]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
1 | /* |
2 | * CDDL HEADER START | |
3 | * | |
4 | * The contents of this file are subject to the terms of the | |
5 | * Common Development and Distribution License (the "License"). | |
6 | * You may not use this file except in compliance with the License. | |
7 | * | |
8 | * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE | |
9 | * or http://www.opensolaris.org/os/licensing. | |
10 | * See the License for the specific language governing permissions | |
11 | * and limitations under the License. | |
12 | * | |
13 | * When distributing Covered Code, include this CDDL HEADER in each | |
14 | * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 | * If applicable, add the following below this CDDL HEADER, with the | |
16 | * fields enclosed by brackets "[]" replaced with your own identifying | |
17 | * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | /* | |
428870ff | 22 | * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. |
492f64e9 | 23 | * Copyright (c) 2011, 2018 by Delphix. All rights reserved. |
a38718a6 | 24 | * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved. |
cc99f275 | 25 | * Copyright (c) 2017, Intel Corporation. |
34dc7c2f BB |
26 | */ |
27 | ||
f1512ee6 | 28 | #include <sys/sysmacros.h> |
34dc7c2f BB |
29 | #include <sys/zfs_context.h> |
30 | #include <sys/fm/fs/zfs.h> | |
31 | #include <sys/spa.h> | |
32 | #include <sys/txg.h> | |
33 | #include <sys/spa_impl.h> | |
34 | #include <sys/vdev_impl.h> | |
35 | #include <sys/zio_impl.h> | |
36 | #include <sys/zio_compress.h> | |
37 | #include <sys/zio_checksum.h> | |
428870ff BB |
38 | #include <sys/dmu_objset.h> |
39 | #include <sys/arc.h> | |
40 | #include <sys/ddt.h> | |
9b67f605 | 41 | #include <sys/blkptr.h> |
b0bc7a84 | 42 | #include <sys/zfeature.h> |
d4a72f23 | 43 | #include <sys/dsl_scan.h> |
3dfb57a3 | 44 | #include <sys/metaslab_impl.h> |
193a37cb | 45 | #include <sys/time.h> |
26ef0cc7 | 46 | #include <sys/trace_zio.h> |
a6255b7f | 47 | #include <sys/abd.h> |
b5256303 | 48 | #include <sys/dsl_crypt.h> |
492f64e9 | 49 | #include <sys/cityhash.h> |
34dc7c2f | 50 | |
34dc7c2f BB |
51 | /* |
52 | * ========================================================================== | |
53 | * I/O type descriptions | |
54 | * ========================================================================== | |
55 | */ | |
e8b96c60 | 56 | const char *zio_type_name[ZIO_TYPES] = { |
3dfb57a3 DB |
57 | /* |
58 | * Note: Linux kernel thread name length is limited | |
59 | * so these names will differ from upstream open zfs. | |
60 | */ | |
451041db | 61 | "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl" |
428870ff | 62 | }; |
34dc7c2f | 63 | |
27f2b90d | 64 | int zio_dva_throttle_enabled = B_TRUE; |
3dfb57a3 | 65 | |
34dc7c2f BB |
66 | /* |
67 | * ========================================================================== | |
68 | * I/O kmem caches | |
69 | * ========================================================================== | |
70 | */ | |
71 | kmem_cache_t *zio_cache; | |
d164b209 | 72 | kmem_cache_t *zio_link_cache; |
34dc7c2f BB |
73 | kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; |
74 | kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
a6255b7f DQ |
75 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
76 | uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
77 | uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
78 | #endif | |
79 | ||
a69052be | 80 | int zio_delay_max = ZIO_DELAY_MAX; |
34dc7c2f | 81 | |
fcff0f35 PD |
82 | #define BP_SPANB(indblkshift, level) \ |
83 | (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) | |
84 | #define COMPARE_META_LEVEL 0x80000000ul | |
55d85d5a GW |
85 | /* |
86 | * The following actions directly effect the spa's sync-to-convergence logic. | |
87 | * The values below define the sync pass when we start performing the action. | |
88 | * Care should be taken when changing these values as they directly impact | |
89 | * spa_sync() performance. Tuning these values may introduce subtle performance | |
90 | * pathologies and should only be done in the context of performance analysis. | |
91 | * These tunables will eventually be removed and replaced with #defines once | |
92 | * enough analysis has been done to determine optimal values. | |
93 | * | |
94 | * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that | |
95 | * regular blocks are not deferred. | |
96 | */ | |
97 | int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ | |
98 | int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */ | |
99 | int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ | |
100 | ||
34dc7c2f | 101 | /* |
b128c09f BB |
102 | * An allocating zio is one that either currently has the DVA allocate |
103 | * stage set or will have it later in its lifetime. | |
34dc7c2f | 104 | */ |
428870ff BB |
105 | #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) |
106 | ||
c409e464 | 107 | int zio_requeue_io_start_cut_in_line = 1; |
428870ff BB |
108 | |
109 | #ifdef ZFS_DEBUG | |
110 | int zio_buf_debug_limit = 16384; | |
111 | #else | |
112 | int zio_buf_debug_limit = 0; | |
113 | #endif | |
34dc7c2f | 114 | |
da6b4005 NB |
115 | static inline void __zio_execute(zio_t *zio); |
116 | ||
3dfb57a3 DB |
117 | static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t); |
118 | ||
34dc7c2f BB |
119 | void |
120 | zio_init(void) | |
121 | { | |
122 | size_t c; | |
123 | vmem_t *data_alloc_arena = NULL; | |
124 | ||
3941503c BB |
125 | zio_cache = kmem_cache_create("zio_cache", |
126 | sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); | |
d164b209 | 127 | zio_link_cache = kmem_cache_create("zio_link_cache", |
6795a698 | 128 | sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); |
34dc7c2f BB |
129 | |
130 | /* | |
131 | * For small buffers, we want a cache for each multiple of | |
f1512ee6 MA |
132 | * SPA_MINBLOCKSIZE. For larger buffers, we want a cache |
133 | * for each quarter-power of 2. | |
34dc7c2f BB |
134 | */ |
135 | for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { | |
136 | size_t size = (c + 1) << SPA_MINBLOCKSHIFT; | |
137 | size_t p2 = size; | |
138 | size_t align = 0; | |
6442f3cf | 139 | size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0; |
34dc7c2f | 140 | |
34328f3c | 141 | #if defined(_ILP32) && defined(_KERNEL) |
f1512ee6 MA |
142 | /* |
143 | * Cache size limited to 1M on 32-bit platforms until ARC | |
144 | * buffers no longer require virtual address space. | |
145 | */ | |
146 | if (size > zfs_max_recordsize) | |
147 | break; | |
148 | #endif | |
149 | ||
150 | while (!ISP2(p2)) | |
34dc7c2f BB |
151 | p2 &= p2 - 1; |
152 | ||
498877ba MA |
153 | #ifndef _KERNEL |
154 | /* | |
155 | * If we are using watchpoints, put each buffer on its own page, | |
156 | * to eliminate the performance overhead of trapping to the | |
157 | * kernel when modifying a non-watched buffer that shares the | |
158 | * page with a watched buffer. | |
159 | */ | |
160 | if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) | |
161 | continue; | |
fcf64f45 BB |
162 | /* |
163 | * Here's the problem - on 4K native devices in userland on | |
164 | * Linux using O_DIRECT, buffers must be 4K aligned or I/O | |
165 | * will fail with EINVAL, causing zdb (and others) to coredump. | |
166 | * Since userland probably doesn't need optimized buffer caches, | |
167 | * we just force 4K alignment on everything. | |
168 | */ | |
169 | align = 8 * SPA_MINBLOCKSIZE; | |
170 | #else | |
24fa2034 | 171 | if (size < PAGESIZE) { |
34dc7c2f | 172 | align = SPA_MINBLOCKSIZE; |
498877ba | 173 | } else if (IS_P2ALIGNED(size, p2 >> 2)) { |
24fa2034 | 174 | align = PAGESIZE; |
34dc7c2f | 175 | } |
fcf64f45 | 176 | #endif |
34dc7c2f BB |
177 | |
178 | if (align != 0) { | |
179 | char name[36]; | |
180 | (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); | |
181 | zio_buf_cache[c] = kmem_cache_create(name, size, | |
6442f3cf | 182 | align, NULL, NULL, NULL, NULL, NULL, cflags); |
34dc7c2f BB |
183 | |
184 | (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); | |
185 | zio_data_buf_cache[c] = kmem_cache_create(name, size, | |
ae6ba3db | 186 | align, NULL, NULL, NULL, NULL, |
6442f3cf | 187 | data_alloc_arena, cflags); |
34dc7c2f BB |
188 | } |
189 | } | |
190 | ||
191 | while (--c != 0) { | |
192 | ASSERT(zio_buf_cache[c] != NULL); | |
193 | if (zio_buf_cache[c - 1] == NULL) | |
194 | zio_buf_cache[c - 1] = zio_buf_cache[c]; | |
195 | ||
196 | ASSERT(zio_data_buf_cache[c] != NULL); | |
197 | if (zio_data_buf_cache[c - 1] == NULL) | |
198 | zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; | |
199 | } | |
200 | ||
34dc7c2f | 201 | zio_inject_init(); |
9759c60f ED |
202 | |
203 | lz4_init(); | |
34dc7c2f BB |
204 | } |
205 | ||
206 | void | |
207 | zio_fini(void) | |
208 | { | |
209 | size_t c; | |
210 | kmem_cache_t *last_cache = NULL; | |
211 | kmem_cache_t *last_data_cache = NULL; | |
212 | ||
213 | for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { | |
f1512ee6 MA |
214 | #ifdef _ILP32 |
215 | /* | |
216 | * Cache size limited to 1M on 32-bit platforms until ARC | |
217 | * buffers no longer require virtual address space. | |
218 | */ | |
219 | if (((c + 1) << SPA_MINBLOCKSHIFT) > zfs_max_recordsize) | |
220 | break; | |
a6255b7f DQ |
221 | #endif |
222 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) | |
223 | if (zio_buf_cache_allocs[c] != zio_buf_cache_frees[c]) | |
224 | (void) printf("zio_fini: [%d] %llu != %llu\n", | |
225 | (int)((c + 1) << SPA_MINBLOCKSHIFT), | |
226 | (long long unsigned)zio_buf_cache_allocs[c], | |
227 | (long long unsigned)zio_buf_cache_frees[c]); | |
f1512ee6 | 228 | #endif |
34dc7c2f BB |
229 | if (zio_buf_cache[c] != last_cache) { |
230 | last_cache = zio_buf_cache[c]; | |
231 | kmem_cache_destroy(zio_buf_cache[c]); | |
232 | } | |
233 | zio_buf_cache[c] = NULL; | |
234 | ||
235 | if (zio_data_buf_cache[c] != last_data_cache) { | |
236 | last_data_cache = zio_data_buf_cache[c]; | |
237 | kmem_cache_destroy(zio_data_buf_cache[c]); | |
238 | } | |
239 | zio_data_buf_cache[c] = NULL; | |
240 | } | |
241 | ||
d164b209 | 242 | kmem_cache_destroy(zio_link_cache); |
34dc7c2f BB |
243 | kmem_cache_destroy(zio_cache); |
244 | ||
245 | zio_inject_fini(); | |
9759c60f ED |
246 | |
247 | lz4_fini(); | |
34dc7c2f BB |
248 | } |
249 | ||
250 | /* | |
251 | * ========================================================================== | |
252 | * Allocate and free I/O buffers | |
253 | * ========================================================================== | |
254 | */ | |
255 | ||
256 | /* | |
257 | * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a | |
258 | * crashdump if the kernel panics, so use it judiciously. Obviously, it's | |
259 | * useful to inspect ZFS metadata, but if possible, we should avoid keeping | |
260 | * excess / transient data in-core during a crashdump. | |
261 | */ | |
262 | void * | |
263 | zio_buf_alloc(size_t size) | |
264 | { | |
265 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
266 | ||
63e3a861 | 267 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
a6255b7f DQ |
268 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
269 | atomic_add_64(&zio_buf_cache_allocs[c], 1); | |
270 | #endif | |
34dc7c2f | 271 | |
efcd79a8 | 272 | return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); |
34dc7c2f BB |
273 | } |
274 | ||
275 | /* | |
276 | * Use zio_data_buf_alloc to allocate data. The data will not appear in a | |
277 | * crashdump if the kernel panics. This exists so that we will limit the amount | |
278 | * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount | |
279 | * of kernel heap dumped to disk when the kernel panics) | |
280 | */ | |
281 | void * | |
282 | zio_data_buf_alloc(size_t size) | |
283 | { | |
284 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
285 | ||
63e3a861 | 286 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f | 287 | |
efcd79a8 | 288 | return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); |
34dc7c2f BB |
289 | } |
290 | ||
291 | void | |
292 | zio_buf_free(void *buf, size_t size) | |
293 | { | |
294 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
295 | ||
63e3a861 | 296 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
a6255b7f DQ |
297 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
298 | atomic_add_64(&zio_buf_cache_frees[c], 1); | |
299 | #endif | |
34dc7c2f BB |
300 | |
301 | kmem_cache_free(zio_buf_cache[c], buf); | |
302 | } | |
303 | ||
304 | void | |
305 | zio_data_buf_free(void *buf, size_t size) | |
306 | { | |
307 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
308 | ||
63e3a861 | 309 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f BB |
310 | |
311 | kmem_cache_free(zio_data_buf_cache[c], buf); | |
312 | } | |
313 | ||
84c07ada GN |
314 | static void |
315 | zio_abd_free(void *abd, size_t size) | |
316 | { | |
317 | abd_free((abd_t *)abd); | |
318 | } | |
319 | ||
34dc7c2f BB |
320 | /* |
321 | * ========================================================================== | |
322 | * Push and pop I/O transform buffers | |
323 | * ========================================================================== | |
324 | */ | |
d3c2ae1c | 325 | void |
a6255b7f | 326 | zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize, |
e9aa730c | 327 | zio_transform_func_t *transform) |
34dc7c2f | 328 | { |
79c76d5b | 329 | zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); |
34dc7c2f | 330 | |
a6255b7f DQ |
331 | /* |
332 | * Ensure that anyone expecting this zio to contain a linear ABD isn't | |
333 | * going to get a nasty surprise when they try to access the data. | |
334 | */ | |
335 | IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data)); | |
336 | ||
337 | zt->zt_orig_abd = zio->io_abd; | |
b128c09f | 338 | zt->zt_orig_size = zio->io_size; |
34dc7c2f | 339 | zt->zt_bufsize = bufsize; |
b128c09f | 340 | zt->zt_transform = transform; |
34dc7c2f BB |
341 | |
342 | zt->zt_next = zio->io_transform_stack; | |
343 | zio->io_transform_stack = zt; | |
344 | ||
a6255b7f | 345 | zio->io_abd = data; |
34dc7c2f BB |
346 | zio->io_size = size; |
347 | } | |
348 | ||
d3c2ae1c | 349 | void |
b128c09f | 350 | zio_pop_transforms(zio_t *zio) |
34dc7c2f | 351 | { |
b128c09f BB |
352 | zio_transform_t *zt; |
353 | ||
354 | while ((zt = zio->io_transform_stack) != NULL) { | |
355 | if (zt->zt_transform != NULL) | |
356 | zt->zt_transform(zio, | |
a6255b7f | 357 | zt->zt_orig_abd, zt->zt_orig_size); |
34dc7c2f | 358 | |
428870ff | 359 | if (zt->zt_bufsize != 0) |
a6255b7f | 360 | abd_free(zio->io_abd); |
34dc7c2f | 361 | |
a6255b7f | 362 | zio->io_abd = zt->zt_orig_abd; |
b128c09f BB |
363 | zio->io_size = zt->zt_orig_size; |
364 | zio->io_transform_stack = zt->zt_next; | |
34dc7c2f | 365 | |
b128c09f | 366 | kmem_free(zt, sizeof (zio_transform_t)); |
34dc7c2f BB |
367 | } |
368 | } | |
369 | ||
b128c09f BB |
370 | /* |
371 | * ========================================================================== | |
b5256303 | 372 | * I/O transform callbacks for subblocks, decompression, and decryption |
b128c09f BB |
373 | * ========================================================================== |
374 | */ | |
375 | static void | |
a6255b7f | 376 | zio_subblock(zio_t *zio, abd_t *data, uint64_t size) |
b128c09f BB |
377 | { |
378 | ASSERT(zio->io_size > size); | |
379 | ||
380 | if (zio->io_type == ZIO_TYPE_READ) | |
a6255b7f | 381 | abd_copy(data, zio->io_abd, size); |
b128c09f BB |
382 | } |
383 | ||
384 | static void | |
a6255b7f | 385 | zio_decompress(zio_t *zio, abd_t *data, uint64_t size) |
b128c09f | 386 | { |
a6255b7f DQ |
387 | if (zio->io_error == 0) { |
388 | void *tmp = abd_borrow_buf(data, size); | |
389 | int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), | |
390 | zio->io_abd, tmp, zio->io_size, size); | |
391 | abd_return_buf_copy(data, tmp, size); | |
392 | ||
c3bd3fb4 TC |
393 | if (zio_injection_enabled && ret == 0) |
394 | ret = zio_handle_fault_injection(zio, EINVAL); | |
395 | ||
a6255b7f DQ |
396 | if (ret != 0) |
397 | zio->io_error = SET_ERROR(EIO); | |
398 | } | |
b128c09f BB |
399 | } |
400 | ||
b5256303 TC |
401 | static void |
402 | zio_decrypt(zio_t *zio, abd_t *data, uint64_t size) | |
403 | { | |
404 | int ret; | |
405 | void *tmp; | |
406 | blkptr_t *bp = zio->io_bp; | |
ae76f45c TC |
407 | spa_t *spa = zio->io_spa; |
408 | uint64_t dsobj = zio->io_bookmark.zb_objset; | |
b5256303 TC |
409 | uint64_t lsize = BP_GET_LSIZE(bp); |
410 | dmu_object_type_t ot = BP_GET_TYPE(bp); | |
411 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
412 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
413 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
414 | boolean_t no_crypt = B_FALSE; | |
415 | ||
416 | ASSERT(BP_USES_CRYPT(bp)); | |
417 | ASSERT3U(size, !=, 0); | |
418 | ||
419 | if (zio->io_error != 0) | |
420 | return; | |
421 | ||
422 | /* | |
423 | * Verify the cksum of MACs stored in an indirect bp. It will always | |
424 | * be possible to verify this since it does not require an encryption | |
425 | * key. | |
426 | */ | |
427 | if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) { | |
428 | zio_crypt_decode_mac_bp(bp, mac); | |
429 | ||
430 | if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) { | |
431 | /* | |
432 | * We haven't decompressed the data yet, but | |
433 | * zio_crypt_do_indirect_mac_checksum() requires | |
434 | * decompressed data to be able to parse out the MACs | |
435 | * from the indirect block. We decompress it now and | |
436 | * throw away the result after we are finished. | |
437 | */ | |
438 | tmp = zio_buf_alloc(lsize); | |
439 | ret = zio_decompress_data(BP_GET_COMPRESS(bp), | |
440 | zio->io_abd, tmp, zio->io_size, lsize); | |
441 | if (ret != 0) { | |
442 | ret = SET_ERROR(EIO); | |
443 | goto error; | |
444 | } | |
445 | ret = zio_crypt_do_indirect_mac_checksum(B_FALSE, | |
446 | tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac); | |
447 | zio_buf_free(tmp, lsize); | |
448 | } else { | |
449 | ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE, | |
450 | zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac); | |
451 | } | |
452 | abd_copy(data, zio->io_abd, size); | |
453 | ||
be9a5c35 TC |
454 | if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) { |
455 | ret = zio_handle_decrypt_injection(spa, | |
456 | &zio->io_bookmark, ot, ECKSUM); | |
457 | } | |
b5256303 TC |
458 | if (ret != 0) |
459 | goto error; | |
460 | ||
461 | return; | |
462 | } | |
463 | ||
464 | /* | |
465 | * If this is an authenticated block, just check the MAC. It would be | |
466 | * nice to separate this out into its own flag, but for the moment | |
467 | * enum zio_flag is out of bits. | |
468 | */ | |
469 | if (BP_IS_AUTHENTICATED(bp)) { | |
470 | if (ot == DMU_OT_OBJSET) { | |
ae76f45c TC |
471 | ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, |
472 | dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp)); | |
b5256303 TC |
473 | } else { |
474 | zio_crypt_decode_mac_bp(bp, mac); | |
ae76f45c TC |
475 | ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, |
476 | zio->io_abd, size, mac); | |
be9a5c35 TC |
477 | if (zio_injection_enabled && ret == 0) { |
478 | ret = zio_handle_decrypt_injection(spa, | |
479 | &zio->io_bookmark, ot, ECKSUM); | |
480 | } | |
b5256303 TC |
481 | } |
482 | abd_copy(data, zio->io_abd, size); | |
483 | ||
484 | if (ret != 0) | |
485 | goto error; | |
486 | ||
487 | return; | |
488 | } | |
489 | ||
490 | zio_crypt_decode_params_bp(bp, salt, iv); | |
491 | ||
492 | if (ot == DMU_OT_INTENT_LOG) { | |
493 | tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t)); | |
494 | zio_crypt_decode_mac_zil(tmp, mac); | |
495 | abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t)); | |
496 | } else { | |
497 | zio_crypt_decode_mac_bp(bp, mac); | |
498 | } | |
499 | ||
be9a5c35 TC |
500 | ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp), |
501 | BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data, | |
502 | zio->io_abd, &no_crypt); | |
b5256303 TC |
503 | if (no_crypt) |
504 | abd_copy(data, zio->io_abd, size); | |
505 | ||
506 | if (ret != 0) | |
507 | goto error; | |
508 | ||
509 | return; | |
510 | ||
511 | error: | |
512 | /* assert that the key was found unless this was speculative */ | |
be9a5c35 | 513 | ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE)); |
b5256303 TC |
514 | |
515 | /* | |
516 | * If there was a decryption / authentication error return EIO as | |
517 | * the io_error. If this was not a speculative zio, create an ereport. | |
518 | */ | |
519 | if (ret == ECKSUM) { | |
a2c2ed1b | 520 | zio->io_error = SET_ERROR(EIO); |
b5256303 | 521 | if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) { |
be9a5c35 | 522 | spa_log_error(spa, &zio->io_bookmark); |
b5256303 | 523 | zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION, |
ae76f45c | 524 | spa, NULL, &zio->io_bookmark, zio, 0, 0); |
b5256303 TC |
525 | } |
526 | } else { | |
527 | zio->io_error = ret; | |
528 | } | |
529 | } | |
530 | ||
b128c09f BB |
531 | /* |
532 | * ========================================================================== | |
533 | * I/O parent/child relationships and pipeline interlocks | |
534 | * ========================================================================== | |
535 | */ | |
d164b209 | 536 | zio_t * |
3dfb57a3 | 537 | zio_walk_parents(zio_t *cio, zio_link_t **zl) |
d164b209 | 538 | { |
d164b209 | 539 | list_t *pl = &cio->io_parent_list; |
b128c09f | 540 | |
3dfb57a3 DB |
541 | *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl); |
542 | if (*zl == NULL) | |
d164b209 BB |
543 | return (NULL); |
544 | ||
3dfb57a3 DB |
545 | ASSERT((*zl)->zl_child == cio); |
546 | return ((*zl)->zl_parent); | |
d164b209 BB |
547 | } |
548 | ||
549 | zio_t * | |
3dfb57a3 | 550 | zio_walk_children(zio_t *pio, zio_link_t **zl) |
d164b209 | 551 | { |
d164b209 BB |
552 | list_t *cl = &pio->io_child_list; |
553 | ||
a8b2e306 TC |
554 | ASSERT(MUTEX_HELD(&pio->io_lock)); |
555 | ||
3dfb57a3 DB |
556 | *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl); |
557 | if (*zl == NULL) | |
d164b209 BB |
558 | return (NULL); |
559 | ||
3dfb57a3 DB |
560 | ASSERT((*zl)->zl_parent == pio); |
561 | return ((*zl)->zl_child); | |
d164b209 BB |
562 | } |
563 | ||
564 | zio_t * | |
565 | zio_unique_parent(zio_t *cio) | |
566 | { | |
3dfb57a3 DB |
567 | zio_link_t *zl = NULL; |
568 | zio_t *pio = zio_walk_parents(cio, &zl); | |
d164b209 | 569 | |
3dfb57a3 | 570 | VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL); |
d164b209 BB |
571 | return (pio); |
572 | } | |
573 | ||
574 | void | |
575 | zio_add_child(zio_t *pio, zio_t *cio) | |
b128c09f | 576 | { |
79c76d5b | 577 | zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); |
d164b209 BB |
578 | |
579 | /* | |
580 | * Logical I/Os can have logical, gang, or vdev children. | |
581 | * Gang I/Os can have gang or vdev children. | |
582 | * Vdev I/Os can only have vdev children. | |
583 | * The following ASSERT captures all of these constraints. | |
584 | */ | |
1ce23dca | 585 | ASSERT3S(cio->io_child_type, <=, pio->io_child_type); |
d164b209 BB |
586 | |
587 | zl->zl_parent = pio; | |
588 | zl->zl_child = cio; | |
589 | ||
b128c09f | 590 | mutex_enter(&pio->io_lock); |
a8b2e306 | 591 | mutex_enter(&cio->io_lock); |
d164b209 BB |
592 | |
593 | ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); | |
594 | ||
1c27024e | 595 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 BB |
596 | pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; |
597 | ||
598 | list_insert_head(&pio->io_child_list, zl); | |
599 | list_insert_head(&cio->io_parent_list, zl); | |
600 | ||
428870ff BB |
601 | pio->io_child_count++; |
602 | cio->io_parent_count++; | |
603 | ||
d164b209 | 604 | mutex_exit(&cio->io_lock); |
a8b2e306 | 605 | mutex_exit(&pio->io_lock); |
b128c09f BB |
606 | } |
607 | ||
34dc7c2f | 608 | static void |
d164b209 | 609 | zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) |
b128c09f | 610 | { |
d164b209 BB |
611 | ASSERT(zl->zl_parent == pio); |
612 | ASSERT(zl->zl_child == cio); | |
b128c09f BB |
613 | |
614 | mutex_enter(&pio->io_lock); | |
a8b2e306 | 615 | mutex_enter(&cio->io_lock); |
d164b209 BB |
616 | |
617 | list_remove(&pio->io_child_list, zl); | |
618 | list_remove(&cio->io_parent_list, zl); | |
619 | ||
428870ff BB |
620 | pio->io_child_count--; |
621 | cio->io_parent_count--; | |
622 | ||
d164b209 | 623 | mutex_exit(&cio->io_lock); |
a8b2e306 | 624 | mutex_exit(&pio->io_lock); |
d164b209 | 625 | kmem_cache_free(zio_link_cache, zl); |
b128c09f BB |
626 | } |
627 | ||
628 | static boolean_t | |
ddc751d5 | 629 | zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait) |
34dc7c2f | 630 | { |
b128c09f BB |
631 | boolean_t waiting = B_FALSE; |
632 | ||
633 | mutex_enter(&zio->io_lock); | |
634 | ASSERT(zio->io_stall == NULL); | |
ddc751d5 GW |
635 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) { |
636 | if (!(ZIO_CHILD_BIT_IS_SET(childbits, c))) | |
637 | continue; | |
638 | ||
639 | uint64_t *countp = &zio->io_children[c][wait]; | |
640 | if (*countp != 0) { | |
641 | zio->io_stage >>= 1; | |
642 | ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN); | |
643 | zio->io_stall = countp; | |
644 | waiting = B_TRUE; | |
645 | break; | |
646 | } | |
b128c09f BB |
647 | } |
648 | mutex_exit(&zio->io_lock); | |
b128c09f BB |
649 | return (waiting); |
650 | } | |
34dc7c2f | 651 | |
bf701a83 BB |
