |
92119cc2 BB |
2271 | fstrans_cookie_t cookie; |
2272 | ||
2273 | cookie = spl_fstrans_mark(); | |
da6b4005 | 2274 | __zio_execute(zio); |
92119cc2 | 2275 | spl_fstrans_unmark(cookie); |
da6b4005 NB |
2276 | } |
2277 | ||
b58986ee BB |
2278 | /* |
2279 | * Used to determine if in the current context the stack is sized large | |
2280 | * enough to allow zio_execute() to be called recursively. A minimum | |
2281 | * stack size of 16K is required to avoid needing to re-dispatch the zio. | |
2282 | */ | |
65c7cc49 | 2283 | static boolean_t |
b58986ee BB |
2284 | zio_execute_stack_check(zio_t *zio) |
2285 | { | |
2286 | #if !defined(HAVE_LARGE_STACKS) | |
2287 | dsl_pool_t *dp = spa_get_dsl(zio->io_spa); | |
2288 | ||
2289 | /* Executing in txg_sync_thread() context. */ | |
2290 | if (dp && curthread == dp->dp_tx.tx_sync_thread) | |
2291 | return (B_TRUE); | |
2292 | ||
2293 | /* Pool initialization outside of zio_taskq context. */ | |
2294 | if (dp && spa_is_initializing(dp->dp_spa) && | |
2295 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) && | |
2296 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH)) | |
2297 | return (B_TRUE); | |
14e4e3cb AZ |
2298 | #else |
2299 | (void) zio; | |
b58986ee BB |
2300 | #endif /* HAVE_LARGE_STACKS */ |
2301 | ||
2302 | return (B_FALSE); | |
2303 | } | |
2304 | ||
da6b4005 NB |
2305 | __attribute__((always_inline)) |
2306 | static inline void | |
2307 | __zio_execute(zio_t *zio) | |
b128c09f | 2308 | { |
3dfb57a3 DB |
2309 | ASSERT3U(zio->io_queued_timestamp, >, 0); |
2310 | ||
b128c09f | 2311 | while (zio->io_stage < ZIO_STAGE_DONE) { |
428870ff BB |
2312 | enum zio_stage pipeline = zio->io_pipeline; |
2313 | enum zio_stage stage = zio->io_stage; | |
62840030 MA |
2314 | |
2315 | zio->io_executor = curthread; | |
34dc7c2f | 2316 | |
b128c09f | 2317 | ASSERT(!MUTEX_HELD(&zio->io_lock)); |
428870ff BB |
2318 | ASSERT(ISP2(stage)); |
2319 | ASSERT(zio->io_stall == NULL); | |
34dc7c2f | 2320 | |
428870ff BB |
2321 | do { |
2322 | stage <<= 1; | |
2323 | } while ((stage & pipeline) == 0); | |
b128c09f BB |
2324 | |
2325 | ASSERT(stage <= ZIO_STAGE_DONE); | |
34dc7c2f BB |
2326 | |
2327 | /* | |
b128c09f BB |
2328 | * If we are in interrupt context and this pipeline stage |
2329 | * will grab a config lock that is held across I/O, | |
428870ff BB |
2330 | * or may wait for an I/O that needs an interrupt thread |
2331 | * to complete, issue async to avoid deadlock. | |
2332 | * | |
2333 | * For VDEV_IO_START, we cut in line so that the io will | |
2334 | * be sent to disk promptly. | |
34dc7c2f | 2335 | */ |
91579709 BB |
2336 | if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && |
2337 | zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { | |
b58986ee BB |
2338 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? |
2339 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
91579709 BB |
2340 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
2341 | return; | |
2342 | } | |
2343 | ||
2344 | /* | |
b58986ee BB |
2345 | * If the current context doesn't have large enough stacks |
2346 | * the zio must be issued asynchronously to prevent overflow. | |
91579709 | 2347 | */ |
b58986ee BB |
2348 | if (zio_execute_stack_check(zio)) { |
2349 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? | |
2350 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
428870ff | 2351 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
b128c09f | 2352 | return; |
34dc7c2f BB |
2353 | } |
2354 | ||
b128c09f | 2355 | zio->io_stage = stage; |
3dfb57a3 | 2356 | zio->io_pipeline_trace |= zio->io_stage; |
34dc7c2f | 2357 | |
62840030 MA |
2358 | /* |
2359 | * The zio pipeline stage returns the next zio to execute | |
2360 | * (typically the same as this one), or NULL if we should | |
2361 | * stop. | |
2362 | */ | |
2363 | zio = zio_pipeline[highbit64(stage) - 1](zio); | |
34dc7c2f | 2364 | |
62840030 MA |
2365 | if (zio == NULL) |
2366 | return; | |
b128c09f | 2367 | } |
34dc7c2f BB |
2368 | } |
2369 | ||
da6b4005 | 2370 | |
b128c09f BB |
2371 | /* |
2372 | * ========================================================================== | |
2373 | * Initiate I/O, either sync or async | |
2374 | * ========================================================================== | |
2375 | */ | |
2376 | int | |
2377 | zio_wait(zio_t *zio) | |
34dc7c2f | 2378 | { |
9cdf7b1f MA |
2379 | /* |
2380 | * Some routines, like zio_free_sync(), may return a NULL zio | |
2381 | * to avoid the performance overhead of creating and then destroying | |
2382 | * an unneeded zio. For the callers' simplicity, we accept a NULL | |
2383 | * zio and ignore it. | |
2384 | */ | |
2385 | if (zio == NULL) | |
2386 | return (0); | |
2387 | ||
8fb1ede1 | 2388 | long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms); |
b128c09f | 2389 | int error; |
34dc7c2f | 2390 | |
1ce23dca PS |
2391 | ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN); |
2392 | ASSERT3P(zio->io_executor, ==, NULL); | |
34dc7c2f | 2393 | |
b128c09f | 2394 | zio->io_waiter = curthread; |
3dfb57a3 DB |
2395 | ASSERT0(zio->io_queued_timestamp); |
2396 | zio->io_queued_timestamp = gethrtime(); | |
34dc7c2f | 2397 | |
3bd4df38 EN |
2398 | if (zio->io_type == ZIO_TYPE_WRITE) { |
2399 | spa_select_allocator(zio); | |
2400 | } | |
da6b4005 | 2401 | __zio_execute(zio); |
34dc7c2f | 2402 | |
b128c09f | 2403 | mutex_enter(&zio->io_lock); |
8fb1ede1 BB |
2404 | while (zio->io_executor != NULL) { |
2405 | error = cv_timedwait_io(&zio->io_cv, &zio->io_lock, | |
2406 | ddi_get_lbolt() + timeout); | |
2407 | ||
2408 | if (zfs_deadman_enabled && error == -1 && | |
2409 | gethrtime() - zio->io_queued_timestamp > | |
2410 | spa_deadman_ziotime(zio->io_spa)) { | |
2411 | mutex_exit(&zio->io_lock); | |
2412 | timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms); | |
2413 | zio_deadman(zio, FTAG); | |
2414 | mutex_enter(&zio->io_lock); | |
2415 | } | |
2416 | } | |
b128c09f | 2417 | mutex_exit(&zio->io_lock); |
34dc7c2f | 2418 | |
b128c09f BB |
2419 | error = zio->io_error; |
2420 | zio_destroy(zio); | |
34dc7c2f | 2421 | |
b128c09f BB |
2422 | return (error); |
2423 | } | |
34dc7c2f | 2424 | |
b128c09f BB |
2425 | void |
2426 | zio_nowait(zio_t *zio) | |
2427 | { | |
9cdf7b1f MA |
2428 | /* |
2429 | * See comment in zio_wait(). | |
2430 | */ | |
2431 | if (zio == NULL) | |
2432 | return; | |
2433 | ||
1ce23dca | 2434 | ASSERT3P(zio->io_executor, ==, NULL); |
34dc7c2f | 2435 | |
d164b209 | 2436 | if (zio->io_child_type == ZIO_CHILD_LOGICAL && |
b035f2b2 | 2437 | list_is_empty(&zio->io_parent_list)) { |
8878261f BB |
2438 | zio_t *pio; |
2439 | ||
34dc7c2f | 2440 | /* |
b128c09f | 2441 | * This is a logical async I/O with no parent to wait for it. |
9babb374 BB |
2442 | * We add it to the spa_async_root_zio "Godfather" I/O which |
2443 | * will ensure they complete prior to unloading the pool. | |
34dc7c2f | 2444 | */ |
b128c09f | 2445 | spa_t *spa = zio->io_spa; |
09eb36ce | 2446 | pio = spa->spa_async_zio_root[CPU_SEQID_UNSTABLE]; |
9babb374 | 2447 | |
8878261f | 2448 | zio_add_child(pio, zio); |
b128c09f | 2449 | } |
34dc7c2f | 2450 | |
3dfb57a3 DB |
2451 | ASSERT0(zio->io_queued_timestamp); |
2452 | zio->io_queued_timestamp = gethrtime(); | |
3bd4df38 EN |
2453 | if (zio->io_type == ZIO_TYPE_WRITE) { |
2454 | spa_select_allocator(zio); | |
2455 | } | |
da6b4005 | 2456 | __zio_execute(zio); |
b128c09f | 2457 | } |
34dc7c2f | 2458 | |
b128c09f BB |
2459 | /* |
2460 | * ========================================================================== | |
1ce23dca | 2461 | * Reexecute, cancel, or suspend/resume failed I/O |
b128c09f BB |
2462 | * ========================================================================== |
2463 | */ | |
34dc7c2f | 2464 | |
b128c09f | 2465 | static void |
23c13c7e | 2466 | zio_reexecute(void *arg) |
b128c09f | 2467 | { |
23c13c7e | 2468 | zio_t *pio = arg; |
3afdc97d | 2469 | zio_t *cio, *cio_next, *gio; |
d164b209 BB |
2470 | |
2471 | ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); | |
2472 | ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); | |
9babb374 BB |
2473 | ASSERT(pio->io_gang_leader == NULL); |
2474 | ASSERT(pio->io_gang_tree == NULL); | |
34dc7c2f | 2475 | |
3afdc97d | 2476 | mutex_enter(&pio->io_lock); |
b128c09f BB |
2477 | pio->io_flags = pio->io_orig_flags; |
2478 | pio->io_stage = pio->io_orig_stage; | |
2479 | pio->io_pipeline = pio->io_orig_pipeline; | |
2480 | pio->io_reexecute = 0; | |
03c6040b | 2481 | pio->io_flags |= ZIO_FLAG_REEXECUTED; |
3dfb57a3 | 2482 | pio->io_pipeline_trace = 0; |
b128c09f | 2483 | pio->io_error = 0; |
3afdc97d AM |
2484 | pio->io_state[ZIO_WAIT_READY] = (pio->io_stage >= ZIO_STAGE_READY) || |
2485 | (pio->io_pipeline & ZIO_STAGE_READY) == 0; | |
2486 | pio->io_state[ZIO_WAIT_DONE] = (pio->io_stage >= ZIO_STAGE_DONE); | |
2487 | zio_link_t *zl = NULL; | |
2488 | while ((gio = zio_walk_parents(pio, &zl)) != NULL) { | |
2489 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) { | |
2490 | gio->io_children[pio->io_child_type][w] += | |
2491 | !pio->io_state[w]; | |
2492 | } | |
2493 | } | |
1c27024e | 2494 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) |
b128c09f | 2495 | pio->io_child_error[c] = 0; |
34dc7c2f | 2496 | |
428870ff BB |
2497 | if (IO_IS_ALLOCATING(pio)) |
2498 | BP_ZERO(pio->io_bp); | |
34dc7c2f | 2499 | |
b128c09f BB |
2500 | /* |
2501 | * As we reexecute pio's children, new children could be created. | |
d164b209 | 2502 | * New children go to the head of pio's io_child_list, however, |
b128c09f | 2503 | * so we will (correctly) not reexecute them. The key is that |
d164b209 BB |
2504 | * the remainder of pio's io_child_list, from 'cio_next' onward, |
2505 | * cannot be affected by any side effects of reexecuting 'cio'. | |
b128c09f | 2506 | */ |
3afdc97d | 2507 | zl = NULL; |
3dfb57a3 DB |
2508 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { |
2509 | cio_next = zio_walk_children(pio, &zl); | |
b128c09f | 2510 | mutex_exit(&pio->io_lock); |
d164b209 | 2511 | zio_reexecute(cio); |
a8b2e306 | 2512 | mutex_enter(&pio->io_lock); |
34dc7c2f | 2513 | } |
a8b2e306 | 2514 | mutex_exit(&pio->io_lock); |
34dc7c2f | 2515 | |
b128c09f BB |
2516 | /* |
2517 | * Now that all children have been reexecuted, execute the parent. | |
9babb374 | 2518 | * We don't reexecute "The Godfather" I/O here as it's the |
9e2c3bb4 | 2519 | * responsibility of the caller to wait on it. |
b128c09f | 2520 | */ |
3dfb57a3 DB |
2521 | if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) { |
2522 | pio->io_queued_timestamp = gethrtime(); | |
da6b4005 | 2523 | __zio_execute(pio); |
3dfb57a3 | 2524 | } |
34dc7c2f BB |
2525 | } |
2526 | ||
b128c09f | 2527 | void |
cec3a0a1 | 2528 | zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason) |
34dc7c2f | 2529 | { |
b128c09f BB |
2530 | if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) |
2531 | fm_panic("Pool '%s' has encountered an uncorrectable I/O " | |
2532 | "failure and the failure mode property for this pool " | |
2533 | "is set to panic.", spa_name(spa)); | |
34dc7c2f | 2534 | |
bf89c199 BB |
2535 | cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O " |
2536 | "failure and has been suspended.\n", spa_name(spa)); | |
2537 | ||
1144586b | 2538 | (void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, |
4f072827 | 2539 | NULL, NULL, 0); |
34dc7c2f | 2540 | |
b128c09f | 2541 | mutex_enter(&spa->spa_suspend_lock); |
34dc7c2f | 2542 | |
b128c09f | 2543 | if (spa->spa_suspend_zio_root == NULL) |
9babb374 BB |
2544 | spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, |
2545 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
2546 | ZIO_FLAG_GODFATHER); | |
34dc7c2f | 2547 | |
cec3a0a1 | 2548 | spa->spa_suspended = reason; |
34dc7c2f | 2549 | |
b128c09f | 2550 | if (zio != NULL) { |
9babb374 | 2551 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f BB |
2552 | ASSERT(zio != spa->spa_suspend_zio_root); |
2553 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
d164b209 | 2554 | ASSERT(zio_unique_parent(zio) == NULL); |
b128c09f BB |
2555 | ASSERT(zio->io_stage == ZIO_STAGE_DONE); |
2556 | zio_add_child(spa->spa_suspend_zio_root, zio); | |
2557 | } | |
34dc7c2f | 2558 | |
b128c09f BB |
2559 | mutex_exit(&spa->spa_suspend_lock); |
2560 | } | |
34dc7c2f | 2561 | |
9babb374 | 2562 | int |
b128c09f BB |
2563 | zio_resume(spa_t *spa) |
2564 | { | |
9babb374 | 2565 | zio_t *pio; |
34dc7c2f BB |
2566 | |
2567 | /* | |
b128c09f | 2568 | * Reexecute all previously suspended i/o. |
34dc7c2f | 2569 | */ |
b128c09f | 2570 | mutex_enter(&spa->spa_suspend_lock); |
cec3a0a1 | 2571 | spa->spa_suspended = ZIO_SUSPEND_NONE; |
b128c09f BB |
2572 | cv_broadcast(&spa->spa_suspend_cv); |
2573 | pio = spa->spa_suspend_zio_root; | |
2574 | spa->spa_suspend_zio_root = NULL; | |
2575 | mutex_exit(&spa->spa_suspend_lock); | |
2576 | ||
2577 | if (pio == NULL) | |
9babb374 | 2578 | return (0); |
34dc7c2f | 2579 | |
9babb374 BB |
2580 | zio_reexecute(pio); |
2581 | return (zio_wait(pio)); | |
b128c09f BB |
2582 | } |
2583 | ||
2584 | void | |
2585 | zio_resume_wait(spa_t *spa) | |
2586 | { | |
2587 | mutex_enter(&spa->spa_suspend_lock); | |
2588 | while (spa_suspended(spa)) | |
2589 | cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); | |
2590 | mutex_exit(&spa->spa_suspend_lock); | |
34dc7c2f BB |
2591 | } |
2592 | ||
2593 | /* | |
2594 | * ========================================================================== | |
b128c09f BB |
2595 | * Gang blocks. |
2596 | * | |
2597 | * A gang block is a collection of small blocks that looks to the DMU | |
2598 | * like one large block. When zio_dva_allocate() cannot find a block | |
2599 | * of the requested size, due to either severe fragmentation or the pool | |
2600 | * being nearly full, it calls zio_write_gang_block() to construct the | |
2601 | * block from smaller fragments. | |
2602 | * | |
2603 | * A gang block consists of a gang header (zio_gbh_phys_t) and up to | |
2604 | * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like | |
2605 | * an indirect block: it's an array of block pointers. It consumes | |
2606 | * only one sector and hence is allocatable regardless of fragmentation. | |
2607 | * The gang header's bps point to its gang members, which hold the data. | |
2608 | * | |
2609 | * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> | |
2610 | * as the verifier to ensure uniqueness of the SHA256 checksum. | |
2611 | * Critically, the gang block bp's blk_cksum is the checksum of the data, | |
2612 | * not the gang header. This ensures that data block signatures (needed for | |
2613 | * deduplication) are independent of how the block is physically stored. | |
2614 | * | |
2615 | * Gang blocks can be nested: a gang member may itself be a gang block. | |
2616 | * Thus every gang block is a tree in which root and all interior nodes are | |
2617 | * gang headers, and the leaves are normal blocks that contain user data. | |
2618 | * The root of the gang tree is called the gang leader. | |
2619 | * | |
2620 | * To perform any operation (read, rewrite, free, claim) on a gang block, | |
2621 | * zio_gang_assemble() first assembles the gang tree (minus data leaves) | |
2622 | * in the io_gang_tree field of the original logical i/o by recursively | |
2623 | * reading the gang leader and all gang headers below it. This yields | |
2624 | * an in-core tree containing the contents of every gang header and the | |
2625 | * bps for every constituent of the gang block. | |
2626 | * | |
2627 | * With the gang tree now assembled, zio_gang_issue() just walks the gang tree | |
2628 | * and invokes a callback on each bp. To free a gang block, zio_gang_issue() | |
2629 | * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. | |
2630 | * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). | |
2631 | * zio_read_gang() is a wrapper around zio_read() that omits reading gang | |
2632 | * headers, since we already have those in io_gang_tree. zio_rewrite_gang() | |
2633 | * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() | |
