Print this page
13065 log when suspending a zpool
Reviewed by: Andy Fiddaman <andy@omniosce.org>
Reviewed by: Toomas Soome <tsoome@me.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/zio.c
+++ new/usr/src/uts/common/fs/zfs/zio.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
|
↓ open down ↓ |
16 lines elided |
↑ open up ↑ |
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 23 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24 24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 25 * Copyright (c) 2014 Integros [integros.com]
26 26 * Copyright (c) 2017, Intel Corporation.
27 + * Copyright 2020 Joyent, Inc.
27 28 */
28 29
29 30 #include <sys/sysmacros.h>
30 31 #include <sys/zfs_context.h>
31 32 #include <sys/fm/fs/zfs.h>
32 33 #include <sys/spa.h>
33 34 #include <sys/txg.h>
34 35 #include <sys/spa_impl.h>
35 36 #include <sys/vdev_impl.h>
36 37 #include <sys/vdev_trim.h>
37 38 #include <sys/zio_impl.h>
38 39 #include <sys/zio_compress.h>
39 40 #include <sys/zio_checksum.h>
40 41 #include <sys/dmu_objset.h>
41 42 #include <sys/arc.h>
42 43 #include <sys/ddt.h>
43 44 #include <sys/blkptr.h>
44 45 #include <sys/zfeature.h>
45 46 #include <sys/time.h>
46 47 #include <sys/dsl_scan.h>
47 48 #include <sys/metaslab_impl.h>
48 49 #include <sys/abd.h>
49 50 #include <sys/cityhash.h>
50 51 #include <sys/dsl_crypt.h>
51 52
52 53 /*
53 54 * ==========================================================================
54 55 * I/O type descriptions
55 56 * ==========================================================================
56 57 */
57 58 const char *zio_type_name[ZIO_TYPES] = {
58 59 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
59 60 "zio_ioctl", "z_trim"
60 61 };
61 62
62 63 boolean_t zio_dva_throttle_enabled = B_TRUE;
63 64
64 65 /*
65 66 * ==========================================================================
66 67 * I/O kmem caches
67 68 * ==========================================================================
68 69 */
69 70 kmem_cache_t *zio_cache;
70 71 kmem_cache_t *zio_link_cache;
71 72 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
72 73 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
73 74
74 75 #ifdef _KERNEL
75 76 extern vmem_t *zio_alloc_arena;
76 77 #endif
77 78
78 79 #define ZIO_PIPELINE_CONTINUE 0x100
79 80 #define ZIO_PIPELINE_STOP 0x101
80 81
81 82 /* Mark IOs as "slow" if they take longer than 30 seconds */
82 83 int zio_slow_io_ms = (30 * MILLISEC);
83 84
84 85 #define BP_SPANB(indblkshift, level) \
85 86 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
86 87 #define COMPARE_META_LEVEL 0x80000000ul
87 88 /*
88 89 * The following actions directly effect the spa's sync-to-convergence logic.
89 90 * The values below define the sync pass when we start performing the action.
90 91 * Care should be taken when changing these values as they directly impact
91 92 * spa_sync() performance. Tuning these values may introduce subtle performance
92 93 * pathologies and should only be done in the context of performance analysis.
93 94 * These tunables will eventually be removed and replaced with #defines once
94 95 * enough analysis has been done to determine optimal values.
95 96 *
96 97 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
97 98 * regular blocks are not deferred.
98 99 */
99 100 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
100 101 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
101 102 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
102 103
103 104 /*
104 105 * An allocating zio is one that either currently has the DVA allocate
105 106 * stage set or will have it later in its lifetime.
106 107 */
107 108 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
108 109
109 110 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
110 111
111 112 #ifdef ZFS_DEBUG
112 113 int zio_buf_debug_limit = 16384;
113 114 #else
114 115 int zio_buf_debug_limit = 0;
115 116 #endif
116 117
117 118 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
118 119
119 120 void
120 121 zio_init(void)
121 122 {
122 123 size_t c;
123 124 vmem_t *data_alloc_arena = NULL;
124 125
125 126 #ifdef _KERNEL
126 127 data_alloc_arena = zio_alloc_arena;
127 128 #endif
128 129 zio_cache = kmem_cache_create("zio_cache",
129 130 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
130 131 zio_link_cache = kmem_cache_create("zio_link_cache",
131 132 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
132 133
133 134 /*
134 135 * For small buffers, we want a cache for each multiple of
135 136 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
136 137 * for each quarter-power of 2.
137 138 */
138 139 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
139 140 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
140 141 size_t p2 = size;
141 142 size_t align = 0;
142 143 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
143 144
144 145 while (!ISP2(p2))
145 146 p2 &= p2 - 1;
146 147
147 148 #ifndef _KERNEL
148 149 /*
149 150 * If we are using watchpoints, put each buffer on its own page,
150 151 * to eliminate the performance overhead of trapping to the
151 152 * kernel when modifying a non-watched buffer that shares the
152 153 * page with a watched buffer.
153 154 */
154 155 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
155 156 continue;
156 157 #endif
157 158 if (size <= 4 * SPA_MINBLOCKSIZE) {
158 159 align = SPA_MINBLOCKSIZE;
159 160 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
160 161 align = MIN(p2 >> 2, PAGESIZE);
161 162 }
162 163
163 164 if (align != 0) {
164 165 char name[36];
165 166 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
166 167 zio_buf_cache[c] = kmem_cache_create(name, size,
167 168 align, NULL, NULL, NULL, NULL, NULL, cflags);
168 169
169 170 /*
170 171 * Since zio_data bufs do not appear in crash dumps, we
171 172 * pass KMC_NOTOUCH so that no allocator metadata is
172 173 * stored with the buffers.
173 174 */
174 175 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
175 176 zio_data_buf_cache[c] = kmem_cache_create(name, size,
176 177 align, NULL, NULL, NULL, NULL, data_alloc_arena,
177 178 cflags | KMC_NOTOUCH);
178 179 }
179 180 }
180 181
181 182 while (--c != 0) {
182 183 ASSERT(zio_buf_cache[c] != NULL);
183 184 if (zio_buf_cache[c - 1] == NULL)
184 185 zio_buf_cache[c - 1] = zio_buf_cache[c];
185 186
186 187 ASSERT(zio_data_buf_cache[c] != NULL);
187 188 if (zio_data_buf_cache[c - 1] == NULL)
188 189 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
189 190 }
190 191
191 192 zio_inject_init();
192 193 }
193 194
194 195 void
195 196 zio_fini(void)
196 197 {
197 198 size_t c;
198 199 kmem_cache_t *last_cache = NULL;
199 200 kmem_cache_t *last_data_cache = NULL;
200 201
201 202 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
202 203 if (zio_buf_cache[c] != last_cache) {
203 204 last_cache = zio_buf_cache[c];
204 205 kmem_cache_destroy(zio_buf_cache[c]);
205 206 }
206 207 zio_buf_cache[c] = NULL;
207 208
208 209 if (zio_data_buf_cache[c] != last_data_cache) {
209 210 last_data_cache = zio_data_buf_cache[c];
210 211 kmem_cache_destroy(zio_data_buf_cache[c]);
211 212 }
212 213 zio_data_buf_cache[c] = NULL;
213 214 }
214 215
215 216 kmem_cache_destroy(zio_link_cache);
216 217 kmem_cache_destroy(zio_cache);
217 218
218 219 zio_inject_fini();
219 220 }
220 221
221 222 /*
222 223 * ==========================================================================
223 224 * Allocate and free I/O buffers
224 225 * ==========================================================================
225 226 */
226 227
227 228 /*
228 229 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
229 230 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
230 231 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
231 232 * excess / transient data in-core during a crashdump.
232 233 */
233 234 void *
234 235 zio_buf_alloc(size_t size)
235 236 {
236 237 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
237 238
238 239 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
239 240
240 241 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
241 242 }
242 243
243 244 /*
244 245 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
245 246 * crashdump if the kernel panics. This exists so that we will limit the amount
246 247 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
247 248 * of kernel heap dumped to disk when the kernel panics)
248 249 */
249 250 void *
250 251 zio_data_buf_alloc(size_t size)
251 252 {
252 253 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
253 254
254 255 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
255 256
256 257 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
257 258 }
258 259
259 260 void
260 261 zio_buf_free(void *buf, size_t size)
261 262 {
262 263 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
263 264
264 265 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
265 266
266 267 kmem_cache_free(zio_buf_cache[c], buf);
267 268 }
268 269
269 270 void
270 271 zio_data_buf_free(void *buf, size_t size)
271 272 {
272 273 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
273 274
274 275 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
275 276
276 277 kmem_cache_free(zio_data_buf_cache[c], buf);
277 278 }
278 279
279 280 /* ARGSUSED */
280 281 static void
281 282 zio_abd_free(void *abd, size_t size)
282 283 {
283 284 abd_free((abd_t *)abd);
284 285 }
285 286
286 287 /*
287 288 * ==========================================================================
288 289 * Push and pop I/O transform buffers
289 290 * ==========================================================================
290 291 */
291 292 void
292 293 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
293 294 zio_transform_func_t *transform)
294 295 {
295 296 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
296 297
297 298 /*
298 299 * Ensure that anyone expecting this zio to contain a linear ABD isn't
299 300 * going to get a nasty surprise when they try to access the data.
300 301 */
301 302 IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
302 303
303 304 zt->zt_orig_abd = zio->io_abd;
304 305 zt->zt_orig_size = zio->io_size;
305 306 zt->zt_bufsize = bufsize;
306 307 zt->zt_transform = transform;
307 308
308 309 zt->zt_next = zio->io_transform_stack;
309 310 zio->io_transform_stack = zt;
310 311
311 312 zio->io_abd = data;
312 313 zio->io_size = size;
313 314 }
314 315
315 316 void
316 317 zio_pop_transforms(zio_t *zio)
317 318 {
318 319 zio_transform_t *zt;
319 320
320 321 while ((zt = zio->io_transform_stack) != NULL) {
321 322 if (zt->zt_transform != NULL)
322 323 zt->zt_transform(zio,
323 324 zt->zt_orig_abd, zt->zt_orig_size);
324 325
325 326 if (zt->zt_bufsize != 0)
326 327 abd_free(zio->io_abd);
327 328
328 329 zio->io_abd = zt->zt_orig_abd;
329 330 zio->io_size = zt->zt_orig_size;
330 331 zio->io_transform_stack = zt->zt_next;
331 332
332 333 kmem_free(zt, sizeof (zio_transform_t));
333 334 }
334 335 }
335 336
336 337 /*
337 338 * ==========================================================================
338 339 * I/O transform callbacks for subblocks, decompression, and decryption
339 340 * ==========================================================================
340 341 */
341 342 static void
342 343 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
343 344 {
344 345 ASSERT(zio->io_size > size);
345 346
346 347 if (zio->io_type == ZIO_TYPE_READ)
347 348 abd_copy(data, zio->io_abd, size);
348 349 }
349 350
350 351 static void
351 352 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
352 353 {
353 354 if (zio->io_error == 0) {
354 355 void *tmp = abd_borrow_buf(data, size);
355 356 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
356 357 zio->io_abd, tmp, zio->io_size, size);
357 358 abd_return_buf_copy(data, tmp, size);
358 359
359 360 if (ret != 0)
360 361 zio->io_error = SET_ERROR(EIO);
361 362 }
362 363 }
363 364
364 365 static void
365 366 zio_decrypt(zio_t *zio, abd_t *data, uint64_t size)
366 367 {
367 368 int ret;
368 369 void *tmp;
369 370 blkptr_t *bp = zio->io_bp;
370 371 spa_t *spa = zio->io_spa;
371 372 uint64_t dsobj = zio->io_bookmark.zb_objset;
372 373 uint64_t lsize = BP_GET_LSIZE(bp);
373 374 dmu_object_type_t ot = BP_GET_TYPE(bp);
374 375 uint8_t salt[ZIO_DATA_SALT_LEN];
375 376 uint8_t iv[ZIO_DATA_IV_LEN];
376 377 uint8_t mac[ZIO_DATA_MAC_LEN];
377 378 boolean_t no_crypt = B_FALSE;
378 379
379 380 ASSERT(BP_USES_CRYPT(bp));
380 381 ASSERT3U(size, !=, 0);
381 382
382 383 if (zio->io_error != 0)
383 384 return;
384 385
385 386 /*
386 387 * Verify the cksum of MACs stored in an indirect bp. It will always
387 388 * be possible to verify this since it does not require an encryption
388 389 * key.
389 390 */
390 391 if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {
391 392 zio_crypt_decode_mac_bp(bp, mac);
392 393
393 394 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
394 395 /*
395 396 * We haven't decompressed the data yet, but
396 397 * zio_crypt_do_indirect_mac_checksum() requires
397 398 * decompressed data to be able to parse out the MACs
398 399 * from the indirect block. We decompress it now and
399 400 * throw away the result after we are finished.
400 401 */
401 402 tmp = zio_buf_alloc(lsize);
402 403 ret = zio_decompress_data(BP_GET_COMPRESS(bp),
403 404 zio->io_abd, tmp, zio->io_size, lsize);
404 405 if (ret != 0) {
405 406 ret = SET_ERROR(EIO);
406 407 goto error;
407 408 }
408 409 ret = zio_crypt_do_indirect_mac_checksum(B_FALSE,
409 410 tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac);
410 411 zio_buf_free(tmp, lsize);
411 412 } else {
412 413 ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
413 414 zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac);
414 415 }
415 416 abd_copy(data, zio->io_abd, size);
416 417
417 418 if (ret != 0)
418 419 goto error;
419 420
420 421 return;
421 422 }
422 423
423 424 /*
424 425 * If this is an authenticated block, just check the MAC. It would be
425 426 * nice to separate this out into its own flag, but for the moment
426 427 * enum zio_flag is out of bits.
427 428 */
428 429 if (BP_IS_AUTHENTICATED(bp)) {
429 430 if (ot == DMU_OT_OBJSET) {
430 431 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa,
431 432 dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp));
432 433 } else {
433 434 zio_crypt_decode_mac_bp(bp, mac);
434 435 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj,
435 436 zio->io_abd, size, mac);
436 437 }
437 438 abd_copy(data, zio->io_abd, size);
438 439
439 440 if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) {
440 441 ret = zio_handle_decrypt_injection(spa,
441 442 &zio->io_bookmark, ot, ECKSUM);
442 443 }
443 444 if (ret != 0)
444 445 goto error;
445 446
446 447 return;
447 448 }
448 449
449 450 zio_crypt_decode_params_bp(bp, salt, iv);
450 451
451 452 if (ot == DMU_OT_INTENT_LOG) {
452 453 tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t));
453 454 zio_crypt_decode_mac_zil(tmp, mac);
454 455 abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t));
455 456 } else {
456 457 zio_crypt_decode_mac_bp(bp, mac);
457 458 }
458 459
459 460 ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp),
460 461 BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data,
461 462 zio->io_abd, &no_crypt);
462 463 if (no_crypt)
463 464 abd_copy(data, zio->io_abd, size);
464 465
465 466 if (ret != 0)
466 467 goto error;
467 468
468 469 return;
469 470
470 471 error:
471 472 /* assert that the key was found unless this was speculative */
472 473 ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE));
473 474
474 475 /*
475 476 * If there was a decryption / authentication error return EIO as
476 477 * the io_error. If this was not a speculative zio, create an ereport.
477 478 */
478 479 if (ret == ECKSUM) {
479 480 zio->io_error = SET_ERROR(EIO);
480 481 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
481 482 spa_log_error(spa, &zio->io_bookmark);
482 483 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
483 484 spa, NULL, &zio->io_bookmark, zio, 0, 0);
484 485 }
485 486 } else {
486 487 zio->io_error = ret;
487 488 }
488 489 }
489 490
490 491 /*
491 492 * ==========================================================================
492 493 * I/O parent/child relationships and pipeline interlocks
493 494 * ==========================================================================
494 495 */
495 496 zio_t *
496 497 zio_walk_parents(zio_t *cio, zio_link_t **zl)
497 498 {
498 499 list_t *pl = &cio->io_parent_list;
499 500
500 501 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
501 502 if (*zl == NULL)
502 503 return (NULL);
503 504
504 505 ASSERT((*zl)->zl_child == cio);
505 506 return ((*zl)->zl_parent);
506 507 }
507 508
508 509 zio_t *
509 510 zio_walk_children(zio_t *pio, zio_link_t **zl)
510 511 {
511 512 list_t *cl = &pio->io_child_list;
512 513
513 514 ASSERT(MUTEX_HELD(&pio->io_lock));
514 515
515 516 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
516 517 if (*zl == NULL)
517 518 return (NULL);
518 519
519 520 ASSERT((*zl)->zl_parent == pio);
520 521 return ((*zl)->zl_child);
521 522 }
522 523
523 524 zio_t *
524 525 zio_unique_parent(zio_t *cio)
525 526 {
526 527 zio_link_t *zl = NULL;
527 528 zio_t *pio = zio_walk_parents(cio, &zl);
528 529
529 530 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
530 531 return (pio);
531 532 }
532 533
533 534 void
534 535 zio_add_child(zio_t *pio, zio_t *cio)
535 536 {
536 537 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
537 538
538 539 /*
539 540 * Logical I/Os can have logical, gang, or vdev children.
540 541 * Gang I/Os can have gang or vdev children.
541 542 * Vdev I/Os can only have vdev children.
542 543 * The following ASSERT captures all of these constraints.
543 544 */
544 545 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
545 546
546 547 zl->zl_parent = pio;
547 548 zl->zl_child = cio;
548 549
549 550 mutex_enter(&pio->io_lock);
550 551 mutex_enter(&cio->io_lock);
551 552
552 553 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
553 554
554 555 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
555 556 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
556 557
557 558 list_insert_head(&pio->io_child_list, zl);
558 559 list_insert_head(&cio->io_parent_list, zl);
559 560
560 561 pio->io_child_count++;
561 562 cio->io_parent_count++;
562 563
563 564 mutex_exit(&cio->io_lock);
564 565 mutex_exit(&pio->io_lock);
565 566 }
566 567
567 568 static void
568 569 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
569 570 {
570 571 ASSERT(zl->zl_parent == pio);
571 572 ASSERT(zl->zl_child == cio);
572 573
573 574 mutex_enter(&pio->io_lock);
574 575 mutex_enter(&cio->io_lock);
575 576
576 577 list_remove(&pio->io_child_list, zl);
577 578 list_remove(&cio->io_parent_list, zl);
578 579
579 580 pio->io_child_count--;
580 581 cio->io_parent_count--;
581 582
582 583 mutex_exit(&cio->io_lock);
583 584 mutex_exit(&pio->io_lock);
584 585
585 586 kmem_cache_free(zio_link_cache, zl);
586 587 }
587 588
588 589 static boolean_t
589 590 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
590 591 {
591 592 boolean_t waiting = B_FALSE;
592 593
593 594 mutex_enter(&zio->io_lock);
594 595 ASSERT(zio->io_stall == NULL);
595 596 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
596 597 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
597 598 continue;
598 599
599 600 uint64_t *countp = &zio->io_children[c][wait];
600 601 if (*countp != 0) {
601 602 zio->io_stage >>= 1;
602 603 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
603 604 zio->io_stall = countp;
604 605 waiting = B_TRUE;
605 606 break;
606 607 }
607 608 }
608 609 mutex_exit(&zio->io_lock);
609 610 return (waiting);
610 611 }
611 612
612 613 static void
613 614 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
614 615 {
615 616 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
616 617 int *errorp = &pio->io_child_error[zio->io_child_type];
617 618
618 619 mutex_enter(&pio->io_lock);
619 620 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
620 621 *errorp = zio_worst_error(*errorp, zio->io_error);
621 622 pio->io_reexecute |= zio->io_reexecute;
622 623 ASSERT3U(*countp, >, 0);
623 624
624 625 (*countp)--;
625 626
626 627 if (*countp == 0 && pio->io_stall == countp) {
627 628 zio_taskq_type_t type =
628 629 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
629 630 ZIO_TASKQ_INTERRUPT;
630 631 pio->io_stall = NULL;
631 632 mutex_exit(&pio->io_lock);
632 633 /*
633 634 * Dispatch the parent zio in its own taskq so that
634 635 * the child can continue to make progress. This also
635 636 * prevents overflowing the stack when we have deeply nested
636 637 * parent-child relationships.
