xref: /titanic_41/usr/src/uts/common/fs/zfs/zio.c (revision 3fb517f786391b507780c78aabb8d98bfea9efe9)
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  */
24 
25 #include <sys/zfs_context.h>
26 #include <sys/fm/fs/zfs.h>
27 #include <sys/spa.h>
28 #include <sys/txg.h>
29 #include <sys/spa_impl.h>
30 #include <sys/vdev_impl.h>
31 #include <sys/zio_impl.h>
32 #include <sys/zio_compress.h>
33 #include <sys/zio_checksum.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/arc.h>
36 #include <sys/ddt.h>
37 
38 /*
39  * ==========================================================================
40  * I/O priority table
41  * ==========================================================================
42  */
43 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
44 	0,	/* ZIO_PRIORITY_NOW		*/
45 	0,	/* ZIO_PRIORITY_SYNC_READ	*/
46 	0,	/* ZIO_PRIORITY_SYNC_WRITE	*/
47 	0,	/* ZIO_PRIORITY_LOG_WRITE	*/
48 	1,	/* ZIO_PRIORITY_CACHE_FILL	*/
49 	1,	/* ZIO_PRIORITY_AGG		*/
50 	4,	/* ZIO_PRIORITY_FREE		*/
51 	4,	/* ZIO_PRIORITY_ASYNC_WRITE	*/
52 	6,	/* ZIO_PRIORITY_ASYNC_READ	*/
53 	10,	/* ZIO_PRIORITY_RESILVER	*/
54 	20,	/* ZIO_PRIORITY_SCRUB		*/
55 };
56 
57 /*
58  * ==========================================================================
59  * I/O type descriptions
60  * ==========================================================================
61  */
62 char *zio_type_name[ZIO_TYPES] = {
63 	"zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
64 	"zio_ioctl"
65 };
66 
67 /*
68  * ==========================================================================
69  * I/O kmem caches
70  * ==========================================================================
71  */
72 kmem_cache_t *zio_cache;
73 kmem_cache_t *zio_link_cache;
74 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
75 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
76 
77 #ifdef _KERNEL
78 extern vmem_t *zio_alloc_arena;
79 #endif
80 
81 /*
82  * An allocating zio is one that either currently has the DVA allocate
83  * stage set or will have it later in its lifetime.
84  */
85 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
86 
87 boolean_t	zio_requeue_io_start_cut_in_line = B_TRUE;
88 
89 #ifdef ZFS_DEBUG
90 int zio_buf_debug_limit = 16384;
91 #else
92 int zio_buf_debug_limit = 0;
93 #endif
94 
95 void
96 zio_init(void)
97 {
98 	size_t c;
99 	vmem_t *data_alloc_arena = NULL;
100 
101 #ifdef _KERNEL
102 	data_alloc_arena = zio_alloc_arena;
103 #endif
104 	zio_cache = kmem_cache_create("zio_cache",
105 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
106 	zio_link_cache = kmem_cache_create("zio_link_cache",
107 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
108 
109 	/*
110 	 * For small buffers, we want a cache for each multiple of
111 	 * SPA_MINBLOCKSIZE.  For medium-size buffers, we want a cache
112 	 * for each quarter-power of 2.  For large buffers, we want
113 	 * a cache for each multiple of PAGESIZE.
114 	 */
115 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
116 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
117 		size_t p2 = size;
118 		size_t align = 0;
119 
120 		while (p2 & (p2 - 1))
121 			p2 &= p2 - 1;
122 
123 		if (size <= 4 * SPA_MINBLOCKSIZE) {
124 			align = SPA_MINBLOCKSIZE;
125 		} else if (P2PHASE(size, PAGESIZE) == 0) {
126 			align = PAGESIZE;
127 		} else if (P2PHASE(size, p2 >> 2) == 0) {
128 			align = p2 >> 2;
129 		}
130 
131 		if (align != 0) {
132 			char name[36];
133 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
134 			zio_buf_cache[c] = kmem_cache_create(name, size,
135 			    align, NULL, NULL, NULL, NULL, NULL,
136 			    size > zio_buf_debug_limit ? KMC_NODEBUG : 0);
137 
138 			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
139 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
140 			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
141 			    size > zio_buf_debug_limit ? KMC_NODEBUG : 0);
142 		}
143 	}
144 
145 	while (--c != 0) {
146 		ASSERT(zio_buf_cache[c] != NULL);
147 		if (zio_buf_cache[c - 1] == NULL)
148 			zio_buf_cache[c - 1] = zio_buf_cache[c];
149 
150 		ASSERT(zio_data_buf_cache[c] != NULL);
151 		if (zio_data_buf_cache[c - 1] == NULL)
152 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
153 	}
154 
155 	zio_inject_init();
156 }
157 
158 void
159 zio_fini(void)
160 {
161 	size_t c;
162 	kmem_cache_t *last_cache = NULL;
163 	kmem_cache_t *last_data_cache = NULL;
164 
165 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
166 		if (zio_buf_cache[c] != last_cache) {
167 			last_cache = zio_buf_cache[c];
168 			kmem_cache_destroy(zio_buf_cache[c]);
169 		}
170 		zio_buf_cache[c] = NULL;
171 
172 		if (zio_data_buf_cache[c] != last_data_cache) {
173 			last_data_cache = zio_data_buf_cache[c];
174 			kmem_cache_destroy(zio_data_buf_cache[c]);
175 		}
176 		zio_data_buf_cache[c] = NULL;
177 	}
178 
179 	kmem_cache_destroy(zio_link_cache);
180 	kmem_cache_destroy(zio_cache);
181 
182 	zio_inject_fini();
183 }
184 
185 /*
186  * ==========================================================================
187  * Allocate and free I/O buffers
188  * ==========================================================================
189  */
190 
191 /*
192  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
193  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
194  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
195  * excess / transient data in-core during a crashdump.
196  */
197 void *
198 zio_buf_alloc(size_t size)
199 {
200 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
201 
202 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
203 
204 	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
205 }
206 
207 /*
208  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
209  * crashdump if the kernel panics.  This exists so that we will limit the amount
210  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
211  * of kernel heap dumped to disk when the kernel panics)
212  */
213 void *
214 zio_data_buf_alloc(size_t size)
215 {
216 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
217 
218 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
219 
220 	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
221 }
222 
223 void
224 zio_buf_free(void *buf, size_t size)
225 {
226 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
227 
228 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
229 
230 	kmem_cache_free(zio_buf_cache[c], buf);
231 }
232 
233 void
234 zio_data_buf_free(void *buf, size_t size)
235 {
236 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
237 
238 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
239 
240 	kmem_cache_free(zio_data_buf_cache[c], buf);
241 }
242 
243 /*
244  * ==========================================================================
245  * Push and pop I/O transform buffers
246  * ==========================================================================
247  */
248 static void
249 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
250 	zio_transform_func_t *transform)
251 {
252 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
253 
254 	zt->zt_orig_data = zio->io_data;
255 	zt->zt_orig_size = zio->io_size;
256 	zt->zt_bufsize = bufsize;
257 	zt->zt_transform = transform;
258 
259 	zt->zt_next = zio->io_transform_stack;
260 	zio->io_transform_stack = zt;
261 
262 	zio->io_data = data;
263 	zio->io_size = size;
264 }
265 
266 static void
267 zio_pop_transforms(zio_t *zio)
268 {
269 	zio_transform_t *zt;
270 
271 	while ((zt = zio->io_transform_stack) != NULL) {
272 		if (zt->zt_transform != NULL)
273 			zt->zt_transform(zio,
274 			    zt->zt_orig_data, zt->zt_orig_size);
275 
276 		if (zt->zt_bufsize != 0)
277 			zio_buf_free(zio->io_data, zt->zt_bufsize);
278 
279 		zio->io_data = zt->zt_orig_data;
280 		zio->io_size = zt->zt_orig_size;
281 		zio->io_transform_stack = zt->zt_next;
282 
283 		kmem_free(zt, sizeof (zio_transform_t));
284 	}
285 }
286 
287 /*
288  * ==========================================================================
289  * I/O transform callbacks for subblocks and decompression
290  * ==========================================================================
291  */
292 static void
293 zio_subblock(zio_t *zio, void *data, uint64_t size)
294 {
295 	ASSERT(zio->io_size > size);
296 
297 	if (zio->io_type == ZIO_TYPE_READ)
298 		bcopy(zio->io_data, data, size);
299 }
300 
301 static void
302 zio_decompress(zio_t *zio, void *data, uint64_t size)
303 {
304 	if (zio->io_error == 0 &&
305 	    zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
306 	    zio->io_data, data, zio->io_size, size) != 0)
307 		zio->io_error = EIO;
308 }
309 
310 /*
311  * ==========================================================================
312  * I/O parent/child relationships and pipeline interlocks
313  * ==========================================================================
314  */
315 /*
316  * NOTE - Callers to zio_walk_parents() and zio_walk_children must
317  *        continue calling these functions until they return NULL.
318  *        Otherwise, the next caller will pick up the list walk in
319  *        some indeterminate state.  (Otherwise every caller would
320  *        have to pass in a cookie to keep the state represented by
321  *        io_walk_link, which gets annoying.)
322  */
323 zio_t *
324 zio_walk_parents(zio_t *cio)
325 {
326 	zio_link_t *zl = cio->io_walk_link;
327 	list_t *pl = &cio->io_parent_list;
328 
329 	zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
330 	cio->io_walk_link = zl;
331 
332 	if (zl == NULL)
333 		return (NULL);
334 
335 	ASSERT(zl->zl_child == cio);
336 	return (zl->zl_parent);
337 }
338 
339 zio_t *
340 zio_walk_children(zio_t *pio)
341 {
342 	zio_link_t *zl = pio->io_walk_link;
343 	list_t *cl = &pio->io_child_list;
344 
345 	zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
346 	pio->io_walk_link = zl;
347 
348 	if (zl == NULL)
349 		return (NULL);
350 
351 	ASSERT(zl->zl_parent == pio);
352 	return (zl->zl_child);
353 }
354 
355 zio_t *
356 zio_unique_parent(zio_t *cio)
357 {
358 	zio_t *pio = zio_walk_parents(cio);
359 
360 	VERIFY(zio_walk_parents(cio) == NULL);
361 	return (pio);
362 }
363 
364 void
365 zio_add_child(zio_t *pio, zio_t *cio)
366 {
367 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
368 
369 	/*
370 	 * Logical I/Os can have logical, gang, or vdev children.
371 	 * Gang I/Os can have gang or vdev children.
372 	 * Vdev I/Os can only have vdev children.
373 	 * The following ASSERT captures all of these constraints.
374 	 */
375 	ASSERT(cio->io_child_type <= pio->io_child_type);
376 
377 	zl->zl_parent = pio;
378 	zl->zl_child = cio;
379 
380 	mutex_enter(&cio->io_lock);
381 	mutex_enter(&pio->io_lock);
382 
383 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
384 
385 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
386 		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
387 
388 	list_insert_head(&pio->io_child_list, zl);
389 	list_insert_head(&cio->io_parent_list, zl);
390 
391 	pio->io_child_count++;
392 	cio->io_parent_count++;
393 
394 	mutex_exit(&pio->io_lock);
395 	mutex_exit(&cio->io_lock);
396 }
397 
398 static void
399 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
400 {
401 	ASSERT(zl->zl_parent == pio);
402 	ASSERT(zl->zl_child == cio);
403 
404 	mutex_enter(&cio->io_lock);
405 	mutex_enter(&pio->io_lock);
406 
407 	list_remove(&pio->io_child_list, zl);
408 	list_remove(&cio->io_parent_list, zl);
409 
410 	pio->io_child_count--;
411 	cio->io_parent_count--;
412 
413 	mutex_exit(&pio->io_lock);
414 	mutex_exit(&cio->io_lock);
415 
416 	kmem_cache_free(zio_link_cache, zl);
417 }
418 
419 static boolean_t
420 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
421 {
422 	uint64_t *countp = &zio->io_children[child][wait];
423 	boolean_t waiting = B_FALSE;
424 
425 	mutex_enter(&zio->io_lock);
426 	ASSERT(zio->io_stall == NULL);
427 	if (*countp != 0) {
428 		zio->io_stage >>= 1;
429 		zio->io_stall = countp;
430 		waiting = B_TRUE;
431 	}
432 	mutex_exit(&zio->io_lock);
433 
434 	return (waiting);
435 }
436 
437 static void
438 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
439 {
440 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
441 	int *errorp = &pio->io_child_error[zio->io_child_type];
442 
443 	mutex_enter(&pio->io_lock);
444 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
445 		*errorp = zio_worst_error(*errorp, zio->io_error);
446 	pio->io_reexecute |= zio->io_reexecute;
447 	ASSERT3U(*countp, >, 0);
448 	if (--*countp == 0 && pio->io_stall == countp) {
449 		pio->io_stall = NULL;
450 		mutex_exit(&pio->io_lock);
451 		zio_execute(pio);
452 	} else {
453 		mutex_exit(&pio->io_lock);
454 	}
455 }
456 
457 static void
458 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
459 {
460 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
461 		zio->io_error = zio->io_child_error[c];
462 }
463 
464 /*
465  * ==========================================================================
466  * Create the various types of I/O (read, write, free, etc)
467  * ==========================================================================
468  */
469 static zio_t *
470 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
471     void *data, uint64_t size, zio_done_func_t *done, void *private,
472     zio_type_t type, int priority, enum zio_flag flags,
473     vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
474     enum zio_stage stage, enum zio_stage pipeline)
475 {
476 	zio_t *zio;
477 
478 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
479 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
480 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
481 
482 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
483 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
484 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
485 
486 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
487 	bzero(zio, sizeof (zio_t));
488 
489 	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
490 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
491 
492 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
493 	    offsetof(zio_link_t, zl_parent_node));
494 	list_create(&zio->io_child_list, sizeof (zio_link_t),
495 	    offsetof(zio_link_t, zl_child_node));
496 
497 	if (vd != NULL)
498 		zio->io_child_type = ZIO_CHILD_VDEV;
499 	else if (flags & ZIO_FLAG_GANG_CHILD)
500 		zio->io_child_type = ZIO_CHILD_GANG;
501 	else if (flags & ZIO_FLAG_DDT_CHILD)
502 		zio->io_child_type = ZIO_CHILD_DDT;
503 	else
504 		zio->io_child_type = ZIO_CHILD_LOGICAL;
505 
506 	if (bp != NULL) {
507 		zio->io_bp = (blkptr_t *)bp;
508 		zio->io_bp_copy = *bp;
509 		zio->io_bp_orig = *bp;
510 		if (type != ZIO_TYPE_WRITE ||
511 		    zio->io_child_type == ZIO_CHILD_DDT)
512 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
513 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
514 			zio->io_logical = zio;
515 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
516 			pipeline |= ZIO_GANG_STAGES;
517 	}
518 
519 	zio->io_spa = spa;
520 	zio->io_txg = txg;
521 	zio->io_done = done;
522 	zio->io_private = private;
523 	zio->io_type = type;
524 	zio->io_priority = priority;
525 	zio->io_vd = vd;
526 	zio->io_offset = offset;
527 	zio->io_orig_data = zio->io_data = data;
528 	zio->io_orig_size = zio->io_size = size;
529 	zio->io_orig_flags = zio->io_flags = flags;
530 	zio->io_orig_stage = zio->io_stage = stage;
531 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
532 
533 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
534 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
535 
536 	if (zb != NULL)
537 		zio->io_bookmark = *zb;
538 
539 	if (pio != NULL) {
540 		if (zio->io_logical == NULL)
541 			zio->io_logical = pio->io_logical;
542 		if (zio->io_child_type == ZIO_CHILD_GANG)
543 			zio->io_gang_leader = pio->io_gang_leader;
544 		zio_add_child(pio, zio);
545 	}
546 
547 	return (zio);
548 }
549 
550 static void
551 zio_destroy(zio_t *zio)
552 {
553 	list_destroy(&zio->io_parent_list);
554 	list_destroy(&zio->io_child_list);
555 	mutex_destroy(&zio->io_lock);
556 	cv_destroy(&zio->io_cv);
557 	kmem_cache_free(zio_cache, zio);
558 }
559 
560 zio_t *
561 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
562     void *private, enum zio_flag flags)
563 {
564 	zio_t *zio;
565 
566 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
567 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
568 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
569 
570 	return (zio);
571 }
572 
573 zio_t *
574 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
575 {
576 	return (zio_null(NULL, spa, NULL, done, private, flags));
577 }
578 
579 zio_t *
580 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
581     void *data, uint64_t size, zio_done_func_t *done, void *private,
582     int priority, enum zio_flag flags, const zbookmark_t *zb)
583 {
584 	zio_t *zio;
585 
586 	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
587 	    data, size, done, private,
588 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
589 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
590 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
591 
592 	return (zio);
593 }
594 
595 zio_t *
596 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
597     void *data, uint64_t size, const zio_prop_t *zp,
598     zio_done_func_t *ready, zio_done_func_t *done, void *private,
599     int priority, enum zio_flag flags, const zbookmark_t *zb)
600 {
601 	zio_t *zio;
602 
603 	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
604 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
605 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
606 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
607 	    zp->zp_type < DMU_OT_NUMTYPES &&
608 	    zp->zp_level < 32 &&
609 	    zp->zp_copies > 0 &&
610 	    zp->zp_copies <= spa_max_replication(spa) &&
611 	    zp->zp_dedup <= 1 &&
612 	    zp->zp_dedup_verify <= 1);
613 
614 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
615 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
616 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
617 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
618 
619 	zio->io_ready = ready;
620 	zio->io_prop = *zp;
621 
622 	return (zio);
623 }
624 
625 zio_t *
626 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
627     uint64_t size, zio_done_func_t *done, void *private, int priority,
628     enum zio_flag flags, zbookmark_t *zb)
629 {
630 	zio_t *zio;
631 
632 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
633 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
634 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
635 
636 	return (zio);
637 }
638 
639 void
640 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
641 {
642 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
643 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
644 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
645 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
646 
647 	zio->io_prop.zp_copies = copies;
648 	zio->io_bp_override = bp;
649 }
650 
651 void
652 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
653 {
654 	bplist_enqueue_deferred(&spa->spa_free_bplist[txg & TXG_MASK], bp);
655 }
656 
657 zio_t *
658 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
659     enum zio_flag flags)
660 {
661 	zio_t *zio;
662 
663 	dprintf_bp(bp, "freeing in txg %llu, pass %u",
664 	    (longlong_t)txg, spa->spa_sync_pass);
665 
666 	ASSERT(!BP_IS_HOLE(bp));
667 	ASSERT(spa_syncing_txg(spa) == txg);
668 	ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
669 
670 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
671 	    NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
672 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
673 
674 	return (zio);
675 }
676 
677 zio_t *
678 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
679     zio_done_func_t *done, void *private, enum zio_flag flags)
680 {
681 	zio_t *zio;
682 
683 	/*
684 	 * A claim is an allocation of a specific block.  Claims are needed
685 	 * to support immediate writes in the intent log.  The issue is that
686 	 * immediate writes contain committed data, but in a txg that was
687 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
688 	 * the intent log claims all blocks that contain immediate write data
689 	 * so that the SPA knows they're in use.
690 	 *
691 	 * All claims *must* be resolved in the first txg -- before the SPA
692 	 * starts allocating blocks -- so that nothing is allocated twice.
693 	 * If txg == 0 we just verify that the block is claimable.
694 	 */
695 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
696 	ASSERT(txg == spa_first_txg(spa) || txg == 0);
697 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
698 
699 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
700 	    done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
701 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
702 
703 	return (zio);
704 }
705 
706 zio_t *
707 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
708     zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
709 {
710 	zio_t *zio;
711 	int c;
712 
713 	if (vd->vdev_children == 0) {
714 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
715 		    ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
716 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
717 
718 		zio->io_cmd = cmd;
719 	} else {
720 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
721 
722 		for (c = 0; c < vd->vdev_children; c++)
723 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
724 			    done, private, priority, flags));
725 	}
726 
727 	return (zio);
728 }
729 
730 zio_t *
731 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
732     void *data, int checksum, zio_done_func_t *done, void *private,
733     int priority, enum zio_flag flags, boolean_t labels)
734 {
735 	zio_t *zio;
736 
737 	ASSERT(vd->vdev_children == 0);
738 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
739 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
740 	ASSERT3U(offset + size, <=, vd->vdev_psize);
741 
742 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
743 	    ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
744 	    ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
745 
746 	zio->io_prop.zp_checksum = checksum;
747 
748 	return (zio);
749 }
750 
751 zio_t *
752 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
753     void *data, int checksum, zio_done_func_t *done, void *private,
754     int priority, enum zio_flag flags, boolean_t labels)
755 {
756 	zio_t *zio;
757 
758 	ASSERT(vd->vdev_children == 0);
759 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
760 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
761 	ASSERT3U(offset + size, <=, vd->vdev_psize);
762 
763 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
764 	    ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
765 	    ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
766 
767 	zio->io_prop.zp_checksum = checksum;
768 
769 	if (zio_checksum_table[checksum].ci_eck) {
770 		/*
771 		 * zec checksums are necessarily destructive -- they modify
772 		 * the end of the write buffer to hold the verifier/checksum.
773 		 * Therefore, we must make a local copy in case the data is
774 		 * being written to multiple places in parallel.
775 		 */
776 		void *wbuf = zio_buf_alloc(size);
777 		bcopy(data, wbuf, size);
778 		zio_push_transform(zio, wbuf, size, size, NULL);
779 	}
780 
781 	return (zio);
782 }
783 
784 /*
785  * Create a child I/O to do some work for us.
786  */
787 zio_t *
788 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
789 	void *data, uint64_t size, int type, int priority, enum zio_flag flags,
790 	zio_done_func_t *done, void *private)
791 {
792 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
793 	zio_t *zio;
794 
795 	ASSERT(vd->vdev_parent ==
796 	    (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
797 
798 	if (type == ZIO_TYPE_READ && bp != NULL) {
799 		/*
800 		 * If we have the bp, then the child should perform the
801 		 * checksum and the parent need not.  This pushes error
802 		 * detection as close to the leaves as possible and
803 		 * eliminates redundant checksums in the interior nodes.
804 		 */
805 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
806 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
807 	}
808 
809 	if (vd->vdev_children == 0)
810 		offset += VDEV_LABEL_START_SIZE;
811 
812 	flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
813 
814 	/*
815 	 * If we've decided to do a repair, the write is not speculative --
816 	 * even if the original read was.
817 	 */
818 	if (flags & ZIO_FLAG_IO_REPAIR)
819 		flags &= ~ZIO_FLAG_SPECULATIVE;
820 
821 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
822 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
823 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
824 
825 	return (zio);
826 }
827 
828 zio_t *
829 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
830 	int type, int priority, enum zio_flag flags,
831 	zio_done_func_t *done, void *private)
832 {
833 	zio_t *zio;
834 
835 	ASSERT(vd->vdev_ops->vdev_op_leaf);
836 
837 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
838 	    data, size, done, private, type, priority,
839 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
840 	    vd, offset, NULL,
841 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
842 
843 	return (zio);
844 }
845 
846 void
847 zio_flush(zio_t *zio, vdev_t *vd)
848 {
849 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
850 	    NULL, NULL, ZIO_PRIORITY_NOW,
851 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
852 }
853 
854 void
855 zio_shrink(zio_t *zio, uint64_t size)
856 {
857 	ASSERT(zio->io_executor == NULL);
858 	ASSERT(zio->io_orig_size == zio->io_size);
859 	ASSERT(size <= zio->io_size);
860 
861 	/*
862 	 * We don't shrink for raidz because of problems with the
863 	 * reconstruction when reading back less than the block size.
864 	 * Note, BP_IS_RAIDZ() assumes no compression.
865 	 */
866 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
867 	if (!BP_IS_RAIDZ(zio->io_bp))
868 		zio->io_orig_size = zio->io_size = size;
869 }
870 
871 /*
872  * ==========================================================================
873  * Prepare to read and write logical blocks
874  * ==========================================================================
875  */
876 
877 static int
878 zio_read_bp_init(zio_t *zio)
879 {
880 	blkptr_t *bp = zio->io_bp;
881 
882 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
883 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
884 	    !(zio->io_flags & ZIO_FLAG_RAW)) {
885 		uint64_t psize = BP_GET_PSIZE(bp);
886 		void *cbuf = zio_buf_alloc(psize);
887 
888 		zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
889 	}
890 
891 	if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
892 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
893 
894 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
895 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
896 
897 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
898 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
899 
900 	return (ZIO_PIPELINE_CONTINUE);
901 }
902 
903 static int
904 zio_write_bp_init(zio_t *zio)
905 {
906 	spa_t *spa = zio->io_spa;
907 	zio_prop_t *zp = &zio->io_prop;
908 	enum zio_compress compress = zp->zp_compress;
909 	blkptr_t *bp = zio->io_bp;
910 	uint64_t lsize = zio->io_size;
911 	uint64_t psize = lsize;
912 	int pass = 1;
913 
914 	/*
915 	 * If our children haven't all reached the ready stage,
916 	 * wait for them and then repeat this pipeline stage.
917 	 */
918 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
919 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
920 		return (ZIO_PIPELINE_STOP);
921 
922 	if (!IO_IS_ALLOCATING(zio))
923 		return (ZIO_PIPELINE_CONTINUE);
924 
925 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
926 
927 	if (zio->io_bp_override) {
928 		ASSERT(bp->blk_birth != zio->io_txg);
929 		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
930 
931 		*bp = *zio->io_bp_override;
932 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
933 
934 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
935 			return (ZIO_PIPELINE_CONTINUE);
936 
937 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
938 		    zp->zp_dedup_verify);
939 
940 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
941 			BP_SET_DEDUP(bp, 1);
942 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
943 			return (ZIO_PIPELINE_CONTINUE);
944 		}
945 		zio->io_bp_override = NULL;
946 		BP_ZERO(bp);
947 	}
948 
949 	if (bp->blk_birth == zio->io_txg) {
950 		/*
951 		 * We're rewriting an existing block, which means we're
952 		 * working on behalf of spa_sync().  For spa_sync() to
953 		 * converge, it must eventually be the case that we don't
954 		 * have to allocate new blocks.  But compression changes
955 		 * the blocksize, which forces a reallocate, and makes
956 		 * convergence take longer.  Therefore, after the first
957 		 * few passes, stop compressing to ensure convergence.
958 		 */
959 		pass = spa_sync_pass(spa);
960 
961 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
962 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
963 		ASSERT(!BP_GET_DEDUP(bp));
964 
965 		if (pass > SYNC_PASS_DONT_COMPRESS)
966 			compress = ZIO_COMPRESS_OFF;
967 
968 		/* Make sure someone doesn't change their mind on overwrites */
969 		ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
970 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
971 	}
972 
973 	if (compress != ZIO_COMPRESS_OFF) {
974 		void *cbuf = zio_buf_alloc(lsize);
975 		psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
976 		if (psize == 0 || psize == lsize) {
977 			compress = ZIO_COMPRESS_OFF;
978 			zio_buf_free(cbuf, lsize);
979 		} else {
980 			ASSERT(psize < lsize);
981 			zio_push_transform(zio, cbuf, psize, lsize, NULL);
982 		}
983 	}
984 
985 	/*
986 	 * The final pass of spa_sync() must be all rewrites, but the first
987 	 * few passes offer a trade-off: allocating blocks defers convergence,
988 	 * but newly allocated blocks are sequential, so they can be written
989 	 * to disk faster.  Therefore, we allow the first few passes of
990 	 * spa_sync() to allocate new blocks, but force rewrites after that.
991 	 * There should only be a handful of blocks after pass 1 in any case.
992 	 */
993 	if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
994 	    pass > SYNC_PASS_REWRITE) {
995 		ASSERT(psize != 0);
996 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
997 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
998 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
999 	} else {
1000 		BP_ZERO(bp);
1001 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1002 	}
1003 
1004 	if (psize == 0) {
1005 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1006 	} else {
1007 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1008 		BP_SET_LSIZE(bp, lsize);
1009 		BP_SET_PSIZE(bp, psize);
1010 		BP_SET_COMPRESS(bp, compress);
1011 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1012 		BP_SET_TYPE(bp, zp->zp_type);
1013 		BP_SET_LEVEL(bp, zp->zp_level);
1014 		BP_SET_DEDUP(bp, zp->zp_dedup);
1015 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1016 		if (zp->zp_dedup) {
1017 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1018 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1019 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1020 		}
1021 	}
1022 
1023 	return (ZIO_PIPELINE_CONTINUE);
1024 }
1025 
1026 static int
1027 zio_free_bp_init(zio_t *zio)
1028 {
1029 	blkptr_t *bp = zio->io_bp;
1030 
1031 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1032 		if (BP_GET_DEDUP(bp))
1033 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1034 		else
1035 			arc_free(zio->io_spa, bp);
1036 	}
1037 
1038 	return (ZIO_PIPELINE_CONTINUE);
1039 }
1040 
1041 /*
1042  * ==========================================================================
1043  * Execute the I/O pipeline
1044  * ==========================================================================
1045  */
1046 
1047 static void
1048 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1049 {
1050 	spa_t *spa = zio->io_spa;
1051 	zio_type_t t = zio->io_type;
1052 	int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0);
1053 
1054 	/*
1055 	 * If we're a config writer or a probe, the normal issue and
1056 	 * interrupt threads may all be blocked waiting for the config lock.
1057 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1058 	 */
1059 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1060 		t = ZIO_TYPE_NULL;
1061 
1062 	/*
1063 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1064 	 */
1065 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1066 		t = ZIO_TYPE_NULL;
1067 
1068 	/*
1069 	 * If this is a high priority I/O, then use the high priority taskq.
1070 	 */
1071 	if (zio->io_priority == ZIO_PRIORITY_NOW &&
1072 	    spa->spa_zio_taskq[t][q + 1] != NULL)
1073 		q++;
1074 
1075 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1076 	(void) taskq_dispatch(spa->spa_zio_taskq[t][q],
1077 	    (task_func_t *)zio_execute, zio, flags);
1078 }
1079 
1080 static boolean_t
1081 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1082 {
1083 	kthread_t *executor = zio->io_executor;
1084 	spa_t *spa = zio->io_spa;
1085 
1086 	for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1087 		if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1088 			return (B_TRUE);
1089 
1090 	return (B_FALSE);
1091 }
1092 
1093 static int
1094 zio_issue_async(zio_t *zio)
1095 {
1096 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1097 
1098 	return (ZIO_PIPELINE_STOP);
1099 }
1100 
1101 void
1102 zio_interrupt(zio_t *zio)
1103 {
1104 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1105 }
1106 
1107 /*
1108  * Execute the I/O pipeline until one of the following occurs:
1109  * (1) the I/O completes; (2) the pipeline stalls waiting for
1110  * dependent child I/Os; (3) the I/O issues, so we're waiting
1111  * for an I/O completion interrupt; (4) the I/O is delegated by
1112  * vdev-level caching or aggregation; (5) the I/O is deferred
1113  * due to vdev-level queueing; (6) the I/O is handed off to
1114  * another thread.  In all cases, the pipeline stops whenever
1115  * there's no CPU work; it never burns a thread in cv_wait().
1116  *
1117  * There's no locking on io_stage because there's no legitimate way
1118  * for multiple threads to be attempting to process the same I/O.
1119  */
1120 static zio_pipe_stage_t *zio_pipeline[];
1121 
1122 void
1123 zio_execute(zio_t *zio)
1124 {
1125 	zio->io_executor = curthread;
1126 
1127 	while (zio->io_stage < ZIO_STAGE_DONE) {
1128 		enum zio_stage pipeline = zio->io_pipeline;
1129 		enum zio_stage stage = zio->io_stage;
1130 		int rv;
1131 
1132 		ASSERT(!MUTEX_HELD(&zio->io_lock));
1133 		ASSERT(ISP2(stage));
1134 		ASSERT(zio->io_stall == NULL);
1135 
1136 		do {
1137 			stage <<= 1;
1138 		} while ((stage & pipeline) == 0);
1139 
1140 		ASSERT(stage <= ZIO_STAGE_DONE);
1141 
1142 		/*
1143 		 * If we are in interrupt context and this pipeline stage
1144 		 * will grab a config lock that is held across I/O,
1145 		 * or may wait for an I/O that needs an interrupt thread
1146 		 * to complete, issue async to avoid deadlock.
1147 		 *
1148 		 * For VDEV_IO_START, we cut in line so that the io will
1149 		 * be sent to disk promptly.
1150 		 */
1151 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1152 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1153 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1154 			    zio_requeue_io_start_cut_in_line : B_FALSE;
1155 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1156 			return;
1157 		}
1158 
1159 		zio->io_stage = stage;
1160 		rv = zio_pipeline[highbit(stage) - 1](zio);
1161 
1162 		if (rv == ZIO_PIPELINE_STOP)
1163 			return;
1164 
1165 		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1166 	}
1167 }
1168 
1169 /*
1170  * ==========================================================================
1171  * Initiate I/O, either sync or async
1172  * ==========================================================================
1173  */
1174 int
1175 zio_wait(zio_t *zio)
1176 {
1177 	int error;
1178 
1179 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1180 	ASSERT(zio->io_executor == NULL);
1181 
1182 	zio->io_waiter = curthread;
1183 
1184 	zio_execute(zio);
1185 
1186 	mutex_enter(&zio->io_lock);
1187 	while (zio->io_executor != NULL)
1188 		cv_wait(&zio->io_cv, &zio->io_lock);
1189 	mutex_exit(&zio->io_lock);
1190 
1191 	error = zio->io_error;
1192 	zio_destroy(zio);
1193 
1194 	return (error);
1195 }
1196 
1197 void
1198 zio_nowait(zio_t *zio)
1199 {
1200 	ASSERT(zio->io_executor == NULL);
1201 
1202 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1203 	    zio_unique_parent(zio) == NULL) {
1204 		/*
1205 		 * This is a logical async I/O with no parent to wait for it.
1206 		 * We add it to the spa_async_root_zio "Godfather" I/O which
1207 		 * will ensure they complete prior to unloading the pool.
1208 		 */
1209 		spa_t *spa = zio->io_spa;
1210 
1211 		zio_add_child(spa->spa_async_zio_root, zio);
1212 	}
1213 
1214 	zio_execute(zio);
1215 }
1216 
1217 /*
1218  * ==========================================================================
1219  * Reexecute or suspend/resume failed I/O
1220  * ==========================================================================
1221  */
1222 
1223 static void
1224 zio_reexecute(zio_t *pio)
1225 {
1226 	zio_t *cio, *cio_next;
1227 
1228 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1229 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1230 	ASSERT(pio->io_gang_leader == NULL);
1231 	ASSERT(pio->io_gang_tree == NULL);
1232 
1233 	pio->io_flags = pio->io_orig_flags;
1234 	pio->io_stage = pio->io_orig_stage;
1235 	pio->io_pipeline = pio->io_orig_pipeline;
1236 	pio->io_reexecute = 0;
1237 	pio->io_error = 0;
1238 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1239 		pio->io_state[w] = 0;
1240 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1241 		pio->io_child_error[c] = 0;
1242 
1243 	if (IO_IS_ALLOCATING(pio))
1244 		BP_ZERO(pio->io_bp);
1245 
1246 	/*
1247 	 * As we reexecute pio's children, new children could be created.
1248 	 * New children go to the head of pio's io_child_list, however,
1249 	 * so we will (correctly) not reexecute them.  The key is that
1250 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1251 	 * cannot be affected by any side effects of reexecuting 'cio'.
1252 	 */
1253 	for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1254 		cio_next = zio_walk_children(pio);
1255 		mutex_enter(&pio->io_lock);
1256 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1257 			pio->io_children[cio->io_child_type][w]++;
1258 		mutex_exit(&pio->io_lock);
1259 		zio_reexecute(cio);
1260 	}
1261 
1262 	/*
1263 	 * Now that all children have been reexecuted, execute the parent.
1264 	 * We don't reexecute "The Godfather" I/O here as it's the
1265 	 * responsibility of the caller to wait on him.
1266 	 */
1267 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1268 		zio_execute(pio);
1269 }
1270 
1271 void
1272 zio_suspend(spa_t *spa, zio_t *zio)
1273 {
1274 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1275 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1276 		    "failure and the failure mode property for this pool "
1277 		    "is set to panic.", spa_name(spa));
1278 
1279 	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1280 
1281 	mutex_enter(&spa->spa_suspend_lock);
1282 
1283 	if (spa->spa_suspend_zio_root == NULL)
1284 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1285 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1286 		    ZIO_FLAG_GODFATHER);
1287 
1288 	spa->spa_suspended = B_TRUE;
1289 
1290 	if (zio != NULL) {
1291 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1292 		ASSERT(zio != spa->spa_suspend_zio_root);
1293 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1294 		ASSERT(zio_unique_parent(zio) == NULL);
1295 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1296 		zio_add_child(spa->spa_suspend_zio_root, zio);
1297 	}
1298 
1299 	mutex_exit(&spa->spa_suspend_lock);
1300 }
1301 
1302 int
1303 zio_resume(spa_t *spa)
1304 {
1305 	zio_t *pio;
1306 
1307 	/*
1308 	 * Reexecute all previously suspended i/o.
1309 	 */
1310 	mutex_enter(&spa->spa_suspend_lock);
1311 	spa->spa_suspended = B_FALSE;
1312 	cv_broadcast(&spa->spa_suspend_cv);
1313 	pio = spa->spa_suspend_zio_root;
1314 	spa->spa_suspend_zio_root = NULL;
1315 	mutex_exit(&spa->spa_suspend_lock);
1316 
1317 	if (pio == NULL)
1318 		return (0);
1319 
1320 	zio_reexecute(pio);
1321 	return (zio_wait(pio));
1322 }
1323 
1324 void
1325 zio_resume_wait(spa_t *spa)
1326 {
1327 	mutex_enter(&spa->spa_suspend_lock);
1328 	while (spa_suspended(spa))
1329 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1330 	mutex_exit(&spa->spa_suspend_lock);
1331 }
1332 
1333 /*
1334  * ==========================================================================
1335  * Gang blocks.
1336  *
1337  * A gang block is a collection of small blocks that looks to the DMU
1338  * like one large block.  When zio_dva_allocate() cannot find a block
1339  * of the requested size, due to either severe fragmentation or the pool
1340  * being nearly full, it calls zio_write_gang_block() to construct the
1341  * block from smaller fragments.
1342  *
1343  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1344  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
1345  * an indirect block: it's an array of block pointers.  It consumes
1346  * only one sector and hence is allocatable regardless of fragmentation.
1347  * The gang header's bps point to its gang members, which hold the data.
1348  *
1349  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1350  * as the verifier to ensure uniqueness of the SHA256 checksum.
1351  * Critically, the gang block bp's blk_cksum is the checksum of the data,
1352  * not the gang header.  This ensures that data block signatures (needed for
1353  * deduplication) are independent of how the block is physically stored.
1354  *
1355  * Gang blocks can be nested: a gang member may itself be a gang block.
1356  * Thus every gang block is a tree in which root and all interior nodes are
1357  * gang headers, and the leaves are normal blocks that contain user data.
1358  * The root of the gang tree is called the gang leader.
1359  *
1360  * To perform any operation (read, rewrite, free, claim) on a gang block,
1361  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1362  * in the io_gang_tree field of the original logical i/o by recursively
1363  * reading the gang leader and all gang headers below it.  This yields
1364  * an in-core tree containing the contents of every gang header and the
1365  * bps for every constituent of the gang block.
1366  *
1367  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1368  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
1369  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1370  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1371  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1372  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
1373  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1374  * of the gang header plus zio_checksum_compute() of the data to update the
1375  * gang header's blk_cksum as described above.
1376  *
1377  * The two-phase assemble/issue model solves the problem of partial failure --
1378  * what if you'd freed part of a gang block but then couldn't read the
1379  * gang header for another part?  Assembling the entire gang tree first
1380  * ensures that all the necessary gang header I/O has succeeded before
1381  * starting the actual work of free, claim, or write.  Once the gang tree
1382  * is assembled, free and claim are in-memory operations that cannot fail.
1383  *
1384  * In the event that a gang write fails, zio_dva_unallocate() walks the
1385  * gang tree to immediately free (i.e. insert back into the space map)
1386  * everything we've allocated.  This ensures that we don't get ENOSPC
1387  * errors during repeated suspend/resume cycles due to a flaky device.
1388  *
1389  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
1390  * the gang tree, we won't modify the block, so we can safely defer the free
1391  * (knowing that the block is still intact).  If we *can* assemble the gang
1392  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1393  * each constituent bp and we can allocate a new block on the next sync pass.
1394  *
1395  * In all cases, the gang tree allows complete recovery from partial failure.
1396  * ==========================================================================
1397  */
1398 
1399 static zio_t *
1400 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1401 {
1402 	if (gn != NULL)
1403 		return (pio);
1404 
1405 	return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1406 	    NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1407 	    &pio->io_bookmark));
1408 }
1409 
1410 zio_t *
1411 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1412 {
1413 	zio_t *zio;
1414 
1415 	if (gn != NULL) {
1416 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1417 		    gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1418 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1419 		/*
1420 		 * As we rewrite each gang header, the pipeline will compute
1421 		 * a new gang block header checksum for it; but no one will
1422 		 * compute a new data checksum, so we do that here.  The one
1423 		 * exception is the gang leader: the pipeline already computed
1424 		 * its data checksum because that stage precedes gang assembly.
1425 		 * (Presently, nothing actually uses interior data checksums;
1426 		 * this is just good hygiene.)
1427 		 */
1428 		if (gn != pio->io_gang_leader->io_gang_tree) {
1429 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1430 			    data, BP_GET_PSIZE(bp));
1431 		}
1432 		/*
1433 		 * If we are here to damage data for testing purposes,
1434 		 * leave the GBH alone so that we can detect the damage.
1435 		 */
1436 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1437 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1438 	} else {
1439 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1440 		    data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1441 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1442 	}
1443 
1444 	return (zio);
1445 }
1446 
1447 /* ARGSUSED */
1448 zio_t *
1449 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1450 {
1451 	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1452 	    ZIO_GANG_CHILD_FLAGS(pio)));
1453 }
1454 
1455 /* ARGSUSED */
1456 zio_t *
1457 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1458 {
1459 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1460 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1461 }
1462 
1463 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1464 	NULL,
1465 	zio_read_gang,
1466 	zio_rewrite_gang,
1467 	zio_free_gang,
1468 	zio_claim_gang,
1469 	NULL
1470 };
1471 
1472 static void zio_gang_tree_assemble_done(zio_t *zio);
1473 
1474 static zio_gang_node_t *
1475 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1476 {
1477 	zio_gang_node_t *gn;
1478 
1479 	ASSERT(*gnpp == NULL);
1480 
1481 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1482 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1483 	*gnpp = gn;
1484 
1485 	return (gn);
1486 }
1487 
1488 static void
1489 zio_gang_node_free(zio_gang_node_t **gnpp)
1490 {
1491 	zio_gang_node_t *gn = *gnpp;
1492 
1493 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1494 		ASSERT(gn->gn_child[g] == NULL);
1495 
1496 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1497 	kmem_free(gn, sizeof (*gn));
1498 	*gnpp = NULL;
1499 }
1500 
1501 static void
1502 zio_gang_tree_free(zio_gang_node_t **gnpp)
1503 {
1504 	zio_gang_node_t *gn = *gnpp;
1505 
1506 	if (gn == NULL)
1507 		return;
1508 
1509 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1510 		zio_gang_tree_free(&gn->gn_child[g]);
1511 
1512 	zio_gang_node_free(gnpp);
1513 }
1514 
1515 static void
1516 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1517 {
1518 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1519 
1520 	ASSERT(gio->io_gang_leader == gio);
1521 	ASSERT(BP_IS_GANG(bp));
1522 
1523 	zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1524 	    SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1525 	    gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1526 }
1527 
1528 static void
1529 zio_gang_tree_assemble_done(zio_t *zio)
1530 {
1531 	zio_t *gio = zio->io_gang_leader;
1532 	zio_gang_node_t *gn = zio->io_private;
1533 	blkptr_t *bp = zio->io_bp;
1534 
1535 	ASSERT(gio == zio_unique_parent(zio));
1536 	ASSERT(zio->io_child_count == 0);
1537 
1538 	if (zio->io_error)
1539 		return;
1540 
1541 	if (BP_SHOULD_BYTESWAP(bp))
1542 		byteswap_uint64_array(zio->io_data, zio->io_size);
1543 
1544 	ASSERT(zio->io_data == gn->gn_gbh);
1545 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1546 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1547 
1548 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1549 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1550 		if (!BP_IS_GANG(gbp))
1551 			continue;
1552 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1553 	}
1554 }
1555 
1556 static void
1557 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1558 {
1559 	zio_t *gio = pio->io_gang_leader;
1560 	zio_t *zio;
1561 
1562 	ASSERT(BP_IS_GANG(bp) == !!gn);
1563 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1564 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1565 
1566 	/*
1567 	 * If you're a gang header, your data is in gn->gn_gbh.
1568 	 * If you're a gang member, your data is in 'data' and gn == NULL.
1569 	 */
1570 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1571 
1572 	if (gn != NULL) {
1573 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1574 
1575 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1576 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1577 			if (BP_IS_HOLE(gbp))
1578 				continue;
1579 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1580 			data = (char *)data + BP_GET_PSIZE(gbp);
1581 		}
1582 	}
1583 
1584 	if (gn == gio->io_gang_tree)
1585 		ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1586 
1587 	if (zio != pio)
1588 		zio_nowait(zio);
1589 }
1590 
1591 static int
1592 zio_gang_assemble(zio_t *zio)
1593 {
1594 	blkptr_t *bp = zio->io_bp;
1595 
1596 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1597 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1598 
1599 	zio->io_gang_leader = zio;
1600 
1601 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1602 
1603 	return (ZIO_PIPELINE_CONTINUE);
1604 }
1605 
1606 static int
1607 zio_gang_issue(zio_t *zio)
1608 {
1609 	blkptr_t *bp = zio->io_bp;
1610 
1611 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1612 		return (ZIO_PIPELINE_STOP);
1613 
1614 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1615 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1616 
1617 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1618 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1619 	else
1620 		zio_gang_tree_free(&zio->io_gang_tree);
1621 
1622 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1623 
1624 	return (ZIO_PIPELINE_CONTINUE);
1625 }
1626 
1627 static void
1628 zio_write_gang_member_ready(zio_t *zio)
1629 {
1630 	zio_t *pio = zio_unique_parent(zio);
1631 	zio_t *gio = zio->io_gang_leader;
1632 	dva_t *cdva = zio->io_bp->blk_dva;
1633 	dva_t *pdva = pio->io_bp->blk_dva;
1634 	uint64_t asize;
1635 
1636 	if (BP_IS_HOLE(zio->io_bp))
1637 		return;
1638 
1639 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1640 
1641 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1642 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1643 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1644 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1645 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1646 
1647 	mutex_enter(&pio->io_lock);
1648 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1649 		ASSERT(DVA_GET_GANG(&pdva[d]));
1650 		asize = DVA_GET_ASIZE(&pdva[d]);
1651 		asize += DVA_GET_ASIZE(&cdva[d]);
1652 		DVA_SET_ASIZE(&pdva[d], asize);
1653 	}
1654 	mutex_exit(&pio->io_lock);
1655 }
1656 
1657 static int
1658 zio_write_gang_block(zio_t *pio)
1659 {
1660 	spa_t *spa = pio->io_spa;
1661 	blkptr_t *bp = pio->io_bp;
1662 	zio_t *gio = pio->io_gang_leader;
1663 	zio_t *zio;
1664 	zio_gang_node_t *gn, **gnpp;
1665 	zio_gbh_phys_t *gbh;
1666 	uint64_t txg = pio->io_txg;
1667 	uint64_t resid = pio->io_size;
1668 	uint64_t lsize;
1669 	int copies = gio->io_prop.zp_copies;
1670 	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1671 	zio_prop_t zp;
1672 	int error;
1673 
1674 	error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1675 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1676 	    METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1677 	if (error) {
1678 		pio->io_error = error;
1679 		return (ZIO_PIPELINE_CONTINUE);
1680 	}
1681 
1682 	if (pio == gio) {
1683 		gnpp = &gio->io_gang_tree;
1684 	} else {
1685 		gnpp = pio->io_private;
1686 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
1687 	}
1688 
1689 	gn = zio_gang_node_alloc(gnpp);
1690 	gbh = gn->gn_gbh;
1691 	bzero(gbh, SPA_GANGBLOCKSIZE);
1692 
1693 	/*
1694 	 * Create the gang header.
1695 	 */
1696 	zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1697 	    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1698 
1699 	/*
1700 	 * Create and nowait the gang children.
1701 	 */
1702 	for (int g = 0; resid != 0; resid -= lsize, g++) {
1703 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1704 		    SPA_MINBLOCKSIZE);
1705 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1706 
1707 		zp.zp_checksum = gio->io_prop.zp_checksum;
1708 		zp.zp_compress = ZIO_COMPRESS_OFF;
1709 		zp.zp_type = DMU_OT_NONE;
1710 		zp.zp_level = 0;
1711 		zp.zp_copies = gio->io_prop.zp_copies;
1712 		zp.zp_dedup = 0;
1713 		zp.zp_dedup_verify = 0;
1714 
1715 		zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1716 		    (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1717 		    zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1718 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1719 		    &pio->io_bookmark));
1720 	}
1721 
1722 	/*
1723 	 * Set pio's pipeline to just wait for zio to finish.
1724 	 */
1725 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1726 
1727 	zio_nowait(zio);
1728 
1729 	return (ZIO_PIPELINE_CONTINUE);
1730 }
1731 
1732 /*
1733  * ==========================================================================
1734  * Dedup
1735  * ==========================================================================
1736  */
1737 static void
1738 zio_ddt_child_read_done(zio_t *zio)
1739 {
1740 	blkptr_t *bp = zio->io_bp;
1741 	ddt_entry_t *dde = zio->io_private;
1742 	ddt_phys_t *ddp;
1743 	zio_t *pio = zio_unique_parent(zio);
1744 
1745 	mutex_enter(&pio->io_lock);
1746 	ddp = ddt_phys_select(dde, bp);
1747 	if (zio->io_error == 0)
1748 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
1749 	if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1750 		dde->dde_repair_data = zio->io_data;
1751 	else
1752 		zio_buf_free(zio->io_data, zio->io_size);
1753 	mutex_exit(&pio->io_lock);
1754 }
1755 
1756 static int
1757 zio_ddt_read_start(zio_t *zio)
1758 {
1759 	blkptr_t *bp = zio->io_bp;
1760 
1761 	ASSERT(BP_GET_DEDUP(bp));
1762 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1763 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1764 
1765 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1766 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1767 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1768 		ddt_phys_t *ddp = dde->dde_phys;
1769 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1770 		blkptr_t blk;
1771 
1772 		ASSERT(zio->io_vsd == NULL);
1773 		zio->io_vsd = dde;
1774 
1775 		if (ddp_self == NULL)
1776 			return (ZIO_PIPELINE_CONTINUE);
1777 
1778 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1779 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1780 				continue;
1781 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1782 			    &blk);
1783 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
1784 			    zio_buf_alloc(zio->io_size), zio->io_size,
1785 			    zio_ddt_child_read_done, dde, zio->io_priority,
1786 			    ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1787 			    &zio->io_bookmark));
1788 		}
1789 		return (ZIO_PIPELINE_CONTINUE);
1790 	}
1791 
1792 	zio_nowait(zio_read(zio, zio->io_spa, bp,
1793 	    zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1794 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1795 
1796 	return (ZIO_PIPELINE_CONTINUE);
1797 }
1798 
1799 static int
1800 zio_ddt_read_done(zio_t *zio)
1801 {
1802 	blkptr_t *bp = zio->io_bp;
1803 
1804 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1805 		return (ZIO_PIPELINE_STOP);
1806 
1807 	ASSERT(BP_GET_DEDUP(bp));
1808 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1809 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1810 
1811 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1812 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1813 		ddt_entry_t *dde = zio->io_vsd;
1814 		if (ddt == NULL) {
1815 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1816 			return (ZIO_PIPELINE_CONTINUE);
1817 		}
1818 		if (dde == NULL) {
1819 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1820 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1821 			return (ZIO_PIPELINE_STOP);
1822 		}
1823 		if (dde->dde_repair_data != NULL) {
1824 			bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1825 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
1826 		}
1827 		ddt_repair_done(ddt, dde);
1828 		zio->io_vsd = NULL;
1829 	}
1830 
1831 	ASSERT(zio->io_vsd == NULL);
1832 
1833 	return (ZIO_PIPELINE_CONTINUE);
1834 }
1835 
1836 static boolean_t
1837 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1838 {
1839 	spa_t *spa = zio->io_spa;
1840 
1841 	/*
1842 	 * Note: we compare the original data, not the transformed data,
1843 	 * because when zio->io_bp is an override bp, we will not have
1844 	 * pushed the I/O transforms.  That's an important optimization
1845 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1846 	 */
1847 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1848 		zio_t *lio = dde->dde_lead_zio[p];
1849 
1850 		if (lio != NULL) {
1851 			return (lio->io_orig_size != zio->io_orig_size ||
1852 			    bcmp(zio->io_orig_data, lio->io_orig_data,
1853 			    zio->io_orig_size) != 0);
1854 		}
1855 	}
1856 
1857 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1858 		ddt_phys_t *ddp = &dde->dde_phys[p];
1859 
1860 		if (ddp->ddp_phys_birth != 0) {
1861 			arc_buf_t *abuf = NULL;
1862 			uint32_t aflags = ARC_WAIT;
1863 			blkptr_t blk = *zio->io_bp;
1864 			int error;
1865 
1866 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1867 
1868 			ddt_exit(ddt);
1869 
1870 			error = arc_read_nolock(NULL, spa, &blk,
1871 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1872 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1873 			    &aflags, &zio->io_bookmark);
1874 
1875 			if (error == 0) {
1876 				if (arc_buf_size(abuf) != zio->io_orig_size ||
1877 				    bcmp(abuf->b_data, zio->io_orig_data,
1878 				    zio->io_orig_size) != 0)
1879 					error = EEXIST;
1880 				VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1881 			}
1882 
1883 			ddt_enter(ddt);
1884 			return (error != 0);
1885 		}
1886 	}
1887 
1888 	return (B_FALSE);
1889 }
1890 
1891 static void
1892 zio_ddt_child_write_ready(zio_t *zio)
1893 {
1894 	int p = zio->io_prop.zp_copies;
1895 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1896 	ddt_entry_t *dde = zio->io_private;
1897 	ddt_phys_t *ddp = &dde->dde_phys[p];
1898 	zio_t *pio;
1899 
1900 	if (zio->io_error)
1901 		return;
1902 
1903 	ddt_enter(ddt);
1904 
1905 	ASSERT(dde->dde_lead_zio[p] == zio);
1906 
1907 	ddt_phys_fill(ddp, zio->io_bp);
1908 
1909 	while ((pio = zio_walk_parents(zio)) != NULL)
1910 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1911 
1912 	ddt_exit(ddt);
1913 }
1914 
1915 static void
1916 zio_ddt_child_write_done(zio_t *zio)
1917 {
1918 	int p = zio->io_prop.zp_copies;
1919 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1920 	ddt_entry_t *dde = zio->io_private;
1921 	ddt_phys_t *ddp = &dde->dde_phys[p];
1922 
1923 	ddt_enter(ddt);
1924 
1925 	ASSERT(ddp->ddp_refcnt == 0);
1926 	ASSERT(dde->dde_lead_zio[p] == zio);
1927 	dde->dde_lead_zio[p] = NULL;
1928 
1929 	if (zio->io_error == 0) {
1930 		while (zio_walk_parents(zio) != NULL)
1931 			ddt_phys_addref(ddp);
1932 	} else {
1933 		ddt_phys_clear(ddp);
1934 	}
1935 
1936 	ddt_exit(ddt);
1937 }
1938 
1939 static void
1940 zio_ddt_ditto_write_done(zio_t *zio)
1941 {
1942 	int p = DDT_PHYS_DITTO;
1943 	zio_prop_t *zp = &zio->io_prop;
1944 	blkptr_t *bp = zio->io_bp;
1945 	ddt_t *ddt = ddt_select(zio->io_spa, bp);
1946 	ddt_entry_t *dde = zio->io_private;
1947 	ddt_phys_t *ddp = &dde->dde_phys[p];
1948 	ddt_key_t *ddk = &dde->dde_key;
1949 
1950 	ddt_enter(ddt);
1951 
1952 	ASSERT(ddp->ddp_refcnt == 0);
1953 	ASSERT(dde->dde_lead_zio[p] == zio);
1954 	dde->dde_lead_zio[p] = NULL;
1955 
1956 	if (zio->io_error == 0) {
1957 		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1958 		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1959 		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1960 		if (ddp->ddp_phys_birth != 0)
1961 			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1962 		ddt_phys_fill(ddp, bp);
1963 	}
1964 
1965 	ddt_exit(ddt);
1966 }
1967 
1968 static int
1969 zio_ddt_write(zio_t *zio)
1970 {
1971 	spa_t *spa = zio->io_spa;
1972 	blkptr_t *bp = zio->io_bp;
1973 	uint64_t txg = zio->io_txg;
1974 	zio_prop_t *zp = &zio->io_prop;
1975 	int p = zp->zp_copies;
1976 	int ditto_copies;
1977 	zio_t *cio = NULL;
1978 	zio_t *dio = NULL;
1979 	ddt_t *ddt = ddt_select(spa, bp);
1980 	ddt_entry_t *dde;
1981 	ddt_phys_t *ddp;
1982 
1983 	ASSERT(BP_GET_DEDUP(bp));
1984 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
1985 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
1986 
1987 	ddt_enter(ddt);
1988 	dde = ddt_lookup(ddt, bp, B_TRUE);
1989 	ddp = &dde->dde_phys[p];
1990 
1991 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
1992 		/*
1993 		 * If we're using a weak checksum, upgrade to a strong checksum
1994 		 * and try again.  If we're already using a strong checksum,
1995 		 * we can't resolve it, so just convert to an ordinary write.
1996 		 * (And automatically e-mail a paper to Nature?)
1997 		 */
1998 		if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
1999 			zp->zp_checksum = spa_dedup_checksum(spa);
2000 			zio_pop_transforms(zio);
2001 			zio->io_stage = ZIO_STAGE_OPEN;
2002 			BP_ZERO(bp);
2003 		} else {
2004 			zp->zp_dedup = 0;
2005 		}
2006 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2007 		ddt_exit(ddt);
2008 		return (ZIO_PIPELINE_CONTINUE);
2009 	}
2010 
2011 	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2012 	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2013 
2014 	if (ditto_copies > ddt_ditto_copies_present(dde) &&
2015 	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2016 		zio_prop_t czp = *zp;
2017 
2018 		czp.zp_copies = ditto_copies;
2019 
2020 		/*
2021 		 * If we arrived here with an override bp, we won't have run
2022 		 * the transform stack, so we won't have the data we need to
2023 		 * generate a child i/o.  So, toss the override bp and restart.
2024 		 * This is safe, because using the override bp is just an
2025 		 * optimization; and it's rare, so the cost doesn't matter.
2026 		 */
2027 		if (zio->io_bp_override) {
2028 			zio_pop_transforms(zio);
2029 			zio->io_stage = ZIO_STAGE_OPEN;
2030 			zio->io_pipeline = ZIO_WRITE_PIPELINE;
2031 			zio->io_bp_override = NULL;
2032 			BP_ZERO(bp);
2033 			ddt_exit(ddt);
2034 			return (ZIO_PIPELINE_CONTINUE);
2035 		}
2036 
2037 		dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2038 		    zio->io_orig_size, &czp, NULL,
2039 		    zio_ddt_ditto_write_done, dde, zio->io_priority,
2040 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2041 
2042 		zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2043 		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2044 	}
2045 
2046 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2047 		if (ddp->ddp_phys_birth != 0)
2048 			ddt_bp_fill(ddp, bp, txg);
2049 		if (dde->dde_lead_zio[p] != NULL)
2050 			zio_add_child(zio, dde->dde_lead_zio[p]);
2051 		else
2052 			ddt_phys_addref(ddp);
2053 	} else if (zio->io_bp_override) {
2054 		ASSERT(bp->blk_birth == txg);
2055 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2056 		ddt_phys_fill(ddp, bp);
2057 		ddt_phys_addref(ddp);
2058 	} else {
2059 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2060 		    zio->io_orig_size, zp, zio_ddt_child_write_ready,
2061 		    zio_ddt_child_write_done, dde, zio->io_priority,
2062 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2063 
2064 		zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2065 		dde->dde_lead_zio[p] = cio;
2066 	}
2067 
2068 	ddt_exit(ddt);
2069 
2070 	if (cio)
2071 		zio_nowait(cio);
2072 	if (dio)
2073 		zio_nowait(dio);
2074 
2075 	return (ZIO_PIPELINE_CONTINUE);
2076 }
2077 
2078 ddt_entry_t *freedde; /* for debugging */
2079 
2080 static int
2081 zio_ddt_free(zio_t *zio)
2082 {
2083 	spa_t *spa = zio->io_spa;
2084 	blkptr_t *bp = zio->io_bp;
2085 	ddt_t *ddt = ddt_select(spa, bp);
2086 	ddt_entry_t *dde;
2087 	ddt_phys_t *ddp;
2088 
2089 	ASSERT(BP_GET_DEDUP(bp));
2090 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2091 
2092 	ddt_enter(ddt);
2093 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2094 	ddp = ddt_phys_select(dde, bp);
2095 	ddt_phys_decref(ddp);
2096 	ddt_exit(ddt);
2097 
2098 	return (ZIO_PIPELINE_CONTINUE);
2099 }
2100 
2101 /*
2102  * ==========================================================================
2103  * Allocate and free blocks
2104  * ==========================================================================
2105  */
2106 static int
2107 zio_dva_allocate(zio_t *zio)
2108 {
2109 	spa_t *spa = zio->io_spa;
2110 	metaslab_class_t *mc = spa_normal_class(spa);
2111 	blkptr_t *bp = zio->io_bp;
2112 	int error;
2113 
2114 	if (zio->io_gang_leader == NULL) {
2115 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2116 		zio->io_gang_leader = zio;
2117 	}
2118 
2119 	ASSERT(BP_IS_HOLE(bp));
2120 	ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2121 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
2122 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2123 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2124 
2125 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
2126 	    zio->io_prop.zp_copies, zio->io_txg, NULL, 0);
2127 
2128 	if (error) {
2129 		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2130 			return (zio_write_gang_block(zio));
2131 		zio->io_error = error;
2132 	}
2133 
2134 	return (ZIO_PIPELINE_CONTINUE);
2135 }
2136 
2137 static int
2138 zio_dva_free(zio_t *zio)
2139 {
2140 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2141 
2142 	return (ZIO_PIPELINE_CONTINUE);
2143 }
2144 
2145 static int
2146 zio_dva_claim(zio_t *zio)
2147 {
2148 	int error;
2149 
2150 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2151 	if (error)
2152 		zio->io_error = error;
2153 
2154 	return (ZIO_PIPELINE_CONTINUE);
2155 }
2156 
2157 /*
2158  * Undo an allocation.  This is used by zio_done() when an I/O fails
2159  * and we want to give back the block we just allocated.
2160  * This handles both normal blocks and gang blocks.
2161  */
2162 static void
2163 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2164 {
2165 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2166 	ASSERT(zio->io_bp_override == NULL);
2167 
2168 	if (!BP_IS_HOLE(bp))
2169 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2170 
2171 	if (gn != NULL) {
2172 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2173 			zio_dva_unallocate(zio, gn->gn_child[g],
2174 			    &gn->gn_gbh->zg_blkptr[g]);
2175 		}
2176 	}
2177 }
2178 
2179 /*
2180  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
2181  */
2182 int
2183 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2184     uint64_t size, boolean_t use_slog)
2185 {
2186 	int error = 1;
2187 
2188 	ASSERT(txg > spa_syncing_txg(spa));
2189 
2190 	if (use_slog)
2191 		error = metaslab_alloc(spa, spa_log_class(spa), size,
2192 		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2193 
2194 	if (error)
2195 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
2196 		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2197 
2198 	if (error == 0) {
2199 		BP_SET_LSIZE(new_bp, size);
2200 		BP_SET_PSIZE(new_bp, size);
2201 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2202 		BP_SET_CHECKSUM(new_bp,
2203 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2204 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2205 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2206 		BP_SET_LEVEL(new_bp, 0);
2207 		BP_SET_DEDUP(new_bp, 0);
2208 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2209 	}
2210 
2211 	return (error);
2212 }
2213 
2214 /*
2215  * Free an intent log block.
2216  */
2217 void
2218 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2219 {
2220 	ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2221 	ASSERT(!BP_IS_GANG(bp));
2222 
2223 	zio_free(spa, txg, bp);
2224 }
2225 
2226 /*
2227  * ==========================================================================
2228  * Read and write to physical devices
2229  * ==========================================================================
2230  */
2231 static int
2232 zio_vdev_io_start(zio_t *zio)
2233 {
2234 	vdev_t *vd = zio->io_vd;
2235 	uint64_t align;
2236 	spa_t *spa = zio->io_spa;
2237 
2238 	ASSERT(zio->io_error == 0);
2239 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2240 
2241 	if (vd == NULL) {
2242 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2243 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2244 
2245 		/*
2246 		 * The mirror_ops handle multiple DVAs in a single BP.
2247 		 */
2248 		return (vdev_mirror_ops.vdev_op_io_start(zio));
2249 	}
2250 
2251 	align = 1ULL << vd->vdev_top->vdev_ashift;
2252 
2253 	if (P2PHASE(zio->io_size, align) != 0) {
2254 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
2255 		char *abuf = zio_buf_alloc(asize);
2256 		ASSERT(vd == vd->vdev_top);
2257 		if (zio->io_type == ZIO_TYPE_WRITE) {
2258 			bcopy(zio->io_data, abuf, zio->io_size);
2259 			bzero(abuf + zio->io_size, asize - zio->io_size);
2260 		}
2261 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2262 	}
2263 
2264 	ASSERT(P2PHASE(zio->io_offset, align) == 0);
2265 	ASSERT(P2PHASE(zio->io_size, align) == 0);
2266 	ASSERT(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2267 
2268 	/*
2269 	 * If this is a repair I/O, and there's no self-healing involved --
2270 	 * that is, we're just resilvering what we expect to resilver --
2271 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2272 	 * This prevents spurious resilvering with nested replication.
2273 	 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2274 	 * A is out of date, we'll read from C+D, then use the data to
2275 	 * resilver A+B -- but we don't actually want to resilver B, just A.
2276 	 * The top-level mirror has no way to know this, so instead we just
2277 	 * discard unnecessary repairs as we work our way down the vdev tree.
2278 	 * The same logic applies to any form of nested replication:
2279 	 * ditto + mirror, RAID-Z + replacing, etc.  This covers them all.
2280 	 */
2281 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2282 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2283 	    zio->io_txg != 0 &&	/* not a delegated i/o */
2284 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2285 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2286 		zio_vdev_io_bypass(zio);
2287 		return (ZIO_PIPELINE_CONTINUE);
2288 	}
2289 
2290 	if (vd->vdev_ops->vdev_op_leaf &&
2291 	    (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2292 
2293 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2294 			return (ZIO_PIPELINE_CONTINUE);
2295 
2296 		if ((zio = vdev_queue_io(zio)) == NULL)
2297 			return (ZIO_PIPELINE_STOP);
2298 
2299 		if (!vdev_accessible(vd, zio)) {
2300 			zio->io_error = ENXIO;
2301 			zio_interrupt(zio);
2302 			return (ZIO_PIPELINE_STOP);
2303 		}
2304 	}
2305 
2306 	return (vd->vdev_ops->vdev_op_io_start(zio));
2307 }
2308 
2309 static int
2310 zio_vdev_io_done(zio_t *zio)
2311 {
2312 	vdev_t *vd = zio->io_vd;
2313 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2314 	boolean_t unexpected_error = B_FALSE;
2315 
2316 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2317 		return (ZIO_PIPELINE_STOP);
2318 
2319 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2320 
2321 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2322 
2323 		vdev_queue_io_done(zio);
2324 
2325 		if (zio->io_type == ZIO_TYPE_WRITE)
2326 			vdev_cache_write(zio);
2327 
2328 		if (zio_injection_enabled && zio->io_error == 0)
2329 			zio->io_error = zio_handle_device_injection(vd,
2330 			    zio, EIO);
2331 
2332 		if (zio_injection_enabled && zio->io_error == 0)
2333 			zio->io_error = zio_handle_label_injection(zio, EIO);
2334 
2335 		if (zio->io_error) {
2336 			if (!vdev_accessible(vd, zio)) {
2337 				zio->io_error = ENXIO;
2338 			} else {
2339 				unexpected_error = B_TRUE;
2340 			}
2341 		}
2342 	}
2343 
2344 	ops->vdev_op_io_done(zio);
2345 
2346 	if (unexpected_error)
2347 		VERIFY(vdev_probe(vd, zio) == NULL);
2348 
2349 	return (ZIO_PIPELINE_CONTINUE);
2350 }
2351 
2352 /*
2353  * For non-raidz ZIOs, we can just copy aside the bad data read from the
2354  * disk, and use that to finish the checksum ereport later.
2355  */
2356 static void
2357 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2358     const void *good_buf)
2359 {
2360 	/* no processing needed */
2361 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2362 }
2363 
2364 /*ARGSUSED*/
2365 void
2366 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2367 {
2368 	void *buf = zio_buf_alloc(zio->io_size);
2369 
2370 	bcopy(zio->io_data, buf, zio->io_size);
2371 
2372 	zcr->zcr_cbinfo = zio->io_size;
2373 	zcr->zcr_cbdata = buf;
2374 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
2375 	zcr->zcr_free = zio_buf_free;
2376 }
2377 
2378 static int
2379 zio_vdev_io_assess(zio_t *zio)
2380 {
2381 	vdev_t *vd = zio->io_vd;
2382 
2383 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2384 		return (ZIO_PIPELINE_STOP);
2385 
2386 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2387 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2388 
2389 	if (zio->io_vsd != NULL) {
2390 		zio->io_vsd_ops->vsd_free(zio);
2391 		zio->io_vsd = NULL;
2392 	}
2393 
2394 	if (zio_injection_enabled && zio->io_error == 0)
2395 		zio->io_error = zio_handle_fault_injection(zio, EIO);
2396 
2397 	/*
2398 	 * If the I/O failed, determine whether we should attempt to retry it.
2399 	 *
2400 	 * On retry, we cut in line in the issue queue, since we don't want
2401 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2402 	 */
2403 	if (zio->io_error && vd == NULL &&
2404 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2405 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
2406 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
2407 		zio->io_error = 0;
2408 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
2409 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2410 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2411 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2412 		    zio_requeue_io_start_cut_in_line);
2413 		return (ZIO_PIPELINE_STOP);
2414 	}
2415 
2416 	/*
2417 	 * If we got an error on a leaf device, convert it to ENXIO
2418 	 * if the device is not accessible at all.
2419 	 */
2420 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2421 	    !vdev_accessible(vd, zio))
2422 		zio->io_error = ENXIO;
2423 
2424 	/*
2425 	 * If we can't write to an interior vdev (mirror or RAID-Z),
2426 	 * set vdev_cant_write so that we stop trying to allocate from it.
2427 	 */
2428 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2429 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2430 		vd->vdev_cant_write = B_TRUE;
2431 
2432 	if (zio->io_error)
2433 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2434 
2435 	return (ZIO_PIPELINE_CONTINUE);
2436 }
2437 
2438 void
2439 zio_vdev_io_reissue(zio_t *zio)
2440 {
2441 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2442 	ASSERT(zio->io_error == 0);
2443 
2444 	zio->io_stage >>= 1;
2445 }
2446 
2447 void
2448 zio_vdev_io_redone(zio_t *zio)
2449 {
2450 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2451 
2452 	zio->io_stage >>= 1;
2453 }
2454 
2455 void
2456 zio_vdev_io_bypass(zio_t *zio)
2457 {
2458 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2459 	ASSERT(zio->io_error == 0);
2460 
2461 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2462 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2463 }
2464 
2465 /*
2466  * ==========================================================================
2467  * Generate and verify checksums
2468  * ==========================================================================
2469  */
2470 static int
2471 zio_checksum_generate(zio_t *zio)
2472 {
2473 	blkptr_t *bp = zio->io_bp;
2474 	enum zio_checksum checksum;
2475 
2476 	if (bp == NULL) {
2477 		/*
2478 		 * This is zio_write_phys().
2479 		 * We're either generating a label checksum, or none at all.
2480 		 */
2481 		checksum = zio->io_prop.zp_checksum;
2482 
2483 		if (checksum == ZIO_CHECKSUM_OFF)
2484 			return (ZIO_PIPELINE_CONTINUE);
2485 
2486 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2487 	} else {
2488 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2489 			ASSERT(!IO_IS_ALLOCATING(zio));
2490 			checksum = ZIO_CHECKSUM_GANG_HEADER;
2491 		} else {
2492 			checksum = BP_GET_CHECKSUM(bp);
2493 		}
2494 	}
2495 
2496 	zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2497 
2498 	return (ZIO_PIPELINE_CONTINUE);
2499 }
2500 
2501 static int
2502 zio_checksum_verify(zio_t *zio)
2503 {
2504 	zio_bad_cksum_t info;
2505 	blkptr_t *bp = zio->io_bp;
2506 	int error;
2507 
2508 	ASSERT(zio->io_vd != NULL);
2509 
2510 	if (bp == NULL) {
2511 		/*
2512 		 * This is zio_read_phys().
2513 		 * We're either verifying a label checksum, or nothing at all.
2514 		 */
2515 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2516 			return (ZIO_PIPELINE_CONTINUE);
2517 
2518 		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2519 	}
2520 
2521 	if ((error = zio_checksum_error(zio, &info)) != 0) {
2522 		zio->io_error = error;
2523 		if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2524 			zfs_ereport_start_checksum(zio->io_spa,
2525 			    zio->io_vd, zio, zio->io_offset,
2526 			    zio->io_size, NULL, &info);
2527 		}
2528 	}
2529 
2530 	return (ZIO_PIPELINE_CONTINUE);
2531 }
2532 
2533 /*
2534  * Called by RAID-Z to ensure we don't compute the checksum twice.
2535  */
2536 void
2537 zio_checksum_verified(zio_t *zio)
2538 {
2539 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2540 }
2541 
2542 /*
2543  * ==========================================================================
2544  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2545  * An error of 0 indictes success.  ENXIO indicates whole-device failure,
2546  * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
2547  * indicate errors that are specific to one I/O, and most likely permanent.
2548  * Any other error is presumed to be worse because we weren't expecting it.
2549  * ==========================================================================
2550  */
2551 int
2552 zio_worst_error(int e1, int e2)
2553 {
2554 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2555 	int r1, r2;
2556 
2557 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2558 		if (e1 == zio_error_rank[r1])
2559 			break;
2560 
2561 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2562 		if (e2 == zio_error_rank[r2])
2563 			break;
2564 
2565 	return (r1 > r2 ? e1 : e2);
2566 }
2567 
2568 /*
2569  * ==========================================================================
2570  * I/O completion
2571  * ==========================================================================
2572  */
2573 static int
2574 zio_ready(zio_t *zio)
2575 {
2576 	blkptr_t *bp = zio->io_bp;
2577 	zio_t *pio, *pio_next;
2578 
2579 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2580 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2581 		return (ZIO_PIPELINE_STOP);
2582 
2583 	if (zio->io_ready) {
2584 		ASSERT(IO_IS_ALLOCATING(zio));
2585 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2586 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2587 
2588 		zio->io_ready(zio);
2589 	}
2590 
2591 	if (bp != NULL && bp != &zio->io_bp_copy)
2592 		zio->io_bp_copy = *bp;
2593 
2594 	if (zio->io_error)
2595 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2596 
2597 	mutex_enter(&zio->io_lock);
2598 	zio->io_state[ZIO_WAIT_READY] = 1;
2599 	pio = zio_walk_parents(zio);
2600 	mutex_exit(&zio->io_lock);
2601 
2602 	/*
2603 	 * As we notify zio's parents, new parents could be added.
2604 	 * New parents go to the head of zio's io_parent_list, however,
2605 	 * so we will (correctly) not notify them.  The remainder of zio's
2606 	 * io_parent_list, from 'pio_next' onward, cannot change because
2607 	 * all parents must wait for us to be done before they can be done.
2608 	 */
2609 	for (; pio != NULL; pio = pio_next) {
2610 		pio_next = zio_walk_parents(zio);
2611 		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2612 	}
2613 
2614 	if (zio->io_flags & ZIO_FLAG_NODATA) {
2615 		if (BP_IS_GANG(bp)) {
2616 			zio->io_flags &= ~ZIO_FLAG_NODATA;
2617 		} else {
2618 			ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2619 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2620 		}
2621 	}
2622 
2623 	if (zio_injection_enabled &&
2624 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
2625 		zio_handle_ignored_writes(zio);
2626 
2627 	return (ZIO_PIPELINE_CONTINUE);
2628 }
2629 
2630 static int
2631 zio_done(zio_t *zio)
2632 {
2633 	spa_t *spa = zio->io_spa;
2634 	zio_t *lio = zio->io_logical;
2635 	blkptr_t *bp = zio->io_bp;
2636 	vdev_t *vd = zio->io_vd;
2637 	uint64_t psize = zio->io_size;
2638 	zio_t *pio, *pio_next;
2639 
2640 	/*
2641 	 * If our children haven't all completed,
2642 	 * wait for them and then repeat this pipeline stage.
2643 	 */
2644 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2645 	    zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2646 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2647 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2648 		return (ZIO_PIPELINE_STOP);
2649 
2650 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2651 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2652 			ASSERT(zio->io_children[c][w] == 0);
2653 
2654 	if (bp != NULL) {
2655 		ASSERT(bp->blk_pad[0] == 0);
2656 		ASSERT(bp->blk_pad[1] == 0);
2657 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2658 		    (bp == zio_unique_parent(zio)->io_bp));
2659 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2660 		    zio->io_bp_override == NULL &&
2661 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2662 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
2663 			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2664 			ASSERT(BP_COUNT_GANG(bp) == 0 ||
2665 			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2666 		}
2667 	}
2668 
2669 	/*
2670 	 * If there were child vdev/gang/ddt errors, they apply to us now.
2671 	 */
2672 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2673 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2674 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2675 
2676 	/*
2677 	 * If the I/O on the transformed data was successful, generate any
2678 	 * checksum reports now while we still have the transformed data.
2679 	 */
2680 	if (zio->io_error == 0) {
2681 		while (zio->io_cksum_report != NULL) {
2682 			zio_cksum_report_t *zcr = zio->io_cksum_report;
2683 			uint64_t align = zcr->zcr_align;
2684 			uint64_t asize = P2ROUNDUP(psize, align);
2685 			char *abuf = zio->io_data;
2686 
2687 			if (asize != psize) {
2688 				abuf = zio_buf_alloc(asize);
2689 				bcopy(zio->io_data, abuf, psize);
2690 				bzero(abuf + psize, asize - psize);
2691 			}
2692 
2693 			zio->io_cksum_report = zcr->zcr_next;
2694 			zcr->zcr_next = NULL;
2695 			zcr->zcr_finish(zcr, abuf);
2696 			zfs_ereport_free_checksum(zcr);
2697 
2698 			if (asize != psize)
2699 				zio_buf_free(abuf, asize);
2700 		}
2701 	}
2702 
2703 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
2704 
2705 	vdev_stat_update(zio, psize);
2706 
2707 	if (zio->io_error) {
2708 		/*
2709 		 * If this I/O is attached to a particular vdev,
2710 		 * generate an error message describing the I/O failure
2711 		 * at the block level.  We ignore these errors if the
2712 		 * device is currently unavailable.
2713 		 */
2714 		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2715 			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2716 
2717 		if ((zio->io_error == EIO || !(zio->io_flags &
2718 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2719 		    zio == lio) {
2720 			/*
2721 			 * For logical I/O requests, tell the SPA to log the
2722 			 * error and generate a logical data ereport.
2723 			 */
2724 			spa_log_error(spa, zio);
2725 			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2726 			    0, 0);
2727 		}
2728 	}
2729 
2730 	if (zio->io_error && zio == lio) {
2731 		/*
2732 		 * Determine whether zio should be reexecuted.  This will
2733 		 * propagate all the way to the root via zio_notify_parent().
2734 		 */
2735 		ASSERT(vd == NULL && bp != NULL);
2736 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2737 
2738 		if (IO_IS_ALLOCATING(zio) &&
2739 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2740 			if (zio->io_error != ENOSPC)
2741 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2742 			else
2743 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2744 		}
2745 
2746 		if ((zio->io_type == ZIO_TYPE_READ ||
2747 		    zio->io_type == ZIO_TYPE_FREE) &&
2748 		    zio->io_error == ENXIO &&
2749 		    spa_load_state(spa) == SPA_LOAD_NONE &&
2750 		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2751 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2752 
2753 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2754 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2755 
2756 		/*
2757 		 * Here is a possibly good place to attempt to do
2758 		 * either combinatorial reconstruction or error correction
2759 		 * based on checksums.  It also might be a good place
2760 		 * to send out preliminary ereports before we suspend
2761 		 * processing.
2762 		 */
2763 	}
2764 
2765 	/*
2766 	 * If there were logical child errors, they apply to us now.
2767 	 * We defer this until now to avoid conflating logical child
2768 	 * errors with errors that happened to the zio itself when
2769 	 * updating vdev stats and reporting FMA events above.
2770 	 */
2771 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2772 
2773 	if ((zio->io_error || zio->io_reexecute) &&
2774 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2775 	    !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2776 		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2777 
2778 	zio_gang_tree_free(&zio->io_gang_tree);
2779 
2780 	/*
2781 	 * Godfather I/Os should never suspend.
2782 	 */
2783 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2784 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2785 		zio->io_reexecute = 0;
2786 
2787 	if (zio->io_reexecute) {
2788 		/*
2789 		 * This is a logical I/O that wants to reexecute.
2790 		 *
2791 		 * Reexecute is top-down.  When an i/o fails, if it's not
2792 		 * the root, it simply notifies its parent and sticks around.
2793 		 * The parent, seeing that it still has children in zio_done(),
2794 		 * does the same.  This percolates all the way up to the root.
2795 		 * The root i/o will reexecute or suspend the entire tree.
2796 		 *
2797 		 * This approach ensures that zio_reexecute() honors
2798 		 * all the original i/o dependency relationships, e.g.
2799 		 * parents not executing until children are ready.
2800 		 */
2801 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2802 
2803 		zio->io_gang_leader = NULL;
2804 
2805 		mutex_enter(&zio->io_lock);
2806 		zio->io_state[ZIO_WAIT_DONE] = 1;
2807 		mutex_exit(&zio->io_lock);
2808 
2809 		/*
2810 		 * "The Godfather" I/O monitors its children but is
2811 		 * not a true parent to them. It will track them through
2812 		 * the pipeline but severs its ties whenever they get into
2813 		 * trouble (e.g. suspended). This allows "The Godfather"
2814 		 * I/O to return status without blocking.
2815 		 */
2816 		for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2817 			zio_link_t *zl = zio->io_walk_link;
2818 			pio_next = zio_walk_parents(zio);
2819 
2820 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2821 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2822 				zio_remove_child(pio, zio, zl);
2823 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2824 			}
2825 		}
2826 
2827 		if ((pio = zio_unique_parent(zio)) != NULL) {
2828 			/*
2829 			 * We're not a root i/o, so there's nothing to do
2830 			 * but notify our parent.  Don't propagate errors
2831 			 * upward since we haven't permanently failed yet.
2832 			 */
2833 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2834 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2835 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2836 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2837 			/*
2838 			 * We'd fail again if we reexecuted now, so suspend
2839 			 * until conditions improve (e.g. device comes online).
2840 			 */
2841 			zio_suspend(spa, zio);
2842 		} else {
2843 			/*
2844 			 * Reexecution is potentially a huge amount of work.
2845 			 * Hand it off to the otherwise-unused claim taskq.
2846 			 */
2847 			(void) taskq_dispatch(
2848 			    spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2849 			    (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2850 		}
2851 		return (ZIO_PIPELINE_STOP);
2852 	}
2853 
2854 	ASSERT(zio->io_child_count == 0);
2855 	ASSERT(zio->io_reexecute == 0);
2856 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2857 
2858 	/*
2859 	 * Report any checksum errors, since the I/O is complete.
2860 	 */
2861 	while (zio->io_cksum_report != NULL) {
2862 		zio_cksum_report_t *zcr = zio->io_cksum_report;
2863 		zio->io_cksum_report = zcr->zcr_next;
2864 		zcr->zcr_next = NULL;
2865 		zcr->zcr_finish(zcr, NULL);
2866 		zfs_ereport_free_checksum(zcr);
2867 	}
2868 
2869 	/*
2870 	 * It is the responsibility of the done callback to ensure that this
2871 	 * particular zio is no longer discoverable for adoption, and as
2872 	 * such, cannot acquire any new parents.
2873 	 */
2874 	if (zio->io_done)
2875 		zio->io_done(zio);
2876 
2877 	mutex_enter(&zio->io_lock);
2878 	zio->io_state[ZIO_WAIT_DONE] = 1;
2879 	mutex_exit(&zio->io_lock);
2880 
2881 	for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2882 		zio_link_t *zl = zio->io_walk_link;
2883 		pio_next = zio_walk_parents(zio);
2884 		zio_remove_child(pio, zio, zl);
2885 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2886 	}
2887 
2888 	if (zio->io_waiter != NULL) {
2889 		mutex_enter(&zio->io_lock);
2890 		zio->io_executor = NULL;
2891 		cv_broadcast(&zio->io_cv);
2892 		mutex_exit(&zio->io_lock);
2893 	} else {
2894 		zio_destroy(zio);
2895 	}
2896 
2897 	return (ZIO_PIPELINE_STOP);
2898 }
2899 
2900 /*
2901  * ==========================================================================
2902  * I/O pipeline definition
2903  * ==========================================================================
2904  */
2905 static zio_pipe_stage_t *zio_pipeline[] = {
2906 	NULL,
2907 	zio_read_bp_init,
2908 	zio_free_bp_init,
2909 	zio_issue_async,
2910 	zio_write_bp_init,
2911 	zio_checksum_generate,
2912 	zio_ddt_read_start,
2913 	zio_ddt_read_done,
2914 	zio_ddt_write,
2915 	zio_ddt_free,
2916 	zio_gang_assemble,
2917 	zio_gang_issue,
2918 	zio_dva_allocate,
2919 	zio_dva_free,
2920 	zio_dva_claim,
2921 	zio_ready,
2922 	zio_vdev_io_start,
2923 	zio_vdev_io_done,
2924 	zio_vdev_io_assess,
2925 	zio_checksum_verify,
2926 	zio_done
2927 };
2928