]>
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> |
3dfb57a3 | 42 | #include <sys/metaslab_impl.h> |
193a37cb | 43 | #include <sys/time.h> |
26ef0cc7 | 44 | #include <sys/trace_zio.h> |
34dc7c2f | 45 | |
34dc7c2f BB |
46 | /* |
47 | * ========================================================================== | |
48 | * I/O type descriptions | |
49 | * ========================================================================== | |
50 | */ | |
e8b96c60 | 51 | const char *zio_type_name[ZIO_TYPES] = { |
3dfb57a3 DB |
52 | /* |
53 | * Note: Linux kernel thread name length is limited | |
54 | * so these names will differ from upstream open zfs. | |
55 | */ | |
451041db | 56 | "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl" |
428870ff | 57 | }; |
34dc7c2f | 58 | |
3dfb57a3 DB |
59 | int zio_dva_throttle_enabled = B_TRUE; |
60 | ||
34dc7c2f BB |
61 | /* |
62 | * ========================================================================== | |
63 | * I/O kmem caches | |
64 | * ========================================================================== | |
65 | */ | |
66 | kmem_cache_t *zio_cache; | |
d164b209 | 67 | kmem_cache_t *zio_link_cache; |
34dc7c2f BB |
68 | kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; |
69 | kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
a69052be | 70 | int zio_delay_max = ZIO_DELAY_MAX; |
34dc7c2f | 71 | |
98b25418 GW |
72 | #define ZIO_PIPELINE_CONTINUE 0x100 |
73 | #define ZIO_PIPELINE_STOP 0x101 | |
74 | ||
fcff0f35 PD |
75 | #define BP_SPANB(indblkshift, level) \ |
76 | (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) | |
77 | #define COMPARE_META_LEVEL 0x80000000ul | |
55d85d5a GW |
78 | /* |
79 | * The following actions directly effect the spa's sync-to-convergence logic. | |
80 | * The values below define the sync pass when we start performing the action. | |
81 | * Care should be taken when changing these values as they directly impact | |
82 | * spa_sync() performance. Tuning these values may introduce subtle performance | |
83 | * pathologies and should only be done in the context of performance analysis. | |
84 | * These tunables will eventually be removed and replaced with #defines once | |
85 | * enough analysis has been done to determine optimal values. | |
86 | * | |
87 | * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that | |
88 | * regular blocks are not deferred. | |
89 | */ | |
90 | int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ | |
91 | int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */ | |
92 | int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ | |
93 | ||
34dc7c2f | 94 | /* |
b128c09f BB |
95 | * An allocating zio is one that either currently has the DVA allocate |
96 | * stage set or will have it later in its lifetime. | |
34dc7c2f | 97 | */ |
428870ff BB |
98 | #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) |
99 | ||
c409e464 | 100 | int zio_requeue_io_start_cut_in_line = 1; |
428870ff BB |
101 | |
102 | #ifdef ZFS_DEBUG | |
103 | int zio_buf_debug_limit = 16384; | |
104 | #else | |
105 | int zio_buf_debug_limit = 0; | |
106 | #endif | |
34dc7c2f | 107 | |
da6b4005 NB |
108 | static inline void __zio_execute(zio_t *zio); |
109 | ||
3dfb57a3 DB |
110 | static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t); |
111 | ||
34dc7c2f BB |
112 | void |
113 | zio_init(void) | |
114 | { | |
115 | size_t c; | |
116 | vmem_t *data_alloc_arena = NULL; | |
117 | ||
3941503c BB |
118 | zio_cache = kmem_cache_create("zio_cache", |
119 | sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); | |
d164b209 | 120 | zio_link_cache = kmem_cache_create("zio_link_cache", |
6795a698 | 121 | sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); |
34dc7c2f BB |
122 | |
123 | /* | |
124 | * For small buffers, we want a cache for each multiple of | |
f1512ee6 MA |
125 | * SPA_MINBLOCKSIZE. For larger buffers, we want a cache |
126 | * for each quarter-power of 2. | |
34dc7c2f BB |
127 | */ |
128 | for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { | |
129 | size_t size = (c + 1) << SPA_MINBLOCKSHIFT; | |
130 | size_t p2 = size; | |
131 | size_t align = 0; | |
6442f3cf | 132 | size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0; |
34dc7c2f | 133 | |
f1512ee6 MA |
134 | #ifdef _ILP32 |
135 | /* | |
136 | * Cache size limited to 1M on 32-bit platforms until ARC | |
137 | * buffers no longer require virtual address space. | |
138 | */ | |
139 | if (size > zfs_max_recordsize) | |
140 | break; | |
141 | #endif | |
142 | ||
143 | while (!ISP2(p2)) | |
34dc7c2f BB |
144 | p2 &= p2 - 1; |
145 | ||
498877ba MA |
146 | #ifndef _KERNEL |
147 | /* | |
148 | * If we are using watchpoints, put each buffer on its own page, | |
149 | * to eliminate the performance overhead of trapping to the | |
150 | * kernel when modifying a non-watched buffer that shares the | |
151 | * page with a watched buffer. | |
152 | */ | |
153 | if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) | |
154 | continue; | |
fcf64f45 BB |
155 | /* |
156 | * Here's the problem - on 4K native devices in userland on | |
157 | * Linux using O_DIRECT, buffers must be 4K aligned or I/O | |
158 | * will fail with EINVAL, causing zdb (and others) to coredump. | |
159 | * Since userland probably doesn't need optimized buffer caches, | |
160 | * we just force 4K alignment on everything. | |
161 | */ | |
162 | align = 8 * SPA_MINBLOCKSIZE; | |
163 | #else | |
34dc7c2f BB |
164 | if (size <= 4 * SPA_MINBLOCKSIZE) { |
165 | align = SPA_MINBLOCKSIZE; | |
498877ba | 166 | } else if (IS_P2ALIGNED(size, p2 >> 2)) { |
f1512ee6 | 167 | align = MIN(p2 >> 2, PAGESIZE); |
34dc7c2f | 168 | } |
fcf64f45 | 169 | #endif |
34dc7c2f BB |
170 | |
171 | if (align != 0) { | |
172 | char name[36]; | |
173 | (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); | |
174 | zio_buf_cache[c] = kmem_cache_create(name, size, | |
6442f3cf | 175 | align, NULL, NULL, NULL, NULL, NULL, cflags); |
34dc7c2f BB |
176 | |
177 | (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); | |
178 | zio_data_buf_cache[c] = kmem_cache_create(name, size, | |
ae6ba3db | 179 | align, NULL, NULL, NULL, NULL, |
6442f3cf | 180 | data_alloc_arena, cflags); |
34dc7c2f BB |
181 | } |
182 | } | |
183 | ||
184 | while (--c != 0) { | |
185 | ASSERT(zio_buf_cache[c] != NULL); | |
186 | if (zio_buf_cache[c - 1] == NULL) | |
187 | zio_buf_cache[c - 1] = zio_buf_cache[c]; | |
188 | ||
189 | ASSERT(zio_data_buf_cache[c] != NULL); | |
190 | if (zio_data_buf_cache[c - 1] == NULL) | |
191 | zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; | |
192 | } | |
193 | ||
34dc7c2f | 194 | zio_inject_init(); |
9759c60f ED |
195 | |
196 | lz4_init(); | |
34dc7c2f BB |
197 | } |
198 | ||
199 | void | |
200 | zio_fini(void) | |
201 | { | |
202 | size_t c; | |
203 | kmem_cache_t *last_cache = NULL; | |
204 | kmem_cache_t *last_data_cache = NULL; | |
205 | ||
206 | for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { | |
f1512ee6 MA |
207 | #ifdef _ILP32 |
208 | /* | |
209 | * Cache size limited to 1M on 32-bit platforms until ARC | |
210 | * buffers no longer require virtual address space. | |
211 | */ | |
212 | if (((c + 1) << SPA_MINBLOCKSHIFT) > zfs_max_recordsize) | |
213 | break; | |
214 | #endif | |
34dc7c2f BB |
215 | if (zio_buf_cache[c] != last_cache) { |
216 | last_cache = zio_buf_cache[c]; | |
217 | kmem_cache_destroy(zio_buf_cache[c]); | |
218 | } | |
219 | zio_buf_cache[c] = NULL; | |
220 | ||
221 | if (zio_data_buf_cache[c] != last_data_cache) { | |
222 | last_data_cache = zio_data_buf_cache[c]; | |
223 | kmem_cache_destroy(zio_data_buf_cache[c]); | |
224 | } | |
225 | zio_data_buf_cache[c] = NULL; | |
226 | } | |
227 | ||
d164b209 | 228 | kmem_cache_destroy(zio_link_cache); |
34dc7c2f BB |
229 | kmem_cache_destroy(zio_cache); |
230 | ||
231 | zio_inject_fini(); | |
9759c60f ED |
232 | |
233 | lz4_fini(); | |
34dc7c2f BB |
234 | } |
235 | ||
236 | /* | |
237 | * ========================================================================== | |
238 | * Allocate and free I/O buffers | |
239 | * ========================================================================== | |
240 | */ | |
241 | ||
242 | /* | |
243 | * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a | |
244 | * crashdump if the kernel panics, so use it judiciously. Obviously, it's | |
245 | * useful to inspect ZFS metadata, but if possible, we should avoid keeping | |
246 | * excess / transient data in-core during a crashdump. | |
247 | */ | |
248 | void * | |
249 | zio_buf_alloc(size_t size) | |
250 | { | |
251 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
252 | ||
63e3a861 | 253 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f | 254 | |
efcd79a8 | 255 | return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); |
34dc7c2f BB |
256 | } |
257 | ||
258 | /* | |
259 | * Use zio_data_buf_alloc to allocate data. The data will not appear in a | |
260 | * crashdump if the kernel panics. This exists so that we will limit the amount | |
261 | * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount | |
262 | * of kernel heap dumped to disk when the kernel panics) | |
263 | */ | |
264 | void * | |
265 | zio_data_buf_alloc(size_t size) | |
266 | { | |
267 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
268 | ||
63e3a861 | 269 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f | 270 | |
efcd79a8 | 271 | return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); |
34dc7c2f BB |
272 | } |
273 | ||
6fe53787 BB |
274 | /* |
275 | * Use zio_buf_alloc_flags when specific allocation flags are needed. e.g. | |
276 | * passing KM_NOSLEEP when it is acceptable for an allocation to fail. | |
277 | */ | |
278 | void * | |
279 | zio_buf_alloc_flags(size_t size, int flags) | |
280 | { | |
281 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
282 | ||
283 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); | |
284 | ||
285 | return (kmem_cache_alloc(zio_buf_cache[c], flags)); | |
286 | } | |
287 | ||
34dc7c2f BB |
288 | void |
289 | zio_buf_free(void *buf, size_t size) | |
290 | { | |
291 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
292 | ||
63e3a861 | 293 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f BB |
294 | |
295 | kmem_cache_free(zio_buf_cache[c], buf); | |
296 | } | |
297 | ||
298 | void | |
299 | zio_data_buf_free(void *buf, size_t size) | |
300 | { | |
301 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
302 | ||
63e3a861 | 303 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f BB |
304 | |
305 | kmem_cache_free(zio_data_buf_cache[c], buf); | |
306 | } | |
307 | ||
308 | /* | |
309 | * ========================================================================== | |
310 | * Push and pop I/O transform buffers | |
311 | * ========================================================================== | |
312 | */ | |
d3c2ae1c | 313 | void |
b128c09f BB |
314 | zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize, |
315 | zio_transform_func_t *transform) | |
34dc7c2f | 316 | { |
79c76d5b | 317 | zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); |
34dc7c2f | 318 | |
b128c09f BB |
319 | zt->zt_orig_data = zio->io_data; |
320 | zt->zt_orig_size = zio->io_size; | |
34dc7c2f | 321 | zt->zt_bufsize = bufsize; |
b128c09f | 322 | zt->zt_transform = transform; |
34dc7c2f BB |
323 | |
324 | zt->zt_next = zio->io_transform_stack; | |
325 | zio->io_transform_stack = zt; | |
326 | ||
327 | zio->io_data = data; | |
328 | zio->io_size = size; | |
329 | } | |
330 | ||
d3c2ae1c | 331 | void |
b128c09f | 332 | zio_pop_transforms(zio_t *zio) |
34dc7c2f | 333 | { |
b128c09f BB |
334 | zio_transform_t *zt; |
335 | ||
336 | while ((zt = zio->io_transform_stack) != NULL) { | |
337 | if (zt->zt_transform != NULL) | |
338 | zt->zt_transform(zio, | |
339 | zt->zt_orig_data, zt->zt_orig_size); | |
34dc7c2f | 340 | |
428870ff BB |
341 | if (zt->zt_bufsize != 0) |
342 | zio_buf_free(zio->io_data, zt->zt_bufsize); | |
34dc7c2f | 343 | |
b128c09f BB |
344 | zio->io_data = zt->zt_orig_data; |
345 | zio->io_size = zt->zt_orig_size; | |
346 | zio->io_transform_stack = zt->zt_next; | |
34dc7c2f | 347 | |
b128c09f | 348 | kmem_free(zt, sizeof (zio_transform_t)); |
34dc7c2f BB |
349 | } |
350 | } | |
351 | ||
b128c09f BB |
352 | /* |
353 | * ========================================================================== | |
354 | * I/O transform callbacks for subblocks and decompression | |
355 | * ========================================================================== | |
356 | */ | |
357 | static void | |
358 | zio_subblock(zio_t *zio, void *data, uint64_t size) | |
359 | { | |
360 | ASSERT(zio->io_size > size); | |
361 | ||
362 | if (zio->io_type == ZIO_TYPE_READ) | |
363 | bcopy(zio->io_data, data, size); | |
364 | } | |
365 | ||
366 | static void | |
367 | zio_decompress(zio_t *zio, void *data, uint64_t size) | |
368 | { | |
369 | if (zio->io_error == 0 && | |
370 | zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), | |
428870ff | 371 | zio->io_data, data, zio->io_size, size) != 0) |
2e528b49 | 372 | zio->io_error = SET_ERROR(EIO); |
b128c09f BB |
373 | } |
374 | ||
375 | /* | |
376 | * ========================================================================== | |
377 | * I/O parent/child relationships and pipeline interlocks | |
378 | * ========================================================================== | |
379 | */ | |
d164b209 | 380 | zio_t * |
3dfb57a3 | 381 | zio_walk_parents(zio_t *cio, zio_link_t **zl) |
d164b209 | 382 | { |
d164b209 | 383 | list_t *pl = &cio->io_parent_list; |
b128c09f | 384 | |
3dfb57a3 DB |
385 | *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl); |
386 | if (*zl == NULL) | |
d164b209 BB |
387 | return (NULL); |
388 | ||
3dfb57a3 DB |
389 | ASSERT((*zl)->zl_child == cio); |
390 | return ((*zl)->zl_parent); | |
d164b209 BB |
391 | } |
392 | ||
393 | zio_t * | |
3dfb57a3 | 394 | zio_walk_children(zio_t *pio, zio_link_t **zl) |
d164b209 | 395 | { |
d164b209 BB |
396 | list_t *cl = &pio->io_child_list; |
397 | ||
3dfb57a3 DB |
398 | *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl); |
399 | if (*zl == NULL) | |
d164b209 BB |
400 | return (NULL); |
401 | ||
3dfb57a3 DB |
402 | ASSERT((*zl)->zl_parent == pio); |
403 | return ((*zl)->zl_child); | |
d164b209 BB |
404 | } |
405 | ||
406 | zio_t * | |
407 | zio_unique_parent(zio_t *cio) | |
408 | { | |
3dfb57a3 DB |
409 | zio_link_t *zl = NULL; |
410 | zio_t *pio = zio_walk_parents(cio, &zl); | |
d164b209 | 411 | |
3dfb57a3 | 412 | VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL); |
d164b209 BB |
413 | return (pio); |
414 | } | |
415 | ||
416 | void | |
417 | zio_add_child(zio_t *pio, zio_t *cio) | |
b128c09f | 418 | { |
79c76d5b | 419 | zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); |
d6320ddb | 420 | int w; |
d164b209 BB |
421 | |
422 | /* | |
423 | * Logical I/Os can have logical, gang, or vdev children. | |
424 | * Gang I/Os can have gang or vdev children. | |
425 | * Vdev I/Os can only have vdev children. | |
426 | * The following ASSERT captures all of these constraints. | |
427 | */ | |
428 | ASSERT(cio->io_child_type <= pio->io_child_type); | |
429 | ||
430 | zl->zl_parent = pio; | |
431 | zl->zl_child = cio; | |
432 | ||
433 | mutex_enter(&cio->io_lock); | |
b128c09f | 434 | mutex_enter(&pio->io_lock); |
d164b209 BB |
435 | |
436 | ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); | |
437 | ||
d6320ddb | 438 | for (w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 BB |
439 | pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; |
440 | ||
441 | list_insert_head(&pio->io_child_list, zl); | |
442 | list_insert_head(&cio->io_parent_list, zl); | |
443 | ||
428870ff BB |
444 | pio->io_child_count++; |
445 | cio->io_parent_count++; | |
446 | ||
b128c09f | 447 | mutex_exit(&pio->io_lock); |
d164b209 | 448 | mutex_exit(&cio->io_lock); |
b128c09f BB |
449 | } |
450 | ||
34dc7c2f | 451 | static void |
d164b209 | 452 | zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) |
b128c09f | 453 | { |
d164b209 BB |
454 | ASSERT(zl->zl_parent == pio); |
455 | ASSERT(zl->zl_child == cio); | |
b128c09f | 456 | |
d164b209 | 457 | mutex_enter(&cio->io_lock); |
b128c09f | 458 | mutex_enter(&pio->io_lock); |
d164b209 BB |
459 | |
460 | list_remove(&pio->io_child_list, zl); | |
461 | list_remove(&cio->io_parent_list, zl); | |
462 | ||
428870ff BB |
463 | pio->io_child_count--; |
464 | cio->io_parent_count--; | |
465 | ||
b128c09f | 466 | mutex_exit(&pio->io_lock); |
d164b209 | 467 | mutex_exit(&cio->io_lock); |
d164b209 | 468 | kmem_cache_free(zio_link_cache, zl); |
b128c09f BB |
469 | } |
470 | ||
471 | static boolean_t | |
472 | zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) | |
34dc7c2f | 473 | { |
b128c09f BB |
474 | uint64_t *countp = &zio->io_children[child][wait]; |
475 | boolean_t waiting = B_FALSE; | |
476 | ||
477 | mutex_enter(&zio->io_lock); | |
478 | ASSERT(zio->io_stall == NULL); | |
479 | if (*countp != 0) { | |
428870ff | 480 | zio->io_stage >>= 1; |
3dfb57a3 | 481 | ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN); |
b128c09f BB |
482 | zio->io_stall = countp; |
483 | waiting = B_TRUE; | |
484 | } | |
485 | mutex_exit(&zio->io_lock); | |
486 | ||
487 | return (waiting); | |
488 | } | |
34dc7c2f | 489 | |
bf701a83 BB |
490 | __attribute__((always_inline)) |
491 | static inline void | |
b128c09f BB |
492 | zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) |
493 | { | |
494 | uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; | |
495 | int *errorp = &pio->io_child_error[zio->io_child_type]; | |
34dc7c2f | 496 | |
b128c09f BB |
497 | mutex_enter(&pio->io_lock); |
498 | if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) | |
499 | *errorp = zio_worst_error(*errorp, zio->io_error); | |
500 | pio->io_reexecute |= zio->io_reexecute; | |
501 | ASSERT3U(*countp, >, 0); | |
e8b96c60 MA |
502 | |
503 | (*countp)--; | |
504 | ||
505 | if (*countp == 0 && pio->io_stall == countp) { | |
3dfb57a3 DB |
506 | zio_taskq_type_t type = |
507 | pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE : | |
508 | ZIO_TASKQ_INTERRUPT; | |
b128c09f BB |
509 | pio->io_stall = NULL; |
510 | mutex_exit(&pio->io_lock); | |
3dfb57a3 DB |
511 | /* |
512 | * Dispatch the parent zio in its own taskq so that | |
513 | * the child can continue to make progress. This also | |
514 | * prevents overflowing the stack when we have deeply nested | |
515 | * parent-child relationships. | |
516 | */ | |
517 | zio_taskq_dispatch(pio, type, B_FALSE); | |
b128c09f BB |
518 | } else { |
519 | mutex_exit(&pio->io_lock); | |
34dc7c2f BB |
520 | } |
521 | } | |
522 | ||
b128c09f BB |
523 | static void |
524 | zio_inherit_child_errors(zio_t *zio, enum zio_child c) | |
525 | { | |
526 | if (zio->io_child_error[c] != 0 && zio->io_error == 0) | |
527 | zio->io_error = zio->io_child_error[c]; | |
528 | } | |
529 | ||
3dfb57a3 DB |
530 | int |
531 | zio_timestamp_compare(const void *x1, const void *x2) | |
532 | { | |
533 | const zio_t *z1 = x1; | |
534 | const zio_t *z2 = x2; | |
535 | int cmp; | |
536 | ||
537 | cmp = AVL_CMP(z1->io_queued_timestamp, z2->io_queued_timestamp); | |
538 | if (likely(cmp)) | |
539 | return (cmp); | |
540 | ||
541 | cmp = AVL_CMP(z1->io_offset, z2->io_offset); | |
542 | if (likely(cmp)) | |
543 | return (cmp); | |
544 | ||
545 | return (AVL_PCMP(z1, z2)); | |
546 | } | |
547 | ||
34dc7c2f BB |
548 | /* |
549 | * ========================================================================== | |
b128c09f | 550 | * Create the various types of I/O (read, write, free, etc) |
34dc7c2f BB |
551 | * ========================================================================== |
552 | */ | |
553 | static zio_t * | |
428870ff | 554 | zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
2aa34383 DK |
555 | void *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done, |
556 | void *private, zio_type_t type, zio_priority_t priority, | |
557 | enum zio_flag flags, vdev_t *vd, uint64_t offset, | |
558 | const zbookmark_phys_t *zb, enum zio_stage stage, | |
559 | enum zio_stage pipeline) | |
34dc7c2f BB |
560 | { |
561 | zio_t *zio; | |
562 | ||
2aa34383 DK |
563 | ASSERT3U(psize, <=, SPA_MAXBLOCKSIZE); |
564 | ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); | |
b128c09f BB |
565 | ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); |
566 | ||
567 | ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); | |
568 | ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); | |
569 | ASSERT(vd || stage == ZIO_STAGE_OPEN); | |
34dc7c2f | 570 | |
2aa34383 DK |
571 | IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0); |
572 | ||
79c76d5b | 573 | zio = kmem_cache_alloc(zio_cache, KM_SLEEP); |
3941503c BB |
574 | bzero(zio, sizeof (zio_t)); |
575 | ||
448d7aaa | 576 | mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL); |
3941503c BB |
577 | cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); |
578 | ||
579 | list_create(&zio->io_parent_list, sizeof (zio_link_t), | |
580 | offsetof(zio_link_t, zl_parent_node)); | |
581 | list_create(&zio->io_child_list, sizeof (zio_link_t), | |
582 | offsetof(zio_link_t, zl_child_node)); | |
d164b209 | 583 | |
b128c09f BB |
584 | if (vd != NULL) |