2634 | * of the gang header plus zio_checksum_compute() of the data to update the | |
2635 | * gang header's blk_cksum as described above. | |
2636 | * | |
2637 | * The two-phase assemble/issue model solves the problem of partial failure -- | |
2638 | * what if you'd freed part of a gang block but then couldn't read the | |
2639 | * gang header for another part? Assembling the entire gang tree first | |
2640 | * ensures that all the necessary gang header I/O has succeeded before | |
2641 | * starting the actual work of free, claim, or write. Once the gang tree | |
2642 | * is assembled, free and claim are in-memory operations that cannot fail. | |
2643 | * | |
2644 | * In the event that a gang write fails, zio_dva_unallocate() walks the | |
2645 | * gang tree to immediately free (i.e. insert back into the space map) | |
2646 | * everything we've allocated. This ensures that we don't get ENOSPC | |
2647 | * errors during repeated suspend/resume cycles due to a flaky device. | |
2648 | * | |
2649 | * Gang rewrites only happen during sync-to-convergence. If we can't assemble | |
2650 | * the gang tree, we won't modify the block, so we can safely defer the free | |
2651 | * (knowing that the block is still intact). If we *can* assemble the gang | |
2652 | * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free | |
2653 | * each constituent bp and we can allocate a new block on the next sync pass. | |
2654 | * | |
2655 | * In all cases, the gang tree allows complete recovery from partial failure. | |
34dc7c2f BB |
2656 | * ========================================================================== |
2657 | */ | |
b128c09f | 2658 | |
a6255b7f DQ |
2659 | static void |
2660 | zio_gang_issue_func_done(zio_t *zio) | |
2661 | { | |
e2af2acc | 2662 | abd_free(zio->io_abd); |
a6255b7f DQ |
2663 | } |
2664 | ||
b128c09f | 2665 | static zio_t * |
a6255b7f DQ |
2666 | zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, |
2667 | uint64_t offset) | |
34dc7c2f | 2668 | { |
b128c09f BB |
2669 | if (gn != NULL) |
2670 | return (pio); | |
34dc7c2f | 2671 | |
a6255b7f DQ |
2672 | return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset), |
2673 | BP_GET_PSIZE(bp), zio_gang_issue_func_done, | |
2674 | NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
b128c09f BB |
2675 | &pio->io_bookmark)); |
2676 | } | |
2677 | ||
a6255b7f DQ |
2678 | static zio_t * |
2679 | zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2680 | uint64_t offset) | |
b128c09f BB |
2681 | { |
2682 | zio_t *zio; | |
2683 | ||
2684 | if (gn != NULL) { | |
a6255b7f DQ |
2685 | abd_t *gbh_abd = |
2686 | abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
b128c09f | 2687 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
a6255b7f DQ |
2688 | gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL, |
2689 | pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
2690 | &pio->io_bookmark); | |
34dc7c2f | 2691 | /* |
b128c09f BB |
2692 | * As we rewrite each gang header, the pipeline will compute |
2693 | * a new gang block header checksum for it; but no one will | |
2694 | * compute a new data checksum, so we do that here. The one | |
2695 | * exception is the gang leader: the pipeline already computed | |
2696 | * its data checksum because that stage precedes gang assembly. | |
2697 | * (Presently, nothing actually uses interior data checksums; | |
2698 | * this is just good hygiene.) | |
34dc7c2f | 2699 | */ |
9babb374 | 2700 | if (gn != pio->io_gang_leader->io_gang_tree) { |
a6255b7f DQ |
2701 | abd_t *buf = abd_get_offset(data, offset); |
2702 | ||
b128c09f | 2703 | zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), |
a6255b7f DQ |
2704 | buf, BP_GET_PSIZE(bp)); |
2705 | ||
e2af2acc | 2706 | abd_free(buf); |
b128c09f | 2707 | } |
428870ff BB |
2708 | /* |
2709 | * If we are here to damage data for testing purposes, | |
2710 | * leave the GBH alone so that we can detect the damage. | |
2711 | */ | |
2712 | if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) | |
2713 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; | |
34dc7c2f | 2714 | } else { |
b128c09f | 2715 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
a6255b7f DQ |
2716 | abd_get_offset(data, offset), BP_GET_PSIZE(bp), |
2717 | zio_gang_issue_func_done, NULL, pio->io_priority, | |
b128c09f | 2718 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); |
34dc7c2f BB |
2719 | } |
2720 | ||
b128c09f BB |
2721 | return (zio); |
2722 | } | |
34dc7c2f | 2723 | |
a6255b7f DQ |
2724 | static zio_t * |
2725 | zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2726 | uint64_t offset) | |
b128c09f | 2727 | { |
14e4e3cb AZ |
2728 | (void) gn, (void) data, (void) offset; |
2729 | ||
9cdf7b1f MA |
2730 | zio_t *zio = zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, |
2731 | ZIO_GANG_CHILD_FLAGS(pio)); | |
2732 | if (zio == NULL) { | |
2733 | zio = zio_null(pio, pio->io_spa, | |
2734 | NULL, NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)); | |
2735 | } | |
2736 | return (zio); | |
34dc7c2f BB |
2737 | } |
2738 | ||
a6255b7f DQ |
2739 | static zio_t * |
2740 | zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2741 | uint64_t offset) | |
34dc7c2f | 2742 | { |
14e4e3cb | 2743 | (void) gn, (void) data, (void) offset; |
b128c09f BB |
2744 | return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, |
2745 | NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); | |
2746 | } | |
2747 | ||
2748 | static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { | |
2749 | NULL, | |
2750 | zio_read_gang, | |
2751 | zio_rewrite_gang, | |
2752 | zio_free_gang, | |
2753 | zio_claim_gang, | |
2754 | NULL | |
2755 | }; | |
34dc7c2f | 2756 | |
b128c09f | 2757 | static void zio_gang_tree_assemble_done(zio_t *zio); |
34dc7c2f | 2758 | |
b128c09f BB |
2759 | static zio_gang_node_t * |
2760 | zio_gang_node_alloc(zio_gang_node_t **gnpp) | |
2761 | { | |
2762 | zio_gang_node_t *gn; | |
34dc7c2f | 2763 | |
b128c09f | 2764 | ASSERT(*gnpp == NULL); |
34dc7c2f | 2765 | |
79c76d5b | 2766 | gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); |
b128c09f BB |
2767 | gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); |
2768 | *gnpp = gn; | |
34dc7c2f | 2769 | |
b128c09f | 2770 | return (gn); |
34dc7c2f BB |
2771 | } |
2772 | ||
34dc7c2f | 2773 | static void |
b128c09f | 2774 | zio_gang_node_free(zio_gang_node_t **gnpp) |
34dc7c2f | 2775 | { |
b128c09f | 2776 | zio_gang_node_t *gn = *gnpp; |
34dc7c2f | 2777 | |
1c27024e | 2778 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f BB |
2779 | ASSERT(gn->gn_child[g] == NULL); |
2780 | ||
2781 | zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
2782 | kmem_free(gn, sizeof (*gn)); | |
2783 | *gnpp = NULL; | |
34dc7c2f BB |
2784 | } |
2785 | ||
b128c09f BB |
2786 | static void |
2787 | zio_gang_tree_free(zio_gang_node_t **gnpp) | |
34dc7c2f | 2788 | { |
b128c09f | 2789 | zio_gang_node_t *gn = *gnpp; |
34dc7c2f | 2790 | |
b128c09f BB |
2791 | if (gn == NULL) |
2792 | return; | |
34dc7c2f | 2793 | |
1c27024e | 2794 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f | 2795 | zio_gang_tree_free(&gn->gn_child[g]); |
34dc7c2f | 2796 | |
b128c09f | 2797 | zio_gang_node_free(gnpp); |
34dc7c2f BB |
2798 | } |
2799 | ||
b128c09f | 2800 | static void |
9babb374 | 2801 | zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) |
34dc7c2f | 2802 | { |
b128c09f | 2803 | zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); |
a6255b7f | 2804 | abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); |
b128c09f | 2805 | |
9babb374 | 2806 | ASSERT(gio->io_gang_leader == gio); |
b128c09f | 2807 | ASSERT(BP_IS_GANG(bp)); |
34dc7c2f | 2808 | |
a6255b7f DQ |
2809 | zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE, |
2810 | zio_gang_tree_assemble_done, gn, gio->io_priority, | |
2811 | ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); | |
b128c09f | 2812 | } |
34dc7c2f | 2813 | |
b128c09f BB |
2814 | static void |
2815 | zio_gang_tree_assemble_done(zio_t *zio) | |
2816 | { | |
9babb374 | 2817 | zio_t *gio = zio->io_gang_leader; |
b128c09f BB |
2818 | zio_gang_node_t *gn = zio->io_private; |
2819 | blkptr_t *bp = zio->io_bp; | |
34dc7c2f | 2820 | |
9babb374 | 2821 | ASSERT(gio == zio_unique_parent(zio)); |
ccec7fbe | 2822 | ASSERT(list_is_empty(&zio->io_child_list)); |
34dc7c2f | 2823 | |
b128c09f BB |
2824 | if (zio->io_error) |
2825 | return; | |
34dc7c2f | 2826 | |
a6255b7f | 2827 | /* this ABD was created from a linear buf in zio_gang_tree_assemble */ |
b128c09f | 2828 | if (BP_SHOULD_BYTESWAP(bp)) |
a6255b7f | 2829 | byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size); |
34dc7c2f | 2830 | |
a6255b7f | 2831 | ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh); |
b128c09f | 2832 | ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); |
428870ff | 2833 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2834 | |
e2af2acc | 2835 | abd_free(zio->io_abd); |
a6255b7f | 2836 | |
1c27024e | 2837 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2838 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2839 | if (!BP_IS_GANG(gbp)) | |
2840 | continue; | |
9babb374 | 2841 | zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); |
b128c09f | 2842 | } |
34dc7c2f BB |
2843 | } |
2844 | ||
b128c09f | 2845 | static void |
a6255b7f DQ |
2846 | zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data, |
2847 | uint64_t offset) | |
34dc7c2f | 2848 | { |
9babb374 | 2849 | zio_t *gio = pio->io_gang_leader; |
b128c09f | 2850 | zio_t *zio; |
34dc7c2f | 2851 | |
b128c09f | 2852 | ASSERT(BP_IS_GANG(bp) == !!gn); |
9babb374 BB |
2853 | ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); |
2854 | ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); | |
34dc7c2f | 2855 | |
b128c09f BB |
2856 | /* |
2857 | * If you're a gang header, your data is in gn->gn_gbh. | |
2858 | * If you're a gang member, your data is in 'data' and gn == NULL. | |
2859 | */ | |
a6255b7f | 2860 | zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset); |
34dc7c2f | 2861 | |
b128c09f | 2862 | if (gn != NULL) { |
428870ff | 2863 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2864 | |
1c27024e | 2865 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2866 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2867 | if (BP_IS_HOLE(gbp)) | |
2868 | continue; | |
a6255b7f DQ |
2869 | zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data, |
2870 | offset); | |
2871 | offset += BP_GET_PSIZE(gbp); | |
b128c09f | 2872 | } |
34dc7c2f BB |
2873 | } |
2874 | ||
9babb374 | 2875 | if (gn == gio->io_gang_tree) |
a6255b7f | 2876 | ASSERT3U(gio->io_size, ==, offset); |
34dc7c2f | 2877 | |
b128c09f BB |
2878 | if (zio != pio) |
2879 | zio_nowait(zio); | |
34dc7c2f BB |
2880 | } |
2881 | ||
62840030 | 2882 | static zio_t * |
b128c09f | 2883 | zio_gang_assemble(zio_t *zio) |
34dc7c2f | 2884 | { |
b128c09f | 2885 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2886 | |
9babb374 BB |
2887 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); |
2888 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
2889 | ||
2890 | zio->io_gang_leader = zio; | |
34dc7c2f | 2891 | |
b128c09f | 2892 | zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); |
34dc7c2f | 2893 | |
62840030 | 2894 | return (zio); |
34dc7c2f BB |
2895 | } |
2896 | ||
62840030 | 2897 | static zio_t * |
b128c09f | 2898 | zio_gang_issue(zio_t *zio) |
34dc7c2f | 2899 | { |
b128c09f | 2900 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2901 | |
ddc751d5 | 2902 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) { |
62840030 | 2903 | return (NULL); |
ddc751d5 | 2904 | } |
34dc7c2f | 2905 | |
9babb374 BB |
2906 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); |
2907 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
34dc7c2f | 2908 | |
b128c09f | 2909 | if (zio->io_child_error[ZIO_CHILD_GANG] == 0) |
a6255b7f DQ |
2910 | zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd, |
2911 | 0); | |
b128c09f | 2912 | else |
9babb374 | 2913 | zio_gang_tree_free(&zio->io_gang_tree); |
34dc7c2f | 2914 | |
b128c09f | 2915 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f | 2916 | |
62840030 | 2917 | return (zio); |
34dc7c2f BB |
2918 | } |
2919 | ||
3bd4df38 EN |
2920 | static void |
2921 | zio_gang_inherit_allocator(zio_t *pio, zio_t *cio) | |
2922 | { | |
2923 | cio->io_allocator = pio->io_allocator; | |
2924 | cio->io_wr_iss_tq = pio->io_wr_iss_tq; | |
2925 | } | |
2926 | ||
34dc7c2f | 2927 | static void |
b128c09f | 2928 | zio_write_gang_member_ready(zio_t *zio) |
34dc7c2f | 2929 | { |
d164b209 | 2930 | zio_t *pio = zio_unique_parent(zio); |
34dc7c2f BB |
2931 | dva_t *cdva = zio->io_bp->blk_dva; |
2932 | dva_t *pdva = pio->io_bp->blk_dva; | |
2933 | uint64_t asize; | |
2a8ba608 | 2934 | zio_t *gio __maybe_unused = zio->io_gang_leader; |
34dc7c2f | 2935 | |
b128c09f BB |
2936 | if (BP_IS_HOLE(zio->io_bp)) |
2937 | return; | |
2938 | ||
2939 | ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); | |
2940 | ||
2941 | ASSERT(zio->io_child_type == ZIO_CHILD_GANG); | |
428870ff BB |
2942 | ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); |
2943 | ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); | |
2944 | ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); | |
14872aaa | 2945 | VERIFY3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); |
34dc7c2f BB |
2946 | |
2947 | mutex_enter(&pio->io_lock); | |
1c27024e | 2948 | for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { |
34dc7c2f BB |
2949 | ASSERT(DVA_GET_GANG(&pdva[d])); |
2950 | asize = DVA_GET_ASIZE(&pdva[d]); | |
2951 | asize += DVA_GET_ASIZE(&cdva[d]); | |
2952 | DVA_SET_ASIZE(&pdva[d], asize); | |
2953 | } | |
2954 | mutex_exit(&pio->io_lock); | |
2955 | } | |
2956 | ||
a6255b7f DQ |
2957 | static void |
2958 | zio_write_gang_done(zio_t *zio) | |
2959 | { | |
c955398b BL |
2960 | /* |
2961 | * The io_abd field will be NULL for a zio with no data. The io_flags | |
2962 | * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't | |
2963 | * check for it here as it is cleared in zio_ready. | |
2964 | */ | |
2965 | if (zio->io_abd != NULL) | |
e2af2acc | 2966 | abd_free(zio->io_abd); |
a6255b7f DQ |
2967 | } |
2968 | ||
62840030 | 2969 | static zio_t * |
aa755b35 | 2970 | zio_write_gang_block(zio_t *pio, metaslab_class_t *mc) |
34dc7c2f | 2971 | { |
b128c09f BB |
2972 | spa_t *spa = pio->io_spa; |
2973 | blkptr_t *bp = pio->io_bp; | |
9babb374 | 2974 | zio_t *gio = pio->io_gang_leader; |
b128c09f BB |
2975 | zio_t *zio; |
2976 | zio_gang_node_t *gn, **gnpp; | |
34dc7c2f | 2977 | zio_gbh_phys_t *gbh; |
a6255b7f | 2978 | abd_t *gbh_abd; |
b128c09f BB |
2979 | uint64_t txg = pio->io_txg; |
2980 | uint64_t resid = pio->io_size; | |
2981 | uint64_t lsize; | |
428870ff | 2982 | int copies = gio->io_prop.zp_copies; |
b128c09f | 2983 | zio_prop_t zp; |
1c27024e | 2984 | int error; |
c955398b | 2985 | boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA); |
b5256303 TC |
2986 | |
2987 | /* | |
14872aaa MA |
2988 | * If one copy was requested, store 2 copies of the GBH, so that we |
2989 | * can still traverse all the data (e.g. to free or scrub) even if a | |
2990 | * block is damaged. Note that we can't store 3 copies of the GBH in | |
2991 | * all cases, e.g. with encryption, which uses DVA[2] for the IV+salt. | |
b5256303 | 2992 | */ |
14872aaa MA |
2993 | int gbh_copies = copies; |
2994 | if (gbh_copies == 1) { | |
2995 | gbh_copies = MIN(2, spa_max_replication(spa)); | |
2996 | } | |
b5256303 | 2997 | |
3bd4df38 | 2998 | ASSERT(ZIO_HAS_ALLOCATOR(pio)); |
1c27024e | 2999 | int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER; |
3dfb57a3 DB |
3000 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
3001 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
c955398b | 3002 | ASSERT(has_data); |
3dfb57a3 DB |
3003 | |
3004 | flags |= METASLAB_ASYNC_ALLOC; | |
f8020c93 AM |
3005 | VERIFY(zfs_refcount_held(&mc->mc_allocator[pio->io_allocator]. |
3006 | mca_alloc_slots, pio)); | |
3dfb57a3 DB |
3007 | |
3008 | /* | |
3009 | * The logical zio has already placed a reservation for | |
3010 | * 'copies' allocation slots but gang blocks may require | |
3011 | * additional copies. These additional copies | |
3012 | * (i.e. gbh_copies - copies) are guaranteed to succeed | |
3013 | * since metaslab_class_throttle_reserve() always allows | |
3014 | * additional reservations for gang blocks. | |
3015 | */ | |
3016 | VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies, | |
492f64e9 | 3017 | pio->io_allocator, pio, flags)); |
3dfb57a3 DB |
3018 | } |
3019 | ||
3020 | error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE, | |
4e21fd06 | 3021 | bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, |
492f64e9 | 3022 | &pio->io_alloc_list, pio, pio->io_allocator); |
34dc7c2f | 3023 | if (error) { |
3dfb57a3 DB |
3024 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
3025 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
c955398b | 3026 | ASSERT(has_data); |
3dfb57a3 DB |
3027 | |
3028 | /* | |
3029 | * If we failed to allocate the gang block header then | |
3030 | * we remove any additional allocation reservations that | |
3031 | * we placed here. The original reservation will | |
3032 | * be removed when the logical I/O goes to the ready | |
3033 | * stage. | |
3034 | */ | |
3035 | metaslab_class_throttle_unreserve(mc, | |
492f64e9 | 3036 | gbh_copies - copies, pio->io_allocator, pio); |
3dfb57a3 DB |
3037 | } |
3038 | ||
b128c09f | 3039 | pio->io_error = error; |
62840030 | 3040 | return (pio); |
34dc7c2f BB |
3041 | } |
3042 | ||
9babb374 BB |
3043 | if (pio == gio) { |
3044 | gnpp = &gio->io_gang_tree; | |
b128c09f BB |
3045 | } else { |
3046 | gnpp = pio->io_private; | |
3047 | ASSERT(pio->io_ready == zio_write_gang_member_ready); | |
34dc7c2f BB |
3048 | } |
3049 | ||
b128c09f BB |
3050 | gn = zio_gang_node_alloc(gnpp); |
3051 | gbh = gn->gn_gbh; | |
861166b0 | 3052 | memset(gbh, 0, SPA_GANGBLOCKSIZE); |
a6255b7f | 3053 | gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE); |
34dc7c2f | 3054 | |
b128c09f BB |
3055 | /* |
3056 | * Create the gang header. | |
3057 | */ | |
a6255b7f DQ |
3058 | zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE, |
3059 | zio_write_gang_done, NULL, pio->io_priority, | |
3060 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); | |
34dc7c2f | 3061 | |
3bd4df38 EN |
3062 | zio_gang_inherit_allocator(pio, zio); |
3063 | ||
b128c09f BB |
3064 | /* |
3065 | * Create and nowait the gang children. | |
3066 | */ | |
1c27024e | 3067 | for (int g = 0; resid != 0; resid -= lsize, g++) { |
b128c09f BB |
3068 | lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), |
3069 | SPA_MINBLOCKSIZE); | |
3070 | ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); | |
3071 | ||
9babb374 | 3072 | zp.zp_checksum = gio->io_prop.zp_checksum; |
b128c09f | 3073 | zp.zp_compress = ZIO_COMPRESS_OFF; |
10b3c7f5 | 3074 | zp.zp_complevel = gio->io_prop.zp_complevel; |
b128c09f BB |
3075 | zp.zp_type = DMU_OT_NONE; |
3076 | zp.zp_level = 0; | |
428870ff | 3077 | zp.zp_copies = gio->io_prop.zp_copies; |
03c6040b GW |
3078 | zp.zp_dedup = B_FALSE; |
3079 | zp.zp_dedup_verify = B_FALSE; | |
3080 | zp.zp_nopwrite = B_FALSE; | |
4807c0ba TC |
3081 | zp.zp_encrypt = gio->io_prop.zp_encrypt; |
3082 | zp.zp_byteorder = gio->io_prop.zp_byteorder; | |
861166b0 AZ |
3083 | memset(zp.zp_salt, 0, ZIO_DATA_SALT_LEN); |
3084 | memset(zp.zp_iv, 0, ZIO_DATA_IV_LEN); | |
3085 | memset(zp.zp_mac, 0, ZIO_DATA_MAC_LEN); | |
b128c09f | 3086 | |
1c27024e | 3087 | zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g], |
c955398b BL |
3088 | has_data ? abd_get_offset(pio->io_abd, pio->io_size - |
3089 | resid) : NULL, lsize, lsize, &zp, | |
ccec7fbe | 3090 | zio_write_gang_member_ready, NULL, |
a6255b7f | 3091 | zio_write_gang_done, &gn->gn_child[g], pio->io_priority, |
3dfb57a3 DB |
3092 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); |
3093 | ||
3bd4df38 EN |
3094 | zio_gang_inherit_allocator(zio, cio); |
3095 | ||
3dfb57a3 DB |
3096 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
3097 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
c955398b | 3098 | ASSERT(has_data); |
3dfb57a3 DB |
3099 | |
3100 | /* | |
3101 | * Gang children won't throttle but we should | |
3102 | * account for their work, so reserve an allocation | |
3103 | * slot for them here. | |
3104 | */ | |
3105 | VERIFY(metaslab_class_throttle_reserve(mc, | |
492f64e9 | 3106 | zp.zp_copies, cio->io_allocator, cio, flags)); |
3dfb57a3 DB |
3107 | } |
3108 | zio_nowait(cio); | |
b128c09f | 3109 | } |
34dc7c2f BB |
3110 | |
3111 | /* | |
b128c09f | 3112 | * Set pio's pipeline to just wait for zio to finish. |
34dc7c2f | 3113 | */ |
b128c09f BB |
3114 | pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
3115 | ||
3116 | zio_nowait(zio); | |
3117 | ||
62840030 | 3118 | return (pio); |
34dc7c2f BB |
3119 | } |
3120 | ||
03c6040b | 3121 | /* |
3c67d83a TH |
3122 | * The zio_nop_write stage in the pipeline determines if allocating a |
3123 | * new bp is necessary. The nopwrite feature can handle writes in | |
3124 | * either syncing or open context (i.e. zil writes) and as a result is | |
3125 | * mutually exclusive with dedup. | |
3126 | * | |
3127 | * By leveraging a cryptographically secure checksum, such as SHA256, we | |
3128 | * can compare the checksums of the new data and the old to determine if | |
3129 | * allocating a new block is required. Note that our requirements for | |
3130 | * cryptographic strength are fairly weak: there can't be any accidental | |
3131 | * hash collisions, but we don't need to be secure against intentional | |
3132 | * (malicious) collisions. To trigger a nopwrite, you have to be able | |
3133 | * to write the file to begin with, and triggering an incorrect (hash | |
3134 | * collision) nopwrite is no worse than simply writing to the file. | |
3135 | * That said, there are no known attacks against the checksum algorithms | |
3136 | * used for nopwrite, assuming that the salt and the checksums | |
3137 | * themselves remain secret. | |
03c6040b | 3138 | */ |
62840030 | 3139 | static zio_t * |
03c6040b GW |
3140 | zio_nop_write(zio_t *zio) |
3141 | { | |
3142 | blkptr_t *bp = zio->io_bp; | |
3143 | blkptr_t *bp_orig = &zio->io_bp_orig; | |
3144 | zio_prop_t *zp = &zio->io_prop; | |
3145 | ||
d7cf06a2 | 3146 | ASSERT(BP_IS_HOLE(bp)); |
03c6040b GW |
3147 | ASSERT(BP_GET_LEVEL(bp) == 0); |
3148 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
3149 | ASSERT(zp->zp_nopwrite); | |
3150 | ASSERT(!zp->zp_dedup); | |
3151 | ASSERT(zio->io_bp_override == NULL); | |
3152 | ASSERT(IO_IS_ALLOCATING(zio)); | |
3153 | ||
3154 | /* | |
3155 | * Check to see if the original bp and the new bp have matching | |
3156 | * characteristics (i.e. same checksum, compression algorithms, etc). | |
3157 | * If they don't then just continue with the pipeline which will | |
3158 | * allocate a new bp. | |
3159 | */ | |
3160 | if (BP_IS_HOLE(bp_orig) || | |
3c67d83a TH |
3161 | !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & |
3162 | ZCHECKSUM_FLAG_NOPWRITE) || | |
b5256303 | 3163 | BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) || |
03c6040b GW |
3164 | BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || |
3165 | BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || | |
3166 | BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || | |
3167 | zp->zp_copies != BP_GET_NDVAS(bp_orig)) | |
62840030 | 3168 | return (zio); |
03c6040b GW |
3169 | |
3170 | /* | |
3171 | * If the checksums match then reset the pipeline so that we | |
3172 | * avoid allocating a new bp and issuing any I/O. | |
3173 | */ | |
3174 | if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { | |
3c67d83a TH |
3175 | ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & |
3176 | ZCHECKSUM_FLAG_NOPWRITE); | |
03c6040b GW |
3177 | ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); |
3178 | ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); | |
3179 | ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); | |
d7cf06a2 | 3180 | ASSERT3U(bp->blk_prop, ==, bp_orig->blk_prop); |
03c6040b | 3181 | |
681a85cb GW |
3182 | /* |
3183 | * If we're overwriting a block that is currently on an | |
3184 | * indirect vdev, then ignore the nopwrite request and | |
3185 | * allow a new block to be allocated on a concrete vdev. | |
3186 | */ | |
3187 | spa_config_enter(zio->io_spa, SCL_VDEV, FTAG, RW_READER); | |
d7cf06a2 GW |
3188 | for (int d = 0; d < BP_GET_NDVAS(bp_orig); d++) { |
3189 | vdev_t *tvd = vdev_lookup_top(zio->io_spa, | |
3190 | DVA_GET_VDEV(&bp_orig->blk_dva[d])); | |
3191 | if (tvd->vdev_ops == &vdev_indirect_ops) { | |
3192 | spa_config_exit(zio->io_spa, SCL_VDEV, FTAG); | |
3193 | return (zio); | |
3194 | } | |
681a85cb GW |
3195 | } |
3196 | spa_config_exit(zio->io_spa, SCL_VDEV, FTAG); | |
3197 | ||
03c6040b GW |
3198 | *bp = *bp_orig; |
3199 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
3200 | zio->io_flags |= ZIO_FLAG_NOPWRITE; | |
3201 | } | |
3202 | ||
62840030 | 3203 | return (zio); |
03c6040b GW |
3204 | } |
3205 | ||
67a1b037 PJD |
3206 | /* |
3207 | * ========================================================================== | |
3208 | * Block Reference Table | |
3209 | * ========================================================================== | |
3210 | */ | |
3211 | static zio_t * | |
3212 | zio_brt_free(zio_t *zio) | |
3213 | { | |
3214 | blkptr_t *bp; | |
3215 | ||
3216 | bp = zio->io_bp; | |
3217 | ||
3218 | if (BP_GET_LEVEL(bp) > 0 || | |
3219 | BP_IS_METADATA(bp) || | |
3220 | !brt_maybe_exists(zio->io_spa, bp)) { | |
3221 | return (zio); | |
3222 | } | |
3223 | ||
3224 | if (!brt_entry_decref(zio->io_spa, bp)) { | |
3225 | /* | |
3226 | * This isn't the last reference, so we cannot free | |
3227 | * the data yet. | |
3228 | */ | |
3229 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
3230 | } | |
3231 | ||
3232 | return (zio); | |
3233 | } | |
3234 | ||
34dc7c2f BB |
3235 | /* |
3236 | * ========================================================================== | |
428870ff | 3237 | * Dedup |
34dc7c2f BB |
3238 | * ========================================================================== |
3239 | */ | |
428870ff BB |
3240 | static void |
3241 | zio_ddt_child_read_done(zio_t *zio) | |
3242 | { | |
3243 | blkptr_t *bp = zio->io_bp; | |
3244 | ddt_entry_t *dde = zio->io_private; | |
3245 | ddt_phys_t *ddp; | |
3246 | zio_t *pio = zio_unique_parent(zio); | |
3247 | ||
3248 | mutex_enter(&pio->io_lock); | |
3249 | ddp = ddt_phys_select(dde, bp); | |
3250 | if (zio->io_error == 0) | |
3251 | ddt_phys_clear(ddp); /* this ddp doesn't need repair */ | |
a6255b7f DQ |
3252 | |
3253 | if (zio->io_error == 0 && dde->dde_repair_abd == NULL) | |
3254 | dde->dde_repair_abd = zio->io_abd; | |
428870ff | 3255 | else |
a6255b7f | 3256 | abd_free(zio->io_abd); |
428870ff BB |
3257 | mutex_exit(&pio->io_lock); |
3258 | } | |
3259 | ||
62840030 | 3260 | static zio_t * |
428870ff BB |
3261 | zio_ddt_read_start(zio_t *zio) |
3262 | { | |
3263 | blkptr_t *bp = zio->io_bp; | |
3264 | ||
3265 | ASSERT(BP_GET_DEDUP(bp)); | |
3266 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
3267 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
3268 | ||
3269 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
3270 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
3271 | ddt_entry_t *dde = ddt_repair_start(ddt, bp); | |
3272 | ddt_phys_t *ddp = dde->dde_phys; | |
3273 | ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); | |
3274 | blkptr_t blk; | |
3275 | ||
3276 | ASSERT(zio->io_vsd == NULL); | |
3277 | zio->io_vsd = dde; | |
3278 | ||
3279 | if (ddp_self == NULL) | |
62840030 | 3280 | return (zio); |
428870ff | 3281 | |
1c27024e | 3282 | for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { |
428870ff BB |
3283 | if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) |
3284 | continue; | |
3285 | ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, | |
3286 | &blk); | |
3287 | zio_nowait(zio_read(zio, zio->io_spa, &blk, | |
a6255b7f DQ |
3288 | abd_alloc_for_io(zio->io_size, B_TRUE), |
3289 | zio->io_size, zio_ddt_child_read_done, dde, | |
3290 | zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) | | |
3291 | ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark)); | |
428870ff | 3292 | } |
62840030 | 3293 | return (zio); |
428870ff BB |
3294 | } |
3295 | ||
3296 | zio_nowait(zio_read(zio, zio->io_spa, bp, | |
a6255b7f | 3297 | zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority, |
428870ff BB |
3298 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); |
3299 | ||
62840030 | 3300 | return (zio); |
428870ff BB |
3301 | } |
3302 | ||
62840030 | 3303 | static zio_t * |
428870ff BB |
3304 | zio_ddt_read_done(zio_t *zio) |
3305 | { | |
3306 | blkptr_t *bp = zio->io_bp; | |
3307 | ||
ddc751d5 | 3308 | if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) { |
62840030 | 3309 | return (NULL); |
ddc751d5 | 3310 | } |
428870ff BB |
3311 | |
3312 | ASSERT(BP_GET_DEDUP(bp)); | |
3313 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
3314 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
3315 | ||
3316 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
3317 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
3318 | ddt_entry_t *dde = zio->io_vsd; | |
3319 | if (ddt == NULL) { | |
3320 | ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); | |
62840030 | 3321 | return (zio); |
428870ff BB |
3322 | } |
3323 | if (dde == NULL) { | |
3324 | zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; | |
3325 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); | |
62840030 | 3326 | return (NULL); |
428870ff | 3327 | } |
a6255b7f DQ |
3328 | if (dde->dde_repair_abd != NULL) { |
3329 | abd_copy(zio->io_abd, dde->dde_repair_abd, | |
3330 | zio->io_size); | |
428870ff BB |
3331 | zio->io_child_error[ZIO_CHILD_DDT] = 0; |
3332 | } | |
3333 | ddt_repair_done(ddt, dde); | |
3334 | zio->io_vsd = NULL; | |
3335 | } | |
3336 | ||
3337 | ASSERT(zio->io_vsd == NULL); | |
3338 | ||
62840030 | 3339 | return (zio); |
428870ff BB |
3340 | } |
3341 | ||
3342 | static boolean_t | |
3343 | zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) | |
3344 | { | |
3345 | spa_t *spa = zio->io_spa; | |
c17bcf83 | 3346 | boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW); |
428870ff | 3347 | |
c17bcf83 | 3348 | ASSERT(!(zio->io_bp_override && do_raw)); |
2aa34383 | 3349 | |
428870ff BB |
3350 | /* |
3351 | * Note: we compare the original data, not the transformed data, | |
3352 | * because when zio->io_bp is an override bp, we will not have | |
3353 | * pushed the I/O transforms. That's an important optimization | |
3354 | * because otherwise we'd compress/encrypt all dmu_sync() data twice. | |
c17bcf83 | 3355 | * However, we should never get a raw, override zio so in these |
b5256303 | 3356 | * cases we can compare the io_abd directly. This is useful because |
c17bcf83 TC |
3357 | * it allows us to do dedup verification even if we don't have access |
3358 | * to the original data (for instance, if the encryption keys aren't | |
3359 | * loaded). | |
428870ff | 3360 | */ |
c17bcf83 | 3361 | |
1c27024e | 3362 | for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
3363 | zio_t *lio = dde->dde_lead_zio[p]; |
3364 | ||
c17bcf83 TC |
3365 | if (lio != NULL && do_raw) { |
3366 | return (lio->io_size != zio->io_size || | |
a6255b7f | 3367 | abd_cmp(zio->io_abd, lio->io_abd) != 0); |
c17bcf83 | 3368 | } else if (lio != NULL) { |
428870ff | 3369 | return (lio->io_orig_size != zio->io_orig_size || |
a6255b7f | 3370 | abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0); |
428870ff BB |
3371 | } |
3372 | } | |
3373 | ||
1c27024e | 3374 | for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
3375 | ddt_phys_t *ddp = &dde->dde_phys[p]; |
3376 | ||
c17bcf83 TC |
3377 | if (ddp->ddp_phys_birth != 0 && do_raw) { |
3378 | blkptr_t blk = *zio->io_bp; | |
3379 | uint64_t psize; | |
a6255b7f | 3380 | abd_t *tmpabd; |
c17bcf83 TC |
3381 | int error; |
3382 | ||
3383 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
3384 | psize = BP_GET_PSIZE(&blk); | |
3385 | ||
3386 | if (psize != zio->io_size) | |
3387 | return (B_TRUE); | |
3388 | ||
3389 | ddt_exit(ddt); | |
3390 | ||
a6255b7f | 3391 | tmpabd = abd_alloc_for_io(psize, B_TRUE); |
c17bcf83 | 3392 | |
a6255b7f | 3393 | error = zio_wait(zio_read(NULL, spa, &blk, tmpabd, |
c17bcf83 TC |
3394 | psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ, |
3395 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
3396 | ZIO_FLAG_RAW, &zio->io_bookmark)); | |
3397 | ||
3398 | if (error == 0) { | |
a6255b7f | 3399 | if (abd_cmp(tmpabd, zio->io_abd) != 0) |
c17bcf83 TC |
3400 | error = SET_ERROR(ENOENT); |
3401 | } | |
3402 | ||
a6255b7f | 3403 | abd_free(tmpabd); |
c17bcf83 TC |
3404 | ddt_enter(ddt); |
3405 | return (error != 0); | |
3406 | } else if (ddp->ddp_phys_birth != 0) { | |
428870ff | 3407 | arc_buf_t *abuf = NULL; |
2a432414 | 3408 | arc_flags_t aflags = ARC_FLAG_WAIT; |
428870ff BB |
3409 | blkptr_t blk = *zio->io_bp; |
3410 | int error; | |
3411 | ||
3412 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
3413 | ||
c17bcf83 TC |
3414 | if (BP_GET_LSIZE(&blk) != zio->io_orig_size) |
3415 | return (B_TRUE); | |
3416 | ||
428870ff BB |
3417 | ddt_exit(ddt); |
3418 | ||
294f6806 | 3419 | error = arc_read(NULL, spa, &blk, |
428870ff BB |
3420 | arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, |
3421 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
3422 | &aflags, &zio->io_bookmark); | |
3423 | ||
3424 | if (error == 0) { | |
a6255b7f | 3425 | if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data, |
428870ff | 3426 | zio->io_orig_size) != 0) |
c17bcf83 | 3427 | error = SET_ERROR(ENOENT); |
d3c2ae1c | 3428 | arc_buf_destroy(abuf, &abuf); |
428870ff BB |
3429 | } |
3430 | ||
3431 | ddt_enter(ddt); | |
3432 | return (error != 0); | |
3433 | } | |
3434 | } | |
3435 | ||
3436 | return (B_FALSE); | |
3437 | } | |
3438 | ||
3439 | static void | |
3440 | zio_ddt_child_write_ready(zio_t *zio) | |
3441 | { | |
3442 | int p = zio->io_prop.zp_copies; | |
3443 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
3444 | ddt_entry_t *dde = zio->io_private; | |
3445 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
3446 | zio_t *pio; | |
3447 | ||
3448 | if (zio->io_error) | |
3449 | return; | |
3450 | ||
3451 | ddt_enter(ddt); | |
3452 | ||
3453 | ASSERT(dde->dde_lead_zio[p] == zio); | |
3454 | ||
3455 | ddt_phys_fill(ddp, zio->io_bp); | |
3456 | ||
1c27024e | 3457 | zio_link_t *zl = NULL; |
3dfb57a3 | 3458 | while ((pio = zio_walk_parents(zio, &zl)) != NULL) |
428870ff BB |
3459 | ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); |
3460 | ||
3461 | ddt_exit(ddt); | |
3462 | } | |
3463 | ||
3464 | static void | |
3465 | zio_ddt_child_write_done(zio_t *zio) | |
3466 | { | |
3467 | int p = zio->io_prop.zp_copies; | |
3468 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
3469 | ddt_entry_t *dde = zio->io_private; | |
3470 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
3471 | ||
3472 | ddt_enter(ddt); | |
3473 | ||
3474 | ASSERT(ddp->ddp_refcnt == 0); | |
3475 | ASSERT(dde->dde_lead_zio[p] == zio); | |
3476 | dde->dde_lead_zio[p] = NULL; | |
3477 | ||
3478 | if (zio->io_error == 0) { | |
3dfb57a3 DB |
3479 | zio_link_t *zl = NULL; |
3480 | while (zio_walk_parents(zio, &zl) != NULL) | |
428870ff BB |
3481 | ddt_phys_addref(ddp); |
3482 | } else { | |
3483 | ddt_phys_clear(ddp); | |
3484 | } | |
3485 | ||
3486 | ddt_exit(ddt); | |
3487 | } | |
3488 | ||
62840030 | 3489 | static zio_t * |
428870ff BB |
3490 | zio_ddt_write(zio_t *zio) |
3491 | { | |
3492 | spa_t *spa = zio->io_spa; | |
3493 | blkptr_t *bp = zio->io_bp; | |
3494 | uint64_t txg = zio->io_txg; | |
3495 | zio_prop_t *zp = &zio->io_prop; | |
3496 | int p = zp->zp_copies; | |
428870ff | 3497 | zio_t *cio = NULL; |
428870ff BB |
3498 | ddt_t *ddt = ddt_select(spa, bp); |
3499 | ddt_entry_t *dde; | |
3500 | ddt_phys_t *ddp; | |
3501 | ||
3502 | ASSERT(BP_GET_DEDUP(bp)); | |
3503 | ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); | |
3504 | ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); | |
c17bcf83 | 3505 | ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW))); |
428870ff BB |
3506 | |
3507 | ddt_enter(ddt); | |
3508 | dde = ddt_lookup(ddt, bp, B_TRUE); | |
3509 | ddp = &dde->dde_phys[p]; | |
3510 | ||
3511 | if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { | |
3512 | /* | |
3513 | * If we're using a weak checksum, upgrade to a strong checksum | |
3514 | * and try again. If we're already using a strong checksum, | |
3515 | * we can't resolve it, so just convert to an ordinary write. | |
3516 | * (And automatically e-mail a paper to Nature?) | |
3517 | */ | |
3c67d83a TH |
3518 | if (!(zio_checksum_table[zp->zp_checksum].ci_flags & |
3519 | ZCHECKSUM_FLAG_DEDUP)) { | |
428870ff BB |
3520 | zp->zp_checksum = spa_dedup_checksum(spa); |
3521 | zio_pop_transforms(zio); | |
3522 | zio->io_stage = ZIO_STAGE_OPEN; | |
3523 | BP_ZERO(bp); | |
3524 | } else { | |
03c6040b | 3525 | zp->zp_dedup = B_FALSE; |
accd6d9d | 3526 | BP_SET_DEDUP(bp, B_FALSE); |
428870ff | 3527 | } |
accd6d9d | 3528 | ASSERT(!BP_GET_DEDUP(bp)); |
428870ff BB |
3529 | zio->io_pipeline = ZIO_WRITE_PIPELINE; |
3530 | ddt_exit(ddt); | |
62840030 | 3531 | return (zio); |
428870ff BB |
3532 | } |
3533 | ||
428870ff BB |
3534 | if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { |
3535 | if (ddp->ddp_phys_birth != 0) | |
3536 | ddt_bp_fill(ddp, bp, txg); | |
3537 | if (dde->dde_lead_zio[p] != NULL) | |
3538 | zio_add_child(zio, dde->dde_lead_zio[p]); | |
3539 | else | |
3540 | ddt_phys_addref(ddp); | |
3541 | } else if (zio->io_bp_override) { | |
3542 | ASSERT(bp->blk_birth == txg); | |
3543 | ASSERT(BP_EQUAL(bp, zio->io_bp_override)); | |
3544 | ddt_phys_fill(ddp, bp); | |
3545 | ddt_phys_addref(ddp); | |
3546 | } else { | |
a6255b7f | 3547 | cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, |
2aa34383 | 3548 | zio->io_orig_size, zio->io_orig_size, zp, |
ccec7fbe | 3549 | zio_ddt_child_write_ready, NULL, |
428870ff BB |
3550 | zio_ddt_child_write_done, dde, zio->io_priority, |
3551 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); | |
3552 | ||
a6255b7f | 3553 | zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL); |
428870ff BB |
3554 | dde->dde_lead_zio[p] = cio; |
3555 | } | |
3556 | ||
3557 | ddt_exit(ddt); | |
3558 | ||
9cdf7b1f | 3559 | zio_nowait(cio); |
428870ff | 3560 | |
62840030 | 3561 | return (zio); |
428870ff BB |
3562 | } |
3563 | ||
27218a32 | 3564 | static ddt_entry_t *freedde; /* for debugging */ |
b128c09f | 3565 | |
62840030 | 3566 | static zio_t * |
428870ff BB |
3567 | zio_ddt_free(zio_t *zio) |
3568 | { | |
3569 | spa_t *spa = zio->io_spa; | |
3570 | blkptr_t *bp = zio->io_bp; | |
3571 | ddt_t *ddt = ddt_select(spa, bp); | |
3572 | ddt_entry_t *dde; | |
3573 | ddt_phys_t *ddp; | |
3574 | ||
3575 | ASSERT(BP_GET_DEDUP(bp)); | |
3576 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
3577 | ||
3578 | ddt_enter(ddt); | |
3579 | freedde = dde = ddt_lookup(ddt, bp, B_TRUE); | |
5dc6af0e BB |
3580 | if (dde) { |
3581 | ddp = ddt_phys_select(dde, bp); | |
3582 | if (ddp) | |
3583 | ddt_phys_decref(ddp); | |
3584 | } | |
428870ff BB |
3585 | ddt_exit(ddt); |
3586 | ||
62840030 | 3587 | return (zio); |
428870ff BB |
3588 | } |
3589 | ||
3590 | /* | |
3591 | * ========================================================================== | |
3592 | * Allocate and free blocks | |
3593 | * ========================================================================== | |
3594 | */ | |
3dfb57a3 DB |
3595 | |
3596 | static zio_t * | |
492f64e9 | 3597 | zio_io_to_allocate(spa_t *spa, int allocator) |
3dfb57a3 DB |
3598 | { |
3599 | zio_t *zio; | |
3600 | ||
1b50749c | 3601 | ASSERT(MUTEX_HELD(&spa->spa_allocs[allocator].spaa_lock)); |
3dfb57a3 | 3602 | |
1b50749c | 3603 | zio = avl_first(&spa->spa_allocs[allocator].spaa_tree); |
3dfb57a3 DB |
3604 | if (zio == NULL) |
3605 | return (NULL); | |
3606 | ||
3607 | ASSERT(IO_IS_ALLOCATING(zio)); | |
3bd4df38 | 3608 | ASSERT(ZIO_HAS_ALLOCATOR(zio)); |
3dfb57a3 DB |
3609 | |
3610 | /* | |
3611 | * Try to place a reservation for this zio. If we're unable to | |
3612 | * reserve then we throttle. | |
3613 | */ | |
492f64e9 | 3614 | ASSERT3U(zio->io_allocator, ==, allocator); |
cc99f275 | 3615 | if (!metaslab_class_throttle_reserve(zio->io_metaslab_class, |
1b50749c | 3616 | zio->io_prop.zp_copies, allocator, zio, 0)) { |
3dfb57a3 DB |
3617 | return (NULL); |
3618 | } | |
3619 | ||
1b50749c | 3620 | avl_remove(&spa->spa_allocs[allocator].spaa_tree, zio); |
3dfb57a3 DB |
3621 | ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE); |
3622 | ||
3623 | return (zio); | |
3624 | } | |
3625 | ||
62840030 | 3626 | static zio_t * |
3dfb57a3 DB |
3627 | zio_dva_throttle(zio_t *zio) |
3628 | { | |
3629 | spa_t *spa = zio->io_spa; | |
3630 | zio_t *nio; | |
cc99f275 DB |
3631 | metaslab_class_t *mc; |
3632 | ||
3633 | /* locate an appropriate allocation class */ | |
3634 | mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type, | |
3635 | zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk); | |
3dfb57a3 DB |
3636 | |
3637 | if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE || | |
cc99f275 | 3638 | !mc->mc_alloc_throttle_enabled || |
3dfb57a3 DB |
3639 | zio->io_child_type == ZIO_CHILD_GANG || |
3640 | zio->io_flags & ZIO_FLAG_NODATA) { | |
62840030 | 3641 | return (zio); |
3dfb57a3 DB |
3642 | } |
3643 | ||
1b50749c | 3644 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); |
3bd4df38 | 3645 | ASSERT(ZIO_HAS_ALLOCATOR(zio)); |
3dfb57a3 | 3646 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); |
3dfb57a3 DB |
3647 | ASSERT3U(zio->io_queued_timestamp, >, 0); |
3648 | ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE); | |
3649 | ||
3bd4df38 | 3650 | int allocator = zio->io_allocator; |
cc99f275 | 3651 | zio->io_metaslab_class = mc; |
1b50749c AM |
3652 | mutex_enter(&spa->spa_allocs[allocator].spaa_lock); |
3653 | avl_add(&spa->spa_allocs[allocator].spaa_tree, zio); | |
3654 | nio = zio_io_to_allocate(spa, allocator); | |
3655 | mutex_exit(&spa->spa_allocs[allocator].spaa_lock); | |
62840030 | 3656 | return (nio); |
3dfb57a3 DB |
3657 | } |
3658 | ||
cc99f275 | 3659 | static void |
492f64e9 | 3660 | zio_allocate_dispatch(spa_t *spa, int allocator) |
3dfb57a3 DB |
3661 | { |
3662 | zio_t *zio; | |
3663 | ||
1b50749c | 3664 | mutex_enter(&spa->spa_allocs[allocator].spaa_lock); |
492f64e9 | 3665 | zio = zio_io_to_allocate(spa, allocator); |
1b50749c | 3666 | mutex_exit(&spa->spa_allocs[allocator].spaa_lock); |
3dfb57a3 DB |
3667 | if (zio == NULL) |
3668 | return; | |
3669 | ||
3670 | ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE); | |
3671 | ASSERT0(zio->io_error); | |
3672 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE); | |
3673 | } | |
3674 | ||
62840030 | 3675 | static zio_t * |
34dc7c2f BB |
3676 | zio_dva_allocate(zio_t *zio) |
3677 | { | |
3678 | spa_t *spa = zio->io_spa; | |
cc99f275 | 3679 | metaslab_class_t *mc; |
34dc7c2f BB |
3680 | blkptr_t *bp = zio->io_bp; |
3681 | int error; | |
6d974228 | 3682 | int flags = 0; |
34dc7c2f | 3683 | |
9babb374 BB |
3684 | if (zio->io_gang_leader == NULL) { |
3685 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
3686 | zio->io_gang_leader = zio; | |
3687 | } | |
3688 | ||
34dc7c2f | 3689 | ASSERT(BP_IS_HOLE(bp)); |
c99c9001 | 3690 | ASSERT0(BP_GET_NDVAS(bp)); |
428870ff BB |
3691 | ASSERT3U(zio->io_prop.zp_copies, >, 0); |
3692 | ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); | |
34dc7c2f BB |
3693 | ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); |
3694 | ||
3dfb57a3 DB |
3695 | if (zio->io_flags & ZIO_FLAG_NODATA) |
3696 | flags |= METASLAB_DONT_THROTTLE; | |
3697 | if (zio->io_flags & ZIO_FLAG_GANG_CHILD) | |
3698 | flags |= METASLAB_GANG_CHILD; | |
3699 | if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) | |
3700 | flags |= METASLAB_ASYNC_ALLOC; | |
3701 | ||
cc99f275 DB |
3702 | /* |
3703 | * if not already chosen, locate an appropriate allocation class | |
3704 | */ | |
3705 | mc = zio->io_metaslab_class; | |
3706 | if (mc == NULL) { | |
3707 | mc = spa_preferred_class(spa, zio->io_size, | |
3708 | zio->io_prop.zp_type, zio->io_prop.zp_level, | |
3709 | zio->io_prop.zp_zpl_smallblk); | |
3710 | zio->io_metaslab_class = mc; | |
3711 | } | |
3712 | ||
aa755b35 MA |
3713 | /* |
3714 | * Try allocating the block in the usual metaslab class. | |
3715 | * If that's full, allocate it in the normal class. | |
3716 | * If that's full, allocate as a gang block, | |
3717 | * and if all are full, the allocation fails (which shouldn't happen). | |
3718 | * | |
3719 | * Note that we do not fall back on embedded slog (ZIL) space, to | |
3720 | * preserve unfragmented slog space, which is critical for decent | |
3721 | * sync write performance. If a log allocation fails, we will fall | |
3722 | * back to spa_sync() which is abysmal for performance. | |
3723 | */ | |
3bd4df38 | 3724 | ASSERT(ZIO_HAS_ALLOCATOR(zio)); |
b128c09f | 3725 | error = metaslab_alloc(spa, mc, zio->io_size, bp, |
4e21fd06 | 3726 | zio->io_prop.zp_copies, zio->io_txg, NULL, flags, |
492f64e9 | 3727 | &zio->io_alloc_list, zio, zio->io_allocator); |
34dc7c2f | 3728 | |
cc99f275 DB |
3729 | /* |
3730 | * Fallback to normal class when an alloc class is full | |
3731 | */ | |
3732 | if (error == ENOSPC && mc != spa_normal_class(spa)) { | |
3733 | /* | |
3734 | * If throttling, transfer reservation over to normal class. | |
3735 | * The io_allocator slot can remain the same even though we | |
3736 | * are switching classes. | |
3737 | */ | |
3738 | if (mc->mc_alloc_throttle_enabled && | |
3739 | (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) { | |
3740 | metaslab_class_throttle_unreserve(mc, | |
3741 | zio->io_prop.zp_copies, zio->io_allocator, zio); | |
3742 | zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING; | |
3743 | ||
aa755b35 MA |
3744 | VERIFY(metaslab_class_throttle_reserve( |
3745 | spa_normal_class(spa), | |
cc99f275 DB |
3746 | zio->io_prop.zp_copies, zio->io_allocator, zio, |
3747 | flags | METASLAB_MUST_RESERVE)); | |
cc99f275 | 3748 | } |
aa755b35 MA |
3749 | zio->io_metaslab_class = mc = spa_normal_class(spa); |
3750 | if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) { | |
3751 | zfs_dbgmsg("%s: metaslab allocation failure, " | |
3752 | "trying normal class: zio %px, size %llu, error %d", | |
8e739b2c RE |
3753 | spa_name(spa), zio, (u_longlong_t)zio->io_size, |
3754 | error); | |
aa755b35 | 3755 | } |
cc99f275 DB |
3756 | |
3757 | error = metaslab_alloc(spa, mc, zio->io_size, bp, | |
3758 | zio->io_prop.zp_copies, zio->io_txg, NULL, flags, | |
3759 | &zio->io_alloc_list, zio, zio->io_allocator); | |
3760 | } | |
3761 | ||
aa755b35 MA |
3762 | if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) { |
3763 | if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) { | |
3764 | zfs_dbgmsg("%s: metaslab allocation failure, " | |
3765 | "trying ganging: zio %px, size %llu, error %d", | |
8e739b2c RE |
3766 | spa_name(spa), zio, (u_longlong_t)zio->io_size, |
3767 | error); | |
aa755b35 MA |
3768 | } |
3769 | return (zio_write_gang_block(zio, mc)); | |
3770 | } | |
3dfb57a3 | 3771 | if (error != 0) { |
aa755b35 MA |
3772 | if (error != ENOSPC || |
3773 | (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC)) { | |
3774 | zfs_dbgmsg("%s: metaslab allocation failure: zio %px, " | |
3775 | "size %llu, error %d", | |
8e739b2c RE |
3776 | spa_name(spa), zio, (u_longlong_t)zio->io_size, |
3777 | error); | |
aa755b35 | 3778 | } |
34dc7c2f BB |
3779 | zio->io_error = error; |
3780 | } | |
3781 | ||
62840030 | 3782 | return (zio); |
34dc7c2f BB |
3783 | } |
3784 | ||
62840030 | 3785 | static zio_t * |
34dc7c2f BB |
3786 | zio_dva_free(zio_t *zio) |
3787 | { | |
b128c09f | 3788 | metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); |
34dc7c2f | 3789 | |
62840030 | 3790 | return (zio); |
