]>
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. |
64fc7762 | 23 | * Copyright (c) 2011, 2017 by Delphix. All rights reserved. |
a38718a6 | 24 | * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved. |
34dc7c2f BB |
25 | */ |
26 | ||
f1512ee6 | 27 | #include <sys/sysmacros.h> |
34dc7c2f BB |
28 | #include <sys/zfs_context.h> |
29 | #include <sys/fm/fs/zfs.h> | |
30 | #include <sys/spa.h> | |
31 | #include <sys/txg.h> | |
32 | #include <sys/spa_impl.h> | |
33 | #include <sys/vdev_impl.h> | |
34 | #include <sys/zio_impl.h> | |
35 | #include <sys/zio_compress.h> | |
36 | #include <sys/zio_checksum.h> | |
428870ff BB |
37 | #include <sys/dmu_objset.h> |
38 | #include <sys/arc.h> | |
39 | #include <sys/ddt.h> | |
9b67f605 | 40 | #include <sys/blkptr.h> |
b0bc7a84 | 41 | #include <sys/zfeature.h> |
d4a72f23 | 42 | #include <sys/dsl_scan.h> |
3dfb57a3 | 43 | #include <sys/metaslab_impl.h> |
193a37cb | 44 | #include <sys/time.h> |
26ef0cc7 | 45 | #include <sys/trace_zio.h> |
a6255b7f | 46 | #include <sys/abd.h> |
b5256303 | 47 | #include <sys/dsl_crypt.h> |
34dc7c2f | 48 | |
34dc7c2f BB |
49 | /* |
50 | * ========================================================================== | |
51 | * I/O type descriptions | |
52 | * ========================================================================== | |
53 | */ | |
e8b96c60 | 54 | const char *zio_type_name[ZIO_TYPES] = { |
3dfb57a3 DB |
55 | /* |
56 | * Note: Linux kernel thread name length is limited | |
57 | * so these names will differ from upstream open zfs. | |
58 | */ | |
451041db | 59 | "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl" |
428870ff | 60 | }; |
34dc7c2f | 61 | |
27f2b90d | 62 | int zio_dva_throttle_enabled = B_TRUE; |
3dfb57a3 | 63 | |
34dc7c2f BB |
64 | /* |
65 | * ========================================================================== | |
66 | * I/O kmem caches | |
67 | * ========================================================================== | |
68 | */ | |
69 | kmem_cache_t *zio_cache; | |
d164b209 | 70 | kmem_cache_t *zio_link_cache; |
34dc7c2f BB |
71 | kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; |
72 | kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
a6255b7f DQ |
73 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
74 | uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
75 | uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
76 | #endif | |
77 | ||
a69052be | 78 | int zio_delay_max = ZIO_DELAY_MAX; |
34dc7c2f | 79 | |
98b25418 GW |
80 | #define ZIO_PIPELINE_CONTINUE 0x100 |
81 | #define ZIO_PIPELINE_STOP 0x101 | |
82 | ||
fcff0f35 PD |
83 | #define BP_SPANB(indblkshift, level) \ |
84 | (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) | |
85 | #define COMPARE_META_LEVEL 0x80000000ul | |
55d85d5a GW |
86 | /* |
87 | * The following actions directly effect the spa's sync-to-convergence logic. | |
88 | * The values below define the sync pass when we start performing the action. | |
89 | * Care should be taken when changing these values as they directly impact | |
90 | * spa_sync() performance. Tuning these values may introduce subtle performance | |
91 | * pathologies and should only be done in the context of performance analysis. | |
92 | * These tunables will eventually be removed and replaced with #defines once | |
93 | * enough analysis has been done to determine optimal values. | |
94 | * | |
95 | * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that | |
96 | * regular blocks are not deferred. | |
97 | */ | |
98 | int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ | |
99 | int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */ | |
100 | int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ | |
101 | ||
34dc7c2f | 102 | /* |
b128c09f BB |
103 | * An allocating zio is one that either currently has the DVA allocate |
104 | * stage set or will have it later in its lifetime. | |
34dc7c2f | 105 | */ |
428870ff BB |
106 | #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) |
107 | ||
c409e464 | 108 | int zio_requeue_io_start_cut_in_line = 1; |
428870ff BB |
109 | |
110 | #ifdef ZFS_DEBUG | |
111 | int zio_buf_debug_limit = 16384; | |
112 | #else | |
113 | int zio_buf_debug_limit = 0; | |
114 | #endif | |
34dc7c2f | 115 | |
da6b4005 NB |
116 | static inline void __zio_execute(zio_t *zio); |
117 | ||
3dfb57a3 DB |
118 | static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t); |
119 | ||
34dc7c2f BB |
120 | void |
121 | zio_init(void) | |
122 | { | |
123 | size_t c; | |
124 | vmem_t *data_alloc_arena = NULL; | |
125 | ||
3941503c BB |
126 | zio_cache = kmem_cache_create("zio_cache", |
127 | sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); | |
d164b209 | 128 | zio_link_cache = kmem_cache_create("zio_link_cache", |
6795a698 | 129 | sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); |
34dc7c2f BB |
130 | |
131 | /* | |
132 | * For small buffers, we want a cache for each multiple of | |
f1512ee6 MA |
133 | * SPA_MINBLOCKSIZE. For larger buffers, we want a cache |
134 | * for each quarter-power of 2. | |
34dc7c2f BB |
135 | */ |
136 | for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { | |
137 | size_t size = (c + 1) << SPA_MINBLOCKSHIFT; | |
138 | size_t p2 = size; | |
139 | size_t align = 0; | |
6442f3cf | 140 | size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0; |
34dc7c2f | 141 | |
34328f3c | 142 | #if defined(_ILP32) && defined(_KERNEL) |
f1512ee6 MA |
143 | /* |
144 | * Cache size limited to 1M on 32-bit platforms until ARC | |
145 | * buffers no longer require virtual address space. | |
146 | */ | |
147 | if (size > zfs_max_recordsize) | |
148 | break; | |
149 | #endif | |
150 | ||
151 | while (!ISP2(p2)) | |
34dc7c2f BB |
152 | p2 &= p2 - 1; |
153 | ||
498877ba MA |
154 | #ifndef _KERNEL |
155 | /* | |
156 | * If we are using watchpoints, put each buffer on its own page, | |
157 | * to eliminate the performance overhead of trapping to the | |
158 | * kernel when modifying a non-watched buffer that shares the | |
159 | * page with a watched buffer. | |
160 | */ | |
161 | if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) | |
162 | continue; | |
fcf64f45 BB |
163 | /* |
164 | * Here's the problem - on 4K native devices in userland on | |
165 | * Linux using O_DIRECT, buffers must be 4K aligned or I/O | |
166 | * will fail with EINVAL, causing zdb (and others) to coredump. | |
167 | * Since userland probably doesn't need optimized buffer caches, | |
168 | * we just force 4K alignment on everything. | |
169 | */ | |
170 | align = 8 * SPA_MINBLOCKSIZE; | |
171 | #else | |
24fa2034 | 172 | if (size < PAGESIZE) { |
34dc7c2f | 173 | align = SPA_MINBLOCKSIZE; |
498877ba | 174 | } else if (IS_P2ALIGNED(size, p2 >> 2)) { |
24fa2034 | 175 | align = PAGESIZE; |
34dc7c2f | 176 | } |
fcf64f45 | 177 | #endif |
34dc7c2f BB |
178 | |
179 | if (align != 0) { | |
180 | char name[36]; | |
181 | (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); | |
182 | zio_buf_cache[c] = kmem_cache_create(name, size, | |
6442f3cf | 183 | align, NULL, NULL, NULL, NULL, NULL, cflags); |
34dc7c2f BB |
184 | |
185 | (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); | |
186 | zio_data_buf_cache[c] = kmem_cache_create(name, size, | |
ae6ba3db | 187 | align, NULL, NULL, NULL, NULL, |
6442f3cf | 188 | data_alloc_arena, cflags); |
34dc7c2f BB |
189 | } |
190 | } | |
191 | ||
192 | while (--c != 0) { | |
193 | ASSERT(zio_buf_cache[c] != NULL); | |
194 | if (zio_buf_cache[c - 1] == NULL) | |
195 | zio_buf_cache[c - 1] = zio_buf_cache[c]; | |
196 | ||
197 | ASSERT(zio_data_buf_cache[c] != NULL); | |
198 | if (zio_data_buf_cache[c - 1] == NULL) | |
199 | zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; | |
200 | } | |
201 | ||
34dc7c2f | 202 | zio_inject_init(); |
9759c60f ED |
203 | |
204 | lz4_init(); | |
34dc7c2f BB |
205 | } |
206 | ||
207 | void | |
208 | zio_fini(void) | |
209 | { | |
210 | size_t c; | |
211 | kmem_cache_t *last_cache = NULL; | |
212 | kmem_cache_t *last_data_cache = NULL; | |
213 | ||
214 | for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { | |
f1512ee6 MA |
215 | #ifdef _ILP32 |
216 | /* | |
217 | * Cache size limited to 1M on 32-bit platforms until ARC | |
218 | * buffers no longer require virtual address space. | |
219 | */ | |
220 | if (((c + 1) << SPA_MINBLOCKSHIFT) > zfs_max_recordsize) | |
221 | break; | |
a6255b7f DQ |
222 | #endif |
223 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) | |
224 | if (zio_buf_cache_allocs[c] != zio_buf_cache_frees[c]) | |
225 | (void) printf("zio_fini: [%d] %llu != %llu\n", | |
226 | (int)((c + 1) << SPA_MINBLOCKSHIFT), | |
227 | (long long unsigned)zio_buf_cache_allocs[c], | |
228 | (long long unsigned)zio_buf_cache_frees[c]); | |
f1512ee6 | 229 | #endif |
34dc7c2f BB |
230 | if (zio_buf_cache[c] != last_cache) { |
231 | last_cache = zio_buf_cache[c]; | |
232 | kmem_cache_destroy(zio_buf_cache[c]); | |
233 | } | |
234 | zio_buf_cache[c] = NULL; | |
235 | ||
236 | if (zio_data_buf_cache[c] != last_data_cache) { | |
237 | last_data_cache = zio_data_buf_cache[c]; | |
238 | kmem_cache_destroy(zio_data_buf_cache[c]); | |
239 | } | |
240 | zio_data_buf_cache[c] = NULL; | |
241 | } | |
242 | ||
d164b209 | 243 | kmem_cache_destroy(zio_link_cache); |
34dc7c2f BB |
244 | kmem_cache_destroy(zio_cache); |
245 | ||
246 | zio_inject_fini(); | |
9759c60f ED |
247 | |
248 | lz4_fini(); | |
34dc7c2f BB |
249 | } |
250 | ||
251 | /* | |
252 | * ========================================================================== | |
253 | * Allocate and free I/O buffers | |
254 | * ========================================================================== | |
255 | */ | |
256 | ||
257 | /* | |
258 | * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a | |
259 | * crashdump if the kernel panics, so use it judiciously. Obviously, it's | |
260 | * useful to inspect ZFS metadata, but if possible, we should avoid keeping | |
261 | * excess / transient data in-core during a crashdump. | |
262 | */ | |
263 | void * | |
264 | zio_buf_alloc(size_t size) | |
265 | { | |
266 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
267 | ||
63e3a861 | 268 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
a6255b7f DQ |
269 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
270 | atomic_add_64(&zio_buf_cache_allocs[c], 1); | |
271 | #endif | |
34dc7c2f | 272 | |
efcd79a8 | 273 | return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); |
34dc7c2f BB |
274 | } |
275 | ||
276 | /* | |
277 | * Use zio_data_buf_alloc to allocate data. The data will not appear in a | |
278 | * crashdump if the kernel panics. This exists so that we will limit the amount | |
279 | * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount | |
280 | * of kernel heap dumped to disk when the kernel panics) | |
281 | */ | |
282 | void * | |
283 | zio_data_buf_alloc(size_t size) | |
284 | { | |
285 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
286 | ||
63e3a861 | 287 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f | 288 | |
efcd79a8 | 289 | return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); |
34dc7c2f BB |
290 | } |
291 | ||
292 | void | |
293 | zio_buf_free(void *buf, size_t size) | |
294 | { | |
295 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
296 | ||
63e3a861 | 297 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
a6255b7f DQ |
298 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
299 | atomic_add_64(&zio_buf_cache_frees[c], 1); | |
300 | #endif | |
34dc7c2f BB |
301 | |
302 | kmem_cache_free(zio_buf_cache[c], buf); | |
303 | } | |
304 | ||
305 | void | |
306 | zio_data_buf_free(void *buf, size_t size) | |
307 | { | |
308 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
309 | ||
63e3a861 | 310 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f BB |
311 | |
312 | kmem_cache_free(zio_data_buf_cache[c], buf); | |
313 | } | |
314 | ||
84c07ada GN |
315 | static void |
316 | zio_abd_free(void *abd, size_t size) | |
317 | { | |
318 | abd_free((abd_t *)abd); | |
319 | } | |
320 | ||
34dc7c2f BB |
321 | /* |
322 | * ========================================================================== | |
323 | * Push and pop I/O transform buffers | |
324 | * ========================================================================== | |
325 | */ | |
d3c2ae1c | 326 | void |
a6255b7f | 327 | zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize, |
e9aa730c | 328 | zio_transform_func_t *transform) |
34dc7c2f | 329 | { |
79c76d5b | 330 | zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); |
34dc7c2f | 331 | |
a6255b7f DQ |
332 | /* |
333 | * Ensure that anyone expecting this zio to contain a linear ABD isn't | |
334 | * going to get a nasty surprise when they try to access the data. | |
335 | */ | |
336 | IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data)); | |
337 | ||
338 | zt->zt_orig_abd = zio->io_abd; | |
b128c09f | 339 | zt->zt_orig_size = zio->io_size; |
34dc7c2f | 340 | zt->zt_bufsize = bufsize; |
b128c09f | 341 | zt->zt_transform = transform; |
34dc7c2f BB |
342 | |
343 | zt->zt_next = zio->io_transform_stack; | |
344 | zio->io_transform_stack = zt; | |
345 | ||
a6255b7f | 346 | zio->io_abd = data; |
34dc7c2f BB |
347 | zio->io_size = size; |
348 | } | |
349 | ||
d3c2ae1c | 350 | void |
b128c09f | 351 | zio_pop_transforms(zio_t *zio) |
34dc7c2f | 352 | { |
b128c09f BB |
353 | zio_transform_t *zt; |
354 | ||
355 | while ((zt = zio->io_transform_stack) != NULL) { | |
356 | if (zt->zt_transform != NULL) | |
357 | zt->zt_transform(zio, | |
a6255b7f | 358 | zt->zt_orig_abd, zt->zt_orig_size); |
34dc7c2f | 359 | |
428870ff | 360 | if (zt->zt_bufsize != 0) |
a6255b7f | 361 | abd_free(zio->io_abd); |
34dc7c2f | 362 | |
a6255b7f | 363 | zio->io_abd = zt->zt_orig_abd; |
b128c09f BB |
364 | zio->io_size = zt->zt_orig_size; |
365 | zio->io_transform_stack = zt->zt_next; | |
34dc7c2f | 366 | |
b128c09f | 367 | kmem_free(zt, sizeof (zio_transform_t)); |
34dc7c2f BB |
368 | } |
369 | } | |
370 | ||
b128c09f BB |
371 | /* |
372 | * ========================================================================== | |
b5256303 | 373 | * I/O transform callbacks for subblocks, decompression, and decryption |
b128c09f BB |
374 | * ========================================================================== |
375 | */ | |
376 | static void | |
a6255b7f | 377 | zio_subblock(zio_t *zio, abd_t *data, uint64_t size) |
b128c09f BB |
378 | { |
379 | ASSERT(zio->io_size > size); | |
380 | ||
381 | if (zio->io_type == ZIO_TYPE_READ) | |
a6255b7f | 382 | abd_copy(data, zio->io_abd, size); |
b128c09f BB |
383 | } |
384 | ||
385 | static void | |
a6255b7f | 386 | zio_decompress(zio_t *zio, abd_t *data, uint64_t size) |
b128c09f | 387 | { |
a6255b7f DQ |
388 | if (zio->io_error == 0) { |
389 | void *tmp = abd_borrow_buf(data, size); | |
390 | int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), | |
391 | zio->io_abd, tmp, zio->io_size, size); | |
392 | abd_return_buf_copy(data, tmp, size); | |
393 | ||
394 | if (ret != 0) | |
395 | zio->io_error = SET_ERROR(EIO); | |
396 | } | |
b128c09f BB |
397 | } |
398 | ||
b5256303 TC |
399 | static void |
400 | zio_decrypt(zio_t *zio, abd_t *data, uint64_t size) | |
401 | { | |
402 | int ret; | |
403 | void *tmp; | |
404 | blkptr_t *bp = zio->io_bp; | |
ae76f45c TC |
405 | spa_t *spa = zio->io_spa; |
406 | uint64_t dsobj = zio->io_bookmark.zb_objset; | |
b5256303 TC |
407 | uint64_t lsize = BP_GET_LSIZE(bp); |
408 | dmu_object_type_t ot = BP_GET_TYPE(bp); | |
409 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
410 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
411 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
412 | boolean_t no_crypt = B_FALSE; | |
413 | ||
414 | ASSERT(BP_USES_CRYPT(bp)); | |
415 | ASSERT3U(size, !=, 0); | |
416 | ||
417 | if (zio->io_error != 0) | |
418 | return; | |
419 | ||
420 | /* | |
421 | * Verify the cksum of MACs stored in an indirect bp. It will always | |
422 | * be possible to verify this since it does not require an encryption | |
423 | * key. | |
424 | */ | |
425 | if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) { | |
426 | zio_crypt_decode_mac_bp(bp, mac); | |
427 | ||
428 | if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) { | |
429 | /* | |
430 | * We haven't decompressed the data yet, but | |
431 | * zio_crypt_do_indirect_mac_checksum() requires | |
432 | * decompressed data to be able to parse out the MACs | |
433 | * from the indirect block. We decompress it now and | |
434 | * throw away the result after we are finished. | |
435 | */ | |
436 | tmp = zio_buf_alloc(lsize); | |
437 | ret = zio_decompress_data(BP_GET_COMPRESS(bp), | |
438 | zio->io_abd, tmp, zio->io_size, lsize); | |
439 | if (ret != 0) { | |
440 | ret = SET_ERROR(EIO); | |
441 | goto error; | |
442 | } | |
443 | ret = zio_crypt_do_indirect_mac_checksum(B_FALSE, | |
444 | tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac); | |
445 | zio_buf_free(tmp, lsize); | |
446 | } else { | |
447 | ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE, | |
448 | zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac); | |
449 | } | |
450 | abd_copy(data, zio->io_abd, size); | |
451 | ||
452 | if (ret != 0) | |
453 | goto error; | |
454 | ||
455 | return; | |
456 | } | |
457 | ||
458 | /* | |
459 | * If this is an authenticated block, just check the MAC. It would be | |
460 | * nice to separate this out into its own flag, but for the moment | |
461 | * enum zio_flag is out of bits. | |
462 | */ | |
463 | if (BP_IS_AUTHENTICATED(bp)) { | |
464 | if (ot == DMU_OT_OBJSET) { | |
ae76f45c TC |
465 | ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, |
466 | dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp)); | |
b5256303 TC |
467 | } else { |
468 | zio_crypt_decode_mac_bp(bp, mac); | |
ae76f45c TC |
469 | ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, |
470 | zio->io_abd, size, mac); | |
b5256303 TC |
471 | } |
472 | abd_copy(data, zio->io_abd, size); | |
473 | ||
474 | if (ret != 0) | |
475 | goto error; | |
476 | ||
477 | return; | |
478 | } | |
479 | ||
480 | zio_crypt_decode_params_bp(bp, salt, iv); | |
481 | ||
482 | if (ot == DMU_OT_INTENT_LOG) { | |
483 | tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t)); | |
484 | zio_crypt_decode_mac_zil(tmp, mac); | |
485 | abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t)); | |
486 | } else { | |
487 | zio_crypt_decode_mac_bp(bp, mac); | |
488 | } | |
489 | ||
ae76f45c TC |
490 | ret = spa_do_crypt_abd(B_FALSE, spa, dsobj, bp, bp->blk_birth, |
491 | size, data, zio->io_abd, iv, mac, salt, &no_crypt); | |
b5256303 TC |
492 | if (no_crypt) |
493 | abd_copy(data, zio->io_abd, size); | |
494 | ||
495 | if (ret != 0) | |
496 | goto error; | |
497 | ||
498 | return; | |
499 | ||
500 | error: | |
501 | /* assert that the key was found unless this was speculative */ | |
502 | ASSERT(ret != ENOENT || (zio->io_flags & ZIO_FLAG_SPECULATIVE)); | |
503 | ||
504 | /* | |
505 | * If there was a decryption / authentication error return EIO as | |
506 | * the io_error. If this was not a speculative zio, create an ereport. | |
507 | */ | |
508 | if (ret == ECKSUM) { | |
509 | ret = SET_ERROR(EIO); | |
510 | if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) { | |
511 | zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION, | |
ae76f45c | 512 | spa, NULL, &zio->io_bookmark, zio, 0, 0); |
b5256303 TC |
513 | } |
514 | } else { | |
515 | zio->io_error = ret; | |
516 | } | |
517 | } | |
518 | ||
b128c09f BB |
519 | /* |
520 | * ========================================================================== | |
521 | * I/O parent/child relationships and pipeline interlocks | |
522 | * ========================================================================== | |
523 | */ | |
d164b209 | 524 | zio_t * |
3dfb57a3 | 525 | zio_walk_parents(zio_t *cio, zio_link_t **zl) |
d164b209 | 526 | { |
d164b209 | 527 | list_t *pl = &cio->io_parent_list; |
b128c09f | 528 | |
3dfb57a3 DB |
529 | *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl); |
530 | if (*zl == NULL) | |
d164b209 BB |
531 | return (NULL); |
532 | ||
3dfb57a3 DB |
533 | ASSERT((*zl)->zl_child == cio); |
534 | return ((*zl)->zl_parent); | |
d164b209 BB |
535 | } |
536 | ||
537 | zio_t * | |
3dfb57a3 | 538 | zio_walk_children(zio_t *pio, zio_link_t **zl) |
d164b209 | 539 | { |
d164b209 BB |
540 | list_t *cl = &pio->io_child_list; |
541 | ||
a8b2e306 TC |
542 | ASSERT(MUTEX_HELD(&pio->io_lock)); |
543 | ||
3dfb57a3 DB |
544 | *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl); |
545 | if (*zl == NULL) | |
d164b209 BB |
546 | return (NULL); |
547 | ||
3dfb57a3 DB |
548 | ASSERT((*zl)->zl_parent == pio); |
549 | return ((*zl)->zl_child); | |
d164b209 BB |
550 | } |
551 | ||
552 | zio_t * | |
553 | zio_unique_parent(zio_t *cio) | |
554 | { | |
3dfb57a3 DB |
555 | zio_link_t *zl = NULL; |
556 | zio_t *pio = zio_walk_parents(cio, &zl); | |
d164b209 | 557 | |
3dfb57a3 | 558 | VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL); |
d164b209 BB |
559 | return (pio); |
560 | } | |
561 | ||
562 | void | |
563 | zio_add_child(zio_t *pio, zio_t *cio) | |
b128c09f | 564 | { |
79c76d5b | 565 | zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); |
d164b209 BB |
566 | |
567 | /* | |
568 | * Logical I/Os can have logical, gang, or vdev children. | |
569 | * Gang I/Os can have gang or vdev children. | |
570 | * Vdev I/Os can only have vdev children. | |
571 | * The following ASSERT captures all of these constraints. | |
572 | */ | |
1ce23dca | 573 | ASSERT3S(cio->io_child_type, <=, pio->io_child_type); |
d164b209 BB |
574 | |
575 | zl->zl_parent = pio; | |
576 | zl->zl_child = cio; | |
577 | ||
b128c09f | 578 | mutex_enter(&pio->io_lock); |
a8b2e306 | 579 | mutex_enter(&cio->io_lock); |
d164b209 BB |
580 | |
581 | ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); | |
582 | ||
1c27024e | 583 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 BB |
584 | pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; |
585 | ||
586 | list_insert_head(&pio->io_child_list, zl); | |
587 | list_insert_head(&cio->io_parent_list, zl); | |
588 | ||
428870ff BB |
589 | pio->io_child_count++; |
590 | cio->io_parent_count++; | |
591 | ||
d164b209 | 592 | mutex_exit(&cio->io_lock); |
a8b2e306 | 593 | mutex_exit(&pio->io_lock); |
b128c09f BB |
594 | } |
595 | ||
34dc7c2f | 596 | static void |
d164b209 | 597 | zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) |
b128c09f | 598 | { |
d164b209 BB |
599 | ASSERT(zl->zl_parent == pio); |
600 | ASSERT(zl->zl_child == cio); | |
b128c09f BB |
601 | |
602 | mutex_enter(&pio->io_lock); | |
a8b2e306 | 603 | mutex_enter(&cio->io_lock); |
d164b209 BB |
604 | |
605 | list_remove(&pio->io_child_list, zl); | |
606 | list_remove(&cio->io_parent_list, zl); | |
607 | ||
428870ff BB |
608 | pio->io_child_count--; |
609 | cio->io_parent_count--; | |
610 | ||
d164b209 | 611 | mutex_exit(&cio->io_lock); |
a8b2e306 | 612 | mutex_exit(&pio->io_lock); |
d164b209 | 613 | kmem_cache_free(zio_link_cache, zl); |
b128c09f BB |
614 | } |
615 | ||
616 | static boolean_t | |
617 | zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) | |
34dc7c2f | 618 | { |
b128c09f BB |
619 | uint64_t *countp = &zio->io_children[child][wait]; |
620 | boolean_t waiting = B_FALSE; | |
621 | ||
622 | mutex_enter(&zio->io_lock); | |
623 | ASSERT(zio->io_stall == NULL); | |
624 | if (*countp != 0) { | |
428870ff | 625 | zio->io_stage >>= 1; |
3dfb57a3 | 626 | ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN); |
b128c09f BB |
627 | zio->io_stall = countp; |
628 | waiting = B_TRUE; | |
629 | } | |
630 | mutex_exit(&zio->io_lock); | |
631 | ||
632 | return (waiting); | |
633 | } | |
34dc7c2f | 634 | |
bf701a83 BB |
635 | __attribute__((always_inline)) |
636 | static inline void | |
b128c09f BB |
637 | zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) |
638 | { | |
639 | uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; | |
640 | int *errorp = &pio->io_child_error[zio->io_child_type]; | |
34dc7c2f | 641 | |
b128c09f BB |
