]>
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. |
bc77ba73 | 23 | * Copyright (c) 2011, 2016 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> |
193a37cb | 42 | #include <sys/time.h> |
26ef0cc7 | 43 | #include <sys/trace_zio.h> |
34dc7c2f | 44 | |
34dc7c2f BB |
45 | /* |
46 | * ========================================================================== | |
47 | * I/O type descriptions | |
48 | * ========================================================================== | |
49 | */ | |
e8b96c60 | 50 | const char *zio_type_name[ZIO_TYPES] = { |
451041db | 51 | "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl" |
428870ff | 52 | }; |
34dc7c2f BB |
53 | |
54 | /* | |
55 | * ========================================================================== | |
56 | * I/O kmem caches | |
57 | * ========================================================================== | |
58 | */ | |
59 | kmem_cache_t *zio_cache; | |
d164b209 | 60 | kmem_cache_t *zio_link_cache; |
34dc7c2f BB |
61 | kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; |
62 | kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
a69052be | 63 | int zio_delay_max = ZIO_DELAY_MAX; |
34dc7c2f | 64 | |
98b25418 GW |
65 | #define ZIO_PIPELINE_CONTINUE 0x100 |
66 | #define ZIO_PIPELINE_STOP 0x101 | |
67 | ||
fcff0f35 PD |
68 | #define BP_SPANB(indblkshift, level) \ |
69 | (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) | |
70 | #define COMPARE_META_LEVEL 0x80000000ul | |
55d85d5a GW |
71 | /* |
72 | * The following actions directly effect the spa's sync-to-convergence logic. | |
73 | * The values below define the sync pass when we start performing the action. | |
74 | * Care should be taken when changing these values as they directly impact | |
75 | * spa_sync() performance. Tuning these values may introduce subtle performance | |
76 | * pathologies and should only be done in the context of performance analysis. | |
77 | * These tunables will eventually be removed and replaced with #defines once | |
78 | * enough analysis has been done to determine optimal values. | |
79 | * | |
80 | * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that | |
81 | * regular blocks are not deferred. | |
82 | */ | |
83 | int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ | |
84 | int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */ | |
85 | int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ | |
86 | ||
34dc7c2f | 87 | /* |
b128c09f BB |
88 | * An allocating zio is one that either currently has the DVA allocate |
89 | * stage set or will have it later in its lifetime. | |
34dc7c2f | 90 | */ |
428870ff BB |
91 | #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) |
92 | ||
c409e464 | 93 | int zio_requeue_io_start_cut_in_line = 1; |
428870ff BB |
94 | |
95 | #ifdef ZFS_DEBUG | |
96 | int zio_buf_debug_limit = 16384; | |
97 | #else | |
98 | int zio_buf_debug_limit = 0; | |
99 | #endif | |
34dc7c2f | 100 | |
da6b4005 NB |
101 | static inline void __zio_execute(zio_t *zio); |
102 | ||
34dc7c2f BB |
103 | void |
104 | zio_init(void) | |
105 | { | |
106 | size_t c; | |
107 | vmem_t *data_alloc_arena = NULL; | |
108 | ||
3941503c BB |
109 | zio_cache = kmem_cache_create("zio_cache", |
110 | sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); | |
d164b209 | 111 | zio_link_cache = kmem_cache_create("zio_link_cache", |
6795a698 | 112 | sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); |
34dc7c2f BB |
113 | |
114 | /* | |
115 | * For small buffers, we want a cache for each multiple of | |
f1512ee6 MA |
116 | * SPA_MINBLOCKSIZE. For larger buffers, we want a cache |
117 | * for each quarter-power of 2. | |
34dc7c2f BB |
118 | */ |
119 | for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { | |
120 | size_t size = (c + 1) << SPA_MINBLOCKSHIFT; | |
121 | size_t p2 = size; | |
122 | size_t align = 0; | |
6442f3cf | 123 | size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0; |
34dc7c2f | 124 | |
f1512ee6 MA |
125 | #ifdef _ILP32 |
126 | /* | |
127 | * Cache size limited to 1M on 32-bit platforms until ARC | |
128 | * buffers no longer require virtual address space. | |
129 | */ | |
130 | if (size > zfs_max_recordsize) | |
131 | break; | |
132 | #endif | |
133 | ||
134 | while (!ISP2(p2)) | |
34dc7c2f BB |
135 | p2 &= p2 - 1; |
136 | ||
498877ba MA |
137 | #ifndef _KERNEL |
138 | /* | |
139 | * If we are using watchpoints, put each buffer on its own page, | |
140 | * to eliminate the performance overhead of trapping to the | |
141 | * kernel when modifying a non-watched buffer that shares the | |
142 | * page with a watched buffer. | |
143 | */ | |
144 | if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) | |
145 | continue; | |
fcf64f45 BB |
146 | /* |
147 | * Here's the problem - on 4K native devices in userland on | |
148 | * Linux using O_DIRECT, buffers must be 4K aligned or I/O | |
149 | * will fail with EINVAL, causing zdb (and others) to coredump. | |
150 | * Since userland probably doesn't need optimized buffer caches, | |
151 | * we just force 4K alignment on everything. | |
152 | */ | |
153 | align = 8 * SPA_MINBLOCKSIZE; | |
154 | #else | |
34dc7c2f BB |
155 | if (size <= 4 * SPA_MINBLOCKSIZE) { |
156 | align = SPA_MINBLOCKSIZE; | |
498877ba | 157 | } else if (IS_P2ALIGNED(size, p2 >> 2)) { |
f1512ee6 | 158 | align = MIN(p2 >> 2, PAGESIZE); |
34dc7c2f | 159 | } |
fcf64f45 | 160 | #endif |
34dc7c2f BB |
161 | |
162 | if (align != 0) { | |
163 | char name[36]; | |
164 | (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); | |
165 | zio_buf_cache[c] = kmem_cache_create(name, size, | |
6442f3cf | 166 | align, NULL, NULL, NULL, NULL, NULL, cflags); |
34dc7c2f BB |
167 | |
168 | (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); | |
169 | zio_data_buf_cache[c] = kmem_cache_create(name, size, | |
ae6ba3db | 170 | align, NULL, NULL, NULL, NULL, |
6442f3cf | 171 | data_alloc_arena, cflags); |
34dc7c2f BB |
172 | } |
173 | } | |
174 | ||
175 | while (--c != 0) { | |
176 | ASSERT(zio_buf_cache[c] != NULL); | |
177 | if (zio_buf_cache[c - 1] == NULL) | |
178 | zio_buf_cache[c - 1] = zio_buf_cache[c]; | |
179 | ||
180 | ASSERT(zio_data_buf_cache[c] != NULL); | |
181 | if (zio_data_buf_cache[c - 1] == NULL) | |
182 | zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; | |
183 | } | |
184 | ||
34dc7c2f | 185 | zio_inject_init(); |
9759c60f ED |
186 | |
187 | lz4_init(); | |
34dc7c2f BB |
188 | } |
189 | ||
190 | void | |
191 | zio_fini(void) | |
192 | { | |
193 | size_t c; | |
194 | kmem_cache_t *last_cache = NULL; | |
195 | kmem_cache_t *last_data_cache = NULL; | |
196 | ||
197 | for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { | |
f1512ee6 MA |
198 | #ifdef _ILP32 |
199 | /* | |
200 | * Cache size limited to 1M on 32-bit platforms until ARC | |
201 | * buffers no longer require virtual address space. | |
202 | */ | |
203 | if (((c + 1) << SPA_MINBLOCKSHIFT) > zfs_max_recordsize) | |
204 | break; | |
205 | #endif | |
34dc7c2f BB |
206 | if (zio_buf_cache[c] != last_cache) { |
207 | last_cache = zio_buf_cache[c]; | |
208 | kmem_cache_destroy(zio_buf_cache[c]); | |
209 | } | |
210 | zio_buf_cache[c] = NULL; | |
211 | ||
212 | if (zio_data_buf_cache[c] != last_data_cache) { | |
213 | last_data_cache = zio_data_buf_cache[c]; | |
214 | kmem_cache_destroy(zio_data_buf_cache[c]); | |
215 | } | |
216 | zio_data_buf_cache[c] = NULL; | |
217 | } | |
218 | ||
d164b209 | 219 | kmem_cache_destroy(zio_link_cache); |
34dc7c2f BB |
220 | kmem_cache_destroy(zio_cache); |
221 | ||
222 | zio_inject_fini(); | |
9759c60f ED |
223 | |
224 | lz4_fini(); | |
34dc7c2f BB |
225 | } |
226 | ||
227 | /* | |
228 | * ========================================================================== | |
229 | * Allocate and free I/O buffers | |
230 | * ========================================================================== | |
231 | */ | |
232 | ||
233 | /* | |
234 | * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a | |
235 | * crashdump if the kernel panics, so use it judiciously. Obviously, it's | |
236 | * useful to inspect ZFS metadata, but if possible, we should avoid keeping | |
237 | * excess / transient data in-core during a crashdump. | |
238 | */ | |
239 | void * | |
240 | zio_buf_alloc(size_t size) | |
241 | { | |
242 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
243 | ||
63e3a861 | 244 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f | 245 | |
efcd79a8 | 246 | return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); |
34dc7c2f BB |
247 | } |
248 | ||
249 | /* | |
250 | * Use zio_data_buf_alloc to allocate data. The data will not appear in a | |
251 | * crashdump if the kernel panics. This exists so that we will limit the amount | |
252 | * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount | |
253 | * of kernel heap dumped to disk when the kernel panics) | |
254 | */ | |
255 | void * | |
256 | zio_data_buf_alloc(size_t size) | |
257 | { | |
258 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
259 | ||
63e3a861 | 260 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f | 261 | |
efcd79a8 | 262 | return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); |
34dc7c2f BB |
263 | } |
264 | ||
6fe53787 BB |
265 | /* |
266 | * Use zio_buf_alloc_flags when specific allocation flags are needed. e.g. | |
267 | * passing KM_NOSLEEP when it is acceptable for an allocation to fail. | |
268 | */ | |
269 | void * | |
270 | zio_buf_alloc_flags(size_t size, int flags) | |
271 | { | |
272 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
273 | ||
274 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); | |
275 | ||
276 | return (kmem_cache_alloc(zio_buf_cache[c], flags)); | |
277 | } | |
278 | ||
34dc7c2f BB |
279 | void |
280 | zio_buf_free(void *buf, size_t size) | |
281 | { | |
282 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
283 | ||
63e3a861 | 284 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f BB |
285 | |
286 | kmem_cache_free(zio_buf_cache[c], buf); | |
287 | } | |
288 | ||
289 | void | |
290 | zio_data_buf_free(void *buf, size_t size) | |
291 | { | |
292 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
293 | ||
63e3a861 | 294 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f BB |
295 | |
296 | kmem_cache_free(zio_data_buf_cache[c], buf); | |
297 | } | |
298 | ||
299 | /* | |
300 | * ========================================================================== | |
301 | * Push and pop I/O transform buffers | |
302 | * ========================================================================== | |
303 | */ | |
d3c2ae1c | 304 | void |
b128c09f BB |
305 | zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize, |
306 | zio_transform_func_t *transform) | |
34dc7c2f | 307 | { |
79c76d5b | 308 | zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); |
34dc7c2f | 309 | |
b128c09f BB |
310 | zt->zt_orig_data = zio->io_data; |
311 | zt->zt_orig_size = zio->io_size; | |
34dc7c2f | 312 | zt->zt_bufsize = bufsize; |
b128c09f | 313 | zt->zt_transform = transform; |
34dc7c2f BB |
314 | |
315 | zt->zt_next = zio->io_transform_stack; | |
316 | zio->io_transform_stack = zt; | |
317 | ||
318 | zio->io_data = data; | |
319 | zio->io_size = size; | |
320 | } | |
321 | ||
d3c2ae1c | 322 | void |
b128c09f | 323 | zio_pop_transforms(zio_t *zio) |
34dc7c2f | 324 | { |
b128c09f BB |
325 | zio_transform_t *zt; |
326 | ||
327 | while ((zt = zio->io_transform_stack) != NULL) { | |
328 | if (zt->zt_transform != NULL) | |
329 | zt->zt_transform(zio, | |
330 | zt->zt_orig_data, zt->zt_orig_size); | |
34dc7c2f | 331 | |
428870ff BB |
332 | if (zt->zt_bufsize != 0) |
333 | zio_buf_free(zio->io_data, zt->zt_bufsize); | |
34dc7c2f | 334 | |
b128c09f BB |
335 | zio->io_data = zt->zt_orig_data; |
336 | zio->io_size = zt->zt_orig_size; | |
337 | zio->io_transform_stack = zt->zt_next; | |
34dc7c2f | 338 | |
b128c09f | 339 | kmem_free(zt, sizeof (zio_transform_t)); |
34dc7c2f BB |
340 | } |
341 | } | |
342 | ||
b128c09f BB |
343 | /* |
344 | * ========================================================================== | |
345 | * I/O transform callbacks for subblocks and decompression | |
346 | * ========================================================================== | |
347 | */ | |
348 | static void | |
349 | zio_subblock(zio_t *zio, void *data, uint64_t size) | |
350 | { | |
351 | ASSERT(zio->io_size > size); | |
352 | ||
353 | if (zio->io_type == ZIO_TYPE_READ) | |
354 | bcopy(zio->io_data, data, size); | |
355 | } | |
356 | ||
357 | static void | |
358 | zio_decompress(zio_t *zio, void *data, uint64_t size) | |
359 | { | |
360 | if (zio->io_error == 0 && | |
361 | zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), | |
428870ff | 362 | zio->io_data, data, zio->io_size, size) != 0) |
2e528b49 | 363 | zio->io_error = SET_ERROR(EIO); |
b128c09f BB |
364 | } |
365 | ||
366 | /* | |
367 | * ========================================================================== | |
368 | * I/O parent/child relationships and pipeline interlocks | |
369 | * ========================================================================== | |
370 | */ | |
d164b209 BB |
371 | /* |
372 | * NOTE - Callers to zio_walk_parents() and zio_walk_children must | |
373 | * continue calling these functions until they return NULL. | |
374 | * Otherwise, the next caller will pick up the list walk in | |
375 | * some indeterminate state. (Otherwise every caller would | |
376 | * have to pass in a cookie to keep the state represented by | |
377 | * io_walk_link, which gets annoying.) | |
378 | */ | |
379 | zio_t * | |
380 | zio_walk_parents(zio_t *cio) | |
381 | { | |
382 | zio_link_t *zl = cio->io_walk_link; | |
383 | list_t *pl = &cio->io_parent_list; | |
b128c09f | 384 | |
d164b209 BB |
385 | zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl); |
386 | cio->io_walk_link = zl; | |
387 | ||
388 | if (zl == NULL) | |
389 | return (NULL); | |
390 | ||
391 | ASSERT(zl->zl_child == cio); | |
392 | return (zl->zl_parent); | |
393 | } | |
394 | ||
395 | zio_t * | |
396 | zio_walk_children(zio_t *pio) | |
397 | { | |
398 | zio_link_t *zl = pio->io_walk_link; | |
399 | list_t *cl = &pio->io_child_list; | |
400 | ||
401 | zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl); | |
402 | pio->io_walk_link = zl; | |
403 | ||
404 | if (zl == NULL) | |
405 | return (NULL); | |
406 | ||
407 | ASSERT(zl->zl_parent == pio); | |
408 | return (zl->zl_child); | |
409 | } | |
410 | ||
411 | zio_t * | |
412 | zio_unique_parent(zio_t *cio) | |
413 | { | |
414 | zio_t *pio = zio_walk_parents(cio); | |
415 | ||
416 | VERIFY(zio_walk_parents(cio) == NULL); | |
417 | return (pio); | |
418 | } | |
419 | ||
420 | void | |
421 | zio_add_child(zio_t *pio, zio_t *cio) | |
b128c09f | 422 | { |
79c76d5b | 423 | zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); |
d6320ddb | 424 | int w; |
d164b209 BB |
425 | |
426 | /* | |
427 | * Logical I/Os can have logical, gang, or vdev children. | |
428 | * Gang I/Os can have gang or vdev children. | |
429 | * Vdev I/Os can only have vdev children. | |
430 | * The following ASSERT captures all of these constraints. | |
431 | */ | |
432 | ASSERT(cio->io_child_type <= pio->io_child_type); | |
433 | ||
434 | zl->zl_parent = pio; | |
435 | zl->zl_child = cio; | |
436 | ||
437 | mutex_enter(&cio->io_lock); | |
b128c09f | 438 | mutex_enter(&pio->io_lock); |
d164b209 BB |
439 | |
440 | ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); | |
441 | ||
d6320ddb | 442 | for (w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 BB |
443 | pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; |
444 | ||
445 | list_insert_head(&pio->io_child_list, zl); | |
446 | list_insert_head(&cio->io_parent_list, zl); | |
447 | ||
428870ff BB |
448 | pio->io_child_count++; |
449 | cio->io_parent_count++; | |
450 | ||
b128c09f | 451 | mutex_exit(&pio->io_lock); |
d164b209 | 452 | mutex_exit(&cio->io_lock); |
b128c09f BB |
453 | } |
454 | ||
34dc7c2f | 455 | static void |
d164b209 | 456 | zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) |
b128c09f | 457 | { |
d164b209 BB |
458 | ASSERT(zl->zl_parent == pio); |
459 | ASSERT(zl->zl_child == cio); | |
b128c09f | 460 | |
d164b209 | 461 | mutex_enter(&cio->io_lock); |
b128c09f | 462 | mutex_enter(&pio->io_lock); |
d164b209 BB |
463 | |
464 | list_remove(&pio->io_child_list, zl); | |
465 | list_remove(&cio->io_parent_list, zl); | |
466 | ||
428870ff BB |
467 | pio->io_child_count--; |
468 | cio->io_parent_count--; | |
469 | ||
b128c09f | 470 | mutex_exit(&pio->io_lock); |
d164b209 BB |
471 | mutex_exit(&cio->io_lock); |
472 | ||
473 | kmem_cache_free(zio_link_cache, zl); | |
b128c09f BB |
474 | } |
475 | ||
476 | static boolean_t | |
477 | zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) | |
34dc7c2f | 478 | { |
b128c09f BB |
479 | uint64_t *countp = &zio->io_children[child][wait]; |
480 | boolean_t waiting = B_FALSE; | |
481 | ||
482 | mutex_enter(&zio->io_lock); | |
483 | ASSERT(zio->io_stall == NULL); | |
484 | if (*countp != 0) { | |
428870ff | 485 | zio->io_stage >>= 1; |
b128c09f BB |
486 | zio->io_stall = countp; |
487 | waiting = B_TRUE; | |
488 | } | |
489 | mutex_exit(&zio->io_lock); | |
490 | ||
491 | return (waiting); | |
492 | } | |
34dc7c2f | 493 | |
bf701a83 BB |
494 | __attribute__((always_inline)) |
495 | static inline void | |
b128c09f BB |
496 | zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) |
497 | { | |
498 | uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; | |
499 | int *errorp = &pio->io_child_error[zio->io_child_type]; | |
34dc7c2f | 500 | |
b128c09f BB |
501 | mutex_enter(&pio->io_lock); |
502 | if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) | |
503 | *errorp = zio_worst_error(*errorp, zio->io_error); | |
504 | pio->io_reexecute |= zio->io_reexecute; | |
505 | ASSERT3U(*countp, >, 0); | |
e8b96c60 MA |
506 | |
507 | (*countp)--; | |
508 | ||
509 | if (*countp == 0 && pio->io_stall == countp) { | |
b128c09f BB |
510 | pio->io_stall = NULL; |
511 | mutex_exit(&pio->io_lock); | |
da6b4005 | 512 | __zio_execute(pio); |
b128c09f BB |
513 | } else { |
514 | mutex_exit(&pio->io_lock); | |
34dc7c2f BB |
515 | } |
516 | } | |
517 | ||
b128c09f BB |
518 | static void |
519 | zio_inherit_child_errors(zio_t *zio, enum zio_child c) | |
520 | { | |
521 | if (zio->io_child_error[c] != 0 && zio->io_error == 0) | |
522 | zio->io_error = zio->io_child_error[c]; | |
523 | } | |
524 | ||
34dc7c2f BB |
525 | /* |
526 | * ========================================================================== | |
b128c09f | 527 | * Create the various types of I/O (read, write, free, etc) |
34dc7c2f BB |
528 | * ========================================================================== |
529 | */ | |
530 | static zio_t * | |
428870ff | 531 | zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
2aa34383 DK |
532 | void *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done, |
533 | void *private, zio_type_t type, zio_priority_t priority, | |
534 | enum zio_flag flags, vdev_t *vd, uint64_t offset, | |
535 | const zbookmark_phys_t *zb, enum zio_stage stage, | |
536 | enum zio_stage pipeline) | |
34dc7c2f BB |
537 | { |
538 | zio_t *zio; | |
539 | ||
2aa34383 DK |
540 | ASSERT3U(psize, <=, SPA_MAXBLOCKSIZE); |
541 | ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); | |
b128c09f BB |
542 | ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); |
543 | ||
544 | ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); | |
545 | ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); | |
546 | ASSERT(vd || stage == ZIO_STAGE_OPEN); | |
34dc7c2f | 547 | |
2aa34383 DK |
548 | IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0); |
549 | ||
79c76d5b | 550 | zio = kmem_cache_alloc(zio_cache, KM_SLEEP); |
3941503c BB |
551 | bzero(zio, sizeof (zio_t)); |
552 | ||
448d7aaa | 553 | mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL); |
3941503c BB |
554 | cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); |
555 | ||
556 | list_create(&zio->io_parent_list, sizeof (zio_link_t), | |
557 | offsetof(zio_link_t, zl_parent_node)); | |
558 | list_create(&zio->io_child_list, sizeof (zio_link_t), | |
559 | offsetof(zio_link_t, zl_child_node)); | |
d164b209 | 560 | |
b128c09f BB |
561 | if (vd != NULL) |
562 | zio->io_child_type = ZIO_CHILD_VDEV; | |
563 | else if (flags & ZIO_FLAG_GANG_CHILD) | |
564 | zio->io_child_type = ZIO_CHILD_GANG; | |
428870ff BB |
565 | else if (flags & ZIO_FLAG_DDT_CHILD) |
566 | zio->io_child_type = ZIO_CHILD_DDT; | |
b128c09f BB |
567 | else |
568 | zio->io_child_type = ZIO_CHILD_LOGICAL; | |
569 | ||
34dc7c2f | 570 | if (bp != NULL) { |
428870ff | 571 | zio->io_bp = (blkptr_t *)bp; |
34dc7c2f BB |
572 | zio->io_bp_copy = *bp; |
573 | zio->io_bp_orig = *bp; | |
428870ff BB |
574 | if (type != ZIO_TYPE_WRITE || |
575 | zio->io_child_type == ZIO_CHILD_DDT) | |
b128c09f | 576 | zio->io_bp = &zio->io_bp_copy; /* so caller can free */ |
9babb374 | 577 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) |
b128c09f | 578 | zio->io_logical = zio; |
9babb374 BB |
579 | if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) |
580 | pipeline |= ZIO_GANG_STAGES; | |
34dc7c2f | 581 | } |
b128c09f BB |
582 | |
583 | zio->io_spa = spa; | |
584 | zio->io_txg = txg; | |
34dc7c2f BB |
585 | zio->io_done = done; |
586 | zio->io_private = private; | |
587 | zio->io_type = type; | |
588 | zio->io_priority = priority; | |
b128c09f BB |
589 | zio->io_vd = vd; |
590 | zio->io_offset = offset; | |
428870ff | 591 | zio->io_orig_data = zio->io_data = data; |
2aa34383 DK |
592 | zio->io_orig_size = zio->io_size = psize; |
593 | zio->io_lsize = lsize; | |
b128c09f BB |
594 | zio->io_orig_flags = zio->io_flags = flags; |
595 | zio->io_orig_stage = zio->io_stage = stage; | |
596 | zio->io_orig_pipeline = zio->io_pipeline = pipeline; | |
34dc7c2f | 597 | |
d164b209 BB |
598 | zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); |
599 | zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); | |
600 | ||
b128c09f BB |
601 | if (zb != NULL) |
602 | zio->io_bookmark = *zb; | |
603 | ||
604 | if (pio != NULL) { | |
b128c09f | 605 | if (zio->io_logical == NULL) |
34dc7c2f | 606 | zio->io_logical = pio->io_logical; |
9babb374 BB |
607 | if (zio->io_child_type == ZIO_CHILD_GANG) |
608 | zio->io_gang_leader = pio->io_gang_leader; | |
b128c09f | 609 | zio_add_child(pio, zio); |
34dc7c2f BB |
610 | } |
611 | ||
a38718a6 GA |
612 | taskq_init_ent(&zio->io_tqent); |
613 | ||
34dc7c2f BB |
614 | return (zio); |
615 | } | |
616 | ||
617 | static void | |
b128c09f | 618 | zio_destroy(zio_t *zio) |
34dc7c2f | 619 | { |
3941503c BB |
620 | list_destroy(&zio->io_parent_list); |
621 | list_destroy(&zio->io_child_list); | |
622 | mutex_destroy(&zio->io_lock); | |
623 | cv_destroy(&zio->io_cv); | |
b128c09f | 624 | kmem_cache_free(zio_cache, zio); |
34dc7c2f BB |
625 | } |
626 | ||
627 | zio_t * | |
d164b209 | 628 | zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, |
428870ff | 629 | void *private, enum zio_flag flags) |
34dc7c2f BB |
630 | { |
631 | zio_t *zio; | |
632 | ||
2aa34383 | 633 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
d164b209 | 634 | ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
b128c09f | 635 | ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); |
34dc7c2f BB |
636 | |
637 | return (zio); | |
638 | } | |
639 | ||
640 | zio_t * | |
428870ff | 641 | zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) |
34dc7c2f | 642 | { |
d164b209 | 643 | return (zio_null(NULL, spa, NULL, done, private, flags)); |
34dc7c2f BB |
644 | } |
645 | ||
63e3a861 MA |
646 | void |
647 | zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp) | |
648 | { | |
649 | int i; | |
650 | ||
651 | if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { | |
652 | zfs_panic_recover("blkptr at %p has invalid TYPE %llu", | |
653 | bp, (longlong_t)BP_GET_TYPE(bp)); | |
654 | } | |
655 | if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS || | |
656 | BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) { | |
657 | zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu", | |
658 | bp, (longlong_t)BP_GET_CHECKSUM(bp)); | |
659 | } | |
660 | if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS || | |
661 | BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) { | |
662 | zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu", | |
663 | bp, (longlong_t)BP_GET_COMPRESS(bp)); | |
664 | } | |
665 | if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
666 | zfs_panic_recover("blkptr at %p has invalid LSIZE %llu", | |
667 | bp, (longlong_t)BP_GET_LSIZE(bp)); | |
668 | } | |
669 | if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
670 | zfs_panic_recover("blkptr at %p has invalid PSIZE %llu", | |
671 | bp, (longlong_t)BP_GET_PSIZE(bp)); | |
672 | } | |
673 | ||
674 | if (BP_IS_EMBEDDED(bp)) { | |
675 | if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) { | |
676 | zfs_panic_recover("blkptr at %p has invalid ETYPE %llu", | |
677 | bp, (longlong_t)BPE_GET_ETYPE(bp)); | |
678 | } | |
679 | } | |
680 | ||
681 | /* | |
682 | * Pool-specific checks. | |
683 | * | |
684 | * Note: it would be nice to verify that the blk_birth and | |
685 | * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() | |
686 | * allows the birth time of log blocks (and dmu_sync()-ed blocks | |
687 | * that are in the log) to be arbitrarily large. | |
688 | */ | |
689 | for (i = 0; i < BP_GET_NDVAS(bp); i++) { | |
690 | uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]); | |
691 | vdev_t *vd; | |
692 | uint64_t offset, asize; | |
693 | if (vdevid >= spa->spa_root_vdev->vdev_children) { | |
694 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
695 | "VDEV %llu", | |
696 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 697 | continue; |
63e3a861 MA |
698 | } |
699 | vd = spa->spa_root_vdev->vdev_child[vdevid]; | |
700 | if (vd == NULL) { | |
701 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
702 | "VDEV %llu", | |
703 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 704 | continue; |
63e3a861 MA |
705 | } |
706 | if (vd->vdev_ops == &vdev_hole_ops) { | |
707 | zfs_panic_recover("blkptr at %p DVA %u has hole " | |
708 | "VDEV %llu", | |
709 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 710 | continue; |
63e3a861 MA |
711 | } |
712 | if (vd->vdev_ops == &vdev_missing_ops) { | |
713 | /* | |
714 | * "missing" vdevs are valid during import, but we | |
715 | * don't have their detailed info (e.g. asize), so | |
716 | * we can't perform any more checks on them. | |
717 | */ | |
718 | continue; | |
719 | } | |
720 | offset = DVA_GET_OFFSET(&bp->blk_dva[i]); | |
721 | asize = DVA_GET_ASIZE(&bp->blk_dva[i]); | |
722 | if (BP_IS_GANG(bp)) | |
723 | asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
724 | if (offset + asize > vd->vdev_asize) { | |
725 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
726 | "OFFSET %llu", | |
727 | bp, i, (longlong_t)offset); | |
728 | } | |
729 | } | |
730 | } | |
731 | ||
34dc7c2f | 732 | zio_t * |
b128c09f BB |
733 | zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, |
734 | void *data, uint64_t size, zio_done_func_t *done, void *private, | |
5dbd68a3 | 735 | zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) |
34dc7c2f BB |
736 | { |
737 | zio_t *zio; | |
738 | ||
63e3a861 MA |
739 | zfs_blkptr_verify(spa, bp); |
740 | ||
428870ff | 741 | zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, |
2aa34383 | 742 | data, size, size, done, private, |
b128c09f | 743 | ZIO_TYPE_READ, priority, flags, NULL, 0, zb, |
428870ff BB |
744 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
745 | ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); | |
34dc7c2f | 746 | |
b128c09f BB |
747 | return (zio); |
748 | } | |
34dc7c2f | 749 | |
34dc7c2f | 750 | zio_t * |
b128c09f | 751 | zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, |
2aa34383 | 752 | void *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp, |
bc77ba73 PD |
753 | zio_done_func_t *ready, zio_done_func_t *children_ready, |
754 | zio_done_func_t *physdone, zio_done_func_t *done, | |
755 | void *private, zio_priority_t priority, enum zio_flag flags, | |
756 | const zbookmark_phys_t *zb) | |
34dc7c2f BB |
757 | { |
758 | zio_t *zio; | |
759 | ||
b128c09f BB |
760 | ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && |
761 | zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && | |
762 | zp->zp_compress >= ZIO_COMPRESS_OFF && | |
763 | zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && | |
9ae529ec | 764 | DMU_OT_IS_VALID(zp->zp_type) && |
b128c09f | 765 | zp->zp_level < 32 && |
428870ff | 766 | zp->zp_copies > 0 && |
03c6040b | 767 | zp->zp_copies <= spa_max_replication(spa)); |
34dc7c2f | 768 | |
2aa34383 | 769 | zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private, |
b128c09f | 770 | ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, |
428870ff BB |
771 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
772 | ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); | |
34dc7c2f BB |
773 | |
774 | zio->io_ready = ready; | |
bc77ba73 | 775 | zio->io_children_ready = children_ready; |
e8b96c60 | 776 | zio->io_physdone = physdone; |
b128c09f | 777 | zio->io_prop = *zp; |
34dc7c2f | 778 | |
9b67f605 MA |
779 | /* |
780 | * Data can be NULL if we are going to call zio_write_override() to | |
781 | * provide the already-allocated BP. But we may need the data to | |
782 | * verify a dedup hit (if requested). In this case, don't try to | |
783 | * dedup (just take the already-allocated BP verbatim). | |
784 | */ | |
785 | if (data == NULL && zio->io_prop.zp_dedup_verify) { | |
786 | zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; | |
787 | } | |
788 | ||
34dc7c2f BB |
789 | return (zio); |
790 | } | |
791 | ||
792 | zio_t * | |
b128c09f | 793 | zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data, |
e8b96c60 | 794 | uint64_t size, zio_done_func_t *done, void *private, |
5dbd68a3 | 795 | zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) |
34dc7c2f BB |
796 | { |
797 | zio_t *zio; | |
798 | ||
2aa34383 | 799 | zio = zio_create(pio, spa, txg, bp, data, size, size, done, private, |
b128c09f BB |
800 | ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, |
801 | ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); | |
34dc7c2f BB |
802 | |
803 | return (zio); | |
804 | } | |
805 | ||
428870ff | 806 | void |
03c6040b | 807 | zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) |
428870ff BB |
808 | { |
809 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
810 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
811 | ASSERT(zio->io_stage == ZIO_STAGE_OPEN); | |
812 | ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); | |
813 | ||
03c6040b GW |
814 | /* |
815 | * We must reset the io_prop to match the values that existed | |
816 | * when the bp was first written by dmu_sync() keeping in mind | |
817 | * that nopwrite and dedup are mutually exclusive. | |
818 | */ | |
819 | zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; | |
820 | zio->io_prop.zp_nopwrite = nopwrite; | |
428870ff BB |
821 | zio->io_prop.zp_copies = copies; |
822 | zio->io_bp_override = bp; | |
823 | } | |
824 | ||
825 | void | |
826 | zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) | |
827 | { | |
9b67f605 MA |
828 | |
829 | /* | |
830 | * The check for EMBEDDED is a performance optimization. We | |
831 | * process the free here (by ignoring it) rather than | |
832 | * putting it on the list and then processing it in zio_free_sync(). | |
833 | */ | |
834 | if (BP_IS_EMBEDDED(bp)) | |
835 | return; | |
13fe0198 | 836 | metaslab_check_free(spa, bp); |
2883cad5 MA |
837 | |
838 | /* | |
839 | * Frees that are for the currently-syncing txg, are not going to be | |
840 | * deferred, and which will not need to do a read (i.e. not GANG or | |
841 | * DEDUP), can be processed immediately. Otherwise, put them on the | |
842 | * in-memory list for later processing. | |
843 | */ | |
844 | if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || | |
845 | txg != spa->spa_syncing_txg || | |
846 | spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { | |
847 | bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); | |
848 | } else { | |
849 | VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0))); | |
850 | } | |
428870ff BB |
851 | } |
852 | ||
34dc7c2f | 853 | zio_t * |
428870ff BB |
854 | zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
855 | enum zio_flag flags) | |
34dc7c2f BB |
856 | { |
857 | zio_t *zio; | |
2883cad5 | 858 | enum zio_stage stage = ZIO_FREE_PIPELINE; |
34dc7c2f | 859 | |
428870ff BB |
860 | ASSERT(!BP_IS_HOLE(bp)); |
861 | ASSERT(spa_syncing_txg(spa) == txg); | |
55d85d5a | 862 | ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); |
34dc7c2f | 863 | |
9b67f605 MA |
864 | if (BP_IS_EMBEDDED(bp)) |
865 | return (zio_null(pio, spa, NULL, NULL, NULL, 0)); | |
866 | ||
13fe0198 | 867 | metaslab_check_free(spa, bp); |
8c841793 | 868 | arc_freed(spa, bp); |
13fe0198 | 869 | |
2883cad5 MA |
870 | /* |
871 | * GANG and DEDUP blocks can induce a read (for the gang block header, | |
872 | * or the DDT), so issue them asynchronously so that this thread is | |
873 | * not tied up. | |
874 | */ | |
875 | if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) | |
876 | stage |= ZIO_STAGE_ISSUE_ASYNC; | |
877 | ||
b128c09f | 878 | zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
2aa34383 DK |
879 | BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, |
880 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage); | |
2883cad5 | 881 | |
34dc7c2f BB |
882 | return (zio); |
883 | } | |
884 | ||
885 | zio_t * | |
428870ff BB |
886 | zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
887 | zio_done_func_t *done, void *private, enum zio_flag flags) | |
34dc7c2f BB |
888 | { |
889 | zio_t *zio; | |
890 | ||
9b67f605 MA |
891 | dprintf_bp(bp, "claiming in txg %llu", txg); |
892 | ||
893 | if (BP_IS_EMBEDDED(bp)) | |
894 | return (zio_null(pio, spa, NULL, NULL, NULL, 0)); | |
895 | ||
34dc7c2f BB |
896 | /* |
897 | * A claim is an allocation of a specific block. Claims are needed | |
898 | * to support immediate writes in the intent log. The issue is that | |
899 | * immediate writes contain committed data, but in a txg that was | |
900 | * *not* committed. Upon opening the pool after an unclean shutdown, | |
901 | * the intent log claims all blocks that contain immediate write data | |
902 | * so that the SPA knows they're in use. | |
903 | * | |
904 | * All claims *must* be resolved in the first txg -- before the SPA | |
905 | * starts allocating blocks -- so that nothing is allocated twice. | |
428870ff | 906 | * If txg == 0 we just verify that the block is claimable. |
34dc7c2f BB |
907 | */ |
908 | ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); | |
428870ff BB |
909 | ASSERT(txg == spa_first_txg(spa) || txg == 0); |
910 | ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ | |
34dc7c2f | 911 | |
b128c09f | 912 | zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
2aa34383 DK |
913 | BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, |
914 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); | |
34dc7c2f BB |
915 | |
916 | return (zio); | |
917 | } | |
918 | ||
919 | zio_t * | |
920 | zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, | |
e8b96c60 | 921 | zio_done_func_t *done, void *private, enum zio_flag flags) |
34dc7c2f BB |
922 | { |
923 | zio_t *zio; | |
924 | int c; | |
925 | ||
926 | if (vd->vdev_children == 0) { | |
2aa34383 | 927 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
e8b96c60 | 928 | ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
34dc7c2f BB |
929 | ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); |
930 | ||
34dc7c2f BB |
931 | zio->io_cmd = cmd; |
932 | } else { | |
d164b209 | 933 | zio = zio_null(pio, spa, NULL, NULL, NULL, flags); |
34dc7c2f BB |
934 | |
935 | for (c = 0; c < vd->vdev_children; c++) | |
936 | zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, | |
e8b96c60 | 937 | done, private, flags)); |
34dc7c2f BB |
938 | } |
939 | ||
940 | return (zio); | |
941 | } | |
942 | ||
34dc7c2f BB |
943 | zio_t * |
944 | zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
945 | void *data, int checksum, zio_done_func_t *done, void *private, | |
e8b96c60 | 946 | zio_priority_t priority, enum zio_flag flags, boolean_t labels) |
34dc7c2f BB |
947 | { |
948 | zio_t *zio; | |
34dc7c2f | 949 | |
b128c09f BB |
950 | ASSERT(vd->vdev_children == 0); |
951 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
952 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
953 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 954 | |
2aa34383 DK |
955 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
956 | private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
957 | offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); | |
34dc7c2f | 958 | |
b128c09f | 959 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f BB |
960 | |
961 | return (zio); | |
962 | } | |
963 | ||
964 | zio_t * | |
965 | zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
966 | void *data, int checksum, zio_done_func_t *done, void *private, | |
e8b96c60 | 967 | zio_priority_t priority, enum zio_flag flags, boolean_t labels) |
34dc7c2f | 968 | { |
34dc7c2f | 969 | zio_t *zio; |
34dc7c2f | 970 | |
b128c09f BB |
971 | ASSERT(vd->vdev_children == 0); |
972 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
973 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
974 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 975 | |
2aa34383 DK |
976 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
977 | private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
978 | offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); | |
34dc7c2f | 979 | |
b128c09f | 980 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f | 981 | |
3c67d83a | 982 | if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { |
34dc7c2f | 983 | /* |
428870ff | 984 | * zec checksums are necessarily destructive -- they modify |
b128c09f | 985 | * the end of the write buffer to hold the verifier/checksum. |
34dc7c2f | 986 | * Therefore, we must make a local copy in case the data is |
b128c09f | 987 | * being written to multiple places in parallel. |
34dc7c2f | 988 | */ |
b128c09f | 989 | void *wbuf = zio_buf_alloc(size); |
34dc7c2f | 990 | bcopy(data, wbuf, size); |
b128c09f | 991 | zio_push_transform(zio, wbuf, size, size, NULL); |
34dc7c2f BB |
992 | } |
993 | ||
994 | return (zio); | |
995 | } | |
996 | ||
997 | /* | |
b128c09f | 998 | * Create a child I/O to do some work for us. |
34dc7c2f BB |
999 | */ |
1000 | zio_t * | |
b128c09f | 1001 | zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, |
e8b96c60 MA |
1002 | void *data, uint64_t size, int type, zio_priority_t priority, |
1003 | enum zio_flag flags, zio_done_func_t *done, void *private) | |
34dc7c2f | 1004 | { |
428870ff | 1005 | enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; |
b128c09f BB |
1006 | zio_t *zio; |
1007 | ||
1008 | ASSERT(vd->vdev_parent == | |
1009 | (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev)); | |
34dc7c2f BB |
1010 | |
1011 | if (type == ZIO_TYPE_READ && bp != NULL) { | |
1012 | /* | |
1013 | * If we have the bp, then the child should perform the | |
1014 | * checksum and the parent need not. This pushes error | |
1015 | * detection as close to the leaves as possible and | |
1016 | * eliminates redundant checksums in the interior nodes. | |
1017 | */ | |
428870ff BB |
1018 | pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; |
1019 | pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; | |
34dc7c2f BB |
1020 | } |
1021 | ||
b128c09f BB |
1022 | if (vd->vdev_children == 0) |
1023 | offset += VDEV_LABEL_START_SIZE; | |
1024 | ||
428870ff BB |
1025 | flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE; |
1026 | ||
1027 | /* | |
1028 | * If we've decided to do a repair, the write is not speculative -- | |
1029 | * even if the original read was. | |
1030 | */ | |
1031 | if (flags & ZIO_FLAG_IO_REPAIR) | |
1032 | flags &= ~ZIO_FLAG_SPECULATIVE; | |
1033 | ||
2aa34383 | 1034 | zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size, |
428870ff BB |
1035 | done, private, type, priority, flags, vd, offset, &pio->io_bookmark, |
1036 | ZIO_STAGE_VDEV_IO_START >> 1, pipeline); | |
34dc7c2f | 1037 | |
e8b96c60 MA |
1038 | zio->io_physdone = pio->io_physdone; |
1039 | if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) | |
1040 | zio->io_logical->io_phys_children++; | |
1041 | ||
b128c09f | 1042 | return (zio); |
34dc7c2f BB |
1043 | } |
1044 | ||
b128c09f BB |
1045 | zio_t * |
1046 | zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size, | |
e8b96c60 | 1047 | int type, zio_priority_t priority, enum zio_flag flags, |
428870ff | 1048 | zio_done_func_t *done, void *private) |
34dc7c2f | 1049 | { |
b128c09f | 1050 | zio_t *zio; |
34dc7c2f | 1051 | |
b128c09f | 1052 | ASSERT(vd->vdev_ops->vdev_op_leaf); |
34dc7c2f | 1053 | |
b128c09f | 1054 | zio = zio_create(NULL, vd->vdev_spa, 0, NULL, |
2aa34383 | 1055 | data, size, size, done, private, type, priority, |
e8b96c60 | 1056 | flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, |
b128c09f | 1057 | vd, offset, NULL, |
428870ff | 1058 | ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); |
34dc7c2f | 1059 | |
b128c09f | 1060 | return (zio); |
34dc7c2f BB |
1061 | } |
1062 | ||
1063 | void | |
b128c09f | 1064 | zio_flush(zio_t *zio, vdev_t *vd) |
34dc7c2f | 1065 | { |
b128c09f | 1066 | zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, |
e8b96c60 | 1067 | NULL, NULL, |
b128c09f | 1068 | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); |
34dc7c2f BB |
1069 | } |
1070 | ||
428870ff BB |
1071 | void |
1072 | zio_shrink(zio_t *zio, uint64_t size) | |
1073 | { | |
1074 | ASSERT(zio->io_executor == NULL); | |
1075 | ASSERT(zio->io_orig_size == zio->io_size); | |
1076 | ASSERT(size <= zio->io_size); | |
1077 | ||
1078 | /* | |
1079 | * We don't shrink for raidz because of problems with the | |
1080 | * reconstruction when reading back less than the block size. | |
1081 | * Note, BP_IS_RAIDZ() assumes no compression. | |
1082 | */ | |
1083 | ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); | |
2aa34383 DK |
1084 | if (!BP_IS_RAIDZ(zio->io_bp)) { |
1085 | /* we are not doing a raw write */ | |
1086 | ASSERT3U(zio->io_size, ==, zio->io_lsize); | |
1087 | zio->io_orig_size = zio->io_size = zio->io_lsize = size; | |
1088 | } | |
428870ff BB |
1089 | } |
1090 | ||
34dc7c2f BB |
1091 | /* |
1092 | * ========================================================================== | |
b128c09f | 1093 | * Prepare to read and write logical blocks |
34dc7c2f BB |
1094 | * ========================================================================== |
1095 | */ | |
b128c09f | 1096 | |
34dc7c2f | 1097 | static int |
b128c09f | 1098 | zio_read_bp_init(zio_t *zio) |
34dc7c2f | 1099 | { |
b128c09f | 1100 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 1101 | |
fb5f0bc8 | 1102 | if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && |
9babb374 BB |
1103 | zio->io_child_type == ZIO_CHILD_LOGICAL && |
1104 | !(zio->io_flags & ZIO_FLAG_RAW)) { | |
9b67f605 MA |
1105 | uint64_t psize = |
1106 | BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); | |
428870ff | 1107 | void *cbuf = zio_buf_alloc(psize); |
b128c09f | 1108 | |
428870ff | 1109 | zio_push_transform(zio, cbuf, psize, psize, zio_decompress); |
34dc7c2f | 1110 | } |
34dc7c2f | 1111 | |
9b67f605 MA |
1112 | if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { |
1113 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1114 | decode_embedded_bp_compressed(bp, zio->io_data); | |
1115 | } else { | |
1116 | ASSERT(!BP_IS_EMBEDDED(bp)); | |
1117 | } | |
1118 | ||
9ae529ec | 1119 | if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) |
b128c09f BB |
1120 | zio->io_flags |= ZIO_FLAG_DONT_CACHE; |
1121 | ||
428870ff BB |
1122 | if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) |
1123 | zio->io_flags |= ZIO_FLAG_DONT_CACHE; | |
1124 | ||
1125 | if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) | |
1126 | zio->io_pipeline = ZIO_DDT_READ_PIPELINE; | |
1127 | ||
b128c09f | 1128 | return (ZIO_PIPELINE_CONTINUE); |
34dc7c2f BB |
1129 | } |
1130 | ||
b128c09f BB |
1131 | static int |
1132 | zio_write_bp_init(zio_t *zio) | |
34dc7c2f | 1133 | { |
428870ff | 1134 | spa_t *spa = zio->io_spa; |
b128c09f | 1135 | zio_prop_t *zp = &zio->io_prop; |
428870ff | 1136 | enum zio_compress compress = zp->zp_compress; |
34dc7c2f | 1137 | blkptr_t *bp = zio->io_bp; |
2aa34383 DK |
1138 | uint64_t lsize = zio->io_lsize; |
1139 | uint64_t psize = zio->io_size; | |
b128c09f | 1140 | int pass = 1; |
34dc7c2f | 1141 | |
2aa34383 DK |
1142 | EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0); |
1143 | ||
b128c09f BB |
1144 | /* |
1145 | * If our children haven't all reached the ready stage, | |
1146 | * wait for them and then repeat this pipeline stage. | |
1147 | */ | |
1148 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || | |
1149 | zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) | |
1150 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f | 1151 | |
b128c09f BB |
1152 | if (!IO_IS_ALLOCATING(zio)) |
1153 | return (ZIO_PIPELINE_CONTINUE); | |
34dc7c2f | 1154 | |
bc77ba73 PD |
1155 | if (zio->io_children_ready != NULL) { |
1156 | /* | |
1157 | * Now that all our children are ready, run the callback | |
1158 | * associated with this zio in case it wants to modify the | |
1159 | * data to be written. | |
1160 | */ | |
1161 | ASSERT3U(zp->zp_level, >, 0); | |
1162 | zio->io_children_ready(zio); | |
1163 | } | |
1164 | ||
428870ff BB |
1165 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1166 | ||
1167 | if (zio->io_bp_override) { | |
1168 | ASSERT(bp->blk_birth != zio->io_txg); | |
1169 | ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); | |
1170 | ||
1171 | *bp = *zio->io_bp_override; | |
1172 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1173 | ||
9b67f605 MA |
1174 | if (BP_IS_EMBEDDED(bp)) |
1175 | return (ZIO_PIPELINE_CONTINUE); | |
1176 | ||
03c6040b GW |
1177 | /* |
1178 | * If we've been overridden and nopwrite is set then | |
1179 | * set the flag accordingly to indicate that a nopwrite | |
1180 | * has already occurred. | |
1181 | */ | |
1182 | if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { | |
1183 | ASSERT(!zp->zp_dedup); | |
1184 | zio->io_flags |= ZIO_FLAG_NOPWRITE; | |
1185 | return (ZIO_PIPELINE_CONTINUE); | |
1186 | } | |
1187 | ||
1188 | ASSERT(!zp->zp_nopwrite); | |
1189 | ||
428870ff BB |
1190 | if (BP_IS_HOLE(bp) || !zp->zp_dedup) |
1191 | return (ZIO_PIPELINE_CONTINUE); | |
1192 | ||
3c67d83a TH |
1193 | ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & |
1194 | ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); | |
428870ff BB |
1195 | |
1196 | if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) { | |
1197 | BP_SET_DEDUP(bp, 1); | |
1198 | zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; | |
1199 | return (ZIO_PIPELINE_CONTINUE); | |
1200 | } | |
5511754b MA |
1201 | zio->io_bp_override = NULL; |
1202 | BP_ZERO(bp); | |
428870ff | 1203 | } |
34dc7c2f | 1204 | |
b0bc7a84 | 1205 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { |
b128c09f BB |
1206 | /* |
1207 | * We're rewriting an existing block, which means we're | |
1208 | * working on behalf of spa_sync(). For spa_sync() to | |
1209 | * converge, it must eventually be the case that we don't | |
1210 | * have to allocate new blocks. But compression changes | |
1211 | * the blocksize, which forces a reallocate, and makes | |
1212 | * convergence take longer. Therefore, after the first | |
1213 | * few passes, stop compressing to ensure convergence. | |
1214 | */ | |
428870ff BB |
1215 | pass = spa_sync_pass(spa); |
1216 | ||
1217 | ASSERT(zio->io_txg == spa_syncing_txg(spa)); | |
1218 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1219 | ASSERT(!BP_GET_DEDUP(bp)); | |
34dc7c2f | 1220 | |
55d85d5a | 1221 | if (pass >= zfs_sync_pass_dont_compress) |
b128c09f | 1222 | compress = ZIO_COMPRESS_OFF; |
34dc7c2f | 1223 | |
b128c09f | 1224 | /* Make sure someone doesn't change their mind on overwrites */ |
9b67f605 | 1225 | ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), |
428870ff | 1226 | spa_max_replication(spa)) == BP_GET_NDVAS(bp)); |
b128c09f | 1227 | } |
34dc7c2f | 1228 | |
2aa34383 DK |
1229 | /* If it's a compressed write that is not raw, compress the buffer. */ |
1230 | if (compress != ZIO_COMPRESS_OFF && psize == lsize) { | |
428870ff BB |
1231 | void *cbuf = zio_buf_alloc(lsize); |
1232 | psize = zio_compress_data(compress, zio->io_data, cbuf, lsize); | |
1233 | if (psize == 0 || psize == lsize) { | |
b128c09f | 1234 | compress = ZIO_COMPRESS_OFF; |
428870ff | 1235 | zio_buf_free(cbuf, lsize); |
9b67f605 MA |
1236 | } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE && |
1237 | zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && | |
1238 | spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { | |
1239 | encode_embedded_bp_compressed(bp, | |
1240 | cbuf, compress, lsize, psize); | |
1241 | BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); | |
1242 | BP_SET_TYPE(bp, zio->io_prop.zp_type); | |
1243 | BP_SET_LEVEL(bp, zio->io_prop.zp_level); | |
1244 | zio_buf_free(cbuf, lsize); | |
1245 | bp->blk_birth = zio->io_txg; | |
1246 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1247 | ASSERT(spa_feature_is_active(spa, | |
1248 | SPA_FEATURE_EMBEDDED_DATA)); | |
1249 | return (ZIO_PIPELINE_CONTINUE); | |
428870ff | 1250 | } else { |
9b67f605 | 1251 | /* |
c3520e7f MA |
1252 | * Round up compressed size up to the ashift |
1253 | * of the smallest-ashift device, and zero the tail. | |
1254 | * This ensures that the compressed size of the BP | |
1255 | * (and thus compressratio property) are correct, | |
1256 | * in that we charge for the padding used to fill out | |
1257 | * the last sector. | |
9b67f605 | 1258 | */ |
c3520e7f MA |
1259 | size_t rounded; |
1260 | ||
1261 | ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT); | |
1262 | ||
1263 | rounded = (size_t)P2ROUNDUP(psize, | |
1264 | 1ULL << spa->spa_min_ashift); | |
1265 | if (rounded >= lsize) { | |
9b67f605 MA |
1266 | compress = ZIO_COMPRESS_OFF; |
1267 | zio_buf_free(cbuf, lsize); | |
c3520e7f | 1268 | psize = lsize; |
9b67f605 | 1269 | } else { |
c3520e7f MA |
1270 | bzero((char *)cbuf + psize, rounded - psize); |
1271 | psize = rounded; | |
9b67f605 MA |
1272 | zio_push_transform(zio, cbuf, |
1273 | psize, lsize, NULL); | |
1274 | } | |
b128c09f | 1275 | } |
2aa34383 DK |
1276 | } else { |
1277 | ASSERT3U(psize, !=, 0); | |
1278 | ||
b128c09f | 1279 | } |
34dc7c2f | 1280 | |
b128c09f BB |
1281 | /* |
1282 | * The final pass of spa_sync() must be all rewrites, but the first | |
1283 | * few passes offer a trade-off: allocating blocks defers convergence, | |
1284 | * but newly allocated blocks are sequential, so they can be written | |
1285 | * to disk faster. Therefore, we allow the first few passes of | |
1286 | * spa_sync() to allocate new blocks, but force rewrites after that. | |
1287 | * There should only be a handful of blocks after pass 1 in any case. | |
1288 | */ | |
b0bc7a84 MG |
1289 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && |
1290 | BP_GET_PSIZE(bp) == psize && | |
55d85d5a | 1291 | pass >= zfs_sync_pass_rewrite) { |
428870ff | 1292 | enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; |
d6320ddb | 1293 | ASSERT(psize != 0); |
b128c09f BB |
1294 | zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; |
1295 | zio->io_flags |= ZIO_FLAG_IO_REWRITE; | |
1296 | } else { | |
1297 | BP_ZERO(bp); | |
1298 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
1299 | } | |
34dc7c2f | 1300 | |
428870ff | 1301 | if (psize == 0) { |
b0bc7a84 MG |
1302 | if (zio->io_bp_orig.blk_birth != 0 && |
1303 | spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { | |
1304 | BP_SET_LSIZE(bp, lsize); | |
1305 | BP_SET_TYPE(bp, zp->zp_type); | |
1306 | BP_SET_LEVEL(bp, zp->zp_level); | |
1307 | BP_SET_BIRTH(bp, zio->io_txg, 0); | |
1308 | } | |
b128c09f BB |
1309 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
1310 | } else { | |
1311 | ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); | |
1312 | BP_SET_LSIZE(bp, lsize); | |
b0bc7a84 MG |
1313 | BP_SET_TYPE(bp, zp->zp_type); |
1314 | BP_SET_LEVEL(bp, zp->zp_level); | |
428870ff | 1315 | BP_SET_PSIZE(bp, psize); |
b128c09f BB |
1316 | BP_SET_COMPRESS(bp, compress); |
1317 | BP_SET_CHECKSUM(bp, zp->zp_checksum); | |
428870ff | 1318 | BP_SET_DEDUP(bp, zp->zp_dedup); |
b128c09f | 1319 | BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); |
428870ff BB |
1320 | if (zp->zp_dedup) { |
1321 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1322 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
1323 | zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; | |
1324 | } | |
03c6040b GW |
1325 | if (zp->zp_nopwrite) { |
1326 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1327 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
1328 | zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; | |
1329 | } | |
428870ff BB |
1330 | } |
1331 | ||
1332 | return (ZIO_PIPELINE_CONTINUE); | |
1333 | } | |
1334 | ||
1335 | static int | |
1336 | zio_free_bp_init(zio_t *zio) | |
1337 | { | |
1338 | blkptr_t *bp = zio->io_bp; | |
1339 | ||
1340 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
1341 | if (BP_GET_DEDUP(bp)) | |
1342 | zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; | |
b128c09f | 1343 | } |
34dc7c2f BB |
1344 | |
1345 | return (ZIO_PIPELINE_CONTINUE); | |
1346 | } | |
1347 | ||
b128c09f BB |
1348 | /* |
1349 | * ========================================================================== | |
1350 | * Execute the I/O pipeline | |
1351 | * ========================================================================== | |
1352 | */ | |
1353 | ||
1354 | static void | |
7ef5e54e | 1355 | zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) |
34dc7c2f | 1356 | { |
428870ff | 1357 | spa_t *spa = zio->io_spa; |
b128c09f | 1358 | zio_type_t t = zio->io_type; |
a38718a6 | 1359 | int flags = (cutinline ? TQ_FRONT : 0); |
34dc7c2f BB |
1360 | |
1361 | /* | |
9babb374 BB |
1362 | * If we're a config writer or a probe, the normal issue and |
1363 | * interrupt threads may all be blocked waiting for the config lock. | |
1364 | * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. | |
34dc7c2f | 1365 | */ |
9babb374 | 1366 | if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) |
b128c09f | 1367 | t = ZIO_TYPE_NULL; |
34dc7c2f BB |
1368 | |
1369 | /* | |
b128c09f | 1370 | * A similar issue exists for the L2ARC write thread until L2ARC 2.0. |
34dc7c2f | 1371 | */ |
b128c09f BB |
1372 | if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) |
1373 | t = ZIO_TYPE_NULL; | |
34dc7c2f | 1374 | |
428870ff | 1375 | /* |
7ef5e54e AL |
1376 | * If this is a high priority I/O, then use the high priority taskq if |
1377 | * available. | |
428870ff BB |
1378 | */ |
1379 | if (zio->io_priority == ZIO_PRIORITY_NOW && | |
7ef5e54e | 1380 | spa->spa_zio_taskq[t][q + 1].stqs_count != 0) |
428870ff BB |
1381 | q++; |
1382 | ||
1383 | ASSERT3U(q, <, ZIO_TASKQ_TYPES); | |
5cc556b4 | 1384 | |
a38718a6 GA |
1385 | /* |
1386 | * NB: We are assuming that the zio can only be dispatched | |
1387 | * to a single taskq at a time. It would be a grievous error | |
1388 | * to dispatch the zio to another taskq at the same time. | |
1389 | */ | |
1390 | ASSERT(taskq_empty_ent(&zio->io_tqent)); | |
7ef5e54e AL |
1391 | spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, |
1392 | flags, &zio->io_tqent); | |
b128c09f | 1393 | } |
34dc7c2f | 1394 | |
b128c09f | 1395 | static boolean_t |
7ef5e54e | 1396 | zio_taskq_member(zio_t *zio, zio_taskq_type_t q) |
b128c09f BB |
1397 | { |
1398 | kthread_t *executor = zio->io_executor; | |
1399 | spa_t *spa = zio->io_spa; | |
d6320ddb | 1400 | zio_type_t t; |
34dc7c2f | 1401 | |
7ef5e54e AL |
1402 | for (t = 0; t < ZIO_TYPES; t++) { |
1403 | spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; | |
1404 | uint_t i; | |
1405 | for (i = 0; i < tqs->stqs_count; i++) { | |
1406 | if (taskq_member(tqs->stqs_taskq[i], executor)) | |
1407 | return (B_TRUE); | |
1408 | } | |
1409 | } | |
34dc7c2f | 1410 | |
b128c09f BB |
1411 | return (B_FALSE); |
1412 | } | |
34dc7c2f | 1413 | |
b128c09f BB |
1414 | static int |
1415 | zio_issue_async(zio_t *zio) | |
1416 | { | |
428870ff | 1417 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); |
b128c09f BB |
1418 | |
1419 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f BB |
1420 | } |
1421 | ||
b128c09f BB |
1422 | void |
1423 | zio_interrupt(zio_t *zio) | |
34dc7c2f | 1424 | { |
428870ff | 1425 | zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); |
b128c09f | 1426 | } |
34dc7c2f | 1427 | |
26ef0cc7 TH |
1428 | void |
1429 | zio_delay_interrupt(zio_t *zio) | |
1430 | { | |
1431 | /* | |
1432 | * The timeout_generic() function isn't defined in userspace, so | |
1433 | * rather than trying to implement the function, the zio delay | |
1434 | * functionality has been disabled for userspace builds. | |
1435 | */ | |
1436 | ||
1437 | #ifdef _KERNEL | |
1438 | /* | |
1439 | * If io_target_timestamp is zero, then no delay has been registered | |
1440 | * for this IO, thus jump to the end of this function and "skip" the | |
1441 | * delay; issuing it directly to the zio layer. | |
1442 | */ | |
1443 | if (zio->io_target_timestamp != 0) { | |
1444 | hrtime_t now = gethrtime(); | |
1445 | ||
1446 | if (now >= zio->io_target_timestamp) { | |
1447 | /* | |
1448 | * This IO has already taken longer than the target | |
1449 | * delay to complete, so we don't want to delay it | |
1450 | * any longer; we "miss" the delay and issue it | |
1451 | * directly to the zio layer. This is likely due to | |
1452 | * the target latency being set to a value less than | |
1453 | * the underlying hardware can satisfy (e.g. delay | |
1454 | * set to 1ms, but the disks take 10ms to complete an | |
1455 | * IO request). | |
1456 | */ | |
1457 | ||
1458 | DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, | |
1459 | hrtime_t, now); | |
1460 | ||
1461 | zio_interrupt(zio); | |
1462 | } else { | |
1463 | taskqid_t tid; | |
1464 | hrtime_t diff = zio->io_target_timestamp - now; | |
1465 | clock_t expire_at_tick = ddi_get_lbolt() + | |
1466 | NSEC_TO_TICK(diff); | |
1467 | ||
1468 | DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, | |
1469 | hrtime_t, now, hrtime_t, diff); | |
1470 | ||
1471 | if (NSEC_TO_TICK(diff) == 0) { | |
1472 | /* Our delay is less than a jiffy - just spin */ | |
1473 | zfs_sleep_until(zio->io_target_timestamp); | |
1474 | } else { | |
1475 | /* | |
1476 | * Use taskq_dispatch_delay() in the place of | |
1477 | * OpenZFS's timeout_generic(). | |
1478 | */ | |
1479 | tid = taskq_dispatch_delay(system_taskq, | |
1480 | (task_func_t *) zio_interrupt, | |
1481 | zio, TQ_NOSLEEP, expire_at_tick); | |
1482 | if (!tid) { | |
1483 | /* | |
1484 | * Couldn't allocate a task. Just | |
1485 | * finish the zio without a delay. | |
1486 | */ | |
1487 | zio_interrupt(zio); | |
1488 | } | |
1489 | } | |
1490 | } | |
1491 | return; | |
1492 | } | |
1493 | #endif | |
1494 | DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); | |
1495 | zio_interrupt(zio); | |
1496 | } | |
1497 | ||
b128c09f BB |
1498 | /* |
1499 | * Execute the I/O pipeline until one of the following occurs: | |
1500 | * (1) the I/O completes; (2) the pipeline stalls waiting for | |
1501 | * dependent child I/Os; (3) the I/O issues, so we're waiting | |
1502 | * for an I/O completion interrupt; (4) the I/O is delegated by | |
1503 | * vdev-level caching or aggregation; (5) the I/O is deferred | |
1504 | * due to vdev-level queueing; (6) the I/O is handed off to | |
1505 | * another thread. In all cases, the pipeline stops whenever | |
8e07b99b | 1506 | * there's no CPU work; it never burns a thread in cv_wait_io(). |
b128c09f BB |
1507 | * |
1508 | * There's no locking on io_stage because there's no legitimate way | |
1509 | * for multiple threads to be attempting to process the same I/O. | |
1510 | */ | |
428870ff | 1511 | static zio_pipe_stage_t *zio_pipeline[]; |
34dc7c2f | 1512 | |
da6b4005 NB |
1513 | /* |
1514 | * zio_execute() is a wrapper around the static function | |
1515 | * __zio_execute() so that we can force __zio_execute() to be | |
1516 | * inlined. This reduces stack overhead which is important | |
1517 | * because __zio_execute() is called recursively in several zio | |
1518 | * code paths. zio_execute() itself cannot be inlined because | |
1519 | * it is externally visible. | |
1520 | */ | |
b128c09f BB |
1521 | void |
1522 | zio_execute(zio_t *zio) | |
da6b4005 | 1523 | { |
92119cc2 BB |
1524 | fstrans_cookie_t cookie; |
1525 | ||
1526 | cookie = spl_fstrans_mark(); | |
da6b4005 | 1527 | __zio_execute(zio); |
92119cc2 | 1528 | spl_fstrans_unmark(cookie); |
da6b4005 NB |
1529 | } |
1530 | ||
b58986ee BB |
1531 | /* |
1532 | * Used to determine if in the current context the stack is sized large | |
1533 | * enough to allow zio_execute() to be called recursively. A minimum | |
1534 | * stack size of 16K is required to avoid needing to re-dispatch the zio. | |
1535 | */ | |
1536 | boolean_t | |
1537 | zio_execute_stack_check(zio_t *zio) | |
1538 | { | |
1539 | #if !defined(HAVE_LARGE_STACKS) | |
1540 | dsl_pool_t *dp = spa_get_dsl(zio->io_spa); | |
1541 | ||
1542 | /* Executing in txg_sync_thread() context. */ | |
1543 | if (dp && curthread == dp->dp_tx.tx_sync_thread) | |
1544 | return (B_TRUE); | |
1545 | ||
1546 | /* Pool initialization outside of zio_taskq context. */ | |
1547 | if (dp && spa_is_initializing(dp->dp_spa) && | |
1548 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) && | |
1549 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH)) | |
1550 | return (B_TRUE); | |
1551 | #endif /* HAVE_LARGE_STACKS */ | |
1552 | ||
1553 | return (B_FALSE); | |
1554 | } | |
1555 | ||
da6b4005 NB |
1556 | __attribute__((always_inline)) |
1557 | static inline void | |
1558 | __zio_execute(zio_t *zio) | |
b128c09f BB |
1559 | { |
1560 | zio->io_executor = curthread; | |
34dc7c2f | 1561 | |
b128c09f | 1562 | while (zio->io_stage < ZIO_STAGE_DONE) { |
428870ff BB |
1563 | enum zio_stage pipeline = zio->io_pipeline; |
1564 | enum zio_stage stage = zio->io_stage; | |
b128c09f | 1565 | int rv; |
34dc7c2f | 1566 | |
b128c09f | 1567 | ASSERT(!MUTEX_HELD(&zio->io_lock)); |
428870ff BB |
1568 | ASSERT(ISP2(stage)); |
1569 | ASSERT(zio->io_stall == NULL); | |
34dc7c2f | 1570 | |
428870ff BB |
1571 | do { |
1572 | stage <<= 1; | |
1573 | } while ((stage & pipeline) == 0); | |
b128c09f BB |
1574 | |
1575 | ASSERT(stage <= ZIO_STAGE_DONE); | |
34dc7c2f BB |
1576 | |
1577 | /* | |
b128c09f BB |
1578 | * If we are in interrupt context and this pipeline stage |
1579 | * will grab a config lock that is held across I/O, | |
428870ff BB |
1580 | * or may wait for an I/O that needs an interrupt thread |
1581 | * to complete, issue async to avoid deadlock. | |
1582 | * | |
1583 | * For VDEV_IO_START, we cut in line so that the io will | |
1584 | * be sent to disk promptly. | |
34dc7c2f | 1585 | */ |
91579709 BB |
1586 | if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && |
1587 | zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { | |
b58986ee BB |
1588 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? |
1589 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
91579709 BB |
1590 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
1591 | return; | |
1592 | } | |
1593 | ||
1594 | /* | |
b58986ee BB |
1595 | * If the current context doesn't have large enough stacks |
1596 | * the zio must be issued asynchronously to prevent overflow. | |
91579709 | 1597 | */ |
b58986ee BB |
1598 | if (zio_execute_stack_check(zio)) { |
1599 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? | |
1600 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
428870ff | 1601 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
b128c09f | 1602 | return; |
34dc7c2f BB |
1603 | } |
1604 | ||
b128c09f | 1605 | zio->io_stage = stage; |
9bd274dd | 1606 | rv = zio_pipeline[highbit64(stage) - 1](zio); |
34dc7c2f | 1607 | |
b128c09f BB |
1608 | if (rv == ZIO_PIPELINE_STOP) |
1609 | return; | |
34dc7c2f | 1610 | |
b128c09f BB |
1611 | ASSERT(rv == ZIO_PIPELINE_CONTINUE); |
1612 | } | |
34dc7c2f BB |
1613 | } |
1614 | ||
da6b4005 | 1615 | |
b128c09f BB |
1616 | /* |
1617 | * ========================================================================== | |
1618 | * Initiate I/O, either sync or async | |
1619 | * ========================================================================== | |
1620 | */ | |
1621 | int | |
1622 | zio_wait(zio_t *zio) | |
34dc7c2f | 1623 | { |
b128c09f | 1624 | int error; |
34dc7c2f | 1625 | |
b128c09f BB |
1626 | ASSERT(zio->io_stage == ZIO_STAGE_OPEN); |
1627 | ASSERT(zio->io_executor == NULL); | |
34dc7c2f | 1628 | |
b128c09f | 1629 | zio->io_waiter = curthread; |
34dc7c2f | 1630 | |
da6b4005 | 1631 | __zio_execute(zio); |
34dc7c2f | 1632 | |
b128c09f | 1633 | mutex_enter(&zio->io_lock); |
72f53c56 | 1634 | while (zio->io_executor != NULL) |
72938d69 | 1635 | cv_wait_io(&zio->io_cv, &zio->io_lock); |
b128c09f | 1636 | mutex_exit(&zio->io_lock); |
34dc7c2f | 1637 | |
b128c09f BB |
1638 | error = zio->io_error; |
1639 | zio_destroy(zio); | |
34dc7c2f | 1640 | |
b128c09f BB |
1641 | return (error); |
1642 | } | |
34dc7c2f | 1643 | |
b128c09f BB |
1644 | void |
1645 | zio_nowait(zio_t *zio) | |
1646 | { | |
1647 | ASSERT(zio->io_executor == NULL); | |
34dc7c2f | 1648 | |
d164b209 BB |
1649 | if (zio->io_child_type == ZIO_CHILD_LOGICAL && |
1650 | zio_unique_parent(zio) == NULL) { | |
8878261f BB |
1651 | zio_t *pio; |
1652 | ||
34dc7c2f | 1653 | /* |
b128c09f | 1654 | * This is a logical async I/O with no parent to wait for it. |
9babb374 BB |
1655 | * We add it to the spa_async_root_zio "Godfather" I/O which |
1656 | * will ensure they complete prior to unloading the pool. | |
34dc7c2f | 1657 | */ |
b128c09f | 1658 | spa_t *spa = zio->io_spa; |
8878261f BB |
1659 | kpreempt_disable(); |
1660 | pio = spa->spa_async_zio_root[CPU_SEQID]; | |
1661 | kpreempt_enable(); | |
9babb374 | 1662 | |
8878261f | 1663 | zio_add_child(pio, zio); |
b128c09f | 1664 | } |
34dc7c2f | 1665 | |
da6b4005 | 1666 | __zio_execute(zio); |
b128c09f | 1667 | } |
34dc7c2f | 1668 | |
b128c09f BB |
1669 | /* |
1670 | * ========================================================================== | |
1671 | * Reexecute or suspend/resume failed I/O | |
1672 | * ========================================================================== | |
1673 | */ | |
34dc7c2f | 1674 | |
b128c09f BB |
1675 | static void |
1676 | zio_reexecute(zio_t *pio) | |
1677 | { | |
d164b209 | 1678 | zio_t *cio, *cio_next; |
d6320ddb | 1679 | int c, w; |
d164b209 BB |
1680 | |
1681 | ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); | |
1682 | ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); | |
9babb374 BB |
1683 | ASSERT(pio->io_gang_leader == NULL); |
1684 | ASSERT(pio->io_gang_tree == NULL); | |
34dc7c2f | 1685 | |
b128c09f BB |
1686 | pio->io_flags = pio->io_orig_flags; |
1687 | pio->io_stage = pio->io_orig_stage; | |
1688 | pio->io_pipeline = pio->io_orig_pipeline; | |
1689 | pio->io_reexecute = 0; | |
03c6040b | 1690 | pio->io_flags |= ZIO_FLAG_REEXECUTED; |
b128c09f | 1691 | pio->io_error = 0; |
d6320ddb | 1692 | for (w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 | 1693 | pio->io_state[w] = 0; |
d6320ddb | 1694 | for (c = 0; c < ZIO_CHILD_TYPES; c++) |
b128c09f | 1695 | pio->io_child_error[c] = 0; |
34dc7c2f | 1696 | |
428870ff BB |
1697 | if (IO_IS_ALLOCATING(pio)) |
1698 | BP_ZERO(pio->io_bp); | |
34dc7c2f | 1699 | |
b128c09f BB |
1700 | /* |
1701 | * As we reexecute pio's children, new children could be created. | |
d164b209 | 1702 | * New children go to the head of pio's io_child_list, however, |
b128c09f | 1703 | * so we will (correctly) not reexecute them. The key is that |
d164b209 BB |
1704 | * the remainder of pio's io_child_list, from 'cio_next' onward, |
1705 | * cannot be affected by any side effects of reexecuting 'cio'. | |
b128c09f | 1706 | */ |
d164b209 BB |
1707 | for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) { |
1708 | cio_next = zio_walk_children(pio); | |
b128c09f | 1709 | mutex_enter(&pio->io_lock); |
d6320ddb | 1710 | for (w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 | 1711 | pio->io_children[cio->io_child_type][w]++; |
b128c09f | 1712 | mutex_exit(&pio->io_lock); |
d164b209 | 1713 | zio_reexecute(cio); |
34dc7c2f | 1714 | } |
34dc7c2f | 1715 | |
b128c09f BB |
1716 | /* |
1717 | * Now that all children have been reexecuted, execute the parent. | |
9babb374 BB |
1718 | * We don't reexecute "The Godfather" I/O here as it's the |
1719 | * responsibility of the caller to wait on him. | |
b128c09f | 1720 | */ |
9babb374 | 1721 | if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) |
da6b4005 | 1722 | __zio_execute(pio); |
34dc7c2f BB |
1723 | } |
1724 | ||
b128c09f BB |
1725 | void |
1726 | zio_suspend(spa_t *spa, zio_t *zio) | |
34dc7c2f | 1727 | { |
b128c09f BB |
1728 | if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) |
1729 | fm_panic("Pool '%s' has encountered an uncorrectable I/O " | |
1730 | "failure and the failure mode property for this pool " | |
1731 | "is set to panic.", spa_name(spa)); | |
34dc7c2f | 1732 | |
bf89c199 BB |
1733 | cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O " |
1734 | "failure and has been suspended.\n", spa_name(spa)); | |
1735 | ||
b128c09f | 1736 | zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); |
34dc7c2f | 1737 | |
b128c09f | 1738 | mutex_enter(&spa->spa_suspend_lock); |
34dc7c2f | 1739 | |
b128c09f | 1740 | if (spa->spa_suspend_zio_root == NULL) |
9babb374 BB |
1741 | spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, |
1742 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
1743 | ZIO_FLAG_GODFATHER); | |
34dc7c2f | 1744 | |
b128c09f | 1745 | spa->spa_suspended = B_TRUE; |
34dc7c2f | 1746 | |
b128c09f | 1747 | if (zio != NULL) { |
9babb374 | 1748 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f BB |
1749 | ASSERT(zio != spa->spa_suspend_zio_root); |
1750 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
d164b209 | 1751 | ASSERT(zio_unique_parent(zio) == NULL); |
b128c09f BB |
1752 | ASSERT(zio->io_stage == ZIO_STAGE_DONE); |
1753 | zio_add_child(spa->spa_suspend_zio_root, zio); | |
1754 | } | |
34dc7c2f | 1755 | |
b128c09f BB |
1756 | mutex_exit(&spa->spa_suspend_lock); |
1757 | } | |
34dc7c2f | 1758 | |
9babb374 | 1759 | int |
b128c09f BB |
1760 | zio_resume(spa_t *spa) |
1761 | { | |
9babb374 | 1762 | zio_t *pio; |
34dc7c2f BB |
1763 | |
1764 | /* | |
b128c09f | 1765 | * Reexecute all previously suspended i/o. |
34dc7c2f | 1766 | */ |
b128c09f BB |
1767 | mutex_enter(&spa->spa_suspend_lock); |
1768 | spa->spa_suspended = B_FALSE; | |
1769 | cv_broadcast(&spa->spa_suspend_cv); | |
1770 | pio = spa->spa_suspend_zio_root; | |
1771 | spa->spa_suspend_zio_root = NULL; | |
1772 | mutex_exit(&spa->spa_suspend_lock); | |
1773 | ||
1774 | if (pio == NULL) | |
9babb374 | 1775 | return (0); |
34dc7c2f | 1776 | |
9babb374 BB |
1777 | zio_reexecute(pio); |
1778 | return (zio_wait(pio)); | |
b128c09f BB |
1779 | } |
1780 | ||
1781 | void | |
1782 | zio_resume_wait(spa_t *spa) | |
1783 | { | |
1784 | mutex_enter(&spa->spa_suspend_lock); | |
1785 | while (spa_suspended(spa)) | |
1786 | cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); | |
1787 | mutex_exit(&spa->spa_suspend_lock); | |
34dc7c2f BB |
1788 | } |
1789 | ||
1790 | /* | |
1791 | * ========================================================================== | |
b128c09f BB |
1792 | * Gang blocks. |
1793 | * | |
1794 | * A gang block is a collection of small blocks that looks to the DMU | |
1795 | * like one large block. When zio_dva_allocate() cannot find a block | |
1796 | * of the requested size, due to either severe fragmentation or the pool | |
1797 | * being nearly full, it calls zio_write_gang_block() to construct the | |
1798 | * block from smaller fragments. | |
1799 | * | |
1800 | * A gang block consists of a gang header (zio_gbh_phys_t) and up to | |
1801 | * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like | |
1802 | * an indirect block: it's an array of block pointers. It consumes | |
1803 | * only one sector and hence is allocatable regardless of fragmentation. | |
1804 | * The gang header's bps point to its gang members, which hold the data. | |
1805 | * | |
1806 | * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> | |
1807 | * as the verifier to ensure uniqueness of the SHA256 checksum. | |
1808 | * Critically, the gang block bp's blk_cksum is the checksum of the data, | |
1809 | * not the gang header. This ensures that data block signatures (needed for | |
1810 | * deduplication) are independent of how the block is physically stored. | |
1811 | * | |
1812 | * Gang blocks can be nested: a gang member may itself be a gang block. | |
1813 | * Thus every gang block is a tree in which root and all interior nodes are | |
1814 | * gang headers, and the leaves are normal blocks that contain user data. | |
1815 | * The root of the gang tree is called the gang leader. | |
1816 | * | |
1817 | * To perform any operation (read, rewrite, free, claim) on a gang block, | |
1818 | * zio_gang_assemble() first assembles the gang tree (minus data leaves) | |
1819 | * in the io_gang_tree field of the original logical i/o by recursively | |
1820 | * reading the gang leader and all gang headers below it. This yields | |
1821 | * an in-core tree containing the contents of every gang header and the | |
1822 | * bps for every constituent of the gang block. | |
1823 | * | |
1824 | * With the gang tree now assembled, zio_gang_issue() just walks the gang tree | |
1825 | * and invokes a callback on each bp. To free a gang block, zio_gang_issue() | |
1826 | * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. | |
1827 | * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). | |
1828 | * zio_read_gang() is a wrapper around zio_read() that omits reading gang | |
1829 | * headers, since we already have those in io_gang_tree. zio_rewrite_gang() | |
1830 | * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() | |
1831 | * of the gang header plus zio_checksum_compute() of the data to update the | |
1832 | * gang header's blk_cksum as described above. | |
1833 | * | |
1834 | * The two-phase assemble/issue model solves the problem of partial failure -- | |
1835 | * what if you'd freed part of a gang block but then couldn't read the | |
1836 | * gang header for another part? Assembling the entire gang tree first | |
1837 | * ensures that all the necessary gang header I/O has succeeded before | |
1838 | * starting the actual work of free, claim, or write. Once the gang tree | |
1839 | * is assembled, free and claim are in-memory operations that cannot fail. | |
1840 | * | |
1841 | * In the event that a gang write fails, zio_dva_unallocate() walks the | |
1842 | * gang tree to immediately free (i.e. insert back into the space map) | |
1843 | * everything we've allocated. This ensures that we don't get ENOSPC | |
1844 | * errors during repeated suspend/resume cycles due to a flaky device. | |
1845 | * | |
1846 | * Gang rewrites only happen during sync-to-convergence. If we can't assemble | |
1847 | * the gang tree, we won't modify the block, so we can safely defer the free | |
1848 | * (knowing that the block is still intact). If we *can* assemble the gang | |
1849 | * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free | |
1850 | * each constituent bp and we can allocate a new block on the next sync pass. | |
1851 | * | |
1852 | * In all cases, the gang tree allows complete recovery from partial failure. | |
34dc7c2f BB |
1853 | * ========================================================================== |
1854 | */ | |
b128c09f BB |
1855 | |
1856 | static zio_t * | |
1857 | zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) | |
34dc7c2f | 1858 | { |
b128c09f BB |
1859 | if (gn != NULL) |
1860 | return (pio); | |
34dc7c2f | 1861 | |
b128c09f BB |
1862 | return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp), |
1863 | NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
1864 | &pio->io_bookmark)); | |
1865 | } | |
1866 | ||
1867 | zio_t * | |
1868 | zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) | |
1869 | { | |
1870 | zio_t *zio; | |
1871 | ||
1872 | if (gn != NULL) { | |
1873 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, | |
1874 | gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority, | |
1875 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); | |
34dc7c2f | 1876 | /* |
b128c09f BB |
1877 | * As we rewrite each gang header, the pipeline will compute |
1878 | * a new gang block header checksum for it; but no one will | |
1879 | * compute a new data checksum, so we do that here. The one | |
1880 | * exception is the gang leader: the pipeline already computed | |
1881 | * its data checksum because that stage precedes gang assembly. | |
1882 | * (Presently, nothing actually uses interior data checksums; | |
1883 | * this is just good hygiene.) | |
34dc7c2f | 1884 | */ |
9babb374 | 1885 | if (gn != pio->io_gang_leader->io_gang_tree) { |
b128c09f BB |
1886 | zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), |
1887 | data, BP_GET_PSIZE(bp)); | |
1888 | } | |
428870ff BB |
1889 | /* |
1890 | * If we are here to damage data for testing purposes, | |
1891 | * leave the GBH alone so that we can detect the damage. | |
1892 | */ | |
1893 | if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) | |
1894 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; | |
34dc7c2f | 1895 | } else { |
b128c09f BB |
1896 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
1897 | data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority, | |
1898 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); | |
34dc7c2f BB |
1899 | } |
1900 | ||
b128c09f BB |
1901 | return (zio); |
1902 | } | |
34dc7c2f | 1903 | |
b128c09f BB |
1904 | /* ARGSUSED */ |
1905 | zio_t * | |
1906 | zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) | |
1907 | { | |
428870ff BB |
1908 | return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, |
1909 | ZIO_GANG_CHILD_FLAGS(pio))); | |
34dc7c2f BB |
1910 | } |
1911 | ||
b128c09f BB |
1912 | /* ARGSUSED */ |
1913 | zio_t * | |
1914 | zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) | |
34dc7c2f | 1915 | { |
b128c09f BB |
1916 | return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, |
1917 | NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); | |
1918 | } | |
1919 | ||
1920 | static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { | |
1921 | NULL, | |
1922 | zio_read_gang, | |
1923 | zio_rewrite_gang, | |
1924 | zio_free_gang, | |
1925 | zio_claim_gang, | |
1926 | NULL | |
1927 | }; | |
34dc7c2f | 1928 | |
b128c09f | 1929 | static void zio_gang_tree_assemble_done(zio_t *zio); |
34dc7c2f | 1930 | |
b128c09f BB |
1931 | static zio_gang_node_t * |
1932 | zio_gang_node_alloc(zio_gang_node_t **gnpp) | |
1933 | { | |
1934 | zio_gang_node_t *gn; | |
34dc7c2f | 1935 | |
b128c09f | 1936 | ASSERT(*gnpp == NULL); |
34dc7c2f | 1937 | |
79c76d5b | 1938 | gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); |
b128c09f BB |
1939 | gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); |
1940 | *gnpp = gn; | |
34dc7c2f | 1941 | |
b128c09f | 1942 | return (gn); |
34dc7c2f BB |
1943 | } |
1944 | ||
34dc7c2f | 1945 | static void |
b128c09f | 1946 | zio_gang_node_free(zio_gang_node_t **gnpp) |
34dc7c2f | 1947 | { |
b128c09f | 1948 | zio_gang_node_t *gn = *gnpp; |
d6320ddb | 1949 | int g; |
34dc7c2f | 1950 | |
d6320ddb | 1951 | for (g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f BB |
1952 | ASSERT(gn->gn_child[g] == NULL); |
1953 | ||
1954 | zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
1955 | kmem_free(gn, sizeof (*gn)); | |
1956 | *gnpp = NULL; | |
34dc7c2f BB |
1957 | } |
1958 | ||
b128c09f BB |
1959 | static void |
1960 | zio_gang_tree_free(zio_gang_node_t **gnpp) | |
34dc7c2f | 1961 | { |
b128c09f | 1962 | zio_gang_node_t *gn = *gnpp; |
d6320ddb | 1963 | int g; |
34dc7c2f | 1964 | |
b128c09f BB |
1965 | if (gn == NULL) |
1966 | return; | |
34dc7c2f | 1967 | |
d6320ddb | 1968 | for (g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f | 1969 | zio_gang_tree_free(&gn->gn_child[g]); |
34dc7c2f | 1970 | |
b128c09f | 1971 | zio_gang_node_free(gnpp); |
34dc7c2f BB |
1972 | } |
1973 | ||
b128c09f | 1974 | static void |
9babb374 | 1975 | zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) |
34dc7c2f | 1976 | { |
b128c09f BB |
1977 | zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); |
1978 | ||
9babb374 | 1979 | ASSERT(gio->io_gang_leader == gio); |
b128c09f | 1980 | ASSERT(BP_IS_GANG(bp)); |
34dc7c2f | 1981 | |
9babb374 | 1982 | zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh, |
b128c09f | 1983 | SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn, |
9babb374 | 1984 | gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); |
b128c09f | 1985 | } |
34dc7c2f | 1986 | |
b128c09f BB |
1987 | static void |
1988 | zio_gang_tree_assemble_done(zio_t *zio) | |
1989 | { | |
9babb374 | 1990 | zio_t *gio = zio->io_gang_leader; |
b128c09f BB |
1991 | zio_gang_node_t *gn = zio->io_private; |
1992 | blkptr_t *bp = zio->io_bp; | |
d6320ddb | 1993 | int g; |
34dc7c2f | 1994 | |
9babb374 | 1995 | ASSERT(gio == zio_unique_parent(zio)); |
428870ff | 1996 | ASSERT(zio->io_child_count == 0); |
34dc7c2f | 1997 | |
b128c09f BB |
1998 | if (zio->io_error) |
1999 | return; | |
34dc7c2f | 2000 | |
b128c09f BB |
2001 | if (BP_SHOULD_BYTESWAP(bp)) |
2002 | byteswap_uint64_array(zio->io_data, zio->io_size); | |
34dc7c2f | 2003 | |
b128c09f BB |
2004 | ASSERT(zio->io_data == gn->gn_gbh); |
2005 | ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); | |
428870ff | 2006 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2007 | |
d6320ddb | 2008 | for (g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2009 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2010 | if (!BP_IS_GANG(gbp)) | |
2011 | continue; | |
9babb374 | 2012 | zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); |
b128c09f | 2013 | } |
34dc7c2f BB |
2014 | } |
2015 | ||
b128c09f BB |
2016 | static void |
2017 | zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data) | |
34dc7c2f | 2018 | { |
9babb374 | 2019 | zio_t *gio = pio->io_gang_leader; |
b128c09f | 2020 | zio_t *zio; |
d6320ddb | 2021 | int g; |
34dc7c2f | 2022 | |
b128c09f | 2023 | ASSERT(BP_IS_GANG(bp) == !!gn); |
9babb374 BB |
2024 | ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); |
2025 | ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); | |
34dc7c2f | 2026 | |
b128c09f BB |
2027 | /* |
2028 | * If you're a gang header, your data is in gn->gn_gbh. | |
2029 | * If you're a gang member, your data is in 'data' and gn == NULL. | |
2030 | */ | |
9babb374 | 2031 | zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data); |
34dc7c2f | 2032 | |
b128c09f | 2033 | if (gn != NULL) { |
428870ff | 2034 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2035 | |
d6320ddb | 2036 | for (g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2037 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2038 | if (BP_IS_HOLE(gbp)) | |
2039 | continue; | |
2040 | zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data); | |
2041 | data = (char *)data + BP_GET_PSIZE(gbp); | |
2042 | } | |
34dc7c2f BB |
2043 | } |
2044 | ||
9babb374 BB |
2045 | if (gn == gio->io_gang_tree) |
2046 | ASSERT3P((char *)gio->io_data + gio->io_size, ==, data); | |
34dc7c2f | 2047 | |
b128c09f BB |
2048 | if (zio != pio) |
2049 | zio_nowait(zio); | |
34dc7c2f BB |
2050 | } |
2051 | ||
2052 | static int | |
b128c09f | 2053 | zio_gang_assemble(zio_t *zio) |
34dc7c2f | 2054 | { |
b128c09f | 2055 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2056 | |
9babb374 BB |
2057 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); |
2058 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
2059 | ||
2060 | zio->io_gang_leader = zio; | |
34dc7c2f | 2061 | |
b128c09f | 2062 | zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); |
34dc7c2f BB |
2063 | |
2064 | return (ZIO_PIPELINE_CONTINUE); | |
2065 | } | |
2066 | ||
2067 | static int | |
b128c09f | 2068 | zio_gang_issue(zio_t *zio) |
34dc7c2f | 2069 | { |
b128c09f | 2070 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2071 | |
b128c09f BB |
2072 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) |
2073 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f | 2074 | |
9babb374 BB |
2075 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); |
2076 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
34dc7c2f | 2077 | |
b128c09f | 2078 | if (zio->io_child_error[ZIO_CHILD_GANG] == 0) |
9babb374 | 2079 | zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data); |
b128c09f | 2080 | else |
9babb374 | 2081 | zio_gang_tree_free(&zio->io_gang_tree); |
34dc7c2f | 2082 | |
b128c09f | 2083 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f BB |
2084 | |
2085 | return (ZIO_PIPELINE_CONTINUE); | |
2086 | } | |
2087 | ||
2088 | static void | |
b128c09f | 2089 | zio_write_gang_member_ready(zio_t *zio) |
34dc7c2f | 2090 | { |
d164b209 | 2091 | zio_t *pio = zio_unique_parent(zio); |
34dc7c2f BB |
2092 | dva_t *cdva = zio->io_bp->blk_dva; |
2093 | dva_t *pdva = pio->io_bp->blk_dva; | |
2094 | uint64_t asize; | |
d6320ddb | 2095 | int d; |
d1d7e268 | 2096 | ASSERTV(zio_t *gio = zio->io_gang_leader); |
34dc7c2f | 2097 | |
b128c09f BB |
2098 | if (BP_IS_HOLE(zio->io_bp)) |
2099 | return; | |
2100 | ||
2101 | ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); | |
2102 | ||
2103 | ASSERT(zio->io_child_type == ZIO_CHILD_GANG); | |
428870ff BB |
2104 | ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); |
2105 | ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); | |
2106 | ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); | |
34dc7c2f | 2107 | ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); |
34dc7c2f BB |
2108 | |
2109 | mutex_enter(&pio->io_lock); | |
d6320ddb | 2110 | for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { |
34dc7c2f BB |
2111 | ASSERT(DVA_GET_GANG(&pdva[d])); |
2112 | asize = DVA_GET_ASIZE(&pdva[d]); | |
2113 | asize += DVA_GET_ASIZE(&cdva[d]); | |
2114 | DVA_SET_ASIZE(&pdva[d], asize); | |
2115 | } | |
2116 | mutex_exit(&pio->io_lock); | |
2117 | } | |
2118 | ||
2119 | static int | |
b128c09f | 2120 | zio_write_gang_block(zio_t *pio) |
34dc7c2f | 2121 | { |
b128c09f BB |
2122 | spa_t *spa = pio->io_spa; |
2123 | blkptr_t *bp = pio->io_bp; | |
9babb374 | 2124 | zio_t *gio = pio->io_gang_leader; |
b128c09f BB |
2125 | zio_t *zio; |
2126 | zio_gang_node_t *gn, **gnpp; | |
34dc7c2f | 2127 | zio_gbh_phys_t *gbh; |
b128c09f BB |
2128 | uint64_t txg = pio->io_txg; |
2129 | uint64_t resid = pio->io_size; | |
2130 | uint64_t lsize; | |
428870ff BB |
2131 | int copies = gio->io_prop.zp_copies; |
2132 | int gbh_copies = MIN(copies + 1, spa_max_replication(spa)); | |
b128c09f | 2133 | zio_prop_t zp; |
d6320ddb | 2134 | int g, error; |
34dc7c2f | 2135 | |
428870ff BB |
2136 | error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE, |
2137 | bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, | |
b128c09f | 2138 | METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER); |
34dc7c2f | 2139 | if (error) { |
b128c09f | 2140 | pio->io_error = error; |
34dc7c2f BB |
2141 | return (ZIO_PIPELINE_CONTINUE); |
2142 | } | |
2143 | ||
9babb374 BB |
2144 | if (pio == gio) { |
2145 | gnpp = &gio->io_gang_tree; | |
b128c09f BB |
2146 | } else { |
2147 | gnpp = pio->io_private; | |
2148 | ASSERT(pio->io_ready == zio_write_gang_member_ready); | |
34dc7c2f BB |
2149 | } |
2150 | ||
b128c09f BB |
2151 | gn = zio_gang_node_alloc(gnpp); |
2152 | gbh = gn->gn_gbh; | |
2153 | bzero(gbh, SPA_GANGBLOCKSIZE); | |
34dc7c2f | 2154 | |
b128c09f BB |
2155 | /* |
2156 | * Create the gang header. | |
2157 | */ | |
2158 | zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL, | |
2159 | pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); | |
34dc7c2f | 2160 | |
b128c09f BB |
2161 | /* |
2162 | * Create and nowait the gang children. | |
2163 | */ | |
d6320ddb | 2164 | for (g = 0; resid != 0; resid -= lsize, g++) { |
b128c09f BB |
2165 | lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), |
2166 | SPA_MINBLOCKSIZE); | |
2167 | ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); | |
2168 | ||
9babb374 | 2169 | zp.zp_checksum = gio->io_prop.zp_checksum; |
b128c09f BB |
2170 | zp.zp_compress = ZIO_COMPRESS_OFF; |
2171 | zp.zp_type = DMU_OT_NONE; | |
2172 | zp.zp_level = 0; | |
428870ff | 2173 | zp.zp_copies = gio->io_prop.zp_copies; |
03c6040b GW |
2174 | zp.zp_dedup = B_FALSE; |
2175 | zp.zp_dedup_verify = B_FALSE; | |
2176 | zp.zp_nopwrite = B_FALSE; | |
b128c09f BB |
2177 | |
2178 | zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g], | |
2aa34383 DK |
2179 | (char *)pio->io_data + (pio->io_size - resid), lsize, |
2180 | lsize, &zp, zio_write_gang_member_ready, NULL, NULL, NULL, | |
bc77ba73 PD |
2181 | &gn->gn_child[g], pio->io_priority, |
2182 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark)); | |
b128c09f | 2183 | } |
34dc7c2f BB |
2184 | |
2185 | /* | |
b128c09f | 2186 | * Set pio's pipeline to just wait for zio to finish. |
34dc7c2f | 2187 | */ |
b128c09f BB |
2188 | pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
2189 | ||
920dd524 ED |
2190 | /* |
2191 | * We didn't allocate this bp, so make sure it doesn't get unmarked. | |
2192 | */ | |
2193 | pio->io_flags &= ~ZIO_FLAG_FASTWRITE; | |
2194 | ||
b128c09f BB |
2195 | zio_nowait(zio); |
2196 | ||
2197 | return (ZIO_PIPELINE_CONTINUE); | |
34dc7c2f BB |
2198 | } |
2199 | ||
03c6040b | 2200 | /* |
3c67d83a TH |
2201 | * The zio_nop_write stage in the pipeline determines if allocating a |
2202 | * new bp is necessary. The nopwrite feature can handle writes in | |
2203 | * either syncing or open context (i.e. zil writes) and as a result is | |
2204 | * mutually exclusive with dedup. | |
2205 | * | |
2206 | * By leveraging a cryptographically secure checksum, such as SHA256, we | |
2207 | * can compare the checksums of the new data and the old to determine if | |
2208 | * allocating a new block is required. Note that our requirements for | |
2209 | * cryptographic strength are fairly weak: there can't be any accidental | |
2210 | * hash collisions, but we don't need to be secure against intentional | |
2211 | * (malicious) collisions. To trigger a nopwrite, you have to be able | |
2212 | * to write the file to begin with, and triggering an incorrect (hash | |
2213 | * collision) nopwrite is no worse than simply writing to the file. | |
2214 | * That said, there are no known attacks against the checksum algorithms | |
2215 | * used for nopwrite, assuming that the salt and the checksums | |
2216 | * themselves remain secret. | |
03c6040b GW |
2217 | */ |
2218 | static int | |
2219 | zio_nop_write(zio_t *zio) | |
2220 | { | |
2221 | blkptr_t *bp = zio->io_bp; | |
2222 | blkptr_t *bp_orig = &zio->io_bp_orig; | |
2223 | zio_prop_t *zp = &zio->io_prop; | |
2224 | ||
2225 | ASSERT(BP_GET_LEVEL(bp) == 0); | |
2226 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
2227 | ASSERT(zp->zp_nopwrite); | |
2228 | ASSERT(!zp->zp_dedup); | |
2229 | ASSERT(zio->io_bp_override == NULL); | |
2230 | ASSERT(IO_IS_ALLOCATING(zio)); | |
2231 | ||
2232 | /* | |
2233 | * Check to see if the original bp and the new bp have matching | |
2234 | * characteristics (i.e. same checksum, compression algorithms, etc). | |
2235 | * If they don't then just continue with the pipeline which will | |
2236 | * allocate a new bp. | |
2237 | */ | |
2238 | if (BP_IS_HOLE(bp_orig) || | |
3c67d83a TH |
2239 | !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & |
2240 | ZCHECKSUM_FLAG_NOPWRITE) || | |
03c6040b GW |
2241 | BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || |
2242 | BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || | |
2243 | BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || | |
2244 | zp->zp_copies != BP_GET_NDVAS(bp_orig)) | |
2245 | return (ZIO_PIPELINE_CONTINUE); | |
2246 | ||
2247 | /* | |
2248 | * If the checksums match then reset the pipeline so that we | |
2249 | * avoid allocating a new bp and issuing any I/O. | |
2250 | */ | |
2251 | if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { | |
3c67d83a TH |
2252 | ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & |
2253 | ZCHECKSUM_FLAG_NOPWRITE); | |
03c6040b GW |
2254 | ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); |
2255 | ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); | |
2256 | ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); | |
2257 | ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, | |
2258 | sizeof (uint64_t)) == 0); | |
2259 | ||
2260 | *bp = *bp_orig; | |
2261 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
2262 | zio->io_flags |= ZIO_FLAG_NOPWRITE; | |
2263 | } | |
2264 | ||
2265 | return (ZIO_PIPELINE_CONTINUE); | |
2266 | } | |
2267 | ||
34dc7c2f BB |
2268 | /* |
2269 | * ========================================================================== | |
428870ff | 2270 | * Dedup |
34dc7c2f BB |
2271 | * ========================================================================== |
2272 | */ | |
428870ff BB |
2273 | static void |
2274 | zio_ddt_child_read_done(zio_t *zio) | |
2275 | { | |
2276 | blkptr_t *bp = zio->io_bp; | |
2277 | ddt_entry_t *dde = zio->io_private; | |
2278 | ddt_phys_t *ddp; | |
2279 | zio_t *pio = zio_unique_parent(zio); | |
2280 | ||
2281 | mutex_enter(&pio->io_lock); | |
2282 | ddp = ddt_phys_select(dde, bp); | |
2283 | if (zio->io_error == 0) | |
2284 | ddt_phys_clear(ddp); /* this ddp doesn't need repair */ | |
2285 | if (zio->io_error == 0 && dde->dde_repair_data == NULL) | |
2286 | dde->dde_repair_data = zio->io_data; | |
2287 | else | |
2288 | zio_buf_free(zio->io_data, zio->io_size); | |
2289 | mutex_exit(&pio->io_lock); | |
2290 | } | |
2291 | ||
2292 | static int | |
2293 | zio_ddt_read_start(zio_t *zio) | |
2294 | { | |
2295 | blkptr_t *bp = zio->io_bp; | |
d6320ddb | 2296 | int p; |
428870ff BB |
2297 | |
2298 | ASSERT(BP_GET_DEDUP(bp)); | |
2299 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
2300 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
2301 | ||
2302 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
2303 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
2304 | ddt_entry_t *dde = ddt_repair_start(ddt, bp); | |
2305 | ddt_phys_t *ddp = dde->dde_phys; | |
2306 | ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); | |
2307 | blkptr_t blk; | |
2308 | ||
2309 | ASSERT(zio->io_vsd == NULL); | |
2310 | zio->io_vsd = dde; | |
2311 | ||
2312 | if (ddp_self == NULL) | |
2313 | return (ZIO_PIPELINE_CONTINUE); | |
2314 | ||
d6320ddb | 2315 | for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { |
428870ff BB |
2316 | if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) |
2317 | continue; | |
2318 | ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, | |
2319 | &blk); | |
2320 | zio_nowait(zio_read(zio, zio->io_spa, &blk, | |
2321 | zio_buf_alloc(zio->io_size), zio->io_size, | |
2322 | zio_ddt_child_read_done, dde, zio->io_priority, | |
2323 | ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE, | |
2324 | &zio->io_bookmark)); | |
2325 | } | |
2326 | return (ZIO_PIPELINE_CONTINUE); | |
2327 | } | |
2328 | ||
2329 | zio_nowait(zio_read(zio, zio->io_spa, bp, | |
2330 | zio->io_data, zio->io_size, NULL, NULL, zio->io_priority, | |
2331 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); | |
2332 | ||
2333 | return (ZIO_PIPELINE_CONTINUE); | |
2334 | } | |
2335 | ||
2336 | static int | |
2337 | zio_ddt_read_done(zio_t *zio) | |
2338 | { | |
2339 | blkptr_t *bp = zio->io_bp; | |
2340 | ||
2341 | if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE)) | |
2342 | return (ZIO_PIPELINE_STOP); | |
2343 | ||
2344 | ASSERT(BP_GET_DEDUP(bp)); | |
2345 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
2346 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
2347 | ||
2348 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
2349 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
2350 | ddt_entry_t *dde = zio->io_vsd; | |
2351 | if (ddt == NULL) { | |
2352 | ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); | |
2353 | return (ZIO_PIPELINE_CONTINUE); | |
2354 | } | |
2355 | if (dde == NULL) { | |
2356 | zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; | |
2357 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); | |
2358 | return (ZIO_PIPELINE_STOP); | |
2359 | } | |
2360 | if (dde->dde_repair_data != NULL) { | |
2361 | bcopy(dde->dde_repair_data, zio->io_data, zio->io_size); | |
2362 | zio->io_child_error[ZIO_CHILD_DDT] = 0; | |
2363 | } | |
2364 | ddt_repair_done(ddt, dde); | |
2365 | zio->io_vsd = NULL; | |
2366 | } | |
2367 | ||
2368 | ASSERT(zio->io_vsd == NULL); | |
2369 | ||
2370 | return (ZIO_PIPELINE_CONTINUE); | |
2371 | } | |
2372 | ||
2373 | static boolean_t | |
2374 | zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) | |
2375 | { | |
2376 | spa_t *spa = zio->io_spa; | |
d6320ddb | 2377 | int p; |
c17bcf83 | 2378 | boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW); |
428870ff | 2379 | |
c17bcf83 | 2380 | ASSERT(!(zio->io_bp_override && do_raw)); |
2aa34383 | 2381 | |
428870ff BB |
2382 | /* |
2383 | * Note: we compare the original data, not the transformed data, | |
2384 | * because when zio->io_bp is an override bp, we will not have | |
2385 | * pushed the I/O transforms. That's an important optimization | |
2386 | * because otherwise we'd compress/encrypt all dmu_sync() data twice. | |
c17bcf83 TC |
2387 | * However, we should never get a raw, override zio so in these |
2388 | * cases we can compare the io_data directly. This is useful because | |
2389 | * it allows us to do dedup verification even if we don't have access | |
2390 | * to the original data (for instance, if the encryption keys aren't | |
2391 | * loaded). | |
428870ff | 2392 | */ |
c17bcf83 | 2393 | |
d6320ddb | 2394 | for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
2395 | zio_t *lio = dde->dde_lead_zio[p]; |
2396 | ||
c17bcf83 TC |
2397 | if (lio != NULL && do_raw) { |
2398 | return (lio->io_size != zio->io_size || | |
2399 | bcmp(zio->io_data, lio->io_data, | |
2400 | zio->io_size) != 0); | |
2401 | } else if (lio != NULL) { | |
428870ff BB |
2402 | return (lio->io_orig_size != zio->io_orig_size || |
2403 | bcmp(zio->io_orig_data, lio->io_orig_data, | |
2404 | zio->io_orig_size) != 0); | |
2405 | } | |
2406 | } | |
2407 | ||
d6320ddb | 2408 | for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
2409 | ddt_phys_t *ddp = &dde->dde_phys[p]; |
2410 | ||
c17bcf83 TC |
2411 | if (ddp->ddp_phys_birth != 0 && do_raw) { |
2412 | blkptr_t blk = *zio->io_bp; | |
2413 | uint64_t psize; | |
2414 | void *tmpbuf; | |
2415 | int error; | |
2416 | ||
2417 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
2418 | psize = BP_GET_PSIZE(&blk); | |
2419 | ||
2420 | if (psize != zio->io_size) | |
2421 | return (B_TRUE); | |
2422 | ||
2423 | ddt_exit(ddt); | |
2424 | ||
2425 | tmpbuf = zio_buf_alloc(psize); | |
2426 | ||
2427 | error = zio_wait(zio_read(NULL, spa, &blk, tmpbuf, | |
2428 | psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ, | |
2429 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
2430 | ZIO_FLAG_RAW, &zio->io_bookmark)); | |
2431 | ||
2432 | if (error == 0) { | |
2433 | if (bcmp(tmpbuf, zio->io_data, psize) != 0) | |
2434 | error = SET_ERROR(ENOENT); | |
2435 | } | |
2436 | ||
2437 | zio_buf_free(tmpbuf, psize); | |
2438 | ddt_enter(ddt); | |
2439 | return (error != 0); | |
2440 | } else if (ddp->ddp_phys_birth != 0) { | |
428870ff | 2441 | arc_buf_t *abuf = NULL; |
2a432414 | 2442 | arc_flags_t aflags = ARC_FLAG_WAIT; |
428870ff BB |
2443 | blkptr_t blk = *zio->io_bp; |
2444 | int error; | |
2445 | ||
2446 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
2447 | ||
c17bcf83 TC |
2448 | if (BP_GET_LSIZE(&blk) != zio->io_orig_size) |
2449 | return (B_TRUE); | |
2450 | ||
428870ff BB |
2451 | ddt_exit(ddt); |
2452 | ||
294f6806 | 2453 | error = arc_read(NULL, spa, &blk, |
428870ff BB |
2454 | arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, |
2455 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
2456 | &aflags, &zio->io_bookmark); | |
2457 | ||
2458 | if (error == 0) { | |
c17bcf83 | 2459 | if (bcmp(abuf->b_data, zio->io_orig_data, |
428870ff | 2460 | zio->io_orig_size) != 0) |
c17bcf83 | 2461 | error = SET_ERROR(ENOENT); |
d3c2ae1c | 2462 | arc_buf_destroy(abuf, &abuf); |
428870ff BB |
2463 | } |
2464 | ||
2465 | ddt_enter(ddt); | |
2466 | return (error != 0); | |
2467 | } | |
2468 | } | |
2469 | ||
2470 | return (B_FALSE); | |
2471 | } | |
2472 | ||
2473 | static void | |
2474 | zio_ddt_child_write_ready(zio_t *zio) | |
2475 | { | |
2476 | int p = zio->io_prop.zp_copies; | |
2477 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
2478 | ddt_entry_t *dde = zio->io_private; | |
2479 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
2480 | zio_t *pio; | |
2481 | ||
2482 | if (zio->io_error) | |
2483 | return; | |
2484 | ||
2485 | ddt_enter(ddt); | |
2486 | ||
2487 | ASSERT(dde->dde_lead_zio[p] == zio); | |
2488 | ||
2489 | ddt_phys_fill(ddp, zio->io_bp); | |
2490 | ||
2491 | while ((pio = zio_walk_parents(zio)) != NULL) | |
2492 | ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); | |
2493 | ||
2494 | ddt_exit(ddt); | |
2495 | } | |
2496 | ||
2497 | static void | |
2498 | zio_ddt_child_write_done(zio_t *zio) | |
2499 | { | |
2500 | int p = zio->io_prop.zp_copies; | |
2501 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
2502 | ddt_entry_t *dde = zio->io_private; | |
2503 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
2504 | ||
2505 | ddt_enter(ddt); | |
2506 | ||
2507 | ASSERT(ddp->ddp_refcnt == 0); | |
2508 | ASSERT(dde->dde_lead_zio[p] == zio); | |
2509 | dde->dde_lead_zio[p] = NULL; | |
2510 | ||
2511 | if (zio->io_error == 0) { | |
2512 | while (zio_walk_parents(zio) != NULL) | |
2513 | ddt_phys_addref(ddp); | |
2514 | } else { | |
2515 | ddt_phys_clear(ddp); | |
2516 | } | |
2517 | ||
2518 | ddt_exit(ddt); | |
2519 | } | |
2520 | ||
2521 | static void | |
2522 | zio_ddt_ditto_write_done(zio_t *zio) | |
2523 | { | |
2524 | int p = DDT_PHYS_DITTO; | |
428870ff BB |
2525 | blkptr_t *bp = zio->io_bp; |
2526 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
2527 | ddt_entry_t *dde = zio->io_private; | |
2528 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
2529 | ddt_key_t *ddk = &dde->dde_key; | |
1fde1e37 | 2530 | ASSERTV(zio_prop_t *zp = &zio->io_prop); |
428870ff BB |
2531 | |
2532 | ddt_enter(ddt); | |
2533 | ||
2534 | ASSERT(ddp->ddp_refcnt == 0); | |
2535 | ASSERT(dde->dde_lead_zio[p] == zio); | |
2536 | dde->dde_lead_zio[p] = NULL; | |
2537 | ||
2538 | if (zio->io_error == 0) { | |
2539 | ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); | |
2540 | ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); | |
2541 | ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); | |
2542 | if (ddp->ddp_phys_birth != 0) | |
2543 | ddt_phys_free(ddt, ddk, ddp, zio->io_txg); | |
2544 | ddt_phys_fill(ddp, bp); | |
2545 | } | |
2546 | ||
2547 | ddt_exit(ddt); | |
2548 | } | |
2549 | ||
2550 | static int | |
2551 | zio_ddt_write(zio_t *zio) | |
2552 | { | |
2553 | spa_t *spa = zio->io_spa; | |
2554 | blkptr_t *bp = zio->io_bp; | |
2555 | uint64_t txg = zio->io_txg; | |
2556 | zio_prop_t *zp = &zio->io_prop; | |
2557 | int p = zp->zp_copies; | |
2558 | int ditto_copies; | |
2559 | zio_t *cio = NULL; | |
2560 | zio_t *dio = NULL; | |
2561 | ddt_t *ddt = ddt_select(spa, bp); | |
2562 | ddt_entry_t *dde; | |
2563 | ddt_phys_t *ddp; | |
2564 | ||
2565 | ASSERT(BP_GET_DEDUP(bp)); | |
2566 | ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); | |
2567 | ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); | |
c17bcf83 | 2568 | ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW))); |
428870ff BB |
2569 | |
2570 | ddt_enter(ddt); | |
2571 | dde = ddt_lookup(ddt, bp, B_TRUE); | |
2572 | ddp = &dde->dde_phys[p]; | |
2573 | ||
2574 | if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { | |
2575 | /* | |
2576 | * If we're using a weak checksum, upgrade to a strong checksum | |
2577 | * and try again. If we're already using a strong checksum, | |
2578 | * we can't resolve it, so just convert to an ordinary write. | |
2579 | * (And automatically e-mail a paper to Nature?) | |
2580 | */ | |
3c67d83a TH |
2581 | if (!(zio_checksum_table[zp->zp_checksum].ci_flags & |
2582 | ZCHECKSUM_FLAG_DEDUP)) { | |
428870ff BB |
2583 | zp->zp_checksum = spa_dedup_checksum(spa); |
2584 | zio_pop_transforms(zio); | |
2585 | zio->io_stage = ZIO_STAGE_OPEN; | |
2586 | BP_ZERO(bp); | |
2587 | } else { | |
03c6040b | 2588 | zp->zp_dedup = B_FALSE; |
428870ff BB |
2589 | } |
2590 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
2591 | ddt_exit(ddt); | |
2592 | return (ZIO_PIPELINE_CONTINUE); | |
2593 | } | |
2594 | ||
2595 | ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); | |
2596 | ASSERT(ditto_copies < SPA_DVAS_PER_BP); | |
2597 | ||
2598 | if (ditto_copies > ddt_ditto_copies_present(dde) && | |
2599 | dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { | |
2600 | zio_prop_t czp = *zp; | |
2601 | ||
2602 | czp.zp_copies = ditto_copies; | |
2603 | ||
2604 | /* | |
2605 | * If we arrived here with an override bp, we won't have run | |
2606 | * the transform stack, so we won't have the data we need to | |
2607 | * generate a child i/o. So, toss the override bp and restart. | |
2608 | * This is safe, because using the override bp is just an | |
2609 | * optimization; and it's rare, so the cost doesn't matter. | |
2610 | */ | |
2611 | if (zio->io_bp_override) { | |
2612 | zio_pop_transforms(zio); | |
2613 | zio->io_stage = ZIO_STAGE_OPEN; | |
2614 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
2615 | zio->io_bp_override = NULL; | |
2616 | BP_ZERO(bp); | |
2617 | ddt_exit(ddt); | |
2618 | return (ZIO_PIPELINE_CONTINUE); | |
2619 | } | |
2620 | ||
2621 | dio = zio_write(zio, spa, txg, bp, zio->io_orig_data, | |
2aa34383 | 2622 | zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL, |
bc77ba73 | 2623 | NULL, zio_ddt_ditto_write_done, dde, zio->io_priority, |
428870ff BB |
2624 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); |
2625 | ||
2626 | zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL); | |
2627 | dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; | |
2628 | } | |
2629 | ||
2630 | if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { | |
2631 | if (ddp->ddp_phys_birth != 0) | |
2632 | ddt_bp_fill(ddp, bp, txg); | |
2633 | if (dde->dde_lead_zio[p] != NULL) | |
2634 | zio_add_child(zio, dde->dde_lead_zio[p]); | |
2635 | else | |
2636 | ddt_phys_addref(ddp); | |
2637 | } else if (zio->io_bp_override) { | |
2638 | ASSERT(bp->blk_birth == txg); | |
2639 | ASSERT(BP_EQUAL(bp, zio->io_bp_override)); | |
2640 | ddt_phys_fill(ddp, bp); | |
2641 | ddt_phys_addref(ddp); | |
2642 | } else { | |
2643 | cio = zio_write(zio, spa, txg, bp, zio->io_orig_data, | |
2aa34383 | 2644 | zio->io_orig_size, zio->io_orig_size, zp, |
bc77ba73 | 2645 | zio_ddt_child_write_ready, NULL, NULL, |
428870ff BB |
2646 | zio_ddt_child_write_done, dde, zio->io_priority, |
2647 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); | |
2648 | ||
2649 | zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL); | |
2650 | dde->dde_lead_zio[p] = cio; | |
2651 | } | |
2652 | ||
2653 | ddt_exit(ddt); | |
2654 | ||
2655 | if (cio) | |
2656 | zio_nowait(cio); | |
2657 | if (dio) | |
2658 | zio_nowait(dio); | |
2659 | ||
2660 | return (ZIO_PIPELINE_CONTINUE); | |
2661 | } | |
2662 | ||
2663 | ddt_entry_t *freedde; /* for debugging */ | |
b128c09f | 2664 | |
428870ff BB |
2665 | static int |
2666 | zio_ddt_free(zio_t *zio) | |
2667 | { | |
2668 | spa_t *spa = zio->io_spa; | |
2669 | blkptr_t *bp = zio->io_bp; | |
2670 | ddt_t *ddt = ddt_select(spa, bp); | |
2671 | ddt_entry_t *dde; | |
2672 | ddt_phys_t *ddp; | |
2673 | ||
2674 | ASSERT(BP_GET_DEDUP(bp)); | |
2675 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
2676 | ||
2677 | ddt_enter(ddt); | |
2678 | freedde = dde = ddt_lookup(ddt, bp, B_TRUE); | |
5dc6af0e BB |
2679 | if (dde) { |
2680 | ddp = ddt_phys_select(dde, bp); | |
2681 | if (ddp) | |
2682 | ddt_phys_decref(ddp); | |
2683 | } | |
428870ff BB |
2684 | ddt_exit(ddt); |
2685 | ||
2686 | return (ZIO_PIPELINE_CONTINUE); | |
2687 | } | |
2688 | ||
2689 | /* | |
2690 | * ========================================================================== | |
2691 | * Allocate and free blocks | |
2692 | * ========================================================================== | |
2693 | */ | |
34dc7c2f BB |
2694 | static int |
2695 | zio_dva_allocate(zio_t *zio) | |
2696 | { | |
2697 | spa_t *spa = zio->io_spa; | |
428870ff | 2698 | metaslab_class_t *mc = spa_normal_class(spa); |
34dc7c2f BB |
2699 | blkptr_t *bp = zio->io_bp; |
2700 | int error; | |
6d974228 | 2701 | int flags = 0; |
34dc7c2f | 2702 | |
9babb374 BB |
2703 | if (zio->io_gang_leader == NULL) { |
2704 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
2705 | zio->io_gang_leader = zio; | |
2706 | } | |
2707 | ||
34dc7c2f | 2708 | ASSERT(BP_IS_HOLE(bp)); |
c99c9001 | 2709 | ASSERT0(BP_GET_NDVAS(bp)); |
428870ff BB |
2710 | ASSERT3U(zio->io_prop.zp_copies, >, 0); |
2711 | ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); | |
34dc7c2f BB |
2712 | ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); |
2713 | ||
6d974228 GW |
2714 | /* |
2715 | * The dump device does not support gang blocks so allocation on | |
2716 | * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid | |
2717 | * the "fast" gang feature. | |
2718 | */ | |
2719 | flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0; | |
2720 | flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ? | |
2721 | METASLAB_GANG_CHILD : 0; | |
920dd524 | 2722 | flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0; |
b128c09f | 2723 | error = metaslab_alloc(spa, mc, zio->io_size, bp, |
6d974228 | 2724 | zio->io_prop.zp_copies, zio->io_txg, NULL, flags); |
34dc7c2f | 2725 | |
b128c09f | 2726 | if (error) { |
6d974228 GW |
2727 | spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, " |
2728 | "size %llu, error %d", spa_name(spa), zio, zio->io_size, | |
2729 | error); | |
b128c09f BB |
2730 | if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) |
2731 | return (zio_write_gang_block(zio)); | |
34dc7c2f BB |
2732 | zio->io_error = error; |
2733 | } | |
2734 | ||
2735 | return (ZIO_PIPELINE_CONTINUE); | |
2736 | } | |
2737 | ||
2738 | static int | |
2739 | zio_dva_free(zio_t *zio) | |
2740 | { | |
b128c09f | 2741 | metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); |
34dc7c2f BB |
2742 | |
2743 | return (ZIO_PIPELINE_CONTINUE); | |
2744 | } | |
2745 | ||
2746 | static int | |
2747 | zio_dva_claim(zio_t *zio) | |
2748 | { | |
b128c09f BB |
2749 | int error; |
2750 | ||
2751 | error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); | |
2752 | if (error) | |
2753 | zio->io_error = error; | |
34dc7c2f BB |
2754 | |
2755 | return (ZIO_PIPELINE_CONTINUE); | |
2756 | } | |
2757 | ||
b128c09f BB |
2758 | /* |
2759 | * Undo an allocation. This is used by zio_done() when an I/O fails | |
2760 | * and we want to give back the block we just allocated. | |
2761 | * This handles both normal blocks and gang blocks. | |
2762 | */ | |
2763 | static void | |
2764 | zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) | |
2765 | { | |
d6320ddb BB |
2766 | int g; |
2767 | ||
b128c09f | 2768 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); |
428870ff | 2769 | ASSERT(zio->io_bp_override == NULL); |
b128c09f BB |
2770 | |
2771 | if (!BP_IS_HOLE(bp)) | |
428870ff | 2772 | metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); |
b128c09f BB |
2773 | |
2774 | if (gn != NULL) { | |
d6320ddb | 2775 | for (g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2776 | zio_dva_unallocate(zio, gn->gn_child[g], |
2777 | &gn->gn_gbh->zg_blkptr[g]); | |
2778 | } | |
2779 | } | |
2780 | } | |
2781 | ||
2782 | /* | |
2783 | * Try to allocate an intent log block. Return 0 on success, errno on failure. | |
2784 | */ | |
2785 | int | |
920dd524 ED |
2786 | zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, uint64_t size, |
2787 | boolean_t use_slog) | |
b128c09f | 2788 | { |
428870ff | 2789 | int error = 1; |
b128c09f | 2790 | |
428870ff BB |
2791 | ASSERT(txg > spa_syncing_txg(spa)); |
2792 | ||
ebf8e3a2 BB |
2793 | /* |
2794 | * ZIL blocks are always contiguous (i.e. not gang blocks) so we | |
2795 | * set the METASLAB_GANG_AVOID flag so that they don't "fast gang" | |
2796 | * when allocating them. | |
2797 | */ | |
2798 | if (use_slog) { | |
428870ff | 2799 | error = metaslab_alloc(spa, spa_log_class(spa), size, |
920dd524 ED |
2800 | new_bp, 1, txg, NULL, |
2801 | METASLAB_FASTWRITE | METASLAB_GANG_AVOID); | |
ebf8e3a2 | 2802 | } |
b128c09f | 2803 | |
ebf8e3a2 | 2804 | if (error) { |
428870ff | 2805 | error = metaslab_alloc(spa, spa_normal_class(spa), size, |
920dd524 | 2806 | new_bp, 1, txg, NULL, |
ac72fac3 | 2807 | METASLAB_FASTWRITE); |
ebf8e3a2 | 2808 | } |
b128c09f BB |
2809 | |
2810 | if (error == 0) { | |
2811 | BP_SET_LSIZE(new_bp, size); | |
2812 | BP_SET_PSIZE(new_bp, size); | |
2813 | BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); | |
428870ff BB |
2814 | BP_SET_CHECKSUM(new_bp, |
2815 | spa_version(spa) >= SPA_VERSION_SLIM_ZIL | |
2816 | ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); | |
b128c09f BB |
2817 | BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); |
2818 | BP_SET_LEVEL(new_bp, 0); | |
428870ff | 2819 | BP_SET_DEDUP(new_bp, 0); |
b128c09f BB |
2820 | BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); |
2821 | } | |
2822 | ||
2823 | return (error); | |
2824 | } | |
2825 | ||
2826 | /* | |
428870ff | 2827 | * Free an intent log block. |
b128c09f BB |
2828 | */ |
2829 | void | |
428870ff | 2830 | zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp) |
b128c09f | 2831 | { |
428870ff | 2832 | ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG); |
b128c09f BB |
2833 | ASSERT(!BP_IS_GANG(bp)); |
2834 | ||
428870ff | 2835 | zio_free(spa, txg, bp); |
b128c09f BB |
2836 | } |
2837 | ||
34dc7c2f BB |
2838 | /* |
2839 | * ========================================================================== | |
2840 | * Read and write to physical devices | |
2841 | * ========================================================================== | |
2842 | */ | |
98b25418 GW |
2843 | |
2844 | ||
2845 | /* | |
2846 | * Issue an I/O to the underlying vdev. Typically the issue pipeline | |
2847 | * stops after this stage and will resume upon I/O completion. | |
2848 | * However, there are instances where the vdev layer may need to | |
2849 | * continue the pipeline when an I/O was not issued. Since the I/O | |
2850 | * that was sent to the vdev layer might be different than the one | |
2851 | * currently active in the pipeline (see vdev_queue_io()), we explicitly | |
2852 | * force the underlying vdev layers to call either zio_execute() or | |
2853 | * zio_interrupt() to ensure that the pipeline continues with the correct I/O. | |
2854 | */ | |
34dc7c2f BB |
2855 | static int |
2856 | zio_vdev_io_start(zio_t *zio) | |
2857 | { | |
2858 | vdev_t *vd = zio->io_vd; | |
34dc7c2f BB |
2859 | uint64_t align; |
2860 | spa_t *spa = zio->io_spa; | |
2861 | ||
193a37cb TH |
2862 | zio->io_delay = 0; |
2863 | ||
b128c09f BB |
2864 | ASSERT(zio->io_error == 0); |
2865 | ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); | |
34dc7c2f | 2866 | |
b128c09f BB |
2867 | if (vd == NULL) { |
2868 | if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
2869 | spa_config_enter(spa, SCL_ZIO, zio, RW_READER); | |
34dc7c2f | 2870 | |
b128c09f BB |
2871 | /* |
2872 | * The mirror_ops handle multiple DVAs in a single BP. | |
2873 | */ | |
98b25418 GW |
2874 | vdev_mirror_ops.vdev_op_io_start(zio); |
2875 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f BB |
2876 | } |
2877 | ||
572e2857 BB |
2878 | /* |
2879 | * We keep track of time-sensitive I/Os so that the scan thread | |
2880 | * can quickly react to certain workloads. In particular, we care | |
2881 | * about non-scrubbing, top-level reads and writes with the following | |
2882 | * characteristics: | |
98b25418 | 2883 | * - synchronous writes of user data to non-slog devices |
572e2857 BB |
2884 | * - any reads of user data |
2885 | * When these conditions are met, adjust the timestamp of spa_last_io | |
2886 | * which allows the scan thread to adjust its workload accordingly. | |
2887 | */ | |
2888 | if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL && | |
2889 | vd == vd->vdev_top && !vd->vdev_islog && | |
2890 | zio->io_bookmark.zb_objset != DMU_META_OBJSET && | |
2891 | zio->io_txg != spa_syncing_txg(spa)) { | |
2892 | uint64_t old = spa->spa_last_io; | |
2893 | uint64_t new = ddi_get_lbolt64(); | |
2894 | if (old != new) | |
2895 | (void) atomic_cas_64(&spa->spa_last_io, old, new); | |
2896 | } | |
2897 | ||
b128c09f BB |
2898 | align = 1ULL << vd->vdev_top->vdev_ashift; |
2899 | ||
b02fe35d AR |
2900 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && |
2901 | P2PHASE(zio->io_size, align) != 0) { | |
2902 | /* Transform logical writes to be a full physical block size. */ | |
34dc7c2f BB |
2903 | uint64_t asize = P2ROUNDUP(zio->io_size, align); |
2904 | char *abuf = zio_buf_alloc(asize); | |
178e73b3 | 2905 | ASSERT(vd == vd->vdev_top); |
34dc7c2f BB |
2906 | if (zio->io_type == ZIO_TYPE_WRITE) { |
2907 | bcopy(zio->io_data, abuf, zio->io_size); | |
2908 | bzero(abuf + zio->io_size, asize - zio->io_size); | |
2909 | } | |
b128c09f | 2910 | zio_push_transform(zio, abuf, asize, asize, zio_subblock); |
34dc7c2f BB |
2911 | } |
2912 | ||
b02fe35d AR |
2913 | /* |
2914 | * If this is not a physical io, make sure that it is properly aligned | |
2915 | * before proceeding. | |
2916 | */ | |
2917 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { | |
2918 | ASSERT0(P2PHASE(zio->io_offset, align)); | |
2919 | ASSERT0(P2PHASE(zio->io_size, align)); | |
2920 | } else { | |
2921 | /* | |
2922 | * For physical writes, we allow 512b aligned writes and assume | |
2923 | * the device will perform a read-modify-write as necessary. | |
2924 | */ | |
2925 | ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); | |
2926 | ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); | |
2927 | } | |
2928 | ||
572e2857 | 2929 | VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); |
fb5f0bc8 BB |
2930 | |
2931 | /* | |
2932 | * If this is a repair I/O, and there's no self-healing involved -- | |
2933 | * that is, we're just resilvering what we expect to resilver -- | |
2934 | * then don't do the I/O unless zio's txg is actually in vd's DTL. | |
2935 | * This prevents spurious resilvering with nested replication. | |
2936 | * For example, given a mirror of mirrors, (A+B)+(C+D), if only | |
2937 | * A is out of date, we'll read from C+D, then use the data to | |
2938 | * resilver A+B -- but we don't actually want to resilver B, just A. | |
2939 | * The top-level mirror has no way to know this, so instead we just | |
2940 | * discard unnecessary repairs as we work our way down the vdev tree. | |
2941 | * The same logic applies to any form of nested replication: | |
2942 | * ditto + mirror, RAID-Z + replacing, etc. This covers them all. | |
2943 | */ | |
2944 | if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && | |
2945 | !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && | |
2946 | zio->io_txg != 0 && /* not a delegated i/o */ | |
2947 | !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { | |
2948 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
fb5f0bc8 BB |
2949 | zio_vdev_io_bypass(zio); |
2950 | return (ZIO_PIPELINE_CONTINUE); | |
2951 | } | |
34dc7c2f | 2952 | |
b128c09f BB |
2953 | if (vd->vdev_ops->vdev_op_leaf && |
2954 | (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) { | |
2955 | ||
b0bc7a84 | 2956 | if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio)) |
d164b209 | 2957 | return (ZIO_PIPELINE_CONTINUE); |
b128c09f BB |
2958 | |
2959 | if ((zio = vdev_queue_io(zio)) == NULL) | |
2960 | return (ZIO_PIPELINE_STOP); | |
2961 | ||
2962 | if (!vdev_accessible(vd, zio)) { | |
2e528b49 | 2963 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
2964 | zio_interrupt(zio); |
2965 | return (ZIO_PIPELINE_STOP); | |
2966 | } | |
b128c09f BB |
2967 | } |
2968 | ||
193a37cb | 2969 | zio->io_delay = gethrtime(); |
98b25418 GW |
2970 | vd->vdev_ops->vdev_op_io_start(zio); |
2971 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f BB |
2972 | } |
2973 | ||
2974 | static int | |
2975 | zio_vdev_io_done(zio_t *zio) | |
2976 | { | |
b128c09f BB |
2977 | vdev_t *vd = zio->io_vd; |
2978 | vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; | |
2979 | boolean_t unexpected_error = B_FALSE; | |
34dc7c2f | 2980 | |
b128c09f BB |
2981 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) |
2982 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f | 2983 | |
b128c09f BB |
2984 | ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); |
2985 | ||
193a37cb TH |
2986 | if (zio->io_delay) |
2987 | zio->io_delay = gethrtime() - zio->io_delay; | |
2988 | ||
b128c09f BB |
2989 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { |
2990 | ||
2991 | vdev_queue_io_done(zio); | |
2992 | ||
2993 | if (zio->io_type == ZIO_TYPE_WRITE) | |
2994 | vdev_cache_write(zio); | |
2995 | ||
2996 | if (zio_injection_enabled && zio->io_error == 0) | |
9babb374 BB |
2997 | zio->io_error = zio_handle_device_injection(vd, |
2998 | zio, EIO); | |
b128c09f BB |
2999 | |
3000 | if (zio_injection_enabled && zio->io_error == 0) | |
3001 | zio->io_error = zio_handle_label_injection(zio, EIO); | |
3002 | ||
3003 | if (zio->io_error) { | |
3004 | if (!vdev_accessible(vd, zio)) { | |
2e528b49 | 3005 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
3006 | } else { |
3007 | unexpected_error = B_TRUE; | |
3008 | } | |
3009 | } | |
3010 | } | |
3011 | ||
3012 | ops->vdev_op_io_done(zio); | |
34dc7c2f | 3013 | |
b128c09f | 3014 | if (unexpected_error) |
d164b209 | 3015 | VERIFY(vdev_probe(vd, zio) == NULL); |
34dc7c2f | 3016 | |
b128c09f | 3017 | return (ZIO_PIPELINE_CONTINUE); |
34dc7c2f BB |
3018 | } |
3019 | ||
428870ff BB |
3020 | /* |
3021 | * For non-raidz ZIOs, we can just copy aside the bad data read from the | |
3022 | * disk, and use that to finish the checksum ereport later. | |
3023 | */ | |
3024 | static void | |
3025 | zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, | |
3026 | const void *good_buf) | |
3027 | { | |
3028 | /* no processing needed */ | |
3029 | zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); | |
3030 | } | |
3031 | ||
3032 | /*ARGSUSED*/ | |
3033 | void | |
3034 | zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) | |
3035 | { | |
3036 | void *buf = zio_buf_alloc(zio->io_size); | |
3037 | ||
3038 | bcopy(zio->io_data, buf, zio->io_size); | |
3039 | ||
3040 | zcr->zcr_cbinfo = zio->io_size; | |
3041 | zcr->zcr_cbdata = buf; | |
3042 | zcr->zcr_finish = zio_vsd_default_cksum_finish; | |
3043 | zcr->zcr_free = zio_buf_free; | |
3044 | } | |
3045 | ||
34dc7c2f BB |
3046 | static int |
3047 | zio_vdev_io_assess(zio_t *zio) | |
3048 | { | |
3049 | vdev_t *vd = zio->io_vd; | |
b128c09f BB |
3050 | |
3051 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) | |
3052 | return (ZIO_PIPELINE_STOP); | |
3053 | ||
3054 | if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
3055 | spa_config_exit(zio->io_spa, SCL_ZIO, zio); | |
3056 | ||
3057 | if (zio->io_vsd != NULL) { | |
428870ff | 3058 | zio->io_vsd_ops->vsd_free(zio); |
b128c09f | 3059 | zio->io_vsd = NULL; |
34dc7c2f BB |
3060 | } |
3061 | ||
b128c09f | 3062 | if (zio_injection_enabled && zio->io_error == 0) |
34dc7c2f BB |
3063 | zio->io_error = zio_handle_fault_injection(zio, EIO); |
3064 | ||
3065 | /* | |
3066 | * If the I/O failed, determine whether we should attempt to retry it. | |
428870ff BB |
3067 | * |
3068 | * On retry, we cut in line in the issue queue, since we don't want | |
3069 | * compression/checksumming/etc. work to prevent our (cheap) IO reissue. | |
34dc7c2f | 3070 | */ |
b128c09f BB |
3071 | if (zio->io_error && vd == NULL && |
3072 | !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { | |
3073 | ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ | |
3074 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ | |
34dc7c2f | 3075 | zio->io_error = 0; |
b128c09f BB |
3076 | zio->io_flags |= ZIO_FLAG_IO_RETRY | |
3077 | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; | |
428870ff BB |
3078 | zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; |
3079 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, | |
3080 | zio_requeue_io_start_cut_in_line); | |
b128c09f | 3081 | return (ZIO_PIPELINE_STOP); |
34dc7c2f BB |
3082 | } |
3083 | ||
b128c09f BB |
3084 | /* |
3085 | * If we got an error on a leaf device, convert it to ENXIO | |
3086 | * if the device is not accessible at all. | |
3087 | */ | |
3088 | if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && | |
3089 | !vdev_accessible(vd, zio)) | |
2e528b49 | 3090 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
3091 | |
3092 | /* | |
3093 | * If we can't write to an interior vdev (mirror or RAID-Z), | |
3094 | * set vdev_cant_write so that we stop trying to allocate from it. | |
3095 | */ | |
3096 | if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && | |
13fe0198 | 3097 | vd != NULL && !vd->vdev_ops->vdev_op_leaf) { |
b128c09f | 3098 | vd->vdev_cant_write = B_TRUE; |
13fe0198 | 3099 | } |
b128c09f BB |
3100 | |
3101 | if (zio->io_error) | |
3102 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
3103 | ||
e8b96c60 MA |
3104 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf && |
3105 | zio->io_physdone != NULL) { | |
3106 | ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); | |
3107 | ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); | |
3108 | zio->io_physdone(zio->io_logical); | |
3109 | } | |
3110 | ||
34dc7c2f BB |
3111 | return (ZIO_PIPELINE_CONTINUE); |
3112 | } | |
3113 | ||
3114 | void | |
3115 | zio_vdev_io_reissue(zio_t *zio) | |
3116 | { | |
3117 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
3118 | ASSERT(zio->io_error == 0); | |
3119 | ||
428870ff | 3120 | zio->io_stage >>= 1; |
34dc7c2f BB |
3121 | } |
3122 | ||
3123 | void | |
3124 | zio_vdev_io_redone(zio_t *zio) | |
3125 | { | |
3126 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); | |
3127 | ||
428870ff | 3128 | zio->io_stage >>= 1; |
34dc7c2f BB |
3129 | } |
3130 | ||
3131 | void | |
3132 | zio_vdev_io_bypass(zio_t *zio) | |
3133 | { | |
3134 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
3135 | ASSERT(zio->io_error == 0); | |
3136 | ||
3137 | zio->io_flags |= ZIO_FLAG_IO_BYPASS; | |
428870ff | 3138 | zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; |
34dc7c2f BB |
3139 | } |
3140 | ||
3141 | /* | |
3142 | * ========================================================================== | |
3143 | * Generate and verify checksums | |
3144 | * ========================================================================== | |
3145 | */ | |
3146 | static int | |
3147 | zio_checksum_generate(zio_t *zio) | |
3148 | { | |
34dc7c2f | 3149 | blkptr_t *bp = zio->io_bp; |
b128c09f | 3150 | enum zio_checksum checksum; |
34dc7c2f | 3151 | |
b128c09f BB |
3152 | if (bp == NULL) { |
3153 | /* | |
3154 | * This is zio_write_phys(). | |
3155 | * We're either generating a label checksum, or none at all. | |
3156 | */ | |
3157 | checksum = zio->io_prop.zp_checksum; | |
34dc7c2f | 3158 | |
b128c09f BB |
3159 | if (checksum == ZIO_CHECKSUM_OFF) |
3160 | return (ZIO_PIPELINE_CONTINUE); | |
3161 | ||
3162 | ASSERT(checksum == ZIO_CHECKSUM_LABEL); | |
3163 | } else { | |
3164 | if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { | |
3165 | ASSERT(!IO_IS_ALLOCATING(zio)); | |
3166 | checksum = ZIO_CHECKSUM_GANG_HEADER; | |
3167 | } else { | |
3168 | checksum = BP_GET_CHECKSUM(bp); | |
3169 | } | |
3170 | } | |
34dc7c2f | 3171 | |
b128c09f | 3172 | zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size); |
34dc7c2f BB |
3173 | |
3174 | return (ZIO_PIPELINE_CONTINUE); | |
3175 | } | |
3176 | ||
3177 | static int | |
b128c09f | 3178 | zio_checksum_verify(zio_t *zio) |
34dc7c2f | 3179 | { |
428870ff | 3180 | zio_bad_cksum_t info; |
b128c09f BB |
3181 | blkptr_t *bp = zio->io_bp; |
3182 | int error; | |
34dc7c2f | 3183 | |
428870ff BB |
3184 | ASSERT(zio->io_vd != NULL); |
3185 | ||
b128c09f BB |
3186 | if (bp == NULL) { |
3187 | /* | |
3188 | * This is zio_read_phys(). | |
3189 | * We're either verifying a label checksum, or nothing at all. | |
3190 | */ | |
3191 | if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) | |
3192 | return (ZIO_PIPELINE_CONTINUE); | |
34dc7c2f | 3193 | |
b128c09f BB |
3194 | ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); |
3195 | } | |
34dc7c2f | 3196 | |
428870ff | 3197 | if ((error = zio_checksum_error(zio, &info)) != 0) { |
b128c09f | 3198 | zio->io_error = error; |
7a3066ff MA |
3199 | if (error == ECKSUM && |
3200 | !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { | |
428870ff BB |
3201 | zfs_ereport_start_checksum(zio->io_spa, |
3202 | zio->io_vd, zio, zio->io_offset, | |
3203 | zio->io_size, NULL, &info); | |
b128c09f | 3204 | } |
34dc7c2f BB |
3205 | } |
3206 | ||
3207 | return (ZIO_PIPELINE_CONTINUE); | |
3208 | } | |
3209 | ||
3210 | /* | |
3211 | * Called by RAID-Z to ensure we don't compute the checksum twice. | |
3212 | */ | |
3213 | void | |
3214 | zio_checksum_verified(zio_t *zio) | |
3215 | { | |
428870ff | 3216 | zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; |
34dc7c2f BB |
3217 | } |
3218 | ||
3219 | /* | |
b128c09f BB |
3220 | * ========================================================================== |
3221 | * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. | |
9b67f605 | 3222 | * An error of 0 indicates success. ENXIO indicates whole-device failure, |
b128c09f BB |
3223 | * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO |
3224 | * indicate errors that are specific to one I/O, and most likely permanent. | |
3225 | * Any other error is presumed to be worse because we weren't expecting it. | |
3226 | * ========================================================================== | |
34dc7c2f | 3227 | */ |
b128c09f BB |
3228 | int |
3229 | zio_worst_error(int e1, int e2) | |
34dc7c2f | 3230 | { |
b128c09f BB |
3231 | static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; |
3232 | int r1, r2; | |
3233 | ||
3234 | for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) | |
3235 | if (e1 == zio_error_rank[r1]) | |
3236 | break; | |
34dc7c2f | 3237 | |
b128c09f BB |
3238 | for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) |
3239 | if (e2 == zio_error_rank[r2]) | |
3240 | break; | |
3241 | ||
3242 | return (r1 > r2 ? e1 : e2); | |
34dc7c2f BB |
3243 | } |
3244 | ||
3245 | /* | |
3246 | * ========================================================================== | |
b128c09f | 3247 | * I/O completion |
34dc7c2f BB |
3248 | * ========================================================================== |
3249 | */ | |
b128c09f BB |
3250 | static int |
3251 | zio_ready(zio_t *zio) | |
34dc7c2f | 3252 | { |
b128c09f | 3253 | blkptr_t *bp = zio->io_bp; |
d164b209 | 3254 | zio_t *pio, *pio_next; |
34dc7c2f | 3255 | |
428870ff BB |
3256 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || |
3257 | zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY)) | |
9babb374 | 3258 | return (ZIO_PIPELINE_STOP); |
34dc7c2f | 3259 | |
9babb374 | 3260 | if (zio->io_ready) { |
b128c09f | 3261 | ASSERT(IO_IS_ALLOCATING(zio)); |
03c6040b GW |
3262 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || |
3263 | (zio->io_flags & ZIO_FLAG_NOPWRITE)); | |
b128c09f | 3264 | ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); |
34dc7c2f | 3265 | |
b128c09f BB |
3266 | zio->io_ready(zio); |
3267 | } | |
34dc7c2f | 3268 | |
b128c09f BB |
3269 | if (bp != NULL && bp != &zio->io_bp_copy) |
3270 | zio->io_bp_copy = *bp; | |
34dc7c2f | 3271 | |
b128c09f BB |
3272 | if (zio->io_error) |
3273 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
34dc7c2f | 3274 | |
d164b209 BB |
3275 | mutex_enter(&zio->io_lock); |
3276 | zio->io_state[ZIO_WAIT_READY] = 1; | |
3277 | pio = zio_walk_parents(zio); | |
3278 | mutex_exit(&zio->io_lock); | |
3279 | ||
3280 | /* | |
3281 | * As we notify zio's parents, new parents could be added. | |
3282 | * New parents go to the head of zio's io_parent_list, however, | |
3283 | * so we will (correctly) not notify them. The remainder of zio's | |
3284 | * io_parent_list, from 'pio_next' onward, cannot change because | |
3285 | * all parents must wait for us to be done before they can be done. | |
3286 | */ | |
3287 | for (; pio != NULL; pio = pio_next) { | |
3288 | pio_next = zio_walk_parents(zio); | |
b128c09f | 3289 | zio_notify_parent(pio, zio, ZIO_WAIT_READY); |
d164b209 | 3290 | } |
34dc7c2f | 3291 | |
428870ff BB |
3292 | if (zio->io_flags & ZIO_FLAG_NODATA) { |
3293 | if (BP_IS_GANG(bp)) { | |
3294 | zio->io_flags &= ~ZIO_FLAG_NODATA; | |
3295 | } else { | |
3296 | ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE); | |
3297 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; | |
3298 | } | |
3299 | } | |
3300 | ||
3301 | if (zio_injection_enabled && | |
3302 | zio->io_spa->spa_syncing_txg == zio->io_txg) | |
3303 | zio_handle_ignored_writes(zio); | |
3304 | ||
b128c09f | 3305 | return (ZIO_PIPELINE_CONTINUE); |
34dc7c2f BB |
3306 | } |
3307 | ||
b128c09f BB |
3308 | static int |
3309 | zio_done(zio_t *zio) | |
34dc7c2f | 3310 | { |
d164b209 | 3311 | zio_t *pio, *pio_next; |
d6320ddb | 3312 | int c, w; |
34dc7c2f | 3313 | |
b128c09f | 3314 | /* |
9babb374 | 3315 | * If our children haven't all completed, |
b128c09f BB |
3316 | * wait for them and then repeat this pipeline stage. |
3317 | */ | |
3318 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || | |
3319 | zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || | |
428870ff | 3320 | zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) || |
b128c09f BB |
3321 | zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) |
3322 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f | 3323 | |
d6320ddb BB |
3324 | for (c = 0; c < ZIO_CHILD_TYPES; c++) |
3325 | for (w = 0; w < ZIO_WAIT_TYPES; w++) | |
b128c09f BB |
3326 | ASSERT(zio->io_children[c][w] == 0); |
3327 | ||
9b67f605 | 3328 | if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) { |
c776b317 BB |
3329 | ASSERT(zio->io_bp->blk_pad[0] == 0); |
3330 | ASSERT(zio->io_bp->blk_pad[1] == 0); | |
d1d7e268 MK |
3331 | ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, |
3332 | sizeof (blkptr_t)) == 0 || | |
c776b317 BB |
3333 | (zio->io_bp == zio_unique_parent(zio)->io_bp)); |
3334 | if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) && | |
428870ff | 3335 | zio->io_bp_override == NULL && |
b128c09f | 3336 | !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { |
c776b317 | 3337 | ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp)); |
d1d7e268 MK |
3338 | ASSERT3U(zio->io_prop.zp_copies, <=, |
3339 | BP_GET_NDVAS(zio->io_bp)); | |
c776b317 | 3340 | ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 || |
d1d7e268 MK |
3341 | (BP_COUNT_GANG(zio->io_bp) == |
3342 | BP_GET_NDVAS(zio->io_bp))); | |
b128c09f | 3343 | } |
03c6040b GW |
3344 | if (zio->io_flags & ZIO_FLAG_NOPWRITE) |
3345 | VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig)); | |
b128c09f BB |
3346 | } |
3347 | ||
3348 | /* | |
428870ff | 3349 | * If there were child vdev/gang/ddt errors, they apply to us now. |
b128c09f BB |
3350 | */ |
3351 | zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); | |
3352 | zio_inherit_child_errors(zio, ZIO_CHILD_GANG); | |
428870ff BB |
3353 | zio_inherit_child_errors(zio, ZIO_CHILD_DDT); |
3354 | ||
3355 | /* | |
3356 | * If the I/O on the transformed data was successful, generate any | |
3357 | * checksum reports now while we still have the transformed data. | |
3358 | */ | |
3359 | if (zio->io_error == 0) { | |
3360 | while (zio->io_cksum_report != NULL) { | |
3361 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
3362 | uint64_t align = zcr->zcr_align; | |
c776b317 | 3363 | uint64_t asize = P2ROUNDUP(zio->io_size, align); |
428870ff BB |
3364 | char *abuf = zio->io_data; |
3365 | ||
c776b317 | 3366 | if (asize != zio->io_size) { |
428870ff | 3367 | abuf = zio_buf_alloc(asize); |
c776b317 | 3368 | bcopy(zio->io_data, abuf, zio->io_size); |
d1d7e268 | 3369 | bzero(abuf+zio->io_size, asize-zio->io_size); |
428870ff BB |
3370 | } |
3371 | ||
3372 | zio->io_cksum_report = zcr->zcr_next; | |
3373 | zcr->zcr_next = NULL; | |
3374 | zcr->zcr_finish(zcr, abuf); | |
3375 | zfs_ereport_free_checksum(zcr); | |
3376 | ||
c776b317 | 3377 | if (asize != zio->io_size) |
428870ff BB |
3378 | zio_buf_free(abuf, asize); |
3379 | } | |
3380 | } | |
b128c09f BB |
3381 | |
3382 | zio_pop_transforms(zio); /* note: may set zio->io_error */ | |
3383 | ||
c776b317 | 3384 | vdev_stat_update(zio, zio->io_size); |
b128c09f | 3385 | |
a69052be | 3386 | /* |
cc92e9d0 | 3387 | * If this I/O is attached to a particular vdev is slow, exceeding |
72f53c56 MJ |
3388 | * 30 seconds to complete, post an error described the I/O delay. |
3389 | * We ignore these errors if the device is currently unavailable. | |
a69052be | 3390 | */ |
193a37cb | 3391 | if (zio->io_delay >= MSEC2NSEC(zio_delay_max)) { |
72f53c56 MJ |
3392 | if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) |
3393 | zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa, | |
d1d7e268 | 3394 | zio->io_vd, zio, 0, 0); |
72f53c56 | 3395 | } |
a69052be | 3396 | |
b128c09f BB |
3397 | if (zio->io_error) { |
3398 | /* | |
3399 | * If this I/O is attached to a particular vdev, | |
3400 | * generate an error message describing the I/O failure | |
3401 | * at the block level. We ignore these errors if the | |
3402 | * device is currently unavailable. | |
3403 | */ | |
c776b317 BB |
3404 | if (zio->io_error != ECKSUM && zio->io_vd != NULL && |
3405 | !vdev_is_dead(zio->io_vd)) | |
3406 | zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa, | |
3407 | zio->io_vd, zio, 0, 0); | |
34dc7c2f | 3408 | |
428870ff BB |
3409 | if ((zio->io_error == EIO || !(zio->io_flags & |
3410 | (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && | |
c776b317 | 3411 | zio == zio->io_logical) { |
b128c09f BB |
3412 | /* |
3413 | * For logical I/O requests, tell the SPA to log the | |
3414 | * error and generate a logical data ereport. | |
3415 | */ | |
c776b317 | 3416 | spa_log_error(zio->io_spa, zio); |
d1d7e268 MK |
3417 | zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, |
3418 | NULL, zio, 0, 0); | |
b128c09f BB |
3419 | } |
3420 | } | |
34dc7c2f | 3421 | |
c776b317 | 3422 | if (zio->io_error && zio == zio->io_logical) { |
b128c09f BB |
3423 | /* |
3424 | * Determine whether zio should be reexecuted. This will | |
3425 | * propagate all the way to the root via zio_notify_parent(). | |
3426 | */ | |
c776b317 | 3427 | ASSERT(zio->io_vd == NULL && zio->io_bp != NULL); |
428870ff | 3428 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); |
b128c09f | 3429 | |
428870ff BB |
3430 | if (IO_IS_ALLOCATING(zio) && |
3431 | !(zio->io_flags & ZIO_FLAG_CANFAIL)) { | |
b128c09f BB |
3432 | if (zio->io_error != ENOSPC) |
3433 | zio->io_reexecute |= ZIO_REEXECUTE_NOW; | |
3434 | else | |
3435 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff | 3436 | } |
b128c09f BB |
3437 | |
3438 | if ((zio->io_type == ZIO_TYPE_READ || | |
3439 | zio->io_type == ZIO_TYPE_FREE) && | |
572e2857 | 3440 | !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && |
b128c09f | 3441 | zio->io_error == ENXIO && |
c776b317 BB |
3442 | spa_load_state(zio->io_spa) == SPA_LOAD_NONE && |
3443 | spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE) | |
b128c09f BB |
3444 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; |
3445 | ||
3446 | if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) | |
3447 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff BB |
3448 | |
3449 | /* | |
3450 | * Here is a possibly good place to attempt to do | |
3451 | * either combinatorial reconstruction or error correction | |
3452 | * based on checksums. It also might be a good place | |
3453 | * to send out preliminary ereports before we suspend | |
3454 | * processing. | |
3455 | */ | |
34dc7c2f BB |
3456 | } |
3457 | ||
3458 | /* | |
b128c09f BB |
3459 | * If there were logical child errors, they apply to us now. |
3460 | * We defer this until now to avoid conflating logical child | |
3461 | * errors with errors that happened to the zio itself when | |
3462 | * updating vdev stats and reporting FMA events above. | |
34dc7c2f | 3463 | */ |
b128c09f | 3464 | zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); |
34dc7c2f | 3465 | |
428870ff BB |
3466 | if ((zio->io_error || zio->io_reexecute) && |
3467 | IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && | |
03c6040b | 3468 | !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) |
c776b317 | 3469 | zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp); |
9babb374 BB |
3470 | |
3471 | zio_gang_tree_free(&zio->io_gang_tree); | |
3472 | ||
3473 | /* | |
3474 | * Godfather I/Os should never suspend. | |
3475 | */ | |
3476 | if ((zio->io_flags & ZIO_FLAG_GODFATHER) && | |
3477 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) | |
3478 | zio->io_reexecute = 0; | |
3479 | ||
b128c09f BB |
3480 | if (zio->io_reexecute) { |
3481 | /* | |
3482 | * This is a logical I/O that wants to reexecute. | |
3483 | * | |
3484 | * Reexecute is top-down. When an i/o fails, if it's not | |
3485 | * the root, it simply notifies its parent and sticks around. | |
3486 | * The parent, seeing that it still has children in zio_done(), | |
3487 | * does the same. This percolates all the way up to the root. | |
3488 | * The root i/o will reexecute or suspend the entire tree. | |
3489 | * | |
3490 | * This approach ensures that zio_reexecute() honors | |
3491 | * all the original i/o dependency relationships, e.g. | |
3492 | * parents not executing until children are ready. | |
3493 | */ | |
3494 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
34dc7c2f | 3495 | |
9babb374 | 3496 | zio->io_gang_leader = NULL; |
b128c09f | 3497 | |
d164b209 BB |
3498 | mutex_enter(&zio->io_lock); |
3499 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
3500 | mutex_exit(&zio->io_lock); | |
3501 | ||
9babb374 BB |
3502 | /* |
3503 | * "The Godfather" I/O monitors its children but is | |
3504 | * not a true parent to them. It will track them through | |
3505 | * the pipeline but severs its ties whenever they get into | |
3506 | * trouble (e.g. suspended). This allows "The Godfather" | |
3507 | * I/O to return status without blocking. | |
3508 | */ | |
3509 | for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { | |
3510 | zio_link_t *zl = zio->io_walk_link; | |
3511 | pio_next = zio_walk_parents(zio); | |
3512 | ||
3513 | if ((pio->io_flags & ZIO_FLAG_GODFATHER) && | |
3514 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { | |
3515 | zio_remove_child(pio, zio, zl); | |
3516 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE); | |
3517 | } | |
3518 | } | |
3519 | ||
d164b209 | 3520 | if ((pio = zio_unique_parent(zio)) != NULL) { |
b128c09f BB |
3521 | /* |
3522 | * We're not a root i/o, so there's nothing to do | |
3523 | * but notify our parent. Don't propagate errors | |
3524 | * upward since we haven't permanently failed yet. | |
3525 | */ | |
9babb374 | 3526 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f BB |
3527 | zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; |
3528 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE); | |
3529 | } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { | |
3530 | /* | |
3531 | * We'd fail again if we reexecuted now, so suspend | |
3532 | * until conditions improve (e.g. device comes online). | |
3533 | */ | |
c776b317 | 3534 | zio_suspend(zio->io_spa, zio); |
b128c09f BB |
3535 | } else { |
3536 | /* | |
3537 | * Reexecution is potentially a huge amount of work. | |
3538 | * Hand it off to the otherwise-unused claim taskq. | |
3539 | */ | |
a38718a6 | 3540 | ASSERT(taskq_empty_ent(&zio->io_tqent)); |
7ef5e54e AL |
3541 | spa_taskq_dispatch_ent(zio->io_spa, |
3542 | ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE, | |
a38718a6 GA |
3543 | (task_func_t *)zio_reexecute, zio, 0, |
3544 | &zio->io_tqent); | |
b128c09f BB |
3545 | } |
3546 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f BB |
3547 | } |
3548 | ||
428870ff | 3549 | ASSERT(zio->io_child_count == 0); |
b128c09f BB |
3550 | ASSERT(zio->io_reexecute == 0); |
3551 | ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); | |
34dc7c2f | 3552 | |
428870ff BB |
3553 | /* |
3554 | * Report any checksum errors, since the I/O is complete. | |
3555 | */ | |
3556 | while (zio->io_cksum_report != NULL) { | |
3557 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
3558 | zio->io_cksum_report = zcr->zcr_next; | |
3559 | zcr->zcr_next = NULL; | |
3560 | zcr->zcr_finish(zcr, NULL); | |
3561 | zfs_ereport_free_checksum(zcr); | |
3562 | } | |
3563 | ||
920dd524 | 3564 | if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp && |
9b67f605 MA |
3565 | !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) && |
3566 | !(zio->io_flags & ZIO_FLAG_NOPWRITE)) { | |
920dd524 ED |
3567 | metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp); |
3568 | } | |
3569 | ||
d164b209 BB |
3570 | /* |
3571 | * It is the responsibility of the done callback to ensure that this | |
3572 | * particular zio is no longer discoverable for adoption, and as | |
3573 | * such, cannot acquire any new parents. | |
3574 | */ | |
b128c09f BB |
3575 | if (zio->io_done) |
3576 | zio->io_done(zio); | |
34dc7c2f | 3577 | |
d164b209 BB |
3578 | mutex_enter(&zio->io_lock); |
3579 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
3580 | mutex_exit(&zio->io_lock); | |
34dc7c2f | 3581 | |
d164b209 BB |
3582 | for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { |
3583 | zio_link_t *zl = zio->io_walk_link; | |
3584 | pio_next = zio_walk_parents(zio); | |
3585 | zio_remove_child(pio, zio, zl); | |
b128c09f BB |
3586 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE); |
3587 | } | |
34dc7c2f | 3588 | |
b128c09f BB |
3589 | if (zio->io_waiter != NULL) { |
3590 | mutex_enter(&zio->io_lock); | |
3591 | zio->io_executor = NULL; | |
3592 | cv_broadcast(&zio->io_cv); | |
3593 | mutex_exit(&zio->io_lock); | |
3594 | } else { | |
3595 | zio_destroy(zio); | |
3596 | } | |
34dc7c2f | 3597 | |
b128c09f | 3598 | return (ZIO_PIPELINE_STOP); |
34dc7c2f BB |
3599 | } |
3600 | ||
3601 | /* | |
b128c09f BB |
3602 | * ========================================================================== |
3603 | * I/O pipeline definition | |
3604 | * ========================================================================== | |
34dc7c2f | 3605 | */ |
428870ff | 3606 | static zio_pipe_stage_t *zio_pipeline[] = { |
b128c09f | 3607 | NULL, |
b128c09f | 3608 | zio_read_bp_init, |
428870ff BB |
3609 | zio_free_bp_init, |
3610 | zio_issue_async, | |
b128c09f BB |
3611 | zio_write_bp_init, |
3612 | zio_checksum_generate, | |
03c6040b | 3613 | zio_nop_write, |
428870ff BB |
3614 | zio_ddt_read_start, |
3615 | zio_ddt_read_done, | |
3616 | zio_ddt_write, | |
3617 | zio_ddt_free, | |
b128c09f BB |
3618 | zio_gang_assemble, |
3619 | zio_gang_issue, | |
3620 | zio_dva_allocate, | |
3621 | zio_dva_free, | |
3622 | zio_dva_claim, | |
3623 | zio_ready, | |
3624 | zio_vdev_io_start, | |
3625 | zio_vdev_io_done, | |
3626 | zio_vdev_io_assess, | |
3627 | zio_checksum_verify, | |
3628 | zio_done | |
3629 | }; | |
c28b2279 | 3630 | |
9ae529ec | 3631 | |
9ae529ec | 3632 | |
9ae529ec | 3633 | |
fcff0f35 PD |
3634 | /* |
3635 | * Compare two zbookmark_phys_t's to see which we would reach first in a | |
3636 | * pre-order traversal of the object tree. | |
3637 | * | |
3638 | * This is simple in every case aside from the meta-dnode object. For all other | |
3639 | * objects, we traverse them in order (object 1 before object 2, and so on). | |
3640 | * However, all of these objects are traversed while traversing object 0, since | |
3641 | * the data it points to is the list of objects. Thus, we need to convert to a | |
3642 | * canonical representation so we can compare meta-dnode bookmarks to | |
3643 | * non-meta-dnode bookmarks. | |
3644 | * | |
3645 | * We do this by calculating "equivalents" for each field of the zbookmark. | |
3646 | * zbookmarks outside of the meta-dnode use their own object and level, and | |
3647 | * calculate the level 0 equivalent (the first L0 blkid that is contained in the | |
3648 | * blocks this bookmark refers to) by multiplying their blkid by their span | |
3649 | * (the number of L0 blocks contained within one block at their level). | |
3650 | * zbookmarks inside the meta-dnode calculate their object equivalent | |
3651 | * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use | |
3652 | * level + 1<<31 (any value larger than a level could ever be) for their level. | |
3653 | * This causes them to always compare before a bookmark in their object | |
3654 | * equivalent, compare appropriately to bookmarks in other objects, and to | |
3655 | * compare appropriately to other bookmarks in the meta-dnode. | |
3656 | */ | |
3657 | int | |
3658 | zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, | |
3659 | const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) | |
3660 | { | |
3661 | /* | |
3662 | * These variables represent the "equivalent" values for the zbookmark, | |
3663 | * after converting zbookmarks inside the meta dnode to their | |
3664 | * normal-object equivalents. | |
3665 | */ | |
3666 | uint64_t zb1obj, zb2obj; | |
3667 | uint64_t zb1L0, zb2L0; | |
3668 | uint64_t zb1level, zb2level; | |
3669 | ||
3670 | if (zb1->zb_object == zb2->zb_object && | |
3671 | zb1->zb_level == zb2->zb_level && | |
3672 | zb1->zb_blkid == zb2->zb_blkid) | |
3673 | return (0); | |
9ae529ec | 3674 | |
fcff0f35 PD |
3675 | /* |
3676 | * BP_SPANB calculates the span in blocks. | |
3677 | */ | |
3678 | zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); | |
3679 | zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); | |
9ae529ec CS |
3680 | |
3681 | if (zb1->zb_object == DMU_META_DNODE_OBJECT) { | |
fcff0f35 PD |
3682 | zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); |
3683 | zb1L0 = 0; | |
3684 | zb1level = zb1->zb_level + COMPARE_META_LEVEL; | |
3685 | } else { | |
3686 | zb1obj = zb1->zb_object; | |
3687 | zb1level = zb1->zb_level; | |
9ae529ec CS |
3688 | } |
3689 | ||
fcff0f35 PD |
3690 | if (zb2->zb_object == DMU_META_DNODE_OBJECT) { |
3691 | zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); | |
3692 | zb2L0 = 0; | |
3693 | zb2level = zb2->zb_level + COMPARE_META_LEVEL; | |
3694 | } else { | |
3695 | zb2obj = zb2->zb_object; | |
3696 | zb2level = zb2->zb_level; | |
3697 | } | |
3698 | ||
3699 | /* Now that we have a canonical representation, do the comparison. */ | |
3700 | if (zb1obj != zb2obj) | |
3701 | return (zb1obj < zb2obj ? -1 : 1); | |
3702 | else if (zb1L0 != zb2L0) | |
3703 | return (zb1L0 < zb2L0 ? -1 : 1); | |
3704 | else if (zb1level != zb2level) | |
3705 | return (zb1level > zb2level ? -1 : 1); | |
3706 | /* | |
3707 | * This can (theoretically) happen if the bookmarks have the same object | |
3708 | * and level, but different blkids, if the block sizes are not the same. | |
3709 | * There is presently no way to change the indirect block sizes | |
3710 | */ | |
3711 | return (0); | |
3712 | } | |
3713 | ||
3714 | /* | |
3715 | * This function checks the following: given that last_block is the place that | |
3716 | * our traversal stopped last time, does that guarantee that we've visited | |
3717 | * every node under subtree_root? Therefore, we can't just use the raw output | |
3718 | * of zbookmark_compare. We have to pass in a modified version of | |
3719 | * subtree_root; by incrementing the block id, and then checking whether | |
3720 | * last_block is before or equal to that, we can tell whether or not having | |
3721 | * visited last_block implies that all of subtree_root's children have been | |
3722 | * visited. | |
3723 | */ | |
3724 | boolean_t | |
3725 | zbookmark_subtree_completed(const dnode_phys_t *dnp, | |
3726 | const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) | |
3727 | { | |
3728 | zbookmark_phys_t mod_zb = *subtree_root; | |
3729 | mod_zb.zb_blkid++; | |
3730 | ASSERT(last_block->zb_level == 0); | |
3731 | ||
3732 | /* The objset_phys_t isn't before anything. */ | |
3733 | if (dnp == NULL) | |
9ae529ec | 3734 | return (B_FALSE); |
fcff0f35 PD |
3735 | |
3736 | /* | |
3737 | * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the | |
3738 | * data block size in sectors, because that variable is only used if | |
3739 | * the bookmark refers to a block in the meta-dnode. Since we don't | |
3740 | * know without examining it what object it refers to, and there's no | |
3741 | * harm in passing in this value in other cases, we always pass it in. | |
3742 | * | |
3743 | * We pass in 0 for the indirect block size shift because zb2 must be | |
3744 | * level 0. The indirect block size is only used to calculate the span | |
3745 | * of the bookmark, but since the bookmark must be level 0, the span is | |
3746 | * always 1, so the math works out. | |
3747 | * | |
3748 | * If you make changes to how the zbookmark_compare code works, be sure | |
3749 | * to make sure that this code still works afterwards. | |
3750 | */ | |
3751 | return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, | |
3752 | 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, | |
3753 | last_block) <= 0); | |
9ae529ec CS |
3754 | } |
3755 | ||
c28b2279 | 3756 | #if defined(_KERNEL) && defined(HAVE_SPL) |
c28b2279 | 3757 | EXPORT_SYMBOL(zio_type_name); |
81971b13 BB |
3758 | EXPORT_SYMBOL(zio_buf_alloc); |
3759 | EXPORT_SYMBOL(zio_data_buf_alloc); | |
6fe53787 | 3760 | EXPORT_SYMBOL(zio_buf_alloc_flags); |
81971b13 BB |
3761 | EXPORT_SYMBOL(zio_buf_free); |
3762 | EXPORT_SYMBOL(zio_data_buf_free); | |
c28b2279 | 3763 | |
a69052be | 3764 | module_param(zio_delay_max, int, 0644); |
c409e464 BB |
3765 | MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event"); |
3766 | ||
3767 | module_param(zio_requeue_io_start_cut_in_line, int, 0644); | |
3768 | MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O"); | |
29dee3ee CP |
3769 | |
3770 | module_param(zfs_sync_pass_deferred_free, int, 0644); | |
3771 | MODULE_PARM_DESC(zfs_sync_pass_deferred_free, | |
d1d7e268 | 3772 | "Defer frees starting in this pass"); |
29dee3ee CP |
3773 | |
3774 | module_param(zfs_sync_pass_dont_compress, int, 0644); | |
3775 | MODULE_PARM_DESC(zfs_sync_pass_dont_compress, | |
d1d7e268 | 3776 | "Don't compress starting in this pass"); |
29dee3ee CP |
3777 | |
3778 | module_param(zfs_sync_pass_rewrite, int, 0644); | |
3779 | MODULE_PARM_DESC(zfs_sync_pass_rewrite, | |
d1d7e268 | 3780 | "Rewrite new bps starting in this pass"); |
c28b2279 | 3781 | #endif |