34dc7c2f BB |
3791 | } |
3792 | ||
62840030 | 3793 | static zio_t * |
34dc7c2f BB |
3794 | zio_dva_claim(zio_t *zio) |
3795 | { | |
b128c09f BB |
3796 | int error; |
3797 | ||
3798 | error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); | |
3799 | if (error) | |
3800 | zio->io_error = error; | |
34dc7c2f | 3801 | |
62840030 | 3802 | return (zio); |
34dc7c2f BB |
3803 | } |
3804 | ||
b128c09f BB |
3805 | /* |
3806 | * Undo an allocation. This is used by zio_done() when an I/O fails | |
3807 | * and we want to give back the block we just allocated. | |
3808 | * This handles both normal blocks and gang blocks. | |
3809 | */ | |
3810 | static void | |
3811 | zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) | |
3812 | { | |
b128c09f | 3813 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); |
428870ff | 3814 | ASSERT(zio->io_bp_override == NULL); |
b128c09f BB |
3815 | |
3816 | if (!BP_IS_HOLE(bp)) | |
428870ff | 3817 | metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); |
b128c09f BB |
3818 | |
3819 | if (gn != NULL) { | |
1c27024e | 3820 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
3821 | zio_dva_unallocate(zio, gn->gn_child[g], |
3822 | &gn->gn_gbh->zg_blkptr[g]); | |
3823 | } | |
3824 | } | |
3825 | } | |
3826 | ||
3827 | /* | |
3828 | * Try to allocate an intent log block. Return 0 on success, errno on failure. | |
3829 | */ | |
3830 | int | |
b5256303 TC |
3831 | zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp, |
3832 | uint64_t size, boolean_t *slog) | |
b128c09f | 3833 | { |
428870ff | 3834 | int error = 1; |
4e21fd06 | 3835 | zio_alloc_list_t io_alloc_list; |
b128c09f | 3836 | |
428870ff BB |
3837 | ASSERT(txg > spa_syncing_txg(spa)); |
3838 | ||
4e21fd06 | 3839 | metaslab_trace_init(&io_alloc_list); |
cc99f275 DB |
3840 | |
3841 | /* | |
3842 | * Block pointer fields are useful to metaslabs for stats and debugging. | |
3843 | * Fill in the obvious ones before calling into metaslab_alloc(). | |
3844 | */ | |
3845 | BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); | |
3846 | BP_SET_PSIZE(new_bp, size); | |
3847 | BP_SET_LEVEL(new_bp, 0); | |
3848 | ||
492f64e9 PD |
3849 | /* |
3850 | * When allocating a zil block, we don't have information about | |
3851 | * the final destination of the block except the objset it's part | |
3852 | * of, so we just hash the objset ID to pick the allocator to get | |
3853 | * some parallelism. | |
3854 | */ | |
b22bab25 | 3855 | int flags = METASLAB_ZIL; |
1b50749c AM |
3856 | int allocator = (uint_t)cityhash4(0, 0, 0, |
3857 | os->os_dsl_dataset->ds_object) % spa->spa_alloc_count; | |
aa755b35 MA |
3858 | error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1, |
3859 | txg, NULL, flags, &io_alloc_list, NULL, allocator); | |
3860 | *slog = (error == 0); | |
3861 | if (error != 0) { | |
3862 | error = metaslab_alloc(spa, spa_embedded_log_class(spa), size, | |
3863 | new_bp, 1, txg, NULL, flags, | |
3864 | &io_alloc_list, NULL, allocator); | |
3865 | } | |
3866 | if (error != 0) { | |
3867 | error = metaslab_alloc(spa, spa_normal_class(spa), size, | |
3868 | new_bp, 1, txg, NULL, flags, | |
3869 | &io_alloc_list, NULL, allocator); | |
ebf8e3a2 | 3870 | } |
4e21fd06 | 3871 | metaslab_trace_fini(&io_alloc_list); |
b128c09f BB |
3872 | |
3873 | if (error == 0) { | |
3874 | BP_SET_LSIZE(new_bp, size); | |
3875 | BP_SET_PSIZE(new_bp, size); | |
3876 | BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); | |
428870ff BB |
3877 | BP_SET_CHECKSUM(new_bp, |
3878 | spa_version(spa) >= SPA_VERSION_SLIM_ZIL | |
3879 | ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); | |
b128c09f BB |
3880 | BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); |
3881 | BP_SET_LEVEL(new_bp, 0); | |
428870ff | 3882 | BP_SET_DEDUP(new_bp, 0); |
b128c09f | 3883 | BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); |
b5256303 TC |
3884 | |
3885 | /* | |
3886 | * encrypted blocks will require an IV and salt. We generate | |
3887 | * these now since we will not be rewriting the bp at | |
3888 | * rewrite time. | |
3889 | */ | |
3890 | if (os->os_encrypted) { | |
3891 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
3892 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
3893 | ||
3894 | BP_SET_CRYPT(new_bp, B_TRUE); | |
3895 | VERIFY0(spa_crypt_get_salt(spa, | |
3896 | dmu_objset_id(os), salt)); | |
3897 | VERIFY0(zio_crypt_generate_iv(iv)); | |
3898 | ||
3899 | zio_crypt_encode_params_bp(new_bp, salt, iv); | |
3900 | } | |
1ce23dca PS |
3901 | } else { |
3902 | zfs_dbgmsg("%s: zil block allocation failure: " | |
8e739b2c RE |
3903 | "size %llu, error %d", spa_name(spa), (u_longlong_t)size, |
3904 | error); | |
b128c09f BB |
3905 | } |
3906 | ||
3907 | return (error); | |
3908 | } | |
3909 | ||
34dc7c2f BB |
3910 | /* |
3911 | * ========================================================================== | |
3912 | * Read and write to physical devices | |
3913 | * ========================================================================== | |
3914 | */ | |
98b25418 | 3915 | |
98b25418 GW |
3916 | /* |
3917 | * Issue an I/O to the underlying vdev. Typically the issue pipeline | |
3918 | * stops after this stage and will resume upon I/O completion. | |
3919 | * However, there are instances where the vdev layer may need to | |
3920 | * continue the pipeline when an I/O was not issued. Since the I/O | |
3921 | * that was sent to the vdev layer might be different than the one | |
3922 | * currently active in the pipeline (see vdev_queue_io()), we explicitly | |
3923 | * force the underlying vdev layers to call either zio_execute() or | |
3924 | * zio_interrupt() to ensure that the pipeline continues with the correct I/O. | |
3925 | */ | |
62840030 | 3926 | static zio_t * |
34dc7c2f BB |
3927 | zio_vdev_io_start(zio_t *zio) |
3928 | { | |
3929 | vdev_t *vd = zio->io_vd; | |
34dc7c2f BB |
3930 | uint64_t align; |
3931 | spa_t *spa = zio->io_spa; | |
3932 | ||
193a37cb TH |
3933 | zio->io_delay = 0; |
3934 | ||
b128c09f BB |
3935 | ASSERT(zio->io_error == 0); |
3936 | ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); | |
34dc7c2f | 3937 | |
b128c09f BB |
3938 | if (vd == NULL) { |
3939 | if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
3940 | spa_config_enter(spa, SCL_ZIO, zio, RW_READER); | |
34dc7c2f | 3941 | |
b128c09f BB |
3942 | /* |
3943 | * The mirror_ops handle multiple DVAs in a single BP. | |
3944 | */ | |
98b25418 | 3945 | vdev_mirror_ops.vdev_op_io_start(zio); |
62840030 | 3946 | return (NULL); |
34dc7c2f BB |
3947 | } |
3948 | ||
3dfb57a3 | 3949 | ASSERT3P(zio->io_logical, !=, zio); |
6cb8e530 PZ |
3950 | if (zio->io_type == ZIO_TYPE_WRITE) { |
3951 | ASSERT(spa->spa_trust_config); | |
3952 | ||
a1d477c2 MA |
3953 | /* |
3954 | * Note: the code can handle other kinds of writes, | |
3955 | * but we don't expect them. | |
3956 | */ | |
2a673e76 | 3957 | if (zio->io_vd->vdev_noalloc) { |
6cb8e530 PZ |
3958 | ASSERT(zio->io_flags & |
3959 | (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL | | |
3960 | ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE)); | |
3961 | } | |
a1d477c2 | 3962 | } |
3dfb57a3 | 3963 | |
b128c09f BB |
3964 | align = 1ULL << vd->vdev_top->vdev_ashift; |
3965 | ||
b02fe35d AR |
3966 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && |
3967 | P2PHASE(zio->io_size, align) != 0) { | |
3968 | /* Transform logical writes to be a full physical block size. */ | |
34dc7c2f | 3969 | uint64_t asize = P2ROUNDUP(zio->io_size, align); |
a6255b7f | 3970 | abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize); |
178e73b3 | 3971 | ASSERT(vd == vd->vdev_top); |
34dc7c2f | 3972 | if (zio->io_type == ZIO_TYPE_WRITE) { |
a6255b7f DQ |
3973 | abd_copy(abuf, zio->io_abd, zio->io_size); |
3974 | abd_zero_off(abuf, zio->io_size, asize - zio->io_size); | |
34dc7c2f | 3975 | } |
b128c09f | 3976 | zio_push_transform(zio, abuf, asize, asize, zio_subblock); |
34dc7c2f BB |
3977 | } |
3978 | ||
b02fe35d AR |
3979 | /* |
3980 | * If this is not a physical io, make sure that it is properly aligned | |
3981 | * before proceeding. | |
3982 | */ | |
3983 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { | |
3984 | ASSERT0(P2PHASE(zio->io_offset, align)); | |
3985 | ASSERT0(P2PHASE(zio->io_size, align)); | |
3986 | } else { | |
3987 | /* | |
3988 | * For physical writes, we allow 512b aligned writes and assume | |
3989 | * the device will perform a read-modify-write as necessary. | |
3990 | */ | |
3991 | ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); | |
3992 | ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); | |
3993 | } | |
3994 | ||
572e2857 | 3995 | VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); |
fb5f0bc8 BB |
3996 | |
3997 | /* | |
3998 | * If this is a repair I/O, and there's no self-healing involved -- | |
3999 | * that is, we're just resilvering what we expect to resilver -- | |
4000 | * then don't do the I/O unless zio's txg is actually in vd's DTL. | |
9e052db4 MA |
4001 | * This prevents spurious resilvering. |
4002 | * | |
4003 | * There are a few ways that we can end up creating these spurious | |
4004 | * resilver i/os: | |
4005 | * | |
4006 | * 1. A resilver i/o will be issued if any DVA in the BP has a | |
4007 | * dirty DTL. The mirror code will issue resilver writes to | |
4008 | * each DVA, including the one(s) that are not on vdevs with dirty | |
4009 | * DTLs. | |
4010 | * | |
4011 | * 2. With nested replication, which happens when we have a | |
4012 | * "replacing" or "spare" vdev that's a child of a mirror or raidz. | |
4013 | * For example, given mirror(replacing(A+B), C), it's likely that | |
4014 | * only A is out of date (it's the new device). In this case, we'll | |
4015 | * read from C, then use the data to resilver A+B -- but we don't | |
4016 | * actually want to resilver B, just A. The top-level mirror has no | |
4017 | * way to know this, so instead we just discard unnecessary repairs | |
4018 | * as we work our way down the vdev tree. | |
4019 | * | |
4020 | * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc. | |
4021 | * The same logic applies to any form of nested replication: ditto | |
4022 | * + mirror, RAID-Z + replacing, etc. | |
4023 | * | |
4024 | * However, indirect vdevs point off to other vdevs which may have | |
4025 | * DTL's, so we never bypass them. The child i/os on concrete vdevs | |
4026 | * will be properly bypassed instead. | |
b2255edc BB |
4027 | * |
4028 | * Leaf DTL_PARTIAL can be empty when a legitimate write comes from | |
4029 | * a dRAID spare vdev. For example, when a dRAID spare is first | |
4030 | * used, its spare blocks need to be written to but the leaf vdev's | |
4031 | * of such blocks can have empty DTL_PARTIAL. | |
4032 | * | |
4033 | * There seemed no clean way to allow such writes while bypassing | |
4034 | * spurious ones. At this point, just avoid all bypassing for dRAID | |
4035 | * for correctness. | |
fb5f0bc8 BB |
4036 | */ |
4037 | if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && | |
4038 | !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && | |
4039 | zio->io_txg != 0 && /* not a delegated i/o */ | |
9e052db4 | 4040 | vd->vdev_ops != &vdev_indirect_ops && |
b2255edc | 4041 | vd->vdev_top->vdev_ops != &vdev_draid_ops && |
fb5f0bc8 BB |
4042 | !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { |
4043 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
fb5f0bc8 | 4044 | zio_vdev_io_bypass(zio); |
62840030 | 4045 | return (zio); |
fb5f0bc8 | 4046 | } |
34dc7c2f | 4047 | |
b2255edc BB |
4048 | /* |
4049 | * Select the next best leaf I/O to process. Distributed spares are | |
4050 | * excluded since they dispatch the I/O directly to a leaf vdev after | |
4051 | * applying the dRAID mapping. | |
4052 | */ | |
4053 | if (vd->vdev_ops->vdev_op_leaf && | |
4054 | vd->vdev_ops != &vdev_draid_spare_ops && | |
4055 | (zio->io_type == ZIO_TYPE_READ || | |
4056 | zio->io_type == ZIO_TYPE_WRITE || | |
4057 | zio->io_type == ZIO_TYPE_TRIM)) { | |
b128c09f | 4058 | |
b128c09f | 4059 | if ((zio = vdev_queue_io(zio)) == NULL) |
62840030 | 4060 | return (NULL); |
b128c09f BB |
4061 | |
4062 | if (!vdev_accessible(vd, zio)) { | |
2e528b49 | 4063 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f | 4064 | zio_interrupt(zio); |
62840030 | 4065 | return (NULL); |
b128c09f | 4066 | } |
67103816 | 4067 | zio->io_delay = gethrtime(); |
b128c09f BB |
4068 | } |
4069 | ||
98b25418 | 4070 | vd->vdev_ops->vdev_op_io_start(zio); |
62840030 | 4071 | return (NULL); |
34dc7c2f BB |
4072 | } |
4073 | ||
62840030 | 4074 | static zio_t * |
34dc7c2f BB |
4075 | zio_vdev_io_done(zio_t *zio) |
4076 | { | |
b128c09f BB |
4077 | vdev_t *vd = zio->io_vd; |
4078 | vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; | |
4079 | boolean_t unexpected_error = B_FALSE; | |
34dc7c2f | 4080 | |
ddc751d5 | 4081 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { |
62840030 | 4082 | return (NULL); |
ddc751d5 | 4083 | } |
34dc7c2f | 4084 | |
1b939560 BB |
4085 | ASSERT(zio->io_type == ZIO_TYPE_READ || |
4086 | zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM); | |
b128c09f | 4087 | |
193a37cb TH |
4088 | if (zio->io_delay) |
4089 | zio->io_delay = gethrtime() - zio->io_delay; | |
4090 | ||
b2255edc BB |
4091 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf && |
4092 | vd->vdev_ops != &vdev_draid_spare_ops) { | |
b128c09f BB |
4093 | vdev_queue_io_done(zio); |
4094 | ||
b128c09f | 4095 | if (zio_injection_enabled && zio->io_error == 0) |
d977122d DB |
4096 | zio->io_error = zio_handle_device_injections(vd, zio, |
4097 | EIO, EILSEQ); | |
b128c09f BB |
4098 | |
4099 | if (zio_injection_enabled && zio->io_error == 0) | |
4100 | zio->io_error = zio_handle_label_injection(zio, EIO); | |
4101 | ||
1b939560 | 4102 | if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) { |
b128c09f | 4103 | if (!vdev_accessible(vd, zio)) { |
2e528b49 | 4104 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
4105 | } else { |
4106 | unexpected_error = B_TRUE; | |
4107 | } | |
4108 | } | |
4109 | } | |
4110 | ||
4111 | ops->vdev_op_io_done(zio); | |
34dc7c2f | 4112 | |
55c12724 | 4113 | if (unexpected_error && vd->vdev_remove_wanted == B_FALSE) |
d164b209 | 4114 | VERIFY(vdev_probe(vd, zio) == NULL); |
34dc7c2f | 4115 | |
62840030 | 4116 | return (zio); |
34dc7c2f BB |
4117 | } |
4118 | ||
a8b2e306 TC |
4119 | /* |
4120 | * This function is used to change the priority of an existing zio that is | |
4121 | * currently in-flight. This is used by the arc to upgrade priority in the | |
4122 | * event that a demand read is made for a block that is currently queued | |
4123 | * as a scrub or async read IO. Otherwise, the high priority read request | |
4124 | * would end up having to wait for the lower priority IO. | |
4125 | */ | |
4126 | void | |
4127 | zio_change_priority(zio_t *pio, zio_priority_t priority) | |
4128 | { | |
4129 | zio_t *cio, *cio_next; | |
4130 | zio_link_t *zl = NULL; | |
4131 | ||
4132 | ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
4133 | ||
4134 | if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) { | |
4135 | vdev_queue_change_io_priority(pio, priority); | |
4136 | } else { | |
4137 | pio->io_priority = priority; | |
4138 | } | |
4139 | ||
4140 | mutex_enter(&pio->io_lock); | |
4141 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { | |
4142 | cio_next = zio_walk_children(pio, &zl); | |
4143 | zio_change_priority(cio, priority); | |
4144 | } | |
4145 | mutex_exit(&pio->io_lock); | |
4146 | } | |
4147 | ||
428870ff BB |
4148 | /* |
4149 | * For non-raidz ZIOs, we can just copy aside the bad data read from the | |
4150 | * disk, and use that to finish the checksum ereport later. | |
4151 | */ | |
4152 | static void | |
4153 | zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, | |
84c07ada | 4154 | const abd_t *good_buf) |
428870ff BB |
4155 | { |
4156 | /* no processing needed */ | |
4157 | zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); | |
4158 | } | |
4159 | ||
428870ff | 4160 | void |
330c6c05 | 4161 | zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr) |
428870ff | 4162 | { |
84c07ada | 4163 | void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size); |
428870ff | 4164 | |
84c07ada | 4165 | abd_copy(abd, zio->io_abd, zio->io_size); |
428870ff BB |
4166 | |
4167 | zcr->zcr_cbinfo = zio->io_size; | |
84c07ada | 4168 | zcr->zcr_cbdata = abd; |
428870ff | 4169 | zcr->zcr_finish = zio_vsd_default_cksum_finish; |
84c07ada | 4170 | zcr->zcr_free = zio_abd_free; |
428870ff BB |
4171 | } |
4172 | ||
62840030 | 4173 | static zio_t * |
34dc7c2f BB |
4174 | zio_vdev_io_assess(zio_t *zio) |
4175 | { | |
4176 | vdev_t *vd = zio->io_vd; | |
b128c09f | 4177 | |
ddc751d5 | 4178 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { |
62840030 | 4179 | return (NULL); |
ddc751d5 | 4180 | } |
b128c09f BB |
4181 | |
4182 | if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
4183 | spa_config_exit(zio->io_spa, SCL_ZIO, zio); | |
4184 | ||
4185 | if (zio->io_vsd != NULL) { | |
428870ff | 4186 | zio->io_vsd_ops->vsd_free(zio); |
b128c09f | 4187 | zio->io_vsd = NULL; |
34dc7c2f BB |
4188 | } |
4189 | ||
b128c09f | 4190 | if (zio_injection_enabled && zio->io_error == 0) |
34dc7c2f BB |
4191 | zio->io_error = zio_handle_fault_injection(zio, EIO); |
4192 | ||
4193 | /* | |
4194 | * If the I/O failed, determine whether we should attempt to retry it. | |
428870ff BB |
4195 | * |
4196 | * On retry, we cut in line in the issue queue, since we don't want | |
4197 | * compression/checksumming/etc. work to prevent our (cheap) IO reissue. | |
34dc7c2f | 4198 | */ |
b128c09f BB |
4199 | if (zio->io_error && vd == NULL && |
4200 | !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { | |
4201 | ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ | |
4202 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ | |
34dc7c2f | 4203 | zio->io_error = 0; |
70ea484e | 4204 | zio->io_flags |= ZIO_FLAG_IO_RETRY | ZIO_FLAG_DONT_AGGREGATE; |
428870ff BB |
4205 | zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; |
4206 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, | |
4207 | zio_requeue_io_start_cut_in_line); | |
62840030 | 4208 | return (NULL); |
34dc7c2f BB |
4209 | } |
4210 | ||
b128c09f BB |
4211 | /* |
4212 | * If we got an error on a leaf device, convert it to ENXIO | |
4213 | * if the device is not accessible at all. | |
4214 | */ | |
4215 | if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && | |
4216 | !vdev_accessible(vd, zio)) | |
2e528b49 | 4217 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
4218 | |
4219 | /* | |
4220 | * If we can't write to an interior vdev (mirror or RAID-Z), | |
4221 | * set vdev_cant_write so that we stop trying to allocate from it. | |
4222 | */ | |
4223 | if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && | |
13fe0198 | 4224 | vd != NULL && !vd->vdev_ops->vdev_op_leaf) { |
2b56a634 MA |
4225 | vdev_dbgmsg(vd, "zio_vdev_io_assess(zio=%px) setting " |
4226 | "cant_write=TRUE due to write failure with ENXIO", | |
4227 | zio); | |
b128c09f | 4228 | vd->vdev_cant_write = B_TRUE; |
13fe0198 | 4229 | } |
b128c09f | 4230 | |
298ec40b GM |
4231 | /* |
4232 | * If a cache flush returns ENOTSUP or ENOTTY, we know that no future | |
1b939560 BB |
4233 | * attempts will ever succeed. In this case we set a persistent |
4234 | * boolean flag so that we don't bother with it in the future. | |
298ec40b GM |
4235 | */ |
4236 | if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) && | |
4237 | zio->io_type == ZIO_TYPE_IOCTL && | |
4238 | zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL) | |
4239 | vd->vdev_nowritecache = B_TRUE; | |
4240 | ||
b128c09f BB |
4241 | if (zio->io_error) |
4242 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
4243 | ||
62840030 | 4244 | return (zio); |
34dc7c2f BB |
4245 | } |
4246 | ||
4247 | void | |
4248 | zio_vdev_io_reissue(zio_t *zio) | |
4249 | { | |
4250 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
4251 | ASSERT(zio->io_error == 0); | |
4252 | ||
428870ff | 4253 | zio->io_stage >>= 1; |
34dc7c2f BB |
4254 | } |
4255 | ||
4256 | void | |
4257 | zio_vdev_io_redone(zio_t *zio) | |
4258 | { | |
4259 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); | |
4260 | ||
428870ff | 4261 | zio->io_stage >>= 1; |
34dc7c2f BB |
4262 | } |
4263 | ||
4264 | void | |
4265 | zio_vdev_io_bypass(zio_t *zio) | |
4266 | { | |
4267 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
4268 | ASSERT(zio->io_error == 0); | |
4269 | ||
4270 | zio->io_flags |= ZIO_FLAG_IO_BYPASS; | |
428870ff | 4271 | zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; |
34dc7c2f BB |
4272 | } |
4273 | ||
b5256303 TC |
4274 | /* |
4275 | * ========================================================================== | |
4276 | * Encrypt and store encryption parameters | |
4277 | * ========================================================================== | |
4278 | */ | |
4279 | ||
4280 | ||
4281 | /* | |
4282 | * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for | |
4283 | * managing the storage of encryption parameters and passing them to the | |
4284 | * lower-level encryption functions. | |
4285 | */ | |
62840030 | 4286 | static zio_t * |
b5256303 TC |
4287 | zio_encrypt(zio_t *zio) |
4288 | { | |
4289 | zio_prop_t *zp = &zio->io_prop; | |
4290 | spa_t *spa = zio->io_spa; | |
4291 | blkptr_t *bp = zio->io_bp; | |
4292 | uint64_t psize = BP_GET_PSIZE(bp); | |
ae76f45c | 4293 | uint64_t dsobj = zio->io_bookmark.zb_objset; |
b5256303 TC |
4294 | dmu_object_type_t ot = BP_GET_TYPE(bp); |
4295 | void *enc_buf = NULL; | |
4296 | abd_t *eabd = NULL; | |
4297 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
4298 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
4299 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
4300 | boolean_t no_crypt = B_FALSE; | |
4301 | ||
4302 | /* the root zio already encrypted the data */ | |
4303 | if (zio->io_child_type == ZIO_CHILD_GANG) | |
62840030 | 4304 | return (zio); |
b5256303 TC |
4305 | |
4306 | /* only ZIL blocks are re-encrypted on rewrite */ | |
4307 | if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG) | |
62840030 | 4308 | return (zio); |
b5256303 TC |
4309 | |
4310 | if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) { | |
4311 | BP_SET_CRYPT(bp, B_FALSE); | |
62840030 | 4312 | return (zio); |
b5256303 TC |
4313 | } |
4314 | ||
4315 | /* if we are doing raw encryption set the provided encryption params */ | |
4316 | if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) { | |
ae76f45c | 4317 | ASSERT0(BP_GET_LEVEL(bp)); |
b5256303 TC |
4318 | BP_SET_CRYPT(bp, B_TRUE); |
4319 | BP_SET_BYTEORDER(bp, zp->zp_byteorder); | |
4320 | if (ot != DMU_OT_OBJSET) | |
4321 | zio_crypt_encode_mac_bp(bp, zp->zp_mac); | |
ae76f45c TC |
4322 | |
4323 | /* dnode blocks must be written out in the provided byteorder */ | |
4324 | if (zp->zp_byteorder != ZFS_HOST_BYTEORDER && | |
4325 | ot == DMU_OT_DNODE) { | |
4326 | void *bswap_buf = zio_buf_alloc(psize); | |
4327 | abd_t *babd = abd_get_from_buf(bswap_buf, psize); | |
4328 | ||
4329 | ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); | |
4330 | abd_copy_to_buf(bswap_buf, zio->io_abd, psize); | |
4331 | dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf, | |
4332 | psize); | |
4333 | ||
4334 | abd_take_ownership_of_buf(babd, B_TRUE); | |
4335 | zio_push_transform(zio, babd, psize, psize, NULL); | |
4336 | } | |
4337 | ||
b5256303 TC |
4338 | if (DMU_OT_IS_ENCRYPTED(ot)) |
4339 | zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv); | |
62840030 | 4340 | return (zio); |
b5256303 TC |
4341 | } |
4342 | ||
4343 | /* indirect blocks only maintain a cksum of the lower level MACs */ | |
4344 | if (BP_GET_LEVEL(bp) > 0) { | |
4345 | BP_SET_CRYPT(bp, B_TRUE); | |
4346 | VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE, | |
4347 | zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp), | |
4348 | mac)); | |
4349 | zio_crypt_encode_mac_bp(bp, mac); | |
62840030 | 4350 | return (zio); |
b5256303 TC |
4351 | } |
4352 | ||
4353 | /* | |
4354 | * Objset blocks are a special case since they have 2 256-bit MACs | |
4355 | * embedded within them. | |
4356 | */ | |
4357 | if (ot == DMU_OT_OBJSET) { | |
4358 | ASSERT0(DMU_OT_IS_ENCRYPTED(ot)); | |
4359 | ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); | |
4360 | BP_SET_CRYPT(bp, B_TRUE); | |
ae76f45c TC |
4361 | VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj, |
4362 | zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp))); | |
62840030 | 4363 | return (zio); |
b5256303 TC |
4364 | } |
4365 | ||
4366 | /* unencrypted object types are only authenticated with a MAC */ | |
4367 | if (!DMU_OT_IS_ENCRYPTED(ot)) { | |
4368 | BP_SET_CRYPT(bp, B_TRUE); | |
ae76f45c TC |
4369 | VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj, |
4370 | zio->io_abd, psize, mac)); | |
b5256303 | 4371 | zio_crypt_encode_mac_bp(bp, mac); |
62840030 | 4372 | return (zio); |
b5256303 TC |
4373 | } |
4374 | ||
4375 | /* | |
4376 | * Later passes of sync-to-convergence may decide to rewrite data | |
4377 | * in place to avoid more disk reallocations. This presents a problem | |
d611989f | 4378 | * for encryption because this constitutes rewriting the new data with |
b5256303 TC |
4379 | * the same encryption key and IV. However, this only applies to blocks |
4380 | * in the MOS (particularly the spacemaps) and we do not encrypt the | |
4381 | * MOS. We assert that the zio is allocating or an intent log write | |
4382 | * to enforce this. | |
4383 | */ | |
4384 | ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG); | |
4385 | ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG); | |
4386 | ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION)); | |
4387 | ASSERT3U(psize, !=, 0); | |
4388 | ||
4389 | enc_buf = zio_buf_alloc(psize); | |
4390 | eabd = abd_get_from_buf(enc_buf, psize); | |
4391 | abd_take_ownership_of_buf(eabd, B_TRUE); | |
4392 | ||
4393 | /* | |
4394 | * For an explanation of what encryption parameters are stored | |
4395 | * where, see the block comment in zio_crypt.c. | |
4396 | */ | |
4397 | if (ot == DMU_OT_INTENT_LOG) { | |
4398 | zio_crypt_decode_params_bp(bp, salt, iv); | |
4399 | } else { | |
4400 | BP_SET_CRYPT(bp, B_TRUE); | |
4401 | } | |
4402 | ||
4403 | /* Perform the encryption. This should not fail */ | |
be9a5c35 TC |
4404 | VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark, |
4405 | BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), | |
4406 | salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt)); | |
b5256303 TC |
4407 | |
4408 | /* encode encryption metadata into the bp */ | |
4409 | if (ot == DMU_OT_INTENT_LOG) { | |
4410 | /* | |
4411 | * ZIL blocks store the MAC in the embedded checksum, so the | |
4412 | * transform must always be applied. | |
4413 | */ | |
4414 | zio_crypt_encode_mac_zil(enc_buf, mac); | |
4415 | zio_push_transform(zio, eabd, psize, psize, NULL); | |
4416 | } else { | |
4417 | BP_SET_CRYPT(bp, B_TRUE); | |
4418 | zio_crypt_encode_params_bp(bp, salt, iv); | |
4419 | zio_crypt_encode_mac_bp(bp, mac); | |
4420 | ||
4421 | if (no_crypt) { | |
4422 | ASSERT3U(ot, ==, DMU_OT_DNODE); | |
4423 | abd_free(eabd); | |
4424 | } else { | |
4425 | zio_push_transform(zio, eabd, psize, psize, NULL); | |
4426 | } | |
4427 | } | |
4428 | ||
62840030 | 4429 | return (zio); |
b5256303 TC |
4430 | } |
4431 | ||
34dc7c2f BB |
4432 | /* |
4433 | * ========================================================================== | |
4434 | * Generate and verify checksums | |
4435 | * ========================================================================== | |
4436 | */ | |
62840030 | 4437 | static zio_t * |
34dc7c2f BB |
4438 | zio_checksum_generate(zio_t *zio) |
4439 | { | |
34dc7c2f | 4440 | blkptr_t *bp = zio->io_bp; |
b128c09f | 4441 | enum zio_checksum checksum; |
34dc7c2f | 4442 | |
b128c09f BB |
4443 | if (bp == NULL) { |
4444 | /* | |
4445 | * This is zio_write_phys(). | |
4446 | * We're either generating a label checksum, or none at all. | |
4447 | */ | |
4448 | checksum = zio->io_prop.zp_checksum; | |
34dc7c2f | 4449 | |
b128c09f | 4450 | if (checksum == ZIO_CHECKSUM_OFF) |
62840030 | 4451 | return (zio); |
b128c09f BB |
4452 | |
4453 | ASSERT(checksum == ZIO_CHECKSUM_LABEL); | |
4454 | } else { | |
4455 | if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { | |
4456 | ASSERT(!IO_IS_ALLOCATING(zio)); | |
4457 | checksum = ZIO_CHECKSUM_GANG_HEADER; | |
4458 | } else { | |
4459 | checksum = BP_GET_CHECKSUM(bp); | |
4460 | } | |
4461 | } | |
34dc7c2f | 4462 | |
a6255b7f | 4463 | zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size); |
34dc7c2f | 4464 | |
62840030 | 4465 | return (zio); |
34dc7c2f BB |
4466 | } |
4467 | ||
62840030 | 4468 | static zio_t * |
b128c09f | 4469 | zio_checksum_verify(zio_t *zio) |
34dc7c2f | 4470 | { |
428870ff | 4471 | zio_bad_cksum_t info; |
b128c09f BB |
4472 | blkptr_t *bp = zio->io_bp; |
4473 | int error; | |
34dc7c2f | 4474 | |
428870ff BB |
4475 | ASSERT(zio->io_vd != NULL); |
4476 | ||
b128c09f BB |
4477 | if (bp == NULL) { |
4478 | /* | |
4479 | * This is zio_read_phys(). | |
4480 | * We're either verifying a label checksum, or nothing at all. | |
4481 | */ | |
4482 | if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) | |
62840030 | 4483 | return (zio); |
34dc7c2f | 4484 | |
b2255edc | 4485 | ASSERT3U(zio->io_prop.zp_checksum, ==, ZIO_CHECKSUM_LABEL); |
b128c09f | 4486 | } |
34dc7c2f | 4487 | |
428870ff | 4488 | if ((error = zio_checksum_error(zio, &info)) != 0) { |
b128c09f | 4489 | zio->io_error = error; |
7a3066ff MA |
4490 | if (error == ECKSUM && |
4491 | !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { | |
03e02e5b DB |
4492 | mutex_enter(&zio->io_vd->vdev_stat_lock); |
4493 | zio->io_vd->vdev_stat.vs_checksum_errors++; | |
4494 | mutex_exit(&zio->io_vd->vdev_stat_lock); | |
7a75f74c RW |
4495 | (void) zfs_ereport_start_checksum(zio->io_spa, |
4496 | zio->io_vd, &zio->io_bookmark, zio, | |
4497 | zio->io_offset, zio->io_size, &info); | |
b128c09f | 4498 | } |
34dc7c2f BB |
4499 | } |
4500 | ||
62840030 | 4501 | return (zio); |
34dc7c2f BB |
4502 | } |
4503 | ||
4504 | /* | |
4505 | * Called by RAID-Z to ensure we don't compute the checksum twice. | |
4506 | */ | |
4507 | void | |
4508 | zio_checksum_verified(zio_t *zio) | |
4509 | { | |
428870ff | 4510 | zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; |
34dc7c2f BB |
4511 | } |
4512 | ||
4513 | /* | |
b128c09f BB |
4514 | * ========================================================================== |
4515 | * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. | |
9b67f605 | 4516 | * An error of 0 indicates success. ENXIO indicates whole-device failure, |
d611989f | 4517 | * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO |
b128c09f BB |
4518 | * indicate errors that are specific to one I/O, and most likely permanent. |
4519 | * Any other error is presumed to be worse because we weren't expecting it. | |
4520 | * ========================================================================== | |
34dc7c2f | 4521 | */ |
b128c09f BB |
4522 | int |
4523 | zio_worst_error(int e1, int e2) | |
34dc7c2f | 4524 | { |
b128c09f BB |
4525 | static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; |
4526 | int r1, r2; | |
4527 | ||
4528 | for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) | |
4529 | if (e1 == zio_error_rank[r1]) | |
4530 | break; | |
34dc7c2f | 4531 | |
b128c09f BB |
4532 | for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) |
4533 | if (e2 == zio_error_rank[r2]) | |
4534 | break; | |
4535 | ||
4536 | return (r1 > r2 ? e1 : e2); | |
34dc7c2f BB |
4537 | } |
4538 | ||
4539 | /* | |
4540 | * ========================================================================== | |
b128c09f | 4541 | * I/O completion |
34dc7c2f BB |
4542 | * ========================================================================== |
4543 | */ | |
62840030 | 4544 | static zio_t * |
b128c09f | 4545 | zio_ready(zio_t *zio) |
34dc7c2f | 4546 | { |
b128c09f | 4547 | blkptr_t *bp = zio->io_bp; |
