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 * Copyright 2020 Joyent, Inc.
28 */
29
30#include <sys/sysmacros.h>
31#include <sys/zfs_context.h>
32#include <sys/fm/fs/zfs.h>
33#include <sys/spa.h>
34#include <sys/txg.h>
35#include <sys/spa_impl.h>
36#include <sys/vdev_impl.h>
37#include <sys/vdev_trim.h>
38#include <sys/zio_impl.h>
39#include <sys/zio_compress.h>
40#include <sys/zio_checksum.h>
41#include <sys/dmu_objset.h>
42#include <sys/arc.h>
43#include <sys/ddt.h>
44#include <sys/blkptr.h>
45#include <sys/zfeature.h>
46#include <sys/time.h>
47#include <sys/dsl_scan.h>
48#include <sys/metaslab_impl.h>
49#include <sys/abd.h>
50#include <sys/cityhash.h>
51#include <sys/dsl_crypt.h>
52
53/*
54 * ==========================================================================
55 * I/O type descriptions
56 * ==========================================================================
57 */
58const char *zio_type_name[ZIO_TYPES] = {
59	"zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
60	"zio_ioctl", "z_trim"
61};
62
63boolean_t zio_dva_throttle_enabled = B_TRUE;
64
65/*
66 * ==========================================================================
67 * I/O kmem caches
68 * ==========================================================================
69 */
70kmem_cache_t *zio_cache;
71kmem_cache_t *zio_link_cache;
72kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
73kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
74
75#ifdef _KERNEL
76extern vmem_t *zio_alloc_arena;
77#endif
78
79#define	ZIO_PIPELINE_CONTINUE		0x100
80#define	ZIO_PIPELINE_STOP		0x101
81
82/* Mark IOs as "slow" if they take longer than 30 seconds */
83int zio_slow_io_ms = (30 * MILLISEC);
84
85#define	BP_SPANB(indblkshift, level) \
86	(((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
87#define	COMPARE_META_LEVEL	0x80000000ul
88/*
89 * The following actions directly effect the spa's sync-to-convergence logic.
90 * The values below define the sync pass when we start performing the action.
91 * Care should be taken when changing these values as they directly impact
92 * spa_sync() performance. Tuning these values may introduce subtle performance
93 * pathologies and should only be done in the context of performance analysis.
94 * These tunables will eventually be removed and replaced with #defines once
95 * enough analysis has been done to determine optimal values.
96 *
97 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
98 * regular blocks are not deferred.
99 */
100int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
101int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
102int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
103
104/*
105 * An allocating zio is one that either currently has the DVA allocate
106 * stage set or will have it later in its lifetime.
107 */
108#define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
109
110boolean_t	zio_requeue_io_start_cut_in_line = B_TRUE;
111
112#ifdef ZFS_DEBUG
113int zio_buf_debug_limit = 16384;
114#else
115int zio_buf_debug_limit = 0;
116#endif
117
118static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
119
120void
121zio_init(void)
122{
123	size_t c;
124	vmem_t *data_alloc_arena = NULL;
125
126#ifdef _KERNEL
127	data_alloc_arena = zio_alloc_arena;
128#endif
129	zio_cache = kmem_cache_create("zio_cache",
130	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
131	zio_link_cache = kmem_cache_create("zio_link_cache",
132	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
133
134	/*
135	 * For small buffers, we want a cache for each multiple of
136	 * SPA_MINBLOCKSIZE.  For larger buffers, we want a cache
137	 * for each quarter-power of 2.
138	 */
139	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
140		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
141		size_t p2 = size;
142		size_t align = 0;
143		size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
144
145		while (!ISP2(p2))
146			p2 &= p2 - 1;
147
148#ifndef _KERNEL
149		/*
150		 * If we are using watchpoints, put each buffer on its own page,
151		 * to eliminate the performance overhead of trapping to the
152		 * kernel when modifying a non-watched buffer that shares the
153		 * page with a watched buffer.
154		 */
155		if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
156			continue;
157#endif
158		if (size <= 4 * SPA_MINBLOCKSIZE) {
159			align = SPA_MINBLOCKSIZE;
160		} else if (IS_P2ALIGNED(size, p2 >> 2)) {
161			align = MIN(p2 >> 2, PAGESIZE);
162		}
163
164		if (align != 0) {
165			char name[36];
166			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
167			zio_buf_cache[c] = kmem_cache_create(name, size,
168			    align, NULL, NULL, NULL, NULL, NULL, cflags);
169
170			/*
171			 * Since zio_data bufs do not appear in crash dumps, we
172			 * pass KMC_NOTOUCH so that no allocator metadata is
173			 * stored with the buffers.
174			 */
175			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
176			zio_data_buf_cache[c] = kmem_cache_create(name, size,
177			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
178			    cflags | KMC_NOTOUCH);
179		}
180	}
181
182	while (--c != 0) {
183		ASSERT(zio_buf_cache[c] != NULL);
184		if (zio_buf_cache[c - 1] == NULL)
185			zio_buf_cache[c - 1] = zio_buf_cache[c];
186
187		ASSERT(zio_data_buf_cache[c] != NULL);
188		if (zio_data_buf_cache[c - 1] == NULL)
189			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
190	}
191
192	zio_inject_init();
193}
194
195void
196zio_fini(void)
197{
198	size_t c;
199	kmem_cache_t *last_cache = NULL;
200	kmem_cache_t *last_data_cache = NULL;
201
202	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
203		if (zio_buf_cache[c] != last_cache) {
204			last_cache = zio_buf_cache[c];
205			kmem_cache_destroy(zio_buf_cache[c]);
206		}
207		zio_buf_cache[c] = NULL;
208
209		if (zio_data_buf_cache[c] != last_data_cache) {
210			last_data_cache = zio_data_buf_cache[c];
211			kmem_cache_destroy(zio_data_buf_cache[c]);
212		}
213		zio_data_buf_cache[c] = NULL;
214	}
215
216	kmem_cache_destroy(zio_link_cache);
217	kmem_cache_destroy(zio_cache);
218
219	zio_inject_fini();
220}
221
222/*
223 * ==========================================================================
224 * Allocate and free I/O buffers
225 * ==========================================================================
226 */
227
228/*
229 * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
230 * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
231 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
232 * excess / transient data in-core during a crashdump.
233 */
234void *
235zio_buf_alloc(size_t size)
236{
237	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
238
239	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
240
241	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
242}
243
244/*
245 * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
246 * crashdump if the kernel panics.  This exists so that we will limit the amount
247 * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
248 * of kernel heap dumped to disk when the kernel panics)
249 */
250void *
251zio_data_buf_alloc(size_t size)
252{
253	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
254
255	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
256
257	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
258}
259
260void
261zio_buf_free(void *buf, size_t size)
262{
263	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
264
265	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
266
267	kmem_cache_free(zio_buf_cache[c], buf);
268}
269
270void
271zio_data_buf_free(void *buf, size_t size)
272{
273	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
274
275	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
276
277	kmem_cache_free(zio_data_buf_cache[c], buf);
278}
279
280/* ARGSUSED */
281static void
282zio_abd_free(void *abd, size_t size)
283{
284	abd_free((abd_t *)abd);
285}
286
287/*
288 * ==========================================================================
289 * Push and pop I/O transform buffers
290 * ==========================================================================
291 */
292void
293zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
294    zio_transform_func_t *transform)
295{
296	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
297
298	/*
299	 * Ensure that anyone expecting this zio to contain a linear ABD isn't
300	 * going to get a nasty surprise when they try to access the data.
301	 */
302	IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
303
304	zt->zt_orig_abd = zio->io_abd;
305	zt->zt_orig_size = zio->io_size;
306	zt->zt_bufsize = bufsize;
307	zt->zt_transform = transform;
308
309	zt->zt_next = zio->io_transform_stack;
310	zio->io_transform_stack = zt;
311
312	zio->io_abd = data;
313	zio->io_size = size;
314}
315
316void
317zio_pop_transforms(zio_t *zio)
318{
319	zio_transform_t *zt;
320
321	while ((zt = zio->io_transform_stack) != NULL) {
322		if (zt->zt_transform != NULL)
323			zt->zt_transform(zio,
324			    zt->zt_orig_abd, zt->zt_orig_size);
325
326		if (zt->zt_bufsize != 0)
327			abd_free(zio->io_abd);
328
329		zio->io_abd = zt->zt_orig_abd;
330		zio->io_size = zt->zt_orig_size;
331		zio->io_transform_stack = zt->zt_next;
332
333		kmem_free(zt, sizeof (zio_transform_t));
334	}
335}
336
337/*
338 * ==========================================================================
339 * I/O transform callbacks for subblocks, decompression, and decryption
340 * ==========================================================================
341 */
342static void
343zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
344{
345	ASSERT(zio->io_size > size);
346
347	if (zio->io_type == ZIO_TYPE_READ)
348		abd_copy(data, zio->io_abd, size);
349}
350
351static void
352zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
353{
354	if (zio->io_error == 0) {
355		void *tmp = abd_borrow_buf(data, size);
356		int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
357		    zio->io_abd, tmp, zio->io_size, size);
358		abd_return_buf_copy(data, tmp, size);
359
360		if (ret != 0)
361			zio->io_error = SET_ERROR(EIO);
362	}
363}
364
365static void
366zio_decrypt(zio_t *zio, abd_t *data, uint64_t size)
367{
368	int ret;
369	void *tmp;
370	blkptr_t *bp = zio->io_bp;
371	spa_t *spa = zio->io_spa;
372	uint64_t dsobj = zio->io_bookmark.zb_objset;
373	uint64_t lsize = BP_GET_LSIZE(bp);
374	dmu_object_type_t ot = BP_GET_TYPE(bp);
375	uint8_t salt[ZIO_DATA_SALT_LEN];
376	uint8_t iv[ZIO_DATA_IV_LEN];
377	uint8_t mac[ZIO_DATA_MAC_LEN];
378	boolean_t no_crypt = B_FALSE;
379
380	ASSERT(BP_USES_CRYPT(bp));
381	ASSERT3U(size, !=, 0);
382
383	if (zio->io_error != 0)
384		return;
385
386	/*
387	 * Verify the cksum of MACs stored in an indirect bp. It will always
388	 * be possible to verify this since it does not require an encryption
389	 * key.
390	 */
391	if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {
392		zio_crypt_decode_mac_bp(bp, mac);
393
394		if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
395			/*
396			 * We haven't decompressed the data yet, but
397			 * zio_crypt_do_indirect_mac_checksum() requires
398			 * decompressed data to be able to parse out the MACs
399			 * from the indirect block. We decompress it now and
400			 * throw away the result after we are finished.
401			 */
402			tmp = zio_buf_alloc(lsize);
403			ret = zio_decompress_data(BP_GET_COMPRESS(bp),
404			    zio->io_abd, tmp, zio->io_size, lsize);
405			if (ret != 0) {
406				ret = SET_ERROR(EIO);
407				goto error;
408			}
409			ret = zio_crypt_do_indirect_mac_checksum(B_FALSE,
410			    tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac);
411			zio_buf_free(tmp, lsize);
412		} else {
413			ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
414			    zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac);
415		}
416		abd_copy(data, zio->io_abd, size);
417
418		if (ret != 0)
419			goto error;
420
421		return;
422	}
423
424	/*
425	 * If this is an authenticated block, just check the MAC. It would be
426	 * nice to separate this out into its own flag, but for the moment
427	 * enum zio_flag is out of bits.
428	 */
429	if (BP_IS_AUTHENTICATED(bp)) {
430		if (ot == DMU_OT_OBJSET) {
431			ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa,
432			    dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp));
433		} else {
434			zio_crypt_decode_mac_bp(bp, mac);
435			ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj,
436			    zio->io_abd, size, mac);
437		}
438		abd_copy(data, zio->io_abd, size);
439
440		if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) {
441			ret = zio_handle_decrypt_injection(spa,
442			    &zio->io_bookmark, ot, ECKSUM);
443		}
444		if (ret != 0)
445			goto error;
446
447		return;
448	}
449
450	zio_crypt_decode_params_bp(bp, salt, iv);
451
452	if (ot == DMU_OT_INTENT_LOG) {
453		tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t));
454		zio_crypt_decode_mac_zil(tmp, mac);
455		abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t));
456	} else {
457		zio_crypt_decode_mac_bp(bp, mac);
458	}
459
460	ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp),
461	    BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data,
462	    zio->io_abd, &no_crypt);
463	if (no_crypt)
464		abd_copy(data, zio->io_abd, size);
465
466	if (ret != 0)
467		goto error;
468
469	return;
470
471error:
472	/* assert that the key was found unless this was speculative */
473	ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE));
474
475	/*
476	 * If there was a decryption / authentication error return EIO as
477	 * the io_error. If this was not a speculative zio, create an ereport.
478	 */
479	if (ret == ECKSUM) {
480		zio->io_error = SET_ERROR(EIO);
481		if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
482			spa_log_error(spa, &zio->io_bookmark);
483			(void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
484			    spa, NULL, &zio->io_bookmark, zio, 0, 0);
485		}
486	} else {
487		zio->io_error = ret;
488	}
489}
490
491/*
492 * ==========================================================================
493 * I/O parent/child relationships and pipeline interlocks
494 * ==========================================================================
495 */
496zio_t *
497zio_walk_parents(zio_t *cio, zio_link_t **zl)
498{
499	list_t *pl = &cio->io_parent_list;
500
501	*zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
502	if (*zl == NULL)
503		return (NULL);
504
505	ASSERT((*zl)->zl_child == cio);
506	return ((*zl)->zl_parent);
507}
508
509zio_t *
510zio_walk_children(zio_t *pio, zio_link_t **zl)
511{
512	list_t *cl = &pio->io_child_list;
513
514	ASSERT(MUTEX_HELD(&pio->io_lock));
515
516	*zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
517	if (*zl == NULL)
518		return (NULL);
519
520	ASSERT((*zl)->zl_parent == pio);
521	return ((*zl)->zl_child);
522}
523
524zio_t *
525zio_unique_parent(zio_t *cio)
526{
527	zio_link_t *zl = NULL;
528	zio_t *pio = zio_walk_parents(cio, &zl);
529
530	VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
531	return (pio);
532}
533
534void
535zio_add_child(zio_t *pio, zio_t *cio)
536{
537	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
538
539	/*
540	 * Logical I/Os can have logical, gang, or vdev children.
541	 * Gang I/Os can have gang or vdev children.
542	 * Vdev I/Os can only have vdev children.
543	 * The following ASSERT captures all of these constraints.
544	 */
545	ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
546
547	zl->zl_parent = pio;
548	zl->zl_child = cio;
549
550	mutex_enter(&pio->io_lock);
551	mutex_enter(&cio->io_lock);
552
553	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
554
555	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
556		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
557
558	list_insert_head(&pio->io_child_list, zl);
559	list_insert_head(&cio->io_parent_list, zl);
560
561	pio->io_child_count++;
562	cio->io_parent_count++;
563
564	mutex_exit(&cio->io_lock);
565	mutex_exit(&pio->io_lock);
566}
567
568static void
569zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
570{
571	ASSERT(zl->zl_parent == pio);
572	ASSERT(zl->zl_child == cio);
573
574	mutex_enter(&pio->io_lock);
575	mutex_enter(&cio->io_lock);
576
577	list_remove(&pio->io_child_list, zl);
578	list_remove(&cio->io_parent_list, zl);
579
580	pio->io_child_count--;
581	cio->io_parent_count--;
582
583	mutex_exit(&cio->io_lock);
584	mutex_exit(&pio->io_lock);
585
586	kmem_cache_free(zio_link_cache, zl);
587}
588
589static boolean_t
590zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
591{
592	boolean_t waiting = B_FALSE;
593
594	mutex_enter(&zio->io_lock);
595	ASSERT(zio->io_stall == NULL);
596	for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
597		if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
598			continue;
599
600		uint64_t *countp = &zio->io_children[c][wait];
601		if (*countp != 0) {
602			zio->io_stage >>= 1;
603			ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
604			zio->io_stall = countp;
605			waiting = B_TRUE;
606			break;
607		}
608	}
609	mutex_exit(&zio->io_lock);
610	return (waiting);
611}
612
613static void
614zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
615{
616	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
617	int *errorp = &pio->io_child_error[zio->io_child_type];
618
619	mutex_enter(&pio->io_lock);
620	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
621		*errorp = zio_worst_error(*errorp, zio->io_error);
622	pio->io_reexecute |= zio->io_reexecute;
623	ASSERT3U(*countp, >, 0);
624
625	(*countp)--;
626
627	if (*countp == 0 && pio->io_stall == countp) {
628		zio_taskq_type_t type =
629		    pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
630		    ZIO_TASKQ_INTERRUPT;
631		pio->io_stall = NULL;
632		mutex_exit(&pio->io_lock);
633		/*
634		 * Dispatch the parent zio in its own taskq so that
635		 * the child can continue to make progress. This also
636		 * prevents overflowing the stack when we have deeply nested
637		 * parent-child relationships.
638		 */
639		zio_taskq_dispatch(pio, type, B_FALSE);
640	} else {
641		mutex_exit(&pio->io_lock);
642	}
643}
644
645static void
646zio_inherit_child_errors(zio_t *zio, enum zio_child c)
647{
648	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
649		zio->io_error = zio->io_child_error[c];
650}
651
652int
653zio_bookmark_compare(const void *x1, const void *x2)
654{
655	const zio_t *z1 = x1;
656	const zio_t *z2 = x2;
657
658	if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
659		return (-1);
660	if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
661		return (1);
662
663	if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
664		return (-1);
665	if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
666		return (1);
667
668	if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
669		return (-1);
670	if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
671		return (1);
672
673	if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
674		return (-1);
675	if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
676		return (1);
677
678	if (z1 < z2)
679		return (-1);
680	if (z1 > z2)
681		return (1);
682
683	return (0);
684}
685
686/*
687 * ==========================================================================
688 * Create the various types of I/O (read, write, free, etc)
689 * ==========================================================================
690 */
691static zio_t *
692zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
693    abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
694    void *private, zio_type_t type, zio_priority_t priority,
695    enum zio_flag flags, vdev_t *vd, uint64_t offset,
696    const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
697{
698	zio_t *zio;
699
700	IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE);
701	ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
702	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
703
704	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
705	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
706	ASSERT(vd || stage == ZIO_STAGE_OPEN);
707
708	IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0);
709
710	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
711	bzero(zio, sizeof (zio_t));
712
713	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
714	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
715
716	list_create(&zio->io_parent_list, sizeof (zio_link_t),
717	    offsetof(zio_link_t, zl_parent_node));
718	list_create(&zio->io_child_list, sizeof (zio_link_t),
719	    offsetof(zio_link_t, zl_child_node));
720	metaslab_trace_init(&zio->io_alloc_list);
721
722	if (vd != NULL)
723		zio->io_child_type = ZIO_CHILD_VDEV;
724	else if (flags & ZIO_FLAG_GANG_CHILD)
725		zio->io_child_type = ZIO_CHILD_GANG;
726	else if (flags & ZIO_FLAG_DDT_CHILD)
727		zio->io_child_type = ZIO_CHILD_DDT;
728	else
729		zio->io_child_type = ZIO_CHILD_LOGICAL;
730
731	if (bp != NULL) {
732		zio->io_bp = (blkptr_t *)bp;
733		zio->io_bp_copy = *bp;
734		zio->io_bp_orig = *bp;
735		if (type != ZIO_TYPE_WRITE ||
736		    zio->io_child_type == ZIO_CHILD_DDT)
737			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
738		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
739			zio->io_logical = zio;
740		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
741			pipeline |= ZIO_GANG_STAGES;
742	}
743
744	zio->io_spa = spa;
745	zio->io_txg = txg;
746	zio->io_done = done;
747	zio->io_private = private;
748	zio->io_type = type;
749	zio->io_priority = priority;
750	zio->io_vd = vd;
751	zio->io_offset = offset;
752	zio->io_orig_abd = zio->io_abd = data;
753	zio->io_orig_size = zio->io_size = psize;
754	zio->io_lsize = lsize;
755	zio->io_orig_flags = zio->io_flags = flags;
756	zio->io_orig_stage = zio->io_stage = stage;
757	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
758	zio->io_pipeline_trace = ZIO_STAGE_OPEN;
759
760	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
761	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
762
763	if (zb != NULL)
764		zio->io_bookmark = *zb;
765
766	if (pio != NULL) {
767		if (zio->io_metaslab_class == NULL)
768			zio->io_metaslab_class = pio->io_metaslab_class;
769		if (zio->io_logical == NULL)
770			zio->io_logical = pio->io_logical;
771		if (zio->io_child_type == ZIO_CHILD_GANG)
772			zio->io_gang_leader = pio->io_gang_leader;
773		zio_add_child(pio, zio);
774	}
775
776	return (zio);
777}
778
779static void
780zio_destroy(zio_t *zio)
781{
782	metaslab_trace_fini(&zio->io_alloc_list);
783	list_destroy(&zio->io_parent_list);
784	list_destroy(&zio->io_child_list);
785	mutex_destroy(&zio->io_lock);
786	cv_destroy(&zio->io_cv);
787	kmem_cache_free(zio_cache, zio);
788}
789
790zio_t *
791zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
792    void *private, enum zio_flag flags)
793{
794	zio_t *zio;
795
796	zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
797	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
798	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
799
800	return (zio);
801}
802
803zio_t *
804zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
805{
806	return (zio_null(NULL, spa, NULL, done, private, flags));
807}
808
809void
810zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
811{
812	if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
813		zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
814		    bp, (longlong_t)BP_GET_TYPE(bp));
815	}
816	if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
817	    BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
818		zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
819		    bp, (longlong_t)BP_GET_CHECKSUM(bp));
820	}
821	if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
822	    BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
823		zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
824		    bp, (longlong_t)BP_GET_COMPRESS(bp));
825	}
826	if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
827		zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
828		    bp, (longlong_t)BP_GET_LSIZE(bp));
829	}
830	if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
831		zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
832		    bp, (longlong_t)BP_GET_PSIZE(bp));
833	}
834
835	if (BP_IS_EMBEDDED(bp)) {
836		if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
837			zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
838			    bp, (longlong_t)BPE_GET_ETYPE(bp));
839		}
840	}
841
842	/*
843	 * Do not verify individual DVAs if the config is not trusted. This
844	 * will be done once the zio is executed in vdev_mirror_map_alloc.
845	 */
846	if (!spa->spa_trust_config)
847		return;
848
849	/*
850	 * Pool-specific checks.
851	 *
852	 * Note: it would be nice to verify that the blk_birth and
853	 * BP_PHYSICAL_BIRTH() are not too large.  However, spa_freeze()
854	 * allows the birth time of log blocks (and dmu_sync()-ed blocks
855	 * that are in the log) to be arbitrarily large.
856	 */
857	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
858		uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
859		if (vdevid >= spa->spa_root_vdev->vdev_children) {
860			zfs_panic_recover("blkptr at %p DVA %u has invalid "
861			    "VDEV %llu",
862			    bp, i, (longlong_t)vdevid);
863			continue;
864		}
865		vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
866		if (vd == NULL) {
867			zfs_panic_recover("blkptr at %p DVA %u has invalid "
868			    "VDEV %llu",
869			    bp, i, (longlong_t)vdevid);
870			continue;
871		}
872		if (vd->vdev_ops == &vdev_hole_ops) {
873			zfs_panic_recover("blkptr at %p DVA %u has hole "
874			    "VDEV %llu",
875			    bp, i, (longlong_t)vdevid);
876			continue;
877		}
878		if (vd->vdev_ops == &vdev_missing_ops) {
879			/*
880			 * "missing" vdevs are valid during import, but we
881			 * don't have their detailed info (e.g. asize), so
882			 * we can't perform any more checks on them.
883			 */
884			continue;
885		}
886		uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
887		uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
888		if (BP_IS_GANG(bp))
889			asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
890		if (offset + asize > vd->vdev_asize) {
891			zfs_panic_recover("blkptr at %p DVA %u has invalid "
892			    "OFFSET %llu",
893			    bp, i, (longlong_t)offset);
894		}
895	}
896}
897
898boolean_t
899zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
900{
901	uint64_t vdevid = DVA_GET_VDEV(dva);
902
903	if (vdevid >= spa->spa_root_vdev->vdev_children)
904		return (B_FALSE);
905
906	vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
907	if (vd == NULL)
908		return (B_FALSE);
909
910	if (vd->vdev_ops == &vdev_hole_ops)
911		return (B_FALSE);
912
913	if (vd->vdev_ops == &vdev_missing_ops) {
914		return (B_FALSE);
915	}
916
917	uint64_t offset = DVA_GET_OFFSET(dva);
918	uint64_t asize = DVA_GET_ASIZE(dva);
919
920	if (BP_IS_GANG(bp))
921		asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
922	if (offset + asize > vd->vdev_asize)
923		return (B_FALSE);
924
925	return (B_TRUE);
926}
927
928zio_t *
929zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
930    abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
931    zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
932{
933	zio_t *zio;
934
935	zfs_blkptr_verify(spa, bp);
936
937	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
938	    data, size, size, done, private,
939	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
940	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
941	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
942
943	return (zio);
944}
945
946zio_t *
947zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
948    abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
949    zio_done_func_t *ready, zio_done_func_t *children_ready,
950    zio_done_func_t *physdone, zio_done_func_t *done,
951    void *private, zio_priority_t priority, enum zio_flag flags,
952    const zbookmark_phys_t *zb)
953{
954	zio_t *zio;
955
956	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
957	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
958	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
959	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
960	    DMU_OT_IS_VALID(zp->zp_type) &&
961	    zp->zp_level < 32 &&
962	    zp->zp_copies > 0 &&
963	    zp->zp_copies <= spa_max_replication(spa));
964
965	zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
966	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
967	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
968	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
969
970	zio->io_ready = ready;
971	zio->io_children_ready = children_ready;
972	zio->io_physdone = physdone;
973	zio->io_prop = *zp;
974
975	/*
976	 * Data can be NULL if we are going to call zio_write_override() to
977	 * provide the already-allocated BP.  But we may need the data to
978	 * verify a dedup hit (if requested).  In this case, don't try to
979	 * dedup (just take the already-allocated BP verbatim). Encrypted
980	 * dedup blocks need data as well so we also disable dedup in this
981	 * case.
982	 */
983	if (data == NULL &&
984	    (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) {
985		zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
986	}
987
988	return (zio);
989}
990
991zio_t *
992zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
993    uint64_t size, zio_done_func_t *done, void *private,
994    zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
995{
996	zio_t *zio;
997
998	zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
999	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
1000	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
1001
1002	return (zio);
1003}
1004
1005void
1006zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
1007{
1008	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1009	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1010	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1011	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
1012
1013	/*
1014	 * We must reset the io_prop to match the values that existed
1015	 * when the bp was first written by dmu_sync() keeping in mind
1016	 * that nopwrite and dedup are mutually exclusive.
1017	 */
1018	zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
1019	zio->io_prop.zp_nopwrite = nopwrite;
1020	zio->io_prop.zp_copies = copies;
1021	zio->io_bp_override = bp;
1022}
1023
1024void
1025zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
1026{
1027
1028	zfs_blkptr_verify(spa, bp);
1029
1030	/*
1031	 * The check for EMBEDDED is a performance optimization.  We
1032	 * process the free here (by ignoring it) rather than
1033	 * putting it on the list and then processing it in zio_free_sync().
1034	 */
1035	if (BP_IS_EMBEDDED(bp))
1036		return;
1037	metaslab_check_free(spa, bp);
1038
1039	/*
1040	 * Frees that are for the currently-syncing txg, are not going to be
1041	 * deferred, and which will not need to do a read (i.e. not GANG or
1042	 * DEDUP), can be processed immediately.  Otherwise, put them on the
1043	 * in-memory list for later processing.
1044	 *
1045	 * Note that we only defer frees after zfs_sync_pass_deferred_free
1046	 * when the log space map feature is disabled. [see relevant comment
1047	 * in spa_sync_iterate_to_convergence()]
1048	 */
1049	if (BP_IS_GANG(bp) ||
1050	    BP_GET_DEDUP(bp) ||
1051	    txg != spa->spa_syncing_txg ||
1052	    (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free &&
1053	    !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))) {
1054		bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1055	} else {
1056		VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
1057	}
1058}
1059
1060zio_t *
1061zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1062    enum zio_flag flags)
1063{
1064	zio_t *zio;
1065	enum zio_stage stage = ZIO_FREE_PIPELINE;
1066
1067	ASSERT(!BP_IS_HOLE(bp));
1068	ASSERT(spa_syncing_txg(spa) == txg);
1069
1070	if (BP_IS_EMBEDDED(bp))
1071		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1072
1073	metaslab_check_free(spa, bp);
1074	arc_freed(spa, bp);
1075	dsl_scan_freed(spa, bp);
1076
1077	/*
1078	 * GANG and DEDUP blocks can induce a read (for the gang block header,
1079	 * or the DDT), so issue them asynchronously so that this thread is
1080	 * not tied up.
1081	 */
1082	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
1083		stage |= ZIO_STAGE_ISSUE_ASYNC;
1084
1085	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1086	    BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1087	    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
1088
1089	return (zio);
1090}
1091
1092zio_t *
1093zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1094    zio_done_func_t *done, void *private, enum zio_flag flags)
1095{
1096	zio_t *zio;
1097
1098	zfs_blkptr_verify(spa, bp);
1099
1100	if (BP_IS_EMBEDDED(bp))
1101		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1102
1103	/*
1104	 * A claim is an allocation of a specific block.  Claims are needed
1105	 * to support immediate writes in the intent log.  The issue is that
1106	 * immediate writes contain committed data, but in a txg that was
1107	 * *not* committed.  Upon opening the pool after an unclean shutdown,
1108	 * the intent log claims all blocks that contain immediate write data
1109	 * so that the SPA knows they're in use.
1110	 *
1111	 * All claims *must* be resolved in the first txg -- before the SPA
1112	 * starts allocating blocks -- so that nothing is allocated twice.
1113	 * If txg == 0 we just verify that the block is claimable.
1114	 */
1115	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1116	    spa_min_claim_txg(spa));
1117	ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1118	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
1119
1120	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1121	    BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1122	    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1123	ASSERT0(zio->io_queued_timestamp);
1124
1125	return (zio);
1126}
1127
1128zio_t *
1129zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
1130    zio_done_func_t *done, void *private, enum zio_flag flags)
1131{
1132	zio_t *zio;
1133	int c;
1134
1135	if (vd->vdev_children == 0) {
1136		zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1137		    ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1138		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1139
1140		zio->io_cmd = cmd;
1141	} else {
1142		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1143
1144		for (c = 0; c < vd->vdev_children; c++)
1145			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1146			    done, private, flags));
1147	}
1148
1149	return (zio);
1150}
1151
1152zio_t *
1153zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1154    zio_done_func_t *done, void *private, zio_priority_t priority,
1155    enum zio_flag flags, enum trim_flag trim_flags)
1156{
1157	zio_t *zio;
1158
1159	ASSERT0(vd->vdev_children);
1160	ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
1161	ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1162	ASSERT3U(size, !=, 0);
1163
1164	zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done,
1165	    private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL,
1166	    vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE);
1167	zio->io_trim_flags = trim_flags;
1168
1169	return (zio);
1170}
1171
1172zio_t *
1173zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1174    abd_t *data, int checksum, zio_done_func_t *done, void *private,
1175    zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1176{
1177	zio_t *zio;
1178
1179	ASSERT(vd->vdev_children == 0);
1180	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1181	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1182	ASSERT3U(offset + size, <=, vd->vdev_psize);
1183
1184	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1185	    private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1186	    offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1187
1188	zio->io_prop.zp_checksum = checksum;
1189
1190	return (zio);
1191}
1192
1193zio_t *
1194zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1195    abd_t *data, int checksum, zio_done_func_t *done, void *private,
1196    zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1197{
1198	zio_t *zio;
1199
1200	ASSERT(vd->vdev_children == 0);
1201	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1202	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1203	ASSERT3U(offset + size, <=, vd->vdev_psize);
1204
1205	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1206	    private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1207	    offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1208
1209	zio->io_prop.zp_checksum = checksum;
1210
1211	if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1212		/*
1213		 * zec checksums are necessarily destructive -- they modify
1214		 * the end of the write buffer to hold the verifier/checksum.
1215		 * Therefore, we must make a local copy in case the data is
1216		 * being written to multiple places in parallel.
1217		 */
1218		abd_t *wbuf = abd_alloc_sametype(data, size);
1219		abd_copy(wbuf, data, size);
1220
1221		zio_push_transform(zio, wbuf, size, size, NULL);
1222	}
1223
1224	return (zio);
1225}
1226
1227/*
1228 * Create a child I/O to do some work for us.
1229 */
1230zio_t *
1231zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1232    abd_t *data, uint64_t size, int type, zio_priority_t priority,
1233    enum zio_flag flags, zio_done_func_t *done, void *private)
1234{
1235	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1236	zio_t *zio;
1237
1238	/*
1239	 * vdev child I/Os do not propagate their error to the parent.
1240	 * Therefore, for correct operation the caller *must* check for
1241	 * and handle the error in the child i/o's done callback.
1242	 * The only exceptions are i/os that we don't care about
1243	 * (OPTIONAL or REPAIR).
1244	 */
1245	ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1246	    done != NULL);
1247
1248	if (type == ZIO_TYPE_READ && bp != NULL) {
1249		/*
1250		 * If we have the bp, then the child should perform the
1251		 * checksum and the parent need not.  This pushes error
1252		 * detection as close to the leaves as possible and
1253		 * eliminates redundant checksums in the interior nodes.
1254		 */
1255		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1256		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1257	}
1258
1259	if (vd->vdev_ops->vdev_op_leaf) {
1260		ASSERT0(vd->vdev_children);
1261		offset += VDEV_LABEL_START_SIZE;
1262	}
1263
1264	flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1265
1266	/*
1267	 * If we've decided to do a repair, the write is not speculative --
1268	 * even if the original read was.
1269	 */
1270	if (flags & ZIO_FLAG_IO_REPAIR)
1271		flags &= ~ZIO_FLAG_SPECULATIVE;
1272
1273	/*
1274	 * If we're creating a child I/O that is not associated with a
1275	 * top-level vdev, then the child zio is not an allocating I/O.
1276	 * If this is a retried I/O then we ignore it since we will
1277	 * have already processed the original allocating I/O.
1278	 */
1279	if (flags & ZIO_FLAG_IO_ALLOCATING &&
1280	    (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1281		ASSERT(pio->io_metaslab_class != NULL);
1282		ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled);
1283		ASSERT(type == ZIO_TYPE_WRITE);
1284		ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1285		ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1286		ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1287		    pio->io_child_type == ZIO_CHILD_GANG);
1288
1289		flags &= ~ZIO_FLAG_IO_ALLOCATING;
1290	}
1291
1292	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1293	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1294	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1295	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1296
1297	zio->io_physdone = pio->io_physdone;
1298	if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1299		zio->io_logical->io_phys_children++;
1300
1301	return (zio);
1302}
1303
1304zio_t *
1305zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1306    zio_type_t type, zio_priority_t priority, enum zio_flag flags,
1307    zio_done_func_t *done, void *private)
1308{
1309	zio_t *zio;
1310
1311	ASSERT(vd->vdev_ops->vdev_op_leaf);
1312
1313	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1314	    data, size, size, done, private, type, priority,
1315	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1316	    vd, offset, NULL,
1317	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1318
1319	return (zio);
1320}
1321
1322void
1323zio_flush(zio_t *zio, vdev_t *vd)
1324{
1325	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1326	    NULL, NULL,
1327	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1328}
1329
1330void
1331zio_shrink(zio_t *zio, uint64_t size)
1332{
1333	ASSERT3P(zio->io_executor, ==, NULL);
1334	ASSERT3P(zio->io_orig_size, ==, zio->io_size);
1335	ASSERT3U(size, <=, zio->io_size);
1336
1337	/*
1338	 * We don't shrink for raidz because of problems with the
1339	 * reconstruction when reading back less than the block size.
1340	 * Note, BP_IS_RAIDZ() assumes no compression.
1341	 */
1342	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1343	if (!BP_IS_RAIDZ(zio->io_bp)) {
1344		/* we are not doing a raw write */
1345		ASSERT3U(zio->io_size, ==, zio->io_lsize);
1346		zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1347	}
1348}
1349
1350/*
1351 * ==========================================================================
1352 * Prepare to read and write logical blocks
1353 * ==========================================================================
1354 */
1355
1356static int
1357zio_read_bp_init(zio_t *zio)
1358{
1359	blkptr_t *bp = zio->io_bp;
1360	uint64_t psize =
1361	    BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1362
1363	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1364
1365	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1366	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
1367	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1368		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1369		    psize, psize, zio_decompress);
1370	}
1371
1372	if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) ||
1373	    BP_HAS_INDIRECT_MAC_CKSUM(bp)) &&
1374	    zio->io_child_type == ZIO_CHILD_LOGICAL) {
1375		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1376		    psize, psize, zio_decrypt);
1377	}
1378
1379	if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1380		int psize = BPE_GET_PSIZE(bp);
1381		void *data = abd_borrow_buf(zio->io_abd, psize);
1382
1383		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1384		decode_embedded_bp_compressed(bp, data);
1385		abd_return_buf_copy(zio->io_abd, data, psize);
1386	} else {
1387		ASSERT(!BP_IS_EMBEDDED(bp));
1388		ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1389	}
1390
1391	if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1392		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1393
1394	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1395		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1396
1397	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1398		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1399
1400	return (ZIO_PIPELINE_CONTINUE);
1401}
1402
1403static int
1404zio_write_bp_init(zio_t *zio)
1405{
1406	if (!IO_IS_ALLOCATING(zio))
1407		return (ZIO_PIPELINE_CONTINUE);
1408
1409	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1410
1411	if (zio->io_bp_override) {
1412		blkptr_t *bp = zio->io_bp;
1413		zio_prop_t *zp = &zio->io_prop;
1414
1415		ASSERT(bp->blk_birth != zio->io_txg);
1416		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1417
1418		*bp = *zio->io_bp_override;
1419		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1420
1421		if (BP_IS_EMBEDDED(bp))
1422			return (ZIO_PIPELINE_CONTINUE);
1423
1424		/*
1425		 * If we've been overridden and nopwrite is set then
1426		 * set the flag accordingly to indicate that a nopwrite
1427		 * has already occurred.
1428		 */
1429		if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1430			ASSERT(!zp->zp_dedup);
1431			ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1432			zio->io_flags |= ZIO_FLAG_NOPWRITE;
1433			return (ZIO_PIPELINE_CONTINUE);
1434		}
1435
1436		ASSERT(!zp->zp_nopwrite);
1437
1438		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1439			return (ZIO_PIPELINE_CONTINUE);
1440
1441		ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1442		    ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1443
1444		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum &&
1445		    !zp->zp_encrypt) {
1446			BP_SET_DEDUP(bp, 1);
1447			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1448			return (ZIO_PIPELINE_CONTINUE);
1449		}
1450
1451		/*
1452		 * We were unable to handle this as an override bp, treat
1453		 * it as a regular write I/O.
1454		 */
1455		zio->io_bp_override = NULL;
1456		*bp = zio->io_bp_orig;
1457		zio->io_pipeline = zio->io_orig_pipeline;
1458	}
1459
1460	return (ZIO_PIPELINE_CONTINUE);
1461}
1462
1463static int
1464zio_write_compress(zio_t *zio)
1465{
1466	spa_t *spa = zio->io_spa;
1467	zio_prop_t *zp = &zio->io_prop;
1468	enum zio_compress compress = zp->zp_compress;
1469	blkptr_t *bp = zio->io_bp;
1470	uint64_t lsize = zio->io_lsize;
1471	uint64_t psize = zio->io_size;
1472	int pass = 1;
1473
1474	/*
1475	 * If our children haven't all reached the ready stage,
1476	 * wait for them and then repeat this pipeline stage.
1477	 */
1478	if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1479	    ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1480		return (ZIO_PIPELINE_STOP);
1481	}
1482
1483	if (!IO_IS_ALLOCATING(zio))
1484		return (ZIO_PIPELINE_CONTINUE);
1485
1486	if (zio->io_children_ready != NULL) {
1487		/*
1488		 * Now that all our children are ready, run the callback
1489		 * associated with this zio in case it wants to modify the
1490		 * data to be written.
1491		 */
1492		ASSERT3U(zp->zp_level, >, 0);
1493		zio->io_children_ready(zio);
1494	}
1495
1496	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1497	ASSERT(zio->io_bp_override == NULL);
1498
1499	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1500		/*
1501		 * We're rewriting an existing block, which means we're
1502		 * working on behalf of spa_sync().  For spa_sync() to
1503		 * converge, it must eventually be the case that we don't
1504		 * have to allocate new blocks.  But compression changes
1505		 * the blocksize, which forces a reallocate, and makes
1506		 * convergence take longer.  Therefore, after the first
1507		 * few passes, stop compressing to ensure convergence.
1508		 */
1509		pass = spa_sync_pass(spa);
1510
1511		ASSERT(zio->io_txg == spa_syncing_txg(spa));
1512		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1513		ASSERT(!BP_GET_DEDUP(bp));
1514
1515		if (pass >= zfs_sync_pass_dont_compress)
1516			compress = ZIO_COMPRESS_OFF;
1517
1518		/* Make sure someone doesn't change their mind on overwrites */
1519		ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1520		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1521	}
1522
1523	/* If it's a compressed write that is not raw, compress the buffer. */
1524	if (compress != ZIO_COMPRESS_OFF &&
1525	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1526		void *cbuf = zio_buf_alloc(lsize);
1527		psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1528		if (psize == 0 || psize == lsize) {
1529			compress = ZIO_COMPRESS_OFF;
1530			zio_buf_free(cbuf, lsize);
1531		} else if (!zp->zp_dedup && !zp->zp_encrypt &&
1532		    psize <= BPE_PAYLOAD_SIZE &&
1533		    zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1534		    spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1535			encode_embedded_bp_compressed(bp,
1536			    cbuf, compress, lsize, psize);
1537			BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1538			BP_SET_TYPE(bp, zio->io_prop.zp_type);
1539			BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1540			zio_buf_free(cbuf, lsize);
1541			bp->blk_birth = zio->io_txg;
1542			zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1543			ASSERT(spa_feature_is_active(spa,
1544			    SPA_FEATURE_EMBEDDED_DATA));
1545			return (ZIO_PIPELINE_CONTINUE);
1546		} else {
1547			/*
1548			 * Round up compressed size up to the ashift
1549			 * of the smallest-ashift device, and zero the tail.
1550			 * This ensures that the compressed size of the BP
1551			 * (and thus compressratio property) are correct,
1552			 * in that we charge for the padding used to fill out
1553			 * the last sector.
1554			 */
1555			ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1556			size_t rounded = (size_t)P2ROUNDUP(psize,
1557			    1ULL << spa->spa_min_ashift);
1558			if (rounded >= lsize) {
1559				compress = ZIO_COMPRESS_OFF;
1560				zio_buf_free(cbuf, lsize);
1561				psize = lsize;
1562			} else {
1563				abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1564				abd_take_ownership_of_buf(cdata, B_TRUE);
1565				abd_zero_off(cdata, psize, rounded - psize);
1566				psize = rounded;
1567				zio_push_transform(zio, cdata,
1568				    psize, lsize, NULL);
1569			}
1570		}
1571
1572		/*
1573		 * We were unable to handle this as an override bp, treat
1574		 * it as a regular write I/O.
1575		 */
1576		zio->io_bp_override = NULL;
1577		*bp = zio->io_bp_orig;
1578		zio->io_pipeline = zio->io_orig_pipeline;
1579
1580	} else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 &&
1581	    zp->zp_type == DMU_OT_DNODE) {
1582		/*
1583		 * The DMU actually relies on the zio layer's compression
1584		 * to free metadnode blocks that have had all contained
1585		 * dnodes freed. As a result, even when doing a raw
1586		 * receive, we must check whether the block can be compressed
1587		 * to a hole.
1588		 */
1589		psize = zio_compress_data(ZIO_COMPRESS_EMPTY,
1590		    zio->io_abd, NULL, lsize);
1591		if (psize == 0)
1592			compress = ZIO_COMPRESS_OFF;
1593	} else {
1594		ASSERT3U(psize, !=, 0);
1595	}
1596
1597	/*
1598	 * The final pass of spa_sync() must be all rewrites, but the first
1599	 * few passes offer a trade-off: allocating blocks defers convergence,
1600	 * but newly allocated blocks are sequential, so they can be written
1601	 * to disk faster.  Therefore, we allow the first few passes of
1602	 * spa_sync() to allocate new blocks, but force rewrites after that.
1603	 * There should only be a handful of blocks after pass 1 in any case.
1604	 */
1605	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1606	    BP_GET_PSIZE(bp) == psize &&
1607	    pass >= zfs_sync_pass_rewrite) {
1608		VERIFY3U(psize, !=, 0);
1609		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1610		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1611		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1612	} else {
1613		BP_ZERO(bp);
1614		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1615	}
1616
1617	if (psize == 0) {
1618		if (zio->io_bp_orig.blk_birth != 0 &&
1619		    spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1620			BP_SET_LSIZE(bp, lsize);
1621			BP_SET_TYPE(bp, zp->zp_type);
1622			BP_SET_LEVEL(bp, zp->zp_level);
1623			BP_SET_BIRTH(bp, zio->io_txg, 0);
1624		}
1625		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1626	} else {
1627		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1628		BP_SET_LSIZE(bp, lsize);
1629		BP_SET_TYPE(bp, zp->zp_type);
1630		BP_SET_LEVEL(bp, zp->zp_level);
1631		BP_SET_PSIZE(bp, psize);
1632		BP_SET_COMPRESS(bp, compress);
1633		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1634		BP_SET_DEDUP(bp, zp->zp_dedup);
1635		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1636		if (zp->zp_dedup) {
1637			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1638			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1639			ASSERT(!zp->zp_encrypt ||
1640			    DMU_OT_IS_ENCRYPTED(zp->zp_type));
1641			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1642		}
1643		if (zp->zp_nopwrite) {
1644			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1645			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1646			zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1647		}
1648	}
1649	return (ZIO_PIPELINE_CONTINUE);
1650}
1651
1652static int
1653zio_free_bp_init(zio_t *zio)
1654{
1655	blkptr_t *bp = zio->io_bp;
1656
1657	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1658		if (BP_GET_DEDUP(bp))
1659			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1660	}
1661
1662	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1663
1664	return (ZIO_PIPELINE_CONTINUE);
1665}
1666
1667/*
1668 * ==========================================================================
1669 * Execute the I/O pipeline
1670 * ==========================================================================
1671 */
1672
1673static void
1674zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1675{
1676	spa_t *spa = zio->io_spa;
1677	zio_type_t t = zio->io_type;
1678	int flags = (cutinline ? TQ_FRONT : 0);
1679
1680	/*
1681	 * If we're a config writer or a probe, the normal issue and
1682	 * interrupt threads may all be blocked waiting for the config lock.
1683	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1684	 */
1685	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1686		t = ZIO_TYPE_NULL;
1687
1688	/*
1689	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1690	 */
1691	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1692		t = ZIO_TYPE_NULL;
1693
1694	/*
1695	 * If this is a high priority I/O, then use the high priority taskq if
1696	 * available.
1697	 */
1698	if ((zio->io_priority == ZIO_PRIORITY_NOW ||
1699	    zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) &&
1700	    spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1701		q++;
1702
1703	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1704
1705	/*
1706	 * NB: We are assuming that the zio can only be dispatched
1707	 * to a single taskq at a time.  It would be a grievous error
1708	 * to dispatch the zio to another taskq at the same time.
1709	 */
1710	ASSERT(zio->io_tqent.tqent_next == NULL);
1711	spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1712	    flags, &zio->io_tqent);
1713}
1714
1715static boolean_t
1716zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1717{
1718	kthread_t *executor = zio->io_executor;
1719	spa_t *spa = zio->io_spa;
1720
1721	for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1722		spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1723		uint_t i;
1724		for (i = 0; i < tqs->stqs_count; i++) {
1725			if (taskq_member(tqs->stqs_taskq[i], executor))
1726				return (B_TRUE);
1727		}
1728	}
1729
1730	return (B_FALSE);
1731}
1732
1733static int
1734zio_issue_async(zio_t *zio)
1735{
1736	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1737
1738	return (ZIO_PIPELINE_STOP);
1739}
1740
1741void
1742zio_interrupt(zio_t *zio)
1743{
1744	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1745}
1746
1747void
1748zio_delay_interrupt(zio_t *zio)
1749{
1750	/*
1751	 * The timeout_generic() function isn't defined in userspace, so
1752	 * rather than trying to implement the function, the zio delay
1753	 * functionality has been disabled for userspace builds.
1754	 */
1755
1756#ifdef _KERNEL
1757	/*
1758	 * If io_target_timestamp is zero, then no delay has been registered
1759	 * for this IO, thus jump to the end of this function and "skip" the
1760	 * delay; issuing it directly to the zio layer.
1761	 */
1762	if (zio->io_target_timestamp != 0) {
1763		hrtime_t now = gethrtime();
1764
1765		if (now >= zio->io_target_timestamp) {
1766			/*
1767			 * This IO has already taken longer than the target
1768			 * delay to complete, so we don't want to delay it
1769			 * any longer; we "miss" the delay and issue it
1770			 * directly to the zio layer. This is likely due to
1771			 * the target latency being set to a value less than
1772			 * the underlying hardware can satisfy (e.g. delay
1773			 * set to 1ms, but the disks take 10ms to complete an
1774			 * IO request).
1775			 */
1776
1777			DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1778			    hrtime_t, now);
1779
1780			zio_interrupt(zio);
1781		} else {
1782			hrtime_t diff = zio->io_target_timestamp - now;
1783
1784			DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1785			    hrtime_t, now, hrtime_t, diff);
1786
1787			(void) timeout_generic(CALLOUT_NORMAL,
1788			    (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1789		}
1790
1791		return;
1792	}
1793#endif
1794
1795	DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1796	zio_interrupt(zio);
1797}
1798
1799/*
1800 * Execute the I/O pipeline until one of the following occurs:
1801 *
1802 *	(1) the I/O completes
1803 *	(2) the pipeline stalls waiting for dependent child I/Os
1804 *	(3) the I/O issues, so we're waiting for an I/O completion interrupt
1805 *	(4) the I/O is delegated by vdev-level caching or aggregation
1806 *	(5) the I/O is deferred due to vdev-level queueing
1807 *	(6) the I/O is handed off to another thread.
1808 *
1809 * In all cases, the pipeline stops whenever there's no CPU work; it never
1810 * burns a thread in cv_wait().
1811 *
1812 * There's no locking on io_stage because there's no legitimate way
1813 * for multiple threads to be attempting to process the same I/O.
1814 */
1815static zio_pipe_stage_t *zio_pipeline[];
1816
1817void
1818zio_execute(zio_t *zio)
1819{
1820	zio->io_executor = curthread;
1821
1822	ASSERT3U(zio->io_queued_timestamp, >, 0);
1823
1824	while (zio->io_stage < ZIO_STAGE_DONE) {
1825		enum zio_stage pipeline = zio->io_pipeline;
1826		enum zio_stage stage = zio->io_stage;
1827		int rv;
1828
1829		ASSERT(!MUTEX_HELD(&zio->io_lock));
1830		ASSERT(ISP2(stage));
1831		ASSERT(zio->io_stall == NULL);
1832
1833		do {
1834			stage <<= 1;
1835		} while ((stage & pipeline) == 0);
1836
1837		ASSERT(stage <= ZIO_STAGE_DONE);
1838
1839		/*
1840		 * If we are in interrupt context and this pipeline stage
1841		 * will grab a config lock that is held across I/O,
1842		 * or may wait for an I/O that needs an interrupt thread
1843		 * to complete, issue async to avoid deadlock.
1844		 *
1845		 * For VDEV_IO_START, we cut in line so that the io will
1846		 * be sent to disk promptly.
1847		 */
1848		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1849		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1850			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1851			    zio_requeue_io_start_cut_in_line : B_FALSE;
1852			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1853			return;
1854		}
1855
1856		zio->io_stage = stage;
1857		zio->io_pipeline_trace |= zio->io_stage;
1858		rv = zio_pipeline[highbit64(stage) - 1](zio);
1859
1860		if (rv == ZIO_PIPELINE_STOP)
1861			return;
1862
1863		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1864	}
1865}
1866
1867/*
1868 * ==========================================================================
1869 * Initiate I/O, either sync or async
1870 * ==========================================================================
1871 */
1872int
1873zio_wait(zio_t *zio)
1874{
1875	int error;
1876
1877	ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1878	ASSERT3P(zio->io_executor, ==, NULL);
1879
1880	zio->io_waiter = curthread;
1881	ASSERT0(zio->io_queued_timestamp);
1882	zio->io_queued_timestamp = gethrtime();
1883
1884	zio_execute(zio);
1885
1886	mutex_enter(&zio->io_lock);
1887	while (zio->io_executor != NULL)
1888		cv_wait(&zio->io_cv, &zio->io_lock);
1889	mutex_exit(&zio->io_lock);
1890
1891	error = zio->io_error;
1892	zio_destroy(zio);
1893
1894	return (error);
1895}
1896
1897void
1898zio_nowait(zio_t *zio)
1899{
1900	ASSERT3P(zio->io_executor, ==, NULL);
1901
1902	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1903	    zio_unique_parent(zio) == NULL) {
1904		/*
1905		 * This is a logical async I/O with no parent to wait for it.
1906		 * We add it to the spa_async_root_zio "Godfather" I/O which
1907		 * will ensure they complete prior to unloading the pool.
1908		 */
1909		spa_t *spa = zio->io_spa;
1910
1911		zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1912	}
1913
1914	ASSERT0(zio->io_queued_timestamp);
1915	zio->io_queued_timestamp = gethrtime();
1916	zio_execute(zio);
1917}
1918
1919/*
1920 * ==========================================================================
1921 * Reexecute, cancel, or suspend/resume failed I/O
1922 * ==========================================================================
1923 */
1924
1925static void
1926zio_reexecute(zio_t *pio)
1927{
1928	zio_t *cio, *cio_next;
1929
1930	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1931	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1932	ASSERT(pio->io_gang_leader == NULL);
1933	ASSERT(pio->io_gang_tree == NULL);
1934
1935	pio->io_flags = pio->io_orig_flags;
1936	pio->io_stage = pio->io_orig_stage;
1937	pio->io_pipeline = pio->io_orig_pipeline;
1938	pio->io_reexecute = 0;
1939	pio->io_flags |= ZIO_FLAG_REEXECUTED;
1940	pio->io_pipeline_trace = 0;
1941	pio->io_error = 0;
1942	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1943		pio->io_state[w] = 0;
1944	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1945		pio->io_child_error[c] = 0;
1946
1947	if (IO_IS_ALLOCATING(pio))
1948		BP_ZERO(pio->io_bp);
1949
1950	/*
1951	 * As we reexecute pio's children, new children could be created.
1952	 * New children go to the head of pio's io_child_list, however,
1953	 * so we will (correctly) not reexecute them.  The key is that
1954	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1955	 * cannot be affected by any side effects of reexecuting 'cio'.
1956	 */
1957	zio_link_t *zl = NULL;
1958	mutex_enter(&pio->io_lock);
1959	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1960		cio_next = zio_walk_children(pio, &zl);
1961		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1962			pio->io_children[cio->io_child_type][w]++;
1963		mutex_exit(&pio->io_lock);
1964		zio_reexecute(cio);
1965		mutex_enter(&pio->io_lock);
1966	}
1967	mutex_exit(&pio->io_lock);
1968
1969	/*
1970	 * Now that all children have been reexecuted, execute the parent.
1971	 * We don't reexecute "The Godfather" I/O here as it's the
1972	 * responsibility of the caller to wait on it.
1973	 */
1974	if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1975		pio->io_queued_timestamp = gethrtime();
1976		zio_execute(pio);
1977	}
1978}
1979
1980void
1981zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
1982{
1983	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1984		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1985		    "failure and the failure mode property for this pool "
1986		    "is set to panic.", spa_name(spa));
1987
1988	cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
1989	    "failure and has been suspended; `zpool clear` will be required "
1990	    "before the pool can be written to.", spa_name(spa));
1991
1992	(void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
1993	    NULL, NULL, 0, 0);
1994
1995	mutex_enter(&spa->spa_suspend_lock);
1996
1997	if (spa->spa_suspend_zio_root == NULL)
1998		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1999		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2000		    ZIO_FLAG_GODFATHER);
2001
2002	spa->spa_suspended = reason;
2003
2004	if (zio != NULL) {
2005		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2006		ASSERT(zio != spa->spa_suspend_zio_root);
2007		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2008		ASSERT(zio_unique_parent(zio) == NULL);
2009		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
2010		zio_add_child(spa->spa_suspend_zio_root, zio);
2011	}
2012
2013	mutex_exit(&spa->spa_suspend_lock);
2014}
2015
2016int
2017zio_resume(spa_t *spa)
2018{
2019	zio_t *pio;
2020
2021	/*
2022	 * Reexecute all previously suspended i/o.
2023	 */
2024	mutex_enter(&spa->spa_suspend_lock);
2025	spa->spa_suspended = ZIO_SUSPEND_NONE;
2026	cv_broadcast(&spa->spa_suspend_cv);
2027	pio = spa->spa_suspend_zio_root;
2028	spa->spa_suspend_zio_root = NULL;
2029	mutex_exit(&spa->spa_suspend_lock);
2030
2031	if (pio == NULL)
2032		return (0);
2033
2034	zio_reexecute(pio);
2035	return (zio_wait(pio));
2036}
2037
2038void
2039zio_resume_wait(spa_t *spa)
2040{
2041	mutex_enter(&spa->spa_suspend_lock);
2042	while (spa_suspended(spa))
2043		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
2044	mutex_exit(&spa->spa_suspend_lock);
2045}
2046
2047/*
2048 * ==========================================================================
2049 * Gang blocks.
2050 *
2051 * A gang block is a collection of small blocks that looks to the DMU
2052 * like one large block.  When zio_dva_allocate() cannot find a block
2053 * of the requested size, due to either severe fragmentation or the pool
2054 * being nearly full, it calls zio_write_gang_block() to construct the
2055 * block from smaller fragments.
2056 *
2057 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2058 * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
2059 * an indirect block: it's an array of block pointers.  It consumes
2060 * only one sector and hence is allocatable regardless of fragmentation.
2061 * The gang header's bps point to its gang members, which hold the data.
2062 *
2063 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2064 * as the verifier to ensure uniqueness of the SHA256 checksum.
2065 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2066 * not the gang header.  This ensures that data block signatures (needed for
2067 * deduplication) are independent of how the block is physically stored.
2068 *
2069 * Gang blocks can be nested: a gang member may itself be a gang block.
2070 * Thus every gang block is a tree in which root and all interior nodes are
2071 * gang headers, and the leaves are normal blocks that contain user data.
2072 * The root of the gang tree is called the gang leader.
2073 *
2074 * To perform any operation (read, rewrite, free, claim) on a gang block,
2075 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2076 * in the io_gang_tree field of the original logical i/o by recursively
2077 * reading the gang leader and all gang headers below it.  This yields
2078 * an in-core tree containing the contents of every gang header and the
2079 * bps for every constituent of the gang block.
2080 *
2081 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2082 * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
2083 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2084 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2085 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2086 * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
2087 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2088 * of the gang header plus zio_checksum_compute() of the data to update the
2089 * gang header's blk_cksum as described above.
2090 *
2091 * The two-phase assemble/issue model solves the problem of partial failure --
2092 * what if you'd freed part of a gang block but then couldn't read the
2093 * gang header for another part?  Assembling the entire gang tree first
2094 * ensures that all the necessary gang header I/O has succeeded before
2095 * starting the actual work of free, claim, or write.  Once the gang tree
2096 * is assembled, free and claim are in-memory operations that cannot fail.
2097 *
2098 * In the event that a gang write fails, zio_dva_unallocate() walks the
2099 * gang tree to immediately free (i.e. insert back into the space map)
2100 * everything we've allocated.  This ensures that we don't get ENOSPC
2101 * errors during repeated suspend/resume cycles due to a flaky device.
2102 *
2103 * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
2104 * the gang tree, we won't modify the block, so we can safely defer the free
2105 * (knowing that the block is still intact).  If we *can* assemble the gang
2106 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2107 * each constituent bp and we can allocate a new block on the next sync pass.
2108 *
2109 * In all cases, the gang tree allows complete recovery from partial failure.
2110 * ==========================================================================
2111 */
2112
2113static void
2114zio_gang_issue_func_done(zio_t *zio)
2115{
2116	abd_put(zio->io_abd);
2117}
2118
2119static zio_t *
2120zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2121    uint64_t offset)
2122{
2123	if (gn != NULL)
2124		return (pio);
2125
2126	return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2127	    BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2128	    NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2129	    &pio->io_bookmark));
2130}
2131
2132static zio_t *
2133zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2134    uint64_t offset)
2135{
2136	zio_t *zio;
2137
2138	if (gn != NULL) {
2139		abd_t *gbh_abd =
2140		    abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2141		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2142		    gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2143		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2144		    &pio->io_bookmark);
2145		/*
2146		 * As we rewrite each gang header, the pipeline will compute
2147		 * a new gang block header checksum for it; but no one will
2148		 * compute a new data checksum, so we do that here.  The one
2149		 * exception is the gang leader: the pipeline already computed
2150		 * its data checksum because that stage precedes gang assembly.
2151		 * (Presently, nothing actually uses interior data checksums;
2152		 * this is just good hygiene.)
2153		 */
2154		if (gn != pio->io_gang_leader->io_gang_tree) {
2155			abd_t *buf = abd_get_offset(data, offset);
2156
2157			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2158			    buf, BP_GET_PSIZE(bp));
2159
2160			abd_put(buf);
2161		}
2162		/*
2163		 * If we are here to damage data for testing purposes,
2164		 * leave the GBH alone so that we can detect the damage.
2165		 */
2166		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2167			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2168	} else {
2169		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2170		    abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2171		    zio_gang_issue_func_done, NULL, pio->io_priority,
2172		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2173	}
2174
2175	return (zio);
2176}
2177
2178/* ARGSUSED */
2179static zio_t *
2180zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2181    uint64_t offset)
2182{
2183	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2184	    ZIO_GANG_CHILD_FLAGS(pio)));
2185}
2186
2187/* ARGSUSED */
2188static zio_t *
2189zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2190    uint64_t offset)
2191{
2192	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2193	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2194}
2195
2196static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2197	NULL,
2198	zio_read_gang,
2199	zio_rewrite_gang,
2200	zio_free_gang,
2201	zio_claim_gang,
2202	NULL
2203};
2204
2205static void zio_gang_tree_assemble_done(zio_t *zio);
2206
2207static zio_gang_node_t *
2208zio_gang_node_alloc(zio_gang_node_t **gnpp)
2209{
2210	zio_gang_node_t *gn;
2211
2212	ASSERT(*gnpp == NULL);
2213
2214	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2215	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2216	*gnpp = gn;
2217
2218	return (gn);
2219}
2220
2221static void
2222zio_gang_node_free(zio_gang_node_t **gnpp)
2223{
2224	zio_gang_node_t *gn = *gnpp;
2225
2226	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2227		ASSERT(gn->gn_child[g] == NULL);
2228
2229	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2230	kmem_free(gn, sizeof (*gn));
2231	*gnpp = NULL;
2232}
2233
2234static void
2235zio_gang_tree_free(zio_gang_node_t **gnpp)
2236{
2237	zio_gang_node_t *gn = *gnpp;
2238
2239	if (gn == NULL)
2240		return;
2241
2242	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2243		zio_gang_tree_free(&gn->gn_child[g]);
2244
2245	zio_gang_node_free(gnpp);
2246}
2247
2248static void
2249zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2250{
2251	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2252	abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2253
2254	ASSERT(gio->io_gang_leader == gio);
2255	ASSERT(BP_IS_GANG(bp));
2256
2257	zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2258	    zio_gang_tree_assemble_done, gn, gio->io_priority,
2259	    ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2260}
2261
2262static void
2263zio_gang_tree_assemble_done(zio_t *zio)
2264{
2265	zio_t *gio = zio->io_gang_leader;
2266	zio_gang_node_t *gn = zio->io_private;
2267	blkptr_t *bp = zio->io_bp;
2268
2269	ASSERT(gio == zio_unique_parent(zio));
2270	ASSERT(zio->io_child_count == 0);
2271
2272	if (zio->io_error)
2273		return;
2274
2275	/* this ABD was created from a linear buf in zio_gang_tree_assemble */
2276	if (BP_SHOULD_BYTESWAP(bp))
2277		byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2278
2279	ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2280	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2281	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2282
2283	abd_put(zio->io_abd);
2284
2285	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2286		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2287		if (!BP_IS_GANG(gbp))
2288			continue;
2289		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2290	}
2291}
2292
2293static void
2294zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2295    uint64_t offset)
2296{
2297	zio_t *gio = pio->io_gang_leader;
2298	zio_t *zio;
2299
2300	ASSERT(BP_IS_GANG(bp) == !!gn);
2301	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2302	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2303
2304	/*
2305	 * If you're a gang header, your data is in gn->gn_gbh.
2306	 * If you're a gang member, your data is in 'data' and gn == NULL.
2307	 */
2308	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2309
2310	if (gn != NULL) {
2311		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2312
2313		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2314			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2315			if (BP_IS_HOLE(gbp))
2316				continue;
2317			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2318			    offset);
2319			offset += BP_GET_PSIZE(gbp);
2320		}
2321	}
2322
2323	if (gn == gio->io_gang_tree)
2324		ASSERT3U(gio->io_size, ==, offset);
2325
2326	if (zio != pio)
2327		zio_nowait(zio);
2328}
2329
2330static int
2331zio_gang_assemble(zio_t *zio)
2332{
2333	blkptr_t *bp = zio->io_bp;
2334
2335	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2336	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2337
2338	zio->io_gang_leader = zio;
2339
2340	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2341
2342	return (ZIO_PIPELINE_CONTINUE);
2343}
2344
2345static int
2346zio_gang_issue(zio_t *zio)
2347{
2348	blkptr_t *bp = zio->io_bp;
2349
2350	if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2351		return (ZIO_PIPELINE_STOP);
2352	}
2353
2354	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2355	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2356
2357	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2358		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2359		    0);
2360	else
2361		zio_gang_tree_free(&zio->io_gang_tree);
2362
2363	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2364
2365	return (ZIO_PIPELINE_CONTINUE);
2366}
2367
2368static void
2369zio_write_gang_member_ready(zio_t *zio)
2370{
2371	zio_t *pio = zio_unique_parent(zio);
2372	zio_t *gio = zio->io_gang_leader;
2373	dva_t *cdva = zio->io_bp->blk_dva;
2374	dva_t *pdva = pio->io_bp->blk_dva;
2375	uint64_t asize;
2376
2377	if (BP_IS_HOLE(zio->io_bp))
2378		return;
2379
2380	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2381
2382	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2383	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2384	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2385	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2386	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2387
2388	mutex_enter(&pio->io_lock);
2389	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2390		ASSERT(DVA_GET_GANG(&pdva[d]));
2391		asize = DVA_GET_ASIZE(&pdva[d]);
2392		asize += DVA_GET_ASIZE(&cdva[d]);
2393		DVA_SET_ASIZE(&pdva[d], asize);
2394	}
2395	mutex_exit(&pio->io_lock);
2396}
2397
2398static void
2399zio_write_gang_done(zio_t *zio)
2400{
2401	/*
2402	 * The io_abd field will be NULL for a zio with no data.  The io_flags
2403	 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2404	 * check for it here as it is cleared in zio_ready.
2405	 */
2406	if (zio->io_abd != NULL)
2407		abd_put(zio->io_abd);
2408}
2409
2410static int
2411zio_write_gang_block(zio_t *pio)
2412{
2413	spa_t *spa = pio->io_spa;
2414	metaslab_class_t *mc = spa_normal_class(spa);
2415	blkptr_t *bp = pio->io_bp;
2416	zio_t *gio = pio->io_gang_leader;
2417	zio_t *zio;
2418	zio_gang_node_t *gn, **gnpp;
2419	zio_gbh_phys_t *gbh;
2420	abd_t *gbh_abd;
2421	uint64_t txg = pio->io_txg;
2422	uint64_t resid = pio->io_size;
2423	uint64_t lsize;
2424	int copies = gio->io_prop.zp_copies;
2425	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2426	zio_prop_t zp;
2427	int error;
2428	boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
2429
2430	/*
2431	 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2432	 * have a third copy.
2433	 */
2434	if (gio->io_prop.zp_encrypt && gbh_copies >= SPA_DVAS_PER_BP)
2435		gbh_copies = SPA_DVAS_PER_BP - 1;
2436
2437	int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2438	if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2439		ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2440		ASSERT(has_data);
2441
2442		flags |= METASLAB_ASYNC_ALLOC;
2443		VERIFY(zfs_refcount_held(&mc->mc_alloc_slots[pio->io_allocator],
2444		    pio));
2445
2446		/*
2447		 * The logical zio has already placed a reservation for
2448		 * 'copies' allocation slots but gang blocks may require
2449		 * additional copies. These additional copies
2450		 * (i.e. gbh_copies - copies) are guaranteed to succeed
2451		 * since metaslab_class_throttle_reserve() always allows
2452		 * additional reservations for gang blocks.
2453		 */
2454		VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2455		    pio->io_allocator, pio, flags));
2456	}
2457
2458	error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2459	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2460	    &pio->io_alloc_list, pio, pio->io_allocator);
2461	if (error) {
2462		if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2463			ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2464			ASSERT(has_data);
2465
2466			/*
2467			 * If we failed to allocate the gang block header then
2468			 * we remove any additional allocation reservations that
2469			 * we placed here. The original reservation will
2470			 * be removed when the logical I/O goes to the ready
2471			 * stage.
2472			 */
2473			metaslab_class_throttle_unreserve(mc,
2474			    gbh_copies - copies, pio->io_allocator, pio);
2475		}
2476		pio->io_error = error;
2477		return (ZIO_PIPELINE_CONTINUE);
2478	}
2479
2480	if (pio == gio) {
2481		gnpp = &gio->io_gang_tree;
2482	} else {
2483		gnpp = pio->io_private;
2484		ASSERT(pio->io_ready == zio_write_gang_member_ready);
2485	}
2486
2487	gn = zio_gang_node_alloc(gnpp);
2488	gbh = gn->gn_gbh;
2489	bzero(gbh, SPA_GANGBLOCKSIZE);
2490	gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2491
2492	/*
2493	 * Create the gang header.
2494	 */
2495	zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2496	    zio_write_gang_done, NULL, pio->io_priority,
2497	    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2498
2499	/*
2500	 * Create and nowait the gang children.
2501	 */
2502	for (int g = 0; resid != 0; resid -= lsize, g++) {
2503		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2504		    SPA_MINBLOCKSIZE);
2505		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2506
2507		zp.zp_checksum = gio->io_prop.zp_checksum;
2508		zp.zp_compress = ZIO_COMPRESS_OFF;
2509		zp.zp_type = DMU_OT_NONE;
2510		zp.zp_level = 0;
2511		zp.zp_copies = gio->io_prop.zp_copies;
2512		zp.zp_dedup = B_FALSE;
2513		zp.zp_dedup_verify = B_FALSE;
2514		zp.zp_nopwrite = B_FALSE;
2515		zp.zp_encrypt = gio->io_prop.zp_encrypt;
2516		zp.zp_byteorder = gio->io_prop.zp_byteorder;
2517		bzero(zp.zp_salt, ZIO_DATA_SALT_LEN);
2518		bzero(zp.zp_iv, ZIO_DATA_IV_LEN);
2519		bzero(zp.zp_mac, ZIO_DATA_MAC_LEN);
2520
2521		zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2522		    has_data ? abd_get_offset(pio->io_abd, pio->io_size -
2523		    resid) : NULL, lsize, lsize, &zp,
2524		    zio_write_gang_member_ready, NULL, NULL,
2525		    zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2526		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2527
2528		if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2529			ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2530			ASSERT(has_data);
2531
2532			/*
2533			 * Gang children won't throttle but we should
2534			 * account for their work, so reserve an allocation
2535			 * slot for them here.
2536			 */
2537			VERIFY(metaslab_class_throttle_reserve(mc,
2538			    zp.zp_copies, cio->io_allocator, cio, flags));
2539		}
2540		zio_nowait(cio);
2541	}
2542
2543	/*
2544	 * Set pio's pipeline to just wait for zio to finish.
2545	 */
2546	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2547
2548	zio_nowait(zio);
2549
2550	return (ZIO_PIPELINE_CONTINUE);
2551}
2552
2553/*
2554 * The zio_nop_write stage in the pipeline determines if allocating a
2555 * new bp is necessary.  The nopwrite feature can handle writes in
2556 * either syncing or open context (i.e. zil writes) and as a result is
2557 * mutually exclusive with dedup.
2558 *
2559 * By leveraging a cryptographically secure checksum, such as SHA256, we
2560 * can compare the checksums of the new data and the old to determine if
2561 * allocating a new block is required.  Note that our requirements for
2562 * cryptographic strength are fairly weak: there can't be any accidental
2563 * hash collisions, but we don't need to be secure against intentional
2564 * (malicious) collisions.  To trigger a nopwrite, you have to be able
2565 * to write the file to begin with, and triggering an incorrect (hash
2566 * collision) nopwrite is no worse than simply writing to the file.
2567 * That said, there are no known attacks against the checksum algorithms
2568 * used for nopwrite, assuming that the salt and the checksums
2569 * themselves remain secret.
2570 */
2571static int
2572zio_nop_write(zio_t *zio)
2573{
2574	blkptr_t *bp = zio->io_bp;
2575	blkptr_t *bp_orig = &zio->io_bp_orig;
2576	zio_prop_t *zp = &zio->io_prop;
2577
2578	ASSERT(BP_GET_LEVEL(bp) == 0);
2579	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2580	ASSERT(zp->zp_nopwrite);
2581	ASSERT(!zp->zp_dedup);
2582	ASSERT(zio->io_bp_override == NULL);
2583	ASSERT(IO_IS_ALLOCATING(zio));
2584
2585	/*
2586	 * Check to see if the original bp and the new bp have matching
2587	 * characteristics (i.e. same checksum, compression algorithms, etc).
2588	 * If they don't then just continue with the pipeline which will
2589	 * allocate a new bp.
2590	 */
2591	if (BP_IS_HOLE(bp_orig) ||
2592	    !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2593	    ZCHECKSUM_FLAG_NOPWRITE) ||
2594	    BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) ||
2595	    BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2596	    BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2597	    BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2598	    zp->zp_copies != BP_GET_NDVAS(bp_orig))
2599		return (ZIO_PIPELINE_CONTINUE);
2600
2601	/*
2602	 * If the checksums match then reset the pipeline so that we
2603	 * avoid allocating a new bp and issuing any I/O.
2604	 */
2605	if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2606		ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2607		    ZCHECKSUM_FLAG_NOPWRITE);
2608		ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2609		ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2610		ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2611		ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2612		    sizeof (uint64_t)) == 0);
2613
2614		*bp = *bp_orig;
2615		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2616		zio->io_flags |= ZIO_FLAG_NOPWRITE;
2617	}
2618
2619	return (ZIO_PIPELINE_CONTINUE);
2620}
2621
2622/*
2623 * ==========================================================================
2624 * Dedup
2625 * ==========================================================================
2626 */
2627static void
2628zio_ddt_child_read_done(zio_t *zio)
2629{
2630	blkptr_t *bp = zio->io_bp;
2631	ddt_entry_t *dde = zio->io_private;
2632	ddt_phys_t *ddp;
2633	zio_t *pio = zio_unique_parent(zio);
2634
2635	mutex_enter(&pio->io_lock);
2636	ddp = ddt_phys_select(dde, bp);
2637	if (zio->io_error == 0)
2638		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
2639
2640	if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2641		dde->dde_repair_abd = zio->io_abd;
2642	else
2643		abd_free(zio->io_abd);
2644	mutex_exit(&pio->io_lock);
2645}
2646
2647static int
2648zio_ddt_read_start(zio_t *zio)
2649{
2650	blkptr_t *bp = zio->io_bp;
2651
2652	ASSERT(BP_GET_DEDUP(bp));
2653	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2654	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2655
2656	if (zio->io_child_error[ZIO_CHILD_DDT]) {
2657		ddt_t *ddt = ddt_select(zio->io_spa, bp);
2658		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2659		ddt_phys_t *ddp = dde->dde_phys;
2660		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2661		blkptr_t blk;
2662
2663		ASSERT(zio->io_vsd == NULL);
2664		zio->io_vsd = dde;
2665
2666		if (ddp_self == NULL)
2667			return (ZIO_PIPELINE_CONTINUE);
2668
2669		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2670			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2671				continue;
2672			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2673			    &blk);
2674			zio_nowait(zio_read(zio, zio->io_spa, &blk,
2675			    abd_alloc_for_io(zio->io_size, B_TRUE),
2676			    zio->io_size, zio_ddt_child_read_done, dde,
2677			    zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2678			    ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2679		}
2680		return (ZIO_PIPELINE_CONTINUE);
2681	}
2682
2683	zio_nowait(zio_read(zio, zio->io_spa, bp,
2684	    zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2685	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2686
2687	return (ZIO_PIPELINE_CONTINUE);
2688}
2689
2690static int
2691zio_ddt_read_done(zio_t *zio)
2692{
2693	blkptr_t *bp = zio->io_bp;
2694
2695	if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2696		return (ZIO_PIPELINE_STOP);
2697	}
2698
2699	ASSERT(BP_GET_DEDUP(bp));
2700	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2701	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2702
2703	if (zio->io_child_error[ZIO_CHILD_DDT]) {
2704		ddt_t *ddt = ddt_select(zio->io_spa, bp);
2705		ddt_entry_t *dde = zio->io_vsd;
2706		if (ddt == NULL) {
2707			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2708			return (ZIO_PIPELINE_CONTINUE);
2709		}
2710		if (dde == NULL) {
2711			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2712			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2713			return (ZIO_PIPELINE_STOP);
2714		}
2715		if (dde->dde_repair_abd != NULL) {
2716			abd_copy(zio->io_abd, dde->dde_repair_abd,
2717			    zio->io_size);
2718			zio->io_child_error[ZIO_CHILD_DDT] = 0;
2719		}
2720		ddt_repair_done(ddt, dde);
2721		zio->io_vsd = NULL;
2722	}
2723
2724	ASSERT(zio->io_vsd == NULL);
2725
2726	return (ZIO_PIPELINE_CONTINUE);
2727}
2728
2729static boolean_t
2730zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2731{
2732	spa_t *spa = zio->io_spa;
2733	boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW);
2734
2735	/* We should never get a raw, override zio */
2736	ASSERT(!(zio->io_bp_override && do_raw));
2737
2738	/*
2739	 * Note: we compare the original data, not the transformed data,
2740	 * because when zio->io_bp is an override bp, we will not have
2741	 * pushed the I/O transforms.  That's an important optimization
2742	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2743	 * However, we should never get a raw, override zio so in these
2744	 * cases we can compare the io_data directly. This is useful because
2745	 * it allows us to do dedup verification even if we don't have access
2746	 * to the original data (for instance, if the encryption keys aren't
2747	 * loaded).
2748	 */
2749
2750	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2751		zio_t *lio = dde->dde_lead_zio[p];
2752
2753		if (lio != NULL && do_raw) {
2754			return (lio->io_size != zio->io_size ||
2755			    abd_cmp(zio->io_abd, lio->io_abd,
2756			    zio->io_size) != 0);
2757		} else if (lio != NULL) {
2758			return (lio->io_orig_size != zio->io_orig_size ||
2759			    abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2760			    zio->io_orig_size) != 0);
2761		}
2762	}
2763
2764	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2765		ddt_phys_t *ddp = &dde->dde_phys[p];
2766
2767		if (ddp->ddp_phys_birth != 0 && do_raw) {
2768			blkptr_t blk = *zio->io_bp;
2769			uint64_t psize;
2770			abd_t *tmpabd;
2771			int error;
2772
2773			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2774			psize = BP_GET_PSIZE(&blk);
2775
2776			if (psize != zio->io_size)
2777				return (B_TRUE);
2778
2779			ddt_exit(ddt);
2780
2781			tmpabd = abd_alloc_for_io(psize, B_TRUE);
2782
2783			error = zio_wait(zio_read(NULL, spa, &blk, tmpabd,
2784			    psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
2785			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2786			    ZIO_FLAG_RAW, &zio->io_bookmark));
2787
2788			if (error == 0) {
2789				if (abd_cmp(tmpabd, zio->io_abd, psize) != 0)
2790					error = SET_ERROR(ENOENT);
2791			}
2792
2793			abd_free(tmpabd);
2794			ddt_enter(ddt);
2795			return (error != 0);
2796		} else if (ddp->ddp_phys_birth != 0) {
2797			arc_buf_t *abuf = NULL;
2798			arc_flags_t aflags = ARC_FLAG_WAIT;
2799			int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2800			blkptr_t blk = *zio->io_bp;
2801			int error;
2802
2803			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2804
2805			if (BP_GET_LSIZE(&blk) != zio->io_orig_size)
2806				return (B_TRUE);
2807
2808			ddt_exit(ddt);
2809
2810			/*
2811			 * Intuitively, it would make more sense to compare
2812			 * io_abd than io_orig_abd in the raw case since you
2813			 * don't want to look at any transformations that have
2814			 * happened to the data. However, for raw I/Os the
2815			 * data will actually be the same in io_abd and
2816			 * io_orig_abd, so all we have to do is issue this as
2817			 * a raw ARC read.
2818			 */
2819			if (do_raw) {
2820				zio_flags |= ZIO_FLAG_RAW;
2821				ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2822				ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2823				    zio->io_size));
2824				ASSERT3P(zio->io_transform_stack, ==, NULL);
2825			}
2826
2827			error = arc_read(NULL, spa, &blk,
2828			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2829			    zio_flags, &aflags, &zio->io_bookmark);
2830
2831			if (error == 0) {
2832				if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2833				    zio->io_orig_size) != 0)
2834					error = SET_ERROR(ENOENT);
2835				arc_buf_destroy(abuf, &abuf);
2836			}
2837
2838			ddt_enter(ddt);
2839			return (error != 0);
2840		}
2841	}
2842
2843	return (B_FALSE);
2844}
2845
2846static void
2847zio_ddt_child_write_ready(zio_t *zio)
2848{
2849	int p = zio->io_prop.zp_copies;
2850	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2851	ddt_entry_t *dde = zio->io_private;
2852	ddt_phys_t *ddp = &dde->dde_phys[p];
2853	zio_t *pio;
2854
2855	if (zio->io_error)
2856		return;
2857
2858	ddt_enter(ddt);
2859
2860	ASSERT(dde->dde_lead_zio[p] == zio);
2861
2862	ddt_phys_fill(ddp, zio->io_bp);
2863
2864	zio_link_t *zl = NULL;
2865	while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2866		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2867
2868	ddt_exit(ddt);
2869}
2870
2871static void
2872zio_ddt_child_write_done(zio_t *zio)
2873{
2874	int p = zio->io_prop.zp_copies;
2875	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2876	ddt_entry_t *dde = zio->io_private;
2877	ddt_phys_t *ddp = &dde->dde_phys[p];
2878
2879	ddt_enter(ddt);
2880
2881	ASSERT(ddp->ddp_refcnt == 0);
2882	ASSERT(dde->dde_lead_zio[p] == zio);
2883	dde->dde_lead_zio[p] = NULL;
2884
2885	if (zio->io_error == 0) {
2886		zio_link_t *zl = NULL;
2887		while (zio_walk_parents(zio, &zl) != NULL)
2888			ddt_phys_addref(ddp);
2889	} else {
2890		ddt_phys_clear(ddp);
2891	}
2892
2893	ddt_exit(ddt);
2894}
2895
2896static void
2897zio_ddt_ditto_write_done(zio_t *zio)
2898{
2899	int p = DDT_PHYS_DITTO;
2900	zio_prop_t *zp = &zio->io_prop;
2901	blkptr_t *bp = zio->io_bp;
2902	ddt_t *ddt = ddt_select(zio->io_spa, bp);
2903	ddt_entry_t *dde = zio->io_private;
2904	ddt_phys_t *ddp = &dde->dde_phys[p];
2905	ddt_key_t *ddk = &dde->dde_key;
2906
2907	ddt_enter(ddt);
2908
2909	ASSERT(ddp->ddp_refcnt == 0);
2910	ASSERT(dde->dde_lead_zio[p] == zio);
2911	dde->dde_lead_zio[p] = NULL;
2912
2913	if (zio->io_error == 0) {
2914		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2915		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2916		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2917		if (ddp->ddp_phys_birth != 0)
2918			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2919		ddt_phys_fill(ddp, bp);
2920	}
2921
2922	ddt_exit(ddt);
2923}
2924
2925static int
2926zio_ddt_write(zio_t *zio)
2927{
2928	spa_t *spa = zio->io_spa;
2929	blkptr_t *bp = zio->io_bp;
2930	uint64_t txg = zio->io_txg;
2931	zio_prop_t *zp = &zio->io_prop;
2932	int p = zp->zp_copies;
2933	int ditto_copies;
2934	zio_t *cio = NULL;
2935	zio_t *dio = NULL;
2936	ddt_t *ddt = ddt_select(spa, bp);
2937	ddt_entry_t *dde;
2938	ddt_phys_t *ddp;
2939
2940	ASSERT(BP_GET_DEDUP(bp));
2941	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2942	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2943	ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2944
2945	ddt_enter(ddt);
2946	dde = ddt_lookup(ddt, bp, B_TRUE);
2947	ddp = &dde->dde_phys[p];
2948
2949	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2950		/*
2951		 * If we're using a weak checksum, upgrade to a strong checksum
2952		 * and try again.  If we're already using a strong checksum,
2953		 * we can't resolve it, so just convert to an ordinary write.
2954		 * (And automatically e-mail a paper to Nature?)
2955		 */
2956		if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2957		    ZCHECKSUM_FLAG_DEDUP)) {
2958			zp->zp_checksum = spa_dedup_checksum(spa);
2959			zio_pop_transforms(zio);
2960			zio->io_stage = ZIO_STAGE_OPEN;
2961			BP_ZERO(bp);
2962		} else {
2963			zp->zp_dedup = B_FALSE;
2964			BP_SET_DEDUP(bp, B_FALSE);
2965		}
2966		ASSERT(!BP_GET_DEDUP(bp));
2967		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2968		ddt_exit(ddt);
2969		return (ZIO_PIPELINE_CONTINUE);
2970	}
2971
2972	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2973	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2974
2975	if (ditto_copies > ddt_ditto_copies_present(dde) &&
2976	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2977		zio_prop_t czp = *zp;
2978
2979		czp.zp_copies = ditto_copies;
2980
2981		/*
2982		 * If we arrived here with an override bp, we won't have run
2983		 * the transform stack, so we won't have the data we need to
2984		 * generate a child i/o.  So, toss the override bp and restart.
2985		 * This is safe, because using the override bp is just an
2986		 * optimization; and it's rare, so the cost doesn't matter.
2987		 */
2988		if (zio->io_bp_override) {
2989			zio_pop_transforms(zio);
2990			zio->io_stage = ZIO_STAGE_OPEN;
2991			zio->io_pipeline = ZIO_WRITE_PIPELINE;
2992			zio->io_bp_override = NULL;
2993			BP_ZERO(bp);
2994			ddt_exit(ddt);
2995			return (ZIO_PIPELINE_CONTINUE);
2996		}
2997
2998		dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2999		    zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
3000		    NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
3001		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3002
3003		zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
3004		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
3005	}
3006
3007	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
3008		if (ddp->ddp_phys_birth != 0)
3009			ddt_bp_fill(ddp, bp, txg);
3010		if (dde->dde_lead_zio[p] != NULL)
3011			zio_add_child(zio, dde->dde_lead_zio[p]);
3012		else
3013			ddt_phys_addref(ddp);
3014	} else if (zio->io_bp_override) {
3015		ASSERT(bp->blk_birth == txg);
3016		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
3017		ddt_phys_fill(ddp, bp);
3018		ddt_phys_addref(ddp);
3019	} else {
3020		cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3021		    zio->io_orig_size, zio->io_orig_size, zp,
3022		    zio_ddt_child_write_ready, NULL, NULL,
3023		    zio_ddt_child_write_done, dde, zio->io_priority,
3024		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3025
3026		zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
3027		dde->dde_lead_zio[p] = cio;
3028	}
3029
3030	ddt_exit(ddt);
3031
3032	if (cio)
3033		zio_nowait(cio);
3034	if (dio)
3035		zio_nowait(dio);
3036
3037	return (ZIO_PIPELINE_CONTINUE);
3038}
3039
3040ddt_entry_t *freedde; /* for debugging */
3041
3042static int
3043zio_ddt_free(zio_t *zio)
3044{
3045	spa_t *spa = zio->io_spa;
3046	blkptr_t *bp = zio->io_bp;
3047	ddt_t *ddt = ddt_select(spa, bp);
3048	ddt_entry_t *dde;
3049	ddt_phys_t *ddp;
3050
3051	ASSERT(BP_GET_DEDUP(bp));
3052	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3053
3054	ddt_enter(ddt);
3055	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
3056	ddp = ddt_phys_select(dde, bp);
3057	ddt_phys_decref(ddp);
3058	ddt_exit(ddt);
3059
3060	return (ZIO_PIPELINE_CONTINUE);
3061}
3062
3063/*
3064 * ==========================================================================
3065 * Allocate and free blocks
3066 * ==========================================================================
3067 */
3068
3069static zio_t *
3070zio_io_to_allocate(spa_t *spa, int allocator)
3071{
3072	zio_t *zio;
3073
3074	ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator]));
3075
3076	zio = avl_first(&spa->spa_alloc_trees[allocator]);
3077	if (zio == NULL)
3078		return (NULL);
3079
3080	ASSERT(IO_IS_ALLOCATING(zio));
3081
3082	/*
3083	 * Try to place a reservation for this zio. If we're unable to
3084	 * reserve then we throttle.
3085	 */
3086	ASSERT3U(zio->io_allocator, ==, allocator);
3087	if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
3088	    zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) {
3089		return (NULL);
3090	}
3091
3092	avl_remove(&spa->spa_alloc_trees[allocator], zio);
3093	ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
3094
3095	return (zio);
3096}
3097
3098static int
3099zio_dva_throttle(zio_t *zio)
3100{
3101	spa_t *spa = zio->io_spa;
3102	zio_t *nio;
3103	metaslab_class_t *mc;
3104
3105	/* locate an appropriate allocation class */
3106	mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type,
3107	    zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk);
3108
3109	if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
3110	    !mc->mc_alloc_throttle_enabled ||
3111	    zio->io_child_type == ZIO_CHILD_GANG ||
3112	    zio->io_flags & ZIO_FLAG_NODATA) {
3113		return (ZIO_PIPELINE_CONTINUE);
3114	}
3115
3116	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3117
3118	ASSERT3U(zio->io_queued_timestamp, >, 0);
3119	ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3120
3121	zbookmark_phys_t *bm = &zio->io_bookmark;
3122	/*
3123	 * We want to try to use as many allocators as possible to help improve
3124	 * performance, but we also want logically adjacent IOs to be physically
3125	 * adjacent to improve sequential read performance. We chunk each object
3126	 * into 2^20 block regions, and then hash based on the objset, object,
3127	 * level, and region to accomplish both of these goals.
3128	 */
3129	zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object,
3130	    bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
3131	mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]);
3132	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3133	zio->io_metaslab_class = mc;
3134	avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio);
3135	nio = zio_io_to_allocate(spa, zio->io_allocator);
3136	mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]);
3137
3138	if (nio == zio)
3139		return (ZIO_PIPELINE_CONTINUE);
3140
3141	if (nio != NULL) {
3142		ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3143		/*
3144		 * We are passing control to a new zio so make sure that
3145		 * it is processed by a different thread. We do this to
3146		 * avoid stack overflows that can occur when parents are
3147		 * throttled and children are making progress. We allow
3148		 * it to go to the head of the taskq since it's already
3149		 * been waiting.
3150		 */
3151		zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
3152	}
3153	return (ZIO_PIPELINE_STOP);
3154}
3155
3156static void
3157zio_allocate_dispatch(spa_t *spa, int allocator)
3158{
3159	zio_t *zio;
3160
3161	mutex_enter(&spa->spa_alloc_locks[allocator]);
3162	zio = zio_io_to_allocate(spa, allocator);
3163	mutex_exit(&spa->spa_alloc_locks[allocator]);
3164	if (zio == NULL)
3165		return;
3166
3167	ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3168	ASSERT0(zio->io_error);
3169	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3170}
3171
3172static int
3173zio_dva_allocate(zio_t *zio)
3174{
3175	spa_t *spa = zio->io_spa;
3176	metaslab_class_t *mc;
3177	blkptr_t *bp = zio->io_bp;
3178	int error;
3179	int flags = 0;
3180
3181	if (zio->io_gang_leader == NULL) {
3182		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3183		zio->io_gang_leader = zio;
3184	}
3185
3186	ASSERT(BP_IS_HOLE(bp));
3187	ASSERT0(BP_GET_NDVAS(bp));
3188	ASSERT3U(zio->io_prop.zp_copies, >, 0);
3189	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3190	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3191
3192	if (zio->io_flags & ZIO_FLAG_NODATA)
3193		flags |= METASLAB_DONT_THROTTLE;
3194	if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
3195		flags |= METASLAB_GANG_CHILD;
3196	if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
3197		flags |= METASLAB_ASYNC_ALLOC;
3198
3199	/*
3200	 * if not already chosen, locate an appropriate allocation class
3201	 */
3202	mc = zio->io_metaslab_class;
3203	if (mc == NULL) {
3204		mc = spa_preferred_class(spa, zio->io_size,
3205		    zio->io_prop.zp_type, zio->io_prop.zp_level,
3206		    zio->io_prop.zp_zpl_smallblk);
3207		zio->io_metaslab_class = mc;
3208	}
3209
3210	error = metaslab_alloc(spa, mc, zio->io_size, bp,
3211	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3212	    &zio->io_alloc_list, zio, zio->io_allocator);
3213
3214	/*
3215	 * Fallback to normal class when an alloc class is full
3216	 */
3217	if (error == ENOSPC && mc != spa_normal_class(spa)) {
3218		/*
3219		 * If throttling, transfer reservation over to normal class.
3220		 * The io_allocator slot can remain the same even though we
3221		 * are switching classes.
3222		 */
3223		if (mc->mc_alloc_throttle_enabled &&
3224		    (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) {
3225			metaslab_class_throttle_unreserve(mc,
3226			    zio->io_prop.zp_copies, zio->io_allocator, zio);
3227			zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING;
3228
3229			mc = spa_normal_class(spa);
3230			VERIFY(metaslab_class_throttle_reserve(mc,
3231			    zio->io_prop.zp_copies, zio->io_allocator, zio,
3232			    flags | METASLAB_MUST_RESERVE));
3233		} else {
3234			mc = spa_normal_class(spa);
3235		}
3236		zio->io_metaslab_class = mc;
3237
3238		error = metaslab_alloc(spa, mc, zio->io_size, bp,
3239		    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3240		    &zio->io_alloc_list, zio, zio->io_allocator);
3241	}
3242
3243	if (error != 0) {
3244		zfs_dbgmsg("%s: metaslab allocation failure: zio %p, "
3245		    "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3246		    error);
3247		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
3248			return (zio_write_gang_block(zio));
3249		zio->io_error = error;
3250	}
3251
3252	return (ZIO_PIPELINE_CONTINUE);
3253}
3254
3255static int
3256zio_dva_free(zio_t *zio)
3257{
3258	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3259
3260	return (ZIO_PIPELINE_CONTINUE);
3261}
3262
3263static int
3264zio_dva_claim(zio_t *zio)
3265{
3266	int error;
3267
3268	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3269	if (error)
3270		zio->io_error = error;
3271
3272	return (ZIO_PIPELINE_CONTINUE);
3273}
3274
3275/*
3276 * Undo an allocation.  This is used by zio_done() when an I/O fails
3277 * and we want to give back the block we just allocated.
3278 * This handles both normal blocks and gang blocks.
3279 */
3280static void
3281zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3282{
3283	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3284	ASSERT(zio->io_bp_override == NULL);
3285
3286	if (!BP_IS_HOLE(bp))
3287		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3288
3289	if (gn != NULL) {
3290		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3291			zio_dva_unallocate(zio, gn->gn_child[g],
3292			    &gn->gn_gbh->zg_blkptr[g]);
3293		}
3294	}
3295}
3296
3297/*
3298 * Try to allocate an intent log block.  Return 0 on success, errno on failure.
3299 */
3300int
3301zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
3302    blkptr_t *old_bp, uint64_t size, boolean_t *slog)
3303{
3304	int error = 1;
3305	zio_alloc_list_t io_alloc_list;
3306
3307	ASSERT(txg > spa_syncing_txg(spa));
3308
3309	metaslab_trace_init(&io_alloc_list);
3310
3311	/*
3312	 * Block pointer fields are useful to metaslabs for stats and debugging.
3313	 * Fill in the obvious ones before calling into metaslab_alloc().
3314	 */
3315	BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3316	BP_SET_PSIZE(new_bp, size);
3317	BP_SET_LEVEL(new_bp, 0);
3318
3319	/*
3320	 * When allocating a zil block, we don't have information about
3321	 * the final destination of the block except the objset it's part
3322	 * of, so we just hash the objset ID to pick the allocator to get
3323	 * some parallelism.
3324	 */
3325	error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3326	    txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL,
3327	    cityhash4(0, 0, 0,
3328	    os->os_dsl_dataset->ds_object) % spa->spa_alloc_count);
3329	if (error == 0) {
3330		*slog = TRUE;
3331	} else {
3332		error = metaslab_alloc(spa, spa_normal_class(spa), size,
3333		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3334		    &io_alloc_list, NULL, cityhash4(0, 0, 0,
3335		    os->os_dsl_dataset->ds_object) % spa->spa_alloc_count);
3336		if (error == 0)
3337			*slog = FALSE;
3338	}
3339	metaslab_trace_fini(&io_alloc_list);
3340
3341	if (error == 0) {
3342		BP_SET_LSIZE(new_bp, size);
3343		BP_SET_PSIZE(new_bp, size);
3344		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3345		BP_SET_CHECKSUM(new_bp,
3346		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3347		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3348		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3349		BP_SET_LEVEL(new_bp, 0);
3350		BP_SET_DEDUP(new_bp, 0);
3351		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3352
3353		/*
3354		 * encrypted blocks will require an IV and salt. We generate
3355		 * these now since we will not be rewriting the bp at
3356		 * rewrite time.
3357		 */
3358		if (os->os_encrypted) {
3359			uint8_t iv[ZIO_DATA_IV_LEN];
3360			uint8_t salt[ZIO_DATA_SALT_LEN];
3361
3362			BP_SET_CRYPT(new_bp, B_TRUE);
3363			VERIFY0(spa_crypt_get_salt(spa,
3364			    dmu_objset_id(os), salt));
3365			VERIFY0(zio_crypt_generate_iv(iv));
3366
3367			zio_crypt_encode_params_bp(new_bp, salt, iv);
3368		}
3369	} else {
3370		zfs_dbgmsg("%s: zil block allocation failure: "
3371		    "size %llu, error %d", spa_name(spa), size, error);
3372	}
3373
3374	return (error);
3375}
3376
3377/*
3378 * ==========================================================================
3379 * Read and write to physical devices
3380 * ==========================================================================
3381 */
3382
3383/*
3384 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3385 * stops after this stage and will resume upon I/O completion.
3386 * However, there are instances where the vdev layer may need to
3387 * continue the pipeline when an I/O was not issued. Since the I/O
3388 * that was sent to the vdev layer might be different than the one
3389 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3390 * force the underlying vdev layers to call either zio_execute() or
3391 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3392 */
3393static int
3394zio_vdev_io_start(zio_t *zio)
3395{
3396	vdev_t *vd = zio->io_vd;
3397	uint64_t align;
3398	spa_t *spa = zio->io_spa;
3399
3400	zio->io_delay = 0;
3401
3402	ASSERT(zio->io_error == 0);
3403	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3404
3405	if (vd == NULL) {
3406		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3407			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3408
3409		/*
3410		 * The mirror_ops handle multiple DVAs in a single BP.
3411		 */
3412		vdev_mirror_ops.vdev_op_io_start(zio);
3413		return (ZIO_PIPELINE_STOP);
3414	}
3415
3416	ASSERT3P(zio->io_logical, !=, zio);
3417	if (zio->io_type == ZIO_TYPE_WRITE) {
3418		ASSERT(spa->spa_trust_config);
3419
3420		/*
3421		 * Note: the code can handle other kinds of writes,
3422		 * but we don't expect them.
3423		 */
3424		if (zio->io_vd->vdev_removing) {
3425			ASSERT(zio->io_flags &
3426			    (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3427			    ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3428		}
3429	}
3430
3431	align = 1ULL << vd->vdev_top->vdev_ashift;
3432
3433	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3434	    P2PHASE(zio->io_size, align) != 0) {
3435		/* Transform logical writes to be a full physical block size. */
3436		uint64_t asize = P2ROUNDUP(zio->io_size, align);
3437		abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3438		ASSERT(vd == vd->vdev_top);
3439		if (zio->io_type == ZIO_TYPE_WRITE) {
3440			abd_copy(abuf, zio->io_abd, zio->io_size);
3441			abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3442		}
3443		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3444	}
3445
3446	/*
3447	 * If this is not a physical io, make sure that it is properly aligned
3448	 * before proceeding.
3449	 */
3450	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3451		ASSERT0(P2PHASE(zio->io_offset, align));
3452		ASSERT0(P2PHASE(zio->io_size, align));
3453	} else {
3454		/*
3455		 * For physical writes, we allow 512b aligned writes and assume
3456		 * the device will perform a read-modify-write as necessary.
3457		 */
3458		ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3459		ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3460	}
3461
3462	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
3463
3464	/*
3465	 * If this is a repair I/O, and there's no self-healing involved --
3466	 * that is, we're just resilvering what we expect to resilver --
3467	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3468	 * This prevents spurious resilvering.
3469	 *
3470	 * There are a few ways that we can end up creating these spurious
3471	 * resilver i/os:
3472	 *
3473	 * 1. A resilver i/o will be issued if any DVA in the BP has a
3474	 * dirty DTL.  The mirror code will issue resilver writes to
3475	 * each DVA, including the one(s) that are not on vdevs with dirty
3476	 * DTLs.
3477	 *
3478	 * 2. With nested replication, which happens when we have a
3479	 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3480	 * For example, given mirror(replacing(A+B), C), it's likely that
3481	 * only A is out of date (it's the new device). In this case, we'll
3482	 * read from C, then use the data to resilver A+B -- but we don't
3483	 * actually want to resilver B, just A. The top-level mirror has no
3484	 * way to know this, so instead we just discard unnecessary repairs
3485	 * as we work our way down the vdev tree.
3486	 *
3487	 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3488	 * The same logic applies to any form of nested replication: ditto
3489	 * + mirror, RAID-Z + replacing, etc.
3490	 *
3491	 * However, indirect vdevs point off to other vdevs which may have
3492	 * DTL's, so we never bypass them.  The child i/os on concrete vdevs
3493	 * will be properly bypassed instead.
3494	 */
3495	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3496	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3497	    zio->io_txg != 0 &&	/* not a delegated i/o */
3498	    vd->vdev_ops != &vdev_indirect_ops &&
3499	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3500		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3501		zio_vdev_io_bypass(zio);
3502		return (ZIO_PIPELINE_CONTINUE);
3503	}
3504
3505	if (vd->vdev_ops->vdev_op_leaf && (zio->io_type == ZIO_TYPE_READ ||
3506	    zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM)) {
3507
3508		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
3509			return (ZIO_PIPELINE_CONTINUE);
3510
3511		if ((zio = vdev_queue_io(zio)) == NULL)
3512			return (ZIO_PIPELINE_STOP);
3513
3514		if (!vdev_accessible(vd, zio)) {
3515			zio->io_error = SET_ERROR(ENXIO);
3516			zio_interrupt(zio);
3517			return (ZIO_PIPELINE_STOP);
3518		}
3519		zio->io_delay = gethrtime();
3520	}
3521
3522	vd->vdev_ops->vdev_op_io_start(zio);
3523	return (ZIO_PIPELINE_STOP);
3524}
3525
3526static int
3527zio_vdev_io_done(zio_t *zio)
3528{
3529	vdev_t *vd = zio->io_vd;
3530	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3531	boolean_t unexpected_error = B_FALSE;
3532
3533	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3534		return (ZIO_PIPELINE_STOP);
3535	}
3536
3537	ASSERT(zio->io_type == ZIO_TYPE_READ ||
3538	    zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM);
3539
3540	if (zio->io_delay)
3541		zio->io_delay = gethrtime() - zio->io_delay;
3542
3543	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
3544
3545		vdev_queue_io_done(zio);
3546
3547		if (zio->io_type == ZIO_TYPE_WRITE)
3548			vdev_cache_write(zio);
3549
3550		if (zio_injection_enabled && zio->io_error == 0)
3551			zio->io_error = zio_handle_device_injection(vd,
3552			    zio, EIO);
3553
3554		if (zio_injection_enabled && zio->io_error == 0)
3555			zio->io_error = zio_handle_label_injection(zio, EIO);
3556
3557		if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) {
3558			if (!vdev_accessible(vd, zio)) {
3559				zio->io_error = SET_ERROR(ENXIO);
3560			} else {
3561				unexpected_error = B_TRUE;
3562			}
3563		}
3564	}
3565
3566	ops->vdev_op_io_done(zio);
3567
3568	if (unexpected_error)
3569		VERIFY(vdev_probe(vd, zio) == NULL);
3570
3571	return (ZIO_PIPELINE_CONTINUE);
3572}
3573
3574/*
3575 * This function is used to change the priority of an existing zio that is
3576 * currently in-flight. This is used by the arc to upgrade priority in the
3577 * event that a demand read is made for a block that is currently queued
3578 * as a scrub or async read IO. Otherwise, the high priority read request
3579 * would end up having to wait for the lower priority IO.
3580 */
3581void
3582zio_change_priority(zio_t *pio, zio_priority_t priority)
3583{
3584	zio_t *cio, *cio_next;
3585	zio_link_t *zl = NULL;
3586
3587	ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3588
3589	if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3590		vdev_queue_change_io_priority(pio, priority);
3591	} else {
3592		pio->io_priority = priority;
3593	}
3594
3595	mutex_enter(&pio->io_lock);
3596	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3597		cio_next = zio_walk_children(pio, &zl);
3598		zio_change_priority(cio, priority);
3599	}
3600	mutex_exit(&pio->io_lock);
3601}
3602
3603/*
3604 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3605 * disk, and use that to finish the checksum ereport later.
3606 */
3607static void
3608zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3609    const abd_t *good_buf)
3610{
3611	/* no processing needed */
3612	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3613}
3614
3615/*ARGSUSED*/
3616void
3617zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3618{
3619	void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size);
3620
3621	abd_copy(abd, zio->io_abd, zio->io_size);
3622
3623	zcr->zcr_cbinfo = zio->io_size;
3624	zcr->zcr_cbdata = abd;
3625	zcr->zcr_finish = zio_vsd_default_cksum_finish;
3626	zcr->zcr_free = zio_abd_free;
3627}
3628
3629static int
3630zio_vdev_io_assess(zio_t *zio)
3631{
3632	vdev_t *vd = zio->io_vd;
3633
3634	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3635		return (ZIO_PIPELINE_STOP);
3636	}
3637
3638	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3639		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3640
3641	if (zio->io_vsd != NULL) {
3642		zio->io_vsd_ops->vsd_free(zio);
3643		zio->io_vsd = NULL;
3644	}
3645
3646	if (zio_injection_enabled && zio->io_error == 0)
3647		zio->io_error = zio_handle_fault_injection(zio, EIO);
3648
3649	/*
3650	 * If the I/O failed, determine whether we should attempt to retry it.
3651	 *
3652	 * On retry, we cut in line in the issue queue, since we don't want
3653	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3654	 */
3655	if (zio->io_error && vd == NULL &&
3656	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3657		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
3658		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
3659		zio->io_error = 0;
3660		zio->io_flags |= ZIO_FLAG_IO_RETRY |
3661		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3662		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3663		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3664		    zio_requeue_io_start_cut_in_line);
3665		return (ZIO_PIPELINE_STOP);
3666	}
3667
3668	/*
3669	 * If we got an error on a leaf device, convert it to ENXIO
3670	 * if the device is not accessible at all.
3671	 */
3672	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3673	    !vdev_accessible(vd, zio))
3674		zio->io_error = SET_ERROR(ENXIO);
3675
3676	/*
3677	 * If we can't write to an interior vdev (mirror or RAID-Z),
3678	 * set vdev_cant_write so that we stop trying to allocate from it.
3679	 */
3680	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3681	    vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3682		vd->vdev_cant_write = B_TRUE;
3683	}
3684
3685	/*
3686	 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3687	 * attempts will ever succeed. In this case we set a persistent
3688	 * boolean flag so that we don't bother with it in the future.
3689	 */
3690	if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3691	    zio->io_type == ZIO_TYPE_IOCTL &&
3692	    zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3693		vd->vdev_nowritecache = B_TRUE;
3694
3695	if (zio->io_error)
3696		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3697
3698	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3699	    zio->io_physdone != NULL) {
3700		ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3701		ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3702		zio->io_physdone(zio->io_logical);
3703	}
3704
3705	return (ZIO_PIPELINE_CONTINUE);
3706}
3707
3708void
3709zio_vdev_io_reissue(zio_t *zio)
3710{
3711	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3712	ASSERT(zio->io_error == 0);
3713
3714	zio->io_stage >>= 1;
3715}
3716
3717void
3718zio_vdev_io_redone(zio_t *zio)
3719{
3720	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3721
3722	zio->io_stage >>= 1;
3723}
3724
3725void
3726zio_vdev_io_bypass(zio_t *zio)
3727{
3728	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3729	ASSERT(zio->io_error == 0);
3730
3731	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3732	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3733}
3734
3735/*
3736 * ==========================================================================
3737 * Encrypt and store encryption parameters
3738 * ==========================================================================
3739 */
3740
3741
3742/*
3743 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
3744 * managing the storage of encryption parameters and passing them to the
3745 * lower-level encryption functions.
3746 */
3747static int
3748zio_encrypt(zio_t *zio)
3749{
3750	zio_prop_t *zp = &zio->io_prop;
3751	spa_t *spa = zio->io_spa;
3752	blkptr_t *bp = zio->io_bp;
3753	uint64_t psize = BP_GET_PSIZE(bp);
3754	uint64_t dsobj = zio->io_bookmark.zb_objset;
3755	dmu_object_type_t ot = BP_GET_TYPE(bp);
3756	void *enc_buf = NULL;
3757	abd_t *eabd = NULL;
3758	uint8_t salt[ZIO_DATA_SALT_LEN];
3759	uint8_t iv[ZIO_DATA_IV_LEN];
3760	uint8_t mac[ZIO_DATA_MAC_LEN];
3761	boolean_t no_crypt = B_FALSE;
3762
3763	/* the root zio already encrypted the data */
3764	if (zio->io_child_type == ZIO_CHILD_GANG)
3765		return (ZIO_PIPELINE_CONTINUE);
3766
3767	/* only ZIL blocks are re-encrypted on rewrite */
3768	if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG)
3769		return (ZIO_PIPELINE_CONTINUE);
3770
3771	if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) {
3772		BP_SET_CRYPT(bp, B_FALSE);
3773		return (ZIO_PIPELINE_CONTINUE);
3774	}
3775
3776	/* if we are doing raw encryption set the provided encryption params */
3777	if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) {
3778		ASSERT0(BP_GET_LEVEL(bp));
3779		BP_SET_CRYPT(bp, B_TRUE);
3780		BP_SET_BYTEORDER(bp, zp->zp_byteorder);
3781		if (ot != DMU_OT_OBJSET)
3782			zio_crypt_encode_mac_bp(bp, zp->zp_mac);
3783
3784		/* dnode blocks must be written out in the provided byteorder */
3785		if (zp->zp_byteorder != ZFS_HOST_BYTEORDER &&
3786		    ot == DMU_OT_DNODE) {
3787			void *bswap_buf = zio_buf_alloc(psize);
3788			abd_t *babd = abd_get_from_buf(bswap_buf, psize);
3789
3790			ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
3791			abd_copy_to_buf(bswap_buf, zio->io_abd, psize);
3792			dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf,
3793			    psize);
3794
3795			abd_take_ownership_of_buf(babd, B_TRUE);
3796			zio_push_transform(zio, babd, psize, psize, NULL);
3797		}
3798
3799		if (DMU_OT_IS_ENCRYPTED(ot))
3800			zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv);
3801		return (ZIO_PIPELINE_CONTINUE);
3802	}
3803
3804	/* indirect blocks only maintain a cksum of the lower level MACs */
3805	if (BP_GET_LEVEL(bp) > 0) {
3806		BP_SET_CRYPT(bp, B_TRUE);
3807		VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE,
3808		    zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp),
3809		    mac));
3810		zio_crypt_encode_mac_bp(bp, mac);
3811		return (ZIO_PIPELINE_CONTINUE);
3812	}
3813
3814	/*
3815	 * Objset blocks are a special case since they have 2 256-bit MACs
3816	 * embedded within them.
3817	 */
3818	if (ot == DMU_OT_OBJSET) {
3819		ASSERT0(DMU_OT_IS_ENCRYPTED(ot));
3820		ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
3821		BP_SET_CRYPT(bp, B_TRUE);
3822		VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj,
3823		    zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp)));
3824		return (ZIO_PIPELINE_CONTINUE);
3825	}
3826
3827	/* unencrypted object types are only authenticated with a MAC */
3828	if (!DMU_OT_IS_ENCRYPTED(ot)) {
3829		BP_SET_CRYPT(bp, B_TRUE);
3830		VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj,
3831		    zio->io_abd, psize, mac));
3832		zio_crypt_encode_mac_bp(bp, mac);
3833		return (ZIO_PIPELINE_CONTINUE);
3834	}
3835
3836	/*
3837	 * Later passes of sync-to-convergence may decide to rewrite data
3838	 * in place to avoid more disk reallocations. This presents a problem
3839	 * for encryption because this consitutes rewriting the new data with
3840	 * the same encryption key and IV. However, this only applies to blocks
3841	 * in the MOS (particularly the spacemaps) and we do not encrypt the
3842	 * MOS. We assert that the zio is allocating or an intent log write
3843	 * to enforce this.
3844	 */
3845	ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG);
3846	ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG);
3847	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION));
3848	ASSERT3U(psize, !=, 0);
3849
3850	enc_buf = zio_buf_alloc(psize);
3851	eabd = abd_get_from_buf(enc_buf, psize);
3852	abd_take_ownership_of_buf(eabd, B_TRUE);
3853
3854	/*
3855	 * For an explanation of what encryption parameters are stored
3856	 * where, see the block comment in zio_crypt.c.
3857	 */
3858	if (ot == DMU_OT_INTENT_LOG) {
3859		zio_crypt_decode_params_bp(bp, salt, iv);
3860	} else {
3861		BP_SET_CRYPT(bp, B_TRUE);
3862	}
3863
3864	/* Perform the encryption. This should not fail */
3865	VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark,
3866	    BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
3867	    salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt));
3868
3869	/* encode encryption metadata into the bp */
3870	if (ot == DMU_OT_INTENT_LOG) {
3871		/*
3872		 * ZIL blocks store the MAC in the embedded checksum, so the
3873		 * transform must always be applied.
3874		 */
3875		zio_crypt_encode_mac_zil(enc_buf, mac);
3876		zio_push_transform(zio, eabd, psize, psize, NULL);
3877	} else {
3878		BP_SET_CRYPT(bp, B_TRUE);
3879		zio_crypt_encode_params_bp(bp, salt, iv);
3880		zio_crypt_encode_mac_bp(bp, mac);
3881
3882		if (no_crypt) {
3883			ASSERT3U(ot, ==, DMU_OT_DNODE);
3884			abd_free(eabd);
3885		} else {
3886			zio_push_transform(zio, eabd, psize, psize, NULL);
3887		}
3888	}
3889
3890	return (ZIO_PIPELINE_CONTINUE);
3891}
3892
3893/*
3894 * ==========================================================================
3895 * Generate and verify checksums
3896 * ==========================================================================
3897 */
3898static int
3899zio_checksum_generate(zio_t *zio)
3900{
3901	blkptr_t *bp = zio->io_bp;
3902	enum zio_checksum checksum;
3903
3904	if (bp == NULL) {
3905		/*
3906		 * This is zio_write_phys().
3907		 * We're either generating a label checksum, or none at all.
3908		 */
3909		checksum = zio->io_prop.zp_checksum;
3910
3911		if (checksum == ZIO_CHECKSUM_OFF)
3912			return (ZIO_PIPELINE_CONTINUE);
3913
3914		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3915	} else {
3916		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3917			ASSERT(!IO_IS_ALLOCATING(zio));
3918			checksum = ZIO_CHECKSUM_GANG_HEADER;
3919		} else {
3920			checksum = BP_GET_CHECKSUM(bp);
3921		}
3922	}
3923
3924	zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
3925
3926	return (ZIO_PIPELINE_CONTINUE);
3927}
3928
3929static int
3930zio_checksum_verify(zio_t *zio)
3931{
3932	zio_bad_cksum_t info;
3933	blkptr_t *bp = zio->io_bp;
3934	int error;
3935
3936	ASSERT(zio->io_vd != NULL);
3937
3938	if (bp == NULL) {
3939		/*
3940		 * This is zio_read_phys().
3941		 * We're either verifying a label checksum, or nothing at all.
3942		 */
3943		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3944			return (ZIO_PIPELINE_CONTINUE);
3945
3946		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3947	}
3948
3949	if ((error = zio_checksum_error(zio, &info)) != 0) {
3950		zio->io_error = error;
3951		if (error == ECKSUM &&
3952		    !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3953			zfs_ereport_start_checksum(zio->io_spa,
3954			    zio->io_vd, &zio->io_bookmark, zio,
3955			    zio->io_offset, zio->io_size, NULL, &info);
3956		}
3957	}
3958
3959	return (ZIO_PIPELINE_CONTINUE);
3960}
3961
3962/*
3963 * Called by RAID-Z to ensure we don't compute the checksum twice.
3964 */
3965void
3966zio_checksum_verified(zio_t *zio)
3967{
3968	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3969}
3970
3971/*
3972 * ==========================================================================
3973 * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3974 * An error of 0 indicates success.  ENXIO indicates whole-device failure,
3975 * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
3976 * indicate errors that are specific to one I/O, and most likely permanent.
3977 * Any other error is presumed to be worse because we weren't expecting it.
3978 * ==========================================================================
3979 */
3980int
3981zio_worst_error(int e1, int e2)
3982{
3983	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3984	int r1, r2;
3985
3986	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3987		if (e1 == zio_error_rank[r1])
3988			break;
3989
3990	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3991		if (e2 == zio_error_rank[r2])
3992			break;
3993
3994	return (r1 > r2 ? e1 : e2);
3995}
3996
3997/*
3998 * ==========================================================================
3999 * I/O completion
4000 * ==========================================================================
4001 */
4002static int
4003zio_ready(zio_t *zio)
4004{
4005	blkptr_t *bp = zio->io_bp;
4006	zio_t *pio, *pio_next;
4007	zio_link_t *zl = NULL;
4008
4009	if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
4010	    ZIO_WAIT_READY)) {
4011		return (ZIO_PIPELINE_STOP);
4012	}
4013
4014	if (zio->io_ready) {
4015		ASSERT(IO_IS_ALLOCATING(zio));
4016		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
4017		    (zio->io_flags & ZIO_FLAG_NOPWRITE));
4018		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
4019
4020		zio->io_ready(zio);
4021	}
4022
4023	if (bp != NULL && bp != &zio->io_bp_copy)
4024		zio->io_bp_copy = *bp;
4025
4026	if (zio->io_error != 0) {
4027		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4028
4029		if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4030			ASSERT(IO_IS_ALLOCATING(zio));
4031			ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4032			ASSERT(zio->io_metaslab_class != NULL);
4033
4034			/*
4035			 * We were unable to allocate anything, unreserve and
4036			 * issue the next I/O to allocate.
4037			 */
4038			metaslab_class_throttle_unreserve(
4039			    zio->io_metaslab_class, zio->io_prop.zp_copies,
4040			    zio->io_allocator, zio);
4041			zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
4042		}
4043	}
4044
4045	mutex_enter(&zio->io_lock);
4046	zio->io_state[ZIO_WAIT_READY] = 1;
4047	pio = zio_walk_parents(zio, &zl);
4048	mutex_exit(&zio->io_lock);
4049
4050	/*
4051	 * As we notify zio's parents, new parents could be added.
4052	 * New parents go to the head of zio's io_parent_list, however,
4053	 * so we will (correctly) not notify them.  The remainder of zio's
4054	 * io_parent_list, from 'pio_next' onward, cannot change because
4055	 * all parents must wait for us to be done before they can be done.
4056	 */
4057	for (; pio != NULL; pio = pio_next) {
4058		pio_next = zio_walk_parents(zio, &zl);
4059		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
4060	}
4061
4062	if (zio->io_flags & ZIO_FLAG_NODATA) {
4063		if (BP_IS_GANG(bp)) {
4064			zio->io_flags &= ~ZIO_FLAG_NODATA;
4065		} else {
4066			ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
4067			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
4068		}
4069	}
4070
4071	if (zio_injection_enabled &&
4072	    zio->io_spa->spa_syncing_txg == zio->io_txg)
4073		zio_handle_ignored_writes(zio);
4074
4075	return (ZIO_PIPELINE_CONTINUE);
4076}
4077
4078/*
4079 * Update the allocation throttle accounting.
4080 */
4081static void
4082zio_dva_throttle_done(zio_t *zio)
4083{
4084	zio_t *lio = zio->io_logical;
4085	zio_t *pio = zio_unique_parent(zio);
4086	vdev_t *vd = zio->io_vd;
4087	int flags = METASLAB_ASYNC_ALLOC;
4088
4089	ASSERT3P(zio->io_bp, !=, NULL);
4090	ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
4091	ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
4092	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
4093	ASSERT(vd != NULL);
4094	ASSERT3P(vd, ==, vd->vdev_top);
4095	ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
4096	ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
4097	ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
4098	ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
4099
4100	/*
4101	 * Parents of gang children can have two flavors -- ones that
4102	 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4103	 * and ones that allocated the constituent blocks. The allocation
4104	 * throttle needs to know the allocating parent zio so we must find
4105	 * it here.
4106	 */
4107	if (pio->io_child_type == ZIO_CHILD_GANG) {
4108		/*
4109		 * If our parent is a rewrite gang child then our grandparent
4110		 * would have been the one that performed the allocation.
4111		 */
4112		if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
4113			pio = zio_unique_parent(pio);
4114		flags |= METASLAB_GANG_CHILD;
4115	}
4116
4117	ASSERT(IO_IS_ALLOCATING(pio));
4118	ASSERT3P(zio, !=, zio->io_logical);
4119	ASSERT(zio->io_logical != NULL);
4120	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4121	ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
4122	ASSERT(zio->io_metaslab_class != NULL);
4123
4124	mutex_enter(&pio->io_lock);
4125	metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
4126	    pio->io_allocator, B_TRUE);
4127	mutex_exit(&pio->io_lock);
4128
4129	metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1,
4130	    pio->io_allocator, pio);
4131
4132	/*
4133	 * Call into the pipeline to see if there is more work that
4134	 * needs to be done. If there is work to be done it will be
4135	 * dispatched to another taskq thread.
4136	 */
4137	zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
4138}
4139
4140static int
4141zio_done(zio_t *zio)
4142{
4143	spa_t *spa = zio->io_spa;
4144	zio_t *lio = zio->io_logical;
4145	blkptr_t *bp = zio->io_bp;
4146	vdev_t *vd = zio->io_vd;
4147	uint64_t psize = zio->io_size;
4148	zio_t *pio, *pio_next;
4149	zio_link_t *zl = NULL;
4150
4151	/*
4152	 * If our children haven't all completed,
4153	 * wait for them and then repeat this pipeline stage.
4154	 */
4155	if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
4156		return (ZIO_PIPELINE_STOP);
4157	}
4158
4159	/*
4160	 * If the allocation throttle is enabled, then update the accounting.
4161	 * We only track child I/Os that are part of an allocating async
4162	 * write. We must do this since the allocation is performed
4163	 * by the logical I/O but the actual write is done by child I/Os.
4164	 */
4165	if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
4166	    zio->io_child_type == ZIO_CHILD_VDEV) {
4167		ASSERT(zio->io_metaslab_class != NULL);
4168		ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled);
4169		zio_dva_throttle_done(zio);
4170	}
4171
4172	/*
4173	 * If the allocation throttle is enabled, verify that
4174	 * we have decremented the refcounts for every I/O that was throttled.
4175	 */
4176	if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4177		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4178		ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4179		ASSERT(bp != NULL);
4180
4181		metaslab_group_alloc_verify(spa, zio->io_bp, zio,
4182		    zio->io_allocator);
4183		VERIFY(zfs_refcount_not_held(
4184		    &zio->io_metaslab_class->mc_alloc_slots[zio->io_allocator],
4185		    zio));
4186	}
4187
4188	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
4189		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
4190			ASSERT(zio->io_children[c][w] == 0);
4191
4192	if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
4193		ASSERT(bp->blk_pad[0] == 0);
4194		ASSERT(bp->blk_pad[1] == 0);
4195		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
4196		    (bp == zio_unique_parent(zio)->io_bp));
4197		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
4198		    zio->io_bp_override == NULL &&
4199		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
4200			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
4201			ASSERT(BP_COUNT_GANG(bp) == 0 ||
4202			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
4203		}
4204		if (zio->io_flags & ZIO_FLAG_NOPWRITE)
4205			VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
4206	}
4207
4208	/*
4209	 * If there were child vdev/gang/ddt errors, they apply to us now.
4210	 */
4211	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
4212	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
4213	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
4214
4215	/*
4216	 * If the I/O on the transformed data was successful, generate any
4217	 * checksum reports now while we still have the transformed data.
4218	 */
4219	if (zio->io_error == 0) {
4220		while (zio->io_cksum_report != NULL) {
4221			zio_cksum_report_t *zcr = zio->io_cksum_report;
4222			uint64_t align = zcr->zcr_align;
4223			uint64_t asize = P2ROUNDUP(psize, align);
4224			abd_t *adata = zio->io_abd;
4225
4226			if (asize != psize) {
4227				adata = abd_alloc(asize, B_TRUE);
4228				abd_copy(adata, zio->io_abd, psize);
4229				abd_zero_off(adata, psize, asize - psize);
4230			}
4231
4232			zio->io_cksum_report = zcr->zcr_next;
4233			zcr->zcr_next = NULL;
4234			zcr->zcr_finish(zcr, adata);
4235			zfs_ereport_free_checksum(zcr);
4236
4237			if (asize != psize)
4238				abd_free(adata);
4239		}
4240	}
4241
4242	zio_pop_transforms(zio);	/* note: may set zio->io_error */
4243
4244	vdev_stat_update(zio, psize);
4245
4246	if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) {
4247		if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) {
4248			/*
4249			 * We want to only increment our slow IO counters if
4250			 * the IO is valid (i.e. not if the drive is removed).
4251			 *
4252			 * zfs_ereport_post() will also do these checks, but
4253			 * it can also have other failures, so we need to
4254			 * increment the slow_io counters independent of it.
4255			 */
4256			if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY,
4257			    zio->io_spa, zio->io_vd, zio)) {
4258				mutex_enter(&zio->io_vd->vdev_stat_lock);
4259				zio->io_vd->vdev_stat.vs_slow_ios++;
4260				mutex_exit(&zio->io_vd->vdev_stat_lock);
4261
4262				(void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY,
4263				    zio->io_spa, zio->io_vd, &zio->io_bookmark,
4264				    zio, 0, 0);
4265			}
4266		}
4267	}
4268
4269	if (zio->io_error) {
4270		/*
4271		 * If this I/O is attached to a particular vdev,
4272		 * generate an error message describing the I/O failure
4273		 * at the block level.  We ignore these errors if the
4274		 * device is currently unavailable.
4275		 */
4276		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
4277			(void) zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd,
4278			    &zio->io_bookmark, zio, 0, 0);
4279
4280		if ((zio->io_error == EIO || !(zio->io_flags &
4281		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
4282		    zio == lio) {
4283			/*
4284			 * For logical I/O requests, tell the SPA to log the
4285			 * error and generate a logical data ereport.
4286			 */
4287			spa_log_error(spa, &zio->io_bookmark);
4288			(void) zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL,
4289			    &zio->io_bookmark, zio, 0, 0);
4290		}
4291	}
4292
4293	if (zio->io_error && zio == lio) {
4294		/*
4295		 * Determine whether zio should be reexecuted.  This will
4296		 * propagate all the way to the root via zio_notify_parent().
4297		 */
4298		ASSERT(vd == NULL && bp != NULL);
4299		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4300
4301		if (IO_IS_ALLOCATING(zio) &&
4302		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
4303			if (zio->io_error != ENOSPC)
4304				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
4305			else
4306				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4307		}
4308
4309		if ((zio->io_type == ZIO_TYPE_READ ||
4310		    zio->io_type == ZIO_TYPE_FREE) &&
4311		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
4312		    zio->io_error == ENXIO &&
4313		    spa_load_state(spa) == SPA_LOAD_NONE &&
4314		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
4315			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4316
4317		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
4318			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4319
4320		/*
4321		 * Here is a possibly good place to attempt to do
4322		 * either combinatorial reconstruction or error correction
4323		 * based on checksums.  It also might be a good place
4324		 * to send out preliminary ereports before we suspend
4325		 * processing.
4326		 */
4327	}
4328
4329	/*
4330	 * If there were logical child errors, they apply to us now.
4331	 * We defer this until now to avoid conflating logical child
4332	 * errors with errors that happened to the zio itself when
4333	 * updating vdev stats and reporting FMA events above.
4334	 */
4335	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
4336
4337	if ((zio->io_error || zio->io_reexecute) &&
4338	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
4339	    !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
4340		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
4341
4342	zio_gang_tree_free(&zio->io_gang_tree);
4343
4344	/*
4345	 * Godfather I/Os should never suspend.
4346	 */
4347	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4348	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4349		zio->io_reexecute = 0;
4350
4351	if (zio->io_reexecute) {
4352		/*
4353		 * This is a logical I/O that wants to reexecute.
4354		 *
4355		 * Reexecute is top-down.  When an i/o fails, if it's not
4356		 * the root, it simply notifies its parent and sticks around.
4357		 * The parent, seeing that it still has children in zio_done(),
4358		 * does the same.  This percolates all the way up to the root.
4359		 * The root i/o will reexecute or suspend the entire tree.
4360		 *
4361		 * This approach ensures that zio_reexecute() honors
4362		 * all the original i/o dependency relationships, e.g.
4363		 * parents not executing until children are ready.
4364		 */
4365		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4366
4367		zio->io_gang_leader = NULL;
4368
4369		mutex_enter(&zio->io_lock);
4370		zio->io_state[ZIO_WAIT_DONE] = 1;
4371		mutex_exit(&zio->io_lock);
4372
4373		/*
4374		 * "The Godfather" I/O monitors its children but is
4375		 * not a true parent to them. It will track them through
4376		 * the pipeline but severs its ties whenever they get into
4377		 * trouble (e.g. suspended). This allows "The Godfather"
4378		 * I/O to return status without blocking.
4379		 */
4380		zl = NULL;
4381		for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4382		    pio = pio_next) {
4383			zio_link_t *remove_zl = zl;
4384			pio_next = zio_walk_parents(zio, &zl);
4385
4386			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4387			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4388				zio_remove_child(pio, zio, remove_zl);
4389				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4390			}
4391		}
4392
4393		if ((pio = zio_unique_parent(zio)) != NULL) {
4394			/*
4395			 * We're not a root i/o, so there's nothing to do
4396			 * but notify our parent.  Don't propagate errors
4397			 * upward since we haven't permanently failed yet.
4398			 */
4399			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4400			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4401			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4402		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4403			/*
4404			 * We'd fail again if we reexecuted now, so suspend
4405			 * until conditions improve (e.g. device comes online).
4406			 */
4407			zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
4408		} else {
4409			/*
4410			 * Reexecution is potentially a huge amount of work.
4411			 * Hand it off to the otherwise-unused claim taskq.
4412			 */
4413			ASSERT(zio->io_tqent.tqent_next == NULL);
4414			spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
4415			    ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
4416			    0, &zio->io_tqent);
4417		}
4418		return (ZIO_PIPELINE_STOP);
4419	}
4420
4421	ASSERT(zio->io_child_count == 0);
4422	ASSERT(zio->io_reexecute == 0);
4423	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4424
4425	/*
4426	 * Report any checksum errors, since the I/O is complete.
4427	 */
4428	while (zio->io_cksum_report != NULL) {
4429		zio_cksum_report_t *zcr = zio->io_cksum_report;
4430		zio->io_cksum_report = zcr->zcr_next;
4431		zcr->zcr_next = NULL;
4432		zcr->zcr_finish(zcr, NULL);
4433		zfs_ereport_free_checksum(zcr);
4434	}
4435
4436	/*
4437	 * It is the responsibility of the done callback to ensure that this
4438	 * particular zio is no longer discoverable for adoption, and as
4439	 * such, cannot acquire any new parents.
4440	 */
4441	if (zio->io_done)
4442		zio->io_done(zio);
4443
4444	mutex_enter(&zio->io_lock);
4445	zio->io_state[ZIO_WAIT_DONE] = 1;
4446	mutex_exit(&zio->io_lock);
4447
4448	zl = NULL;
4449	for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4450		zio_link_t *remove_zl = zl;
4451		pio_next = zio_walk_parents(zio, &zl);
4452		zio_remove_child(pio, zio, remove_zl);
4453		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4454	}
4455
4456	if (zio->io_waiter != NULL) {
4457		mutex_enter(&zio->io_lock);
4458		zio->io_executor = NULL;
4459		cv_broadcast(&zio->io_cv);
4460		mutex_exit(&zio->io_lock);
4461	} else {
4462		zio_destroy(zio);
4463	}
4464
4465	return (ZIO_PIPELINE_STOP);
4466}
4467
4468/*
4469 * ==========================================================================
4470 * I/O pipeline definition
4471 * ==========================================================================
4472 */
4473static zio_pipe_stage_t *zio_pipeline[] = {
4474	NULL,
4475	zio_read_bp_init,
4476	zio_write_bp_init,
4477	zio_free_bp_init,
4478	zio_issue_async,
4479	zio_write_compress,
4480	zio_encrypt,
4481	zio_checksum_generate,
4482	zio_nop_write,
4483	zio_ddt_read_start,
4484	zio_ddt_read_done,
4485	zio_ddt_write,
4486	zio_ddt_free,
4487	zio_gang_assemble,
4488	zio_gang_issue,
4489	zio_dva_throttle,
4490	zio_dva_allocate,
4491	zio_dva_free,
4492	zio_dva_claim,
4493	zio_ready,
4494	zio_vdev_io_start,
4495	zio_vdev_io_done,
4496	zio_vdev_io_assess,
4497	zio_checksum_verify,
4498	zio_done
4499};
4500
4501
4502
4503
4504/*
4505 * Compare two zbookmark_phys_t's to see which we would reach first in a
4506 * pre-order traversal of the object tree.
4507 *
4508 * This is simple in every case aside from the meta-dnode object. For all other
4509 * objects, we traverse them in order (object 1 before object 2, and so on).
4510 * However, all of these objects are traversed while traversing object 0, since
4511 * the data it points to is the list of objects.  Thus, we need to convert to a
4512 * canonical representation so we can compare meta-dnode bookmarks to
4513 * non-meta-dnode bookmarks.
4514 *
4515 * We do this by calculating "equivalents" for each field of the zbookmark.
4516 * zbookmarks outside of the meta-dnode use their own object and level, and
4517 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4518 * blocks this bookmark refers to) by multiplying their blkid by their span
4519 * (the number of L0 blocks contained within one block at their level).
4520 * zbookmarks inside the meta-dnode calculate their object equivalent
4521 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4522 * level + 1<<31 (any value larger than a level could ever be) for their level.
4523 * This causes them to always compare before a bookmark in their object
4524 * equivalent, compare appropriately to bookmarks in other objects, and to
4525 * compare appropriately to other bookmarks in the meta-dnode.
4526 */
4527int
4528zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4529    const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4530{
4531	/*
4532	 * These variables represent the "equivalent" values for the zbookmark,
4533	 * after converting zbookmarks inside the meta dnode to their
4534	 * normal-object equivalents.
4535	 */
4536	uint64_t zb1obj, zb2obj;
4537	uint64_t zb1L0, zb2L0;
4538	uint64_t zb1level, zb2level;
4539
4540	if (zb1->zb_object == zb2->zb_object &&
4541	    zb1->zb_level == zb2->zb_level &&
4542	    zb1->zb_blkid == zb2->zb_blkid)
4543		return (0);
4544
4545	/*
4546	 * BP_SPANB calculates the span in blocks.
4547	 */
4548	zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4549	zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4550
4551	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4552		zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4553		zb1L0 = 0;
4554		zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4555	} else {
4556		zb1obj = zb1->zb_object;
4557		zb1level = zb1->zb_level;
4558	}
4559
4560	if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4561		zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4562		zb2L0 = 0;
4563		zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4564	} else {
4565		zb2obj = zb2->zb_object;
4566		zb2level = zb2->zb_level;
4567	}
4568
4569	/* Now that we have a canonical representation, do the comparison. */
4570	if (zb1obj != zb2obj)
4571		return (zb1obj < zb2obj ? -1 : 1);
4572	else if (zb1L0 != zb2L0)
4573		return (zb1L0 < zb2L0 ? -1 : 1);
4574	else if (zb1level != zb2level)
4575		return (zb1level > zb2level ? -1 : 1);
4576	/*
4577	 * This can (theoretically) happen if the bookmarks have the same object
4578	 * and level, but different blkids, if the block sizes are not the same.
4579	 * There is presently no way to change the indirect block sizes
4580	 */
4581	return (0);
4582}
4583
4584/*
4585 *  This function checks the following: given that last_block is the place that
4586 *  our traversal stopped last time, does that guarantee that we've visited
4587 *  every node under subtree_root?  Therefore, we can't just use the raw output
4588 *  of zbookmark_compare.  We have to pass in a modified version of
4589 *  subtree_root; by incrementing the block id, and then checking whether
4590 *  last_block is before or equal to that, we can tell whether or not having
4591 *  visited last_block implies that all of subtree_root's children have been
4592 *  visited.
4593 */
4594boolean_t
4595zbookmark_subtree_completed(const dnode_phys_t *dnp,
4596    const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4597{
4598	zbookmark_phys_t mod_zb = *subtree_root;
4599	mod_zb.zb_blkid++;
4600	ASSERT(last_block->zb_level == 0);
4601
4602	/* The objset_phys_t isn't before anything. */
4603	if (dnp == NULL)
4604		return (B_FALSE);
4605
4606	/*
4607	 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4608	 * data block size in sectors, because that variable is only used if
4609	 * the bookmark refers to a block in the meta-dnode.  Since we don't
4610	 * know without examining it what object it refers to, and there's no
4611	 * harm in passing in this value in other cases, we always pass it in.
4612	 *
4613	 * We pass in 0 for the indirect block size shift because zb2 must be
4614	 * level 0.  The indirect block size is only used to calculate the span
4615	 * of the bookmark, but since the bookmark must be level 0, the span is
4616	 * always 1, so the math works out.
4617	 *
4618	 * If you make changes to how the zbookmark_compare code works, be sure
4619	 * to make sure that this code still works afterwards.
4620	 */
4621	return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4622	    1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
4623	    last_block) <= 0);
4624}
4625