637 638 */
638 639 zio_taskq_dispatch(pio, type, B_FALSE);
639 640 } else {
640 641 mutex_exit(&pio->io_lock);
641 642 }
642 643 }
643 644
644 645 static void
645 646 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
646 647 {
647 648 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
648 649 zio->io_error = zio->io_child_error[c];
649 650 }
650 651
651 652 int
652 653 zio_bookmark_compare(const void *x1, const void *x2)
653 654 {
654 655 const zio_t *z1 = x1;
655 656 const zio_t *z2 = x2;
656 657
657 658 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
658 659 return (-1);
659 660 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
660 661 return (1);
661 662
662 663 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
663 664 return (-1);
664 665 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
665 666 return (1);
666 667
667 668 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
668 669 return (-1);
669 670 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
670 671 return (1);
671 672
672 673 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
673 674 return (-1);
674 675 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
675 676 return (1);
676 677
677 678 if (z1 < z2)
678 679 return (-1);
679 680 if (z1 > z2)
680 681 return (1);
681 682
682 683 return (0);
683 684 }
684 685
685 686 /*
686 687 * ==========================================================================
687 688 * Create the various types of I/O (read, write, free, etc)
688 689 * ==========================================================================
689 690 */
690 691 static zio_t *
691 692 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
692 693 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
693 694 void *private, zio_type_t type, zio_priority_t priority,
694 695 enum zio_flag flags, vdev_t *vd, uint64_t offset,
695 696 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
696 697 {
697 698 zio_t *zio;
698 699
699 700 IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE);
700 701 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
701 702 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
702 703
703 704 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
704 705 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
705 706 ASSERT(vd || stage == ZIO_STAGE_OPEN);
706 707
707 708 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0);
708 709
709 710 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
710 711 bzero(zio, sizeof (zio_t));
711 712
712 713 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
713 714 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
714 715
715 716 list_create(&zio->io_parent_list, sizeof (zio_link_t),
716 717 offsetof(zio_link_t, zl_parent_node));
717 718 list_create(&zio->io_child_list, sizeof (zio_link_t),
718 719 offsetof(zio_link_t, zl_child_node));
719 720 metaslab_trace_init(&zio->io_alloc_list);
720 721
721 722 if (vd != NULL)
722 723 zio->io_child_type = ZIO_CHILD_VDEV;
723 724 else if (flags & ZIO_FLAG_GANG_CHILD)
724 725 zio->io_child_type = ZIO_CHILD_GANG;
725 726 else if (flags & ZIO_FLAG_DDT_CHILD)
726 727 zio->io_child_type = ZIO_CHILD_DDT;
727 728 else
728 729 zio->io_child_type = ZIO_CHILD_LOGICAL;
729 730
730 731 if (bp != NULL) {
731 732 zio->io_bp = (blkptr_t *)bp;
732 733 zio->io_bp_copy = *bp;
733 734 zio->io_bp_orig = *bp;
734 735 if (type != ZIO_TYPE_WRITE ||
735 736 zio->io_child_type == ZIO_CHILD_DDT)
736 737 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
737 738 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
738 739 zio->io_logical = zio;
739 740 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
740 741 pipeline |= ZIO_GANG_STAGES;
741 742 }
742 743
743 744 zio->io_spa = spa;
744 745 zio->io_txg = txg;
745 746 zio->io_done = done;
746 747 zio->io_private = private;
747 748 zio->io_type = type;
748 749 zio->io_priority = priority;
749 750 zio->io_vd = vd;
750 751 zio->io_offset = offset;
751 752 zio->io_orig_abd = zio->io_abd = data;
752 753 zio->io_orig_size = zio->io_size = psize;
753 754 zio->io_lsize = lsize;
754 755 zio->io_orig_flags = zio->io_flags = flags;
755 756 zio->io_orig_stage = zio->io_stage = stage;
756 757 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
757 758 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
758 759
759 760 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
760 761 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
761 762
762 763 if (zb != NULL)
763 764 zio->io_bookmark = *zb;
764 765
765 766 if (pio != NULL) {
766 767 if (zio->io_metaslab_class == NULL)
767 768 zio->io_metaslab_class = pio->io_metaslab_class;
768 769 if (zio->io_logical == NULL)
769 770 zio->io_logical = pio->io_logical;
770 771 if (zio->io_child_type == ZIO_CHILD_GANG)
771 772 zio->io_gang_leader = pio->io_gang_leader;
772 773 zio_add_child(pio, zio);
773 774 }
774 775
775 776 return (zio);
776 777 }
777 778
778 779 static void
779 780 zio_destroy(zio_t *zio)
780 781 {
781 782 metaslab_trace_fini(&zio->io_alloc_list);
782 783 list_destroy(&zio->io_parent_list);
783 784 list_destroy(&zio->io_child_list);
784 785 mutex_destroy(&zio->io_lock);
785 786 cv_destroy(&zio->io_cv);
786 787 kmem_cache_free(zio_cache, zio);
787 788 }
788 789
789 790 zio_t *
790 791 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
791 792 void *private, enum zio_flag flags)
792 793 {
793 794 zio_t *zio;
794 795
795 796 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
796 797 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
797 798 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
798 799
799 800 return (zio);
800 801 }
801 802
802 803 zio_t *
803 804 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
804 805 {
805 806 return (zio_null(NULL, spa, NULL, done, private, flags));
806 807 }
807 808
808 809 void
809 810 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
810 811 {
811 812 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
812 813 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
813 814 bp, (longlong_t)BP_GET_TYPE(bp));
814 815 }
815 816 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
816 817 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
817 818 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
818 819 bp, (longlong_t)BP_GET_CHECKSUM(bp));
819 820 }
820 821 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
821 822 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
822 823 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
823 824 bp, (longlong_t)BP_GET_COMPRESS(bp));
824 825 }
825 826 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
826 827 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
827 828 bp, (longlong_t)BP_GET_LSIZE(bp));
828 829 }
829 830 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
830 831 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
831 832 bp, (longlong_t)BP_GET_PSIZE(bp));
832 833 }
833 834
834 835 if (BP_IS_EMBEDDED(bp)) {
835 836 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
836 837 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
837 838 bp, (longlong_t)BPE_GET_ETYPE(bp));
838 839 }
839 840 }
840 841
841 842 /*
842 843 * Do not verify individual DVAs if the config is not trusted. This
843 844 * will be done once the zio is executed in vdev_mirror_map_alloc.
844 845 */
845 846 if (!spa->spa_trust_config)
846 847 return;
847 848
848 849 /*
849 850 * Pool-specific checks.
850 851 *
851 852 * Note: it would be nice to verify that the blk_birth and
852 853 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
853 854 * allows the birth time of log blocks (and dmu_sync()-ed blocks
854 855 * that are in the log) to be arbitrarily large.
855 856 */
856 857 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
857 858 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
858 859 if (vdevid >= spa->spa_root_vdev->vdev_children) {
859 860 zfs_panic_recover("blkptr at %p DVA %u has invalid "
860 861 "VDEV %llu",
861 862 bp, i, (longlong_t)vdevid);
862 863 continue;
863 864 }
864 865 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
865 866 if (vd == NULL) {
866 867 zfs_panic_recover("blkptr at %p DVA %u has invalid "
867 868 "VDEV %llu",
868 869 bp, i, (longlong_t)vdevid);
869 870 continue;
870 871 }
871 872 if (vd->vdev_ops == &vdev_hole_ops) {
872 873 zfs_panic_recover("blkptr at %p DVA %u has hole "
873 874 "VDEV %llu",
874 875 bp, i, (longlong_t)vdevid);
875 876 continue;
876 877 }
877 878 if (vd->vdev_ops == &vdev_missing_ops) {
878 879 /*
879 880 * "missing" vdevs are valid during import, but we
880 881 * don't have their detailed info (e.g. asize), so
881 882 * we can't perform any more checks on them.
882 883 */
883 884 continue;
884 885 }
885 886 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
886 887 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
887 888 if (BP_IS_GANG(bp))
888 889 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
889 890 if (offset + asize > vd->vdev_asize) {
890 891 zfs_panic_recover("blkptr at %p DVA %u has invalid "
891 892 "OFFSET %llu",
892 893 bp, i, (longlong_t)offset);
893 894 }
894 895 }
895 896 }
896 897
897 898 boolean_t
898 899 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
899 900 {
900 901 uint64_t vdevid = DVA_GET_VDEV(dva);
901 902
902 903 if (vdevid >= spa->spa_root_vdev->vdev_children)
903 904 return (B_FALSE);
904 905
905 906 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
906 907 if (vd == NULL)
907 908 return (B_FALSE);
908 909
909 910 if (vd->vdev_ops == &vdev_hole_ops)
910 911 return (B_FALSE);
911 912
912 913 if (vd->vdev_ops == &vdev_missing_ops) {
913 914 return (B_FALSE);
914 915 }
915 916
916 917 uint64_t offset = DVA_GET_OFFSET(dva);
917 918 uint64_t asize = DVA_GET_ASIZE(dva);
918 919
919 920 if (BP_IS_GANG(bp))
920 921 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
921 922 if (offset + asize > vd->vdev_asize)
922 923 return (B_FALSE);
923 924
924 925 return (B_TRUE);
925 926 }
926 927
927 928 zio_t *
928 929 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
929 930 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
930 931 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
931 932 {
932 933 zio_t *zio;
933 934
934 935 zfs_blkptr_verify(spa, bp);
935 936
936 937 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
937 938 data, size, size, done, private,
938 939 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
939 940 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
940 941 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
941 942
942 943 return (zio);
943 944 }
944 945
945 946 zio_t *
946 947 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
947 948 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
948 949 zio_done_func_t *ready, zio_done_func_t *children_ready,
949 950 zio_done_func_t *physdone, zio_done_func_t *done,
950 951 void *private, zio_priority_t priority, enum zio_flag flags,
951 952 const zbookmark_phys_t *zb)
952 953 {
953 954 zio_t *zio;
954 955
955 956 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
956 957 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
957 958 zp->zp_compress >= ZIO_COMPRESS_OFF &&
958 959 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
959 960 DMU_OT_IS_VALID(zp->zp_type) &&
960 961 zp->zp_level < 32 &&
961 962 zp->zp_copies > 0 &&
962 963 zp->zp_copies <= spa_max_replication(spa));
963 964
964 965 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
965 966 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
966 967 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
967 968 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
968 969
969 970 zio->io_ready = ready;
970 971 zio->io_children_ready = children_ready;
971 972 zio->io_physdone = physdone;
972 973 zio->io_prop = *zp;
973 974
974 975 /*
975 976 * Data can be NULL if we are going to call zio_write_override() to
976 977 * provide the already-allocated BP. But we may need the data to
977 978 * verify a dedup hit (if requested). In this case, don't try to
978 979 * dedup (just take the already-allocated BP verbatim). Encrypted
979 980 * dedup blocks need data as well so we also disable dedup in this
980 981 * case.
981 982 */
982 983 if (data == NULL &&
983 984 (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) {
984 985 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
985 986 }
986 987
987 988 return (zio);
988 989 }
989 990
990 991 zio_t *
991 992 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
992 993 uint64_t size, zio_done_func_t *done, void *private,
993 994 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
994 995 {
995 996 zio_t *zio;
996 997
997 998 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
998 999 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
999 1000 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
1000 1001
1001 1002 return (zio);
1002 1003 }
1003 1004
1004 1005 void
1005 1006 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
1006 1007 {
1007 1008 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1008 1009 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1009 1010 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1010 1011 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
1011 1012
1012 1013 /*
1013 1014 * We must reset the io_prop to match the values that existed
1014 1015 * when the bp was first written by dmu_sync() keeping in mind
1015 1016 * that nopwrite and dedup are mutually exclusive.
1016 1017 */
1017 1018 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
1018 1019 zio->io_prop.zp_nopwrite = nopwrite;
1019 1020 zio->io_prop.zp_copies = copies;
1020 1021 zio->io_bp_override = bp;
1021 1022 }
1022 1023
1023 1024 void
1024 1025 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
1025 1026 {
1026 1027
1027 1028 zfs_blkptr_verify(spa, bp);
1028 1029
1029 1030 /*
1030 1031 * The check for EMBEDDED is a performance optimization. We
1031 1032 * process the free here (by ignoring it) rather than
1032 1033 * putting it on the list and then processing it in zio_free_sync().
1033 1034 */
1034 1035 if (BP_IS_EMBEDDED(bp))
1035 1036 return;
1036 1037 metaslab_check_free(spa, bp);
1037 1038
1038 1039 /*
1039 1040 * Frees that are for the currently-syncing txg, are not going to be
1040 1041 * deferred, and which will not need to do a read (i.e. not GANG or
1041 1042 * DEDUP), can be processed immediately. Otherwise, put them on the
1042 1043 * in-memory list for later processing.
1043 1044 *
1044 1045 * Note that we only defer frees after zfs_sync_pass_deferred_free
1045 1046 * when the log space map feature is disabled. [see relevant comment
1046 1047 * in spa_sync_iterate_to_convergence()]
1047 1048 */
1048 1049 if (BP_IS_GANG(bp) ||
1049 1050 BP_GET_DEDUP(bp) ||
1050 1051 txg != spa->spa_syncing_txg ||
1051 1052 (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free &&
1052 1053 !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))) {
1053 1054 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1054 1055 } else {
1055 1056 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
1056 1057 }
1057 1058 }
1058 1059
1059 1060 zio_t *
1060 1061 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1061 1062 enum zio_flag flags)
1062 1063 {
1063 1064 zio_t *zio;
1064 1065 enum zio_stage stage = ZIO_FREE_PIPELINE;
1065 1066
1066 1067 ASSERT(!BP_IS_HOLE(bp));
1067 1068 ASSERT(spa_syncing_txg(spa) == txg);
1068 1069
1069 1070 if (BP_IS_EMBEDDED(bp))
1070 1071 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1071 1072
1072 1073 metaslab_check_free(spa, bp);
1073 1074 arc_freed(spa, bp);
1074 1075 dsl_scan_freed(spa, bp);
1075 1076
1076 1077 /*
1077 1078 * GANG and DEDUP blocks can induce a read (for the gang block header,
1078 1079 * or the DDT), so issue them asynchronously so that this thread is
1079 1080 * not tied up.
1080 1081 */
1081 1082 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
1082 1083 stage |= ZIO_STAGE_ISSUE_ASYNC;
1083 1084
1084 1085 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1085 1086 BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1086 1087 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
1087 1088
1088 1089 return (zio);
1089 1090 }
1090 1091
1091 1092 zio_t *
1092 1093 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1093 1094 zio_done_func_t *done, void *private, enum zio_flag flags)
1094 1095 {
1095 1096 zio_t *zio;
1096 1097
1097 1098 zfs_blkptr_verify(spa, bp);
1098 1099
1099 1100 if (BP_IS_EMBEDDED(bp))
1100 1101 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1101 1102
1102 1103 /*
1103 1104 * A claim is an allocation of a specific block. Claims are needed
1104 1105 * to support immediate writes in the intent log. The issue is that
1105 1106 * immediate writes contain committed data, but in a txg that was
1106 1107 * *not* committed. Upon opening the pool after an unclean shutdown,
1107 1108 * the intent log claims all blocks that contain immediate write data
1108 1109 * so that the SPA knows they're in use.
1109 1110 *
1110 1111 * All claims *must* be resolved in the first txg -- before the SPA
1111 1112 * starts allocating blocks -- so that nothing is allocated twice.
1112 1113 * If txg == 0 we just verify that the block is claimable.
1113 1114 */
1114 1115 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1115 1116 spa_min_claim_txg(spa));
1116 1117 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1117 1118 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
1118 1119
1119 1120 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1120 1121 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1121 1122 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1122 1123 ASSERT0(zio->io_queued_timestamp);
1123 1124
1124 1125 return (zio);
1125 1126 }
1126 1127
1127 1128 zio_t *
1128 1129 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
1129 1130 zio_done_func_t *done, void *private, enum zio_flag flags)
1130 1131 {
1131 1132 zio_t *zio;
1132 1133 int c;
1133 1134
1134 1135 if (vd->vdev_children == 0) {
1135 1136 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1136 1137 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1137 1138 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1138 1139
1139 1140 zio->io_cmd = cmd;
1140 1141 } else {
1141 1142 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1142 1143
1143 1144 for (c = 0; c < vd->vdev_children; c++)
1144 1145 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1145 1146 done, private, flags));
1146 1147 }
1147 1148
1148 1149 return (zio);
1149 1150 }
1150 1151
1151 1152 zio_t *
1152 1153 zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1153 1154 zio_done_func_t *done, void *private, zio_priority_t priority,
1154 1155 enum zio_flag flags, enum trim_flag trim_flags)
1155 1156 {
1156 1157 zio_t *zio;
1157 1158
1158 1159 ASSERT0(vd->vdev_children);
1159 1160 ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
1160 1161 ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1161 1162 ASSERT3U(size, !=, 0);
1162 1163
1163 1164 zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done,
1164 1165 private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL,
1165 1166 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE);
1166 1167 zio->io_trim_flags = trim_flags;
1167 1168
1168 1169 return (zio);
1169 1170 }
1170 1171
1171 1172 zio_t *
1172 1173 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1173 1174 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1174 1175 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1175 1176 {
1176 1177 zio_t *zio;
1177 1178
1178 1179 ASSERT(vd->vdev_children == 0);
1179 1180 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1180 1181 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1181 1182 ASSERT3U(offset + size, <=, vd->vdev_psize);
1182 1183
1183 1184 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1184 1185 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1185 1186 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1186 1187
1187 1188 zio->io_prop.zp_checksum = checksum;
1188 1189
1189 1190 return (zio);
1190 1191 }
1191 1192
1192 1193 zio_t *
1193 1194 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1194 1195 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1195 1196 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1196 1197 {
1197 1198 zio_t *zio;
1198 1199
1199 1200 ASSERT(vd->vdev_children == 0);
1200 1201 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1201 1202 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1202 1203 ASSERT3U(offset + size, <=, vd->vdev_psize);
1203 1204
1204 1205 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1205 1206 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1206 1207 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1207 1208
1208 1209 zio->io_prop.zp_checksum = checksum;
1209 1210
1210 1211 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1211 1212 /*
1212 1213 * zec checksums are necessarily destructive -- they modify
1213 1214 * the end of the write buffer to hold the verifier/checksum.
1214 1215 * Therefore, we must make a local copy in case the data is
1215 1216 * being written to multiple places in parallel.
1216 1217 */
1217 1218 abd_t *wbuf = abd_alloc_sametype(data, size);
1218 1219 abd_copy(wbuf, data, size);
1219 1220
1220 1221 zio_push_transform(zio, wbuf, size, size, NULL);
1221 1222 }
1222 1223
1223 1224 return (zio);
1224 1225 }
1225 1226
1226 1227 /*
1227 1228 * Create a child I/O to do some work for us.
1228 1229 */
1229 1230 zio_t *
1230 1231 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1231 1232 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1232 1233 enum zio_flag flags, zio_done_func_t *done, void *private)
1233 1234 {
1234 1235 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1235 1236 zio_t *zio;
1236 1237
1237 1238 /*
1238 1239 * vdev child I/Os do not propagate their error to the parent.
1239 1240 * Therefore, for correct operation the caller *must* check for
1240 1241 * and handle the error in the child i/o's done callback.
1241 1242 * The only exceptions are i/os that we don't care about
1242 1243 * (OPTIONAL or REPAIR).
1243 1244 */
1244 1245 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1245 1246 done != NULL);
1246 1247
1247 1248 if (type == ZIO_TYPE_READ && bp != NULL) {
1248 1249 /*
1249 1250 * If we have the bp, then the child should perform the
1250 1251 * checksum and the parent need not. This pushes error
1251 1252 * detection as close to the leaves as possible and
1252 1253 * eliminates redundant checksums in the interior nodes.
1253 1254 */
1254 1255 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1255 1256 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1256 1257 }
1257 1258
1258 1259 if (vd->vdev_ops->vdev_op_leaf) {
1259 1260 ASSERT0(vd->vdev_children);
1260 1261 offset += VDEV_LABEL_START_SIZE;
1261 1262 }
1262 1263
1263 1264 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1264 1265
1265 1266 /*
1266 1267 * If we've decided to do a repair, the write is not speculative --
1267 1268 * even if the original read was.
1268 1269 */
1269 1270 if (flags & ZIO_FLAG_IO_REPAIR)
1270 1271 flags &= ~ZIO_FLAG_SPECULATIVE;
1271 1272
1272 1273 /*
1273 1274 * If we're creating a child I/O that is not associated with a
1274 1275 * top-level vdev, then the child zio is not an allocating I/O.
1275 1276 * If this is a retried I/O then we ignore it since we will
1276 1277 * have already processed the original allocating I/O.
1277 1278 */
1278 1279 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1279 1280 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1280 1281 ASSERT(pio->io_metaslab_class != NULL);
1281 1282 ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled);
1282 1283 ASSERT(type == ZIO_TYPE_WRITE);
1283 1284 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1284 1285 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1285 1286 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1286 1287 pio->io_child_type == ZIO_CHILD_GANG);
1287 1288
1288 1289 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1289 1290 }
1290 1291
1291 1292 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1292 1293 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1293 1294 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1294 1295 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1295 1296
1296 1297 zio->io_physdone = pio->io_physdone;
1297 1298 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1298 1299 zio->io_logical->io_phys_children++;
1299 1300
1300 1301 return (zio);
1301 1302 }
1302 1303
1303 1304 zio_t *
1304 1305 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1305 1306 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
1306 1307 zio_done_func_t *done, void *private)
1307 1308 {
1308 1309 zio_t *zio;
1309 1310
1310 1311 ASSERT(vd->vdev_ops->vdev_op_leaf);
1311 1312
1312 1313 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1313 1314 data, size, size, done, private, type, priority,
1314 1315 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1315 1316 vd, offset, NULL,
1316 1317 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1317 1318
1318 1319 return (zio);
1319 1320 }
1320 1321
1321 1322 void
1322 1323 zio_flush(zio_t *zio, vdev_t *vd)
1323 1324 {
1324 1325 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1325 1326 NULL, NULL,
1326 1327 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1327 1328 }
1328 1329
1329 1330 void
1330 1331 zio_shrink(zio_t *zio, uint64_t size)
1331 1332 {
1332 1333 ASSERT3P(zio->io_executor, ==, NULL);
1333 1334 ASSERT3P(zio->io_orig_size, ==, zio->io_size);
1334 1335 ASSERT3U(size, <=, zio->io_size);
1335 1336
1336 1337 /*
1337 1338 * We don't shrink for raidz because of problems with the
1338 1339 * reconstruction when reading back less than the block size.
