Print this page
3742 zfs comments need cleaner, more consistent style
Submitted by: Will Andrews <willa@spectralogic.com>
Submitted by: Alan Somers <alans@spectralogic.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Eric Schrock <eric.schrock@delphix.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
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) 2013 by Delphix. All rights reserved.
24 24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 25 */
26 26
27 27 #include <sys/zfs_context.h>
28 28 #include <sys/fm/fs/zfs.h>
29 29 #include <sys/spa.h>
30 30 #include <sys/txg.h>
31 31 #include <sys/spa_impl.h>
32 32 #include <sys/vdev_impl.h>
33 33 #include <sys/zio_impl.h>
34 34 #include <sys/zio_compress.h>
35 35 #include <sys/zio_checksum.h>
36 36 #include <sys/dmu_objset.h>
37 37 #include <sys/arc.h>
38 38 #include <sys/ddt.h>
39 39
40 40 /*
41 41 * ==========================================================================
42 42 * I/O priority table
43 43 * ==========================================================================
44 44 */
45 45 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
46 46 0, /* ZIO_PRIORITY_NOW */
47 47 0, /* ZIO_PRIORITY_SYNC_READ */
48 48 0, /* ZIO_PRIORITY_SYNC_WRITE */
49 49 0, /* ZIO_PRIORITY_LOG_WRITE */
50 50 1, /* ZIO_PRIORITY_CACHE_FILL */
51 51 1, /* ZIO_PRIORITY_AGG */
52 52 4, /* ZIO_PRIORITY_FREE */
53 53 4, /* ZIO_PRIORITY_ASYNC_WRITE */
54 54 6, /* ZIO_PRIORITY_ASYNC_READ */
55 55 10, /* ZIO_PRIORITY_RESILVER */
56 56 20, /* ZIO_PRIORITY_SCRUB */
57 57 2, /* ZIO_PRIORITY_DDT_PREFETCH */
58 58 };
59 59
60 60 /*
61 61 * ==========================================================================
62 62 * I/O type descriptions
63 63 * ==========================================================================
64 64 */
65 65 char *zio_type_name[ZIO_TYPES] = {
66 66 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
67 67 "zio_ioctl"
68 68 };
69 69
70 70 /*
71 71 * ==========================================================================
72 72 * I/O kmem caches
73 73 * ==========================================================================
74 74 */
75 75 kmem_cache_t *zio_cache;
76 76 kmem_cache_t *zio_link_cache;
77 77 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
78 78 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
79 79
80 80 #ifdef _KERNEL
81 81 extern vmem_t *zio_alloc_arena;
82 82 #endif
83 83 extern int zfs_mg_alloc_failures;
84 84
85 85 /*
86 86 * The following actions directly effect the spa's sync-to-convergence logic.
87 87 * The values below define the sync pass when we start performing the action.
88 88 * Care should be taken when changing these values as they directly impact
89 89 * spa_sync() performance. Tuning these values may introduce subtle performance
90 90 * pathologies and should only be done in the context of performance analysis.
91 91 * These tunables will eventually be removed and replaced with #defines once
92 92 * enough analysis has been done to determine optimal values.
93 93 *
94 94 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
95 95 * regular blocks are not deferred.
96 96 */
97 97 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
98 98 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
99 99 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
100 100
101 101 /*
102 102 * An allocating zio is one that either currently has the DVA allocate
103 103 * stage set or will have it later in its lifetime.
104 104 */
105 105 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
106 106
107 107 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
108 108
109 109 #ifdef ZFS_DEBUG
110 110 int zio_buf_debug_limit = 16384;
111 111 #else
112 112 int zio_buf_debug_limit = 0;
113 113 #endif
114 114
115 115 void
116 116 zio_init(void)
117 117 {
118 118 size_t c;
119 119 vmem_t *data_alloc_arena = NULL;
120 120
121 121 #ifdef _KERNEL
122 122 data_alloc_arena = zio_alloc_arena;
123 123 #endif
124 124 zio_cache = kmem_cache_create("zio_cache",
125 125 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
126 126 zio_link_cache = kmem_cache_create("zio_link_cache",
127 127 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
128 128
129 129 /*
130 130 * For small buffers, we want a cache for each multiple of
131 131 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
132 132 * for each quarter-power of 2. For large buffers, we want
133 133 * a cache for each multiple of PAGESIZE.
134 134 */
135 135 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
136 136 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
137 137 size_t p2 = size;
138 138 size_t align = 0;
139 139 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
140 140
141 141 while (p2 & (p2 - 1))
142 142 p2 &= p2 - 1;
143 143
144 144 #ifndef _KERNEL
145 145 /*
146 146 * If we are using watchpoints, put each buffer on its own page,
147 147 * to eliminate the performance overhead of trapping to the
148 148 * kernel when modifying a non-watched buffer that shares the
149 149 * page with a watched buffer.
150 150 */
151 151 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
152 152 continue;
153 153 #endif
154 154 if (size <= 4 * SPA_MINBLOCKSIZE) {
155 155 align = SPA_MINBLOCKSIZE;
156 156 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
157 157 align = PAGESIZE;
158 158 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
159 159 align = p2 >> 2;
160 160 }
161 161
162 162 if (align != 0) {
163 163 char name[36];
164 164 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
165 165 zio_buf_cache[c] = kmem_cache_create(name, size,
166 166 align, NULL, NULL, NULL, NULL, NULL, cflags);
167 167
168 168 /*
169 169 * Since zio_data bufs do not appear in crash dumps, we
170 170 * pass KMC_NOTOUCH so that no allocator metadata is
171 171 * stored with the buffers.
172 172 */
173 173 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
174 174 zio_data_buf_cache[c] = kmem_cache_create(name, size,
175 175 align, NULL, NULL, NULL, NULL, data_alloc_arena,
176 176 cflags | KMC_NOTOUCH);
177 177 }
178 178 }
179 179
180 180 while (--c != 0) {
181 181 ASSERT(zio_buf_cache[c] != NULL);
182 182 if (zio_buf_cache[c - 1] == NULL)
183 183 zio_buf_cache[c - 1] = zio_buf_cache[c];
184 184
185 185 ASSERT(zio_data_buf_cache[c] != NULL);
186 186 if (zio_data_buf_cache[c - 1] == NULL)
187 187 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
188 188 }
189 189
190 190 /*
191 191 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
192 192 * to fail 3 times per txg or 8 failures, whichever is greater.
193 193 */
194 194 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
195 195
196 196 zio_inject_init();
197 197 }
198 198
199 199 void
200 200 zio_fini(void)
201 201 {
202 202 size_t c;
203 203 kmem_cache_t *last_cache = NULL;
204 204 kmem_cache_t *last_data_cache = NULL;
205 205
206 206 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
207 207 if (zio_buf_cache[c] != last_cache) {
208 208 last_cache = zio_buf_cache[c];
209 209 kmem_cache_destroy(zio_buf_cache[c]);
210 210 }
211 211 zio_buf_cache[c] = NULL;
212 212
213 213 if (zio_data_buf_cache[c] != last_data_cache) {
214 214 last_data_cache = zio_data_buf_cache[c];
215 215 kmem_cache_destroy(zio_data_buf_cache[c]);
216 216 }
217 217 zio_data_buf_cache[c] = NULL;
218 218 }
219 219
220 220 kmem_cache_destroy(zio_link_cache);
221 221 kmem_cache_destroy(zio_cache);
222 222
223 223 zio_inject_fini();
224 224 }
225 225
226 226 /*
227 227 * ==========================================================================
228 228 * Allocate and free I/O buffers
229 229 * ==========================================================================
230 230 */
231 231
232 232 /*
233 233 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
234 234 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
235 235 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
236 236 * excess / transient data in-core during a crashdump.
237 237 */
238 238 void *
239 239 zio_buf_alloc(size_t size)
240 240 {
241 241 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
242 242
243 243 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
244 244
245 245 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
246 246 }
247 247
248 248 /*
249 249 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
250 250 * crashdump if the kernel panics. This exists so that we will limit the amount
251 251 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
252 252 * of kernel heap dumped to disk when the kernel panics)
253 253 */
254 254 void *
255 255 zio_data_buf_alloc(size_t size)
256 256 {
257 257 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
258 258
259 259 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
260 260
261 261 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
262 262 }
263 263
264 264 void
265 265 zio_buf_free(void *buf, size_t size)
266 266 {
267 267 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
268 268
269 269 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
270 270
271 271 kmem_cache_free(zio_buf_cache[c], buf);
272 272 }
273 273
274 274 void
275 275 zio_data_buf_free(void *buf, size_t size)
276 276 {
277 277 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
278 278
279 279 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
280 280
281 281 kmem_cache_free(zio_data_buf_cache[c], buf);
282 282 }
283 283
284 284 /*
285 285 * ==========================================================================
286 286 * Push and pop I/O transform buffers
287 287 * ==========================================================================
288 288 */
289 289 static void
290 290 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
291 291 zio_transform_func_t *transform)
292 292 {
293 293 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
294 294
295 295 zt->zt_orig_data = zio->io_data;
296 296 zt->zt_orig_size = zio->io_size;
297 297 zt->zt_bufsize = bufsize;
298 298 zt->zt_transform = transform;
299 299
300 300 zt->zt_next = zio->io_transform_stack;
301 301 zio->io_transform_stack = zt;
302 302
303 303 zio->io_data = data;
304 304 zio->io_size = size;
305 305 }
306 306
307 307 static void
308 308 zio_pop_transforms(zio_t *zio)
309 309 {
310 310 zio_transform_t *zt;
311 311
312 312 while ((zt = zio->io_transform_stack) != NULL) {
313 313 if (zt->zt_transform != NULL)
314 314 zt->zt_transform(zio,
315 315 zt->zt_orig_data, zt->zt_orig_size);
316 316
317 317 if (zt->zt_bufsize != 0)
318 318 zio_buf_free(zio->io_data, zt->zt_bufsize);
319 319
320 320 zio->io_data = zt->zt_orig_data;
321 321 zio->io_size = zt->zt_orig_size;
322 322 zio->io_transform_stack = zt->zt_next;
323 323
324 324 kmem_free(zt, sizeof (zio_transform_t));
325 325 }
326 326 }
327 327
328 328 /*
329 329 * ==========================================================================
330 330 * I/O transform callbacks for subblocks and decompression
331 331 * ==========================================================================
332 332 */
333 333 static void
334 334 zio_subblock(zio_t *zio, void *data, uint64_t size)
335 335 {
336 336 ASSERT(zio->io_size > size);
337 337
338 338 if (zio->io_type == ZIO_TYPE_READ)
339 339 bcopy(zio->io_data, data, size);
340 340 }
341 341
342 342 static void
343 343 zio_decompress(zio_t *zio, void *data, uint64_t size)
344 344 {
345 345 if (zio->io_error == 0 &&
346 346 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
347 347 zio->io_data, data, zio->io_size, size) != 0)
348 348 zio->io_error = SET_ERROR(EIO);
349 349 }
350 350
351 351 /*
352 352 * ==========================================================================
353 353 * I/O parent/child relationships and pipeline interlocks
354 354 * ==========================================================================
355 355 */
356 356 /*
357 357 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
358 358 * continue calling these functions until they return NULL.
359 359 * Otherwise, the next caller will pick up the list walk in
360 360 * some indeterminate state. (Otherwise every caller would
361 361 * have to pass in a cookie to keep the state represented by
362 362 * io_walk_link, which gets annoying.)
363 363 */
364 364 zio_t *
365 365 zio_walk_parents(zio_t *cio)
366 366 {
367 367 zio_link_t *zl = cio->io_walk_link;
368 368 list_t *pl = &cio->io_parent_list;
369 369
370 370 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
371 371 cio->io_walk_link = zl;
372 372
373 373 if (zl == NULL)
374 374 return (NULL);
375 375
376 376 ASSERT(zl->zl_child == cio);
377 377 return (zl->zl_parent);
378 378 }
379 379
380 380 zio_t *
381 381 zio_walk_children(zio_t *pio)
382 382 {
383 383 zio_link_t *zl = pio->io_walk_link;
384 384 list_t *cl = &pio->io_child_list;
385 385
386 386 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
387 387 pio->io_walk_link = zl;
388 388
389 389 if (zl == NULL)
390 390 return (NULL);
391 391
392 392 ASSERT(zl->zl_parent == pio);
393 393 return (zl->zl_child);
394 394 }
395 395
396 396 zio_t *
397 397 zio_unique_parent(zio_t *cio)
398 398 {
399 399 zio_t *pio = zio_walk_parents(cio);
400 400
401 401 VERIFY(zio_walk_parents(cio) == NULL);
402 402 return (pio);
403 403 }
404 404
405 405 void
406 406 zio_add_child(zio_t *pio, zio_t *cio)
407 407 {
408 408 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
409 409
410 410 /*
411 411 * Logical I/Os can have logical, gang, or vdev children.
412 412 * Gang I/Os can have gang or vdev children.
413 413 * Vdev I/Os can only have vdev children.
414 414 * The following ASSERT captures all of these constraints.
