xref: /titanic_44/usr/src/uts/common/fs/zfs/zio.c (revision ad09f8b827db90c9a0093f0b6382803fa64a5fd1)
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 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/zfs_context.h>
27 #include <sys/fm/fs/zfs.h>
28 #include <sys/spa.h>
29 #include <sys/txg.h>
30 #include <sys/spa_impl.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/zio_impl.h>
33 #include <sys/zio_compress.h>
34 #include <sys/zio_checksum.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/arc.h>
37 #include <sys/ddt.h>
38 
39 /*
40  * ==========================================================================
41  * I/O priority table
42  * ==========================================================================
43  */
44 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
45 	0,	/* ZIO_PRIORITY_NOW		*/
46 	0,	/* ZIO_PRIORITY_SYNC_READ	*/
47 	0,	/* ZIO_PRIORITY_SYNC_WRITE	*/
48 	6,	/* ZIO_PRIORITY_ASYNC_READ	*/
49 	4,	/* ZIO_PRIORITY_ASYNC_WRITE	*/
50 	4,	/* ZIO_PRIORITY_FREE		*/
51 	0,	/* ZIO_PRIORITY_CACHE_FILL	*/
52 	0,	/* ZIO_PRIORITY_LOG_WRITE	*/
53 	10,	/* ZIO_PRIORITY_RESILVER	*/
54 	20,	/* ZIO_PRIORITY_SCRUB		*/
55 };
56 
57 /*
58  * ==========================================================================
59  * I/O type descriptions
60  * ==========================================================================
61  */
62 char *zio_type_name[ZIO_TYPES] = {
63 	"null", "read", "write", "free", "claim", "ioctl" };
64 
65 /*
66  * ==========================================================================
67  * I/O kmem caches
68  * ==========================================================================
69  */
70 kmem_cache_t *zio_cache;
71 kmem_cache_t *zio_link_cache;
72 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
73 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
74 
75 #ifdef _KERNEL
76 extern vmem_t *zio_alloc_arena;
77 #endif
78 
79 /*
80  * An allocating zio is one that either currently has the DVA allocate
81  * stage set or will have it later in its lifetime.
82  */
83 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
84 
85 #ifdef ZFS_DEBUG
86 int zio_buf_debug_limit = 16384;
87 #else
88 int zio_buf_debug_limit = 0;
89 #endif
90 
91 void
92 zio_init(void)
93 {
94 	size_t c;
95 	vmem_t *data_alloc_arena = NULL;
96 
97 #ifdef _KERNEL
98 	data_alloc_arena = zio_alloc_arena;
99 #endif
100 	zio_cache = kmem_cache_create("zio_cache",
101 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
102 	zio_link_cache = kmem_cache_create("zio_link_cache",
103 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
104 
105 	/*
106 	 * For small buffers, we want a cache for each multiple of
107 	 * SPA_MINBLOCKSIZE.  For medium-size buffers, we want a cache
108 	 * for each quarter-power of 2.  For large buffers, we want
109 	 * a cache for each multiple of PAGESIZE.
110 	 */
111 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
112 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
113 		size_t p2 = size;
114 		size_t align = 0;
115 
116 		while (p2 & (p2 - 1))
117 			p2 &= p2 - 1;
118 
119 		if (size <= 4 * SPA_MINBLOCKSIZE) {
120 			align = SPA_MINBLOCKSIZE;
121 		} else if (P2PHASE(size, PAGESIZE) == 0) {
122 			align = PAGESIZE;
123 		} else if (P2PHASE(size, p2 >> 2) == 0) {
124 			align = p2 >> 2;
125 		}
126 
127 		if (align != 0) {
128 			char name[36];
129 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
130 			zio_buf_cache[c] = kmem_cache_create(name, size,
131 			    align, NULL, NULL, NULL, NULL, NULL,
132 			    size > zio_buf_debug_limit ? KMC_NODEBUG : 0);
133 
134 			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
135 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
136 			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
137 			    size > zio_buf_debug_limit ? KMC_NODEBUG : 0);
138 		}
139 	}
140 
141 	while (--c != 0) {
142 		ASSERT(zio_buf_cache[c] != NULL);
143 		if (zio_buf_cache[c - 1] == NULL)
144 			zio_buf_cache[c - 1] = zio_buf_cache[c];
145 
146 		ASSERT(zio_data_buf_cache[c] != NULL);
147 		if (zio_data_buf_cache[c - 1] == NULL)
148 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
149 	}
150 
151 	zio_inject_init();
152 }
153 
154 void
155 zio_fini(void)
156 {
157 	size_t c;
158 	kmem_cache_t *last_cache = NULL;
159 	kmem_cache_t *last_data_cache = NULL;
160 
161 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
162 		if (zio_buf_cache[c] != last_cache) {
163 			last_cache = zio_buf_cache[c];
164 			kmem_cache_destroy(zio_buf_cache[c]);
165 		}
166 		zio_buf_cache[c] = NULL;
167 
168 		if (zio_data_buf_cache[c] != last_data_cache) {
169 			last_data_cache = zio_data_buf_cache[c];
170 			kmem_cache_destroy(zio_data_buf_cache[c]);
171 		}
172 		zio_data_buf_cache[c] = NULL;
173 	}
174 
175 	kmem_cache_destroy(zio_link_cache);
176 	kmem_cache_destroy(zio_cache);
177 
178 	zio_inject_fini();
179 }
180 
181 /*
182  * ==========================================================================
183  * Allocate and free I/O buffers
184  * ==========================================================================
185  */
186 
187 /*
188  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
189  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
190  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
191  * excess / transient data in-core during a crashdump.
192  */
193 void *
194 zio_buf_alloc(size_t size)
195 {
196 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
197 
198 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
199 
200 	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
201 }
202 
203 /*
204  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
205  * crashdump if the kernel panics.  This exists so that we will limit the amount
206  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
207  * of kernel heap dumped to disk when the kernel panics)
208  */
209 void *
210 zio_data_buf_alloc(size_t size)
211 {
212 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
213 
214 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
215 
216 	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
217 }
218 
219 void
220 zio_buf_free(void *buf, size_t size)
221 {
222 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
223 
224 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
225 
226 	kmem_cache_free(zio_buf_cache[c], buf);
227 }
228 
229 void
230 zio_data_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_data_buf_cache[c], buf);
237 }
238 
239 /*
240  * ==========================================================================
241  * Push and pop I/O transform buffers
242  * ==========================================================================
243  */
244 static void
245 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
246 	zio_transform_func_t *transform)
247 {
248 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
249 
250 	zt->zt_orig_data = zio->io_data;
251 	zt->zt_orig_size = zio->io_size;
252 	zt->zt_bufsize = bufsize;
253 	zt->zt_transform = transform;
254 
255 	zt->zt_next = zio->io_transform_stack;
256 	zio->io_transform_stack = zt;
257 
258 	zio->io_data = data;
259 	zio->io_size = size;
260 }
261 
262 static void
263 zio_pop_transforms(zio_t *zio)
264 {
265 	zio_transform_t *zt;
266 
267 	while ((zt = zio->io_transform_stack) != NULL) {
268 		if (zt->zt_transform != NULL)
269 			zt->zt_transform(zio,
270 			    zt->zt_orig_data, zt->zt_orig_size);
271 
272 		if (zt->zt_bufsize != 0)
273 			zio_buf_free(zio->io_data, zt->zt_bufsize);
274 
275 		zio->io_data = zt->zt_orig_data;
276 		zio->io_size = zt->zt_orig_size;
277 		zio->io_transform_stack = zt->zt_next;
278 
279 		kmem_free(zt, sizeof (zio_transform_t));
280 	}
281 }
282 
283 /*
284  * ==========================================================================
285  * I/O transform callbacks for subblocks and decompression
286  * ==========================================================================
287  */
288 static void
289 zio_subblock(zio_t *zio, void *data, uint64_t size)
290 {
291 	ASSERT(zio->io_size > size);
292 
293 	if (zio->io_type == ZIO_TYPE_READ)
294 		bcopy(zio->io_data, data, size);
295 }
296 
297 static void
298 zio_decompress(zio_t *zio, void *data, uint64_t size)
299 {
300 	if (zio->io_error == 0 &&
301 	    zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
302 	    zio->io_data, data, zio->io_size, size) != 0)
303 		zio->io_error = EIO;
304 }
305 
306 /*
307  * ==========================================================================
308  * I/O parent/child relationships and pipeline interlocks
309  * ==========================================================================
310  */
311 /*
312  * NOTE - Callers to zio_walk_parents() and zio_walk_children must
313  *        continue calling these functions until they return NULL.
314  *        Otherwise, the next caller will pick up the list walk in
315  *        some indeterminate state.  (Otherwise every caller would
316  *        have to pass in a cookie to keep the state represented by
317  *        io_walk_link, which gets annoying.)
318  */
319 zio_t *
320 zio_walk_parents(zio_t *cio)
321 {
322 	zio_link_t *zl = cio->io_walk_link;
323 	list_t *pl = &cio->io_parent_list;
324 
325 	zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
326 	cio->io_walk_link = zl;
327 
328 	if (zl == NULL)
329 		return (NULL);
330 
331 	ASSERT(zl->zl_child == cio);
332 	return (zl->zl_parent);
333 }
334 
335 zio_t *
336 zio_walk_children(zio_t *pio)
337 {
338 	zio_link_t *zl = pio->io_walk_link;
339 	list_t *cl = &pio->io_child_list;
340 
341 	zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
342 	pio->io_walk_link = zl;
343 
344 	if (zl == NULL)
345 		return (NULL);
346 
347 	ASSERT(zl->zl_parent == pio);
348 	return (zl->zl_child);
349 }
350 
351 zio_t *
352 zio_unique_parent(zio_t *cio)
353 {
354 	zio_t *pio = zio_walk_parents(cio);
355 
356 	VERIFY(zio_walk_parents(cio) == NULL);
357 	return (pio);
358 }
359 
360 void
361 zio_add_child(zio_t *pio, zio_t *cio)
362 {
363 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
364 
365 	/*
366 	 * Logical I/Os can have logical, gang, or vdev children.
367 	 * Gang I/Os can have gang or vdev children.
368 	 * Vdev I/Os can only have vdev children.
369 	 * The following ASSERT captures all of these constraints.
370 	 */
371 	ASSERT(cio->io_child_type <= pio->io_child_type);
372 
373 	zl->zl_parent = pio;
374 	zl->zl_child = cio;
375 
376 	mutex_enter(&cio->io_lock);
377 	mutex_enter(&pio->io_lock);
378 
379 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
380 
381 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
382 		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
383 
384 	list_insert_head(&pio->io_child_list, zl);
385 	list_insert_head(&cio->io_parent_list, zl);
386 
387 	pio->io_child_count++;
388 	cio->io_parent_count++;
389 
390 	mutex_exit(&pio->io_lock);
391 	mutex_exit(&cio->io_lock);
392 }
393 
394 static void
395 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
396 {
397 	ASSERT(zl->zl_parent == pio);
398 	ASSERT(zl->zl_child == cio);
399 
400 	mutex_enter(&cio->io_lock);
401 	mutex_enter(&pio->io_lock);
402 
403 	list_remove(&pio->io_child_list, zl);
404 	list_remove(&cio->io_parent_list, zl);
405 
406 	pio->io_child_count--;
407 	cio->io_parent_count--;
408 
409 	mutex_exit(&pio->io_lock);
410 	mutex_exit(&cio->io_lock);
411 
412 	kmem_cache_free(zio_link_cache, zl);
413 }
414 
415 static boolean_t
416 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
417 {
418 	uint64_t *countp = &zio->io_children[child][wait];
419 	boolean_t waiting = B_FALSE;
420 
421 	mutex_enter(&zio->io_lock);
422 	ASSERT(zio->io_stall == NULL);
423 	if (*countp != 0) {
424 		zio->io_stage >>= 1;
425 		zio->io_stall = countp;
426 		waiting = B_TRUE;
427 	}
428 	mutex_exit(&zio->io_lock);
429 
430 	return (waiting);
431 }
432 
433 static void
434 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
435 {
436 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
437 	int *errorp = &pio->io_child_error[zio->io_child_type];
438 
439 	mutex_enter(&pio->io_lock);
440 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
441 		*errorp = zio_worst_error(*errorp, zio->io_error);
442 	pio->io_reexecute |= zio->io_reexecute;
443 	ASSERT3U(*countp, >, 0);
444 	if (--*countp == 0 && pio->io_stall == countp) {
445 		pio->io_stall = NULL;
446 		mutex_exit(&pio->io_lock);
447 		zio_execute(pio);
448 	} else {
449 		mutex_exit(&pio->io_lock);
450 	}
451 }
452 
453 static void
454 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
455 {
456 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
457 		zio->io_error = zio->io_child_error[c];
458 }
459 
460 /*
461  * ==========================================================================
462  * Create the various types of I/O (read, write, free, etc)
463  * ==========================================================================
464  */
465 static zio_t *
466 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
467     void *data, uint64_t size, zio_done_func_t *done, void *private,
468     zio_type_t type, int priority, enum zio_flag flags,
469     vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
470     enum zio_stage stage, enum zio_stage pipeline)
471 {
472 	zio_t *zio;
473 
474 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
475 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
476 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
477 
478 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
479 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
480 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
481 
482 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
483 	bzero(zio, sizeof (zio_t));
484 
485 	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
486 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
487 
488 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
489 	    offsetof(zio_link_t, zl_parent_node));
490 	list_create(&zio->io_child_list, sizeof (zio_link_t),
491 	    offsetof(zio_link_t, zl_child_node));
492 
493 	if (vd != NULL)
494 		zio->io_child_type = ZIO_CHILD_VDEV;
495 	else if (flags & ZIO_FLAG_GANG_CHILD)
496 		zio->io_child_type = ZIO_CHILD_GANG;
497 	else if (flags & ZIO_FLAG_DDT_CHILD)
498 		zio->io_child_type = ZIO_CHILD_DDT;
499 	else
500 		zio->io_child_type = ZIO_CHILD_LOGICAL;
501 
502 	if (bp != NULL) {
503 		zio->io_bp = (blkptr_t *)bp;
504 		zio->io_bp_copy = *bp;
505 		zio->io_bp_orig = *bp;
506 		if (type != ZIO_TYPE_WRITE ||
507 		    zio->io_child_type == ZIO_CHILD_DDT)
508 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
509 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
510 			zio->io_logical = zio;
511 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
512 			pipeline |= ZIO_GANG_STAGES;
513 	}
514 
515 	zio->io_spa = spa;
516 	zio->io_txg = txg;
517 	zio->io_done = done;
518 	zio->io_private = private;
519 	zio->io_type = type;
520 	zio->io_priority = priority;
521 	zio->io_vd = vd;
522 	zio->io_offset = offset;
523 	zio->io_orig_data = zio->io_data = data;
524 	zio->io_orig_size = zio->io_size = size;
525 	zio->io_orig_flags = zio->io_flags = flags;
526 	zio->io_orig_stage = zio->io_stage = stage;
527 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
528 
529 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
530 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
531 
532 	if (zb != NULL)
533 		zio->io_bookmark = *zb;
534 
535 	if (pio != NULL) {
536 		if (zio->io_logical == NULL)
537 			zio->io_logical = pio->io_logical;
538 		if (zio->io_child_type == ZIO_CHILD_GANG)
539 			zio->io_gang_leader = pio->io_gang_leader;
540 		zio_add_child(pio, zio);
541 	}
542 
543 	return (zio);
544 }
545 
546 static void
547 zio_destroy(zio_t *zio)
548 {
549 	list_destroy(&zio->io_parent_list);
550 	list_destroy(&zio->io_child_list);
551 	mutex_destroy(&zio->io_lock);
552 	cv_destroy(&zio->io_cv);
553 	kmem_cache_free(zio_cache, zio);
554 }
555 
556 zio_t *
557 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
558     void *private, enum zio_flag flags)
559 {
560 	zio_t *zio;
561 
562 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
563 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
564 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
565 
566 	return (zio);
567 }
568 
569 zio_t *
570 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
571 {
572 	return (zio_null(NULL, spa, NULL, done, private, flags));
573 }
574 
575 zio_t *
576 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
577     void *data, uint64_t size, zio_done_func_t *done, void *private,
578     int priority, enum zio_flag flags, const zbookmark_t *zb)
579 {
580 	zio_t *zio;
581 
582 	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
583 	    data, size, done, private,
584 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
585 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
586 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
587 
588 	return (zio);
589 }
590 
591 zio_t *
592 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
593     void *data, uint64_t size, const zio_prop_t *zp,
594     zio_done_func_t *ready, zio_done_func_t *done, void *private,
595     int priority, enum zio_flag flags, const zbookmark_t *zb)
596 {
597 	zio_t *zio;
598 
599 	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
600 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
601 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
602 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
603 	    zp->zp_type < DMU_OT_NUMTYPES &&
604 	    zp->zp_level < 32 &&
605 	    zp->zp_copies > 0 &&
606 	    zp->zp_copies <= spa_max_replication(spa) &&
607 	    zp->zp_dedup <= 1 &&
608 	    zp->zp_dedup_verify <= 1);
609 
610 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
611 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
612 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
613 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
614 
615 	zio->io_ready = ready;
616 	zio->io_prop = *zp;
617 
618 	return (zio);
619 }
620 
621 zio_t *
622 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
623     uint64_t size, zio_done_func_t *done, void *private, int priority,
624     enum zio_flag flags, zbookmark_t *zb)
625 {
626 	zio_t *zio;
627 
628 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
629 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
630 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
631 
632 	return (zio);
633 }
634 
635 void
636 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
637 {
638 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
639 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
640 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
641 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
642 
643 	zio->io_prop.zp_copies = copies;
644 	zio->io_bp_override = bp;
645 }
646 
647 void
648 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
649 {
650 	bplist_enqueue_deferred(&spa->spa_free_bplist[txg & TXG_MASK], bp);
651 }
652 
653 zio_t *
654 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
655     enum zio_flag flags)
656 {
657 	zio_t *zio;
658 
659 	ASSERT(!BP_IS_HOLE(bp));
660 	ASSERT(spa_syncing_txg(spa) == txg);
661 	ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
662 
663 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
664 	    NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
665 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
666 
667 	return (zio);
668 }
669 
670 zio_t *
671 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
672     zio_done_func_t *done, void *private, enum zio_flag flags)
673 {
674 	zio_t *zio;
675 
676 	/*
677 	 * A claim is an allocation of a specific block.  Claims are needed
678 	 * to support immediate writes in the intent log.  The issue is that
679 	 * immediate writes contain committed data, but in a txg that was
680 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
681 	 * the intent log claims all blocks that contain immediate write data
682 	 * so that the SPA knows they're in use.
683 	 *
684 	 * All claims *must* be resolved in the first txg -- before the SPA
685 	 * starts allocating blocks -- so that nothing is allocated twice.
686 	 * If txg == 0 we just verify that the block is claimable.
687 	 */
688 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
689 	ASSERT(txg == spa_first_txg(spa) || txg == 0);
690 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
691 
692 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
693 	    done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
694 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
695 
696 	return (zio);
697 }
698 
699 zio_t *
700 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
701     zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
702 {
703 	zio_t *zio;
704 	int c;
705 
706 	if (vd->vdev_children == 0) {
707 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
708 		    ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
709 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
710 
711 		zio->io_cmd = cmd;
712 	} else {
713 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
714 
715 		for (c = 0; c < vd->vdev_children; c++)
716 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
717 			    done, private, priority, flags));
718 	}
719 
720 	return (zio);
721 }
722 
723 zio_t *
724 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
725     void *data, int checksum, zio_done_func_t *done, void *private,
726     int priority, enum zio_flag flags, boolean_t labels)
727 {
728 	zio_t *zio;
729 
730 	ASSERT(vd->vdev_children == 0);
731 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
732 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
733 	ASSERT3U(offset + size, <=, vd->vdev_psize);
734 
735 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
736 	    ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
737 	    ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
738 
739 	zio->io_prop.zp_checksum = checksum;
740 
741 	return (zio);
742 }
743 
744 zio_t *
745 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
746     void *data, int checksum, zio_done_func_t *done, void *private,
747     int priority, enum zio_flag flags, boolean_t labels)
748 {
749 	zio_t *zio;
750 
751 	ASSERT(vd->vdev_children == 0);
752 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
753 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
754 	ASSERT3U(offset + size, <=, vd->vdev_psize);
755 
756 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
757 	    ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
758 	    ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
759 
760 	zio->io_prop.zp_checksum = checksum;
761 
762 	if (zio_checksum_table[checksum].ci_zbt) {
763 		/*
764 		 * zbt checksums are necessarily destructive -- they modify
765 		 * the end of the write buffer to hold the verifier/checksum.
766 		 * Therefore, we must make a local copy in case the data is
767 		 * being written to multiple places in parallel.
768 		 */
769 		void *wbuf = zio_buf_alloc(size);
770 		bcopy(data, wbuf, size);
771 		zio_push_transform(zio, wbuf, size, size, NULL);
772 	}
773 
774 	return (zio);
775 }
776 
777 /*
778  * Create a child I/O to do some work for us.
779  */
780 zio_t *
781 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
782 	void *data, uint64_t size, int type, int priority, enum zio_flag flags,
783 	zio_done_func_t *done, void *private)
784 {
785 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
786 	zio_t *zio;
787 
788 	ASSERT(vd->vdev_parent ==
789 	    (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
790 
791 	if (type == ZIO_TYPE_READ && bp != NULL) {
792 		/*
793 		 * If we have the bp, then the child should perform the
794 		 * checksum and the parent need not.  This pushes error
795 		 * detection as close to the leaves as possible and
796 		 * eliminates redundant checksums in the interior nodes.
797 		 */
798 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
799 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
800 	}
801 
802 	if (vd->vdev_children == 0)
803 		offset += VDEV_LABEL_START_SIZE;
804 
805 	flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
806 
807 	/*
808 	 * If we've decided to do a repair, the write is not speculative --
809 	 * even if the original read was.
810 	 */
811 	if (flags & ZIO_FLAG_IO_REPAIR)
812 		flags &= ~ZIO_FLAG_SPECULATIVE;
813 
814 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
815 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
816 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
817 
818 	return (zio);
819 }
820 
821 zio_t *
822 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
823 	int type, int priority, enum zio_flag flags,
824 	zio_done_func_t *done, void *private)
825 {
826 	zio_t *zio;
827 
828 	ASSERT(vd->vdev_ops->vdev_op_leaf);
829 
830 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
831 	    data, size, done, private, type, priority,
832 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
833 	    vd, offset, NULL,
834 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
835 
836 	return (zio);
837 }
838 
839 void
840 zio_flush(zio_t *zio, vdev_t *vd)
841 {
842 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
843 	    NULL, NULL, ZIO_PRIORITY_NOW,
844 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
845 }
846 
847 /*
848  * ==========================================================================
849  * Prepare to read and write logical blocks
850  * ==========================================================================
851  */
852 
853 static int
854 zio_read_bp_init(zio_t *zio)
855 {
856 	blkptr_t *bp = zio->io_bp;
857 
858 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
859 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
860 	    !(zio->io_flags & ZIO_FLAG_RAW)) {
861 		uint64_t psize = BP_GET_PSIZE(bp);
862 		void *cbuf = zio_buf_alloc(psize);
863 
864 		zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
865 	}
866 
867 	if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
868 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
869 
870 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
871 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
872 
873 	return (ZIO_PIPELINE_CONTINUE);
874 }
875 
876 static int
877 zio_write_bp_init(zio_t *zio)
878 {
879 	spa_t *spa = zio->io_spa;
880 	zio_prop_t *zp = &zio->io_prop;
881 	enum zio_compress compress = zp->zp_compress;
882 	blkptr_t *bp = zio->io_bp;
883 	uint64_t lsize = zio->io_size;
884 	uint64_t psize = lsize;
885 	int pass = 1;
886 
887 	/*
888 	 * If our children haven't all reached the ready stage,
889 	 * wait for them and then repeat this pipeline stage.
890 	 */
891 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
892 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
893 		return (ZIO_PIPELINE_STOP);
894 
895 	if (!IO_IS_ALLOCATING(zio))
896 		return (ZIO_PIPELINE_CONTINUE);
897 
898 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
899 
900 	if (zio->io_bp_override) {
901 		ASSERT(bp->blk_birth != zio->io_txg);
902 		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
903 
904 		*bp = *zio->io_bp_override;
905 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
906 
907 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
908 			return (ZIO_PIPELINE_CONTINUE);
909 
910 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
911 		    zp->zp_dedup_verify);
912 
913 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
914 			BP_SET_DEDUP(bp, 1);
915 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
916 			return (ZIO_PIPELINE_CONTINUE);
917 		}
918 		zio->io_bp_override = NULL;
919 		BP_ZERO(bp);
920 	}
921 
922 	if (bp->blk_birth == zio->io_txg) {
923 		/*
924 		 * We're rewriting an existing block, which means we're
925 		 * working on behalf of spa_sync().  For spa_sync() to
926 		 * converge, it must eventually be the case that we don't
927 		 * have to allocate new blocks.  But compression changes
928 		 * the blocksize, which forces a reallocate, and makes
929 		 * convergence take longer.  Therefore, after the first
930 		 * few passes, stop compressing to ensure convergence.
931 		 */
932 		pass = spa_sync_pass(spa);
933 
934 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
935 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
936 		ASSERT(!BP_GET_DEDUP(bp));
937 
938 		if (pass > SYNC_PASS_DONT_COMPRESS)
939 			compress = ZIO_COMPRESS_OFF;
940 
941 		/* Make sure someone doesn't change their mind on overwrites */
942 		ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
943 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
944 	}
945 
946 	if (compress != ZIO_COMPRESS_OFF) {
947 		void *cbuf = zio_buf_alloc(lsize);
948 		psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
949 		if (psize == 0 || psize == lsize) {
950 			compress = ZIO_COMPRESS_OFF;
951 			zio_buf_free(cbuf, lsize);
952 		} else {
953 			ASSERT(psize < lsize);
954 			zio_push_transform(zio, cbuf, psize, lsize, NULL);
955 		}
956 	}
957 
958 	/*
959 	 * The final pass of spa_sync() must be all rewrites, but the first
960 	 * few passes offer a trade-off: allocating blocks defers convergence,
961 	 * but newly allocated blocks are sequential, so they can be written
962 	 * to disk faster.  Therefore, we allow the first few passes of
963 	 * spa_sync() to allocate new blocks, but force rewrites after that.
964 	 * There should only be a handful of blocks after pass 1 in any case.
965 	 */
966 	if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
967 	    pass > SYNC_PASS_REWRITE) {
968 		ASSERT(psize != 0);
969 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
970 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
971 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
972 	} else {
973 		BP_ZERO(bp);
974 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
975 	}
976 
977 	if (psize == 0) {
978 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
979 	} else {
980 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
981 		BP_SET_LSIZE(bp, lsize);
982 		BP_SET_PSIZE(bp, psize);
983 		BP_SET_COMPRESS(bp, compress);
984 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
985 		BP_SET_TYPE(bp, zp->zp_type);
986 		BP_SET_LEVEL(bp, zp->zp_level);
987 		BP_SET_DEDUP(bp, zp->zp_dedup);
988 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
989 		if (zp->zp_dedup) {
990 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
991 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
992 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
993 		}
994 	}
995 
996 	return (ZIO_PIPELINE_CONTINUE);
997 }
998 
999 static int
1000 zio_free_bp_init(zio_t *zio)
1001 {
1002 	blkptr_t *bp = zio->io_bp;
1003 
1004 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1005 		if (BP_GET_DEDUP(bp))
1006 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1007 		else
1008 			arc_free(zio->io_spa, bp);
1009 	}
1010 
1011 	return (ZIO_PIPELINE_CONTINUE);
1012 }
1013 
1014 /*
1015  * ==========================================================================
1016  * Execute the I/O pipeline
1017  * ==========================================================================
1018  */
1019 
1020 static void
1021 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q)
1022 {
1023 	zio_type_t t = zio->io_type;
1024 
1025 	/*
1026 	 * If we're a config writer or a probe, the normal issue and
1027 	 * interrupt threads may all be blocked waiting for the config lock.
1028 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1029 	 */
1030 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1031 		t = ZIO_TYPE_NULL;
1032 
1033 	/*
1034 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1035 	 */
1036 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1037 		t = ZIO_TYPE_NULL;
1038 
1039 	(void) taskq_dispatch(zio->io_spa->spa_zio_taskq[t][q],
1040 	    (task_func_t *)zio_execute, zio, TQ_SLEEP);
1041 }
1042 
1043 static boolean_t
1044 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1045 {
1046 	kthread_t *executor = zio->io_executor;
1047 	spa_t *spa = zio->io_spa;
1048 
1049 	for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1050 		if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1051 			return (B_TRUE);
1052 
1053 	return (B_FALSE);
1054 }
1055 
1056 static int
1057 zio_issue_async(zio_t *zio)
1058 {
1059 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE);
1060 
1061 	return (ZIO_PIPELINE_STOP);
1062 }
1063 
1064 void
1065 zio_interrupt(zio_t *zio)
1066 {
1067 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT);
1068 }
1069 
1070 /*
1071  * Execute the I/O pipeline until one of the following occurs:
1072  * (1) the I/O completes; (2) the pipeline stalls waiting for
1073  * dependent child I/Os; (3) the I/O issues, so we're waiting
1074  * for an I/O completion interrupt; (4) the I/O is delegated by
1075  * vdev-level caching or aggregation; (5) the I/O is deferred
1076  * due to vdev-level queueing; (6) the I/O is handed off to
1077  * another thread.  In all cases, the pipeline stops whenever
1078  * there's no CPU work; it never burns a thread in cv_wait().
1079  *
1080  * There's no locking on io_stage because there's no legitimate way
1081  * for multiple threads to be attempting to process the same I/O.
1082  */
1083 static zio_pipe_stage_t *zio_pipeline[];
1084 
1085 void
1086 zio_execute(zio_t *zio)
1087 {
1088 	zio->io_executor = curthread;
1089 
1090 	while (zio->io_stage < ZIO_STAGE_DONE) {
1091 		enum zio_stage pipeline = zio->io_pipeline;
1092 		enum zio_stage stage = zio->io_stage;
1093 		int rv;
1094 
1095 		ASSERT(!MUTEX_HELD(&zio->io_lock));
1096 		ASSERT(ISP2(stage));
1097 		ASSERT(zio->io_stall == NULL);
1098 
1099 		do {
1100 			stage <<= 1;
1101 		} while ((stage & pipeline) == 0);
1102 
1103 		ASSERT(stage <= ZIO_STAGE_DONE);
1104 
1105 		/*
1106 		 * If we are in interrupt context and this pipeline stage
1107 		 * will grab a config lock that is held across I/O,
1108 		 * or may wait for an I/O that needs an interrupt thread
1109 		 * to complete, issue async to avoid deadlock.
1110 		 */
1111 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1112 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1113 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE);
1114 			return;
1115 		}
1116 
1117 		zio->io_stage = stage;
1118 		rv = zio_pipeline[highbit(stage) - 1](zio);
1119 
1120 		if (rv == ZIO_PIPELINE_STOP)
1121 			return;
1122 
1123 		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1124 	}
1125 }
1126 
1127 /*
1128  * ==========================================================================
1129  * Initiate I/O, either sync or async
1130  * ==========================================================================
1131  */
1132 int
1133 zio_wait(zio_t *zio)
1134 {
1135 	int error;
1136 
1137 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1138 	ASSERT(zio->io_executor == NULL);
1139 
1140 	zio->io_waiter = curthread;
1141 
1142 	zio_execute(zio);
1143 
1144 	mutex_enter(&zio->io_lock);
1145 	while (zio->io_executor != NULL)
1146 		cv_wait(&zio->io_cv, &zio->io_lock);
1147 	mutex_exit(&zio->io_lock);
1148 
1149 	error = zio->io_error;
1150 	zio_destroy(zio);
1151 
1152 	return (error);
1153 }
1154 
1155 void
1156 zio_nowait(zio_t *zio)
1157 {
1158 	ASSERT(zio->io_executor == NULL);
1159 
1160 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1161 	    zio_unique_parent(zio) == NULL) {
1162 		/*
1163 		 * This is a logical async I/O with no parent to wait for it.
1164 		 * We add it to the spa_async_root_zio "Godfather" I/O which
1165 		 * will ensure they complete prior to unloading the pool.
1166 		 */
1167 		spa_t *spa = zio->io_spa;
1168 
1169 		zio_add_child(spa->spa_async_zio_root, zio);
1170 	}
1171 
1172 	zio_execute(zio);
1173 }
1174 
1175 /*
1176  * ==========================================================================
1177  * Reexecute or suspend/resume failed I/O
1178  * ==========================================================================
1179  */
1180 
1181 static void
1182 zio_reexecute(zio_t *pio)
1183 {
1184 	zio_t *cio, *cio_next;
1185 
1186 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1187 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1188 	ASSERT(pio->io_gang_leader == NULL);
1189 	ASSERT(pio->io_gang_tree == NULL);
1190 
1191 	pio->io_flags = pio->io_orig_flags;
1192 	pio->io_stage = pio->io_orig_stage;
1193 	pio->io_pipeline = pio->io_orig_pipeline;
1194 	pio->io_reexecute = 0;
1195 	pio->io_error = 0;
1196 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1197 		pio->io_state[w] = 0;
1198 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1199 		pio->io_child_error[c] = 0;
1200 
1201 	if (IO_IS_ALLOCATING(pio))
1202 		BP_ZERO(pio->io_bp);
1203 
1204 	/*
1205 	 * As we reexecute pio's children, new children could be created.
1206 	 * New children go to the head of pio's io_child_list, however,
1207 	 * so we will (correctly) not reexecute them.  The key is that
1208 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1209 	 * cannot be affected by any side effects of reexecuting 'cio'.
1210 	 */
1211 	for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1212 		cio_next = zio_walk_children(pio);
1213 		mutex_enter(&pio->io_lock);
1214 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1215 			pio->io_children[cio->io_child_type][w]++;
1216 		mutex_exit(&pio->io_lock);
1217 		zio_reexecute(cio);
1218 	}
1219 
1220 	/*
1221 	 * Now that all children have been reexecuted, execute the parent.
1222 	 * We don't reexecute "The Godfather" I/O here as it's the
1223 	 * responsibility of the caller to wait on him.
1224 	 */
1225 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1226 		zio_execute(pio);
1227 }
1228 
1229 void
1230 zio_suspend(spa_t *spa, zio_t *zio)
1231 {
1232 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1233 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1234 		    "failure and the failure mode property for this pool "
1235 		    "is set to panic.", spa_name(spa));
1236 
1237 	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1238 
1239 	mutex_enter(&spa->spa_suspend_lock);
1240 
1241 	if (spa->spa_suspend_zio_root == NULL)
1242 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1243 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1244 		    ZIO_FLAG_GODFATHER);
1245 
1246 	spa->spa_suspended = B_TRUE;
1247 
1248 	if (zio != NULL) {
1249 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1250 		ASSERT(zio != spa->spa_suspend_zio_root);
1251 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1252 		ASSERT(zio_unique_parent(zio) == NULL);
1253 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1254 		zio_add_child(spa->spa_suspend_zio_root, zio);
1255 	}
1256 
1257 	mutex_exit(&spa->spa_suspend_lock);
1258 }
1259 
1260 int
1261 zio_resume(spa_t *spa)
1262 {
1263 	zio_t *pio;
1264 
1265 	/*
1266 	 * Reexecute all previously suspended i/o.
1267 	 */
1268 	mutex_enter(&spa->spa_suspend_lock);
1269 	spa->spa_suspended = B_FALSE;
1270 	cv_broadcast(&spa->spa_suspend_cv);
1271 	pio = spa->spa_suspend_zio_root;
1272 	spa->spa_suspend_zio_root = NULL;
1273 	mutex_exit(&spa->spa_suspend_lock);
1274 
1275 	if (pio == NULL)
1276 		return (0);
1277 
1278 	zio_reexecute(pio);
1279 	return (zio_wait(pio));
1280 }
1281 
1282 void
1283 zio_resume_wait(spa_t *spa)
1284 {
1285 	mutex_enter(&spa->spa_suspend_lock);
1286 	while (spa_suspended(spa))
1287 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1288 	mutex_exit(&spa->spa_suspend_lock);
1289 }
1290 
1291 /*
1292  * ==========================================================================
1293  * Gang blocks.
1294  *
1295  * A gang block is a collection of small blocks that looks to the DMU
1296  * like one large block.  When zio_dva_allocate() cannot find a block
1297  * of the requested size, due to either severe fragmentation or the pool
1298  * being nearly full, it calls zio_write_gang_block() to construct the
1299  * block from smaller fragments.
1300  *
1301  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1302  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
1303  * an indirect block: it's an array of block pointers.  It consumes
1304  * only one sector and hence is allocatable regardless of fragmentation.
1305  * The gang header's bps point to its gang members, which hold the data.
1306  *
1307  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1308  * as the verifier to ensure uniqueness of the SHA256 checksum.
1309  * Critically, the gang block bp's blk_cksum is the checksum of the data,
1310  * not the gang header.  This ensures that data block signatures (needed for
1311  * deduplication) are independent of how the block is physically stored.
1312  *
1313  * Gang blocks can be nested: a gang member may itself be a gang block.
1314  * Thus every gang block is a tree in which root and all interior nodes are
1315  * gang headers, and the leaves are normal blocks that contain user data.
1316  * The root of the gang tree is called the gang leader.
1317  *
1318  * To perform any operation (read, rewrite, free, claim) on a gang block,
1319  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1320  * in the io_gang_tree field of the original logical i/o by recursively
1321  * reading the gang leader and all gang headers below it.  This yields
1322  * an in-core tree containing the contents of every gang header and the
1323  * bps for every constituent of the gang block.
1324  *
1325  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1326  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
1327  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1328  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1329  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1330  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
1331  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1332  * of the gang header plus zio_checksum_compute() of the data to update the
1333  * gang header's blk_cksum as described above.
1334  *
1335  * The two-phase assemble/issue model solves the problem of partial failure --
1336  * what if you'd freed part of a gang block but then couldn't read the
1337  * gang header for another part?  Assembling the entire gang tree first
1338  * ensures that all the necessary gang header I/O has succeeded before
1339  * starting the actual work of free, claim, or write.  Once the gang tree
1340  * is assembled, free and claim are in-memory operations that cannot fail.
1341  *
1342  * In the event that a gang write fails, zio_dva_unallocate() walks the
1343  * gang tree to immediately free (i.e. insert back into the space map)
1344  * everything we've allocated.  This ensures that we don't get ENOSPC
1345  * errors during repeated suspend/resume cycles due to a flaky device.
1346  *
1347  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
1348  * the gang tree, we won't modify the block, so we can safely defer the free
1349  * (knowing that the block is still intact).  If we *can* assemble the gang
1350  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1351  * each constituent bp and we can allocate a new block on the next sync pass.
1352  *
1353  * In all cases, the gang tree allows complete recovery from partial failure.
1354  * ==========================================================================
1355  */
1356 
1357 static zio_t *
1358 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1359 {
1360 	if (gn != NULL)
1361 		return (pio);
1362 
1363 	return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1364 	    NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1365 	    &pio->io_bookmark));
1366 }
1367 
1368 zio_t *
1369 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1370 {
1371 	zio_t *zio;
1372 
1373 	if (gn != NULL) {
1374 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1375 		    gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1376 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1377 		/*
1378 		 * As we rewrite each gang header, the pipeline will compute
1379 		 * a new gang block header checksum for it; but no one will
1380 		 * compute a new data checksum, so we do that here.  The one
1381 		 * exception is the gang leader: the pipeline already computed
1382 		 * its data checksum because that stage precedes gang assembly.
1383 		 * (Presently, nothing actually uses interior data checksums;
1384 		 * this is just good hygiene.)
1385 		 */
1386 		if (gn != pio->io_gang_leader->io_gang_tree) {
1387 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1388 			    data, BP_GET_PSIZE(bp));
1389 		}
1390 		/*
1391 		 * If we are here to damage data for testing purposes,
1392 		 * leave the GBH alone so that we can detect the damage.
1393 		 */
1394 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1395 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1396 	} else {
1397 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1398 		    data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1399 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1400 	}
1401 
1402 	return (zio);
1403 }
1404 
1405 /* ARGSUSED */
1406 zio_t *
1407 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1408 {
1409 	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1410 	    ZIO_GANG_CHILD_FLAGS(pio)));
1411 }
1412 
1413 /* ARGSUSED */
1414 zio_t *
1415 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1416 {
1417 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1418 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1419 }
1420 
1421 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1422 	NULL,
1423 	zio_read_gang,
1424 	zio_rewrite_gang,
1425 	zio_free_gang,
1426 	zio_claim_gang,
1427 	NULL
1428 };
1429 
1430 static void zio_gang_tree_assemble_done(zio_t *zio);
1431 
1432 static zio_gang_node_t *
1433 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1434 {
1435 	zio_gang_node_t *gn;
1436 
1437 	ASSERT(*gnpp == NULL);
1438 
1439 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1440 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1441 	*gnpp = gn;
1442 
1443 	return (gn);
1444 }
1445 
1446 static void
1447 zio_gang_node_free(zio_gang_node_t **gnpp)
1448 {
1449 	zio_gang_node_t *gn = *gnpp;
1450 
1451 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1452 		ASSERT(gn->gn_child[g] == NULL);
1453 
1454 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1455 	kmem_free(gn, sizeof (*gn));
1456 	*gnpp = NULL;
1457 }
1458 
1459 static void
1460 zio_gang_tree_free(zio_gang_node_t **gnpp)
1461 {
1462 	zio_gang_node_t *gn = *gnpp;
1463 
1464 	if (gn == NULL)
1465 		return;
1466 
1467 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1468 		zio_gang_tree_free(&gn->gn_child[g]);
1469 
1470 	zio_gang_node_free(gnpp);
1471 }
1472 
1473 static void
1474 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1475 {
1476 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1477 
1478 	ASSERT(gio->io_gang_leader == gio);
1479 	ASSERT(BP_IS_GANG(bp));
1480 
1481 	zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1482 	    SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1483 	    gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1484 }
1485 
1486 static void
1487 zio_gang_tree_assemble_done(zio_t *zio)
1488 {
1489 	zio_t *gio = zio->io_gang_leader;
1490 	zio_gang_node_t *gn = zio->io_private;
1491 	blkptr_t *bp = zio->io_bp;
1492 
1493 	ASSERT(gio == zio_unique_parent(zio));
1494 	ASSERT(zio->io_child_count == 0);
1495 
1496 	if (zio->io_error)
1497 		return;
1498 
1499 	if (BP_SHOULD_BYTESWAP(bp))
1500 		byteswap_uint64_array(zio->io_data, zio->io_size);
1501 
1502 	ASSERT(zio->io_data == gn->gn_gbh);
1503 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1504 	ASSERT(gn->gn_gbh->zg_tail.zbt_magic == ZBT_MAGIC);
1505 
1506 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1507 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1508 		if (!BP_IS_GANG(gbp))
1509 			continue;
1510 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1511 	}
1512 }
1513 
1514 static void
1515 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1516 {
1517 	zio_t *gio = pio->io_gang_leader;
1518 	zio_t *zio;
1519 
1520 	ASSERT(BP_IS_GANG(bp) == !!gn);
1521 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1522 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1523 
1524 	/*
1525 	 * If you're a gang header, your data is in gn->gn_gbh.
1526 	 * If you're a gang member, your data is in 'data' and gn == NULL.
1527 	 */
1528 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1529 
1530 	if (gn != NULL) {
1531 		ASSERT(gn->gn_gbh->zg_tail.zbt_magic == ZBT_MAGIC);
1532 
1533 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1534 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1535 			if (BP_IS_HOLE(gbp))
1536 				continue;
1537 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1538 			data = (char *)data + BP_GET_PSIZE(gbp);
1539 		}
1540 	}
1541 
1542 	if (gn == gio->io_gang_tree)
1543 		ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1544 
1545 	if (zio != pio)
1546 		zio_nowait(zio);
1547 }
1548 
1549 static int
1550 zio_gang_assemble(zio_t *zio)
1551 {
1552 	blkptr_t *bp = zio->io_bp;
1553 
1554 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1555 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1556 
1557 	zio->io_gang_leader = zio;
1558 
1559 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1560 
1561 	return (ZIO_PIPELINE_CONTINUE);
1562 }
1563 
1564 static int
1565 zio_gang_issue(zio_t *zio)
1566 {
1567 	blkptr_t *bp = zio->io_bp;
1568 
1569 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1570 		return (ZIO_PIPELINE_STOP);
1571 
1572 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1573 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1574 
1575 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1576 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1577 	else
1578 		zio_gang_tree_free(&zio->io_gang_tree);
1579 
1580 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1581 
1582 	return (ZIO_PIPELINE_CONTINUE);
1583 }
1584 
1585 static void
1586 zio_write_gang_member_ready(zio_t *zio)
1587 {
1588 	zio_t *pio = zio_unique_parent(zio);
1589 	zio_t *gio = zio->io_gang_leader;
1590 	dva_t *cdva = zio->io_bp->blk_dva;
1591 	dva_t *pdva = pio->io_bp->blk_dva;
1592 	uint64_t asize;
1593 
1594 	if (BP_IS_HOLE(zio->io_bp))
1595 		return;
1596 
1597 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1598 
1599 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1600 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1601 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1602 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1603 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1604 
1605 	mutex_enter(&pio->io_lock);
1606 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1607 		ASSERT(DVA_GET_GANG(&pdva[d]));
1608 		asize = DVA_GET_ASIZE(&pdva[d]);
1609 		asize += DVA_GET_ASIZE(&cdva[d]);
1610 		DVA_SET_ASIZE(&pdva[d], asize);
1611 	}
1612 	mutex_exit(&pio->io_lock);
1613 }
1614 
1615 static int
1616 zio_write_gang_block(zio_t *pio)
1617 {
1618 	spa_t *spa = pio->io_spa;
1619 	blkptr_t *bp = pio->io_bp;
1620 	zio_t *gio = pio->io_gang_leader;
1621 	zio_t *zio;
1622 	zio_gang_node_t *gn, **gnpp;
1623 	zio_gbh_phys_t *gbh;
1624 	uint64_t txg = pio->io_txg;
1625 	uint64_t resid = pio->io_size;
1626 	uint64_t lsize;
1627 	int copies = gio->io_prop.zp_copies;
1628 	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1629 	zio_prop_t zp;
1630 	int error;
1631 
1632 	error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1633 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1634 	    METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1635 	if (error) {
1636 		pio->io_error = error;
1637 		return (ZIO_PIPELINE_CONTINUE);
1638 	}
1639 
1640 	if (pio == gio) {
1641 		gnpp = &gio->io_gang_tree;
1642 	} else {
1643 		gnpp = pio->io_private;
1644 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
1645 	}
1646 
1647 	gn = zio_gang_node_alloc(gnpp);
1648 	gbh = gn->gn_gbh;
1649 	bzero(gbh, SPA_GANGBLOCKSIZE);
1650 
1651 	/*
1652 	 * Create the gang header.
1653 	 */
1654 	zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1655 	    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1656 
1657 	/*
1658 	 * Create and nowait the gang children.
1659 	 */
1660 	for (int g = 0; resid != 0; resid -= lsize, g++) {
1661 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1662 		    SPA_MINBLOCKSIZE);
1663 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1664 
1665 		zp.zp_checksum = gio->io_prop.zp_checksum;
1666 		zp.zp_compress = ZIO_COMPRESS_OFF;
1667 		zp.zp_type = DMU_OT_NONE;
1668 		zp.zp_level = 0;
1669 		zp.zp_copies = gio->io_prop.zp_copies;
1670 		zp.zp_dedup = 0;
1671 		zp.zp_dedup_verify = 0;
1672 
1673 		zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1674 		    (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1675 		    zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1676 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1677 		    &pio->io_bookmark));
1678 	}
1679 
1680 	/*
1681 	 * Set pio's pipeline to just wait for zio to finish.
1682 	 */
1683 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1684 
1685 	zio_nowait(zio);
1686 
1687 	return (ZIO_PIPELINE_CONTINUE);
1688 }
1689 
1690 /*
1691  * ==========================================================================
1692  * Dedup
1693  * ==========================================================================
1694  */
1695 static void
1696 zio_ddt_child_read_done(zio_t *zio)
1697 {
1698 	blkptr_t *bp = zio->io_bp;
1699 	ddt_entry_t *dde = zio->io_private;
1700 	ddt_phys_t *ddp;
1701 	zio_t *pio = zio_unique_parent(zio);
1702 
1703 	mutex_enter(&pio->io_lock);
1704 	ddp = ddt_phys_select(dde, bp);
1705 	if (zio->io_error == 0)
1706 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
1707 	if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1708 		dde->dde_repair_data = zio->io_data;
1709 	else
1710 		zio_buf_free(zio->io_data, zio->io_size);
1711 	mutex_exit(&pio->io_lock);
1712 }
1713 
1714 static int
1715 zio_ddt_read_start(zio_t *zio)
1716 {
1717 	blkptr_t *bp = zio->io_bp;
1718 
1719 	ASSERT(BP_GET_DEDUP(bp));
1720 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1721 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1722 
1723 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1724 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1725 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1726 		ddt_phys_t *ddp = dde->dde_phys;
1727 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1728 		blkptr_t blk;
1729 
1730 		ASSERT(zio->io_vsd == NULL);
1731 		zio->io_vsd = dde;
1732 
1733 		if (ddp_self == NULL)
1734 			return (ZIO_PIPELINE_CONTINUE);
1735 
1736 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1737 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1738 				continue;
1739 			ddt_bp_create(ddt, &dde->dde_key, ddp, &blk);
1740 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
1741 			    zio_buf_alloc(zio->io_size), zio->io_size,
1742 			    zio_ddt_child_read_done, dde, zio->io_priority,
1743 			    ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1744 			    &zio->io_bookmark));
1745 		}
1746 		return (ZIO_PIPELINE_CONTINUE);
1747 	}
1748 
1749 	zio_nowait(zio_read(zio, zio->io_spa, bp,
1750 	    zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1751 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1752 
1753 	return (ZIO_PIPELINE_CONTINUE);
1754 }
1755 
1756 static int
1757 zio_ddt_read_done(zio_t *zio)
1758 {
1759 	blkptr_t *bp = zio->io_bp;
1760 
1761 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1762 		return (ZIO_PIPELINE_STOP);
1763 
1764 	ASSERT(BP_GET_DEDUP(bp));
1765 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1766 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1767 
1768 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1769 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1770 		ddt_entry_t *dde = zio->io_vsd;
1771 		if (ddt == NULL) {
1772 			ASSERT(zio->io_spa->spa_load_state != SPA_LOAD_NONE);
1773 			return (ZIO_PIPELINE_CONTINUE);
1774 		}
1775 		if (dde == NULL) {
1776 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1777 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE);
1778 			return (ZIO_PIPELINE_STOP);
1779 		}
1780 		if (dde->dde_repair_data != NULL) {
1781 			bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1782 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
1783 		}
1784 		ddt_repair_done(ddt, dde);
1785 		zio->io_vsd = NULL;
1786 	}
1787 
1788 	ASSERT(zio->io_vsd == NULL);
1789 
1790 	return (ZIO_PIPELINE_CONTINUE);
1791 }
1792 
1793 static boolean_t
1794 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1795 {
1796 	spa_t *spa = zio->io_spa;
1797 
1798 	/*
1799 	 * Note: we compare the original data, not the transformed data,
1800 	 * because when zio->io_bp is an override bp, we will not have
1801 	 * pushed the I/O transforms.  That's an important optimization
1802 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1803 	 */
1804 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1805 		zio_t *lio = dde->dde_lead_zio[p];
1806 
1807 		if (lio != NULL) {
1808 			return (lio->io_orig_size != zio->io_orig_size ||
1809 			    bcmp(zio->io_orig_data, lio->io_orig_data,
1810 			    zio->io_orig_size) != 0);
1811 		}
1812 	}
1813 
1814 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1815 		ddt_phys_t *ddp = &dde->dde_phys[p];
1816 
1817 		if (ddp->ddp_phys_birth != 0) {
1818 			arc_buf_t *abuf = NULL;
1819 			uint32_t aflags = ARC_WAIT;
1820 			blkptr_t blk = *zio->io_bp;
1821 			int error;
1822 
1823 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1824 
1825 			ddt_exit(ddt);
1826 
1827 			error = arc_read_nolock(NULL, spa, &blk,
1828 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1829 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1830 			    &aflags, &zio->io_bookmark);
1831 
1832 			if (error == 0) {
1833 				if (arc_buf_size(abuf) != zio->io_orig_size ||
1834 				    bcmp(abuf->b_data, zio->io_orig_data,
1835 				    zio->io_orig_size) != 0)
1836 					error = EEXIST;
1837 				VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1838 			}
1839 
1840 			ddt_enter(ddt);
1841 			return (error != 0);
1842 		}
1843 	}
1844 
1845 	return (B_FALSE);
1846 }
1847 
1848 static void
1849 zio_ddt_child_write_ready(zio_t *zio)
1850 {
1851 	int p = zio->io_prop.zp_copies;
1852 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1853 	ddt_entry_t *dde = zio->io_private;
1854 	ddt_phys_t *ddp = &dde->dde_phys[p];
1855 	zio_t *pio;
1856 
1857 	if (zio->io_error)
1858 		return;
1859 
1860 	ddt_enter(ddt);
1861 
1862 	ASSERT(dde->dde_lead_zio[p] == zio);
1863 
1864 	ddt_phys_fill(ddp, zio->io_bp);
1865 
1866 	while ((pio = zio_walk_parents(zio)) != NULL)
1867 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1868 
1869 	ddt_exit(ddt);
1870 }
1871 
1872 static void
1873 zio_ddt_child_write_done(zio_t *zio)
1874 {
1875 	int p = zio->io_prop.zp_copies;
1876 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1877 	ddt_entry_t *dde = zio->io_private;
1878 	ddt_phys_t *ddp = &dde->dde_phys[p];
1879 
1880 	ddt_enter(ddt);
1881 
1882 	ASSERT(ddp->ddp_refcnt == 0);
1883 	ASSERT(dde->dde_lead_zio[p] == zio);
1884 	dde->dde_lead_zio[p] = NULL;
1885 
1886 	if (zio->io_error == 0) {
1887 		while (zio_walk_parents(zio) != NULL)
1888 			ddt_phys_addref(ddp);
1889 	} else {
1890 		ddt_phys_clear(ddp);
1891 	}
1892 
1893 	ddt_exit(ddt);
1894 }
1895 
1896 static void
1897 zio_ddt_ditto_write_done(zio_t *zio)
1898 {
1899 	int p = DDT_PHYS_DITTO;
1900 	zio_prop_t *zp = &zio->io_prop;
1901 	blkptr_t *bp = zio->io_bp;
1902 	ddt_t *ddt = ddt_select(zio->io_spa, bp);
1903 	ddt_entry_t *dde = zio->io_private;
1904 	ddt_phys_t *ddp = &dde->dde_phys[p];
1905 	ddt_key_t *ddk = &dde->dde_key;
1906 
1907 	ddt_enter(ddt);
1908 
1909 	ASSERT(ddp->ddp_refcnt == 0);
1910 	ASSERT(dde->dde_lead_zio[p] == zio);
1911 	dde->dde_lead_zio[p] = NULL;
1912 
1913 	if (zio->io_error == 0) {
1914 		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1915 		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1916 		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1917 		if (ddp->ddp_phys_birth != 0)
1918 			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1919 		ddt_phys_fill(ddp, bp);
1920 	}
1921 
1922 	ddt_exit(ddt);
1923 }
1924 
1925 static int
1926 zio_ddt_write(zio_t *zio)
1927 {
1928 	spa_t *spa = zio->io_spa;
1929 	blkptr_t *bp = zio->io_bp;
1930 	uint64_t txg = zio->io_txg;
1931 	zio_prop_t *zp = &zio->io_prop;
1932 	int p = zp->zp_copies;
1933 	int ditto_copies;
1934 	zio_t *cio = NULL;
1935 	zio_t *dio = NULL;
1936 	ddt_t *ddt = ddt_select(spa, bp);
1937 	ddt_entry_t *dde;
1938 	ddt_phys_t *ddp;
1939 
1940 	ASSERT(BP_GET_DEDUP(bp));
1941 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
1942 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
1943 
1944 	ddt_enter(ddt);
1945 	dde = ddt_lookup(ddt, bp, B_TRUE);
1946 	ddp = &dde->dde_phys[p];
1947 
1948 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
1949 		/*
1950 		 * If we're using a weak checksum, upgrade to a strong checksum
1951 		 * and try again.  If we're already using a strong checksum,
1952 		 * we can't resolve it, so just convert to an ordinary write.
1953 		 * (And automatically e-mail a paper to Nature?)
1954 		 */
1955 		if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
1956 			zp->zp_checksum = spa_dedup_checksum(spa);
1957 			zio_pop_transforms(zio);
1958 			zio->io_stage = ZIO_STAGE_OPEN;
1959 			BP_ZERO(bp);
1960 		} else {
1961 			zp->zp_dedup = 0;
1962 		}
1963 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1964 		ddt_exit(ddt);
1965 		return (ZIO_PIPELINE_CONTINUE);
1966 	}
1967 
1968 	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
1969 	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
1970 
1971 	if (ditto_copies > ddt_ditto_copies_present(dde) &&
1972 	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
1973 		zio_prop_t czp = *zp;
1974 
1975 		czp.zp_copies = ditto_copies;
1976 
1977 		/*
1978 		 * If we arrived here with an override bp, we won't have run
1979 		 * the transform stack, so we won't have the data we need to
1980 		 * generate a child i/o.  So, toss the override bp and restart.
1981 		 * This is safe, because using the override bp is just an
1982 		 * optimization; and it's rare, so the cost doesn't matter.
1983 		 */
1984 		if (zio->io_bp_override) {
1985 			zio_pop_transforms(zio);
1986 			zio->io_stage = ZIO_STAGE_OPEN;
1987 			zio->io_pipeline = ZIO_WRITE_PIPELINE;
1988 			zio->io_bp_override = NULL;
1989 			BP_ZERO(bp);
1990 			ddt_exit(ddt);
1991 			return (ZIO_PIPELINE_CONTINUE);
1992 		}
1993 
1994 		dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
1995 		    zio->io_orig_size, &czp, NULL,
1996 		    zio_ddt_ditto_write_done, dde, zio->io_priority,
1997 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
1998 
1999 		zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2000 		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2001 	}
2002 
2003 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2004 		if (ddp->ddp_phys_birth != 0)
2005 			ddt_bp_fill(ddp, bp, txg);
2006 		if (dde->dde_lead_zio[p] != NULL)
2007 			zio_add_child(zio, dde->dde_lead_zio[p]);
2008 		else
2009 			ddt_phys_addref(ddp);
2010 	} else if (zio->io_bp_override) {
2011 		ASSERT(bp->blk_birth == txg);
2012 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2013 		ddt_phys_fill(ddp, bp);
2014 		ddt_phys_addref(ddp);
2015 	} else {
2016 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2017 		    zio->io_orig_size, zp, zio_ddt_child_write_ready,
2018 		    zio_ddt_child_write_done, dde, zio->io_priority,
2019 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2020 
2021 		zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2022 		dde->dde_lead_zio[p] = cio;
2023 	}
2024 
2025 	ddt_exit(ddt);
2026 
2027 	if (cio)
2028 		zio_nowait(cio);
2029 	if (dio)
2030 		zio_nowait(dio);
2031 
2032 	return (ZIO_PIPELINE_CONTINUE);
2033 }
2034 
2035 static int
2036 zio_ddt_free(zio_t *zio)
2037 {
2038 	spa_t *spa = zio->io_spa;
2039 	blkptr_t *bp = zio->io_bp;
2040 	ddt_t *ddt = ddt_select(spa, bp);
2041 	ddt_entry_t *dde;
2042 	ddt_phys_t *ddp;
2043 
2044 	ASSERT(BP_GET_DEDUP(bp));
2045 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2046 
2047 	ddt_enter(ddt);
2048 	dde = ddt_lookup(ddt, bp, B_TRUE);
2049 	ddp = ddt_phys_select(dde, bp);
2050 	ddt_phys_decref(ddp);
2051 	ddt_exit(ddt);
2052 
2053 	return (ZIO_PIPELINE_CONTINUE);
2054 }
2055 
2056 /*
2057  * ==========================================================================
2058  * Allocate and free blocks
2059  * ==========================================================================
2060  */
2061 static int
2062 zio_dva_allocate(zio_t *zio)
2063 {
2064 	spa_t *spa = zio->io_spa;
2065 	metaslab_class_t *mc = spa_normal_class(spa);
2066 	blkptr_t *bp = zio->io_bp;
2067 	int error;
2068 
2069 	if (zio->io_gang_leader == NULL) {
2070 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2071 		zio->io_gang_leader = zio;
2072 	}
2073 
2074 	ASSERT(BP_IS_HOLE(bp));
2075 	ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2076 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
2077 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2078 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2079 
2080 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
2081 	    zio->io_prop.zp_copies, zio->io_txg, NULL, 0);
2082 
2083 	if (error) {
2084 		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2085 			return (zio_write_gang_block(zio));
2086 		zio->io_error = error;
2087 	}
2088 
2089 	return (ZIO_PIPELINE_CONTINUE);
2090 }
2091 
2092 static int
2093 zio_dva_free(zio_t *zio)
2094 {
2095 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2096 
2097 	return (ZIO_PIPELINE_CONTINUE);
2098 }
2099 
2100 static int
2101 zio_dva_claim(zio_t *zio)
2102 {
2103 	int error;
2104 
2105 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2106 	if (error)
2107 		zio->io_error = error;
2108 
2109 	return (ZIO_PIPELINE_CONTINUE);
2110 }
2111 
2112 /*
2113  * Undo an allocation.  This is used by zio_done() when an I/O fails
2114  * and we want to give back the block we just allocated.
2115  * This handles both normal blocks and gang blocks.
2116  */
2117 static void
2118 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2119 {
2120 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2121 	ASSERT(zio->io_bp_override == NULL);
2122 
2123 	if (!BP_IS_HOLE(bp))
2124 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2125 
2126 	if (gn != NULL) {
2127 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2128 			zio_dva_unallocate(zio, gn->gn_child[g],
2129 			    &gn->gn_gbh->zg_blkptr[g]);
2130 		}
2131 	}
2132 }
2133 
2134 /*
2135  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
2136  */
2137 int
2138 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2139     uint64_t size, boolean_t use_slog)
2140 {
2141 	int error = 1;
2142 
2143 	ASSERT(txg > spa_syncing_txg(spa));
2144 
2145 	if (use_slog)
2146 		error = metaslab_alloc(spa, spa_log_class(spa), size,
2147 		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2148 
2149 	if (error)
2150 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
2151 		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2152 
2153 	if (error == 0) {
2154 		BP_SET_LSIZE(new_bp, size);
2155 		BP_SET_PSIZE(new_bp, size);
2156 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2157 		BP_SET_CHECKSUM(new_bp, ZIO_CHECKSUM_ZILOG);
2158 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2159 		BP_SET_LEVEL(new_bp, 0);
2160 		BP_SET_DEDUP(new_bp, 0);
2161 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2162 	}
2163 
2164 	return (error);
2165 }
2166 
2167 /*
2168  * Free an intent log block.
2169  */
2170 void
2171 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2172 {
2173 	ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2174 	ASSERT(!BP_IS_GANG(bp));
2175 
2176 	zio_free(spa, txg, bp);
2177 }
2178 
2179 /*
2180  * ==========================================================================
2181  * Read and write to physical devices
2182  * ==========================================================================
2183  */
2184 static int
2185 zio_vdev_io_start(zio_t *zio)
2186 {
2187 	vdev_t *vd = zio->io_vd;
2188 	uint64_t align;
2189 	spa_t *spa = zio->io_spa;
2190 
2191 	ASSERT(zio->io_error == 0);
2192 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2193 
2194 	if (vd == NULL) {
2195 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2196 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2197 
2198 		/*
2199 		 * The mirror_ops handle multiple DVAs in a single BP.
2200 		 */
2201 		return (vdev_mirror_ops.vdev_op_io_start(zio));
2202 	}
2203 
2204 	align = 1ULL << vd->vdev_top->vdev_ashift;
2205 
2206 	if (P2PHASE(zio->io_size, align) != 0) {
2207 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
2208 		char *abuf = zio_buf_alloc(asize);
2209 		ASSERT(vd == vd->vdev_top);
2210 		if (zio->io_type == ZIO_TYPE_WRITE) {
2211 			bcopy(zio->io_data, abuf, zio->io_size);
2212 			bzero(abuf + zio->io_size, asize - zio->io_size);
2213 		}
2214 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2215 	}
2216 
2217 	ASSERT(P2PHASE(zio->io_offset, align) == 0);
2218 	ASSERT(P2PHASE(zio->io_size, align) == 0);
2219 	ASSERT(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2220 
2221 	/*
2222 	 * If this is a repair I/O, and there's no self-healing involved --
2223 	 * that is, we're just resilvering what we expect to resilver --
2224 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2225 	 * This prevents spurious resilvering with nested replication.
2226 	 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2227 	 * A is out of date, we'll read from C+D, then use the data to
2228 	 * resilver A+B -- but we don't actually want to resilver B, just A.
2229 	 * The top-level mirror has no way to know this, so instead we just
2230 	 * discard unnecessary repairs as we work our way down the vdev tree.
2231 	 * The same logic applies to any form of nested replication:
2232 	 * ditto + mirror, RAID-Z + replacing, etc.  This covers them all.
2233 	 */
2234 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2235 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2236 	    zio->io_txg != 0 &&	/* not a delegated i/o */
2237 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2238 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2239 		zio_vdev_io_bypass(zio);
2240 		return (ZIO_PIPELINE_CONTINUE);
2241 	}
2242 
2243 	if (vd->vdev_ops->vdev_op_leaf &&
2244 	    (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2245 
2246 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2247 			return (ZIO_PIPELINE_CONTINUE);
2248 
2249 		if ((zio = vdev_queue_io(zio)) == NULL)
2250 			return (ZIO_PIPELINE_STOP);
2251 
2252 		if (!vdev_accessible(vd, zio)) {
2253 			zio->io_error = ENXIO;
2254 			zio_interrupt(zio);
2255 			return (ZIO_PIPELINE_STOP);
2256 		}
2257 	}
2258 
2259 	return (vd->vdev_ops->vdev_op_io_start(zio));
2260 }
2261 
2262 static int
2263 zio_vdev_io_done(zio_t *zio)
2264 {
2265 	vdev_t *vd = zio->io_vd;
2266 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2267 	boolean_t unexpected_error = B_FALSE;
2268 
2269 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2270 		return (ZIO_PIPELINE_STOP);
2271 
2272 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2273 
2274 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2275 
2276 		vdev_queue_io_done(zio);
2277 
2278 		if (zio->io_type == ZIO_TYPE_WRITE)
2279 			vdev_cache_write(zio);
2280 
2281 		if (zio_injection_enabled && zio->io_error == 0)
2282 			zio->io_error = zio_handle_device_injection(vd,
2283 			    zio, EIO);
2284 
2285 		if (zio_injection_enabled && zio->io_error == 0)
2286 			zio->io_error = zio_handle_label_injection(zio, EIO);
2287 
2288 		if (zio->io_error) {
2289 			if (!vdev_accessible(vd, zio)) {
2290 				zio->io_error = ENXIO;
2291 			} else {
2292 				unexpected_error = B_TRUE;
2293 			}
2294 		}
2295 	}
2296 
2297 	ops->vdev_op_io_done(zio);
2298 
2299 	if (unexpected_error)
2300 		VERIFY(vdev_probe(vd, zio) == NULL);
2301 
2302 	return (ZIO_PIPELINE_CONTINUE);
2303 }
2304 
2305 /*
2306  * For non-raidz ZIOs, we can just copy aside the bad data read from the
2307  * disk, and use that to finish the checksum ereport later.
2308  */
2309 static void
2310 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2311     const void *good_buf)
2312 {
2313 	/* no processing needed */
2314 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2315 }
2316 
2317 /*ARGSUSED*/
2318 void
2319 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2320 {
2321 	void *buf = zio_buf_alloc(zio->io_size);
2322 
2323 	bcopy(zio->io_data, buf, zio->io_size);
2324 
2325 	zcr->zcr_cbinfo = zio->io_size;
2326 	zcr->zcr_cbdata = buf;
2327 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
2328 	zcr->zcr_free = zio_buf_free;
2329 }
2330 
2331 static int
2332 zio_vdev_io_assess(zio_t *zio)
2333 {
2334 	vdev_t *vd = zio->io_vd;
2335 
2336 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2337 		return (ZIO_PIPELINE_STOP);
2338 
2339 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2340 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2341 
2342 	if (zio->io_vsd != NULL) {
2343 		zio->io_vsd_ops->vsd_free(zio);
2344 		zio->io_vsd = NULL;
2345 	}
2346 
2347 	if (zio_injection_enabled && zio->io_error == 0)
2348 		zio->io_error = zio_handle_fault_injection(zio, EIO);
2349 
2350 	/*
2351 	 * If the I/O failed, determine whether we should attempt to retry it.
2352 	 */
2353 	if (zio->io_error && vd == NULL &&
2354 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2355 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
2356 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
2357 		zio->io_error = 0;
2358 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
2359 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2360 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2361 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE);
2362 		return (ZIO_PIPELINE_STOP);
2363 	}
2364 
2365 	/*
2366 	 * If we got an error on a leaf device, convert it to ENXIO
2367 	 * if the device is not accessible at all.
2368 	 */
2369 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2370 	    !vdev_accessible(vd, zio))
2371 		zio->io_error = ENXIO;
2372 
2373 	/*
2374 	 * If we can't write to an interior vdev (mirror or RAID-Z),
2375 	 * set vdev_cant_write so that we stop trying to allocate from it.
2376 	 */
2377 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2378 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2379 		vd->vdev_cant_write = B_TRUE;
2380 
2381 	if (zio->io_error)
2382 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2383 
2384 	return (ZIO_PIPELINE_CONTINUE);
2385 }
2386 
2387 void
2388 zio_vdev_io_reissue(zio_t *zio)
2389 {
2390 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2391 	ASSERT(zio->io_error == 0);
2392 
2393 	zio->io_stage >>= 1;
2394 }
2395 
2396 void
2397 zio_vdev_io_redone(zio_t *zio)
2398 {
2399 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2400 
2401 	zio->io_stage >>= 1;
2402 }
2403 
2404 void
2405 zio_vdev_io_bypass(zio_t *zio)
2406 {
2407 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2408 	ASSERT(zio->io_error == 0);
2409 
2410 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2411 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2412 }
2413 
2414 /*
2415  * ==========================================================================
2416  * Generate and verify checksums
2417  * ==========================================================================
2418  */
2419 static int
2420 zio_checksum_generate(zio_t *zio)
2421 {
2422 	blkptr_t *bp = zio->io_bp;
2423 	enum zio_checksum checksum;
2424 
2425 	if (bp == NULL) {
2426 		/*
2427 		 * This is zio_write_phys().
2428 		 * We're either generating a label checksum, or none at all.
2429 		 */
2430 		checksum = zio->io_prop.zp_checksum;
2431 
2432 		if (checksum == ZIO_CHECKSUM_OFF)
2433 			return (ZIO_PIPELINE_CONTINUE);
2434 
2435 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2436 	} else {
2437 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2438 			ASSERT(!IO_IS_ALLOCATING(zio));
2439 			checksum = ZIO_CHECKSUM_GANG_HEADER;
2440 		} else {
2441 			checksum = BP_GET_CHECKSUM(bp);
2442 		}
2443 	}
2444 
2445 	zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2446 
2447 	return (ZIO_PIPELINE_CONTINUE);
2448 }
2449 
2450 static int
2451 zio_checksum_verify(zio_t *zio)
2452 {
2453 	zio_bad_cksum_t info;
2454 	blkptr_t *bp = zio->io_bp;
2455 	int error;
2456 
2457 	ASSERT(zio->io_vd != NULL);
2458 
2459 	if (bp == NULL) {
2460 		/*
2461 		 * This is zio_read_phys().
2462 		 * We're either verifying a label checksum, or nothing at all.
2463 		 */
2464 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2465 			return (ZIO_PIPELINE_CONTINUE);
2466 
2467 		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2468 	}
2469 
2470 	if ((error = zio_checksum_error(zio, &info)) != 0) {
2471 		zio->io_error = error;
2472 		if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2473 			zfs_ereport_start_checksum(zio->io_spa,
2474 			    zio->io_vd, zio, zio->io_offset,
2475 			    zio->io_size, NULL, &info);
2476 		}
2477 	}
2478 
2479 	return (ZIO_PIPELINE_CONTINUE);
2480 }
2481 
2482 /*
2483  * Called by RAID-Z to ensure we don't compute the checksum twice.
2484  */
2485 void
2486 zio_checksum_verified(zio_t *zio)
2487 {
2488 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2489 }
2490 
2491 /*
2492  * ==========================================================================
2493  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2494  * An error of 0 indictes success.  ENXIO indicates whole-device failure,
2495  * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
2496  * indicate errors that are specific to one I/O, and most likely permanent.
2497  * Any other error is presumed to be worse because we weren't expecting it.
2498  * ==========================================================================
2499  */
2500 int
2501 zio_worst_error(int e1, int e2)
2502 {
2503 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2504 	int r1, r2;
2505 
2506 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2507 		if (e1 == zio_error_rank[r1])
2508 			break;
2509 
2510 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2511 		if (e2 == zio_error_rank[r2])
2512 			break;
2513 
2514 	return (r1 > r2 ? e1 : e2);
2515 }
2516 
2517 /*
2518  * ==========================================================================
2519  * I/O completion
2520  * ==========================================================================
2521  */
2522 static int
2523 zio_ready(zio_t *zio)
2524 {
2525 	blkptr_t *bp = zio->io_bp;
2526 	zio_t *pio, *pio_next;
2527 
2528 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2529 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2530 		return (ZIO_PIPELINE_STOP);
2531 
2532 	if (zio->io_ready) {
2533 		ASSERT(IO_IS_ALLOCATING(zio));
2534 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2535 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2536 
2537 		zio->io_ready(zio);
2538 	}
2539 
2540 	if (bp != NULL && bp != &zio->io_bp_copy)
2541 		zio->io_bp_copy = *bp;
2542 
2543 	if (zio->io_error)
2544 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2545 
2546 	mutex_enter(&zio->io_lock);
2547 	zio->io_state[ZIO_WAIT_READY] = 1;
2548 	pio = zio_walk_parents(zio);
2549 	mutex_exit(&zio->io_lock);
2550 
2551 	/*
2552 	 * As we notify zio's parents, new parents could be added.
2553 	 * New parents go to the head of zio's io_parent_list, however,
2554 	 * so we will (correctly) not notify them.  The remainder of zio's
2555 	 * io_parent_list, from 'pio_next' onward, cannot change because
2556 	 * all parents must wait for us to be done before they can be done.
2557 	 */
2558 	for (; pio != NULL; pio = pio_next) {
2559 		pio_next = zio_walk_parents(zio);
2560 		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2561 	}
2562 
2563 	if (zio->io_flags & ZIO_FLAG_NODATA) {
2564 		if (BP_IS_GANG(bp)) {
2565 			zio->io_flags &= ~ZIO_FLAG_NODATA;
2566 		} else {
2567 			ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2568 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2569 		}
2570 	}
2571 
2572 	if (zio_injection_enabled &&
2573 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
2574 		zio_handle_ignored_writes(zio);
2575 
2576 	return (ZIO_PIPELINE_CONTINUE);
2577 }
2578 
2579 static int
2580 zio_done(zio_t *zio)
2581 {
2582 	spa_t *spa = zio->io_spa;
2583 	zio_t *lio = zio->io_logical;
2584 	blkptr_t *bp = zio->io_bp;
2585 	vdev_t *vd = zio->io_vd;
2586 	uint64_t psize = zio->io_size;
2587 	zio_t *pio, *pio_next;
2588 
2589 	/*
2590 	 * If our children haven't all completed,
2591 	 * wait for them and then repeat this pipeline stage.
2592 	 */
2593 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2594 	    zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2595 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2596 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2597 		return (ZIO_PIPELINE_STOP);
2598 
2599 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2600 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2601 			ASSERT(zio->io_children[c][w] == 0);
2602 
2603 	if (bp != NULL) {
2604 		ASSERT(bp->blk_pad[0] == 0);
2605 		ASSERT(bp->blk_pad[1] == 0);
2606 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2607 		    (bp == zio_unique_parent(zio)->io_bp));
2608 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2609 		    zio->io_bp_override == NULL &&
2610 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2611 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
2612 			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2613 			ASSERT(BP_COUNT_GANG(bp) == 0 ||
2614 			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2615 		}
2616 	}
2617 
2618 	/*
2619 	 * If there were child vdev/gang/ddt errors, they apply to us now.
2620 	 */
2621 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2622 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2623 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2624 
2625 	/*
2626 	 * If the I/O on the transformed data was successful, generate any
2627 	 * checksum reports now while we still have the transformed data.
2628 	 */
2629 	if (zio->io_error == 0) {
2630 		while (zio->io_cksum_report != NULL) {
2631 			zio_cksum_report_t *zcr = zio->io_cksum_report;
2632 			uint64_t align = zcr->zcr_align;
2633 			uint64_t asize = P2ROUNDUP(psize, align);
2634 			char *abuf = zio->io_data;
2635 
2636 			if (asize != psize) {
2637 				abuf = zio_buf_alloc(asize);
2638 				bcopy(zio->io_data, abuf, psize);
2639 				bzero(abuf + psize, asize - psize);
2640 			}
2641 
2642 			zio->io_cksum_report = zcr->zcr_next;
2643 			zcr->zcr_next = NULL;
2644 			zcr->zcr_finish(zcr, abuf);
2645 			zfs_ereport_free_checksum(zcr);
2646 
2647 			if (asize != psize)
2648 				zio_buf_free(abuf, asize);
2649 		}
2650 	}
2651 
2652 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
2653 
2654 	vdev_stat_update(zio, psize);
2655 
2656 	if (zio->io_error) {
2657 		/*
2658 		 * If this I/O is attached to a particular vdev,
2659 		 * generate an error message describing the I/O failure
2660 		 * at the block level.  We ignore these errors if the
2661 		 * device is currently unavailable.
2662 		 */
2663 		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2664 			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2665 
2666 		if ((zio->io_error == EIO || !(zio->io_flags &
2667 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2668 		    zio == lio) {
2669 			/*
2670 			 * For logical I/O requests, tell the SPA to log the
2671 			 * error and generate a logical data ereport.
2672 			 */
2673 			spa_log_error(spa, zio);
2674 			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2675 			    0, 0);
2676 		}
2677 	}
2678 
2679 	if (zio->io_error && zio == lio) {
2680 		/*
2681 		 * Determine whether zio should be reexecuted.  This will
2682 		 * propagate all the way to the root via zio_notify_parent().
2683 		 */
2684 		ASSERT(vd == NULL && bp != NULL);
2685 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2686 
2687 		if (IO_IS_ALLOCATING(zio) &&
2688 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2689 			if (zio->io_error != ENOSPC)
2690 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2691 			else
2692 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2693 		}
2694 
2695 		if ((zio->io_type == ZIO_TYPE_READ ||
2696 		    zio->io_type == ZIO_TYPE_FREE) &&
2697 		    zio->io_error == ENXIO &&
2698 		    spa->spa_load_state == SPA_LOAD_NONE &&
2699 		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2700 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2701 
2702 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2703 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2704 
2705 		/*
2706 		 * Here is a possibly good place to attempt to do
2707 		 * either combinatorial reconstruction or error correction
2708 		 * based on checksums.  It also might be a good place
2709 		 * to send out preliminary ereports before we suspend
2710 		 * processing.
2711 		 */
2712 	}
2713 
2714 	/*
2715 	 * If there were logical child errors, they apply to us now.
2716 	 * We defer this until now to avoid conflating logical child
2717 	 * errors with errors that happened to the zio itself when
2718 	 * updating vdev stats and reporting FMA events above.
2719 	 */
2720 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2721 
2722 	if ((zio->io_error || zio->io_reexecute) &&
2723 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2724 	    !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2725 		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2726 
2727 	zio_gang_tree_free(&zio->io_gang_tree);
2728 
2729 	/*
2730 	 * Godfather I/Os should never suspend.
2731 	 */
2732 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2733 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2734 		zio->io_reexecute = 0;
2735 
2736 	if (zio->io_reexecute) {
2737 		/*
2738 		 * This is a logical I/O that wants to reexecute.
2739 		 *
2740 		 * Reexecute is top-down.  When an i/o fails, if it's not
2741 		 * the root, it simply notifies its parent and sticks around.
2742 		 * The parent, seeing that it still has children in zio_done(),
2743 		 * does the same.  This percolates all the way up to the root.
2744 		 * The root i/o will reexecute or suspend the entire tree.
2745 		 *
2746 		 * This approach ensures that zio_reexecute() honors
2747 		 * all the original i/o dependency relationships, e.g.
2748 		 * parents not executing until children are ready.
2749 		 */
2750 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2751 
2752 		zio->io_gang_leader = NULL;
2753 
2754 		mutex_enter(&zio->io_lock);
2755 		zio->io_state[ZIO_WAIT_DONE] = 1;
2756 		mutex_exit(&zio->io_lock);
2757 
2758 		/*
2759 		 * "The Godfather" I/O monitors its children but is
2760 		 * not a true parent to them. It will track them through
2761 		 * the pipeline but severs its ties whenever they get into
2762 		 * trouble (e.g. suspended). This allows "The Godfather"
2763 		 * I/O to return status without blocking.
2764 		 */
2765 		for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2766 			zio_link_t *zl = zio->io_walk_link;
2767 			pio_next = zio_walk_parents(zio);
2768 
2769 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2770 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2771 				zio_remove_child(pio, zio, zl);
2772 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2773 			}
2774 		}
2775 
2776 		if ((pio = zio_unique_parent(zio)) != NULL) {
2777 			/*
2778 			 * We're not a root i/o, so there's nothing to do
2779 			 * but notify our parent.  Don't propagate errors
2780 			 * upward since we haven't permanently failed yet.
2781 			 */
2782 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2783 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2784 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2785 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2786 			/*
2787 			 * We'd fail again if we reexecuted now, so suspend
2788 			 * until conditions improve (e.g. device comes online).
2789 			 */
2790 			zio_suspend(spa, zio);
2791 		} else {
2792 			/*
2793 			 * Reexecution is potentially a huge amount of work.
2794 			 * Hand it off to the otherwise-unused claim taskq.
2795 			 */
2796 			(void) taskq_dispatch(
2797 			    spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2798 			    (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2799 		}
2800 		return (ZIO_PIPELINE_STOP);
2801 	}
2802 
2803 	ASSERT(zio->io_child_count == 0);
2804 	ASSERT(zio->io_reexecute == 0);
2805 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2806 
2807 	/*
2808 	 * Report any checksum errors, since the I/O is complete.
2809 	 */
2810 	while (zio->io_cksum_report != NULL) {
2811 		zio_cksum_report_t *zcr = zio->io_cksum_report;
2812 		zio->io_cksum_report = zcr->zcr_next;
2813 		zcr->zcr_next = NULL;
2814 		zcr->zcr_finish(zcr, NULL);
2815 		zfs_ereport_free_checksum(zcr);
2816 	}
2817 
2818 	/*
2819 	 * It is the responsibility of the done callback to ensure that this
2820 	 * particular zio is no longer discoverable for adoption, and as
2821 	 * such, cannot acquire any new parents.
2822 	 */
2823 	if (zio->io_done)
2824 		zio->io_done(zio);
2825 
2826 	mutex_enter(&zio->io_lock);
2827 	zio->io_state[ZIO_WAIT_DONE] = 1;
2828 	mutex_exit(&zio->io_lock);
2829 
2830 	for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2831 		zio_link_t *zl = zio->io_walk_link;
2832 		pio_next = zio_walk_parents(zio);
2833 		zio_remove_child(pio, zio, zl);
2834 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2835 	}
2836 
2837 	if (zio->io_waiter != NULL) {
2838 		mutex_enter(&zio->io_lock);
2839 		zio->io_executor = NULL;
2840 		cv_broadcast(&zio->io_cv);
2841 		mutex_exit(&zio->io_lock);
2842 	} else {
2843 		zio_destroy(zio);
2844 	}
2845 
2846 	return (ZIO_PIPELINE_STOP);
2847 }
2848 
2849 /*
2850  * ==========================================================================
2851  * I/O pipeline definition
2852  * ==========================================================================
2853  */
2854 static zio_pipe_stage_t *zio_pipeline[] = {
2855 	NULL,
2856 	zio_read_bp_init,
2857 	zio_free_bp_init,
2858 	zio_issue_async,
2859 	zio_write_bp_init,
2860 	zio_checksum_generate,
2861 	zio_ddt_read_start,
2862 	zio_ddt_read_done,
2863 	zio_ddt_write,
2864 	zio_ddt_free,
2865 	zio_gang_assemble,
2866 	zio_gang_issue,
2867 	zio_dva_allocate,
2868 	zio_dva_free,
2869 	zio_dva_claim,
2870 	zio_ready,
2871 	zio_vdev_io_start,
2872 	zio_vdev_io_done,
2873 	zio_vdev_io_assess,
2874 	zio_checksum_verify,
2875 	zio_done
2876 };
2877