642 | mutex_enter(&pio->io_lock); |
643 | if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) | |
644 | *errorp = zio_worst_error(*errorp, zio->io_error); | |
645 | pio->io_reexecute |= zio->io_reexecute; | |
646 | ASSERT3U(*countp, >, 0); | |
e8b96c60 MA |
647 | |
648 | (*countp)--; | |
649 | ||
650 | if (*countp == 0 && pio->io_stall == countp) { | |
3dfb57a3 DB |
651 | zio_taskq_type_t type = |
652 | pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE : | |
653 | ZIO_TASKQ_INTERRUPT; | |
b128c09f BB |
654 | pio->io_stall = NULL; |
655 | mutex_exit(&pio->io_lock); | |
3dfb57a3 DB |
656 | /* |
657 | * Dispatch the parent zio in its own taskq so that | |
658 | * the child can continue to make progress. This also | |
659 | * prevents overflowing the stack when we have deeply nested | |
660 | * parent-child relationships. | |
661 | */ | |
662 | zio_taskq_dispatch(pio, type, B_FALSE); | |
b128c09f BB |
663 | } else { |
664 | mutex_exit(&pio->io_lock); | |
34dc7c2f BB |
665 | } |
666 | } | |
667 | ||
b128c09f BB |
668 | static void |
669 | zio_inherit_child_errors(zio_t *zio, enum zio_child c) | |
670 | { | |
671 | if (zio->io_child_error[c] != 0 && zio->io_error == 0) | |
672 | zio->io_error = zio->io_child_error[c]; | |
673 | } | |
674 | ||
3dfb57a3 | 675 | int |
64fc7762 | 676 | zio_bookmark_compare(const void *x1, const void *x2) |
3dfb57a3 DB |
677 | { |
678 | const zio_t *z1 = x1; | |
679 | const zio_t *z2 = x2; | |
3dfb57a3 | 680 | |
64fc7762 MA |
681 | if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset) |
682 | return (-1); | |
683 | if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset) | |
684 | return (1); | |
3dfb57a3 | 685 | |
64fc7762 MA |
686 | if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object) |
687 | return (-1); | |
688 | if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object) | |
689 | return (1); | |
3dfb57a3 | 690 | |
64fc7762 MA |
691 | if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level) |
692 | return (-1); | |
693 | if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level) | |
694 | return (1); | |
695 | ||
696 | if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid) | |
697 | return (-1); | |
698 | if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid) | |
699 | return (1); | |
700 | ||
701 | if (z1 < z2) | |
702 | return (-1); | |
703 | if (z1 > z2) | |
704 | return (1); | |
705 | ||
706 | return (0); | |
3dfb57a3 DB |
707 | } |
708 | ||
34dc7c2f BB |
709 | /* |
710 | * ========================================================================== | |
b128c09f | 711 | * Create the various types of I/O (read, write, free, etc) |
34dc7c2f BB |
712 | * ========================================================================== |
713 | */ | |
714 | static zio_t * | |
428870ff | 715 | zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
a6255b7f | 716 | abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done, |
2aa34383 DK |
717 | void *private, zio_type_t type, zio_priority_t priority, |
718 | enum zio_flag flags, vdev_t *vd, uint64_t offset, | |
719 | const zbookmark_phys_t *zb, enum zio_stage stage, | |
720 | enum zio_stage pipeline) | |
34dc7c2f BB |
721 | { |
722 | zio_t *zio; | |
723 | ||
2aa34383 DK |
724 | ASSERT3U(psize, <=, SPA_MAXBLOCKSIZE); |
725 | ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); | |
b128c09f BB |
726 | ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); |
727 | ||
728 | ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); | |
729 | ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); | |
730 | ASSERT(vd || stage == ZIO_STAGE_OPEN); | |
34dc7c2f | 731 | |
b5256303 | 732 | IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0); |
2aa34383 | 733 | |
79c76d5b | 734 | zio = kmem_cache_alloc(zio_cache, KM_SLEEP); |
3941503c BB |
735 | bzero(zio, sizeof (zio_t)); |
736 | ||
448d7aaa | 737 | mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL); |
3941503c BB |
738 | cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); |
739 | ||
740 | list_create(&zio->io_parent_list, sizeof (zio_link_t), | |
741 | offsetof(zio_link_t, zl_parent_node)); | |
742 | list_create(&zio->io_child_list, sizeof (zio_link_t), | |
743 | offsetof(zio_link_t, zl_child_node)); | |
4e21fd06 | 744 | metaslab_trace_init(&zio->io_alloc_list); |
d164b209 | 745 | |
b128c09f BB |
746 | if (vd != NULL) |
747 | zio->io_child_type = ZIO_CHILD_VDEV; | |
748 | else if (flags & ZIO_FLAG_GANG_CHILD) | |
749 | zio->io_child_type = ZIO_CHILD_GANG; | |
428870ff BB |
750 | else if (flags & ZIO_FLAG_DDT_CHILD) |
751 | zio->io_child_type = ZIO_CHILD_DDT; | |
b128c09f BB |
752 | else |
753 | zio->io_child_type = ZIO_CHILD_LOGICAL; | |
754 | ||
34dc7c2f | 755 | if (bp != NULL) { |
428870ff | 756 | zio->io_bp = (blkptr_t *)bp; |
34dc7c2f BB |
757 | zio->io_bp_copy = *bp; |
758 | zio->io_bp_orig = *bp; | |
428870ff BB |
759 | if (type != ZIO_TYPE_WRITE || |
760 | zio->io_child_type == ZIO_CHILD_DDT) | |
b128c09f | 761 | zio->io_bp = &zio->io_bp_copy; /* so caller can free */ |
9babb374 | 762 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) |
b128c09f | 763 | zio->io_logical = zio; |
9babb374 BB |
764 | if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) |
765 | pipeline |= ZIO_GANG_STAGES; | |
34dc7c2f | 766 | } |
b128c09f BB |
767 | |
768 | zio->io_spa = spa; | |
769 | zio->io_txg = txg; | |
34dc7c2f BB |
770 | zio->io_done = done; |
771 | zio->io_private = private; | |
772 | zio->io_type = type; | |
773 | zio->io_priority = priority; | |
b128c09f BB |
774 | zio->io_vd = vd; |
775 | zio->io_offset = offset; | |
a6255b7f | 776 | zio->io_orig_abd = zio->io_abd = data; |
2aa34383 DK |
777 | zio->io_orig_size = zio->io_size = psize; |
778 | zio->io_lsize = lsize; | |
b128c09f BB |
779 | zio->io_orig_flags = zio->io_flags = flags; |
780 | zio->io_orig_stage = zio->io_stage = stage; | |
781 | zio->io_orig_pipeline = zio->io_pipeline = pipeline; | |
3dfb57a3 | 782 | zio->io_pipeline_trace = ZIO_STAGE_OPEN; |
34dc7c2f | 783 | |
d164b209 BB |
784 | zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); |
785 | zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); | |
786 | ||
b128c09f BB |
787 | if (zb != NULL) |
788 | zio->io_bookmark = *zb; | |
789 | ||
790 | if (pio != NULL) { | |
b128c09f | 791 | if (zio->io_logical == NULL) |
34dc7c2f | 792 | zio->io_logical = pio->io_logical; |
9babb374 BB |
793 | if (zio->io_child_type == ZIO_CHILD_GANG) |
794 | zio->io_gang_leader = pio->io_gang_leader; | |
b128c09f | 795 | zio_add_child(pio, zio); |
34dc7c2f BB |
796 | } |
797 | ||
a38718a6 GA |
798 | taskq_init_ent(&zio->io_tqent); |
799 | ||
34dc7c2f BB |
800 | return (zio); |
801 | } | |
802 | ||
803 | static void | |
b128c09f | 804 | zio_destroy(zio_t *zio) |
34dc7c2f | 805 | { |
4e21fd06 | 806 | metaslab_trace_fini(&zio->io_alloc_list); |
3941503c BB |
807 | list_destroy(&zio->io_parent_list); |
808 | list_destroy(&zio->io_child_list); | |
809 | mutex_destroy(&zio->io_lock); | |
810 | cv_destroy(&zio->io_cv); | |
b128c09f | 811 | kmem_cache_free(zio_cache, zio); |
34dc7c2f BB |
812 | } |
813 | ||
814 | zio_t * | |
d164b209 | 815 | zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, |
428870ff | 816 | void *private, enum zio_flag flags) |
34dc7c2f BB |
817 | { |
818 | zio_t *zio; | |
819 | ||
2aa34383 | 820 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
d164b209 | 821 | ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
b128c09f | 822 | ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); |
34dc7c2f BB |
823 | |
824 | return (zio); | |
825 | } | |
826 | ||
827 | zio_t * | |
428870ff | 828 | zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) |
34dc7c2f | 829 | { |
d164b209 | 830 | return (zio_null(NULL, spa, NULL, done, private, flags)); |
34dc7c2f BB |
831 | } |
832 | ||
63e3a861 MA |
833 | void |
834 | zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp) | |
835 | { | |
63e3a861 MA |
836 | if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { |
837 | zfs_panic_recover("blkptr at %p has invalid TYPE %llu", | |
838 | bp, (longlong_t)BP_GET_TYPE(bp)); | |
839 | } | |
840 | if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS || | |
841 | BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) { | |
842 | zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu", | |
843 | bp, (longlong_t)BP_GET_CHECKSUM(bp)); | |
844 | } | |
845 | if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS || | |
846 | BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) { | |
847 | zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu", | |
848 | bp, (longlong_t)BP_GET_COMPRESS(bp)); | |
849 | } | |
850 | if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
851 | zfs_panic_recover("blkptr at %p has invalid LSIZE %llu", | |
852 | bp, (longlong_t)BP_GET_LSIZE(bp)); | |
853 | } | |
854 | if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
855 | zfs_panic_recover("blkptr at %p has invalid PSIZE %llu", | |
856 | bp, (longlong_t)BP_GET_PSIZE(bp)); | |
857 | } | |
858 | ||
859 | if (BP_IS_EMBEDDED(bp)) { | |
860 | if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) { | |
861 | zfs_panic_recover("blkptr at %p has invalid ETYPE %llu", | |
862 | bp, (longlong_t)BPE_GET_ETYPE(bp)); | |
863 | } | |
864 | } | |
865 | ||
866 | /* | |
867 | * Pool-specific checks. | |
868 | * | |
869 | * Note: it would be nice to verify that the blk_birth and | |
870 | * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() | |
871 | * allows the birth time of log blocks (and dmu_sync()-ed blocks | |
872 | * that are in the log) to be arbitrarily large. | |
873 | */ | |
1c27024e | 874 | for (int i = 0; i < BP_GET_NDVAS(bp); i++) { |
63e3a861 | 875 | uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]); |
1c27024e | 876 | |
63e3a861 MA |
877 | if (vdevid >= spa->spa_root_vdev->vdev_children) { |
878 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
879 | "VDEV %llu", | |
880 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 881 | continue; |
63e3a861 | 882 | } |
1c27024e | 883 | vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; |
63e3a861 MA |
884 | if (vd == NULL) { |
885 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
886 | "VDEV %llu", | |
887 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 888 | continue; |
63e3a861 MA |
889 | } |
890 | if (vd->vdev_ops == &vdev_hole_ops) { | |
891 | zfs_panic_recover("blkptr at %p DVA %u has hole " | |
892 | "VDEV %llu", | |
893 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 894 | continue; |
63e3a861 MA |
895 | } |
896 | if (vd->vdev_ops == &vdev_missing_ops) { | |
897 | /* | |
898 | * "missing" vdevs are valid during import, but we | |
899 | * don't have their detailed info (e.g. asize), so | |
900 | * we can't perform any more checks on them. | |
901 | */ | |
902 | continue; | |
903 | } | |
1c27024e DB |
904 | uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]); |
905 | uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]); | |
63e3a861 MA |
906 | if (BP_IS_GANG(bp)) |
907 | asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
908 | if (offset + asize > vd->vdev_asize) { | |
909 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
910 | "OFFSET %llu", | |
911 | bp, i, (longlong_t)offset); | |
912 | } | |
913 | } | |
914 | } | |
915 | ||
34dc7c2f | 916 | zio_t * |
b128c09f | 917 | zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, |
a6255b7f | 918 | abd_t *data, uint64_t size, zio_done_func_t *done, void *private, |
5dbd68a3 | 919 | zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) |
34dc7c2f BB |
920 | { |
921 | zio_t *zio; | |
922 | ||
63e3a861 MA |
923 | zfs_blkptr_verify(spa, bp); |
924 | ||
428870ff | 925 | zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, |
2aa34383 | 926 | data, size, size, done, private, |
b128c09f | 927 | ZIO_TYPE_READ, priority, flags, NULL, 0, zb, |
428870ff BB |
928 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
929 | ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); | |
34dc7c2f | 930 | |
b128c09f BB |
931 | return (zio); |
932 | } | |
34dc7c2f | 933 | |
34dc7c2f | 934 | zio_t * |
b128c09f | 935 | zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, |
a6255b7f | 936 | abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp, |
bc77ba73 PD |
937 | zio_done_func_t *ready, zio_done_func_t *children_ready, |
938 | zio_done_func_t *physdone, zio_done_func_t *done, | |
939 | void *private, zio_priority_t priority, enum zio_flag flags, | |
940 | const zbookmark_phys_t *zb) | |
34dc7c2f BB |
941 | { |
942 | zio_t *zio; | |
943 | ||
b128c09f BB |
944 | ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && |
945 | zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && | |
946 | zp->zp_compress >= ZIO_COMPRESS_OFF && | |
947 | zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && | |
9ae529ec | 948 | DMU_OT_IS_VALID(zp->zp_type) && |
b128c09f | 949 | zp->zp_level < 32 && |
428870ff | 950 | zp->zp_copies > 0 && |
03c6040b | 951 | zp->zp_copies <= spa_max_replication(spa)); |
34dc7c2f | 952 | |
2aa34383 | 953 | zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private, |
b128c09f | 954 | ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, |
428870ff BB |
955 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
956 | ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); | |
34dc7c2f BB |
957 | |
958 | zio->io_ready = ready; | |
bc77ba73 | 959 | zio->io_children_ready = children_ready; |
e8b96c60 | 960 | zio->io_physdone = physdone; |
b128c09f | 961 | zio->io_prop = *zp; |
34dc7c2f | 962 | |
9b67f605 MA |
963 | /* |
964 | * Data can be NULL if we are going to call zio_write_override() to | |
965 | * provide the already-allocated BP. But we may need the data to | |
966 | * verify a dedup hit (if requested). In this case, don't try to | |
b5256303 TC |
967 | * dedup (just take the already-allocated BP verbatim). Encrypted |
968 | * dedup blocks need data as well so we also disable dedup in this | |
969 | * case. | |
9b67f605 | 970 | */ |
b5256303 TC |
971 | if (data == NULL && |
972 | (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) { | |
9b67f605 MA |
973 | zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; |
974 | } | |
975 | ||
34dc7c2f BB |
976 | return (zio); |
977 | } | |
978 | ||
979 | zio_t * | |
a6255b7f | 980 | zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data, |
e8b96c60 | 981 | uint64_t size, zio_done_func_t *done, void *private, |
5dbd68a3 | 982 | zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) |
34dc7c2f BB |
983 | { |
984 | zio_t *zio; | |
985 | ||
2aa34383 | 986 | zio = zio_create(pio, spa, txg, bp, data, size, size, done, private, |
3dfb57a3 | 987 | ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb, |
b128c09f | 988 | ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); |
34dc7c2f BB |
989 | |
990 | return (zio); | |
991 | } | |
992 | ||
428870ff | 993 | void |
03c6040b | 994 | zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) |
428870ff BB |
995 | { |
996 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
997 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
998 | ASSERT(zio->io_stage == ZIO_STAGE_OPEN); | |
999 | ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); | |
1000 | ||
03c6040b GW |
1001 | /* |
1002 | * We must reset the io_prop to match the values that existed | |
1003 | * when the bp was first written by dmu_sync() keeping in mind | |
1004 | * that nopwrite and dedup are mutually exclusive. | |
1005 | */ | |
1006 | zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; | |
1007 | zio->io_prop.zp_nopwrite = nopwrite; | |
428870ff BB |
1008 | zio->io_prop.zp_copies = copies; |
1009 | zio->io_bp_override = bp; | |
1010 | } | |
1011 | ||
1012 | void | |
1013 | zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) | |
1014 | { | |
9b67f605 MA |
1015 | |
1016 | /* | |
1017 | * The check for EMBEDDED is a performance optimization. We | |
1018 | * process the free here (by ignoring it) rather than | |
1019 | * putting it on the list and then processing it in zio_free_sync(). | |
1020 | */ | |
1021 | if (BP_IS_EMBEDDED(bp)) | |
1022 | return; | |
13fe0198 | 1023 | metaslab_check_free(spa, bp); |
2883cad5 MA |
1024 | |
1025 | /* | |
1026 | * Frees that are for the currently-syncing txg, are not going to be | |
1027 | * deferred, and which will not need to do a read (i.e. not GANG or | |
1028 | * DEDUP), can be processed immediately. Otherwise, put them on the | |
1029 | * in-memory list for later processing. | |
1030 | */ | |
1031 | if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || | |
1032 | txg != spa->spa_syncing_txg || | |
1033 | spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { | |
1034 | bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); | |
1035 | } else { | |
1036 | VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0))); | |
1037 | } | |
428870ff BB |
1038 | } |
1039 | ||
34dc7c2f | 1040 | zio_t * |
428870ff BB |
1041 | zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
1042 | enum zio_flag flags) | |
34dc7c2f BB |
1043 | { |
1044 | zio_t *zio; | |
2883cad5 | 1045 | enum zio_stage stage = ZIO_FREE_PIPELINE; |
34dc7c2f | 1046 | |
428870ff BB |
1047 | ASSERT(!BP_IS_HOLE(bp)); |
1048 | ASSERT(spa_syncing_txg(spa) == txg); | |
55d85d5a | 1049 | ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); |
34dc7c2f | 1050 | |
9b67f605 MA |
1051 | if (BP_IS_EMBEDDED(bp)) |
1052 | return (zio_null(pio, spa, NULL, NULL, NULL, 0)); | |
1053 | ||
13fe0198 | 1054 | metaslab_check_free(spa, bp); |
8c841793 | 1055 | arc_freed(spa, bp); |
d4a72f23 | 1056 | dsl_scan_freed(spa, bp); |
13fe0198 | 1057 | |
2883cad5 MA |
1058 | /* |
1059 | * GANG and DEDUP blocks can induce a read (for the gang block header, | |
1060 | * or the DDT), so issue them asynchronously so that this thread is | |
1061 | * not tied up. | |
1062 | */ | |
1063 | if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) | |
1064 | stage |= ZIO_STAGE_ISSUE_ASYNC; | |
1065 | ||
b128c09f | 1066 | zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
2aa34383 DK |
1067 | BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, |
1068 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage); | |
2883cad5 | 1069 | |
34dc7c2f BB |
1070 | return (zio); |
1071 | } | |
1072 | ||
1073 | zio_t * | |
428870ff BB |
1074 | zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
1075 | zio_done_func_t *done, void *private, enum zio_flag flags) | |
34dc7c2f BB |
1076 | { |
1077 | zio_t *zio; | |
1078 | ||
9b67f605 MA |
1079 | dprintf_bp(bp, "claiming in txg %llu", txg); |
1080 | ||
1081 | if (BP_IS_EMBEDDED(bp)) | |
1082 | return (zio_null(pio, spa, NULL, NULL, NULL, 0)); | |
1083 | ||
34dc7c2f BB |
1084 | /* |
1085 | * A claim is an allocation of a specific block. Claims are needed | |
1086 | * to support immediate writes in the intent log. The issue is that | |
1087 | * immediate writes contain committed data, but in a txg that was | |
1088 | * *not* committed. Upon opening the pool after an unclean shutdown, | |
1089 | * the intent log claims all blocks that contain immediate write data | |
1090 | * so that the SPA knows they're in use. | |
1091 | * | |
1092 | * All claims *must* be resolved in the first txg -- before the SPA | |
1093 | * starts allocating blocks -- so that nothing is allocated twice. | |
428870ff | 1094 | * If txg == 0 we just verify that the block is claimable. |
34dc7c2f BB |
1095 | */ |
1096 | ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); | |
428870ff BB |
1097 | ASSERT(txg == spa_first_txg(spa) || txg == 0); |
1098 | ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ | |
34dc7c2f | 1099 | |
b128c09f | 1100 | zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
2aa34383 DK |
1101 | BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, |
1102 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); | |
3dfb57a3 | 1103 | ASSERT0(zio->io_queued_timestamp); |
34dc7c2f BB |
1104 | |
1105 | return (zio); | |
1106 | } | |
1107 | ||
1108 | zio_t * | |
1109 | zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, | |
e8b96c60 | 1110 | zio_done_func_t *done, void *private, enum zio_flag flags) |
34dc7c2f BB |
1111 | { |
1112 | zio_t *zio; | |
1113 | int c; | |
1114 | ||
1115 | if (vd->vdev_children == 0) { | |
2aa34383 | 1116 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
e8b96c60 | 1117 | ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
34dc7c2f BB |
1118 | ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); |
1119 | ||
34dc7c2f BB |
1120 | zio->io_cmd = cmd; |
1121 | } else { | |
d164b209 | 1122 | zio = zio_null(pio, spa, NULL, NULL, NULL, flags); |
34dc7c2f BB |
1123 | |
1124 | for (c = 0; c < vd->vdev_children; c++) | |
1125 | zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, | |
e8b96c60 | 1126 | done, private, flags)); |
34dc7c2f BB |
1127 | } |
1128 | ||
1129 | return (zio); | |
1130 | } | |
1131 | ||
34dc7c2f BB |
1132 | zio_t * |
1133 | zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
a6255b7f | 1134 | abd_t *data, int checksum, zio_done_func_t *done, void *private, |
e8b96c60 | 1135 | zio_priority_t priority, enum zio_flag flags, boolean_t labels) |
34dc7c2f BB |
1136 | { |
1137 | zio_t *zio; | |
34dc7c2f | 1138 | |
b128c09f BB |
1139 | ASSERT(vd->vdev_children == 0); |
1140 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
1141 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
1142 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 1143 | |
2aa34383 DK |
1144 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
1145 | private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
1146 | offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); | |
34dc7c2f | 1147 | |
b128c09f | 1148 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f BB |
1149 | |
1150 | return (zio); | |
1151 | } | |
1152 | ||
1153 | zio_t * | |
1154 | zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
a6255b7f | 1155 | abd_t *data, int checksum, zio_done_func_t *done, void *private, |
e8b96c60 | 1156 | zio_priority_t priority, enum zio_flag flags, boolean_t labels) |
34dc7c2f | 1157 | { |
34dc7c2f | 1158 | zio_t *zio; |
34dc7c2f | 1159 | |
b128c09f BB |
1160 | ASSERT(vd->vdev_children == 0); |
1161 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
1162 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
1163 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 1164 | |
2aa34383 DK |
1165 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
1166 | private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
1167 | offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); | |
34dc7c2f | 1168 | |
b128c09f | 1169 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f | 1170 | |
3c67d83a | 1171 | if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { |
34dc7c2f | 1172 | /* |
428870ff | 1173 | * zec checksums are necessarily destructive -- they modify |
b128c09f | 1174 | * the end of the write buffer to hold the verifier/checksum. |
34dc7c2f | 1175 | * Therefore, we must make a local copy in case the data is |
b128c09f | 1176 | * being written to multiple places in parallel. |
34dc7c2f | 1177 | */ |
a6255b7f DQ |
1178 | abd_t *wbuf = abd_alloc_sametype(data, size); |
1179 | abd_copy(wbuf, data, size); | |
1180 | ||
b128c09f | 1181 | zio_push_transform(zio, wbuf, size, size, NULL); |
34dc7c2f BB |
1182 | } |
1183 | ||
1184 | return (zio); | |
1185 | } | |
1186 | ||
1187 | /* | |
b128c09f | 1188 | * Create a child I/O to do some work for us. |
34dc7c2f BB |
1189 | */ |
1190 | zio_t * | |
b128c09f | 1191 | zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, |
4ea3f864 GM |
1192 | abd_t *data, uint64_t size, int type, zio_priority_t priority, |
1193 | enum zio_flag flags, zio_done_func_t *done, void *private) | |
34dc7c2f | 1194 | { |
428870ff | 1195 | enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; |
b128c09f BB |
1196 | zio_t *zio; |
1197 | ||
1198 | ASSERT(vd->vdev_parent == | |
1199 | (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev)); | |
34dc7c2f BB |
1200 | |
1201 | if (type == ZIO_TYPE_READ && bp != NULL) { | |
1202 | /* | |
1203 | * If we have the bp, then the child should perform the | |
1204 | * checksum and the parent need not. This pushes error | |
1205 | * detection as close to the leaves as possible and | |
1206 | * eliminates redundant checksums in the interior nodes. | |
1207 | */ | |
428870ff BB |
1208 | pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; |
1209 | pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; | |
34dc7c2f BB |
1210 | } |
1211 | ||
b128c09f BB |
1212 | if (vd->vdev_children == 0) |
1213 | offset += VDEV_LABEL_START_SIZE; | |
1214 | ||
428870ff BB |
1215 | flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE; |
1216 | ||
1217 | /* | |
1218 | * If we've decided to do a repair, the write is not speculative -- | |
1219 | * even if the original read was. | |
1220 | */ | |
1221 | if (flags & ZIO_FLAG_IO_REPAIR) | |
1222 | flags &= ~ZIO_FLAG_SPECULATIVE; | |
1223 | ||
3dfb57a3 DB |
1224 | /* |
1225 | * If we're creating a child I/O that is not associated with a | |
1226 | * top-level vdev, then the child zio is not an allocating I/O. | |
1227 | * If this is a retried I/O then we ignore it since we will | |
1228 | * have already processed the original allocating I/O. | |
1229 | */ | |
1230 | if (flags & ZIO_FLAG_IO_ALLOCATING && | |
1231 | (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) { | |
6d00b5e1 | 1232 | ASSERTV(metaslab_class_t *mc = spa_normal_class(pio->io_spa)); |
3dfb57a3 DB |
1233 | |
1234 | ASSERT(mc->mc_alloc_throttle_enabled); | |
1235 | ASSERT(type == ZIO_TYPE_WRITE); | |
1236 | ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE); | |
1237 | ASSERT(!(flags & ZIO_FLAG_IO_REPAIR)); | |
1238 | ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) || | |
1239 | pio->io_child_type == ZIO_CHILD_GANG); | |
1240 | ||
1241 | flags &= ~ZIO_FLAG_IO_ALLOCATING; | |
1242 | } | |
1243 | ||
1244 | ||
2aa34383 | 1245 | zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size, |
428870ff BB |
1246 | done, private, type, priority, flags, vd, offset, &pio->io_bookmark, |
1247 | ZIO_STAGE_VDEV_IO_START >> 1, pipeline); | |
3dfb57a3 | 1248 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); |
34dc7c2f | 1249 | |
e8b96c60 MA |
1250 | zio->io_physdone = pio->io_physdone; |
1251 | if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) | |
1252 | zio->io_logical->io_phys_children++; | |
1253 | ||
b128c09f | 1254 | return (zio); |
34dc7c2f BB |
1255 | } |
1256 | ||
b128c09f | 1257 | zio_t * |
a6255b7f | 1258 | zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size, |
e9aa730c GM |
1259 | int type, zio_priority_t priority, enum zio_flag flags, |
1260 | zio_done_func_t *done, void *private) | |
34dc7c2f | 1261 | { |
b128c09f | 1262 | zio_t *zio; |
34dc7c2f | 1263 | |
b128c09f | 1264 | ASSERT(vd->vdev_ops->vdev_op_leaf); |
34dc7c2f | 1265 | |
b128c09f | 1266 | zio = zio_create(NULL, vd->vdev_spa, 0, NULL, |
2aa34383 | 1267 | data, size, size, done, private, type, priority, |
e8b96c60 | 1268 | flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, |
b128c09f | 1269 | vd, offset, NULL, |
428870ff | 1270 | ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); |
34dc7c2f | 1271 | |
b128c09f | 1272 | return (zio); |
34dc7c2f BB |
1273 | } |
1274 | ||
1275 | void | |
b128c09f | 1276 | zio_flush(zio_t *zio, vdev_t *vd) |
34dc7c2f | 1277 | { |
b128c09f | 1278 | zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, |
e8b96c60 | 1279 | NULL, NULL, |
b128c09f | 1280 | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); |
34dc7c2f BB |
1281 | } |
1282 | ||
428870ff BB |
1283 | void |
1284 | zio_shrink(zio_t *zio, uint64_t size) | |
1285 | { | |
1ce23dca PS |
1286 | ASSERT3P(zio->io_executor, ==, NULL); |
1287 | ASSERT3U(zio->io_orig_size, ==, zio->io_size); | |
1288 | ASSERT3U(size, <=, zio->io_size); | |
428870ff BB |
1289 | |
1290 | /* | |
1291 | * We don't shrink for raidz because of problems with the | |
1292 | * reconstruction when reading back less than the block size. | |
1293 | * Note, BP_IS_RAIDZ() assumes no compression. | |
1294 | */ | |
1295 | ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); | |
2aa34383 DK |
1296 | if (!BP_IS_RAIDZ(zio->io_bp)) { |
1297 | /* we are not doing a raw write */ | |
1298 | ASSERT3U(zio->io_size, ==, zio->io_lsize); | |
1299 | zio->io_orig_size = zio->io_size = zio->io_lsize = size; | |
1300 | } | |
428870ff BB |
1301 | } |
1302 | ||
34dc7c2f BB |
1303 | /* |
1304 | * ========================================================================== | |
b128c09f | 1305 | * Prepare to read and write logical blocks |
34dc7c2f BB |
1306 | * ========================================================================== |
1307 | */ | |
b128c09f | 1308 | |
34dc7c2f | 1309 | static int |
b128c09f | 1310 | zio_read_bp_init(zio_t *zio) |
34dc7c2f | 1311 | { |
b128c09f | 1312 | blkptr_t *bp = zio->io_bp; |
b5256303 TC |
1313 | uint64_t psize = |
1314 | BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); | |
34dc7c2f | 1315 | |
fb5f0bc8 | 1316 | if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && |
9babb374 | 1317 | zio->io_child_type == ZIO_CHILD_LOGICAL && |
b5256303 | 1318 | !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { |
a6255b7f DQ |
1319 | zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), |
1320 | psize, psize, zio_decompress); | |
34dc7c2f | 1321 | } |
34dc7c2f | 1322 | |
b5256303 TC |
1323 | if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) || |
1324 | BP_HAS_INDIRECT_MAC_CKSUM(bp)) && | |
1325 | zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
1326 | zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), | |
1327 | psize, psize, zio_decrypt); | |
1328 | } | |
1329 | ||
9b67f605 | 1330 | if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { |
a6255b7f DQ |
1331 | int psize = BPE_GET_PSIZE(bp); |
1332 | void *data = abd_borrow_buf(zio->io_abd, psize); | |
1333 | ||
9b67f605 | 1334 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
a6255b7f DQ |
1335 | decode_embedded_bp_compressed(bp, data); |
1336 | abd_return_buf_copy(zio->io_abd, data, psize); | |
9b67f605 MA |
1337 | } else { |
1338 | ASSERT(!BP_IS_EMBEDDED(bp)); | |
1339 | } | |
1340 | ||
9ae529ec | 1341 | if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) |
b128c09f BB |
1342 | zio->io_flags |= ZIO_FLAG_DONT_CACHE; |
1343 | ||
428870ff BB |
1344 | if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) |
1345 | zio->io_flags |= ZIO_FLAG_DONT_CACHE; | |
1346 | ||
1347 | if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) | |
1348 | zio->io_pipeline = ZIO_DDT_READ_PIPELINE; | |
1349 | ||
b128c09f | 1350 | return (ZIO_PIPELINE_CONTINUE); |
34dc7c2f BB |
1351 | } |
1352 | ||
b128c09f BB |
1353 | static int |
1354 | zio_write_bp_init(zio_t *zio) | |
34dc7c2f | 1355 | { |
b128c09f BB |
1356 | if (!IO_IS_ALLOCATING(zio)) |
1357 | return (ZIO_PIPELINE_CONTINUE); | |
34dc7c2f | 1358 | |
428870ff BB |
1359 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1360 | ||
1361 | if (zio->io_bp_override) { | |
3dfb57a3 DB |
1362 | blkptr_t *bp = zio->io_bp; |
1363 | zio_prop_t *zp = &zio->io_prop; | |
1364 | ||
428870ff BB |
1365 | ASSERT(bp->blk_birth != zio->io_txg); |
1366 | ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); | |
1367 | ||
1368 | *bp = *zio->io_bp_override; | |
1369 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1370 | ||
9b67f605 MA |
1371 | if (BP_IS_EMBEDDED(bp)) |
1372 | return (ZIO_PIPELINE_CONTINUE); | |
1373 | ||
03c6040b GW |
1374 | /* |
1375 | * If we've been overridden and nopwrite is set then | |
1376 | * set the flag accordingly to indicate that a nopwrite | |
1377 | * has already occurred. | |
1378 | */ | |
1379 | if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { | |
1380 | ASSERT(!zp->zp_dedup); | |
3dfb57a3 | 1381 | ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum); |
03c6040b GW |
1382 | zio->io_flags |= ZIO_FLAG_NOPWRITE; |
1383 | return (ZIO_PIPELINE_CONTINUE); | |
1384 | } | |
1385 | ||
1386 | ASSERT(!zp->zp_nopwrite); | |
1387 | ||
428870ff BB |
1388 | if (BP_IS_HOLE(bp) || !zp->zp_dedup) |
1389 | return (ZIO_PIPELINE_CONTINUE); | |
1390 | ||
3c67d83a TH |
1391 | ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & |
1392 | ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); | |
428870ff | 1393 | |
b5256303 TC |
1394 | if (BP_GET_CHECKSUM(bp) == zp->zp_checksum && |
1395 | !zp->zp_encrypt) { | |
428870ff BB |
1396 | BP_SET_DEDUP(bp, 1); |
1397 | zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; | |
1398 | return (ZIO_PIPELINE_CONTINUE); | |
1399 | } | |
3dfb57a3 DB |
1400 | |
1401 | /* | |
1402 | * We were unable to handle this as an override bp, treat | |
1403 | * it as a regular write I/O. | |
1404 | */ | |
5511754b | 1405 | zio->io_bp_override = NULL; |
3dfb57a3 DB |
1406 | *bp = zio->io_bp_orig; |
1407 | zio->io_pipeline = zio->io_orig_pipeline; | |
1408 | } | |
1409 | ||
1410 | return (ZIO_PIPELINE_CONTINUE); | |
1411 | } | |
1412 | ||
1413 | static int | |
1414 | zio_write_compress(zio_t *zio) | |
1415 | { | |
1416 | spa_t *spa = zio->io_spa; | |
1417 | zio_prop_t *zp = &zio->io_prop; | |
1418 | enum zio_compress compress = zp->zp_compress; | |
1419 | blkptr_t *bp = zio->io_bp; | |
1420 | uint64_t lsize = zio->io_lsize; | |
1421 | uint64_t psize = zio->io_size; | |
1422 | int pass = 1; | |
1423 | ||
3dfb57a3 DB |
1424 | /* |
1425 | * If our children haven't all reached the ready stage, | |
1426 | * wait for them and then repeat this pipeline stage. | |
1427 | */ | |
1428 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || | |
1429 | zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) | |
1430 | return (ZIO_PIPELINE_STOP); | |
1431 | ||
1432 | if (!IO_IS_ALLOCATING(zio)) | |
1433 | return (ZIO_PIPELINE_CONTINUE); | |
1434 | ||
1435 | if (zio->io_children_ready != NULL) { | |
1436 | /* | |
1437 | * Now that all our children are ready, run the callback | |
1438 | * associated with this zio in case it wants to modify the | |
1439 | * data to be written. | |
1440 | */ | |
1441 | ASSERT3U(zp->zp_level, >, 0); | |
1442 | zio->io_children_ready(zio); | |
428870ff | 1443 | } |
34dc7c2f | 1444 | |
3dfb57a3 DB |
1445 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1446 | ASSERT(zio->io_bp_override == NULL); | |
1447 | ||
b0bc7a84 | 1448 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { |
b128c09f BB |
1449 | /* |
1450 | * We're rewriting an existing block, which means we're | |
1451 | * working on behalf of spa_sync(). For spa_sync() to | |
1452 | * converge, it must eventually be the case that we don't | |
1453 | * have to allocate new blocks. But compression changes | |
1454 | * the blocksize, which forces a reallocate, and makes | |
1455 | * convergence take longer. Therefore, after the first | |
1456 | * few passes, stop compressing to ensure convergence. | |
1457 | */ | |
428870ff BB |
1458 | pass = spa_sync_pass(spa); |
1459 | ||
1460 | ASSERT(zio->io_txg == spa_syncing_txg(spa)); | |
1461 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1462 | ASSERT(!BP_GET_DEDUP(bp)); | |
34dc7c2f | 1463 | |
55d85d5a | 1464 | if (pass >= zfs_sync_pass_dont_compress) |
b128c09f | 1465 | compress = ZIO_COMPRESS_OFF; |
34dc7c2f | 1466 | |
b128c09f | 1467 | /* Make sure someone doesn't change their mind on overwrites */ |
9b67f605 | 1468 | ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), |
428870ff | 1469 | spa_max_replication(spa)) == BP_GET_NDVAS(bp)); |
b128c09f | 1470 | } |
34dc7c2f | 1471 | |
2aa34383 | 1472 | /* If it's a compressed write that is not raw, compress the buffer. */ |
b5256303 TC |
1473 | if (compress != ZIO_COMPRESS_OFF && |
1474 | !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { | |
428870ff | 1475 | void *cbuf = zio_buf_alloc(lsize); |
a6255b7f | 1476 | psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize); |
428870ff | 1477 | if (psize == 0 || psize == lsize) { |
b128c09f | 1478 | compress = ZIO_COMPRESS_OFF; |
428870ff | 1479 | zio_buf_free(cbuf, lsize); |
b5256303 TC |
1480 | } else if (!zp->zp_dedup && !zp->zp_encrypt && |
1481 | psize <= BPE_PAYLOAD_SIZE && | |
9b67f605 MA |
1482 | zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && |
1483 | spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { | |
1484 | encode_embedded_bp_compressed(bp, | |
1485 | cbuf, compress, lsize, psize); | |
1486 | BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); | |
1487 | BP_SET_TYPE(bp, zio->io_prop.zp_type); | |
1488 | BP_SET_LEVEL(bp, zio->io_prop.zp_level); | |
1489 | zio_buf_free(cbuf, lsize); | |
1490 | bp->blk_birth = zio->io_txg; | |
1491 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1492 | ASSERT(spa_feature_is_active(spa, | |
1493 | SPA_FEATURE_EMBEDDED_DATA)); | |
1494 | return (ZIO_PIPELINE_CONTINUE); | |
428870ff | 1495 | } else { |
9b67f605 | 1496 | /* |
c3520e7f MA |
1497 | * Round up compressed size up to the ashift |
1498 | * of the smallest-ashift device, and zero the tail. | |
1499 | * This ensures that the compressed size of the BP | |
1500 | * (and thus compressratio property) are correct, | |
1501 | * in that we charge for the padding used to fill out | |
1502 | * the last sector. | |
9b67f605 | 1503 | */ |
c3520e7f | 1504 | ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT); |
1c27024e | 1505 | size_t rounded = (size_t)P2ROUNDUP(psize, |
c3520e7f MA |
1506 | 1ULL << spa->spa_min_ashift); |
1507 | if (rounded >= lsize) { | |
9b67f605 MA |
1508 | compress = ZIO_COMPRESS_OFF; |
1509 | zio_buf_free(cbuf, lsize); | |
c3520e7f | 1510 | psize = lsize; |
9b67f605 | 1511 | } else { |
a6255b7f DQ |
1512 | abd_t *cdata = abd_get_from_buf(cbuf, lsize); |
1513 | abd_take_ownership_of_buf(cdata, B_TRUE); | |
1514 | abd_zero_off(cdata, psize, rounded - psize); | |
c3520e7f | 1515 | psize = rounded; |
a6255b7f | 1516 | zio_push_transform(zio, cdata, |
9b67f605 MA |
1517 | psize, lsize, NULL); |
1518 | } | |
b128c09f | 1519 | } |
3dfb57a3 DB |
1520 | |
1521 | /* | |
1522 | * We were unable to handle this as an override bp, treat | |
1523 | * it as a regular write I/O. | |
1524 | */ | |
1525 | zio->io_bp_override = NULL; | |
1526 | *bp = zio->io_bp_orig; | |
1527 | zio->io_pipeline = zio->io_orig_pipeline; | |
1528 | ||
2aa34383 DK |
1529 | } else { |
1530 | ASSERT3U(psize, !=, 0); | |
1531 | ||
b128c09f | 1532 | } |
34dc7c2f | 1533 | |
b128c09f BB |
1534 | /* |
1535 | * The final pass of spa_sync() must be all rewrites, but the first | |
1536 | * few passes offer a trade-off: allocating blocks defers convergence, | |
1537 | * but newly allocated blocks are sequential, so they can be written | |
1538 | * to disk faster. Therefore, we allow the first few passes of | |
1539 | * spa_sync() to allocate new blocks, but force rewrites after that. | |
1540 | * There should only be a handful of blocks after pass 1 in any case. | |
1541 | */ | |
b0bc7a84 MG |
1542 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && |
1543 | BP_GET_PSIZE(bp) == psize && | |
55d85d5a | 1544 | pass >= zfs_sync_pass_rewrite) { |
d6320ddb | 1545 | ASSERT(psize != 0); |
1c27024e | 1546 | enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; |
b128c09f BB |
1547 | zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; |
1548 | zio->io_flags |= ZIO_FLAG_IO_REWRITE; | |
1549 | } else { | |
1550 | BP_ZERO(bp); | |
1551 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
1552 | } | |
34dc7c2f | 1553 | |
428870ff | 1554 | if (psize == 0) { |
b0bc7a84 MG |
1555 | if (zio->io_bp_orig.blk_birth != 0 && |
1556 | spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { | |
1557 | BP_SET_LSIZE(bp, lsize); | |
1558 | BP_SET_TYPE(bp, zp->zp_type); | |
1559 | BP_SET_LEVEL(bp, zp->zp_level); | |
1560 | BP_SET_BIRTH(bp, zio->io_txg, 0); | |
1561 | } | |
b128c09f BB |
1562 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
1563 | } else { | |
1564 | ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); | |
1565 | BP_SET_LSIZE(bp, lsize); | |
b0bc7a84 MG |
1566 | BP_SET_TYPE(bp, zp->zp_type); |
1567 | BP_SET_LEVEL(bp, zp->zp_level); | |
428870ff | 1568 | BP_SET_PSIZE(bp, psize); |
b128c09f BB |
1569 | BP_SET_COMPRESS(bp, compress); |
1570 | BP_SET_CHECKSUM(bp, zp->zp_checksum); | |
428870ff | 1571 | BP_SET_DEDUP(bp, zp->zp_dedup); |
b128c09f | 1572 | BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); |
428870ff BB |
1573 | if (zp->zp_dedup) { |
1574 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1575 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
b5256303 TC |
1576 | ASSERT(!zp->zp_encrypt || |
1577 | DMU_OT_IS_ENCRYPTED(zp->zp_type)); | |
428870ff BB |
1578 | zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; |
1579 | } | |
03c6040b GW |
1580 | if (zp->zp_nopwrite) { |
1581 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1582 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
1583 | zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; | |
1584 | } | |
428870ff | 1585 | } |
428870ff BB |
1586 | return (ZIO_PIPELINE_CONTINUE); |
1587 | } | |
1588 | ||
1589 | static int | |
1590 | zio_free_bp_init(zio_t *zio) | |
1591 | { | |
1592 | blkptr_t *bp = zio->io_bp; | |
1593 | ||
1594 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
1595 | if (BP_GET_DEDUP(bp)) | |
1596 | zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; | |
b128c09f | 1597 | } |
34dc7c2f BB |
1598 | |
1599 | return (ZIO_PIPELINE_CONTINUE); | |
1600 | } | |
1601 | ||
b128c09f BB |
1602 | /* |
1603 | * ========================================================================== | |
1604 | * Execute the I/O pipeline | |
1605 | * ========================================================================== | |
1606 | */ | |
1607 | ||
1608 | static void | |
7ef5e54e | 1609 | zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) |
34dc7c2f | 1610 | { |
428870ff | 1611 | spa_t *spa = zio->io_spa; |
b128c09f | 1612 | zio_type_t t = zio->io_type; |
a38718a6 | 1613 | int flags = (cutinline ? TQ_FRONT : 0); |
34dc7c2f BB |
1614 | |
1615 | /* | |
9babb374 BB |
1616 | * If we're a config writer or a probe, the normal issue and |
1617 | * interrupt threads may all be blocked waiting for the config lock. | |
1618 | * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. | |
34dc7c2f | 1619 | */ |
9babb374 | 1620 | if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) |
b128c09f | 1621 | t = ZIO_TYPE_NULL; |
34dc7c2f BB |
1622 | |
1623 | /* | |
b128c09f | 1624 | * A similar issue exists for the L2ARC write thread until L2ARC 2.0. |
34dc7c2f | 1625 | */ |
b128c09f BB |
1626 | if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) |
1627 | t = ZIO_TYPE_NULL; | |
34dc7c2f | 1628 | |
428870ff | 1629 | /* |
7ef5e54e AL |
1630 | * If this is a high priority I/O, then use the high priority taskq if |
1631 | * available. | |
428870ff BB |
1632 | */ |
1633 | if (zio->io_priority == ZIO_PRIORITY_NOW && | |
7ef5e54e | 1634 | spa->spa_zio_taskq[t][q + 1].stqs_count != 0) |
428870ff BB |
1635 | q++; |
1636 | ||
1637 | ASSERT3U(q, <, ZIO_TASKQ_TYPES); | |
5cc556b4 | 1638 | |
a38718a6 GA |
1639 | /* |
1640 | * NB: We are assuming that the zio can only be dispatched | |
1641 | * to a single taskq at a time. It would be a grievous error | |
1642 | * to dispatch the zio to another taskq at the same time. | |
1643 | */ | |
1644 | ASSERT(taskq_empty_ent(&zio->io_tqent)); | |
7ef5e54e AL |
1645 | spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, |
1646 | flags, &zio->io_tqent); | |
b128c09f | 1647 | } |
34dc7c2f | 1648 | |
b128c09f | 1649 | static boolean_t |
7ef5e54e | 1650 | zio_taskq_member(zio_t *zio, zio_taskq_type_t q) |
b128c09f BB |
1651 | { |
1652 | kthread_t *executor = zio->io_executor; | |
1653 | spa_t *spa = zio->io_spa; | |
34dc7c2f | 1654 | |
1c27024e | 1655 | for (zio_type_t t = 0; t < ZIO_TYPES; t++) { |
7ef5e54e AL |
1656 | spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; |
1657 | uint_t i; | |
1658 | for (i = 0; i < tqs->stqs_count; i++) { | |
1659 | if (taskq_member(tqs->stqs_taskq[i], executor)) | |
1660 | return (B_TRUE); | |
1661 | } | |
1662 | } | |
34dc7c2f | 1663 | |
b128c09f BB |
1664 | return (B_FALSE); |
1665 | } | |
34dc7c2f | 1666 | |
b128c09f BB |
1667 | static int |
1668 | zio_issue_async(zio_t *zio) | |
1669 | { | |
428870ff | 1670 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); |
b128c09f BB |
1671 | |
1672 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f BB |
1673 | } |
1674 | ||
b128c09f BB |
1675 | void |
1676 | zio_interrupt(zio_t *zio) | |
34dc7c2f | 1677 | { |
428870ff | 1678 | zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); |
b128c09f | 1679 | } |
34dc7c2f | 1680 | |
26ef0cc7 TH |
1681 | void |
1682 | zio_delay_interrupt(zio_t *zio) | |
1683 | { | |
1684 | /* | |
1685 | * The timeout_generic() function isn't defined in userspace, so | |
1686 | * rather than trying to implement the function, the zio delay | |
1687 | * functionality has been disabled for userspace builds. | |
1688 | */ | |
1689 | ||
1690 | #ifdef _KERNEL | |
1691 | /* | |
1692 | * If io_target_timestamp is zero, then no delay has been registered | |
1693 | * for this IO, thus jump to the end of this function and "skip" the | |
1694 | * delay; issuing it directly to the zio layer. | |
1695 | */ | |
1696 | if (zio->io_target_timestamp != 0) { | |
1697 | hrtime_t now = gethrtime(); | |
1698 | ||
1699 | if (now >= zio->io_target_timestamp) { | |
1700 | /* | |
1701 | * This IO has already taken longer than the target | |
1702 | * delay to complete, so we don't want to delay it | |
1703 | * any longer; we "miss" the delay and issue it | |
1704 | * directly to the zio layer. This is likely due to | |
1705 | * the target latency being set to a value less than | |
1706 | * the underlying hardware can satisfy (e.g. delay | |
1707 | * set to 1ms, but the disks take 10ms to complete an | |
1708 | * IO request). | |
1709 | */ | |
1710 | ||
1711 | DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, | |
1712 | hrtime_t, now); | |
1713 | ||
1714 | zio_interrupt(zio); | |
1715 | } else { | |
1716 | taskqid_t tid; | |
1717 | hrtime_t diff = zio->io_target_timestamp - now; | |
1718 | clock_t expire_at_tick = ddi_get_lbolt() + | |
1719 | NSEC_TO_TICK(diff); | |
1720 | ||
1721 | DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, | |
1722 | hrtime_t, now, hrtime_t, diff); | |
1723 | ||
1724 | if (NSEC_TO_TICK(diff) == 0) { | |
1725 | /* Our delay is less than a jiffy - just spin */ | |
1726 | zfs_sleep_until(zio->io_target_timestamp); | |
1727 | } else { | |
1728 | /* | |
1729 | * Use taskq_dispatch_delay() in the place of | |
1730 | * OpenZFS's timeout_generic(). | |
1731 | */ | |
1732 | tid = taskq_dispatch_delay(system_taskq, | |
02730c33 | 1733 | (task_func_t *)zio_interrupt, |
26ef0cc7 | 1734 | zio, TQ_NOSLEEP, expire_at_tick); |
48d3eb40 | 1735 | if (tid == TASKQID_INVALID) { |
26ef0cc7 TH |
1736 | /* |
1737 | * Couldn't allocate a task. Just | |
1738 | * finish the zio without a delay. | |
1739 | */ | |
1740 | zio_interrupt(zio); | |
1741 | } | |
1742 | } | |
1743 | } | |
1744 | return; | |
1745 | } | |
1746 | #endif | |
1747 | DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); | |
1748 | zio_interrupt(zio); | |
1749 | } | |
1750 | ||
8fb1ede1 BB |
1751 | static void |
1752 | zio_deadman_impl(zio_t *pio) | |
1753 | { | |
1754 | zio_t *cio, *cio_next; | |
1755 | zio_link_t *zl = NULL; | |
1756 | vdev_t *vd = pio->io_vd; | |
1757 | ||
1758 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { | |
1759 | vdev_queue_t *vq = &vd->vdev_queue; | |
1760 | zbookmark_phys_t *zb = &pio->io_bookmark; | |
1761 | uint64_t delta = gethrtime() - pio->io_timestamp; | |
1762 | uint64_t failmode = spa_get_deadman_failmode(pio->io_spa); | |
1763 | ||
1764 | zfs_dbgmsg("slow zio: zio=%p timestamp=%llu " | |
1765 | "delta=%llu queued=%llu io=%llu " | |
1766 | "path=%s last=%llu " | |
1767 | "type=%d priority=%d flags=0x%x " | |
1768 | "stage=0x%x pipeline=0x%x pipeline-trace=0x%x " | |
1769 | "objset=%llu object=%llu level=%llu blkid=%llu " | |
1770 | "offset=%llu size=%llu error=%d", | |
