xref: /titanic_51/usr/src/uts/common/fs/zfs/zio.c (revision 90b1de135fcfa7ce4adc9138a885aa94bbcef04f)
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, 2014 by Delphix. All rights reserved.
24  * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25  */
26 
27 #include <sys/sysmacros.h>
28 #include <sys/zfs_context.h>
29 #include <sys/fm/fs/zfs.h>
30 #include <sys/spa.h>
31 #include <sys/txg.h>
32 #include <sys/spa_impl.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio_impl.h>
35 #include <sys/zio_compress.h>
36 #include <sys/zio_checksum.h>
37 #include <sys/dmu_objset.h>
38 #include <sys/arc.h>
39 #include <sys/ddt.h>
40 #include <sys/blkptr.h>
41 #include <sys/zfeature.h>
42 
43 /*
44  * ==========================================================================
45  * I/O type descriptions
46  * ==========================================================================
47  */
48 const char *zio_type_name[ZIO_TYPES] = {
49 	"zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
50 	"zio_ioctl"
51 };
52 
53 /*
54  * ==========================================================================
55  * I/O kmem caches
56  * ==========================================================================
57  */
58 kmem_cache_t *zio_cache;
59 kmem_cache_t *zio_link_cache;
60 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
61 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
62 
63 #ifdef _KERNEL
64 extern vmem_t *zio_alloc_arena;
65 #endif
66 
67 #define	ZIO_PIPELINE_CONTINUE		0x100
68 #define	ZIO_PIPELINE_STOP		0x101
69 
70 /*
71  * The following actions directly effect the spa's sync-to-convergence logic.
72  * The values below define the sync pass when we start performing the action.
73  * Care should be taken when changing these values as they directly impact
74  * spa_sync() performance. Tuning these values may introduce subtle performance
75  * pathologies and should only be done in the context of performance analysis.
76  * These tunables will eventually be removed and replaced with #defines once
77  * enough analysis has been done to determine optimal values.
78  *
79  * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
80  * regular blocks are not deferred.
81  */
82 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
83 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
84 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
85 
86 /*
87  * An allocating zio is one that either currently has the DVA allocate
88  * stage set or will have it later in its lifetime.
89  */
90 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
91 
92 boolean_t	zio_requeue_io_start_cut_in_line = B_TRUE;
93 
94 #ifdef ZFS_DEBUG
95 int zio_buf_debug_limit = 16384;
96 #else
97 int zio_buf_debug_limit = 0;
98 #endif
99 
100 void
101 zio_init(void)
102 {
103 	size_t c;
104 	vmem_t *data_alloc_arena = NULL;
105 
106 #ifdef _KERNEL
107 	data_alloc_arena = zio_alloc_arena;
108 #endif
109 	zio_cache = kmem_cache_create("zio_cache",
110 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
111 	zio_link_cache = kmem_cache_create("zio_link_cache",
112 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
113 
114 	/*
115 	 * For small buffers, we want a cache for each multiple of
116 	 * SPA_MINBLOCKSIZE.  For larger buffers, we want a cache
117 	 * for each quarter-power of 2.
118 	 */
119 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
120 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
121 		size_t p2 = size;
122 		size_t align = 0;
123 		size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
124 
125 		while (!ISP2(p2))
126 			p2 &= p2 - 1;
127 
128 #ifndef _KERNEL
129 		/*
130 		 * If we are using watchpoints, put each buffer on its own page,
131 		 * to eliminate the performance overhead of trapping to the
132 		 * kernel when modifying a non-watched buffer that shares the
133 		 * page with a watched buffer.
134 		 */
135 		if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
136 			continue;
137 #endif
138 		if (size <= 4 * SPA_MINBLOCKSIZE) {
139 			align = SPA_MINBLOCKSIZE;
140 		} else if (IS_P2ALIGNED(size, p2 >> 2)) {
141 			align = MIN(p2 >> 2, PAGESIZE);
142 		}
143 
144 		if (align != 0) {
145 			char name[36];
146 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
147 			zio_buf_cache[c] = kmem_cache_create(name, size,
148 			    align, NULL, NULL, NULL, NULL, NULL, cflags);
149 
150 			/*
151 			 * Since zio_data bufs do not appear in crash dumps, we
152 			 * pass KMC_NOTOUCH so that no allocator metadata is
153 			 * stored with the buffers.
154 			 */
155 			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
156 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
157 			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
158 			    cflags | KMC_NOTOUCH);
159 		}
160 	}
161 
162 	while (--c != 0) {
163 		ASSERT(zio_buf_cache[c] != NULL);
164 		if (zio_buf_cache[c - 1] == NULL)
165 			zio_buf_cache[c - 1] = zio_buf_cache[c];
166 
167 		ASSERT(zio_data_buf_cache[c] != NULL);
168 		if (zio_data_buf_cache[c - 1] == NULL)
169 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
170 	}
171 
172 	zio_inject_init();
173 }
174 
175 void
176 zio_fini(void)
177 {
178 	size_t c;
179 	kmem_cache_t *last_cache = NULL;
180 	kmem_cache_t *last_data_cache = NULL;
181 
182 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
183 		if (zio_buf_cache[c] != last_cache) {
184 			last_cache = zio_buf_cache[c];
185 			kmem_cache_destroy(zio_buf_cache[c]);
186 		}
187 		zio_buf_cache[c] = NULL;
188 
189 		if (zio_data_buf_cache[c] != last_data_cache) {
190 			last_data_cache = zio_data_buf_cache[c];
191 			kmem_cache_destroy(zio_data_buf_cache[c]);
192 		}
193 		zio_data_buf_cache[c] = NULL;
194 	}
195 
196 	kmem_cache_destroy(zio_link_cache);
197 	kmem_cache_destroy(zio_cache);
198 
199 	zio_inject_fini();
200 }
201 
202 /*
203  * ==========================================================================
204  * Allocate and free I/O buffers
205  * ==========================================================================
206  */
207 
208 /*
209  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
210  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
211  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
212  * excess / transient data in-core during a crashdump.
213  */
214 void *
215 zio_buf_alloc(size_t size)
216 {
217 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
218 
219 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
220 
221 	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
222 }
223 
224 /*
225  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
226  * crashdump if the kernel panics.  This exists so that we will limit the amount
227  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
228  * of kernel heap dumped to disk when the kernel panics)
229  */
230 void *
231 zio_data_buf_alloc(size_t size)
232 {
233 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
234 
235 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
236 
237 	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
238 }
239 
240 void
241 zio_buf_free(void *buf, size_t size)
242 {
243 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
244 
245 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
246 
247 	kmem_cache_free(zio_buf_cache[c], buf);
248 }
249 
250 void
251 zio_data_buf_free(void *buf, size_t size)
252 {
253 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
254 
255 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
256 
257 	kmem_cache_free(zio_data_buf_cache[c], buf);
258 }
259 
260 /*
261  * ==========================================================================
262  * Push and pop I/O transform buffers
263  * ==========================================================================
264  */
265 static void
266 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
267 	zio_transform_func_t *transform)
268 {
269 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
270 
271 	zt->zt_orig_data = zio->io_data;
272 	zt->zt_orig_size = zio->io_size;
273 	zt->zt_bufsize = bufsize;
274 	zt->zt_transform = transform;
275 
276 	zt->zt_next = zio->io_transform_stack;
277 	zio->io_transform_stack = zt;
278 
279 	zio->io_data = data;
280 	zio->io_size = size;
281 }
282 
283 static void
284 zio_pop_transforms(zio_t *zio)
285 {
286 	zio_transform_t *zt;
287 
288 	while ((zt = zio->io_transform_stack) != NULL) {
289 		if (zt->zt_transform != NULL)
290 			zt->zt_transform(zio,
291 			    zt->zt_orig_data, zt->zt_orig_size);
292 
293 		if (zt->zt_bufsize != 0)
294 			zio_buf_free(zio->io_data, zt->zt_bufsize);
295 
296 		zio->io_data = zt->zt_orig_data;
297 		zio->io_size = zt->zt_orig_size;
298 		zio->io_transform_stack = zt->zt_next;
299 
300 		kmem_free(zt, sizeof (zio_transform_t));
301 	}
302 }
303 
304 /*
305  * ==========================================================================
306  * I/O transform callbacks for subblocks and decompression
307  * ==========================================================================
308  */
309 static void
310 zio_subblock(zio_t *zio, void *data, uint64_t size)
311 {
312 	ASSERT(zio->io_size > size);
313 
314 	if (zio->io_type == ZIO_TYPE_READ)
315 		bcopy(zio->io_data, data, size);
316 }
317 
318 static void
319 zio_decompress(zio_t *zio, void *data, uint64_t size)
320 {
321 	if (zio->io_error == 0 &&
322 	    zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
323 	    zio->io_data, data, zio->io_size, size) != 0)
324 		zio->io_error = SET_ERROR(EIO);
325 }
326 
327 /*
328  * ==========================================================================
329  * I/O parent/child relationships and pipeline interlocks
330  * ==========================================================================
331  */
332 /*
333  * NOTE - Callers to zio_walk_parents() and zio_walk_children must
334  *        continue calling these functions until they return NULL.
335  *        Otherwise, the next caller will pick up the list walk in
336  *        some indeterminate state.  (Otherwise every caller would
337  *        have to pass in a cookie to keep the state represented by
338  *        io_walk_link, which gets annoying.)
339  */
340 zio_t *
341 zio_walk_parents(zio_t *cio)
342 {
343 	zio_link_t *zl = cio->io_walk_link;
344 	list_t *pl = &cio->io_parent_list;
345 
346 	zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
347 	cio->io_walk_link = zl;
348 
349 	if (zl == NULL)
350 		return (NULL);
351 
352 	ASSERT(zl->zl_child == cio);
353 	return (zl->zl_parent);
354 }
355 
356 zio_t *
357 zio_walk_children(zio_t *pio)
358 {
359 	zio_link_t *zl = pio->io_walk_link;
360 	list_t *cl = &pio->io_child_list;
361 
362 	zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
363 	pio->io_walk_link = zl;
364 
365 	if (zl == NULL)
366 		return (NULL);
367 
368 	ASSERT(zl->zl_parent == pio);
369 	return (zl->zl_child);
370 }
371 
372 zio_t *
373 zio_unique_parent(zio_t *cio)
374 {
375 	zio_t *pio = zio_walk_parents(cio);
376 
377 	VERIFY(zio_walk_parents(cio) == NULL);
378 	return (pio);
379 }
380 
381 void
382 zio_add_child(zio_t *pio, zio_t *cio)
383 {
384 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
385 
386 	/*
387 	 * Logical I/Os can have logical, gang, or vdev children.
388 	 * Gang I/Os can have gang or vdev children.
389 	 * Vdev I/Os can only have vdev children.
390 	 * The following ASSERT captures all of these constraints.
391 	 */
392 	ASSERT(cio->io_child_type <= pio->io_child_type);
393 
394 	zl->zl_parent = pio;
395 	zl->zl_child = cio;
396 
397 	mutex_enter(&cio->io_lock);
398 	mutex_enter(&pio->io_lock);
399 
400 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
401 
402 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
403 		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
404 
405 	list_insert_head(&pio->io_child_list, zl);
406 	list_insert_head(&cio->io_parent_list, zl);
407 
408 	pio->io_child_count++;
409 	cio->io_parent_count++;
410 
411 	mutex_exit(&pio->io_lock);
412 	mutex_exit(&cio->io_lock);
413 }
414 
415 static void
416 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
417 {
418 	ASSERT(zl->zl_parent == pio);
419 	ASSERT(zl->zl_child == cio);
420 
421 	mutex_enter(&cio->io_lock);
422 	mutex_enter(&pio->io_lock);
423 
424 	list_remove(&pio->io_child_list, zl);
425 	list_remove(&cio->io_parent_list, zl);
426 
427 	pio->io_child_count--;
428 	cio->io_parent_count--;
429 
430 	mutex_exit(&pio->io_lock);
431 	mutex_exit(&cio->io_lock);
432 
433 	kmem_cache_free(zio_link_cache, zl);
434 }
435 
436 static boolean_t
437 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
438 {
439 	uint64_t *countp = &zio->io_children[child][wait];
440 	boolean_t waiting = B_FALSE;
441 
442 	mutex_enter(&zio->io_lock);
443 	ASSERT(zio->io_stall == NULL);
444 	if (*countp != 0) {
445 		zio->io_stage >>= 1;
446 		zio->io_stall = countp;
447 		waiting = B_TRUE;
448 	}
449 	mutex_exit(&zio->io_lock);
450 
451 	return (waiting);
452 }
453 
454 static void
455 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
456 {
457 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
458 	int *errorp = &pio->io_child_error[zio->io_child_type];
459 
460 	mutex_enter(&pio->io_lock);
461 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
462 		*errorp = zio_worst_error(*errorp, zio->io_error);
463 	pio->io_reexecute |= zio->io_reexecute;
464 	ASSERT3U(*countp, >, 0);
465 
466 	(*countp)--;
467 
468 	if (*countp == 0 && pio->io_stall == countp) {
469 		pio->io_stall = NULL;
470 		mutex_exit(&pio->io_lock);
471 		zio_execute(pio);
472 	} else {
473 		mutex_exit(&pio->io_lock);
474 	}
475 }
476 
477 static void
478 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
479 {
480 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
481 		zio->io_error = zio->io_child_error[c];
482 }
483 
484 /*
485  * ==========================================================================
486  * Create the various types of I/O (read, write, free, etc)
487  * ==========================================================================
488  */
489 static zio_t *
490 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
491     void *data, uint64_t size, zio_done_func_t *done, void *private,
492     zio_type_t type, zio_priority_t priority, enum zio_flag flags,
493     vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
494     enum zio_stage stage, enum zio_stage pipeline)
495 {
496 	zio_t *zio;
497 
498 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
499 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
500 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
501 
502 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
503 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
504 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
505 
506 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
507 	bzero(zio, sizeof (zio_t));
508 
509 	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
510 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
511 
512 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
513 	    offsetof(zio_link_t, zl_parent_node));
514 	list_create(&zio->io_child_list, sizeof (zio_link_t),
515 	    offsetof(zio_link_t, zl_child_node));
516 
517 	if (vd != NULL)
518 		zio->io_child_type = ZIO_CHILD_VDEV;
519 	else if (flags & ZIO_FLAG_GANG_CHILD)
520 		zio->io_child_type = ZIO_CHILD_GANG;
521 	else if (flags & ZIO_FLAG_DDT_CHILD)
522 		zio->io_child_type = ZIO_CHILD_DDT;
523 	else
524 		zio->io_child_type = ZIO_CHILD_LOGICAL;
525 
526 	if (bp != NULL) {
527 		zio->io_bp = (blkptr_t *)bp;
528 		zio->io_bp_copy = *bp;
529 		zio->io_bp_orig = *bp;
530 		if (type != ZIO_TYPE_WRITE ||
531 		    zio->io_child_type == ZIO_CHILD_DDT)
532 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
533 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
534 			zio->io_logical = zio;
535 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
536 			pipeline |= ZIO_GANG_STAGES;
537 	}
538 
539 	zio->io_spa = spa;
540 	zio->io_txg = txg;
541 	zio->io_done = done;
542 	zio->io_private = private;
543 	zio->io_type = type;
544 	zio->io_priority = priority;
545 	zio->io_vd = vd;
546 	zio->io_offset = offset;
547 	zio->io_orig_data = zio->io_data = data;
548 	zio->io_orig_size = zio->io_size = size;
549 	zio->io_orig_flags = zio->io_flags = flags;
550 	zio->io_orig_stage = zio->io_stage = stage;
551 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
552 
553 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
554 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
555 
556 	if (zb != NULL)
557 		zio->io_bookmark = *zb;
558 
559 	if (pio != NULL) {
560 		if (zio->io_logical == NULL)
561 			zio->io_logical = pio->io_logical;
562 		if (zio->io_child_type == ZIO_CHILD_GANG)
563 			zio->io_gang_leader = pio->io_gang_leader;
564 		zio_add_child(pio, zio);
565 	}
566 
567 	return (zio);
568 }
569 
570 static void
571 zio_destroy(zio_t *zio)
572 {
573 	list_destroy(&zio->io_parent_list);
574 	list_destroy(&zio->io_child_list);
575 	mutex_destroy(&zio->io_lock);
576 	cv_destroy(&zio->io_cv);
577 	kmem_cache_free(zio_cache, zio);
578 }
579 
580 zio_t *
581 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
582     void *private, enum zio_flag flags)
583 {
584 	zio_t *zio;
585 
586 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
587 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
588 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
589 
590 	return (zio);
591 }
592 
593 zio_t *
594 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
595 {
596 	return (zio_null(NULL, spa, NULL, done, private, flags));
597 }
598 
599 void
600 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
601 {
602 	if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
603 		zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
604 		    bp, (longlong_t)BP_GET_TYPE(bp));
605 	}
606 	if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
607 	    BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
608 		zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
609 		    bp, (longlong_t)BP_GET_CHECKSUM(bp));
610 	}
611 	if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
612 	    BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
613 		zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
614 		    bp, (longlong_t)BP_GET_COMPRESS(bp));
615 	}
616 	if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
617 		zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
618 		    bp, (longlong_t)BP_GET_LSIZE(bp));
619 	}
620 	if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
621 		zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
622 		    bp, (longlong_t)BP_GET_PSIZE(bp));
623 	}
624 
625 	if (BP_IS_EMBEDDED(bp)) {
626 		if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
627 			zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
628 			    bp, (longlong_t)BPE_GET_ETYPE(bp));
629 		}
630 	}
631 
632 	/*
633 	 * Pool-specific checks.
634 	 *
635 	 * Note: it would be nice to verify that the blk_birth and
636 	 * BP_PHYSICAL_BIRTH() are not too large.  However, spa_freeze()
637 	 * allows the birth time of log blocks (and dmu_sync()-ed blocks
638 	 * that are in the log) to be arbitrarily large.
639 	 */
640 	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
641 		uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
642 		if (vdevid >= spa->spa_root_vdev->vdev_children) {
643 			zfs_panic_recover("blkptr at %p DVA %u has invalid "
644 			    "VDEV %llu",
645 			    bp, i, (longlong_t)vdevid);
646 		}
647 		vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
648 		if (vd == NULL) {
649 			zfs_panic_recover("blkptr at %p DVA %u has invalid "
650 			    "VDEV %llu",
651 			    bp, i, (longlong_t)vdevid);
652 		}
653 		if (vd->vdev_ops == &vdev_hole_ops) {
654 			zfs_panic_recover("blkptr at %p DVA %u has hole "
655 			    "VDEV %llu",
656 			    bp, i, (longlong_t)vdevid);
657 
658 		}
659 		if (vd->vdev_ops == &vdev_missing_ops) {
660 			/*
661 			 * "missing" vdevs are valid during import, but we
662 			 * don't have their detailed info (e.g. asize), so
663 			 * we can't perform any more checks on them.
664 			 */
665 			continue;
666 		}
667 		uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
668 		uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
669 		if (BP_IS_GANG(bp))
670 			asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
671 		if (offset + asize > vd->vdev_asize) {
672 			zfs_panic_recover("blkptr at %p DVA %u has invalid "
673 			    "OFFSET %llu",
674 			    bp, i, (longlong_t)offset);
675 		}
676 	}
677 }
678 
679 zio_t *
680 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
681     void *data, uint64_t size, zio_done_func_t *done, void *private,
682     zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
683 {
684 	zio_t *zio;
685 
686 	zfs_blkptr_verify(spa, bp);
687 
688 	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
689 	    data, size, done, private,
690 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
691 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
692 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
693 
694 	return (zio);
695 }
696 
697 zio_t *
698 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
699     void *data, uint64_t size, const zio_prop_t *zp,
700     zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
701     void *private,
702     zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
703 {
704 	zio_t *zio;
705 
706 	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
707 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
708 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
709 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
710 	    DMU_OT_IS_VALID(zp->zp_type) &&
711 	    zp->zp_level < 32 &&
712 	    zp->zp_copies > 0 &&
713 	    zp->zp_copies <= spa_max_replication(spa));
714 
715 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
716 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
717 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
718 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
719 
720 	zio->io_ready = ready;
721 	zio->io_physdone = physdone;
722 	zio->io_prop = *zp;
723 
724 	/*
725 	 * Data can be NULL if we are going to call zio_write_override() to
726 	 * provide the already-allocated BP.  But we may need the data to
727 	 * verify a dedup hit (if requested).  In this case, don't try to
728 	 * dedup (just take the already-allocated BP verbatim).
729 	 */
730 	if (data == NULL && zio->io_prop.zp_dedup_verify) {
731 		zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
732 	}
733 
734 	return (zio);
735 }
736 
737 zio_t *
738 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
739     uint64_t size, zio_done_func_t *done, void *private,
740     zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
741 {
742 	zio_t *zio;
743 
744 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
745 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
746 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
747 
748 	return (zio);
749 }
750 
751 void
752 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
753 {
754 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
755 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
756 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
757 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
758 
759 	/*
760 	 * We must reset the io_prop to match the values that existed
761 	 * when the bp was first written by dmu_sync() keeping in mind
762 	 * that nopwrite and dedup are mutually exclusive.
763 	 */
764 	zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
765 	zio->io_prop.zp_nopwrite = nopwrite;
766 	zio->io_prop.zp_copies = copies;
767 	zio->io_bp_override = bp;
768 }
769 
770 void
771 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
772 {
773 
774 	/*
775 	 * The check for EMBEDDED is a performance optimization.  We
776 	 * process the free here (by ignoring it) rather than
777 	 * putting it on the list and then processing it in zio_free_sync().
778 	 */
779 	if (BP_IS_EMBEDDED(bp))
780 		return;
781 	metaslab_check_free(spa, bp);
782 
783 	/*
784 	 * Frees that are for the currently-syncing txg, are not going to be
785 	 * deferred, and which will not need to do a read (i.e. not GANG or
786 	 * DEDUP), can be processed immediately.  Otherwise, put them on the
787 	 * in-memory list for later processing.
788 	 */
789 	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
790 	    txg != spa->spa_syncing_txg ||
791 	    spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
792 		bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
793 	} else {
794 		VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
795 	}
796 }
797 
798 zio_t *
799 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
800     enum zio_flag flags)
801 {
802 	zio_t *zio;
803 	enum zio_stage stage = ZIO_FREE_PIPELINE;
804 
805 	ASSERT(!BP_IS_HOLE(bp));
806 	ASSERT(spa_syncing_txg(spa) == txg);
807 	ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
808 
809 	if (BP_IS_EMBEDDED(bp))
810 		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
811 
812 	metaslab_check_free(spa, bp);
813 	arc_freed(spa, bp);
814 
815 	/*
816 	 * GANG and DEDUP blocks can induce a read (for the gang block header,
817 	 * or the DDT), so issue them asynchronously so that this thread is
818 	 * not tied up.
819 	 */
820 	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
821 		stage |= ZIO_STAGE_ISSUE_ASYNC;
822 
823 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
824 	    NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
825 	    NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
826 
827 	return (zio);
828 }
829 
830 zio_t *
831 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
832     zio_done_func_t *done, void *private, enum zio_flag flags)
833 {
834 	zio_t *zio;
835 
836 	dprintf_bp(bp, "claiming in txg %llu", txg);
837 
838 	if (BP_IS_EMBEDDED(bp))
839 		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
840 
841 	/*
842 	 * A claim is an allocation of a specific block.  Claims are needed
843 	 * to support immediate writes in the intent log.  The issue is that
844 	 * immediate writes contain committed data, but in a txg that was
845 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
846 	 * the intent log claims all blocks that contain immediate write data
847 	 * so that the SPA knows they're in use.
848 	 *
849 	 * All claims *must* be resolved in the first txg -- before the SPA
850 	 * starts allocating blocks -- so that nothing is allocated twice.
851 	 * If txg == 0 we just verify that the block is claimable.
852 	 */
853 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
854 	ASSERT(txg == spa_first_txg(spa) || txg == 0);
855 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
856 
857 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
858 	    done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
859 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
860 
861 	return (zio);
862 }
863 
864 zio_t *
865 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
866     zio_done_func_t *done, void *private, enum zio_flag flags)
867 {
868 	zio_t *zio;
869 	int c;
870 
871 	if (vd->vdev_children == 0) {
872 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
873 		    ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
874 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
875 
876 		zio->io_cmd = cmd;
877 	} else {
878 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
879 
880 		for (c = 0; c < vd->vdev_children; c++)
881 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
882 			    done, private, flags));
883 	}
884 
885 	return (zio);
886 }
887 
888 zio_t *
889 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
890     void *data, int checksum, zio_done_func_t *done, void *private,
891     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
892 {
893 	zio_t *zio;
894 
895 	ASSERT(vd->vdev_children == 0);
896 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
897 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
898 	ASSERT3U(offset + size, <=, vd->vdev_psize);
899 
900 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
901 	    ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
902 	    NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
903 
904 	zio->io_prop.zp_checksum = checksum;
905 
906 	return (zio);
907 }
908 
909 zio_t *
910 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
911     void *data, int checksum, zio_done_func_t *done, void *private,
912     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
913 {
914 	zio_t *zio;
915 
916 	ASSERT(vd->vdev_children == 0);
917 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
918 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
919 	ASSERT3U(offset + size, <=, vd->vdev_psize);
920 
921 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
922 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
923 	    NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
924 
925 	zio->io_prop.zp_checksum = checksum;
926 
927 	if (zio_checksum_table[checksum].ci_eck) {
928 		/*
929 		 * zec checksums are necessarily destructive -- they modify
930 		 * the end of the write buffer to hold the verifier/checksum.
931 		 * Therefore, we must make a local copy in case the data is
932 		 * being written to multiple places in parallel.
933 		 */
934 		void *wbuf = zio_buf_alloc(size);
935 		bcopy(data, wbuf, size);
936 		zio_push_transform(zio, wbuf, size, size, NULL);
937 	}
938 
939 	return (zio);
940 }
941 
942 /*
943  * Create a child I/O to do some work for us.
944  */
945 zio_t *
946 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
947 	void *data, uint64_t size, int type, zio_priority_t priority,
948 	enum zio_flag flags, zio_done_func_t *done, void *private)
949 {
950 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
951 	zio_t *zio;
952 
953 	ASSERT(vd->vdev_parent ==
954 	    (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
955 
956 	if (type == ZIO_TYPE_READ && bp != NULL) {
957 		/*
958 		 * If we have the bp, then the child should perform the
959 		 * checksum and the parent need not.  This pushes error
960 		 * detection as close to the leaves as possible and
961 		 * eliminates redundant checksums in the interior nodes.
962 		 */
963 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
964 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
965 	}
966 
967 	if (vd->vdev_children == 0)
968 		offset += VDEV_LABEL_START_SIZE;
969 
970 	flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
971 
972 	/*
973 	 * If we've decided to do a repair, the write is not speculative --
974 	 * even if the original read was.
975 	 */
976 	if (flags & ZIO_FLAG_IO_REPAIR)
977 		flags &= ~ZIO_FLAG_SPECULATIVE;
978 
979 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
980 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
981 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
982 
983 	zio->io_physdone = pio->io_physdone;
984 	if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
985 		zio->io_logical->io_phys_children++;
986 
987 	return (zio);
988 }
989 
990 zio_t *
991 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
992 	int type, zio_priority_t priority, enum zio_flag flags,
993 	zio_done_func_t *done, void *private)
994 {
995 	zio_t *zio;
996 
997 	ASSERT(vd->vdev_ops->vdev_op_leaf);
998 
999 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1000 	    data, size, done, private, type, priority,
1001 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1002 	    vd, offset, NULL,
1003 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1004 
1005 	return (zio);
1006 }
1007 
1008 void
1009 zio_flush(zio_t *zio, vdev_t *vd)
1010 {
1011 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1012 	    NULL, NULL,
1013 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1014 }
1015 
1016 void
1017 zio_shrink(zio_t *zio, uint64_t size)
1018 {
1019 	ASSERT(zio->io_executor == NULL);
1020 	ASSERT(zio->io_orig_size == zio->io_size);
1021 	ASSERT(size <= zio->io_size);
1022 
1023 	/*
1024 	 * We don't shrink for raidz because of problems with the
1025 	 * reconstruction when reading back less than the block size.
1026 	 * Note, BP_IS_RAIDZ() assumes no compression.
1027 	 */
1028 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1029 	if (!BP_IS_RAIDZ(zio->io_bp))
1030 		zio->io_orig_size = zio->io_size = size;
1031 }
1032 
1033 /*
1034  * ==========================================================================
1035  * Prepare to read and write logical blocks
1036  * ==========================================================================
1037  */
1038 
1039 static int
1040 zio_read_bp_init(zio_t *zio)
1041 {
1042 	blkptr_t *bp = zio->io_bp;
1043 
1044 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1045 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
1046 	    !(zio->io_flags & ZIO_FLAG_RAW)) {
1047 		uint64_t psize =
1048 		    BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1049 		void *cbuf = zio_buf_alloc(psize);
1050 
1051 		zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1052 	}
1053 
1054 	if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1055 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1056 		decode_embedded_bp_compressed(bp, zio->io_data);
1057 	} else {
1058 		ASSERT(!BP_IS_EMBEDDED(bp));
1059 	}
1060 
1061 	if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1062 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1063 
1064 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1065 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1066 
1067 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1068 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1069 
1070 	return (ZIO_PIPELINE_CONTINUE);
1071 }
1072 
1073 static int
1074 zio_write_bp_init(zio_t *zio)
1075 {
1076 	spa_t *spa = zio->io_spa;
1077 	zio_prop_t *zp = &zio->io_prop;
1078 	enum zio_compress compress = zp->zp_compress;
1079 	blkptr_t *bp = zio->io_bp;
1080 	uint64_t lsize = zio->io_size;
1081 	uint64_t psize = lsize;
1082 	int pass = 1;
1083 
1084 	/*
1085 	 * If our children haven't all reached the ready stage,
1086 	 * wait for them and then repeat this pipeline stage.
1087 	 */
1088 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1089 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1090 		return (ZIO_PIPELINE_STOP);
1091 
1092 	if (!IO_IS_ALLOCATING(zio))
1093 		return (ZIO_PIPELINE_CONTINUE);
1094 
1095 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1096 
1097 	if (zio->io_bp_override) {
1098 		ASSERT(bp->blk_birth != zio->io_txg);
1099 		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1100 
1101 		*bp = *zio->io_bp_override;
1102 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1103 
1104 		if (BP_IS_EMBEDDED(bp))
1105 			return (ZIO_PIPELINE_CONTINUE);
1106 
1107 		/*
1108 		 * If we've been overridden and nopwrite is set then
1109 		 * set the flag accordingly to indicate that a nopwrite
1110 		 * has already occurred.
1111 		 */
1112 		if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1113 			ASSERT(!zp->zp_dedup);
1114 			zio->io_flags |= ZIO_FLAG_NOPWRITE;
1115 			return (ZIO_PIPELINE_CONTINUE);
1116 		}
1117 
1118 		ASSERT(!zp->zp_nopwrite);
1119 
1120 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1121 			return (ZIO_PIPELINE_CONTINUE);
1122 
1123 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1124 		    zp->zp_dedup_verify);
1125 
1126 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1127 			BP_SET_DEDUP(bp, 1);
1128 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1129 			return (ZIO_PIPELINE_CONTINUE);
1130 		}
1131 		zio->io_bp_override = NULL;
1132 		BP_ZERO(bp);
1133 	}
1134 
1135 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1136 		/*
1137 		 * We're rewriting an existing block, which means we're
1138 		 * working on behalf of spa_sync().  For spa_sync() to
1139 		 * converge, it must eventually be the case that we don't
1140 		 * have to allocate new blocks.  But compression changes
1141 		 * the blocksize, which forces a reallocate, and makes
1142 		 * convergence take longer.  Therefore, after the first
1143 		 * few passes, stop compressing to ensure convergence.
1144 		 */
1145 		pass = spa_sync_pass(spa);
1146 
1147 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
1148 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1149 		ASSERT(!BP_GET_DEDUP(bp));
1150 
1151 		if (pass >= zfs_sync_pass_dont_compress)
1152 			compress = ZIO_COMPRESS_OFF;
1153 
1154 		/* Make sure someone doesn't change their mind on overwrites */
1155 		ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1156 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1157 	}
1158 
1159 	if (compress != ZIO_COMPRESS_OFF) {
1160 		void *cbuf = zio_buf_alloc(lsize);
1161 		psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1162 		if (psize == 0 || psize == lsize) {
1163 			compress = ZIO_COMPRESS_OFF;
1164 			zio_buf_free(cbuf, lsize);
1165 		} else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1166 		    zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1167 		    spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1168 			encode_embedded_bp_compressed(bp,
1169 			    cbuf, compress, lsize, psize);
1170 			BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1171 			BP_SET_TYPE(bp, zio->io_prop.zp_type);
1172 			BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1173 			zio_buf_free(cbuf, lsize);
1174 			bp->blk_birth = zio->io_txg;
1175 			zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1176 			ASSERT(spa_feature_is_active(spa,
1177 			    SPA_FEATURE_EMBEDDED_DATA));
1178 			return (ZIO_PIPELINE_CONTINUE);
1179 		} else {
1180 			/*
1181 			 * Round up compressed size to MINBLOCKSIZE and
1182 			 * zero the tail.
1183 			 */
1184 			size_t rounded =
1185 			    P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE);
1186 			if (rounded > psize) {
1187 				bzero((char *)cbuf + psize, rounded - psize);
1188 				psize = rounded;
1189 			}
1190 			if (psize == lsize) {
1191 				compress = ZIO_COMPRESS_OFF;
1192 				zio_buf_free(cbuf, lsize);
1193 			} else {
1194 				zio_push_transform(zio, cbuf,
1195 				    psize, lsize, NULL);
1196 			}
1197 		}
1198 	}
1199 
1200 	/*
1201 	 * The final pass of spa_sync() must be all rewrites, but the first
1202 	 * few passes offer a trade-off: allocating blocks defers convergence,
1203 	 * but newly allocated blocks are sequential, so they can be written
1204 	 * to disk faster.  Therefore, we allow the first few passes of
1205 	 * spa_sync() to allocate new blocks, but force rewrites after that.
1206 	 * There should only be a handful of blocks after pass 1 in any case.
1207 	 */
1208 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1209 	    BP_GET_PSIZE(bp) == psize &&
1210 	    pass >= zfs_sync_pass_rewrite) {
1211 		ASSERT(psize != 0);
1212 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1213 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1214 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1215 	} else {
1216 		BP_ZERO(bp);
1217 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1218 	}
1219 
1220 	if (psize == 0) {
1221 		if (zio->io_bp_orig.blk_birth != 0 &&
1222 		    spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1223 			BP_SET_LSIZE(bp, lsize);
1224 			BP_SET_TYPE(bp, zp->zp_type);
1225 			BP_SET_LEVEL(bp, zp->zp_level);
1226 			BP_SET_BIRTH(bp, zio->io_txg, 0);
1227 		}
1228 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1229 	} else {
1230 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1231 		BP_SET_LSIZE(bp, lsize);
1232 		BP_SET_TYPE(bp, zp->zp_type);
1233 		BP_SET_LEVEL(bp, zp->zp_level);
1234 		BP_SET_PSIZE(bp, psize);
1235 		BP_SET_COMPRESS(bp, compress);
1236 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1237 		BP_SET_DEDUP(bp, zp->zp_dedup);
1238 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1239 		if (zp->zp_dedup) {
1240 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1241 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1242 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1243 		}
1244 		if (zp->zp_nopwrite) {
1245 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1246 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1247 			zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1248 		}
1249 	}
1250 
1251 	return (ZIO_PIPELINE_CONTINUE);
1252 }
1253 
1254 static int
1255 zio_free_bp_init(zio_t *zio)
1256 {
1257 	blkptr_t *bp = zio->io_bp;
1258 
1259 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1260 		if (BP_GET_DEDUP(bp))
1261 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1262 	}
1263 
1264 	return (ZIO_PIPELINE_CONTINUE);
1265 }
1266 
1267 /*
1268  * ==========================================================================
1269  * Execute the I/O pipeline
1270  * ==========================================================================
1271  */
1272 
1273 static void
1274 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1275 {
1276 	spa_t *spa = zio->io_spa;
1277 	zio_type_t t = zio->io_type;
1278 	int flags = (cutinline ? TQ_FRONT : 0);
1279 
1280 	/*
1281 	 * If we're a config writer or a probe, the normal issue and
1282 	 * interrupt threads may all be blocked waiting for the config lock.
1283 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1284 	 */
1285 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1286 		t = ZIO_TYPE_NULL;
1287 
1288 	/*
1289 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1290 	 */
1291 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1292 		t = ZIO_TYPE_NULL;
1293 
1294 	/*
1295 	 * If this is a high priority I/O, then use the high priority taskq if
1296 	 * available.
1297 	 */
1298 	if (zio->io_priority == ZIO_PRIORITY_NOW &&
1299 	    spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1300 		q++;
1301 
1302 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1303 
1304 	/*
1305 	 * NB: We are assuming that the zio can only be dispatched
1306 	 * to a single taskq at a time.  It would be a grievous error
1307 	 * to dispatch the zio to another taskq at the same time.
1308 	 */
1309 	ASSERT(zio->io_tqent.tqent_next == NULL);
1310 	spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1311 	    flags, &zio->io_tqent);
1312 }
1313 
1314 static boolean_t
1315 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1316 {
1317 	kthread_t *executor = zio->io_executor;
1318 	spa_t *spa = zio->io_spa;
1319 
1320 	for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1321 		spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1322 		uint_t i;
1323 		for (i = 0; i < tqs->stqs_count; i++) {
1324 			if (taskq_member(tqs->stqs_taskq[i], executor))
1325 				return (B_TRUE);
1326 		}
1327 	}
1328 
1329 	return (B_FALSE);
1330 }
1331 
1332 static int
1333 zio_issue_async(zio_t *zio)
1334 {
1335 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1336 
1337 	return (ZIO_PIPELINE_STOP);
1338 }
1339 
1340 void
1341 zio_interrupt(zio_t *zio)
1342 {
1343 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1344 }
1345 
1346 /*
1347  * Execute the I/O pipeline until one of the following occurs:
1348  *
1349  *	(1) the I/O completes
1350  *	(2) the pipeline stalls waiting for dependent child I/Os
1351  *	(3) the I/O issues, so we're waiting for an I/O completion interrupt
1352  *	(4) the I/O is delegated by vdev-level caching or aggregation
1353  *	(5) the I/O is deferred due to vdev-level queueing
1354  *	(6) the I/O is handed off to another thread.
1355  *
1356  * In all cases, the pipeline stops whenever there's no CPU work; it never
1357  * burns a thread in cv_wait().
1358  *
1359  * There's no locking on io_stage because there's no legitimate way
1360  * for multiple threads to be attempting to process the same I/O.
1361  */
1362 static zio_pipe_stage_t *zio_pipeline[];
1363 
1364 void
1365 zio_execute(zio_t *zio)
1366 {
1367 	zio->io_executor = curthread;
1368 
1369 	while (zio->io_stage < ZIO_STAGE_DONE) {
1370 		enum zio_stage pipeline = zio->io_pipeline;
1371 		enum zio_stage stage = zio->io_stage;
1372 		int rv;
1373 
1374 		ASSERT(!MUTEX_HELD(&zio->io_lock));
1375 		ASSERT(ISP2(stage));
1376 		ASSERT(zio->io_stall == NULL);
1377 
1378 		do {
1379 			stage <<= 1;
1380 		} while ((stage & pipeline) == 0);
1381 
1382 		ASSERT(stage <= ZIO_STAGE_DONE);
1383 
1384 		/*
1385 		 * If we are in interrupt context and this pipeline stage
1386 		 * will grab a config lock that is held across I/O,
1387 		 * or may wait for an I/O that needs an interrupt thread
1388 		 * to complete, issue async to avoid deadlock.
1389 		 *
1390 		 * For VDEV_IO_START, we cut in line so that the io will
1391 		 * be sent to disk promptly.
1392 		 */
1393 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1394 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1395 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1396 			    zio_requeue_io_start_cut_in_line : B_FALSE;
1397 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1398 			return;
1399 		}
1400 
1401 		zio->io_stage = stage;
1402 		rv = zio_pipeline[highbit64(stage) - 1](zio);
1403 
1404 		if (rv == ZIO_PIPELINE_STOP)
1405 			return;
1406 
1407 		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1408 	}
1409 }
1410 
1411 /*
1412  * ==========================================================================
1413  * Initiate I/O, either sync or async
1414  * ==========================================================================
1415  */
1416 int
1417 zio_wait(zio_t *zio)
1418 {
1419 	int error;
1420 
1421 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1422 	ASSERT(zio->io_executor == NULL);
1423 
1424 	zio->io_waiter = curthread;
1425 
1426 	zio_execute(zio);
1427 
1428 	mutex_enter(&zio->io_lock);
1429 	while (zio->io_executor != NULL)
1430 		cv_wait(&zio->io_cv, &zio->io_lock);
1431 	mutex_exit(&zio->io_lock);
1432 
1433 	error = zio->io_error;
1434 	zio_destroy(zio);
1435 
1436 	return (error);
1437 }
1438 
1439 void
1440 zio_nowait(zio_t *zio)
1441 {
1442 	ASSERT(zio->io_executor == NULL);
1443 
1444 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1445 	    zio_unique_parent(zio) == NULL) {
1446 		/*
1447 		 * This is a logical async I/O with no parent to wait for it.
1448 		 * We add it to the spa_async_root_zio "Godfather" I/O which
1449 		 * will ensure they complete prior to unloading the pool.
1450 		 */
1451 		spa_t *spa = zio->io_spa;
1452 
1453 		zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1454 	}
1455 
1456 	zio_execute(zio);
1457 }
1458 
1459 /*
1460  * ==========================================================================
1461  * Reexecute or suspend/resume failed I/O
1462  * ==========================================================================
1463  */
1464 
1465 static void
1466 zio_reexecute(zio_t *pio)
1467 {
1468 	zio_t *cio, *cio_next;
1469 
1470 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1471 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1472 	ASSERT(pio->io_gang_leader == NULL);
1473 	ASSERT(pio->io_gang_tree == NULL);
1474 
1475 	pio->io_flags = pio->io_orig_flags;
1476 	pio->io_stage = pio->io_orig_stage;
1477 	pio->io_pipeline = pio->io_orig_pipeline;
1478 	pio->io_reexecute = 0;
1479 	pio->io_flags |= ZIO_FLAG_REEXECUTED;
1480 	pio->io_error = 0;
1481 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1482 		pio->io_state[w] = 0;
1483 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1484 		pio->io_child_error[c] = 0;
1485 
1486 	if (IO_IS_ALLOCATING(pio))
1487 		BP_ZERO(pio->io_bp);
1488 
1489 	/*
1490 	 * As we reexecute pio's children, new children could be created.
1491 	 * New children go to the head of pio's io_child_list, however,
1492 	 * so we will (correctly) not reexecute them.  The key is that
1493 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1494 	 * cannot be affected by any side effects of reexecuting 'cio'.
1495 	 */
1496 	for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1497 		cio_next = zio_walk_children(pio);
1498 		mutex_enter(&pio->io_lock);
1499 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1500 			pio->io_children[cio->io_child_type][w]++;
1501 		mutex_exit(&pio->io_lock);
1502 		zio_reexecute(cio);
1503 	}
1504 
1505 	/*
1506 	 * Now that all children have been reexecuted, execute the parent.
1507 	 * We don't reexecute "The Godfather" I/O here as it's the
1508 	 * responsibility of the caller to wait on him.
1509 	 */
1510 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1511 		zio_execute(pio);
1512 }
1513 
1514 void
1515 zio_suspend(spa_t *spa, zio_t *zio)
1516 {
1517 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1518 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1519 		    "failure and the failure mode property for this pool "
1520 		    "is set to panic.", spa_name(spa));
1521 
1522 	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1523 
1524 	mutex_enter(&spa->spa_suspend_lock);
1525 
1526 	if (spa->spa_suspend_zio_root == NULL)
1527 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1528 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1529 		    ZIO_FLAG_GODFATHER);
1530 
1531 	spa->spa_suspended = B_TRUE;
1532 
1533 	if (zio != NULL) {
1534 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1535 		ASSERT(zio != spa->spa_suspend_zio_root);
1536 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1537 		ASSERT(zio_unique_parent(zio) == NULL);
1538 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1539 		zio_add_child(spa->spa_suspend_zio_root, zio);
1540 	}
1541 
1542 	mutex_exit(&spa->spa_suspend_lock);
1543 }
1544 
1545 int
1546 zio_resume(spa_t *spa)
1547 {
1548 	zio_t *pio;
1549 
1550 	/*
1551 	 * Reexecute all previously suspended i/o.
1552 	 */
1553 	mutex_enter(&spa->spa_suspend_lock);
1554 	spa->spa_suspended = B_FALSE;
1555 	cv_broadcast(&spa->spa_suspend_cv);
1556 	pio = spa->spa_suspend_zio_root;
1557 	spa->spa_suspend_zio_root = NULL;
1558 	mutex_exit(&spa->spa_suspend_lock);
1559 
1560 	if (pio == NULL)
1561 		return (0);
1562 
1563 	zio_reexecute(pio);
1564 	return (zio_wait(pio));
1565 }
1566 
1567 void
1568 zio_resume_wait(spa_t *spa)
1569 {
1570 	mutex_enter(&spa->spa_suspend_lock);
1571 	while (spa_suspended(spa))
1572 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1573 	mutex_exit(&spa->spa_suspend_lock);
1574 }
1575 
1576 /*
1577  * ==========================================================================
1578  * Gang blocks.
1579  *
1580  * A gang block is a collection of small blocks that looks to the DMU
1581  * like one large block.  When zio_dva_allocate() cannot find a block
1582  * of the requested size, due to either severe fragmentation or the pool
1583  * being nearly full, it calls zio_write_gang_block() to construct the
1584  * block from smaller fragments.
1585  *
1586  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1587  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
1588  * an indirect block: it's an array of block pointers.  It consumes
1589  * only one sector and hence is allocatable regardless of fragmentation.
1590  * The gang header's bps point to its gang members, which hold the data.
1591  *
1592  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1593  * as the verifier to ensure uniqueness of the SHA256 checksum.
1594  * Critically, the gang block bp's blk_cksum is the checksum of the data,
1595  * not the gang header.  This ensures that data block signatures (needed for
1596  * deduplication) are independent of how the block is physically stored.
1597  *
1598  * Gang blocks can be nested: a gang member may itself be a gang block.
1599  * Thus every gang block is a tree in which root and all interior nodes are
1600  * gang headers, and the leaves are normal blocks that contain user data.
1601  * The root of the gang tree is called the gang leader.
1602  *
1603  * To perform any operation (read, rewrite, free, claim) on a gang block,
1604  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1605  * in the io_gang_tree field of the original logical i/o by recursively
1606  * reading the gang leader and all gang headers below it.  This yields
1607  * an in-core tree containing the contents of every gang header and the
1608  * bps for every constituent of the gang block.
1609  *
1610  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1611  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
1612  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1613  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1614  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1615  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
1616  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1617  * of the gang header plus zio_checksum_compute() of the data to update the
1618  * gang header's blk_cksum as described above.
1619  *
1620  * The two-phase assemble/issue model solves the problem of partial failure --
1621  * what if you'd freed part of a gang block but then couldn't read the
1622  * gang header for another part?  Assembling the entire gang tree first
1623  * ensures that all the necessary gang header I/O has succeeded before
1624  * starting the actual work of free, claim, or write.  Once the gang tree
1625  * is assembled, free and claim are in-memory operations that cannot fail.
1626  *
1627  * In the event that a gang write fails, zio_dva_unallocate() walks the
1628  * gang tree to immediately free (i.e. insert back into the space map)
1629  * everything we've allocated.  This ensures that we don't get ENOSPC
1630  * errors during repeated suspend/resume cycles due to a flaky device.
1631  *
1632  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
1633  * the gang tree, we won't modify the block, so we can safely defer the free
1634  * (knowing that the block is still intact).  If we *can* assemble the gang
1635  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1636  * each constituent bp and we can allocate a new block on the next sync pass.
1637  *
1638  * In all cases, the gang tree allows complete recovery from partial failure.
1639  * ==========================================================================
1640  */
1641 
1642 static zio_t *
1643 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1644 {
1645 	if (gn != NULL)
1646 		return (pio);
1647 
1648 	return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1649 	    NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1650 	    &pio->io_bookmark));
1651 }
1652 
1653 zio_t *
1654 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1655 {
1656 	zio_t *zio;
1657 
1658 	if (gn != NULL) {
1659 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1660 		    gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1661 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1662 		/*
1663 		 * As we rewrite each gang header, the pipeline will compute
1664 		 * a new gang block header checksum for it; but no one will
1665 		 * compute a new data checksum, so we do that here.  The one
1666 		 * exception is the gang leader: the pipeline already computed
1667 		 * its data checksum because that stage precedes gang assembly.
1668 		 * (Presently, nothing actually uses interior data checksums;
1669 		 * this is just good hygiene.)
1670 		 */
1671 		if (gn != pio->io_gang_leader->io_gang_tree) {
1672 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1673 			    data, BP_GET_PSIZE(bp));
1674 		}
1675 		/*
1676 		 * If we are here to damage data for testing purposes,
1677 		 * leave the GBH alone so that we can detect the damage.
1678 		 */
1679 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1680 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1681 	} else {
1682 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1683 		    data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1684 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1685 	}
1686 
1687 	return (zio);
1688 }
1689 
1690 /* ARGSUSED */
1691 zio_t *
1692 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1693 {
1694 	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1695 	    ZIO_GANG_CHILD_FLAGS(pio)));
1696 }
1697 
1698 /* ARGSUSED */
1699 zio_t *
1700 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1701 {
1702 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1703 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1704 }
1705 
1706 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1707 	NULL,
1708 	zio_read_gang,
1709 	zio_rewrite_gang,
1710 	zio_free_gang,
1711 	zio_claim_gang,
1712 	NULL
1713 };
1714 
1715 static void zio_gang_tree_assemble_done(zio_t *zio);
1716 
1717 static zio_gang_node_t *
1718 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1719 {
1720 	zio_gang_node_t *gn;
1721 
1722 	ASSERT(*gnpp == NULL);
1723 
1724 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1725 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1726 	*gnpp = gn;
1727 
1728 	return (gn);
1729 }
1730 
1731 static void
1732 zio_gang_node_free(zio_gang_node_t **gnpp)
1733 {
1734 	zio_gang_node_t *gn = *gnpp;
1735 
1736 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1737 		ASSERT(gn->gn_child[g] == NULL);
1738 
1739 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1740 	kmem_free(gn, sizeof (*gn));
1741 	*gnpp = NULL;
1742 }
1743 
1744 static void
1745 zio_gang_tree_free(zio_gang_node_t **gnpp)
1746 {
1747 	zio_gang_node_t *gn = *gnpp;
1748 
1749 	if (gn == NULL)
1750 		return;
1751 
1752 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1753 		zio_gang_tree_free(&gn->gn_child[g]);
1754 
1755 	zio_gang_node_free(gnpp);
1756 }
1757 
1758 static void
1759 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1760 {
1761 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1762 
1763 	ASSERT(gio->io_gang_leader == gio);
1764 	ASSERT(BP_IS_GANG(bp));
1765 
1766 	zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1767 	    SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1768 	    gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1769 }
1770 
1771 static void
1772 zio_gang_tree_assemble_done(zio_t *zio)
1773 {
1774 	zio_t *gio = zio->io_gang_leader;
1775 	zio_gang_node_t *gn = zio->io_private;
1776 	blkptr_t *bp = zio->io_bp;
1777 
1778 	ASSERT(gio == zio_unique_parent(zio));
1779 	ASSERT(zio->io_child_count == 0);
1780 
1781 	if (zio->io_error)
1782 		return;
1783 
1784 	if (BP_SHOULD_BYTESWAP(bp))
1785 		byteswap_uint64_array(zio->io_data, zio->io_size);
1786 
1787 	ASSERT(zio->io_data == gn->gn_gbh);
1788 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1789 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1790 
1791 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1792 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1793 		if (!BP_IS_GANG(gbp))
1794 			continue;
1795 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1796 	}
1797 }
1798 
1799 static void
1800 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1801 {
1802 	zio_t *gio = pio->io_gang_leader;
1803 	zio_t *zio;
1804 
1805 	ASSERT(BP_IS_GANG(bp) == !!gn);
1806 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1807 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1808 
1809 	/*
1810 	 * If you're a gang header, your data is in gn->gn_gbh.
1811 	 * If you're a gang member, your data is in 'data' and gn == NULL.
1812 	 */
1813 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1814 
1815 	if (gn != NULL) {
1816 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1817 
1818 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1819 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1820 			if (BP_IS_HOLE(gbp))
1821 				continue;
1822 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1823 			data = (char *)data + BP_GET_PSIZE(gbp);
1824 		}
1825 	}
1826 
1827 	if (gn == gio->io_gang_tree)
1828 		ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1829 
1830 	if (zio != pio)
1831 		zio_nowait(zio);
1832 }
1833 
1834 static int
1835 zio_gang_assemble(zio_t *zio)
1836 {
1837 	blkptr_t *bp = zio->io_bp;
1838 
1839 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1840 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1841 
1842 	zio->io_gang_leader = zio;
1843 
1844 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1845 
1846 	return (ZIO_PIPELINE_CONTINUE);
1847 }
1848 
1849 static int
1850 zio_gang_issue(zio_t *zio)
1851 {
1852 	blkptr_t *bp = zio->io_bp;
1853 
1854 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1855 		return (ZIO_PIPELINE_STOP);
1856 
1857 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1858 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1859 
1860 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1861 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1862 	else
1863 		zio_gang_tree_free(&zio->io_gang_tree);
1864 
1865 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1866 
1867 	return (ZIO_PIPELINE_CONTINUE);
1868 }
1869 
1870 static void
1871 zio_write_gang_member_ready(zio_t *zio)
1872 {
1873 	zio_t *pio = zio_unique_parent(zio);
1874 	zio_t *gio = zio->io_gang_leader;
1875 	dva_t *cdva = zio->io_bp->blk_dva;
1876 	dva_t *pdva = pio->io_bp->blk_dva;
1877 	uint64_t asize;
1878 
1879 	if (BP_IS_HOLE(zio->io_bp))
1880 		return;
1881 
1882 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1883 
1884 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1885 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1886 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1887 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1888 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1889 
1890 	mutex_enter(&pio->io_lock);
1891 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1892 		ASSERT(DVA_GET_GANG(&pdva[d]));
1893 		asize = DVA_GET_ASIZE(&pdva[d]);
1894 		asize += DVA_GET_ASIZE(&cdva[d]);
1895 		DVA_SET_ASIZE(&pdva[d], asize);
1896 	}
1897 	mutex_exit(&pio->io_lock);
1898 }
1899 
1900 static int
1901 zio_write_gang_block(zio_t *pio)
1902 {
1903 	spa_t *spa = pio->io_spa;
1904 	blkptr_t *bp = pio->io_bp;
1905 	zio_t *gio = pio->io_gang_leader;
1906 	zio_t *zio;
1907 	zio_gang_node_t *gn, **gnpp;
1908 	zio_gbh_phys_t *gbh;
1909 	uint64_t txg = pio->io_txg;
1910 	uint64_t resid = pio->io_size;
1911 	uint64_t lsize;
1912 	int copies = gio->io_prop.zp_copies;
1913 	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1914 	zio_prop_t zp;
1915 	int error;
1916 
1917 	error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1918 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1919 	    METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1920 	if (error) {
1921 		pio->io_error = error;
1922 		return (ZIO_PIPELINE_CONTINUE);
1923 	}
1924 
1925 	if (pio == gio) {
1926 		gnpp = &gio->io_gang_tree;
1927 	} else {
1928 		gnpp = pio->io_private;
1929 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
1930 	}
1931 
1932 	gn = zio_gang_node_alloc(gnpp);
1933 	gbh = gn->gn_gbh;
1934 	bzero(gbh, SPA_GANGBLOCKSIZE);
1935 
1936 	/*
1937 	 * Create the gang header.
1938 	 */
1939 	zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1940 	    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1941 
1942 	/*
1943 	 * Create and nowait the gang children.
1944 	 */
1945 	for (int g = 0; resid != 0; resid -= lsize, g++) {
1946 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1947 		    SPA_MINBLOCKSIZE);
1948 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1949 
1950 		zp.zp_checksum = gio->io_prop.zp_checksum;
1951 		zp.zp_compress = ZIO_COMPRESS_OFF;
1952 		zp.zp_type = DMU_OT_NONE;
1953 		zp.zp_level = 0;
1954 		zp.zp_copies = gio->io_prop.zp_copies;
1955 		zp.zp_dedup = B_FALSE;
1956 		zp.zp_dedup_verify = B_FALSE;
1957 		zp.zp_nopwrite = B_FALSE;
1958 
1959 		zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1960 		    (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1961 		    zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1962 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1963 		    &pio->io_bookmark));
1964 	}
1965 
1966 	/*
1967 	 * Set pio's pipeline to just wait for zio to finish.
1968 	 */
1969 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1970 
1971 	zio_nowait(zio);
1972 
1973 	return (ZIO_PIPELINE_CONTINUE);
1974 }
1975 
1976 /*
1977  * The zio_nop_write stage in the pipeline determines if allocating
1978  * a new bp is necessary.  By leveraging a cryptographically secure checksum,
1979  * such as SHA256, we can compare the checksums of the new data and the old
1980  * to determine if allocating a new block is required.  The nopwrite
1981  * feature can handle writes in either syncing or open context (i.e. zil
1982  * writes) and as a result is mutually exclusive with dedup.
1983  */
1984 static int
1985 zio_nop_write(zio_t *zio)
1986 {
1987 	blkptr_t *bp = zio->io_bp;
1988 	blkptr_t *bp_orig = &zio->io_bp_orig;
1989 	zio_prop_t *zp = &zio->io_prop;
1990 
1991 	ASSERT(BP_GET_LEVEL(bp) == 0);
1992 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1993 	ASSERT(zp->zp_nopwrite);
1994 	ASSERT(!zp->zp_dedup);
1995 	ASSERT(zio->io_bp_override == NULL);
1996 	ASSERT(IO_IS_ALLOCATING(zio));
1997 
1998 	/*
1999 	 * Check to see if the original bp and the new bp have matching
2000 	 * characteristics (i.e. same checksum, compression algorithms, etc).
2001 	 * If they don't then just continue with the pipeline which will
2002 	 * allocate a new bp.
2003 	 */
2004 	if (BP_IS_HOLE(bp_orig) ||
2005 	    !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
2006 	    BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2007 	    BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2008 	    BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2009 	    zp->zp_copies != BP_GET_NDVAS(bp_orig))
2010 		return (ZIO_PIPELINE_CONTINUE);
2011 
2012 	/*
2013 	 * If the checksums match then reset the pipeline so that we
2014 	 * avoid allocating a new bp and issuing any I/O.
2015 	 */
2016 	if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2017 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
2018 		ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2019 		ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2020 		ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2021 		ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2022 		    sizeof (uint64_t)) == 0);
2023 
2024 		*bp = *bp_orig;
2025 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2026 		zio->io_flags |= ZIO_FLAG_NOPWRITE;
2027 	}
2028 
2029 	return (ZIO_PIPELINE_CONTINUE);
2030 }
2031 
2032 /*
2033  * ==========================================================================
2034  * Dedup
2035  * ==========================================================================
2036  */
2037 static void
2038 zio_ddt_child_read_done(zio_t *zio)
2039 {
2040 	blkptr_t *bp = zio->io_bp;
2041 	ddt_entry_t *dde = zio->io_private;
2042 	ddt_phys_t *ddp;
2043 	zio_t *pio = zio_unique_parent(zio);
2044 
2045 	mutex_enter(&pio->io_lock);
2046 	ddp = ddt_phys_select(dde, bp);
2047 	if (zio->io_error == 0)
2048 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
2049 	if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2050 		dde->dde_repair_data = zio->io_data;
2051 	else
2052 		zio_buf_free(zio->io_data, zio->io_size);
2053 	mutex_exit(&pio->io_lock);
2054 }
2055 
2056 static int
2057 zio_ddt_read_start(zio_t *zio)
2058 {
2059 	blkptr_t *bp = zio->io_bp;
2060 
2061 	ASSERT(BP_GET_DEDUP(bp));
2062 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2063 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2064 
2065 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
2066 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
2067 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2068 		ddt_phys_t *ddp = dde->dde_phys;
2069 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2070 		blkptr_t blk;
2071 
2072 		ASSERT(zio->io_vsd == NULL);
2073 		zio->io_vsd = dde;
2074 
2075 		if (ddp_self == NULL)
2076 			return (ZIO_PIPELINE_CONTINUE);
2077 
2078 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2079 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2080 				continue;
2081 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2082 			    &blk);
2083 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
2084 			    zio_buf_alloc(zio->io_size), zio->io_size,
2085 			    zio_ddt_child_read_done, dde, zio->io_priority,
2086 			    ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2087 			    &zio->io_bookmark));
2088 		}
2089 		return (ZIO_PIPELINE_CONTINUE);
2090 	}
2091 
2092 	zio_nowait(zio_read(zio, zio->io_spa, bp,
2093 	    zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2094 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2095 
2096 	return (ZIO_PIPELINE_CONTINUE);
2097 }
2098 
2099 static int
2100 zio_ddt_read_done(zio_t *zio)
2101 {
2102 	blkptr_t *bp = zio->io_bp;
2103 
2104 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2105 		return (ZIO_PIPELINE_STOP);
2106 
2107 	ASSERT(BP_GET_DEDUP(bp));
2108 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2109 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2110 
2111 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
2112 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
2113 		ddt_entry_t *dde = zio->io_vsd;
2114 		if (ddt == NULL) {
2115 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2116 			return (ZIO_PIPELINE_CONTINUE);
2117 		}
2118 		if (dde == NULL) {
2119 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2120 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2121 			return (ZIO_PIPELINE_STOP);
2122 		}
2123 		if (dde->dde_repair_data != NULL) {
2124 			bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2125 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
2126 		}
2127 		ddt_repair_done(ddt, dde);
2128 		zio->io_vsd = NULL;
2129 	}
2130 
2131 	ASSERT(zio->io_vsd == NULL);
2132 
2133 	return (ZIO_PIPELINE_CONTINUE);
2134 }
2135 
2136 static boolean_t
2137 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2138 {
2139 	spa_t *spa = zio->io_spa;
2140 
2141 	/*
2142 	 * Note: we compare the original data, not the transformed data,
2143 	 * because when zio->io_bp is an override bp, we will not have
2144 	 * pushed the I/O transforms.  That's an important optimization
2145 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2146 	 */
2147 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2148 		zio_t *lio = dde->dde_lead_zio[p];
2149 
2150 		if (lio != NULL) {
2151 			return (lio->io_orig_size != zio->io_orig_size ||
2152 			    bcmp(zio->io_orig_data, lio->io_orig_data,
2153 			    zio->io_orig_size) != 0);
2154 		}
2155 	}
2156 
2157 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2158 		ddt_phys_t *ddp = &dde->dde_phys[p];
2159 
2160 		if (ddp->ddp_phys_birth != 0) {
2161 			arc_buf_t *abuf = NULL;
2162 			uint32_t aflags = ARC_WAIT;
2163 			blkptr_t blk = *zio->io_bp;
2164 			int error;
2165 
2166 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2167 
2168 			ddt_exit(ddt);
2169 
2170 			error = arc_read(NULL, spa, &blk,
2171 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2172 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2173 			    &aflags, &zio->io_bookmark);
2174 
2175 			if (error == 0) {
2176 				if (arc_buf_size(abuf) != zio->io_orig_size ||
2177 				    bcmp(abuf->b_data, zio->io_orig_data,
2178 				    zio->io_orig_size) != 0)
2179 					error = SET_ERROR(EEXIST);
2180 				VERIFY(arc_buf_remove_ref(abuf, &abuf));
2181 			}
2182 
2183 			ddt_enter(ddt);
2184 			return (error != 0);
2185 		}
2186 	}
2187 
2188 	return (B_FALSE);
2189 }
2190 
2191 static void
2192 zio_ddt_child_write_ready(zio_t *zio)
2193 {
2194 	int p = zio->io_prop.zp_copies;
2195 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2196 	ddt_entry_t *dde = zio->io_private;
2197 	ddt_phys_t *ddp = &dde->dde_phys[p];
2198 	zio_t *pio;
2199 
2200 	if (zio->io_error)
2201 		return;
2202 
2203 	ddt_enter(ddt);
2204 
2205 	ASSERT(dde->dde_lead_zio[p] == zio);
2206 
2207 	ddt_phys_fill(ddp, zio->io_bp);
2208 
2209 	while ((pio = zio_walk_parents(zio)) != NULL)
2210 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2211 
2212 	ddt_exit(ddt);
2213 }
2214 
2215 static void
2216 zio_ddt_child_write_done(zio_t *zio)
2217 {
2218 	int p = zio->io_prop.zp_copies;
2219 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2220 	ddt_entry_t *dde = zio->io_private;
2221 	ddt_phys_t *ddp = &dde->dde_phys[p];
2222 
2223 	ddt_enter(ddt);
2224 
2225 	ASSERT(ddp->ddp_refcnt == 0);
2226 	ASSERT(dde->dde_lead_zio[p] == zio);
2227 	dde->dde_lead_zio[p] = NULL;
2228 
2229 	if (zio->io_error == 0) {
2230 		while (zio_walk_parents(zio) != NULL)
2231 			ddt_phys_addref(ddp);
2232 	} else {
2233 		ddt_phys_clear(ddp);
2234 	}
2235 
2236 	ddt_exit(ddt);
2237 }
2238 
2239 static void
2240 zio_ddt_ditto_write_done(zio_t *zio)
2241 {
2242 	int p = DDT_PHYS_DITTO;
2243 	zio_prop_t *zp = &zio->io_prop;
2244 	blkptr_t *bp = zio->io_bp;
2245 	ddt_t *ddt = ddt_select(zio->io_spa, bp);
2246 	ddt_entry_t *dde = zio->io_private;
2247 	ddt_phys_t *ddp = &dde->dde_phys[p];
2248 	ddt_key_t *ddk = &dde->dde_key;
2249 
2250 	ddt_enter(ddt);
2251 
2252 	ASSERT(ddp->ddp_refcnt == 0);
2253 	ASSERT(dde->dde_lead_zio[p] == zio);
2254 	dde->dde_lead_zio[p] = NULL;
2255 
2256 	if (zio->io_error == 0) {
2257 		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2258 		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2259 		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2260 		if (ddp->ddp_phys_birth != 0)
2261 			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2262 		ddt_phys_fill(ddp, bp);
2263 	}
2264 
2265 	ddt_exit(ddt);
2266 }
2267 
2268 static int
2269 zio_ddt_write(zio_t *zio)
2270 {
2271 	spa_t *spa = zio->io_spa;
2272 	blkptr_t *bp = zio->io_bp;
2273 	uint64_t txg = zio->io_txg;
2274 	zio_prop_t *zp = &zio->io_prop;
2275 	int p = zp->zp_copies;
2276 	int ditto_copies;
2277 	zio_t *cio = NULL;
2278 	zio_t *dio = NULL;
2279 	ddt_t *ddt = ddt_select(spa, bp);
2280 	ddt_entry_t *dde;
2281 	ddt_phys_t *ddp;
2282 
2283 	ASSERT(BP_GET_DEDUP(bp));
2284 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2285 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2286 
2287 	ddt_enter(ddt);
2288 	dde = ddt_lookup(ddt, bp, B_TRUE);
2289 	ddp = &dde->dde_phys[p];
2290 
2291 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2292 		/*
2293 		 * If we're using a weak checksum, upgrade to a strong checksum
2294 		 * and try again.  If we're already using a strong checksum,
2295 		 * we can't resolve it, so just convert to an ordinary write.
2296 		 * (And automatically e-mail a paper to Nature?)
2297 		 */
2298 		if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2299 			zp->zp_checksum = spa_dedup_checksum(spa);
2300 			zio_pop_transforms(zio);
2301 			zio->io_stage = ZIO_STAGE_OPEN;
2302 			BP_ZERO(bp);
2303 		} else {
2304 			zp->zp_dedup = B_FALSE;
2305 		}
2306 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2307 		ddt_exit(ddt);
2308 		return (ZIO_PIPELINE_CONTINUE);
2309 	}
2310 
2311 	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2312 	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2313 
2314 	if (ditto_copies > ddt_ditto_copies_present(dde) &&
2315 	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2316 		zio_prop_t czp = *zp;
2317 
2318 		czp.zp_copies = ditto_copies;
2319 
2320 		/*
2321 		 * If we arrived here with an override bp, we won't have run
2322 		 * the transform stack, so we won't have the data we need to
2323 		 * generate a child i/o.  So, toss the override bp and restart.
2324 		 * This is safe, because using the override bp is just an
2325 		 * optimization; and it's rare, so the cost doesn't matter.
2326 		 */
2327 		if (zio->io_bp_override) {
2328 			zio_pop_transforms(zio);
2329 			zio->io_stage = ZIO_STAGE_OPEN;
2330 			zio->io_pipeline = ZIO_WRITE_PIPELINE;
2331 			zio->io_bp_override = NULL;
2332 			BP_ZERO(bp);
2333 			ddt_exit(ddt);
2334 			return (ZIO_PIPELINE_CONTINUE);
2335 		}
2336 
2337 		dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2338 		    zio->io_orig_size, &czp, NULL, NULL,
2339 		    zio_ddt_ditto_write_done, dde, zio->io_priority,
2340 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2341 
2342 		zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2343 		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2344 	}
2345 
2346 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2347 		if (ddp->ddp_phys_birth != 0)
2348 			ddt_bp_fill(ddp, bp, txg);
2349 		if (dde->dde_lead_zio[p] != NULL)
2350 			zio_add_child(zio, dde->dde_lead_zio[p]);
2351 		else
2352 			ddt_phys_addref(ddp);
2353 	} else if (zio->io_bp_override) {
2354 		ASSERT(bp->blk_birth == txg);
2355 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2356 		ddt_phys_fill(ddp, bp);
2357 		ddt_phys_addref(ddp);
2358 	} else {
2359 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2360 		    zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2361 		    zio_ddt_child_write_done, dde, zio->io_priority,
2362 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2363 
2364 		zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2365 		dde->dde_lead_zio[p] = cio;
2366 	}
2367 
2368 	ddt_exit(ddt);
2369 
2370 	if (cio)
2371 		zio_nowait(cio);
2372 	if (dio)
2373 		zio_nowait(dio);
2374 
2375 	return (ZIO_PIPELINE_CONTINUE);
2376 }
2377 
2378 ddt_entry_t *freedde; /* for debugging */
2379 
2380 static int
2381 zio_ddt_free(zio_t *zio)
2382 {
2383 	spa_t *spa = zio->io_spa;
2384 	blkptr_t *bp = zio->io_bp;
2385 	ddt_t *ddt = ddt_select(spa, bp);
2386 	ddt_entry_t *dde;
2387 	ddt_phys_t *ddp;
2388 
2389 	ASSERT(BP_GET_DEDUP(bp));
2390 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2391 
2392 	ddt_enter(ddt);
2393 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2394 	ddp = ddt_phys_select(dde, bp);
2395 	ddt_phys_decref(ddp);
2396 	ddt_exit(ddt);
2397 
2398 	return (ZIO_PIPELINE_CONTINUE);
2399 }
2400 
2401 /*
2402  * ==========================================================================
2403  * Allocate and free blocks
2404  * ==========================================================================
2405  */
2406 static int
2407 zio_dva_allocate(zio_t *zio)
2408 {
2409 	spa_t *spa = zio->io_spa;
2410 	metaslab_class_t *mc = spa_normal_class(spa);
2411 	blkptr_t *bp = zio->io_bp;
2412 	int error;
2413 	int flags = 0;
2414 
2415 	if (zio->io_gang_leader == NULL) {
2416 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2417 		zio->io_gang_leader = zio;
2418 	}
2419 
2420 	ASSERT(BP_IS_HOLE(bp));
2421 	ASSERT0(BP_GET_NDVAS(bp));
2422 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
2423 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2424 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2425 
2426 	/*
2427 	 * The dump device does not support gang blocks so allocation on
2428 	 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2429 	 * the "fast" gang feature.
2430 	 */
2431 	flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2432 	flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2433 	    METASLAB_GANG_CHILD : 0;
2434 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
2435 	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2436 
2437 	if (error) {
2438 		spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2439 		    "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2440 		    error);
2441 		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2442 			return (zio_write_gang_block(zio));
2443 		zio->io_error = error;
2444 	}
2445 
2446 	return (ZIO_PIPELINE_CONTINUE);
2447 }
2448 
2449 static int
2450 zio_dva_free(zio_t *zio)
2451 {
2452 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2453 
2454 	return (ZIO_PIPELINE_CONTINUE);
2455 }
2456 
2457 static int
2458 zio_dva_claim(zio_t *zio)
2459 {
2460 	int error;
2461 
2462 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2463 	if (error)
2464 		zio->io_error = error;
2465 
2466 	return (ZIO_PIPELINE_CONTINUE);
2467 }
2468 
2469 /*
2470  * Undo an allocation.  This is used by zio_done() when an I/O fails
2471  * and we want to give back the block we just allocated.
2472  * This handles both normal blocks and gang blocks.
2473  */
2474 static void
2475 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2476 {
2477 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2478 	ASSERT(zio->io_bp_override == NULL);
2479 
2480 	if (!BP_IS_HOLE(bp))
2481 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2482 
2483 	if (gn != NULL) {
2484 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2485 			zio_dva_unallocate(zio, gn->gn_child[g],
2486 			    &gn->gn_gbh->zg_blkptr[g]);
2487 		}
2488 	}
2489 }
2490 
2491 /*
2492  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
2493  */
2494 int
2495 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2496     uint64_t size, boolean_t use_slog)
2497 {
2498 	int error = 1;
2499 
2500 	ASSERT(txg > spa_syncing_txg(spa));
2501 
2502 	/*
2503 	 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2504 	 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2505 	 * when allocating them.
2506 	 */
2507 	if (use_slog) {
2508 		error = metaslab_alloc(spa, spa_log_class(spa), size,
2509 		    new_bp, 1, txg, old_bp,
2510 		    METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2511 	}
2512 
2513 	if (error) {
2514 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
2515 		    new_bp, 1, txg, old_bp,
2516 		    METASLAB_HINTBP_AVOID);
2517 	}
2518 
2519 	if (error == 0) {
2520 		BP_SET_LSIZE(new_bp, size);
2521 		BP_SET_PSIZE(new_bp, size);
2522 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2523 		BP_SET_CHECKSUM(new_bp,
2524 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2525 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2526 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2527 		BP_SET_LEVEL(new_bp, 0);
2528 		BP_SET_DEDUP(new_bp, 0);
2529 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2530 	}
2531 
2532 	return (error);
2533 }
2534 
2535 /*
2536  * Free an intent log block.
2537  */
2538 void
2539 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2540 {
2541 	ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2542 	ASSERT(!BP_IS_GANG(bp));
2543 
2544 	zio_free(spa, txg, bp);
2545 }
2546 
2547 /*
2548  * ==========================================================================
2549  * Read and write to physical devices
2550  * ==========================================================================
2551  */
2552 
2553 
2554 /*
2555  * Issue an I/O to the underlying vdev. Typically the issue pipeline
2556  * stops after this stage and will resume upon I/O completion.
2557  * However, there are instances where the vdev layer may need to
2558  * continue the pipeline when an I/O was not issued. Since the I/O
2559  * that was sent to the vdev layer might be different than the one
2560  * currently active in the pipeline (see vdev_queue_io()), we explicitly
2561  * force the underlying vdev layers to call either zio_execute() or
2562  * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
2563  */
2564 static int
2565 zio_vdev_io_start(zio_t *zio)
2566 {
2567 	vdev_t *vd = zio->io_vd;
2568 	uint64_t align;
2569 	spa_t *spa = zio->io_spa;
2570 
2571 	ASSERT(zio->io_error == 0);
2572 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2573 
2574 	if (vd == NULL) {
2575 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2576 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2577 
2578 		/*
2579 		 * The mirror_ops handle multiple DVAs in a single BP.
2580 		 */
2581 		vdev_mirror_ops.vdev_op_io_start(zio);
2582 		return (ZIO_PIPELINE_STOP);
2583 	}
2584 
2585 	/*
2586 	 * We keep track of time-sensitive I/Os so that the scan thread
2587 	 * can quickly react to certain workloads.  In particular, we care
2588 	 * about non-scrubbing, top-level reads and writes with the following
2589 	 * characteristics:
2590 	 *	- synchronous writes of user data to non-slog devices
2591 	 *	- any reads of user data
2592 	 * When these conditions are met, adjust the timestamp of spa_last_io
2593 	 * which allows the scan thread to adjust its workload accordingly.
2594 	 */
2595 	if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2596 	    vd == vd->vdev_top && !vd->vdev_islog &&
2597 	    zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2598 	    zio->io_txg != spa_syncing_txg(spa)) {
2599 		uint64_t old = spa->spa_last_io;
2600 		uint64_t new = ddi_get_lbolt64();
2601 		if (old != new)
2602 			(void) atomic_cas_64(&spa->spa_last_io, old, new);
2603 	}
2604 
2605 	align = 1ULL << vd->vdev_top->vdev_ashift;
2606 
2607 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
2608 	    P2PHASE(zio->io_size, align) != 0) {
2609 		/* Transform logical writes to be a full physical block size. */
2610 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
2611 		char *abuf = zio_buf_alloc(asize);
2612 		ASSERT(vd == vd->vdev_top);
2613 		if (zio->io_type == ZIO_TYPE_WRITE) {
2614 			bcopy(zio->io_data, abuf, zio->io_size);
2615 			bzero(abuf + zio->io_size, asize - zio->io_size);
2616 		}
2617 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2618 	}
2619 
2620 	/*
2621 	 * If this is not a physical io, make sure that it is properly aligned
2622 	 * before proceeding.
2623 	 */
2624 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2625 		ASSERT0(P2PHASE(zio->io_offset, align));
2626 		ASSERT0(P2PHASE(zio->io_size, align));
2627 	} else {
2628 		/*
2629 		 * For physical writes, we allow 512b aligned writes and assume
2630 		 * the device will perform a read-modify-write as necessary.
2631 		 */
2632 		ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2633 		ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2634 	}
2635 
2636 	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2637 
2638 	/*
2639 	 * If this is a repair I/O, and there's no self-healing involved --
2640 	 * that is, we're just resilvering what we expect to resilver --
2641 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2642 	 * This prevents spurious resilvering with nested replication.
2643 	 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2644 	 * A is out of date, we'll read from C+D, then use the data to
2645 	 * resilver A+B -- but we don't actually want to resilver B, just A.
2646 	 * The top-level mirror has no way to know this, so instead we just
2647 	 * discard unnecessary repairs as we work our way down the vdev tree.
2648 	 * The same logic applies to any form of nested replication:
2649 	 * ditto + mirror, RAID-Z + replacing, etc.  This covers them all.
2650 	 */
2651 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2652 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2653 	    zio->io_txg != 0 &&	/* not a delegated i/o */
2654 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2655 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2656 		zio_vdev_io_bypass(zio);
2657 		return (ZIO_PIPELINE_CONTINUE);
2658 	}
2659 
2660 	if (vd->vdev_ops->vdev_op_leaf &&
2661 	    (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2662 
2663 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
2664 			return (ZIO_PIPELINE_CONTINUE);
2665 
2666 		if ((zio = vdev_queue_io(zio)) == NULL)
2667 			return (ZIO_PIPELINE_STOP);
2668 
2669 		if (!vdev_accessible(vd, zio)) {
2670 			zio->io_error = SET_ERROR(ENXIO);
2671 			zio_interrupt(zio);
2672 			return (ZIO_PIPELINE_STOP);
2673 		}
2674 	}
2675 
2676 	vd->vdev_ops->vdev_op_io_start(zio);
2677 	return (ZIO_PIPELINE_STOP);
2678 }
2679 
2680 static int
2681 zio_vdev_io_done(zio_t *zio)
2682 {
2683 	vdev_t *vd = zio->io_vd;
2684 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2685 	boolean_t unexpected_error = B_FALSE;
2686 
2687 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2688 		return (ZIO_PIPELINE_STOP);
2689 
2690 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2691 
2692 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2693 
2694 		vdev_queue_io_done(zio);
2695 
2696 		if (zio->io_type == ZIO_TYPE_WRITE)
2697 			vdev_cache_write(zio);
2698 
2699 		if (zio_injection_enabled && zio->io_error == 0)
2700 			zio->io_error = zio_handle_device_injection(vd,
2701 			    zio, EIO);
2702 
2703 		if (zio_injection_enabled && zio->io_error == 0)
2704 			zio->io_error = zio_handle_label_injection(zio, EIO);
2705 
2706 		if (zio->io_error) {
2707 			if (!vdev_accessible(vd, zio)) {
2708 				zio->io_error = SET_ERROR(ENXIO);
2709 			} else {
2710 				unexpected_error = B_TRUE;
2711 			}
2712 		}
2713 	}
2714 
2715 	ops->vdev_op_io_done(zio);
2716 
2717 	if (unexpected_error)
2718 		VERIFY(vdev_probe(vd, zio) == NULL);
2719 
2720 	return (ZIO_PIPELINE_CONTINUE);
2721 }
2722 
2723 /*
2724  * For non-raidz ZIOs, we can just copy aside the bad data read from the
2725  * disk, and use that to finish the checksum ereport later.
2726  */
2727 static void
2728 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2729     const void *good_buf)
2730 {
2731 	/* no processing needed */
2732 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2733 }
2734 
2735 /*ARGSUSED*/
2736 void
2737 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2738 {
2739 	void *buf = zio_buf_alloc(zio->io_size);
2740 
2741 	bcopy(zio->io_data, buf, zio->io_size);
2742 
2743 	zcr->zcr_cbinfo = zio->io_size;
2744 	zcr->zcr_cbdata = buf;
2745 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
2746 	zcr->zcr_free = zio_buf_free;
2747 }
2748 
2749 static int
2750 zio_vdev_io_assess(zio_t *zio)
2751 {
2752 	vdev_t *vd = zio->io_vd;
2753 
2754 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2755 		return (ZIO_PIPELINE_STOP);
2756 
2757 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2758 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2759 
2760 	if (zio->io_vsd != NULL) {
2761 		zio->io_vsd_ops->vsd_free(zio);
2762 		zio->io_vsd = NULL;
2763 	}
2764 
2765 	if (zio_injection_enabled && zio->io_error == 0)
2766 		zio->io_error = zio_handle_fault_injection(zio, EIO);
2767 
2768 	/*
2769 	 * If the I/O failed, determine whether we should attempt to retry it.
2770 	 *
2771 	 * On retry, we cut in line in the issue queue, since we don't want
2772 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2773 	 */
2774 	if (zio->io_error && vd == NULL &&
2775 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2776 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
2777 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
2778 		zio->io_error = 0;
2779 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
2780 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2781 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2782 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2783 		    zio_requeue_io_start_cut_in_line);
2784 		return (ZIO_PIPELINE_STOP);
2785 	}
2786 
2787 	/*
2788 	 * If we got an error on a leaf device, convert it to ENXIO
2789 	 * if the device is not accessible at all.
2790 	 */
2791 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2792 	    !vdev_accessible(vd, zio))
2793 		zio->io_error = SET_ERROR(ENXIO);
2794 
2795 	/*
2796 	 * If we can't write to an interior vdev (mirror or RAID-Z),
2797 	 * set vdev_cant_write so that we stop trying to allocate from it.
2798 	 */
2799 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2800 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2801 		vd->vdev_cant_write = B_TRUE;
2802 	}
2803 
2804 	if (zio->io_error)
2805 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2806 
2807 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2808 	    zio->io_physdone != NULL) {
2809 		ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2810 		ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2811 		zio->io_physdone(zio->io_logical);
2812 	}
2813 
2814 	return (ZIO_PIPELINE_CONTINUE);
2815 }
2816 
2817 void
2818 zio_vdev_io_reissue(zio_t *zio)
2819 {
2820 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2821 	ASSERT(zio->io_error == 0);
2822 
2823 	zio->io_stage >>= 1;
2824 }
2825 
2826 void
2827 zio_vdev_io_redone(zio_t *zio)
2828 {
2829 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2830 
2831 	zio->io_stage >>= 1;
2832 }
2833 
2834 void
2835 zio_vdev_io_bypass(zio_t *zio)
2836 {
2837 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2838 	ASSERT(zio->io_error == 0);
2839 
2840 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2841 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2842 }
2843 
2844 /*
2845  * ==========================================================================
2846  * Generate and verify checksums
2847  * ==========================================================================
2848  */
2849 static int
2850 zio_checksum_generate(zio_t *zio)
2851 {
2852 	blkptr_t *bp = zio->io_bp;
2853 	enum zio_checksum checksum;
2854 
2855 	if (bp == NULL) {
2856 		/*
2857 		 * This is zio_write_phys().
2858 		 * We're either generating a label checksum, or none at all.
2859 		 */
2860 		checksum = zio->io_prop.zp_checksum;
2861 
2862 		if (checksum == ZIO_CHECKSUM_OFF)
2863 			return (ZIO_PIPELINE_CONTINUE);
2864 
2865 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2866 	} else {
2867 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2868 			ASSERT(!IO_IS_ALLOCATING(zio));
2869 			checksum = ZIO_CHECKSUM_GANG_HEADER;
2870 		} else {
2871 			checksum = BP_GET_CHECKSUM(bp);
2872 		}
2873 	}
2874 
2875 	zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2876 
2877 	return (ZIO_PIPELINE_CONTINUE);
2878 }
2879 
2880 static int
2881 zio_checksum_verify(zio_t *zio)
2882 {
2883 	zio_bad_cksum_t info;
2884 	blkptr_t *bp = zio->io_bp;
2885 	int error;
2886 
2887 	ASSERT(zio->io_vd != NULL);
2888 
2889 	if (bp == NULL) {
2890 		/*
2891 		 * This is zio_read_phys().
2892 		 * We're either verifying a label checksum, or nothing at all.
2893 		 */
2894 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2895 			return (ZIO_PIPELINE_CONTINUE);
2896 
2897 		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2898 	}
2899 
2900 	if ((error = zio_checksum_error(zio, &info)) != 0) {
2901 		zio->io_error = error;
2902 		if (error == ECKSUM &&
2903 		    !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2904 			zfs_ereport_start_checksum(zio->io_spa,
2905 			    zio->io_vd, zio, zio->io_offset,
2906 			    zio->io_size, NULL, &info);
2907 		}
2908 	}
2909 
2910 	return (ZIO_PIPELINE_CONTINUE);
2911 }
2912 
2913 /*
2914  * Called by RAID-Z to ensure we don't compute the checksum twice.
2915  */
2916 void
2917 zio_checksum_verified(zio_t *zio)
2918 {
2919 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2920 }
2921 
2922 /*
2923  * ==========================================================================
2924  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2925  * An error of 0 indicates success.  ENXIO indicates whole-device failure,
2926  * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
2927  * indicate errors that are specific to one I/O, and most likely permanent.
2928  * Any other error is presumed to be worse because we weren't expecting it.
2929  * ==========================================================================
2930  */
2931 int
2932 zio_worst_error(int e1, int e2)
2933 {
2934 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2935 	int r1, r2;
2936 
2937 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2938 		if (e1 == zio_error_rank[r1])
2939 			break;
2940 
2941 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2942 		if (e2 == zio_error_rank[r2])
2943 			break;
2944 
2945 	return (r1 > r2 ? e1 : e2);
2946 }
2947 
2948 /*
2949  * ==========================================================================
2950  * I/O completion
2951  * ==========================================================================
2952  */
2953 static int
2954 zio_ready(zio_t *zio)
2955 {
2956 	blkptr_t *bp = zio->io_bp;
2957 	zio_t *pio, *pio_next;
2958 
2959 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2960 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2961 		return (ZIO_PIPELINE_STOP);
2962 
2963 	if (zio->io_ready) {
2964 		ASSERT(IO_IS_ALLOCATING(zio));
2965 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2966 		    (zio->io_flags & ZIO_FLAG_NOPWRITE));
2967 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2968 
2969 		zio->io_ready(zio);
2970 	}
2971 
2972 	if (bp != NULL && bp != &zio->io_bp_copy)
2973 		zio->io_bp_copy = *bp;
2974 
2975 	if (zio->io_error)
2976 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2977 
2978 	mutex_enter(&zio->io_lock);
2979 	zio->io_state[ZIO_WAIT_READY] = 1;
2980 	pio = zio_walk_parents(zio);
2981 	mutex_exit(&zio->io_lock);
2982 
2983 	/*
2984 	 * As we notify zio's parents, new parents could be added.
2985 	 * New parents go to the head of zio's io_parent_list, however,
2986 	 * so we will (correctly) not notify them.  The remainder of zio's
2987 	 * io_parent_list, from 'pio_next' onward, cannot change because
2988 	 * all parents must wait for us to be done before they can be done.
2989 	 */
2990 	for (; pio != NULL; pio = pio_next) {
2991 		pio_next = zio_walk_parents(zio);
2992 		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2993 	}
2994 
2995 	if (zio->io_flags & ZIO_FLAG_NODATA) {
2996 		if (BP_IS_GANG(bp)) {
2997 			zio->io_flags &= ~ZIO_FLAG_NODATA;
2998 		} else {
2999 			ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3000 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3001 		}
3002 	}
3003 
3004 	if (zio_injection_enabled &&
3005 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
3006 		zio_handle_ignored_writes(zio);
3007 
3008 	return (ZIO_PIPELINE_CONTINUE);
3009 }
3010 
3011 static int
3012 zio_done(zio_t *zio)
3013 {
3014 	spa_t *spa = zio->io_spa;
3015 	zio_t *lio = zio->io_logical;
3016 	blkptr_t *bp = zio->io_bp;
3017 	vdev_t *vd = zio->io_vd;
3018 	uint64_t psize = zio->io_size;
3019 	zio_t *pio, *pio_next;
3020 
3021 	/*
3022 	 * If our children haven't all completed,
3023 	 * wait for them and then repeat this pipeline stage.
3024 	 */
3025 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3026 	    zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3027 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3028 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3029 		return (ZIO_PIPELINE_STOP);
3030 
3031 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3032 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3033 			ASSERT(zio->io_children[c][w] == 0);
3034 
3035 	if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3036 		ASSERT(bp->blk_pad[0] == 0);
3037 		ASSERT(bp->blk_pad[1] == 0);
3038 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3039 		    (bp == zio_unique_parent(zio)->io_bp));
3040 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3041 		    zio->io_bp_override == NULL &&
3042 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3043 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
3044 			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3045 			ASSERT(BP_COUNT_GANG(bp) == 0 ||
3046 			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3047 		}
3048 		if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3049 			VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3050 	}
3051 
3052 	/*
3053 	 * If there were child vdev/gang/ddt errors, they apply to us now.
3054 	 */
3055 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3056 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3057 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3058 
3059 	/*
3060 	 * If the I/O on the transformed data was successful, generate any
3061 	 * checksum reports now while we still have the transformed data.
3062 	 */
3063 	if (zio->io_error == 0) {
3064 		while (zio->io_cksum_report != NULL) {
3065 			zio_cksum_report_t *zcr = zio->io_cksum_report;
3066 			uint64_t align = zcr->zcr_align;
3067 			uint64_t asize = P2ROUNDUP(psize, align);
3068 			char *abuf = zio->io_data;
3069 
3070 			if (asize != psize) {
3071 				abuf = zio_buf_alloc(asize);
3072 				bcopy(zio->io_data, abuf, psize);
3073 				bzero(abuf + psize, asize - psize);
3074 			}
3075 
3076 			zio->io_cksum_report = zcr->zcr_next;
3077 			zcr->zcr_next = NULL;
3078 			zcr->zcr_finish(zcr, abuf);
3079 			zfs_ereport_free_checksum(zcr);
3080 
3081 			if (asize != psize)
3082 				zio_buf_free(abuf, asize);
3083 		}
3084 	}
3085 
3086 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
3087 
3088 	vdev_stat_update(zio, psize);
3089 
3090 	if (zio->io_error) {
3091 		/*
3092 		 * If this I/O is attached to a particular vdev,
3093 		 * generate an error message describing the I/O failure
3094 		 * at the block level.  We ignore these errors if the
3095 		 * device is currently unavailable.
3096 		 */
3097 		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3098 			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3099 
3100 		if ((zio->io_error == EIO || !(zio->io_flags &
3101 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3102 		    zio == lio) {
3103 			/*
3104 			 * For logical I/O requests, tell the SPA to log the
3105 			 * error and generate a logical data ereport.
3106 			 */
3107 			spa_log_error(spa, zio);
3108 			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3109 			    0, 0);
3110 		}
3111 	}
3112 
3113 	if (zio->io_error && zio == lio) {
3114 		/*
3115 		 * Determine whether zio should be reexecuted.  This will
3116 		 * propagate all the way to the root via zio_notify_parent().
3117 		 */
3118 		ASSERT(vd == NULL && bp != NULL);
3119 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3120 
3121 		if (IO_IS_ALLOCATING(zio) &&
3122 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3123 			if (zio->io_error != ENOSPC)
3124 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3125 			else
3126 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3127 		}
3128 
3129 		if ((zio->io_type == ZIO_TYPE_READ ||
3130 		    zio->io_type == ZIO_TYPE_FREE) &&
3131 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3132 		    zio->io_error == ENXIO &&
3133 		    spa_load_state(spa) == SPA_LOAD_NONE &&
3134 		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3135 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3136 
3137 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3138 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3139 
3140 		/*
3141 		 * Here is a possibly good place to attempt to do
3142 		 * either combinatorial reconstruction or error correction
3143 		 * based on checksums.  It also might be a good place
3144 		 * to send out preliminary ereports before we suspend
3145 		 * processing.
3146 		 */
3147 	}
3148 
3149 	/*
3150 	 * If there were logical child errors, they apply to us now.
3151 	 * We defer this until now to avoid conflating logical child
3152 	 * errors with errors that happened to the zio itself when
3153 	 * updating vdev stats and reporting FMA events above.
3154 	 */
3155 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3156 
3157 	if ((zio->io_error || zio->io_reexecute) &&
3158 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3159 	    !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3160 		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3161 
3162 	zio_gang_tree_free(&zio->io_gang_tree);
3163 
3164 	/*
3165 	 * Godfather I/Os should never suspend.
3166 	 */
3167 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3168 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3169 		zio->io_reexecute = 0;
3170 
3171 	if (zio->io_reexecute) {
3172 		/*
3173 		 * This is a logical I/O that wants to reexecute.
3174 		 *
3175 		 * Reexecute is top-down.  When an i/o fails, if it's not
3176 		 * the root, it simply notifies its parent and sticks around.
3177 		 * The parent, seeing that it still has children in zio_done(),
3178 		 * does the same.  This percolates all the way up to the root.
3179 		 * The root i/o will reexecute or suspend the entire tree.
3180 		 *
3181 		 * This approach ensures that zio_reexecute() honors
3182 		 * all the original i/o dependency relationships, e.g.
3183 		 * parents not executing until children are ready.
3184 		 */
3185 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3186 
3187 		zio->io_gang_leader = NULL;
3188 
3189 		mutex_enter(&zio->io_lock);
3190 		zio->io_state[ZIO_WAIT_DONE] = 1;
3191 		mutex_exit(&zio->io_lock);
3192 
3193 		/*
3194 		 * "The Godfather" I/O monitors its children but is
3195 		 * not a true parent to them. It will track them through
3196 		 * the pipeline but severs its ties whenever they get into
3197 		 * trouble (e.g. suspended). This allows "The Godfather"
3198 		 * I/O to return status without blocking.
3199 		 */
3200 		for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3201 			zio_link_t *zl = zio->io_walk_link;
3202 			pio_next = zio_walk_parents(zio);
3203 
3204 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3205 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3206 				zio_remove_child(pio, zio, zl);
3207 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3208 			}
3209 		}
3210 
3211 		if ((pio = zio_unique_parent(zio)) != NULL) {
3212 			/*
3213 			 * We're not a root i/o, so there's nothing to do
3214 			 * but notify our parent.  Don't propagate errors
3215 			 * upward since we haven't permanently failed yet.
3216 			 */
3217 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3218 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3219 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3220 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3221 			/*
3222 			 * We'd fail again if we reexecuted now, so suspend
3223 			 * until conditions improve (e.g. device comes online).
3224 			 */
3225 			zio_suspend(spa, zio);
3226 		} else {
3227 			/*
3228 			 * Reexecution is potentially a huge amount of work.
3229 			 * Hand it off to the otherwise-unused claim taskq.
3230 			 */
3231 			ASSERT(zio->io_tqent.tqent_next == NULL);
3232 			spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3233 			    ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3234 			    0, &zio->io_tqent);
3235 		}
3236 		return (ZIO_PIPELINE_STOP);
3237 	}
3238 
3239 	ASSERT(zio->io_child_count == 0);
3240 	ASSERT(zio->io_reexecute == 0);
3241 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3242 
3243 	/*
3244 	 * Report any checksum errors, since the I/O is complete.
3245 	 */
3246 	while (zio->io_cksum_report != NULL) {
3247 		zio_cksum_report_t *zcr = zio->io_cksum_report;
3248 		zio->io_cksum_report = zcr->zcr_next;
3249 		zcr->zcr_next = NULL;
3250 		zcr->zcr_finish(zcr, NULL);
3251 		zfs_ereport_free_checksum(zcr);
3252 	}
3253 
3254 	/*
3255 	 * It is the responsibility of the done callback to ensure that this
3256 	 * particular zio is no longer discoverable for adoption, and as
3257 	 * such, cannot acquire any new parents.
3258 	 */
3259 	if (zio->io_done)
3260 		zio->io_done(zio);
3261 
3262 	mutex_enter(&zio->io_lock);
3263 	zio->io_state[ZIO_WAIT_DONE] = 1;
3264 	mutex_exit(&zio->io_lock);
3265 
3266 	for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3267 		zio_link_t *zl = zio->io_walk_link;
3268 		pio_next = zio_walk_parents(zio);
3269 		zio_remove_child(pio, zio, zl);
3270 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3271 	}
3272 
3273 	if (zio->io_waiter != NULL) {
3274 		mutex_enter(&zio->io_lock);
3275 		zio->io_executor = NULL;
3276 		cv_broadcast(&zio->io_cv);
3277 		mutex_exit(&zio->io_lock);
3278 	} else {
3279 		zio_destroy(zio);
3280 	}
3281 
3282 	return (ZIO_PIPELINE_STOP);
3283 }
3284 
3285 /*
3286  * ==========================================================================
3287  * I/O pipeline definition
3288  * ==========================================================================
3289  */
3290 static zio_pipe_stage_t *zio_pipeline[] = {
3291 	NULL,
3292 	zio_read_bp_init,
3293 	zio_free_bp_init,
3294 	zio_issue_async,
3295 	zio_write_bp_init,
3296 	zio_checksum_generate,
3297 	zio_nop_write,
3298 	zio_ddt_read_start,
3299 	zio_ddt_read_done,
3300 	zio_ddt_write,
3301 	zio_ddt_free,
3302 	zio_gang_assemble,
3303 	zio_gang_issue,
3304 	zio_dva_allocate,
3305 	zio_dva_free,
3306 	zio_dva_claim,
3307 	zio_ready,
3308 	zio_vdev_io_start,
3309 	zio_vdev_io_done,
3310 	zio_vdev_io_assess,
3311 	zio_checksum_verify,
3312 	zio_done
3313 };
3314 
3315 /* dnp is the dnode for zb1->zb_object */
3316 boolean_t
3317 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1,
3318     const zbookmark_phys_t *zb2)
3319 {
3320 	uint64_t zb1nextL0, zb2thisobj;
3321 
3322 	ASSERT(zb1->zb_objset == zb2->zb_objset);
3323 	ASSERT(zb2->zb_level == 0);
3324 
3325 	/* The objset_phys_t isn't before anything. */
3326 	if (dnp == NULL)
3327 		return (B_FALSE);
3328 
3329 	zb1nextL0 = (zb1->zb_blkid + 1) <<
3330 	    ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3331 
3332 	zb2thisobj = zb2->zb_object ? zb2->zb_object :
3333 	    zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3334 
3335 	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3336 		uint64_t nextobj = zb1nextL0 *
3337 		    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3338 		return (nextobj <= zb2thisobj);
3339 	}
3340 
3341 	if (zb1->zb_object < zb2thisobj)
3342 		return (B_TRUE);
3343 	if (zb1->zb_object > zb2thisobj)
3344 		return (B_FALSE);
3345 	if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3346 		return (B_FALSE);
3347 	return (zb1nextL0 <= zb2->zb_blkid);
3348 }
3349