1339 1340 * Note, BP_IS_RAIDZ() assumes no compression.
1340 1341 */
1341 1342 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1342 1343 if (!BP_IS_RAIDZ(zio->io_bp)) {
1343 1344 /* we are not doing a raw write */
1344 1345 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1345 1346 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1346 1347 }
1347 1348 }
1348 1349
1349 1350 /*
1350 1351 * ==========================================================================
1351 1352 * Prepare to read and write logical blocks
1352 1353 * ==========================================================================
1353 1354 */
1354 1355
1355 1356 static int
1356 1357 zio_read_bp_init(zio_t *zio)
1357 1358 {
1358 1359 blkptr_t *bp = zio->io_bp;
1359 1360 uint64_t psize =
1360 1361 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1361 1362
1362 1363 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1363 1364
1364 1365 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1365 1366 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1366 1367 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1367 1368 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1368 1369 psize, psize, zio_decompress);
1369 1370 }
1370 1371
1371 1372 if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) ||
1372 1373 BP_HAS_INDIRECT_MAC_CKSUM(bp)) &&
1373 1374 zio->io_child_type == ZIO_CHILD_LOGICAL) {
1374 1375 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1375 1376 psize, psize, zio_decrypt);
1376 1377 }
1377 1378
1378 1379 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1379 1380 int psize = BPE_GET_PSIZE(bp);
1380 1381 void *data = abd_borrow_buf(zio->io_abd, psize);
1381 1382
1382 1383 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1383 1384 decode_embedded_bp_compressed(bp, data);
1384 1385 abd_return_buf_copy(zio->io_abd, data, psize);
1385 1386 } else {
1386 1387 ASSERT(!BP_IS_EMBEDDED(bp));
1387 1388 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1388 1389 }
1389 1390
1390 1391 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1391 1392 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1392 1393
1393 1394 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1394 1395 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1395 1396
1396 1397 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1397 1398 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1398 1399
1399 1400 return (ZIO_PIPELINE_CONTINUE);
1400 1401 }
1401 1402
1402 1403 static int
1403 1404 zio_write_bp_init(zio_t *zio)
1404 1405 {
1405 1406 if (!IO_IS_ALLOCATING(zio))
1406 1407 return (ZIO_PIPELINE_CONTINUE);
1407 1408
1408 1409 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1409 1410
1410 1411 if (zio->io_bp_override) {
1411 1412 blkptr_t *bp = zio->io_bp;
1412 1413 zio_prop_t *zp = &zio->io_prop;
1413 1414
1414 1415 ASSERT(bp->blk_birth != zio->io_txg);
1415 1416 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1416 1417
1417 1418 *bp = *zio->io_bp_override;
1418 1419 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1419 1420
1420 1421 if (BP_IS_EMBEDDED(bp))
1421 1422 return (ZIO_PIPELINE_CONTINUE);
1422 1423
1423 1424 /*
1424 1425 * If we've been overridden and nopwrite is set then
1425 1426 * set the flag accordingly to indicate that a nopwrite
1426 1427 * has already occurred.
1427 1428 */
1428 1429 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1429 1430 ASSERT(!zp->zp_dedup);
1430 1431 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1431 1432 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1432 1433 return (ZIO_PIPELINE_CONTINUE);
1433 1434 }
1434 1435
1435 1436 ASSERT(!zp->zp_nopwrite);
1436 1437
1437 1438 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1438 1439 return (ZIO_PIPELINE_CONTINUE);
1439 1440
1440 1441 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1441 1442 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1442 1443
1443 1444 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum &&
1444 1445 !zp->zp_encrypt) {
1445 1446 BP_SET_DEDUP(bp, 1);
1446 1447 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1447 1448 return (ZIO_PIPELINE_CONTINUE);
1448 1449 }
1449 1450
1450 1451 /*
1451 1452 * We were unable to handle this as an override bp, treat
1452 1453 * it as a regular write I/O.
1453 1454 */
1454 1455 zio->io_bp_override = NULL;
1455 1456 *bp = zio->io_bp_orig;
1456 1457 zio->io_pipeline = zio->io_orig_pipeline;
1457 1458 }
1458 1459
1459 1460 return (ZIO_PIPELINE_CONTINUE);
1460 1461 }
1461 1462
1462 1463 static int
1463 1464 zio_write_compress(zio_t *zio)
1464 1465 {
1465 1466 spa_t *spa = zio->io_spa;
1466 1467 zio_prop_t *zp = &zio->io_prop;
1467 1468 enum zio_compress compress = zp->zp_compress;
1468 1469 blkptr_t *bp = zio->io_bp;
1469 1470 uint64_t lsize = zio->io_lsize;
1470 1471 uint64_t psize = zio->io_size;
1471 1472 int pass = 1;
1472 1473
1473 1474 /*
1474 1475 * If our children haven't all reached the ready stage,
1475 1476 * wait for them and then repeat this pipeline stage.
1476 1477 */
1477 1478 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1478 1479 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1479 1480 return (ZIO_PIPELINE_STOP);
1480 1481 }
1481 1482
1482 1483 if (!IO_IS_ALLOCATING(zio))
1483 1484 return (ZIO_PIPELINE_CONTINUE);
1484 1485
1485 1486 if (zio->io_children_ready != NULL) {
1486 1487 /*
1487 1488 * Now that all our children are ready, run the callback
1488 1489 * associated with this zio in case it wants to modify the
1489 1490 * data to be written.
1490 1491 */
1491 1492 ASSERT3U(zp->zp_level, >, 0);
1492 1493 zio->io_children_ready(zio);
1493 1494 }
1494 1495
1495 1496 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1496 1497 ASSERT(zio->io_bp_override == NULL);
1497 1498
1498 1499 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1499 1500 /*
1500 1501 * We're rewriting an existing block, which means we're
1501 1502 * working on behalf of spa_sync(). For spa_sync() to
1502 1503 * converge, it must eventually be the case that we don't
1503 1504 * have to allocate new blocks. But compression changes
1504 1505 * the blocksize, which forces a reallocate, and makes
1505 1506 * convergence take longer. Therefore, after the first
1506 1507 * few passes, stop compressing to ensure convergence.
1507 1508 */
1508 1509 pass = spa_sync_pass(spa);
1509 1510
1510 1511 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1511 1512 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1512 1513 ASSERT(!BP_GET_DEDUP(bp));
1513 1514
1514 1515 if (pass >= zfs_sync_pass_dont_compress)
1515 1516 compress = ZIO_COMPRESS_OFF;
1516 1517
1517 1518 /* Make sure someone doesn't change their mind on overwrites */
1518 1519 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1519 1520 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1520 1521 }
1521 1522
1522 1523 /* If it's a compressed write that is not raw, compress the buffer. */
1523 1524 if (compress != ZIO_COMPRESS_OFF &&
1524 1525 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1525 1526 void *cbuf = zio_buf_alloc(lsize);
1526 1527 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1527 1528 if (psize == 0 || psize == lsize) {
1528 1529 compress = ZIO_COMPRESS_OFF;
1529 1530 zio_buf_free(cbuf, lsize);
1530 1531 } else if (!zp->zp_dedup && !zp->zp_encrypt &&
1531 1532 psize <= BPE_PAYLOAD_SIZE &&
1532 1533 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1533 1534 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1534 1535 encode_embedded_bp_compressed(bp,
1535 1536 cbuf, compress, lsize, psize);
1536 1537 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1537 1538 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1538 1539 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1539 1540 zio_buf_free(cbuf, lsize);
1540 1541 bp->blk_birth = zio->io_txg;
1541 1542 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1542 1543 ASSERT(spa_feature_is_active(spa,
1543 1544 SPA_FEATURE_EMBEDDED_DATA));
1544 1545 return (ZIO_PIPELINE_CONTINUE);
1545 1546 } else {
1546 1547 /*
1547 1548 * Round up compressed size up to the ashift
1548 1549 * of the smallest-ashift device, and zero the tail.
1549 1550 * This ensures that the compressed size of the BP
1550 1551 * (and thus compressratio property) are correct,
1551 1552 * in that we charge for the padding used to fill out
1552 1553 * the last sector.
1553 1554 */
1554 1555 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1555 1556 size_t rounded = (size_t)P2ROUNDUP(psize,
1556 1557 1ULL << spa->spa_min_ashift);
1557 1558 if (rounded >= lsize) {
1558 1559 compress = ZIO_COMPRESS_OFF;
1559 1560 zio_buf_free(cbuf, lsize);
1560 1561 psize = lsize;
1561 1562 } else {
1562 1563 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1563 1564 abd_take_ownership_of_buf(cdata, B_TRUE);
1564 1565 abd_zero_off(cdata, psize, rounded - psize);
1565 1566 psize = rounded;
1566 1567 zio_push_transform(zio, cdata,
1567 1568 psize, lsize, NULL);
1568 1569 }
1569 1570 }
1570 1571
1571 1572 /*
1572 1573 * We were unable to handle this as an override bp, treat
1573 1574 * it as a regular write I/O.
1574 1575 */
1575 1576 zio->io_bp_override = NULL;
1576 1577 *bp = zio->io_bp_orig;
1577 1578 zio->io_pipeline = zio->io_orig_pipeline;
1578 1579
1579 1580 } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 &&
1580 1581 zp->zp_type == DMU_OT_DNODE) {
1581 1582 /*
1582 1583 * The DMU actually relies on the zio layer's compression
1583 1584 * to free metadnode blocks that have had all contained
1584 1585 * dnodes freed. As a result, even when doing a raw
1585 1586 * receive, we must check whether the block can be compressed
1586 1587 * to a hole.
1587 1588 */
1588 1589 psize = zio_compress_data(ZIO_COMPRESS_EMPTY,
1589 1590 zio->io_abd, NULL, lsize);
1590 1591 if (psize == 0)
1591 1592 compress = ZIO_COMPRESS_OFF;
1592 1593 } else {
1593 1594 ASSERT3U(psize, !=, 0);
1594 1595 }
1595 1596
1596 1597 /*
1597 1598 * The final pass of spa_sync() must be all rewrites, but the first
1598 1599 * few passes offer a trade-off: allocating blocks defers convergence,
1599 1600 * but newly allocated blocks are sequential, so they can be written
1600 1601 * to disk faster. Therefore, we allow the first few passes of
1601 1602 * spa_sync() to allocate new blocks, but force rewrites after that.
1602 1603 * There should only be a handful of blocks after pass 1 in any case.
1603 1604 */
1604 1605 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1605 1606 BP_GET_PSIZE(bp) == psize &&
1606 1607 pass >= zfs_sync_pass_rewrite) {
1607 1608 VERIFY3U(psize, !=, 0);
1608 1609 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1609 1610 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1610 1611 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1611 1612 } else {
1612 1613 BP_ZERO(bp);
1613 1614 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1614 1615 }
1615 1616
1616 1617 if (psize == 0) {
1617 1618 if (zio->io_bp_orig.blk_birth != 0 &&
1618 1619 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1619 1620 BP_SET_LSIZE(bp, lsize);
1620 1621 BP_SET_TYPE(bp, zp->zp_type);
1621 1622 BP_SET_LEVEL(bp, zp->zp_level);
1622 1623 BP_SET_BIRTH(bp, zio->io_txg, 0);
1623 1624 }
1624 1625 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1625 1626 } else {
1626 1627 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1627 1628 BP_SET_LSIZE(bp, lsize);
1628 1629 BP_SET_TYPE(bp, zp->zp_type);
1629 1630 BP_SET_LEVEL(bp, zp->zp_level);
1630 1631 BP_SET_PSIZE(bp, psize);
1631 1632 BP_SET_COMPRESS(bp, compress);
1632 1633 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1633 1634 BP_SET_DEDUP(bp, zp->zp_dedup);
1634 1635 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1635 1636 if (zp->zp_dedup) {
1636 1637 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1637 1638 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1638 1639 ASSERT(!zp->zp_encrypt ||
1639 1640 DMU_OT_IS_ENCRYPTED(zp->zp_type));
1640 1641 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1641 1642 }
1642 1643 if (zp->zp_nopwrite) {
1643 1644 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1644 1645 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1645 1646 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1646 1647 }
1647 1648 }
1648 1649 return (ZIO_PIPELINE_CONTINUE);
1649 1650 }
1650 1651
1651 1652 static int
1652 1653 zio_free_bp_init(zio_t *zio)
1653 1654 {
1654 1655 blkptr_t *bp = zio->io_bp;
1655 1656
1656 1657 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1657 1658 if (BP_GET_DEDUP(bp))
1658 1659 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1659 1660 }
1660 1661
1661 1662 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1662 1663
1663 1664 return (ZIO_PIPELINE_CONTINUE);
1664 1665 }
1665 1666
1666 1667 /*
1667 1668 * ==========================================================================
1668 1669 * Execute the I/O pipeline
1669 1670 * ==========================================================================
1670 1671 */
1671 1672
1672 1673 static void
1673 1674 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1674 1675 {
1675 1676 spa_t *spa = zio->io_spa;
1676 1677 zio_type_t t = zio->io_type;
1677 1678 int flags = (cutinline ? TQ_FRONT : 0);
1678 1679
1679 1680 /*
1680 1681 * If we're a config writer or a probe, the normal issue and
1681 1682 * interrupt threads may all be blocked waiting for the config lock.
1682 1683 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1683 1684 */
1684 1685 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1685 1686 t = ZIO_TYPE_NULL;
1686 1687
1687 1688 /*
1688 1689 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1689 1690 */
1690 1691 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1691 1692 t = ZIO_TYPE_NULL;
1692 1693
1693 1694 /*
1694 1695 * If this is a high priority I/O, then use the high priority taskq if
1695 1696 * available.
1696 1697 */
1697 1698 if ((zio->io_priority == ZIO_PRIORITY_NOW ||
1698 1699 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) &&
1699 1700 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1700 1701 q++;
1701 1702
1702 1703 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1703 1704
1704 1705 /*
1705 1706 * NB: We are assuming that the zio can only be dispatched
1706 1707 * to a single taskq at a time. It would be a grievous error
1707 1708 * to dispatch the zio to another taskq at the same time.
1708 1709 */
1709 1710 ASSERT(zio->io_tqent.tqent_next == NULL);
1710 1711 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1711 1712 flags, &zio->io_tqent);
1712 1713 }
1713 1714
1714 1715 static boolean_t
1715 1716 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1716 1717 {
1717 1718 kthread_t *executor = zio->io_executor;
1718 1719 spa_t *spa = zio->io_spa;
1719 1720
1720 1721 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1721 1722 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1722 1723 uint_t i;
1723 1724 for (i = 0; i < tqs->stqs_count; i++) {
1724 1725 if (taskq_member(tqs->stqs_taskq[i], executor))
1725 1726 return (B_TRUE);
1726 1727 }
1727 1728 }
1728 1729
1729 1730 return (B_FALSE);
1730 1731 }
1731 1732
1732 1733 static int
1733 1734 zio_issue_async(zio_t *zio)
1734 1735 {
1735 1736 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1736 1737
1737 1738 return (ZIO_PIPELINE_STOP);
1738 1739 }
1739 1740
1740 1741 void
1741 1742 zio_interrupt(zio_t *zio)
1742 1743 {
1743 1744 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1744 1745 }
1745 1746
1746 1747 void
1747 1748 zio_delay_interrupt(zio_t *zio)
1748 1749 {
1749 1750 /*
1750 1751 * The timeout_generic() function isn't defined in userspace, so
1751 1752 * rather than trying to implement the function, the zio delay
1752 1753 * functionality has been disabled for userspace builds.
1753 1754 */
1754 1755
1755 1756 #ifdef _KERNEL
1756 1757 /*
1757 1758 * If io_target_timestamp is zero, then no delay has been registered
1758 1759 * for this IO, thus jump to the end of this function and "skip" the
1759 1760 * delay; issuing it directly to the zio layer.
1760 1761 */
1761 1762 if (zio->io_target_timestamp != 0) {
1762 1763 hrtime_t now = gethrtime();
1763 1764
1764 1765 if (now >= zio->io_target_timestamp) {
1765 1766 /*
1766 1767 * This IO has already taken longer than the target
1767 1768 * delay to complete, so we don't want to delay it
1768 1769 * any longer; we "miss" the delay and issue it
1769 1770 * directly to the zio layer. This is likely due to
1770 1771 * the target latency being set to a value less than
1771 1772 * the underlying hardware can satisfy (e.g. delay
1772 1773 * set to 1ms, but the disks take 10ms to complete an
1773 1774 * IO request).
1774 1775 */
1775 1776
1776 1777 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1777 1778 hrtime_t, now);
1778 1779
1779 1780 zio_interrupt(zio);
1780 1781 } else {
1781 1782 hrtime_t diff = zio->io_target_timestamp - now;
1782 1783
1783 1784 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1784 1785 hrtime_t, now, hrtime_t, diff);
1785 1786
1786 1787 (void) timeout_generic(CALLOUT_NORMAL,
1787 1788 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1788 1789 }
1789 1790
1790 1791 return;
1791 1792 }
1792 1793 #endif
1793 1794
1794 1795 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1795 1796 zio_interrupt(zio);
1796 1797 }
1797 1798
1798 1799 /*
1799 1800 * Execute the I/O pipeline until one of the following occurs:
1800 1801 *
1801 1802 * (1) the I/O completes
1802 1803 * (2) the pipeline stalls waiting for dependent child I/Os
1803 1804 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1804 1805 * (4) the I/O is delegated by vdev-level caching or aggregation
1805 1806 * (5) the I/O is deferred due to vdev-level queueing
1806 1807 * (6) the I/O is handed off to another thread.
1807 1808 *
1808 1809 * In all cases, the pipeline stops whenever there's no CPU work; it never
1809 1810 * burns a thread in cv_wait().
1810 1811 *
1811 1812 * There's no locking on io_stage because there's no legitimate way
1812 1813 * for multiple threads to be attempting to process the same I/O.
1813 1814 */
1814 1815 static zio_pipe_stage_t *zio_pipeline[];
1815 1816
1816 1817 void
1817 1818 zio_execute(zio_t *zio)
1818 1819 {
1819 1820 zio->io_executor = curthread;
1820 1821
1821 1822 ASSERT3U(zio->io_queued_timestamp, >, 0);
1822 1823
1823 1824 while (zio->io_stage < ZIO_STAGE_DONE) {
1824 1825 enum zio_stage pipeline = zio->io_pipeline;
1825 1826 enum zio_stage stage = zio->io_stage;
1826 1827 int rv;
1827 1828
1828 1829 ASSERT(!MUTEX_HELD(&zio->io_lock));
1829 1830 ASSERT(ISP2(stage));
1830 1831 ASSERT(zio->io_stall == NULL);
1831 1832
1832 1833 do {
1833 1834 stage <<= 1;
1834 1835 } while ((stage & pipeline) == 0);
1835 1836
1836 1837 ASSERT(stage <= ZIO_STAGE_DONE);
1837 1838
1838 1839 /*
1839 1840 * If we are in interrupt context and this pipeline stage
1840 1841 * will grab a config lock that is held across I/O,
1841 1842 * or may wait for an I/O that needs an interrupt thread
1842 1843 * to complete, issue async to avoid deadlock.
1843 1844 *
1844 1845 * For VDEV_IO_START, we cut in line so that the io will
1845 1846 * be sent to disk promptly.
1846 1847 */
1847 1848 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1848 1849 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1849 1850 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1850 1851 zio_requeue_io_start_cut_in_line : B_FALSE;
1851 1852 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1852 1853 return;
1853 1854 }
1854 1855
1855 1856 zio->io_stage = stage;
1856 1857 zio->io_pipeline_trace |= zio->io_stage;
1857 1858 rv = zio_pipeline[highbit64(stage) - 1](zio);
1858 1859
1859 1860 if (rv == ZIO_PIPELINE_STOP)
1860 1861 return;
1861 1862
1862 1863 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1863 1864 }
1864 1865 }
1865 1866
1866 1867 /*
1867 1868 * ==========================================================================
1868 1869 * Initiate I/O, either sync or async
1869 1870 * ==========================================================================
1870 1871 */
1871 1872 int
1872 1873 zio_wait(zio_t *zio)
1873 1874 {
1874 1875 int error;
1875 1876
1876 1877 ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1877 1878 ASSERT3P(zio->io_executor, ==, NULL);
1878 1879
1879 1880 zio->io_waiter = curthread;
1880 1881 ASSERT0(zio->io_queued_timestamp);
1881 1882 zio->io_queued_timestamp = gethrtime();
1882 1883
1883 1884 zio_execute(zio);
1884 1885
1885 1886 mutex_enter(&zio->io_lock);
1886 1887 while (zio->io_executor != NULL)
1887 1888 cv_wait(&zio->io_cv, &zio->io_lock);
1888 1889 mutex_exit(&zio->io_lock);
1889 1890
1890 1891 error = zio->io_error;
1891 1892 zio_destroy(zio);
1892 1893
1893 1894 return (error);
1894 1895 }
1895 1896
1896 1897 void
1897 1898 zio_nowait(zio_t *zio)
1898 1899 {
1899 1900 ASSERT3P(zio->io_executor, ==, NULL);
1900 1901
1901 1902 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1902 1903 zio_unique_parent(zio) == NULL) {
1903 1904 /*
1904 1905 * This is a logical async I/O with no parent to wait for it.
1905 1906 * We add it to the spa_async_root_zio "Godfather" I/O which
1906 1907 * will ensure they complete prior to unloading the pool.
1907 1908 */
1908 1909 spa_t *spa = zio->io_spa;
1909 1910
1910 1911 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1911 1912 }
1912 1913
1913 1914 ASSERT0(zio->io_queued_timestamp);
1914 1915 zio->io_queued_timestamp = gethrtime();
1915 1916 zio_execute(zio);
1916 1917 }
1917 1918
1918 1919 /*
1919 1920 * ==========================================================================
1920 1921 * Reexecute, cancel, or suspend/resume failed I/O
1921 1922 * ==========================================================================
1922 1923 */
1923 1924
1924 1925 static void
1925 1926 zio_reexecute(zio_t *pio)
1926 1927 {
1927 1928 zio_t *cio, *cio_next;
1928 1929
1929 1930 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1930 1931 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1931 1932 ASSERT(pio->io_gang_leader == NULL);
1932 1933 ASSERT(pio->io_gang_tree == NULL);
1933 1934
1934 1935 pio->io_flags = pio->io_orig_flags;
1935 1936 pio->io_stage = pio->io_orig_stage;
1936 1937 pio->io_pipeline = pio->io_orig_pipeline;
1937 1938 pio->io_reexecute = 0;
1938 1939 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1939 1940 pio->io_pipeline_trace = 0;
1940 1941 pio->io_error = 0;
1941 1942 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1942 1943 pio->io_state[w] = 0;
1943 1944 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1944 1945 pio->io_child_error[c] = 0;
1945 1946
1946 1947 if (IO_IS_ALLOCATING(pio))
1947 1948 BP_ZERO(pio->io_bp);
1948 1949
1949 1950 /*
1950 1951 * As we reexecute pio's children, new children could be created.
1951 1952 * New children go to the head of pio's io_child_list, however,
1952 1953 * so we will (correctly) not reexecute them. The key is that
1953 1954 * the remainder of pio's io_child_list, from 'cio_next' onward,
1954 1955 * cannot be affected by any side effects of reexecuting 'cio'.
1955 1956 */
1956 1957 zio_link_t *zl = NULL;
1957 1958 mutex_enter(&pio->io_lock);
1958 1959 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1959 1960 cio_next = zio_walk_children(pio, &zl);
1960 1961 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1961 1962 pio->io_children[cio->io_child_type][w]++;
1962 1963 mutex_exit(&pio->io_lock);
1963 1964 zio_reexecute(cio);
1964 1965 mutex_enter(&pio->io_lock);
1965 1966 }
1966 1967 mutex_exit(&pio->io_lock);
1967 1968
1968 1969 /*
1969 1970 * Now that all children have been reexecuted, execute the parent.
1970 1971 * We don't reexecute "The Godfather" I/O here as it's the
1971 1972 * responsibility of the caller to wait on it.
1972 1973 */
1973 1974 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1974 1975 pio->io_queued_timestamp = gethrtime();
1975 1976 zio_execute(pio);
1976 1977 }
|
↓ open down ↓ |
1940 lines elided |
↑ open up ↑ |
1977 1978 }
1978 1979
1979 1980 void
1980 1981 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
1981 1982 {
1982 1983 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1983 1984 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1984 1985 "failure and the failure mode property for this pool "
1985 1986 "is set to panic.", spa_name(spa));
1986 1987
1988 + cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
1989 + "failure and has been suspended; `zpool clear` will be required "
1990 + "before the pool can be written to.", spa_name(spa));
1991 +
1987 1992 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
1988 1993 NULL, NULL, 0, 0);
1989 1994
1990 1995 mutex_enter(&spa->spa_suspend_lock);
1991 1996
1992 1997 if (spa->spa_suspend_zio_root == NULL)
1993 1998 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1994 1999 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1995 2000 ZIO_FLAG_GODFATHER);
1996 2001
1997 2002 spa->spa_suspended = reason;
1998 2003
1999 2004 if (zio != NULL) {
2000 2005 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2001 2006 ASSERT(zio != spa->spa_suspend_zio_root);
2002 2007 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2003 2008 ASSERT(zio_unique_parent(zio) == NULL);
2004 2009 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
2005 2010 zio_add_child(spa->spa_suspend_zio_root, zio);
2006 2011 }
2007 2012
2008 2013 mutex_exit(&spa->spa_suspend_lock);
2009 2014 }
2010 2015
2011 2016 int
2012 2017 zio_resume(spa_t *spa)
2013 2018 {
2014 2019 zio_t *pio;
2015 2020
2016 2021 /*
2017 2022 * Reexecute all previously suspended i/o.
2018 2023 */
2019 2024 mutex_enter(&spa->spa_suspend_lock);
2020 2025 spa->spa_suspended = ZIO_SUSPEND_NONE;
2021 2026 cv_broadcast(&spa->spa_suspend_cv);
2022 2027 pio = spa->spa_suspend_zio_root;
2023 2028 spa->spa_suspend_zio_root = NULL;
2024 2029 mutex_exit(&spa->spa_suspend_lock);
2025 2030
2026 2031 if (pio == NULL)
2027 2032 return (0);
2028 2033
2029 2034 zio_reexecute(pio);
2030 2035 return (zio_wait(pio));
2031 2036 }
2032 2037
2033 2038 void
2034 2039 zio_resume_wait(spa_t *spa)
2035 2040 {
2036 2041 mutex_enter(&spa->spa_suspend_lock);
2037 2042 while (spa_suspended(spa))
2038 2043 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
2039 2044 mutex_exit(&spa->spa_suspend_lock);
2040 2045 }
2041 2046
2042 2047 /*
2043 2048 * ==========================================================================
2044 2049 * Gang blocks.
2045 2050 *
2046 2051 * A gang block is a collection of small blocks that looks to the DMU
2047 2052 * like one large block. When zio_dva_allocate() cannot find a block
2048 2053 * of the requested size, due to either severe fragmentation or the pool
2049 2054 * being nearly full, it calls zio_write_gang_block() to construct the
2050 2055 * block from smaller fragments.
2051 2056 *
2052 2057 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2053 2058 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2054 2059 * an indirect block: it's an array of block pointers. It consumes
2055 2060 * only one sector and hence is allocatable regardless of fragmentation.
2056 2061 * The gang header's bps point to its gang members, which hold the data.
2057 2062 *
2058 2063 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2059 2064 * as the verifier to ensure uniqueness of the SHA256 checksum.
2060 2065 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2061 2066 * not the gang header. This ensures that data block signatures (needed for
2062 2067 * deduplication) are independent of how the block is physically stored.
2063 2068 *
2064 2069 * Gang blocks can be nested: a gang member may itself be a gang block.
2065 2070 * Thus every gang block is a tree in which root and all interior nodes are
2066 2071 * gang headers, and the leaves are normal blocks that contain user data.
2067 2072 * The root of the gang tree is called the gang leader.
2068 2073 *
2069 2074 * To perform any operation (read, rewrite, free, claim) on a gang block,
2070 2075 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2071 2076 * in the io_gang_tree field of the original logical i/o by recursively
2072 2077 * reading the gang leader and all gang headers below it. This yields
2073 2078 * an in-core tree containing the contents of every gang header and the
2074 2079 * bps for every constituent of the gang block.
2075 2080 *
2076 2081 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2077 2082 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2078 2083 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2079 2084 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2080 2085 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2081 2086 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2082 2087 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2083 2088 * of the gang header plus zio_checksum_compute() of the data to update the
2084 2089 * gang header's blk_cksum as described above.
2085 2090 *
2086 2091 * The two-phase assemble/issue model solves the problem of partial failure --
2087 2092 * what if you'd freed part of a gang block but then couldn't read the
2088 2093 * gang header for another part? Assembling the entire gang tree first
2089 2094 * ensures that all the necessary gang header I/O has succeeded before
2090 2095 * starting the actual work of free, claim, or write. Once the gang tree
2091 2096 * is assembled, free and claim are in-memory operations that cannot fail.
2092 2097 *
2093 2098 * In the event that a gang write fails, zio_dva_unallocate() walks the
2094 2099 * gang tree to immediately free (i.e. insert back into the space map)
2095 2100 * everything we've allocated. This ensures that we don't get ENOSPC
2096 2101 * errors during repeated suspend/resume cycles due to a flaky device.
2097 2102 *
2098 2103 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2099 2104 * the gang tree, we won't modify the block, so we can safely defer the free
2100 2105 * (knowing that the block is still intact). If we *can* assemble the gang
2101 2106 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2102 2107 * each constituent bp and we can allocate a new block on the next sync pass.
2103 2108 *
2104 2109 * In all cases, the gang tree allows complete recovery from partial failure.
2105 2110 * ==========================================================================
2106 2111 */
2107 2112
2108 2113 static void
2109 2114 zio_gang_issue_func_done(zio_t *zio)
2110 2115 {
2111 2116 abd_put(zio->io_abd);
2112 2117 }
2113 2118
2114 2119 static zio_t *
2115 2120 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2116 2121 uint64_t offset)
2117 2122 {
2118 2123 if (gn != NULL)
2119 2124 return (pio);
2120 2125
2121 2126 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2122 2127 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2123 2128 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2124 2129 &pio->io_bookmark));
2125 2130 }
2126 2131
2127 2132 static zio_t *
2128 2133 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2129 2134 uint64_t offset)
2130 2135 {
2131 2136 zio_t *zio;
2132 2137
2133 2138 if (gn != NULL) {
2134 2139 abd_t *gbh_abd =
2135 2140 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2136 2141 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2137 2142 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2138 2143 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2139 2144 &pio->io_bookmark);
2140 2145 /*
2141 2146 * As we rewrite each gang header, the pipeline will compute
2142 2147 * a new gang block header checksum for it; but no one will
2143 2148 * compute a new data checksum, so we do that here. The one
2144 2149 * exception is the gang leader: the pipeline already computed
2145 2150 * its data checksum because that stage precedes gang assembly.
2146 2151 * (Presently, nothing actually uses interior data checksums;
2147 2152 * this is just good hygiene.)
2148 2153 */
2149 2154 if (gn != pio->io_gang_leader->io_gang_tree) {
2150 2155 abd_t *buf = abd_get_offset(data, offset);
2151 2156
2152 2157 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2153 2158 buf, BP_GET_PSIZE(bp));
2154 2159
2155 2160 abd_put(buf);
2156 2161 }
2157 2162 /*
2158 2163 * If we are here to damage data for testing purposes,
2159 2164 * leave the GBH alone so that we can detect the damage.
2160 2165 */
2161 2166 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2162 2167 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2163 2168 } else {
2164 2169 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2165 2170 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2166 2171 zio_gang_issue_func_done, NULL, pio->io_priority,
2167 2172 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2168 2173 }
2169 2174
2170 2175 return (zio);
2171 2176 }
2172 2177
2173 2178 /* ARGSUSED */
2174 2179 static zio_t *
2175 2180 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2176 2181 uint64_t offset)
2177 2182 {
2178 2183 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2179 2184 ZIO_GANG_CHILD_FLAGS(pio)));
2180 2185 }
2181 2186
2182 2187 /* ARGSUSED */
2183 2188 static zio_t *
2184 2189 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2185 2190 uint64_t offset)
2186 2191 {
2187 2192 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2188 2193 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2189 2194 }
2190 2195
2191 2196 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2192 2197 NULL,
2193 2198 zio_read_gang,
2194 2199 zio_rewrite_gang,
2195 2200 zio_free_gang,
2196 2201 zio_claim_gang,
2197 2202 NULL
2198 2203 };
2199 2204
2200 2205 static void zio_gang_tree_assemble_done(zio_t *zio);
2201 2206
2202 2207 static zio_gang_node_t *
2203 2208 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2204 2209 {
2205 2210 zio_gang_node_t *gn;
2206 2211
2207 2212 ASSERT(*gnpp == NULL);
2208 2213
2209 2214 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2210 2215 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2211 2216 *gnpp = gn;
2212 2217
2213 2218 return (gn);
2214 2219 }
2215 2220
2216 2221 static void
2217 2222 zio_gang_node_free(zio_gang_node_t **gnpp)
2218 2223 {
2219 2224 zio_gang_node_t *gn = *gnpp;
2220 2225
2221 2226 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2222 2227 ASSERT(gn->gn_child[g] == NULL);
2223 2228
2224 2229 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2225 2230 kmem_free(gn, sizeof (*gn));
2226 2231 *gnpp = NULL;
2227 2232 }
2228 2233
2229 2234 static void
2230 2235 zio_gang_tree_free(zio_gang_node_t **gnpp)
2231 2236 {
2232 2237 zio_gang_node_t *gn = *gnpp;
2233 2238
2234 2239 if (gn == NULL)
2235 2240 return;
2236 2241
2237 2242 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2238 2243 zio_gang_tree_free(&gn->gn_child[g]);
2239 2244
2240 2245 zio_gang_node_free(gnpp);
2241 2246 }
2242 2247
2243 2248 static void
2244 2249 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2245 2250 {
2246 2251 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2247 2252 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2248 2253
2249 2254 ASSERT(gio->io_gang_leader == gio);
2250 2255 ASSERT(BP_IS_GANG(bp));
2251 2256
2252 2257 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2253 2258 zio_gang_tree_assemble_done, gn, gio->io_priority,
2254 2259 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2255 2260 }
2256 2261
2257 2262 static void
2258 2263 zio_gang_tree_assemble_done(zio_t *zio)
2259 2264 {
2260 2265 zio_t *gio = zio->io_gang_leader;
2261 2266 zio_gang_node_t *gn = zio->io_private;
2262 2267 blkptr_t *bp = zio->io_bp;
2263 2268
2264 2269 ASSERT(gio == zio_unique_parent(zio));
2265 2270 ASSERT(zio->io_child_count == 0);
2266 2271
2267 2272 if (zio->io_error)
2268 2273 return;
2269 2274
2270 2275 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2271 2276 if (BP_SHOULD_BYTESWAP(bp))
2272 2277 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2273 2278
2274 2279 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2275 2280 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2276 2281 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2277 2282
2278 2283 abd_put(zio->io_abd);
2279 2284
2280 2285 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2281 2286 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2282 2287 if (!BP_IS_GANG(gbp))
2283 2288 continue;
2284 2289 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2285 2290 }
2286 2291 }
2287 2292
2288 2293 static void
2289 2294 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2290 2295 uint64_t offset)
2291 2296 {
2292 2297 zio_t *gio = pio->io_gang_leader;
2293 2298 zio_t *zio;
2294 2299
2295 2300 ASSERT(BP_IS_GANG(bp) == !!gn);
2296 2301 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2297 2302 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2298 2303
2299 2304 /*
2300 2305 * If you're a gang header, your data is in gn->gn_gbh.
2301 2306 * If you're a gang member, your data is in 'data' and gn == NULL.
2302 2307 */
2303 2308 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2304 2309
2305 2310 if (gn != NULL) {
2306 2311 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2307 2312
2308 2313 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2309 2314 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2310 2315 if (BP_IS_HOLE(gbp))
2311 2316 continue;
2312 2317 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2313 2318 offset);
2314 2319 offset += BP_GET_PSIZE(gbp);
2315 2320 }
2316 2321 }
2317 2322
2318 2323 if (gn == gio->io_gang_tree)
2319 2324 ASSERT3U(gio->io_size, ==, offset);
2320 2325
2321 2326 if (zio != pio)
2322 2327 zio_nowait(zio);
2323 2328 }
2324 2329
2325 2330 static int
2326 2331 zio_gang_assemble(zio_t *zio)
2327 2332 {
2328 2333 blkptr_t *bp = zio->io_bp;
2329 2334
2330 2335 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2331 2336 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2332 2337
2333 2338 zio->io_gang_leader = zio;
2334 2339
2335 2340 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2336 2341
2337 2342 return (ZIO_PIPELINE_CONTINUE);
2338 2343 }
2339 2344
2340 2345 static int
2341 2346 zio_gang_issue(zio_t *zio)
2342 2347 {
2343 2348 blkptr_t *bp = zio->io_bp;
2344 2349
2345 2350 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2346 2351 return (ZIO_PIPELINE_STOP);
2347 2352 }
2348 2353
2349 2354 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2350 2355 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2351 2356
2352 2357 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2353 2358 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2354 2359 0);
2355 2360 else
2356 2361 zio_gang_tree_free(&zio->io_gang_tree);
2357 2362
2358 2363 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2359 2364
2360 2365 return (ZIO_PIPELINE_CONTINUE);
2361 2366 }
2362 2367
2363 2368 static void
2364 2369 zio_write_gang_member_ready(zio_t *zio)
2365 2370 {
2366 2371 zio_t *pio = zio_unique_parent(zio);
2367 2372 zio_t *gio = zio->io_gang_leader;
2368 2373 dva_t *cdva = zio->io_bp->blk_dva;
2369 2374 dva_t *pdva = pio->io_bp->blk_dva;
2370 2375 uint64_t asize;
2371 2376
2372 2377 if (BP_IS_HOLE(zio->io_bp))
2373 2378 return;
2374 2379
2375 2380 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2376 2381
2377 2382 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2378 2383 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2379 2384 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2380 2385 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2381 2386 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2382 2387
2383 2388 mutex_enter(&pio->io_lock);
2384 2389 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2385 2390 ASSERT(DVA_GET_GANG(&pdva[d]));
2386 2391 asize = DVA_GET_ASIZE(&pdva[d]);
2387 2392 asize += DVA_GET_ASIZE(&cdva[d]);
2388 2393 DVA_SET_ASIZE(&pdva[d], asize);
2389 2394 }
2390 2395 mutex_exit(&pio->io_lock);
2391 2396 }
2392 2397
2393 2398 static void
2394 2399 zio_write_gang_done(zio_t *zio)
2395 2400 {
2396 2401 /*
2397 2402 * The io_abd field will be NULL for a zio with no data. The io_flags
2398 2403 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2399 2404 * check for it here as it is cleared in zio_ready.
2400 2405 */
2401 2406 if (zio->io_abd != NULL)
2402 2407 abd_put(zio->io_abd);
2403 2408 }
2404 2409
2405 2410 static int
2406 2411 zio_write_gang_block(zio_t *pio)
2407 2412 {
2408 2413 spa_t *spa = pio->io_spa;
2409 2414 metaslab_class_t *mc = spa_normal_class(spa);
2410 2415 blkptr_t *bp = pio->io_bp;
2411 2416 zio_t *gio = pio->io_gang_leader;
2412 2417 zio_t *zio;
2413 2418 zio_gang_node_t *gn, **gnpp;
2414 2419 zio_gbh_phys_t *gbh;
2415 2420 abd_t *gbh_abd;
2416 2421 uint64_t txg = pio->io_txg;
2417 2422 uint64_t resid = pio->io_size;
2418 2423 uint64_t lsize;
2419 2424 int copies = gio->io_prop.zp_copies;
2420 2425 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2421 2426 zio_prop_t zp;
2422 2427 int error;
2423 2428 boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
2424 2429
2425 2430 /*
2426 2431 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2427 2432 * have a third copy.
2428 2433 */
2429 2434 if (gio->io_prop.zp_encrypt && gbh_copies >= SPA_DVAS_PER_BP)
2430 2435 gbh_copies = SPA_DVAS_PER_BP - 1;
2431 2436
2432 2437 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2433 2438 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2434 2439 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2435 2440 ASSERT(has_data);
2436 2441
2437 2442 flags |= METASLAB_ASYNC_ALLOC;
2438 2443 VERIFY(zfs_refcount_held(&mc->mc_alloc_slots[pio->io_allocator],
2439 2444 pio));
2440 2445
2441 2446 /*
2442 2447 * The logical zio has already placed a reservation for
2443 2448 * 'copies' allocation slots but gang blocks may require
2444 2449 * additional copies. These additional copies
2445 2450 * (i.e. gbh_copies - copies) are guaranteed to succeed
2446 2451 * since metaslab_class_throttle_reserve() always allows
2447 2452 * additional reservations for gang blocks.
2448 2453 */
2449 2454 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2450 2455 pio->io_allocator, pio, flags));
2451 2456 }
2452 2457
2453 2458 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2454 2459 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2455 2460 &pio->io_alloc_list, pio, pio->io_allocator);
2456 2461 if (error) {
2457 2462 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2458 2463 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2459 2464 ASSERT(has_data);
2460 2465
2461 2466 /*
2462 2467 * If we failed to allocate the gang block header then
2463 2468 * we remove any additional allocation reservations that
2464 2469 * we placed here. The original reservation will
2465 2470 * be removed when the logical I/O goes to the ready
2466 2471 * stage.
