xref: /freebsd/sys/contrib/openzfs/module/zfs/zio.c (revision 0d48d1ffe0446cd2f87ce02555e3d17772ae7284)
1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3  * CDDL HEADER START
4  *
5  * The contents of this file are subject to the terms of the
6  * Common Development and Distribution License (the "License").
7  * You may not use this file except in compliance with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or https://opensource.org/licenses/CDDL-1.0.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2011, 2022 by Delphix. All rights reserved.
25  * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
26  * Copyright (c) 2017, Intel Corporation.
27  * Copyright (c) 2019, 2023, 2024, 2025, Klara, Inc.
28  * Copyright (c) 2019, Allan Jude
29  * Copyright (c) 2021, Datto, Inc.
30  * Copyright (c) 2021, 2024 by George Melikov. All rights reserved.
31  */
32 
33 #include <sys/sysmacros.h>
34 #include <sys/zfs_context.h>
35 #include <sys/fm/fs/zfs.h>
36 #include <sys/spa.h>
37 #include <sys/txg.h>
38 #include <sys/spa_impl.h>
39 #include <sys/vdev_impl.h>
40 #include <sys/vdev_trim.h>
41 #include <sys/zio_impl.h>
42 #include <sys/zio_compress.h>
43 #include <sys/zio_checksum.h>
44 #include <sys/dmu_objset.h>
45 #include <sys/arc.h>
46 #include <sys/brt.h>
47 #include <sys/ddt.h>
48 #include <sys/blkptr.h>
49 #include <sys/zfeature.h>
50 #include <sys/dsl_scan.h>
51 #include <sys/metaslab_impl.h>
52 #include <sys/time.h>
53 #include <sys/trace_zfs.h>
54 #include <sys/abd.h>
55 #include <sys/dsl_crypt.h>
56 #include <cityhash.h>
57 
58 /*
59  * ==========================================================================
60  * I/O type descriptions
61  * ==========================================================================
62  */
63 const char *const zio_type_name[ZIO_TYPES] = {
64 	/*
65 	 * Note: Linux kernel thread name length is limited
66 	 * so these names will differ from upstream open zfs.
67 	 */
68 	"z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_flush", "z_trim"
69 };
70 
71 int zio_dva_throttle_enabled = B_TRUE;
72 static int zio_deadman_log_all = B_FALSE;
73 
74 /*
75  * ==========================================================================
76  * I/O kmem caches
77  * ==========================================================================
78  */
79 static kmem_cache_t *zio_cache;
80 static kmem_cache_t *zio_link_cache;
81 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
82 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
83 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
84 static uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
85 static uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
86 #endif
87 
88 /* Mark IOs as "slow" if they take longer than 30 seconds */
89 static uint_t zio_slow_io_ms = (30 * MILLISEC);
90 
91 #define	BP_SPANB(indblkshift, level) \
92 	(((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
93 #define	COMPARE_META_LEVEL	0x80000000ul
94 /*
95  * The following actions directly effect the spa's sync-to-convergence logic.
96  * The values below define the sync pass when we start performing the action.
97  * Care should be taken when changing these values as they directly impact
98  * spa_sync() performance. Tuning these values may introduce subtle performance
99  * pathologies and should only be done in the context of performance analysis.
100  * These tunables will eventually be removed and replaced with #defines once
101  * enough analysis has been done to determine optimal values.
102  *
103  * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
104  * regular blocks are not deferred.
105  *
106  * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable
107  * compression (including of metadata).  In practice, we don't have this
108  * many sync passes, so this has no effect.
109  *
110  * The original intent was that disabling compression would help the sync
111  * passes to converge. However, in practice disabling compression increases
112  * the average number of sync passes, because when we turn compression off, a
113  * lot of block's size will change and thus we have to re-allocate (not
114  * overwrite) them. It also increases the number of 128KB allocations (e.g.
115  * for indirect blocks and spacemaps) because these will not be compressed.
116  * The 128K allocations are especially detrimental to performance on highly
117  * fragmented systems, which may have very few free segments of this size,
118  * and may need to load new metaslabs to satisfy 128K allocations.
119  */
120 
121 /* defer frees starting in this pass */
122 uint_t zfs_sync_pass_deferred_free = 2;
123 
124 /* don't compress starting in this pass */
125 static uint_t zfs_sync_pass_dont_compress = 8;
126 
127 /* rewrite new bps starting in this pass */
128 static uint_t zfs_sync_pass_rewrite = 2;
129 
130 /*
131  * An allocating zio is one that either currently has the DVA allocate
132  * stage set or will have it later in its lifetime.
133  */
134 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
135 
136 /*
137  * Enable smaller cores by excluding metadata
138  * allocations as well.
139  */
140 int zio_exclude_metadata = 0;
141 static int zio_requeue_io_start_cut_in_line = 1;
142 
143 #ifdef ZFS_DEBUG
144 static const int zio_buf_debug_limit = 16384;
145 #else
146 static const int zio_buf_debug_limit = 0;
147 #endif
148 
149 typedef struct zio_stats {
150 	kstat_named_t ziostat_total_allocations;
151 	kstat_named_t ziostat_alloc_class_fallbacks;
152 	kstat_named_t ziostat_gang_writes;
153 	kstat_named_t ziostat_gang_multilevel;
154 } zio_stats_t;
155 
156 static zio_stats_t zio_stats = {
157 	{ "total_allocations",	KSTAT_DATA_UINT64 },
158 	{ "alloc_class_fallbacks",	KSTAT_DATA_UINT64 },
159 	{ "gang_writes",	KSTAT_DATA_UINT64 },
160 	{ "gang_multilevel",	KSTAT_DATA_UINT64 },
161 };
162 
163 struct {
164 	wmsum_t ziostat_total_allocations;
165 	wmsum_t ziostat_alloc_class_fallbacks;
166 	wmsum_t ziostat_gang_writes;
167 	wmsum_t ziostat_gang_multilevel;
168 } ziostat_sums;
169 
170 #define	ZIOSTAT_BUMP(stat)	wmsum_add(&ziostat_sums.stat, 1);
171 
172 static kstat_t *zio_ksp;
173 
174 static inline void __zio_execute(zio_t *zio);
175 
176 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
177 
178 static int
179 zio_kstats_update(kstat_t *ksp, int rw)
180 {
181 	zio_stats_t *zs = ksp->ks_data;
182 	if (rw == KSTAT_WRITE)
183 		return (EACCES);
184 
185 	zs->ziostat_total_allocations.value.ui64 =
186 	    wmsum_value(&ziostat_sums.ziostat_total_allocations);
187 	zs->ziostat_alloc_class_fallbacks.value.ui64 =
188 	    wmsum_value(&ziostat_sums.ziostat_alloc_class_fallbacks);
189 	zs->ziostat_gang_writes.value.ui64 =
190 	    wmsum_value(&ziostat_sums.ziostat_gang_writes);
191 	zs->ziostat_gang_multilevel.value.ui64 =
192 	    wmsum_value(&ziostat_sums.ziostat_gang_multilevel);
193 	return (0);
194 }
195 
196 void
197 zio_init(void)
198 {
199 	size_t c;
200 
201 	zio_cache = kmem_cache_create("zio_cache",
202 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
203 	zio_link_cache = kmem_cache_create("zio_link_cache",
204 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
205 
206 	wmsum_init(&ziostat_sums.ziostat_total_allocations, 0);
207 	wmsum_init(&ziostat_sums.ziostat_alloc_class_fallbacks, 0);
208 	wmsum_init(&ziostat_sums.ziostat_gang_writes, 0);
209 	wmsum_init(&ziostat_sums.ziostat_gang_multilevel, 0);
210 	zio_ksp = kstat_create("zfs", 0, "zio_stats",
211 	    "misc", KSTAT_TYPE_NAMED, sizeof (zio_stats) /
212 	    sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
213 	if (zio_ksp != NULL) {
214 		zio_ksp->ks_data = &zio_stats;
215 		zio_ksp->ks_update = zio_kstats_update;
216 		kstat_install(zio_ksp);
217 	}
218 
219 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
220 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
221 		size_t align, cflags, data_cflags;
222 		char name[32];
223 
224 		/*
225 		 * Create cache for each half-power of 2 size, starting from
226 		 * SPA_MINBLOCKSIZE.  It should give us memory space efficiency
227 		 * of ~7/8, sufficient for transient allocations mostly using
228 		 * these caches.
229 		 */
230 		size_t p2 = size;
231 		while (!ISP2(p2))
232 			p2 &= p2 - 1;
233 		if (!IS_P2ALIGNED(size, p2 / 2))
234 			continue;
235 
236 #ifndef _KERNEL
237 		/*
238 		 * If we are using watchpoints, put each buffer on its own page,
239 		 * to eliminate the performance overhead of trapping to the
240 		 * kernel when modifying a non-watched buffer that shares the
241 		 * page with a watched buffer.
242 		 */
243 		if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
244 			continue;
245 #endif
246 
247 		if (IS_P2ALIGNED(size, PAGESIZE))
248 			align = PAGESIZE;
249 		else
250 			align = 1 << (highbit64(size ^ (size - 1)) - 1);
251 
252 		cflags = (zio_exclude_metadata || size > zio_buf_debug_limit) ?
253 		    KMC_NODEBUG : 0;
254 		data_cflags = KMC_NODEBUG;
255 		if (abd_size_alloc_linear(size)) {
256 			cflags |= KMC_RECLAIMABLE;
257 			data_cflags |= KMC_RECLAIMABLE;
258 		}
259 		if (cflags == data_cflags) {
260 			/*
261 			 * Resulting kmem caches would be identical.
262 			 * Save memory by creating only one.
263 			 */
264 			(void) snprintf(name, sizeof (name),
265 			    "zio_buf_comb_%lu", (ulong_t)size);
266 			zio_buf_cache[c] = kmem_cache_create(name, size, align,
267 			    NULL, NULL, NULL, NULL, NULL, cflags);
268 			zio_data_buf_cache[c] = zio_buf_cache[c];
269 			continue;
270 		}
271 		(void) snprintf(name, sizeof (name), "zio_buf_%lu",
272 		    (ulong_t)size);
273 		zio_buf_cache[c] = kmem_cache_create(name, size, align,
274 		    NULL, NULL, NULL, NULL, NULL, cflags);
275 
276 		(void) snprintf(name, sizeof (name), "zio_data_buf_%lu",
277 		    (ulong_t)size);
278 		zio_data_buf_cache[c] = kmem_cache_create(name, size, align,
279 		    NULL, NULL, NULL, NULL, NULL, data_cflags);
280 	}
281 
282 	while (--c != 0) {
283 		ASSERT(zio_buf_cache[c] != NULL);
284 		if (zio_buf_cache[c - 1] == NULL)
285 			zio_buf_cache[c - 1] = zio_buf_cache[c];
286 
287 		ASSERT(zio_data_buf_cache[c] != NULL);
288 		if (zio_data_buf_cache[c - 1] == NULL)
289 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
290 	}
291 
292 	zio_inject_init();
293 
294 	lz4_init();
295 }
296 
297 void
298 zio_fini(void)
299 {
300 	size_t n = SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT;
301 
302 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
303 	for (size_t i = 0; i < n; i++) {
304 		if (zio_buf_cache_allocs[i] != zio_buf_cache_frees[i])
305 			(void) printf("zio_fini: [%d] %llu != %llu\n",
306 			    (int)((i + 1) << SPA_MINBLOCKSHIFT),
307 			    (long long unsigned)zio_buf_cache_allocs[i],
308 			    (long long unsigned)zio_buf_cache_frees[i]);
309 	}
310 #endif
311 
312 	/*
313 	 * The same kmem cache can show up multiple times in both zio_buf_cache
314 	 * and zio_data_buf_cache. Do a wasteful but trivially correct scan to
315 	 * sort it out.
316 	 */
317 	for (size_t i = 0; i < n; i++) {
318 		kmem_cache_t *cache = zio_buf_cache[i];
319 		if (cache == NULL)
320 			continue;
321 		for (size_t j = i; j < n; j++) {
322 			if (cache == zio_buf_cache[j])
323 				zio_buf_cache[j] = NULL;
324 			if (cache == zio_data_buf_cache[j])
325 				zio_data_buf_cache[j] = NULL;
326 		}
327 		kmem_cache_destroy(cache);
328 	}
329 
330 	for (size_t i = 0; i < n; i++) {
331 		kmem_cache_t *cache = zio_data_buf_cache[i];
332 		if (cache == NULL)
333 			continue;
334 		for (size_t j = i; j < n; j++) {
335 			if (cache == zio_data_buf_cache[j])
336 				zio_data_buf_cache[j] = NULL;
337 		}
338 		kmem_cache_destroy(cache);
339 	}
340 
341 	for (size_t i = 0; i < n; i++) {
342 		VERIFY3P(zio_buf_cache[i], ==, NULL);
343 		VERIFY3P(zio_data_buf_cache[i], ==, NULL);
344 	}
345 
346 	if (zio_ksp != NULL) {
347 		kstat_delete(zio_ksp);
348 		zio_ksp = NULL;
349 	}
350 
351 	wmsum_fini(&ziostat_sums.ziostat_total_allocations);
352 	wmsum_fini(&ziostat_sums.ziostat_alloc_class_fallbacks);
353 	wmsum_fini(&ziostat_sums.ziostat_gang_writes);
354 	wmsum_fini(&ziostat_sums.ziostat_gang_multilevel);
355 
356 	kmem_cache_destroy(zio_link_cache);
357 	kmem_cache_destroy(zio_cache);
358 
359 	zio_inject_fini();
360 
361 	lz4_fini();
362 }
363 
364 /*
365  * ==========================================================================
366  * Allocate and free I/O buffers
367  * ==========================================================================
368  */
369 
370 #if defined(ZFS_DEBUG) && defined(_KERNEL)
371 #define	ZFS_ZIO_BUF_CANARY	1
372 #endif
373 
374 #ifdef ZFS_ZIO_BUF_CANARY
375 static const ulong_t zio_buf_canary = (ulong_t)0xdeadc0dedead210b;
376 
377 /*
378  * Use empty space after the buffer to detect overflows.
379  *
380  * Since zio_init() creates kmem caches only for certain set of buffer sizes,
381  * allocations of different sizes may have some unused space after the data.
382  * Filling part of that space with a known pattern on allocation and checking
383  * it on free should allow us to detect some buffer overflows.
384  */
385 static void
386 zio_buf_put_canary(ulong_t *p, size_t size, kmem_cache_t **cache, size_t c)
387 {
388 	size_t off = P2ROUNDUP(size, sizeof (ulong_t));
389 	ulong_t *canary = p + off / sizeof (ulong_t);
390 	size_t asize = (c + 1) << SPA_MINBLOCKSHIFT;
391 	if (c + 1 < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT &&
392 	    cache[c] == cache[c + 1])
393 		asize = (c + 2) << SPA_MINBLOCKSHIFT;
394 	for (; off < asize; canary++, off += sizeof (ulong_t))
395 		*canary = zio_buf_canary;
396 }
397 
398 static void
399 zio_buf_check_canary(ulong_t *p, size_t size, kmem_cache_t **cache, size_t c)
400 {
401 	size_t off = P2ROUNDUP(size, sizeof (ulong_t));
402 	ulong_t *canary = p + off / sizeof (ulong_t);
403 	size_t asize = (c + 1) << SPA_MINBLOCKSHIFT;
404 	if (c + 1 < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT &&
405 	    cache[c] == cache[c + 1])
406 		asize = (c + 2) << SPA_MINBLOCKSHIFT;
407 	for (; off < asize; canary++, off += sizeof (ulong_t)) {
408 		if (unlikely(*canary != zio_buf_canary)) {
409 			PANIC("ZIO buffer overflow %p (%zu) + %zu %#lx != %#lx",
410 			    p, size, (canary - p) * sizeof (ulong_t),
411 			    *canary, zio_buf_canary);
412 		}
413 	}
414 }
415 #endif
416 
417 /*
418  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
419  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
420  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
421  * excess / transient data in-core during a crashdump.
422  */
423 void *
424 zio_buf_alloc(size_t size)
425 {
426 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
427 
428 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
429 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
430 	atomic_add_64(&zio_buf_cache_allocs[c], 1);
431 #endif
432 
433 	void *p = kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE);
434 #ifdef ZFS_ZIO_BUF_CANARY
435 	zio_buf_put_canary(p, size, zio_buf_cache, c);
436 #endif
437 	return (p);
438 }
439 
440 /*
441  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
442  * crashdump if the kernel panics.  This exists so that we will limit the amount
443  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
444  * of kernel heap dumped to disk when the kernel panics)
445  */
446 void *
447 zio_data_buf_alloc(size_t size)
448 {
449 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
450 
451 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
452 
453 	void *p = kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE);
454 #ifdef ZFS_ZIO_BUF_CANARY
455 	zio_buf_put_canary(p, size, zio_data_buf_cache, c);
456 #endif
457 	return (p);
458 }
459 
460 void
461 zio_buf_free(void *buf, size_t size)
462 {
463 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
464 
465 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
466 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
467 	atomic_add_64(&zio_buf_cache_frees[c], 1);
468 #endif
469 
470 #ifdef ZFS_ZIO_BUF_CANARY
471 	zio_buf_check_canary(buf, size, zio_buf_cache, c);
472 #endif
473 	kmem_cache_free(zio_buf_cache[c], buf);
474 }
475 
476 void
477 zio_data_buf_free(void *buf, size_t size)
478 {
479 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
480 
481 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
482 
483 #ifdef ZFS_ZIO_BUF_CANARY
484 	zio_buf_check_canary(buf, size, zio_data_buf_cache, c);
485 #endif
486 	kmem_cache_free(zio_data_buf_cache[c], buf);
487 }
488 
489 static void
490 zio_abd_free(void *abd, size_t size)
491 {
492 	(void) size;
493 	abd_free((abd_t *)abd);
494 }
495 
496 /*
497  * ==========================================================================
498  * Push and pop I/O transform buffers
499  * ==========================================================================
500  */
501 void
502 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
503     zio_transform_func_t *transform)
504 {
505 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
506 
507 	zt->zt_orig_abd = zio->io_abd;
508 	zt->zt_orig_size = zio->io_size;
509 	zt->zt_bufsize = bufsize;
510 	zt->zt_transform = transform;
511 
512 	zt->zt_next = zio->io_transform_stack;
513 	zio->io_transform_stack = zt;
514 
515 	zio->io_abd = data;
516 	zio->io_size = size;
517 }
518 
519 void
520 zio_pop_transforms(zio_t *zio)
521 {
522 	zio_transform_t *zt;
523 
524 	while ((zt = zio->io_transform_stack) != NULL) {
525 		if (zt->zt_transform != NULL)
526 			zt->zt_transform(zio,
527 			    zt->zt_orig_abd, zt->zt_orig_size);
528 
529 		if (zt->zt_bufsize != 0)
530 			abd_free(zio->io_abd);
531 
532 		zio->io_abd = zt->zt_orig_abd;
533 		zio->io_size = zt->zt_orig_size;
534 		zio->io_transform_stack = zt->zt_next;
535 
536 		kmem_free(zt, sizeof (zio_transform_t));
537 	}
538 }
539 
540 /*
541  * ==========================================================================
542  * I/O transform callbacks for subblocks, decompression, and decryption
543  * ==========================================================================
544  */
545 static void
546 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
547 {
548 	ASSERT(zio->io_size > size);
549 
550 	if (zio->io_type == ZIO_TYPE_READ)
551 		abd_copy(data, zio->io_abd, size);
552 }
553 
554 static void
555 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
556 {
557 	if (zio->io_error == 0) {
558 		int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
559 		    zio->io_abd, data, zio->io_size, size,
560 		    &zio->io_prop.zp_complevel);
561 
562 		if (zio_injection_enabled && ret == 0)
563 			ret = zio_handle_fault_injection(zio, EINVAL);
564 
565 		if (ret != 0)
566 			zio->io_error = SET_ERROR(EIO);
567 	}
568 }
569 
570 static void
571 zio_decrypt(zio_t *zio, abd_t *data, uint64_t size)
572 {
573 	int ret;
574 	void *tmp;
575 	blkptr_t *bp = zio->io_bp;
576 	spa_t *spa = zio->io_spa;
577 	uint64_t dsobj = zio->io_bookmark.zb_objset;
578 	uint64_t lsize = BP_GET_LSIZE(bp);
579 	dmu_object_type_t ot = BP_GET_TYPE(bp);
580 	uint8_t salt[ZIO_DATA_SALT_LEN];
581 	uint8_t iv[ZIO_DATA_IV_LEN];
582 	uint8_t mac[ZIO_DATA_MAC_LEN];
583 	boolean_t no_crypt = B_FALSE;
584 
585 	ASSERT(BP_USES_CRYPT(bp));
586 	ASSERT3U(size, !=, 0);
587 
588 	if (zio->io_error != 0)
589 		return;
590 
591 	/*
592 	 * Verify the cksum of MACs stored in an indirect bp. It will always
593 	 * be possible to verify this since it does not require an encryption
594 	 * key.
595 	 */
596 	if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {
597 		zio_crypt_decode_mac_bp(bp, mac);
598 
599 		if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
600 			/*
601 			 * We haven't decompressed the data yet, but
602 			 * zio_crypt_do_indirect_mac_checksum() requires
603 			 * decompressed data to be able to parse out the MACs
604 			 * from the indirect block. We decompress it now and
605 			 * throw away the result after we are finished.
606 			 */
607 			abd_t *abd = abd_alloc_linear(lsize, B_TRUE);
608 			ret = zio_decompress_data(BP_GET_COMPRESS(bp),
609 			    zio->io_abd, abd, zio->io_size, lsize,
610 			    &zio->io_prop.zp_complevel);
611 			if (ret != 0) {
612 				abd_free(abd);
613 				ret = SET_ERROR(EIO);
614 				goto error;
615 			}
616 			ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
617 			    abd, lsize, BP_SHOULD_BYTESWAP(bp), mac);
618 			abd_free(abd);
619 		} else {
620 			ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
621 			    zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac);
622 		}
623 		abd_copy(data, zio->io_abd, size);
624 
625 		if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) {
626 			ret = zio_handle_decrypt_injection(spa,
627 			    &zio->io_bookmark, ot, ECKSUM);
628 		}
629 		if (ret != 0)
630 			goto error;
631 
632 		return;
633 	}
634 
635 	/*
636 	 * If this is an authenticated block, just check the MAC. It would be
637 	 * nice to separate this out into its own flag, but when this was done,
638 	 * we had run out of bits in what is now zio_flag_t. Future cleanup
639 	 * could make this a flag bit.
640 	 */
641 	if (BP_IS_AUTHENTICATED(bp)) {
642 		if (ot == DMU_OT_OBJSET) {
643 			ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa,
644 			    dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp));
645 		} else {
646 			zio_crypt_decode_mac_bp(bp, mac);
647 			ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj,
648 			    zio->io_abd, size, mac);
649 			if (zio_injection_enabled && ret == 0) {
650 				ret = zio_handle_decrypt_injection(spa,
651 				    &zio->io_bookmark, ot, ECKSUM);
652 			}
653 		}
654 		abd_copy(data, zio->io_abd, size);
655 
656 		if (ret != 0)
657 			goto error;
658 
659 		return;
660 	}
661 
662 	zio_crypt_decode_params_bp(bp, salt, iv);
663 
664 	if (ot == DMU_OT_INTENT_LOG) {
665 		tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t));
666 		zio_crypt_decode_mac_zil(tmp, mac);
667 		abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t));
668 	} else {
669 		zio_crypt_decode_mac_bp(bp, mac);
670 	}
671 
672 	ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp),
673 	    BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data,
674 	    zio->io_abd, &no_crypt);
675 	if (no_crypt)
676 		abd_copy(data, zio->io_abd, size);
677 
678 	if (ret != 0)
679 		goto error;
680 
681 	return;
682 
683 error:
684 	/* assert that the key was found unless this was speculative */
685 	ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE));
686 
687 	/*
688 	 * If there was a decryption / authentication error return EIO as
689 	 * the io_error. If this was not a speculative zio, create an ereport.
690 	 */
691 	if (ret == ECKSUM) {
692 		zio->io_error = SET_ERROR(EIO);
693 		if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
694 			spa_log_error(spa, &zio->io_bookmark,
695 			    BP_GET_LOGICAL_BIRTH(zio->io_bp));
696 			(void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
697 			    spa, NULL, &zio->io_bookmark, zio, 0);
698 		}
699 	} else {
700 		zio->io_error = ret;
701 	}
702 }
703 
704 /*
705  * ==========================================================================
706  * I/O parent/child relationships and pipeline interlocks
707  * ==========================================================================
708  */
709 zio_t *
710 zio_walk_parents(zio_t *cio, zio_link_t **zl)
711 {
712 	list_t *pl = &cio->io_parent_list;
713 
714 	*zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
715 	if (*zl == NULL)
716 		return (NULL);
717 
718 	ASSERT((*zl)->zl_child == cio);
719 	return ((*zl)->zl_parent);
720 }
721 
722 zio_t *
723 zio_walk_children(zio_t *pio, zio_link_t **zl)
724 {
725 	list_t *cl = &pio->io_child_list;
726 
727 	ASSERT(MUTEX_HELD(&pio->io_lock));
728 
729 	*zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
730 	if (*zl == NULL)
731 		return (NULL);
732 
733 	ASSERT((*zl)->zl_parent == pio);
734 	return ((*zl)->zl_child);
735 }
736 
737 zio_t *
738 zio_unique_parent(zio_t *cio)
739 {
740 	zio_link_t *zl = NULL;
741 	zio_t *pio = zio_walk_parents(cio, &zl);
742 
743 	VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
744 	return (pio);
745 }
746 
747 void
748 zio_add_child(zio_t *pio, zio_t *cio)
749 {
750 	/*
751 	 * Logical I/Os can have logical, gang, or vdev children.
752 	 * Gang I/Os can have gang or vdev children.
753 	 * Vdev I/Os can only have vdev children.
754 	 * The following ASSERT captures all of these constraints.
755 	 */
756 	ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
757 
758 	/* Parent should not have READY stage if child doesn't have it. */
759 	IMPLY((cio->io_pipeline & ZIO_STAGE_READY) == 0 &&
760 	    (cio->io_child_type != ZIO_CHILD_VDEV),
761 	    (pio->io_pipeline & ZIO_STAGE_READY) == 0);
762 
763 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
764 	zl->zl_parent = pio;
765 	zl->zl_child = cio;
766 
767 	mutex_enter(&pio->io_lock);
768 	mutex_enter(&cio->io_lock);
769 
770 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
771 
772 	uint64_t *countp = pio->io_children[cio->io_child_type];
773 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
774 		countp[w] += !cio->io_state[w];
775 
776 	list_insert_head(&pio->io_child_list, zl);
777 	list_insert_head(&cio->io_parent_list, zl);
778 
779 	mutex_exit(&cio->io_lock);
780 	mutex_exit(&pio->io_lock);
781 }
782 
783 void
784 zio_add_child_first(zio_t *pio, zio_t *cio)
785 {
786 	/*
787 	 * Logical I/Os can have logical, gang, or vdev children.
788 	 * Gang I/Os can have gang or vdev children.
789 	 * Vdev I/Os can only have vdev children.
790 	 * The following ASSERT captures all of these constraints.
791 	 */
792 	ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
793 
794 	/* Parent should not have READY stage if child doesn't have it. */
795 	IMPLY((cio->io_pipeline & ZIO_STAGE_READY) == 0 &&
796 	    (cio->io_child_type != ZIO_CHILD_VDEV),
797 	    (pio->io_pipeline & ZIO_STAGE_READY) == 0);
798 
799 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
800 	zl->zl_parent = pio;
801 	zl->zl_child = cio;
802 
803 	ASSERT(list_is_empty(&cio->io_parent_list));
804 	list_insert_head(&cio->io_parent_list, zl);
805 
806 	mutex_enter(&pio->io_lock);
807 
808 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
809 
810 	uint64_t *countp = pio->io_children[cio->io_child_type];
811 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
812 		countp[w] += !cio->io_state[w];
813 
814 	list_insert_head(&pio->io_child_list, zl);
815 
816 	mutex_exit(&pio->io_lock);
817 }
818 
819 static void
820 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
821 {
822 	ASSERT(zl->zl_parent == pio);
823 	ASSERT(zl->zl_child == cio);
824 
825 	mutex_enter(&pio->io_lock);
826 	mutex_enter(&cio->io_lock);
827 
828 	list_remove(&pio->io_child_list, zl);
829 	list_remove(&cio->io_parent_list, zl);
830 
831 	mutex_exit(&cio->io_lock);
832 	mutex_exit(&pio->io_lock);
833 	kmem_cache_free(zio_link_cache, zl);
834 }
835 
836 static boolean_t
837 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
838 {
839 	boolean_t waiting = B_FALSE;
840 
841 	mutex_enter(&zio->io_lock);
842 	ASSERT(zio->io_stall == NULL);
843 	for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
844 		if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
845 			continue;
846 
847 		uint64_t *countp = &zio->io_children[c][wait];
848 		if (*countp != 0) {
849 			zio->io_stage >>= 1;
850 			ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
851 			zio->io_stall = countp;
852 			waiting = B_TRUE;
853 			break;
854 		}
855 	}
856 	mutex_exit(&zio->io_lock);
857 	return (waiting);
858 }
859 
860 __attribute__((always_inline))
861 static inline void
862 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait,
863     zio_t **next_to_executep)
864 {
865 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
866 	int *errorp = &pio->io_child_error[zio->io_child_type];
867 
868 	mutex_enter(&pio->io_lock);
869 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
870 		*errorp = zio_worst_error(*errorp, zio->io_error);
871 	pio->io_reexecute |= zio->io_reexecute;
872 	ASSERT3U(*countp, >, 0);
873 
874 	/*
875 	 * Propogate the Direct I/O checksum verify failure to the parent.
876 	 */
877 	if (zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR)
878 		pio->io_flags |= ZIO_FLAG_DIO_CHKSUM_ERR;
879 
880 	(*countp)--;
881 
882 	if (*countp == 0 && pio->io_stall == countp) {
883 		zio_taskq_type_t type =
884 		    pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
885 		    ZIO_TASKQ_INTERRUPT;
886 		pio->io_stall = NULL;
887 		mutex_exit(&pio->io_lock);
888 
889 		/*
890 		 * If we can tell the caller to execute this parent next, do
891 		 * so. We do this if the parent's zio type matches the child's
892 		 * type, or if it's a zio_null() with no done callback, and so
893 		 * has no actual work to do. Otherwise dispatch the parent zio
894 		 * in its own taskq.
895 		 *
896 		 * Having the caller execute the parent when possible reduces
897 		 * locking on the zio taskq's, reduces context switch
898 		 * overhead, and has no recursion penalty.  Note that one
899 		 * read from disk typically causes at least 3 zio's: a
900 		 * zio_null(), the logical zio_read(), and then a physical
901 		 * zio.  When the physical ZIO completes, we are able to call
902 		 * zio_done() on all 3 of these zio's from one invocation of
903 		 * zio_execute() by returning the parent back to
904 		 * zio_execute().  Since the parent isn't executed until this
905 		 * thread returns back to zio_execute(), the caller should do
906 		 * so promptly.
907 		 *
908 		 * In other cases, dispatching the parent prevents
909 		 * overflowing the stack when we have deeply nested
910 		 * parent-child relationships, as we do with the "mega zio"
911 		 * of writes for spa_sync(), and the chain of ZIL blocks.
912 		 */
913 		if (next_to_executep != NULL && *next_to_executep == NULL &&
914 		    (pio->io_type == zio->io_type ||
915 		    (pio->io_type == ZIO_TYPE_NULL && !pio->io_done))) {
916 			*next_to_executep = pio;
917 		} else {
918 			zio_taskq_dispatch(pio, type, B_FALSE);
919 		}
920 	} else {
921 		mutex_exit(&pio->io_lock);
922 	}
923 }
924 
925 static void
926 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
927 {
928 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
929 		zio->io_error = zio->io_child_error[c];
930 }
931 
932 int
933 zio_bookmark_compare(const void *x1, const void *x2)
934 {
935 	const zio_t *z1 = x1;
936 	const zio_t *z2 = x2;
937 
938 	if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
939 		return (-1);
940 	if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
941 		return (1);
942 
943 	if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
944 		return (-1);
945 	if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
946 		return (1);
947 
948 	if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
949 		return (-1);
950 	if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
951 		return (1);
952 
953 	if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
954 		return (-1);
955 	if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
956 		return (1);
957 
958 	if (z1 < z2)
959 		return (-1);
960 	if (z1 > z2)
961 		return (1);
962 
963 	return (0);
964 }
965 
966 /*
967  * ==========================================================================
968  * Create the various types of I/O (read, write, free, etc)
969  * ==========================================================================
970  */
971 static zio_t *
972 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
973     abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
974     void *private, zio_type_t type, zio_priority_t priority,
975     zio_flag_t flags, vdev_t *vd, uint64_t offset,
976     const zbookmark_phys_t *zb, enum zio_stage stage,
977     enum zio_stage pipeline)
978 {
979 	zio_t *zio;
980 
981 	IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE);
982 	ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
983 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
984 
985 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
986 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
987 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
988 
989 	IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0);
990 
991 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
992 	memset(zio, 0, sizeof (zio_t));
993 
994 	mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL);
995 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
996 
997 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
998 	    offsetof(zio_link_t, zl_parent_node));
999 	list_create(&zio->io_child_list, sizeof (zio_link_t),
1000 	    offsetof(zio_link_t, zl_child_node));
1001 	metaslab_trace_init(&zio->io_alloc_list);
1002 
1003 	if (vd != NULL)
1004 		zio->io_child_type = ZIO_CHILD_VDEV;
1005 	else if (flags & ZIO_FLAG_GANG_CHILD)
1006 		zio->io_child_type = ZIO_CHILD_GANG;
1007 	else if (flags & ZIO_FLAG_DDT_CHILD)
1008 		zio->io_child_type = ZIO_CHILD_DDT;
1009 	else
1010 		zio->io_child_type = ZIO_CHILD_LOGICAL;
1011 
1012 	if (bp != NULL) {
1013 		if (type != ZIO_TYPE_WRITE ||
1014 		    zio->io_child_type == ZIO_CHILD_DDT) {
1015 			zio->io_bp_copy = *bp;
1016 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
1017 		} else {
1018 			zio->io_bp = (blkptr_t *)bp;
1019 		}
1020 		zio->io_bp_orig = *bp;
1021 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
1022 			zio->io_logical = zio;
1023 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
1024 			pipeline |= ZIO_GANG_STAGES;
1025 		if (flags & ZIO_FLAG_PREALLOCATED) {
1026 			BP_ZERO_DVAS(zio->io_bp);
1027 			BP_SET_BIRTH(zio->io_bp, 0, 0);
1028 		}
1029 	}
1030 
1031 	zio->io_spa = spa;
1032 	zio->io_txg = txg;
1033 	zio->io_done = done;
1034 	zio->io_private = private;
1035 	zio->io_type = type;
1036 	zio->io_priority = priority;
1037 	zio->io_vd = vd;
1038 	zio->io_offset = offset;
1039 	zio->io_orig_abd = zio->io_abd = data;
1040 	zio->io_orig_size = zio->io_size = psize;
1041 	zio->io_lsize = lsize;
1042 	zio->io_orig_flags = zio->io_flags = flags;
1043 	zio->io_orig_stage = zio->io_stage = stage;
1044 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
1045 	zio->io_pipeline_trace = ZIO_STAGE_OPEN;
1046 	zio->io_allocator = ZIO_ALLOCATOR_NONE;
1047 
1048 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY) ||
1049 	    (pipeline & ZIO_STAGE_READY) == 0;
1050 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
1051 
1052 	if (zb != NULL)
1053 		zio->io_bookmark = *zb;
1054 
1055 	if (pio != NULL) {
1056 		zio->io_metaslab_class = pio->io_metaslab_class;
1057 		if (zio->io_logical == NULL)
1058 			zio->io_logical = pio->io_logical;
1059 		if (zio->io_child_type == ZIO_CHILD_GANG)
1060 			zio->io_gang_leader = pio->io_gang_leader;
1061 		zio_add_child_first(pio, zio);
1062 	}
1063 
1064 	taskq_init_ent(&zio->io_tqent);
1065 
1066 	return (zio);
1067 }
1068 
1069 void
1070 zio_destroy(zio_t *zio)
1071 {
1072 	metaslab_trace_fini(&zio->io_alloc_list);
1073 	list_destroy(&zio->io_parent_list);
1074 	list_destroy(&zio->io_child_list);
1075 	mutex_destroy(&zio->io_lock);
1076 	cv_destroy(&zio->io_cv);
1077 	kmem_cache_free(zio_cache, zio);
1078 }
1079 
1080 /*
1081  * ZIO intended to be between others.  Provides synchronization at READY
1082  * and DONE pipeline stages and calls the respective callbacks.
1083  */
1084 zio_t *
1085 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
1086     void *private, zio_flag_t flags)
1087 {
1088 	zio_t *zio;
1089 
1090 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1091 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1092 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
1093 
1094 	return (zio);
1095 }
1096 
1097 /*
1098  * ZIO intended to be a root of a tree.  Unlike null ZIO does not have a
1099  * READY pipeline stage (is ready on creation), so it should not be used
1100  * as child of any ZIO that may need waiting for grandchildren READY stage
1101  * (any other ZIO type).
1102  */
1103 zio_t *
1104 zio_root(spa_t *spa, zio_done_func_t *done, void *private, zio_flag_t flags)
1105 {
1106 	zio_t *zio;
1107 
1108 	zio = zio_create(NULL, spa, 0, NULL, NULL, 0, 0, done, private,
1109 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, NULL, 0, NULL,
1110 	    ZIO_STAGE_OPEN, ZIO_ROOT_PIPELINE);
1111 
1112 	return (zio);
1113 }
1114 
1115 static int
1116 zfs_blkptr_verify_log(spa_t *spa, const blkptr_t *bp,
1117     enum blk_verify_flag blk_verify, const char *fmt, ...)
1118 {
1119 	va_list adx;
1120 	char buf[256];
1121 
1122 	va_start(adx, fmt);
1123 	(void) vsnprintf(buf, sizeof (buf), fmt, adx);
1124 	va_end(adx);
1125 
1126 	zfs_dbgmsg("bad blkptr at %px: "
1127 	    "DVA[0]=%#llx/%#llx "
1128 	    "DVA[1]=%#llx/%#llx "
1129 	    "DVA[2]=%#llx/%#llx "
1130 	    "prop=%#llx "
1131 	    "pad=%#llx,%#llx "
1132 	    "phys_birth=%#llx "
1133 	    "birth=%#llx "
1134 	    "fill=%#llx "
1135 	    "cksum=%#llx/%#llx/%#llx/%#llx",
1136 	    bp,
1137 	    (long long)bp->blk_dva[0].dva_word[0],
1138 	    (long long)bp->blk_dva[0].dva_word[1],
1139 	    (long long)bp->blk_dva[1].dva_word[0],
1140 	    (long long)bp->blk_dva[1].dva_word[1],
1141 	    (long long)bp->blk_dva[2].dva_word[0],
1142 	    (long long)bp->blk_dva[2].dva_word[1],
1143 	    (long long)bp->blk_prop,
1144 	    (long long)bp->blk_pad[0],
1145 	    (long long)bp->blk_pad[1],
1146 	    (long long)BP_GET_PHYSICAL_BIRTH(bp),
1147 	    (long long)BP_GET_LOGICAL_BIRTH(bp),
1148 	    (long long)bp->blk_fill,
1149 	    (long long)bp->blk_cksum.zc_word[0],
1150 	    (long long)bp->blk_cksum.zc_word[1],
1151 	    (long long)bp->blk_cksum.zc_word[2],
1152 	    (long long)bp->blk_cksum.zc_word[3]);
1153 	switch (blk_verify) {
1154 	case BLK_VERIFY_HALT:
1155 		zfs_panic_recover("%s: %s", spa_name(spa), buf);
1156 		break;
1157 	case BLK_VERIFY_LOG:
1158 		zfs_dbgmsg("%s: %s", spa_name(spa), buf);
1159 		break;
1160 	case BLK_VERIFY_ONLY:
1161 		break;
1162 	}
1163 
1164 	return (1);
1165 }
1166 
1167 /*
1168  * Verify the block pointer fields contain reasonable values.  This means
1169  * it only contains known object types, checksum/compression identifiers,
1170  * block sizes within the maximum allowed limits, valid DVAs, etc.
1171  *
1172  * If everything checks out 0 is returned.  The zfs_blkptr_verify
1173  * argument controls the behavior when an invalid field is detected.
1174  *
1175  * Values for blk_verify_flag:
1176  *   BLK_VERIFY_ONLY: evaluate the block
1177  *   BLK_VERIFY_LOG: evaluate the block and log problems
1178  *   BLK_VERIFY_HALT: call zfs_panic_recover on error
1179  *
1180  * Values for blk_config_flag:
1181  *   BLK_CONFIG_HELD: caller holds SCL_VDEV for writer
1182  *   BLK_CONFIG_NEEDED: caller holds no config lock, SCL_VDEV will be
1183  *   obtained for reader
1184  *   BLK_CONFIG_SKIP: skip checks which require SCL_VDEV, for better
1185  *   performance
1186  */
1187 int
1188 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp,
1189     enum blk_config_flag blk_config, enum blk_verify_flag blk_verify)
1190 {
1191 	int errors = 0;
1192 
1193 	if (unlikely(!DMU_OT_IS_VALID(BP_GET_TYPE(bp)))) {
1194 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1195 		    "blkptr at %px has invalid TYPE %llu",
1196 		    bp, (longlong_t)BP_GET_TYPE(bp));
1197 	}
1198 	if (unlikely(BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS)) {
1199 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1200 		    "blkptr at %px has invalid COMPRESS %llu",
1201 		    bp, (longlong_t)BP_GET_COMPRESS(bp));
1202 	}
1203 	if (unlikely(BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE)) {
1204 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1205 		    "blkptr at %px has invalid LSIZE %llu",
1206 		    bp, (longlong_t)BP_GET_LSIZE(bp));
1207 	}
1208 	if (BP_IS_EMBEDDED(bp)) {
1209 		if (unlikely(BPE_GET_ETYPE(bp) >= NUM_BP_EMBEDDED_TYPES)) {
1210 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1211 			    "blkptr at %px has invalid ETYPE %llu",
1212 			    bp, (longlong_t)BPE_GET_ETYPE(bp));
1213 		}
1214 		if (unlikely(BPE_GET_PSIZE(bp) > BPE_PAYLOAD_SIZE)) {
1215 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1216 			    "blkptr at %px has invalid PSIZE %llu",
1217 			    bp, (longlong_t)BPE_GET_PSIZE(bp));
1218 		}
1219 		return (errors ? ECKSUM : 0);
1220 	} else if (BP_IS_HOLE(bp)) {
1221 		/*
1222 		 * Holes are allowed (expected, even) to have no DVAs, no
1223 		 * checksum, and no psize.
1224 		 */
1225 		return (errors ? ECKSUM : 0);
1226 	} else if (unlikely(!DVA_IS_VALID(&bp->blk_dva[0]))) {
1227 		/* Non-hole, non-embedded BPs _must_ have at least one DVA */
1228 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1229 		    "blkptr at %px has no valid DVAs", bp);
1230 	}
1231 	if (unlikely(BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS)) {
1232 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1233 		    "blkptr at %px has invalid CHECKSUM %llu",
1234 		    bp, (longlong_t)BP_GET_CHECKSUM(bp));
1235 	}
1236 	if (unlikely(BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE)) {
1237 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1238 		    "blkptr at %px has invalid PSIZE %llu",
1239 		    bp, (longlong_t)BP_GET_PSIZE(bp));
1240 	}
1241 
1242 	/*
1243 	 * Do not verify individual DVAs if the config is not trusted. This
1244 	 * will be done once the zio is executed in vdev_mirror_map_alloc.
1245 	 */
1246 	if (unlikely(!spa->spa_trust_config))
1247 		return (errors ? ECKSUM : 0);
1248 
1249 	switch (blk_config) {
1250 	case BLK_CONFIG_HELD:
1251 		ASSERT(spa_config_held(spa, SCL_VDEV, RW_WRITER));
1252 		break;
1253 	case BLK_CONFIG_NEEDED:
1254 		spa_config_enter(spa, SCL_VDEV, bp, RW_READER);
1255 		break;
1256 	case BLK_CONFIG_NEEDED_TRY:
1257 		if (!spa_config_tryenter(spa, SCL_VDEV, bp, RW_READER))
1258 			return (EBUSY);
1259 		break;
1260 	case BLK_CONFIG_SKIP:
1261 		return (errors ? ECKSUM : 0);
1262 	default:
1263 		panic("invalid blk_config %u", blk_config);
1264 	}
1265 
1266 	/*
1267 	 * Pool-specific checks.
1268 	 *
1269 	 * Note: it would be nice to verify that the logical birth
1270 	 * and physical birth are not too large.  However,
1271 	 * spa_freeze() allows the birth time of log blocks (and
1272 	 * dmu_sync()-ed blocks that are in the log) to be arbitrarily
1273 	 * large.
1274 	 */
1275 	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
1276 		const dva_t *dva = &bp->blk_dva[i];
1277 		uint64_t vdevid = DVA_GET_VDEV(dva);
1278 
1279 		if (unlikely(vdevid >= spa->spa_root_vdev->vdev_children)) {
1280 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1281 			    "blkptr at %px DVA %u has invalid VDEV %llu",
1282 			    bp, i, (longlong_t)vdevid);
1283 			continue;
1284 		}
1285 		vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1286 		if (unlikely(vd == NULL)) {
1287 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1288 			    "blkptr at %px DVA %u has invalid VDEV %llu",
1289 			    bp, i, (longlong_t)vdevid);
1290 			continue;
1291 		}
1292 		if (unlikely(vd->vdev_ops == &vdev_hole_ops)) {
1293 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1294 			    "blkptr at %px DVA %u has hole VDEV %llu",
1295 			    bp, i, (longlong_t)vdevid);
1296 			continue;
1297 		}
1298 		if (vd->vdev_ops == &vdev_missing_ops) {
1299 			/*
1300 			 * "missing" vdevs are valid during import, but we
1301 			 * don't have their detailed info (e.g. asize), so
1302 			 * we can't perform any more checks on them.
1303 			 */
1304 			continue;
1305 		}
1306 		uint64_t offset = DVA_GET_OFFSET(dva);
1307 		uint64_t asize = DVA_GET_ASIZE(dva);
1308 		if (DVA_GET_GANG(dva))
1309 			asize = vdev_gang_header_asize(vd);
1310 		if (unlikely(offset + asize > vd->vdev_asize)) {
1311 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1312 			    "blkptr at %px DVA %u has invalid OFFSET %llu",
1313 			    bp, i, (longlong_t)offset);
1314 		}
1315 	}
1316 	if (blk_config == BLK_CONFIG_NEEDED || blk_config ==
1317 	    BLK_CONFIG_NEEDED_TRY)
1318 		spa_config_exit(spa, SCL_VDEV, bp);
1319 
1320 	return (errors ? ECKSUM : 0);
1321 }
1322 
1323 boolean_t
1324 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
1325 {
1326 	(void) bp;
1327 	uint64_t vdevid = DVA_GET_VDEV(dva);
1328 
1329 	if (vdevid >= spa->spa_root_vdev->vdev_children)
1330 		return (B_FALSE);
1331 
1332 	vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1333 	if (vd == NULL)
1334 		return (B_FALSE);
1335 
1336 	if (vd->vdev_ops == &vdev_hole_ops)
1337 		return (B_FALSE);
1338 
1339 	if (vd->vdev_ops == &vdev_missing_ops) {
1340 		return (B_FALSE);
1341 	}
1342 
1343 	uint64_t offset = DVA_GET_OFFSET(dva);
1344 	uint64_t asize = DVA_GET_ASIZE(dva);
1345 
1346 	if (DVA_GET_GANG(dva))
1347 		asize = vdev_gang_header_asize(vd);
1348 	if (offset + asize > vd->vdev_asize)
1349 		return (B_FALSE);
1350 
1351 	return (B_TRUE);
1352 }
1353 
1354 zio_t *
1355 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
1356     abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
1357     zio_priority_t priority, zio_flag_t flags, const zbookmark_phys_t *zb)
1358 {
1359 	zio_t *zio;
1360 
1361 	zio = zio_create(pio, spa, BP_GET_BIRTH(bp), bp,
1362 	    data, size, size, done, private,
1363 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
1364 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1365 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
1366 
1367 	return (zio);
1368 }
1369 
1370 zio_t *
1371 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
1372     abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
1373     zio_done_func_t *ready, zio_done_func_t *children_ready,
1374     zio_done_func_t *done, void *private, zio_priority_t priority,
1375     zio_flag_t flags, const zbookmark_phys_t *zb)
1376 {
1377 	zio_t *zio;
1378 	enum zio_stage pipeline = zp->zp_direct_write == B_TRUE ?
1379 	    ZIO_DIRECT_WRITE_PIPELINE : (flags & ZIO_FLAG_DDT_CHILD) ?
1380 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE;
1381 
1382 
1383 	zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
1384 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
1385 	    ZIO_STAGE_OPEN, pipeline);
1386 
1387 	zio->io_ready = ready;
1388 	zio->io_children_ready = children_ready;
1389 	zio->io_prop = *zp;
1390 
1391 	/*
1392 	 * Data can be NULL if we are going to call zio_write_override() to
1393 	 * provide the already-allocated BP.  But we may need the data to
1394 	 * verify a dedup hit (if requested).  In this case, don't try to
1395 	 * dedup (just take the already-allocated BP verbatim). Encrypted
1396 	 * dedup blocks need data as well so we also disable dedup in this
1397 	 * case.
1398 	 */
1399 	if (data == NULL &&
1400 	    (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) {
1401 		zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
1402 	}
1403 
1404 	return (zio);
1405 }
1406 
1407 zio_t *
1408 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
1409     uint64_t size, zio_done_func_t *done, void *private,
1410     zio_priority_t priority, zio_flag_t flags, zbookmark_phys_t *zb)
1411 {
1412 	zio_t *zio;
1413 
1414 	zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
1415 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
1416 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
1417 
1418 	return (zio);
1419 }
1420 
1421 void
1422 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, int gang_copies,
1423     boolean_t nopwrite, boolean_t brtwrite)
1424 {
1425 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1426 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1427 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1428 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
1429 	ASSERT(!brtwrite || !nopwrite);
1430 
1431 	/*
1432 	 * We must reset the io_prop to match the values that existed
1433 	 * when the bp was first written by dmu_sync() keeping in mind
1434 	 * that nopwrite and dedup are mutually exclusive.
1435 	 */
1436 	zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
1437 	zio->io_prop.zp_nopwrite = nopwrite;
1438 	zio->io_prop.zp_brtwrite = brtwrite;
1439 	zio->io_prop.zp_copies = copies;
1440 	zio->io_prop.zp_gang_copies = gang_copies;
1441 	zio->io_bp_override = bp;
1442 }
1443 
1444 void
1445 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
1446 {
1447 
1448 	(void) zfs_blkptr_verify(spa, bp, BLK_CONFIG_NEEDED, BLK_VERIFY_HALT);
1449 
1450 	/*
1451 	 * The check for EMBEDDED is a performance optimization.  We
1452 	 * process the free here (by ignoring it) rather than
1453 	 * putting it on the list and then processing it in zio_free_sync().
1454 	 */
1455 	if (BP_IS_EMBEDDED(bp))
1456 		return;
1457 
1458 	/*
1459 	 * Frees that are for the currently-syncing txg, are not going to be
1460 	 * deferred, and which will not need to do a read (i.e. not GANG or
1461 	 * DEDUP), can be processed immediately.  Otherwise, put them on the
1462 	 * in-memory list for later processing.
1463 	 *
1464 	 * Note that we only defer frees after zfs_sync_pass_deferred_free
1465 	 * when the log space map feature is disabled. [see relevant comment
1466 	 * in spa_sync_iterate_to_convergence()]
1467 	 */
1468 	if (BP_IS_GANG(bp) ||
1469 	    BP_GET_DEDUP(bp) ||
1470 	    txg != spa->spa_syncing_txg ||
1471 	    (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free &&
1472 	    !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) ||
1473 	    brt_maybe_exists(spa, bp)) {
1474 		metaslab_check_free(spa, bp);
1475 		bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1476 	} else {
1477 		VERIFY3P(zio_free_sync(NULL, spa, txg, bp, 0), ==, NULL);
1478 	}
1479 }
1480 
1481 /*
1482  * To improve performance, this function may return NULL if we were able
1483  * to do the free immediately.  This avoids the cost of creating a zio
1484  * (and linking it to the parent, etc).
1485  */
1486 zio_t *
1487 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1488     zio_flag_t flags)
1489 {
1490 	ASSERT(!BP_IS_HOLE(bp));
1491 	ASSERT(spa_syncing_txg(spa) == txg);
1492 
1493 	if (BP_IS_EMBEDDED(bp))
1494 		return (NULL);
1495 
1496 	metaslab_check_free(spa, bp);
1497 	arc_freed(spa, bp);
1498 	dsl_scan_freed(spa, bp);
1499 
1500 	if (BP_IS_GANG(bp) ||
1501 	    BP_GET_DEDUP(bp) ||
1502 	    brt_maybe_exists(spa, bp)) {
1503 		/*
1504 		 * GANG, DEDUP and BRT blocks can induce a read (for the gang
1505 		 * block header, the DDT or the BRT), so issue them
1506 		 * asynchronously so that this thread is not tied up.
1507 		 */
1508 		enum zio_stage stage =
1509 		    ZIO_FREE_PIPELINE | ZIO_STAGE_ISSUE_ASYNC;
1510 
1511 		return (zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1512 		    BP_GET_PSIZE(bp), NULL, NULL,
1513 		    ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1514 		    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage));
1515 	} else {
1516 		metaslab_free(spa, bp, txg, B_FALSE);
1517 		return (NULL);
1518 	}
1519 }
1520 
1521 zio_t *
1522 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1523     zio_done_func_t *done, void *private, zio_flag_t flags)
1524 {
1525 	zio_t *zio;
1526 
1527 	(void) zfs_blkptr_verify(spa, bp, (flags & ZIO_FLAG_CONFIG_WRITER) ?
1528 	    BLK_CONFIG_HELD : BLK_CONFIG_NEEDED, BLK_VERIFY_HALT);
1529 
1530 	if (BP_IS_EMBEDDED(bp))
1531 		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1532 
1533 	/*
1534 	 * A claim is an allocation of a specific block.  Claims are needed
1535 	 * to support immediate writes in the intent log.  The issue is that
1536 	 * immediate writes contain committed data, but in a txg that was
1537 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
1538 	 * the intent log claims all blocks that contain immediate write data
1539 	 * so that the SPA knows they're in use.
1540 	 *
1541 	 * All claims *must* be resolved in the first txg -- before the SPA
1542 	 * starts allocating blocks -- so that nothing is allocated twice.
1543 	 * If txg == 0 we just verify that the block is claimable.
1544 	 */
1545 	ASSERT3U(BP_GET_LOGICAL_BIRTH(&spa->spa_uberblock.ub_rootbp), <,
1546 	    spa_min_claim_txg(spa));
1547 	ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1548 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(8) */
1549 
1550 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1551 	    BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1552 	    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1553 	ASSERT0(zio->io_queued_timestamp);
1554 
1555 	return (zio);
1556 }
1557 
1558 zio_t *
1559 zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1560     zio_done_func_t *done, void *private, zio_priority_t priority,
1561     zio_flag_t flags, enum trim_flag trim_flags)
1562 {
1563 	zio_t *zio;
1564 
1565 	ASSERT0(vd->vdev_children);
1566 	ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
1567 	ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1568 	ASSERT3U(size, !=, 0);
1569 
1570 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done,
1571 	    private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL,
1572 	    vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE);
1573 	zio->io_trim_flags = trim_flags;
1574 
1575 	return (zio);
1576 }
1577 
1578 zio_t *
1579 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1580     abd_t *data, int checksum, zio_done_func_t *done, void *private,
1581     zio_priority_t priority, zio_flag_t flags, boolean_t labels)
1582 {
1583 	zio_t *zio;
1584 
1585 	ASSERT(vd->vdev_children == 0);
1586 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1587 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1588 	ASSERT3U(offset + size, <=, vd->vdev_psize);
1589 
1590 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1591 	    private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1592 	    offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1593 
1594 	zio->io_prop.zp_checksum = checksum;
1595 
1596 	return (zio);
1597 }
1598 
1599 zio_t *
1600 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1601     abd_t *data, int checksum, zio_done_func_t *done, void *private,
1602     zio_priority_t priority, zio_flag_t flags, boolean_t labels)
1603 {
1604 	zio_t *zio;
1605 
1606 	ASSERT(vd->vdev_children == 0);
1607 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1608 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1609 	ASSERT3U(offset + size, <=, vd->vdev_psize);
1610 
1611 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1612 	    private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1613 	    offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1614 
1615 	zio->io_prop.zp_checksum = checksum;
1616 
1617 	if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1618 		/*
1619 		 * zec checksums are necessarily destructive -- they modify
1620 		 * the end of the write buffer to hold the verifier/checksum.
1621 		 * Therefore, we must make a local copy in case the data is
1622 		 * being written to multiple places in parallel.
1623 		 */
1624 		abd_t *wbuf = abd_alloc_sametype(data, size);
1625 		abd_copy(wbuf, data, size);
1626 
1627 		zio_push_transform(zio, wbuf, size, size, NULL);
1628 	}
1629 
1630 	return (zio);
1631 }
1632 
1633 /*
1634  * Create a child I/O to do some work for us.
1635  */
1636 zio_t *
1637 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1638     abd_t *data, uint64_t size, int type, zio_priority_t priority,
1639     zio_flag_t flags, zio_done_func_t *done, void *private)
1640 {
1641 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1642 	zio_t *zio;
1643 
1644 	/*
1645 	 * vdev child I/Os do not propagate their error to the parent.
1646 	 * Therefore, for correct operation the caller *must* check for
1647 	 * and handle the error in the child i/o's done callback.
1648 	 * The only exceptions are i/os that we don't care about
1649 	 * (OPTIONAL or REPAIR).
1650 	 */
1651 	ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1652 	    done != NULL);
1653 
1654 	if (type == ZIO_TYPE_READ && bp != NULL) {
1655 		/*
1656 		 * If we have the bp, then the child should perform the
1657 		 * checksum and the parent need not.  This pushes error
1658 		 * detection as close to the leaves as possible and
1659 		 * eliminates redundant checksums in the interior nodes.
1660 		 */
1661 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1662 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1663 		/*
1664 		 * We never allow the mirror VDEV to attempt reading from any
1665 		 * additional data copies after the first Direct I/O checksum
1666 		 * verify failure. This is to avoid bad data being written out
1667 		 * through the mirror during self healing. See comment in
1668 		 * vdev_mirror_io_done() for more details.
1669 		 */
1670 		ASSERT0(pio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR);
1671 	} else if (type == ZIO_TYPE_WRITE &&
1672 	    pio->io_prop.zp_direct_write == B_TRUE) {
1673 		/*
1674 		 * By default we only will verify checksums for Direct I/O
1675 		 * writes for Linux. FreeBSD is able to place user pages under
1676 		 * write protection before issuing them to the ZIO pipeline.
1677 		 *
1678 		 * Checksum validation errors will only be reported through
1679 		 * the top-level VDEV, which is set by this child ZIO.
1680 		 */
1681 		ASSERT3P(bp, !=, NULL);
1682 		ASSERT3U(pio->io_child_type, ==, ZIO_CHILD_LOGICAL);
1683 		pipeline |= ZIO_STAGE_DIO_CHECKSUM_VERIFY;
1684 	}
1685 
1686 	if (vd->vdev_ops->vdev_op_leaf) {
1687 		ASSERT0(vd->vdev_children);
1688 		offset += VDEV_LABEL_START_SIZE;
1689 	}
1690 
1691 	flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1692 
1693 	/*
1694 	 * If we've decided to do a repair, the write is not speculative --
1695 	 * even if the original read was.
1696 	 */
1697 	if (flags & ZIO_FLAG_IO_REPAIR)
1698 		flags &= ~ZIO_FLAG_SPECULATIVE;
1699 
1700 	/*
1701 	 * If we're creating a child I/O that is not associated with a
1702 	 * top-level vdev, then the child zio is not an allocating I/O.
1703 	 * If this is a retried I/O then we ignore it since we will
1704 	 * have already processed the original allocating I/O.
1705 	 */
1706 	if (flags & ZIO_FLAG_IO_ALLOCATING &&
1707 	    (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1708 		ASSERT(pio->io_metaslab_class != NULL);
1709 		ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled);
1710 		ASSERT(type == ZIO_TYPE_WRITE);
1711 		ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1712 		ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1713 		ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1714 		    pio->io_child_type == ZIO_CHILD_GANG);
1715 
1716 		flags &= ~ZIO_FLAG_IO_ALLOCATING;
1717 	}
1718 
1719 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1720 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1721 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1722 	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1723 
1724 	return (zio);
1725 }
1726 
1727 zio_t *
1728 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1729     zio_type_t type, zio_priority_t priority, zio_flag_t flags,
1730     zio_done_func_t *done, void *private)
1731 {
1732 	zio_t *zio;
1733 
1734 	ASSERT(vd->vdev_ops->vdev_op_leaf);
1735 
1736 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1737 	    data, size, size, done, private, type, priority,
1738 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1739 	    vd, offset, NULL,
1740 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1741 
1742 	return (zio);
1743 }
1744 
1745 
1746 /*
1747  * Send a flush command to the given vdev. Unlike most zio creation functions,
1748  * the flush zios are issued immediately. You can wait on pio to pause until
1749  * the flushes complete.
1750  */
1751 void
1752 zio_flush(zio_t *pio, vdev_t *vd)
1753 {
1754 	const zio_flag_t flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
1755 	    ZIO_FLAG_DONT_RETRY;
1756 
1757 	if (vd->vdev_nowritecache)
1758 		return;
1759 
1760 	if (vd->vdev_children == 0) {
1761 		zio_nowait(zio_create(pio, vd->vdev_spa, 0, NULL, NULL, 0, 0,
1762 		    NULL, NULL, ZIO_TYPE_FLUSH, ZIO_PRIORITY_NOW, flags, vd, 0,
1763 		    NULL, ZIO_STAGE_OPEN, ZIO_FLUSH_PIPELINE));
1764 	} else {
1765 		for (uint64_t c = 0; c < vd->vdev_children; c++)
1766 			zio_flush(pio, vd->vdev_child[c]);
1767 	}
1768 }
1769 
1770 void
1771 zio_shrink(zio_t *zio, uint64_t size)
1772 {
1773 	ASSERT3P(zio->io_executor, ==, NULL);
1774 	ASSERT3U(zio->io_orig_size, ==, zio->io_size);
1775 	ASSERT3U(size, <=, zio->io_size);
1776 
1777 	/*
1778 	 * We don't shrink for raidz because of problems with the
1779 	 * reconstruction when reading back less than the block size.
1780 	 * Note, BP_IS_RAIDZ() assumes no compression.
1781 	 */
1782 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1783 	if (!BP_IS_RAIDZ(zio->io_bp)) {
1784 		/* we are not doing a raw write */
1785 		ASSERT3U(zio->io_size, ==, zio->io_lsize);
1786 		zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1787 	}
1788 }
1789 
1790 /*
1791  * Round provided allocation size up to a value that can be allocated
1792  * by at least some vdev(s) in the pool with minimum or no additional
1793  * padding and without extra space usage on others
1794  */
1795 static uint64_t
1796 zio_roundup_alloc_size(spa_t *spa, uint64_t size)
1797 {
1798 	if (size > spa->spa_min_alloc)
1799 		return (roundup(size, spa->spa_gcd_alloc));
1800 	return (spa->spa_min_alloc);
1801 }
1802 
1803 size_t
1804 zio_get_compression_max_size(enum zio_compress compress, uint64_t gcd_alloc,
1805     uint64_t min_alloc, size_t s_len)
1806 {
1807 	size_t d_len;
1808 
1809 	/* minimum 12.5% must be saved (legacy value, may be changed later) */
1810 	d_len = s_len - (s_len >> 3);
1811 
1812 	/* ZLE can't use exactly d_len bytes, it needs more, so ignore it */
1813 	if (compress == ZIO_COMPRESS_ZLE)
1814 		return (d_len);
1815 
1816 	d_len = d_len - d_len % gcd_alloc;
1817 
1818 	if (d_len < min_alloc)
1819 		return (BPE_PAYLOAD_SIZE);
1820 	return (d_len);
1821 }
1822 
1823 /*
1824  * ==========================================================================
1825  * Prepare to read and write logical blocks
1826  * ==========================================================================
1827  */
1828 
1829 static zio_t *
1830 zio_read_bp_init(zio_t *zio)
1831 {
1832 	blkptr_t *bp = zio->io_bp;
1833 	uint64_t psize =
1834 	    BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1835 
1836 	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1837 
1838 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1839 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
1840 	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1841 		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1842 		    psize, psize, zio_decompress);
1843 	}
1844 
1845 	if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) ||
1846 	    BP_HAS_INDIRECT_MAC_CKSUM(bp)) &&
1847 	    zio->io_child_type == ZIO_CHILD_LOGICAL) {
1848 		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1849 		    psize, psize, zio_decrypt);
1850 	}
1851 
1852 	if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1853 		int psize = BPE_GET_PSIZE(bp);
1854 		void *data = abd_borrow_buf(zio->io_abd, psize);
1855 
1856 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1857 		decode_embedded_bp_compressed(bp, data);
1858 		abd_return_buf_copy(zio->io_abd, data, psize);
1859 	} else {
1860 		ASSERT(!BP_IS_EMBEDDED(bp));
1861 	}
1862 
1863 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1864 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1865 
1866 	return (zio);
1867 }
1868 
1869 static zio_t *
1870 zio_write_bp_init(zio_t *zio)
1871 {
1872 	if (!IO_IS_ALLOCATING(zio))
1873 		return (zio);
1874 
1875 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1876 
1877 	if (zio->io_bp_override) {
1878 		blkptr_t *bp = zio->io_bp;
1879 		zio_prop_t *zp = &zio->io_prop;
1880 
1881 		ASSERT(BP_GET_LOGICAL_BIRTH(bp) != zio->io_txg);
1882 
1883 		*bp = *zio->io_bp_override;
1884 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1885 
1886 		if (zp->zp_brtwrite)
1887 			return (zio);
1888 
1889 		ASSERT(!BP_GET_DEDUP(zio->io_bp_override));
1890 
1891 		if (BP_IS_EMBEDDED(bp))
1892 			return (zio);
1893 
1894 		/*
1895 		 * If we've been overridden and nopwrite is set then
1896 		 * set the flag accordingly to indicate that a nopwrite
1897 		 * has already occurred.
1898 		 */
1899 		if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1900 			ASSERT(!zp->zp_dedup);
1901 			ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1902 			zio->io_flags |= ZIO_FLAG_NOPWRITE;
1903 			return (zio);
1904 		}
1905 
1906 		ASSERT(!zp->zp_nopwrite);
1907 
1908 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1909 			return (zio);
1910 
1911 		ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1912 		    ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1913 
1914 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum &&
1915 		    !zp->zp_encrypt) {
1916 			BP_SET_DEDUP(bp, 1);
1917 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1918 			return (zio);
1919 		}
1920 
1921 		/*
1922 		 * We were unable to handle this as an override bp, treat
1923 		 * it as a regular write I/O.
1924 		 */
1925 		zio->io_bp_override = NULL;
1926 		*bp = zio->io_bp_orig;
1927 		zio->io_pipeline = zio->io_orig_pipeline;
1928 	}
1929 
1930 	return (zio);
1931 }
1932 
1933 static zio_t *
1934 zio_write_compress(zio_t *zio)
1935 {
1936 	spa_t *spa = zio->io_spa;
1937 	zio_prop_t *zp = &zio->io_prop;
1938 	enum zio_compress compress = zp->zp_compress;
1939 	blkptr_t *bp = zio->io_bp;
1940 	uint64_t lsize = zio->io_lsize;
1941 	uint64_t psize = zio->io_size;
1942 	uint32_t pass = 1;
1943 
1944 	/*
1945 	 * If our children haven't all reached the ready stage,
1946 	 * wait for them and then repeat this pipeline stage.
1947 	 */
1948 	if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1949 	    ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1950 		return (NULL);
1951 	}
1952 
1953 	if (!IO_IS_ALLOCATING(zio))
1954 		return (zio);
1955 
1956 	if (zio->io_children_ready != NULL) {
1957 		/*
1958 		 * Now that all our children are ready, run the callback
1959 		 * associated with this zio in case it wants to modify the
1960 		 * data to be written.
1961 		 */
1962 		ASSERT3U(zp->zp_level, >, 0);
1963 		zio->io_children_ready(zio);
1964 	}
1965 
1966 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1967 	ASSERT(zio->io_bp_override == NULL);
1968 
1969 	if (!BP_IS_HOLE(bp) && BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg) {
1970 		/*
1971 		 * We're rewriting an existing block, which means we're
1972 		 * working on behalf of spa_sync().  For spa_sync() to
1973 		 * converge, it must eventually be the case that we don't
1974 		 * have to allocate new blocks.  But compression changes
1975 		 * the blocksize, which forces a reallocate, and makes
1976 		 * convergence take longer.  Therefore, after the first
1977 		 * few passes, stop compressing to ensure convergence.
1978 		 */
1979 		pass = spa_sync_pass(spa);
1980 
1981 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
1982 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1983 		ASSERT(!BP_GET_DEDUP(bp));
1984 
1985 		if (pass >= zfs_sync_pass_dont_compress)
1986 			compress = ZIO_COMPRESS_OFF;
1987 
1988 		/* Make sure someone doesn't change their mind on overwrites */
1989 		ASSERT(BP_IS_EMBEDDED(bp) || BP_IS_GANG(bp) ||
1990 		    MIN(zp->zp_copies, spa_max_replication(spa))
1991 		    == BP_GET_NDVAS(bp));
1992 	}
1993 
1994 	/* If it's a compressed write that is not raw, compress the buffer. */
1995 	if (compress != ZIO_COMPRESS_OFF &&
1996 	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1997 		abd_t *cabd = NULL;
1998 		if (abd_cmp_zero(zio->io_abd, lsize) == 0)
1999 			psize = 0;
2000 		else if (compress == ZIO_COMPRESS_EMPTY)
2001 			psize = lsize;
2002 		else
2003 			psize = zio_compress_data(compress, zio->io_abd, &cabd,
2004 			    lsize,
2005 			    zio_get_compression_max_size(compress,
2006 			    spa->spa_gcd_alloc, spa->spa_min_alloc, lsize),
2007 			    zp->zp_complevel);
2008 		if (psize == 0) {
2009 			compress = ZIO_COMPRESS_OFF;
2010 		} else if (psize >= lsize) {
2011 			compress = ZIO_COMPRESS_OFF;
2012 			if (cabd != NULL)
2013 				abd_free(cabd);
2014 		} else if (psize <= BPE_PAYLOAD_SIZE && !zp->zp_encrypt &&
2015 		    zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
2016 		    spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
2017 			void *cbuf = abd_borrow_buf_copy(cabd, lsize);
2018 			encode_embedded_bp_compressed(bp,
2019 			    cbuf, compress, lsize, psize);
2020 			BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
2021 			BP_SET_TYPE(bp, zio->io_prop.zp_type);
2022 			BP_SET_LEVEL(bp, zio->io_prop.zp_level);
2023 			abd_return_buf(cabd, cbuf, lsize);
2024 			abd_free(cabd);
2025 			BP_SET_LOGICAL_BIRTH(bp, zio->io_txg);
2026 			zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2027 			ASSERT(spa_feature_is_active(spa,
2028 			    SPA_FEATURE_EMBEDDED_DATA));
2029 			return (zio);
2030 		} else {
2031 			/*
2032 			 * Round compressed size up to the minimum allocation
2033 			 * size of the smallest-ashift device, and zero the
2034 			 * tail. This ensures that the compressed size of the
2035 			 * BP (and thus compressratio property) are correct,
2036 			 * in that we charge for the padding used to fill out
2037 			 * the last sector.
2038 			 */
2039 			size_t rounded = (size_t)zio_roundup_alloc_size(spa,
2040 			    psize);
2041 			if (rounded >= lsize) {
2042 				compress = ZIO_COMPRESS_OFF;
2043 				abd_free(cabd);
2044 				psize = lsize;
2045 			} else {
2046 				abd_zero_off(cabd, psize, rounded - psize);
2047 				psize = rounded;
2048 				zio_push_transform(zio, cabd,
2049 				    psize, lsize, NULL);
2050 			}
2051 		}
2052 
2053 		/*
2054 		 * We were unable to handle this as an override bp, treat
2055 		 * it as a regular write I/O.
2056 		 */
2057 		zio->io_bp_override = NULL;
2058 		*bp = zio->io_bp_orig;
2059 		zio->io_pipeline = zio->io_orig_pipeline;
2060 
2061 	} else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 &&
2062 	    zp->zp_type == DMU_OT_DNODE) {
2063 		/*
2064 		 * The DMU actually relies on the zio layer's compression
2065 		 * to free metadnode blocks that have had all contained
2066 		 * dnodes freed. As a result, even when doing a raw
2067 		 * receive, we must check whether the block can be compressed
2068 		 * to a hole.
2069 		 */
2070 		if (abd_cmp_zero(zio->io_abd, lsize) == 0) {
2071 			psize = 0;
2072 			compress = ZIO_COMPRESS_OFF;
2073 		} else {
2074 			psize = lsize;
2075 		}
2076 	} else if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS &&
2077 	    !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) {
2078 		/*
2079 		 * If we are raw receiving an encrypted dataset we should not
2080 		 * take this codepath because it will change the on-disk block
2081 		 * and decryption will fail.
2082 		 */
2083 		size_t rounded = MIN((size_t)zio_roundup_alloc_size(spa, psize),
2084 		    lsize);
2085 
2086 		if (rounded != psize) {
2087 			abd_t *cdata = abd_alloc_linear(rounded, B_TRUE);
2088 			abd_zero_off(cdata, psize, rounded - psize);
2089 			abd_copy_off(cdata, zio->io_abd, 0, 0, psize);
2090 			psize = rounded;
2091 			zio_push_transform(zio, cdata,
2092 			    psize, rounded, NULL);
2093 		}
2094 	} else {
2095 		ASSERT3U(psize, !=, 0);
2096 	}
2097 
2098 	/*
2099 	 * The final pass of spa_sync() must be all rewrites, but the first
2100 	 * few passes offer a trade-off: allocating blocks defers convergence,
2101 	 * but newly allocated blocks are sequential, so they can be written
2102 	 * to disk faster.  Therefore, we allow the first few passes of
2103 	 * spa_sync() to allocate new blocks, but force rewrites after that.
2104 	 * There should only be a handful of blocks after pass 1 in any case.
2105 	 */
2106 	if (!BP_IS_HOLE(bp) && BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg &&
2107 	    BP_GET_PSIZE(bp) == psize &&
2108 	    pass >= zfs_sync_pass_rewrite) {
2109 		VERIFY3U(psize, !=, 0);
2110 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
2111 
2112 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
2113 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
2114 	} else {
2115 		BP_ZERO(bp);
2116 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2117 	}
2118 
2119 	if (psize == 0) {
2120 		if (BP_GET_LOGICAL_BIRTH(&zio->io_bp_orig) != 0 &&
2121 		    spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
2122 			BP_SET_LSIZE(bp, lsize);
2123 			BP_SET_TYPE(bp, zp->zp_type);
2124 			BP_SET_LEVEL(bp, zp->zp_level);
2125 			BP_SET_BIRTH(bp, zio->io_txg, 0);
2126 		}
2127 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2128 	} else {
2129 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
2130 		BP_SET_LSIZE(bp, lsize);
2131 		BP_SET_TYPE(bp, zp->zp_type);
2132 		BP_SET_LEVEL(bp, zp->zp_level);
2133 		BP_SET_PSIZE(bp, psize);
2134 		BP_SET_COMPRESS(bp, compress);
2135 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
2136 		BP_SET_DEDUP(bp, zp->zp_dedup);
2137 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
2138 		if (zp->zp_dedup) {
2139 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2140 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2141 			ASSERT(!zp->zp_encrypt ||
2142 			    DMU_OT_IS_ENCRYPTED(zp->zp_type));
2143 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
2144 		}
2145 		if (zp->zp_nopwrite) {
2146 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2147 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2148 			zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
2149 		}
2150 	}
2151 	return (zio);
2152 }
2153 
2154 static zio_t *
2155 zio_free_bp_init(zio_t *zio)
2156 {
2157 	blkptr_t *bp = zio->io_bp;
2158 
2159 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
2160 		if (BP_GET_DEDUP(bp))
2161 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
2162 	}
2163 
2164 	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
2165 
2166 	return (zio);
2167 }
2168 
2169 /*
2170  * ==========================================================================
2171  * Execute the I/O pipeline
2172  * ==========================================================================
2173  */
2174 
2175 static void
2176 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
2177 {
2178 	spa_t *spa = zio->io_spa;
2179 	zio_type_t t = zio->io_type;
2180 
2181 	/*
2182 	 * If we're a config writer or a probe, the normal issue and
2183 	 * interrupt threads may all be blocked waiting for the config lock.
2184 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
2185 	 */
2186 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
2187 		t = ZIO_TYPE_NULL;
2188 
2189 	/*
2190 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
2191 	 */
2192 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
2193 		t = ZIO_TYPE_NULL;
2194 
2195 	/*
2196 	 * If this is a high priority I/O, then use the high priority taskq if
2197 	 * available or cut the line otherwise.
2198 	 */
2199 	if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) {
2200 		if (spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
2201 			q++;
2202 		else
2203 			cutinline = B_TRUE;
2204 	}
2205 
2206 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
2207 
2208 	spa_taskq_dispatch(spa, t, q, zio_execute, zio, cutinline);
2209 }
2210 
2211 static boolean_t
2212 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
2213 {
2214 	spa_t *spa = zio->io_spa;
2215 
2216 	taskq_t *tq = taskq_of_curthread();
2217 
2218 	for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
2219 		spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
2220 		uint_t i;
2221 		for (i = 0; i < tqs->stqs_count; i++) {
2222 			if (tqs->stqs_taskq[i] == tq)
2223 				return (B_TRUE);
2224 		}
2225 	}
2226 
2227 	return (B_FALSE);
2228 }
2229 
2230 static zio_t *
2231 zio_issue_async(zio_t *zio)
2232 {
2233 	ASSERT((zio->io_type != ZIO_TYPE_WRITE) || ZIO_HAS_ALLOCATOR(zio));
2234 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2235 	return (NULL);
2236 }
2237 
2238 void
2239 zio_interrupt(void *zio)
2240 {
2241 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
2242 }
2243 
2244 void
2245 zio_delay_interrupt(zio_t *zio)
2246 {
2247 	/*
2248 	 * The timeout_generic() function isn't defined in userspace, so
2249 	 * rather than trying to implement the function, the zio delay
2250 	 * functionality has been disabled for userspace builds.
2251 	 */
2252 
2253 #ifdef _KERNEL
2254 	/*
2255 	 * If io_target_timestamp is zero, then no delay has been registered
2256 	 * for this IO, thus jump to the end of this function and "skip" the
2257 	 * delay; issuing it directly to the zio layer.
2258 	 */
2259 	if (zio->io_target_timestamp != 0) {
2260 		hrtime_t now = gethrtime();
2261 
2262 		if (now >= zio->io_target_timestamp) {
2263 			/*
2264 			 * This IO has already taken longer than the target
2265 			 * delay to complete, so we don't want to delay it
2266 			 * any longer; we "miss" the delay and issue it
2267 			 * directly to the zio layer. This is likely due to
2268 			 * the target latency being set to a value less than
2269 			 * the underlying hardware can satisfy (e.g. delay
2270 			 * set to 1ms, but the disks take 10ms to complete an
2271 			 * IO request).
2272 			 */
2273 
2274 			DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
2275 			    hrtime_t, now);
2276 
2277 			zio_interrupt(zio);
2278 		} else {
2279 			taskqid_t tid;
2280 			hrtime_t diff = zio->io_target_timestamp - now;
2281 			int ticks = MAX(1, NSEC_TO_TICK(diff));
2282 			clock_t expire_at_tick = ddi_get_lbolt() + ticks;
2283 
2284 			DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
2285 			    hrtime_t, now, hrtime_t, diff);
2286 
2287 			tid = taskq_dispatch_delay(system_taskq, zio_interrupt,
2288 			    zio, TQ_NOSLEEP, expire_at_tick);
2289 			if (tid == TASKQID_INVALID) {
2290 				/*
2291 				 * Couldn't allocate a task.  Just finish the
2292 				 * zio without a delay.
2293 				 */
2294 				zio_interrupt(zio);
2295 			}
2296 		}
2297 		return;
2298 	}
2299 #endif
2300 	DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
2301 	zio_interrupt(zio);
2302 }
2303 
2304 static void
2305 zio_deadman_impl(zio_t *pio, int ziodepth)
2306 {
2307 	zio_t *cio, *cio_next;
2308 	zio_link_t *zl = NULL;
2309 	vdev_t *vd = pio->io_vd;
2310 	uint64_t failmode = spa_get_deadman_failmode(pio->io_spa);
2311 
2312 	if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) {
2313 		vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL;
2314 		zbookmark_phys_t *zb = &pio->io_bookmark;
2315 		uint64_t delta = gethrtime() - pio->io_timestamp;
2316 
2317 		zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu "
2318 		    "delta=%llu queued=%llu io=%llu "
2319 		    "path=%s "
2320 		    "last=%llu type=%d "
2321 		    "priority=%d flags=0x%llx stage=0x%x "
2322 		    "pipeline=0x%x pipeline-trace=0x%x "
2323 		    "objset=%llu object=%llu "
2324 		    "level=%llu blkid=%llu "
2325 		    "offset=%llu size=%llu "
2326 		    "error=%d",
2327 		    ziodepth, pio, pio->io_timestamp,
2328 		    (u_longlong_t)delta, pio->io_delta, pio->io_delay,
2329 		    vd ? vd->vdev_path : "NULL",
2330 		    vq ? vq->vq_io_complete_ts : 0, pio->io_type,
2331 		    pio->io_priority, (u_longlong_t)pio->io_flags,
2332 		    pio->io_stage, pio->io_pipeline, pio->io_pipeline_trace,
2333 		    (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object,
2334 		    (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid,
2335 		    (u_longlong_t)pio->io_offset, (u_longlong_t)pio->io_size,
2336 		    pio->io_error);
2337 		(void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN,
2338 		    pio->io_spa, vd, zb, pio, 0);
2339 	}
2340 
2341 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2342 	    list_is_empty(&pio->io_child_list) &&
2343 	    failmode == ZIO_FAILURE_MODE_CONTINUE &&
2344 	    taskq_empty_ent(&pio->io_tqent) &&
2345 	    pio->io_queue_state == ZIO_QS_ACTIVE) {
2346 		pio->io_error = EINTR;
2347 		zio_interrupt(pio);
2348 	}
2349 
2350 	mutex_enter(&pio->io_lock);
2351 	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2352 		cio_next = zio_walk_children(pio, &zl);
2353 		zio_deadman_impl(cio, ziodepth + 1);
2354 	}
2355 	mutex_exit(&pio->io_lock);
2356 }
2357 
2358 /*
2359  * Log the critical information describing this zio and all of its children
2360  * using the zfs_dbgmsg() interface then post deadman event for the ZED.
2361  */
2362 void
2363 zio_deadman(zio_t *pio, const char *tag)
2364 {
2365 	spa_t *spa = pio->io_spa;
2366 	char *name = spa_name(spa);
2367 
2368 	if (!zfs_deadman_enabled || spa_suspended(spa))
2369 		return;
2370 
2371 	zio_deadman_impl(pio, 0);
2372 
2373 	switch (spa_get_deadman_failmode(spa)) {
2374 	case ZIO_FAILURE_MODE_WAIT:
2375 		zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name);
2376 		break;
2377 
2378 	case ZIO_FAILURE_MODE_CONTINUE:
2379 		zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name);
2380 		break;
2381 
2382 	case ZIO_FAILURE_MODE_PANIC:
2383 		fm_panic("%s determined I/O to pool '%s' is hung.", tag, name);
2384 		break;
2385 	}
2386 }
2387 
2388 /*
2389  * Execute the I/O pipeline until one of the following occurs:
2390  * (1) the I/O completes; (2) the pipeline stalls waiting for
2391  * dependent child I/Os; (3) the I/O issues, so we're waiting
2392  * for an I/O completion interrupt; (4) the I/O is delegated by
2393  * vdev-level caching or aggregation; (5) the I/O is deferred
2394  * due to vdev-level queueing; (6) the I/O is handed off to
2395  * another thread.  In all cases, the pipeline stops whenever
2396  * there's no CPU work; it never burns a thread in cv_wait_io().
2397  *
2398  * There's no locking on io_stage because there's no legitimate way
2399  * for multiple threads to be attempting to process the same I/O.
2400  */
2401 static zio_pipe_stage_t *zio_pipeline[];
2402 
2403 /*
2404  * zio_execute() is a wrapper around the static function
2405  * __zio_execute() so that we can force  __zio_execute() to be
2406  * inlined.  This reduces stack overhead which is important
2407  * because __zio_execute() is called recursively in several zio
2408  * code paths.  zio_execute() itself cannot be inlined because
2409  * it is externally visible.
2410  */
2411 void
2412 zio_execute(void *zio)
2413 {
2414 	fstrans_cookie_t cookie;
2415 
2416 	cookie = spl_fstrans_mark();
2417 	__zio_execute(zio);
2418 	spl_fstrans_unmark(cookie);
2419 }
2420 
2421 /*
2422  * Used to determine if in the current context the stack is sized large
2423  * enough to allow zio_execute() to be called recursively.  A minimum
2424  * stack size of 16K is required to avoid needing to re-dispatch the zio.
2425  */
2426 static boolean_t
2427 zio_execute_stack_check(zio_t *zio)
2428 {
2429 #if !defined(HAVE_LARGE_STACKS)
2430 	dsl_pool_t *dp = spa_get_dsl(zio->io_spa);
2431 
2432 	/* Executing in txg_sync_thread() context. */
2433 	if (dp && curthread == dp->dp_tx.tx_sync_thread)
2434 		return (B_TRUE);
2435 
2436 	/* Pool initialization outside of zio_taskq context. */
2437 	if (dp && spa_is_initializing(dp->dp_spa) &&
2438 	    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
2439 	    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))
2440 		return (B_TRUE);
2441 #else
2442 	(void) zio;
2443 #endif /* HAVE_LARGE_STACKS */
2444 
2445 	return (B_FALSE);
2446 }
2447 
2448 __attribute__((always_inline))
2449 static inline void
2450 __zio_execute(zio_t *zio)
2451 {
2452 	ASSERT3U(zio->io_queued_timestamp, >, 0);
2453 
2454 	while (zio->io_stage < ZIO_STAGE_DONE) {
2455 		enum zio_stage pipeline = zio->io_pipeline;
2456 		enum zio_stage stage = zio->io_stage;
2457 
2458 		zio->io_executor = curthread;
2459 
2460 		ASSERT(!MUTEX_HELD(&zio->io_lock));
2461 		ASSERT(ISP2(stage));
2462 		ASSERT(zio->io_stall == NULL);
2463 
2464 		do {
2465 			stage <<= 1;
2466 		} while ((stage & pipeline) == 0);
2467 
2468 		ASSERT(stage <= ZIO_STAGE_DONE);
2469 
2470 		/*
2471 		 * If we are in interrupt context and this pipeline stage
2472 		 * will grab a config lock that is held across I/O,
2473 		 * or may wait for an I/O that needs an interrupt thread
2474 		 * to complete, issue async to avoid deadlock.
2475 		 *
2476 		 * For VDEV_IO_START, we cut in line so that the io will
2477 		 * be sent to disk promptly.
2478 		 */
2479 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
2480 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
2481 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2482 			    zio_requeue_io_start_cut_in_line : B_FALSE;
2483 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2484 			return;
2485 		}
2486 
2487 		/*
2488 		 * If the current context doesn't have large enough stacks
2489 		 * the zio must be issued asynchronously to prevent overflow.
2490 		 */
2491 		if (zio_execute_stack_check(zio)) {
2492 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2493 			    zio_requeue_io_start_cut_in_line : B_FALSE;
2494 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2495 			return;
2496 		}
2497 
2498 		zio->io_stage = stage;
2499 		zio->io_pipeline_trace |= zio->io_stage;
2500 
2501 		/*
2502 		 * The zio pipeline stage returns the next zio to execute
2503 		 * (typically the same as this one), or NULL if we should
2504 		 * stop.
2505 		 */
2506 		zio = zio_pipeline[highbit64(stage) - 1](zio);
2507 
2508 		if (zio == NULL)
2509 			return;
2510 	}
2511 }
2512 
2513 
2514 /*
2515  * ==========================================================================
2516  * Initiate I/O, either sync or async
2517  * ==========================================================================
2518  */
2519 int
2520 zio_wait(zio_t *zio)
2521 {
2522 	/*
2523 	 * Some routines, like zio_free_sync(), may return a NULL zio
2524 	 * to avoid the performance overhead of creating and then destroying
2525 	 * an unneeded zio.  For the callers' simplicity, we accept a NULL
2526 	 * zio and ignore it.
2527 	 */
2528 	if (zio == NULL)
2529 		return (0);
2530 
2531 	long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms);
2532 	int error;
2533 
2534 	ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN);
2535 	ASSERT3P(zio->io_executor, ==, NULL);
2536 
2537 	zio->io_waiter = curthread;
2538 	ASSERT0(zio->io_queued_timestamp);
2539 	zio->io_queued_timestamp = gethrtime();
2540 
2541 	if (zio->io_type == ZIO_TYPE_WRITE) {
2542 		spa_select_allocator(zio);
2543 	}
2544 	__zio_execute(zio);
2545 
2546 	mutex_enter(&zio->io_lock);
2547 	while (zio->io_executor != NULL) {
2548 		error = cv_timedwait_io(&zio->io_cv, &zio->io_lock,
2549 		    ddi_get_lbolt() + timeout);
2550 
2551 		if (zfs_deadman_enabled && error == -1 &&
2552 		    gethrtime() - zio->io_queued_timestamp >
2553 		    spa_deadman_ziotime(zio->io_spa)) {
2554 			mutex_exit(&zio->io_lock);
2555 			timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms);
2556 			zio_deadman(zio, FTAG);
2557 			mutex_enter(&zio->io_lock);
2558 		}
2559 	}
2560 	mutex_exit(&zio->io_lock);
2561 
2562 	error = zio->io_error;
2563 	zio_destroy(zio);
2564 
2565 	return (error);
2566 }
2567 
2568 void
2569 zio_nowait(zio_t *zio)
2570 {
2571 	/*
2572 	 * See comment in zio_wait().
2573 	 */
2574 	if (zio == NULL)
2575 		return;
2576 
2577 	ASSERT3P(zio->io_executor, ==, NULL);
2578 
2579 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
2580 	    list_is_empty(&zio->io_parent_list)) {
2581 		zio_t *pio;
2582 
2583 		/*
2584 		 * This is a logical async I/O with no parent to wait for it.
2585 		 * We add it to the spa_async_root_zio "Godfather" I/O which
2586 		 * will ensure they complete prior to unloading the pool.
2587 		 */
2588 		spa_t *spa = zio->io_spa;
2589 		pio = spa->spa_async_zio_root[CPU_SEQID_UNSTABLE];
2590 
2591 		zio_add_child(pio, zio);
2592 	}
2593 
2594 	ASSERT0(zio->io_queued_timestamp);
2595 	zio->io_queued_timestamp = gethrtime();
2596 	if (zio->io_type == ZIO_TYPE_WRITE) {
2597 		spa_select_allocator(zio);
2598 	}
2599 	__zio_execute(zio);
2600 }
2601 
2602 /*
2603  * ==========================================================================
2604  * Reexecute, cancel, or suspend/resume failed I/O
2605  * ==========================================================================
2606  */
2607 
2608 static void
2609 zio_reexecute(void *arg)
2610 {
2611 	zio_t *pio = arg;
2612 	zio_t *cio, *cio_next, *gio;
2613 
2614 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
2615 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
2616 	ASSERT(pio->io_gang_leader == NULL);
2617 	ASSERT(pio->io_gang_tree == NULL);
2618 
2619 	mutex_enter(&pio->io_lock);
2620 	pio->io_flags = pio->io_orig_flags;
2621 	pio->io_stage = pio->io_orig_stage;
2622 	pio->io_pipeline = pio->io_orig_pipeline;
2623 	pio->io_reexecute = 0;
2624 	pio->io_flags |= ZIO_FLAG_REEXECUTED;
2625 	pio->io_pipeline_trace = 0;
2626 	pio->io_error = 0;
2627 	pio->io_state[ZIO_WAIT_READY] = (pio->io_stage >= ZIO_STAGE_READY) ||
2628 	    (pio->io_pipeline & ZIO_STAGE_READY) == 0;
2629 	pio->io_state[ZIO_WAIT_DONE] = (pio->io_stage >= ZIO_STAGE_DONE);
2630 
2631 	/*
2632 	 * It's possible for a failed ZIO to be a descendant of more than one
2633 	 * ZIO tree. When reexecuting it, we have to be sure to add its wait
2634 	 * states to all parent wait counts.
2635 	 *
2636 	 * Those parents, in turn, may have other children that are currently
2637 	 * active, usually because they've already been reexecuted after
2638 	 * resuming. Those children may be executing and may call
2639 	 * zio_notify_parent() at the same time as we're updating our parent's
2640 	 * counts. To avoid races while updating the counts, we take
2641 	 * gio->io_lock before each update.
2642 	 */
2643 	zio_link_t *zl = NULL;
2644 	while ((gio = zio_walk_parents(pio, &zl)) != NULL) {
2645 		mutex_enter(&gio->io_lock);
2646 		for (int w = 0; w < ZIO_WAIT_TYPES; w++) {
2647 			gio->io_children[pio->io_child_type][w] +=
2648 			    !pio->io_state[w];
2649 		}
2650 		mutex_exit(&gio->io_lock);
2651 	}
2652 
2653 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2654 		pio->io_child_error[c] = 0;
2655 
2656 	if (IO_IS_ALLOCATING(pio))
2657 		BP_ZERO(pio->io_bp);
2658 
2659 	/*
2660 	 * As we reexecute pio's children, new children could be created.
2661 	 * New children go to the head of pio's io_child_list, however,
2662 	 * so we will (correctly) not reexecute them.  The key is that
2663 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
2664 	 * cannot be affected by any side effects of reexecuting 'cio'.
2665 	 */
2666 	zl = NULL;
2667 	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2668 		cio_next = zio_walk_children(pio, &zl);
2669 		mutex_exit(&pio->io_lock);
2670 		zio_reexecute(cio);
2671 		mutex_enter(&pio->io_lock);
2672 	}
2673 	mutex_exit(&pio->io_lock);
2674 
2675 	/*
2676 	 * Now that all children have been reexecuted, execute the parent.
2677 	 * We don't reexecute "The Godfather" I/O here as it's the
2678 	 * responsibility of the caller to wait on it.
2679 	 */
2680 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
2681 		pio->io_queued_timestamp = gethrtime();
2682 		__zio_execute(pio);
2683 	}
2684 }
2685 
2686 void
2687 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
2688 {
2689 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
2690 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2691 		    "failure and the failure mode property for this pool "
2692 		    "is set to panic.", spa_name(spa));
2693 
2694 	if (reason != ZIO_SUSPEND_MMP) {
2695 		cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable "
2696 		    "I/O failure and has been suspended.", spa_name(spa));
2697 	}
2698 
2699 	(void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
2700 	    NULL, NULL, 0);
2701 
2702 	mutex_enter(&spa->spa_suspend_lock);
2703 
2704 	if (spa->spa_suspend_zio_root == NULL)
2705 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
2706 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2707 		    ZIO_FLAG_GODFATHER);
2708 
2709 	spa->spa_suspended = reason;
2710 
2711 	if (zio != NULL) {
2712 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2713 		ASSERT(zio != spa->spa_suspend_zio_root);
2714 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2715 		ASSERT(zio_unique_parent(zio) == NULL);
2716 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
2717 		zio_add_child(spa->spa_suspend_zio_root, zio);
2718 	}
2719 
2720 	mutex_exit(&spa->spa_suspend_lock);
2721 
2722 	txg_wait_kick(spa->spa_dsl_pool);
2723 }
2724 
2725 int
2726 zio_resume(spa_t *spa)
2727 {
2728 	zio_t *pio;
2729 
2730 	/*
2731 	 * Reexecute all previously suspended i/o.
2732 	 */
2733 	mutex_enter(&spa->spa_suspend_lock);
2734 	if (spa->spa_suspended != ZIO_SUSPEND_NONE)
2735 		cmn_err(CE_WARN, "Pool '%s' was suspended and is being "
2736 		    "resumed. Failed I/O will be retried.",
2737 		    spa_name(spa));
2738 	spa->spa_suspended = ZIO_SUSPEND_NONE;
2739 	cv_broadcast(&spa->spa_suspend_cv);
2740 	pio = spa->spa_suspend_zio_root;
2741 	spa->spa_suspend_zio_root = NULL;
2742 	mutex_exit(&spa->spa_suspend_lock);
2743 
2744 	if (pio == NULL)
2745 		return (0);
2746 
2747 	zio_reexecute(pio);
2748 	return (zio_wait(pio));
2749 }
2750 
2751 void
2752 zio_resume_wait(spa_t *spa)
2753 {
2754 	mutex_enter(&spa->spa_suspend_lock);
2755 	while (spa_suspended(spa))
2756 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
2757 	mutex_exit(&spa->spa_suspend_lock);
2758 }
2759 
2760 /*
2761  * ==========================================================================
2762  * Gang blocks.
2763  *
2764  * A gang block is a collection of small blocks that looks to the DMU
2765  * like one large block.  When zio_dva_allocate() cannot find a block
2766  * of the requested size, due to either severe fragmentation or the pool
2767  * being nearly full, it calls zio_write_gang_block() to construct the
2768  * block from smaller fragments.
2769  *
2770  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2771  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
2772  * an indirect block: it's an array of block pointers.  It consumes
2773  * only one sector and hence is allocatable regardless of fragmentation.
2774  * The gang header's bps point to its gang members, which hold the data.
2775  *
2776  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2777  * as the verifier to ensure uniqueness of the SHA256 checksum.
2778  * Critically, the gang block bp's blk_cksum is the checksum of the data,
2779  * not the gang header.  This ensures that data block signatures (needed for
2780  * deduplication) are independent of how the block is physically stored.
2781  *
2782  * Gang blocks can be nested: a gang member may itself be a gang block.
2783  * Thus every gang block is a tree in which root and all interior nodes are
2784  * gang headers, and the leaves are normal blocks that contain user data.
2785  * The root of the gang tree is called the gang leader.
2786  *
2787  * To perform any operation (read, rewrite, free, claim) on a gang block,
2788  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2789  * in the io_gang_tree field of the original logical i/o by recursively
2790  * reading the gang leader and all gang headers below it.  This yields
2791  * an in-core tree containing the contents of every gang header and the
2792  * bps for every constituent of the gang block.
2793  *
2794  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2795  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
2796  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2797  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2798  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2799  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
2800  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2801  * of the gang header plus zio_checksum_compute() of the data to update the
2802  * gang header's blk_cksum as described above.
2803  *
2804  * The two-phase assemble/issue model solves the problem of partial failure --
2805  * what if you'd freed part of a gang block but then couldn't read the
2806  * gang header for another part?  Assembling the entire gang tree first
2807  * ensures that all the necessary gang header I/O has succeeded before
2808  * starting the actual work of free, claim, or write.  Once the gang tree
2809  * is assembled, free and claim are in-memory operations that cannot fail.
2810  *
2811  * In the event that a gang write fails, zio_dva_unallocate() walks the
2812  * gang tree to immediately free (i.e. insert back into the space map)
2813  * everything we've allocated.  This ensures that we don't get ENOSPC
2814  * errors during repeated suspend/resume cycles due to a flaky device.
2815  *
2816  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
2817  * the gang tree, we won't modify the block, so we can safely defer the free
2818  * (knowing that the block is still intact).  If we *can* assemble the gang
2819  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2820  * each constituent bp and we can allocate a new block on the next sync pass.
2821  *
2822  * In all cases, the gang tree allows complete recovery from partial failure.
2823  * ==========================================================================
2824  */
2825 
2826 static void
2827 zio_gang_issue_func_done(zio_t *zio)
2828 {
2829 	abd_free(zio->io_abd);
2830 }
2831 
2832 static zio_t *
2833 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2834     uint64_t offset)
2835 {
2836 	if (gn != NULL)
2837 		return (pio);
2838 
2839 	return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2840 	    BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2841 	    NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2842 	    &pio->io_bookmark));
2843 }
2844 
2845 static zio_t *
2846 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2847     uint64_t offset)
2848 {
2849 	zio_t *zio;
2850 
2851 	if (gn != NULL) {
2852 		abd_t *gbh_abd =
2853 		    abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2854 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2855 		    gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2856 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2857 		    &pio->io_bookmark);
2858 		/*
2859 		 * As we rewrite each gang header, the pipeline will compute
2860 		 * a new gang block header checksum for it; but no one will
2861 		 * compute a new data checksum, so we do that here.  The one
2862 		 * exception is the gang leader: the pipeline already computed
2863 		 * its data checksum because that stage precedes gang assembly.
2864 		 * (Presently, nothing actually uses interior data checksums;
2865 		 * this is just good hygiene.)
2866 		 */
2867 		if (gn != pio->io_gang_leader->io_gang_tree) {
2868 			abd_t *buf = abd_get_offset(data, offset);
2869 
2870 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2871 			    buf, BP_GET_PSIZE(bp));
2872 
2873 			abd_free(buf);
2874 		}
2875 		/*
2876 		 * If we are here to damage data for testing purposes,
2877 		 * leave the GBH alone so that we can detect the damage.
2878 		 */
2879 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2880 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2881 	} else {
2882 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2883 		    abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2884 		    zio_gang_issue_func_done, NULL, pio->io_priority,
2885 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2886 	}
2887 
2888 	return (zio);
2889 }
2890 
2891 static zio_t *
2892 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2893     uint64_t offset)
2894 {
2895 	(void) gn, (void) data, (void) offset;
2896 
2897 	zio_t *zio = zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2898 	    ZIO_GANG_CHILD_FLAGS(pio));
2899 	if (zio == NULL) {
2900 		zio = zio_null(pio, pio->io_spa,
2901 		    NULL, NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio));
2902 	}
2903 	return (zio);
2904 }
2905 
2906 static zio_t *
2907 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2908     uint64_t offset)
2909 {
2910 	(void) gn, (void) data, (void) offset;
2911 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2912 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2913 }
2914 
2915 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2916 	NULL,
2917 	zio_read_gang,
2918 	zio_rewrite_gang,
2919 	zio_free_gang,
2920 	zio_claim_gang,
2921 	NULL
2922 };
2923 
2924 static void zio_gang_tree_assemble_done(zio_t *zio);
2925 
2926 static zio_gang_node_t *
2927 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2928 {
2929 	zio_gang_node_t *gn;
2930 
2931 	ASSERT(*gnpp == NULL);
2932 
2933 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2934 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2935 	*gnpp = gn;
2936 
2937 	return (gn);
2938 }
2939 
2940 static void
2941 zio_gang_node_free(zio_gang_node_t **gnpp)
2942 {
2943 	zio_gang_node_t *gn = *gnpp;
2944 
2945 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2946 		ASSERT(gn->gn_child[g] == NULL);
2947 
2948 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2949 	kmem_free(gn, sizeof (*gn));
2950 	*gnpp = NULL;
2951 }
2952 
2953 static void
2954 zio_gang_tree_free(zio_gang_node_t **gnpp)
2955 {
2956 	zio_gang_node_t *gn = *gnpp;
2957 
2958 	if (gn == NULL)
2959 		return;
2960 
2961 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2962 		zio_gang_tree_free(&gn->gn_child[g]);
2963 
2964 	zio_gang_node_free(gnpp);
2965 }
2966 
2967 static void
2968 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2969 {
2970 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2971 	abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2972 
2973 	ASSERT(gio->io_gang_leader == gio);
2974 	ASSERT(BP_IS_GANG(bp));
2975 
2976 	zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2977 	    zio_gang_tree_assemble_done, gn, gio->io_priority,
2978 	    ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2979 }
2980 
2981 static void
2982 zio_gang_tree_assemble_done(zio_t *zio)
2983 {
2984 	zio_t *gio = zio->io_gang_leader;
2985 	zio_gang_node_t *gn = zio->io_private;
2986 	blkptr_t *bp = zio->io_bp;
2987 
2988 	ASSERT(gio == zio_unique_parent(zio));
2989 	ASSERT(list_is_empty(&zio->io_child_list));
2990 
2991 	if (zio->io_error)
2992 		return;
2993 
2994 	/* this ABD was created from a linear buf in zio_gang_tree_assemble */
2995 	if (BP_SHOULD_BYTESWAP(bp))
2996 		byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2997 
2998 	ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2999 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
3000 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
3001 
3002 	abd_free(zio->io_abd);
3003 
3004 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3005 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
3006 		if (!BP_IS_GANG(gbp))
3007 			continue;
3008 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
3009 	}
3010 }
3011 
3012 static void
3013 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
3014     uint64_t offset)
3015 {
3016 	zio_t *gio = pio->io_gang_leader;
3017 	zio_t *zio;
3018 
3019 	ASSERT(BP_IS_GANG(bp) == !!gn);
3020 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
3021 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
3022 
3023 	/*
3024 	 * If you're a gang header, your data is in gn->gn_gbh.
3025 	 * If you're a gang member, your data is in 'data' and gn == NULL.
3026 	 */
3027 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
3028 
3029 	if (gn != NULL) {
3030 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
3031 
3032 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3033 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
3034 			if (BP_IS_HOLE(gbp))
3035 				continue;
3036 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
3037 			    offset);
3038 			offset += BP_GET_PSIZE(gbp);
3039 		}
3040 	}
3041 
3042 	if (gn == gio->io_gang_tree)
3043 		ASSERT3U(gio->io_size, ==, offset);
3044 
3045 	if (zio != pio)
3046 		zio_nowait(zio);
3047 }
3048 
3049 static zio_t *
3050 zio_gang_assemble(zio_t *zio)
3051 {
3052 	blkptr_t *bp = zio->io_bp;
3053 
3054 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
3055 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3056 
3057 	zio->io_gang_leader = zio;
3058 
3059 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
3060 
3061 	return (zio);
3062 }
3063 
3064 static zio_t *
3065 zio_gang_issue(zio_t *zio)
3066 {
3067 	blkptr_t *bp = zio->io_bp;
3068 
3069 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
3070 		return (NULL);
3071 	}
3072 
3073 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
3074 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3075 
3076 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
3077 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
3078 		    0);
3079 	else
3080 		zio_gang_tree_free(&zio->io_gang_tree);
3081 
3082 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3083 
3084 	return (zio);
3085 }
3086 
3087 static void
3088 zio_gang_inherit_allocator(zio_t *pio, zio_t *cio)
3089 {
3090 	cio->io_allocator = pio->io_allocator;
3091 }
3092 
3093 static void
3094 zio_write_gang_member_ready(zio_t *zio)
3095 {
3096 	zio_t *pio = zio_unique_parent(zio);
3097 	dva_t *cdva = zio->io_bp->blk_dva;
3098 	dva_t *pdva = pio->io_bp->blk_dva;
3099 	uint64_t asize;
3100 	zio_t *gio __maybe_unused = zio->io_gang_leader;
3101 
3102 	if (BP_IS_HOLE(zio->io_bp))
3103 		return;
3104 
3105 	/*
3106 	 * If we're getting direct-invoked from zio_write_gang_block(),
3107 	 * the bp_orig will be set.
3108 	 */
3109 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig) ||
3110 	    zio->io_flags & ZIO_FLAG_PREALLOCATED);
3111 
3112 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
3113 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
3114 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
3115 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
3116 	VERIFY3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
3117 
3118 	mutex_enter(&pio->io_lock);
3119 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
3120 		ASSERT(DVA_GET_GANG(&pdva[d]));
3121 		asize = DVA_GET_ASIZE(&pdva[d]);
3122 		asize += DVA_GET_ASIZE(&cdva[d]);
3123 		DVA_SET_ASIZE(&pdva[d], asize);
3124 	}
3125 	mutex_exit(&pio->io_lock);
3126 }
3127 
3128 static void
3129 zio_write_gang_done(zio_t *zio)
3130 {
3131 	/*
3132 	 * The io_abd field will be NULL for a zio with no data.  The io_flags
3133 	 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
3134 	 * check for it here as it is cleared in zio_ready.
3135 	 */
3136 	if (zio->io_abd != NULL)
3137 		abd_free(zio->io_abd);
3138 }
3139 
3140 static zio_t *
3141 zio_write_gang_block(zio_t *pio, metaslab_class_t *mc)
3142 {
3143 	spa_t *spa = pio->io_spa;
3144 	blkptr_t *bp = pio->io_bp;
3145 	zio_t *gio = pio->io_gang_leader;
3146 	zio_t *zio;
3147 	zio_gang_node_t *gn, **gnpp;
3148 	zio_gbh_phys_t *gbh;
3149 	abd_t *gbh_abd;
3150 	uint64_t txg = pio->io_txg;
3151 	uint64_t resid = pio->io_size;
3152 	zio_prop_t zp;
3153 	int error;
3154 	boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
3155 
3156 	/*
3157 	 * Store multiple copies of the GBH, so that we can still traverse
3158 	 * all the data (e.g. to free or scrub) even if a block is damaged.
3159 	 * This value respects the redundant_metadata property.
3160 	 */
3161 	int gbh_copies = gio->io_prop.zp_gang_copies;
3162 	if (gbh_copies == 0) {
3163 		/*
3164 		 * This should only happen in the case where we're filling in
3165 		 * DDT entries for a parent that wants more copies than the DDT
3166 		 * has.  In that case, we cannot gang without creating a mixed
3167 		 * blkptr, which is illegal.
3168 		 */
3169 		ASSERT3U(gio->io_child_type, ==, ZIO_CHILD_DDT);
3170 		pio->io_error = EAGAIN;
3171 		return (pio);
3172 	}
3173 	ASSERT3S(gbh_copies, >, 0);
3174 	ASSERT3S(gbh_copies, <=, SPA_DVAS_PER_BP);
3175 
3176 	ASSERT(ZIO_HAS_ALLOCATOR(pio));
3177 	int flags = METASLAB_GANG_HEADER;
3178 	if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3179 		ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3180 		ASSERT(has_data);
3181 
3182 		flags |= METASLAB_ASYNC_ALLOC;
3183 	}
3184 
3185 	error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
3186 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
3187 	    &pio->io_alloc_list, pio->io_allocator, pio);
3188 	if (error) {
3189 		pio->io_error = error;
3190 		return (pio);
3191 	}
3192 
3193 	if (pio == gio) {
3194 		gnpp = &gio->io_gang_tree;
3195 	} else {
3196 		gnpp = pio->io_private;
3197 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
3198 	}
3199 
3200 	gn = zio_gang_node_alloc(gnpp);
3201 	gbh = gn->gn_gbh;
3202 	memset(gbh, 0, SPA_GANGBLOCKSIZE);
3203 	gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
3204 
3205 	/*
3206 	 * Create the gang header.
3207 	 */
3208 	zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
3209 	    zio_write_gang_done, NULL, pio->io_priority,
3210 	    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
3211 
3212 	zio_gang_inherit_allocator(pio, zio);
3213 	if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3214 		boolean_t more;
3215 		VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies,
3216 		    zio, B_TRUE, &more));
3217 	}
3218 
3219 	/*
3220 	 * Create and nowait the gang children. First, we try to do
3221 	 * opportunistic allocations. If that fails to generate enough
3222 	 * space, we fall back to normal zio_write calls for nested gang.
3223 	 */
3224 	for (int g = 0; resid != 0; g++) {
3225 		flags &= METASLAB_ASYNC_ALLOC;
3226 		flags |= METASLAB_GANG_CHILD;
3227 		zp.zp_checksum = gio->io_prop.zp_checksum;
3228 		zp.zp_compress = ZIO_COMPRESS_OFF;
3229 		zp.zp_complevel = gio->io_prop.zp_complevel;
3230 		zp.zp_type = zp.zp_storage_type = DMU_OT_NONE;
3231 		zp.zp_level = 0;
3232 		zp.zp_copies = gio->io_prop.zp_copies;
3233 		zp.zp_gang_copies = gio->io_prop.zp_gang_copies;
3234 		zp.zp_dedup = B_FALSE;
3235 		zp.zp_dedup_verify = B_FALSE;
3236 		zp.zp_nopwrite = B_FALSE;
3237 		zp.zp_encrypt = gio->io_prop.zp_encrypt;
3238 		zp.zp_byteorder = gio->io_prop.zp_byteorder;
3239 		zp.zp_direct_write = B_FALSE;
3240 		memset(zp.zp_salt, 0, ZIO_DATA_SALT_LEN);
3241 		memset(zp.zp_iv, 0, ZIO_DATA_IV_LEN);
3242 		memset(zp.zp_mac, 0, ZIO_DATA_MAC_LEN);
3243 
3244 		uint64_t min_size = zio_roundup_alloc_size(spa,
3245 		    resid / (SPA_GBH_NBLKPTRS - g));
3246 		min_size = MIN(min_size, resid);
3247 		bp = &gbh->zg_blkptr[g];
3248 
3249 		zio_alloc_list_t cio_list;
3250 		metaslab_trace_init(&cio_list);
3251 		uint64_t allocated_size = UINT64_MAX;
3252 		error = metaslab_alloc_range(spa, mc, min_size, resid,
3253 		    bp, gio->io_prop.zp_copies, txg, NULL,
3254 		    flags, &cio_list, zio->io_allocator, NULL, &allocated_size);
3255 
3256 		boolean_t allocated = error == 0;
3257 
3258 		uint64_t psize = allocated ? MIN(resid, allocated_size) :
3259 		    min_size;
3260 
3261 		zio_t *cio = zio_write(zio, spa, txg, bp, has_data ?
3262 		    abd_get_offset(pio->io_abd, pio->io_size - resid) : NULL,
3263 		    psize, psize, &zp, zio_write_gang_member_ready, NULL,
3264 		    zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
3265 		    ZIO_GANG_CHILD_FLAGS(pio) |
3266 		    (allocated ? ZIO_FLAG_PREALLOCATED : 0), &pio->io_bookmark);
3267 
3268 		resid -= psize;
3269 		zio_gang_inherit_allocator(zio, cio);
3270 		if (allocated) {
3271 			metaslab_trace_move(&cio_list, &cio->io_alloc_list);
3272 			metaslab_group_alloc_increment_all(spa,
3273 			    &cio->io_bp_orig, zio->io_allocator, flags, psize,
3274 			    cio);
3275 		}
3276 		/*
3277 		 * We do not reserve for the child writes, since we already
3278 		 * reserved for the parent.  Unreserve though will be called
3279 		 * for individual children.  We can do this since sum of all
3280 		 * child's physical sizes is equal to parent's physical size.
3281 		 * It would not work for potentially bigger allocation sizes.
3282 		 */
3283 
3284 		zio_nowait(cio);
3285 	}
3286 
3287 	/*
3288 	 * Set pio's pipeline to just wait for zio to finish.
3289 	 */
3290 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3291 
3292 	zio_nowait(zio);
3293 
3294 	return (pio);
3295 }
3296 
3297 /*
3298  * The zio_nop_write stage in the pipeline determines if allocating a
3299  * new bp is necessary.  The nopwrite feature can handle writes in
3300  * either syncing or open context (i.e. zil writes) and as a result is
3301  * mutually exclusive with dedup.
3302  *
3303  * By leveraging a cryptographically secure checksum, such as SHA256, we
3304  * can compare the checksums of the new data and the old to determine if
3305  * allocating a new block is required.  Note that our requirements for
3306  * cryptographic strength are fairly weak: there can't be any accidental
3307  * hash collisions, but we don't need to be secure against intentional
3308  * (malicious) collisions.  To trigger a nopwrite, you have to be able
3309  * to write the file to begin with, and triggering an incorrect (hash
3310  * collision) nopwrite is no worse than simply writing to the file.
3311  * That said, there are no known attacks against the checksum algorithms
3312  * used for nopwrite, assuming that the salt and the checksums
3313  * themselves remain secret.
3314  */
3315 static zio_t *
3316 zio_nop_write(zio_t *zio)
3317 {
3318 	blkptr_t *bp = zio->io_bp;
3319 	blkptr_t *bp_orig = &zio->io_bp_orig;
3320 	zio_prop_t *zp = &zio->io_prop;
3321 
3322 	ASSERT(BP_IS_HOLE(bp));
3323 	ASSERT(BP_GET_LEVEL(bp) == 0);
3324 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
3325 	ASSERT(zp->zp_nopwrite);
3326 	ASSERT(!zp->zp_dedup);
3327 	ASSERT(zio->io_bp_override == NULL);
3328 	ASSERT(IO_IS_ALLOCATING(zio));
3329 
3330 	/*
3331 	 * Check to see if the original bp and the new bp have matching
3332 	 * characteristics (i.e. same checksum, compression algorithms, etc).
3333 	 * If they don't then just continue with the pipeline which will
3334 	 * allocate a new bp.
3335 	 */
3336 	if (BP_IS_HOLE(bp_orig) ||
3337 	    !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
3338 	    ZCHECKSUM_FLAG_NOPWRITE) ||
3339 	    BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) ||
3340 	    BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
3341 	    BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
3342 	    BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
3343 	    zp->zp_copies != BP_GET_NDVAS(bp_orig))
3344 		return (zio);
3345 
3346 	/*
3347 	 * If the checksums match then reset the pipeline so that we
3348 	 * avoid allocating a new bp and issuing any I/O.
3349 	 */
3350 	if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
3351 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
3352 		    ZCHECKSUM_FLAG_NOPWRITE);
3353 		ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
3354 		ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
3355 		ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
3356 		ASSERT3U(bp->blk_prop, ==, bp_orig->blk_prop);
3357 
3358 		/*
3359 		 * If we're overwriting a block that is currently on an
3360 		 * indirect vdev, then ignore the nopwrite request and
3361 		 * allow a new block to be allocated on a concrete vdev.
3362 		 */
3363 		spa_config_enter(zio->io_spa, SCL_VDEV, FTAG, RW_READER);
3364 		for (int d = 0; d < BP_GET_NDVAS(bp_orig); d++) {
3365 			vdev_t *tvd = vdev_lookup_top(zio->io_spa,
3366 			    DVA_GET_VDEV(&bp_orig->blk_dva[d]));
3367 			if (tvd->vdev_ops == &vdev_indirect_ops) {
3368 				spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3369 				return (zio);
3370 			}
3371 		}
3372 		spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3373 
3374 		*bp = *bp_orig;
3375 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3376 		zio->io_flags |= ZIO_FLAG_NOPWRITE;
3377 	}
3378 
3379 	return (zio);
3380 }
3381 
3382 /*
3383  * ==========================================================================
3384  * Block Reference Table
3385  * ==========================================================================
3386  */
3387 static zio_t *
3388 zio_brt_free(zio_t *zio)
3389 {
3390 	blkptr_t *bp;
3391 
3392 	bp = zio->io_bp;
3393 
3394 	if (BP_GET_LEVEL(bp) > 0 ||
3395 	    BP_IS_METADATA(bp) ||
3396 	    !brt_maybe_exists(zio->io_spa, bp)) {
3397 		return (zio);
3398 	}
3399 
3400 	if (!brt_entry_decref(zio->io_spa, bp)) {
3401 		/*
3402 		 * This isn't the last reference, so we cannot free
3403 		 * the data yet.
3404 		 */
3405 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3406 	}
3407 
3408 	return (zio);
3409 }
3410 
3411 /*
3412  * ==========================================================================
3413  * Dedup
3414  * ==========================================================================
3415  */
3416 static void
3417 zio_ddt_child_read_done(zio_t *zio)
3418 {
3419 	blkptr_t *bp = zio->io_bp;
3420 	ddt_t *ddt;
3421 	ddt_entry_t *dde = zio->io_private;
3422 	zio_t *pio = zio_unique_parent(zio);
3423 
3424 	mutex_enter(&pio->io_lock);
3425 	ddt = ddt_select(zio->io_spa, bp);
3426 
3427 	if (zio->io_error == 0) {
3428 		ddt_phys_variant_t v = ddt_phys_select(ddt, dde, bp);
3429 		/* this phys variant doesn't need repair */
3430 		ddt_phys_clear(dde->dde_phys, v);
3431 	}
3432 
3433 	if (zio->io_error == 0 && dde->dde_io->dde_repair_abd == NULL)
3434 		dde->dde_io->dde_repair_abd = zio->io_abd;
3435 	else
3436 		abd_free(zio->io_abd);
3437 	mutex_exit(&pio->io_lock);
3438 }
3439 
3440 static zio_t *
3441 zio_ddt_read_start(zio_t *zio)
3442 {
3443 	blkptr_t *bp = zio->io_bp;
3444 
3445 	ASSERT(BP_GET_DEDUP(bp));
3446 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3447 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3448 
3449 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
3450 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
3451 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
3452 		ddt_phys_variant_t v_self = ddt_phys_select(ddt, dde, bp);
3453 		ddt_univ_phys_t *ddp = dde->dde_phys;
3454 		blkptr_t blk;
3455 
3456 		ASSERT(zio->io_vsd == NULL);
3457 		zio->io_vsd = dde;
3458 
3459 		if (v_self == DDT_PHYS_NONE)
3460 			return (zio);
3461 
3462 		/* issue I/O for the other copies */
3463 		for (int p = 0; p < DDT_NPHYS(ddt); p++) {
3464 			ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p);
3465 
3466 			if (ddt_phys_birth(ddp, v) == 0 || v == v_self)
3467 				continue;
3468 
3469 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key,
3470 			    ddp, v, &blk);
3471 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
3472 			    abd_alloc_for_io(zio->io_size, B_TRUE),
3473 			    zio->io_size, zio_ddt_child_read_done, dde,
3474 			    zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
3475 			    ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
3476 		}
3477 		return (zio);
3478 	}
3479 
3480 	zio_nowait(zio_read(zio, zio->io_spa, bp,
3481 	    zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
3482 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
3483 
3484 	return (zio);
3485 }
3486 
3487 static zio_t *
3488 zio_ddt_read_done(zio_t *zio)
3489 {
3490 	blkptr_t *bp = zio->io_bp;
3491 
3492 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
3493 		return (NULL);
3494 	}
3495 
3496 	ASSERT(BP_GET_DEDUP(bp));
3497 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3498 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3499 
3500 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
3501 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
3502 		ddt_entry_t *dde = zio->io_vsd;
3503 		if (ddt == NULL) {
3504 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
3505 			return (zio);
3506 		}
3507 		if (dde == NULL) {
3508 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
3509 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
3510 			return (NULL);
3511 		}
3512 		if (dde->dde_io->dde_repair_abd != NULL) {
3513 			abd_copy(zio->io_abd, dde->dde_io->dde_repair_abd,
3514 			    zio->io_size);
3515 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
3516 		}
3517 		ddt_repair_done(ddt, dde);
3518 		zio->io_vsd = NULL;
3519 	}
3520 
3521 	ASSERT(zio->io_vsd == NULL);
3522 
3523 	return (zio);
3524 }
3525 
3526 static boolean_t
3527 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
3528 {
3529 	spa_t *spa = zio->io_spa;
3530 	boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW);
3531 
3532 	ASSERT(!(zio->io_bp_override && do_raw));
3533 
3534 	/*
3535 	 * Note: we compare the original data, not the transformed data,
3536 	 * because when zio->io_bp is an override bp, we will not have
3537 	 * pushed the I/O transforms.  That's an important optimization
3538 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
3539 	 * However, we should never get a raw, override zio so in these
3540 	 * cases we can compare the io_abd directly. This is useful because
3541 	 * it allows us to do dedup verification even if we don't have access
3542 	 * to the original data (for instance, if the encryption keys aren't
3543 	 * loaded).
3544 	 */
3545 
3546 	for (int p = 0; p < DDT_NPHYS(ddt); p++) {
3547 		if (DDT_PHYS_IS_DITTO(ddt, p))
3548 			continue;
3549 
3550 		if (dde->dde_io == NULL)
3551 			continue;
3552 
3553 		zio_t *lio = dde->dde_io->dde_lead_zio[p];
3554 		if (lio == NULL)
3555 			continue;
3556 
3557 		if (do_raw)
3558 			return (lio->io_size != zio->io_size ||
3559 			    abd_cmp(zio->io_abd, lio->io_abd) != 0);
3560 
3561 		return (lio->io_orig_size != zio->io_orig_size ||
3562 		    abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0);
3563 	}
3564 
3565 	for (int p = 0; p < DDT_NPHYS(ddt); p++) {
3566 		ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p);
3567 		uint64_t phys_birth = ddt_phys_birth(dde->dde_phys, v);
3568 
3569 		if (phys_birth != 0 && do_raw) {
3570 			blkptr_t blk = *zio->io_bp;
3571 			uint64_t psize;
3572 			abd_t *tmpabd;
3573 			int error;
3574 
3575 			ddt_bp_fill(dde->dde_phys, v, &blk, phys_birth);
3576 			psize = BP_GET_PSIZE(&blk);
3577 
3578 			if (psize != zio->io_size)
3579 				return (B_TRUE);
3580 
3581 			ddt_exit(ddt);
3582 
3583 			tmpabd = abd_alloc_for_io(psize, B_TRUE);
3584 
3585 			error = zio_wait(zio_read(NULL, spa, &blk, tmpabd,
3586 			    psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
3587 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
3588 			    ZIO_FLAG_RAW, &zio->io_bookmark));
3589 
3590 			if (error == 0) {
3591 				if (abd_cmp(tmpabd, zio->io_abd) != 0)
3592 					error = SET_ERROR(ENOENT);
3593 			}
3594 
3595 			abd_free(tmpabd);
3596 			ddt_enter(ddt);
3597 			return (error != 0);
3598 		} else if (phys_birth != 0) {
3599 			arc_buf_t *abuf = NULL;
3600 			arc_flags_t aflags = ARC_FLAG_WAIT;
3601 			blkptr_t blk = *zio->io_bp;
3602 			int error;
3603 
3604 			ddt_bp_fill(dde->dde_phys, v, &blk, phys_birth);
3605 
3606 			if (BP_GET_LSIZE(&blk) != zio->io_orig_size)
3607 				return (B_TRUE);
3608 
3609 			ddt_exit(ddt);
3610 
3611 			error = arc_read(NULL, spa, &blk,
3612 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
3613 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
3614 			    &aflags, &zio->io_bookmark);
3615 
3616 			if (error == 0) {
3617 				if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
3618 				    zio->io_orig_size) != 0)
3619 					error = SET_ERROR(ENOENT);
3620 				arc_buf_destroy(abuf, &abuf);
3621 			}
3622 
3623 			ddt_enter(ddt);
3624 			return (error != 0);
3625 		}
3626 	}
3627 
3628 	return (B_FALSE);
3629 }
3630 
3631 static void
3632 zio_ddt_child_write_done(zio_t *zio)
3633 {
3634 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3635 	ddt_entry_t *dde = zio->io_private;
3636 
3637 	zio_link_t *zl = NULL;
3638 	ASSERT3P(zio_walk_parents(zio, &zl), !=, NULL);
3639 
3640 	int p = DDT_PHYS_FOR_COPIES(ddt, zio->io_prop.zp_copies);
3641 	ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p);
3642 	ddt_univ_phys_t *ddp = dde->dde_phys;
3643 
3644 	ddt_enter(ddt);
3645 
3646 	/* we're the lead, so once we're done there's no one else outstanding */
3647 	if (dde->dde_io->dde_lead_zio[p] == zio)
3648 		dde->dde_io->dde_lead_zio[p] = NULL;
3649 
3650 	ddt_univ_phys_t *orig = &dde->dde_io->dde_orig_phys;
3651 
3652 	if (zio->io_error != 0) {
3653 		/*
3654 		 * The write failed, so we're about to abort the entire IO
3655 		 * chain. We need to revert the entry back to what it was at
3656 		 * the last time it was successfully extended.
3657 		 */
3658 		ddt_phys_unextend(ddp, orig, v);
3659 		ddt_phys_clear(orig, v);
3660 
3661 		ddt_exit(ddt);
3662 		return;
3663 	}
3664 
3665 	/*
3666 	 * Add references for all dedup writes that were waiting on the
3667 	 * physical one, skipping any other physical writes that are waiting.
3668 	 */
3669 	zio_t *pio;
3670 	zl = NULL;
3671 	while ((pio = zio_walk_parents(zio, &zl)) != NULL) {
3672 		if (!(pio->io_flags & ZIO_FLAG_DDT_CHILD))
3673 			ddt_phys_addref(ddp, v);
3674 	}
3675 
3676 	/*
3677 	 * We've successfully added new DVAs to the entry. Clear the saved
3678 	 * state or, if there's still outstanding IO, remember it so we can
3679 	 * revert to a known good state if that IO fails.
3680 	 */
3681 	if (dde->dde_io->dde_lead_zio[p] == NULL)
3682 		ddt_phys_clear(orig, v);
3683 	else
3684 		ddt_phys_copy(orig, ddp, v);
3685 
3686 	ddt_exit(ddt);
3687 }
3688 
3689 static void
3690 zio_ddt_child_write_ready(zio_t *zio)
3691 {
3692 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3693 	ddt_entry_t *dde = zio->io_private;
3694 
3695 	zio_link_t *zl = NULL;
3696 	ASSERT3P(zio_walk_parents(zio, &zl), !=, NULL);
3697 
3698 	int p = DDT_PHYS_FOR_COPIES(ddt, zio->io_prop.zp_copies);
3699 	ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p);
3700 
3701 	if (ddt_phys_is_gang(dde->dde_phys, v)) {
3702 		for (int i = 0; i < BP_GET_NDVAS(zio->io_bp); i++) {
3703 			dva_t *d = &zio->io_bp->blk_dva[i];
3704 			metaslab_group_alloc_decrement(zio->io_spa,
3705 			    DVA_GET_VDEV(d), zio->io_allocator,
3706 			    METASLAB_ASYNC_ALLOC, zio->io_size, zio);
3707 		}
3708 		zio->io_error = EAGAIN;
3709 	}
3710 
3711 	if (zio->io_error != 0)
3712 		return;
3713 
3714 	ddt_enter(ddt);
3715 
3716 	ddt_phys_extend(dde->dde_phys, v, zio->io_bp);
3717 
3718 	zio_t *pio;
3719 	zl = NULL;
3720 	while ((pio = zio_walk_parents(zio, &zl)) != NULL) {
3721 		if (!(pio->io_flags & ZIO_FLAG_DDT_CHILD))
3722 			ddt_bp_fill(dde->dde_phys, v, pio->io_bp, zio->io_txg);
3723 	}
3724 
3725 	ddt_exit(ddt);
3726 }
3727 
3728 static zio_t *
3729 zio_ddt_write(zio_t *zio)
3730 {
3731 	spa_t *spa = zio->io_spa;
3732 	blkptr_t *bp = zio->io_bp;
3733 	uint64_t txg = zio->io_txg;
3734 	zio_prop_t *zp = &zio->io_prop;
3735 	ddt_t *ddt = ddt_select(spa, bp);
3736 	ddt_entry_t *dde;
3737 
3738 	ASSERT(BP_GET_DEDUP(bp));
3739 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
3740 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
3741 	ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
3742 	/*
3743 	 * Deduplication will not take place for Direct I/O writes. The
3744 	 * ddt_tree will be emptied in syncing context. Direct I/O writes take
3745 	 * place in the open-context. Direct I/O write can not attempt to
3746 	 * modify the ddt_tree while issuing out a write.
3747 	 */
3748 	ASSERT3B(zio->io_prop.zp_direct_write, ==, B_FALSE);
3749 
3750 	ddt_enter(ddt);
3751 	/*
3752 	 * Search DDT for matching entry.  Skip DVAs verification here, since
3753 	 * they can go only from override, and once we get here the override
3754 	 * pointer can't have "D" flag to be confused with pruned DDT entries.
3755 	 */
3756 	IMPLY(zio->io_bp_override, !BP_GET_DEDUP(zio->io_bp_override));
3757 	dde = ddt_lookup(ddt, bp, B_FALSE);
3758 	if (dde == NULL) {
3759 		/* DDT size is over its quota so no new entries */
3760 		zp->zp_dedup = B_FALSE;
3761 		BP_SET_DEDUP(bp, B_FALSE);
3762 		if (zio->io_bp_override == NULL)
3763 			zio->io_pipeline = ZIO_WRITE_PIPELINE;
3764 		ddt_exit(ddt);
3765 		return (zio);
3766 	}
3767 
3768 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
3769 		/*
3770 		 * If we're using a weak checksum, upgrade to a strong checksum
3771 		 * and try again.  If we're already using a strong checksum,
3772 		 * we can't resolve it, so just convert to an ordinary write.
3773 		 * (And automatically e-mail a paper to Nature?)
3774 		 */
3775 		if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
3776 		    ZCHECKSUM_FLAG_DEDUP)) {
3777 			zp->zp_checksum = spa_dedup_checksum(spa);
3778 			zio_pop_transforms(zio);
3779 			zio->io_stage = ZIO_STAGE_OPEN;
3780 			BP_ZERO(bp);
3781 		} else {
3782 			zp->zp_dedup = B_FALSE;
3783 			BP_SET_DEDUP(bp, B_FALSE);
3784 		}
3785 		ASSERT(!BP_GET_DEDUP(bp));
3786 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
3787 		ddt_exit(ddt);
3788 		return (zio);
3789 	}
3790 
3791 	int p = DDT_PHYS_FOR_COPIES(ddt, zp->zp_copies);
3792 	ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p);
3793 	ddt_univ_phys_t *ddp = dde->dde_phys;
3794 
3795 	/*
3796 	 * In the common cases, at this point we have a regular BP with no
3797 	 * allocated DVAs, and the corresponding DDT entry for its checksum.
3798 	 * Our goal is to fill the BP with enough DVAs to satisfy its copies=
3799 	 * requirement.
3800 	 *
3801 	 * One of three things needs to happen to fulfill this:
3802 	 *
3803 	 * - if the DDT entry has enough DVAs to satisfy the BP, we just copy
3804 	 *   them out of the entry and return;
3805 	 *
3806 	 * - if the DDT entry has no DVAs (ie its brand new), then we have to
3807 	 *   issue the write as normal so that DVAs can be allocated and the
3808 	 *   data land on disk. We then copy the DVAs into the DDT entry on
3809 	 *   return.
3810 	 *
3811 	 * - if the DDT entry has some DVAs, but too few, we have to issue the
3812 	 *   write, adjusted to have allocate fewer copies. When it returns, we
3813 	 *   add the new DVAs to the DDT entry, and update the BP to have the
3814 	 *   full amount it originally requested.
3815 	 *
3816 	 * In all cases, if there's already a writing IO in flight, we need to
3817 	 * defer the action until after the write is done. If our action is to
3818 	 * write, we need to adjust our request for additional DVAs to match
3819 	 * what will be in the DDT entry after it completes. In this way every
3820 	 * IO can be guaranteed to recieve enough DVAs simply by joining the
3821 	 * end of the chain and letting the sequence play out.
3822 	 */
3823 
3824 	/*
3825 	 * Number of DVAs in the DDT entry. If the BP is encrypted we ignore
3826 	 * the third one as normal.
3827 	 */
3828 	int have_dvas = ddt_phys_dva_count(ddp, v, BP_IS_ENCRYPTED(bp));
3829 	IMPLY(have_dvas == 0, ddt_phys_birth(ddp, v) == 0);
3830 	boolean_t is_ganged = ddt_phys_is_gang(ddp, v);
3831 
3832 	/* Number of DVAs requested by the IO. */
3833 	uint8_t need_dvas = zp->zp_copies;
3834 	/* Number of DVAs in outstanding writes for this dde. */
3835 	uint8_t parent_dvas = 0;
3836 
3837 	/*
3838 	 * What we do next depends on whether or not there's IO outstanding that
3839 	 * will update this entry.
3840 	 */
3841 	if (dde->dde_io == NULL || dde->dde_io->dde_lead_zio[p] == NULL) {
3842 		/*
3843 		 * No IO outstanding, so we only need to worry about ourselves.
3844 		 */
3845 
3846 		/*
3847 		 * Override BPs bring their own DVAs and their own problems.
3848 		 */
3849 		if (zio->io_bp_override) {
3850 			/*
3851 			 * For a brand-new entry, all the work has been done
3852 			 * for us, and we can just fill it out from the provided
3853 			 * block and leave.
3854 			 */
3855 			if (have_dvas == 0) {
3856 				ASSERT(BP_GET_LOGICAL_BIRTH(bp) == txg);
3857 				ASSERT(BP_EQUAL(bp, zio->io_bp_override));
3858 				ddt_phys_extend(ddp, v, bp);
3859 				ddt_phys_addref(ddp, v);
3860 				ddt_exit(ddt);
3861 				return (zio);
3862 			}
3863 
3864 			/*
3865 			 * If we already have this entry, then we want to treat
3866 			 * it like a regular write. To do this we just wipe
3867 			 * them out and proceed like a regular write.
3868 			 *
3869 			 * Even if there are some DVAs in the entry, we still
3870 			 * have to clear them out. We can't use them to fill
3871 			 * out the dedup entry, as they are all referenced
3872 			 * together by a bp already on disk, and will be freed
3873 			 * as a group.
3874 			 */
3875 			BP_ZERO_DVAS(bp);
3876 			BP_SET_BIRTH(bp, 0, 0);
3877 		}
3878 
3879 		/*
3880 		 * If there are enough DVAs in the entry to service our request,
3881 		 * then we can just use them as-is.
3882 		 */
3883 		if (have_dvas >= need_dvas) {
3884 			ddt_bp_fill(ddp, v, bp, txg);
3885 			ddt_phys_addref(ddp, v);
3886 			ddt_exit(ddt);
3887 			return (zio);
3888 		}
3889 
3890 		/*
3891 		 * Otherwise, we have to issue IO to fill the entry up to the
3892 		 * amount we need.
3893 		 */
3894 		need_dvas -= have_dvas;
3895 	} else {
3896 		/*
3897 		 * There's a write in-flight. If there's already enough DVAs on
3898 		 * the entry, then either there were already enough to start
3899 		 * with, or the in-flight IO is between READY and DONE, and so
3900 		 * has extended the entry with new DVAs. Either way, we don't
3901 		 * need to do anything, we can just slot in behind it.
3902 		 */
3903 
3904 		if (zio->io_bp_override) {
3905 			/*
3906 			 * If there's a write out, then we're soon going to
3907 			 * have our own copies of this block, so clear out the
3908 			 * override block and treat it as a regular dedup
3909 			 * write. See comment above.
3910 			 */
3911 			BP_ZERO_DVAS(bp);
3912 			BP_SET_BIRTH(bp, 0, 0);
3913 		}
3914 
3915 		if (have_dvas >= need_dvas) {
3916 			/*
3917 			 * A minor point: there might already be enough
3918 			 * committed DVAs in the entry to service our request,
3919 			 * but we don't know which are completed and which are
3920 			 * allocated but not yet written. In this case, should
3921 			 * the IO for the new DVAs fail, we will be on the end
3922 			 * of the IO chain and will also recieve an error, even
3923 			 * though our request could have been serviced.
3924 			 *
3925 			 * This is an extremely rare case, as it requires the
3926 			 * original block to be copied with a request for a
3927 			 * larger number of DVAs, then copied again requesting
3928 			 * the same (or already fulfilled) number of DVAs while
3929 			 * the first request is active, and then that first
3930 			 * request errors. In return, the logic required to
3931 			 * catch and handle it is complex. For now, I'm just
3932 			 * not going to bother with it.
3933 			 */
3934 
3935 			/*
3936 			 * We always fill the bp here as we may have arrived
3937 			 * after the in-flight write has passed READY, and so
3938 			 * missed out.
3939 			 */
3940 			ddt_bp_fill(ddp, v, bp, txg);
3941 			zio_add_child(zio, dde->dde_io->dde_lead_zio[p]);
3942 			ddt_exit(ddt);
3943 			return (zio);
3944 		}
3945 
3946 		/*
3947 		 * There's not enough in the entry yet, so we need to look at
3948 		 * the write in-flight and see how many DVAs it will have once
3949 		 * it completes.
3950 		 *
3951 		 * The in-flight write has potentially had its copies request
3952 		 * reduced (if we're filling out an existing entry), so we need
3953 		 * to reach in and get the original write to find out what it is
3954 		 * expecting.
3955 		 *
3956 		 * Note that the parent of the lead zio will always have the
3957 		 * highest zp_copies of any zio in the chain, because ones that
3958 		 * can be serviced without additional IO are always added to
3959 		 * the back of the chain.
3960 		 */
3961 		zio_link_t *zl = NULL;
3962 		zio_t *pio =
3963 		    zio_walk_parents(dde->dde_io->dde_lead_zio[p], &zl);
3964 		ASSERT(pio);
3965 		parent_dvas = pio->io_prop.zp_copies;
3966 
3967 		if (parent_dvas >= need_dvas) {
3968 			zio_add_child(zio, dde->dde_io->dde_lead_zio[p]);
3969 			ddt_exit(ddt);
3970 			return (zio);
3971 		}
3972 
3973 		/*
3974 		 * Still not enough, so we will need to issue to get the
3975 		 * shortfall.
3976 		 */
3977 		need_dvas -= parent_dvas;
3978 	}
3979 
3980 	if (is_ganged) {
3981 		zp->zp_dedup = B_FALSE;
3982 		BP_SET_DEDUP(bp, B_FALSE);
3983 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
3984 		ddt_exit(ddt);
3985 		return (zio);
3986 	}
3987 
3988 	/*
3989 	 * We need to write. We will create a new write with the copies
3990 	 * property adjusted to match the number of DVAs we need to need to
3991 	 * grow the DDT entry by to satisfy the request.
3992 	 */
3993 	zio_prop_t czp = *zp;
3994 	if (have_dvas > 0 || parent_dvas > 0) {
3995 		czp.zp_copies = need_dvas;
3996 		czp.zp_gang_copies = 0;
3997 	} else {
3998 		ASSERT3U(czp.zp_copies, ==, need_dvas);
3999 	}
4000 
4001 	zio_t *cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
4002 	    zio->io_orig_size, zio->io_orig_size, &czp,
4003 	    zio_ddt_child_write_ready, NULL,
4004 	    zio_ddt_child_write_done, dde, zio->io_priority,
4005 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
4006 
4007 	zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
4008 
4009 	/*
4010 	 * We are the new lead zio, because our parent has the highest
4011 	 * zp_copies that has been requested for this entry so far.
4012 	 */
4013 	ddt_alloc_entry_io(dde);
4014 	if (dde->dde_io->dde_lead_zio[p] == NULL) {
4015 		/*
4016 		 * First time out, take a copy of the stable entry to revert
4017 		 * to if there's an error (see zio_ddt_child_write_done())
4018 		 */
4019 		ddt_phys_copy(&dde->dde_io->dde_orig_phys, dde->dde_phys, v);
4020 	} else {
4021 		/*
4022 		 * Make the existing chain our child, because it cannot
4023 		 * complete until we have.
4024 		 */
4025 		zio_add_child(cio, dde->dde_io->dde_lead_zio[p]);
4026 	}
4027 	dde->dde_io->dde_lead_zio[p] = cio;
4028 
4029 	ddt_exit(ddt);
4030 
4031 	zio_nowait(cio);
4032 
4033 	return (zio);
4034 }
4035 
4036 static ddt_entry_t *freedde; /* for debugging */
4037 
4038 static zio_t *
4039 zio_ddt_free(zio_t *zio)
4040 {
4041 	spa_t *spa = zio->io_spa;
4042 	blkptr_t *bp = zio->io_bp;
4043 	ddt_t *ddt = ddt_select(spa, bp);
4044 	ddt_entry_t *dde = NULL;
4045 
4046 	ASSERT(BP_GET_DEDUP(bp));
4047 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4048 
4049 	ddt_enter(ddt);
4050 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
4051 	if (dde) {
4052 		ddt_phys_variant_t v = ddt_phys_select(ddt, dde, bp);
4053 		if (v != DDT_PHYS_NONE)
4054 			ddt_phys_decref(dde->dde_phys, v);
4055 	}
4056 	ddt_exit(ddt);
4057 
4058 	/*
4059 	 * When no entry was found, it must have been pruned,
4060 	 * so we can free it now instead of decrementing the
4061 	 * refcount in the DDT.
4062 	 */
4063 	if (!dde) {
4064 		BP_SET_DEDUP(bp, 0);
4065 		zio->io_pipeline |= ZIO_STAGE_DVA_FREE;
4066 	}
4067 
4068 	return (zio);
4069 }
4070 
4071 /*
4072  * ==========================================================================
4073  * Allocate and free blocks
4074  * ==========================================================================
4075  */
4076 
4077 static zio_t *
4078 zio_io_to_allocate(metaslab_class_allocator_t *mca, boolean_t *more)
4079 {
4080 	zio_t *zio;
4081 
4082 	ASSERT(MUTEX_HELD(&mca->mca_lock));
4083 
4084 	zio = avl_first(&mca->mca_tree);
4085 	if (zio == NULL) {
4086 		*more = B_FALSE;
4087 		return (NULL);
4088 	}
4089 
4090 	ASSERT(IO_IS_ALLOCATING(zio));
4091 	ASSERT(ZIO_HAS_ALLOCATOR(zio));
4092 
4093 	/*
4094 	 * Try to place a reservation for this zio. If we're unable to
4095 	 * reserve then we throttle.
4096 	 */
4097 	if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
4098 	    zio->io_prop.zp_copies, zio, B_FALSE, more)) {
4099 		return (NULL);
4100 	}
4101 
4102 	avl_remove(&mca->mca_tree, zio);
4103 	ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
4104 
4105 	if (avl_is_empty(&mca->mca_tree))
4106 		*more = B_FALSE;
4107 	return (zio);
4108 }
4109 
4110 static zio_t *
4111 zio_dva_throttle(zio_t *zio)
4112 {
4113 	spa_t *spa = zio->io_spa;
4114 	zio_t *nio;
4115 	metaslab_class_t *mc;
4116 	boolean_t more;
4117 
4118 	/*
4119 	 * If not already chosen, choose an appropriate allocation class.
4120 	 */
4121 	mc = zio->io_metaslab_class;
4122 	if (mc == NULL)
4123 		mc = spa_preferred_class(spa, zio);
4124 
4125 	if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
4126 	    !mc->mc_alloc_throttle_enabled ||
4127 	    zio->io_child_type == ZIO_CHILD_GANG ||
4128 	    zio->io_flags & ZIO_FLAG_NODATA) {
4129 		return (zio);
4130 	}
4131 
4132 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4133 	ASSERT(ZIO_HAS_ALLOCATOR(zio));
4134 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
4135 	ASSERT3U(zio->io_queued_timestamp, >, 0);
4136 	ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
4137 
4138 	zio->io_metaslab_class = mc;
4139 	metaslab_class_allocator_t *mca = &mc->mc_allocator[zio->io_allocator];
4140 	mutex_enter(&mca->mca_lock);
4141 	avl_add(&mca->mca_tree, zio);
4142 	nio = zio_io_to_allocate(mca, &more);
4143 	mutex_exit(&mca->mca_lock);
4144 	return (nio);
4145 }
4146 
4147 static void
4148 zio_allocate_dispatch(metaslab_class_t *mc, int allocator)
4149 {
4150 	metaslab_class_allocator_t *mca = &mc->mc_allocator[allocator];
4151 	zio_t *zio;
4152 	boolean_t more;
4153 
4154 	do {
4155 		mutex_enter(&mca->mca_lock);
4156 		zio = zio_io_to_allocate(mca, &more);
4157 		mutex_exit(&mca->mca_lock);
4158 		if (zio == NULL)
4159 			return;
4160 
4161 		ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
4162 		ASSERT0(zio->io_error);
4163 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
4164 	} while (more);
4165 }
4166 
4167 static zio_t *
4168 zio_dva_allocate(zio_t *zio)
4169 {
4170 	spa_t *spa = zio->io_spa;
4171 	metaslab_class_t *mc;
4172 	blkptr_t *bp = zio->io_bp;
4173 	int error;
4174 	int flags = 0;
4175 
4176 	if (zio->io_gang_leader == NULL) {
4177 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
4178 		zio->io_gang_leader = zio;
4179 	}
4180 	if (zio->io_flags & ZIO_FLAG_PREALLOCATED) {
4181 		ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_GANG);
4182 		memcpy(zio->io_bp->blk_dva, zio->io_bp_orig.blk_dva,
4183 		    3 * sizeof (dva_t));
4184 		BP_SET_BIRTH(zio->io_bp, BP_GET_LOGICAL_BIRTH(&zio->io_bp_orig),
4185 		    BP_GET_PHYSICAL_BIRTH(&zio->io_bp_orig));
4186 		return (zio);
4187 	}
4188 
4189 	ASSERT(BP_IS_HOLE(bp));
4190 	ASSERT0(BP_GET_NDVAS(bp));
4191 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
4192 
4193 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
4194 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
4195 
4196 	if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
4197 		flags |= METASLAB_GANG_CHILD;
4198 	if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
4199 		flags |= METASLAB_ASYNC_ALLOC;
4200 
4201 	/*
4202 	 * If not already chosen, choose an appropriate allocation class.
4203 	 */
4204 	mc = zio->io_metaslab_class;
4205 	if (mc == NULL) {
4206 		mc = spa_preferred_class(spa, zio);
4207 		zio->io_metaslab_class = mc;
4208 	}
4209 	ZIOSTAT_BUMP(ziostat_total_allocations);
4210 
4211 again:
4212 	/*
4213 	 * Try allocating the block in the usual metaslab class.
4214 	 * If that's full, allocate it in the normal class.
4215 	 * If that's full, allocate as a gang block,
4216 	 * and if all are full, the allocation fails (which shouldn't happen).
4217 	 *
4218 	 * Note that we do not fall back on embedded slog (ZIL) space, to
4219 	 * preserve unfragmented slog space, which is critical for decent
4220 	 * sync write performance.  If a log allocation fails, we will fall
4221 	 * back to spa_sync() which is abysmal for performance.
4222 	 */
4223 	ASSERT(ZIO_HAS_ALLOCATOR(zio));
4224 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
4225 	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
4226 	    &zio->io_alloc_list, zio->io_allocator, zio);
4227 
4228 	/*
4229 	 * Fallback to normal class when an alloc class is full
4230 	 */
4231 	if (error == ENOSPC && mc != spa_normal_class(spa)) {
4232 		/*
4233 		 * When the dedup or special class is spilling into the  normal
4234 		 * class, there can still be significant space available due
4235 		 * to deferred frees that are in-flight.  We track the txg when
4236 		 * this occurred and back off adding new DDT entries for a few
4237 		 * txgs to allow the free blocks to be processed.
4238 		 */
4239 		if ((mc == spa_dedup_class(spa) || (spa_special_has_ddt(spa) &&
4240 		    mc == spa_special_class(spa))) &&
4241 		    spa->spa_dedup_class_full_txg != zio->io_txg) {
4242 			spa->spa_dedup_class_full_txg = zio->io_txg;
4243 			zfs_dbgmsg("%s[%d]: %s class spilling, req size %d, "
4244 			    "%llu allocated of %llu",
4245 			    spa_name(spa), (int)zio->io_txg,
4246 			    mc == spa_dedup_class(spa) ? "dedup" : "special",
4247 			    (int)zio->io_size,
4248 			    (u_longlong_t)metaslab_class_get_alloc(mc),
4249 			    (u_longlong_t)metaslab_class_get_space(mc));
4250 		}
4251 
4252 		/*
4253 		 * If we are holding old class reservation, drop it.
4254 		 * Dispatch the next ZIO(s) there if some are waiting.
4255 		 */
4256 		if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4257 			if (metaslab_class_throttle_unreserve(mc,
4258 			    zio->io_prop.zp_copies, zio)) {
4259 				zio_allocate_dispatch(zio->io_metaslab_class,
4260 				    zio->io_allocator);
4261 			}
4262 			zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING;
4263 		}
4264 
4265 		if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) {
4266 			zfs_dbgmsg("%s: metaslab allocation failure, "
4267 			    "trying normal class: zio %px, size %llu, error %d",
4268 			    spa_name(spa), zio, (u_longlong_t)zio->io_size,
4269 			    error);
4270 		}
4271 		zio->io_metaslab_class = mc = spa_normal_class(spa);
4272 		ZIOSTAT_BUMP(ziostat_alloc_class_fallbacks);
4273 
4274 		/*
4275 		 * If normal class uses throttling, return to that pipeline
4276 		 * stage.  Otherwise just do another allocation attempt.
4277 		 */
4278 		if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
4279 		    mc->mc_alloc_throttle_enabled &&
4280 		    zio->io_child_type != ZIO_CHILD_GANG &&
4281 		    !(zio->io_flags & ZIO_FLAG_NODATA)) {
4282 			zio->io_stage = ZIO_STAGE_DVA_THROTTLE >> 1;
4283 			return (zio);
4284 		}
4285 		goto again;
4286 	}
4287 
4288 	if (error == ENOSPC && zio->io_size > spa->spa_min_alloc) {
4289 		if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) {
4290 			zfs_dbgmsg("%s: metaslab allocation failure, "
4291 			    "trying ganging: zio %px, size %llu, error %d",
4292 			    spa_name(spa), zio, (u_longlong_t)zio->io_size,
4293 			    error);
4294 		}
4295 		ZIOSTAT_BUMP(ziostat_gang_writes);
4296 		if (flags & METASLAB_GANG_CHILD)
4297 			ZIOSTAT_BUMP(ziostat_gang_multilevel);
4298 		return (zio_write_gang_block(zio, mc));
4299 	}
4300 	if (error != 0) {
4301 		if (error != ENOSPC ||
4302 		    (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC)) {
4303 			zfs_dbgmsg("%s: metaslab allocation failure: zio %px, "
4304 			    "size %llu, error %d",
4305 			    spa_name(spa), zio, (u_longlong_t)zio->io_size,
4306 			    error);
4307 		}
4308 		zio->io_error = error;
4309 	}
4310 
4311 	return (zio);
4312 }
4313 
4314 static zio_t *
4315 zio_dva_free(zio_t *zio)
4316 {
4317 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
4318 
4319 	return (zio);
4320 }
4321 
4322 static zio_t *
4323 zio_dva_claim(zio_t *zio)
4324 {
4325 	int error;
4326 
4327 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
4328 	if (error)
4329 		zio->io_error = error;
4330 
4331 	return (zio);
4332 }
4333 
4334 /*
4335  * Undo an allocation.  This is used by zio_done() when an I/O fails
4336  * and we want to give back the block we just allocated.
4337  * This handles both normal blocks and gang blocks.
4338  */
4339 static void
4340 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
4341 {
4342 	ASSERT(BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg || BP_IS_HOLE(bp));
4343 	ASSERT(zio->io_bp_override == NULL);
4344 
4345 	if (!BP_IS_HOLE(bp)) {
4346 		metaslab_free(zio->io_spa, bp, BP_GET_LOGICAL_BIRTH(bp),
4347 		    B_TRUE);
4348 	}
4349 
4350 	if (gn != NULL) {
4351 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
4352 			zio_dva_unallocate(zio, gn->gn_child[g],
4353 			    &gn->gn_gbh->zg_blkptr[g]);
4354 		}
4355 	}
4356 }
4357 
4358 /*
4359  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
4360  */
4361 int
4362 zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
4363     uint64_t size, boolean_t *slog)
4364 {
4365 	int error = 1;
4366 	zio_alloc_list_t io_alloc_list;
4367 
4368 	ASSERT(txg > spa_syncing_txg(spa));
4369 
4370 	metaslab_trace_init(&io_alloc_list);
4371 
4372 	/*
4373 	 * Block pointer fields are useful to metaslabs for stats and debugging.
4374 	 * Fill in the obvious ones before calling into metaslab_alloc().
4375 	 */
4376 	BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
4377 	BP_SET_PSIZE(new_bp, size);
4378 	BP_SET_LEVEL(new_bp, 0);
4379 
4380 	/*
4381 	 * When allocating a zil block, we don't have information about
4382 	 * the final destination of the block except the objset it's part
4383 	 * of, so we just hash the objset ID to pick the allocator to get
4384 	 * some parallelism.
4385 	 */
4386 	int flags = METASLAB_ZIL;
4387 	int allocator = (uint_t)cityhash1(os->os_dsl_dataset->ds_object)
4388 	    % spa->spa_alloc_count;
4389 	ZIOSTAT_BUMP(ziostat_total_allocations);
4390 	error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
4391 	    txg, NULL, flags, &io_alloc_list, allocator, NULL);
4392 	*slog = (error == 0);
4393 	if (error != 0) {
4394 		error = metaslab_alloc(spa, spa_embedded_log_class(spa), size,
4395 		    new_bp, 1, txg, NULL, flags, &io_alloc_list, allocator,
4396 		    NULL);
4397 	}
4398 	if (error != 0) {
4399 		ZIOSTAT_BUMP(ziostat_alloc_class_fallbacks);
4400 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
4401 		    new_bp, 1, txg, NULL, flags, &io_alloc_list, allocator,
4402 		    NULL);
4403 	}
4404 	metaslab_trace_fini(&io_alloc_list);
4405 
4406 	if (error == 0) {
4407 		BP_SET_LSIZE(new_bp, size);
4408 		BP_SET_PSIZE(new_bp, size);
4409 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
4410 		BP_SET_CHECKSUM(new_bp,
4411 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
4412 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
4413 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
4414 		BP_SET_LEVEL(new_bp, 0);
4415 		BP_SET_DEDUP(new_bp, 0);
4416 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
4417 
4418 		/*
4419 		 * encrypted blocks will require an IV and salt. We generate
4420 		 * these now since we will not be rewriting the bp at
4421 		 * rewrite time.
4422 		 */
4423 		if (os->os_encrypted) {
4424 			uint8_t iv[ZIO_DATA_IV_LEN];
4425 			uint8_t salt[ZIO_DATA_SALT_LEN];
4426 
4427 			BP_SET_CRYPT(new_bp, B_TRUE);
4428 			VERIFY0(spa_crypt_get_salt(spa,
4429 			    dmu_objset_id(os), salt));
4430 			VERIFY0(zio_crypt_generate_iv(iv));
4431 
4432 			zio_crypt_encode_params_bp(new_bp, salt, iv);
4433 		}
4434 	} else {
4435 		zfs_dbgmsg("%s: zil block allocation failure: "
4436 		    "size %llu, error %d", spa_name(spa), (u_longlong_t)size,
4437 		    error);
4438 	}
4439 
4440 	return (error);
4441 }
4442 
4443 /*
4444  * ==========================================================================
4445  * Read and write to physical devices
4446  * ==========================================================================
4447  */
4448 
4449 /*
4450  * Issue an I/O to the underlying vdev. Typically the issue pipeline
4451  * stops after this stage and will resume upon I/O completion.
4452  * However, there are instances where the vdev layer may need to
4453  * continue the pipeline when an I/O was not issued. Since the I/O
4454  * that was sent to the vdev layer might be different than the one
4455  * currently active in the pipeline (see vdev_queue_io()), we explicitly
4456  * force the underlying vdev layers to call either zio_execute() or
4457  * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
4458  */
4459 static zio_t *
4460 zio_vdev_io_start(zio_t *zio)
4461 {
4462 	vdev_t *vd = zio->io_vd;
4463 	uint64_t align;
4464 	spa_t *spa = zio->io_spa;
4465 
4466 	zio->io_delay = 0;
4467 
4468 	ASSERT(zio->io_error == 0);
4469 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
4470 
4471 	if (vd == NULL) {
4472 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
4473 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
4474 
4475 		/*
4476 		 * The mirror_ops handle multiple DVAs in a single BP.
4477 		 */
4478 		vdev_mirror_ops.vdev_op_io_start(zio);
4479 		return (NULL);
4480 	}
4481 
4482 	ASSERT3P(zio->io_logical, !=, zio);
4483 	if (zio->io_type == ZIO_TYPE_WRITE) {
4484 		ASSERT(spa->spa_trust_config);
4485 
4486 		/*
4487 		 * Note: the code can handle other kinds of writes,
4488 		 * but we don't expect them.
4489 		 */
4490 		if (zio->io_vd->vdev_noalloc) {
4491 			ASSERT(zio->io_flags &
4492 			    (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
4493 			    ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
4494 		}
4495 	}
4496 
4497 	align = 1ULL << vd->vdev_top->vdev_ashift;
4498 
4499 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
4500 	    P2PHASE(zio->io_size, align) != 0) {
4501 		/* Transform logical writes to be a full physical block size. */
4502 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
4503 		abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
4504 		ASSERT(vd == vd->vdev_top);
4505 		if (zio->io_type == ZIO_TYPE_WRITE) {
4506 			abd_copy(abuf, zio->io_abd, zio->io_size);
4507 			abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
4508 		}
4509 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
4510 	}
4511 
4512 	/*
4513 	 * If this is not a physical io, make sure that it is properly aligned
4514 	 * before proceeding.
4515 	 */
4516 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
4517 		ASSERT0(P2PHASE(zio->io_offset, align));
4518 		ASSERT0(P2PHASE(zio->io_size, align));
4519 	} else {
4520 		/*
4521 		 * For physical writes, we allow 512b aligned writes and assume
4522 		 * the device will perform a read-modify-write as necessary.
4523 		 */
4524 		ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
4525 		ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
4526 	}
4527 
4528 	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
4529 
4530 	/*
4531 	 * If this is a repair I/O, and there's no self-healing involved --
4532 	 * that is, we're just resilvering what we expect to resilver --
4533 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
4534 	 * This prevents spurious resilvering.
4535 	 *
4536 	 * There are a few ways that we can end up creating these spurious
4537 	 * resilver i/os:
4538 	 *
4539 	 * 1. A resilver i/o will be issued if any DVA in the BP has a
4540 	 * dirty DTL.  The mirror code will issue resilver writes to
4541 	 * each DVA, including the one(s) that are not on vdevs with dirty
4542 	 * DTLs.
4543 	 *
4544 	 * 2. With nested replication, which happens when we have a
4545 	 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
4546 	 * For example, given mirror(replacing(A+B), C), it's likely that
4547 	 * only A is out of date (it's the new device). In this case, we'll
4548 	 * read from C, then use the data to resilver A+B -- but we don't
4549 	 * actually want to resilver B, just A. The top-level mirror has no
4550 	 * way to know this, so instead we just discard unnecessary repairs
4551 	 * as we work our way down the vdev tree.
4552 	 *
4553 	 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
4554 	 * The same logic applies to any form of nested replication: ditto
4555 	 * + mirror, RAID-Z + replacing, etc.
4556 	 *
4557 	 * However, indirect vdevs point off to other vdevs which may have
4558 	 * DTL's, so we never bypass them.  The child i/os on concrete vdevs
4559 	 * will be properly bypassed instead.
4560 	 *
4561 	 * Leaf DTL_PARTIAL can be empty when a legitimate write comes from
4562 	 * a dRAID spare vdev. For example, when a dRAID spare is first
4563 	 * used, its spare blocks need to be written to but the leaf vdev's
4564 	 * of such blocks can have empty DTL_PARTIAL.
4565 	 *
4566 	 * There seemed no clean way to allow such writes while bypassing
4567 	 * spurious ones. At this point, just avoid all bypassing for dRAID
4568 	 * for correctness.
4569 	 */
4570 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
4571 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
4572 	    zio->io_txg != 0 &&	/* not a delegated i/o */
4573 	    vd->vdev_ops != &vdev_indirect_ops &&
4574 	    vd->vdev_top->vdev_ops != &vdev_draid_ops &&
4575 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
4576 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4577 		zio_vdev_io_bypass(zio);
4578 		return (zio);
4579 	}
4580 
4581 	/*
4582 	 * Select the next best leaf I/O to process.  Distributed spares are
4583 	 * excluded since they dispatch the I/O directly to a leaf vdev after
4584 	 * applying the dRAID mapping.
4585 	 */
4586 	if (vd->vdev_ops->vdev_op_leaf &&
4587 	    vd->vdev_ops != &vdev_draid_spare_ops &&
4588 	    (zio->io_type == ZIO_TYPE_READ ||
4589 	    zio->io_type == ZIO_TYPE_WRITE ||
4590 	    zio->io_type == ZIO_TYPE_TRIM)) {
4591 
4592 		if ((zio = vdev_queue_io(zio)) == NULL)
4593 			return (NULL);
4594 
4595 		if (!vdev_accessible(vd, zio)) {
4596 			zio->io_error = SET_ERROR(ENXIO);
4597 			zio_interrupt(zio);
4598 			return (NULL);
4599 		}
4600 		zio->io_delay = gethrtime();
4601 
4602 		if (zio_handle_device_injection(vd, zio, ENOSYS) != 0) {
4603 			/*
4604 			 * "no-op" injections return success, but do no actual
4605 			 * work. Just return it.
4606 			 */
4607 			zio_delay_interrupt(zio);
4608 			return (NULL);
4609 		}
4610 	}
4611 
4612 	vd->vdev_ops->vdev_op_io_start(zio);
4613 	return (NULL);
4614 }
4615 
4616 static zio_t *
4617 zio_vdev_io_done(zio_t *zio)
4618 {
4619 	vdev_t *vd = zio->io_vd;
4620 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
4621 	boolean_t unexpected_error = B_FALSE;
4622 
4623 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
4624 		return (NULL);
4625 	}
4626 
4627 	ASSERT(zio->io_type == ZIO_TYPE_READ ||
4628 	    zio->io_type == ZIO_TYPE_WRITE ||
4629 	    zio->io_type == ZIO_TYPE_FLUSH ||
4630 	    zio->io_type == ZIO_TYPE_TRIM);
4631 
4632 	if (zio->io_delay)
4633 		zio->io_delay = gethrtime() - zio->io_delay;
4634 
4635 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
4636 	    vd->vdev_ops != &vdev_draid_spare_ops) {
4637 		if (zio->io_type != ZIO_TYPE_FLUSH)
4638 			vdev_queue_io_done(zio);
4639 
4640 		if (zio_injection_enabled && zio->io_error == 0)
4641 			zio->io_error = zio_handle_device_injections(vd, zio,
4642 			    EIO, EILSEQ);
4643 
4644 		if (zio_injection_enabled && zio->io_error == 0)
4645 			zio->io_error = zio_handle_label_injection(zio, EIO);
4646 
4647 		if (zio->io_error && zio->io_type != ZIO_TYPE_FLUSH &&
4648 		    zio->io_type != ZIO_TYPE_TRIM) {
4649 			if (!vdev_accessible(vd, zio)) {
4650 				zio->io_error = SET_ERROR(ENXIO);
4651 			} else {
4652 				unexpected_error = B_TRUE;
4653 			}
4654 		}
4655 	}
4656 
4657 	ops->vdev_op_io_done(zio);
4658 
4659 	if (unexpected_error && vd->vdev_remove_wanted == B_FALSE)
4660 		VERIFY(vdev_probe(vd, zio) == NULL);
4661 
4662 	return (zio);
4663 }
4664 
4665 /*
4666  * This function is used to change the priority of an existing zio that is
4667  * currently in-flight. This is used by the arc to upgrade priority in the
4668  * event that a demand read is made for a block that is currently queued
4669  * as a scrub or async read IO. Otherwise, the high priority read request
4670  * would end up having to wait for the lower priority IO.
4671  */
4672 void
4673 zio_change_priority(zio_t *pio, zio_priority_t priority)
4674 {
4675 	zio_t *cio, *cio_next;
4676 	zio_link_t *zl = NULL;
4677 
4678 	ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
4679 
4680 	if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
4681 		vdev_queue_change_io_priority(pio, priority);
4682 	} else {
4683 		pio->io_priority = priority;
4684 	}
4685 
4686 	mutex_enter(&pio->io_lock);
4687 	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
4688 		cio_next = zio_walk_children(pio, &zl);
4689 		zio_change_priority(cio, priority);
4690 	}
4691 	mutex_exit(&pio->io_lock);
4692 }
4693 
4694 /*
4695  * For non-raidz ZIOs, we can just copy aside the bad data read from the
4696  * disk, and use that to finish the checksum ereport later.
4697  */
4698 static void
4699 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
4700     const abd_t *good_buf)
4701 {
4702 	/* no processing needed */
4703 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
4704 }
4705 
4706 void
4707 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr)
4708 {
4709 	void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size);
4710 
4711 	abd_copy(abd, zio->io_abd, zio->io_size);
4712 
4713 	zcr->zcr_cbinfo = zio->io_size;
4714 	zcr->zcr_cbdata = abd;
4715 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
4716 	zcr->zcr_free = zio_abd_free;
4717 }
4718 
4719 static zio_t *
4720 zio_vdev_io_assess(zio_t *zio)
4721 {
4722 	vdev_t *vd = zio->io_vd;
4723 
4724 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
4725 		return (NULL);
4726 	}
4727 
4728 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
4729 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
4730 
4731 	if (zio->io_vsd != NULL) {
4732 		zio->io_vsd_ops->vsd_free(zio);
4733 		zio->io_vsd = NULL;
4734 	}
4735 
4736 	/*
4737 	 * If a Direct I/O operation has a checksum verify error then this I/O
4738 	 * should not attempt to be issued again.
4739 	 */
4740 	if (zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR) {
4741 		if (zio->io_type == ZIO_TYPE_WRITE) {
4742 			ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_LOGICAL);
4743 			ASSERT3U(zio->io_error, ==, EIO);
4744 		}
4745 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4746 		return (zio);
4747 	}
4748 
4749 	if (zio_injection_enabled && zio->io_error == 0)
4750 		zio->io_error = zio_handle_fault_injection(zio, EIO);
4751 
4752 	/*
4753 	 * If the I/O failed, determine whether we should attempt to retry it.
4754 	 *
4755 	 * On retry, we cut in line in the issue queue, since we don't want
4756 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
4757 	 */
4758 	if (zio->io_error && vd == NULL &&
4759 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
4760 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
4761 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
4762 		zio->io_error = 0;
4763 		zio->io_flags |= ZIO_FLAG_IO_RETRY | ZIO_FLAG_DONT_AGGREGATE;
4764 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
4765 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
4766 		    zio_requeue_io_start_cut_in_line);
4767 		return (NULL);
4768 	}
4769 
4770 	/*
4771 	 * If we got an error on a leaf device, convert it to ENXIO
4772 	 * if the device is not accessible at all.
4773 	 */
4774 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
4775 	    !vdev_accessible(vd, zio))
4776 		zio->io_error = SET_ERROR(ENXIO);
4777 
4778 	/*
4779 	 * If we can't write to an interior vdev (mirror or RAID-Z),
4780 	 * set vdev_cant_write so that we stop trying to allocate from it.
4781 	 */
4782 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
4783 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
4784 		vdev_dbgmsg(vd, "zio_vdev_io_assess(zio=%px) setting "
4785 		    "cant_write=TRUE due to write failure with ENXIO",
4786 		    zio);
4787 		vd->vdev_cant_write = B_TRUE;
4788 	}
4789 
4790 	/*
4791 	 * If a cache flush returns ENOTSUP we know that no future
4792 	 * attempts will ever succeed. In this case we set a persistent
4793 	 * boolean flag so that we don't bother with it in the future, and
4794 	 * then we act like the flush succeeded.
4795 	 */
4796 	if (zio->io_error == ENOTSUP && zio->io_type == ZIO_TYPE_FLUSH &&
4797 	    vd != NULL) {
4798 		vd->vdev_nowritecache = B_TRUE;
4799 		zio->io_error = 0;
4800 	}
4801 
4802 	if (zio->io_error)
4803 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4804 
4805 	return (zio);
4806 }
4807 
4808 void
4809 zio_vdev_io_reissue(zio_t *zio)
4810 {
4811 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4812 	ASSERT(zio->io_error == 0);
4813 
4814 	zio->io_stage >>= 1;
4815 }
4816 
4817 void
4818 zio_vdev_io_redone(zio_t *zio)
4819 {
4820 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
4821 
4822 	zio->io_stage >>= 1;
4823 }
4824 
4825 void
4826 zio_vdev_io_bypass(zio_t *zio)
4827 {
4828 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4829 	ASSERT(zio->io_error == 0);
4830 
4831 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
4832 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
4833 }
4834 
4835 /*
4836  * ==========================================================================
4837  * Encrypt and store encryption parameters
4838  * ==========================================================================
4839  */
4840 
4841 
4842 /*
4843  * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
4844  * managing the storage of encryption parameters and passing them to the
4845  * lower-level encryption functions.
4846  */
4847 static zio_t *
4848 zio_encrypt(zio_t *zio)
4849 {
4850 	zio_prop_t *zp = &zio->io_prop;
4851 	spa_t *spa = zio->io_spa;
4852 	blkptr_t *bp = zio->io_bp;
4853 	uint64_t psize = BP_GET_PSIZE(bp);
4854 	uint64_t dsobj = zio->io_bookmark.zb_objset;
4855 	dmu_object_type_t ot = BP_GET_TYPE(bp);
4856 	void *enc_buf = NULL;
4857 	abd_t *eabd = NULL;
4858 	uint8_t salt[ZIO_DATA_SALT_LEN];
4859 	uint8_t iv[ZIO_DATA_IV_LEN];
4860 	uint8_t mac[ZIO_DATA_MAC_LEN];
4861 	boolean_t no_crypt = B_FALSE;
4862 
4863 	/* the root zio already encrypted the data */
4864 	if (zio->io_child_type == ZIO_CHILD_GANG)
4865 		return (zio);
4866 
4867 	/* only ZIL blocks are re-encrypted on rewrite */
4868 	if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG)
4869 		return (zio);
4870 
4871 	if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) {
4872 		BP_SET_CRYPT(bp, B_FALSE);
4873 		return (zio);
4874 	}
4875 
4876 	/* if we are doing raw encryption set the provided encryption params */
4877 	if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) {
4878 		ASSERT0(BP_GET_LEVEL(bp));
4879 		BP_SET_CRYPT(bp, B_TRUE);
4880 		BP_SET_BYTEORDER(bp, zp->zp_byteorder);
4881 		if (ot != DMU_OT_OBJSET)
4882 			zio_crypt_encode_mac_bp(bp, zp->zp_mac);
4883 
4884 		/* dnode blocks must be written out in the provided byteorder */
4885 		if (zp->zp_byteorder != ZFS_HOST_BYTEORDER &&
4886 		    ot == DMU_OT_DNODE) {
4887 			void *bswap_buf = zio_buf_alloc(psize);
4888 			abd_t *babd = abd_get_from_buf(bswap_buf, psize);
4889 
4890 			ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4891 			abd_copy_to_buf(bswap_buf, zio->io_abd, psize);
4892 			dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf,
4893 			    psize);
4894 
4895 			abd_take_ownership_of_buf(babd, B_TRUE);
4896 			zio_push_transform(zio, babd, psize, psize, NULL);
4897 		}
4898 
4899 		if (DMU_OT_IS_ENCRYPTED(ot))
4900 			zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv);
4901 		return (zio);
4902 	}
4903 
4904 	/* indirect blocks only maintain a cksum of the lower level MACs */
4905 	if (BP_GET_LEVEL(bp) > 0) {
4906 		BP_SET_CRYPT(bp, B_TRUE);
4907 		VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE,
4908 		    zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp),
4909 		    mac));
4910 		zio_crypt_encode_mac_bp(bp, mac);
4911 		return (zio);
4912 	}
4913 
4914 	/*
4915 	 * Objset blocks are a special case since they have 2 256-bit MACs
4916 	 * embedded within them.
4917 	 */
4918 	if (ot == DMU_OT_OBJSET) {
4919 		ASSERT0(DMU_OT_IS_ENCRYPTED(ot));
4920 		ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4921 		BP_SET_CRYPT(bp, B_TRUE);
4922 		VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj,
4923 		    zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp)));
4924 		return (zio);
4925 	}
4926 
4927 	/* unencrypted object types are only authenticated with a MAC */
4928 	if (!DMU_OT_IS_ENCRYPTED(ot)) {
4929 		BP_SET_CRYPT(bp, B_TRUE);
4930 		VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj,
4931 		    zio->io_abd, psize, mac));
4932 		zio_crypt_encode_mac_bp(bp, mac);
4933 		return (zio);
4934 	}
4935 
4936 	/*
4937 	 * Later passes of sync-to-convergence may decide to rewrite data
4938 	 * in place to avoid more disk reallocations. This presents a problem
4939 	 * for encryption because this constitutes rewriting the new data with
4940 	 * the same encryption key and IV. However, this only applies to blocks
4941 	 * in the MOS (particularly the spacemaps) and we do not encrypt the
4942 	 * MOS. We assert that the zio is allocating or an intent log write
4943 	 * to enforce this.
4944 	 */
4945 	ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG);
4946 	ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG);
4947 	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION));
4948 	ASSERT3U(psize, !=, 0);
4949 
4950 	enc_buf = zio_buf_alloc(psize);
4951 	eabd = abd_get_from_buf(enc_buf, psize);
4952 	abd_take_ownership_of_buf(eabd, B_TRUE);
4953 
4954 	/*
4955 	 * For an explanation of what encryption parameters are stored
4956 	 * where, see the block comment in zio_crypt.c.
4957 	 */
4958 	if (ot == DMU_OT_INTENT_LOG) {
4959 		zio_crypt_decode_params_bp(bp, salt, iv);
4960 	} else {
4961 		BP_SET_CRYPT(bp, B_TRUE);
4962 	}
4963 
4964 	/* Perform the encryption. This should not fail */
4965 	VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark,
4966 	    BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
4967 	    salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt));
4968 
4969 	/* encode encryption metadata into the bp */
4970 	if (ot == DMU_OT_INTENT_LOG) {
4971 		/*
4972 		 * ZIL blocks store the MAC in the embedded checksum, so the
4973 		 * transform must always be applied.
4974 		 */
4975 		zio_crypt_encode_mac_zil(enc_buf, mac);
4976 		zio_push_transform(zio, eabd, psize, psize, NULL);
4977 	} else {
4978 		BP_SET_CRYPT(bp, B_TRUE);
4979 		zio_crypt_encode_params_bp(bp, salt, iv);
4980 		zio_crypt_encode_mac_bp(bp, mac);
4981 
4982 		if (no_crypt) {
4983 			ASSERT3U(ot, ==, DMU_OT_DNODE);
4984 			abd_free(eabd);
4985 		} else {
4986 			zio_push_transform(zio, eabd, psize, psize, NULL);
4987 		}
4988 	}
4989 
4990 	return (zio);
4991 }
4992 
4993 /*
4994  * ==========================================================================
4995  * Generate and verify checksums
4996  * ==========================================================================
4997  */
4998 static zio_t *
4999 zio_checksum_generate(zio_t *zio)
5000 {
5001 	blkptr_t *bp = zio->io_bp;
5002 	enum zio_checksum checksum;
5003 
5004 	if (bp == NULL) {
5005 		/*
5006 		 * This is zio_write_phys().
5007 		 * We're either generating a label checksum, or none at all.
5008 		 */
5009 		checksum = zio->io_prop.zp_checksum;
5010 
5011 		if (checksum == ZIO_CHECKSUM_OFF)
5012 			return (zio);
5013 
5014 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
5015 	} else {
5016 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
5017 			ASSERT(!IO_IS_ALLOCATING(zio));
5018 			checksum = ZIO_CHECKSUM_GANG_HEADER;
5019 		} else {
5020 			checksum = BP_GET_CHECKSUM(bp);
5021 		}
5022 	}
5023 
5024 	zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
5025 
5026 	return (zio);
5027 }
5028 
5029 static zio_t *
5030 zio_checksum_verify(zio_t *zio)
5031 {
5032 	zio_bad_cksum_t info;
5033 	blkptr_t *bp = zio->io_bp;
5034 	int error;
5035 
5036 	ASSERT(zio->io_vd != NULL);
5037 
5038 	if (bp == NULL) {
5039 		/*
5040 		 * This is zio_read_phys().
5041 		 * We're either verifying a label checksum, or nothing at all.
5042 		 */
5043 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
5044 			return (zio);
5045 
5046 		ASSERT3U(zio->io_prop.zp_checksum, ==, ZIO_CHECKSUM_LABEL);
5047 	}
5048 
5049 	ASSERT0(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR);
5050 	IMPLY(zio->io_flags & ZIO_FLAG_DIO_READ,
5051 	    !(zio->io_flags & ZIO_FLAG_SPECULATIVE));
5052 
5053 	if ((error = zio_checksum_error(zio, &info)) != 0) {
5054 		zio->io_error = error;
5055 		if (error == ECKSUM &&
5056 		    !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
5057 			if (zio->io_flags & ZIO_FLAG_DIO_READ) {
5058 				zio->io_flags |= ZIO_FLAG_DIO_CHKSUM_ERR;
5059 				zio_t *pio = zio_unique_parent(zio);
5060 				/*
5061 				 * Any Direct I/O read that has a checksum
5062 				 * error must be treated as suspicous as the
5063 				 * contents of the buffer could be getting
5064 				 * manipulated while the I/O is taking place.
5065 				 *
5066 				 * The checksum verify error will only be
5067 				 * reported here for disk and file VDEV's and
5068 				 * will be reported on those that the failure
5069 				 * occurred on. Other types of VDEV's report the
5070 				 * verify failure in their own code paths.
5071 				 */
5072 				if (pio->io_child_type == ZIO_CHILD_LOGICAL) {
5073 					zio_dio_chksum_verify_error_report(zio);
5074 				}
5075 			} else {
5076 				mutex_enter(&zio->io_vd->vdev_stat_lock);
5077 				zio->io_vd->vdev_stat.vs_checksum_errors++;
5078 				mutex_exit(&zio->io_vd->vdev_stat_lock);
5079 				(void) zfs_ereport_start_checksum(zio->io_spa,
5080 				    zio->io_vd, &zio->io_bookmark, zio,
5081 				    zio->io_offset, zio->io_size, &info);
5082 			}
5083 		}
5084 	}
5085 
5086 	return (zio);
5087 }
5088 
5089 static zio_t *
5090 zio_dio_checksum_verify(zio_t *zio)
5091 {
5092 	zio_t *pio = zio_unique_parent(zio);
5093 	int error;
5094 
5095 	ASSERT3P(zio->io_vd, !=, NULL);
5096 	ASSERT3P(zio->io_bp, !=, NULL);
5097 	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
5098 	ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
5099 	ASSERT3B(pio->io_prop.zp_direct_write, ==, B_TRUE);
5100 	ASSERT3U(pio->io_child_type, ==, ZIO_CHILD_LOGICAL);
5101 
5102 	if (zfs_vdev_direct_write_verify == 0 || zio->io_error != 0)
5103 		goto out;
5104 
5105 	if ((error = zio_checksum_error(zio, NULL)) != 0) {
5106 		zio->io_error = error;
5107 		if (error == ECKSUM) {
5108 			zio->io_flags |= ZIO_FLAG_DIO_CHKSUM_ERR;
5109 			zio_dio_chksum_verify_error_report(zio);
5110 		}
5111 	}
5112 
5113 out:
5114 	return (zio);
5115 }
5116 
5117 
5118 /*
5119  * Called by RAID-Z to ensure we don't compute the checksum twice.
5120  */
5121 void
5122 zio_checksum_verified(zio_t *zio)
5123 {
5124 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
5125 }
5126 
5127 /*
5128  * Report Direct I/O checksum verify error and create ZED event.
5129  */
5130 void
5131 zio_dio_chksum_verify_error_report(zio_t *zio)
5132 {
5133 	ASSERT(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR);
5134 
5135 	if (zio->io_child_type == ZIO_CHILD_LOGICAL)
5136 		return;
5137 
5138 	mutex_enter(&zio->io_vd->vdev_stat_lock);
5139 	zio->io_vd->vdev_stat.vs_dio_verify_errors++;
5140 	mutex_exit(&zio->io_vd->vdev_stat_lock);
5141 	if (zio->io_type == ZIO_TYPE_WRITE) {
5142 		/*
5143 		 * Convert checksum error for writes into EIO.
5144 		 */
5145 		zio->io_error = SET_ERROR(EIO);
5146 		/*
5147 		 * Report dio_verify_wr ZED event.
5148 		 */
5149 		(void) zfs_ereport_post(FM_EREPORT_ZFS_DIO_VERIFY_WR,
5150 		    zio->io_spa,  zio->io_vd, &zio->io_bookmark, zio, 0);
5151 	} else {
5152 		/*
5153 		 * Report dio_verify_rd ZED event.
5154 		 */
5155 		(void) zfs_ereport_post(FM_EREPORT_ZFS_DIO_VERIFY_RD,
5156 		    zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0);
5157 	}
5158 }
5159 
5160 /*
5161  * ==========================================================================
5162  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
5163  * An error of 0 indicates success.  ENXIO indicates whole-device failure,
5164  * which may be transient (e.g. unplugged) or permanent.  ECKSUM and EIO
5165  * indicate errors that are specific to one I/O, and most likely permanent.
5166  * Any other error is presumed to be worse because we weren't expecting it.
5167  * ==========================================================================
5168  */
5169 int
5170 zio_worst_error(int e1, int e2)
5171 {
5172 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
5173 	int r1, r2;
5174 
5175 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
5176 		if (e1 == zio_error_rank[r1])
5177 			break;
5178 
5179 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
5180 		if (e2 == zio_error_rank[r2])
5181 			break;
5182 
5183 	return (r1 > r2 ? e1 : e2);
5184 }
5185 
5186 /*
5187  * ==========================================================================
5188  * I/O completion
5189  * ==========================================================================
5190  */
5191 static zio_t *
5192 zio_ready(zio_t *zio)
5193 {
5194 	blkptr_t *bp = zio->io_bp;
5195 	zio_t *pio, *pio_next;
5196 	zio_link_t *zl = NULL;
5197 
5198 	if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
5199 	    ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT, ZIO_WAIT_READY)) {
5200 		return (NULL);
5201 	}
5202 
5203 	if (zio->io_ready) {
5204 		ASSERT(IO_IS_ALLOCATING(zio));
5205 		ASSERT(BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg ||
5206 		    BP_IS_HOLE(bp) || (zio->io_flags & ZIO_FLAG_NOPWRITE));
5207 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
5208 
5209 		zio->io_ready(zio);
5210 	}
5211 
5212 #ifdef ZFS_DEBUG
5213 	if (bp != NULL && bp != &zio->io_bp_copy)
5214 		zio->io_bp_copy = *bp;
5215 #endif
5216 
5217 	if (zio->io_error != 0) {
5218 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
5219 
5220 		if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
5221 			ASSERT(IO_IS_ALLOCATING(zio));
5222 			ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
5223 			ASSERT(zio->io_metaslab_class != NULL);
5224 			ASSERT(ZIO_HAS_ALLOCATOR(zio));
5225 
5226 			/*
5227 			 * We were unable to allocate anything, unreserve and
5228 			 * issue the next I/O to allocate.
5229 			 */
5230 			if (metaslab_class_throttle_unreserve(
5231 			    zio->io_metaslab_class, zio->io_prop.zp_copies,
5232 			    zio)) {
5233 				zio_allocate_dispatch(zio->io_metaslab_class,
5234 				    zio->io_allocator);
5235 			}
5236 		}
5237 	}
5238 
5239 	mutex_enter(&zio->io_lock);
5240 	zio->io_state[ZIO_WAIT_READY] = 1;
5241 	pio = zio_walk_parents(zio, &zl);
5242 	mutex_exit(&zio->io_lock);
5243 
5244 	/*
5245 	 * As we notify zio's parents, new parents could be added.
5246 	 * New parents go to the head of zio's io_parent_list, however,
5247 	 * so we will (correctly) not notify them.  The remainder of zio's
5248 	 * io_parent_list, from 'pio_next' onward, cannot change because
5249 	 * all parents must wait for us to be done before they can be done.
5250 	 */
5251 	for (; pio != NULL; pio = pio_next) {
5252 		pio_next = zio_walk_parents(zio, &zl);
5253 		zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL);
5254 	}
5255 
5256 	if (zio->io_flags & ZIO_FLAG_NODATA) {
5257 		if (bp != NULL && BP_IS_GANG(bp)) {
5258 			zio->io_flags &= ~ZIO_FLAG_NODATA;
5259 		} else {
5260 			ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
5261 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
5262 		}
5263 	}
5264 
5265 	if (zio_injection_enabled &&
5266 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
5267 		zio_handle_ignored_writes(zio);
5268 
5269 	return (zio);
5270 }
5271 
5272 /*
5273  * Update the allocation throttle accounting.
5274  */
5275 static void
5276 zio_dva_throttle_done(zio_t *zio)
5277 {
5278 	zio_t *pio = zio_unique_parent(zio);
5279 	vdev_t *vd = zio->io_vd;
5280 	int flags = METASLAB_ASYNC_ALLOC;
5281 	const void *tag = pio;
5282 
5283 	ASSERT3P(zio->io_bp, !=, NULL);
5284 	ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
5285 	ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
5286 	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
5287 	ASSERT(vd != NULL);
5288 	ASSERT3P(vd, ==, vd->vdev_top);
5289 	ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY));
5290 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
5291 	ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
5292 
5293 	/*
5294 	 * Parents of gang children can have two flavors -- ones that allocated
5295 	 * the gang header (will have ZIO_FLAG_IO_REWRITE set) and ones that
5296 	 * allocated the constituent blocks.  The first use their parent as tag.
5297 	 */
5298 	if (pio->io_child_type == ZIO_CHILD_GANG &&
5299 	    (pio->io_flags & ZIO_FLAG_IO_REWRITE))
5300 		tag = zio_unique_parent(pio);
5301 
5302 	ASSERT(IO_IS_ALLOCATING(pio) || (pio->io_child_type == ZIO_CHILD_GANG &&
5303 	    (pio->io_flags & ZIO_FLAG_IO_REWRITE)));
5304 	ASSERT(ZIO_HAS_ALLOCATOR(pio));
5305 	ASSERT3P(zio, !=, zio->io_logical);
5306 	ASSERT(zio->io_logical != NULL);
5307 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
5308 	ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
5309 	ASSERT(zio->io_metaslab_class != NULL);
5310 	ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled);
5311 
5312 	metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id,
5313 	    pio->io_allocator, flags, pio->io_size, tag);
5314 
5315 	if (metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1, pio)) {
5316 		zio_allocate_dispatch(zio->io_metaslab_class,
5317 		    pio->io_allocator);
5318 	}
5319 }
5320 
5321 static zio_t *
5322 zio_done(zio_t *zio)
5323 {
5324 	/*
5325 	 * Always attempt to keep stack usage minimal here since
5326 	 * we can be called recursively up to 19 levels deep.
5327 	 */
5328 	const uint64_t psize = zio->io_size;
5329 	zio_t *pio, *pio_next;
5330 	zio_link_t *zl = NULL;
5331 
5332 	/*
5333 	 * If our children haven't all completed,
5334 	 * wait for them and then repeat this pipeline stage.
5335 	 */
5336 	if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
5337 		return (NULL);
5338 	}
5339 
5340 	/*
5341 	 * If the allocation throttle is enabled, then update the accounting.
5342 	 * We only track child I/Os that are part of an allocating async
5343 	 * write. We must do this since the allocation is performed
5344 	 * by the logical I/O but the actual write is done by child I/Os.
5345 	 */
5346 	if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
5347 	    zio->io_child_type == ZIO_CHILD_VDEV)
5348 		zio_dva_throttle_done(zio);
5349 
5350 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
5351 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
5352 			ASSERT(zio->io_children[c][w] == 0);
5353 
5354 	if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) {
5355 		ASSERT(zio->io_bp->blk_pad[0] == 0);
5356 		ASSERT(zio->io_bp->blk_pad[1] == 0);
5357 		ASSERT(memcmp(zio->io_bp, &zio->io_bp_copy,
5358 		    sizeof (blkptr_t)) == 0 ||
5359 		    (zio->io_bp == zio_unique_parent(zio)->io_bp));
5360 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
5361 		    zio->io_bp_override == NULL &&
5362 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
5363 			ASSERT3U(zio->io_prop.zp_copies, <=,
5364 			    BP_GET_NDVAS(zio->io_bp));
5365 			ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
5366 			    (BP_COUNT_GANG(zio->io_bp) ==
5367 			    BP_GET_NDVAS(zio->io_bp)));
5368 		}
5369 		if (zio->io_flags & ZIO_FLAG_NOPWRITE)
5370 			VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig));
5371 	}
5372 
5373 	/*
5374 	 * If there were child vdev/gang/ddt errors, they apply to us now.
5375 	 */
5376 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
5377 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
5378 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
5379 
5380 	/*
5381 	 * If the I/O on the transformed data was successful, generate any
5382 	 * checksum reports now while we still have the transformed data.
5383 	 */
5384 	if (zio->io_error == 0) {
5385 		while (zio->io_cksum_report != NULL) {
5386 			zio_cksum_report_t *zcr = zio->io_cksum_report;
5387 			uint64_t align = zcr->zcr_align;
5388 			uint64_t asize = P2ROUNDUP(psize, align);
5389 			abd_t *adata = zio->io_abd;
5390 
5391 			if (adata != NULL && asize != psize) {
5392 				adata = abd_alloc(asize, B_TRUE);
5393 				abd_copy(adata, zio->io_abd, psize);
5394 				abd_zero_off(adata, psize, asize - psize);
5395 			}
5396 
5397 			zio->io_cksum_report = zcr->zcr_next;
5398 			zcr->zcr_next = NULL;
5399 			zcr->zcr_finish(zcr, adata);
5400 			zfs_ereport_free_checksum(zcr);
5401 
5402 			if (adata != NULL && asize != psize)
5403 				abd_free(adata);
5404 		}
5405 	}
5406 
5407 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
5408 
5409 	vdev_stat_update(zio, psize);
5410 
5411 	/*
5412 	 * If this I/O is attached to a particular vdev is slow, exceeding
5413 	 * 30 seconds to complete, post an error described the I/O delay.
5414 	 * We ignore these errors if the device is currently unavailable.
5415 	 */
5416 	if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) {
5417 		if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) {
5418 			/*
5419 			 * We want to only increment our slow IO counters if
5420 			 * the IO is valid (i.e. not if the drive is removed).
5421 			 *
5422 			 * zfs_ereport_post() will also do these checks, but
5423 			 * it can also ratelimit and have other failures, so we
5424 			 * need to increment the slow_io counters independent
5425 			 * of it.
5426 			 */
5427 			if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY,
5428 			    zio->io_spa, zio->io_vd, zio)) {
5429 				mutex_enter(&zio->io_vd->vdev_stat_lock);
5430 				zio->io_vd->vdev_stat.vs_slow_ios++;
5431 				mutex_exit(&zio->io_vd->vdev_stat_lock);
5432 
5433 				(void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY,
5434 				    zio->io_spa, zio->io_vd, &zio->io_bookmark,
5435 				    zio, 0);
5436 			}
5437 		}
5438 	}
5439 
5440 	if (zio->io_error) {
5441 		/*
5442 		 * If this I/O is attached to a particular vdev,
5443 		 * generate an error message describing the I/O failure
5444 		 * at the block level.  We ignore these errors if the
5445 		 * device is currently unavailable.
5446 		 */
5447 		if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
5448 		    !vdev_is_dead(zio->io_vd) &&
5449 		    !(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR)) {
5450 			int ret = zfs_ereport_post(FM_EREPORT_ZFS_IO,
5451 			    zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0);
5452 			if (ret != EALREADY) {
5453 				mutex_enter(&zio->io_vd->vdev_stat_lock);
5454 				if (zio->io_type == ZIO_TYPE_READ)
5455 					zio->io_vd->vdev_stat.vs_read_errors++;
5456 				else if (zio->io_type == ZIO_TYPE_WRITE)
5457 					zio->io_vd->vdev_stat.vs_write_errors++;
5458 				mutex_exit(&zio->io_vd->vdev_stat_lock);
5459 			}
5460 		}
5461 
5462 		if ((zio->io_error == EIO || !(zio->io_flags &
5463 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
5464 		    !(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR) &&
5465 		    zio == zio->io_logical) {
5466 			/*
5467 			 * For logical I/O requests, tell the SPA to log the
5468 			 * error and generate a logical data ereport.
5469 			 */
5470 			spa_log_error(zio->io_spa, &zio->io_bookmark,
5471 			    BP_GET_LOGICAL_BIRTH(zio->io_bp));
5472 			(void) zfs_ereport_post(FM_EREPORT_ZFS_DATA,
5473 			    zio->io_spa, NULL, &zio->io_bookmark, zio, 0);
5474 		}
5475 	}
5476 
5477 	if (zio->io_error && zio == zio->io_logical) {
5478 
5479 		/*
5480 		 * A DDT child tried to create a mixed gang/non-gang BP. We're
5481 		 * going to have to just retry as a non-dedup IO.
5482 		 */
5483 		if (zio->io_error == EAGAIN && IO_IS_ALLOCATING(zio) &&
5484 		    zio->io_prop.zp_dedup) {
5485 			zio->io_reexecute |= ZIO_REEXECUTE_NOW;
5486 			zio->io_prop.zp_dedup = B_FALSE;
5487 		}
5488 		/*
5489 		 * Determine whether zio should be reexecuted.  This will
5490 		 * propagate all the way to the root via zio_notify_parent().
5491 		 */
5492 		ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
5493 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
5494 
5495 		if (IO_IS_ALLOCATING(zio) &&
5496 		    !(zio->io_flags & ZIO_FLAG_CANFAIL) &&
5497 		    !(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR)) {
5498 			if (zio->io_error != ENOSPC)
5499 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
5500 			else
5501 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
5502 		}
5503 
5504 		if ((zio->io_type == ZIO_TYPE_READ ||
5505 		    zio->io_type == ZIO_TYPE_FREE) &&
5506 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
5507 		    zio->io_error == ENXIO &&
5508 		    spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
5509 		    spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
5510 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
5511 
5512 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
5513 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
5514 
5515 		/*
5516 		 * Here is a possibly good place to attempt to do
5517 		 * either combinatorial reconstruction or error correction
5518 		 * based on checksums.  It also might be a good place
5519 		 * to send out preliminary ereports before we suspend
5520 		 * processing.
5521 		 */
5522 	}
5523 
5524 	/*
5525 	 * If there were logical child errors, they apply to us now.
5526 	 * We defer this until now to avoid conflating logical child
5527 	 * errors with errors that happened to the zio itself when
5528 	 * updating vdev stats and reporting FMA events above.
5529 	 */
5530 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
5531 
5532 	if ((zio->io_error || zio->io_reexecute) &&
5533 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
5534 	    !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
5535 		zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
5536 
5537 	zio_gang_tree_free(&zio->io_gang_tree);
5538 
5539 	/*
5540 	 * Godfather I/Os should never suspend.
5541 	 */
5542 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
5543 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
5544 		zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND;
5545 
5546 	if (zio->io_reexecute) {
5547 		/*
5548 		 * A Direct I/O operation that has a checksum verify error
5549 		 * should not attempt to reexecute. Instead, the error should
5550 		 * just be propagated back.
5551 		 */
5552 		ASSERT(!(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR));
5553 
5554 		/*
5555 		 * This is a logical I/O that wants to reexecute.
5556 		 *
5557 		 * Reexecute is top-down.  When an i/o fails, if it's not
5558 		 * the root, it simply notifies its parent and sticks around.
5559 		 * The parent, seeing that it still has children in zio_done(),
5560 		 * does the same.  This percolates all the way up to the root.
5561 		 * The root i/o will reexecute or suspend the entire tree.
5562 		 *
5563 		 * This approach ensures that zio_reexecute() honors
5564 		 * all the original i/o dependency relationships, e.g.
5565 		 * parents not executing until children are ready.
5566 		 */
5567 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
5568 
5569 		zio->io_gang_leader = NULL;
5570 
5571 		mutex_enter(&zio->io_lock);
5572 		zio->io_state[ZIO_WAIT_DONE] = 1;
5573 		mutex_exit(&zio->io_lock);
5574 
5575 		/*
5576 		 * "The Godfather" I/O monitors its children but is
5577 		 * not a true parent to them. It will track them through
5578 		 * the pipeline but severs its ties whenever they get into
5579 		 * trouble (e.g. suspended). This allows "The Godfather"
5580 		 * I/O to return status without blocking.
5581 		 */
5582 		zl = NULL;
5583 		for (pio = zio_walk_parents(zio, &zl); pio != NULL;
5584 		    pio = pio_next) {
5585 			zio_link_t *remove_zl = zl;
5586 			pio_next = zio_walk_parents(zio, &zl);
5587 
5588 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
5589 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
5590 				zio_remove_child(pio, zio, remove_zl);
5591 				/*
5592 				 * This is a rare code path, so we don't
5593 				 * bother with "next_to_execute".
5594 				 */
5595 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE,
5596 				    NULL);
5597 			}
5598 		}
5599 
5600 		if ((pio = zio_unique_parent(zio)) != NULL) {
5601 			/*
5602 			 * We're not a root i/o, so there's nothing to do
5603 			 * but notify our parent.  Don't propagate errors
5604 			 * upward since we haven't permanently failed yet.
5605 			 */
5606 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
5607 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
5608 			/*
5609 			 * This is a rare code path, so we don't bother with
5610 			 * "next_to_execute".
5611 			 */
5612 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL);
5613 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
5614 			/*
5615 			 * We'd fail again if we reexecuted now, so suspend
5616 			 * until conditions improve (e.g. device comes online).
5617 			 */
5618 			zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
5619 		} else {
5620 			/*
5621 			 * Reexecution is potentially a huge amount of work.
5622 			 * Hand it off to the otherwise-unused claim taskq.
5623 			 */
5624 			spa_taskq_dispatch(zio->io_spa,
5625 			    ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
5626 			    zio_reexecute, zio, B_FALSE);
5627 		}
5628 		return (NULL);
5629 	}
5630 
5631 	ASSERT(list_is_empty(&zio->io_child_list));
5632 	ASSERT(zio->io_reexecute == 0);
5633 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
5634 
5635 	/*
5636 	 * Report any checksum errors, since the I/O is complete.
5637 	 */
5638 	while (zio->io_cksum_report != NULL) {
5639 		zio_cksum_report_t *zcr = zio->io_cksum_report;
5640 		zio->io_cksum_report = zcr->zcr_next;
5641 		zcr->zcr_next = NULL;
5642 		zcr->zcr_finish(zcr, NULL);
5643 		zfs_ereport_free_checksum(zcr);
5644 	}
5645 
5646 	/*
5647 	 * It is the responsibility of the done callback to ensure that this
5648 	 * particular zio is no longer discoverable for adoption, and as
5649 	 * such, cannot acquire any new parents.
5650 	 */
5651 	if (zio->io_done)
5652 		zio->io_done(zio);
5653 
5654 	mutex_enter(&zio->io_lock);
5655 	zio->io_state[ZIO_WAIT_DONE] = 1;
5656 	mutex_exit(&zio->io_lock);
5657 
5658 	/*
5659 	 * We are done executing this zio.  We may want to execute a parent
5660 	 * next.  See the comment in zio_notify_parent().
5661 	 */
5662 	zio_t *next_to_execute = NULL;
5663 	zl = NULL;
5664 	for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
5665 		zio_link_t *remove_zl = zl;
5666 		pio_next = zio_walk_parents(zio, &zl);
5667 		zio_remove_child(pio, zio, remove_zl);
5668 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute);
5669 	}
5670 
5671 	if (zio->io_waiter != NULL) {
5672 		mutex_enter(&zio->io_lock);
5673 		zio->io_executor = NULL;
5674 		cv_broadcast(&zio->io_cv);
5675 		mutex_exit(&zio->io_lock);
5676 	} else {
5677 		zio_destroy(zio);
5678 	}
5679 
5680 	return (next_to_execute);
5681 }
5682 
5683 /*
5684  * ==========================================================================
5685  * I/O pipeline definition
5686  * ==========================================================================
5687  */
5688 static zio_pipe_stage_t *zio_pipeline[] = {
5689 	NULL,
5690 	zio_read_bp_init,
5691 	zio_write_bp_init,
5692 	zio_free_bp_init,
5693 	zio_issue_async,
5694 	zio_write_compress,
5695 	zio_encrypt,
5696 	zio_checksum_generate,
5697 	zio_nop_write,
5698 	zio_brt_free,
5699 	zio_ddt_read_start,
5700 	zio_ddt_read_done,
5701 	zio_ddt_write,
5702 	zio_ddt_free,
5703 	zio_gang_assemble,
5704 	zio_gang_issue,
5705 	zio_dva_throttle,
5706 	zio_dva_allocate,
5707 	zio_dva_free,
5708 	zio_dva_claim,
5709 	zio_ready,
5710 	zio_vdev_io_start,
5711 	zio_vdev_io_done,
5712 	zio_vdev_io_assess,
5713 	zio_checksum_verify,
5714 	zio_dio_checksum_verify,
5715 	zio_done
5716 };
5717 
5718 
5719 
5720 
5721 /*
5722  * Compare two zbookmark_phys_t's to see which we would reach first in a
5723  * pre-order traversal of the object tree.
5724  *
5725  * This is simple in every case aside from the meta-dnode object. For all other
5726  * objects, we traverse them in order (object 1 before object 2, and so on).
5727  * However, all of these objects are traversed while traversing object 0, since
5728  * the data it points to is the list of objects.  Thus, we need to convert to a
5729  * canonical representation so we can compare meta-dnode bookmarks to
5730  * non-meta-dnode bookmarks.
5731  *
5732  * We do this by calculating "equivalents" for each field of the zbookmark.
5733  * zbookmarks outside of the meta-dnode use their own object and level, and
5734  * calculate the level 0 equivalent (the first L0 blkid that is contained in the
5735  * blocks this bookmark refers to) by multiplying their blkid by their span
5736  * (the number of L0 blocks contained within one block at their level).
5737  * zbookmarks inside the meta-dnode calculate their object equivalent
5738  * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
5739  * level + 1<<31 (any value larger than a level could ever be) for their level.
5740  * This causes them to always compare before a bookmark in their object
5741  * equivalent, compare appropriately to bookmarks in other objects, and to
5742  * compare appropriately to other bookmarks in the meta-dnode.
5743  */
5744 int
5745 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
5746     const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
5747 {
5748 	/*
5749 	 * These variables represent the "equivalent" values for the zbookmark,
5750 	 * after converting zbookmarks inside the meta dnode to their
5751 	 * normal-object equivalents.
5752 	 */
5753 	uint64_t zb1obj, zb2obj;
5754 	uint64_t zb1L0, zb2L0;
5755 	uint64_t zb1level, zb2level;
5756 
5757 	if (zb1->zb_object == zb2->zb_object &&
5758 	    zb1->zb_level == zb2->zb_level &&
5759 	    zb1->zb_blkid == zb2->zb_blkid)
5760 		return (0);
5761 
5762 	IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT);
5763 	IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT);
5764 
5765 	/*
5766 	 * BP_SPANB calculates the span in blocks.
5767 	 */
5768 	zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
5769 	zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
5770 
5771 	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
5772 		zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
5773 		zb1L0 = 0;
5774 		zb1level = zb1->zb_level + COMPARE_META_LEVEL;
5775 	} else {
5776 		zb1obj = zb1->zb_object;
5777 		zb1level = zb1->zb_level;
5778 	}
5779 
5780 	if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
5781 		zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
5782 		zb2L0 = 0;
5783 		zb2level = zb2->zb_level + COMPARE_META_LEVEL;
5784 	} else {
5785 		zb2obj = zb2->zb_object;
5786 		zb2level = zb2->zb_level;
5787 	}
5788 
5789 	/* Now that we have a canonical representation, do the comparison. */
5790 	if (zb1obj != zb2obj)
5791 		return (zb1obj < zb2obj ? -1 : 1);
5792 	else if (zb1L0 != zb2L0)
5793 		return (zb1L0 < zb2L0 ? -1 : 1);
5794 	else if (zb1level != zb2level)
5795 		return (zb1level > zb2level ? -1 : 1);
5796 	/*
5797 	 * This can (theoretically) happen if the bookmarks have the same object
5798 	 * and level, but different blkids, if the block sizes are not the same.
5799 	 * There is presently no way to change the indirect block sizes
5800 	 */
5801 	return (0);
5802 }
5803 
5804 /*
5805  *  This function checks the following: given that last_block is the place that
5806  *  our traversal stopped last time, does that guarantee that we've visited
5807  *  every node under subtree_root?  Therefore, we can't just use the raw output
5808  *  of zbookmark_compare.  We have to pass in a modified version of
5809  *  subtree_root; by incrementing the block id, and then checking whether
5810  *  last_block is before or equal to that, we can tell whether or not having
5811  *  visited last_block implies that all of subtree_root's children have been
5812  *  visited.
5813  */
5814 boolean_t
5815 zbookmark_subtree_completed(const dnode_phys_t *dnp,
5816     const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
5817 {
5818 	zbookmark_phys_t mod_zb = *subtree_root;
5819 	mod_zb.zb_blkid++;
5820 	ASSERT0(last_block->zb_level);
5821 
5822 	/* The objset_phys_t isn't before anything. */
5823 	if (dnp == NULL)
5824 		return (B_FALSE);
5825 
5826 	/*
5827 	 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
5828 	 * data block size in sectors, because that variable is only used if
5829 	 * the bookmark refers to a block in the meta-dnode.  Since we don't
5830 	 * know without examining it what object it refers to, and there's no
5831 	 * harm in passing in this value in other cases, we always pass it in.
5832 	 *
5833 	 * We pass in 0 for the indirect block size shift because zb2 must be
5834 	 * level 0.  The indirect block size is only used to calculate the span
5835 	 * of the bookmark, but since the bookmark must be level 0, the span is
5836 	 * always 1, so the math works out.
5837 	 *
5838 	 * If you make changes to how the zbookmark_compare code works, be sure
5839 	 * to make sure that this code still works afterwards.
5840 	 */
5841 	return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
5842 	    1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
5843 	    last_block) <= 0);
5844 }
5845 
5846 /*
5847  * This function is similar to zbookmark_subtree_completed(), but returns true
5848  * if subtree_root is equal or ahead of last_block, i.e. still to be done.
5849  */
5850 boolean_t
5851 zbookmark_subtree_tbd(const dnode_phys_t *dnp,
5852     const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
5853 {
5854 	ASSERT0(last_block->zb_level);
5855 	if (dnp == NULL)
5856 		return (B_FALSE);
5857 	return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
5858 	    1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, subtree_root,
5859 	    last_block) >= 0);
5860 }
5861 
5862 EXPORT_SYMBOL(zio_type_name);
5863 EXPORT_SYMBOL(zio_buf_alloc);
5864 EXPORT_SYMBOL(zio_data_buf_alloc);
5865 EXPORT_SYMBOL(zio_buf_free);
5866 EXPORT_SYMBOL(zio_data_buf_free);
5867 
5868 ZFS_MODULE_PARAM(zfs_zio, zio_, slow_io_ms, INT, ZMOD_RW,
5869 	"Max I/O completion time (milliseconds) before marking it as slow");
5870 
5871 ZFS_MODULE_PARAM(zfs_zio, zio_, requeue_io_start_cut_in_line, INT, ZMOD_RW,
5872 	"Prioritize requeued I/O");
5873 
5874 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_deferred_free,  UINT, ZMOD_RW,
5875 	"Defer frees starting in this pass");
5876 
5877 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_dont_compress, UINT, ZMOD_RW,
5878 	"Don't compress starting in this pass");
5879 
5880 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_rewrite, UINT, ZMOD_RW,
5881 	"Rewrite new bps starting in this pass");
5882 
5883 ZFS_MODULE_PARAM(zfs_zio, zio_, dva_throttle_enabled, INT, ZMOD_RW,
5884 	"Throttle block allocations in the ZIO pipeline");
5885 
5886 ZFS_MODULE_PARAM(zfs_zio, zio_, deadman_log_all, INT, ZMOD_RW,
5887 	"Log all slow ZIOs, not just those with vdevs");
5888