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, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
26 * Copyright (c) 2017, Intel Corporation.
27 */
28
29#include <sys/sysmacros.h>
30#include <sys/zfs_context.h>
31#include <sys/fm/fs/zfs.h>
32#include <sys/spa.h>
33#include <sys/txg.h>
34#include <sys/spa_impl.h>
35#include <sys/vdev_impl.h>
36#include <sys/zio_impl.h>
37#include <sys/zio_compress.h>
38#include <sys/zio_checksum.h>
39#include <sys/dmu_objset.h>
40#include <sys/arc.h>
41#include <sys/ddt.h>
42#include <sys/blkptr.h>
43#include <sys/zfeature.h>
44#include <sys/dsl_scan.h>
45#include <sys/metaslab_impl.h>
46#include <sys/abd.h>
47#include <sys/cityhash.h>
48#include <sys/dsl_crypt.h>
49
50/*
51 * ==========================================================================
52 * I/O type descriptions
53 * ==========================================================================
54 */
55const char *zio_type_name[ZIO_TYPES] = {
56	"zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
57	"zio_ioctl"
58};
59
60boolean_t zio_dva_throttle_enabled = B_TRUE;
61
62/*
63 * ==========================================================================
64 * I/O kmem caches
65 * ==========================================================================
66 */
67kmem_cache_t *zio_cache;
68kmem_cache_t *zio_link_cache;
69kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
70kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
71
72#ifdef _KERNEL
73extern vmem_t *zio_alloc_arena;
74#endif
75
76#define	ZIO_PIPELINE_CONTINUE		0x100
77#define	ZIO_PIPELINE_STOP		0x101
78
79#define	BP_SPANB(indblkshift, level) \
80	(((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
81#define	COMPARE_META_LEVEL	0x80000000ul
82/*
83 * The following actions directly effect the spa's sync-to-convergence logic.
84 * The values below define the sync pass when we start performing the action.
85 * Care should be taken when changing these values as they directly impact
86 * spa_sync() performance. Tuning these values may introduce subtle performance
87 * pathologies and should only be done in the context of performance analysis.
88 * These tunables will eventually be removed and replaced with #defines once
89 * enough analysis has been done to determine optimal values.
90 *
91 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
92 * regular blocks are not deferred.
93 */
94int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
95int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
96int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
97
98/*
99 * An allocating zio is one that either currently has the DVA allocate
100 * stage set or will have it later in its lifetime.
101 */
102#define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
103
104boolean_t	zio_requeue_io_start_cut_in_line = B_TRUE;
105
106#ifdef ZFS_DEBUG
107int zio_buf_debug_limit = 16384;
108#else
109int zio_buf_debug_limit = 0;
110#endif
111
112static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
113
114void
115zio_init(void)
116{
117	size_t c;
118	vmem_t *data_alloc_arena = NULL;
119
120#ifdef _KERNEL
121	data_alloc_arena = zio_alloc_arena;
122#endif
123	zio_cache = kmem_cache_create("zio_cache",
124	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
125	zio_link_cache = kmem_cache_create("zio_link_cache",
126	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
127
128	/*
129	 * For small buffers, we want a cache for each multiple of
130	 * SPA_MINBLOCKSIZE.  For larger buffers, we want a cache
131	 * for each quarter-power of 2.
132	 */
133	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
134		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
135		size_t p2 = size;
136		size_t align = 0;
137		size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
138
139		while (!ISP2(p2))
140			p2 &= p2 - 1;
141
142#ifndef _KERNEL
143		/*
144		 * If we are using watchpoints, put each buffer on its own page,
145		 * to eliminate the performance overhead of trapping to the
146		 * kernel when modifying a non-watched buffer that shares the
147		 * page with a watched buffer.
148		 */
149		if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
150			continue;
151#endif
152		if (size <= 4 * SPA_MINBLOCKSIZE) {
153			align = SPA_MINBLOCKSIZE;
154		} else if (IS_P2ALIGNED(size, p2 >> 2)) {
155			align = MIN(p2 >> 2, PAGESIZE);
156		}
157
158		if (align != 0) {
159			char name[36];
160			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
161			zio_buf_cache[c] = kmem_cache_create(name, size,
162			    align, NULL, NULL, NULL, NULL, NULL, cflags);
163
164			/*
165			 * Since zio_data bufs do not appear in crash dumps, we
166			 * pass KMC_NOTOUCH so that no allocator metadata is
167			 * stored with the buffers.
168			 */
169			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
170			zio_data_buf_cache[c] = kmem_cache_create(name, size,
171			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
172			    cflags | KMC_NOTOUCH);
173		}
174	}
175
176	while (--c != 0) {
177		ASSERT(zio_buf_cache[c] != NULL);
178		if (zio_buf_cache[c - 1] == NULL)
179			zio_buf_cache[c - 1] = zio_buf_cache[c];
180
181		ASSERT(zio_data_buf_cache[c] != NULL);
182		if (zio_data_buf_cache[c - 1] == NULL)
183			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
184	}
185
186	zio_inject_init();
187}
188
189void
190zio_fini(void)
191{
192	size_t c;
193	kmem_cache_t *last_cache = NULL;
194	kmem_cache_t *last_data_cache = NULL;
195
196	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
197		if (zio_buf_cache[c] != last_cache) {
198			last_cache = zio_buf_cache[c];
199			kmem_cache_destroy(zio_buf_cache[c]);
200		}
201		zio_buf_cache[c] = NULL;
202
203		if (zio_data_buf_cache[c] != last_data_cache) {
204			last_data_cache = zio_data_buf_cache[c];
205			kmem_cache_destroy(zio_data_buf_cache[c]);
206		}
207		zio_data_buf_cache[c] = NULL;
208	}
209
210	kmem_cache_destroy(zio_link_cache);
211	kmem_cache_destroy(zio_cache);
212
213	zio_inject_fini();
214}
215
216/*
217 * ==========================================================================
218 * Allocate and free I/O buffers
219 * ==========================================================================
220 */
221
222/*
223 * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
224 * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
225 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
226 * excess / transient data in-core during a crashdump.
227 */
228void *
229zio_buf_alloc(size_t size)
230{
231	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
232
233	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
234
235	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
236}
237
238/*
239 * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
240 * crashdump if the kernel panics.  This exists so that we will limit the amount
241 * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
242 * of kernel heap dumped to disk when the kernel panics)
243 */
244void *
245zio_data_buf_alloc(size_t size)
246{
247	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
248
249	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
250
251	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
252}
253
254void
255zio_buf_free(void *buf, size_t size)
256{
257	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
258
259	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
260
261	kmem_cache_free(zio_buf_cache[c], buf);
262}
263
264void
265zio_data_buf_free(void *buf, size_t size)
266{
267	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
268
269	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
270
271	kmem_cache_free(zio_data_buf_cache[c], buf);
272}
273
274/* ARGSUSED */
275static void
276zio_abd_free(void *abd, size_t size)
277{
278	abd_free((abd_t *)abd);
279}
280
281/*
282 * ==========================================================================
283 * Push and pop I/O transform buffers
284 * ==========================================================================
285 */
286void
287zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
288    zio_transform_func_t *transform)
289{
290	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
291
292	/*
293	 * Ensure that anyone expecting this zio to contain a linear ABD isn't
294	 * going to get a nasty surprise when they try to access the data.
295	 */
296	IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
297
298	zt->zt_orig_abd = zio->io_abd;
299	zt->zt_orig_size = zio->io_size;
300	zt->zt_bufsize = bufsize;
301	zt->zt_transform = transform;
302
303	zt->zt_next = zio->io_transform_stack;
304	zio->io_transform_stack = zt;
305
306	zio->io_abd = data;
307	zio->io_size = size;
308}
309
310void
311zio_pop_transforms(zio_t *zio)
312{
313	zio_transform_t *zt;
314
315	while ((zt = zio->io_transform_stack) != NULL) {
316		if (zt->zt_transform != NULL)
317			zt->zt_transform(zio,
318			    zt->zt_orig_abd, zt->zt_orig_size);
319
320		if (zt->zt_bufsize != 0)
321			abd_free(zio->io_abd);
322
323		zio->io_abd = zt->zt_orig_abd;
324		zio->io_size = zt->zt_orig_size;
325		zio->io_transform_stack = zt->zt_next;
326
327		kmem_free(zt, sizeof (zio_transform_t));
328	}
329}
330
331/*
332 * ==========================================================================
333 * I/O transform callbacks for subblocks, decompression, and decryption
334 * ==========================================================================
335 */
336static void
337zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
338{
339	ASSERT(zio->io_size > size);
340
341	if (zio->io_type == ZIO_TYPE_READ)
342		abd_copy(data, zio->io_abd, size);
343}
344
345static void
346zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
347{
348	if (zio->io_error == 0) {
349		void *tmp = abd_borrow_buf(data, size);
350		int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
351		    zio->io_abd, tmp, zio->io_size, size);
352		abd_return_buf_copy(data, tmp, size);
353
354		if (ret != 0)
355			zio->io_error = SET_ERROR(EIO);
356	}
357}
358
359static void
360zio_decrypt(zio_t *zio, abd_t *data, uint64_t size)
361{
362	int ret;
363	void *tmp;
364	blkptr_t *bp = zio->io_bp;
365	spa_t *spa = zio->io_spa;
366	uint64_t dsobj = zio->io_bookmark.zb_objset;
367	uint64_t lsize = BP_GET_LSIZE(bp);
368	dmu_object_type_t ot = BP_GET_TYPE(bp);
369	uint8_t salt[ZIO_DATA_SALT_LEN];
370	uint8_t iv[ZIO_DATA_IV_LEN];
371	uint8_t mac[ZIO_DATA_MAC_LEN];
372	boolean_t no_crypt = B_FALSE;
373
374	ASSERT(BP_USES_CRYPT(bp));
375	ASSERT3U(size, !=, 0);
376
377	if (zio->io_error != 0)
378		return;
379
380	/*
381	 * Verify the cksum of MACs stored in an indirect bp. It will always
382	 * be possible to verify this since it does not require an encryption
383	 * key.
384	 */
385	if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {
386		zio_crypt_decode_mac_bp(bp, mac);
387
388		if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
389			/*
390			 * We haven't decompressed the data yet, but
391			 * zio_crypt_do_indirect_mac_checksum() requires
392			 * decompressed data to be able to parse out the MACs
393			 * from the indirect block. We decompress it now and
394			 * throw away the result after we are finished.
395			 */
396			tmp = zio_buf_alloc(lsize);
397			ret = zio_decompress_data(BP_GET_COMPRESS(bp),
398			    zio->io_abd, tmp, zio->io_size, lsize);
399			if (ret != 0) {
400				ret = SET_ERROR(EIO);
401				goto error;
402			}
403			ret = zio_crypt_do_indirect_mac_checksum(B_FALSE,
404			    tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac);
405			zio_buf_free(tmp, lsize);
406		} else {
407			ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
408			    zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac);
409		}
410		abd_copy(data, zio->io_abd, size);
411
412		if (ret != 0)
413			goto error;
414
415		return;
416	}
417
418	/*
419	 * If this is an authenticated block, just check the MAC. It would be
420	 * nice to separate this out into its own flag, but for the moment
421	 * enum zio_flag is out of bits.
422	 */
423	if (BP_IS_AUTHENTICATED(bp)) {
424		if (ot == DMU_OT_OBJSET) {
425			ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa,
426			    dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp));
427		} else {
428			zio_crypt_decode_mac_bp(bp, mac);
429			ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj,
430			    zio->io_abd, size, mac);
431		}
432		abd_copy(data, zio->io_abd, size);
433
434		if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) {
435			ret = zio_handle_decrypt_injection(spa,
436			    &zio->io_bookmark, ot, ECKSUM);
437		}
438		if (ret != 0)
439			goto error;
440
441		return;
442	}
443
444	zio_crypt_decode_params_bp(bp, salt, iv);
445
446	if (ot == DMU_OT_INTENT_LOG) {
447		tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t));
448		zio_crypt_decode_mac_zil(tmp, mac);
449		abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t));
450	} else {
451		zio_crypt_decode_mac_bp(bp, mac);
452	}
453
454	ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp),
455	    BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data,
456	    zio->io_abd, &no_crypt);
457	if (no_crypt)
458		abd_copy(data, zio->io_abd, size);
459
460	if (ret != 0)
461		goto error;
462
463	return;
464
465error:
466	/* assert that the key was found unless this was speculative */
467	ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE));
468
469	/*
470	 * If there was a decryption / authentication error return EIO as
471	 * the io_error. If this was not a speculative zio, create an ereport.
472	 */
473	if (ret == ECKSUM) {
474		zio->io_error = SET_ERROR(EIO);
475		if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
476			spa_log_error(spa, &zio->io_bookmark);
477			zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
478			    spa, NULL, &zio->io_bookmark, zio, 0, 0);
479		}
480	} else {
481		zio->io_error = ret;
482	}
483}
484
485/*
486 * ==========================================================================
487 * I/O parent/child relationships and pipeline interlocks
488 * ==========================================================================
489 */
490zio_t *
491zio_walk_parents(zio_t *cio, zio_link_t **zl)
492{
493	list_t *pl = &cio->io_parent_list;
494
495	*zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
496	if (*zl == NULL)
497		return (NULL);
498
499	ASSERT((*zl)->zl_child == cio);
500	return ((*zl)->zl_parent);
501}
502
503zio_t *
504zio_walk_children(zio_t *pio, zio_link_t **zl)
505{
506	list_t *cl = &pio->io_child_list;
507
508	ASSERT(MUTEX_HELD(&pio->io_lock));
509
510	*zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
511	if (*zl == NULL)
512		return (NULL);
513
514	ASSERT((*zl)->zl_parent == pio);
515	return ((*zl)->zl_child);
516}
517
518zio_t *
519zio_unique_parent(zio_t *cio)
520{
521	zio_link_t *zl = NULL;
522	zio_t *pio = zio_walk_parents(cio, &zl);
523
524	VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
525	return (pio);
526}
527
528void
529zio_add_child(zio_t *pio, zio_t *cio)
530{
531	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
532
533	/*
534	 * Logical I/Os can have logical, gang, or vdev children.
535	 * Gang I/Os can have gang or vdev children.
536	 * Vdev I/Os can only have vdev children.
537	 * The following ASSERT captures all of these constraints.
538	 */
539	ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
540
541	zl->zl_parent = pio;
542	zl->zl_child = cio;
543
544	mutex_enter(&pio->io_lock);
545	mutex_enter(&cio->io_lock);
546
547	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
548
549	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
550		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
551
552	list_insert_head(&pio->io_child_list, zl);
553	list_insert_head(&cio->io_parent_list, zl);
554
555	pio->io_child_count++;
556	cio->io_parent_count++;
557
558	mutex_exit(&cio->io_lock);
559	mutex_exit(&pio->io_lock);
560}
561
562static void
563zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
564{
565	ASSERT(zl->zl_parent == pio);
566	ASSERT(zl->zl_child == cio);
567
568	mutex_enter(&pio->io_lock);
569	mutex_enter(&cio->io_lock);
570
571	list_remove(&pio->io_child_list, zl);
572	list_remove(&cio->io_parent_list, zl);
573
574	pio->io_child_count--;
575	cio->io_parent_count--;
576
577	mutex_exit(&cio->io_lock);
578	mutex_exit(&pio->io_lock);
579
580	kmem_cache_free(zio_link_cache, zl);
581}
582
583static boolean_t
584zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
585{
586	boolean_t waiting = B_FALSE;
587
588	mutex_enter(&zio->io_lock);
589	ASSERT(zio->io_stall == NULL);
590	for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
591		if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
592			continue;
593
594		uint64_t *countp = &zio->io_children[c][wait];
595		if (*countp != 0) {
596			zio->io_stage >>= 1;
597			ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
598			zio->io_stall = countp;
599			waiting = B_TRUE;
600			break;
601		}
602	}
603	mutex_exit(&zio->io_lock);
604	return (waiting);
605}
606
607static void
608zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
609{
610	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
611	int *errorp = &pio->io_child_error[zio->io_child_type];
612
613	mutex_enter(&pio->io_lock);
614	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
615		*errorp = zio_worst_error(*errorp, zio->io_error);
616	pio->io_reexecute |= zio->io_reexecute;
617	ASSERT3U(*countp, >, 0);
618
619	(*countp)--;
620
621	if (*countp == 0 && pio->io_stall == countp) {
622		zio_taskq_type_t type =
623		    pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
624		    ZIO_TASKQ_INTERRUPT;
625		pio->io_stall = NULL;
626		mutex_exit(&pio->io_lock);
627		/*
628		 * Dispatch the parent zio in its own taskq so that
629		 * the child can continue to make progress. This also
630		 * prevents overflowing the stack when we have deeply nested
631		 * parent-child relationships.
632		 */
633		zio_taskq_dispatch(pio, type, B_FALSE);
634	} else {
635		mutex_exit(&pio->io_lock);
636	}
637}
638
639static void
640zio_inherit_child_errors(zio_t *zio, enum zio_child c)
641{
642	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
643		zio->io_error = zio->io_child_error[c];
644}
645
646int
647zio_bookmark_compare(const void *x1, const void *x2)
648{
649	const zio_t *z1 = x1;
650	const zio_t *z2 = x2;
651
652	if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
653		return (-1);
654	if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
655		return (1);
656
657	if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
658		return (-1);
659	if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
660		return (1);
661
662	if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
663		return (-1);
664	if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
665		return (1);
666
667	if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
668		return (-1);
669	if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
670		return (1);
671
672	if (z1 < z2)
673		return (-1);
674	if (z1 > z2)
675		return (1);
676
677	return (0);
678}
679
680/*
681 * ==========================================================================
682 * Create the various types of I/O (read, write, free, etc)
683 * ==========================================================================
684 */
685static zio_t *
686zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
687    abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
688    void *private, zio_type_t type, zio_priority_t priority,
689    enum zio_flag flags, vdev_t *vd, uint64_t offset,
690    const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
691{
692	zio_t *zio;
693
694	ASSERT3U(psize, <=, SPA_MAXBLOCKSIZE);
695	ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
696	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
697
698	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
699	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
700	ASSERT(vd || stage == ZIO_STAGE_OPEN);
701
702	IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0);
703
704	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
705	bzero(zio, sizeof (zio_t));
706
707	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
708	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
709
710	list_create(&zio->io_parent_list, sizeof (zio_link_t),
711	    offsetof(zio_link_t, zl_parent_node));
712	list_create(&zio->io_child_list, sizeof (zio_link_t),
713	    offsetof(zio_link_t, zl_child_node));
714	metaslab_trace_init(&zio->io_alloc_list);
715
716	if (vd != NULL)
717		zio->io_child_type = ZIO_CHILD_VDEV;
718	else if (flags & ZIO_FLAG_GANG_CHILD)
719		zio->io_child_type = ZIO_CHILD_GANG;
720	else if (flags & ZIO_FLAG_DDT_CHILD)
721		zio->io_child_type = ZIO_CHILD_DDT;
722	else
723		zio->io_child_type = ZIO_CHILD_LOGICAL;
724
725	if (bp != NULL) {
726		zio->io_bp = (blkptr_t *)bp;
727		zio->io_bp_copy = *bp;
728		zio->io_bp_orig = *bp;
729		if (type != ZIO_TYPE_WRITE ||
730		    zio->io_child_type == ZIO_CHILD_DDT)
731			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
732		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
733			zio->io_logical = zio;
734		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
735			pipeline |= ZIO_GANG_STAGES;
736	}
737
738	zio->io_spa = spa;
739	zio->io_txg = txg;
740	zio->io_done = done;
741	zio->io_private = private;
742	zio->io_type = type;
743	zio->io_priority = priority;
744	zio->io_vd = vd;
745	zio->io_offset = offset;
746	zio->io_orig_abd = zio->io_abd = data;
747	zio->io_orig_size = zio->io_size = psize;
748	zio->io_lsize = lsize;
749	zio->io_orig_flags = zio->io_flags = flags;
750	zio->io_orig_stage = zio->io_stage = stage;
751	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
752	zio->io_pipeline_trace = ZIO_STAGE_OPEN;
753
754	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
755	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
756
757	if (zb != NULL)
758		zio->io_bookmark = *zb;
759
760	if (pio != NULL) {
761		if (zio->io_metaslab_class == NULL)
762			zio->io_metaslab_class = pio->io_metaslab_class;
763		if (zio->io_logical == NULL)
764			zio->io_logical = pio->io_logical;
765		if (zio->io_child_type == ZIO_CHILD_GANG)
766			zio->io_gang_leader = pio->io_gang_leader;
767		zio_add_child(pio, zio);
768	}
769
770	return (zio);
771}
772
773static void
774zio_destroy(zio_t *zio)
775{
776	metaslab_trace_fini(&zio->io_alloc_list);
777	list_destroy(&zio->io_parent_list);
778	list_destroy(&zio->io_child_list);
779	mutex_destroy(&zio->io_lock);
780	cv_destroy(&zio->io_cv);
781	kmem_cache_free(zio_cache, zio);
782}
783
784zio_t *
785zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
786    void *private, enum zio_flag flags)
787{
788	zio_t *zio;
789
790	zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
791	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
792	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
793
794	return (zio);
795}
796
797zio_t *
798zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
799{
800	return (zio_null(NULL, spa, NULL, done, private, flags));
801}
802
803void
804zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
805{
806	if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
807		zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
808		    bp, (longlong_t)BP_GET_TYPE(bp));
809	}
810	if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
811	    BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
812		zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
813		    bp, (longlong_t)BP_GET_CHECKSUM(bp));
814	}
815	if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
816	    BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
817		zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
818		    bp, (longlong_t)BP_GET_COMPRESS(bp));
819	}
820	if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
821		zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
822		    bp, (longlong_t)BP_GET_LSIZE(bp));
823	}
824	if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
825		zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
826		    bp, (longlong_t)BP_GET_PSIZE(bp));
827	}
828
829	if (BP_IS_EMBEDDED(bp)) {
830		if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
831			zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
832			    bp, (longlong_t)BPE_GET_ETYPE(bp));
833		}
834	}
835
836	/*
837	 * Do not verify individual DVAs if the config is not trusted. This
838	 * will be done once the zio is executed in vdev_mirror_map_alloc.
839	 */
840	if (!spa->spa_trust_config)
841		return;
842
843	/*
844	 * Pool-specific checks.
845	 *
846	 * Note: it would be nice to verify that the blk_birth and
847	 * BP_PHYSICAL_BIRTH() are not too large.  However, spa_freeze()
848	 * allows the birth time of log blocks (and dmu_sync()-ed blocks
849	 * that are in the log) to be arbitrarily large.
850	 */
851	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
852		uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
853		if (vdevid >= spa->spa_root_vdev->vdev_children) {
854			zfs_panic_recover("blkptr at %p DVA %u has invalid "
855			    "VDEV %llu",
856			    bp, i, (longlong_t)vdevid);
857			continue;
858		}
859		vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
860		if (vd == NULL) {
861			zfs_panic_recover("blkptr at %p DVA %u has invalid "
862			    "VDEV %llu",
863			    bp, i, (longlong_t)vdevid);
864			continue;
865		}
866		if (vd->vdev_ops == &vdev_hole_ops) {
867			zfs_panic_recover("blkptr at %p DVA %u has hole "
868			    "VDEV %llu",
869			    bp, i, (longlong_t)vdevid);
870			continue;
871		}
872		if (vd->vdev_ops == &vdev_missing_ops) {
873			/*
874			 * "missing" vdevs are valid during import, but we
875			 * don't have their detailed info (e.g. asize), so
876			 * we can't perform any more checks on them.
877			 */
878			continue;
879		}
880		uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
881		uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
882		if (BP_IS_GANG(bp))
883			asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
884		if (offset + asize > vd->vdev_asize) {
885			zfs_panic_recover("blkptr at %p DVA %u has invalid "
886			    "OFFSET %llu",
887			    bp, i, (longlong_t)offset);
888		}
889	}
890}
891
892boolean_t
893zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
894{
895	uint64_t vdevid = DVA_GET_VDEV(dva);
896
897	if (vdevid >= spa->spa_root_vdev->vdev_children)
898		return (B_FALSE);
899
900	vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
901	if (vd == NULL)
902		return (B_FALSE);
903
904	if (vd->vdev_ops == &vdev_hole_ops)
905		return (B_FALSE);
906
907	if (vd->vdev_ops == &vdev_missing_ops) {
908		return (B_FALSE);
909	}
910
911	uint64_t offset = DVA_GET_OFFSET(dva);
912	uint64_t asize = DVA_GET_ASIZE(dva);
913
914	if (BP_IS_GANG(bp))
915		asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
916	if (offset + asize > vd->vdev_asize)
917		return (B_FALSE);
918
919	return (B_TRUE);
920}
921
922zio_t *
923zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
924    abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
925    zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
926{
927	zio_t *zio;
928
929	zfs_blkptr_verify(spa, bp);
930
931	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
932	    data, size, size, done, private,
933	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
934	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
935	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
936
937	return (zio);
938}
939
940zio_t *
941zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
942    abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
943    zio_done_func_t *ready, zio_done_func_t *children_ready,
944    zio_done_func_t *physdone, zio_done_func_t *done,
945    void *private, zio_priority_t priority, enum zio_flag flags,
946    const zbookmark_phys_t *zb)
947{
948	zio_t *zio;
949
950	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
951	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
952	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
953	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
954	    DMU_OT_IS_VALID(zp->zp_type) &&
955	    zp->zp_level < 32 &&
956	    zp->zp_copies > 0 &&
957	    zp->zp_copies <= spa_max_replication(spa));
958
959	zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
960	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
961	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
962	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
963
964	zio->io_ready = ready;
965	zio->io_children_ready = children_ready;
966	zio->io_physdone = physdone;
967	zio->io_prop = *zp;
968
969	/*
970	 * Data can be NULL if we are going to call zio_write_override() to
971	 * provide the already-allocated BP.  But we may need the data to
972	 * verify a dedup hit (if requested).  In this case, don't try to
973	 * dedup (just take the already-allocated BP verbatim). Encrypted
974	 * dedup blocks need data as well so we also disable dedup in this
975	 * case.
976	 */
977	if (data == NULL &&
978	    (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) {
979		zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
980	}
981
982	return (zio);
983}
984
985zio_t *
986zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
987    uint64_t size, zio_done_func_t *done, void *private,
988    zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
989{
990	zio_t *zio;
991
992	zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
993	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
994	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
995
996	return (zio);
997}
998
999void
1000zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
1001{
1002	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1003	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1004	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1005	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
1006
1007	/*
1008	 * We must reset the io_prop to match the values that existed
1009	 * when the bp was first written by dmu_sync() keeping in mind
1010	 * that nopwrite and dedup are mutually exclusive.
1011	 */
1012	zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
1013	zio->io_prop.zp_nopwrite = nopwrite;
1014	zio->io_prop.zp_copies = copies;
1015	zio->io_bp_override = bp;
1016}
1017
1018void
1019zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
1020{
1021
1022	zfs_blkptr_verify(spa, bp);
1023
1024	/*
1025	 * The check for EMBEDDED is a performance optimization.  We
1026	 * process the free here (by ignoring it) rather than
1027	 * putting it on the list and then processing it in zio_free_sync().
1028	 */
1029	if (BP_IS_EMBEDDED(bp))
1030		return;
1031	metaslab_check_free(spa, bp);
1032
1033	/*
1034	 * Frees that are for the currently-syncing txg, are not going to be
1035	 * deferred, and which will not need to do a read (i.e. not GANG or
1036	 * DEDUP), can be processed immediately.  Otherwise, put them on the
1037	 * in-memory list for later processing.
1038	 */
1039	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
1040	    txg != spa->spa_syncing_txg ||
1041	    spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
1042		bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1043	} else {
1044		VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
1045	}
1046}
1047
1048zio_t *
1049zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1050    enum zio_flag flags)
1051{
1052	zio_t *zio;
1053	enum zio_stage stage = ZIO_FREE_PIPELINE;
1054
1055	ASSERT(!BP_IS_HOLE(bp));
1056	ASSERT(spa_syncing_txg(spa) == txg);
1057	ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
1058
1059	if (BP_IS_EMBEDDED(bp))
1060		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1061
1062	metaslab_check_free(spa, bp);
1063	arc_freed(spa, bp);
1064	dsl_scan_freed(spa, bp);
1065
1066	/*
1067	 * GANG and DEDUP blocks can induce a read (for the gang block header,
1068	 * or the DDT), so issue them asynchronously so that this thread is
1069	 * not tied up.
1070	 */
1071	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
1072		stage |= ZIO_STAGE_ISSUE_ASYNC;
1073
1074	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1075	    BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1076	    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
1077
1078	return (zio);
1079}
1080
1081zio_t *
1082zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1083    zio_done_func_t *done, void *private, enum zio_flag flags)
1084{
1085	zio_t *zio;
1086
1087	zfs_blkptr_verify(spa, bp);
1088
1089	if (BP_IS_EMBEDDED(bp))
1090		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1091
1092	/*
1093	 * A claim is an allocation of a specific block.  Claims are needed
1094	 * to support immediate writes in the intent log.  The issue is that
1095	 * immediate writes contain committed data, but in a txg that was
1096	 * *not* committed.  Upon opening the pool after an unclean shutdown,
1097	 * the intent log claims all blocks that contain immediate write data
1098	 * so that the SPA knows they're in use.
1099	 *
1100	 * All claims *must* be resolved in the first txg -- before the SPA
1101	 * starts allocating blocks -- so that nothing is allocated twice.
1102	 * If txg == 0 we just verify that the block is claimable.
1103	 */
1104	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1105	    spa_min_claim_txg(spa));
1106	ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1107	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
1108
1109	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1110	    BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1111	    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1112	ASSERT0(zio->io_queued_timestamp);
1113
1114	return (zio);
1115}
1116
1117zio_t *
1118zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
1119    zio_done_func_t *done, void *private, enum zio_flag flags)
1120{
1121	zio_t *zio;
1122	int c;
1123
1124	if (vd->vdev_children == 0) {
1125		zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1126		    ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1127		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1128
1129		zio->io_cmd = cmd;
1130	} else {
1131		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1132
1133		for (c = 0; c < vd->vdev_children; c++)
1134			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1135			    done, private, flags));
1136	}
1137
1138	return (zio);
1139}
1140
1141zio_t *
1142zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1143    abd_t *data, int checksum, zio_done_func_t *done, void *private,
1144    zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1145{
1146	zio_t *zio;
1147
1148	ASSERT(vd->vdev_children == 0);
1149	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1150	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1151	ASSERT3U(offset + size, <=, vd->vdev_psize);
1152
1153	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1154	    private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1155	    offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1156
1157	zio->io_prop.zp_checksum = checksum;
1158
1159	return (zio);
1160}
1161
1162zio_t *
1163zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1164    abd_t *data, int checksum, zio_done_func_t *done, void *private,
1165    zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1166{
1167	zio_t *zio;
1168
1169	ASSERT(vd->vdev_children == 0);
1170	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1171	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1172	ASSERT3U(offset + size, <=, vd->vdev_psize);
1173
1174	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1175	    private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1176	    offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1177
1178	zio->io_prop.zp_checksum = checksum;
1179
1180	if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1181		/*
1182		 * zec checksums are necessarily destructive -- they modify
1183		 * the end of the write buffer to hold the verifier/checksum.
1184		 * Therefore, we must make a local copy in case the data is
1185		 * being written to multiple places in parallel.
1186		 */
1187		abd_t *wbuf = abd_alloc_sametype(data, size);
1188		abd_copy(wbuf, data, size);
1189
1190		zio_push_transform(zio, wbuf, size, size, NULL);
1191	}
1192
1193	return (zio);
1194}
1195
1196/*
1197 * Create a child I/O to do some work for us.
1198 */
1199zio_t *
1200zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1201    abd_t *data, uint64_t size, int type, zio_priority_t priority,
1202    enum zio_flag flags, zio_done_func_t *done, void *private)
1203{
1204	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1205	zio_t *zio;
1206
1207	/*
1208	 * vdev child I/Os do not propagate their error to the parent.
1209	 * Therefore, for correct operation the caller *must* check for
1210	 * and handle the error in the child i/o's done callback.
1211	 * The only exceptions are i/os that we don't care about
1212	 * (OPTIONAL or REPAIR).
1213	 */
1214	ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1215	    done != NULL);
1216
1217	if (type == ZIO_TYPE_READ && bp != NULL) {
1218		/*
1219		 * If we have the bp, then the child should perform the
1220		 * checksum and the parent need not.  This pushes error
1221		 * detection as close to the leaves as possible and
1222		 * eliminates redundant checksums in the interior nodes.
1223		 */
1224		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1225		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1226	}
1227
1228	if (vd->vdev_ops->vdev_op_leaf) {
1229		ASSERT0(vd->vdev_children);
1230		offset += VDEV_LABEL_START_SIZE;
1231	}
1232
1233	flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1234
1235	/*
1236	 * If we've decided to do a repair, the write is not speculative --
1237	 * even if the original read was.
1238	 */
1239	if (flags & ZIO_FLAG_IO_REPAIR)
1240		flags &= ~ZIO_FLAG_SPECULATIVE;
1241
1242	/*
1243	 * If we're creating a child I/O that is not associated with a
1244	 * top-level vdev, then the child zio is not an allocating I/O.
1245	 * If this is a retried I/O then we ignore it since we will
1246	 * have already processed the original allocating I/O.
1247	 */
1248	if (flags & ZIO_FLAG_IO_ALLOCATING &&
1249	    (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1250		ASSERT(pio->io_metaslab_class != NULL);
1251		ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled);
1252		ASSERT(type == ZIO_TYPE_WRITE);
1253		ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1254		ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1255		ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1256		    pio->io_child_type == ZIO_CHILD_GANG);
1257
1258		flags &= ~ZIO_FLAG_IO_ALLOCATING;
1259	}
1260
1261	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1262	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1263	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1264	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1265
1266	zio->io_physdone = pio->io_physdone;
1267	if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1268		zio->io_logical->io_phys_children++;
1269
1270	return (zio);
1271}
1272
1273zio_t *
1274zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1275    zio_type_t type, zio_priority_t priority, enum zio_flag flags,
1276    zio_done_func_t *done, void *private)
1277{
1278	zio_t *zio;
1279
1280	ASSERT(vd->vdev_ops->vdev_op_leaf);
1281
1282	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1283	    data, size, size, done, private, type, priority,
1284	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1285	    vd, offset, NULL,
1286	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1287
1288	return (zio);
1289}
1290
1291void
1292zio_flush(zio_t *zio, vdev_t *vd)
1293{
1294	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1295	    NULL, NULL,
1296	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1297}
1298
1299void
1300zio_shrink(zio_t *zio, uint64_t size)
1301{
1302	ASSERT3P(zio->io_executor, ==, NULL);
1303	ASSERT3P(zio->io_orig_size, ==, zio->io_size);
1304	ASSERT3U(size, <=, zio->io_size);
1305
1306	/*
1307	 * We don't shrink for raidz because of problems with the
1308	 * reconstruction when reading back less than the block size.
1309	 * Note, BP_IS_RAIDZ() assumes no compression.
1310	 */
1311	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1312	if (!BP_IS_RAIDZ(zio->io_bp)) {
1313		/* we are not doing a raw write */
1314		ASSERT3U(zio->io_size, ==, zio->io_lsize);
1315		zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1316	}
1317}
1318
1319/*
1320 * ==========================================================================
1321 * Prepare to read and write logical blocks
1322 * ==========================================================================
1323 */
1324
1325static int
1326zio_read_bp_init(zio_t *zio)
1327{
1328	blkptr_t *bp = zio->io_bp;
1329	uint64_t psize =
1330	    BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1331
1332	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1333
1334	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1335	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
1336	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1337		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1338		    psize, psize, zio_decompress);
1339	}
1340
1341	if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) ||
1342	    BP_HAS_INDIRECT_MAC_CKSUM(bp)) &&
1343	    zio->io_child_type == ZIO_CHILD_LOGICAL) {
1344		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1345		    psize, psize, zio_decrypt);
1346	}
1347
1348	if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1349		int psize = BPE_GET_PSIZE(bp);
1350		void *data = abd_borrow_buf(zio->io_abd, psize);
1351
1352		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1353		decode_embedded_bp_compressed(bp, data);
1354		abd_return_buf_copy(zio->io_abd, data, psize);
1355	} else {
1356		ASSERT(!BP_IS_EMBEDDED(bp));
1357		ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1358	}
1359
1360	if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1361		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1362
1363	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1364		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1365
1366	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1367		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1368
1369	return (ZIO_PIPELINE_CONTINUE);
1370}
1371
1372static int
1373zio_write_bp_init(zio_t *zio)
1374{
1375	if (!IO_IS_ALLOCATING(zio))
1376		return (ZIO_PIPELINE_CONTINUE);
1377
1378	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1379
1380	if (zio->io_bp_override) {
1381		blkptr_t *bp = zio->io_bp;
1382		zio_prop_t *zp = &zio->io_prop;
1383
1384		ASSERT(bp->blk_birth != zio->io_txg);
1385		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1386
1387		*bp = *zio->io_bp_override;
1388		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1389
1390		if (BP_IS_EMBEDDED(bp))
1391			return (ZIO_PIPELINE_CONTINUE);
1392
1393		/*
1394		 * If we've been overridden and nopwrite is set then
1395		 * set the flag accordingly to indicate that a nopwrite
1396		 * has already occurred.
1397		 */
1398		if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1399			ASSERT(!zp->zp_dedup);
1400			ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1401			zio->io_flags |= ZIO_FLAG_NOPWRITE;
1402			return (ZIO_PIPELINE_CONTINUE);
1403		}
1404
1405		ASSERT(!zp->zp_nopwrite);
1406
1407		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1408			return (ZIO_PIPELINE_CONTINUE);
1409
1410		ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1411		    ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1412
1413		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum &&
1414		    !zp->zp_encrypt) {
1415			BP_SET_DEDUP(bp, 1);
1416			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1417			return (ZIO_PIPELINE_CONTINUE);
1418		}
1419
1420		/*
1421		 * We were unable to handle this as an override bp, treat
1422		 * it as a regular write I/O.
1423		 */
1424		zio->io_bp_override = NULL;
1425		*bp = zio->io_bp_orig;
1426		zio->io_pipeline = zio->io_orig_pipeline;
1427	}
1428
1429	return (ZIO_PIPELINE_CONTINUE);
1430}
1431
1432static int
1433zio_write_compress(zio_t *zio)
1434{
1435	spa_t *spa = zio->io_spa;
1436	zio_prop_t *zp = &zio->io_prop;
1437	enum zio_compress compress = zp->zp_compress;
1438	blkptr_t *bp = zio->io_bp;
1439	uint64_t lsize = zio->io_lsize;
1440	uint64_t psize = zio->io_size;
1441	int pass = 1;
1442
1443	/*
1444	 * If our children haven't all reached the ready stage,
1445	 * wait for them and then repeat this pipeline stage.
1446	 */
1447	if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1448	    ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1449		return (ZIO_PIPELINE_STOP);
1450	}
1451
1452	if (!IO_IS_ALLOCATING(zio))
1453		return (ZIO_PIPELINE_CONTINUE);
1454
1455	if (zio->io_children_ready != NULL) {
1456		/*
1457		 * Now that all our children are ready, run the callback
1458		 * associated with this zio in case it wants to modify the
1459		 * data to be written.
1460		 */
1461		ASSERT3U(zp->zp_level, >, 0);
1462		zio->io_children_ready(zio);
1463	}
1464
1465	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1466	ASSERT(zio->io_bp_override == NULL);
1467
1468	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1469		/*
1470		 * We're rewriting an existing block, which means we're
1471		 * working on behalf of spa_sync().  For spa_sync() to
1472		 * converge, it must eventually be the case that we don't
1473		 * have to allocate new blocks.  But compression changes
1474		 * the blocksize, which forces a reallocate, and makes
1475		 * convergence take longer.  Therefore, after the first
1476		 * few passes, stop compressing to ensure convergence.
1477		 */
1478		pass = spa_sync_pass(spa);
1479
1480		ASSERT(zio->io_txg == spa_syncing_txg(spa));
1481		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1482		ASSERT(!BP_GET_DEDUP(bp));
1483
1484		if (pass >= zfs_sync_pass_dont_compress)
1485			compress = ZIO_COMPRESS_OFF;
1486
1487		/* Make sure someone doesn't change their mind on overwrites */
1488		ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1489		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1490	}
1491
1492	/* If it's a compressed write that is not raw, compress the buffer. */
1493	if (compress != ZIO_COMPRESS_OFF &&
1494	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1495		void *cbuf = zio_buf_alloc(lsize);
1496		psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1497		if (psize == 0 || psize == lsize) {
1498			compress = ZIO_COMPRESS_OFF;
1499			zio_buf_free(cbuf, lsize);
1500		} else if (!zp->zp_dedup && !zp->zp_encrypt &&
1501		    psize <= BPE_PAYLOAD_SIZE &&
1502		    zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1503		    spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1504			encode_embedded_bp_compressed(bp,
1505			    cbuf, compress, lsize, psize);
1506			BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1507			BP_SET_TYPE(bp, zio->io_prop.zp_type);
1508			BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1509			zio_buf_free(cbuf, lsize);
1510			bp->blk_birth = zio->io_txg;
1511			zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1512			ASSERT(spa_feature_is_active(spa,
1513			    SPA_FEATURE_EMBEDDED_DATA));
1514			return (ZIO_PIPELINE_CONTINUE);
1515		} else {
1516			/*
1517			 * Round up compressed size up to the ashift
1518			 * of the smallest-ashift device, and zero the tail.
1519			 * This ensures that the compressed size of the BP
1520			 * (and thus compressratio property) are correct,
1521			 * in that we charge for the padding used to fill out
1522			 * the last sector.
1523			 */
1524			ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1525			size_t rounded = (size_t)P2ROUNDUP(psize,
1526			    1ULL << spa->spa_min_ashift);
1527			if (rounded >= lsize) {
1528				compress = ZIO_COMPRESS_OFF;
1529				zio_buf_free(cbuf, lsize);
1530				psize = lsize;
1531			} else {
1532				abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1533				abd_take_ownership_of_buf(cdata, B_TRUE);
1534				abd_zero_off(cdata, psize, rounded - psize);
1535				psize = rounded;
1536				zio_push_transform(zio, cdata,
1537				    psize, lsize, NULL);
1538			}
1539		}
1540
1541		/*
1542		 * We were unable to handle this as an override bp, treat
1543		 * it as a regular write I/O.
1544		 */
1545		zio->io_bp_override = NULL;
1546		*bp = zio->io_bp_orig;
1547		zio->io_pipeline = zio->io_orig_pipeline;
1548
1549	} else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 &&
1550	    zp->zp_type == DMU_OT_DNODE) {
1551		/*
1552		 * The DMU actually relies on the zio layer's compression
1553		 * to free metadnode blocks that have had all contained
1554		 * dnodes freed. As a result, even when doing a raw
1555		 * receive, we must check whether the block can be compressed
1556		 * to a hole.
1557		 */
1558		psize = zio_compress_data(ZIO_COMPRESS_EMPTY,
1559		    zio->io_abd, NULL, lsize);
1560		if (psize == 0)
1561			compress = ZIO_COMPRESS_OFF;
1562	} else {
1563		ASSERT3U(psize, !=, 0);
1564	}
1565
1566	/*
1567	 * The final pass of spa_sync() must be all rewrites, but the first
1568	 * few passes offer a trade-off: allocating blocks defers convergence,
1569	 * but newly allocated blocks are sequential, so they can be written
1570	 * to disk faster.  Therefore, we allow the first few passes of
1571	 * spa_sync() to allocate new blocks, but force rewrites after that.
1572	 * There should only be a handful of blocks after pass 1 in any case.
1573	 */
1574	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1575	    BP_GET_PSIZE(bp) == psize &&
1576	    pass >= zfs_sync_pass_rewrite) {
1577		VERIFY3U(psize, !=, 0);
1578		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1579		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1580		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1581	} else {
1582		BP_ZERO(bp);
1583		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1584	}
1585
1586	if (psize == 0) {
1587		if (zio->io_bp_orig.blk_birth != 0 &&
1588		    spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1589			BP_SET_LSIZE(bp, lsize);
1590			BP_SET_TYPE(bp, zp->zp_type);
1591			BP_SET_LEVEL(bp, zp->zp_level);
1592			BP_SET_BIRTH(bp, zio->io_txg, 0);
1593		}
1594		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1595	} else {
1596		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1597		BP_SET_LSIZE(bp, lsize);
1598		BP_SET_TYPE(bp, zp->zp_type);
1599		BP_SET_LEVEL(bp, zp->zp_level);
1600		BP_SET_PSIZE(bp, psize);
1601		BP_SET_COMPRESS(bp, compress);
1602		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1603		BP_SET_DEDUP(bp, zp->zp_dedup);
1604		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1605		if (zp->zp_dedup) {
1606			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1607			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1608			ASSERT(!zp->zp_encrypt ||
1609			    DMU_OT_IS_ENCRYPTED(zp->zp_type));
1610			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1611		}
1612		if (zp->zp_nopwrite) {
1613			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1614			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1615			zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1616		}
1617	}
1618	return (ZIO_PIPELINE_CONTINUE);
1619}
1620
1621static int
1622zio_free_bp_init(zio_t *zio)
1623{
1624	blkptr_t *bp = zio->io_bp;
1625
1626	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1627		if (BP_GET_DEDUP(bp))
1628			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1629	}
1630
1631	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1632
1633	return (ZIO_PIPELINE_CONTINUE);
1634}
1635
1636/*
1637 * ==========================================================================
1638 * Execute the I/O pipeline
1639 * ==========================================================================
1640 */
1641
1642static void
1643zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1644{
1645	spa_t *spa = zio->io_spa;
1646	zio_type_t t = zio->io_type;
1647	int flags = (cutinline ? TQ_FRONT : 0);
1648
1649	/*
1650	 * If we're a config writer or a probe, the normal issue and
1651	 * interrupt threads may all be blocked waiting for the config lock.
1652	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1653	 */
1654	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1655		t = ZIO_TYPE_NULL;
1656
1657	/*
1658	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1659	 */
1660	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1661		t = ZIO_TYPE_NULL;
1662
1663	/*
1664	 * If this is a high priority I/O, then use the high priority taskq if
1665	 * available.
1666	 */
1667	if ((zio->io_priority == ZIO_PRIORITY_NOW ||
1668	    zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) &&
1669	    spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1670		q++;
1671
1672	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1673
1674	/*
1675	 * NB: We are assuming that the zio can only be dispatched
1676	 * to a single taskq at a time.  It would be a grievous error
1677	 * to dispatch the zio to another taskq at the same time.
1678	 */
1679	ASSERT(zio->io_tqent.tqent_next == NULL);
1680	spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1681	    flags, &zio->io_tqent);
1682}
1683
1684static boolean_t
1685zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1686{
1687	kthread_t *executor = zio->io_executor;
1688	spa_t *spa = zio->io_spa;
1689
1690	for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1691		spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1692		uint_t i;
1693		for (i = 0; i < tqs->stqs_count; i++) {
1694			if (taskq_member(tqs->stqs_taskq[i], executor))
1695				return (B_TRUE);
1696		}
1697	}
1698
1699	return (B_FALSE);
1700}
1701
1702static int
1703zio_issue_async(zio_t *zio)
1704{
1705	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1706
1707	return (ZIO_PIPELINE_STOP);
1708}
1709
1710void
1711zio_interrupt(zio_t *zio)
1712{
1713	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1714}
1715
1716void
1717zio_delay_interrupt(zio_t *zio)
1718{
1719	/*
1720	 * The timeout_generic() function isn't defined in userspace, so
1721	 * rather than trying to implement the function, the zio delay
1722	 * functionality has been disabled for userspace builds.
1723	 */
1724
1725#ifdef _KERNEL
1726	/*
1727	 * If io_target_timestamp is zero, then no delay has been registered
1728	 * for this IO, thus jump to the end of this function and "skip" the
1729	 * delay; issuing it directly to the zio layer.
1730	 */
1731	if (zio->io_target_timestamp != 0) {
1732		hrtime_t now = gethrtime();
1733
1734		if (now >= zio->io_target_timestamp) {
1735			/*
1736			 * This IO has already taken longer than the target
1737			 * delay to complete, so we don't want to delay it
1738			 * any longer; we "miss" the delay and issue it
1739			 * directly to the zio layer. This is likely due to
1740			 * the target latency being set to a value less than
1741			 * the underlying hardware can satisfy (e.g. delay
1742			 * set to 1ms, but the disks take 10ms to complete an
1743			 * IO request).
1744			 */
1745
1746			DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1747			    hrtime_t, now);
1748
1749			zio_interrupt(zio);
1750		} else {
1751			hrtime_t diff = zio->io_target_timestamp - now;
1752
1753			DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1754			    hrtime_t, now, hrtime_t, diff);
1755
1756			(void) timeout_generic(CALLOUT_NORMAL,
1757			    (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1758		}
1759
1760		return;
1761	}
1762#endif
1763
1764	DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1765	zio_interrupt(zio);
1766}
1767
1768/*
1769 * Execute the I/O pipeline until one of the following occurs:
1770 *
1771 *	(1) the I/O completes
1772 *	(2) the pipeline stalls waiting for dependent child I/Os
1773 *	(3) the I/O issues, so we're waiting for an I/O completion interrupt
1774 *	(4) the I/O is delegated by vdev-level caching or aggregation
1775 *	(5) the I/O is deferred due to vdev-level queueing
1776 *	(6) the I/O is handed off to another thread.
1777 *
1778 * In all cases, the pipeline stops whenever there's no CPU work; it never
1779 * burns a thread in cv_wait().
1780 *
1781 * There's no locking on io_stage because there's no legitimate way
1782 * for multiple threads to be attempting to process the same I/O.
1783 */
1784static zio_pipe_stage_t *zio_pipeline[];
1785
1786void
1787zio_execute(zio_t *zio)
1788{
1789	zio->io_executor = curthread;
1790
1791	ASSERT3U(zio->io_queued_timestamp, >, 0);
1792
1793	while (zio->io_stage < ZIO_STAGE_DONE) {
1794		enum zio_stage pipeline = zio->io_pipeline;
1795		enum zio_stage stage = zio->io_stage;
1796		int rv;
1797
1798		ASSERT(!MUTEX_HELD(&zio->io_lock));
1799		ASSERT(ISP2(stage));
1800		ASSERT(zio->io_stall == NULL);
1801
1802		do {
1803			stage <<= 1;
1804		} while ((stage & pipeline) == 0);
1805
1806		ASSERT(stage <= ZIO_STAGE_DONE);
1807
1808		/*
1809		 * If we are in interrupt context and this pipeline stage
1810		 * will grab a config lock that is held across I/O,
1811		 * or may wait for an I/O that needs an interrupt thread
1812		 * to complete, issue async to avoid deadlock.
1813		 *
1814		 * For VDEV_IO_START, we cut in line so that the io will
1815		 * be sent to disk promptly.
1816		 */
1817		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1818		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1819			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1820			    zio_requeue_io_start_cut_in_line : B_FALSE;
1821			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1822			return;
1823		}
1824
1825		zio->io_stage = stage;
1826		zio->io_pipeline_trace |= zio->io_stage;
1827		rv = zio_pipeline[highbit64(stage) - 1](zio);
1828
1829		if (rv == ZIO_PIPELINE_STOP)
1830			return;
1831
1832		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1833	}
1834}
1835
1836/*
1837 * ==========================================================================
1838 * Initiate I/O, either sync or async
1839 * ==========================================================================
1840 */
1841int
1842zio_wait(zio_t *zio)
1843{
1844	int error;
1845
1846	ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1847	ASSERT3P(zio->io_executor, ==, NULL);
1848
1849	zio->io_waiter = curthread;
1850	ASSERT0(zio->io_queued_timestamp);
1851	zio->io_queued_timestamp = gethrtime();
1852
1853	zio_execute(zio);
1854
1855	mutex_enter(&zio->io_lock);
1856	while (zio->io_executor != NULL)
1857		cv_wait(&zio->io_cv, &zio->io_lock);
1858	mutex_exit(&zio->io_lock);
1859
1860	error = zio->io_error;
1861	zio_destroy(zio);
1862
1863	return (error);
1864}
1865
1866void
1867zio_nowait(zio_t *zio)
1868{
1869	ASSERT3P(zio->io_executor, ==, NULL);
1870
1871	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1872	    zio_unique_parent(zio) == NULL) {
1873		/*
1874		 * This is a logical async I/O with no parent to wait for it.
1875		 * We add it to the spa_async_root_zio "Godfather" I/O which
1876		 * will ensure they complete prior to unloading the pool.
1877		 */
1878		spa_t *spa = zio->io_spa;
1879
1880		zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1881	}
1882
1883	ASSERT0(zio->io_queued_timestamp);
1884	zio->io_queued_timestamp = gethrtime();
1885	zio_execute(zio);
1886}
1887
1888/*
1889 * ==========================================================================
1890 * Reexecute, cancel, or suspend/resume failed I/O
1891 * ==========================================================================
1892 */
1893
1894static void
1895zio_reexecute(zio_t *pio)
1896{
1897	zio_t *cio, *cio_next;
1898
1899	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1900	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1901	ASSERT(pio->io_gang_leader == NULL);
1902	ASSERT(pio->io_gang_tree == NULL);
1903
1904	pio->io_flags = pio->io_orig_flags;
1905	pio->io_stage = pio->io_orig_stage;
1906	pio->io_pipeline = pio->io_orig_pipeline;
1907	pio->io_reexecute = 0;
1908	pio->io_flags |= ZIO_FLAG_REEXECUTED;
1909	pio->io_pipeline_trace = 0;
1910	pio->io_error = 0;
1911	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1912		pio->io_state[w] = 0;
1913	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1914		pio->io_child_error[c] = 0;
1915
1916	if (IO_IS_ALLOCATING(pio))
1917		BP_ZERO(pio->io_bp);
1918
1919	/*
1920	 * As we reexecute pio's children, new children could be created.
1921	 * New children go to the head of pio's io_child_list, however,
1922	 * so we will (correctly) not reexecute them.  The key is that
1923	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1924	 * cannot be affected by any side effects of reexecuting 'cio'.
1925	 */
1926	zio_link_t *zl = NULL;
1927	mutex_enter(&pio->io_lock);
1928	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1929		cio_next = zio_walk_children(pio, &zl);
1930		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1931			pio->io_children[cio->io_child_type][w]++;
1932		mutex_exit(&pio->io_lock);
1933		zio_reexecute(cio);
1934		mutex_enter(&pio->io_lock);
1935	}
1936	mutex_exit(&pio->io_lock);
1937
1938	/*
1939	 * Now that all children have been reexecuted, execute the parent.
1940	 * We don't reexecute "The Godfather" I/O here as it's the
1941	 * responsibility of the caller to wait on it.
1942	 */
1943	if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1944		pio->io_queued_timestamp = gethrtime();
1945		zio_execute(pio);
1946	}
1947}
1948
1949void
1950zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
1951{
1952	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1953		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1954		    "failure and the failure mode property for this pool "
1955		    "is set to panic.", spa_name(spa));
1956
1957	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
1958	    NULL, NULL, 0, 0);
1959
1960	mutex_enter(&spa->spa_suspend_lock);
1961
1962	if (spa->spa_suspend_zio_root == NULL)
1963		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1964		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1965		    ZIO_FLAG_GODFATHER);
1966
1967	spa->spa_suspended = reason;
1968
1969	if (zio != NULL) {
1970		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1971		ASSERT(zio != spa->spa_suspend_zio_root);
1972		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1973		ASSERT(zio_unique_parent(zio) == NULL);
1974		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1975		zio_add_child(spa->spa_suspend_zio_root, zio);
1976	}
1977
1978	mutex_exit(&spa->spa_suspend_lock);
1979}
1980
1981int
1982zio_resume(spa_t *spa)
1983{
1984	zio_t *pio;
1985
1986	/*
1987	 * Reexecute all previously suspended i/o.
1988	 */
1989	mutex_enter(&spa->spa_suspend_lock);
1990	spa->spa_suspended = ZIO_SUSPEND_NONE;
1991	cv_broadcast(&spa->spa_suspend_cv);
1992	pio = spa->spa_suspend_zio_root;
1993	spa->spa_suspend_zio_root = NULL;
1994	mutex_exit(&spa->spa_suspend_lock);
1995
1996	if (pio == NULL)
1997		return (0);
1998
1999	zio_reexecute(pio);
2000	return (zio_wait(pio));
2001}
2002
2003void
2004zio_resume_wait(spa_t *spa)
2005{
2006	mutex_enter(&spa->spa_suspend_lock);
2007	while (spa_suspended(spa))
2008		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
2009	mutex_exit(&spa->spa_suspend_lock);
2010}
2011
2012/*
2013 * ==========================================================================
2014 * Gang blocks.
2015 *
2016 * A gang block is a collection of small blocks that looks to the DMU
2017 * like one large block.  When zio_dva_allocate() cannot find a block
2018 * of the requested size, due to either severe fragmentation or the pool
2019 * being nearly full, it calls zio_write_gang_block() to construct the
2020 * block from smaller fragments.
2021 *
2022 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2023 * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
2024 * an indirect block: it's an array of block pointers.  It consumes
2025 * only one sector and hence is allocatable regardless of fragmentation.
2026 * The gang header's bps point to its gang members, which hold the data.
2027 *
2028 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2029 * as the verifier to ensure uniqueness of the SHA256 checksum.
2030 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2031 * not the gang header.  This ensures that data block signatures (needed for
2032 * deduplication) are independent of how the block is physically stored.
2033 *
2034 * Gang blocks can be nested: a gang member may itself be a gang block.
2035 * Thus every gang block is a tree in which root and all interior nodes are
2036 * gang headers, and the leaves are normal blocks that contain user data.
2037 * The root of the gang tree is called the gang leader.
2038 *
2039 * To perform any operation (read, rewrite, free, claim) on a gang block,
2040 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2041 * in the io_gang_tree field of the original logical i/o by recursively
2042 * reading the gang leader and all gang headers below it.  This yields
2043 * an in-core tree containing the contents of every gang header and the
2044 * bps for every constituent of the gang block.
2045 *
2046 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2047 * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
2048 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2049 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2050 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2051 * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
2052 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2053 * of the gang header plus zio_checksum_compute() of the data to update the
2054 * gang header's blk_cksum as described above.
2055 *
2056 * The two-phase assemble/issue model solves the problem of partial failure --
2057 * what if you'd freed part of a gang block but then couldn't read the
2058 * gang header for another part?  Assembling the entire gang tree first
2059 * ensures that all the necessary gang header I/O has succeeded before
2060 * starting the actual work of free, claim, or write.  Once the gang tree
2061 * is assembled, free and claim are in-memory operations that cannot fail.
2062 *
2063 * In the event that a gang write fails, zio_dva_unallocate() walks the
2064 * gang tree to immediately free (i.e. insert back into the space map)
2065 * everything we've allocated.  This ensures that we don't get ENOSPC
2066 * errors during repeated suspend/resume cycles due to a flaky device.
2067 *
2068 * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
2069 * the gang tree, we won't modify the block, so we can safely defer the free
2070 * (knowing that the block is still intact).  If we *can* assemble the gang
2071 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2072 * each constituent bp and we can allocate a new block on the next sync pass.
2073 *
2074 * In all cases, the gang tree allows complete recovery from partial failure.
2075 * ==========================================================================
2076 */
2077
2078static void
2079zio_gang_issue_func_done(zio_t *zio)
2080{
2081	abd_put(zio->io_abd);
2082}
2083
2084static zio_t *
2085zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2086    uint64_t offset)
2087{
2088	if (gn != NULL)
2089		return (pio);
2090
2091	return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2092	    BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2093	    NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2094	    &pio->io_bookmark));
2095}
2096
2097static zio_t *
2098zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2099    uint64_t offset)
2100{
2101	zio_t *zio;
2102
2103	if (gn != NULL) {
2104		abd_t *gbh_abd =
2105		    abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2106		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2107		    gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2108		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2109		    &pio->io_bookmark);
2110		/*
2111		 * As we rewrite each gang header, the pipeline will compute
2112		 * a new gang block header checksum for it; but no one will
2113		 * compute a new data checksum, so we do that here.  The one
2114		 * exception is the gang leader: the pipeline already computed
2115		 * its data checksum because that stage precedes gang assembly.
2116		 * (Presently, nothing actually uses interior data checksums;
2117		 * this is just good hygiene.)
2118		 */
2119		if (gn != pio->io_gang_leader->io_gang_tree) {
2120			abd_t *buf = abd_get_offset(data, offset);
2121
2122			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2123			    buf, BP_GET_PSIZE(bp));
2124
2125			abd_put(buf);
2126		}
2127		/*
2128		 * If we are here to damage data for testing purposes,
2129		 * leave the GBH alone so that we can detect the damage.
2130		 */
2131		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2132			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2133	} else {
2134		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2135		    abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2136		    zio_gang_issue_func_done, NULL, pio->io_priority,
2137		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2138	}
2139
2140	return (zio);
2141}
2142
2143/* ARGSUSED */
2144static zio_t *
2145zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2146    uint64_t offset)
2147{
2148	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2149	    ZIO_GANG_CHILD_FLAGS(pio)));
2150}
2151
2152/* ARGSUSED */
2153static zio_t *
2154zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2155    uint64_t offset)
2156{
2157	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2158	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2159}
2160
2161static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2162	NULL,
2163	zio_read_gang,
2164	zio_rewrite_gang,
2165	zio_free_gang,
2166	zio_claim_gang,
2167	NULL
2168};
2169
2170static void zio_gang_tree_assemble_done(zio_t *zio);
2171
2172static zio_gang_node_t *
2173zio_gang_node_alloc(zio_gang_node_t **gnpp)
2174{
2175	zio_gang_node_t *gn;
2176
2177	ASSERT(*gnpp == NULL);
2178
2179	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2180	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2181	*gnpp = gn;
2182
2183	return (gn);
2184}
2185
2186static void
2187zio_gang_node_free(zio_gang_node_t **gnpp)
2188{
2189	zio_gang_node_t *gn = *gnpp;
2190
2191	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2192		ASSERT(gn->gn_child[g] == NULL);
2193
2194	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2195	kmem_free(gn, sizeof (*gn));
2196	*gnpp = NULL;
2197}
2198
2199static void
2200zio_gang_tree_free(zio_gang_node_t **gnpp)
2201{
2202	zio_gang_node_t *gn = *gnpp;
2203
2204	if (gn == NULL)
2205		return;
2206
2207	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2208		zio_gang_tree_free(&gn->gn_child[g]);
2209
2210	zio_gang_node_free(gnpp);
2211}
2212
2213static void
2214zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2215{
2216	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2217	abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2218
2219	ASSERT(gio->io_gang_leader == gio);
2220	ASSERT(BP_IS_GANG(bp));
2221
2222	zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2223	    zio_gang_tree_assemble_done, gn, gio->io_priority,
2224	    ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2225}
2226
2227static void
2228zio_gang_tree_assemble_done(zio_t *zio)
2229{
2230	zio_t *gio = zio->io_gang_leader;
2231	zio_gang_node_t *gn = zio->io_private;
2232	blkptr_t *bp = zio->io_bp;
2233
2234	ASSERT(gio == zio_unique_parent(zio));
2235	ASSERT(zio->io_child_count == 0);
2236
2237	if (zio->io_error)
2238		return;
2239
2240	/* this ABD was created from a linear buf in zio_gang_tree_assemble */
2241	if (BP_SHOULD_BYTESWAP(bp))
2242		byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2243
2244	ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2245	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2246	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2247
2248	abd_put(zio->io_abd);
2249
2250	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2251		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2252		if (!BP_IS_GANG(gbp))
2253			continue;
2254		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2255	}
2256}
2257
2258static void
2259zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2260    uint64_t offset)
2261{
2262	zio_t *gio = pio->io_gang_leader;
2263	zio_t *zio;
2264
2265	ASSERT(BP_IS_GANG(bp) == !!gn);
2266	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2267	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2268
2269	/*
2270	 * If you're a gang header, your data is in gn->gn_gbh.
2271	 * If you're a gang member, your data is in 'data' and gn == NULL.
2272	 */
2273	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2274
2275	if (gn != NULL) {
2276		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2277
2278		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2279			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2280			if (BP_IS_HOLE(gbp))
2281				continue;
2282			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2283			    offset);
2284			offset += BP_GET_PSIZE(gbp);
2285		}
2286	}
2287
2288	if (gn == gio->io_gang_tree)
2289		ASSERT3U(gio->io_size, ==, offset);
2290
2291	if (zio != pio)
2292		zio_nowait(zio);
2293}
2294
2295static int
2296zio_gang_assemble(zio_t *zio)
2297{
2298	blkptr_t *bp = zio->io_bp;
2299
2300	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2301	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2302
2303	zio->io_gang_leader = zio;
2304
2305	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2306
2307	return (ZIO_PIPELINE_CONTINUE);
2308}
2309
2310static int
2311zio_gang_issue(zio_t *zio)
2312{
2313	blkptr_t *bp = zio->io_bp;
2314
2315	if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2316		return (ZIO_PIPELINE_STOP);
2317	}
2318
2319	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2320	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2321
2322	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2323		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2324		    0);
2325	else
2326		zio_gang_tree_free(&zio->io_gang_tree);
2327
2328	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2329
2330	return (ZIO_PIPELINE_CONTINUE);
2331}
2332
2333static void
2334zio_write_gang_member_ready(zio_t *zio)
2335{
2336	zio_t *pio = zio_unique_parent(zio);
2337	zio_t *gio = zio->io_gang_leader;
2338	dva_t *cdva = zio->io_bp->blk_dva;
2339	dva_t *pdva = pio->io_bp->blk_dva;
2340	uint64_t asize;
2341
2342	if (BP_IS_HOLE(zio->io_bp))
2343		return;
2344
2345	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2346
2347	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2348	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2349	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2350	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2351	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2352
2353	mutex_enter(&pio->io_lock);
2354	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2355		ASSERT(DVA_GET_GANG(&pdva[d]));
2356		asize = DVA_GET_ASIZE(&pdva[d]);
2357		asize += DVA_GET_ASIZE(&cdva[d]);
2358		DVA_SET_ASIZE(&pdva[d], asize);
2359	}
2360	mutex_exit(&pio->io_lock);
2361}
2362
2363static void
2364zio_write_gang_done(zio_t *zio)
2365{
2366	/*
2367	 * The io_abd field will be NULL for a zio with no data.  The io_flags
2368	 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2369	 * check for it here as it is cleared in zio_ready.
2370	 */
2371	if (zio->io_abd != NULL)
2372		abd_put(zio->io_abd);
2373}
2374
2375static int
2376zio_write_gang_block(zio_t *pio)
2377{
2378	spa_t *spa = pio->io_spa;
2379	metaslab_class_t *mc = spa_normal_class(spa);
2380	blkptr_t *bp = pio->io_bp;
2381	zio_t *gio = pio->io_gang_leader;
2382	zio_t *zio;
2383	zio_gang_node_t *gn, **gnpp;
2384	zio_gbh_phys_t *gbh;
2385	abd_t *gbh_abd;
2386	uint64_t txg = pio->io_txg;
2387	uint64_t resid = pio->io_size;
2388	uint64_t lsize;
2389	int copies = gio->io_prop.zp_copies;
2390	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2391	zio_prop_t zp;
2392	int error;
2393	boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
2394
2395	/*
2396	 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2397	 * have a third copy.
2398	 */
2399	if (gio->io_prop.zp_encrypt && gbh_copies >= SPA_DVAS_PER_BP)
2400		gbh_copies = SPA_DVAS_PER_BP - 1;
2401
2402	int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2403	if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2404		ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2405		ASSERT(has_data);
2406
2407		flags |= METASLAB_ASYNC_ALLOC;
2408		VERIFY(zfs_refcount_held(&mc->mc_alloc_slots[pio->io_allocator],
2409		    pio));
2410
2411		/*
2412		 * The logical zio has already placed a reservation for
2413		 * 'copies' allocation slots but gang blocks may require
2414		 * additional copies. These additional copies
2415		 * (i.e. gbh_copies - copies) are guaranteed to succeed
2416		 * since metaslab_class_throttle_reserve() always allows
2417		 * additional reservations for gang blocks.
2418		 */
2419		VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2420		    pio->io_allocator, pio, flags));
2421	}
2422
2423	error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2424	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2425	    &pio->io_alloc_list, pio, pio->io_allocator);
2426	if (error) {
2427		if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2428			ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2429			ASSERT(has_data);
2430
2431			/*
2432			 * If we failed to allocate the gang block header then
2433			 * we remove any additional allocation reservations that
2434			 * we placed here. The original reservation will
2435			 * be removed when the logical I/O goes to the ready
2436			 * stage.
2437			 */
2438			metaslab_class_throttle_unreserve(mc,
2439			    gbh_copies - copies, pio->io_allocator, pio);
2440		}
2441		pio->io_error = error;
2442		return (ZIO_PIPELINE_CONTINUE);
2443	}
2444
2445	if (pio == gio) {
2446		gnpp = &gio->io_gang_tree;
2447	} else {
2448		gnpp = pio->io_private;
2449		ASSERT(pio->io_ready == zio_write_gang_member_ready);
2450	}
2451
2452	gn = zio_gang_node_alloc(gnpp);
2453	gbh = gn->gn_gbh;
2454	bzero(gbh, SPA_GANGBLOCKSIZE);
2455	gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2456
2457	/*
2458	 * Create the gang header.
2459	 */
2460	zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2461	    zio_write_gang_done, NULL, pio->io_priority,
2462	    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2463
2464	/*
2465	 * Create and nowait the gang children.
2466	 */
2467	for (int g = 0; resid != 0; resid -= lsize, g++) {
2468		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2469		    SPA_MINBLOCKSIZE);
2470		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2471
2472		zp.zp_checksum = gio->io_prop.zp_checksum;
2473		zp.zp_compress = ZIO_COMPRESS_OFF;
2474		zp.zp_type = DMU_OT_NONE;
2475		zp.zp_level = 0;
2476		zp.zp_copies = gio->io_prop.zp_copies;
2477		zp.zp_dedup = B_FALSE;
2478		zp.zp_dedup_verify = B_FALSE;
2479		zp.zp_nopwrite = B_FALSE;
2480		zp.zp_encrypt = gio->io_prop.zp_encrypt;
2481		zp.zp_byteorder = gio->io_prop.zp_byteorder;
2482		bzero(zp.zp_salt, ZIO_DATA_SALT_LEN);
2483		bzero(zp.zp_iv, ZIO_DATA_IV_LEN);
2484		bzero(zp.zp_mac, ZIO_DATA_MAC_LEN);
2485
2486		zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2487		    has_data ? abd_get_offset(pio->io_abd, pio->io_size -
2488		    resid) : NULL, lsize, lsize, &zp,
2489		    zio_write_gang_member_ready, NULL, NULL,
2490		    zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2491		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2492
2493		if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2494			ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2495			ASSERT(has_data);
2496
2497			/*
2498			 * Gang children won't throttle but we should
2499			 * account for their work, so reserve an allocation
2500			 * slot for them here.
2501			 */
2502			VERIFY(metaslab_class_throttle_reserve(mc,
2503			    zp.zp_copies, cio->io_allocator, cio, flags));
2504		}
2505		zio_nowait(cio);
2506	}
2507
2508	/*
2509	 * Set pio's pipeline to just wait for zio to finish.
2510	 */
2511	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2512
2513	zio_nowait(zio);
2514
2515	return (ZIO_PIPELINE_CONTINUE);
2516}
2517
2518/*
2519 * The zio_nop_write stage in the pipeline determines if allocating a
2520 * new bp is necessary.  The nopwrite feature can handle writes in
2521 * either syncing or open context (i.e. zil writes) and as a result is
2522 * mutually exclusive with dedup.
2523 *
2524 * By leveraging a cryptographically secure checksum, such as SHA256, we
2525 * can compare the checksums of the new data and the old to determine if
2526 * allocating a new block is required.  Note that our requirements for
2527 * cryptographic strength are fairly weak: there can't be any accidental
2528 * hash collisions, but we don't need to be secure against intentional
2529 * (malicious) collisions.  To trigger a nopwrite, you have to be able
2530 * to write the file to begin with, and triggering an incorrect (hash
2531 * collision) nopwrite is no worse than simply writing to the file.
2532 * That said, there are no known attacks against the checksum algorithms
2533 * used for nopwrite, assuming that the salt and the checksums
2534 * themselves remain secret.
2535 */
2536static int
2537zio_nop_write(zio_t *zio)
2538{
2539	blkptr_t *bp = zio->io_bp;
2540	blkptr_t *bp_orig = &zio->io_bp_orig;
2541	zio_prop_t *zp = &zio->io_prop;
2542
2543	ASSERT(BP_GET_LEVEL(bp) == 0);
2544	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2545	ASSERT(zp->zp_nopwrite);
2546	ASSERT(!zp->zp_dedup);
2547	ASSERT(zio->io_bp_override == NULL);
2548	ASSERT(IO_IS_ALLOCATING(zio));
2549
2550	/*
2551	 * Check to see if the original bp and the new bp have matching
2552	 * characteristics (i.e. same checksum, compression algorithms, etc).
2553	 * If they don't then just continue with the pipeline which will
2554	 * allocate a new bp.
2555	 */
2556	if (BP_IS_HOLE(bp_orig) ||
2557	    !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2558	    ZCHECKSUM_FLAG_NOPWRITE) ||
2559	    BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) ||
2560	    BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2561	    BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2562	    BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2563	    zp->zp_copies != BP_GET_NDVAS(bp_orig))
2564		return (ZIO_PIPELINE_CONTINUE);
2565
2566	/*
2567	 * If the checksums match then reset the pipeline so that we
2568	 * avoid allocating a new bp and issuing any I/O.
2569	 */
2570	if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2571		ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2572		    ZCHECKSUM_FLAG_NOPWRITE);
2573		ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2574		ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2575		ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2576		ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2577		    sizeof (uint64_t)) == 0);
2578
2579		*bp = *bp_orig;
2580		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2581		zio->io_flags |= ZIO_FLAG_NOPWRITE;
2582	}
2583
2584	return (ZIO_PIPELINE_CONTINUE);
2585}
2586
2587/*
2588 * ==========================================================================
2589 * Dedup
2590 * ==========================================================================
2591 */
2592static void
2593zio_ddt_child_read_done(zio_t *zio)
2594{
2595	blkptr_t *bp = zio->io_bp;
2596	ddt_entry_t *dde = zio->io_private;
2597	ddt_phys_t *ddp;
2598	zio_t *pio = zio_unique_parent(zio);
2599
2600	mutex_enter(&pio->io_lock);
2601	ddp = ddt_phys_select(dde, bp);
2602	if (zio->io_error == 0)
2603		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
2604
2605	if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2606		dde->dde_repair_abd = zio->io_abd;
2607	else
2608		abd_free(zio->io_abd);
2609	mutex_exit(&pio->io_lock);
2610}
2611
2612static int
2613zio_ddt_read_start(zio_t *zio)
2614{
2615	blkptr_t *bp = zio->io_bp;
2616
2617	ASSERT(BP_GET_DEDUP(bp));
2618	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2619	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2620
2621	if (zio->io_child_error[ZIO_CHILD_DDT]) {
2622		ddt_t *ddt = ddt_select(zio->io_spa, bp);
2623		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2624		ddt_phys_t *ddp = dde->dde_phys;
2625		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2626		blkptr_t blk;
2627
2628		ASSERT(zio->io_vsd == NULL);
2629		zio->io_vsd = dde;
2630
2631		if (ddp_self == NULL)
2632			return (ZIO_PIPELINE_CONTINUE);
2633
2634		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2635			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2636				continue;
2637			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2638			    &blk);
2639			zio_nowait(zio_read(zio, zio->io_spa, &blk,
2640			    abd_alloc_for_io(zio->io_size, B_TRUE),
2641			    zio->io_size, zio_ddt_child_read_done, dde,
2642			    zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2643			    ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2644		}
2645		return (ZIO_PIPELINE_CONTINUE);
2646	}
2647
2648	zio_nowait(zio_read(zio, zio->io_spa, bp,
2649	    zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2650	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2651
2652	return (ZIO_PIPELINE_CONTINUE);
2653}
2654
2655static int
2656zio_ddt_read_done(zio_t *zio)
2657{
2658	blkptr_t *bp = zio->io_bp;
2659
2660	if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2661		return (ZIO_PIPELINE_STOP);
2662	}
2663
2664	ASSERT(BP_GET_DEDUP(bp));
2665	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2666	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2667
2668	if (zio->io_child_error[ZIO_CHILD_DDT]) {
2669		ddt_t *ddt = ddt_select(zio->io_spa, bp);
2670		ddt_entry_t *dde = zio->io_vsd;
2671		if (ddt == NULL) {
2672			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2673			return (ZIO_PIPELINE_CONTINUE);
2674		}
2675		if (dde == NULL) {
2676			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2677			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2678			return (ZIO_PIPELINE_STOP);
2679		}
2680		if (dde->dde_repair_abd != NULL) {
2681			abd_copy(zio->io_abd, dde->dde_repair_abd,
2682			    zio->io_size);
2683			zio->io_child_error[ZIO_CHILD_DDT] = 0;
2684		}
2685		ddt_repair_done(ddt, dde);
2686		zio->io_vsd = NULL;
2687	}
2688
2689	ASSERT(zio->io_vsd == NULL);
2690
2691	return (ZIO_PIPELINE_CONTINUE);
2692}
2693
2694static boolean_t
2695zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2696{
2697	spa_t *spa = zio->io_spa;
2698	boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW);
2699
2700	/* We should never get a raw, override zio */
2701	ASSERT(!(zio->io_bp_override && do_raw));
2702
2703	/*
2704	 * Note: we compare the original data, not the transformed data,
2705	 * because when zio->io_bp is an override bp, we will not have
2706	 * pushed the I/O transforms.  That's an important optimization
2707	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2708	 * However, we should never get a raw, override zio so in these
2709	 * cases we can compare the io_data directly. This is useful because
2710	 * it allows us to do dedup verification even if we don't have access
2711	 * to the original data (for instance, if the encryption keys aren't
2712	 * loaded).
2713	 */
2714
2715	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2716		zio_t *lio = dde->dde_lead_zio[p];
2717
2718		if (lio != NULL && do_raw) {
2719			return (lio->io_size != zio->io_size ||
2720			    abd_cmp(zio->io_abd, lio->io_abd,
2721			    zio->io_size) != 0);
2722		} else if (lio != NULL) {
2723			return (lio->io_orig_size != zio->io_orig_size ||
2724			    abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2725			    zio->io_orig_size) != 0);
2726		}
2727	}
2728
2729	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2730		ddt_phys_t *ddp = &dde->dde_phys[p];
2731
2732		if (ddp->ddp_phys_birth != 0 && do_raw) {
2733			blkptr_t blk = *zio->io_bp;
2734			uint64_t psize;
2735			abd_t *tmpabd;
2736			int error;
2737
2738			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2739			psize = BP_GET_PSIZE(&blk);
2740
2741			if (psize != zio->io_size)
2742				return (B_TRUE);
2743
2744			ddt_exit(ddt);
2745
2746			tmpabd = abd_alloc_for_io(psize, B_TRUE);
2747
2748			error = zio_wait(zio_read(NULL, spa, &blk, tmpabd,
2749			    psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
2750			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2751			    ZIO_FLAG_RAW, &zio->io_bookmark));
2752
2753			if (error == 0) {
2754				if (abd_cmp(tmpabd, zio->io_abd, psize) != 0)
2755					error = SET_ERROR(ENOENT);
2756			}
2757
2758			abd_free(tmpabd);
2759			ddt_enter(ddt);
2760			return (error != 0);
2761		} else if (ddp->ddp_phys_birth != 0) {
2762			arc_buf_t *abuf = NULL;
2763			arc_flags_t aflags = ARC_FLAG_WAIT;
2764			int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2765			blkptr_t blk = *zio->io_bp;
2766			int error;
2767
2768			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2769
2770			if (BP_GET_LSIZE(&blk) != zio->io_orig_size)
2771				return (B_TRUE);
2772
2773			ddt_exit(ddt);
2774
2775			/*
2776			 * Intuitively, it would make more sense to compare
2777			 * io_abd than io_orig_abd in the raw case since you
2778			 * don't want to look at any transformations that have
2779			 * happened to the data. However, for raw I/Os the
2780			 * data will actually be the same in io_abd and
2781			 * io_orig_abd, so all we have to do is issue this as
2782			 * a raw ARC read.
2783			 */
2784			if (do_raw) {
2785				zio_flags |= ZIO_FLAG_RAW;
2786				ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2787				ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2788				    zio->io_size));
2789				ASSERT3P(zio->io_transform_stack, ==, NULL);
2790			}
2791
2792			error = arc_read(NULL, spa, &blk,
2793			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2794			    zio_flags, &aflags, &zio->io_bookmark);
2795
2796			if (error == 0) {
2797				if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2798				    zio->io_orig_size) != 0)
2799					error = SET_ERROR(ENOENT);
2800				arc_buf_destroy(abuf, &abuf);
2801			}
2802
2803			ddt_enter(ddt);
2804			return (error != 0);
2805		}
2806	}
2807
2808	return (B_FALSE);
2809}
2810
2811static void
2812zio_ddt_child_write_ready(zio_t *zio)
2813{
2814	int p = zio->io_prop.zp_copies;
2815	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2816	ddt_entry_t *dde = zio->io_private;
2817	ddt_phys_t *ddp = &dde->dde_phys[p];
2818	zio_t *pio;
2819
2820	if (zio->io_error)
2821		return;
2822
2823	ddt_enter(ddt);
2824
2825	ASSERT(dde->dde_lead_zio[p] == zio);
2826
2827	ddt_phys_fill(ddp, zio->io_bp);
2828
2829	zio_link_t *zl = NULL;
2830	while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2831		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2832
2833	ddt_exit(ddt);
2834}
2835
2836static void
2837zio_ddt_child_write_done(zio_t *zio)
2838{
2839	int p = zio->io_prop.zp_copies;
2840	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2841	ddt_entry_t *dde = zio->io_private;
2842	ddt_phys_t *ddp = &dde->dde_phys[p];
2843
2844	ddt_enter(ddt);
2845
2846	ASSERT(ddp->ddp_refcnt == 0);
2847	ASSERT(dde->dde_lead_zio[p] == zio);
2848	dde->dde_lead_zio[p] = NULL;
2849
2850	if (zio->io_error == 0) {
2851		zio_link_t *zl = NULL;
2852		while (zio_walk_parents(zio, &zl) != NULL)
2853			ddt_phys_addref(ddp);
2854	} else {
2855		ddt_phys_clear(ddp);
2856	}
2857
2858	ddt_exit(ddt);
2859}
2860
2861static void
2862zio_ddt_ditto_write_done(zio_t *zio)
2863{
2864	int p = DDT_PHYS_DITTO;
2865	zio_prop_t *zp = &zio->io_prop;
2866	blkptr_t *bp = zio->io_bp;
2867	ddt_t *ddt = ddt_select(zio->io_spa, bp);
2868	ddt_entry_t *dde = zio->io_private;
2869	ddt_phys_t *ddp = &dde->dde_phys[p];
2870	ddt_key_t *ddk = &dde->dde_key;
2871
2872	ddt_enter(ddt);
2873
2874	ASSERT(ddp->ddp_refcnt == 0);
2875	ASSERT(dde->dde_lead_zio[p] == zio);
2876	dde->dde_lead_zio[p] = NULL;
2877
2878	if (zio->io_error == 0) {
2879		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2880		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2881		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2882		if (ddp->ddp_phys_birth != 0)
2883			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2884		ddt_phys_fill(ddp, bp);
2885	}
2886
2887	ddt_exit(ddt);
2888}
2889
2890static int
2891zio_ddt_write(zio_t *zio)
2892{
2893	spa_t *spa = zio->io_spa;
2894	blkptr_t *bp = zio->io_bp;
2895	uint64_t txg = zio->io_txg;
2896	zio_prop_t *zp = &zio->io_prop;
2897	int p = zp->zp_copies;
2898	int ditto_copies;
2899	zio_t *cio = NULL;
2900	zio_t *dio = NULL;
2901	ddt_t *ddt = ddt_select(spa, bp);
2902	ddt_entry_t *dde;
2903	ddt_phys_t *ddp;
2904
2905	ASSERT(BP_GET_DEDUP(bp));
2906	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2907	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2908	ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2909
2910	ddt_enter(ddt);
2911	dde = ddt_lookup(ddt, bp, B_TRUE);
2912	ddp = &dde->dde_phys[p];
2913
2914	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2915		/*
2916		 * If we're using a weak checksum, upgrade to a strong checksum
2917		 * and try again.  If we're already using a strong checksum,
2918		 * we can't resolve it, so just convert to an ordinary write.
2919		 * (And automatically e-mail a paper to Nature?)
2920		 */
2921		if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2922		    ZCHECKSUM_FLAG_DEDUP)) {
2923			zp->zp_checksum = spa_dedup_checksum(spa);
2924			zio_pop_transforms(zio);
2925			zio->io_stage = ZIO_STAGE_OPEN;
2926			BP_ZERO(bp);
2927		} else {
2928			zp->zp_dedup = B_FALSE;
2929			BP_SET_DEDUP(bp, B_FALSE);
2930		}
2931		ASSERT(!BP_GET_DEDUP(bp));
2932		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2933		ddt_exit(ddt);
2934		return (ZIO_PIPELINE_CONTINUE);
2935	}
2936
2937	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2938	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2939
2940	if (ditto_copies > ddt_ditto_copies_present(dde) &&
2941	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2942		zio_prop_t czp = *zp;
2943
2944		czp.zp_copies = ditto_copies;
2945
2946		/*
2947		 * If we arrived here with an override bp, we won't have run
2948		 * the transform stack, so we won't have the data we need to
2949		 * generate a child i/o.  So, toss the override bp and restart.
2950		 * This is safe, because using the override bp is just an
2951		 * optimization; and it's rare, so the cost doesn't matter.
2952		 */
2953		if (zio->io_bp_override) {
2954			zio_pop_transforms(zio);
2955			zio->io_stage = ZIO_STAGE_OPEN;
2956			zio->io_pipeline = ZIO_WRITE_PIPELINE;
2957			zio->io_bp_override = NULL;
2958			BP_ZERO(bp);
2959			ddt_exit(ddt);
2960			return (ZIO_PIPELINE_CONTINUE);
2961		}
2962
2963		dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2964		    zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2965		    NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2966		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2967
2968		zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2969		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2970	}
2971
2972	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2973		if (ddp->ddp_phys_birth != 0)
2974			ddt_bp_fill(ddp, bp, txg);
2975		if (dde->dde_lead_zio[p] != NULL)
2976			zio_add_child(zio, dde->dde_lead_zio[p]);
2977		else
2978			ddt_phys_addref(ddp);
2979	} else if (zio->io_bp_override) {
2980		ASSERT(bp->blk_birth == txg);
2981		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2982		ddt_phys_fill(ddp, bp);
2983		ddt_phys_addref(ddp);
2984	} else {
2985		cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2986		    zio->io_orig_size, zio->io_orig_size, zp,
2987		    zio_ddt_child_write_ready, NULL, NULL,
2988		    zio_ddt_child_write_done, dde, zio->io_priority,
2989		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2990
2991		zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
2992		dde->dde_lead_zio[p] = cio;
2993	}
2994
2995	ddt_exit(ddt);
2996
2997	if (cio)
2998		zio_nowait(cio);
2999	if (dio)
3000		zio_nowait(dio);
3001
3002	return (ZIO_PIPELINE_CONTINUE);
3003}
3004
3005ddt_entry_t *freedde; /* for debugging */
3006
3007static int
3008zio_ddt_free(zio_t *zio)
3009{
3010	spa_t *spa = zio->io_spa;
3011	blkptr_t *bp = zio->io_bp;
3012	ddt_t *ddt = ddt_select(spa, bp);
3013	ddt_entry_t *dde;
3014	ddt_phys_t *ddp;
3015
3016	ASSERT(BP_GET_DEDUP(bp));
3017	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3018
3019	ddt_enter(ddt);
3020	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
3021	ddp = ddt_phys_select(dde, bp);
3022	ddt_phys_decref(ddp);
3023	ddt_exit(ddt);
3024
3025	return (ZIO_PIPELINE_CONTINUE);
3026}
3027
3028/*
3029 * ==========================================================================
3030 * Allocate and free blocks
3031 * ==========================================================================
3032 */
3033
3034static zio_t *
3035zio_io_to_allocate(spa_t *spa, int allocator)
3036{
3037	zio_t *zio;
3038
3039	ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator]));
3040
3041	zio = avl_first(&spa->spa_alloc_trees[allocator]);
3042	if (zio == NULL)
3043		return (NULL);
3044
3045	ASSERT(IO_IS_ALLOCATING(zio));
3046
3047	/*
3048	 * Try to place a reservation for this zio. If we're unable to
3049	 * reserve then we throttle.
3050	 */
3051	ASSERT3U(zio->io_allocator, ==, allocator);
3052	if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
3053	    zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) {
3054		return (NULL);
3055	}
3056
3057	avl_remove(&spa->spa_alloc_trees[allocator], zio);
3058	ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
3059
3060	return (zio);
3061}
3062
3063static int
3064zio_dva_throttle(zio_t *zio)
3065{
3066	spa_t *spa = zio->io_spa;
3067	zio_t *nio;
3068	metaslab_class_t *mc;
3069
3070	/* locate an appropriate allocation class */
3071	mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type,
3072	    zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk);
3073
3074	if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
3075	    !mc->mc_alloc_throttle_enabled ||
3076	    zio->io_child_type == ZIO_CHILD_GANG ||
3077	    zio->io_flags & ZIO_FLAG_NODATA) {
3078		return (ZIO_PIPELINE_CONTINUE);
3079	}
3080
3081	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3082
3083	ASSERT3U(zio->io_queued_timestamp, >, 0);
3084	ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3085
3086	zbookmark_phys_t *bm = &zio->io_bookmark;
3087	/*
3088	 * We want to try to use as many allocators as possible to help improve
3089	 * performance, but we also want logically adjacent IOs to be physically
3090	 * adjacent to improve sequential read performance. We chunk each object
3091	 * into 2^20 block regions, and then hash based on the objset, object,
3092	 * level, and region to accomplish both of these goals.
3093	 */
3094	zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object,
3095	    bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
3096	mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]);
3097	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3098	zio->io_metaslab_class = mc;
3099	avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio);
3100	nio = zio_io_to_allocate(spa, zio->io_allocator);
3101	mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]);
3102
3103	if (nio == zio)
3104		return (ZIO_PIPELINE_CONTINUE);
3105
3106	if (nio != NULL) {
3107		ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3108		/*
3109		 * We are passing control to a new zio so make sure that
3110		 * it is processed by a different thread. We do this to
3111		 * avoid stack overflows that can occur when parents are
3112		 * throttled and children are making progress. We allow
3113		 * it to go to the head of the taskq since it's already
3114		 * been waiting.
3115		 */
3116		zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
3117	}
3118	return (ZIO_PIPELINE_STOP);
3119}
3120
3121static void
3122zio_allocate_dispatch(spa_t *spa, int allocator)
3123{
3124	zio_t *zio;
3125
3126	mutex_enter(&spa->spa_alloc_locks[allocator]);
3127	zio = zio_io_to_allocate(spa, allocator);
3128	mutex_exit(&spa->spa_alloc_locks[allocator]);
3129	if (zio == NULL)
3130		return;
3131
3132	ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3133	ASSERT0(zio->io_error);
3134	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3135}
3136
3137static int
3138zio_dva_allocate(zio_t *zio)
3139{
3140	spa_t *spa = zio->io_spa;
3141	metaslab_class_t *mc;
3142	blkptr_t *bp = zio->io_bp;
3143	int error;
3144	int flags = 0;
3145
3146	if (zio->io_gang_leader == NULL) {
3147		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3148		zio->io_gang_leader = zio;
3149	}
3150
3151	ASSERT(BP_IS_HOLE(bp));
3152	ASSERT0(BP_GET_NDVAS(bp));
3153	ASSERT3U(zio->io_prop.zp_copies, >, 0);
3154	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3155	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3156
3157	if (zio->io_flags & ZIO_FLAG_NODATA)
3158		flags |= METASLAB_DONT_THROTTLE;
3159	if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
3160		flags |= METASLAB_GANG_CHILD;
3161	if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
3162		flags |= METASLAB_ASYNC_ALLOC;
3163
3164	/*
3165	 * if not already chosen, locate an appropriate allocation class
3166	 */
3167	mc = zio->io_metaslab_class;
3168	if (mc == NULL) {
3169		mc = spa_preferred_class(spa, zio->io_size,
3170		    zio->io_prop.zp_type, zio->io_prop.zp_level,
3171		    zio->io_prop.zp_zpl_smallblk);
3172		zio->io_metaslab_class = mc;
3173	}
3174
3175	error = metaslab_alloc(spa, mc, zio->io_size, bp,
3176	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3177	    &zio->io_alloc_list, zio, zio->io_allocator);
3178
3179	/*
3180	 * Fallback to normal class when an alloc class is full
3181	 */
3182	if (error == ENOSPC && mc != spa_normal_class(spa)) {
3183		/*
3184		 * If throttling, transfer reservation over to normal class.
3185		 * The io_allocator slot can remain the same even though we
3186		 * are switching classes.
3187		 */
3188		if (mc->mc_alloc_throttle_enabled &&
3189		    (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) {
3190			metaslab_class_throttle_unreserve(mc,
3191			    zio->io_prop.zp_copies, zio->io_allocator, zio);
3192			zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING;
3193
3194			mc = spa_normal_class(spa);
3195			VERIFY(metaslab_class_throttle_reserve(mc,
3196			    zio->io_prop.zp_copies, zio->io_allocator, zio,
3197			    flags | METASLAB_MUST_RESERVE));
3198		} else {
3199			mc = spa_normal_class(spa);
3200		}
3201		zio->io_metaslab_class = mc;
3202
3203		error = metaslab_alloc(spa, mc, zio->io_size, bp,
3204		    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3205		    &zio->io_alloc_list, zio, zio->io_allocator);
3206	}
3207
3208	if (error != 0) {
3209		zfs_dbgmsg("%s: metaslab allocation failure: zio %p, "
3210		    "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3211		    error);
3212		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
3213			return (zio_write_gang_block(zio));
3214		zio->io_error = error;
3215	}
3216
3217	return (ZIO_PIPELINE_CONTINUE);
3218}
3219
3220static int
3221zio_dva_free(zio_t *zio)
3222{
3223	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3224
3225	return (ZIO_PIPELINE_CONTINUE);
3226}
3227
3228static int
3229zio_dva_claim(zio_t *zio)
3230{
3231	int error;
3232
3233	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3234	if (error)
3235		zio->io_error = error;
3236
3237	return (ZIO_PIPELINE_CONTINUE);
3238}
3239
3240/*
3241 * Undo an allocation.  This is used by zio_done() when an I/O fails
3242 * and we want to give back the block we just allocated.
3243 * This handles both normal blocks and gang blocks.
3244 */
3245static void
3246zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3247{
3248	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3249	ASSERT(zio->io_bp_override == NULL);
3250
3251	if (!BP_IS_HOLE(bp))
3252		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3253
3254	if (gn != NULL) {
3255		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3256			zio_dva_unallocate(zio, gn->gn_child[g],
3257			    &gn->gn_gbh->zg_blkptr[g]);
3258		}
3259	}
3260}
3261
3262/*
3263 * Try to allocate an intent log block.  Return 0 on success, errno on failure.
3264 */
3265int
3266zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
3267    blkptr_t *old_bp, uint64_t size, boolean_t *slog)
3268{
3269	int error = 1;
3270	zio_alloc_list_t io_alloc_list;
3271
3272	ASSERT(txg > spa_syncing_txg(spa));
3273
3274	metaslab_trace_init(&io_alloc_list);
3275
3276	/*
3277	 * Block pointer fields are useful to metaslabs for stats and debugging.
3278	 * Fill in the obvious ones before calling into metaslab_alloc().
3279	 */
3280	BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3281	BP_SET_PSIZE(new_bp, size);
3282	BP_SET_LEVEL(new_bp, 0);
3283
3284	/*
3285	 * When allocating a zil block, we don't have information about
3286	 * the final destination of the block except the objset it's part
3287	 * of, so we just hash the objset ID to pick the allocator to get
3288	 * some parallelism.
3289	 */
3290	error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3291	    txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL,
3292	    cityhash4(0, 0, 0,
3293	    os->os_dsl_dataset->ds_object) % spa->spa_alloc_count);
3294	if (error == 0) {
3295		*slog = TRUE;
3296	} else {
3297		error = metaslab_alloc(spa, spa_normal_class(spa), size,
3298		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3299		    &io_alloc_list, NULL, cityhash4(0, 0, 0,
3300		    os->os_dsl_dataset->ds_object) % spa->spa_alloc_count);
3301		if (error == 0)
3302			*slog = FALSE;
3303	}
3304	metaslab_trace_fini(&io_alloc_list);
3305
3306	if (error == 0) {
3307		BP_SET_LSIZE(new_bp, size);
3308		BP_SET_PSIZE(new_bp, size);
3309		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3310		BP_SET_CHECKSUM(new_bp,
3311		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3312		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3313		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3314		BP_SET_LEVEL(new_bp, 0);
3315		BP_SET_DEDUP(new_bp, 0);
3316		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3317
3318		/*
3319		 * encrypted blocks will require an IV and salt. We generate
3320		 * these now since we will not be rewriting the bp at
3321		 * rewrite time.
3322		 */
3323		if (os->os_encrypted) {
3324			uint8_t iv[ZIO_DATA_IV_LEN];
3325			uint8_t salt[ZIO_DATA_SALT_LEN];
3326
3327			BP_SET_CRYPT(new_bp, B_TRUE);
3328			VERIFY0(spa_crypt_get_salt(spa,
3329			    dmu_objset_id(os), salt));
3330			VERIFY0(zio_crypt_generate_iv(iv));
3331
3332			zio_crypt_encode_params_bp(new_bp, salt, iv);
3333		}
3334	} else {
3335		zfs_dbgmsg("%s: zil block allocation failure: "
3336		    "size %llu, error %d", spa_name(spa), size, error);
3337	}
3338
3339	return (error);
3340}
3341
3342/*
3343 * ==========================================================================
3344 * Read and write to physical devices
3345 * ==========================================================================
3346 */
3347
3348
3349/*
3350 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3351 * stops after this stage and will resume upon I/O completion.
3352 * However, there are instances where the vdev layer may need to
3353 * continue the pipeline when an I/O was not issued. Since the I/O
3354 * that was sent to the vdev layer might be different than the one
3355 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3356 * force the underlying vdev layers to call either zio_execute() or
3357 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3358 */
3359static int
3360zio_vdev_io_start(zio_t *zio)
3361{
3362	vdev_t *vd = zio->io_vd;
3363	uint64_t align;
3364	spa_t *spa = zio->io_spa;
3365
3366	ASSERT(zio->io_error == 0);
3367	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3368
3369	if (vd == NULL) {
3370		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3371			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3372
3373		/*
3374		 * The mirror_ops handle multiple DVAs in a single BP.
3375		 */
3376		vdev_mirror_ops.vdev_op_io_start(zio);
3377		return (ZIO_PIPELINE_STOP);
3378	}
3379
3380	ASSERT3P(zio->io_logical, !=, zio);
3381	if (zio->io_type == ZIO_TYPE_WRITE) {
3382		ASSERT(spa->spa_trust_config);
3383
3384		/*
3385		 * Note: the code can handle other kinds of writes,
3386		 * but we don't expect them.
3387		 */
3388		if (zio->io_vd->vdev_removing) {
3389			ASSERT(zio->io_flags &
3390			    (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3391			    ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3392		}
3393	}
3394
3395	align = 1ULL << vd->vdev_top->vdev_ashift;
3396
3397	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3398	    P2PHASE(zio->io_size, align) != 0) {
3399		/* Transform logical writes to be a full physical block size. */
3400		uint64_t asize = P2ROUNDUP(zio->io_size, align);
3401		abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3402		ASSERT(vd == vd->vdev_top);
3403		if (zio->io_type == ZIO_TYPE_WRITE) {
3404			abd_copy(abuf, zio->io_abd, zio->io_size);
3405			abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3406		}
3407		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3408	}
3409
3410	/*
3411	 * If this is not a physical io, make sure that it is properly aligned
3412	 * before proceeding.
3413	 */
3414	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3415		ASSERT0(P2PHASE(zio->io_offset, align));
3416		ASSERT0(P2PHASE(zio->io_size, align));
3417	} else {
3418		/*
3419		 * For physical writes, we allow 512b aligned writes and assume
3420		 * the device will perform a read-modify-write as necessary.
3421		 */
3422		ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3423		ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3424	}
3425
3426	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
3427
3428	/*
3429	 * If this is a repair I/O, and there's no self-healing involved --
3430	 * that is, we're just resilvering what we expect to resilver --
3431	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3432	 * This prevents spurious resilvering.
3433	 *
3434	 * There are a few ways that we can end up creating these spurious
3435	 * resilver i/os:
3436	 *
3437	 * 1. A resilver i/o will be issued if any DVA in the BP has a
3438	 * dirty DTL.  The mirror code will issue resilver writes to
3439	 * each DVA, including the one(s) that are not on vdevs with dirty
3440	 * DTLs.
3441	 *
3442	 * 2. With nested replication, which happens when we have a
3443	 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3444	 * For example, given mirror(replacing(A+B), C), it's likely that
3445	 * only A is out of date (it's the new device). In this case, we'll
3446	 * read from C, then use the data to resilver A+B -- but we don't
3447	 * actually want to resilver B, just A. The top-level mirror has no
3448	 * way to know this, so instead we just discard unnecessary repairs
3449	 * as we work our way down the vdev tree.
3450	 *
3451	 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3452	 * The same logic applies to any form of nested replication: ditto
3453	 * + mirror, RAID-Z + replacing, etc.
3454	 *
3455	 * However, indirect vdevs point off to other vdevs which may have
3456	 * DTL's, so we never bypass them.  The child i/os on concrete vdevs
3457	 * will be properly bypassed instead.
3458	 */
3459	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3460	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3461	    zio->io_txg != 0 &&	/* not a delegated i/o */
3462	    vd->vdev_ops != &vdev_indirect_ops &&
3463	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3464		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3465		zio_vdev_io_bypass(zio);
3466		return (ZIO_PIPELINE_CONTINUE);
3467	}
3468
3469	if (vd->vdev_ops->vdev_op_leaf &&
3470	    (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
3471
3472		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
3473			return (ZIO_PIPELINE_CONTINUE);
3474
3475		if ((zio = vdev_queue_io(zio)) == NULL)
3476			return (ZIO_PIPELINE_STOP);
3477
3478		if (!vdev_accessible(vd, zio)) {
3479			zio->io_error = SET_ERROR(ENXIO);
3480			zio_interrupt(zio);
3481			return (ZIO_PIPELINE_STOP);
3482		}
3483	}
3484
3485	vd->vdev_ops->vdev_op_io_start(zio);
3486	return (ZIO_PIPELINE_STOP);
3487}
3488
3489static int
3490zio_vdev_io_done(zio_t *zio)
3491{
3492	vdev_t *vd = zio->io_vd;
3493	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3494	boolean_t unexpected_error = B_FALSE;
3495
3496	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3497		return (ZIO_PIPELINE_STOP);
3498	}
3499
3500	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
3501
3502	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
3503
3504		vdev_queue_io_done(zio);
3505
3506		if (zio->io_type == ZIO_TYPE_WRITE)
3507			vdev_cache_write(zio);
3508
3509		if (zio_injection_enabled && zio->io_error == 0)
3510			zio->io_error = zio_handle_device_injection(vd,
3511			    zio, EIO);
3512
3513		if (zio_injection_enabled && zio->io_error == 0)
3514			zio->io_error = zio_handle_label_injection(zio, EIO);
3515
3516		if (zio->io_error) {
3517			if (!vdev_accessible(vd, zio)) {
3518				zio->io_error = SET_ERROR(ENXIO);
3519			} else {
3520				unexpected_error = B_TRUE;
3521			}
3522		}
3523	}
3524
3525	ops->vdev_op_io_done(zio);
3526
3527	if (unexpected_error)
3528		VERIFY(vdev_probe(vd, zio) == NULL);
3529
3530	return (ZIO_PIPELINE_CONTINUE);
3531}
3532
3533/*
3534 * This function is used to change the priority of an existing zio that is
3535 * currently in-flight. This is used by the arc to upgrade priority in the
3536 * event that a demand read is made for a block that is currently queued
3537 * as a scrub or async read IO. Otherwise, the high priority read request
3538 * would end up having to wait for the lower priority IO.
3539 */
3540void
3541zio_change_priority(zio_t *pio, zio_priority_t priority)
3542{
3543	zio_t *cio, *cio_next;
3544	zio_link_t *zl = NULL;
3545
3546	ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3547
3548	if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3549		vdev_queue_change_io_priority(pio, priority);
3550	} else {
3551		pio->io_priority = priority;
3552	}
3553
3554	mutex_enter(&pio->io_lock);
3555	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3556		cio_next = zio_walk_children(pio, &zl);
3557		zio_change_priority(cio, priority);
3558	}
3559	mutex_exit(&pio->io_lock);
3560}
3561
3562/*
3563 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3564 * disk, and use that to finish the checksum ereport later.
3565 */
3566static void
3567zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3568    const abd_t *good_buf)
3569{
3570	/* no processing needed */
3571	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3572}
3573
3574/*ARGSUSED*/
3575void
3576zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3577{
3578	void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size);
3579
3580	abd_copy(abd, zio->io_abd, zio->io_size);
3581
3582	zcr->zcr_cbinfo = zio->io_size;
3583	zcr->zcr_cbdata = abd;
3584	zcr->zcr_finish = zio_vsd_default_cksum_finish;
3585	zcr->zcr_free = zio_abd_free;
3586}
3587
3588static int
3589zio_vdev_io_assess(zio_t *zio)
3590{
3591	vdev_t *vd = zio->io_vd;
3592
3593	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3594		return (ZIO_PIPELINE_STOP);
3595	}
3596
3597	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3598		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3599
3600	if (zio->io_vsd != NULL) {
3601		zio->io_vsd_ops->vsd_free(zio);
3602		zio->io_vsd = NULL;
3603	}
3604
3605	if (zio_injection_enabled && zio->io_error == 0)
3606		zio->io_error = zio_handle_fault_injection(zio, EIO);
3607
3608	/*
3609	 * If the I/O failed, determine whether we should attempt to retry it.
3610	 *
3611	 * On retry, we cut in line in the issue queue, since we don't want
3612	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3613	 */
3614	if (zio->io_error && vd == NULL &&
3615	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3616		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
3617		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
3618		zio->io_error = 0;
3619		zio->io_flags |= ZIO_FLAG_IO_RETRY |
3620		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3621		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3622		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3623		    zio_requeue_io_start_cut_in_line);
3624		return (ZIO_PIPELINE_STOP);
3625	}
3626
3627	/*
3628	 * If we got an error on a leaf device, convert it to ENXIO
3629	 * if the device is not accessible at all.
3630	 */
3631	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3632	    !vdev_accessible(vd, zio))
3633		zio->io_error = SET_ERROR(ENXIO);
3634
3635	/*
3636	 * If we can't write to an interior vdev (mirror or RAID-Z),
3637	 * set vdev_cant_write so that we stop trying to allocate from it.
3638	 */
3639	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3640	    vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3641		vd->vdev_cant_write = B_TRUE;
3642	}
3643
3644	/*
3645	 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3646	 * attempts will ever succeed. In this case we set a persistent bit so
3647	 * that we don't bother with it in the future.
3648	 */
3649	if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3650	    zio->io_type == ZIO_TYPE_IOCTL &&
3651	    zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3652		vd->vdev_nowritecache = B_TRUE;
3653
3654	if (zio->io_error)
3655		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3656
3657	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3658	    zio->io_physdone != NULL) {
3659		ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3660		ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3661		zio->io_physdone(zio->io_logical);
3662	}
3663
3664	return (ZIO_PIPELINE_CONTINUE);
3665}
3666
3667void
3668zio_vdev_io_reissue(zio_t *zio)
3669{
3670	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3671	ASSERT(zio->io_error == 0);
3672
3673	zio->io_stage >>= 1;
3674}
3675
3676void
3677zio_vdev_io_redone(zio_t *zio)
3678{
3679	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3680
3681	zio->io_stage >>= 1;
3682}
3683
3684void
3685zio_vdev_io_bypass(zio_t *zio)
3686{
3687	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3688	ASSERT(zio->io_error == 0);
3689
3690	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3691	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3692}
3693
3694/*
3695 * ==========================================================================
3696 * Encrypt and store encryption parameters
3697 * ==========================================================================
3698 */
3699
3700
3701/*
3702 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
3703 * managing the storage of encryption parameters and passing them to the
3704 * lower-level encryption functions.
3705 */
3706static int
3707zio_encrypt(zio_t *zio)
3708{
3709	zio_prop_t *zp = &zio->io_prop;
3710	spa_t *spa = zio->io_spa;
3711	blkptr_t *bp = zio->io_bp;
3712	uint64_t psize = BP_GET_PSIZE(bp);
3713	uint64_t dsobj = zio->io_bookmark.zb_objset;
3714	dmu_object_type_t ot = BP_GET_TYPE(bp);
3715	void *enc_buf = NULL;
3716	abd_t *eabd = NULL;
3717	uint8_t salt[ZIO_DATA_SALT_LEN];
3718	uint8_t iv[ZIO_DATA_IV_LEN];
3719	uint8_t mac[ZIO_DATA_MAC_LEN];
3720	boolean_t no_crypt = B_FALSE;
3721
3722	/* the root zio already encrypted the data */
3723	if (zio->io_child_type == ZIO_CHILD_GANG)
3724		return (ZIO_PIPELINE_CONTINUE);
3725
3726	/* only ZIL blocks are re-encrypted on rewrite */
3727	if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG)
3728		return (ZIO_PIPELINE_CONTINUE);
3729
3730	if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) {
3731		BP_SET_CRYPT(bp, B_FALSE);
3732		return (ZIO_PIPELINE_CONTINUE);
3733	}
3734
3735	/* if we are doing raw encryption set the provided encryption params */
3736	if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) {
3737		ASSERT0(BP_GET_LEVEL(bp));
3738		BP_SET_CRYPT(bp, B_TRUE);
3739		BP_SET_BYTEORDER(bp, zp->zp_byteorder);
3740		if (ot != DMU_OT_OBJSET)
3741			zio_crypt_encode_mac_bp(bp, zp->zp_mac);
3742
3743		/* dnode blocks must be written out in the provided byteorder */
3744		if (zp->zp_byteorder != ZFS_HOST_BYTEORDER &&
3745		    ot == DMU_OT_DNODE) {
3746			void *bswap_buf = zio_buf_alloc(psize);
3747			abd_t *babd = abd_get_from_buf(bswap_buf, psize);
3748
3749			ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
3750			abd_copy_to_buf(bswap_buf, zio->io_abd, psize);
3751			dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf,
3752			    psize);
3753
3754			abd_take_ownership_of_buf(babd, B_TRUE);
3755			zio_push_transform(zio, babd, psize, psize, NULL);
3756		}
3757
3758		if (DMU_OT_IS_ENCRYPTED(ot))
3759			zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv);
3760		return (ZIO_PIPELINE_CONTINUE);
3761	}
3762
3763	/* indirect blocks only maintain a cksum of the lower level MACs */
3764	if (BP_GET_LEVEL(bp) > 0) {
3765		BP_SET_CRYPT(bp, B_TRUE);
3766		VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE,
3767		    zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp),
3768		    mac));
3769		zio_crypt_encode_mac_bp(bp, mac);
3770		return (ZIO_PIPELINE_CONTINUE);
3771	}
3772
3773	/*
3774	 * Objset blocks are a special case since they have 2 256-bit MACs
3775	 * embedded within them.
3776	 */
3777	if (ot == DMU_OT_OBJSET) {
3778		ASSERT0(DMU_OT_IS_ENCRYPTED(ot));
3779		ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
3780		BP_SET_CRYPT(bp, B_TRUE);
3781		VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj,
3782		    zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp)));
3783		return (ZIO_PIPELINE_CONTINUE);
3784	}
3785
3786	/* unencrypted object types are only authenticated with a MAC */
3787	if (!DMU_OT_IS_ENCRYPTED(ot)) {
3788		BP_SET_CRYPT(bp, B_TRUE);
3789		VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj,
3790		    zio->io_abd, psize, mac));
3791		zio_crypt_encode_mac_bp(bp, mac);
3792		return (ZIO_PIPELINE_CONTINUE);
3793	}
3794
3795	/*
3796	 * Later passes of sync-to-convergence may decide to rewrite data
3797	 * in place to avoid more disk reallocations. This presents a problem
3798	 * for encryption because this consitutes rewriting the new data with
3799	 * the same encryption key and IV. However, this only applies to blocks
3800	 * in the MOS (particularly the spacemaps) and we do not encrypt the
3801	 * MOS. We assert that the zio is allocating or an intent log write
3802	 * to enforce this.
3803	 */
3804	ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG);
3805	ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG);
3806	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION));
3807	ASSERT3U(psize, !=, 0);
3808
3809	enc_buf = zio_buf_alloc(psize);
3810	eabd = abd_get_from_buf(enc_buf, psize);
3811	abd_take_ownership_of_buf(eabd, B_TRUE);
3812
3813	/*
3814	 * For an explanation of what encryption parameters are stored
3815	 * where, see the block comment in zio_crypt.c.
3816	 */
3817	if (ot == DMU_OT_INTENT_LOG) {
3818		zio_crypt_decode_params_bp(bp, salt, iv);
3819	} else {
3820		BP_SET_CRYPT(bp, B_TRUE);
3821	}
3822
3823	/* Perform the encryption. This should not fail */
3824	VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark,
3825	    BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
3826	    salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt));
3827
3828	/* encode encryption metadata into the bp */
3829	if (ot == DMU_OT_INTENT_LOG) {
3830		/*
3831		 * ZIL blocks store the MAC in the embedded checksum, so the
3832		 * transform must always be applied.
3833		 */
3834		zio_crypt_encode_mac_zil(enc_buf, mac);
3835		zio_push_transform(zio, eabd, psize, psize, NULL);
3836	} else {
3837		BP_SET_CRYPT(bp, B_TRUE);
3838		zio_crypt_encode_params_bp(bp, salt, iv);
3839		zio_crypt_encode_mac_bp(bp, mac);
3840
3841		if (no_crypt) {
3842			ASSERT3U(ot, ==, DMU_OT_DNODE);
3843			abd_free(eabd);
3844		} else {
3845			zio_push_transform(zio, eabd, psize, psize, NULL);
3846		}
3847	}
3848
3849	return (ZIO_PIPELINE_CONTINUE);
3850}
3851
3852/*
3853 * ==========================================================================
3854 * Generate and verify checksums
3855 * ==========================================================================
3856 */
3857static int
3858zio_checksum_generate(zio_t *zio)
3859{
3860	blkptr_t *bp = zio->io_bp;
3861	enum zio_checksum checksum;
3862
3863	if (bp == NULL) {
3864		/*
3865		 * This is zio_write_phys().
3866		 * We're either generating a label checksum, or none at all.
3867		 */
3868		checksum = zio->io_prop.zp_checksum;
3869
3870		if (checksum == ZIO_CHECKSUM_OFF)
3871			return (ZIO_PIPELINE_CONTINUE);
3872
3873		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3874	} else {
3875		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3876			ASSERT(!IO_IS_ALLOCATING(zio));
3877			checksum = ZIO_CHECKSUM_GANG_HEADER;
3878		} else {
3879			checksum = BP_GET_CHECKSUM(bp);
3880		}
3881	}
3882
3883	zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
3884
3885	return (ZIO_PIPELINE_CONTINUE);
3886}
3887
3888static int
3889zio_checksum_verify(zio_t *zio)
3890{
3891	zio_bad_cksum_t info;
3892	blkptr_t *bp = zio->io_bp;
3893	int error;
3894
3895	ASSERT(zio->io_vd != NULL);
3896
3897	if (bp == NULL) {
3898		/*
3899		 * This is zio_read_phys().
3900		 * We're either verifying a label checksum, or nothing at all.
3901		 */
3902		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3903			return (ZIO_PIPELINE_CONTINUE);
3904
3905		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3906	}
3907
3908	if ((error = zio_checksum_error(zio, &info)) != 0) {
3909		zio->io_error = error;
3910		if (error == ECKSUM &&
3911		    !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3912			zfs_ereport_start_checksum(zio->io_spa,
3913			    zio->io_vd, &zio->io_bookmark, zio,
3914			    zio->io_offset, zio->io_size, NULL, &info);
3915		}
3916	}
3917
3918	return (ZIO_PIPELINE_CONTINUE);
3919}
3920
3921/*
3922 * Called by RAID-Z to ensure we don't compute the checksum twice.
3923 */
3924void
3925zio_checksum_verified(zio_t *zio)
3926{
3927	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3928}
3929
3930/*
3931 * ==========================================================================
3932 * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3933 * An error of 0 indicates success.  ENXIO indicates whole-device failure,
3934 * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
3935 * indicate errors that are specific to one I/O, and most likely permanent.
3936 * Any other error is presumed to be worse because we weren't expecting it.
3937 * ==========================================================================
3938 */
3939int
3940zio_worst_error(int e1, int e2)
3941{
3942	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3943	int r1, r2;
3944
3945	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3946		if (e1 == zio_error_rank[r1])
3947			break;
3948
3949	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3950		if (e2 == zio_error_rank[r2])
3951			break;
3952
3953	return (r1 > r2 ? e1 : e2);
3954}
3955
3956/*
3957 * ==========================================================================
3958 * I/O completion
3959 * ==========================================================================
3960 */
3961static int
3962zio_ready(zio_t *zio)
3963{
3964	blkptr_t *bp = zio->io_bp;
3965	zio_t *pio, *pio_next;
3966	zio_link_t *zl = NULL;
3967
3968	if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
3969	    ZIO_WAIT_READY)) {
3970		return (ZIO_PIPELINE_STOP);
3971	}
3972
3973	if (zio->io_ready) {
3974		ASSERT(IO_IS_ALLOCATING(zio));
3975		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3976		    (zio->io_flags & ZIO_FLAG_NOPWRITE));
3977		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3978
3979		zio->io_ready(zio);
3980	}
3981
3982	if (bp != NULL && bp != &zio->io_bp_copy)
3983		zio->io_bp_copy = *bp;
3984
3985	if (zio->io_error != 0) {
3986		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3987
3988		if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3989			ASSERT(IO_IS_ALLOCATING(zio));
3990			ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3991			ASSERT(zio->io_metaslab_class != NULL);
3992
3993			/*
3994			 * We were unable to allocate anything, unreserve and
3995			 * issue the next I/O to allocate.
3996			 */
3997			metaslab_class_throttle_unreserve(
3998			    zio->io_metaslab_class, zio->io_prop.zp_copies,
3999			    zio->io_allocator, zio);
4000			zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
4001		}
4002	}
4003
4004	mutex_enter(&zio->io_lock);
4005	zio->io_state[ZIO_WAIT_READY] = 1;
4006	pio = zio_walk_parents(zio, &zl);
4007	mutex_exit(&zio->io_lock);
4008
4009	/*
4010	 * As we notify zio's parents, new parents could be added.
4011	 * New parents go to the head of zio's io_parent_list, however,
4012	 * so we will (correctly) not notify them.  The remainder of zio's
4013	 * io_parent_list, from 'pio_next' onward, cannot change because
4014	 * all parents must wait for us to be done before they can be done.
4015	 */
4016	for (; pio != NULL; pio = pio_next) {
4017		pio_next = zio_walk_parents(zio, &zl);
4018		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
4019	}
4020
4021	if (zio->io_flags & ZIO_FLAG_NODATA) {
4022		if (BP_IS_GANG(bp)) {
4023			zio->io_flags &= ~ZIO_FLAG_NODATA;
4024		} else {
4025			ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
4026			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
4027		}
4028	}
4029
4030	if (zio_injection_enabled &&
4031	    zio->io_spa->spa_syncing_txg == zio->io_txg)
4032		zio_handle_ignored_writes(zio);
4033
4034	return (ZIO_PIPELINE_CONTINUE);
4035}
4036
4037/*
4038 * Update the allocation throttle accounting.
4039 */
4040static void
4041zio_dva_throttle_done(zio_t *zio)
4042{
4043	zio_t *lio = zio->io_logical;
4044	zio_t *pio = zio_unique_parent(zio);
4045	vdev_t *vd = zio->io_vd;
4046	int flags = METASLAB_ASYNC_ALLOC;
4047
4048	ASSERT3P(zio->io_bp, !=, NULL);
4049	ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
4050	ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
4051	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
4052	ASSERT(vd != NULL);
4053	ASSERT3P(vd, ==, vd->vdev_top);
4054	ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
4055	ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
4056	ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
4057	ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
4058
4059	/*
4060	 * Parents of gang children can have two flavors -- ones that
4061	 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4062	 * and ones that allocated the constituent blocks. The allocation
4063	 * throttle needs to know the allocating parent zio so we must find
4064	 * it here.
4065	 */
4066	if (pio->io_child_type == ZIO_CHILD_GANG) {
4067		/*
4068		 * If our parent is a rewrite gang child then our grandparent
4069		 * would have been the one that performed the allocation.
4070		 */
4071		if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
4072			pio = zio_unique_parent(pio);
4073		flags |= METASLAB_GANG_CHILD;
4074	}
4075
4076	ASSERT(IO_IS_ALLOCATING(pio));
4077	ASSERT3P(zio, !=, zio->io_logical);
4078	ASSERT(zio->io_logical != NULL);
4079	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4080	ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
4081	ASSERT(zio->io_metaslab_class != NULL);
4082
4083	mutex_enter(&pio->io_lock);
4084	metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
4085	    pio->io_allocator, B_TRUE);
4086	mutex_exit(&pio->io_lock);
4087
4088	metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1,
4089	    pio->io_allocator, pio);
4090
4091	/*
4092	 * Call into the pipeline to see if there is more work that
4093	 * needs to be done. If there is work to be done it will be
4094	 * dispatched to another taskq thread.
4095	 */
4096	zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
4097}
4098
4099static int
4100zio_done(zio_t *zio)
4101{
4102	spa_t *spa = zio->io_spa;
4103	zio_t *lio = zio->io_logical;
4104	blkptr_t *bp = zio->io_bp;
4105	vdev_t *vd = zio->io_vd;
4106	uint64_t psize = zio->io_size;
4107	zio_t *pio, *pio_next;
4108	zio_link_t *zl = NULL;
4109
4110	/*
4111	 * If our children haven't all completed,
4112	 * wait for them and then repeat this pipeline stage.
4113	 */
4114	if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
4115		return (ZIO_PIPELINE_STOP);
4116	}
4117
4118	/*
4119	 * If the allocation throttle is enabled, then update the accounting.
4120	 * We only track child I/Os that are part of an allocating async
4121	 * write. We must do this since the allocation is performed
4122	 * by the logical I/O but the actual write is done by child I/Os.
4123	 */
4124	if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
4125	    zio->io_child_type == ZIO_CHILD_VDEV) {
4126		ASSERT(zio->io_metaslab_class != NULL);
4127		ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled);
4128		zio_dva_throttle_done(zio);
4129	}
4130
4131	/*
4132	 * If the allocation throttle is enabled, verify that
4133	 * we have decremented the refcounts for every I/O that was throttled.
4134	 */
4135	if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4136		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4137		ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4138		ASSERT(bp != NULL);
4139
4140		metaslab_group_alloc_verify(spa, zio->io_bp, zio,
4141		    zio->io_allocator);
4142		VERIFY(zfs_refcount_not_held(
4143		    &zio->io_metaslab_class->mc_alloc_slots[zio->io_allocator],
4144		    zio));
4145	}
4146
4147	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
4148		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
4149			ASSERT(zio->io_children[c][w] == 0);
4150
4151	if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
4152		ASSERT(bp->blk_pad[0] == 0);
4153		ASSERT(bp->blk_pad[1] == 0);
4154		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
4155		    (bp == zio_unique_parent(zio)->io_bp));
4156		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
4157		    zio->io_bp_override == NULL &&
4158		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
4159			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
4160			ASSERT(BP_COUNT_GANG(bp) == 0 ||
4161			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
4162		}
4163		if (zio->io_flags & ZIO_FLAG_NOPWRITE)
4164			VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
4165	}
4166
4167	/*
4168	 * If there were child vdev/gang/ddt errors, they apply to us now.
4169	 */
4170	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
4171	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
4172	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
4173
4174	/*
4175	 * If the I/O on the transformed data was successful, generate any
4176	 * checksum reports now while we still have the transformed data.
4177	 */
4178	if (zio->io_error == 0) {
4179		while (zio->io_cksum_report != NULL) {
4180			zio_cksum_report_t *zcr = zio->io_cksum_report;
4181			uint64_t align = zcr->zcr_align;
4182			uint64_t asize = P2ROUNDUP(psize, align);
4183			abd_t *adata = zio->io_abd;
4184
4185			if (asize != psize) {
4186				adata = abd_alloc(asize, B_TRUE);
4187				abd_copy(adata, zio->io_abd, psize);
4188				abd_zero_off(adata, psize, asize - psize);
4189			}
4190
4191			zio->io_cksum_report = zcr->zcr_next;
4192			zcr->zcr_next = NULL;
4193			zcr->zcr_finish(zcr, adata);
4194			zfs_ereport_free_checksum(zcr);
4195
4196			if (asize != psize)
4197				abd_free(adata);
4198		}
4199	}
4200
4201	zio_pop_transforms(zio);	/* note: may set zio->io_error */
4202
4203	vdev_stat_update(zio, psize);
4204
4205	if (zio->io_error) {
4206		/*
4207		 * If this I/O is attached to a particular vdev,
4208		 * generate an error message describing the I/O failure
4209		 * at the block level.  We ignore these errors if the
4210		 * device is currently unavailable.
4211		 */
4212		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
4213			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd,
4214			    &zio->io_bookmark, zio, 0, 0);
4215
4216		if ((zio->io_error == EIO || !(zio->io_flags &
4217		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
4218		    zio == lio) {
4219			/*
4220			 * For logical I/O requests, tell the SPA to log the
4221			 * error and generate a logical data ereport.
4222			 */
4223			spa_log_error(spa, &zio->io_bookmark);
4224			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL,
4225			    &zio->io_bookmark, zio, 0, 0);
4226		}
4227	}
4228
4229	if (zio->io_error && zio == lio) {
4230		/*
4231		 * Determine whether zio should be reexecuted.  This will
4232		 * propagate all the way to the root via zio_notify_parent().
4233		 */
4234		ASSERT(vd == NULL && bp != NULL);
4235		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4236
4237		if (IO_IS_ALLOCATING(zio) &&
4238		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
4239			if (zio->io_error != ENOSPC)
4240				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
4241			else
4242				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4243		}
4244
4245		if ((zio->io_type == ZIO_TYPE_READ ||
4246		    zio->io_type == ZIO_TYPE_FREE) &&
4247		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
4248		    zio->io_error == ENXIO &&
4249		    spa_load_state(spa) == SPA_LOAD_NONE &&
4250		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
4251			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4252
4253		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
4254			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4255
4256		/*
4257		 * Here is a possibly good place to attempt to do
4258		 * either combinatorial reconstruction or error correction
4259		 * based on checksums.  It also might be a good place
4260		 * to send out preliminary ereports before we suspend
4261		 * processing.
4262		 */
4263	}
4264
4265	/*
4266	 * If there were logical child errors, they apply to us now.
4267	 * We defer this until now to avoid conflating logical child
4268	 * errors with errors that happened to the zio itself when
4269	 * updating vdev stats and reporting FMA events above.
4270	 */
4271	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
4272
4273	if ((zio->io_error || zio->io_reexecute) &&
4274	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
4275	    !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
4276		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
4277
4278	zio_gang_tree_free(&zio->io_gang_tree);
4279
4280	/*
4281	 * Godfather I/Os should never suspend.
4282	 */
4283	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4284	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4285		zio->io_reexecute = 0;
4286
4287	if (zio->io_reexecute) {
4288		/*
4289		 * This is a logical I/O that wants to reexecute.
4290		 *
4291		 * Reexecute is top-down.  When an i/o fails, if it's not
4292		 * the root, it simply notifies its parent and sticks around.
4293		 * The parent, seeing that it still has children in zio_done(),
4294		 * does the same.  This percolates all the way up to the root.
4295		 * The root i/o will reexecute or suspend the entire tree.
4296		 *
4297		 * This approach ensures that zio_reexecute() honors
4298		 * all the original i/o dependency relationships, e.g.
4299		 * parents not executing until children are ready.
4300		 */
4301		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4302
4303		zio->io_gang_leader = NULL;
4304
4305		mutex_enter(&zio->io_lock);
4306		zio->io_state[ZIO_WAIT_DONE] = 1;
4307		mutex_exit(&zio->io_lock);
4308
4309		/*
4310		 * "The Godfather" I/O monitors its children but is
4311		 * not a true parent to them. It will track them through
4312		 * the pipeline but severs its ties whenever they get into
4313		 * trouble (e.g. suspended). This allows "The Godfather"
4314		 * I/O to return status without blocking.
4315		 */
4316		zl = NULL;
4317		for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4318		    pio = pio_next) {
4319			zio_link_t *remove_zl = zl;
4320			pio_next = zio_walk_parents(zio, &zl);
4321
4322			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4323			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4324				zio_remove_child(pio, zio, remove_zl);
4325				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4326			}
4327		}
4328
4329		if ((pio = zio_unique_parent(zio)) != NULL) {
4330			/*
4331			 * We're not a root i/o, so there's nothing to do
4332			 * but notify our parent.  Don't propagate errors
4333			 * upward since we haven't permanently failed yet.
4334			 */
4335			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4336			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4337			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4338		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4339			/*
4340			 * We'd fail again if we reexecuted now, so suspend
4341			 * until conditions improve (e.g. device comes online).
4342			 */
4343			zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
4344		} else {
4345			/*
4346			 * Reexecution is potentially a huge amount of work.
4347			 * Hand it off to the otherwise-unused claim taskq.
4348			 */
4349			ASSERT(zio->io_tqent.tqent_next == NULL);
4350			spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
4351			    ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
4352			    0, &zio->io_tqent);
4353		}
4354		return (ZIO_PIPELINE_STOP);
4355	}
4356
4357	ASSERT(zio->io_child_count == 0);
4358	ASSERT(zio->io_reexecute == 0);
4359	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4360
4361	/*
4362	 * Report any checksum errors, since the I/O is complete.
4363	 */
4364	while (zio->io_cksum_report != NULL) {
4365		zio_cksum_report_t *zcr = zio->io_cksum_report;
4366		zio->io_cksum_report = zcr->zcr_next;
4367		zcr->zcr_next = NULL;
4368		zcr->zcr_finish(zcr, NULL);
4369		zfs_ereport_free_checksum(zcr);
4370	}
4371
4372	/*
4373	 * It is the responsibility of the done callback to ensure that this
4374	 * particular zio is no longer discoverable for adoption, and as
4375	 * such, cannot acquire any new parents.
4376	 */
4377	if (zio->io_done)
4378		zio->io_done(zio);
4379
4380	mutex_enter(&zio->io_lock);
4381	zio->io_state[ZIO_WAIT_DONE] = 1;
4382	mutex_exit(&zio->io_lock);
4383
4384	zl = NULL;
4385	for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4386		zio_link_t *remove_zl = zl;
4387		pio_next = zio_walk_parents(zio, &zl);
4388		zio_remove_child(pio, zio, remove_zl);
4389		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4390	}
4391
4392	if (zio->io_waiter != NULL) {
4393		mutex_enter(&zio->io_lock);
4394		zio->io_executor = NULL;
4395		cv_broadcast(&zio->io_cv);
4396		mutex_exit(&zio->io_lock);
4397	} else {
4398		zio_destroy(zio);
4399	}
4400
4401	return (ZIO_PIPELINE_STOP);
4402}
4403
4404/*
4405 * ==========================================================================
4406 * I/O pipeline definition
4407 * ==========================================================================
4408 */
4409static zio_pipe_stage_t *zio_pipeline[] = {
4410	NULL,
4411	zio_read_bp_init,
4412	zio_write_bp_init,
4413	zio_free_bp_init,
4414	zio_issue_async,
4415	zio_write_compress,
4416	zio_encrypt,
4417	zio_checksum_generate,
4418	zio_nop_write,
4419	zio_ddt_read_start,
4420	zio_ddt_read_done,
4421	zio_ddt_write,
4422	zio_ddt_free,
4423	zio_gang_assemble,
4424	zio_gang_issue,
4425	zio_dva_throttle,
4426	zio_dva_allocate,
4427	zio_dva_free,
4428	zio_dva_claim,
4429	zio_ready,
4430	zio_vdev_io_start,
4431	zio_vdev_io_done,
4432	zio_vdev_io_assess,
4433	zio_checksum_verify,
4434	zio_done
4435};
4436
4437
4438
4439
4440/*
4441 * Compare two zbookmark_phys_t's to see which we would reach first in a
4442 * pre-order traversal of the object tree.
4443 *
4444 * This is simple in every case aside from the meta-dnode object. For all other
4445 * objects, we traverse them in order (object 1 before object 2, and so on).
4446 * However, all of these objects are traversed while traversing object 0, since
4447 * the data it points to is the list of objects.  Thus, we need to convert to a
4448 * canonical representation so we can compare meta-dnode bookmarks to
4449 * non-meta-dnode bookmarks.
4450 *
4451 * We do this by calculating "equivalents" for each field of the zbookmark.
4452 * zbookmarks outside of the meta-dnode use their own object and level, and
4453 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4454 * blocks this bookmark refers to) by multiplying their blkid by their span
4455 * (the number of L0 blocks contained within one block at their level).
4456 * zbookmarks inside the meta-dnode calculate their object equivalent
4457 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4458 * level + 1<<31 (any value larger than a level could ever be) for their level.
4459 * This causes them to always compare before a bookmark in their object
4460 * equivalent, compare appropriately to bookmarks in other objects, and to
4461 * compare appropriately to other bookmarks in the meta-dnode.
4462 */
4463int
4464zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4465    const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4466{
4467	/*
4468	 * These variables represent the "equivalent" values for the zbookmark,
4469	 * after converting zbookmarks inside the meta dnode to their
4470	 * normal-object equivalents.
4471	 */
4472	uint64_t zb1obj, zb2obj;
4473	uint64_t zb1L0, zb2L0;
4474	uint64_t zb1level, zb2level;
4475
4476	if (zb1->zb_object == zb2->zb_object &&
4477	    zb1->zb_level == zb2->zb_level &&
4478	    zb1->zb_blkid == zb2->zb_blkid)
4479		return (0);
4480
4481	/*
4482	 * BP_SPANB calculates the span in blocks.
4483	 */
4484	zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4485	zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4486
4487	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4488		zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4489		zb1L0 = 0;
4490		zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4491	} else {
4492		zb1obj = zb1->zb_object;
4493		zb1level = zb1->zb_level;
4494	}
4495
4496	if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4497		zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4498		zb2L0 = 0;
4499		zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4500	} else {
4501		zb2obj = zb2->zb_object;
4502		zb2level = zb2->zb_level;
4503	}
4504
4505	/* Now that we have a canonical representation, do the comparison. */
4506	if (zb1obj != zb2obj)
4507		return (zb1obj < zb2obj ? -1 : 1);
4508	else if (zb1L0 != zb2L0)
4509		return (zb1L0 < zb2L0 ? -1 : 1);
4510	else if (zb1level != zb2level)
4511		return (zb1level > zb2level ? -1 : 1);
4512	/*
4513	 * This can (theoretically) happen if the bookmarks have the same object
4514	 * and level, but different blkids, if the block sizes are not the same.
4515	 * There is presently no way to change the indirect block sizes
4516	 */
4517	return (0);
4518}
4519
4520/*
4521 *  This function checks the following: given that last_block is the place that
4522 *  our traversal stopped last time, does that guarantee that we've visited
4523 *  every node under subtree_root?  Therefore, we can't just use the raw output
4524 *  of zbookmark_compare.  We have to pass in a modified version of
4525 *  subtree_root; by incrementing the block id, and then checking whether
4526 *  last_block is before or equal to that, we can tell whether or not having
4527 *  visited last_block implies that all of subtree_root's children have been
4528 *  visited.
4529 */
4530boolean_t
4531zbookmark_subtree_completed(const dnode_phys_t *dnp,
4532    const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4533{
4534	zbookmark_phys_t mod_zb = *subtree_root;
4535	mod_zb.zb_blkid++;
4536	ASSERT(last_block->zb_level == 0);
4537
4538	/* The objset_phys_t isn't before anything. */
4539	if (dnp == NULL)
4540		return (B_FALSE);
4541
4542	/*
4543	 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4544	 * data block size in sectors, because that variable is only used if
4545	 * the bookmark refers to a block in the meta-dnode.  Since we don't
4546	 * know without examining it what object it refers to, and there's no
4547	 * harm in passing in this value in other cases, we always pass it in.
4548	 *
4549	 * We pass in 0 for the indirect block size shift because zb2 must be
4550	 * level 0.  The indirect block size is only used to calculate the span
4551	 * of the bookmark, but since the bookmark must be level 0, the span is
4552	 * always 1, so the math works out.
4553	 *
4554	 * If you make changes to how the zbookmark_compare code works, be sure
4555	 * to make sure that this code still works afterwards.
4556	 */
4557	return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4558	    1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
4559	    last_block) <= 0);
4560}
4561