415 415 */
416 416 ASSERT(cio->io_child_type <= pio->io_child_type);
417 417
418 418 zl->zl_parent = pio;
419 419 zl->zl_child = cio;
420 420
421 421 mutex_enter(&cio->io_lock);
422 422 mutex_enter(&pio->io_lock);
423 423
424 424 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
425 425
426 426 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
427 427 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
428 428
429 429 list_insert_head(&pio->io_child_list, zl);
430 430 list_insert_head(&cio->io_parent_list, zl);
431 431
432 432 pio->io_child_count++;
433 433 cio->io_parent_count++;
434 434
435 435 mutex_exit(&pio->io_lock);
436 436 mutex_exit(&cio->io_lock);
437 437 }
438 438
439 439 static void
440 440 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
441 441 {
442 442 ASSERT(zl->zl_parent == pio);
443 443 ASSERT(zl->zl_child == cio);
444 444
445 445 mutex_enter(&cio->io_lock);
446 446 mutex_enter(&pio->io_lock);
447 447
448 448 list_remove(&pio->io_child_list, zl);
449 449 list_remove(&cio->io_parent_list, zl);
450 450
451 451 pio->io_child_count--;
452 452 cio->io_parent_count--;
453 453
454 454 mutex_exit(&pio->io_lock);
455 455 mutex_exit(&cio->io_lock);
456 456
457 457 kmem_cache_free(zio_link_cache, zl);
458 458 }
459 459
460 460 static boolean_t
461 461 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
462 462 {
463 463 uint64_t *countp = &zio->io_children[child][wait];
464 464 boolean_t waiting = B_FALSE;
465 465
466 466 mutex_enter(&zio->io_lock);
467 467 ASSERT(zio->io_stall == NULL);
468 468 if (*countp != 0) {
469 469 zio->io_stage >>= 1;
470 470 zio->io_stall = countp;
471 471 waiting = B_TRUE;
472 472 }
473 473 mutex_exit(&zio->io_lock);
474 474
475 475 return (waiting);
476 476 }
477 477
478 478 static void
479 479 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
480 480 {
481 481 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
482 482 int *errorp = &pio->io_child_error[zio->io_child_type];
483 483
484 484 mutex_enter(&pio->io_lock);
485 485 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
486 486 *errorp = zio_worst_error(*errorp, zio->io_error);
487 487 pio->io_reexecute |= zio->io_reexecute;
488 488 ASSERT3U(*countp, >, 0);
489 489 if (--*countp == 0 && pio->io_stall == countp) {
490 490 pio->io_stall = NULL;
491 491 mutex_exit(&pio->io_lock);
492 492 zio_execute(pio);
493 493 } else {
494 494 mutex_exit(&pio->io_lock);
495 495 }
496 496 }
497 497
498 498 static void
499 499 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
500 500 {
501 501 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
502 502 zio->io_error = zio->io_child_error[c];
503 503 }
504 504
505 505 /*
506 506 * ==========================================================================
507 507 * Create the various types of I/O (read, write, free, etc)
508 508 * ==========================================================================
509 509 */
510 510 static zio_t *
511 511 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
512 512 void *data, uint64_t size, zio_done_func_t *done, void *private,
513 513 zio_type_t type, int priority, enum zio_flag flags,
514 514 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
515 515 enum zio_stage stage, enum zio_stage pipeline)
516 516 {
517 517 zio_t *zio;
518 518
519 519 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
520 520 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
521 521 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
522 522
523 523 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
524 524 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
525 525 ASSERT(vd || stage == ZIO_STAGE_OPEN);
526 526
527 527 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
528 528 bzero(zio, sizeof (zio_t));
529 529
530 530 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
531 531 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
532 532
533 533 list_create(&zio->io_parent_list, sizeof (zio_link_t),
534 534 offsetof(zio_link_t, zl_parent_node));
535 535 list_create(&zio->io_child_list, sizeof (zio_link_t),
536 536 offsetof(zio_link_t, zl_child_node));
537 537
538 538 if (vd != NULL)
539 539 zio->io_child_type = ZIO_CHILD_VDEV;
540 540 else if (flags & ZIO_FLAG_GANG_CHILD)
541 541 zio->io_child_type = ZIO_CHILD_GANG;
542 542 else if (flags & ZIO_FLAG_DDT_CHILD)
543 543 zio->io_child_type = ZIO_CHILD_DDT;
544 544 else
545 545 zio->io_child_type = ZIO_CHILD_LOGICAL;
546 546
547 547 if (bp != NULL) {
548 548 zio->io_bp = (blkptr_t *)bp;
549 549 zio->io_bp_copy = *bp;
550 550 zio->io_bp_orig = *bp;
551 551 if (type != ZIO_TYPE_WRITE ||
552 552 zio->io_child_type == ZIO_CHILD_DDT)
553 553 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
554 554 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
555 555 zio->io_logical = zio;
556 556 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
557 557 pipeline |= ZIO_GANG_STAGES;
558 558 }
559 559
560 560 zio->io_spa = spa;
561 561 zio->io_txg = txg;
562 562 zio->io_done = done;
563 563 zio->io_private = private;
564 564 zio->io_type = type;
565 565 zio->io_priority = priority;
566 566 zio->io_vd = vd;
567 567 zio->io_offset = offset;
568 568 zio->io_orig_data = zio->io_data = data;
569 569 zio->io_orig_size = zio->io_size = size;
570 570 zio->io_orig_flags = zio->io_flags = flags;
571 571 zio->io_orig_stage = zio->io_stage = stage;
572 572 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
573 573
574 574 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
575 575 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
576 576
577 577 if (zb != NULL)
578 578 zio->io_bookmark = *zb;
579 579
580 580 if (pio != NULL) {
581 581 if (zio->io_logical == NULL)
582 582 zio->io_logical = pio->io_logical;
583 583 if (zio->io_child_type == ZIO_CHILD_GANG)
584 584 zio->io_gang_leader = pio->io_gang_leader;
585 585 zio_add_child(pio, zio);
586 586 }
587 587
588 588 return (zio);
589 589 }
590 590
591 591 static void
592 592 zio_destroy(zio_t *zio)
593 593 {
594 594 list_destroy(&zio->io_parent_list);
595 595 list_destroy(&zio->io_child_list);
596 596 mutex_destroy(&zio->io_lock);
597 597 cv_destroy(&zio->io_cv);
598 598 kmem_cache_free(zio_cache, zio);
599 599 }
600 600
601 601 zio_t *
602 602 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
603 603 void *private, enum zio_flag flags)
604 604 {
605 605 zio_t *zio;
606 606
607 607 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
608 608 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
609 609 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
610 610
611 611 return (zio);
612 612 }
613 613
614 614 zio_t *
615 615 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
616 616 {
617 617 return (zio_null(NULL, spa, NULL, done, private, flags));
618 618 }
619 619
620 620 zio_t *
621 621 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
622 622 void *data, uint64_t size, zio_done_func_t *done, void *private,
623 623 int priority, enum zio_flag flags, const zbookmark_t *zb)
624 624 {
625 625 zio_t *zio;
626 626
627 627 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
628 628 data, size, done, private,
629 629 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
630 630 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
631 631 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
632 632
633 633 return (zio);
634 634 }
635 635
636 636 zio_t *
637 637 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
638 638 void *data, uint64_t size, const zio_prop_t *zp,
639 639 zio_done_func_t *ready, zio_done_func_t *done, void *private,
640 640 int priority, enum zio_flag flags, const zbookmark_t *zb)
641 641 {
642 642 zio_t *zio;
643 643
644 644 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
645 645 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
646 646 zp->zp_compress >= ZIO_COMPRESS_OFF &&
647 647 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
648 648 DMU_OT_IS_VALID(zp->zp_type) &&
649 649 zp->zp_level < 32 &&
650 650 zp->zp_copies > 0 &&
651 651 zp->zp_copies <= spa_max_replication(spa));
652 652
653 653 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
654 654 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
655 655 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
656 656 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
657 657
658 658 zio->io_ready = ready;
659 659 zio->io_prop = *zp;
660 660
661 661 return (zio);
662 662 }
663 663
664 664 zio_t *
665 665 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
666 666 uint64_t size, zio_done_func_t *done, void *private, int priority,
667 667 enum zio_flag flags, zbookmark_t *zb)
668 668 {
669 669 zio_t *zio;
670 670
671 671 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
672 672 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
673 673 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
674 674
675 675 return (zio);
676 676 }
677 677
678 678 void
679 679 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
680 680 {
681 681 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
682 682 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
683 683 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
684 684 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
685 685
686 686 /*
687 687 * We must reset the io_prop to match the values that existed
688 688 * when the bp was first written by dmu_sync() keeping in mind
689 689 * that nopwrite and dedup are mutually exclusive.
690 690 */
691 691 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
692 692 zio->io_prop.zp_nopwrite = nopwrite;
693 693 zio->io_prop.zp_copies = copies;
694 694 zio->io_bp_override = bp;
695 695 }
696 696
697 697 void
698 698 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
699 699 {
700 700 metaslab_check_free(spa, bp);
701 701 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
702 702 }
703 703
704 704 zio_t *
705 705 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
706 706 enum zio_flag flags)
707 707 {
708 708 zio_t *zio;
709 709
710 710 dprintf_bp(bp, "freeing in txg %llu, pass %u",
711 711 (longlong_t)txg, spa->spa_sync_pass);
712 712
713 713 ASSERT(!BP_IS_HOLE(bp));
714 714 ASSERT(spa_syncing_txg(spa) == txg);
715 715 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
716 716
717 717 metaslab_check_free(spa, bp);
718 718
719 719 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
720 720 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
721 721 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
722 722
723 723 return (zio);
724 724 }
725 725
726 726 zio_t *
727 727 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
728 728 zio_done_func_t *done, void *private, enum zio_flag flags)
729 729 {
730 730 zio_t *zio;
731 731
732 732 /*
733 733 * A claim is an allocation of a specific block. Claims are needed
734 734 * to support immediate writes in the intent log. The issue is that
735 735 * immediate writes contain committed data, but in a txg that was
736 736 * *not* committed. Upon opening the pool after an unclean shutdown,
737 737 * the intent log claims all blocks that contain immediate write data
738 738 * so that the SPA knows they're in use.
739 739 *
740 740 * All claims *must* be resolved in the first txg -- before the SPA
741 741 * starts allocating blocks -- so that nothing is allocated twice.
742 742 * If txg == 0 we just verify that the block is claimable.
743 743 */
744 744 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
745 745 ASSERT(txg == spa_first_txg(spa) || txg == 0);
746 746 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
747 747
748 748 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
749 749 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
750 750 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
751 751
752 752 return (zio);
753 753 }
754 754
755 755 zio_t *
756 756 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
757 757 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
758 758 {
759 759 zio_t *zio;
760 760 int c;
761 761
762 762 if (vd->vdev_children == 0) {
763 763 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
764 764 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
765 765 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
766 766
767 767 zio->io_cmd = cmd;
768 768 } else {
769 769 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
770 770
771 771 for (c = 0; c < vd->vdev_children; c++)
772 772 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
773 773 done, private, priority, flags));
774 774 }
775 775
776 776 return (zio);
777 777 }
778 778
779 779 zio_t *
780 780 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
781 781 void *data, int checksum, zio_done_func_t *done, void *private,
782 782 int priority, enum zio_flag flags, boolean_t labels)
783 783 {
784 784 zio_t *zio;
785 785
786 786 ASSERT(vd->vdev_children == 0);
787 787 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
788 788 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
789 789 ASSERT3U(offset + size, <=, vd->vdev_psize);
790 790
791 791 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
792 792 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
793 793 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
794 794
795 795 zio->io_prop.zp_checksum = checksum;
796 796
797 797 return (zio);
798 798 }
799 799
800 800 zio_t *
801 801 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
802 802 void *data, int checksum, zio_done_func_t *done, void *private,
803 803 int priority, enum zio_flag flags, boolean_t labels)
804 804 {
805 805 zio_t *zio;
806 806
807 807 ASSERT(vd->vdev_children == 0);
808 808 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
809 809 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
810 810 ASSERT3U(offset + size, <=, vd->vdev_psize);
811 811
812 812 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
813 813 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
814 814 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
815 815
816 816 zio->io_prop.zp_checksum = checksum;
817 817
818 818 if (zio_checksum_table[checksum].ci_eck) {
819 819 /*
820 820 * zec checksums are necessarily destructive -- they modify
821 821 * the end of the write buffer to hold the verifier/checksum.
822 822 * Therefore, we must make a local copy in case the data is
823 823 * being written to multiple places in parallel.
824 824 */
825 825 void *wbuf = zio_buf_alloc(size);
826 826 bcopy(data, wbuf, size);
827 827 zio_push_transform(zio, wbuf, size, size, NULL);
828 828 }
829 829
830 830 return (zio);
831 831 }
832 832
833 833 /*
834 834 * Create a child I/O to do some work for us.
835 835 */
836 836 zio_t *
837 837 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
838 838 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
839 839 zio_done_func_t *done, void *private)
840 840 {
841 841 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
842 842 zio_t *zio;
843 843
844 844 ASSERT(vd->vdev_parent ==
845 845 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
846 846
847 847 if (type == ZIO_TYPE_READ && bp != NULL) {
848 848 /*
849 849 * If we have the bp, then the child should perform the
850 850 * checksum and the parent need not. This pushes error
851 851 * detection as close to the leaves as possible and
852 852 * eliminates redundant checksums in the interior nodes.
853 853 */
854 854 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
855 855 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
856 856 }
857 857
858 858 if (vd->vdev_children == 0)
859 859 offset += VDEV_LABEL_START_SIZE;
860 860
861 861 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
862 862
863 863 /*
864 864 * If we've decided to do a repair, the write is not speculative --
865 865 * even if the original read was.
866 866 */
867 867 if (flags & ZIO_FLAG_IO_REPAIR)
868 868 flags &= ~ZIO_FLAG_SPECULATIVE;
869 869
870 870 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
871 871 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
872 872 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
873 873
874 874 return (zio);
875 875 }
876 876
877 877 zio_t *
878 878 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
879 879 int type, int priority, enum zio_flag flags,
880 880 zio_done_func_t *done, void *private)
881 881 {
882 882 zio_t *zio;
883 883
884 884 ASSERT(vd->vdev_ops->vdev_op_leaf);
885 885
886 886 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
887 887 data, size, done, private, type, priority,
888 888 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
889 889 vd, offset, NULL,
890 890 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
891 891
892 892 return (zio);
893 893 }
894 894
895 895 void
896 896 zio_flush(zio_t *zio, vdev_t *vd)
897 897 {
898 898 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
899 899 NULL, NULL, ZIO_PRIORITY_NOW,
900 900 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
901 901 }
902 902
903 903 void
904 904 zio_shrink(zio_t *zio, uint64_t size)
905 905 {
906 906 ASSERT(zio->io_executor == NULL);
907 907 ASSERT(zio->io_orig_size == zio->io_size);
908 908 ASSERT(size <= zio->io_size);
909 909
910 910 /*
911 911 * We don't shrink for raidz because of problems with the
912 912 * reconstruction when reading back less than the block size.
913 913 * Note, BP_IS_RAIDZ() assumes no compression.
914 914 */
915 915 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
916 916 if (!BP_IS_RAIDZ(zio->io_bp))
917 917 zio->io_orig_size = zio->io_size = size;
918 918 }
919 919
920 920 /*
921 921 * ==========================================================================
922 922 * Prepare to read and write logical blocks
923 923 * ==========================================================================
924 924 */
925 925
926 926 static int
927 927 zio_read_bp_init(zio_t *zio)
928 928 {
929 929 blkptr_t *bp = zio->io_bp;
930 930
931 931 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
932 932 zio->io_child_type == ZIO_CHILD_LOGICAL &&
933 933 !(zio->io_flags & ZIO_FLAG_RAW)) {
934 934 uint64_t psize = BP_GET_PSIZE(bp);
935 935 void *cbuf = zio_buf_alloc(psize);
936 936
937 937 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
938 938 }
939 939
940 940 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
941 941 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
942 942
943 943 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
944 944 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
945 945
946 946 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
947 947 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
948 948
949 949 return (ZIO_PIPELINE_CONTINUE);
950 950 }
951 951
952 952 static int
953 953 zio_write_bp_init(zio_t *zio)
954 954 {
955 955 spa_t *spa = zio->io_spa;
956 956 zio_prop_t *zp = &zio->io_prop;
957 957 enum zio_compress compress = zp->zp_compress;
958 958 blkptr_t *bp = zio->io_bp;
959 959 uint64_t lsize = zio->io_size;
960 960 uint64_t psize = lsize;
961 961 int pass = 1;
962 962
963 963 /*
964 964 * If our children haven't all reached the ready stage,
965 965 * wait for them and then repeat this pipeline stage.
966 966 */
967 967 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
968 968 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
969 969 return (ZIO_PIPELINE_STOP);
970 970
971 971 if (!IO_IS_ALLOCATING(zio))
972 972 return (ZIO_PIPELINE_CONTINUE);
973 973
974 974 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
975 975
976 976 if (zio->io_bp_override) {
977 977 ASSERT(bp->blk_birth != zio->io_txg);
978 978 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
979 979
980 980 *bp = *zio->io_bp_override;
981 981 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
982 982
983 983 /*
984 984 * If we've been overridden and nopwrite is set then
985 985 * set the flag accordingly to indicate that a nopwrite
986 986 * has already occurred.
987 987 */
988 988 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
989 989 ASSERT(!zp->zp_dedup);
990 990 zio->io_flags |= ZIO_FLAG_NOPWRITE;
991 991 return (ZIO_PIPELINE_CONTINUE);
992 992 }
993 993
994 994 ASSERT(!zp->zp_nopwrite);
995 995
996 996 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
997 997 return (ZIO_PIPELINE_CONTINUE);
998 998
999 999 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1000 1000 zp->zp_dedup_verify);
1001 1001
1002 1002 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1003 1003 BP_SET_DEDUP(bp, 1);
1004 1004 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1005 1005 return (ZIO_PIPELINE_CONTINUE);
1006 1006 }
1007 1007 zio->io_bp_override = NULL;
1008 1008 BP_ZERO(bp);
1009 1009 }
1010 1010
1011 1011 if (bp->blk_birth == zio->io_txg) {
1012 1012 /*
1013 1013 * We're rewriting an existing block, which means we're
1014 1014 * working on behalf of spa_sync(). For spa_sync() to
1015 1015 * converge, it must eventually be the case that we don't
1016 1016 * have to allocate new blocks. But compression changes
1017 1017 * the blocksize, which forces a reallocate, and makes
1018 1018 * convergence take longer. Therefore, after the first
1019 1019 * few passes, stop compressing to ensure convergence.
1020 1020 */
1021 1021 pass = spa_sync_pass(spa);
1022 1022
1023 1023 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1024 1024 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1025 1025 ASSERT(!BP_GET_DEDUP(bp));
1026 1026
1027 1027 if (pass >= zfs_sync_pass_dont_compress)
1028 1028 compress = ZIO_COMPRESS_OFF;
1029 1029
1030 1030 /* Make sure someone doesn't change their mind on overwrites */
1031 1031 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1032 1032 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1033 1033 }
1034 1034
1035 1035 if (compress != ZIO_COMPRESS_OFF) {
1036 1036 void *cbuf = zio_buf_alloc(lsize);
1037 1037 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1038 1038 if (psize == 0 || psize == lsize) {
1039 1039 compress = ZIO_COMPRESS_OFF;
1040 1040 zio_buf_free(cbuf, lsize);
1041 1041 } else {
1042 1042 ASSERT(psize < lsize);
1043 1043 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1044 1044 }
1045 1045 }
1046 1046
1047 1047 /*
1048 1048 * The final pass of spa_sync() must be all rewrites, but the first
1049 1049 * few passes offer a trade-off: allocating blocks defers convergence,
1050 1050 * but newly allocated blocks are sequential, so they can be written
1051 1051 * to disk faster. Therefore, we allow the first few passes of
1052 1052 * spa_sync() to allocate new blocks, but force rewrites after that.
1053 1053 * There should only be a handful of blocks after pass 1 in any case.
1054 1054 */
1055 1055 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1056 1056 pass >= zfs_sync_pass_rewrite) {
1057 1057 ASSERT(psize != 0);
1058 1058 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1059 1059 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1060 1060 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1061 1061 } else {
1062 1062 BP_ZERO(bp);
1063 1063 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1064 1064 }
1065 1065
1066 1066 if (psize == 0) {
1067 1067 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1068 1068 } else {
1069 1069 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1070 1070 BP_SET_LSIZE(bp, lsize);
1071 1071 BP_SET_PSIZE(bp, psize);
1072 1072 BP_SET_COMPRESS(bp, compress);
1073 1073 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1074 1074 BP_SET_TYPE(bp, zp->zp_type);
1075 1075 BP_SET_LEVEL(bp, zp->zp_level);
1076 1076 BP_SET_DEDUP(bp, zp->zp_dedup);
1077 1077 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1078 1078 if (zp->zp_dedup) {
1079 1079 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1080 1080 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1081 1081 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1082 1082 }
1083 1083 if (zp->zp_nopwrite) {
1084 1084 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1085 1085 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1086 1086 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1087 1087 }
1088 1088 }
1089 1089
1090 1090 return (ZIO_PIPELINE_CONTINUE);
1091 1091 }
1092 1092
1093 1093 static int
1094 1094 zio_free_bp_init(zio_t *zio)
1095 1095 {
1096 1096 blkptr_t *bp = zio->io_bp;
1097 1097
1098 1098 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1099 1099 if (BP_GET_DEDUP(bp))
1100 1100 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1101 1101 }
1102 1102
1103 1103 return (ZIO_PIPELINE_CONTINUE);
1104 1104 }
1105 1105
1106 1106 /*
1107 1107 * ==========================================================================
1108 1108 * Execute the I/O pipeline
1109 1109 * ==========================================================================
1110 1110 */
1111 1111
1112 1112 static void
1113 1113 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1114 1114 {
1115 1115 spa_t *spa = zio->io_spa;
1116 1116 zio_type_t t = zio->io_type;
1117 1117 int flags = (cutinline ? TQ_FRONT : 0);
1118 1118
1119 1119 /*
1120 1120 * If we're a config writer or a probe, the normal issue and
1121 1121 * interrupt threads may all be blocked waiting for the config lock.
1122 1122 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1123 1123 */
1124 1124 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1125 1125 t = ZIO_TYPE_NULL;
1126 1126
1127 1127 /*
1128 1128 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1129 1129 */
1130 1130 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1131 1131 t = ZIO_TYPE_NULL;
1132 1132
1133 1133 /*
1134 1134 * If this is a high priority I/O, then use the high priority taskq if
1135 1135 * available.
1136 1136 */
1137 1137 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1138 1138 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1139 1139 q++;
1140 1140
1141 1141 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1142 1142
1143 1143 /*
1144 1144 * NB: We are assuming that the zio can only be dispatched
1145 1145 * to a single taskq at a time. It would be a grievous error
1146 1146 * to dispatch the zio to another taskq at the same time.
1147 1147 */
1148 1148 ASSERT(zio->io_tqent.tqent_next == NULL);
1149 1149 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1150 1150 flags, &zio->io_tqent);
1151 1151 }
1152 1152
1153 1153 static boolean_t
1154 1154 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1155 1155 {
1156 1156 kthread_t *executor = zio->io_executor;
1157 1157 spa_t *spa = zio->io_spa;
1158 1158
1159 1159 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1160 1160 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1161 1161 uint_t i;
1162 1162 for (i = 0; i < tqs->stqs_count; i++) {
1163 1163 if (taskq_member(tqs->stqs_taskq[i], executor))
1164 1164 return (B_TRUE);
1165 1165 }
1166 1166 }
1167 1167
1168 1168 return (B_FALSE);
1169 1169 }
1170 1170
1171 1171 static int
1172 1172 zio_issue_async(zio_t *zio)
1173 1173 {
1174 1174 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1175 1175
1176 1176 return (ZIO_PIPELINE_STOP);
|
↓ open down ↓ |
1176 lines elided |
↑ open up ↑ |
1177 1177 }
1178 1178
1179 1179 void
1180 1180 zio_interrupt(zio_t *zio)
1181 1181 {
1182 1182 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1183 1183 }
1184 1184
1185 1185 /*
1186 1186 * Execute the I/O pipeline until one of the following occurs:
1187 - * (1) the I/O completes; (2) the pipeline stalls waiting for
1188 - * dependent child I/Os; (3) the I/O issues, so we're waiting
1189 - * for an I/O completion interrupt; (4) the I/O is delegated by
1190 - * vdev-level caching or aggregation; (5) the I/O is deferred
1191 - * due to vdev-level queueing; (6) the I/O is handed off to
1192 - * another thread. In all cases, the pipeline stops whenever
1193 - * there's no CPU work; it never burns a thread in cv_wait().
1187 + *
1188 + * (1) the I/O completes
1189 + * (2) the pipeline stalls waiting for dependent child I/Os
1190 + * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1191 + * (4) the I/O is delegated by vdev-level caching or aggregation
1192 + * (5) the I/O is deferred due to vdev-level queueing
1193 + * (6) the I/O is handed off to another thread.
1194 + *
1195 + * In all cases, the pipeline stops whenever there's no CPU work; it never
1196 + * burns a thread in cv_wait().
1194 1197 *
1195 1198 * There's no locking on io_stage because there's no legitimate way
1196 1199 * for multiple threads to be attempting to process the same I/O.
1197 1200 */
1198 1201 static zio_pipe_stage_t *zio_pipeline[];
1199 1202
1200 1203 void
1201 1204 zio_execute(zio_t *zio)
1202 1205 {
1203 1206 zio->io_executor = curthread;
1204 1207
1205 1208 while (zio->io_stage < ZIO_STAGE_DONE) {
1206 1209 enum zio_stage pipeline = zio->io_pipeline;
1207 1210 enum zio_stage stage = zio->io_stage;
1208 1211 int rv;
1209 1212
1210 1213 ASSERT(!MUTEX_HELD(&zio->io_lock));
1211 1214 ASSERT(ISP2(stage));
1212 1215 ASSERT(zio->io_stall == NULL);
1213 1216
1214 1217 do {
1215 1218 stage <<= 1;
1216 1219 } while ((stage & pipeline) == 0);
1217 1220
1218 1221 ASSERT(stage <= ZIO_STAGE_DONE);
1219 1222
1220 1223 /*
1221 1224 * If we are in interrupt context and this pipeline stage
1222 1225 * will grab a config lock that is held across I/O,
1223 1226 * or may wait for an I/O that needs an interrupt thread
1224 1227 * to complete, issue async to avoid deadlock.
1225 1228 *
1226 1229 * For VDEV_IO_START, we cut in line so that the io will
1227 1230 * be sent to disk promptly.
1228 1231 */
1229 1232 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1230 1233 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1231 1234 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1232 1235 zio_requeue_io_start_cut_in_line : B_FALSE;
1233 1236 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1234 1237 return;
1235 1238 }
1236 1239
1237 1240 zio->io_stage = stage;
1238 1241 rv = zio_pipeline[highbit(stage) - 1](zio);
1239 1242
1240 1243 if (rv == ZIO_PIPELINE_STOP)
1241 1244 return;
1242 1245
1243 1246 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1244 1247 }
1245 1248 }
1246 1249
1247 1250 /*
1248 1251 * ==========================================================================
1249 1252 * Initiate I/O, either sync or async
1250 1253 * ==========================================================================
1251 1254 */
1252 1255 int
1253 1256 zio_wait(zio_t *zio)
1254 1257 {
1255 1258 int error;
1256 1259
1257 1260 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1258 1261 ASSERT(zio->io_executor == NULL);
1259 1262
1260 1263 zio->io_waiter = curthread;
1261 1264
1262 1265 zio_execute(zio);
1263 1266
1264 1267 mutex_enter(&zio->io_lock);
1265 1268 while (zio->io_executor != NULL)
1266 1269 cv_wait(&zio->io_cv, &zio->io_lock);
1267 1270 mutex_exit(&zio->io_lock);
1268 1271
1269 1272 error = zio->io_error;
1270 1273 zio_destroy(zio);
1271 1274
1272 1275 return (error);
1273 1276 }
1274 1277
1275 1278 void
1276 1279 zio_nowait(zio_t *zio)
1277 1280 {
1278 1281 ASSERT(zio->io_executor == NULL);
1279 1282
1280 1283 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1281 1284 zio_unique_parent(zio) == NULL) {
1282 1285 /*
1283 1286 * This is a logical async I/O with no parent to wait for it.
1284 1287 * We add it to the spa_async_root_zio "Godfather" I/O which
1285 1288 * will ensure they complete prior to unloading the pool.
1286 1289 */
1287 1290 spa_t *spa = zio->io_spa;
1288 1291
1289 1292 zio_add_child(spa->spa_async_zio_root, zio);
1290 1293 }
1291 1294
1292 1295 zio_execute(zio);
1293 1296 }
1294 1297
1295 1298 /*
1296 1299 * ==========================================================================
1297 1300 * Reexecute or suspend/resume failed I/O
1298 1301 * ==========================================================================
1299 1302 */
1300 1303
1301 1304 static void
1302 1305 zio_reexecute(zio_t *pio)
1303 1306 {
1304 1307 zio_t *cio, *cio_next;
1305 1308
1306 1309 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1307 1310 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1308 1311 ASSERT(pio->io_gang_leader == NULL);
1309 1312 ASSERT(pio->io_gang_tree == NULL);
1310 1313
1311 1314 pio->io_flags = pio->io_orig_flags;
1312 1315 pio->io_stage = pio->io_orig_stage;
1313 1316 pio->io_pipeline = pio->io_orig_pipeline;
1314 1317 pio->io_reexecute = 0;
1315 1318 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1316 1319 pio->io_error = 0;
1317 1320 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1318 1321 pio->io_state[w] = 0;
1319 1322 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1320 1323 pio->io_child_error[c] = 0;
1321 1324
1322 1325 if (IO_IS_ALLOCATING(pio))
1323 1326 BP_ZERO(pio->io_bp);
1324 1327
1325 1328 /*
1326 1329 * As we reexecute pio's children, new children could be created.
1327 1330 * New children go to the head of pio's io_child_list, however,
1328 1331 * so we will (correctly) not reexecute them. The key is that
1329 1332 * the remainder of pio's io_child_list, from 'cio_next' onward,
1330 1333 * cannot be affected by any side effects of reexecuting 'cio'.
1331 1334 */
1332 1335 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1333 1336 cio_next = zio_walk_children(pio);
1334 1337 mutex_enter(&pio->io_lock);
1335 1338 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1336 1339 pio->io_children[cio->io_child_type][w]++;
1337 1340 mutex_exit(&pio->io_lock);
1338 1341 zio_reexecute(cio);
1339 1342 }
1340 1343
1341 1344 /*
1342 1345 * Now that all children have been reexecuted, execute the parent.
1343 1346 * We don't reexecute "The Godfather" I/O here as it's the
1344 1347 * responsibility of the caller to wait on him.
1345 1348 */
1346 1349 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1347 1350 zio_execute(pio);
1348 1351 }
1349 1352
1350 1353 void
1351 1354 zio_suspend(spa_t *spa, zio_t *zio)
1352 1355 {
1353 1356 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1354 1357 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1355 1358 "failure and the failure mode property for this pool "
1356 1359 "is set to panic.", spa_name(spa));
1357 1360
1358 1361 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1359 1362
1360 1363 mutex_enter(&spa->spa_suspend_lock);
1361 1364
1362 1365 if (spa->spa_suspend_zio_root == NULL)
1363 1366 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1364 1367 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1365 1368 ZIO_FLAG_GODFATHER);
1366 1369
1367 1370 spa->spa_suspended = B_TRUE;
1368 1371
1369 1372 if (zio != NULL) {
1370 1373 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1371 1374 ASSERT(zio != spa->spa_suspend_zio_root);
1372 1375 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1373 1376 ASSERT(zio_unique_parent(zio) == NULL);
1374 1377 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1375 1378 zio_add_child(spa->spa_suspend_zio_root, zio);
1376 1379 }
1377 1380
1378 1381 mutex_exit(&spa->spa_suspend_lock);
1379 1382 }
1380 1383
1381 1384 int
1382 1385 zio_resume(spa_t *spa)
1383 1386 {
1384 1387 zio_t *pio;
1385 1388
1386 1389 /*
1387 1390 * Reexecute all previously suspended i/o.
1388 1391 */
1389 1392 mutex_enter(&spa->spa_suspend_lock);
1390 1393 spa->spa_suspended = B_FALSE;
1391 1394 cv_broadcast(&spa->spa_suspend_cv);
1392 1395 pio = spa->spa_suspend_zio_root;
1393 1396 spa->spa_suspend_zio_root = NULL;
1394 1397 mutex_exit(&spa->spa_suspend_lock);
1395 1398
1396 1399 if (pio == NULL)
1397 1400 return (0);
1398 1401
1399 1402 zio_reexecute(pio);
1400 1403 return (zio_wait(pio));
1401 1404 }
1402 1405
1403 1406 void
1404 1407 zio_resume_wait(spa_t *spa)
1405 1408 {
1406 1409 mutex_enter(&spa->spa_suspend_lock);
1407 1410 while (spa_suspended(spa))
1408 1411 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1409 1412 mutex_exit(&spa->spa_suspend_lock);
1410 1413 }
1411 1414
1412 1415 /*
1413 1416 * ==========================================================================
1414 1417 * Gang blocks.
1415 1418 *
1416 1419 * A gang block is a collection of small blocks that looks to the DMU
1417 1420 * like one large block. When zio_dva_allocate() cannot find a block
1418 1421 * of the requested size, due to either severe fragmentation or the pool
1419 1422 * being nearly full, it calls zio_write_gang_block() to construct the
1420 1423 * block from smaller fragments.
1421 1424 *
1422 1425 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1423 1426 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1424 1427 * an indirect block: it's an array of block pointers. It consumes
1425 1428 * only one sector and hence is allocatable regardless of fragmentation.
1426 1429 * The gang header's bps point to its gang members, which hold the data.
1427 1430 *
1428 1431 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1429 1432 * as the verifier to ensure uniqueness of the SHA256 checksum.
1430 1433 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1431 1434 * not the gang header. This ensures that data block signatures (needed for
1432 1435 * deduplication) are independent of how the block is physically stored.
1433 1436 *
1434 1437 * Gang blocks can be nested: a gang member may itself be a gang block.
1435 1438 * Thus every gang block is a tree in which root and all interior nodes are
1436 1439 * gang headers, and the leaves are normal blocks that contain user data.
1437 1440 * The root of the gang tree is called the gang leader.
1438 1441 *
1439 1442 * To perform any operation (read, rewrite, free, claim) on a gang block,
1440 1443 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1441 1444 * in the io_gang_tree field of the original logical i/o by recursively
1442 1445 * reading the gang leader and all gang headers below it. This yields
1443 1446 * an in-core tree containing the contents of every gang header and the
1444 1447 * bps for every constituent of the gang block.
1445 1448 *
1446 1449 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1447 1450 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1448 1451 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1449 1452 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1450 1453 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1451 1454 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1452 1455 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1453 1456 * of the gang header plus zio_checksum_compute() of the data to update the
1454 1457 * gang header's blk_cksum as described above.
1455 1458 *
1456 1459 * The two-phase assemble/issue model solves the problem of partial failure --
1457 1460 * what if you'd freed part of a gang block but then couldn't read the
1458 1461 * gang header for another part? Assembling the entire gang tree first
1459 1462 * ensures that all the necessary gang header I/O has succeeded before
1460 1463 * starting the actual work of free, claim, or write. Once the gang tree
1461 1464 * is assembled, free and claim are in-memory operations that cannot fail.
1462 1465 *
1463 1466 * In the event that a gang write fails, zio_dva_unallocate() walks the
1464 1467 * gang tree to immediately free (i.e. insert back into the space map)
1465 1468 * everything we've allocated. This ensures that we don't get ENOSPC
1466 1469 * errors during repeated suspend/resume cycles due to a flaky device.
1467 1470 *
1468 1471 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1469 1472 * the gang tree, we won't modify the block, so we can safely defer the free
1470 1473 * (knowing that the block is still intact). If we *can* assemble the gang
1471 1474 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1472 1475 * each constituent bp and we can allocate a new block on the next sync pass.
1473 1476 *
1474 1477 * In all cases, the gang tree allows complete recovery from partial failure.
1475 1478 * ==========================================================================
1476 1479 */
1477 1480
1478 1481 static zio_t *
1479 1482 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1480 1483 {
1481 1484 if (gn != NULL)
1482 1485 return (pio);
1483 1486
1484 1487 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1485 1488 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1486 1489 &pio->io_bookmark));
1487 1490 }
1488 1491
1489 1492 zio_t *
1490 1493 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1491 1494 {
1492 1495 zio_t *zio;
1493 1496
1494 1497 if (gn != NULL) {
1495 1498 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1496 1499 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1497 1500 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1498 1501 /*
1499 1502 * As we rewrite each gang header, the pipeline will compute
1500 1503 * a new gang block header checksum for it; but no one will
1501 1504 * compute a new data checksum, so we do that here. The one
1502 1505 * exception is the gang leader: the pipeline already computed
1503 1506 * its data checksum because that stage precedes gang assembly.
1504 1507 * (Presently, nothing actually uses interior data checksums;
1505 1508 * this is just good hygiene.)
1506 1509 */
1507 1510 if (gn != pio->io_gang_leader->io_gang_tree) {
1508 1511 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1509 1512 data, BP_GET_PSIZE(bp));
1510 1513 }
1511 1514 /*
1512 1515 * If we are here to damage data for testing purposes,
1513 1516 * leave the GBH alone so that we can detect the damage.
1514 1517 */
1515 1518 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1516 1519 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1517 1520 } else {
1518 1521 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1519 1522 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1520 1523 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1521 1524 }
1522 1525
1523 1526 return (zio);
1524 1527 }
1525 1528
1526 1529 /* ARGSUSED */
1527 1530 zio_t *
1528 1531 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1529 1532 {
1530 1533 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1531 1534 ZIO_GANG_CHILD_FLAGS(pio)));
1532 1535 }
1533 1536
1534 1537 /* ARGSUSED */
1535 1538 zio_t *
1536 1539 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1537 1540 {
1538 1541 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1539 1542 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1540 1543 }
1541 1544
1542 1545 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1543 1546 NULL,
1544 1547 zio_read_gang,
1545 1548 zio_rewrite_gang,
1546 1549 zio_free_gang,
1547 1550 zio_claim_gang,
1548 1551 NULL
1549 1552 };
1550 1553
1551 1554 static void zio_gang_tree_assemble_done(zio_t *zio);
1552 1555
1553 1556 static zio_gang_node_t *
1554 1557 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1555 1558 {
1556 1559 zio_gang_node_t *gn;
1557 1560
1558 1561 ASSERT(*gnpp == NULL);
1559 1562
1560 1563 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1561 1564 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1562 1565 *gnpp = gn;
1563 1566
1564 1567 return (gn);
1565 1568 }
1566 1569
1567 1570 static void
1568 1571 zio_gang_node_free(zio_gang_node_t **gnpp)
1569 1572 {
1570 1573 zio_gang_node_t *gn = *gnpp;
1571 1574
1572 1575 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1573 1576 ASSERT(gn->gn_child[g] == NULL);
1574 1577
1575 1578 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1576 1579 kmem_free(gn, sizeof (*gn));
1577 1580 *gnpp = NULL;
1578 1581 }
1579 1582
1580 1583 static void
1581 1584 zio_gang_tree_free(zio_gang_node_t **gnpp)
1582 1585 {
1583 1586 zio_gang_node_t *gn = *gnpp;
1584 1587
1585 1588 if (gn == NULL)
1586 1589 return;
1587 1590
1588 1591 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1589 1592 zio_gang_tree_free(&gn->gn_child[g]);
1590 1593
1591 1594 zio_gang_node_free(gnpp);
1592 1595 }
1593 1596
1594 1597 static void
1595 1598 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1596 1599 {
1597 1600 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1598 1601
1599 1602 ASSERT(gio->io_gang_leader == gio);
1600 1603 ASSERT(BP_IS_GANG(bp));
1601 1604
1602 1605 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1603 1606 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1604 1607 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1605 1608 }
1606 1609
1607 1610 static void
1608 1611 zio_gang_tree_assemble_done(zio_t *zio)
1609 1612 {
1610 1613 zio_t *gio = zio->io_gang_leader;
1611 1614 zio_gang_node_t *gn = zio->io_private;
1612 1615 blkptr_t *bp = zio->io_bp;
1613 1616
1614 1617 ASSERT(gio == zio_unique_parent(zio));
1615 1618 ASSERT(zio->io_child_count == 0);
1616 1619
1617 1620 if (zio->io_error)
1618 1621 return;
1619 1622
1620 1623 if (BP_SHOULD_BYTESWAP(bp))
1621 1624 byteswap_uint64_array(zio->io_data, zio->io_size);
1622 1625
1623 1626 ASSERT(zio->io_data == gn->gn_gbh);
1624 1627 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1625 1628 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1626 1629
1627 1630 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1628 1631 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1629 1632 if (!BP_IS_GANG(gbp))
1630 1633 continue;
1631 1634 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1632 1635 }
1633 1636 }
1634 1637
1635 1638 static void
1636 1639 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1637 1640 {
1638 1641 zio_t *gio = pio->io_gang_leader;
1639 1642 zio_t *zio;
1640 1643
1641 1644 ASSERT(BP_IS_GANG(bp) == !!gn);
1642 1645 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1643 1646 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1644 1647
1645 1648 /*
1646 1649 * If you're a gang header, your data is in gn->gn_gbh.
