xref: /titanic_44/usr/src/uts/common/fs/zfs/zio.c (revision fd435bccec40cb841f712413108d8577422fd9ab)
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 #define	BP_SPANB(indblkshift, level) \
71 	(((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
72 #define	COMPARE_META_LEVEL	0x80000000ul
73 /*
74  * The following actions directly effect the spa's sync-to-convergence logic.
75  * The values below define the sync pass when we start performing the action.
76  * Care should be taken when changing these values as they directly impact
77  * spa_sync() performance. Tuning these values may introduce subtle performance
78  * pathologies and should only be done in the context of performance analysis.
79  * These tunables will eventually be removed and replaced with #defines once
80  * enough analysis has been done to determine optimal values.
81  *
82  * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
83  * regular blocks are not deferred.
84  */
85 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
86 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
87 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
88 
89 /*
90  * An allocating zio is one that either currently has the DVA allocate
91  * stage set or will have it later in its lifetime.
92  */
93 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
94 
95 boolean_t	zio_requeue_io_start_cut_in_line = B_TRUE;
96 
97 #ifdef ZFS_DEBUG
98 int zio_buf_debug_limit = 16384;
99 #else
100 int zio_buf_debug_limit = 0;
101 #endif
102 
103 void
104 zio_init(void)
105 {
106 	size_t c;
107 	vmem_t *data_alloc_arena = NULL;
108 
109 #ifdef _KERNEL
110 	data_alloc_arena = zio_alloc_arena;
111 #endif
112 	zio_cache = kmem_cache_create("zio_cache",
113 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
114 	zio_link_cache = kmem_cache_create("zio_link_cache",
115 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
116 
117 	/*
118 	 * For small buffers, we want a cache for each multiple of
119 	 * SPA_MINBLOCKSIZE.  For larger buffers, we want a cache
120 	 * for each quarter-power of 2.
121 	 */
122 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
123 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
124 		size_t p2 = size;
125 		size_t align = 0;
126 		size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
127 
128 		while (!ISP2(p2))
129 			p2 &= p2 - 1;
130 
131 #ifndef _KERNEL
132 		/*
133 		 * If we are using watchpoints, put each buffer on its own page,
134 		 * to eliminate the performance overhead of trapping to the
135 		 * kernel when modifying a non-watched buffer that shares the
136 		 * page with a watched buffer.
137 		 */
138 		if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
139 			continue;
140 #endif
141 		if (size <= 4 * SPA_MINBLOCKSIZE) {
142 			align = SPA_MINBLOCKSIZE;
143 		} else if (IS_P2ALIGNED(size, p2 >> 2)) {
144 			align = MIN(p2 >> 2, PAGESIZE);
145 		}
146 
147 		if (align != 0) {
148 			char name[36];
149 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
150 			zio_buf_cache[c] = kmem_cache_create(name, size,
151 			    align, NULL, NULL, NULL, NULL, NULL, cflags);
152 
153 			/*
154 			 * Since zio_data bufs do not appear in crash dumps, we
155 			 * pass KMC_NOTOUCH so that no allocator metadata is
156 			 * stored with the buffers.
157 			 */
158 			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
159 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
160 			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
161 			    cflags | KMC_NOTOUCH);
162 		}
163 	}
164 
165 	while (--c != 0) {
166 		ASSERT(zio_buf_cache[c] != NULL);
167 		if (zio_buf_cache[c - 1] == NULL)
168 			zio_buf_cache[c - 1] = zio_buf_cache[c];
169 
170 		ASSERT(zio_data_buf_cache[c] != NULL);
171 		if (zio_data_buf_cache[c - 1] == NULL)
172 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
173 	}
174 
175 	zio_inject_init();
176 }
177 
178 void
179 zio_fini(void)
180 {
181 	size_t c;
182 	kmem_cache_t *last_cache = NULL;
183 	kmem_cache_t *last_data_cache = NULL;
184 
185 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
186 		if (zio_buf_cache[c] != last_cache) {
187 			last_cache = zio_buf_cache[c];
188 			kmem_cache_destroy(zio_buf_cache[c]);
189 		}
190 		zio_buf_cache[c] = NULL;
191 
192 		if (zio_data_buf_cache[c] != last_data_cache) {
193 			last_data_cache = zio_data_buf_cache[c];
194 			kmem_cache_destroy(zio_data_buf_cache[c]);
195 		}
196 		zio_data_buf_cache[c] = NULL;
197 	}
198 
199 	kmem_cache_destroy(zio_link_cache);
200 	kmem_cache_destroy(zio_cache);
201 
202 	zio_inject_fini();
203 }
204 
205 /*
206  * ==========================================================================
207  * Allocate and free I/O buffers
208  * ==========================================================================
209  */
210 
211 /*
212  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
213  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
214  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
215  * excess / transient data in-core during a crashdump.
216  */
217 void *
218 zio_buf_alloc(size_t size)
219 {
220 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
221 
222 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
223 
224 	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
225 }
226 
227 /*
228  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
229  * crashdump if the kernel panics.  This exists so that we will limit the amount
230  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
231  * of kernel heap dumped to disk when the kernel panics)
232  */
233 void *
234 zio_data_buf_alloc(size_t size)
235 {
236 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
237 
238 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
239 
240 	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
241 }
242 
243 void
244 zio_buf_free(void *buf, size_t size)
245 {
246 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
247 
248 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
249 
250 	kmem_cache_free(zio_buf_cache[c], buf);
251 }
252 
253 void
254 zio_data_buf_free(void *buf, size_t size)
255 {
256 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
257 
258 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
259 
260 	kmem_cache_free(zio_data_buf_cache[c], buf);
261 }
262 
263 /*
264  * ==========================================================================
265  * Push and pop I/O transform buffers
266  * ==========================================================================
267  */
268 static void
269 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
270 	zio_transform_func_t *transform)
271 {
272 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
273 
274 	zt->zt_orig_data = zio->io_data;
275 	zt->zt_orig_size = zio->io_size;
276 	zt->zt_bufsize = bufsize;
277 	zt->zt_transform = transform;
278 
279 	zt->zt_next = zio->io_transform_stack;
280 	zio->io_transform_stack = zt;
281 
282 	zio->io_data = data;
283 	zio->io_size = size;
284 }
285 
286 static void
287 zio_pop_transforms(zio_t *zio)
288 {
289 	zio_transform_t *zt;
290 
291 	while ((zt = zio->io_transform_stack) != NULL) {
292 		if (zt->zt_transform != NULL)
293 			zt->zt_transform(zio,
294 			    zt->zt_orig_data, zt->zt_orig_size);
295 
296 		if (zt->zt_bufsize != 0)
297 			zio_buf_free(zio->io_data, zt->zt_bufsize);
298 
299 		zio->io_data = zt->zt_orig_data;
300 		zio->io_size = zt->zt_orig_size;
301 		zio->io_transform_stack = zt->zt_next;
302 
303 		kmem_free(zt, sizeof (zio_transform_t));
304 	}
305 }
306 
307 /*
308  * ==========================================================================
309  * I/O transform callbacks for subblocks and decompression
310  * ==========================================================================
311  */
312 static void
313 zio_subblock(zio_t *zio, void *data, uint64_t size)
314 {
315 	ASSERT(zio->io_size > size);
316 
317 	if (zio->io_type == ZIO_TYPE_READ)
318 		bcopy(zio->io_data, data, size);
319 }
320 
321 static void
322 zio_decompress(zio_t *zio, void *data, uint64_t size)
323 {
324 	if (zio->io_error == 0 &&
325 	    zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
326 	    zio->io_data, data, zio->io_size, size) != 0)
327 		zio->io_error = SET_ERROR(EIO);
328 }
329 
330 /*
331  * ==========================================================================
332  * I/O parent/child relationships and pipeline interlocks
333  * ==========================================================================
334  */
335 /*
336  * NOTE - Callers to zio_walk_parents() and zio_walk_children must
337  *        continue calling these functions until they return NULL.
338  *        Otherwise, the next caller will pick up the list walk in
339  *        some indeterminate state.  (Otherwise every caller would
340  *        have to pass in a cookie to keep the state represented by
341  *        io_walk_link, which gets annoying.)
342  */
343 zio_t *
344 zio_walk_parents(zio_t *cio)
345 {
346 	zio_link_t *zl = cio->io_walk_link;
347 	list_t *pl = &cio->io_parent_list;
348 
349 	zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
350 	cio->io_walk_link = zl;
351 
352 	if (zl == NULL)
353 		return (NULL);
354 
355 	ASSERT(zl->zl_child == cio);
356 	return (zl->zl_parent);
357 }
358 
359 zio_t *
360 zio_walk_children(zio_t *pio)
361 {
362 	zio_link_t *zl = pio->io_walk_link;
363 	list_t *cl = &pio->io_child_list;
364 
365 	zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
366 	pio->io_walk_link = zl;
367 
368 	if (zl == NULL)
369 		return (NULL);
370 
371 	ASSERT(zl->zl_parent == pio);
372 	return (zl->zl_child);
373 }
374 
375 zio_t *
376 zio_unique_parent(zio_t *cio)
377 {
378 	zio_t *pio = zio_walk_parents(cio);
379 
380 	VERIFY(zio_walk_parents(cio) == NULL);
381 	return (pio);
382 }
383 
384 void
385 zio_add_child(zio_t *pio, zio_t *cio)
386 {
387 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
388 
389 	/*
390 	 * Logical I/Os can have logical, gang, or vdev children.
391 	 * Gang I/Os can have gang or vdev children.
392 	 * Vdev I/Os can only have vdev children.
393 	 * The following ASSERT captures all of these constraints.
394 	 */
395 	ASSERT(cio->io_child_type <= pio->io_child_type);
396 
397 	zl->zl_parent = pio;
398 	zl->zl_child = cio;
399 
400 	mutex_enter(&cio->io_lock);
401 	mutex_enter(&pio->io_lock);
402 
403 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
404 
405 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
406 		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
407 
408 	list_insert_head(&pio->io_child_list, zl);
409 	list_insert_head(&cio->io_parent_list, zl);
410 
411 	pio->io_child_count++;
412 	cio->io_parent_count++;
413 
414 	mutex_exit(&pio->io_lock);
415 	mutex_exit(&cio->io_lock);
416 }
417 
418 static void
419 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
420 {
421 	ASSERT(zl->zl_parent == pio);
422 	ASSERT(zl->zl_child == cio);
423 
424 	mutex_enter(&cio->io_lock);
425 	mutex_enter(&pio->io_lock);
426 
427 	list_remove(&pio->io_child_list, zl);
428 	list_remove(&cio->io_parent_list, zl);
429 
430 	pio->io_child_count--;
431 	cio->io_parent_count--;
432 
433 	mutex_exit(&pio->io_lock);
434 	mutex_exit(&cio->io_lock);
435 
436 	kmem_cache_free(zio_link_cache, zl);
437 }
438 
439 static boolean_t
440 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
441 {
442 	uint64_t *countp = &zio->io_children[child][wait];
443 	boolean_t waiting = B_FALSE;
444 
445 	mutex_enter(&zio->io_lock);
446 	ASSERT(zio->io_stall == NULL);
447 	if (*countp != 0) {
448 		zio->io_stage >>= 1;
449 		zio->io_stall = countp;
450 		waiting = B_TRUE;
451 	}
452 	mutex_exit(&zio->io_lock);
453 
454 	return (waiting);
455 }
456 
457 static void
458 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
459 {
460 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
461 	int *errorp = &pio->io_child_error[zio->io_child_type];
462 
463 	mutex_enter(&pio->io_lock);
464 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
465 		*errorp = zio_worst_error(*errorp, zio->io_error);
466 	pio->io_reexecute |= zio->io_reexecute;
467 	ASSERT3U(*countp, >, 0);
468 
469 	(*countp)--;
470 
471 	if (*countp == 0 && pio->io_stall == countp) {
472 		pio->io_stall = NULL;
473 		mutex_exit(&pio->io_lock);
474 		zio_execute(pio);
475 	} else {
476 		mutex_exit(&pio->io_lock);
477 	}
478 }
479 
480 static void
481 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
482 {
483 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
484 		zio->io_error = zio->io_child_error[c];
485 }
486 
487 /*
488  * ==========================================================================
489  * Create the various types of I/O (read, write, free, etc)
490  * ==========================================================================
491  */
492 static zio_t *
493 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
494     void *data, uint64_t size, zio_done_func_t *done, void *private,
495     zio_type_t type, zio_priority_t priority, enum zio_flag flags,
496     vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
497     enum zio_stage stage, enum zio_stage pipeline)
498 {
499 	zio_t *zio;
500 
501 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
502 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
503 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
504 
505 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
506 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
507 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
508 
509 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
510 	bzero(zio, sizeof (zio_t));
511 
512 	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
513 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
514 
515 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
516 	    offsetof(zio_link_t, zl_parent_node));
517 	list_create(&zio->io_child_list, sizeof (zio_link_t),
518 	    offsetof(zio_link_t, zl_child_node));
519 
520 	if (vd != NULL)
521 		zio->io_child_type = ZIO_CHILD_VDEV;
522 	else if (flags & ZIO_FLAG_GANG_CHILD)
523 		zio->io_child_type = ZIO_CHILD_GANG;
524 	else if (flags & ZIO_FLAG_DDT_CHILD)
525 		zio->io_child_type = ZIO_CHILD_DDT;
526 	else
527 		zio->io_child_type = ZIO_CHILD_LOGICAL;
528 
529 	if (bp != NULL) {
530 		zio->io_bp = (blkptr_t *)bp;
531 		zio->io_bp_copy = *bp;
532 		zio->io_bp_orig = *bp;
533 		if (type != ZIO_TYPE_WRITE ||
534 		    zio->io_child_type == ZIO_CHILD_DDT)
535 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
536 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
537 			zio->io_logical = zio;
538 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
539 			pipeline |= ZIO_GANG_STAGES;
540 	}
541 
542 	zio->io_spa = spa;
543 	zio->io_txg = txg;
544 	zio->io_done = done;
545 	zio->io_private = private;
546 	zio->io_type = type;
547 	zio->io_priority = priority;
548 	zio->io_vd = vd;
549 	zio->io_offset = offset;
550 	zio->io_orig_data = zio->io_data = data;
551 	zio->io_orig_size = zio->io_size = size;
552 	zio->io_orig_flags = zio->io_flags = flags;
553 	zio->io_orig_stage = zio->io_stage = stage;
554 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
555 
556 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
557 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
558 
559 	if (zb != NULL)
560 		zio->io_bookmark = *zb;
561 
562 	if (pio != NULL) {
563 		if (zio->io_logical == NULL)
564 			zio->io_logical = pio->io_logical;
565 		if (zio->io_child_type == ZIO_CHILD_GANG)
566 			zio->io_gang_leader = pio->io_gang_leader;
567 		zio_add_child(pio, zio);
568 	}
569 
570 	return (zio);
571 }
572 
573 static void
574 zio_destroy(zio_t *zio)
575 {
576 	list_destroy(&zio->io_parent_list);
577 	list_destroy(&zio->io_child_list);
578 	mutex_destroy(&zio->io_lock);
579 	cv_destroy(&zio->io_cv);
580 	kmem_cache_free(zio_cache, zio);
581 }
582 
583 zio_t *
584 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
585     void *private, enum zio_flag flags)
586 {
587 	zio_t *zio;
588 
589 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
590 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
591 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
592 
593 	return (zio);
594 }
595 
596 zio_t *
597 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
598 {
599 	return (zio_null(NULL, spa, NULL, done, private, flags));
600 }
601 
602 void
603 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
604 {
605 	if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
606 		zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
607 		    bp, (longlong_t)BP_GET_TYPE(bp));
608 	}
609 	if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
610 	    BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
611 		zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
612 		    bp, (longlong_t)BP_GET_CHECKSUM(bp));
613 	}
614 	if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
615 	    BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
616 		zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
617 		    bp, (longlong_t)BP_GET_COMPRESS(bp));
618 	}
619 	if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
620 		zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
621 		    bp, (longlong_t)BP_GET_LSIZE(bp));
622 	}
623 	if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
624 		zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
625 		    bp, (longlong_t)BP_GET_PSIZE(bp));
626 	}
627 
628 	if (BP_IS_EMBEDDED(bp)) {
629 		if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
630 			zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
631 			    bp, (longlong_t)BPE_GET_ETYPE(bp));
632 		}
633 	}
634 
635 	/*
636 	 * Pool-specific checks.
637 	 *
638 	 * Note: it would be nice to verify that the blk_birth and
639 	 * BP_PHYSICAL_BIRTH() are not too large.  However, spa_freeze()
640 	 * allows the birth time of log blocks (and dmu_sync()-ed blocks
641 	 * that are in the log) to be arbitrarily large.
642 	 */
643 	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
644 		uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
645 		if (vdevid >= spa->spa_root_vdev->vdev_children) {
646 			zfs_panic_recover("blkptr at %p DVA %u has invalid "
647 			    "VDEV %llu",
648 			    bp, i, (longlong_t)vdevid);
649 			continue;
650 		}
651 		vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
652 		if (vd == NULL) {
653 			zfs_panic_recover("blkptr at %p DVA %u has invalid "
654 			    "VDEV %llu",
655 			    bp, i, (longlong_t)vdevid);
656 			continue;
657 		}
658 		if (vd->vdev_ops == &vdev_hole_ops) {
659 			zfs_panic_recover("blkptr at %p DVA %u has hole "
660 			    "VDEV %llu",
661 			    bp, i, (longlong_t)vdevid);
662 			continue;
663 		}
664 		if (vd->vdev_ops == &vdev_missing_ops) {
665 			/*
666 			 * "missing" vdevs are valid during import, but we
667 			 * don't have their detailed info (e.g. asize), so
668 			 * we can't perform any more checks on them.
669 			 */
670 			continue;
671 		}
672 		uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
673 		uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
674 		if (BP_IS_GANG(bp))
675 			asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
676 		if (offset + asize > vd->vdev_asize) {
677 			zfs_panic_recover("blkptr at %p DVA %u has invalid "
678 			    "OFFSET %llu",
679 			    bp, i, (longlong_t)offset);
680 		}
681 	}
682 }
683 
684 zio_t *
685 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
686     void *data, uint64_t size, zio_done_func_t *done, void *private,
687     zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
688 {
689 	zio_t *zio;
690 
691 	zfs_blkptr_verify(spa, bp);
692 
693 	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
694 	    data, size, done, private,
695 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
696 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
697 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
698 
699 	return (zio);
700 }
701 
702 zio_t *
703 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
704     void *data, uint64_t size, const zio_prop_t *zp,
705     zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
706     void *private,
707     zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
708 {
709 	zio_t *zio;
710 
711 	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
712 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
713 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
714 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
715 	    DMU_OT_IS_VALID(zp->zp_type) &&
716 	    zp->zp_level < 32 &&
717 	    zp->zp_copies > 0 &&
718 	    zp->zp_copies <= spa_max_replication(spa));
719 
720 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
721 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
722 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
723 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
724 
725 	zio->io_ready = ready;
726 	zio->io_physdone = physdone;
727 	zio->io_prop = *zp;
728 
729 	/*
730 	 * Data can be NULL if we are going to call zio_write_override() to
731 	 * provide the already-allocated BP.  But we may need the data to
732 	 * verify a dedup hit (if requested).  In this case, don't try to
733 	 * dedup (just take the already-allocated BP verbatim).
734 	 */
735 	if (data == NULL && zio->io_prop.zp_dedup_verify) {
736 		zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
737 	}
738 
739 	return (zio);
740 }
741 
742 zio_t *
743 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
744     uint64_t size, zio_done_func_t *done, void *private,
745     zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
746 {
747 	zio_t *zio;
748 
749 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
750 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
751 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
752 
753 	return (zio);
754 }
755 
756 void
757 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
758 {
759 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
760 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
761 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
762 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
763 
764 	/*
765 	 * We must reset the io_prop to match the values that existed
766 	 * when the bp was first written by dmu_sync() keeping in mind
767 	 * that nopwrite and dedup are mutually exclusive.
768 	 */
769 	zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
770 	zio->io_prop.zp_nopwrite = nopwrite;
771 	zio->io_prop.zp_copies = copies;
772 	zio->io_bp_override = bp;
773 }
774 
775 void
776 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
777 {
778 
779 	/*
780 	 * The check for EMBEDDED is a performance optimization.  We
781 	 * process the free here (by ignoring it) rather than
782 	 * putting it on the list and then processing it in zio_free_sync().
783 	 */
784 	if (BP_IS_EMBEDDED(bp))
785 		return;
786 	metaslab_check_free(spa, bp);
787 
788 	/*
789 	 * Frees that are for the currently-syncing txg, are not going to be
790 	 * deferred, and which will not need to do a read (i.e. not GANG or
791 	 * DEDUP), can be processed immediately.  Otherwise, put them on the
792 	 * in-memory list for later processing.
793 	 */
794 	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
795 	    txg != spa->spa_syncing_txg ||
796 	    spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
797 		bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
798 	} else {
799 		VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
800 	}
801 }
802 
803 zio_t *
804 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
805     enum zio_flag flags)
806 {
807 	zio_t *zio;
808 	enum zio_stage stage = ZIO_FREE_PIPELINE;
809 
810 	ASSERT(!BP_IS_HOLE(bp));
811 	ASSERT(spa_syncing_txg(spa) == txg);
812 	ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
813 
814 	if (BP_IS_EMBEDDED(bp))
815 		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
816 
817 	metaslab_check_free(spa, bp);
818 	arc_freed(spa, bp);
819 
820 	/*
821 	 * GANG and DEDUP blocks can induce a read (for the gang block header,
822 	 * or the DDT), so issue them asynchronously so that this thread is
823 	 * not tied up.
824 	 */
825 	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
826 		stage |= ZIO_STAGE_ISSUE_ASYNC;
827 
828 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
829 	    NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
830 	    NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
831 
832 	return (zio);
833 }
834 
835 zio_t *
836 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
837     zio_done_func_t *done, void *private, enum zio_flag flags)
838 {
839 	zio_t *zio;
840 
841 	dprintf_bp(bp, "claiming in txg %llu", txg);
842 
843 	if (BP_IS_EMBEDDED(bp))
844 		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
845 
846 	/*
847 	 * A claim is an allocation of a specific block.  Claims are needed
848 	 * to support immediate writes in the intent log.  The issue is that
849 	 * immediate writes contain committed data, but in a txg that was
850 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
851 	 * the intent log claims all blocks that contain immediate write data
852 	 * so that the SPA knows they're in use.
853 	 *
854 	 * All claims *must* be resolved in the first txg -- before the SPA
855 	 * starts allocating blocks -- so that nothing is allocated twice.
856 	 * If txg == 0 we just verify that the block is claimable.
857 	 */
858 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
859 	ASSERT(txg == spa_first_txg(spa) || txg == 0);
860 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
861 
862 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
863 	    done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
864 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
865 
866 	return (zio);
867 }
868 
869 zio_t *
870 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
871     zio_done_func_t *done, void *private, enum zio_flag flags)
872 {
873 	zio_t *zio;
874 	int c;
875 
876 	if (vd->vdev_children == 0) {
877 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
878 		    ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
879 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
880 
881 		zio->io_cmd = cmd;
882 	} else {
883 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
884 
885 		for (c = 0; c < vd->vdev_children; c++)
886 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
887 			    done, private, flags));
888 	}
889 
890 	return (zio);
891 }
892 
893 zio_t *
894 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
895     void *data, int checksum, zio_done_func_t *done, void *private,
896     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
897 {
898 	zio_t *zio;
899 
900 	ASSERT(vd->vdev_children == 0);
901 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
902 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
903 	ASSERT3U(offset + size, <=, vd->vdev_psize);
904 
905 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
906 	    ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
907 	    NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
908 
909 	zio->io_prop.zp_checksum = checksum;
910 
911 	return (zio);
912 }
913 
914 zio_t *
915 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
916     void *data, int checksum, zio_done_func_t *done, void *private,
917     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
918 {
919 	zio_t *zio;
920 
921 	ASSERT(vd->vdev_children == 0);
922 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
923 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
924 	ASSERT3U(offset + size, <=, vd->vdev_psize);
925 
926 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
927 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
928 	    NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
929 
930 	zio->io_prop.zp_checksum = checksum;
931 
932 	if (zio_checksum_table[checksum].ci_eck) {
933 		/*
934 		 * zec checksums are necessarily destructive -- they modify
935 		 * the end of the write buffer to hold the verifier/checksum.
936 		 * Therefore, we must make a local copy in case the data is
937 		 * being written to multiple places in parallel.
938 		 */
939 		void *wbuf = zio_buf_alloc(size);
940 		bcopy(data, wbuf, size);
941 		zio_push_transform(zio, wbuf, size, size, NULL);
942 	}
943 
944 	return (zio);
945 }
946 
947 /*
948  * Create a child I/O to do some work for us.
949  */
950 zio_t *
951 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
952 	void *data, uint64_t size, int type, zio_priority_t priority,
953 	enum zio_flag flags, zio_done_func_t *done, void *private)
954 {
955 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
956 	zio_t *zio;
957 
958 	ASSERT(vd->vdev_parent ==
959 	    (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
960 
961 	if (type == ZIO_TYPE_READ && bp != NULL) {
962 		/*
963 		 * If we have the bp, then the child should perform the
964 		 * checksum and the parent need not.  This pushes error
965 		 * detection as close to the leaves as possible and
966 		 * eliminates redundant checksums in the interior nodes.
967 		 */
968 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
969 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
970 	}
971 
972 	if (vd->vdev_children == 0)
973 		offset += VDEV_LABEL_START_SIZE;
974 
975 	flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
976 
977 	/*
978 	 * If we've decided to do a repair, the write is not speculative --
979 	 * even if the original read was.
980 	 */
981 	if (flags & ZIO_FLAG_IO_REPAIR)
982 		flags &= ~ZIO_FLAG_SPECULATIVE;
983 
984 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
985 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
986 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
987 
988 	zio->io_physdone = pio->io_physdone;
989 	if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
990 		zio->io_logical->io_phys_children++;
991 
992 	return (zio);
993 }
994 
995 zio_t *
996 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
997 	int type, zio_priority_t priority, enum zio_flag flags,
998 	zio_done_func_t *done, void *private)
999 {
1000 	zio_t *zio;
1001 
1002 	ASSERT(vd->vdev_ops->vdev_op_leaf);
1003 
1004 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1005 	    data, size, done, private, type, priority,
1006 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1007 	    vd, offset, NULL,
1008 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1009 
1010 	return (zio);
1011 }
1012 
1013 void
1014 zio_flush(zio_t *zio, vdev_t *vd)
1015 {
1016 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1017 	    NULL, NULL,
1018 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1019 }
1020 
1021 void
1022 zio_shrink(zio_t *zio, uint64_t size)
1023 {
1024 	ASSERT(zio->io_executor == NULL);
1025 	ASSERT(zio->io_orig_size == zio->io_size);
1026 	ASSERT(size <= zio->io_size);
1027 
1028 	/*
1029 	 * We don't shrink for raidz because of problems with the
1030 	 * reconstruction when reading back less than the block size.
1031 	 * Note, BP_IS_RAIDZ() assumes no compression.
1032 	 */
1033 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1034 	if (!BP_IS_RAIDZ(zio->io_bp))
1035 		zio->io_orig_size = zio->io_size = size;
1036 }
1037 
1038 /*
1039  * ==========================================================================
1040  * Prepare to read and write logical blocks
1041  * ==========================================================================
1042  */
1043 
1044 static int
1045 zio_read_bp_init(zio_t *zio)
1046 {
1047 	blkptr_t *bp = zio->io_bp;
1048 
1049 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1050 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
1051 	    !(zio->io_flags & ZIO_FLAG_RAW)) {
1052 		uint64_t psize =
1053 		    BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1054 		void *cbuf = zio_buf_alloc(psize);
1055 
1056 		zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1057 	}
1058 
1059 	if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1060 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1061 		decode_embedded_bp_compressed(bp, zio->io_data);
1062 	} else {
1063 		ASSERT(!BP_IS_EMBEDDED(bp));
1064 	}
1065 
1066 	if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1067 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1068 
1069 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1070 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1071 
1072 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1073 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1074 
1075 	return (ZIO_PIPELINE_CONTINUE);
1076 }
1077 
1078 static int
1079 zio_write_bp_init(zio_t *zio)
1080 {
1081 	spa_t *spa = zio->io_spa;
1082 	zio_prop_t *zp = &zio->io_prop;
1083 	enum zio_compress compress = zp->zp_compress;
1084 	blkptr_t *bp = zio->io_bp;
1085 	uint64_t lsize = zio->io_size;
1086 	uint64_t psize = lsize;
1087 	int pass = 1;
1088 
1089 	/*
1090 	 * If our children haven't all reached the ready stage,
1091 	 * wait for them and then repeat this pipeline stage.
1092 	 */
1093 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1094 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1095 		return (ZIO_PIPELINE_STOP);
1096 
1097 	if (!IO_IS_ALLOCATING(zio))
1098 		return (ZIO_PIPELINE_CONTINUE);
1099 
1100 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1101 
1102 	if (zio->io_bp_override) {
1103 		ASSERT(bp->blk_birth != zio->io_txg);
1104 		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1105 
1106 		*bp = *zio->io_bp_override;
1107 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1108 
1109 		if (BP_IS_EMBEDDED(bp))
1110 			return (ZIO_PIPELINE_CONTINUE);
1111 
1112 		/*
1113 		 * If we've been overridden and nopwrite is set then
1114 		 * set the flag accordingly to indicate that a nopwrite
1115 		 * has already occurred.
1116 		 */
1117 		if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1118 			ASSERT(!zp->zp_dedup);
1119 			zio->io_flags |= ZIO_FLAG_NOPWRITE;
1120 			return (ZIO_PIPELINE_CONTINUE);
1121 		}
1122 
1123 		ASSERT(!zp->zp_nopwrite);
1124 
1125 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1126 			return (ZIO_PIPELINE_CONTINUE);
1127 
1128 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1129 		    zp->zp_dedup_verify);
1130 
1131 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1132 			BP_SET_DEDUP(bp, 1);
1133 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1134 			return (ZIO_PIPELINE_CONTINUE);
1135 		}
1136 	}
1137 
1138 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1139 		/*
1140 		 * We're rewriting an existing block, which means we're
1141 		 * working on behalf of spa_sync().  For spa_sync() to
1142 		 * converge, it must eventually be the case that we don't
1143 		 * have to allocate new blocks.  But compression changes
1144 		 * the blocksize, which forces a reallocate, and makes
1145 		 * convergence take longer.  Therefore, after the first
1146 		 * few passes, stop compressing to ensure convergence.
1147 		 */
1148 		pass = spa_sync_pass(spa);
1149 
1150 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
1151 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1152 		ASSERT(!BP_GET_DEDUP(bp));
1153 
1154 		if (pass >= zfs_sync_pass_dont_compress)
1155 			compress = ZIO_COMPRESS_OFF;
1156 
1157 		/* Make sure someone doesn't change their mind on overwrites */
1158 		ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1159 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1160 	}
1161 
1162 	if (compress != ZIO_COMPRESS_OFF) {
1163 		void *cbuf = zio_buf_alloc(lsize);
1164 		psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1165 		if (psize == 0 || psize == lsize) {
1166 			compress = ZIO_COMPRESS_OFF;
1167 			zio_buf_free(cbuf, lsize);
1168 		} else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1169 		    zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1170 		    spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1171 			encode_embedded_bp_compressed(bp,
1172 			    cbuf, compress, lsize, psize);
1173 			BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1174 			BP_SET_TYPE(bp, zio->io_prop.zp_type);
1175 			BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1176 			zio_buf_free(cbuf, lsize);
1177 			bp->blk_birth = zio->io_txg;
1178 			zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1179 			ASSERT(spa_feature_is_active(spa,
1180 			    SPA_FEATURE_EMBEDDED_DATA));
1181 			return (ZIO_PIPELINE_CONTINUE);
1182 		} else {
1183 			/*
1184 			 * Round up compressed size up to the ashift
1185 			 * of the smallest-ashift device, and zero the tail.
1186 			 * This ensures that the compressed size of the BP
1187 			 * (and thus compressratio property) are correct,
1188 			 * in that we charge for the padding used to fill out
1189 			 * the last sector.
1190 			 */
1191 			ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1192 			size_t rounded = (size_t)P2ROUNDUP(psize,
1193 			    1ULL << spa->spa_min_ashift);
1194 			if (rounded >= lsize) {
1195 				compress = ZIO_COMPRESS_OFF;
1196 				zio_buf_free(cbuf, lsize);
1197 				psize = lsize;
1198 			} else {
1199 				bzero((char *)cbuf + psize, rounded - psize);
1200 				psize = rounded;
1201 				zio_push_transform(zio, cbuf,
1202 				    psize, lsize, NULL);
1203 			}
1204 		}
1205 	}
1206 
1207 	/*
1208 	 * The final pass of spa_sync() must be all rewrites, but the first
1209 	 * few passes offer a trade-off: allocating blocks defers convergence,
1210 	 * but newly allocated blocks are sequential, so they can be written
1211 	 * to disk faster.  Therefore, we allow the first few passes of
1212 	 * spa_sync() to allocate new blocks, but force rewrites after that.
1213 	 * There should only be a handful of blocks after pass 1 in any case.
1214 	 */
1215 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1216 	    BP_GET_PSIZE(bp) == psize &&
1217 	    pass >= zfs_sync_pass_rewrite) {
1218 		ASSERT(psize != 0);
1219 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1220 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1221 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1222 	} else {
1223 		BP_ZERO(bp);
1224 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1225 	}
1226 
1227 	if (psize == 0) {
1228 		if (zio->io_bp_orig.blk_birth != 0 &&
1229 		    spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1230 			BP_SET_LSIZE(bp, lsize);
1231 			BP_SET_TYPE(bp, zp->zp_type);
1232 			BP_SET_LEVEL(bp, zp->zp_level);
1233 			BP_SET_BIRTH(bp, zio->io_txg, 0);
1234 		}
1235 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1236 	} else {
1237 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1238 		BP_SET_LSIZE(bp, lsize);
1239 		BP_SET_TYPE(bp, zp->zp_type);
1240 		BP_SET_LEVEL(bp, zp->zp_level);
1241 		BP_SET_PSIZE(bp, psize);
1242 		BP_SET_COMPRESS(bp, compress);
1243 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1244 		BP_SET_DEDUP(bp, zp->zp_dedup);
1245 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1246 		if (zp->zp_dedup) {
1247 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1248 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1249 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1250 		}
1251 		if (zp->zp_nopwrite) {
1252 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1253 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1254 			zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1255 		}
1256 	}
1257 
1258 	return (ZIO_PIPELINE_CONTINUE);
1259 }
1260 
1261 static int
1262 zio_free_bp_init(zio_t *zio)
1263 {
1264 	blkptr_t *bp = zio->io_bp;
1265 
1266 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1267 		if (BP_GET_DEDUP(bp))
1268 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1269 	}
1270 
1271 	return (ZIO_PIPELINE_CONTINUE);
1272 }
1273 
1274 /*
1275  * ==========================================================================
1276  * Execute the I/O pipeline
1277  * ==========================================================================
1278  */
1279 
1280 static void
1281 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1282 {
1283 	spa_t *spa = zio->io_spa;
1284 	zio_type_t t = zio->io_type;
1285 	int flags = (cutinline ? TQ_FRONT : 0);
1286 
1287 	/*
1288 	 * If we're a config writer or a probe, the normal issue and
1289 	 * interrupt threads may all be blocked waiting for the config lock.
1290 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1291 	 */
1292 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1293 		t = ZIO_TYPE_NULL;
1294 
1295 	/*
1296 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1297 	 */
1298 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1299 		t = ZIO_TYPE_NULL;
1300 
1301 	/*
1302 	 * If this is a high priority I/O, then use the high priority taskq if
1303 	 * available.
1304 	 */
1305 	if (zio->io_priority == ZIO_PRIORITY_NOW &&
1306 	    spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1307 		q++;
1308 
1309 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1310 
1311 	/*
1312 	 * NB: We are assuming that the zio can only be dispatched
1313 	 * to a single taskq at a time.  It would be a grievous error
1314 	 * to dispatch the zio to another taskq at the same time.
1315 	 */
1316 	ASSERT(zio->io_tqent.tqent_next == NULL);
1317 	spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1318 	    flags, &zio->io_tqent);
1319 }
1320 
1321 static boolean_t
1322 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1323 {
1324 	kthread_t *executor = zio->io_executor;
1325 	spa_t *spa = zio->io_spa;
1326 
1327 	for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1328 		spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1329 		uint_t i;
1330 		for (i = 0; i < tqs->stqs_count; i++) {
1331 			if (taskq_member(tqs->stqs_taskq[i], executor))
1332 				return (B_TRUE);
1333 		}
1334 	}
1335 
1336 	return (B_FALSE);
1337 }
1338 
1339 static int
1340 zio_issue_async(zio_t *zio)
1341 {
1342 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1343 
1344 	return (ZIO_PIPELINE_STOP);
1345 }
1346 
1347 void
1348 zio_interrupt(zio_t *zio)
1349 {
1350 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1351 }
1352 
1353 /*
1354  * Execute the I/O pipeline until one of the following occurs:
1355  *
1356  *	(1) the I/O completes
1357  *	(2) the pipeline stalls waiting for dependent child I/Os
1358  *	(3) the I/O issues, so we're waiting for an I/O completion interrupt
1359  *	(4) the I/O is delegated by vdev-level caching or aggregation
1360  *	(5) the I/O is deferred due to vdev-level queueing
1361  *	(6) the I/O is handed off to another thread.
1362  *
1363  * In all cases, the pipeline stops whenever there's no CPU work; it never
1364  * burns a thread in cv_wait().
1365  *
1366  * There's no locking on io_stage because there's no legitimate way
1367  * for multiple threads to be attempting to process the same I/O.
1368  */
1369 static zio_pipe_stage_t *zio_pipeline[];
1370 
1371 void
1372 zio_execute(zio_t *zio)
1373 {
1374 	zio->io_executor = curthread;
1375 
1376 	while (zio->io_stage < ZIO_STAGE_DONE) {
1377 		enum zio_stage pipeline = zio->io_pipeline;
1378 		enum zio_stage stage = zio->io_stage;
1379 		int rv;
1380 
1381 		ASSERT(!MUTEX_HELD(&zio->io_lock));
1382 		ASSERT(ISP2(stage));
1383 		ASSERT(zio->io_stall == NULL);
1384 
1385 		do {
1386 			stage <<= 1;
1387 		} while ((stage & pipeline) == 0);
1388 
1389 		ASSERT(stage <= ZIO_STAGE_DONE);
1390 
1391 		/*
1392 		 * If we are in interrupt context and this pipeline stage
1393 		 * will grab a config lock that is held across I/O,
1394 		 * or may wait for an I/O that needs an interrupt thread
1395 		 * to complete, issue async to avoid deadlock.
1396 		 *
1397 		 * For VDEV_IO_START, we cut in line so that the io will
1398 		 * be sent to disk promptly.
1399 		 */
1400 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1401 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1402 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1403 			    zio_requeue_io_start_cut_in_line : B_FALSE;
1404 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1405 			return;
1406 		}
1407 
1408 		zio->io_stage = stage;
1409 		rv = zio_pipeline[highbit64(stage) - 1](zio);
1410 
1411 		if (rv == ZIO_PIPELINE_STOP)
1412 			return;
1413 
1414 		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1415 	}
1416 }
1417 
1418 /*
1419  * ==========================================================================
1420  * Initiate I/O, either sync or async
1421  * ==========================================================================
1422  */
1423 int
1424 zio_wait(zio_t *zio)
1425 {
1426 	int error;
1427 
1428 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1429 	ASSERT(zio->io_executor == NULL);
1430 
1431 	zio->io_waiter = curthread;
1432 
1433 	zio_execute(zio);
1434 
1435 	mutex_enter(&zio->io_lock);
1436 	while (zio->io_executor != NULL)
1437 		cv_wait(&zio->io_cv, &zio->io_lock);
1438 	mutex_exit(&zio->io_lock);
1439 
1440 	error = zio->io_error;
1441 	zio_destroy(zio);
1442 
1443 	return (error);
1444 }
1445 
1446 void
1447 zio_nowait(zio_t *zio)
1448 {
1449 	ASSERT(zio->io_executor == NULL);
1450 
1451 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1452 	    zio_unique_parent(zio) == NULL) {
1453 		/*
1454 		 * This is a logical async I/O with no parent to wait for it.
1455 		 * We add it to the spa_async_root_zio "Godfather" I/O which
1456 		 * will ensure they complete prior to unloading the pool.
1457 		 */
1458 		spa_t *spa = zio->io_spa;
1459 
1460 		zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1461 	}
1462 
1463 	zio_execute(zio);
1464 }
1465 
1466 /*
1467  * ==========================================================================
1468  * Reexecute or suspend/resume failed I/O
1469  * ==========================================================================
1470  */
1471 
1472 static void
1473 zio_reexecute(zio_t *pio)
1474 {
1475 	zio_t *cio, *cio_next;
1476 
1477 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1478 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1479 	ASSERT(pio->io_gang_leader == NULL);
1480 	ASSERT(pio->io_gang_tree == NULL);
1481 
1482 	pio->io_flags = pio->io_orig_flags;
1483 	pio->io_stage = pio->io_orig_stage;
1484 	pio->io_pipeline = pio->io_orig_pipeline;
1485 	pio->io_reexecute = 0;
1486 	pio->io_flags |= ZIO_FLAG_REEXECUTED;
1487 	pio->io_error = 0;
1488 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1489 		pio->io_state[w] = 0;
1490 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1491 		pio->io_child_error[c] = 0;
1492 
1493 	if (IO_IS_ALLOCATING(pio))
1494 		BP_ZERO(pio->io_bp);
1495 
1496 	/*
1497 	 * As we reexecute pio's children, new children could be created.
1498 	 * New children go to the head of pio's io_child_list, however,
1499 	 * so we will (correctly) not reexecute them.  The key is that
1500 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1501 	 * cannot be affected by any side effects of reexecuting 'cio'.
1502 	 */
1503 	for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1504 		cio_next = zio_walk_children(pio);
1505 		mutex_enter(&pio->io_lock);
1506 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1507 			pio->io_children[cio->io_child_type][w]++;
1508 		mutex_exit(&pio->io_lock);
1509 		zio_reexecute(cio);
1510 	}
1511 
1512 	/*
1513 	 * Now that all children have been reexecuted, execute the parent.
1514 	 * We don't reexecute "The Godfather" I/O here as it's the
1515 	 * responsibility of the caller to wait on him.
1516 	 */
1517 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1518 		zio_execute(pio);
1519 }
1520 
1521 void
1522 zio_suspend(spa_t *spa, zio_t *zio)
1523 {
1524 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1525 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1526 		    "failure and the failure mode property for this pool "
1527 		    "is set to panic.", spa_name(spa));
1528 
1529 	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1530 
1531 	mutex_enter(&spa->spa_suspend_lock);
1532 
1533 	if (spa->spa_suspend_zio_root == NULL)
1534 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1535 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1536 		    ZIO_FLAG_GODFATHER);
1537 
1538 	spa->spa_suspended = B_TRUE;
1539 
1540 	if (zio != NULL) {
1541 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1542 		ASSERT(zio != spa->spa_suspend_zio_root);
1543 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1544 		ASSERT(zio_unique_parent(zio) == NULL);
1545 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1546 		zio_add_child(spa->spa_suspend_zio_root, zio);
1547 	}
1548 
1549 	mutex_exit(&spa->spa_suspend_lock);
1550 }
1551 
1552 int
1553 zio_resume(spa_t *spa)
1554 {
1555 	zio_t *pio;
1556 
1557 	/*
1558 	 * Reexecute all previously suspended i/o.
1559 	 */
1560 	mutex_enter(&spa->spa_suspend_lock);
1561 	spa->spa_suspended = B_FALSE;
1562 	cv_broadcast(&spa->spa_suspend_cv);
1563 	pio = spa->spa_suspend_zio_root;
1564 	spa->spa_suspend_zio_root = NULL;
1565 	mutex_exit(&spa->spa_suspend_lock);
1566 
1567 	if (pio == NULL)
1568 		return (0);
1569 
1570 	zio_reexecute(pio);
1571 	return (zio_wait(pio));
1572 }
1573 
1574 void
1575 zio_resume_wait(spa_t *spa)
1576 {
1577 	mutex_enter(&spa->spa_suspend_lock);
1578 	while (spa_suspended(spa))
1579 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1580 	mutex_exit(&spa->spa_suspend_lock);
1581 }
1582 
1583 /*
1584  * ==========================================================================
1585  * Gang blocks.
1586  *
1587  * A gang block is a collection of small blocks that looks to the DMU
1588  * like one large block.  When zio_dva_allocate() cannot find a block
1589  * of the requested size, due to either severe fragmentation or the pool
1590  * being nearly full, it calls zio_write_gang_block() to construct the
1591  * block from smaller fragments.
1592  *
1593  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1594  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
1595  * an indirect block: it's an array of block pointers.  It consumes
1596  * only one sector and hence is allocatable regardless of fragmentation.
1597  * The gang header's bps point to its gang members, which hold the data.
1598  *
1599  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1600  * as the verifier to ensure uniqueness of the SHA256 checksum.
1601  * Critically, the gang block bp's blk_cksum is the checksum of the data,
1602  * not the gang header.  This ensures that data block signatures (needed for
1603  * deduplication) are independent of how the block is physically stored.
1604  *
1605  * Gang blocks can be nested: a gang member may itself be a gang block.
1606  * Thus every gang block is a tree in which root and all interior nodes are
1607  * gang headers, and the leaves are normal blocks that contain user data.
1608  * The root of the gang tree is called the gang leader.
1609  *
1610  * To perform any operation (read, rewrite, free, claim) on a gang block,
1611  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1612  * in the io_gang_tree field of the original logical i/o by recursively
1613  * reading the gang leader and all gang headers below it.  This yields
1614  * an in-core tree containing the contents of every gang header and the
1615  * bps for every constituent of the gang block.
1616  *
1617  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1618  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
1619  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1620  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1621  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1622  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
1623  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1624  * of the gang header plus zio_checksum_compute() of the data to update the
1625  * gang header's blk_cksum as described above.
1626  *
1627  * The two-phase assemble/issue model solves the problem of partial failure --
1628  * what if you'd freed part of a gang block but then couldn't read the
1629  * gang header for another part?  Assembling the entire gang tree first
1630  * ensures that all the necessary gang header I/O has succeeded before
1631  * starting the actual work of free, claim, or write.  Once the gang tree
1632  * is assembled, free and claim are in-memory operations that cannot fail.
1633  *
1634  * In the event that a gang write fails, zio_dva_unallocate() walks the
1635  * gang tree to immediately free (i.e. insert back into the space map)
1636  * everything we've allocated.  This ensures that we don't get ENOSPC
1637  * errors during repeated suspend/resume cycles due to a flaky device.
1638  *
1639  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
1640  * the gang tree, we won't modify the block, so we can safely defer the free
1641  * (knowing that the block is still intact).  If we *can* assemble the gang
1642  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1643  * each constituent bp and we can allocate a new block on the next sync pass.
1644  *
1645  * In all cases, the gang tree allows complete recovery from partial failure.
1646  * ==========================================================================
1647  */
1648 
1649 static zio_t *
1650 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1651 {
1652 	if (gn != NULL)
1653 		return (pio);
1654 
1655 	return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1656 	    NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1657 	    &pio->io_bookmark));
1658 }
1659 
1660 zio_t *
1661 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1662 {
1663 	zio_t *zio;
1664 
1665 	if (gn != NULL) {
1666 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1667 		    gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1668 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1669 		/*
1670 		 * As we rewrite each gang header, the pipeline will compute
1671 		 * a new gang block header checksum for it; but no one will
1672 		 * compute a new data checksum, so we do that here.  The one
1673 		 * exception is the gang leader: the pipeline already computed
1674 		 * its data checksum because that stage precedes gang assembly.
1675 		 * (Presently, nothing actually uses interior data checksums;
1676 		 * this is just good hygiene.)
1677 		 */
1678 		if (gn != pio->io_gang_leader->io_gang_tree) {
1679 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1680 			    data, BP_GET_PSIZE(bp));
1681 		}
1682 		/*
1683 		 * If we are here to damage data for testing purposes,
1684 		 * leave the GBH alone so that we can detect the damage.
1685 		 */
1686 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1687 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1688 	} else {
1689 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1690 		    data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1691 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1692 	}
1693 
1694 	return (zio);
1695 }
1696 
1697 /* ARGSUSED */
1698 zio_t *
1699 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1700 {
1701 	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1702 	    ZIO_GANG_CHILD_FLAGS(pio)));
1703 }
1704 
1705 /* ARGSUSED */
1706 zio_t *
1707 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1708 {
1709 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1710 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1711 }
1712 
1713 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1714 	NULL,
1715 	zio_read_gang,
1716 	zio_rewrite_gang,
1717 	zio_free_gang,
1718 	zio_claim_gang,
1719 	NULL
1720 };
1721 
1722 static void zio_gang_tree_assemble_done(zio_t *zio);
1723 
1724 static zio_gang_node_t *
1725 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1726 {
1727 	zio_gang_node_t *gn;
1728 
1729 	ASSERT(*gnpp == NULL);
1730 
1731 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1732 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1733 	*gnpp = gn;
1734 
1735 	return (gn);
1736 }
1737 
1738 static void
1739 zio_gang_node_free(zio_gang_node_t **gnpp)
1740 {
1741 	zio_gang_node_t *gn = *gnpp;
1742 
1743 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1744 		ASSERT(gn->gn_child[g] == NULL);
1745 
1746 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1747 	kmem_free(gn, sizeof (*gn));
1748 	*gnpp = NULL;
1749 }
1750 
1751 static void
1752 zio_gang_tree_free(zio_gang_node_t **gnpp)
1753 {
1754 	zio_gang_node_t *gn = *gnpp;
1755 
1756 	if (gn == NULL)
1757 		return;
1758 
1759 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1760 		zio_gang_tree_free(&gn->gn_child[g]);
1761 
1762 	zio_gang_node_free(gnpp);
1763 }
1764 
1765 static void
1766 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1767 {
1768 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1769 
1770 	ASSERT(gio->io_gang_leader == gio);
1771 	ASSERT(BP_IS_GANG(bp));
1772 
1773 	zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1774 	    SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1775 	    gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1776 }
1777 
1778 static void
1779 zio_gang_tree_assemble_done(zio_t *zio)
1780 {
1781 	zio_t *gio = zio->io_gang_leader;
1782 	zio_gang_node_t *gn = zio->io_private;
1783 	blkptr_t *bp = zio->io_bp;
1784 
1785 	ASSERT(gio == zio_unique_parent(zio));
1786 	ASSERT(zio->io_child_count == 0);
1787 
1788 	if (zio->io_error)
1789 		return;
1790 
1791 	if (BP_SHOULD_BYTESWAP(bp))
1792 		byteswap_uint64_array(zio->io_data, zio->io_size);
1793 
1794 	ASSERT(zio->io_data == gn->gn_gbh);
1795 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1796 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1797 
1798 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1799 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1800 		if (!BP_IS_GANG(gbp))
1801 			continue;
1802 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1803 	}
1804 }
1805 
1806 static void
1807 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1808 {
1809 	zio_t *gio = pio->io_gang_leader;
1810 	zio_t *zio;
1811 
1812 	ASSERT(BP_IS_GANG(bp) == !!gn);
1813 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1814 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1815 
1816 	/*
1817 	 * If you're a gang header, your data is in gn->gn_gbh.
1818 	 * If you're a gang member, your data is in 'data' and gn == NULL.
1819 	 */
1820 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1821 
1822 	if (gn != NULL) {
1823 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1824 
1825 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1826 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1827 			if (BP_IS_HOLE(gbp))
1828 				continue;
1829 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1830 			data = (char *)data + BP_GET_PSIZE(gbp);
1831 		}
1832 	}
1833 
1834 	if (gn == gio->io_gang_tree)
1835 		ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1836 
1837 	if (zio != pio)
1838 		zio_nowait(zio);
1839 }
1840 
1841 static int
1842 zio_gang_assemble(zio_t *zio)
1843 {
1844 	blkptr_t *bp = zio->io_bp;
1845 
1846 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1847 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1848 
1849 	zio->io_gang_leader = zio;
1850 
1851 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1852 
1853 	return (ZIO_PIPELINE_CONTINUE);
1854 }
1855 
1856 static int
1857 zio_gang_issue(zio_t *zio)
1858 {
1859 	blkptr_t *bp = zio->io_bp;
1860 
1861 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1862 		return (ZIO_PIPELINE_STOP);
1863 
1864 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1865 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1866 
1867 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1868 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1869 	else
1870 		zio_gang_tree_free(&zio->io_gang_tree);
1871 
1872 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1873 
1874 	return (ZIO_PIPELINE_CONTINUE);
1875 }
1876 
1877 static void
1878 zio_write_gang_member_ready(zio_t *zio)
1879 {
1880 	zio_t *pio = zio_unique_parent(zio);
1881 	zio_t *gio = zio->io_gang_leader;
1882 	dva_t *cdva = zio->io_bp->blk_dva;
1883 	dva_t *pdva = pio->io_bp->blk_dva;
1884 	uint64_t asize;
1885 
1886 	if (BP_IS_HOLE(zio->io_bp))
1887 		return;
1888 
1889 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1890 
1891 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1892 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1893 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1894 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1895 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1896 
1897 	mutex_enter(&pio->io_lock);
1898 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1899 		ASSERT(DVA_GET_GANG(&pdva[d]));
1900 		asize = DVA_GET_ASIZE(&pdva[d]);
1901 		asize += DVA_GET_ASIZE(&cdva[d]);
1902 		DVA_SET_ASIZE(&pdva[d], asize);
1903 	}
1904 	mutex_exit(&pio->io_lock);
1905 }
1906 
1907 static int
1908 zio_write_gang_block(zio_t *pio)
1909 {
1910 	spa_t *spa = pio->io_spa;
1911 	blkptr_t *bp = pio->io_bp;
1912 	zio_t *gio = pio->io_gang_leader;
1913 	zio_t *zio;
1914 	zio_gang_node_t *gn, **gnpp;
1915 	zio_gbh_phys_t *gbh;
1916 	uint64_t txg = pio->io_txg;
1917 	uint64_t resid = pio->io_size;
1918 	uint64_t lsize;
1919 	int copies = gio->io_prop.zp_copies;
1920 	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1921 	zio_prop_t zp;
1922 	int error;
1923 
1924 	error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1925 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1926 	    METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1927 	if (error) {
1928 		pio->io_error = error;
1929 		return (ZIO_PIPELINE_CONTINUE);
1930 	}
1931 
1932 	if (pio == gio) {
1933 		gnpp = &gio->io_gang_tree;
1934 	} else {
1935 		gnpp = pio->io_private;
1936 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
1937 	}
1938 
1939 	gn = zio_gang_node_alloc(gnpp);
1940 	gbh = gn->gn_gbh;
1941 	bzero(gbh, SPA_GANGBLOCKSIZE);
1942 
1943 	/*
1944 	 * Create the gang header.
1945 	 */
1946 	zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1947 	    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1948 
1949 	/*
1950 	 * Create and nowait the gang children.
1951 	 */
1952 	for (int g = 0; resid != 0; resid -= lsize, g++) {
1953 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1954 		    SPA_MINBLOCKSIZE);
1955 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1956 
1957 		zp.zp_checksum = gio->io_prop.zp_checksum;
1958 		zp.zp_compress = ZIO_COMPRESS_OFF;
1959 		zp.zp_type = DMU_OT_NONE;
1960 		zp.zp_level = 0;
1961 		zp.zp_copies = gio->io_prop.zp_copies;
1962 		zp.zp_dedup = B_FALSE;
1963 		zp.zp_dedup_verify = B_FALSE;
1964 		zp.zp_nopwrite = B_FALSE;
1965 
1966 		zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1967 		    (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1968 		    zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1969 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1970 		    &pio->io_bookmark));
1971 	}
1972 
1973 	/*
1974 	 * Set pio's pipeline to just wait for zio to finish.
1975 	 */
1976 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1977 
1978 	zio_nowait(zio);
1979 
1980 	return (ZIO_PIPELINE_CONTINUE);
1981 }
1982 
1983 /*
1984  * The zio_nop_write stage in the pipeline determines if allocating
1985  * a new bp is necessary.  By leveraging a cryptographically secure checksum,
1986  * such as SHA256, we can compare the checksums of the new data and the old
1987  * to determine if allocating a new block is required.  The nopwrite
1988  * feature can handle writes in either syncing or open context (i.e. zil
1989  * writes) and as a result is mutually exclusive with dedup.
1990  */
1991 static int
1992 zio_nop_write(zio_t *zio)
1993 {
1994 	blkptr_t *bp = zio->io_bp;
1995 	blkptr_t *bp_orig = &zio->io_bp_orig;
1996 	zio_prop_t *zp = &zio->io_prop;
1997 
1998 	ASSERT(BP_GET_LEVEL(bp) == 0);
1999 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2000 	ASSERT(zp->zp_nopwrite);
2001 	ASSERT(!zp->zp_dedup);
2002 	ASSERT(zio->io_bp_override == NULL);
2003 	ASSERT(IO_IS_ALLOCATING(zio));
2004 
2005 	/*
2006 	 * Check to see if the original bp and the new bp have matching
2007 	 * characteristics (i.e. same checksum, compression algorithms, etc).
2008 	 * If they don't then just continue with the pipeline which will
2009 	 * allocate a new bp.
2010 	 */
2011 	if (BP_IS_HOLE(bp_orig) ||
2012 	    !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
2013 	    BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2014 	    BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2015 	    BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2016 	    zp->zp_copies != BP_GET_NDVAS(bp_orig))
2017 		return (ZIO_PIPELINE_CONTINUE);
2018 
2019 	/*
2020 	 * If the checksums match then reset the pipeline so that we
2021 	 * avoid allocating a new bp and issuing any I/O.
2022 	 */
2023 	if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2024 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
2025 		ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2026 		ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2027 		ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2028 		ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2029 		    sizeof (uint64_t)) == 0);
2030 
2031 		*bp = *bp_orig;
2032 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2033 		zio->io_flags |= ZIO_FLAG_NOPWRITE;
2034 	}
2035 
2036 	return (ZIO_PIPELINE_CONTINUE);
2037 }
2038 
2039 /*
2040  * ==========================================================================
2041  * Dedup
2042  * ==========================================================================
2043  */
2044 static void
2045 zio_ddt_child_read_done(zio_t *zio)
2046 {
2047 	blkptr_t *bp = zio->io_bp;
2048 	ddt_entry_t *dde = zio->io_private;
2049 	ddt_phys_t *ddp;
2050 	zio_t *pio = zio_unique_parent(zio);
2051 
2052 	mutex_enter(&pio->io_lock);
2053 	ddp = ddt_phys_select(dde, bp);
2054 	if (zio->io_error == 0)
2055 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
2056 	if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2057 		dde->dde_repair_data = zio->io_data;
2058 	else
2059 		zio_buf_free(zio->io_data, zio->io_size);
2060 	mutex_exit(&pio->io_lock);
2061 }
2062 
2063 static int
2064 zio_ddt_read_start(zio_t *zio)
2065 {
2066 	blkptr_t *bp = zio->io_bp;
2067 
2068 	ASSERT(BP_GET_DEDUP(bp));
2069 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2070 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2071 
2072 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
2073 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
2074 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2075 		ddt_phys_t *ddp = dde->dde_phys;
2076 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2077 		blkptr_t blk;
2078 
2079 		ASSERT(zio->io_vsd == NULL);
2080 		zio->io_vsd = dde;
2081 
2082 		if (ddp_self == NULL)
2083 			return (ZIO_PIPELINE_CONTINUE);
2084 
2085 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2086 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2087 				continue;
2088 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2089 			    &blk);
2090 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
2091 			    zio_buf_alloc(zio->io_size), zio->io_size,
2092 			    zio_ddt_child_read_done, dde, zio->io_priority,
2093 			    ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2094 			    &zio->io_bookmark));
2095 		}
2096 		return (ZIO_PIPELINE_CONTINUE);
2097 	}
2098 
2099 	zio_nowait(zio_read(zio, zio->io_spa, bp,
2100 	    zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2101 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2102 
2103 	return (ZIO_PIPELINE_CONTINUE);
2104 }
2105 
2106 static int
2107 zio_ddt_read_done(zio_t *zio)
2108 {
2109 	blkptr_t *bp = zio->io_bp;
2110 
2111 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2112 		return (ZIO_PIPELINE_STOP);
2113 
2114 	ASSERT(BP_GET_DEDUP(bp));
2115 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2116 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2117 
2118 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
2119 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
2120 		ddt_entry_t *dde = zio->io_vsd;
2121 		if (ddt == NULL) {
2122 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2123 			return (ZIO_PIPELINE_CONTINUE);
2124 		}
2125 		if (dde == NULL) {
2126 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2127 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2128 			return (ZIO_PIPELINE_STOP);
2129 		}
2130 		if (dde->dde_repair_data != NULL) {
2131 			bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2132 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
2133 		}
2134 		ddt_repair_done(ddt, dde);
2135 		zio->io_vsd = NULL;
2136 	}
2137 
2138 	ASSERT(zio->io_vsd == NULL);
2139 
2140 	return (ZIO_PIPELINE_CONTINUE);
2141 }
2142 
2143 static boolean_t
2144 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2145 {
2146 	spa_t *spa = zio->io_spa;
2147 
2148 	/*
2149 	 * Note: we compare the original data, not the transformed data,
2150 	 * because when zio->io_bp is an override bp, we will not have
2151 	 * pushed the I/O transforms.  That's an important optimization
2152 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2153 	 */
2154 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2155 		zio_t *lio = dde->dde_lead_zio[p];
2156 
2157 		if (lio != NULL) {
2158 			return (lio->io_orig_size != zio->io_orig_size ||
2159 			    bcmp(zio->io_orig_data, lio->io_orig_data,
2160 			    zio->io_orig_size) != 0);
2161 		}
2162 	}
2163 
2164 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2165 		ddt_phys_t *ddp = &dde->dde_phys[p];
2166 
2167 		if (ddp->ddp_phys_birth != 0) {
2168 			arc_buf_t *abuf = NULL;
2169 			arc_flags_t aflags = ARC_FLAG_WAIT;
2170 			blkptr_t blk = *zio->io_bp;
2171 			int error;
2172 
2173 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2174 
2175 			ddt_exit(ddt);
2176 
2177 			error = arc_read(NULL, spa, &blk,
2178 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2179 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2180 			    &aflags, &zio->io_bookmark);
2181 
2182 			if (error == 0) {
2183 				if (arc_buf_size(abuf) != zio->io_orig_size ||
2184 				    bcmp(abuf->b_data, zio->io_orig_data,
2185 				    zio->io_orig_size) != 0)
2186 					error = SET_ERROR(EEXIST);
2187 				VERIFY(arc_buf_remove_ref(abuf, &abuf));
2188 			}
2189 
2190 			ddt_enter(ddt);
2191 			return (error != 0);
2192 		}
2193 	}
2194 
2195 	return (B_FALSE);
2196 }
2197 
2198 static void
2199 zio_ddt_child_write_ready(zio_t *zio)
2200 {
2201 	int p = zio->io_prop.zp_copies;
2202 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2203 	ddt_entry_t *dde = zio->io_private;
2204 	ddt_phys_t *ddp = &dde->dde_phys[p];
2205 	zio_t *pio;
2206 
2207 	if (zio->io_error)
2208 		return;
2209 
2210 	ddt_enter(ddt);
2211 
2212 	ASSERT(dde->dde_lead_zio[p] == zio);
2213 
2214 	ddt_phys_fill(ddp, zio->io_bp);
2215 
2216 	while ((pio = zio_walk_parents(zio)) != NULL)
2217 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2218 
2219 	ddt_exit(ddt);
2220 }
2221 
2222 static void
2223 zio_ddt_child_write_done(zio_t *zio)
2224 {
2225 	int p = zio->io_prop.zp_copies;
2226 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2227 	ddt_entry_t *dde = zio->io_private;
2228 	ddt_phys_t *ddp = &dde->dde_phys[p];
2229 
2230 	ddt_enter(ddt);
2231 
2232 	ASSERT(ddp->ddp_refcnt == 0);
2233 	ASSERT(dde->dde_lead_zio[p] == zio);
2234 	dde->dde_lead_zio[p] = NULL;
2235 
2236 	if (zio->io_error == 0) {
2237 		while (zio_walk_parents(zio) != NULL)
2238 			ddt_phys_addref(ddp);
2239 	} else {
2240 		ddt_phys_clear(ddp);
2241 	}
2242 
2243 	ddt_exit(ddt);
2244 }
2245 
2246 static void
2247 zio_ddt_ditto_write_done(zio_t *zio)
2248 {
2249 	int p = DDT_PHYS_DITTO;
2250 	zio_prop_t *zp = &zio->io_prop;
2251 	blkptr_t *bp = zio->io_bp;
2252 	ddt_t *ddt = ddt_select(zio->io_spa, bp);
2253 	ddt_entry_t *dde = zio->io_private;
2254 	ddt_phys_t *ddp = &dde->dde_phys[p];
2255 	ddt_key_t *ddk = &dde->dde_key;
2256 
2257 	ddt_enter(ddt);
2258 
2259 	ASSERT(ddp->ddp_refcnt == 0);
2260 	ASSERT(dde->dde_lead_zio[p] == zio);
2261 	dde->dde_lead_zio[p] = NULL;
2262 
2263 	if (zio->io_error == 0) {
2264 		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2265 		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2266 		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2267 		if (ddp->ddp_phys_birth != 0)
2268 			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2269 		ddt_phys_fill(ddp, bp);
2270 	}
2271 
2272 	ddt_exit(ddt);
2273 }
2274 
2275 static int
2276 zio_ddt_write(zio_t *zio)
2277 {
2278 	spa_t *spa = zio->io_spa;
2279 	blkptr_t *bp = zio->io_bp;
2280 	uint64_t txg = zio->io_txg;
2281 	zio_prop_t *zp = &zio->io_prop;
2282 	int p = zp->zp_copies;
2283 	int ditto_copies;
2284 	zio_t *cio = NULL;
2285 	zio_t *dio = NULL;
2286 	ddt_t *ddt = ddt_select(spa, bp);
2287 	ddt_entry_t *dde;
2288 	ddt_phys_t *ddp;
2289 
2290 	ASSERT(BP_GET_DEDUP(bp));
2291 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2292 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2293 
2294 	ddt_enter(ddt);
2295 	dde = ddt_lookup(ddt, bp, B_TRUE);
2296 	ddp = &dde->dde_phys[p];
2297 
2298 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2299 		/*
2300 		 * If we're using a weak checksum, upgrade to a strong checksum
2301 		 * and try again.  If we're already using a strong checksum,
2302 		 * we can't resolve it, so just convert to an ordinary write.
2303 		 * (And automatically e-mail a paper to Nature?)
2304 		 */
2305 		if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2306 			zp->zp_checksum = spa_dedup_checksum(spa);
2307 			zio_pop_transforms(zio);
2308 			zio->io_stage = ZIO_STAGE_OPEN;
2309 			BP_ZERO(bp);
2310 		} else {
2311 			zp->zp_dedup = B_FALSE;
2312 		}
2313 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2314 		ddt_exit(ddt);
2315 		return (ZIO_PIPELINE_CONTINUE);
2316 	}
2317 
2318 	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2319 	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2320 
2321 	if (ditto_copies > ddt_ditto_copies_present(dde) &&
2322 	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2323 		zio_prop_t czp = *zp;
2324 
2325 		czp.zp_copies = ditto_copies;
2326 
2327 		/*
2328 		 * If we arrived here with an override bp, we won't have run
2329 		 * the transform stack, so we won't have the data we need to
2330 		 * generate a child i/o.  So, toss the override bp and restart.
2331 		 * This is safe, because using the override bp is just an
2332 		 * optimization; and it's rare, so the cost doesn't matter.
2333 		 */
2334 		if (zio->io_bp_override) {
2335 			zio_pop_transforms(zio);
2336 			zio->io_stage = ZIO_STAGE_OPEN;
2337 			zio->io_pipeline = ZIO_WRITE_PIPELINE;
2338 			zio->io_bp_override = NULL;
2339 			BP_ZERO(bp);
2340 			ddt_exit(ddt);
2341 			return (ZIO_PIPELINE_CONTINUE);
2342 		}
2343 
2344 		dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2345 		    zio->io_orig_size, &czp, NULL, NULL,
2346 		    zio_ddt_ditto_write_done, dde, zio->io_priority,
2347 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2348 
2349 		zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2350 		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2351 	}
2352 
2353 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2354 		if (ddp->ddp_phys_birth != 0)
2355 			ddt_bp_fill(ddp, bp, txg);
2356 		if (dde->dde_lead_zio[p] != NULL)
2357 			zio_add_child(zio, dde->dde_lead_zio[p]);
2358 		else
2359 			ddt_phys_addref(ddp);
2360 	} else if (zio->io_bp_override) {
2361 		ASSERT(bp->blk_birth == txg);
2362 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2363 		ddt_phys_fill(ddp, bp);
2364 		ddt_phys_addref(ddp);
2365 	} else {
2366 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2367 		    zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2368 		    zio_ddt_child_write_done, dde, zio->io_priority,
2369 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2370 
2371 		zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2372 		dde->dde_lead_zio[p] = cio;
2373 	}
2374 
2375 	ddt_exit(ddt);
2376 
2377 	if (cio)
2378 		zio_nowait(cio);
2379 	if (dio)
2380 		zio_nowait(dio);
2381 
2382 	return (ZIO_PIPELINE_CONTINUE);
2383 }
2384 
2385 ddt_entry_t *freedde; /* for debugging */
2386 
2387 static int
2388 zio_ddt_free(zio_t *zio)
2389 {
2390 	spa_t *spa = zio->io_spa;
2391 	blkptr_t *bp = zio->io_bp;
2392 	ddt_t *ddt = ddt_select(spa, bp);
2393 	ddt_entry_t *dde;
2394 	ddt_phys_t *ddp;
2395 
2396 	ASSERT(BP_GET_DEDUP(bp));
2397 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2398 
2399 	ddt_enter(ddt);
2400 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2401 	ddp = ddt_phys_select(dde, bp);
2402 	ddt_phys_decref(ddp);
2403 	ddt_exit(ddt);
2404 
2405 	return (ZIO_PIPELINE_CONTINUE);
2406 }
2407 
2408 /*
2409  * ==========================================================================
2410  * Allocate and free blocks
2411  * ==========================================================================
2412  */
2413 static int
2414 zio_dva_allocate(zio_t *zio)
2415 {
2416 	spa_t *spa = zio->io_spa;
2417 	metaslab_class_t *mc = spa_normal_class(spa);
2418 	blkptr_t *bp = zio->io_bp;
2419 	int error;
2420 	int flags = 0;
2421 
2422 	if (zio->io_gang_leader == NULL) {
2423 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2424 		zio->io_gang_leader = zio;
2425 	}
2426 
2427 	ASSERT(BP_IS_HOLE(bp));
2428 	ASSERT0(BP_GET_NDVAS(bp));
2429 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
2430 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2431 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2432 
2433 	/*
2434 	 * The dump device does not support gang blocks so allocation on
2435 	 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2436 	 * the "fast" gang feature.
2437 	 */
2438 	flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2439 	flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2440 	    METASLAB_GANG_CHILD : 0;
2441 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
2442 	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2443 
2444 	if (error) {
2445 		spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2446 		    "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2447 		    error);
2448 		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2449 			return (zio_write_gang_block(zio));
2450 		zio->io_error = error;
2451 	}
2452 
2453 	return (ZIO_PIPELINE_CONTINUE);
2454 }
2455 
2456 static int
2457 zio_dva_free(zio_t *zio)
2458 {
2459 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2460 
2461 	return (ZIO_PIPELINE_CONTINUE);
2462 }
2463 
2464 static int
2465 zio_dva_claim(zio_t *zio)
2466 {
2467 	int error;
2468 
2469 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2470 	if (error)
2471 		zio->io_error = error;
2472 
2473 	return (ZIO_PIPELINE_CONTINUE);
2474 }
2475 
2476 /*
2477  * Undo an allocation.  This is used by zio_done() when an I/O fails
2478  * and we want to give back the block we just allocated.
2479  * This handles both normal blocks and gang blocks.
2480  */
2481 static void
2482 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2483 {
2484 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2485 	ASSERT(zio->io_bp_override == NULL);
2486 
2487 	if (!BP_IS_HOLE(bp))
2488 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2489 
2490 	if (gn != NULL) {
2491 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2492 			zio_dva_unallocate(zio, gn->gn_child[g],
2493 			    &gn->gn_gbh->zg_blkptr[g]);
2494 		}
2495 	}
2496 }
2497 
2498 /*
2499  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
2500  */
2501 int
2502 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2503     uint64_t size, boolean_t use_slog)
2504 {
2505 	int error = 1;
2506 
2507 	ASSERT(txg > spa_syncing_txg(spa));
2508 
2509 	/*
2510 	 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2511 	 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2512 	 * when allocating them.
2513 	 */
2514 	if (use_slog) {
2515 		error = metaslab_alloc(spa, spa_log_class(spa), size,
2516 		    new_bp, 1, txg, old_bp,
2517 		    METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2518 	}
2519 
2520 	if (error) {
2521 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
2522 		    new_bp, 1, txg, old_bp,
2523 		    METASLAB_HINTBP_AVOID);
2524 	}
2525 
2526 	if (error == 0) {
2527 		BP_SET_LSIZE(new_bp, size);
2528 		BP_SET_PSIZE(new_bp, size);
2529 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2530 		BP_SET_CHECKSUM(new_bp,
2531 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2532 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2533 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2534 		BP_SET_LEVEL(new_bp, 0);
2535 		BP_SET_DEDUP(new_bp, 0);
2536 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2537 	}
2538 
2539 	return (error);
2540 }
2541 
2542 /*
2543  * Free an intent log block.
2544  */
2545 void
2546 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2547 {
2548 	ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2549 	ASSERT(!BP_IS_GANG(bp));
2550 
2551 	zio_free(spa, txg, bp);
2552 }
2553 
2554 /*
2555  * ==========================================================================
2556  * Read and write to physical devices
2557  * ==========================================================================
2558  */
2559 
2560 
2561 /*
2562  * Issue an I/O to the underlying vdev. Typically the issue pipeline
2563  * stops after this stage and will resume upon I/O completion.
2564  * However, there are instances where the vdev layer may need to
2565  * continue the pipeline when an I/O was not issued. Since the I/O
2566  * that was sent to the vdev layer might be different than the one
2567  * currently active in the pipeline (see vdev_queue_io()), we explicitly
2568  * force the underlying vdev layers to call either zio_execute() or
2569  * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
2570  */
2571 static int
2572 zio_vdev_io_start(zio_t *zio)
2573 {
2574 	vdev_t *vd = zio->io_vd;
2575 	uint64_t align;
2576 	spa_t *spa = zio->io_spa;
2577 
2578 	ASSERT(zio->io_error == 0);
2579 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2580 
2581 	if (vd == NULL) {
2582 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2583 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2584 
2585 		/*
2586 		 * The mirror_ops handle multiple DVAs in a single BP.
2587 		 */
2588 		vdev_mirror_ops.vdev_op_io_start(zio);
2589 		return (ZIO_PIPELINE_STOP);
2590 	}
2591 
2592 	/*
2593 	 * We keep track of time-sensitive I/Os so that the scan thread
2594 	 * can quickly react to certain workloads.  In particular, we care
2595 	 * about non-scrubbing, top-level reads and writes with the following
2596 	 * characteristics:
2597 	 *	- synchronous writes of user data to non-slog devices
2598 	 *	- any reads of user data
2599 	 * When these conditions are met, adjust the timestamp of spa_last_io
2600 	 * which allows the scan thread to adjust its workload accordingly.
2601 	 */
2602 	if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2603 	    vd == vd->vdev_top && !vd->vdev_islog &&
2604 	    zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2605 	    zio->io_txg != spa_syncing_txg(spa)) {
2606 		uint64_t old = spa->spa_last_io;
2607 		uint64_t new = ddi_get_lbolt64();
2608 		if (old != new)
2609 			(void) atomic_cas_64(&spa->spa_last_io, old, new);
2610 	}
2611 
2612 	align = 1ULL << vd->vdev_top->vdev_ashift;
2613 
2614 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
2615 	    P2PHASE(zio->io_size, align) != 0) {
2616 		/* Transform logical writes to be a full physical block size. */
2617 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
2618 		char *abuf = zio_buf_alloc(asize);
2619 		ASSERT(vd == vd->vdev_top);
2620 		if (zio->io_type == ZIO_TYPE_WRITE) {
2621 			bcopy(zio->io_data, abuf, zio->io_size);
2622 			bzero(abuf + zio->io_size, asize - zio->io_size);
2623 		}
2624 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2625 	}
2626 
2627 	/*
2628 	 * If this is not a physical io, make sure that it is properly aligned
2629 	 * before proceeding.
2630 	 */
2631 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2632 		ASSERT0(P2PHASE(zio->io_offset, align));
2633 		ASSERT0(P2PHASE(zio->io_size, align));
2634 	} else {
2635 		/*
2636 		 * For physical writes, we allow 512b aligned writes and assume
2637 		 * the device will perform a read-modify-write as necessary.
2638 		 */
2639 		ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2640 		ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2641 	}
2642 
2643 	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2644 
2645 	/*
2646 	 * If this is a repair I/O, and there's no self-healing involved --
2647 	 * that is, we're just resilvering what we expect to resilver --
2648 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2649 	 * This prevents spurious resilvering with nested replication.
2650 	 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2651 	 * A is out of date, we'll read from C+D, then use the data to
2652 	 * resilver A+B -- but we don't actually want to resilver B, just A.
2653 	 * The top-level mirror has no way to know this, so instead we just
2654 	 * discard unnecessary repairs as we work our way down the vdev tree.
2655 	 * The same logic applies to any form of nested replication:
2656 	 * ditto + mirror, RAID-Z + replacing, etc.  This covers them all.
2657 	 */
2658 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2659 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2660 	    zio->io_txg != 0 &&	/* not a delegated i/o */
2661 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2662 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2663 		zio_vdev_io_bypass(zio);
2664 		return (ZIO_PIPELINE_CONTINUE);
2665 	}
2666 
2667 	if (vd->vdev_ops->vdev_op_leaf &&
2668 	    (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2669 
2670 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
2671 			return (ZIO_PIPELINE_CONTINUE);
2672 
2673 		if ((zio = vdev_queue_io(zio)) == NULL)
2674 			return (ZIO_PIPELINE_STOP);
2675 
2676 		if (!vdev_accessible(vd, zio)) {
2677 			zio->io_error = SET_ERROR(ENXIO);
2678 			zio_interrupt(zio);
2679 			return (ZIO_PIPELINE_STOP);
2680 		}
2681 	}
2682 
2683 	vd->vdev_ops->vdev_op_io_start(zio);
2684 	return (ZIO_PIPELINE_STOP);
2685 }
2686 
2687 static int
2688 zio_vdev_io_done(zio_t *zio)
2689 {
2690 	vdev_t *vd = zio->io_vd;
2691 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2692 	boolean_t unexpected_error = B_FALSE;
2693 
2694 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2695 		return (ZIO_PIPELINE_STOP);
2696 
2697 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2698 
2699 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2700 
2701 		vdev_queue_io_done(zio);
2702 
2703 		if (zio->io_type == ZIO_TYPE_WRITE)
2704 			vdev_cache_write(zio);
2705 
2706 		if (zio_injection_enabled && zio->io_error == 0)
2707 			zio->io_error = zio_handle_device_injection(vd,
2708 			    zio, EIO);
2709 
2710 		if (zio_injection_enabled && zio->io_error == 0)
2711 			zio->io_error = zio_handle_label_injection(zio, EIO);
2712 
2713 		if (zio->io_error) {
2714 			if (!vdev_accessible(vd, zio)) {
2715 				zio->io_error = SET_ERROR(ENXIO);
2716 			} else {
2717 				unexpected_error = B_TRUE;
2718 			}
2719 		}
2720 	}
2721 
2722 	ops->vdev_op_io_done(zio);
2723 
2724 	if (unexpected_error)
2725 		VERIFY(vdev_probe(vd, zio) == NULL);
2726 
2727 	return (ZIO_PIPELINE_CONTINUE);
2728 }
2729 
2730 /*
2731  * For non-raidz ZIOs, we can just copy aside the bad data read from the
2732  * disk, and use that to finish the checksum ereport later.
2733  */
2734 static void
2735 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2736     const void *good_buf)
2737 {
2738 	/* no processing needed */
2739 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2740 }
2741 
2742 /*ARGSUSED*/
2743 void
2744 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2745 {
2746 	void *buf = zio_buf_alloc(zio->io_size);
2747 
2748 	bcopy(zio->io_data, buf, zio->io_size);
2749 
2750 	zcr->zcr_cbinfo = zio->io_size;
2751 	zcr->zcr_cbdata = buf;
2752 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
2753 	zcr->zcr_free = zio_buf_free;
2754 }
2755 
2756 static int
2757 zio_vdev_io_assess(zio_t *zio)
2758 {
2759 	vdev_t *vd = zio->io_vd;
2760 
2761 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2762 		return (ZIO_PIPELINE_STOP);
2763 
2764 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2765 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2766 
2767 	if (zio->io_vsd != NULL) {
2768 		zio->io_vsd_ops->vsd_free(zio);
2769 		zio->io_vsd = NULL;
2770 	}
2771 
2772 	if (zio_injection_enabled && zio->io_error == 0)
2773 		zio->io_error = zio_handle_fault_injection(zio, EIO);
2774 
2775 	/*
2776 	 * If the I/O failed, determine whether we should attempt to retry it.
2777 	 *
2778 	 * On retry, we cut in line in the issue queue, since we don't want
2779 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2780 	 */
2781 	if (zio->io_error && vd == NULL &&
2782 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2783 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
2784 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
2785 		zio->io_error = 0;
2786 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
2787 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2788 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2789 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2790 		    zio_requeue_io_start_cut_in_line);
2791 		return (ZIO_PIPELINE_STOP);
2792 	}
2793 
2794 	/*
2795 	 * If we got an error on a leaf device, convert it to ENXIO
2796 	 * if the device is not accessible at all.
2797 	 */
2798 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2799 	    !vdev_accessible(vd, zio))
2800 		zio->io_error = SET_ERROR(ENXIO);
2801 
2802 	/*
2803 	 * If we can't write to an interior vdev (mirror or RAID-Z),
2804 	 * set vdev_cant_write so that we stop trying to allocate from it.
2805 	 */
2806 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2807 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2808 		vd->vdev_cant_write = B_TRUE;
2809 	}
2810 
2811 	if (zio->io_error)
2812 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2813 
2814 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2815 	    zio->io_physdone != NULL) {
2816 		ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2817 		ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2818 		zio->io_physdone(zio->io_logical);
2819 	}
2820 
2821 	return (ZIO_PIPELINE_CONTINUE);
2822 }
2823 
2824 void
2825 zio_vdev_io_reissue(zio_t *zio)
2826 {
2827 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2828 	ASSERT(zio->io_error == 0);
2829 
2830 	zio->io_stage >>= 1;
2831 }
2832 
2833 void
2834 zio_vdev_io_redone(zio_t *zio)
2835 {
2836 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2837 
2838 	zio->io_stage >>= 1;
2839 }
2840 
2841 void
2842 zio_vdev_io_bypass(zio_t *zio)
2843 {
2844 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2845 	ASSERT(zio->io_error == 0);
2846 
2847 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2848 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2849 }
2850 
2851 /*
2852  * ==========================================================================
2853  * Generate and verify checksums
2854  * ==========================================================================
2855  */
2856 static int
2857 zio_checksum_generate(zio_t *zio)
2858 {
2859 	blkptr_t *bp = zio->io_bp;
2860 	enum zio_checksum checksum;
2861 
2862 	if (bp == NULL) {
2863 		/*
2864 		 * This is zio_write_phys().
2865 		 * We're either generating a label checksum, or none at all.
2866 		 */
2867 		checksum = zio->io_prop.zp_checksum;
2868 
2869 		if (checksum == ZIO_CHECKSUM_OFF)
2870 			return (ZIO_PIPELINE_CONTINUE);
2871 
2872 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2873 	} else {
2874 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2875 			ASSERT(!IO_IS_ALLOCATING(zio));
2876 			checksum = ZIO_CHECKSUM_GANG_HEADER;
2877 		} else {
2878 			checksum = BP_GET_CHECKSUM(bp);
2879 		}
2880 	}
2881 
2882 	zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2883 
2884 	return (ZIO_PIPELINE_CONTINUE);
2885 }
2886 
2887 static int
2888 zio_checksum_verify(zio_t *zio)
2889 {
2890 	zio_bad_cksum_t info;
2891 	blkptr_t *bp = zio->io_bp;
2892 	int error;
2893 
2894 	ASSERT(zio->io_vd != NULL);
2895 
2896 	if (bp == NULL) {
2897 		/*
2898 		 * This is zio_read_phys().
2899 		 * We're either verifying a label checksum, or nothing at all.
2900 		 */
2901 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2902 			return (ZIO_PIPELINE_CONTINUE);
2903 
2904 		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2905 	}
2906 
2907 	if ((error = zio_checksum_error(zio, &info)) != 0) {
2908 		zio->io_error = error;
2909 		if (error == ECKSUM &&
2910 		    !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2911 			zfs_ereport_start_checksum(zio->io_spa,
2912 			    zio->io_vd, zio, zio->io_offset,
2913 			    zio->io_size, NULL, &info);
2914 		}
2915 	}
2916 
2917 	return (ZIO_PIPELINE_CONTINUE);
2918 }
2919 
2920 /*
2921  * Called by RAID-Z to ensure we don't compute the checksum twice.
2922  */
2923 void
2924 zio_checksum_verified(zio_t *zio)
2925 {
2926 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2927 }
2928 
2929 /*
2930  * ==========================================================================
2931  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2932  * An error of 0 indicates success.  ENXIO indicates whole-device failure,
2933  * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
2934  * indicate errors that are specific to one I/O, and most likely permanent.
2935  * Any other error is presumed to be worse because we weren't expecting it.
2936  * ==========================================================================
2937  */
2938 int
2939 zio_worst_error(int e1, int e2)
2940 {
2941 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2942 	int r1, r2;
2943 
2944 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2945 		if (e1 == zio_error_rank[r1])
2946 			break;
2947 
2948 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2949 		if (e2 == zio_error_rank[r2])
2950 			break;
2951 
2952 	return (r1 > r2 ? e1 : e2);
2953 }
2954 
2955 /*
2956  * ==========================================================================
2957  * I/O completion
2958  * ==========================================================================
2959  */
2960 static int
2961 zio_ready(zio_t *zio)
2962 {
2963 	blkptr_t *bp = zio->io_bp;
2964 	zio_t *pio, *pio_next;
2965 
2966 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2967 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2968 		return (ZIO_PIPELINE_STOP);
2969 
2970 	if (zio->io_ready) {
2971 		ASSERT(IO_IS_ALLOCATING(zio));
2972 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2973 		    (zio->io_flags & ZIO_FLAG_NOPWRITE));
2974 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2975 
2976 		zio->io_ready(zio);
2977 	}
2978 
2979 	if (bp != NULL && bp != &zio->io_bp_copy)
2980 		zio->io_bp_copy = *bp;
2981 
2982 	if (zio->io_error)
2983 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2984 
2985 	mutex_enter(&zio->io_lock);
2986 	zio->io_state[ZIO_WAIT_READY] = 1;
2987 	pio = zio_walk_parents(zio);
2988 	mutex_exit(&zio->io_lock);
2989 
2990 	/*
2991 	 * As we notify zio's parents, new parents could be added.
2992 	 * New parents go to the head of zio's io_parent_list, however,
2993 	 * so we will (correctly) not notify them.  The remainder of zio's
2994 	 * io_parent_list, from 'pio_next' onward, cannot change because
2995 	 * all parents must wait for us to be done before they can be done.
2996 	 */
2997 	for (; pio != NULL; pio = pio_next) {
2998 		pio_next = zio_walk_parents(zio);
2999 		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3000 	}
3001 
3002 	if (zio->io_flags & ZIO_FLAG_NODATA) {
3003 		if (BP_IS_GANG(bp)) {
3004 			zio->io_flags &= ~ZIO_FLAG_NODATA;
3005 		} else {
3006 			ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3007 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3008 		}
3009 	}
3010 
3011 	if (zio_injection_enabled &&
3012 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
3013 		zio_handle_ignored_writes(zio);
3014 
3015 	return (ZIO_PIPELINE_CONTINUE);
3016 }
3017 
3018 static int
3019 zio_done(zio_t *zio)
3020 {
3021 	spa_t *spa = zio->io_spa;
3022 	zio_t *lio = zio->io_logical;
3023 	blkptr_t *bp = zio->io_bp;
3024 	vdev_t *vd = zio->io_vd;
3025 	uint64_t psize = zio->io_size;
3026 	zio_t *pio, *pio_next;
3027 
3028 	/*
3029 	 * If our children haven't all completed,
3030 	 * wait for them and then repeat this pipeline stage.
3031 	 */
3032 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3033 	    zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3034 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3035 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3036 		return (ZIO_PIPELINE_STOP);
3037 
3038 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3039 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3040 			ASSERT(zio->io_children[c][w] == 0);
3041 
3042 	if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3043 		ASSERT(bp->blk_pad[0] == 0);
3044 		ASSERT(bp->blk_pad[1] == 0);
3045 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3046 		    (bp == zio_unique_parent(zio)->io_bp));
3047 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3048 		    zio->io_bp_override == NULL &&
3049 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3050 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
3051 			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3052 			ASSERT(BP_COUNT_GANG(bp) == 0 ||
3053 			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3054 		}
3055 		if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3056 			VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3057 	}
3058 
3059 	/*
3060 	 * If there were child vdev/gang/ddt errors, they apply to us now.
3061 	 */
3062 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3063 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3064 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3065 
3066 	/*
3067 	 * If the I/O on the transformed data was successful, generate any
3068 	 * checksum reports now while we still have the transformed data.
3069 	 */
3070 	if (zio->io_error == 0) {
3071 		while (zio->io_cksum_report != NULL) {
3072 			zio_cksum_report_t *zcr = zio->io_cksum_report;
3073 			uint64_t align = zcr->zcr_align;
3074 			uint64_t asize = P2ROUNDUP(psize, align);
3075 			char *abuf = zio->io_data;
3076 
3077 			if (asize != psize) {
3078 				abuf = zio_buf_alloc(asize);
3079 				bcopy(zio->io_data, abuf, psize);
3080 				bzero(abuf + psize, asize - psize);
3081 			}
3082 
3083 			zio->io_cksum_report = zcr->zcr_next;
3084 			zcr->zcr_next = NULL;
3085 			zcr->zcr_finish(zcr, abuf);
3086 			zfs_ereport_free_checksum(zcr);
3087 
3088 			if (asize != psize)
3089 				zio_buf_free(abuf, asize);
3090 		}
3091 	}
3092 
3093 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
3094 
3095 	vdev_stat_update(zio, psize);
3096 
3097 	if (zio->io_error) {
3098 		/*
3099 		 * If this I/O is attached to a particular vdev,
3100 		 * generate an error message describing the I/O failure
3101 		 * at the block level.  We ignore these errors if the
3102 		 * device is currently unavailable.
3103 		 */
3104 		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3105 			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3106 
3107 		if ((zio->io_error == EIO || !(zio->io_flags &
3108 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3109 		    zio == lio) {
3110 			/*
3111 			 * For logical I/O requests, tell the SPA to log the
3112 			 * error and generate a logical data ereport.
3113 			 */
3114 			spa_log_error(spa, zio);
3115 			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3116 			    0, 0);
3117 		}
3118 	}
3119 
3120 	if (zio->io_error && zio == lio) {
3121 		/*
3122 		 * Determine whether zio should be reexecuted.  This will
3123 		 * propagate all the way to the root via zio_notify_parent().
3124 		 */
3125 		ASSERT(vd == NULL && bp != NULL);
3126 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3127 
3128 		if (IO_IS_ALLOCATING(zio) &&
3129 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3130 			if (zio->io_error != ENOSPC)
3131 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3132 			else
3133 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3134 		}
3135 
3136 		if ((zio->io_type == ZIO_TYPE_READ ||
3137 		    zio->io_type == ZIO_TYPE_FREE) &&
3138 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3139 		    zio->io_error == ENXIO &&
3140 		    spa_load_state(spa) == SPA_LOAD_NONE &&
3141 		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3142 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3143 
3144 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3145 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3146 
3147 		/*
3148 		 * Here is a possibly good place to attempt to do
3149 		 * either combinatorial reconstruction or error correction
3150 		 * based on checksums.  It also might be a good place
3151 		 * to send out preliminary ereports before we suspend
3152 		 * processing.
3153 		 */
3154 	}
3155 
3156 	/*
3157 	 * If there were logical child errors, they apply to us now.
3158 	 * We defer this until now to avoid conflating logical child
3159 	 * errors with errors that happened to the zio itself when
3160 	 * updating vdev stats and reporting FMA events above.
3161 	 */
3162 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3163 
3164 	if ((zio->io_error || zio->io_reexecute) &&
3165 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3166 	    !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3167 		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3168 
3169 	zio_gang_tree_free(&zio->io_gang_tree);
3170 
3171 	/*
3172 	 * Godfather I/Os should never suspend.
3173 	 */
3174 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3175 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3176 		zio->io_reexecute = 0;
3177 
3178 	if (zio->io_reexecute) {
3179 		/*
3180 		 * This is a logical I/O that wants to reexecute.
3181 		 *
3182 		 * Reexecute is top-down.  When an i/o fails, if it's not
3183 		 * the root, it simply notifies its parent and sticks around.
3184 		 * The parent, seeing that it still has children in zio_done(),
3185 		 * does the same.  This percolates all the way up to the root.
3186 		 * The root i/o will reexecute or suspend the entire tree.
3187 		 *
3188 		 * This approach ensures that zio_reexecute() honors
3189 		 * all the original i/o dependency relationships, e.g.
3190 		 * parents not executing until children are ready.
3191 		 */
3192 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3193 
3194 		zio->io_gang_leader = NULL;
3195 
3196 		mutex_enter(&zio->io_lock);
3197 		zio->io_state[ZIO_WAIT_DONE] = 1;
3198 		mutex_exit(&zio->io_lock);
3199 
3200 		/*
3201 		 * "The Godfather" I/O monitors its children but is
3202 		 * not a true parent to them. It will track them through
3203 		 * the pipeline but severs its ties whenever they get into
3204 		 * trouble (e.g. suspended). This allows "The Godfather"
3205 		 * I/O to return status without blocking.
3206 		 */
3207 		for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3208 			zio_link_t *zl = zio->io_walk_link;
3209 			pio_next = zio_walk_parents(zio);
3210 
3211 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3212 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3213 				zio_remove_child(pio, zio, zl);
3214 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3215 			}
3216 		}
3217 
3218 		if ((pio = zio_unique_parent(zio)) != NULL) {
3219 			/*
3220 			 * We're not a root i/o, so there's nothing to do
3221 			 * but notify our parent.  Don't propagate errors
3222 			 * upward since we haven't permanently failed yet.
3223 			 */
3224 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3225 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3226 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3227 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3228 			/*
3229 			 * We'd fail again if we reexecuted now, so suspend
3230 			 * until conditions improve (e.g. device comes online).
3231 			 */
3232 			zio_suspend(spa, zio);
3233 		} else {
3234 			/*
3235 			 * Reexecution is potentially a huge amount of work.
3236 			 * Hand it off to the otherwise-unused claim taskq.
3237 			 */
3238 			ASSERT(zio->io_tqent.tqent_next == NULL);
3239 			spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3240 			    ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3241 			    0, &zio->io_tqent);
3242 		}
3243 		return (ZIO_PIPELINE_STOP);
3244 	}
3245 
3246 	ASSERT(zio->io_child_count == 0);
3247 	ASSERT(zio->io_reexecute == 0);
3248 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3249 
3250 	/*
3251 	 * Report any checksum errors, since the I/O is complete.
3252 	 */
3253 	while (zio->io_cksum_report != NULL) {
3254 		zio_cksum_report_t *zcr = zio->io_cksum_report;
3255 		zio->io_cksum_report = zcr->zcr_next;
3256 		zcr->zcr_next = NULL;
3257 		zcr->zcr_finish(zcr, NULL);
3258 		zfs_ereport_free_checksum(zcr);
3259 	}
3260 
3261 	/*
3262 	 * It is the responsibility of the done callback to ensure that this
3263 	 * particular zio is no longer discoverable for adoption, and as
3264 	 * such, cannot acquire any new parents.
3265 	 */
3266 	if (zio->io_done)
3267 		zio->io_done(zio);
3268 
3269 	mutex_enter(&zio->io_lock);
3270 	zio->io_state[ZIO_WAIT_DONE] = 1;
3271 	mutex_exit(&zio->io_lock);
3272 
3273 	for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3274 		zio_link_t *zl = zio->io_walk_link;
3275 		pio_next = zio_walk_parents(zio);
3276 		zio_remove_child(pio, zio, zl);
3277 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3278 	}
3279 
3280 	if (zio->io_waiter != NULL) {
3281 		mutex_enter(&zio->io_lock);
3282 		zio->io_executor = NULL;
3283 		cv_broadcast(&zio->io_cv);
3284 		mutex_exit(&zio->io_lock);
3285 	} else {
3286 		zio_destroy(zio);
3287 	}
3288 
3289 	return (ZIO_PIPELINE_STOP);
3290 }
3291 
3292 /*
3293  * ==========================================================================
3294  * I/O pipeline definition
3295  * ==========================================================================
3296  */
3297 static zio_pipe_stage_t *zio_pipeline[] = {
3298 	NULL,
3299 	zio_read_bp_init,
3300 	zio_free_bp_init,
3301 	zio_issue_async,
3302 	zio_write_bp_init,
3303 	zio_checksum_generate,
3304 	zio_nop_write,
3305 	zio_ddt_read_start,
3306 	zio_ddt_read_done,
3307 	zio_ddt_write,
3308 	zio_ddt_free,
3309 	zio_gang_assemble,
3310 	zio_gang_issue,
3311 	zio_dva_allocate,
3312 	zio_dva_free,
3313 	zio_dva_claim,
3314 	zio_ready,
3315 	zio_vdev_io_start,
3316 	zio_vdev_io_done,
3317 	zio_vdev_io_assess,
3318 	zio_checksum_verify,
3319 	zio_done
3320 };
3321 
3322 
3323 
3324 
3325 /*
3326  * Compare two zbookmark_phys_t's to see which we would reach first in a
3327  * pre-order traversal of the object tree.
3328  *
3329  * This is simple in every case aside from the meta-dnode object. For all other
3330  * objects, we traverse them in order (object 1 before object 2, and so on).
3331  * However, all of these objects are traversed while traversing object 0, since
3332  * the data it points to is the list of objects.  Thus, we need to convert to a
3333  * canonical representation so we can compare meta-dnode bookmarks to
3334  * non-meta-dnode bookmarks.
3335  *
3336  * We do this by calculating "equivalents" for each field of the zbookmark.
3337  * zbookmarks outside of the meta-dnode use their own object and level, and
3338  * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3339  * blocks this bookmark refers to) by multiplying their blkid by their span
3340  * (the number of L0 blocks contained within one block at their level).
3341  * zbookmarks inside the meta-dnode calculate their object equivalent
3342  * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3343  * level + 1<<31 (any value larger than a level could ever be) for their level.
3344  * This causes them to always compare before a bookmark in their object
3345  * equivalent, compare appropriately to bookmarks in other objects, and to
3346  * compare appropriately to other bookmarks in the meta-dnode.
3347  */
3348 int
3349 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
3350     const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
3351 {
3352 	/*
3353 	 * These variables represent the "equivalent" values for the zbookmark,
3354 	 * after converting zbookmarks inside the meta dnode to their
3355 	 * normal-object equivalents.
3356 	 */
3357 	uint64_t zb1obj, zb2obj;
3358 	uint64_t zb1L0, zb2L0;
3359 	uint64_t zb1level, zb2level;
3360 
3361 	if (zb1->zb_object == zb2->zb_object &&
3362 	    zb1->zb_level == zb2->zb_level &&
3363 	    zb1->zb_blkid == zb2->zb_blkid)
3364 		return (0);
3365 
3366 	/*
3367 	 * BP_SPANB calculates the span in blocks.
3368 	 */
3369 	zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
3370 	zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
3371 
3372 	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3373 		zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3374 		zb1L0 = 0;
3375 		zb1level = zb1->zb_level + COMPARE_META_LEVEL;
3376 	} else {
3377 		zb1obj = zb1->zb_object;
3378 		zb1level = zb1->zb_level;
3379 	}
3380 
3381 	if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
3382 		zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3383 		zb2L0 = 0;
3384 		zb2level = zb2->zb_level + COMPARE_META_LEVEL;
3385 	} else {
3386 		zb2obj = zb2->zb_object;
3387 		zb2level = zb2->zb_level;
3388 	}
3389 
3390 	/* Now that we have a canonical representation, do the comparison. */
3391 	if (zb1obj != zb2obj)
3392 		return (zb1obj < zb2obj ? -1 : 1);
3393 	else if (zb1L0 != zb2L0)
3394 		return (zb1L0 < zb2L0 ? -1 : 1);
3395 	else if (zb1level != zb2level)
3396 		return (zb1level > zb2level ? -1 : 1);
3397 	/*
3398 	 * This can (theoretically) happen if the bookmarks have the same object
3399 	 * and level, but different blkids, if the block sizes are not the same.
3400 	 * There is presently no way to change the indirect block sizes
3401 	 */
3402 	return (0);
3403 }
3404 
3405 /*
3406  *  This function checks the following: given that last_block is the place that
3407  *  our traversal stopped last time, does that guarantee that we've visited
3408  *  every node under subtree_root?  Therefore, we can't just use the raw output
3409  *  of zbookmark_compare.  We have to pass in a modified version of
3410  *  subtree_root; by incrementing the block id, and then checking whether
3411  *  last_block is before or equal to that, we can tell whether or not having
3412  *  visited last_block implies that all of subtree_root's children have been
3413  *  visited.
3414  */
3415 boolean_t
3416 zbookmark_subtree_completed(const dnode_phys_t *dnp,
3417     const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
3418 {
3419 	zbookmark_phys_t mod_zb = *subtree_root;
3420 	mod_zb.zb_blkid++;
3421 	ASSERT(last_block->zb_level == 0);
3422 
3423 	/* The objset_phys_t isn't before anything. */
3424 	if (dnp == NULL)
3425 		return (B_FALSE);
3426 
3427 	/*
3428 	 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
3429 	 * data block size in sectors, because that variable is only used if
3430 	 * the bookmark refers to a block in the meta-dnode.  Since we don't
3431 	 * know without examining it what object it refers to, and there's no
3432 	 * harm in passing in this value in other cases, we always pass it in.
3433 	 *
3434 	 * We pass in 0 for the indirect block size shift because zb2 must be
3435 	 * level 0.  The indirect block size is only used to calculate the span
3436 	 * of the bookmark, but since the bookmark must be level 0, the span is
3437 	 * always 1, so the math works out.
3438 	 *
3439 	 * If you make changes to how the zbookmark_compare code works, be sure
3440 	 * to make sure that this code still works afterwards.
3441 	 */
3442 	return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
3443 	    1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
3444 	    last_block) <= 0);
3445 }
3446