1771 | pio, pio->io_timestamp, | |
1772 | delta, pio->io_delta, pio->io_delay, | |
1773 | vd->vdev_path, vq->vq_io_complete_ts, | |
1774 | pio->io_type, pio->io_priority, pio->io_flags, | |
1775 | pio->io_state, pio->io_pipeline, pio->io_pipeline_trace, | |
1776 | zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid, | |
1777 | pio->io_offset, pio->io_size, pio->io_error); | |
1778 | zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN, | |
1779 | pio->io_spa, vd, zb, pio, 0, 0); | |
1780 | ||
1781 | if (failmode == ZIO_FAILURE_MODE_CONTINUE && | |
1782 | taskq_empty_ent(&pio->io_tqent)) { | |
1783 | zio_interrupt(pio); | |
1784 | } | |
1785 | } | |
1786 | ||
1787 | mutex_enter(&pio->io_lock); | |
1788 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { | |
1789 | cio_next = zio_walk_children(pio, &zl); | |
1790 | zio_deadman_impl(cio); | |
1791 | } | |
1792 | mutex_exit(&pio->io_lock); | |
1793 | } | |
1794 | ||
1795 | /* | |
1796 | * Log the critical information describing this zio and all of its children | |
1797 | * using the zfs_dbgmsg() interface then post deadman event for the ZED. | |
1798 | */ | |
1799 | void | |
1800 | zio_deadman(zio_t *pio, char *tag) | |
1801 | { | |
1802 | spa_t *spa = pio->io_spa; | |
1803 | char *name = spa_name(spa); | |
1804 | ||
1805 | if (!zfs_deadman_enabled || spa_suspended(spa)) | |
1806 | return; | |
1807 | ||
1808 | zio_deadman_impl(pio); | |
1809 | ||
1810 | switch (spa_get_deadman_failmode(spa)) { | |
1811 | case ZIO_FAILURE_MODE_WAIT: | |
1812 | zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name); | |
1813 | break; | |
1814 | ||
1815 | case ZIO_FAILURE_MODE_CONTINUE: | |
1816 | zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name); | |
1817 | break; | |
1818 | ||
1819 | case ZIO_FAILURE_MODE_PANIC: | |
1820 | fm_panic("%s determined I/O to pool '%s' is hung.", tag, name); | |
1821 | break; | |
1822 | } | |
1823 | } | |
1824 | ||
b128c09f BB |
1825 | /* |
1826 | * Execute the I/O pipeline until one of the following occurs: | |
1827 | * (1) the I/O completes; (2) the pipeline stalls waiting for | |
1828 | * dependent child I/Os; (3) the I/O issues, so we're waiting | |
1829 | * for an I/O completion interrupt; (4) the I/O is delegated by | |
1830 | * vdev-level caching or aggregation; (5) the I/O is deferred | |
1831 | * due to vdev-level queueing; (6) the I/O is handed off to | |
1832 | * another thread. In all cases, the pipeline stops whenever | |
8e07b99b | 1833 | * there's no CPU work; it never burns a thread in cv_wait_io(). |
b128c09f BB |
1834 | * |
1835 | * There's no locking on io_stage because there's no legitimate way | |
1836 | * for multiple threads to be attempting to process the same I/O. | |
1837 | */ | |
428870ff | 1838 | static zio_pipe_stage_t *zio_pipeline[]; |
34dc7c2f | 1839 | |
da6b4005 NB |
1840 | /* |
1841 | * zio_execute() is a wrapper around the static function | |
1842 | * __zio_execute() so that we can force __zio_execute() to be | |
1843 | * inlined. This reduces stack overhead which is important | |
1844 | * because __zio_execute() is called recursively in several zio | |
1845 | * code paths. zio_execute() itself cannot be inlined because | |
1846 | * it is externally visible. | |
1847 | */ | |
b128c09f BB |
1848 | void |
1849 | zio_execute(zio_t *zio) | |
da6b4005 | 1850 | { |
92119cc2 BB |
1851 | fstrans_cookie_t cookie; |
1852 | ||
1853 | cookie = spl_fstrans_mark(); | |
da6b4005 | 1854 | __zio_execute(zio); |
92119cc2 | 1855 | spl_fstrans_unmark(cookie); |
da6b4005 NB |
1856 | } |
1857 | ||
b58986ee BB |
1858 | /* |
1859 | * Used to determine if in the current context the stack is sized large | |
1860 | * enough to allow zio_execute() to be called recursively. A minimum | |
1861 | * stack size of 16K is required to avoid needing to re-dispatch the zio. | |
1862 | */ | |
1863 | boolean_t | |
1864 | zio_execute_stack_check(zio_t *zio) | |
1865 | { | |
1866 | #if !defined(HAVE_LARGE_STACKS) | |
1867 | dsl_pool_t *dp = spa_get_dsl(zio->io_spa); | |
1868 | ||
1869 | /* Executing in txg_sync_thread() context. */ | |
1870 | if (dp && curthread == dp->dp_tx.tx_sync_thread) | |
1871 | return (B_TRUE); | |
1872 | ||
1873 | /* Pool initialization outside of zio_taskq context. */ | |
1874 | if (dp && spa_is_initializing(dp->dp_spa) && | |
1875 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) && | |
1876 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH)) | |
1877 | return (B_TRUE); | |
1878 | #endif /* HAVE_LARGE_STACKS */ | |
1879 | ||
1880 | return (B_FALSE); | |
1881 | } | |
1882 | ||
da6b4005 NB |
1883 | __attribute__((always_inline)) |
1884 | static inline void | |
1885 | __zio_execute(zio_t *zio) | |
b128c09f BB |
1886 | { |
1887 | zio->io_executor = curthread; | |
34dc7c2f | 1888 | |
3dfb57a3 DB |
1889 | ASSERT3U(zio->io_queued_timestamp, >, 0); |
1890 | ||
b128c09f | 1891 | while (zio->io_stage < ZIO_STAGE_DONE) { |
428870ff BB |
1892 | enum zio_stage pipeline = zio->io_pipeline; |
1893 | enum zio_stage stage = zio->io_stage; | |
b128c09f | 1894 | int rv; |
34dc7c2f | 1895 | |
b128c09f | 1896 | ASSERT(!MUTEX_HELD(&zio->io_lock)); |
428870ff BB |
1897 | ASSERT(ISP2(stage)); |
1898 | ASSERT(zio->io_stall == NULL); | |
34dc7c2f | 1899 | |
428870ff BB |
1900 | do { |
1901 | stage <<= 1; | |
1902 | } while ((stage & pipeline) == 0); | |
b128c09f BB |
1903 | |
1904 | ASSERT(stage <= ZIO_STAGE_DONE); | |
34dc7c2f BB |
1905 | |
1906 | /* | |
b128c09f BB |
1907 | * If we are in interrupt context and this pipeline stage |
1908 | * will grab a config lock that is held across I/O, | |
428870ff BB |
1909 | * or may wait for an I/O that needs an interrupt thread |
1910 | * to complete, issue async to avoid deadlock. | |
1911 | * | |
1912 | * For VDEV_IO_START, we cut in line so that the io will | |
1913 | * be sent to disk promptly. | |
34dc7c2f | 1914 | */ |
91579709 BB |
1915 | if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && |
1916 | zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { | |
b58986ee BB |
1917 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? |
1918 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
91579709 BB |
1919 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
1920 | return; | |
1921 | } | |
1922 | ||
1923 | /* | |
b58986ee BB |
1924 | * If the current context doesn't have large enough stacks |
1925 | * the zio must be issued asynchronously to prevent overflow. | |
91579709 | 1926 | */ |
b58986ee BB |
1927 | if (zio_execute_stack_check(zio)) { |
1928 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? | |
1929 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
428870ff | 1930 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
b128c09f | 1931 | return; |
34dc7c2f BB |
1932 | } |
1933 | ||
b128c09f | 1934 | zio->io_stage = stage; |
3dfb57a3 | 1935 | zio->io_pipeline_trace |= zio->io_stage; |
9bd274dd | 1936 | rv = zio_pipeline[highbit64(stage) - 1](zio); |
34dc7c2f | 1937 | |
b128c09f BB |
1938 | if (rv == ZIO_PIPELINE_STOP) |
1939 | return; | |
34dc7c2f | 1940 | |
b128c09f BB |
1941 | ASSERT(rv == ZIO_PIPELINE_CONTINUE); |
1942 | } | |
34dc7c2f BB |
1943 | } |
1944 | ||
da6b4005 | 1945 | |
b128c09f BB |
1946 | /* |
1947 | * ========================================================================== | |
1948 | * Initiate I/O, either sync or async | |
1949 | * ========================================================================== | |
1950 | */ | |
1951 | int | |
1952 | zio_wait(zio_t *zio) | |
34dc7c2f | 1953 | { |
8fb1ede1 | 1954 | long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms); |
b128c09f | 1955 | int error; |
34dc7c2f | 1956 | |
1ce23dca PS |
1957 | ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN); |
1958 | ASSERT3P(zio->io_executor, ==, NULL); | |
34dc7c2f | 1959 | |
b128c09f | 1960 | zio->io_waiter = curthread; |
3dfb57a3 DB |
1961 | ASSERT0(zio->io_queued_timestamp); |
1962 | zio->io_queued_timestamp = gethrtime(); | |
34dc7c2f | 1963 | |
da6b4005 | 1964 | __zio_execute(zio); |
34dc7c2f | 1965 | |
b128c09f | 1966 | mutex_enter(&zio->io_lock); |
8fb1ede1 BB |
1967 | while (zio->io_executor != NULL) { |
1968 | error = cv_timedwait_io(&zio->io_cv, &zio->io_lock, | |
1969 | ddi_get_lbolt() + timeout); | |
1970 | ||
1971 | if (zfs_deadman_enabled && error == -1 && | |
1972 | gethrtime() - zio->io_queued_timestamp > | |
1973 | spa_deadman_ziotime(zio->io_spa)) { | |
1974 | mutex_exit(&zio->io_lock); | |
1975 | timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms); | |
1976 | zio_deadman(zio, FTAG); | |
1977 | mutex_enter(&zio->io_lock); | |
1978 | } | |
1979 | } | |
b128c09f | 1980 | mutex_exit(&zio->io_lock); |
34dc7c2f | 1981 | |
b128c09f BB |
1982 | error = zio->io_error; |
1983 | zio_destroy(zio); | |
34dc7c2f | 1984 | |
b128c09f BB |
1985 | return (error); |
1986 | } | |
34dc7c2f | 1987 | |
b128c09f BB |
1988 | void |
1989 | zio_nowait(zio_t *zio) | |
1990 | { | |
1ce23dca | 1991 | ASSERT3P(zio->io_executor, ==, NULL); |
34dc7c2f | 1992 | |
d164b209 BB |
1993 | if (zio->io_child_type == ZIO_CHILD_LOGICAL && |
1994 | zio_unique_parent(zio) == NULL) { | |
8878261f BB |
1995 | zio_t *pio; |
1996 | ||
34dc7c2f | 1997 | /* |
b128c09f | 1998 | * This is a logical async I/O with no parent to wait for it. |
9babb374 BB |
1999 | * We add it to the spa_async_root_zio "Godfather" I/O which |
2000 | * will ensure they complete prior to unloading the pool. | |
34dc7c2f | 2001 | */ |
b128c09f | 2002 | spa_t *spa = zio->io_spa; |
8878261f BB |
2003 | kpreempt_disable(); |
2004 | pio = spa->spa_async_zio_root[CPU_SEQID]; | |
2005 | kpreempt_enable(); | |
9babb374 | 2006 | |
8878261f | 2007 | zio_add_child(pio, zio); |
b128c09f | 2008 | } |
34dc7c2f | 2009 | |
3dfb57a3 DB |
2010 | ASSERT0(zio->io_queued_timestamp); |
2011 | zio->io_queued_timestamp = gethrtime(); | |
da6b4005 | 2012 | __zio_execute(zio); |
b128c09f | 2013 | } |
34dc7c2f | 2014 | |
b128c09f BB |
2015 | /* |
2016 | * ========================================================================== | |
1ce23dca | 2017 | * Reexecute, cancel, or suspend/resume failed I/O |
b128c09f BB |
2018 | * ========================================================================== |
2019 | */ | |
34dc7c2f | 2020 | |
b128c09f BB |
2021 | static void |
2022 | zio_reexecute(zio_t *pio) | |
2023 | { | |
d164b209 BB |
2024 | zio_t *cio, *cio_next; |
2025 | ||
2026 | ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); | |
2027 | ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); | |
9babb374 BB |
2028 | ASSERT(pio->io_gang_leader == NULL); |
2029 | ASSERT(pio->io_gang_tree == NULL); | |
34dc7c2f | 2030 | |
b128c09f BB |
2031 | pio->io_flags = pio->io_orig_flags; |
2032 | pio->io_stage = pio->io_orig_stage; | |
2033 | pio->io_pipeline = pio->io_orig_pipeline; | |
2034 | pio->io_reexecute = 0; | |
03c6040b | 2035 | pio->io_flags |= ZIO_FLAG_REEXECUTED; |
3dfb57a3 | 2036 | pio->io_pipeline_trace = 0; |
b128c09f | 2037 | pio->io_error = 0; |
1c27024e | 2038 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 | 2039 | pio->io_state[w] = 0; |
1c27024e | 2040 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) |
b128c09f | 2041 | pio->io_child_error[c] = 0; |
34dc7c2f | 2042 | |
428870ff BB |
2043 | if (IO_IS_ALLOCATING(pio)) |
2044 | BP_ZERO(pio->io_bp); | |
34dc7c2f | 2045 | |
b128c09f BB |
2046 | /* |
2047 | * As we reexecute pio's children, new children could be created. | |
d164b209 | 2048 | * New children go to the head of pio's io_child_list, however, |
b128c09f | 2049 | * so we will (correctly) not reexecute them. The key is that |
d164b209 BB |
2050 | * the remainder of pio's io_child_list, from 'cio_next' onward, |
2051 | * cannot be affected by any side effects of reexecuting 'cio'. | |
b128c09f | 2052 | */ |
1c27024e | 2053 | zio_link_t *zl = NULL; |
a8b2e306 | 2054 | mutex_enter(&pio->io_lock); |
3dfb57a3 DB |
2055 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { |
2056 | cio_next = zio_walk_children(pio, &zl); | |
1c27024e | 2057 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 | 2058 | pio->io_children[cio->io_child_type][w]++; |
b128c09f | 2059 | mutex_exit(&pio->io_lock); |
d164b209 | 2060 | zio_reexecute(cio); |
a8b2e306 | 2061 | mutex_enter(&pio->io_lock); |
34dc7c2f | 2062 | } |
a8b2e306 | 2063 | mutex_exit(&pio->io_lock); |
34dc7c2f | 2064 | |
b128c09f BB |
2065 | /* |
2066 | * Now that all children have been reexecuted, execute the parent. | |
9babb374 | 2067 | * We don't reexecute "The Godfather" I/O here as it's the |
9e2c3bb4 | 2068 | * responsibility of the caller to wait on it. |
b128c09f | 2069 | */ |
3dfb57a3 DB |
2070 | if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) { |
2071 | pio->io_queued_timestamp = gethrtime(); | |
da6b4005 | 2072 | __zio_execute(pio); |
3dfb57a3 | 2073 | } |
34dc7c2f BB |
2074 | } |
2075 | ||
b128c09f BB |
2076 | void |
2077 | zio_suspend(spa_t *spa, zio_t *zio) | |
34dc7c2f | 2078 | { |
b128c09f BB |
2079 | if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) |
2080 | fm_panic("Pool '%s' has encountered an uncorrectable I/O " | |
2081 | "failure and the failure mode property for this pool " | |
2082 | "is set to panic.", spa_name(spa)); | |
34dc7c2f | 2083 | |
bf89c199 BB |
2084 | cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O " |
2085 | "failure and has been suspended.\n", spa_name(spa)); | |
2086 | ||
b5256303 TC |
2087 | zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, |
2088 | NULL, NULL, 0, 0); | |
34dc7c2f | 2089 | |
b128c09f | 2090 | mutex_enter(&spa->spa_suspend_lock); |
34dc7c2f | 2091 | |
b128c09f | 2092 | if (spa->spa_suspend_zio_root == NULL) |
9babb374 BB |
2093 | spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, |
2094 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
2095 | ZIO_FLAG_GODFATHER); | |
34dc7c2f | 2096 | |
b128c09f | 2097 | spa->spa_suspended = B_TRUE; |
34dc7c2f | 2098 | |
b128c09f | 2099 | if (zio != NULL) { |
9babb374 | 2100 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f BB |
2101 | ASSERT(zio != spa->spa_suspend_zio_root); |
2102 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
d164b209 | 2103 | ASSERT(zio_unique_parent(zio) == NULL); |
b128c09f BB |
2104 | ASSERT(zio->io_stage == ZIO_STAGE_DONE); |
2105 | zio_add_child(spa->spa_suspend_zio_root, zio); | |
2106 | } | |
34dc7c2f | 2107 | |
b128c09f BB |
2108 | mutex_exit(&spa->spa_suspend_lock); |
2109 | } | |
34dc7c2f | 2110 | |
9babb374 | 2111 | int |
b128c09f BB |
2112 | zio_resume(spa_t *spa) |
2113 | { | |
9babb374 | 2114 | zio_t *pio; |
34dc7c2f BB |
2115 | |
2116 | /* | |
b128c09f | 2117 | * Reexecute all previously suspended i/o. |
34dc7c2f | 2118 | */ |
b128c09f BB |
2119 | mutex_enter(&spa->spa_suspend_lock); |
2120 | spa->spa_suspended = B_FALSE; | |
2121 | cv_broadcast(&spa->spa_suspend_cv); | |
2122 | pio = spa->spa_suspend_zio_root; | |
2123 | spa->spa_suspend_zio_root = NULL; | |
2124 | mutex_exit(&spa->spa_suspend_lock); | |
2125 | ||
2126 | if (pio == NULL) | |
9babb374 | 2127 | return (0); |
34dc7c2f | 2128 | |
9babb374 BB |
2129 | zio_reexecute(pio); |
2130 | return (zio_wait(pio)); | |
b128c09f BB |
2131 | } |
2132 | ||
2133 | void | |
2134 | zio_resume_wait(spa_t *spa) | |
2135 | { | |
2136 | mutex_enter(&spa->spa_suspend_lock); | |
2137 | while (spa_suspended(spa)) | |
2138 | cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); | |
2139 | mutex_exit(&spa->spa_suspend_lock); | |
34dc7c2f BB |
2140 | } |
2141 | ||
2142 | /* | |
2143 | * ========================================================================== | |
b128c09f BB |
2144 | * Gang blocks. |
2145 | * | |
2146 | * A gang block is a collection of small blocks that looks to the DMU | |
2147 | * like one large block. When zio_dva_allocate() cannot find a block | |
2148 | * of the requested size, due to either severe fragmentation or the pool | |
2149 | * being nearly full, it calls zio_write_gang_block() to construct the | |
2150 | * block from smaller fragments. | |
2151 | * | |
2152 | * A gang block consists of a gang header (zio_gbh_phys_t) and up to | |
2153 | * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like | |
2154 | * an indirect block: it's an array of block pointers. It consumes | |
2155 | * only one sector and hence is allocatable regardless of fragmentation. | |
2156 | * The gang header's bps point to its gang members, which hold the data. | |
2157 | * | |
2158 | * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> | |
2159 | * as the verifier to ensure uniqueness of the SHA256 checksum. | |
2160 | * Critically, the gang block bp's blk_cksum is the checksum of the data, | |
2161 | * not the gang header. This ensures that data block signatures (needed for | |
2162 | * deduplication) are independent of how the block is physically stored. | |
2163 | * | |
2164 | * Gang blocks can be nested: a gang member may itself be a gang block. | |
2165 | * Thus every gang block is a tree in which root and all interior nodes are | |
2166 | * gang headers, and the leaves are normal blocks that contain user data. | |
2167 | * The root of the gang tree is called the gang leader. | |
2168 | * | |
2169 | * To perform any operation (read, rewrite, free, claim) on a gang block, | |
2170 | * zio_gang_assemble() first assembles the gang tree (minus data leaves) | |
2171 | * in the io_gang_tree field of the original logical i/o by recursively | |
2172 | * reading the gang leader and all gang headers below it. This yields | |
2173 | * an in-core tree containing the contents of every gang header and the | |
2174 | * bps for every constituent of the gang block. | |
2175 | * | |
2176 | * With the gang tree now assembled, zio_gang_issue() just walks the gang tree | |
2177 | * and invokes a callback on each bp. To free a gang block, zio_gang_issue() | |
2178 | * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. | |
2179 | * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). | |
2180 | * zio_read_gang() is a wrapper around zio_read() that omits reading gang | |
2181 | * headers, since we already have those in io_gang_tree. zio_rewrite_gang() | |
2182 | * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() | |
2183 | * of the gang header plus zio_checksum_compute() of the data to update the | |
2184 | * gang header's blk_cksum as described above. | |
2185 | * | |
2186 | * The two-phase assemble/issue model solves the problem of partial failure -- | |
2187 | * what if you'd freed part of a gang block but then couldn't read the | |
2188 | * gang header for another part? Assembling the entire gang tree first | |
2189 | * ensures that all the necessary gang header I/O has succeeded before | |
2190 | * starting the actual work of free, claim, or write. Once the gang tree | |
2191 | * is assembled, free and claim are in-memory operations that cannot fail. | |
2192 | * | |
2193 | * In the event that a gang write fails, zio_dva_unallocate() walks the | |
2194 | * gang tree to immediately free (i.e. insert back into the space map) | |
2195 | * everything we've allocated. This ensures that we don't get ENOSPC | |
2196 | * errors during repeated suspend/resume cycles due to a flaky device. | |
2197 | * | |
2198 | * Gang rewrites only happen during sync-to-convergence. If we can't assemble | |
2199 | * the gang tree, we won't modify the block, so we can safely defer the free | |
2200 | * (knowing that the block is still intact). If we *can* assemble the gang | |
2201 | * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free | |
2202 | * each constituent bp and we can allocate a new block on the next sync pass. | |
2203 | * | |
2204 | * In all cases, the gang tree allows complete recovery from partial failure. | |
34dc7c2f BB |
2205 | * ========================================================================== |
2206 | */ | |
b128c09f | 2207 | |
a6255b7f DQ |
2208 | static void |
2209 | zio_gang_issue_func_done(zio_t *zio) | |
2210 | { | |
2211 | abd_put(zio->io_abd); | |
2212 | } | |
2213 | ||
b128c09f | 2214 | static zio_t * |
a6255b7f DQ |
2215 | zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, |
2216 | uint64_t offset) | |
34dc7c2f | 2217 | { |
b128c09f BB |
2218 | if (gn != NULL) |
2219 | return (pio); | |
34dc7c2f | 2220 | |
a6255b7f DQ |
2221 | return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset), |
2222 | BP_GET_PSIZE(bp), zio_gang_issue_func_done, | |
2223 | NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
b128c09f BB |
2224 | &pio->io_bookmark)); |
2225 | } | |
2226 | ||
a6255b7f DQ |
2227 | static zio_t * |
2228 | zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2229 | uint64_t offset) | |
b128c09f BB |
2230 | { |
2231 | zio_t *zio; | |
2232 | ||
2233 | if (gn != NULL) { | |
a6255b7f DQ |
2234 | abd_t *gbh_abd = |
2235 | abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
b128c09f | 2236 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
a6255b7f DQ |
2237 | gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL, |
2238 | pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
2239 | &pio->io_bookmark); | |
34dc7c2f | 2240 | /* |
b128c09f BB |
2241 | * As we rewrite each gang header, the pipeline will compute |
2242 | * a new gang block header checksum for it; but no one will | |
2243 | * compute a new data checksum, so we do that here. The one | |
2244 | * exception is the gang leader: the pipeline already computed | |
2245 | * its data checksum because that stage precedes gang assembly. | |
2246 | * (Presently, nothing actually uses interior data checksums; | |
2247 | * this is just good hygiene.) | |
34dc7c2f | 2248 | */ |
9babb374 | 2249 | if (gn != pio->io_gang_leader->io_gang_tree) { |
a6255b7f DQ |
2250 | abd_t *buf = abd_get_offset(data, offset); |
2251 | ||
b128c09f | 2252 | zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), |
a6255b7f DQ |
2253 | buf, BP_GET_PSIZE(bp)); |
2254 | ||
2255 | abd_put(buf); | |
b128c09f | 2256 | } |
428870ff BB |
2257 | /* |
2258 | * If we are here to damage data for testing purposes, | |
2259 | * leave the GBH alone so that we can detect the damage. | |
2260 | */ | |
2261 | if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) | |
2262 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; | |
34dc7c2f | 2263 | } else { |
b128c09f | 2264 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
a6255b7f DQ |
2265 | abd_get_offset(data, offset), BP_GET_PSIZE(bp), |
2266 | zio_gang_issue_func_done, NULL, pio->io_priority, | |
b128c09f | 2267 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); |
34dc7c2f BB |
2268 | } |
2269 | ||
b128c09f BB |
2270 | return (zio); |
2271 | } | |
34dc7c2f | 2272 | |
b128c09f | 2273 | /* ARGSUSED */ |
a6255b7f DQ |
2274 | static zio_t * |
2275 | zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2276 | uint64_t offset) | |
b128c09f | 2277 | { |
428870ff BB |
2278 | return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, |
2279 | ZIO_GANG_CHILD_FLAGS(pio))); | |
34dc7c2f BB |
2280 | } |
2281 | ||
b128c09f | 2282 | /* ARGSUSED */ |
a6255b7f DQ |
2283 | static zio_t * |
2284 | zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2285 | uint64_t offset) | |
34dc7c2f | 2286 | { |
b128c09f BB |
2287 | return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, |
2288 | NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); | |
2289 | } | |
2290 | ||
2291 | static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { | |
2292 | NULL, | |
2293 | zio_read_gang, | |
2294 | zio_rewrite_gang, | |
2295 | zio_free_gang, | |
2296 | zio_claim_gang, | |
2297 | NULL | |
2298 | }; | |
34dc7c2f | 2299 | |
b128c09f | 2300 | static void zio_gang_tree_assemble_done(zio_t *zio); |
34dc7c2f | 2301 | |
b128c09f BB |
2302 | static zio_gang_node_t * |
2303 | zio_gang_node_alloc(zio_gang_node_t **gnpp) | |
2304 | { | |
2305 | zio_gang_node_t *gn; | |
34dc7c2f | 2306 | |
b128c09f | 2307 | ASSERT(*gnpp == NULL); |
34dc7c2f | 2308 | |
79c76d5b | 2309 | gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); |
b128c09f BB |
2310 | gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); |
2311 | *gnpp = gn; | |
34dc7c2f | 2312 | |
b128c09f | 2313 | return (gn); |
34dc7c2f BB |
2314 | } |
2315 | ||
34dc7c2f | 2316 | static void |
b128c09f | 2317 | zio_gang_node_free(zio_gang_node_t **gnpp) |
34dc7c2f | 2318 | { |
b128c09f | 2319 | zio_gang_node_t *gn = *gnpp; |
34dc7c2f | 2320 | |
1c27024e | 2321 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f BB |
2322 | ASSERT(gn->gn_child[g] == NULL); |
2323 | ||
2324 | zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
2325 | kmem_free(gn, sizeof (*gn)); | |
2326 | *gnpp = NULL; | |
34dc7c2f BB |
2327 | } |
2328 | ||
b128c09f BB |
2329 | static void |
2330 | zio_gang_tree_free(zio_gang_node_t **gnpp) | |
34dc7c2f | 2331 | { |
b128c09f | 2332 | zio_gang_node_t *gn = *gnpp; |
34dc7c2f | 2333 | |
b128c09f BB |
2334 | if (gn == NULL) |
2335 | return; | |
34dc7c2f | 2336 | |
1c27024e | 2337 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f | 2338 | zio_gang_tree_free(&gn->gn_child[g]); |
34dc7c2f | 2339 | |
b128c09f | 2340 | zio_gang_node_free(gnpp); |
34dc7c2f BB |
2341 | } |
2342 | ||
b128c09f | 2343 | static void |
9babb374 | 2344 | zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) |
34dc7c2f | 2345 | { |
b128c09f | 2346 | zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); |
a6255b7f | 2347 | abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); |
b128c09f | 2348 | |
9babb374 | 2349 | ASSERT(gio->io_gang_leader == gio); |
b128c09f | 2350 | ASSERT(BP_IS_GANG(bp)); |
34dc7c2f | 2351 | |
a6255b7f DQ |
2352 | zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE, |
2353 | zio_gang_tree_assemble_done, gn, gio->io_priority, | |
2354 | ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); | |
b128c09f | 2355 | } |
34dc7c2f | 2356 | |
b128c09f BB |
2357 | static void |
2358 | zio_gang_tree_assemble_done(zio_t *zio) | |
2359 | { | |
9babb374 | 2360 | zio_t *gio = zio->io_gang_leader; |
b128c09f BB |
2361 | zio_gang_node_t *gn = zio->io_private; |
2362 | blkptr_t *bp = zio->io_bp; | |
34dc7c2f | 2363 | |
9babb374 | 2364 | ASSERT(gio == zio_unique_parent(zio)); |
428870ff | 2365 | ASSERT(zio->io_child_count == 0); |
34dc7c2f | 2366 | |
b128c09f BB |
2367 | if (zio->io_error) |
2368 | return; | |
34dc7c2f | 2369 | |
a6255b7f | 2370 | /* this ABD was created from a linear buf in zio_gang_tree_assemble */ |
b128c09f | 2371 | if (BP_SHOULD_BYTESWAP(bp)) |
a6255b7f | 2372 | byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size); |
34dc7c2f | 2373 | |
a6255b7f | 2374 | ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh); |
b128c09f | 2375 | ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); |
428870ff | 2376 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2377 | |
a6255b7f DQ |
2378 | abd_put(zio->io_abd); |
2379 | ||
1c27024e | 2380 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2381 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2382 | if (!BP_IS_GANG(gbp)) | |
2383 | continue; | |
9babb374 | 2384 | zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); |
b128c09f | 2385 | } |
34dc7c2f BB |
2386 | } |
2387 | ||
b128c09f | 2388 | static void |
a6255b7f DQ |
2389 | zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data, |
2390 | uint64_t offset) | |
34dc7c2f | 2391 | { |
9babb374 | 2392 | zio_t *gio = pio->io_gang_leader; |
b128c09f | 2393 | zio_t *zio; |
34dc7c2f | 2394 | |
b128c09f | 2395 | ASSERT(BP_IS_GANG(bp) == !!gn); |
9babb374 BB |
2396 | ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); |
2397 | ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); | |
34dc7c2f | 2398 | |
b128c09f BB |
2399 | /* |
2400 | * If you're a gang header, your data is in gn->gn_gbh. | |
2401 | * If you're a gang member, your data is in 'data' and gn == NULL. | |
2402 | */ | |
a6255b7f | 2403 | zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset); |
34dc7c2f | 2404 | |
b128c09f | 2405 | if (gn != NULL) { |
428870ff | 2406 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2407 | |
1c27024e | 2408 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2409 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2410 | if (BP_IS_HOLE(gbp)) | |
2411 | continue; | |
a6255b7f DQ |
2412 | zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data, |
2413 | offset); | |
2414 | offset += BP_GET_PSIZE(gbp); | |
b128c09f | 2415 | } |
34dc7c2f BB |
2416 | } |
2417 | ||
9babb374 | 2418 | if (gn == gio->io_gang_tree) |
a6255b7f | 2419 | ASSERT3U(gio->io_size, ==, offset); |
34dc7c2f | 2420 | |
b128c09f BB |
2421 | if (zio != pio) |
2422 | zio_nowait(zio); | |
34dc7c2f BB |
2423 | } |
2424 | ||
2425 | static int | |
b128c09f | 2426 | zio_gang_assemble(zio_t *zio) |
34dc7c2f | 2427 | { |
b128c09f | 2428 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2429 | |
9babb374 BB |
2430 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); |
2431 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
2432 | ||
2433 | zio->io_gang_leader = zio; | |
34dc7c2f | 2434 | |
b128c09f | 2435 | zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); |
34dc7c2f BB |
2436 | |
2437 | return (ZIO_PIPELINE_CONTINUE); | |
2438 | } | |
2439 | ||
2440 | static int | |
b128c09f | 2441 | zio_gang_issue(zio_t *zio) |
34dc7c2f | 2442 | { |
b128c09f | 2443 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2444 | |
b128c09f BB |
2445 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) |
2446 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f | 2447 | |
9babb374 BB |
2448 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); |
2449 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
34dc7c2f | 2450 | |
b128c09f | 2451 | if (zio->io_child_error[ZIO_CHILD_GANG] == 0) |
a6255b7f DQ |
2452 | zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd, |
2453 | 0); | |
b128c09f | 2454 | else |
9babb374 | 2455 | zio_gang_tree_free(&zio->io_gang_tree); |
34dc7c2f | 2456 | |
b128c09f | 2457 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f BB |
2458 | |
2459 | return (ZIO_PIPELINE_CONTINUE); | |
2460 | } | |
2461 | ||
2462 | static void | |
b128c09f | 2463 | zio_write_gang_member_ready(zio_t *zio) |
34dc7c2f | 2464 | { |
d164b209 | 2465 | zio_t *pio = zio_unique_parent(zio); |
34dc7c2f BB |
2466 | dva_t *cdva = zio->io_bp->blk_dva; |
2467 | dva_t *pdva = pio->io_bp->blk_dva; | |
2468 | uint64_t asize; | |
d1d7e268 | 2469 | ASSERTV(zio_t *gio = zio->io_gang_leader); |
34dc7c2f | 2470 | |
b128c09f BB |
2471 | if (BP_IS_HOLE(zio->io_bp)) |
2472 | return; | |
2473 | ||
2474 | ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); | |
2475 | ||
2476 | ASSERT(zio->io_child_type == ZIO_CHILD_GANG); | |
428870ff BB |
2477 | ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); |
2478 | ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); | |
2479 | ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); | |
34dc7c2f | 2480 | ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); |
34dc7c2f BB |
2481 | |
2482 | mutex_enter(&pio->io_lock); | |
1c27024e | 2483 | for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { |
34dc7c2f BB |
2484 | ASSERT(DVA_GET_GANG(&pdva[d])); |
2485 | asize = DVA_GET_ASIZE(&pdva[d]); | |
2486 | asize += DVA_GET_ASIZE(&cdva[d]); | |
2487 | DVA_SET_ASIZE(&pdva[d], asize); | |
2488 | } | |
2489 | mutex_exit(&pio->io_lock); | |
2490 | } | |
2491 | ||
a6255b7f DQ |
2492 | static void |
2493 | zio_write_gang_done(zio_t *zio) | |
2494 | { | |
2495 | abd_put(zio->io_abd); | |
2496 | } | |
2497 | ||
34dc7c2f | 2498 | static int |
b128c09f | 2499 | zio_write_gang_block(zio_t *pio) |
34dc7c2f | 2500 | { |
b128c09f | 2501 | spa_t *spa = pio->io_spa; |
3dfb57a3 | 2502 | metaslab_class_t *mc = spa_normal_class(spa); |
b128c09f | 2503 | blkptr_t *bp = pio->io_bp; |
9babb374 | 2504 | zio_t *gio = pio->io_gang_leader; |
b128c09f BB |
2505 | zio_t *zio; |
2506 | zio_gang_node_t *gn, **gnpp; | |
34dc7c2f | 2507 | zio_gbh_phys_t *gbh; |
a6255b7f | 2508 | abd_t *gbh_abd; |
b128c09f BB |
2509 | uint64_t txg = pio->io_txg; |
2510 | uint64_t resid = pio->io_size; | |
2511 | uint64_t lsize; | |
428870ff | 2512 | int copies = gio->io_prop.zp_copies; |
b5256303 | 2513 | int gbh_copies; |
b128c09f | 2514 | zio_prop_t zp; |
1c27024e | 2515 | int error; |
b5256303 TC |
2516 | |
2517 | /* | |
2518 | * encrypted blocks need DVA[2] free so encrypted gang headers can't | |
2519 | * have a third copy. | |
2520 | */ | |
2521 | gbh_copies = MIN(copies + 1, spa_max_replication(spa)); | |
2522 | if (gio->io_prop.zp_encrypt && gbh_copies >= SPA_DVAS_PER_BP) | |
2523 | gbh_copies = SPA_DVAS_PER_BP - 1; | |
2524 | ||
1c27024e | 2525 | int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER; |
3dfb57a3 DB |
2526 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
2527 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
2528 | ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); | |
2529 | ||
2530 | flags |= METASLAB_ASYNC_ALLOC; | |
2531 | VERIFY(refcount_held(&mc->mc_alloc_slots, pio)); | |
2532 | ||
2533 | /* | |
2534 | * The logical zio has already placed a reservation for | |
2535 | * 'copies' allocation slots but gang blocks may require | |
2536 | * additional copies. These additional copies | |
2537 | * (i.e. gbh_copies - copies) are guaranteed to succeed | |
2538 | * since metaslab_class_throttle_reserve() always allows | |
2539 | * additional reservations for gang blocks. | |
2540 | */ | |
2541 | VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies, | |
2542 | pio, flags)); | |
2543 | } | |
2544 | ||
2545 | error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE, | |
4e21fd06 DB |
2546 | bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, |
2547 | &pio->io_alloc_list, pio); | |
34dc7c2f | 2548 | if (error) { |
3dfb57a3 DB |
2549 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
2550 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
2551 | ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); | |
2552 | ||
2553 | /* | |
2554 | * If we failed to allocate the gang block header then | |
2555 | * we remove any additional allocation reservations that | |
2556 | * we placed here. The original reservation will | |
2557 | * be removed when the logical I/O goes to the ready | |
2558 | * stage. | |
2559 | */ | |
2560 | metaslab_class_throttle_unreserve(mc, | |
2561 | gbh_copies - copies, pio); | |
2562 | } | |
2563 | ||
b128c09f | 2564 | pio->io_error = error; |
34dc7c2f BB |
2565 | return (ZIO_PIPELINE_CONTINUE); |
2566 | } | |
2567 | ||
9babb374 BB |
2568 | if (pio == gio) { |
2569 | gnpp = &gio->io_gang_tree; | |
b128c09f BB |
2570 | } else { |
2571 | gnpp = pio->io_private; | |
2572 | ASSERT(pio->io_ready == zio_write_gang_member_ready); | |
34dc7c2f BB |
2573 | } |
2574 | ||
b128c09f BB |
2575 | gn = zio_gang_node_alloc(gnpp); |
2576 | gbh = gn->gn_gbh; | |
2577 | bzero(gbh, SPA_GANGBLOCKSIZE); | |
a6255b7f | 2578 | gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE); |
34dc7c2f | 2579 | |
b128c09f BB |
2580 | /* |
2581 | * Create the gang header. | |
2582 | */ | |
a6255b7f DQ |
2583 | zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE, |
2584 | zio_write_gang_done, NULL, pio->io_priority, | |
2585 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); | |
34dc7c2f | 2586 | |
b128c09f BB |
2587 | /* |
2588 | * Create and nowait the gang children. | |
2589 | */ | |
1c27024e | 2590 | for (int g = 0; resid != 0; resid -= lsize, g++) { |
b128c09f BB |
2591 | lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), |
2592 | SPA_MINBLOCKSIZE); | |
2593 | ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); | |
2594 | ||
9babb374 | 2595 | zp.zp_checksum = gio->io_prop.zp_checksum; |
b128c09f BB |
2596 | zp.zp_compress = ZIO_COMPRESS_OFF; |
2597 | zp.zp_type = DMU_OT_NONE; | |
2598 | zp.zp_level = 0; | |
428870ff | 2599 | zp.zp_copies = gio->io_prop.zp_copies; |
03c6040b GW |
2600 | zp.zp_dedup = B_FALSE; |
2601 | zp.zp_dedup_verify = B_FALSE; | |
2602 | zp.zp_nopwrite = B_FALSE; | |
4807c0ba TC |
2603 | zp.zp_encrypt = gio->io_prop.zp_encrypt; |
2604 | zp.zp_byteorder = gio->io_prop.zp_byteorder; | |
b5256303 TC |
2605 | bzero(zp.zp_salt, ZIO_DATA_SALT_LEN); |
2606 | bzero(zp.zp_iv, ZIO_DATA_IV_LEN); | |
2607 | bzero(zp.zp_mac, ZIO_DATA_MAC_LEN); | |
b128c09f | 2608 | |
1c27024e | 2609 | zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g], |
a6255b7f DQ |
2610 | abd_get_offset(pio->io_abd, pio->io_size - resid), lsize, |
2611 | lsize, &zp, zio_write_gang_member_ready, NULL, NULL, | |
2612 | zio_write_gang_done, &gn->gn_child[g], pio->io_priority, | |
3dfb57a3 DB |
2613 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); |
2614 | ||
2615 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { | |
2616 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
2617 | ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); | |
2618 | ||
2619 | /* | |
2620 | * Gang children won't throttle but we should | |
2621 | * account for their work, so reserve an allocation | |
2622 | * slot for them here. | |
2623 | */ | |
2624 | VERIFY(metaslab_class_throttle_reserve(mc, | |
2625 | zp.zp_copies, cio, flags)); | |
2626 | } | |
2627 | zio_nowait(cio); | |
b128c09f | 2628 | } |
34dc7c2f BB |
2629 | |
2630 | /* | |
b128c09f | 2631 | * Set pio's pipeline to just wait for zio to finish. |
34dc7c2f | 2632 | */ |
b128c09f BB |
2633 | pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
2634 | ||
920dd524 ED |
2635 | /* |
2636 | * We didn't allocate this bp, so make sure it doesn't get unmarked. | |
2637 | */ | |
2638 | pio->io_flags &= ~ZIO_FLAG_FASTWRITE; | |
2639 | ||
b128c09f BB |
2640 | zio_nowait(zio); |
2641 | ||
2642 | return (ZIO_PIPELINE_CONTINUE); | |
34dc7c2f BB |
2643 | } |
2644 | ||
03c6040b | 2645 | /* |
3c67d83a TH |
2646 | * The zio_nop_write stage in the pipeline determines if allocating a |
2647 | * new bp is necessary. The nopwrite feature can handle writes in | |
2648 | * either syncing or open context (i.e. zil writes) and as a result is | |
2649 | * mutually exclusive with dedup. | |
2650 | * | |
2651 | * By leveraging a cryptographically secure checksum, such as SHA256, we | |
2652 | * can compare the checksums of the new data and the old to determine if | |
2653 | * allocating a new block is required. Note that our requirements for | |
2654 | * cryptographic strength are fairly weak: there can't be any accidental | |
2655 | * hash collisions, but we don't need to be secure against intentional | |
2656 | * (malicious) collisions. To trigger a nopwrite, you have to be able | |
2657 | * to write the file to begin with, and triggering an incorrect (hash | |
2658 | * collision) nopwrite is no worse than simply writing to the file. | |
2659 | * That said, there are no known attacks against the checksum algorithms | |
2660 | * used for nopwrite, assuming that the salt and the checksums | |
2661 | * themselves remain secret. | |
03c6040b GW |
2662 | */ |
2663 | static int | |
2664 | zio_nop_write(zio_t *zio) | |
2665 | { | |
2666 | blkptr_t *bp = zio->io_bp; | |
2667 | blkptr_t *bp_orig = &zio->io_bp_orig; | |
2668 | zio_prop_t *zp = &zio->io_prop; | |
2669 | ||
2670 | ASSERT(BP_GET_LEVEL(bp) == 0); | |
2671 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
2672 | ASSERT(zp->zp_nopwrite); | |
2673 | ASSERT(!zp->zp_dedup); | |
2674 | ASSERT(zio->io_bp_override == NULL); | |
2675 | ASSERT(IO_IS_ALLOCATING(zio)); | |
2676 | ||
2677 | /* | |
2678 | * Check to see if the original bp and the new bp have matching | |
2679 | * characteristics (i.e. same checksum, compression algorithms, etc). | |
2680 | * If they don't then just continue with the pipeline which will | |
2681 | * allocate a new bp. | |
2682 | */ | |
2683 | if (BP_IS_HOLE(bp_orig) || | |
3c67d83a TH |
2684 | !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & |
2685 | ZCHECKSUM_FLAG_NOPWRITE) || | |
b5256303 | 2686 | BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) || |
03c6040b GW |
2687 | BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || |
2688 | BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || | |
2689 | BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || | |
2690 | zp->zp_copies != BP_GET_NDVAS(bp_orig)) | |
2691 | return (ZIO_PIPELINE_CONTINUE); | |
2692 | ||
2693 | /* | |
2694 | * If the checksums match then reset the pipeline so that we | |
2695 | * avoid allocating a new bp and issuing any I/O. | |
2696 | */ | |
2697 | if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { | |
3c67d83a TH |
2698 | ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & |
2699 | ZCHECKSUM_FLAG_NOPWRITE); | |
03c6040b GW |
2700 | ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); |
2701 | ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); | |
2702 | ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); | |
2703 | ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, | |
2704 | sizeof (uint64_t)) == 0); | |
2705 | ||
2706 | *bp = *bp_orig; | |
2707 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
2708 | zio->io_flags |= ZIO_FLAG_NOPWRITE; | |
2709 | } | |
2710 | ||
2711 | return (ZIO_PIPELINE_CONTINUE); | |
2712 | } | |
2713 | ||
34dc7c2f BB |
2714 | /* |
2715 | * ========================================================================== | |
428870ff | 2716 | * Dedup |
34dc7c2f BB |
2717 | * ========================================================================== |
2718 | */ | |
428870ff BB |
2719 | static void |
2720 | zio_ddt_child_read_done(zio_t *zio) | |
2721 | { | |
2722 | blkptr_t *bp = zio->io_bp; | |
2723 | ddt_entry_t *dde = zio->io_private; | |
2724 | ddt_phys_t *ddp; | |
2725 | zio_t *pio = zio_unique_parent(zio); | |
2726 | ||
2727 | mutex_enter(&pio->io_lock); | |
2728 | ddp = ddt_phys_select(dde, bp); | |
2729 | if (zio->io_error == 0) | |
2730 | ddt_phys_clear(ddp); /* this ddp doesn't need repair */ | |
a6255b7f DQ |
2731 | |
2732 | if (zio->io_error == 0 && dde->dde_repair_abd == NULL) | |
2733 | dde->dde_repair_abd = zio->io_abd; | |
428870ff | 2734 | else |
a6255b7f | 2735 | abd_free(zio->io_abd); |
428870ff BB |
2736 | mutex_exit(&pio->io_lock); |
2737 | } | |
2738 | ||
2739 | static int | |
2740 | zio_ddt_read_start(zio_t *zio) | |
2741 | { | |
2742 | blkptr_t *bp = zio->io_bp; | |
2743 | ||
2744 | ASSERT(BP_GET_DEDUP(bp)); | |
2745 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
2746 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
2747 | ||
2748 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
2749 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
2750 | ddt_entry_t *dde = ddt_repair_start(ddt, bp); | |
2751 | ddt_phys_t *ddp = dde->dde_phys; | |
2752 | ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); | |
2753 | blkptr_t blk; | |
2754 | ||
2755 | ASSERT(zio->io_vsd == NULL); | |
2756 | zio->io_vsd = dde; | |
2757 | ||
2758 | if (ddp_self == NULL) | |
2759 | return (ZIO_PIPELINE_CONTINUE); | |
2760 | ||
1c27024e | 2761 | for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { |
428870ff BB |
2762 | if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) |
2763 | continue; | |
2764 | ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, | |
2765 | &blk); | |
2766 | zio_nowait(zio_read(zio, zio->io_spa, &blk, | |
a6255b7f DQ |
2767 | abd_alloc_for_io(zio->io_size, B_TRUE), |
2768 | zio->io_size, zio_ddt_child_read_done, dde, | |
2769 | zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) | | |
2770 | ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark)); | |
428870ff BB |
2771 | } |
2772 | return (ZIO_PIPELINE_CONTINUE); | |
2773 | } | |
2774 | ||
2775 | zio_nowait(zio_read(zio, zio->io_spa, bp, | |
a6255b7f | 2776 | zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority, |
428870ff BB |
2777 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); |
2778 | ||
2779 | return (ZIO_PIPELINE_CONTINUE); | |
2780 | } | |
2781 | ||
2782 | static int | |
2783 | zio_ddt_read_done(zio_t *zio) | |
2784 | { | |
2785 | blkptr_t *bp = zio->io_bp; | |
2786 | ||
2787 | if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE)) | |
2788 | return (ZIO_PIPELINE_STOP); | |
2789 | ||
2790 | ASSERT(BP_GET_DEDUP(bp)); | |
2791 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
2792 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
2793 | ||
2794 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
2795 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
2796 | ddt_entry_t *dde = zio->io_vsd; | |
2797 | if (ddt == NULL) { | |
2798 | ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); | |
2799 | return (ZIO_PIPELINE_CONTINUE); | |
2800 | } | |
2801 | if (dde == NULL) { | |
2802 | zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; | |
2803 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); | |
2804 | return (ZIO_PIPELINE_STOP); | |
2805 | } | |
a6255b7f DQ |
2806 | if (dde->dde_repair_abd != NULL) { |
2807 | abd_copy(zio->io_abd, dde->dde_repair_abd, | |
2808 | zio->io_size); | |
428870ff BB |
2809 | zio->io_child_error[ZIO_CHILD_DDT] = 0; |
2810 | } | |
2811 | ddt_repair_done(ddt, dde); | |
2812 | zio->io_vsd = NULL; | |
2813 | } | |
2814 | ||
2815 | ASSERT(zio->io_vsd == NULL); | |
2816 | ||
2817 | return (ZIO_PIPELINE_CONTINUE); | |
2818 | } | |
2819 | ||
2820 | static boolean_t | |
2821 | zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) | |
2822 | { | |
2823 | spa_t *spa = zio->io_spa; | |
c17bcf83 | 2824 | boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW); |
428870ff | 2825 | |
c17bcf83 | 2826 | ASSERT(!(zio->io_bp_override && do_raw)); |
2aa34383 | 2827 | |
428870ff BB |
2828 | /* |
2829 | * Note: we compare the original data, not the transformed data, | |
2830 | * because when zio->io_bp is an override bp, we will not have | |
2831 | * pushed the I/O transforms. That's an important optimization | |
2832 | * because otherwise we'd compress/encrypt all dmu_sync() data twice. | |
c17bcf83 | 2833 | * However, we should never get a raw, override zio so in these |
b5256303 | 2834 | * cases we can compare the io_abd directly. This is useful because |
c17bcf83 TC |
2835 | * it allows us to do dedup verification even if we don't have access |
2836 | * to the original data (for instance, if the encryption keys aren't | |
2837 | * loaded). | |
428870ff | 2838 | */ |
c17bcf83 | 2839 | |
1c27024e | 2840 | for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
2841 | zio_t *lio = dde->dde_lead_zio[p]; |
2842 | ||
c17bcf83 TC |
2843 | if (lio != NULL && do_raw) { |
2844 | return (lio->io_size != zio->io_size || | |
a6255b7f | 2845 | abd_cmp(zio->io_abd, lio->io_abd) != 0); |
c17bcf83 | 2846 | } else if (lio != NULL) { |
428870ff | 2847 | return (lio->io_orig_size != zio->io_orig_size || |
a6255b7f | 2848 | abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0); |
428870ff BB |
2849 | } |
2850 | } | |
2851 | ||
1c27024e | 2852 | for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
2853 | ddt_phys_t *ddp = &dde->dde_phys[p]; |
2854 | ||
c17bcf83 TC |
2855 | if (ddp->ddp_phys_birth != 0 && do_raw) { |
2856 | blkptr_t blk = *zio->io_bp; | |
2857 | uint64_t psize; | |
a6255b7f | 2858 | abd_t *tmpabd; |
c17bcf83 TC |
2859 | int error; |
2860 | ||
2861 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
2862 | psize = BP_GET_PSIZE(&blk); | |
2863 | ||
2864 | if (psize != zio->io_size) | |
2865 | return (B_TRUE); | |
2866 | ||
2867 | ddt_exit(ddt); | |
2868 | ||
a6255b7f | 2869 | tmpabd = abd_alloc_for_io(psize, B_TRUE); |
c17bcf83 | 2870 | |
a6255b7f | 2871 | error = zio_wait(zio_read(NULL, spa, &blk, tmpabd, |
c17bcf83 TC |
2872 | psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ, |
2873 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
2874 | ZIO_FLAG_RAW, &zio->io_bookmark)); | |
2875 | ||
2876 | if (error == 0) { | |
a6255b7f | 2877 | if (abd_cmp(tmpabd, zio->io_abd) != 0) |
c17bcf83 TC |
2878 | error = SET_ERROR(ENOENT); |
2879 | } | |
2880 | ||
a6255b7f | 2881 | abd_free(tmpabd); |
c17bcf83 TC |
2882 | ddt_enter(ddt); |
2883 | return (error != 0); | |
2884 | } else if (ddp->ddp_phys_birth != 0) { | |
428870ff | 2885 | arc_buf_t *abuf = NULL; |
2a432414 | 2886 | arc_flags_t aflags = ARC_FLAG_WAIT; |
428870ff BB |
2887 | blkptr_t blk = *zio->io_bp; |
2888 | int error; | |
2889 | ||
2890 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
2891 | ||
c17bcf83 TC |
2892 | if (BP_GET_LSIZE(&blk) != zio->io_orig_size) |
2893 | return (B_TRUE); | |
2894 | ||
428870ff BB |
2895 | ddt_exit(ddt); |
2896 | ||
294f6806 | 2897 | error = arc_read(NULL, spa, &blk, |
428870ff BB |
2898 | arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, |
2899 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
2900 | &aflags, &zio->io_bookmark); | |
2901 | ||
2902 | if (error == 0) { | |
a6255b7f | 2903 | if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data, |
428870ff | 2904 | zio->io_orig_size) != 0) |
c17bcf83 | 2905 | error = SET_ERROR(ENOENT); |
d3c2ae1c | 2906 | arc_buf_destroy(abuf, &abuf); |
428870ff BB |
2907 | } |
2908 | ||
2909 | ddt_enter(ddt); | |
2910 | return (error != 0); | |
2911 | } | |
2912 | } | |
2913 | ||
2914 | return (B_FALSE); | |
2915 | } | |
2916 | ||
2917 | static void | |
2918 | zio_ddt_child_write_ready(zio_t *zio) | |
2919 | { | |
2920 | int p = zio->io_prop.zp_copies; | |
2921 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
2922 | ddt_entry_t *dde = zio->io_private; | |
2923 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
2924 | zio_t *pio; | |
2925 | ||
2926 | if (zio->io_error) | |
2927 | return; | |
2928 | ||
2929 | ddt_enter(ddt); | |
2930 | ||
2931 | ASSERT(dde->dde_lead_zio[p] == zio); | |
2932 | ||
2933 | ddt_phys_fill(ddp, zio->io_bp); | |
2934 | ||
1c27024e | 2935 | zio_link_t *zl = NULL; |
3dfb57a3 | 2936 | while ((pio = zio_walk_parents(zio, &zl)) != NULL) |
428870ff BB |
2937 | ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); |
2938 | ||
2939 | ddt_exit(ddt); | |
2940 | } | |
2941 | ||
2942 | static void | |
2943 | zio_ddt_child_write_done(zio_t *zio) | |
2944 | { | |
2945 | int p = zio->io_prop.zp_copies; | |
2946 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
2947 | ddt_entry_t *dde = zio->io_private; | |
2948 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
2949 | ||
2950 | ddt_enter(ddt); | |
2951 | ||
2952 | ASSERT(ddp->ddp_refcnt == 0); | |
2953 | ASSERT(dde->dde_lead_zio[p] == zio); | |
2954 | dde->dde_lead_zio[p] = NULL; | |
2955 | ||
2956 | if (zio->io_error == 0) { | |
3dfb57a3 DB |
2957 | zio_link_t *zl = NULL; |
2958 | while (zio_walk_parents(zio, &zl) != NULL) | |
428870ff BB |
2959 | ddt_phys_addref(ddp); |
2960 | } else { | |
2961 | ddt_phys_clear(ddp); | |
2962 | } | |
2963 | ||
2964 | ddt_exit(ddt); | |
2965 | } | |
2966 | ||
2967 | static void | |
2968 | zio_ddt_ditto_write_done(zio_t *zio) | |
2969 | { | |
2970 | int p = DDT_PHYS_DITTO; | |
1c27024e | 2971 | ASSERTV(zio_prop_t *zp = &zio->io_prop); |
428870ff BB |
2972 | blkptr_t *bp = zio->io_bp; |
2973 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
2974 | ddt_entry_t *dde = zio->io_private; | |
2975 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
2976 | ddt_key_t *ddk = &dde->dde_key; | |
2977 | ||
2978 | ddt_enter(ddt); | |
2979 | ||
2980 | ASSERT(ddp->ddp_refcnt == 0); | |
2981 | ASSERT(dde->dde_lead_zio[p] == zio); | |
2982 | dde->dde_lead_zio[p] = NULL; | |
2983 | ||
2984 | if (zio->io_error == 0) { | |
2985 | ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); | |
2986 | ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); | |
2987 | ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); | |
2988 | if (ddp->ddp_phys_birth != 0) | |
2989 | ddt_phys_free(ddt, ddk, ddp, zio->io_txg); | |
2990 | ddt_phys_fill(ddp, bp); | |
2991 | } | |
2992 | ||
2993 | ddt_exit(ddt); | |
2994 | } | |
2995 | ||
2996 | static int | |
2997 | zio_ddt_write(zio_t *zio) | |
2998 | { | |
2999 | spa_t *spa = zio->io_spa; | |
3000 | blkptr_t *bp = zio->io_bp; | |
3001 | uint64_t txg = zio->io_txg; | |
3002 | zio_prop_t *zp = &zio->io_prop; | |
3003 | int p = zp->zp_copies; | |
3004 | int ditto_copies; | |
3005 | zio_t *cio = NULL; | |
3006 | zio_t *dio = NULL; | |
3007 | ddt_t *ddt = ddt_select(spa, bp); | |
3008 | ddt_entry_t *dde; | |
3009 | ddt_phys_t *ddp; | |
3010 | ||
3011 | ASSERT(BP_GET_DEDUP(bp)); | |
3012 | ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); | |
3013 | ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); | |
c17bcf83 | 3014 | ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW))); |
428870ff BB |
3015 | |
3016 | ddt_enter(ddt); | |
3017 | dde = ddt_lookup(ddt, bp, B_TRUE); | |
3018 | ddp = &dde->dde_phys[p]; | |
3019 | ||
3020 | if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { | |
3021 | /* | |
3022 | * If we're using a weak checksum, upgrade to a strong checksum | |
3023 | * and try again. If we're already using a strong checksum, | |
3024 | * we can't resolve it, so just convert to an ordinary write. | |
3025 | * (And automatically e-mail a paper to Nature?) | |
3026 | */ | |
3c67d83a TH |
3027 | if (!(zio_checksum_table[zp->zp_checksum].ci_flags & |
3028 | ZCHECKSUM_FLAG_DEDUP)) { | |
428870ff BB |
3029 | zp->zp_checksum = spa_dedup_checksum(spa); |
3030 | zio_pop_transforms(zio); | |
3031 | zio->io_stage = ZIO_STAGE_OPEN; | |
3032 | BP_ZERO(bp); | |
3033 | } else { | |
03c6040b | 3034 | zp->zp_dedup = B_FALSE; |
428870ff BB |
3035 | } |
3036 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
3037 | ddt_exit(ddt); | |
3038 | return (ZIO_PIPELINE_CONTINUE); | |
3039 | } | |
3040 | ||
3041 | ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); | |
3042 | ASSERT(ditto_copies < SPA_DVAS_PER_BP); | |
3043 | ||
3044 | if (ditto_copies > ddt_ditto_copies_present(dde) && | |
3045 | dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { | |
3046 | zio_prop_t czp = *zp; | |
3047 | ||
3048 | czp.zp_copies = ditto_copies; | |
3049 | ||
3050 | /* | |
3051 | * If we arrived here with an override bp, we won't have run | |
3052 | * the transform stack, so we won't have the data we need to | |
3053 | * generate a child i/o. So, toss the override bp and restart. | |
3054 | * This is safe, because using the override bp is just an | |
3055 | * optimization; and it's rare, so the cost doesn't matter. | |
3056 | */ | |
3057 | if (zio->io_bp_override) { | |
3058 | zio_pop_transforms(zio); | |
3059 | zio->io_stage = ZIO_STAGE_OPEN; | |
3060 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
3061 | zio->io_bp_override = NULL; | |
3062 | BP_ZERO(bp); | |
3063 | ddt_exit(ddt); | |
3064 | return (ZIO_PIPELINE_CONTINUE); | |
3065 | } | |
3066 | ||
a6255b7f | 3067 | dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, |
2aa34383 | 3068 | zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL, |
bc77ba73 | 3069 | NULL, zio_ddt_ditto_write_done, dde, zio->io_priority, |
428870ff BB |
3070 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); |
3071 | ||
a6255b7f | 3072 | zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL); |
428870ff BB |
3073 | dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; |
3074 | } | |
3075 | ||
3076 | if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { | |
3077 | if (ddp->ddp_phys_birth != 0) | |
3078 | ddt_bp_fill(ddp, bp, txg); | |
3079 | if (dde->dde_lead_zio[p] != NULL) | |
3080 | zio_add_child(zio, dde->dde_lead_zio[p]); | |
3081 | else | |
3082 | ddt_phys_addref(ddp); | |
3083 | } else if (zio->io_bp_override) { | |
3084 | ASSERT(bp->blk_birth == txg); | |
3085 | ASSERT(BP_EQUAL(bp, zio->io_bp_override)); | |
3086 | ddt_phys_fill(ddp, bp); | |
3087 | ddt_phys_addref(ddp); | |
3088 | } else { | |
a6255b7f | 3089 | cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, |
2aa34383 | 3090 | zio->io_orig_size, zio->io_orig_size, zp, |
bc77ba73 | 3091 | zio_ddt_child_write_ready, NULL, NULL, |
428870ff BB |
3092 | zio_ddt_child_write_done, dde, zio->io_priority, |
3093 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); | |
3094 | ||
a6255b7f | 3095 | zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL); |
428870ff BB |
3096 | dde->dde_lead_zio[p] = cio; |
3097 | } | |
3098 | ||
3099 | ddt_exit(ddt); | |
3100 | ||
3101 | if (cio) | |
3102 | zio_nowait(cio); | |
3103 | if (dio) | |
3104 | zio_nowait(dio); | |
3105 | ||
3106 | return (ZIO_PIPELINE_CONTINUE); | |
3107 | } | |
3108 | ||
3109 | ddt_entry_t *freedde; /* for debugging */ | |
b128c09f | 3110 | |
428870ff BB |
3111 | static int |
3112 | zio_ddt_free(zio_t *zio) | |
3113 | { | |
3114 | spa_t *spa = zio->io_spa; | |
3115 | blkptr_t *bp = zio->io_bp; | |
3116 | ddt_t *ddt = ddt_select(spa, bp); | |
3117 | ddt_entry_t *dde; | |
3118 | ddt_phys_t *ddp; | |
3119 | ||
3120 | ASSERT(BP_GET_DEDUP(bp)); | |
3121 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
3122 | ||
3123 | ddt_enter(ddt); | |
3124 | freedde = dde = ddt_lookup(ddt, bp, B_TRUE); | |
5dc6af0e BB |
3125 | if (dde) { |
3126 | ddp = ddt_phys_select(dde, bp); | |
3127 | if (ddp) | |
3128 | ddt_phys_decref(ddp); | |
3129 | } | |
428870ff BB |
3130 | ddt_exit(ddt); |
3131 | ||
3132 | return (ZIO_PIPELINE_CONTINUE); | |
3133 | } | |
3134 | ||
3135 | /* | |
3136 | * ========================================================================== | |
3137 | * Allocate and free blocks | |
3138 | * ========================================================================== | |
3139 | */ | |
3dfb57a3 DB |
3140 | |
3141 | static zio_t * | |
3142 | zio_io_to_allocate(spa_t *spa) | |
3143 | { | |
3144 | zio_t *zio; | |
3145 | ||
3146 | ASSERT(MUTEX_HELD(&spa->spa_alloc_lock)); | |
3147 | ||
3148 | zio = avl_first(&spa->spa_alloc_tree); | |
3149 | if (zio == NULL) | |
3150 | return (NULL); | |
3151 | ||
3152 | ASSERT(IO_IS_ALLOCATING(zio)); | |
3153 | ||
3154 | /* | |
3155 | * Try to place a reservation for this zio. If we're unable to | |
3156 | * reserve then we throttle. | |
3157 | */ | |
3158 | if (!metaslab_class_throttle_reserve(spa_normal_class(spa), | |
3159 | zio->io_prop.zp_copies, zio, 0)) { | |
3160 | return (NULL); | |
3161 | } | |
3162 | ||
3163 | avl_remove(&spa->spa_alloc_tree, zio); | |
3164 | ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE); | |
3165 | ||
3166 | return (zio); | |
3167 | } | |
3168 | ||
3169 | static int | |
3170 | zio_dva_throttle(zio_t *zio) | |
3171 | { | |
3172 | spa_t *spa = zio->io_spa; | |
3173 | zio_t *nio; | |
3174 | ||
3175 | if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE || | |
3176 | !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled || | |
3177 | zio->io_child_type == ZIO_CHILD_GANG || | |
3178 | zio->io_flags & ZIO_FLAG_NODATA) { | |
3179 | return (ZIO_PIPELINE_CONTINUE); | |
3180 | } | |
3181 | ||
3182 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
3183 | ||
3184 | ASSERT3U(zio->io_queued_timestamp, >, 0); | |
3185 | ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE); | |
3186 | ||
3187 | mutex_enter(&spa->spa_alloc_lock); | |
3188 | ||
3189 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
3190 | avl_add(&spa->spa_alloc_tree, zio); | |
3191 | ||
3192 | nio = zio_io_to_allocate(zio->io_spa); | |
3193 | mutex_exit(&spa->spa_alloc_lock); | |
3194 | ||
3195 | if (nio == zio) | |
3196 | return (ZIO_PIPELINE_CONTINUE); | |
3197 | ||
3198 | if (nio != NULL) { | |
3dfb57a3 DB |
3199 | ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE); |
3200 | /* | |
3201 | * We are passing control to a new zio so make sure that | |
3202 | * it is processed by a different thread. We do this to | |
3203 | * avoid stack overflows that can occur when parents are | |
3204 | * throttled and children are making progress. We allow | |
3205 | * it to go to the head of the taskq since it's already | |
3206 | * been waiting. | |
3207 | */ | |
3208 | zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE); | |
3209 | } | |
3210 | return (ZIO_PIPELINE_STOP); | |
3211 | } | |
3212 | ||
3213 | void | |
3214 | zio_allocate_dispatch(spa_t *spa) | |
3215 | { | |
3216 | zio_t *zio; | |
3217 | ||
3218 | mutex_enter(&spa->spa_alloc_lock); | |
3219 | zio = zio_io_to_allocate(spa); | |
3220 | mutex_exit(&spa->spa_alloc_lock); | |
3221 | if (zio == NULL) | |
3222 | return; | |
3223 | ||
3224 | ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE); | |
3225 | ASSERT0(zio->io_error); | |
3226 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE); | |
3227 | } | |
3228 | ||
34dc7c2f BB |
3229 | static int |
3230 | zio_dva_allocate(zio_t *zio) | |
3231 | { | |
3232 | spa_t *spa = zio->io_spa; | |
428870ff | 3233 | metaslab_class_t *mc = spa_normal_class(spa); |
34dc7c2f BB |
3234 | blkptr_t *bp = zio->io_bp; |
3235 | int error; | |
6d974228 | 3236 | int flags = 0; |
34dc7c2f | 3237 | |
9babb374 BB |
3238 | if (zio->io_gang_leader == NULL) { |
3239 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
3240 | zio->io_gang_leader = zio; | |
3241 | } | |
3242 | ||
34dc7c2f | 3243 | ASSERT(BP_IS_HOLE(bp)); |
c99c9001 | 3244 | ASSERT0(BP_GET_NDVAS(bp)); |
428870ff BB |
3245 | ASSERT3U(zio->io_prop.zp_copies, >, 0); |
3246 | ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); | |
34dc7c2f BB |
3247 | ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); |
3248 | ||
920dd524 | 3249 | flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0; |
3dfb57a3 DB |
3250 | if (zio->io_flags & ZIO_FLAG_NODATA) |
3251 | flags |= METASLAB_DONT_THROTTLE; | |
3252 | if (zio->io_flags & ZIO_FLAG_GANG_CHILD) | |
3253 | flags |= METASLAB_GANG_CHILD; | |
3254 | if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) | |
3255 | flags |= METASLAB_ASYNC_ALLOC; | |
3256 | ||
b128c09f | 3257 | error = metaslab_alloc(spa, mc, zio->io_size, bp, |
4e21fd06 DB |
3258 | zio->io_prop.zp_copies, zio->io_txg, NULL, flags, |
3259 | &zio->io_alloc_list, zio); | |
34dc7c2f | 3260 | |
3dfb57a3 | 3261 | if (error != 0) { |
6d974228 GW |
3262 | spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, " |
3263 | "size %llu, error %d", spa_name(spa), zio, zio->io_size, | |
3264 | error); | |
b128c09f BB |
3265 | if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) |
3266 | return (zio_write_gang_block(zio)); | |
34dc7c2f BB |
3267 | zio->io_error = error; |
3268 | } | |
3269 | ||
3270 | return (ZIO_PIPELINE_CONTINUE); | |
3271 | } | |
3272 | ||
3273 | static int | |
3274 | zio_dva_free(zio_t *zio) | |
3275 | { | |
b128c09f | 3276 | metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); |
34dc7c2f BB |
3277 | |
3278 | return (ZIO_PIPELINE_CONTINUE); | |
3279 | } | |
3280 | ||
3281 | static int | |
3282 | zio_dva_claim(zio_t *zio) | |
3283 | { | |
b128c09f BB |
3284 | int error; |
3285 | ||
3286 | error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); | |
3287 | if (error) | |
3288 | zio->io_error = error; | |
34dc7c2f BB |
3289 | |
3290 | return (ZIO_PIPELINE_CONTINUE); | |
3291 | } | |
3292 | ||
b128c09f BB |
3293 | /* |
3294 | * Undo an allocation. This is used by zio_done() when an I/O fails | |
3295 | * and we want to give back the block we just allocated. | |
3296 | * This handles both normal blocks and gang blocks. | |
3297 | */ | |
3298 | static void | |
3299 | zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) | |
3300 | { | |
b128c09f | 3301 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); |
428870ff | 3302 | ASSERT(zio->io_bp_override == NULL); |
b128c09f BB |
3303 | |
3304 | if (!BP_IS_HOLE(bp)) | |
428870ff | 3305 | metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); |
b128c09f BB |
3306 | |
3307 | if (gn != NULL) { | |
1c27024e | 3308 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
3309 | zio_dva_unallocate(zio, gn->gn_child[g], |
3310 | &gn->gn_gbh->zg_blkptr[g]); | |
3311 | } | |
3312 | } | |
3313 | } | |
3314 | ||
3315 | /* | |
3316 | * Try to allocate an intent log block. Return 0 on success, errno on failure. | |
3317 | */ | |
3318 | int | |
b5256303 TC |
3319 | zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp, |
3320 | uint64_t size, boolean_t *slog) | |
b128c09f | 3321 | { |
428870ff | 3322 | int error = 1; |
4e21fd06 | 3323 | zio_alloc_list_t io_alloc_list; |
b128c09f | 3324 | |
428870ff BB |
3325 | ASSERT(txg > spa_syncing_txg(spa)); |
3326 | ||
4e21fd06 | 3327 | metaslab_trace_init(&io_alloc_list); |
1b7c1e5c GDN |
3328 | error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1, |
3329 | txg, NULL, METASLAB_FASTWRITE, &io_alloc_list, NULL); | |
3330 | if (error == 0) { | |
3331 | *slog = TRUE; | |
3332 | } else { | |
428870ff | 3333 | error = metaslab_alloc(spa, spa_normal_class(spa), size, |
4e21fd06 DB |
3334 | new_bp, 1, txg, NULL, METASLAB_FASTWRITE, |
3335 | &io_alloc_list, NULL); | |
1b7c1e5c GDN |
3336 | if (error == 0) |
3337 | *slog = FALSE; | |
ebf8e3a2 | 3338 | } |
4e21fd06 | 3339 | metaslab_trace_fini(&io_alloc_list); |
b128c09f BB |
3340 | |
3341 | if (error == 0) { | |
3342 | BP_SET_LSIZE(new_bp, size); | |
3343 | BP_SET_PSIZE(new_bp, size); | |
3344 | BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); | |
428870ff BB |
3345 | BP_SET_CHECKSUM(new_bp, |
3346 | spa_version(spa) >= SPA_VERSION_SLIM_ZIL | |
3347 | ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); | |
b128c09f BB |
3348 | BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); |
3349 | BP_SET_LEVEL(new_bp, 0); | |
428870ff | 3350 | BP_SET_DEDUP(new_bp, 0); |
b128c09f | 3351 | BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); |
b5256303 TC |
3352 | |
3353 | /* | |
3354 | * encrypted blocks will require an IV and salt. We generate | |
3355 | * these now since we will not be rewriting the bp at | |
3356 | * rewrite time. | |
3357 | */ | |
3358 | if (os->os_encrypted) { | |
3359 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
3360 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
3361 | ||
3362 | BP_SET_CRYPT(new_bp, B_TRUE); | |
3363 | VERIFY0(spa_crypt_get_salt(spa, | |
3364 | dmu_objset_id(os), salt)); | |
3365 | VERIFY0(zio_crypt_generate_iv(iv)); | |
3366 | ||
3367 | zio_crypt_encode_params_bp(new_bp, salt, iv); | |
3368 | } | |
1ce23dca PS |
3369 | } else { |
3370 | zfs_dbgmsg("%s: zil block allocation failure: " | |
3371 | "size %llu, error %d", spa_name(spa), size, error); | |
b128c09f BB |
3372 | } |
3373 | ||
3374 | return (error); | |
3375 | } | |
3376 | ||
3377 | /* | |
428870ff | 3378 | * Free an intent log block. |
b128c09f BB |
3379 | */ |
3380 | void | |
428870ff | 3381 | zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp) |
b128c09f | 3382 | { |
428870ff | 3383 | ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG); |
b128c09f BB |
3384 | ASSERT(!BP_IS_GANG(bp)); |
3385 | ||
428870ff | 3386 | zio_free(spa, txg, bp); |
b128c09f BB |
3387 | } |
3388 | ||
34dc7c2f BB |
3389 | /* |
3390 | * ========================================================================== | |
3391 | * Read and write to physical devices | |
3392 | * ========================================================================== | |
3393 | */ | |
98b25418 GW |
3394 | |
3395 | ||
3396 | /* | |
3397 | * Issue an I/O to the underlying vdev. Typically the issue pipeline | |
3398 | * stops after this stage and will resume upon I/O completion. | |
3399 | * However, there are instances where the vdev layer may need to | |
3400 | * continue the pipeline when an I/O was not issued. Since the I/O | |
3401 | * that was sent to the vdev layer might be different than the one | |
3402 | * currently active in the pipeline (see vdev_queue_io()), we explicitly | |
3403 | * force the underlying vdev layers to call either zio_execute() or | |
3404 | * zio_interrupt() to ensure that the pipeline continues with the correct I/O. | |
3405 | */ | |
34dc7c2f BB |
3406 | static int |
3407 | zio_vdev_io_start(zio_t *zio) | |
3408 | { | |
3409 | vdev_t *vd = zio->io_vd; | |
34dc7c2f BB |
3410 | uint64_t align; |
3411 | spa_t *spa = zio->io_spa; | |
3412 | ||
193a37cb TH |
3413 | zio->io_delay = 0; |
3414 | ||
b128c09f BB |
3415 | ASSERT(zio->io_error == 0); |
3416 | ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); | |
34dc7c2f | 3417 | |
b128c09f BB |
3418 | if (vd == NULL) { |
3419 | if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
3420 | spa_config_enter(spa, SCL_ZIO, zio, RW_READER); | |
34dc7c2f | 3421 | |
b128c09f BB |
3422 | /* |
3423 | * The mirror_ops handle multiple DVAs in a single BP. | |
3424 | */ | |
98b25418 GW |
3425 | vdev_mirror_ops.vdev_op_io_start(zio); |
3426 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f BB |
3427 | } |
3428 | ||
3dfb57a3 DB |
3429 | ASSERT3P(zio->io_logical, !=, zio); |
3430 | ||
b128c09f BB |
3431 | align = 1ULL << vd->vdev_top->vdev_ashift; |
3432 | ||
b02fe35d AR |
3433 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && |
3434 | P2PHASE(zio->io_size, align) != 0) { | |
3435 | /* Transform logical writes to be a full physical block size. */ | |
34dc7c2f | 3436 | uint64_t asize = P2ROUNDUP(zio->io_size, align); |
a6255b7f | 3437 | abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize); |
178e73b3 | 3438 | ASSERT(vd == vd->vdev_top); |
34dc7c2f | 3439 | if (zio->io_type == ZIO_TYPE_WRITE) { |
a6255b7f DQ |
3440 | abd_copy(abuf, zio->io_abd, zio->io_size); |
3441 | abd_zero_off(abuf, zio->io_size, asize - zio->io_size); | |
34dc7c2f | 3442 | } |
b128c09f | 3443 | zio_push_transform(zio, abuf, asize, asize, zio_subblock); |
34dc7c2f BB |
3444 | } |
3445 | ||
b02fe35d AR |
3446 | /* |
3447 | * If this is not a physical io, make sure that it is properly aligned | |
3448 | * before proceeding. | |
3449 | */ | |
3450 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { | |
3451 | ASSERT0(P2PHASE(zio->io_offset, align)); | |
3452 | ASSERT0(P2PHASE(zio->io_size, align)); | |
3453 | } else { | |
3454 | /* | |
3455 | * For physical writes, we allow 512b aligned writes and assume | |
3456 | * the device will perform a read-modify-write as necessary. | |
3457 | */ | |
3458 | ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); | |
3459 | ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); | |
3460 | } | |
3461 | ||
572e2857 | 3462 | VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); |
fb5f0bc8 BB |
3463 | |
3464 | /* | |
3465 | * If this is a repair I/O, and there's no self-healing involved -- | |
3466 | * that is, we're just resilvering what we expect to resilver -- | |
3467 | * then don't do the I/O unless zio's txg is actually in vd's DTL. | |
3468 | * This prevents spurious resilvering with nested replication. | |
3469 | * For example, given a mirror of mirrors, (A+B)+(C+D), if only | |
3470 | * A is out of date, we'll read from C+D, then use the data to | |
3471 | * resilver A+B -- but we don't actually want to resilver B, just A. | |
3472 | * The top-level mirror has no way to know this, so instead we just | |
3473 | * discard unnecessary repairs as we work our way down the vdev tree. | |
3474 | * The same logic applies to any form of nested replication: | |
3475 | * ditto + mirror, RAID-Z + replacing, etc. This covers them all. | |
3476 | */ | |
3477 | if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && | |
3478 | !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && | |
3479 | zio->io_txg != 0 && /* not a delegated i/o */ | |
3480 | !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { | |
3481 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
fb5f0bc8 BB |
3482 | zio_vdev_io_bypass(zio); |
3483 | return (ZIO_PIPELINE_CONTINUE); | |
3484 | } | |
34dc7c2f | 3485 | |
b128c09f BB |
3486 | if (vd->vdev_ops->vdev_op_leaf && |
3487 | (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) { | |
3488 | ||
b0bc7a84 | 3489 | if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio)) |
d164b209 | 3490 | return (ZIO_PIPELINE_CONTINUE); |
b128c09f BB |
3491 | |
3492 | if ((zio = vdev_queue_io(zio)) == NULL) | |
3493 | return (ZIO_PIPELINE_STOP); | |
3494 | ||
3495 | if (!vdev_accessible(vd, zio)) { | |
2e528b49 | 3496 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
3497 | zio_interrupt(zio); |
3498 | return (ZIO_PIPELINE_STOP); | |
3499 | } | |
67103816 | 3500 | zio->io_delay = gethrtime(); |
b128c09f BB |
3501 | } |
3502 | ||
98b25418 GW |
3503 | vd->vdev_ops->vdev_op_io_start(zio); |
3504 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f BB |
3505 | } |
3506 | ||
3507 | static int | |
3508 | zio_vdev_io_done(zio_t *zio) | |
3509 | { | |
b128c09f BB |
3510 | vdev_t *vd = zio->io_vd; |
3511 | vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; | |