585 | zio->io_child_type = ZIO_CHILD_VDEV; | |
586 | else if (flags & ZIO_FLAG_GANG_CHILD) | |
587 | zio->io_child_type = ZIO_CHILD_GANG; | |
428870ff BB |
588 | else if (flags & ZIO_FLAG_DDT_CHILD) |
589 | zio->io_child_type = ZIO_CHILD_DDT; | |
b128c09f BB |
590 | else |
591 | zio->io_child_type = ZIO_CHILD_LOGICAL; | |
592 | ||
34dc7c2f | 593 | if (bp != NULL) { |
428870ff | 594 | zio->io_bp = (blkptr_t *)bp; |
34dc7c2f BB |
595 | zio->io_bp_copy = *bp; |
596 | zio->io_bp_orig = *bp; | |
428870ff BB |
597 | if (type != ZIO_TYPE_WRITE || |
598 | zio->io_child_type == ZIO_CHILD_DDT) | |
b128c09f | 599 | zio->io_bp = &zio->io_bp_copy; /* so caller can free */ |
9babb374 | 600 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) |
b128c09f | 601 | zio->io_logical = zio; |
9babb374 BB |
602 | if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) |
603 | pipeline |= ZIO_GANG_STAGES; | |
34dc7c2f | 604 | } |
b128c09f BB |
605 | |
606 | zio->io_spa = spa; | |
607 | zio->io_txg = txg; | |
34dc7c2f BB |
608 | zio->io_done = done; |
609 | zio->io_private = private; | |
610 | zio->io_type = type; | |
611 | zio->io_priority = priority; | |
b128c09f BB |
612 | zio->io_vd = vd; |
613 | zio->io_offset = offset; | |
428870ff | 614 | zio->io_orig_data = zio->io_data = data; |
2aa34383 DK |
615 | zio->io_orig_size = zio->io_size = psize; |
616 | zio->io_lsize = lsize; | |
b128c09f BB |
617 | zio->io_orig_flags = zio->io_flags = flags; |
618 | zio->io_orig_stage = zio->io_stage = stage; | |
619 | zio->io_orig_pipeline = zio->io_pipeline = pipeline; | |
3dfb57a3 | 620 | zio->io_pipeline_trace = ZIO_STAGE_OPEN; |
34dc7c2f | 621 | |
d164b209 BB |
622 | zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); |
623 | zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); | |
624 | ||
b128c09f BB |
625 | if (zb != NULL) |
626 | zio->io_bookmark = *zb; | |
627 | ||
628 | if (pio != NULL) { | |
b128c09f | 629 | if (zio->io_logical == NULL) |
34dc7c2f | 630 | zio->io_logical = pio->io_logical; |
9babb374 BB |
631 | if (zio->io_child_type == ZIO_CHILD_GANG) |
632 | zio->io_gang_leader = pio->io_gang_leader; | |
b128c09f | 633 | zio_add_child(pio, zio); |
34dc7c2f BB |
634 | } |
635 | ||
a38718a6 GA |
636 | taskq_init_ent(&zio->io_tqent); |
637 | ||
34dc7c2f BB |
638 | return (zio); |
639 | } | |
640 | ||
641 | static void | |
b128c09f | 642 | zio_destroy(zio_t *zio) |
34dc7c2f | 643 | { |
3941503c BB |
644 | list_destroy(&zio->io_parent_list); |
645 | list_destroy(&zio->io_child_list); | |
646 | mutex_destroy(&zio->io_lock); | |
647 | cv_destroy(&zio->io_cv); | |
b128c09f | 648 | kmem_cache_free(zio_cache, zio); |
34dc7c2f BB |
649 | } |
650 | ||
651 | zio_t * | |
d164b209 | 652 | zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, |
428870ff | 653 | void *private, enum zio_flag flags) |
34dc7c2f BB |
654 | { |
655 | zio_t *zio; | |
656 | ||
2aa34383 | 657 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
d164b209 | 658 | ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
b128c09f | 659 | ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); |
34dc7c2f BB |
660 | |
661 | return (zio); | |
662 | } | |
663 | ||
664 | zio_t * | |
428870ff | 665 | zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) |
34dc7c2f | 666 | { |
d164b209 | 667 | return (zio_null(NULL, spa, NULL, done, private, flags)); |
34dc7c2f BB |
668 | } |
669 | ||
63e3a861 MA |
670 | void |
671 | zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp) | |
672 | { | |
673 | int i; | |
674 | ||
675 | if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { | |
676 | zfs_panic_recover("blkptr at %p has invalid TYPE %llu", | |
677 | bp, (longlong_t)BP_GET_TYPE(bp)); | |
678 | } | |
679 | if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS || | |
680 | BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) { | |
681 | zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu", | |
682 | bp, (longlong_t)BP_GET_CHECKSUM(bp)); | |
683 | } | |
684 | if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS || | |
685 | BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) { | |
686 | zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu", | |
687 | bp, (longlong_t)BP_GET_COMPRESS(bp)); | |
688 | } | |
689 | if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
690 | zfs_panic_recover("blkptr at %p has invalid LSIZE %llu", | |
691 | bp, (longlong_t)BP_GET_LSIZE(bp)); | |
692 | } | |
693 | if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
694 | zfs_panic_recover("blkptr at %p has invalid PSIZE %llu", | |
695 | bp, (longlong_t)BP_GET_PSIZE(bp)); | |
696 | } | |
697 | ||
698 | if (BP_IS_EMBEDDED(bp)) { | |
699 | if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) { | |
700 | zfs_panic_recover("blkptr at %p has invalid ETYPE %llu", | |
701 | bp, (longlong_t)BPE_GET_ETYPE(bp)); | |
702 | } | |
703 | } | |
704 | ||
705 | /* | |
706 | * Pool-specific checks. | |
707 | * | |
708 | * Note: it would be nice to verify that the blk_birth and | |
709 | * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() | |
710 | * allows the birth time of log blocks (and dmu_sync()-ed blocks | |
711 | * that are in the log) to be arbitrarily large. | |
712 | */ | |
713 | for (i = 0; i < BP_GET_NDVAS(bp); i++) { | |
714 | uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]); | |
715 | vdev_t *vd; | |
716 | uint64_t offset, asize; | |
717 | if (vdevid >= spa->spa_root_vdev->vdev_children) { | |
718 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
719 | "VDEV %llu", | |
720 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 721 | continue; |
63e3a861 MA |
722 | } |
723 | vd = spa->spa_root_vdev->vdev_child[vdevid]; | |
724 | if (vd == NULL) { | |
725 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
726 | "VDEV %llu", | |
727 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 728 | continue; |
63e3a861 MA |
729 | } |
730 | if (vd->vdev_ops == &vdev_hole_ops) { | |
731 | zfs_panic_recover("blkptr at %p DVA %u has hole " | |
732 | "VDEV %llu", | |
733 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 734 | continue; |
63e3a861 MA |
735 | } |
736 | if (vd->vdev_ops == &vdev_missing_ops) { | |
737 | /* | |
738 | * "missing" vdevs are valid during import, but we | |
739 | * don't have their detailed info (e.g. asize), so | |
740 | * we can't perform any more checks on them. | |
741 | */ | |
742 | continue; | |
743 | } | |
744 | offset = DVA_GET_OFFSET(&bp->blk_dva[i]); | |
745 | asize = DVA_GET_ASIZE(&bp->blk_dva[i]); | |
746 | if (BP_IS_GANG(bp)) | |
747 | asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
748 | if (offset + asize > vd->vdev_asize) { | |
749 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
750 | "OFFSET %llu", | |
751 | bp, i, (longlong_t)offset); | |
752 | } | |
753 | } | |
754 | } | |
755 | ||
34dc7c2f | 756 | zio_t * |
b128c09f BB |
757 | zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, |
758 | void *data, uint64_t size, zio_done_func_t *done, void *private, | |
5dbd68a3 | 759 | zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) |
34dc7c2f BB |
760 | { |
761 | zio_t *zio; | |
762 | ||
63e3a861 MA |
763 | zfs_blkptr_verify(spa, bp); |
764 | ||
428870ff | 765 | zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, |
2aa34383 | 766 | data, size, size, done, private, |
b128c09f | 767 | ZIO_TYPE_READ, priority, flags, NULL, 0, zb, |
428870ff BB |
768 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
769 | ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); | |
34dc7c2f | 770 | |
b128c09f BB |
771 | return (zio); |
772 | } | |
34dc7c2f | 773 | |
34dc7c2f | 774 | zio_t * |
b128c09f | 775 | zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, |
2aa34383 | 776 | void *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp, |
bc77ba73 PD |
777 | zio_done_func_t *ready, zio_done_func_t *children_ready, |
778 | zio_done_func_t *physdone, zio_done_func_t *done, | |
779 | void *private, zio_priority_t priority, enum zio_flag flags, | |
780 | const zbookmark_phys_t *zb) | |
34dc7c2f BB |
781 | { |
782 | zio_t *zio; | |
783 | ||
b128c09f BB |
784 | ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && |
785 | zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && | |
786 | zp->zp_compress >= ZIO_COMPRESS_OFF && | |
787 | zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && | |
9ae529ec | 788 | DMU_OT_IS_VALID(zp->zp_type) && |
b128c09f | 789 | zp->zp_level < 32 && |
428870ff | 790 | zp->zp_copies > 0 && |
03c6040b | 791 | zp->zp_copies <= spa_max_replication(spa)); |
34dc7c2f | 792 | |
2aa34383 | 793 | zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private, |
b128c09f | 794 | ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, |
428870ff BB |
795 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
796 | ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); | |
34dc7c2f BB |
797 | |
798 | zio->io_ready = ready; | |
bc77ba73 | 799 | zio->io_children_ready = children_ready; |
e8b96c60 | 800 | zio->io_physdone = physdone; |
b128c09f | 801 | zio->io_prop = *zp; |
34dc7c2f | 802 | |
9b67f605 MA |
803 | /* |
804 | * Data can be NULL if we are going to call zio_write_override() to | |
805 | * provide the already-allocated BP. But we may need the data to | |
806 | * verify a dedup hit (if requested). In this case, don't try to | |
807 | * dedup (just take the already-allocated BP verbatim). | |
808 | */ | |
809 | if (data == NULL && zio->io_prop.zp_dedup_verify) { | |
810 | zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; | |
811 | } | |
812 | ||
34dc7c2f BB |
813 | return (zio); |
814 | } | |
815 | ||
816 | zio_t * | |
b128c09f | 817 | zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data, |
e8b96c60 | 818 | uint64_t size, zio_done_func_t *done, void *private, |
5dbd68a3 | 819 | zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) |
34dc7c2f BB |
820 | { |
821 | zio_t *zio; | |
822 | ||
2aa34383 | 823 | zio = zio_create(pio, spa, txg, bp, data, size, size, done, private, |
3dfb57a3 | 824 | ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb, |
b128c09f | 825 | ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); |
34dc7c2f BB |
826 | |
827 | return (zio); | |
828 | } | |
829 | ||
428870ff | 830 | void |
03c6040b | 831 | zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) |
428870ff BB |
832 | { |
833 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
834 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
835 | ASSERT(zio->io_stage == ZIO_STAGE_OPEN); | |
836 | ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); | |
837 | ||
03c6040b GW |
838 | /* |
839 | * We must reset the io_prop to match the values that existed | |
840 | * when the bp was first written by dmu_sync() keeping in mind | |
841 | * that nopwrite and dedup are mutually exclusive. | |
842 | */ | |
843 | zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; | |
844 | zio->io_prop.zp_nopwrite = nopwrite; | |
428870ff BB |
845 | zio->io_prop.zp_copies = copies; |
846 | zio->io_bp_override = bp; | |
847 | } | |
848 | ||
849 | void | |
850 | zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) | |
851 | { | |
9b67f605 MA |
852 | |
853 | /* | |
854 | * The check for EMBEDDED is a performance optimization. We | |
855 | * process the free here (by ignoring it) rather than | |
856 | * putting it on the list and then processing it in zio_free_sync(). | |
857 | */ | |
858 | if (BP_IS_EMBEDDED(bp)) | |
859 | return; | |
13fe0198 | 860 | metaslab_check_free(spa, bp); |
2883cad5 MA |
861 | |
862 | /* | |
863 | * Frees that are for the currently-syncing txg, are not going to be | |
864 | * deferred, and which will not need to do a read (i.e. not GANG or | |
865 | * DEDUP), can be processed immediately. Otherwise, put them on the | |
866 | * in-memory list for later processing. | |
867 | */ | |
868 | if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || | |
869 | txg != spa->spa_syncing_txg || | |
870 | spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { | |
871 | bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); | |
872 | } else { | |
873 | VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0))); | |
874 | } | |
428870ff BB |
875 | } |
876 | ||
34dc7c2f | 877 | zio_t * |
428870ff BB |
878 | zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
879 | enum zio_flag flags) | |
34dc7c2f BB |
880 | { |
881 | zio_t *zio; | |
2883cad5 | 882 | enum zio_stage stage = ZIO_FREE_PIPELINE; |
34dc7c2f | 883 | |
428870ff BB |
884 | ASSERT(!BP_IS_HOLE(bp)); |
885 | ASSERT(spa_syncing_txg(spa) == txg); | |
55d85d5a | 886 | ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); |
34dc7c2f | 887 | |
9b67f605 MA |
888 | if (BP_IS_EMBEDDED(bp)) |
889 | return (zio_null(pio, spa, NULL, NULL, NULL, 0)); | |
890 | ||
13fe0198 | 891 | metaslab_check_free(spa, bp); |
8c841793 | 892 | arc_freed(spa, bp); |
13fe0198 | 893 | |
2883cad5 MA |
894 | /* |
895 | * GANG and DEDUP blocks can induce a read (for the gang block header, | |
896 | * or the DDT), so issue them asynchronously so that this thread is | |
897 | * not tied up. | |
898 | */ | |
899 | if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) | |
900 | stage |= ZIO_STAGE_ISSUE_ASYNC; | |
901 | ||
b128c09f | 902 | zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
2aa34383 DK |
903 | BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, |
904 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage); | |
2883cad5 | 905 | |
34dc7c2f BB |
906 | return (zio); |
907 | } | |
908 | ||
909 | zio_t * | |
428870ff BB |
910 | zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
911 | zio_done_func_t *done, void *private, enum zio_flag flags) | |
34dc7c2f BB |
912 | { |
913 | zio_t *zio; | |
914 | ||
9b67f605 MA |
915 | dprintf_bp(bp, "claiming in txg %llu", txg); |
916 | ||
917 | if (BP_IS_EMBEDDED(bp)) | |
918 | return (zio_null(pio, spa, NULL, NULL, NULL, 0)); | |
919 | ||
34dc7c2f BB |
920 | /* |
921 | * A claim is an allocation of a specific block. Claims are needed | |
922 | * to support immediate writes in the intent log. The issue is that | |
923 | * immediate writes contain committed data, but in a txg that was | |
924 | * *not* committed. Upon opening the pool after an unclean shutdown, | |
925 | * the intent log claims all blocks that contain immediate write data | |
926 | * so that the SPA knows they're in use. | |
927 | * | |
928 | * All claims *must* be resolved in the first txg -- before the SPA | |
929 | * starts allocating blocks -- so that nothing is allocated twice. | |
428870ff | 930 | * If txg == 0 we just verify that the block is claimable. |
34dc7c2f BB |
931 | */ |
932 | ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); | |
428870ff BB |
933 | ASSERT(txg == spa_first_txg(spa) || txg == 0); |
934 | ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ | |
34dc7c2f | 935 | |
b128c09f | 936 | zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
2aa34383 DK |
937 | BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, |
938 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); | |
3dfb57a3 | 939 | ASSERT0(zio->io_queued_timestamp); |
34dc7c2f BB |
940 | |
941 | return (zio); | |
942 | } | |
943 | ||
944 | zio_t * | |
945 | zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, | |
e8b96c60 | 946 | zio_done_func_t *done, void *private, enum zio_flag flags) |
34dc7c2f BB |
947 | { |
948 | zio_t *zio; | |
949 | int c; | |
950 | ||
951 | if (vd->vdev_children == 0) { | |
2aa34383 | 952 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
e8b96c60 | 953 | ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
34dc7c2f BB |
954 | ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); |
955 | ||
34dc7c2f BB |
956 | zio->io_cmd = cmd; |
957 | } else { | |
d164b209 | 958 | zio = zio_null(pio, spa, NULL, NULL, NULL, flags); |
34dc7c2f BB |
959 | |
960 | for (c = 0; c < vd->vdev_children; c++) | |
961 | zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, | |
e8b96c60 | 962 | done, private, flags)); |
34dc7c2f BB |
963 | } |
964 | ||
965 | return (zio); | |
966 | } | |
967 | ||
34dc7c2f BB |
968 | zio_t * |
969 | zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
970 | void *data, int checksum, zio_done_func_t *done, void *private, | |
e8b96c60 | 971 | zio_priority_t priority, enum zio_flag flags, boolean_t labels) |
34dc7c2f BB |
972 | { |
973 | zio_t *zio; | |
34dc7c2f | 974 | |
b128c09f BB |
975 | ASSERT(vd->vdev_children == 0); |
976 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
977 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
978 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 979 | |
2aa34383 DK |
980 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
981 | private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
982 | offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); | |
34dc7c2f | 983 | |
b128c09f | 984 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f BB |
985 | |
986 | return (zio); | |
987 | } | |
988 | ||
989 | zio_t * | |
990 | zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
991 | void *data, int checksum, zio_done_func_t *done, void *private, | |
e8b96c60 | 992 | zio_priority_t priority, enum zio_flag flags, boolean_t labels) |
34dc7c2f | 993 | { |
34dc7c2f | 994 | zio_t *zio; |
34dc7c2f | 995 | |
b128c09f BB |
996 | ASSERT(vd->vdev_children == 0); |
997 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
998 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
999 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 1000 | |
2aa34383 DK |
1001 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
1002 | private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
1003 | offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); | |
34dc7c2f | 1004 | |
b128c09f | 1005 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f | 1006 | |
3c67d83a | 1007 | if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { |
34dc7c2f | 1008 | /* |
428870ff | 1009 | * zec checksums are necessarily destructive -- they modify |
b128c09f | 1010 | * the end of the write buffer to hold the verifier/checksum. |
34dc7c2f | 1011 | * Therefore, we must make a local copy in case the data is |
b128c09f | 1012 | * being written to multiple places in parallel. |
34dc7c2f | 1013 | */ |
b128c09f | 1014 | void *wbuf = zio_buf_alloc(size); |
34dc7c2f | 1015 | bcopy(data, wbuf, size); |
b128c09f | 1016 | zio_push_transform(zio, wbuf, size, size, NULL); |
34dc7c2f BB |
1017 | } |
1018 | ||
1019 | return (zio); | |
1020 | } | |
1021 | ||
1022 | /* | |
b128c09f | 1023 | * Create a child I/O to do some work for us. |
34dc7c2f BB |
1024 | */ |
1025 | zio_t * | |
b128c09f | 1026 | zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, |
e8b96c60 MA |
1027 | void *data, uint64_t size, int type, zio_priority_t priority, |
1028 | enum zio_flag flags, zio_done_func_t *done, void *private) | |
34dc7c2f | 1029 | { |
428870ff | 1030 | enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; |
b128c09f BB |
1031 | zio_t *zio; |
1032 | ||
1033 | ASSERT(vd->vdev_parent == | |
1034 | (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev)); | |
34dc7c2f BB |
1035 | |
1036 | if (type == ZIO_TYPE_READ && bp != NULL) { | |
1037 | /* | |
1038 | * If we have the bp, then the child should perform the | |
1039 | * checksum and the parent need not. This pushes error | |
1040 | * detection as close to the leaves as possible and | |
1041 | * eliminates redundant checksums in the interior nodes. | |
1042 | */ | |
428870ff BB |
1043 | pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; |
1044 | pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; | |
34dc7c2f BB |
1045 | } |
1046 | ||
b128c09f BB |
1047 | if (vd->vdev_children == 0) |
1048 | offset += VDEV_LABEL_START_SIZE; | |
1049 | ||
428870ff BB |
1050 | flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE; |
1051 | ||
1052 | /* | |
1053 | * If we've decided to do a repair, the write is not speculative -- | |
1054 | * even if the original read was. | |
1055 | */ | |
1056 | if (flags & ZIO_FLAG_IO_REPAIR) | |
1057 | flags &= ~ZIO_FLAG_SPECULATIVE; | |
1058 | ||
3dfb57a3 DB |
1059 | /* |
1060 | * If we're creating a child I/O that is not associated with a | |
1061 | * top-level vdev, then the child zio is not an allocating I/O. | |
1062 | * If this is a retried I/O then we ignore it since we will | |
1063 | * have already processed the original allocating I/O. | |
1064 | */ | |
1065 | if (flags & ZIO_FLAG_IO_ALLOCATING && | |
1066 | (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) { | |
1067 | metaslab_class_t *mc = spa_normal_class(pio->io_spa); | |
1068 | ||
1069 | ASSERT(mc->mc_alloc_throttle_enabled); | |
1070 | ASSERT(type == ZIO_TYPE_WRITE); | |
1071 | ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE); | |
1072 | ASSERT(!(flags & ZIO_FLAG_IO_REPAIR)); | |
1073 | ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) || | |
1074 | pio->io_child_type == ZIO_CHILD_GANG); | |
1075 | ||
1076 | flags &= ~ZIO_FLAG_IO_ALLOCATING; | |
1077 | } | |
1078 | ||
1079 | ||
2aa34383 | 1080 | zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size, |
428870ff BB |
1081 | done, private, type, priority, flags, vd, offset, &pio->io_bookmark, |
1082 | ZIO_STAGE_VDEV_IO_START >> 1, pipeline); | |
3dfb57a3 | 1083 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); |
34dc7c2f | 1084 | |
e8b96c60 MA |
1085 | zio->io_physdone = pio->io_physdone; |
1086 | if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) | |
1087 | zio->io_logical->io_phys_children++; | |
1088 | ||
b128c09f | 1089 | return (zio); |
34dc7c2f BB |
1090 | } |
1091 | ||
b128c09f BB |
1092 | zio_t * |
1093 | zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size, | |
e8b96c60 | 1094 | int type, zio_priority_t priority, enum zio_flag flags, |
428870ff | 1095 | zio_done_func_t *done, void *private) |
34dc7c2f | 1096 | { |
b128c09f | 1097 | zio_t *zio; |
34dc7c2f | 1098 | |
b128c09f | 1099 | ASSERT(vd->vdev_ops->vdev_op_leaf); |
34dc7c2f | 1100 | |
b128c09f | 1101 | zio = zio_create(NULL, vd->vdev_spa, 0, NULL, |
2aa34383 | 1102 | data, size, size, done, private, type, priority, |
e8b96c60 | 1103 | flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, |
b128c09f | 1104 | vd, offset, NULL, |
428870ff | 1105 | ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); |
34dc7c2f | 1106 | |
b128c09f | 1107 | return (zio); |
34dc7c2f BB |
1108 | } |
1109 | ||
1110 | void | |
b128c09f | 1111 | zio_flush(zio_t *zio, vdev_t *vd) |
34dc7c2f | 1112 | { |
b128c09f | 1113 | zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, |
e8b96c60 | 1114 | NULL, NULL, |
b128c09f | 1115 | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); |
34dc7c2f BB |
1116 | } |
1117 | ||
428870ff BB |
1118 | void |
1119 | zio_shrink(zio_t *zio, uint64_t size) | |
1120 | { | |
1121 | ASSERT(zio->io_executor == NULL); | |
1122 | ASSERT(zio->io_orig_size == zio->io_size); | |
1123 | ASSERT(size <= zio->io_size); | |
1124 | ||
1125 | /* | |
1126 | * We don't shrink for raidz because of problems with the | |
1127 | * reconstruction when reading back less than the block size. | |
1128 | * Note, BP_IS_RAIDZ() assumes no compression. | |
1129 | */ | |
1130 | ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); | |
2aa34383 DK |
1131 | if (!BP_IS_RAIDZ(zio->io_bp)) { |
1132 | /* we are not doing a raw write */ | |
1133 | ASSERT3U(zio->io_size, ==, zio->io_lsize); | |
1134 | zio->io_orig_size = zio->io_size = zio->io_lsize = size; | |
1135 | } | |
428870ff BB |
1136 | } |
1137 | ||
34dc7c2f BB |
1138 | /* |
1139 | * ========================================================================== | |
b128c09f | 1140 | * Prepare to read and write logical blocks |
34dc7c2f BB |
1141 | * ========================================================================== |
1142 | */ | |
b128c09f | 1143 | |
34dc7c2f | 1144 | static int |
b128c09f | 1145 | zio_read_bp_init(zio_t *zio) |
34dc7c2f | 1146 | { |
b128c09f | 1147 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 1148 | |
fb5f0bc8 | 1149 | if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && |
9babb374 BB |
1150 | zio->io_child_type == ZIO_CHILD_LOGICAL && |
1151 | !(zio->io_flags & ZIO_FLAG_RAW)) { | |
9b67f605 MA |
1152 | uint64_t psize = |
1153 | BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); | |
428870ff | 1154 | void *cbuf = zio_buf_alloc(psize); |
b128c09f | 1155 | |
428870ff | 1156 | zio_push_transform(zio, cbuf, psize, psize, zio_decompress); |
34dc7c2f | 1157 | } |
34dc7c2f | 1158 | |
9b67f605 MA |
1159 | if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { |
1160 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1161 | decode_embedded_bp_compressed(bp, zio->io_data); | |
1162 | } else { | |
1163 | ASSERT(!BP_IS_EMBEDDED(bp)); | |
1164 | } | |
1165 | ||
9ae529ec | 1166 | if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) |
b128c09f BB |
1167 | zio->io_flags |= ZIO_FLAG_DONT_CACHE; |
1168 | ||
428870ff BB |
1169 | if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) |
1170 | zio->io_flags |= ZIO_FLAG_DONT_CACHE; | |
1171 | ||
1172 | if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) | |
1173 | zio->io_pipeline = ZIO_DDT_READ_PIPELINE; | |
1174 | ||
b128c09f | 1175 | return (ZIO_PIPELINE_CONTINUE); |
34dc7c2f BB |
1176 | } |
1177 | ||
b128c09f BB |
1178 | static int |
1179 | zio_write_bp_init(zio_t *zio) | |
34dc7c2f | 1180 | { |
34dc7c2f | 1181 | |
b128c09f BB |
1182 | if (!IO_IS_ALLOCATING(zio)) |
1183 | return (ZIO_PIPELINE_CONTINUE); | |
34dc7c2f | 1184 | |
428870ff BB |
1185 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1186 | ||
1187 | if (zio->io_bp_override) { | |
3dfb57a3 DB |
1188 | blkptr_t *bp = zio->io_bp; |
1189 | zio_prop_t *zp = &zio->io_prop; | |
1190 | ||
428870ff BB |
1191 | ASSERT(bp->blk_birth != zio->io_txg); |
1192 | ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); | |
1193 | ||
1194 | *bp = *zio->io_bp_override; | |
1195 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1196 | ||
9b67f605 MA |
1197 | if (BP_IS_EMBEDDED(bp)) |
1198 | return (ZIO_PIPELINE_CONTINUE); | |
1199 | ||
03c6040b GW |
1200 | /* |
1201 | * If we've been overridden and nopwrite is set then | |
1202 | * set the flag accordingly to indicate that a nopwrite | |
1203 | * has already occurred. | |
1204 | */ | |
1205 | if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { | |
1206 | ASSERT(!zp->zp_dedup); | |
3dfb57a3 | 1207 | ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum); |
03c6040b GW |
1208 | zio->io_flags |= ZIO_FLAG_NOPWRITE; |
1209 | return (ZIO_PIPELINE_CONTINUE); | |
1210 | } | |
1211 | ||
1212 | ASSERT(!zp->zp_nopwrite); | |
1213 | ||
428870ff BB |
1214 | if (BP_IS_HOLE(bp) || !zp->zp_dedup) |
1215 | return (ZIO_PIPELINE_CONTINUE); | |
1216 | ||
3c67d83a TH |
1217 | ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & |
1218 | ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); | |
428870ff BB |
1219 | |
1220 | if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) { | |
1221 | BP_SET_DEDUP(bp, 1); | |
1222 | zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; | |
1223 | return (ZIO_PIPELINE_CONTINUE); | |
1224 | } | |
3dfb57a3 DB |
1225 | |
1226 | /* | |
1227 | * We were unable to handle this as an override bp, treat | |
1228 | * it as a regular write I/O. | |
1229 | */ | |
5511754b | 1230 | zio->io_bp_override = NULL; |
3dfb57a3 DB |
1231 | *bp = zio->io_bp_orig; |
1232 | zio->io_pipeline = zio->io_orig_pipeline; | |
1233 | } | |
1234 | ||
1235 | return (ZIO_PIPELINE_CONTINUE); | |
1236 | } | |
1237 | ||
1238 | static int | |
1239 | zio_write_compress(zio_t *zio) | |
1240 | { | |
1241 | spa_t *spa = zio->io_spa; | |
1242 | zio_prop_t *zp = &zio->io_prop; | |
1243 | enum zio_compress compress = zp->zp_compress; | |
1244 | blkptr_t *bp = zio->io_bp; | |
1245 | uint64_t lsize = zio->io_lsize; | |
1246 | uint64_t psize = zio->io_size; | |
1247 | int pass = 1; | |
1248 | ||
1249 | EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0); | |
1250 | ||
1251 | /* | |
1252 | * If our children haven't all reached the ready stage, | |
1253 | * wait for them and then repeat this pipeline stage. | |
1254 | */ | |
1255 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || | |
1256 | zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) | |
1257 | return (ZIO_PIPELINE_STOP); | |
1258 | ||
1259 | if (!IO_IS_ALLOCATING(zio)) | |
1260 | return (ZIO_PIPELINE_CONTINUE); | |
1261 | ||
1262 | if (zio->io_children_ready != NULL) { | |
1263 | /* | |
1264 | * Now that all our children are ready, run the callback | |
1265 | * associated with this zio in case it wants to modify the | |
1266 | * data to be written. | |
1267 | */ | |
1268 | ASSERT3U(zp->zp_level, >, 0); | |
1269 | zio->io_children_ready(zio); | |
428870ff | 1270 | } |
34dc7c2f | 1271 | |
3dfb57a3 DB |
1272 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1273 | ASSERT(zio->io_bp_override == NULL); | |
1274 | ||
b0bc7a84 | 1275 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { |
b128c09f BB |
1276 | /* |
1277 | * We're rewriting an existing block, which means we're | |
1278 | * working on behalf of spa_sync(). For spa_sync() to | |
1279 | * converge, it must eventually be the case that we don't | |
1280 | * have to allocate new blocks. But compression changes | |
1281 | * the blocksize, which forces a reallocate, and makes | |
1282 | * convergence take longer. Therefore, after the first | |
1283 | * few passes, stop compressing to ensure convergence. | |
1284 | */ | |
428870ff BB |
1285 | pass = spa_sync_pass(spa); |
1286 | ||
1287 | ASSERT(zio->io_txg == spa_syncing_txg(spa)); | |
1288 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1289 | ASSERT(!BP_GET_DEDUP(bp)); | |
34dc7c2f | 1290 | |
55d85d5a | 1291 | if (pass >= zfs_sync_pass_dont_compress) |
b128c09f | 1292 | compress = ZIO_COMPRESS_OFF; |
34dc7c2f | 1293 | |
b128c09f | 1294 | /* Make sure someone doesn't change their mind on overwrites */ |
9b67f605 | 1295 | ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), |
428870ff | 1296 | spa_max_replication(spa)) == BP_GET_NDVAS(bp)); |
b128c09f | 1297 | } |
34dc7c2f | 1298 | |
2aa34383 DK |
1299 | /* If it's a compressed write that is not raw, compress the buffer. */ |
1300 | if (compress != ZIO_COMPRESS_OFF && psize == lsize) { | |
428870ff BB |
1301 | void *cbuf = zio_buf_alloc(lsize); |
1302 | psize = zio_compress_data(compress, zio->io_data, cbuf, lsize); | |
1303 | if (psize == 0 || psize == lsize) { | |
b128c09f | 1304 | compress = ZIO_COMPRESS_OFF; |
428870ff | 1305 | zio_buf_free(cbuf, lsize); |
9b67f605 MA |
1306 | } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE && |
1307 | zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && | |
1308 | spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { | |
1309 | encode_embedded_bp_compressed(bp, | |
1310 | cbuf, compress, lsize, psize); | |
1311 | BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); | |
1312 | BP_SET_TYPE(bp, zio->io_prop.zp_type); | |
1313 | BP_SET_LEVEL(bp, zio->io_prop.zp_level); | |
1314 | zio_buf_free(cbuf, lsize); | |
1315 | bp->blk_birth = zio->io_txg; | |
1316 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1317 | ASSERT(spa_feature_is_active(spa, | |
1318 | SPA_FEATURE_EMBEDDED_DATA)); | |
1319 | return (ZIO_PIPELINE_CONTINUE); | |
428870ff | 1320 | } else { |
9b67f605 | 1321 | /* |
c3520e7f MA |
1322 | * Round up compressed size up to the ashift |
1323 | * of the smallest-ashift device, and zero the tail. | |
1324 | * This ensures that the compressed size of the BP | |
1325 | * (and thus compressratio property) are correct, | |
1326 | * in that we charge for the padding used to fill out | |
1327 | * the last sector. | |
9b67f605 | 1328 | */ |
c3520e7f MA |
1329 | size_t rounded; |
1330 | ||
1331 | ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT); | |
1332 | ||
1333 | rounded = (size_t)P2ROUNDUP(psize, | |
1334 | 1ULL << spa->spa_min_ashift); | |
1335 | if (rounded >= lsize) { | |
9b67f605 MA |
1336 | compress = ZIO_COMPRESS_OFF; |
1337 | zio_buf_free(cbuf, lsize); | |
c3520e7f | 1338 | psize = lsize; |
9b67f605 | 1339 | } else { |
c3520e7f MA |
1340 | bzero((char *)cbuf + psize, rounded - psize); |
1341 | psize = rounded; | |
9b67f605 MA |
1342 | zio_push_transform(zio, cbuf, |
1343 | psize, lsize, NULL); | |
1344 | } | |
b128c09f | 1345 | } |
3dfb57a3 DB |
1346 | |
1347 | /* | |
1348 | * We were unable to handle this as an override bp, treat | |
1349 | * it as a regular write I/O. | |
1350 | */ | |
1351 | zio->io_bp_override = NULL; | |
1352 | *bp = zio->io_bp_orig; | |
1353 | zio->io_pipeline = zio->io_orig_pipeline; | |
1354 | ||
2aa34383 DK |
1355 | } else { |
1356 | ASSERT3U(psize, !=, 0); | |
1357 | ||
b128c09f | 1358 | } |
34dc7c2f | 1359 | |
b128c09f BB |
1360 | /* |
1361 | * The final pass of spa_sync() must be all rewrites, but the first | |
1362 | * few passes offer a trade-off: allocating blocks defers convergence, | |
1363 | * but newly allocated blocks are sequential, so they can be written | |
1364 | * to disk faster. Therefore, we allow the first few passes of | |
1365 | * spa_sync() to allocate new blocks, but force rewrites after that. | |
1366 | * There should only be a handful of blocks after pass 1 in any case. | |
1367 | */ | |
b0bc7a84 MG |
1368 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && |
1369 | BP_GET_PSIZE(bp) == psize && | |
55d85d5a | 1370 | pass >= zfs_sync_pass_rewrite) { |
428870ff | 1371 | enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; |
d6320ddb | 1372 | ASSERT(psize != 0); |
b128c09f BB |
1373 | zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; |
1374 | zio->io_flags |= ZIO_FLAG_IO_REWRITE; | |
1375 | } else { | |
1376 | BP_ZERO(bp); | |
1377 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
1378 | } | |
34dc7c2f | 1379 | |
428870ff | 1380 | if (psize == 0) { |
b0bc7a84 MG |
1381 | if (zio->io_bp_orig.blk_birth != 0 && |
1382 | spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { | |
1383 | BP_SET_LSIZE(bp, lsize); | |
1384 | BP_SET_TYPE(bp, zp->zp_type); | |
1385 | BP_SET_LEVEL(bp, zp->zp_level); | |
1386 | BP_SET_BIRTH(bp, zio->io_txg, 0); | |
1387 | } | |
b128c09f BB |
1388 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
1389 | } else { | |
1390 | ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); | |
1391 | BP_SET_LSIZE(bp, lsize); | |
b0bc7a84 MG |
1392 | BP_SET_TYPE(bp, zp->zp_type); |
1393 | BP_SET_LEVEL(bp, zp->zp_level); | |
428870ff | 1394 | BP_SET_PSIZE(bp, psize); |
b128c09f BB |
1395 | BP_SET_COMPRESS(bp, compress); |
1396 | BP_SET_CHECKSUM(bp, zp->zp_checksum); | |
428870ff | 1397 | BP_SET_DEDUP(bp, zp->zp_dedup); |
b128c09f | 1398 | BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); |
428870ff BB |
1399 | if (zp->zp_dedup) { |
1400 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1401 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
1402 | zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; | |
1403 | } | |
03c6040b GW |
1404 | if (zp->zp_nopwrite) { |
1405 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1406 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
1407 | zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; | |
1408 | } | |
428870ff | 1409 | } |
428870ff BB |
1410 | return (ZIO_PIPELINE_CONTINUE); |
1411 | } | |
1412 | ||
1413 | static int | |
1414 | zio_free_bp_init(zio_t *zio) | |
1415 | { | |
1416 | blkptr_t *bp = zio->io_bp; | |
1417 | ||
1418 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
1419 | if (BP_GET_DEDUP(bp)) | |
1420 | zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; | |
b128c09f | 1421 | } |
34dc7c2f BB |
1422 | |
1423 | return (ZIO_PIPELINE_CONTINUE); | |
1424 | } | |
1425 | ||
b128c09f BB |
1426 | /* |
1427 | * ========================================================================== | |
1428 | * Execute the I/O pipeline | |
1429 | * ========================================================================== | |
1430 | */ | |
1431 | ||
1432 | static void | |
7ef5e54e | 1433 | zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) |
34dc7c2f | 1434 | { |
428870ff | 1435 | spa_t *spa = zio->io_spa; |
b128c09f | 1436 | zio_type_t t = zio->io_type; |
a38718a6 | 1437 | int flags = (cutinline ? TQ_FRONT : 0); |
34dc7c2f BB |
1438 | |
1439 | /* | |
9babb374 BB |
1440 | * If we're a config writer or a probe, the normal issue and |
1441 | * interrupt threads may all be blocked waiting for the config lock. | |
1442 | * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. | |
34dc7c2f | 1443 | */ |
9babb374 | 1444 | if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) |
b128c09f | 1445 | t = ZIO_TYPE_NULL; |
34dc7c2f BB |
1446 | |
1447 | /* | |
b128c09f | 1448 | * A similar issue exists for the L2ARC write thread until L2ARC 2.0. |
34dc7c2f | 1449 | */ |
b128c09f BB |
1450 | if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) |
1451 | t = ZIO_TYPE_NULL; | |
34dc7c2f | 1452 | |
428870ff | 1453 | /* |
7ef5e54e AL |
1454 | * If this is a high priority I/O, then use the high priority taskq if |
1455 | * available. | |
428870ff BB |
1456 | */ |
1457 | if (zio->io_priority == ZIO_PRIORITY_NOW && | |
7ef5e54e | 1458 | spa->spa_zio_taskq[t][q + 1].stqs_count != 0) |
428870ff BB |
1459 | q++; |
1460 | ||
1461 | ASSERT3U(q, <, ZIO_TASKQ_TYPES); | |
5cc556b4 | 1462 | |
a38718a6 GA |
1463 | /* |
1464 | * NB: We are assuming that the zio can only be dispatched | |
1465 | * to a single taskq at a time. It would be a grievous error | |
1466 | * to dispatch the zio to another taskq at the same time. | |
1467 | */ | |
1468 | ASSERT(taskq_empty_ent(&zio->io_tqent)); | |
7ef5e54e AL |
1469 | spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, |
1470 | flags, &zio->io_tqent); | |
b128c09f | 1471 | } |
34dc7c2f | 1472 | |
b128c09f | 1473 | static boolean_t |
7ef5e54e | 1474 | zio_taskq_member(zio_t *zio, zio_taskq_type_t q) |
b128c09f BB |
1475 | { |
1476 | kthread_t *executor = zio->io_executor; | |
1477 | spa_t *spa = zio->io_spa; | |
d6320ddb | 1478 | zio_type_t t; |
34dc7c2f | 1479 | |
7ef5e54e AL |
1480 | for (t = 0; t < ZIO_TYPES; t++) { |
1481 | spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; | |
1482 | uint_t i; | |
1483 | for (i = 0; i < tqs->stqs_count; i++) { | |
1484 | if (taskq_member(tqs->stqs_taskq[i], executor)) | |
1485 | return (B_TRUE); | |
1486 | } | |
1487 | } | |
34dc7c2f | 1488 | |
b128c09f BB |
1489 | return (B_FALSE); |
1490 | } | |
34dc7c2f | 1491 | |
b128c09f BB |
1492 | static int |
1493 | zio_issue_async(zio_t *zio) | |
1494 | { | |
428870ff | 1495 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); |
b128c09f BB |
1496 | |
1497 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f BB |
1498 | } |
1499 | ||
b128c09f BB |
1500 | void |
1501 | zio_interrupt(zio_t *zio) | |
34dc7c2f | 1502 | { |
428870ff | 1503 | zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); |
b128c09f | 1504 | } |
34dc7c2f | 1505 | |
26ef0cc7 TH |
1506 | void |
1507 | zio_delay_interrupt(zio_t *zio) | |
1508 | { | |
1509 | /* | |
1510 | * The timeout_generic() function isn't defined in userspace, so | |
1511 | * rather than trying to implement the function, the zio delay | |
1512 | * functionality has been disabled for userspace builds. | |
1513 | */ | |
1514 | ||
1515 | #ifdef _KERNEL | |
1516 | /* | |
1517 | * If io_target_timestamp is zero, then no delay has been registered | |
1518 | * for this IO, thus jump to the end of this function and "skip" the | |
1519 | * delay; issuing it directly to the zio layer. | |
1520 | */ | |
1521 | if (zio->io_target_timestamp != 0) { | |
1522 | hrtime_t now = gethrtime(); | |
1523 | ||
1524 | if (now >= zio->io_target_timestamp) { | |
1525 | /* | |
1526 | * This IO has already taken longer than the target | |
1527 | * delay to complete, so we don't want to delay it | |
1528 | * any longer; we "miss" the delay and issue it | |
1529 | * directly to the zio layer. This is likely due to | |
1530 | * the target latency being set to a value less than | |
1531 | * the underlying hardware can satisfy (e.g. delay | |
1532 | * set to 1ms, but the disks take 10ms to complete an | |
1533 | * IO request). | |
1534 | */ | |
1535 | ||
1536 | DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, | |
1537 | hrtime_t, now); | |
1538 | ||
1539 | zio_interrupt(zio); | |
1540 | } else { | |
1541 | taskqid_t tid; | |
1542 | hrtime_t diff = zio->io_target_timestamp - now; | |
1543 | clock_t expire_at_tick = ddi_get_lbolt() + | |
1544 | NSEC_TO_TICK(diff); | |
1545 | ||
1546 | DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, | |
1547 | hrtime_t, now, hrtime_t, diff); | |
1548 | ||
1549 | if (NSEC_TO_TICK(diff) == 0) { | |
1550 | /* Our delay is less than a jiffy - just spin */ | |
1551 | zfs_sleep_until(zio->io_target_timestamp); | |
1552 | } else { | |
1553 | /* | |
1554 | * Use taskq_dispatch_delay() in the place of | |
1555 | * OpenZFS's timeout_generic(). | |
1556 | */ | |
1557 | tid = taskq_dispatch_delay(system_taskq, | |
1558 | (task_func_t *) zio_interrupt, | |
1559 | zio, TQ_NOSLEEP, expire_at_tick); | |
1560 | if (!tid) { | |
1561 | /* | |
1562 | * Couldn't allocate a task. Just | |
1563 | * finish the zio without a delay. | |
1564 | */ | |
1565 | zio_interrupt(zio); | |
1566 | } | |
1567 | } | |
1568 | } | |
1569 | return; | |
1570 | } | |
1571 | #endif | |
1572 | DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); | |
1573 | zio_interrupt(zio); | |
1574 | } | |
1575 | ||
b128c09f BB |
1576 | /* |
1577 | * Execute the I/O pipeline until one of the following occurs: | |
1578 | * (1) the I/O completes; (2) the pipeline stalls waiting for | |
1579 | * dependent child I/Os; (3) the I/O issues, so we're waiting | |
1580 | * for an I/O completion interrupt; (4) the I/O is delegated by | |
1581 | * vdev-level caching or aggregation; (5) the I/O is deferred | |
1582 | * due to vdev-level queueing; (6) the I/O is handed off to | |
1583 | * another thread. In all cases, the pipeline stops whenever | |
8e07b99b | 1584 | * there's no CPU work; it never burns a thread in cv_wait_io(). |
b128c09f BB |
1585 | * |
1586 | * There's no locking on io_stage because there's no legitimate way | |
1587 | * for multiple threads to be attempting to process the same I/O. | |
1588 | */ | |
428870ff | 1589 | static zio_pipe_stage_t *zio_pipeline[]; |
34dc7c2f | 1590 | |
da6b4005 NB |
1591 | /* |
1592 | * zio_execute() is a wrapper around the static function | |
