1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2013 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 */
26
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
29 #include <sys/spa.h>
30 #include <sys/txg.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio_impl.h>
34 #include <sys/zio_compress.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/arc.h>
38 #include <sys/ddt.h>
39
40 /*
41 * ==========================================================================
42 * I/O priority table
43 * ==========================================================================
44 */
45 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
46 0, /* ZIO_PRIORITY_NOW */
47 0, /* ZIO_PRIORITY_SYNC_READ */
48 0, /* ZIO_PRIORITY_SYNC_WRITE */
49 0, /* ZIO_PRIORITY_LOG_WRITE */
50 1, /* ZIO_PRIORITY_CACHE_FILL */
51 1, /* ZIO_PRIORITY_AGG */
52 4, /* ZIO_PRIORITY_FREE */
53 4, /* ZIO_PRIORITY_ASYNC_WRITE */
54 6, /* ZIO_PRIORITY_ASYNC_READ */
55 10, /* ZIO_PRIORITY_RESILVER */
56 20, /* ZIO_PRIORITY_SCRUB */
57 2, /* ZIO_PRIORITY_DDT_PREFETCH */
58 };
59
60 /*
61 * ==========================================================================
62 * I/O type descriptions
63 * ==========================================================================
64 */
65 char *zio_type_name[ZIO_TYPES] = {
66 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
67 "zio_ioctl"
68 };
69
70 /*
71 * ==========================================================================
72 * I/O kmem caches
73 * ==========================================================================
74 */
75 kmem_cache_t *zio_cache;
76 kmem_cache_t *zio_link_cache;
77 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
78 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
79
80 #ifdef _KERNEL
81 extern vmem_t *zio_alloc_arena;
82 #endif
83 extern int zfs_mg_alloc_failures;
84
85 /*
86 * The following actions directly effect the spa's sync-to-convergence logic.
87 * The values below define the sync pass when we start performing the action.
88 * Care should be taken when changing these values as they directly impact
89 * spa_sync() performance. Tuning these values may introduce subtle performance
90 * pathologies and should only be done in the context of performance analysis.
91 * These tunables will eventually be removed and replaced with #defines once
92 * enough analysis has been done to determine optimal values.
93 *
94 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
95 * regular blocks are not deferred.
96 */
97 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
98 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
99 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
100
101 /*
102 * An allocating zio is one that either currently has the DVA allocate
103 * stage set or will have it later in its lifetime.
104 */
105 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
106
107 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
108
109 #ifdef ZFS_DEBUG
110 int zio_buf_debug_limit = 16384;
111 #else
112 int zio_buf_debug_limit = 0;
113 #endif
114
115 void
116 zio_init(void)
117 {
118 size_t c;
119 vmem_t *data_alloc_arena = NULL;
120
121 #ifdef _KERNEL
122 data_alloc_arena = zio_alloc_arena;
123 #endif
124 zio_cache = kmem_cache_create("zio_cache",
125 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
126 zio_link_cache = kmem_cache_create("zio_link_cache",
127 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
128
129 /*
130 * For small buffers, we want a cache for each multiple of
131 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
132 * for each quarter-power of 2. For large buffers, we want
133 * a cache for each multiple of PAGESIZE.
134 */
135 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
136 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
137 size_t p2 = size;
138 size_t align = 0;
139 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
140
141 while (p2 & (p2 - 1))
142 p2 &= p2 - 1;
143
144 #ifndef _KERNEL
145 /*
146 * If we are using watchpoints, put each buffer on its own page,
147 * to eliminate the performance overhead of trapping to the
148 * kernel when modifying a non-watched buffer that shares the
149 * page with a watched buffer.
150 */
151 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
152 continue;
153 #endif
154 if (size <= 4 * SPA_MINBLOCKSIZE) {
155 align = SPA_MINBLOCKSIZE;
156 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
157 align = PAGESIZE;
158 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
159 align = p2 >> 2;
160 }
161
162 if (align != 0) {
163 char name[36];
164 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
165 zio_buf_cache[c] = kmem_cache_create(name, size,
166 align, NULL, NULL, NULL, NULL, NULL, cflags);
167
168 /*
169 * Since zio_data bufs do not appear in crash dumps, we
170 * pass KMC_NOTOUCH so that no allocator metadata is
171 * stored with the buffers.
172 */
173 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
174 zio_data_buf_cache[c] = kmem_cache_create(name, size,
175 align, NULL, NULL, NULL, NULL, data_alloc_arena,
176 cflags | KMC_NOTOUCH);
177 }
178 }
179
180 while (--c != 0) {
181 ASSERT(zio_buf_cache[c] != NULL);
182 if (zio_buf_cache[c - 1] == NULL)
183 zio_buf_cache[c - 1] = zio_buf_cache[c];
184
185 ASSERT(zio_data_buf_cache[c] != NULL);
186 if (zio_data_buf_cache[c - 1] == NULL)
187 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
188 }
189
190 /*
191 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
192 * to fail 3 times per txg or 8 failures, whichever is greater.
193 */
194 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
195
196 zio_inject_init();
197 }
198
199 void
200 zio_fini(void)
201 {
202 size_t c;
203 kmem_cache_t *last_cache = NULL;
204 kmem_cache_t *last_data_cache = NULL;
205
206 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
207 if (zio_buf_cache[c] != last_cache) {
208 last_cache = zio_buf_cache[c];
209 kmem_cache_destroy(zio_buf_cache[c]);
210 }
211 zio_buf_cache[c] = NULL;
212
213 if (zio_data_buf_cache[c] != last_data_cache) {
214 last_data_cache = zio_data_buf_cache[c];
215 kmem_cache_destroy(zio_data_buf_cache[c]);
216 }
217 zio_data_buf_cache[c] = NULL;
218 }
219
220 kmem_cache_destroy(zio_link_cache);
221 kmem_cache_destroy(zio_cache);
222
223 zio_inject_fini();
224 }
225
226 /*
227 * ==========================================================================
228 * Allocate and free I/O buffers
229 * ==========================================================================
230 */
231
232 /*
233 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
234 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
235 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
236 * excess / transient data in-core during a crashdump.
237 */
238 void *
239 zio_buf_alloc(size_t size)
240 {
241 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
242
243 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
244
245 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
246 }
247
248 /*
249 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
250 * crashdump if the kernel panics. This exists so that we will limit the amount
251 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
252 * of kernel heap dumped to disk when the kernel panics)
253 */
254 void *
255 zio_data_buf_alloc(size_t size)
256 {
257 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
258
259 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
260
261 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
262 }
263
264 void
265 zio_buf_free(void *buf, size_t size)
266 {
267 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
268
269 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
270
271 kmem_cache_free(zio_buf_cache[c], buf);
272 }
273
274 void
275 zio_data_buf_free(void *buf, size_t size)
276 {
277 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
278
279 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
280
281 kmem_cache_free(zio_data_buf_cache[c], buf);
282 }
283
284 /*
285 * ==========================================================================
286 * Push and pop I/O transform buffers
287 * ==========================================================================
288 */
289 static void
290 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
291 zio_transform_func_t *transform)
292 {
293 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
294
295 zt->zt_orig_data = zio->io_data;
296 zt->zt_orig_size = zio->io_size;
297 zt->zt_bufsize = bufsize;
298 zt->zt_transform = transform;
299
300 zt->zt_next = zio->io_transform_stack;
301 zio->io_transform_stack = zt;
302
303 zio->io_data = data;
304 zio->io_size = size;
305 }
306
307 static void
308 zio_pop_transforms(zio_t *zio)
309 {
310 zio_transform_t *zt;
311
312 while ((zt = zio->io_transform_stack) != NULL) {
313 if (zt->zt_transform != NULL)
314 zt->zt_transform(zio,
315 zt->zt_orig_data, zt->zt_orig_size);
316
317 if (zt->zt_bufsize != 0)
318 zio_buf_free(zio->io_data, zt->zt_bufsize);
319
320 zio->io_data = zt->zt_orig_data;
321 zio->io_size = zt->zt_orig_size;
322 zio->io_transform_stack = zt->zt_next;
323
324 kmem_free(zt, sizeof (zio_transform_t));
325 }
326 }
327
328 /*
329 * ==========================================================================
330 * I/O transform callbacks for subblocks and decompression
331 * ==========================================================================
332 */
333 static void
334 zio_subblock(zio_t *zio, void *data, uint64_t size)
335 {
336 ASSERT(zio->io_size > size);
337
338 if (zio->io_type == ZIO_TYPE_READ)
339 bcopy(zio->io_data, data, size);
340 }
341
342 static void
343 zio_decompress(zio_t *zio, void *data, uint64_t size)
344 {
345 if (zio->io_error == 0 &&
346 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
347 zio->io_data, data, zio->io_size, size) != 0)
348 zio->io_error = SET_ERROR(EIO);
349 }
350
351 /*
352 * ==========================================================================
353 * I/O parent/child relationships and pipeline interlocks
354 * ==========================================================================
355 */
356 /*
357 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
358 * continue calling these functions until they return NULL.
359 * Otherwise, the next caller will pick up the list walk in
360 * some indeterminate state. (Otherwise every caller would
361 * have to pass in a cookie to keep the state represented by
362 * io_walk_link, which gets annoying.)
363 */
364 zio_t *
365 zio_walk_parents(zio_t *cio)
366 {
367 zio_link_t *zl = cio->io_walk_link;
368 list_t *pl = &cio->io_parent_list;
369
370 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
371 cio->io_walk_link = zl;
372
373 if (zl == NULL)
374 return (NULL);
375
376 ASSERT(zl->zl_child == cio);
377 return (zl->zl_parent);
378 }
379
380 zio_t *
381 zio_walk_children(zio_t *pio)
382 {
383 zio_link_t *zl = pio->io_walk_link;
384 list_t *cl = &pio->io_child_list;
385
386 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
387 pio->io_walk_link = zl;
388
389 if (zl == NULL)
390 return (NULL);
391
392 ASSERT(zl->zl_parent == pio);
393 return (zl->zl_child);
394 }
395
396 zio_t *
397 zio_unique_parent(zio_t *cio)
398 {
399 zio_t *pio = zio_walk_parents(cio);
400
401 VERIFY(zio_walk_parents(cio) == NULL);
402 return (pio);
403 }
404
405 void
406 zio_add_child(zio_t *pio, zio_t *cio)
407 {
408 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
409
410 /*
411 * Logical I/Os can have logical, gang, or vdev children.
412 * Gang I/Os can have gang or vdev children.
413 * Vdev I/Os can only have vdev children.
414 * The following ASSERT captures all of these constraints.
415 */
416 ASSERT(cio->io_child_type <= pio->io_child_type);
417
418 zl->zl_parent = pio;
419 zl->zl_child = cio;
420
421 mutex_enter(&cio->io_lock);
422 mutex_enter(&pio->io_lock);
423
424 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
425
426 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
427 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
428
429 list_insert_head(&pio->io_child_list, zl);
430 list_insert_head(&cio->io_parent_list, zl);
431
432 pio->io_child_count++;
433 cio->io_parent_count++;
434
435 mutex_exit(&pio->io_lock);
436 mutex_exit(&cio->io_lock);
437 }
438
439 static void
440 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
441 {
442 ASSERT(zl->zl_parent == pio);
443 ASSERT(zl->zl_child == cio);
444
445 mutex_enter(&cio->io_lock);
446 mutex_enter(&pio->io_lock);
447
448 list_remove(&pio->io_child_list, zl);
449 list_remove(&cio->io_parent_list, zl);
450
451 pio->io_child_count--;
452 cio->io_parent_count--;
453
454 mutex_exit(&pio->io_lock);
455 mutex_exit(&cio->io_lock);
456
457 kmem_cache_free(zio_link_cache, zl);
458 }
459
460 static boolean_t
461 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
462 {
463 uint64_t *countp = &zio->io_children[child][wait];
464 boolean_t waiting = B_FALSE;
465
466 mutex_enter(&zio->io_lock);
467 ASSERT(zio->io_stall == NULL);
468 if (*countp != 0) {
469 zio->io_stage >>= 1;
470 zio->io_stall = countp;
471 waiting = B_TRUE;
472 }
473 mutex_exit(&zio->io_lock);
474
475 return (waiting);
476 }
477
478 static void
479 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
480 {
481 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
482 int *errorp = &pio->io_child_error[zio->io_child_type];
483
484 mutex_enter(&pio->io_lock);
485 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
486 *errorp = zio_worst_error(*errorp, zio->io_error);
487 pio->io_reexecute |= zio->io_reexecute;
488 ASSERT3U(*countp, >, 0);
489 if (--*countp == 0 && pio->io_stall == countp) {
490 pio->io_stall = NULL;
491 mutex_exit(&pio->io_lock);
492 zio_execute(pio);
493 } else {
494 mutex_exit(&pio->io_lock);
495 }
496 }
497
498 static void
499 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
500 {
501 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
502 zio->io_error = zio->io_child_error[c];
503 }
504
505 /*
506 * ==========================================================================
507 * Create the various types of I/O (read, write, free, etc)
508 * ==========================================================================
509 */
510 static zio_t *
511 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
512 void *data, uint64_t size, zio_done_func_t *done, void *private,
513 zio_type_t type, int priority, enum zio_flag flags,
514 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
515 enum zio_stage stage, enum zio_stage pipeline)
516 {
517 zio_t *zio;
518
519 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
520 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
521 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
522
523 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
524 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
525 ASSERT(vd || stage == ZIO_STAGE_OPEN);
526
527 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
528 bzero(zio, sizeof (zio_t));
529
530 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
531 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
532
533 list_create(&zio->io_parent_list, sizeof (zio_link_t),
534 offsetof(zio_link_t, zl_parent_node));
535 list_create(&zio->io_child_list, sizeof (zio_link_t),
536 offsetof(zio_link_t, zl_child_node));
537
538 if (vd != NULL)
539 zio->io_child_type = ZIO_CHILD_VDEV;
540 else if (flags & ZIO_FLAG_GANG_CHILD)
541 zio->io_child_type = ZIO_CHILD_GANG;
542 else if (flags & ZIO_FLAG_DDT_CHILD)
543 zio->io_child_type = ZIO_CHILD_DDT;
544 else
545 zio->io_child_type = ZIO_CHILD_LOGICAL;
546
547 if (bp != NULL) {
548 zio->io_bp = (blkptr_t *)bp;
549 zio->io_bp_copy = *bp;
550 zio->io_bp_orig = *bp;
551 if (type != ZIO_TYPE_WRITE ||
552 zio->io_child_type == ZIO_CHILD_DDT)
553 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
554 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
555 zio->io_logical = zio;
556 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
557 pipeline |= ZIO_GANG_STAGES;
558 }
559
560 zio->io_spa = spa;
561 zio->io_txg = txg;
562 zio->io_done = done;
563 zio->io_private = private;
564 zio->io_type = type;
565 zio->io_priority = priority;
566 zio->io_vd = vd;
567 zio->io_offset = offset;
568 zio->io_orig_data = zio->io_data = data;
569 zio->io_orig_size = zio->io_size = size;
570 zio->io_orig_flags = zio->io_flags = flags;
571 zio->io_orig_stage = zio->io_stage = stage;
572 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
573
574 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
575 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
576
577 if (zb != NULL)
578 zio->io_bookmark = *zb;
579
580 if (pio != NULL) {
581 if (zio->io_logical == NULL)
582 zio->io_logical = pio->io_logical;
583 if (zio->io_child_type == ZIO_CHILD_GANG)
584 zio->io_gang_leader = pio->io_gang_leader;
585 zio_add_child(pio, zio);
586 }
587
588 return (zio);
589 }
590
591 static void
592 zio_destroy(zio_t *zio)
593 {
594 list_destroy(&zio->io_parent_list);
595 list_destroy(&zio->io_child_list);
596 mutex_destroy(&zio->io_lock);
597 cv_destroy(&zio->io_cv);
598 kmem_cache_free(zio_cache, zio);
599 }
600
601 zio_t *
602 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
603 void *private, enum zio_flag flags)
604 {
605 zio_t *zio;
606
607 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
608 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
609 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
610
611 return (zio);
612 }
613
614 zio_t *
615 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
616 {
617 return (zio_null(NULL, spa, NULL, done, private, flags));
618 }
619
620 zio_t *
621 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
622 void *data, uint64_t size, zio_done_func_t *done, void *private,
623 int priority, enum zio_flag flags, const zbookmark_t *zb)
624 {
625 zio_t *zio;
626
627 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
628 data, size, done, private,
629 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
630 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
631 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
632
633 return (zio);
634 }
635
636 zio_t *
637 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
638 void *data, uint64_t size, const zio_prop_t *zp,
639 zio_done_func_t *ready, zio_done_func_t *done, void *private,
640 int priority, enum zio_flag flags, const zbookmark_t *zb)
641 {
642 zio_t *zio;
643
644 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
645 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
646 zp->zp_compress >= ZIO_COMPRESS_OFF &&
647 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
648 DMU_OT_IS_VALID(zp->zp_type) &&
649 zp->zp_level < 32 &&
650 zp->zp_copies > 0 &&
651 zp->zp_copies <= spa_max_replication(spa));
652
653 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
654 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
655 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
656 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
657
658 zio->io_ready = ready;
659 zio->io_prop = *zp;
660
661 return (zio);
662 }
663
664 zio_t *
665 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
666 uint64_t size, zio_done_func_t *done, void *private, int priority,
667 enum zio_flag flags, zbookmark_t *zb)
668 {
669 zio_t *zio;
670
671 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
672 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
673 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
674
675 return (zio);
676 }
677
678 void
679 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
680 {
681 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
682 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
683 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
684 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
685
686 /*
687 * We must reset the io_prop to match the values that existed
688 * when the bp was first written by dmu_sync() keeping in mind
689 * that nopwrite and dedup are mutually exclusive.
690 */
691 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
692 zio->io_prop.zp_nopwrite = nopwrite;
693 zio->io_prop.zp_copies = copies;
694 zio->io_bp_override = bp;
695 }
696
697 void
698 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
699 {
700 metaslab_check_free(spa, bp);
701
702 /*
703 * Frees that are for the currently-syncing txg, are not going to be
704 * deferred, and which will not need to do a read (i.e. not GANG or
705 * DEDUP), can be processed immediately. Otherwise, put them on the
706 * in-memory list for later processing.
707 */
708 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
709 txg != spa->spa_syncing_txg ||
710 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
711 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
712 } else {
713 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
714 }
715 }
716
717 zio_t *
718 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
719 enum zio_flag flags)
720 {
721 zio_t *zio;
722 enum zio_stage stage = ZIO_FREE_PIPELINE;
723
724 dprintf_bp(bp, "freeing in txg %llu, pass %u",
725 (longlong_t)txg, spa->spa_sync_pass);
726
727 ASSERT(!BP_IS_HOLE(bp));
728 ASSERT(spa_syncing_txg(spa) == txg);
729 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
730
731 metaslab_check_free(spa, bp);
732 arc_freed(spa, bp);
733
734 /*
735 * GANG and DEDUP blocks can induce a read (for the gang block header,
736 * or the DDT), so issue them asynchronously so that this thread is
737 * not tied up.
738 */
739 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
740 stage |= ZIO_STAGE_ISSUE_ASYNC;
741
742 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
743 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
744 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
745
746
747 return (zio);
748 }
749
750 zio_t *
751 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
752 zio_done_func_t *done, void *private, enum zio_flag flags)
753 {
754 zio_t *zio;
755
756 /*
757 * A claim is an allocation of a specific block. Claims are needed
758 * to support immediate writes in the intent log. The issue is that
759 * immediate writes contain committed data, but in a txg that was
760 * *not* committed. Upon opening the pool after an unclean shutdown,
761 * the intent log claims all blocks that contain immediate write data
762 * so that the SPA knows they're in use.
763 *
764 * All claims *must* be resolved in the first txg -- before the SPA
765 * starts allocating blocks -- so that nothing is allocated twice.
766 * If txg == 0 we just verify that the block is claimable.
767 */
768 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
769 ASSERT(txg == spa_first_txg(spa) || txg == 0);
770 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
771
772 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
773 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
774 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
775
776 return (zio);
777 }
778
779 zio_t *
780 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
781 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
782 {
783 zio_t *zio;
784 int c;
785
786 if (vd->vdev_children == 0) {
787 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
788 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
789 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
790
791 zio->io_cmd = cmd;
792 } else {
793 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
794
795 for (c = 0; c < vd->vdev_children; c++)
796 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
797 done, private, priority, flags));
798 }
799
800 return (zio);
801 }
802
803 zio_t *
804 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
805 void *data, int checksum, zio_done_func_t *done, void *private,
806 int priority, enum zio_flag flags, boolean_t labels)
807 {
808 zio_t *zio;
809
810 ASSERT(vd->vdev_children == 0);
811 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
812 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
813 ASSERT3U(offset + size, <=, vd->vdev_psize);
814
815 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
816 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
817 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
818
819 zio->io_prop.zp_checksum = checksum;
820
821 return (zio);
822 }
823
824 zio_t *
825 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
826 void *data, int checksum, zio_done_func_t *done, void *private,
827 int priority, enum zio_flag flags, boolean_t labels)
828 {
829 zio_t *zio;
830
831 ASSERT(vd->vdev_children == 0);
832 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
833 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
834 ASSERT3U(offset + size, <=, vd->vdev_psize);
835
836 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
837 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
838 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
839
840 zio->io_prop.zp_checksum = checksum;
841
842 if (zio_checksum_table[checksum].ci_eck) {
843 /*
844 * zec checksums are necessarily destructive -- they modify
845 * the end of the write buffer to hold the verifier/checksum.
846 * Therefore, we must make a local copy in case the data is
847 * being written to multiple places in parallel.
848 */
849 void *wbuf = zio_buf_alloc(size);
850 bcopy(data, wbuf, size);
851 zio_push_transform(zio, wbuf, size, size, NULL);
852 }
853
854 return (zio);
855 }
856
857 /*
858 * Create a child I/O to do some work for us.
859 */
860 zio_t *
861 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
862 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
863 zio_done_func_t *done, void *private)
864 {
865 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
866 zio_t *zio;
867
868 ASSERT(vd->vdev_parent ==
869 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
870
871 if (type == ZIO_TYPE_READ && bp != NULL) {
872 /*
873 * If we have the bp, then the child should perform the
874 * checksum and the parent need not. This pushes error
875 * detection as close to the leaves as possible and
876 * eliminates redundant checksums in the interior nodes.
877 */
878 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
879 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
880 }
881
882 if (vd->vdev_children == 0)
883 offset += VDEV_LABEL_START_SIZE;
884
885 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
886
887 /*
888 * If we've decided to do a repair, the write is not speculative --
889 * even if the original read was.
890 */
891 if (flags & ZIO_FLAG_IO_REPAIR)
892 flags &= ~ZIO_FLAG_SPECULATIVE;
893
894 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
895 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
896 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
897
898 return (zio);
899 }
900
901 zio_t *
902 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
903 int type, int priority, enum zio_flag flags,
904 zio_done_func_t *done, void *private)
905 {
906 zio_t *zio;
907
908 ASSERT(vd->vdev_ops->vdev_op_leaf);
909
910 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
911 data, size, done, private, type, priority,
912 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
913 vd, offset, NULL,
914 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
915
916 return (zio);
917 }
918
919 void
920 zio_flush(zio_t *zio, vdev_t *vd)
921 {
922 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
923 NULL, NULL, ZIO_PRIORITY_NOW,
924 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
925 }
926
927 void
928 zio_shrink(zio_t *zio, uint64_t size)
929 {
930 ASSERT(zio->io_executor == NULL);
931 ASSERT(zio->io_orig_size == zio->io_size);
932 ASSERT(size <= zio->io_size);
933
934 /*
935 * We don't shrink for raidz because of problems with the
936 * reconstruction when reading back less than the block size.
937 * Note, BP_IS_RAIDZ() assumes no compression.
938 */
939 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
940 if (!BP_IS_RAIDZ(zio->io_bp))
941 zio->io_orig_size = zio->io_size = size;
942 }
943
944 /*
945 * ==========================================================================
946 * Prepare to read and write logical blocks
947 * ==========================================================================
948 */
949
950 static int
951 zio_read_bp_init(zio_t *zio)
952 {
953 blkptr_t *bp = zio->io_bp;
954
955 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
956 zio->io_child_type == ZIO_CHILD_LOGICAL &&
957 !(zio->io_flags & ZIO_FLAG_RAW)) {
958 uint64_t psize = BP_GET_PSIZE(bp);
959 void *cbuf = zio_buf_alloc(psize);
960
961 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
962 }
963
964 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
965 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
966
967 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
968 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
969
970 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
971 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
972
973 return (ZIO_PIPELINE_CONTINUE);
974 }
975
976 static int
977 zio_write_bp_init(zio_t *zio)
978 {
979 spa_t *spa = zio->io_spa;
980 zio_prop_t *zp = &zio->io_prop;
981 enum zio_compress compress = zp->zp_compress;
982 blkptr_t *bp = zio->io_bp;
983 uint64_t lsize = zio->io_size;
984 uint64_t psize = lsize;
985 int pass = 1;
986
987 /*
988 * If our children haven't all reached the ready stage,
989 * wait for them and then repeat this pipeline stage.
990 */
991 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
992 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
993 return (ZIO_PIPELINE_STOP);
994
995 if (!IO_IS_ALLOCATING(zio))
996 return (ZIO_PIPELINE_CONTINUE);
997
998 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
999
1000 if (zio->io_bp_override) {
1001 ASSERT(bp->blk_birth != zio->io_txg);
1002 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1003
1004 *bp = *zio->io_bp_override;
1005 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1006
1007 /*
1008 * If we've been overridden and nopwrite is set then
1009 * set the flag accordingly to indicate that a nopwrite
1010 * has already occurred.
1011 */
1012 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1013 ASSERT(!zp->zp_dedup);
1014 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1015 return (ZIO_PIPELINE_CONTINUE);
1016 }
1017
1018 ASSERT(!zp->zp_nopwrite);
1019
1020 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1021 return (ZIO_PIPELINE_CONTINUE);
1022
1023 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1024 zp->zp_dedup_verify);
1025
1026 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1027 BP_SET_DEDUP(bp, 1);
1028 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1029 return (ZIO_PIPELINE_CONTINUE);
1030 }
1031 zio->io_bp_override = NULL;
1032 BP_ZERO(bp);
1033 }
1034
1035 if (bp->blk_birth == zio->io_txg) {
1036 /*
1037 * We're rewriting an existing block, which means we're
1038 * working on behalf of spa_sync(). For spa_sync() to
1039 * converge, it must eventually be the case that we don't
1040 * have to allocate new blocks. But compression changes
1041 * the blocksize, which forces a reallocate, and makes
1042 * convergence take longer. Therefore, after the first
1043 * few passes, stop compressing to ensure convergence.
1044 */
1045 pass = spa_sync_pass(spa);
1046
1047 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1048 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1049 ASSERT(!BP_GET_DEDUP(bp));
1050
1051 if (pass >= zfs_sync_pass_dont_compress)
1052 compress = ZIO_COMPRESS_OFF;
1053
1054 /* Make sure someone doesn't change their mind on overwrites */
1055 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1056 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1057 }
1058
1059 if (compress != ZIO_COMPRESS_OFF) {
1060 void *cbuf = zio_buf_alloc(lsize);
1061 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1062 if (psize == 0 || psize == lsize) {
1063 compress = ZIO_COMPRESS_OFF;
1064 zio_buf_free(cbuf, lsize);
1065 } else {
1066 ASSERT(psize < lsize);
1067 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1068 }
1069 }
1070
1071 /*
1072 * The final pass of spa_sync() must be all rewrites, but the first
1073 * few passes offer a trade-off: allocating blocks defers convergence,
1074 * but newly allocated blocks are sequential, so they can be written
1075 * to disk faster. Therefore, we allow the first few passes of
1076 * spa_sync() to allocate new blocks, but force rewrites after that.
1077 * There should only be a handful of blocks after pass 1 in any case.
1078 */
1079 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1080 pass >= zfs_sync_pass_rewrite) {
1081 ASSERT(psize != 0);
1082 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1083 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1084 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1085 } else {
1086 BP_ZERO(bp);
1087 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1088 }
1089
1090 if (psize == 0) {
1091 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1092 } else {
1093 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1094 BP_SET_LSIZE(bp, lsize);
1095 BP_SET_PSIZE(bp, psize);
1096 BP_SET_COMPRESS(bp, compress);
1097 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1098 BP_SET_TYPE(bp, zp->zp_type);
1099 BP_SET_LEVEL(bp, zp->zp_level);
1100 BP_SET_DEDUP(bp, zp->zp_dedup);
1101 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1102 if (zp->zp_dedup) {
1103 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1104 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1105 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1106 }
1107 if (zp->zp_nopwrite) {
1108 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1109 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1110 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1111 }
1112 }
1113
1114 return (ZIO_PIPELINE_CONTINUE);
1115 }
1116
1117 static int
1118 zio_free_bp_init(zio_t *zio)
1119 {
1120 blkptr_t *bp = zio->io_bp;
1121
1122 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1123 if (BP_GET_DEDUP(bp))
1124 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1125 }
1126
1127 return (ZIO_PIPELINE_CONTINUE);
1128 }
1129
1130 /*
1131 * ==========================================================================
1132 * Execute the I/O pipeline
1133 * ==========================================================================
1134 */
1135
1136 static void
1137 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1138 {
1139 spa_t *spa = zio->io_spa;
1140 zio_type_t t = zio->io_type;
1141 int flags = (cutinline ? TQ_FRONT : 0);
1142
1143 /*
1144 * If we're a config writer or a probe, the normal issue and
1145 * interrupt threads may all be blocked waiting for the config lock.
1146 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1147 */
1148 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1149 t = ZIO_TYPE_NULL;
1150
1151 /*
1152 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1153 */
1154 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1155 t = ZIO_TYPE_NULL;
1156
1157 /*
1158 * If this is a high priority I/O, then use the high priority taskq if
1159 * available.
1160 */
1161 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1162 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1163 q++;
1164
1165 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1166
1167 /*
1168 * NB: We are assuming that the zio can only be dispatched
1169 * to a single taskq at a time. It would be a grievous error
1170 * to dispatch the zio to another taskq at the same time.
1171 */
1172 ASSERT(zio->io_tqent.tqent_next == NULL);
1173 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1174 flags, &zio->io_tqent);
1175 }
1176
1177 static boolean_t
1178 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1179 {
1180 kthread_t *executor = zio->io_executor;
1181 spa_t *spa = zio->io_spa;
1182
1183 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1184 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1185 uint_t i;
1186 for (i = 0; i < tqs->stqs_count; i++) {
1187 if (taskq_member(tqs->stqs_taskq[i], executor))
1188 return (B_TRUE);
1189 }
1190 }
1191
1192 return (B_FALSE);
1193 }
1194
1195 static int
1196 zio_issue_async(zio_t *zio)
1197 {
1198 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1199
1200 return (ZIO_PIPELINE_STOP);
1201 }
1202
1203 void
1204 zio_interrupt(zio_t *zio)
1205 {
1206 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1207 }
1208
1209 /*
1210 * Execute the I/O pipeline until one of the following occurs:
1211 *
1212 * (1) the I/O completes
1213 * (2) the pipeline stalls waiting for dependent child I/Os
1214 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1215 * (4) the I/O is delegated by vdev-level caching or aggregation
1216 * (5) the I/O is deferred due to vdev-level queueing
1217 * (6) the I/O is handed off to another thread.
1218 *
1219 * In all cases, the pipeline stops whenever there's no CPU work; it never
1220 * burns a thread in cv_wait().
1221 *
1222 * There's no locking on io_stage because there's no legitimate way
1223 * for multiple threads to be attempting to process the same I/O.
1224 */
1225 static zio_pipe_stage_t *zio_pipeline[];
1226
1227 void
1228 zio_execute(zio_t *zio)
1229 {
1230 zio->io_executor = curthread;
1231
1232 while (zio->io_stage < ZIO_STAGE_DONE) {
1233 enum zio_stage pipeline = zio->io_pipeline;
1234 enum zio_stage stage = zio->io_stage;
1235 int rv;
1236
1237 ASSERT(!MUTEX_HELD(&zio->io_lock));
1238 ASSERT(ISP2(stage));
1239 ASSERT(zio->io_stall == NULL);
1240
1241 do {
1242 stage <<= 1;
1243 } while ((stage & pipeline) == 0);
1244
1245 ASSERT(stage <= ZIO_STAGE_DONE);
1246
1247 /*
1248 * If we are in interrupt context and this pipeline stage
1249 * will grab a config lock that is held across I/O,
1250 * or may wait for an I/O that needs an interrupt thread
1251 * to complete, issue async to avoid deadlock.
1252 *
1253 * For VDEV_IO_START, we cut in line so that the io will
1254 * be sent to disk promptly.
1255 */
1256 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1257 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1258 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1259 zio_requeue_io_start_cut_in_line : B_FALSE;
1260 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1261 return;
1262 }
1263
1264 zio->io_stage = stage;
1265 rv = zio_pipeline[highbit(stage) - 1](zio);
1266
1267 if (rv == ZIO_PIPELINE_STOP)
1268 return;
1269
1270 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1271 }
1272 }
1273
1274 /*
1275 * ==========================================================================
1276 * Initiate I/O, either sync or async
1277 * ==========================================================================
1278 */
1279 int
1280 zio_wait(zio_t *zio)
1281 {
1282 int error;
1283
1284 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1285 ASSERT(zio->io_executor == NULL);
1286
1287 zio->io_waiter = curthread;
1288
1289 zio_execute(zio);
1290
1291 mutex_enter(&zio->io_lock);
1292 while (zio->io_executor != NULL)
1293 cv_wait(&zio->io_cv, &zio->io_lock);
1294 mutex_exit(&zio->io_lock);
1295
1296 error = zio->io_error;
1297 zio_destroy(zio);
1298
1299 return (error);
1300 }
1301
1302 void
1303 zio_nowait(zio_t *zio)
1304 {
1305 ASSERT(zio->io_executor == NULL);
1306
1307 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1308 zio_unique_parent(zio) == NULL) {
1309 /*
1310 * This is a logical async I/O with no parent to wait for it.
1311 * We add it to the spa_async_root_zio "Godfather" I/O which
1312 * will ensure they complete prior to unloading the pool.
1313 */
1314 spa_t *spa = zio->io_spa;
1315
1316 zio_add_child(spa->spa_async_zio_root, zio);
1317 }
1318
1319 zio_execute(zio);
1320 }
1321
1322 /*
1323 * ==========================================================================
1324 * Reexecute or suspend/resume failed I/O
1325 * ==========================================================================
1326 */
1327
1328 static void
1329 zio_reexecute(zio_t *pio)
1330 {
1331 zio_t *cio, *cio_next;
1332
1333 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1334 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1335 ASSERT(pio->io_gang_leader == NULL);
1336 ASSERT(pio->io_gang_tree == NULL);
1337
1338 pio->io_flags = pio->io_orig_flags;
1339 pio->io_stage = pio->io_orig_stage;
1340 pio->io_pipeline = pio->io_orig_pipeline;
1341 pio->io_reexecute = 0;
1342 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1343 pio->io_error = 0;
1344 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1345 pio->io_state[w] = 0;
1346 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1347 pio->io_child_error[c] = 0;
1348
1349 if (IO_IS_ALLOCATING(pio))
1350 BP_ZERO(pio->io_bp);
1351
1352 /*
1353 * As we reexecute pio's children, new children could be created.
1354 * New children go to the head of pio's io_child_list, however,
1355 * so we will (correctly) not reexecute them. The key is that
1356 * the remainder of pio's io_child_list, from 'cio_next' onward,
1357 * cannot be affected by any side effects of reexecuting 'cio'.
1358 */
1359 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1360 cio_next = zio_walk_children(pio);
1361 mutex_enter(&pio->io_lock);
1362 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1363 pio->io_children[cio->io_child_type][w]++;
1364 mutex_exit(&pio->io_lock);
1365 zio_reexecute(cio);
1366 }
1367
1368 /*
1369 * Now that all children have been reexecuted, execute the parent.
1370 * We don't reexecute "The Godfather" I/O here as it's the
1371 * responsibility of the caller to wait on him.
1372 */
1373 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1374 zio_execute(pio);
1375 }
1376
1377 void
1378 zio_suspend(spa_t *spa, zio_t *zio)
1379 {
1380 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1381 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1382 "failure and the failure mode property for this pool "
1383 "is set to panic.", spa_name(spa));
1384
1385 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1386
1387 mutex_enter(&spa->spa_suspend_lock);
1388
1389 if (spa->spa_suspend_zio_root == NULL)
1390 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1391 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1392 ZIO_FLAG_GODFATHER);
1393
1394 spa->spa_suspended = B_TRUE;
1395
1396 if (zio != NULL) {
1397 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1398 ASSERT(zio != spa->spa_suspend_zio_root);
1399 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1400 ASSERT(zio_unique_parent(zio) == NULL);
1401 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1402 zio_add_child(spa->spa_suspend_zio_root, zio);
1403 }
1404
1405 mutex_exit(&spa->spa_suspend_lock);
1406 }
1407
1408 int
1409 zio_resume(spa_t *spa)
1410 {
1411 zio_t *pio;
1412
1413 /*
1414 * Reexecute all previously suspended i/o.
1415 */
1416 mutex_enter(&spa->spa_suspend_lock);
1417 spa->spa_suspended = B_FALSE;
1418 cv_broadcast(&spa->spa_suspend_cv);
1419 pio = spa->spa_suspend_zio_root;
1420 spa->spa_suspend_zio_root = NULL;
1421 mutex_exit(&spa->spa_suspend_lock);
1422
1423 if (pio == NULL)
1424 return (0);
1425
1426 zio_reexecute(pio);
1427 return (zio_wait(pio));
1428 }
1429
1430 void
1431 zio_resume_wait(spa_t *spa)
1432 {
1433 mutex_enter(&spa->spa_suspend_lock);
1434 while (spa_suspended(spa))
1435 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1436 mutex_exit(&spa->spa_suspend_lock);
1437 }
1438
1439 /*
1440 * ==========================================================================
1441 * Gang blocks.
1442 *
1443 * A gang block is a collection of small blocks that looks to the DMU
1444 * like one large block. When zio_dva_allocate() cannot find a block
1445 * of the requested size, due to either severe fragmentation or the pool
1446 * being nearly full, it calls zio_write_gang_block() to construct the
1447 * block from smaller fragments.
1448 *
1449 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1450 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1451 * an indirect block: it's an array of block pointers. It consumes
1452 * only one sector and hence is allocatable regardless of fragmentation.
1453 * The gang header's bps point to its gang members, which hold the data.
1454 *
1455 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1456 * as the verifier to ensure uniqueness of the SHA256 checksum.
1457 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1458 * not the gang header. This ensures that data block signatures (needed for
1459 * deduplication) are independent of how the block is physically stored.
1460 *
1461 * Gang blocks can be nested: a gang member may itself be a gang block.
1462 * Thus every gang block is a tree in which root and all interior nodes are
1463 * gang headers, and the leaves are normal blocks that contain user data.
1464 * The root of the gang tree is called the gang leader.
1465 *
1466 * To perform any operation (read, rewrite, free, claim) on a gang block,
1467 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1468 * in the io_gang_tree field of the original logical i/o by recursively
1469 * reading the gang leader and all gang headers below it. This yields
1470 * an in-core tree containing the contents of every gang header and the
1471 * bps for every constituent of the gang block.
1472 *
1473 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1474 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1475 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1476 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1477 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1478 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1479 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1480 * of the gang header plus zio_checksum_compute() of the data to update the
1481 * gang header's blk_cksum as described above.
1482 *
1483 * The two-phase assemble/issue model solves the problem of partial failure --
1484 * what if you'd freed part of a gang block but then couldn't read the
1485 * gang header for another part? Assembling the entire gang tree first
1486 * ensures that all the necessary gang header I/O has succeeded before
1487 * starting the actual work of free, claim, or write. Once the gang tree
1488 * is assembled, free and claim are in-memory operations that cannot fail.
1489 *
1490 * In the event that a gang write fails, zio_dva_unallocate() walks the
1491 * gang tree to immediately free (i.e. insert back into the space map)
1492 * everything we've allocated. This ensures that we don't get ENOSPC
1493 * errors during repeated suspend/resume cycles due to a flaky device.
1494 *
1495 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1496 * the gang tree, we won't modify the block, so we can safely defer the free
1497 * (knowing that the block is still intact). If we *can* assemble the gang
1498 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1499 * each constituent bp and we can allocate a new block on the next sync pass.
1500 *
1501 * In all cases, the gang tree allows complete recovery from partial failure.
1502 * ==========================================================================
1503 */
1504
1505 static zio_t *
1506 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1507 {
1508 if (gn != NULL)
1509 return (pio);
1510
1511 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1512 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1513 &pio->io_bookmark));
1514 }
1515
1516 zio_t *
1517 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1518 {
1519 zio_t *zio;
1520
1521 if (gn != NULL) {
1522 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1523 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1524 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1525 /*
1526 * As we rewrite each gang header, the pipeline will compute
1527 * a new gang block header checksum for it; but no one will
1528 * compute a new data checksum, so we do that here. The one
1529 * exception is the gang leader: the pipeline already computed
1530 * its data checksum because that stage precedes gang assembly.
1531 * (Presently, nothing actually uses interior data checksums;
1532 * this is just good hygiene.)
1533 */
1534 if (gn != pio->io_gang_leader->io_gang_tree) {
1535 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1536 data, BP_GET_PSIZE(bp));
1537 }
1538 /*
1539 * If we are here to damage data for testing purposes,
1540 * leave the GBH alone so that we can detect the damage.
1541 */
1542 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1543 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1544 } else {
1545 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1546 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1547 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1548 }
1549
1550 return (zio);
1551 }
1552
1553 /* ARGSUSED */
1554 zio_t *
1555 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1556 {
1557 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1558 ZIO_GANG_CHILD_FLAGS(pio)));
1559 }
1560
1561 /* ARGSUSED */
1562 zio_t *
1563 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1564 {
1565 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1566 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1567 }
1568
1569 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1570 NULL,
1571 zio_read_gang,
1572 zio_rewrite_gang,
1573 zio_free_gang,
1574 zio_claim_gang,
1575 NULL
1576 };
1577
1578 static void zio_gang_tree_assemble_done(zio_t *zio);
1579
1580 static zio_gang_node_t *
1581 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1582 {
1583 zio_gang_node_t *gn;
1584
1585 ASSERT(*gnpp == NULL);
1586
1587 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1588 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1589 *gnpp = gn;
1590
1591 return (gn);
1592 }
1593
1594 static void
1595 zio_gang_node_free(zio_gang_node_t **gnpp)
1596 {
1597 zio_gang_node_t *gn = *gnpp;
1598
1599 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1600 ASSERT(gn->gn_child[g] == NULL);
1601
1602 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1603 kmem_free(gn, sizeof (*gn));
1604 *gnpp = NULL;
1605 }
1606
1607 static void
1608 zio_gang_tree_free(zio_gang_node_t **gnpp)
1609 {
1610 zio_gang_node_t *gn = *gnpp;
1611
1612 if (gn == NULL)
1613 return;
1614
1615 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1616 zio_gang_tree_free(&gn->gn_child[g]);
1617
1618 zio_gang_node_free(gnpp);
1619 }
1620
1621 static void
1622 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1623 {
1624 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1625
1626 ASSERT(gio->io_gang_leader == gio);
1627 ASSERT(BP_IS_GANG(bp));
1628
1629 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1630 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1631 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1632 }
1633
1634 static void
1635 zio_gang_tree_assemble_done(zio_t *zio)
1636 {
1637 zio_t *gio = zio->io_gang_leader;
1638 zio_gang_node_t *gn = zio->io_private;
1639 blkptr_t *bp = zio->io_bp;
1640
1641 ASSERT(gio == zio_unique_parent(zio));
1642 ASSERT(zio->io_child_count == 0);
1643
1644 if (zio->io_error)
1645 return;
1646
1647 if (BP_SHOULD_BYTESWAP(bp))
1648 byteswap_uint64_array(zio->io_data, zio->io_size);
1649
1650 ASSERT(zio->io_data == gn->gn_gbh);
1651 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1652 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1653
1654 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1655 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1656 if (!BP_IS_GANG(gbp))
1657 continue;
1658 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1659 }
1660 }
1661
1662 static void
1663 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1664 {
1665 zio_t *gio = pio->io_gang_leader;
1666 zio_t *zio;
1667
1668 ASSERT(BP_IS_GANG(bp) == !!gn);
1669 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1670 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1671
1672 /*
1673 * If you're a gang header, your data is in gn->gn_gbh.
1674 * If you're a gang member, your data is in 'data' and gn == NULL.
1675 */
1676 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1677
1678 if (gn != NULL) {
1679 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1680
1681 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1682 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1683 if (BP_IS_HOLE(gbp))
1684 continue;
1685 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1686 data = (char *)data + BP_GET_PSIZE(gbp);
1687 }
1688 }
1689
1690 if (gn == gio->io_gang_tree)
1691 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1692
1693 if (zio != pio)
1694 zio_nowait(zio);
1695 }
1696
1697 static int
1698 zio_gang_assemble(zio_t *zio)
1699 {
1700 blkptr_t *bp = zio->io_bp;
1701
1702 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1703 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1704
1705 zio->io_gang_leader = zio;
1706
1707 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1708
1709 return (ZIO_PIPELINE_CONTINUE);
1710 }
1711
1712 static int
1713 zio_gang_issue(zio_t *zio)
1714 {
1715 blkptr_t *bp = zio->io_bp;
1716
1717 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1718 return (ZIO_PIPELINE_STOP);
1719
1720 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1721 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1722
1723 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1724 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1725 else
1726 zio_gang_tree_free(&zio->io_gang_tree);
1727
1728 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1729
1730 return (ZIO_PIPELINE_CONTINUE);
1731 }
1732
1733 static void
1734 zio_write_gang_member_ready(zio_t *zio)
1735 {
1736 zio_t *pio = zio_unique_parent(zio);
1737 zio_t *gio = zio->io_gang_leader;
1738 dva_t *cdva = zio->io_bp->blk_dva;
1739 dva_t *pdva = pio->io_bp->blk_dva;
1740 uint64_t asize;
1741
1742 if (BP_IS_HOLE(zio->io_bp))
1743 return;
1744
1745 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1746
1747 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1748 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1749 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1750 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1751 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1752
1753 mutex_enter(&pio->io_lock);
1754 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1755 ASSERT(DVA_GET_GANG(&pdva[d]));
1756 asize = DVA_GET_ASIZE(&pdva[d]);
1757 asize += DVA_GET_ASIZE(&cdva[d]);
1758 DVA_SET_ASIZE(&pdva[d], asize);
1759 }
1760 mutex_exit(&pio->io_lock);
1761 }
1762
1763 static int
1764 zio_write_gang_block(zio_t *pio)
1765 {
1766 spa_t *spa = pio->io_spa;
1767 blkptr_t *bp = pio->io_bp;
1768 zio_t *gio = pio->io_gang_leader;
1769 zio_t *zio;
1770 zio_gang_node_t *gn, **gnpp;
1771 zio_gbh_phys_t *gbh;
1772 uint64_t txg = pio->io_txg;
1773 uint64_t resid = pio->io_size;
1774 uint64_t lsize;
1775 int copies = gio->io_prop.zp_copies;
1776 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1777 zio_prop_t zp;
1778 int error;
1779
1780 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1781 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1782 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1783 if (error) {
1784 pio->io_error = error;
1785 return (ZIO_PIPELINE_CONTINUE);
1786 }
1787
1788 if (pio == gio) {
1789 gnpp = &gio->io_gang_tree;
1790 } else {
1791 gnpp = pio->io_private;
1792 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1793 }
1794
1795 gn = zio_gang_node_alloc(gnpp);
1796 gbh = gn->gn_gbh;
1797 bzero(gbh, SPA_GANGBLOCKSIZE);
1798
1799 /*
1800 * Create the gang header.
1801 */
1802 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1803 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1804
1805 /*
1806 * Create and nowait the gang children.
1807 */
1808 for (int g = 0; resid != 0; resid -= lsize, g++) {
1809 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1810 SPA_MINBLOCKSIZE);
1811 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1812
1813 zp.zp_checksum = gio->io_prop.zp_checksum;
1814 zp.zp_compress = ZIO_COMPRESS_OFF;
1815 zp.zp_type = DMU_OT_NONE;
1816 zp.zp_level = 0;
1817 zp.zp_copies = gio->io_prop.zp_copies;
1818 zp.zp_dedup = B_FALSE;
1819 zp.zp_dedup_verify = B_FALSE;
1820 zp.zp_nopwrite = B_FALSE;
1821
1822 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1823 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1824 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1825 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1826 &pio->io_bookmark));
1827 }
1828
1829 /*
1830 * Set pio's pipeline to just wait for zio to finish.
1831 */
1832 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1833
1834 zio_nowait(zio);
1835
1836 return (ZIO_PIPELINE_CONTINUE);
1837 }
1838
1839 /*
1840 * The zio_nop_write stage in the pipeline determines if allocating
1841 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1842 * such as SHA256, we can compare the checksums of the new data and the old
1843 * to determine if allocating a new block is required. The nopwrite
1844 * feature can handle writes in either syncing or open context (i.e. zil
1845 * writes) and as a result is mutually exclusive with dedup.
1846 */
1847 static int
1848 zio_nop_write(zio_t *zio)
1849 {
1850 blkptr_t *bp = zio->io_bp;
1851 blkptr_t *bp_orig = &zio->io_bp_orig;
1852 zio_prop_t *zp = &zio->io_prop;
1853
1854 ASSERT(BP_GET_LEVEL(bp) == 0);
1855 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1856 ASSERT(zp->zp_nopwrite);
1857 ASSERT(!zp->zp_dedup);
1858 ASSERT(zio->io_bp_override == NULL);
1859 ASSERT(IO_IS_ALLOCATING(zio));
1860
1861 /*
1862 * Check to see if the original bp and the new bp have matching
1863 * characteristics (i.e. same checksum, compression algorithms, etc).
1864 * If they don't then just continue with the pipeline which will
1865 * allocate a new bp.
1866 */
1867 if (BP_IS_HOLE(bp_orig) ||
1868 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1869 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1870 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1871 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1872 zp->zp_copies != BP_GET_NDVAS(bp_orig))
1873 return (ZIO_PIPELINE_CONTINUE);
1874
1875 /*
1876 * If the checksums match then reset the pipeline so that we
1877 * avoid allocating a new bp and issuing any I/O.
1878 */
1879 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1880 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1881 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1882 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1883 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1884 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1885 sizeof (uint64_t)) == 0);
1886
1887 *bp = *bp_orig;
1888 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1889 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1890 }
1891
1892 return (ZIO_PIPELINE_CONTINUE);
1893 }
1894
1895 /*
1896 * ==========================================================================
1897 * Dedup
1898 * ==========================================================================
1899 */
1900 static void
1901 zio_ddt_child_read_done(zio_t *zio)
1902 {
1903 blkptr_t *bp = zio->io_bp;
1904 ddt_entry_t *dde = zio->io_private;
1905 ddt_phys_t *ddp;
1906 zio_t *pio = zio_unique_parent(zio);
1907
1908 mutex_enter(&pio->io_lock);
1909 ddp = ddt_phys_select(dde, bp);
1910 if (zio->io_error == 0)
1911 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1912 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1913 dde->dde_repair_data = zio->io_data;
1914 else
1915 zio_buf_free(zio->io_data, zio->io_size);
1916 mutex_exit(&pio->io_lock);
1917 }
1918
1919 static int
1920 zio_ddt_read_start(zio_t *zio)
1921 {
1922 blkptr_t *bp = zio->io_bp;
1923
1924 ASSERT(BP_GET_DEDUP(bp));
1925 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1926 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1927
1928 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1929 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1930 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1931 ddt_phys_t *ddp = dde->dde_phys;
1932 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1933 blkptr_t blk;
1934
1935 ASSERT(zio->io_vsd == NULL);
1936 zio->io_vsd = dde;
1937
1938 if (ddp_self == NULL)
1939 return (ZIO_PIPELINE_CONTINUE);
1940
1941 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1942 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1943 continue;
1944 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1945 &blk);
1946 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1947 zio_buf_alloc(zio->io_size), zio->io_size,
1948 zio_ddt_child_read_done, dde, zio->io_priority,
1949 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1950 &zio->io_bookmark));
1951 }
1952 return (ZIO_PIPELINE_CONTINUE);
1953 }
1954
1955 zio_nowait(zio_read(zio, zio->io_spa, bp,
1956 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1957 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1958
1959 return (ZIO_PIPELINE_CONTINUE);
1960 }
1961
1962 static int
1963 zio_ddt_read_done(zio_t *zio)
1964 {
1965 blkptr_t *bp = zio->io_bp;
1966
1967 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1968 return (ZIO_PIPELINE_STOP);
1969
1970 ASSERT(BP_GET_DEDUP(bp));
1971 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1972 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1973
1974 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1975 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1976 ddt_entry_t *dde = zio->io_vsd;
1977 if (ddt == NULL) {
1978 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1979 return (ZIO_PIPELINE_CONTINUE);
1980 }
1981 if (dde == NULL) {
1982 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1983 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1984 return (ZIO_PIPELINE_STOP);
1985 }
1986 if (dde->dde_repair_data != NULL) {
1987 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1988 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1989 }
1990 ddt_repair_done(ddt, dde);
1991 zio->io_vsd = NULL;
1992 }
1993
1994 ASSERT(zio->io_vsd == NULL);
1995
1996 return (ZIO_PIPELINE_CONTINUE);
1997 }
1998
1999 static boolean_t
2000 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2001 {
2002 spa_t *spa = zio->io_spa;
2003
2004 /*
2005 * Note: we compare the original data, not the transformed data,
2006 * because when zio->io_bp is an override bp, we will not have
2007 * pushed the I/O transforms. That's an important optimization
2008 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2009 */
2010 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2011 zio_t *lio = dde->dde_lead_zio[p];
2012
2013 if (lio != NULL) {
2014 return (lio->io_orig_size != zio->io_orig_size ||
2015 bcmp(zio->io_orig_data, lio->io_orig_data,
2016 zio->io_orig_size) != 0);
2017 }
2018 }
2019
2020 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2021 ddt_phys_t *ddp = &dde->dde_phys[p];
2022
2023 if (ddp->ddp_phys_birth != 0) {
2024 arc_buf_t *abuf = NULL;
2025 uint32_t aflags = ARC_WAIT;
2026 blkptr_t blk = *zio->io_bp;
2027 int error;
2028
2029 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2030
2031 ddt_exit(ddt);
2032
2033 error = arc_read(NULL, spa, &blk,
2034 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2035 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2036 &aflags, &zio->io_bookmark);
2037
2038 if (error == 0) {
2039 if (arc_buf_size(abuf) != zio->io_orig_size ||
2040 bcmp(abuf->b_data, zio->io_orig_data,
2041 zio->io_orig_size) != 0)
2042 error = SET_ERROR(EEXIST);
2043 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2044 }
2045
2046 ddt_enter(ddt);
2047 return (error != 0);
2048 }
2049 }
2050
2051 return (B_FALSE);
2052 }
2053
2054 static void
2055 zio_ddt_child_write_ready(zio_t *zio)
2056 {
2057 int p = zio->io_prop.zp_copies;
2058 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2059 ddt_entry_t *dde = zio->io_private;
2060 ddt_phys_t *ddp = &dde->dde_phys[p];
2061 zio_t *pio;
2062
2063 if (zio->io_error)
2064 return;
2065
2066 ddt_enter(ddt);
2067
2068 ASSERT(dde->dde_lead_zio[p] == zio);
2069
2070 ddt_phys_fill(ddp, zio->io_bp);
2071
2072 while ((pio = zio_walk_parents(zio)) != NULL)
2073 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2074
2075 ddt_exit(ddt);
2076 }
2077
2078 static void
2079 zio_ddt_child_write_done(zio_t *zio)
2080 {
2081 int p = zio->io_prop.zp_copies;
2082 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2083 ddt_entry_t *dde = zio->io_private;
2084 ddt_phys_t *ddp = &dde->dde_phys[p];
2085
2086 ddt_enter(ddt);
2087
2088 ASSERT(ddp->ddp_refcnt == 0);
2089 ASSERT(dde->dde_lead_zio[p] == zio);
2090 dde->dde_lead_zio[p] = NULL;
2091
2092 if (zio->io_error == 0) {
2093 while (zio_walk_parents(zio) != NULL)
2094 ddt_phys_addref(ddp);
2095 } else {
2096 ddt_phys_clear(ddp);
2097 }
2098
2099 ddt_exit(ddt);
2100 }
2101
2102 static void
2103 zio_ddt_ditto_write_done(zio_t *zio)
2104 {
2105 int p = DDT_PHYS_DITTO;
2106 zio_prop_t *zp = &zio->io_prop;
2107 blkptr_t *bp = zio->io_bp;
2108 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2109 ddt_entry_t *dde = zio->io_private;
2110 ddt_phys_t *ddp = &dde->dde_phys[p];
2111 ddt_key_t *ddk = &dde->dde_key;
2112
2113 ddt_enter(ddt);
2114
2115 ASSERT(ddp->ddp_refcnt == 0);
2116 ASSERT(dde->dde_lead_zio[p] == zio);
2117 dde->dde_lead_zio[p] = NULL;
2118
2119 if (zio->io_error == 0) {
2120 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2121 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2122 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2123 if (ddp->ddp_phys_birth != 0)
2124 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2125 ddt_phys_fill(ddp, bp);
2126 }
2127
2128 ddt_exit(ddt);
2129 }
2130
2131 static int
2132 zio_ddt_write(zio_t *zio)
2133 {
2134 spa_t *spa = zio->io_spa;
2135 blkptr_t *bp = zio->io_bp;
2136 uint64_t txg = zio->io_txg;
2137 zio_prop_t *zp = &zio->io_prop;
2138 int p = zp->zp_copies;
2139 int ditto_copies;
2140 zio_t *cio = NULL;
2141 zio_t *dio = NULL;
2142 ddt_t *ddt = ddt_select(spa, bp);
2143 ddt_entry_t *dde;
2144 ddt_phys_t *ddp;
2145
2146 ASSERT(BP_GET_DEDUP(bp));
2147 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2148 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2149
2150 ddt_enter(ddt);
2151 dde = ddt_lookup(ddt, bp, B_TRUE);
2152 ddp = &dde->dde_phys[p];
2153
2154 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2155 /*
2156 * If we're using a weak checksum, upgrade to a strong checksum
2157 * and try again. If we're already using a strong checksum,
2158 * we can't resolve it, so just convert to an ordinary write.
2159 * (And automatically e-mail a paper to Nature?)
2160 */
2161 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2162 zp->zp_checksum = spa_dedup_checksum(spa);
2163 zio_pop_transforms(zio);
2164 zio->io_stage = ZIO_STAGE_OPEN;
2165 BP_ZERO(bp);
2166 } else {
2167 zp->zp_dedup = B_FALSE;
2168 }
2169 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2170 ddt_exit(ddt);
2171 return (ZIO_PIPELINE_CONTINUE);
2172 }
2173
2174 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2175 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2176
2177 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2178 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2179 zio_prop_t czp = *zp;
2180
2181 czp.zp_copies = ditto_copies;
2182
2183 /*
2184 * If we arrived here with an override bp, we won't have run
2185 * the transform stack, so we won't have the data we need to
2186 * generate a child i/o. So, toss the override bp and restart.
2187 * This is safe, because using the override bp is just an
2188 * optimization; and it's rare, so the cost doesn't matter.
2189 */
2190 if (zio->io_bp_override) {
2191 zio_pop_transforms(zio);
2192 zio->io_stage = ZIO_STAGE_OPEN;
2193 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2194 zio->io_bp_override = NULL;
2195 BP_ZERO(bp);
2196 ddt_exit(ddt);
2197 return (ZIO_PIPELINE_CONTINUE);
2198 }
2199
2200 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2201 zio->io_orig_size, &czp, NULL,
2202 zio_ddt_ditto_write_done, dde, zio->io_priority,
2203 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2204
2205 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2206 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2207 }
2208
2209 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2210 if (ddp->ddp_phys_birth != 0)
2211 ddt_bp_fill(ddp, bp, txg);
2212 if (dde->dde_lead_zio[p] != NULL)
2213 zio_add_child(zio, dde->dde_lead_zio[p]);
2214 else
2215 ddt_phys_addref(ddp);
2216 } else if (zio->io_bp_override) {
2217 ASSERT(bp->blk_birth == txg);
2218 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2219 ddt_phys_fill(ddp, bp);
2220 ddt_phys_addref(ddp);
2221 } else {
2222 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2223 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2224 zio_ddt_child_write_done, dde, zio->io_priority,
2225 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2226
2227 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2228 dde->dde_lead_zio[p] = cio;
2229 }
2230
2231 ddt_exit(ddt);
2232
2233 if (cio)
2234 zio_nowait(cio);
2235 if (dio)
2236 zio_nowait(dio);
2237
2238 return (ZIO_PIPELINE_CONTINUE);
2239 }
2240
2241 ddt_entry_t *freedde; /* for debugging */
2242
2243 static int
2244 zio_ddt_free(zio_t *zio)
2245 {
2246 spa_t *spa = zio->io_spa;
2247 blkptr_t *bp = zio->io_bp;
2248 ddt_t *ddt = ddt_select(spa, bp);
2249 ddt_entry_t *dde;
2250 ddt_phys_t *ddp;
2251
2252 ASSERT(BP_GET_DEDUP(bp));
2253 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2254
2255 ddt_enter(ddt);
2256 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2257 ddp = ddt_phys_select(dde, bp);
2258 ddt_phys_decref(ddp);
2259 ddt_exit(ddt);
2260
2261 return (ZIO_PIPELINE_CONTINUE);
2262 }
2263
2264 /*
2265 * ==========================================================================
2266 * Allocate and free blocks
2267 * ==========================================================================
2268 */
2269 static int
2270 zio_dva_allocate(zio_t *zio)
2271 {
2272 spa_t *spa = zio->io_spa;
2273 metaslab_class_t *mc = spa_normal_class(spa);
2274 blkptr_t *bp = zio->io_bp;
2275 int error;
2276 int flags = 0;
2277
2278 if (zio->io_gang_leader == NULL) {
2279 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2280 zio->io_gang_leader = zio;
2281 }
2282
2283 ASSERT(BP_IS_HOLE(bp));
2284 ASSERT0(BP_GET_NDVAS(bp));
2285 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2286 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2287 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2288
2289 /*
2290 * The dump device does not support gang blocks so allocation on
2291 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2292 * the "fast" gang feature.
2293 */
2294 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2295 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2296 METASLAB_GANG_CHILD : 0;
2297 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2298 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2299
2300 if (error) {
2301 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2302 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2303 error);
2304 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2305 return (zio_write_gang_block(zio));
2306 zio->io_error = error;
2307 }
2308
2309 return (ZIO_PIPELINE_CONTINUE);
2310 }
2311
2312 static int
2313 zio_dva_free(zio_t *zio)
2314 {
2315 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2316
2317 return (ZIO_PIPELINE_CONTINUE);
2318 }
2319
2320 static int
2321 zio_dva_claim(zio_t *zio)
2322 {
2323 int error;
2324
2325 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2326 if (error)
2327 zio->io_error = error;
2328
2329 return (ZIO_PIPELINE_CONTINUE);
2330 }
2331
2332 /*
2333 * Undo an allocation. This is used by zio_done() when an I/O fails
2334 * and we want to give back the block we just allocated.
2335 * This handles both normal blocks and gang blocks.
2336 */
2337 static void
2338 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2339 {
2340 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2341 ASSERT(zio->io_bp_override == NULL);
2342
2343 if (!BP_IS_HOLE(bp))
2344 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2345
2346 if (gn != NULL) {
2347 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2348 zio_dva_unallocate(zio, gn->gn_child[g],
2349 &gn->gn_gbh->zg_blkptr[g]);
2350 }
2351 }
2352 }
2353
2354 /*
2355 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2356 */
2357 int
2358 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2359 uint64_t size, boolean_t use_slog)
2360 {
2361 int error = 1;
2362
2363 ASSERT(txg > spa_syncing_txg(spa));
2364
2365 /*
2366 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2367 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2368 * when allocating them.
2369 */
2370 if (use_slog) {
2371 error = metaslab_alloc(spa, spa_log_class(spa), size,
2372 new_bp, 1, txg, old_bp,
2373 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2374 }
2375
2376 if (error) {
2377 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2378 new_bp, 1, txg, old_bp,
2379 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2380 }
2381
2382 if (error == 0) {
2383 BP_SET_LSIZE(new_bp, size);
2384 BP_SET_PSIZE(new_bp, size);
2385 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2386 BP_SET_CHECKSUM(new_bp,
2387 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2388 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2389 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2390 BP_SET_LEVEL(new_bp, 0);
2391 BP_SET_DEDUP(new_bp, 0);
2392 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2393 }
2394
2395 return (error);
2396 }
2397
2398 /*
2399 * Free an intent log block.
2400 */
2401 void
2402 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2403 {
2404 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2405 ASSERT(!BP_IS_GANG(bp));
2406
2407 zio_free(spa, txg, bp);
2408 }
2409
2410 /*
2411 * ==========================================================================
2412 * Read and write to physical devices
2413 * ==========================================================================
2414 */
2415 static int
2416 zio_vdev_io_start(zio_t *zio)
2417 {
2418 vdev_t *vd = zio->io_vd;
2419 uint64_t align;
2420 spa_t *spa = zio->io_spa;
2421
2422 ASSERT(zio->io_error == 0);
2423 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2424
2425 if (vd == NULL) {
2426 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2427 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2428
2429 /*
2430 * The mirror_ops handle multiple DVAs in a single BP.
2431 */
2432 return (vdev_mirror_ops.vdev_op_io_start(zio));
2433 }
2434
2435 /*
2436 * We keep track of time-sensitive I/Os so that the scan thread
2437 * can quickly react to certain workloads. In particular, we care
2438 * about non-scrubbing, top-level reads and writes with the following
2439 * characteristics:
2440 * - synchronous writes of user data to non-slog devices
2441 * - any reads of user data
2442 * When these conditions are met, adjust the timestamp of spa_last_io
2443 * which allows the scan thread to adjust its workload accordingly.
2444 */
2445 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2446 vd == vd->vdev_top && !vd->vdev_islog &&
2447 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2448 zio->io_txg != spa_syncing_txg(spa)) {
2449 uint64_t old = spa->spa_last_io;
2450 uint64_t new = ddi_get_lbolt64();
2451 if (old != new)
2452 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2453 }
2454
2455 align = 1ULL << vd->vdev_top->vdev_ashift;
2456
2457 if (P2PHASE(zio->io_size, align) != 0) {
2458 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2459 char *abuf = zio_buf_alloc(asize);
2460 ASSERT(vd == vd->vdev_top);
2461 if (zio->io_type == ZIO_TYPE_WRITE) {
2462 bcopy(zio->io_data, abuf, zio->io_size);
2463 bzero(abuf + zio->io_size, asize - zio->io_size);
2464 }
2465 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2466 }
2467
2468 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2469 ASSERT(P2PHASE(zio->io_size, align) == 0);
2470 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2471
2472 /*
2473 * If this is a repair I/O, and there's no self-healing involved --
2474 * that is, we're just resilvering what we expect to resilver --
2475 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2476 * This prevents spurious resilvering with nested replication.
2477 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2478 * A is out of date, we'll read from C+D, then use the data to
2479 * resilver A+B -- but we don't actually want to resilver B, just A.
2480 * The top-level mirror has no way to know this, so instead we just
2481 * discard unnecessary repairs as we work our way down the vdev tree.
2482 * The same logic applies to any form of nested replication:
2483 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2484 */
2485 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2486 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2487 zio->io_txg != 0 && /* not a delegated i/o */
2488 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2489 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2490 zio_vdev_io_bypass(zio);
2491 return (ZIO_PIPELINE_CONTINUE);
2492 }
2493
2494 if (vd->vdev_ops->vdev_op_leaf &&
2495 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2496
2497 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2498 return (ZIO_PIPELINE_CONTINUE);
2499
2500 if ((zio = vdev_queue_io(zio)) == NULL)
2501 return (ZIO_PIPELINE_STOP);
2502
2503 if (!vdev_accessible(vd, zio)) {
2504 zio->io_error = SET_ERROR(ENXIO);
2505 zio_interrupt(zio);
2506 return (ZIO_PIPELINE_STOP);
2507 }
2508 }
2509
2510 return (vd->vdev_ops->vdev_op_io_start(zio));
2511 }
2512
2513 static int
2514 zio_vdev_io_done(zio_t *zio)
2515 {
2516 vdev_t *vd = zio->io_vd;
2517 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2518 boolean_t unexpected_error = B_FALSE;
2519
2520 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2521 return (ZIO_PIPELINE_STOP);
2522
2523 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2524
2525 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2526
2527 vdev_queue_io_done(zio);
2528
2529 if (zio->io_type == ZIO_TYPE_WRITE)
2530 vdev_cache_write(zio);
2531
2532 if (zio_injection_enabled && zio->io_error == 0)
2533 zio->io_error = zio_handle_device_injection(vd,
2534 zio, EIO);
2535
2536 if (zio_injection_enabled && zio->io_error == 0)
2537 zio->io_error = zio_handle_label_injection(zio, EIO);
2538
2539 if (zio->io_error) {
2540 if (!vdev_accessible(vd, zio)) {
2541 zio->io_error = SET_ERROR(ENXIO);
2542 } else {
2543 unexpected_error = B_TRUE;
2544 }
2545 }
2546 }
2547
2548 ops->vdev_op_io_done(zio);
2549
2550 if (unexpected_error)
2551 VERIFY(vdev_probe(vd, zio) == NULL);
2552
2553 return (ZIO_PIPELINE_CONTINUE);
2554 }
2555
2556 /*
2557 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2558 * disk, and use that to finish the checksum ereport later.
2559 */
2560 static void
2561 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2562 const void *good_buf)
2563 {
2564 /* no processing needed */
2565 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2566 }
2567
2568 /*ARGSUSED*/
2569 void
2570 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2571 {
2572 void *buf = zio_buf_alloc(zio->io_size);
2573
2574 bcopy(zio->io_data, buf, zio->io_size);
2575
2576 zcr->zcr_cbinfo = zio->io_size;
2577 zcr->zcr_cbdata = buf;
2578 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2579 zcr->zcr_free = zio_buf_free;
2580 }
2581
2582 static int
2583 zio_vdev_io_assess(zio_t *zio)
2584 {
2585 vdev_t *vd = zio->io_vd;
2586
2587 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2588 return (ZIO_PIPELINE_STOP);
2589
2590 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2591 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2592
2593 if (zio->io_vsd != NULL) {
2594 zio->io_vsd_ops->vsd_free(zio);
2595 zio->io_vsd = NULL;
2596 }
2597
2598 if (zio_injection_enabled && zio->io_error == 0)
2599 zio->io_error = zio_handle_fault_injection(zio, EIO);
2600
2601 /*
2602 * If the I/O failed, determine whether we should attempt to retry it.
2603 *
2604 * On retry, we cut in line in the issue queue, since we don't want
2605 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2606 */
2607 if (zio->io_error && vd == NULL &&
2608 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2609 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2610 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2611 zio->io_error = 0;
2612 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2613 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2614 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2615 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2616 zio_requeue_io_start_cut_in_line);
2617 return (ZIO_PIPELINE_STOP);
2618 }
2619
2620 /*
2621 * If we got an error on a leaf device, convert it to ENXIO
2622 * if the device is not accessible at all.
2623 */
2624 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2625 !vdev_accessible(vd, zio))
2626 zio->io_error = SET_ERROR(ENXIO);
2627
2628 /*
2629 * If we can't write to an interior vdev (mirror or RAID-Z),
2630 * set vdev_cant_write so that we stop trying to allocate from it.
2631 */
2632 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2633 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2634 vd->vdev_cant_write = B_TRUE;
2635 }
2636
2637 if (zio->io_error)
2638 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2639
2640 return (ZIO_PIPELINE_CONTINUE);
2641 }
2642
2643 void
2644 zio_vdev_io_reissue(zio_t *zio)
2645 {
2646 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2647 ASSERT(zio->io_error == 0);
2648
2649 zio->io_stage >>= 1;
2650 }
2651
2652 void
2653 zio_vdev_io_redone(zio_t *zio)
2654 {
2655 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2656
2657 zio->io_stage >>= 1;
2658 }
2659
2660 void
2661 zio_vdev_io_bypass(zio_t *zio)
2662 {
2663 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2664 ASSERT(zio->io_error == 0);
2665
2666 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2667 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2668 }
2669
2670 /*
2671 * ==========================================================================
2672 * Generate and verify checksums
2673 * ==========================================================================
2674 */
2675 static int
2676 zio_checksum_generate(zio_t *zio)
2677 {
2678 blkptr_t *bp = zio->io_bp;
2679 enum zio_checksum checksum;
2680
2681 if (bp == NULL) {
2682 /*
2683 * This is zio_write_phys().
2684 * We're either generating a label checksum, or none at all.
2685 */
2686 checksum = zio->io_prop.zp_checksum;
2687
2688 if (checksum == ZIO_CHECKSUM_OFF)
2689 return (ZIO_PIPELINE_CONTINUE);
2690
2691 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2692 } else {
2693 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2694 ASSERT(!IO_IS_ALLOCATING(zio));
2695 checksum = ZIO_CHECKSUM_GANG_HEADER;
2696 } else {
2697 checksum = BP_GET_CHECKSUM(bp);
2698 }
2699 }
2700
2701 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2702
2703 return (ZIO_PIPELINE_CONTINUE);
2704 }
2705
2706 static int
2707 zio_checksum_verify(zio_t *zio)
2708 {
2709 zio_bad_cksum_t info;
2710 blkptr_t *bp = zio->io_bp;
2711 int error;
2712
2713 ASSERT(zio->io_vd != NULL);
2714
2715 if (bp == NULL) {
2716 /*
2717 * This is zio_read_phys().
2718 * We're either verifying a label checksum, or nothing at all.
2719 */
2720 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2721 return (ZIO_PIPELINE_CONTINUE);
2722
2723 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2724 }
2725
2726 if ((error = zio_checksum_error(zio, &info)) != 0) {
2727 zio->io_error = error;
2728 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2729 zfs_ereport_start_checksum(zio->io_spa,
2730 zio->io_vd, zio, zio->io_offset,
2731 zio->io_size, NULL, &info);
2732 }
2733 }
2734
2735 return (ZIO_PIPELINE_CONTINUE);
2736 }
2737
2738 /*
2739 * Called by RAID-Z to ensure we don't compute the checksum twice.
2740 */
2741 void
2742 zio_checksum_verified(zio_t *zio)
2743 {
2744 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2745 }
2746
2747 /*
2748 * ==========================================================================
2749 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2750 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2751 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2752 * indicate errors that are specific to one I/O, and most likely permanent.
2753 * Any other error is presumed to be worse because we weren't expecting it.
2754 * ==========================================================================
2755 */
2756 int
2757 zio_worst_error(int e1, int e2)
2758 {
2759 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2760 int r1, r2;
2761
2762 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2763 if (e1 == zio_error_rank[r1])
2764 break;
2765
2766 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2767 if (e2 == zio_error_rank[r2])
2768 break;
2769
2770 return (r1 > r2 ? e1 : e2);
2771 }
2772
2773 /*
2774 * ==========================================================================
2775 * I/O completion
2776 * ==========================================================================
2777 */
2778 static int
2779 zio_ready(zio_t *zio)
2780 {
2781 blkptr_t *bp = zio->io_bp;
2782 zio_t *pio, *pio_next;
2783
2784 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2785 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2786 return (ZIO_PIPELINE_STOP);
2787
2788 if (zio->io_ready) {
2789 ASSERT(IO_IS_ALLOCATING(zio));
2790 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2791 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2792 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2793
2794 zio->io_ready(zio);
2795 }
2796
2797 if (bp != NULL && bp != &zio->io_bp_copy)
2798 zio->io_bp_copy = *bp;
2799
2800 if (zio->io_error)
2801 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2802
2803 mutex_enter(&zio->io_lock);
2804 zio->io_state[ZIO_WAIT_READY] = 1;
2805 pio = zio_walk_parents(zio);
2806 mutex_exit(&zio->io_lock);
2807
2808 /*
2809 * As we notify zio's parents, new parents could be added.
2810 * New parents go to the head of zio's io_parent_list, however,
2811 * so we will (correctly) not notify them. The remainder of zio's
2812 * io_parent_list, from 'pio_next' onward, cannot change because
2813 * all parents must wait for us to be done before they can be done.
2814 */
2815 for (; pio != NULL; pio = pio_next) {
2816 pio_next = zio_walk_parents(zio);
2817 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2818 }
2819
2820 if (zio->io_flags & ZIO_FLAG_NODATA) {
2821 if (BP_IS_GANG(bp)) {
2822 zio->io_flags &= ~ZIO_FLAG_NODATA;
2823 } else {
2824 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2825 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2826 }
2827 }
2828
2829 if (zio_injection_enabled &&
2830 zio->io_spa->spa_syncing_txg == zio->io_txg)
2831 zio_handle_ignored_writes(zio);
2832
2833 return (ZIO_PIPELINE_CONTINUE);
2834 }
2835
2836 static int
2837 zio_done(zio_t *zio)
2838 {
2839 spa_t *spa = zio->io_spa;
2840 zio_t *lio = zio->io_logical;
2841 blkptr_t *bp = zio->io_bp;
2842 vdev_t *vd = zio->io_vd;
2843 uint64_t psize = zio->io_size;
2844 zio_t *pio, *pio_next;
2845
2846 /*
2847 * If our children haven't all completed,
2848 * wait for them and then repeat this pipeline stage.
2849 */
2850 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2851 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2852 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2853 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2854 return (ZIO_PIPELINE_STOP);
2855
2856 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2857 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2858 ASSERT(zio->io_children[c][w] == 0);
2859
2860 if (bp != NULL) {
2861 ASSERT(bp->blk_pad[0] == 0);
2862 ASSERT(bp->blk_pad[1] == 0);
2863 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2864 (bp == zio_unique_parent(zio)->io_bp));
2865 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2866 zio->io_bp_override == NULL &&
2867 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2868 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2869 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2870 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2871 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2872 }
2873 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
2874 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
2875 }
2876
2877 /*
2878 * If there were child vdev/gang/ddt errors, they apply to us now.
2879 */
2880 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2881 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2882 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2883
2884 /*
2885 * If the I/O on the transformed data was successful, generate any
2886 * checksum reports now while we still have the transformed data.
2887 */
2888 if (zio->io_error == 0) {
2889 while (zio->io_cksum_report != NULL) {
2890 zio_cksum_report_t *zcr = zio->io_cksum_report;
2891 uint64_t align = zcr->zcr_align;
2892 uint64_t asize = P2ROUNDUP(psize, align);
2893 char *abuf = zio->io_data;
2894
2895 if (asize != psize) {
2896 abuf = zio_buf_alloc(asize);
2897 bcopy(zio->io_data, abuf, psize);
2898 bzero(abuf + psize, asize - psize);
2899 }
2900
2901 zio->io_cksum_report = zcr->zcr_next;
2902 zcr->zcr_next = NULL;
2903 zcr->zcr_finish(zcr, abuf);
2904 zfs_ereport_free_checksum(zcr);
2905
2906 if (asize != psize)
2907 zio_buf_free(abuf, asize);
2908 }
2909 }
2910
2911 zio_pop_transforms(zio); /* note: may set zio->io_error */
2912
2913 vdev_stat_update(zio, psize);
2914
2915 if (zio->io_error) {
2916 /*
2917 * If this I/O is attached to a particular vdev,
2918 * generate an error message describing the I/O failure
2919 * at the block level. We ignore these errors if the
2920 * device is currently unavailable.
2921 */
2922 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2923 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2924
2925 if ((zio->io_error == EIO || !(zio->io_flags &
2926 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2927 zio == lio) {
2928 /*
2929 * For logical I/O requests, tell the SPA to log the
2930 * error and generate a logical data ereport.
2931 */
2932 spa_log_error(spa, zio);
2933 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2934 0, 0);
2935 }
2936 }
2937
2938 if (zio->io_error && zio == lio) {
2939 /*
2940 * Determine whether zio should be reexecuted. This will
2941 * propagate all the way to the root via zio_notify_parent().
2942 */
2943 ASSERT(vd == NULL && bp != NULL);
2944 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2945
2946 if (IO_IS_ALLOCATING(zio) &&
2947 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2948 if (zio->io_error != ENOSPC)
2949 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2950 else
2951 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2952 }
2953
2954 if ((zio->io_type == ZIO_TYPE_READ ||
2955 zio->io_type == ZIO_TYPE_FREE) &&
2956 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2957 zio->io_error == ENXIO &&
2958 spa_load_state(spa) == SPA_LOAD_NONE &&
2959 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2960 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2961
2962 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2963 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2964
2965 /*
2966 * Here is a possibly good place to attempt to do
2967 * either combinatorial reconstruction or error correction
2968 * based on checksums. It also might be a good place
2969 * to send out preliminary ereports before we suspend
2970 * processing.
2971 */
2972 }
2973
2974 /*
2975 * If there were logical child errors, they apply to us now.
2976 * We defer this until now to avoid conflating logical child
2977 * errors with errors that happened to the zio itself when
2978 * updating vdev stats and reporting FMA events above.
2979 */
2980 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2981
2982 if ((zio->io_error || zio->io_reexecute) &&
2983 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2984 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
2985 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2986
2987 zio_gang_tree_free(&zio->io_gang_tree);
2988
2989 /*
2990 * Godfather I/Os should never suspend.
2991 */
2992 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2993 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2994 zio->io_reexecute = 0;
2995
2996 if (zio->io_reexecute) {
2997 /*
2998 * This is a logical I/O that wants to reexecute.
2999 *
3000 * Reexecute is top-down. When an i/o fails, if it's not
3001 * the root, it simply notifies its parent and sticks around.
3002 * The parent, seeing that it still has children in zio_done(),
3003 * does the same. This percolates all the way up to the root.
3004 * The root i/o will reexecute or suspend the entire tree.
3005 *
3006 * This approach ensures that zio_reexecute() honors
3007 * all the original i/o dependency relationships, e.g.
3008 * parents not executing until children are ready.
3009 */
3010 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3011
3012 zio->io_gang_leader = NULL;
3013
3014 mutex_enter(&zio->io_lock);
3015 zio->io_state[ZIO_WAIT_DONE] = 1;
3016 mutex_exit(&zio->io_lock);
3017
3018 /*
3019 * "The Godfather" I/O monitors its children but is
3020 * not a true parent to them. It will track them through
3021 * the pipeline but severs its ties whenever they get into
3022 * trouble (e.g. suspended). This allows "The Godfather"
3023 * I/O to return status without blocking.
3024 */
3025 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3026 zio_link_t *zl = zio->io_walk_link;
3027 pio_next = zio_walk_parents(zio);
3028
3029 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3030 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3031 zio_remove_child(pio, zio, zl);
3032 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3033 }
3034 }
3035
3036 if ((pio = zio_unique_parent(zio)) != NULL) {
3037 /*
3038 * We're not a root i/o, so there's nothing to do
3039 * but notify our parent. Don't propagate errors
3040 * upward since we haven't permanently failed yet.
3041 */
3042 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3043 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3044 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3045 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3046 /*
3047 * We'd fail again if we reexecuted now, so suspend
3048 * until conditions improve (e.g. device comes online).
3049 */
3050 zio_suspend(spa, zio);
3051 } else {
3052 /*
3053 * Reexecution is potentially a huge amount of work.
3054 * Hand it off to the otherwise-unused claim taskq.
3055 */
3056 ASSERT(zio->io_tqent.tqent_next == NULL);
3057 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3058 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3059 0, &zio->io_tqent);
3060 }
3061 return (ZIO_PIPELINE_STOP);
3062 }
3063
3064 ASSERT(zio->io_child_count == 0);
3065 ASSERT(zio->io_reexecute == 0);
3066 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3067
3068 /*
3069 * Report any checksum errors, since the I/O is complete.
3070 */
3071 while (zio->io_cksum_report != NULL) {
3072 zio_cksum_report_t *zcr = zio->io_cksum_report;
3073 zio->io_cksum_report = zcr->zcr_next;
3074 zcr->zcr_next = NULL;
3075 zcr->zcr_finish(zcr, NULL);
3076 zfs_ereport_free_checksum(zcr);
3077 }
3078
3079 /*
3080 * It is the responsibility of the done callback to ensure that this
3081 * particular zio is no longer discoverable for adoption, and as
3082 * such, cannot acquire any new parents.
3083 */
3084 if (zio->io_done)
3085 zio->io_done(zio);
3086
3087 mutex_enter(&zio->io_lock);
3088 zio->io_state[ZIO_WAIT_DONE] = 1;
3089 mutex_exit(&zio->io_lock);
3090
3091 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3092 zio_link_t *zl = zio->io_walk_link;
3093 pio_next = zio_walk_parents(zio);
3094 zio_remove_child(pio, zio, zl);
3095 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3096 }
3097
3098 if (zio->io_waiter != NULL) {
3099 mutex_enter(&zio->io_lock);
3100 zio->io_executor = NULL;
3101 cv_broadcast(&zio->io_cv);
3102 mutex_exit(&zio->io_lock);
3103 } else {
3104 zio_destroy(zio);
3105 }
3106
3107 return (ZIO_PIPELINE_STOP);
3108 }
3109
3110 /*
3111 * ==========================================================================
3112 * I/O pipeline definition
3113 * ==========================================================================
3114 */
3115 static zio_pipe_stage_t *zio_pipeline[] = {
3116 NULL,
3117 zio_read_bp_init,
3118 zio_free_bp_init,
3119 zio_issue_async,
3120 zio_write_bp_init,
3121 zio_checksum_generate,
3122 zio_nop_write,
3123 zio_ddt_read_start,
3124 zio_ddt_read_done,
3125 zio_ddt_write,
3126 zio_ddt_free,
3127 zio_gang_assemble,
3128 zio_gang_issue,
3129 zio_dva_allocate,
3130 zio_dva_free,
3131 zio_dva_claim,
3132 zio_ready,
3133 zio_vdev_io_start,
3134 zio_vdev_io_done,
3135 zio_vdev_io_assess,
3136 zio_checksum_verify,
3137 zio_done
3138 };
3139
3140 /* dnp is the dnode for zb1->zb_object */
3141 boolean_t
3142 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3143 const zbookmark_t *zb2)
3144 {
3145 uint64_t zb1nextL0, zb2thisobj;
3146
3147 ASSERT(zb1->zb_objset == zb2->zb_objset);
3148 ASSERT(zb2->zb_level == 0);
3149
3150 /*
3151 * A bookmark in the deadlist is considered to be after
3152 * everything else.
3153 */
3154 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3155 return (B_TRUE);
3156
3157 /* The objset_phys_t isn't before anything. */
3158 if (dnp == NULL)
3159 return (B_FALSE);
3160
3161 zb1nextL0 = (zb1->zb_blkid + 1) <<
3162 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3163
3164 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3165 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3166
3167 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3168 uint64_t nextobj = zb1nextL0 *
3169 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3170 return (nextobj <= zb2thisobj);
3171 }
3172
3173 if (zb1->zb_object < zb2thisobj)
3174 return (B_TRUE);
3175 if (zb1->zb_object > zb2thisobj)
3176 return (B_FALSE);
3177 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3178 return (B_FALSE);
3179 return (zb1nextL0 <= zb2->zb_blkid);
3180 }