Print this page
Possibility to physically reserve space without writing leaf blocks
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) 2011, 2014 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 #include <sys/blkptr.h>
40 40 #include <sys/zfeature.h>
41 41
42 42 /*
43 43 * ==========================================================================
44 44 * I/O type descriptions
45 45 * ==========================================================================
46 46 */
47 47 const char *zio_type_name[ZIO_TYPES] = {
48 48 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
49 49 "zio_ioctl"
50 50 };
51 51
52 52 /*
53 53 * ==========================================================================
54 54 * I/O kmem caches
55 55 * ==========================================================================
56 56 */
57 57 kmem_cache_t *zio_cache;
58 58 kmem_cache_t *zio_link_cache;
59 59 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
60 60 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
61 61
62 62 #ifdef _KERNEL
63 63 extern vmem_t *zio_alloc_arena;
64 64 #endif
65 65
66 66 /*
67 67 * The following actions directly effect the spa's sync-to-convergence logic.
68 68 * The values below define the sync pass when we start performing the action.
69 69 * Care should be taken when changing these values as they directly impact
70 70 * spa_sync() performance. Tuning these values may introduce subtle performance
71 71 * pathologies and should only be done in the context of performance analysis.
72 72 * These tunables will eventually be removed and replaced with #defines once
73 73 * enough analysis has been done to determine optimal values.
74 74 *
75 75 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
76 76 * regular blocks are not deferred.
77 77 */
78 78 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
79 79 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
80 80 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
81 81
82 82 /*
83 83 * An allocating zio is one that either currently has the DVA allocate
84 84 * stage set or will have it later in its lifetime.
85 85 */
86 86 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
87 87
88 88 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
89 89
90 90 #ifdef ZFS_DEBUG
91 91 int zio_buf_debug_limit = 16384;
92 92 #else
93 93 int zio_buf_debug_limit = 0;
94 94 #endif
95 95
96 96 void
97 97 zio_init(void)
98 98 {
99 99 size_t c;
100 100 vmem_t *data_alloc_arena = NULL;
101 101
102 102 #ifdef _KERNEL
103 103 data_alloc_arena = zio_alloc_arena;
104 104 #endif
105 105 zio_cache = kmem_cache_create("zio_cache",
106 106 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
107 107 zio_link_cache = kmem_cache_create("zio_link_cache",
108 108 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
109 109
110 110 /*
111 111 * For small buffers, we want a cache for each multiple of
112 112 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
113 113 * for each quarter-power of 2. For large buffers, we want
114 114 * a cache for each multiple of PAGESIZE.
115 115 */
116 116 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
117 117 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
118 118 size_t p2 = size;
119 119 size_t align = 0;
120 120 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
121 121
122 122 while (p2 & (p2 - 1))
123 123 p2 &= p2 - 1;
124 124
125 125 #ifndef _KERNEL
126 126 /*
127 127 * If we are using watchpoints, put each buffer on its own page,
128 128 * to eliminate the performance overhead of trapping to the
129 129 * kernel when modifying a non-watched buffer that shares the
130 130 * page with a watched buffer.
131 131 */
132 132 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
133 133 continue;
134 134 #endif
135 135 if (size <= 4 * SPA_MINBLOCKSIZE) {
136 136 align = SPA_MINBLOCKSIZE;
137 137 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
138 138 align = PAGESIZE;
139 139 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
140 140 align = p2 >> 2;
141 141 }
142 142
143 143 if (align != 0) {
144 144 char name[36];
145 145 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
146 146 zio_buf_cache[c] = kmem_cache_create(name, size,
147 147 align, NULL, NULL, NULL, NULL, NULL, cflags);
148 148
149 149 /*
150 150 * Since zio_data bufs do not appear in crash dumps, we
151 151 * pass KMC_NOTOUCH so that no allocator metadata is
152 152 * stored with the buffers.
153 153 */
154 154 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
155 155 zio_data_buf_cache[c] = kmem_cache_create(name, size,
156 156 align, NULL, NULL, NULL, NULL, data_alloc_arena,
157 157 cflags | KMC_NOTOUCH);
158 158 }
159 159 }
160 160
161 161 while (--c != 0) {
162 162 ASSERT(zio_buf_cache[c] != NULL);
163 163 if (zio_buf_cache[c - 1] == NULL)
164 164 zio_buf_cache[c - 1] = zio_buf_cache[c];
165 165
166 166 ASSERT(zio_data_buf_cache[c] != NULL);
167 167 if (zio_data_buf_cache[c - 1] == NULL)
168 168 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
169 169 }
170 170
171 171 zio_inject_init();
172 172 }
173 173
174 174 void
175 175 zio_fini(void)
176 176 {
177 177 size_t c;
178 178 kmem_cache_t *last_cache = NULL;
179 179 kmem_cache_t *last_data_cache = NULL;
180 180
181 181 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
182 182 if (zio_buf_cache[c] != last_cache) {
183 183 last_cache = zio_buf_cache[c];
184 184 kmem_cache_destroy(zio_buf_cache[c]);
185 185 }
186 186 zio_buf_cache[c] = NULL;
187 187
188 188 if (zio_data_buf_cache[c] != last_data_cache) {
189 189 last_data_cache = zio_data_buf_cache[c];
190 190 kmem_cache_destroy(zio_data_buf_cache[c]);
191 191 }
192 192 zio_data_buf_cache[c] = NULL;
193 193 }
194 194
195 195 kmem_cache_destroy(zio_link_cache);
196 196 kmem_cache_destroy(zio_cache);
197 197
198 198 zio_inject_fini();
199 199 }
200 200
201 201 /*
202 202 * ==========================================================================
203 203 * Allocate and free I/O buffers
204 204 * ==========================================================================
205 205 */
206 206
207 207 /*
208 208 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
209 209 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
210 210 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
211 211 * excess / transient data in-core during a crashdump.
212 212 */
213 213 void *
214 214 zio_buf_alloc(size_t size)
215 215 {
216 216 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
217 217
218 218 ASSERT3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
219 219
220 220 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
221 221 }
222 222
223 223 /*
224 224 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
225 225 * crashdump if the kernel panics. This exists so that we will limit the amount
226 226 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
227 227 * of kernel heap dumped to disk when the kernel panics)
228 228 */
229 229 void *
230 230 zio_data_buf_alloc(size_t size)
231 231 {
232 232 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
233 233
234 234 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
235 235
236 236 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
237 237 }
238 238
239 239 void
240 240 zio_buf_free(void *buf, size_t size)
241 241 {
242 242 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
243 243
244 244 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
245 245
246 246 kmem_cache_free(zio_buf_cache[c], buf);
247 247 }
248 248
249 249 void
250 250 zio_data_buf_free(void *buf, size_t size)
251 251 {
252 252 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
253 253
254 254 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
255 255
256 256 kmem_cache_free(zio_data_buf_cache[c], buf);
257 257 }
258 258
259 259 /*
260 260 * ==========================================================================
261 261 * Push and pop I/O transform buffers
262 262 * ==========================================================================
263 263 */
264 264 static void
265 265 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
266 266 zio_transform_func_t *transform)
267 267 {
268 268 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
269 269
270 270 zt->zt_orig_data = zio->io_data;
271 271 zt->zt_orig_size = zio->io_size;
272 272 zt->zt_bufsize = bufsize;
273 273 zt->zt_transform = transform;
274 274
275 275 zt->zt_next = zio->io_transform_stack;
276 276 zio->io_transform_stack = zt;
277 277
278 278 zio->io_data = data;
279 279 zio->io_size = size;
280 280 }
281 281
282 282 static void
283 283 zio_pop_transforms(zio_t *zio)
284 284 {
285 285 zio_transform_t *zt;
286 286
287 287 while ((zt = zio->io_transform_stack) != NULL) {
288 288 if (zt->zt_transform != NULL)
289 289 zt->zt_transform(zio,
290 290 zt->zt_orig_data, zt->zt_orig_size);
291 291
292 292 if (zt->zt_bufsize != 0)
293 293 zio_buf_free(zio->io_data, zt->zt_bufsize);
294 294
295 295 zio->io_data = zt->zt_orig_data;
296 296 zio->io_size = zt->zt_orig_size;
297 297 zio->io_transform_stack = zt->zt_next;
298 298
299 299 kmem_free(zt, sizeof (zio_transform_t));
300 300 }
301 301 }
302 302
303 303 /*
304 304 * ==========================================================================
305 305 * I/O transform callbacks for subblocks and decompression
306 306 * ==========================================================================
307 307 */
308 308 static void
309 309 zio_subblock(zio_t *zio, void *data, uint64_t size)
310 310 {
311 311 ASSERT(zio->io_size > size);
312 312
313 313 if (zio->io_type == ZIO_TYPE_READ)
314 314 bcopy(zio->io_data, data, size);
315 315 }
316 316
317 317 static void
318 318 zio_decompress(zio_t *zio, void *data, uint64_t size)
319 319 {
320 320 if (zio->io_error == 0 &&
321 321 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
322 322 zio->io_data, data, zio->io_size, size) != 0)
323 323 zio->io_error = SET_ERROR(EIO);
324 324 }
325 325
326 326 /*
327 327 * ==========================================================================
328 328 * I/O parent/child relationships and pipeline interlocks
329 329 * ==========================================================================
330 330 */
331 331 /*
332 332 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
333 333 * continue calling these functions until they return NULL.
334 334 * Otherwise, the next caller will pick up the list walk in
335 335 * some indeterminate state. (Otherwise every caller would
336 336 * have to pass in a cookie to keep the state represented by
337 337 * io_walk_link, which gets annoying.)
338 338 */
339 339 zio_t *
340 340 zio_walk_parents(zio_t *cio)
341 341 {
342 342 zio_link_t *zl = cio->io_walk_link;
343 343 list_t *pl = &cio->io_parent_list;
344 344
345 345 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
346 346 cio->io_walk_link = zl;
347 347
348 348 if (zl == NULL)
349 349 return (NULL);
350 350
351 351 ASSERT(zl->zl_child == cio);
352 352 return (zl->zl_parent);
353 353 }
354 354
355 355 zio_t *
356 356 zio_walk_children(zio_t *pio)
357 357 {
358 358 zio_link_t *zl = pio->io_walk_link;
359 359 list_t *cl = &pio->io_child_list;
360 360
361 361 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
362 362 pio->io_walk_link = zl;
363 363
364 364 if (zl == NULL)
365 365 return (NULL);
366 366
367 367 ASSERT(zl->zl_parent == pio);
368 368 return (zl->zl_child);
369 369 }
370 370
371 371 zio_t *
372 372 zio_unique_parent(zio_t *cio)
373 373 {
374 374 zio_t *pio = zio_walk_parents(cio);
375 375
376 376 VERIFY(zio_walk_parents(cio) == NULL);
377 377 return (pio);
378 378 }
379 379
380 380 void
381 381 zio_add_child(zio_t *pio, zio_t *cio)
382 382 {
383 383 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
384 384
385 385 /*
386 386 * Logical I/Os can have logical, gang, or vdev children.
387 387 * Gang I/Os can have gang or vdev children.
388 388 * Vdev I/Os can only have vdev children.
389 389 * The following ASSERT captures all of these constraints.
390 390 */
391 391 ASSERT(cio->io_child_type <= pio->io_child_type);
392 392
393 393 zl->zl_parent = pio;
394 394 zl->zl_child = cio;
395 395
396 396 mutex_enter(&cio->io_lock);
397 397 mutex_enter(&pio->io_lock);
398 398
399 399 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
400 400
401 401 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
402 402 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
403 403
404 404 list_insert_head(&pio->io_child_list, zl);
405 405 list_insert_head(&cio->io_parent_list, zl);
406 406
407 407 pio->io_child_count++;
408 408 cio->io_parent_count++;
409 409
410 410 mutex_exit(&pio->io_lock);
411 411 mutex_exit(&cio->io_lock);
412 412 }
413 413
414 414 static void
415 415 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
416 416 {
417 417 ASSERT(zl->zl_parent == pio);
418 418 ASSERT(zl->zl_child == cio);
419 419
420 420 mutex_enter(&cio->io_lock);
421 421 mutex_enter(&pio->io_lock);
422 422
423 423 list_remove(&pio->io_child_list, zl);
424 424 list_remove(&cio->io_parent_list, zl);
425 425
426 426 pio->io_child_count--;
427 427 cio->io_parent_count--;
428 428
429 429 mutex_exit(&pio->io_lock);
430 430 mutex_exit(&cio->io_lock);
431 431
432 432 kmem_cache_free(zio_link_cache, zl);
433 433 }
434 434
435 435 static boolean_t
436 436 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
437 437 {
438 438 uint64_t *countp = &zio->io_children[child][wait];
439 439 boolean_t waiting = B_FALSE;
440 440
441 441 mutex_enter(&zio->io_lock);
442 442 ASSERT(zio->io_stall == NULL);
443 443 if (*countp != 0) {
444 444 zio->io_stage >>= 1;
445 445 zio->io_stall = countp;
446 446 waiting = B_TRUE;
447 447 }
448 448 mutex_exit(&zio->io_lock);
449 449
450 450 return (waiting);
451 451 }
452 452
453 453 static void
454 454 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
455 455 {
456 456 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
457 457 int *errorp = &pio->io_child_error[zio->io_child_type];
458 458
459 459 mutex_enter(&pio->io_lock);
460 460 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
461 461 *errorp = zio_worst_error(*errorp, zio->io_error);
462 462 pio->io_reexecute |= zio->io_reexecute;
463 463 ASSERT3U(*countp, >, 0);
464 464
465 465 (*countp)--;
466 466
467 467 if (*countp == 0 && pio->io_stall == countp) {
468 468 pio->io_stall = NULL;
469 469 mutex_exit(&pio->io_lock);
470 470 zio_execute(pio);
471 471 } else {
472 472 mutex_exit(&pio->io_lock);
473 473 }
474 474 }
475 475
476 476 static void
477 477 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
478 478 {
479 479 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
480 480 zio->io_error = zio->io_child_error[c];
481 481 }
482 482
483 483 /*
484 484 * ==========================================================================
485 485 * Create the various types of I/O (read, write, free, etc)
486 486 * ==========================================================================
487 487 */
488 488 static zio_t *
489 489 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
490 490 void *data, uint64_t size, zio_done_func_t *done, void *private,
491 491 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
492 492 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
493 493 enum zio_stage stage, enum zio_stage pipeline)
494 494 {
495 495 zio_t *zio;
496 496
497 497 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
498 498 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
499 499 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
500 500
501 501 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
502 502 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
503 503 ASSERT(vd || stage == ZIO_STAGE_OPEN);
504 504
505 505 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
506 506 bzero(zio, sizeof (zio_t));
507 507
508 508 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
509 509 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
510 510
511 511 list_create(&zio->io_parent_list, sizeof (zio_link_t),
512 512 offsetof(zio_link_t, zl_parent_node));
513 513 list_create(&zio->io_child_list, sizeof (zio_link_t),
514 514 offsetof(zio_link_t, zl_child_node));
515 515
516 516 if (vd != NULL)
517 517 zio->io_child_type = ZIO_CHILD_VDEV;
518 518 else if (flags & ZIO_FLAG_GANG_CHILD)
519 519 zio->io_child_type = ZIO_CHILD_GANG;
520 520 else if (flags & ZIO_FLAG_DDT_CHILD)
521 521 zio->io_child_type = ZIO_CHILD_DDT;
522 522 else
523 523 zio->io_child_type = ZIO_CHILD_LOGICAL;
524 524
525 525 if (bp != NULL) {
526 526 zio->io_bp = (blkptr_t *)bp;
527 527 zio->io_bp_copy = *bp;
528 528 zio->io_bp_orig = *bp;
529 529 if (type != ZIO_TYPE_WRITE ||
530 530 zio->io_child_type == ZIO_CHILD_DDT)
531 531 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
532 532 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
533 533 zio->io_logical = zio;
534 534 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
535 535 pipeline |= ZIO_GANG_STAGES;
536 536 }
537 537
538 538 zio->io_spa = spa;
539 539 zio->io_txg = txg;
540 540 zio->io_done = done;
541 541 zio->io_private = private;
542 542 zio->io_type = type;
543 543 zio->io_priority = priority;
544 544 zio->io_vd = vd;
545 545 zio->io_offset = offset;
546 546 zio->io_orig_data = zio->io_data = data;
547 547 zio->io_orig_size = zio->io_size = size;
548 548 zio->io_orig_flags = zio->io_flags = flags;
549 549 zio->io_orig_stage = zio->io_stage = stage;
550 550 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
551 551
552 552 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
553 553 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
554 554
555 555 if (zb != NULL)
556 556 zio->io_bookmark = *zb;
557 557
558 558 if (pio != NULL) {
559 559 if (zio->io_logical == NULL)
560 560 zio->io_logical = pio->io_logical;
561 561 if (zio->io_child_type == ZIO_CHILD_GANG)
562 562 zio->io_gang_leader = pio->io_gang_leader;
563 563 zio_add_child(pio, zio);
564 564 }
565 565
566 566 return (zio);
567 567 }
568 568
569 569 static void
570 570 zio_destroy(zio_t *zio)
571 571 {
572 572 list_destroy(&zio->io_parent_list);
573 573 list_destroy(&zio->io_child_list);
574 574 mutex_destroy(&zio->io_lock);
575 575 cv_destroy(&zio->io_cv);
576 576 kmem_cache_free(zio_cache, zio);
577 577 }
578 578
579 579 zio_t *
580 580 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
581 581 void *private, enum zio_flag flags)
582 582 {
583 583 zio_t *zio;
584 584
585 585 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
586 586 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
587 587 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
588 588
589 589 return (zio);
590 590 }
591 591
592 592 zio_t *
593 593 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
594 594 {
595 595 return (zio_null(NULL, spa, NULL, done, private, flags));
596 596 }
597 597
598 598 zio_t *
599 599 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
600 600 void *data, uint64_t size, zio_done_func_t *done, void *private,
601 601 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
602 602 {
603 603 zio_t *zio;
604 604
605 605 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
606 606 data, size, done, private,
607 607 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
608 608 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
609 609 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
610 610
611 611 return (zio);
612 612 }
613 613
614 614 zio_t *
615 615 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
616 616 void *data, uint64_t size, const zio_prop_t *zp,
617 617 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
618 618 void *private,
619 619 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
620 620 {
621 621 zio_t *zio;
622 622
623 623 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
624 624 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
625 625 zp->zp_compress >= ZIO_COMPRESS_OFF &&
626 626 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
627 627 DMU_OT_IS_VALID(zp->zp_type) &&
628 628 zp->zp_level < 32 &&
629 629 zp->zp_copies > 0 &&
630 630 zp->zp_copies <= spa_max_replication(spa));
631 631
632 632 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
633 633 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
634 634 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
635 635 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
636 636
637 637 zio->io_ready = ready;
638 638 zio->io_physdone = physdone;
639 639 zio->io_prop = *zp;
640 640
641 641 /*
642 642 * Data can be NULL if we are going to call zio_write_override() to
643 643 * provide the already-allocated BP. But we may need the data to
644 644 * verify a dedup hit (if requested). In this case, don't try to
645 645 * dedup (just take the already-allocated BP verbatim).
646 646 */
647 647 if (data == NULL && zio->io_prop.zp_dedup_verify) {
648 648 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
649 649 }
650 650
651 651 return (zio);
652 652 }
653 653
654 654 zio_t *
655 655 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
656 656 uint64_t size, zio_done_func_t *done, void *private,
657 657 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
658 658 {
659 659 zio_t *zio;
660 660
661 661 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
662 662 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
663 663 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
664 664
665 665 return (zio);
666 666 }
667 667
668 668 void
669 669 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
670 670 {
671 671 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
672 672 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
673 673 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
674 674 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
675 675
676 676 /*
677 677 * We must reset the io_prop to match the values that existed
678 678 * when the bp was first written by dmu_sync() keeping in mind
679 679 * that nopwrite and dedup are mutually exclusive.
680 680 */
681 681 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
682 682 zio->io_prop.zp_nopwrite = nopwrite;
683 683 zio->io_prop.zp_copies = copies;
684 684 zio->io_bp_override = bp;
685 685 }
686 686
687 687 void
688 688 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
689 689 {
690 690
691 691 /*
692 692 * The check for EMBEDDED is a performance optimization. We
693 693 * process the free here (by ignoring it) rather than
694 694 * putting it on the list and then processing it in zio_free_sync().
695 695 */
696 696 if (BP_IS_EMBEDDED(bp))
697 697 return;
698 698 metaslab_check_free(spa, bp);
699 699
700 700 /*
701 701 * Frees that are for the currently-syncing txg, are not going to be
702 702 * deferred, and which will not need to do a read (i.e. not GANG or
703 703 * DEDUP), can be processed immediately. Otherwise, put them on the
704 704 * in-memory list for later processing.
705 705 */
706 706 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
707 707 txg != spa->spa_syncing_txg ||
708 708 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
709 709 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
710 710 } else {
711 711 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
712 712 }
713 713 }
714 714
715 715 zio_t *
716 716 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
717 717 enum zio_flag flags)
718 718 {
719 719 zio_t *zio;
720 720 enum zio_stage stage = ZIO_FREE_PIPELINE;
721 721
722 722 ASSERT(!BP_IS_HOLE(bp));
723 723 ASSERT(spa_syncing_txg(spa) == txg);
724 724 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
725 725
726 726 if (BP_IS_EMBEDDED(bp))
727 727 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
728 728
729 729 metaslab_check_free(spa, bp);
730 730 arc_freed(spa, bp);
731 731
732 732 /*
733 733 * GANG and DEDUP blocks can induce a read (for the gang block header,
734 734 * or the DDT), so issue them asynchronously so that this thread is
735 735 * not tied up.
736 736 */
737 737 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
738 738 stage |= ZIO_STAGE_ISSUE_ASYNC;
739 739
740 740 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
741 741 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
742 742 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
743 743
744 744 return (zio);
745 745 }
746 746
747 747 zio_t *
748 748 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
749 749 zio_done_func_t *done, void *private, enum zio_flag flags)
750 750 {
751 751 zio_t *zio;
752 752
753 753 dprintf_bp(bp, "claiming in txg %llu", txg);
754 754
755 755 if (BP_IS_EMBEDDED(bp))
756 756 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
757 757
758 758 /*
759 759 * A claim is an allocation of a specific block. Claims are needed
760 760 * to support immediate writes in the intent log. The issue is that
761 761 * immediate writes contain committed data, but in a txg that was
762 762 * *not* committed. Upon opening the pool after an unclean shutdown,
763 763 * the intent log claims all blocks that contain immediate write data
764 764 * so that the SPA knows they're in use.
765 765 *
766 766 * All claims *must* be resolved in the first txg -- before the SPA
767 767 * starts allocating blocks -- so that nothing is allocated twice.
768 768 * If txg == 0 we just verify that the block is claimable.
769 769 */
770 770 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
771 771 ASSERT(txg == spa_first_txg(spa) || txg == 0);
772 772 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
773 773
774 774 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
775 775 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
776 776 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
777 777
778 778 return (zio);
779 779 }
780 780
781 781 zio_t *
782 782 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
783 783 zio_done_func_t *done, void *private, enum zio_flag flags)
784 784 {
785 785 zio_t *zio;
786 786 int c;
787 787
788 788 if (vd->vdev_children == 0) {
789 789 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
790 790 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
791 791 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
792 792
793 793 zio->io_cmd = cmd;
794 794 } else {
795 795 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
796 796
797 797 for (c = 0; c < vd->vdev_children; c++)
798 798 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
799 799 done, private, flags));
800 800 }
801 801
802 802 return (zio);
803 803 }
804 804
805 805 zio_t *
806 806 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
807 807 void *data, int checksum, zio_done_func_t *done, void *private,
808 808 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
809 809 {
810 810 zio_t *zio;
811 811
812 812 ASSERT(vd->vdev_children == 0);
813 813 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
814 814 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
815 815 ASSERT3U(offset + size, <=, vd->vdev_psize);
816 816
817 817 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
818 818 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
819 819 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
820 820
821 821 zio->io_prop.zp_checksum = checksum;
822 822
823 823 return (zio);
824 824 }
825 825
826 826 zio_t *
827 827 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
828 828 void *data, int checksum, zio_done_func_t *done, void *private,
829 829 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
830 830 {
831 831 zio_t *zio;
832 832
833 833 ASSERT(vd->vdev_children == 0);
834 834 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
835 835 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
836 836 ASSERT3U(offset + size, <=, vd->vdev_psize);
837 837
838 838 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
839 839 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
840 840 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
841 841
842 842 zio->io_prop.zp_checksum = checksum;
843 843
844 844 if (zio_checksum_table[checksum].ci_eck) {
845 845 /*
846 846 * zec checksums are necessarily destructive -- they modify
847 847 * the end of the write buffer to hold the verifier/checksum.
848 848 * Therefore, we must make a local copy in case the data is
849 849 * being written to multiple places in parallel.
850 850 */
851 851 void *wbuf = zio_buf_alloc(size);
852 852 bcopy(data, wbuf, size);
853 853 zio_push_transform(zio, wbuf, size, size, NULL);
854 854 }
855 855
856 856 return (zio);
857 857 }
858 858
859 859 /*
860 860 * Create a child I/O to do some work for us.
861 861 */
862 862 zio_t *
863 863 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
864 864 void *data, uint64_t size, int type, zio_priority_t priority,
865 865 enum zio_flag flags, zio_done_func_t *done, void *private)
866 866 {
867 867 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
868 868 zio_t *zio;
869 869
870 870 ASSERT(vd->vdev_parent ==
871 871 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
872 872
873 873 if (type == ZIO_TYPE_READ && bp != NULL) {
874 874 /*
875 875 * If we have the bp, then the child should perform the
876 876 * checksum and the parent need not. This pushes error
877 877 * detection as close to the leaves as possible and
878 878 * eliminates redundant checksums in the interior nodes.
879 879 */
880 880 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
881 881 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
882 882 }
883 883
884 884 if (vd->vdev_children == 0)
885 885 offset += VDEV_LABEL_START_SIZE;
886 886
887 887 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
888 888
889 889 /*
890 890 * If we've decided to do a repair, the write is not speculative --
891 891 * even if the original read was.
892 892 */
893 893 if (flags & ZIO_FLAG_IO_REPAIR)
894 894 flags &= ~ZIO_FLAG_SPECULATIVE;
895 895
896 896 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
897 897 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
898 898 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
899 899
900 900 zio->io_physdone = pio->io_physdone;
901 901 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
902 902 zio->io_logical->io_phys_children++;
903 903
904 904 return (zio);
905 905 }
906 906
907 907 zio_t *
908 908 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
909 909 int type, zio_priority_t priority, enum zio_flag flags,
910 910 zio_done_func_t *done, void *private)
911 911 {
912 912 zio_t *zio;
913 913
914 914 ASSERT(vd->vdev_ops->vdev_op_leaf);
915 915
916 916 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
917 917 data, size, done, private, type, priority,
918 918 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
919 919 vd, offset, NULL,
920 920 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
921 921
922 922 return (zio);
923 923 }
924 924
925 925 void
926 926 zio_flush(zio_t *zio, vdev_t *vd)
927 927 {
928 928 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
929 929 NULL, NULL,
930 930 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
931 931 }
932 932
933 933 void
934 934 zio_shrink(zio_t *zio, uint64_t size)
935 935 {
936 936 ASSERT(zio->io_executor == NULL);
937 937 ASSERT(zio->io_orig_size == zio->io_size);
938 938 ASSERT(size <= zio->io_size);
939 939
940 940 /*
941 941 * We don't shrink for raidz because of problems with the
942 942 * reconstruction when reading back less than the block size.
943 943 * Note, BP_IS_RAIDZ() assumes no compression.
944 944 */
945 945 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
946 946 if (!BP_IS_RAIDZ(zio->io_bp))
947 947 zio->io_orig_size = zio->io_size = size;
948 948 }
949 949
950 950 /*
↓ open down ↓ |
950 lines elided |
↑ open up ↑ |
951 951 * ==========================================================================
952 952 * Prepare to read and write logical blocks
953 953 * ==========================================================================
954 954 */
955 955
956 956 static int
957 957 zio_read_bp_init(zio_t *zio)
958 958 {
959 959 blkptr_t *bp = zio->io_bp;
960 960
961 + if (!BP_IS_EMBEDDED(bp) && BP_GET_PROP_RESERVATION(bp)) {
962 + memset(zio->io_orig_data, 0, zio->io_orig_size);
963 + zio->io_pipeline = ZIO_INTERLOCK_STAGES;
964 + return (ZIO_PIPELINE_CONTINUE);
965 + }
966 +
961 967 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
962 968 zio->io_child_type == ZIO_CHILD_LOGICAL &&
963 969 !(zio->io_flags & ZIO_FLAG_RAW)) {
964 970 uint64_t psize =
965 971 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
966 972 void *cbuf = zio_buf_alloc(psize);
967 973
968 974 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
969 975 }
970 976
971 977 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
972 978 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
973 979 decode_embedded_bp_compressed(bp, zio->io_data);
974 980 } else {
975 981 ASSERT(!BP_IS_EMBEDDED(bp));
976 982 }
977 983
978 984 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
979 985 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
980 986
981 987 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
982 988 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
983 989
984 990 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
985 991 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
↓ open down ↓ |
15 lines elided |
↑ open up ↑ |
986 992
987 993 return (ZIO_PIPELINE_CONTINUE);
988 994 }
989 995
990 996 static int
991 997 zio_write_bp_init(zio_t *zio)
992 998 {
993 999 spa_t *spa = zio->io_spa;
994 1000 zio_prop_t *zp = &zio->io_prop;
995 1001 enum zio_compress compress = zp->zp_compress;
1002 + enum zio_checksum checksum = zp->zp_checksum;
1003 + uint8_t dedup = zp->zp_dedup;
996 1004 blkptr_t *bp = zio->io_bp;
997 1005 uint64_t lsize = zio->io_size;
998 1006 uint64_t psize = lsize;
999 1007 int pass = 1;
1000 1008
1001 1009 /*
1002 1010 * If our children haven't all reached the ready stage,
1003 1011 * wait for them and then repeat this pipeline stage.
1004 1012 */
1005 1013 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1006 1014 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1007 1015 return (ZIO_PIPELINE_STOP);
1008 1016
1009 1017 if (!IO_IS_ALLOCATING(zio))
1010 1018 return (ZIO_PIPELINE_CONTINUE);
1011 1019
1012 1020 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1013 1021
1022 + if (zp->zp_zero_write && !(zio->io_pipeline & ZIO_GANG_STAGES)) {
1023 + dedup = B_FALSE;
1024 + compress = ZIO_COMPRESS_OFF;
1025 + checksum = ZIO_CHECKSUM_OFF;
1026 + }
1027 +
1014 1028 if (zio->io_bp_override) {
1015 1029 ASSERT(bp->blk_birth != zio->io_txg);
1016 1030 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1017 1031
1018 1032 *bp = *zio->io_bp_override;
1019 1033 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1020 1034
1021 1035 if (BP_IS_EMBEDDED(bp))
1022 1036 return (ZIO_PIPELINE_CONTINUE);
1023 1037
1024 1038 /*
1025 1039 * If we've been overridden and nopwrite is set then
1026 1040 * set the flag accordingly to indicate that a nopwrite
↓ open down ↓ |
3 lines elided |
↑ open up ↑ |
1027 1041 * has already occurred.
1028 1042 */
1029 1043 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1030 1044 ASSERT(!zp->zp_dedup);
1031 1045 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1032 1046 return (ZIO_PIPELINE_CONTINUE);
1033 1047 }
1034 1048
1035 1049 ASSERT(!zp->zp_nopwrite);
1036 1050
1037 - if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1051 + if (BP_IS_HOLE(bp) || !dedup)
1038 1052 return (ZIO_PIPELINE_CONTINUE);
1039 1053
1040 - ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1054 + ASSERT(zio_checksum_table[checksum].ci_dedup ||
1041 1055 zp->zp_dedup_verify);
1042 1056
1043 - if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1057 + if (BP_GET_CHECKSUM(bp) == checksum) {
1044 1058 BP_SET_DEDUP(bp, 1);
1045 1059 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1046 1060 return (ZIO_PIPELINE_CONTINUE);
1047 1061 }
1048 1062 zio->io_bp_override = NULL;
1049 1063 BP_ZERO(bp);
1050 1064 }
1051 1065
1052 1066 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1053 1067 /*
1054 1068 * We're rewriting an existing block, which means we're
1055 1069 * working on behalf of spa_sync(). For spa_sync() to
1056 1070 * converge, it must eventually be the case that we don't
1057 1071 * have to allocate new blocks. But compression changes
1058 1072 * the blocksize, which forces a reallocate, and makes
1059 1073 * convergence take longer. Therefore, after the first
1060 1074 * few passes, stop compressing to ensure convergence.
1061 1075 */
1062 1076 pass = spa_sync_pass(spa);
1063 1077
1064 1078 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1065 1079 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1066 1080 ASSERT(!BP_GET_DEDUP(bp));
1067 1081
1068 1082 if (pass >= zfs_sync_pass_dont_compress)
1069 1083 compress = ZIO_COMPRESS_OFF;
1070 1084
1071 1085 /* Make sure someone doesn't change their mind on overwrites */
1072 1086 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1073 1087 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1074 1088 }
1075 1089
1076 1090 if (compress != ZIO_COMPRESS_OFF) {
1077 1091 void *cbuf = zio_buf_alloc(lsize);
1078 1092 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1079 1093 if (psize == 0 || psize == lsize) {
1080 1094 compress = ZIO_COMPRESS_OFF;
1081 1095 zio_buf_free(cbuf, lsize);
1082 1096 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1083 1097 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1084 1098 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1085 1099 encode_embedded_bp_compressed(bp,
1086 1100 cbuf, compress, lsize, psize);
1087 1101 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1088 1102 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1089 1103 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1090 1104 zio_buf_free(cbuf, lsize);
1091 1105 bp->blk_birth = zio->io_txg;
1092 1106 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1093 1107 ASSERT(spa_feature_is_active(spa,
1094 1108 SPA_FEATURE_EMBEDDED_DATA));
1095 1109 return (ZIO_PIPELINE_CONTINUE);
1096 1110 } else {
1097 1111 /*
1098 1112 * Round up compressed size to MINBLOCKSIZE and
1099 1113 * zero the tail.
1100 1114 */
1101 1115 size_t rounded =
1102 1116 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE);
1103 1117 if (rounded > psize) {
1104 1118 bzero((char *)cbuf + psize, rounded - psize);
1105 1119 psize = rounded;
1106 1120 }
1107 1121 if (psize == lsize) {
1108 1122 compress = ZIO_COMPRESS_OFF;
1109 1123 zio_buf_free(cbuf, lsize);
1110 1124 } else {
1111 1125 zio_push_transform(zio, cbuf,
1112 1126 psize, lsize, NULL);
1113 1127 }
1114 1128 }
1115 1129 }
1116 1130
1117 1131 /*
1118 1132 * The final pass of spa_sync() must be all rewrites, but the first
1119 1133 * few passes offer a trade-off: allocating blocks defers convergence,
1120 1134 * but newly allocated blocks are sequential, so they can be written
1121 1135 * to disk faster. Therefore, we allow the first few passes of
1122 1136 * spa_sync() to allocate new blocks, but force rewrites after that.
1123 1137 * There should only be a handful of blocks after pass 1 in any case.
1124 1138 */
1125 1139 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1126 1140 BP_GET_PSIZE(bp) == psize &&
1127 1141 pass >= zfs_sync_pass_rewrite) {
1128 1142 ASSERT(psize != 0);
1129 1143 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1130 1144 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1131 1145 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1132 1146 } else {
1133 1147 BP_ZERO(bp);
1134 1148 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1135 1149 }
1136 1150
1137 1151 if (psize == 0) {
1138 1152 if (zio->io_bp_orig.blk_birth != 0 &&
1139 1153 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1140 1154 BP_SET_LSIZE(bp, lsize);
1141 1155 BP_SET_TYPE(bp, zp->zp_type);
1142 1156 BP_SET_LEVEL(bp, zp->zp_level);
1143 1157 BP_SET_BIRTH(bp, zio->io_txg, 0);
1144 1158 }
1145 1159 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
↓ open down ↓ |
92 lines elided |
↑ open up ↑ |
1146 1160 } else {
1147 1161 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1148 1162 BP_SET_LSIZE(bp, lsize);
1149 1163 BP_SET_TYPE(bp, zp->zp_type);
1150 1164 BP_SET_LEVEL(bp, zp->zp_level);
1151 1165 BP_SET_PSIZE(bp, psize);
1152 1166 BP_SET_COMPRESS(bp, compress);
1153 1167 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1154 1168 BP_SET_DEDUP(bp, zp->zp_dedup);
1155 1169 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1170 + if (zp->zp_zero_write && !(zio->io_pipeline & ZIO_GANG_STAGES)) {
1171 + boolean_t need_allocate = B_FALSE;
1172 + if (zio->io_pipeline & ZIO_STAGE_DVA_ALLOCATE)
1173 + need_allocate = B_TRUE;
1174 + zio->io_pipeline = ZIO_INTERLOCK_STAGES;
1175 + if (need_allocate)
1176 + zio->io_pipeline |= ZIO_STAGE_DVA_ALLOCATE;
1177 + BP_SET_PROP_RESERVATION(bp, 1);
1178 + } else {
1179 + BP_SET_PROP_RESERVATION(bp, 0);
1180 + }
1156 1181 if (zp->zp_dedup) {
1157 1182 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1158 1183 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1159 1184 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1160 1185 }
1161 1186 if (zp->zp_nopwrite) {
1162 1187 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1163 1188 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1164 1189 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1165 1190 }
1166 1191 }
1167 1192
1168 1193 return (ZIO_PIPELINE_CONTINUE);
1169 1194 }
1170 1195
1171 1196 static int
1172 1197 zio_free_bp_init(zio_t *zio)
1173 1198 {
1174 1199 blkptr_t *bp = zio->io_bp;
1175 1200
1176 1201 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1177 1202 if (BP_GET_DEDUP(bp))
1178 1203 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1179 1204 }
1180 1205
1181 1206 return (ZIO_PIPELINE_CONTINUE);
1182 1207 }
1183 1208
1184 1209 /*
1185 1210 * ==========================================================================
1186 1211 * Execute the I/O pipeline
1187 1212 * ==========================================================================
1188 1213 */
1189 1214
1190 1215 static void
1191 1216 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1192 1217 {
1193 1218 spa_t *spa = zio->io_spa;
1194 1219 zio_type_t t = zio->io_type;
1195 1220 int flags = (cutinline ? TQ_FRONT : 0);
1196 1221
1197 1222 /*
1198 1223 * If we're a config writer or a probe, the normal issue and
1199 1224 * interrupt threads may all be blocked waiting for the config lock.
1200 1225 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1201 1226 */
1202 1227 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1203 1228 t = ZIO_TYPE_NULL;
1204 1229
1205 1230 /*
1206 1231 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1207 1232 */
1208 1233 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1209 1234 t = ZIO_TYPE_NULL;
1210 1235
1211 1236 /*
1212 1237 * If this is a high priority I/O, then use the high priority taskq if
1213 1238 * available.
1214 1239 */
1215 1240 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1216 1241 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1217 1242 q++;
1218 1243
1219 1244 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1220 1245
1221 1246 /*
1222 1247 * NB: We are assuming that the zio can only be dispatched
1223 1248 * to a single taskq at a time. It would be a grievous error
1224 1249 * to dispatch the zio to another taskq at the same time.
1225 1250 */
1226 1251 ASSERT(zio->io_tqent.tqent_next == NULL);
1227 1252 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1228 1253 flags, &zio->io_tqent);
1229 1254 }
1230 1255
1231 1256 static boolean_t
1232 1257 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1233 1258 {
1234 1259 kthread_t *executor = zio->io_executor;
1235 1260 spa_t *spa = zio->io_spa;
1236 1261
1237 1262 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1238 1263 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1239 1264 uint_t i;
1240 1265 for (i = 0; i < tqs->stqs_count; i++) {
1241 1266 if (taskq_member(tqs->stqs_taskq[i], executor))
1242 1267 return (B_TRUE);
1243 1268 }
1244 1269 }
1245 1270
1246 1271 return (B_FALSE);
1247 1272 }
1248 1273
1249 1274 static int
1250 1275 zio_issue_async(zio_t *zio)
1251 1276 {
1252 1277 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1253 1278
1254 1279 return (ZIO_PIPELINE_STOP);
1255 1280 }
1256 1281
1257 1282 void
1258 1283 zio_interrupt(zio_t *zio)
1259 1284 {
1260 1285 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1261 1286 }
1262 1287
1263 1288 /*
1264 1289 * Execute the I/O pipeline until one of the following occurs:
1265 1290 *
1266 1291 * (1) the I/O completes
1267 1292 * (2) the pipeline stalls waiting for dependent child I/Os
1268 1293 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1269 1294 * (4) the I/O is delegated by vdev-level caching or aggregation
1270 1295 * (5) the I/O is deferred due to vdev-level queueing
1271 1296 * (6) the I/O is handed off to another thread.
1272 1297 *
1273 1298 * In all cases, the pipeline stops whenever there's no CPU work; it never
1274 1299 * burns a thread in cv_wait().
1275 1300 *
1276 1301 * There's no locking on io_stage because there's no legitimate way
1277 1302 * for multiple threads to be attempting to process the same I/O.
1278 1303 */
1279 1304 static zio_pipe_stage_t *zio_pipeline[];
1280 1305
1281 1306 void
1282 1307 zio_execute(zio_t *zio)
1283 1308 {
1284 1309 zio->io_executor = curthread;
1285 1310
1286 1311 while (zio->io_stage < ZIO_STAGE_DONE) {
1287 1312 enum zio_stage pipeline = zio->io_pipeline;
1288 1313 enum zio_stage stage = zio->io_stage;
1289 1314 int rv;
1290 1315
1291 1316 ASSERT(!MUTEX_HELD(&zio->io_lock));
1292 1317 ASSERT(ISP2(stage));
1293 1318 ASSERT(zio->io_stall == NULL);
1294 1319
1295 1320 do {
1296 1321 stage <<= 1;
1297 1322 } while ((stage & pipeline) == 0);
1298 1323
1299 1324 ASSERT(stage <= ZIO_STAGE_DONE);
1300 1325
1301 1326 /*
1302 1327 * If we are in interrupt context and this pipeline stage
1303 1328 * will grab a config lock that is held across I/O,
1304 1329 * or may wait for an I/O that needs an interrupt thread
1305 1330 * to complete, issue async to avoid deadlock.
1306 1331 *
1307 1332 * For VDEV_IO_START, we cut in line so that the io will
1308 1333 * be sent to disk promptly.
1309 1334 */
1310 1335 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1311 1336 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1312 1337 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1313 1338 zio_requeue_io_start_cut_in_line : B_FALSE;
1314 1339 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1315 1340 return;
1316 1341 }
1317 1342
1318 1343 zio->io_stage = stage;
1319 1344 rv = zio_pipeline[highbit64(stage) - 1](zio);
1320 1345
1321 1346 if (rv == ZIO_PIPELINE_STOP)
1322 1347 return;
1323 1348
1324 1349 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1325 1350 }
1326 1351 }
1327 1352
1328 1353 /*
1329 1354 * ==========================================================================
1330 1355 * Initiate I/O, either sync or async
1331 1356 * ==========================================================================
1332 1357 */
1333 1358 int
1334 1359 zio_wait(zio_t *zio)
1335 1360 {
1336 1361 int error;
1337 1362
1338 1363 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1339 1364 ASSERT(zio->io_executor == NULL);
1340 1365
1341 1366 zio->io_waiter = curthread;
1342 1367
1343 1368 zio_execute(zio);
1344 1369
1345 1370 mutex_enter(&zio->io_lock);
1346 1371 while (zio->io_executor != NULL)
1347 1372 cv_wait(&zio->io_cv, &zio->io_lock);
1348 1373 mutex_exit(&zio->io_lock);
1349 1374
1350 1375 error = zio->io_error;
1351 1376 zio_destroy(zio);
1352 1377
1353 1378 return (error);
1354 1379 }
1355 1380
1356 1381 void
1357 1382 zio_nowait(zio_t *zio)
1358 1383 {
1359 1384 ASSERT(zio->io_executor == NULL);
1360 1385
1361 1386 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1362 1387 zio_unique_parent(zio) == NULL) {
1363 1388 /*
1364 1389 * This is a logical async I/O with no parent to wait for it.
1365 1390 * We add it to the spa_async_root_zio "Godfather" I/O which
1366 1391 * will ensure they complete prior to unloading the pool.
1367 1392 */
1368 1393 spa_t *spa = zio->io_spa;
1369 1394
1370 1395 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1371 1396 }
1372 1397
1373 1398 zio_execute(zio);
1374 1399 }
1375 1400
1376 1401 /*
1377 1402 * ==========================================================================
1378 1403 * Reexecute or suspend/resume failed I/O
1379 1404 * ==========================================================================
1380 1405 */
1381 1406
1382 1407 static void
1383 1408 zio_reexecute(zio_t *pio)
1384 1409 {
1385 1410 zio_t *cio, *cio_next;
1386 1411
1387 1412 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1388 1413 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1389 1414 ASSERT(pio->io_gang_leader == NULL);
1390 1415 ASSERT(pio->io_gang_tree == NULL);
1391 1416
1392 1417 pio->io_flags = pio->io_orig_flags;
1393 1418 pio->io_stage = pio->io_orig_stage;
1394 1419 pio->io_pipeline = pio->io_orig_pipeline;
1395 1420 pio->io_reexecute = 0;
1396 1421 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1397 1422 pio->io_error = 0;
1398 1423 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1399 1424 pio->io_state[w] = 0;
1400 1425 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1401 1426 pio->io_child_error[c] = 0;
1402 1427
1403 1428 if (IO_IS_ALLOCATING(pio))
1404 1429 BP_ZERO(pio->io_bp);
1405 1430
1406 1431 /*
1407 1432 * As we reexecute pio's children, new children could be created.
1408 1433 * New children go to the head of pio's io_child_list, however,
1409 1434 * so we will (correctly) not reexecute them. The key is that
1410 1435 * the remainder of pio's io_child_list, from 'cio_next' onward,
1411 1436 * cannot be affected by any side effects of reexecuting 'cio'.
1412 1437 */
1413 1438 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1414 1439 cio_next = zio_walk_children(pio);
1415 1440 mutex_enter(&pio->io_lock);
1416 1441 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1417 1442 pio->io_children[cio->io_child_type][w]++;
1418 1443 mutex_exit(&pio->io_lock);
1419 1444 zio_reexecute(cio);
1420 1445 }
1421 1446
1422 1447 /*
1423 1448 * Now that all children have been reexecuted, execute the parent.
1424 1449 * We don't reexecute "The Godfather" I/O here as it's the
1425 1450 * responsibility of the caller to wait on him.
1426 1451 */
1427 1452 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1428 1453 zio_execute(pio);
1429 1454 }
1430 1455
1431 1456 void
1432 1457 zio_suspend(spa_t *spa, zio_t *zio)
1433 1458 {
1434 1459 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1435 1460 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1436 1461 "failure and the failure mode property for this pool "
1437 1462 "is set to panic.", spa_name(spa));
1438 1463
1439 1464 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1440 1465
1441 1466 mutex_enter(&spa->spa_suspend_lock);
1442 1467
1443 1468 if (spa->spa_suspend_zio_root == NULL)
1444 1469 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1445 1470 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1446 1471 ZIO_FLAG_GODFATHER);
1447 1472
1448 1473 spa->spa_suspended = B_TRUE;
1449 1474
1450 1475 if (zio != NULL) {
1451 1476 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1452 1477 ASSERT(zio != spa->spa_suspend_zio_root);
1453 1478 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1454 1479 ASSERT(zio_unique_parent(zio) == NULL);
1455 1480 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1456 1481 zio_add_child(spa->spa_suspend_zio_root, zio);
1457 1482 }
1458 1483
1459 1484 mutex_exit(&spa->spa_suspend_lock);
1460 1485 }
1461 1486
1462 1487 int
1463 1488 zio_resume(spa_t *spa)
1464 1489 {
1465 1490 zio_t *pio;
1466 1491
1467 1492 /*
1468 1493 * Reexecute all previously suspended i/o.
1469 1494 */
1470 1495 mutex_enter(&spa->spa_suspend_lock);
1471 1496 spa->spa_suspended = B_FALSE;
1472 1497 cv_broadcast(&spa->spa_suspend_cv);
1473 1498 pio = spa->spa_suspend_zio_root;
1474 1499 spa->spa_suspend_zio_root = NULL;
1475 1500 mutex_exit(&spa->spa_suspend_lock);
1476 1501
1477 1502 if (pio == NULL)
1478 1503 return (0);
1479 1504
1480 1505 zio_reexecute(pio);
1481 1506 return (zio_wait(pio));
1482 1507 }
1483 1508
1484 1509 void
1485 1510 zio_resume_wait(spa_t *spa)
1486 1511 {
1487 1512 mutex_enter(&spa->spa_suspend_lock);
1488 1513 while (spa_suspended(spa))
1489 1514 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1490 1515 mutex_exit(&spa->spa_suspend_lock);
1491 1516 }
1492 1517
1493 1518 /*
1494 1519 * ==========================================================================
1495 1520 * Gang blocks.
1496 1521 *
1497 1522 * A gang block is a collection of small blocks that looks to the DMU
1498 1523 * like one large block. When zio_dva_allocate() cannot find a block
1499 1524 * of the requested size, due to either severe fragmentation or the pool
1500 1525 * being nearly full, it calls zio_write_gang_block() to construct the
1501 1526 * block from smaller fragments.
1502 1527 *
1503 1528 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1504 1529 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1505 1530 * an indirect block: it's an array of block pointers. It consumes
1506 1531 * only one sector and hence is allocatable regardless of fragmentation.
1507 1532 * The gang header's bps point to its gang members, which hold the data.
1508 1533 *
1509 1534 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1510 1535 * as the verifier to ensure uniqueness of the SHA256 checksum.
1511 1536 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1512 1537 * not the gang header. This ensures that data block signatures (needed for
1513 1538 * deduplication) are independent of how the block is physically stored.
1514 1539 *
1515 1540 * Gang blocks can be nested: a gang member may itself be a gang block.
1516 1541 * Thus every gang block is a tree in which root and all interior nodes are
1517 1542 * gang headers, and the leaves are normal blocks that contain user data.
1518 1543 * The root of the gang tree is called the gang leader.
1519 1544 *
1520 1545 * To perform any operation (read, rewrite, free, claim) on a gang block,
1521 1546 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1522 1547 * in the io_gang_tree field of the original logical i/o by recursively
1523 1548 * reading the gang leader and all gang headers below it. This yields
1524 1549 * an in-core tree containing the contents of every gang header and the
1525 1550 * bps for every constituent of the gang block.
1526 1551 *
1527 1552 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1528 1553 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1529 1554 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1530 1555 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1531 1556 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1532 1557 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1533 1558 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1534 1559 * of the gang header plus zio_checksum_compute() of the data to update the
1535 1560 * gang header's blk_cksum as described above.
1536 1561 *
1537 1562 * The two-phase assemble/issue model solves the problem of partial failure --
1538 1563 * what if you'd freed part of a gang block but then couldn't read the
1539 1564 * gang header for another part? Assembling the entire gang tree first
1540 1565 * ensures that all the necessary gang header I/O has succeeded before
1541 1566 * starting the actual work of free, claim, or write. Once the gang tree
1542 1567 * is assembled, free and claim are in-memory operations that cannot fail.
1543 1568 *
1544 1569 * In the event that a gang write fails, zio_dva_unallocate() walks the
1545 1570 * gang tree to immediately free (i.e. insert back into the space map)
1546 1571 * everything we've allocated. This ensures that we don't get ENOSPC
1547 1572 * errors during repeated suspend/resume cycles due to a flaky device.
1548 1573 *
1549 1574 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1550 1575 * the gang tree, we won't modify the block, so we can safely defer the free
1551 1576 * (knowing that the block is still intact). If we *can* assemble the gang
1552 1577 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1553 1578 * each constituent bp and we can allocate a new block on the next sync pass.
1554 1579 *
1555 1580 * In all cases, the gang tree allows complete recovery from partial failure.
1556 1581 * ==========================================================================
1557 1582 */
1558 1583
1559 1584 static zio_t *
1560 1585 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1561 1586 {
1562 1587 if (gn != NULL)
1563 1588 return (pio);
1564 1589
1565 1590 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1566 1591 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1567 1592 &pio->io_bookmark));
1568 1593 }
1569 1594
1570 1595 zio_t *
1571 1596 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1572 1597 {
1573 1598 zio_t *zio;
1574 1599
1575 1600 if (gn != NULL) {
1576 1601 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1577 1602 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1578 1603 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1579 1604 /*
1580 1605 * As we rewrite each gang header, the pipeline will compute
1581 1606 * a new gang block header checksum for it; but no one will
1582 1607 * compute a new data checksum, so we do that here. The one
1583 1608 * exception is the gang leader: the pipeline already computed
1584 1609 * its data checksum because that stage precedes gang assembly.
1585 1610 * (Presently, nothing actually uses interior data checksums;
1586 1611 * this is just good hygiene.)
1587 1612 */
1588 1613 if (gn != pio->io_gang_leader->io_gang_tree) {
1589 1614 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1590 1615 data, BP_GET_PSIZE(bp));
1591 1616 }
1592 1617 /*
1593 1618 * If we are here to damage data for testing purposes,
1594 1619 * leave the GBH alone so that we can detect the damage.
1595 1620 */
1596 1621 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1597 1622 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1598 1623 } else {
1599 1624 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1600 1625 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1601 1626 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1602 1627 }
1603 1628
1604 1629 return (zio);
1605 1630 }
1606 1631
1607 1632 /* ARGSUSED */
1608 1633 zio_t *
1609 1634 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1610 1635 {
1611 1636 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1612 1637 ZIO_GANG_CHILD_FLAGS(pio)));
1613 1638 }
1614 1639
1615 1640 /* ARGSUSED */
1616 1641 zio_t *
1617 1642 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1618 1643 {
1619 1644 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1620 1645 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1621 1646 }
1622 1647
1623 1648 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1624 1649 NULL,
1625 1650 zio_read_gang,
1626 1651 zio_rewrite_gang,
1627 1652 zio_free_gang,
1628 1653 zio_claim_gang,
1629 1654 NULL
1630 1655 };
1631 1656
1632 1657 static void zio_gang_tree_assemble_done(zio_t *zio);
1633 1658
1634 1659 static zio_gang_node_t *
1635 1660 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1636 1661 {
1637 1662 zio_gang_node_t *gn;
1638 1663
1639 1664 ASSERT(*gnpp == NULL);
1640 1665
1641 1666 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1642 1667 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1643 1668 *gnpp = gn;
1644 1669
1645 1670 return (gn);
1646 1671 }
1647 1672
1648 1673 static void
1649 1674 zio_gang_node_free(zio_gang_node_t **gnpp)
1650 1675 {
1651 1676 zio_gang_node_t *gn = *gnpp;
1652 1677
1653 1678 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1654 1679 ASSERT(gn->gn_child[g] == NULL);
1655 1680
1656 1681 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1657 1682 kmem_free(gn, sizeof (*gn));
1658 1683 *gnpp = NULL;
1659 1684 }
1660 1685
1661 1686 static void
1662 1687 zio_gang_tree_free(zio_gang_node_t **gnpp)
1663 1688 {
1664 1689 zio_gang_node_t *gn = *gnpp;
1665 1690
1666 1691 if (gn == NULL)
1667 1692 return;
1668 1693
1669 1694 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1670 1695 zio_gang_tree_free(&gn->gn_child[g]);
1671 1696
1672 1697 zio_gang_node_free(gnpp);
1673 1698 }
1674 1699
1675 1700 static void
1676 1701 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1677 1702 {
1678 1703 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1679 1704
1680 1705 ASSERT(gio->io_gang_leader == gio);
1681 1706 ASSERT(BP_IS_GANG(bp));
1682 1707
1683 1708 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1684 1709 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1685 1710 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1686 1711 }
1687 1712
1688 1713 static void
1689 1714 zio_gang_tree_assemble_done(zio_t *zio)
1690 1715 {
1691 1716 zio_t *gio = zio->io_gang_leader;
1692 1717 zio_gang_node_t *gn = zio->io_private;
1693 1718 blkptr_t *bp = zio->io_bp;
1694 1719
1695 1720 ASSERT(gio == zio_unique_parent(zio));
1696 1721 ASSERT(zio->io_child_count == 0);
1697 1722
1698 1723 if (zio->io_error)
1699 1724 return;
1700 1725
1701 1726 if (BP_SHOULD_BYTESWAP(bp))
1702 1727 byteswap_uint64_array(zio->io_data, zio->io_size);
1703 1728
1704 1729 ASSERT(zio->io_data == gn->gn_gbh);
1705 1730 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1706 1731 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1707 1732
1708 1733 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1709 1734 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1710 1735 if (!BP_IS_GANG(gbp))
1711 1736 continue;
1712 1737 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1713 1738 }
1714 1739 }
1715 1740
1716 1741 static void
1717 1742 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1718 1743 {
1719 1744 zio_t *gio = pio->io_gang_leader;
1720 1745 zio_t *zio;
1721 1746
1722 1747 ASSERT(BP_IS_GANG(bp) == !!gn);
1723 1748 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1724 1749 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1725 1750
1726 1751 /*
1727 1752 * If you're a gang header, your data is in gn->gn_gbh.
1728 1753 * If you're a gang member, your data is in 'data' and gn == NULL.
1729 1754 */
1730 1755 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1731 1756
1732 1757 if (gn != NULL) {
1733 1758 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1734 1759
1735 1760 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1736 1761 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1737 1762 if (BP_IS_HOLE(gbp))
1738 1763 continue;
1739 1764 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1740 1765 data = (char *)data + BP_GET_PSIZE(gbp);
1741 1766 }
1742 1767 }
1743 1768
1744 1769 if (gn == gio->io_gang_tree)
1745 1770 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1746 1771
1747 1772 if (zio != pio)
1748 1773 zio_nowait(zio);
1749 1774 }
1750 1775
1751 1776 static int
1752 1777 zio_gang_assemble(zio_t *zio)
1753 1778 {
1754 1779 blkptr_t *bp = zio->io_bp;
1755 1780
1756 1781 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1757 1782 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1758 1783
1759 1784 zio->io_gang_leader = zio;
1760 1785
1761 1786 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1762 1787
1763 1788 return (ZIO_PIPELINE_CONTINUE);
1764 1789 }
1765 1790
1766 1791 static int
1767 1792 zio_gang_issue(zio_t *zio)
1768 1793 {
1769 1794 blkptr_t *bp = zio->io_bp;
1770 1795
1771 1796 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1772 1797 return (ZIO_PIPELINE_STOP);
1773 1798
1774 1799 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1775 1800 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1776 1801
1777 1802 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1778 1803 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1779 1804 else
1780 1805 zio_gang_tree_free(&zio->io_gang_tree);
1781 1806
1782 1807 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1783 1808
1784 1809 return (ZIO_PIPELINE_CONTINUE);
1785 1810 }
1786 1811
1787 1812 static void
1788 1813 zio_write_gang_member_ready(zio_t *zio)
1789 1814 {
1790 1815 zio_t *pio = zio_unique_parent(zio);
1791 1816 zio_t *gio = zio->io_gang_leader;
1792 1817 dva_t *cdva = zio->io_bp->blk_dva;
1793 1818 dva_t *pdva = pio->io_bp->blk_dva;
1794 1819 uint64_t asize;
1795 1820
1796 1821 if (BP_IS_HOLE(zio->io_bp))
1797 1822 return;
1798 1823
1799 1824 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1800 1825
1801 1826 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1802 1827 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1803 1828 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1804 1829 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1805 1830 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1806 1831
1807 1832 mutex_enter(&pio->io_lock);
1808 1833 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1809 1834 ASSERT(DVA_GET_GANG(&pdva[d]));
1810 1835 asize = DVA_GET_ASIZE(&pdva[d]);
1811 1836 asize += DVA_GET_ASIZE(&cdva[d]);
1812 1837 DVA_SET_ASIZE(&pdva[d], asize);
1813 1838 }
1814 1839 mutex_exit(&pio->io_lock);
1815 1840 }
1816 1841
1817 1842 static int
1818 1843 zio_write_gang_block(zio_t *pio)
1819 1844 {
1820 1845 spa_t *spa = pio->io_spa;
1821 1846 blkptr_t *bp = pio->io_bp;
1822 1847 zio_t *gio = pio->io_gang_leader;
1823 1848 zio_t *zio;
1824 1849 zio_gang_node_t *gn, **gnpp;
1825 1850 zio_gbh_phys_t *gbh;
1826 1851 uint64_t txg = pio->io_txg;
1827 1852 uint64_t resid = pio->io_size;
1828 1853 uint64_t lsize;
1829 1854 int copies = gio->io_prop.zp_copies;
1830 1855 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1831 1856 zio_prop_t zp;
1832 1857 int error;
1833 1858
1834 1859 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1835 1860 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1836 1861 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1837 1862 if (error) {
1838 1863 pio->io_error = error;
1839 1864 return (ZIO_PIPELINE_CONTINUE);
1840 1865 }
1841 1866
1842 1867 if (pio == gio) {
1843 1868 gnpp = &gio->io_gang_tree;
1844 1869 } else {
1845 1870 gnpp = pio->io_private;
1846 1871 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1847 1872 }
1848 1873
1849 1874 gn = zio_gang_node_alloc(gnpp);
1850 1875 gbh = gn->gn_gbh;
1851 1876 bzero(gbh, SPA_GANGBLOCKSIZE);
1852 1877
1853 1878 /*
1854 1879 * Create the gang header.
1855 1880 */
1856 1881 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1857 1882 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1858 1883
1859 1884 /*
1860 1885 * Create and nowait the gang children.
1861 1886 */
1862 1887 for (int g = 0; resid != 0; resid -= lsize, g++) {
1863 1888 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
↓ open down ↓ |
698 lines elided |
↑ open up ↑ |
1864 1889 SPA_MINBLOCKSIZE);
1865 1890 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1866 1891
1867 1892 zp.zp_checksum = gio->io_prop.zp_checksum;
1868 1893 zp.zp_compress = ZIO_COMPRESS_OFF;
1869 1894 zp.zp_type = DMU_OT_NONE;
1870 1895 zp.zp_level = 0;
1871 1896 zp.zp_copies = gio->io_prop.zp_copies;
1872 1897 zp.zp_dedup = B_FALSE;
1873 1898 zp.zp_dedup_verify = B_FALSE;
1899 + zp.zp_zero_write = B_FALSE;
1874 1900 zp.zp_nopwrite = B_FALSE;
1875 1901
1876 1902 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1877 1903 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1878 1904 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1879 1905 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1880 1906 &pio->io_bookmark));
1881 1907 }
1882 1908
1883 1909 /*
1884 1910 * Set pio's pipeline to just wait for zio to finish.
1885 1911 */
1886 1912 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1887 1913
1888 1914 zio_nowait(zio);
1889 1915
1890 1916 return (ZIO_PIPELINE_CONTINUE);
1891 1917 }
1892 1918
1893 1919 /*
1894 1920 * The zio_nop_write stage in the pipeline determines if allocating
1895 1921 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1896 1922 * such as SHA256, we can compare the checksums of the new data and the old
1897 1923 * to determine if allocating a new block is required. The nopwrite
1898 1924 * feature can handle writes in either syncing or open context (i.e. zil
1899 1925 * writes) and as a result is mutually exclusive with dedup.
1900 1926 */
1901 1927 static int
1902 1928 zio_nop_write(zio_t *zio)
1903 1929 {
1904 1930 blkptr_t *bp = zio->io_bp;
1905 1931 blkptr_t *bp_orig = &zio->io_bp_orig;
1906 1932 zio_prop_t *zp = &zio->io_prop;
1907 1933
1908 1934 ASSERT(BP_GET_LEVEL(bp) == 0);
1909 1935 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1910 1936 ASSERT(zp->zp_nopwrite);
1911 1937 ASSERT(!zp->zp_dedup);
1912 1938 ASSERT(zio->io_bp_override == NULL);
1913 1939 ASSERT(IO_IS_ALLOCATING(zio));
1914 1940
1915 1941 /*
1916 1942 * Check to see if the original bp and the new bp have matching
1917 1943 * characteristics (i.e. same checksum, compression algorithms, etc).
1918 1944 * If they don't then just continue with the pipeline which will
1919 1945 * allocate a new bp.
1920 1946 */
1921 1947 if (BP_IS_HOLE(bp_orig) ||
1922 1948 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1923 1949 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1924 1950 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1925 1951 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1926 1952 zp->zp_copies != BP_GET_NDVAS(bp_orig))
1927 1953 return (ZIO_PIPELINE_CONTINUE);
1928 1954
1929 1955 /*
1930 1956 * If the checksums match then reset the pipeline so that we
1931 1957 * avoid allocating a new bp and issuing any I/O.
1932 1958 */
1933 1959 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1934 1960 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1935 1961 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1936 1962 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1937 1963 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1938 1964 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1939 1965 sizeof (uint64_t)) == 0);
1940 1966
1941 1967 *bp = *bp_orig;
1942 1968 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1943 1969 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1944 1970 }
1945 1971
1946 1972 return (ZIO_PIPELINE_CONTINUE);
1947 1973 }
1948 1974
1949 1975 /*
1950 1976 * ==========================================================================
1951 1977 * Dedup
1952 1978 * ==========================================================================
1953 1979 */
1954 1980 static void
1955 1981 zio_ddt_child_read_done(zio_t *zio)
1956 1982 {
1957 1983 blkptr_t *bp = zio->io_bp;
1958 1984 ddt_entry_t *dde = zio->io_private;
1959 1985 ddt_phys_t *ddp;
1960 1986 zio_t *pio = zio_unique_parent(zio);
1961 1987
1962 1988 mutex_enter(&pio->io_lock);
1963 1989 ddp = ddt_phys_select(dde, bp);
1964 1990 if (zio->io_error == 0)
1965 1991 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1966 1992 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1967 1993 dde->dde_repair_data = zio->io_data;
1968 1994 else
1969 1995 zio_buf_free(zio->io_data, zio->io_size);
1970 1996 mutex_exit(&pio->io_lock);
1971 1997 }
1972 1998
1973 1999 static int
1974 2000 zio_ddt_read_start(zio_t *zio)
1975 2001 {
1976 2002 blkptr_t *bp = zio->io_bp;
1977 2003
1978 2004 ASSERT(BP_GET_DEDUP(bp));
1979 2005 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1980 2006 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1981 2007
1982 2008 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1983 2009 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1984 2010 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1985 2011 ddt_phys_t *ddp = dde->dde_phys;
1986 2012 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1987 2013 blkptr_t blk;
1988 2014
1989 2015 ASSERT(zio->io_vsd == NULL);
1990 2016 zio->io_vsd = dde;
1991 2017
1992 2018 if (ddp_self == NULL)
1993 2019 return (ZIO_PIPELINE_CONTINUE);
1994 2020
1995 2021 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1996 2022 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1997 2023 continue;
1998 2024 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1999 2025 &blk);
2000 2026 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2001 2027 zio_buf_alloc(zio->io_size), zio->io_size,
2002 2028 zio_ddt_child_read_done, dde, zio->io_priority,
2003 2029 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2004 2030 &zio->io_bookmark));
2005 2031 }
2006 2032 return (ZIO_PIPELINE_CONTINUE);
2007 2033 }
2008 2034
2009 2035 zio_nowait(zio_read(zio, zio->io_spa, bp,
2010 2036 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2011 2037 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2012 2038
2013 2039 return (ZIO_PIPELINE_CONTINUE);
2014 2040 }
2015 2041
2016 2042 static int
2017 2043 zio_ddt_read_done(zio_t *zio)
2018 2044 {
2019 2045 blkptr_t *bp = zio->io_bp;
2020 2046
2021 2047 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2022 2048 return (ZIO_PIPELINE_STOP);
2023 2049
2024 2050 ASSERT(BP_GET_DEDUP(bp));
2025 2051 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2026 2052 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2027 2053
2028 2054 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2029 2055 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2030 2056 ddt_entry_t *dde = zio->io_vsd;
2031 2057 if (ddt == NULL) {
2032 2058 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2033 2059 return (ZIO_PIPELINE_CONTINUE);
2034 2060 }
2035 2061 if (dde == NULL) {
2036 2062 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2037 2063 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2038 2064 return (ZIO_PIPELINE_STOP);
2039 2065 }
2040 2066 if (dde->dde_repair_data != NULL) {
2041 2067 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2042 2068 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2043 2069 }
2044 2070 ddt_repair_done(ddt, dde);
2045 2071 zio->io_vsd = NULL;
2046 2072 }
2047 2073
2048 2074 ASSERT(zio->io_vsd == NULL);
2049 2075
2050 2076 return (ZIO_PIPELINE_CONTINUE);
2051 2077 }
2052 2078
2053 2079 static boolean_t
2054 2080 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2055 2081 {
2056 2082 spa_t *spa = zio->io_spa;
2057 2083
2058 2084 /*
2059 2085 * Note: we compare the original data, not the transformed data,
2060 2086 * because when zio->io_bp is an override bp, we will not have
2061 2087 * pushed the I/O transforms. That's an important optimization
2062 2088 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2063 2089 */
2064 2090 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2065 2091 zio_t *lio = dde->dde_lead_zio[p];
2066 2092
2067 2093 if (lio != NULL) {
2068 2094 return (lio->io_orig_size != zio->io_orig_size ||
2069 2095 bcmp(zio->io_orig_data, lio->io_orig_data,
2070 2096 zio->io_orig_size) != 0);
2071 2097 }
2072 2098 }
2073 2099
2074 2100 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2075 2101 ddt_phys_t *ddp = &dde->dde_phys[p];
2076 2102
2077 2103 if (ddp->ddp_phys_birth != 0) {
2078 2104 arc_buf_t *abuf = NULL;
2079 2105 uint32_t aflags = ARC_WAIT;
2080 2106 blkptr_t blk = *zio->io_bp;
2081 2107 int error;
2082 2108
2083 2109 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2084 2110
2085 2111 ddt_exit(ddt);
2086 2112
2087 2113 error = arc_read(NULL, spa, &blk,
2088 2114 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2089 2115 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2090 2116 &aflags, &zio->io_bookmark);
2091 2117
2092 2118 if (error == 0) {
2093 2119 if (arc_buf_size(abuf) != zio->io_orig_size ||
2094 2120 bcmp(abuf->b_data, zio->io_orig_data,
2095 2121 zio->io_orig_size) != 0)
2096 2122 error = SET_ERROR(EEXIST);
2097 2123 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2098 2124 }
2099 2125
2100 2126 ddt_enter(ddt);
2101 2127 return (error != 0);
2102 2128 }
2103 2129 }
2104 2130
2105 2131 return (B_FALSE);
2106 2132 }
2107 2133
2108 2134 static void
2109 2135 zio_ddt_child_write_ready(zio_t *zio)
2110 2136 {
2111 2137 int p = zio->io_prop.zp_copies;
2112 2138 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2113 2139 ddt_entry_t *dde = zio->io_private;
2114 2140 ddt_phys_t *ddp = &dde->dde_phys[p];
2115 2141 zio_t *pio;
2116 2142
2117 2143 if (zio->io_error)
2118 2144 return;
2119 2145
2120 2146 ddt_enter(ddt);
2121 2147
2122 2148 ASSERT(dde->dde_lead_zio[p] == zio);
2123 2149
2124 2150 ddt_phys_fill(ddp, zio->io_bp);
2125 2151
2126 2152 while ((pio = zio_walk_parents(zio)) != NULL)
2127 2153 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2128 2154
2129 2155 ddt_exit(ddt);
2130 2156 }
2131 2157
2132 2158 static void
2133 2159 zio_ddt_child_write_done(zio_t *zio)
2134 2160 {
2135 2161 int p = zio->io_prop.zp_copies;
2136 2162 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2137 2163 ddt_entry_t *dde = zio->io_private;
2138 2164 ddt_phys_t *ddp = &dde->dde_phys[p];
2139 2165
2140 2166 ddt_enter(ddt);
2141 2167
2142 2168 ASSERT(ddp->ddp_refcnt == 0);
2143 2169 ASSERT(dde->dde_lead_zio[p] == zio);
2144 2170 dde->dde_lead_zio[p] = NULL;
2145 2171
2146 2172 if (zio->io_error == 0) {
2147 2173 while (zio_walk_parents(zio) != NULL)
2148 2174 ddt_phys_addref(ddp);
2149 2175 } else {
2150 2176 ddt_phys_clear(ddp);
2151 2177 }
2152 2178
2153 2179 ddt_exit(ddt);
2154 2180 }
2155 2181
2156 2182 static void
2157 2183 zio_ddt_ditto_write_done(zio_t *zio)
2158 2184 {
2159 2185 int p = DDT_PHYS_DITTO;
2160 2186 zio_prop_t *zp = &zio->io_prop;
2161 2187 blkptr_t *bp = zio->io_bp;
2162 2188 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2163 2189 ddt_entry_t *dde = zio->io_private;
2164 2190 ddt_phys_t *ddp = &dde->dde_phys[p];
2165 2191 ddt_key_t *ddk = &dde->dde_key;
2166 2192
2167 2193 ddt_enter(ddt);
2168 2194
2169 2195 ASSERT(ddp->ddp_refcnt == 0);
2170 2196 ASSERT(dde->dde_lead_zio[p] == zio);
2171 2197 dde->dde_lead_zio[p] = NULL;
2172 2198
2173 2199 if (zio->io_error == 0) {
2174 2200 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2175 2201 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2176 2202 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2177 2203 if (ddp->ddp_phys_birth != 0)
2178 2204 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2179 2205 ddt_phys_fill(ddp, bp);
2180 2206 }
2181 2207
2182 2208 ddt_exit(ddt);
2183 2209 }
2184 2210
2185 2211 static int
2186 2212 zio_ddt_write(zio_t *zio)
2187 2213 {
2188 2214 spa_t *spa = zio->io_spa;
2189 2215 blkptr_t *bp = zio->io_bp;
2190 2216 uint64_t txg = zio->io_txg;
2191 2217 zio_prop_t *zp = &zio->io_prop;
2192 2218 int p = zp->zp_copies;
2193 2219 int ditto_copies;
2194 2220 zio_t *cio = NULL;
2195 2221 zio_t *dio = NULL;
2196 2222 ddt_t *ddt = ddt_select(spa, bp);
2197 2223 ddt_entry_t *dde;
2198 2224 ddt_phys_t *ddp;
2199 2225
2200 2226 ASSERT(BP_GET_DEDUP(bp));
2201 2227 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2202 2228 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2203 2229
2204 2230 ddt_enter(ddt);
2205 2231 dde = ddt_lookup(ddt, bp, B_TRUE);
2206 2232 ddp = &dde->dde_phys[p];
2207 2233
2208 2234 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2209 2235 /*
2210 2236 * If we're using a weak checksum, upgrade to a strong checksum
2211 2237 * and try again. If we're already using a strong checksum,
2212 2238 * we can't resolve it, so just convert to an ordinary write.
2213 2239 * (And automatically e-mail a paper to Nature?)
2214 2240 */
2215 2241 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2216 2242 zp->zp_checksum = spa_dedup_checksum(spa);
2217 2243 zio_pop_transforms(zio);
2218 2244 zio->io_stage = ZIO_STAGE_OPEN;
2219 2245 BP_ZERO(bp);
2220 2246 } else {
2221 2247 zp->zp_dedup = B_FALSE;
2222 2248 }
2223 2249 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2224 2250 ddt_exit(ddt);
2225 2251 return (ZIO_PIPELINE_CONTINUE);
2226 2252 }
2227 2253
2228 2254 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2229 2255 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2230 2256
2231 2257 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2232 2258 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2233 2259 zio_prop_t czp = *zp;
2234 2260
2235 2261 czp.zp_copies = ditto_copies;
2236 2262
2237 2263 /*
2238 2264 * If we arrived here with an override bp, we won't have run
2239 2265 * the transform stack, so we won't have the data we need to
2240 2266 * generate a child i/o. So, toss the override bp and restart.
2241 2267 * This is safe, because using the override bp is just an
2242 2268 * optimization; and it's rare, so the cost doesn't matter.
2243 2269 */
2244 2270 if (zio->io_bp_override) {
2245 2271 zio_pop_transforms(zio);
2246 2272 zio->io_stage = ZIO_STAGE_OPEN;
2247 2273 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2248 2274 zio->io_bp_override = NULL;
2249 2275 BP_ZERO(bp);
2250 2276 ddt_exit(ddt);
2251 2277 return (ZIO_PIPELINE_CONTINUE);
2252 2278 }
2253 2279
2254 2280 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2255 2281 zio->io_orig_size, &czp, NULL, NULL,
2256 2282 zio_ddt_ditto_write_done, dde, zio->io_priority,
2257 2283 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2258 2284
2259 2285 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2260 2286 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2261 2287 }
2262 2288
2263 2289 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2264 2290 if (ddp->ddp_phys_birth != 0)
2265 2291 ddt_bp_fill(ddp, bp, txg);
2266 2292 if (dde->dde_lead_zio[p] != NULL)
2267 2293 zio_add_child(zio, dde->dde_lead_zio[p]);
2268 2294 else
2269 2295 ddt_phys_addref(ddp);
2270 2296 } else if (zio->io_bp_override) {
2271 2297 ASSERT(bp->blk_birth == txg);
2272 2298 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2273 2299 ddt_phys_fill(ddp, bp);
2274 2300 ddt_phys_addref(ddp);
2275 2301 } else {
2276 2302 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2277 2303 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2278 2304 zio_ddt_child_write_done, dde, zio->io_priority,
2279 2305 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2280 2306
2281 2307 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2282 2308 dde->dde_lead_zio[p] = cio;
2283 2309 }
2284 2310
2285 2311 ddt_exit(ddt);
2286 2312
2287 2313 if (cio)
2288 2314 zio_nowait(cio);
2289 2315 if (dio)
2290 2316 zio_nowait(dio);
2291 2317
2292 2318 return (ZIO_PIPELINE_CONTINUE);
2293 2319 }
2294 2320
2295 2321 ddt_entry_t *freedde; /* for debugging */
2296 2322
2297 2323 static int
2298 2324 zio_ddt_free(zio_t *zio)
2299 2325 {
2300 2326 spa_t *spa = zio->io_spa;
2301 2327 blkptr_t *bp = zio->io_bp;
2302 2328 ddt_t *ddt = ddt_select(spa, bp);
2303 2329 ddt_entry_t *dde;
2304 2330 ddt_phys_t *ddp;
2305 2331
2306 2332 ASSERT(BP_GET_DEDUP(bp));
2307 2333 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2308 2334
2309 2335 ddt_enter(ddt);
2310 2336 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2311 2337 ddp = ddt_phys_select(dde, bp);
2312 2338 ddt_phys_decref(ddp);
2313 2339 ddt_exit(ddt);
2314 2340
2315 2341 return (ZIO_PIPELINE_CONTINUE);
2316 2342 }
2317 2343
2318 2344 /*
2319 2345 * ==========================================================================
2320 2346 * Allocate and free blocks
2321 2347 * ==========================================================================
2322 2348 */
2323 2349 static int
2324 2350 zio_dva_allocate(zio_t *zio)
2325 2351 {
2326 2352 spa_t *spa = zio->io_spa;
2327 2353 metaslab_class_t *mc = spa_normal_class(spa);
2328 2354 blkptr_t *bp = zio->io_bp;
2329 2355 int error;
2330 2356 int flags = 0;
2331 2357
2332 2358 if (zio->io_gang_leader == NULL) {
2333 2359 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2334 2360 zio->io_gang_leader = zio;
2335 2361 }
2336 2362
2337 2363 ASSERT(BP_IS_HOLE(bp));
2338 2364 ASSERT0(BP_GET_NDVAS(bp));
2339 2365 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2340 2366 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2341 2367 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2342 2368
2343 2369 /*
2344 2370 * The dump device does not support gang blocks so allocation on
2345 2371 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2346 2372 * the "fast" gang feature.
2347 2373 */
2348 2374 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2349 2375 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2350 2376 METASLAB_GANG_CHILD : 0;
2351 2377 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2352 2378 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2353 2379
2354 2380 if (error) {
2355 2381 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2356 2382 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2357 2383 error);
2358 2384 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2359 2385 return (zio_write_gang_block(zio));
2360 2386 zio->io_error = error;
2361 2387 }
2362 2388
2363 2389 return (ZIO_PIPELINE_CONTINUE);
2364 2390 }
2365 2391
2366 2392 static int
2367 2393 zio_dva_free(zio_t *zio)
2368 2394 {
2369 2395 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2370 2396
2371 2397 return (ZIO_PIPELINE_CONTINUE);
2372 2398 }
2373 2399
2374 2400 static int
2375 2401 zio_dva_claim(zio_t *zio)
2376 2402 {
2377 2403 int error;
2378 2404
2379 2405 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2380 2406 if (error)
2381 2407 zio->io_error = error;
2382 2408
2383 2409 return (ZIO_PIPELINE_CONTINUE);
2384 2410 }
2385 2411
2386 2412 /*
2387 2413 * Undo an allocation. This is used by zio_done() when an I/O fails
2388 2414 * and we want to give back the block we just allocated.
2389 2415 * This handles both normal blocks and gang blocks.
2390 2416 */
2391 2417 static void
2392 2418 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2393 2419 {
2394 2420 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2395 2421 ASSERT(zio->io_bp_override == NULL);
2396 2422
2397 2423 if (!BP_IS_HOLE(bp))
2398 2424 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2399 2425
2400 2426 if (gn != NULL) {
2401 2427 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2402 2428 zio_dva_unallocate(zio, gn->gn_child[g],
2403 2429 &gn->gn_gbh->zg_blkptr[g]);
2404 2430 }
2405 2431 }
2406 2432 }
2407 2433
2408 2434 /*
2409 2435 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2410 2436 */
2411 2437 int
2412 2438 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2413 2439 uint64_t size, boolean_t use_slog)
2414 2440 {
2415 2441 int error = 1;
2416 2442
2417 2443 ASSERT(txg > spa_syncing_txg(spa));
2418 2444
2419 2445 /*
2420 2446 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2421 2447 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2422 2448 * when allocating them.
2423 2449 */
2424 2450 if (use_slog) {
2425 2451 error = metaslab_alloc(spa, spa_log_class(spa), size,
2426 2452 new_bp, 1, txg, old_bp,
2427 2453 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2428 2454 }
2429 2455
2430 2456 if (error) {
2431 2457 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2432 2458 new_bp, 1, txg, old_bp,
2433 2459 METASLAB_HINTBP_AVOID);
2434 2460 }
2435 2461
2436 2462 if (error == 0) {
2437 2463 BP_SET_LSIZE(new_bp, size);
2438 2464 BP_SET_PSIZE(new_bp, size);
2439 2465 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2440 2466 BP_SET_CHECKSUM(new_bp,
2441 2467 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2442 2468 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2443 2469 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2444 2470 BP_SET_LEVEL(new_bp, 0);
2445 2471 BP_SET_DEDUP(new_bp, 0);
2446 2472 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2447 2473 }
2448 2474
2449 2475 return (error);
2450 2476 }
2451 2477
2452 2478 /*
2453 2479 * Free an intent log block.
2454 2480 */
2455 2481 void
2456 2482 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2457 2483 {
2458 2484 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2459 2485 ASSERT(!BP_IS_GANG(bp));
2460 2486
2461 2487 zio_free(spa, txg, bp);
2462 2488 }
2463 2489
2464 2490 /*
2465 2491 * ==========================================================================
2466 2492 * Read and write to physical devices
2467 2493 * ==========================================================================
2468 2494 */
2469 2495 static int
2470 2496 zio_vdev_io_start(zio_t *zio)
2471 2497 {
2472 2498 vdev_t *vd = zio->io_vd;
2473 2499 uint64_t align;
2474 2500 spa_t *spa = zio->io_spa;
2475 2501
2476 2502 ASSERT(zio->io_error == 0);
2477 2503 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2478 2504
2479 2505 if (vd == NULL) {
2480 2506 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2481 2507 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2482 2508
2483 2509 /*
2484 2510 * The mirror_ops handle multiple DVAs in a single BP.
2485 2511 */
2486 2512 return (vdev_mirror_ops.vdev_op_io_start(zio));
2487 2513 }
2488 2514
2489 2515 /*
2490 2516 * We keep track of time-sensitive I/Os so that the scan thread
2491 2517 * can quickly react to certain workloads. In particular, we care
2492 2518 * about non-scrubbing, top-level reads and writes with the following
2493 2519 * characteristics:
2494 2520 * - synchronous writes of user data to non-slog devices
2495 2521 * - any reads of user data
2496 2522 * When these conditions are met, adjust the timestamp of spa_last_io
2497 2523 * which allows the scan thread to adjust its workload accordingly.
2498 2524 */
2499 2525 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2500 2526 vd == vd->vdev_top && !vd->vdev_islog &&
2501 2527 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2502 2528 zio->io_txg != spa_syncing_txg(spa)) {
2503 2529 uint64_t old = spa->spa_last_io;
2504 2530 uint64_t new = ddi_get_lbolt64();
2505 2531 if (old != new)
2506 2532 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2507 2533 }
2508 2534
2509 2535 align = 1ULL << vd->vdev_top->vdev_ashift;
2510 2536
2511 2537 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
2512 2538 P2PHASE(zio->io_size, align) != 0) {
2513 2539 /* Transform logical writes to be a full physical block size. */
2514 2540 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2515 2541 char *abuf = zio_buf_alloc(asize);
2516 2542 ASSERT(vd == vd->vdev_top);
2517 2543 if (zio->io_type == ZIO_TYPE_WRITE) {
2518 2544 bcopy(zio->io_data, abuf, zio->io_size);
2519 2545 bzero(abuf + zio->io_size, asize - zio->io_size);
2520 2546 }
2521 2547 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2522 2548 }
2523 2549
2524 2550 /*
2525 2551 * If this is not a physical io, make sure that it is properly aligned
2526 2552 * before proceeding.
2527 2553 */
2528 2554 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2529 2555 ASSERT0(P2PHASE(zio->io_offset, align));
2530 2556 ASSERT0(P2PHASE(zio->io_size, align));
2531 2557 } else {
2532 2558 /*
2533 2559 * For physical writes, we allow 512b aligned writes and assume
2534 2560 * the device will perform a read-modify-write as necessary.
2535 2561 */
2536 2562 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2537 2563 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2538 2564 }
2539 2565
2540 2566 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2541 2567
2542 2568 /*
2543 2569 * If this is a repair I/O, and there's no self-healing involved --
2544 2570 * that is, we're just resilvering what we expect to resilver --
2545 2571 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2546 2572 * This prevents spurious resilvering with nested replication.
2547 2573 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2548 2574 * A is out of date, we'll read from C+D, then use the data to
2549 2575 * resilver A+B -- but we don't actually want to resilver B, just A.
2550 2576 * The top-level mirror has no way to know this, so instead we just
2551 2577 * discard unnecessary repairs as we work our way down the vdev tree.
2552 2578 * The same logic applies to any form of nested replication:
2553 2579 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2554 2580 */
2555 2581 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2556 2582 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2557 2583 zio->io_txg != 0 && /* not a delegated i/o */
2558 2584 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2559 2585 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2560 2586 zio_vdev_io_bypass(zio);
2561 2587 return (ZIO_PIPELINE_CONTINUE);
2562 2588 }
2563 2589
2564 2590 if (vd->vdev_ops->vdev_op_leaf &&
2565 2591 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2566 2592
2567 2593 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
2568 2594 return (ZIO_PIPELINE_CONTINUE);
2569 2595
2570 2596 if ((zio = vdev_queue_io(zio)) == NULL)
2571 2597 return (ZIO_PIPELINE_STOP);
2572 2598
2573 2599 if (!vdev_accessible(vd, zio)) {
2574 2600 zio->io_error = SET_ERROR(ENXIO);
2575 2601 zio_interrupt(zio);
2576 2602 return (ZIO_PIPELINE_STOP);
2577 2603 }
2578 2604 }
2579 2605
2580 2606 return (vd->vdev_ops->vdev_op_io_start(zio));
2581 2607 }
2582 2608
2583 2609 static int
2584 2610 zio_vdev_io_done(zio_t *zio)
2585 2611 {
2586 2612 vdev_t *vd = zio->io_vd;
2587 2613 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2588 2614 boolean_t unexpected_error = B_FALSE;
2589 2615
2590 2616 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2591 2617 return (ZIO_PIPELINE_STOP);
2592 2618
2593 2619 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2594 2620
2595 2621 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2596 2622
2597 2623 vdev_queue_io_done(zio);
2598 2624
2599 2625 if (zio->io_type == ZIO_TYPE_WRITE)
2600 2626 vdev_cache_write(zio);
2601 2627
2602 2628 if (zio_injection_enabled && zio->io_error == 0)
2603 2629 zio->io_error = zio_handle_device_injection(vd,
2604 2630 zio, EIO);
2605 2631
2606 2632 if (zio_injection_enabled && zio->io_error == 0)
2607 2633 zio->io_error = zio_handle_label_injection(zio, EIO);
2608 2634
2609 2635 if (zio->io_error) {
2610 2636 if (!vdev_accessible(vd, zio)) {
2611 2637 zio->io_error = SET_ERROR(ENXIO);
2612 2638 } else {
2613 2639 unexpected_error = B_TRUE;
2614 2640 }
2615 2641 }
2616 2642 }
2617 2643
2618 2644 ops->vdev_op_io_done(zio);
2619 2645
2620 2646 if (unexpected_error)
2621 2647 VERIFY(vdev_probe(vd, zio) == NULL);
2622 2648
2623 2649 return (ZIO_PIPELINE_CONTINUE);
2624 2650 }
2625 2651
2626 2652 /*
2627 2653 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2628 2654 * disk, and use that to finish the checksum ereport later.
2629 2655 */
2630 2656 static void
2631 2657 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2632 2658 const void *good_buf)
2633 2659 {
2634 2660 /* no processing needed */
2635 2661 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2636 2662 }
2637 2663
2638 2664 /*ARGSUSED*/
2639 2665 void
2640 2666 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2641 2667 {
2642 2668 void *buf = zio_buf_alloc(zio->io_size);
2643 2669
2644 2670 bcopy(zio->io_data, buf, zio->io_size);
2645 2671
2646 2672 zcr->zcr_cbinfo = zio->io_size;
2647 2673 zcr->zcr_cbdata = buf;
2648 2674 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2649 2675 zcr->zcr_free = zio_buf_free;
2650 2676 }
2651 2677
2652 2678 static int
2653 2679 zio_vdev_io_assess(zio_t *zio)
2654 2680 {
2655 2681 vdev_t *vd = zio->io_vd;
2656 2682
2657 2683 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2658 2684 return (ZIO_PIPELINE_STOP);
2659 2685
2660 2686 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2661 2687 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2662 2688
2663 2689 if (zio->io_vsd != NULL) {
2664 2690 zio->io_vsd_ops->vsd_free(zio);
2665 2691 zio->io_vsd = NULL;
2666 2692 }
2667 2693
2668 2694 if (zio_injection_enabled && zio->io_error == 0)
2669 2695 zio->io_error = zio_handle_fault_injection(zio, EIO);
2670 2696
2671 2697 /*
2672 2698 * If the I/O failed, determine whether we should attempt to retry it.
2673 2699 *
2674 2700 * On retry, we cut in line in the issue queue, since we don't want
2675 2701 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2676 2702 */
2677 2703 if (zio->io_error && vd == NULL &&
2678 2704 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2679 2705 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2680 2706 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2681 2707 zio->io_error = 0;
2682 2708 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2683 2709 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2684 2710 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2685 2711 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2686 2712 zio_requeue_io_start_cut_in_line);
2687 2713 return (ZIO_PIPELINE_STOP);
2688 2714 }
2689 2715
2690 2716 /*
2691 2717 * If we got an error on a leaf device, convert it to ENXIO
2692 2718 * if the device is not accessible at all.
2693 2719 */
2694 2720 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2695 2721 !vdev_accessible(vd, zio))
2696 2722 zio->io_error = SET_ERROR(ENXIO);
2697 2723
2698 2724 /*
2699 2725 * If we can't write to an interior vdev (mirror or RAID-Z),
2700 2726 * set vdev_cant_write so that we stop trying to allocate from it.
2701 2727 */
2702 2728 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2703 2729 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2704 2730 vd->vdev_cant_write = B_TRUE;
2705 2731 }
2706 2732
2707 2733 if (zio->io_error)
2708 2734 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2709 2735
2710 2736 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2711 2737 zio->io_physdone != NULL) {
2712 2738 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2713 2739 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2714 2740 zio->io_physdone(zio->io_logical);
2715 2741 }
2716 2742
2717 2743 return (ZIO_PIPELINE_CONTINUE);
2718 2744 }
2719 2745
2720 2746 void
2721 2747 zio_vdev_io_reissue(zio_t *zio)
2722 2748 {
2723 2749 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2724 2750 ASSERT(zio->io_error == 0);
2725 2751
2726 2752 zio->io_stage >>= 1;
2727 2753 }
2728 2754
2729 2755 void
2730 2756 zio_vdev_io_redone(zio_t *zio)
2731 2757 {
2732 2758 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2733 2759
2734 2760 zio->io_stage >>= 1;
2735 2761 }
2736 2762
2737 2763 void
2738 2764 zio_vdev_io_bypass(zio_t *zio)
2739 2765 {
2740 2766 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2741 2767 ASSERT(zio->io_error == 0);
2742 2768
2743 2769 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2744 2770 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2745 2771 }
2746 2772
2747 2773 /*
2748 2774 * ==========================================================================
2749 2775 * Generate and verify checksums
2750 2776 * ==========================================================================
2751 2777 */
2752 2778 static int
2753 2779 zio_checksum_generate(zio_t *zio)
2754 2780 {
2755 2781 blkptr_t *bp = zio->io_bp;
2756 2782 enum zio_checksum checksum;
2757 2783
2758 2784 if (bp == NULL) {
2759 2785 /*
2760 2786 * This is zio_write_phys().
2761 2787 * We're either generating a label checksum, or none at all.
2762 2788 */
2763 2789 checksum = zio->io_prop.zp_checksum;
2764 2790
2765 2791 if (checksum == ZIO_CHECKSUM_OFF)
2766 2792 return (ZIO_PIPELINE_CONTINUE);
2767 2793
2768 2794 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2769 2795 } else {
2770 2796 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2771 2797 ASSERT(!IO_IS_ALLOCATING(zio));
2772 2798 checksum = ZIO_CHECKSUM_GANG_HEADER;
2773 2799 } else {
2774 2800 checksum = BP_GET_CHECKSUM(bp);
2775 2801 }
2776 2802 }
2777 2803
2778 2804 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2779 2805
2780 2806 return (ZIO_PIPELINE_CONTINUE);
2781 2807 }
2782 2808
2783 2809 static int
2784 2810 zio_checksum_verify(zio_t *zio)
2785 2811 {
2786 2812 zio_bad_cksum_t info;
2787 2813 blkptr_t *bp = zio->io_bp;
2788 2814 int error;
2789 2815
2790 2816 ASSERT(zio->io_vd != NULL);
2791 2817
2792 2818 if (bp == NULL) {
2793 2819 /*
2794 2820 * This is zio_read_phys().
2795 2821 * We're either verifying a label checksum, or nothing at all.
2796 2822 */
2797 2823 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2798 2824 return (ZIO_PIPELINE_CONTINUE);
2799 2825
2800 2826 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2801 2827 }
2802 2828
2803 2829 if ((error = zio_checksum_error(zio, &info)) != 0) {
2804 2830 zio->io_error = error;
2805 2831 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2806 2832 zfs_ereport_start_checksum(zio->io_spa,
2807 2833 zio->io_vd, zio, zio->io_offset,
2808 2834 zio->io_size, NULL, &info);
2809 2835 }
2810 2836 }
2811 2837
2812 2838 return (ZIO_PIPELINE_CONTINUE);
2813 2839 }
2814 2840
2815 2841 /*
2816 2842 * Called by RAID-Z to ensure we don't compute the checksum twice.
2817 2843 */
2818 2844 void
2819 2845 zio_checksum_verified(zio_t *zio)
2820 2846 {
2821 2847 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2822 2848 }
2823 2849
2824 2850 /*
2825 2851 * ==========================================================================
2826 2852 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2827 2853 * An error of 0 indicates success. ENXIO indicates whole-device failure,
2828 2854 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2829 2855 * indicate errors that are specific to one I/O, and most likely permanent.
2830 2856 * Any other error is presumed to be worse because we weren't expecting it.
2831 2857 * ==========================================================================
2832 2858 */
2833 2859 int
2834 2860 zio_worst_error(int e1, int e2)
2835 2861 {
2836 2862 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2837 2863 int r1, r2;
2838 2864
2839 2865 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2840 2866 if (e1 == zio_error_rank[r1])
2841 2867 break;
2842 2868
2843 2869 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2844 2870 if (e2 == zio_error_rank[r2])
2845 2871 break;
2846 2872
2847 2873 return (r1 > r2 ? e1 : e2);
2848 2874 }
2849 2875
2850 2876 /*
2851 2877 * ==========================================================================
2852 2878 * I/O completion
2853 2879 * ==========================================================================
2854 2880 */
2855 2881 static int
2856 2882 zio_ready(zio_t *zio)
2857 2883 {
2858 2884 blkptr_t *bp = zio->io_bp;
2859 2885 zio_t *pio, *pio_next;
2860 2886
2861 2887 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2862 2888 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2863 2889 return (ZIO_PIPELINE_STOP);
2864 2890
2865 2891 if (zio->io_ready) {
2866 2892 ASSERT(IO_IS_ALLOCATING(zio));
2867 2893 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2868 2894 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2869 2895 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2870 2896
2871 2897 zio->io_ready(zio);
2872 2898 }
2873 2899
2874 2900 if (bp != NULL && bp != &zio->io_bp_copy)
2875 2901 zio->io_bp_copy = *bp;
2876 2902
2877 2903 if (zio->io_error)
2878 2904 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2879 2905
2880 2906 mutex_enter(&zio->io_lock);
2881 2907 zio->io_state[ZIO_WAIT_READY] = 1;
2882 2908 pio = zio_walk_parents(zio);
2883 2909 mutex_exit(&zio->io_lock);
2884 2910
2885 2911 /*
2886 2912 * As we notify zio's parents, new parents could be added.