1647 1650 * If you're a gang member, your data is in 'data' and gn == NULL.
1648 1651 */
1649 1652 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1650 1653
1651 1654 if (gn != NULL) {
1652 1655 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1653 1656
1654 1657 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1655 1658 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1656 1659 if (BP_IS_HOLE(gbp))
1657 1660 continue;
1658 1661 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1659 1662 data = (char *)data + BP_GET_PSIZE(gbp);
1660 1663 }
1661 1664 }
1662 1665
1663 1666 if (gn == gio->io_gang_tree)
1664 1667 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1665 1668
1666 1669 if (zio != pio)
1667 1670 zio_nowait(zio);
1668 1671 }
1669 1672
1670 1673 static int
1671 1674 zio_gang_assemble(zio_t *zio)
1672 1675 {
1673 1676 blkptr_t *bp = zio->io_bp;
1674 1677
1675 1678 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1676 1679 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1677 1680
1678 1681 zio->io_gang_leader = zio;
1679 1682
1680 1683 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1681 1684
1682 1685 return (ZIO_PIPELINE_CONTINUE);
1683 1686 }
1684 1687
1685 1688 static int
1686 1689 zio_gang_issue(zio_t *zio)
1687 1690 {
1688 1691 blkptr_t *bp = zio->io_bp;
1689 1692
1690 1693 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1691 1694 return (ZIO_PIPELINE_STOP);
1692 1695
1693 1696 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1694 1697 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1695 1698
1696 1699 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1697 1700 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1698 1701 else
1699 1702 zio_gang_tree_free(&zio->io_gang_tree);
1700 1703
1701 1704 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1702 1705
1703 1706 return (ZIO_PIPELINE_CONTINUE);
1704 1707 }
1705 1708
1706 1709 static void
1707 1710 zio_write_gang_member_ready(zio_t *zio)
1708 1711 {
1709 1712 zio_t *pio = zio_unique_parent(zio);
1710 1713 zio_t *gio = zio->io_gang_leader;
1711 1714 dva_t *cdva = zio->io_bp->blk_dva;
1712 1715 dva_t *pdva = pio->io_bp->blk_dva;
1713 1716 uint64_t asize;
1714 1717
1715 1718 if (BP_IS_HOLE(zio->io_bp))
1716 1719 return;
1717 1720
1718 1721 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1719 1722
1720 1723 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1721 1724 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1722 1725 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1723 1726 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1724 1727 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1725 1728
1726 1729 mutex_enter(&pio->io_lock);
1727 1730 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1728 1731 ASSERT(DVA_GET_GANG(&pdva[d]));
1729 1732 asize = DVA_GET_ASIZE(&pdva[d]);
1730 1733 asize += DVA_GET_ASIZE(&cdva[d]);
1731 1734 DVA_SET_ASIZE(&pdva[d], asize);
1732 1735 }
1733 1736 mutex_exit(&pio->io_lock);
1734 1737 }
1735 1738
1736 1739 static int
1737 1740 zio_write_gang_block(zio_t *pio)
1738 1741 {
1739 1742 spa_t *spa = pio->io_spa;
1740 1743 blkptr_t *bp = pio->io_bp;
1741 1744 zio_t *gio = pio->io_gang_leader;
1742 1745 zio_t *zio;
1743 1746 zio_gang_node_t *gn, **gnpp;
1744 1747 zio_gbh_phys_t *gbh;
1745 1748 uint64_t txg = pio->io_txg;
1746 1749 uint64_t resid = pio->io_size;
1747 1750 uint64_t lsize;
1748 1751 int copies = gio->io_prop.zp_copies;
1749 1752 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1750 1753 zio_prop_t zp;
1751 1754 int error;
1752 1755
1753 1756 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1754 1757 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1755 1758 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1756 1759 if (error) {
1757 1760 pio->io_error = error;
1758 1761 return (ZIO_PIPELINE_CONTINUE);
1759 1762 }
1760 1763
1761 1764 if (pio == gio) {
1762 1765 gnpp = &gio->io_gang_tree;
1763 1766 } else {
1764 1767 gnpp = pio->io_private;
1765 1768 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1766 1769 }
1767 1770
1768 1771 gn = zio_gang_node_alloc(gnpp);
1769 1772 gbh = gn->gn_gbh;
1770 1773 bzero(gbh, SPA_GANGBLOCKSIZE);
1771 1774
1772 1775 /*
1773 1776 * Create the gang header.
1774 1777 */
1775 1778 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1776 1779 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1777 1780
1778 1781 /*
1779 1782 * Create and nowait the gang children.
1780 1783 */
1781 1784 for (int g = 0; resid != 0; resid -= lsize, g++) {
1782 1785 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1783 1786 SPA_MINBLOCKSIZE);
1784 1787 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1785 1788
1786 1789 zp.zp_checksum = gio->io_prop.zp_checksum;
1787 1790 zp.zp_compress = ZIO_COMPRESS_OFF;
1788 1791 zp.zp_type = DMU_OT_NONE;
1789 1792 zp.zp_level = 0;
1790 1793 zp.zp_copies = gio->io_prop.zp_copies;
1791 1794 zp.zp_dedup = B_FALSE;
1792 1795 zp.zp_dedup_verify = B_FALSE;
1793 1796 zp.zp_nopwrite = B_FALSE;
1794 1797
1795 1798 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1796 1799 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1797 1800 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1798 1801 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1799 1802 &pio->io_bookmark));
1800 1803 }
1801 1804
1802 1805 /*
1803 1806 * Set pio's pipeline to just wait for zio to finish.
1804 1807 */
1805 1808 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1806 1809
1807 1810 zio_nowait(zio);
1808 1811
1809 1812 return (ZIO_PIPELINE_CONTINUE);
1810 1813 }
1811 1814
1812 1815 /*
1813 1816 * The zio_nop_write stage in the pipeline determines if allocating
1814 1817 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1815 1818 * such as SHA256, we can compare the checksums of the new data and the old
1816 1819 * to determine if allocating a new block is required. The nopwrite
1817 1820 * feature can handle writes in either syncing or open context (i.e. zil
1818 1821 * writes) and as a result is mutually exclusive with dedup.
1819 1822 */
1820 1823 static int
1821 1824 zio_nop_write(zio_t *zio)
1822 1825 {
1823 1826 blkptr_t *bp = zio->io_bp;
1824 1827 blkptr_t *bp_orig = &zio->io_bp_orig;
1825 1828 zio_prop_t *zp = &zio->io_prop;
1826 1829
1827 1830 ASSERT(BP_GET_LEVEL(bp) == 0);
1828 1831 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1829 1832 ASSERT(zp->zp_nopwrite);
1830 1833 ASSERT(!zp->zp_dedup);
1831 1834 ASSERT(zio->io_bp_override == NULL);
1832 1835 ASSERT(IO_IS_ALLOCATING(zio));
1833 1836
1834 1837 /*
1835 1838 * Check to see if the original bp and the new bp have matching
1836 1839 * characteristics (i.e. same checksum, compression algorithms, etc).
1837 1840 * If they don't then just continue with the pipeline which will
1838 1841 * allocate a new bp.
1839 1842 */
1840 1843 if (BP_IS_HOLE(bp_orig) ||
1841 1844 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1842 1845 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1843 1846 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1844 1847 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1845 1848 zp->zp_copies != BP_GET_NDVAS(bp_orig))
1846 1849 return (ZIO_PIPELINE_CONTINUE);
1847 1850
1848 1851 /*
1849 1852 * If the checksums match then reset the pipeline so that we
1850 1853 * avoid allocating a new bp and issuing any I/O.
1851 1854 */
1852 1855 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1853 1856 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1854 1857 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1855 1858 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1856 1859 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1857 1860 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1858 1861 sizeof (uint64_t)) == 0);
1859 1862
1860 1863 *bp = *bp_orig;
1861 1864 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1862 1865 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1863 1866 }
1864 1867
1865 1868 return (ZIO_PIPELINE_CONTINUE);
1866 1869 }
1867 1870
1868 1871 /*
1869 1872 * ==========================================================================
1870 1873 * Dedup
1871 1874 * ==========================================================================
1872 1875 */
1873 1876 static void
1874 1877 zio_ddt_child_read_done(zio_t *zio)
1875 1878 {
1876 1879 blkptr_t *bp = zio->io_bp;
1877 1880 ddt_entry_t *dde = zio->io_private;
1878 1881 ddt_phys_t *ddp;
1879 1882 zio_t *pio = zio_unique_parent(zio);
1880 1883
1881 1884 mutex_enter(&pio->io_lock);
1882 1885 ddp = ddt_phys_select(dde, bp);
1883 1886 if (zio->io_error == 0)
1884 1887 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1885 1888 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1886 1889 dde->dde_repair_data = zio->io_data;
1887 1890 else
1888 1891 zio_buf_free(zio->io_data, zio->io_size);
1889 1892 mutex_exit(&pio->io_lock);
1890 1893 }
1891 1894
1892 1895 static int
1893 1896 zio_ddt_read_start(zio_t *zio)
1894 1897 {
1895 1898 blkptr_t *bp = zio->io_bp;
1896 1899
1897 1900 ASSERT(BP_GET_DEDUP(bp));
1898 1901 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1899 1902 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1900 1903
1901 1904 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1902 1905 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1903 1906 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1904 1907 ddt_phys_t *ddp = dde->dde_phys;
1905 1908 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1906 1909 blkptr_t blk;
1907 1910
1908 1911 ASSERT(zio->io_vsd == NULL);
1909 1912 zio->io_vsd = dde;
1910 1913
1911 1914 if (ddp_self == NULL)
1912 1915 return (ZIO_PIPELINE_CONTINUE);
1913 1916
1914 1917 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1915 1918 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1916 1919 continue;
1917 1920 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1918 1921 &blk);
1919 1922 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1920 1923 zio_buf_alloc(zio->io_size), zio->io_size,
1921 1924 zio_ddt_child_read_done, dde, zio->io_priority,
1922 1925 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1923 1926 &zio->io_bookmark));
1924 1927 }
1925 1928 return (ZIO_PIPELINE_CONTINUE);
1926 1929 }
1927 1930
1928 1931 zio_nowait(zio_read(zio, zio->io_spa, bp,
1929 1932 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1930 1933 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1931 1934
1932 1935 return (ZIO_PIPELINE_CONTINUE);
1933 1936 }
1934 1937
1935 1938 static int
1936 1939 zio_ddt_read_done(zio_t *zio)
1937 1940 {
1938 1941 blkptr_t *bp = zio->io_bp;
1939 1942
1940 1943 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1941 1944 return (ZIO_PIPELINE_STOP);
1942 1945
1943 1946 ASSERT(BP_GET_DEDUP(bp));
1944 1947 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1945 1948 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1946 1949
1947 1950 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1948 1951 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1949 1952 ddt_entry_t *dde = zio->io_vsd;
1950 1953 if (ddt == NULL) {
1951 1954 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1952 1955 return (ZIO_PIPELINE_CONTINUE);
1953 1956 }
1954 1957 if (dde == NULL) {
1955 1958 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1956 1959 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1957 1960 return (ZIO_PIPELINE_STOP);
1958 1961 }
1959 1962 if (dde->dde_repair_data != NULL) {
1960 1963 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1961 1964 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1962 1965 }
1963 1966 ddt_repair_done(ddt, dde);
1964 1967 zio->io_vsd = NULL;
1965 1968 }
1966 1969
1967 1970 ASSERT(zio->io_vsd == NULL);
1968 1971
1969 1972 return (ZIO_PIPELINE_CONTINUE);
1970 1973 }
1971 1974
1972 1975 static boolean_t
1973 1976 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1974 1977 {
1975 1978 spa_t *spa = zio->io_spa;
1976 1979
1977 1980 /*
1978 1981 * Note: we compare the original data, not the transformed data,
1979 1982 * because when zio->io_bp is an override bp, we will not have
1980 1983 * pushed the I/O transforms. That's an important optimization
1981 1984 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1982 1985 */
1983 1986 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1984 1987 zio_t *lio = dde->dde_lead_zio[p];
1985 1988
1986 1989 if (lio != NULL) {
1987 1990 return (lio->io_orig_size != zio->io_orig_size ||
1988 1991 bcmp(zio->io_orig_data, lio->io_orig_data,
1989 1992 zio->io_orig_size) != 0);
1990 1993 }
1991 1994 }
1992 1995
1993 1996 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1994 1997 ddt_phys_t *ddp = &dde->dde_phys[p];
1995 1998
1996 1999 if (ddp->ddp_phys_birth != 0) {
1997 2000 arc_buf_t *abuf = NULL;
1998 2001 uint32_t aflags = ARC_WAIT;
1999 2002 blkptr_t blk = *zio->io_bp;
2000 2003 int error;
2001 2004
2002 2005 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2003 2006
2004 2007 ddt_exit(ddt);
2005 2008
2006 2009 error = arc_read(NULL, spa, &blk,
2007 2010 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2008 2011 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2009 2012 &aflags, &zio->io_bookmark);
2010 2013
2011 2014 if (error == 0) {
2012 2015 if (arc_buf_size(abuf) != zio->io_orig_size ||
2013 2016 bcmp(abuf->b_data, zio->io_orig_data,
2014 2017 zio->io_orig_size) != 0)
2015 2018 error = SET_ERROR(EEXIST);
2016 2019 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2017 2020 }
2018 2021
2019 2022 ddt_enter(ddt);
2020 2023 return (error != 0);
2021 2024 }
2022 2025 }
2023 2026
2024 2027 return (B_FALSE);
2025 2028 }
2026 2029
2027 2030 static void
2028 2031 zio_ddt_child_write_ready(zio_t *zio)
2029 2032 {
2030 2033 int p = zio->io_prop.zp_copies;
2031 2034 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2032 2035 ddt_entry_t *dde = zio->io_private;
2033 2036 ddt_phys_t *ddp = &dde->dde_phys[p];
2034 2037 zio_t *pio;
2035 2038
2036 2039 if (zio->io_error)
2037 2040 return;
2038 2041
2039 2042 ddt_enter(ddt);
2040 2043
2041 2044 ASSERT(dde->dde_lead_zio[p] == zio);
2042 2045
2043 2046 ddt_phys_fill(ddp, zio->io_bp);
2044 2047
2045 2048 while ((pio = zio_walk_parents(zio)) != NULL)
2046 2049 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2047 2050
2048 2051 ddt_exit(ddt);
2049 2052 }
2050 2053
2051 2054 static void
2052 2055 zio_ddt_child_write_done(zio_t *zio)
2053 2056 {
2054 2057 int p = zio->io_prop.zp_copies;
2055 2058 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2056 2059 ddt_entry_t *dde = zio->io_private;
2057 2060 ddt_phys_t *ddp = &dde->dde_phys[p];
2058 2061
2059 2062 ddt_enter(ddt);
2060 2063
2061 2064 ASSERT(ddp->ddp_refcnt == 0);
2062 2065 ASSERT(dde->dde_lead_zio[p] == zio);
2063 2066 dde->dde_lead_zio[p] = NULL;
2064 2067
2065 2068 if (zio->io_error == 0) {
2066 2069 while (zio_walk_parents(zio) != NULL)
2067 2070 ddt_phys_addref(ddp);
2068 2071 } else {
2069 2072 ddt_phys_clear(ddp);
2070 2073 }
2071 2074
2072 2075 ddt_exit(ddt);
2073 2076 }
2074 2077
2075 2078 static void
2076 2079 zio_ddt_ditto_write_done(zio_t *zio)
2077 2080 {
2078 2081 int p = DDT_PHYS_DITTO;
2079 2082 zio_prop_t *zp = &zio->io_prop;
2080 2083 blkptr_t *bp = zio->io_bp;
2081 2084 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2082 2085 ddt_entry_t *dde = zio->io_private;
2083 2086 ddt_phys_t *ddp = &dde->dde_phys[p];
2084 2087 ddt_key_t *ddk = &dde->dde_key;
2085 2088
2086 2089 ddt_enter(ddt);
2087 2090
2088 2091 ASSERT(ddp->ddp_refcnt == 0);
2089 2092 ASSERT(dde->dde_lead_zio[p] == zio);
2090 2093 dde->dde_lead_zio[p] = NULL;
2091 2094
2092 2095 if (zio->io_error == 0) {
2093 2096 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2094 2097 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2095 2098 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2096 2099 if (ddp->ddp_phys_birth != 0)
2097 2100 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2098 2101 ddt_phys_fill(ddp, bp);
2099 2102 }
2100 2103
2101 2104 ddt_exit(ddt);
2102 2105 }
2103 2106
2104 2107 static int
2105 2108 zio_ddt_write(zio_t *zio)
2106 2109 {
2107 2110 spa_t *spa = zio->io_spa;
2108 2111 blkptr_t *bp = zio->io_bp;
2109 2112 uint64_t txg = zio->io_txg;
2110 2113 zio_prop_t *zp = &zio->io_prop;
2111 2114 int p = zp->zp_copies;
2112 2115 int ditto_copies;
2113 2116 zio_t *cio = NULL;
2114 2117 zio_t *dio = NULL;
2115 2118 ddt_t *ddt = ddt_select(spa, bp);
2116 2119 ddt_entry_t *dde;
2117 2120 ddt_phys_t *ddp;
2118 2121
2119 2122 ASSERT(BP_GET_DEDUP(bp));
2120 2123 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2121 2124 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2122 2125
2123 2126 ddt_enter(ddt);
2124 2127 dde = ddt_lookup(ddt, bp, B_TRUE);
2125 2128 ddp = &dde->dde_phys[p];
2126 2129
2127 2130 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2128 2131 /*
2129 2132 * If we're using a weak checksum, upgrade to a strong checksum
2130 2133 * and try again. If we're already using a strong checksum,
2131 2134 * we can't resolve it, so just convert to an ordinary write.
2132 2135 * (And automatically e-mail a paper to Nature?)
2133 2136 */
2134 2137 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2135 2138 zp->zp_checksum = spa_dedup_checksum(spa);
2136 2139 zio_pop_transforms(zio);
2137 2140 zio->io_stage = ZIO_STAGE_OPEN;
2138 2141 BP_ZERO(bp);
2139 2142 } else {
2140 2143 zp->zp_dedup = B_FALSE;
2141 2144 }
2142 2145 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2143 2146 ddt_exit(ddt);
2144 2147 return (ZIO_PIPELINE_CONTINUE);
2145 2148 }
2146 2149
2147 2150 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2148 2151 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2149 2152
2150 2153 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2151 2154 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2152 2155 zio_prop_t czp = *zp;
2153 2156
2154 2157 czp.zp_copies = ditto_copies;
2155 2158
2156 2159 /*
2157 2160 * If we arrived here with an override bp, we won't have run
2158 2161 * the transform stack, so we won't have the data we need to
2159 2162 * generate a child i/o. So, toss the override bp and restart.
2160 2163 * This is safe, because using the override bp is just an
2161 2164 * optimization; and it's rare, so the cost doesn't matter.
2162 2165 */
2163 2166 if (zio->io_bp_override) {
2164 2167 zio_pop_transforms(zio);
2165 2168 zio->io_stage = ZIO_STAGE_OPEN;
2166 2169 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2167 2170 zio->io_bp_override = NULL;
2168 2171 BP_ZERO(bp);
2169 2172 ddt_exit(ddt);
2170 2173 return (ZIO_PIPELINE_CONTINUE);
2171 2174 }
2172 2175
2173 2176 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2174 2177 zio->io_orig_size, &czp, NULL,
2175 2178 zio_ddt_ditto_write_done, dde, zio->io_priority,
2176 2179 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2177 2180
2178 2181 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2179 2182 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2180 2183 }
2181 2184
2182 2185 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2183 2186 if (ddp->ddp_phys_birth != 0)
2184 2187 ddt_bp_fill(ddp, bp, txg);
2185 2188 if (dde->dde_lead_zio[p] != NULL)
2186 2189 zio_add_child(zio, dde->dde_lead_zio[p]);
2187 2190 else
2188 2191 ddt_phys_addref(ddp);
2189 2192 } else if (zio->io_bp_override) {
2190 2193 ASSERT(bp->blk_birth == txg);
2191 2194 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2192 2195 ddt_phys_fill(ddp, bp);
2193 2196 ddt_phys_addref(ddp);
2194 2197 } else {
2195 2198 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2196 2199 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2197 2200 zio_ddt_child_write_done, dde, zio->io_priority,
2198 2201 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2199 2202
2200 2203 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2201 2204 dde->dde_lead_zio[p] = cio;
2202 2205 }
2203 2206
2204 2207 ddt_exit(ddt);
2205 2208
2206 2209 if (cio)
2207 2210 zio_nowait(cio);
2208 2211 if (dio)
2209 2212 zio_nowait(dio);
2210 2213
2211 2214 return (ZIO_PIPELINE_CONTINUE);
2212 2215 }
2213 2216
2214 2217 ddt_entry_t *freedde; /* for debugging */
2215 2218
2216 2219 static int
2217 2220 zio_ddt_free(zio_t *zio)
2218 2221 {
2219 2222 spa_t *spa = zio->io_spa;
2220 2223 blkptr_t *bp = zio->io_bp;
2221 2224 ddt_t *ddt = ddt_select(spa, bp);
2222 2225 ddt_entry_t *dde;
2223 2226 ddt_phys_t *ddp;
2224 2227
2225 2228 ASSERT(BP_GET_DEDUP(bp));
2226 2229 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2227 2230
2228 2231 ddt_enter(ddt);
2229 2232 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2230 2233 ddp = ddt_phys_select(dde, bp);
2231 2234 ddt_phys_decref(ddp);
2232 2235 ddt_exit(ddt);
2233 2236
2234 2237 return (ZIO_PIPELINE_CONTINUE);
2235 2238 }
2236 2239
2237 2240 /*
2238 2241 * ==========================================================================
2239 2242 * Allocate and free blocks
2240 2243 * ==========================================================================
2241 2244 */
2242 2245 static int
2243 2246 zio_dva_allocate(zio_t *zio)
2244 2247 {
2245 2248 spa_t *spa = zio->io_spa;
2246 2249 metaslab_class_t *mc = spa_normal_class(spa);
2247 2250 blkptr_t *bp = zio->io_bp;
2248 2251 int error;
2249 2252 int flags = 0;
2250 2253
2251 2254 if (zio->io_gang_leader == NULL) {
2252 2255 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2253 2256 zio->io_gang_leader = zio;
2254 2257 }
2255 2258
2256 2259 ASSERT(BP_IS_HOLE(bp));
2257 2260 ASSERT0(BP_GET_NDVAS(bp));
2258 2261 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2259 2262 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2260 2263 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2261 2264
2262 2265 /*
2263 2266 * The dump device does not support gang blocks so allocation on
2264 2267 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2265 2268 * the "fast" gang feature.
2266 2269 */
2267 2270 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2268 2271 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2269 2272 METASLAB_GANG_CHILD : 0;
2270 2273 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2271 2274 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2272 2275
2273 2276 if (error) {
2274 2277 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2275 2278 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2276 2279 error);
2277 2280 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2278 2281 return (zio_write_gang_block(zio));
2279 2282 zio->io_error = error;
2280 2283 }
2281 2284
2282 2285 return (ZIO_PIPELINE_CONTINUE);
2283 2286 }
2284 2287
2285 2288 static int
2286 2289 zio_dva_free(zio_t *zio)
2287 2290 {
2288 2291 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2289 2292
2290 2293 return (ZIO_PIPELINE_CONTINUE);
2291 2294 }
2292 2295
2293 2296 static int
2294 2297 zio_dva_claim(zio_t *zio)
2295 2298 {
2296 2299 int error;
2297 2300
2298 2301 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2299 2302 if (error)
2300 2303 zio->io_error = error;
2301 2304
2302 2305 return (ZIO_PIPELINE_CONTINUE);
2303 2306 }
2304 2307
2305 2308 /*
2306 2309 * Undo an allocation. This is used by zio_done() when an I/O fails
2307 2310 * and we want to give back the block we just allocated.
2308 2311 * This handles both normal blocks and gang blocks.
2309 2312 */
2310 2313 static void
2311 2314 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2312 2315 {
2313 2316 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2314 2317 ASSERT(zio->io_bp_override == NULL);
2315 2318
2316 2319 if (!BP_IS_HOLE(bp))
2317 2320 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2318 2321
2319 2322 if (gn != NULL) {
2320 2323 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2321 2324 zio_dva_unallocate(zio, gn->gn_child[g],
2322 2325 &gn->gn_gbh->zg_blkptr[g]);
2323 2326 }
2324 2327 }
2325 2328 }
2326 2329
2327 2330 /*
2328 2331 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2329 2332 */
2330 2333 int
2331 2334 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2332 2335 uint64_t size, boolean_t use_slog)
2333 2336 {
2334 2337 int error = 1;
2335 2338
2336 2339 ASSERT(txg > spa_syncing_txg(spa));
2337 2340
2338 2341 /*
2339 2342 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2340 2343 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2341 2344 * when allocating them.
2342 2345 */
2343 2346 if (use_slog) {
2344 2347 error = metaslab_alloc(spa, spa_log_class(spa), size,
2345 2348 new_bp, 1, txg, old_bp,
2346 2349 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2347 2350 }
2348 2351
2349 2352 if (error) {
2350 2353 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2351 2354 new_bp, 1, txg, old_bp,
2352 2355 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2353 2356 }
2354 2357
2355 2358 if (error == 0) {
2356 2359 BP_SET_LSIZE(new_bp, size);
2357 2360 BP_SET_PSIZE(new_bp, size);
2358 2361 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2359 2362 BP_SET_CHECKSUM(new_bp,
2360 2363 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2361 2364 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2362 2365 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2363 2366 BP_SET_LEVEL(new_bp, 0);
2364 2367 BP_SET_DEDUP(new_bp, 0);
2365 2368 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2366 2369 }
2367 2370
2368 2371 return (error);
2369 2372 }
2370 2373
2371 2374 /*
2372 2375 * Free an intent log block.
2373 2376 */
2374 2377 void
2375 2378 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2376 2379 {
2377 2380 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2378 2381 ASSERT(!BP_IS_GANG(bp));
2379 2382
2380 2383 zio_free(spa, txg, bp);
2381 2384 }
2382 2385
2383 2386 /*
2384 2387 * ==========================================================================
2385 2388 * Read and write to physical devices
2386 2389 * ==========================================================================
2387 2390 */
2388 2391 static int
2389 2392 zio_vdev_io_start(zio_t *zio)
2390 2393 {
2391 2394 vdev_t *vd = zio->io_vd;
2392 2395 uint64_t align;
2393 2396 spa_t *spa = zio->io_spa;
2394 2397
2395 2398 ASSERT(zio->io_error == 0);
2396 2399 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2397 2400
2398 2401 if (vd == NULL) {
2399 2402 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2400 2403 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2401 2404
2402 2405 /*
2403 2406 * The mirror_ops handle multiple DVAs in a single BP.
