xref: /titanic_41/usr/src/uts/common/fs/zfs/zio.c (revision 0db3240d392634cfff2f95fb6da34b56b8dc574f)
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 2010 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 	0,	/* ZIO_PRIORITY_LOG_WRITE	*/
49 	1,	/* ZIO_PRIORITY_CACHE_FILL	*/
50 	1,	/* ZIO_PRIORITY_AGG		*/
51 	4,	/* ZIO_PRIORITY_FREE		*/
52 	4,	/* ZIO_PRIORITY_ASYNC_WRITE	*/
53 	6,	/* ZIO_PRIORITY_ASYNC_READ	*/
54 	10,	/* ZIO_PRIORITY_RESILVER	*/
55 	20,	/* ZIO_PRIORITY_SCRUB		*/
56 };
57 
58 /*
59  * ==========================================================================
60  * I/O type descriptions
61  * ==========================================================================
62  */
63 char *zio_type_name[ZIO_TYPES] = {
64 	"zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
65 	"zio_ioctl"
66 };
67 
68 /*
69  * ==========================================================================
70  * I/O kmem caches
71  * ==========================================================================
72  */
73 kmem_cache_t *zio_cache;
74 kmem_cache_t *zio_link_cache;
75 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
76 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
77 
78 #ifdef _KERNEL
79 extern vmem_t *zio_alloc_arena;
80 #endif
81 
82 /*
83  * An allocating zio is one that either currently has the DVA allocate
84  * stage set or will have it later in its lifetime.
85  */
86 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
87 
88 boolean_t	zio_requeue_io_start_cut_in_line = B_TRUE;
89 
90 #ifdef ZFS_DEBUG
91 int zio_buf_debug_limit = 16384;
92 #else
93 int zio_buf_debug_limit = 0;
94 #endif
95 
96 void
97 zio_init(void)
98 {
99 	size_t c;
100 	vmem_t *data_alloc_arena = NULL;
101 
102 #ifdef _KERNEL
103 	data_alloc_arena = zio_alloc_arena;
104 #endif
105 	zio_cache = kmem_cache_create("zio_cache",
106 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
107 	zio_link_cache = kmem_cache_create("zio_link_cache",
108 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
109 
110 	/*
111 	 * For small buffers, we want a cache for each multiple of
112 	 * SPA_MINBLOCKSIZE.  For medium-size buffers, we want a cache
113 	 * for each quarter-power of 2.  For large buffers, we want
114 	 * a cache for each multiple of PAGESIZE.
115 	 */
116 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
117 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
118 		size_t p2 = size;
119 		size_t align = 0;
120 
121 		while (p2 & (p2 - 1))
122 			p2 &= p2 - 1;
123 
124 		if (size <= 4 * SPA_MINBLOCKSIZE) {
125 			align = SPA_MINBLOCKSIZE;
126 		} else if (P2PHASE(size, PAGESIZE) == 0) {
127 			align = PAGESIZE;
128 		} else if (P2PHASE(size, p2 >> 2) == 0) {
129 			align = p2 >> 2;
130 		}
131 
132 		if (align != 0) {
133 			char name[36];
134 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
135 			zio_buf_cache[c] = kmem_cache_create(name, size,
136 			    align, NULL, NULL, NULL, NULL, NULL,
137 			    size > zio_buf_debug_limit ? KMC_NODEBUG : 0);
138 
139 			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
140 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
141 			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
142 			    size > zio_buf_debug_limit ? KMC_NODEBUG : 0);
143 		}
144 	}
145 
146 	while (--c != 0) {
147 		ASSERT(zio_buf_cache[c] != NULL);
148 		if (zio_buf_cache[c - 1] == NULL)
149 			zio_buf_cache[c - 1] = zio_buf_cache[c];
150 
151 		ASSERT(zio_data_buf_cache[c] != NULL);
152 		if (zio_data_buf_cache[c - 1] == NULL)
153 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
154 	}
155 
156 	zio_inject_init();
157 }
158 
159 void
160 zio_fini(void)
161 {
162 	size_t c;
163 	kmem_cache_t *last_cache = NULL;
164 	kmem_cache_t *last_data_cache = NULL;
165 
166 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
167 		if (zio_buf_cache[c] != last_cache) {
168 			last_cache = zio_buf_cache[c];
169 			kmem_cache_destroy(zio_buf_cache[c]);
170 		}
171 		zio_buf_cache[c] = NULL;
172 
173 		if (zio_data_buf_cache[c] != last_data_cache) {
174 			last_data_cache = zio_data_buf_cache[c];
175 			kmem_cache_destroy(zio_data_buf_cache[c]);
176 		}
177 		zio_data_buf_cache[c] = NULL;
178 	}
179 
180 	kmem_cache_destroy(zio_link_cache);
181 	kmem_cache_destroy(zio_cache);
182 
183 	zio_inject_fini();
184 }
185 
186 /*
187  * ==========================================================================
188  * Allocate and free I/O buffers
189  * ==========================================================================
190  */
191 
192 /*
193  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
194  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
195  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
196  * excess / transient data in-core during a crashdump.
197  */
198 void *
199 zio_buf_alloc(size_t size)
200 {
201 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
202 
203 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
204 
205 	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
206 }
207 
208 /*
209  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
210  * crashdump if the kernel panics.  This exists so that we will limit the amount
211  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
212  * of kernel heap dumped to disk when the kernel panics)
213  */
214 void *
215 zio_data_buf_alloc(size_t size)
216 {
217 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
218 
219 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
220 
221 	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
222 }
223 
224 void
225 zio_buf_free(void *buf, size_t size)
226 {
227 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
228 
229 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
230 
231 	kmem_cache_free(zio_buf_cache[c], buf);
232 }
233 
234 void
235 zio_data_buf_free(void *buf, size_t size)
236 {
237 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
238 
239 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
240 
241 	kmem_cache_free(zio_data_buf_cache[c], buf);
242 }
243 
244 /*
245  * ==========================================================================
246  * Push and pop I/O transform buffers
247  * ==========================================================================
248  */
249 static void
250 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
251 	zio_transform_func_t *transform)
252 {
253 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
254 
255 	zt->zt_orig_data = zio->io_data;
256 	zt->zt_orig_size = zio->io_size;
257 	zt->zt_bufsize = bufsize;
258 	zt->zt_transform = transform;
259 
260 	zt->zt_next = zio->io_transform_stack;
261 	zio->io_transform_stack = zt;
262 
263 	zio->io_data = data;
264 	zio->io_size = size;
265 }
266 
267 static void
268 zio_pop_transforms(zio_t *zio)
269 {
270 	zio_transform_t *zt;
271 
272 	while ((zt = zio->io_transform_stack) != NULL) {
273 		if (zt->zt_transform != NULL)
274 			zt->zt_transform(zio,
275 			    zt->zt_orig_data, zt->zt_orig_size);
276 
277 		if (zt->zt_bufsize != 0)
278 			zio_buf_free(zio->io_data, zt->zt_bufsize);
279 
280 		zio->io_data = zt->zt_orig_data;
281 		zio->io_size = zt->zt_orig_size;
282 		zio->io_transform_stack = zt->zt_next;
283 
284 		kmem_free(zt, sizeof (zio_transform_t));
285 	}
286 }
287 
288 /*
289  * ==========================================================================
290  * I/O transform callbacks for subblocks and decompression
291  * ==========================================================================
292  */
293 static void
294 zio_subblock(zio_t *zio, void *data, uint64_t size)
295 {
296 	ASSERT(zio->io_size > size);
297 
298 	if (zio->io_type == ZIO_TYPE_READ)
299 		bcopy(zio->io_data, data, size);
300 }
301 
302 static void
303 zio_decompress(zio_t *zio, void *data, uint64_t size)
304 {
305 	if (zio->io_error == 0 &&
306 	    zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
307 	    zio->io_data, data, zio->io_size, size) != 0)
308 		zio->io_error = EIO;
309 }
310 
311 /*
312  * ==========================================================================
313  * I/O parent/child relationships and pipeline interlocks
314  * ==========================================================================
315  */
316 /*
317  * NOTE - Callers to zio_walk_parents() and zio_walk_children must
318  *        continue calling these functions until they return NULL.
319  *        Otherwise, the next caller will pick up the list walk in
320  *        some indeterminate state.  (Otherwise every caller would
321  *        have to pass in a cookie to keep the state represented by
322  *        io_walk_link, which gets annoying.)
323  */
324 zio_t *
325 zio_walk_parents(zio_t *cio)
326 {
327 	zio_link_t *zl = cio->io_walk_link;
328 	list_t *pl = &cio->io_parent_list;
329 
330 	zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
331 	cio->io_walk_link = zl;
332 
333 	if (zl == NULL)
334 		return (NULL);
335 
336 	ASSERT(zl->zl_child == cio);
337 	return (zl->zl_parent);
338 }
339 
340 zio_t *
341 zio_walk_children(zio_t *pio)
342 {
343 	zio_link_t *zl = pio->io_walk_link;
344 	list_t *cl = &pio->io_child_list;
345 
346 	zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
347 	pio->io_walk_link = zl;
348 
349 	if (zl == NULL)
350 		return (NULL);
351 
352 	ASSERT(zl->zl_parent == pio);
353 	return (zl->zl_child);
354 }
355 
356 zio_t *
357 zio_unique_parent(zio_t *cio)
358 {
359 	zio_t *pio = zio_walk_parents(cio);
360 
361 	VERIFY(zio_walk_parents(cio) == NULL);
362 	return (pio);
363 }
364 
365 void
366 zio_add_child(zio_t *pio, zio_t *cio)
367 {
368 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
369 
370 	/*
371 	 * Logical I/Os can have logical, gang, or vdev children.
372 	 * Gang I/Os can have gang or vdev children.
373 	 * Vdev I/Os can only have vdev children.
374 	 * The following ASSERT captures all of these constraints.
375 	 */
376 	ASSERT(cio->io_child_type <= pio->io_child_type);
377 
378 	zl->zl_parent = pio;
379 	zl->zl_child = cio;
380 
381 	mutex_enter(&cio->io_lock);
382 	mutex_enter(&pio->io_lock);
383 
384 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
385 
386 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
387 		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
388 
389 	list_insert_head(&pio->io_child_list, zl);
390 	list_insert_head(&cio->io_parent_list, zl);
391 
392 	pio->io_child_count++;
393 	cio->io_parent_count++;
394 
395 	mutex_exit(&pio->io_lock);
396 	mutex_exit(&cio->io_lock);
397 }
398 
399 static void
400 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
401 {
402 	ASSERT(zl->zl_parent == pio);
403 	ASSERT(zl->zl_child == cio);
404 
405 	mutex_enter(&cio->io_lock);
406 	mutex_enter(&pio->io_lock);
407 
408 	list_remove(&pio->io_child_list, zl);
409 	list_remove(&cio->io_parent_list, zl);
410 
411 	pio->io_child_count--;
412 	cio->io_parent_count--;
413 
414 	mutex_exit(&pio->io_lock);
415 	mutex_exit(&cio->io_lock);
416 
417 	kmem_cache_free(zio_link_cache, zl);
418 }
419 
420 static boolean_t
421 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
422 {
423 	uint64_t *countp = &zio->io_children[child][wait];
424 	boolean_t waiting = B_FALSE;
425 
426 	mutex_enter(&zio->io_lock);
427 	ASSERT(zio->io_stall == NULL);
428 	if (*countp != 0) {
429 		zio->io_stage >>= 1;
430 		zio->io_stall = countp;
431 		waiting = B_TRUE;
432 	}
433 	mutex_exit(&zio->io_lock);
434 
435 	return (waiting);
436 }
437 
438 static void
439 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
440 {
441 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
442 	int *errorp = &pio->io_child_error[zio->io_child_type];
443 
444 	mutex_enter(&pio->io_lock);
445 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
446 		*errorp = zio_worst_error(*errorp, zio->io_error);
447 	pio->io_reexecute |= zio->io_reexecute;
448 	ASSERT3U(*countp, >, 0);
449 	if (--*countp == 0 && pio->io_stall == countp) {
450 		pio->io_stall = NULL;
451 		mutex_exit(&pio->io_lock);
452 		zio_execute(pio);
453 	} else {
454 		mutex_exit(&pio->io_lock);
455 	}
456 }
457 
458 static void
459 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
460 {
461 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
462 		zio->io_error = zio->io_child_error[c];
463 }
464 
465 /*
466  * ==========================================================================
467  * Create the various types of I/O (read, write, free, etc)
468  * ==========================================================================
469  */
470 static zio_t *
471 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
472     void *data, uint64_t size, zio_done_func_t *done, void *private,
473     zio_type_t type, int priority, enum zio_flag flags,
474     vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
475     enum zio_stage stage, enum zio_stage pipeline)
476 {
477 	zio_t *zio;
478 
479 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
480 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
481 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
482 
483 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
484 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
485 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
486 
487 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
488 	bzero(zio, sizeof (zio_t));
489 
490 	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
491 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
492 
493 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
494 	    offsetof(zio_link_t, zl_parent_node));
495 	list_create(&zio->io_child_list, sizeof (zio_link_t),
496 	    offsetof(zio_link_t, zl_child_node));
497 
498 	if (vd != NULL)
499 		zio->io_child_type = ZIO_CHILD_VDEV;
500 	else if (flags & ZIO_FLAG_GANG_CHILD)
501 		zio->io_child_type = ZIO_CHILD_GANG;
502 	else if (flags & ZIO_FLAG_DDT_CHILD)
503 		zio->io_child_type = ZIO_CHILD_DDT;
504 	else
505 		zio->io_child_type = ZIO_CHILD_LOGICAL;
506 
507 	if (bp != NULL) {
508 		zio->io_bp = (blkptr_t *)bp;
509 		zio->io_bp_copy = *bp;
510 		zio->io_bp_orig = *bp;
511 		if (type != ZIO_TYPE_WRITE ||
512 		    zio->io_child_type == ZIO_CHILD_DDT)
513 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
514 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
515 			zio->io_logical = zio;
516 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
517 			pipeline |= ZIO_GANG_STAGES;
518 	}
519 
520 	zio->io_spa = spa;
521 	zio->io_txg = txg;
522 	zio->io_done = done;
523 	zio->io_private = private;
524 	zio->io_type = type;
525 	zio->io_priority = priority;
526 	zio->io_vd = vd;
527 	zio->io_offset = offset;
528 	zio->io_orig_data = zio->io_data = data;
529 	zio->io_orig_size = zio->io_size = size;
530 	zio->io_orig_flags = zio->io_flags = flags;
531 	zio->io_orig_stage = zio->io_stage = stage;
532 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
533 
534 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
535 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
536 
537 	if (zb != NULL)
538 		zio->io_bookmark = *zb;
539 
540 	if (pio != NULL) {
541 		if (zio->io_logical == NULL)
542 			zio->io_logical = pio->io_logical;
543 		if (zio->io_child_type == ZIO_CHILD_GANG)
544 			zio->io_gang_leader = pio->io_gang_leader;
545 		zio_add_child(pio, zio);
546 	}
547 
548 	return (zio);
549 }
550 
551 static void
552 zio_destroy(zio_t *zio)
553 {
554 	list_destroy(&zio->io_parent_list);
555 	list_destroy(&zio->io_child_list);
556 	mutex_destroy(&zio->io_lock);
557 	cv_destroy(&zio->io_cv);
558 	kmem_cache_free(zio_cache, zio);
559 }
560 
561 zio_t *
562 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
563     void *private, enum zio_flag flags)
564 {
565 	zio_t *zio;
566 
567 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
568 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
569 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
570 
571 	return (zio);
572 }
573 
574 zio_t *
575 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
576 {
577 	return (zio_null(NULL, spa, NULL, done, private, flags));
578 }
579 
580 zio_t *
581 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
582     void *data, uint64_t size, zio_done_func_t *done, void *private,
583     int priority, enum zio_flag flags, const zbookmark_t *zb)
584 {
585 	zio_t *zio;
586 
587 	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
588 	    data, size, done, private,
589 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
590 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
591 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
592 
593 	return (zio);
594 }
595 
596 zio_t *
597 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
598     void *data, uint64_t size, const zio_prop_t *zp,
599     zio_done_func_t *ready, zio_done_func_t *done, void *private,
600     int priority, enum zio_flag flags, const zbookmark_t *zb)
601 {
602 	zio_t *zio;
603 
604 	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
605 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
606 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
607 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
608 	    zp->zp_type < DMU_OT_NUMTYPES &&
609 	    zp->zp_level < 32 &&
610 	    zp->zp_copies > 0 &&
611 	    zp->zp_copies <= spa_max_replication(spa) &&
612 	    zp->zp_dedup <= 1 &&
613 	    zp->zp_dedup_verify <= 1);
614 
615 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
616 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
617 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
618 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
619 
620 	zio->io_ready = ready;
621 	zio->io_prop = *zp;
622 
623 	return (zio);
624 }
625 
626 zio_t *
627 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
628     uint64_t size, zio_done_func_t *done, void *private, int priority,
629     enum zio_flag flags, zbookmark_t *zb)
630 {
631 	zio_t *zio;
632 
633 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
634 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
635 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
636 
637 	return (zio);
638 }
639 
640 void
641 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
642 {
643 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
644 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
645 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
646 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
647 
648 	zio->io_prop.zp_copies = copies;
649 	zio->io_bp_override = bp;
650 }
651 
652 void
653 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
654 {
655 	bplist_enqueue_deferred(&spa->spa_free_bplist[txg & TXG_MASK], bp);
656 }
657 
658 zio_t *
659 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
660     enum zio_flag flags)
661 {
662 	zio_t *zio;
663 
664 	ASSERT(!BP_IS_HOLE(bp));
665 	ASSERT(spa_syncing_txg(spa) == txg);
666 	ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
667 
668 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
669 	    NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
670 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
671 
672 	return (zio);
673 }
674 
675 zio_t *
676 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
677     zio_done_func_t *done, void *private, enum zio_flag flags)
678 {
679 	zio_t *zio;
680 
681 	/*
682 	 * A claim is an allocation of a specific block.  Claims are needed
683 	 * to support immediate writes in the intent log.  The issue is that
684 	 * immediate writes contain committed data, but in a txg that was
685 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
686 	 * the intent log claims all blocks that contain immediate write data
687 	 * so that the SPA knows they're in use.
688 	 *
689 	 * All claims *must* be resolved in the first txg -- before the SPA
690 	 * starts allocating blocks -- so that nothing is allocated twice.
691 	 * If txg == 0 we just verify that the block is claimable.
692 	 */
693 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
694 	ASSERT(txg == spa_first_txg(spa) || txg == 0);
695 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
696 
697 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
698 	    done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
699 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
700 
701 	return (zio);
702 }
703 
704 zio_t *
705 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
706     zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
707 {
708 	zio_t *zio;
709 	int c;
710 
711 	if (vd->vdev_children == 0) {
712 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
713 		    ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
714 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
715 
716 		zio->io_cmd = cmd;
717 	} else {
718 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
719 
720 		for (c = 0; c < vd->vdev_children; c++)
721 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
722 			    done, private, priority, flags));
723 	}
724 
725 	return (zio);
726 }
727 
728 zio_t *
729 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
730     void *data, int checksum, zio_done_func_t *done, void *private,
731     int priority, enum zio_flag flags, boolean_t labels)
732 {
733 	zio_t *zio;
734 
735 	ASSERT(vd->vdev_children == 0);
736 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
737 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
738 	ASSERT3U(offset + size, <=, vd->vdev_psize);
739 
740 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
741 	    ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
742 	    ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
743 
744 	zio->io_prop.zp_checksum = checksum;
745 
746 	return (zio);
747 }
748 
749 zio_t *
750 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
751     void *data, int checksum, zio_done_func_t *done, void *private,
752     int priority, enum zio_flag flags, boolean_t labels)
753 {
754 	zio_t *zio;
755 
756 	ASSERT(vd->vdev_children == 0);
757 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
758 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
759 	ASSERT3U(offset + size, <=, vd->vdev_psize);
760 
761 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
762 	    ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
763 	    ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
764 
765 	zio->io_prop.zp_checksum = checksum;
766 
767 	if (zio_checksum_table[checksum].ci_eck) {
768 		/*
769 		 * zec checksums are necessarily destructive -- they modify
770 		 * the end of the write buffer to hold the verifier/checksum.
771 		 * Therefore, we must make a local copy in case the data is
772 		 * being written to multiple places in parallel.
773 		 */
774 		void *wbuf = zio_buf_alloc(size);
775 		bcopy(data, wbuf, size);
776 		zio_push_transform(zio, wbuf, size, size, NULL);
777 	}
778 
779 	return (zio);
780 }
781 
782 /*
783  * Create a child I/O to do some work for us.
784  */
785 zio_t *
786 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
787 	void *data, uint64_t size, int type, int priority, enum zio_flag flags,
788 	zio_done_func_t *done, void *private)
789 {
790 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
791 	zio_t *zio;
792 
793 	ASSERT(vd->vdev_parent ==
794 	    (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
795 
796 	if (type == ZIO_TYPE_READ && bp != NULL) {
797 		/*
798 		 * If we have the bp, then the child should perform the
799 		 * checksum and the parent need not.  This pushes error
800 		 * detection as close to the leaves as possible and
801 		 * eliminates redundant checksums in the interior nodes.
802 		 */
803 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
804 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
805 	}
806 
807 	if (vd->vdev_children == 0)
808 		offset += VDEV_LABEL_START_SIZE;
809 
810 	flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
811 
812 	/*
813 	 * If we've decided to do a repair, the write is not speculative --
814 	 * even if the original read was.
815 	 */
816 	if (flags & ZIO_FLAG_IO_REPAIR)
817 		flags &= ~ZIO_FLAG_SPECULATIVE;
818 
819 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
820 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
821 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
822 
823 	return (zio);
824 }
825 
826 zio_t *
827 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
828 	int type, int priority, enum zio_flag flags,
829 	zio_done_func_t *done, void *private)
830 {
831 	zio_t *zio;
832 
833 	ASSERT(vd->vdev_ops->vdev_op_leaf);
834 
835 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
836 	    data, size, done, private, type, priority,
837 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
838 	    vd, offset, NULL,
839 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
840 
841 	return (zio);
842 }
843 
844 void
845 zio_flush(zio_t *zio, vdev_t *vd)
846 {
847 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
848 	    NULL, NULL, ZIO_PRIORITY_NOW,
849 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
850 }
851 
852 void
853 zio_shrink(zio_t *zio, uint64_t size)
854 {
855 	ASSERT(zio->io_executor == NULL);
856 	ASSERT(zio->io_orig_size == zio->io_size);
857 	ASSERT(size <= zio->io_size);
858 
859 	/*
860 	 * We don't shrink for raidz because of problems with the
861 	 * reconstruction when reading back less than the block size.
862 	 * Note, BP_IS_RAIDZ() assumes no compression.
863 	 */
864 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
865 	if (!BP_IS_RAIDZ(zio->io_bp))
866 		zio->io_orig_size = zio->io_size = size;
867 }
868 
869 /*
870  * ==========================================================================
871  * Prepare to read and write logical blocks
872  * ==========================================================================
873  */
874 
875 static int
876 zio_read_bp_init(zio_t *zio)
877 {
878 	blkptr_t *bp = zio->io_bp;
879 
880 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
881 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
882 	    !(zio->io_flags & ZIO_FLAG_RAW)) {
883 		uint64_t psize = BP_GET_PSIZE(bp);
884 		void *cbuf = zio_buf_alloc(psize);
885 
886 		zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
887 	}
888 
889 	if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
890 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
891 
892 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
893 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
894 
895 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
896 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
897 
898 	return (ZIO_PIPELINE_CONTINUE);
899 }
900 
901 static int
902 zio_write_bp_init(zio_t *zio)
903 {
904 	spa_t *spa = zio->io_spa;
905 	zio_prop_t *zp = &zio->io_prop;
906 	enum zio_compress compress = zp->zp_compress;
907 	blkptr_t *bp = zio->io_bp;
908 	uint64_t lsize = zio->io_size;
909 	uint64_t psize = lsize;
910 	int pass = 1;
911 
912 	/*
913 	 * If our children haven't all reached the ready stage,
914 	 * wait for them and then repeat this pipeline stage.
915 	 */
916 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
917 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
918 		return (ZIO_PIPELINE_STOP);
919 
920 	if (!IO_IS_ALLOCATING(zio))
921 		return (ZIO_PIPELINE_CONTINUE);
922 
923 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
924 
925 	if (zio->io_bp_override) {
926 		ASSERT(bp->blk_birth != zio->io_txg);
927 		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
928 
929 		*bp = *zio->io_bp_override;
930 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
931 
932 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
933 			return (ZIO_PIPELINE_CONTINUE);
934 
935 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
936 		    zp->zp_dedup_verify);
937 
938 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
939 			BP_SET_DEDUP(bp, 1);
940 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
941 			return (ZIO_PIPELINE_CONTINUE);
942 		}
943 		zio->io_bp_override = NULL;
944 		BP_ZERO(bp);
945 	}
946 
947 	if (bp->blk_birth == zio->io_txg) {
948 		/*
949 		 * We're rewriting an existing block, which means we're
950 		 * working on behalf of spa_sync().  For spa_sync() to
951 		 * converge, it must eventually be the case that we don't
952 		 * have to allocate new blocks.  But compression changes
953 		 * the blocksize, which forces a reallocate, and makes
954 		 * convergence take longer.  Therefore, after the first
955 		 * few passes, stop compressing to ensure convergence.
956 		 */
957 		pass = spa_sync_pass(spa);
958 
959 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
960 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
961 		ASSERT(!BP_GET_DEDUP(bp));
962 
963 		if (pass > SYNC_PASS_DONT_COMPRESS)
964 			compress = ZIO_COMPRESS_OFF;
965 
966 		/* Make sure someone doesn't change their mind on overwrites */
967 		ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
968 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
969 	}
970 
971 	if (compress != ZIO_COMPRESS_OFF) {
972 		void *cbuf = zio_buf_alloc(lsize);
973 		psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
974 		if (psize == 0 || psize == lsize) {
975 			compress = ZIO_COMPRESS_OFF;
976 			zio_buf_free(cbuf, lsize);
977 		} else {
978 			ASSERT(psize < lsize);
979 			zio_push_transform(zio, cbuf, psize, lsize, NULL);
980 		}
981 	}
982 
983 	/*
984 	 * The final pass of spa_sync() must be all rewrites, but the first
985 	 * few passes offer a trade-off: allocating blocks defers convergence,
986 	 * but newly allocated blocks are sequential, so they can be written
987 	 * to disk faster.  Therefore, we allow the first few passes of
988 	 * spa_sync() to allocate new blocks, but force rewrites after that.
989 	 * There should only be a handful of blocks after pass 1 in any case.
990 	 */
991 	if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
992 	    pass > SYNC_PASS_REWRITE) {
993 		ASSERT(psize != 0);
994 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
995 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
996 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
997 	} else {
998 		BP_ZERO(bp);
999 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1000 	}
1001 
1002 	if (psize == 0) {
1003 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1004 	} else {
1005 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1006 		BP_SET_LSIZE(bp, lsize);
1007 		BP_SET_PSIZE(bp, psize);
1008 		BP_SET_COMPRESS(bp, compress);
1009 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1010 		BP_SET_TYPE(bp, zp->zp_type);
1011 		BP_SET_LEVEL(bp, zp->zp_level);
1012 		BP_SET_DEDUP(bp, zp->zp_dedup);
1013 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1014 		if (zp->zp_dedup) {
1015 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1016 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1017 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1018 		}
1019 	}
1020 
1021 	return (ZIO_PIPELINE_CONTINUE);
1022 }
1023 
1024 static int
1025 zio_free_bp_init(zio_t *zio)
1026 {
1027 	blkptr_t *bp = zio->io_bp;
1028 
1029 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1030 		if (BP_GET_DEDUP(bp))
1031 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1032 		else
1033 			arc_free(zio->io_spa, bp);
1034 	}
1035 
1036 	return (ZIO_PIPELINE_CONTINUE);
1037 }
1038 
1039 /*
1040  * ==========================================================================
1041  * Execute the I/O pipeline
1042  * ==========================================================================
1043  */
1044 
1045 static void
1046 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1047 {
1048 	spa_t *spa = zio->io_spa;
1049 	zio_type_t t = zio->io_type;
1050 	int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0);
1051 
1052 	/*
1053 	 * If we're a config writer or a probe, the normal issue and
1054 	 * interrupt threads may all be blocked waiting for the config lock.
1055 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1056 	 */
1057 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1058 		t = ZIO_TYPE_NULL;
1059 
1060 	/*
1061 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1062 	 */
1063 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1064 		t = ZIO_TYPE_NULL;
1065 
1066 	/*
1067 	 * If this is a high priority I/O, then use the high priority taskq.
1068 	 */
1069 	if (zio->io_priority == ZIO_PRIORITY_NOW &&
1070 	    spa->spa_zio_taskq[t][q + 1] != NULL)
1071 		q++;
1072 
1073 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1074 	(void) taskq_dispatch(spa->spa_zio_taskq[t][q],
1075 	    (task_func_t *)zio_execute, zio, flags);
1076 }
1077 
1078 static boolean_t
1079 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1080 {
1081 	kthread_t *executor = zio->io_executor;
1082 	spa_t *spa = zio->io_spa;
1083 
1084 	for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1085 		if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1086 			return (B_TRUE);
1087 
1088 	return (B_FALSE);
1089 }
1090 
1091 static int
1092 zio_issue_async(zio_t *zio)
1093 {
1094 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1095 
1096 	return (ZIO_PIPELINE_STOP);
1097 }
1098 
1099 void
1100 zio_interrupt(zio_t *zio)
1101 {
1102 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1103 }
1104 
1105 /*
1106  * Execute the I/O pipeline until one of the following occurs:
1107  * (1) the I/O completes; (2) the pipeline stalls waiting for
1108  * dependent child I/Os; (3) the I/O issues, so we're waiting
1109  * for an I/O completion interrupt; (4) the I/O is delegated by
1110  * vdev-level caching or aggregation; (5) the I/O is deferred
1111  * due to vdev-level queueing; (6) the I/O is handed off to
1112  * another thread.  In all cases, the pipeline stops whenever
1113  * there's no CPU work; it never burns a thread in cv_wait().
1114  *
1115  * There's no locking on io_stage because there's no legitimate way
1116  * for multiple threads to be attempting to process the same I/O.
1117  */
1118 static zio_pipe_stage_t *zio_pipeline[];
1119 
1120 void
1121 zio_execute(zio_t *zio)
1122 {
1123 	zio->io_executor = curthread;
1124 
1125 	while (zio->io_stage < ZIO_STAGE_DONE) {
1126 		enum zio_stage pipeline = zio->io_pipeline;
1127 		enum zio_stage stage = zio->io_stage;
1128 		int rv;
1129 
1130 		ASSERT(!MUTEX_HELD(&zio->io_lock));
1131 		ASSERT(ISP2(stage));
1132 		ASSERT(zio->io_stall == NULL);
1133 
1134 		do {
1135 			stage <<= 1;
1136 		} while ((stage & pipeline) == 0);
1137 
1138 		ASSERT(stage <= ZIO_STAGE_DONE);
1139 
1140 		/*
1141 		 * If we are in interrupt context and this pipeline stage
1142 		 * will grab a config lock that is held across I/O,
1143 		 * or may wait for an I/O that needs an interrupt thread
1144 		 * to complete, issue async to avoid deadlock.
1145 		 *
1146 		 * For VDEV_IO_START, we cut in line so that the io will
1147 		 * be sent to disk promptly.
1148 		 */
1149 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1150 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1151 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1152 			    zio_requeue_io_start_cut_in_line : B_FALSE;
1153 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1154 			return;
1155 		}
1156 
1157 		zio->io_stage = stage;
1158 		rv = zio_pipeline[highbit(stage) - 1](zio);
1159 
1160 		if (rv == ZIO_PIPELINE_STOP)
1161 			return;
1162 
1163 		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1164 	}
1165 }
1166 
1167 /*
1168  * ==========================================================================
1169  * Initiate I/O, either sync or async
1170  * ==========================================================================
1171  */
1172 int
1173 zio_wait(zio_t *zio)
1174 {
1175 	int error;
1176 
1177 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1178 	ASSERT(zio->io_executor == NULL);
1179 
1180 	zio->io_waiter = curthread;
1181 
1182 	zio_execute(zio);
1183 
1184 	mutex_enter(&zio->io_lock);
1185 	while (zio->io_executor != NULL)
1186 		cv_wait(&zio->io_cv, &zio->io_lock);
1187 	mutex_exit(&zio->io_lock);
1188 
1189 	error = zio->io_error;
1190 	zio_destroy(zio);
1191 
1192 	return (error);
1193 }
1194 
1195 void
1196 zio_nowait(zio_t *zio)
1197 {
1198 	ASSERT(zio->io_executor == NULL);
1199 
1200 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1201 	    zio_unique_parent(zio) == NULL) {
1202 		/*
1203 		 * This is a logical async I/O with no parent to wait for it.
1204 		 * We add it to the spa_async_root_zio "Godfather" I/O which
1205 		 * will ensure they complete prior to unloading the pool.
1206 		 */
1207 		spa_t *spa = zio->io_spa;
1208 
1209 		zio_add_child(spa->spa_async_zio_root, zio);
1210 	}
1211 
1212 	zio_execute(zio);
1213 }
1214 
1215 /*
1216  * ==========================================================================
1217  * Reexecute or suspend/resume failed I/O
1218  * ==========================================================================
1219  */
1220 
1221 static void
1222 zio_reexecute(zio_t *pio)
1223 {
1224 	zio_t *cio, *cio_next;
1225 
1226 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1227 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1228 	ASSERT(pio->io_gang_leader == NULL);
1229 	ASSERT(pio->io_gang_tree == NULL);
1230 
1231 	pio->io_flags = pio->io_orig_flags;
1232 	pio->io_stage = pio->io_orig_stage;
1233 	pio->io_pipeline = pio->io_orig_pipeline;
1234 	pio->io_reexecute = 0;
1235 	pio->io_error = 0;
1236 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1237 		pio->io_state[w] = 0;
1238 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1239 		pio->io_child_error[c] = 0;
1240 
1241 	if (IO_IS_ALLOCATING(pio))
1242 		BP_ZERO(pio->io_bp);
1243 
1244 	/*
1245 	 * As we reexecute pio's children, new children could be created.
1246 	 * New children go to the head of pio's io_child_list, however,
1247 	 * so we will (correctly) not reexecute them.  The key is that
1248 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1249 	 * cannot be affected by any side effects of reexecuting 'cio'.
1250 	 */
1251 	for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1252 		cio_next = zio_walk_children(pio);
1253 		mutex_enter(&pio->io_lock);
1254 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1255 			pio->io_children[cio->io_child_type][w]++;
1256 		mutex_exit(&pio->io_lock);
1257 		zio_reexecute(cio);
1258 	}
1259 
1260 	/*
1261 	 * Now that all children have been reexecuted, execute the parent.
1262 	 * We don't reexecute "The Godfather" I/O here as it's the
1263 	 * responsibility of the caller to wait on him.
1264 	 */
1265 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1266 		zio_execute(pio);
1267 }
1268 
1269 void
1270 zio_suspend(spa_t *spa, zio_t *zio)
1271 {
1272 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1273 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1274 		    "failure and the failure mode property for this pool "
1275 		    "is set to panic.", spa_name(spa));
1276 
1277 	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1278 
1279 	mutex_enter(&spa->spa_suspend_lock);
1280 
1281 	if (spa->spa_suspend_zio_root == NULL)
1282 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1283 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1284 		    ZIO_FLAG_GODFATHER);
1285 
1286 	spa->spa_suspended = B_TRUE;
1287 
1288 	if (zio != NULL) {
1289 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1290 		ASSERT(zio != spa->spa_suspend_zio_root);
1291 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1292 		ASSERT(zio_unique_parent(zio) == NULL);
1293 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1294 		zio_add_child(spa->spa_suspend_zio_root, zio);
1295 	}
1296 
1297 	mutex_exit(&spa->spa_suspend_lock);
1298 }
1299 
1300 int
1301 zio_resume(spa_t *spa)
1302 {
1303 	zio_t *pio;
1304 
1305 	/*
1306 	 * Reexecute all previously suspended i/o.
1307 	 */
1308 	mutex_enter(&spa->spa_suspend_lock);
1309 	spa->spa_suspended = B_FALSE;
1310 	cv_broadcast(&spa->spa_suspend_cv);
1311 	pio = spa->spa_suspend_zio_root;
1312 	spa->spa_suspend_zio_root = NULL;
1313 	mutex_exit(&spa->spa_suspend_lock);
1314 
1315 	if (pio == NULL)
1316 		return (0);
1317 
1318 	zio_reexecute(pio);
1319 	return (zio_wait(pio));
1320 }
1321 
1322 void
1323 zio_resume_wait(spa_t *spa)
1324 {
1325 	mutex_enter(&spa->spa_suspend_lock);
1326 	while (spa_suspended(spa))
1327 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1328 	mutex_exit(&spa->spa_suspend_lock);
1329 }
1330 
1331 /*
1332  * ==========================================================================
1333  * Gang blocks.
1334  *
1335  * A gang block is a collection of small blocks that looks to the DMU
1336  * like one large block.  When zio_dva_allocate() cannot find a block
1337  * of the requested size, due to either severe fragmentation or the pool
1338  * being nearly full, it calls zio_write_gang_block() to construct the
1339  * block from smaller fragments.
1340  *
1341  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1342  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
1343  * an indirect block: it's an array of block pointers.  It consumes
1344  * only one sector and hence is allocatable regardless of fragmentation.
1345  * The gang header's bps point to its gang members, which hold the data.
1346  *
1347  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1348  * as the verifier to ensure uniqueness of the SHA256 checksum.
1349  * Critically, the gang block bp's blk_cksum is the checksum of the data,
1350  * not the gang header.  This ensures that data block signatures (needed for
1351  * deduplication) are independent of how the block is physically stored.
1352  *
1353  * Gang blocks can be nested: a gang member may itself be a gang block.
1354  * Thus every gang block is a tree in which root and all interior nodes are
1355  * gang headers, and the leaves are normal blocks that contain user data.
1356  * The root of the gang tree is called the gang leader.
1357  *
1358  * To perform any operation (read, rewrite, free, claim) on a gang block,
1359  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1360  * in the io_gang_tree field of the original logical i/o by recursively
1361  * reading the gang leader and all gang headers below it.  This yields
1362  * an in-core tree containing the contents of every gang header and the
1363  * bps for every constituent of the gang block.
1364  *
1365  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1366  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
1367  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1368  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1369  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1370  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
1371  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1372  * of the gang header plus zio_checksum_compute() of the data to update the
1373  * gang header's blk_cksum as described above.
1374  *
1375  * The two-phase assemble/issue model solves the problem of partial failure --
1376  * what if you'd freed part of a gang block but then couldn't read the
1377  * gang header for another part?  Assembling the entire gang tree first
1378  * ensures that all the necessary gang header I/O has succeeded before
1379  * starting the actual work of free, claim, or write.  Once the gang tree
1380  * is assembled, free and claim are in-memory operations that cannot fail.
1381  *
1382  * In the event that a gang write fails, zio_dva_unallocate() walks the
1383  * gang tree to immediately free (i.e. insert back into the space map)
1384  * everything we've allocated.  This ensures that we don't get ENOSPC
1385  * errors during repeated suspend/resume cycles due to a flaky device.
1386  *
1387  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
1388  * the gang tree, we won't modify the block, so we can safely defer the free
1389  * (knowing that the block is still intact).  If we *can* assemble the gang
1390  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1391  * each constituent bp and we can allocate a new block on the next sync pass.
1392  *
1393  * In all cases, the gang tree allows complete recovery from partial failure.
1394  * ==========================================================================
1395  */
1396 
1397 static zio_t *
1398 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1399 {
1400 	if (gn != NULL)
1401 		return (pio);
1402 
1403 	return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1404 	    NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1405 	    &pio->io_bookmark));
1406 }
1407 
1408 zio_t *
1409 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1410 {
1411 	zio_t *zio;
1412 
1413 	if (gn != NULL) {
1414 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1415 		    gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1416 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1417 		/*
1418 		 * As we rewrite each gang header, the pipeline will compute
1419 		 * a new gang block header checksum for it; but no one will
1420 		 * compute a new data checksum, so we do that here.  The one
1421 		 * exception is the gang leader: the pipeline already computed
1422 		 * its data checksum because that stage precedes gang assembly.
1423 		 * (Presently, nothing actually uses interior data checksums;
1424 		 * this is just good hygiene.)
1425 		 */
1426 		if (gn != pio->io_gang_leader->io_gang_tree) {
1427 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1428 			    data, BP_GET_PSIZE(bp));
1429 		}
1430 		/*
1431 		 * If we are here to damage data for testing purposes,
1432 		 * leave the GBH alone so that we can detect the damage.
1433 		 */
1434 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1435 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1436 	} else {
1437 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1438 		    data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1439 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1440 	}
1441 
1442 	return (zio);
1443 }
1444 
1445 /* ARGSUSED */
1446 zio_t *
1447 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1448 {
1449 	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1450 	    ZIO_GANG_CHILD_FLAGS(pio)));
1451 }
1452 
1453 /* ARGSUSED */
1454 zio_t *
1455 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1456 {
1457 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1458 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1459 }
1460 
1461 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1462 	NULL,
1463 	zio_read_gang,
1464 	zio_rewrite_gang,
1465 	zio_free_gang,
1466 	zio_claim_gang,
1467 	NULL
1468 };
1469 
1470 static void zio_gang_tree_assemble_done(zio_t *zio);
1471 
1472 static zio_gang_node_t *
1473 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1474 {
1475 	zio_gang_node_t *gn;
1476 
1477 	ASSERT(*gnpp == NULL);
1478 
1479 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1480 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1481 	*gnpp = gn;
1482 
1483 	return (gn);
1484 }
1485 
1486 static void
1487 zio_gang_node_free(zio_gang_node_t **gnpp)
1488 {
1489 	zio_gang_node_t *gn = *gnpp;
1490 
1491 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1492 		ASSERT(gn->gn_child[g] == NULL);
1493 
1494 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1495 	kmem_free(gn, sizeof (*gn));
1496 	*gnpp = NULL;
1497 }
1498 
1499 static void
1500 zio_gang_tree_free(zio_gang_node_t **gnpp)
1501 {
1502 	zio_gang_node_t *gn = *gnpp;
1503 
1504 	if (gn == NULL)
1505 		return;
1506 
1507 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1508 		zio_gang_tree_free(&gn->gn_child[g]);
1509 
1510 	zio_gang_node_free(gnpp);
1511 }
1512 
1513 static void
1514 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1515 {
1516 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1517 
1518 	ASSERT(gio->io_gang_leader == gio);
1519 	ASSERT(BP_IS_GANG(bp));
1520 
1521 	zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1522 	    SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1523 	    gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1524 }
1525 
1526 static void
1527 zio_gang_tree_assemble_done(zio_t *zio)
1528 {
1529 	zio_t *gio = zio->io_gang_leader;
1530 	zio_gang_node_t *gn = zio->io_private;
1531 	blkptr_t *bp = zio->io_bp;
1532 
1533 	ASSERT(gio == zio_unique_parent(zio));
1534 	ASSERT(zio->io_child_count == 0);
1535 
1536 	if (zio->io_error)
1537 		return;
1538 
1539 	if (BP_SHOULD_BYTESWAP(bp))
1540 		byteswap_uint64_array(zio->io_data, zio->io_size);
1541 
1542 	ASSERT(zio->io_data == gn->gn_gbh);
1543 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1544 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1545 
1546 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1547 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1548 		if (!BP_IS_GANG(gbp))
1549 			continue;
1550 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1551 	}
1552 }
1553 
1554 static void
1555 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1556 {
1557 	zio_t *gio = pio->io_gang_leader;
1558 	zio_t *zio;
1559 
1560 	ASSERT(BP_IS_GANG(bp) == !!gn);
1561 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1562 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1563 
1564 	/*
1565 	 * If you're a gang header, your data is in gn->gn_gbh.
1566 	 * If you're a gang member, your data is in 'data' and gn == NULL.
1567 	 */
1568 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1569 
1570 	if (gn != NULL) {
1571 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1572 
1573 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1574 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1575 			if (BP_IS_HOLE(gbp))
1576 				continue;
1577 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1578 			data = (char *)data + BP_GET_PSIZE(gbp);
1579 		}
1580 	}
1581 
1582 	if (gn == gio->io_gang_tree)
1583 		ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1584 
1585 	if (zio != pio)
1586 		zio_nowait(zio);
1587 }
1588 
1589 static int
1590 zio_gang_assemble(zio_t *zio)
1591 {
1592 	blkptr_t *bp = zio->io_bp;
1593 
1594 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1595 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1596 
1597 	zio->io_gang_leader = zio;
1598 
1599 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1600 
1601 	return (ZIO_PIPELINE_CONTINUE);
1602 }
1603 
1604 static int
1605 zio_gang_issue(zio_t *zio)
1606 {
1607 	blkptr_t *bp = zio->io_bp;
1608 
1609 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1610 		return (ZIO_PIPELINE_STOP);
1611 
1612 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1613 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1614 
1615 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1616 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1617 	else
1618 		zio_gang_tree_free(&zio->io_gang_tree);
1619 
1620 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1621 
1622 	return (ZIO_PIPELINE_CONTINUE);
1623 }
1624 
1625 static void
1626 zio_write_gang_member_ready(zio_t *zio)
1627 {
1628 	zio_t *pio = zio_unique_parent(zio);
1629 	zio_t *gio = zio->io_gang_leader;
1630 	dva_t *cdva = zio->io_bp->blk_dva;
1631 	dva_t *pdva = pio->io_bp->blk_dva;
1632 	uint64_t asize;
1633 
1634 	if (BP_IS_HOLE(zio->io_bp))
1635 		return;
1636 
1637 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1638 
1639 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1640 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1641 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1642 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1643 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1644 
1645 	mutex_enter(&pio->io_lock);
1646 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1647 		ASSERT(DVA_GET_GANG(&pdva[d]));
1648 		asize = DVA_GET_ASIZE(&pdva[d]);
1649 		asize += DVA_GET_ASIZE(&cdva[d]);
1650 		DVA_SET_ASIZE(&pdva[d], asize);
1651 	}
1652 	mutex_exit(&pio->io_lock);
1653 }
1654 
1655 static int
1656 zio_write_gang_block(zio_t *pio)
1657 {
1658 	spa_t *spa = pio->io_spa;
1659 	blkptr_t *bp = pio->io_bp;
1660 	zio_t *gio = pio->io_gang_leader;
1661 	zio_t *zio;
1662 	zio_gang_node_t *gn, **gnpp;
1663 	zio_gbh_phys_t *gbh;
1664 	uint64_t txg = pio->io_txg;
1665 	uint64_t resid = pio->io_size;
1666 	uint64_t lsize;
1667 	int copies = gio->io_prop.zp_copies;
1668 	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1669 	zio_prop_t zp;
1670 	int error;
1671 
1672 	error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1673 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1674 	    METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1675 	if (error) {
1676 		pio->io_error = error;
1677 		return (ZIO_PIPELINE_CONTINUE);
1678 	}
1679 
1680 	if (pio == gio) {
1681 		gnpp = &gio->io_gang_tree;
1682 	} else {
1683 		gnpp = pio->io_private;
1684 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
1685 	}
1686 
1687 	gn = zio_gang_node_alloc(gnpp);
1688 	gbh = gn->gn_gbh;
1689 	bzero(gbh, SPA_GANGBLOCKSIZE);
1690 
1691 	/*
1692 	 * Create the gang header.
1693 	 */
1694 	zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1695 	    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1696 
1697 	/*
1698 	 * Create and nowait the gang children.
1699 	 */
1700 	for (int g = 0; resid != 0; resid -= lsize, g++) {
1701 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1702 		    SPA_MINBLOCKSIZE);
1703 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1704 
1705 		zp.zp_checksum = gio->io_prop.zp_checksum;
1706 		zp.zp_compress = ZIO_COMPRESS_OFF;
1707 		zp.zp_type = DMU_OT_NONE;
1708 		zp.zp_level = 0;
1709 		zp.zp_copies = gio->io_prop.zp_copies;
1710 		zp.zp_dedup = 0;
1711 		zp.zp_dedup_verify = 0;
1712 
1713 		zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1714 		    (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1715 		    zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1716 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1717 		    &pio->io_bookmark));
1718 	}
1719 
1720 	/*
1721 	 * Set pio's pipeline to just wait for zio to finish.
1722 	 */
1723 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1724 
1725 	zio_nowait(zio);
1726 
1727 	return (ZIO_PIPELINE_CONTINUE);
1728 }
1729 
1730 /*
1731  * ==========================================================================
1732  * Dedup
1733  * ==========================================================================
1734  */
1735 static void
1736 zio_ddt_child_read_done(zio_t *zio)
1737 {
1738 	blkptr_t *bp = zio->io_bp;
1739 	ddt_entry_t *dde = zio->io_private;
1740 	ddt_phys_t *ddp;
1741 	zio_t *pio = zio_unique_parent(zio);
1742 
1743 	mutex_enter(&pio->io_lock);
1744 	ddp = ddt_phys_select(dde, bp);
1745 	if (zio->io_error == 0)
1746 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
1747 	if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1748 		dde->dde_repair_data = zio->io_data;
1749 	else
1750 		zio_buf_free(zio->io_data, zio->io_size);
1751 	mutex_exit(&pio->io_lock);
1752 }
1753 
1754 static int
1755 zio_ddt_read_start(zio_t *zio)
1756 {
1757 	blkptr_t *bp = zio->io_bp;
1758 
1759 	ASSERT(BP_GET_DEDUP(bp));
1760 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1761 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1762 
1763 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1764 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1765 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1766 		ddt_phys_t *ddp = dde->dde_phys;
1767 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1768 		blkptr_t blk;
1769 
1770 		ASSERT(zio->io_vsd == NULL);
1771 		zio->io_vsd = dde;
1772 
1773 		if (ddp_self == NULL)
1774 			return (ZIO_PIPELINE_CONTINUE);
1775 
1776 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1777 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1778 				continue;
1779 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1780 			    &blk);
1781 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
1782 			    zio_buf_alloc(zio->io_size), zio->io_size,
1783 			    zio_ddt_child_read_done, dde, zio->io_priority,
1784 			    ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1785 			    &zio->io_bookmark));
1786 		}
1787 		return (ZIO_PIPELINE_CONTINUE);
1788 	}
1789 
1790 	zio_nowait(zio_read(zio, zio->io_spa, bp,
1791 	    zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1792 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1793 
1794 	return (ZIO_PIPELINE_CONTINUE);
1795 }
1796 
1797 static int
1798 zio_ddt_read_done(zio_t *zio)
1799 {
1800 	blkptr_t *bp = zio->io_bp;
1801 
1802 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1803 		return (ZIO_PIPELINE_STOP);
1804 
1805 	ASSERT(BP_GET_DEDUP(bp));
1806 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1807 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1808 
1809 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1810 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1811 		ddt_entry_t *dde = zio->io_vsd;
1812 		if (ddt == NULL) {
1813 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1814 			return (ZIO_PIPELINE_CONTINUE);
1815 		}
1816 		if (dde == NULL) {
1817 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1818 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1819 			return (ZIO_PIPELINE_STOP);
1820 		}
1821 		if (dde->dde_repair_data != NULL) {
1822 			bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1823 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
1824 		}
1825 		ddt_repair_done(ddt, dde);
1826 		zio->io_vsd = NULL;
1827 	}
1828 
1829 	ASSERT(zio->io_vsd == NULL);
1830 
1831 	return (ZIO_PIPELINE_CONTINUE);
1832 }
1833 
1834 static boolean_t
1835 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1836 {
1837 	spa_t *spa = zio->io_spa;
1838 
1839 	/*
1840 	 * Note: we compare the original data, not the transformed data,
1841 	 * because when zio->io_bp is an override bp, we will not have
1842 	 * pushed the I/O transforms.  That's an important optimization
1843 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1844 	 */
1845 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1846 		zio_t *lio = dde->dde_lead_zio[p];
1847 
1848 		if (lio != NULL) {
1849 			return (lio->io_orig_size != zio->io_orig_size ||
1850 			    bcmp(zio->io_orig_data, lio->io_orig_data,
1851 			    zio->io_orig_size) != 0);
1852 		}
1853 	}
1854 
1855 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1856 		ddt_phys_t *ddp = &dde->dde_phys[p];
1857 
1858 		if (ddp->ddp_phys_birth != 0) {
1859 			arc_buf_t *abuf = NULL;
1860 			uint32_t aflags = ARC_WAIT;
1861 			blkptr_t blk = *zio->io_bp;
1862 			int error;
1863 
1864 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1865 
1866 			ddt_exit(ddt);
1867 
1868 			error = arc_read_nolock(NULL, spa, &blk,
1869 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1870 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1871 			    &aflags, &zio->io_bookmark);
1872 
1873 			if (error == 0) {
1874 				if (arc_buf_size(abuf) != zio->io_orig_size ||
1875 				    bcmp(abuf->b_data, zio->io_orig_data,
1876 				    zio->io_orig_size) != 0)
1877 					error = EEXIST;
1878 				VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1879 			}
1880 
1881 			ddt_enter(ddt);
1882 			return (error != 0);
1883 		}
1884 	}
1885 
1886 	return (B_FALSE);
1887 }
1888 
1889 static void
1890 zio_ddt_child_write_ready(zio_t *zio)
1891 {
1892 	int p = zio->io_prop.zp_copies;
1893 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1894 	ddt_entry_t *dde = zio->io_private;
1895 	ddt_phys_t *ddp = &dde->dde_phys[p];
1896 	zio_t *pio;
1897 
1898 	if (zio->io_error)
1899 		return;
1900 
1901 	ddt_enter(ddt);
1902 
1903 	ASSERT(dde->dde_lead_zio[p] == zio);
1904 
1905 	ddt_phys_fill(ddp, zio->io_bp);
1906 
1907 	while ((pio = zio_walk_parents(zio)) != NULL)
1908 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1909 
1910 	ddt_exit(ddt);
1911 }
1912 
1913 static void
1914 zio_ddt_child_write_done(zio_t *zio)
1915 {
1916 	int p = zio->io_prop.zp_copies;
1917 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1918 	ddt_entry_t *dde = zio->io_private;
1919 	ddt_phys_t *ddp = &dde->dde_phys[p];
1920 
1921 	ddt_enter(ddt);
1922 
1923 	ASSERT(ddp->ddp_refcnt == 0);
1924 	ASSERT(dde->dde_lead_zio[p] == zio);
1925 	dde->dde_lead_zio[p] = NULL;
1926 
1927 	if (zio->io_error == 0) {
1928 		while (zio_walk_parents(zio) != NULL)
1929 			ddt_phys_addref(ddp);
1930 	} else {
1931 		ddt_phys_clear(ddp);
1932 	}
1933 
1934 	ddt_exit(ddt);
1935 }
1936 
1937 static void
1938 zio_ddt_ditto_write_done(zio_t *zio)
1939 {
1940 	int p = DDT_PHYS_DITTO;
1941 	zio_prop_t *zp = &zio->io_prop;
1942 	blkptr_t *bp = zio->io_bp;
1943 	ddt_t *ddt = ddt_select(zio->io_spa, bp);
1944 	ddt_entry_t *dde = zio->io_private;
1945 	ddt_phys_t *ddp = &dde->dde_phys[p];
1946 	ddt_key_t *ddk = &dde->dde_key;
1947 
1948 	ddt_enter(ddt);
1949 
1950 	ASSERT(ddp->ddp_refcnt == 0);
1951 	ASSERT(dde->dde_lead_zio[p] == zio);
1952 	dde->dde_lead_zio[p] = NULL;
1953 
1954 	if (zio->io_error == 0) {
1955 		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1956 		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1957 		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1958 		if (ddp->ddp_phys_birth != 0)
1959 			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1960 		ddt_phys_fill(ddp, bp);
1961 	}
1962 
1963 	ddt_exit(ddt);
1964 }
1965 
1966 static int
1967 zio_ddt_write(zio_t *zio)
1968 {
1969 	spa_t *spa = zio->io_spa;
1970 	blkptr_t *bp = zio->io_bp;
1971 	uint64_t txg = zio->io_txg;
1972 	zio_prop_t *zp = &zio->io_prop;
1973 	int p = zp->zp_copies;
1974 	int ditto_copies;
1975 	zio_t *cio = NULL;
1976 	zio_t *dio = NULL;
1977 	ddt_t *ddt = ddt_select(spa, bp);
1978 	ddt_entry_t *dde;
1979 	ddt_phys_t *ddp;
1980 
1981 	ASSERT(BP_GET_DEDUP(bp));
1982 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
1983 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
1984 
1985 	ddt_enter(ddt);
1986 	dde = ddt_lookup(ddt, bp, B_TRUE);
1987 	ddp = &dde->dde_phys[p];
1988 
1989 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
1990 		/*
1991 		 * If we're using a weak checksum, upgrade to a strong checksum
1992 		 * and try again.  If we're already using a strong checksum,
1993 		 * we can't resolve it, so just convert to an ordinary write.
1994 		 * (And automatically e-mail a paper to Nature?)
1995 		 */
1996 		if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
1997 			zp->zp_checksum = spa_dedup_checksum(spa);
1998 			zio_pop_transforms(zio);
1999 			zio->io_stage = ZIO_STAGE_OPEN;
2000 			BP_ZERO(bp);
2001 		} else {
2002 			zp->zp_dedup = 0;
2003 		}
2004 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2005 		ddt_exit(ddt);
2006 		return (ZIO_PIPELINE_CONTINUE);
2007 	}
2008 
2009 	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2010 	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2011 
2012 	if (ditto_copies > ddt_ditto_copies_present(dde) &&
2013 	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2014 		zio_prop_t czp = *zp;
2015 
2016 		czp.zp_copies = ditto_copies;
2017 
2018 		/*
2019 		 * If we arrived here with an override bp, we won't have run
2020 		 * the transform stack, so we won't have the data we need to
2021 		 * generate a child i/o.  So, toss the override bp and restart.
2022 		 * This is safe, because using the override bp is just an
2023 		 * optimization; and it's rare, so the cost doesn't matter.
2024 		 */
2025 		if (zio->io_bp_override) {
2026 			zio_pop_transforms(zio);
2027 			zio->io_stage = ZIO_STAGE_OPEN;
2028 			zio->io_pipeline = ZIO_WRITE_PIPELINE;
2029 			zio->io_bp_override = NULL;
2030 			BP_ZERO(bp);
2031 			ddt_exit(ddt);
2032 			return (ZIO_PIPELINE_CONTINUE);
2033 		}
2034 
2035 		dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2036 		    zio->io_orig_size, &czp, NULL,
2037 		    zio_ddt_ditto_write_done, dde, zio->io_priority,
2038 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2039 
2040 		zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2041 		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2042 	}
2043 
2044 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2045 		if (ddp->ddp_phys_birth != 0)
2046 			ddt_bp_fill(ddp, bp, txg);
2047 		if (dde->dde_lead_zio[p] != NULL)
2048 			zio_add_child(zio, dde->dde_lead_zio[p]);
2049 		else
2050 			ddt_phys_addref(ddp);
2051 	} else if (zio->io_bp_override) {
2052 		ASSERT(bp->blk_birth == txg);
2053 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2054 		ddt_phys_fill(ddp, bp);
2055 		ddt_phys_addref(ddp);
2056 	} else {
2057 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2058 		    zio->io_orig_size, zp, zio_ddt_child_write_ready,
2059 		    zio_ddt_child_write_done, dde, zio->io_priority,
2060 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2061 
2062 		zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2063 		dde->dde_lead_zio[p] = cio;
2064 	}
2065 
2066 	ddt_exit(ddt);
2067 
2068 	if (cio)
2069 		zio_nowait(cio);
2070 	if (dio)
2071 		zio_nowait(dio);
2072 
2073 	return (ZIO_PIPELINE_CONTINUE);
2074 }
2075 
2076 static int
2077 zio_ddt_free(zio_t *zio)
2078 {
2079 	spa_t *spa = zio->io_spa;
2080 	blkptr_t *bp = zio->io_bp;
2081 	ddt_t *ddt = ddt_select(spa, bp);
2082 	ddt_entry_t *dde;
2083 	ddt_phys_t *ddp;
2084 
2085 	ASSERT(BP_GET_DEDUP(bp));
2086 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2087 
2088 	ddt_enter(ddt);
2089 	dde = ddt_lookup(ddt, bp, B_TRUE);
2090 	ddp = ddt_phys_select(dde, bp);
2091 	ddt_phys_decref(ddp);
2092 	ddt_exit(ddt);
2093 
2094 	return (ZIO_PIPELINE_CONTINUE);
2095 }
2096 
2097 /*
2098  * ==========================================================================
2099  * Allocate and free blocks
2100  * ==========================================================================
2101  */
2102 static int
2103 zio_dva_allocate(zio_t *zio)
2104 {
2105 	spa_t *spa = zio->io_spa;
2106 	metaslab_class_t *mc = spa_normal_class(spa);
2107 	blkptr_t *bp = zio->io_bp;
2108 	int error;
2109 
2110 	if (zio->io_gang_leader == NULL) {
2111 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2112 		zio->io_gang_leader = zio;
2113 	}
2114 
2115 	ASSERT(BP_IS_HOLE(bp));
2116 	ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2117 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
2118 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2119 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2120 
2121 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
2122 	    zio->io_prop.zp_copies, zio->io_txg, NULL, 0);
2123 
2124 	if (error) {
2125 		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2126 			return (zio_write_gang_block(zio));
2127 		zio->io_error = error;
2128 	}
2129 
2130 	return (ZIO_PIPELINE_CONTINUE);
2131 }
2132 
2133 static int
2134 zio_dva_free(zio_t *zio)
2135 {
2136 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2137 
2138 	return (ZIO_PIPELINE_CONTINUE);
2139 }
2140 
2141 static int
2142 zio_dva_claim(zio_t *zio)
2143 {
2144 	int error;
2145 
2146 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2147 	if (error)
2148 		zio->io_error = error;
2149 
2150 	return (ZIO_PIPELINE_CONTINUE);
2151 }
2152 
2153 /*
2154  * Undo an allocation.  This is used by zio_done() when an I/O fails
2155  * and we want to give back the block we just allocated.
2156  * This handles both normal blocks and gang blocks.
2157  */
2158 static void
2159 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2160 {
2161 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2162 	ASSERT(zio->io_bp_override == NULL);
2163 
2164 	if (!BP_IS_HOLE(bp))
2165 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2166 
2167 	if (gn != NULL) {
2168 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2169 			zio_dva_unallocate(zio, gn->gn_child[g],
2170 			    &gn->gn_gbh->zg_blkptr[g]);
2171 		}
2172 	}
2173 }
2174 
2175 /*
2176  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
2177  */
2178 int
2179 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2180     uint64_t size, boolean_t use_slog)
2181 {
2182 	int error = 1;
2183 
2184 	ASSERT(txg > spa_syncing_txg(spa));
2185 
2186 	if (use_slog)
2187 		error = metaslab_alloc(spa, spa_log_class(spa), size,
2188 		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2189 
2190 	if (error)
2191 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
2192 		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2193 
2194 	if (error == 0) {
2195 		BP_SET_LSIZE(new_bp, size);
2196 		BP_SET_PSIZE(new_bp, size);
2197 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2198 		BP_SET_CHECKSUM(new_bp,
2199 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2200 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2201 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2202 		BP_SET_LEVEL(new_bp, 0);
2203 		BP_SET_DEDUP(new_bp, 0);
2204 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2205 	}
2206 
2207 	return (error);
2208 }
2209 
2210 /*
2211  * Free an intent log block.
2212  */
2213 void
2214 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2215 {
2216 	ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2217 	ASSERT(!BP_IS_GANG(bp));
2218 
2219 	zio_free(spa, txg, bp);
2220 }
2221 
2222 /*
2223  * ==========================================================================
2224  * Read and write to physical devices
2225  * ==========================================================================
2226  */
2227 static int
2228 zio_vdev_io_start(zio_t *zio)
2229 {
2230 	vdev_t *vd = zio->io_vd;
2231 	uint64_t align;
2232 	spa_t *spa = zio->io_spa;
2233 
2234 	ASSERT(zio->io_error == 0);
2235 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2236 
2237 	if (vd == NULL) {
2238 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2239 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2240 
2241 		/*
2242 		 * The mirror_ops handle multiple DVAs in a single BP.
2243 		 */
2244 		return (vdev_mirror_ops.vdev_op_io_start(zio));
2245 	}
2246 
2247 	align = 1ULL << vd->vdev_top->vdev_ashift;
2248 
2249 	if (P2PHASE(zio->io_size, align) != 0) {
2250 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
2251 		char *abuf = zio_buf_alloc(asize);
2252 		ASSERT(vd == vd->vdev_top);
2253 		if (zio->io_type == ZIO_TYPE_WRITE) {
2254 			bcopy(zio->io_data, abuf, zio->io_size);
2255 			bzero(abuf + zio->io_size, asize - zio->io_size);
2256 		}
2257 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2258 	}
2259 
2260 	ASSERT(P2PHASE(zio->io_offset, align) == 0);
2261 	ASSERT(P2PHASE(zio->io_size, align) == 0);
2262 	ASSERT(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2263 
2264 	/*
2265 	 * If this is a repair I/O, and there's no self-healing involved --
2266 	 * that is, we're just resilvering what we expect to resilver --
2267 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2268 	 * This prevents spurious resilvering with nested replication.
2269 	 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2270 	 * A is out of date, we'll read from C+D, then use the data to
2271 	 * resilver A+B -- but we don't actually want to resilver B, just A.
2272 	 * The top-level mirror has no way to know this, so instead we just
2273 	 * discard unnecessary repairs as we work our way down the vdev tree.
2274 	 * The same logic applies to any form of nested replication:
2275 	 * ditto + mirror, RAID-Z + replacing, etc.  This covers them all.
2276 	 */
2277 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2278 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2279 	    zio->io_txg != 0 &&	/* not a delegated i/o */
2280 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2281 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2282 		zio_vdev_io_bypass(zio);
2283 		return (ZIO_PIPELINE_CONTINUE);
2284 	}
2285 
2286 	if (vd->vdev_ops->vdev_op_leaf &&
2287 	    (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2288 
2289 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2290 			return (ZIO_PIPELINE_CONTINUE);
2291 
2292 		if ((zio = vdev_queue_io(zio)) == NULL)
2293 			return (ZIO_PIPELINE_STOP);
2294 
2295 		if (!vdev_accessible(vd, zio)) {
2296 			zio->io_error = ENXIO;
2297 			zio_interrupt(zio);
2298 			return (ZIO_PIPELINE_STOP);
2299 		}
2300 	}
2301 
2302 	return (vd->vdev_ops->vdev_op_io_start(zio));
2303 }
2304 
2305 static int
2306 zio_vdev_io_done(zio_t *zio)
2307 {
2308 	vdev_t *vd = zio->io_vd;
2309 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2310 	boolean_t unexpected_error = B_FALSE;
2311 
2312 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2313 		return (ZIO_PIPELINE_STOP);
2314 
2315 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2316 
2317 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2318 
2319 		vdev_queue_io_done(zio);
2320 
2321 		if (zio->io_type == ZIO_TYPE_WRITE)
2322 			vdev_cache_write(zio);
2323 
2324 		if (zio_injection_enabled && zio->io_error == 0)
2325 			zio->io_error = zio_handle_device_injection(vd,
2326 			    zio, EIO);
2327 
2328 		if (zio_injection_enabled && zio->io_error == 0)
2329 			zio->io_error = zio_handle_label_injection(zio, EIO);
2330 
2331 		if (zio->io_error) {
2332 			if (!vdev_accessible(vd, zio)) {
2333 				zio->io_error = ENXIO;
2334 			} else {
2335 				unexpected_error = B_TRUE;
2336 			}
2337 		}
2338 	}
2339 
2340 	ops->vdev_op_io_done(zio);
2341 
2342 	if (unexpected_error)
2343 		VERIFY(vdev_probe(vd, zio) == NULL);
2344 
2345 	return (ZIO_PIPELINE_CONTINUE);
2346 }
2347 
2348 /*
2349  * For non-raidz ZIOs, we can just copy aside the bad data read from the
2350  * disk, and use that to finish the checksum ereport later.
2351  */
2352 static void
2353 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2354     const void *good_buf)
2355 {
2356 	/* no processing needed */
2357 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2358 }
2359 
2360 /*ARGSUSED*/
2361 void
2362 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2363 {
2364 	void *buf = zio_buf_alloc(zio->io_size);
2365 
2366 	bcopy(zio->io_data, buf, zio->io_size);
2367 
2368 	zcr->zcr_cbinfo = zio->io_size;
2369 	zcr->zcr_cbdata = buf;
2370 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
2371 	zcr->zcr_free = zio_buf_free;
2372 }
2373 
2374 static int
2375 zio_vdev_io_assess(zio_t *zio)
2376 {
2377 	vdev_t *vd = zio->io_vd;
2378 
2379 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2380 		return (ZIO_PIPELINE_STOP);
2381 
2382 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2383 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2384 
2385 	if (zio->io_vsd != NULL) {
2386 		zio->io_vsd_ops->vsd_free(zio);
2387 		zio->io_vsd = NULL;
2388 	}
2389 
2390 	if (zio_injection_enabled && zio->io_error == 0)
2391 		zio->io_error = zio_handle_fault_injection(zio, EIO);
2392 
2393 	/*
2394 	 * If the I/O failed, determine whether we should attempt to retry it.
2395 	 *
2396 	 * On retry, we cut in line in the issue queue, since we don't want
2397 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2398 	 */
2399 	if (zio->io_error && vd == NULL &&
2400 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2401 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
2402 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
2403 		zio->io_error = 0;
2404 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
2405 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2406 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2407 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2408 		    zio_requeue_io_start_cut_in_line);
2409 		return (ZIO_PIPELINE_STOP);
2410 	}
2411 
2412 	/*
2413 	 * If we got an error on a leaf device, convert it to ENXIO
2414 	 * if the device is not accessible at all.
2415 	 */
2416 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2417 	    !vdev_accessible(vd, zio))
2418 		zio->io_error = ENXIO;
2419 
2420 	/*
2421 	 * If we can't write to an interior vdev (mirror or RAID-Z),
2422 	 * set vdev_cant_write so that we stop trying to allocate from it.
2423 	 */
2424 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2425 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2426 		vd->vdev_cant_write = B_TRUE;
2427 
2428 	if (zio->io_error)
2429 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2430 
2431 	return (ZIO_PIPELINE_CONTINUE);
2432 }
2433 
2434 void
2435 zio_vdev_io_reissue(zio_t *zio)
2436 {
2437 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2438 	ASSERT(zio->io_error == 0);
2439 
2440 	zio->io_stage >>= 1;
2441 }
2442 
2443 void
2444 zio_vdev_io_redone(zio_t *zio)
2445 {
2446 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2447 
2448 	zio->io_stage >>= 1;
2449 }
2450 
2451 void
2452 zio_vdev_io_bypass(zio_t *zio)
2453 {
2454 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2455 	ASSERT(zio->io_error == 0);
2456 
2457 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2458 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2459 }
2460 
2461 /*
2462  * ==========================================================================
2463  * Generate and verify checksums
2464  * ==========================================================================
2465  */
2466 static int
2467 zio_checksum_generate(zio_t *zio)
2468 {
2469 	blkptr_t *bp = zio->io_bp;
2470 	enum zio_checksum checksum;
2471 
2472 	if (bp == NULL) {
2473 		/*
2474 		 * This is zio_write_phys().
2475 		 * We're either generating a label checksum, or none at all.
2476 		 */
2477 		checksum = zio->io_prop.zp_checksum;
2478 
2479 		if (checksum == ZIO_CHECKSUM_OFF)
2480 			return (ZIO_PIPELINE_CONTINUE);
2481 
2482 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2483 	} else {
2484 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2485 			ASSERT(!IO_IS_ALLOCATING(zio));
2486 			checksum = ZIO_CHECKSUM_GANG_HEADER;
2487 		} else {
2488 			checksum = BP_GET_CHECKSUM(bp);
2489 		}
2490 	}
2491 
2492 	zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2493 
2494 	return (ZIO_PIPELINE_CONTINUE);
2495 }
2496 
2497 static int
2498 zio_checksum_verify(zio_t *zio)
2499 {
2500 	zio_bad_cksum_t info;
2501 	blkptr_t *bp = zio->io_bp;
2502 	int error;
2503 
2504 	ASSERT(zio->io_vd != NULL);
2505 
2506 	if (bp == NULL) {
2507 		/*
2508 		 * This is zio_read_phys().
2509 		 * We're either verifying a label checksum, or nothing at all.
2510 		 */
2511 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2512 			return (ZIO_PIPELINE_CONTINUE);
2513 
2514 		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2515 	}
2516 
2517 	if ((error = zio_checksum_error(zio, &info)) != 0) {
2518 		zio->io_error = error;
2519 		if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2520 			zfs_ereport_start_checksum(zio->io_spa,
2521 			    zio->io_vd, zio, zio->io_offset,
2522 			    zio->io_size, NULL, &info);
2523 		}
2524 	}
2525 
2526 	return (ZIO_PIPELINE_CONTINUE);
2527 }
2528 
2529 /*
2530  * Called by RAID-Z to ensure we don't compute the checksum twice.
2531  */
2532 void
2533 zio_checksum_verified(zio_t *zio)
2534 {
2535 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2536 }
2537 
2538 /*
2539  * ==========================================================================
2540  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2541  * An error of 0 indictes success.  ENXIO indicates whole-device failure,
2542  * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
2543  * indicate errors that are specific to one I/O, and most likely permanent.
2544  * Any other error is presumed to be worse because we weren't expecting it.
2545  * ==========================================================================
2546  */
2547 int
2548 zio_worst_error(int e1, int e2)
2549 {
2550 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2551 	int r1, r2;
2552 
2553 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2554 		if (e1 == zio_error_rank[r1])
2555 			break;
2556 
2557 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2558 		if (e2 == zio_error_rank[r2])
2559 			break;
2560 
2561 	return (r1 > r2 ? e1 : e2);
2562 }
2563 
2564 /*
2565  * ==========================================================================
2566  * I/O completion
2567  * ==========================================================================
2568  */
2569 static int
2570 zio_ready(zio_t *zio)
2571 {
2572 	blkptr_t *bp = zio->io_bp;
2573 	zio_t *pio, *pio_next;
2574 
2575 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2576 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2577 		return (ZIO_PIPELINE_STOP);
2578 
2579 	if (zio->io_ready) {
2580 		ASSERT(IO_IS_ALLOCATING(zio));
2581 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2582 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2583 
2584 		zio->io_ready(zio);
2585 	}
2586 
2587 	if (bp != NULL && bp != &zio->io_bp_copy)
2588 		zio->io_bp_copy = *bp;
2589 
2590 	if (zio->io_error)
2591 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2592 
2593 	mutex_enter(&zio->io_lock);
2594 	zio->io_state[ZIO_WAIT_READY] = 1;
2595 	pio = zio_walk_parents(zio);
2596 	mutex_exit(&zio->io_lock);
2597 
2598 	/*
2599 	 * As we notify zio's parents, new parents could be added.
2600 	 * New parents go to the head of zio's io_parent_list, however,
2601 	 * so we will (correctly) not notify them.  The remainder of zio's
2602 	 * io_parent_list, from 'pio_next' onward, cannot change because
2603 	 * all parents must wait for us to be done before they can be done.
2604 	 */
2605 	for (; pio != NULL; pio = pio_next) {
2606 		pio_next = zio_walk_parents(zio);
2607 		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2608 	}
2609 
2610 	if (zio->io_flags & ZIO_FLAG_NODATA) {
2611 		if (BP_IS_GANG(bp)) {
2612 			zio->io_flags &= ~ZIO_FLAG_NODATA;
2613 		} else {
2614 			ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2615 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2616 		}
2617 	}
2618 
2619 	if (zio_injection_enabled &&
2620 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
2621 		zio_handle_ignored_writes(zio);
2622 
2623 	return (ZIO_PIPELINE_CONTINUE);
2624 }
2625 
2626 static int
2627 zio_done(zio_t *zio)
2628 {
2629 	spa_t *spa = zio->io_spa;
2630 	zio_t *lio = zio->io_logical;
2631 	blkptr_t *bp = zio->io_bp;
2632 	vdev_t *vd = zio->io_vd;
2633 	uint64_t psize = zio->io_size;
2634 	zio_t *pio, *pio_next;
2635 
2636 	/*
2637 	 * If our children haven't all completed,
2638 	 * wait for them and then repeat this pipeline stage.
2639 	 */
2640 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2641 	    zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2642 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2643 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2644 		return (ZIO_PIPELINE_STOP);
2645 
2646 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2647 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2648 			ASSERT(zio->io_children[c][w] == 0);
2649 
2650 	if (bp != NULL) {
2651 		ASSERT(bp->blk_pad[0] == 0);
2652 		ASSERT(bp->blk_pad[1] == 0);
2653 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2654 		    (bp == zio_unique_parent(zio)->io_bp));
2655 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2656 		    zio->io_bp_override == NULL &&
2657 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2658 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
2659 			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2660 			ASSERT(BP_COUNT_GANG(bp) == 0 ||
2661 			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2662 		}
2663 	}
2664 
2665 	/*
2666 	 * If there were child vdev/gang/ddt errors, they apply to us now.
2667 	 */
2668 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2669 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2670 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2671 
2672 	/*
2673 	 * If the I/O on the transformed data was successful, generate any
2674 	 * checksum reports now while we still have the transformed data.
2675 	 */
2676 	if (zio->io_error == 0) {
2677 		while (zio->io_cksum_report != NULL) {
2678 			zio_cksum_report_t *zcr = zio->io_cksum_report;
2679 			uint64_t align = zcr->zcr_align;
2680 			uint64_t asize = P2ROUNDUP(psize, align);
2681 			char *abuf = zio->io_data;
2682 
2683 			if (asize != psize) {
2684 				abuf = zio_buf_alloc(asize);
2685 				bcopy(zio->io_data, abuf, psize);
2686 				bzero(abuf + psize, asize - psize);
2687 			}
2688 
2689 			zio->io_cksum_report = zcr->zcr_next;
2690 			zcr->zcr_next = NULL;
2691 			zcr->zcr_finish(zcr, abuf);
2692 			zfs_ereport_free_checksum(zcr);
2693 
2694 			if (asize != psize)
2695 				zio_buf_free(abuf, asize);
2696 		}
2697 	}
2698 
2699 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
2700 
2701 	vdev_stat_update(zio, psize);
2702 
2703 	if (zio->io_error) {
2704 		/*
2705 		 * If this I/O is attached to a particular vdev,
2706 		 * generate an error message describing the I/O failure
2707 		 * at the block level.  We ignore these errors if the
2708 		 * device is currently unavailable.
2709 		 */
2710 		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2711 			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2712 
2713 		if ((zio->io_error == EIO || !(zio->io_flags &
2714 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2715 		    zio == lio) {
2716 			/*
2717 			 * For logical I/O requests, tell the SPA to log the
2718 			 * error and generate a logical data ereport.
2719 			 */
2720 			spa_log_error(spa, zio);
2721 			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2722 			    0, 0);
2723 		}
2724 	}
2725 
2726 	if (zio->io_error && zio == lio) {
2727 		/*
2728 		 * Determine whether zio should be reexecuted.  This will
2729 		 * propagate all the way to the root via zio_notify_parent().
2730 		 */
2731 		ASSERT(vd == NULL && bp != NULL);
2732 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2733 
2734 		if (IO_IS_ALLOCATING(zio) &&
2735 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2736 			if (zio->io_error != ENOSPC)
2737 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2738 			else
2739 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2740 		}
2741 
2742 		if ((zio->io_type == ZIO_TYPE_READ ||
2743 		    zio->io_type == ZIO_TYPE_FREE) &&
2744 		    zio->io_error == ENXIO &&
2745 		    spa_load_state(spa) == SPA_LOAD_NONE &&
2746 		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2747 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2748 
2749 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2750 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2751 
2752 		/*
2753 		 * Here is a possibly good place to attempt to do
2754 		 * either combinatorial reconstruction or error correction
2755 		 * based on checksums.  It also might be a good place
2756 		 * to send out preliminary ereports before we suspend
2757 		 * processing.
2758 		 */
2759 	}
2760 
2761 	/*
2762 	 * If there were logical child errors, they apply to us now.
2763 	 * We defer this until now to avoid conflating logical child
2764 	 * errors with errors that happened to the zio itself when
2765 	 * updating vdev stats and reporting FMA events above.
2766 	 */
2767 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2768 
2769 	if ((zio->io_error || zio->io_reexecute) &&
2770 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2771 	    !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2772 		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2773 
2774 	zio_gang_tree_free(&zio->io_gang_tree);
2775 
2776 	/*
2777 	 * Godfather I/Os should never suspend.
2778 	 */
2779 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2780 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2781 		zio->io_reexecute = 0;
2782 
2783 	if (zio->io_reexecute) {
2784 		/*
2785 		 * This is a logical I/O that wants to reexecute.
2786 		 *
2787 		 * Reexecute is top-down.  When an i/o fails, if it's not
2788 		 * the root, it simply notifies its parent and sticks around.
2789 		 * The parent, seeing that it still has children in zio_done(),
2790 		 * does the same.  This percolates all the way up to the root.
2791 		 * The root i/o will reexecute or suspend the entire tree.
2792 		 *
2793 		 * This approach ensures that zio_reexecute() honors
2794 		 * all the original i/o dependency relationships, e.g.
2795 		 * parents not executing until children are ready.
2796 		 */
2797 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2798 
2799 		zio->io_gang_leader = NULL;
2800 
2801 		mutex_enter(&zio->io_lock);
2802 		zio->io_state[ZIO_WAIT_DONE] = 1;
2803 		mutex_exit(&zio->io_lock);
2804 
2805 		/*
2806 		 * "The Godfather" I/O monitors its children but is
2807 		 * not a true parent to them. It will track them through
2808 		 * the pipeline but severs its ties whenever they get into
2809 		 * trouble (e.g. suspended). This allows "The Godfather"
2810 		 * I/O to return status without blocking.
2811 		 */
2812 		for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2813 			zio_link_t *zl = zio->io_walk_link;
2814 			pio_next = zio_walk_parents(zio);
2815 
2816 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2817 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2818 				zio_remove_child(pio, zio, zl);
2819 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2820 			}
2821 		}
2822 
2823 		if ((pio = zio_unique_parent(zio)) != NULL) {
2824 			/*
2825 			 * We're not a root i/o, so there's nothing to do
2826 			 * but notify our parent.  Don't propagate errors
2827 			 * upward since we haven't permanently failed yet.
2828 			 */
2829 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2830 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2831 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2832 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2833 			/*
2834 			 * We'd fail again if we reexecuted now, so suspend
2835 			 * until conditions improve (e.g. device comes online).
2836 			 */
2837 			zio_suspend(spa, zio);
2838 		} else {
2839 			/*
2840 			 * Reexecution is potentially a huge amount of work.
2841 			 * Hand it off to the otherwise-unused claim taskq.
2842 			 */
2843 			(void) taskq_dispatch(
2844 			    spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2845 			    (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2846 		}
2847 		return (ZIO_PIPELINE_STOP);
2848 	}
2849 
2850 	ASSERT(zio->io_child_count == 0);
2851 	ASSERT(zio->io_reexecute == 0);
2852 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2853 
2854 	/*
2855 	 * Report any checksum errors, since the I/O is complete.
2856 	 */
2857 	while (zio->io_cksum_report != NULL) {
2858 		zio_cksum_report_t *zcr = zio->io_cksum_report;
2859 		zio->io_cksum_report = zcr->zcr_next;
2860 		zcr->zcr_next = NULL;
2861 		zcr->zcr_finish(zcr, NULL);
2862 		zfs_ereport_free_checksum(zcr);
2863 	}
2864 
2865 	/*
2866 	 * It is the responsibility of the done callback to ensure that this
2867 	 * particular zio is no longer discoverable for adoption, and as
2868 	 * such, cannot acquire any new parents.
2869 	 */
2870 	if (zio->io_done)
2871 		zio->io_done(zio);
2872 
2873 	mutex_enter(&zio->io_lock);
2874 	zio->io_state[ZIO_WAIT_DONE] = 1;
2875 	mutex_exit(&zio->io_lock);
2876 
2877 	for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2878 		zio_link_t *zl = zio->io_walk_link;
2879 		pio_next = zio_walk_parents(zio);
2880 		zio_remove_child(pio, zio, zl);
2881 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2882 	}
2883 
2884 	if (zio->io_waiter != NULL) {
2885 		mutex_enter(&zio->io_lock);
2886 		zio->io_executor = NULL;
2887 		cv_broadcast(&zio->io_cv);
2888 		mutex_exit(&zio->io_lock);
2889 	} else {
2890 		zio_destroy(zio);
2891 	}
2892 
2893 	return (ZIO_PIPELINE_STOP);
2894 }
2895 
2896 /*
2897  * ==========================================================================
2898  * I/O pipeline definition
2899  * ==========================================================================
2900  */
2901 static zio_pipe_stage_t *zio_pipeline[] = {
2902 	NULL,
2903 	zio_read_bp_init,
2904 	zio_free_bp_init,
2905 	zio_issue_async,
2906 	zio_write_bp_init,
2907 	zio_checksum_generate,
2908 	zio_ddt_read_start,
2909 	zio_ddt_read_done,
2910 	zio_ddt_write,
2911 	zio_ddt_free,
2912 	zio_gang_assemble,
2913 	zio_gang_issue,
2914 	zio_dva_allocate,
2915 	zio_dva_free,
2916 	zio_dva_claim,
2917 	zio_ready,
2918 	zio_vdev_io_start,
2919 	zio_vdev_io_done,
2920 	zio_vdev_io_assess,
2921 	zio_checksum_verify,
2922 	zio_done
2923 };
2924