2887 2913 * New parents go to the head of zio's io_parent_list, however,
2888 2914 * so we will (correctly) not notify them. The remainder of zio's
2889 2915 * io_parent_list, from 'pio_next' onward, cannot change because
2890 2916 * all parents must wait for us to be done before they can be done.
2891 2917 */
2892 2918 for (; pio != NULL; pio = pio_next) {
2893 2919 pio_next = zio_walk_parents(zio);
2894 2920 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2895 2921 }
2896 2922
2897 2923 if (zio->io_flags & ZIO_FLAG_NODATA) {
2898 2924 if (BP_IS_GANG(bp)) {
2899 2925 zio->io_flags &= ~ZIO_FLAG_NODATA;
2900 2926 } else {
2901 2927 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2902 2928 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2903 2929 }
2904 2930 }
2905 2931
2906 2932 if (zio_injection_enabled &&
2907 2933 zio->io_spa->spa_syncing_txg == zio->io_txg)
2908 2934 zio_handle_ignored_writes(zio);
2909 2935
2910 2936 return (ZIO_PIPELINE_CONTINUE);
2911 2937 }
2912 2938
2913 2939 static int
2914 2940 zio_done(zio_t *zio)
2915 2941 {
2916 2942 spa_t *spa = zio->io_spa;
2917 2943 zio_t *lio = zio->io_logical;
2918 2944 blkptr_t *bp = zio->io_bp;
2919 2945 vdev_t *vd = zio->io_vd;
2920 2946 uint64_t psize = zio->io_size;
2921 2947 zio_t *pio, *pio_next;
2922 2948
2923 2949 /*
2924 2950 * If our children haven't all completed,
2925 2951 * wait for them and then repeat this pipeline stage.
2926 2952 */
2927 2953 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2928 2954 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2929 2955 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2930 2956 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2931 2957 return (ZIO_PIPELINE_STOP);
2932 2958
2933 2959 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2934 2960 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2935 2961 ASSERT(zio->io_children[c][w] == 0);
2936 2962
2937 2963 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
2938 2964 ASSERT(bp->blk_pad[0] == 0);
2939 2965 ASSERT(bp->blk_pad[1] == 0);
2940 2966 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2941 2967 (bp == zio_unique_parent(zio)->io_bp));
2942 2968 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2943 2969 zio->io_bp_override == NULL &&
2944 2970 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2945 2971 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2946 2972 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2947 2973 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2948 2974 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2949 2975 }
2950 2976 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
2951 2977 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
2952 2978 }
2953 2979
2954 2980 /*
2955 2981 * If there were child vdev/gang/ddt errors, they apply to us now.
2956 2982 */
2957 2983 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2958 2984 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2959 2985 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2960 2986
2961 2987 /*
2962 2988 * If the I/O on the transformed data was successful, generate any
2963 2989 * checksum reports now while we still have the transformed data.
2964 2990 */
2965 2991 if (zio->io_error == 0) {
2966 2992 while (zio->io_cksum_report != NULL) {
2967 2993 zio_cksum_report_t *zcr = zio->io_cksum_report;
2968 2994 uint64_t align = zcr->zcr_align;
2969 2995 uint64_t asize = P2ROUNDUP(psize, align);
2970 2996 char *abuf = zio->io_data;
2971 2997
2972 2998 if (asize != psize) {
2973 2999 abuf = zio_buf_alloc(asize);
2974 3000 bcopy(zio->io_data, abuf, psize);
2975 3001 bzero(abuf + psize, asize - psize);
2976 3002 }
2977 3003
2978 3004 zio->io_cksum_report = zcr->zcr_next;
2979 3005 zcr->zcr_next = NULL;
2980 3006 zcr->zcr_finish(zcr, abuf);
2981 3007 zfs_ereport_free_checksum(zcr);
2982 3008
2983 3009 if (asize != psize)
2984 3010 zio_buf_free(abuf, asize);
2985 3011 }
2986 3012 }
2987 3013
2988 3014 zio_pop_transforms(zio); /* note: may set zio->io_error */
2989 3015
2990 3016 vdev_stat_update(zio, psize);
2991 3017
2992 3018 if (zio->io_error) {
2993 3019 /*
2994 3020 * If this I/O is attached to a particular vdev,
2995 3021 * generate an error message describing the I/O failure
2996 3022 * at the block level. We ignore these errors if the
2997 3023 * device is currently unavailable.
2998 3024 */
2999 3025 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3000 3026 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3001 3027
3002 3028 if ((zio->io_error == EIO || !(zio->io_flags &
3003 3029 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3004 3030 zio == lio) {
3005 3031 /*
3006 3032 * For logical I/O requests, tell the SPA to log the
3007 3033 * error and generate a logical data ereport.
3008 3034 */
3009 3035 spa_log_error(spa, zio);
3010 3036 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3011 3037 0, 0);
3012 3038 }
3013 3039 }
3014 3040
3015 3041 if (zio->io_error && zio == lio) {
3016 3042 /*
3017 3043 * Determine whether zio should be reexecuted. This will
3018 3044 * propagate all the way to the root via zio_notify_parent().
3019 3045 */
3020 3046 ASSERT(vd == NULL && bp != NULL);
3021 3047 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3022 3048
3023 3049 if (IO_IS_ALLOCATING(zio) &&
3024 3050 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3025 3051 if (zio->io_error != ENOSPC)
3026 3052 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3027 3053 else
3028 3054 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3029 3055 }
3030 3056
3031 3057 if ((zio->io_type == ZIO_TYPE_READ ||
3032 3058 zio->io_type == ZIO_TYPE_FREE) &&
3033 3059 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3034 3060 zio->io_error == ENXIO &&
3035 3061 spa_load_state(spa) == SPA_LOAD_NONE &&
3036 3062 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3037 3063 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3038 3064
3039 3065 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3040 3066 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3041 3067
3042 3068 /*
3043 3069 * Here is a possibly good place to attempt to do
3044 3070 * either combinatorial reconstruction or error correction
3045 3071 * based on checksums. It also might be a good place
3046 3072 * to send out preliminary ereports before we suspend
3047 3073 * processing.
3048 3074 */
3049 3075 }
3050 3076
3051 3077 /*
3052 3078 * If there were logical child errors, they apply to us now.
3053 3079 * We defer this until now to avoid conflating logical child
3054 3080 * errors with errors that happened to the zio itself when
3055 3081 * updating vdev stats and reporting FMA events above.
3056 3082 */
3057 3083 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3058 3084
3059 3085 if ((zio->io_error || zio->io_reexecute) &&
3060 3086 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3061 3087 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3062 3088 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3063 3089
3064 3090 zio_gang_tree_free(&zio->io_gang_tree);
3065 3091
3066 3092 /*
3067 3093 * Godfather I/Os should never suspend.
3068 3094 */
3069 3095 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3070 3096 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3071 3097 zio->io_reexecute = 0;
3072 3098
3073 3099 if (zio->io_reexecute) {
3074 3100 /*
3075 3101 * This is a logical I/O that wants to reexecute.
3076 3102 *
3077 3103 * Reexecute is top-down. When an i/o fails, if it's not
3078 3104 * the root, it simply notifies its parent and sticks around.
3079 3105 * The parent, seeing that it still has children in zio_done(),
3080 3106 * does the same. This percolates all the way up to the root.
3081 3107 * The root i/o will reexecute or suspend the entire tree.
3082 3108 *
3083 3109 * This approach ensures that zio_reexecute() honors
3084 3110 * all the original i/o dependency relationships, e.g.
3085 3111 * parents not executing until children are ready.
3086 3112 */
3087 3113 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3088 3114
3089 3115 zio->io_gang_leader = NULL;
3090 3116
3091 3117 mutex_enter(&zio->io_lock);
3092 3118 zio->io_state[ZIO_WAIT_DONE] = 1;
3093 3119 mutex_exit(&zio->io_lock);
3094 3120
3095 3121 /*
3096 3122 * "The Godfather" I/O monitors its children but is
3097 3123 * not a true parent to them. It will track them through
3098 3124 * the pipeline but severs its ties whenever they get into
3099 3125 * trouble (e.g. suspended). This allows "The Godfather"
3100 3126 * I/O to return status without blocking.
3101 3127 */
3102 3128 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3103 3129 zio_link_t *zl = zio->io_walk_link;
3104 3130 pio_next = zio_walk_parents(zio);
3105 3131
3106 3132 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3107 3133 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3108 3134 zio_remove_child(pio, zio, zl);
3109 3135 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3110 3136 }
3111 3137 }
3112 3138
3113 3139 if ((pio = zio_unique_parent(zio)) != NULL) {
3114 3140 /*
3115 3141 * We're not a root i/o, so there's nothing to do
3116 3142 * but notify our parent. Don't propagate errors
3117 3143 * upward since we haven't permanently failed yet.
3118 3144 */
3119 3145 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3120 3146 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3121 3147 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3122 3148 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3123 3149 /*
3124 3150 * We'd fail again if we reexecuted now, so suspend
3125 3151 * until conditions improve (e.g. device comes online).
3126 3152 */
3127 3153 zio_suspend(spa, zio);
3128 3154 } else {
3129 3155 /*
3130 3156 * Reexecution is potentially a huge amount of work.
3131 3157 * Hand it off to the otherwise-unused claim taskq.
3132 3158 */
3133 3159 ASSERT(zio->io_tqent.tqent_next == NULL);
3134 3160 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3135 3161 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3136 3162 0, &zio->io_tqent);
3137 3163 }
3138 3164 return (ZIO_PIPELINE_STOP);
3139 3165 }
3140 3166
3141 3167 ASSERT(zio->io_child_count == 0);
3142 3168 ASSERT(zio->io_reexecute == 0);
3143 3169 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3144 3170
3145 3171 /*
3146 3172 * Report any checksum errors, since the I/O is complete.
3147 3173 */
3148 3174 while (zio->io_cksum_report != NULL) {
3149 3175 zio_cksum_report_t *zcr = zio->io_cksum_report;
3150 3176 zio->io_cksum_report = zcr->zcr_next;
3151 3177 zcr->zcr_next = NULL;
3152 3178 zcr->zcr_finish(zcr, NULL);
3153 3179 zfs_ereport_free_checksum(zcr);
3154 3180 }
3155 3181
3156 3182 /*
3157 3183 * It is the responsibility of the done callback to ensure that this
3158 3184 * particular zio is no longer discoverable for adoption, and as
3159 3185 * such, cannot acquire any new parents.
3160 3186 */
3161 3187 if (zio->io_done)
3162 3188 zio->io_done(zio);
3163 3189
3164 3190 mutex_enter(&zio->io_lock);
3165 3191 zio->io_state[ZIO_WAIT_DONE] = 1;
3166 3192 mutex_exit(&zio->io_lock);
3167 3193
3168 3194 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3169 3195 zio_link_t *zl = zio->io_walk_link;
3170 3196 pio_next = zio_walk_parents(zio);
3171 3197 zio_remove_child(pio, zio, zl);
3172 3198 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3173 3199 }
3174 3200
3175 3201 if (zio->io_waiter != NULL) {
3176 3202 mutex_enter(&zio->io_lock);
3177 3203 zio->io_executor = NULL;
3178 3204 cv_broadcast(&zio->io_cv);
3179 3205 mutex_exit(&zio->io_lock);
3180 3206 } else {
3181 3207 zio_destroy(zio);
3182 3208 }
3183 3209
3184 3210 return (ZIO_PIPELINE_STOP);
3185 3211 }
3186 3212
3187 3213 /*
3188 3214 * ==========================================================================
3189 3215 * I/O pipeline definition
3190 3216 * ==========================================================================
3191 3217 */
3192 3218 static zio_pipe_stage_t *zio_pipeline[] = {
3193 3219 NULL,
3194 3220 zio_read_bp_init,
3195 3221 zio_free_bp_init,
3196 3222 zio_issue_async,
3197 3223 zio_write_bp_init,
3198 3224 zio_checksum_generate,
3199 3225 zio_nop_write,
3200 3226 zio_ddt_read_start,
3201 3227 zio_ddt_read_done,
3202 3228 zio_ddt_write,
3203 3229 zio_ddt_free,
3204 3230 zio_gang_assemble,
3205 3231 zio_gang_issue,
3206 3232 zio_dva_allocate,
3207 3233 zio_dva_free,
3208 3234 zio_dva_claim,
3209 3235 zio_ready,
3210 3236 zio_vdev_io_start,
3211 3237 zio_vdev_io_done,
3212 3238 zio_vdev_io_assess,
3213 3239 zio_checksum_verify,
3214 3240 zio_done
3215 3241 };
3216 3242
3217 3243 /* dnp is the dnode for zb1->zb_object */
3218 3244 boolean_t
3219 3245 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1,
3220 3246 const zbookmark_phys_t *zb2)
3221 3247 {
3222 3248 uint64_t zb1nextL0, zb2thisobj;
3223 3249
3224 3250 ASSERT(zb1->zb_objset == zb2->zb_objset);
3225 3251 ASSERT(zb2->zb_level == 0);
3226 3252
3227 3253 /* The objset_phys_t isn't before anything. */
3228 3254 if (dnp == NULL)
3229 3255 return (B_FALSE);
3230 3256
3231 3257 zb1nextL0 = (zb1->zb_blkid + 1) <<
3232 3258 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3233 3259
3234 3260 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3235 3261 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3236 3262
3237 3263 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3238 3264 uint64_t nextobj = zb1nextL0 *
3239 3265 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3240 3266 return (nextobj <= zb2thisobj);
3241 3267 }
3242 3268
3243 3269 if (zb1->zb_object < zb2thisobj)
3244 3270 return (B_TRUE);
3245 3271 if (zb1->zb_object > zb2thisobj)
3246 3272 return (B_FALSE);
3247 3273 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3248 3274 return (B_FALSE);
3249 3275 return (zb1nextL0 <= zb2->zb_blkid);
3250 3276 }
↓ open down ↓ |
1367 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX