1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
26 * Copyright (c) 2017, Intel Corporation.
27 */
28
29#include <sys/sysmacros.h>
30#include <sys/zfs_context.h>
31#include <sys/fm/fs/zfs.h>
32#include <sys/spa.h>
33#include <sys/txg.h>
34#include <sys/spa_impl.h>
35#include <sys/vdev_impl.h>
36#include <sys/vdev_trim.h>
37#include <sys/zio_impl.h>
38#include <sys/zio_compress.h>
39#include <sys/zio_checksum.h>
40#include <sys/dmu_objset.h>
41#include <sys/arc.h>
42#include <sys/ddt.h>
43#include <sys/blkptr.h>
44#include <sys/zfeature.h>
45#include <sys/dsl_scan.h>
46#include <sys/metaslab_impl.h>
47#include <sys/abd.h>
48#include <sys/cityhash.h>
49#include <sys/dsl_crypt.h>
50
51/*
52 * ==========================================================================
53 * I/O type descriptions
54 * ==========================================================================
55 */
56const char *zio_type_name[ZIO_TYPES] = {
57	"zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
58	"zio_ioctl", "z_trim"
59};
60
61boolean_t zio_dva_throttle_enabled = B_TRUE;
62
63/*
64 * ==========================================================================
65 * I/O kmem caches
66 * ==========================================================================
67 */
68kmem_cache_t *zio_cache;
69kmem_cache_t *zio_link_cache;
70kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
71kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
72
73#ifdef _KERNEL
74extern vmem_t *zio_alloc_arena;
75#endif
76
77#define	ZIO_PIPELINE_CONTINUE		0x100
78#define	ZIO_PIPELINE_STOP		0x101
79
80#define	BP_SPANB(indblkshift, level) \
81	(((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
82#define	COMPARE_META_LEVEL	0x80000000ul
83/*
84 * The following actions directly effect the spa's sync-to-convergence logic.
85 * The values below define the sync pass when we start performing the action.
86 * Care should be taken when changing these values as they directly impact
87 * spa_sync() performance. Tuning these values may introduce subtle performance
88 * pathologies and should only be done in the context of performance analysis.
89 * These tunables will eventually be removed and replaced with #defines once
90 * enough analysis has been done to determine optimal values.
91 *
92 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
93 * regular blocks are not deferred.
94 */
95int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
96int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
97int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
98
99/*
100 * An allocating zio is one that either currently has the DVA allocate
101 * stage set or will have it later in its lifetime.
102 */
103#define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
104
105boolean_t	zio_requeue_io_start_cut_in_line = B_TRUE;
106
107#ifdef ZFS_DEBUG
108int zio_buf_debug_limit = 16384;
109#else
110int zio_buf_debug_limit = 0;
111#endif
112
113static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
114
115void
116zio_init(void)
117{
118	size_t c;
119	vmem_t *data_alloc_arena = NULL;
120
121#ifdef _KERNEL
122	data_alloc_arena = zio_alloc_arena;
123#endif
124	zio_cache = kmem_cache_create("zio_cache",
125	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
126	zio_link_cache = kmem_cache_create("zio_link_cache",
127	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
128
129	/*
130	 * For small buffers, we want a cache for each multiple of
131	 * SPA_MINBLOCKSIZE.  For larger buffers, we want a cache
132	 * for each quarter-power of 2.
133	 */
134	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
135		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
136		size_t p2 = size;
137		size_t align = 0;
138		size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
139
140		while (!ISP2(p2))
141			p2 &= p2 - 1;
142
143#ifndef _KERNEL
144		/*
145		 * If we are using watchpoints, put each buffer on its own page,
146		 * to eliminate the performance overhead of trapping to the
147		 * kernel when modifying a non-watched buffer that shares the
148		 * page with a watched buffer.
149		 */
150		if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
151			continue;
152#endif
153		if (size <= 4 * SPA_MINBLOCKSIZE) {
154			align = SPA_MINBLOCKSIZE;
155		} else if (IS_P2ALIGNED(size, p2 >> 2)) {
156			align = MIN(p2 >> 2, PAGESIZE);
157		}
158
159		if (align != 0) {
160			char name[36];
161			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
162			zio_buf_cache[c] = kmem_cache_create(name, size,
163			    align, NULL, NULL, NULL, NULL, NULL, cflags);
164
165			/*
166			 * Since zio_data bufs do not appear in crash dumps, we
167			 * pass KMC_NOTOUCH so that no allocator metadata is
168			 * stored with the buffers.
169			 */
170			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
171			zio_data_buf_cache[c] = kmem_cache_create(name, size,
172			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
173			    cflags | KMC_NOTOUCH);
174		}
175	}
176
177	while (--c != 0) {
178		ASSERT(zio_buf_cache[c] != NULL);
179		if (zio_buf_cache[c - 1] == NULL)
180			zio_buf_cache[c - 1] = zio_buf_cache[c];
181
182		ASSERT(zio_data_buf_cache[c] != NULL);
183		if (zio_data_buf_cache[c - 1] == NULL)
184			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
185	}
186
187	zio_inject_init();
188}
189
190void
191zio_fini(void)
192{
193	size_t c;
194	kmem_cache_t *last_cache = NULL;
195	kmem_cache_t *last_data_cache = NULL;
196
197	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
198		if (zio_buf_cache[c] != last_cache) {
199			last_cache = zio_buf_cache[c];
200			kmem_cache_destroy(zio_buf_cache[c]);
201		}
202		zio_buf_cache[c] = NULL;
203
204		if (zio_data_buf_cache[c] != last_data_cache) {
205			last_data_cache = zio_data_buf_cache[c];
206			kmem_cache_destroy(zio_data_buf_cache[c]);
207		}
208		zio_data_buf_cache[c] = NULL;
209	}
210
211	kmem_cache_destroy(zio_link_cache);
212	kmem_cache_destroy(zio_cache);
213
214	zio_inject_fini();
215}
216
217/*
218 * ==========================================================================
219 * Allocate and free I/O buffers
220 * ==========================================================================
221 */
222
223/*
224 * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
225 * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
226 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
227 * excess / transient data in-core during a crashdump.
228 */
229void *
230zio_buf_alloc(size_t size)
231{
232	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
233
234	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
235
236	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
237}
238
239/*
240 * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
241 * crashdump if the kernel panics.  This exists so that we will limit the amount
242 * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
243 * of kernel heap dumped to disk when the kernel panics)
244 */
245void *
246zio_data_buf_alloc(size_t size)
247{
248	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
249
250	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
251
252	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
253}
254
255void
256zio_buf_free(void *buf, size_t size)
257{
258	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
259
260	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
261
262	kmem_cache_free(zio_buf_cache[c], buf);
263}
264
265void
266zio_data_buf_free(void *buf, size_t size)
267{
268	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
269
270	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
271
272	kmem_cache_free(zio_data_buf_cache[c], buf);
273}
274
275/* ARGSUSED */
276static void
277zio_abd_free(void *abd, size_t size)
278{
279	abd_free((abd_t *)abd);
280}
281
282/*
283 * ==========================================================================
284 * Push and pop I/O transform buffers
285 * ==========================================================================
286 */
287void
288zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
289    zio_transform_func_t *transform)
290{
291	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
292
293	/*
294	 * Ensure that anyone expecting this zio to contain a linear ABD isn't
295	 * going to get a nasty surprise when they try to access the data.
296	 */
297	IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
298
299	zt->zt_orig_abd = zio->io_abd;
300	zt->zt_orig_size = zio->io_size;
301	zt->zt_bufsize = bufsize;
302	zt->zt_transform = transform;
303
304	zt->zt_next = zio->io_transform_stack;
305	zio->io_transform_stack = zt;
306
307	zio->io_abd = data;
308	zio->io_size = size;
309}
310
311void
312zio_pop_transforms(zio_t *zio)
313{
314	zio_transform_t *zt;
315
316	while ((zt = zio->io_transform_stack) != NULL) {
317		if (zt->zt_transform != NULL)
318			zt->zt_transform(zio,
319			    zt->zt_orig_abd, zt->zt_orig_size);
320
321		if (zt->zt_bufsize != 0)
322			abd_free(zio->io_abd);
323
324		zio->io_abd = zt->zt_orig_abd;
325		zio->io_size = zt->zt_orig_size;
326		zio->io_transform_stack = zt->zt_next;
327
328		kmem_free(zt, sizeof (zio_transform_t));
329	}
330}
331
332/*
333 * ==========================================================================
334 * I/O transform callbacks for subblocks, decompression, and decryption
335 * ==========================================================================
336 */
337static void
338zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
339{
340	ASSERT(zio->io_size > size);
341
342	if (zio->io_type == ZIO_TYPE_READ)
343		abd_copy(data, zio->io_abd, size);
344}
345
346static void
347zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
348{
349	if (zio->io_error == 0) {
350		void *tmp = abd_borrow_buf(data, size);
351		int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
352		    zio->io_abd, tmp, zio->io_size, size);
353		abd_return_buf_copy(data, tmp, size);
354
355		if (ret != 0)
356			zio->io_error = SET_ERROR(EIO);
357	}
358}
359
360static void
361zio_decrypt(zio_t *zio, abd_t *data, uint64_t size)
362{
363	int ret;
364	void *tmp;
365	blkptr_t *bp = zio->io_bp;
366	spa_t *spa = zio->io_spa;
367	uint64_t dsobj = zio->io_bookmark.zb_objset;
368	uint64_t lsize = BP_GET_LSIZE(bp);
369	dmu_object_type_t ot = BP_GET_TYPE(bp);
370	uint8_t salt[ZIO_DATA_SALT_LEN];
371	uint8_t iv[ZIO_DATA_IV_LEN];
372	uint8_t mac[ZIO_DATA_MAC_LEN];
373	boolean_t no_crypt = B_FALSE;
374
375	ASSERT(BP_USES_CRYPT(bp));
376	ASSERT3U(size, !=, 0);
377
378	if (zio->io_error != 0)
379		return;
380
381	/*
382	 * Verify the cksum of MACs stored in an indirect bp. It will always
383	 * be possible to verify this since it does not require an encryption
384	 * key.
385	 */
386	if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {
387		zio_crypt_decode_mac_bp(bp, mac);
388
389		if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
390			/*
391			 * We haven't decompressed the data yet, but
392			 * zio_crypt_do_indirect_mac_checksum() requires
393			 * decompressed data to be able to parse out the MACs
394			 * from the indirect block. We decompress it now and
395			 * throw away the result after we are finished.
396			 */
397			tmp = zio_buf_alloc(lsize);
398			ret = zio_decompress_data(BP_GET_COMPRESS(bp),
399			    zio->io_abd, tmp, zio->io_size, lsize);
400			if (ret != 0) {
401				ret = SET_ERROR(EIO);
402				goto error;
403			}
404			ret = zio_crypt_do_indirect_mac_checksum(B_FALSE,
405			    tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac);
406			zio_buf_free(tmp, lsize);
407		} else {
408			ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
409			    zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac);
410		}
411		abd_copy(data, zio->io_abd, size);
412
413		if (ret != 0)
414			goto error;
415
416		return;
417	}
418
419	/*
420	 * If this is an authenticated block, just check the MAC. It would be
421	 * nice to separate this out into its own flag, but for the moment
422	 * enum zio_flag is out of bits.
423	 */
424	if (BP_IS_AUTHENTICATED(bp)) {
425		if (ot == DMU_OT_OBJSET) {
426			ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa,
427			    dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp));
428		} else {
429			zio_crypt_decode_mac_bp(bp, mac);
430			ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj,
431			    zio->io_abd, size, mac);
432		}
433		abd_copy(data, zio->io_abd, size);
434
435		if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) {
436			ret = zio_handle_decrypt_injection(spa,
437			    &zio->io_bookmark, ot, ECKSUM);
438		}
439		if (ret != 0)
440			goto error;
441
442		return;
443	}
444
445	zio_crypt_decode_params_bp(bp, salt, iv);
446
447	if (ot == DMU_OT_INTENT_LOG) {
448		tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t));
449		zio_crypt_decode_mac_zil(tmp, mac);
450		abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t));
451	} else {
452		zio_crypt_decode_mac_bp(bp, mac);
453	}
454
455	ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp),
456	    BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data,
457	    zio->io_abd, &no_crypt);
458	if (no_crypt)
459		abd_copy(data, zio->io_abd, size);
460
461	if (ret != 0)
462		goto error;
463
464	return;
465
466error:
467	/* assert that the key was found unless this was speculative */
468	ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE));
469
470	/*
471	 * If there was a decryption / authentication error return EIO as
472	 * the io_error. If this was not a speculative zio, create an ereport.
473	 */
474	if (ret == ECKSUM) {
475		zio->io_error = SET_ERROR(EIO);
476		if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
477			spa_log_error(spa, &zio->io_bookmark);
478			zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
479			    spa, NULL, &zio->io_bookmark, zio, 0, 0);
480		}
481	} else {
482		zio->io_error = ret;
483	}
484}
485
486/*
487 * ==========================================================================
488 * I/O parent/child relationships and pipeline interlocks
489 * ==========================================================================
490 */
491zio_t *
492zio_walk_parents(zio_t *cio, zio_link_t **zl)
493{
494	list_t *pl = &cio->io_parent_list;
495
496	*zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
497	if (*zl == NULL)
498		return (NULL);
499
500	ASSERT((*zl)->zl_child == cio);
501	return ((*zl)->zl_parent);
502}
503
504zio_t *
505zio_walk_children(zio_t *pio, zio_link_t **zl)
506{
507	list_t *cl = &pio->io_child_list;
508
509	ASSERT(MUTEX_HELD(&pio->io_lock));
510
511	*zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
512	if (*zl == NULL)
513		return (NULL);
514
515	ASSERT((*zl)->zl_parent == pio);
516	return ((*zl)->zl_child);
517}
518
519zio_t *
520zio_unique_parent(zio_t *cio)
521{
522	zio_link_t *zl = NULL;
523	zio_t *pio = zio_walk_parents(cio, &zl);
524
525	VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
526	return (pio);
527}
528
529void
530zio_add_child(zio_t *pio, zio_t *cio)
531{
532	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
533
534	/*
535	 * Logical I/Os can have logical, gang, or vdev children.
536	 * Gang I/Os can have gang or vdev children.
537	 * Vdev I/Os can only have vdev children.
538	 * The following ASSERT captures all of these constraints.
539	 */
540	ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
541
542	zl->zl_parent = pio;
543	zl->zl_child = cio;
544
545	mutex_enter(&pio->io_lock);
546	mutex_enter(&cio->io_lock);
547
548	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
549
550	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
551		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
552
553	list_insert_head(&pio->io_child_list, zl);
554	list_insert_head(&cio->io_parent_list, zl);
555
556	pio->io_child_count++;
557	cio->io_parent_count++;
558
559	mutex_exit(&cio->io_lock);
560	mutex_exit(&pio->io_lock);
561}
562
563static void
564zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
565{
566	ASSERT(zl->zl_parent == pio);
567	ASSERT(zl->zl_child == cio);
568
569	mutex_enter(&pio->io_lock);
570	mutex_enter(&cio->io_lock);
571
572	list_remove(&pio->io_child_list, zl);
573	list_remove(&cio->io_parent_list, zl);
574
575	pio->io_child_count--;
576	cio->io_parent_count--;
577
578	mutex_exit(&cio->io_lock);
579	mutex_exit(&pio->io_lock);
580
581	kmem_cache_free(zio_link_cache, zl);
582}
583
584static boolean_t
585zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
586{
587	boolean_t waiting = B_FALSE;
588
589	mutex_enter(&zio->io_lock);
590	ASSERT(zio->io_stall == NULL);
591	for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
592		if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
593			continue;
594
595		uint64_t *countp = &zio->io_children[c][wait];
596		if (*countp != 0) {
597			zio->io_stage >>= 1;
598			ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
599			zio->io_stall = countp;
600			waiting = B_TRUE;
601			break;
602		}
603	}
604	mutex_exit(&zio->io_lock);
605	return (waiting);
606}
607
608static void
609zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
610{
611	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
612	int *errorp = &pio->io_child_error[zio->io_child_type];
613
614	mutex_enter(&pio->io_lock);
615	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
616		*errorp = zio_worst_error(*errorp, zio->io_error);
617	pio->io_reexecute |= zio->io_reexecute;
618	ASSERT3U(*countp, >, 0);
619
620	(*countp)--;
621
622	if (*countp == 0 && pio->io_stall == countp) {
623		zio_taskq_type_t type =
624		    pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
625		    ZIO_TASKQ_INTERRUPT;
626		pio->io_stall = NULL;
627		mutex_exit(&pio->io_lock);
628		/*
629		 * Dispatch the parent zio in its own taskq so that
630		 * the child can continue to make progress. This also
631		 * prevents overflowing the stack when we have deeply nested
632		 * parent-child relationships.
633		 */
634		zio_taskq_dispatch(pio, type, B_FALSE);
635	} else {
636		mutex_exit(&pio->io_lock);
637	}
638}
639
640static void
641zio_inherit_child_errors(zio_t *zio, enum zio_child c)
642{
643	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
644		zio->io_error = zio->io_child_error[c];
645}
646
647int
648zio_bookmark_compare(const void *x1, const void *x2)
649{
650	const zio_t *z1 = x1;
651	const zio_t *z2 = x2;
652
653	if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
654		return (-1);
655	if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
656		return (1);
657
658	if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
659		return (-1);
660	if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
661		return (1);
662
663	if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
664		return (-1);
665	if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
666		return (1);
667
668	if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
669		return (-1);
670	if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
671		return (1);
672
673	if (z1 < z2)
674		return (-1);
675	if (z1 > z2)
676		return (1);
677
678	return (0);
679}
680
681/*
682 * ==========================================================================
683 * Create the various types of I/O (read, write, free, etc)
684 * ==========================================================================
685 */
686static zio_t *
687zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
688    abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
689    void *private, zio_type_t type, zio_priority_t priority,
690    enum zio_flag flags, vdev_t *vd, uint64_t offset,
691    const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
692{
693	zio_t *zio;
694
695	IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE);
696	ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
697	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
698
699	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
700	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
701	ASSERT(vd || stage == ZIO_STAGE_OPEN);
702
703	IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0);
704
705	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
706	bzero(zio, sizeof (zio_t));
707
708	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
709	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
710
711	list_create(&zio->io_parent_list, sizeof (zio_link_t),
712	    offsetof(zio_link_t, zl_parent_node));
713	list_create(&zio->io_child_list, sizeof (zio_link_t),
714	    offsetof(zio_link_t, zl_child_node));
715	metaslab_trace_init(&zio->io_alloc_list);
716
717	if (vd != NULL)
718		zio->io_child_type = ZIO_CHILD_VDEV;
719	else if (flags & ZIO_FLAG_GANG_CHILD)
720		zio->io_child_type = ZIO_CHILD_GANG;
721	else if (flags & ZIO_FLAG_DDT_CHILD)
722		zio->io_child_type = ZIO_CHILD_DDT;
723	else
724		zio->io_child_type = ZIO_CHILD_LOGICAL;
725
726	if (bp != NULL) {
727		zio->io_bp = (blkptr_t *)bp;
728		zio->io_bp_copy = *bp;
729		zio->io_bp_orig = *bp;
730		if (type != ZIO_TYPE_WRITE ||
731		    zio->io_child_type == ZIO_CHILD_DDT)
732			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
733		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
734			zio->io_logical = zio;
735		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
736			pipeline |= ZIO_GANG_STAGES;
737	}
738
739	zio->io_spa = spa;
740	zio->io_txg = txg;
741	zio->io_done = done;
742	zio->io_private = private;
743	zio->io_type = type;
744	zio->io_priority = priority;
745	zio->io_vd = vd;
746	zio->io_offset = offset;
747	zio->io_orig_abd = zio->io_abd = data;
748	zio->io_orig_size = zio->io_size = psize;
749	zio->io_lsize = lsize;
750	zio->io_orig_flags = zio->io_flags = flags;
751	zio->io_orig_stage = zio->io_stage = stage;
752	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
753	zio->io_pipeline_trace = ZIO_STAGE_OPEN;
754
755	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
756	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
757
758	if (zb != NULL)
759		zio->io_bookmark = *zb;
760
761	if (pio != NULL) {
762		if (zio->io_metaslab_class == NULL)
763			zio->io_metaslab_class = pio->io_metaslab_class;
764		if (zio->io_logical == NULL)
765			zio->io_logical = pio->io_logical;
766		if (zio->io_child_type == ZIO_CHILD_GANG)
767			zio->io_gang_leader = pio->io_gang_leader;
768		zio_add_child(pio, zio);
769	}
770
771	return (zio);
772}
773
774static void
775zio_destroy(zio_t *zio)
776{
777	metaslab_trace_fini(&zio->io_alloc_list);
778	list_destroy(&zio->io_parent_list);
779	list_destroy(&zio->io_child_list);
780	mutex_destroy(&zio->io_lock);
781	cv_destroy(&zio->io_cv);
782	kmem_cache_free(zio_cache, zio);
783}
784
785zio_t *
786zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
787    void *private, enum zio_flag flags)
788{
789	zio_t *zio;
790
791	zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
792	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
793	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
794
795	return (zio);
796}
797
798zio_t *
799zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
800{
801	return (zio_null(NULL, spa, NULL, done, private, flags));
802}
803
804void
805zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
806{
807	if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
808		zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
809		    bp, (longlong_t)BP_GET_TYPE(bp));
810	}
811	if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
812	    BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
813		zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
814		    bp, (longlong_t)BP_GET_CHECKSUM(bp));
815	}
816	if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
817	    BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
818		zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
819		    bp, (longlong_t)BP_GET_COMPRESS(bp));
820	}
821	if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
822		zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
823		    bp, (longlong_t)BP_GET_LSIZE(bp));
824	}
825	if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
826		zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
827		    bp, (longlong_t)BP_GET_PSIZE(bp));
828	}
829
830	if (BP_IS_EMBEDDED(bp)) {
831		if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
832			zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
833			    bp, (longlong_t)BPE_GET_ETYPE(bp));
834		}
835	}
836
837	/*
838	 * Do not verify individual DVAs if the config is not trusted. This
839	 * will be done once the zio is executed in vdev_mirror_map_alloc.
840	 */
841	if (!spa->spa_trust_config)
842		return;
843
844	/*
845	 * Pool-specific checks.
846	 *
847	 * Note: it would be nice to verify that the blk_birth and
848	 * BP_PHYSICAL_BIRTH() are not too large.  However, spa_freeze()
849	 * allows the birth time of log blocks (and dmu_sync()-ed blocks
850	 * that are in the log) to be arbitrarily large.
851	 */
852	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
853		uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
854		if (vdevid >= spa->spa_root_vdev->vdev_children) {
855			zfs_panic_recover("blkptr at %p DVA %u has invalid "
856			    "VDEV %llu",
857			    bp, i, (longlong_t)vdevid);
858			continue;
859		}
860		vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
861		if (vd == NULL) {
862			zfs_panic_recover("blkptr at %p DVA %u has invalid "
863			    "VDEV %llu",
864			    bp, i, (longlong_t)vdevid);
865			continue;
866		}
867		if (vd->vdev_ops == &vdev_hole_ops) {
868			zfs_panic_recover("blkptr at %p DVA %u has hole "
869			    "VDEV %llu",
870			    bp, i, (longlong_t)vdevid);
871			continue;
872		}
873		if (vd->vdev_ops == &vdev_missing_ops) {
874			/*
875			 * "missing" vdevs are valid during import, but we
876			 * don't have their detailed info (e.g. asize), so
877			 * we can't perform any more checks on them.
878			 */
879			continue;
880		}
881		uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
882		uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
883		if (BP_IS_GANG(bp))
884			asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
885		if (offset + asize > vd->vdev_asize) {
886			zfs_panic_recover("blkptr at %p DVA %u has invalid "
887			    "OFFSET %llu",
888			    bp, i, (longlong_t)offset);
889		}
890	}
891}
892
893boolean_t
894zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
895{
896	uint64_t vdevid = DVA_GET_VDEV(dva);
897
898	if (vdevid >= spa->spa_root_vdev->vdev_children)
899		return (B_FALSE);
900
901	vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
902	if (vd == NULL)
903		return (B_FALSE);
904
905	if (vd->vdev_ops == &vdev_hole_ops)
906		return (B_FALSE);
907
908	if (vd->vdev_ops == &vdev_missing_ops) {
909		return (B_FALSE);
910	}
911
912	uint64_t offset = DVA_GET_OFFSET(dva);
913	uint64_t asize = DVA_GET_ASIZE(dva);
914
915	if (BP_IS_GANG(bp))
916		asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
917	if (offset + asize > vd->vdev_asize)
918		return (B_FALSE);
919
920	return (B_TRUE);
921}
922
923zio_t *
924zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
925    abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
926    zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
927{
928	zio_t *zio;
929
930	zfs_blkptr_verify(spa, bp);
931
932	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
933	    data, size, size, done, private,
934	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
935	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
936	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
937
938	return (zio);
939}
940
941zio_t *
942zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
943    abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
944    zio_done_func_t *ready, zio_done_func_t *children_ready,
945    zio_done_func_t *physdone, zio_done_func_t *done,
946    void *private, zio_priority_t priority, enum zio_flag flags,
947    const zbookmark_phys_t *zb)
948{
949	zio_t *zio;
950
951	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
952	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
953	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
954	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
955	    DMU_OT_IS_VALID(zp->zp_type) &&
956	    zp->zp_level < 32 &&
957	    zp->zp_copies > 0 &&
958	    zp->zp_copies <= spa_max_replication(spa));
959
960	zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
961	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
962	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
963	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
964
965	zio->io_ready = ready;
966	zio->io_children_ready = children_ready;
967	zio->io_physdone = physdone;
968	zio->io_prop = *zp;
969
970	/*
971	 * Data can be NULL if we are going to call zio_write_override() to
972	 * provide the already-allocated BP.  But we may need the data to
973	 * verify a dedup hit (if requested).  In this case, don't try to
974	 * dedup (just take the already-allocated BP verbatim). Encrypted
975	 * dedup blocks need data as well so we also disable dedup in this
976	 * case.
977	 */
978	if (data == NULL &&
979	    (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) {
980		zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
981	}
982
983	return (zio);
984}
985
986zio_t *
987zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
988    uint64_t size, zio_done_func_t *done, void *private,
989    zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
990{
991	zio_t *zio;
992
993	zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
994	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
995	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
996
997	return (zio);
998}
999
1000void
1001zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
1002{
1003	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1004	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1005	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1006	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
1007
1008	/*
1009	 * We must reset the io_prop to match the values that existed
1010	 * when the bp was first written by dmu_sync() keeping in mind
1011	 * that nopwrite and dedup are mutually exclusive.
1012	 */
1013	zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
1014	zio->io_prop.zp_nopwrite = nopwrite;
1015	zio->io_prop.zp_copies = copies;
1016	zio->io_bp_override = bp;
1017}
1018
1019void
1020zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
1021{
1022
1023	zfs_blkptr_verify(spa, bp);
1024
1025	/*
1026	 * The check for EMBEDDED is a performance optimization.  We
1027	 * process the free here (by ignoring it) rather than
1028	 * putting it on the list and then processing it in zio_free_sync().
1029	 */
1030	if (BP_IS_EMBEDDED(bp))
1031		return;
1032	metaslab_check_free(spa, bp);
1033
1034	/*
1035	 * Frees that are for the currently-syncing txg, are not going to be
1036	 * deferred, and which will not need to do a read (i.e. not GANG or
1037	 * DEDUP), can be processed immediately.  Otherwise, put them on the
1038	 * in-memory list for later processing.
1039	 *
1040	 * Note that we only defer frees after zfs_sync_pass_deferred_free
1041	 * when the log space map feature is disabled. [see relevant comment
1042	 * in spa_sync_iterate_to_convergence()]
1043	 */
1044	if (BP_IS_GANG(bp) ||
1045	    BP_GET_DEDUP(bp) ||
1046	    txg != spa->spa_syncing_txg ||
1047	    (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free &&
1048	    !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))) {
1049		bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1050	} else {
1051		VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
1052	}
1053}
1054
1055zio_t *
1056zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1057    enum zio_flag flags)
1058{
1059	zio_t *zio;
1060	enum zio_stage stage = ZIO_FREE_PIPELINE;
1061
1062	ASSERT(!BP_IS_HOLE(bp));
1063	ASSERT(spa_syncing_txg(spa) == txg);
1064
1065	if (BP_IS_EMBEDDED(bp))
1066		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1067
1068	metaslab_check_free(spa, bp);
1069	arc_freed(spa, bp);
1070	dsl_scan_freed(spa, bp);
1071
1072	/*
1073	 * GANG and DEDUP blocks can induce a read (for the gang block header,
1074	 * or the DDT), so issue them asynchronously so that this thread is
1075	 * not tied up.
1076	 */
1077	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
1078		stage |= ZIO_STAGE_ISSUE_ASYNC;
1079
1080	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1081	    BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1082	    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
1083
1084	return (zio);
1085}
1086
1087zio_t *
1088zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1089    zio_done_func_t *done, void *private, enum zio_flag flags)
1090{
1091	zio_t *zio;
1092
1093	zfs_blkptr_verify(spa, bp);
1094
1095	if (BP_IS_EMBEDDED(bp))
1096		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1097
1098	/*
1099	 * A claim is an allocation of a specific block.  Claims are needed
1100	 * to support immediate writes in the intent log.  The issue is that
1101	 * immediate writes contain committed data, but in a txg that was
1102	 * *not* committed.  Upon opening the pool after an unclean shutdown,
1103	 * the intent log claims all blocks that contain immediate write data
1104	 * so that the SPA knows they're in use.
1105	 *
1106	 * All claims *must* be resolved in the first txg -- before the SPA
1107	 * starts allocating blocks -- so that nothing is allocated twice.
1108	 * If txg == 0 we just verify that the block is claimable.
1109	 */
1110	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1111	    spa_min_claim_txg(spa));
1112	ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1113	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
1114
1115	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1116	    BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1117	    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1118	ASSERT0(zio->io_queued_timestamp);
1119
1120	return (zio);
1121}
1122
1123zio_t *
1124zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
1125    zio_done_func_t *done, void *private, enum zio_flag flags)
1126{
1127	zio_t *zio;
1128	int c;
1129
1130	if (vd->vdev_children == 0) {
1131		zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1132		    ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1133		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1134
1135		zio->io_cmd = cmd;
1136	} else {
1137		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1138
1139		for (c = 0; c < vd->vdev_children; c++)
1140			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1141			    done, private, flags));
1142	}
1143
1144	return (zio);
1145}
1146
1147zio_t *
1148zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1149    zio_done_func_t *done, void *private, zio_priority_t priority,
1150    enum zio_flag flags, enum trim_flag trim_flags)
1151{
1152	zio_t *zio;
1153
1154	ASSERT0(vd->vdev_children);
1155	ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
1156	ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1157	ASSERT3U(size, !=, 0);
1158
1159	zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done,
1160	    private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL,
1161	    vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE);
1162	zio->io_trim_flags = trim_flags;
1163
1164	return (zio);
1165}
1166
1167zio_t *
1168zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1169    abd_t *data, int checksum, zio_done_func_t *done, void *private,
1170    zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1171{
1172	zio_t *zio;
1173
1174	ASSERT(vd->vdev_children == 0);
1175	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1176	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1177	ASSERT3U(offset + size, <=, vd->vdev_psize);
1178
1179	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1180	    private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1181	    offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1182
1183	zio->io_prop.zp_checksum = checksum;
1184
1185	return (zio);
1186}
1187
1188zio_t *
1189zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1190    abd_t *data, int checksum, zio_done_func_t *done, void *private,
1191    zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1192{
1193	zio_t *zio;
1194
1195	ASSERT(vd->vdev_children == 0);
1196	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1197	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1198	ASSERT3U(offset + size, <=, vd->vdev_psize);
1199
1200	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1201	    private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1202	    offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1203
1204	zio->io_prop.zp_checksum = checksum;
1205
1206	if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1207		/*
1208		 * zec checksums are necessarily destructive -- they modify
1209		 * the end of the write buffer to hold the verifier/checksum.
1210		 * Therefore, we must make a local copy in case the data is
1211		 * being written to multiple places in parallel.
1212		 */
1213		abd_t *wbuf = abd_alloc_sametype(data, size);
1214		abd_copy(wbuf, data, size);
1215
1216		zio_push_transform(zio, wbuf, size, size, NULL);
1217	}
1218
1219	return (zio);
1220}
1221
1222/*
1223 * Create a child I/O to do some work for us.
1224 */
1225zio_t *
1226zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1227    abd_t *data, uint64_t size, int type, zio_priority_t priority,
1228    enum zio_flag flags, zio_done_func_t *done, void *private)
1229{
1230	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1231	zio_t *zio;
1232
1233	/*
1234	 * vdev child I/Os do not propagate their error to the parent.
1235	 * Therefore, for correct operation the caller *must* check for
1236	 * and handle the error in the child i/o's done callback.
1237	 * The only exceptions are i/os that we don't care about
1238	 * (OPTIONAL or REPAIR).
1239	 */
1240	ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1241	    done != NULL);
1242
1243	if (type == ZIO_TYPE_READ && bp != NULL) {
1244		/*
1245		 * If we have the bp, then the child should perform the
1246		 * checksum and the parent need not.  This pushes error
1247		 * detection as close to the leaves as possible and
1248		 * eliminates redundant checksums in the interior nodes.
1249		 */
1250		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1251		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1252	}
1253
1254	if (vd->vdev_ops->vdev_op_leaf) {
1255		ASSERT0(vd->vdev_children);
1256		offset += VDEV_LABEL_START_SIZE;
1257	}
1258
1259	flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1260
1261	/*
1262	 * If we've decided to do a repair, the write is not speculative --
1263	 * even if the original read was.
1264	 */
1265	if (flags & ZIO_FLAG_IO_REPAIR)
1266		flags &= ~ZIO_FLAG_SPECULATIVE;
1267
1268	/*
1269	 * If we're creating a child I/O that is not associated with a
1270	 * top-level vdev, then the child zio is not an allocating I/O.
1271	 * If this is a retried I/O then we ignore it since we will
1272	 * have already processed the original allocating I/O.
1273	 */
1274	if (flags & ZIO_FLAG_IO_ALLOCATING &&
1275	    (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1276		ASSERT(pio->io_metaslab_class != NULL);
1277		ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled);
1278		ASSERT(type == ZIO_TYPE_WRITE);
1279		ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1280		ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1281		ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1282		    pio->io_child_type == ZIO_CHILD_GANG);
1283
1284		flags &= ~ZIO_FLAG_IO_ALLOCATING;
1285	}
1286
1287	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1288	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1289	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1290	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1291
1292	zio->io_physdone = pio->io_physdone;
1293	if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1294		zio->io_logical->io_phys_children++;
1295
1296	return (zio);
1297}
1298
1299zio_t *
1300zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1301    zio_type_t type, zio_priority_t priority, enum zio_flag flags,
1302    zio_done_func_t *done, void *private)
1303{
1304	zio_t *zio;
1305
1306	ASSERT(vd->vdev_ops->vdev_op_leaf);
1307
1308	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1309	    data, size, size, done, private, type, priority,
1310	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1311	    vd, offset, NULL,
1312	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1313
1314	return (zio);
1315}
1316
1317void
1318zio_flush(zio_t *zio, vdev_t *vd)
1319{
1320	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1321	    NULL, NULL,
1322	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1323}
1324
1325void
1326zio_shrink(zio_t *zio, uint64_t size)
1327{
1328	ASSERT3P(zio->io_executor, ==, NULL);
1329	ASSERT3P(zio->io_orig_size, ==, zio->io_size);
1330	ASSERT3U(size, <=, zio->io_size);
1331
1332	/*
1333	 * We don't shrink for raidz because of problems with the
1334	 * reconstruction when reading back less than the block size.
1335	 * Note, BP_IS_RAIDZ() assumes no compression.
1336	 */
1337	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1338	if (!BP_IS_RAIDZ(zio->io_bp)) {
1339		/* we are not doing a raw write */
1340		ASSERT3U(zio->io_size, ==, zio->io_lsize);
1341		zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1342	}
1343}
1344
1345/*
1346 * ==========================================================================
1347 * Prepare to read and write logical blocks
1348 * ==========================================================================
1349 */
1350
1351static int
1352zio_read_bp_init(zio_t *zio)
1353{
1354	blkptr_t *bp = zio->io_bp;
1355	uint64_t psize =
1356	    BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1357
1358	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1359
1360	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1361	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
1362	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1363		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1364		    psize, psize, zio_decompress);
1365	}
1366
1367	if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) ||
1368	    BP_HAS_INDIRECT_MAC_CKSUM(bp)) &&
1369	    zio->io_child_type == ZIO_CHILD_LOGICAL) {
1370		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1371		    psize, psize, zio_decrypt);
1372	}
1373
1374	if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1375		int psize = BPE_GET_PSIZE(bp);
1376		void *data = abd_borrow_buf(zio->io_abd, psize);
1377
1378		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1379		decode_embedded_bp_compressed(bp, data);
1380		abd_return_buf_copy(zio->io_abd, data, psize);
1381	} else {
1382		ASSERT(!BP_IS_EMBEDDED(bp));
1383		ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1384	}
1385
1386	if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1387		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1388
1389	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1390		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1391
1392	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1393		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1394
1395	return (ZIO_PIPELINE_CONTINUE);
1396}
1397
1398static int
1399zio_write_bp_init(zio_t *zio)
1400{
1401	if (!IO_IS_ALLOCATING(zio))
1402		return (ZIO_PIPELINE_CONTINUE);
1403
1404	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1405
1406	if (zio->io_bp_override) {
1407		blkptr_t *bp = zio->io_bp;
1408		zio_prop_t *zp = &zio->io_prop;
1409
1410		ASSERT(bp->blk_birth != zio->io_txg);
1411		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1412
1413		*bp = *zio->io_bp_override;
1414		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1415
1416		if (BP_IS_EMBEDDED(bp))
1417			return (ZIO_PIPELINE_CONTINUE);
1418
1419		/*
1420		 * If we've been overridden and nopwrite is set then
1421		 * set the flag accordingly to indicate that a nopwrite
1422		 * has already occurred.
1423		 */
1424		if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1425			ASSERT(!zp->zp_dedup);
1426			ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1427			zio->io_flags |= ZIO_FLAG_NOPWRITE;
1428			return (ZIO_PIPELINE_CONTINUE);
1429		}
1430
1431		ASSERT(!zp->zp_nopwrite);
1432
1433		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1434			return (ZIO_PIPELINE_CONTINUE);
1435
1436		ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1437		    ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1438
1439		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum &&
1440		    !zp->zp_encrypt) {
1441			BP_SET_DEDUP(bp, 1);
1442			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1443			return (ZIO_PIPELINE_CONTINUE);
1444		}
1445
1446		/*
1447		 * We were unable to handle this as an override bp, treat
1448		 * it as a regular write I/O.
1449		 */
1450		zio->io_bp_override = NULL;
1451		*bp = zio->io_bp_orig;
1452		zio->io_pipeline = zio->io_orig_pipeline;
1453	}
1454
1455	return (ZIO_PIPELINE_CONTINUE);
1456}
1457
1458static int
1459zio_write_compress(zio_t *zio)
1460{
1461	spa_t *spa = zio->io_spa;
1462	zio_prop_t *zp = &zio->io_prop;
1463	enum zio_compress compress = zp->zp_compress;
1464	blkptr_t *bp = zio->io_bp;
1465	uint64_t lsize = zio->io_lsize;
1466	uint64_t psize = zio->io_size;
1467	int pass = 1;
1468
1469	/*
1470	 * If our children haven't all reached the ready stage,
1471	 * wait for them and then repeat this pipeline stage.
1472	 */
1473	if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1474	    ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1475		return (ZIO_PIPELINE_STOP);
1476	}
1477
1478	if (!IO_IS_ALLOCATING(zio))
1479		return (ZIO_PIPELINE_CONTINUE);
1480
1481	if (zio->io_children_ready != NULL) {
1482		/*
1483		 * Now that all our children are ready, run the callback
1484		 * associated with this zio in case it wants to modify the
1485		 * data to be written.
1486		 */
1487		ASSERT3U(zp->zp_level, >, 0);
1488		zio->io_children_ready(zio);
1489	}
1490
1491	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1492	ASSERT(zio->io_bp_override == NULL);
1493
1494	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1495		/*
1496		 * We're rewriting an existing block, which means we're
1497		 * working on behalf of spa_sync().  For spa_sync() to
1498		 * converge, it must eventually be the case that we don't
1499		 * have to allocate new blocks.  But compression changes
1500		 * the blocksize, which forces a reallocate, and makes
1501		 * convergence take longer.  Therefore, after the first
1502		 * few passes, stop compressing to ensure convergence.
1503		 */
1504		pass = spa_sync_pass(spa);
1505
1506		ASSERT(zio->io_txg == spa_syncing_txg(spa));
1507		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1508		ASSERT(!BP_GET_DEDUP(bp));
1509
1510		if (pass >= zfs_sync_pass_dont_compress)
1511			compress = ZIO_COMPRESS_OFF;
1512
1513		/* Make sure someone doesn't change their mind on overwrites */
1514		ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1515		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1516	}
1517
1518	/* If it's a compressed write that is not raw, compress the buffer. */
1519	if (compress != ZIO_COMPRESS_OFF &&
1520	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1521		void *cbuf = zio_buf_alloc(lsize);
1522		psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1523		if (psize == 0 || psize == lsize) {
1524			compress = ZIO_COMPRESS_OFF;
1525			zio_buf_free(cbuf, lsize);
1526		} else if (!zp->zp_dedup && !zp->zp_encrypt &&
1527		    psize <= BPE_PAYLOAD_SIZE &&
1528		    zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1529		    spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1530			encode_embedded_bp_compressed(bp,
1531			    cbuf, compress, lsize, psize);
1532			BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1533			BP_SET_TYPE(bp, zio->io_prop.zp_type);
1534			BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1535			zio_buf_free(cbuf, lsize);
1536			bp->blk_birth = zio->io_txg;
1537			zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1538			ASSERT(spa_feature_is_active(spa,
1539			    SPA_FEATURE_EMBEDDED_DATA));
1540			return (ZIO_PIPELINE_CONTINUE);
1541		} else {
1542			/*
1543			 * Round up compressed size up to the ashift
1544			 * of the smallest-ashift device, and zero the tail.
1545			 * This ensures that the compressed size of the BP
1546			 * (and thus compressratio property) are correct,
1547			 * in that we charge for the padding used to fill out
1548			 * the last sector.
1549			 */
1550			ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1551			size_t rounded = (size_t)P2ROUNDUP(psize,
1552			    1ULL << spa->spa_min_ashift);
1553			if (rounded >= lsize) {
1554				compress = ZIO_COMPRESS_OFF;
1555				zio_buf_free(cbuf, lsize);
1556				psize = lsize;
1557			} else {
1558				abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1559				abd_take_ownership_of_buf(cdata, B_TRUE);
1560				abd_zero_off(cdata, psize, rounded - psize);
1561				psize = rounded;
1562				zio_push_transform(zio, cdata,
1563				    psize, lsize, NULL);
1564			}
1565		}
1566
1567		/*
1568		 * We were unable to handle this as an override bp, treat
1569		 * it as a regular write I/O.
1570		 */
1571		zio->io_bp_override = NULL;
1572		*bp = zio->io_bp_orig;
1573		zio->io_pipeline = zio->io_orig_pipeline;
1574
1575	} else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 &&
1576	    zp->zp_type == DMU_OT_DNODE) {
1577		/*
1578		 * The DMU actually relies on the zio layer's compression
1579		 * to free metadnode blocks that have had all contained
1580		 * dnodes freed. As a result, even when doing a raw
1581		 * receive, we must check whether the block can be compressed
1582		 * to a hole.
1583		 */
1584		psize = zio_compress_data(ZIO_COMPRESS_EMPTY,
1585		    zio->io_abd, NULL, lsize);
1586		if (psize == 0)
1587			compress = ZIO_COMPRESS_OFF;
1588	} else {
1589		ASSERT3U(psize, !=, 0);
1590	}
1591
1592	/*
1593	 * The final pass of spa_sync() must be all rewrites, but the first
1594	 * few passes offer a trade-off: allocating blocks defers convergence,
1595	 * but newly allocated blocks are sequential, so they can be written
1596	 * to disk faster.  Therefore, we allow the first few passes of
1597	 * spa_sync() to allocate new blocks, but force rewrites after that.
1598	 * There should only be a handful of blocks after pass 1 in any case.
1599	 */
1600	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1601	    BP_GET_PSIZE(bp) == psize &&
1602	    pass >= zfs_sync_pass_rewrite) {
1603		VERIFY3U(psize, !=, 0);
1604		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1605		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1606		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1607	} else {
1608		BP_ZERO(bp);
1609		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1610	}
1611
1612	if (psize == 0) {
1613		if (zio->io_bp_orig.blk_birth != 0 &&
1614		    spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1615			BP_SET_LSIZE(bp, lsize);
1616			BP_SET_TYPE(bp, zp->zp_type);
1617			BP_SET_LEVEL(bp, zp->zp_level);
1618			BP_SET_BIRTH(bp, zio->io_txg, 0);
1619		}
1620		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1621	} else {
1622		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1623		BP_SET_LSIZE(bp, lsize);
1624		BP_SET_TYPE(bp, zp->zp_type);
1625		BP_SET_LEVEL(bp, zp->zp_level);
1626		BP_SET_PSIZE(bp, psize);
1627		BP_SET_COMPRESS(bp, compress);
1628		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1629		BP_SET_DEDUP(bp, zp->zp_dedup);
1630		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1631		if (zp->zp_dedup) {
1632			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1633			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1634			ASSERT(!zp->zp_encrypt ||
1635			    DMU_OT_IS_ENCRYPTED(zp->zp_type));
1636			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1637		}
1638		if (zp->zp_nopwrite) {
1639			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1640			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1641			zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1642		}
1643	}
1644	return (ZIO_PIPELINE_CONTINUE);
1645}
1646
1647static int
1648zio_free_bp_init(zio_t *zio)
1649{
1650	blkptr_t *bp = zio->io_bp;
1651
1652	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1653		if (BP_GET_DEDUP(bp))
1654			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1655	}
1656
1657	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1658
1659	return (ZIO_PIPELINE_CONTINUE);
1660}
1661
1662/*
1663 * ==========================================================================
1664 * Execute the I/O pipeline
1665 * ==========================================================================
1666 */
1667
1668static void
1669zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1670{
1671	spa_t *spa = zio->io_spa;
1672	zio_type_t t = zio->io_type;
1673	int flags = (cutinline ? TQ_FRONT : 0);
1674
1675	/*
1676	 * If we're a config writer or a probe, the normal issue and
1677	 * interrupt threads may all be blocked waiting for the config lock.
1678	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1679	 */
1680	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1681		t = ZIO_TYPE_NULL;
1682
1683	/*
1684	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1685	 */
1686	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1687		t = ZIO_TYPE_NULL;
1688
1689	/*
1690	 * If this is a high priority I/O, then use the high priority taskq if
1691	 * available.
1692	 */
1693	if ((zio->io_priority == ZIO_PRIORITY_NOW ||
1694	    zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) &&
1695	    spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1696		q++;
1697
1698	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1699
1700	/*
1701	 * NB: We are assuming that the zio can only be dispatched
1702	 * to a single taskq at a time.  It would be a grievous error
1703	 * to dispatch the zio to another taskq at the same time.
1704	 */
1705	ASSERT(zio->io_tqent.tqent_next == NULL);
1706	spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1707	    flags, &zio->io_tqent);
1708}
1709
1710static boolean_t
1711zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1712{
1713	kthread_t *executor = zio->io_executor;
1714	spa_t *spa = zio->io_spa;
1715
1716	for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1717		spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1718		uint_t i;
1719		for (i = 0; i < tqs->stqs_count; i++) {
1720			if (taskq_member(tqs->stqs_taskq[i], executor))
1721				return (B_TRUE);
1722		}
1723	}
1724
1725	return (B_FALSE);
1726}
1727
1728static int
1729zio_issue_async(zio_t *zio)
1730{
1731	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1732
1733	return (ZIO_PIPELINE_STOP);
1734}
1735
1736void
1737zio_interrupt(zio_t *zio)
1738{
1739	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1740}
1741
1742void
1743zio_delay_interrupt(zio_t *zio)
1744{
1745	/*
1746	 * The timeout_generic() function isn't defined in userspace, so
1747	 * rather than trying to implement the function, the zio delay
1748	 * functionality has been disabled for userspace builds.
1749	 */
1750
1751#ifdef _KERNEL
1752	/*
1753	 * If io_target_timestamp is zero, then no delay has been registered
1754	 * for this IO, thus jump to the end of this function and "skip" the
1755	 * delay; issuing it directly to the zio layer.
1756	 */
1757	if (zio->io_target_timestamp != 0) {
1758		hrtime_t now = gethrtime();
1759
1760		if (now >= zio->io_target_timestamp) {
1761			/*
1762			 * This IO has already taken longer than the target
1763			 * delay to complete, so we don't want to delay it
1764			 * any longer; we "miss" the delay and issue it
1765			 * directly to the zio layer. This is likely due to
1766			 * the target latency being set to a value less than
1767			 * the underlying hardware can satisfy (e.g. delay
1768			 * set to 1ms, but the disks take 10ms to complete an
1769			 * IO request).
1770			 */
1771
1772			DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1773			    hrtime_t, now);
1774
1775			zio_interrupt(zio);
1776		} else {
1777			hrtime_t diff = zio->io_target_timestamp - now;
1778
1779			DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1780			    hrtime_t, now, hrtime_t, diff);
1781
1782			(void) timeout_generic(CALLOUT_NORMAL,
1783			    (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1784		}
1785
1786		return;
1787	}
1788#endif
1789
1790	DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1791	zio_interrupt(zio);
1792}
1793
1794/*
1795 * Execute the I/O pipeline until one of the following occurs:
1796 *
1797 *	(1) the I/O completes
1798 *	(2) the pipeline stalls waiting for dependent child I/Os
1799 *	(3) the I/O issues, so we're waiting for an I/O completion interrupt
1800 *	(4) the I/O is delegated by vdev-level caching or aggregation
1801 *	(5) the I/O is deferred due to vdev-level queueing
1802 *	(6) the I/O is handed off to another thread.
1803 *
1804 * In all cases, the pipeline stops whenever there's no CPU work; it never
1805 * burns a thread in cv_wait().
1806 *
1807 * There's no locking on io_stage because there's no legitimate way
1808 * for multiple threads to be attempting to process the same I/O.
1809 */
1810static zio_pipe_stage_t *zio_pipeline[];
1811
1812void
1813zio_execute(zio_t *zio)
1814{
1815	zio->io_executor = curthread;
1816
1817	ASSERT3U(zio->io_queued_timestamp, >, 0);
1818
1819	while (zio->io_stage < ZIO_STAGE_DONE) {
1820		enum zio_stage pipeline = zio->io_pipeline;
1821		enum zio_stage stage = zio->io_stage;
1822		int rv;
1823
1824		ASSERT(!MUTEX_HELD(&zio->io_lock));
1825		ASSERT(ISP2(stage));
1826		ASSERT(zio->io_stall == NULL);
1827
1828		do {
1829			stage <<= 1;
1830		} while ((stage & pipeline) == 0);
1831
1832		ASSERT(stage <= ZIO_STAGE_DONE);
1833
1834		/*
1835		 * If we are in interrupt context and this pipeline stage
1836		 * will grab a config lock that is held across I/O,
1837		 * or may wait for an I/O that needs an interrupt thread
1838		 * to complete, issue async to avoid deadlock.
1839		 *
1840		 * For VDEV_IO_START, we cut in line so that the io will
1841		 * be sent to disk promptly.
1842		 */
1843		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1844		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1845			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1846			    zio_requeue_io_start_cut_in_line : B_FALSE;
1847			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1848			return;
1849		}
1850
1851		zio->io_stage = stage;
1852		zio->io_pipeline_trace |= zio->io_stage;
1853		rv = zio_pipeline[highbit64(stage) - 1](zio);
1854
1855		if (rv == ZIO_PIPELINE_STOP)
1856			return;
1857
1858		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1859	}
1860}
1861
1862/*
1863 * ==========================================================================
1864 * Initiate I/O, either sync or async
1865 * ==========================================================================
1866 */
1867int
1868zio_wait(zio_t *zio)
1869{
1870	int error;
1871
1872	ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1873	ASSERT3P(zio->io_executor, ==, NULL);
1874
1875	zio->io_waiter = curthread;
1876	ASSERT0(zio->io_queued_timestamp);
1877	zio->io_queued_timestamp = gethrtime();
1878
1879	zio_execute(zio);
1880
1881	mutex_enter(&zio->io_lock);
1882	while (zio->io_executor != NULL)
1883		cv_wait(&zio->io_cv, &zio->io_lock);
1884	mutex_exit(&zio->io_lock);
1885
1886	error = zio->io_error;
1887	zio_destroy(zio);
1888
1889	return (error);
1890}
1891
1892void
1893zio_nowait(zio_t *zio)
1894{
1895	ASSERT3P(zio->io_executor, ==, NULL);
1896
1897	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1898	    zio_unique_parent(zio) == NULL) {
1899		/*
1900		 * This is a logical async I/O with no parent to wait for it.
1901		 * We add it to the spa_async_root_zio "Godfather" I/O which
1902		 * will ensure they complete prior to unloading the pool.
1903		 */
1904		spa_t *spa = zio->io_spa;
1905
1906		zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1907	}
1908
1909	ASSERT0(zio->io_queued_timestamp);
1910	zio->io_queued_timestamp = gethrtime();
1911	zio_execute(zio);
1912}
1913
1914/*
1915 * ==========================================================================
1916 * Reexecute, cancel, or suspend/resume failed I/O
1917 * ==========================================================================
1918 */
1919
1920static void
1921zio_reexecute(zio_t *pio)
1922{
1923	zio_t *cio, *cio_next;
1924
1925	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1926	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1927	ASSERT(pio->io_gang_leader == NULL);
1928	ASSERT(pio->io_gang_tree == NULL);
1929
1930	pio->io_flags = pio->io_orig_flags;
1931	pio->io_stage = pio->io_orig_stage;
1932	pio->io_pipeline = pio->io_orig_pipeline;
1933	pio->io_reexecute = 0;
1934	pio->io_flags |= ZIO_FLAG_REEXECUTED;
1935	pio->io_pipeline_trace = 0;
1936	pio->io_error = 0;
1937	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1938		pio->io_state[w] = 0;
1939	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1940		pio->io_child_error[c] = 0;
1941
1942	if (IO_IS_ALLOCATING(pio))
1943		BP_ZERO(pio->io_bp);
1944
1945	/*
1946	 * As we reexecute pio's children, new children could be created.
1947	 * New children go to the head of pio's io_child_list, however,
1948	 * so we will (correctly) not reexecute them.  The key is that
1949	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1950	 * cannot be affected by any side effects of reexecuting 'cio'.
1951	 */
1952	zio_link_t *zl = NULL;
1953	mutex_enter(&pio->io_lock);
1954	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1955		cio_next = zio_walk_children(pio, &zl);
1956		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1957			pio->io_children[cio->io_child_type][w]++;
1958		mutex_exit(&pio->io_lock);
1959		zio_reexecute(cio);
1960		mutex_enter(&pio->io_lock);
1961	}
1962	mutex_exit(&pio->io_lock);
1963
1964	/*
1965	 * Now that all children have been reexecuted, execute the parent.
1966	 * We don't reexecute "The Godfather" I/O here as it's the
1967	 * responsibility of the caller to wait on it.
1968	 */
1969	if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1970		pio->io_queued_timestamp = gethrtime();
1971		zio_execute(pio);
1972	}
1973}
1974
1975void
1976zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
1977{
1978	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1979		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1980		    "failure and the failure mode property for this pool "
1981		    "is set to panic.", spa_name(spa));
1982
1983	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
1984	    NULL, NULL, 0, 0);
1985
1986	mutex_enter(&spa->spa_suspend_lock);
1987
1988	if (spa->spa_suspend_zio_root == NULL)
1989		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1990		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1991		    ZIO_FLAG_GODFATHER);
1992
1993	spa->spa_suspended = reason;
1994
1995	if (zio != NULL) {
1996		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1997		ASSERT(zio != spa->spa_suspend_zio_root);
1998		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1999		ASSERT(zio_unique_parent(zio) == NULL);
2000		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
2001		zio_add_child(spa->spa_suspend_zio_root, zio);
2002	}
2003
2004	mutex_exit(&spa->spa_suspend_lock);
2005}
2006
2007int
2008zio_resume(spa_t *spa)
2009{
2010	zio_t *pio;
2011
2012	/*
2013	 * Reexecute all previously suspended i/o.
2014	 */
2015	mutex_enter(&spa->spa_suspend_lock);
2016	spa->spa_suspended = ZIO_SUSPEND_NONE;
2017	cv_broadcast(&spa->spa_suspend_cv);
2018	pio = spa->spa_suspend_zio_root;
2019	spa->spa_suspend_zio_root = NULL;
2020	mutex_exit(&spa->spa_suspend_lock);
2021
2022	if (pio == NULL)
2023		return (0);
2024
2025	zio_reexecute(pio);
2026	return (zio_wait(pio));
2027}
2028
2029void
2030zio_resume_wait(spa_t *spa)
2031{
2032	mutex_enter(&spa->spa_suspend_lock);
2033	while (spa_suspended(spa))
2034		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
2035	mutex_exit(&spa->spa_suspend_lock);
2036}
2037
2038/*
2039 * ==========================================================================
2040 * Gang blocks.
2041 *
2042 * A gang block is a collection of small blocks that looks to the DMU
2043 * like one large block.  When zio_dva_allocate() cannot find a block
2044 * of the requested size, due to either severe fragmentation or the pool
2045 * being nearly full, it calls zio_write_gang_block() to construct the
2046 * block from smaller fragments.
2047 *
2048 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2049 * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
2050 * an indirect block: it's an array of block pointers.  It consumes
2051 * only one sector and hence is allocatable regardless of fragmentation.
2052 * The gang header's bps point to its gang members, which hold the data.
2053 *
2054 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2055 * as the verifier to ensure uniqueness of the SHA256 checksum.
2056 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2057 * not the gang header.  This ensures that data block signatures (needed for
2058 * deduplication) are independent of how the block is physically stored.
2059 *
2060 * Gang blocks can be nested: a gang member may itself be a gang block.
2061 * Thus every gang block is a tree in which root and all interior nodes are
2062 * gang headers, and the leaves are normal blocks that contain user data.
2063 * The root of the gang tree is called the gang leader.
2064 *
2065 * To perform any operation (read, rewrite, free, claim) on a gang block,
2066 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2067 * in the io_gang_tree field of the original logical i/o by recursively
2068 * reading the gang leader and all gang headers below it.  This yields
2069 * an in-core tree containing the contents of every gang header and the
2070 * bps for every constituent of the gang block.
2071 *
2072 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2073 * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
2074 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2075 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2076 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2077 * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
2078 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2079 * of the gang header plus zio_checksum_compute() of the data to update the
2080 * gang header's blk_cksum as described above.
2081 *
2082 * The two-phase assemble/issue model solves the problem of partial failure --
2083 * what if you'd freed part of a gang block but then couldn't read the
2084 * gang header for another part?  Assembling the entire gang tree first
2085 * ensures that all the necessary gang header I/O has succeeded before
2086 * starting the actual work of free, claim, or write.  Once the gang tree
2087 * is assembled, free and claim are in-memory operations that cannot fail.
2088 *
2089 * In the event that a gang write fails, zio_dva_unallocate() walks the
2090 * gang tree to immediately free (i.e. insert back into the space map)
2091 * everything we've allocated.  This ensures that we don't get ENOSPC
2092 * errors during repeated suspend/resume cycles due to a flaky device.
2093 *
2094 * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
2095 * the gang tree, we won't modify the block, so we can safely defer the free
2096 * (knowing that the block is still intact).  If we *can* assemble the gang
2097 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2098 * each constituent bp and we can allocate a new block on the next sync pass.
2099 *
2100 * In all cases, the gang tree allows complete recovery from partial failure.
2101 * ==========================================================================
2102 */
2103
2104static void
2105zio_gang_issue_func_done(zio_t *zio)
2106{
2107	abd_put(zio->io_abd);
2108}
2109
2110static zio_t *
2111zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2112    uint64_t offset)
2113{
2114	if (gn != NULL)
2115		return (pio);
2116
2117	return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2118	    BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2119	    NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2120	    &pio->io_bookmark));
2121}
2122
2123static zio_t *
2124zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2125    uint64_t offset)
2126{
2127	zio_t *zio;
2128
2129	if (gn != NULL) {
2130		abd_t *gbh_abd =
2131		    abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2132		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2133		    gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2134		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2135		    &pio->io_bookmark);
2136		/*
2137		 * As we rewrite each gang header, the pipeline will compute
2138		 * a new gang block header checksum for it; but no one will
2139		 * compute a new data checksum, so we do that here.  The one
2140		 * exception is the gang leader: the pipeline already computed
2141		 * its data checksum because that stage precedes gang assembly.
2142		 * (Presently, nothing actually uses interior data checksums;
2143		 * this is just good hygiene.)
2144		 */
2145		if (gn != pio->io_gang_leader->io_gang_tree) {
2146			abd_t *buf = abd_get_offset(data, offset);
2147
2148			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2149			    buf, BP_GET_PSIZE(bp));
2150
2151			abd_put(buf);
2152		}
2153		/*
2154		 * If we are here to damage data for testing purposes,
2155		 * leave the GBH alone so that we can detect the damage.
2156		 */
2157		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2158			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2159	} else {
2160		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2161		    abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2162		    zio_gang_issue_func_done, NULL, pio->io_priority,
2163		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2164	}
2165
2166	return (zio);
2167}
2168
2169/* ARGSUSED */
2170static zio_t *
2171zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2172    uint64_t offset)
2173{
2174	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2175	    ZIO_GANG_CHILD_FLAGS(pio)));
2176}
2177
2178/* ARGSUSED */
2179static zio_t *
2180zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2181    uint64_t offset)
2182{
2183	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2184	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2185}
2186
2187static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2188	NULL,
2189	zio_read_gang,
2190	zio_rewrite_gang,
2191	zio_free_gang,
2192	zio_claim_gang,
2193	NULL
2194};
2195
2196static void zio_gang_tree_assemble_done(zio_t *zio);
2197
2198static zio_gang_node_t *
2199zio_gang_node_alloc(zio_gang_node_t **gnpp)
2200{
2201	zio_gang_node_t *gn;
2202
2203	ASSERT(*gnpp == NULL);
2204
2205	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2206	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2207	*gnpp = gn;
2208
2209	return (gn);
2210}
2211
2212static void
2213zio_gang_node_free(zio_gang_node_t **gnpp)
2214{
2215	zio_gang_node_t *gn = *gnpp;
2216
2217	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2218		ASSERT(gn->gn_child[g] == NULL);
2219
2220	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2221	kmem_free(gn, sizeof (*gn));
2222	*gnpp = NULL;
2223}
2224
2225static void
2226zio_gang_tree_free(zio_gang_node_t **gnpp)
2227{
2228	zio_gang_node_t *gn = *gnpp;
2229
2230	if (gn == NULL)
2231		return;
2232
2233	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2234		zio_gang_tree_free(&gn->gn_child[g]);
2235
2236	zio_gang_node_free(gnpp);
2237}
2238
2239static void
2240zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2241{
2242	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2243	abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2244
2245	ASSERT(gio->io_gang_leader == gio);
2246	ASSERT(BP_IS_GANG(bp));
2247
2248	zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2249	    zio_gang_tree_assemble_done, gn, gio->io_priority,
2250	    ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2251}
2252
2253static void
2254zio_gang_tree_assemble_done(zio_t *zio)
2255{
2256	zio_t *gio = zio->io_gang_leader;
2257	zio_gang_node_t *gn = zio->io_private;
2258	blkptr_t *bp = zio->io_bp;
2259
2260	ASSERT(gio == zio_unique_parent(zio));
2261	ASSERT(zio->io_child_count == 0);
2262
2263	if (zio->io_error)
2264		return;
2265
2266	/* this ABD was created from a linear buf in zio_gang_tree_assemble */
2267	if (BP_SHOULD_BYTESWAP(bp))
2268		byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2269
2270	ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2271	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2272	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2273
2274	abd_put(zio->io_abd);
2275
2276	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2277		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2278		if (!BP_IS_GANG(gbp))
2279			continue;
2280		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2281	}
2282}
2283
2284static void
2285zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2286    uint64_t offset)
2287{
2288	zio_t *gio = pio->io_gang_leader;
2289	zio_t *zio;
2290
2291	ASSERT(BP_IS_GANG(bp) == !!gn);
2292	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2293	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2294
2295	/*
2296	 * If you're a gang header, your data is in gn->gn_gbh.
2297	 * If you're a gang member, your data is in 'data' and gn == NULL.
2298	 */
2299	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2300
2301	if (gn != NULL) {
2302		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2303
2304		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2305			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2306			if (BP_IS_HOLE(gbp))
2307				continue;
2308			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2309			    offset);
2310			offset += BP_GET_PSIZE(gbp);
2311		}
2312	}
2313
2314	if (gn == gio->io_gang_tree)
2315		ASSERT3U(gio->io_size, ==, offset);
2316
2317	if (zio != pio)
2318		zio_nowait(zio);
2319}
2320
2321static int
2322zio_gang_assemble(zio_t *zio)
2323{
2324	blkptr_t *bp = zio->io_bp;
2325
2326	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2327	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2328
2329	zio->io_gang_leader = zio;
2330
2331	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2332
2333	return (ZIO_PIPELINE_CONTINUE);
2334}
2335
2336static int
2337zio_gang_issue(zio_t *zio)
2338{
2339	blkptr_t *bp = zio->io_bp;
2340
2341	if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2342		return (ZIO_PIPELINE_STOP);
2343	}
2344
2345	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2346	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2347
2348	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2349		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2350		    0);
2351	else
2352		zio_gang_tree_free(&zio->io_gang_tree);
2353
2354	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2355
2356	return (ZIO_PIPELINE_CONTINUE);
2357}
2358
2359static void
2360zio_write_gang_member_ready(zio_t *zio)
2361{
2362	zio_t *pio = zio_unique_parent(zio);
2363	zio_t *gio = zio->io_gang_leader;
2364	dva_t *cdva = zio->io_bp->blk_dva;
2365	dva_t *pdva = pio->io_bp->blk_dva;
2366	uint64_t asize;
2367
2368	if (BP_IS_HOLE(zio->io_bp))
2369		return;
2370
2371	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2372
2373	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2374	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2375	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2376	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2377	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2378
2379	mutex_enter(&pio->io_lock);
2380	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2381		ASSERT(DVA_GET_GANG(&pdva[d]));
2382		asize = DVA_GET_ASIZE(&pdva[d]);
2383		asize += DVA_GET_ASIZE(&cdva[d]);
2384		DVA_SET_ASIZE(&pdva[d], asize);
2385	}
2386	mutex_exit(&pio->io_lock);
2387}
2388
2389static void
2390zio_write_gang_done(zio_t *zio)
2391{
2392	/*
2393	 * The io_abd field will be NULL for a zio with no data.  The io_flags
2394	 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2395	 * check for it here as it is cleared in zio_ready.
2396	 */
2397	if (zio->io_abd != NULL)
2398		abd_put(zio->io_abd);
2399}
2400
2401static int
2402zio_write_gang_block(zio_t *pio)
2403{
2404	spa_t *spa = pio->io_spa;
2405	metaslab_class_t *mc = spa_normal_class(spa);
2406	blkptr_t *bp = pio->io_bp;
2407	zio_t *gio = pio->io_gang_leader;
2408	zio_t *zio;
2409	zio_gang_node_t *gn, **gnpp;
2410	zio_gbh_phys_t *gbh;
2411	abd_t *gbh_abd;
2412	uint64_t txg = pio->io_txg;
2413	uint64_t resid = pio->io_size;
2414	uint64_t lsize;
2415	int copies = gio->io_prop.zp_copies;
2416	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2417	zio_prop_t zp;
2418	int error;
2419	boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
2420
2421	/*
2422	 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2423	 * have a third copy.
2424	 */
2425	if (gio->io_prop.zp_encrypt && gbh_copies >= SPA_DVAS_PER_BP)
2426		gbh_copies = SPA_DVAS_PER_BP - 1;
2427
2428	int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2429	if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2430		ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2431		ASSERT(has_data);
2432
2433		flags |= METASLAB_ASYNC_ALLOC;
2434		VERIFY(zfs_refcount_held(&mc->mc_alloc_slots[pio->io_allocator],
2435		    pio));
2436
2437		/*
2438		 * The logical zio has already placed a reservation for
2439		 * 'copies' allocation slots but gang blocks may require
2440		 * additional copies. These additional copies
2441		 * (i.e. gbh_copies - copies) are guaranteed to succeed
2442		 * since metaslab_class_throttle_reserve() always allows
2443		 * additional reservations for gang blocks.
2444		 */
2445		VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2446		    pio->io_allocator, pio, flags));
2447	}
2448
2449	error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2450	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2451	    &pio->io_alloc_list, pio, pio->io_allocator);
2452	if (error) {
2453		if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2454			ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2455			ASSERT(has_data);
2456
2457			/*
2458			 * If we failed to allocate the gang block header then
2459			 * we remove any additional allocation reservations that
2460			 * we placed here. The original reservation will
2461			 * be removed when the logical I/O goes to the ready
2462			 * stage.
2463			 */
2464			metaslab_class_throttle_unreserve(mc,
2465			    gbh_copies - copies, pio->io_allocator, pio);
2466		}
2467		pio->io_error = error;
2468		return (ZIO_PIPELINE_CONTINUE);
2469	}
2470
2471	if (pio == gio) {
2472		gnpp = &gio->io_gang_tree;
2473	} else {
2474		gnpp = pio->io_private;
2475		ASSERT(pio->io_ready == zio_write_gang_member_ready);
2476	}
2477
2478	gn = zio_gang_node_alloc(gnpp);
2479	gbh = gn->gn_gbh;
2480	bzero(gbh, SPA_GANGBLOCKSIZE);
2481	gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2482
2483	/*
2484	 * Create the gang header.
2485	 */
2486	zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2487	    zio_write_gang_done, NULL, pio->io_priority,
2488	    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2489
2490	/*
2491	 * Create and nowait the gang children.
2492	 */
2493	for (int g = 0; resid != 0; resid -= lsize, g++) {
2494		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2495		    SPA_MINBLOCKSIZE);
2496		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2497
2498		zp.zp_checksum = gio->io_prop.zp_checksum;
2499		zp.zp_compress = ZIO_COMPRESS_OFF;
2500		zp.zp_type = DMU_OT_NONE;
2501		zp.zp_level = 0;
2502		zp.zp_copies = gio->io_prop.zp_copies;
2503		zp.zp_dedup = B_FALSE;
2504		zp.zp_dedup_verify = B_FALSE;
2505		zp.zp_nopwrite = B_FALSE;
2506		zp.zp_encrypt = gio->io_prop.zp_encrypt;
2507		zp.zp_byteorder = gio->io_prop.zp_byteorder;
2508		bzero(zp.zp_salt, ZIO_DATA_SALT_LEN);
2509		bzero(zp.zp_iv, ZIO_DATA_IV_LEN);
2510		bzero(zp.zp_mac, ZIO_DATA_MAC_LEN);
2511
2512		zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2513		    has_data ? abd_get_offset(pio->io_abd, pio->io_size -
2514		    resid) : NULL, lsize, lsize, &zp,
2515		    zio_write_gang_member_ready, NULL, NULL,
2516		    zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2517		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2518
2519		if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2520			ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2521			ASSERT(has_data);
2522
2523			/*
2524			 * Gang children won't throttle but we should
2525			 * account for their work, so reserve an allocation
2526			 * slot for them here.
2527			 */
2528			VERIFY(metaslab_class_throttle_reserve(mc,
2529			    zp.zp_copies, cio->io_allocator, cio, flags));
2530		}
2531		zio_nowait(cio);
2532	}
2533
2534	/*
2535	 * Set pio's pipeline to just wait for zio to finish.
2536	 */
2537	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2538
2539	zio_nowait(zio);
2540
2541	return (ZIO_PIPELINE_CONTINUE);
2542}
2543
2544/*
2545 * The zio_nop_write stage in the pipeline determines if allocating a
2546 * new bp is necessary.  The nopwrite feature can handle writes in
2547 * either syncing or open context (i.e. zil writes) and as a result is
2548 * mutually exclusive with dedup.
2549 *
2550 * By leveraging a cryptographically secure checksum, such as SHA256, we
2551 * can compare the checksums of the new data and the old to determine if
2552 * allocating a new block is required.  Note that our requirements for
2553 * cryptographic strength are fairly weak: there can't be any accidental
2554 * hash collisions, but we don't need to be secure against intentional
2555 * (malicious) collisions.  To trigger a nopwrite, you have to be able
2556 * to write the file to begin with, and triggering an incorrect (hash
2557 * collision) nopwrite is no worse than simply writing to the file.
2558 * That said, there are no known attacks against the checksum algorithms
2559 * used for nopwrite, assuming that the salt and the checksums
2560 * themselves remain secret.
2561 */
2562static int
2563zio_nop_write(zio_t *zio)
2564{
2565	blkptr_t *bp = zio->io_bp;
2566	blkptr_t *bp_orig = &zio->io_bp_orig;
2567	zio_prop_t *zp = &zio->io_prop;
2568
2569	ASSERT(BP_GET_LEVEL(bp) == 0);
2570	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2571	ASSERT(zp->zp_nopwrite);
2572	ASSERT(!zp->zp_dedup);
2573	ASSERT(zio->io_bp_override == NULL);
2574	ASSERT(IO_IS_ALLOCATING(zio));
2575
2576	/*
2577	 * Check to see if the original bp and the new bp have matching
2578	 * characteristics (i.e. same checksum, compression algorithms, etc).
2579	 * If they don't then just continue with the pipeline which will
2580	 * allocate a new bp.
2581	 */
2582	if (BP_IS_HOLE(bp_orig) ||
2583	    !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2584	    ZCHECKSUM_FLAG_NOPWRITE) ||
2585	    BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) ||
2586	    BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2587	    BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2588	    BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2589	    zp->zp_copies != BP_GET_NDVAS(bp_orig))
2590		return (ZIO_PIPELINE_CONTINUE);
2591
2592	/*
2593	 * If the checksums match then reset the pipeline so that we
2594	 * avoid allocating a new bp and issuing any I/O.
2595	 */
2596	if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2597		ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2598		    ZCHECKSUM_FLAG_NOPWRITE);
2599		ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2600		ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2601		ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2602		ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2603		    sizeof (uint64_t)) == 0);
2604
2605		*bp = *bp_orig;
2606		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2607		zio->io_flags |= ZIO_FLAG_NOPWRITE;
2608	}
2609
2610	return (ZIO_PIPELINE_CONTINUE);
2611}
2612
2613/*
2614 * ==========================================================================
2615 * Dedup
2616 * ==========================================================================
2617 */
2618static void
2619zio_ddt_child_read_done(zio_t *zio)
2620{
2621	blkptr_t *bp = zio->io_bp;
2622	ddt_entry_t *dde = zio->io_private;
2623	ddt_phys_t *ddp;
2624	zio_t *pio = zio_unique_parent(zio);
2625
2626	mutex_enter(&pio->io_lock);
2627	ddp = ddt_phys_select(dde, bp);
2628	if (zio->io_error == 0)
2629		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
2630
2631	if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2632		dde->dde_repair_abd = zio->io_abd;
2633	else
2634		abd_free(zio->io_abd);
2635	mutex_exit(&pio->io_lock);
2636}
2637
2638static int
2639zio_ddt_read_start(zio_t *zio)
2640{
2641	blkptr_t *bp = zio->io_bp;
2642
2643	ASSERT(BP_GET_DEDUP(bp));
2644	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2645	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2646
2647	if (zio->io_child_error[ZIO_CHILD_DDT]) {
2648		ddt_t *ddt = ddt_select(zio->io_spa, bp);
2649		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2650		ddt_phys_t *ddp = dde->dde_phys;
2651		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2652		blkptr_t blk;
2653
2654		ASSERT(zio->io_vsd == NULL);
2655		zio->io_vsd = dde;
2656
2657		if (ddp_self == NULL)
2658			return (ZIO_PIPELINE_CONTINUE);
2659
2660		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2661			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2662				continue;
2663			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2664			    &blk);
2665			zio_nowait(zio_read(zio, zio->io_spa, &blk,
2666			    abd_alloc_for_io(zio->io_size, B_TRUE),
2667			    zio->io_size, zio_ddt_child_read_done, dde,
2668			    zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2669			    ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2670		}
2671		return (ZIO_PIPELINE_CONTINUE);
2672	}
2673
2674	zio_nowait(zio_read(zio, zio->io_spa, bp,
2675	    zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2676	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2677
2678	return (ZIO_PIPELINE_CONTINUE);
2679}
2680
2681static int
2682zio_ddt_read_done(zio_t *zio)
2683{
2684	blkptr_t *bp = zio->io_bp;
2685
2686	if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2687		return (ZIO_PIPELINE_STOP);
2688	}
2689
2690	ASSERT(BP_GET_DEDUP(bp));
2691	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2692	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2693
2694	if (zio->io_child_error[ZIO_CHILD_DDT]) {
2695		ddt_t *ddt = ddt_select(zio->io_spa, bp);
2696		ddt_entry_t *dde = zio->io_vsd;
2697		if (ddt == NULL) {
2698			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2699			return (ZIO_PIPELINE_CONTINUE);
2700		}
2701		if (dde == NULL) {
2702			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2703			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2704			return (ZIO_PIPELINE_STOP);
2705		}
2706		if (dde->dde_repair_abd != NULL) {
2707			abd_copy(zio->io_abd, dde->dde_repair_abd,
2708			    zio->io_size);
2709			zio->io_child_error[ZIO_CHILD_DDT] = 0;
2710		}
2711		ddt_repair_done(ddt, dde);
2712		zio->io_vsd = NULL;
2713	}
2714
2715	ASSERT(zio->io_vsd == NULL);
2716
2717	return (ZIO_PIPELINE_CONTINUE);
2718}
2719
2720static boolean_t
2721zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2722{
2723	spa_t *spa = zio->io_spa;
2724	boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW);
2725
2726	/* We should never get a raw, override zio */
2727	ASSERT(!(zio->io_bp_override && do_raw));
2728
2729	/*
2730	 * Note: we compare the original data, not the transformed data,
2731	 * because when zio->io_bp is an override bp, we will not have
2732	 * pushed the I/O transforms.  That's an important optimization
2733	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2734	 * However, we should never get a raw, override zio so in these
2735	 * cases we can compare the io_data directly. This is useful because
2736	 * it allows us to do dedup verification even if we don't have access
2737	 * to the original data (for instance, if the encryption keys aren't
2738	 * loaded).
2739	 */
2740
2741	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2742		zio_t *lio = dde->dde_lead_zio[p];
2743
2744		if (lio != NULL && do_raw) {
2745			return (lio->io_size != zio->io_size ||
2746			    abd_cmp(zio->io_abd, lio->io_abd,
2747			    zio->io_size) != 0);
2748		} else if (lio != NULL) {
2749			return (lio->io_orig_size != zio->io_orig_size ||
2750			    abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2751			    zio->io_orig_size) != 0);
2752		}
2753	}
2754
2755	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2756		ddt_phys_t *ddp = &dde->dde_phys[p];
2757
2758		if (ddp->ddp_phys_birth != 0 && do_raw) {
2759			blkptr_t blk = *zio->io_bp;
2760			uint64_t psize;
2761			abd_t *tmpabd;
2762			int error;
2763
2764			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2765			psize = BP_GET_PSIZE(&blk);
2766
2767			if (psize != zio->io_size)
2768				return (B_TRUE);
2769
2770			ddt_exit(ddt);
2771
2772			tmpabd = abd_alloc_for_io(psize, B_TRUE);
2773
2774			error = zio_wait(zio_read(NULL, spa, &blk, tmpabd,
2775			    psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
2776			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2777			    ZIO_FLAG_RAW, &zio->io_bookmark));
2778
2779			if (error == 0) {
2780				if (abd_cmp(tmpabd, zio->io_abd, psize) != 0)
2781					error = SET_ERROR(ENOENT);
2782			}
2783
2784			abd_free(tmpabd);
2785			ddt_enter(ddt);
2786			return (error != 0);
2787		} else if (ddp->ddp_phys_birth != 0) {
2788			arc_buf_t *abuf = NULL;
2789			arc_flags_t aflags = ARC_FLAG_WAIT;
2790			int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2791			blkptr_t blk = *zio->io_bp;
2792			int error;
2793
2794			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2795
2796			if (BP_GET_LSIZE(&blk) != zio->io_orig_size)
2797				return (B_TRUE);
2798
2799			ddt_exit(ddt);
2800
2801			/*
2802			 * Intuitively, it would make more sense to compare
2803			 * io_abd than io_orig_abd in the raw case since you
2804			 * don't want to look at any transformations that have
2805			 * happened to the data. However, for raw I/Os the
2806			 * data will actually be the same in io_abd and
2807			 * io_orig_abd, so all we have to do is issue this as
2808			 * a raw ARC read.
2809			 */
2810			if (do_raw) {
2811				zio_flags |= ZIO_FLAG_RAW;
2812				ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2813				ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2814				    zio->io_size));
2815				ASSERT3P(zio->io_transform_stack, ==, NULL);
2816			}
2817
2818			error = arc_read(NULL, spa, &blk,
2819			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2820			    zio_flags, &aflags, &zio->io_bookmark);
2821
2822			if (error == 0) {
2823				if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2824				    zio->io_orig_size) != 0)
2825					error = SET_ERROR(ENOENT);
2826				arc_buf_destroy(abuf, &abuf);
2827			}
2828
2829			ddt_enter(ddt);
2830			return (error != 0);
2831		}
2832	}
2833
2834	return (B_FALSE);
2835}
2836
2837static void
2838zio_ddt_child_write_ready(zio_t *zio)
2839{
2840	int p = zio->io_prop.zp_copies;
2841	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2842	ddt_entry_t *dde = zio->io_private;
2843	ddt_phys_t *ddp = &dde->dde_phys[p];
2844	zio_t *pio;
2845
2846	if (zio->io_error)
2847		return;
2848
2849	ddt_enter(ddt);
2850
2851	ASSERT(dde->dde_lead_zio[p] == zio);
2852
2853	ddt_phys_fill(ddp, zio->io_bp);
2854
2855	zio_link_t *zl = NULL;
2856	while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2857		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2858
2859	ddt_exit(ddt);
2860}
2861
2862static void
2863zio_ddt_child_write_done(zio_t *zio)
2864{
2865	int p = zio->io_prop.zp_copies;
2866	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2867	ddt_entry_t *dde = zio->io_private;
2868	ddt_phys_t *ddp = &dde->dde_phys[p];
2869
2870	ddt_enter(ddt);
2871
2872	ASSERT(ddp->ddp_refcnt == 0);
2873	ASSERT(dde->dde_lead_zio[p] == zio);
2874	dde->dde_lead_zio[p] = NULL;
2875
2876	if (zio->io_error == 0) {
2877		zio_link_t *zl = NULL;
2878		while (zio_walk_parents(zio, &zl) != NULL)
2879			ddt_phys_addref(ddp);
2880	} else {
2881		ddt_phys_clear(ddp);
2882	}
2883
2884	ddt_exit(ddt);
2885}
2886
2887static void
2888zio_ddt_ditto_write_done(zio_t *zio)
2889{
2890	int p = DDT_PHYS_DITTO;
2891	zio_prop_t *zp = &zio->io_prop;
2892	blkptr_t *bp = zio->io_bp;
2893	ddt_t *ddt = ddt_select(zio->io_spa, bp);
2894	ddt_entry_t *dde = zio->io_private;
2895	ddt_phys_t *ddp = &dde->dde_phys[p];
2896	ddt_key_t *ddk = &dde->dde_key;
2897
2898	ddt_enter(ddt);
2899
2900	ASSERT(ddp->ddp_refcnt == 0);
2901	ASSERT(dde->dde_lead_zio[p] == zio);
2902	dde->dde_lead_zio[p] = NULL;
2903
2904	if (zio->io_error == 0) {
2905		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2906		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2907		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2908		if (ddp->ddp_phys_birth != 0)
2909			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2910		ddt_phys_fill(ddp, bp);
2911	}
2912
2913	ddt_exit(ddt);
2914}
2915
2916static int
2917zio_ddt_write(zio_t *zio)
2918{
2919	spa_t *spa = zio->io_spa;
2920	blkptr_t *bp = zio->io_bp;
2921	uint64_t txg = zio->io_txg;
2922	zio_prop_t *zp = &zio->io_prop;
2923	int p = zp->zp_copies;
2924	int ditto_copies;
2925	zio_t *cio = NULL;
2926	zio_t *dio = NULL;
2927	ddt_t *ddt = ddt_select(spa, bp);
2928	ddt_entry_t *dde;
2929	ddt_phys_t *ddp;
2930
2931	ASSERT(BP_GET_DEDUP(bp));
2932	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2933	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2934	ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2935
2936	ddt_enter(ddt);
2937	dde = ddt_lookup(ddt, bp, B_TRUE);
2938	ddp = &dde->dde_phys[p];
2939
2940	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2941		/*
2942		 * If we're using a weak checksum, upgrade to a strong checksum
2943		 * and try again.  If we're already using a strong checksum,
2944		 * we can't resolve it, so just convert to an ordinary write.
2945		 * (And automatically e-mail a paper to Nature?)
2946		 */
2947		if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2948		    ZCHECKSUM_FLAG_DEDUP)) {
2949			zp->zp_checksum = spa_dedup_checksum(spa);
2950			zio_pop_transforms(zio);
2951			zio->io_stage = ZIO_STAGE_OPEN;
2952			BP_ZERO(bp);
2953		} else {
2954			zp->zp_dedup = B_FALSE;
2955			BP_SET_DEDUP(bp, B_FALSE);
2956		}
2957		ASSERT(!BP_GET_DEDUP(bp));
2958		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2959		ddt_exit(ddt);
2960		return (ZIO_PIPELINE_CONTINUE);
2961	}
2962
2963	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2964	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2965
2966	if (ditto_copies > ddt_ditto_copies_present(dde) &&
2967	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2968		zio_prop_t czp = *zp;
2969
2970		czp.zp_copies = ditto_copies;
2971
2972		/*
2973		 * If we arrived here with an override bp, we won't have run
2974		 * the transform stack, so we won't have the data we need to
2975		 * generate a child i/o.  So, toss the override bp and restart.
2976		 * This is safe, because using the override bp is just an
2977		 * optimization; and it's rare, so the cost doesn't matter.
2978		 */
2979		if (zio->io_bp_override) {
2980			zio_pop_transforms(zio);
2981			zio->io_stage = ZIO_STAGE_OPEN;
2982			zio->io_pipeline = ZIO_WRITE_PIPELINE;
2983			zio->io_bp_override = NULL;
2984			BP_ZERO(bp);
2985			ddt_exit(ddt);
2986			return (ZIO_PIPELINE_CONTINUE);
2987		}
2988
2989		dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2990		    zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2991		    NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2992		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2993
2994		zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2995		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2996	}
2997
2998	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2999		if (ddp->ddp_phys_birth != 0)
3000			ddt_bp_fill(ddp, bp, txg);
3001		if (dde->dde_lead_zio[p] != NULL)
3002			zio_add_child(zio, dde->dde_lead_zio[p]);
3003		else
3004			ddt_phys_addref(ddp);
3005	} else if (zio->io_bp_override) {
3006		ASSERT(bp->blk_birth == txg);
3007		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
3008		ddt_phys_fill(ddp, bp);
3009		ddt_phys_addref(ddp);
3010	} else {
3011		cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3012		    zio->io_orig_size, zio->io_orig_size, zp,
3013		    zio_ddt_child_write_ready, NULL, NULL,
3014		    zio_ddt_child_write_done, dde, zio->io_priority,
3015		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3016
3017		zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
3018		dde->dde_lead_zio[p] = cio;
3019	}
3020
3021	ddt_exit(ddt);
3022
3023	if (cio)
3024		zio_nowait(cio);
3025	if (dio)
3026		zio_nowait(dio);
3027
3028	return (ZIO_PIPELINE_CONTINUE);
3029}
3030
3031ddt_entry_t *freedde; /* for debugging */
3032
3033static int
3034zio_ddt_free(zio_t *zio)
3035{
3036	spa_t *spa = zio->io_spa;
3037	blkptr_t *bp = zio->io_bp;
3038	ddt_t *ddt = ddt_select(spa, bp);
3039	ddt_entry_t *dde;
3040	ddt_phys_t *ddp;
3041
3042	ASSERT(BP_GET_DEDUP(bp));
3043	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3044
3045	ddt_enter(ddt);
3046	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
3047	ddp = ddt_phys_select(dde, bp);
3048	ddt_phys_decref(ddp);
3049	ddt_exit(ddt);
3050
3051	return (ZIO_PIPELINE_CONTINUE);
3052}
3053
3054/*
3055 * ==========================================================================
3056 * Allocate and free blocks
3057 * ==========================================================================
3058 */
3059
3060static zio_t *
3061zio_io_to_allocate(spa_t *spa, int allocator)
3062{
3063	zio_t *zio;
3064
3065	ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator]));
3066
3067	zio = avl_first(&spa->spa_alloc_trees[allocator]);
3068	if (zio == NULL)
3069		return (NULL);
3070
3071	ASSERT(IO_IS_ALLOCATING(zio));
3072
3073	/*
3074	 * Try to place a reservation for this zio. If we're unable to
3075	 * reserve then we throttle.
3076	 */
3077	ASSERT3U(zio->io_allocator, ==, allocator);
3078	if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
3079	    zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) {
3080		return (NULL);
3081	}
3082
3083	avl_remove(&spa->spa_alloc_trees[allocator], zio);
3084	ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
3085
3086	return (zio);
3087}
3088
3089static int
3090zio_dva_throttle(zio_t *zio)
3091{
3092	spa_t *spa = zio->io_spa;
3093	zio_t *nio;
3094	metaslab_class_t *mc;
3095
3096	/* locate an appropriate allocation class */
3097	mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type,
3098	    zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk);
3099
3100	if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
3101	    !mc->mc_alloc_throttle_enabled ||
3102	    zio->io_child_type == ZIO_CHILD_GANG ||
3103	    zio->io_flags & ZIO_FLAG_NODATA) {
3104		return (ZIO_PIPELINE_CONTINUE);
3105	}
3106
3107	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3108
3109	ASSERT3U(zio->io_queued_timestamp, >, 0);
3110	ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3111
3112	zbookmark_phys_t *bm = &zio->io_bookmark;
3113	/*
3114	 * We want to try to use as many allocators as possible to help improve
3115	 * performance, but we also want logically adjacent IOs to be physically
3116	 * adjacent to improve sequential read performance. We chunk each object
3117	 * into 2^20 block regions, and then hash based on the objset, object,
3118	 * level, and region to accomplish both of these goals.
3119	 */
3120	zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object,
3121	    bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
3122	mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]);
3123	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3124	zio->io_metaslab_class = mc;
3125	avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio);
3126	nio = zio_io_to_allocate(spa, zio->io_allocator);
3127	mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]);
3128
3129	if (nio == zio)
3130		return (ZIO_PIPELINE_CONTINUE);
3131
3132	if (nio != NULL) {
3133		ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3134		/*
3135		 * We are passing control to a new zio so make sure that
3136		 * it is processed by a different thread. We do this to
3137		 * avoid stack overflows that can occur when parents are
3138		 * throttled and children are making progress. We allow
3139		 * it to go to the head of the taskq since it's already
3140		 * been waiting.
3141		 */
3142		zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
3143	}
3144	return (ZIO_PIPELINE_STOP);
3145}
3146
3147static void
3148zio_allocate_dispatch(spa_t *spa, int allocator)
3149{
3150	zio_t *zio;
3151
3152	mutex_enter(&spa->spa_alloc_locks[allocator]);
3153	zio = zio_io_to_allocate(spa, allocator);
3154	mutex_exit(&spa->spa_alloc_locks[allocator]);
3155	if (zio == NULL)
3156		return;
3157
3158	ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3159	ASSERT0(zio->io_error);
3160	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3161}
3162
3163static int
3164zio_dva_allocate(zio_t *zio)
3165{
3166	spa_t *spa = zio->io_spa;
3167	metaslab_class_t *mc;
3168	blkptr_t *bp = zio->io_bp;
3169	int error;
3170	int flags = 0;
3171
3172	if (zio->io_gang_leader == NULL) {
3173		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3174		zio->io_gang_leader = zio;
3175	}
3176
3177	ASSERT(BP_IS_HOLE(bp));
3178	ASSERT0(BP_GET_NDVAS(bp));
3179	ASSERT3U(zio->io_prop.zp_copies, >, 0);
3180	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3181	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3182
3183	if (zio->io_flags & ZIO_FLAG_NODATA)
3184		flags |= METASLAB_DONT_THROTTLE;
3185	if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
3186		flags |= METASLAB_GANG_CHILD;
3187	if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
3188		flags |= METASLAB_ASYNC_ALLOC;
3189
3190	/*
3191	 * if not already chosen, locate an appropriate allocation class
3192	 */
3193	mc = zio->io_metaslab_class;
3194	if (mc == NULL) {
3195		mc = spa_preferred_class(spa, zio->io_size,
3196		    zio->io_prop.zp_type, zio->io_prop.zp_level,
3197		    zio->io_prop.zp_zpl_smallblk);
3198		zio->io_metaslab_class = mc;
3199	}
3200
3201	error = metaslab_alloc(spa, mc, zio->io_size, bp,
3202	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3203	    &zio->io_alloc_list, zio, zio->io_allocator);
3204
3205	/*
3206	 * Fallback to normal class when an alloc class is full
3207	 */
3208	if (error == ENOSPC && mc != spa_normal_class(spa)) {
3209		/*
3210		 * If throttling, transfer reservation over to normal class.
3211		 * The io_allocator slot can remain the same even though we
3212		 * are switching classes.
3213		 */
3214		if (mc->mc_alloc_throttle_enabled &&
3215		    (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) {
3216			metaslab_class_throttle_unreserve(mc,
3217			    zio->io_prop.zp_copies, zio->io_allocator, zio);
3218			zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING;
3219
3220			mc = spa_normal_class(spa);
3221			VERIFY(metaslab_class_throttle_reserve(mc,
3222			    zio->io_prop.zp_copies, zio->io_allocator, zio,
3223			    flags | METASLAB_MUST_RESERVE));
3224		} else {
3225			mc = spa_normal_class(spa);
3226		}
3227		zio->io_metaslab_class = mc;
3228
3229		error = metaslab_alloc(spa, mc, zio->io_size, bp,
3230		    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3231		    &zio->io_alloc_list, zio, zio->io_allocator);
3232	}
3233
3234	if (error != 0) {
3235		zfs_dbgmsg("%s: metaslab allocation failure: zio %p, "
3236		    "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3237		    error);
3238		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
3239			return (zio_write_gang_block(zio));
3240		zio->io_error = error;
3241	}
3242
3243	return (ZIO_PIPELINE_CONTINUE);
3244}
3245
3246static int
3247zio_dva_free(zio_t *zio)
3248{
3249	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3250
3251	return (ZIO_PIPELINE_CONTINUE);
3252}
3253
3254static int
3255zio_dva_claim(zio_t *zio)
3256{
3257	int error;
3258
3259	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3260	if (error)
3261		zio->io_error = error;
3262
3263	return (ZIO_PIPELINE_CONTINUE);
3264}
3265
3266/*
3267 * Undo an allocation.  This is used by zio_done() when an I/O fails
3268 * and we want to give back the block we just allocated.
3269 * This handles both normal blocks and gang blocks.
3270 */
3271static void
3272zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3273{
3274	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3275	ASSERT(zio->io_bp_override == NULL);
3276
3277	if (!BP_IS_HOLE(bp))
3278		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3279
3280	if (gn != NULL) {
3281		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3282			zio_dva_unallocate(zio, gn->gn_child[g],
3283			    &gn->gn_gbh->zg_blkptr[g]);
3284		}
3285	}
3286}
3287
3288/*
3289 * Try to allocate an intent log block.  Return 0 on success, errno on failure.
3290 */
3291int
3292zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
3293    blkptr_t *old_bp, uint64_t size, boolean_t *slog)
3294{
3295	int error = 1;
3296	zio_alloc_list_t io_alloc_list;
3297
3298	ASSERT(txg > spa_syncing_txg(spa));
3299
3300	metaslab_trace_init(&io_alloc_list);
3301
3302	/*
3303	 * Block pointer fields are useful to metaslabs for stats and debugging.
3304	 * Fill in the obvious ones before calling into metaslab_alloc().
3305	 */
3306	BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3307	BP_SET_PSIZE(new_bp, size);
3308	BP_SET_LEVEL(new_bp, 0);
3309
3310	/*
3311	 * When allocating a zil block, we don't have information about
3312	 * the final destination of the block except the objset it's part
3313	 * of, so we just hash the objset ID to pick the allocator to get
3314	 * some parallelism.
3315	 */
3316	error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3317	    txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL,
3318	    cityhash4(0, 0, 0,
3319	    os->os_dsl_dataset->ds_object) % spa->spa_alloc_count);
3320	if (error == 0) {
3321		*slog = TRUE;
3322	} else {
3323		error = metaslab_alloc(spa, spa_normal_class(spa), size,
3324		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3325		    &io_alloc_list, NULL, cityhash4(0, 0, 0,
3326		    os->os_dsl_dataset->ds_object) % spa->spa_alloc_count);
3327		if (error == 0)
3328			*slog = FALSE;
3329	}
3330	metaslab_trace_fini(&io_alloc_list);
3331
3332	if (error == 0) {
3333		BP_SET_LSIZE(new_bp, size);
3334		BP_SET_PSIZE(new_bp, size);
3335		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3336		BP_SET_CHECKSUM(new_bp,
3337		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3338		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3339		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3340		BP_SET_LEVEL(new_bp, 0);
3341		BP_SET_DEDUP(new_bp, 0);
3342		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3343
3344		/*
3345		 * encrypted blocks will require an IV and salt. We generate
3346		 * these now since we will not be rewriting the bp at
3347		 * rewrite time.
3348		 */
3349		if (os->os_encrypted) {
3350			uint8_t iv[ZIO_DATA_IV_LEN];
3351			uint8_t salt[ZIO_DATA_SALT_LEN];
3352
3353			BP_SET_CRYPT(new_bp, B_TRUE);
3354			VERIFY0(spa_crypt_get_salt(spa,
3355			    dmu_objset_id(os), salt));
3356			VERIFY0(zio_crypt_generate_iv(iv));
3357
3358			zio_crypt_encode_params_bp(new_bp, salt, iv);
3359		}
3360	} else {
3361		zfs_dbgmsg("%s: zil block allocation failure: "
3362		    "size %llu, error %d", spa_name(spa), size, error);
3363	}
3364
3365	return (error);
3366}
3367
3368/*
3369 * ==========================================================================
3370 * Read and write to physical devices
3371 * ==========================================================================
3372 */
3373
3374/*
3375 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3376 * stops after this stage and will resume upon I/O completion.
3377 * However, there are instances where the vdev layer may need to
3378 * continue the pipeline when an I/O was not issued. Since the I/O
3379 * that was sent to the vdev layer might be different than the one
3380 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3381 * force the underlying vdev layers to call either zio_execute() or
3382 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3383 */
3384static int
3385zio_vdev_io_start(zio_t *zio)
3386{
3387	vdev_t *vd = zio->io_vd;
3388	uint64_t align;
3389	spa_t *spa = zio->io_spa;
3390
3391	ASSERT(zio->io_error == 0);
3392	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3393
3394	if (vd == NULL) {
3395		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3396			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3397
3398		/*
3399		 * The mirror_ops handle multiple DVAs in a single BP.
3400		 */
3401		vdev_mirror_ops.vdev_op_io_start(zio);
3402		return (ZIO_PIPELINE_STOP);
3403	}
3404
3405	ASSERT3P(zio->io_logical, !=, zio);
3406	if (zio->io_type == ZIO_TYPE_WRITE) {
3407		ASSERT(spa->spa_trust_config);
3408
3409		/*
3410		 * Note: the code can handle other kinds of writes,
3411		 * but we don't expect them.
3412		 */
3413		if (zio->io_vd->vdev_removing) {
3414			ASSERT(zio->io_flags &
3415			    (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3416			    ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3417		}
3418	}
3419
3420	align = 1ULL << vd->vdev_top->vdev_ashift;
3421
3422	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3423	    P2PHASE(zio->io_size, align) != 0) {
3424		/* Transform logical writes to be a full physical block size. */
3425		uint64_t asize = P2ROUNDUP(zio->io_size, align);
3426		abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3427		ASSERT(vd == vd->vdev_top);
3428		if (zio->io_type == ZIO_TYPE_WRITE) {
3429			abd_copy(abuf, zio->io_abd, zio->io_size);
3430			abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3431		}
3432		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3433	}
3434
3435	/*
3436	 * If this is not a physical io, make sure that it is properly aligned
3437	 * before proceeding.
3438	 */
3439	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3440		ASSERT0(P2PHASE(zio->io_offset, align));
3441		ASSERT0(P2PHASE(zio->io_size, align));
3442	} else {
3443		/*
3444		 * For physical writes, we allow 512b aligned writes and assume
3445		 * the device will perform a read-modify-write as necessary.
3446		 */
3447		ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3448		ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3449	}
3450
3451	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
3452
3453	/*
3454	 * If this is a repair I/O, and there's no self-healing involved --
3455	 * that is, we're just resilvering what we expect to resilver --
3456	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3457	 * This prevents spurious resilvering.
3458	 *
3459	 * There are a few ways that we can end up creating these spurious
3460	 * resilver i/os:
3461	 *
3462	 * 1. A resilver i/o will be issued if any DVA in the BP has a
3463	 * dirty DTL.  The mirror code will issue resilver writes to
3464	 * each DVA, including the one(s) that are not on vdevs with dirty
3465	 * DTLs.
3466	 *
3467	 * 2. With nested replication, which happens when we have a
3468	 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3469	 * For example, given mirror(replacing(A+B), C), it's likely that
3470	 * only A is out of date (it's the new device). In this case, we'll
3471	 * read from C, then use the data to resilver A+B -- but we don't
3472	 * actually want to resilver B, just A. The top-level mirror has no
3473	 * way to know this, so instead we just discard unnecessary repairs
3474	 * as we work our way down the vdev tree.
3475	 *
3476	 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3477	 * The same logic applies to any form of nested replication: ditto
3478	 * + mirror, RAID-Z + replacing, etc.
3479	 *
3480	 * However, indirect vdevs point off to other vdevs which may have
3481	 * DTL's, so we never bypass them.  The child i/os on concrete vdevs
3482	 * will be properly bypassed instead.
3483	 */
3484	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3485	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3486	    zio->io_txg != 0 &&	/* not a delegated i/o */
3487	    vd->vdev_ops != &vdev_indirect_ops &&
3488	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3489		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3490		zio_vdev_io_bypass(zio);
3491		return (ZIO_PIPELINE_CONTINUE);
3492	}
3493
3494	if (vd->vdev_ops->vdev_op_leaf && (zio->io_type == ZIO_TYPE_READ ||
3495	    zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM)) {
3496
3497		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
3498			return (ZIO_PIPELINE_CONTINUE);
3499
3500		if ((zio = vdev_queue_io(zio)) == NULL)
3501			return (ZIO_PIPELINE_STOP);
3502
3503		if (!vdev_accessible(vd, zio)) {
3504			zio->io_error = SET_ERROR(ENXIO);
3505			zio_interrupt(zio);
3506			return (ZIO_PIPELINE_STOP);
3507		}
3508	}
3509
3510	vd->vdev_ops->vdev_op_io_start(zio);
3511	return (ZIO_PIPELINE_STOP);
3512}
3513
3514static int
3515zio_vdev_io_done(zio_t *zio)
3516{
3517	vdev_t *vd = zio->io_vd;
3518	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3519	boolean_t unexpected_error = B_FALSE;
3520
3521	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3522		return (ZIO_PIPELINE_STOP);
3523	}
3524
3525	ASSERT(zio->io_type == ZIO_TYPE_READ ||
3526	    zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM);
3527
3528	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
3529
3530		vdev_queue_io_done(zio);
3531
3532		if (zio->io_type == ZIO_TYPE_WRITE)
3533			vdev_cache_write(zio);
3534
3535		if (zio_injection_enabled && zio->io_error == 0)
3536			zio->io_error = zio_handle_device_injection(vd,
3537			    zio, EIO);
3538
3539		if (zio_injection_enabled && zio->io_error == 0)
3540			zio->io_error = zio_handle_label_injection(zio, EIO);
3541
3542		if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) {
3543			if (!vdev_accessible(vd, zio)) {
3544				zio->io_error = SET_ERROR(ENXIO);
3545			} else {
3546				unexpected_error = B_TRUE;
3547			}
3548		}
3549	}
3550
3551	ops->vdev_op_io_done(zio);
3552
3553	if (unexpected_error)
3554		VERIFY(vdev_probe(vd, zio) == NULL);
3555
3556	return (ZIO_PIPELINE_CONTINUE);
3557}
3558
3559/*
3560 * This function is used to change the priority of an existing zio that is
3561 * currently in-flight. This is used by the arc to upgrade priority in the
3562 * event that a demand read is made for a block that is currently queued
3563 * as a scrub or async read IO. Otherwise, the high priority read request
3564 * would end up having to wait for the lower priority IO.
3565 */
3566void
3567zio_change_priority(zio_t *pio, zio_priority_t priority)
3568{
3569	zio_t *cio, *cio_next;
3570	zio_link_t *zl = NULL;
3571
3572	ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3573
3574	if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3575		vdev_queue_change_io_priority(pio, priority);
3576	} else {
3577		pio->io_priority = priority;
3578	}
3579
3580	mutex_enter(&pio->io_lock);
3581	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3582		cio_next = zio_walk_children(pio, &zl);
3583		zio_change_priority(cio, priority);
3584	}
3585	mutex_exit(&pio->io_lock);
3586}
3587
3588/*
3589 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3590 * disk, and use that to finish the checksum ereport later.
3591 */
3592static void
3593zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3594    const abd_t *good_buf)
3595{
3596	/* no processing needed */
3597	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3598}
3599
3600/*ARGSUSED*/
3601void
3602zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3603{
3604	void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size);
3605
3606	abd_copy(abd, zio->io_abd, zio->io_size);
3607
3608	zcr->zcr_cbinfo = zio->io_size;
3609	zcr->zcr_cbdata = abd;
3610	zcr->zcr_finish = zio_vsd_default_cksum_finish;
3611	zcr->zcr_free = zio_abd_free;
3612}
3613
3614static int
3615zio_vdev_io_assess(zio_t *zio)
3616{
3617	vdev_t *vd = zio->io_vd;
3618
3619	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3620		return (ZIO_PIPELINE_STOP);
3621	}
3622
3623	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3624		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3625
3626	if (zio->io_vsd != NULL) {
3627		zio->io_vsd_ops->vsd_free(zio);
3628		zio->io_vsd = NULL;
3629	}
3630
3631	if (zio_injection_enabled && zio->io_error == 0)
3632		zio->io_error = zio_handle_fault_injection(zio, EIO);
3633
3634	/*
3635	 * If the I/O failed, determine whether we should attempt to retry it.
3636	 *
3637	 * On retry, we cut in line in the issue queue, since we don't want
3638	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3639	 */
3640	if (zio->io_error && vd == NULL &&
3641	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3642		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
3643		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
3644		zio->io_error = 0;
3645		zio->io_flags |= ZIO_FLAG_IO_RETRY |
3646		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3647		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3648		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3649		    zio_requeue_io_start_cut_in_line);
3650		return (ZIO_PIPELINE_STOP);
3651	}
3652
3653	/*
3654	 * If we got an error on a leaf device, convert it to ENXIO
3655	 * if the device is not accessible at all.
3656	 */
3657	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3658	    !vdev_accessible(vd, zio))
3659		zio->io_error = SET_ERROR(ENXIO);
3660
3661	/*
3662	 * If we can't write to an interior vdev (mirror or RAID-Z),
3663	 * set vdev_cant_write so that we stop trying to allocate from it.
3664	 */
3665	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3666	    vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3667		vd->vdev_cant_write = B_TRUE;
3668	}
3669
3670	/*
3671	 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3672	 * attempts will ever succeed. In this case we set a persistent
3673	 * boolean flag so that we don't bother with it in the future.
3674	 */
3675	if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3676	    zio->io_type == ZIO_TYPE_IOCTL &&
3677	    zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3678		vd->vdev_nowritecache = B_TRUE;
3679
3680	if (zio->io_error)
3681		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3682
3683	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3684	    zio->io_physdone != NULL) {
3685		ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3686		ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3687		zio->io_physdone(zio->io_logical);
3688	}
3689
3690	return (ZIO_PIPELINE_CONTINUE);
3691}
3692
3693void
3694zio_vdev_io_reissue(zio_t *zio)
3695{
3696	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3697	ASSERT(zio->io_error == 0);
3698
3699	zio->io_stage >>= 1;
3700}
3701
3702void
3703zio_vdev_io_redone(zio_t *zio)
3704{
3705	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3706
3707	zio->io_stage >>= 1;
3708}
3709
3710void
3711zio_vdev_io_bypass(zio_t *zio)
3712{
3713	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3714	ASSERT(zio->io_error == 0);
3715
3716	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3717	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3718}
3719
3720/*
3721 * ==========================================================================
3722 * Encrypt and store encryption parameters
3723 * ==========================================================================
3724 */
3725
3726
3727/*
3728 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
3729 * managing the storage of encryption parameters and passing them to the
3730 * lower-level encryption functions.
3731 */
3732static int
3733zio_encrypt(zio_t *zio)
3734{
3735	zio_prop_t *zp = &zio->io_prop;
3736	spa_t *spa = zio->io_spa;
3737	blkptr_t *bp = zio->io_bp;
3738	uint64_t psize = BP_GET_PSIZE(bp);
3739	uint64_t dsobj = zio->io_bookmark.zb_objset;
3740	dmu_object_type_t ot = BP_GET_TYPE(bp);
3741	void *enc_buf = NULL;
3742	abd_t *eabd = NULL;
3743	uint8_t salt[ZIO_DATA_SALT_LEN];
3744	uint8_t iv[ZIO_DATA_IV_LEN];
3745	uint8_t mac[ZIO_DATA_MAC_LEN];
3746	boolean_t no_crypt = B_FALSE;
3747
3748	/* the root zio already encrypted the data */
3749	if (zio->io_child_type == ZIO_CHILD_GANG)
3750		return (ZIO_PIPELINE_CONTINUE);
3751
3752	/* only ZIL blocks are re-encrypted on rewrite */
3753	if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG)
3754		return (ZIO_PIPELINE_CONTINUE);
3755
3756	if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) {
3757		BP_SET_CRYPT(bp, B_FALSE);
3758		return (ZIO_PIPELINE_CONTINUE);
3759	}
3760
3761	/* if we are doing raw encryption set the provided encryption params */
3762	if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) {
3763		ASSERT0(BP_GET_LEVEL(bp));
3764		BP_SET_CRYPT(bp, B_TRUE);
3765		BP_SET_BYTEORDER(bp, zp->zp_byteorder);
3766		if (ot != DMU_OT_OBJSET)
3767			zio_crypt_encode_mac_bp(bp, zp->zp_mac);
3768
3769		/* dnode blocks must be written out in the provided byteorder */
3770		if (zp->zp_byteorder != ZFS_HOST_BYTEORDER &&
3771		    ot == DMU_OT_DNODE) {
3772			void *bswap_buf = zio_buf_alloc(psize);
3773			abd_t *babd = abd_get_from_buf(bswap_buf, psize);
3774
3775			ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
3776			abd_copy_to_buf(bswap_buf, zio->io_abd, psize);
3777			dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf,
3778			    psize);
3779
3780			abd_take_ownership_of_buf(babd, B_TRUE);
3781			zio_push_transform(zio, babd, psize, psize, NULL);
3782		}
3783
3784		if (DMU_OT_IS_ENCRYPTED(ot))
3785			zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv);
3786		return (ZIO_PIPELINE_CONTINUE);
3787	}
3788
3789	/* indirect blocks only maintain a cksum of the lower level MACs */
3790	if (BP_GET_LEVEL(bp) > 0) {
3791		BP_SET_CRYPT(bp, B_TRUE);
3792		VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE,
3793		    zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp),
3794		    mac));
3795		zio_crypt_encode_mac_bp(bp, mac);
3796		return (ZIO_PIPELINE_CONTINUE);
3797	}
3798
3799	/*
3800	 * Objset blocks are a special case since they have 2 256-bit MACs
3801	 * embedded within them.
3802	 */
3803	if (ot == DMU_OT_OBJSET) {
3804		ASSERT0(DMU_OT_IS_ENCRYPTED(ot));
3805		ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
3806		BP_SET_CRYPT(bp, B_TRUE);
3807		VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj,
3808		    zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp)));
3809		return (ZIO_PIPELINE_CONTINUE);
3810	}
3811
3812	/* unencrypted object types are only authenticated with a MAC */
3813	if (!DMU_OT_IS_ENCRYPTED(ot)) {
3814		BP_SET_CRYPT(bp, B_TRUE);
3815		VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj,
3816		    zio->io_abd, psize, mac));
3817		zio_crypt_encode_mac_bp(bp, mac);
3818		return (ZIO_PIPELINE_CONTINUE);
3819	}
3820
3821	/*
3822	 * Later passes of sync-to-convergence may decide to rewrite data
3823	 * in place to avoid more disk reallocations. This presents a problem
3824	 * for encryption because this consitutes rewriting the new data with
3825	 * the same encryption key and IV. However, this only applies to blocks
3826	 * in the MOS (particularly the spacemaps) and we do not encrypt the
3827	 * MOS. We assert that the zio is allocating or an intent log write
3828	 * to enforce this.
3829	 */
3830	ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG);
3831	ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG);
3832	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION));
3833	ASSERT3U(psize, !=, 0);
3834
3835	enc_buf = zio_buf_alloc(psize);
3836	eabd = abd_get_from_buf(enc_buf, psize);
3837	abd_take_ownership_of_buf(eabd, B_TRUE);
3838
3839	/*
3840	 * For an explanation of what encryption parameters are stored
3841	 * where, see the block comment in zio_crypt.c.
3842	 */
3843	if (ot == DMU_OT_INTENT_LOG) {
3844		zio_crypt_decode_params_bp(bp, salt, iv);
3845	} else {
3846		BP_SET_CRYPT(bp, B_TRUE);
3847	}
3848
3849	/* Perform the encryption. This should not fail */
3850	VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark,
3851	    BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
3852	    salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt));
3853
3854	/* encode encryption metadata into the bp */
3855	if (ot == DMU_OT_INTENT_LOG) {
3856		/*
3857		 * ZIL blocks store the MAC in the embedded checksum, so the
3858		 * transform must always be applied.
3859		 */
3860		zio_crypt_encode_mac_zil(enc_buf, mac);
3861		zio_push_transform(zio, eabd, psize, psize, NULL);
3862	} else {
3863		BP_SET_CRYPT(bp, B_TRUE);
3864		zio_crypt_encode_params_bp(bp, salt, iv);
3865		zio_crypt_encode_mac_bp(bp, mac);
3866
3867		if (no_crypt) {
3868			ASSERT3U(ot, ==, DMU_OT_DNODE);
3869			abd_free(eabd);
3870		} else {
3871			zio_push_transform(zio, eabd, psize, psize, NULL);
3872		}
3873	}
3874
3875	return (ZIO_PIPELINE_CONTINUE);
3876}
3877
3878/*
3879 * ==========================================================================
3880 * Generate and verify checksums
3881 * ==========================================================================
3882 */
3883static int
3884zio_checksum_generate(zio_t *zio)
3885{
3886	blkptr_t *bp = zio->io_bp;
3887	enum zio_checksum checksum;
3888
3889	if (bp == NULL) {
3890		/*
3891		 * This is zio_write_phys().
3892		 * We're either generating a label checksum, or none at all.
3893		 */
3894		checksum = zio->io_prop.zp_checksum;
3895
3896		if (checksum == ZIO_CHECKSUM_OFF)
3897			return (ZIO_PIPELINE_CONTINUE);
3898
3899		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3900	} else {
3901		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3902			ASSERT(!IO_IS_ALLOCATING(zio));
3903			checksum = ZIO_CHECKSUM_GANG_HEADER;
3904		} else {
3905			checksum = BP_GET_CHECKSUM(bp);
3906		}
3907	}
3908
3909	zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
3910
3911	return (ZIO_PIPELINE_CONTINUE);
3912}
3913
3914static int
3915zio_checksum_verify(zio_t *zio)
3916{
3917	zio_bad_cksum_t info;
3918	blkptr_t *bp = zio->io_bp;
3919	int error;
3920
3921	ASSERT(zio->io_vd != NULL);
3922
3923	if (bp == NULL) {
3924		/*
3925		 * This is zio_read_phys().
3926		 * We're either verifying a label checksum, or nothing at all.
3927		 */
3928		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3929			return (ZIO_PIPELINE_CONTINUE);
3930
3931		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3932	}
3933
3934	if ((error = zio_checksum_error(zio, &info)) != 0) {
3935		zio->io_error = error;
3936		if (error == ECKSUM &&
3937		    !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3938			zfs_ereport_start_checksum(zio->io_spa,
3939			    zio->io_vd, &zio->io_bookmark, zio,
3940			    zio->io_offset, zio->io_size, NULL, &info);
3941		}
3942	}
3943
3944	return (ZIO_PIPELINE_CONTINUE);
3945}
3946
3947/*
3948 * Called by RAID-Z to ensure we don't compute the checksum twice.
3949 */
3950void
3951zio_checksum_verified(zio_t *zio)
3952{
3953	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3954}
3955
3956/*
3957 * ==========================================================================
3958 * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3959 * An error of 0 indicates success.  ENXIO indicates whole-device failure,
3960 * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
3961 * indicate errors that are specific to one I/O, and most likely permanent.
3962 * Any other error is presumed to be worse because we weren't expecting it.
3963 * ==========================================================================
3964 */
3965int
3966zio_worst_error(int e1, int e2)
3967{
3968	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3969	int r1, r2;
3970
3971	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3972		if (e1 == zio_error_rank[r1])
3973			break;
3974
3975	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3976		if (e2 == zio_error_rank[r2])
3977			break;
3978
3979	return (r1 > r2 ? e1 : e2);
3980}
3981
3982/*
3983 * ==========================================================================
3984 * I/O completion
3985 * ==========================================================================
3986 */
3987static int
3988zio_ready(zio_t *zio)
3989{
3990	blkptr_t *bp = zio->io_bp;
3991	zio_t *pio, *pio_next;
3992	zio_link_t *zl = NULL;
3993
3994	if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
3995	    ZIO_WAIT_READY)) {
3996		return (ZIO_PIPELINE_STOP);
3997	}
3998
3999	if (zio->io_ready) {
4000		ASSERT(IO_IS_ALLOCATING(zio));
4001		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
4002		    (zio->io_flags & ZIO_FLAG_NOPWRITE));
4003		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
4004
4005		zio->io_ready(zio);
4006	}
4007
4008	if (bp != NULL && bp != &zio->io_bp_copy)
4009		zio->io_bp_copy = *bp;
4010
4011	if (zio->io_error != 0) {
4012		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4013
4014		if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4015			ASSERT(IO_IS_ALLOCATING(zio));
4016			ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4017			ASSERT(zio->io_metaslab_class != NULL);
4018
4019			/*
4020			 * We were unable to allocate anything, unreserve and
4021			 * issue the next I/O to allocate.
4022			 */
4023			metaslab_class_throttle_unreserve(
4024			    zio->io_metaslab_class, zio->io_prop.zp_copies,
4025			    zio->io_allocator, zio);
4026			zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
4027		}
4028	}
4029
4030	mutex_enter(&zio->io_lock);
4031	zio->io_state[ZIO_WAIT_READY] = 1;
4032	pio = zio_walk_parents(zio, &zl);
4033	mutex_exit(&zio->io_lock);
4034
4035	/*
4036	 * As we notify zio's parents, new parents could be added.
4037	 * New parents go to the head of zio's io_parent_list, however,
4038	 * so we will (correctly) not notify them.  The remainder of zio's
4039	 * io_parent_list, from 'pio_next' onward, cannot change because
4040	 * all parents must wait for us to be done before they can be done.
4041	 */
4042	for (; pio != NULL; pio = pio_next) {
4043		pio_next = zio_walk_parents(zio, &zl);
4044		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
4045	}
4046
4047	if (zio->io_flags & ZIO_FLAG_NODATA) {
4048		if (BP_IS_GANG(bp)) {
4049			zio->io_flags &= ~ZIO_FLAG_NODATA;
4050		} else {
4051			ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
4052			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
4053		}
4054	}
4055
4056	if (zio_injection_enabled &&
4057	    zio->io_spa->spa_syncing_txg == zio->io_txg)
4058		zio_handle_ignored_writes(zio);
4059
4060	return (ZIO_PIPELINE_CONTINUE);
4061}
4062
4063/*
4064 * Update the allocation throttle accounting.
4065 */
4066static void
4067zio_dva_throttle_done(zio_t *zio)
4068{
4069	zio_t *lio = zio->io_logical;
4070	zio_t *pio = zio_unique_parent(zio);
4071	vdev_t *vd = zio->io_vd;
4072	int flags = METASLAB_ASYNC_ALLOC;
4073
4074	ASSERT3P(zio->io_bp, !=, NULL);
4075	ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
4076	ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
4077	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
4078	ASSERT(vd != NULL);
4079	ASSERT3P(vd, ==, vd->vdev_top);
4080	ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
4081	ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
4082	ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
4083	ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
4084
4085	/*
4086	 * Parents of gang children can have two flavors -- ones that
4087	 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4088	 * and ones that allocated the constituent blocks. The allocation
4089	 * throttle needs to know the allocating parent zio so we must find
4090	 * it here.
4091	 */
4092	if (pio->io_child_type == ZIO_CHILD_GANG) {
4093		/*
4094		 * If our parent is a rewrite gang child then our grandparent
4095		 * would have been the one that performed the allocation.
4096		 */
4097		if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
4098			pio = zio_unique_parent(pio);
4099		flags |= METASLAB_GANG_CHILD;
4100	}
4101
4102	ASSERT(IO_IS_ALLOCATING(pio));
4103	ASSERT3P(zio, !=, zio->io_logical);
4104	ASSERT(zio->io_logical != NULL);
4105	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4106	ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
4107	ASSERT(zio->io_metaslab_class != NULL);
4108
4109	mutex_enter(&pio->io_lock);
4110	metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
4111	    pio->io_allocator, B_TRUE);
4112	mutex_exit(&pio->io_lock);
4113
4114	metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1,
4115	    pio->io_allocator, pio);
4116
4117	/*
4118	 * Call into the pipeline to see if there is more work that
4119	 * needs to be done. If there is work to be done it will be
4120	 * dispatched to another taskq thread.
4121	 */
4122	zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
4123}
4124
4125static int
4126zio_done(zio_t *zio)
4127{
4128	spa_t *spa = zio->io_spa;
4129	zio_t *lio = zio->io_logical;
4130	blkptr_t *bp = zio->io_bp;
4131	vdev_t *vd = zio->io_vd;
4132	uint64_t psize = zio->io_size;
4133	zio_t *pio, *pio_next;
4134	zio_link_t *zl = NULL;
4135
4136	/*
4137	 * If our children haven't all completed,
4138	 * wait for them and then repeat this pipeline stage.
4139	 */
4140	if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
4141		return (ZIO_PIPELINE_STOP);
4142	}
4143
4144	/*
4145	 * If the allocation throttle is enabled, then update the accounting.
4146	 * We only track child I/Os that are part of an allocating async
4147	 * write. We must do this since the allocation is performed
4148	 * by the logical I/O but the actual write is done by child I/Os.
4149	 */
4150	if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
4151	    zio->io_child_type == ZIO_CHILD_VDEV) {
4152		ASSERT(zio->io_metaslab_class != NULL);
4153		ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled);
4154		zio_dva_throttle_done(zio);
4155	}
4156
4157	/*
4158	 * If the allocation throttle is enabled, verify that
4159	 * we have decremented the refcounts for every I/O that was throttled.
4160	 */
4161	if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4162		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4163		ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4164		ASSERT(bp != NULL);
4165
4166		metaslab_group_alloc_verify(spa, zio->io_bp, zio,
4167		    zio->io_allocator);
4168		VERIFY(zfs_refcount_not_held(
4169		    &zio->io_metaslab_class->mc_alloc_slots[zio->io_allocator],
4170		    zio));
4171	}
4172
4173	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
4174		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
4175			ASSERT(zio->io_children[c][w] == 0);
4176
4177	if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
4178		ASSERT(bp->blk_pad[0] == 0);
4179		ASSERT(bp->blk_pad[1] == 0);
4180		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
4181		    (bp == zio_unique_parent(zio)->io_bp));
4182		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
4183		    zio->io_bp_override == NULL &&
4184		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
4185			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
4186			ASSERT(BP_COUNT_GANG(bp) == 0 ||
4187			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
4188		}
4189		if (zio->io_flags & ZIO_FLAG_NOPWRITE)
4190			VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
4191	}
4192
4193	/*
4194	 * If there were child vdev/gang/ddt errors, they apply to us now.
4195	 */
4196	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
4197	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
4198	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
4199
4200	/*
4201	 * If the I/O on the transformed data was successful, generate any
4202	 * checksum reports now while we still have the transformed data.
4203	 */
4204	if (zio->io_error == 0) {
4205		while (zio->io_cksum_report != NULL) {
4206			zio_cksum_report_t *zcr = zio->io_cksum_report;
4207			uint64_t align = zcr->zcr_align;
4208			uint64_t asize = P2ROUNDUP(psize, align);
4209			abd_t *adata = zio->io_abd;
4210
4211			if (asize != psize) {
4212				adata = abd_alloc(asize, B_TRUE);
4213				abd_copy(adata, zio->io_abd, psize);
4214				abd_zero_off(adata, psize, asize - psize);
4215			}
4216
4217			zio->io_cksum_report = zcr->zcr_next;
4218			zcr->zcr_next = NULL;
4219			zcr->zcr_finish(zcr, adata);
4220			zfs_ereport_free_checksum(zcr);
4221
4222			if (asize != psize)
4223				abd_free(adata);
4224		}
4225	}
4226
4227	zio_pop_transforms(zio);	/* note: may set zio->io_error */
4228
4229	vdev_stat_update(zio, psize);
4230
4231	if (zio->io_error) {
4232		/*
4233		 * If this I/O is attached to a particular vdev,
4234		 * generate an error message describing the I/O failure
4235		 * at the block level.  We ignore these errors if the
4236		 * device is currently unavailable.
4237		 */
4238		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
4239			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd,
4240			    &zio->io_bookmark, zio, 0, 0);
4241
4242		if ((zio->io_error == EIO || !(zio->io_flags &
4243		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
4244		    zio == lio) {
4245			/*
4246			 * For logical I/O requests, tell the SPA to log the
4247			 * error and generate a logical data ereport.
4248			 */
4249			spa_log_error(spa, &zio->io_bookmark);
4250			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL,
4251			    &zio->io_bookmark, zio, 0, 0);
4252		}
4253	}
4254
4255	if (zio->io_error && zio == lio) {
4256		/*
4257		 * Determine whether zio should be reexecuted.  This will
4258		 * propagate all the way to the root via zio_notify_parent().
4259		 */
4260		ASSERT(vd == NULL && bp != NULL);
4261		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4262
4263		if (IO_IS_ALLOCATING(zio) &&
4264		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
4265			if (zio->io_error != ENOSPC)
4266				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
4267			else
4268				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4269		}
4270
4271		if ((zio->io_type == ZIO_TYPE_READ ||
4272		    zio->io_type == ZIO_TYPE_FREE) &&
4273		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
4274		    zio->io_error == ENXIO &&
4275		    spa_load_state(spa) == SPA_LOAD_NONE &&
4276		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
4277			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4278
4279		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
4280			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4281
4282		/*
4283		 * Here is a possibly good place to attempt to do
4284		 * either combinatorial reconstruction or error correction
4285		 * based on checksums.  It also might be a good place
4286		 * to send out preliminary ereports before we suspend
4287		 * processing.
4288		 */
4289	}
4290
4291	/*
4292	 * If there were logical child errors, they apply to us now.
4293	 * We defer this until now to avoid conflating logical child
4294	 * errors with errors that happened to the zio itself when
4295	 * updating vdev stats and reporting FMA events above.
4296	 */
4297	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
4298
4299	if ((zio->io_error || zio->io_reexecute) &&
4300	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
4301	    !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
4302		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
4303
4304	zio_gang_tree_free(&zio->io_gang_tree);
4305
4306	/*
4307	 * Godfather I/Os should never suspend.
4308	 */
4309	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4310	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4311		zio->io_reexecute = 0;
4312
4313	if (zio->io_reexecute) {
4314		/*
4315		 * This is a logical I/O that wants to reexecute.
4316		 *
4317		 * Reexecute is top-down.  When an i/o fails, if it's not
4318		 * the root, it simply notifies its parent and sticks around.
4319		 * The parent, seeing that it still has children in zio_done(),
4320		 * does the same.  This percolates all the way up to the root.
4321		 * The root i/o will reexecute or suspend the entire tree.
4322		 *
4323		 * This approach ensures that zio_reexecute() honors
4324		 * all the original i/o dependency relationships, e.g.
4325		 * parents not executing until children are ready.
4326		 */
4327		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4328
4329		zio->io_gang_leader = NULL;
4330
4331		mutex_enter(&zio->io_lock);
4332		zio->io_state[ZIO_WAIT_DONE] = 1;
4333		mutex_exit(&zio->io_lock);
4334
4335		/*
4336		 * "The Godfather" I/O monitors its children but is
4337		 * not a true parent to them. It will track them through
4338		 * the pipeline but severs its ties whenever they get into
4339		 * trouble (e.g. suspended). This allows "The Godfather"
4340		 * I/O to return status without blocking.
4341		 */
4342		zl = NULL;
4343		for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4344		    pio = pio_next) {
4345			zio_link_t *remove_zl = zl;
4346			pio_next = zio_walk_parents(zio, &zl);
4347
4348			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4349			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4350				zio_remove_child(pio, zio, remove_zl);
4351				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4352			}
4353		}
4354
4355		if ((pio = zio_unique_parent(zio)) != NULL) {
4356			/*
4357			 * We're not a root i/o, so there's nothing to do
4358			 * but notify our parent.  Don't propagate errors
4359			 * upward since we haven't permanently failed yet.
4360			 */
4361			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4362			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4363			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4364		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4365			/*
4366			 * We'd fail again if we reexecuted now, so suspend
4367			 * until conditions improve (e.g. device comes online).
4368			 */
4369			zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
4370		} else {
4371			/*
4372			 * Reexecution is potentially a huge amount of work.
4373			 * Hand it off to the otherwise-unused claim taskq.
4374			 */
4375			ASSERT(zio->io_tqent.tqent_next == NULL);
4376			spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
4377			    ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
4378			    0, &zio->io_tqent);
4379		}
4380		return (ZIO_PIPELINE_STOP);
4381	}
4382
4383	ASSERT(zio->io_child_count == 0);
4384	ASSERT(zio->io_reexecute == 0);
4385	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4386
4387	/*
4388	 * Report any checksum errors, since the I/O is complete.
4389	 */
4390	while (zio->io_cksum_report != NULL) {
4391		zio_cksum_report_t *zcr = zio->io_cksum_report;
4392		zio->io_cksum_report = zcr->zcr_next;
4393		zcr->zcr_next = NULL;
4394		zcr->zcr_finish(zcr, NULL);
4395		zfs_ereport_free_checksum(zcr);
4396	}
4397
4398	/*
4399	 * It is the responsibility of the done callback to ensure that this
4400	 * particular zio is no longer discoverable for adoption, and as
4401	 * such, cannot acquire any new parents.
4402	 */
4403	if (zio->io_done)
4404		zio->io_done(zio);
4405
4406	mutex_enter(&zio->io_lock);
4407	zio->io_state[ZIO_WAIT_DONE] = 1;
4408	mutex_exit(&zio->io_lock);
4409
4410	zl = NULL;
4411	for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4412		zio_link_t *remove_zl = zl;
4413		pio_next = zio_walk_parents(zio, &zl);
4414		zio_remove_child(pio, zio, remove_zl);
4415		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4416	}
4417
4418	if (zio->io_waiter != NULL) {
4419		mutex_enter(&zio->io_lock);
4420		zio->io_executor = NULL;
4421		cv_broadcast(&zio->io_cv);
4422		mutex_exit(&zio->io_lock);
4423	} else {
4424		zio_destroy(zio);
4425	}
4426
4427	return (ZIO_PIPELINE_STOP);
4428}
4429
4430/*
4431 * ==========================================================================
4432 * I/O pipeline definition
4433 * ==========================================================================
4434 */
4435static zio_pipe_stage_t *zio_pipeline[] = {
4436	NULL,
4437	zio_read_bp_init,
4438	zio_write_bp_init,
4439	zio_free_bp_init,
4440	zio_issue_async,
4441	zio_write_compress,
4442	zio_encrypt,
4443	zio_checksum_generate,
4444	zio_nop_write,
4445	zio_ddt_read_start,
4446	zio_ddt_read_done,
4447	zio_ddt_write,
4448	zio_ddt_free,
4449	zio_gang_assemble,
4450	zio_gang_issue,
4451	zio_dva_throttle,
4452	zio_dva_allocate,
4453	zio_dva_free,
4454	zio_dva_claim,
4455	zio_ready,
4456	zio_vdev_io_start,
4457	zio_vdev_io_done,
4458	zio_vdev_io_assess,
4459	zio_checksum_verify,
4460	zio_done
4461};
4462
4463
4464
4465
4466/*
4467 * Compare two zbookmark_phys_t's to see which we would reach first in a
4468 * pre-order traversal of the object tree.
4469 *
4470 * This is simple in every case aside from the meta-dnode object. For all other
4471 * objects, we traverse them in order (object 1 before object 2, and so on).
4472 * However, all of these objects are traversed while traversing object 0, since
4473 * the data it points to is the list of objects.  Thus, we need to convert to a
4474 * canonical representation so we can compare meta-dnode bookmarks to
4475 * non-meta-dnode bookmarks.
4476 *
4477 * We do this by calculating "equivalents" for each field of the zbookmark.
4478 * zbookmarks outside of the meta-dnode use their own object and level, and
4479 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4480 * blocks this bookmark refers to) by multiplying their blkid by their span
4481 * (the number of L0 blocks contained within one block at their level).
4482 * zbookmarks inside the meta-dnode calculate their object equivalent
4483 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4484 * level + 1<<31 (any value larger than a level could ever be) for their level.
4485 * This causes them to always compare before a bookmark in their object
4486 * equivalent, compare appropriately to bookmarks in other objects, and to
4487 * compare appropriately to other bookmarks in the meta-dnode.
4488 */
4489int
4490zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4491    const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4492{
4493	/*
4494	 * These variables represent the "equivalent" values for the zbookmark,
4495	 * after converting zbookmarks inside the meta dnode to their
4496	 * normal-object equivalents.
4497	 */
4498	uint64_t zb1obj, zb2obj;
4499	uint64_t zb1L0, zb2L0;
4500	uint64_t zb1level, zb2level;
4501
4502	if (zb1->zb_object == zb2->zb_object &&
4503	    zb1->zb_level == zb2->zb_level &&
4504	    zb1->zb_blkid == zb2->zb_blkid)
4505		return (0);
4506
4507	/*
4508	 * BP_SPANB calculates the span in blocks.
4509	 */
4510	zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4511	zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4512
4513	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4514		zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4515		zb1L0 = 0;
4516		zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4517	} else {
4518		zb1obj = zb1->zb_object;
4519		zb1level = zb1->zb_level;
4520	}
4521
4522	if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4523		zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4524		zb2L0 = 0;
4525		zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4526	} else {
4527		zb2obj = zb2->zb_object;
4528		zb2level = zb2->zb_level;
4529	}
4530
4531	/* Now that we have a canonical representation, do the comparison. */
4532	if (zb1obj != zb2obj)
4533		return (zb1obj < zb2obj ? -1 : 1);
4534	else if (zb1L0 != zb2L0)
4535		return (zb1L0 < zb2L0 ? -1 : 1);
4536	else if (zb1level != zb2level)
4537		return (zb1level > zb2level ? -1 : 1);
4538	/*
4539	 * This can (theoretically) happen if the bookmarks have the same object
4540	 * and level, but different blkids, if the block sizes are not the same.
4541	 * There is presently no way to change the indirect block sizes
4542	 */
4543	return (0);
4544}
4545
4546/*
4547 *  This function checks the following: given that last_block is the place that
4548 *  our traversal stopped last time, does that guarantee that we've visited
4549 *  every node under subtree_root?  Therefore, we can't just use the raw output
4550 *  of zbookmark_compare.  We have to pass in a modified version of
4551 *  subtree_root; by incrementing the block id, and then checking whether
4552 *  last_block is before or equal to that, we can tell whether or not having
4553 *  visited last_block implies that all of subtree_root's children have been
4554 *  visited.
4555 */
4556boolean_t
4557zbookmark_subtree_completed(const dnode_phys_t *dnp,
4558    const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4559{
4560	zbookmark_phys_t mod_zb = *subtree_root;
4561	mod_zb.zb_blkid++;
4562	ASSERT(last_block->zb_level == 0);
4563
4564	/* The objset_phys_t isn't before anything. */
4565	if (dnp == NULL)
4566		return (B_FALSE);
4567
4568	/*
4569	 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4570	 * data block size in sectors, because that variable is only used if
4571	 * the bookmark refers to a block in the meta-dnode.  Since we don't
4572	 * know without examining it what object it refers to, and there's no
4573	 * harm in passing in this value in other cases, we always pass it in.
4574	 *
4575	 * We pass in 0 for the indirect block size shift because zb2 must be
4576	 * level 0.  The indirect block size is only used to calculate the span
4577	 * of the bookmark, but since the bookmark must be level 0, the span is
4578	 * always 1, so the math works out.
4579	 *
4580	 * If you make changes to how the zbookmark_compare code works, be sure
4581	 * to make sure that this code still works afterwards.
4582	 */
4583	return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4584	    1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
4585	    last_block) <= 0);
4586}
4587