2467 2472 */
2468 2473 metaslab_class_throttle_unreserve(mc,
2469 2474 gbh_copies - copies, pio->io_allocator, pio);
2470 2475 }
2471 2476 pio->io_error = error;
2472 2477 return (ZIO_PIPELINE_CONTINUE);
2473 2478 }
2474 2479
2475 2480 if (pio == gio) {
2476 2481 gnpp = &gio->io_gang_tree;
2477 2482 } else {
2478 2483 gnpp = pio->io_private;
2479 2484 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2480 2485 }
2481 2486
2482 2487 gn = zio_gang_node_alloc(gnpp);
2483 2488 gbh = gn->gn_gbh;
2484 2489 bzero(gbh, SPA_GANGBLOCKSIZE);
2485 2490 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2486 2491
2487 2492 /*
2488 2493 * Create the gang header.
2489 2494 */
2490 2495 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2491 2496 zio_write_gang_done, NULL, pio->io_priority,
2492 2497 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2493 2498
2494 2499 /*
2495 2500 * Create and nowait the gang children.
2496 2501 */
2497 2502 for (int g = 0; resid != 0; resid -= lsize, g++) {
2498 2503 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2499 2504 SPA_MINBLOCKSIZE);
2500 2505 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2501 2506
2502 2507 zp.zp_checksum = gio->io_prop.zp_checksum;
2503 2508 zp.zp_compress = ZIO_COMPRESS_OFF;
2504 2509 zp.zp_type = DMU_OT_NONE;
2505 2510 zp.zp_level = 0;
2506 2511 zp.zp_copies = gio->io_prop.zp_copies;
2507 2512 zp.zp_dedup = B_FALSE;
2508 2513 zp.zp_dedup_verify = B_FALSE;
2509 2514 zp.zp_nopwrite = B_FALSE;
2510 2515 zp.zp_encrypt = gio->io_prop.zp_encrypt;
2511 2516 zp.zp_byteorder = gio->io_prop.zp_byteorder;
2512 2517 bzero(zp.zp_salt, ZIO_DATA_SALT_LEN);
2513 2518 bzero(zp.zp_iv, ZIO_DATA_IV_LEN);
2514 2519 bzero(zp.zp_mac, ZIO_DATA_MAC_LEN);
2515 2520
2516 2521 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2517 2522 has_data ? abd_get_offset(pio->io_abd, pio->io_size -
2518 2523 resid) : NULL, lsize, lsize, &zp,
2519 2524 zio_write_gang_member_ready, NULL, NULL,
2520 2525 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2521 2526 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2522 2527
2523 2528 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2524 2529 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2525 2530 ASSERT(has_data);
2526 2531
2527 2532 /*
2528 2533 * Gang children won't throttle but we should
2529 2534 * account for their work, so reserve an allocation
2530 2535 * slot for them here.
2531 2536 */
2532 2537 VERIFY(metaslab_class_throttle_reserve(mc,
2533 2538 zp.zp_copies, cio->io_allocator, cio, flags));
2534 2539 }
2535 2540 zio_nowait(cio);
2536 2541 }
2537 2542
2538 2543 /*
2539 2544 * Set pio's pipeline to just wait for zio to finish.
2540 2545 */
2541 2546 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2542 2547
2543 2548 zio_nowait(zio);
2544 2549
2545 2550 return (ZIO_PIPELINE_CONTINUE);
2546 2551 }
2547 2552
2548 2553 /*
2549 2554 * The zio_nop_write stage in the pipeline determines if allocating a
2550 2555 * new bp is necessary. The nopwrite feature can handle writes in
2551 2556 * either syncing or open context (i.e. zil writes) and as a result is
2552 2557 * mutually exclusive with dedup.
2553 2558 *
2554 2559 * By leveraging a cryptographically secure checksum, such as SHA256, we
2555 2560 * can compare the checksums of the new data and the old to determine if
2556 2561 * allocating a new block is required. Note that our requirements for
2557 2562 * cryptographic strength are fairly weak: there can't be any accidental
2558 2563 * hash collisions, but we don't need to be secure against intentional
2559 2564 * (malicious) collisions. To trigger a nopwrite, you have to be able
2560 2565 * to write the file to begin with, and triggering an incorrect (hash
2561 2566 * collision) nopwrite is no worse than simply writing to the file.
2562 2567 * That said, there are no known attacks against the checksum algorithms
2563 2568 * used for nopwrite, assuming that the salt and the checksums
2564 2569 * themselves remain secret.
2565 2570 */
2566 2571 static int
2567 2572 zio_nop_write(zio_t *zio)
2568 2573 {
2569 2574 blkptr_t *bp = zio->io_bp;
2570 2575 blkptr_t *bp_orig = &zio->io_bp_orig;
2571 2576 zio_prop_t *zp = &zio->io_prop;
2572 2577
2573 2578 ASSERT(BP_GET_LEVEL(bp) == 0);
2574 2579 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2575 2580 ASSERT(zp->zp_nopwrite);
2576 2581 ASSERT(!zp->zp_dedup);
2577 2582 ASSERT(zio->io_bp_override == NULL);
2578 2583 ASSERT(IO_IS_ALLOCATING(zio));
2579 2584
2580 2585 /*
2581 2586 * Check to see if the original bp and the new bp have matching
2582 2587 * characteristics (i.e. same checksum, compression algorithms, etc).
2583 2588 * If they don't then just continue with the pipeline which will
2584 2589 * allocate a new bp.
2585 2590 */
2586 2591 if (BP_IS_HOLE(bp_orig) ||
2587 2592 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2588 2593 ZCHECKSUM_FLAG_NOPWRITE) ||
2589 2594 BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) ||
2590 2595 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2591 2596 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2592 2597 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2593 2598 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2594 2599 return (ZIO_PIPELINE_CONTINUE);
2595 2600
2596 2601 /*
2597 2602 * If the checksums match then reset the pipeline so that we
2598 2603 * avoid allocating a new bp and issuing any I/O.
2599 2604 */
2600 2605 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2601 2606 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2602 2607 ZCHECKSUM_FLAG_NOPWRITE);
2603 2608 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2604 2609 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2605 2610 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2606 2611 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2607 2612 sizeof (uint64_t)) == 0);
2608 2613
2609 2614 *bp = *bp_orig;
2610 2615 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2611 2616 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2612 2617 }
2613 2618
2614 2619 return (ZIO_PIPELINE_CONTINUE);
2615 2620 }
2616 2621
2617 2622 /*
2618 2623 * ==========================================================================
2619 2624 * Dedup
2620 2625 * ==========================================================================
2621 2626 */
2622 2627 static void
2623 2628 zio_ddt_child_read_done(zio_t *zio)
2624 2629 {
2625 2630 blkptr_t *bp = zio->io_bp;
2626 2631 ddt_entry_t *dde = zio->io_private;
2627 2632 ddt_phys_t *ddp;
2628 2633 zio_t *pio = zio_unique_parent(zio);
2629 2634
2630 2635 mutex_enter(&pio->io_lock);
2631 2636 ddp = ddt_phys_select(dde, bp);
2632 2637 if (zio->io_error == 0)
2633 2638 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2634 2639
2635 2640 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2636 2641 dde->dde_repair_abd = zio->io_abd;
2637 2642 else
2638 2643 abd_free(zio->io_abd);
2639 2644 mutex_exit(&pio->io_lock);
2640 2645 }
2641 2646
2642 2647 static int
2643 2648 zio_ddt_read_start(zio_t *zio)
2644 2649 {
2645 2650 blkptr_t *bp = zio->io_bp;
2646 2651
2647 2652 ASSERT(BP_GET_DEDUP(bp));
2648 2653 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2649 2654 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2650 2655
2651 2656 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2652 2657 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2653 2658 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2654 2659 ddt_phys_t *ddp = dde->dde_phys;
2655 2660 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2656 2661 blkptr_t blk;
2657 2662
2658 2663 ASSERT(zio->io_vsd == NULL);
2659 2664 zio->io_vsd = dde;
2660 2665
2661 2666 if (ddp_self == NULL)
2662 2667 return (ZIO_PIPELINE_CONTINUE);
2663 2668
2664 2669 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2665 2670 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2666 2671 continue;
2667 2672 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2668 2673 &blk);
2669 2674 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2670 2675 abd_alloc_for_io(zio->io_size, B_TRUE),
2671 2676 zio->io_size, zio_ddt_child_read_done, dde,
2672 2677 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2673 2678 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2674 2679 }
2675 2680 return (ZIO_PIPELINE_CONTINUE);
2676 2681 }
2677 2682
2678 2683 zio_nowait(zio_read(zio, zio->io_spa, bp,
2679 2684 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2680 2685 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2681 2686
2682 2687 return (ZIO_PIPELINE_CONTINUE);
2683 2688 }
2684 2689
2685 2690 static int
2686 2691 zio_ddt_read_done(zio_t *zio)
2687 2692 {
2688 2693 blkptr_t *bp = zio->io_bp;
2689 2694
2690 2695 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2691 2696 return (ZIO_PIPELINE_STOP);
2692 2697 }
2693 2698
2694 2699 ASSERT(BP_GET_DEDUP(bp));
2695 2700 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2696 2701 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2697 2702
2698 2703 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2699 2704 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2700 2705 ddt_entry_t *dde = zio->io_vsd;
2701 2706 if (ddt == NULL) {
2702 2707 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2703 2708 return (ZIO_PIPELINE_CONTINUE);
2704 2709 }
2705 2710 if (dde == NULL) {
2706 2711 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2707 2712 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2708 2713 return (ZIO_PIPELINE_STOP);
2709 2714 }
2710 2715 if (dde->dde_repair_abd != NULL) {
2711 2716 abd_copy(zio->io_abd, dde->dde_repair_abd,
2712 2717 zio->io_size);
2713 2718 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2714 2719 }
2715 2720 ddt_repair_done(ddt, dde);
2716 2721 zio->io_vsd = NULL;
2717 2722 }
2718 2723
2719 2724 ASSERT(zio->io_vsd == NULL);
2720 2725
2721 2726 return (ZIO_PIPELINE_CONTINUE);
2722 2727 }
2723 2728
2724 2729 static boolean_t
2725 2730 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2726 2731 {
2727 2732 spa_t *spa = zio->io_spa;
2728 2733 boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW);
2729 2734
2730 2735 /* We should never get a raw, override zio */
2731 2736 ASSERT(!(zio->io_bp_override && do_raw));
2732 2737
2733 2738 /*
2734 2739 * Note: we compare the original data, not the transformed data,
2735 2740 * because when zio->io_bp is an override bp, we will not have
2736 2741 * pushed the I/O transforms. That's an important optimization
2737 2742 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2738 2743 * However, we should never get a raw, override zio so in these
2739 2744 * cases we can compare the io_data directly. This is useful because
2740 2745 * it allows us to do dedup verification even if we don't have access
2741 2746 * to the original data (for instance, if the encryption keys aren't
2742 2747 * loaded).
2743 2748 */
2744 2749
2745 2750 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2746 2751 zio_t *lio = dde->dde_lead_zio[p];
2747 2752
2748 2753 if (lio != NULL && do_raw) {
2749 2754 return (lio->io_size != zio->io_size ||
2750 2755 abd_cmp(zio->io_abd, lio->io_abd,
2751 2756 zio->io_size) != 0);
2752 2757 } else if (lio != NULL) {
2753 2758 return (lio->io_orig_size != zio->io_orig_size ||
2754 2759 abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2755 2760 zio->io_orig_size) != 0);
2756 2761 }
2757 2762 }
2758 2763
2759 2764 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2760 2765 ddt_phys_t *ddp = &dde->dde_phys[p];
2761 2766
2762 2767 if (ddp->ddp_phys_birth != 0 && do_raw) {
2763 2768 blkptr_t blk = *zio->io_bp;
2764 2769 uint64_t psize;
2765 2770 abd_t *tmpabd;
2766 2771 int error;
2767 2772
2768 2773 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2769 2774 psize = BP_GET_PSIZE(&blk);
2770 2775
2771 2776 if (psize != zio->io_size)
2772 2777 return (B_TRUE);
2773 2778
2774 2779 ddt_exit(ddt);
2775 2780
2776 2781 tmpabd = abd_alloc_for_io(psize, B_TRUE);
2777 2782
2778 2783 error = zio_wait(zio_read(NULL, spa, &blk, tmpabd,
2779 2784 psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
2780 2785 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2781 2786 ZIO_FLAG_RAW, &zio->io_bookmark));
2782 2787
2783 2788 if (error == 0) {
2784 2789 if (abd_cmp(tmpabd, zio->io_abd, psize) != 0)
2785 2790 error = SET_ERROR(ENOENT);
2786 2791 }
2787 2792
2788 2793 abd_free(tmpabd);
2789 2794 ddt_enter(ddt);
2790 2795 return (error != 0);
2791 2796 } else if (ddp->ddp_phys_birth != 0) {
2792 2797 arc_buf_t *abuf = NULL;
2793 2798 arc_flags_t aflags = ARC_FLAG_WAIT;
2794 2799 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2795 2800 blkptr_t blk = *zio->io_bp;
2796 2801 int error;
2797 2802
2798 2803 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2799 2804
2800 2805 if (BP_GET_LSIZE(&blk) != zio->io_orig_size)
2801 2806 return (B_TRUE);
2802 2807
2803 2808 ddt_exit(ddt);
2804 2809
2805 2810 /*
2806 2811 * Intuitively, it would make more sense to compare
2807 2812 * io_abd than io_orig_abd in the raw case since you
2808 2813 * don't want to look at any transformations that have
2809 2814 * happened to the data. However, for raw I/Os the
2810 2815 * data will actually be the same in io_abd and
2811 2816 * io_orig_abd, so all we have to do is issue this as
2812 2817 * a raw ARC read.
2813 2818 */
2814 2819 if (do_raw) {
2815 2820 zio_flags |= ZIO_FLAG_RAW;
2816 2821 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2817 2822 ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2818 2823 zio->io_size));
2819 2824 ASSERT3P(zio->io_transform_stack, ==, NULL);
2820 2825 }
2821 2826
2822 2827 error = arc_read(NULL, spa, &blk,
2823 2828 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2824 2829 zio_flags, &aflags, &zio->io_bookmark);
2825 2830
2826 2831 if (error == 0) {
2827 2832 if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2828 2833 zio->io_orig_size) != 0)
2829 2834 error = SET_ERROR(ENOENT);
2830 2835 arc_buf_destroy(abuf, &abuf);
2831 2836 }
2832 2837
2833 2838 ddt_enter(ddt);
2834 2839 return (error != 0);
2835 2840 }
2836 2841 }
2837 2842
2838 2843 return (B_FALSE);
2839 2844 }
2840 2845
2841 2846 static void
2842 2847 zio_ddt_child_write_ready(zio_t *zio)
2843 2848 {
2844 2849 int p = zio->io_prop.zp_copies;
2845 2850 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2846 2851 ddt_entry_t *dde = zio->io_private;
2847 2852 ddt_phys_t *ddp = &dde->dde_phys[p];
2848 2853 zio_t *pio;
2849 2854
2850 2855 if (zio->io_error)
2851 2856 return;
2852 2857
2853 2858 ddt_enter(ddt);
2854 2859
2855 2860 ASSERT(dde->dde_lead_zio[p] == zio);
2856 2861
2857 2862 ddt_phys_fill(ddp, zio->io_bp);
2858 2863
2859 2864 zio_link_t *zl = NULL;
2860 2865 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2861 2866 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2862 2867
2863 2868 ddt_exit(ddt);
2864 2869 }
2865 2870
2866 2871 static void
2867 2872 zio_ddt_child_write_done(zio_t *zio)
2868 2873 {
2869 2874 int p = zio->io_prop.zp_copies;
2870 2875 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2871 2876 ddt_entry_t *dde = zio->io_private;
2872 2877 ddt_phys_t *ddp = &dde->dde_phys[p];
2873 2878
2874 2879 ddt_enter(ddt);
2875 2880
2876 2881 ASSERT(ddp->ddp_refcnt == 0);
2877 2882 ASSERT(dde->dde_lead_zio[p] == zio);
2878 2883 dde->dde_lead_zio[p] = NULL;
2879 2884
2880 2885 if (zio->io_error == 0) {
2881 2886 zio_link_t *zl = NULL;
2882 2887 while (zio_walk_parents(zio, &zl) != NULL)
2883 2888 ddt_phys_addref(ddp);
2884 2889 } else {
2885 2890 ddt_phys_clear(ddp);
2886 2891 }
2887 2892
2888 2893 ddt_exit(ddt);
2889 2894 }
2890 2895
2891 2896 static void
2892 2897 zio_ddt_ditto_write_done(zio_t *zio)
2893 2898 {
2894 2899 int p = DDT_PHYS_DITTO;
2895 2900 zio_prop_t *zp = &zio->io_prop;
2896 2901 blkptr_t *bp = zio->io_bp;
2897 2902 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2898 2903 ddt_entry_t *dde = zio->io_private;
2899 2904 ddt_phys_t *ddp = &dde->dde_phys[p];
2900 2905 ddt_key_t *ddk = &dde->dde_key;
2901 2906
2902 2907 ddt_enter(ddt);
2903 2908
2904 2909 ASSERT(ddp->ddp_refcnt == 0);
2905 2910 ASSERT(dde->dde_lead_zio[p] == zio);
2906 2911 dde->dde_lead_zio[p] = NULL;
2907 2912
2908 2913 if (zio->io_error == 0) {
2909 2914 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2910 2915 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2911 2916 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2912 2917 if (ddp->ddp_phys_birth != 0)
2913 2918 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2914 2919 ddt_phys_fill(ddp, bp);
2915 2920 }
2916 2921
2917 2922 ddt_exit(ddt);
2918 2923 }
2919 2924
2920 2925 static int
2921 2926 zio_ddt_write(zio_t *zio)
2922 2927 {
2923 2928 spa_t *spa = zio->io_spa;
2924 2929 blkptr_t *bp = zio->io_bp;
2925 2930 uint64_t txg = zio->io_txg;
2926 2931 zio_prop_t *zp = &zio->io_prop;
2927 2932 int p = zp->zp_copies;
2928 2933 int ditto_copies;
2929 2934 zio_t *cio = NULL;
2930 2935 zio_t *dio = NULL;
2931 2936 ddt_t *ddt = ddt_select(spa, bp);
2932 2937 ddt_entry_t *dde;
2933 2938 ddt_phys_t *ddp;
2934 2939
2935 2940 ASSERT(BP_GET_DEDUP(bp));
2936 2941 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2937 2942 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2938 2943 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2939 2944
2940 2945 ddt_enter(ddt);
2941 2946 dde = ddt_lookup(ddt, bp, B_TRUE);
2942 2947 ddp = &dde->dde_phys[p];
2943 2948
2944 2949 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2945 2950 /*
2946 2951 * If we're using a weak checksum, upgrade to a strong checksum
2947 2952 * and try again. If we're already using a strong checksum,
2948 2953 * we can't resolve it, so just convert to an ordinary write.
2949 2954 * (And automatically e-mail a paper to Nature?)
2950 2955 */
2951 2956 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2952 2957 ZCHECKSUM_FLAG_DEDUP)) {
2953 2958 zp->zp_checksum = spa_dedup_checksum(spa);
2954 2959 zio_pop_transforms(zio);
2955 2960 zio->io_stage = ZIO_STAGE_OPEN;
2956 2961 BP_ZERO(bp);
2957 2962 } else {
2958 2963 zp->zp_dedup = B_FALSE;
2959 2964 BP_SET_DEDUP(bp, B_FALSE);
2960 2965 }
2961 2966 ASSERT(!BP_GET_DEDUP(bp));
2962 2967 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2963 2968 ddt_exit(ddt);
2964 2969 return (ZIO_PIPELINE_CONTINUE);
2965 2970 }
2966 2971
2967 2972 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2968 2973 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2969 2974
2970 2975 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2971 2976 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2972 2977 zio_prop_t czp = *zp;
2973 2978
2974 2979 czp.zp_copies = ditto_copies;
2975 2980
2976 2981 /*
2977 2982 * If we arrived here with an override bp, we won't have run
2978 2983 * the transform stack, so we won't have the data we need to
2979 2984 * generate a child i/o. So, toss the override bp and restart.
2980 2985 * This is safe, because using the override bp is just an
2981 2986 * optimization; and it's rare, so the cost doesn't matter.