652 | __attribute__((always_inline)) |
653 | static inline void | |
62840030 MA |
654 | zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait, |
655 | zio_t **next_to_executep) | |
b128c09f BB |
656 | { |
657 | uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; | |
658 | int *errorp = &pio->io_child_error[zio->io_child_type]; | |
34dc7c2f | 659 | |
b128c09f BB |
660 | mutex_enter(&pio->io_lock); |
661 | if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) | |
662 | *errorp = zio_worst_error(*errorp, zio->io_error); | |
663 | pio->io_reexecute |= zio->io_reexecute; | |
664 | ASSERT3U(*countp, >, 0); | |
e8b96c60 MA |
665 | |
666 | (*countp)--; | |
667 | ||
668 | if (*countp == 0 && pio->io_stall == countp) { | |
3dfb57a3 DB |
669 | zio_taskq_type_t type = |
670 | pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE : | |
671 | ZIO_TASKQ_INTERRUPT; | |
b128c09f BB |
672 | pio->io_stall = NULL; |
673 | mutex_exit(&pio->io_lock); | |
62840030 | 674 | |
3dfb57a3 | 675 | /* |
62840030 MA |
676 | * If we can tell the caller to execute this parent next, do |
677 | * so. Otherwise dispatch the parent zio as its own task. | |
678 | * | |
679 | * Having the caller execute the parent when possible reduces | |
680 | * locking on the zio taskq's, reduces context switch | |
681 | * overhead, and has no recursion penalty. Note that one | |
682 | * read from disk typically causes at least 3 zio's: a | |
683 | * zio_null(), the logical zio_read(), and then a physical | |
684 | * zio. When the physical ZIO completes, we are able to call | |
685 | * zio_done() on all 3 of these zio's from one invocation of | |
686 | * zio_execute() by returning the parent back to | |
687 | * zio_execute(). Since the parent isn't executed until this | |
688 | * thread returns back to zio_execute(), the caller should do | |
689 | * so promptly. | |
690 | * | |
691 | * In other cases, dispatching the parent prevents | |
692 | * overflowing the stack when we have deeply nested | |
693 | * parent-child relationships, as we do with the "mega zio" | |
694 | * of writes for spa_sync(), and the chain of ZIL blocks. | |
3dfb57a3 | 695 | */ |
62840030 MA |
696 | if (next_to_executep != NULL && *next_to_executep == NULL) { |
697 | *next_to_executep = pio; | |
698 | } else { | |
699 | zio_taskq_dispatch(pio, type, B_FALSE); | |
700 | } | |
b128c09f BB |
701 | } else { |
702 | mutex_exit(&pio->io_lock); | |
34dc7c2f BB |
703 | } |
704 | } | |
705 | ||
b128c09f BB |
706 | static void |
707 | zio_inherit_child_errors(zio_t *zio, enum zio_child c) | |
708 | { | |
709 | if (zio->io_child_error[c] != 0 && zio->io_error == 0) | |
710 | zio->io_error = zio->io_child_error[c]; | |
711 | } | |
712 | ||
3dfb57a3 | 713 | int |
64fc7762 | 714 | zio_bookmark_compare(const void *x1, const void *x2) |
3dfb57a3 DB |
715 | { |
716 | const zio_t *z1 = x1; | |
717 | const zio_t *z2 = x2; | |
3dfb57a3 | 718 | |
64fc7762 MA |
719 | if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset) |
720 | return (-1); | |
721 | if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset) | |
722 | return (1); | |
3dfb57a3 | 723 | |
64fc7762 MA |
724 | if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object) |
725 | return (-1); | |
726 | if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object) | |
727 | return (1); | |
3dfb57a3 | 728 | |
64fc7762 MA |
729 | if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level) |
730 | return (-1); | |
731 | if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level) | |
732 | return (1); | |
733 | ||
734 | if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid) | |
735 | return (-1); | |
736 | if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid) | |
737 | return (1); | |
738 | ||
739 | if (z1 < z2) | |
740 | return (-1); | |
741 | if (z1 > z2) | |
742 | return (1); | |
743 | ||
744 | return (0); | |
3dfb57a3 DB |
745 | } |
746 | ||
34dc7c2f BB |
747 | /* |
748 | * ========================================================================== | |
b128c09f | 749 | * Create the various types of I/O (read, write, free, etc) |
34dc7c2f BB |
750 | * ========================================================================== |
751 | */ | |
752 | static zio_t * | |
428870ff | 753 | zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
a6255b7f | 754 | abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done, |
2aa34383 DK |
755 | void *private, zio_type_t type, zio_priority_t priority, |
756 | enum zio_flag flags, vdev_t *vd, uint64_t offset, | |
757 | const zbookmark_phys_t *zb, enum zio_stage stage, | |
758 | enum zio_stage pipeline) | |
34dc7c2f BB |
759 | { |
760 | zio_t *zio; | |
761 | ||
2aa34383 DK |
762 | ASSERT3U(psize, <=, SPA_MAXBLOCKSIZE); |
763 | ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); | |
b128c09f BB |
764 | ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); |
765 | ||
766 | ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); | |
767 | ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); | |
768 | ASSERT(vd || stage == ZIO_STAGE_OPEN); | |
34dc7c2f | 769 | |
b5256303 | 770 | IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0); |
2aa34383 | 771 | |
79c76d5b | 772 | zio = kmem_cache_alloc(zio_cache, KM_SLEEP); |
3941503c BB |
773 | bzero(zio, sizeof (zio_t)); |
774 | ||
448d7aaa | 775 | mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL); |
3941503c BB |
776 | cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); |
777 | ||
778 | list_create(&zio->io_parent_list, sizeof (zio_link_t), | |
779 | offsetof(zio_link_t, zl_parent_node)); | |
780 | list_create(&zio->io_child_list, sizeof (zio_link_t), | |
781 | offsetof(zio_link_t, zl_child_node)); | |
4e21fd06 | 782 | metaslab_trace_init(&zio->io_alloc_list); |
d164b209 | 783 | |
b128c09f BB |
784 | if (vd != NULL) |
785 | zio->io_child_type = ZIO_CHILD_VDEV; | |
786 | else if (flags & ZIO_FLAG_GANG_CHILD) | |
787 | zio->io_child_type = ZIO_CHILD_GANG; | |
428870ff BB |
788 | else if (flags & ZIO_FLAG_DDT_CHILD) |
789 | zio->io_child_type = ZIO_CHILD_DDT; | |
b128c09f BB |
790 | else |
791 | zio->io_child_type = ZIO_CHILD_LOGICAL; | |
792 | ||
34dc7c2f | 793 | if (bp != NULL) { |
428870ff | 794 | zio->io_bp = (blkptr_t *)bp; |
34dc7c2f BB |
795 | zio->io_bp_copy = *bp; |
796 | zio->io_bp_orig = *bp; | |
428870ff BB |
797 | if (type != ZIO_TYPE_WRITE || |
798 | zio->io_child_type == ZIO_CHILD_DDT) | |
b128c09f | 799 | zio->io_bp = &zio->io_bp_copy; /* so caller can free */ |
9babb374 | 800 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) |
b128c09f | 801 | zio->io_logical = zio; |
9babb374 BB |
802 | if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) |
803 | pipeline |= ZIO_GANG_STAGES; | |
34dc7c2f | 804 | } |
b128c09f BB |
805 | |
806 | zio->io_spa = spa; | |
807 | zio->io_txg = txg; | |
34dc7c2f BB |
808 | zio->io_done = done; |
809 | zio->io_private = private; | |
810 | zio->io_type = type; | |
811 | zio->io_priority = priority; | |
b128c09f BB |
812 | zio->io_vd = vd; |
813 | zio->io_offset = offset; | |
a6255b7f | 814 | zio->io_orig_abd = zio->io_abd = data; |
2aa34383 DK |
815 | zio->io_orig_size = zio->io_size = psize; |
816 | zio->io_lsize = lsize; | |
b128c09f BB |
817 | zio->io_orig_flags = zio->io_flags = flags; |
818 | zio->io_orig_stage = zio->io_stage = stage; | |
819 | zio->io_orig_pipeline = zio->io_pipeline = pipeline; | |
3dfb57a3 | 820 | zio->io_pipeline_trace = ZIO_STAGE_OPEN; |
34dc7c2f | 821 | |
d164b209 BB |
822 | zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); |
823 | zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); | |
824 | ||
b128c09f BB |
825 | if (zb != NULL) |
826 | zio->io_bookmark = *zb; | |
827 | ||
828 | if (pio != NULL) { | |
cc99f275 DB |
829 | if (zio->io_metaslab_class == NULL) |
830 | zio->io_metaslab_class = pio->io_metaslab_class; | |
b128c09f | 831 | if (zio->io_logical == NULL) |
34dc7c2f | 832 | zio->io_logical = pio->io_logical; |
9babb374 BB |
833 | if (zio->io_child_type == ZIO_CHILD_GANG) |
834 | zio->io_gang_leader = pio->io_gang_leader; | |
b128c09f | 835 | zio_add_child(pio, zio); |
34dc7c2f BB |
836 | } |
837 | ||
a38718a6 GA |
838 | taskq_init_ent(&zio->io_tqent); |
839 | ||
34dc7c2f BB |
840 | return (zio); |
841 | } | |
842 | ||
843 | static void | |
b128c09f | 844 | zio_destroy(zio_t *zio) |
34dc7c2f | 845 | { |
4e21fd06 | 846 | metaslab_trace_fini(&zio->io_alloc_list); |
3941503c BB |
847 | list_destroy(&zio->io_parent_list); |
848 | list_destroy(&zio->io_child_list); | |
849 | mutex_destroy(&zio->io_lock); | |
850 | cv_destroy(&zio->io_cv); | |
b128c09f | 851 | kmem_cache_free(zio_cache, zio); |
34dc7c2f BB |
852 | } |
853 | ||
854 | zio_t * | |
d164b209 | 855 | zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, |
428870ff | 856 | void *private, enum zio_flag flags) |
34dc7c2f BB |
857 | { |
858 | zio_t *zio; | |
859 | ||
2aa34383 | 860 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
d164b209 | 861 | ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
b128c09f | 862 | ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); |
34dc7c2f BB |
863 | |
864 | return (zio); | |
865 | } | |
866 | ||
867 | zio_t * | |
428870ff | 868 | zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) |
34dc7c2f | 869 | { |
d164b209 | 870 | return (zio_null(NULL, spa, NULL, done, private, flags)); |
34dc7c2f BB |
871 | } |
872 | ||
63e3a861 MA |
873 | void |
874 | zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp) | |
875 | { | |
63e3a861 MA |
876 | if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { |
877 | zfs_panic_recover("blkptr at %p has invalid TYPE %llu", | |
878 | bp, (longlong_t)BP_GET_TYPE(bp)); | |
879 | } | |
880 | if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS || | |
881 | BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) { | |
882 | zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu", | |
883 | bp, (longlong_t)BP_GET_CHECKSUM(bp)); | |
884 | } | |
885 | if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS || | |
886 | BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) { | |
887 | zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu", | |
888 | bp, (longlong_t)BP_GET_COMPRESS(bp)); | |
889 | } | |
890 | if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
891 | zfs_panic_recover("blkptr at %p has invalid LSIZE %llu", | |
892 | bp, (longlong_t)BP_GET_LSIZE(bp)); | |
893 | } | |
894 | if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
895 | zfs_panic_recover("blkptr at %p has invalid PSIZE %llu", | |
896 | bp, (longlong_t)BP_GET_PSIZE(bp)); | |
897 | } | |
898 | ||
899 | if (BP_IS_EMBEDDED(bp)) { | |
900 | if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) { | |
901 | zfs_panic_recover("blkptr at %p has invalid ETYPE %llu", | |
902 | bp, (longlong_t)BPE_GET_ETYPE(bp)); | |
903 | } | |
904 | } | |
905 | ||
6cb8e530 PZ |
906 | /* |
907 | * Do not verify individual DVAs if the config is not trusted. This | |
908 | * will be done once the zio is executed in vdev_mirror_map_alloc. | |
909 | */ | |
910 | if (!spa->spa_trust_config) | |
911 | return; | |
912 | ||
63e3a861 MA |
913 | /* |
914 | * Pool-specific checks. | |
915 | * | |
916 | * Note: it would be nice to verify that the blk_birth and | |
917 | * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() | |
918 | * allows the birth time of log blocks (and dmu_sync()-ed blocks | |
919 | * that are in the log) to be arbitrarily large. | |
920 | */ | |
1c27024e | 921 | for (int i = 0; i < BP_GET_NDVAS(bp); i++) { |
63e3a861 | 922 | uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]); |
1c27024e | 923 | |
63e3a861 MA |
924 | if (vdevid >= spa->spa_root_vdev->vdev_children) { |
925 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
926 | "VDEV %llu", | |
927 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 928 | continue; |
63e3a861 | 929 | } |
1c27024e | 930 | vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; |
63e3a861 MA |
931 | if (vd == NULL) { |
932 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
933 | "VDEV %llu", | |
934 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 935 | continue; |
63e3a861 MA |
936 | } |
937 | if (vd->vdev_ops == &vdev_hole_ops) { | |
938 | zfs_panic_recover("blkptr at %p DVA %u has hole " | |
939 | "VDEV %llu", | |
940 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 941 | continue; |
63e3a861 MA |
942 | } |
943 | if (vd->vdev_ops == &vdev_missing_ops) { | |
944 | /* | |
945 | * "missing" vdevs are valid during import, but we | |
946 | * don't have their detailed info (e.g. asize), so | |
947 | * we can't perform any more checks on them. | |
948 | */ | |
949 | continue; | |
950 | } | |
1c27024e DB |
951 | uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]); |
952 | uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]); | |
63e3a861 MA |
953 | if (BP_IS_GANG(bp)) |
954 | asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
955 | if (offset + asize > vd->vdev_asize) { | |
956 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
957 | "OFFSET %llu", | |
958 | bp, i, (longlong_t)offset); | |
959 | } | |
960 | } | |
961 | } | |
962 | ||
6cb8e530 PZ |
963 | boolean_t |
964 | zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp) | |
965 | { | |
966 | uint64_t vdevid = DVA_GET_VDEV(dva); | |
967 | ||
968 | if (vdevid >= spa->spa_root_vdev->vdev_children) | |
969 | return (B_FALSE); | |
970 | ||
971 | vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; | |
972 | if (vd == NULL) | |
973 | return (B_FALSE); | |
974 | ||
975 | if (vd->vdev_ops == &vdev_hole_ops) | |
976 | return (B_FALSE); | |
977 | ||
978 | if (vd->vdev_ops == &vdev_missing_ops) { | |
979 | return (B_FALSE); | |
980 | } | |
981 | ||
982 | uint64_t offset = DVA_GET_OFFSET(dva); | |
983 | uint64_t asize = DVA_GET_ASIZE(dva); | |
984 | ||
985 | if (BP_IS_GANG(bp)) | |
986 | asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
987 | if (offset + asize > vd->vdev_asize) | |
988 | return (B_FALSE); | |
989 | ||
990 | return (B_TRUE); | |
991 | } | |
992 | ||
34dc7c2f | 993 | zio_t * |
b128c09f | 994 | zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, |
a6255b7f | 995 | abd_t *data, uint64_t size, zio_done_func_t *done, void *private, |
5dbd68a3 | 996 | zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) |
34dc7c2f BB |
997 | { |
998 | zio_t *zio; | |
999 | ||
63e3a861 MA |
1000 | zfs_blkptr_verify(spa, bp); |
1001 | ||
428870ff | 1002 | zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, |
2aa34383 | 1003 | data, size, size, done, private, |
b128c09f | 1004 | ZIO_TYPE_READ, priority, flags, NULL, 0, zb, |
428870ff BB |
1005 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
1006 | ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); | |
34dc7c2f | 1007 | |
b128c09f BB |
1008 | return (zio); |
1009 | } | |
34dc7c2f | 1010 | |
34dc7c2f | 1011 | zio_t * |
b128c09f | 1012 | zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, |
a6255b7f | 1013 | abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp, |
bc77ba73 PD |
1014 | zio_done_func_t *ready, zio_done_func_t *children_ready, |
1015 | zio_done_func_t *physdone, zio_done_func_t *done, | |
1016 | void *private, zio_priority_t priority, enum zio_flag flags, | |
1017 | const zbookmark_phys_t *zb) | |
34dc7c2f BB |
1018 | { |
1019 | zio_t *zio; | |
1020 | ||
b128c09f BB |
1021 | ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && |
1022 | zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && | |
1023 | zp->zp_compress >= ZIO_COMPRESS_OFF && | |
1024 | zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && | |
9ae529ec | 1025 | DMU_OT_IS_VALID(zp->zp_type) && |
b128c09f | 1026 | zp->zp_level < 32 && |
428870ff | 1027 | zp->zp_copies > 0 && |
03c6040b | 1028 | zp->zp_copies <= spa_max_replication(spa)); |
34dc7c2f | 1029 | |
2aa34383 | 1030 | zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private, |
b128c09f | 1031 | ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, |
428870ff BB |
1032 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
1033 | ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); | |
34dc7c2f BB |
1034 | |
1035 | zio->io_ready = ready; | |
bc77ba73 | 1036 | zio->io_children_ready = children_ready; |
e8b96c60 | 1037 | zio->io_physdone = physdone; |
b128c09f | 1038 | zio->io_prop = *zp; |
34dc7c2f | 1039 | |
9b67f605 MA |
1040 | /* |
1041 | * Data can be NULL if we are going to call zio_write_override() to | |
1042 | * provide the already-allocated BP. But we may need the data to | |
1043 | * verify a dedup hit (if requested). In this case, don't try to | |
b5256303 TC |
1044 | * dedup (just take the already-allocated BP verbatim). Encrypted |
1045 | * dedup blocks need data as well so we also disable dedup in this | |
1046 | * case. | |
9b67f605 | 1047 | */ |
b5256303 TC |
1048 | if (data == NULL && |
1049 | (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) { | |
9b67f605 MA |
1050 | zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; |
1051 | } | |
1052 | ||
34dc7c2f BB |
1053 | return (zio); |
1054 | } | |
1055 | ||
1056 | zio_t * | |
a6255b7f | 1057 | zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data, |
e8b96c60 | 1058 | uint64_t size, zio_done_func_t *done, void *private, |
5dbd68a3 | 1059 | zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) |
34dc7c2f BB |
1060 | { |
1061 | zio_t *zio; | |
1062 | ||
2aa34383 | 1063 | zio = zio_create(pio, spa, txg, bp, data, size, size, done, private, |
3dfb57a3 | 1064 | ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb, |
b128c09f | 1065 | ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); |
34dc7c2f BB |
1066 | |
1067 | return (zio); | |
1068 | } | |
1069 | ||
428870ff | 1070 | void |
03c6040b | 1071 | zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) |
428870ff BB |
1072 | { |
1073 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
1074 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1075 | ASSERT(zio->io_stage == ZIO_STAGE_OPEN); | |
1076 | ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); | |
1077 | ||
03c6040b GW |
1078 | /* |
1079 | * We must reset the io_prop to match the values that existed | |
1080 | * when the bp was first written by dmu_sync() keeping in mind | |
1081 | * that nopwrite and dedup are mutually exclusive. | |
1082 | */ | |
1083 | zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; | |
1084 | zio->io_prop.zp_nopwrite = nopwrite; | |
428870ff BB |
1085 | zio->io_prop.zp_copies = copies; |
1086 | zio->io_bp_override = bp; | |
1087 | } | |
1088 | ||
1089 | void | |
1090 | zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) | |
1091 | { | |
9b67f605 | 1092 | |
a1d477c2 MA |
1093 | zfs_blkptr_verify(spa, bp); |
1094 | ||
9b67f605 MA |
1095 | /* |
1096 | * The check for EMBEDDED is a performance optimization. We | |
1097 | * process the free here (by ignoring it) rather than | |
1098 | * putting it on the list and then processing it in zio_free_sync(). | |
1099 | */ | |
1100 | if (BP_IS_EMBEDDED(bp)) | |
1101 | return; | |
13fe0198 | 1102 | metaslab_check_free(spa, bp); |
2883cad5 MA |
1103 | |
1104 | /* | |
1105 | * Frees that are for the currently-syncing txg, are not going to be | |
1106 | * deferred, and which will not need to do a read (i.e. not GANG or | |
1107 | * DEDUP), can be processed immediately. Otherwise, put them on the | |
1108 | * in-memory list for later processing. | |
1109 | */ | |
1110 | if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || | |
1111 | txg != spa->spa_syncing_txg || | |
1112 | spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { | |
1113 | bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); | |
1114 | } else { | |
1115 | VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0))); | |
1116 | } | |
428870ff BB |
1117 | } |
1118 | ||
34dc7c2f | 1119 | zio_t * |
428870ff BB |
1120 | zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
1121 | enum zio_flag flags) | |
34dc7c2f BB |
1122 | { |
1123 | zio_t *zio; | |
2883cad5 | 1124 | enum zio_stage stage = ZIO_FREE_PIPELINE; |
34dc7c2f | 1125 | |
428870ff BB |
1126 | ASSERT(!BP_IS_HOLE(bp)); |
1127 | ASSERT(spa_syncing_txg(spa) == txg); | |
55d85d5a | 1128 | ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); |
34dc7c2f | 1129 | |
9b67f605 MA |
1130 | if (BP_IS_EMBEDDED(bp)) |
1131 | return (zio_null(pio, spa, NULL, NULL, NULL, 0)); | |
1132 | ||
13fe0198 | 1133 | metaslab_check_free(spa, bp); |
8c841793 | 1134 | arc_freed(spa, bp); |
d4a72f23 | 1135 | dsl_scan_freed(spa, bp); |
13fe0198 | 1136 | |
2883cad5 MA |
1137 | /* |
1138 | * GANG and DEDUP blocks can induce a read (for the gang block header, | |
1139 | * or the DDT), so issue them asynchronously so that this thread is | |
1140 | * not tied up. | |
1141 | */ | |
1142 | if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) | |
1143 | stage |= ZIO_STAGE_ISSUE_ASYNC; | |
1144 | ||
b128c09f | 1145 | zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
2aa34383 DK |
1146 | BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, |
1147 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage); | |
2883cad5 | 1148 | |
34dc7c2f BB |
1149 | return (zio); |
1150 | } | |
1151 | ||
1152 | zio_t * | |
428870ff BB |
1153 | zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
1154 | zio_done_func_t *done, void *private, enum zio_flag flags) | |
34dc7c2f BB |
1155 | { |
1156 | zio_t *zio; | |
1157 | ||
a1d477c2 | 1158 | zfs_blkptr_verify(spa, bp); |
9b67f605 MA |
1159 | |
1160 | if (BP_IS_EMBEDDED(bp)) | |
1161 | return (zio_null(pio, spa, NULL, NULL, NULL, 0)); | |
1162 | ||
34dc7c2f BB |
1163 | /* |
1164 | * A claim is an allocation of a specific block. Claims are needed | |
1165 | * to support immediate writes in the intent log. The issue is that | |
1166 | * immediate writes contain committed data, but in a txg that was | |
1167 | * *not* committed. Upon opening the pool after an unclean shutdown, | |
1168 | * the intent log claims all blocks that contain immediate write data | |
1169 | * so that the SPA knows they're in use. | |
1170 | * | |
1171 | * All claims *must* be resolved in the first txg -- before the SPA | |
1172 | * starts allocating blocks -- so that nothing is allocated twice. | |
428870ff | 1173 | * If txg == 0 we just verify that the block is claimable. |
34dc7c2f | 1174 | */ |
d2734cce SD |
1175 | ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, |
1176 | spa_min_claim_txg(spa)); | |
1177 | ASSERT(txg == spa_min_claim_txg(spa) || txg == 0); | |
428870ff | 1178 | ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ |
34dc7c2f | 1179 | |
b128c09f | 1180 | zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
2aa34383 DK |
1181 | BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, |
1182 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); | |
3dfb57a3 | 1183 | ASSERT0(zio->io_queued_timestamp); |
34dc7c2f BB |
1184 | |
1185 | return (zio); | |
1186 | } | |
1187 | ||
1188 | zio_t * | |
1189 | zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, | |
e8b96c60 | 1190 | zio_done_func_t *done, void *private, enum zio_flag flags) |
34dc7c2f BB |
1191 | { |
1192 | zio_t *zio; | |
1193 | int c; | |
1194 | ||
1195 | if (vd->vdev_children == 0) { | |
2aa34383 | 1196 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
e8b96c60 | 1197 | ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
34dc7c2f BB |
1198 | ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); |
1199 | ||
34dc7c2f BB |
1200 | zio->io_cmd = cmd; |
1201 | } else { | |
d164b209 | 1202 | zio = zio_null(pio, spa, NULL, NULL, NULL, flags); |
34dc7c2f BB |
1203 | |
1204 | for (c = 0; c < vd->vdev_children; c++) | |
1205 | zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, | |
e8b96c60 | 1206 | done, private, flags)); |
34dc7c2f BB |
1207 | } |
1208 | ||
1209 | return (zio); | |
1210 | } | |
1211 | ||
34dc7c2f BB |
1212 | zio_t * |
1213 | zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
a6255b7f | 1214 | abd_t *data, int checksum, zio_done_func_t *done, void *private, |
e8b96c60 | 1215 | zio_priority_t priority, enum zio_flag flags, boolean_t labels) |
34dc7c2f BB |
1216 | { |
1217 | zio_t *zio; | |
34dc7c2f | 1218 | |
b128c09f BB |
1219 | ASSERT(vd->vdev_children == 0); |
1220 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
1221 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
1222 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 1223 | |
2aa34383 DK |
1224 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
1225 | private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
1226 | offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); | |
34dc7c2f | 1227 | |
b128c09f | 1228 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f BB |
1229 | |
1230 | return (zio); | |
1231 | } | |
1232 | ||
1233 | zio_t * | |
1234 | zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
a6255b7f | 1235 | abd_t *data, int checksum, zio_done_func_t *done, void *private, |
e8b96c60 | 1236 | zio_priority_t priority, enum zio_flag flags, boolean_t labels) |
34dc7c2f | 1237 | { |
34dc7c2f | 1238 | zio_t *zio; |
34dc7c2f | 1239 | |
b128c09f BB |
1240 | ASSERT(vd->vdev_children == 0); |
1241 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
1242 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
1243 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 1244 | |
2aa34383 DK |
1245 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
1246 | private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
1247 | offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); | |
34dc7c2f | 1248 | |
b128c09f | 1249 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f | 1250 | |
3c67d83a | 1251 | if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { |
34dc7c2f | 1252 | /* |
428870ff | 1253 | * zec checksums are necessarily destructive -- they modify |
b128c09f | 1254 | * the end of the write buffer to hold the verifier/checksum. |
34dc7c2f | 1255 | * Therefore, we must make a local copy in case the data is |
b128c09f | 1256 | * being written to multiple places in parallel. |
34dc7c2f | 1257 | */ |
a6255b7f DQ |
1258 | abd_t *wbuf = abd_alloc_sametype(data, size); |
1259 | abd_copy(wbuf, data, size); | |
1260 | ||
b128c09f | 1261 | zio_push_transform(zio, wbuf, size, size, NULL); |
34dc7c2f BB |
1262 | } |
1263 | ||
1264 | return (zio); | |
1265 | } | |
1266 | ||
1267 | /* | |
b128c09f | 1268 | * Create a child I/O to do some work for us. |
34dc7c2f BB |
1269 | */ |
1270 | zio_t * | |
b128c09f | 1271 | zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, |
4ea3f864 GM |
1272 | abd_t *data, uint64_t size, int type, zio_priority_t priority, |
1273 | enum zio_flag flags, zio_done_func_t *done, void *private) | |
34dc7c2f | 1274 | { |
428870ff | 1275 | enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; |
b128c09f BB |
1276 | zio_t *zio; |
1277 | ||
a1d477c2 MA |
1278 | /* |
1279 | * vdev child I/Os do not propagate their error to the parent. | |
1280 | * Therefore, for correct operation the caller *must* check for | |
1281 | * and handle the error in the child i/o's done callback. | |
1282 | * The only exceptions are i/os that we don't care about | |
1283 | * (OPTIONAL or REPAIR). | |
1284 | */ | |
1285 | ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) || | |
1286 | done != NULL); | |
1287 | ||
34dc7c2f BB |
1288 | if (type == ZIO_TYPE_READ && bp != NULL) { |
1289 | /* | |
1290 | * If we have the bp, then the child should perform the | |
1291 | * checksum and the parent need not. This pushes error | |
1292 | * detection as close to the leaves as possible and | |
1293 | * eliminates redundant checksums in the interior nodes. | |
1294 | */ | |
428870ff BB |
1295 | pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; |
1296 | pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; | |
34dc7c2f BB |
1297 | } |
1298 | ||
a1d477c2 MA |
1299 | if (vd->vdev_ops->vdev_op_leaf) { |
1300 | ASSERT0(vd->vdev_children); | |
b128c09f | 1301 | offset += VDEV_LABEL_START_SIZE; |
a1d477c2 | 1302 | } |
b128c09f | 1303 | |
a1d477c2 | 1304 | flags |= ZIO_VDEV_CHILD_FLAGS(pio); |
428870ff BB |
1305 | |
1306 | /* | |
1307 | * If we've decided to do a repair, the write is not speculative -- | |
1308 | * even if the original read was. | |
1309 | */ | |
1310 | if (flags & ZIO_FLAG_IO_REPAIR) | |
1311 | flags &= ~ZIO_FLAG_SPECULATIVE; | |
1312 | ||
3dfb57a3 DB |
1313 | /* |
1314 | * If we're creating a child I/O that is not associated with a | |
1315 | * top-level vdev, then the child zio is not an allocating I/O. | |
1316 | * If this is a retried I/O then we ignore it since we will | |
1317 | * have already processed the original allocating I/O. | |
1318 | */ | |
1319 | if (flags & ZIO_FLAG_IO_ALLOCATING && | |
1320 | (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) { | |
cc99f275 DB |
1321 | ASSERT(pio->io_metaslab_class != NULL); |
1322 | ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled); | |
3dfb57a3 DB |
1323 | ASSERT(type == ZIO_TYPE_WRITE); |
1324 | ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE); | |
1325 | ASSERT(!(flags & ZIO_FLAG_IO_REPAIR)); | |
1326 | ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) || | |
1327 | pio->io_child_type == ZIO_CHILD_GANG); | |
1328 | ||
1329 | flags &= ~ZIO_FLAG_IO_ALLOCATING; | |
1330 | } | |
1331 | ||
1332 | ||
2aa34383 | 1333 | zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size, |
428870ff BB |
1334 | done, private, type, priority, flags, vd, offset, &pio->io_bookmark, |
1335 | ZIO_STAGE_VDEV_IO_START >> 1, pipeline); | |
3dfb57a3 | 1336 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); |
34dc7c2f | 1337 | |
e8b96c60 MA |
1338 | zio->io_physdone = pio->io_physdone; |
1339 | if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) | |
1340 | zio->io_logical->io_phys_children++; | |
1341 | ||
b128c09f | 1342 | return (zio); |
34dc7c2f BB |
1343 | } |
1344 | ||
b128c09f | 1345 | zio_t * |
a6255b7f | 1346 | zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size, |
9e052db4 | 1347 | zio_type_t type, zio_priority_t priority, enum zio_flag flags, |
e9aa730c | 1348 | zio_done_func_t *done, void *private) |
34dc7c2f | 1349 | { |
b128c09f | 1350 | zio_t *zio; |
34dc7c2f | 1351 | |
b128c09f | 1352 | ASSERT(vd->vdev_ops->vdev_op_leaf); |
34dc7c2f | 1353 | |
b128c09f | 1354 | zio = zio_create(NULL, vd->vdev_spa, 0, NULL, |
2aa34383 | 1355 | data, size, size, done, private, type, priority, |
e8b96c60 | 1356 | flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, |
b128c09f | 1357 | vd, offset, NULL, |
428870ff | 1358 | ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); |
34dc7c2f | 1359 | |
b128c09f | 1360 | return (zio); |
34dc7c2f BB |
1361 | } |
1362 | ||
1363 | void | |
b128c09f | 1364 | zio_flush(zio_t *zio, vdev_t *vd) |
34dc7c2f | 1365 | { |
b128c09f | 1366 | zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, |
e8b96c60 | 1367 | NULL, NULL, |
b128c09f | 1368 | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); |
34dc7c2f BB |
1369 | } |
1370 | ||
428870ff BB |
1371 | void |
1372 | zio_shrink(zio_t *zio, uint64_t size) | |
1373 | { | |
1ce23dca PS |
1374 | ASSERT3P(zio->io_executor, ==, NULL); |
1375 | ASSERT3U(zio->io_orig_size, ==, zio->io_size); | |
1376 | ASSERT3U(size, <=, zio->io_size); | |
428870ff BB |
1377 | |
1378 | /* | |
1379 | * We don't shrink for raidz because of problems with the | |
1380 | * reconstruction when reading back less than the block size. | |
1381 | * Note, BP_IS_RAIDZ() assumes no compression. | |
1382 | */ | |
1383 | ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); | |
2aa34383 DK |
1384 | if (!BP_IS_RAIDZ(zio->io_bp)) { |
1385 | /* we are not doing a raw write */ | |
1386 | ASSERT3U(zio->io_size, ==, zio->io_lsize); | |
1387 | zio->io_orig_size = zio->io_size = zio->io_lsize = size; | |
1388 | } | |
428870ff BB |
1389 | } |
1390 | ||
34dc7c2f BB |
1391 | /* |
1392 | * ========================================================================== | |
b128c09f | 1393 | * Prepare to read and write logical blocks |
34dc7c2f BB |
1394 | * ========================================================================== |
1395 | */ | |
b128c09f | 1396 | |
62840030 | 1397 | static zio_t * |
b128c09f | 1398 | zio_read_bp_init(zio_t *zio) |
34dc7c2f | 1399 | { |
b128c09f | 1400 | blkptr_t *bp = zio->io_bp; |
b5256303 TC |
1401 | uint64_t psize = |
1402 | BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); | |
34dc7c2f | 1403 | |
a1d477c2 MA |
1404 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
1405 | ||
fb5f0bc8 | 1406 | if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && |
9babb374 | 1407 | zio->io_child_type == ZIO_CHILD_LOGICAL && |
b5256303 | 1408 | !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { |
a6255b7f DQ |
1409 | zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), |
1410 | psize, psize, zio_decompress); | |
34dc7c2f | 1411 | } |
34dc7c2f | 1412 | |
b5256303 TC |
1413 | if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) || |
1414 | BP_HAS_INDIRECT_MAC_CKSUM(bp)) && | |
1415 | zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
1416 | zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), | |
1417 | psize, psize, zio_decrypt); | |
1418 | } | |
1419 | ||
9b67f605 | 1420 | if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { |
a6255b7f DQ |
1421 | int psize = BPE_GET_PSIZE(bp); |
1422 | void *data = abd_borrow_buf(zio->io_abd, psize); | |
1423 | ||
9b67f605 | 1424 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
a6255b7f DQ |
1425 | decode_embedded_bp_compressed(bp, data); |
1426 | abd_return_buf_copy(zio->io_abd, data, psize); | |
9b67f605 MA |
1427 | } else { |
1428 | ASSERT(!BP_IS_EMBEDDED(bp)); | |
a1d477c2 | 1429 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
9b67f605 MA |
1430 | } |
1431 | ||
9ae529ec | 1432 | if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) |
b128c09f BB |
1433 | zio->io_flags |= ZIO_FLAG_DONT_CACHE; |
1434 | ||
428870ff BB |
1435 | if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) |
1436 | zio->io_flags |= ZIO_FLAG_DONT_CACHE; | |
1437 | ||
1438 | if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) | |
1439 | zio->io_pipeline = ZIO_DDT_READ_PIPELINE; | |
1440 | ||
62840030 | 1441 | return (zio); |
34dc7c2f BB |
1442 | } |
1443 | ||
62840030 | 1444 | static zio_t * |
b128c09f | 1445 | zio_write_bp_init(zio_t *zio) |
34dc7c2f | 1446 | { |
b128c09f | 1447 | if (!IO_IS_ALLOCATING(zio)) |
62840030 | 1448 | return (zio); |
34dc7c2f | 1449 | |
428870ff BB |
1450 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1451 | ||
1452 | if (zio->io_bp_override) { | |
3dfb57a3 DB |
1453 | blkptr_t *bp = zio->io_bp; |
1454 | zio_prop_t *zp = &zio->io_prop; | |
1455 | ||
428870ff BB |
1456 | ASSERT(bp->blk_birth != zio->io_txg); |
1457 | ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); | |
1458 | ||
1459 | *bp = *zio->io_bp_override; | |
1460 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1461 | ||
9b67f605 | 1462 | if (BP_IS_EMBEDDED(bp)) |
62840030 | 1463 | return (zio); |
9b67f605 | 1464 | |
03c6040b GW |
1465 | /* |
1466 | * If we've been overridden and nopwrite is set then | |
1467 | * set the flag accordingly to indicate that a nopwrite | |
1468 | * has already occurred. | |
1469 | */ | |
1470 | if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { | |
1471 | ASSERT(!zp->zp_dedup); | |
3dfb57a3 | 1472 | ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum); |
03c6040b | 1473 | zio->io_flags |= ZIO_FLAG_NOPWRITE; |
62840030 | 1474 | return (zio); |
03c6040b GW |
1475 | } |
1476 | ||
1477 | ASSERT(!zp->zp_nopwrite); | |
1478 | ||
428870ff | 1479 | if (BP_IS_HOLE(bp) || !zp->zp_dedup) |
62840030 | 1480 | return (zio); |
428870ff | 1481 | |
3c67d83a TH |
1482 | ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & |
1483 | ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); | |
428870ff | 1484 | |
b5256303 TC |
1485 | if (BP_GET_CHECKSUM(bp) == zp->zp_checksum && |
1486 | !zp->zp_encrypt) { | |
428870ff BB |
1487 | BP_SET_DEDUP(bp, 1); |
1488 | zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; | |
62840030 | 1489 | return (zio); |
428870ff | 1490 | } |
3dfb57a3 DB |
1491 | |
1492 | /* | |
1493 | * We were unable to handle this as an override bp, treat | |
1494 | * it as a regular write I/O. | |
1495 | */ | |
5511754b | 1496 | zio->io_bp_override = NULL; |
3dfb57a3 DB |
1497 | *bp = zio->io_bp_orig; |
1498 | zio->io_pipeline = zio->io_orig_pipeline; | |
1499 | } | |
1500 | ||
62840030 | 1501 | return (zio); |
3dfb57a3 DB |
1502 | } |
1503 | ||
62840030 | 1504 | static zio_t * |
3dfb57a3 DB |
1505 | zio_write_compress(zio_t *zio) |
1506 | { | |
1507 | spa_t *spa = zio->io_spa; | |
1508 | zio_prop_t *zp = &zio->io_prop; | |
1509 | enum zio_compress compress = zp->zp_compress; | |
1510 | blkptr_t *bp = zio->io_bp; | |
1511 | uint64_t lsize = zio->io_lsize; | |
1512 | uint64_t psize = zio->io_size; | |
1513 | int pass = 1; | |
1514 | ||
3dfb57a3 DB |
1515 | /* |
1516 | * If our children haven't all reached the ready stage, | |
1517 | * wait for them and then repeat this pipeline stage. | |
1518 | */ | |
ddc751d5 GW |
1519 | if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT | |
1520 | ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) { | |
62840030 | 1521 | return (NULL); |
ddc751d5 | 1522 | } |
3dfb57a3 DB |
1523 | |
1524 | if (!IO_IS_ALLOCATING(zio)) | |
62840030 | 1525 | return (zio); |
3dfb57a3 DB |
1526 | |
1527 | if (zio->io_children_ready != NULL) { | |
1528 | /* | |
1529 | * Now that all our children are ready, run the callback | |
1530 | * associated with this zio in case it wants to modify the | |
1531 | * data to be written. | |
1532 | */ | |
1533 | ASSERT3U(zp->zp_level, >, 0); | |
1534 | zio->io_children_ready(zio); | |
428870ff | 1535 | } |
34dc7c2f | 1536 | |
3dfb57a3 DB |
1537 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1538 | ASSERT(zio->io_bp_override == NULL); | |
1539 | ||
b0bc7a84 | 1540 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { |
b128c09f BB |
1541 | /* |
1542 | * We're rewriting an existing block, which means we're | |
1543 | * working on behalf of spa_sync(). For spa_sync() to | |
1544 | * converge, it must eventually be the case that we don't | |
1545 | * have to allocate new blocks. But compression changes | |
1546 | * the blocksize, which forces a reallocate, and makes | |
1547 | * convergence take longer. Therefore, after the first | |
1548 | * few passes, stop compressing to ensure convergence. | |
1549 | */ | |
428870ff BB |
1550 | pass = spa_sync_pass(spa); |
1551 | ||
1552 | ASSERT(zio->io_txg == spa_syncing_txg(spa)); | |
1553 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1554 | ASSERT(!BP_GET_DEDUP(bp)); | |
34dc7c2f | 1555 | |
55d85d5a | 1556 | if (pass >= zfs_sync_pass_dont_compress) |
b128c09f | 1557 | compress = ZIO_COMPRESS_OFF; |
34dc7c2f | 1558 | |
b128c09f | 1559 | /* Make sure someone doesn't change their mind on overwrites */ |
9b67f605 | 1560 | ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), |
428870ff | 1561 | spa_max_replication(spa)) == BP_GET_NDVAS(bp)); |
b128c09f | 1562 | } |
34dc7c2f | 1563 | |
2aa34383 | 1564 | /* If it's a compressed write that is not raw, compress the buffer. */ |
b5256303 TC |
1565 | if (compress != ZIO_COMPRESS_OFF && |
1566 | !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { | |
428870ff | 1567 | void *cbuf = zio_buf_alloc(lsize); |
a6255b7f | 1568 | psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize); |
428870ff | 1569 | if (psize == 0 || psize == lsize) { |
b128c09f | 1570 | compress = ZIO_COMPRESS_OFF; |
428870ff | 1571 | zio_buf_free(cbuf, lsize); |
b5256303 TC |
1572 | } else if (!zp->zp_dedup && !zp->zp_encrypt && |
1573 | psize <= BPE_PAYLOAD_SIZE && | |
9b67f605 MA |
1574 | zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && |
1575 | spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { | |
1576 | encode_embedded_bp_compressed(bp, | |
1577 | cbuf, compress, lsize, psize); | |
1578 | BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); | |
1579 | BP_SET_TYPE(bp, zio->io_prop.zp_type); | |
1580 | BP_SET_LEVEL(bp, zio->io_prop.zp_level); | |
1581 | zio_buf_free(cbuf, lsize); | |
1582 | bp->blk_birth = zio->io_txg; | |
1583 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1584 | ASSERT(spa_feature_is_active(spa, | |
1585 | SPA_FEATURE_EMBEDDED_DATA)); | |
62840030 | 1586 | return (zio); |
428870ff | 1587 | } else { |
9b67f605 | 1588 | /* |
c3520e7f MA |
1589 | * Round up compressed size up to the ashift |
1590 | * of the smallest-ashift device, and zero the tail. | |
1591 | * This ensures that the compressed size of the BP | |
1592 | * (and thus compressratio property) are correct, | |
1593 | * in that we charge for the padding used to fill out | |
1594 | * the last sector. | |
9b67f605 | 1595 | */ |
c3520e7f | 1596 | ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT); |
1c27024e | 1597 | size_t rounded = (size_t)P2ROUNDUP(psize, |
c3520e7f MA |
1598 | 1ULL << spa->spa_min_ashift); |
1599 | if (rounded >= lsize) { | |
9b67f605 MA |
1600 | compress = ZIO_COMPRESS_OFF; |
1601 | zio_buf_free(cbuf, lsize); | |
c3520e7f | 1602 | psize = lsize; |
9b67f605 | 1603 | } else { |
a6255b7f DQ |
1604 | abd_t *cdata = abd_get_from_buf(cbuf, lsize); |
1605 | abd_take_ownership_of_buf(cdata, B_TRUE); | |
1606 | abd_zero_off(cdata, psize, rounded - psize); | |
c3520e7f | 1607 | psize = rounded; |
a6255b7f | 1608 | zio_push_transform(zio, cdata, |
9b67f605 MA |
1609 | psize, lsize, NULL); |
1610 | } | |
b128c09f | 1611 | } |
3dfb57a3 DB |
1612 | |
1613 | /* | |
1614 | * We were unable to handle this as an override bp, treat | |
1615 | * it as a regular write I/O. | |
1616 | */ | |
1617 | zio->io_bp_override = NULL; | |
1618 | *bp = zio->io_bp_orig; | |
1619 | zio->io_pipeline = zio->io_orig_pipeline; | |
1620 | ||
b1d21733 TC |
1621 | } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 && |
1622 | zp->zp_type == DMU_OT_DNODE) { | |
1623 | /* | |
1624 | * The DMU actually relies on the zio layer's compression | |
1625 | * to free metadnode blocks that have had all contained | |
1626 | * dnodes freed. As a result, even when doing a raw | |
1627 | * receive, we must check whether the block can be compressed | |
1628 | * to a hole. | |
1629 | */ | |
1630 | psize = zio_compress_data(ZIO_COMPRESS_EMPTY, | |
1631 | zio->io_abd, NULL, lsize); | |
1632 | if (psize == 0) | |
1633 | compress = ZIO_COMPRESS_OFF; | |
2aa34383 DK |
1634 | } else { |
1635 | ASSERT3U(psize, !=, 0); | |
b128c09f | 1636 | } |
34dc7c2f | 1637 | |
b128c09f BB |
1638 | /* |
1639 | * The final pass of spa_sync() must be all rewrites, but the first | |
1640 | * few passes offer a trade-off: allocating blocks defers convergence, | |
1641 | * but newly allocated blocks are sequential, so they can be written | |
1642 | * to disk faster. Therefore, we allow the first few passes of | |
1643 | * spa_sync() to allocate new blocks, but force rewrites after that. | |
1644 | * There should only be a handful of blocks after pass 1 in any case. | |
1645 | */ | |
b0bc7a84 MG |
1646 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && |
1647 | BP_GET_PSIZE(bp) == psize && | |
55d85d5a | 1648 | pass >= zfs_sync_pass_rewrite) { |
cc99f275 | 1649 | VERIFY3U(psize, !=, 0); |
1c27024e | 1650 | enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; |
cc99f275 | 1651 | |
b128c09f BB |
1652 | zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; |
1653 | zio->io_flags |= ZIO_FLAG_IO_REWRITE; | |
1654 | } else { | |
1655 | BP_ZERO(bp); | |
1656 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
1657 | } | |
34dc7c2f | 1658 | |
428870ff | 1659 | if (psize == 0) { |
b0bc7a84 MG |
1660 | if (zio->io_bp_orig.blk_birth != 0 && |
1661 | spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { | |
1662 | BP_SET_LSIZE(bp, lsize); | |
1663 | BP_SET_TYPE(bp, zp->zp_type); | |
1664 | BP_SET_LEVEL(bp, zp->zp_level); | |
1665 | BP_SET_BIRTH(bp, zio->io_txg, 0); | |
1666 | } | |
b128c09f BB |
1667 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
1668 | } else { | |
1669 | ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); | |
1670 | BP_SET_LSIZE(bp, lsize); | |
b0bc7a84 MG |
1671 | BP_SET_TYPE(bp, zp->zp_type); |
1672 | BP_SET_LEVEL(bp, zp->zp_level); | |
428870ff | 1673 | BP_SET_PSIZE(bp, psize); |
b128c09f BB |
1674 | BP_SET_COMPRESS(bp, compress); |
1675 | BP_SET_CHECKSUM(bp, zp->zp_checksum); | |
428870ff | 1676 | BP_SET_DEDUP(bp, zp->zp_dedup); |
b128c09f | 1677 | BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); |
428870ff BB |
1678 | if (zp->zp_dedup) { |
1679 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1680 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
b5256303 TC |
1681 | ASSERT(!zp->zp_encrypt || |
1682 | DMU_OT_IS_ENCRYPTED(zp->zp_type)); | |
428870ff BB |
1683 | zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; |
1684 | } | |
03c6040b GW |
1685 | if (zp->zp_nopwrite) { |
1686 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1687 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
1688 | zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; | |
1689 | } | |
428870ff | 1690 | } |
62840030 | 1691 | return (zio); |
428870ff BB |
1692 | } |
1693 | ||
62840030 | 1694 | static zio_t * |
428870ff BB |
1695 | zio_free_bp_init(zio_t *zio) |
1696 | { | |
1697 | blkptr_t *bp = zio->io_bp; | |
1698 | ||
1699 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
1700 | if (BP_GET_DEDUP(bp)) | |
1701 | zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; | |
b128c09f | 1702 | } |
34dc7c2f | 1703 | |
a1d477c2 MA |
1704 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
1705 | ||
62840030 | 1706 | return (zio); |
34dc7c2f BB |
1707 | } |
1708 | ||
b128c09f BB |
1709 | /* |
1710 | * ========================================================================== | |
1711 | * Execute the I/O pipeline | |
1712 | * ========================================================================== | |
1713 | */ | |
1714 | ||
1715 | static void | |
7ef5e54e | 1716 | zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) |
34dc7c2f | 1717 | { |
428870ff | 1718 | spa_t *spa = zio->io_spa; |
b128c09f | 1719 | zio_type_t t = zio->io_type; |
a38718a6 | 1720 | int flags = (cutinline ? TQ_FRONT : 0); |
34dc7c2f BB |
1721 | |
1722 | /* | |
9babb374 BB |
1723 | * If we're a config writer or a probe, the normal issue and |
1724 | * interrupt threads may all be blocked waiting for the config lock. | |
1725 | * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. | |
34dc7c2f | 1726 | */ |
9babb374 | 1727 | if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) |
b128c09f | 1728 | t = ZIO_TYPE_NULL; |
34dc7c2f BB |
1729 | |
1730 | /* | |
b128c09f | 1731 | * A similar issue exists for the L2ARC write thread until L2ARC 2.0. |
34dc7c2f | 1732 | */ |
b128c09f BB |
1733 | if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) |
1734 | t = ZIO_TYPE_NULL; | |
34dc7c2f | 1735 | |
428870ff | 1736 | /* |
7ef5e54e AL |
1737 | * If this is a high priority I/O, then use the high priority taskq if |
1738 | * available. | |
428870ff BB |
1739 | */ |
1740 | if (zio->io_priority == ZIO_PRIORITY_NOW && | |
7ef5e54e | 1741 | spa->spa_zio_taskq[t][q + 1].stqs_count != 0) |
428870ff BB |
1742 | q++; |
1743 | ||
1744 | ASSERT3U(q, <, ZIO_TASKQ_TYPES); | |
5cc556b4 | 1745 | |
a38718a6 GA |
1746 | /* |
1747 | * NB: We are assuming that the zio can only be dispatched | |
1748 | * to a single taskq at a time. It would be a grievous error | |
1749 | * to dispatch the zio to another taskq at the same time. | |
1750 | */ | |
1751 | ASSERT(taskq_empty_ent(&zio->io_tqent)); | |
7ef5e54e AL |
1752 | spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, |
1753 | flags, &zio->io_tqent); | |
b128c09f | 1754 | } |
34dc7c2f | 1755 | |
b128c09f | 1756 | static boolean_t |
7ef5e54e | 1757 | zio_taskq_member(zio_t *zio, zio_taskq_type_t q) |
b128c09f BB |
1758 | { |
1759 | kthread_t *executor = zio->io_executor; | |
1760 | spa_t *spa = zio->io_spa; | |
34dc7c2f | 1761 | |
1c27024e | 1762 | for (zio_type_t t = 0; t < ZIO_TYPES; t++) { |
7ef5e54e AL |
1763 | spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; |
1764 | uint_t i; | |
1765 | for (i = 0; i < tqs->stqs_count; i++) { | |
1766 | if (taskq_member(tqs->stqs_taskq[i], executor)) | |
1767 | return (B_TRUE); | |
1768 | } | |
1769 | } | |
34dc7c2f | 1770 | |
b128c09f BB |
1771 | return (B_FALSE); |
1772 | } | |
34dc7c2f | 1773 | |
62840030 | 1774 | static zio_t * |
b128c09f BB |
1775 | zio_issue_async(zio_t *zio) |
1776 | { | |
428870ff | 1777 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); |
b128c09f | 1778 | |
62840030 | 1779 | return (NULL); |
34dc7c2f BB |
1780 | } |
1781 | ||
b128c09f BB |
1782 | void |
1783 | zio_interrupt(zio_t *zio) | |
34dc7c2f | 1784 | { |
428870ff | 1785 | zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); |
b128c09f | 1786 | } |
34dc7c2f | 1787 | |
26ef0cc7 TH |
1788 | void |
1789 | zio_delay_interrupt(zio_t *zio) | |
1790 | { | |
1791 | /* | |
1792 | * The timeout_generic() function isn't defined in userspace, so | |
1793 | * rather than trying to implement the function, the zio delay | |
1794 | * functionality has been disabled for userspace builds. | |
1795 | */ | |
1796 | ||
1797 | #ifdef _KERNEL | |
1798 | /* | |
1799 | * If io_target_timestamp is zero, then no delay has been registered | |
1800 | * for this IO, thus jump to the end of this function and "skip" the | |
1801 | * delay; issuing it directly to the zio layer. | |
1802 | */ | |
1803 | if (zio->io_target_timestamp != 0) { | |
1804 | hrtime_t now = gethrtime(); | |
1805 | ||
1806 | if (now >= zio->io_target_timestamp) { | |
1807 | /* | |
1808 | * This IO has already taken longer than the target | |
1809 | * delay to complete, so we don't want to delay it | |
1810 | * any longer; we "miss" the delay and issue it | |
1811 | * directly to the zio layer. This is likely due to | |
1812 | * the target latency being set to a value less than | |
1813 | * the underlying hardware can satisfy (e.g. delay | |
1814 | * set to 1ms, but the disks take 10ms to complete an | |
1815 | * IO request). | |
1816 | */ | |
1817 | ||
1818 | DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, | |
1819 | hrtime_t, now); | |
1820 | ||
1821 | zio_interrupt(zio); | |
1822 | } else { | |
1823 | taskqid_t tid; | |
1824 | hrtime_t diff = zio->io_target_timestamp - now; | |
1825 | clock_t expire_at_tick = ddi_get_lbolt() + | |
1826 | NSEC_TO_TICK(diff); | |
1827 | ||
1828 | DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, | |
1829 | hrtime_t, now, hrtime_t, diff); | |
1830 | ||
1831 | if (NSEC_TO_TICK(diff) == 0) { | |
1832 | /* Our delay is less than a jiffy - just spin */ | |
1833 | zfs_sleep_until(zio->io_target_timestamp); | |
1834 | } else { | |
1835 | /* | |
1836 | * Use taskq_dispatch_delay() in the place of | |
1837 | * OpenZFS's timeout_generic(). | |
1838 | */ | |
1839 | tid = taskq_dispatch_delay(system_taskq, | |
02730c33 | 1840 | (task_func_t *)zio_interrupt, |
26ef0cc7 | 1841 | zio, TQ_NOSLEEP, expire_at_tick); |
48d3eb40 | 1842 | if (tid == TASKQID_INVALID) { |
26ef0cc7 TH |
1843 | /* |
1844 | * Couldn't allocate a task. Just | |
1845 | * finish the zio without a delay. | |
1846 | */ | |
1847 | zio_interrupt(zio); | |
1848 | } | |
1849 | } | |
1850 | } | |
1851 | return; | |
1852 | } | |
1853 | #endif | |
1854 | DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); | |
1855 | zio_interrupt(zio); | |
1856 | } | |
1857 | ||
8fb1ede1 BB |
1858 | static void |
1859 | zio_deadman_impl(zio_t *pio) | |
1860 | { | |
1861 | zio_t *cio, *cio_next; | |
1862 | zio_link_t *zl = NULL; | |
1863 | vdev_t *vd = pio->io_vd; | |
1864 | ||
1865 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { | |
1866 | vdev_queue_t *vq = &vd->vdev_queue; | |
1867 | zbookmark_phys_t *zb = &pio->io_bookmark; | |
1868 | uint64_t delta = gethrtime() - pio->io_timestamp; | |
1869 | uint64_t failmode = spa_get_deadman_failmode(pio->io_spa); | |
1870 | ||
1871 | zfs_dbgmsg("slow zio: zio=%p timestamp=%llu " | |
1872 | "delta=%llu queued=%llu io=%llu " | |
1873 | "path=%s last=%llu " | |
1874 | "type=%d priority=%d flags=0x%x " | |
1875 | "stage=0x%x pipeline=0x%x pipeline-trace=0x%x " | |
1876 | "objset=%llu object=%llu level=%llu blkid=%llu " | |
1877 | "offset=%llu size=%llu error=%d", | |
1878 | pio, pio->io_timestamp, | |
1879 | delta, pio->io_delta, pio->io_delay, | |
1880 | vd->vdev_path, vq->vq_io_complete_ts, | |
1881 | pio->io_type, pio->io_priority, pio->io_flags, | |
1882 | pio->io_state, pio->io_pipeline, pio->io_pipeline_trace, | |
1883 | zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid, | |
1884 | pio->io_offset, pio->io_size, pio->io_error); | |
1885 | zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN, | |
1886 | pio->io_spa, vd, zb, pio, 0, 0); | |
1887 | ||
1888 | if (failmode == ZIO_FAILURE_MODE_CONTINUE && | |
1889 | taskq_empty_ent(&pio->io_tqent)) { | |
1890 | zio_interrupt(pio); | |
1891 | } | |
1892 | } | |
1893 | ||
1894 | mutex_enter(&pio->io_lock); | |
1895 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { | |
1896 | cio_next = zio_walk_children(pio, &zl); | |
1897 | zio_deadman_impl(cio); | |
1898 | } | |
1899 | mutex_exit(&pio->io_lock); | |
1900 | } | |
1901 | ||
1902 | /* | |
1903 | * Log the critical information describing this zio and all of its children | |
1904 | * using the zfs_dbgmsg() interface then post deadman event for the ZED. | |
1905 | */ | |
1906 | void | |
1907 | zio_deadman(zio_t *pio, char *tag) | |
1908 | { | |
1909 | spa_t *spa = pio->io_spa; | |
1910 | char *name = spa_name(spa); | |
1911 | ||
1912 | if (!zfs_deadman_enabled || spa_suspended(spa)) | |
1913 | return; | |
1914 | ||
1915 | zio_deadman_impl(pio); | |
1916 | ||
1917 | switch (spa_get_deadman_failmode(spa)) { | |
1918 | case ZIO_FAILURE_MODE_WAIT: | |
1919 | zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name); | |
1920 | break; | |
1921 | ||
1922 | case ZIO_FAILURE_MODE_CONTINUE: | |
1923 | zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name); | |
1924 | break; | |
1925 | ||
1926 | case ZIO_FAILURE_MODE_PANIC: | |
1927 | fm_panic("%s determined I/O to pool '%s' is hung.", tag, name); | |
1928 | break; | |
1929 | } | |
1930 | } | |
1931 | ||
b128c09f BB |
1932 | /* |
1933 | * Execute the I/O pipeline until one of the following occurs: | |
1934 | * (1) the I/O completes; (2) the pipeline stalls waiting for | |
1935 | * dependent child I/Os; (3) the I/O issues, so we're waiting | |
1936 | * for an I/O completion interrupt; (4) the I/O is delegated by | |
1937 | * vdev-level caching or aggregation; (5) the I/O is deferred | |
1938 | * due to vdev-level queueing; (6) the I/O is handed off to | |
1939 | * another thread. In all cases, the pipeline stops whenever | |
8e07b99b | 1940 | * there's no CPU work; it never burns a thread in cv_wait_io(). |
b128c09f BB |
1941 | * |
1942 | * There's no locking on io_stage because there's no legitimate way | |
1943 | * for multiple threads to be attempting to process the same I/O. | |
1944 | */ | |
428870ff | 1945 | static zio_pipe_stage_t *zio_pipeline[]; |
34dc7c2f | 1946 | |
da6b4005 NB |
1947 | /* |
1948 | * zio_execute() is a wrapper around the static function | |
1949 | * __zio_execute() so that we can force __zio_execute() to be | |
1950 | * inlined. This reduces stack overhead which is important | |
1951 | * because __zio_execute() is called recursively in several zio | |
1952 | * code paths. zio_execute() itself cannot be inlined because | |
1953 | * it is externally visible. | |
1954 | */ | |
b128c09f BB |
1955 | void |
1956 | zio_execute(zio_t *zio) | |
da6b4005 | 1957 | { |
92119cc2 BB |
1958 | fstrans_cookie_t cookie; |
1959 | ||
1960 | cookie = spl_fstrans_mark(); | |
da6b4005 | 1961 | __zio_execute(zio); |
92119cc2 | 1962 | spl_fstrans_unmark(cookie); |
da6b4005 NB |
1963 | } |
1964 | ||
b58986ee BB |
1965 | /* |
1966 | * Used to determine if in the current context the stack is sized large | |
1967 | * enough to allow zio_execute() to be called recursively. A minimum | |
1968 | * stack size of 16K is required to avoid needing to re-dispatch the zio. | |
1969 | */ | |
1970 | boolean_t | |
1971 | zio_execute_stack_check(zio_t *zio) | |
1972 | { | |
1973 | #if !defined(HAVE_LARGE_STACKS) | |
1974 | dsl_pool_t *dp = spa_get_dsl(zio->io_spa); | |
1975 | ||
1976 | /* Executing in txg_sync_thread() context. */ | |
1977 | if (dp && curthread == dp->dp_tx.tx_sync_thread) | |
1978 | return (B_TRUE); | |
1979 | ||
1980 | /* Pool initialization outside of zio_taskq context. */ | |
1981 | if (dp && spa_is_initializing(dp->dp_spa) && | |
1982 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) && | |
1983 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH)) | |
1984 | return (B_TRUE); | |
1985 | #endif /* HAVE_LARGE_STACKS */ | |
1986 | ||
1987 | return (B_FALSE); | |
1988 | } | |
1989 | ||
da6b4005 NB |
1990 | __attribute__((always_inline)) |
1991 | static inline void | |
1992 | __zio_execute(zio_t *zio) | |
b128c09f | 1993 | { |
3dfb57a3 DB |
1994 | ASSERT3U(zio->io_queued_timestamp, >, 0); |
1995 | ||
b128c09f | 1996 | while (zio->io_stage < ZIO_STAGE_DONE) { |
428870ff BB |
1997 | enum zio_stage pipeline = zio->io_pipeline; |
1998 | enum zio_stage stage = zio->io_stage; | |
62840030 MA |
1999 | |
2000 | zio->io_executor = curthread; | |
34dc7c2f | 2001 | |
b128c09f | 2002 | ASSERT(!MUTEX_HELD(&zio->io_lock)); |
428870ff BB |
2003 | ASSERT(ISP2(stage)); |
2004 | ASSERT(zio->io_stall == NULL); | |
34dc7c2f | 2005 | |
428870ff BB |
2006 | do { |
2007 | stage <<= 1; | |
2008 | } while ((stage & pipeline) == 0); | |
b128c09f BB |
2009 | |
2010 | ASSERT(stage <= ZIO_STAGE_DONE); | |
34dc7c2f BB |
2011 | |
2012 | /* | |
b128c09f BB |
2013 | * If we are in interrupt context and this pipeline stage |
2014 | * will grab a config lock that is held across I/O, | |
428870ff BB |
2015 | * or may wait for an I/O that needs an interrupt thread |
2016 | * to complete, issue async to avoid deadlock. | |
2017 | * | |
2018 | * For VDEV_IO_START, we cut in line so that the io will | |
2019 | * be sent to disk promptly. | |
34dc7c2f | 2020 | */ |
91579709 BB |
2021 | if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && |
2022 | zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { | |
b58986ee BB |
2023 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? |
2024 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
91579709 BB |
2025 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
2026 | return; | |
2027 | } | |
2028 | ||
2029 | /* | |
b58986ee BB |
2030 | * If the current context doesn't have large enough stacks |
2031 | * the zio must be issued asynchronously to prevent overflow. | |
91579709 | 2032 | */ |
b58986ee BB |
2033 | if (zio_execute_stack_check(zio)) { |
2034 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? | |
2035 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
428870ff | 2036 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
b128c09f | 2037 | return; |
34dc7c2f BB |
2038 | } |
2039 | ||
b128c09f | 2040 | zio->io_stage = stage; |
3dfb57a3 | 2041 | zio->io_pipeline_trace |= zio->io_stage; |
34dc7c2f | 2042 | |
62840030 MA |
2043 | /* |
2044 | * The zio pipeline stage returns the next zio to execute | |
2045 | * (typically the same as this one), or NULL if we should | |
2046 | * stop. | |
2047 | */ | |
2048 | zio = zio_pipeline[highbit64(stage) - 1](zio); | |
34dc7c2f | 2049 | |
62840030 MA |
2050 | if (zio == NULL) |
2051 | return; | |
b128c09f | 2052 | } |
34dc7c2f BB |
2053 | } |
2054 | ||
da6b4005 | 2055 | |
b128c09f BB |
2056 | /* |
2057 | * ========================================================================== | |
2058 | * Initiate I/O, either sync or async | |
2059 | * ========================================================================== | |
2060 | */ | |
2061 | int | |
2062 | zio_wait(zio_t *zio) | |
34dc7c2f | 2063 | { |
8fb1ede1 | 2064 | long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms); |
b128c09f | 2065 | int error; |
34dc7c2f | 2066 | |
1ce23dca PS |
2067 | ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN); |
2068 | ASSERT3P(zio->io_executor, ==, NULL); | |
34dc7c2f | 2069 | |
b128c09f | 2070 | zio->io_waiter = curthread; |
3dfb57a3 DB |
2071 | ASSERT0(zio->io_queued_timestamp); |
2072 | zio->io_queued_timestamp = gethrtime(); | |
34dc7c2f | 2073 | |
da6b4005 | 2074 | __zio_execute(zio); |
34dc7c2f | 2075 | |
b128c09f | 2076 | mutex_enter(&zio->io_lock); |
8fb1ede1 BB |
2077 | while (zio->io_executor != NULL) { |
2078 | error = cv_timedwait_io(&zio->io_cv, &zio->io_lock, | |
2079 | ddi_get_lbolt() + timeout); | |
2080 | ||
2081 | if (zfs_deadman_enabled && error == -1 && | |
2082 | gethrtime() - zio->io_queued_timestamp > | |
2083 | spa_deadman_ziotime(zio->io_spa)) { | |
2084 | mutex_exit(&zio->io_lock); | |
2085 | timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms); | |
2086 | zio_deadman(zio, FTAG); | |
2087 | mutex_enter(&zio->io_lock); | |
2088 | } | |
2089 | } | |
b128c09f | 2090 | mutex_exit(&zio->io_lock); |
34dc7c2f | 2091 | |
b128c09f BB |
2092 | error = zio->io_error; |
2093 | zio_destroy(zio); | |
34dc7c2f | 2094 | |
b128c09f BB |
2095 | return (error); |
2096 | } | |
34dc7c2f | 2097 | |
b128c09f BB |
2098 | void |
2099 | zio_nowait(zio_t *zio) | |
2100 | { | |
1ce23dca | 2101 | ASSERT3P(zio->io_executor, ==, NULL); |
34dc7c2f | 2102 | |
d164b209 BB |
2103 | if (zio->io_child_type == ZIO_CHILD_LOGICAL && |
2104 | zio_unique_parent(zio) == NULL) { | |
8878261f BB |
2105 | zio_t *pio; |
2106 | ||
34dc7c2f | 2107 | /* |
b128c09f | 2108 | * This is a logical async I/O with no parent to wait for it. |
9babb374 BB |
2109 | * We add it to the spa_async_root_zio "Godfather" I/O which |
2110 | * will ensure they complete prior to unloading the pool. | |
34dc7c2f | 2111 | */ |
b128c09f | 2112 | spa_t *spa = zio->io_spa; |
8878261f BB |
2113 | kpreempt_disable(); |
2114 | pio = spa->spa_async_zio_root[CPU_SEQID]; | |
2115 | kpreempt_enable(); | |
9babb374 | 2116 | |
8878261f | 2117 | zio_add_child(pio, zio); |
b128c09f | 2118 | } |
34dc7c2f | 2119 | |
3dfb57a3 DB |
2120 | ASSERT0(zio->io_queued_timestamp); |
2121 | zio->io_queued_timestamp = gethrtime(); | |
da6b4005 | 2122 | __zio_execute(zio); |
b128c09f | 2123 | } |
34dc7c2f | 2124 | |
b128c09f BB |
2125 | /* |
2126 | * ========================================================================== | |
1ce23dca | 2127 | * Reexecute, cancel, or suspend/resume failed I/O |
b128c09f BB |
2128 | * ========================================================================== |
2129 | */ | |
34dc7c2f | 2130 | |
b128c09f BB |
2131 | static void |
2132 | zio_reexecute(zio_t *pio) | |
2133 | { | |
d164b209 BB |
2134 | zio_t *cio, *cio_next; |
2135 | ||
2136 | ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); | |
2137 | ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); | |
9babb374 BB |
2138 | ASSERT(pio->io_gang_leader == NULL); |
2139 | ASSERT(pio->io_gang_tree == NULL); | |
34dc7c2f | 2140 | |
b128c09f BB |
2141 | pio->io_flags = pio->io_orig_flags; |
2142 | pio->io_stage = pio->io_orig_stage; | |
2143 | pio->io_pipeline = pio->io_orig_pipeline; | |
2144 | pio->io_reexecute = 0; | |
03c6040b | 2145 | pio->io_flags |= ZIO_FLAG_REEXECUTED; |
3dfb57a3 | 2146 | pio->io_pipeline_trace = 0; |
b128c09f | 2147 | pio->io_error = 0; |
1c27024e | 2148 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 | 2149 | pio->io_state[w] = 0; |
1c27024e | 2150 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) |
b128c09f | 2151 | pio->io_child_error[c] = 0; |
34dc7c2f | 2152 | |
428870ff BB |
2153 | if (IO_IS_ALLOCATING(pio)) |
2154 | BP_ZERO(pio->io_bp); | |
34dc7c2f | 2155 | |
b128c09f BB |
2156 | /* |
2157 | * As we reexecute pio's children, new children could be created. | |
d164b209 | 2158 | * New children go to the head of pio's io_child_list, however, |
b128c09f | 2159 | * so we will (correctly) not reexecute them. The key is that |
d164b209 BB |
2160 | * the remainder of pio's io_child_list, from 'cio_next' onward, |
2161 | * cannot be affected by any side effects of reexecuting 'cio'. | |
b128c09f | 2162 | */ |
1c27024e | 2163 | zio_link_t *zl = NULL; |
a8b2e306 | 2164 | mutex_enter(&pio->io_lock); |
3dfb57a3 DB |
2165 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { |
2166 | cio_next = zio_walk_children(pio, &zl); | |
1c27024e | 2167 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 | 2168 | pio->io_children[cio->io_child_type][w]++; |
b128c09f | 2169 | mutex_exit(&pio->io_lock); |
d164b209 | 2170 | zio_reexecute(cio); |
a8b2e306 | 2171 | mutex_enter(&pio->io_lock); |
34dc7c2f | 2172 | } |
a8b2e306 | 2173 | mutex_exit(&pio->io_lock); |
34dc7c2f | 2174 | |
b128c09f BB |
2175 | /* |
2176 | * Now that all children have been reexecuted, execute the parent. | |
9babb374 | 2177 | * We don't reexecute "The Godfather" I/O here as it's the |
9e2c3bb4 | 2178 | * responsibility of the caller to wait on it. |
b128c09f | 2179 | */ |
3dfb57a3 DB |
2180 | if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) { |
2181 | pio->io_queued_timestamp = gethrtime(); | |
da6b4005 | 2182 | __zio_execute(pio); |
3dfb57a3 | 2183 | } |
34dc7c2f BB |
2184 | } |
2185 | ||
b128c09f | 2186 | void |
cec3a0a1 | 2187 | zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason) |
34dc7c2f | 2188 | { |
b128c09f BB |
2189 | if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) |
2190 | fm_panic("Pool '%s' has encountered an uncorrectable I/O " | |
2191 | "failure and the failure mode property for this pool " | |
2192 | "is set to panic.", spa_name(spa)); | |
34dc7c2f | 2193 | |
bf89c199 BB |
2194 | cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O " |
2195 | "failure and has been suspended.\n", spa_name(spa)); | |
2196 | ||
b5256303 TC |
2197 | zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, |
2198 | NULL, NULL, 0, 0); | |
34dc7c2f | 2199 | |
b128c09f | 2200 | mutex_enter(&spa->spa_suspend_lock); |
34dc7c2f | 2201 | |
b128c09f | 2202 | if (spa->spa_suspend_zio_root == NULL) |
9babb374 BB |
2203 | spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, |
2204 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
2205 | ZIO_FLAG_GODFATHER); | |
34dc7c2f | 2206 | |
cec3a0a1 | 2207 | spa->spa_suspended = reason; |
34dc7c2f | 2208 | |
b128c09f | 2209 | if (zio != NULL) { |
9babb374 | 2210 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f BB |
2211 | ASSERT(zio != spa->spa_suspend_zio_root); |
2212 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
d164b209 | 2213 | ASSERT(zio_unique_parent(zio) == NULL); |
b128c09f BB |
2214 | ASSERT(zio->io_stage == ZIO_STAGE_DONE); |
2215 | zio_add_child(spa->spa_suspend_zio_root, zio); | |
2216 | } | |
34dc7c2f | 2217 | |
b128c09f BB |
2218 | mutex_exit(&spa->spa_suspend_lock); |
2219 | } | |
34dc7c2f | 2220 | |
9babb374 | 2221 | int |
b128c09f BB |
2222 | zio_resume(spa_t *spa) |
2223 | { | |
9babb374 | 2224 | zio_t *pio; |
34dc7c2f BB |
2225 | |
2226 | /* | |
b128c09f | 2227 | * Reexecute all previously suspended i/o. |
34dc7c2f | 2228 | */ |
b128c09f | 2229 | mutex_enter(&spa->spa_suspend_lock); |
cec3a0a1 | 2230 | spa->spa_suspended = ZIO_SUSPEND_NONE; |
b128c09f BB |
2231 | cv_broadcast(&spa->spa_suspend_cv); |
2232 | pio = spa->spa_suspend_zio_root; | |
2233 | spa->spa_suspend_zio_root = NULL; | |
2234 | mutex_exit(&spa->spa_suspend_lock); | |
2235 | ||
2236 | if (pio == NULL) | |
9babb374 | 2237 | return (0); |
34dc7c2f | 2238 | |
9babb374 BB |
2239 | zio_reexecute(pio); |
2240 | return (zio_wait(pio)); | |
b128c09f BB |
2241 | } |
2242 | ||
2243 | void | |
2244 | zio_resume_wait(spa_t *spa) | |
2245 | { | |
2246 | mutex_enter(&spa->spa_suspend_lock); | |
2247 | while (spa_suspended(spa)) | |
2248 | cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); | |
2249 | mutex_exit(&spa->spa_suspend_lock); | |
34dc7c2f BB |
2250 | } |
2251 | ||
2252 | /* | |
2253 | * ========================================================================== | |
b128c09f BB |
2254 | * Gang blocks. |
2255 | * | |
2256 | * A gang block is a collection of small blocks that looks to the DMU | |
2257 | * like one large block. When zio_dva_allocate() cannot find a block | |
2258 | * of the requested size, due to either severe fragmentation or the pool | |
2259 | * being nearly full, it calls zio_write_gang_block() to construct the | |
2260 | * block from smaller fragments. | |
2261 | * | |
2262 | * A gang block consists of a gang header (zio_gbh_phys_t) and up to | |
2263 | * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like | |
2264 | * an indirect block: it's an array of block pointers. It consumes | |
2265 | * only one sector and hence is allocatable regardless of fragmentation. | |
2266 | * The gang header's bps point to its gang members, which hold the data. | |
2267 | * | |
2268 | * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> | |
2269 | * as the verifier to ensure uniqueness of the SHA256 checksum. | |
2270 | * Critically, the gang block bp's blk_cksum is the checksum of the data, | |
2271 | * not the gang header. This ensures that data block signatures (needed for | |
2272 | * deduplication) are independent of how the block is physically stored. | |
2273 | * | |
2274 | * Gang blocks can be nested: a gang member may itself be a gang block. | |
2275 | * Thus every gang block is a tree in which root and all interior nodes are | |
2276 | * gang headers, and the leaves are normal blocks that contain user data. | |
2277 | * The root of the gang tree is called the gang leader. | |
2278 | * | |
2279 | * To perform any operation (read, rewrite, free, claim) on a gang block, | |
2280 | * zio_gang_assemble() first assembles the gang tree (minus data leaves) | |
2281 | * in the io_gang_tree field of the original logical i/o by recursively | |
2282 | * reading the gang leader and all gang headers below it. This yields | |
2283 | * an in-core tree containing the contents of every gang header and the | |
2284 | * bps for every constituent of the gang block. | |
2285 | * | |
2286 | * With the gang tree now assembled, zio_gang_issue() just walks the gang tree | |
2287 | * and invokes a callback on each bp. To free a gang block, zio_gang_issue() | |
2288 | * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. | |
2289 | * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). | |
2290 | * zio_read_gang() is a wrapper around zio_read() that omits reading gang | |
2291 | * headers, since we already have those in io_gang_tree. zio_rewrite_gang() | |
2292 | * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() | |
2293 | * of the gang header plus zio_checksum_compute() of the data to update the | |
2294 | * gang header's blk_cksum as described above. | |
2295 | * | |
2296 | * The two-phase assemble/issue model solves the problem of partial failure -- | |
2297 | * what if you'd freed part of a gang block but then couldn't read the | |
2298 | * gang header for another part? Assembling the entire gang tree first | |
2299 | * ensures that all the necessary gang header I/O has succeeded before | |
2300 | * starting the actual work of free, claim, or write. Once the gang tree | |
2301 | * is assembled, free and claim are in-memory operations that cannot fail. | |
2302 | * | |
2303 | * In the event that a gang write fails, zio_dva_unallocate() walks the | |
2304 | * gang tree to immediately free (i.e. insert back into the space map) | |
2305 | * everything we've allocated. This ensures that we don't get ENOSPC | |
2306 | * errors during repeated suspend/resume cycles due to a flaky device. | |
2307 | * | |
2308 | * Gang rewrites only happen during sync-to-convergence. If we can't assemble | |
2309 | * the gang tree, we won't modify the block, so we can safely defer the free | |
2310 | * (knowing that the block is still intact). If we *can* assemble the gang | |
2311 | * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free | |
2312 | * each constituent bp and we can allocate a new block on the next sync pass. | |
2313 | * | |
2314 | * In all cases, the gang tree allows complete recovery from partial failure. | |
34dc7c2f BB |
2315 | * ========================================================================== |
2316 | */ | |
b128c09f | 2317 | |
a6255b7f DQ |
2318 | static void |
2319 | zio_gang_issue_func_done(zio_t *zio) | |
2320 | { | |
2321 | abd_put(zio->io_abd); | |
2322 | } | |
2323 | ||
b128c09f | 2324 | static zio_t * |
a6255b7f DQ |
2325 | zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, |
2326 | uint64_t offset) | |
34dc7c2f | 2327 | { |
b128c09f BB |
2328 | if (gn != NULL) |
2329 | return (pio); | |
34dc7c2f | 2330 | |
a6255b7f DQ |
2331 | return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset), |
2332 | BP_GET_PSIZE(bp), zio_gang_issue_func_done, | |
2333 | NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
b128c09f BB |
2334 | &pio->io_bookmark)); |
2335 | } | |
2336 | ||
a6255b7f DQ |
2337 | static zio_t * |
2338 | zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2339 | uint64_t offset) | |
b128c09f BB |
2340 | { |
2341 | zio_t *zio; | |
2342 | ||
2343 | if (gn != NULL) { | |
a6255b7f DQ |
2344 | abd_t *gbh_abd = |
2345 | abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
b128c09f | 2346 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
a6255b7f DQ |
2347 | gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL, |
2348 | pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
2349 | &pio->io_bookmark); | |
34dc7c2f | 2350 | /* |
b128c09f BB |
2351 | * As we rewrite each gang header, the pipeline will compute |
2352 | * a new gang block header checksum for it; but no one will | |
2353 | * compute a new data checksum, so we do that here. The one | |
2354 | * exception is the gang leader: the pipeline already computed | |
2355 | * its data checksum because that stage precedes gang assembly. | |
2356 | * (Presently, nothing actually uses interior data checksums; | |
2357 | * this is just good hygiene.) | |
34dc7c2f | 2358 | */ |
9babb374 | 2359 | if (gn != pio->io_gang_leader->io_gang_tree) { |
a6255b7f DQ |
2360 | abd_t *buf = abd_get_offset(data, offset); |
2361 | ||
b128c09f | 2362 | zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), |
a6255b7f DQ |
2363 | buf, BP_GET_PSIZE(bp)); |
2364 | ||
2365 | abd_put(buf); | |
b128c09f | 2366 | } |
428870ff BB |
2367 | /* |
2368 | * If we are here to damage data for testing purposes, | |
2369 | * leave the GBH alone so that we can detect the damage. | |
2370 | */ | |
2371 | if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) | |
2372 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; | |
34dc7c2f | 2373 | } else { |
b128c09f | 2374 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
a6255b7f DQ |
2375 | abd_get_offset(data, offset), BP_GET_PSIZE(bp), |
2376 | zio_gang_issue_func_done, NULL, pio->io_priority, | |
b128c09f | 2377 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); |
34dc7c2f BB |
2378 | } |
2379 | ||
b128c09f BB |
2380 | return (zio); |
2381 | } | |
34dc7c2f | 2382 | |
b128c09f | 2383 | /* ARGSUSED */ |
a6255b7f DQ |
2384 | static zio_t * |
2385 | zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2386 | uint64_t offset) | |
b128c09f | 2387 | { |
428870ff BB |
2388 | return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, |
2389 | ZIO_GANG_CHILD_FLAGS(pio))); | |
34dc7c2f BB |
2390 | } |
2391 | ||
b128c09f | 2392 | /* ARGSUSED */ |
a6255b7f DQ |
2393 | static zio_t * |
2394 | zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2395 | uint64_t offset) | |
34dc7c2f | 2396 | { |
b128c09f BB |
2397 | return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, |
2398 | NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); | |
2399 | } | |
2400 | ||
2401 | static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { | |
2402 | NULL, | |
2403 | zio_read_gang, | |
2404 | zio_rewrite_gang, | |
2405 | zio_free_gang, | |
2406 | zio_claim_gang, | |
2407 | NULL | |
2408 | }; | |
34dc7c2f | 2409 | |
b128c09f | 2410 | static void zio_gang_tree_assemble_done(zio_t *zio); |
34dc7c2f | 2411 | |
b128c09f BB |
2412 | static zio_gang_node_t * |
2413 | zio_gang_node_alloc(zio_gang_node_t **gnpp) | |
2414 | { | |
2415 | zio_gang_node_t *gn; | |
34dc7c2f | 2416 | |
b128c09f | 2417 | ASSERT(*gnpp == NULL); |
34dc7c2f | 2418 | |
79c76d5b | 2419 | gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); |
b128c09f BB |
2420 | gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); |
