xref: /titanic_51/usr/src/uts/common/fs/zfs/zio.c (revision 58865bb7f764a6ca11f3057bee77153724ebb239)
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  * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
24  * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25  */
26 
27 #include <sys/sysmacros.h>
28 #include <sys/zfs_context.h>
29 #include <sys/fm/fs/zfs.h>
30 #include <sys/spa.h>
31 #include <sys/txg.h>
32 #include <sys/spa_impl.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio_impl.h>
35 #include <sys/zio_compress.h>
36 #include <sys/zio_checksum.h>
37 #include <sys/dmu_objset.h>
38 #include <sys/arc.h>
39 #include <sys/ddt.h>
40 #include <sys/blkptr.h>
41 #include <sys/zfeature.h>
42 
43 /*
44  * ==========================================================================
45  * I/O type descriptions
46  * ==========================================================================
47  */
48 const char *zio_type_name[ZIO_TYPES] = {
49 	"zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
50 	"zio_ioctl"
51 };
52 
53 /*
54  * ==========================================================================
55  * I/O kmem caches
56  * ==========================================================================
57  */
58 kmem_cache_t *zio_cache;
59 kmem_cache_t *zio_link_cache;
60 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
61 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
62 
63 #ifdef _KERNEL
64 extern vmem_t *zio_alloc_arena;
65 #endif
66 
67 #define	ZIO_PIPELINE_CONTINUE		0x100
68 #define	ZIO_PIPELINE_STOP		0x101
69 
70 /*
71  * The following actions directly effect the spa's sync-to-convergence logic.
72  * The values below define the sync pass when we start performing the action.
73  * Care should be taken when changing these values as they directly impact
74  * spa_sync() performance. Tuning these values may introduce subtle performance
75  * pathologies and should only be done in the context of performance analysis.
76  * These tunables will eventually be removed and replaced with #defines once
77  * enough analysis has been done to determine optimal values.
78  *
79  * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
80  * regular blocks are not deferred.
81  */
82 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
83 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
84 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
85 
86 /*
87  * An allocating zio is one that either currently has the DVA allocate
88  * stage set or will have it later in its lifetime.
89  */
90 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
91 
92 boolean_t	zio_requeue_io_start_cut_in_line = B_TRUE;
93 
94 #ifdef ZFS_DEBUG
95 int zio_buf_debug_limit = 16384;
96 #else
97 int zio_buf_debug_limit = 0;
98 #endif
99 
100 void
101 zio_init(void)
102 {
103 	size_t c;
104 	vmem_t *data_alloc_arena = NULL;
105 
106 #ifdef _KERNEL
107 	data_alloc_arena = zio_alloc_arena;
108 #endif
109 	zio_cache = kmem_cache_create("zio_cache",
110 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
111 	zio_link_cache = kmem_cache_create("zio_link_cache",
112 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
113 
114 	/*
115 	 * For small buffers, we want a cache for each multiple of
116 	 * SPA_MINBLOCKSIZE.  For larger buffers, we want a cache
117 	 * for each quarter-power of 2.
118 	 */
119 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
120 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
121 		size_t p2 = size;
122 		size_t align = 0;
123 		size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
124 
125 		while (!ISP2(p2))
126 			p2 &= p2 - 1;
127 
128 #ifndef _KERNEL
129 		/*
130 		 * If we are using watchpoints, put each buffer on its own page,
131 		 * to eliminate the performance overhead of trapping to the
132 		 * kernel when modifying a non-watched buffer that shares the
133 		 * page with a watched buffer.
134 		 */
135 		if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
136 			continue;
137 #endif
138 		if (size <= 4 * SPA_MINBLOCKSIZE) {
139 			align = SPA_MINBLOCKSIZE;
140 		} else if (IS_P2ALIGNED(size, p2 >> 2)) {
141 			align = MIN(p2 >> 2, PAGESIZE);
142 		}
143 
144 		if (align != 0) {
145 			char name[36];
146 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
147 			zio_buf_cache[c] = kmem_cache_create(name, size,
148 			    align, NULL, NULL, NULL, NULL, NULL, cflags);
149 
150 			/*
151 			 * Since zio_data bufs do not appear in crash dumps, we
152 			 * pass KMC_NOTOUCH so that no allocator metadata is
153 			 * stored with the buffers.
154 			 */
155 			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
156 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
157 			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
158 			    cflags | KMC_NOTOUCH);
159 		}
160 	}
161 
162 	while (--c != 0) {
163 		ASSERT(zio_buf_cache[c] != NULL);
164 		if (zio_buf_cache[c - 1] == NULL)
165 			zio_buf_cache[c - 1] = zio_buf_cache[c];
166 
167 		ASSERT(zio_data_buf_cache[c] != NULL);
168 		if (zio_data_buf_cache[c - 1] == NULL)
169 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
170 	}
171 
172 	zio_inject_init();
173 }
174 
175 void
176 zio_fini(void)
177 {
178 	size_t c;
179 	kmem_cache_t *last_cache = NULL;
180 	kmem_cache_t *last_data_cache = NULL;
181 
182 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
183 		if (zio_buf_cache[c] != last_cache) {
184 			last_cache = zio_buf_cache[c];
185 			kmem_cache_destroy(zio_buf_cache[c]);
186 		}
187 		zio_buf_cache[c] = NULL;
188 
189 		if (zio_data_buf_cache[c] != last_data_cache) {
190 			last_data_cache = zio_data_buf_cache[c];
191 			kmem_cache_destroy(zio_data_buf_cache[c]);
192 		}
193 		zio_data_buf_cache[c] = NULL;
194 	}
195 
196 	kmem_cache_destroy(zio_link_cache);
197 	kmem_cache_destroy(zio_cache);
198 
199 	zio_inject_fini();
200 }
201 
202 /*
203  * ==========================================================================
204  * Allocate and free I/O buffers
205  * ==========================================================================
206  */
207 
208 /*
209  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
210  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
211  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
212  * excess / transient data in-core during a crashdump.
213  */
214 void *
215 zio_buf_alloc(size_t size)
216 {
217 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
218 
219 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
220 
221 	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
222 }
223 
224 /*
225  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
226  * crashdump if the kernel panics.  This exists so that we will limit the amount
227  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
228  * of kernel heap dumped to disk when the kernel panics)
229  */
230 void *
231 zio_data_buf_alloc(size_t size)
232 {
233 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
234 
235 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
236 
237 	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
238 }
239 
240 void
241 zio_buf_free(void *buf, size_t size)
242 {
243 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
244 
245 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
246 
247 	kmem_cache_free(zio_buf_cache[c], buf);
248 }
249 
250 void
251 zio_data_buf_free(void *buf, size_t size)
252 {
253 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
254 
255 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
256 
257 	kmem_cache_free(zio_data_buf_cache[c], buf);
258 }
259 
260 /*
261  * ==========================================================================
262  * Push and pop I/O transform buffers
263  * ==========================================================================
264  */
265 static void
266 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
267 	zio_transform_func_t *transform)
268 {
269 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
270 
271 	zt->zt_orig_data = zio->io_data;
272 	zt->zt_orig_size = zio->io_size;
273 	zt->zt_bufsize = bufsize;
274 	zt->zt_transform = transform;
275 
276 	zt->zt_next = zio->io_transform_stack;
277 	zio->io_transform_stack = zt;
278 
279 	zio->io_data = data;
280 	zio->io_size = size;
281 }
282 
283 static void
284 zio_pop_transforms(zio_t *zio)
285 {
286 	zio_transform_t *zt;
287 
288 	while ((zt = zio->io_transform_stack) != NULL) {
289 		if (zt->zt_transform != NULL)
290 			zt->zt_transform(zio,
291 			    zt->zt_orig_data, zt->zt_orig_size);
292 
293 		if (zt->zt_bufsize != 0)
294 			zio_buf_free(zio->io_data, zt->zt_bufsize);
295 
296 		zio->io_data = zt->zt_orig_data;
297 		zio->io_size = zt->zt_orig_size;
298 		zio->io_transform_stack = zt->zt_next;
299 
300 		kmem_free(zt, sizeof (zio_transform_t));
301 	}
302 }
303 
304 /*
305  * ==========================================================================
306  * I/O transform callbacks for subblocks and decompression
307  * ==========================================================================
308  */
309 static void
310 zio_subblock(zio_t *zio, void *data, uint64_t size)
311 {
312 	ASSERT(zio->io_size > size);
313 
314 	if (zio->io_type == ZIO_TYPE_READ)
315 		bcopy(zio->io_data, data, size);
316 }
317 
318 static void
319 zio_decompress(zio_t *zio, void *data, uint64_t size)
320 {
321 	if (zio->io_error == 0 &&
322 	    zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
323 	    zio->io_data, data, zio->io_size, size) != 0)
324 		zio->io_error = SET_ERROR(EIO);
325 }
326 
327 /*
328  * ==========================================================================
329  * I/O parent/child relationships and pipeline interlocks
330  * ==========================================================================
331  */
332 /*
333  * NOTE - Callers to zio_walk_parents() and zio_walk_children must
334  *        continue calling these functions until they return NULL.
335  *        Otherwise, the next caller will pick up the list walk in
336  *        some indeterminate state.  (Otherwise every caller would
337  *        have to pass in a cookie to keep the state represented by
338  *        io_walk_link, which gets annoying.)
339  */
340 zio_t *
341 zio_walk_parents(zio_t *cio)
342 {
343 	zio_link_t *zl = cio->io_walk_link;
344 	list_t *pl = &cio->io_parent_list;
345 
346 	zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
347 	cio->io_walk_link = zl;
348 
349 	if (zl == NULL)
350 		return (NULL);
351 
352 	ASSERT(zl->zl_child == cio);
353 	return (zl->zl_parent);
354 }
355 
356 zio_t *
357 zio_walk_children(zio_t *pio)
358 {
359 	zio_link_t *zl = pio->io_walk_link;
360 	list_t *cl = &pio->io_child_list;
361 
362 	zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
363 	pio->io_walk_link = zl;
364 
365 	if (zl == NULL)
366 		return (NULL);
367 
368 	ASSERT(zl->zl_parent == pio);
369 	return (zl->zl_child);
370 }
371 
372 zio_t *
373 zio_unique_parent(zio_t *cio)
374 {
375 	zio_t *pio = zio_walk_parents(cio);
376 
377 	VERIFY(zio_walk_parents(cio) == NULL);
378 	return (pio);
379 }
380 
381 void
382 zio_add_child(zio_t *pio, zio_t *cio)
383 {
384 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
385 
386 	/*
387 	 * Logical I/Os can have logical, gang, or vdev children.
388 	 * Gang I/Os can have gang or vdev children.
389 	 * Vdev I/Os can only have vdev children.
390 	 * The following ASSERT captures all of these constraints.
391 	 */
392 	ASSERT(cio->io_child_type <= pio->io_child_type);
393 
394 	zl->zl_parent = pio;
395 	zl->zl_child = cio;
396 
397 	mutex_enter(&cio->io_lock);
398 	mutex_enter(&pio->io_lock);
399 
400 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
401 
402 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
403 		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
404 
405 	list_insert_head(&pio->io_child_list, zl);
406 	list_insert_head(&cio->io_parent_list, zl);
407 
408 	pio->io_child_count++;
409 	cio->io_parent_count++;
410 
411 	mutex_exit(&pio->io_lock);
412 	mutex_exit(&cio->io_lock);
413 }
414 
415 static void
416 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
417 {
418 	ASSERT(zl->zl_parent == pio);
419 	ASSERT(zl->zl_child == cio);
420 
421 	mutex_enter(&cio->io_lock);
422 	mutex_enter(&pio->io_lock);
423 
424 	list_remove(&pio->io_child_list, zl);
425 	list_remove(&cio->io_parent_list, zl);
426 
427 	pio->io_child_count--;
428 	cio->io_parent_count--;
429 
430 	mutex_exit(&pio->io_lock);
431 	mutex_exit(&cio->io_lock);
432 
433 	kmem_cache_free(zio_link_cache, zl);
434 }
435 
436 static boolean_t
437 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
438 {
439 	uint64_t *countp = &zio->io_children[child][wait];
440 	boolean_t waiting = B_FALSE;
441 
442 	mutex_enter(&zio->io_lock);
443 	ASSERT(zio->io_stall == NULL);
444 	if (*countp != 0) {
445 		zio->io_stage >>= 1;
446 		zio->io_stall = countp;
447 		waiting = B_TRUE;
448 	}
449 	mutex_exit(&zio->io_lock);
450 
451 	return (waiting);
452 }
453 
454 static void
455 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
456 {
457 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
458 	int *errorp = &pio->io_child_error[zio->io_child_type];
459 
460 	mutex_enter(&pio->io_lock);
461 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
462 		*errorp = zio_worst_error(*errorp, zio->io_error);
463 	pio->io_reexecute |= zio->io_reexecute;
464 	ASSERT3U(*countp, >, 0);
465 
466 	(*countp)--;
467 
468 	if (*countp == 0 && pio->io_stall == countp) {
469 		pio->io_stall = NULL;
470 		mutex_exit(&pio->io_lock);
471 		zio_execute(pio);
472 	} else {
473 		mutex_exit(&pio->io_lock);
474 	}
475 }
476 
477 static void
478 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
479 {
480 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
481 		zio->io_error = zio->io_child_error[c];
482 }
483 
484 /*
485  * ==========================================================================
486  * Create the various types of I/O (read, write, free, etc)
487  * ==========================================================================
488  */
489 static zio_t *
490 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
491     void *data, uint64_t size, zio_done_func_t *done, void *private,
492     zio_type_t type, zio_priority_t priority, enum zio_flag flags,
493     vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
494     enum zio_stage stage, enum zio_stage pipeline)
495 {
496 	zio_t *zio;
497 
498 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
499 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
500 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
501 
502 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
503 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
504 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
505 
506 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
507 	bzero(zio, sizeof (zio_t));
508 
509 	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
510 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
511 
512 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
513 	    offsetof(zio_link_t, zl_parent_node));
514 	list_create(&zio->io_child_list, sizeof (zio_link_t),
515 	    offsetof(zio_link_t, zl_child_node));
516 
517 	if (vd != NULL)
518 		zio->io_child_type = ZIO_CHILD_VDEV;
519 	else if (flags & ZIO_FLAG_GANG_CHILD)
520 		zio->io_child_type = ZIO_CHILD_GANG;
521 	else if (flags & ZIO_FLAG_DDT_CHILD)
522 		zio->io_child_type = ZIO_CHILD_DDT;
523 	else
524 		zio->io_child_type = ZIO_CHILD_LOGICAL;
525 
526 	if (bp != NULL) {
527 		zio->io_bp = (blkptr_t *)bp;
528 		zio->io_bp_copy = *bp;
529 		zio->io_bp_orig = *bp;
530 		if (type != ZIO_TYPE_WRITE ||
531 		    zio->io_child_type == ZIO_CHILD_DDT)
532 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
533 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
534 			zio->io_logical = zio;
535 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
536 			pipeline |= ZIO_GANG_STAGES;
537 	}
538 
539 	zio->io_spa = spa;
540 	zio->io_txg = txg;
541 	zio->io_done = done;
542 	zio->io_private = private;
543 	zio->io_type = type;
544 	zio->io_priority = priority;
545 	zio->io_vd = vd;
546 	zio->io_offset = offset;
547 	zio->io_orig_data = zio->io_data = data;
548 	zio->io_orig_size = zio->io_size = size;
549 	zio->io_orig_flags = zio->io_flags = flags;
550 	zio->io_orig_stage = zio->io_stage = stage;
551 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
552 
553 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
554 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
555 
556 	if (zb != NULL)
557 		zio->io_bookmark = *zb;
558 
559 	if (pio != NULL) {
560 		if (zio->io_logical == NULL)
561 			zio->io_logical = pio->io_logical;
562 		if (zio->io_child_type == ZIO_CHILD_GANG)
563 			zio->io_gang_leader = pio->io_gang_leader;
564 		zio_add_child(pio, zio);
565 	}
566 
567 	return (zio);
568 }
569 
570 static void
571 zio_destroy(zio_t *zio)
572 {
573 	list_destroy(&zio->io_parent_list);
574 	list_destroy(&zio->io_child_list);
575 	mutex_destroy(&zio->io_lock);
576 	cv_destroy(&zio->io_cv);
577 	kmem_cache_free(zio_cache, zio);
578 }
579 
580 zio_t *
581 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
582     void *private, enum zio_flag flags)
583 {
584 	zio_t *zio;
585 
586 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
587 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
588 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
589 
590 	return (zio);
591 }
592 
593 zio_t *
594 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
595 {
596 	return (zio_null(NULL, spa, NULL, done, private, flags));
597 }
598 
599 void
600 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
601 {
602 	if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
603 		zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
604 		    bp, (longlong_t)BP_GET_TYPE(bp));
605 	}
606 	if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
607 	    BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
608 		zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
609 		    bp, (longlong_t)BP_GET_CHECKSUM(bp));
610 	}
611 	if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
612 	    BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
613 		zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
614 		    bp, (longlong_t)BP_GET_COMPRESS(bp));
615 	}
616 	if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
617 		zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
618 		    bp, (longlong_t)BP_GET_LSIZE(bp));
619 	}
620 	if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
621 		zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
622 		    bp, (longlong_t)BP_GET_PSIZE(bp));
623 	}
624 
625 	if (BP_IS_EMBEDDED(bp)) {
626 		if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
627 			zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
628 			    bp, (longlong_t)BPE_GET_ETYPE(bp));
629 		}
630 	}
631 
632 	/*
633 	 * Pool-specific checks.
634 	 *
635 	 * Note: it would be nice to verify that the blk_birth and
636 	 * BP_PHYSICAL_BIRTH() are not too large.  However, spa_freeze()
637 	 * allows the birth time of log blocks (and dmu_sync()-ed blocks
638 	 * that are in the log) to be arbitrarily large.
639 	 */
640 	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
641 		uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
642 		if (vdevid >= spa->spa_root_vdev->vdev_children) {
643 			zfs_panic_recover("blkptr at %p DVA %u has invalid "
644 			    "VDEV %llu",
645 			    bp, i, (longlong_t)vdevid);
646 		}
647 		vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
648 		if (vd == NULL) {
649 			zfs_panic_recover("blkptr at %p DVA %u has invalid "
650 			    "VDEV %llu",
651 			    bp, i, (longlong_t)vdevid);
652 		}
653 		if (vd->vdev_ops == &vdev_hole_ops) {
654 			zfs_panic_recover("blkptr at %p DVA %u has hole "
655 			    "VDEV %llu",
656 			    bp, i, (longlong_t)vdevid);
657 
658 		}
659 		if (vd->vdev_ops == &vdev_missing_ops) {
660 			/*
661 			 * "missing" vdevs are valid during import, but we
662 			 * don't have their detailed info (e.g. asize), so
663 			 * we can't perform any more checks on them.
664 			 */
665 			continue;
666 		}
667 		uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
668 		uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
669 		if (BP_IS_GANG(bp))
670 			asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
671 		if (offset + asize > vd->vdev_asize) {
672 			zfs_panic_recover("blkptr at %p DVA %u has invalid "
673 			    "OFFSET %llu",
674 			    bp, i, (longlong_t)offset);
675 		}
676 	}
677 }
678 
679 zio_t *
680 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
681     void *data, uint64_t size, zio_done_func_t *done, void *private,
682     zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
683 {
684 	zio_t *zio;
685 
686 	zfs_blkptr_verify(spa, bp);
687 
688 	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
689 	    data, size, done, private,
690 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
691 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
692 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
693 
694 	return (zio);
695 }
696 
697 zio_t *
698 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
699     void *data, uint64_t size, const zio_prop_t *zp,
700     zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
701     void *private,
702     zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
703 {
704 	zio_t *zio;
705 
706 	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
707 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
708 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
709 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
710 	    DMU_OT_IS_VALID(zp->zp_type) &&
711 	    zp->zp_level < 32 &&
712 	    zp->zp_copies > 0 &&
713 	    zp->zp_copies <= spa_max_replication(spa));
714 
715 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
716 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
717 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
718 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
719 
720 	zio->io_ready = ready;
721 	zio->io_physdone = physdone;
722 	zio->io_prop = *zp;
723 
724 	/*
725 	 * Data can be NULL if we are going to call zio_write_override() to
726 	 * provide the already-allocated BP.  But we may need the data to
727 	 * verify a dedup hit (if requested).  In this case, don't try to
728 	 * dedup (just take the already-allocated BP verbatim).
729 	 */
730 	if (data == NULL && zio->io_prop.zp_dedup_verify) {
731 		zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
732 	}
733 
734 	return (zio);
735 }
736 
737 zio_t *
738 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
739     uint64_t size, zio_done_func_t *done, void *private,
740     zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
741 {
742 	zio_t *zio;
743 
744 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
745 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
746 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
747 
748 	return (zio);
749 }
750 
751 void
752 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
753 {
754 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
755 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
756 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
757 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
758 
759 	/*
760 	 * We must reset the io_prop to match the values that existed
761 	 * when the bp was first written by dmu_sync() keeping in mind
762 	 * that nopwrite and dedup are mutually exclusive.
763 	 */
764 	zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
765 	zio->io_prop.zp_nopwrite = nopwrite;
766 	zio->io_prop.zp_copies = copies;
767 	zio->io_bp_override = bp;
768 }
769 
770 void
771 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
772 {
773 
774 	/*
775 	 * The check for EMBEDDED is a performance optimization.  We
776 	 * process the free here (by ignoring it) rather than
777 	 * putting it on the list and then processing it in zio_free_sync().
778 	 */
779 	if (BP_IS_EMBEDDED(bp))
780 		return;
781 	metaslab_check_free(spa, bp);
782 
783 	/*
784 	 * Frees that are for the currently-syncing txg, are not going to be
785 	 * deferred, and which will not need to do a read (i.e. not GANG or
786 	 * DEDUP), can be processed immediately.  Otherwise, put them on the
787 	 * in-memory list for later processing.
788 	 */
789 	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
790 	    txg != spa->spa_syncing_txg ||
791 	    spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
792 		bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
793 	} else {
794 		VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
795 	}
796 }
797 
798 zio_t *
799 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
800     enum zio_flag flags)
801 {
802 	zio_t *zio;
803 	enum zio_stage stage = ZIO_FREE_PIPELINE;
804 
805 	ASSERT(!BP_IS_HOLE(bp));
806 	ASSERT(spa_syncing_txg(spa) == txg);
807 	ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
808 
809 	if (BP_IS_EMBEDDED(bp))
810 		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
811 
812 	metaslab_check_free(spa, bp);
813 	arc_freed(spa, bp);
814 
815 	/*
816 	 * GANG and DEDUP blocks can induce a read (for the gang block header,
817 	 * or the DDT), so issue them asynchronously so that this thread is
818 	 * not tied up.
819 	 */
820 	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
821 		stage |= ZIO_STAGE_ISSUE_ASYNC;
822 
823 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
824 	    NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
825 	    NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
826 
827 	return (zio);
828 }
829 
830 zio_t *
831 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
832     zio_done_func_t *done, void *private, enum zio_flag flags)
833 {
834 	zio_t *zio;
835 
836 	dprintf_bp(bp, "claiming in txg %llu", txg);
837 
838 	if (BP_IS_EMBEDDED(bp))
839 		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
840 
841 	/*
842 	 * A claim is an allocation of a specific block.  Claims are needed
843 	 * to support immediate writes in the intent log.  The issue is that
844 	 * immediate writes contain committed data, but in a txg that was
845 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
846 	 * the intent log claims all blocks that contain immediate write data
847 	 * so that the SPA knows they're in use.
848 	 *
849 	 * All claims *must* be resolved in the first txg -- before the SPA
850 	 * starts allocating blocks -- so that nothing is allocated twice.
851 	 * If txg == 0 we just verify that the block is claimable.
852 	 */
853 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
854 	ASSERT(txg == spa_first_txg(spa) || txg == 0);
855 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
856 
857 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
858 	    done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
859 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
860 
861 	return (zio);
862 }
863 
864 zio_t *
865 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
866     zio_done_func_t *done, void *private, enum zio_flag flags)
867 {
868 	zio_t *zio;
869 	int c;
870 
871 	if (vd->vdev_children == 0) {
872 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
873 		    ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
874 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
875 
876 		zio->io_cmd = cmd;
877 	} else {
878 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
879 
880 		for (c = 0; c < vd->vdev_children; c++)
881 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
882 			    done, private, flags));
883 	}
884 
885 	return (zio);
886 }
887 
888 zio_t *
889 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
890     void *data, int checksum, zio_done_func_t *done, void *private,
891     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
892 {
893 	zio_t *zio;
894 
895 	ASSERT(vd->vdev_children == 0);
896 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
897 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
898 	ASSERT3U(offset + size, <=, vd->vdev_psize);
899 
900 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
901 	    ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
902 	    NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
903 
904 	zio->io_prop.zp_checksum = checksum;
905 
906 	return (zio);
907 }
908 
909 zio_t *
910 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
911     void *data, int checksum, zio_done_func_t *done, void *private,
912     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
913 {
914 	zio_t *zio;
915 
916 	ASSERT(vd->vdev_children == 0);
917 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
918 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
919 	ASSERT3U(offset + size, <=, vd->vdev_psize);
920 
921 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
922 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
923 	    NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
924 
925 	zio->io_prop.zp_checksum = checksum;
926 
927 	if (zio_checksum_table[checksum].ci_eck) {
928 		/*
929 		 * zec checksums are necessarily destructive -- they modify
930 		 * the end of the write buffer to hold the verifier/checksum.
931 		 * Therefore, we must make a local copy in case the data is
932 		 * being written to multiple places in parallel.
933 		 */
934 		void *wbuf = zio_buf_alloc(size);
935 		bcopy(data, wbuf, size);
936 		zio_push_transform(zio, wbuf, size, size, NULL);
937 	}
938 
939 	return (zio);
940 }
941 
942 /*
943  * Create a child I/O to do some work for us.
944  */
945 zio_t *
946 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
947 	void *data, uint64_t size, int type, zio_priority_t priority,
948 	enum zio_flag flags, zio_done_func_t *done, void *private)
949 {
950 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
951 	zio_t *zio;
952 
953 	ASSERT(vd->vdev_parent ==
954 	    (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
955 
956 	if (type == ZIO_TYPE_READ && bp != NULL) {
957 		/*
958 		 * If we have the bp, then the child should perform the
959 		 * checksum and the parent need not.  This pushes error
960 		 * detection as close to the leaves as possible and
961 		 * eliminates redundant checksums in the interior nodes.
962 		 */
963 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
964 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
965 	}
966 
967 	if (vd->vdev_children == 0)
968 		offset += VDEV_LABEL_START_SIZE;
969 
970 	flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
971 
972 	/*
973 	 * If we've decided to do a repair, the write is not speculative --
974 	 * even if the original read was.
975 	 */
976 	if (flags & ZIO_FLAG_IO_REPAIR)
977 		flags &= ~ZIO_FLAG_SPECULATIVE;
978 
979 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
980 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
981 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
982 
983 	zio->io_physdone = pio->io_physdone;
984 	if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
985 		zio->io_logical->io_phys_children++;
986 
987 	return (zio);
988 }
989 
990 zio_t *
991 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
992 	int type, zio_priority_t priority, enum zio_flag flags,
993 	zio_done_func_t *done, void *private)
994 {
995 	zio_t *zio;
996 
997 	ASSERT(vd->vdev_ops->vdev_op_leaf);
998 
999 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1000 	    data, size, done, private, type, priority,
1001 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1002 	    vd, offset, NULL,
1003 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1004 
1005 	return (zio);
1006 }
1007 
1008 void
1009 zio_flush(zio_t *zio, vdev_t *vd)
1010 {
1011 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1012 	    NULL, NULL,
1013 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1014 }
1015 
1016 void
1017 zio_shrink(zio_t *zio, uint64_t size)
1018 {
1019 	ASSERT(zio->io_executor == NULL);
1020 	ASSERT(zio->io_orig_size == zio->io_size);
1021 	ASSERT(size <= zio->io_size);
1022 
1023 	/*
1024 	 * We don't shrink for raidz because of problems with the
1025 	 * reconstruction when reading back less than the block size.
1026 	 * Note, BP_IS_RAIDZ() assumes no compression.
1027 	 */
1028 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1029 	if (!BP_IS_RAIDZ(zio->io_bp))
1030 		zio->io_orig_size = zio->io_size = size;
1031 }
1032 
1033 /*
1034  * ==========================================================================
1035  * Prepare to read and write logical blocks
1036  * ==========================================================================
1037  */
1038 
1039 static int
1040 zio_read_bp_init(zio_t *zio)
1041 {
1042 	blkptr_t *bp = zio->io_bp;
1043 
1044 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1045 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
1046 	    !(zio->io_flags & ZIO_FLAG_RAW)) {
1047 		uint64_t psize =
1048 		    BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1049 		void *cbuf = zio_buf_alloc(psize);
1050 
1051 		zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1052 	}
1053 
1054 	if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1055 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1056 		decode_embedded_bp_compressed(bp, zio->io_data);
1057 	} else {
1058 		ASSERT(!BP_IS_EMBEDDED(bp));
1059 	}
1060 
1061 	if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1062 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1063 
1064 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1065 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1066 
1067 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1068 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1069 
1070 	return (ZIO_PIPELINE_CONTINUE);
1071 }
1072 
1073 static int
1074 zio_write_bp_init(zio_t *zio)
1075 {
1076 	spa_t *spa = zio->io_spa;
1077 	zio_prop_t *zp = &zio->io_prop;
1078 	enum zio_compress compress = zp->zp_compress;
1079 	blkptr_t *bp = zio->io_bp;
1080 	uint64_t lsize = zio->io_size;
1081 	uint64_t psize = lsize;
1082 	int pass = 1;
1083 
1084 	/*
1085 	 * If our children haven't all reached the ready stage,
1086 	 * wait for them and then repeat this pipeline stage.
1087 	 */
1088 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1089 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1090 		return (ZIO_PIPELINE_STOP);
1091 
1092 	if (!IO_IS_ALLOCATING(zio))
1093 		return (ZIO_PIPELINE_CONTINUE);
1094 
1095 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1096 
1097 	if (zio->io_bp_override) {
1098 		ASSERT(bp->blk_birth != zio->io_txg);
1099 		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1100 
1101 		*bp = *zio->io_bp_override;
1102 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1103 
1104 		if (BP_IS_EMBEDDED(bp))
1105 			return (ZIO_PIPELINE_CONTINUE);
1106 
1107 		/*
1108 		 * If we've been overridden and nopwrite is set then
1109 		 * set the flag accordingly to indicate that a nopwrite
1110 		 * has already occurred.
1111 		 */
1112 		if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1113 			ASSERT(!zp->zp_dedup);
1114 			zio->io_flags |= ZIO_FLAG_NOPWRITE;
1115 			return (ZIO_PIPELINE_CONTINUE);
1116 		}
1117 
1118 		ASSERT(!zp->zp_nopwrite);
1119 
1120 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1121 			return (ZIO_PIPELINE_CONTINUE);
1122 
1123 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1124 		    zp->zp_dedup_verify);
1125 
1126 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1127 			BP_SET_DEDUP(bp, 1);
1128 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1129 			return (ZIO_PIPELINE_CONTINUE);
1130 		}
1131 	}
1132 
1133 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1134 		/*
1135 		 * We're rewriting an existing block, which means we're
1136 		 * working on behalf of spa_sync().  For spa_sync() to
1137 		 * converge, it must eventually be the case that we don't
1138 		 * have to allocate new blocks.  But compression changes
1139 		 * the blocksize, which forces a reallocate, and makes
1140 		 * convergence take longer.  Therefore, after the first
1141 		 * few passes, stop compressing to ensure convergence.
1142 		 */
1143 		pass = spa_sync_pass(spa);
1144 
1145 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
1146 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1147 		ASSERT(!BP_GET_DEDUP(bp));
1148 
1149 		if (pass >= zfs_sync_pass_dont_compress)
1150 			compress = ZIO_COMPRESS_OFF;
1151 
1152 		/* Make sure someone doesn't change their mind on overwrites */
1153 		ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1154 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1155 	}
1156 
1157 	if (compress != ZIO_COMPRESS_OFF) {
1158 		void *cbuf = zio_buf_alloc(lsize);
1159 		psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1160 		if (psize == 0 || psize == lsize) {
1161 			compress = ZIO_COMPRESS_OFF;
1162 			zio_buf_free(cbuf, lsize);
1163 		} else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1164 		    zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1165 		    spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1166 			encode_embedded_bp_compressed(bp,
1167 			    cbuf, compress, lsize, psize);
1168 			BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1169 			BP_SET_TYPE(bp, zio->io_prop.zp_type);
1170 			BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1171 			zio_buf_free(cbuf, lsize);
1172 			bp->blk_birth = zio->io_txg;
1173 			zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1174 			ASSERT(spa_feature_is_active(spa,
1175 			    SPA_FEATURE_EMBEDDED_DATA));
1176 			return (ZIO_PIPELINE_CONTINUE);
1177 		} else {
1178 			/*
1179 			 * Round up compressed size up to the ashift
1180 			 * of the smallest-ashift device, and zero the tail.
1181 			 * This ensures that the compressed size of the BP
1182 			 * (and thus compressratio property) are correct,
1183 			 * in that we charge for the padding used to fill out
1184 			 * the last sector.
1185 			 */
1186 			ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1187 			size_t rounded = (size_t)P2ROUNDUP(psize,
1188 			    1ULL << spa->spa_min_ashift);
1189 			if (rounded >= lsize) {
1190 				compress = ZIO_COMPRESS_OFF;
1191 				zio_buf_free(cbuf, lsize);
1192 				psize = lsize;
1193 			} else {
1194 				bzero((char *)cbuf + psize, rounded - psize);
1195 				psize = rounded;
1196 				zio_push_transform(zio, cbuf,
1197 				    psize, lsize, NULL);
1198 			}
1199 		}
1200 	}
1201 
1202 	/*
1203 	 * The final pass of spa_sync() must be all rewrites, but the first
1204 	 * few passes offer a trade-off: allocating blocks defers convergence,
1205 	 * but newly allocated blocks are sequential, so they can be written
1206 	 * to disk faster.  Therefore, we allow the first few passes of
1207 	 * spa_sync() to allocate new blocks, but force rewrites after that.
1208 	 * There should only be a handful of blocks after pass 1 in any case.
1209 	 */
1210 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1211 	    BP_GET_PSIZE(bp) == psize &&
1212 	    pass >= zfs_sync_pass_rewrite) {
1213 		ASSERT(psize != 0);
1214 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1215 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1216 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1217 	} else {
1218 		BP_ZERO(bp);
1219 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1220 	}
1221 
1222 	if (psize == 0) {
1223 		if (zio->io_bp_orig.blk_birth != 0 &&
1224 		    spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1225 			BP_SET_LSIZE(bp, lsize);
1226 			BP_SET_TYPE(bp, zp->zp_type);
1227 			BP_SET_LEVEL(bp, zp->zp_level);
1228 			BP_SET_BIRTH(bp, zio->io_txg, 0);
1229 		}
1230 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1231 	} else {
1232 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1233 		BP_SET_LSIZE(bp, lsize);
1234 		BP_SET_TYPE(bp, zp->zp_type);
1235 		BP_SET_LEVEL(bp, zp->zp_level);
1236 		BP_SET_PSIZE(bp, psize);
1237 		BP_SET_COMPRESS(bp, compress);
1238 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1239 		BP_SET_DEDUP(bp, zp->zp_dedup);
1240 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1241 		if (zp->zp_dedup) {
1242 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1243 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1244 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1245 		}
1246 		if (zp->zp_nopwrite) {
1247 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1248 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1249 			zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1250 		}
1251 	}
1252 
1253 	return (ZIO_PIPELINE_CONTINUE);
1254 }
1255 
1256 static int
1257 zio_free_bp_init(zio_t *zio)
1258 {
1259 	blkptr_t *bp = zio->io_bp;
1260 
1261 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1262 		if (BP_GET_DEDUP(bp))
1263 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1264 	}
1265 
1266 	return (ZIO_PIPELINE_CONTINUE);
1267 }
1268 
1269 /*
1270  * ==========================================================================
1271  * Execute the I/O pipeline
1272  * ==========================================================================
1273  */
1274 
1275 static void
1276 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1277 {
1278 	spa_t *spa = zio->io_spa;
1279 	zio_type_t t = zio->io_type;
1280 	int flags = (cutinline ? TQ_FRONT : 0);
1281 
1282 	/*
1283 	 * If we're a config writer or a probe, the normal issue and
1284 	 * interrupt threads may all be blocked waiting for the config lock.
1285 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1286 	 */
1287 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1288 		t = ZIO_TYPE_NULL;
1289 
1290 	/*
1291 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1292 	 */
1293 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1294 		t = ZIO_TYPE_NULL;
1295 
1296 	/*
1297 	 * If this is a high priority I/O, then use the high priority taskq if
1298 	 * available.
1299 	 */
1300 	if (zio->io_priority == ZIO_PRIORITY_NOW &&
1301 	    spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1302 		q++;
1303 
1304 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1305 
1306 	/*
1307 	 * NB: We are assuming that the zio can only be dispatched
1308 	 * to a single taskq at a time.  It would be a grievous error
1309 	 * to dispatch the zio to another taskq at the same time.
1310 	 */
1311 	ASSERT(zio->io_tqent.tqent_next == NULL);
1312 	spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1313 	    flags, &zio->io_tqent);
1314 }
1315 
1316 static boolean_t
1317 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1318 {
1319 	kthread_t *executor = zio->io_executor;
1320 	spa_t *spa = zio->io_spa;
1321 
1322 	for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1323 		spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1324 		uint_t i;
1325 		for (i = 0; i < tqs->stqs_count; i++) {
1326 			if (taskq_member(tqs->stqs_taskq[i], executor))
1327 				return (B_TRUE);
1328 		}
1329 	}
1330 
1331 	return (B_FALSE);
1332 }
1333 
1334 static int
1335 zio_issue_async(zio_t *zio)
1336 {
1337 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1338 
1339 	return (ZIO_PIPELINE_STOP);
1340 }
1341 
1342 void
1343 zio_interrupt(zio_t *zio)
1344 {
1345 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1346 }
1347 
1348 /*
1349  * Execute the I/O pipeline until one of the following occurs:
1350  *
1351  *	(1) the I/O completes
1352  *	(2) the pipeline stalls waiting for dependent child I/Os
1353  *	(3) the I/O issues, so we're waiting for an I/O completion interrupt
1354  *	(4) the I/O is delegated by vdev-level caching or aggregation
1355  *	(5) the I/O is deferred due to vdev-level queueing
1356  *	(6) the I/O is handed off to another thread.
1357  *
1358  * In all cases, the pipeline stops whenever there's no CPU work; it never
1359  * burns a thread in cv_wait().
1360  *
1361  * There's no locking on io_stage because there's no legitimate way
1362  * for multiple threads to be attempting to process the same I/O.
1363  */
1364 static zio_pipe_stage_t *zio_pipeline[];
1365 
1366 void
1367 zio_execute(zio_t *zio)
1368 {
1369 	zio->io_executor = curthread;
1370 
1371 	while (zio->io_stage < ZIO_STAGE_DONE) {
1372 		enum zio_stage pipeline = zio->io_pipeline;
1373 		enum zio_stage stage = zio->io_stage;
1374 		int rv;
1375 
1376 		ASSERT(!MUTEX_HELD(&zio->io_lock));
1377 		ASSERT(ISP2(stage));
1378 		ASSERT(zio->io_stall == NULL);
1379 
1380 		do {
1381 			stage <<= 1;
1382 		} while ((stage & pipeline) == 0);
1383 
1384 		ASSERT(stage <= ZIO_STAGE_DONE);
1385 
1386 		/*
1387 		 * If we are in interrupt context and this pipeline stage
1388 		 * will grab a config lock that is held across I/O,
1389 		 * or may wait for an I/O that needs an interrupt thread
1390 		 * to complete, issue async to avoid deadlock.
1391 		 *
1392 		 * For VDEV_IO_START, we cut in line so that the io will
1393 		 * be sent to disk promptly.
1394 		 */
1395 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1396 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1397 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1398 			    zio_requeue_io_start_cut_in_line : B_FALSE;
1399 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1400 			return;
1401 		}
1402 
1403 		zio->io_stage = stage;
1404 		rv = zio_pipeline[highbit64(stage) - 1](zio);
1405 
1406 		if (rv == ZIO_PIPELINE_STOP)
1407 			return;
1408 
1409 		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1410 	}
1411 }
1412 
1413 /*
1414  * ==========================================================================
1415  * Initiate I/O, either sync or async
1416  * ==========================================================================
1417  */
1418 int
1419 zio_wait(zio_t *zio)
1420 {
1421 	int error;
1422 
1423 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1424 	ASSERT(zio->io_executor == NULL);
1425 
1426 	zio->io_waiter = curthread;
1427 
1428 	zio_execute(zio);
1429 
1430 	mutex_enter(&zio->io_lock);
1431 	while (zio->io_executor != NULL)
1432 		cv_wait(&zio->io_cv, &zio->io_lock);
1433 	mutex_exit(&zio->io_lock);
1434 
1435 	error = zio->io_error;
1436 	zio_destroy(zio);
1437 
1438 	return (error);
1439 }
1440 
1441 void
1442 zio_nowait(zio_t *zio)
1443 {
1444 	ASSERT(zio->io_executor == NULL);
1445 
1446 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1447 	    zio_unique_parent(zio) == NULL) {
1448 		/*
1449 		 * This is a logical async I/O with no parent to wait for it.
1450 		 * We add it to the spa_async_root_zio "Godfather" I/O which
1451 		 * will ensure they complete prior to unloading the pool.
1452 		 */
1453 		spa_t *spa = zio->io_spa;
1454 
1455 		zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1456 	}
1457 
1458 	zio_execute(zio);
1459 }
1460 
1461 /*
1462  * ==========================================================================
1463  * Reexecute or suspend/resume failed I/O
1464  * ==========================================================================
1465  */
1466 
1467 static void
1468 zio_reexecute(zio_t *pio)
1469 {
1470 	zio_t *cio, *cio_next;
1471 
1472 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1473 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1474 	ASSERT(pio->io_gang_leader == NULL);
1475 	ASSERT(pio->io_gang_tree == NULL);
1476 
1477 	pio->io_flags = pio->io_orig_flags;
1478 	pio->io_stage = pio->io_orig_stage;
1479 	pio->io_pipeline = pio->io_orig_pipeline;
1480 	pio->io_reexecute = 0;
1481 	pio->io_flags |= ZIO_FLAG_REEXECUTED;
1482 	pio->io_error = 0;
1483 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1484 		pio->io_state[w] = 0;
1485 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1486 		pio->io_child_error[c] = 0;
1487 
1488 	if (IO_IS_ALLOCATING(pio))
1489 		BP_ZERO(pio->io_bp);
1490 
1491 	/*
1492 	 * As we reexecute pio's children, new children could be created.
1493 	 * New children go to the head of pio's io_child_list, however,
1494 	 * so we will (correctly) not reexecute them.  The key is that
1495 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1496 	 * cannot be affected by any side effects of reexecuting 'cio'.
1497 	 */
1498 	for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1499 		cio_next = zio_walk_children(pio);
1500 		mutex_enter(&pio->io_lock);
1501 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1502 			pio->io_children[cio->io_child_type][w]++;
1503 		mutex_exit(&pio->io_lock);
1504 		zio_reexecute(cio);
1505 	}
1506 
1507 	/*
1508 	 * Now that all children have been reexecuted, execute the parent.
1509 	 * We don't reexecute "The Godfather" I/O here as it's the
1510 	 * responsibility of the caller to wait on him.
1511 	 */
1512 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1513 		zio_execute(pio);
1514 }
1515 
1516 void
1517 zio_suspend(spa_t *spa, zio_t *zio)
1518 {
1519 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1520 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1521 		    "failure and the failure mode property for this pool "
1522 		    "is set to panic.", spa_name(spa));
1523 
1524 	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1525 
1526 	mutex_enter(&spa->spa_suspend_lock);
1527 
1528 	if (spa->spa_suspend_zio_root == NULL)
1529 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1530 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1531 		    ZIO_FLAG_GODFATHER);
1532 
1533 	spa->spa_suspended = B_TRUE;
1534 
1535 	if (zio != NULL) {
1536 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1537 		ASSERT(zio != spa->spa_suspend_zio_root);
1538 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1539 		ASSERT(zio_unique_parent(zio) == NULL);
1540 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1541 		zio_add_child(spa->spa_suspend_zio_root, zio);
1542 	}
1543 
1544 	mutex_exit(&spa->spa_suspend_lock);
1545 }
1546 
1547 int
1548 zio_resume(spa_t *spa)
1549 {
1550 	zio_t *pio;
1551 
1552 	/*
1553 	 * Reexecute all previously suspended i/o.
1554 	 */
1555 	mutex_enter(&spa->spa_suspend_lock);
1556 	spa->spa_suspended = B_FALSE;
1557 	cv_broadcast(&spa->spa_suspend_cv);
1558 	pio = spa->spa_suspend_zio_root;
1559 	spa->spa_suspend_zio_root = NULL;
1560 	mutex_exit(&spa->spa_suspend_lock);
1561 
1562 	if (pio == NULL)
1563 		return (0);
1564 
1565 	zio_reexecute(pio);
1566 	return (zio_wait(pio));
1567 }
1568 
1569 void
1570 zio_resume_wait(spa_t *spa)
1571 {
1572 	mutex_enter(&spa->spa_suspend_lock);
1573 	while (spa_suspended(spa))
1574 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1575 	mutex_exit(&spa->spa_suspend_lock);
1576 }
1577 
1578 /*
1579  * ==========================================================================
1580  * Gang blocks.
1581  *
1582  * A gang block is a collection of small blocks that looks to the DMU
1583  * like one large block.  When zio_dva_allocate() cannot find a block
1584  * of the requested size, due to either severe fragmentation or the pool
1585  * being nearly full, it calls zio_write_gang_block() to construct the
1586  * block from smaller fragments.
1587  *
1588  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1589  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
1590  * an indirect block: it's an array of block pointers.  It consumes
1591  * only one sector and hence is allocatable regardless of fragmentation.
1592  * The gang header's bps point to its gang members, which hold the data.
1593  *
1594  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1595  * as the verifier to ensure uniqueness of the SHA256 checksum.
1596  * Critically, the gang block bp's blk_cksum is the checksum of the data,
1597  * not the gang header.  This ensures that data block signatures (needed for
1598  * deduplication) are independent of how the block is physically stored.
1599  *
1600  * Gang blocks can be nested: a gang member may itself be a gang block.
1601  * Thus every gang block is a tree in which root and all interior nodes are
1602  * gang headers, and the leaves are normal blocks that contain user data.
1603  * The root of the gang tree is called the gang leader.
1604  *
1605  * To perform any operation (read, rewrite, free, claim) on a gang block,
1606  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1607  * in the io_gang_tree field of the original logical i/o by recursively
1608  * reading the gang leader and all gang headers below it.  This yields
1609  * an in-core tree containing the contents of every gang header and the
1610  * bps for every constituent of the gang block.
1611  *
1612  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1613  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
1614  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1615  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1616  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1617  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
1618  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1619  * of the gang header plus zio_checksum_compute() of the data to update the
1620  * gang header's blk_cksum as described above.
1621  *
1622  * The two-phase assemble/issue model solves the problem of partial failure --
1623  * what if you'd freed part of a gang block but then couldn't read the
1624  * gang header for another part?  Assembling the entire gang tree first
1625  * ensures that all the necessary gang header I/O has succeeded before
1626  * starting the actual work of free, claim, or write.  Once the gang tree
1627  * is assembled, free and claim are in-memory operations that cannot fail.
1628  *
1629  * In the event that a gang write fails, zio_dva_unallocate() walks the
1630  * gang tree to immediately free (i.e. insert back into the space map)
1631  * everything we've allocated.  This ensures that we don't get ENOSPC
1632  * errors during repeated suspend/resume cycles due to a flaky device.
1633  *
1634  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
1635  * the gang tree, we won't modify the block, so we can safely defer the free
1636  * (knowing that the block is still intact).  If we *can* assemble the gang
1637  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1638  * each constituent bp and we can allocate a new block on the next sync pass.
1639  *
1640  * In all cases, the gang tree allows complete recovery from partial failure.
1641  * ==========================================================================
1642  */
1643 
1644 static zio_t *
1645 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1646 {
1647 	if (gn != NULL)
1648 		return (pio);
1649 
1650 	return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1651 	    NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1652 	    &pio->io_bookmark));
1653 }
1654 
1655 zio_t *
1656 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1657 {
1658 	zio_t *zio;
1659 
1660 	if (gn != NULL) {
1661 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1662 		    gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1663 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1664 		/*
1665 		 * As we rewrite each gang header, the pipeline will compute
1666 		 * a new gang block header checksum for it; but no one will
1667 		 * compute a new data checksum, so we do that here.  The one
1668 		 * exception is the gang leader: the pipeline already computed
1669 		 * its data checksum because that stage precedes gang assembly.
1670 		 * (Presently, nothing actually uses interior data checksums;
1671 		 * this is just good hygiene.)
1672 		 */
1673 		if (gn != pio->io_gang_leader->io_gang_tree) {
1674 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1675 			    data, BP_GET_PSIZE(bp));
1676 		}
1677 		/*
1678 		 * If we are here to damage data for testing purposes,
1679 		 * leave the GBH alone so that we can detect the damage.
1680 		 */
1681 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1682 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1683 	} else {
1684 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1685 		    data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1686 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1687 	}
1688 
1689 	return (zio);
1690 }
1691 
1692 /* ARGSUSED */
1693 zio_t *
1694 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1695 {
1696 	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1697 	    ZIO_GANG_CHILD_FLAGS(pio)));
1698 }
1699 
1700 /* ARGSUSED */
1701 zio_t *
1702 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1703 {
1704 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1705 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1706 }
1707 
1708 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1709 	NULL,
1710 	zio_read_gang,
1711 	zio_rewrite_gang,
1712 	zio_free_gang,
1713 	zio_claim_gang,
1714 	NULL
1715 };
1716 
1717 static void zio_gang_tree_assemble_done(zio_t *zio);
1718 
1719 static zio_gang_node_t *
1720 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1721 {
1722 	zio_gang_node_t *gn;
1723 
1724 	ASSERT(*gnpp == NULL);
1725 
1726 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1727 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1728 	*gnpp = gn;
1729 
1730 	return (gn);
1731 }
1732 
1733 static void
1734 zio_gang_node_free(zio_gang_node_t **gnpp)
1735 {
1736 	zio_gang_node_t *gn = *gnpp;
1737 
1738 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1739 		ASSERT(gn->gn_child[g] == NULL);
1740 
1741 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1742 	kmem_free(gn, sizeof (*gn));
1743 	*gnpp = NULL;
1744 }
1745 
1746 static void
1747 zio_gang_tree_free(zio_gang_node_t **gnpp)
1748 {
1749 	zio_gang_node_t *gn = *gnpp;
1750 
1751 	if (gn == NULL)
1752 		return;
1753 
1754 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1755 		zio_gang_tree_free(&gn->gn_child[g]);
1756 
1757 	zio_gang_node_free(gnpp);
1758 }
1759 
1760 static void
1761 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1762 {
1763 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1764 
1765 	ASSERT(gio->io_gang_leader == gio);
1766 	ASSERT(BP_IS_GANG(bp));
1767 
1768 	zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1769 	    SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1770 	    gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1771 }
1772 
1773 static void
1774 zio_gang_tree_assemble_done(zio_t *zio)
1775 {
1776 	zio_t *gio = zio->io_gang_leader;
1777 	zio_gang_node_t *gn = zio->io_private;
1778 	blkptr_t *bp = zio->io_bp;
1779 
1780 	ASSERT(gio == zio_unique_parent(zio));
1781 	ASSERT(zio->io_child_count == 0);
1782 
1783 	if (zio->io_error)
1784 		return;
1785 
1786 	if (BP_SHOULD_BYTESWAP(bp))
1787 		byteswap_uint64_array(zio->io_data, zio->io_size);
1788 
1789 	ASSERT(zio->io_data == gn->gn_gbh);
1790 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1791 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1792 
1793 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1794 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1795 		if (!BP_IS_GANG(gbp))
1796 			continue;
1797 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1798 	}
1799 }
1800 
1801 static void
1802 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1803 {
1804 	zio_t *gio = pio->io_gang_leader;
1805 	zio_t *zio;
1806 
1807 	ASSERT(BP_IS_GANG(bp) == !!gn);
1808 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1809 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1810 
1811 	/*
1812 	 * If you're a gang header, your data is in gn->gn_gbh.
1813 	 * If you're a gang member, your data is in 'data' and gn == NULL.
1814 	 */
1815 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1816 
1817 	if (gn != NULL) {
1818 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1819 
1820 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1821 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1822 			if (BP_IS_HOLE(gbp))
1823 				continue;
1824 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1825 			data = (char *)data + BP_GET_PSIZE(gbp);
1826 		}
1827 	}
1828 
1829 	if (gn == gio->io_gang_tree)
1830 		ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1831 
1832 	if (zio != pio)
1833 		zio_nowait(zio);
1834 }
1835 
1836 static int
1837 zio_gang_assemble(zio_t *zio)
1838 {
1839 	blkptr_t *bp = zio->io_bp;
1840 
1841 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1842 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1843 
1844 	zio->io_gang_leader = zio;
1845 
1846 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1847 
1848 	return (ZIO_PIPELINE_CONTINUE);
1849 }
1850 
1851 static int
1852 zio_gang_issue(zio_t *zio)
1853 {
1854 	blkptr_t *bp = zio->io_bp;
1855 
1856 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1857 		return (ZIO_PIPELINE_STOP);
1858 
1859 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1860 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1861 
1862 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1863 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1864 	else
1865 		zio_gang_tree_free(&zio->io_gang_tree);
1866 
1867 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1868 
1869 	return (ZIO_PIPELINE_CONTINUE);
1870 }
1871 
1872 static void
1873 zio_write_gang_member_ready(zio_t *zio)
1874 {
1875 	zio_t *pio = zio_unique_parent(zio);
1876 	zio_t *gio = zio->io_gang_leader;
1877 	dva_t *cdva = zio->io_bp->blk_dva;
1878 	dva_t *pdva = pio->io_bp->blk_dva;
1879 	uint64_t asize;
1880 
1881 	if (BP_IS_HOLE(zio->io_bp))
1882 		return;
1883 
1884 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1885 
1886 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1887 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1888 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1889 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1890 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1891 
1892 	mutex_enter(&pio->io_lock);
1893 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1894 		ASSERT(DVA_GET_GANG(&pdva[d]));
1895 		asize = DVA_GET_ASIZE(&pdva[d]);
1896 		asize += DVA_GET_ASIZE(&cdva[d]);
1897 		DVA_SET_ASIZE(&pdva[d], asize);
1898 	}
1899 	mutex_exit(&pio->io_lock);
1900 }
1901 
1902 static int
1903 zio_write_gang_block(zio_t *pio)
1904 {
1905 	spa_t *spa = pio->io_spa;
1906 	blkptr_t *bp = pio->io_bp;
1907 	zio_t *gio = pio->io_gang_leader;
1908 	zio_t *zio;
1909 	zio_gang_node_t *gn, **gnpp;
1910 	zio_gbh_phys_t *gbh;
1911 	uint64_t txg = pio->io_txg;
1912 	uint64_t resid = pio->io_size;
1913 	uint64_t lsize;
1914 	int copies = gio->io_prop.zp_copies;
1915 	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1916 	zio_prop_t zp;
1917 	int error;
1918 
1919 	error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1920 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1921 	    METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1922 	if (error) {
1923 		pio->io_error = error;
1924 		return (ZIO_PIPELINE_CONTINUE);
1925 	}
1926 
1927 	if (pio == gio) {
1928 		gnpp = &gio->io_gang_tree;
1929 	} else {
1930 		gnpp = pio->io_private;
1931 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
1932 	}
1933 
1934 	gn = zio_gang_node_alloc(gnpp);
1935 	gbh = gn->gn_gbh;
1936 	bzero(gbh, SPA_GANGBLOCKSIZE);
1937 
1938 	/*
1939 	 * Create the gang header.
1940 	 */
1941 	zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1942 	    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1943 
1944 	/*
1945 	 * Create and nowait the gang children.
1946 	 */
1947 	for (int g = 0; resid != 0; resid -= lsize, g++) {
1948 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1949 		    SPA_MINBLOCKSIZE);
1950 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1951 
1952 		zp.zp_checksum = gio->io_prop.zp_checksum;
1953 		zp.zp_compress = ZIO_COMPRESS_OFF;
1954 		zp.zp_type = DMU_OT_NONE;
1955 		zp.zp_level = 0;
1956 		zp.zp_copies = gio->io_prop.zp_copies;
1957 		zp.zp_dedup = B_FALSE;
1958 		zp.zp_dedup_verify = B_FALSE;
1959 		zp.zp_nopwrite = B_FALSE;
1960 
1961 		zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1962 		    (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1963 		    zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1964 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1965 		    &pio->io_bookmark));
1966 	}
1967 
1968 	/*
1969 	 * Set pio's pipeline to just wait for zio to finish.
1970 	 */
1971 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1972 
1973 	zio_nowait(zio);
1974 
1975 	return (ZIO_PIPELINE_CONTINUE);
1976 }
1977 
1978 /*
1979  * The zio_nop_write stage in the pipeline determines if allocating
1980  * a new bp is necessary.  By leveraging a cryptographically secure checksum,
1981  * such as SHA256, we can compare the checksums of the new data and the old
1982  * to determine if allocating a new block is required.  The nopwrite
1983  * feature can handle writes in either syncing or open context (i.e. zil
1984  * writes) and as a result is mutually exclusive with dedup.
1985  */
1986 static int
1987 zio_nop_write(zio_t *zio)
1988 {
1989 	blkptr_t *bp = zio->io_bp;
1990 	blkptr_t *bp_orig = &zio->io_bp_orig;
1991 	zio_prop_t *zp = &zio->io_prop;
1992 
1993 	ASSERT(BP_GET_LEVEL(bp) == 0);
1994 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1995 	ASSERT(zp->zp_nopwrite);
1996 	ASSERT(!zp->zp_dedup);
1997 	ASSERT(zio->io_bp_override == NULL);
1998 	ASSERT(IO_IS_ALLOCATING(zio));
1999 
2000 	/*
2001 	 * Check to see if the original bp and the new bp have matching
2002 	 * characteristics (i.e. same checksum, compression algorithms, etc).
2003 	 * If they don't then just continue with the pipeline which will
2004 	 * allocate a new bp.
2005 	 */
2006 	if (BP_IS_HOLE(bp_orig) ||
2007 	    !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
2008 	    BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2009 	    BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2010 	    BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2011 	    zp->zp_copies != BP_GET_NDVAS(bp_orig))
2012 		return (ZIO_PIPELINE_CONTINUE);
2013 
2014 	/*
2015 	 * If the checksums match then reset the pipeline so that we
2016 	 * avoid allocating a new bp and issuing any I/O.
2017 	 */
2018 	if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2019 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
2020 		ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2021 		ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2022 		ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2023 		ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2024 		    sizeof (uint64_t)) == 0);
2025 
2026 		*bp = *bp_orig;
2027 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2028 		zio->io_flags |= ZIO_FLAG_NOPWRITE;
2029 	}
2030 
2031 	return (ZIO_PIPELINE_CONTINUE);
2032 }
2033 
2034 /*
2035  * ==========================================================================
2036  * Dedup
2037  * ==========================================================================
2038  */
2039 static void
2040 zio_ddt_child_read_done(zio_t *zio)
2041 {
2042 	blkptr_t *bp = zio->io_bp;
2043 	ddt_entry_t *dde = zio->io_private;
2044 	ddt_phys_t *ddp;
2045 	zio_t *pio = zio_unique_parent(zio);
2046 
2047 	mutex_enter(&pio->io_lock);
2048 	ddp = ddt_phys_select(dde, bp);
2049 	if (zio->io_error == 0)
2050 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
2051 	if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2052 		dde->dde_repair_data = zio->io_data;
2053 	else
2054 		zio_buf_free(zio->io_data, zio->io_size);
2055 	mutex_exit(&pio->io_lock);
2056 }
2057 
2058 static int
2059 zio_ddt_read_start(zio_t *zio)
2060 {
2061 	blkptr_t *bp = zio->io_bp;
2062 
2063 	ASSERT(BP_GET_DEDUP(bp));
2064 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2065 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2066 
2067 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
2068 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
2069 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2070 		ddt_phys_t *ddp = dde->dde_phys;
2071 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2072 		blkptr_t blk;
2073 
2074 		ASSERT(zio->io_vsd == NULL);
2075 		zio->io_vsd = dde;
2076 
2077 		if (ddp_self == NULL)
2078 			return (ZIO_PIPELINE_CONTINUE);
2079 
2080 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2081 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2082 				continue;
2083 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2084 			    &blk);
2085 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
2086 			    zio_buf_alloc(zio->io_size), zio->io_size,
2087 			    zio_ddt_child_read_done, dde, zio->io_priority,
2088 			    ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2089 			    &zio->io_bookmark));
2090 		}
2091 		return (ZIO_PIPELINE_CONTINUE);
2092 	}
2093 
2094 	zio_nowait(zio_read(zio, zio->io_spa, bp,
2095 	    zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2096 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2097 
2098 	return (ZIO_PIPELINE_CONTINUE);
2099 }
2100 
2101 static int
2102 zio_ddt_read_done(zio_t *zio)
2103 {
2104 	blkptr_t *bp = zio->io_bp;
2105 
2106 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2107 		return (ZIO_PIPELINE_STOP);
2108 
2109 	ASSERT(BP_GET_DEDUP(bp));
2110 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2111 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2112 
2113 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
2114 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
2115 		ddt_entry_t *dde = zio->io_vsd;
2116 		if (ddt == NULL) {
2117 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2118 			return (ZIO_PIPELINE_CONTINUE);
2119 		}
2120 		if (dde == NULL) {
2121 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2122 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2123 			return (ZIO_PIPELINE_STOP);
2124 		}
2125 		if (dde->dde_repair_data != NULL) {
2126 			bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2127 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
2128 		}
2129 		ddt_repair_done(ddt, dde);
2130 		zio->io_vsd = NULL;
2131 	}
2132 
2133 	ASSERT(zio->io_vsd == NULL);
2134 
2135 	return (ZIO_PIPELINE_CONTINUE);
2136 }
2137 
2138 static boolean_t
2139 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2140 {
2141 	spa_t *spa = zio->io_spa;
2142 
2143 	/*
2144 	 * Note: we compare the original data, not the transformed data,
2145 	 * because when zio->io_bp is an override bp, we will not have
2146 	 * pushed the I/O transforms.  That's an important optimization
2147 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2148 	 */
2149 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2150 		zio_t *lio = dde->dde_lead_zio[p];
2151 
2152 		if (lio != NULL) {
2153 			return (lio->io_orig_size != zio->io_orig_size ||
2154 			    bcmp(zio->io_orig_data, lio->io_orig_data,
2155 			    zio->io_orig_size) != 0);
2156 		}
2157 	}
2158 
2159 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2160 		ddt_phys_t *ddp = &dde->dde_phys[p];
2161 
2162 		if (ddp->ddp_phys_birth != 0) {
2163 			arc_buf_t *abuf = NULL;
2164 			arc_flags_t aflags = ARC_FLAG_WAIT;
2165 			blkptr_t blk = *zio->io_bp;
2166 			int error;
2167 
2168 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2169 
2170 			ddt_exit(ddt);
2171 
2172 			error = arc_read(NULL, spa, &blk,
2173 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2174 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2175 			    &aflags, &zio->io_bookmark);
2176 
2177 			if (error == 0) {
2178 				if (arc_buf_size(abuf) != zio->io_orig_size ||
2179 				    bcmp(abuf->b_data, zio->io_orig_data,
2180 				    zio->io_orig_size) != 0)
2181 					error = SET_ERROR(EEXIST);
2182 				VERIFY(arc_buf_remove_ref(abuf, &abuf));
2183 			}
2184 
2185 			ddt_enter(ddt);
2186 			return (error != 0);
2187 		}
2188 	}
2189 
2190 	return (B_FALSE);
2191 }
2192 
2193 static void
2194 zio_ddt_child_write_ready(zio_t *zio)
2195 {
2196 	int p = zio->io_prop.zp_copies;
2197 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2198 	ddt_entry_t *dde = zio->io_private;
2199 	ddt_phys_t *ddp = &dde->dde_phys[p];
2200 	zio_t *pio;
2201 
2202 	if (zio->io_error)
2203 		return;
2204 
2205 	ddt_enter(ddt);
2206 
2207 	ASSERT(dde->dde_lead_zio[p] == zio);
2208 
2209 	ddt_phys_fill(ddp, zio->io_bp);
2210 
2211 	while ((pio = zio_walk_parents(zio)) != NULL)
2212 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2213 
2214 	ddt_exit(ddt);
2215 }
2216 
2217 static void
2218 zio_ddt_child_write_done(zio_t *zio)
2219 {
2220 	int p = zio->io_prop.zp_copies;
2221 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2222 	ddt_entry_t *dde = zio->io_private;
2223 	ddt_phys_t *ddp = &dde->dde_phys[p];
2224 
2225 	ddt_enter(ddt);
2226 
2227 	ASSERT(ddp->ddp_refcnt == 0);
2228 	ASSERT(dde->dde_lead_zio[p] == zio);
2229 	dde->dde_lead_zio[p] = NULL;
2230 
2231 	if (zio->io_error == 0) {
2232 		while (zio_walk_parents(zio) != NULL)
2233 			ddt_phys_addref(ddp);
2234 	} else {
2235 		ddt_phys_clear(ddp);
2236 	}
2237 
2238 	ddt_exit(ddt);
2239 }
2240 
2241 static void
2242 zio_ddt_ditto_write_done(zio_t *zio)
2243 {
2244 	int p = DDT_PHYS_DITTO;
2245 	zio_prop_t *zp = &zio->io_prop;
2246 	blkptr_t *bp = zio->io_bp;
2247 	ddt_t *ddt = ddt_select(zio->io_spa, bp);
2248 	ddt_entry_t *dde = zio->io_private;
2249 	ddt_phys_t *ddp = &dde->dde_phys[p];
2250 	ddt_key_t *ddk = &dde->dde_key;
2251 
2252 	ddt_enter(ddt);
2253 
2254 	ASSERT(ddp->ddp_refcnt == 0);
2255 	ASSERT(dde->dde_lead_zio[p] == zio);
2256 	dde->dde_lead_zio[p] = NULL;
2257 
2258 	if (zio->io_error == 0) {
2259 		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2260 		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2261 		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2262 		if (ddp->ddp_phys_birth != 0)
2263 			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2264 		ddt_phys_fill(ddp, bp);
2265 	}
2266 
2267 	ddt_exit(ddt);
2268 }
2269 
2270 static int
2271 zio_ddt_write(zio_t *zio)
2272 {
2273 	spa_t *spa = zio->io_spa;
2274 	blkptr_t *bp = zio->io_bp;
2275 	uint64_t txg = zio->io_txg;
2276 	zio_prop_t *zp = &zio->io_prop;
2277 	int p = zp->zp_copies;
2278 	int ditto_copies;
2279 	zio_t *cio = NULL;
2280 	zio_t *dio = NULL;
2281 	ddt_t *ddt = ddt_select(spa, bp);
2282 	ddt_entry_t *dde;
2283 	ddt_phys_t *ddp;
2284 
2285 	ASSERT(BP_GET_DEDUP(bp));
2286 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2287 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2288 
2289 	ddt_enter(ddt);
2290 	dde = ddt_lookup(ddt, bp, B_TRUE);
2291 	ddp = &dde->dde_phys[p];
2292 
2293 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2294 		/*
2295 		 * If we're using a weak checksum, upgrade to a strong checksum
2296 		 * and try again.  If we're already using a strong checksum,
2297 		 * we can't resolve it, so just convert to an ordinary write.
2298 		 * (And automatically e-mail a paper to Nature?)
2299 		 */
2300 		if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2301 			zp->zp_checksum = spa_dedup_checksum(spa);
2302 			zio_pop_transforms(zio);
2303 			zio->io_stage = ZIO_STAGE_OPEN;
2304 			BP_ZERO(bp);
2305 		} else {
2306 			zp->zp_dedup = B_FALSE;
2307 		}
2308 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2309 		ddt_exit(ddt);
2310 		return (ZIO_PIPELINE_CONTINUE);
2311 	}
2312 
2313 	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2314 	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2315 
2316 	if (ditto_copies > ddt_ditto_copies_present(dde) &&
2317 	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2318 		zio_prop_t czp = *zp;
2319 
2320 		czp.zp_copies = ditto_copies;
2321 
2322 		/*
2323 		 * If we arrived here with an override bp, we won't have run
2324 		 * the transform stack, so we won't have the data we need to
2325 		 * generate a child i/o.  So, toss the override bp and restart.
2326 		 * This is safe, because using the override bp is just an
2327 		 * optimization; and it's rare, so the cost doesn't matter.
2328 		 */
2329 		if (zio->io_bp_override) {
2330 			zio_pop_transforms(zio);
2331 			zio->io_stage = ZIO_STAGE_OPEN;
2332 			zio->io_pipeline = ZIO_WRITE_PIPELINE;
2333 			zio->io_bp_override = NULL;
2334 			BP_ZERO(bp);
2335 			ddt_exit(ddt);
2336 			return (ZIO_PIPELINE_CONTINUE);
2337 		}
2338 
2339 		dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2340 		    zio->io_orig_size, &czp, NULL, NULL,
2341 		    zio_ddt_ditto_write_done, dde, zio->io_priority,
2342 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2343 
2344 		zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2345 		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2346 	}
2347 
2348 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2349 		if (ddp->ddp_phys_birth != 0)
2350 			ddt_bp_fill(ddp, bp, txg);
2351 		if (dde->dde_lead_zio[p] != NULL)
2352 			zio_add_child(zio, dde->dde_lead_zio[p]);
2353 		else
2354 			ddt_phys_addref(ddp);
2355 	} else if (zio->io_bp_override) {
2356 		ASSERT(bp->blk_birth == txg);
2357 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2358 		ddt_phys_fill(ddp, bp);
2359 		ddt_phys_addref(ddp);
2360 	} else {
2361 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2362 		    zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2363 		    zio_ddt_child_write_done, dde, zio->io_priority,
2364 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2365 
2366 		zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2367 		dde->dde_lead_zio[p] = cio;
2368 	}
2369 
2370 	ddt_exit(ddt);
2371 
2372 	if (cio)
2373 		zio_nowait(cio);
2374 	if (dio)
2375 		zio_nowait(dio);
2376 
2377 	return (ZIO_PIPELINE_CONTINUE);
2378 }
2379 
2380 ddt_entry_t *freedde; /* for debugging */
2381 
2382 static int
2383 zio_ddt_free(zio_t *zio)
2384 {
2385 	spa_t *spa = zio->io_spa;
2386 	blkptr_t *bp = zio->io_bp;
2387 	ddt_t *ddt = ddt_select(spa, bp);
2388 	ddt_entry_t *dde;
2389 	ddt_phys_t *ddp;
2390 
2391 	ASSERT(BP_GET_DEDUP(bp));
2392 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2393 
2394 	ddt_enter(ddt);
2395 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2396 	ddp = ddt_phys_select(dde, bp);
2397 	ddt_phys_decref(ddp);
2398 	ddt_exit(ddt);
2399 
2400 	return (ZIO_PIPELINE_CONTINUE);
2401 }
2402 
2403 /*
2404  * ==========================================================================
2405  * Allocate and free blocks
2406  * ==========================================================================
2407  */
2408 static int
2409 zio_dva_allocate(zio_t *zio)
2410 {
2411 	spa_t *spa = zio->io_spa;
2412 	metaslab_class_t *mc = spa_normal_class(spa);
2413 	blkptr_t *bp = zio->io_bp;
2414 	int error;
2415 	int flags = 0;
2416 
2417 	if (zio->io_gang_leader == NULL) {
2418 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2419 		zio->io_gang_leader = zio;
2420 	}
2421 
2422 	ASSERT(BP_IS_HOLE(bp));
2423 	ASSERT0(BP_GET_NDVAS(bp));
2424 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
2425 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2426 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2427 
2428 	/*
2429 	 * The dump device does not support gang blocks so allocation on
2430 	 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2431 	 * the "fast" gang feature.
2432 	 */
2433 	flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2434 	flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2435 	    METASLAB_GANG_CHILD : 0;
2436 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
2437 	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2438 
2439 	if (error) {
2440 		spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2441 		    "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2442 		    error);
2443 		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2444 			return (zio_write_gang_block(zio));
2445 		zio->io_error = error;
2446 	}
2447 
2448 	return (ZIO_PIPELINE_CONTINUE);
2449 }
2450 
2451 static int
2452 zio_dva_free(zio_t *zio)
2453 {
2454 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2455 
2456 	return (ZIO_PIPELINE_CONTINUE);
2457 }
2458 
2459 static int
2460 zio_dva_claim(zio_t *zio)
2461 {
2462 	int error;
2463 
2464 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2465 	if (error)
2466 		zio->io_error = error;
2467 
2468 	return (ZIO_PIPELINE_CONTINUE);
2469 }
2470 
2471 /*
2472  * Undo an allocation.  This is used by zio_done() when an I/O fails
2473  * and we want to give back the block we just allocated.
2474  * This handles both normal blocks and gang blocks.
2475  */
2476 static void
2477 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2478 {
2479 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2480 	ASSERT(zio->io_bp_override == NULL);
2481 
2482 	if (!BP_IS_HOLE(bp))
2483 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2484 
2485 	if (gn != NULL) {
2486 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2487 			zio_dva_unallocate(zio, gn->gn_child[g],
2488 			    &gn->gn_gbh->zg_blkptr[g]);
2489 		}
2490 	}
2491 }
2492 
2493 /*
2494  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
2495  */
2496 int
2497 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2498     uint64_t size, boolean_t use_slog)
2499 {
2500 	int error = 1;
2501 
2502 	ASSERT(txg > spa_syncing_txg(spa));
2503 
2504 	/*
2505 	 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2506 	 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2507 	 * when allocating them.
2508 	 */
2509 	if (use_slog) {
2510 		error = metaslab_alloc(spa, spa_log_class(spa), size,
2511 		    new_bp, 1, txg, old_bp,
2512 		    METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2513 	}
2514 
2515 	if (error) {
2516 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
2517 		    new_bp, 1, txg, old_bp,
2518 		    METASLAB_HINTBP_AVOID);
2519 	}
2520 
2521 	if (error == 0) {
2522 		BP_SET_LSIZE(new_bp, size);
2523 		BP_SET_PSIZE(new_bp, size);
2524 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2525 		BP_SET_CHECKSUM(new_bp,
2526 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2527 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2528 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2529 		BP_SET_LEVEL(new_bp, 0);
2530 		BP_SET_DEDUP(new_bp, 0);
2531 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2532 	}
2533 
2534 	return (error);
2535 }
2536 
2537 /*
2538  * Free an intent log block.
2539  */
2540 void
2541 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2542 {
2543 	ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2544 	ASSERT(!BP_IS_GANG(bp));
2545 
2546 	zio_free(spa, txg, bp);
2547 }
2548 
2549 /*
2550  * ==========================================================================
2551  * Read and write to physical devices
2552  * ==========================================================================
2553  */
2554 
2555 
2556 /*
2557  * Issue an I/O to the underlying vdev. Typically the issue pipeline
2558  * stops after this stage and will resume upon I/O completion.
2559  * However, there are instances where the vdev layer may need to
2560  * continue the pipeline when an I/O was not issued. Since the I/O
2561  * that was sent to the vdev layer might be different than the one
2562  * currently active in the pipeline (see vdev_queue_io()), we explicitly
2563  * force the underlying vdev layers to call either zio_execute() or
2564  * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
2565  */
2566 static int
2567 zio_vdev_io_start(zio_t *zio)
2568 {
2569 	vdev_t *vd = zio->io_vd;
2570 	uint64_t align;
2571 	spa_t *spa = zio->io_spa;
2572 
2573 	ASSERT(zio->io_error == 0);
2574 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2575 
2576 	if (vd == NULL) {
2577 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2578 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2579 
2580 		/*
2581 		 * The mirror_ops handle multiple DVAs in a single BP.
2582 		 */
2583 		vdev_mirror_ops.vdev_op_io_start(zio);
2584 		return (ZIO_PIPELINE_STOP);
2585 	}
2586 
2587 	/*
2588 	 * We keep track of time-sensitive I/Os so that the scan thread
2589 	 * can quickly react to certain workloads.  In particular, we care
2590 	 * about non-scrubbing, top-level reads and writes with the following
2591 	 * characteristics:
2592 	 *	- synchronous writes of user data to non-slog devices
2593 	 *	- any reads of user data
2594 	 * When these conditions are met, adjust the timestamp of spa_last_io
2595 	 * which allows the scan thread to adjust its workload accordingly.
2596 	 */
2597 	if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2598 	    vd == vd->vdev_top && !vd->vdev_islog &&
2599 	    zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2600 	    zio->io_txg != spa_syncing_txg(spa)) {
2601 		uint64_t old = spa->spa_last_io;
2602 		uint64_t new = ddi_get_lbolt64();
2603 		if (old != new)
2604 			(void) atomic_cas_64(&spa->spa_last_io, old, new);
2605 	}
2606 
2607 	align = 1ULL << vd->vdev_top->vdev_ashift;
2608 
2609 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
2610 	    P2PHASE(zio->io_size, align) != 0) {
2611 		/* Transform logical writes to be a full physical block size. */
2612 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
2613 		char *abuf = zio_buf_alloc(asize);
2614 		ASSERT(vd == vd->vdev_top);
2615 		if (zio->io_type == ZIO_TYPE_WRITE) {
2616 			bcopy(zio->io_data, abuf, zio->io_size);
2617 			bzero(abuf + zio->io_size, asize - zio->io_size);
2618 		}
2619 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2620 	}
2621 
2622 	/*
2623 	 * If this is not a physical io, make sure that it is properly aligned
2624 	 * before proceeding.
2625 	 */
2626 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2627 		ASSERT0(P2PHASE(zio->io_offset, align));
2628 		ASSERT0(P2PHASE(zio->io_size, align));
2629 	} else {
2630 		/*
2631 		 * For physical writes, we allow 512b aligned writes and assume
2632 		 * the device will perform a read-modify-write as necessary.
2633 		 */
2634 		ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2635 		ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2636 	}
2637 
2638 	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2639 
2640 	/*
2641 	 * If this is a repair I/O, and there's no self-healing involved --
2642 	 * that is, we're just resilvering what we expect to resilver --
2643 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2644 	 * This prevents spurious resilvering with nested replication.
2645 	 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2646 	 * A is out of date, we'll read from C+D, then use the data to
2647 	 * resilver A+B -- but we don't actually want to resilver B, just A.
2648 	 * The top-level mirror has no way to know this, so instead we just
2649 	 * discard unnecessary repairs as we work our way down the vdev tree.
2650 	 * The same logic applies to any form of nested replication:
2651 	 * ditto + mirror, RAID-Z + replacing, etc.  This covers them all.
2652 	 */
2653 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2654 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2655 	    zio->io_txg != 0 &&	/* not a delegated i/o */
2656 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2657 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2658 		zio_vdev_io_bypass(zio);
2659 		return (ZIO_PIPELINE_CONTINUE);
2660 	}
2661 
2662 	if (vd->vdev_ops->vdev_op_leaf &&
2663 	    (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2664 
2665 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
2666 			return (ZIO_PIPELINE_CONTINUE);
2667 
2668 		if ((zio = vdev_queue_io(zio)) == NULL)
2669 			return (ZIO_PIPELINE_STOP);
2670 
2671 		if (!vdev_accessible(vd, zio)) {
2672 			zio->io_error = SET_ERROR(ENXIO);
2673 			zio_interrupt(zio);
2674 			return (ZIO_PIPELINE_STOP);
2675 		}
2676 	}
2677 
2678 	vd->vdev_ops->vdev_op_io_start(zio);
2679 	return (ZIO_PIPELINE_STOP);
2680 }
2681 
2682 static int
2683 zio_vdev_io_done(zio_t *zio)
2684 {
2685 	vdev_t *vd = zio->io_vd;
2686 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2687 	boolean_t unexpected_error = B_FALSE;
2688 
2689 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2690 		return (ZIO_PIPELINE_STOP);
2691 
2692 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2693 
2694 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2695 
2696 		vdev_queue_io_done(zio);
2697 
2698 		if (zio->io_type == ZIO_TYPE_WRITE)
2699 			vdev_cache_write(zio);
2700 
2701 		if (zio_injection_enabled && zio->io_error == 0)
2702 			zio->io_error = zio_handle_device_injection(vd,
2703 			    zio, EIO);
2704 
2705 		if (zio_injection_enabled && zio->io_error == 0)
2706 			zio->io_error = zio_handle_label_injection(zio, EIO);
2707 
2708 		if (zio->io_error) {
2709 			if (!vdev_accessible(vd, zio)) {
2710 				zio->io_error = SET_ERROR(ENXIO);
2711 			} else {
2712 				unexpected_error = B_TRUE;
2713 			}
2714 		}
2715 	}
2716 
2717 	ops->vdev_op_io_done(zio);
2718 
2719 	if (unexpected_error)
2720 		VERIFY(vdev_probe(vd, zio) == NULL);
2721 
2722 	return (ZIO_PIPELINE_CONTINUE);
2723 }
2724 
2725 /*
2726  * For non-raidz ZIOs, we can just copy aside the bad data read from the
2727  * disk, and use that to finish the checksum ereport later.
2728  */
2729 static void
2730 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2731     const void *good_buf)
2732 {
2733 	/* no processing needed */
2734 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2735 }
2736 
2737 /*ARGSUSED*/
2738 void
2739 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2740 {
2741 	void *buf = zio_buf_alloc(zio->io_size);
2742 
2743 	bcopy(zio->io_data, buf, zio->io_size);
2744 
2745 	zcr->zcr_cbinfo = zio->io_size;
2746 	zcr->zcr_cbdata = buf;
2747 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
2748 	zcr->zcr_free = zio_buf_free;
2749 }
2750 
2751 static int
2752 zio_vdev_io_assess(zio_t *zio)
2753 {
2754 	vdev_t *vd = zio->io_vd;
2755 
2756 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2757 		return (ZIO_PIPELINE_STOP);
2758 
2759 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2760 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2761 
2762 	if (zio->io_vsd != NULL) {
2763 		zio->io_vsd_ops->vsd_free(zio);
2764 		zio->io_vsd = NULL;
2765 	}
2766 
2767 	if (zio_injection_enabled && zio->io_error == 0)
2768 		zio->io_error = zio_handle_fault_injection(zio, EIO);
2769 
2770 	/*
2771 	 * If the I/O failed, determine whether we should attempt to retry it.
2772 	 *
2773 	 * On retry, we cut in line in the issue queue, since we don't want
2774 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2775 	 */
2776 	if (zio->io_error && vd == NULL &&
2777 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2778 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
2779 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
2780 		zio->io_error = 0;
2781 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
2782 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2783 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2784 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2785 		    zio_requeue_io_start_cut_in_line);
2786 		return (ZIO_PIPELINE_STOP);
2787 	}
2788 
2789 	/*
2790 	 * If we got an error on a leaf device, convert it to ENXIO
2791 	 * if the device is not accessible at all.
2792 	 */
2793 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2794 	    !vdev_accessible(vd, zio))
2795 		zio->io_error = SET_ERROR(ENXIO);
2796 
2797 	/*
2798 	 * If we can't write to an interior vdev (mirror or RAID-Z),
2799 	 * set vdev_cant_write so that we stop trying to allocate from it.
2800 	 */
2801 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2802 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2803 		vd->vdev_cant_write = B_TRUE;
2804 	}
2805 
2806 	if (zio->io_error)
2807 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2808 
2809 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2810 	    zio->io_physdone != NULL) {
2811 		ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2812 		ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2813 		zio->io_physdone(zio->io_logical);
2814 	}
2815 
2816 	return (ZIO_PIPELINE_CONTINUE);
2817 }
2818 
2819 void
2820 zio_vdev_io_reissue(zio_t *zio)
2821 {
2822 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2823 	ASSERT(zio->io_error == 0);
2824 
2825 	zio->io_stage >>= 1;
2826 }
2827 
2828 void
2829 zio_vdev_io_redone(zio_t *zio)
2830 {
2831 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2832 
2833 	zio->io_stage >>= 1;
2834 }
2835 
2836 void
2837 zio_vdev_io_bypass(zio_t *zio)
2838 {
2839 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2840 	ASSERT(zio->io_error == 0);
2841 
2842 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2843 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2844 }
2845 
2846 /*
2847  * ==========================================================================
2848  * Generate and verify checksums
2849  * ==========================================================================
2850  */
2851 static int
2852 zio_checksum_generate(zio_t *zio)
2853 {
2854 	blkptr_t *bp = zio->io_bp;
2855 	enum zio_checksum checksum;
2856 
2857 	if (bp == NULL) {
2858 		/*
2859 		 * This is zio_write_phys().
2860 		 * We're either generating a label checksum, or none at all.
2861 		 */
2862 		checksum = zio->io_prop.zp_checksum;
2863 
2864 		if (checksum == ZIO_CHECKSUM_OFF)
2865 			return (ZIO_PIPELINE_CONTINUE);
2866 
2867 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2868 	} else {
2869 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2870 			ASSERT(!IO_IS_ALLOCATING(zio));
2871 			checksum = ZIO_CHECKSUM_GANG_HEADER;
2872 		} else {
2873 			checksum = BP_GET_CHECKSUM(bp);
2874 		}
2875 	}
2876 
2877 	zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2878 
2879 	return (ZIO_PIPELINE_CONTINUE);
2880 }
2881 
2882 static int
2883 zio_checksum_verify(zio_t *zio)
2884 {
2885 	zio_bad_cksum_t info;
2886 	blkptr_t *bp = zio->io_bp;
2887 	int error;
2888 
2889 	ASSERT(zio->io_vd != NULL);
2890 
2891 	if (bp == NULL) {
2892 		/*
2893 		 * This is zio_read_phys().
2894 		 * We're either verifying a label checksum, or nothing at all.
2895 		 */
2896 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2897 			return (ZIO_PIPELINE_CONTINUE);
2898 
2899 		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2900 	}
2901 
2902 	if ((error = zio_checksum_error(zio, &info)) != 0) {
2903 		zio->io_error = error;
2904 		if (error == ECKSUM &&
2905 		    !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2906 			zfs_ereport_start_checksum(zio->io_spa,
2907 			    zio->io_vd, zio, zio->io_offset,
2908 			    zio->io_size, NULL, &info);
2909 		}
2910 	}
2911 
2912 	return (ZIO_PIPELINE_CONTINUE);
2913 }
2914 
2915 /*
2916  * Called by RAID-Z to ensure we don't compute the checksum twice.
2917  */
2918 void
2919 zio_checksum_verified(zio_t *zio)
2920 {
2921 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2922 }
2923 
2924 /*
2925  * ==========================================================================
2926  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2927  * An error of 0 indicates success.  ENXIO indicates whole-device failure,
2928  * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
2929  * indicate errors that are specific to one I/O, and most likely permanent.
2930  * Any other error is presumed to be worse because we weren't expecting it.
2931  * ==========================================================================
2932  */
2933 int
2934 zio_worst_error(int e1, int e2)
2935 {
2936 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2937 	int r1, r2;
2938 
2939 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2940 		if (e1 == zio_error_rank[r1])
2941 			break;
2942 
2943 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2944 		if (e2 == zio_error_rank[r2])
2945 			break;
2946 
2947 	return (r1 > r2 ? e1 : e2);
2948 }
2949 
2950 /*
2951  * ==========================================================================
2952  * I/O completion
2953  * ==========================================================================
2954  */
2955 static int
2956 zio_ready(zio_t *zio)
2957 {
2958 	blkptr_t *bp = zio->io_bp;
2959 	zio_t *pio, *pio_next;
2960 
2961 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2962 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2963 		return (ZIO_PIPELINE_STOP);
2964 
2965 	if (zio->io_ready) {
2966 		ASSERT(IO_IS_ALLOCATING(zio));
2967 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2968 		    (zio->io_flags & ZIO_FLAG_NOPWRITE));
2969 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2970 
2971 		zio->io_ready(zio);
2972 	}
2973 
2974 	if (bp != NULL && bp != &zio->io_bp_copy)
2975 		zio->io_bp_copy = *bp;
2976 
2977 	if (zio->io_error)
2978 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2979 
2980 	mutex_enter(&zio->io_lock);
2981 	zio->io_state[ZIO_WAIT_READY] = 1;
2982 	pio = zio_walk_parents(zio);
2983 	mutex_exit(&zio->io_lock);
2984 
2985 	/*
2986 	 * As we notify zio's parents, new parents could be added.
2987 	 * New parents go to the head of zio's io_parent_list, however,
2988 	 * so we will (correctly) not notify them.  The remainder of zio's
2989 	 * io_parent_list, from 'pio_next' onward, cannot change because
2990 	 * all parents must wait for us to be done before they can be done.
2991 	 */
2992 	for (; pio != NULL; pio = pio_next) {
2993 		pio_next = zio_walk_parents(zio);
2994 		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2995 	}
2996 
2997 	if (zio->io_flags & ZIO_FLAG_NODATA) {
2998 		if (BP_IS_GANG(bp)) {
2999 			zio->io_flags &= ~ZIO_FLAG_NODATA;
3000 		} else {
3001 			ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3002 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3003 		}
3004 	}
3005 
3006 	if (zio_injection_enabled &&
3007 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
3008 		zio_handle_ignored_writes(zio);
3009 
3010 	return (ZIO_PIPELINE_CONTINUE);
3011 }
3012 
3013 static int
3014 zio_done(zio_t *zio)
3015 {
3016 	spa_t *spa = zio->io_spa;
3017 	zio_t *lio = zio->io_logical;
3018 	blkptr_t *bp = zio->io_bp;
3019 	vdev_t *vd = zio->io_vd;
3020 	uint64_t psize = zio->io_size;
3021 	zio_t *pio, *pio_next;
3022 
3023 	/*
3024 	 * If our children haven't all completed,
3025 	 * wait for them and then repeat this pipeline stage.
3026 	 */
3027 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3028 	    zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3029 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3030 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3031 		return (ZIO_PIPELINE_STOP);
3032 
3033 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3034 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3035 			ASSERT(zio->io_children[c][w] == 0);
3036 
3037 	if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3038 		ASSERT(bp->blk_pad[0] == 0);
3039 		ASSERT(bp->blk_pad[1] == 0);
3040 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3041 		    (bp == zio_unique_parent(zio)->io_bp));
3042 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3043 		    zio->io_bp_override == NULL &&
3044 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3045 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
3046 			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3047 			ASSERT(BP_COUNT_GANG(bp) == 0 ||
3048 			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3049 		}
3050 		if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3051 			VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3052 	}
3053 
3054 	/*
3055 	 * If there were child vdev/gang/ddt errors, they apply to us now.
3056 	 */
3057 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3058 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3059 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3060 
3061 	/*
3062 	 * If the I/O on the transformed data was successful, generate any
3063 	 * checksum reports now while we still have the transformed data.
3064 	 */
3065 	if (zio->io_error == 0) {
3066 		while (zio->io_cksum_report != NULL) {
3067 			zio_cksum_report_t *zcr = zio->io_cksum_report;
3068 			uint64_t align = zcr->zcr_align;
3069 			uint64_t asize = P2ROUNDUP(psize, align);
3070 			char *abuf = zio->io_data;
3071 
3072 			if (asize != psize) {
3073 				abuf = zio_buf_alloc(asize);
3074 				bcopy(zio->io_data, abuf, psize);
3075 				bzero(abuf + psize, asize - psize);
3076 			}
3077 
3078 			zio->io_cksum_report = zcr->zcr_next;
3079 			zcr->zcr_next = NULL;
3080 			zcr->zcr_finish(zcr, abuf);
3081 			zfs_ereport_free_checksum(zcr);
3082 
3083 			if (asize != psize)
3084 				zio_buf_free(abuf, asize);
3085 		}
3086 	}
3087 
3088 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
3089 
3090 	vdev_stat_update(zio, psize);
3091 
3092 	if (zio->io_error) {
3093 		/*
3094 		 * If this I/O is attached to a particular vdev,
3095 		 * generate an error message describing the I/O failure
3096 		 * at the block level.  We ignore these errors if the
3097 		 * device is currently unavailable.
3098 		 */
3099 		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3100 			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3101 
3102 		if ((zio->io_error == EIO || !(zio->io_flags &
3103 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3104 		    zio == lio) {
3105 			/*
3106 			 * For logical I/O requests, tell the SPA to log the
3107 			 * error and generate a logical data ereport.
3108 			 */
3109 			spa_log_error(spa, zio);
3110 			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3111 			    0, 0);
3112 		}
3113 	}
3114 
3115 	if (zio->io_error && zio == lio) {
3116 		/*
3117 		 * Determine whether zio should be reexecuted.  This will
3118 		 * propagate all the way to the root via zio_notify_parent().
3119 		 */
3120 		ASSERT(vd == NULL && bp != NULL);
3121 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3122 
3123 		if (IO_IS_ALLOCATING(zio) &&
3124 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3125 			if (zio->io_error != ENOSPC)
3126 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3127 			else
3128 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3129 		}
3130 
3131 		if ((zio->io_type == ZIO_TYPE_READ ||
3132 		    zio->io_type == ZIO_TYPE_FREE) &&
3133 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3134 		    zio->io_error == ENXIO &&
3135 		    spa_load_state(spa) == SPA_LOAD_NONE &&
3136 		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3137 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3138 
3139 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3140 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3141 
3142 		/*
3143 		 * Here is a possibly good place to attempt to do
3144 		 * either combinatorial reconstruction or error correction
3145 		 * based on checksums.  It also might be a good place
3146 		 * to send out preliminary ereports before we suspend
3147 		 * processing.
3148 		 */
3149 	}
3150 
3151 	/*
3152 	 * If there were logical child errors, they apply to us now.
3153 	 * We defer this until now to avoid conflating logical child
3154 	 * errors with errors that happened to the zio itself when
3155 	 * updating vdev stats and reporting FMA events above.
3156 	 */
3157 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3158 
3159 	if ((zio->io_error || zio->io_reexecute) &&
3160 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3161 	    !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3162 		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3163 
3164 	zio_gang_tree_free(&zio->io_gang_tree);
3165 
3166 	/*
3167 	 * Godfather I/Os should never suspend.
3168 	 */
3169 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3170 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3171 		zio->io_reexecute = 0;
3172 
3173 	if (zio->io_reexecute) {
3174 		/*
3175 		 * This is a logical I/O that wants to reexecute.
3176 		 *
3177 		 * Reexecute is top-down.  When an i/o fails, if it's not
3178 		 * the root, it simply notifies its parent and sticks around.
3179 		 * The parent, seeing that it still has children in zio_done(),
3180 		 * does the same.  This percolates all the way up to the root.
3181 		 * The root i/o will reexecute or suspend the entire tree.
3182 		 *
3183 		 * This approach ensures that zio_reexecute() honors
3184 		 * all the original i/o dependency relationships, e.g.
3185 		 * parents not executing until children are ready.
3186 		 */
3187 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3188 
3189 		zio->io_gang_leader = NULL;
3190 
3191 		mutex_enter(&zio->io_lock);
3192 		zio->io_state[ZIO_WAIT_DONE] = 1;
3193 		mutex_exit(&zio->io_lock);
3194 
3195 		/*
3196 		 * "The Godfather" I/O monitors its children but is
3197 		 * not a true parent to them. It will track them through
3198 		 * the pipeline but severs its ties whenever they get into
3199 		 * trouble (e.g. suspended). This allows "The Godfather"
3200 		 * I/O to return status without blocking.
3201 		 */
3202 		for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3203 			zio_link_t *zl = zio->io_walk_link;
3204 			pio_next = zio_walk_parents(zio);
3205 
3206 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3207 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3208 				zio_remove_child(pio, zio, zl);
3209 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3210 			}
3211 		}
3212 
3213 		if ((pio = zio_unique_parent(zio)) != NULL) {
3214 			/*
3215 			 * We're not a root i/o, so there's nothing to do
3216 			 * but notify our parent.  Don't propagate errors
3217 			 * upward since we haven't permanently failed yet.
3218 			 */
3219 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3220 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3221 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3222 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3223 			/*
3224 			 * We'd fail again if we reexecuted now, so suspend
3225 			 * until conditions improve (e.g. device comes online).
3226 			 */
3227 			zio_suspend(spa, zio);
3228 		} else {
3229 			/*
3230 			 * Reexecution is potentially a huge amount of work.
3231 			 * Hand it off to the otherwise-unused claim taskq.
3232 			 */
3233 			ASSERT(zio->io_tqent.tqent_next == NULL);
3234 			spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3235 			    ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3236 			    0, &zio->io_tqent);
3237 		}
3238 		return (ZIO_PIPELINE_STOP);
3239 	}
3240 
3241 	ASSERT(zio->io_child_count == 0);
3242 	ASSERT(zio->io_reexecute == 0);
3243 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3244 
3245 	/*
3246 	 * Report any checksum errors, since the I/O is complete.
3247 	 */
3248 	while (zio->io_cksum_report != NULL) {
3249 		zio_cksum_report_t *zcr = zio->io_cksum_report;
3250 		zio->io_cksum_report = zcr->zcr_next;
3251 		zcr->zcr_next = NULL;
3252 		zcr->zcr_finish(zcr, NULL);
3253 		zfs_ereport_free_checksum(zcr);
3254 	}
3255 
3256 	/*
3257 	 * It is the responsibility of the done callback to ensure that this
3258 	 * particular zio is no longer discoverable for adoption, and as
3259 	 * such, cannot acquire any new parents.
3260 	 */
3261 	if (zio->io_done)
3262 		zio->io_done(zio);
3263 
3264 	mutex_enter(&zio->io_lock);
3265 	zio->io_state[ZIO_WAIT_DONE] = 1;
3266 	mutex_exit(&zio->io_lock);
3267 
3268 	for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3269 		zio_link_t *zl = zio->io_walk_link;
3270 		pio_next = zio_walk_parents(zio);
3271 		zio_remove_child(pio, zio, zl);
3272 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3273 	}
3274 
3275 	if (zio->io_waiter != NULL) {
3276 		mutex_enter(&zio->io_lock);
3277 		zio->io_executor = NULL;
3278 		cv_broadcast(&zio->io_cv);
3279 		mutex_exit(&zio->io_lock);
3280 	} else {
3281 		zio_destroy(zio);
3282 	}
3283 
3284 	return (ZIO_PIPELINE_STOP);
3285 }
3286 
3287 /*
3288  * ==========================================================================
3289  * I/O pipeline definition
3290  * ==========================================================================
3291  */
3292 static zio_pipe_stage_t *zio_pipeline[] = {
3293 	NULL,
3294 	zio_read_bp_init,
3295 	zio_free_bp_init,
3296 	zio_issue_async,
3297 	zio_write_bp_init,
3298 	zio_checksum_generate,
3299 	zio_nop_write,
3300 	zio_ddt_read_start,
3301 	zio_ddt_read_done,
3302 	zio_ddt_write,
3303 	zio_ddt_free,
3304 	zio_gang_assemble,
3305 	zio_gang_issue,
3306 	zio_dva_allocate,
3307 	zio_dva_free,
3308 	zio_dva_claim,
3309 	zio_ready,
3310 	zio_vdev_io_start,
3311 	zio_vdev_io_done,
3312 	zio_vdev_io_assess,
3313 	zio_checksum_verify,
3314 	zio_done
3315 };
3316 
3317 /* dnp is the dnode for zb1->zb_object */
3318 boolean_t
3319 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1,
3320     const zbookmark_phys_t *zb2)
3321 {
3322 	uint64_t zb1nextL0, zb2thisobj;
3323 
3324 	ASSERT(zb1->zb_objset == zb2->zb_objset);
3325 	ASSERT(zb2->zb_level == 0);
3326 
3327 	/* The objset_phys_t isn't before anything. */
3328 	if (dnp == NULL)
3329 		return (B_FALSE);
3330 
3331 	zb1nextL0 = (zb1->zb_blkid + 1) <<
3332 	    ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3333 
3334 	zb2thisobj = zb2->zb_object ? zb2->zb_object :
3335 	    zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3336 
3337 	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3338 		uint64_t nextobj = zb1nextL0 *
3339 		    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3340 		return (nextobj <= zb2thisobj);
3341 	}
3342 
3343 	if (zb1->zb_object < zb2thisobj)
3344 		return (B_TRUE);
3345 	if (zb1->zb_object > zb2thisobj)
3346 		return (B_FALSE);
3347 	if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3348 		return (B_FALSE);
3349 	return (zb1nextL0 <= zb2->zb_blkid);
3350 }
3351