d164b209 | 4548 | zio_t *pio, *pio_next; |
3dfb57a3 | 4549 | zio_link_t *zl = NULL; |
34dc7c2f | 4550 | |
eda3fcd5 AM |
4551 | if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT | |
4552 | ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT, ZIO_WAIT_READY)) { | |
62840030 | 4553 | return (NULL); |
ddc751d5 | 4554 | } |
34dc7c2f | 4555 | |
9babb374 | 4556 | if (zio->io_ready) { |
b128c09f | 4557 | ASSERT(IO_IS_ALLOCATING(zio)); |
03c6040b GW |
4558 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || |
4559 | (zio->io_flags & ZIO_FLAG_NOPWRITE)); | |
b128c09f | 4560 | ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); |
34dc7c2f | 4561 | |
b128c09f BB |
4562 | zio->io_ready(zio); |
4563 | } | |
34dc7c2f | 4564 | |
b4a08730 | 4565 | #ifdef ZFS_DEBUG |
b128c09f BB |
4566 | if (bp != NULL && bp != &zio->io_bp_copy) |
4567 | zio->io_bp_copy = *bp; | |
b4a08730 | 4568 | #endif |
34dc7c2f | 4569 | |
3dfb57a3 | 4570 | if (zio->io_error != 0) { |
b128c09f | 4571 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f | 4572 | |
3dfb57a3 DB |
4573 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
4574 | ASSERT(IO_IS_ALLOCATING(zio)); | |
4575 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
cc99f275 | 4576 | ASSERT(zio->io_metaslab_class != NULL); |
3bd4df38 | 4577 | ASSERT(ZIO_HAS_ALLOCATOR(zio)); |
cc99f275 | 4578 | |
3dfb57a3 DB |
4579 | /* |
4580 | * We were unable to allocate anything, unreserve and | |
4581 | * issue the next I/O to allocate. | |
4582 | */ | |
4583 | metaslab_class_throttle_unreserve( | |
cc99f275 DB |
4584 | zio->io_metaslab_class, zio->io_prop.zp_copies, |
4585 | zio->io_allocator, zio); | |
492f64e9 | 4586 | zio_allocate_dispatch(zio->io_spa, zio->io_allocator); |
3dfb57a3 DB |
4587 | } |
4588 | } | |
4589 | ||
d164b209 BB |
4590 | mutex_enter(&zio->io_lock); |
4591 | zio->io_state[ZIO_WAIT_READY] = 1; | |
3dfb57a3 | 4592 | pio = zio_walk_parents(zio, &zl); |
d164b209 BB |
4593 | mutex_exit(&zio->io_lock); |
4594 | ||
4595 | /* | |
4596 | * As we notify zio's parents, new parents could be added. | |
4597 | * New parents go to the head of zio's io_parent_list, however, | |
4598 | * so we will (correctly) not notify them. The remainder of zio's | |
4599 | * io_parent_list, from 'pio_next' onward, cannot change because | |
4600 | * all parents must wait for us to be done before they can be done. | |
4601 | */ | |
4602 | for (; pio != NULL; pio = pio_next) { | |
3dfb57a3 | 4603 | pio_next = zio_walk_parents(zio, &zl); |
62840030 | 4604 | zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL); |
d164b209 | 4605 | } |
34dc7c2f | 4606 | |
428870ff | 4607 | if (zio->io_flags & ZIO_FLAG_NODATA) { |
7cb67d62 | 4608 | if (bp != NULL && BP_IS_GANG(bp)) { |
428870ff BB |
4609 | zio->io_flags &= ~ZIO_FLAG_NODATA; |
4610 | } else { | |
a6255b7f | 4611 | ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE); |
428870ff BB |
4612 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; |
4613 | } | |
4614 | } | |
4615 | ||
4616 | if (zio_injection_enabled && | |
4617 | zio->io_spa->spa_syncing_txg == zio->io_txg) | |
4618 | zio_handle_ignored_writes(zio); | |
4619 | ||
62840030 | 4620 | return (zio); |
34dc7c2f BB |
4621 | } |
4622 | ||
3dfb57a3 DB |
4623 | /* |
4624 | * Update the allocation throttle accounting. | |
4625 | */ | |
4626 | static void | |
4627 | zio_dva_throttle_done(zio_t *zio) | |
4628 | { | |
2a8ba608 | 4629 | zio_t *lio __maybe_unused = zio->io_logical; |
3dfb57a3 DB |
4630 | zio_t *pio = zio_unique_parent(zio); |
4631 | vdev_t *vd = zio->io_vd; | |
4632 | int flags = METASLAB_ASYNC_ALLOC; | |
4633 | ||
4634 | ASSERT3P(zio->io_bp, !=, NULL); | |
4635 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); | |
4636 | ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE); | |
4637 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); | |
4638 | ASSERT(vd != NULL); | |
4639 | ASSERT3P(vd, ==, vd->vdev_top); | |
21df134f SB |
4640 | ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY)); |
4641 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); | |
3dfb57a3 DB |
4642 | ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING); |
4643 | ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
4644 | ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA)); | |
4645 | ||
4646 | /* | |
4647 | * Parents of gang children can have two flavors -- ones that | |
4648 | * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set) | |
4649 | * and ones that allocated the constituent blocks. The allocation | |
4650 | * throttle needs to know the allocating parent zio so we must find | |
4651 | * it here. | |
4652 | */ | |
4653 | if (pio->io_child_type == ZIO_CHILD_GANG) { | |
4654 | /* | |
4655 | * If our parent is a rewrite gang child then our grandparent | |
4656 | * would have been the one that performed the allocation. | |
4657 | */ | |
4658 | if (pio->io_flags & ZIO_FLAG_IO_REWRITE) | |
4659 | pio = zio_unique_parent(pio); | |
4660 | flags |= METASLAB_GANG_CHILD; | |
4661 | } | |
4662 | ||
4663 | ASSERT(IO_IS_ALLOCATING(pio)); | |
3bd4df38 | 4664 | ASSERT(ZIO_HAS_ALLOCATOR(pio)); |
3dfb57a3 DB |
4665 | ASSERT3P(zio, !=, zio->io_logical); |
4666 | ASSERT(zio->io_logical != NULL); | |
4667 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); | |
4668 | ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE); | |
cc99f275 | 4669 | ASSERT(zio->io_metaslab_class != NULL); |
3dfb57a3 DB |
4670 | |
4671 | mutex_enter(&pio->io_lock); | |
492f64e9 PD |
4672 | metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags, |
4673 | pio->io_allocator, B_TRUE); | |
3dfb57a3 DB |
4674 | mutex_exit(&pio->io_lock); |
4675 | ||
cc99f275 DB |
4676 | metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1, |
4677 | pio->io_allocator, pio); | |
3dfb57a3 DB |
4678 | |
4679 | /* | |
4680 | * Call into the pipeline to see if there is more work that | |
4681 | * needs to be done. If there is work to be done it will be | |
4682 | * dispatched to another taskq thread. | |
4683 | */ | |
492f64e9 | 4684 | zio_allocate_dispatch(zio->io_spa, pio->io_allocator); |
3dfb57a3 DB |
4685 | } |
4686 | ||
62840030 | 4687 | static zio_t * |
b128c09f | 4688 | zio_done(zio_t *zio) |
34dc7c2f | 4689 | { |
3dfb57a3 DB |
4690 | /* |
4691 | * Always attempt to keep stack usage minimal here since | |
d611989f | 4692 | * we can be called recursively up to 19 levels deep. |
3dfb57a3 | 4693 | */ |
84c07ada | 4694 | const uint64_t psize = zio->io_size; |
d164b209 | 4695 | zio_t *pio, *pio_next; |
3dfb57a3 | 4696 | zio_link_t *zl = NULL; |
34dc7c2f | 4697 | |
b128c09f | 4698 | /* |
9babb374 | 4699 | * If our children haven't all completed, |
b128c09f BB |
4700 | * wait for them and then repeat this pipeline stage. |
4701 | */ | |
ddc751d5 | 4702 | if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) { |
62840030 | 4703 | return (NULL); |
ddc751d5 | 4704 | } |
34dc7c2f | 4705 | |
3dfb57a3 DB |
4706 | /* |
4707 | * If the allocation throttle is enabled, then update the accounting. | |
4708 | * We only track child I/Os that are part of an allocating async | |
4709 | * write. We must do this since the allocation is performed | |
4710 | * by the logical I/O but the actual write is done by child I/Os. | |
4711 | */ | |
4712 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING && | |
4713 | zio->io_child_type == ZIO_CHILD_VDEV) { | |
cc99f275 DB |
4714 | ASSERT(zio->io_metaslab_class != NULL); |
4715 | ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled); | |
3dfb57a3 DB |
4716 | zio_dva_throttle_done(zio); |
4717 | } | |
4718 | ||
4719 | /* | |
4720 | * If the allocation throttle is enabled, verify that | |
4721 | * we have decremented the refcounts for every I/O that was throttled. | |
4722 | */ | |
4723 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { | |
4724 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
4725 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
4726 | ASSERT(zio->io_bp != NULL); | |
3bd4df38 | 4727 | ASSERT(ZIO_HAS_ALLOCATOR(zio)); |
cc99f275 | 4728 | |
492f64e9 PD |
4729 | metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio, |
4730 | zio->io_allocator); | |
f8020c93 AM |
4731 | VERIFY(zfs_refcount_not_held(&zio->io_metaslab_class-> |
4732 | mc_allocator[zio->io_allocator].mca_alloc_slots, zio)); | |
3dfb57a3 DB |
4733 | } |
4734 | ||
4735 | ||
1c27024e DB |
4736 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) |
4737 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) | |
b128c09f BB |
4738 | ASSERT(zio->io_children[c][w] == 0); |
4739 | ||
9b67f605 | 4740 | if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) { |
c776b317 BB |
4741 | ASSERT(zio->io_bp->blk_pad[0] == 0); |
4742 | ASSERT(zio->io_bp->blk_pad[1] == 0); | |
861166b0 | 4743 | ASSERT(memcmp(zio->io_bp, &zio->io_bp_copy, |
d1d7e268 | 4744 | sizeof (blkptr_t)) == 0 || |
c776b317 BB |
4745 | (zio->io_bp == zio_unique_parent(zio)->io_bp)); |
4746 | if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) && | |
428870ff | 4747 | zio->io_bp_override == NULL && |
b128c09f | 4748 | !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { |
d1d7e268 MK |
4749 | ASSERT3U(zio->io_prop.zp_copies, <=, |
4750 | BP_GET_NDVAS(zio->io_bp)); | |
c776b317 | 4751 | ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 || |
d1d7e268 MK |
4752 | (BP_COUNT_GANG(zio->io_bp) == |
4753 | BP_GET_NDVAS(zio->io_bp))); | |
b128c09f | 4754 | } |
03c6040b GW |
4755 | if (zio->io_flags & ZIO_FLAG_NOPWRITE) |
4756 | VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig)); | |
b128c09f BB |
4757 | } |
4758 | ||
4759 | /* | |
428870ff | 4760 | * If there were child vdev/gang/ddt errors, they apply to us now. |
b128c09f BB |
4761 | */ |
4762 | zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); | |
4763 | zio_inherit_child_errors(zio, ZIO_CHILD_GANG); | |
428870ff BB |
4764 | zio_inherit_child_errors(zio, ZIO_CHILD_DDT); |
4765 | ||
4766 | /* | |
4767 | * If the I/O on the transformed data was successful, generate any | |
4768 | * checksum reports now while we still have the transformed data. | |
4769 | */ | |
4770 | if (zio->io_error == 0) { | |
4771 | while (zio->io_cksum_report != NULL) { | |
4772 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
4773 | uint64_t align = zcr->zcr_align; | |
a6255b7f | 4774 | uint64_t asize = P2ROUNDUP(psize, align); |
a6255b7f DQ |
4775 | abd_t *adata = zio->io_abd; |
4776 | ||
f2286383 | 4777 | if (adata != NULL && asize != psize) { |
84c07ada | 4778 | adata = abd_alloc(asize, B_TRUE); |
a6255b7f DQ |
4779 | abd_copy(adata, zio->io_abd, psize); |
4780 | abd_zero_off(adata, psize, asize - psize); | |
428870ff BB |
4781 | } |
4782 | ||
4783 | zio->io_cksum_report = zcr->zcr_next; | |
4784 | zcr->zcr_next = NULL; | |
84c07ada | 4785 | zcr->zcr_finish(zcr, adata); |
428870ff BB |
4786 | zfs_ereport_free_checksum(zcr); |
4787 | ||
f2286383 | 4788 | if (adata != NULL && asize != psize) |
a6255b7f | 4789 | abd_free(adata); |
428870ff BB |
4790 | } |
4791 | } | |
b128c09f BB |
4792 | |
4793 | zio_pop_transforms(zio); /* note: may set zio->io_error */ | |
4794 | ||
a6255b7f | 4795 | vdev_stat_update(zio, psize); |
b128c09f | 4796 | |
a69052be | 4797 | /* |
cc92e9d0 | 4798 | * If this I/O is attached to a particular vdev is slow, exceeding |
72f53c56 MJ |
4799 | * 30 seconds to complete, post an error described the I/O delay. |
4800 | * We ignore these errors if the device is currently unavailable. | |
a69052be | 4801 | */ |
ad796b8a TH |
4802 | if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) { |
4803 | if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) { | |
4804 | /* | |
4805 | * We want to only increment our slow IO counters if | |
4806 | * the IO is valid (i.e. not if the drive is removed). | |
4807 | * | |
4808 | * zfs_ereport_post() will also do these checks, but | |
4809 | * it can also ratelimit and have other failures, so we | |
4810 | * need to increment the slow_io counters independent | |
4811 | * of it. | |
4812 | */ | |
4813 | if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY, | |
4814 | zio->io_spa, zio->io_vd, zio)) { | |
4815 | mutex_enter(&zio->io_vd->vdev_stat_lock); | |
4816 | zio->io_vd->vdev_stat.vs_slow_ios++; | |
4817 | mutex_exit(&zio->io_vd->vdev_stat_lock); | |
4818 | ||
1144586b | 4819 | (void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY, |
ad796b8a | 4820 | zio->io_spa, zio->io_vd, &zio->io_bookmark, |
4f072827 | 4821 | zio, 0); |
ad796b8a TH |
4822 | } |
4823 | } | |
72f53c56 | 4824 | } |
a69052be | 4825 | |
b128c09f BB |
4826 | if (zio->io_error) { |
4827 | /* | |
4828 | * If this I/O is attached to a particular vdev, | |
4829 | * generate an error message describing the I/O failure | |
4830 | * at the block level. We ignore these errors if the | |
4831 | * device is currently unavailable. | |
4832 | */ | |
c776b317 | 4833 | if (zio->io_error != ECKSUM && zio->io_vd != NULL && |
2bbec1c9 | 4834 | !vdev_is_dead(zio->io_vd)) { |
4f072827 DB |
4835 | int ret = zfs_ereport_post(FM_EREPORT_ZFS_IO, |
4836 | zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0); | |
4837 | if (ret != EALREADY) { | |
4838 | mutex_enter(&zio->io_vd->vdev_stat_lock); | |
4839 | if (zio->io_type == ZIO_TYPE_READ) | |
4840 | zio->io_vd->vdev_stat.vs_read_errors++; | |
4841 | else if (zio->io_type == ZIO_TYPE_WRITE) | |
4842 | zio->io_vd->vdev_stat.vs_write_errors++; | |
4843 | mutex_exit(&zio->io_vd->vdev_stat_lock); | |
2bbec1c9 | 4844 | } |
2bbec1c9 | 4845 | } |
34dc7c2f | 4846 | |
428870ff BB |
4847 | if ((zio->io_error == EIO || !(zio->io_flags & |
4848 | (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && | |
c776b317 | 4849 | zio == zio->io_logical) { |
b128c09f BB |
4850 | /* |
4851 | * For logical I/O requests, tell the SPA to log the | |
4852 | * error and generate a logical data ereport. | |
4853 | */ | |
431083f7 GA |
4854 | spa_log_error(zio->io_spa, &zio->io_bookmark, |
4855 | &zio->io_bp->blk_birth); | |
1144586b | 4856 | (void) zfs_ereport_post(FM_EREPORT_ZFS_DATA, |
4f072827 | 4857 | zio->io_spa, NULL, &zio->io_bookmark, zio, 0); |
b128c09f BB |
4858 | } |
4859 | } | |
34dc7c2f | 4860 | |
c776b317 | 4861 | if (zio->io_error && zio == zio->io_logical) { |
b128c09f BB |
4862 | /* |
4863 | * Determine whether zio should be reexecuted. This will | |
4864 | * propagate all the way to the root via zio_notify_parent(). | |
4865 | */ | |
c776b317 | 4866 | ASSERT(zio->io_vd == NULL && zio->io_bp != NULL); |
428870ff | 4867 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); |
b128c09f | 4868 | |
428870ff BB |
4869 | if (IO_IS_ALLOCATING(zio) && |
4870 | !(zio->io_flags & ZIO_FLAG_CANFAIL)) { | |
b128c09f BB |
4871 | if (zio->io_error != ENOSPC) |
4872 | zio->io_reexecute |= ZIO_REEXECUTE_NOW; | |
4873 | else | |
4874 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff | 4875 | } |
b128c09f BB |
4876 | |
4877 | if ((zio->io_type == ZIO_TYPE_READ || | |
4878 | zio->io_type == ZIO_TYPE_FREE) && | |
572e2857 | 4879 | !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && |
b128c09f | 4880 | zio->io_error == ENXIO && |
c776b317 BB |
4881 | spa_load_state(zio->io_spa) == SPA_LOAD_NONE && |
4882 | spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE) | |
b128c09f BB |
4883 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; |
4884 | ||
4885 | if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) | |
4886 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff BB |
4887 | |
4888 | /* | |
4889 | * Here is a possibly good place to attempt to do | |
4890 | * either combinatorial reconstruction or error correction | |
4891 | * based on checksums. It also might be a good place | |
4892 | * to send out preliminary ereports before we suspend | |
4893 | * processing. | |
4894 | */ | |
34dc7c2f BB |
4895 | } |
4896 | ||
4897 | /* | |
b128c09f BB |
4898 | * If there were logical child errors, they apply to us now. |
4899 | * We defer this until now to avoid conflating logical child | |
4900 | * errors with errors that happened to the zio itself when | |
4901 | * updating vdev stats and reporting FMA events above. | |
34dc7c2f | 4902 | */ |
b128c09f | 4903 | zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); |
34dc7c2f | 4904 | |
428870ff BB |
4905 | if ((zio->io_error || zio->io_reexecute) && |
4906 | IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && | |
03c6040b | 4907 | !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) |
c776b317 | 4908 | zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp); |
9babb374 BB |
4909 | |
4910 | zio_gang_tree_free(&zio->io_gang_tree); | |
4911 | ||
4912 | /* | |
4913 | * Godfather I/Os should never suspend. | |
4914 | */ | |
4915 | if ((zio->io_flags & ZIO_FLAG_GODFATHER) && | |
4916 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) | |
a32494d2 | 4917 | zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND; |
9babb374 | 4918 | |
b128c09f BB |
4919 | if (zio->io_reexecute) { |
4920 | /* | |
4921 | * This is a logical I/O that wants to reexecute. | |
4922 | * | |
4923 | * Reexecute is top-down. When an i/o fails, if it's not | |
4924 | * the root, it simply notifies its parent and sticks around. | |
4925 | * The parent, seeing that it still has children in zio_done(), | |
4926 | * does the same. This percolates all the way up to the root. | |
4927 | * The root i/o will reexecute or suspend the entire tree. | |
4928 | * | |
4929 | * This approach ensures that zio_reexecute() honors | |
4930 | * all the original i/o dependency relationships, e.g. | |
4931 | * parents not executing until children are ready. | |
4932 | */ | |
4933 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
34dc7c2f | 4934 | |
9babb374 | 4935 | zio->io_gang_leader = NULL; |
b128c09f | 4936 | |
d164b209 BB |
4937 | mutex_enter(&zio->io_lock); |
4938 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
4939 | mutex_exit(&zio->io_lock); | |
4940 | ||
9babb374 BB |
4941 | /* |
4942 | * "The Godfather" I/O monitors its children but is | |
4943 | * not a true parent to them. It will track them through | |
4944 | * the pipeline but severs its ties whenever they get into | |
4945 | * trouble (e.g. suspended). This allows "The Godfather" | |
4946 | * I/O to return status without blocking. | |
4947 | */ | |
3dfb57a3 DB |
4948 | zl = NULL; |
4949 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; | |
4950 | pio = pio_next) { | |
4951 | zio_link_t *remove_zl = zl; | |
4952 | pio_next = zio_walk_parents(zio, &zl); | |
9babb374 BB |
4953 | |
4954 | if ((pio->io_flags & ZIO_FLAG_GODFATHER) && | |
4955 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { | |
3dfb57a3 | 4956 | zio_remove_child(pio, zio, remove_zl); |
62840030 MA |
4957 | /* |
4958 | * This is a rare code path, so we don't | |
4959 | * bother with "next_to_execute". | |
4960 | */ | |
4961 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, | |
4962 | NULL); | |
9babb374 BB |
4963 | } |
4964 | } | |
4965 | ||
d164b209 | 4966 | if ((pio = zio_unique_parent(zio)) != NULL) { |
b128c09f BB |
4967 | /* |
4968 | * We're not a root i/o, so there's nothing to do | |
4969 | * but notify our parent. Don't propagate errors | |
4970 | * upward since we haven't permanently failed yet. | |
4971 | */ | |
9babb374 | 4972 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f | 4973 | zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; |
62840030 MA |
4974 | /* |
4975 | * This is a rare code path, so we don't bother with | |
4976 | * "next_to_execute". | |
4977 | */ | |
4978 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL); | |
b128c09f BB |
4979 | } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { |
4980 | /* | |
4981 | * We'd fail again if we reexecuted now, so suspend | |
4982 | * until conditions improve (e.g. device comes online). | |
4983 | */ | |
cec3a0a1 | 4984 | zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR); |
b128c09f BB |
4985 | } else { |
4986 | /* | |
4987 | * Reexecution is potentially a huge amount of work. | |
4988 | * Hand it off to the otherwise-unused claim taskq. | |
4989 | */ | |
a38718a6 | 4990 | ASSERT(taskq_empty_ent(&zio->io_tqent)); |
7ef5e54e AL |
4991 | spa_taskq_dispatch_ent(zio->io_spa, |
4992 | ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE, | |
3bd4df38 | 4993 | zio_reexecute, zio, 0, &zio->io_tqent, NULL); |
b128c09f | 4994 | } |
62840030 | 4995 | return (NULL); |
34dc7c2f BB |
4996 | } |
4997 | ||
ccec7fbe | 4998 | ASSERT(list_is_empty(&zio->io_child_list)); |
b128c09f BB |
4999 | ASSERT(zio->io_reexecute == 0); |
5000 | ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); | |
34dc7c2f | 5001 | |
428870ff BB |
5002 | /* |
5003 | * Report any checksum errors, since the I/O is complete. | |
5004 | */ | |
5005 | while (zio->io_cksum_report != NULL) { | |
5006 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
5007 | zio->io_cksum_report = zcr->zcr_next; | |
5008 | zcr->zcr_next = NULL; | |
5009 | zcr->zcr_finish(zcr, NULL); | |
5010 | zfs_ereport_free_checksum(zcr); | |
5011 | } | |
5012 | ||
d164b209 BB |
5013 | /* |
5014 | * It is the responsibility of the done callback to ensure that this | |
5015 | * particular zio is no longer discoverable for adoption, and as | |
5016 | * such, cannot acquire any new parents. | |
5017 | */ | |
b128c09f BB |
5018 | if (zio->io_done) |
5019 | zio->io_done(zio); | |
34dc7c2f | 5020 | |
d164b209 BB |
5021 | mutex_enter(&zio->io_lock); |
5022 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
5023 | mutex_exit(&zio->io_lock); | |
34dc7c2f | 5024 | |
62840030 MA |
5025 | /* |
5026 | * We are done executing this zio. We may want to execute a parent | |
5027 | * next. See the comment in zio_notify_parent(). | |
5028 | */ | |
5029 | zio_t *next_to_execute = NULL; | |
3dfb57a3 DB |
5030 | zl = NULL; |
5031 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) { | |
5032 | zio_link_t *remove_zl = zl; | |
5033 | pio_next = zio_walk_parents(zio, &zl); | |
5034 | zio_remove_child(pio, zio, remove_zl); | |
62840030 | 5035 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute); |
b128c09f | 5036 | } |
34dc7c2f | 5037 | |
b128c09f BB |
5038 | if (zio->io_waiter != NULL) { |
5039 | mutex_enter(&zio->io_lock); | |
5040 | zio->io_executor = NULL; | |
5041 | cv_broadcast(&zio->io_cv); | |
5042 | mutex_exit(&zio->io_lock); | |
5043 | } else { | |
5044 | zio_destroy(zio); | |
5045 | } | |
34dc7c2f | 5046 | |
62840030 | 5047 | return (next_to_execute); |
34dc7c2f BB |
5048 | } |
5049 | ||
5050 | /* | |
b128c09f BB |
5051 | * ========================================================================== |
5052 | * I/O pipeline definition | |
5053 | * ========================================================================== | |
34dc7c2f | 5054 | */ |
428870ff | 5055 | static zio_pipe_stage_t *zio_pipeline[] = { |
b128c09f | 5056 | NULL, |
b128c09f | 5057 | zio_read_bp_init, |
3dfb57a3 | 5058 | zio_write_bp_init, |
428870ff BB |
5059 | zio_free_bp_init, |
5060 | zio_issue_async, | |
3dfb57a3 | 5061 | zio_write_compress, |
b5256303 | 5062 | zio_encrypt, |
b128c09f | 5063 | zio_checksum_generate, |
03c6040b | 5064 | zio_nop_write, |
67a1b037 | 5065 | zio_brt_free, |
428870ff BB |
5066 | zio_ddt_read_start, |
5067 | zio_ddt_read_done, | |
5068 | zio_ddt_write, | |
5069 | zio_ddt_free, | |
b128c09f BB |
5070 | zio_gang_assemble, |
5071 | zio_gang_issue, | |
3dfb57a3 | 5072 | zio_dva_throttle, |
b128c09f BB |
5073 | zio_dva_allocate, |
5074 | zio_dva_free, | |
5075 | zio_dva_claim, | |
5076 | zio_ready, | |
5077 | zio_vdev_io_start, | |
5078 | zio_vdev_io_done, | |
5079 | zio_vdev_io_assess, | |
5080 | zio_checksum_verify, | |
5081 | zio_done | |
5082 | }; | |
c28b2279 | 5083 | |
9ae529ec | 5084 | |
9ae529ec | 5085 | |
9ae529ec | 5086 | |
fcff0f35 PD |
5087 | /* |
5088 | * Compare two zbookmark_phys_t's to see which we would reach first in a | |
5089 | * pre-order traversal of the object tree. | |
5090 | * | |
5091 | * This is simple in every case aside from the meta-dnode object. For all other | |
5092 | * objects, we traverse them in order (object 1 before object 2, and so on). | |
5093 | * However, all of these objects are traversed while traversing object 0, since | |
5094 | * the data it points to is the list of objects. Thus, we need to convert to a | |
5095 | * canonical representation so we can compare meta-dnode bookmarks to | |
5096 | * non-meta-dnode bookmarks. | |
5097 | * | |
5098 | * We do this by calculating "equivalents" for each field of the zbookmark. | |
5099 | * zbookmarks outside of the meta-dnode use their own object and level, and | |
5100 | * calculate the level 0 equivalent (the first L0 blkid that is contained in the | |
5101 | * blocks this bookmark refers to) by multiplying their blkid by their span | |
5102 | * (the number of L0 blocks contained within one block at their level). | |
5103 | * zbookmarks inside the meta-dnode calculate their object equivalent | |
5104 | * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use | |
5105 | * level + 1<<31 (any value larger than a level could ever be) for their level. | |
5106 | * This causes them to always compare before a bookmark in their object | |
5107 | * equivalent, compare appropriately to bookmarks in other objects, and to | |
5108 | * compare appropriately to other bookmarks in the meta-dnode. | |
5109 | */ | |
5110 | int | |
5111 | zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, | |
5112 | const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) | |
5113 | { | |
5114 | /* | |
5115 | * These variables represent the "equivalent" values for the zbookmark, | |
5116 | * after converting zbookmarks inside the meta dnode to their | |
5117 | * normal-object equivalents. | |
5118 | */ | |
5119 | uint64_t zb1obj, zb2obj; | |
5120 | uint64_t zb1L0, zb2L0; | |
5121 | uint64_t zb1level, zb2level; | |
5122 | ||
5123 | if (zb1->zb_object == zb2->zb_object && | |
5124 | zb1->zb_level == zb2->zb_level && | |
5125 | zb1->zb_blkid == zb2->zb_blkid) | |
5126 | return (0); | |
9ae529ec | 5127 | |
30af21b0 PD |
5128 | IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT); |
5129 | IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT); | |
5130 | ||
fcff0f35 PD |
5131 | /* |
5132 | * BP_SPANB calculates the span in blocks. | |
5133 | */ | |
5134 | zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); | |
5135 | zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); | |
9ae529ec CS |
5136 | |
5137 | if (zb1->zb_object == DMU_META_DNODE_OBJECT) { | |
fcff0f35 PD |
5138 | zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); |
5139 | zb1L0 = 0; | |
5140 | zb1level = zb1->zb_level + COMPARE_META_LEVEL; | |
5141 | } else { | |
5142 | zb1obj = zb1->zb_object; | |
5143 | zb1level = zb1->zb_level; | |
9ae529ec CS |
5144 | } |
5145 | ||
fcff0f35 PD |
5146 | if (zb2->zb_object == DMU_META_DNODE_OBJECT) { |
5147 | zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); | |
5148 | zb2L0 = 0; | |
5149 | zb2level = zb2->zb_level + COMPARE_META_LEVEL; | |
5150 | } else { | |
5151 | zb2obj = zb2->zb_object; | |
5152 | zb2level = zb2->zb_level; | |
5153 | } | |
5154 | ||
5155 | /* Now that we have a canonical representation, do the comparison. */ | |
5156 | if (zb1obj != zb2obj) | |
5157 | return (zb1obj < zb2obj ? -1 : 1); | |
5158 | else if (zb1L0 != zb2L0) | |
5159 | return (zb1L0 < zb2L0 ? -1 : 1); | |
5160 | else if (zb1level != zb2level) | |
5161 | return (zb1level > zb2level ? -1 : 1); | |
5162 | /* | |
5163 | * This can (theoretically) happen if the bookmarks have the same object | |
5164 | * and level, but different blkids, if the block sizes are not the same. | |
5165 | * There is presently no way to change the indirect block sizes | |
5166 | */ | |
5167 | return (0); | |
5168 | } | |
5169 | ||
5170 | /* | |
5171 | * This function checks the following: given that last_block is the place that | |
5172 | * our traversal stopped last time, does that guarantee that we've visited | |
5173 | * every node under subtree_root? Therefore, we can't just use the raw output | |
5174 | * of zbookmark_compare. We have to pass in a modified version of | |
5175 | * subtree_root; by incrementing the block id, and then checking whether | |
5176 | * last_block is before or equal to that, we can tell whether or not having | |
5177 | * visited last_block implies that all of subtree_root's children have been | |
5178 | * visited. | |
5179 | */ | |
5180 | boolean_t | |
5181 | zbookmark_subtree_completed(const dnode_phys_t *dnp, | |
5182 | const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) | |
5183 | { | |
5184 | zbookmark_phys_t mod_zb = *subtree_root; | |
5185 | mod_zb.zb_blkid++; | |
33dba8c7 | 5186 | ASSERT0(last_block->zb_level); |
fcff0f35 PD |
5187 | |
5188 | /* The objset_phys_t isn't before anything. */ | |
5189 | if (dnp == NULL) | |
9ae529ec | 5190 | return (B_FALSE); |
fcff0f35 PD |
5191 | |
5192 | /* | |
5193 | * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the | |
5194 | * data block size in sectors, because that variable is only used if | |
5195 | * the bookmark refers to a block in the meta-dnode. Since we don't | |
5196 | * know without examining it what object it refers to, and there's no | |
5197 | * harm in passing in this value in other cases, we always pass it in. | |
5198 | * | |
5199 | * We pass in 0 for the indirect block size shift because zb2 must be | |
5200 | * level 0. The indirect block size is only used to calculate the span | |
5201 | * of the bookmark, but since the bookmark must be level 0, the span is | |
5202 | * always 1, so the math works out. | |
5203 | * | |
5204 | * If you make changes to how the zbookmark_compare code works, be sure | |
5205 | * to make sure that this code still works afterwards. | |
5206 | */ | |
5207 | return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, | |
5208 | 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, | |
5209 | last_block) <= 0); | |
9ae529ec CS |
5210 | } |
5211 | ||
33dba8c7 AM |
5212 | /* |
5213 | * This function is similar to zbookmark_subtree_completed(), but returns true | |
5214 | * if subtree_root is equal or ahead of last_block, i.e. still to be done. | |
5215 | */ | |
5216 | boolean_t | |
5217 | zbookmark_subtree_tbd(const dnode_phys_t *dnp, | |
5218 | const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) | |
5219 | { | |
5220 | ASSERT0(last_block->zb_level); | |
5221 | if (dnp == NULL) | |
5222 | return (B_FALSE); | |
5223 | return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, | |
5224 | 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, subtree_root, | |
5225 | last_block) >= 0); | |
5226 | } | |
5227 | ||
c28b2279 | 5228 | EXPORT_SYMBOL(zio_type_name); |
81971b13 BB |
5229 | EXPORT_SYMBOL(zio_buf_alloc); |
5230 | EXPORT_SYMBOL(zio_data_buf_alloc); | |
5231 | EXPORT_SYMBOL(zio_buf_free); | |
5232 | EXPORT_SYMBOL(zio_data_buf_free); | |
c28b2279 | 5233 | |
03fdcb9a | 5234 | ZFS_MODULE_PARAM(zfs_zio, zio_, slow_io_ms, INT, ZMOD_RW, |
ad796b8a | 5235 | "Max I/O completion time (milliseconds) before marking it as slow"); |
c409e464 | 5236 | |
03fdcb9a MM |
5237 | ZFS_MODULE_PARAM(zfs_zio, zio_, requeue_io_start_cut_in_line, INT, ZMOD_RW, |
5238 | "Prioritize requeued I/O"); | |
29dee3ee | 5239 | |
fdc2d303 | 5240 | ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_deferred_free, UINT, ZMOD_RW, |
d1d7e268 | 5241 | "Defer frees starting in this pass"); |
29dee3ee | 5242 | |
fdc2d303 | 5243 | ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_dont_compress, UINT, ZMOD_RW, |
d1d7e268 | 5244 | "Don't compress starting in this pass"); |
29dee3ee | 5245 | |
fdc2d303 | 5246 | ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_rewrite, UINT, ZMOD_RW, |
d1d7e268 | 5247 | "Rewrite new bps starting in this pass"); |
3dfb57a3 | 5248 | |
03fdcb9a | 5249 | ZFS_MODULE_PARAM(zfs_zio, zio_, dva_throttle_enabled, INT, ZMOD_RW, |
3dfb57a3 | 5250 | "Throttle block allocations in the ZIO pipeline"); |
638dd5f4 | 5251 | |
03fdcb9a | 5252 | ZFS_MODULE_PARAM(zfs_zio, zio_, deadman_log_all, INT, ZMOD_RW, |
638dd5f4 | 5253 | "Log all slow ZIOs, not just those with vdevs"); |