2421 | *gnpp = gn; | |
34dc7c2f | 2422 | |
b128c09f | 2423 | return (gn); |
34dc7c2f BB |
2424 | } |
2425 | ||
34dc7c2f | 2426 | static void |
b128c09f | 2427 | zio_gang_node_free(zio_gang_node_t **gnpp) |
34dc7c2f | 2428 | { |
b128c09f | 2429 | zio_gang_node_t *gn = *gnpp; |
34dc7c2f | 2430 | |
1c27024e | 2431 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f BB |
2432 | ASSERT(gn->gn_child[g] == NULL); |
2433 | ||
2434 | zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
2435 | kmem_free(gn, sizeof (*gn)); | |
2436 | *gnpp = NULL; | |
34dc7c2f BB |
2437 | } |
2438 | ||
b128c09f BB |
2439 | static void |
2440 | zio_gang_tree_free(zio_gang_node_t **gnpp) | |
34dc7c2f | 2441 | { |
b128c09f | 2442 | zio_gang_node_t *gn = *gnpp; |
34dc7c2f | 2443 | |
b128c09f BB |
2444 | if (gn == NULL) |
2445 | return; | |
34dc7c2f | 2446 | |
1c27024e | 2447 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f | 2448 | zio_gang_tree_free(&gn->gn_child[g]); |
34dc7c2f | 2449 | |
b128c09f | 2450 | zio_gang_node_free(gnpp); |
34dc7c2f BB |
2451 | } |
2452 | ||
b128c09f | 2453 | static void |
9babb374 | 2454 | zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) |
34dc7c2f | 2455 | { |
b128c09f | 2456 | zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); |
a6255b7f | 2457 | abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); |
b128c09f | 2458 | |
9babb374 | 2459 | ASSERT(gio->io_gang_leader == gio); |
b128c09f | 2460 | ASSERT(BP_IS_GANG(bp)); |
34dc7c2f | 2461 | |
a6255b7f DQ |
2462 | zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE, |
2463 | zio_gang_tree_assemble_done, gn, gio->io_priority, | |
2464 | ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); | |
b128c09f | 2465 | } |
34dc7c2f | 2466 | |
b128c09f BB |
2467 | static void |
2468 | zio_gang_tree_assemble_done(zio_t *zio) | |
2469 | { | |
9babb374 | 2470 | zio_t *gio = zio->io_gang_leader; |
b128c09f BB |
2471 | zio_gang_node_t *gn = zio->io_private; |
2472 | blkptr_t *bp = zio->io_bp; | |
34dc7c2f | 2473 | |
9babb374 | 2474 | ASSERT(gio == zio_unique_parent(zio)); |
428870ff | 2475 | ASSERT(zio->io_child_count == 0); |
34dc7c2f | 2476 | |
b128c09f BB |
2477 | if (zio->io_error) |
2478 | return; | |
34dc7c2f | 2479 | |
a6255b7f | 2480 | /* this ABD was created from a linear buf in zio_gang_tree_assemble */ |
b128c09f | 2481 | if (BP_SHOULD_BYTESWAP(bp)) |
a6255b7f | 2482 | byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size); |
34dc7c2f | 2483 | |
a6255b7f | 2484 | ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh); |
b128c09f | 2485 | ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); |
428870ff | 2486 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2487 | |
a6255b7f DQ |
2488 | abd_put(zio->io_abd); |
2489 | ||
1c27024e | 2490 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2491 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2492 | if (!BP_IS_GANG(gbp)) | |
2493 | continue; | |
9babb374 | 2494 | zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); |
b128c09f | 2495 | } |
34dc7c2f BB |
2496 | } |
2497 | ||
b128c09f | 2498 | static void |
a6255b7f DQ |
2499 | zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data, |
2500 | uint64_t offset) | |
34dc7c2f | 2501 | { |
9babb374 | 2502 | zio_t *gio = pio->io_gang_leader; |
b128c09f | 2503 | zio_t *zio; |
34dc7c2f | 2504 | |
b128c09f | 2505 | ASSERT(BP_IS_GANG(bp) == !!gn); |
9babb374 BB |
2506 | ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); |
2507 | ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); | |
34dc7c2f | 2508 | |
b128c09f BB |
2509 | /* |
2510 | * If you're a gang header, your data is in gn->gn_gbh. | |
2511 | * If you're a gang member, your data is in 'data' and gn == NULL. | |
2512 | */ | |
a6255b7f | 2513 | zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset); |
34dc7c2f | 2514 | |
b128c09f | 2515 | if (gn != NULL) { |
428870ff | 2516 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2517 | |
1c27024e | 2518 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2519 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2520 | if (BP_IS_HOLE(gbp)) | |
2521 | continue; | |
a6255b7f DQ |
2522 | zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data, |
2523 | offset); | |
2524 | offset += BP_GET_PSIZE(gbp); | |
b128c09f | 2525 | } |
34dc7c2f BB |
2526 | } |
2527 | ||
9babb374 | 2528 | if (gn == gio->io_gang_tree) |
a6255b7f | 2529 | ASSERT3U(gio->io_size, ==, offset); |
34dc7c2f | 2530 | |
b128c09f BB |
2531 | if (zio != pio) |
2532 | zio_nowait(zio); | |
34dc7c2f BB |
2533 | } |
2534 | ||
62840030 | 2535 | static zio_t * |
b128c09f | 2536 | zio_gang_assemble(zio_t *zio) |
34dc7c2f | 2537 | { |
b128c09f | 2538 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2539 | |
9babb374 BB |
2540 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); |
2541 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
2542 | ||
2543 | zio->io_gang_leader = zio; | |
34dc7c2f | 2544 | |
b128c09f | 2545 | zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); |
34dc7c2f | 2546 | |
62840030 | 2547 | return (zio); |
34dc7c2f BB |
2548 | } |
2549 | ||
62840030 | 2550 | static zio_t * |
b128c09f | 2551 | zio_gang_issue(zio_t *zio) |
34dc7c2f | 2552 | { |
b128c09f | 2553 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2554 | |
ddc751d5 | 2555 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) { |
62840030 | 2556 | return (NULL); |
ddc751d5 | 2557 | } |
34dc7c2f | 2558 | |
9babb374 BB |
2559 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); |
2560 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
34dc7c2f | 2561 | |
b128c09f | 2562 | if (zio->io_child_error[ZIO_CHILD_GANG] == 0) |
a6255b7f DQ |
2563 | zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd, |
2564 | 0); | |
b128c09f | 2565 | else |
9babb374 | 2566 | zio_gang_tree_free(&zio->io_gang_tree); |
34dc7c2f | 2567 | |
b128c09f | 2568 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f | 2569 | |
62840030 | 2570 | return (zio); |
34dc7c2f BB |
2571 | } |
2572 | ||
2573 | static void | |
b128c09f | 2574 | zio_write_gang_member_ready(zio_t *zio) |
34dc7c2f | 2575 | { |
d164b209 | 2576 | zio_t *pio = zio_unique_parent(zio); |
34dc7c2f BB |
2577 | dva_t *cdva = zio->io_bp->blk_dva; |
2578 | dva_t *pdva = pio->io_bp->blk_dva; | |
2579 | uint64_t asize; | |
d1d7e268 | 2580 | ASSERTV(zio_t *gio = zio->io_gang_leader); |
34dc7c2f | 2581 | |
b128c09f BB |
2582 | if (BP_IS_HOLE(zio->io_bp)) |
2583 | return; | |
2584 | ||
2585 | ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); | |
2586 | ||
2587 | ASSERT(zio->io_child_type == ZIO_CHILD_GANG); | |
428870ff BB |
2588 | ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); |
2589 | ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); | |
2590 | ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); | |
34dc7c2f | 2591 | ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); |
34dc7c2f BB |
2592 | |
2593 | mutex_enter(&pio->io_lock); | |
1c27024e | 2594 | for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { |
34dc7c2f BB |
2595 | ASSERT(DVA_GET_GANG(&pdva[d])); |
2596 | asize = DVA_GET_ASIZE(&pdva[d]); | |
2597 | asize += DVA_GET_ASIZE(&cdva[d]); | |
2598 | DVA_SET_ASIZE(&pdva[d], asize); | |
2599 | } | |
2600 | mutex_exit(&pio->io_lock); | |
2601 | } | |
2602 | ||
a6255b7f DQ |
2603 | static void |
2604 | zio_write_gang_done(zio_t *zio) | |
2605 | { | |
2606 | abd_put(zio->io_abd); | |
2607 | } | |
2608 | ||
62840030 | 2609 | static zio_t * |
b128c09f | 2610 | zio_write_gang_block(zio_t *pio) |
34dc7c2f | 2611 | { |
b128c09f | 2612 | spa_t *spa = pio->io_spa; |
3dfb57a3 | 2613 | metaslab_class_t *mc = spa_normal_class(spa); |
b128c09f | 2614 | blkptr_t *bp = pio->io_bp; |
9babb374 | 2615 | zio_t *gio = pio->io_gang_leader; |
b128c09f BB |
2616 | zio_t *zio; |
2617 | zio_gang_node_t *gn, **gnpp; | |
34dc7c2f | 2618 | zio_gbh_phys_t *gbh; |
a6255b7f | 2619 | abd_t *gbh_abd; |
b128c09f BB |
2620 | uint64_t txg = pio->io_txg; |
2621 | uint64_t resid = pio->io_size; | |
2622 | uint64_t lsize; | |
428870ff | 2623 | int copies = gio->io_prop.zp_copies; |
b5256303 | 2624 | int gbh_copies; |
b128c09f | 2625 | zio_prop_t zp; |
1c27024e | 2626 | int error; |
b5256303 TC |
2627 | |
2628 | /* | |
2629 | * encrypted blocks need DVA[2] free so encrypted gang headers can't | |
2630 | * have a third copy. | |
2631 | */ | |
2632 | gbh_copies = MIN(copies + 1, spa_max_replication(spa)); | |
2633 | if (gio->io_prop.zp_encrypt && gbh_copies >= SPA_DVAS_PER_BP) | |
2634 | gbh_copies = SPA_DVAS_PER_BP - 1; | |
2635 | ||
1c27024e | 2636 | int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER; |
3dfb57a3 DB |
2637 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
2638 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
2639 | ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); | |
2640 | ||
2641 | flags |= METASLAB_ASYNC_ALLOC; | |
492f64e9 PD |
2642 | VERIFY(refcount_held(&mc->mc_alloc_slots[pio->io_allocator], |
2643 | pio)); | |
3dfb57a3 DB |
2644 | |
2645 | /* | |
2646 | * The logical zio has already placed a reservation for | |
2647 | * 'copies' allocation slots but gang blocks may require | |
2648 | * additional copies. These additional copies | |
2649 | * (i.e. gbh_copies - copies) are guaranteed to succeed | |
2650 | * since metaslab_class_throttle_reserve() always allows | |
2651 | * additional reservations for gang blocks. | |
2652 | */ | |
2653 | VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies, | |
492f64e9 | 2654 | pio->io_allocator, pio, flags)); |
3dfb57a3 DB |
2655 | } |
2656 | ||
2657 | error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE, | |
4e21fd06 | 2658 | bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, |
492f64e9 | 2659 | &pio->io_alloc_list, pio, pio->io_allocator); |
34dc7c2f | 2660 | if (error) { |
3dfb57a3 DB |
2661 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
2662 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
2663 | ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); | |
2664 | ||
2665 | /* | |
2666 | * If we failed to allocate the gang block header then | |
2667 | * we remove any additional allocation reservations that | |
2668 | * we placed here. The original reservation will | |
2669 | * be removed when the logical I/O goes to the ready | |
2670 | * stage. | |
2671 | */ | |
2672 | metaslab_class_throttle_unreserve(mc, | |
492f64e9 | 2673 | gbh_copies - copies, pio->io_allocator, pio); |
3dfb57a3 DB |
2674 | } |
2675 | ||
b128c09f | 2676 | pio->io_error = error; |
62840030 | 2677 | return (pio); |
34dc7c2f BB |
2678 | } |
2679 | ||
9babb374 BB |
2680 | if (pio == gio) { |
2681 | gnpp = &gio->io_gang_tree; | |
b128c09f BB |
2682 | } else { |
2683 | gnpp = pio->io_private; | |
2684 | ASSERT(pio->io_ready == zio_write_gang_member_ready); | |
34dc7c2f BB |
2685 | } |
2686 | ||
b128c09f BB |
2687 | gn = zio_gang_node_alloc(gnpp); |
2688 | gbh = gn->gn_gbh; | |
2689 | bzero(gbh, SPA_GANGBLOCKSIZE); | |
a6255b7f | 2690 | gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE); |
34dc7c2f | 2691 | |
b128c09f BB |
2692 | /* |
2693 | * Create the gang header. | |
2694 | */ | |
a6255b7f DQ |
2695 | zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE, |
2696 | zio_write_gang_done, NULL, pio->io_priority, | |
2697 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); | |
34dc7c2f | 2698 | |
b128c09f BB |
2699 | /* |
2700 | * Create and nowait the gang children. | |
2701 | */ | |
1c27024e | 2702 | for (int g = 0; resid != 0; resid -= lsize, g++) { |
b128c09f BB |
2703 | lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), |
2704 | SPA_MINBLOCKSIZE); | |
2705 | ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); | |
2706 | ||
9babb374 | 2707 | zp.zp_checksum = gio->io_prop.zp_checksum; |
b128c09f BB |
2708 | zp.zp_compress = ZIO_COMPRESS_OFF; |
2709 | zp.zp_type = DMU_OT_NONE; | |
2710 | zp.zp_level = 0; | |
428870ff | 2711 | zp.zp_copies = gio->io_prop.zp_copies; |
03c6040b GW |
2712 | zp.zp_dedup = B_FALSE; |
2713 | zp.zp_dedup_verify = B_FALSE; | |
2714 | zp.zp_nopwrite = B_FALSE; | |
4807c0ba TC |
2715 | zp.zp_encrypt = gio->io_prop.zp_encrypt; |
2716 | zp.zp_byteorder = gio->io_prop.zp_byteorder; | |
b5256303 TC |
2717 | bzero(zp.zp_salt, ZIO_DATA_SALT_LEN); |
2718 | bzero(zp.zp_iv, ZIO_DATA_IV_LEN); | |
2719 | bzero(zp.zp_mac, ZIO_DATA_MAC_LEN); | |
b128c09f | 2720 | |
1c27024e | 2721 | zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g], |
a6255b7f DQ |
2722 | abd_get_offset(pio->io_abd, pio->io_size - resid), lsize, |
2723 | lsize, &zp, zio_write_gang_member_ready, NULL, NULL, | |
2724 | zio_write_gang_done, &gn->gn_child[g], pio->io_priority, | |
3dfb57a3 DB |
2725 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); |
2726 | ||
2727 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { | |
2728 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
2729 | ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); | |
2730 | ||
2731 | /* | |
2732 | * Gang children won't throttle but we should | |
2733 | * account for their work, so reserve an allocation | |
2734 | * slot for them here. | |
2735 | */ | |
2736 | VERIFY(metaslab_class_throttle_reserve(mc, | |
492f64e9 | 2737 | zp.zp_copies, cio->io_allocator, cio, flags)); |
3dfb57a3 DB |
2738 | } |
2739 | zio_nowait(cio); | |
b128c09f | 2740 | } |
34dc7c2f BB |
2741 | |
2742 | /* | |
b128c09f | 2743 | * Set pio's pipeline to just wait for zio to finish. |
34dc7c2f | 2744 | */ |
b128c09f BB |
2745 | pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
2746 | ||
920dd524 ED |
2747 | /* |
2748 | * We didn't allocate this bp, so make sure it doesn't get unmarked. | |
2749 | */ | |
2750 | pio->io_flags &= ~ZIO_FLAG_FASTWRITE; | |
2751 | ||
b128c09f BB |
2752 | zio_nowait(zio); |
2753 | ||
62840030 | 2754 | return (pio); |
34dc7c2f BB |
2755 | } |
2756 | ||
03c6040b | 2757 | /* |
3c67d83a TH |
2758 | * The zio_nop_write stage in the pipeline determines if allocating a |
2759 | * new bp is necessary. The nopwrite feature can handle writes in | |
2760 | * either syncing or open context (i.e. zil writes) and as a result is | |
2761 | * mutually exclusive with dedup. | |
2762 | * | |
2763 | * By leveraging a cryptographically secure checksum, such as SHA256, we | |
2764 | * can compare the checksums of the new data and the old to determine if | |
2765 | * allocating a new block is required. Note that our requirements for | |
2766 | * cryptographic strength are fairly weak: there can't be any accidental | |
2767 | * hash collisions, but we don't need to be secure against intentional | |
2768 | * (malicious) collisions. To trigger a nopwrite, you have to be able | |
2769 | * to write the file to begin with, and triggering an incorrect (hash | |
2770 | * collision) nopwrite is no worse than simply writing to the file. | |
2771 | * That said, there are no known attacks against the checksum algorithms | |
2772 | * used for nopwrite, assuming that the salt and the checksums | |
2773 | * themselves remain secret. | |
03c6040b | 2774 | */ |
62840030 | 2775 | static zio_t * |
03c6040b GW |
2776 | zio_nop_write(zio_t *zio) |
2777 | { | |
2778 | blkptr_t *bp = zio->io_bp; | |
2779 | blkptr_t *bp_orig = &zio->io_bp_orig; | |
2780 | zio_prop_t *zp = &zio->io_prop; | |
2781 | ||
2782 | ASSERT(BP_GET_LEVEL(bp) == 0); | |
2783 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
2784 | ASSERT(zp->zp_nopwrite); | |
2785 | ASSERT(!zp->zp_dedup); | |
2786 | ASSERT(zio->io_bp_override == NULL); | |
2787 | ASSERT(IO_IS_ALLOCATING(zio)); | |
2788 | ||
2789 | /* | |
2790 | * Check to see if the original bp and the new bp have matching | |
2791 | * characteristics (i.e. same checksum, compression algorithms, etc). | |
2792 | * If they don't then just continue with the pipeline which will | |
2793 | * allocate a new bp. | |
2794 | */ | |
2795 | if (BP_IS_HOLE(bp_orig) || | |
3c67d83a TH |
2796 | !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & |
2797 | ZCHECKSUM_FLAG_NOPWRITE) || | |
b5256303 | 2798 | BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) || |
03c6040b GW |
2799 | BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || |
2800 | BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || | |
2801 | BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || | |
2802 | zp->zp_copies != BP_GET_NDVAS(bp_orig)) | |
62840030 | 2803 | return (zio); |
03c6040b GW |
2804 | |
2805 | /* | |
2806 | * If the checksums match then reset the pipeline so that we | |
2807 | * avoid allocating a new bp and issuing any I/O. | |
2808 | */ | |
2809 | if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { | |
3c67d83a TH |
2810 | ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & |
2811 | ZCHECKSUM_FLAG_NOPWRITE); | |
03c6040b GW |
2812 | ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); |
2813 | ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); | |
2814 | ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); | |
2815 | ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, | |
2816 | sizeof (uint64_t)) == 0); | |
2817 | ||
2818 | *bp = *bp_orig; | |
2819 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
2820 | zio->io_flags |= ZIO_FLAG_NOPWRITE; | |
2821 | } | |
2822 | ||
62840030 | 2823 | return (zio); |
03c6040b GW |
2824 | } |
2825 | ||
34dc7c2f BB |
2826 | /* |
2827 | * ========================================================================== | |
428870ff | 2828 | * Dedup |
34dc7c2f BB |
2829 | * ========================================================================== |
2830 | */ | |
428870ff BB |
2831 | static void |
2832 | zio_ddt_child_read_done(zio_t *zio) | |
2833 | { | |
2834 | blkptr_t *bp = zio->io_bp; | |
2835 | ddt_entry_t *dde = zio->io_private; | |
2836 | ddt_phys_t *ddp; | |
2837 | zio_t *pio = zio_unique_parent(zio); | |
2838 | ||
2839 | mutex_enter(&pio->io_lock); | |
2840 | ddp = ddt_phys_select(dde, bp); | |
2841 | if (zio->io_error == 0) | |
2842 | ddt_phys_clear(ddp); /* this ddp doesn't need repair */ | |
a6255b7f DQ |
2843 | |
2844 | if (zio->io_error == 0 && dde->dde_repair_abd == NULL) | |
2845 | dde->dde_repair_abd = zio->io_abd; | |
428870ff | 2846 | else |
a6255b7f | 2847 | abd_free(zio->io_abd); |
428870ff BB |
2848 | mutex_exit(&pio->io_lock); |
2849 | } | |
2850 | ||
62840030 | 2851 | static zio_t * |
428870ff BB |
2852 | zio_ddt_read_start(zio_t *zio) |
2853 | { | |
2854 | blkptr_t *bp = zio->io_bp; | |
2855 | ||
2856 | ASSERT(BP_GET_DEDUP(bp)); | |
2857 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
2858 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
2859 | ||
2860 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
2861 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
2862 | ddt_entry_t *dde = ddt_repair_start(ddt, bp); | |
2863 | ddt_phys_t *ddp = dde->dde_phys; | |
2864 | ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); | |
2865 | blkptr_t blk; | |
2866 | ||
2867 | ASSERT(zio->io_vsd == NULL); | |
2868 | zio->io_vsd = dde; | |
2869 | ||
2870 | if (ddp_self == NULL) | |
62840030 | 2871 | return (zio); |
428870ff | 2872 | |
1c27024e | 2873 | for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { |
428870ff BB |
2874 | if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) |
2875 | continue; | |
2876 | ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, | |
2877 | &blk); | |
2878 | zio_nowait(zio_read(zio, zio->io_spa, &blk, | |
a6255b7f DQ |
2879 | abd_alloc_for_io(zio->io_size, B_TRUE), |
2880 | zio->io_size, zio_ddt_child_read_done, dde, | |
2881 | zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) | | |
2882 | ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark)); | |
428870ff | 2883 | } |
62840030 | 2884 | return (zio); |
428870ff BB |
2885 | } |
2886 | ||
2887 | zio_nowait(zio_read(zio, zio->io_spa, bp, | |
a6255b7f | 2888 | zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority, |
428870ff BB |
2889 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); |
2890 | ||
62840030 | 2891 | return (zio); |
428870ff BB |
2892 | } |
2893 | ||
62840030 | 2894 | static zio_t * |
428870ff BB |
2895 | zio_ddt_read_done(zio_t *zio) |
2896 | { | |
2897 | blkptr_t *bp = zio->io_bp; | |
2898 | ||
ddc751d5 | 2899 | if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) { |
62840030 | 2900 | return (NULL); |
ddc751d5 | 2901 | } |
428870ff BB |
2902 | |
2903 | ASSERT(BP_GET_DEDUP(bp)); | |
2904 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
2905 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
2906 | ||
2907 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
2908 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
2909 | ddt_entry_t *dde = zio->io_vsd; | |
2910 | if (ddt == NULL) { | |
2911 | ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); | |
62840030 | 2912 | return (zio); |
428870ff BB |
2913 | } |
2914 | if (dde == NULL) { | |
2915 | zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; | |
2916 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); | |
62840030 | 2917 | return (NULL); |
428870ff | 2918 | } |
a6255b7f DQ |
2919 | if (dde->dde_repair_abd != NULL) { |
2920 | abd_copy(zio->io_abd, dde->dde_repair_abd, | |
2921 | zio->io_size); | |
428870ff BB |
2922 | zio->io_child_error[ZIO_CHILD_DDT] = 0; |
2923 | } | |
2924 | ddt_repair_done(ddt, dde); | |
2925 | zio->io_vsd = NULL; | |
2926 | } | |
2927 | ||
2928 | ASSERT(zio->io_vsd == NULL); | |
2929 | ||
62840030 | 2930 | return (zio); |
428870ff BB |
2931 | } |
2932 | ||
2933 | static boolean_t | |
2934 | zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) | |
2935 | { | |
2936 | spa_t *spa = zio->io_spa; | |
c17bcf83 | 2937 | boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW); |
428870ff | 2938 | |
c17bcf83 | 2939 | ASSERT(!(zio->io_bp_override && do_raw)); |
2aa34383 | 2940 | |
428870ff BB |
2941 | /* |
2942 | * Note: we compare the original data, not the transformed data, | |
2943 | * because when zio->io_bp is an override bp, we will not have | |
2944 | * pushed the I/O transforms. That's an important optimization | |
2945 | * because otherwise we'd compress/encrypt all dmu_sync() data twice. | |
c17bcf83 | 2946 | * However, we should never get a raw, override zio so in these |
b5256303 | 2947 | * cases we can compare the io_abd directly. This is useful because |
c17bcf83 TC |
2948 | * it allows us to do dedup verification even if we don't have access |
2949 | * to the original data (for instance, if the encryption keys aren't | |
2950 | * loaded). | |
428870ff | 2951 | */ |
c17bcf83 | 2952 | |
1c27024e | 2953 | for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
2954 | zio_t *lio = dde->dde_lead_zio[p]; |
2955 | ||
c17bcf83 TC |
2956 | if (lio != NULL && do_raw) { |
2957 | return (lio->io_size != zio->io_size || | |
a6255b7f | 2958 | abd_cmp(zio->io_abd, lio->io_abd) != 0); |
c17bcf83 | 2959 | } else if (lio != NULL) { |
428870ff | 2960 | return (lio->io_orig_size != zio->io_orig_size || |
a6255b7f | 2961 | abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0); |
428870ff BB |
2962 | } |
2963 | } | |
2964 | ||
1c27024e | 2965 | for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
2966 | ddt_phys_t *ddp = &dde->dde_phys[p]; |
2967 | ||
c17bcf83 TC |
2968 | if (ddp->ddp_phys_birth != 0 && do_raw) { |
2969 | blkptr_t blk = *zio->io_bp; | |
2970 | uint64_t psize; | |
a6255b7f | 2971 | abd_t *tmpabd; |
c17bcf83 TC |
2972 | int error; |
2973 | ||
2974 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
2975 | psize = BP_GET_PSIZE(&blk); | |
2976 | ||
2977 | if (psize != zio->io_size) | |
2978 | return (B_TRUE); | |
2979 | ||
2980 | ddt_exit(ddt); | |
2981 | ||
a6255b7f | 2982 | tmpabd = abd_alloc_for_io(psize, B_TRUE); |
c17bcf83 | 2983 | |
a6255b7f | 2984 | error = zio_wait(zio_read(NULL, spa, &blk, tmpabd, |
c17bcf83 TC |
2985 | psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ, |
2986 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
2987 | ZIO_FLAG_RAW, &zio->io_bookmark)); | |
2988 | ||
2989 | if (error == 0) { | |
a6255b7f | 2990 | if (abd_cmp(tmpabd, zio->io_abd) != 0) |
c17bcf83 TC |
2991 | error = SET_ERROR(ENOENT); |
2992 | } | |
2993 | ||
a6255b7f | 2994 | abd_free(tmpabd); |
c17bcf83 TC |
2995 | ddt_enter(ddt); |
2996 | return (error != 0); | |
2997 | } else if (ddp->ddp_phys_birth != 0) { | |
428870ff | 2998 | arc_buf_t *abuf = NULL; |
2a432414 | 2999 | arc_flags_t aflags = ARC_FLAG_WAIT; |
428870ff BB |
3000 | blkptr_t blk = *zio->io_bp; |
3001 | int error; | |
3002 | ||
3003 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
3004 | ||
c17bcf83 TC |
3005 | if (BP_GET_LSIZE(&blk) != zio->io_orig_size) |
3006 | return (B_TRUE); | |
3007 | ||
428870ff BB |
3008 | ddt_exit(ddt); |
3009 | ||
294f6806 | 3010 | error = arc_read(NULL, spa, &blk, |
428870ff BB |
3011 | arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, |
3012 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
3013 | &aflags, &zio->io_bookmark); | |
3014 | ||
3015 | if (error == 0) { | |
a6255b7f | 3016 | if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data, |
428870ff | 3017 | zio->io_orig_size) != 0) |
c17bcf83 | 3018 | error = SET_ERROR(ENOENT); |
d3c2ae1c | 3019 | arc_buf_destroy(abuf, &abuf); |
428870ff BB |
3020 | } |
3021 | ||
3022 | ddt_enter(ddt); | |
3023 | return (error != 0); | |
3024 | } | |
3025 | } | |
3026 | ||
3027 | return (B_FALSE); | |
3028 | } | |
3029 | ||
3030 | static void | |
3031 | zio_ddt_child_write_ready(zio_t *zio) | |
3032 | { | |
3033 | int p = zio->io_prop.zp_copies; | |
3034 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
3035 | ddt_entry_t *dde = zio->io_private; | |
3036 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
3037 | zio_t *pio; | |
3038 | ||
3039 | if (zio->io_error) | |
3040 | return; | |
3041 | ||
3042 | ddt_enter(ddt); | |
3043 | ||
3044 | ASSERT(dde->dde_lead_zio[p] == zio); | |
3045 | ||
3046 | ddt_phys_fill(ddp, zio->io_bp); | |
3047 | ||
1c27024e | 3048 | zio_link_t *zl = NULL; |
3dfb57a3 | 3049 | while ((pio = zio_walk_parents(zio, &zl)) != NULL) |
428870ff BB |
3050 | ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); |
3051 | ||
3052 | ddt_exit(ddt); | |
3053 | } | |
3054 | ||
3055 | static void | |
3056 | zio_ddt_child_write_done(zio_t *zio) | |
3057 | { | |
3058 | int p = zio->io_prop.zp_copies; | |
3059 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
3060 | ddt_entry_t *dde = zio->io_private; | |
3061 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
3062 | ||
3063 | ddt_enter(ddt); | |
3064 | ||
3065 | ASSERT(ddp->ddp_refcnt == 0); | |
3066 | ASSERT(dde->dde_lead_zio[p] == zio); | |
3067 | dde->dde_lead_zio[p] = NULL; | |
3068 | ||
3069 | if (zio->io_error == 0) { | |
3dfb57a3 DB |
3070 | zio_link_t *zl = NULL; |
3071 | while (zio_walk_parents(zio, &zl) != NULL) | |
428870ff BB |
3072 | ddt_phys_addref(ddp); |
3073 | } else { | |
3074 | ddt_phys_clear(ddp); | |
3075 | } | |
3076 | ||
3077 | ddt_exit(ddt); | |
3078 | } | |
3079 | ||
3080 | static void | |
3081 | zio_ddt_ditto_write_done(zio_t *zio) | |
3082 | { | |
3083 | int p = DDT_PHYS_DITTO; | |
1c27024e | 3084 | ASSERTV(zio_prop_t *zp = &zio->io_prop); |
428870ff BB |
3085 | blkptr_t *bp = zio->io_bp; |
3086 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
3087 | ddt_entry_t *dde = zio->io_private; | |
3088 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
3089 | ddt_key_t *ddk = &dde->dde_key; | |
3090 | ||
3091 | ddt_enter(ddt); | |
3092 | ||
3093 | ASSERT(ddp->ddp_refcnt == 0); | |
3094 | ASSERT(dde->dde_lead_zio[p] == zio); | |
3095 | dde->dde_lead_zio[p] = NULL; | |
3096 | ||
3097 | if (zio->io_error == 0) { | |
3098 | ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); | |
3099 | ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); | |
3100 | ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); | |
3101 | if (ddp->ddp_phys_birth != 0) | |
3102 | ddt_phys_free(ddt, ddk, ddp, zio->io_txg); | |
3103 | ddt_phys_fill(ddp, bp); | |
3104 | } | |
3105 | ||
3106 | ddt_exit(ddt); | |
3107 | } | |
3108 | ||
62840030 | 3109 | static zio_t * |
428870ff BB |
3110 | zio_ddt_write(zio_t *zio) |
3111 | { | |
3112 | spa_t *spa = zio->io_spa; | |
3113 | blkptr_t *bp = zio->io_bp; | |
3114 | uint64_t txg = zio->io_txg; | |
3115 | zio_prop_t *zp = &zio->io_prop; | |
3116 | int p = zp->zp_copies; | |
3117 | int ditto_copies; | |
3118 | zio_t *cio = NULL; | |
3119 | zio_t *dio = NULL; | |
3120 | ddt_t *ddt = ddt_select(spa, bp); | |
3121 | ddt_entry_t *dde; | |
3122 | ddt_phys_t *ddp; | |
3123 | ||
3124 | ASSERT(BP_GET_DEDUP(bp)); | |
3125 | ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); | |
3126 | ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); | |
c17bcf83 | 3127 | ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW))); |
428870ff BB |
3128 | |
3129 | ddt_enter(ddt); | |
3130 | dde = ddt_lookup(ddt, bp, B_TRUE); | |
3131 | ddp = &dde->dde_phys[p]; | |
3132 | ||
3133 | if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { | |
3134 | /* | |
3135 | * If we're using a weak checksum, upgrade to a strong checksum | |
3136 | * and try again. If we're already using a strong checksum, | |
3137 | * we can't resolve it, so just convert to an ordinary write. | |
3138 | * (And automatically e-mail a paper to Nature?) | |
3139 | */ | |
3c67d83a TH |
3140 | if (!(zio_checksum_table[zp->zp_checksum].ci_flags & |
3141 | ZCHECKSUM_FLAG_DEDUP)) { | |
428870ff BB |
3142 | zp->zp_checksum = spa_dedup_checksum(spa); |
3143 | zio_pop_transforms(zio); | |
3144 | zio->io_stage = ZIO_STAGE_OPEN; | |
3145 | BP_ZERO(bp); | |
3146 | } else { | |
03c6040b | 3147 | zp->zp_dedup = B_FALSE; |
428870ff BB |
3148 | } |
3149 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
3150 | ddt_exit(ddt); | |
62840030 | 3151 | return (zio); |
428870ff BB |
3152 | } |
3153 | ||
3154 | ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); | |
3155 | ASSERT(ditto_copies < SPA_DVAS_PER_BP); | |
3156 | ||
3157 | if (ditto_copies > ddt_ditto_copies_present(dde) && | |
3158 | dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { | |
3159 | zio_prop_t czp = *zp; | |
3160 | ||
3161 | czp.zp_copies = ditto_copies; | |
3162 | ||
3163 | /* | |
3164 | * If we arrived here with an override bp, we won't have run | |
3165 | * the transform stack, so we won't have the data we need to | |
3166 | * generate a child i/o. So, toss the override bp and restart. | |
3167 | * This is safe, because using the override bp is just an | |
3168 | * optimization; and it's rare, so the cost doesn't matter. | |
3169 | */ | |
3170 | if (zio->io_bp_override) { | |
3171 | zio_pop_transforms(zio); | |
3172 | zio->io_stage = ZIO_STAGE_OPEN; | |
3173 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
3174 | zio->io_bp_override = NULL; | |
3175 | BP_ZERO(bp); | |
3176 | ddt_exit(ddt); | |
62840030 | 3177 | return (zio); |
428870ff BB |
3178 | } |
3179 | ||
a6255b7f | 3180 | dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, |
2aa34383 | 3181 | zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL, |
bc77ba73 | 3182 | NULL, zio_ddt_ditto_write_done, dde, zio->io_priority, |
428870ff BB |
3183 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); |
3184 | ||
a6255b7f | 3185 | zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL); |
428870ff BB |
3186 | dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; |
3187 | } | |
3188 | ||
3189 | if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { | |
3190 | if (ddp->ddp_phys_birth != 0) | |
3191 | ddt_bp_fill(ddp, bp, txg); | |
3192 | if (dde->dde_lead_zio[p] != NULL) | |
3193 | zio_add_child(zio, dde->dde_lead_zio[p]); | |
3194 | else | |
3195 | ddt_phys_addref(ddp); | |
3196 | } else if (zio->io_bp_override) { | |
3197 | ASSERT(bp->blk_birth == txg); | |
3198 | ASSERT(BP_EQUAL(bp, zio->io_bp_override)); | |
3199 | ddt_phys_fill(ddp, bp); | |
3200 | ddt_phys_addref(ddp); | |
3201 | } else { | |
a6255b7f | 3202 | cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, |
2aa34383 | 3203 | zio->io_orig_size, zio->io_orig_size, zp, |
bc77ba73 | 3204 | zio_ddt_child_write_ready, NULL, NULL, |
428870ff BB |
3205 | zio_ddt_child_write_done, dde, zio->io_priority, |
3206 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); | |
3207 | ||
a6255b7f | 3208 | zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL); |
428870ff BB |
3209 | dde->dde_lead_zio[p] = cio; |
3210 | } | |
3211 | ||
3212 | ddt_exit(ddt); | |
3213 | ||
3214 | if (cio) | |
3215 | zio_nowait(cio); | |
3216 | if (dio) | |
3217 | zio_nowait(dio); | |
3218 | ||
62840030 | 3219 | return (zio); |
428870ff BB |
3220 | } |
3221 | ||
3222 | ddt_entry_t *freedde; /* for debugging */ | |
b128c09f | 3223 | |
62840030 | 3224 | static zio_t * |
428870ff BB |
3225 | zio_ddt_free(zio_t *zio) |
3226 | { | |
3227 | spa_t *spa = zio->io_spa; | |
3228 | blkptr_t *bp = zio->io_bp; | |
3229 | ddt_t *ddt = ddt_select(spa, bp); | |
3230 | ddt_entry_t *dde; | |
3231 | ddt_phys_t *ddp; | |
3232 | ||
3233 | ASSERT(BP_GET_DEDUP(bp)); | |
3234 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
3235 | ||
3236 | ddt_enter(ddt); | |
3237 | freedde = dde = ddt_lookup(ddt, bp, B_TRUE); | |
5dc6af0e BB |
3238 | if (dde) { |
3239 | ddp = ddt_phys_select(dde, bp); | |
3240 | if (ddp) | |
3241 | ddt_phys_decref(ddp); | |
3242 | } | |
428870ff BB |
3243 | ddt_exit(ddt); |
3244 | ||
62840030 | 3245 | return (zio); |
428870ff BB |
3246 | } |
3247 | ||
3248 | /* | |
3249 | * ========================================================================== | |
3250 | * Allocate and free blocks | |
3251 | * ========================================================================== | |
3252 | */ | |
3dfb57a3 DB |
3253 | |
3254 | static zio_t * | |
492f64e9 | 3255 | zio_io_to_allocate(spa_t *spa, int allocator) |
3dfb57a3 DB |
3256 | { |
3257 | zio_t *zio; | |
3258 | ||
492f64e9 | 3259 | ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator])); |
3dfb57a3 | 3260 | |
492f64e9 | 3261 | zio = avl_first(&spa->spa_alloc_trees[allocator]); |
3dfb57a3 DB |
3262 | if (zio == NULL) |
3263 | return (NULL); | |
3264 | ||
3265 | ASSERT(IO_IS_ALLOCATING(zio)); | |
3266 | ||
3267 | /* | |
3268 | * Try to place a reservation for this zio. If we're unable to | |
3269 | * reserve then we throttle. | |
3270 | */ | |
492f64e9 | 3271 | ASSERT3U(zio->io_allocator, ==, allocator); |
cc99f275 | 3272 | if (!metaslab_class_throttle_reserve(zio->io_metaslab_class, |
492f64e9 | 3273 | zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) { |
3dfb57a3 DB |
3274 | return (NULL); |
3275 | } | |
3276 | ||
492f64e9 | 3277 | avl_remove(&spa->spa_alloc_trees[allocator], zio); |
3dfb57a3 DB |
3278 | ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE); |
3279 | ||
3280 | return (zio); | |
3281 | } | |
3282 | ||
62840030 | 3283 | static zio_t * |
3dfb57a3 DB |
3284 | zio_dva_throttle(zio_t *zio) |
3285 | { | |
3286 | spa_t *spa = zio->io_spa; | |
3287 | zio_t *nio; | |
cc99f275 DB |
3288 | metaslab_class_t *mc; |
3289 | ||
3290 | /* locate an appropriate allocation class */ | |
3291 | mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type, | |
3292 | zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk); | |
3dfb57a3 DB |
3293 | |
3294 | if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE || | |
cc99f275 | 3295 | !mc->mc_alloc_throttle_enabled || |
3dfb57a3 DB |
3296 | zio->io_child_type == ZIO_CHILD_GANG || |
3297 | zio->io_flags & ZIO_FLAG_NODATA) { | |
62840030 | 3298 | return (zio); |
3dfb57a3 DB |
3299 | } |
3300 | ||
3301 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
3302 | ||
3303 | ASSERT3U(zio->io_queued_timestamp, >, 0); | |
3304 | ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE); | |
3305 | ||
492f64e9 PD |
3306 | zbookmark_phys_t *bm = &zio->io_bookmark; |
3307 | /* | |
3308 | * We want to try to use as many allocators as possible to help improve | |
3309 | * performance, but we also want logically adjacent IOs to be physically | |
3310 | * adjacent to improve sequential read performance. We chunk each object | |
3311 | * into 2^20 block regions, and then hash based on the objset, object, | |
3312 | * level, and region to accomplish both of these goals. | |
3313 | */ | |
3314 | zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object, | |
3315 | bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count; | |
3316 | mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]); | |
3dfb57a3 | 3317 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); |
cc99f275 | 3318 | zio->io_metaslab_class = mc; |
492f64e9 | 3319 | avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio); |
cc99f275 | 3320 | nio = zio_io_to_allocate(spa, zio->io_allocator); |
492f64e9 | 3321 | mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]); |
62840030 | 3322 | return (nio); |
3dfb57a3 DB |
3323 | } |
3324 | ||
cc99f275 | 3325 | static void |
492f64e9 | 3326 | zio_allocate_dispatch(spa_t *spa, int allocator) |
3dfb57a3 DB |
3327 | { |
3328 | zio_t *zio; | |
3329 | ||
492f64e9 PD |
3330 | mutex_enter(&spa->spa_alloc_locks[allocator]); |
3331 | zio = zio_io_to_allocate(spa, allocator); | |
3332 | mutex_exit(&spa->spa_alloc_locks[allocator]); | |
3dfb57a3 DB |
3333 | if (zio == NULL) |
3334 | return; | |
3335 | ||
3336 | ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE); | |
3337 | ASSERT0(zio->io_error); | |
3338 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE); | |
3339 | } | |
3340 | ||
62840030 | 3341 | static zio_t * |
34dc7c2f BB |
3342 | zio_dva_allocate(zio_t *zio) |
3343 | { | |
3344 | spa_t *spa = zio->io_spa; | |
cc99f275 | 3345 | metaslab_class_t *mc; |
34dc7c2f BB |
3346 | blkptr_t *bp = zio->io_bp; |
3347 | int error; | |
6d974228 | 3348 | int flags = 0; |
34dc7c2f | 3349 | |
9babb374 BB |
3350 | if (zio->io_gang_leader == NULL) { |
3351 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
3352 | zio->io_gang_leader = zio; | |
3353 | } | |
3354 | ||
34dc7c2f | 3355 | ASSERT(BP_IS_HOLE(bp)); |
c99c9001 | 3356 | ASSERT0(BP_GET_NDVAS(bp)); |
428870ff BB |
3357 | ASSERT3U(zio->io_prop.zp_copies, >, 0); |
3358 | ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); | |
34dc7c2f BB |
3359 | ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); |
3360 | ||
920dd524 | 3361 | flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0; |
3dfb57a3 DB |
3362 | if (zio->io_flags & ZIO_FLAG_NODATA) |
3363 | flags |= METASLAB_DONT_THROTTLE; | |
3364 | if (zio->io_flags & ZIO_FLAG_GANG_CHILD) | |
3365 | flags |= METASLAB_GANG_CHILD; | |
3366 | if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) | |
3367 | flags |= METASLAB_ASYNC_ALLOC; | |
3368 | ||
cc99f275 DB |
3369 | /* |
3370 | * if not already chosen, locate an appropriate allocation class | |
3371 | */ | |
3372 | mc = zio->io_metaslab_class; | |
3373 | if (mc == NULL) { | |
3374 | mc = spa_preferred_class(spa, zio->io_size, | |
3375 | zio->io_prop.zp_type, zio->io_prop.zp_level, | |
3376 | zio->io_prop.zp_zpl_smallblk); | |
3377 | zio->io_metaslab_class = mc; | |
3378 | } | |
3379 | ||
b128c09f | 3380 | error = metaslab_alloc(spa, mc, zio->io_size, bp, |
4e21fd06 | 3381 | zio->io_prop.zp_copies, zio->io_txg, NULL, flags, |
492f64e9 | 3382 | &zio->io_alloc_list, zio, zio->io_allocator); |
34dc7c2f | 3383 | |
cc99f275 DB |
3384 | /* |
3385 | * Fallback to normal class when an alloc class is full | |
3386 | */ | |
3387 | if (error == ENOSPC && mc != spa_normal_class(spa)) { | |
3388 | /* | |
3389 | * If throttling, transfer reservation over to normal class. | |
3390 | * The io_allocator slot can remain the same even though we | |
3391 | * are switching classes. | |
3392 | */ | |
3393 | if (mc->mc_alloc_throttle_enabled && | |
3394 | (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) { | |
3395 | metaslab_class_throttle_unreserve(mc, | |
3396 | zio->io_prop.zp_copies, zio->io_allocator, zio); | |
3397 | zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING; | |
3398 | ||
3399 | mc = spa_normal_class(spa); | |
3400 | VERIFY(metaslab_class_throttle_reserve(mc, | |
3401 | zio->io_prop.zp_copies, zio->io_allocator, zio, | |
3402 | flags | METASLAB_MUST_RESERVE)); | |
3403 | } else { | |
3404 | mc = spa_normal_class(spa); | |
3405 | } | |
3406 | zio->io_metaslab_class = mc; | |
3407 | ||
3408 | error = metaslab_alloc(spa, mc, zio->io_size, bp, | |
3409 | zio->io_prop.zp_copies, zio->io_txg, NULL, flags, | |
3410 | &zio->io_alloc_list, zio, zio->io_allocator); | |
3411 | } | |
3412 | ||
3dfb57a3 | 3413 | if (error != 0) { |
964c2d69 | 3414 | zfs_dbgmsg("%s: metaslab allocation failure: zio %p, " |
6d974228 GW |
3415 | "size %llu, error %d", spa_name(spa), zio, zio->io_size, |
3416 | error); | |
b128c09f BB |
3417 | if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) |
3418 | return (zio_write_gang_block(zio)); | |
34dc7c2f BB |
3419 | zio->io_error = error; |
3420 | } | |
3421 | ||
62840030 | 3422 | return (zio); |
34dc7c2f BB |
3423 | } |
3424 | ||
62840030 | 3425 | static zio_t * |
34dc7c2f BB |
3426 | zio_dva_free(zio_t *zio) |
3427 | { | |
b128c09f | 3428 | metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); |
34dc7c2f | 3429 | |
62840030 | 3430 | return (zio); |
34dc7c2f BB |
3431 | } |
3432 | ||
62840030 | 3433 | static zio_t * |
34dc7c2f BB |
3434 | zio_dva_claim(zio_t *zio) |
3435 | { | |
b128c09f BB |
3436 | int error; |
3437 | ||
3438 | error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); | |
3439 | if (error) | |
3440 | zio->io_error = error; | |
34dc7c2f | 3441 | |
62840030 | 3442 | return (zio); |
34dc7c2f BB |
3443 | } |
3444 | ||
b128c09f BB |
3445 | /* |
3446 | * Undo an allocation. This is used by zio_done() when an I/O fails | |
3447 | * and we want to give back the block we just allocated. | |
3448 | * This handles both normal blocks and gang blocks. | |
3449 | */ | |
3450 | static void | |
3451 | zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) | |
3452 | { | |
b128c09f | 3453 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); |
428870ff | 3454 | ASSERT(zio->io_bp_override == NULL); |
b128c09f BB |
3455 | |
3456 | if (!BP_IS_HOLE(bp)) | |
428870ff | 3457 | metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); |
b128c09f BB |
3458 | |
3459 | if (gn != NULL) { | |
1c27024e | 3460 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
3461 | zio_dva_unallocate(zio, gn->gn_child[g], |
3462 | &gn->gn_gbh->zg_blkptr[g]); | |
3463 | } | |
3464 | } | |
3465 | } | |
3466 | ||
3467 | /* | |
3468 | * Try to allocate an intent log block. Return 0 on success, errno on failure. | |
3469 | */ | |
3470 | int | |
b5256303 TC |
3471 | zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp, |
3472 | uint64_t size, boolean_t *slog) | |
b128c09f | 3473 | { |
428870ff | 3474 | int error = 1; |
4e21fd06 | 3475 | zio_alloc_list_t io_alloc_list; |
b128c09f | 3476 | |
428870ff BB |
3477 | ASSERT(txg > spa_syncing_txg(spa)); |
3478 | ||
4e21fd06 | 3479 | metaslab_trace_init(&io_alloc_list); |
cc99f275 DB |
3480 | |
3481 | /* | |
3482 | * Block pointer fields are useful to metaslabs for stats and debugging. | |
3483 | * Fill in the obvious ones before calling into metaslab_alloc(). | |
3484 | */ | |
3485 | BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); | |
3486 | BP_SET_PSIZE(new_bp, size); | |
3487 | BP_SET_LEVEL(new_bp, 0); | |
3488 | ||
492f64e9 PD |
3489 | /* |
3490 | * When allocating a zil block, we don't have information about | |
3491 | * the final destination of the block except the objset it's part | |
3492 | * of, so we just hash the objset ID to pick the allocator to get | |
3493 | * some parallelism. | |
3494 | */ | |
1b7c1e5c | 3495 | error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1, |
492f64e9 PD |
3496 | txg, NULL, METASLAB_FASTWRITE, &io_alloc_list, NULL, |
3497 | cityhash4(0, 0, 0, os->os_dsl_dataset->ds_object) % | |
3498 | spa->spa_alloc_count); | |
1b7c1e5c GDN |
3499 | if (error == 0) { |
3500 | *slog = TRUE; | |
3501 | } else { | |
428870ff | 3502 | error = metaslab_alloc(spa, spa_normal_class(spa), size, |
4e21fd06 | 3503 | new_bp, 1, txg, NULL, METASLAB_FASTWRITE, |
492f64e9 PD |
3504 | &io_alloc_list, NULL, cityhash4(0, 0, 0, |
3505 | os->os_dsl_dataset->ds_object) % spa->spa_alloc_count); | |
1b7c1e5c GDN |
3506 | if (error == 0) |
3507 | *slog = FALSE; | |
ebf8e3a2 | 3508 | } |
4e21fd06 | 3509 | metaslab_trace_fini(&io_alloc_list); |
b128c09f BB |
3510 | |
3511 | if (error == 0) { | |
3512 | BP_SET_LSIZE(new_bp, size); | |
3513 | BP_SET_PSIZE(new_bp, size); | |
3514 | BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); | |
428870ff BB |
3515 | BP_SET_CHECKSUM(new_bp, |
3516 | spa_version(spa) >= SPA_VERSION_SLIM_ZIL | |
3517 | ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); | |
b128c09f BB |
3518 | BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); |
3519 | BP_SET_LEVEL(new_bp, 0); | |
428870ff | 3520 | BP_SET_DEDUP(new_bp, 0); |
b128c09f | 3521 | BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); |
b5256303 TC |
3522 | |
3523 | /* | |
3524 | * encrypted blocks will require an IV and salt. We generate | |
3525 | * these now since we will not be rewriting the bp at | |
3526 | * rewrite time. | |
3527 | */ | |
3528 | if (os->os_encrypted) { | |
3529 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
3530 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
3531 | ||
3532 | BP_SET_CRYPT(new_bp, B_TRUE); | |
3533 | VERIFY0(spa_crypt_get_salt(spa, | |
3534 | dmu_objset_id(os), salt)); | |
3535 | VERIFY0(zio_crypt_generate_iv(iv)); | |
3536 | ||
3537 | zio_crypt_encode_params_bp(new_bp, salt, iv); | |
3538 | } | |
1ce23dca PS |
3539 | } else { |
3540 | zfs_dbgmsg("%s: zil block allocation failure: " | |
3541 | "size %llu, error %d", spa_name(spa), size, error); | |
b128c09f BB |
3542 | } |
3543 | ||
3544 | return (error); | |
3545 | } | |
3546 | ||
34dc7c2f BB |
3547 | /* |
3548 | * ========================================================================== | |
3549 | * Read and write to physical devices | |
3550 | * ========================================================================== | |
3551 | */ | |
98b25418 GW |
3552 | |
3553 | ||
3554 | /* | |
3555 | * Issue an I/O to the underlying vdev. Typically the issue pipeline | |
3556 | * stops after this stage and will resume upon I/O completion. | |
3557 | * However, there are instances where the vdev layer may need to | |
3558 | * continue the pipeline when an I/O was not issued. Since the I/O | |
3559 | * that was sent to the vdev layer might be different than the one | |
3560 | * currently active in the pipeline (see vdev_queue_io()), we explicitly | |
3561 | * force the underlying vdev layers to call either zio_execute() or | |
3562 | * zio_interrupt() to ensure that the pipeline continues with the correct I/O. | |
3563 | */ | |
62840030 | 3564 | static zio_t * |
34dc7c2f BB |
3565 | zio_vdev_io_start(zio_t *zio) |
3566 | { | |
3567 | vdev_t *vd = zio->io_vd; | |
34dc7c2f BB |
3568 | uint64_t align; |
3569 | spa_t *spa = zio->io_spa; | |
3570 | ||
193a37cb TH |
3571 | zio->io_delay = 0; |
3572 | ||
b128c09f BB |
3573 | ASSERT(zio->io_error == 0); |
3574 | ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); | |
34dc7c2f | 3575 | |
b128c09f BB |
3576 | if (vd == NULL) { |
3577 | if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
3578 | spa_config_enter(spa, SCL_ZIO, zio, RW_READER); | |
34dc7c2f | 3579 | |
b128c09f BB |
3580 | /* |
3581 | * The mirror_ops handle multiple DVAs in a single BP. | |
3582 | */ | |
98b25418 | 3583 | vdev_mirror_ops.vdev_op_io_start(zio); |
62840030 | 3584 | return (NULL); |
34dc7c2f BB |
3585 | } |
3586 | ||
3dfb57a3 | 3587 | ASSERT3P(zio->io_logical, !=, zio); |
6cb8e530 PZ |
3588 | if (zio->io_type == ZIO_TYPE_WRITE) { |
3589 | ASSERT(spa->spa_trust_config); | |
3590 | ||
a1d477c2 MA |
3591 | /* |
3592 | * Note: the code can handle other kinds of writes, | |
3593 | * but we don't expect them. | |
3594 | */ | |
6cb8e530 PZ |
3595 | if (zio->io_vd->vdev_removing) { |
3596 | ASSERT(zio->io_flags & | |
3597 | (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL | | |
3598 | ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE)); | |
3599 | } | |
a1d477c2 | 3600 | } |
3dfb57a3 | 3601 | |
b128c09f BB |
3602 | align = 1ULL << vd->vdev_top->vdev_ashift; |
3603 | ||
b02fe35d AR |
3604 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && |
3605 | P2PHASE(zio->io_size, align) != 0) { | |
3606 | /* Transform logical writes to be a full physical block size. */ | |
34dc7c2f | 3607 | uint64_t asize = P2ROUNDUP(zio->io_size, align); |
a6255b7f | 3608 | abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize); |
178e73b3 | 3609 | ASSERT(vd == vd->vdev_top); |
34dc7c2f | 3610 | if (zio->io_type == ZIO_TYPE_WRITE) { |
a6255b7f DQ |
3611 | abd_copy(abuf, zio->io_abd, zio->io_size); |
3612 | abd_zero_off(abuf, zio->io_size, asize - zio->io_size); | |
34dc7c2f | 3613 | } |
b128c09f | 3614 | zio_push_transform(zio, abuf, asize, asize, zio_subblock); |
34dc7c2f BB |
3615 | } |
3616 | ||
b02fe35d AR |
3617 | /* |
3618 | * If this is not a physical io, make sure that it is properly aligned | |
3619 | * before proceeding. | |
3620 | */ | |
3621 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { | |
3622 | ASSERT0(P2PHASE(zio->io_offset, align)); | |
3623 | ASSERT0(P2PHASE(zio->io_size, align)); | |
3624 | } else { | |
3625 | /* | |
3626 | * For physical writes, we allow 512b aligned writes and assume | |
3627 | * the device will perform a read-modify-write as necessary. | |
3628 | */ | |
3629 | ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); | |
3630 | ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); | |
3631 | } | |
3632 | ||
572e2857 | 3633 | VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); |
fb5f0bc8 BB |
3634 | |
3635 | /* | |
3636 | * If this is a repair I/O, and there's no self-healing involved -- | |
3637 | * that is, we're just resilvering what we expect to resilver -- | |
3638 | * then don't do the I/O unless zio's txg is actually in vd's DTL. | |
9e052db4 MA |
3639 | * This prevents spurious resilvering. |
3640 | * | |
3641 | * There are a few ways that we can end up creating these spurious | |
3642 | * resilver i/os: | |
3643 | * | |
3644 | * 1. A resilver i/o will be issued if any DVA in the BP has a | |
3645 | * dirty DTL. The mirror code will issue resilver writes to | |
3646 | * each DVA, including the one(s) that are not on vdevs with dirty | |
3647 | * DTLs. | |
3648 | * | |
3649 | * 2. With nested replication, which happens when we have a | |
3650 | * "replacing" or "spare" vdev that's a child of a mirror or raidz. | |
3651 | * For example, given mirror(replacing(A+B), C), it's likely that | |
3652 | * only A is out of date (it's the new device). In this case, we'll | |
3653 | * read from C, then use the data to resilver A+B -- but we don't | |
3654 | * actually want to resilver B, just A. The top-level mirror has no | |
3655 | * way to know this, so instead we just discard unnecessary repairs | |
3656 | * as we work our way down the vdev tree. | |
3657 | * | |
3658 | * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc. | |
3659 | * The same logic applies to any form of nested replication: ditto | |
3660 | * + mirror, RAID-Z + replacing, etc. | |
3661 | * | |
3662 | * However, indirect vdevs point off to other vdevs which may have | |
3663 | * DTL's, so we never bypass them. The child i/os on concrete vdevs | |
3664 | * will be properly bypassed instead. | |
fb5f0bc8 BB |
3665 | */ |
3666 | if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && | |
3667 | !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && | |
3668 | zio->io_txg != 0 && /* not a delegated i/o */ | |
9e052db4 | 3669 | vd->vdev_ops != &vdev_indirect_ops && |
fb5f0bc8 BB |
3670 | !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { |
3671 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
fb5f0bc8 | 3672 | zio_vdev_io_bypass(zio); |
62840030 | 3673 | return (zio); |
fb5f0bc8 | 3674 | } |
34dc7c2f | 3675 | |
b128c09f BB |
3676 | if (vd->vdev_ops->vdev_op_leaf && |
3677 | (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) { | |
3678 | ||
b0bc7a84 | 3679 | if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio)) |
62840030 | 3680 | return (zio); |
b128c09f BB |
3681 | |
3682 | if ((zio = vdev_queue_io(zio)) == NULL) | |
62840030 | 3683 | return (NULL); |
b128c09f BB |
3684 | |
3685 | if (!vdev_accessible(vd, zio)) { | |
2e528b49 | 3686 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f | 3687 | zio_interrupt(zio); |
62840030 | 3688 | return (NULL); |
b128c09f | 3689 | } |
67103816 | 3690 | zio->io_delay = gethrtime(); |
b128c09f BB |
3691 | } |
3692 | ||
98b25418 | 3693 | vd->vdev_ops->vdev_op_io_start(zio); |
62840030 | 3694 | return (NULL); |
34dc7c2f BB |
3695 | } |
3696 | ||
62840030 | 3697 | static zio_t * |
34dc7c2f BB |
3698 | zio_vdev_io_done(zio_t *zio) |
3699 | { | |
b128c09f BB |
3700 | vdev_t *vd = zio->io_vd; |
3701 | vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; | |
3702 | boolean_t unexpected_error = B_FALSE; | |
34dc7c2f | 3703 | |
ddc751d5 | 3704 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { |
62840030 | 3705 | return (NULL); |
ddc751d5 | 3706 | } |
34dc7c2f | 3707 | |
b128c09f BB |
3708 | ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); |
3709 | ||
193a37cb TH |
3710 | if (zio->io_delay) |
3711 | zio->io_delay = gethrtime() - zio->io_delay; | |
3712 | ||
b128c09f BB |
3713 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { |
3714 | ||
3715 | vdev_queue_io_done(zio); | |
3716 | ||
3717 | if (zio->io_type == ZIO_TYPE_WRITE) | |
3718 | vdev_cache_write(zio); | |
3719 | ||
3720 | if (zio_injection_enabled && zio->io_error == 0) | |
d977122d DB |
3721 | zio->io_error = zio_handle_device_injections(vd, zio, |
3722 | EIO, EILSEQ); | |
b128c09f BB |
3723 | |
3724 | if (zio_injection_enabled && zio->io_error == 0) | |
3725 | zio->io_error = zio_handle_label_injection(zio, EIO); | |
3726 | ||
3727 | if (zio->io_error) { | |
3728 | if (!vdev_accessible(vd, zio)) { | |
2e528b49 | 3729 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
3730 | } else { |
3731 | unexpected_error = B_TRUE; | |
3732 | } | |
3733 | } | |
3734 | } | |
3735 | ||
3736 | ops->vdev_op_io_done(zio); | |
34dc7c2f | 3737 | |
f43615d0 | 3738 | if (unexpected_error) |
d164b209 | 3739 | VERIFY(vdev_probe(vd, zio) == NULL); |
34dc7c2f | 3740 | |
62840030 | 3741 | return (zio); |
34dc7c2f BB |
3742 | } |
3743 | ||
a8b2e306 TC |
3744 | /* |
3745 | * This function is used to change the priority of an existing zio that is | |
3746 | * currently in-flight. This is used by the arc to upgrade priority in the | |
3747 | * event that a demand read is made for a block that is currently queued | |
3748 | * as a scrub or async read IO. Otherwise, the high priority read request | |
3749 | * would end up having to wait for the lower priority IO. | |
3750 | */ | |
3751 | void | |
3752 | zio_change_priority(zio_t *pio, zio_priority_t priority) | |
3753 | { | |
3754 | zio_t *cio, *cio_next; | |
3755 | zio_link_t *zl = NULL; | |
3756 | ||
3757 | ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
3758 | ||
3759 | if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) { | |
3760 | vdev_queue_change_io_priority(pio, priority); | |
3761 | } else { | |
3762 | pio->io_priority = priority; | |
3763 | } | |
3764 | ||
3765 | mutex_enter(&pio->io_lock); | |
3766 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { | |
3767 | cio_next = zio_walk_children(pio, &zl); | |
3768 | zio_change_priority(cio, priority); | |
3769 | } | |
3770 | mutex_exit(&pio->io_lock); | |
3771 | } | |
3772 | ||
428870ff BB |
3773 | /* |
3774 | * For non-raidz ZIOs, we can just copy aside the bad data read from the | |
3775 | * disk, and use that to finish the checksum ereport later. | |
3776 | */ | |
3777 | static void | |
3778 | zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, | |
84c07ada | 3779 | const abd_t *good_buf) |
428870ff BB |
3780 | { |
3781 | /* no processing needed */ | |
3782 | zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); | |
3783 | } | |
3784 | ||
3785 | /*ARGSUSED*/ | |
3786 | void | |
3787 | zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) | |
3788 | { | |
84c07ada | 3789 | void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size); |
428870ff | 3790 | |
84c07ada | 3791 | abd_copy(abd, zio->io_abd, zio->io_size); |
428870ff BB |
3792 | |
3793 | zcr->zcr_cbinfo = zio->io_size; | |
84c07ada | 3794 | zcr->zcr_cbdata = abd; |
428870ff | 3795 | zcr->zcr_finish = zio_vsd_default_cksum_finish; |
84c07ada | 3796 | zcr->zcr_free = zio_abd_free; |
428870ff BB |
3797 | } |
3798 | ||
62840030 | 3799 | static zio_t * |
34dc7c2f BB |
3800 | zio_vdev_io_assess(zio_t *zio) |
3801 | { | |
3802 | vdev_t *vd = zio->io_vd; | |
b128c09f | 3803 | |
ddc751d5 | 3804 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { |
62840030 | 3805 | return (NULL); |
ddc751d5 | 3806 | } |
b128c09f BB |
3807 | |
3808 | if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
3809 | spa_config_exit(zio->io_spa, SCL_ZIO, zio); | |
3810 | ||
3811 | if (zio->io_vsd != NULL) { | |
428870ff | 3812 | zio->io_vsd_ops->vsd_free(zio); |
b128c09f | 3813 | zio->io_vsd = NULL; |
34dc7c2f BB |
3814 | } |
3815 | ||
b128c09f | 3816 | if (zio_injection_enabled && zio->io_error == 0) |
34dc7c2f BB |
3817 | zio->io_error = zio_handle_fault_injection(zio, EIO); |
3818 | ||
3819 | /* | |
3820 | * If the I/O failed, determine whether we should attempt to retry it. | |
428870ff BB |
3821 | * |
3822 | * On retry, we cut in line in the issue queue, since we don't want | |
3823 | * compression/checksumming/etc. work to prevent our (cheap) IO reissue. | |
34dc7c2f | 3824 | */ |
b128c09f BB |
3825 | if (zio->io_error && vd == NULL && |
3826 | !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { | |
3827 | ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ | |
3828 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ | |
34dc7c2f | 3829 | zio->io_error = 0; |
b128c09f BB |
3830 | zio->io_flags |= ZIO_FLAG_IO_RETRY | |
3831 | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; | |
428870ff BB |
3832 | zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; |
3833 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, | |
3834 | zio_requeue_io_start_cut_in_line); | |
62840030 | 3835 | return (NULL); |
34dc7c2f BB |
3836 | } |
3837 | ||
b128c09f BB |
3838 | /* |
3839 | * If we got an error on a leaf device, convert it to ENXIO | |
3840 | * if the device is not accessible at all. | |
3841 | */ | |
3842 | if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && | |
3843 | !vdev_accessible(vd, zio)) | |
2e528b49 | 3844 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
3845 | |
3846 | /* | |
3847 | * If we can't write to an interior vdev (mirror or RAID-Z), | |
3848 | * set vdev_cant_write so that we stop trying to allocate from it. | |
3849 | */ | |
3850 | if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && | |
13fe0198 | 3851 | vd != NULL && !vd->vdev_ops->vdev_op_leaf) { |
b128c09f | 3852 | vd->vdev_cant_write = B_TRUE; |
13fe0198 | 3853 | } |
b128c09f | 3854 | |
298ec40b GM |
3855 | /* |
3856 | * If a cache flush returns ENOTSUP or ENOTTY, we know that no future | |
3857 | * attempts will ever succeed. In this case we set a persistent bit so | |
3858 | * that we don't bother with it in the future. | |
3859 | */ | |
3860 | if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) && | |
3861 | zio->io_type == ZIO_TYPE_IOCTL && | |
3862 | zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL) | |
3863 | vd->vdev_nowritecache = B_TRUE; | |
3864 | ||
b128c09f BB |
3865 | if (zio->io_error) |
3866 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
3867 | ||
e8b96c60 MA |
3868 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf && |
3869 | zio->io_physdone != NULL) { | |
3870 | ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); | |
3871 | ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); | |
3872 | zio->io_physdone(zio->io_logical); | |
3873 | } | |
3874 | ||
62840030 | 3875 | return (zio); |
34dc7c2f BB |
3876 | } |
3877 | ||
3878 | void | |
3879 | zio_vdev_io_reissue(zio_t *zio) | |
3880 | { | |
3881 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
3882 | ASSERT(zio->io_error == 0); | |
3883 | ||
428870ff | 3884 | zio->io_stage >>= 1; |
34dc7c2f BB |
3885 | } |
3886 | ||
3887 | void | |
3888 | zio_vdev_io_redone(zio_t *zio) | |
3889 | { | |
3890 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); | |
3891 | ||
428870ff | 3892 | zio->io_stage >>= 1; |
34dc7c2f BB |
3893 | } |
3894 | ||
3895 | void | |
3896 | zio_vdev_io_bypass(zio_t *zio) | |
3897 | { | |
3898 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
3899 | ASSERT(zio->io_error == 0); | |
3900 | ||
3901 | zio->io_flags |= ZIO_FLAG_IO_BYPASS; | |
428870ff | 3902 | zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; |
34dc7c2f BB |
3903 | } |
3904 | ||
b5256303 TC |
3905 | /* |
3906 | * ========================================================================== | |
3907 | * Encrypt and store encryption parameters | |
3908 | * ========================================================================== | |
3909 | */ | |
3910 | ||
3911 | ||
3912 | /* | |
3913 | * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for | |
3914 | * managing the storage of encryption parameters and passing them to the | |
3915 | * lower-level encryption functions. | |
3916 | */ | |
62840030 | 3917 | static zio_t * |
b5256303 TC |
3918 | zio_encrypt(zio_t *zio) |
3919 | { | |
3920 | zio_prop_t *zp = &zio->io_prop; | |
3921 | spa_t *spa = zio->io_spa; | |
3922 | blkptr_t *bp = zio->io_bp; | |
3923 | uint64_t psize = BP_GET_PSIZE(bp); | |
ae76f45c | 3924 | uint64_t dsobj = zio->io_bookmark.zb_objset; |
b5256303 TC |
3925 | dmu_object_type_t ot = BP_GET_TYPE(bp); |
3926 | void *enc_buf = NULL; | |
3927 | abd_t *eabd = NULL; | |
3928 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
3929 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
3930 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
3931 | boolean_t no_crypt = B_FALSE; | |
3932 | ||
3933 | /* the root zio already encrypted the data */ | |
3934 | if (zio->io_child_type == ZIO_CHILD_GANG) | |
62840030 | 3935 | return (zio); |
b5256303 TC |
3936 | |
3937 | /* only ZIL blocks are re-encrypted on rewrite */ | |
3938 | if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG) | |
62840030 | 3939 | return (zio); |
b5256303 TC |
3940 | |
3941 | if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) { | |
3942 | BP_SET_CRYPT(bp, B_FALSE); | |
62840030 | 3943 | return (zio); |
b5256303 TC |
3944 | } |
3945 | ||
3946 | /* if we are doing raw encryption set the provided encryption params */ | |
3947 | if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) { | |
ae76f45c | 3948 | ASSERT0(BP_GET_LEVEL(bp)); |
b5256303 TC |
3949 | BP_SET_CRYPT(bp, B_TRUE); |
3950 | BP_SET_BYTEORDER(bp, zp->zp_byteorder); | |
3951 | if (ot != DMU_OT_OBJSET) | |
3952 | zio_crypt_encode_mac_bp(bp, zp->zp_mac); | |
ae76f45c TC |
3953 | |
3954 | /* dnode blocks must be written out in the provided byteorder */ | |
3955 | if (zp->zp_byteorder != ZFS_HOST_BYTEORDER && | |
3956 | ot == DMU_OT_DNODE) { | |
3957 | void *bswap_buf = zio_buf_alloc(psize); | |
3958 | abd_t *babd = abd_get_from_buf(bswap_buf, psize); | |
3959 | ||
3960 | ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); | |
3961 | abd_copy_to_buf(bswap_buf, zio->io_abd, psize); | |
3962 | dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf, | |
3963 | psize); | |
3964 | ||
3965 | abd_take_ownership_of_buf(babd, B_TRUE); | |
3966 | zio_push_transform(zio, babd, psize, psize, NULL); | |
3967 | } | |
3968 | ||
b5256303 TC |
3969 | if (DMU_OT_IS_ENCRYPTED(ot)) |
3970 | zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv); | |
62840030 | 3971 | return (zio); |
b5256303 TC |
3972 | } |
3973 | ||
3974 | /* indirect blocks only maintain a cksum of the lower level MACs */ | |
3975 | if (BP_GET_LEVEL(bp) > 0) { | |
3976 | BP_SET_CRYPT(bp, B_TRUE); | |
3977 | VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE, | |
3978 | zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp), | |
3979 | mac)); | |
3980 | zio_crypt_encode_mac_bp(bp, mac); | |
62840030 | 3981 | return (zio); |
b5256303 TC |
3982 | } |
3983 | ||
3984 | /* | |
3985 | * Objset blocks are a special case since they have 2 256-bit MACs | |
3986 | * embedded within them. | |
3987 | */ | |
3988 | if (ot == DMU_OT_OBJSET) { | |
3989 | ASSERT0(DMU_OT_IS_ENCRYPTED(ot)); | |
3990 | ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); | |
3991 | BP_SET_CRYPT(bp, B_TRUE); | |
ae76f45c TC |
3992 | VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj, |
3993 | zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp))); | |
62840030 | 3994 | return (zio); |
b5256303 TC |
3995 | } |
3996 | ||
3997 | /* unencrypted object types are only authenticated with a MAC */ | |
3998 | if (!DMU_OT_IS_ENCRYPTED(ot)) { | |
3999 | BP_SET_CRYPT(bp, B_TRUE); | |
ae76f45c TC |
4000 | VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj, |
4001 | zio->io_abd, psize, mac)); | |
b5256303 | 4002 | zio_crypt_encode_mac_bp(bp, mac); |
62840030 | 4003 | return (zio); |
b5256303 TC |
4004 | } |
4005 | ||
4006 | /* | |
4007 | * Later passes of sync-to-convergence may decide to rewrite data | |
4008 | * in place to avoid more disk reallocations. This presents a problem | |
4009 | * for encryption because this consitutes rewriting the new data with | |
4010 | * the same encryption key and IV. However, this only applies to blocks | |
4011 | * in the MOS (particularly the spacemaps) and we do not encrypt the | |
4012 | * MOS. We assert that the zio is allocating or an intent log write | |
4013 | * to enforce this. | |
4014 | */ | |
4015 | ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG); | |
4016 | ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG); | |
4017 | ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION)); | |
4018 | ASSERT3U(psize, !=, 0); | |
4019 | ||
4020 | enc_buf = zio_buf_alloc(psize); | |
4021 | eabd = abd_get_from_buf(enc_buf, psize); | |
4022 | abd_take_ownership_of_buf(eabd, B_TRUE); | |
4023 | ||
4024 | /* | |
4025 | * For an explanation of what encryption parameters are stored | |
4026 | * where, see the block comment in zio_crypt.c. | |
4027 | */ | |
4028 | if (ot == DMU_OT_INTENT_LOG) { | |
4029 | zio_crypt_decode_params_bp(bp, salt, iv); | |
4030 | } else { | |
4031 | BP_SET_CRYPT(bp, B_TRUE); | |
4032 | } | |
4033 | ||
4034 | /* Perform the encryption. This should not fail */ | |
be9a5c35 TC |
4035 | VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark, |
4036 | BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), | |
4037 | salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt)); | |
b5256303 TC |
4038 | |
4039 | /* encode encryption metadata into the bp */ | |
4040 | if (ot == DMU_OT_INTENT_LOG) { | |
4041 | /* | |
4042 | * ZIL blocks store the MAC in the embedded checksum, so the | |
4043 | * transform must always be applied. | |
4044 | */ | |
4045 | zio_crypt_encode_mac_zil(enc_buf, mac); | |
4046 | zio_push_transform(zio, eabd, psize, psize, NULL); | |
4047 | } else { | |
4048 | BP_SET_CRYPT(bp, B_TRUE); | |
4049 | zio_crypt_encode_params_bp(bp, salt, iv); | |
4050 | zio_crypt_encode_mac_bp(bp, mac); | |
4051 | ||
4052 | if (no_crypt) { | |
4053 | ASSERT3U(ot, ==, DMU_OT_DNODE); | |
4054 | abd_free(eabd); | |
4055 | } else { | |
4056 | zio_push_transform(zio, eabd, psize, psize, NULL); | |
4057 | } | |
4058 | } | |
4059 | ||
62840030 | 4060 | return (zio); |
b5256303 TC |
4061 | } |
4062 | ||
34dc7c2f BB |
4063 | /* |
4064 | * ========================================================================== | |
4065 | * Generate and verify checksums | |
4066 | * ========================================================================== | |
4067 | */ | |
62840030 | 4068 | static zio_t * |
34dc7c2f BB |
4069 | zio_checksum_generate(zio_t *zio) |
4070 | { | |
34dc7c2f | 4071 | blkptr_t *bp = zio->io_bp; |
b128c09f | 4072 | enum zio_checksum checksum; |
34dc7c2f | 4073 | |
b128c09f BB |
4074 | if (bp == NULL) { |
4075 | /* | |
4076 | * This is zio_write_phys(). | |
4077 | * We're either generating a label checksum, or none at all. | |
4078 | */ | |
4079 | checksum = zio->io_prop.zp_checksum; | |
34dc7c2f | 4080 | |
b128c09f | 4081 | if (checksum == ZIO_CHECKSUM_OFF) |
62840030 | 4082 | return (zio); |
b128c09f BB |
4083 | |
4084 | ASSERT(checksum == ZIO_CHECKSUM_LABEL); | |
4085 | } else { | |
4086 | if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { | |
4087 | ASSERT(!IO_IS_ALLOCATING(zio)); | |
4088 | checksum = ZIO_CHECKSUM_GANG_HEADER; | |
4089 | } else { | |
4090 | checksum = BP_GET_CHECKSUM(bp); | |
4091 | } | |
4092 | } | |
34dc7c2f | 4093 | |
a6255b7f | 4094 | zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size); |
34dc7c2f | 4095 | |
62840030 | 4096 | return (zio); |
34dc7c2f BB |
4097 | } |
4098 | ||
62840030 | 4099 | static zio_t * |
b128c09f | 4100 | zio_checksum_verify(zio_t *zio) |
34dc7c2f | 4101 | { |
428870ff | 4102 | zio_bad_cksum_t info; |
b128c09f BB |
4103 | blkptr_t *bp = zio->io_bp; |
4104 | int error; | |
34dc7c2f | 4105 | |
428870ff BB |
4106 | ASSERT(zio->io_vd != NULL); |
4107 | ||
b128c09f BB |
4108 | if (bp == NULL) { |
4109 | /* | |
4110 | * This is zio_read_phys(). | |
4111 | * We're either verifying a label checksum, or nothing at all. | |
4112 | */ | |
4113 | if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) | |
62840030 | 4114 | return (zio); |
34dc7c2f | 4115 | |
b128c09f BB |
4116 | ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); |
4117 | } | |
34dc7c2f | 4118 | |
428870ff | 4119 | if ((error = zio_checksum_error(zio, &info)) != 0) { |
b128c09f | 4120 | zio->io_error = error; |
7a3066ff MA |
4121 | if (error == ECKSUM && |
4122 | !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { | |
428870ff | 4123 | zfs_ereport_start_checksum(zio->io_spa, |
b5256303 TC |
4124 | zio->io_vd, &zio->io_bookmark, zio, |
4125 | zio->io_offset, zio->io_size, NULL, &info); | |
b128c09f | 4126 | } |
34dc7c2f BB |
4127 | } |
4128 | ||
62840030 | 4129 | return (zio); |
34dc7c2f BB |
4130 | } |
4131 | ||
4132 | /* | |
4133 | * Called by RAID-Z to ensure we don't compute the checksum twice. | |
4134 | */ | |
4135 | void | |
4136 | zio_checksum_verified(zio_t *zio) | |
4137 | { | |
428870ff | 4138 | zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; |
34dc7c2f BB |
4139 | } |
4140 | ||
4141 | /* | |
b128c09f BB |
4142 | * ========================================================================== |
4143 | * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. | |
9b67f605 | 4144 | * An error of 0 indicates success. ENXIO indicates whole-device failure, |
b128c09f BB |
4145 | * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO |
4146 | * indicate errors that are specific to one I/O, and most likely permanent. | |
4147 | * Any other error is presumed to be worse because we weren't expecting it. | |
4148 | * ========================================================================== | |
34dc7c2f | 4149 | */ |
b128c09f BB |
4150 | int |
4151 | zio_worst_error(int e1, int e2) | |
34dc7c2f | 4152 | { |
b128c09f BB |
4153 | static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; |
4154 | int r1, r2; | |
4155 | ||
4156 | for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) | |
4157 | if (e1 == zio_error_rank[r1]) | |
4158 | break; | |
34dc7c2f | 4159 | |
b128c09f BB |
4160 | for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) |
4161 | if (e2 == zio_error_rank[r2]) | |
4162 | break; | |
4163 | ||
4164 | return (r1 > r2 ? e1 : e2); | |
34dc7c2f BB |
4165 | } |
4166 | ||
4167 | /* | |
4168 | * ========================================================================== | |
b128c09f | 4169 | * I/O completion |
34dc7c2f BB |
4170 | * ========================================================================== |
4171 | */ | |
62840030 | 4172 | static zio_t * |
b128c09f | 4173 | zio_ready(zio_t *zio) |
34dc7c2f | 4174 | { |
b128c09f | 4175 | blkptr_t *bp = zio->io_bp; |
d164b209 | 4176 | zio_t *pio, *pio_next; |
3dfb57a3 | 4177 | zio_link_t *zl = NULL; |
34dc7c2f | 4178 | |
ddc751d5 GW |
4179 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT, |
4180 | ZIO_WAIT_READY)) { | |
62840030 | 4181 | return (NULL); |
ddc751d5 | 4182 | } |
34dc7c2f | 4183 | |
9babb374 | 4184 | if (zio->io_ready) { |
b128c09f | 4185 | ASSERT(IO_IS_ALLOCATING(zio)); |
03c6040b GW |
4186 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || |
4187 | (zio->io_flags & ZIO_FLAG_NOPWRITE)); | |
b128c09f | 4188 | ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); |
34dc7c2f | 4189 | |
b128c09f BB |
4190 | zio->io_ready(zio); |
4191 | } | |
34dc7c2f | 4192 | |
b128c09f BB |
4193 | if (bp != NULL && bp != &zio->io_bp_copy) |
4194 | zio->io_bp_copy = *bp; | |
34dc7c2f | 4195 | |
3dfb57a3 | 4196 | if (zio->io_error != 0) { |
b128c09f | 4197 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f | 4198 | |
3dfb57a3 DB |
4199 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
4200 | ASSERT(IO_IS_ALLOCATING(zio)); | |
4201 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
cc99f275 DB |
4202 | ASSERT(zio->io_metaslab_class != NULL); |
4203 | ||
3dfb57a3 DB |
4204 | /* |
4205 | * We were unable to allocate anything, unreserve and | |
4206 | * issue the next I/O to allocate. | |
4207 | */ | |
4208 | metaslab_class_throttle_unreserve( | |
cc99f275 DB |
4209 | zio->io_metaslab_class, zio->io_prop.zp_copies, |
4210 | zio->io_allocator, zio); | |
492f64e9 | 4211 | zio_allocate_dispatch(zio->io_spa, zio->io_allocator); |
3dfb57a3 DB |
4212 | } |
4213 | } | |
4214 | ||
d164b209 BB |
4215 | mutex_enter(&zio->io_lock); |
4216 | zio->io_state[ZIO_WAIT_READY] = 1; | |
3dfb57a3 | 4217 | pio = zio_walk_parents(zio, &zl); |
d164b209 BB |
4218 | mutex_exit(&zio->io_lock); |
4219 | ||
4220 | /* | |
4221 | * As we notify zio's parents, new parents could be added. | |
4222 | * New parents go to the head of zio's io_parent_list, however, | |
4223 | * so we will (correctly) not notify them. The remainder of zio's | |
4224 | * io_parent_list, from 'pio_next' onward, cannot change because | |
4225 | * all parents must wait for us to be done before they can be done. | |
4226 | */ | |
4227 | for (; pio != NULL; pio = pio_next) { | |
3dfb57a3 | 4228 | pio_next = zio_walk_parents(zio, &zl); |
62840030 | 4229 | zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL); |
d164b209 | 4230 | } |
34dc7c2f | 4231 | |
428870ff BB |
4232 | if (zio->io_flags & ZIO_FLAG_NODATA) { |
4233 | if (BP_IS_GANG(bp)) { | |
4234 | zio->io_flags &= ~ZIO_FLAG_NODATA; | |
4235 | } else { | |
a6255b7f | 4236 | ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE); |
428870ff BB |
4237 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; |
4238 | } | |
4239 | } | |
4240 | ||
4241 | if (zio_injection_enabled && | |
4242 | zio->io_spa->spa_syncing_txg == zio->io_txg) | |
4243 | zio_handle_ignored_writes(zio); | |
4244 | ||
62840030 | 4245 | return (zio); |
34dc7c2f BB |
4246 | } |
4247 | ||
3dfb57a3 DB |
4248 | /* |
4249 | * Update the allocation throttle accounting. | |
4250 | */ | |
4251 | static void | |
4252 | zio_dva_throttle_done(zio_t *zio) | |
4253 | { | |
1c27024e | 4254 | ASSERTV(zio_t *lio = zio->io_logical); |
3dfb57a3 DB |
4255 | zio_t *pio = zio_unique_parent(zio); |
4256 | vdev_t *vd = zio->io_vd; | |
4257 | int flags = METASLAB_ASYNC_ALLOC; | |
4258 | ||
4259 | ASSERT3P(zio->io_bp, !=, NULL); | |
4260 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); | |
4261 | ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE); | |
4262 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); | |
4263 | ASSERT(vd != NULL); | |
4264 | ASSERT3P(vd, ==, vd->vdev_top); | |
21df134f SB |
4265 | ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY)); |
4266 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); | |
3dfb57a3 DB |
4267 | ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING); |
4268 | ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
4269 | ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA)); | |
4270 | ||
4271 | /* | |
4272 | * Parents of gang children can have two flavors -- ones that | |
4273 | * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set) | |
4274 | * and ones that allocated the constituent blocks. The allocation | |
4275 | * throttle needs to know the allocating parent zio so we must find | |
4276 | * it here. | |
4277 | */ | |
4278 | if (pio->io_child_type == ZIO_CHILD_GANG) { | |
4279 | /* | |
4280 | * If our parent is a rewrite gang child then our grandparent | |
4281 | * would have been the one that performed the allocation. | |
4282 | */ | |
4283 | if (pio->io_flags & ZIO_FLAG_IO_REWRITE) | |
4284 | pio = zio_unique_parent(pio); | |
4285 | flags |= METASLAB_GANG_CHILD; | |
4286 | } | |
4287 | ||
4288 | ASSERT(IO_IS_ALLOCATING(pio)); | |
4289 | ASSERT3P(zio, !=, zio->io_logical); | |
4290 | ASSERT(zio->io_logical != NULL); | |
4291 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); | |
4292 | ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE); | |
cc99f275 | 4293 | ASSERT(zio->io_metaslab_class != NULL); |
3dfb57a3 DB |
4294 | |
4295 | mutex_enter(&pio->io_lock); | |
492f64e9 PD |
4296 | metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags, |
4297 | pio->io_allocator, B_TRUE); | |
3dfb57a3 DB |
4298 | mutex_exit(&pio->io_lock); |
4299 | ||
cc99f275 DB |
4300 | metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1, |
4301 | pio->io_allocator, pio); | |
3dfb57a3 DB |
4302 | |
4303 | /* | |
4304 | * Call into the pipeline to see if there is more work that | |
4305 | * needs to be done. If there is work to be done it will be | |
4306 | * dispatched to another taskq thread. | |
4307 | */ | |
492f64e9 | 4308 | zio_allocate_dispatch(zio->io_spa, pio->io_allocator); |
3dfb57a3 DB |
4309 | } |
4310 | ||
62840030 | 4311 | static zio_t * |
b128c09f | 4312 | zio_done(zio_t *zio) |
34dc7c2f | 4313 | { |
3dfb57a3 DB |
4314 | /* |
4315 | * Always attempt to keep stack usage minimal here since | |
4316 | * we can be called recurisvely up to 19 levels deep. | |
4317 | */ | |
84c07ada | 4318 | const uint64_t psize = zio->io_size; |
d164b209 | 4319 | zio_t *pio, *pio_next; |
3dfb57a3 | 4320 | zio_link_t *zl = NULL; |
34dc7c2f | 4321 | |
b128c09f | 4322 | /* |
9babb374 | 4323 | * If our children haven't all completed, |
b128c09f BB |
4324 | * wait for them and then repeat this pipeline stage. |
4325 | */ | |
ddc751d5 | 4326 | if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) { |
62840030 | 4327 | return (NULL); |
ddc751d5 | 4328 | } |
34dc7c2f | 4329 | |
3dfb57a3 DB |
4330 | /* |
4331 | * If the allocation throttle is enabled, then update the accounting. | |
4332 | * We only track child I/Os that are part of an allocating async | |
4333 | * write. We must do this since the allocation is performed | |
4334 | * by the logical I/O but the actual write is done by child I/Os. | |
4335 | */ | |
4336 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING && | |
4337 | zio->io_child_type == ZIO_CHILD_VDEV) { | |
cc99f275 DB |
4338 | ASSERT(zio->io_metaslab_class != NULL); |
4339 | ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled); | |
3dfb57a3 DB |
4340 | zio_dva_throttle_done(zio); |
4341 | } | |
4342 | ||
4343 | /* | |
4344 | * If the allocation throttle is enabled, verify that | |
4345 | * we have decremented the refcounts for every I/O that was throttled. | |
4346 | */ | |
4347 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { | |
4348 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
4349 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
4350 | ASSERT(zio->io_bp != NULL); | |
cc99f275 | 4351 | |
492f64e9 PD |
4352 | metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio, |
4353 | zio->io_allocator); | |
cc99f275 DB |
4354 | VERIFY(refcount_not_held( |
4355 | &zio->io_metaslab_class->mc_alloc_slots[zio->io_allocator], | |
492f64e9 | 4356 | zio)); |
3dfb57a3 DB |
4357 | } |
4358 | ||
4359 | ||
1c27024e DB |
4360 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) |
4361 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) | |
b128c09f BB |
4362 | ASSERT(zio->io_children[c][w] == 0); |
4363 | ||
9b67f605 | 4364 | if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) { |
c776b317 BB |
4365 | ASSERT(zio->io_bp->blk_pad[0] == 0); |
4366 | ASSERT(zio->io_bp->blk_pad[1] == 0); | |
d1d7e268 MK |
4367 | ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, |
4368 | sizeof (blkptr_t)) == 0 || | |
c776b317 BB |
4369 | (zio->io_bp == zio_unique_parent(zio)->io_bp)); |
4370 | if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) && | |
428870ff | 4371 | zio->io_bp_override == NULL && |
b128c09f | 4372 | !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { |
d1d7e268 MK |
4373 | ASSERT3U(zio->io_prop.zp_copies, <=, |
4374 | BP_GET_NDVAS(zio->io_bp)); | |
c776b317 | 4375 | ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 || |
d1d7e268 MK |
4376 | (BP_COUNT_GANG(zio->io_bp) == |
4377 | BP_GET_NDVAS(zio->io_bp))); | |
b128c09f | 4378 | } |
03c6040b GW |
4379 | if (zio->io_flags & ZIO_FLAG_NOPWRITE) |
4380 | VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig)); | |
b128c09f BB |
4381 | } |
4382 | ||
4383 | /* | |
428870ff | 4384 | * If there were child vdev/gang/ddt errors, they apply to us now. |
b128c09f BB |
4385 | */ |
4386 | zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); | |
4387 | zio_inherit_child_errors(zio, ZIO_CHILD_GANG); | |
428870ff BB |
4388 | zio_inherit_child_errors(zio, ZIO_CHILD_DDT); |
4389 | ||
4390 | /* | |
4391 | * If the I/O on the transformed data was successful, generate any | |
4392 | * checksum reports now while we still have the transformed data. | |
4393 | */ | |
4394 | if (zio->io_error == 0) { | |
4395 | while (zio->io_cksum_report != NULL) { | |
4396 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
4397 | uint64_t align = zcr->zcr_align; | |
a6255b7f | 4398 | uint64_t asize = P2ROUNDUP(psize, align); |
a6255b7f DQ |
4399 | abd_t *adata = zio->io_abd; |
4400 | ||
4401 | if (asize != psize) { | |
84c07ada | 4402 | adata = abd_alloc(asize, B_TRUE); |
a6255b7f DQ |
4403 | abd_copy(adata, zio->io_abd, psize); |
4404 | abd_zero_off(adata, psize, asize - psize); | |
428870ff BB |
4405 | } |
4406 | ||
4407 | zio->io_cksum_report = zcr->zcr_next; | |
4408 | zcr->zcr_next = NULL; | |
84c07ada | 4409 | zcr->zcr_finish(zcr, adata); |
428870ff BB |
4410 | zfs_ereport_free_checksum(zcr); |
4411 | ||
a6255b7f DQ |
4412 | if (asize != psize) |
4413 | abd_free(adata); | |
428870ff BB |
4414 | } |
4415 | } | |
b128c09f BB |
4416 | |
4417 | zio_pop_transforms(zio); /* note: may set zio->io_error */ | |
4418 | ||
a6255b7f | 4419 | vdev_stat_update(zio, psize); |
b128c09f | 4420 | |
a69052be | 4421 | /* |
cc92e9d0 | 4422 | * If this I/O is attached to a particular vdev is slow, exceeding |
72f53c56 MJ |
4423 | * 30 seconds to complete, post an error described the I/O delay. |
4424 | * We ignore these errors if the device is currently unavailable. | |
a69052be | 4425 | */ |
193a37cb | 4426 | if (zio->io_delay >= MSEC2NSEC(zio_delay_max)) { |
72f53c56 MJ |
4427 | if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) |
4428 | zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa, | |
b5256303 | 4429 | zio->io_vd, &zio->io_bookmark, zio, 0, 0); |
72f53c56 | 4430 | } |
a69052be | 4431 | |
b128c09f BB |
4432 | if (zio->io_error) { |
4433 | /* | |
4434 | * If this I/O is attached to a particular vdev, | |
4435 | * generate an error message describing the I/O failure | |
4436 | * at the block level. We ignore these errors if the | |
4437 | * device is currently unavailable. | |
4438 | */ | |
c776b317 | 4439 | if (zio->io_error != ECKSUM && zio->io_vd != NULL && |
02730c33 | 4440 | !vdev_is_dead(zio->io_vd)) |
c776b317 | 4441 | zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa, |
b5256303 | 4442 | zio->io_vd, &zio->io_bookmark, zio, 0, 0); |
34dc7c2f | 4443 | |
428870ff BB |
4444 | if ((zio->io_error == EIO || !(zio->io_flags & |
4445 | (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && | |
c776b317 | 4446 | zio == zio->io_logical) { |
b128c09f BB |
4447 | /* |
4448 | * For logical I/O requests, tell the SPA to log the | |
4449 | * error and generate a logical data ereport. | |
4450 | */ | |
b5256303 | 4451 | spa_log_error(zio->io_spa, &zio->io_bookmark); |
d1d7e268 | 4452 | zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, |
b5256303 | 4453 | NULL, &zio->io_bookmark, zio, 0, 0); |
b128c09f BB |
4454 | } |
4455 | } | |
34dc7c2f | 4456 | |
c776b317 | 4457 | if (zio->io_error && zio == zio->io_logical) { |
b128c09f BB |
4458 | /* |
4459 | * Determine whether zio should be reexecuted. This will | |
4460 | * propagate all the way to the root via zio_notify_parent(). | |
4461 | */ | |
c776b317 | 4462 | ASSERT(zio->io_vd == NULL && zio->io_bp != NULL); |
428870ff | 4463 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); |
b128c09f | 4464 | |
428870ff BB |
4465 | if (IO_IS_ALLOCATING(zio) && |
4466 | !(zio->io_flags & ZIO_FLAG_CANFAIL)) { | |
b128c09f BB |
4467 | if (zio->io_error != ENOSPC) |
4468 | zio->io_reexecute |= ZIO_REEXECUTE_NOW; | |
4469 | else | |
4470 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff | 4471 | } |
b128c09f BB |
4472 | |
4473 | if ((zio->io_type == ZIO_TYPE_READ || | |
4474 | zio->io_type == ZIO_TYPE_FREE) && | |
572e2857 | 4475 | !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && |
b128c09f | 4476 | zio->io_error == ENXIO && |
c776b317 BB |
4477 | spa_load_state(zio->io_spa) == SPA_LOAD_NONE && |
4478 | spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE) | |
b128c09f BB |
4479 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; |
4480 | ||
4481 | if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) | |
4482 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff BB |
4483 | |
4484 | /* | |
4485 | * Here is a possibly good place to attempt to do | |
4486 | * either combinatorial reconstruction or error correction | |
4487 | * based on checksums. It also might be a good place | |
4488 | * to send out preliminary ereports before we suspend | |
4489 | * processing. | |
4490 | */ | |
34dc7c2f BB |
4491 | } |
4492 | ||
4493 | /* | |
b128c09f BB |
4494 | * If there were logical child errors, they apply to us now. |
4495 | * We defer this until now to avoid conflating logical child | |
4496 | * errors with errors that happened to the zio itself when | |
4497 | * updating vdev stats and reporting FMA events above. | |
34dc7c2f | 4498 | */ |
b128c09f | 4499 | zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); |
34dc7c2f | 4500 | |
428870ff BB |
4501 | if ((zio->io_error || zio->io_reexecute) && |
4502 | IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && | |
03c6040b | 4503 | !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) |
c776b317 | 4504 | zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp); |
9babb374 BB |
4505 | |
4506 | zio_gang_tree_free(&zio->io_gang_tree); | |
4507 | ||
4508 | /* | |
4509 | * Godfather I/Os should never suspend. | |
4510 | */ | |
4511 | if ((zio->io_flags & ZIO_FLAG_GODFATHER) && | |
4512 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) | |
a32494d2 | 4513 | zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND; |
9babb374 | 4514 | |
b128c09f BB |
4515 | if (zio->io_reexecute) { |
4516 | /* | |
4517 | * This is a logical I/O that wants to reexecute. | |
4518 | * | |
4519 | * Reexecute is top-down. When an i/o fails, if it's not | |
4520 | * the root, it simply notifies its parent and sticks around. | |
4521 | * The parent, seeing that it still has children in zio_done(), | |
4522 | * does the same. This percolates all the way up to the root. | |
4523 | * The root i/o will reexecute or suspend the entire tree. | |
4524 | * | |
4525 | * This approach ensures that zio_reexecute() honors | |
4526 | * all the original i/o dependency relationships, e.g. | |
4527 | * parents not executing until children are ready. | |
4528 | */ | |
4529 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
34dc7c2f | 4530 | |
9babb374 | 4531 | zio->io_gang_leader = NULL; |
b128c09f | 4532 | |
d164b209 BB |
4533 | mutex_enter(&zio->io_lock); |
4534 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
4535 | mutex_exit(&zio->io_lock); | |
4536 | ||
9babb374 BB |
4537 | /* |
4538 | * "The Godfather" I/O monitors its children but is | |
4539 | * not a true parent to them. It will track them through | |
4540 | * the pipeline but severs its ties whenever they get into | |
4541 | * trouble (e.g. suspended). This allows "The Godfather" | |
4542 | * I/O to return status without blocking. | |
4543 | */ | |
3dfb57a3 DB |
4544 | zl = NULL; |
4545 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; | |
4546 | pio = pio_next) { | |
4547 | zio_link_t *remove_zl = zl; | |
4548 | pio_next = zio_walk_parents(zio, &zl); | |
9babb374 BB |
4549 | |
4550 | if ((pio->io_flags & ZIO_FLAG_GODFATHER) && | |
4551 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { | |
3dfb57a3 | 4552 | zio_remove_child(pio, zio, remove_zl); |
62840030 MA |
4553 | /* |
4554 | * This is a rare code path, so we don't | |
4555 | * bother with "next_to_execute". | |
4556 | */ | |
4557 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, | |
4558 | NULL); | |
9babb374 BB |
4559 | } |
4560 | } | |
4561 | ||
d164b209 | 4562 | if ((pio = zio_unique_parent(zio)) != NULL) { |
b128c09f BB |
4563 | /* |
4564 | * We're not a root i/o, so there's nothing to do | |
4565 | * but notify our parent. Don't propagate errors | |
4566 | * upward since we haven't permanently failed yet. | |
4567 | */ | |
9babb374 | 4568 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f | 4569 | zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; |
62840030 MA |
4570 | /* |
4571 | * This is a rare code path, so we don't bother with | |
4572 | * "next_to_execute". | |
4573 | */ | |
4574 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL); | |
b128c09f BB |
4575 | } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { |
4576 | /* | |
4577 | * We'd fail again if we reexecuted now, so suspend | |
4578 | * until conditions improve (e.g. device comes online). | |
4579 | */ | |
cec3a0a1 | 4580 | zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR); |
b128c09f BB |
4581 | } else { |
4582 | /* | |
4583 | * Reexecution is potentially a huge amount of work. | |
4584 | * Hand it off to the otherwise-unused claim taskq. | |
4585 | */ | |
a38718a6 | 4586 | ASSERT(taskq_empty_ent(&zio->io_tqent)); |
7ef5e54e AL |
4587 | spa_taskq_dispatch_ent(zio->io_spa, |
4588 | ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE, | |
a38718a6 GA |
4589 | (task_func_t *)zio_reexecute, zio, 0, |
4590 | &zio->io_tqent); | |
b128c09f | 4591 | } |
62840030 | 4592 | return (NULL); |
34dc7c2f BB |
4593 | } |
4594 | ||
428870ff | 4595 | ASSERT(zio->io_child_count == 0); |
b128c09f BB |
4596 | ASSERT(zio->io_reexecute == 0); |
4597 | ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); | |
34dc7c2f | 4598 | |
428870ff BB |
4599 | /* |
4600 | * Report any checksum errors, since the I/O is complete. | |
4601 | */ | |
4602 | while (zio->io_cksum_report != NULL) { | |
4603 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
4604 | zio->io_cksum_report = zcr->zcr_next; | |
4605 | zcr->zcr_next = NULL; | |
4606 | zcr->zcr_finish(zcr, NULL); | |
4607 | zfs_ereport_free_checksum(zcr); | |
4608 | } | |
4609 | ||
920dd524 | 4610 | if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp && |
9b67f605 MA |
4611 | !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) && |
4612 | !(zio->io_flags & ZIO_FLAG_NOPWRITE)) { | |
920dd524 ED |
4613 | metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp); |
4614 | } | |
4615 | ||
d164b209 BB |
4616 | /* |
4617 | * It is the responsibility of the done callback to ensure that this | |
4618 | * particular zio is no longer discoverable for adoption, and as | |
4619 | * such, cannot acquire any new parents. | |
4620 | */ | |
b128c09f BB |
4621 | if (zio->io_done) |
4622 | zio->io_done(zio); | |
34dc7c2f | 4623 | |
d164b209 BB |
4624 | mutex_enter(&zio->io_lock); |
4625 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
4626 | mutex_exit(&zio->io_lock); | |
34dc7c2f | 4627 | |
62840030 MA |
4628 | /* |
4629 | * We are done executing this zio. We may want to execute a parent | |
4630 | * next. See the comment in zio_notify_parent(). | |
4631 | */ | |
4632 | zio_t *next_to_execute = NULL; | |
3dfb57a3 DB |
4633 | zl = NULL; |
4634 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) { | |
4635 | zio_link_t *remove_zl = zl; | |
4636 | pio_next = zio_walk_parents(zio, &zl); | |
4637 | zio_remove_child(pio, zio, remove_zl); | |
62840030 | 4638 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute); |
b128c09f | 4639 | } |
34dc7c2f | 4640 | |
b128c09f BB |
4641 | if (zio->io_waiter != NULL) { |
4642 | mutex_enter(&zio->io_lock); | |
4643 | zio->io_executor = NULL; | |
4644 | cv_broadcast(&zio->io_cv); | |
4645 | mutex_exit(&zio->io_lock); | |
4646 | } else { | |
4647 | zio_destroy(zio); | |
4648 | } | |
34dc7c2f | 4649 | |
62840030 | 4650 | return (next_to_execute); |
34dc7c2f BB |
4651 | } |
4652 | ||
4653 | /* | |
b128c09f BB |
4654 | * ========================================================================== |
4655 | * I/O pipeline definition | |
4656 | * ========================================================================== | |
34dc7c2f | 4657 | */ |
428870ff | 4658 | static zio_pipe_stage_t *zio_pipeline[] = { |
b128c09f | 4659 | NULL, |
b128c09f | 4660 | zio_read_bp_init, |
3dfb57a3 | 4661 | zio_write_bp_init, |
428870ff BB |
4662 | zio_free_bp_init, |
4663 | zio_issue_async, | |
3dfb57a3 | 4664 | zio_write_compress, |
b5256303 | 4665 | zio_encrypt, |
b128c09f | 4666 | zio_checksum_generate, |
03c6040b | 4667 | zio_nop_write, |
428870ff BB |
4668 | zio_ddt_read_start, |
4669 | zio_ddt_read_done, | |
4670 | zio_ddt_write, | |
4671 | zio_ddt_free, | |
b128c09f BB |
4672 | zio_gang_assemble, |
4673 | zio_gang_issue, | |
3dfb57a3 | 4674 | zio_dva_throttle, |
b128c09f BB |
4675 | zio_dva_allocate, |
4676 | zio_dva_free, | |
4677 | zio_dva_claim, | |
4678 | zio_ready, | |
4679 | zio_vdev_io_start, | |
4680 | zio_vdev_io_done, | |
4681 | zio_vdev_io_assess, | |
4682 | zio_checksum_verify, | |
4683 | zio_done | |
4684 | }; | |
c28b2279 | 4685 | |
9ae529ec | 4686 | |
9ae529ec | 4687 | |
9ae529ec | 4688 | |
fcff0f35 PD |
4689 | /* |
4690 | * Compare two zbookmark_phys_t's to see which we would reach first in a | |
4691 | * pre-order traversal of the object tree. | |
4692 | * | |
4693 | * This is simple in every case aside from the meta-dnode object. For all other | |
4694 | * objects, we traverse them in order (object 1 before object 2, and so on). | |
4695 | * However, all of these objects are traversed while traversing object 0, since | |
4696 | * the data it points to is the list of objects. Thus, we need to convert to a | |
4697 | * canonical representation so we can compare meta-dnode bookmarks to | |
4698 | * non-meta-dnode bookmarks. | |
4699 | * | |
4700 | * We do this by calculating "equivalents" for each field of the zbookmark. | |
4701 | * zbookmarks outside of the meta-dnode use their own object and level, and | |
4702 | * calculate the level 0 equivalent (the first L0 blkid that is contained in the | |
4703 | * blocks this bookmark refers to) by multiplying their blkid by their span | |
4704 | * (the number of L0 blocks contained within one block at their level). | |
4705 | * zbookmarks inside the meta-dnode calculate their object equivalent | |
4706 | * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use | |
4707 | * level + 1<<31 (any value larger than a level could ever be) for their level. | |
4708 | * This causes them to always compare before a bookmark in their object | |
4709 | * equivalent, compare appropriately to bookmarks in other objects, and to | |
4710 | * compare appropriately to other bookmarks in the meta-dnode. | |
4711 | */ | |
4712 | int | |
4713 | zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, | |
4714 | const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) | |
4715 | { | |
4716 | /* | |
4717 | * These variables represent the "equivalent" values for the zbookmark, | |
4718 | * after converting zbookmarks inside the meta dnode to their | |
4719 | * normal-object equivalents. | |
4720 | */ | |
4721 | uint64_t zb1obj, zb2obj; | |
4722 | uint64_t zb1L0, zb2L0; | |
4723 | uint64_t zb1level, zb2level; | |
4724 | ||
4725 | if (zb1->zb_object == zb2->zb_object && | |
4726 | zb1->zb_level == zb2->zb_level && | |
4727 | zb1->zb_blkid == zb2->zb_blkid) | |
4728 | return (0); | |
9ae529ec | 4729 | |
fcff0f35 PD |
4730 | /* |
4731 | * BP_SPANB calculates the span in blocks. | |
4732 | */ | |
4733 | zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); | |
4734 | zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); | |
9ae529ec CS |
4735 | |
4736 | if (zb1->zb_object == DMU_META_DNODE_OBJECT) { | |
fcff0f35 PD |
4737 | zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); |
4738 | zb1L0 = 0; | |
4739 | zb1level = zb1->zb_level + COMPARE_META_LEVEL; | |
4740 | } else { | |
4741 | zb1obj = zb1->zb_object; | |
4742 | zb1level = zb1->zb_level; | |
9ae529ec CS |
4743 | } |
4744 | ||
fcff0f35 PD |
4745 | if (zb2->zb_object == DMU_META_DNODE_OBJECT) { |
4746 | zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); | |
4747 | zb2L0 = 0; | |
4748 | zb2level = zb2->zb_level + COMPARE_META_LEVEL; | |
4749 | } else { | |
4750 | zb2obj = zb2->zb_object; | |
4751 | zb2level = zb2->zb_level; | |
4752 | } | |
4753 | ||
4754 | /* Now that we have a canonical representation, do the comparison. */ | |
4755 | if (zb1obj != zb2obj) | |
4756 | return (zb1obj < zb2obj ? -1 : 1); | |
4757 | else if (zb1L0 != zb2L0) | |
4758 | return (zb1L0 < zb2L0 ? -1 : 1); | |
4759 | else if (zb1level != zb2level) | |
4760 | return (zb1level > zb2level ? -1 : 1); | |
4761 | /* | |
4762 | * This can (theoretically) happen if the bookmarks have the same object | |
4763 | * and level, but different blkids, if the block sizes are not the same. | |
4764 | * There is presently no way to change the indirect block sizes | |
4765 | */ | |
4766 | return (0); | |
4767 | } | |
4768 | ||
4769 | /* | |
4770 | * This function checks the following: given that last_block is the place that | |
4771 | * our traversal stopped last time, does that guarantee that we've visited | |
4772 | * every node under subtree_root? Therefore, we can't just use the raw output | |
4773 | * of zbookmark_compare. We have to pass in a modified version of | |
4774 | * subtree_root; by incrementing the block id, and then checking whether | |
4775 | * last_block is before or equal to that, we can tell whether or not having | |
4776 | * visited last_block implies that all of subtree_root's children have been | |
4777 | * visited. | |
4778 | */ | |
4779 | boolean_t | |
4780 | zbookmark_subtree_completed(const dnode_phys_t *dnp, | |
4781 | const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) | |
4782 | { | |
4783 | zbookmark_phys_t mod_zb = *subtree_root; | |
4784 | mod_zb.zb_blkid++; | |
4785 | ASSERT(last_block->zb_level == 0); | |
4786 | ||
4787 | /* The objset_phys_t isn't before anything. */ | |
4788 | if (dnp == NULL) | |
9ae529ec | 4789 | return (B_FALSE); |
fcff0f35 PD |
4790 | |
4791 | /* | |
4792 | * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the | |
4793 | * data block size in sectors, because that variable is only used if | |
4794 | * the bookmark refers to a block in the meta-dnode. Since we don't | |
4795 | * know without examining it what object it refers to, and there's no | |
4796 | * harm in passing in this value in other cases, we always pass it in. | |
4797 | * | |
4798 | * We pass in 0 for the indirect block size shift because zb2 must be | |
4799 | * level 0. The indirect block size is only used to calculate the span | |
4800 | * of the bookmark, but since the bookmark must be level 0, the span is | |
4801 | * always 1, so the math works out. | |
4802 | * | |
4803 | * If you make changes to how the zbookmark_compare code works, be sure | |
4804 | * to make sure that this code still works afterwards. | |
4805 | */ | |
4806 | return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, | |
4807 | 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, | |
4808 | last_block) <= 0); | |
9ae529ec CS |
4809 | } |
4810 | ||
93ce2b4c | 4811 | #if defined(_KERNEL) |
c28b2279 | 4812 | EXPORT_SYMBOL(zio_type_name); |
81971b13 BB |
4813 | EXPORT_SYMBOL(zio_buf_alloc); |
4814 | EXPORT_SYMBOL(zio_data_buf_alloc); | |
4815 | EXPORT_SYMBOL(zio_buf_free); | |
4816 | EXPORT_SYMBOL(zio_data_buf_free); | |
c28b2279 | 4817 | |
a69052be | 4818 | module_param(zio_delay_max, int, 0644); |
c409e464 BB |
4819 | MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event"); |
4820 | ||
4821 | module_param(zio_requeue_io_start_cut_in_line, int, 0644); | |
4822 | MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O"); | |
29dee3ee CP |
4823 | |
4824 | module_param(zfs_sync_pass_deferred_free, int, 0644); | |
4825 | MODULE_PARM_DESC(zfs_sync_pass_deferred_free, | |
d1d7e268 | 4826 | "Defer frees starting in this pass"); |
29dee3ee CP |
4827 | |
4828 | module_param(zfs_sync_pass_dont_compress, int, 0644); | |
4829 | MODULE_PARM_DESC(zfs_sync_pass_dont_compress, | |
d1d7e268 | 4830 | "Don't compress starting in this pass"); |
29dee3ee CP |
4831 | |
4832 | module_param(zfs_sync_pass_rewrite, int, 0644); | |
4833 | MODULE_PARM_DESC(zfs_sync_pass_rewrite, | |
d1d7e268 | 4834 | "Rewrite new bps starting in this pass"); |
3dfb57a3 DB |
4835 | |
4836 | module_param(zio_dva_throttle_enabled, int, 0644); | |
4837 | MODULE_PARM_DESC(zio_dva_throttle_enabled, | |
4838 | "Throttle block allocations in the ZIO pipeline"); | |
c28b2279 | 4839 | #endif |