2404 2407 */
2405 2408 return (vdev_mirror_ops.vdev_op_io_start(zio));
2406 2409 }
2407 2410
2408 2411 /*
2409 2412 * We keep track of time-sensitive I/Os so that the scan thread
2410 2413 * can quickly react to certain workloads. In particular, we care
2411 2414 * about non-scrubbing, top-level reads and writes with the following
2412 2415 * characteristics:
2413 2416 * - synchronous writes of user data to non-slog devices
2414 2417 * - any reads of user data
2415 2418 * When these conditions are met, adjust the timestamp of spa_last_io
2416 2419 * which allows the scan thread to adjust its workload accordingly.
2417 2420 */
2418 2421 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2419 2422 vd == vd->vdev_top && !vd->vdev_islog &&
2420 2423 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2421 2424 zio->io_txg != spa_syncing_txg(spa)) {
2422 2425 uint64_t old = spa->spa_last_io;
2423 2426 uint64_t new = ddi_get_lbolt64();
2424 2427 if (old != new)
2425 2428 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2426 2429 }
2427 2430
2428 2431 align = 1ULL << vd->vdev_top->vdev_ashift;
2429 2432
2430 2433 if (P2PHASE(zio->io_size, align) != 0) {
2431 2434 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2432 2435 char *abuf = zio_buf_alloc(asize);
2433 2436 ASSERT(vd == vd->vdev_top);
2434 2437 if (zio->io_type == ZIO_TYPE_WRITE) {
2435 2438 bcopy(zio->io_data, abuf, zio->io_size);
2436 2439 bzero(abuf + zio->io_size, asize - zio->io_size);
2437 2440 }
2438 2441 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2439 2442 }
2440 2443
2441 2444 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2442 2445 ASSERT(P2PHASE(zio->io_size, align) == 0);
2443 2446 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2444 2447
2445 2448 /*
2446 2449 * If this is a repair I/O, and there's no self-healing involved --
2447 2450 * that is, we're just resilvering what we expect to resilver --
2448 2451 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2449 2452 * This prevents spurious resilvering with nested replication.
2450 2453 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2451 2454 * A is out of date, we'll read from C+D, then use the data to
2452 2455 * resilver A+B -- but we don't actually want to resilver B, just A.
2453 2456 * The top-level mirror has no way to know this, so instead we just
2454 2457 * discard unnecessary repairs as we work our way down the vdev tree.
2455 2458 * The same logic applies to any form of nested replication:
2456 2459 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2457 2460 */
2458 2461 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2459 2462 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2460 2463 zio->io_txg != 0 && /* not a delegated i/o */
2461 2464 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2462 2465 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2463 2466 zio_vdev_io_bypass(zio);
2464 2467 return (ZIO_PIPELINE_CONTINUE);
2465 2468 }
2466 2469
2467 2470 if (vd->vdev_ops->vdev_op_leaf &&
2468 2471 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2469 2472
2470 2473 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2471 2474 return (ZIO_PIPELINE_CONTINUE);
2472 2475
2473 2476 if ((zio = vdev_queue_io(zio)) == NULL)
2474 2477 return (ZIO_PIPELINE_STOP);
2475 2478
2476 2479 if (!vdev_accessible(vd, zio)) {
2477 2480 zio->io_error = SET_ERROR(ENXIO);
2478 2481 zio_interrupt(zio);
2479 2482 return (ZIO_PIPELINE_STOP);
2480 2483 }
2481 2484 }
2482 2485
2483 2486 return (vd->vdev_ops->vdev_op_io_start(zio));
2484 2487 }
2485 2488
2486 2489 static int
2487 2490 zio_vdev_io_done(zio_t *zio)
2488 2491 {
2489 2492 vdev_t *vd = zio->io_vd;
2490 2493 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2491 2494 boolean_t unexpected_error = B_FALSE;
2492 2495
2493 2496 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2494 2497 return (ZIO_PIPELINE_STOP);
2495 2498
2496 2499 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2497 2500
2498 2501 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2499 2502
2500 2503 vdev_queue_io_done(zio);
2501 2504
2502 2505 if (zio->io_type == ZIO_TYPE_WRITE)
2503 2506 vdev_cache_write(zio);
2504 2507
2505 2508 if (zio_injection_enabled && zio->io_error == 0)
2506 2509 zio->io_error = zio_handle_device_injection(vd,
2507 2510 zio, EIO);
2508 2511
2509 2512 if (zio_injection_enabled && zio->io_error == 0)
2510 2513 zio->io_error = zio_handle_label_injection(zio, EIO);
2511 2514
2512 2515 if (zio->io_error) {
2513 2516 if (!vdev_accessible(vd, zio)) {
2514 2517 zio->io_error = SET_ERROR(ENXIO);
2515 2518 } else {
2516 2519 unexpected_error = B_TRUE;
2517 2520 }
2518 2521 }
2519 2522 }
2520 2523
2521 2524 ops->vdev_op_io_done(zio);
2522 2525
2523 2526 if (unexpected_error)
2524 2527 VERIFY(vdev_probe(vd, zio) == NULL);
2525 2528
2526 2529 return (ZIO_PIPELINE_CONTINUE);
2527 2530 }
2528 2531
2529 2532 /*
2530 2533 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2531 2534 * disk, and use that to finish the checksum ereport later.
2532 2535 */
2533 2536 static void
2534 2537 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2535 2538 const void *good_buf)
2536 2539 {
2537 2540 /* no processing needed */
2538 2541 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2539 2542 }
2540 2543
2541 2544 /*ARGSUSED*/
2542 2545 void
2543 2546 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2544 2547 {
2545 2548 void *buf = zio_buf_alloc(zio->io_size);
2546 2549
2547 2550 bcopy(zio->io_data, buf, zio->io_size);
2548 2551
2549 2552 zcr->zcr_cbinfo = zio->io_size;
2550 2553 zcr->zcr_cbdata = buf;
2551 2554 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2552 2555 zcr->zcr_free = zio_buf_free;
2553 2556 }
2554 2557
2555 2558 static int
2556 2559 zio_vdev_io_assess(zio_t *zio)
2557 2560 {
2558 2561 vdev_t *vd = zio->io_vd;
2559 2562
2560 2563 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2561 2564 return (ZIO_PIPELINE_STOP);
2562 2565
2563 2566 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2564 2567 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2565 2568
2566 2569 if (zio->io_vsd != NULL) {
2567 2570 zio->io_vsd_ops->vsd_free(zio);
2568 2571 zio->io_vsd = NULL;
2569 2572 }
2570 2573
2571 2574 if (zio_injection_enabled && zio->io_error == 0)
2572 2575 zio->io_error = zio_handle_fault_injection(zio, EIO);
2573 2576
2574 2577 /*
2575 2578 * If the I/O failed, determine whether we should attempt to retry it.
2576 2579 *
2577 2580 * On retry, we cut in line in the issue queue, since we don't want
2578 2581 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2579 2582 */
2580 2583 if (zio->io_error && vd == NULL &&
2581 2584 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2582 2585 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2583 2586 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2584 2587 zio->io_error = 0;
2585 2588 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2586 2589 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2587 2590 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2588 2591 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2589 2592 zio_requeue_io_start_cut_in_line);
2590 2593 return (ZIO_PIPELINE_STOP);
2591 2594 }
2592 2595
2593 2596 /*
2594 2597 * If we got an error on a leaf device, convert it to ENXIO
2595 2598 * if the device is not accessible at all.
2596 2599 */
2597 2600 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2598 2601 !vdev_accessible(vd, zio))
2599 2602 zio->io_error = SET_ERROR(ENXIO);
2600 2603
2601 2604 /*
2602 2605 * If we can't write to an interior vdev (mirror or RAID-Z),
2603 2606 * set vdev_cant_write so that we stop trying to allocate from it.
2604 2607 */
2605 2608 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2606 2609 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2607 2610 vd->vdev_cant_write = B_TRUE;
2608 2611 }
2609 2612
2610 2613 if (zio->io_error)
2611 2614 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2612 2615
2613 2616 return (ZIO_PIPELINE_CONTINUE);
2614 2617 }
2615 2618
2616 2619 void
2617 2620 zio_vdev_io_reissue(zio_t *zio)
2618 2621 {
2619 2622 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2620 2623 ASSERT(zio->io_error == 0);
2621 2624
2622 2625 zio->io_stage >>= 1;
2623 2626 }
2624 2627
2625 2628 void
2626 2629 zio_vdev_io_redone(zio_t *zio)
2627 2630 {
2628 2631 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2629 2632
2630 2633 zio->io_stage >>= 1;
2631 2634 }
2632 2635
2633 2636 void
2634 2637 zio_vdev_io_bypass(zio_t *zio)
2635 2638 {
2636 2639 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2637 2640 ASSERT(zio->io_error == 0);
2638 2641
2639 2642 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2640 2643 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2641 2644 }
2642 2645
2643 2646 /*
2644 2647 * ==========================================================================
2645 2648 * Generate and verify checksums
2646 2649 * ==========================================================================
2647 2650 */
2648 2651 static int
2649 2652 zio_checksum_generate(zio_t *zio)
2650 2653 {
2651 2654 blkptr_t *bp = zio->io_bp;
2652 2655 enum zio_checksum checksum;
2653 2656
2654 2657 if (bp == NULL) {
2655 2658 /*
2656 2659 * This is zio_write_phys().
2657 2660 * We're either generating a label checksum, or none at all.
2658 2661 */
2659 2662 checksum = zio->io_prop.zp_checksum;
2660 2663
2661 2664 if (checksum == ZIO_CHECKSUM_OFF)
2662 2665 return (ZIO_PIPELINE_CONTINUE);
2663 2666
2664 2667 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2665 2668 } else {
2666 2669 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2667 2670 ASSERT(!IO_IS_ALLOCATING(zio));
2668 2671 checksum = ZIO_CHECKSUM_GANG_HEADER;
2669 2672 } else {
2670 2673 checksum = BP_GET_CHECKSUM(bp);
2671 2674 }
2672 2675 }
2673 2676
2674 2677 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2675 2678
2676 2679 return (ZIO_PIPELINE_CONTINUE);
2677 2680 }
2678 2681
2679 2682 static int
2680 2683 zio_checksum_verify(zio_t *zio)
2681 2684 {
2682 2685 zio_bad_cksum_t info;
2683 2686 blkptr_t *bp = zio->io_bp;
2684 2687 int error;
2685 2688
2686 2689 ASSERT(zio->io_vd != NULL);
2687 2690
2688 2691 if (bp == NULL) {
2689 2692 /*
2690 2693 * This is zio_read_phys().
2691 2694 * We're either verifying a label checksum, or nothing at all.
2692 2695 */
2693 2696 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2694 2697 return (ZIO_PIPELINE_CONTINUE);
2695 2698
2696 2699 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2697 2700 }
2698 2701
2699 2702 if ((error = zio_checksum_error(zio, &info)) != 0) {
2700 2703 zio->io_error = error;
2701 2704 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2702 2705 zfs_ereport_start_checksum(zio->io_spa,
2703 2706 zio->io_vd, zio, zio->io_offset,
2704 2707 zio->io_size, NULL, &info);
2705 2708 }
2706 2709 }
2707 2710
2708 2711 return (ZIO_PIPELINE_CONTINUE);
2709 2712 }
2710 2713
2711 2714 /*
2712 2715 * Called by RAID-Z to ensure we don't compute the checksum twice.
2713 2716 */
2714 2717 void
2715 2718 zio_checksum_verified(zio_t *zio)
2716 2719 {
2717 2720 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2718 2721 }
2719 2722
2720 2723 /*
2721 2724 * ==========================================================================
2722 2725 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2723 2726 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2724 2727 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2725 2728 * indicate errors that are specific to one I/O, and most likely permanent.
2726 2729 * Any other error is presumed to be worse because we weren't expecting it.
2727 2730 * ==========================================================================
2728 2731 */
2729 2732 int
2730 2733 zio_worst_error(int e1, int e2)
2731 2734 {
2732 2735 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2733 2736 int r1, r2;
2734 2737
2735 2738 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2736 2739 if (e1 == zio_error_rank[r1])
2737 2740 break;
2738 2741
2739 2742 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2740 2743 if (e2 == zio_error_rank[r2])
2741 2744 break;
2742 2745
2743 2746 return (r1 > r2 ? e1 : e2);
2744 2747 }
2745 2748
2746 2749 /*
2747 2750 * ==========================================================================
2748 2751 * I/O completion
2749 2752 * ==========================================================================
2750 2753 */
2751 2754 static int
2752 2755 zio_ready(zio_t *zio)
2753 2756 {
2754 2757 blkptr_t *bp = zio->io_bp;
2755 2758 zio_t *pio, *pio_next;
2756 2759
2757 2760 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2758 2761 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2759 2762 return (ZIO_PIPELINE_STOP);
2760 2763
2761 2764 if (zio->io_ready) {
2762 2765 ASSERT(IO_IS_ALLOCATING(zio));
2763 2766 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2764 2767 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2765 2768 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2766 2769
2767 2770 zio->io_ready(zio);
2768 2771 }
2769 2772
2770 2773 if (bp != NULL && bp != &zio->io_bp_copy)
2771 2774 zio->io_bp_copy = *bp;
2772 2775
2773 2776 if (zio->io_error)
2774 2777 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2775 2778
2776 2779 mutex_enter(&zio->io_lock);
2777 2780 zio->io_state[ZIO_WAIT_READY] = 1;
2778 2781 pio = zio_walk_parents(zio);
2779 2782 mutex_exit(&zio->io_lock);
2780 2783
2781 2784 /*
2782 2785 * As we notify zio's parents, new parents could be added.
2783 2786 * New parents go to the head of zio's io_parent_list, however,
2784 2787 * so we will (correctly) not notify them. The remainder of zio's
2785 2788 * io_parent_list, from 'pio_next' onward, cannot change because
2786 2789 * all parents must wait for us to be done before they can be done.
2787 2790 */
2788 2791 for (; pio != NULL; pio = pio_next) {
2789 2792 pio_next = zio_walk_parents(zio);
2790 2793 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2791 2794 }
2792 2795
2793 2796 if (zio->io_flags & ZIO_FLAG_NODATA) {
2794 2797 if (BP_IS_GANG(bp)) {
2795 2798 zio->io_flags &= ~ZIO_FLAG_NODATA;
2796 2799 } else {
2797 2800 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2798 2801 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2799 2802 }
2800 2803 }
2801 2804
2802 2805 if (zio_injection_enabled &&
2803 2806 zio->io_spa->spa_syncing_txg == zio->io_txg)
2804 2807 zio_handle_ignored_writes(zio);
2805 2808
2806 2809 return (ZIO_PIPELINE_CONTINUE);
2807 2810 }
2808 2811
2809 2812 static int
2810 2813 zio_done(zio_t *zio)
2811 2814 {
2812 2815 spa_t *spa = zio->io_spa;
2813 2816 zio_t *lio = zio->io_logical;
2814 2817 blkptr_t *bp = zio->io_bp;
2815 2818 vdev_t *vd = zio->io_vd;
2816 2819 uint64_t psize = zio->io_size;
2817 2820 zio_t *pio, *pio_next;
2818 2821
2819 2822 /*
2820 2823 * If our children haven't all completed,
2821 2824 * wait for them and then repeat this pipeline stage.
2822 2825 */
2823 2826 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2824 2827 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2825 2828 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2826 2829 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2827 2830 return (ZIO_PIPELINE_STOP);
2828 2831
2829 2832 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2830 2833 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2831 2834 ASSERT(zio->io_children[c][w] == 0);
2832 2835
2833 2836 if (bp != NULL) {
2834 2837 ASSERT(bp->blk_pad[0] == 0);
2835 2838 ASSERT(bp->blk_pad[1] == 0);
2836 2839 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2837 2840 (bp == zio_unique_parent(zio)->io_bp));
2838 2841 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2839 2842 zio->io_bp_override == NULL &&
2840 2843 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2841 2844 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2842 2845 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2843 2846 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2844 2847 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2845 2848 }
2846 2849 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
2847 2850 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
2848 2851 }
2849 2852
2850 2853 /*
2851 2854 * If there were child vdev/gang/ddt errors, they apply to us now.
2852 2855 */
2853 2856 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2854 2857 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2855 2858 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2856 2859
2857 2860 /*
2858 2861 * If the I/O on the transformed data was successful, generate any
2859 2862 * checksum reports now while we still have the transformed data.
2860 2863 */
2861 2864 if (zio->io_error == 0) {
2862 2865 while (zio->io_cksum_report != NULL) {
2863 2866 zio_cksum_report_t *zcr = zio->io_cksum_report;
2864 2867 uint64_t align = zcr->zcr_align;
2865 2868 uint64_t asize = P2ROUNDUP(psize, align);
2866 2869 char *abuf = zio->io_data;
2867 2870
2868 2871 if (asize != psize) {
2869 2872 abuf = zio_buf_alloc(asize);
2870 2873 bcopy(zio->io_data, abuf, psize);
2871 2874 bzero(abuf + psize, asize - psize);
2872 2875 }
2873 2876
2874 2877 zio->io_cksum_report = zcr->zcr_next;
2875 2878 zcr->zcr_next = NULL;
2876 2879 zcr->zcr_finish(zcr, abuf);
2877 2880 zfs_ereport_free_checksum(zcr);
2878 2881
2879 2882 if (asize != psize)
2880 2883 zio_buf_free(abuf, asize);
2881 2884 }
2882 2885 }
2883 2886
2884 2887 zio_pop_transforms(zio); /* note: may set zio->io_error */
2885 2888
2886 2889 vdev_stat_update(zio, psize);
2887 2890
2888 2891 if (zio->io_error) {
2889 2892 /*
2890 2893 * If this I/O is attached to a particular vdev,
2891 2894 * generate an error message describing the I/O failure
2892 2895 * at the block level. We ignore these errors if the
2893 2896 * device is currently unavailable.
2894 2897 */
2895 2898 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2896 2899 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2897 2900
2898 2901 if ((zio->io_error == EIO || !(zio->io_flags &
2899 2902 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2900 2903 zio == lio) {
2901 2904 /*
2902 2905 * For logical I/O requests, tell the SPA to log the
2903 2906 * error and generate a logical data ereport.
2904 2907 */
2905 2908 spa_log_error(spa, zio);
2906 2909 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2907 2910 0, 0);
2908 2911 }
2909 2912 }
2910 2913
2911 2914 if (zio->io_error && zio == lio) {
2912 2915 /*
2913 2916 * Determine whether zio should be reexecuted. This will
2914 2917 * propagate all the way to the root via zio_notify_parent().
2915 2918 */
2916 2919 ASSERT(vd == NULL && bp != NULL);
2917 2920 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2918 2921
2919 2922 if (IO_IS_ALLOCATING(zio) &&
2920 2923 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2921 2924 if (zio->io_error != ENOSPC)
2922 2925 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2923 2926 else
2924 2927 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2925 2928 }
2926 2929
2927 2930 if ((zio->io_type == ZIO_TYPE_READ ||
2928 2931 zio->io_type == ZIO_TYPE_FREE) &&
2929 2932 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2930 2933 zio->io_error == ENXIO &&
2931 2934 spa_load_state(spa) == SPA_LOAD_NONE &&
2932 2935 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2933 2936 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2934 2937
2935 2938 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2936 2939 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2937 2940
2938 2941 /*
2939 2942 * Here is a possibly good place to attempt to do
2940 2943 * either combinatorial reconstruction or error correction
2941 2944 * based on checksums. It also might be a good place
2942 2945 * to send out preliminary ereports before we suspend
2943 2946 * processing.
2944 2947 */
2945 2948 }
2946 2949
2947 2950 /*
2948 2951 * If there were logical child errors, they apply to us now.
2949 2952 * We defer this until now to avoid conflating logical child
2950 2953 * errors with errors that happened to the zio itself when
2951 2954 * updating vdev stats and reporting FMA events above.
2952 2955 */
2953 2956 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2954 2957
2955 2958 if ((zio->io_error || zio->io_reexecute) &&
2956 2959 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2957 2960 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
2958 2961 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2959 2962
2960 2963 zio_gang_tree_free(&zio->io_gang_tree);
2961 2964
2962 2965 /*
2963 2966 * Godfather I/Os should never suspend.
2964 2967 */
2965 2968 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2966 2969 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2967 2970 zio->io_reexecute = 0;
2968 2971
2969 2972 if (zio->io_reexecute) {
2970 2973 /*
2971 2974 * This is a logical I/O that wants to reexecute.
2972 2975 *
2973 2976 * Reexecute is top-down. When an i/o fails, if it's not
2974 2977 * the root, it simply notifies its parent and sticks around.
2975 2978 * The parent, seeing that it still has children in zio_done(),
2976 2979 * does the same. This percolates all the way up to the root.
2977 2980 * The root i/o will reexecute or suspend the entire tree.
2978 2981 *
2979 2982 * This approach ensures that zio_reexecute() honors
2980 2983 * all the original i/o dependency relationships, e.g.
2981 2984 * parents not executing until children are ready.
2982 2985 */
2983 2986 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2984 2987
2985 2988 zio->io_gang_leader = NULL;
2986 2989
2987 2990 mutex_enter(&zio->io_lock);
2988 2991 zio->io_state[ZIO_WAIT_DONE] = 1;
2989 2992 mutex_exit(&zio->io_lock);
2990 2993
2991 2994 /*
2992 2995 * "The Godfather" I/O monitors its children but is
2993 2996 * not a true parent to them. It will track them through
2994 2997 * the pipeline but severs its ties whenever they get into
2995 2998 * trouble (e.g. suspended). This allows "The Godfather"
2996 2999 * I/O to return status without blocking.
2997 3000 */
2998 3001 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2999 3002 zio_link_t *zl = zio->io_walk_link;
3000 3003 pio_next = zio_walk_parents(zio);
3001 3004
3002 3005 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3003 3006 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3004 3007 zio_remove_child(pio, zio, zl);
3005 3008 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3006 3009 }
3007 3010 }
3008 3011
3009 3012 if ((pio = zio_unique_parent(zio)) != NULL) {
3010 3013 /*
3011 3014 * We're not a root i/o, so there's nothing to do
3012 3015 * but notify our parent. Don't propagate errors
3013 3016 * upward since we haven't permanently failed yet.
3014 3017 */
3015 3018 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3016 3019 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3017 3020 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3018 3021 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3019 3022 /*
3020 3023 * We'd fail again if we reexecuted now, so suspend
3021 3024 * until conditions improve (e.g. device comes online).
3022 3025 */
3023 3026 zio_suspend(spa, zio);
3024 3027 } else {
3025 3028 /*
3026 3029 * Reexecution is potentially a huge amount of work.
3027 3030 * Hand it off to the otherwise-unused claim taskq.
3028 3031 */
3029 3032 ASSERT(zio->io_tqent.tqent_next == NULL);
3030 3033 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3031 3034 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3032 3035 0, &zio->io_tqent);
3033 3036 }
3034 3037 return (ZIO_PIPELINE_STOP);
3035 3038 }
3036 3039
3037 3040 ASSERT(zio->io_child_count == 0);
3038 3041 ASSERT(zio->io_reexecute == 0);
3039 3042 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3040 3043
3041 3044 /*
3042 3045 * Report any checksum errors, since the I/O is complete.
3043 3046 */
3044 3047 while (zio->io_cksum_report != NULL) {
3045 3048 zio_cksum_report_t *zcr = zio->io_cksum_report;
3046 3049 zio->io_cksum_report = zcr->zcr_next;
3047 3050 zcr->zcr_next = NULL;
3048 3051 zcr->zcr_finish(zcr, NULL);
3049 3052 zfs_ereport_free_checksum(zcr);
3050 3053 }
3051 3054
3052 3055 /*
3053 3056 * It is the responsibility of the done callback to ensure that this
3054 3057 * particular zio is no longer discoverable for adoption, and as
3055 3058 * such, cannot acquire any new parents.
3056 3059 */
3057 3060 if (zio->io_done)
3058 3061 zio->io_done(zio);
3059 3062
3060 3063 mutex_enter(&zio->io_lock);
3061 3064 zio->io_state[ZIO_WAIT_DONE] = 1;
3062 3065 mutex_exit(&zio->io_lock);
3063 3066
3064 3067 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3065 3068 zio_link_t *zl = zio->io_walk_link;
3066 3069 pio_next = zio_walk_parents(zio);
3067 3070 zio_remove_child(pio, zio, zl);
3068 3071 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3069 3072 }
3070 3073
3071 3074 if (zio->io_waiter != NULL) {
3072 3075 mutex_enter(&zio->io_lock);
3073 3076 zio->io_executor = NULL;
3074 3077 cv_broadcast(&zio->io_cv);
3075 3078 mutex_exit(&zio->io_lock);
3076 3079 } else {
3077 3080 zio_destroy(zio);
3078 3081 }
3079 3082
3080 3083 return (ZIO_PIPELINE_STOP);
3081 3084 }
3082 3085
3083 3086 /*
3084 3087 * ==========================================================================
3085 3088 * I/O pipeline definition
3086 3089 * ==========================================================================
3087 3090 */
3088 3091 static zio_pipe_stage_t *zio_pipeline[] = {
3089 3092 NULL,
3090 3093 zio_read_bp_init,
3091 3094 zio_free_bp_init,
3092 3095 zio_issue_async,
3093 3096 zio_write_bp_init,
3094 3097 zio_checksum_generate,
3095 3098 zio_nop_write,
3096 3099 zio_ddt_read_start,
3097 3100 zio_ddt_read_done,
3098 3101 zio_ddt_write,
3099 3102 zio_ddt_free,
3100 3103 zio_gang_assemble,
3101 3104 zio_gang_issue,
3102 3105 zio_dva_allocate,
3103 3106 zio_dva_free,
3104 3107 zio_dva_claim,
3105 3108 zio_ready,
3106 3109 zio_vdev_io_start,
3107 3110 zio_vdev_io_done,
3108 3111 zio_vdev_io_assess,
3109 3112 zio_checksum_verify,
3110 3113 zio_done
3111 3114 };
3112 3115
3113 3116 /* dnp is the dnode for zb1->zb_object */
3114 3117 boolean_t
3115 3118 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3116 3119 const zbookmark_t *zb2)
3117 3120 {
3118 3121 uint64_t zb1nextL0, zb2thisobj;
3119 3122
3120 3123 ASSERT(zb1->zb_objset == zb2->zb_objset);
3121 3124 ASSERT(zb2->zb_level == 0);
3122 3125
3123 3126 /*
3124 3127 * A bookmark in the deadlist is considered to be after
3125 3128 * everything else.
3126 3129 */
3127 3130 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3128 3131 return (B_TRUE);
3129 3132
3130 3133 /* The objset_phys_t isn't before anything. */
3131 3134 if (dnp == NULL)
3132 3135 return (B_FALSE);
3133 3136
3134 3137 zb1nextL0 = (zb1->zb_blkid + 1) <<
3135 3138 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3136 3139
3137 3140 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3138 3141 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3139 3142
3140 3143 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3141 3144 uint64_t nextobj = zb1nextL0 *
3142 3145 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3143 3146 return (nextobj <= zb2thisobj);
3144 3147 }
3145 3148
3146 3149 if (zb1->zb_object < zb2thisobj)
3147 3150 return (B_TRUE);
3148 3151 if (zb1->zb_object > zb2thisobj)
3149 3152 return (B_FALSE);
3150 3153 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3151 3154 return (B_FALSE);
3152 3155 return (zb1nextL0 <= zb2->zb_blkid);
3153 3156 }
|
↓ open down ↓ |
1950 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX