xref: /titanic_51/usr/src/uts/common/fs/zfs/zio.c (revision dc32d872cbeb56532bcea030255db9cd79bac7da)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
24  * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25  */
26 
27 #include <sys/sysmacros.h>
28 #include <sys/zfs_context.h>
29 #include <sys/fm/fs/zfs.h>
30 #include <sys/spa.h>
31 #include <sys/txg.h>
32 #include <sys/spa_impl.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio_impl.h>
35 #include <sys/zio_compress.h>
36 #include <sys/zio_checksum.h>
37 #include <sys/dmu_objset.h>
38 #include <sys/arc.h>
39 #include <sys/ddt.h>
40 #include <sys/blkptr.h>
41 #include <sys/zfeature.h>
42 
43 /*
44  * ==========================================================================
45  * I/O type descriptions
46  * ==========================================================================
47  */
48 const char *zio_type_name[ZIO_TYPES] = {
49 	"zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
50 	"zio_ioctl"
51 };
52 
53 /*
54  * ==========================================================================
55  * I/O kmem caches
56  * ==========================================================================
57  */
58 kmem_cache_t *zio_cache;
59 kmem_cache_t *zio_link_cache;
60 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
61 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
62 
63 #ifdef _KERNEL
64 extern vmem_t *zio_alloc_arena;
65 #endif
66 
67 #define	ZIO_PIPELINE_CONTINUE		0x100
68 #define	ZIO_PIPELINE_STOP		0x101
69 
70 #define	BP_SPANB(indblkshift, level) \
71 	(((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
72 #define	COMPARE_META_LEVEL	0x80000000ul
73 /*
74  * The following actions directly effect the spa's sync-to-convergence logic.
75  * The values below define the sync pass when we start performing the action.
76  * Care should be taken when changing these values as they directly impact
77  * spa_sync() performance. Tuning these values may introduce subtle performance
78  * pathologies and should only be done in the context of performance analysis.
79  * These tunables will eventually be removed and replaced with #defines once
80  * enough analysis has been done to determine optimal values.
81  *
82  * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
83  * regular blocks are not deferred.
84  */
85 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
86 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
87 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
88 
89 /*
90  * An allocating zio is one that either currently has the DVA allocate
91  * stage set or will have it later in its lifetime.
92  */
93 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
94 
95 boolean_t	zio_requeue_io_start_cut_in_line = B_TRUE;
96 
97 #ifdef ZFS_DEBUG
98 int zio_buf_debug_limit = 16384;
99 #else
100 int zio_buf_debug_limit = 0;
101 #endif
102 
103 void
104 zio_init(void)
105 {
106 	size_t c;
107 	vmem_t *data_alloc_arena = NULL;
108 
109 #ifdef _KERNEL
110 	data_alloc_arena = zio_alloc_arena;
111 #endif
112 	zio_cache = kmem_cache_create("zio_cache",
113 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
114 	zio_link_cache = kmem_cache_create("zio_link_cache",
115 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
116 
117 	/*
118 	 * For small buffers, we want a cache for each multiple of
119 	 * SPA_MINBLOCKSIZE.  For larger buffers, we want a cache
120 	 * for each quarter-power of 2.
121 	 */
122 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
123 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
124 		size_t p2 = size;
125 		size_t align = 0;
126 		size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
127 
128 		while (!ISP2(p2))
129 			p2 &= p2 - 1;
130 
131 #ifndef _KERNEL
132 		/*
133 		 * If we are using watchpoints, put each buffer on its own page,
134 		 * to eliminate the performance overhead of trapping to the
135 		 * kernel when modifying a non-watched buffer that shares the
136 		 * page with a watched buffer.
137 		 */
138 		if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
139 			continue;
140 #endif
141 		if (size <= 4 * SPA_MINBLOCKSIZE) {
142 			align = SPA_MINBLOCKSIZE;
143 		} else if (IS_P2ALIGNED(size, p2 >> 2)) {
144 			align = MIN(p2 >> 2, PAGESIZE);
145 		}
146 
147 		if (align != 0) {
148 			char name[36];
149 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
150 			zio_buf_cache[c] = kmem_cache_create(name, size,
151 			    align, NULL, NULL, NULL, NULL, NULL, cflags);
152 
153 			/*
154 			 * Since zio_data bufs do not appear in crash dumps, we
155 			 * pass KMC_NOTOUCH so that no allocator metadata is
156 			 * stored with the buffers.
157 			 */
158 			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
159 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
160 			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
161 			    cflags | KMC_NOTOUCH);
162 		}
163 	}
164 
165 	while (--c != 0) {
166 		ASSERT(zio_buf_cache[c] != NULL);
167 		if (zio_buf_cache[c - 1] == NULL)
168 			zio_buf_cache[c - 1] = zio_buf_cache[c];
169 
170 		ASSERT(zio_data_buf_cache[c] != NULL);
171 		if (zio_data_buf_cache[c - 1] == NULL)
172 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
173 	}
174 
175 	zio_inject_init();
176 }
177 
178 void
179 zio_fini(void)
180 {
181 	size_t c;
182 	kmem_cache_t *last_cache = NULL;
183 	kmem_cache_t *last_data_cache = NULL;
184 
185 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
186 		if (zio_buf_cache[c] != last_cache) {
187 			last_cache = zio_buf_cache[c];
188 			kmem_cache_destroy(zio_buf_cache[c]);
189 		}
190 		zio_buf_cache[c] = NULL;
191 
192 		if (zio_data_buf_cache[c] != last_data_cache) {
193 			last_data_cache = zio_data_buf_cache[c];
194 			kmem_cache_destroy(zio_data_buf_cache[c]);
195 		}
196 		zio_data_buf_cache[c] = NULL;
197 	}
198 
199 	kmem_cache_destroy(zio_link_cache);
200 	kmem_cache_destroy(zio_cache);
201 
202 	zio_inject_fini();
203 }
204 
205 /*
206  * ==========================================================================
207  * Allocate and free I/O buffers
208  * ==========================================================================
209  */
210 
211 /*
212  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
213  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
214  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
215  * excess / transient data in-core during a crashdump.
216  */
217 void *
218 zio_buf_alloc(size_t size)
219 {
220 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
221 
222 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
223 
224 	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
225 }
226 
227 /*
228  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
229  * crashdump if the kernel panics.  This exists so that we will limit the amount
230  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
231  * of kernel heap dumped to disk when the kernel panics)
232  */
233 void *
234 zio_data_buf_alloc(size_t size)
235 {
236 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
237 
238 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
239 
240 	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
241 }
242 
243 void
244 zio_buf_free(void *buf, size_t size)
245 {
246 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
247 
248 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
249 
250 	kmem_cache_free(zio_buf_cache[c], buf);
251 }
252 
253 void
254 zio_data_buf_free(void *buf, size_t size)
255 {
256 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
257 
258 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
259 
260 	kmem_cache_free(zio_data_buf_cache[c], buf);
261 }
262 
263 /*
264  * ==========================================================================
265  * Push and pop I/O transform buffers
266  * ==========================================================================
267  */
268 static void
269 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
270     zio_transform_func_t *transform)
271 {
272 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
273 
274 	zt->zt_orig_data = zio->io_data;
275 	zt->zt_orig_size = zio->io_size;
276 	zt->zt_bufsize = bufsize;
277 	zt->zt_transform = transform;
278 
279 	zt->zt_next = zio->io_transform_stack;
280 	zio->io_transform_stack = zt;
281 
282 	zio->io_data = data;
283 	zio->io_size = size;
284 }
285 
286 static void
287 zio_pop_transforms(zio_t *zio)
288 {
289 	zio_transform_t *zt;
290 
291 	while ((zt = zio->io_transform_stack) != NULL) {
292 		if (zt->zt_transform != NULL)
293 			zt->zt_transform(zio,
294 			    zt->zt_orig_data, zt->zt_orig_size);
295 
296 		if (zt->zt_bufsize != 0)
297 			zio_buf_free(zio->io_data, zt->zt_bufsize);
298 
299 		zio->io_data = zt->zt_orig_data;
300 		zio->io_size = zt->zt_orig_size;
301 		zio->io_transform_stack = zt->zt_next;
302 
303 		kmem_free(zt, sizeof (zio_transform_t));
304 	}
305 }
306 
307 /*
308  * ==========================================================================
309  * I/O transform callbacks for subblocks and decompression
310  * ==========================================================================
311  */
312 static void
313 zio_subblock(zio_t *zio, void *data, uint64_t size)
314 {
315 	ASSERT(zio->io_size > size);
316 
317 	if (zio->io_type == ZIO_TYPE_READ)
318 		bcopy(zio->io_data, data, size);
319 }
320 
321 static void
322 zio_decompress(zio_t *zio, void *data, uint64_t size)
323 {
324 	if (zio->io_error == 0 &&
325 	    zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
326 	    zio->io_data, data, zio->io_size, size) != 0)
327 		zio->io_error = SET_ERROR(EIO);
328 }
329 
330 /*
331  * ==========================================================================
332  * I/O parent/child relationships and pipeline interlocks
333  * ==========================================================================
334  */
335 /*
336  * NOTE - Callers to zio_walk_parents() and zio_walk_children must
337  *        continue calling these functions until they return NULL.
338  *        Otherwise, the next caller will pick up the list walk in
339  *        some indeterminate state.  (Otherwise every caller would
340  *        have to pass in a cookie to keep the state represented by
341  *        io_walk_link, which gets annoying.)
342  */
343 zio_t *
344 zio_walk_parents(zio_t *cio)
345 {
346 	zio_link_t *zl = cio->io_walk_link;
347 	list_t *pl = &cio->io_parent_list;
348 
349 	zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
350 	cio->io_walk_link = zl;
351 
352 	if (zl == NULL)
353 		return (NULL);
354 
355 	ASSERT(zl->zl_child == cio);
356 	return (zl->zl_parent);
357 }
358 
359 zio_t *
360 zio_walk_children(zio_t *pio)
361 {
362 	zio_link_t *zl = pio->io_walk_link;
363 	list_t *cl = &pio->io_child_list;
364 
365 	zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
366 	pio->io_walk_link = zl;
367 
368 	if (zl == NULL)
369 		return (NULL);
370 
371 	ASSERT(zl->zl_parent == pio);
372 	return (zl->zl_child);
373 }
374 
375 zio_t *
376 zio_unique_parent(zio_t *cio)
377 {
378 	zio_t *pio = zio_walk_parents(cio);
379 
380 	VERIFY(zio_walk_parents(cio) == NULL);
381 	return (pio);
382 }
383 
384 void
385 zio_add_child(zio_t *pio, zio_t *cio)
386 {
387 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
388 
389 	/*
390 	 * Logical I/Os can have logical, gang, or vdev children.
391 	 * Gang I/Os can have gang or vdev children.
392 	 * Vdev I/Os can only have vdev children.
393 	 * The following ASSERT captures all of these constraints.
394 	 */
395 	ASSERT(cio->io_child_type <= pio->io_child_type);
396 
397 	zl->zl_parent = pio;
398 	zl->zl_child = cio;
399 
400 	mutex_enter(&cio->io_lock);
401 	mutex_enter(&pio->io_lock);
402 
403 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
404 
405 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
406 		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
407 
408 	list_insert_head(&pio->io_child_list, zl);
409 	list_insert_head(&cio->io_parent_list, zl);
410 
411 	pio->io_child_count++;
412 	cio->io_parent_count++;
413 
414 	mutex_exit(&pio->io_lock);
415 	mutex_exit(&cio->io_lock);
416 }
417 
418 static void
419 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
420 {
421 	ASSERT(zl->zl_parent == pio);
422 	ASSERT(zl->zl_child == cio);
423 
424 	mutex_enter(&cio->io_lock);
425 	mutex_enter(&pio->io_lock);
426 
427 	list_remove(&pio->io_child_list, zl);
428 	list_remove(&cio->io_parent_list, zl);
429 
430 	pio->io_child_count--;
431 	cio->io_parent_count--;
432 
433 	mutex_exit(&pio->io_lock);
434 	mutex_exit(&cio->io_lock);
435 
436 	kmem_cache_free(zio_link_cache, zl);
437 }
438 
439 static boolean_t
440 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
441 {
442 	uint64_t *countp = &zio->io_children[child][wait];
443 	boolean_t waiting = B_FALSE;
444 
445 	mutex_enter(&zio->io_lock);
446 	ASSERT(zio->io_stall == NULL);
447 	if (*countp != 0) {
448 		zio->io_stage >>= 1;
449 		zio->io_stall = countp;
450 		waiting = B_TRUE;
451 	}
452 	mutex_exit(&zio->io_lock);
453 
454 	return (waiting);
455 }
456 
457 static void
458 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
459 {
460 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
461 	int *errorp = &pio->io_child_error[zio->io_child_type];
462 
463 	mutex_enter(&pio->io_lock);
464 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
465 		*errorp = zio_worst_error(*errorp, zio->io_error);
466 	pio->io_reexecute |= zio->io_reexecute;
467 	ASSERT3U(*countp, >, 0);
468 
469 	(*countp)--;
470 
471 	if (*countp == 0 && pio->io_stall == countp) {
472 		pio->io_stall = NULL;
473 		mutex_exit(&pio->io_lock);
474 		zio_execute(pio);
475 	} else {
476 		mutex_exit(&pio->io_lock);
477 	}
478 }
479 
480 static void
481 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
482 {
483 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
484 		zio->io_error = zio->io_child_error[c];
485 }
486 
487 /*
488  * ==========================================================================
489  * Create the various types of I/O (read, write, free, etc)
490  * ==========================================================================
491  */
492 static zio_t *
493 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
494     void *data, uint64_t size, zio_done_func_t *done, void *private,
495     zio_type_t type, zio_priority_t priority, enum zio_flag flags,
496     vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
497     enum zio_stage stage, enum zio_stage pipeline)
498 {
499 	zio_t *zio;
500 
501 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
502 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
503 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
504 
505 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
506 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
507 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
508 
509 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
510 	bzero(zio, sizeof (zio_t));
511 
512 	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
513 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
514 
515 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
516 	    offsetof(zio_link_t, zl_parent_node));
517 	list_create(&zio->io_child_list, sizeof (zio_link_t),
518 	    offsetof(zio_link_t, zl_child_node));
519 
520 	if (vd != NULL)
521 		zio->io_child_type = ZIO_CHILD_VDEV;
522 	else if (flags & ZIO_FLAG_GANG_CHILD)
523 		zio->io_child_type = ZIO_CHILD_GANG;
524 	else if (flags & ZIO_FLAG_DDT_CHILD)
525 		zio->io_child_type = ZIO_CHILD_DDT;
526 	else
527 		zio->io_child_type = ZIO_CHILD_LOGICAL;
528 
529 	if (bp != NULL) {
530 		zio->io_bp = (blkptr_t *)bp;
531 		zio->io_bp_copy = *bp;
532 		zio->io_bp_orig = *bp;
533 		if (type != ZIO_TYPE_WRITE ||
534 		    zio->io_child_type == ZIO_CHILD_DDT)
535 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
536 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
537 			zio->io_logical = zio;
538 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
539 			pipeline |= ZIO_GANG_STAGES;
540 	}
541 
542 	zio->io_spa = spa;
543 	zio->io_txg = txg;
544 	zio->io_done = done;
545 	zio->io_private = private;
546 	zio->io_type = type;
547 	zio->io_priority = priority;
548 	zio->io_vd = vd;
549 	zio->io_offset = offset;
550 	zio->io_orig_data = zio->io_data = data;
551 	zio->io_orig_size = zio->io_size = size;
552 	zio->io_orig_flags = zio->io_flags = flags;
553 	zio->io_orig_stage = zio->io_stage = stage;
554 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
555 
556 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
557 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
558 
559 	if (zb != NULL)
560 		zio->io_bookmark = *zb;
561 
562 	if (pio != NULL) {
563 		if (zio->io_logical == NULL)
564 			zio->io_logical = pio->io_logical;
565 		if (zio->io_child_type == ZIO_CHILD_GANG)
566 			zio->io_gang_leader = pio->io_gang_leader;
567 		zio_add_child(pio, zio);
568 	}
569 
570 	return (zio);
571 }
572 
573 static void
574 zio_destroy(zio_t *zio)
575 {
576 	list_destroy(&zio->io_parent_list);
577 	list_destroy(&zio->io_child_list);
578 	mutex_destroy(&zio->io_lock);
579 	cv_destroy(&zio->io_cv);
580 	kmem_cache_free(zio_cache, zio);
581 }
582 
583 zio_t *
584 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
585     void *private, enum zio_flag flags)
586 {
587 	zio_t *zio;
588 
589 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
590 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
591 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
592 
593 	return (zio);
594 }
595 
596 zio_t *
597 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
598 {
599 	return (zio_null(NULL, spa, NULL, done, private, flags));
600 }
601 
602 void
603 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
604 {
605 	if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
606 		zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
607 		    bp, (longlong_t)BP_GET_TYPE(bp));
608 	}
609 	if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
610 	    BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
611 		zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
612 		    bp, (longlong_t)BP_GET_CHECKSUM(bp));
613 	}
614 	if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
615 	    BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
616 		zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
617 		    bp, (longlong_t)BP_GET_COMPRESS(bp));
618 	}
619 	if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
620 		zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
621 		    bp, (longlong_t)BP_GET_LSIZE(bp));
622 	}
623 	if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
624 		zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
625 		    bp, (longlong_t)BP_GET_PSIZE(bp));
626 	}
627 
628 	if (BP_IS_EMBEDDED(bp)) {
629 		if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
630 			zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
631 			    bp, (longlong_t)BPE_GET_ETYPE(bp));
632 		}
633 	}
634 
635 	/*
636 	 * Pool-specific checks.
637 	 *
638 	 * Note: it would be nice to verify that the blk_birth and
639 	 * BP_PHYSICAL_BIRTH() are not too large.  However, spa_freeze()
640 	 * allows the birth time of log blocks (and dmu_sync()-ed blocks
641 	 * that are in the log) to be arbitrarily large.
642 	 */
643 	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
644 		uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
645 		if (vdevid >= spa->spa_root_vdev->vdev_children) {
646 			zfs_panic_recover("blkptr at %p DVA %u has invalid "
647 			    "VDEV %llu",
648 			    bp, i, (longlong_t)vdevid);
649 			continue;
650 		}
651 		vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
652 		if (vd == NULL) {
653 			zfs_panic_recover("blkptr at %p DVA %u has invalid "
654 			    "VDEV %llu",
655 			    bp, i, (longlong_t)vdevid);
656 			continue;
657 		}
658 		if (vd->vdev_ops == &vdev_hole_ops) {
659 			zfs_panic_recover("blkptr at %p DVA %u has hole "
660 			    "VDEV %llu",
661 			    bp, i, (longlong_t)vdevid);
662 			continue;
663 		}
664 		if (vd->vdev_ops == &vdev_missing_ops) {
665 			/*
666 			 * "missing" vdevs are valid during import, but we
667 			 * don't have their detailed info (e.g. asize), so
668 			 * we can't perform any more checks on them.
669 			 */
670 			continue;
671 		}
672 		uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
673 		uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
674 		if (BP_IS_GANG(bp))
675 			asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
676 		if (offset + asize > vd->vdev_asize) {
677 			zfs_panic_recover("blkptr at %p DVA %u has invalid "
678 			    "OFFSET %llu",
679 			    bp, i, (longlong_t)offset);
680 		}
681 	}
682 }
683 
684 zio_t *
685 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
686     void *data, uint64_t size, zio_done_func_t *done, void *private,
687     zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
688 {
689 	zio_t *zio;
690 
691 	zfs_blkptr_verify(spa, bp);
692 
693 	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
694 	    data, size, done, private,
695 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
696 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
697 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
698 
699 	return (zio);
700 }
701 
702 zio_t *
703 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
704     void *data, uint64_t size, const zio_prop_t *zp,
705     zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
706     void *private,
707     zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
708 {
709 	zio_t *zio;
710 
711 	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
712 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
713 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
714 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
715 	    DMU_OT_IS_VALID(zp->zp_type) &&
716 	    zp->zp_level < 32 &&
717 	    zp->zp_copies > 0 &&
718 	    zp->zp_copies <= spa_max_replication(spa));
719 
720 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
721 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
722 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
723 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
724 
725 	zio->io_ready = ready;
726 	zio->io_physdone = physdone;
727 	zio->io_prop = *zp;
728 
729 	/*
730 	 * Data can be NULL if we are going to call zio_write_override() to
731 	 * provide the already-allocated BP.  But we may need the data to
732 	 * verify a dedup hit (if requested).  In this case, don't try to
733 	 * dedup (just take the already-allocated BP verbatim).
734 	 */
735 	if (data == NULL && zio->io_prop.zp_dedup_verify) {
736 		zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
737 	}
738 
739 	return (zio);
740 }
741 
742 zio_t *
743 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
744     uint64_t size, zio_done_func_t *done, void *private,
745     zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
746 {
747 	zio_t *zio;
748 
749 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
750 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
751 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
752 
753 	return (zio);
754 }
755 
756 void
757 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
758 {
759 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
760 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
761 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
762 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
763 
764 	/*
765 	 * We must reset the io_prop to match the values that existed
766 	 * when the bp was first written by dmu_sync() keeping in mind
767 	 * that nopwrite and dedup are mutually exclusive.
768 	 */
769 	zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
770 	zio->io_prop.zp_nopwrite = nopwrite;
771 	zio->io_prop.zp_copies = copies;
772 	zio->io_bp_override = bp;
773 }
774 
775 void
776 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
777 {
778 
779 	/*
780 	 * The check for EMBEDDED is a performance optimization.  We
781 	 * process the free here (by ignoring it) rather than
782 	 * putting it on the list and then processing it in zio_free_sync().
783 	 */
784 	if (BP_IS_EMBEDDED(bp))
785 		return;
786 	metaslab_check_free(spa, bp);
787 
788 	/*
789 	 * Frees that are for the currently-syncing txg, are not going to be
790 	 * deferred, and which will not need to do a read (i.e. not GANG or
791 	 * DEDUP), can be processed immediately.  Otherwise, put them on the
792 	 * in-memory list for later processing.
793 	 */
794 	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
795 	    txg != spa->spa_syncing_txg ||
796 	    spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
797 		bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
798 	} else {
799 		VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
800 	}
801 }
802 
803 zio_t *
804 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
805     enum zio_flag flags)
806 {
807 	zio_t *zio;
808 	enum zio_stage stage = ZIO_FREE_PIPELINE;
809 
810 	ASSERT(!BP_IS_HOLE(bp));
811 	ASSERT(spa_syncing_txg(spa) == txg);
812 	ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
813 
814 	if (BP_IS_EMBEDDED(bp))
815 		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
816 
817 	metaslab_check_free(spa, bp);
818 	arc_freed(spa, bp);
819 
820 	/*
821 	 * GANG and DEDUP blocks can induce a read (for the gang block header,
822 	 * or the DDT), so issue them asynchronously so that this thread is
823 	 * not tied up.
824 	 */
825 	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
826 		stage |= ZIO_STAGE_ISSUE_ASYNC;
827 
828 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
829 	    NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
830 	    NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
831 
832 	return (zio);
833 }
834 
835 zio_t *
836 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
837     zio_done_func_t *done, void *private, enum zio_flag flags)
838 {
839 	zio_t *zio;
840 
841 	dprintf_bp(bp, "claiming in txg %llu", txg);
842 
843 	if (BP_IS_EMBEDDED(bp))
844 		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
845 
846 	/*
847 	 * A claim is an allocation of a specific block.  Claims are needed
848 	 * to support immediate writes in the intent log.  The issue is that
849 	 * immediate writes contain committed data, but in a txg that was
850 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
851 	 * the intent log claims all blocks that contain immediate write data
852 	 * so that the SPA knows they're in use.
853 	 *
854 	 * All claims *must* be resolved in the first txg -- before the SPA
855 	 * starts allocating blocks -- so that nothing is allocated twice.
856 	 * If txg == 0 we just verify that the block is claimable.
857 	 */
858 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
859 	ASSERT(txg == spa_first_txg(spa) || txg == 0);
860 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
861 
862 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
863 	    done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
864 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
865 
866 	return (zio);
867 }
868 
869 zio_t *
870 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
871     zio_done_func_t *done, void *private, enum zio_flag flags)
872 {
873 	zio_t *zio;
874 	int c;
875 
876 	if (vd->vdev_children == 0) {
877 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
878 		    ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
879 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
880 
881 		zio->io_cmd = cmd;
882 	} else {
883 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
884 
885 		for (c = 0; c < vd->vdev_children; c++)
886 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
887 			    done, private, flags));
888 	}
889 
890 	return (zio);
891 }
892 
893 zio_t *
894 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
895     void *data, int checksum, zio_done_func_t *done, void *private,
896     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
897 {
898 	zio_t *zio;
899 
900 	ASSERT(vd->vdev_children == 0);
901 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
902 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
903 	ASSERT3U(offset + size, <=, vd->vdev_psize);
904 
905 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
906 	    ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
907 	    NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
908 
909 	zio->io_prop.zp_checksum = checksum;
910 
911 	return (zio);
912 }
913 
914 zio_t *
915 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
916     void *data, int checksum, zio_done_func_t *done, void *private,
917     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
918 {
919 	zio_t *zio;
920 
921 	ASSERT(vd->vdev_children == 0);
922 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
923 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
924 	ASSERT3U(offset + size, <=, vd->vdev_psize);
925 
926 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
927 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
928 	    NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
929 
930 	zio->io_prop.zp_checksum = checksum;
931 
932 	if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
933 		/*
934 		 * zec checksums are necessarily destructive -- they modify
935 		 * the end of the write buffer to hold the verifier/checksum.
936 		 * Therefore, we must make a local copy in case the data is
937 		 * being written to multiple places in parallel.
938 		 */
939 		void *wbuf = zio_buf_alloc(size);
940 		bcopy(data, wbuf, size);
941 		zio_push_transform(zio, wbuf, size, size, NULL);
942 	}
943 
944 	return (zio);
945 }
946 
947 /*
948  * Create a child I/O to do some work for us.
949  */
950 zio_t *
951 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
952 	void *data, uint64_t size, int type, zio_priority_t priority,
953 	enum zio_flag flags, zio_done_func_t *done, void *private)
954 {
955 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
956 	zio_t *zio;
957 
958 	ASSERT(vd->vdev_parent ==
959 	    (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
960 
961 	if (type == ZIO_TYPE_READ && bp != NULL) {
962 		/*
963 		 * If we have the bp, then the child should perform the
964 		 * checksum and the parent need not.  This pushes error
965 		 * detection as close to the leaves as possible and
966 		 * eliminates redundant checksums in the interior nodes.
967 		 */
968 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
969 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
970 	}
971 
972 	if (vd->vdev_children == 0)
973 		offset += VDEV_LABEL_START_SIZE;
974 
975 	flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
976 
977 	/*
978 	 * If we've decided to do a repair, the write is not speculative --
979 	 * even if the original read was.
980 	 */
981 	if (flags & ZIO_FLAG_IO_REPAIR)
982 		flags &= ~ZIO_FLAG_SPECULATIVE;
983 
984 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
985 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
986 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
987 
988 	zio->io_physdone = pio->io_physdone;
989 	if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
990 		zio->io_logical->io_phys_children++;
991 
992 	return (zio);
993 }
994 
995 zio_t *
996 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
997     int type, zio_priority_t priority, enum zio_flag flags,
998     zio_done_func_t *done, void *private)
999 {
1000 	zio_t *zio;
1001 
1002 	ASSERT(vd->vdev_ops->vdev_op_leaf);
1003 
1004 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1005 	    data, size, done, private, type, priority,
1006 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1007 	    vd, offset, NULL,
1008 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1009 
1010 	return (zio);
1011 }
1012 
1013 void
1014 zio_flush(zio_t *zio, vdev_t *vd)
1015 {
1016 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1017 	    NULL, NULL,
1018 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1019 }
1020 
1021 void
1022 zio_shrink(zio_t *zio, uint64_t size)
1023 {
1024 	ASSERT(zio->io_executor == NULL);
1025 	ASSERT(zio->io_orig_size == zio->io_size);
1026 	ASSERT(size <= zio->io_size);
1027 
1028 	/*
1029 	 * We don't shrink for raidz because of problems with the
1030 	 * reconstruction when reading back less than the block size.
1031 	 * Note, BP_IS_RAIDZ() assumes no compression.
1032 	 */
1033 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1034 	if (!BP_IS_RAIDZ(zio->io_bp))
1035 		zio->io_orig_size = zio->io_size = size;
1036 }
1037 
1038 /*
1039  * ==========================================================================
1040  * Prepare to read and write logical blocks
1041  * ==========================================================================
1042  */
1043 
1044 static int
1045 zio_read_bp_init(zio_t *zio)
1046 {
1047 	blkptr_t *bp = zio->io_bp;
1048 
1049 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1050 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
1051 	    !(zio->io_flags & ZIO_FLAG_RAW)) {
1052 		uint64_t psize =
1053 		    BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1054 		void *cbuf = zio_buf_alloc(psize);
1055 
1056 		zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1057 	}
1058 
1059 	if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1060 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1061 		decode_embedded_bp_compressed(bp, zio->io_data);
1062 	} else {
1063 		ASSERT(!BP_IS_EMBEDDED(bp));
1064 	}
1065 
1066 	if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1067 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1068 
1069 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1070 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1071 
1072 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1073 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1074 
1075 	return (ZIO_PIPELINE_CONTINUE);
1076 }
1077 
1078 static int
1079 zio_write_bp_init(zio_t *zio)
1080 {
1081 	spa_t *spa = zio->io_spa;
1082 	zio_prop_t *zp = &zio->io_prop;
1083 	enum zio_compress compress = zp->zp_compress;
1084 	blkptr_t *bp = zio->io_bp;
1085 	uint64_t lsize = zio->io_size;
1086 	uint64_t psize = lsize;
1087 	int pass = 1;
1088 
1089 	/*
1090 	 * If our children haven't all reached the ready stage,
1091 	 * wait for them and then repeat this pipeline stage.
1092 	 */
1093 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1094 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1095 		return (ZIO_PIPELINE_STOP);
1096 
1097 	if (!IO_IS_ALLOCATING(zio))
1098 		return (ZIO_PIPELINE_CONTINUE);
1099 
1100 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1101 
1102 	if (zio->io_bp_override) {
1103 		ASSERT(bp->blk_birth != zio->io_txg);
1104 		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1105 
1106 		*bp = *zio->io_bp_override;
1107 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1108 
1109 		if (BP_IS_EMBEDDED(bp))
1110 			return (ZIO_PIPELINE_CONTINUE);
1111 
1112 		/*
1113 		 * If we've been overridden and nopwrite is set then
1114 		 * set the flag accordingly to indicate that a nopwrite
1115 		 * has already occurred.
1116 		 */
1117 		if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1118 			ASSERT(!zp->zp_dedup);
1119 			zio->io_flags |= ZIO_FLAG_NOPWRITE;
1120 			return (ZIO_PIPELINE_CONTINUE);
1121 		}
1122 
1123 		ASSERT(!zp->zp_nopwrite);
1124 
1125 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1126 			return (ZIO_PIPELINE_CONTINUE);
1127 
1128 		ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1129 		    ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1130 
1131 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1132 			BP_SET_DEDUP(bp, 1);
1133 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1134 			return (ZIO_PIPELINE_CONTINUE);
1135 		}
1136 		zio->io_bp_override = NULL;
1137 		BP_ZERO(bp);
1138 	}
1139 
1140 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1141 		/*
1142 		 * We're rewriting an existing block, which means we're
1143 		 * working on behalf of spa_sync().  For spa_sync() to
1144 		 * converge, it must eventually be the case that we don't
1145 		 * have to allocate new blocks.  But compression changes
1146 		 * the blocksize, which forces a reallocate, and makes
1147 		 * convergence take longer.  Therefore, after the first
1148 		 * few passes, stop compressing to ensure convergence.
1149 		 */
1150 		pass = spa_sync_pass(spa);
1151 
1152 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
1153 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1154 		ASSERT(!BP_GET_DEDUP(bp));
1155 
1156 		if (pass >= zfs_sync_pass_dont_compress)
1157 			compress = ZIO_COMPRESS_OFF;
1158 
1159 		/* Make sure someone doesn't change their mind on overwrites */
1160 		ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1161 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1162 	}
1163 
1164 	if (compress != ZIO_COMPRESS_OFF) {
1165 		void *cbuf = zio_buf_alloc(lsize);
1166 		psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1167 		if (psize == 0 || psize == lsize) {
1168 			compress = ZIO_COMPRESS_OFF;
1169 			zio_buf_free(cbuf, lsize);
1170 		} else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1171 		    zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1172 		    spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1173 			encode_embedded_bp_compressed(bp,
1174 			    cbuf, compress, lsize, psize);
1175 			BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1176 			BP_SET_TYPE(bp, zio->io_prop.zp_type);
1177 			BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1178 			zio_buf_free(cbuf, lsize);
1179 			bp->blk_birth = zio->io_txg;
1180 			zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1181 			ASSERT(spa_feature_is_active(spa,
1182 			    SPA_FEATURE_EMBEDDED_DATA));
1183 			return (ZIO_PIPELINE_CONTINUE);
1184 		} else {
1185 			/*
1186 			 * Round up compressed size up to the ashift
1187 			 * of the smallest-ashift device, and zero the tail.
1188 			 * This ensures that the compressed size of the BP
1189 			 * (and thus compressratio property) are correct,
1190 			 * in that we charge for the padding used to fill out
1191 			 * the last sector.
1192 			 */
1193 			ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1194 			size_t rounded = (size_t)P2ROUNDUP(psize,
1195 			    1ULL << spa->spa_min_ashift);
1196 			if (rounded >= lsize) {
1197 				compress = ZIO_COMPRESS_OFF;
1198 				zio_buf_free(cbuf, lsize);
1199 				psize = lsize;
1200 			} else {
1201 				bzero((char *)cbuf + psize, rounded - psize);
1202 				psize = rounded;
1203 				zio_push_transform(zio, cbuf,
1204 				    psize, lsize, NULL);
1205 			}
1206 		}
1207 	}
1208 
1209 	/*
1210 	 * The final pass of spa_sync() must be all rewrites, but the first
1211 	 * few passes offer a trade-off: allocating blocks defers convergence,
1212 	 * but newly allocated blocks are sequential, so they can be written
1213 	 * to disk faster.  Therefore, we allow the first few passes of
1214 	 * spa_sync() to allocate new blocks, but force rewrites after that.
1215 	 * There should only be a handful of blocks after pass 1 in any case.
1216 	 */
1217 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1218 	    BP_GET_PSIZE(bp) == psize &&
1219 	    pass >= zfs_sync_pass_rewrite) {
1220 		ASSERT(psize != 0);
1221 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1222 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1223 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1224 	} else {
1225 		BP_ZERO(bp);
1226 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1227 	}
1228 
1229 	if (psize == 0) {
1230 		if (zio->io_bp_orig.blk_birth != 0 &&
1231 		    spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1232 			BP_SET_LSIZE(bp, lsize);
1233 			BP_SET_TYPE(bp, zp->zp_type);
1234 			BP_SET_LEVEL(bp, zp->zp_level);
1235 			BP_SET_BIRTH(bp, zio->io_txg, 0);
1236 		}
1237 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1238 	} else {
1239 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1240 		BP_SET_LSIZE(bp, lsize);
1241 		BP_SET_TYPE(bp, zp->zp_type);
1242 		BP_SET_LEVEL(bp, zp->zp_level);
1243 		BP_SET_PSIZE(bp, psize);
1244 		BP_SET_COMPRESS(bp, compress);
1245 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1246 		BP_SET_DEDUP(bp, zp->zp_dedup);
1247 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1248 		if (zp->zp_dedup) {
1249 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1250 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1251 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1252 		}
1253 		if (zp->zp_nopwrite) {
1254 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1255 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1256 			zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1257 		}
1258 	}
1259 
1260 	return (ZIO_PIPELINE_CONTINUE);
1261 }
1262 
1263 static int
1264 zio_free_bp_init(zio_t *zio)
1265 {
1266 	blkptr_t *bp = zio->io_bp;
1267 
1268 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1269 		if (BP_GET_DEDUP(bp))
1270 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1271 	}
1272 
1273 	return (ZIO_PIPELINE_CONTINUE);
1274 }
1275 
1276 /*
1277  * ==========================================================================
1278  * Execute the I/O pipeline
1279  * ==========================================================================
1280  */
1281 
1282 static void
1283 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1284 {
1285 	spa_t *spa = zio->io_spa;
1286 	zio_type_t t = zio->io_type;
1287 	int flags = (cutinline ? TQ_FRONT : 0);
1288 
1289 	/*
1290 	 * If we're a config writer or a probe, the normal issue and
1291 	 * interrupt threads may all be blocked waiting for the config lock.
1292 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1293 	 */
1294 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1295 		t = ZIO_TYPE_NULL;
1296 
1297 	/*
1298 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1299 	 */
1300 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1301 		t = ZIO_TYPE_NULL;
1302 
1303 	/*
1304 	 * If this is a high priority I/O, then use the high priority taskq if
1305 	 * available.
1306 	 */
1307 	if (zio->io_priority == ZIO_PRIORITY_NOW &&
1308 	    spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1309 		q++;
1310 
1311 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1312 
1313 	/*
1314 	 * NB: We are assuming that the zio can only be dispatched
1315 	 * to a single taskq at a time.  It would be a grievous error
1316 	 * to dispatch the zio to another taskq at the same time.
1317 	 */
1318 	ASSERT(zio->io_tqent.tqent_next == NULL);
1319 	spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1320 	    flags, &zio->io_tqent);
1321 }
1322 
1323 static boolean_t
1324 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1325 {
1326 	kthread_t *executor = zio->io_executor;
1327 	spa_t *spa = zio->io_spa;
1328 
1329 	for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1330 		spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1331 		uint_t i;
1332 		for (i = 0; i < tqs->stqs_count; i++) {
1333 			if (taskq_member(tqs->stqs_taskq[i], executor))
1334 				return (B_TRUE);
1335 		}
1336 	}
1337 
1338 	return (B_FALSE);
1339 }
1340 
1341 static int
1342 zio_issue_async(zio_t *zio)
1343 {
1344 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1345 
1346 	return (ZIO_PIPELINE_STOP);
1347 }
1348 
1349 void
1350 zio_interrupt(zio_t *zio)
1351 {
1352 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1353 }
1354 
1355 /*
1356  * Execute the I/O pipeline until one of the following occurs:
1357  *
1358  *	(1) the I/O completes
1359  *	(2) the pipeline stalls waiting for dependent child I/Os
1360  *	(3) the I/O issues, so we're waiting for an I/O completion interrupt
1361  *	(4) the I/O is delegated by vdev-level caching or aggregation
1362  *	(5) the I/O is deferred due to vdev-level queueing
1363  *	(6) the I/O is handed off to another thread.
1364  *
1365  * In all cases, the pipeline stops whenever there's no CPU work; it never
1366  * burns a thread in cv_wait().
1367  *
1368  * There's no locking on io_stage because there's no legitimate way
1369  * for multiple threads to be attempting to process the same I/O.
1370  */
1371 static zio_pipe_stage_t *zio_pipeline[];
1372 
1373 void
1374 zio_execute(zio_t *zio)
1375 {
1376 	zio->io_executor = curthread;
1377 
1378 	while (zio->io_stage < ZIO_STAGE_DONE) {
1379 		enum zio_stage pipeline = zio->io_pipeline;
1380 		enum zio_stage stage = zio->io_stage;
1381 		int rv;
1382 
1383 		ASSERT(!MUTEX_HELD(&zio->io_lock));
1384 		ASSERT(ISP2(stage));
1385 		ASSERT(zio->io_stall == NULL);
1386 
1387 		do {
1388 			stage <<= 1;
1389 		} while ((stage & pipeline) == 0);
1390 
1391 		ASSERT(stage <= ZIO_STAGE_DONE);
1392 
1393 		/*
1394 		 * If we are in interrupt context and this pipeline stage
1395 		 * will grab a config lock that is held across I/O,
1396 		 * or may wait for an I/O that needs an interrupt thread
1397 		 * to complete, issue async to avoid deadlock.
1398 		 *
1399 		 * For VDEV_IO_START, we cut in line so that the io will
1400 		 * be sent to disk promptly.
1401 		 */
1402 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1403 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1404 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1405 			    zio_requeue_io_start_cut_in_line : B_FALSE;
1406 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1407 			return;
1408 		}
1409 
1410 		zio->io_stage = stage;
1411 		rv = zio_pipeline[highbit64(stage) - 1](zio);
1412 
1413 		if (rv == ZIO_PIPELINE_STOP)
1414 			return;
1415 
1416 		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1417 	}
1418 }
1419 
1420 /*
1421  * ==========================================================================
1422  * Initiate I/O, either sync or async
1423  * ==========================================================================
1424  */
1425 int
1426 zio_wait(zio_t *zio)
1427 {
1428 	int error;
1429 
1430 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1431 	ASSERT(zio->io_executor == NULL);
1432 
1433 	zio->io_waiter = curthread;
1434 
1435 	zio_execute(zio);
1436 
1437 	mutex_enter(&zio->io_lock);
1438 	while (zio->io_executor != NULL)
1439 		cv_wait(&zio->io_cv, &zio->io_lock);
1440 	mutex_exit(&zio->io_lock);
1441 
1442 	error = zio->io_error;
1443 	zio_destroy(zio);
1444 
1445 	return (error);
1446 }
1447 
1448 void
1449 zio_nowait(zio_t *zio)
1450 {
1451 	ASSERT(zio->io_executor == NULL);
1452 
1453 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1454 	    zio_unique_parent(zio) == NULL) {
1455 		/*
1456 		 * This is a logical async I/O with no parent to wait for it.
1457 		 * We add it to the spa_async_root_zio "Godfather" I/O which
1458 		 * will ensure they complete prior to unloading the pool.
1459 		 */
1460 		spa_t *spa = zio->io_spa;
1461 
1462 		zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1463 	}
1464 
1465 	zio_execute(zio);
1466 }
1467 
1468 /*
1469  * ==========================================================================
1470  * Reexecute or suspend/resume failed I/O
1471  * ==========================================================================
1472  */
1473 
1474 static void
1475 zio_reexecute(zio_t *pio)
1476 {
1477 	zio_t *cio, *cio_next;
1478 
1479 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1480 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1481 	ASSERT(pio->io_gang_leader == NULL);
1482 	ASSERT(pio->io_gang_tree == NULL);
1483 
1484 	pio->io_flags = pio->io_orig_flags;
1485 	pio->io_stage = pio->io_orig_stage;
1486 	pio->io_pipeline = pio->io_orig_pipeline;
1487 	pio->io_reexecute = 0;
1488 	pio->io_flags |= ZIO_FLAG_REEXECUTED;
1489 	pio->io_error = 0;
1490 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1491 		pio->io_state[w] = 0;
1492 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1493 		pio->io_child_error[c] = 0;
1494 
1495 	if (IO_IS_ALLOCATING(pio))
1496 		BP_ZERO(pio->io_bp);
1497 
1498 	/*
1499 	 * As we reexecute pio's children, new children could be created.
1500 	 * New children go to the head of pio's io_child_list, however,
1501 	 * so we will (correctly) not reexecute them.  The key is that
1502 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1503 	 * cannot be affected by any side effects of reexecuting 'cio'.
1504 	 */
1505 	for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1506 		cio_next = zio_walk_children(pio);
1507 		mutex_enter(&pio->io_lock);
1508 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1509 			pio->io_children[cio->io_child_type][w]++;
1510 		mutex_exit(&pio->io_lock);
1511 		zio_reexecute(cio);
1512 	}
1513 
1514 	/*
1515 	 * Now that all children have been reexecuted, execute the parent.
1516 	 * We don't reexecute "The Godfather" I/O here as it's the
1517 	 * responsibility of the caller to wait on him.
1518 	 */
1519 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1520 		zio_execute(pio);
1521 }
1522 
1523 void
1524 zio_suspend(spa_t *spa, zio_t *zio)
1525 {
1526 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1527 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1528 		    "failure and the failure mode property for this pool "
1529 		    "is set to panic.", spa_name(spa));
1530 
1531 	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1532 
1533 	mutex_enter(&spa->spa_suspend_lock);
1534 
1535 	if (spa->spa_suspend_zio_root == NULL)
1536 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1537 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1538 		    ZIO_FLAG_GODFATHER);
1539 
1540 	spa->spa_suspended = B_TRUE;
1541 
1542 	if (zio != NULL) {
1543 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1544 		ASSERT(zio != spa->spa_suspend_zio_root);
1545 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1546 		ASSERT(zio_unique_parent(zio) == NULL);
1547 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1548 		zio_add_child(spa->spa_suspend_zio_root, zio);
1549 	}
1550 
1551 	mutex_exit(&spa->spa_suspend_lock);
1552 }
1553 
1554 int
1555 zio_resume(spa_t *spa)
1556 {
1557 	zio_t *pio;
1558 
1559 	/*
1560 	 * Reexecute all previously suspended i/o.
1561 	 */
1562 	mutex_enter(&spa->spa_suspend_lock);
1563 	spa->spa_suspended = B_FALSE;
1564 	cv_broadcast(&spa->spa_suspend_cv);
1565 	pio = spa->spa_suspend_zio_root;
1566 	spa->spa_suspend_zio_root = NULL;
1567 	mutex_exit(&spa->spa_suspend_lock);
1568 
1569 	if (pio == NULL)
1570 		return (0);
1571 
1572 	zio_reexecute(pio);
1573 	return (zio_wait(pio));
1574 }
1575 
1576 void
1577 zio_resume_wait(spa_t *spa)
1578 {
1579 	mutex_enter(&spa->spa_suspend_lock);
1580 	while (spa_suspended(spa))
1581 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1582 	mutex_exit(&spa->spa_suspend_lock);
1583 }
1584 
1585 /*
1586  * ==========================================================================
1587  * Gang blocks.
1588  *
1589  * A gang block is a collection of small blocks that looks to the DMU
1590  * like one large block.  When zio_dva_allocate() cannot find a block
1591  * of the requested size, due to either severe fragmentation or the pool
1592  * being nearly full, it calls zio_write_gang_block() to construct the
1593  * block from smaller fragments.
1594  *
1595  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1596  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
1597  * an indirect block: it's an array of block pointers.  It consumes
1598  * only one sector and hence is allocatable regardless of fragmentation.
1599  * The gang header's bps point to its gang members, which hold the data.
1600  *
1601  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1602  * as the verifier to ensure uniqueness of the SHA256 checksum.
1603  * Critically, the gang block bp's blk_cksum is the checksum of the data,
1604  * not the gang header.  This ensures that data block signatures (needed for
1605  * deduplication) are independent of how the block is physically stored.
1606  *
1607  * Gang blocks can be nested: a gang member may itself be a gang block.
1608  * Thus every gang block is a tree in which root and all interior nodes are
1609  * gang headers, and the leaves are normal blocks that contain user data.
1610  * The root of the gang tree is called the gang leader.
1611  *
1612  * To perform any operation (read, rewrite, free, claim) on a gang block,
1613  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1614  * in the io_gang_tree field of the original logical i/o by recursively
1615  * reading the gang leader and all gang headers below it.  This yields
1616  * an in-core tree containing the contents of every gang header and the
1617  * bps for every constituent of the gang block.
1618  *
1619  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1620  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
1621  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1622  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1623  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1624  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
1625  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1626  * of the gang header plus zio_checksum_compute() of the data to update the
1627  * gang header's blk_cksum as described above.
1628  *
1629  * The two-phase assemble/issue model solves the problem of partial failure --
1630  * what if you'd freed part of a gang block but then couldn't read the
1631  * gang header for another part?  Assembling the entire gang tree first
1632  * ensures that all the necessary gang header I/O has succeeded before
1633  * starting the actual work of free, claim, or write.  Once the gang tree
1634  * is assembled, free and claim are in-memory operations that cannot fail.
1635  *
1636  * In the event that a gang write fails, zio_dva_unallocate() walks the
1637  * gang tree to immediately free (i.e. insert back into the space map)
1638  * everything we've allocated.  This ensures that we don't get ENOSPC
1639  * errors during repeated suspend/resume cycles due to a flaky device.
1640  *
1641  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
1642  * the gang tree, we won't modify the block, so we can safely defer the free
1643  * (knowing that the block is still intact).  If we *can* assemble the gang
1644  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1645  * each constituent bp and we can allocate a new block on the next sync pass.
1646  *
1647  * In all cases, the gang tree allows complete recovery from partial failure.
1648  * ==========================================================================
1649  */
1650 
1651 static zio_t *
1652 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1653 {
1654 	if (gn != NULL)
1655 		return (pio);
1656 
1657 	return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1658 	    NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1659 	    &pio->io_bookmark));
1660 }
1661 
1662 zio_t *
1663 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1664 {
1665 	zio_t *zio;
1666 
1667 	if (gn != NULL) {
1668 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1669 		    gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1670 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1671 		/*
1672 		 * As we rewrite each gang header, the pipeline will compute
1673 		 * a new gang block header checksum for it; but no one will
1674 		 * compute a new data checksum, so we do that here.  The one
1675 		 * exception is the gang leader: the pipeline already computed
1676 		 * its data checksum because that stage precedes gang assembly.
1677 		 * (Presently, nothing actually uses interior data checksums;
1678 		 * this is just good hygiene.)
1679 		 */
1680 		if (gn != pio->io_gang_leader->io_gang_tree) {
1681 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1682 			    data, BP_GET_PSIZE(bp));
1683 		}
1684 		/*
1685 		 * If we are here to damage data for testing purposes,
1686 		 * leave the GBH alone so that we can detect the damage.
1687 		 */
1688 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1689 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1690 	} else {
1691 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1692 		    data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1693 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1694 	}
1695 
1696 	return (zio);
1697 }
1698 
1699 /* ARGSUSED */
1700 zio_t *
1701 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1702 {
1703 	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1704 	    ZIO_GANG_CHILD_FLAGS(pio)));
1705 }
1706 
1707 /* ARGSUSED */
1708 zio_t *
1709 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1710 {
1711 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1712 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1713 }
1714 
1715 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1716 	NULL,
1717 	zio_read_gang,
1718 	zio_rewrite_gang,
1719 	zio_free_gang,
1720 	zio_claim_gang,
1721 	NULL
1722 };
1723 
1724 static void zio_gang_tree_assemble_done(zio_t *zio);
1725 
1726 static zio_gang_node_t *
1727 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1728 {
1729 	zio_gang_node_t *gn;
1730 
1731 	ASSERT(*gnpp == NULL);
1732 
1733 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1734 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1735 	*gnpp = gn;
1736 
1737 	return (gn);
1738 }
1739 
1740 static void
1741 zio_gang_node_free(zio_gang_node_t **gnpp)
1742 {
1743 	zio_gang_node_t *gn = *gnpp;
1744 
1745 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1746 		ASSERT(gn->gn_child[g] == NULL);
1747 
1748 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1749 	kmem_free(gn, sizeof (*gn));
1750 	*gnpp = NULL;
1751 }
1752 
1753 static void
1754 zio_gang_tree_free(zio_gang_node_t **gnpp)
1755 {
1756 	zio_gang_node_t *gn = *gnpp;
1757 
1758 	if (gn == NULL)
1759 		return;
1760 
1761 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1762 		zio_gang_tree_free(&gn->gn_child[g]);
1763 
1764 	zio_gang_node_free(gnpp);
1765 }
1766 
1767 static void
1768 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1769 {
1770 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1771 
1772 	ASSERT(gio->io_gang_leader == gio);
1773 	ASSERT(BP_IS_GANG(bp));
1774 
1775 	zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1776 	    SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1777 	    gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1778 }
1779 
1780 static void
1781 zio_gang_tree_assemble_done(zio_t *zio)
1782 {
1783 	zio_t *gio = zio->io_gang_leader;
1784 	zio_gang_node_t *gn = zio->io_private;
1785 	blkptr_t *bp = zio->io_bp;
1786 
1787 	ASSERT(gio == zio_unique_parent(zio));
1788 	ASSERT(zio->io_child_count == 0);
1789 
1790 	if (zio->io_error)
1791 		return;
1792 
1793 	if (BP_SHOULD_BYTESWAP(bp))
1794 		byteswap_uint64_array(zio->io_data, zio->io_size);
1795 
1796 	ASSERT(zio->io_data == gn->gn_gbh);
1797 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1798 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1799 
1800 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1801 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1802 		if (!BP_IS_GANG(gbp))
1803 			continue;
1804 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1805 	}
1806 }
1807 
1808 static void
1809 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1810 {
1811 	zio_t *gio = pio->io_gang_leader;
1812 	zio_t *zio;
1813 
1814 	ASSERT(BP_IS_GANG(bp) == !!gn);
1815 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1816 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1817 
1818 	/*
1819 	 * If you're a gang header, your data is in gn->gn_gbh.
1820 	 * If you're a gang member, your data is in 'data' and gn == NULL.
1821 	 */
1822 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1823 
1824 	if (gn != NULL) {
1825 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1826 
1827 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1828 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1829 			if (BP_IS_HOLE(gbp))
1830 				continue;
1831 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1832 			data = (char *)data + BP_GET_PSIZE(gbp);
1833 		}
1834 	}
1835 
1836 	if (gn == gio->io_gang_tree)
1837 		ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1838 
1839 	if (zio != pio)
1840 		zio_nowait(zio);
1841 }
1842 
1843 static int
1844 zio_gang_assemble(zio_t *zio)
1845 {
1846 	blkptr_t *bp = zio->io_bp;
1847 
1848 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1849 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1850 
1851 	zio->io_gang_leader = zio;
1852 
1853 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1854 
1855 	return (ZIO_PIPELINE_CONTINUE);
1856 }
1857 
1858 static int
1859 zio_gang_issue(zio_t *zio)
1860 {
1861 	blkptr_t *bp = zio->io_bp;
1862 
1863 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1864 		return (ZIO_PIPELINE_STOP);
1865 
1866 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1867 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1868 
1869 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1870 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1871 	else
1872 		zio_gang_tree_free(&zio->io_gang_tree);
1873 
1874 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1875 
1876 	return (ZIO_PIPELINE_CONTINUE);
1877 }
1878 
1879 static void
1880 zio_write_gang_member_ready(zio_t *zio)
1881 {
1882 	zio_t *pio = zio_unique_parent(zio);
1883 	zio_t *gio = zio->io_gang_leader;
1884 	dva_t *cdva = zio->io_bp->blk_dva;
1885 	dva_t *pdva = pio->io_bp->blk_dva;
1886 	uint64_t asize;
1887 
1888 	if (BP_IS_HOLE(zio->io_bp))
1889 		return;
1890 
1891 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1892 
1893 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1894 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1895 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1896 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1897 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1898 
1899 	mutex_enter(&pio->io_lock);
1900 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1901 		ASSERT(DVA_GET_GANG(&pdva[d]));
1902 		asize = DVA_GET_ASIZE(&pdva[d]);
1903 		asize += DVA_GET_ASIZE(&cdva[d]);
1904 		DVA_SET_ASIZE(&pdva[d], asize);
1905 	}
1906 	mutex_exit(&pio->io_lock);
1907 }
1908 
1909 static int
1910 zio_write_gang_block(zio_t *pio)
1911 {
1912 	spa_t *spa = pio->io_spa;
1913 	blkptr_t *bp = pio->io_bp;
1914 	zio_t *gio = pio->io_gang_leader;
1915 	zio_t *zio;
1916 	zio_gang_node_t *gn, **gnpp;
1917 	zio_gbh_phys_t *gbh;
1918 	uint64_t txg = pio->io_txg;
1919 	uint64_t resid = pio->io_size;
1920 	uint64_t lsize;
1921 	int copies = gio->io_prop.zp_copies;
1922 	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1923 	zio_prop_t zp;
1924 	int error;
1925 
1926 	error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1927 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1928 	    METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1929 	if (error) {
1930 		pio->io_error = error;
1931 		return (ZIO_PIPELINE_CONTINUE);
1932 	}
1933 
1934 	if (pio == gio) {
1935 		gnpp = &gio->io_gang_tree;
1936 	} else {
1937 		gnpp = pio->io_private;
1938 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
1939 	}
1940 
1941 	gn = zio_gang_node_alloc(gnpp);
1942 	gbh = gn->gn_gbh;
1943 	bzero(gbh, SPA_GANGBLOCKSIZE);
1944 
1945 	/*
1946 	 * Create the gang header.
1947 	 */
1948 	zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1949 	    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1950 
1951 	/*
1952 	 * Create and nowait the gang children.
1953 	 */
1954 	for (int g = 0; resid != 0; resid -= lsize, g++) {
1955 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1956 		    SPA_MINBLOCKSIZE);
1957 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1958 
1959 		zp.zp_checksum = gio->io_prop.zp_checksum;
1960 		zp.zp_compress = ZIO_COMPRESS_OFF;
1961 		zp.zp_type = DMU_OT_NONE;
1962 		zp.zp_level = 0;
1963 		zp.zp_copies = gio->io_prop.zp_copies;
1964 		zp.zp_dedup = B_FALSE;
1965 		zp.zp_dedup_verify = B_FALSE;
1966 		zp.zp_nopwrite = B_FALSE;
1967 
1968 		zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1969 		    (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1970 		    zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1971 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1972 		    &pio->io_bookmark));
1973 	}
1974 
1975 	/*
1976 	 * Set pio's pipeline to just wait for zio to finish.
1977 	 */
1978 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1979 
1980 	zio_nowait(zio);
1981 
1982 	return (ZIO_PIPELINE_CONTINUE);
1983 }
1984 
1985 /*
1986  * The zio_nop_write stage in the pipeline determines if allocating a
1987  * new bp is necessary.  The nopwrite feature can handle writes in
1988  * either syncing or open context (i.e. zil writes) and as a result is
1989  * mutually exclusive with dedup.
1990  *
1991  * By leveraging a cryptographically secure checksum, such as SHA256, we
1992  * can compare the checksums of the new data and the old to determine if
1993  * allocating a new block is required.  Note that our requirements for
1994  * cryptographic strength are fairly weak: there can't be any accidental
1995  * hash collisions, but we don't need to be secure against intentional
1996  * (malicious) collisions.  To trigger a nopwrite, you have to be able
1997  * to write the file to begin with, and triggering an incorrect (hash
1998  * collision) nopwrite is no worse than simply writing to the file.
1999  * That said, there are no known attacks against the checksum algorithms
2000  * used for nopwrite, assuming that the salt and the checksums
2001  * themselves remain secret.
2002  */
2003 static int
2004 zio_nop_write(zio_t *zio)
2005 {
2006 	blkptr_t *bp = zio->io_bp;
2007 	blkptr_t *bp_orig = &zio->io_bp_orig;
2008 	zio_prop_t *zp = &zio->io_prop;
2009 
2010 	ASSERT(BP_GET_LEVEL(bp) == 0);
2011 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2012 	ASSERT(zp->zp_nopwrite);
2013 	ASSERT(!zp->zp_dedup);
2014 	ASSERT(zio->io_bp_override == NULL);
2015 	ASSERT(IO_IS_ALLOCATING(zio));
2016 
2017 	/*
2018 	 * Check to see if the original bp and the new bp have matching
2019 	 * characteristics (i.e. same checksum, compression algorithms, etc).
2020 	 * If they don't then just continue with the pipeline which will
2021 	 * allocate a new bp.
2022 	 */
2023 	if (BP_IS_HOLE(bp_orig) ||
2024 	    !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2025 	    ZCHECKSUM_FLAG_NOPWRITE) ||
2026 	    BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2027 	    BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2028 	    BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2029 	    zp->zp_copies != BP_GET_NDVAS(bp_orig))
2030 		return (ZIO_PIPELINE_CONTINUE);
2031 
2032 	/*
2033 	 * If the checksums match then reset the pipeline so that we
2034 	 * avoid allocating a new bp and issuing any I/O.
2035 	 */
2036 	if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2037 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2038 		    ZCHECKSUM_FLAG_NOPWRITE);
2039 		ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2040 		ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2041 		ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2042 		ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2043 		    sizeof (uint64_t)) == 0);
2044 
2045 		*bp = *bp_orig;
2046 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2047 		zio->io_flags |= ZIO_FLAG_NOPWRITE;
2048 	}
2049 
2050 	return (ZIO_PIPELINE_CONTINUE);
2051 }
2052 
2053 /*
2054  * ==========================================================================
2055  * Dedup
2056  * ==========================================================================
2057  */
2058 static void
2059 zio_ddt_child_read_done(zio_t *zio)
2060 {
2061 	blkptr_t *bp = zio->io_bp;
2062 	ddt_entry_t *dde = zio->io_private;
2063 	ddt_phys_t *ddp;
2064 	zio_t *pio = zio_unique_parent(zio);
2065 
2066 	mutex_enter(&pio->io_lock);
2067 	ddp = ddt_phys_select(dde, bp);
2068 	if (zio->io_error == 0)
2069 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
2070 	if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2071 		dde->dde_repair_data = zio->io_data;
2072 	else
2073 		zio_buf_free(zio->io_data, zio->io_size);
2074 	mutex_exit(&pio->io_lock);
2075 }
2076 
2077 static int
2078 zio_ddt_read_start(zio_t *zio)
2079 {
2080 	blkptr_t *bp = zio->io_bp;
2081 
2082 	ASSERT(BP_GET_DEDUP(bp));
2083 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2084 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2085 
2086 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
2087 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
2088 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2089 		ddt_phys_t *ddp = dde->dde_phys;
2090 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2091 		blkptr_t blk;
2092 
2093 		ASSERT(zio->io_vsd == NULL);
2094 		zio->io_vsd = dde;
2095 
2096 		if (ddp_self == NULL)
2097 			return (ZIO_PIPELINE_CONTINUE);
2098 
2099 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2100 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2101 				continue;
2102 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2103 			    &blk);
2104 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
2105 			    zio_buf_alloc(zio->io_size), zio->io_size,
2106 			    zio_ddt_child_read_done, dde, zio->io_priority,
2107 			    ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2108 			    &zio->io_bookmark));
2109 		}
2110 		return (ZIO_PIPELINE_CONTINUE);
2111 	}
2112 
2113 	zio_nowait(zio_read(zio, zio->io_spa, bp,
2114 	    zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2115 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2116 
2117 	return (ZIO_PIPELINE_CONTINUE);
2118 }
2119 
2120 static int
2121 zio_ddt_read_done(zio_t *zio)
2122 {
2123 	blkptr_t *bp = zio->io_bp;
2124 
2125 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2126 		return (ZIO_PIPELINE_STOP);
2127 
2128 	ASSERT(BP_GET_DEDUP(bp));
2129 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2130 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2131 
2132 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
2133 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
2134 		ddt_entry_t *dde = zio->io_vsd;
2135 		if (ddt == NULL) {
2136 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2137 			return (ZIO_PIPELINE_CONTINUE);
2138 		}
2139 		if (dde == NULL) {
2140 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2141 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2142 			return (ZIO_PIPELINE_STOP);
2143 		}
2144 		if (dde->dde_repair_data != NULL) {
2145 			bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2146 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
2147 		}
2148 		ddt_repair_done(ddt, dde);
2149 		zio->io_vsd = NULL;
2150 	}
2151 
2152 	ASSERT(zio->io_vsd == NULL);
2153 
2154 	return (ZIO_PIPELINE_CONTINUE);
2155 }
2156 
2157 static boolean_t
2158 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2159 {
2160 	spa_t *spa = zio->io_spa;
2161 
2162 	/*
2163 	 * Note: we compare the original data, not the transformed data,
2164 	 * because when zio->io_bp is an override bp, we will not have
2165 	 * pushed the I/O transforms.  That's an important optimization
2166 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2167 	 */
2168 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2169 		zio_t *lio = dde->dde_lead_zio[p];
2170 
2171 		if (lio != NULL) {
2172 			return (lio->io_orig_size != zio->io_orig_size ||
2173 			    bcmp(zio->io_orig_data, lio->io_orig_data,
2174 			    zio->io_orig_size) != 0);
2175 		}
2176 	}
2177 
2178 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2179 		ddt_phys_t *ddp = &dde->dde_phys[p];
2180 
2181 		if (ddp->ddp_phys_birth != 0) {
2182 			arc_buf_t *abuf = NULL;
2183 			arc_flags_t aflags = ARC_FLAG_WAIT;
2184 			blkptr_t blk = *zio->io_bp;
2185 			int error;
2186 
2187 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2188 
2189 			ddt_exit(ddt);
2190 
2191 			error = arc_read(NULL, spa, &blk,
2192 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2193 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2194 			    &aflags, &zio->io_bookmark);
2195 
2196 			if (error == 0) {
2197 				if (arc_buf_size(abuf) != zio->io_orig_size ||
2198 				    bcmp(abuf->b_data, zio->io_orig_data,
2199 				    zio->io_orig_size) != 0)
2200 					error = SET_ERROR(EEXIST);
2201 				VERIFY(arc_buf_remove_ref(abuf, &abuf));
2202 			}
2203 
2204 			ddt_enter(ddt);
2205 			return (error != 0);
2206 		}
2207 	}
2208 
2209 	return (B_FALSE);
2210 }
2211 
2212 static void
2213 zio_ddt_child_write_ready(zio_t *zio)
2214 {
2215 	int p = zio->io_prop.zp_copies;
2216 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2217 	ddt_entry_t *dde = zio->io_private;
2218 	ddt_phys_t *ddp = &dde->dde_phys[p];
2219 	zio_t *pio;
2220 
2221 	if (zio->io_error)
2222 		return;
2223 
2224 	ddt_enter(ddt);
2225 
2226 	ASSERT(dde->dde_lead_zio[p] == zio);
2227 
2228 	ddt_phys_fill(ddp, zio->io_bp);
2229 
2230 	while ((pio = zio_walk_parents(zio)) != NULL)
2231 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2232 
2233 	ddt_exit(ddt);
2234 }
2235 
2236 static void
2237 zio_ddt_child_write_done(zio_t *zio)
2238 {
2239 	int p = zio->io_prop.zp_copies;
2240 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2241 	ddt_entry_t *dde = zio->io_private;
2242 	ddt_phys_t *ddp = &dde->dde_phys[p];
2243 
2244 	ddt_enter(ddt);
2245 
2246 	ASSERT(ddp->ddp_refcnt == 0);
2247 	ASSERT(dde->dde_lead_zio[p] == zio);
2248 	dde->dde_lead_zio[p] = NULL;
2249 
2250 	if (zio->io_error == 0) {
2251 		while (zio_walk_parents(zio) != NULL)
2252 			ddt_phys_addref(ddp);
2253 	} else {
2254 		ddt_phys_clear(ddp);
2255 	}
2256 
2257 	ddt_exit(ddt);
2258 }
2259 
2260 static void
2261 zio_ddt_ditto_write_done(zio_t *zio)
2262 {
2263 	int p = DDT_PHYS_DITTO;
2264 	zio_prop_t *zp = &zio->io_prop;
2265 	blkptr_t *bp = zio->io_bp;
2266 	ddt_t *ddt = ddt_select(zio->io_spa, bp);
2267 	ddt_entry_t *dde = zio->io_private;
2268 	ddt_phys_t *ddp = &dde->dde_phys[p];
2269 	ddt_key_t *ddk = &dde->dde_key;
2270 
2271 	ddt_enter(ddt);
2272 
2273 	ASSERT(ddp->ddp_refcnt == 0);
2274 	ASSERT(dde->dde_lead_zio[p] == zio);
2275 	dde->dde_lead_zio[p] = NULL;
2276 
2277 	if (zio->io_error == 0) {
2278 		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2279 		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2280 		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2281 		if (ddp->ddp_phys_birth != 0)
2282 			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2283 		ddt_phys_fill(ddp, bp);
2284 	}
2285 
2286 	ddt_exit(ddt);
2287 }
2288 
2289 static int
2290 zio_ddt_write(zio_t *zio)
2291 {
2292 	spa_t *spa = zio->io_spa;
2293 	blkptr_t *bp = zio->io_bp;
2294 	uint64_t txg = zio->io_txg;
2295 	zio_prop_t *zp = &zio->io_prop;
2296 	int p = zp->zp_copies;
2297 	int ditto_copies;
2298 	zio_t *cio = NULL;
2299 	zio_t *dio = NULL;
2300 	ddt_t *ddt = ddt_select(spa, bp);
2301 	ddt_entry_t *dde;
2302 	ddt_phys_t *ddp;
2303 
2304 	ASSERT(BP_GET_DEDUP(bp));
2305 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2306 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2307 
2308 	ddt_enter(ddt);
2309 	dde = ddt_lookup(ddt, bp, B_TRUE);
2310 	ddp = &dde->dde_phys[p];
2311 
2312 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2313 		/*
2314 		 * If we're using a weak checksum, upgrade to a strong checksum
2315 		 * and try again.  If we're already using a strong checksum,
2316 		 * we can't resolve it, so just convert to an ordinary write.
2317 		 * (And automatically e-mail a paper to Nature?)
2318 		 */
2319 		if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2320 		    ZCHECKSUM_FLAG_DEDUP)) {
2321 			zp->zp_checksum = spa_dedup_checksum(spa);
2322 			zio_pop_transforms(zio);
2323 			zio->io_stage = ZIO_STAGE_OPEN;
2324 			BP_ZERO(bp);
2325 		} else {
2326 			zp->zp_dedup = B_FALSE;
2327 		}
2328 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2329 		ddt_exit(ddt);
2330 		return (ZIO_PIPELINE_CONTINUE);
2331 	}
2332 
2333 	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2334 	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2335 
2336 	if (ditto_copies > ddt_ditto_copies_present(dde) &&
2337 	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2338 		zio_prop_t czp = *zp;
2339 
2340 		czp.zp_copies = ditto_copies;
2341 
2342 		/*
2343 		 * If we arrived here with an override bp, we won't have run
2344 		 * the transform stack, so we won't have the data we need to
2345 		 * generate a child i/o.  So, toss the override bp and restart.
2346 		 * This is safe, because using the override bp is just an
2347 		 * optimization; and it's rare, so the cost doesn't matter.
2348 		 */
2349 		if (zio->io_bp_override) {
2350 			zio_pop_transforms(zio);
2351 			zio->io_stage = ZIO_STAGE_OPEN;
2352 			zio->io_pipeline = ZIO_WRITE_PIPELINE;
2353 			zio->io_bp_override = NULL;
2354 			BP_ZERO(bp);
2355 			ddt_exit(ddt);
2356 			return (ZIO_PIPELINE_CONTINUE);
2357 		}
2358 
2359 		dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2360 		    zio->io_orig_size, &czp, NULL, NULL,
2361 		    zio_ddt_ditto_write_done, dde, zio->io_priority,
2362 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2363 
2364 		zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2365 		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2366 	}
2367 
2368 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2369 		if (ddp->ddp_phys_birth != 0)
2370 			ddt_bp_fill(ddp, bp, txg);
2371 		if (dde->dde_lead_zio[p] != NULL)
2372 			zio_add_child(zio, dde->dde_lead_zio[p]);
2373 		else
2374 			ddt_phys_addref(ddp);
2375 	} else if (zio->io_bp_override) {
2376 		ASSERT(bp->blk_birth == txg);
2377 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2378 		ddt_phys_fill(ddp, bp);
2379 		ddt_phys_addref(ddp);
2380 	} else {
2381 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2382 		    zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2383 		    zio_ddt_child_write_done, dde, zio->io_priority,
2384 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2385 
2386 		zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2387 		dde->dde_lead_zio[p] = cio;
2388 	}
2389 
2390 	ddt_exit(ddt);
2391 
2392 	if (cio)
2393 		zio_nowait(cio);
2394 	if (dio)
2395 		zio_nowait(dio);
2396 
2397 	return (ZIO_PIPELINE_CONTINUE);
2398 }
2399 
2400 ddt_entry_t *freedde; /* for debugging */
2401 
2402 static int
2403 zio_ddt_free(zio_t *zio)
2404 {
2405 	spa_t *spa = zio->io_spa;
2406 	blkptr_t *bp = zio->io_bp;
2407 	ddt_t *ddt = ddt_select(spa, bp);
2408 	ddt_entry_t *dde;
2409 	ddt_phys_t *ddp;
2410 
2411 	ASSERT(BP_GET_DEDUP(bp));
2412 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2413 
2414 	ddt_enter(ddt);
2415 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2416 	ddp = ddt_phys_select(dde, bp);
2417 	ddt_phys_decref(ddp);
2418 	ddt_exit(ddt);
2419 
2420 	return (ZIO_PIPELINE_CONTINUE);
2421 }
2422 
2423 /*
2424  * ==========================================================================
2425  * Allocate and free blocks
2426  * ==========================================================================
2427  */
2428 static int
2429 zio_dva_allocate(zio_t *zio)
2430 {
2431 	spa_t *spa = zio->io_spa;
2432 	metaslab_class_t *mc = spa_normal_class(spa);
2433 	blkptr_t *bp = zio->io_bp;
2434 	int error;
2435 	int flags = 0;
2436 
2437 	if (zio->io_gang_leader == NULL) {
2438 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2439 		zio->io_gang_leader = zio;
2440 	}
2441 
2442 	ASSERT(BP_IS_HOLE(bp));
2443 	ASSERT0(BP_GET_NDVAS(bp));
2444 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
2445 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2446 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2447 
2448 	/*
2449 	 * The dump device does not support gang blocks so allocation on
2450 	 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2451 	 * the "fast" gang feature.
2452 	 */
2453 	flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2454 	flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2455 	    METASLAB_GANG_CHILD : 0;
2456 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
2457 	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2458 
2459 	if (error) {
2460 		spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2461 		    "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2462 		    error);
2463 		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2464 			return (zio_write_gang_block(zio));
2465 		zio->io_error = error;
2466 	}
2467 
2468 	return (ZIO_PIPELINE_CONTINUE);
2469 }
2470 
2471 static int
2472 zio_dva_free(zio_t *zio)
2473 {
2474 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2475 
2476 	return (ZIO_PIPELINE_CONTINUE);
2477 }
2478 
2479 static int
2480 zio_dva_claim(zio_t *zio)
2481 {
2482 	int error;
2483 
2484 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2485 	if (error)
2486 		zio->io_error = error;
2487 
2488 	return (ZIO_PIPELINE_CONTINUE);
2489 }
2490 
2491 /*
2492  * Undo an allocation.  This is used by zio_done() when an I/O fails
2493  * and we want to give back the block we just allocated.
2494  * This handles both normal blocks and gang blocks.
2495  */
2496 static void
2497 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2498 {
2499 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2500 	ASSERT(zio->io_bp_override == NULL);
2501 
2502 	if (!BP_IS_HOLE(bp))
2503 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2504 
2505 	if (gn != NULL) {
2506 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2507 			zio_dva_unallocate(zio, gn->gn_child[g],
2508 			    &gn->gn_gbh->zg_blkptr[g]);
2509 		}
2510 	}
2511 }
2512 
2513 /*
2514  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
2515  */
2516 int
2517 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2518     uint64_t size, boolean_t use_slog)
2519 {
2520 	int error = 1;
2521 
2522 	ASSERT(txg > spa_syncing_txg(spa));
2523 
2524 	/*
2525 	 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2526 	 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2527 	 * when allocating them.
2528 	 */
2529 	if (use_slog) {
2530 		error = metaslab_alloc(spa, spa_log_class(spa), size,
2531 		    new_bp, 1, txg, old_bp,
2532 		    METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2533 	}
2534 
2535 	if (error) {
2536 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
2537 		    new_bp, 1, txg, old_bp,
2538 		    METASLAB_HINTBP_AVOID);
2539 	}
2540 
2541 	if (error == 0) {
2542 		BP_SET_LSIZE(new_bp, size);
2543 		BP_SET_PSIZE(new_bp, size);
2544 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2545 		BP_SET_CHECKSUM(new_bp,
2546 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2547 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2548 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2549 		BP_SET_LEVEL(new_bp, 0);
2550 		BP_SET_DEDUP(new_bp, 0);
2551 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2552 	}
2553 
2554 	return (error);
2555 }
2556 
2557 /*
2558  * Free an intent log block.
2559  */
2560 void
2561 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2562 {
2563 	ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2564 	ASSERT(!BP_IS_GANG(bp));
2565 
2566 	zio_free(spa, txg, bp);
2567 }
2568 
2569 /*
2570  * ==========================================================================
2571  * Read and write to physical devices
2572  * ==========================================================================
2573  */
2574 
2575 
2576 /*
2577  * Issue an I/O to the underlying vdev. Typically the issue pipeline
2578  * stops after this stage and will resume upon I/O completion.
2579  * However, there are instances where the vdev layer may need to
2580  * continue the pipeline when an I/O was not issued. Since the I/O
2581  * that was sent to the vdev layer might be different than the one
2582  * currently active in the pipeline (see vdev_queue_io()), we explicitly
2583  * force the underlying vdev layers to call either zio_execute() or
2584  * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
2585  */
2586 static int
2587 zio_vdev_io_start(zio_t *zio)
2588 {
2589 	vdev_t *vd = zio->io_vd;
2590 	uint64_t align;
2591 	spa_t *spa = zio->io_spa;
2592 
2593 	ASSERT(zio->io_error == 0);
2594 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2595 
2596 	if (vd == NULL) {
2597 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2598 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2599 
2600 		/*
2601 		 * The mirror_ops handle multiple DVAs in a single BP.
2602 		 */
2603 		vdev_mirror_ops.vdev_op_io_start(zio);
2604 		return (ZIO_PIPELINE_STOP);
2605 	}
2606 
2607 	/*
2608 	 * We keep track of time-sensitive I/Os so that the scan thread
2609 	 * can quickly react to certain workloads.  In particular, we care
2610 	 * about non-scrubbing, top-level reads and writes with the following
2611 	 * characteristics:
2612 	 *	- synchronous writes of user data to non-slog devices
2613 	 *	- any reads of user data
2614 	 * When these conditions are met, adjust the timestamp of spa_last_io
2615 	 * which allows the scan thread to adjust its workload accordingly.
2616 	 */
2617 	if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2618 	    vd == vd->vdev_top && !vd->vdev_islog &&
2619 	    zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2620 	    zio->io_txg != spa_syncing_txg(spa)) {
2621 		uint64_t old = spa->spa_last_io;
2622 		uint64_t new = ddi_get_lbolt64();
2623 		if (old != new)
2624 			(void) atomic_cas_64(&spa->spa_last_io, old, new);
2625 	}
2626 
2627 	align = 1ULL << vd->vdev_top->vdev_ashift;
2628 
2629 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
2630 	    P2PHASE(zio->io_size, align) != 0) {
2631 		/* Transform logical writes to be a full physical block size. */
2632 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
2633 		char *abuf = zio_buf_alloc(asize);
2634 		ASSERT(vd == vd->vdev_top);
2635 		if (zio->io_type == ZIO_TYPE_WRITE) {
2636 			bcopy(zio->io_data, abuf, zio->io_size);
2637 			bzero(abuf + zio->io_size, asize - zio->io_size);
2638 		}
2639 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2640 	}
2641 
2642 	/*
2643 	 * If this is not a physical io, make sure that it is properly aligned
2644 	 * before proceeding.
2645 	 */
2646 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2647 		ASSERT0(P2PHASE(zio->io_offset, align));
2648 		ASSERT0(P2PHASE(zio->io_size, align));
2649 	} else {
2650 		/*
2651 		 * For physical writes, we allow 512b aligned writes and assume
2652 		 * the device will perform a read-modify-write as necessary.
2653 		 */
2654 		ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2655 		ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2656 	}
2657 
2658 	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2659 
2660 	/*
2661 	 * If this is a repair I/O, and there's no self-healing involved --
2662 	 * that is, we're just resilvering what we expect to resilver --
2663 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2664 	 * This prevents spurious resilvering with nested replication.
2665 	 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2666 	 * A is out of date, we'll read from C+D, then use the data to
2667 	 * resilver A+B -- but we don't actually want to resilver B, just A.
2668 	 * The top-level mirror has no way to know this, so instead we just
2669 	 * discard unnecessary repairs as we work our way down the vdev tree.
2670 	 * The same logic applies to any form of nested replication:
2671 	 * ditto + mirror, RAID-Z + replacing, etc.  This covers them all.
2672 	 */
2673 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2674 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2675 	    zio->io_txg != 0 &&	/* not a delegated i/o */
2676 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2677 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2678 		zio_vdev_io_bypass(zio);
2679 		return (ZIO_PIPELINE_CONTINUE);
2680 	}
2681 
2682 	if (vd->vdev_ops->vdev_op_leaf &&
2683 	    (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2684 
2685 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
2686 			return (ZIO_PIPELINE_CONTINUE);
2687 
2688 		if ((zio = vdev_queue_io(zio)) == NULL)
2689 			return (ZIO_PIPELINE_STOP);
2690 
2691 		if (!vdev_accessible(vd, zio)) {
2692 			zio->io_error = SET_ERROR(ENXIO);
2693 			zio_interrupt(zio);
2694 			return (ZIO_PIPELINE_STOP);
2695 		}
2696 	}
2697 
2698 	vd->vdev_ops->vdev_op_io_start(zio);
2699 	return (ZIO_PIPELINE_STOP);
2700 }
2701 
2702 static int
2703 zio_vdev_io_done(zio_t *zio)
2704 {
2705 	vdev_t *vd = zio->io_vd;
2706 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2707 	boolean_t unexpected_error = B_FALSE;
2708 
2709 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2710 		return (ZIO_PIPELINE_STOP);
2711 
2712 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2713 
2714 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2715 
2716 		vdev_queue_io_done(zio);
2717 
2718 		if (zio->io_type == ZIO_TYPE_WRITE)
2719 			vdev_cache_write(zio);
2720 
2721 		if (zio_injection_enabled && zio->io_error == 0)
2722 			zio->io_error = zio_handle_device_injection(vd,
2723 			    zio, EIO);
2724 
2725 		if (zio_injection_enabled && zio->io_error == 0)
2726 			zio->io_error = zio_handle_label_injection(zio, EIO);
2727 
2728 		if (zio->io_error) {
2729 			if (!vdev_accessible(vd, zio)) {
2730 				zio->io_error = SET_ERROR(ENXIO);
2731 			} else {
2732 				unexpected_error = B_TRUE;
2733 			}
2734 		}
2735 	}
2736 
2737 	ops->vdev_op_io_done(zio);
2738 
2739 	if (unexpected_error)
2740 		VERIFY(vdev_probe(vd, zio) == NULL);
2741 
2742 	return (ZIO_PIPELINE_CONTINUE);
2743 }
2744 
2745 /*
2746  * For non-raidz ZIOs, we can just copy aside the bad data read from the
2747  * disk, and use that to finish the checksum ereport later.
2748  */
2749 static void
2750 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2751     const void *good_buf)
2752 {
2753 	/* no processing needed */
2754 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2755 }
2756 
2757 /*ARGSUSED*/
2758 void
2759 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2760 {
2761 	void *buf = zio_buf_alloc(zio->io_size);
2762 
2763 	bcopy(zio->io_data, buf, zio->io_size);
2764 
2765 	zcr->zcr_cbinfo = zio->io_size;
2766 	zcr->zcr_cbdata = buf;
2767 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
2768 	zcr->zcr_free = zio_buf_free;
2769 }
2770 
2771 static int
2772 zio_vdev_io_assess(zio_t *zio)
2773 {
2774 	vdev_t *vd = zio->io_vd;
2775 
2776 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2777 		return (ZIO_PIPELINE_STOP);
2778 
2779 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2780 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2781 
2782 	if (zio->io_vsd != NULL) {
2783 		zio->io_vsd_ops->vsd_free(zio);
2784 		zio->io_vsd = NULL;
2785 	}
2786 
2787 	if (zio_injection_enabled && zio->io_error == 0)
2788 		zio->io_error = zio_handle_fault_injection(zio, EIO);
2789 
2790 	/*
2791 	 * If the I/O failed, determine whether we should attempt to retry it.
2792 	 *
2793 	 * On retry, we cut in line in the issue queue, since we don't want
2794 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2795 	 */
2796 	if (zio->io_error && vd == NULL &&
2797 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2798 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
2799 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
2800 		zio->io_error = 0;
2801 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
2802 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2803 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2804 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2805 		    zio_requeue_io_start_cut_in_line);
2806 		return (ZIO_PIPELINE_STOP);
2807 	}
2808 
2809 	/*
2810 	 * If we got an error on a leaf device, convert it to ENXIO
2811 	 * if the device is not accessible at all.
2812 	 */
2813 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2814 	    !vdev_accessible(vd, zio))
2815 		zio->io_error = SET_ERROR(ENXIO);
2816 
2817 	/*
2818 	 * If we can't write to an interior vdev (mirror or RAID-Z),
2819 	 * set vdev_cant_write so that we stop trying to allocate from it.
2820 	 */
2821 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2822 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2823 		vd->vdev_cant_write = B_TRUE;
2824 	}
2825 
2826 	if (zio->io_error)
2827 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2828 
2829 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2830 	    zio->io_physdone != NULL) {
2831 		ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2832 		ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2833 		zio->io_physdone(zio->io_logical);
2834 	}
2835 
2836 	return (ZIO_PIPELINE_CONTINUE);
2837 }
2838 
2839 void
2840 zio_vdev_io_reissue(zio_t *zio)
2841 {
2842 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2843 	ASSERT(zio->io_error == 0);
2844 
2845 	zio->io_stage >>= 1;
2846 }
2847 
2848 void
2849 zio_vdev_io_redone(zio_t *zio)
2850 {
2851 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2852 
2853 	zio->io_stage >>= 1;
2854 }
2855 
2856 void
2857 zio_vdev_io_bypass(zio_t *zio)
2858 {
2859 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2860 	ASSERT(zio->io_error == 0);
2861 
2862 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2863 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2864 }
2865 
2866 /*
2867  * ==========================================================================
2868  * Generate and verify checksums
2869  * ==========================================================================
2870  */
2871 static int
2872 zio_checksum_generate(zio_t *zio)
2873 {
2874 	blkptr_t *bp = zio->io_bp;
2875 	enum zio_checksum checksum;
2876 
2877 	if (bp == NULL) {
2878 		/*
2879 		 * This is zio_write_phys().
2880 		 * We're either generating a label checksum, or none at all.
2881 		 */
2882 		checksum = zio->io_prop.zp_checksum;
2883 
2884 		if (checksum == ZIO_CHECKSUM_OFF)
2885 			return (ZIO_PIPELINE_CONTINUE);
2886 
2887 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2888 	} else {
2889 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2890 			ASSERT(!IO_IS_ALLOCATING(zio));
2891 			checksum = ZIO_CHECKSUM_GANG_HEADER;
2892 		} else {
2893 			checksum = BP_GET_CHECKSUM(bp);
2894 		}
2895 	}
2896 
2897 	zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2898 
2899 	return (ZIO_PIPELINE_CONTINUE);
2900 }
2901 
2902 static int
2903 zio_checksum_verify(zio_t *zio)
2904 {
2905 	zio_bad_cksum_t info;
2906 	blkptr_t *bp = zio->io_bp;
2907 	int error;
2908 
2909 	ASSERT(zio->io_vd != NULL);
2910 
2911 	if (bp == NULL) {
2912 		/*
2913 		 * This is zio_read_phys().
2914 		 * We're either verifying a label checksum, or nothing at all.
2915 		 */
2916 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2917 			return (ZIO_PIPELINE_CONTINUE);
2918 
2919 		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2920 	}
2921 
2922 	if ((error = zio_checksum_error(zio, &info)) != 0) {
2923 		zio->io_error = error;
2924 		if (error == ECKSUM &&
2925 		    !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2926 			zfs_ereport_start_checksum(zio->io_spa,
2927 			    zio->io_vd, zio, zio->io_offset,
2928 			    zio->io_size, NULL, &info);
2929 		}
2930 	}
2931 
2932 	return (ZIO_PIPELINE_CONTINUE);
2933 }
2934 
2935 /*
2936  * Called by RAID-Z to ensure we don't compute the checksum twice.
2937  */
2938 void
2939 zio_checksum_verified(zio_t *zio)
2940 {
2941 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2942 }
2943 
2944 /*
2945  * ==========================================================================
2946  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2947  * An error of 0 indicates success.  ENXIO indicates whole-device failure,
2948  * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
2949  * indicate errors that are specific to one I/O, and most likely permanent.
2950  * Any other error is presumed to be worse because we weren't expecting it.
2951  * ==========================================================================
2952  */
2953 int
2954 zio_worst_error(int e1, int e2)
2955 {
2956 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2957 	int r1, r2;
2958 
2959 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2960 		if (e1 == zio_error_rank[r1])
2961 			break;
2962 
2963 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2964 		if (e2 == zio_error_rank[r2])
2965 			break;
2966 
2967 	return (r1 > r2 ? e1 : e2);
2968 }
2969 
2970 /*
2971  * ==========================================================================
2972  * I/O completion
2973  * ==========================================================================
2974  */
2975 static int
2976 zio_ready(zio_t *zio)
2977 {
2978 	blkptr_t *bp = zio->io_bp;
2979 	zio_t *pio, *pio_next;
2980 
2981 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2982 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2983 		return (ZIO_PIPELINE_STOP);
2984 
2985 	if (zio->io_ready) {
2986 		ASSERT(IO_IS_ALLOCATING(zio));
2987 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2988 		    (zio->io_flags & ZIO_FLAG_NOPWRITE));
2989 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2990 
2991 		zio->io_ready(zio);
2992 	}
2993 
2994 	if (bp != NULL && bp != &zio->io_bp_copy)
2995 		zio->io_bp_copy = *bp;
2996 
2997 	if (zio->io_error)
2998 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2999 
3000 	mutex_enter(&zio->io_lock);
3001 	zio->io_state[ZIO_WAIT_READY] = 1;
3002 	pio = zio_walk_parents(zio);
3003 	mutex_exit(&zio->io_lock);
3004 
3005 	/*
3006 	 * As we notify zio's parents, new parents could be added.
3007 	 * New parents go to the head of zio's io_parent_list, however,
3008 	 * so we will (correctly) not notify them.  The remainder of zio's
3009 	 * io_parent_list, from 'pio_next' onward, cannot change because
3010 	 * all parents must wait for us to be done before they can be done.
3011 	 */
3012 	for (; pio != NULL; pio = pio_next) {
3013 		pio_next = zio_walk_parents(zio);
3014 		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3015 	}
3016 
3017 	if (zio->io_flags & ZIO_FLAG_NODATA) {
3018 		if (BP_IS_GANG(bp)) {
3019 			zio->io_flags &= ~ZIO_FLAG_NODATA;
3020 		} else {
3021 			ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3022 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3023 		}
3024 	}
3025 
3026 	if (zio_injection_enabled &&
3027 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
3028 		zio_handle_ignored_writes(zio);
3029 
3030 	return (ZIO_PIPELINE_CONTINUE);
3031 }
3032 
3033 static int
3034 zio_done(zio_t *zio)
3035 {
3036 	spa_t *spa = zio->io_spa;
3037 	zio_t *lio = zio->io_logical;
3038 	blkptr_t *bp = zio->io_bp;
3039 	vdev_t *vd = zio->io_vd;
3040 	uint64_t psize = zio->io_size;
3041 	zio_t *pio, *pio_next;
3042 
3043 	/*
3044 	 * If our children haven't all completed,
3045 	 * wait for them and then repeat this pipeline stage.
3046 	 */
3047 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3048 	    zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3049 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3050 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3051 		return (ZIO_PIPELINE_STOP);
3052 
3053 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3054 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3055 			ASSERT(zio->io_children[c][w] == 0);
3056 
3057 	if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3058 		ASSERT(bp->blk_pad[0] == 0);
3059 		ASSERT(bp->blk_pad[1] == 0);
3060 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3061 		    (bp == zio_unique_parent(zio)->io_bp));
3062 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3063 		    zio->io_bp_override == NULL &&
3064 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3065 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
3066 			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3067 			ASSERT(BP_COUNT_GANG(bp) == 0 ||
3068 			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3069 		}
3070 		if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3071 			VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3072 	}
3073 
3074 	/*
3075 	 * If there were child vdev/gang/ddt errors, they apply to us now.
3076 	 */
3077 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3078 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3079 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3080 
3081 	/*
3082 	 * If the I/O on the transformed data was successful, generate any
3083 	 * checksum reports now while we still have the transformed data.
3084 	 */
3085 	if (zio->io_error == 0) {
3086 		while (zio->io_cksum_report != NULL) {
3087 			zio_cksum_report_t *zcr = zio->io_cksum_report;
3088 			uint64_t align = zcr->zcr_align;
3089 			uint64_t asize = P2ROUNDUP(psize, align);
3090 			char *abuf = zio->io_data;
3091 
3092 			if (asize != psize) {
3093 				abuf = zio_buf_alloc(asize);
3094 				bcopy(zio->io_data, abuf, psize);
3095 				bzero(abuf + psize, asize - psize);
3096 			}
3097 
3098 			zio->io_cksum_report = zcr->zcr_next;
3099 			zcr->zcr_next = NULL;
3100 			zcr->zcr_finish(zcr, abuf);
3101 			zfs_ereport_free_checksum(zcr);
3102 
3103 			if (asize != psize)
3104 				zio_buf_free(abuf, asize);
3105 		}
3106 	}
3107 
3108 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
3109 
3110 	vdev_stat_update(zio, psize);
3111 
3112 	if (zio->io_error) {
3113 		/*
3114 		 * If this I/O is attached to a particular vdev,
3115 		 * generate an error message describing the I/O failure
3116 		 * at the block level.  We ignore these errors if the
3117 		 * device is currently unavailable.
3118 		 */
3119 		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3120 			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3121 
3122 		if ((zio->io_error == EIO || !(zio->io_flags &
3123 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3124 		    zio == lio) {
3125 			/*
3126 			 * For logical I/O requests, tell the SPA to log the
3127 			 * error and generate a logical data ereport.
3128 			 */
3129 			spa_log_error(spa, zio);
3130 			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3131 			    0, 0);
3132 		}
3133 	}
3134 
3135 	if (zio->io_error && zio == lio) {
3136 		/*
3137 		 * Determine whether zio should be reexecuted.  This will
3138 		 * propagate all the way to the root via zio_notify_parent().
3139 		 */
3140 		ASSERT(vd == NULL && bp != NULL);
3141 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3142 
3143 		if (IO_IS_ALLOCATING(zio) &&
3144 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3145 			if (zio->io_error != ENOSPC)
3146 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3147 			else
3148 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3149 		}
3150 
3151 		if ((zio->io_type == ZIO_TYPE_READ ||
3152 		    zio->io_type == ZIO_TYPE_FREE) &&
3153 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3154 		    zio->io_error == ENXIO &&
3155 		    spa_load_state(spa) == SPA_LOAD_NONE &&
3156 		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3157 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3158 
3159 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3160 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3161 
3162 		/*
3163 		 * Here is a possibly good place to attempt to do
3164 		 * either combinatorial reconstruction or error correction
3165 		 * based on checksums.  It also might be a good place
3166 		 * to send out preliminary ereports before we suspend
3167 		 * processing.
3168 		 */
3169 	}
3170 
3171 	/*
3172 	 * If there were logical child errors, they apply to us now.
3173 	 * We defer this until now to avoid conflating logical child
3174 	 * errors with errors that happened to the zio itself when
3175 	 * updating vdev stats and reporting FMA events above.
3176 	 */
3177 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3178 
3179 	if ((zio->io_error || zio->io_reexecute) &&
3180 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3181 	    !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3182 		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3183 
3184 	zio_gang_tree_free(&zio->io_gang_tree);
3185 
3186 	/*
3187 	 * Godfather I/Os should never suspend.
3188 	 */
3189 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3190 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3191 		zio->io_reexecute = 0;
3192 
3193 	if (zio->io_reexecute) {
3194 		/*
3195 		 * This is a logical I/O that wants to reexecute.
3196 		 *
3197 		 * Reexecute is top-down.  When an i/o fails, if it's not
3198 		 * the root, it simply notifies its parent and sticks around.
3199 		 * The parent, seeing that it still has children in zio_done(),
3200 		 * does the same.  This percolates all the way up to the root.
3201 		 * The root i/o will reexecute or suspend the entire tree.
3202 		 *
3203 		 * This approach ensures that zio_reexecute() honors
3204 		 * all the original i/o dependency relationships, e.g.
3205 		 * parents not executing until children are ready.
3206 		 */
3207 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3208 
3209 		zio->io_gang_leader = NULL;
3210 
3211 		mutex_enter(&zio->io_lock);
3212 		zio->io_state[ZIO_WAIT_DONE] = 1;
3213 		mutex_exit(&zio->io_lock);
3214 
3215 		/*
3216 		 * "The Godfather" I/O monitors its children but is
3217 		 * not a true parent to them. It will track them through
3218 		 * the pipeline but severs its ties whenever they get into
3219 		 * trouble (e.g. suspended). This allows "The Godfather"
3220 		 * I/O to return status without blocking.
3221 		 */
3222 		for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3223 			zio_link_t *zl = zio->io_walk_link;
3224 			pio_next = zio_walk_parents(zio);
3225 
3226 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3227 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3228 				zio_remove_child(pio, zio, zl);
3229 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3230 			}
3231 		}
3232 
3233 		if ((pio = zio_unique_parent(zio)) != NULL) {
3234 			/*
3235 			 * We're not a root i/o, so there's nothing to do
3236 			 * but notify our parent.  Don't propagate errors
3237 			 * upward since we haven't permanently failed yet.
3238 			 */
3239 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3240 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3241 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3242 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3243 			/*
3244 			 * We'd fail again if we reexecuted now, so suspend
3245 			 * until conditions improve (e.g. device comes online).
3246 			 */
3247 			zio_suspend(spa, zio);
3248 		} else {
3249 			/*
3250 			 * Reexecution is potentially a huge amount of work.
3251 			 * Hand it off to the otherwise-unused claim taskq.
3252 			 */
3253 			ASSERT(zio->io_tqent.tqent_next == NULL);
3254 			spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3255 			    ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3256 			    0, &zio->io_tqent);
3257 		}
3258 		return (ZIO_PIPELINE_STOP);
3259 	}
3260 
3261 	ASSERT(zio->io_child_count == 0);
3262 	ASSERT(zio->io_reexecute == 0);
3263 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3264 
3265 	/*
3266 	 * Report any checksum errors, since the I/O is complete.
3267 	 */
3268 	while (zio->io_cksum_report != NULL) {
3269 		zio_cksum_report_t *zcr = zio->io_cksum_report;
3270 		zio->io_cksum_report = zcr->zcr_next;
3271 		zcr->zcr_next = NULL;
3272 		zcr->zcr_finish(zcr, NULL);
3273 		zfs_ereport_free_checksum(zcr);
3274 	}
3275 
3276 	/*
3277 	 * It is the responsibility of the done callback to ensure that this
3278 	 * particular zio is no longer discoverable for adoption, and as
3279 	 * such, cannot acquire any new parents.
3280 	 */
3281 	if (zio->io_done)
3282 		zio->io_done(zio);
3283 
3284 	mutex_enter(&zio->io_lock);
3285 	zio->io_state[ZIO_WAIT_DONE] = 1;
3286 	mutex_exit(&zio->io_lock);
3287 
3288 	for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3289 		zio_link_t *zl = zio->io_walk_link;
3290 		pio_next = zio_walk_parents(zio);
3291 		zio_remove_child(pio, zio, zl);
3292 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3293 	}
3294 
3295 	if (zio->io_waiter != NULL) {
3296 		mutex_enter(&zio->io_lock);
3297 		zio->io_executor = NULL;
3298 		cv_broadcast(&zio->io_cv);
3299 		mutex_exit(&zio->io_lock);
3300 	} else {
3301 		zio_destroy(zio);
3302 	}
3303 
3304 	return (ZIO_PIPELINE_STOP);
3305 }
3306 
3307 /*
3308  * ==========================================================================
3309  * I/O pipeline definition
3310  * ==========================================================================
3311  */
3312 static zio_pipe_stage_t *zio_pipeline[] = {
3313 	NULL,
3314 	zio_read_bp_init,
3315 	zio_free_bp_init,
3316 	zio_issue_async,
3317 	zio_write_bp_init,
3318 	zio_checksum_generate,
3319 	zio_nop_write,
3320 	zio_ddt_read_start,
3321 	zio_ddt_read_done,
3322 	zio_ddt_write,
3323 	zio_ddt_free,
3324 	zio_gang_assemble,
3325 	zio_gang_issue,
3326 	zio_dva_allocate,
3327 	zio_dva_free,
3328 	zio_dva_claim,
3329 	zio_ready,
3330 	zio_vdev_io_start,
3331 	zio_vdev_io_done,
3332 	zio_vdev_io_assess,
3333 	zio_checksum_verify,
3334 	zio_done
3335 };
3336 
3337 
3338 
3339 
3340 /*
3341  * Compare two zbookmark_phys_t's to see which we would reach first in a
3342  * pre-order traversal of the object tree.
3343  *
3344  * This is simple in every case aside from the meta-dnode object. For all other
3345  * objects, we traverse them in order (object 1 before object 2, and so on).
3346  * However, all of these objects are traversed while traversing object 0, since
3347  * the data it points to is the list of objects.  Thus, we need to convert to a
3348  * canonical representation so we can compare meta-dnode bookmarks to
3349  * non-meta-dnode bookmarks.
3350  *
3351  * We do this by calculating "equivalents" for each field of the zbookmark.
3352  * zbookmarks outside of the meta-dnode use their own object and level, and
3353  * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3354  * blocks this bookmark refers to) by multiplying their blkid by their span
3355  * (the number of L0 blocks contained within one block at their level).
3356  * zbookmarks inside the meta-dnode calculate their object equivalent
3357  * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3358  * level + 1<<31 (any value larger than a level could ever be) for their level.
3359  * This causes them to always compare before a bookmark in their object
3360  * equivalent, compare appropriately to bookmarks in other objects, and to
3361  * compare appropriately to other bookmarks in the meta-dnode.
3362  */
3363 int
3364 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
3365     const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
3366 {
3367 	/*
3368 	 * These variables represent the "equivalent" values for the zbookmark,
3369 	 * after converting zbookmarks inside the meta dnode to their
3370 	 * normal-object equivalents.
3371 	 */
3372 	uint64_t zb1obj, zb2obj;
3373 	uint64_t zb1L0, zb2L0;
3374 	uint64_t zb1level, zb2level;
3375 
3376 	if (zb1->zb_object == zb2->zb_object &&
3377 	    zb1->zb_level == zb2->zb_level &&
3378 	    zb1->zb_blkid == zb2->zb_blkid)
3379 		return (0);
3380 
3381 	/*
3382 	 * BP_SPANB calculates the span in blocks.
3383 	 */
3384 	zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
3385 	zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
3386 
3387 	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3388 		zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3389 		zb1L0 = 0;
3390 		zb1level = zb1->zb_level + COMPARE_META_LEVEL;
3391 	} else {
3392 		zb1obj = zb1->zb_object;
3393 		zb1level = zb1->zb_level;
3394 	}
3395 
3396 	if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
3397 		zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3398 		zb2L0 = 0;
3399 		zb2level = zb2->zb_level + COMPARE_META_LEVEL;
3400 	} else {
3401 		zb2obj = zb2->zb_object;
3402 		zb2level = zb2->zb_level;
3403 	}
3404 
3405 	/* Now that we have a canonical representation, do the comparison. */
3406 	if (zb1obj != zb2obj)
3407 		return (zb1obj < zb2obj ? -1 : 1);
3408 	else if (zb1L0 != zb2L0)
3409 		return (zb1L0 < zb2L0 ? -1 : 1);
3410 	else if (zb1level != zb2level)
3411 		return (zb1level > zb2level ? -1 : 1);
3412 	/*
3413 	 * This can (theoretically) happen if the bookmarks have the same object
3414 	 * and level, but different blkids, if the block sizes are not the same.
3415 	 * There is presently no way to change the indirect block sizes
3416 	 */
3417 	return (0);
3418 }
3419 
3420 /*
3421  *  This function checks the following: given that last_block is the place that
3422  *  our traversal stopped last time, does that guarantee that we've visited
3423  *  every node under subtree_root?  Therefore, we can't just use the raw output
3424  *  of zbookmark_compare.  We have to pass in a modified version of
3425  *  subtree_root; by incrementing the block id, and then checking whether
3426  *  last_block is before or equal to that, we can tell whether or not having
3427  *  visited last_block implies that all of subtree_root's children have been
3428  *  visited.
3429  */
3430 boolean_t
3431 zbookmark_subtree_completed(const dnode_phys_t *dnp,
3432     const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
3433 {
3434 	zbookmark_phys_t mod_zb = *subtree_root;
3435 	mod_zb.zb_blkid++;
3436 	ASSERT(last_block->zb_level == 0);
3437 
3438 	/* The objset_phys_t isn't before anything. */
3439 	if (dnp == NULL)
3440 		return (B_FALSE);
3441 
3442 	/*
3443 	 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
3444 	 * data block size in sectors, because that variable is only used if
3445 	 * the bookmark refers to a block in the meta-dnode.  Since we don't
3446 	 * know without examining it what object it refers to, and there's no
3447 	 * harm in passing in this value in other cases, we always pass it in.
3448 	 *
3449 	 * We pass in 0 for the indirect block size shift because zb2 must be
3450 	 * level 0.  The indirect block size is only used to calculate the span
3451 	 * of the bookmark, but since the bookmark must be level 0, the span is
3452 	 * always 1, so the math works out.
3453 	 *
3454 	 * If you make changes to how the zbookmark_compare code works, be sure
3455 	 * to make sure that this code still works afterwards.
3456 	 */
3457 	return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
3458 	    1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
3459 	    last_block) <= 0);
3460 }
3461