3512 | boolean_t unexpected_error = B_FALSE; | |
34dc7c2f | 3513 | |
b128c09f BB |
3514 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) |
3515 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f | 3516 | |
b128c09f BB |
3517 | ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); |
3518 | ||
193a37cb TH |
3519 | if (zio->io_delay) |
3520 | zio->io_delay = gethrtime() - zio->io_delay; | |
3521 | ||
b128c09f BB |
3522 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { |
3523 | ||
3524 | vdev_queue_io_done(zio); | |
3525 | ||
3526 | if (zio->io_type == ZIO_TYPE_WRITE) | |
3527 | vdev_cache_write(zio); | |
3528 | ||
3529 | if (zio_injection_enabled && zio->io_error == 0) | |
d977122d DB |
3530 | zio->io_error = zio_handle_device_injections(vd, zio, |
3531 | EIO, EILSEQ); | |
b128c09f BB |
3532 | |
3533 | if (zio_injection_enabled && zio->io_error == 0) | |
3534 | zio->io_error = zio_handle_label_injection(zio, EIO); | |
3535 | ||
3536 | if (zio->io_error) { | |
3537 | if (!vdev_accessible(vd, zio)) { | |
2e528b49 | 3538 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
3539 | } else { |
3540 | unexpected_error = B_TRUE; | |
3541 | } | |
3542 | } | |
3543 | } | |
3544 | ||
3545 | ops->vdev_op_io_done(zio); | |
34dc7c2f | 3546 | |
f43615d0 | 3547 | if (unexpected_error) |
d164b209 | 3548 | VERIFY(vdev_probe(vd, zio) == NULL); |
34dc7c2f | 3549 | |
b128c09f | 3550 | return (ZIO_PIPELINE_CONTINUE); |
34dc7c2f BB |
3551 | } |
3552 | ||
a8b2e306 TC |
3553 | /* |
3554 | * This function is used to change the priority of an existing zio that is | |
3555 | * currently in-flight. This is used by the arc to upgrade priority in the | |
3556 | * event that a demand read is made for a block that is currently queued | |
3557 | * as a scrub or async read IO. Otherwise, the high priority read request | |
3558 | * would end up having to wait for the lower priority IO. | |
3559 | */ | |
3560 | void | |
3561 | zio_change_priority(zio_t *pio, zio_priority_t priority) | |
3562 | { | |
3563 | zio_t *cio, *cio_next; | |
3564 | zio_link_t *zl = NULL; | |
3565 | ||
3566 | ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
3567 | ||
3568 | if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) { | |
3569 | vdev_queue_change_io_priority(pio, priority); | |
3570 | } else { | |
3571 | pio->io_priority = priority; | |
3572 | } | |
3573 | ||
3574 | mutex_enter(&pio->io_lock); | |
3575 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { | |
3576 | cio_next = zio_walk_children(pio, &zl); | |
3577 | zio_change_priority(cio, priority); | |
3578 | } | |
3579 | mutex_exit(&pio->io_lock); | |
3580 | } | |
3581 | ||
428870ff BB |
3582 | /* |
3583 | * For non-raidz ZIOs, we can just copy aside the bad data read from the | |
3584 | * disk, and use that to finish the checksum ereport later. | |
3585 | */ | |
3586 | static void | |
3587 | zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, | |
84c07ada | 3588 | const abd_t *good_buf) |
428870ff BB |
3589 | { |
3590 | /* no processing needed */ | |
3591 | zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); | |
3592 | } | |
3593 | ||
3594 | /*ARGSUSED*/ | |
3595 | void | |
3596 | zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) | |
3597 | { | |
84c07ada | 3598 | void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size); |
428870ff | 3599 | |
84c07ada | 3600 | abd_copy(abd, zio->io_abd, zio->io_size); |
428870ff BB |
3601 | |
3602 | zcr->zcr_cbinfo = zio->io_size; | |
84c07ada | 3603 | zcr->zcr_cbdata = abd; |
428870ff | 3604 | zcr->zcr_finish = zio_vsd_default_cksum_finish; |
84c07ada | 3605 | zcr->zcr_free = zio_abd_free; |
428870ff BB |
3606 | } |
3607 | ||
34dc7c2f BB |
3608 | static int |
3609 | zio_vdev_io_assess(zio_t *zio) | |
3610 | { | |
3611 | vdev_t *vd = zio->io_vd; | |
b128c09f BB |
3612 | |
3613 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) | |
3614 | return (ZIO_PIPELINE_STOP); | |
3615 | ||
3616 | if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
3617 | spa_config_exit(zio->io_spa, SCL_ZIO, zio); | |
3618 | ||
3619 | if (zio->io_vsd != NULL) { | |
428870ff | 3620 | zio->io_vsd_ops->vsd_free(zio); |
b128c09f | 3621 | zio->io_vsd = NULL; |
34dc7c2f BB |
3622 | } |
3623 | ||
b128c09f | 3624 | if (zio_injection_enabled && zio->io_error == 0) |
34dc7c2f BB |
3625 | zio->io_error = zio_handle_fault_injection(zio, EIO); |
3626 | ||
3627 | /* | |
3628 | * If the I/O failed, determine whether we should attempt to retry it. | |
428870ff BB |
3629 | * |
3630 | * On retry, we cut in line in the issue queue, since we don't want | |
3631 | * compression/checksumming/etc. work to prevent our (cheap) IO reissue. | |
34dc7c2f | 3632 | */ |
b128c09f BB |
3633 | if (zio->io_error && vd == NULL && |
3634 | !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { | |
3635 | ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ | |
3636 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ | |
34dc7c2f | 3637 | zio->io_error = 0; |
b128c09f BB |
3638 | zio->io_flags |= ZIO_FLAG_IO_RETRY | |
3639 | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; | |
428870ff BB |
3640 | zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; |
3641 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, | |
3642 | zio_requeue_io_start_cut_in_line); | |
b128c09f | 3643 | return (ZIO_PIPELINE_STOP); |
34dc7c2f BB |
3644 | } |
3645 | ||
b128c09f BB |
3646 | /* |
3647 | * If we got an error on a leaf device, convert it to ENXIO | |
3648 | * if the device is not accessible at all. | |
3649 | */ | |
3650 | if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && | |
3651 | !vdev_accessible(vd, zio)) | |
2e528b49 | 3652 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
3653 | |
3654 | /* | |
3655 | * If we can't write to an interior vdev (mirror or RAID-Z), | |
3656 | * set vdev_cant_write so that we stop trying to allocate from it. | |
3657 | */ | |
3658 | if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && | |
13fe0198 | 3659 | vd != NULL && !vd->vdev_ops->vdev_op_leaf) { |
b128c09f | 3660 | vd->vdev_cant_write = B_TRUE; |
13fe0198 | 3661 | } |
b128c09f | 3662 | |
298ec40b GM |
3663 | /* |
3664 | * If a cache flush returns ENOTSUP or ENOTTY, we know that no future | |
3665 | * attempts will ever succeed. In this case we set a persistent bit so | |
3666 | * that we don't bother with it in the future. | |
3667 | */ | |
3668 | if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) && | |
3669 | zio->io_type == ZIO_TYPE_IOCTL && | |
3670 | zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL) | |
3671 | vd->vdev_nowritecache = B_TRUE; | |
3672 | ||
b128c09f BB |
3673 | if (zio->io_error) |
3674 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
3675 | ||
e8b96c60 MA |
3676 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf && |
3677 | zio->io_physdone != NULL) { | |
3678 | ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); | |
3679 | ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); | |
3680 | zio->io_physdone(zio->io_logical); | |
3681 | } | |
3682 | ||
34dc7c2f BB |
3683 | return (ZIO_PIPELINE_CONTINUE); |
3684 | } | |
3685 | ||
3686 | void | |
3687 | zio_vdev_io_reissue(zio_t *zio) | |
3688 | { | |
3689 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
3690 | ASSERT(zio->io_error == 0); | |
3691 | ||
428870ff | 3692 | zio->io_stage >>= 1; |
34dc7c2f BB |
3693 | } |
3694 | ||
3695 | void | |
3696 | zio_vdev_io_redone(zio_t *zio) | |
3697 | { | |
3698 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); | |
3699 | ||
428870ff | 3700 | zio->io_stage >>= 1; |
34dc7c2f BB |
3701 | } |
3702 | ||
3703 | void | |
3704 | zio_vdev_io_bypass(zio_t *zio) | |
3705 | { | |
3706 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
3707 | ASSERT(zio->io_error == 0); | |
3708 | ||
3709 | zio->io_flags |= ZIO_FLAG_IO_BYPASS; | |
428870ff | 3710 | zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; |
34dc7c2f BB |
3711 | } |
3712 | ||
b5256303 TC |
3713 | /* |
3714 | * ========================================================================== | |
3715 | * Encrypt and store encryption parameters | |
3716 | * ========================================================================== | |
3717 | */ | |
3718 | ||
3719 | ||
3720 | /* | |
3721 | * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for | |
3722 | * managing the storage of encryption parameters and passing them to the | |
3723 | * lower-level encryption functions. | |
3724 | */ | |
3725 | static int | |
3726 | zio_encrypt(zio_t *zio) | |
3727 | { | |
3728 | zio_prop_t *zp = &zio->io_prop; | |
3729 | spa_t *spa = zio->io_spa; | |
3730 | blkptr_t *bp = zio->io_bp; | |
3731 | uint64_t psize = BP_GET_PSIZE(bp); | |
ae76f45c | 3732 | uint64_t dsobj = zio->io_bookmark.zb_objset; |
b5256303 TC |
3733 | dmu_object_type_t ot = BP_GET_TYPE(bp); |
3734 | void *enc_buf = NULL; | |
3735 | abd_t *eabd = NULL; | |
3736 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
3737 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
3738 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
3739 | boolean_t no_crypt = B_FALSE; | |
3740 | ||
3741 | /* the root zio already encrypted the data */ | |
3742 | if (zio->io_child_type == ZIO_CHILD_GANG) | |
3743 | return (ZIO_PIPELINE_CONTINUE); | |
3744 | ||
3745 | /* only ZIL blocks are re-encrypted on rewrite */ | |
3746 | if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG) | |
3747 | return (ZIO_PIPELINE_CONTINUE); | |
3748 | ||
3749 | if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) { | |
3750 | BP_SET_CRYPT(bp, B_FALSE); | |
3751 | return (ZIO_PIPELINE_CONTINUE); | |
3752 | } | |
3753 | ||
3754 | /* if we are doing raw encryption set the provided encryption params */ | |
3755 | if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) { | |
ae76f45c | 3756 | ASSERT0(BP_GET_LEVEL(bp)); |
b5256303 TC |
3757 | BP_SET_CRYPT(bp, B_TRUE); |
3758 | BP_SET_BYTEORDER(bp, zp->zp_byteorder); | |
3759 | if (ot != DMU_OT_OBJSET) | |
3760 | zio_crypt_encode_mac_bp(bp, zp->zp_mac); | |
ae76f45c TC |
3761 | |
3762 | /* dnode blocks must be written out in the provided byteorder */ | |
3763 | if (zp->zp_byteorder != ZFS_HOST_BYTEORDER && | |
3764 | ot == DMU_OT_DNODE) { | |
3765 | void *bswap_buf = zio_buf_alloc(psize); | |
3766 | abd_t *babd = abd_get_from_buf(bswap_buf, psize); | |
3767 | ||
3768 | ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); | |
3769 | abd_copy_to_buf(bswap_buf, zio->io_abd, psize); | |
3770 | dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf, | |
3771 | psize); | |
3772 | ||
3773 | abd_take_ownership_of_buf(babd, B_TRUE); | |
3774 | zio_push_transform(zio, babd, psize, psize, NULL); | |
3775 | } | |
3776 | ||
b5256303 TC |
3777 | if (DMU_OT_IS_ENCRYPTED(ot)) |
3778 | zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv); | |
3779 | return (ZIO_PIPELINE_CONTINUE); | |
3780 | } | |
3781 | ||
3782 | /* indirect blocks only maintain a cksum of the lower level MACs */ | |
3783 | if (BP_GET_LEVEL(bp) > 0) { | |
3784 | BP_SET_CRYPT(bp, B_TRUE); | |
3785 | VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE, | |
3786 | zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp), | |
3787 | mac)); | |
3788 | zio_crypt_encode_mac_bp(bp, mac); | |
3789 | return (ZIO_PIPELINE_CONTINUE); | |
3790 | } | |
3791 | ||
3792 | /* | |
3793 | * Objset blocks are a special case since they have 2 256-bit MACs | |
3794 | * embedded within them. | |
3795 | */ | |
3796 | if (ot == DMU_OT_OBJSET) { | |
3797 | ASSERT0(DMU_OT_IS_ENCRYPTED(ot)); | |
3798 | ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); | |
3799 | BP_SET_CRYPT(bp, B_TRUE); | |
ae76f45c TC |
3800 | VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj, |
3801 | zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp))); | |
b5256303 TC |
3802 | return (ZIO_PIPELINE_CONTINUE); |
3803 | } | |
3804 | ||
3805 | /* unencrypted object types are only authenticated with a MAC */ | |
3806 | if (!DMU_OT_IS_ENCRYPTED(ot)) { | |
3807 | BP_SET_CRYPT(bp, B_TRUE); | |
ae76f45c TC |
3808 | VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj, |
3809 | zio->io_abd, psize, mac)); | |
b5256303 TC |
3810 | zio_crypt_encode_mac_bp(bp, mac); |
3811 | return (ZIO_PIPELINE_CONTINUE); | |
3812 | } | |
3813 | ||
3814 | /* | |
3815 | * Later passes of sync-to-convergence may decide to rewrite data | |
3816 | * in place to avoid more disk reallocations. This presents a problem | |
3817 | * for encryption because this consitutes rewriting the new data with | |
3818 | * the same encryption key and IV. However, this only applies to blocks | |
3819 | * in the MOS (particularly the spacemaps) and we do not encrypt the | |
3820 | * MOS. We assert that the zio is allocating or an intent log write | |
3821 | * to enforce this. | |
3822 | */ | |
3823 | ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG); | |
3824 | ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG); | |
3825 | ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION)); | |
3826 | ASSERT3U(psize, !=, 0); | |
3827 | ||
3828 | enc_buf = zio_buf_alloc(psize); | |
3829 | eabd = abd_get_from_buf(enc_buf, psize); | |
3830 | abd_take_ownership_of_buf(eabd, B_TRUE); | |
3831 | ||
3832 | /* | |
3833 | * For an explanation of what encryption parameters are stored | |
3834 | * where, see the block comment in zio_crypt.c. | |
3835 | */ | |
3836 | if (ot == DMU_OT_INTENT_LOG) { | |
3837 | zio_crypt_decode_params_bp(bp, salt, iv); | |
3838 | } else { | |
3839 | BP_SET_CRYPT(bp, B_TRUE); | |
3840 | } | |
3841 | ||
3842 | /* Perform the encryption. This should not fail */ | |
ae76f45c TC |
3843 | VERIFY0(spa_do_crypt_abd(B_TRUE, spa, dsobj, bp, zio->io_txg, |
3844 | psize, zio->io_abd, eabd, iv, mac, salt, &no_crypt)); | |
b5256303 TC |
3845 | |
3846 | /* encode encryption metadata into the bp */ | |
3847 | if (ot == DMU_OT_INTENT_LOG) { | |
3848 | /* | |
3849 | * ZIL blocks store the MAC in the embedded checksum, so the | |
3850 | * transform must always be applied. | |
3851 | */ | |
3852 | zio_crypt_encode_mac_zil(enc_buf, mac); | |
3853 | zio_push_transform(zio, eabd, psize, psize, NULL); | |
3854 | } else { | |
3855 | BP_SET_CRYPT(bp, B_TRUE); | |
3856 | zio_crypt_encode_params_bp(bp, salt, iv); | |
3857 | zio_crypt_encode_mac_bp(bp, mac); | |
3858 | ||
3859 | if (no_crypt) { | |
3860 | ASSERT3U(ot, ==, DMU_OT_DNODE); | |
3861 | abd_free(eabd); | |
3862 | } else { | |
3863 | zio_push_transform(zio, eabd, psize, psize, NULL); | |
3864 | } | |
3865 | } | |
3866 | ||
3867 | return (ZIO_PIPELINE_CONTINUE); | |
3868 | } | |
3869 | ||
34dc7c2f BB |
3870 | /* |
3871 | * ========================================================================== | |
3872 | * Generate and verify checksums | |
3873 | * ========================================================================== | |
3874 | */ | |
3875 | static int | |
3876 | zio_checksum_generate(zio_t *zio) | |
3877 | { | |
34dc7c2f | 3878 | blkptr_t *bp = zio->io_bp; |
b128c09f | 3879 | enum zio_checksum checksum; |
34dc7c2f | 3880 | |
b128c09f BB |
3881 | if (bp == NULL) { |
3882 | /* | |
3883 | * This is zio_write_phys(). | |
3884 | * We're either generating a label checksum, or none at all. | |
3885 | */ | |
3886 | checksum = zio->io_prop.zp_checksum; | |
34dc7c2f | 3887 | |
b128c09f BB |
3888 | if (checksum == ZIO_CHECKSUM_OFF) |
3889 | return (ZIO_PIPELINE_CONTINUE); | |
3890 | ||
3891 | ASSERT(checksum == ZIO_CHECKSUM_LABEL); | |
3892 | } else { | |
3893 | if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { | |
3894 | ASSERT(!IO_IS_ALLOCATING(zio)); | |
3895 | checksum = ZIO_CHECKSUM_GANG_HEADER; | |
3896 | } else { | |
3897 | checksum = BP_GET_CHECKSUM(bp); | |
3898 | } | |
3899 | } | |
34dc7c2f | 3900 | |
a6255b7f | 3901 | zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size); |
34dc7c2f BB |
3902 | |
3903 | return (ZIO_PIPELINE_CONTINUE); | |
3904 | } | |
3905 | ||
3906 | static int | |
b128c09f | 3907 | zio_checksum_verify(zio_t *zio) |
34dc7c2f | 3908 | { |
428870ff | 3909 | zio_bad_cksum_t info; |
b128c09f BB |
3910 | blkptr_t *bp = zio->io_bp; |
3911 | int error; | |
34dc7c2f | 3912 | |
428870ff BB |
3913 | ASSERT(zio->io_vd != NULL); |
3914 | ||
b128c09f BB |
3915 | if (bp == NULL) { |
3916 | /* | |
3917 | * This is zio_read_phys(). | |
3918 | * We're either verifying a label checksum, or nothing at all. | |
3919 | */ | |
3920 | if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) | |
3921 | return (ZIO_PIPELINE_CONTINUE); | |
34dc7c2f | 3922 | |
b128c09f BB |
3923 | ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); |
3924 | } | |
34dc7c2f | 3925 | |
428870ff | 3926 | if ((error = zio_checksum_error(zio, &info)) != 0) { |
b128c09f | 3927 | zio->io_error = error; |
7a3066ff MA |
3928 | if (error == ECKSUM && |
3929 | !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { | |
428870ff | 3930 | zfs_ereport_start_checksum(zio->io_spa, |
b5256303 TC |
3931 | zio->io_vd, &zio->io_bookmark, zio, |
3932 | zio->io_offset, zio->io_size, NULL, &info); | |
b128c09f | 3933 | } |
34dc7c2f BB |
3934 | } |
3935 | ||
3936 | return (ZIO_PIPELINE_CONTINUE); | |
3937 | } | |
3938 | ||
3939 | /* | |
3940 | * Called by RAID-Z to ensure we don't compute the checksum twice. | |
3941 | */ | |
3942 | void | |
3943 | zio_checksum_verified(zio_t *zio) | |
3944 | { | |
428870ff | 3945 | zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; |
34dc7c2f BB |
3946 | } |
3947 | ||
3948 | /* | |
b128c09f BB |
3949 | * ========================================================================== |
3950 | * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. | |
9b67f605 | 3951 | * An error of 0 indicates success. ENXIO indicates whole-device failure, |
b128c09f BB |
3952 | * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO |
3953 | * indicate errors that are specific to one I/O, and most likely permanent. | |
3954 | * Any other error is presumed to be worse because we weren't expecting it. | |
3955 | * ========================================================================== | |
34dc7c2f | 3956 | */ |
b128c09f BB |
3957 | int |
3958 | zio_worst_error(int e1, int e2) | |
34dc7c2f | 3959 | { |
b128c09f BB |
3960 | static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; |
3961 | int r1, r2; | |
3962 | ||
3963 | for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) | |
3964 | if (e1 == zio_error_rank[r1]) | |
3965 | break; | |
34dc7c2f | 3966 | |
b128c09f BB |
3967 | for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) |
3968 | if (e2 == zio_error_rank[r2]) | |
3969 | break; | |
3970 | ||
3971 | return (r1 > r2 ? e1 : e2); | |
34dc7c2f BB |
3972 | } |
3973 | ||
3974 | /* | |
3975 | * ========================================================================== | |
b128c09f | 3976 | * I/O completion |
34dc7c2f BB |
3977 | * ========================================================================== |
3978 | */ | |
b128c09f BB |
3979 | static int |
3980 | zio_ready(zio_t *zio) | |
34dc7c2f | 3981 | { |
b128c09f | 3982 | blkptr_t *bp = zio->io_bp; |
d164b209 | 3983 | zio_t *pio, *pio_next; |
3dfb57a3 | 3984 | zio_link_t *zl = NULL; |
34dc7c2f | 3985 | |
428870ff BB |
3986 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || |
3987 | zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY)) | |
9babb374 | 3988 | return (ZIO_PIPELINE_STOP); |
34dc7c2f | 3989 | |
9babb374 | 3990 | if (zio->io_ready) { |
b128c09f | 3991 | ASSERT(IO_IS_ALLOCATING(zio)); |
03c6040b GW |
3992 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || |
3993 | (zio->io_flags & ZIO_FLAG_NOPWRITE)); | |
b128c09f | 3994 | ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); |
34dc7c2f | 3995 | |
b128c09f BB |
3996 | zio->io_ready(zio); |
3997 | } | |
34dc7c2f | 3998 | |
b128c09f BB |
3999 | if (bp != NULL && bp != &zio->io_bp_copy) |
4000 | zio->io_bp_copy = *bp; | |
34dc7c2f | 4001 | |
3dfb57a3 | 4002 | if (zio->io_error != 0) { |
b128c09f | 4003 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f | 4004 | |
3dfb57a3 DB |
4005 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
4006 | ASSERT(IO_IS_ALLOCATING(zio)); | |
4007 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
4008 | /* | |
4009 | * We were unable to allocate anything, unreserve and | |
4010 | * issue the next I/O to allocate. | |
4011 | */ | |
4012 | metaslab_class_throttle_unreserve( | |
4013 | spa_normal_class(zio->io_spa), | |
4014 | zio->io_prop.zp_copies, zio); | |
4015 | zio_allocate_dispatch(zio->io_spa); | |
4016 | } | |
4017 | } | |
4018 | ||
d164b209 BB |
4019 | mutex_enter(&zio->io_lock); |
4020 | zio->io_state[ZIO_WAIT_READY] = 1; | |
3dfb57a3 | 4021 | pio = zio_walk_parents(zio, &zl); |
d164b209 BB |
4022 | mutex_exit(&zio->io_lock); |
4023 | ||
4024 | /* | |
4025 | * As we notify zio's parents, new parents could be added. | |
4026 | * New parents go to the head of zio's io_parent_list, however, | |
4027 | * so we will (correctly) not notify them. The remainder of zio's | |
4028 | * io_parent_list, from 'pio_next' onward, cannot change because | |
4029 | * all parents must wait for us to be done before they can be done. | |
4030 | */ | |
4031 | for (; pio != NULL; pio = pio_next) { | |
3dfb57a3 | 4032 | pio_next = zio_walk_parents(zio, &zl); |
b128c09f | 4033 | zio_notify_parent(pio, zio, ZIO_WAIT_READY); |
d164b209 | 4034 | } |
34dc7c2f | 4035 | |
428870ff BB |
4036 | if (zio->io_flags & ZIO_FLAG_NODATA) { |
4037 | if (BP_IS_GANG(bp)) { | |
4038 | zio->io_flags &= ~ZIO_FLAG_NODATA; | |
4039 | } else { | |
a6255b7f | 4040 | ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE); |
428870ff BB |
4041 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; |
4042 | } | |
4043 | } | |
4044 | ||
4045 | if (zio_injection_enabled && | |
4046 | zio->io_spa->spa_syncing_txg == zio->io_txg) | |
4047 | zio_handle_ignored_writes(zio); | |
4048 | ||
b128c09f | 4049 | return (ZIO_PIPELINE_CONTINUE); |
34dc7c2f BB |
4050 | } |
4051 | ||
3dfb57a3 DB |
4052 | /* |
4053 | * Update the allocation throttle accounting. | |
4054 | */ | |
4055 | static void | |
4056 | zio_dva_throttle_done(zio_t *zio) | |
4057 | { | |
1c27024e | 4058 | ASSERTV(zio_t *lio = zio->io_logical); |
3dfb57a3 DB |
4059 | zio_t *pio = zio_unique_parent(zio); |
4060 | vdev_t *vd = zio->io_vd; | |
4061 | int flags = METASLAB_ASYNC_ALLOC; | |
4062 | ||
4063 | ASSERT3P(zio->io_bp, !=, NULL); | |
4064 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); | |
4065 | ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE); | |
4066 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); | |
4067 | ASSERT(vd != NULL); | |
4068 | ASSERT3P(vd, ==, vd->vdev_top); | |
21df134f SB |
4069 | ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY)); |
4070 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); | |
3dfb57a3 DB |
4071 | ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING); |
4072 | ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
4073 | ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA)); | |
4074 | ||
4075 | /* | |
4076 | * Parents of gang children can have two flavors -- ones that | |
4077 | * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set) | |
4078 | * and ones that allocated the constituent blocks. The allocation | |
4079 | * throttle needs to know the allocating parent zio so we must find | |
4080 | * it here. | |
4081 | */ | |
4082 | if (pio->io_child_type == ZIO_CHILD_GANG) { | |
4083 | /* | |
4084 | * If our parent is a rewrite gang child then our grandparent | |
4085 | * would have been the one that performed the allocation. | |
4086 | */ | |
4087 | if (pio->io_flags & ZIO_FLAG_IO_REWRITE) | |
4088 | pio = zio_unique_parent(pio); | |
4089 | flags |= METASLAB_GANG_CHILD; | |
4090 | } | |
4091 | ||
4092 | ASSERT(IO_IS_ALLOCATING(pio)); | |
4093 | ASSERT3P(zio, !=, zio->io_logical); | |
4094 | ASSERT(zio->io_logical != NULL); | |
4095 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); | |
4096 | ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE); | |
4097 | ||
4098 | mutex_enter(&pio->io_lock); | |
4099 | metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags); | |
4100 | mutex_exit(&pio->io_lock); | |
4101 | ||
4102 | metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa), | |
4103 | 1, pio); | |
4104 | ||
4105 | /* | |
4106 | * Call into the pipeline to see if there is more work that | |
4107 | * needs to be done. If there is work to be done it will be | |
4108 | * dispatched to another taskq thread. | |
4109 | */ | |
4110 | zio_allocate_dispatch(zio->io_spa); | |
4111 | } | |
4112 | ||
b128c09f BB |
4113 | static int |
4114 | zio_done(zio_t *zio) | |
34dc7c2f | 4115 | { |
3dfb57a3 DB |
4116 | /* |
4117 | * Always attempt to keep stack usage minimal here since | |
4118 | * we can be called recurisvely up to 19 levels deep. | |
4119 | */ | |
84c07ada | 4120 | const uint64_t psize = zio->io_size; |
d164b209 | 4121 | zio_t *pio, *pio_next; |
3dfb57a3 | 4122 | zio_link_t *zl = NULL; |
34dc7c2f | 4123 | |
b128c09f | 4124 | /* |
9babb374 | 4125 | * If our children haven't all completed, |
b128c09f BB |
4126 | * wait for them and then repeat this pipeline stage. |
4127 | */ | |
4128 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || | |
4129 | zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || | |
428870ff | 4130 | zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) || |
b128c09f BB |
4131 | zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) |
4132 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f | 4133 | |
3dfb57a3 DB |
4134 | /* |
4135 | * If the allocation throttle is enabled, then update the accounting. | |
4136 | * We only track child I/Os that are part of an allocating async | |
4137 | * write. We must do this since the allocation is performed | |
4138 | * by the logical I/O but the actual write is done by child I/Os. | |
4139 | */ | |
4140 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING && | |
4141 | zio->io_child_type == ZIO_CHILD_VDEV) { | |
4142 | ASSERT(spa_normal_class( | |
4143 | zio->io_spa)->mc_alloc_throttle_enabled); | |
4144 | zio_dva_throttle_done(zio); | |
4145 | } | |
4146 | ||
4147 | /* | |
4148 | * If the allocation throttle is enabled, verify that | |
4149 | * we have decremented the refcounts for every I/O that was throttled. | |
4150 | */ | |
4151 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { | |
4152 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
4153 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
4154 | ASSERT(zio->io_bp != NULL); | |
4155 | metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio); | |
4156 | VERIFY(refcount_not_held( | |
4157 | &(spa_normal_class(zio->io_spa)->mc_alloc_slots), zio)); | |
4158 | } | |
4159 | ||
4160 | ||
1c27024e DB |
4161 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) |
4162 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) | |
b128c09f BB |
4163 | ASSERT(zio->io_children[c][w] == 0); |
4164 | ||
9b67f605 | 4165 | if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) { |
c776b317 BB |
4166 | ASSERT(zio->io_bp->blk_pad[0] == 0); |
4167 | ASSERT(zio->io_bp->blk_pad[1] == 0); | |
d1d7e268 MK |
4168 | ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, |
4169 | sizeof (blkptr_t)) == 0 || | |
c776b317 BB |
4170 | (zio->io_bp == zio_unique_parent(zio)->io_bp)); |
4171 | if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) && | |
428870ff | 4172 | zio->io_bp_override == NULL && |
b128c09f | 4173 | !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { |
d1d7e268 MK |
4174 | ASSERT3U(zio->io_prop.zp_copies, <=, |
4175 | BP_GET_NDVAS(zio->io_bp)); | |
c776b317 | 4176 | ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 || |
d1d7e268 MK |
4177 | (BP_COUNT_GANG(zio->io_bp) == |
4178 | BP_GET_NDVAS(zio->io_bp))); | |
b128c09f | 4179 | } |
03c6040b GW |
4180 | if (zio->io_flags & ZIO_FLAG_NOPWRITE) |
4181 | VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig)); | |
b128c09f BB |
4182 | } |
4183 | ||
4184 | /* | |
428870ff | 4185 | * If there were child vdev/gang/ddt errors, they apply to us now. |
b128c09f BB |
4186 | */ |
4187 | zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); | |
4188 | zio_inherit_child_errors(zio, ZIO_CHILD_GANG); | |
428870ff BB |
4189 | zio_inherit_child_errors(zio, ZIO_CHILD_DDT); |
4190 | ||
4191 | /* | |
4192 | * If the I/O on the transformed data was successful, generate any | |
4193 | * checksum reports now while we still have the transformed data. | |
4194 | */ | |
4195 | if (zio->io_error == 0) { | |
4196 | while (zio->io_cksum_report != NULL) { | |
4197 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
4198 | uint64_t align = zcr->zcr_align; | |
a6255b7f | 4199 | uint64_t asize = P2ROUNDUP(psize, align); |
a6255b7f DQ |
4200 | abd_t *adata = zio->io_abd; |
4201 | ||
4202 | if (asize != psize) { | |
84c07ada | 4203 | adata = abd_alloc(asize, B_TRUE); |
a6255b7f DQ |
4204 | abd_copy(adata, zio->io_abd, psize); |
4205 | abd_zero_off(adata, psize, asize - psize); | |
428870ff BB |
4206 | } |
4207 | ||
4208 | zio->io_cksum_report = zcr->zcr_next; | |
4209 | zcr->zcr_next = NULL; | |
84c07ada | 4210 | zcr->zcr_finish(zcr, adata); |
428870ff BB |
4211 | zfs_ereport_free_checksum(zcr); |
4212 | ||
a6255b7f DQ |
4213 | if (asize != psize) |
4214 | abd_free(adata); | |
428870ff BB |
4215 | } |
4216 | } | |
b128c09f BB |
4217 | |
4218 | zio_pop_transforms(zio); /* note: may set zio->io_error */ | |
4219 | ||
a6255b7f | 4220 | vdev_stat_update(zio, psize); |
b128c09f | 4221 | |
a69052be | 4222 | /* |
cc92e9d0 | 4223 | * If this I/O is attached to a particular vdev is slow, exceeding |
72f53c56 MJ |
4224 | * 30 seconds to complete, post an error described the I/O delay. |
4225 | * We ignore these errors if the device is currently unavailable. | |
a69052be | 4226 | */ |
193a37cb | 4227 | if (zio->io_delay >= MSEC2NSEC(zio_delay_max)) { |
72f53c56 MJ |
4228 | if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) |
4229 | zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa, | |
b5256303 | 4230 | zio->io_vd, &zio->io_bookmark, zio, 0, 0); |
72f53c56 | 4231 | } |
a69052be | 4232 | |
b128c09f BB |
4233 | if (zio->io_error) { |
4234 | /* | |
4235 | * If this I/O is attached to a particular vdev, | |
4236 | * generate an error message describing the I/O failure | |
4237 | * at the block level. We ignore these errors if the | |
4238 | * device is currently unavailable. | |
4239 | */ | |
c776b317 | 4240 | if (zio->io_error != ECKSUM && zio->io_vd != NULL && |
02730c33 | 4241 | !vdev_is_dead(zio->io_vd)) |
c776b317 | 4242 | zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa, |
b5256303 | 4243 | zio->io_vd, &zio->io_bookmark, zio, 0, 0); |
34dc7c2f | 4244 | |
428870ff BB |
4245 | if ((zio->io_error == EIO || !(zio->io_flags & |
4246 | (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && | |
c776b317 | 4247 | zio == zio->io_logical) { |
b128c09f BB |
4248 | /* |
4249 | * For logical I/O requests, tell the SPA to log the | |
4250 | * error and generate a logical data ereport. | |
4251 | */ | |
b5256303 | 4252 | spa_log_error(zio->io_spa, &zio->io_bookmark); |
d1d7e268 | 4253 | zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, |
b5256303 | 4254 | NULL, &zio->io_bookmark, zio, 0, 0); |
b128c09f BB |
4255 | } |
4256 | } | |
34dc7c2f | 4257 | |
c776b317 | 4258 | if (zio->io_error && zio == zio->io_logical) { |
b128c09f BB |
4259 | /* |
4260 | * Determine whether zio should be reexecuted. This will | |
4261 | * propagate all the way to the root via zio_notify_parent(). | |
4262 | */ | |
c776b317 | 4263 | ASSERT(zio->io_vd == NULL && zio->io_bp != NULL); |
428870ff | 4264 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); |
b128c09f | 4265 | |
428870ff BB |
4266 | if (IO_IS_ALLOCATING(zio) && |
4267 | !(zio->io_flags & ZIO_FLAG_CANFAIL)) { | |
b128c09f BB |
4268 | if (zio->io_error != ENOSPC) |
4269 | zio->io_reexecute |= ZIO_REEXECUTE_NOW; | |
4270 | else | |
4271 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff | 4272 | } |
b128c09f BB |
4273 | |
4274 | if ((zio->io_type == ZIO_TYPE_READ || | |
4275 | zio->io_type == ZIO_TYPE_FREE) && | |
572e2857 | 4276 | !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && |
b128c09f | 4277 | zio->io_error == ENXIO && |
c776b317 BB |
4278 | spa_load_state(zio->io_spa) == SPA_LOAD_NONE && |
4279 | spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE) | |
b128c09f BB |
4280 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; |
4281 | ||
4282 | if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) | |
4283 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff BB |
4284 | |
4285 | /* | |
4286 | * Here is a possibly good place to attempt to do | |
4287 | * either combinatorial reconstruction or error correction | |
4288 | * based on checksums. It also might be a good place | |
4289 | * to send out preliminary ereports before we suspend | |
4290 | * processing. | |
4291 | */ | |
34dc7c2f BB |
4292 | } |
4293 | ||
4294 | /* | |
b128c09f BB |
4295 | * If there were logical child errors, they apply to us now. |
4296 | * We defer this until now to avoid conflating logical child | |
4297 | * errors with errors that happened to the zio itself when | |
4298 | * updating vdev stats and reporting FMA events above. | |
34dc7c2f | 4299 | */ |
b128c09f | 4300 | zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); |
34dc7c2f | 4301 | |
428870ff BB |
4302 | if ((zio->io_error || zio->io_reexecute) && |
4303 | IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && | |
03c6040b | 4304 | !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) |
c776b317 | 4305 | zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp); |
9babb374 BB |
4306 | |
4307 | zio_gang_tree_free(&zio->io_gang_tree); | |
4308 | ||
4309 | /* | |
4310 | * Godfather I/Os should never suspend. | |
4311 | */ | |
4312 | if ((zio->io_flags & ZIO_FLAG_GODFATHER) && | |
4313 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) | |
a32494d2 | 4314 | zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND; |
9babb374 | 4315 | |
b128c09f BB |
4316 | if (zio->io_reexecute) { |
4317 | /* | |
4318 | * This is a logical I/O that wants to reexecute. | |
4319 | * | |
4320 | * Reexecute is top-down. When an i/o fails, if it's not | |
4321 | * the root, it simply notifies its parent and sticks around. | |
4322 | * The parent, seeing that it still has children in zio_done(), | |
4323 | * does the same. This percolates all the way up to the root. | |
4324 | * The root i/o will reexecute or suspend the entire tree. | |
4325 | * | |
4326 | * This approach ensures that zio_reexecute() honors | |
4327 | * all the original i/o dependency relationships, e.g. | |
4328 | * parents not executing until children are ready. | |
4329 | */ | |
4330 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
34dc7c2f | 4331 | |
9babb374 | 4332 | zio->io_gang_leader = NULL; |
b128c09f | 4333 | |
d164b209 BB |
4334 | mutex_enter(&zio->io_lock); |
4335 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
4336 | mutex_exit(&zio->io_lock); | |
4337 | ||
9babb374 BB |
4338 | /* |
4339 | * "The Godfather" I/O monitors its children but is | |
4340 | * not a true parent to them. It will track them through | |
4341 | * the pipeline but severs its ties whenever they get into | |
4342 | * trouble (e.g. suspended). This allows "The Godfather" | |
4343 | * I/O to return status without blocking. | |
4344 | */ | |
3dfb57a3 DB |
4345 | zl = NULL; |
4346 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; | |
4347 | pio = pio_next) { | |
4348 | zio_link_t *remove_zl = zl; | |
4349 | pio_next = zio_walk_parents(zio, &zl); | |
9babb374 BB |
4350 | |
4351 | if ((pio->io_flags & ZIO_FLAG_GODFATHER) && | |
4352 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { | |
3dfb57a3 | 4353 | zio_remove_child(pio, zio, remove_zl); |
9babb374 BB |
4354 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE); |
4355 | } | |
4356 | } | |
4357 | ||
d164b209 | 4358 | if ((pio = zio_unique_parent(zio)) != NULL) { |
b128c09f BB |
4359 | /* |
4360 | * We're not a root i/o, so there's nothing to do | |
4361 | * but notify our parent. Don't propagate errors | |
4362 | * upward since we haven't permanently failed yet. | |
4363 | */ | |
9babb374 | 4364 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f BB |
4365 | zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; |
4366 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE); | |
4367 | } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { | |
4368 | /* | |
4369 | * We'd fail again if we reexecuted now, so suspend | |
4370 | * until conditions improve (e.g. device comes online). | |
4371 | */ | |
c776b317 | 4372 | zio_suspend(zio->io_spa, zio); |
b128c09f BB |
4373 | } else { |
4374 | /* | |
4375 | * Reexecution is potentially a huge amount of work. | |
4376 | * Hand it off to the otherwise-unused claim taskq. | |
4377 | */ | |
a38718a6 | 4378 | ASSERT(taskq_empty_ent(&zio->io_tqent)); |
7ef5e54e AL |
4379 | spa_taskq_dispatch_ent(zio->io_spa, |
4380 | ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE, | |
a38718a6 GA |
4381 | (task_func_t *)zio_reexecute, zio, 0, |
4382 | &zio->io_tqent); | |
b128c09f BB |
4383 | } |
4384 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f BB |
4385 | } |
4386 | ||
428870ff | 4387 | ASSERT(zio->io_child_count == 0); |
b128c09f BB |
4388 | ASSERT(zio->io_reexecute == 0); |
4389 | ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); | |
34dc7c2f | 4390 | |
428870ff BB |
4391 | /* |
4392 | * Report any checksum errors, since the I/O is complete. | |
4393 | */ | |
4394 | while (zio->io_cksum_report != NULL) { | |
4395 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
4396 | zio->io_cksum_report = zcr->zcr_next; | |
4397 | zcr->zcr_next = NULL; | |
4398 | zcr->zcr_finish(zcr, NULL); | |
4399 | zfs_ereport_free_checksum(zcr); | |
4400 | } | |
4401 | ||
920dd524 | 4402 | if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp && |
9b67f605 MA |
4403 | !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) && |
4404 | !(zio->io_flags & ZIO_FLAG_NOPWRITE)) { | |
920dd524 ED |
4405 | metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp); |
4406 | } | |
4407 | ||
d164b209 BB |
4408 | /* |
4409 | * It is the responsibility of the done callback to ensure that this | |
4410 | * particular zio is no longer discoverable for adoption, and as | |
4411 | * such, cannot acquire any new parents. | |
4412 | */ | |
b128c09f BB |
4413 | if (zio->io_done) |
4414 | zio->io_done(zio); | |
34dc7c2f | 4415 | |
d164b209 BB |
4416 | mutex_enter(&zio->io_lock); |
4417 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
4418 | mutex_exit(&zio->io_lock); | |
34dc7c2f | 4419 | |
3dfb57a3 DB |
4420 | zl = NULL; |
4421 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) { | |
4422 | zio_link_t *remove_zl = zl; | |
4423 | pio_next = zio_walk_parents(zio, &zl); | |
4424 | zio_remove_child(pio, zio, remove_zl); | |
b128c09f BB |
4425 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE); |
4426 | } | |
34dc7c2f | 4427 | |
b128c09f BB |
4428 | if (zio->io_waiter != NULL) { |
4429 | mutex_enter(&zio->io_lock); | |
4430 | zio->io_executor = NULL; | |
4431 | cv_broadcast(&zio->io_cv); | |
4432 | mutex_exit(&zio->io_lock); | |
4433 | } else { | |
4434 | zio_destroy(zio); | |
4435 | } | |
34dc7c2f | 4436 | |
b128c09f | 4437 | return (ZIO_PIPELINE_STOP); |
34dc7c2f BB |
4438 | } |
4439 | ||
4440 | /* | |
b128c09f BB |
4441 | * ========================================================================== |
4442 | * I/O pipeline definition | |
4443 | * ========================================================================== | |
34dc7c2f | 4444 | */ |
428870ff | 4445 | static zio_pipe_stage_t *zio_pipeline[] = { |
b128c09f | 4446 | NULL, |
b128c09f | 4447 | zio_read_bp_init, |
3dfb57a3 | 4448 | zio_write_bp_init, |
428870ff BB |
4449 | zio_free_bp_init, |
4450 | zio_issue_async, | |
3dfb57a3 | 4451 | zio_write_compress, |
b5256303 | 4452 | zio_encrypt, |
b128c09f | 4453 | zio_checksum_generate, |
03c6040b | 4454 | zio_nop_write, |
428870ff BB |
4455 | zio_ddt_read_start, |
4456 | zio_ddt_read_done, | |
4457 | zio_ddt_write, | |
4458 | zio_ddt_free, | |
b128c09f BB |
4459 | zio_gang_assemble, |
4460 | zio_gang_issue, | |
3dfb57a3 | 4461 | zio_dva_throttle, |
b128c09f BB |
4462 | zio_dva_allocate, |
4463 | zio_dva_free, | |
4464 | zio_dva_claim, | |
4465 | zio_ready, | |
4466 | zio_vdev_io_start, | |
4467 | zio_vdev_io_done, | |
4468 | zio_vdev_io_assess, | |
4469 | zio_checksum_verify, | |
4470 | zio_done | |
4471 | }; | |
c28b2279 | 4472 | |
9ae529ec | 4473 | |
9ae529ec | 4474 | |
9ae529ec | 4475 | |
fcff0f35 PD |
4476 | /* |
4477 | * Compare two zbookmark_phys_t's to see which we would reach first in a | |
4478 | * pre-order traversal of the object tree. | |
4479 | * | |
4480 | * This is simple in every case aside from the meta-dnode object. For all other | |
4481 | * objects, we traverse them in order (object 1 before object 2, and so on). | |
4482 | * However, all of these objects are traversed while traversing object 0, since | |
4483 | * the data it points to is the list of objects. Thus, we need to convert to a | |
4484 | * canonical representation so we can compare meta-dnode bookmarks to | |
4485 | * non-meta-dnode bookmarks. | |
4486 | * | |
4487 | * We do this by calculating "equivalents" for each field of the zbookmark. | |
4488 | * zbookmarks outside of the meta-dnode use their own object and level, and | |
4489 | * calculate the level 0 equivalent (the first L0 blkid that is contained in the | |
4490 | * blocks this bookmark refers to) by multiplying their blkid by their span | |
4491 | * (the number of L0 blocks contained within one block at their level). | |
4492 | * zbookmarks inside the meta-dnode calculate their object equivalent | |
4493 | * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use | |
4494 | * level + 1<<31 (any value larger than a level could ever be) for their level. | |
4495 | * This causes them to always compare before a bookmark in their object | |
4496 | * equivalent, compare appropriately to bookmarks in other objects, and to | |
4497 | * compare appropriately to other bookmarks in the meta-dnode. | |
4498 | */ | |
4499 | int | |
4500 | zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, | |
4501 | const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) | |
4502 | { | |
4503 | /* | |
4504 | * These variables represent the "equivalent" values for the zbookmark, | |
4505 | * after converting zbookmarks inside the meta dnode to their | |
4506 | * normal-object equivalents. | |
4507 | */ | |
4508 | uint64_t zb1obj, zb2obj; | |
4509 | uint64_t zb1L0, zb2L0; | |
4510 | uint64_t zb1level, zb2level; | |
4511 | ||
4512 | if (zb1->zb_object == zb2->zb_object && | |
4513 | zb1->zb_level == zb2->zb_level && | |
4514 | zb1->zb_blkid == zb2->zb_blkid) | |
4515 | return (0); | |
9ae529ec | 4516 | |
fcff0f35 PD |
4517 | /* |
4518 | * BP_SPANB calculates the span in blocks. | |
4519 | */ | |
4520 | zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); | |
4521 | zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); | |
9ae529ec CS |
4522 | |
4523 | if (zb1->zb_object == DMU_META_DNODE_OBJECT) { | |
fcff0f35 PD |
4524 | zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); |
4525 | zb1L0 = 0; | |
4526 | zb1level = zb1->zb_level + COMPARE_META_LEVEL; | |
4527 | } else { | |
4528 | zb1obj = zb1->zb_object; | |
4529 | zb1level = zb1->zb_level; | |
9ae529ec CS |
4530 | } |
4531 | ||
fcff0f35 PD |
4532 | if (zb2->zb_object == DMU_META_DNODE_OBJECT) { |
4533 | zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); | |
4534 | zb2L0 = 0; | |
4535 | zb2level = zb2->zb_level + COMPARE_META_LEVEL; | |
4536 | } else { | |
4537 | zb2obj = zb2->zb_object; | |
4538 | zb2level = zb2->zb_level; | |
4539 | } | |
4540 | ||
4541 | /* Now that we have a canonical representation, do the comparison. */ | |
4542 | if (zb1obj != zb2obj) | |
4543 | return (zb1obj < zb2obj ? -1 : 1); | |
4544 | else if (zb1L0 != zb2L0) | |
4545 | return (zb1L0 < zb2L0 ? -1 : 1); | |
4546 | else if (zb1level != zb2level) | |
4547 | return (zb1level > zb2level ? -1 : 1); | |
4548 | /* | |
4549 | * This can (theoretically) happen if the bookmarks have the same object | |
4550 | * and level, but different blkids, if the block sizes are not the same. | |
4551 | * There is presently no way to change the indirect block sizes | |
4552 | */ | |
4553 | return (0); | |
4554 | } | |
4555 | ||
4556 | /* | |
4557 | * This function checks the following: given that last_block is the place that | |
4558 | * our traversal stopped last time, does that guarantee that we've visited | |
4559 | * every node under subtree_root? Therefore, we can't just use the raw output | |
4560 | * of zbookmark_compare. We have to pass in a modified version of | |
4561 | * subtree_root; by incrementing the block id, and then checking whether | |
4562 | * last_block is before or equal to that, we can tell whether or not having | |
4563 | * visited last_block implies that all of subtree_root's children have been | |
4564 | * visited. | |
4565 | */ | |
4566 | boolean_t | |
4567 | zbookmark_subtree_completed(const dnode_phys_t *dnp, | |
4568 | const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) | |
4569 | { | |
4570 | zbookmark_phys_t mod_zb = *subtree_root; | |
4571 | mod_zb.zb_blkid++; | |
4572 | ASSERT(last_block->zb_level == 0); | |
4573 | ||
4574 | /* The objset_phys_t isn't before anything. */ | |
4575 | if (dnp == NULL) | |
9ae529ec | 4576 | return (B_FALSE); |
fcff0f35 PD |
4577 | |
4578 | /* | |
4579 | * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the | |
4580 | * data block size in sectors, because that variable is only used if | |
4581 | * the bookmark refers to a block in the meta-dnode. Since we don't | |
4582 | * know without examining it what object it refers to, and there's no | |
4583 | * harm in passing in this value in other cases, we always pass it in. | |
4584 | * | |
4585 | * We pass in 0 for the indirect block size shift because zb2 must be | |
4586 | * level 0. The indirect block size is only used to calculate the span | |
4587 | * of the bookmark, but since the bookmark must be level 0, the span is | |
4588 | * always 1, so the math works out. | |
4589 | * | |
4590 | * If you make changes to how the zbookmark_compare code works, be sure | |
4591 | * to make sure that this code still works afterwards. | |
4592 | */ | |
4593 | return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, | |
4594 | 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, | |
4595 | last_block) <= 0); | |
9ae529ec CS |
4596 | } |
4597 | ||
c28b2279 | 4598 | #if defined(_KERNEL) && defined(HAVE_SPL) |
c28b2279 | 4599 | EXPORT_SYMBOL(zio_type_name); |
81971b13 BB |
4600 | EXPORT_SYMBOL(zio_buf_alloc); |
4601 | EXPORT_SYMBOL(zio_data_buf_alloc); | |
4602 | EXPORT_SYMBOL(zio_buf_free); | |
4603 | EXPORT_SYMBOL(zio_data_buf_free); | |
c28b2279 | 4604 | |
a69052be | 4605 | module_param(zio_delay_max, int, 0644); |
c409e464 BB |
4606 | MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event"); |
4607 | ||
4608 | module_param(zio_requeue_io_start_cut_in_line, int, 0644); | |
4609 | MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O"); | |
29dee3ee CP |
4610 | |
4611 | module_param(zfs_sync_pass_deferred_free, int, 0644); | |
4612 | MODULE_PARM_DESC(zfs_sync_pass_deferred_free, | |
d1d7e268 | 4613 | "Defer frees starting in this pass"); |
29dee3ee CP |
4614 | |
4615 | module_param(zfs_sync_pass_dont_compress, int, 0644); | |
4616 | MODULE_PARM_DESC(zfs_sync_pass_dont_compress, | |
d1d7e268 | 4617 | "Don't compress starting in this pass"); |
29dee3ee CP |
4618 | |
4619 | module_param(zfs_sync_pass_rewrite, int, 0644); | |
4620 | MODULE_PARM_DESC(zfs_sync_pass_rewrite, | |
d1d7e268 | 4621 | "Rewrite new bps starting in this pass"); |
3dfb57a3 DB |
4622 | |
4623 | module_param(zio_dva_throttle_enabled, int, 0644); | |
4624 | MODULE_PARM_DESC(zio_dva_throttle_enabled, | |
4625 | "Throttle block allocations in the ZIO pipeline"); | |
c28b2279 | 4626 | #endif |