2982 2987 */
2983 2988 if (zio->io_bp_override) {
2984 2989 zio_pop_transforms(zio);
2985 2990 zio->io_stage = ZIO_STAGE_OPEN;
2986 2991 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2987 2992 zio->io_bp_override = NULL;
2988 2993 BP_ZERO(bp);
2989 2994 ddt_exit(ddt);
2990 2995 return (ZIO_PIPELINE_CONTINUE);
2991 2996 }
2992 2997
2993 2998 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2994 2999 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2995 3000 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2996 3001 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2997 3002
2998 3003 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2999 3004 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
3000 3005 }
3001 3006
3002 3007 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
3003 3008 if (ddp->ddp_phys_birth != 0)
3004 3009 ddt_bp_fill(ddp, bp, txg);
3005 3010 if (dde->dde_lead_zio[p] != NULL)
3006 3011 zio_add_child(zio, dde->dde_lead_zio[p]);
3007 3012 else
3008 3013 ddt_phys_addref(ddp);
3009 3014 } else if (zio->io_bp_override) {
3010 3015 ASSERT(bp->blk_birth == txg);
3011 3016 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
3012 3017 ddt_phys_fill(ddp, bp);
3013 3018 ddt_phys_addref(ddp);
3014 3019 } else {
3015 3020 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3016 3021 zio->io_orig_size, zio->io_orig_size, zp,
3017 3022 zio_ddt_child_write_ready, NULL, NULL,
3018 3023 zio_ddt_child_write_done, dde, zio->io_priority,
3019 3024 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3020 3025
3021 3026 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
3022 3027 dde->dde_lead_zio[p] = cio;
3023 3028 }
3024 3029
3025 3030 ddt_exit(ddt);
3026 3031
3027 3032 if (cio)
3028 3033 zio_nowait(cio);
3029 3034 if (dio)
3030 3035 zio_nowait(dio);
3031 3036
3032 3037 return (ZIO_PIPELINE_CONTINUE);
3033 3038 }
3034 3039
3035 3040 ddt_entry_t *freedde; /* for debugging */
3036 3041
3037 3042 static int
3038 3043 zio_ddt_free(zio_t *zio)
3039 3044 {
3040 3045 spa_t *spa = zio->io_spa;
3041 3046 blkptr_t *bp = zio->io_bp;
3042 3047 ddt_t *ddt = ddt_select(spa, bp);
3043 3048 ddt_entry_t *dde;
3044 3049 ddt_phys_t *ddp;
3045 3050
3046 3051 ASSERT(BP_GET_DEDUP(bp));
3047 3052 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3048 3053
3049 3054 ddt_enter(ddt);
3050 3055 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
3051 3056 ddp = ddt_phys_select(dde, bp);
3052 3057 ddt_phys_decref(ddp);
3053 3058 ddt_exit(ddt);
3054 3059
3055 3060 return (ZIO_PIPELINE_CONTINUE);
3056 3061 }
3057 3062
3058 3063 /*
3059 3064 * ==========================================================================
3060 3065 * Allocate and free blocks
3061 3066 * ==========================================================================
3062 3067 */
3063 3068
3064 3069 static zio_t *
3065 3070 zio_io_to_allocate(spa_t *spa, int allocator)
3066 3071 {
3067 3072 zio_t *zio;
3068 3073
3069 3074 ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator]));
3070 3075
3071 3076 zio = avl_first(&spa->spa_alloc_trees[allocator]);
3072 3077 if (zio == NULL)
3073 3078 return (NULL);
3074 3079
3075 3080 ASSERT(IO_IS_ALLOCATING(zio));
3076 3081
3077 3082 /*
3078 3083 * Try to place a reservation for this zio. If we're unable to
3079 3084 * reserve then we throttle.
3080 3085 */
3081 3086 ASSERT3U(zio->io_allocator, ==, allocator);
3082 3087 if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
3083 3088 zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) {
3084 3089 return (NULL);
3085 3090 }
3086 3091
3087 3092 avl_remove(&spa->spa_alloc_trees[allocator], zio);
3088 3093 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
3089 3094
3090 3095 return (zio);
3091 3096 }
3092 3097
3093 3098 static int
3094 3099 zio_dva_throttle(zio_t *zio)
3095 3100 {
3096 3101 spa_t *spa = zio->io_spa;
3097 3102 zio_t *nio;
3098 3103 metaslab_class_t *mc;
3099 3104
3100 3105 /* locate an appropriate allocation class */
3101 3106 mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type,
3102 3107 zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk);
3103 3108
3104 3109 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
3105 3110 !mc->mc_alloc_throttle_enabled ||
3106 3111 zio->io_child_type == ZIO_CHILD_GANG ||
3107 3112 zio->io_flags & ZIO_FLAG_NODATA) {
3108 3113 return (ZIO_PIPELINE_CONTINUE);
3109 3114 }
3110 3115
3111 3116 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3112 3117
3113 3118 ASSERT3U(zio->io_queued_timestamp, >, 0);
3114 3119 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3115 3120
3116 3121 zbookmark_phys_t *bm = &zio->io_bookmark;
3117 3122 /*
3118 3123 * We want to try to use as many allocators as possible to help improve
3119 3124 * performance, but we also want logically adjacent IOs to be physically
3120 3125 * adjacent to improve sequential read performance. We chunk each object
3121 3126 * into 2^20 block regions, and then hash based on the objset, object,
3122 3127 * level, and region to accomplish both of these goals.
3123 3128 */
3124 3129 zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object,
3125 3130 bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
3126 3131 mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]);
3127 3132 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3128 3133 zio->io_metaslab_class = mc;
3129 3134 avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio);
3130 3135 nio = zio_io_to_allocate(spa, zio->io_allocator);
3131 3136 mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]);
3132 3137
3133 3138 if (nio == zio)
3134 3139 return (ZIO_PIPELINE_CONTINUE);
3135 3140
3136 3141 if (nio != NULL) {
3137 3142 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3138 3143 /*
3139 3144 * We are passing control to a new zio so make sure that
3140 3145 * it is processed by a different thread. We do this to
3141 3146 * avoid stack overflows that can occur when parents are
3142 3147 * throttled and children are making progress. We allow
3143 3148 * it to go to the head of the taskq since it's already
3144 3149 * been waiting.
3145 3150 */
3146 3151 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
3147 3152 }
3148 3153 return (ZIO_PIPELINE_STOP);
3149 3154 }
3150 3155
3151 3156 static void
3152 3157 zio_allocate_dispatch(spa_t *spa, int allocator)
3153 3158 {
3154 3159 zio_t *zio;
3155 3160
3156 3161 mutex_enter(&spa->spa_alloc_locks[allocator]);
3157 3162 zio = zio_io_to_allocate(spa, allocator);
3158 3163 mutex_exit(&spa->spa_alloc_locks[allocator]);
3159 3164 if (zio == NULL)
3160 3165 return;
3161 3166
3162 3167 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3163 3168 ASSERT0(zio->io_error);
3164 3169 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3165 3170 }
3166 3171
3167 3172 static int
3168 3173 zio_dva_allocate(zio_t *zio)
3169 3174 {
3170 3175 spa_t *spa = zio->io_spa;
3171 3176 metaslab_class_t *mc;
3172 3177 blkptr_t *bp = zio->io_bp;
3173 3178 int error;
3174 3179 int flags = 0;
3175 3180
3176 3181 if (zio->io_gang_leader == NULL) {
3177 3182 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3178 3183 zio->io_gang_leader = zio;
3179 3184 }
3180 3185
3181 3186 ASSERT(BP_IS_HOLE(bp));
3182 3187 ASSERT0(BP_GET_NDVAS(bp));
3183 3188 ASSERT3U(zio->io_prop.zp_copies, >, 0);
3184 3189 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3185 3190 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3186 3191
3187 3192 if (zio->io_flags & ZIO_FLAG_NODATA)
3188 3193 flags |= METASLAB_DONT_THROTTLE;
3189 3194 if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
3190 3195 flags |= METASLAB_GANG_CHILD;
3191 3196 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
3192 3197 flags |= METASLAB_ASYNC_ALLOC;
3193 3198
3194 3199 /*
3195 3200 * if not already chosen, locate an appropriate allocation class
3196 3201 */
3197 3202 mc = zio->io_metaslab_class;
3198 3203 if (mc == NULL) {
3199 3204 mc = spa_preferred_class(spa, zio->io_size,
3200 3205 zio->io_prop.zp_type, zio->io_prop.zp_level,
3201 3206 zio->io_prop.zp_zpl_smallblk);
3202 3207 zio->io_metaslab_class = mc;
3203 3208 }
3204 3209
3205 3210 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3206 3211 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3207 3212 &zio->io_alloc_list, zio, zio->io_allocator);
3208 3213
3209 3214 /*
3210 3215 * Fallback to normal class when an alloc class is full
3211 3216 */
3212 3217 if (error == ENOSPC && mc != spa_normal_class(spa)) {
3213 3218 /*
3214 3219 * If throttling, transfer reservation over to normal class.
3215 3220 * The io_allocator slot can remain the same even though we
3216 3221 * are switching classes.
3217 3222 */
3218 3223 if (mc->mc_alloc_throttle_enabled &&
3219 3224 (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) {
3220 3225 metaslab_class_throttle_unreserve(mc,
3221 3226 zio->io_prop.zp_copies, zio->io_allocator, zio);
3222 3227 zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING;
3223 3228
3224 3229 mc = spa_normal_class(spa);
3225 3230 VERIFY(metaslab_class_throttle_reserve(mc,
3226 3231 zio->io_prop.zp_copies, zio->io_allocator, zio,
3227 3232 flags | METASLAB_MUST_RESERVE));
3228 3233 } else {
3229 3234 mc = spa_normal_class(spa);
3230 3235 }
3231 3236 zio->io_metaslab_class = mc;
3232 3237
3233 3238 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3234 3239 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3235 3240 &zio->io_alloc_list, zio, zio->io_allocator);
3236 3241 }
3237 3242
3238 3243 if (error != 0) {
3239 3244 zfs_dbgmsg("%s: metaslab allocation failure: zio %p, "
3240 3245 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3241 3246 error);
3242 3247 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
3243 3248 return (zio_write_gang_block(zio));
3244 3249 zio->io_error = error;
3245 3250 }
3246 3251
3247 3252 return (ZIO_PIPELINE_CONTINUE);
3248 3253 }
3249 3254
3250 3255 static int
3251 3256 zio_dva_free(zio_t *zio)
3252 3257 {
3253 3258 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3254 3259
3255 3260 return (ZIO_PIPELINE_CONTINUE);
3256 3261 }
3257 3262
3258 3263 static int
3259 3264 zio_dva_claim(zio_t *zio)
3260 3265 {
3261 3266 int error;
3262 3267
3263 3268 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3264 3269 if (error)
3265 3270 zio->io_error = error;
3266 3271
3267 3272 return (ZIO_PIPELINE_CONTINUE);
3268 3273 }
3269 3274
3270 3275 /*
3271 3276 * Undo an allocation. This is used by zio_done() when an I/O fails
3272 3277 * and we want to give back the block we just allocated.
3273 3278 * This handles both normal blocks and gang blocks.
3274 3279 */
3275 3280 static void
3276 3281 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3277 3282 {
3278 3283 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3279 3284 ASSERT(zio->io_bp_override == NULL);
3280 3285
3281 3286 if (!BP_IS_HOLE(bp))
3282 3287 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3283 3288
3284 3289 if (gn != NULL) {
3285 3290 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3286 3291 zio_dva_unallocate(zio, gn->gn_child[g],
3287 3292 &gn->gn_gbh->zg_blkptr[g]);
3288 3293 }
3289 3294 }
3290 3295 }
3291 3296
3292 3297 /*
3293 3298 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3294 3299 */
3295 3300 int
3296 3301 zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
3297 3302 blkptr_t *old_bp, uint64_t size, boolean_t *slog)
3298 3303 {
3299 3304 int error = 1;
3300 3305 zio_alloc_list_t io_alloc_list;
3301 3306
3302 3307 ASSERT(txg > spa_syncing_txg(spa));
3303 3308
3304 3309 metaslab_trace_init(&io_alloc_list);
3305 3310
3306 3311 /*
3307 3312 * Block pointer fields are useful to metaslabs for stats and debugging.
3308 3313 * Fill in the obvious ones before calling into metaslab_alloc().
3309 3314 */
3310 3315 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3311 3316 BP_SET_PSIZE(new_bp, size);
3312 3317 BP_SET_LEVEL(new_bp, 0);
3313 3318
3314 3319 /*
3315 3320 * When allocating a zil block, we don't have information about
3316 3321 * the final destination of the block except the objset it's part
3317 3322 * of, so we just hash the objset ID to pick the allocator to get
3318 3323 * some parallelism.
3319 3324 */
3320 3325 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3321 3326 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL,
3322 3327 cityhash4(0, 0, 0,
3323 3328 os->os_dsl_dataset->ds_object) % spa->spa_alloc_count);
3324 3329 if (error == 0) {
3325 3330 *slog = TRUE;
3326 3331 } else {
3327 3332 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3328 3333 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3329 3334 &io_alloc_list, NULL, cityhash4(0, 0, 0,
3330 3335 os->os_dsl_dataset->ds_object) % spa->spa_alloc_count);
3331 3336 if (error == 0)
3332 3337 *slog = FALSE;
3333 3338 }
3334 3339 metaslab_trace_fini(&io_alloc_list);
3335 3340
3336 3341 if (error == 0) {
3337 3342 BP_SET_LSIZE(new_bp, size);
3338 3343 BP_SET_PSIZE(new_bp, size);
3339 3344 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3340 3345 BP_SET_CHECKSUM(new_bp,
3341 3346 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3342 3347 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3343 3348 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3344 3349 BP_SET_LEVEL(new_bp, 0);
3345 3350 BP_SET_DEDUP(new_bp, 0);
3346 3351 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3347 3352
3348 3353 /*
3349 3354 * encrypted blocks will require an IV and salt. We generate
3350 3355 * these now since we will not be rewriting the bp at
3351 3356 * rewrite time.
3352 3357 */
3353 3358 if (os->os_encrypted) {
3354 3359 uint8_t iv[ZIO_DATA_IV_LEN];
3355 3360 uint8_t salt[ZIO_DATA_SALT_LEN];
3356 3361
3357 3362 BP_SET_CRYPT(new_bp, B_TRUE);
3358 3363 VERIFY0(spa_crypt_get_salt(spa,
3359 3364 dmu_objset_id(os), salt));
3360 3365 VERIFY0(zio_crypt_generate_iv(iv));
3361 3366
3362 3367 zio_crypt_encode_params_bp(new_bp, salt, iv);
3363 3368 }
3364 3369 } else {
3365 3370 zfs_dbgmsg("%s: zil block allocation failure: "
3366 3371 "size %llu, error %d", spa_name(spa), size, error);
3367 3372 }
3368 3373
3369 3374 return (error);
3370 3375 }
3371 3376
3372 3377 /*
3373 3378 * ==========================================================================
3374 3379 * Read and write to physical devices
3375 3380 * ==========================================================================
3376 3381 */
3377 3382
3378 3383 /*
3379 3384 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3380 3385 * stops after this stage and will resume upon I/O completion.
3381 3386 * However, there are instances where the vdev layer may need to
3382 3387 * continue the pipeline when an I/O was not issued. Since the I/O
3383 3388 * that was sent to the vdev layer might be different than the one
3384 3389 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3385 3390 * force the underlying vdev layers to call either zio_execute() or
3386 3391 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3387 3392 */
3388 3393 static int
3389 3394 zio_vdev_io_start(zio_t *zio)
3390 3395 {
3391 3396 vdev_t *vd = zio->io_vd;
3392 3397 uint64_t align;
3393 3398 spa_t *spa = zio->io_spa;
3394 3399
3395 3400 zio->io_delay = 0;
3396 3401
3397 3402 ASSERT(zio->io_error == 0);
3398 3403 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3399 3404
3400 3405 if (vd == NULL) {
3401 3406 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3402 3407 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3403 3408
3404 3409 /*
3405 3410 * The mirror_ops handle multiple DVAs in a single BP.
3406 3411 */
3407 3412 vdev_mirror_ops.vdev_op_io_start(zio);
3408 3413 return (ZIO_PIPELINE_STOP);
3409 3414 }
3410 3415
3411 3416 ASSERT3P(zio->io_logical, !=, zio);
3412 3417 if (zio->io_type == ZIO_TYPE_WRITE) {
3413 3418 ASSERT(spa->spa_trust_config);
3414 3419
3415 3420 /*
3416 3421 * Note: the code can handle other kinds of writes,
3417 3422 * but we don't expect them.
3418 3423 */
3419 3424 if (zio->io_vd->vdev_removing) {
3420 3425 ASSERT(zio->io_flags &
3421 3426 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3422 3427 ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3423 3428 }
3424 3429 }
3425 3430
3426 3431 align = 1ULL << vd->vdev_top->vdev_ashift;
3427 3432
3428 3433 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3429 3434 P2PHASE(zio->io_size, align) != 0) {
3430 3435 /* Transform logical writes to be a full physical block size. */
3431 3436 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3432 3437 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3433 3438 ASSERT(vd == vd->vdev_top);
3434 3439 if (zio->io_type == ZIO_TYPE_WRITE) {
3435 3440 abd_copy(abuf, zio->io_abd, zio->io_size);
3436 3441 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3437 3442 }
3438 3443 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3439 3444 }
3440 3445
3441 3446 /*
3442 3447 * If this is not a physical io, make sure that it is properly aligned
3443 3448 * before proceeding.
3444 3449 */
3445 3450 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3446 3451 ASSERT0(P2PHASE(zio->io_offset, align));
3447 3452 ASSERT0(P2PHASE(zio->io_size, align));
3448 3453 } else {
3449 3454 /*
3450 3455 * For physical writes, we allow 512b aligned writes and assume
3451 3456 * the device will perform a read-modify-write as necessary.
3452 3457 */
3453 3458 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3454 3459 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3455 3460 }
3456 3461
3457 3462 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
3458 3463
3459 3464 /*
3460 3465 * If this is a repair I/O, and there's no self-healing involved --
3461 3466 * that is, we're just resilvering what we expect to resilver --
3462 3467 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3463 3468 * This prevents spurious resilvering.
3464 3469 *
3465 3470 * There are a few ways that we can end up creating these spurious
3466 3471 * resilver i/os:
3467 3472 *
3468 3473 * 1. A resilver i/o will be issued if any DVA in the BP has a
3469 3474 * dirty DTL. The mirror code will issue resilver writes to
3470 3475 * each DVA, including the one(s) that are not on vdevs with dirty
3471 3476 * DTLs.
3472 3477 *
3473 3478 * 2. With nested replication, which happens when we have a
3474 3479 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3475 3480 * For example, given mirror(replacing(A+B), C), it's likely that
3476 3481 * only A is out of date (it's the new device). In this case, we'll
3477 3482 * read from C, then use the data to resilver A+B -- but we don't
3478 3483 * actually want to resilver B, just A. The top-level mirror has no
3479 3484 * way to know this, so instead we just discard unnecessary repairs
3480 3485 * as we work our way down the vdev tree.
3481 3486 *
3482 3487 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3483 3488 * The same logic applies to any form of nested replication: ditto
3484 3489 * + mirror, RAID-Z + replacing, etc.
3485 3490 *
3486 3491 * However, indirect vdevs point off to other vdevs which may have
3487 3492 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3488 3493 * will be properly bypassed instead.
3489 3494 */
3490 3495 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3491 3496 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3492 3497 zio->io_txg != 0 && /* not a delegated i/o */
3493 3498 vd->vdev_ops != &vdev_indirect_ops &&
3494 3499 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3495 3500 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3496 3501 zio_vdev_io_bypass(zio);
3497 3502 return (ZIO_PIPELINE_CONTINUE);
3498 3503 }
3499 3504
3500 3505 if (vd->vdev_ops->vdev_op_leaf && (zio->io_type == ZIO_TYPE_READ ||
3501 3506 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM)) {
3502 3507
3503 3508 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
3504 3509 return (ZIO_PIPELINE_CONTINUE);
3505 3510
3506 3511 if ((zio = vdev_queue_io(zio)) == NULL)
3507 3512 return (ZIO_PIPELINE_STOP);
3508 3513
3509 3514 if (!vdev_accessible(vd, zio)) {
3510 3515 zio->io_error = SET_ERROR(ENXIO);
3511 3516 zio_interrupt(zio);
3512 3517 return (ZIO_PIPELINE_STOP);
3513 3518 }
3514 3519 zio->io_delay = gethrtime();
3515 3520 }
3516 3521
3517 3522 vd->vdev_ops->vdev_op_io_start(zio);
3518 3523 return (ZIO_PIPELINE_STOP);
3519 3524 }
3520 3525
3521 3526 static int
3522 3527 zio_vdev_io_done(zio_t *zio)
3523 3528 {
3524 3529 vdev_t *vd = zio->io_vd;
3525 3530 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3526 3531 boolean_t unexpected_error = B_FALSE;
3527 3532
3528 3533 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3529 3534 return (ZIO_PIPELINE_STOP);
3530 3535 }
3531 3536
3532 3537 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3533 3538 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM);
3534 3539
3535 3540 if (zio->io_delay)
3536 3541 zio->io_delay = gethrtime() - zio->io_delay;
3537 3542
3538 3543 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
3539 3544
3540 3545 vdev_queue_io_done(zio);
3541 3546
3542 3547 if (zio->io_type == ZIO_TYPE_WRITE)
3543 3548 vdev_cache_write(zio);
3544 3549
3545 3550 if (zio_injection_enabled && zio->io_error == 0)
3546 3551 zio->io_error = zio_handle_device_injection(vd,
3547 3552 zio, EIO);
3548 3553
3549 3554 if (zio_injection_enabled && zio->io_error == 0)
3550 3555 zio->io_error = zio_handle_label_injection(zio, EIO);
3551 3556
3552 3557 if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) {
3553 3558 if (!vdev_accessible(vd, zio)) {
3554 3559 zio->io_error = SET_ERROR(ENXIO);
3555 3560 } else {
3556 3561 unexpected_error = B_TRUE;
3557 3562 }
3558 3563 }
3559 3564 }
3560 3565
3561 3566 ops->vdev_op_io_done(zio);
3562 3567
3563 3568 if (unexpected_error)
3564 3569 VERIFY(vdev_probe(vd, zio) == NULL);
3565 3570
3566 3571 return (ZIO_PIPELINE_CONTINUE);
3567 3572 }
3568 3573
3569 3574 /*
3570 3575 * This function is used to change the priority of an existing zio that is
3571 3576 * currently in-flight. This is used by the arc to upgrade priority in the
3572 3577 * event that a demand read is made for a block that is currently queued
3573 3578 * as a scrub or async read IO. Otherwise, the high priority read request
3574 3579 * would end up having to wait for the lower priority IO.
3575 3580 */
3576 3581 void
3577 3582 zio_change_priority(zio_t *pio, zio_priority_t priority)
3578 3583 {
3579 3584 zio_t *cio, *cio_next;
3580 3585 zio_link_t *zl = NULL;
3581 3586
3582 3587 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3583 3588
3584 3589 if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3585 3590 vdev_queue_change_io_priority(pio, priority);
3586 3591 } else {
3587 3592 pio->io_priority = priority;
3588 3593 }
3589 3594
3590 3595 mutex_enter(&pio->io_lock);
3591 3596 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3592 3597 cio_next = zio_walk_children(pio, &zl);
3593 3598 zio_change_priority(cio, priority);
3594 3599 }
3595 3600 mutex_exit(&pio->io_lock);
3596 3601 }
3597 3602
3598 3603 /*
3599 3604 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3600 3605 * disk, and use that to finish the checksum ereport later.
3601 3606 */
3602 3607 static void
3603 3608 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3604 3609 const abd_t *good_buf)
3605 3610 {
3606 3611 /* no processing needed */
3607 3612 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3608 3613 }
3609 3614
3610 3615 /*ARGSUSED*/
3611 3616 void
3612 3617 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3613 3618 {
3614 3619 void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size);
3615 3620
3616 3621 abd_copy(abd, zio->io_abd, zio->io_size);
3617 3622
3618 3623 zcr->zcr_cbinfo = zio->io_size;
3619 3624 zcr->zcr_cbdata = abd;
3620 3625 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3621 3626 zcr->zcr_free = zio_abd_free;
3622 3627 }
3623 3628
3624 3629 static int
3625 3630 zio_vdev_io_assess(zio_t *zio)
3626 3631 {
3627 3632 vdev_t *vd = zio->io_vd;
3628 3633
3629 3634 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3630 3635 return (ZIO_PIPELINE_STOP);
3631 3636 }
3632 3637
3633 3638 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3634 3639 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3635 3640
3636 3641 if (zio->io_vsd != NULL) {
3637 3642 zio->io_vsd_ops->vsd_free(zio);
3638 3643 zio->io_vsd = NULL;
3639 3644 }
3640 3645
3641 3646 if (zio_injection_enabled && zio->io_error == 0)
3642 3647 zio->io_error = zio_handle_fault_injection(zio, EIO);
3643 3648
3644 3649 /*
3645 3650 * If the I/O failed, determine whether we should attempt to retry it.
3646 3651 *
3647 3652 * On retry, we cut in line in the issue queue, since we don't want
3648 3653 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3649 3654 */
3650 3655 if (zio->io_error && vd == NULL &&
3651 3656 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3652 3657 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3653 3658 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3654 3659 zio->io_error = 0;
3655 3660 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3656 3661 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3657 3662 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3658 3663 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3659 3664 zio_requeue_io_start_cut_in_line);
3660 3665 return (ZIO_PIPELINE_STOP);
3661 3666 }
3662 3667
3663 3668 /*
3664 3669 * If we got an error on a leaf device, convert it to ENXIO
3665 3670 * if the device is not accessible at all.
3666 3671 */
3667 3672 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3668 3673 !vdev_accessible(vd, zio))
3669 3674 zio->io_error = SET_ERROR(ENXIO);
3670 3675
3671 3676 /*
3672 3677 * If we can't write to an interior vdev (mirror or RAID-Z),
3673 3678 * set vdev_cant_write so that we stop trying to allocate from it.
3674 3679 */
3675 3680 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3676 3681 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3677 3682 vd->vdev_cant_write = B_TRUE;
3678 3683 }
3679 3684
3680 3685 /*
3681 3686 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3682 3687 * attempts will ever succeed. In this case we set a persistent
3683 3688 * boolean flag so that we don't bother with it in the future.
3684 3689 */
3685 3690 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3686 3691 zio->io_type == ZIO_TYPE_IOCTL &&
3687 3692 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3688 3693 vd->vdev_nowritecache = B_TRUE;
3689 3694
3690 3695 if (zio->io_error)
3691 3696 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3692 3697
3693 3698 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3694 3699 zio->io_physdone != NULL) {
3695 3700 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3696 3701 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3697 3702 zio->io_physdone(zio->io_logical);
3698 3703 }
3699 3704
3700 3705 return (ZIO_PIPELINE_CONTINUE);
3701 3706 }
3702 3707
3703 3708 void
3704 3709 zio_vdev_io_reissue(zio_t *zio)
3705 3710 {
3706 3711 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3707 3712 ASSERT(zio->io_error == 0);
3708 3713
3709 3714 zio->io_stage >>= 1;
3710 3715 }
3711 3716
3712 3717 void
3713 3718 zio_vdev_io_redone(zio_t *zio)
3714 3719 {
3715 3720 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3716 3721
3717 3722 zio->io_stage >>= 1;
3718 3723 }
3719 3724
3720 3725 void
3721 3726 zio_vdev_io_bypass(zio_t *zio)
3722 3727 {
3723 3728 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3724 3729 ASSERT(zio->io_error == 0);
3725 3730
3726 3731 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3727 3732 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3728 3733 }
3729 3734
3730 3735 /*
3731 3736 * ==========================================================================
3732 3737 * Encrypt and store encryption parameters
3733 3738 * ==========================================================================
3734 3739 */
3735 3740
3736 3741
3737 3742 /*
3738 3743 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
3739 3744 * managing the storage of encryption parameters and passing them to the
3740 3745 * lower-level encryption functions.
3741 3746 */
3742 3747 static int
3743 3748 zio_encrypt(zio_t *zio)
3744 3749 {
3745 3750 zio_prop_t *zp = &zio->io_prop;
3746 3751 spa_t *spa = zio->io_spa;
3747 3752 blkptr_t *bp = zio->io_bp;
3748 3753 uint64_t psize = BP_GET_PSIZE(bp);
3749 3754 uint64_t dsobj = zio->io_bookmark.zb_objset;
3750 3755 dmu_object_type_t ot = BP_GET_TYPE(bp);
3751 3756 void *enc_buf = NULL;
3752 3757 abd_t *eabd = NULL;
3753 3758 uint8_t salt[ZIO_DATA_SALT_LEN];
3754 3759 uint8_t iv[ZIO_DATA_IV_LEN];
3755 3760 uint8_t mac[ZIO_DATA_MAC_LEN];
3756 3761 boolean_t no_crypt = B_FALSE;
3757 3762
3758 3763 /* the root zio already encrypted the data */
3759 3764 if (zio->io_child_type == ZIO_CHILD_GANG)
3760 3765 return (ZIO_PIPELINE_CONTINUE);
3761 3766
3762 3767 /* only ZIL blocks are re-encrypted on rewrite */
3763 3768 if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG)
3764 3769 return (ZIO_PIPELINE_CONTINUE);
3765 3770
3766 3771 if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) {
3767 3772 BP_SET_CRYPT(bp, B_FALSE);
3768 3773 return (ZIO_PIPELINE_CONTINUE);
3769 3774 }
3770 3775
3771 3776 /* if we are doing raw encryption set the provided encryption params */
3772 3777 if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) {
3773 3778 ASSERT0(BP_GET_LEVEL(bp));
3774 3779 BP_SET_CRYPT(bp, B_TRUE);
3775 3780 BP_SET_BYTEORDER(bp, zp->zp_byteorder);
3776 3781 if (ot != DMU_OT_OBJSET)
3777 3782 zio_crypt_encode_mac_bp(bp, zp->zp_mac);
3778 3783
3779 3784 /* dnode blocks must be written out in the provided byteorder */
3780 3785 if (zp->zp_byteorder != ZFS_HOST_BYTEORDER &&
3781 3786 ot == DMU_OT_DNODE) {
3782 3787 void *bswap_buf = zio_buf_alloc(psize);
3783 3788 abd_t *babd = abd_get_from_buf(bswap_buf, psize);
3784 3789
3785 3790 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
3786 3791 abd_copy_to_buf(bswap_buf, zio->io_abd, psize);
3787 3792 dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf,
3788 3793 psize);
3789 3794
3790 3795 abd_take_ownership_of_buf(babd, B_TRUE);
3791 3796 zio_push_transform(zio, babd, psize, psize, NULL);
3792 3797 }
3793 3798
3794 3799 if (DMU_OT_IS_ENCRYPTED(ot))
3795 3800 zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv);
3796 3801 return (ZIO_PIPELINE_CONTINUE);
3797 3802 }
3798 3803
3799 3804 /* indirect blocks only maintain a cksum of the lower level MACs */
3800 3805 if (BP_GET_LEVEL(bp) > 0) {
3801 3806 BP_SET_CRYPT(bp, B_TRUE);
3802 3807 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE,
3803 3808 zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp),
3804 3809 mac));
3805 3810 zio_crypt_encode_mac_bp(bp, mac);
3806 3811 return (ZIO_PIPELINE_CONTINUE);
3807 3812 }
3808 3813
3809 3814 /*
3810 3815 * Objset blocks are a special case since they have 2 256-bit MACs
3811 3816 * embedded within them.
3812 3817 */
3813 3818 if (ot == DMU_OT_OBJSET) {
3814 3819 ASSERT0(DMU_OT_IS_ENCRYPTED(ot));
3815 3820 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
3816 3821 BP_SET_CRYPT(bp, B_TRUE);
3817 3822 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj,
3818 3823 zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp)));
3819 3824 return (ZIO_PIPELINE_CONTINUE);
3820 3825 }
3821 3826
3822 3827 /* unencrypted object types are only authenticated with a MAC */
3823 3828 if (!DMU_OT_IS_ENCRYPTED(ot)) {
3824 3829 BP_SET_CRYPT(bp, B_TRUE);
3825 3830 VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj,
3826 3831 zio->io_abd, psize, mac));
3827 3832 zio_crypt_encode_mac_bp(bp, mac);
3828 3833 return (ZIO_PIPELINE_CONTINUE);
3829 3834 }
3830 3835
3831 3836 /*
3832 3837 * Later passes of sync-to-convergence may decide to rewrite data
3833 3838 * in place to avoid more disk reallocations. This presents a problem
3834 3839 * for encryption because this consitutes rewriting the new data with
3835 3840 * the same encryption key and IV. However, this only applies to blocks
3836 3841 * in the MOS (particularly the spacemaps) and we do not encrypt the
3837 3842 * MOS. We assert that the zio is allocating or an intent log write
3838 3843 * to enforce this.
3839 3844 */
3840 3845 ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG);
3841 3846 ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG);
3842 3847 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION));
3843 3848 ASSERT3U(psize, !=, 0);
3844 3849
3845 3850 enc_buf = zio_buf_alloc(psize);
3846 3851 eabd = abd_get_from_buf(enc_buf, psize);
3847 3852 abd_take_ownership_of_buf(eabd, B_TRUE);
3848 3853
3849 3854 /*
3850 3855 * For an explanation of what encryption parameters are stored
3851 3856 * where, see the block comment in zio_crypt.c.
3852 3857 */
3853 3858 if (ot == DMU_OT_INTENT_LOG) {
3854 3859 zio_crypt_decode_params_bp(bp, salt, iv);
3855 3860 } else {
3856 3861 BP_SET_CRYPT(bp, B_TRUE);
3857 3862 }
3858 3863
3859 3864 /* Perform the encryption. This should not fail */
3860 3865 VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark,
3861 3866 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
3862 3867 salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt));
3863 3868
3864 3869 /* encode encryption metadata into the bp */
3865 3870 if (ot == DMU_OT_INTENT_LOG) {
3866 3871 /*
3867 3872 * ZIL blocks store the MAC in the embedded checksum, so the
3868 3873 * transform must always be applied.
3869 3874 */
3870 3875 zio_crypt_encode_mac_zil(enc_buf, mac);
3871 3876 zio_push_transform(zio, eabd, psize, psize, NULL);
3872 3877 } else {
3873 3878 BP_SET_CRYPT(bp, B_TRUE);
3874 3879 zio_crypt_encode_params_bp(bp, salt, iv);
3875 3880 zio_crypt_encode_mac_bp(bp, mac);
3876 3881
3877 3882 if (no_crypt) {
3878 3883 ASSERT3U(ot, ==, DMU_OT_DNODE);
3879 3884 abd_free(eabd);
3880 3885 } else {
3881 3886 zio_push_transform(zio, eabd, psize, psize, NULL);
3882 3887 }
3883 3888 }
3884 3889
3885 3890 return (ZIO_PIPELINE_CONTINUE);
3886 3891 }
3887 3892
3888 3893 /*
3889 3894 * ==========================================================================
3890 3895 * Generate and verify checksums
3891 3896 * ==========================================================================
3892 3897 */
3893 3898 static int
3894 3899 zio_checksum_generate(zio_t *zio)
3895 3900 {
3896 3901 blkptr_t *bp = zio->io_bp;
3897 3902 enum zio_checksum checksum;
3898 3903
3899 3904 if (bp == NULL) {
3900 3905 /*
3901 3906 * This is zio_write_phys().
3902 3907 * We're either generating a label checksum, or none at all.
3903 3908 */
3904 3909 checksum = zio->io_prop.zp_checksum;
3905 3910
3906 3911 if (checksum == ZIO_CHECKSUM_OFF)
3907 3912 return (ZIO_PIPELINE_CONTINUE);
3908 3913
3909 3914 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3910 3915 } else {
3911 3916 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3912 3917 ASSERT(!IO_IS_ALLOCATING(zio));
3913 3918 checksum = ZIO_CHECKSUM_GANG_HEADER;
3914 3919 } else {
3915 3920 checksum = BP_GET_CHECKSUM(bp);
3916 3921 }
3917 3922 }
3918 3923
3919 3924 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
3920 3925
3921 3926 return (ZIO_PIPELINE_CONTINUE);
3922 3927 }
3923 3928
3924 3929 static int
3925 3930 zio_checksum_verify(zio_t *zio)
3926 3931 {
3927 3932 zio_bad_cksum_t info;
3928 3933 blkptr_t *bp = zio->io_bp;
3929 3934 int error;
3930 3935
3931 3936 ASSERT(zio->io_vd != NULL);
3932 3937
3933 3938 if (bp == NULL) {
3934 3939 /*
3935 3940 * This is zio_read_phys().
3936 3941 * We're either verifying a label checksum, or nothing at all.
3937 3942 */
3938 3943 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3939 3944 return (ZIO_PIPELINE_CONTINUE);
3940 3945
3941 3946 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3942 3947 }
3943 3948
3944 3949 if ((error = zio_checksum_error(zio, &info)) != 0) {
3945 3950 zio->io_error = error;
3946 3951 if (error == ECKSUM &&
3947 3952 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3948 3953 zfs_ereport_start_checksum(zio->io_spa,
3949 3954 zio->io_vd, &zio->io_bookmark, zio,
3950 3955 zio->io_offset, zio->io_size, NULL, &info);
3951 3956 }
3952 3957 }
3953 3958
3954 3959 return (ZIO_PIPELINE_CONTINUE);
3955 3960 }
3956 3961
3957 3962 /*
3958 3963 * Called by RAID-Z to ensure we don't compute the checksum twice.
3959 3964 */
3960 3965 void
3961 3966 zio_checksum_verified(zio_t *zio)
3962 3967 {
3963 3968 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3964 3969 }
3965 3970
3966 3971 /*
3967 3972 * ==========================================================================
3968 3973 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3969 3974 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3970 3975 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3971 3976 * indicate errors that are specific to one I/O, and most likely permanent.
3972 3977 * Any other error is presumed to be worse because we weren't expecting it.
3973 3978 * ==========================================================================
3974 3979 */
3975 3980 int
3976 3981 zio_worst_error(int e1, int e2)
3977 3982 {
3978 3983 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3979 3984 int r1, r2;
3980 3985
3981 3986 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3982 3987 if (e1 == zio_error_rank[r1])
3983 3988 break;
3984 3989
3985 3990 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3986 3991 if (e2 == zio_error_rank[r2])
3987 3992 break;
3988 3993
3989 3994 return (r1 > r2 ? e1 : e2);
3990 3995 }
3991 3996
3992 3997 /*
3993 3998 * ==========================================================================
3994 3999 * I/O completion
3995 4000 * ==========================================================================
3996 4001 */
3997 4002 static int
3998 4003 zio_ready(zio_t *zio)
3999 4004 {
4000 4005 blkptr_t *bp = zio->io_bp;
4001 4006 zio_t *pio, *pio_next;
4002 4007 zio_link_t *zl = NULL;
4003 4008
4004 4009 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
4005 4010 ZIO_WAIT_READY)) {
4006 4011 return (ZIO_PIPELINE_STOP);
4007 4012 }
4008 4013
4009 4014 if (zio->io_ready) {
4010 4015 ASSERT(IO_IS_ALLOCATING(zio));
4011 4016 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
4012 4017 (zio->io_flags & ZIO_FLAG_NOPWRITE));
4013 4018 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
4014 4019
4015 4020 zio->io_ready(zio);
4016 4021 }
4017 4022
4018 4023 if (bp != NULL && bp != &zio->io_bp_copy)
4019 4024 zio->io_bp_copy = *bp;
4020 4025
4021 4026 if (zio->io_error != 0) {
4022 4027 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4023 4028
4024 4029 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4025 4030 ASSERT(IO_IS_ALLOCATING(zio));
4026 4031 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4027 4032 ASSERT(zio->io_metaslab_class != NULL);
4028 4033
4029 4034 /*
4030 4035 * We were unable to allocate anything, unreserve and
4031 4036 * issue the next I/O to allocate.
4032 4037 */
4033 4038 metaslab_class_throttle_unreserve(
4034 4039 zio->io_metaslab_class, zio->io_prop.zp_copies,
4035 4040 zio->io_allocator, zio);
4036 4041 zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
4037 4042 }
4038 4043 }
4039 4044
4040 4045 mutex_enter(&zio->io_lock);
4041 4046 zio->io_state[ZIO_WAIT_READY] = 1;
4042 4047 pio = zio_walk_parents(zio, &zl);
4043 4048 mutex_exit(&zio->io_lock);
4044 4049
4045 4050 /*
4046 4051 * As we notify zio's parents, new parents could be added.
4047 4052 * New parents go to the head of zio's io_parent_list, however,
4048 4053 * so we will (correctly) not notify them. The remainder of zio's
4049 4054 * io_parent_list, from 'pio_next' onward, cannot change because
4050 4055 * all parents must wait for us to be done before they can be done.
4051 4056 */
4052 4057 for (; pio != NULL; pio = pio_next) {
4053 4058 pio_next = zio_walk_parents(zio, &zl);
4054 4059 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
4055 4060 }
4056 4061
4057 4062 if (zio->io_flags & ZIO_FLAG_NODATA) {
4058 4063 if (BP_IS_GANG(bp)) {
4059 4064 zio->io_flags &= ~ZIO_FLAG_NODATA;
4060 4065 } else {
4061 4066 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
4062 4067 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
4063 4068 }
4064 4069 }
4065 4070
4066 4071 if (zio_injection_enabled &&
4067 4072 zio->io_spa->spa_syncing_txg == zio->io_txg)
4068 4073 zio_handle_ignored_writes(zio);
4069 4074
4070 4075 return (ZIO_PIPELINE_CONTINUE);
4071 4076 }
4072 4077
4073 4078 /*
4074 4079 * Update the allocation throttle accounting.
4075 4080 */
4076 4081 static void
4077 4082 zio_dva_throttle_done(zio_t *zio)
4078 4083 {
4079 4084 zio_t *lio = zio->io_logical;
4080 4085 zio_t *pio = zio_unique_parent(zio);
4081 4086 vdev_t *vd = zio->io_vd;
4082 4087 int flags = METASLAB_ASYNC_ALLOC;
4083 4088
4084 4089 ASSERT3P(zio->io_bp, !=, NULL);
4085 4090 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
4086 4091 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
4087 4092 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
4088 4093 ASSERT(vd != NULL);
4089 4094 ASSERT3P(vd, ==, vd->vdev_top);
4090 4095 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
4091 4096 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
4092 4097 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
4093 4098 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
4094 4099
4095 4100 /*
4096 4101 * Parents of gang children can have two flavors -- ones that
4097 4102 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4098 4103 * and ones that allocated the constituent blocks. The allocation
4099 4104 * throttle needs to know the allocating parent zio so we must find
4100 4105 * it here.
4101 4106 */
4102 4107 if (pio->io_child_type == ZIO_CHILD_GANG) {
4103 4108 /*
4104 4109 * If our parent is a rewrite gang child then our grandparent
4105 4110 * would have been the one that performed the allocation.
4106 4111 */
4107 4112 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
4108 4113 pio = zio_unique_parent(pio);
4109 4114 flags |= METASLAB_GANG_CHILD;
4110 4115 }
4111 4116
4112 4117 ASSERT(IO_IS_ALLOCATING(pio));
4113 4118 ASSERT3P(zio, !=, zio->io_logical);
4114 4119 ASSERT(zio->io_logical != NULL);
4115 4120 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4116 4121 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
4117 4122 ASSERT(zio->io_metaslab_class != NULL);
4118 4123
4119 4124 mutex_enter(&pio->io_lock);
4120 4125 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
4121 4126 pio->io_allocator, B_TRUE);
4122 4127 mutex_exit(&pio->io_lock);
4123 4128
4124 4129 metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1,
4125 4130 pio->io_allocator, pio);
4126 4131
4127 4132 /*
4128 4133 * Call into the pipeline to see if there is more work that
4129 4134 * needs to be done. If there is work to be done it will be
4130 4135 * dispatched to another taskq thread.
4131 4136 */
4132 4137 zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
4133 4138 }
4134 4139
4135 4140 static int
4136 4141 zio_done(zio_t *zio)
4137 4142 {
4138 4143 spa_t *spa = zio->io_spa;
4139 4144 zio_t *lio = zio->io_logical;
4140 4145 blkptr_t *bp = zio->io_bp;
4141 4146 vdev_t *vd = zio->io_vd;
4142 4147 uint64_t psize = zio->io_size;
4143 4148 zio_t *pio, *pio_next;
4144 4149 zio_link_t *zl = NULL;
4145 4150
4146 4151 /*
4147 4152 * If our children haven't all completed,
4148 4153 * wait for them and then repeat this pipeline stage.
4149 4154 */
4150 4155 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
4151 4156 return (ZIO_PIPELINE_STOP);
4152 4157 }
4153 4158
4154 4159 /*
4155 4160 * If the allocation throttle is enabled, then update the accounting.
4156 4161 * We only track child I/Os that are part of an allocating async
4157 4162 * write. We must do this since the allocation is performed
4158 4163 * by the logical I/O but the actual write is done by child I/Os.
4159 4164 */
4160 4165 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
4161 4166 zio->io_child_type == ZIO_CHILD_VDEV) {
4162 4167 ASSERT(zio->io_metaslab_class != NULL);
4163 4168 ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled);
4164 4169 zio_dva_throttle_done(zio);
4165 4170 }
4166 4171
4167 4172 /*
4168 4173 * If the allocation throttle is enabled, verify that
4169 4174 * we have decremented the refcounts for every I/O that was throttled.
4170 4175 */
4171 4176 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4172 4177 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4173 4178 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4174 4179 ASSERT(bp != NULL);
4175 4180
4176 4181 metaslab_group_alloc_verify(spa, zio->io_bp, zio,
4177 4182 zio->io_allocator);
4178 4183 VERIFY(zfs_refcount_not_held(
4179 4184 &zio->io_metaslab_class->mc_alloc_slots[zio->io_allocator],
4180 4185 zio));
4181 4186 }
4182 4187
4183 4188 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
4184 4189 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
4185 4190 ASSERT(zio->io_children[c][w] == 0);
4186 4191
4187 4192 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
4188 4193 ASSERT(bp->blk_pad[0] == 0);
4189 4194 ASSERT(bp->blk_pad[1] == 0);
4190 4195 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
4191 4196 (bp == zio_unique_parent(zio)->io_bp));
4192 4197 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
4193 4198 zio->io_bp_override == NULL &&
4194 4199 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
4195 4200 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
4196 4201 ASSERT(BP_COUNT_GANG(bp) == 0 ||
4197 4202 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
4198 4203 }
4199 4204 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
4200 4205 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
4201 4206 }
4202 4207
4203 4208 /*
4204 4209 * If there were child vdev/gang/ddt errors, they apply to us now.
4205 4210 */
4206 4211 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
4207 4212 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
4208 4213 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
4209 4214
4210 4215 /*
4211 4216 * If the I/O on the transformed data was successful, generate any
4212 4217 * checksum reports now while we still have the transformed data.
4213 4218 */
4214 4219 if (zio->io_error == 0) {
4215 4220 while (zio->io_cksum_report != NULL) {
4216 4221 zio_cksum_report_t *zcr = zio->io_cksum_report;
4217 4222 uint64_t align = zcr->zcr_align;
4218 4223 uint64_t asize = P2ROUNDUP(psize, align);
4219 4224 abd_t *adata = zio->io_abd;
4220 4225
4221 4226 if (asize != psize) {
4222 4227 adata = abd_alloc(asize, B_TRUE);
4223 4228 abd_copy(adata, zio->io_abd, psize);
4224 4229 abd_zero_off(adata, psize, asize - psize);
4225 4230 }
4226 4231
4227 4232 zio->io_cksum_report = zcr->zcr_next;
4228 4233 zcr->zcr_next = NULL;
4229 4234 zcr->zcr_finish(zcr, adata);
4230 4235 zfs_ereport_free_checksum(zcr);
4231 4236
4232 4237 if (asize != psize)
4233 4238 abd_free(adata);
4234 4239 }
4235 4240 }
4236 4241
4237 4242 zio_pop_transforms(zio); /* note: may set zio->io_error */
4238 4243
4239 4244 vdev_stat_update(zio, psize);
4240 4245
4241 4246 if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) {
4242 4247 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) {
4243 4248 /*
4244 4249 * We want to only increment our slow IO counters if
4245 4250 * the IO is valid (i.e. not if the drive is removed).
4246 4251 *
4247 4252 * zfs_ereport_post() will also do these checks, but
4248 4253 * it can also have other failures, so we need to
4249 4254 * increment the slow_io counters independent of it.
4250 4255 */
4251 4256 if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY,
4252 4257 zio->io_spa, zio->io_vd, zio)) {
4253 4258 mutex_enter(&zio->io_vd->vdev_stat_lock);
4254 4259 zio->io_vd->vdev_stat.vs_slow_ios++;
4255 4260 mutex_exit(&zio->io_vd->vdev_stat_lock);
4256 4261
4257 4262 zfs_ereport_post(FM_EREPORT_ZFS_DELAY,
4258 4263 zio->io_spa, zio->io_vd, &zio->io_bookmark,
4259 4264 zio, 0, 0);
4260 4265 }
4261 4266 }
4262 4267 }
4263 4268
4264 4269 if (zio->io_error) {
4265 4270 /*
4266 4271 * If this I/O is attached to a particular vdev,
4267 4272 * generate an error message describing the I/O failure
4268 4273 * at the block level. We ignore these errors if the
4269 4274 * device is currently unavailable.
4270 4275 */
4271 4276 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
4272 4277 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd,
4273 4278 &zio->io_bookmark, zio, 0, 0);
4274 4279
4275 4280 if ((zio->io_error == EIO || !(zio->io_flags &
4276 4281 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
4277 4282 zio == lio) {
4278 4283 /*
4279 4284 * For logical I/O requests, tell the SPA to log the
4280 4285 * error and generate a logical data ereport.
4281 4286 */
4282 4287 spa_log_error(spa, &zio->io_bookmark);
4283 4288 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL,
4284 4289 &zio->io_bookmark, zio, 0, 0);
4285 4290 }
4286 4291 }
4287 4292
4288 4293 if (zio->io_error && zio == lio) {
4289 4294 /*
4290 4295 * Determine whether zio should be reexecuted. This will
4291 4296 * propagate all the way to the root via zio_notify_parent().
4292 4297 */
4293 4298 ASSERT(vd == NULL && bp != NULL);
4294 4299 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4295 4300
4296 4301 if (IO_IS_ALLOCATING(zio) &&
4297 4302 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
4298 4303 if (zio->io_error != ENOSPC)
4299 4304 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
4300 4305 else
4301 4306 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4302 4307 }
4303 4308
4304 4309 if ((zio->io_type == ZIO_TYPE_READ ||
4305 4310 zio->io_type == ZIO_TYPE_FREE) &&
4306 4311 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
4307 4312 zio->io_error == ENXIO &&
4308 4313 spa_load_state(spa) == SPA_LOAD_NONE &&
4309 4314 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
4310 4315 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4311 4316
4312 4317 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
4313 4318 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4314 4319
4315 4320 /*
4316 4321 * Here is a possibly good place to attempt to do
4317 4322 * either combinatorial reconstruction or error correction
4318 4323 * based on checksums. It also might be a good place
4319 4324 * to send out preliminary ereports before we suspend
4320 4325 * processing.
4321 4326 */
4322 4327 }
4323 4328
4324 4329 /*
4325 4330 * If there were logical child errors, they apply to us now.
4326 4331 * We defer this until now to avoid conflating logical child
4327 4332 * errors with errors that happened to the zio itself when
4328 4333 * updating vdev stats and reporting FMA events above.
4329 4334 */
4330 4335 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
4331 4336
4332 4337 if ((zio->io_error || zio->io_reexecute) &&
4333 4338 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
4334 4339 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
4335 4340 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
4336 4341
4337 4342 zio_gang_tree_free(&zio->io_gang_tree);
4338 4343
4339 4344 /*
4340 4345 * Godfather I/Os should never suspend.
4341 4346 */
4342 4347 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4343 4348 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4344 4349 zio->io_reexecute = 0;
4345 4350
4346 4351 if (zio->io_reexecute) {
4347 4352 /*
4348 4353 * This is a logical I/O that wants to reexecute.
4349 4354 *
4350 4355 * Reexecute is top-down. When an i/o fails, if it's not
4351 4356 * the root, it simply notifies its parent and sticks around.
4352 4357 * The parent, seeing that it still has children in zio_done(),
4353 4358 * does the same. This percolates all the way up to the root.
4354 4359 * The root i/o will reexecute or suspend the entire tree.
4355 4360 *
4356 4361 * This approach ensures that zio_reexecute() honors
4357 4362 * all the original i/o dependency relationships, e.g.
4358 4363 * parents not executing until children are ready.
4359 4364 */
4360 4365 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4361 4366
4362 4367 zio->io_gang_leader = NULL;
4363 4368
4364 4369 mutex_enter(&zio->io_lock);
4365 4370 zio->io_state[ZIO_WAIT_DONE] = 1;
4366 4371 mutex_exit(&zio->io_lock);
4367 4372
4368 4373 /*
4369 4374 * "The Godfather" I/O monitors its children but is
4370 4375 * not a true parent to them. It will track them through
4371 4376 * the pipeline but severs its ties whenever they get into
4372 4377 * trouble (e.g. suspended). This allows "The Godfather"
4373 4378 * I/O to return status without blocking.
4374 4379 */
4375 4380 zl = NULL;
4376 4381 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4377 4382 pio = pio_next) {
4378 4383 zio_link_t *remove_zl = zl;
4379 4384 pio_next = zio_walk_parents(zio, &zl);
4380 4385
4381 4386 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4382 4387 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4383 4388 zio_remove_child(pio, zio, remove_zl);
4384 4389 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4385 4390 }
4386 4391 }
4387 4392
4388 4393 if ((pio = zio_unique_parent(zio)) != NULL) {
4389 4394 /*
4390 4395 * We're not a root i/o, so there's nothing to do
4391 4396 * but notify our parent. Don't propagate errors
4392 4397 * upward since we haven't permanently failed yet.
4393 4398 */
4394 4399 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4395 4400 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4396 4401 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4397 4402 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4398 4403 /*
4399 4404 * We'd fail again if we reexecuted now, so suspend
4400 4405 * until conditions improve (e.g. device comes online).
4401 4406 */
4402 4407 zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
4403 4408 } else {
4404 4409 /*
4405 4410 * Reexecution is potentially a huge amount of work.
4406 4411 * Hand it off to the otherwise-unused claim taskq.
4407 4412 */
4408 4413 ASSERT(zio->io_tqent.tqent_next == NULL);
4409 4414 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
4410 4415 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
4411 4416 0, &zio->io_tqent);
4412 4417 }
4413 4418 return (ZIO_PIPELINE_STOP);
4414 4419 }
4415 4420
4416 4421 ASSERT(zio->io_child_count == 0);
4417 4422 ASSERT(zio->io_reexecute == 0);
4418 4423 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4419 4424
4420 4425 /*
4421 4426 * Report any checksum errors, since the I/O is complete.
4422 4427 */
4423 4428 while (zio->io_cksum_report != NULL) {
4424 4429 zio_cksum_report_t *zcr = zio->io_cksum_report;
4425 4430 zio->io_cksum_report = zcr->zcr_next;
4426 4431 zcr->zcr_next = NULL;
4427 4432 zcr->zcr_finish(zcr, NULL);
4428 4433 zfs_ereport_free_checksum(zcr);
4429 4434 }
4430 4435
4431 4436 /*
4432 4437 * It is the responsibility of the done callback to ensure that this
4433 4438 * particular zio is no longer discoverable for adoption, and as
4434 4439 * such, cannot acquire any new parents.
4435 4440 */
4436 4441 if (zio->io_done)
4437 4442 zio->io_done(zio);
4438 4443
4439 4444 mutex_enter(&zio->io_lock);
4440 4445 zio->io_state[ZIO_WAIT_DONE] = 1;
4441 4446 mutex_exit(&zio->io_lock);
4442 4447
4443 4448 zl = NULL;
4444 4449 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4445 4450 zio_link_t *remove_zl = zl;
4446 4451 pio_next = zio_walk_parents(zio, &zl);
4447 4452 zio_remove_child(pio, zio, remove_zl);
4448 4453 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4449 4454 }
4450 4455
4451 4456 if (zio->io_waiter != NULL) {
4452 4457 mutex_enter(&zio->io_lock);
4453 4458 zio->io_executor = NULL;
4454 4459 cv_broadcast(&zio->io_cv);
4455 4460 mutex_exit(&zio->io_lock);
4456 4461 } else {
4457 4462 zio_destroy(zio);
4458 4463 }
4459 4464
4460 4465 return (ZIO_PIPELINE_STOP);
4461 4466 }
4462 4467
4463 4468 /*
4464 4469 * ==========================================================================
4465 4470 * I/O pipeline definition
4466 4471 * ==========================================================================
4467 4472 */
4468 4473 static zio_pipe_stage_t *zio_pipeline[] = {
4469 4474 NULL,
4470 4475 zio_read_bp_init,
4471 4476 zio_write_bp_init,
4472 4477 zio_free_bp_init,
4473 4478 zio_issue_async,
4474 4479 zio_write_compress,
4475 4480 zio_encrypt,
4476 4481 zio_checksum_generate,
4477 4482 zio_nop_write,
4478 4483 zio_ddt_read_start,
4479 4484 zio_ddt_read_done,
4480 4485 zio_ddt_write,
4481 4486 zio_ddt_free,
4482 4487 zio_gang_assemble,
4483 4488 zio_gang_issue,
4484 4489 zio_dva_throttle,
4485 4490 zio_dva_allocate,
4486 4491 zio_dva_free,
4487 4492 zio_dva_claim,
4488 4493 zio_ready,
4489 4494 zio_vdev_io_start,
4490 4495 zio_vdev_io_done,
4491 4496 zio_vdev_io_assess,
4492 4497 zio_checksum_verify,
4493 4498 zio_done
4494 4499 };
4495 4500
4496 4501
4497 4502
4498 4503
4499 4504 /*
4500 4505 * Compare two zbookmark_phys_t's to see which we would reach first in a
4501 4506 * pre-order traversal of the object tree.
4502 4507 *
4503 4508 * This is simple in every case aside from the meta-dnode object. For all other
4504 4509 * objects, we traverse them in order (object 1 before object 2, and so on).
4505 4510 * However, all of these objects are traversed while traversing object 0, since
4506 4511 * the data it points to is the list of objects. Thus, we need to convert to a
4507 4512 * canonical representation so we can compare meta-dnode bookmarks to
4508 4513 * non-meta-dnode bookmarks.
4509 4514 *
4510 4515 * We do this by calculating "equivalents" for each field of the zbookmark.
4511 4516 * zbookmarks outside of the meta-dnode use their own object and level, and
4512 4517 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4513 4518 * blocks this bookmark refers to) by multiplying their blkid by their span
4514 4519 * (the number of L0 blocks contained within one block at their level).
4515 4520 * zbookmarks inside the meta-dnode calculate their object equivalent
4516 4521 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4517 4522 * level + 1<<31 (any value larger than a level could ever be) for their level.
4518 4523 * This causes them to always compare before a bookmark in their object
4519 4524 * equivalent, compare appropriately to bookmarks in other objects, and to
4520 4525 * compare appropriately to other bookmarks in the meta-dnode.
4521 4526 */
4522 4527 int
4523 4528 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4524 4529 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4525 4530 {
4526 4531 /*
4527 4532 * These variables represent the "equivalent" values for the zbookmark,
4528 4533 * after converting zbookmarks inside the meta dnode to their
4529 4534 * normal-object equivalents.
4530 4535 */
4531 4536 uint64_t zb1obj, zb2obj;
4532 4537 uint64_t zb1L0, zb2L0;
4533 4538 uint64_t zb1level, zb2level;
4534 4539
4535 4540 if (zb1->zb_object == zb2->zb_object &&
4536 4541 zb1->zb_level == zb2->zb_level &&
4537 4542 zb1->zb_blkid == zb2->zb_blkid)
4538 4543 return (0);
4539 4544
4540 4545 /*
4541 4546 * BP_SPANB calculates the span in blocks.
4542 4547 */
4543 4548 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4544 4549 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4545 4550
4546 4551 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4547 4552 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4548 4553 zb1L0 = 0;
4549 4554 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4550 4555 } else {
4551 4556 zb1obj = zb1->zb_object;
4552 4557 zb1level = zb1->zb_level;
4553 4558 }
4554 4559
4555 4560 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4556 4561 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4557 4562 zb2L0 = 0;
4558 4563 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4559 4564 } else {
4560 4565 zb2obj = zb2->zb_object;
4561 4566 zb2level = zb2->zb_level;
4562 4567 }
4563 4568
4564 4569 /* Now that we have a canonical representation, do the comparison. */
4565 4570 if (zb1obj != zb2obj)
4566 4571 return (zb1obj < zb2obj ? -1 : 1);
4567 4572 else if (zb1L0 != zb2L0)
4568 4573 return (zb1L0 < zb2L0 ? -1 : 1);
4569 4574 else if (zb1level != zb2level)
4570 4575 return (zb1level > zb2level ? -1 : 1);
4571 4576 /*
4572 4577 * This can (theoretically) happen if the bookmarks have the same object
4573 4578 * and level, but different blkids, if the block sizes are not the same.
4574 4579 * There is presently no way to change the indirect block sizes
4575 4580 */
4576 4581 return (0);
4577 4582 }
4578 4583
4579 4584 /*
4580 4585 * This function checks the following: given that last_block is the place that
4581 4586 * our traversal stopped last time, does that guarantee that we've visited
4582 4587 * every node under subtree_root? Therefore, we can't just use the raw output
4583 4588 * of zbookmark_compare. We have to pass in a modified version of
4584 4589 * subtree_root; by incrementing the block id, and then checking whether
4585 4590 * last_block is before or equal to that, we can tell whether or not having
4586 4591 * visited last_block implies that all of subtree_root's children have been
4587 4592 * visited.
4588 4593 */
4589 4594 boolean_t
4590 4595 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4591 4596 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4592 4597 {
4593 4598 zbookmark_phys_t mod_zb = *subtree_root;
4594 4599 mod_zb.zb_blkid++;
4595 4600 ASSERT(last_block->zb_level == 0);
4596 4601
4597 4602 /* The objset_phys_t isn't before anything. */
4598 4603 if (dnp == NULL)
4599 4604 return (B_FALSE);
4600 4605
4601 4606 /*
4602 4607 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4603 4608 * data block size in sectors, because that variable is only used if
4604 4609 * the bookmark refers to a block in the meta-dnode. Since we don't
4605 4610 * know without examining it what object it refers to, and there's no
4606 4611 * harm in passing in this value in other cases, we always pass it in.
4607 4612 *
4608 4613 * We pass in 0 for the indirect block size shift because zb2 must be
4609 4614 * level 0. The indirect block size is only used to calculate the span
4610 4615 * of the bookmark, but since the bookmark must be level 0, the span is
4611 4616 * always 1, so the math works out.
4612 4617 *
4613 4618 * If you make changes to how the zbookmark_compare code works, be sure
4614 4619 * to make sure that this code still works afterwards.
4615 4620 */
4616 4621 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4617 4622 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
4618 4623 last_block) <= 0);
4619 4624 }
|
↓ open down ↓ |
2623 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX