xref: /titanic_52/usr/src/uts/common/fs/zfs/zio.c (revision efd4c9b63ad77503c101fc6c2ed8ba96c9d52964)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 #include <sys/zfs_context.h>
26 #include <sys/fm/fs/zfs.h>
27 #include <sys/spa.h>
28 #include <sys/txg.h>
29 #include <sys/spa_impl.h>
30 #include <sys/vdev_impl.h>
31 #include <sys/zio_impl.h>
32 #include <sys/zio_compress.h>
33 #include <sys/zio_checksum.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/arc.h>
36 #include <sys/ddt.h>
37 
38 /*
39  * ==========================================================================
40  * I/O priority table
41  * ==========================================================================
42  */
43 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
44 	0,	/* ZIO_PRIORITY_NOW		*/
45 	0,	/* ZIO_PRIORITY_SYNC_READ	*/
46 	0,	/* ZIO_PRIORITY_SYNC_WRITE	*/
47 	0,	/* ZIO_PRIORITY_LOG_WRITE	*/
48 	1,	/* ZIO_PRIORITY_CACHE_FILL	*/
49 	1,	/* ZIO_PRIORITY_AGG		*/
50 	4,	/* ZIO_PRIORITY_FREE		*/
51 	4,	/* ZIO_PRIORITY_ASYNC_WRITE	*/
52 	6,	/* ZIO_PRIORITY_ASYNC_READ	*/
53 	10,	/* ZIO_PRIORITY_RESILVER	*/
54 	20,	/* ZIO_PRIORITY_SCRUB		*/
55 	2,	/* ZIO_PRIORITY_DDT_PREFETCH	*/
56 };
57 
58 /*
59  * ==========================================================================
60  * I/O type descriptions
61  * ==========================================================================
62  */
63 char *zio_type_name[ZIO_TYPES] = {
64 	"zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
65 	"zio_ioctl"
66 };
67 
68 /*
69  * ==========================================================================
70  * I/O kmem caches
71  * ==========================================================================
72  */
73 kmem_cache_t *zio_cache;
74 kmem_cache_t *zio_link_cache;
75 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
76 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
77 
78 #ifdef _KERNEL
79 extern vmem_t *zio_alloc_arena;
80 #endif
81 
82 /*
83  * An allocating zio is one that either currently has the DVA allocate
84  * stage set or will have it later in its lifetime.
85  */
86 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
87 
88 boolean_t	zio_requeue_io_start_cut_in_line = B_TRUE;
89 
90 #ifdef ZFS_DEBUG
91 int zio_buf_debug_limit = 16384;
92 #else
93 int zio_buf_debug_limit = 0;
94 #endif
95 
96 void
97 zio_init(void)
98 {
99 	size_t c;
100 	vmem_t *data_alloc_arena = NULL;
101 
102 #ifdef _KERNEL
103 	data_alloc_arena = zio_alloc_arena;
104 #endif
105 	zio_cache = kmem_cache_create("zio_cache",
106 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
107 	zio_link_cache = kmem_cache_create("zio_link_cache",
108 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
109 
110 	/*
111 	 * For small buffers, we want a cache for each multiple of
112 	 * SPA_MINBLOCKSIZE.  For medium-size buffers, we want a cache
113 	 * for each quarter-power of 2.  For large buffers, we want
114 	 * a cache for each multiple of PAGESIZE.
115 	 */
116 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
117 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
118 		size_t p2 = size;
119 		size_t align = 0;
120 
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_append(&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 	dprintf_bp(bp, "freeing in txg %llu, pass %u",
665 	    (longlong_t)txg, spa->spa_sync_pass);
666 
667 	ASSERT(!BP_IS_HOLE(bp));
668 	ASSERT(spa_syncing_txg(spa) == txg);
669 	ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
670 
671 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
672 	    NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
673 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
674 
675 	return (zio);
676 }
677 
678 zio_t *
679 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
680     zio_done_func_t *done, void *private, enum zio_flag flags)
681 {
682 	zio_t *zio;
683 
684 	/*
685 	 * A claim is an allocation of a specific block.  Claims are needed
686 	 * to support immediate writes in the intent log.  The issue is that
687 	 * immediate writes contain committed data, but in a txg that was
688 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
689 	 * the intent log claims all blocks that contain immediate write data
690 	 * so that the SPA knows they're in use.
691 	 *
692 	 * All claims *must* be resolved in the first txg -- before the SPA
693 	 * starts allocating blocks -- so that nothing is allocated twice.
694 	 * If txg == 0 we just verify that the block is claimable.
695 	 */
696 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
697 	ASSERT(txg == spa_first_txg(spa) || txg == 0);
698 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
699 
700 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
701 	    done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
702 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
703 
704 	return (zio);
705 }
706 
707 zio_t *
708 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
709     zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
710 {
711 	zio_t *zio;
712 	int c;
713 
714 	if (vd->vdev_children == 0) {
715 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
716 		    ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
717 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
718 
719 		zio->io_cmd = cmd;
720 	} else {
721 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
722 
723 		for (c = 0; c < vd->vdev_children; c++)
724 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
725 			    done, private, priority, flags));
726 	}
727 
728 	return (zio);
729 }
730 
731 zio_t *
732 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
733     void *data, int checksum, zio_done_func_t *done, void *private,
734     int priority, enum zio_flag flags, boolean_t labels)
735 {
736 	zio_t *zio;
737 
738 	ASSERT(vd->vdev_children == 0);
739 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
740 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
741 	ASSERT3U(offset + size, <=, vd->vdev_psize);
742 
743 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
744 	    ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
745 	    ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
746 
747 	zio->io_prop.zp_checksum = checksum;
748 
749 	return (zio);
750 }
751 
752 zio_t *
753 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
754     void *data, int checksum, zio_done_func_t *done, void *private,
755     int priority, enum zio_flag flags, boolean_t labels)
756 {
757 	zio_t *zio;
758 
759 	ASSERT(vd->vdev_children == 0);
760 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
761 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
762 	ASSERT3U(offset + size, <=, vd->vdev_psize);
763 
764 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
765 	    ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
766 	    ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
767 
768 	zio->io_prop.zp_checksum = checksum;
769 
770 	if (zio_checksum_table[checksum].ci_eck) {
771 		/*
772 		 * zec checksums are necessarily destructive -- they modify
773 		 * the end of the write buffer to hold the verifier/checksum.
774 		 * Therefore, we must make a local copy in case the data is
775 		 * being written to multiple places in parallel.
776 		 */
777 		void *wbuf = zio_buf_alloc(size);
778 		bcopy(data, wbuf, size);
779 		zio_push_transform(zio, wbuf, size, size, NULL);
780 	}
781 
782 	return (zio);
783 }
784 
785 /*
786  * Create a child I/O to do some work for us.
787  */
788 zio_t *
789 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
790 	void *data, uint64_t size, int type, int priority, enum zio_flag flags,
791 	zio_done_func_t *done, void *private)
792 {
793 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
794 	zio_t *zio;
795 
796 	ASSERT(vd->vdev_parent ==
797 	    (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
798 
799 	if (type == ZIO_TYPE_READ && bp != NULL) {
800 		/*
801 		 * If we have the bp, then the child should perform the
802 		 * checksum and the parent need not.  This pushes error
803 		 * detection as close to the leaves as possible and
804 		 * eliminates redundant checksums in the interior nodes.
805 		 */
806 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
807 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
808 	}
809 
810 	if (vd->vdev_children == 0)
811 		offset += VDEV_LABEL_START_SIZE;
812 
813 	flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
814 
815 	/*
816 	 * If we've decided to do a repair, the write is not speculative --
817 	 * even if the original read was.
818 	 */
819 	if (flags & ZIO_FLAG_IO_REPAIR)
820 		flags &= ~ZIO_FLAG_SPECULATIVE;
821 
822 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
823 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
824 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
825 
826 	return (zio);
827 }
828 
829 zio_t *
830 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
831 	int type, int priority, enum zio_flag flags,
832 	zio_done_func_t *done, void *private)
833 {
834 	zio_t *zio;
835 
836 	ASSERT(vd->vdev_ops->vdev_op_leaf);
837 
838 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
839 	    data, size, done, private, type, priority,
840 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
841 	    vd, offset, NULL,
842 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
843 
844 	return (zio);
845 }
846 
847 void
848 zio_flush(zio_t *zio, vdev_t *vd)
849 {
850 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
851 	    NULL, NULL, ZIO_PRIORITY_NOW,
852 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
853 }
854 
855 void
856 zio_shrink(zio_t *zio, uint64_t size)
857 {
858 	ASSERT(zio->io_executor == NULL);
859 	ASSERT(zio->io_orig_size == zio->io_size);
860 	ASSERT(size <= zio->io_size);
861 
862 	/*
863 	 * We don't shrink for raidz because of problems with the
864 	 * reconstruction when reading back less than the block size.
865 	 * Note, BP_IS_RAIDZ() assumes no compression.
866 	 */
867 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
868 	if (!BP_IS_RAIDZ(zio->io_bp))
869 		zio->io_orig_size = zio->io_size = size;
870 }
871 
872 /*
873  * ==========================================================================
874  * Prepare to read and write logical blocks
875  * ==========================================================================
876  */
877 
878 static int
879 zio_read_bp_init(zio_t *zio)
880 {
881 	blkptr_t *bp = zio->io_bp;
882 
883 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
884 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
885 	    !(zio->io_flags & ZIO_FLAG_RAW)) {
886 		uint64_t psize = BP_GET_PSIZE(bp);
887 		void *cbuf = zio_buf_alloc(psize);
888 
889 		zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
890 	}
891 
892 	if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
893 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
894 
895 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
896 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
897 
898 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
899 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
900 
901 	return (ZIO_PIPELINE_CONTINUE);
902 }
903 
904 static int
905 zio_write_bp_init(zio_t *zio)
906 {
907 	spa_t *spa = zio->io_spa;
908 	zio_prop_t *zp = &zio->io_prop;
909 	enum zio_compress compress = zp->zp_compress;
910 	blkptr_t *bp = zio->io_bp;
911 	uint64_t lsize = zio->io_size;
912 	uint64_t psize = lsize;
913 	int pass = 1;
914 
915 	/*
916 	 * If our children haven't all reached the ready stage,
917 	 * wait for them and then repeat this pipeline stage.
918 	 */
919 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
920 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
921 		return (ZIO_PIPELINE_STOP);
922 
923 	if (!IO_IS_ALLOCATING(zio))
924 		return (ZIO_PIPELINE_CONTINUE);
925 
926 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
927 
928 	if (zio->io_bp_override) {
929 		ASSERT(bp->blk_birth != zio->io_txg);
930 		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
931 
932 		*bp = *zio->io_bp_override;
933 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
934 
935 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
936 			return (ZIO_PIPELINE_CONTINUE);
937 
938 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
939 		    zp->zp_dedup_verify);
940 
941 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
942 			BP_SET_DEDUP(bp, 1);
943 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
944 			return (ZIO_PIPELINE_CONTINUE);
945 		}
946 		zio->io_bp_override = NULL;
947 		BP_ZERO(bp);
948 	}
949 
950 	if (bp->blk_birth == zio->io_txg) {
951 		/*
952 		 * We're rewriting an existing block, which means we're
953 		 * working on behalf of spa_sync().  For spa_sync() to
954 		 * converge, it must eventually be the case that we don't
955 		 * have to allocate new blocks.  But compression changes
956 		 * the blocksize, which forces a reallocate, and makes
957 		 * convergence take longer.  Therefore, after the first
958 		 * few passes, stop compressing to ensure convergence.
959 		 */
960 		pass = spa_sync_pass(spa);
961 
962 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
963 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
964 		ASSERT(!BP_GET_DEDUP(bp));
965 
966 		if (pass > SYNC_PASS_DONT_COMPRESS)
967 			compress = ZIO_COMPRESS_OFF;
968 
969 		/* Make sure someone doesn't change their mind on overwrites */
970 		ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
971 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
972 	}
973 
974 	if (compress != ZIO_COMPRESS_OFF) {
975 		void *cbuf = zio_buf_alloc(lsize);
976 		psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
977 		if (psize == 0 || psize == lsize) {
978 			compress = ZIO_COMPRESS_OFF;
979 			zio_buf_free(cbuf, lsize);
980 		} else {
981 			ASSERT(psize < lsize);
982 			zio_push_transform(zio, cbuf, psize, lsize, NULL);
983 		}
984 	}
985 
986 	/*
987 	 * The final pass of spa_sync() must be all rewrites, but the first
988 	 * few passes offer a trade-off: allocating blocks defers convergence,
989 	 * but newly allocated blocks are sequential, so they can be written
990 	 * to disk faster.  Therefore, we allow the first few passes of
991 	 * spa_sync() to allocate new blocks, but force rewrites after that.
992 	 * There should only be a handful of blocks after pass 1 in any case.
993 	 */
994 	if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
995 	    pass > SYNC_PASS_REWRITE) {
996 		ASSERT(psize != 0);
997 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
998 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
999 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1000 	} else {
1001 		BP_ZERO(bp);
1002 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1003 	}
1004 
1005 	if (psize == 0) {
1006 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1007 	} else {
1008 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1009 		BP_SET_LSIZE(bp, lsize);
1010 		BP_SET_PSIZE(bp, psize);
1011 		BP_SET_COMPRESS(bp, compress);
1012 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1013 		BP_SET_TYPE(bp, zp->zp_type);
1014 		BP_SET_LEVEL(bp, zp->zp_level);
1015 		BP_SET_DEDUP(bp, zp->zp_dedup);
1016 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1017 		if (zp->zp_dedup) {
1018 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1019 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1020 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1021 		}
1022 	}
1023 
1024 	return (ZIO_PIPELINE_CONTINUE);
1025 }
1026 
1027 static int
1028 zio_free_bp_init(zio_t *zio)
1029 {
1030 	blkptr_t *bp = zio->io_bp;
1031 
1032 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1033 		if (BP_GET_DEDUP(bp))
1034 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1035 	}
1036 
1037 	return (ZIO_PIPELINE_CONTINUE);
1038 }
1039 
1040 /*
1041  * ==========================================================================
1042  * Execute the I/O pipeline
1043  * ==========================================================================
1044  */
1045 
1046 static void
1047 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1048 {
1049 	spa_t *spa = zio->io_spa;
1050 	zio_type_t t = zio->io_type;
1051 	int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0);
1052 
1053 	/*
1054 	 * If we're a config writer or a probe, the normal issue and
1055 	 * interrupt threads may all be blocked waiting for the config lock.
1056 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1057 	 */
1058 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1059 		t = ZIO_TYPE_NULL;
1060 
1061 	/*
1062 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1063 	 */
1064 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1065 		t = ZIO_TYPE_NULL;
1066 
1067 	/*
1068 	 * If this is a high priority I/O, then use the high priority taskq.
1069 	 */
1070 	if (zio->io_priority == ZIO_PRIORITY_NOW &&
1071 	    spa->spa_zio_taskq[t][q + 1] != NULL)
1072 		q++;
1073 
1074 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1075 	(void) taskq_dispatch(spa->spa_zio_taskq[t][q],
1076 	    (task_func_t *)zio_execute, zio, flags);
1077 }
1078 
1079 static boolean_t
1080 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1081 {
1082 	kthread_t *executor = zio->io_executor;
1083 	spa_t *spa = zio->io_spa;
1084 
1085 	for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1086 		if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1087 			return (B_TRUE);
1088 
1089 	return (B_FALSE);
1090 }
1091 
1092 static int
1093 zio_issue_async(zio_t *zio)
1094 {
1095 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1096 
1097 	return (ZIO_PIPELINE_STOP);
1098 }
1099 
1100 void
1101 zio_interrupt(zio_t *zio)
1102 {
1103 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1104 }
1105 
1106 /*
1107  * Execute the I/O pipeline until one of the following occurs:
1108  * (1) the I/O completes; (2) the pipeline stalls waiting for
1109  * dependent child I/Os; (3) the I/O issues, so we're waiting
1110  * for an I/O completion interrupt; (4) the I/O is delegated by
1111  * vdev-level caching or aggregation; (5) the I/O is deferred
1112  * due to vdev-level queueing; (6) the I/O is handed off to
1113  * another thread.  In all cases, the pipeline stops whenever
1114  * there's no CPU work; it never burns a thread in cv_wait().
1115  *
1116  * There's no locking on io_stage because there's no legitimate way
1117  * for multiple threads to be attempting to process the same I/O.
1118  */
1119 static zio_pipe_stage_t *zio_pipeline[];
1120 
1121 void
1122 zio_execute(zio_t *zio)
1123 {
1124 	zio->io_executor = curthread;
1125 
1126 	while (zio->io_stage < ZIO_STAGE_DONE) {
1127 		enum zio_stage pipeline = zio->io_pipeline;
1128 		enum zio_stage stage = zio->io_stage;
1129 		int rv;
1130 
1131 		ASSERT(!MUTEX_HELD(&zio->io_lock));
1132 		ASSERT(ISP2(stage));
1133 		ASSERT(zio->io_stall == NULL);
1134 
1135 		do {
1136 			stage <<= 1;
1137 		} while ((stage & pipeline) == 0);
1138 
1139 		ASSERT(stage <= ZIO_STAGE_DONE);
1140 
1141 		/*
1142 		 * If we are in interrupt context and this pipeline stage
1143 		 * will grab a config lock that is held across I/O,
1144 		 * or may wait for an I/O that needs an interrupt thread
1145 		 * to complete, issue async to avoid deadlock.
1146 		 *
1147 		 * For VDEV_IO_START, we cut in line so that the io will
1148 		 * be sent to disk promptly.
1149 		 */
1150 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1151 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1152 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1153 			    zio_requeue_io_start_cut_in_line : B_FALSE;
1154 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1155 			return;
1156 		}
1157 
1158 		zio->io_stage = stage;
1159 		rv = zio_pipeline[highbit(stage) - 1](zio);
1160 
1161 		if (rv == ZIO_PIPELINE_STOP)
1162 			return;
1163 
1164 		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1165 	}
1166 }
1167 
1168 /*
1169  * ==========================================================================
1170  * Initiate I/O, either sync or async
1171  * ==========================================================================
1172  */
1173 int
1174 zio_wait(zio_t *zio)
1175 {
1176 	int error;
1177 
1178 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1179 	ASSERT(zio->io_executor == NULL);
1180 
1181 	zio->io_waiter = curthread;
1182 
1183 	zio_execute(zio);
1184 
1185 	mutex_enter(&zio->io_lock);
1186 	while (zio->io_executor != NULL)
1187 		cv_wait(&zio->io_cv, &zio->io_lock);
1188 	mutex_exit(&zio->io_lock);
1189 
1190 	error = zio->io_error;
1191 	zio_destroy(zio);
1192 
1193 	return (error);
1194 }
1195 
1196 void
1197 zio_nowait(zio_t *zio)
1198 {
1199 	ASSERT(zio->io_executor == NULL);
1200 
1201 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1202 	    zio_unique_parent(zio) == NULL) {
1203 		/*
1204 		 * This is a logical async I/O with no parent to wait for it.
1205 		 * We add it to the spa_async_root_zio "Godfather" I/O which
1206 		 * will ensure they complete prior to unloading the pool.
1207 		 */
1208 		spa_t *spa = zio->io_spa;
1209 
1210 		zio_add_child(spa->spa_async_zio_root, zio);
1211 	}
1212 
1213 	zio_execute(zio);
1214 }
1215 
1216 /*
1217  * ==========================================================================
1218  * Reexecute or suspend/resume failed I/O
1219  * ==========================================================================
1220  */
1221 
1222 static void
1223 zio_reexecute(zio_t *pio)
1224 {
1225 	zio_t *cio, *cio_next;
1226 
1227 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1228 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1229 	ASSERT(pio->io_gang_leader == NULL);
1230 	ASSERT(pio->io_gang_tree == NULL);
1231 
1232 	pio->io_flags = pio->io_orig_flags;
1233 	pio->io_stage = pio->io_orig_stage;
1234 	pio->io_pipeline = pio->io_orig_pipeline;
1235 	pio->io_reexecute = 0;
1236 	pio->io_error = 0;
1237 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1238 		pio->io_state[w] = 0;
1239 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1240 		pio->io_child_error[c] = 0;
1241 
1242 	if (IO_IS_ALLOCATING(pio))
1243 		BP_ZERO(pio->io_bp);
1244 
1245 	/*
1246 	 * As we reexecute pio's children, new children could be created.
1247 	 * New children go to the head of pio's io_child_list, however,
1248 	 * so we will (correctly) not reexecute them.  The key is that
1249 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1250 	 * cannot be affected by any side effects of reexecuting 'cio'.
1251 	 */
1252 	for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1253 		cio_next = zio_walk_children(pio);
1254 		mutex_enter(&pio->io_lock);
1255 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1256 			pio->io_children[cio->io_child_type][w]++;
1257 		mutex_exit(&pio->io_lock);
1258 		zio_reexecute(cio);
1259 	}
1260 
1261 	/*
1262 	 * Now that all children have been reexecuted, execute the parent.
1263 	 * We don't reexecute "The Godfather" I/O here as it's the
1264 	 * responsibility of the caller to wait on him.
1265 	 */
1266 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1267 		zio_execute(pio);
1268 }
1269 
1270 void
1271 zio_suspend(spa_t *spa, zio_t *zio)
1272 {
1273 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1274 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1275 		    "failure and the failure mode property for this pool "
1276 		    "is set to panic.", spa_name(spa));
1277 
1278 	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1279 
1280 	mutex_enter(&spa->spa_suspend_lock);
1281 
1282 	if (spa->spa_suspend_zio_root == NULL)
1283 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1284 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1285 		    ZIO_FLAG_GODFATHER);
1286 
1287 	spa->spa_suspended = B_TRUE;
1288 
1289 	if (zio != NULL) {
1290 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1291 		ASSERT(zio != spa->spa_suspend_zio_root);
1292 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1293 		ASSERT(zio_unique_parent(zio) == NULL);
1294 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1295 		zio_add_child(spa->spa_suspend_zio_root, zio);
1296 	}
1297 
1298 	mutex_exit(&spa->spa_suspend_lock);
1299 }
1300 
1301 int
1302 zio_resume(spa_t *spa)
1303 {
1304 	zio_t *pio;
1305 
1306 	/*
1307 	 * Reexecute all previously suspended i/o.
1308 	 */
1309 	mutex_enter(&spa->spa_suspend_lock);
1310 	spa->spa_suspended = B_FALSE;
1311 	cv_broadcast(&spa->spa_suspend_cv);
1312 	pio = spa->spa_suspend_zio_root;
1313 	spa->spa_suspend_zio_root = NULL;
1314 	mutex_exit(&spa->spa_suspend_lock);
1315 
1316 	if (pio == NULL)
1317 		return (0);
1318 
1319 	zio_reexecute(pio);
1320 	return (zio_wait(pio));
1321 }
1322 
1323 void
1324 zio_resume_wait(spa_t *spa)
1325 {
1326 	mutex_enter(&spa->spa_suspend_lock);
1327 	while (spa_suspended(spa))
1328 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1329 	mutex_exit(&spa->spa_suspend_lock);
1330 }
1331 
1332 /*
1333  * ==========================================================================
1334  * Gang blocks.
1335  *
1336  * A gang block is a collection of small blocks that looks to the DMU
1337  * like one large block.  When zio_dva_allocate() cannot find a block
1338  * of the requested size, due to either severe fragmentation or the pool
1339  * being nearly full, it calls zio_write_gang_block() to construct the
1340  * block from smaller fragments.
1341  *
1342  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1343  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
1344  * an indirect block: it's an array of block pointers.  It consumes
1345  * only one sector and hence is allocatable regardless of fragmentation.
1346  * The gang header's bps point to its gang members, which hold the data.
1347  *
1348  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1349  * as the verifier to ensure uniqueness of the SHA256 checksum.
1350  * Critically, the gang block bp's blk_cksum is the checksum of the data,
1351  * not the gang header.  This ensures that data block signatures (needed for
1352  * deduplication) are independent of how the block is physically stored.
1353  *
1354  * Gang blocks can be nested: a gang member may itself be a gang block.
1355  * Thus every gang block is a tree in which root and all interior nodes are
1356  * gang headers, and the leaves are normal blocks that contain user data.
1357  * The root of the gang tree is called the gang leader.
1358  *
1359  * To perform any operation (read, rewrite, free, claim) on a gang block,
1360  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1361  * in the io_gang_tree field of the original logical i/o by recursively
1362  * reading the gang leader and all gang headers below it.  This yields
1363  * an in-core tree containing the contents of every gang header and the
1364  * bps for every constituent of the gang block.
1365  *
1366  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1367  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
1368  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1369  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1370  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1371  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
1372  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1373  * of the gang header plus zio_checksum_compute() of the data to update the
1374  * gang header's blk_cksum as described above.
1375  *
1376  * The two-phase assemble/issue model solves the problem of partial failure --
1377  * what if you'd freed part of a gang block but then couldn't read the
1378  * gang header for another part?  Assembling the entire gang tree first
1379  * ensures that all the necessary gang header I/O has succeeded before
1380  * starting the actual work of free, claim, or write.  Once the gang tree
1381  * is assembled, free and claim are in-memory operations that cannot fail.
1382  *
1383  * In the event that a gang write fails, zio_dva_unallocate() walks the
1384  * gang tree to immediately free (i.e. insert back into the space map)
1385  * everything we've allocated.  This ensures that we don't get ENOSPC
1386  * errors during repeated suspend/resume cycles due to a flaky device.
1387  *
1388  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
1389  * the gang tree, we won't modify the block, so we can safely defer the free
1390  * (knowing that the block is still intact).  If we *can* assemble the gang
1391  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1392  * each constituent bp and we can allocate a new block on the next sync pass.
1393  *
1394  * In all cases, the gang tree allows complete recovery from partial failure.
1395  * ==========================================================================
1396  */
1397 
1398 static zio_t *
1399 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1400 {
1401 	if (gn != NULL)
1402 		return (pio);
1403 
1404 	return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1405 	    NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1406 	    &pio->io_bookmark));
1407 }
1408 
1409 zio_t *
1410 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1411 {
1412 	zio_t *zio;
1413 
1414 	if (gn != NULL) {
1415 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1416 		    gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1417 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1418 		/*
1419 		 * As we rewrite each gang header, the pipeline will compute
1420 		 * a new gang block header checksum for it; but no one will
1421 		 * compute a new data checksum, so we do that here.  The one
1422 		 * exception is the gang leader: the pipeline already computed
1423 		 * its data checksum because that stage precedes gang assembly.
1424 		 * (Presently, nothing actually uses interior data checksums;
1425 		 * this is just good hygiene.)
1426 		 */
1427 		if (gn != pio->io_gang_leader->io_gang_tree) {
1428 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1429 			    data, BP_GET_PSIZE(bp));
1430 		}
1431 		/*
1432 		 * If we are here to damage data for testing purposes,
1433 		 * leave the GBH alone so that we can detect the damage.
1434 		 */
1435 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1436 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1437 	} else {
1438 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1439 		    data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1440 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1441 	}
1442 
1443 	return (zio);
1444 }
1445 
1446 /* ARGSUSED */
1447 zio_t *
1448 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1449 {
1450 	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1451 	    ZIO_GANG_CHILD_FLAGS(pio)));
1452 }
1453 
1454 /* ARGSUSED */
1455 zio_t *
1456 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1457 {
1458 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1459 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1460 }
1461 
1462 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1463 	NULL,
1464 	zio_read_gang,
1465 	zio_rewrite_gang,
1466 	zio_free_gang,
1467 	zio_claim_gang,
1468 	NULL
1469 };
1470 
1471 static void zio_gang_tree_assemble_done(zio_t *zio);
1472 
1473 static zio_gang_node_t *
1474 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1475 {
1476 	zio_gang_node_t *gn;
1477 
1478 	ASSERT(*gnpp == NULL);
1479 
1480 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1481 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1482 	*gnpp = gn;
1483 
1484 	return (gn);
1485 }
1486 
1487 static void
1488 zio_gang_node_free(zio_gang_node_t **gnpp)
1489 {
1490 	zio_gang_node_t *gn = *gnpp;
1491 
1492 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1493 		ASSERT(gn->gn_child[g] == NULL);
1494 
1495 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1496 	kmem_free(gn, sizeof (*gn));
1497 	*gnpp = NULL;
1498 }
1499 
1500 static void
1501 zio_gang_tree_free(zio_gang_node_t **gnpp)
1502 {
1503 	zio_gang_node_t *gn = *gnpp;
1504 
1505 	if (gn == NULL)
1506 		return;
1507 
1508 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1509 		zio_gang_tree_free(&gn->gn_child[g]);
1510 
1511 	zio_gang_node_free(gnpp);
1512 }
1513 
1514 static void
1515 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1516 {
1517 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1518 
1519 	ASSERT(gio->io_gang_leader == gio);
1520 	ASSERT(BP_IS_GANG(bp));
1521 
1522 	zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1523 	    SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1524 	    gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1525 }
1526 
1527 static void
1528 zio_gang_tree_assemble_done(zio_t *zio)
1529 {
1530 	zio_t *gio = zio->io_gang_leader;
1531 	zio_gang_node_t *gn = zio->io_private;
1532 	blkptr_t *bp = zio->io_bp;
1533 
1534 	ASSERT(gio == zio_unique_parent(zio));
1535 	ASSERT(zio->io_child_count == 0);
1536 
1537 	if (zio->io_error)
1538 		return;
1539 
1540 	if (BP_SHOULD_BYTESWAP(bp))
1541 		byteswap_uint64_array(zio->io_data, zio->io_size);
1542 
1543 	ASSERT(zio->io_data == gn->gn_gbh);
1544 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1545 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1546 
1547 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1548 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1549 		if (!BP_IS_GANG(gbp))
1550 			continue;
1551 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1552 	}
1553 }
1554 
1555 static void
1556 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1557 {
1558 	zio_t *gio = pio->io_gang_leader;
1559 	zio_t *zio;
1560 
1561 	ASSERT(BP_IS_GANG(bp) == !!gn);
1562 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1563 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1564 
1565 	/*
1566 	 * If you're a gang header, your data is in gn->gn_gbh.
1567 	 * If you're a gang member, your data is in 'data' and gn == NULL.
1568 	 */
1569 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1570 
1571 	if (gn != NULL) {
1572 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1573 
1574 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1575 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1576 			if (BP_IS_HOLE(gbp))
1577 				continue;
1578 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1579 			data = (char *)data + BP_GET_PSIZE(gbp);
1580 		}
1581 	}
1582 
1583 	if (gn == gio->io_gang_tree)
1584 		ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1585 
1586 	if (zio != pio)
1587 		zio_nowait(zio);
1588 }
1589 
1590 static int
1591 zio_gang_assemble(zio_t *zio)
1592 {
1593 	blkptr_t *bp = zio->io_bp;
1594 
1595 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1596 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1597 
1598 	zio->io_gang_leader = zio;
1599 
1600 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1601 
1602 	return (ZIO_PIPELINE_CONTINUE);
1603 }
1604 
1605 static int
1606 zio_gang_issue(zio_t *zio)
1607 {
1608 	blkptr_t *bp = zio->io_bp;
1609 
1610 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1611 		return (ZIO_PIPELINE_STOP);
1612 
1613 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1614 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1615 
1616 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1617 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1618 	else
1619 		zio_gang_tree_free(&zio->io_gang_tree);
1620 
1621 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1622 
1623 	return (ZIO_PIPELINE_CONTINUE);
1624 }
1625 
1626 static void
1627 zio_write_gang_member_ready(zio_t *zio)
1628 {
1629 	zio_t *pio = zio_unique_parent(zio);
1630 	zio_t *gio = zio->io_gang_leader;
1631 	dva_t *cdva = zio->io_bp->blk_dva;
1632 	dva_t *pdva = pio->io_bp->blk_dva;
1633 	uint64_t asize;
1634 
1635 	if (BP_IS_HOLE(zio->io_bp))
1636 		return;
1637 
1638 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1639 
1640 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1641 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1642 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1643 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1644 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1645 
1646 	mutex_enter(&pio->io_lock);
1647 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1648 		ASSERT(DVA_GET_GANG(&pdva[d]));
1649 		asize = DVA_GET_ASIZE(&pdva[d]);
1650 		asize += DVA_GET_ASIZE(&cdva[d]);
1651 		DVA_SET_ASIZE(&pdva[d], asize);
1652 	}
1653 	mutex_exit(&pio->io_lock);
1654 }
1655 
1656 static int
1657 zio_write_gang_block(zio_t *pio)
1658 {
1659 	spa_t *spa = pio->io_spa;
1660 	blkptr_t *bp = pio->io_bp;
1661 	zio_t *gio = pio->io_gang_leader;
1662 	zio_t *zio;
1663 	zio_gang_node_t *gn, **gnpp;
1664 	zio_gbh_phys_t *gbh;
1665 	uint64_t txg = pio->io_txg;
1666 	uint64_t resid = pio->io_size;
1667 	uint64_t lsize;
1668 	int copies = gio->io_prop.zp_copies;
1669 	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1670 	zio_prop_t zp;
1671 	int error;
1672 
1673 	error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1674 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1675 	    METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1676 	if (error) {
1677 		pio->io_error = error;
1678 		return (ZIO_PIPELINE_CONTINUE);
1679 	}
1680 
1681 	if (pio == gio) {
1682 		gnpp = &gio->io_gang_tree;
1683 	} else {
1684 		gnpp = pio->io_private;
1685 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
1686 	}
1687 
1688 	gn = zio_gang_node_alloc(gnpp);
1689 	gbh = gn->gn_gbh;
1690 	bzero(gbh, SPA_GANGBLOCKSIZE);
1691 
1692 	/*
1693 	 * Create the gang header.
1694 	 */
1695 	zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1696 	    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1697 
1698 	/*
1699 	 * Create and nowait the gang children.
1700 	 */
1701 	for (int g = 0; resid != 0; resid -= lsize, g++) {
1702 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1703 		    SPA_MINBLOCKSIZE);
1704 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1705 
1706 		zp.zp_checksum = gio->io_prop.zp_checksum;
1707 		zp.zp_compress = ZIO_COMPRESS_OFF;
1708 		zp.zp_type = DMU_OT_NONE;
1709 		zp.zp_level = 0;
1710 		zp.zp_copies = gio->io_prop.zp_copies;
1711 		zp.zp_dedup = 0;
1712 		zp.zp_dedup_verify = 0;
1713 
1714 		zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1715 		    (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1716 		    zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1717 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1718 		    &pio->io_bookmark));
1719 	}
1720 
1721 	/*
1722 	 * Set pio's pipeline to just wait for zio to finish.
1723 	 */
1724 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1725 
1726 	zio_nowait(zio);
1727 
1728 	return (ZIO_PIPELINE_CONTINUE);
1729 }
1730 
1731 /*
1732  * ==========================================================================
1733  * Dedup
1734  * ==========================================================================
1735  */
1736 static void
1737 zio_ddt_child_read_done(zio_t *zio)
1738 {
1739 	blkptr_t *bp = zio->io_bp;
1740 	ddt_entry_t *dde = zio->io_private;
1741 	ddt_phys_t *ddp;
1742 	zio_t *pio = zio_unique_parent(zio);
1743 
1744 	mutex_enter(&pio->io_lock);
1745 	ddp = ddt_phys_select(dde, bp);
1746 	if (zio->io_error == 0)
1747 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
1748 	if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1749 		dde->dde_repair_data = zio->io_data;
1750 	else
1751 		zio_buf_free(zio->io_data, zio->io_size);
1752 	mutex_exit(&pio->io_lock);
1753 }
1754 
1755 static int
1756 zio_ddt_read_start(zio_t *zio)
1757 {
1758 	blkptr_t *bp = zio->io_bp;
1759 
1760 	ASSERT(BP_GET_DEDUP(bp));
1761 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1762 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1763 
1764 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1765 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1766 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1767 		ddt_phys_t *ddp = dde->dde_phys;
1768 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1769 		blkptr_t blk;
1770 
1771 		ASSERT(zio->io_vsd == NULL);
1772 		zio->io_vsd = dde;
1773 
1774 		if (ddp_self == NULL)
1775 			return (ZIO_PIPELINE_CONTINUE);
1776 
1777 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1778 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1779 				continue;
1780 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1781 			    &blk);
1782 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
1783 			    zio_buf_alloc(zio->io_size), zio->io_size,
1784 			    zio_ddt_child_read_done, dde, zio->io_priority,
1785 			    ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1786 			    &zio->io_bookmark));
1787 		}
1788 		return (ZIO_PIPELINE_CONTINUE);
1789 	}
1790 
1791 	zio_nowait(zio_read(zio, zio->io_spa, bp,
1792 	    zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1793 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1794 
1795 	return (ZIO_PIPELINE_CONTINUE);
1796 }
1797 
1798 static int
1799 zio_ddt_read_done(zio_t *zio)
1800 {
1801 	blkptr_t *bp = zio->io_bp;
1802 
1803 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1804 		return (ZIO_PIPELINE_STOP);
1805 
1806 	ASSERT(BP_GET_DEDUP(bp));
1807 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1808 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1809 
1810 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1811 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1812 		ddt_entry_t *dde = zio->io_vsd;
1813 		if (ddt == NULL) {
1814 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1815 			return (ZIO_PIPELINE_CONTINUE);
1816 		}
1817 		if (dde == NULL) {
1818 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1819 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1820 			return (ZIO_PIPELINE_STOP);
1821 		}
1822 		if (dde->dde_repair_data != NULL) {
1823 			bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1824 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
1825 		}
1826 		ddt_repair_done(ddt, dde);
1827 		zio->io_vsd = NULL;
1828 	}
1829 
1830 	ASSERT(zio->io_vsd == NULL);
1831 
1832 	return (ZIO_PIPELINE_CONTINUE);
1833 }
1834 
1835 static boolean_t
1836 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1837 {
1838 	spa_t *spa = zio->io_spa;
1839 
1840 	/*
1841 	 * Note: we compare the original data, not the transformed data,
1842 	 * because when zio->io_bp is an override bp, we will not have
1843 	 * pushed the I/O transforms.  That's an important optimization
1844 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1845 	 */
1846 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1847 		zio_t *lio = dde->dde_lead_zio[p];
1848 
1849 		if (lio != NULL) {
1850 			return (lio->io_orig_size != zio->io_orig_size ||
1851 			    bcmp(zio->io_orig_data, lio->io_orig_data,
1852 			    zio->io_orig_size) != 0);
1853 		}
1854 	}
1855 
1856 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1857 		ddt_phys_t *ddp = &dde->dde_phys[p];
1858 
1859 		if (ddp->ddp_phys_birth != 0) {
1860 			arc_buf_t *abuf = NULL;
1861 			uint32_t aflags = ARC_WAIT;
1862 			blkptr_t blk = *zio->io_bp;
1863 			int error;
1864 
1865 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1866 
1867 			ddt_exit(ddt);
1868 
1869 			error = arc_read_nolock(NULL, spa, &blk,
1870 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1871 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1872 			    &aflags, &zio->io_bookmark);
1873 
1874 			if (error == 0) {
1875 				if (arc_buf_size(abuf) != zio->io_orig_size ||
1876 				    bcmp(abuf->b_data, zio->io_orig_data,
1877 				    zio->io_orig_size) != 0)
1878 					error = EEXIST;
1879 				VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1880 			}
1881 
1882 			ddt_enter(ddt);
1883 			return (error != 0);
1884 		}
1885 	}
1886 
1887 	return (B_FALSE);
1888 }
1889 
1890 static void
1891 zio_ddt_child_write_ready(zio_t *zio)
1892 {
1893 	int p = zio->io_prop.zp_copies;
1894 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1895 	ddt_entry_t *dde = zio->io_private;
1896 	ddt_phys_t *ddp = &dde->dde_phys[p];
1897 	zio_t *pio;
1898 
1899 	if (zio->io_error)
1900 		return;
1901 
1902 	ddt_enter(ddt);
1903 
1904 	ASSERT(dde->dde_lead_zio[p] == zio);
1905 
1906 	ddt_phys_fill(ddp, zio->io_bp);
1907 
1908 	while ((pio = zio_walk_parents(zio)) != NULL)
1909 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1910 
1911 	ddt_exit(ddt);
1912 }
1913 
1914 static void
1915 zio_ddt_child_write_done(zio_t *zio)
1916 {
1917 	int p = zio->io_prop.zp_copies;
1918 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1919 	ddt_entry_t *dde = zio->io_private;
1920 	ddt_phys_t *ddp = &dde->dde_phys[p];
1921 
1922 	ddt_enter(ddt);
1923 
1924 	ASSERT(ddp->ddp_refcnt == 0);
1925 	ASSERT(dde->dde_lead_zio[p] == zio);
1926 	dde->dde_lead_zio[p] = NULL;
1927 
1928 	if (zio->io_error == 0) {
1929 		while (zio_walk_parents(zio) != NULL)
1930 			ddt_phys_addref(ddp);
1931 	} else {
1932 		ddt_phys_clear(ddp);
1933 	}
1934 
1935 	ddt_exit(ddt);
1936 }
1937 
1938 static void
1939 zio_ddt_ditto_write_done(zio_t *zio)
1940 {
1941 	int p = DDT_PHYS_DITTO;
1942 	zio_prop_t *zp = &zio->io_prop;
1943 	blkptr_t *bp = zio->io_bp;
1944 	ddt_t *ddt = ddt_select(zio->io_spa, bp);
1945 	ddt_entry_t *dde = zio->io_private;
1946 	ddt_phys_t *ddp = &dde->dde_phys[p];
1947 	ddt_key_t *ddk = &dde->dde_key;
1948 
1949 	ddt_enter(ddt);
1950 
1951 	ASSERT(ddp->ddp_refcnt == 0);
1952 	ASSERT(dde->dde_lead_zio[p] == zio);
1953 	dde->dde_lead_zio[p] = NULL;
1954 
1955 	if (zio->io_error == 0) {
1956 		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1957 		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1958 		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1959 		if (ddp->ddp_phys_birth != 0)
1960 			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1961 		ddt_phys_fill(ddp, bp);
1962 	}
1963 
1964 	ddt_exit(ddt);
1965 }
1966 
1967 static int
1968 zio_ddt_write(zio_t *zio)
1969 {
1970 	spa_t *spa = zio->io_spa;
1971 	blkptr_t *bp = zio->io_bp;
1972 	uint64_t txg = zio->io_txg;
1973 	zio_prop_t *zp = &zio->io_prop;
1974 	int p = zp->zp_copies;
1975 	int ditto_copies;
1976 	zio_t *cio = NULL;
1977 	zio_t *dio = NULL;
1978 	ddt_t *ddt = ddt_select(spa, bp);
1979 	ddt_entry_t *dde;
1980 	ddt_phys_t *ddp;
1981 
1982 	ASSERT(BP_GET_DEDUP(bp));
1983 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
1984 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
1985 
1986 	ddt_enter(ddt);
1987 	dde = ddt_lookup(ddt, bp, B_TRUE);
1988 	ddp = &dde->dde_phys[p];
1989 
1990 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
1991 		/*
1992 		 * If we're using a weak checksum, upgrade to a strong checksum
1993 		 * and try again.  If we're already using a strong checksum,
1994 		 * we can't resolve it, so just convert to an ordinary write.
1995 		 * (And automatically e-mail a paper to Nature?)
1996 		 */
1997 		if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
1998 			zp->zp_checksum = spa_dedup_checksum(spa);
1999 			zio_pop_transforms(zio);
2000 			zio->io_stage = ZIO_STAGE_OPEN;
2001 			BP_ZERO(bp);
2002 		} else {
2003 			zp->zp_dedup = 0;
2004 		}
2005 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2006 		ddt_exit(ddt);
2007 		return (ZIO_PIPELINE_CONTINUE);
2008 	}
2009 
2010 	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2011 	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2012 
2013 	if (ditto_copies > ddt_ditto_copies_present(dde) &&
2014 	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2015 		zio_prop_t czp = *zp;
2016 
2017 		czp.zp_copies = ditto_copies;
2018 
2019 		/*
2020 		 * If we arrived here with an override bp, we won't have run
2021 		 * the transform stack, so we won't have the data we need to
2022 		 * generate a child i/o.  So, toss the override bp and restart.
2023 		 * This is safe, because using the override bp is just an
2024 		 * optimization; and it's rare, so the cost doesn't matter.
2025 		 */
2026 		if (zio->io_bp_override) {
2027 			zio_pop_transforms(zio);
2028 			zio->io_stage = ZIO_STAGE_OPEN;
2029 			zio->io_pipeline = ZIO_WRITE_PIPELINE;
2030 			zio->io_bp_override = NULL;
2031 			BP_ZERO(bp);
2032 			ddt_exit(ddt);
2033 			return (ZIO_PIPELINE_CONTINUE);
2034 		}
2035 
2036 		dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2037 		    zio->io_orig_size, &czp, NULL,
2038 		    zio_ddt_ditto_write_done, dde, zio->io_priority,
2039 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2040 
2041 		zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2042 		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2043 	}
2044 
2045 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2046 		if (ddp->ddp_phys_birth != 0)
2047 			ddt_bp_fill(ddp, bp, txg);
2048 		if (dde->dde_lead_zio[p] != NULL)
2049 			zio_add_child(zio, dde->dde_lead_zio[p]);
2050 		else
2051 			ddt_phys_addref(ddp);
2052 	} else if (zio->io_bp_override) {
2053 		ASSERT(bp->blk_birth == txg);
2054 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2055 		ddt_phys_fill(ddp, bp);
2056 		ddt_phys_addref(ddp);
2057 	} else {
2058 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2059 		    zio->io_orig_size, zp, zio_ddt_child_write_ready,
2060 		    zio_ddt_child_write_done, dde, zio->io_priority,
2061 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2062 
2063 		zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2064 		dde->dde_lead_zio[p] = cio;
2065 	}
2066 
2067 	ddt_exit(ddt);
2068 
2069 	if (cio)
2070 		zio_nowait(cio);
2071 	if (dio)
2072 		zio_nowait(dio);
2073 
2074 	return (ZIO_PIPELINE_CONTINUE);
2075 }
2076 
2077 ddt_entry_t *freedde; /* for debugging */
2078 
2079 static int
2080 zio_ddt_free(zio_t *zio)
2081 {
2082 	spa_t *spa = zio->io_spa;
2083 	blkptr_t *bp = zio->io_bp;
2084 	ddt_t *ddt = ddt_select(spa, bp);
2085 	ddt_entry_t *dde;
2086 	ddt_phys_t *ddp;
2087 
2088 	ASSERT(BP_GET_DEDUP(bp));
2089 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2090 
2091 	ddt_enter(ddt);
2092 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2093 	ddp = ddt_phys_select(dde, bp);
2094 	ddt_phys_decref(ddp);
2095 	ddt_exit(ddt);
2096 
2097 	return (ZIO_PIPELINE_CONTINUE);
2098 }
2099 
2100 /*
2101  * ==========================================================================
2102  * Allocate and free blocks
2103  * ==========================================================================
2104  */
2105 static int
2106 zio_dva_allocate(zio_t *zio)
2107 {
2108 	spa_t *spa = zio->io_spa;
2109 	metaslab_class_t *mc = spa_normal_class(spa);
2110 	blkptr_t *bp = zio->io_bp;
2111 	int error;
2112 
2113 	if (zio->io_gang_leader == NULL) {
2114 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2115 		zio->io_gang_leader = zio;
2116 	}
2117 
2118 	ASSERT(BP_IS_HOLE(bp));
2119 	ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2120 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
2121 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2122 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2123 
2124 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
2125 	    zio->io_prop.zp_copies, zio->io_txg, NULL, 0);
2126 
2127 	if (error) {
2128 		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2129 			return (zio_write_gang_block(zio));
2130 		zio->io_error = error;
2131 	}
2132 
2133 	return (ZIO_PIPELINE_CONTINUE);
2134 }
2135 
2136 static int
2137 zio_dva_free(zio_t *zio)
2138 {
2139 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2140 
2141 	return (ZIO_PIPELINE_CONTINUE);
2142 }
2143 
2144 static int
2145 zio_dva_claim(zio_t *zio)
2146 {
2147 	int error;
2148 
2149 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2150 	if (error)
2151 		zio->io_error = error;
2152 
2153 	return (ZIO_PIPELINE_CONTINUE);
2154 }
2155 
2156 /*
2157  * Undo an allocation.  This is used by zio_done() when an I/O fails
2158  * and we want to give back the block we just allocated.
2159  * This handles both normal blocks and gang blocks.
2160  */
2161 static void
2162 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2163 {
2164 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2165 	ASSERT(zio->io_bp_override == NULL);
2166 
2167 	if (!BP_IS_HOLE(bp))
2168 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2169 
2170 	if (gn != NULL) {
2171 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2172 			zio_dva_unallocate(zio, gn->gn_child[g],
2173 			    &gn->gn_gbh->zg_blkptr[g]);
2174 		}
2175 	}
2176 }
2177 
2178 /*
2179  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
2180  */
2181 int
2182 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2183     uint64_t size, boolean_t use_slog)
2184 {
2185 	int error = 1;
2186 
2187 	ASSERT(txg > spa_syncing_txg(spa));
2188 
2189 	if (use_slog)
2190 		error = metaslab_alloc(spa, spa_log_class(spa), size,
2191 		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2192 
2193 	if (error)
2194 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
2195 		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2196 
2197 	if (error == 0) {
2198 		BP_SET_LSIZE(new_bp, size);
2199 		BP_SET_PSIZE(new_bp, size);
2200 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2201 		BP_SET_CHECKSUM(new_bp,
2202 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2203 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2204 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2205 		BP_SET_LEVEL(new_bp, 0);
2206 		BP_SET_DEDUP(new_bp, 0);
2207 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2208 	}
2209 
2210 	return (error);
2211 }
2212 
2213 /*
2214  * Free an intent log block.
2215  */
2216 void
2217 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2218 {
2219 	ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2220 	ASSERT(!BP_IS_GANG(bp));
2221 
2222 	zio_free(spa, txg, bp);
2223 }
2224 
2225 /*
2226  * ==========================================================================
2227  * Read and write to physical devices
2228  * ==========================================================================
2229  */
2230 static int
2231 zio_vdev_io_start(zio_t *zio)
2232 {
2233 	vdev_t *vd = zio->io_vd;
2234 	uint64_t align;
2235 	spa_t *spa = zio->io_spa;
2236 
2237 	ASSERT(zio->io_error == 0);
2238 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2239 
2240 	if (vd == NULL) {
2241 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2242 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2243 
2244 		/*
2245 		 * The mirror_ops handle multiple DVAs in a single BP.
2246 		 */
2247 		return (vdev_mirror_ops.vdev_op_io_start(zio));
2248 	}
2249 
2250 	/*
2251 	 * We keep track of time-sensitive I/Os so that the scan thread
2252 	 * can quickly react to certain workloads.  In particular, we care
2253 	 * about non-scrubbing, top-level reads and writes with the following
2254 	 * characteristics:
2255 	 * 	- synchronous writes of user data to non-slog devices
2256 	 *	- any reads of user data
2257 	 * When these conditions are met, adjust the timestamp of spa_last_io
2258 	 * which allows the scan thread to adjust its workload accordingly.
2259 	 */
2260 	if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2261 	    vd == vd->vdev_top && !vd->vdev_islog &&
2262 	    zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2263 	    zio->io_txg != spa_syncing_txg(spa)) {
2264 		uint64_t old = spa->spa_last_io;
2265 		uint64_t new = ddi_get_lbolt64();
2266 		if (old != new)
2267 			(void) atomic_cas_64(&spa->spa_last_io, old, new);
2268 	}
2269 
2270 	align = 1ULL << vd->vdev_top->vdev_ashift;
2271 
2272 	if (P2PHASE(zio->io_size, align) != 0) {
2273 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
2274 		char *abuf = zio_buf_alloc(asize);
2275 		ASSERT(vd == vd->vdev_top);
2276 		if (zio->io_type == ZIO_TYPE_WRITE) {
2277 			bcopy(zio->io_data, abuf, zio->io_size);
2278 			bzero(abuf + zio->io_size, asize - zio->io_size);
2279 		}
2280 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2281 	}
2282 
2283 	ASSERT(P2PHASE(zio->io_offset, align) == 0);
2284 	ASSERT(P2PHASE(zio->io_size, align) == 0);
2285 	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2286 
2287 	/*
2288 	 * If this is a repair I/O, and there's no self-healing involved --
2289 	 * that is, we're just resilvering what we expect to resilver --
2290 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2291 	 * This prevents spurious resilvering with nested replication.
2292 	 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2293 	 * A is out of date, we'll read from C+D, then use the data to
2294 	 * resilver A+B -- but we don't actually want to resilver B, just A.
2295 	 * The top-level mirror has no way to know this, so instead we just
2296 	 * discard unnecessary repairs as we work our way down the vdev tree.
2297 	 * The same logic applies to any form of nested replication:
2298 	 * ditto + mirror, RAID-Z + replacing, etc.  This covers them all.
2299 	 */
2300 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2301 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2302 	    zio->io_txg != 0 &&	/* not a delegated i/o */
2303 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2304 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2305 		zio_vdev_io_bypass(zio);
2306 		return (ZIO_PIPELINE_CONTINUE);
2307 	}
2308 
2309 	if (vd->vdev_ops->vdev_op_leaf &&
2310 	    (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2311 
2312 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2313 			return (ZIO_PIPELINE_CONTINUE);
2314 
2315 		if ((zio = vdev_queue_io(zio)) == NULL)
2316 			return (ZIO_PIPELINE_STOP);
2317 
2318 		if (!vdev_accessible(vd, zio)) {
2319 			zio->io_error = ENXIO;
2320 			zio_interrupt(zio);
2321 			return (ZIO_PIPELINE_STOP);
2322 		}
2323 	}
2324 
2325 	return (vd->vdev_ops->vdev_op_io_start(zio));
2326 }
2327 
2328 static int
2329 zio_vdev_io_done(zio_t *zio)
2330 {
2331 	vdev_t *vd = zio->io_vd;
2332 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2333 	boolean_t unexpected_error = B_FALSE;
2334 
2335 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2336 		return (ZIO_PIPELINE_STOP);
2337 
2338 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2339 
2340 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2341 
2342 		vdev_queue_io_done(zio);
2343 
2344 		if (zio->io_type == ZIO_TYPE_WRITE)
2345 			vdev_cache_write(zio);
2346 
2347 		if (zio_injection_enabled && zio->io_error == 0)
2348 			zio->io_error = zio_handle_device_injection(vd,
2349 			    zio, EIO);
2350 
2351 		if (zio_injection_enabled && zio->io_error == 0)
2352 			zio->io_error = zio_handle_label_injection(zio, EIO);
2353 
2354 		if (zio->io_error) {
2355 			if (!vdev_accessible(vd, zio)) {
2356 				zio->io_error = ENXIO;
2357 			} else {
2358 				unexpected_error = B_TRUE;
2359 			}
2360 		}
2361 	}
2362 
2363 	ops->vdev_op_io_done(zio);
2364 
2365 	if (unexpected_error)
2366 		VERIFY(vdev_probe(vd, zio) == NULL);
2367 
2368 	return (ZIO_PIPELINE_CONTINUE);
2369 }
2370 
2371 /*
2372  * For non-raidz ZIOs, we can just copy aside the bad data read from the
2373  * disk, and use that to finish the checksum ereport later.
2374  */
2375 static void
2376 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2377     const void *good_buf)
2378 {
2379 	/* no processing needed */
2380 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2381 }
2382 
2383 /*ARGSUSED*/
2384 void
2385 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2386 {
2387 	void *buf = zio_buf_alloc(zio->io_size);
2388 
2389 	bcopy(zio->io_data, buf, zio->io_size);
2390 
2391 	zcr->zcr_cbinfo = zio->io_size;
2392 	zcr->zcr_cbdata = buf;
2393 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
2394 	zcr->zcr_free = zio_buf_free;
2395 }
2396 
2397 static int
2398 zio_vdev_io_assess(zio_t *zio)
2399 {
2400 	vdev_t *vd = zio->io_vd;
2401 
2402 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2403 		return (ZIO_PIPELINE_STOP);
2404 
2405 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2406 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2407 
2408 	if (zio->io_vsd != NULL) {
2409 		zio->io_vsd_ops->vsd_free(zio);
2410 		zio->io_vsd = NULL;
2411 	}
2412 
2413 	if (zio_injection_enabled && zio->io_error == 0)
2414 		zio->io_error = zio_handle_fault_injection(zio, EIO);
2415 
2416 	/*
2417 	 * If the I/O failed, determine whether we should attempt to retry it.
2418 	 *
2419 	 * On retry, we cut in line in the issue queue, since we don't want
2420 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2421 	 */
2422 	if (zio->io_error && vd == NULL &&
2423 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2424 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
2425 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
2426 		zio->io_error = 0;
2427 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
2428 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2429 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2430 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2431 		    zio_requeue_io_start_cut_in_line);
2432 		return (ZIO_PIPELINE_STOP);
2433 	}
2434 
2435 	/*
2436 	 * If we got an error on a leaf device, convert it to ENXIO
2437 	 * if the device is not accessible at all.
2438 	 */
2439 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2440 	    !vdev_accessible(vd, zio))
2441 		zio->io_error = ENXIO;
2442 
2443 	/*
2444 	 * If we can't write to an interior vdev (mirror or RAID-Z),
2445 	 * set vdev_cant_write so that we stop trying to allocate from it.
2446 	 */
2447 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2448 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2449 		vd->vdev_cant_write = B_TRUE;
2450 
2451 	if (zio->io_error)
2452 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2453 
2454 	return (ZIO_PIPELINE_CONTINUE);
2455 }
2456 
2457 void
2458 zio_vdev_io_reissue(zio_t *zio)
2459 {
2460 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2461 	ASSERT(zio->io_error == 0);
2462 
2463 	zio->io_stage >>= 1;
2464 }
2465 
2466 void
2467 zio_vdev_io_redone(zio_t *zio)
2468 {
2469 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2470 
2471 	zio->io_stage >>= 1;
2472 }
2473 
2474 void
2475 zio_vdev_io_bypass(zio_t *zio)
2476 {
2477 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2478 	ASSERT(zio->io_error == 0);
2479 
2480 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2481 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2482 }
2483 
2484 /*
2485  * ==========================================================================
2486  * Generate and verify checksums
2487  * ==========================================================================
2488  */
2489 static int
2490 zio_checksum_generate(zio_t *zio)
2491 {
2492 	blkptr_t *bp = zio->io_bp;
2493 	enum zio_checksum checksum;
2494 
2495 	if (bp == NULL) {
2496 		/*
2497 		 * This is zio_write_phys().
2498 		 * We're either generating a label checksum, or none at all.
2499 		 */
2500 		checksum = zio->io_prop.zp_checksum;
2501 
2502 		if (checksum == ZIO_CHECKSUM_OFF)
2503 			return (ZIO_PIPELINE_CONTINUE);
2504 
2505 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2506 	} else {
2507 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2508 			ASSERT(!IO_IS_ALLOCATING(zio));
2509 			checksum = ZIO_CHECKSUM_GANG_HEADER;
2510 		} else {
2511 			checksum = BP_GET_CHECKSUM(bp);
2512 		}
2513 	}
2514 
2515 	zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2516 
2517 	return (ZIO_PIPELINE_CONTINUE);
2518 }
2519 
2520 static int
2521 zio_checksum_verify(zio_t *zio)
2522 {
2523 	zio_bad_cksum_t info;
2524 	blkptr_t *bp = zio->io_bp;
2525 	int error;
2526 
2527 	ASSERT(zio->io_vd != NULL);
2528 
2529 	if (bp == NULL) {
2530 		/*
2531 		 * This is zio_read_phys().
2532 		 * We're either verifying a label checksum, or nothing at all.
2533 		 */
2534 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2535 			return (ZIO_PIPELINE_CONTINUE);
2536 
2537 		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2538 	}
2539 
2540 	if ((error = zio_checksum_error(zio, &info)) != 0) {
2541 		zio->io_error = error;
2542 		if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2543 			zfs_ereport_start_checksum(zio->io_spa,
2544 			    zio->io_vd, zio, zio->io_offset,
2545 			    zio->io_size, NULL, &info);
2546 		}
2547 	}
2548 
2549 	return (ZIO_PIPELINE_CONTINUE);
2550 }
2551 
2552 /*
2553  * Called by RAID-Z to ensure we don't compute the checksum twice.
2554  */
2555 void
2556 zio_checksum_verified(zio_t *zio)
2557 {
2558 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2559 }
2560 
2561 /*
2562  * ==========================================================================
2563  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2564  * An error of 0 indictes success.  ENXIO indicates whole-device failure,
2565  * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
2566  * indicate errors that are specific to one I/O, and most likely permanent.
2567  * Any other error is presumed to be worse because we weren't expecting it.
2568  * ==========================================================================
2569  */
2570 int
2571 zio_worst_error(int e1, int e2)
2572 {
2573 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2574 	int r1, r2;
2575 
2576 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2577 		if (e1 == zio_error_rank[r1])
2578 			break;
2579 
2580 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2581 		if (e2 == zio_error_rank[r2])
2582 			break;
2583 
2584 	return (r1 > r2 ? e1 : e2);
2585 }
2586 
2587 /*
2588  * ==========================================================================
2589  * I/O completion
2590  * ==========================================================================
2591  */
2592 static int
2593 zio_ready(zio_t *zio)
2594 {
2595 	blkptr_t *bp = zio->io_bp;
2596 	zio_t *pio, *pio_next;
2597 
2598 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2599 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2600 		return (ZIO_PIPELINE_STOP);
2601 
2602 	if (zio->io_ready) {
2603 		ASSERT(IO_IS_ALLOCATING(zio));
2604 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2605 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2606 
2607 		zio->io_ready(zio);
2608 	}
2609 
2610 	if (bp != NULL && bp != &zio->io_bp_copy)
2611 		zio->io_bp_copy = *bp;
2612 
2613 	if (zio->io_error)
2614 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2615 
2616 	mutex_enter(&zio->io_lock);
2617 	zio->io_state[ZIO_WAIT_READY] = 1;
2618 	pio = zio_walk_parents(zio);
2619 	mutex_exit(&zio->io_lock);
2620 
2621 	/*
2622 	 * As we notify zio's parents, new parents could be added.
2623 	 * New parents go to the head of zio's io_parent_list, however,
2624 	 * so we will (correctly) not notify them.  The remainder of zio's
2625 	 * io_parent_list, from 'pio_next' onward, cannot change because
2626 	 * all parents must wait for us to be done before they can be done.
2627 	 */
2628 	for (; pio != NULL; pio = pio_next) {
2629 		pio_next = zio_walk_parents(zio);
2630 		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2631 	}
2632 
2633 	if (zio->io_flags & ZIO_FLAG_NODATA) {
2634 		if (BP_IS_GANG(bp)) {
2635 			zio->io_flags &= ~ZIO_FLAG_NODATA;
2636 		} else {
2637 			ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2638 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2639 		}
2640 	}
2641 
2642 	if (zio_injection_enabled &&
2643 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
2644 		zio_handle_ignored_writes(zio);
2645 
2646 	return (ZIO_PIPELINE_CONTINUE);
2647 }
2648 
2649 static int
2650 zio_done(zio_t *zio)
2651 {
2652 	spa_t *spa = zio->io_spa;
2653 	zio_t *lio = zio->io_logical;
2654 	blkptr_t *bp = zio->io_bp;
2655 	vdev_t *vd = zio->io_vd;
2656 	uint64_t psize = zio->io_size;
2657 	zio_t *pio, *pio_next;
2658 
2659 	/*
2660 	 * If our children haven't all completed,
2661 	 * wait for them and then repeat this pipeline stage.
2662 	 */
2663 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2664 	    zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2665 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2666 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2667 		return (ZIO_PIPELINE_STOP);
2668 
2669 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2670 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2671 			ASSERT(zio->io_children[c][w] == 0);
2672 
2673 	if (bp != NULL) {
2674 		ASSERT(bp->blk_pad[0] == 0);
2675 		ASSERT(bp->blk_pad[1] == 0);
2676 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2677 		    (bp == zio_unique_parent(zio)->io_bp));
2678 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2679 		    zio->io_bp_override == NULL &&
2680 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2681 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
2682 			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2683 			ASSERT(BP_COUNT_GANG(bp) == 0 ||
2684 			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2685 		}
2686 	}
2687 
2688 	/*
2689 	 * If there were child vdev/gang/ddt errors, they apply to us now.
2690 	 */
2691 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2692 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2693 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2694 
2695 	/*
2696 	 * If the I/O on the transformed data was successful, generate any
2697 	 * checksum reports now while we still have the transformed data.
2698 	 */
2699 	if (zio->io_error == 0) {
2700 		while (zio->io_cksum_report != NULL) {
2701 			zio_cksum_report_t *zcr = zio->io_cksum_report;
2702 			uint64_t align = zcr->zcr_align;
2703 			uint64_t asize = P2ROUNDUP(psize, align);
2704 			char *abuf = zio->io_data;
2705 
2706 			if (asize != psize) {
2707 				abuf = zio_buf_alloc(asize);
2708 				bcopy(zio->io_data, abuf, psize);
2709 				bzero(abuf + psize, asize - psize);
2710 			}
2711 
2712 			zio->io_cksum_report = zcr->zcr_next;
2713 			zcr->zcr_next = NULL;
2714 			zcr->zcr_finish(zcr, abuf);
2715 			zfs_ereport_free_checksum(zcr);
2716 
2717 			if (asize != psize)
2718 				zio_buf_free(abuf, asize);
2719 		}
2720 	}
2721 
2722 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
2723 
2724 	vdev_stat_update(zio, psize);
2725 
2726 	if (zio->io_error) {
2727 		/*
2728 		 * If this I/O is attached to a particular vdev,
2729 		 * generate an error message describing the I/O failure
2730 		 * at the block level.  We ignore these errors if the
2731 		 * device is currently unavailable.
2732 		 */
2733 		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2734 			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2735 
2736 		if ((zio->io_error == EIO || !(zio->io_flags &
2737 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2738 		    zio == lio) {
2739 			/*
2740 			 * For logical I/O requests, tell the SPA to log the
2741 			 * error and generate a logical data ereport.
2742 			 */
2743 			spa_log_error(spa, zio);
2744 			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2745 			    0, 0);
2746 		}
2747 	}
2748 
2749 	if (zio->io_error && zio == lio) {
2750 		/*
2751 		 * Determine whether zio should be reexecuted.  This will
2752 		 * propagate all the way to the root via zio_notify_parent().
2753 		 */
2754 		ASSERT(vd == NULL && bp != NULL);
2755 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2756 
2757 		if (IO_IS_ALLOCATING(zio) &&
2758 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2759 			if (zio->io_error != ENOSPC)
2760 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2761 			else
2762 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2763 		}
2764 
2765 		if ((zio->io_type == ZIO_TYPE_READ ||
2766 		    zio->io_type == ZIO_TYPE_FREE) &&
2767 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2768 		    zio->io_error == ENXIO &&
2769 		    spa_load_state(spa) == SPA_LOAD_NONE &&
2770 		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2771 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2772 
2773 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2774 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2775 
2776 		/*
2777 		 * Here is a possibly good place to attempt to do
2778 		 * either combinatorial reconstruction or error correction
2779 		 * based on checksums.  It also might be a good place
2780 		 * to send out preliminary ereports before we suspend
2781 		 * processing.
2782 		 */
2783 	}
2784 
2785 	/*
2786 	 * If there were logical child errors, they apply to us now.
2787 	 * We defer this until now to avoid conflating logical child
2788 	 * errors with errors that happened to the zio itself when
2789 	 * updating vdev stats and reporting FMA events above.
2790 	 */
2791 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2792 
2793 	if ((zio->io_error || zio->io_reexecute) &&
2794 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2795 	    !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2796 		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2797 
2798 	zio_gang_tree_free(&zio->io_gang_tree);
2799 
2800 	/*
2801 	 * Godfather I/Os should never suspend.
2802 	 */
2803 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2804 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2805 		zio->io_reexecute = 0;
2806 
2807 	if (zio->io_reexecute) {
2808 		/*
2809 		 * This is a logical I/O that wants to reexecute.
2810 		 *
2811 		 * Reexecute is top-down.  When an i/o fails, if it's not
2812 		 * the root, it simply notifies its parent and sticks around.
2813 		 * The parent, seeing that it still has children in zio_done(),
2814 		 * does the same.  This percolates all the way up to the root.
2815 		 * The root i/o will reexecute or suspend the entire tree.
2816 		 *
2817 		 * This approach ensures that zio_reexecute() honors
2818 		 * all the original i/o dependency relationships, e.g.
2819 		 * parents not executing until children are ready.
2820 		 */
2821 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2822 
2823 		zio->io_gang_leader = NULL;
2824 
2825 		mutex_enter(&zio->io_lock);
2826 		zio->io_state[ZIO_WAIT_DONE] = 1;
2827 		mutex_exit(&zio->io_lock);
2828 
2829 		/*
2830 		 * "The Godfather" I/O monitors its children but is
2831 		 * not a true parent to them. It will track them through
2832 		 * the pipeline but severs its ties whenever they get into
2833 		 * trouble (e.g. suspended). This allows "The Godfather"
2834 		 * I/O to return status without blocking.
2835 		 */
2836 		for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2837 			zio_link_t *zl = zio->io_walk_link;
2838 			pio_next = zio_walk_parents(zio);
2839 
2840 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2841 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2842 				zio_remove_child(pio, zio, zl);
2843 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2844 			}
2845 		}
2846 
2847 		if ((pio = zio_unique_parent(zio)) != NULL) {
2848 			/*
2849 			 * We're not a root i/o, so there's nothing to do
2850 			 * but notify our parent.  Don't propagate errors
2851 			 * upward since we haven't permanently failed yet.
2852 			 */
2853 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2854 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2855 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2856 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2857 			/*
2858 			 * We'd fail again if we reexecuted now, so suspend
2859 			 * until conditions improve (e.g. device comes online).
2860 			 */
2861 			zio_suspend(spa, zio);
2862 		} else {
2863 			/*
2864 			 * Reexecution is potentially a huge amount of work.
2865 			 * Hand it off to the otherwise-unused claim taskq.
2866 			 */
2867 			(void) taskq_dispatch(
2868 			    spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2869 			    (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2870 		}
2871 		return (ZIO_PIPELINE_STOP);
2872 	}
2873 
2874 	ASSERT(zio->io_child_count == 0);
2875 	ASSERT(zio->io_reexecute == 0);
2876 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2877 
2878 	/*
2879 	 * Report any checksum errors, since the I/O is complete.
2880 	 */
2881 	while (zio->io_cksum_report != NULL) {
2882 		zio_cksum_report_t *zcr = zio->io_cksum_report;
2883 		zio->io_cksum_report = zcr->zcr_next;
2884 		zcr->zcr_next = NULL;
2885 		zcr->zcr_finish(zcr, NULL);
2886 		zfs_ereport_free_checksum(zcr);
2887 	}
2888 
2889 	/*
2890 	 * It is the responsibility of the done callback to ensure that this
2891 	 * particular zio is no longer discoverable for adoption, and as
2892 	 * such, cannot acquire any new parents.
2893 	 */
2894 	if (zio->io_done)
2895 		zio->io_done(zio);
2896 
2897 	mutex_enter(&zio->io_lock);
2898 	zio->io_state[ZIO_WAIT_DONE] = 1;
2899 	mutex_exit(&zio->io_lock);
2900 
2901 	for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2902 		zio_link_t *zl = zio->io_walk_link;
2903 		pio_next = zio_walk_parents(zio);
2904 		zio_remove_child(pio, zio, zl);
2905 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2906 	}
2907 
2908 	if (zio->io_waiter != NULL) {
2909 		mutex_enter(&zio->io_lock);
2910 		zio->io_executor = NULL;
2911 		cv_broadcast(&zio->io_cv);
2912 		mutex_exit(&zio->io_lock);
2913 	} else {
2914 		zio_destroy(zio);
2915 	}
2916 
2917 	return (ZIO_PIPELINE_STOP);
2918 }
2919 
2920 /*
2921  * ==========================================================================
2922  * I/O pipeline definition
2923  * ==========================================================================
2924  */
2925 static zio_pipe_stage_t *zio_pipeline[] = {
2926 	NULL,
2927 	zio_read_bp_init,
2928 	zio_free_bp_init,
2929 	zio_issue_async,
2930 	zio_write_bp_init,
2931 	zio_checksum_generate,
2932 	zio_ddt_read_start,
2933 	zio_ddt_read_done,
2934 	zio_ddt_write,
2935 	zio_ddt_free,
2936 	zio_gang_assemble,
2937 	zio_gang_issue,
2938 	zio_dva_allocate,
2939 	zio_dva_free,
2940 	zio_dva_claim,
2941 	zio_ready,
2942 	zio_vdev_io_start,
2943 	zio_vdev_io_done,
2944 	zio_vdev_io_assess,
2945 	zio_checksum_verify,
2946 	zio_done
2947 };
2948