1593 | * __zio_execute() so that we can force __zio_execute() to be | |
1594 | * inlined. This reduces stack overhead which is important | |
1595 | * because __zio_execute() is called recursively in several zio | |
1596 | * code paths. zio_execute() itself cannot be inlined because | |
1597 | * it is externally visible. | |
1598 | */ | |
b128c09f BB |
1599 | void |
1600 | zio_execute(zio_t *zio) | |
da6b4005 | 1601 | { |
92119cc2 BB |
1602 | fstrans_cookie_t cookie; |
1603 | ||
1604 | cookie = spl_fstrans_mark(); | |
da6b4005 | 1605 | __zio_execute(zio); |
92119cc2 | 1606 | spl_fstrans_unmark(cookie); |
da6b4005 NB |
1607 | } |
1608 | ||
b58986ee BB |
1609 | /* |
1610 | * Used to determine if in the current context the stack is sized large | |
1611 | * enough to allow zio_execute() to be called recursively. A minimum | |
1612 | * stack size of 16K is required to avoid needing to re-dispatch the zio. | |
1613 | */ | |
1614 | boolean_t | |
1615 | zio_execute_stack_check(zio_t *zio) | |
1616 | { | |
1617 | #if !defined(HAVE_LARGE_STACKS) | |
1618 | dsl_pool_t *dp = spa_get_dsl(zio->io_spa); | |
1619 | ||
1620 | /* Executing in txg_sync_thread() context. */ | |
1621 | if (dp && curthread == dp->dp_tx.tx_sync_thread) | |
1622 | return (B_TRUE); | |
1623 | ||
1624 | /* Pool initialization outside of zio_taskq context. */ | |
1625 | if (dp && spa_is_initializing(dp->dp_spa) && | |
1626 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) && | |
1627 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH)) | |
1628 | return (B_TRUE); | |
1629 | #endif /* HAVE_LARGE_STACKS */ | |
1630 | ||
1631 | return (B_FALSE); | |
1632 | } | |
1633 | ||
da6b4005 NB |
1634 | __attribute__((always_inline)) |
1635 | static inline void | |
1636 | __zio_execute(zio_t *zio) | |
b128c09f BB |
1637 | { |
1638 | zio->io_executor = curthread; | |
34dc7c2f | 1639 | |
3dfb57a3 DB |
1640 | ASSERT3U(zio->io_queued_timestamp, >, 0); |
1641 | ||
b128c09f | 1642 | while (zio->io_stage < ZIO_STAGE_DONE) { |
428870ff BB |
1643 | enum zio_stage pipeline = zio->io_pipeline; |
1644 | enum zio_stage stage = zio->io_stage; | |
b128c09f | 1645 | int rv; |
34dc7c2f | 1646 | |
b128c09f | 1647 | ASSERT(!MUTEX_HELD(&zio->io_lock)); |
428870ff BB |
1648 | ASSERT(ISP2(stage)); |
1649 | ASSERT(zio->io_stall == NULL); | |
34dc7c2f | 1650 | |
428870ff BB |
1651 | do { |
1652 | stage <<= 1; | |
1653 | } while ((stage & pipeline) == 0); | |
b128c09f BB |
1654 | |
1655 | ASSERT(stage <= ZIO_STAGE_DONE); | |
34dc7c2f BB |
1656 | |
1657 | /* | |
b128c09f BB |
1658 | * If we are in interrupt context and this pipeline stage |
1659 | * will grab a config lock that is held across I/O, | |
428870ff BB |
1660 | * or may wait for an I/O that needs an interrupt thread |
1661 | * to complete, issue async to avoid deadlock. | |
1662 | * | |
1663 | * For VDEV_IO_START, we cut in line so that the io will | |
1664 | * be sent to disk promptly. | |
34dc7c2f | 1665 | */ |
91579709 BB |
1666 | if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && |
1667 | zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { | |
b58986ee BB |
1668 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? |
1669 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
91579709 BB |
1670 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
1671 | return; | |
1672 | } | |
1673 | ||
1674 | /* | |
b58986ee BB |
1675 | * If the current context doesn't have large enough stacks |
1676 | * the zio must be issued asynchronously to prevent overflow. | |
91579709 | 1677 | */ |
b58986ee BB |
1678 | if (zio_execute_stack_check(zio)) { |
1679 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? | |
1680 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
428870ff | 1681 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
b128c09f | 1682 | return; |
34dc7c2f BB |
1683 | } |
1684 | ||
b128c09f | 1685 | zio->io_stage = stage; |
3dfb57a3 | 1686 | zio->io_pipeline_trace |= zio->io_stage; |
9bd274dd | 1687 | rv = zio_pipeline[highbit64(stage) - 1](zio); |
34dc7c2f | 1688 | |
b128c09f BB |
1689 | if (rv == ZIO_PIPELINE_STOP) |
1690 | return; | |
34dc7c2f | 1691 | |
b128c09f BB |
1692 | ASSERT(rv == ZIO_PIPELINE_CONTINUE); |
1693 | } | |
34dc7c2f BB |
1694 | } |
1695 | ||
da6b4005 | 1696 | |
b128c09f BB |
1697 | /* |
1698 | * ========================================================================== | |
1699 | * Initiate I/O, either sync or async | |
1700 | * ========================================================================== | |
1701 | */ | |
1702 | int | |
1703 | zio_wait(zio_t *zio) | |
34dc7c2f | 1704 | { |
b128c09f | 1705 | int error; |
34dc7c2f | 1706 | |
b128c09f BB |
1707 | ASSERT(zio->io_stage == ZIO_STAGE_OPEN); |
1708 | ASSERT(zio->io_executor == NULL); | |
34dc7c2f | 1709 | |
b128c09f | 1710 | zio->io_waiter = curthread; |
3dfb57a3 DB |
1711 | ASSERT0(zio->io_queued_timestamp); |
1712 | zio->io_queued_timestamp = gethrtime(); | |
34dc7c2f | 1713 | |
da6b4005 | 1714 | __zio_execute(zio); |
34dc7c2f | 1715 | |
b128c09f | 1716 | mutex_enter(&zio->io_lock); |
72f53c56 | 1717 | while (zio->io_executor != NULL) |
72938d69 | 1718 | cv_wait_io(&zio->io_cv, &zio->io_lock); |
b128c09f | 1719 | mutex_exit(&zio->io_lock); |
34dc7c2f | 1720 | |
b128c09f BB |
1721 | error = zio->io_error; |
1722 | zio_destroy(zio); | |
34dc7c2f | 1723 | |
b128c09f BB |
1724 | return (error); |
1725 | } | |
34dc7c2f | 1726 | |
b128c09f BB |
1727 | void |
1728 | zio_nowait(zio_t *zio) | |
1729 | { | |
1730 | ASSERT(zio->io_executor == NULL); | |
34dc7c2f | 1731 | |
d164b209 BB |
1732 | if (zio->io_child_type == ZIO_CHILD_LOGICAL && |
1733 | zio_unique_parent(zio) == NULL) { | |
8878261f BB |
1734 | zio_t *pio; |
1735 | ||
34dc7c2f | 1736 | /* |
b128c09f | 1737 | * This is a logical async I/O with no parent to wait for it. |
9babb374 BB |
1738 | * We add it to the spa_async_root_zio "Godfather" I/O which |
1739 | * will ensure they complete prior to unloading the pool. | |
34dc7c2f | 1740 | */ |
b128c09f | 1741 | spa_t *spa = zio->io_spa; |
8878261f BB |
1742 | kpreempt_disable(); |
1743 | pio = spa->spa_async_zio_root[CPU_SEQID]; | |
1744 | kpreempt_enable(); | |
9babb374 | 1745 | |
8878261f | 1746 | zio_add_child(pio, zio); |
b128c09f | 1747 | } |
34dc7c2f | 1748 | |
3dfb57a3 DB |
1749 | ASSERT0(zio->io_queued_timestamp); |
1750 | zio->io_queued_timestamp = gethrtime(); | |
da6b4005 | 1751 | __zio_execute(zio); |
b128c09f | 1752 | } |
34dc7c2f | 1753 | |
b128c09f BB |
1754 | /* |
1755 | * ========================================================================== | |
1756 | * Reexecute or suspend/resume failed I/O | |
1757 | * ========================================================================== | |
1758 | */ | |
34dc7c2f | 1759 | |
b128c09f BB |
1760 | static void |
1761 | zio_reexecute(zio_t *pio) | |
1762 | { | |
d164b209 | 1763 | zio_t *cio, *cio_next; |
d6320ddb | 1764 | int c, w; |
3dfb57a3 | 1765 | zio_link_t *zl = NULL; |
d164b209 BB |
1766 | |
1767 | ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); | |
1768 | ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); | |
9babb374 BB |
1769 | ASSERT(pio->io_gang_leader == NULL); |
1770 | ASSERT(pio->io_gang_tree == NULL); | |
34dc7c2f | 1771 | |
b128c09f BB |
1772 | pio->io_flags = pio->io_orig_flags; |
1773 | pio->io_stage = pio->io_orig_stage; | |
1774 | pio->io_pipeline = pio->io_orig_pipeline; | |
1775 | pio->io_reexecute = 0; | |
03c6040b | 1776 | pio->io_flags |= ZIO_FLAG_REEXECUTED; |
3dfb57a3 | 1777 | pio->io_pipeline_trace = 0; |
b128c09f | 1778 | pio->io_error = 0; |
d6320ddb | 1779 | for (w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 | 1780 | pio->io_state[w] = 0; |
d6320ddb | 1781 | for (c = 0; c < ZIO_CHILD_TYPES; c++) |
b128c09f | 1782 | pio->io_child_error[c] = 0; |
34dc7c2f | 1783 | |
428870ff BB |
1784 | if (IO_IS_ALLOCATING(pio)) |
1785 | BP_ZERO(pio->io_bp); | |
34dc7c2f | 1786 | |
b128c09f BB |
1787 | /* |
1788 | * As we reexecute pio's children, new children could be created. | |
d164b209 | 1789 | * New children go to the head of pio's io_child_list, however, |
b128c09f | 1790 | * so we will (correctly) not reexecute them. The key is that |
d164b209 BB |
1791 | * the remainder of pio's io_child_list, from 'cio_next' onward, |
1792 | * cannot be affected by any side effects of reexecuting 'cio'. | |
b128c09f | 1793 | */ |
3dfb57a3 DB |
1794 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { |
1795 | cio_next = zio_walk_children(pio, &zl); | |
b128c09f | 1796 | mutex_enter(&pio->io_lock); |
d6320ddb | 1797 | for (w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 | 1798 | pio->io_children[cio->io_child_type][w]++; |
b128c09f | 1799 | mutex_exit(&pio->io_lock); |
d164b209 | 1800 | zio_reexecute(cio); |
34dc7c2f | 1801 | } |
34dc7c2f | 1802 | |
b128c09f BB |
1803 | /* |
1804 | * Now that all children have been reexecuted, execute the parent. | |
9babb374 BB |
1805 | * We don't reexecute "The Godfather" I/O here as it's the |
1806 | * responsibility of the caller to wait on him. | |
b128c09f | 1807 | */ |
3dfb57a3 DB |
1808 | if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) { |
1809 | pio->io_queued_timestamp = gethrtime(); | |
da6b4005 | 1810 | __zio_execute(pio); |
3dfb57a3 | 1811 | } |
34dc7c2f BB |
1812 | } |
1813 | ||
b128c09f BB |
1814 | void |
1815 | zio_suspend(spa_t *spa, zio_t *zio) | |
34dc7c2f | 1816 | { |
b128c09f BB |
1817 | if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) |
1818 | fm_panic("Pool '%s' has encountered an uncorrectable I/O " | |
1819 | "failure and the failure mode property for this pool " | |
1820 | "is set to panic.", spa_name(spa)); | |
34dc7c2f | 1821 | |
bf89c199 BB |
1822 | cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O " |
1823 | "failure and has been suspended.\n", spa_name(spa)); | |
1824 | ||
b128c09f | 1825 | zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); |
34dc7c2f | 1826 | |
b128c09f | 1827 | mutex_enter(&spa->spa_suspend_lock); |
34dc7c2f | 1828 | |
b128c09f | 1829 | if (spa->spa_suspend_zio_root == NULL) |
9babb374 BB |
1830 | spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, |
1831 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
1832 | ZIO_FLAG_GODFATHER); | |
34dc7c2f | 1833 | |
b128c09f | 1834 | spa->spa_suspended = B_TRUE; |
34dc7c2f | 1835 | |
b128c09f | 1836 | if (zio != NULL) { |
9babb374 | 1837 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f BB |
1838 | ASSERT(zio != spa->spa_suspend_zio_root); |
1839 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
d164b209 | 1840 | ASSERT(zio_unique_parent(zio) == NULL); |
b128c09f BB |
1841 | ASSERT(zio->io_stage == ZIO_STAGE_DONE); |
1842 | zio_add_child(spa->spa_suspend_zio_root, zio); | |
1843 | } | |
34dc7c2f | 1844 | |
b128c09f BB |
1845 | mutex_exit(&spa->spa_suspend_lock); |
1846 | } | |
34dc7c2f | 1847 | |
9babb374 | 1848 | int |
b128c09f BB |
1849 | zio_resume(spa_t *spa) |
1850 | { | |
9babb374 | 1851 | zio_t *pio; |
34dc7c2f BB |
1852 | |
1853 | /* | |
b128c09f | 1854 | * Reexecute all previously suspended i/o. |
34dc7c2f | 1855 | */ |
b128c09f BB |
1856 | mutex_enter(&spa->spa_suspend_lock); |
1857 | spa->spa_suspended = B_FALSE; | |
1858 | cv_broadcast(&spa->spa_suspend_cv); | |
1859 | pio = spa->spa_suspend_zio_root; | |
1860 | spa->spa_suspend_zio_root = NULL; | |
1861 | mutex_exit(&spa->spa_suspend_lock); | |
1862 | ||
1863 | if (pio == NULL) | |
9babb374 | 1864 | return (0); |
34dc7c2f | 1865 | |
9babb374 BB |
1866 | zio_reexecute(pio); |
1867 | return (zio_wait(pio)); | |
b128c09f BB |
1868 | } |
1869 | ||
1870 | void | |
1871 | zio_resume_wait(spa_t *spa) | |
1872 | { | |
1873 | mutex_enter(&spa->spa_suspend_lock); | |
1874 | while (spa_suspended(spa)) | |
1875 | cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); | |
1876 | mutex_exit(&spa->spa_suspend_lock); | |
34dc7c2f BB |
1877 | } |
1878 | ||
1879 | /* | |
1880 | * ========================================================================== | |
b128c09f BB |
1881 | * Gang blocks. |
1882 | * | |
1883 | * A gang block is a collection of small blocks that looks to the DMU | |
1884 | * like one large block. When zio_dva_allocate() cannot find a block | |
1885 | * of the requested size, due to either severe fragmentation or the pool | |
1886 | * being nearly full, it calls zio_write_gang_block() to construct the | |
1887 | * block from smaller fragments. | |
1888 | * | |
1889 | * A gang block consists of a gang header (zio_gbh_phys_t) and up to | |
1890 | * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like | |
1891 | * an indirect block: it's an array of block pointers. It consumes | |
1892 | * only one sector and hence is allocatable regardless of fragmentation. | |
1893 | * The gang header's bps point to its gang members, which hold the data. | |
1894 | * | |
1895 | * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> | |
1896 | * as the verifier to ensure uniqueness of the SHA256 checksum. | |
1897 | * Critically, the gang block bp's blk_cksum is the checksum of the data, | |
1898 | * not the gang header. This ensures that data block signatures (needed for | |
1899 | * deduplication) are independent of how the block is physically stored. | |
1900 | * | |
1901 | * Gang blocks can be nested: a gang member may itself be a gang block. | |
1902 | * Thus every gang block is a tree in which root and all interior nodes are | |
1903 | * gang headers, and the leaves are normal blocks that contain user data. | |
1904 | * The root of the gang tree is called the gang leader. | |
1905 | * | |
1906 | * To perform any operation (read, rewrite, free, claim) on a gang block, | |
1907 | * zio_gang_assemble() first assembles the gang tree (minus data leaves) | |
1908 | * in the io_gang_tree field of the original logical i/o by recursively | |
1909 | * reading the gang leader and all gang headers below it. This yields | |
1910 | * an in-core tree containing the contents of every gang header and the | |
1911 | * bps for every constituent of the gang block. | |
1912 | * | |
1913 | * With the gang tree now assembled, zio_gang_issue() just walks the gang tree | |
1914 | * and invokes a callback on each bp. To free a gang block, zio_gang_issue() | |
1915 | * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. | |
1916 | * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). | |
1917 | * zio_read_gang() is a wrapper around zio_read() that omits reading gang | |
1918 | * headers, since we already have those in io_gang_tree. zio_rewrite_gang() | |
1919 | * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() | |
1920 | * of the gang header plus zio_checksum_compute() of the data to update the | |
1921 | * gang header's blk_cksum as described above. | |
1922 | * | |
1923 | * The two-phase assemble/issue model solves the problem of partial failure -- | |
1924 | * what if you'd freed part of a gang block but then couldn't read the | |
1925 | * gang header for another part? Assembling the entire gang tree first | |
1926 | * ensures that all the necessary gang header I/O has succeeded before | |
1927 | * starting the actual work of free, claim, or write. Once the gang tree | |
1928 | * is assembled, free and claim are in-memory operations that cannot fail. | |
1929 | * | |
1930 | * In the event that a gang write fails, zio_dva_unallocate() walks the | |
1931 | * gang tree to immediately free (i.e. insert back into the space map) | |
1932 | * everything we've allocated. This ensures that we don't get ENOSPC | |
1933 | * errors during repeated suspend/resume cycles due to a flaky device. | |
1934 | * | |
1935 | * Gang rewrites only happen during sync-to-convergence. If we can't assemble | |
1936 | * the gang tree, we won't modify the block, so we can safely defer the free | |
1937 | * (knowing that the block is still intact). If we *can* assemble the gang | |
1938 | * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free | |
1939 | * each constituent bp and we can allocate a new block on the next sync pass. | |
1940 | * | |
1941 | * In all cases, the gang tree allows complete recovery from partial failure. | |
34dc7c2f BB |
1942 | * ========================================================================== |
1943 | */ | |
b128c09f BB |
1944 | |
1945 | static zio_t * | |
1946 | zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) | |
34dc7c2f | 1947 | { |
b128c09f BB |
1948 | if (gn != NULL) |
1949 | return (pio); | |
34dc7c2f | 1950 | |
b128c09f BB |
1951 | return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp), |
1952 | NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
1953 | &pio->io_bookmark)); | |
1954 | } | |
1955 | ||
1956 | zio_t * | |
1957 | zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) | |
1958 | { | |
1959 | zio_t *zio; | |
1960 | ||
1961 | if (gn != NULL) { | |
1962 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, | |
1963 | gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority, | |
1964 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); | |
34dc7c2f | 1965 | /* |
b128c09f BB |
1966 | * As we rewrite each gang header, the pipeline will compute |
1967 | * a new gang block header checksum for it; but no one will | |
1968 | * compute a new data checksum, so we do that here. The one | |
1969 | * exception is the gang leader: the pipeline already computed | |
1970 | * its data checksum because that stage precedes gang assembly. | |
1971 | * (Presently, nothing actually uses interior data checksums; | |
1972 | * this is just good hygiene.) | |
34dc7c2f | 1973 | */ |
9babb374 | 1974 | if (gn != pio->io_gang_leader->io_gang_tree) { |
b128c09f BB |
1975 | zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), |
1976 | data, BP_GET_PSIZE(bp)); | |
1977 | } | |
428870ff BB |
1978 | /* |
1979 | * If we are here to damage data for testing purposes, | |
1980 | * leave the GBH alone so that we can detect the damage. | |
1981 | */ | |
1982 | if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) | |
1983 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; | |
34dc7c2f | 1984 | } else { |
b128c09f BB |
1985 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
1986 | data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority, | |
1987 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); | |
34dc7c2f BB |
1988 | } |
1989 | ||
b128c09f BB |
1990 | return (zio); |
1991 | } | |
34dc7c2f | 1992 | |
b128c09f BB |
1993 | /* ARGSUSED */ |
1994 | zio_t * | |
1995 | zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) | |
1996 | { | |
428870ff BB |
1997 | return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, |
1998 | ZIO_GANG_CHILD_FLAGS(pio))); | |
34dc7c2f BB |
1999 | } |
2000 | ||
b128c09f BB |
2001 | /* ARGSUSED */ |
2002 | zio_t * | |
2003 | zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) | |
34dc7c2f | 2004 | { |
b128c09f BB |
2005 | return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, |
2006 | NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); | |
2007 | } | |
2008 | ||
2009 | static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { | |
2010 | NULL, | |
2011 | zio_read_gang, | |
2012 | zio_rewrite_gang, | |
2013 | zio_free_gang, | |
2014 | zio_claim_gang, | |
2015 | NULL | |
2016 | }; | |
34dc7c2f | 2017 | |
b128c09f | 2018 | static void zio_gang_tree_assemble_done(zio_t *zio); |
34dc7c2f | 2019 | |
b128c09f BB |
2020 | static zio_gang_node_t * |
2021 | zio_gang_node_alloc(zio_gang_node_t **gnpp) | |
2022 | { | |
2023 | zio_gang_node_t *gn; | |
34dc7c2f | 2024 | |
b128c09f | 2025 | ASSERT(*gnpp == NULL); |
34dc7c2f | 2026 | |
79c76d5b | 2027 | gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); |
b128c09f BB |
2028 | gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); |
2029 | *gnpp = gn; | |
34dc7c2f | 2030 | |
b128c09f | 2031 | return (gn); |
34dc7c2f BB |
2032 | } |
2033 | ||
34dc7c2f | 2034 | static void |
b128c09f | 2035 | zio_gang_node_free(zio_gang_node_t **gnpp) |
34dc7c2f | 2036 | { |
b128c09f | 2037 | zio_gang_node_t *gn = *gnpp; |
d6320ddb | 2038 | int g; |
34dc7c2f | 2039 | |
d6320ddb | 2040 | for (g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f BB |
2041 | ASSERT(gn->gn_child[g] == NULL); |
2042 | ||
2043 | zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
2044 | kmem_free(gn, sizeof (*gn)); | |
2045 | *gnpp = NULL; | |
34dc7c2f BB |
2046 | } |
2047 | ||
b128c09f BB |
2048 | static void |
2049 | zio_gang_tree_free(zio_gang_node_t **gnpp) | |
34dc7c2f | 2050 | { |
b128c09f | 2051 | zio_gang_node_t *gn = *gnpp; |
d6320ddb | 2052 | int g; |
34dc7c2f | 2053 | |
b128c09f BB |
2054 | if (gn == NULL) |
2055 | return; | |
34dc7c2f | 2056 | |
d6320ddb | 2057 | for (g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f | 2058 | zio_gang_tree_free(&gn->gn_child[g]); |
34dc7c2f | 2059 | |
b128c09f | 2060 | zio_gang_node_free(gnpp); |
34dc7c2f BB |
2061 | } |
2062 | ||
b128c09f | 2063 | static void |
9babb374 | 2064 | zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) |
34dc7c2f | 2065 | { |
b128c09f BB |
2066 | zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); |
2067 | ||
9babb374 | 2068 | ASSERT(gio->io_gang_leader == gio); |
b128c09f | 2069 | ASSERT(BP_IS_GANG(bp)); |
34dc7c2f | 2070 | |
9babb374 | 2071 | zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh, |
b128c09f | 2072 | SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn, |
9babb374 | 2073 | gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); |
b128c09f | 2074 | } |
34dc7c2f | 2075 | |
b128c09f BB |
2076 | static void |
2077 | zio_gang_tree_assemble_done(zio_t *zio) | |
2078 | { | |
9babb374 | 2079 | zio_t *gio = zio->io_gang_leader; |
b128c09f BB |
2080 | zio_gang_node_t *gn = zio->io_private; |
2081 | blkptr_t *bp = zio->io_bp; | |
d6320ddb | 2082 | int g; |
34dc7c2f | 2083 | |
9babb374 | 2084 | ASSERT(gio == zio_unique_parent(zio)); |
428870ff | 2085 | ASSERT(zio->io_child_count == 0); |
34dc7c2f | 2086 | |
b128c09f BB |
2087 | if (zio->io_error) |
2088 | return; | |
34dc7c2f | 2089 | |
b128c09f BB |
2090 | if (BP_SHOULD_BYTESWAP(bp)) |
2091 | byteswap_uint64_array(zio->io_data, zio->io_size); | |
34dc7c2f | 2092 | |
b128c09f BB |
2093 | ASSERT(zio->io_data == gn->gn_gbh); |
2094 | ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); | |
428870ff | 2095 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2096 | |
d6320ddb | 2097 | for (g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2098 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2099 | if (!BP_IS_GANG(gbp)) | |
2100 | continue; | |
9babb374 | 2101 | zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); |
b128c09f | 2102 | } |
34dc7c2f BB |
2103 | } |
2104 | ||
b128c09f BB |
2105 | static void |
2106 | zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data) | |
34dc7c2f | 2107 | { |
9babb374 | 2108 | zio_t *gio = pio->io_gang_leader; |
b128c09f | 2109 | zio_t *zio; |
d6320ddb | 2110 | int g; |
34dc7c2f | 2111 | |
b128c09f | 2112 | ASSERT(BP_IS_GANG(bp) == !!gn); |
9babb374 BB |
2113 | ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); |
2114 | ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); | |
34dc7c2f | 2115 | |
b128c09f BB |
2116 | /* |
2117 | * If you're a gang header, your data is in gn->gn_gbh. | |
2118 | * If you're a gang member, your data is in 'data' and gn == NULL. | |
2119 | */ | |
9babb374 | 2120 | zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data); |
34dc7c2f | 2121 | |
b128c09f | 2122 | if (gn != NULL) { |
428870ff | 2123 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2124 | |
d6320ddb | 2125 | for (g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2126 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2127 | if (BP_IS_HOLE(gbp)) | |
2128 | continue; | |
2129 | zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data); | |
2130 | data = (char *)data + BP_GET_PSIZE(gbp); | |
2131 | } | |
34dc7c2f BB |
2132 | } |
2133 | ||
9babb374 BB |
2134 | if (gn == gio->io_gang_tree) |
2135 | ASSERT3P((char *)gio->io_data + gio->io_size, ==, data); | |
34dc7c2f | 2136 | |
b128c09f BB |
2137 | if (zio != pio) |
2138 | zio_nowait(zio); | |
34dc7c2f BB |
2139 | } |
2140 | ||
2141 | static int | |
b128c09f | 2142 | zio_gang_assemble(zio_t *zio) |
34dc7c2f | 2143 | { |
b128c09f | 2144 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2145 | |
9babb374 BB |
2146 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); |
2147 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
2148 | ||
2149 | zio->io_gang_leader = zio; | |
34dc7c2f | 2150 | |
b128c09f | 2151 | zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); |
34dc7c2f BB |
2152 | |
2153 | return (ZIO_PIPELINE_CONTINUE); | |
2154 | } | |
2155 | ||
2156 | static int | |
b128c09f | 2157 | zio_gang_issue(zio_t *zio) |
34dc7c2f | 2158 | { |
b128c09f | 2159 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2160 | |
b128c09f BB |
2161 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) |
2162 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f | 2163 | |
9babb374 BB |
2164 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); |
2165 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
34dc7c2f | 2166 | |
b128c09f | 2167 | if (zio->io_child_error[ZIO_CHILD_GANG] == 0) |
9babb374 | 2168 | zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data); |
b128c09f | 2169 | else |
9babb374 | 2170 | zio_gang_tree_free(&zio->io_gang_tree); |
34dc7c2f | 2171 | |
b128c09f | 2172 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f BB |
2173 | |
2174 | return (ZIO_PIPELINE_CONTINUE); | |
2175 | } | |
2176 | ||
2177 | static void | |
b128c09f | 2178 | zio_write_gang_member_ready(zio_t *zio) |
34dc7c2f | 2179 | { |
d164b209 | 2180 | zio_t *pio = zio_unique_parent(zio); |
34dc7c2f BB |
2181 | dva_t *cdva = zio->io_bp->blk_dva; |
2182 | dva_t *pdva = pio->io_bp->blk_dva; | |
2183 | uint64_t asize; | |
d6320ddb | 2184 | int d; |
d1d7e268 | 2185 | ASSERTV(zio_t *gio = zio->io_gang_leader); |
34dc7c2f | 2186 | |
b128c09f BB |
2187 | if (BP_IS_HOLE(zio->io_bp)) |
2188 | return; | |
2189 | ||
2190 | ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); | |
2191 | ||
2192 | ASSERT(zio->io_child_type == ZIO_CHILD_GANG); | |
428870ff BB |
2193 | ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); |
2194 | ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); | |
2195 | ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); | |
34dc7c2f | 2196 | ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); |
34dc7c2f BB |
2197 | |
2198 | mutex_enter(&pio->io_lock); | |
d6320ddb | 2199 | for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { |
34dc7c2f BB |
2200 | ASSERT(DVA_GET_GANG(&pdva[d])); |
2201 | asize = DVA_GET_ASIZE(&pdva[d]); | |
2202 | asize += DVA_GET_ASIZE(&cdva[d]); | |
2203 | DVA_SET_ASIZE(&pdva[d], asize); | |
2204 | } | |
2205 | mutex_exit(&pio->io_lock); | |
2206 | } | |
2207 | ||
2208 | static int | |
b128c09f | 2209 | zio_write_gang_block(zio_t *pio) |
34dc7c2f | 2210 | { |
b128c09f | 2211 | spa_t *spa = pio->io_spa; |
3dfb57a3 | 2212 | metaslab_class_t *mc = spa_normal_class(spa); |
b128c09f | 2213 | blkptr_t *bp = pio->io_bp; |
9babb374 | 2214 | zio_t *gio = pio->io_gang_leader; |
b128c09f BB |
2215 | zio_t *zio; |
2216 | zio_gang_node_t *gn, **gnpp; | |
34dc7c2f | 2217 | zio_gbh_phys_t *gbh; |
b128c09f BB |
2218 | uint64_t txg = pio->io_txg; |
2219 | uint64_t resid = pio->io_size; | |
2220 | uint64_t lsize; | |
428870ff BB |
2221 | int copies = gio->io_prop.zp_copies; |
2222 | int gbh_copies = MIN(copies + 1, spa_max_replication(spa)); | |
b128c09f | 2223 | zio_prop_t zp; |
d6320ddb | 2224 | int g, error; |
34dc7c2f | 2225 | |
3dfb57a3 DB |
2226 | int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER; |
2227 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { | |
2228 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
2229 | ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); | |
2230 | ||
2231 | flags |= METASLAB_ASYNC_ALLOC; | |
2232 | VERIFY(refcount_held(&mc->mc_alloc_slots, pio)); | |
2233 | ||
2234 | /* | |
2235 | * The logical zio has already placed a reservation for | |
2236 | * 'copies' allocation slots but gang blocks may require | |
2237 | * additional copies. These additional copies | |
2238 | * (i.e. gbh_copies - copies) are guaranteed to succeed | |
2239 | * since metaslab_class_throttle_reserve() always allows | |
2240 | * additional reservations for gang blocks. | |
2241 | */ | |
2242 | VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies, | |
2243 | pio, flags)); | |
2244 | } | |
2245 | ||
2246 | error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE, | |
2247 | bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, pio); | |
34dc7c2f | 2248 | if (error) { |
3dfb57a3 DB |
2249 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
2250 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
2251 | ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); | |
2252 | ||
2253 | /* | |
2254 | * If we failed to allocate the gang block header then | |
2255 | * we remove any additional allocation reservations that | |
2256 | * we placed here. The original reservation will | |
2257 | * be removed when the logical I/O goes to the ready | |
2258 | * stage. | |
2259 | */ | |
2260 | metaslab_class_throttle_unreserve(mc, | |
2261 | gbh_copies - copies, pio); | |
2262 | } | |
2263 | ||
b128c09f | 2264 | pio->io_error = error; |
34dc7c2f BB |
2265 | return (ZIO_PIPELINE_CONTINUE); |
2266 | } | |
2267 | ||
9babb374 BB |
2268 | if (pio == gio) { |
2269 | gnpp = &gio->io_gang_tree; | |
b128c09f BB |
2270 | } else { |
2271 | gnpp = pio->io_private; | |
2272 | ASSERT(pio->io_ready == zio_write_gang_member_ready); | |
34dc7c2f BB |
2273 | } |
2274 | ||
b128c09f BB |
2275 | gn = zio_gang_node_alloc(gnpp); |
2276 | gbh = gn->gn_gbh; | |
2277 | bzero(gbh, SPA_GANGBLOCKSIZE); | |
34dc7c2f | 2278 | |
b128c09f BB |
2279 | /* |
2280 | * Create the gang header. | |
2281 | */ | |
2282 | zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL, | |
2283 | pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); | |
34dc7c2f | 2284 | |
b128c09f BB |
2285 | /* |
2286 | * Create and nowait the gang children. | |
2287 | */ | |
d6320ddb | 2288 | for (g = 0; resid != 0; resid -= lsize, g++) { |
3dfb57a3 DB |
2289 | zio_t *cio; |
2290 | ||
b128c09f BB |
2291 | lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), |
2292 | SPA_MINBLOCKSIZE); | |
2293 | ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); | |
2294 | ||
9babb374 | 2295 | zp.zp_checksum = gio->io_prop.zp_checksum; |
b128c09f BB |
2296 | zp.zp_compress = ZIO_COMPRESS_OFF; |
2297 | zp.zp_type = DMU_OT_NONE; | |
2298 | zp.zp_level = 0; | |
428870ff | 2299 | zp.zp_copies = gio->io_prop.zp_copies; |
03c6040b GW |
2300 | zp.zp_dedup = B_FALSE; |
2301 | zp.zp_dedup_verify = B_FALSE; | |
2302 | zp.zp_nopwrite = B_FALSE; | |
b128c09f | 2303 | |
3dfb57a3 | 2304 | cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g], |
2aa34383 DK |
2305 | (char *)pio->io_data + (pio->io_size - resid), lsize, |
2306 | lsize, &zp, zio_write_gang_member_ready, NULL, NULL, NULL, | |
bc77ba73 | 2307 | &gn->gn_child[g], pio->io_priority, |
3dfb57a3 DB |
2308 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); |
2309 | ||
2310 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { | |
2311 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
2312 | ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); | |
2313 | ||
2314 | /* | |
2315 | * Gang children won't throttle but we should | |
2316 | * account for their work, so reserve an allocation | |
2317 | * slot for them here. | |
2318 | */ | |
2319 | VERIFY(metaslab_class_throttle_reserve(mc, | |
2320 | zp.zp_copies, cio, flags)); | |
2321 | } | |
2322 | zio_nowait(cio); | |
2323 | ||
b128c09f | 2324 | } |
34dc7c2f BB |
2325 | |
2326 | /* | |
b128c09f | 2327 | * Set pio's pipeline to just wait for zio to finish. |
34dc7c2f | 2328 | */ |
b128c09f BB |
2329 | pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
2330 | ||
920dd524 ED |
2331 | /* |
2332 | * We didn't allocate this bp, so make sure it doesn't get unmarked. | |
2333 | */ | |
2334 | pio->io_flags &= ~ZIO_FLAG_FASTWRITE; | |
2335 | ||
b128c09f BB |
2336 | zio_nowait(zio); |
2337 | ||
2338 | return (ZIO_PIPELINE_CONTINUE); | |
34dc7c2f BB |
2339 | } |
2340 | ||
03c6040b | 2341 | /* |
3c67d83a TH |
2342 | * The zio_nop_write stage in the pipeline determines if allocating a |
2343 | * new bp is necessary. The nopwrite feature can handle writes in | |
2344 | * either syncing or open context (i.e. zil writes) and as a result is | |
2345 | * mutually exclusive with dedup. | |
2346 | * | |
2347 | * By leveraging a cryptographically secure checksum, such as SHA256, we | |
2348 | * can compare the checksums of the new data and the old to determine if | |
2349 | * allocating a new block is required. Note that our requirements for | |
2350 | * cryptographic strength are fairly weak: there can't be any accidental | |
2351 | * hash collisions, but we don't need to be secure against intentional | |
2352 | * (malicious) collisions. To trigger a nopwrite, you have to be able | |
2353 | * to write the file to begin with, and triggering an incorrect (hash | |
2354 | * collision) nopwrite is no worse than simply writing to the file. | |
2355 | * That said, there are no known attacks against the checksum algorithms | |
2356 | * used for nopwrite, assuming that the salt and the checksums | |
2357 | * themselves remain secret. | |
03c6040b GW |
2358 | */ |
2359 | static int | |
2360 | zio_nop_write(zio_t *zio) | |
2361 | { | |
2362 | blkptr_t *bp = zio->io_bp; | |
2363 | blkptr_t *bp_orig = &zio->io_bp_orig; | |
2364 | zio_prop_t *zp = &zio->io_prop; | |
2365 | ||
2366 | ASSERT(BP_GET_LEVEL(bp) == 0); | |
2367 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
2368 | ASSERT(zp->zp_nopwrite); | |
2369 | ASSERT(!zp->zp_dedup); | |
2370 | ASSERT(zio->io_bp_override == NULL); | |
2371 | ASSERT(IO_IS_ALLOCATING(zio)); | |
2372 | ||
2373 | /* | |
2374 | * Check to see if the original bp and the new bp have matching | |
2375 | * characteristics (i.e. same checksum, compression algorithms, etc). | |
2376 | * If they don't then just continue with the pipeline which will | |
2377 | * allocate a new bp. | |
2378 | */ | |
2379 | if (BP_IS_HOLE(bp_orig) || | |
3c67d83a TH |
2380 | !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & |
2381 | ZCHECKSUM_FLAG_NOPWRITE) || | |
03c6040b GW |
2382 | BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || |
2383 | BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || | |
2384 | BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || | |
2385 | zp->zp_copies != BP_GET_NDVAS(bp_orig)) | |
2386 | return (ZIO_PIPELINE_CONTINUE); | |
2387 | ||
2388 | /* | |
2389 | * If the checksums match then reset the pipeline so that we | |
2390 | * avoid allocating a new bp and issuing any I/O. | |
2391 | */ | |
2392 | if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { | |
3c67d83a TH |
2393 | ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & |
2394 | ZCHECKSUM_FLAG_NOPWRITE); | |
03c6040b GW |
2395 | ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); |
2396 | ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); | |
2397 | ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); | |
2398 | ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, | |
2399 | sizeof (uint64_t)) == 0); | |
2400 | ||
2401 | *bp = *bp_orig; | |
2402 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
2403 | zio->io_flags |= ZIO_FLAG_NOPWRITE; | |
2404 | } | |
2405 | ||
2406 | return (ZIO_PIPELINE_CONTINUE); | |
2407 | } | |
2408 | ||
34dc7c2f BB |
2409 | /* |
2410 | * ========================================================================== | |
428870ff | 2411 | * Dedup |
34dc7c2f BB |
2412 | * ========================================================================== |
2413 | */ | |
428870ff BB |
2414 | static void |
2415 | zio_ddt_child_read_done(zio_t *zio) | |
2416 | { | |
2417 | blkptr_t *bp = zio->io_bp; | |
2418 | ddt_entry_t *dde = zio->io_private; | |
2419 | ddt_phys_t *ddp; | |
2420 | zio_t *pio = zio_unique_parent(zio); | |
2421 | ||
2422 | mutex_enter(&pio->io_lock); | |
2423 | ddp = ddt_phys_select(dde, bp); | |
2424 | if (zio->io_error == 0) | |
2425 | ddt_phys_clear(ddp); /* this ddp doesn't need repair */ | |
2426 | if (zio->io_error == 0 && dde->dde_repair_data == NULL) | |
2427 | dde->dde_repair_data = zio->io_data; | |
2428 | else | |
2429 | zio_buf_free(zio->io_data, zio->io_size); | |
2430 | mutex_exit(&pio->io_lock); | |
2431 | } | |
2432 | ||
2433 | static int | |
2434 | zio_ddt_read_start(zio_t *zio) | |
2435 | { | |
2436 | blkptr_t *bp = zio->io_bp; | |
d6320ddb | 2437 | int p; |
428870ff BB |
2438 | |
2439 | ASSERT(BP_GET_DEDUP(bp)); | |
2440 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
2441 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
2442 | ||
2443 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
2444 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
2445 | ddt_entry_t *dde = ddt_repair_start(ddt, bp); | |
2446 | ddt_phys_t *ddp = dde->dde_phys; | |
2447 | ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); | |
2448 | blkptr_t blk; | |
2449 | ||
2450 | ASSERT(zio->io_vsd == NULL); | |
2451 | zio->io_vsd = dde; | |
2452 | ||
2453 | if (ddp_self == NULL) | |
2454 | return (ZIO_PIPELINE_CONTINUE); | |
2455 | ||
d6320ddb | 2456 | for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { |
428870ff BB |
2457 | if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) |
2458 | continue; | |
2459 | ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, | |
2460 | &blk); | |
2461 | zio_nowait(zio_read(zio, zio->io_spa, &blk, | |
2462 | zio_buf_alloc(zio->io_size), zio->io_size, | |
2463 | zio_ddt_child_read_done, dde, zio->io_priority, | |
2464 | ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE, | |
2465 | &zio->io_bookmark)); | |
2466 | } | |
2467 | return (ZIO_PIPELINE_CONTINUE); | |
2468 | } | |
2469 | ||
2470 | zio_nowait(zio_read(zio, zio->io_spa, bp, | |
2471 | zio->io_data, zio->io_size, NULL, NULL, zio->io_priority, | |
2472 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); | |
2473 | ||
2474 | return (ZIO_PIPELINE_CONTINUE); | |
2475 | } | |
2476 | ||
2477 | static int | |
2478 | zio_ddt_read_done(zio_t *zio) | |
2479 | { | |
2480 | blkptr_t *bp = zio->io_bp; | |
2481 | ||
2482 | if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE)) | |
2483 | return (ZIO_PIPELINE_STOP); | |
2484 | ||
2485 | ASSERT(BP_GET_DEDUP(bp)); | |
2486 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
2487 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
2488 | ||
2489 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
2490 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
2491 | ddt_entry_t *dde = zio->io_vsd; | |
2492 | if (ddt == NULL) { | |
2493 | ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); | |
2494 | return (ZIO_PIPELINE_CONTINUE); | |
2495 | } | |
2496 | if (dde == NULL) { | |
2497 | zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; | |
2498 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); | |
2499 | return (ZIO_PIPELINE_STOP); | |
2500 | } | |
2501 | if (dde->dde_repair_data != NULL) { | |
2502 | bcopy(dde->dde_repair_data, zio->io_data, zio->io_size); | |
2503 | zio->io_child_error[ZIO_CHILD_DDT] = 0; | |
2504 | } | |
2505 | ddt_repair_done(ddt, dde); | |
2506 | zio->io_vsd = NULL; | |
2507 | } | |
2508 | ||
2509 | ASSERT(zio->io_vsd == NULL); | |
2510 | ||
2511 | return (ZIO_PIPELINE_CONTINUE); | |
2512 | } | |
2513 | ||
2514 | static boolean_t | |
2515 | zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) | |
2516 | { | |
2517 | spa_t *spa = zio->io_spa; | |
d6320ddb | 2518 | int p; |
c17bcf83 | 2519 | boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW); |
428870ff | 2520 | |
c17bcf83 | 2521 | ASSERT(!(zio->io_bp_override && do_raw)); |
2aa34383 | 2522 | |
428870ff BB |
2523 | /* |
2524 | * Note: we compare the original data, not the transformed data, | |
2525 | * because when zio->io_bp is an override bp, we will not have | |
2526 | * pushed the I/O transforms. That's an important optimization | |
2527 | * because otherwise we'd compress/encrypt all dmu_sync() data twice. | |
c17bcf83 TC |
2528 | * However, we should never get a raw, override zio so in these |
2529 | * cases we can compare the io_data directly. This is useful because | |
2530 | * it allows us to do dedup verification even if we don't have access | |
2531 | * to the original data (for instance, if the encryption keys aren't | |
2532 | * loaded). | |
428870ff | 2533 | */ |
c17bcf83 | 2534 | |
d6320ddb | 2535 | for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
2536 | zio_t *lio = dde->dde_lead_zio[p]; |
2537 | ||
c17bcf83 TC |
2538 | if (lio != NULL && do_raw) { |
2539 | return (lio->io_size != zio->io_size || | |
2540 | bcmp(zio->io_data, lio->io_data, | |
2541 | zio->io_size) != 0); | |
2542 | } else if (lio != NULL) { | |
428870ff BB |
2543 | return (lio->io_orig_size != zio->io_orig_size || |
2544 | bcmp(zio->io_orig_data, lio->io_orig_data, | |
2545 | zio->io_orig_size) != 0); | |
2546 | } | |
2547 | } | |
2548 | ||
d6320ddb | 2549 | for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
2550 | ddt_phys_t *ddp = &dde->dde_phys[p]; |
2551 | ||
c17bcf83 TC |
2552 | if (ddp->ddp_phys_birth != 0 && do_raw) { |
2553 | blkptr_t blk = *zio->io_bp; | |
2554 | uint64_t psize; | |
2555 | void *tmpbuf; | |
2556 | int error; | |
2557 | ||
2558 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
2559 | psize = BP_GET_PSIZE(&blk); | |
2560 | ||
2561 | if (psize != zio->io_size) | |
2562 | return (B_TRUE); | |
2563 | ||
2564 | ddt_exit(ddt); | |
2565 | ||
2566 | tmpbuf = zio_buf_alloc(psize); | |
2567 | ||
2568 | error = zio_wait(zio_read(NULL, spa, &blk, tmpbuf, | |
2569 | psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ, | |
2570 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
2571 | ZIO_FLAG_RAW, &zio->io_bookmark)); | |
2572 | ||
2573 | if (error == 0) { | |
2574 | if (bcmp(tmpbuf, zio->io_data, psize) != 0) | |
2575 | error = SET_ERROR(ENOENT); | |
2576 | } | |
2577 | ||
2578 | zio_buf_free(tmpbuf, psize); | |
2579 | ddt_enter(ddt); | |
2580 | return (error != 0); | |
2581 | } else if (ddp->ddp_phys_birth != 0) { | |
428870ff | 2582 | arc_buf_t *abuf = NULL; |
2a432414 | 2583 | arc_flags_t aflags = ARC_FLAG_WAIT; |
428870ff BB |
2584 | blkptr_t blk = *zio->io_bp; |
2585 | int error; | |
2586 | ||
2587 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
2588 | ||
c17bcf83 TC |
2589 | if (BP_GET_LSIZE(&blk) != zio->io_orig_size) |
2590 | return (B_TRUE); | |
2591 | ||
428870ff BB |
2592 | ddt_exit(ddt); |
2593 | ||
294f6806 | 2594 | error = arc_read(NULL, spa, &blk, |
428870ff BB |
2595 | arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, |
2596 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
2597 | &aflags, &zio->io_bookmark); | |
2598 | ||
2599 | if (error == 0) { | |
c17bcf83 | 2600 | if (bcmp(abuf->b_data, zio->io_orig_data, |
428870ff | 2601 | zio->io_orig_size) != 0) |
c17bcf83 | 2602 | error = SET_ERROR(ENOENT); |
d3c2ae1c | 2603 | arc_buf_destroy(abuf, &abuf); |
428870ff BB |
2604 | } |
2605 | ||
2606 | ddt_enter(ddt); | |
2607 | return (error != 0); | |
2608 | } | |
2609 | } | |
2610 | ||
2611 | return (B_FALSE); | |
2612 | } | |
2613 | ||
2614 | static void | |
2615 | zio_ddt_child_write_ready(zio_t *zio) | |
2616 | { | |
2617 | int p = zio->io_prop.zp_copies; | |
2618 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
2619 | ddt_entry_t *dde = zio->io_private; | |
2620 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
2621 | zio_t *pio; | |
3dfb57a3 | 2622 | zio_link_t *zl; |
428870ff BB |
2623 | |
2624 | if (zio->io_error) | |
2625 | return; | |
2626 | ||
2627 | ddt_enter(ddt); | |
2628 | ||
2629 | ASSERT(dde->dde_lead_zio[p] == zio); | |
2630 | ||
2631 | ddt_phys_fill(ddp, zio->io_bp); | |
2632 | ||
3dfb57a3 DB |
2633 | zl = NULL; |
2634 | while ((pio = zio_walk_parents(zio, &zl)) != NULL) | |
428870ff BB |
2635 | ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); |
2636 | ||
2637 | ddt_exit(ddt); | |
2638 | } | |
2639 | ||
2640 | static void | |
2641 | zio_ddt_child_write_done(zio_t *zio) | |
2642 | { | |
2643 | int p = zio->io_prop.zp_copies; | |
2644 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
2645 | ddt_entry_t *dde = zio->io_private; | |
2646 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
2647 | ||
2648 | ddt_enter(ddt); | |
2649 | ||
2650 | ASSERT(ddp->ddp_refcnt == 0); | |
2651 | ASSERT(dde->dde_lead_zio[p] == zio); | |
2652 | dde->dde_lead_zio[p] = NULL; | |
2653 | ||
2654 | if (zio->io_error == 0) { | |
3dfb57a3 DB |
2655 | zio_link_t *zl = NULL; |
2656 | while (zio_walk_parents(zio, &zl) != NULL) | |
428870ff BB |
2657 | ddt_phys_addref(ddp); |
2658 | } else { | |
2659 | ddt_phys_clear(ddp); | |
2660 | } | |
2661 | ||
2662 | ddt_exit(ddt); | |
2663 | } | |
2664 | ||
2665 | static void | |
2666 | zio_ddt_ditto_write_done(zio_t *zio) | |
2667 | { | |
2668 | int p = DDT_PHYS_DITTO; | |
428870ff BB |
2669 | blkptr_t *bp = zio->io_bp; |
2670 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
2671 | ddt_entry_t *dde = zio->io_private; | |
2672 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
2673 | ddt_key_t *ddk = &dde->dde_key; | |
1fde1e37 | 2674 | ASSERTV(zio_prop_t *zp = &zio->io_prop); |
428870ff BB |
2675 | |
2676 | ddt_enter(ddt); | |
2677 | ||
2678 | ASSERT(ddp->ddp_refcnt == 0); | |
2679 | ASSERT(dde->dde_lead_zio[p] == zio); | |
2680 | dde->dde_lead_zio[p] = NULL; | |
2681 | ||
2682 | if (zio->io_error == 0) { | |
2683 | ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); | |
2684 | ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); | |
2685 | ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); | |
2686 | if (ddp->ddp_phys_birth != 0) | |
2687 | ddt_phys_free(ddt, ddk, ddp, zio->io_txg); | |
2688 | ddt_phys_fill(ddp, bp); | |
2689 | } | |
2690 | ||
2691 | ddt_exit(ddt); | |
2692 | } | |
2693 | ||
2694 | static int | |
2695 | zio_ddt_write(zio_t *zio) | |
2696 | { | |
2697 | spa_t *spa = zio->io_spa; | |
2698 | blkptr_t *bp = zio->io_bp; | |
2699 | uint64_t txg = zio->io_txg; | |
2700 | zio_prop_t *zp = &zio->io_prop; | |
2701 | int p = zp->zp_copies; | |
2702 | int ditto_copies; | |
2703 | zio_t *cio = NULL; | |
2704 | zio_t *dio = NULL; | |
2705 | ddt_t *ddt = ddt_select(spa, bp); | |
2706 | ddt_entry_t *dde; | |
2707 | ddt_phys_t *ddp; | |
2708 | ||
2709 | ASSERT(BP_GET_DEDUP(bp)); | |
2710 | ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); | |
2711 | ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); | |
c17bcf83 | 2712 | ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW))); |
428870ff BB |
2713 | |
2714 | ddt_enter(ddt); | |
2715 | dde = ddt_lookup(ddt, bp, B_TRUE); | |
2716 | ddp = &dde->dde_phys[p]; | |
2717 | ||
2718 | if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { | |
2719 | /* | |
2720 | * If we're using a weak checksum, upgrade to a strong checksum | |
2721 | * and try again. If we're already using a strong checksum, | |
2722 | * we can't resolve it, so just convert to an ordinary write. | |
2723 | * (And automatically e-mail a paper to Nature?) | |
2724 | */ | |
3c67d83a TH |
2725 | if (!(zio_checksum_table[zp->zp_checksum].ci_flags & |
2726 | ZCHECKSUM_FLAG_DEDUP)) { | |
428870ff BB |
2727 | zp->zp_checksum = spa_dedup_checksum(spa); |
2728 | zio_pop_transforms(zio); | |
2729 | zio->io_stage = ZIO_STAGE_OPEN; | |
2730 | BP_ZERO(bp); | |
2731 | } else { | |
03c6040b | 2732 | zp->zp_dedup = B_FALSE; |
428870ff BB |
2733 | } |
2734 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
2735 | ddt_exit(ddt); | |
2736 | return (ZIO_PIPELINE_CONTINUE); | |
2737 | } | |
2738 | ||
2739 | ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); | |
2740 | ASSERT(ditto_copies < SPA_DVAS_PER_BP); | |
2741 | ||
2742 | if (ditto_copies > ddt_ditto_copies_present(dde) && | |
2743 | dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { | |
2744 | zio_prop_t czp = *zp; | |
2745 | ||
2746 | czp.zp_copies = ditto_copies; | |
2747 | ||
2748 | /* | |
2749 | * If we arrived here with an override bp, we won't have run | |
2750 | * the transform stack, so we won't have the data we need to | |
2751 | * generate a child i/o. So, toss the override bp and restart. | |
2752 | * This is safe, because using the override bp is just an | |
2753 | * optimization; and it's rare, so the cost doesn't matter. | |
2754 | */ | |
2755 | if (zio->io_bp_override) { | |
2756 | zio_pop_transforms(zio); | |
2757 | zio->io_stage = ZIO_STAGE_OPEN; | |
2758 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
2759 | zio->io_bp_override = NULL; | |
2760 | BP_ZERO(bp); | |
2761 | ddt_exit(ddt); | |
2762 | return (ZIO_PIPELINE_CONTINUE); | |
2763 | } | |
2764 | ||
2765 | dio = zio_write(zio, spa, txg, bp, zio->io_orig_data, | |
2aa34383 | 2766 | zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL, |
bc77ba73 | 2767 | NULL, zio_ddt_ditto_write_done, dde, zio->io_priority, |
428870ff BB |
2768 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); |
2769 | ||
2770 | zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL); | |
2771 | dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; | |
2772 | } | |
2773 | ||
2774 | if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { | |
2775 | if (ddp->ddp_phys_birth != 0) | |
2776 | ddt_bp_fill(ddp, bp, txg); | |
2777 | if (dde->dde_lead_zio[p] != NULL) | |
2778 | zio_add_child(zio, dde->dde_lead_zio[p]); | |
2779 | else | |
2780 | ddt_phys_addref(ddp); | |
2781 | } else if (zio->io_bp_override) { | |
2782 | ASSERT(bp->blk_birth == txg); | |
2783 | ASSERT(BP_EQUAL(bp, zio->io_bp_override)); | |
2784 | ddt_phys_fill(ddp, bp); | |
2785 | ddt_phys_addref(ddp); | |
2786 | } else { | |
2787 | cio = zio_write(zio, spa, txg, bp, zio->io_orig_data, | |
2aa34383 | 2788 | zio->io_orig_size, zio->io_orig_size, zp, |
bc77ba73 | 2789 | zio_ddt_child_write_ready, NULL, NULL, |
428870ff BB |
2790 | zio_ddt_child_write_done, dde, zio->io_priority, |
2791 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); | |
2792 | ||
2793 | zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL); | |
2794 | dde->dde_lead_zio[p] = cio; | |
2795 | } | |
2796 | ||
2797 | ddt_exit(ddt); | |
2798 | ||
2799 | if (cio) | |
2800 | zio_nowait(cio); | |
2801 | if (dio) | |
2802 | zio_nowait(dio); | |
2803 | ||
2804 | return (ZIO_PIPELINE_CONTINUE); | |
2805 | } | |
2806 | ||
2807 | ddt_entry_t *freedde; /* for debugging */ | |
b128c09f | 2808 | |
428870ff BB |
2809 | static int |
2810 | zio_ddt_free(zio_t *zio) | |
2811 | { | |
2812 | spa_t *spa = zio->io_spa; | |
2813 | blkptr_t *bp = zio->io_bp; | |
2814 | ddt_t *ddt = ddt_select(spa, bp); | |
2815 | ddt_entry_t *dde; | |
2816 | ddt_phys_t *ddp; | |
2817 | ||
2818 | ASSERT(BP_GET_DEDUP(bp)); | |
2819 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
2820 | ||
2821 | ddt_enter(ddt); | |
2822 | freedde = dde = ddt_lookup(ddt, bp, B_TRUE); | |
5dc6af0e BB |
2823 | if (dde) { |
2824 | ddp = ddt_phys_select(dde, bp); | |
2825 | if (ddp) | |
2826 | ddt_phys_decref(ddp); | |
2827 | } | |
428870ff BB |
2828 | ddt_exit(ddt); |
2829 | ||
2830 | return (ZIO_PIPELINE_CONTINUE); | |
2831 | } | |
2832 | ||
2833 | /* | |
2834 | * ========================================================================== | |
2835 | * Allocate and free blocks | |
2836 | * ========================================================================== | |
2837 | */ | |
3dfb57a3 DB |
2838 | |
2839 | static zio_t * | |
2840 | zio_io_to_allocate(spa_t *spa) | |
2841 | { | |
2842 | zio_t *zio; | |
2843 | ||
2844 | ASSERT(MUTEX_HELD(&spa->spa_alloc_lock)); | |
2845 | ||
2846 | zio = avl_first(&spa->spa_alloc_tree); | |
2847 | if (zio == NULL) | |
2848 | return (NULL); | |
2849 | ||
2850 | ASSERT(IO_IS_ALLOCATING(zio)); | |
2851 | ||
2852 | /* | |
2853 | * Try to place a reservation for this zio. If we're unable to | |
2854 | * reserve then we throttle. | |
2855 | */ | |
2856 | if (!metaslab_class_throttle_reserve(spa_normal_class(spa), | |
2857 | zio->io_prop.zp_copies, zio, 0)) { | |
2858 | return (NULL); | |
2859 | } | |
2860 | ||
2861 | avl_remove(&spa->spa_alloc_tree, zio); | |
2862 | ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE); | |
2863 | ||
2864 | return (zio); | |
2865 | } | |
2866 | ||
2867 | static int | |
2868 | zio_dva_throttle(zio_t *zio) | |
2869 | { | |
2870 | spa_t *spa = zio->io_spa; | |
2871 | zio_t *nio; | |
2872 | ||
2873 | if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE || | |
2874 | !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled || | |
2875 | zio->io_child_type == ZIO_CHILD_GANG || | |
2876 | zio->io_flags & ZIO_FLAG_NODATA) { | |
2877 | return (ZIO_PIPELINE_CONTINUE); | |
2878 | } | |
2879 | ||
2880 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
2881 | ||
2882 | ASSERT3U(zio->io_queued_timestamp, >, 0); | |
2883 | ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE); | |
2884 | ||
2885 | mutex_enter(&spa->spa_alloc_lock); | |
2886 | ||
2887 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
2888 | avl_add(&spa->spa_alloc_tree, zio); | |
2889 | ||
2890 | nio = zio_io_to_allocate(zio->io_spa); | |
2891 | mutex_exit(&spa->spa_alloc_lock); | |
2892 | ||
2893 | if (nio == zio) | |
2894 | return (ZIO_PIPELINE_CONTINUE); | |
2895 | ||
2896 | if (nio != NULL) { | |
2897 | ASSERT3U(nio->io_queued_timestamp, <=, | |
2898 | zio->io_queued_timestamp); | |
2899 | ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE); | |
2900 | /* | |
2901 | * We are passing control to a new zio so make sure that | |
2902 | * it is processed by a different thread. We do this to | |
2903 | * avoid stack overflows that can occur when parents are | |
2904 | * throttled and children are making progress. We allow | |
2905 | * it to go to the head of the taskq since it's already | |
2906 | * been waiting. | |
2907 | */ | |
2908 | zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE); | |
2909 | } | |
2910 | return (ZIO_PIPELINE_STOP); | |
2911 | } | |
2912 | ||
2913 | void | |
2914 | zio_allocate_dispatch(spa_t *spa) | |
2915 | { | |
2916 | zio_t *zio; | |
2917 | ||
2918 | mutex_enter(&spa->spa_alloc_lock); | |
2919 | zio = zio_io_to_allocate(spa); | |
2920 | mutex_exit(&spa->spa_alloc_lock); | |
2921 | if (zio == NULL) | |
2922 | return; | |
2923 | ||
2924 | ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE); | |
2925 | ASSERT0(zio->io_error); | |
2926 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE); | |
2927 | } | |
2928 | ||
34dc7c2f BB |
2929 | static int |
2930 | zio_dva_allocate(zio_t *zio) | |
2931 | { | |
2932 | spa_t *spa = zio->io_spa; | |
428870ff | 2933 | metaslab_class_t *mc = spa_normal_class(spa); |
34dc7c2f BB |
2934 | blkptr_t *bp = zio->io_bp; |
2935 | int error; | |
6d974228 | 2936 | int flags = 0; |
34dc7c2f | 2937 | |
9babb374 BB |
2938 | if (zio->io_gang_leader == NULL) { |
2939 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
2940 | zio->io_gang_leader = zio; | |
2941 | } | |
2942 | ||
34dc7c2f | 2943 | ASSERT(BP_IS_HOLE(bp)); |
c99c9001 | 2944 | ASSERT0(BP_GET_NDVAS(bp)); |
428870ff BB |
2945 | ASSERT3U(zio->io_prop.zp_copies, >, 0); |
2946 | ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); | |
34dc7c2f BB |
2947 | ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); |
2948 | ||
920dd524 | 2949 | flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0; |
3dfb57a3 DB |
2950 | if (zio->io_flags & ZIO_FLAG_NODATA) |
2951 | flags |= METASLAB_DONT_THROTTLE; | |
2952 | if (zio->io_flags & ZIO_FLAG_GANG_CHILD) | |
2953 | flags |= METASLAB_GANG_CHILD; | |
2954 | if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) | |
2955 | flags |= METASLAB_ASYNC_ALLOC; | |
2956 | ||
b128c09f | 2957 | error = metaslab_alloc(spa, mc, zio->io_size, bp, |
3dfb57a3 | 2958 | zio->io_prop.zp_copies, zio->io_txg, NULL, flags, zio); |
34dc7c2f | 2959 | |
3dfb57a3 | 2960 | if (error != 0) { |
6d974228 GW |
2961 | spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, " |
2962 | "size %llu, error %d", spa_name(spa), zio, zio->io_size, | |
2963 | error); | |
b128c09f BB |
2964 | if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) |
2965 | return (zio_write_gang_block(zio)); | |
34dc7c2f BB |
2966 | zio->io_error = error; |
2967 | } | |
2968 | ||
2969 | return (ZIO_PIPELINE_CONTINUE); | |
2970 | } | |
2971 | ||
2972 | static int | |
2973 | zio_dva_free(zio_t *zio) | |
2974 | { | |
b128c09f | 2975 | metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); |
34dc7c2f BB |
2976 | |
2977 | return (ZIO_PIPELINE_CONTINUE); | |
2978 | } | |
2979 | ||
2980 | static int | |
2981 | zio_dva_claim(zio_t *zio) | |
2982 | { | |
b128c09f BB |
2983 | int error; |
2984 | ||
2985 | error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); | |
2986 | if (error) | |
2987 | zio->io_error = error; | |
34dc7c2f BB |
2988 | |
2989 | return (ZIO_PIPELINE_CONTINUE); | |
2990 | } | |
2991 | ||
b128c09f BB |
2992 | /* |
2993 | * Undo an allocation. This is used by zio_done() when an I/O fails | |
2994 | * and we want to give back the block we just allocated. | |
2995 | * This handles both normal blocks and gang blocks. | |
2996 | */ | |
2997 | static void | |
2998 | zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) | |
2999 | { | |
d6320ddb BB |
3000 | int g; |
3001 | ||
b128c09f | 3002 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); |
428870ff | 3003 | ASSERT(zio->io_bp_override == NULL); |
b128c09f BB |
3004 | |
3005 | if (!BP_IS_HOLE(bp)) | |
428870ff | 3006 | metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); |
b128c09f BB |
3007 | |
3008 | if (gn != NULL) { | |
d6320ddb | 3009 | for (g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
3010 | zio_dva_unallocate(zio, gn->gn_child[g], |
3011 | &gn->gn_gbh->zg_blkptr[g]); | |
3012 | } | |
3013 | } | |
3014 | } | |
3015 | ||
3016 | /* | |
3017 | * Try to allocate an intent log block. Return 0 on success, errno on failure. | |
3018 | */ | |
3019 | int | |
920dd524 ED |
3020 | zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, uint64_t size, |
3021 | boolean_t use_slog) | |
b128c09f | 3022 | { |
428870ff | 3023 | int error = 1; |
b128c09f | 3024 | |
428870ff BB |
3025 | ASSERT(txg > spa_syncing_txg(spa)); |
3026 | ||
ebf8e3a2 | 3027 | if (use_slog) { |
428870ff | 3028 | error = metaslab_alloc(spa, spa_log_class(spa), size, |
3dfb57a3 | 3029 | new_bp, 1, txg, NULL, METASLAB_FASTWRITE, NULL); |
ebf8e3a2 | 3030 | } |
b128c09f | 3031 | |
ebf8e3a2 | 3032 | if (error) { |
428870ff | 3033 | error = metaslab_alloc(spa, spa_normal_class(spa), size, |
3dfb57a3 | 3034 | new_bp, 1, txg, NULL, METASLAB_FASTWRITE, NULL); |
ebf8e3a2 | 3035 | } |
b128c09f BB |
3036 | |
3037 | if (error == 0) { | |
3038 | BP_SET_LSIZE(new_bp, size); | |
3039 | BP_SET_PSIZE(new_bp, size); | |
3040 | BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); | |
428870ff BB |
3041 | BP_SET_CHECKSUM(new_bp, |
3042 | spa_version(spa) >= SPA_VERSION_SLIM_ZIL | |
3043 | ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); | |
b128c09f BB |
3044 | BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); |
3045 | BP_SET_LEVEL(new_bp, 0); | |
428870ff | 3046 | BP_SET_DEDUP(new_bp, 0); |
b128c09f BB |
3047 | BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); |
3048 | } | |
3049 | ||
3050 | return (error); | |
3051 | } | |
3052 | ||
3053 | /* | |
428870ff | 3054 | * Free an intent log block. |
b128c09f BB |
3055 | */ |
3056 | void | |
428870ff | 3057 | zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp) |
b128c09f | 3058 | { |
428870ff | 3059 | ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG); |
b128c09f BB |
3060 | ASSERT(!BP_IS_GANG(bp)); |
3061 | ||
428870ff | 3062 | zio_free(spa, txg, bp); |
b128c09f BB |
3063 | } |
3064 | ||
34dc7c2f BB |
3065 | /* |
3066 | * ========================================================================== | |
3067 | * Read and write to physical devices | |
3068 | * ========================================================================== | |
3069 | */ | |
98b25418 GW |
3070 | |
3071 | ||
3072 | /* | |
3073 | * Issue an I/O to the underlying vdev. Typically the issue pipeline | |
3074 | * stops after this stage and will resume upon I/O completion. | |
3075 | * However, there are instances where the vdev layer may need to | |
3076 | * continue the pipeline when an I/O was not issued. Since the I/O | |
3077 | * that was sent to the vdev layer might be different than the one | |
3078 | * currently active in the pipeline (see vdev_queue_io()), we explicitly | |
3079 | * force the underlying vdev layers to call either zio_execute() or | |
3080 | * zio_interrupt() to ensure that the pipeline continues with the correct I/O. | |
3081 | */ | |
34dc7c2f BB |
3082 | static int |
3083 | zio_vdev_io_start(zio_t *zio) | |
3084 | { | |
3085 | vdev_t *vd = zio->io_vd; | |
34dc7c2f BB |
3086 | uint64_t align; |
3087 | spa_t *spa = zio->io_spa; | |
3088 | ||
193a37cb TH |
3089 | zio->io_delay = 0; |
3090 | ||
b128c09f BB |
3091 | ASSERT(zio->io_error == 0); |
3092 | ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); | |
34dc7c2f | 3093 | |
b128c09f BB |
3094 | if (vd == NULL) { |
3095 | if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
3096 | spa_config_enter(spa, SCL_ZIO, zio, RW_READER); | |
34dc7c2f | 3097 | |
b128c09f BB |
3098 | /* |
3099 | * The mirror_ops handle multiple DVAs in a single BP. | |
3100 | */ | |
98b25418 GW |
3101 | vdev_mirror_ops.vdev_op_io_start(zio); |
3102 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f BB |
3103 | } |
3104 | ||
3dfb57a3 DB |
3105 | ASSERT3P(zio->io_logical, !=, zio); |
3106 | ||
572e2857 BB |
3107 | /* |
3108 | * We keep track of time-sensitive I/Os so that the scan thread | |
3109 | * can quickly react to certain workloads. In particular, we care | |
3110 | * about non-scrubbing, top-level reads and writes with the following | |
3111 | * characteristics: | |
98b25418 | 3112 | * - synchronous writes of user data to non-slog devices |
572e2857 BB |
3113 | * - any reads of user data |
3114 | * When these conditions are met, adjust the timestamp of spa_last_io | |
3115 | * which allows the scan thread to adjust its workload accordingly. | |
3116 | */ | |
3117 | if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL && | |
3118 | vd == vd->vdev_top && !vd->vdev_islog && | |
3119 | zio->io_bookmark.zb_objset != DMU_META_OBJSET && | |
3120 | zio->io_txg != spa_syncing_txg(spa)) { | |
3121 | uint64_t old = spa->spa_last_io; | |
3122 | uint64_t new = ddi_get_lbolt64(); | |
3123 | if (old != new) | |
3124 | (void) atomic_cas_64(&spa->spa_last_io, old, new); | |
3125 | } | |
3126 | ||
b128c09f BB |
3127 | align = 1ULL << vd->vdev_top->vdev_ashift; |
3128 | ||
b02fe35d AR |
3129 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && |
3130 | P2PHASE(zio->io_size, align) != 0) { | |
3131 | /* Transform logical writes to be a full physical block size. */ | |
34dc7c2f BB |
3132 | uint64_t asize = P2ROUNDUP(zio->io_size, align); |
3133 | char *abuf = zio_buf_alloc(asize); | |
178e73b3 | 3134 | ASSERT(vd == vd->vdev_top); |
34dc7c2f BB |
3135 | if (zio->io_type == ZIO_TYPE_WRITE) { |
3136 | bcopy(zio->io_data, abuf, zio->io_size); | |
3137 | bzero(abuf + zio->io_size, asize - zio->io_size); | |
3138 | } | |
b128c09f | 3139 | zio_push_transform(zio, abuf, asize, asize, zio_subblock); |
34dc7c2f BB |
3140 | } |
3141 | ||
b02fe35d AR |
3142 | /* |
3143 | * If this is not a physical io, make sure that it is properly aligned | |
3144 | * before proceeding. | |
3145 | */ | |
3146 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { | |
3147 | ASSERT0(P2PHASE(zio->io_offset, align)); | |
3148 | ASSERT0(P2PHASE(zio->io_size, align)); | |
3149 | } else { | |
3150 | /* | |
3151 | * For physical writes, we allow 512b aligned writes and assume | |
3152 | * the device will perform a read-modify-write as necessary. | |
3153 | */ | |
3154 | ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); | |
3155 | ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); | |
3156 | } | |
3157 | ||
572e2857 | 3158 | VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); |
fb5f0bc8 BB |
3159 | |
3160 | /* | |
3161 | * If this is a repair I/O, and there's no self-healing involved -- | |
3162 | * that is, we're just resilvering what we expect to resilver -- | |
3163 | * then don't do the I/O unless zio's txg is actually in vd's DTL. | |
3164 | * This prevents spurious resilvering with nested replication. | |
3165 | * For example, given a mirror of mirrors, (A+B)+(C+D), if only | |
3166 | * A is out of date, we'll read from C+D, then use the data to | |
3167 | * resilver A+B -- but we don't actually want to resilver B, just A. | |
3168 | * The top-level mirror has no way to know this, so instead we just | |
3169 | * discard unnecessary repairs as we work our way down the vdev tree. | |
3170 | * The same logic applies to any form of nested replication: | |
3171 | * ditto + mirror, RAID-Z + replacing, etc. This covers them all. | |
3172 | */ | |
3173 | if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && | |
3174 | !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && | |
3175 | zio->io_txg != 0 && /* not a delegated i/o */ | |
3176 | !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { | |
3177 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
fb5f0bc8 BB |
3178 | zio_vdev_io_bypass(zio); |
3179 | return (ZIO_PIPELINE_CONTINUE); | |
3180 | } | |
34dc7c2f | 3181 | |
b128c09f BB |
3182 | if (vd->vdev_ops->vdev_op_leaf && |
3183 | (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) { | |
3184 | ||
b0bc7a84 | 3185 | if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio)) |
d164b209 | 3186 | return (ZIO_PIPELINE_CONTINUE); |
b128c09f BB |
3187 | |
3188 | if ((zio = vdev_queue_io(zio)) == NULL) | |
3189 | return (ZIO_PIPELINE_STOP); | |
3190 | ||
3191 | if (!vdev_accessible(vd, zio)) { | |
2e528b49 | 3192 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
3193 | zio_interrupt(zio); |
3194 | return (ZIO_PIPELINE_STOP); | |
3195 | } | |
b128c09f BB |
3196 | } |
3197 | ||
193a37cb | 3198 | zio->io_delay = gethrtime(); |
98b25418 GW |
3199 | vd->vdev_ops->vdev_op_io_start(zio); |
3200 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f BB |
3201 | } |
3202 | ||
3203 | static int | |
3204 | zio_vdev_io_done(zio_t *zio) | |
3205 | { | |
b128c09f BB |
3206 | vdev_t *vd = zio->io_vd; |
3207 | vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; | |
3208 | boolean_t unexpected_error = B_FALSE; | |
34dc7c2f | 3209 | |
b128c09f BB |
3210 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) |
3211 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f | 3212 | |
b128c09f BB |
3213 | ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); |
3214 | ||
193a37cb TH |
3215 | if (zio->io_delay) |
3216 | zio->io_delay = gethrtime() - zio->io_delay; | |
3217 | ||
b128c09f BB |
3218 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { |
3219 | ||
3220 | vdev_queue_io_done(zio); | |
3221 | ||
3222 | if (zio->io_type == ZIO_TYPE_WRITE) | |
3223 | vdev_cache_write(zio); | |
3224 | ||
3225 | if (zio_injection_enabled && zio->io_error == 0) | |
9babb374 BB |
3226 | zio->io_error = zio_handle_device_injection(vd, |
3227 | zio, EIO); | |
b128c09f BB |
3228 | |
3229 | if (zio_injection_enabled && zio->io_error == 0) | |
3230 | zio->io_error = zio_handle_label_injection(zio, EIO); | |
3231 | ||
3232 | if (zio->io_error) { | |
3233 | if (!vdev_accessible(vd, zio)) { | |
2e528b49 | 3234 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
3235 | } else { |
3236 | unexpected_error = B_TRUE; | |
3237 | } | |
3238 | } | |
3239 | } | |
3240 | ||
3241 | ops->vdev_op_io_done(zio); | |
34dc7c2f | 3242 | |
b128c09f | 3243 | if (unexpected_error) |
d164b209 | 3244 | VERIFY(vdev_probe(vd, zio) == NULL); |
34dc7c2f | 3245 | |
b128c09f | 3246 | return (ZIO_PIPELINE_CONTINUE); |
34dc7c2f BB |
3247 | } |
3248 | ||
428870ff BB |
3249 | /* |
3250 | * For non-raidz ZIOs, we can just copy aside the bad data read from the | |
3251 | * disk, and use that to finish the checksum ereport later. | |
3252 | */ | |
3253 | static void | |
3254 | zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, | |
3255 | const void *good_buf) | |
3256 | { | |
3257 | /* no processing needed */ | |
3258 | zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); | |
3259 | } | |
3260 | ||
3261 | /*ARGSUSED*/ | |
3262 | void | |
3263 | zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) | |
3264 | { | |
3265 | void *buf = zio_buf_alloc(zio->io_size); | |
3266 | ||
3267 | bcopy(zio->io_data, buf, zio->io_size); | |
3268 | ||
3269 | zcr->zcr_cbinfo = zio->io_size; | |
3270 | zcr->zcr_cbdata = buf; | |
3271 | zcr->zcr_finish = zio_vsd_default_cksum_finish; | |
3272 | zcr->zcr_free = zio_buf_free; | |
3273 | } | |
3274 | ||
34dc7c2f BB |
3275 | static int |
3276 | zio_vdev_io_assess(zio_t *zio) | |
3277 | { | |
3278 | vdev_t *vd = zio->io_vd; | |
b128c09f BB |
3279 | |
3280 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) | |
3281 | return (ZIO_PIPELINE_STOP); | |
3282 | ||
3283 | if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
3284 | spa_config_exit(zio->io_spa, SCL_ZIO, zio); | |
3285 | ||
3286 | if (zio->io_vsd != NULL) { | |
428870ff | 3287 | zio->io_vsd_ops->vsd_free(zio); |
b128c09f | 3288 | zio->io_vsd = NULL; |
34dc7c2f BB |
3289 | } |
3290 | ||
b128c09f | 3291 | if (zio_injection_enabled && zio->io_error == 0) |
34dc7c2f BB |
3292 | zio->io_error = zio_handle_fault_injection(zio, EIO); |
3293 | ||
3294 | /* | |
3295 | * If the I/O failed, determine whether we should attempt to retry it. | |
428870ff BB |
3296 | * |
3297 | * On retry, we cut in line in the issue queue, since we don't want | |
3298 | * compression/checksumming/etc. work to prevent our (cheap) IO reissue. | |
34dc7c2f | 3299 | */ |
b128c09f BB |
3300 | if (zio->io_error && vd == NULL && |
3301 | !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { | |
3302 | ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ | |
3303 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ | |
34dc7c2f | 3304 | zio->io_error = 0; |
b128c09f BB |
3305 | zio->io_flags |= ZIO_FLAG_IO_RETRY | |
3306 | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; | |
428870ff BB |
3307 | zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; |
3308 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, | |
3309 | zio_requeue_io_start_cut_in_line); | |
b128c09f | 3310 | return (ZIO_PIPELINE_STOP); |
34dc7c2f BB |
3311 | } |
3312 | ||
b128c09f BB |
3313 | /* |
3314 | * If we got an error on a leaf device, convert it to ENXIO | |
3315 | * if the device is not accessible at all. | |
3316 | */ | |
3317 | if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && | |
3318 | !vdev_accessible(vd, zio)) | |
2e528b49 | 3319 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
3320 | |
3321 | /* | |
3322 | * If we can't write to an interior vdev (mirror or RAID-Z), | |
3323 | * set vdev_cant_write so that we stop trying to allocate from it. | |
3324 | */ | |
3325 | if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && | |
13fe0198 | 3326 | vd != NULL && !vd->vdev_ops->vdev_op_leaf) { |
b128c09f | 3327 | vd->vdev_cant_write = B_TRUE; |
13fe0198 | 3328 | } |
b128c09f BB |
3329 | |
3330 | if (zio->io_error) | |
3331 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
3332 | ||
e8b96c60 MA |
3333 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf && |
3334 | zio->io_physdone != NULL) { | |
3335 | ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); | |
3336 | ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); | |
3337 | zio->io_physdone(zio->io_logical); | |
3338 | } | |
3339 | ||
34dc7c2f BB |
3340 | return (ZIO_PIPELINE_CONTINUE); |
3341 | } | |
3342 | ||
3343 | void | |
3344 | zio_vdev_io_reissue(zio_t *zio) | |
3345 | { | |
3346 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
3347 | ASSERT(zio->io_error == 0); | |
3348 | ||
428870ff | 3349 | zio->io_stage >>= 1; |
34dc7c2f BB |
3350 | } |
3351 | ||
3352 | void | |
3353 | zio_vdev_io_redone(zio_t *zio) | |
3354 | { | |
3355 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); | |
3356 | ||
428870ff | 3357 | zio->io_stage >>= 1; |
34dc7c2f BB |
3358 | } |
3359 | ||
3360 | void | |
3361 | zio_vdev_io_bypass(zio_t *zio) | |
3362 | { | |
3363 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
3364 | ASSERT(zio->io_error == 0); | |
3365 | ||
3366 | zio->io_flags |= ZIO_FLAG_IO_BYPASS; | |
428870ff | 3367 | zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; |
34dc7c2f BB |
3368 | } |
3369 | ||
3370 | /* | |
3371 | * ========================================================================== | |
3372 | * Generate and verify checksums | |
3373 | * ========================================================================== | |
3374 | */ | |
3375 | static int | |
3376 | zio_checksum_generate(zio_t *zio) | |
3377 | { | |
34dc7c2f | 3378 | blkptr_t *bp = zio->io_bp; |
b128c09f | 3379 | enum zio_checksum checksum; |
34dc7c2f | 3380 | |
b128c09f BB |
3381 | if (bp == NULL) { |
3382 | /* | |
3383 | * This is zio_write_phys(). | |
3384 | * We're either generating a label checksum, or none at all. | |
3385 | */ | |
3386 | checksum = zio->io_prop.zp_checksum; | |
34dc7c2f | 3387 | |
b128c09f BB |
3388 | if (checksum == ZIO_CHECKSUM_OFF) |
3389 | return (ZIO_PIPELINE_CONTINUE); | |
3390 | ||
3391 | ASSERT(checksum == ZIO_CHECKSUM_LABEL); | |
3392 | } else { | |
3393 | if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { | |
3394 | ASSERT(!IO_IS_ALLOCATING(zio)); | |
3395 | checksum = ZIO_CHECKSUM_GANG_HEADER; | |
3396 | } else { | |
3397 | checksum = BP_GET_CHECKSUM(bp); | |
3398 | } | |
3399 | } | |
34dc7c2f | 3400 | |
b128c09f | 3401 | zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size); |
34dc7c2f BB |
3402 | |
3403 | return (ZIO_PIPELINE_CONTINUE); | |
3404 | } | |
3405 | ||
3406 | static int | |
b128c09f | 3407 | zio_checksum_verify(zio_t *zio) |
34dc7c2f | 3408 | { |
428870ff | 3409 | zio_bad_cksum_t info; |
b128c09f BB |
3410 | blkptr_t *bp = zio->io_bp; |
3411 | int error; | |
34dc7c2f | 3412 | |
428870ff BB |
3413 | ASSERT(zio->io_vd != NULL); |
3414 | ||
b128c09f BB |
3415 | if (bp == NULL) { |
3416 | /* | |
3417 | * This is zio_read_phys(). | |
3418 | * We're either verifying a label checksum, or nothing at all. | |
3419 | */ | |
3420 | if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) | |
3421 | return (ZIO_PIPELINE_CONTINUE); | |
34dc7c2f | 3422 | |
b128c09f BB |
3423 | ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); |
3424 | } | |
34dc7c2f | 3425 | |
428870ff | 3426 | if ((error = zio_checksum_error(zio, &info)) != 0) { |
b128c09f | 3427 | zio->io_error = error; |
7a3066ff MA |
3428 | if (error == ECKSUM && |
3429 | !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { | |
428870ff BB |
3430 | zfs_ereport_start_checksum(zio->io_spa, |
3431 | zio->io_vd, zio, zio->io_offset, | |
3432 | zio->io_size, NULL, &info); | |
b128c09f | 3433 | } |
34dc7c2f BB |
3434 | } |
3435 | ||
3436 | return (ZIO_PIPELINE_CONTINUE); | |
3437 | } | |
3438 | ||
3439 | /* | |
3440 | * Called by RAID-Z to ensure we don't compute the checksum twice. | |
3441 | */ | |
3442 | void | |
3443 | zio_checksum_verified(zio_t *zio) | |
3444 | { | |
428870ff | 3445 | zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; |
34dc7c2f BB |
3446 | } |
3447 | ||
3448 | /* | |
b128c09f BB |
3449 | * ========================================================================== |
3450 | * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. | |
9b67f605 | 3451 | * An error of 0 indicates success. ENXIO indicates whole-device failure, |
b128c09f BB |
3452 | * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO |
3453 | * indicate errors that are specific to one I/O, and most likely permanent. | |
3454 | * Any other error is presumed to be worse because we weren't expecting it. | |
3455 | * ========================================================================== | |
34dc7c2f | 3456 | */ |
b128c09f BB |
3457 | int |
3458 | zio_worst_error(int e1, int e2) | |
34dc7c2f | 3459 | { |
b128c09f BB |
3460 | static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; |
3461 | int r1, r2; | |
3462 | ||
3463 | for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) | |
3464 | if (e1 == zio_error_rank[r1]) | |
3465 | break; | |
34dc7c2f | 3466 | |
b128c09f BB |
3467 | for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) |
3468 | if (e2 == zio_error_rank[r2]) | |
3469 | break; | |
3470 | ||
3471 | return (r1 > r2 ? e1 : e2); | |
34dc7c2f BB |
3472 | } |
3473 | ||
3474 | /* | |
3475 | * ========================================================================== | |
b128c09f | 3476 | * I/O completion |
34dc7c2f BB |
3477 | * ========================================================================== |
3478 | */ | |
b128c09f BB |
3479 | static int |
3480 | zio_ready(zio_t *zio) | |
34dc7c2f | 3481 | { |
b128c09f | 3482 | blkptr_t *bp = zio->io_bp; |
d164b209 | 3483 | zio_t *pio, *pio_next; |
3dfb57a3 | 3484 | zio_link_t *zl = NULL; |
34dc7c2f | 3485 | |
428870ff BB |
3486 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || |
3487 | zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY)) | |
9babb374 | 3488 | return (ZIO_PIPELINE_STOP); |
34dc7c2f | 3489 | |
9babb374 | 3490 | if (zio->io_ready) { |
b128c09f | 3491 | ASSERT(IO_IS_ALLOCATING(zio)); |
03c6040b GW |
3492 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || |
3493 | (zio->io_flags & ZIO_FLAG_NOPWRITE)); | |
b128c09f | 3494 | ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); |
34dc7c2f | 3495 | |
b128c09f BB |
3496 | zio->io_ready(zio); |
3497 | } | |
34dc7c2f | 3498 | |
b128c09f BB |
3499 | if (bp != NULL && bp != &zio->io_bp_copy) |
3500 | zio->io_bp_copy = *bp; | |
34dc7c2f | 3501 | |
3dfb57a3 | 3502 | if (zio->io_error != 0) { |
b128c09f | 3503 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f | 3504 | |
3dfb57a3 DB |
3505 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
3506 | ASSERT(IO_IS_ALLOCATING(zio)); | |
3507 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
3508 | /* | |
3509 | * We were unable to allocate anything, unreserve and | |
3510 | * issue the next I/O to allocate. | |
3511 | */ | |
3512 | metaslab_class_throttle_unreserve( | |
3513 | spa_normal_class(zio->io_spa), | |
3514 | zio->io_prop.zp_copies, zio); | |
3515 | zio_allocate_dispatch(zio->io_spa); | |
3516 | } | |
3517 | } | |
3518 | ||
d164b209 BB |
3519 | mutex_enter(&zio->io_lock); |
3520 | zio->io_state[ZIO_WAIT_READY] = 1; | |
3dfb57a3 | 3521 | pio = zio_walk_parents(zio, &zl); |
d164b209 BB |
3522 | mutex_exit(&zio->io_lock); |
3523 | ||
3524 | /* | |
3525 | * As we notify zio's parents, new parents could be added. | |
3526 | * New parents go to the head of zio's io_parent_list, however, | |
3527 | * so we will (correctly) not notify them. The remainder of zio's | |
3528 | * io_parent_list, from 'pio_next' onward, cannot change because | |
3529 | * all parents must wait for us to be done before they can be done. | |
3530 | */ | |
3531 | for (; pio != NULL; pio = pio_next) { | |
3dfb57a3 | 3532 | pio_next = zio_walk_parents(zio, &zl); |
b128c09f | 3533 | zio_notify_parent(pio, zio, ZIO_WAIT_READY); |
d164b209 | 3534 | } |
34dc7c2f | 3535 | |
428870ff BB |
3536 | if (zio->io_flags & ZIO_FLAG_NODATA) { |
3537 | if (BP_IS_GANG(bp)) { | |
3538 | zio->io_flags &= ~ZIO_FLAG_NODATA; | |
3539 | } else { | |
3540 | ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE); | |
3541 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; | |
3542 | } | |
3543 | } | |
3544 | ||
3545 | if (zio_injection_enabled && | |
3546 | zio->io_spa->spa_syncing_txg == zio->io_txg) | |
3547 | zio_handle_ignored_writes(zio); | |
3548 | ||
b128c09f | 3549 | return (ZIO_PIPELINE_CONTINUE); |
34dc7c2f BB |
3550 | } |
3551 | ||
3dfb57a3 DB |
3552 | /* |
3553 | * Update the allocation throttle accounting. | |
3554 | */ | |
3555 | static void | |
3556 | zio_dva_throttle_done(zio_t *zio) | |
3557 | { | |
3558 | zio_t *lio = zio->io_logical; | |
3559 | zio_t *pio = zio_unique_parent(zio); | |
3560 | vdev_t *vd = zio->io_vd; | |
3561 | int flags = METASLAB_ASYNC_ALLOC; | |
3562 | ||
3563 | ASSERT3P(zio->io_bp, !=, NULL); | |
3564 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); | |
3565 | ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE); | |
3566 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); | |
3567 | ASSERT(vd != NULL); | |
3568 | ASSERT3P(vd, ==, vd->vdev_top); | |
3569 | ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY))); | |
3570 | ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING); | |
3571 | ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
3572 | ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA)); | |
3573 | ||
3574 | /* | |
3575 | * Parents of gang children can have two flavors -- ones that | |
3576 | * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set) | |
3577 | * and ones that allocated the constituent blocks. The allocation | |
3578 | * throttle needs to know the allocating parent zio so we must find | |
3579 | * it here. | |
3580 | */ | |
3581 | if (pio->io_child_type == ZIO_CHILD_GANG) { | |
3582 | /* | |
3583 | * If our parent is a rewrite gang child then our grandparent | |
3584 | * would have been the one that performed the allocation. | |
3585 | */ | |
3586 | if (pio->io_flags & ZIO_FLAG_IO_REWRITE) | |
3587 | pio = zio_unique_parent(pio); | |
3588 | flags |= METASLAB_GANG_CHILD; | |
3589 | } | |
3590 | ||
3591 | ASSERT(IO_IS_ALLOCATING(pio)); | |
3592 | ASSERT3P(zio, !=, zio->io_logical); | |
3593 | ASSERT(zio->io_logical != NULL); | |
3594 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); | |
3595 | ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE); | |
3596 | ||
3597 | mutex_enter(&pio->io_lock); | |
3598 | metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags); | |
3599 | mutex_exit(&pio->io_lock); | |
3600 | ||
3601 | metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa), | |
3602 | 1, pio); | |
3603 | ||
3604 | /* | |
3605 | * Call into the pipeline to see if there is more work that | |
3606 | * needs to be done. If there is work to be done it will be | |
3607 | * dispatched to another taskq thread. | |
3608 | */ | |
3609 | zio_allocate_dispatch(zio->io_spa); | |
3610 | } | |
3611 | ||
b128c09f BB |
3612 | static int |
3613 | zio_done(zio_t *zio) | |
34dc7c2f | 3614 | { |
3dfb57a3 DB |
3615 | /* |
3616 | * Always attempt to keep stack usage minimal here since | |
3617 | * we can be called recurisvely up to 19 levels deep. | |
3618 | */ | |
d164b209 | 3619 | zio_t *pio, *pio_next; |
d6320ddb | 3620 | int c, w; |
3dfb57a3 | 3621 | zio_link_t *zl = NULL; |
34dc7c2f | 3622 | |
b128c09f | 3623 | /* |
9babb374 | 3624 | * If our children haven't all completed, |
b128c09f BB |
3625 | * wait for them and then repeat this pipeline stage. |
3626 | */ | |
3627 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || | |
3628 | zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || | |
428870ff | 3629 | zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) || |
b128c09f BB |
3630 | zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) |
3631 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f | 3632 | |
3dfb57a3 DB |
3633 | /* |
3634 | * If the allocation throttle is enabled, then update the accounting. | |
3635 | * We only track child I/Os that are part of an allocating async | |
3636 | * write. We must do this since the allocation is performed | |
3637 | * by the logical I/O but the actual write is done by child I/Os. | |
3638 | */ | |
3639 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING && | |
3640 | zio->io_child_type == ZIO_CHILD_VDEV) { | |
3641 | ASSERT(spa_normal_class( | |
3642 | zio->io_spa)->mc_alloc_throttle_enabled); | |
3643 | zio_dva_throttle_done(zio); | |
3644 | } | |
3645 | ||
3646 | /* | |
3647 | * If the allocation throttle is enabled, verify that | |
3648 | * we have decremented the refcounts for every I/O that was throttled. | |
3649 | */ | |
3650 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { | |
3651 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
3652 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
3653 | ASSERT(zio->io_bp != NULL); | |
3654 | metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio); | |
3655 | VERIFY(refcount_not_held( | |
3656 | &(spa_normal_class(zio->io_spa)->mc_alloc_slots), zio)); | |
3657 | } | |
3658 | ||
3659 | ||
d6320ddb BB |
3660 | for (c = 0; c < ZIO_CHILD_TYPES; c++) |
3661 | for (w = 0; w < ZIO_WAIT_TYPES; w++) | |
b128c09f BB |
3662 | ASSERT(zio->io_children[c][w] == 0); |
3663 | ||
9b67f605 | 3664 | if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) { |
c776b317 BB |
3665 | ASSERT(zio->io_bp->blk_pad[0] == 0); |
3666 | ASSERT(zio->io_bp->blk_pad[1] == 0); | |
d1d7e268 MK |
3667 | ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, |
3668 | sizeof (blkptr_t)) == 0 || | |
c776b317 BB |
3669 | (zio->io_bp == zio_unique_parent(zio)->io_bp)); |
3670 | if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) && | |
428870ff | 3671 | zio->io_bp_override == NULL && |
b128c09f | 3672 | !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { |
c776b317 | 3673 | ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp)); |
d1d7e268 MK |
3674 | ASSERT3U(zio->io_prop.zp_copies, <=, |
3675 | BP_GET_NDVAS(zio->io_bp)); | |
c776b317 | 3676 | ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 || |
d1d7e268 MK |
3677 | (BP_COUNT_GANG(zio->io_bp) == |
3678 | BP_GET_NDVAS(zio->io_bp))); | |
b128c09f | 3679 | } |
03c6040b GW |
3680 | if (zio->io_flags & ZIO_FLAG_NOPWRITE) |
3681 | VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig)); | |
b128c09f BB |
3682 | } |
3683 | ||
3684 | /* | |
428870ff | 3685 | * If there were child vdev/gang/ddt errors, they apply to us now. |
b128c09f BB |
3686 | */ |
3687 | zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); | |
3688 | zio_inherit_child_errors(zio, ZIO_CHILD_GANG); | |
428870ff BB |
3689 | zio_inherit_child_errors(zio, ZIO_CHILD_DDT); |
3690 | ||
3691 | /* | |
3692 | * If the I/O on the transformed data was successful, generate any | |
3693 | * checksum reports now while we still have the transformed data. | |
3694 | */ | |
3695 | if (zio->io_error == 0) { | |
3696 | while (zio->io_cksum_report != NULL) { | |
3697 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
3698 | uint64_t align = zcr->zcr_align; | |
c776b317 | 3699 | uint64_t asize = P2ROUNDUP(zio->io_size, align); |
428870ff BB |
3700 | char *abuf = zio->io_data; |
3701 | ||
c776b317 | 3702 | if (asize != zio->io_size) { |
428870ff | 3703 | abuf = zio_buf_alloc(asize); |
c776b317 | 3704 | bcopy(zio->io_data, abuf, zio->io_size); |
d1d7e268 | 3705 | bzero(abuf+zio->io_size, asize-zio->io_size); |
428870ff BB |
3706 | } |
3707 | ||
3708 | zio->io_cksum_report = zcr->zcr_next; | |
3709 | zcr->zcr_next = NULL; | |
3710 | zcr->zcr_finish(zcr, abuf); | |
3711 | zfs_ereport_free_checksum(zcr); | |
3712 | ||
c776b317 | 3713 | if (asize != zio->io_size) |
428870ff BB |
3714 | zio_buf_free(abuf, asize); |
3715 | } | |
3716 | } | |
b128c09f BB |
3717 | |
3718 | zio_pop_transforms(zio); /* note: may set zio->io_error */ | |
3719 | ||
c776b317 | 3720 | vdev_stat_update(zio, zio->io_size); |
b128c09f | 3721 | |
a69052be | 3722 | /* |
cc92e9d0 | 3723 | * If this I/O is attached to a particular vdev is slow, exceeding |
72f53c56 MJ |
3724 | * 30 seconds to complete, post an error described the I/O delay. |
3725 | * We ignore these errors if the device is currently unavailable. | |
a69052be | 3726 | */ |
193a37cb | 3727 | if (zio->io_delay >= MSEC2NSEC(zio_delay_max)) { |
72f53c56 MJ |
3728 | if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) |
3729 | zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa, | |
d1d7e268 | 3730 | zio->io_vd, zio, 0, 0); |
72f53c56 | 3731 | } |
a69052be | 3732 | |
b128c09f BB |
3733 | if (zio->io_error) { |
3734 | /* | |
3735 | * If this I/O is attached to a particular vdev, | |
3736 | * generate an error message describing the I/O failure | |
3737 | * at the block level. We ignore these errors if the | |
3738 | * device is currently unavailable. | |
3739 | */ | |
c776b317 BB |
3740 | if (zio->io_error != ECKSUM && zio->io_vd != NULL && |
3741 | !vdev_is_dead(zio->io_vd)) | |
3742 | zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa, | |
3743 | zio->io_vd, zio, 0, 0); | |
34dc7c2f | 3744 | |
428870ff BB |
3745 | if ((zio->io_error == EIO || !(zio->io_flags & |
3746 | (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && | |
c776b317 | 3747 | zio == zio->io_logical) { |
b128c09f BB |
3748 | /* |
3749 | * For logical I/O requests, tell the SPA to log the | |
3750 | * error and generate a logical data ereport. | |
3751 | */ | |
c776b317 | 3752 | spa_log_error(zio->io_spa, zio); |
d1d7e268 MK |
3753 | zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, |
3754 | NULL, zio, 0, 0); | |
b128c09f BB |
3755 | } |
3756 | } | |
34dc7c2f | 3757 | |
c776b317 | 3758 | if (zio->io_error && zio == zio->io_logical) { |
b128c09f BB |
3759 | /* |
3760 | * Determine whether zio should be reexecuted. This will | |
3761 | * propagate all the way to the root via zio_notify_parent(). | |
3762 | */ | |
c776b317 | 3763 | ASSERT(zio->io_vd == NULL && zio->io_bp != NULL); |
428870ff | 3764 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); |
b128c09f | 3765 | |
428870ff BB |
3766 | if (IO_IS_ALLOCATING(zio) && |
3767 | !(zio->io_flags & ZIO_FLAG_CANFAIL)) { | |
b128c09f BB |
3768 | if (zio->io_error != ENOSPC) |
3769 | zio->io_reexecute |= ZIO_REEXECUTE_NOW; | |
3770 | else | |
3771 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff | 3772 | } |
b128c09f BB |
3773 | |
3774 | if ((zio->io_type == ZIO_TYPE_READ || | |
3775 | zio->io_type == ZIO_TYPE_FREE) && | |
572e2857 | 3776 | !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && |
b128c09f | 3777 | zio->io_error == ENXIO && |
c776b317 BB |
3778 | spa_load_state(zio->io_spa) == SPA_LOAD_NONE && |
3779 | spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE) | |
b128c09f BB |
3780 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; |
3781 | ||
3782 | if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) | |
3783 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff BB |
3784 | |
3785 | /* | |
3786 | * Here is a possibly good place to attempt to do | |
3787 | * either combinatorial reconstruction or error correction | |
3788 | * based on checksums. It also might be a good place | |
3789 | * to send out preliminary ereports before we suspend | |
3790 | * processing. | |
3791 | */ | |
34dc7c2f BB |
3792 | } |
3793 | ||
3794 | /* | |
b128c09f BB |
3795 | * If there were logical child errors, they apply to us now. |
3796 | * We defer this until now to avoid conflating logical child | |
3797 | * errors with errors that happened to the zio itself when | |
3798 | * updating vdev stats and reporting FMA events above. | |
34dc7c2f | 3799 | */ |
b128c09f | 3800 | zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); |
34dc7c2f | 3801 | |
428870ff BB |
3802 | if ((zio->io_error || zio->io_reexecute) && |
3803 | IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && | |
03c6040b | 3804 | !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) |
c776b317 | 3805 | zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp); |
9babb374 BB |
3806 | |
3807 | zio_gang_tree_free(&zio->io_gang_tree); | |
3808 | ||
3809 | /* | |
3810 | * Godfather I/Os should never suspend. | |
3811 | */ | |
3812 | if ((zio->io_flags & ZIO_FLAG_GODFATHER) && | |
3813 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) | |
3814 | zio->io_reexecute = 0; | |
3815 | ||
b128c09f BB |
3816 | if (zio->io_reexecute) { |
3817 | /* | |
3818 | * This is a logical I/O that wants to reexecute. | |
3819 | * | |
3820 | * Reexecute is top-down. When an i/o fails, if it's not | |
3821 | * the root, it simply notifies its parent and sticks around. | |
3822 | * The parent, seeing that it still has children in zio_done(), | |
3823 | * does the same. This percolates all the way up to the root. | |
3824 | * The root i/o will reexecute or suspend the entire tree. | |
3825 | * | |
3826 | * This approach ensures that zio_reexecute() honors | |
3827 | * all the original i/o dependency relationships, e.g. | |
3828 | * parents not executing until children are ready. | |
3829 | */ | |
3830 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
34dc7c2f | 3831 | |
9babb374 | 3832 | zio->io_gang_leader = NULL; |
b128c09f | 3833 | |
d164b209 BB |
3834 | mutex_enter(&zio->io_lock); |
3835 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
3836 | mutex_exit(&zio->io_lock); | |
3837 | ||
9babb374 BB |
3838 | /* |
3839 | * "The Godfather" I/O monitors its children but is | |
3840 | * not a true parent to them. It will track them through | |
3841 | * the pipeline but severs its ties whenever they get into | |
3842 | * trouble (e.g. suspended). This allows "The Godfather" | |
3843 | * I/O to return status without blocking. | |
3844 | */ | |
3dfb57a3 DB |
3845 | zl = NULL; |
3846 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; | |
3847 | pio = pio_next) { | |
3848 | zio_link_t *remove_zl = zl; | |
3849 | pio_next = zio_walk_parents(zio, &zl); | |
9babb374 BB |
3850 | |
3851 | if ((pio->io_flags & ZIO_FLAG_GODFATHER) && | |
3852 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { | |
3dfb57a3 | 3853 | zio_remove_child(pio, zio, remove_zl); |
9babb374 BB |
3854 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE); |
3855 | } | |
3856 | } | |
3857 | ||
d164b209 | 3858 | if ((pio = zio_unique_parent(zio)) != NULL) { |
b128c09f BB |
3859 | /* |
3860 | * We're not a root i/o, so there's nothing to do | |
3861 | * but notify our parent. Don't propagate errors | |
3862 | * upward since we haven't permanently failed yet. | |
3863 | */ | |
9babb374 | 3864 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f BB |
3865 | zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; |
3866 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE); | |
3867 | } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { | |
3868 | /* | |
3869 | * We'd fail again if we reexecuted now, so suspend | |
3870 | * until conditions improve (e.g. device comes online). | |
3871 | */ | |
c776b317 | 3872 | zio_suspend(zio->io_spa, zio); |
b128c09f BB |
3873 | } else { |
3874 | /* | |
3875 | * Reexecution is potentially a huge amount of work. | |
3876 | * Hand it off to the otherwise-unused claim taskq. | |
3877 | */ | |
a38718a6 | 3878 | ASSERT(taskq_empty_ent(&zio->io_tqent)); |
7ef5e54e AL |
3879 | spa_taskq_dispatch_ent(zio->io_spa, |
3880 | ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE, | |
a38718a6 GA |
3881 | (task_func_t *)zio_reexecute, zio, 0, |
3882 | &zio->io_tqent); | |
b128c09f BB |
3883 | } |
3884 | return (ZIO_PIPELINE_STOP); | |
34dc7c2f BB |
3885 | } |
3886 | ||
428870ff | 3887 | ASSERT(zio->io_child_count == 0); |
b128c09f BB |
3888 | ASSERT(zio->io_reexecute == 0); |
3889 | ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); | |
34dc7c2f | 3890 | |
428870ff BB |
3891 | /* |
3892 | * Report any checksum errors, since the I/O is complete. | |
3893 | */ | |
3894 | while (zio->io_cksum_report != NULL) { | |
3895 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
3896 | zio->io_cksum_report = zcr->zcr_next; | |
3897 | zcr->zcr_next = NULL; | |
3898 | zcr->zcr_finish(zcr, NULL); | |
3899 | zfs_ereport_free_checksum(zcr); | |
3900 | } | |
3901 | ||
920dd524 | 3902 | if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp && |
9b67f605 MA |
3903 | !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) && |
3904 | !(zio->io_flags & ZIO_FLAG_NOPWRITE)) { | |
920dd524 ED |
3905 | metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp); |
3906 | } | |
3907 | ||
d164b209 BB |
3908 | /* |
3909 | * It is the responsibility of the done callback to ensure that this | |
3910 | * particular zio is no longer discoverable for adoption, and as | |
3911 | * such, cannot acquire any new parents. | |
3912 | */ | |
b128c09f BB |
3913 | if (zio->io_done) |
3914 | zio->io_done(zio); | |
34dc7c2f | 3915 | |
d164b209 BB |
3916 | mutex_enter(&zio->io_lock); |
3917 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
3918 | mutex_exit(&zio->io_lock); | |
34dc7c2f | 3919 | |
3dfb57a3 DB |
3920 | zl = NULL; |
3921 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) { | |
3922 | zio_link_t *remove_zl = zl; | |
3923 | pio_next = zio_walk_parents(zio, &zl); | |
3924 | zio_remove_child(pio, zio, remove_zl); | |
b128c09f BB |
3925 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE); |
3926 | } | |
34dc7c2f | 3927 | |
b128c09f BB |
3928 | if (zio->io_waiter != NULL) { |
3929 | mutex_enter(&zio->io_lock); | |
3930 | zio->io_executor = NULL; | |
3931 | cv_broadcast(&zio->io_cv); | |
3932 | mutex_exit(&zio->io_lock); | |
3933 | } else { | |
3934 | zio_destroy(zio); | |
3935 | } | |
34dc7c2f | 3936 | |
b128c09f | 3937 | return (ZIO_PIPELINE_STOP); |
34dc7c2f BB |
3938 | } |
3939 | ||
3940 | /* | |
b128c09f BB |
3941 | * ========================================================================== |
3942 | * I/O pipeline definition | |
3943 | * ========================================================================== | |
34dc7c2f | 3944 | */ |
428870ff | 3945 | static zio_pipe_stage_t *zio_pipeline[] = { |
b128c09f | 3946 | NULL, |
b128c09f | 3947 | zio_read_bp_init, |
3dfb57a3 | 3948 | zio_write_bp_init, |
428870ff BB |
3949 | zio_free_bp_init, |
3950 | zio_issue_async, | |
3dfb57a3 | 3951 | zio_write_compress, |
b128c09f | 3952 | zio_checksum_generate, |
03c6040b | 3953 | zio_nop_write, |
428870ff BB |
3954 | zio_ddt_read_start, |
3955 | zio_ddt_read_done, | |
3956 | zio_ddt_write, | |
3957 | zio_ddt_free, | |
b128c09f BB |
3958 | zio_gang_assemble, |
3959 | zio_gang_issue, | |
3dfb57a3 | 3960 | zio_dva_throttle, |
b128c09f BB |
3961 | zio_dva_allocate, |
3962 | zio_dva_free, | |
3963 | zio_dva_claim, | |
3964 | zio_ready, | |
3965 | zio_vdev_io_start, | |
3966 | zio_vdev_io_done, | |
3967 | zio_vdev_io_assess, | |
3968 | zio_checksum_verify, | |
3969 | zio_done | |
3970 | }; | |
c28b2279 | 3971 | |
9ae529ec | 3972 | |
9ae529ec | 3973 | |
9ae529ec | 3974 | |
fcff0f35 PD |
3975 | /* |
3976 | * Compare two zbookmark_phys_t's to see which we would reach first in a | |
3977 | * pre-order traversal of the object tree. | |
3978 | * | |
3979 | * This is simple in every case aside from the meta-dnode object. For all other | |
3980 | * objects, we traverse them in order (object 1 before object 2, and so on). | |
3981 | * However, all of these objects are traversed while traversing object 0, since | |
3982 | * the data it points to is the list of objects. Thus, we need to convert to a | |
3983 | * canonical representation so we can compare meta-dnode bookmarks to | |
3984 | * non-meta-dnode bookmarks. | |
3985 | * | |
3986 | * We do this by calculating "equivalents" for each field of the zbookmark. | |
3987 | * zbookmarks outside of the meta-dnode use their own object and level, and | |
3988 | * calculate the level 0 equivalent (the first L0 blkid that is contained in the | |
3989 | * blocks this bookmark refers to) by multiplying their blkid by their span | |
3990 | * (the number of L0 blocks contained within one block at their level). | |
3991 | * zbookmarks inside the meta-dnode calculate their object equivalent | |
3992 | * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use | |
3993 | * level + 1<<31 (any value larger than a level could ever be) for their level. | |
3994 | * This causes them to always compare before a bookmark in their object | |
3995 | * equivalent, compare appropriately to bookmarks in other objects, and to | |
3996 | * compare appropriately to other bookmarks in the meta-dnode. | |
3997 | */ | |
3998 | int | |
3999 | zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, | |
4000 | const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) | |
4001 | { | |
4002 | /* | |
4003 | * These variables represent the "equivalent" values for the zbookmark, | |
4004 | * after converting zbookmarks inside the meta dnode to their | |
4005 | * normal-object equivalents. | |
4006 | */ | |
4007 | uint64_t zb1obj, zb2obj; | |
4008 | uint64_t zb1L0, zb2L0; | |
4009 | uint64_t zb1level, zb2level; | |
4010 | ||
4011 | if (zb1->zb_object == zb2->zb_object && | |
4012 | zb1->zb_level == zb2->zb_level && | |
4013 | zb1->zb_blkid == zb2->zb_blkid) | |
4014 | return (0); | |
9ae529ec | 4015 | |
fcff0f35 PD |
4016 | /* |
4017 | * BP_SPANB calculates the span in blocks. | |
4018 | */ | |
4019 | zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); | |
4020 | zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); | |
9ae529ec CS |
4021 | |
4022 | if (zb1->zb_object == DMU_META_DNODE_OBJECT) { | |
fcff0f35 PD |
4023 | zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); |
4024 | zb1L0 = 0; | |
4025 | zb1level = zb1->zb_level + COMPARE_META_LEVEL; | |
4026 | } else { | |
4027 | zb1obj = zb1->zb_object; | |
4028 | zb1level = zb1->zb_level; | |
9ae529ec CS |
4029 | } |
4030 | ||
fcff0f35 PD |
4031 | if (zb2->zb_object == DMU_META_DNODE_OBJECT) { |
4032 | zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); | |
4033 | zb2L0 = 0; | |
4034 | zb2level = zb2->zb_level + COMPARE_META_LEVEL; | |
4035 | } else { | |
4036 | zb2obj = zb2->zb_object; | |
4037 | zb2level = zb2->zb_level; | |
4038 | } | |
4039 | ||
4040 | /* Now that we have a canonical representation, do the comparison. */ | |
4041 | if (zb1obj != zb2obj) | |
4042 | return (zb1obj < zb2obj ? -1 : 1); | |
4043 | else if (zb1L0 != zb2L0) | |
4044 | return (zb1L0 < zb2L0 ? -1 : 1); | |
4045 | else if (zb1level != zb2level) | |
4046 | return (zb1level > zb2level ? -1 : 1); | |
4047 | /* | |
4048 | * This can (theoretically) happen if the bookmarks have the same object | |
4049 | * and level, but different blkids, if the block sizes are not the same. | |
4050 | * There is presently no way to change the indirect block sizes | |
4051 | */ | |
4052 | return (0); | |
4053 | } | |
4054 | ||
4055 | /* | |
4056 | * This function checks the following: given that last_block is the place that | |
4057 | * our traversal stopped last time, does that guarantee that we've visited | |
4058 | * every node under subtree_root? Therefore, we can't just use the raw output | |
4059 | * of zbookmark_compare. We have to pass in a modified version of | |
4060 | * subtree_root; by incrementing the block id, and then checking whether | |
4061 | * last_block is before or equal to that, we can tell whether or not having | |
4062 | * visited last_block implies that all of subtree_root's children have been | |
4063 | * visited. | |
4064 | */ | |
4065 | boolean_t | |
4066 | zbookmark_subtree_completed(const dnode_phys_t *dnp, | |
4067 | const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) | |
4068 | { | |
4069 | zbookmark_phys_t mod_zb = *subtree_root; | |
4070 | mod_zb.zb_blkid++; | |
4071 | ASSERT(last_block->zb_level == 0); | |
4072 | ||
4073 | /* The objset_phys_t isn't before anything. */ | |
4074 | if (dnp == NULL) | |
9ae529ec | 4075 | return (B_FALSE); |
fcff0f35 PD |
4076 | |
4077 | /* | |
4078 | * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the | |
4079 | * data block size in sectors, because that variable is only used if | |
4080 | * the bookmark refers to a block in the meta-dnode. Since we don't | |
4081 | * know without examining it what object it refers to, and there's no | |
4082 | * harm in passing in this value in other cases, we always pass it in. | |
4083 | * | |
4084 | * We pass in 0 for the indirect block size shift because zb2 must be | |
4085 | * level 0. The indirect block size is only used to calculate the span | |
4086 | * of the bookmark, but since the bookmark must be level 0, the span is | |
4087 | * always 1, so the math works out. | |
4088 | * | |
4089 | * If you make changes to how the zbookmark_compare code works, be sure | |
4090 | * to make sure that this code still works afterwards. | |
4091 | */ | |
4092 | return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, | |
4093 | 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, | |
4094 | last_block) <= 0); | |
9ae529ec CS |
4095 | } |
4096 | ||
c28b2279 | 4097 | #if defined(_KERNEL) && defined(HAVE_SPL) |
c28b2279 | 4098 | EXPORT_SYMBOL(zio_type_name); |
81971b13 BB |
4099 | EXPORT_SYMBOL(zio_buf_alloc); |
4100 | EXPORT_SYMBOL(zio_data_buf_alloc); | |
6fe53787 | 4101 | EXPORT_SYMBOL(zio_buf_alloc_flags); |
81971b13 BB |
4102 | EXPORT_SYMBOL(zio_buf_free); |
4103 | EXPORT_SYMBOL(zio_data_buf_free); | |
c28b2279 | 4104 | |
a69052be | 4105 | module_param(zio_delay_max, int, 0644); |
c409e464 BB |
4106 | MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event"); |
4107 | ||
4108 | module_param(zio_requeue_io_start_cut_in_line, int, 0644); | |
4109 | MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O"); | |
29dee3ee CP |
4110 | |
4111 | module_param(zfs_sync_pass_deferred_free, int, 0644); | |
4112 | MODULE_PARM_DESC(zfs_sync_pass_deferred_free, | |
d1d7e268 | 4113 | "Defer frees starting in this pass"); |
29dee3ee CP |
4114 | |
4115 | module_param(zfs_sync_pass_dont_compress, int, 0644); | |
4116 | MODULE_PARM_DESC(zfs_sync_pass_dont_compress, | |
d1d7e268 | 4117 | "Don't compress starting in this pass"); |
29dee3ee CP |
4118 | |
4119 | module_param(zfs_sync_pass_rewrite, int, 0644); | |
4120 | MODULE_PARM_DESC(zfs_sync_pass_rewrite, | |
d1d7e268 | 4121 | "Rewrite new bps starting in this pass"); |
3dfb57a3 DB |
4122 | |
4123 | module_param(zio_dva_throttle_enabled, int, 0644); | |
4124 | MODULE_PARM_DESC(zio_dva_throttle_enabled, | |
4125 | "Throttle block allocations in the ZIO pipeline"); | |
c28b2279 | 4126 | #endif |