xref: /illumos-gate/usr/src/uts/common/fs/zfs/zio.c (revision 83eaeac78ef2f69de16c2fecd3077c0ee9269743)
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 	2,	/* ZIO_PRIORITY_DDT_PREFETCH	*/
56 };
57 
58 /*
59  * ==========================================================================
60  * I/O type descriptions
61  * ==========================================================================
62  */
63 char *zio_type_name[ZIO_TYPES] = {
64 	"zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
65 	"zio_ioctl"
66 };
67 
68 /*
69  * ==========================================================================
70  * I/O kmem caches
71  * ==========================================================================
72  */
73 kmem_cache_t *zio_cache;
74 kmem_cache_t *zio_link_cache;
75 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
76 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
77 
78 #ifdef _KERNEL
79 extern vmem_t *zio_alloc_arena;
80 #endif
81 
82 /*
83  * An allocating zio is one that either currently has the DVA allocate
84  * stage set or will have it later in its lifetime.
85  */
86 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
87 
88 boolean_t	zio_requeue_io_start_cut_in_line = B_TRUE;
89 
90 #ifdef ZFS_DEBUG
91 int zio_buf_debug_limit = 16384;
92 #else
93 int zio_buf_debug_limit = 0;
94 #endif
95 
96 void
97 zio_init(void)
98 {
99 	size_t c;
100 	vmem_t *data_alloc_arena = NULL;
101 
102 #ifdef _KERNEL
103 	data_alloc_arena = zio_alloc_arena;
104 #endif
105 	zio_cache = kmem_cache_create("zio_cache",
106 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
107 	zio_link_cache = kmem_cache_create("zio_link_cache",
108 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
109 
110 	/*
111 	 * For small buffers, we want a cache for each multiple of
112 	 * SPA_MINBLOCKSIZE.  For medium-size buffers, we want a cache
113 	 * for each quarter-power of 2.  For large buffers, we want
114 	 * a cache for each multiple of PAGESIZE.
115 	 */
116 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
117 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
118 		size_t p2 = size;
119 		size_t align = 0;
120 		size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
121 
122 		while (p2 & (p2 - 1))
123 			p2 &= p2 - 1;
124 
125 		if (size <= 4 * SPA_MINBLOCKSIZE) {
126 			align = SPA_MINBLOCKSIZE;
127 		} else if (P2PHASE(size, PAGESIZE) == 0) {
128 			align = PAGESIZE;
129 		} else if (P2PHASE(size, p2 >> 2) == 0) {
130 			align = p2 >> 2;
131 		}
132 
133 		if (align != 0) {
134 			char name[36];
135 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
136 			zio_buf_cache[c] = kmem_cache_create(name, size,
137 			    align, NULL, NULL, NULL, NULL, NULL, cflags);
138 
139 			/*
140 			 * Since zio_data bufs do not appear in crash dumps, we
141 			 * pass KMC_NOTOUCH so that no allocator metadata is
142 			 * stored with the buffers.
143 			 */
144 			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
145 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
146 			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
147 			    cflags | KMC_NOTOUCH);
148 		}
149 	}
150 
151 	while (--c != 0) {
152 		ASSERT(zio_buf_cache[c] != NULL);
153 		if (zio_buf_cache[c - 1] == NULL)
154 			zio_buf_cache[c - 1] = zio_buf_cache[c];
155 
156 		ASSERT(zio_data_buf_cache[c] != NULL);
157 		if (zio_data_buf_cache[c - 1] == NULL)
158 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
159 	}
160 
161 	zio_inject_init();
162 }
163 
164 void
165 zio_fini(void)
166 {
167 	size_t c;
168 	kmem_cache_t *last_cache = NULL;
169 	kmem_cache_t *last_data_cache = NULL;
170 
171 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
172 		if (zio_buf_cache[c] != last_cache) {
173 			last_cache = zio_buf_cache[c];
174 			kmem_cache_destroy(zio_buf_cache[c]);
175 		}
176 		zio_buf_cache[c] = NULL;
177 
178 		if (zio_data_buf_cache[c] != last_data_cache) {
179 			last_data_cache = zio_data_buf_cache[c];
180 			kmem_cache_destroy(zio_data_buf_cache[c]);
181 		}
182 		zio_data_buf_cache[c] = NULL;
183 	}
184 
185 	kmem_cache_destroy(zio_link_cache);
186 	kmem_cache_destroy(zio_cache);
187 
188 	zio_inject_fini();
189 }
190 
191 /*
192  * ==========================================================================
193  * Allocate and free I/O buffers
194  * ==========================================================================
195  */
196 
197 /*
198  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
199  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
200  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
201  * excess / transient data in-core during a crashdump.
202  */
203 void *
204 zio_buf_alloc(size_t size)
205 {
206 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
207 
208 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
209 
210 	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
211 }
212 
213 /*
214  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
215  * crashdump if the kernel panics.  This exists so that we will limit the amount
216  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
217  * of kernel heap dumped to disk when the kernel panics)
218  */
219 void *
220 zio_data_buf_alloc(size_t size)
221 {
222 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
223 
224 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
225 
226 	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
227 }
228 
229 void
230 zio_buf_free(void *buf, size_t size)
231 {
232 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
233 
234 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
235 
236 	kmem_cache_free(zio_buf_cache[c], buf);
237 }
238 
239 void
240 zio_data_buf_free(void *buf, size_t size)
241 {
242 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
243 
244 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
245 
246 	kmem_cache_free(zio_data_buf_cache[c], buf);
247 }
248 
249 /*
250  * ==========================================================================
251  * Push and pop I/O transform buffers
252  * ==========================================================================
253  */
254 static void
255 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
256 	zio_transform_func_t *transform)
257 {
258 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
259 
260 	zt->zt_orig_data = zio->io_data;
261 	zt->zt_orig_size = zio->io_size;
262 	zt->zt_bufsize = bufsize;
263 	zt->zt_transform = transform;
264 
265 	zt->zt_next = zio->io_transform_stack;
266 	zio->io_transform_stack = zt;
267 
268 	zio->io_data = data;
269 	zio->io_size = size;
270 }
271 
272 static void
273 zio_pop_transforms(zio_t *zio)
274 {
275 	zio_transform_t *zt;
276 
277 	while ((zt = zio->io_transform_stack) != NULL) {
278 		if (zt->zt_transform != NULL)
279 			zt->zt_transform(zio,
280 			    zt->zt_orig_data, zt->zt_orig_size);
281 
282 		if (zt->zt_bufsize != 0)
283 			zio_buf_free(zio->io_data, zt->zt_bufsize);
284 
285 		zio->io_data = zt->zt_orig_data;
286 		zio->io_size = zt->zt_orig_size;
287 		zio->io_transform_stack = zt->zt_next;
288 
289 		kmem_free(zt, sizeof (zio_transform_t));
290 	}
291 }
292 
293 /*
294  * ==========================================================================
295  * I/O transform callbacks for subblocks and decompression
296  * ==========================================================================
297  */
298 static void
299 zio_subblock(zio_t *zio, void *data, uint64_t size)
300 {
301 	ASSERT(zio->io_size > size);
302 
303 	if (zio->io_type == ZIO_TYPE_READ)
304 		bcopy(zio->io_data, data, size);
305 }
306 
307 static void
308 zio_decompress(zio_t *zio, void *data, uint64_t size)
309 {
310 	if (zio->io_error == 0 &&
311 	    zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
312 	    zio->io_data, data, zio->io_size, size) != 0)
313 		zio->io_error = EIO;
314 }
315 
316 /*
317  * ==========================================================================
318  * I/O parent/child relationships and pipeline interlocks
319  * ==========================================================================
320  */
321 /*
322  * NOTE - Callers to zio_walk_parents() and zio_walk_children must
323  *        continue calling these functions until they return NULL.
324  *        Otherwise, the next caller will pick up the list walk in
325  *        some indeterminate state.  (Otherwise every caller would
326  *        have to pass in a cookie to keep the state represented by
327  *        io_walk_link, which gets annoying.)
328  */
329 zio_t *
330 zio_walk_parents(zio_t *cio)
331 {
332 	zio_link_t *zl = cio->io_walk_link;
333 	list_t *pl = &cio->io_parent_list;
334 
335 	zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
336 	cio->io_walk_link = zl;
337 
338 	if (zl == NULL)
339 		return (NULL);
340 
341 	ASSERT(zl->zl_child == cio);
342 	return (zl->zl_parent);
343 }
344 
345 zio_t *
346 zio_walk_children(zio_t *pio)
347 {
348 	zio_link_t *zl = pio->io_walk_link;
349 	list_t *cl = &pio->io_child_list;
350 
351 	zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
352 	pio->io_walk_link = zl;
353 
354 	if (zl == NULL)
355 		return (NULL);
356 
357 	ASSERT(zl->zl_parent == pio);
358 	return (zl->zl_child);
359 }
360 
361 zio_t *
362 zio_unique_parent(zio_t *cio)
363 {
364 	zio_t *pio = zio_walk_parents(cio);
365 
366 	VERIFY(zio_walk_parents(cio) == NULL);
367 	return (pio);
368 }
369 
370 void
371 zio_add_child(zio_t *pio, zio_t *cio)
372 {
373 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
374 
375 	/*
376 	 * Logical I/Os can have logical, gang, or vdev children.
377 	 * Gang I/Os can have gang or vdev children.
378 	 * Vdev I/Os can only have vdev children.
379 	 * The following ASSERT captures all of these constraints.
380 	 */
381 	ASSERT(cio->io_child_type <= pio->io_child_type);
382 
383 	zl->zl_parent = pio;
384 	zl->zl_child = cio;
385 
386 	mutex_enter(&cio->io_lock);
387 	mutex_enter(&pio->io_lock);
388 
389 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
390 
391 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
392 		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
393 
394 	list_insert_head(&pio->io_child_list, zl);
395 	list_insert_head(&cio->io_parent_list, zl);
396 
397 	pio->io_child_count++;
398 	cio->io_parent_count++;
399 
400 	mutex_exit(&pio->io_lock);
401 	mutex_exit(&cio->io_lock);
402 }
403 
404 static void
405 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
406 {
407 	ASSERT(zl->zl_parent == pio);
408 	ASSERT(zl->zl_child == cio);
409 
410 	mutex_enter(&cio->io_lock);
411 	mutex_enter(&pio->io_lock);
412 
413 	list_remove(&pio->io_child_list, zl);
414 	list_remove(&cio->io_parent_list, zl);
415 
416 	pio->io_child_count--;
417 	cio->io_parent_count--;
418 
419 	mutex_exit(&pio->io_lock);
420 	mutex_exit(&cio->io_lock);
421 
422 	kmem_cache_free(zio_link_cache, zl);
423 }
424 
425 static boolean_t
426 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
427 {
428 	uint64_t *countp = &zio->io_children[child][wait];
429 	boolean_t waiting = B_FALSE;
430 
431 	mutex_enter(&zio->io_lock);
432 	ASSERT(zio->io_stall == NULL);
433 	if (*countp != 0) {
434 		zio->io_stage >>= 1;
435 		zio->io_stall = countp;
436 		waiting = B_TRUE;
437 	}
438 	mutex_exit(&zio->io_lock);
439 
440 	return (waiting);
441 }
442 
443 static void
444 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
445 {
446 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
447 	int *errorp = &pio->io_child_error[zio->io_child_type];
448 
449 	mutex_enter(&pio->io_lock);
450 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
451 		*errorp = zio_worst_error(*errorp, zio->io_error);
452 	pio->io_reexecute |= zio->io_reexecute;
453 	ASSERT3U(*countp, >, 0);
454 	if (--*countp == 0 && pio->io_stall == countp) {
455 		pio->io_stall = NULL;
456 		mutex_exit(&pio->io_lock);
457 		zio_execute(pio);
458 	} else {
459 		mutex_exit(&pio->io_lock);
460 	}
461 }
462 
463 static void
464 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
465 {
466 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
467 		zio->io_error = zio->io_child_error[c];
468 }
469 
470 /*
471  * ==========================================================================
472  * Create the various types of I/O (read, write, free, etc)
473  * ==========================================================================
474  */
475 static zio_t *
476 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
477     void *data, uint64_t size, zio_done_func_t *done, void *private,
478     zio_type_t type, int priority, enum zio_flag flags,
479     vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
480     enum zio_stage stage, enum zio_stage pipeline)
481 {
482 	zio_t *zio;
483 
484 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
485 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
486 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
487 
488 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
489 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
490 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
491 
492 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
493 	bzero(zio, sizeof (zio_t));
494 
495 	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
496 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
497 
498 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
499 	    offsetof(zio_link_t, zl_parent_node));
500 	list_create(&zio->io_child_list, sizeof (zio_link_t),
501 	    offsetof(zio_link_t, zl_child_node));
502 
503 	if (vd != NULL)
504 		zio->io_child_type = ZIO_CHILD_VDEV;
505 	else if (flags & ZIO_FLAG_GANG_CHILD)
506 		zio->io_child_type = ZIO_CHILD_GANG;
507 	else if (flags & ZIO_FLAG_DDT_CHILD)
508 		zio->io_child_type = ZIO_CHILD_DDT;
509 	else
510 		zio->io_child_type = ZIO_CHILD_LOGICAL;
511 
512 	if (bp != NULL) {
513 		zio->io_bp = (blkptr_t *)bp;
514 		zio->io_bp_copy = *bp;
515 		zio->io_bp_orig = *bp;
516 		if (type != ZIO_TYPE_WRITE ||
517 		    zio->io_child_type == ZIO_CHILD_DDT)
518 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
519 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
520 			zio->io_logical = zio;
521 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
522 			pipeline |= ZIO_GANG_STAGES;
523 	}
524 
525 	zio->io_spa = spa;
526 	zio->io_txg = txg;
527 	zio->io_done = done;
528 	zio->io_private = private;
529 	zio->io_type = type;
530 	zio->io_priority = priority;
531 	zio->io_vd = vd;
532 	zio->io_offset = offset;
533 	zio->io_orig_data = zio->io_data = data;
534 	zio->io_orig_size = zio->io_size = size;
535 	zio->io_orig_flags = zio->io_flags = flags;
536 	zio->io_orig_stage = zio->io_stage = stage;
537 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
538 
539 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
540 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
541 
542 	if (zb != NULL)
543 		zio->io_bookmark = *zb;
544 
545 	if (pio != NULL) {
546 		if (zio->io_logical == NULL)
547 			zio->io_logical = pio->io_logical;
548 		if (zio->io_child_type == ZIO_CHILD_GANG)
549 			zio->io_gang_leader = pio->io_gang_leader;
550 		zio_add_child(pio, zio);
551 	}
552 
553 	return (zio);
554 }
555 
556 static void
557 zio_destroy(zio_t *zio)
558 {
559 	list_destroy(&zio->io_parent_list);
560 	list_destroy(&zio->io_child_list);
561 	mutex_destroy(&zio->io_lock);
562 	cv_destroy(&zio->io_cv);
563 	kmem_cache_free(zio_cache, zio);
564 }
565 
566 zio_t *
567 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
568     void *private, enum zio_flag flags)
569 {
570 	zio_t *zio;
571 
572 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
573 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
574 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
575 
576 	return (zio);
577 }
578 
579 zio_t *
580 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
581 {
582 	return (zio_null(NULL, spa, NULL, done, private, flags));
583 }
584 
585 zio_t *
586 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
587     void *data, uint64_t size, zio_done_func_t *done, void *private,
588     int priority, enum zio_flag flags, const zbookmark_t *zb)
589 {
590 	zio_t *zio;
591 
592 	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
593 	    data, size, done, private,
594 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
595 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
596 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
597 
598 	return (zio);
599 }
600 
601 zio_t *
602 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
603     void *data, uint64_t size, const zio_prop_t *zp,
604     zio_done_func_t *ready, zio_done_func_t *done, void *private,
605     int priority, enum zio_flag flags, const zbookmark_t *zb)
606 {
607 	zio_t *zio;
608 
609 	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
610 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
611 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
612 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
613 	    zp->zp_type < DMU_OT_NUMTYPES &&
614 	    zp->zp_level < 32 &&
615 	    zp->zp_copies > 0 &&
616 	    zp->zp_copies <= spa_max_replication(spa) &&
617 	    zp->zp_dedup <= 1 &&
618 	    zp->zp_dedup_verify <= 1);
619 
620 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
621 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
622 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
623 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
624 
625 	zio->io_ready = ready;
626 	zio->io_prop = *zp;
627 
628 	return (zio);
629 }
630 
631 zio_t *
632 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
633     uint64_t size, zio_done_func_t *done, void *private, int priority,
634     enum zio_flag flags, zbookmark_t *zb)
635 {
636 	zio_t *zio;
637 
638 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
639 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
640 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
641 
642 	return (zio);
643 }
644 
645 void
646 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
647 {
648 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
649 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
650 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
651 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
652 
653 	zio->io_prop.zp_copies = copies;
654 	zio->io_bp_override = bp;
655 }
656 
657 void
658 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
659 {
660 	bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
661 }
662 
663 zio_t *
664 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
665     enum zio_flag flags)
666 {
667 	zio_t *zio;
668 
669 	dprintf_bp(bp, "freeing in txg %llu, pass %u",
670 	    (longlong_t)txg, spa->spa_sync_pass);
671 
672 	ASSERT(!BP_IS_HOLE(bp));
673 	ASSERT(spa_syncing_txg(spa) == txg);
674 	ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
675 
676 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
677 	    NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
678 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
679 
680 	return (zio);
681 }
682 
683 zio_t *
684 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
685     zio_done_func_t *done, void *private, enum zio_flag flags)
686 {
687 	zio_t *zio;
688 
689 	/*
690 	 * A claim is an allocation of a specific block.  Claims are needed
691 	 * to support immediate writes in the intent log.  The issue is that
692 	 * immediate writes contain committed data, but in a txg that was
693 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
694 	 * the intent log claims all blocks that contain immediate write data
695 	 * so that the SPA knows they're in use.
696 	 *
697 	 * All claims *must* be resolved in the first txg -- before the SPA
698 	 * starts allocating blocks -- so that nothing is allocated twice.
699 	 * If txg == 0 we just verify that the block is claimable.
700 	 */
701 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
702 	ASSERT(txg == spa_first_txg(spa) || txg == 0);
703 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
704 
705 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
706 	    done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
707 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
708 
709 	return (zio);
710 }
711 
712 zio_t *
713 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
714     zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
715 {
716 	zio_t *zio;
717 	int c;
718 
719 	if (vd->vdev_children == 0) {
720 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
721 		    ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
722 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
723 
724 		zio->io_cmd = cmd;
725 	} else {
726 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
727 
728 		for (c = 0; c < vd->vdev_children; c++)
729 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
730 			    done, private, priority, flags));
731 	}
732 
733 	return (zio);
734 }
735 
736 zio_t *
737 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
738     void *data, int checksum, zio_done_func_t *done, void *private,
739     int priority, enum zio_flag flags, boolean_t labels)
740 {
741 	zio_t *zio;
742 
743 	ASSERT(vd->vdev_children == 0);
744 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
745 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
746 	ASSERT3U(offset + size, <=, vd->vdev_psize);
747 
748 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
749 	    ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
750 	    ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
751 
752 	zio->io_prop.zp_checksum = checksum;
753 
754 	return (zio);
755 }
756 
757 zio_t *
758 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
759     void *data, int checksum, zio_done_func_t *done, void *private,
760     int priority, enum zio_flag flags, boolean_t labels)
761 {
762 	zio_t *zio;
763 
764 	ASSERT(vd->vdev_children == 0);
765 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
766 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
767 	ASSERT3U(offset + size, <=, vd->vdev_psize);
768 
769 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
770 	    ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
771 	    ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
772 
773 	zio->io_prop.zp_checksum = checksum;
774 
775 	if (zio_checksum_table[checksum].ci_eck) {
776 		/*
777 		 * zec checksums are necessarily destructive -- they modify
778 		 * the end of the write buffer to hold the verifier/checksum.
779 		 * Therefore, we must make a local copy in case the data is
780 		 * being written to multiple places in parallel.
781 		 */
782 		void *wbuf = zio_buf_alloc(size);
783 		bcopy(data, wbuf, size);
784 		zio_push_transform(zio, wbuf, size, size, NULL);
785 	}
786 
787 	return (zio);
788 }
789 
790 /*
791  * Create a child I/O to do some work for us.
792  */
793 zio_t *
794 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
795 	void *data, uint64_t size, int type, int priority, enum zio_flag flags,
796 	zio_done_func_t *done, void *private)
797 {
798 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
799 	zio_t *zio;
800 
801 	ASSERT(vd->vdev_parent ==
802 	    (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
803 
804 	if (type == ZIO_TYPE_READ && bp != NULL) {
805 		/*
806 		 * If we have the bp, then the child should perform the
807 		 * checksum and the parent need not.  This pushes error
808 		 * detection as close to the leaves as possible and
809 		 * eliminates redundant checksums in the interior nodes.
810 		 */
811 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
812 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
813 	}
814 
815 	if (vd->vdev_children == 0)
816 		offset += VDEV_LABEL_START_SIZE;
817 
818 	flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
819 
820 	/*
821 	 * If we've decided to do a repair, the write is not speculative --
822 	 * even if the original read was.
823 	 */
824 	if (flags & ZIO_FLAG_IO_REPAIR)
825 		flags &= ~ZIO_FLAG_SPECULATIVE;
826 
827 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
828 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
829 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
830 
831 	return (zio);
832 }
833 
834 zio_t *
835 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
836 	int type, int priority, enum zio_flag flags,
837 	zio_done_func_t *done, void *private)
838 {
839 	zio_t *zio;
840 
841 	ASSERT(vd->vdev_ops->vdev_op_leaf);
842 
843 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
844 	    data, size, done, private, type, priority,
845 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
846 	    vd, offset, NULL,
847 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
848 
849 	return (zio);
850 }
851 
852 void
853 zio_flush(zio_t *zio, vdev_t *vd)
854 {
855 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
856 	    NULL, NULL, ZIO_PRIORITY_NOW,
857 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
858 }
859 
860 void
861 zio_shrink(zio_t *zio, uint64_t size)
862 {
863 	ASSERT(zio->io_executor == NULL);
864 	ASSERT(zio->io_orig_size == zio->io_size);
865 	ASSERT(size <= zio->io_size);
866 
867 	/*
868 	 * We don't shrink for raidz because of problems with the
869 	 * reconstruction when reading back less than the block size.
870 	 * Note, BP_IS_RAIDZ() assumes no compression.
871 	 */
872 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
873 	if (!BP_IS_RAIDZ(zio->io_bp))
874 		zio->io_orig_size = zio->io_size = size;
875 }
876 
877 /*
878  * ==========================================================================
879  * Prepare to read and write logical blocks
880  * ==========================================================================
881  */
882 
883 static int
884 zio_read_bp_init(zio_t *zio)
885 {
886 	blkptr_t *bp = zio->io_bp;
887 
888 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
889 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
890 	    !(zio->io_flags & ZIO_FLAG_RAW)) {
891 		uint64_t psize = BP_GET_PSIZE(bp);
892 		void *cbuf = zio_buf_alloc(psize);
893 
894 		zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
895 	}
896 
897 	if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
898 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
899 
900 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
901 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
902 
903 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
904 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
905 
906 	return (ZIO_PIPELINE_CONTINUE);
907 }
908 
909 static int
910 zio_write_bp_init(zio_t *zio)
911 {
912 	spa_t *spa = zio->io_spa;
913 	zio_prop_t *zp = &zio->io_prop;
914 	enum zio_compress compress = zp->zp_compress;
915 	blkptr_t *bp = zio->io_bp;
916 	uint64_t lsize = zio->io_size;
917 	uint64_t psize = lsize;
918 	int pass = 1;
919 
920 	/*
921 	 * If our children haven't all reached the ready stage,
922 	 * wait for them and then repeat this pipeline stage.
923 	 */
924 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
925 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
926 		return (ZIO_PIPELINE_STOP);
927 
928 	if (!IO_IS_ALLOCATING(zio))
929 		return (ZIO_PIPELINE_CONTINUE);
930 
931 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
932 
933 	if (zio->io_bp_override) {
934 		ASSERT(bp->blk_birth != zio->io_txg);
935 		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
936 
937 		*bp = *zio->io_bp_override;
938 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
939 
940 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
941 			return (ZIO_PIPELINE_CONTINUE);
942 
943 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
944 		    zp->zp_dedup_verify);
945 
946 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
947 			BP_SET_DEDUP(bp, 1);
948 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
949 			return (ZIO_PIPELINE_CONTINUE);
950 		}
951 		zio->io_bp_override = NULL;
952 		BP_ZERO(bp);
953 	}
954 
955 	if (bp->blk_birth == zio->io_txg) {
956 		/*
957 		 * We're rewriting an existing block, which means we're
958 		 * working on behalf of spa_sync().  For spa_sync() to
959 		 * converge, it must eventually be the case that we don't
960 		 * have to allocate new blocks.  But compression changes
961 		 * the blocksize, which forces a reallocate, and makes
962 		 * convergence take longer.  Therefore, after the first
963 		 * few passes, stop compressing to ensure convergence.
964 		 */
965 		pass = spa_sync_pass(spa);
966 
967 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
968 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
969 		ASSERT(!BP_GET_DEDUP(bp));
970 
971 		if (pass > SYNC_PASS_DONT_COMPRESS)
972 			compress = ZIO_COMPRESS_OFF;
973 
974 		/* Make sure someone doesn't change their mind on overwrites */
975 		ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
976 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
977 	}
978 
979 	if (compress != ZIO_COMPRESS_OFF) {
980 		void *cbuf = zio_buf_alloc(lsize);
981 		psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
982 		if (psize == 0 || psize == lsize) {
983 			compress = ZIO_COMPRESS_OFF;
984 			zio_buf_free(cbuf, lsize);
985 		} else {
986 			ASSERT(psize < lsize);
987 			zio_push_transform(zio, cbuf, psize, lsize, NULL);
988 		}
989 	}
990 
991 	/*
992 	 * The final pass of spa_sync() must be all rewrites, but the first
993 	 * few passes offer a trade-off: allocating blocks defers convergence,
994 	 * but newly allocated blocks are sequential, so they can be written
995 	 * to disk faster.  Therefore, we allow the first few passes of
996 	 * spa_sync() to allocate new blocks, but force rewrites after that.
997 	 * There should only be a handful of blocks after pass 1 in any case.
998 	 */
999 	if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1000 	    pass > SYNC_PASS_REWRITE) {
1001 		ASSERT(psize != 0);
1002 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1003 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1004 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1005 	} else {
1006 		BP_ZERO(bp);
1007 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1008 	}
1009 
1010 	if (psize == 0) {
1011 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1012 	} else {
1013 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1014 		BP_SET_LSIZE(bp, lsize);
1015 		BP_SET_PSIZE(bp, psize);
1016 		BP_SET_COMPRESS(bp, compress);
1017 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1018 		BP_SET_TYPE(bp, zp->zp_type);
1019 		BP_SET_LEVEL(bp, zp->zp_level);
1020 		BP_SET_DEDUP(bp, zp->zp_dedup);
1021 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1022 		if (zp->zp_dedup) {
1023 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1024 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1025 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1026 		}
1027 	}
1028 
1029 	return (ZIO_PIPELINE_CONTINUE);
1030 }
1031 
1032 static int
1033 zio_free_bp_init(zio_t *zio)
1034 {
1035 	blkptr_t *bp = zio->io_bp;
1036 
1037 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1038 		if (BP_GET_DEDUP(bp))
1039 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1040 	}
1041 
1042 	return (ZIO_PIPELINE_CONTINUE);
1043 }
1044 
1045 /*
1046  * ==========================================================================
1047  * Execute the I/O pipeline
1048  * ==========================================================================
1049  */
1050 
1051 static void
1052 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1053 {
1054 	spa_t *spa = zio->io_spa;
1055 	zio_type_t t = zio->io_type;
1056 	int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0);
1057 
1058 	/*
1059 	 * If we're a config writer or a probe, the normal issue and
1060 	 * interrupt threads may all be blocked waiting for the config lock.
1061 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1062 	 */
1063 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1064 		t = ZIO_TYPE_NULL;
1065 
1066 	/*
1067 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1068 	 */
1069 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1070 		t = ZIO_TYPE_NULL;
1071 
1072 	/*
1073 	 * If this is a high priority I/O, then use the high priority taskq.
1074 	 */
1075 	if (zio->io_priority == ZIO_PRIORITY_NOW &&
1076 	    spa->spa_zio_taskq[t][q + 1] != NULL)
1077 		q++;
1078 
1079 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1080 	(void) taskq_dispatch(spa->spa_zio_taskq[t][q],
1081 	    (task_func_t *)zio_execute, zio, flags);
1082 }
1083 
1084 static boolean_t
1085 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1086 {
1087 	kthread_t *executor = zio->io_executor;
1088 	spa_t *spa = zio->io_spa;
1089 
1090 	for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1091 		if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1092 			return (B_TRUE);
1093 
1094 	return (B_FALSE);
1095 }
1096 
1097 static int
1098 zio_issue_async(zio_t *zio)
1099 {
1100 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1101 
1102 	return (ZIO_PIPELINE_STOP);
1103 }
1104 
1105 void
1106 zio_interrupt(zio_t *zio)
1107 {
1108 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1109 }
1110 
1111 /*
1112  * Execute the I/O pipeline until one of the following occurs:
1113  * (1) the I/O completes; (2) the pipeline stalls waiting for
1114  * dependent child I/Os; (3) the I/O issues, so we're waiting
1115  * for an I/O completion interrupt; (4) the I/O is delegated by
1116  * vdev-level caching or aggregation; (5) the I/O is deferred
1117  * due to vdev-level queueing; (6) the I/O is handed off to
1118  * another thread.  In all cases, the pipeline stops whenever
1119  * there's no CPU work; it never burns a thread in cv_wait().
1120  *
1121  * There's no locking on io_stage because there's no legitimate way
1122  * for multiple threads to be attempting to process the same I/O.
1123  */
1124 static zio_pipe_stage_t *zio_pipeline[];
1125 
1126 void
1127 zio_execute(zio_t *zio)
1128 {
1129 	zio->io_executor = curthread;
1130 
1131 	while (zio->io_stage < ZIO_STAGE_DONE) {
1132 		enum zio_stage pipeline = zio->io_pipeline;
1133 		enum zio_stage stage = zio->io_stage;
1134 		int rv;
1135 
1136 		ASSERT(!MUTEX_HELD(&zio->io_lock));
1137 		ASSERT(ISP2(stage));
1138 		ASSERT(zio->io_stall == NULL);
1139 
1140 		do {
1141 			stage <<= 1;
1142 		} while ((stage & pipeline) == 0);
1143 
1144 		ASSERT(stage <= ZIO_STAGE_DONE);
1145 
1146 		/*
1147 		 * If we are in interrupt context and this pipeline stage
1148 		 * will grab a config lock that is held across I/O,
1149 		 * or may wait for an I/O that needs an interrupt thread
1150 		 * to complete, issue async to avoid deadlock.
1151 		 *
1152 		 * For VDEV_IO_START, we cut in line so that the io will
1153 		 * be sent to disk promptly.
1154 		 */
1155 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1156 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1157 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1158 			    zio_requeue_io_start_cut_in_line : B_FALSE;
1159 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1160 			return;
1161 		}
1162 
1163 		zio->io_stage = stage;
1164 		rv = zio_pipeline[highbit(stage) - 1](zio);
1165 
1166 		if (rv == ZIO_PIPELINE_STOP)
1167 			return;
1168 
1169 		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1170 	}
1171 }
1172 
1173 /*
1174  * ==========================================================================
1175  * Initiate I/O, either sync or async
1176  * ==========================================================================
1177  */
1178 int
1179 zio_wait(zio_t *zio)
1180 {
1181 	int error;
1182 
1183 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1184 	ASSERT(zio->io_executor == NULL);
1185 
1186 	zio->io_waiter = curthread;
1187 
1188 	zio_execute(zio);
1189 
1190 	mutex_enter(&zio->io_lock);
1191 	while (zio->io_executor != NULL)
1192 		cv_wait(&zio->io_cv, &zio->io_lock);
1193 	mutex_exit(&zio->io_lock);
1194 
1195 	error = zio->io_error;
1196 	zio_destroy(zio);
1197 
1198 	return (error);
1199 }
1200 
1201 void
1202 zio_nowait(zio_t *zio)
1203 {
1204 	ASSERT(zio->io_executor == NULL);
1205 
1206 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1207 	    zio_unique_parent(zio) == NULL) {
1208 		/*
1209 		 * This is a logical async I/O with no parent to wait for it.
1210 		 * We add it to the spa_async_root_zio "Godfather" I/O which
1211 		 * will ensure they complete prior to unloading the pool.
1212 		 */
1213 		spa_t *spa = zio->io_spa;
1214 
1215 		zio_add_child(spa->spa_async_zio_root, zio);
1216 	}
1217 
1218 	zio_execute(zio);
1219 }
1220 
1221 /*
1222  * ==========================================================================
1223  * Reexecute or suspend/resume failed I/O
1224  * ==========================================================================
1225  */
1226 
1227 static void
1228 zio_reexecute(zio_t *pio)
1229 {
1230 	zio_t *cio, *cio_next;
1231 
1232 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1233 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1234 	ASSERT(pio->io_gang_leader == NULL);
1235 	ASSERT(pio->io_gang_tree == NULL);
1236 
1237 	pio->io_flags = pio->io_orig_flags;
1238 	pio->io_stage = pio->io_orig_stage;
1239 	pio->io_pipeline = pio->io_orig_pipeline;
1240 	pio->io_reexecute = 0;
1241 	pio->io_error = 0;
1242 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1243 		pio->io_state[w] = 0;
1244 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1245 		pio->io_child_error[c] = 0;
1246 
1247 	if (IO_IS_ALLOCATING(pio))
1248 		BP_ZERO(pio->io_bp);
1249 
1250 	/*
1251 	 * As we reexecute pio's children, new children could be created.
1252 	 * New children go to the head of pio's io_child_list, however,
1253 	 * so we will (correctly) not reexecute them.  The key is that
1254 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1255 	 * cannot be affected by any side effects of reexecuting 'cio'.
1256 	 */
1257 	for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1258 		cio_next = zio_walk_children(pio);
1259 		mutex_enter(&pio->io_lock);
1260 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1261 			pio->io_children[cio->io_child_type][w]++;
1262 		mutex_exit(&pio->io_lock);
1263 		zio_reexecute(cio);
1264 	}
1265 
1266 	/*
1267 	 * Now that all children have been reexecuted, execute the parent.
1268 	 * We don't reexecute "The Godfather" I/O here as it's the
1269 	 * responsibility of the caller to wait on him.
1270 	 */
1271 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1272 		zio_execute(pio);
1273 }
1274 
1275 void
1276 zio_suspend(spa_t *spa, zio_t *zio)
1277 {
1278 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1279 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1280 		    "failure and the failure mode property for this pool "
1281 		    "is set to panic.", spa_name(spa));
1282 
1283 	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1284 
1285 	mutex_enter(&spa->spa_suspend_lock);
1286 
1287 	if (spa->spa_suspend_zio_root == NULL)
1288 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1289 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1290 		    ZIO_FLAG_GODFATHER);
1291 
1292 	spa->spa_suspended = B_TRUE;
1293 
1294 	if (zio != NULL) {
1295 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1296 		ASSERT(zio != spa->spa_suspend_zio_root);
1297 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1298 		ASSERT(zio_unique_parent(zio) == NULL);
1299 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1300 		zio_add_child(spa->spa_suspend_zio_root, zio);
1301 	}
1302 
1303 	mutex_exit(&spa->spa_suspend_lock);
1304 }
1305 
1306 int
1307 zio_resume(spa_t *spa)
1308 {
1309 	zio_t *pio;
1310 
1311 	/*
1312 	 * Reexecute all previously suspended i/o.
1313 	 */
1314 	mutex_enter(&spa->spa_suspend_lock);
1315 	spa->spa_suspended = B_FALSE;
1316 	cv_broadcast(&spa->spa_suspend_cv);
1317 	pio = spa->spa_suspend_zio_root;
1318 	spa->spa_suspend_zio_root = NULL;
1319 	mutex_exit(&spa->spa_suspend_lock);
1320 
1321 	if (pio == NULL)
1322 		return (0);
1323 
1324 	zio_reexecute(pio);
1325 	return (zio_wait(pio));
1326 }
1327 
1328 void
1329 zio_resume_wait(spa_t *spa)
1330 {
1331 	mutex_enter(&spa->spa_suspend_lock);
1332 	while (spa_suspended(spa))
1333 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1334 	mutex_exit(&spa->spa_suspend_lock);
1335 }
1336 
1337 /*
1338  * ==========================================================================
1339  * Gang blocks.
1340  *
1341  * A gang block is a collection of small blocks that looks to the DMU
1342  * like one large block.  When zio_dva_allocate() cannot find a block
1343  * of the requested size, due to either severe fragmentation or the pool
1344  * being nearly full, it calls zio_write_gang_block() to construct the
1345  * block from smaller fragments.
1346  *
1347  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1348  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
1349  * an indirect block: it's an array of block pointers.  It consumes
1350  * only one sector and hence is allocatable regardless of fragmentation.
1351  * The gang header's bps point to its gang members, which hold the data.
1352  *
1353  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1354  * as the verifier to ensure uniqueness of the SHA256 checksum.
1355  * Critically, the gang block bp's blk_cksum is the checksum of the data,
1356  * not the gang header.  This ensures that data block signatures (needed for
1357  * deduplication) are independent of how the block is physically stored.
1358  *
1359  * Gang blocks can be nested: a gang member may itself be a gang block.
1360  * Thus every gang block is a tree in which root and all interior nodes are
1361  * gang headers, and the leaves are normal blocks that contain user data.
1362  * The root of the gang tree is called the gang leader.
1363  *
1364  * To perform any operation (read, rewrite, free, claim) on a gang block,
1365  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1366  * in the io_gang_tree field of the original logical i/o by recursively
1367  * reading the gang leader and all gang headers below it.  This yields
1368  * an in-core tree containing the contents of every gang header and the
1369  * bps for every constituent of the gang block.
1370  *
1371  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1372  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
1373  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1374  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1375  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1376  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
1377  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1378  * of the gang header plus zio_checksum_compute() of the data to update the
1379  * gang header's blk_cksum as described above.
1380  *
1381  * The two-phase assemble/issue model solves the problem of partial failure --
1382  * what if you'd freed part of a gang block but then couldn't read the
1383  * gang header for another part?  Assembling the entire gang tree first
1384  * ensures that all the necessary gang header I/O has succeeded before
1385  * starting the actual work of free, claim, or write.  Once the gang tree
1386  * is assembled, free and claim are in-memory operations that cannot fail.
1387  *
1388  * In the event that a gang write fails, zio_dva_unallocate() walks the
1389  * gang tree to immediately free (i.e. insert back into the space map)
1390  * everything we've allocated.  This ensures that we don't get ENOSPC
1391  * errors during repeated suspend/resume cycles due to a flaky device.
1392  *
1393  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
1394  * the gang tree, we won't modify the block, so we can safely defer the free
1395  * (knowing that the block is still intact).  If we *can* assemble the gang
1396  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1397  * each constituent bp and we can allocate a new block on the next sync pass.
1398  *
1399  * In all cases, the gang tree allows complete recovery from partial failure.
1400  * ==========================================================================
1401  */
1402 
1403 static zio_t *
1404 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1405 {
1406 	if (gn != NULL)
1407 		return (pio);
1408 
1409 	return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1410 	    NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1411 	    &pio->io_bookmark));
1412 }
1413 
1414 zio_t *
1415 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1416 {
1417 	zio_t *zio;
1418 
1419 	if (gn != NULL) {
1420 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1421 		    gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1422 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1423 		/*
1424 		 * As we rewrite each gang header, the pipeline will compute
1425 		 * a new gang block header checksum for it; but no one will
1426 		 * compute a new data checksum, so we do that here.  The one
1427 		 * exception is the gang leader: the pipeline already computed
1428 		 * its data checksum because that stage precedes gang assembly.
1429 		 * (Presently, nothing actually uses interior data checksums;
1430 		 * this is just good hygiene.)
1431 		 */
1432 		if (gn != pio->io_gang_leader->io_gang_tree) {
1433 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1434 			    data, BP_GET_PSIZE(bp));
1435 		}
1436 		/*
1437 		 * If we are here to damage data for testing purposes,
1438 		 * leave the GBH alone so that we can detect the damage.
1439 		 */
1440 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1441 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1442 	} else {
1443 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1444 		    data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1445 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1446 	}
1447 
1448 	return (zio);
1449 }
1450 
1451 /* ARGSUSED */
1452 zio_t *
1453 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1454 {
1455 	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1456 	    ZIO_GANG_CHILD_FLAGS(pio)));
1457 }
1458 
1459 /* ARGSUSED */
1460 zio_t *
1461 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1462 {
1463 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1464 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1465 }
1466 
1467 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1468 	NULL,
1469 	zio_read_gang,
1470 	zio_rewrite_gang,
1471 	zio_free_gang,
1472 	zio_claim_gang,
1473 	NULL
1474 };
1475 
1476 static void zio_gang_tree_assemble_done(zio_t *zio);
1477 
1478 static zio_gang_node_t *
1479 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1480 {
1481 	zio_gang_node_t *gn;
1482 
1483 	ASSERT(*gnpp == NULL);
1484 
1485 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1486 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1487 	*gnpp = gn;
1488 
1489 	return (gn);
1490 }
1491 
1492 static void
1493 zio_gang_node_free(zio_gang_node_t **gnpp)
1494 {
1495 	zio_gang_node_t *gn = *gnpp;
1496 
1497 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1498 		ASSERT(gn->gn_child[g] == NULL);
1499 
1500 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1501 	kmem_free(gn, sizeof (*gn));
1502 	*gnpp = NULL;
1503 }
1504 
1505 static void
1506 zio_gang_tree_free(zio_gang_node_t **gnpp)
1507 {
1508 	zio_gang_node_t *gn = *gnpp;
1509 
1510 	if (gn == NULL)
1511 		return;
1512 
1513 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1514 		zio_gang_tree_free(&gn->gn_child[g]);
1515 
1516 	zio_gang_node_free(gnpp);
1517 }
1518 
1519 static void
1520 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1521 {
1522 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1523 
1524 	ASSERT(gio->io_gang_leader == gio);
1525 	ASSERT(BP_IS_GANG(bp));
1526 
1527 	zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1528 	    SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1529 	    gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1530 }
1531 
1532 static void
1533 zio_gang_tree_assemble_done(zio_t *zio)
1534 {
1535 	zio_t *gio = zio->io_gang_leader;
1536 	zio_gang_node_t *gn = zio->io_private;
1537 	blkptr_t *bp = zio->io_bp;
1538 
1539 	ASSERT(gio == zio_unique_parent(zio));
1540 	ASSERT(zio->io_child_count == 0);
1541 
1542 	if (zio->io_error)
1543 		return;
1544 
1545 	if (BP_SHOULD_BYTESWAP(bp))
1546 		byteswap_uint64_array(zio->io_data, zio->io_size);
1547 
1548 	ASSERT(zio->io_data == gn->gn_gbh);
1549 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1550 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1551 
1552 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1553 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1554 		if (!BP_IS_GANG(gbp))
1555 			continue;
1556 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1557 	}
1558 }
1559 
1560 static void
1561 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1562 {
1563 	zio_t *gio = pio->io_gang_leader;
1564 	zio_t *zio;
1565 
1566 	ASSERT(BP_IS_GANG(bp) == !!gn);
1567 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1568 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1569 
1570 	/*
1571 	 * If you're a gang header, your data is in gn->gn_gbh.
1572 	 * If you're a gang member, your data is in 'data' and gn == NULL.
1573 	 */
1574 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1575 
1576 	if (gn != NULL) {
1577 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1578 
1579 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1580 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1581 			if (BP_IS_HOLE(gbp))
1582 				continue;
1583 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1584 			data = (char *)data + BP_GET_PSIZE(gbp);
1585 		}
1586 	}
1587 
1588 	if (gn == gio->io_gang_tree)
1589 		ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1590 
1591 	if (zio != pio)
1592 		zio_nowait(zio);
1593 }
1594 
1595 static int
1596 zio_gang_assemble(zio_t *zio)
1597 {
1598 	blkptr_t *bp = zio->io_bp;
1599 
1600 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1601 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1602 
1603 	zio->io_gang_leader = zio;
1604 
1605 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1606 
1607 	return (ZIO_PIPELINE_CONTINUE);
1608 }
1609 
1610 static int
1611 zio_gang_issue(zio_t *zio)
1612 {
1613 	blkptr_t *bp = zio->io_bp;
1614 
1615 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1616 		return (ZIO_PIPELINE_STOP);
1617 
1618 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1619 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1620 
1621 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1622 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1623 	else
1624 		zio_gang_tree_free(&zio->io_gang_tree);
1625 
1626 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1627 
1628 	return (ZIO_PIPELINE_CONTINUE);
1629 }
1630 
1631 static void
1632 zio_write_gang_member_ready(zio_t *zio)
1633 {
1634 	zio_t *pio = zio_unique_parent(zio);
1635 	zio_t *gio = zio->io_gang_leader;
1636 	dva_t *cdva = zio->io_bp->blk_dva;
1637 	dva_t *pdva = pio->io_bp->blk_dva;
1638 	uint64_t asize;
1639 
1640 	if (BP_IS_HOLE(zio->io_bp))
1641 		return;
1642 
1643 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1644 
1645 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1646 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1647 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1648 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1649 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1650 
1651 	mutex_enter(&pio->io_lock);
1652 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1653 		ASSERT(DVA_GET_GANG(&pdva[d]));
1654 		asize = DVA_GET_ASIZE(&pdva[d]);
1655 		asize += DVA_GET_ASIZE(&cdva[d]);
1656 		DVA_SET_ASIZE(&pdva[d], asize);
1657 	}
1658 	mutex_exit(&pio->io_lock);
1659 }
1660 
1661 static int
1662 zio_write_gang_block(zio_t *pio)
1663 {
1664 	spa_t *spa = pio->io_spa;
1665 	blkptr_t *bp = pio->io_bp;
1666 	zio_t *gio = pio->io_gang_leader;
1667 	zio_t *zio;
1668 	zio_gang_node_t *gn, **gnpp;
1669 	zio_gbh_phys_t *gbh;
1670 	uint64_t txg = pio->io_txg;
1671 	uint64_t resid = pio->io_size;
1672 	uint64_t lsize;
1673 	int copies = gio->io_prop.zp_copies;
1674 	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1675 	zio_prop_t zp;
1676 	int error;
1677 
1678 	error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1679 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1680 	    METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1681 	if (error) {
1682 		pio->io_error = error;
1683 		return (ZIO_PIPELINE_CONTINUE);
1684 	}
1685 
1686 	if (pio == gio) {
1687 		gnpp = &gio->io_gang_tree;
1688 	} else {
1689 		gnpp = pio->io_private;
1690 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
1691 	}
1692 
1693 	gn = zio_gang_node_alloc(gnpp);
1694 	gbh = gn->gn_gbh;
1695 	bzero(gbh, SPA_GANGBLOCKSIZE);
1696 
1697 	/*
1698 	 * Create the gang header.
1699 	 */
1700 	zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1701 	    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1702 
1703 	/*
1704 	 * Create and nowait the gang children.
1705 	 */
1706 	for (int g = 0; resid != 0; resid -= lsize, g++) {
1707 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1708 		    SPA_MINBLOCKSIZE);
1709 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1710 
1711 		zp.zp_checksum = gio->io_prop.zp_checksum;
1712 		zp.zp_compress = ZIO_COMPRESS_OFF;
1713 		zp.zp_type = DMU_OT_NONE;
1714 		zp.zp_level = 0;
1715 		zp.zp_copies = gio->io_prop.zp_copies;
1716 		zp.zp_dedup = 0;
1717 		zp.zp_dedup_verify = 0;
1718 
1719 		zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1720 		    (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1721 		    zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1722 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1723 		    &pio->io_bookmark));
1724 	}
1725 
1726 	/*
1727 	 * Set pio's pipeline to just wait for zio to finish.
1728 	 */
1729 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1730 
1731 	zio_nowait(zio);
1732 
1733 	return (ZIO_PIPELINE_CONTINUE);
1734 }
1735 
1736 /*
1737  * ==========================================================================
1738  * Dedup
1739  * ==========================================================================
1740  */
1741 static void
1742 zio_ddt_child_read_done(zio_t *zio)
1743 {
1744 	blkptr_t *bp = zio->io_bp;
1745 	ddt_entry_t *dde = zio->io_private;
1746 	ddt_phys_t *ddp;
1747 	zio_t *pio = zio_unique_parent(zio);
1748 
1749 	mutex_enter(&pio->io_lock);
1750 	ddp = ddt_phys_select(dde, bp);
1751 	if (zio->io_error == 0)
1752 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
1753 	if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1754 		dde->dde_repair_data = zio->io_data;
1755 	else
1756 		zio_buf_free(zio->io_data, zio->io_size);
1757 	mutex_exit(&pio->io_lock);
1758 }
1759 
1760 static int
1761 zio_ddt_read_start(zio_t *zio)
1762 {
1763 	blkptr_t *bp = zio->io_bp;
1764 
1765 	ASSERT(BP_GET_DEDUP(bp));
1766 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1767 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1768 
1769 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1770 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1771 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1772 		ddt_phys_t *ddp = dde->dde_phys;
1773 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1774 		blkptr_t blk;
1775 
1776 		ASSERT(zio->io_vsd == NULL);
1777 		zio->io_vsd = dde;
1778 
1779 		if (ddp_self == NULL)
1780 			return (ZIO_PIPELINE_CONTINUE);
1781 
1782 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1783 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1784 				continue;
1785 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1786 			    &blk);
1787 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
1788 			    zio_buf_alloc(zio->io_size), zio->io_size,
1789 			    zio_ddt_child_read_done, dde, zio->io_priority,
1790 			    ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1791 			    &zio->io_bookmark));
1792 		}
1793 		return (ZIO_PIPELINE_CONTINUE);
1794 	}
1795 
1796 	zio_nowait(zio_read(zio, zio->io_spa, bp,
1797 	    zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1798 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1799 
1800 	return (ZIO_PIPELINE_CONTINUE);
1801 }
1802 
1803 static int
1804 zio_ddt_read_done(zio_t *zio)
1805 {
1806 	blkptr_t *bp = zio->io_bp;
1807 
1808 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1809 		return (ZIO_PIPELINE_STOP);
1810 
1811 	ASSERT(BP_GET_DEDUP(bp));
1812 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1813 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1814 
1815 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1816 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1817 		ddt_entry_t *dde = zio->io_vsd;
1818 		if (ddt == NULL) {
1819 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1820 			return (ZIO_PIPELINE_CONTINUE);
1821 		}
1822 		if (dde == NULL) {
1823 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1824 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1825 			return (ZIO_PIPELINE_STOP);
1826 		}
1827 		if (dde->dde_repair_data != NULL) {
1828 			bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1829 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
1830 		}
1831 		ddt_repair_done(ddt, dde);
1832 		zio->io_vsd = NULL;
1833 	}
1834 
1835 	ASSERT(zio->io_vsd == NULL);
1836 
1837 	return (ZIO_PIPELINE_CONTINUE);
1838 }
1839 
1840 static boolean_t
1841 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1842 {
1843 	spa_t *spa = zio->io_spa;
1844 
1845 	/*
1846 	 * Note: we compare the original data, not the transformed data,
1847 	 * because when zio->io_bp is an override bp, we will not have
1848 	 * pushed the I/O transforms.  That's an important optimization
1849 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1850 	 */
1851 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1852 		zio_t *lio = dde->dde_lead_zio[p];
1853 
1854 		if (lio != NULL) {
1855 			return (lio->io_orig_size != zio->io_orig_size ||
1856 			    bcmp(zio->io_orig_data, lio->io_orig_data,
1857 			    zio->io_orig_size) != 0);
1858 		}
1859 	}
1860 
1861 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1862 		ddt_phys_t *ddp = &dde->dde_phys[p];
1863 
1864 		if (ddp->ddp_phys_birth != 0) {
1865 			arc_buf_t *abuf = NULL;
1866 			uint32_t aflags = ARC_WAIT;
1867 			blkptr_t blk = *zio->io_bp;
1868 			int error;
1869 
1870 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1871 
1872 			ddt_exit(ddt);
1873 
1874 			error = arc_read_nolock(NULL, spa, &blk,
1875 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1876 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1877 			    &aflags, &zio->io_bookmark);
1878 
1879 			if (error == 0) {
1880 				if (arc_buf_size(abuf) != zio->io_orig_size ||
1881 				    bcmp(abuf->b_data, zio->io_orig_data,
1882 				    zio->io_orig_size) != 0)
1883 					error = EEXIST;
1884 				VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1885 			}
1886 
1887 			ddt_enter(ddt);
1888 			return (error != 0);
1889 		}
1890 	}
1891 
1892 	return (B_FALSE);
1893 }
1894 
1895 static void
1896 zio_ddt_child_write_ready(zio_t *zio)
1897 {
1898 	int p = zio->io_prop.zp_copies;
1899 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1900 	ddt_entry_t *dde = zio->io_private;
1901 	ddt_phys_t *ddp = &dde->dde_phys[p];
1902 	zio_t *pio;
1903 
1904 	if (zio->io_error)
1905 		return;
1906 
1907 	ddt_enter(ddt);
1908 
1909 	ASSERT(dde->dde_lead_zio[p] == zio);
1910 
1911 	ddt_phys_fill(ddp, zio->io_bp);
1912 
1913 	while ((pio = zio_walk_parents(zio)) != NULL)
1914 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1915 
1916 	ddt_exit(ddt);
1917 }
1918 
1919 static void
1920 zio_ddt_child_write_done(zio_t *zio)
1921 {
1922 	int p = zio->io_prop.zp_copies;
1923 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1924 	ddt_entry_t *dde = zio->io_private;
1925 	ddt_phys_t *ddp = &dde->dde_phys[p];
1926 
1927 	ddt_enter(ddt);
1928 
1929 	ASSERT(ddp->ddp_refcnt == 0);
1930 	ASSERT(dde->dde_lead_zio[p] == zio);
1931 	dde->dde_lead_zio[p] = NULL;
1932 
1933 	if (zio->io_error == 0) {
1934 		while (zio_walk_parents(zio) != NULL)
1935 			ddt_phys_addref(ddp);
1936 	} else {
1937 		ddt_phys_clear(ddp);
1938 	}
1939 
1940 	ddt_exit(ddt);
1941 }
1942 
1943 static void
1944 zio_ddt_ditto_write_done(zio_t *zio)
1945 {
1946 	int p = DDT_PHYS_DITTO;
1947 	zio_prop_t *zp = &zio->io_prop;
1948 	blkptr_t *bp = zio->io_bp;
1949 	ddt_t *ddt = ddt_select(zio->io_spa, bp);
1950 	ddt_entry_t *dde = zio->io_private;
1951 	ddt_phys_t *ddp = &dde->dde_phys[p];
1952 	ddt_key_t *ddk = &dde->dde_key;
1953 
1954 	ddt_enter(ddt);
1955 
1956 	ASSERT(ddp->ddp_refcnt == 0);
1957 	ASSERT(dde->dde_lead_zio[p] == zio);
1958 	dde->dde_lead_zio[p] = NULL;
1959 
1960 	if (zio->io_error == 0) {
1961 		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1962 		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1963 		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1964 		if (ddp->ddp_phys_birth != 0)
1965 			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1966 		ddt_phys_fill(ddp, bp);
1967 	}
1968 
1969 	ddt_exit(ddt);
1970 }
1971 
1972 static int
1973 zio_ddt_write(zio_t *zio)
1974 {
1975 	spa_t *spa = zio->io_spa;
1976 	blkptr_t *bp = zio->io_bp;
1977 	uint64_t txg = zio->io_txg;
1978 	zio_prop_t *zp = &zio->io_prop;
1979 	int p = zp->zp_copies;
1980 	int ditto_copies;
1981 	zio_t *cio = NULL;
1982 	zio_t *dio = NULL;
1983 	ddt_t *ddt = ddt_select(spa, bp);
1984 	ddt_entry_t *dde;
1985 	ddt_phys_t *ddp;
1986 
1987 	ASSERT(BP_GET_DEDUP(bp));
1988 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
1989 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
1990 
1991 	ddt_enter(ddt);
1992 	dde = ddt_lookup(ddt, bp, B_TRUE);
1993 	ddp = &dde->dde_phys[p];
1994 
1995 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
1996 		/*
1997 		 * If we're using a weak checksum, upgrade to a strong checksum
1998 		 * and try again.  If we're already using a strong checksum,
1999 		 * we can't resolve it, so just convert to an ordinary write.
2000 		 * (And automatically e-mail a paper to Nature?)
2001 		 */
2002 		if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2003 			zp->zp_checksum = spa_dedup_checksum(spa);
2004 			zio_pop_transforms(zio);
2005 			zio->io_stage = ZIO_STAGE_OPEN;
2006 			BP_ZERO(bp);
2007 		} else {
2008 			zp->zp_dedup = 0;
2009 		}
2010 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2011 		ddt_exit(ddt);
2012 		return (ZIO_PIPELINE_CONTINUE);
2013 	}
2014 
2015 	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2016 	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2017 
2018 	if (ditto_copies > ddt_ditto_copies_present(dde) &&
2019 	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2020 		zio_prop_t czp = *zp;
2021 
2022 		czp.zp_copies = ditto_copies;
2023 
2024 		/*
2025 		 * If we arrived here with an override bp, we won't have run
2026 		 * the transform stack, so we won't have the data we need to
2027 		 * generate a child i/o.  So, toss the override bp and restart.
2028 		 * This is safe, because using the override bp is just an
2029 		 * optimization; and it's rare, so the cost doesn't matter.
2030 		 */
2031 		if (zio->io_bp_override) {
2032 			zio_pop_transforms(zio);
2033 			zio->io_stage = ZIO_STAGE_OPEN;
2034 			zio->io_pipeline = ZIO_WRITE_PIPELINE;
2035 			zio->io_bp_override = NULL;
2036 			BP_ZERO(bp);
2037 			ddt_exit(ddt);
2038 			return (ZIO_PIPELINE_CONTINUE);
2039 		}
2040 
2041 		dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2042 		    zio->io_orig_size, &czp, NULL,
2043 		    zio_ddt_ditto_write_done, dde, zio->io_priority,
2044 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2045 
2046 		zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2047 		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2048 	}
2049 
2050 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2051 		if (ddp->ddp_phys_birth != 0)
2052 			ddt_bp_fill(ddp, bp, txg);
2053 		if (dde->dde_lead_zio[p] != NULL)
2054 			zio_add_child(zio, dde->dde_lead_zio[p]);
2055 		else
2056 			ddt_phys_addref(ddp);
2057 	} else if (zio->io_bp_override) {
2058 		ASSERT(bp->blk_birth == txg);
2059 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2060 		ddt_phys_fill(ddp, bp);
2061 		ddt_phys_addref(ddp);
2062 	} else {
2063 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2064 		    zio->io_orig_size, zp, zio_ddt_child_write_ready,
2065 		    zio_ddt_child_write_done, dde, zio->io_priority,
2066 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2067 
2068 		zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2069 		dde->dde_lead_zio[p] = cio;
2070 	}
2071 
2072 	ddt_exit(ddt);
2073 
2074 	if (cio)
2075 		zio_nowait(cio);
2076 	if (dio)
2077 		zio_nowait(dio);
2078 
2079 	return (ZIO_PIPELINE_CONTINUE);
2080 }
2081 
2082 ddt_entry_t *freedde; /* for debugging */
2083 
2084 static int
2085 zio_ddt_free(zio_t *zio)
2086 {
2087 	spa_t *spa = zio->io_spa;
2088 	blkptr_t *bp = zio->io_bp;
2089 	ddt_t *ddt = ddt_select(spa, bp);
2090 	ddt_entry_t *dde;
2091 	ddt_phys_t *ddp;
2092 
2093 	ASSERT(BP_GET_DEDUP(bp));
2094 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2095 
2096 	ddt_enter(ddt);
2097 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2098 	ddp = ddt_phys_select(dde, bp);
2099 	ddt_phys_decref(ddp);
2100 	ddt_exit(ddt);
2101 
2102 	return (ZIO_PIPELINE_CONTINUE);
2103 }
2104 
2105 /*
2106  * ==========================================================================
2107  * Allocate and free blocks
2108  * ==========================================================================
2109  */
2110 static int
2111 zio_dva_allocate(zio_t *zio)
2112 {
2113 	spa_t *spa = zio->io_spa;
2114 	metaslab_class_t *mc = spa_normal_class(spa);
2115 	blkptr_t *bp = zio->io_bp;
2116 	int error;
2117 
2118 	if (zio->io_gang_leader == NULL) {
2119 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2120 		zio->io_gang_leader = zio;
2121 	}
2122 
2123 	ASSERT(BP_IS_HOLE(bp));
2124 	ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2125 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
2126 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2127 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2128 
2129 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
2130 	    zio->io_prop.zp_copies, zio->io_txg, NULL, 0);
2131 
2132 	if (error) {
2133 		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2134 			return (zio_write_gang_block(zio));
2135 		zio->io_error = error;
2136 	}
2137 
2138 	return (ZIO_PIPELINE_CONTINUE);
2139 }
2140 
2141 static int
2142 zio_dva_free(zio_t *zio)
2143 {
2144 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2145 
2146 	return (ZIO_PIPELINE_CONTINUE);
2147 }
2148 
2149 static int
2150 zio_dva_claim(zio_t *zio)
2151 {
2152 	int error;
2153 
2154 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2155 	if (error)
2156 		zio->io_error = error;
2157 
2158 	return (ZIO_PIPELINE_CONTINUE);
2159 }
2160 
2161 /*
2162  * Undo an allocation.  This is used by zio_done() when an I/O fails
2163  * and we want to give back the block we just allocated.
2164  * This handles both normal blocks and gang blocks.
2165  */
2166 static void
2167 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2168 {
2169 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2170 	ASSERT(zio->io_bp_override == NULL);
2171 
2172 	if (!BP_IS_HOLE(bp))
2173 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2174 
2175 	if (gn != NULL) {
2176 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2177 			zio_dva_unallocate(zio, gn->gn_child[g],
2178 			    &gn->gn_gbh->zg_blkptr[g]);
2179 		}
2180 	}
2181 }
2182 
2183 /*
2184  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
2185  */
2186 int
2187 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2188     uint64_t size, boolean_t use_slog)
2189 {
2190 	int error = 1;
2191 
2192 	ASSERT(txg > spa_syncing_txg(spa));
2193 
2194 	if (use_slog)
2195 		error = metaslab_alloc(spa, spa_log_class(spa), size,
2196 		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2197 
2198 	if (error)
2199 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
2200 		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2201 
2202 	if (error == 0) {
2203 		BP_SET_LSIZE(new_bp, size);
2204 		BP_SET_PSIZE(new_bp, size);
2205 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2206 		BP_SET_CHECKSUM(new_bp,
2207 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2208 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2209 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2210 		BP_SET_LEVEL(new_bp, 0);
2211 		BP_SET_DEDUP(new_bp, 0);
2212 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2213 	}
2214 
2215 	return (error);
2216 }
2217 
2218 /*
2219  * Free an intent log block.
2220  */
2221 void
2222 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2223 {
2224 	ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2225 	ASSERT(!BP_IS_GANG(bp));
2226 
2227 	zio_free(spa, txg, bp);
2228 }
2229 
2230 /*
2231  * ==========================================================================
2232  * Read and write to physical devices
2233  * ==========================================================================
2234  */
2235 static int
2236 zio_vdev_io_start(zio_t *zio)
2237 {
2238 	vdev_t *vd = zio->io_vd;
2239 	uint64_t align;
2240 	spa_t *spa = zio->io_spa;
2241 
2242 	ASSERT(zio->io_error == 0);
2243 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2244 
2245 	if (vd == NULL) {
2246 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2247 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2248 
2249 		/*
2250 		 * The mirror_ops handle multiple DVAs in a single BP.
2251 		 */
2252 		return (vdev_mirror_ops.vdev_op_io_start(zio));
2253 	}
2254 
2255 	/*
2256 	 * We keep track of time-sensitive I/Os so that the scan thread
2257 	 * can quickly react to certain workloads.  In particular, we care
2258 	 * about non-scrubbing, top-level reads and writes with the following
2259 	 * characteristics:
2260 	 * 	- synchronous writes of user data to non-slog devices
2261 	 *	- any reads of user data
2262 	 * When these conditions are met, adjust the timestamp of spa_last_io
2263 	 * which allows the scan thread to adjust its workload accordingly.
2264 	 */
2265 	if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2266 	    vd == vd->vdev_top && !vd->vdev_islog &&
2267 	    zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2268 	    zio->io_txg != spa_syncing_txg(spa)) {
2269 		uint64_t old = spa->spa_last_io;
2270 		uint64_t new = ddi_get_lbolt64();
2271 		if (old != new)
2272 			(void) atomic_cas_64(&spa->spa_last_io, old, new);
2273 	}
2274 
2275 	align = 1ULL << vd->vdev_top->vdev_ashift;
2276 
2277 	if (P2PHASE(zio->io_size, align) != 0) {
2278 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
2279 		char *abuf = zio_buf_alloc(asize);
2280 		ASSERT(vd == vd->vdev_top);
2281 		if (zio->io_type == ZIO_TYPE_WRITE) {
2282 			bcopy(zio->io_data, abuf, zio->io_size);
2283 			bzero(abuf + zio->io_size, asize - zio->io_size);
2284 		}
2285 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2286 	}
2287 
2288 	ASSERT(P2PHASE(zio->io_offset, align) == 0);
2289 	ASSERT(P2PHASE(zio->io_size, align) == 0);
2290 	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2291 
2292 	/*
2293 	 * If this is a repair I/O, and there's no self-healing involved --
2294 	 * that is, we're just resilvering what we expect to resilver --
2295 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2296 	 * This prevents spurious resilvering with nested replication.
2297 	 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2298 	 * A is out of date, we'll read from C+D, then use the data to
2299 	 * resilver A+B -- but we don't actually want to resilver B, just A.
2300 	 * The top-level mirror has no way to know this, so instead we just
2301 	 * discard unnecessary repairs as we work our way down the vdev tree.
2302 	 * The same logic applies to any form of nested replication:
2303 	 * ditto + mirror, RAID-Z + replacing, etc.  This covers them all.
2304 	 */
2305 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2306 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2307 	    zio->io_txg != 0 &&	/* not a delegated i/o */
2308 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2309 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2310 		zio_vdev_io_bypass(zio);
2311 		return (ZIO_PIPELINE_CONTINUE);
2312 	}
2313 
2314 	if (vd->vdev_ops->vdev_op_leaf &&
2315 	    (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2316 
2317 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2318 			return (ZIO_PIPELINE_CONTINUE);
2319 
2320 		if ((zio = vdev_queue_io(zio)) == NULL)
2321 			return (ZIO_PIPELINE_STOP);
2322 
2323 		if (!vdev_accessible(vd, zio)) {
2324 			zio->io_error = ENXIO;
2325 			zio_interrupt(zio);
2326 			return (ZIO_PIPELINE_STOP);
2327 		}
2328 	}
2329 
2330 	return (vd->vdev_ops->vdev_op_io_start(zio));
2331 }
2332 
2333 static int
2334 zio_vdev_io_done(zio_t *zio)
2335 {
2336 	vdev_t *vd = zio->io_vd;
2337 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2338 	boolean_t unexpected_error = B_FALSE;
2339 
2340 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2341 		return (ZIO_PIPELINE_STOP);
2342 
2343 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2344 
2345 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2346 
2347 		vdev_queue_io_done(zio);
2348 
2349 		if (zio->io_type == ZIO_TYPE_WRITE)
2350 			vdev_cache_write(zio);
2351 
2352 		if (zio_injection_enabled && zio->io_error == 0)
2353 			zio->io_error = zio_handle_device_injection(vd,
2354 			    zio, EIO);
2355 
2356 		if (zio_injection_enabled && zio->io_error == 0)
2357 			zio->io_error = zio_handle_label_injection(zio, EIO);
2358 
2359 		if (zio->io_error) {
2360 			if (!vdev_accessible(vd, zio)) {
2361 				zio->io_error = ENXIO;
2362 			} else {
2363 				unexpected_error = B_TRUE;
2364 			}
2365 		}
2366 	}
2367 
2368 	ops->vdev_op_io_done(zio);
2369 
2370 	if (unexpected_error)
2371 		VERIFY(vdev_probe(vd, zio) == NULL);
2372 
2373 	return (ZIO_PIPELINE_CONTINUE);
2374 }
2375 
2376 /*
2377  * For non-raidz ZIOs, we can just copy aside the bad data read from the
2378  * disk, and use that to finish the checksum ereport later.
2379  */
2380 static void
2381 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2382     const void *good_buf)
2383 {
2384 	/* no processing needed */
2385 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2386 }
2387 
2388 /*ARGSUSED*/
2389 void
2390 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2391 {
2392 	void *buf = zio_buf_alloc(zio->io_size);
2393 
2394 	bcopy(zio->io_data, buf, zio->io_size);
2395 
2396 	zcr->zcr_cbinfo = zio->io_size;
2397 	zcr->zcr_cbdata = buf;
2398 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
2399 	zcr->zcr_free = zio_buf_free;
2400 }
2401 
2402 static int
2403 zio_vdev_io_assess(zio_t *zio)
2404 {
2405 	vdev_t *vd = zio->io_vd;
2406 
2407 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2408 		return (ZIO_PIPELINE_STOP);
2409 
2410 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2411 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2412 
2413 	if (zio->io_vsd != NULL) {
2414 		zio->io_vsd_ops->vsd_free(zio);
2415 		zio->io_vsd = NULL;
2416 	}
2417 
2418 	if (zio_injection_enabled && zio->io_error == 0)
2419 		zio->io_error = zio_handle_fault_injection(zio, EIO);
2420 
2421 	/*
2422 	 * If the I/O failed, determine whether we should attempt to retry it.
2423 	 *
2424 	 * On retry, we cut in line in the issue queue, since we don't want
2425 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2426 	 */
2427 	if (zio->io_error && vd == NULL &&
2428 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2429 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
2430 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
2431 		zio->io_error = 0;
2432 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
2433 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2434 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2435 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2436 		    zio_requeue_io_start_cut_in_line);
2437 		return (ZIO_PIPELINE_STOP);
2438 	}
2439 
2440 	/*
2441 	 * If we got an error on a leaf device, convert it to ENXIO
2442 	 * if the device is not accessible at all.
2443 	 */
2444 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2445 	    !vdev_accessible(vd, zio))
2446 		zio->io_error = ENXIO;
2447 
2448 	/*
2449 	 * If we can't write to an interior vdev (mirror or RAID-Z),
2450 	 * set vdev_cant_write so that we stop trying to allocate from it.
2451 	 */
2452 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2453 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2454 		vd->vdev_cant_write = B_TRUE;
2455 
2456 	if (zio->io_error)
2457 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2458 
2459 	return (ZIO_PIPELINE_CONTINUE);
2460 }
2461 
2462 void
2463 zio_vdev_io_reissue(zio_t *zio)
2464 {
2465 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2466 	ASSERT(zio->io_error == 0);
2467 
2468 	zio->io_stage >>= 1;
2469 }
2470 
2471 void
2472 zio_vdev_io_redone(zio_t *zio)
2473 {
2474 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2475 
2476 	zio->io_stage >>= 1;
2477 }
2478 
2479 void
2480 zio_vdev_io_bypass(zio_t *zio)
2481 {
2482 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2483 	ASSERT(zio->io_error == 0);
2484 
2485 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2486 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2487 }
2488 
2489 /*
2490  * ==========================================================================
2491  * Generate and verify checksums
2492  * ==========================================================================
2493  */
2494 static int
2495 zio_checksum_generate(zio_t *zio)
2496 {
2497 	blkptr_t *bp = zio->io_bp;
2498 	enum zio_checksum checksum;
2499 
2500 	if (bp == NULL) {
2501 		/*
2502 		 * This is zio_write_phys().
2503 		 * We're either generating a label checksum, or none at all.
2504 		 */
2505 		checksum = zio->io_prop.zp_checksum;
2506 
2507 		if (checksum == ZIO_CHECKSUM_OFF)
2508 			return (ZIO_PIPELINE_CONTINUE);
2509 
2510 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2511 	} else {
2512 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2513 			ASSERT(!IO_IS_ALLOCATING(zio));
2514 			checksum = ZIO_CHECKSUM_GANG_HEADER;
2515 		} else {
2516 			checksum = BP_GET_CHECKSUM(bp);
2517 		}
2518 	}
2519 
2520 	zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2521 
2522 	return (ZIO_PIPELINE_CONTINUE);
2523 }
2524 
2525 static int
2526 zio_checksum_verify(zio_t *zio)
2527 {
2528 	zio_bad_cksum_t info;
2529 	blkptr_t *bp = zio->io_bp;
2530 	int error;
2531 
2532 	ASSERT(zio->io_vd != NULL);
2533 
2534 	if (bp == NULL) {
2535 		/*
2536 		 * This is zio_read_phys().
2537 		 * We're either verifying a label checksum, or nothing at all.
2538 		 */
2539 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2540 			return (ZIO_PIPELINE_CONTINUE);
2541 
2542 		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2543 	}
2544 
2545 	if ((error = zio_checksum_error(zio, &info)) != 0) {
2546 		zio->io_error = error;
2547 		if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2548 			zfs_ereport_start_checksum(zio->io_spa,
2549 			    zio->io_vd, zio, zio->io_offset,
2550 			    zio->io_size, NULL, &info);
2551 		}
2552 	}
2553 
2554 	return (ZIO_PIPELINE_CONTINUE);
2555 }
2556 
2557 /*
2558  * Called by RAID-Z to ensure we don't compute the checksum twice.
2559  */
2560 void
2561 zio_checksum_verified(zio_t *zio)
2562 {
2563 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2564 }
2565 
2566 /*
2567  * ==========================================================================
2568  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2569  * An error of 0 indictes success.  ENXIO indicates whole-device failure,
2570  * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
2571  * indicate errors that are specific to one I/O, and most likely permanent.
2572  * Any other error is presumed to be worse because we weren't expecting it.
2573  * ==========================================================================
2574  */
2575 int
2576 zio_worst_error(int e1, int e2)
2577 {
2578 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2579 	int r1, r2;
2580 
2581 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2582 		if (e1 == zio_error_rank[r1])
2583 			break;
2584 
2585 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2586 		if (e2 == zio_error_rank[r2])
2587 			break;
2588 
2589 	return (r1 > r2 ? e1 : e2);
2590 }
2591 
2592 /*
2593  * ==========================================================================
2594  * I/O completion
2595  * ==========================================================================
2596  */
2597 static int
2598 zio_ready(zio_t *zio)
2599 {
2600 	blkptr_t *bp = zio->io_bp;
2601 	zio_t *pio, *pio_next;
2602 
2603 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2604 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2605 		return (ZIO_PIPELINE_STOP);
2606 
2607 	if (zio->io_ready) {
2608 		ASSERT(IO_IS_ALLOCATING(zio));
2609 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2610 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2611 
2612 		zio->io_ready(zio);
2613 	}
2614 
2615 	if (bp != NULL && bp != &zio->io_bp_copy)
2616 		zio->io_bp_copy = *bp;
2617 
2618 	if (zio->io_error)
2619 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2620 
2621 	mutex_enter(&zio->io_lock);
2622 	zio->io_state[ZIO_WAIT_READY] = 1;
2623 	pio = zio_walk_parents(zio);
2624 	mutex_exit(&zio->io_lock);
2625 
2626 	/*
2627 	 * As we notify zio's parents, new parents could be added.
2628 	 * New parents go to the head of zio's io_parent_list, however,
2629 	 * so we will (correctly) not notify them.  The remainder of zio's
2630 	 * io_parent_list, from 'pio_next' onward, cannot change because
2631 	 * all parents must wait for us to be done before they can be done.
2632 	 */
2633 	for (; pio != NULL; pio = pio_next) {
2634 		pio_next = zio_walk_parents(zio);
2635 		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2636 	}
2637 
2638 	if (zio->io_flags & ZIO_FLAG_NODATA) {
2639 		if (BP_IS_GANG(bp)) {
2640 			zio->io_flags &= ~ZIO_FLAG_NODATA;
2641 		} else {
2642 			ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2643 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2644 		}
2645 	}
2646 
2647 	if (zio_injection_enabled &&
2648 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
2649 		zio_handle_ignored_writes(zio);
2650 
2651 	return (ZIO_PIPELINE_CONTINUE);
2652 }
2653 
2654 static int
2655 zio_done(zio_t *zio)
2656 {
2657 	spa_t *spa = zio->io_spa;
2658 	zio_t *lio = zio->io_logical;
2659 	blkptr_t *bp = zio->io_bp;
2660 	vdev_t *vd = zio->io_vd;
2661 	uint64_t psize = zio->io_size;
2662 	zio_t *pio, *pio_next;
2663 
2664 	/*
2665 	 * If our children haven't all completed,
2666 	 * wait for them and then repeat this pipeline stage.
2667 	 */
2668 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2669 	    zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2670 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2671 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2672 		return (ZIO_PIPELINE_STOP);
2673 
2674 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2675 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2676 			ASSERT(zio->io_children[c][w] == 0);
2677 
2678 	if (bp != NULL) {
2679 		ASSERT(bp->blk_pad[0] == 0);
2680 		ASSERT(bp->blk_pad[1] == 0);
2681 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2682 		    (bp == zio_unique_parent(zio)->io_bp));
2683 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2684 		    zio->io_bp_override == NULL &&
2685 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2686 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
2687 			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2688 			ASSERT(BP_COUNT_GANG(bp) == 0 ||
2689 			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2690 		}
2691 	}
2692 
2693 	/*
2694 	 * If there were child vdev/gang/ddt errors, they apply to us now.
2695 	 */
2696 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2697 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2698 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2699 
2700 	/*
2701 	 * If the I/O on the transformed data was successful, generate any
2702 	 * checksum reports now while we still have the transformed data.
2703 	 */
2704 	if (zio->io_error == 0) {
2705 		while (zio->io_cksum_report != NULL) {
2706 			zio_cksum_report_t *zcr = zio->io_cksum_report;
2707 			uint64_t align = zcr->zcr_align;
2708 			uint64_t asize = P2ROUNDUP(psize, align);
2709 			char *abuf = zio->io_data;
2710 
2711 			if (asize != psize) {
2712 				abuf = zio_buf_alloc(asize);
2713 				bcopy(zio->io_data, abuf, psize);
2714 				bzero(abuf + psize, asize - psize);
2715 			}
2716 
2717 			zio->io_cksum_report = zcr->zcr_next;
2718 			zcr->zcr_next = NULL;
2719 			zcr->zcr_finish(zcr, abuf);
2720 			zfs_ereport_free_checksum(zcr);
2721 
2722 			if (asize != psize)
2723 				zio_buf_free(abuf, asize);
2724 		}
2725 	}
2726 
2727 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
2728 
2729 	vdev_stat_update(zio, psize);
2730 
2731 	if (zio->io_error) {
2732 		/*
2733 		 * If this I/O is attached to a particular vdev,
2734 		 * generate an error message describing the I/O failure
2735 		 * at the block level.  We ignore these errors if the
2736 		 * device is currently unavailable.
2737 		 */
2738 		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2739 			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2740 
2741 		if ((zio->io_error == EIO || !(zio->io_flags &
2742 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2743 		    zio == lio) {
2744 			/*
2745 			 * For logical I/O requests, tell the SPA to log the
2746 			 * error and generate a logical data ereport.
2747 			 */
2748 			spa_log_error(spa, zio);
2749 			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2750 			    0, 0);
2751 		}
2752 	}
2753 
2754 	if (zio->io_error && zio == lio) {
2755 		/*
2756 		 * Determine whether zio should be reexecuted.  This will
2757 		 * propagate all the way to the root via zio_notify_parent().
2758 		 */
2759 		ASSERT(vd == NULL && bp != NULL);
2760 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2761 
2762 		if (IO_IS_ALLOCATING(zio) &&
2763 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2764 			if (zio->io_error != ENOSPC)
2765 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2766 			else
2767 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2768 		}
2769 
2770 		if ((zio->io_type == ZIO_TYPE_READ ||
2771 		    zio->io_type == ZIO_TYPE_FREE) &&
2772 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2773 		    zio->io_error == ENXIO &&
2774 		    spa_load_state(spa) == SPA_LOAD_NONE &&
2775 		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2776 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2777 
2778 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2779 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2780 
2781 		/*
2782 		 * Here is a possibly good place to attempt to do
2783 		 * either combinatorial reconstruction or error correction
2784 		 * based on checksums.  It also might be a good place
2785 		 * to send out preliminary ereports before we suspend
2786 		 * processing.
2787 		 */
2788 	}
2789 
2790 	/*
2791 	 * If there were logical child errors, they apply to us now.
2792 	 * We defer this until now to avoid conflating logical child
2793 	 * errors with errors that happened to the zio itself when
2794 	 * updating vdev stats and reporting FMA events above.
2795 	 */
2796 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2797 
2798 	if ((zio->io_error || zio->io_reexecute) &&
2799 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2800 	    !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2801 		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2802 
2803 	zio_gang_tree_free(&zio->io_gang_tree);
2804 
2805 	/*
2806 	 * Godfather I/Os should never suspend.
2807 	 */
2808 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2809 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2810 		zio->io_reexecute = 0;
2811 
2812 	if (zio->io_reexecute) {
2813 		/*
2814 		 * This is a logical I/O that wants to reexecute.
2815 		 *
2816 		 * Reexecute is top-down.  When an i/o fails, if it's not
2817 		 * the root, it simply notifies its parent and sticks around.
2818 		 * The parent, seeing that it still has children in zio_done(),
2819 		 * does the same.  This percolates all the way up to the root.
2820 		 * The root i/o will reexecute or suspend the entire tree.
2821 		 *
2822 		 * This approach ensures that zio_reexecute() honors
2823 		 * all the original i/o dependency relationships, e.g.
2824 		 * parents not executing until children are ready.
2825 		 */
2826 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2827 
2828 		zio->io_gang_leader = NULL;
2829 
2830 		mutex_enter(&zio->io_lock);
2831 		zio->io_state[ZIO_WAIT_DONE] = 1;
2832 		mutex_exit(&zio->io_lock);
2833 
2834 		/*
2835 		 * "The Godfather" I/O monitors its children but is
2836 		 * not a true parent to them. It will track them through
2837 		 * the pipeline but severs its ties whenever they get into
2838 		 * trouble (e.g. suspended). This allows "The Godfather"
2839 		 * I/O to return status without blocking.
2840 		 */
2841 		for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2842 			zio_link_t *zl = zio->io_walk_link;
2843 			pio_next = zio_walk_parents(zio);
2844 
2845 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2846 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2847 				zio_remove_child(pio, zio, zl);
2848 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2849 			}
2850 		}
2851 
2852 		if ((pio = zio_unique_parent(zio)) != NULL) {
2853 			/*
2854 			 * We're not a root i/o, so there's nothing to do
2855 			 * but notify our parent.  Don't propagate errors
2856 			 * upward since we haven't permanently failed yet.
2857 			 */
2858 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2859 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2860 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2861 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2862 			/*
2863 			 * We'd fail again if we reexecuted now, so suspend
2864 			 * until conditions improve (e.g. device comes online).
2865 			 */
2866 			zio_suspend(spa, zio);
2867 		} else {
2868 			/*
2869 			 * Reexecution is potentially a huge amount of work.
2870 			 * Hand it off to the otherwise-unused claim taskq.
2871 			 */
2872 			(void) taskq_dispatch(
2873 			    spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2874 			    (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2875 		}
2876 		return (ZIO_PIPELINE_STOP);
2877 	}
2878 
2879 	ASSERT(zio->io_child_count == 0);
2880 	ASSERT(zio->io_reexecute == 0);
2881 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2882 
2883 	/*
2884 	 * Report any checksum errors, since the I/O is complete.
2885 	 */
2886 	while (zio->io_cksum_report != NULL) {
2887 		zio_cksum_report_t *zcr = zio->io_cksum_report;
2888 		zio->io_cksum_report = zcr->zcr_next;
2889 		zcr->zcr_next = NULL;
2890 		zcr->zcr_finish(zcr, NULL);
2891 		zfs_ereport_free_checksum(zcr);
2892 	}
2893 
2894 	/*
2895 	 * It is the responsibility of the done callback to ensure that this
2896 	 * particular zio is no longer discoverable for adoption, and as
2897 	 * such, cannot acquire any new parents.
2898 	 */
2899 	if (zio->io_done)
2900 		zio->io_done(zio);
2901 
2902 	mutex_enter(&zio->io_lock);
2903 	zio->io_state[ZIO_WAIT_DONE] = 1;
2904 	mutex_exit(&zio->io_lock);
2905 
2906 	for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2907 		zio_link_t *zl = zio->io_walk_link;
2908 		pio_next = zio_walk_parents(zio);
2909 		zio_remove_child(pio, zio, zl);
2910 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2911 	}
2912 
2913 	if (zio->io_waiter != NULL) {
2914 		mutex_enter(&zio->io_lock);
2915 		zio->io_executor = NULL;
2916 		cv_broadcast(&zio->io_cv);
2917 		mutex_exit(&zio->io_lock);
2918 	} else {
2919 		zio_destroy(zio);
2920 	}
2921 
2922 	return (ZIO_PIPELINE_STOP);
2923 }
2924 
2925 /*
2926  * ==========================================================================
2927  * I/O pipeline definition
2928  * ==========================================================================
2929  */
2930 static zio_pipe_stage_t *zio_pipeline[] = {
2931 	NULL,
2932 	zio_read_bp_init,
2933 	zio_free_bp_init,
2934 	zio_issue_async,
2935 	zio_write_bp_init,
2936 	zio_checksum_generate,
2937 	zio_ddt_read_start,
2938 	zio_ddt_read_done,
2939 	zio_ddt_write,
2940 	zio_ddt_free,
2941 	zio_gang_assemble,
2942 	zio_gang_issue,
2943 	zio_dva_allocate,
2944 	zio_dva_free,
2945 	zio_dva_claim,
2946 	zio_ready,
2947 	zio_vdev_io_start,
2948 	zio_vdev_io_done,
2949 	zio_vdev_io_assess,
2950 	zio_checksum_verify,
2951 	zio_done
2952 };
2953