xref: /freebsd/sys/contrib/openzfs/module/zfs/zio.c (revision 0ce9a414adc33af29607adbd81e0760e014fcd76)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or https://opensource.org/licenses/CDDL-1.0.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2011, 2022 by Delphix. All rights reserved.
24  * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25  * Copyright (c) 2017, Intel Corporation.
26  * Copyright (c) 2019, 2023, 2024, 2025, Klara, Inc.
27  * Copyright (c) 2019, Allan Jude
28  * Copyright (c) 2021, Datto, Inc.
29  * Copyright (c) 2021, 2024 by George Melikov. All rights reserved.
30  */
31 
32 #include <sys/sysmacros.h>
33 #include <sys/zfs_context.h>
34 #include <sys/fm/fs/zfs.h>
35 #include <sys/spa.h>
36 #include <sys/txg.h>
37 #include <sys/spa_impl.h>
38 #include <sys/vdev_impl.h>
39 #include <sys/vdev_trim.h>
40 #include <sys/zio_impl.h>
41 #include <sys/zio_compress.h>
42 #include <sys/zio_checksum.h>
43 #include <sys/dmu_objset.h>
44 #include <sys/arc.h>
45 #include <sys/brt.h>
46 #include <sys/ddt.h>
47 #include <sys/blkptr.h>
48 #include <sys/zfeature.h>
49 #include <sys/dsl_scan.h>
50 #include <sys/metaslab_impl.h>
51 #include <sys/time.h>
52 #include <sys/trace_zfs.h>
53 #include <sys/abd.h>
54 #include <sys/dsl_crypt.h>
55 #include <cityhash.h>
56 
57 /*
58  * ==========================================================================
59  * I/O type descriptions
60  * ==========================================================================
61  */
62 const char *const zio_type_name[ZIO_TYPES] = {
63 	/*
64 	 * Note: Linux kernel thread name length is limited
65 	 * so these names will differ from upstream open zfs.
66 	 */
67 	"z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_flush", "z_trim"
68 };
69 
70 int zio_dva_throttle_enabled = B_TRUE;
71 static int zio_deadman_log_all = B_FALSE;
72 
73 /*
74  * ==========================================================================
75  * I/O kmem caches
76  * ==========================================================================
77  */
78 static kmem_cache_t *zio_cache;
79 static kmem_cache_t *zio_link_cache;
80 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
81 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
82 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
83 static uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
84 static uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
85 #endif
86 
87 /* Mark IOs as "slow" if they take longer than 30 seconds */
88 static uint_t zio_slow_io_ms = (30 * MILLISEC);
89 
90 #define	BP_SPANB(indblkshift, level) \
91 	(((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
92 #define	COMPARE_META_LEVEL	0x80000000ul
93 /*
94  * The following actions directly effect the spa's sync-to-convergence logic.
95  * The values below define the sync pass when we start performing the action.
96  * Care should be taken when changing these values as they directly impact
97  * spa_sync() performance. Tuning these values may introduce subtle performance
98  * pathologies and should only be done in the context of performance analysis.
99  * These tunables will eventually be removed and replaced with #defines once
100  * enough analysis has been done to determine optimal values.
101  *
102  * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
103  * regular blocks are not deferred.
104  *
105  * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable
106  * compression (including of metadata).  In practice, we don't have this
107  * many sync passes, so this has no effect.
108  *
109  * The original intent was that disabling compression would help the sync
110  * passes to converge. However, in practice disabling compression increases
111  * the average number of sync passes, because when we turn compression off, a
112  * lot of block's size will change and thus we have to re-allocate (not
113  * overwrite) them. It also increases the number of 128KB allocations (e.g.
114  * for indirect blocks and spacemaps) because these will not be compressed.
115  * The 128K allocations are especially detrimental to performance on highly
116  * fragmented systems, which may have very few free segments of this size,
117  * and may need to load new metaslabs to satisfy 128K allocations.
118  */
119 
120 /* defer frees starting in this pass */
121 uint_t zfs_sync_pass_deferred_free = 2;
122 
123 /* don't compress starting in this pass */
124 static uint_t zfs_sync_pass_dont_compress = 8;
125 
126 /* rewrite new bps starting in this pass */
127 static uint_t zfs_sync_pass_rewrite = 2;
128 
129 /*
130  * An allocating zio is one that either currently has the DVA allocate
131  * stage set or will have it later in its lifetime.
132  */
133 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
134 
135 /*
136  * Enable smaller cores by excluding metadata
137  * allocations as well.
138  */
139 int zio_exclude_metadata = 0;
140 static int zio_requeue_io_start_cut_in_line = 1;
141 
142 #ifdef ZFS_DEBUG
143 static const int zio_buf_debug_limit = 16384;
144 #else
145 static const int zio_buf_debug_limit = 0;
146 #endif
147 
148 typedef struct zio_stats {
149 	kstat_named_t ziostat_total_allocations;
150 	kstat_named_t ziostat_alloc_class_fallbacks;
151 	kstat_named_t ziostat_gang_writes;
152 	kstat_named_t ziostat_gang_multilevel;
153 } zio_stats_t;
154 
155 static zio_stats_t zio_stats = {
156 	{ "total_allocations",	KSTAT_DATA_UINT64 },
157 	{ "alloc_class_fallbacks",	KSTAT_DATA_UINT64 },
158 	{ "gang_writes",	KSTAT_DATA_UINT64 },
159 	{ "gang_multilevel",	KSTAT_DATA_UINT64 },
160 };
161 
162 struct {
163 	wmsum_t ziostat_total_allocations;
164 	wmsum_t ziostat_alloc_class_fallbacks;
165 	wmsum_t ziostat_gang_writes;
166 	wmsum_t ziostat_gang_multilevel;
167 } ziostat_sums;
168 
169 #define	ZIOSTAT_BUMP(stat)	wmsum_add(&ziostat_sums.stat, 1);
170 
171 static kstat_t *zio_ksp;
172 
173 static inline void __zio_execute(zio_t *zio);
174 
175 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
176 
177 static int
178 zio_kstats_update(kstat_t *ksp, int rw)
179 {
180 	zio_stats_t *zs = ksp->ks_data;
181 	if (rw == KSTAT_WRITE)
182 		return (EACCES);
183 
184 	zs->ziostat_total_allocations.value.ui64 =
185 	    wmsum_value(&ziostat_sums.ziostat_total_allocations);
186 	zs->ziostat_alloc_class_fallbacks.value.ui64 =
187 	    wmsum_value(&ziostat_sums.ziostat_alloc_class_fallbacks);
188 	zs->ziostat_gang_writes.value.ui64 =
189 	    wmsum_value(&ziostat_sums.ziostat_gang_writes);
190 	zs->ziostat_gang_multilevel.value.ui64 =
191 	    wmsum_value(&ziostat_sums.ziostat_gang_multilevel);
192 	return (0);
193 }
194 
195 void
196 zio_init(void)
197 {
198 	size_t c;
199 
200 	zio_cache = kmem_cache_create("zio_cache",
201 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
202 	zio_link_cache = kmem_cache_create("zio_link_cache",
203 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
204 
205 	wmsum_init(&ziostat_sums.ziostat_total_allocations, 0);
206 	wmsum_init(&ziostat_sums.ziostat_alloc_class_fallbacks, 0);
207 	wmsum_init(&ziostat_sums.ziostat_gang_writes, 0);
208 	wmsum_init(&ziostat_sums.ziostat_gang_multilevel, 0);
209 	zio_ksp = kstat_create("zfs", 0, "zio_stats",
210 	    "misc", KSTAT_TYPE_NAMED, sizeof (zio_stats) /
211 	    sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
212 	if (zio_ksp != NULL) {
213 		zio_ksp->ks_data = &zio_stats;
214 		zio_ksp->ks_update = zio_kstats_update;
215 		kstat_install(zio_ksp);
216 	}
217 
218 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
219 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
220 		size_t align, cflags, data_cflags;
221 		char name[32];
222 
223 		/*
224 		 * Create cache for each half-power of 2 size, starting from
225 		 * SPA_MINBLOCKSIZE.  It should give us memory space efficiency
226 		 * of ~7/8, sufficient for transient allocations mostly using
227 		 * these caches.
228 		 */
229 		size_t p2 = size;
230 		while (!ISP2(p2))
231 			p2 &= p2 - 1;
232 		if (!IS_P2ALIGNED(size, p2 / 2))
233 			continue;
234 
235 #ifndef _KERNEL
236 		/*
237 		 * If we are using watchpoints, put each buffer on its own page,
238 		 * to eliminate the performance overhead of trapping to the
239 		 * kernel when modifying a non-watched buffer that shares the
240 		 * page with a watched buffer.
241 		 */
242 		if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
243 			continue;
244 #endif
245 
246 		if (IS_P2ALIGNED(size, PAGESIZE))
247 			align = PAGESIZE;
248 		else
249 			align = 1 << (highbit64(size ^ (size - 1)) - 1);
250 
251 		cflags = (zio_exclude_metadata || size > zio_buf_debug_limit) ?
252 		    KMC_NODEBUG : 0;
253 		data_cflags = KMC_NODEBUG;
254 		if (abd_size_alloc_linear(size)) {
255 			cflags |= KMC_RECLAIMABLE;
256 			data_cflags |= KMC_RECLAIMABLE;
257 		}
258 		if (cflags == data_cflags) {
259 			/*
260 			 * Resulting kmem caches would be identical.
261 			 * Save memory by creating only one.
262 			 */
263 			(void) snprintf(name, sizeof (name),
264 			    "zio_buf_comb_%lu", (ulong_t)size);
265 			zio_buf_cache[c] = kmem_cache_create(name, size, align,
266 			    NULL, NULL, NULL, NULL, NULL, cflags);
267 			zio_data_buf_cache[c] = zio_buf_cache[c];
268 			continue;
269 		}
270 		(void) snprintf(name, sizeof (name), "zio_buf_%lu",
271 		    (ulong_t)size);
272 		zio_buf_cache[c] = kmem_cache_create(name, size, align,
273 		    NULL, NULL, NULL, NULL, NULL, cflags);
274 
275 		(void) snprintf(name, sizeof (name), "zio_data_buf_%lu",
276 		    (ulong_t)size);
277 		zio_data_buf_cache[c] = kmem_cache_create(name, size, align,
278 		    NULL, NULL, NULL, NULL, NULL, data_cflags);
279 	}
280 
281 	while (--c != 0) {
282 		ASSERT(zio_buf_cache[c] != NULL);
283 		if (zio_buf_cache[c - 1] == NULL)
284 			zio_buf_cache[c - 1] = zio_buf_cache[c];
285 
286 		ASSERT(zio_data_buf_cache[c] != NULL);
287 		if (zio_data_buf_cache[c - 1] == NULL)
288 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
289 	}
290 
291 	zio_inject_init();
292 
293 	lz4_init();
294 }
295 
296 void
297 zio_fini(void)
298 {
299 	size_t n = SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT;
300 
301 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
302 	for (size_t i = 0; i < n; i++) {
303 		if (zio_buf_cache_allocs[i] != zio_buf_cache_frees[i])
304 			(void) printf("zio_fini: [%d] %llu != %llu\n",
305 			    (int)((i + 1) << SPA_MINBLOCKSHIFT),
306 			    (long long unsigned)zio_buf_cache_allocs[i],
307 			    (long long unsigned)zio_buf_cache_frees[i]);
308 	}
309 #endif
310 
311 	/*
312 	 * The same kmem cache can show up multiple times in both zio_buf_cache
313 	 * and zio_data_buf_cache. Do a wasteful but trivially correct scan to
314 	 * sort it out.
315 	 */
316 	for (size_t i = 0; i < n; i++) {
317 		kmem_cache_t *cache = zio_buf_cache[i];
318 		if (cache == NULL)
319 			continue;
320 		for (size_t j = i; j < n; j++) {
321 			if (cache == zio_buf_cache[j])
322 				zio_buf_cache[j] = NULL;
323 			if (cache == zio_data_buf_cache[j])
324 				zio_data_buf_cache[j] = NULL;
325 		}
326 		kmem_cache_destroy(cache);
327 	}
328 
329 	for (size_t i = 0; i < n; i++) {
330 		kmem_cache_t *cache = zio_data_buf_cache[i];
331 		if (cache == NULL)
332 			continue;
333 		for (size_t j = i; j < n; j++) {
334 			if (cache == zio_data_buf_cache[j])
335 				zio_data_buf_cache[j] = NULL;
336 		}
337 		kmem_cache_destroy(cache);
338 	}
339 
340 	for (size_t i = 0; i < n; i++) {
341 		VERIFY3P(zio_buf_cache[i], ==, NULL);
342 		VERIFY3P(zio_data_buf_cache[i], ==, NULL);
343 	}
344 
345 	if (zio_ksp != NULL) {
346 		kstat_delete(zio_ksp);
347 		zio_ksp = NULL;
348 	}
349 
350 	wmsum_fini(&ziostat_sums.ziostat_total_allocations);
351 	wmsum_fini(&ziostat_sums.ziostat_alloc_class_fallbacks);
352 	wmsum_fini(&ziostat_sums.ziostat_gang_writes);
353 	wmsum_fini(&ziostat_sums.ziostat_gang_multilevel);
354 
355 	kmem_cache_destroy(zio_link_cache);
356 	kmem_cache_destroy(zio_cache);
357 
358 	zio_inject_fini();
359 
360 	lz4_fini();
361 }
362 
363 /*
364  * ==========================================================================
365  * Allocate and free I/O buffers
366  * ==========================================================================
367  */
368 
369 #if defined(ZFS_DEBUG) && defined(_KERNEL)
370 #define	ZFS_ZIO_BUF_CANARY	1
371 #endif
372 
373 #ifdef ZFS_ZIO_BUF_CANARY
374 static const ulong_t zio_buf_canary = (ulong_t)0xdeadc0dedead210b;
375 
376 /*
377  * Use empty space after the buffer to detect overflows.
378  *
379  * Since zio_init() creates kmem caches only for certain set of buffer sizes,
380  * allocations of different sizes may have some unused space after the data.
381  * Filling part of that space with a known pattern on allocation and checking
382  * it on free should allow us to detect some buffer overflows.
383  */
384 static void
385 zio_buf_put_canary(ulong_t *p, size_t size, kmem_cache_t **cache, size_t c)
386 {
387 	size_t off = P2ROUNDUP(size, sizeof (ulong_t));
388 	ulong_t *canary = p + off / sizeof (ulong_t);
389 	size_t asize = (c + 1) << SPA_MINBLOCKSHIFT;
390 	if (c + 1 < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT &&
391 	    cache[c] == cache[c + 1])
392 		asize = (c + 2) << SPA_MINBLOCKSHIFT;
393 	for (; off < asize; canary++, off += sizeof (ulong_t))
394 		*canary = zio_buf_canary;
395 }
396 
397 static void
398 zio_buf_check_canary(ulong_t *p, size_t size, kmem_cache_t **cache, size_t c)
399 {
400 	size_t off = P2ROUNDUP(size, sizeof (ulong_t));
401 	ulong_t *canary = p + off / sizeof (ulong_t);
402 	size_t asize = (c + 1) << SPA_MINBLOCKSHIFT;
403 	if (c + 1 < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT &&
404 	    cache[c] == cache[c + 1])
405 		asize = (c + 2) << SPA_MINBLOCKSHIFT;
406 	for (; off < asize; canary++, off += sizeof (ulong_t)) {
407 		if (unlikely(*canary != zio_buf_canary)) {
408 			PANIC("ZIO buffer overflow %p (%zu) + %zu %#lx != %#lx",
409 			    p, size, (canary - p) * sizeof (ulong_t),
410 			    *canary, zio_buf_canary);
411 		}
412 	}
413 }
414 #endif
415 
416 /*
417  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
418  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
419  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
420  * excess / transient data in-core during a crashdump.
421  */
422 void *
423 zio_buf_alloc(size_t size)
424 {
425 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
426 
427 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
428 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
429 	atomic_add_64(&zio_buf_cache_allocs[c], 1);
430 #endif
431 
432 	void *p = kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE);
433 #ifdef ZFS_ZIO_BUF_CANARY
434 	zio_buf_put_canary(p, size, zio_buf_cache, c);
435 #endif
436 	return (p);
437 }
438 
439 /*
440  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
441  * crashdump if the kernel panics.  This exists so that we will limit the amount
442  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
443  * of kernel heap dumped to disk when the kernel panics)
444  */
445 void *
446 zio_data_buf_alloc(size_t size)
447 {
448 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
449 
450 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
451 
452 	void *p = kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE);
453 #ifdef ZFS_ZIO_BUF_CANARY
454 	zio_buf_put_canary(p, size, zio_data_buf_cache, c);
455 #endif
456 	return (p);
457 }
458 
459 void
460 zio_buf_free(void *buf, size_t size)
461 {
462 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
463 
464 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
465 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
466 	atomic_add_64(&zio_buf_cache_frees[c], 1);
467 #endif
468 
469 #ifdef ZFS_ZIO_BUF_CANARY
470 	zio_buf_check_canary(buf, size, zio_buf_cache, c);
471 #endif
472 	kmem_cache_free(zio_buf_cache[c], buf);
473 }
474 
475 void
476 zio_data_buf_free(void *buf, size_t size)
477 {
478 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
479 
480 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
481 
482 #ifdef ZFS_ZIO_BUF_CANARY
483 	zio_buf_check_canary(buf, size, zio_data_buf_cache, c);
484 #endif
485 	kmem_cache_free(zio_data_buf_cache[c], buf);
486 }
487 
488 static void
489 zio_abd_free(void *abd, size_t size)
490 {
491 	(void) size;
492 	abd_free((abd_t *)abd);
493 }
494 
495 /*
496  * ==========================================================================
497  * Push and pop I/O transform buffers
498  * ==========================================================================
499  */
500 void
501 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
502     zio_transform_func_t *transform)
503 {
504 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
505 
506 	zt->zt_orig_abd = zio->io_abd;
507 	zt->zt_orig_size = zio->io_size;
508 	zt->zt_bufsize = bufsize;
509 	zt->zt_transform = transform;
510 
511 	zt->zt_next = zio->io_transform_stack;
512 	zio->io_transform_stack = zt;
513 
514 	zio->io_abd = data;
515 	zio->io_size = size;
516 }
517 
518 void
519 zio_pop_transforms(zio_t *zio)
520 {
521 	zio_transform_t *zt;
522 
523 	while ((zt = zio->io_transform_stack) != NULL) {
524 		if (zt->zt_transform != NULL)
525 			zt->zt_transform(zio,
526 			    zt->zt_orig_abd, zt->zt_orig_size);
527 
528 		if (zt->zt_bufsize != 0)
529 			abd_free(zio->io_abd);
530 
531 		zio->io_abd = zt->zt_orig_abd;
532 		zio->io_size = zt->zt_orig_size;
533 		zio->io_transform_stack = zt->zt_next;
534 
535 		kmem_free(zt, sizeof (zio_transform_t));
536 	}
537 }
538 
539 /*
540  * ==========================================================================
541  * I/O transform callbacks for subblocks, decompression, and decryption
542  * ==========================================================================
543  */
544 static void
545 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
546 {
547 	ASSERT(zio->io_size > size);
548 
549 	if (zio->io_type == ZIO_TYPE_READ)
550 		abd_copy(data, zio->io_abd, size);
551 }
552 
553 static void
554 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
555 {
556 	if (zio->io_error == 0) {
557 		int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
558 		    zio->io_abd, data, zio->io_size, size,
559 		    &zio->io_prop.zp_complevel);
560 
561 		if (zio_injection_enabled && ret == 0)
562 			ret = zio_handle_fault_injection(zio, EINVAL);
563 
564 		if (ret != 0)
565 			zio->io_error = SET_ERROR(EIO);
566 	}
567 }
568 
569 static void
570 zio_decrypt(zio_t *zio, abd_t *data, uint64_t size)
571 {
572 	int ret;
573 	void *tmp;
574 	blkptr_t *bp = zio->io_bp;
575 	spa_t *spa = zio->io_spa;
576 	uint64_t dsobj = zio->io_bookmark.zb_objset;
577 	uint64_t lsize = BP_GET_LSIZE(bp);
578 	dmu_object_type_t ot = BP_GET_TYPE(bp);
579 	uint8_t salt[ZIO_DATA_SALT_LEN];
580 	uint8_t iv[ZIO_DATA_IV_LEN];
581 	uint8_t mac[ZIO_DATA_MAC_LEN];
582 	boolean_t no_crypt = B_FALSE;
583 
584 	ASSERT(BP_USES_CRYPT(bp));
585 	ASSERT3U(size, !=, 0);
586 
587 	if (zio->io_error != 0)
588 		return;
589 
590 	/*
591 	 * Verify the cksum of MACs stored in an indirect bp. It will always
592 	 * be possible to verify this since it does not require an encryption
593 	 * key.
594 	 */
595 	if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {
596 		zio_crypt_decode_mac_bp(bp, mac);
597 
598 		if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
599 			/*
600 			 * We haven't decompressed the data yet, but
601 			 * zio_crypt_do_indirect_mac_checksum() requires
602 			 * decompressed data to be able to parse out the MACs
603 			 * from the indirect block. We decompress it now and
604 			 * throw away the result after we are finished.
605 			 */
606 			abd_t *abd = abd_alloc_linear(lsize, B_TRUE);
607 			ret = zio_decompress_data(BP_GET_COMPRESS(bp),
608 			    zio->io_abd, abd, zio->io_size, lsize,
609 			    &zio->io_prop.zp_complevel);
610 			if (ret != 0) {
611 				abd_free(abd);
612 				ret = SET_ERROR(EIO);
613 				goto error;
614 			}
615 			ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
616 			    abd, lsize, BP_SHOULD_BYTESWAP(bp), mac);
617 			abd_free(abd);
618 		} else {
619 			ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
620 			    zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac);
621 		}
622 		abd_copy(data, zio->io_abd, size);
623 
624 		if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) {
625 			ret = zio_handle_decrypt_injection(spa,
626 			    &zio->io_bookmark, ot, ECKSUM);
627 		}
628 		if (ret != 0)
629 			goto error;
630 
631 		return;
632 	}
633 
634 	/*
635 	 * If this is an authenticated block, just check the MAC. It would be
636 	 * nice to separate this out into its own flag, but when this was done,
637 	 * we had run out of bits in what is now zio_flag_t. Future cleanup
638 	 * could make this a flag bit.
639 	 */
640 	if (BP_IS_AUTHENTICATED(bp)) {
641 		if (ot == DMU_OT_OBJSET) {
642 			ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa,
643 			    dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp));
644 		} else {
645 			zio_crypt_decode_mac_bp(bp, mac);
646 			ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj,
647 			    zio->io_abd, size, mac);
648 			if (zio_injection_enabled && ret == 0) {
649 				ret = zio_handle_decrypt_injection(spa,
650 				    &zio->io_bookmark, ot, ECKSUM);
651 			}
652 		}
653 		abd_copy(data, zio->io_abd, size);
654 
655 		if (ret != 0)
656 			goto error;
657 
658 		return;
659 	}
660 
661 	zio_crypt_decode_params_bp(bp, salt, iv);
662 
663 	if (ot == DMU_OT_INTENT_LOG) {
664 		tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t));
665 		zio_crypt_decode_mac_zil(tmp, mac);
666 		abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t));
667 	} else {
668 		zio_crypt_decode_mac_bp(bp, mac);
669 	}
670 
671 	ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp),
672 	    BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data,
673 	    zio->io_abd, &no_crypt);
674 	if (no_crypt)
675 		abd_copy(data, zio->io_abd, size);
676 
677 	if (ret != 0)
678 		goto error;
679 
680 	return;
681 
682 error:
683 	/* assert that the key was found unless this was speculative */
684 	ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE));
685 
686 	/*
687 	 * If there was a decryption / authentication error return EIO as
688 	 * the io_error. If this was not a speculative zio, create an ereport.
689 	 */
690 	if (ret == ECKSUM) {
691 		zio->io_error = SET_ERROR(EIO);
692 		if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
693 			spa_log_error(spa, &zio->io_bookmark,
694 			    BP_GET_LOGICAL_BIRTH(zio->io_bp));
695 			(void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
696 			    spa, NULL, &zio->io_bookmark, zio, 0);
697 		}
698 	} else {
699 		zio->io_error = ret;
700 	}
701 }
702 
703 /*
704  * ==========================================================================
705  * I/O parent/child relationships and pipeline interlocks
706  * ==========================================================================
707  */
708 zio_t *
709 zio_walk_parents(zio_t *cio, zio_link_t **zl)
710 {
711 	list_t *pl = &cio->io_parent_list;
712 
713 	*zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
714 	if (*zl == NULL)
715 		return (NULL);
716 
717 	ASSERT((*zl)->zl_child == cio);
718 	return ((*zl)->zl_parent);
719 }
720 
721 zio_t *
722 zio_walk_children(zio_t *pio, zio_link_t **zl)
723 {
724 	list_t *cl = &pio->io_child_list;
725 
726 	ASSERT(MUTEX_HELD(&pio->io_lock));
727 
728 	*zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
729 	if (*zl == NULL)
730 		return (NULL);
731 
732 	ASSERT((*zl)->zl_parent == pio);
733 	return ((*zl)->zl_child);
734 }
735 
736 zio_t *
737 zio_unique_parent(zio_t *cio)
738 {
739 	zio_link_t *zl = NULL;
740 	zio_t *pio = zio_walk_parents(cio, &zl);
741 
742 	VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
743 	return (pio);
744 }
745 
746 void
747 zio_add_child(zio_t *pio, zio_t *cio)
748 {
749 	/*
750 	 * Logical I/Os can have logical, gang, or vdev children.
751 	 * Gang I/Os can have gang or vdev children.
752 	 * Vdev I/Os can only have vdev children.
753 	 * The following ASSERT captures all of these constraints.
754 	 */
755 	ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
756 
757 	/* Parent should not have READY stage if child doesn't have it. */
758 	IMPLY((cio->io_pipeline & ZIO_STAGE_READY) == 0 &&
759 	    (cio->io_child_type != ZIO_CHILD_VDEV),
760 	    (pio->io_pipeline & ZIO_STAGE_READY) == 0);
761 
762 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
763 	zl->zl_parent = pio;
764 	zl->zl_child = cio;
765 
766 	mutex_enter(&pio->io_lock);
767 	mutex_enter(&cio->io_lock);
768 
769 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
770 
771 	uint64_t *countp = pio->io_children[cio->io_child_type];
772 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
773 		countp[w] += !cio->io_state[w];
774 
775 	list_insert_head(&pio->io_child_list, zl);
776 	list_insert_head(&cio->io_parent_list, zl);
777 
778 	mutex_exit(&cio->io_lock);
779 	mutex_exit(&pio->io_lock);
780 }
781 
782 void
783 zio_add_child_first(zio_t *pio, zio_t *cio)
784 {
785 	/*
786 	 * Logical I/Os can have logical, gang, or vdev children.
787 	 * Gang I/Os can have gang or vdev children.
788 	 * Vdev I/Os can only have vdev children.
789 	 * The following ASSERT captures all of these constraints.
790 	 */
791 	ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
792 
793 	/* Parent should not have READY stage if child doesn't have it. */
794 	IMPLY((cio->io_pipeline & ZIO_STAGE_READY) == 0 &&
795 	    (cio->io_child_type != ZIO_CHILD_VDEV),
796 	    (pio->io_pipeline & ZIO_STAGE_READY) == 0);
797 
798 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
799 	zl->zl_parent = pio;
800 	zl->zl_child = cio;
801 
802 	ASSERT(list_is_empty(&cio->io_parent_list));
803 	list_insert_head(&cio->io_parent_list, zl);
804 
805 	mutex_enter(&pio->io_lock);
806 
807 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
808 
809 	uint64_t *countp = pio->io_children[cio->io_child_type];
810 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
811 		countp[w] += !cio->io_state[w];
812 
813 	list_insert_head(&pio->io_child_list, zl);
814 
815 	mutex_exit(&pio->io_lock);
816 }
817 
818 static void
819 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
820 {
821 	ASSERT(zl->zl_parent == pio);
822 	ASSERT(zl->zl_child == cio);
823 
824 	mutex_enter(&pio->io_lock);
825 	mutex_enter(&cio->io_lock);
826 
827 	list_remove(&pio->io_child_list, zl);
828 	list_remove(&cio->io_parent_list, zl);
829 
830 	mutex_exit(&cio->io_lock);
831 	mutex_exit(&pio->io_lock);
832 	kmem_cache_free(zio_link_cache, zl);
833 }
834 
835 static boolean_t
836 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
837 {
838 	boolean_t waiting = B_FALSE;
839 
840 	mutex_enter(&zio->io_lock);
841 	ASSERT(zio->io_stall == NULL);
842 	for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
843 		if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
844 			continue;
845 
846 		uint64_t *countp = &zio->io_children[c][wait];
847 		if (*countp != 0) {
848 			zio->io_stage >>= 1;
849 			ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
850 			zio->io_stall = countp;
851 			waiting = B_TRUE;
852 			break;
853 		}
854 	}
855 	mutex_exit(&zio->io_lock);
856 	return (waiting);
857 }
858 
859 __attribute__((always_inline))
860 static inline void
861 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait,
862     zio_t **next_to_executep)
863 {
864 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
865 	int *errorp = &pio->io_child_error[zio->io_child_type];
866 
867 	mutex_enter(&pio->io_lock);
868 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
869 		*errorp = zio_worst_error(*errorp, zio->io_error);
870 	pio->io_reexecute |= zio->io_reexecute;
871 	ASSERT3U(*countp, >, 0);
872 
873 	/*
874 	 * Propogate the Direct I/O checksum verify failure to the parent.
875 	 */
876 	if (zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR)
877 		pio->io_flags |= ZIO_FLAG_DIO_CHKSUM_ERR;
878 
879 	(*countp)--;
880 
881 	if (*countp == 0 && pio->io_stall == countp) {
882 		zio_taskq_type_t type =
883 		    pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
884 		    ZIO_TASKQ_INTERRUPT;
885 		pio->io_stall = NULL;
886 		mutex_exit(&pio->io_lock);
887 
888 		/*
889 		 * If we can tell the caller to execute this parent next, do
890 		 * so. We do this if the parent's zio type matches the child's
891 		 * type, or if it's a zio_null() with no done callback, and so
892 		 * has no actual work to do. Otherwise dispatch the parent zio
893 		 * in its own taskq.
894 		 *
895 		 * Having the caller execute the parent when possible reduces
896 		 * locking on the zio taskq's, reduces context switch
897 		 * overhead, and has no recursion penalty.  Note that one
898 		 * read from disk typically causes at least 3 zio's: a
899 		 * zio_null(), the logical zio_read(), and then a physical
900 		 * zio.  When the physical ZIO completes, we are able to call
901 		 * zio_done() on all 3 of these zio's from one invocation of
902 		 * zio_execute() by returning the parent back to
903 		 * zio_execute().  Since the parent isn't executed until this
904 		 * thread returns back to zio_execute(), the caller should do
905 		 * so promptly.
906 		 *
907 		 * In other cases, dispatching the parent prevents
908 		 * overflowing the stack when we have deeply nested
909 		 * parent-child relationships, as we do with the "mega zio"
910 		 * of writes for spa_sync(), and the chain of ZIL blocks.
911 		 */
912 		if (next_to_executep != NULL && *next_to_executep == NULL &&
913 		    (pio->io_type == zio->io_type ||
914 		    (pio->io_type == ZIO_TYPE_NULL && !pio->io_done))) {
915 			*next_to_executep = pio;
916 		} else {
917 			zio_taskq_dispatch(pio, type, B_FALSE);
918 		}
919 	} else {
920 		mutex_exit(&pio->io_lock);
921 	}
922 }
923 
924 static void
925 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
926 {
927 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
928 		zio->io_error = zio->io_child_error[c];
929 }
930 
931 int
932 zio_bookmark_compare(const void *x1, const void *x2)
933 {
934 	const zio_t *z1 = x1;
935 	const zio_t *z2 = x2;
936 
937 	if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
938 		return (-1);
939 	if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
940 		return (1);
941 
942 	if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
943 		return (-1);
944 	if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
945 		return (1);
946 
947 	if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
948 		return (-1);
949 	if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
950 		return (1);
951 
952 	if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
953 		return (-1);
954 	if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
955 		return (1);
956 
957 	if (z1 < z2)
958 		return (-1);
959 	if (z1 > z2)
960 		return (1);
961 
962 	return (0);
963 }
964 
965 /*
966  * ==========================================================================
967  * Create the various types of I/O (read, write, free, etc)
968  * ==========================================================================
969  */
970 static zio_t *
971 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
972     abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
973     void *private, zio_type_t type, zio_priority_t priority,
974     zio_flag_t flags, vdev_t *vd, uint64_t offset,
975     const zbookmark_phys_t *zb, enum zio_stage stage,
976     enum zio_stage pipeline)
977 {
978 	zio_t *zio;
979 
980 	IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE);
981 	ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
982 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
983 
984 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
985 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
986 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
987 
988 	IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0);
989 
990 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
991 	memset(zio, 0, sizeof (zio_t));
992 
993 	mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL);
994 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
995 
996 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
997 	    offsetof(zio_link_t, zl_parent_node));
998 	list_create(&zio->io_child_list, sizeof (zio_link_t),
999 	    offsetof(zio_link_t, zl_child_node));
1000 	metaslab_trace_init(&zio->io_alloc_list);
1001 
1002 	if (vd != NULL)
1003 		zio->io_child_type = ZIO_CHILD_VDEV;
1004 	else if (flags & ZIO_FLAG_GANG_CHILD)
1005 		zio->io_child_type = ZIO_CHILD_GANG;
1006 	else if (flags & ZIO_FLAG_DDT_CHILD)
1007 		zio->io_child_type = ZIO_CHILD_DDT;
1008 	else
1009 		zio->io_child_type = ZIO_CHILD_LOGICAL;
1010 
1011 	if (bp != NULL) {
1012 		if (type != ZIO_TYPE_WRITE ||
1013 		    zio->io_child_type == ZIO_CHILD_DDT) {
1014 			zio->io_bp_copy = *bp;
1015 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
1016 		} else {
1017 			zio->io_bp = (blkptr_t *)bp;
1018 		}
1019 		zio->io_bp_orig = *bp;
1020 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
1021 			zio->io_logical = zio;
1022 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
1023 			pipeline |= ZIO_GANG_STAGES;
1024 	}
1025 
1026 	zio->io_spa = spa;
1027 	zio->io_txg = txg;
1028 	zio->io_done = done;
1029 	zio->io_private = private;
1030 	zio->io_type = type;
1031 	zio->io_priority = priority;
1032 	zio->io_vd = vd;
1033 	zio->io_offset = offset;
1034 	zio->io_orig_abd = zio->io_abd = data;
1035 	zio->io_orig_size = zio->io_size = psize;
1036 	zio->io_lsize = lsize;
1037 	zio->io_orig_flags = zio->io_flags = flags;
1038 	zio->io_orig_stage = zio->io_stage = stage;
1039 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
1040 	zio->io_pipeline_trace = ZIO_STAGE_OPEN;
1041 	zio->io_allocator = ZIO_ALLOCATOR_NONE;
1042 
1043 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY) ||
1044 	    (pipeline & ZIO_STAGE_READY) == 0;
1045 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
1046 
1047 	if (zb != NULL)
1048 		zio->io_bookmark = *zb;
1049 
1050 	if (pio != NULL) {
1051 		zio->io_metaslab_class = pio->io_metaslab_class;
1052 		if (zio->io_logical == NULL)
1053 			zio->io_logical = pio->io_logical;
1054 		if (zio->io_child_type == ZIO_CHILD_GANG)
1055 			zio->io_gang_leader = pio->io_gang_leader;
1056 		zio_add_child_first(pio, zio);
1057 	}
1058 
1059 	taskq_init_ent(&zio->io_tqent);
1060 
1061 	return (zio);
1062 }
1063 
1064 void
1065 zio_destroy(zio_t *zio)
1066 {
1067 	metaslab_trace_fini(&zio->io_alloc_list);
1068 	list_destroy(&zio->io_parent_list);
1069 	list_destroy(&zio->io_child_list);
1070 	mutex_destroy(&zio->io_lock);
1071 	cv_destroy(&zio->io_cv);
1072 	kmem_cache_free(zio_cache, zio);
1073 }
1074 
1075 /*
1076  * ZIO intended to be between others.  Provides synchronization at READY
1077  * and DONE pipeline stages and calls the respective callbacks.
1078  */
1079 zio_t *
1080 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
1081     void *private, zio_flag_t flags)
1082 {
1083 	zio_t *zio;
1084 
1085 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1086 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1087 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
1088 
1089 	return (zio);
1090 }
1091 
1092 /*
1093  * ZIO intended to be a root of a tree.  Unlike null ZIO does not have a
1094  * READY pipeline stage (is ready on creation), so it should not be used
1095  * as child of any ZIO that may need waiting for grandchildren READY stage
1096  * (any other ZIO type).
1097  */
1098 zio_t *
1099 zio_root(spa_t *spa, zio_done_func_t *done, void *private, zio_flag_t flags)
1100 {
1101 	zio_t *zio;
1102 
1103 	zio = zio_create(NULL, spa, 0, NULL, NULL, 0, 0, done, private,
1104 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, NULL, 0, NULL,
1105 	    ZIO_STAGE_OPEN, ZIO_ROOT_PIPELINE);
1106 
1107 	return (zio);
1108 }
1109 
1110 static int
1111 zfs_blkptr_verify_log(spa_t *spa, const blkptr_t *bp,
1112     enum blk_verify_flag blk_verify, const char *fmt, ...)
1113 {
1114 	va_list adx;
1115 	char buf[256];
1116 
1117 	va_start(adx, fmt);
1118 	(void) vsnprintf(buf, sizeof (buf), fmt, adx);
1119 	va_end(adx);
1120 
1121 	zfs_dbgmsg("bad blkptr at %px: "
1122 	    "DVA[0]=%#llx/%#llx "
1123 	    "DVA[1]=%#llx/%#llx "
1124 	    "DVA[2]=%#llx/%#llx "
1125 	    "prop=%#llx "
1126 	    "pad=%#llx,%#llx "
1127 	    "phys_birth=%#llx "
1128 	    "birth=%#llx "
1129 	    "fill=%#llx "
1130 	    "cksum=%#llx/%#llx/%#llx/%#llx",
1131 	    bp,
1132 	    (long long)bp->blk_dva[0].dva_word[0],
1133 	    (long long)bp->blk_dva[0].dva_word[1],
1134 	    (long long)bp->blk_dva[1].dva_word[0],
1135 	    (long long)bp->blk_dva[1].dva_word[1],
1136 	    (long long)bp->blk_dva[2].dva_word[0],
1137 	    (long long)bp->blk_dva[2].dva_word[1],
1138 	    (long long)bp->blk_prop,
1139 	    (long long)bp->blk_pad[0],
1140 	    (long long)bp->blk_pad[1],
1141 	    (long long)BP_GET_PHYSICAL_BIRTH(bp),
1142 	    (long long)BP_GET_LOGICAL_BIRTH(bp),
1143 	    (long long)bp->blk_fill,
1144 	    (long long)bp->blk_cksum.zc_word[0],
1145 	    (long long)bp->blk_cksum.zc_word[1],
1146 	    (long long)bp->blk_cksum.zc_word[2],
1147 	    (long long)bp->blk_cksum.zc_word[3]);
1148 	switch (blk_verify) {
1149 	case BLK_VERIFY_HALT:
1150 		zfs_panic_recover("%s: %s", spa_name(spa), buf);
1151 		break;
1152 	case BLK_VERIFY_LOG:
1153 		zfs_dbgmsg("%s: %s", spa_name(spa), buf);
1154 		break;
1155 	case BLK_VERIFY_ONLY:
1156 		break;
1157 	}
1158 
1159 	return (1);
1160 }
1161 
1162 /*
1163  * Verify the block pointer fields contain reasonable values.  This means
1164  * it only contains known object types, checksum/compression identifiers,
1165  * block sizes within the maximum allowed limits, valid DVAs, etc.
1166  *
1167  * If everything checks out B_TRUE is returned.  The zfs_blkptr_verify
1168  * argument controls the behavior when an invalid field is detected.
1169  *
1170  * Values for blk_verify_flag:
1171  *   BLK_VERIFY_ONLY: evaluate the block
1172  *   BLK_VERIFY_LOG: evaluate the block and log problems
1173  *   BLK_VERIFY_HALT: call zfs_panic_recover on error
1174  *
1175  * Values for blk_config_flag:
1176  *   BLK_CONFIG_HELD: caller holds SCL_VDEV for writer
1177  *   BLK_CONFIG_NEEDED: caller holds no config lock, SCL_VDEV will be
1178  *   obtained for reader
1179  *   BLK_CONFIG_SKIP: skip checks which require SCL_VDEV, for better
1180  *   performance
1181  */
1182 boolean_t
1183 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp,
1184     enum blk_config_flag blk_config, enum blk_verify_flag blk_verify)
1185 {
1186 	int errors = 0;
1187 
1188 	if (unlikely(!DMU_OT_IS_VALID(BP_GET_TYPE(bp)))) {
1189 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1190 		    "blkptr at %px has invalid TYPE %llu",
1191 		    bp, (longlong_t)BP_GET_TYPE(bp));
1192 	}
1193 	if (unlikely(BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS)) {
1194 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1195 		    "blkptr at %px has invalid COMPRESS %llu",
1196 		    bp, (longlong_t)BP_GET_COMPRESS(bp));
1197 	}
1198 	if (unlikely(BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE)) {
1199 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1200 		    "blkptr at %px has invalid LSIZE %llu",
1201 		    bp, (longlong_t)BP_GET_LSIZE(bp));
1202 	}
1203 	if (BP_IS_EMBEDDED(bp)) {
1204 		if (unlikely(BPE_GET_ETYPE(bp) >= NUM_BP_EMBEDDED_TYPES)) {
1205 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1206 			    "blkptr at %px has invalid ETYPE %llu",
1207 			    bp, (longlong_t)BPE_GET_ETYPE(bp));
1208 		}
1209 		if (unlikely(BPE_GET_PSIZE(bp) > BPE_PAYLOAD_SIZE)) {
1210 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1211 			    "blkptr at %px has invalid PSIZE %llu",
1212 			    bp, (longlong_t)BPE_GET_PSIZE(bp));
1213 		}
1214 		return (errors == 0);
1215 	}
1216 	if (unlikely(BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS)) {
1217 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1218 		    "blkptr at %px has invalid CHECKSUM %llu",
1219 		    bp, (longlong_t)BP_GET_CHECKSUM(bp));
1220 	}
1221 	if (unlikely(BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE)) {
1222 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1223 		    "blkptr at %px has invalid PSIZE %llu",
1224 		    bp, (longlong_t)BP_GET_PSIZE(bp));
1225 	}
1226 
1227 	/*
1228 	 * Do not verify individual DVAs if the config is not trusted. This
1229 	 * will be done once the zio is executed in vdev_mirror_map_alloc.
1230 	 */
1231 	if (unlikely(!spa->spa_trust_config))
1232 		return (errors == 0);
1233 
1234 	switch (blk_config) {
1235 	case BLK_CONFIG_HELD:
1236 		ASSERT(spa_config_held(spa, SCL_VDEV, RW_WRITER));
1237 		break;
1238 	case BLK_CONFIG_NEEDED:
1239 		spa_config_enter(spa, SCL_VDEV, bp, RW_READER);
1240 		break;
1241 	case BLK_CONFIG_SKIP:
1242 		return (errors == 0);
1243 	default:
1244 		panic("invalid blk_config %u", blk_config);
1245 	}
1246 
1247 	/*
1248 	 * Pool-specific checks.
1249 	 *
1250 	 * Note: it would be nice to verify that the logical birth
1251 	 * and physical birth are not too large.  However,
1252 	 * spa_freeze() allows the birth time of log blocks (and
1253 	 * dmu_sync()-ed blocks that are in the log) to be arbitrarily
1254 	 * large.
1255 	 */
1256 	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
1257 		const dva_t *dva = &bp->blk_dva[i];
1258 		uint64_t vdevid = DVA_GET_VDEV(dva);
1259 
1260 		if (unlikely(vdevid >= spa->spa_root_vdev->vdev_children)) {
1261 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1262 			    "blkptr at %px DVA %u has invalid VDEV %llu",
1263 			    bp, i, (longlong_t)vdevid);
1264 			continue;
1265 		}
1266 		vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1267 		if (unlikely(vd == NULL)) {
1268 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1269 			    "blkptr at %px DVA %u has invalid VDEV %llu",
1270 			    bp, i, (longlong_t)vdevid);
1271 			continue;
1272 		}
1273 		if (unlikely(vd->vdev_ops == &vdev_hole_ops)) {
1274 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1275 			    "blkptr at %px DVA %u has hole VDEV %llu",
1276 			    bp, i, (longlong_t)vdevid);
1277 			continue;
1278 		}
1279 		if (vd->vdev_ops == &vdev_missing_ops) {
1280 			/*
1281 			 * "missing" vdevs are valid during import, but we
1282 			 * don't have their detailed info (e.g. asize), so
1283 			 * we can't perform any more checks on them.
1284 			 */
1285 			continue;
1286 		}
1287 		uint64_t offset = DVA_GET_OFFSET(dva);
1288 		uint64_t asize = DVA_GET_ASIZE(dva);
1289 		if (DVA_GET_GANG(dva))
1290 			asize = vdev_gang_header_asize(vd);
1291 		if (unlikely(offset + asize > vd->vdev_asize)) {
1292 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1293 			    "blkptr at %px DVA %u has invalid OFFSET %llu",
1294 			    bp, i, (longlong_t)offset);
1295 		}
1296 	}
1297 	if (blk_config == BLK_CONFIG_NEEDED)
1298 		spa_config_exit(spa, SCL_VDEV, bp);
1299 
1300 	return (errors == 0);
1301 }
1302 
1303 boolean_t
1304 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
1305 {
1306 	(void) bp;
1307 	uint64_t vdevid = DVA_GET_VDEV(dva);
1308 
1309 	if (vdevid >= spa->spa_root_vdev->vdev_children)
1310 		return (B_FALSE);
1311 
1312 	vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1313 	if (vd == NULL)
1314 		return (B_FALSE);
1315 
1316 	if (vd->vdev_ops == &vdev_hole_ops)
1317 		return (B_FALSE);
1318 
1319 	if (vd->vdev_ops == &vdev_missing_ops) {
1320 		return (B_FALSE);
1321 	}
1322 
1323 	uint64_t offset = DVA_GET_OFFSET(dva);
1324 	uint64_t asize = DVA_GET_ASIZE(dva);
1325 
1326 	if (DVA_GET_GANG(dva))
1327 		asize = vdev_gang_header_asize(vd);
1328 	if (offset + asize > vd->vdev_asize)
1329 		return (B_FALSE);
1330 
1331 	return (B_TRUE);
1332 }
1333 
1334 zio_t *
1335 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
1336     abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
1337     zio_priority_t priority, zio_flag_t flags, const zbookmark_phys_t *zb)
1338 {
1339 	zio_t *zio;
1340 
1341 	zio = zio_create(pio, spa, BP_GET_BIRTH(bp), bp,
1342 	    data, size, size, done, private,
1343 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
1344 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1345 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
1346 
1347 	return (zio);
1348 }
1349 
1350 zio_t *
1351 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
1352     abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
1353     zio_done_func_t *ready, zio_done_func_t *children_ready,
1354     zio_done_func_t *done, void *private, zio_priority_t priority,
1355     zio_flag_t flags, const zbookmark_phys_t *zb)
1356 {
1357 	zio_t *zio;
1358 	enum zio_stage pipeline = zp->zp_direct_write == B_TRUE ?
1359 	    ZIO_DIRECT_WRITE_PIPELINE : (flags & ZIO_FLAG_DDT_CHILD) ?
1360 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE;
1361 
1362 
1363 	zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
1364 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
1365 	    ZIO_STAGE_OPEN, pipeline);
1366 
1367 	zio->io_ready = ready;
1368 	zio->io_children_ready = children_ready;
1369 	zio->io_prop = *zp;
1370 
1371 	/*
1372 	 * Data can be NULL if we are going to call zio_write_override() to
1373 	 * provide the already-allocated BP.  But we may need the data to
1374 	 * verify a dedup hit (if requested).  In this case, don't try to
1375 	 * dedup (just take the already-allocated BP verbatim). Encrypted
1376 	 * dedup blocks need data as well so we also disable dedup in this
1377 	 * case.
1378 	 */
1379 	if (data == NULL &&
1380 	    (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) {
1381 		zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
1382 	}
1383 
1384 	return (zio);
1385 }
1386 
1387 zio_t *
1388 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
1389     uint64_t size, zio_done_func_t *done, void *private,
1390     zio_priority_t priority, zio_flag_t flags, zbookmark_phys_t *zb)
1391 {
1392 	zio_t *zio;
1393 
1394 	zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
1395 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
1396 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
1397 
1398 	return (zio);
1399 }
1400 
1401 void
1402 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite,
1403     boolean_t brtwrite)
1404 {
1405 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1406 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1407 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1408 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
1409 	ASSERT(!brtwrite || !nopwrite);
1410 
1411 	/*
1412 	 * We must reset the io_prop to match the values that existed
1413 	 * when the bp was first written by dmu_sync() keeping in mind
1414 	 * that nopwrite and dedup are mutually exclusive.
1415 	 */
1416 	zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
1417 	zio->io_prop.zp_nopwrite = nopwrite;
1418 	zio->io_prop.zp_brtwrite = brtwrite;
1419 	zio->io_prop.zp_copies = copies;
1420 	zio->io_bp_override = bp;
1421 }
1422 
1423 void
1424 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
1425 {
1426 
1427 	(void) zfs_blkptr_verify(spa, bp, BLK_CONFIG_NEEDED, BLK_VERIFY_HALT);
1428 
1429 	/*
1430 	 * The check for EMBEDDED is a performance optimization.  We
1431 	 * process the free here (by ignoring it) rather than
1432 	 * putting it on the list and then processing it in zio_free_sync().
1433 	 */
1434 	if (BP_IS_EMBEDDED(bp))
1435 		return;
1436 
1437 	/*
1438 	 * Frees that are for the currently-syncing txg, are not going to be
1439 	 * deferred, and which will not need to do a read (i.e. not GANG or
1440 	 * DEDUP), can be processed immediately.  Otherwise, put them on the
1441 	 * in-memory list for later processing.
1442 	 *
1443 	 * Note that we only defer frees after zfs_sync_pass_deferred_free
1444 	 * when the log space map feature is disabled. [see relevant comment
1445 	 * in spa_sync_iterate_to_convergence()]
1446 	 */
1447 	if (BP_IS_GANG(bp) ||
1448 	    BP_GET_DEDUP(bp) ||
1449 	    txg != spa->spa_syncing_txg ||
1450 	    (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free &&
1451 	    !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) ||
1452 	    brt_maybe_exists(spa, bp)) {
1453 		metaslab_check_free(spa, bp);
1454 		bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1455 	} else {
1456 		VERIFY3P(zio_free_sync(NULL, spa, txg, bp, 0), ==, NULL);
1457 	}
1458 }
1459 
1460 /*
1461  * To improve performance, this function may return NULL if we were able
1462  * to do the free immediately.  This avoids the cost of creating a zio
1463  * (and linking it to the parent, etc).
1464  */
1465 zio_t *
1466 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1467     zio_flag_t flags)
1468 {
1469 	ASSERT(!BP_IS_HOLE(bp));
1470 	ASSERT(spa_syncing_txg(spa) == txg);
1471 
1472 	if (BP_IS_EMBEDDED(bp))
1473 		return (NULL);
1474 
1475 	metaslab_check_free(spa, bp);
1476 	arc_freed(spa, bp);
1477 	dsl_scan_freed(spa, bp);
1478 
1479 	if (BP_IS_GANG(bp) ||
1480 	    BP_GET_DEDUP(bp) ||
1481 	    brt_maybe_exists(spa, bp)) {
1482 		/*
1483 		 * GANG, DEDUP and BRT blocks can induce a read (for the gang
1484 		 * block header, the DDT or the BRT), so issue them
1485 		 * asynchronously so that this thread is not tied up.
1486 		 */
1487 		enum zio_stage stage =
1488 		    ZIO_FREE_PIPELINE | ZIO_STAGE_ISSUE_ASYNC;
1489 
1490 		return (zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1491 		    BP_GET_PSIZE(bp), NULL, NULL,
1492 		    ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1493 		    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage));
1494 	} else {
1495 		metaslab_free(spa, bp, txg, B_FALSE);
1496 		return (NULL);
1497 	}
1498 }
1499 
1500 zio_t *
1501 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1502     zio_done_func_t *done, void *private, zio_flag_t flags)
1503 {
1504 	zio_t *zio;
1505 
1506 	(void) zfs_blkptr_verify(spa, bp, (flags & ZIO_FLAG_CONFIG_WRITER) ?
1507 	    BLK_CONFIG_HELD : BLK_CONFIG_NEEDED, BLK_VERIFY_HALT);
1508 
1509 	if (BP_IS_EMBEDDED(bp))
1510 		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1511 
1512 	/*
1513 	 * A claim is an allocation of a specific block.  Claims are needed
1514 	 * to support immediate writes in the intent log.  The issue is that
1515 	 * immediate writes contain committed data, but in a txg that was
1516 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
1517 	 * the intent log claims all blocks that contain immediate write data
1518 	 * so that the SPA knows they're in use.
1519 	 *
1520 	 * All claims *must* be resolved in the first txg -- before the SPA
1521 	 * starts allocating blocks -- so that nothing is allocated twice.
1522 	 * If txg == 0 we just verify that the block is claimable.
1523 	 */
1524 	ASSERT3U(BP_GET_LOGICAL_BIRTH(&spa->spa_uberblock.ub_rootbp), <,
1525 	    spa_min_claim_txg(spa));
1526 	ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1527 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(8) */
1528 
1529 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1530 	    BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1531 	    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1532 	ASSERT0(zio->io_queued_timestamp);
1533 
1534 	return (zio);
1535 }
1536 
1537 zio_t *
1538 zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1539     zio_done_func_t *done, void *private, zio_priority_t priority,
1540     zio_flag_t flags, enum trim_flag trim_flags)
1541 {
1542 	zio_t *zio;
1543 
1544 	ASSERT0(vd->vdev_children);
1545 	ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
1546 	ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1547 	ASSERT3U(size, !=, 0);
1548 
1549 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done,
1550 	    private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL,
1551 	    vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE);
1552 	zio->io_trim_flags = trim_flags;
1553 
1554 	return (zio);
1555 }
1556 
1557 zio_t *
1558 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1559     abd_t *data, int checksum, zio_done_func_t *done, void *private,
1560     zio_priority_t priority, zio_flag_t flags, boolean_t labels)
1561 {
1562 	zio_t *zio;
1563 
1564 	ASSERT(vd->vdev_children == 0);
1565 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1566 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1567 	ASSERT3U(offset + size, <=, vd->vdev_psize);
1568 
1569 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1570 	    private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1571 	    offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1572 
1573 	zio->io_prop.zp_checksum = checksum;
1574 
1575 	return (zio);
1576 }
1577 
1578 zio_t *
1579 zio_write_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_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1592 	    offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1593 
1594 	zio->io_prop.zp_checksum = checksum;
1595 
1596 	if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1597 		/*
1598 		 * zec checksums are necessarily destructive -- they modify
1599 		 * the end of the write buffer to hold the verifier/checksum.
1600 		 * Therefore, we must make a local copy in case the data is
1601 		 * being written to multiple places in parallel.
1602 		 */
1603 		abd_t *wbuf = abd_alloc_sametype(data, size);
1604 		abd_copy(wbuf, data, size);
1605 
1606 		zio_push_transform(zio, wbuf, size, size, NULL);
1607 	}
1608 
1609 	return (zio);
1610 }
1611 
1612 /*
1613  * Create a child I/O to do some work for us.
1614  */
1615 zio_t *
1616 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1617     abd_t *data, uint64_t size, int type, zio_priority_t priority,
1618     zio_flag_t flags, zio_done_func_t *done, void *private)
1619 {
1620 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1621 	zio_t *zio;
1622 
1623 	/*
1624 	 * vdev child I/Os do not propagate their error to the parent.
1625 	 * Therefore, for correct operation the caller *must* check for
1626 	 * and handle the error in the child i/o's done callback.
1627 	 * The only exceptions are i/os that we don't care about
1628 	 * (OPTIONAL or REPAIR).
1629 	 */
1630 	ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1631 	    done != NULL);
1632 
1633 	if (type == ZIO_TYPE_READ && bp != NULL) {
1634 		/*
1635 		 * If we have the bp, then the child should perform the
1636 		 * checksum and the parent need not.  This pushes error
1637 		 * detection as close to the leaves as possible and
1638 		 * eliminates redundant checksums in the interior nodes.
1639 		 */
1640 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1641 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1642 		/*
1643 		 * We never allow the mirror VDEV to attempt reading from any
1644 		 * additional data copies after the first Direct I/O checksum
1645 		 * verify failure. This is to avoid bad data being written out
1646 		 * through the mirror during self healing. See comment in
1647 		 * vdev_mirror_io_done() for more details.
1648 		 */
1649 		ASSERT0(pio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR);
1650 	} else if (type == ZIO_TYPE_WRITE &&
1651 	    pio->io_prop.zp_direct_write == B_TRUE) {
1652 		/*
1653 		 * By default we only will verify checksums for Direct I/O
1654 		 * writes for Linux. FreeBSD is able to place user pages under
1655 		 * write protection before issuing them to the ZIO pipeline.
1656 		 *
1657 		 * Checksum validation errors will only be reported through
1658 		 * the top-level VDEV, which is set by this child ZIO.
1659 		 */
1660 		ASSERT3P(bp, !=, NULL);
1661 		ASSERT3U(pio->io_child_type, ==, ZIO_CHILD_LOGICAL);
1662 		pipeline |= ZIO_STAGE_DIO_CHECKSUM_VERIFY;
1663 	}
1664 
1665 	if (vd->vdev_ops->vdev_op_leaf) {
1666 		ASSERT0(vd->vdev_children);
1667 		offset += VDEV_LABEL_START_SIZE;
1668 	}
1669 
1670 	flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1671 
1672 	/*
1673 	 * If we've decided to do a repair, the write is not speculative --
1674 	 * even if the original read was.
1675 	 */
1676 	if (flags & ZIO_FLAG_IO_REPAIR)
1677 		flags &= ~ZIO_FLAG_SPECULATIVE;
1678 
1679 	/*
1680 	 * If we're creating a child I/O that is not associated with a
1681 	 * top-level vdev, then the child zio is not an allocating I/O.
1682 	 * If this is a retried I/O then we ignore it since we will
1683 	 * have already processed the original allocating I/O.
1684 	 */
1685 	if (flags & ZIO_FLAG_IO_ALLOCATING &&
1686 	    (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1687 		ASSERT(pio->io_metaslab_class != NULL);
1688 		ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled);
1689 		ASSERT(type == ZIO_TYPE_WRITE);
1690 		ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1691 		ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1692 		ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1693 		    pio->io_child_type == ZIO_CHILD_GANG);
1694 
1695 		flags &= ~ZIO_FLAG_IO_ALLOCATING;
1696 	}
1697 
1698 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1699 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1700 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1701 	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1702 
1703 	return (zio);
1704 }
1705 
1706 zio_t *
1707 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1708     zio_type_t type, zio_priority_t priority, zio_flag_t flags,
1709     zio_done_func_t *done, void *private)
1710 {
1711 	zio_t *zio;
1712 
1713 	ASSERT(vd->vdev_ops->vdev_op_leaf);
1714 
1715 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1716 	    data, size, size, done, private, type, priority,
1717 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1718 	    vd, offset, NULL,
1719 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1720 
1721 	return (zio);
1722 }
1723 
1724 
1725 /*
1726  * Send a flush command to the given vdev. Unlike most zio creation functions,
1727  * the flush zios are issued immediately. You can wait on pio to pause until
1728  * the flushes complete.
1729  */
1730 void
1731 zio_flush(zio_t *pio, vdev_t *vd)
1732 {
1733 	const zio_flag_t flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
1734 	    ZIO_FLAG_DONT_RETRY;
1735 
1736 	if (vd->vdev_nowritecache)
1737 		return;
1738 
1739 	if (vd->vdev_children == 0) {
1740 		zio_nowait(zio_create(pio, vd->vdev_spa, 0, NULL, NULL, 0, 0,
1741 		    NULL, NULL, ZIO_TYPE_FLUSH, ZIO_PRIORITY_NOW, flags, vd, 0,
1742 		    NULL, ZIO_STAGE_OPEN, ZIO_FLUSH_PIPELINE));
1743 	} else {
1744 		for (uint64_t c = 0; c < vd->vdev_children; c++)
1745 			zio_flush(pio, vd->vdev_child[c]);
1746 	}
1747 }
1748 
1749 void
1750 zio_shrink(zio_t *zio, uint64_t size)
1751 {
1752 	ASSERT3P(zio->io_executor, ==, NULL);
1753 	ASSERT3U(zio->io_orig_size, ==, zio->io_size);
1754 	ASSERT3U(size, <=, zio->io_size);
1755 
1756 	/*
1757 	 * We don't shrink for raidz because of problems with the
1758 	 * reconstruction when reading back less than the block size.
1759 	 * Note, BP_IS_RAIDZ() assumes no compression.
1760 	 */
1761 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1762 	if (!BP_IS_RAIDZ(zio->io_bp)) {
1763 		/* we are not doing a raw write */
1764 		ASSERT3U(zio->io_size, ==, zio->io_lsize);
1765 		zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1766 	}
1767 }
1768 
1769 /*
1770  * Round provided allocation size up to a value that can be allocated
1771  * by at least some vdev(s) in the pool with minimum or no additional
1772  * padding and without extra space usage on others
1773  */
1774 static uint64_t
1775 zio_roundup_alloc_size(spa_t *spa, uint64_t size)
1776 {
1777 	if (size > spa->spa_min_alloc)
1778 		return (roundup(size, spa->spa_gcd_alloc));
1779 	return (spa->spa_min_alloc);
1780 }
1781 
1782 size_t
1783 zio_get_compression_max_size(enum zio_compress compress, uint64_t gcd_alloc,
1784     uint64_t min_alloc, size_t s_len)
1785 {
1786 	size_t d_len;
1787 
1788 	/* minimum 12.5% must be saved (legacy value, may be changed later) */
1789 	d_len = s_len - (s_len >> 3);
1790 
1791 	/* ZLE can't use exactly d_len bytes, it needs more, so ignore it */
1792 	if (compress == ZIO_COMPRESS_ZLE)
1793 		return (d_len);
1794 
1795 	d_len = d_len - d_len % gcd_alloc;
1796 
1797 	if (d_len < min_alloc)
1798 		return (BPE_PAYLOAD_SIZE);
1799 	return (d_len);
1800 }
1801 
1802 /*
1803  * ==========================================================================
1804  * Prepare to read and write logical blocks
1805  * ==========================================================================
1806  */
1807 
1808 static zio_t *
1809 zio_read_bp_init(zio_t *zio)
1810 {
1811 	blkptr_t *bp = zio->io_bp;
1812 	uint64_t psize =
1813 	    BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1814 
1815 	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1816 
1817 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1818 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
1819 	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1820 		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1821 		    psize, psize, zio_decompress);
1822 	}
1823 
1824 	if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) ||
1825 	    BP_HAS_INDIRECT_MAC_CKSUM(bp)) &&
1826 	    zio->io_child_type == ZIO_CHILD_LOGICAL) {
1827 		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1828 		    psize, psize, zio_decrypt);
1829 	}
1830 
1831 	if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1832 		int psize = BPE_GET_PSIZE(bp);
1833 		void *data = abd_borrow_buf(zio->io_abd, psize);
1834 
1835 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1836 		decode_embedded_bp_compressed(bp, data);
1837 		abd_return_buf_copy(zio->io_abd, data, psize);
1838 	} else {
1839 		ASSERT(!BP_IS_EMBEDDED(bp));
1840 	}
1841 
1842 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1843 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1844 
1845 	return (zio);
1846 }
1847 
1848 static zio_t *
1849 zio_write_bp_init(zio_t *zio)
1850 {
1851 	if (!IO_IS_ALLOCATING(zio))
1852 		return (zio);
1853 
1854 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1855 
1856 	if (zio->io_bp_override) {
1857 		blkptr_t *bp = zio->io_bp;
1858 		zio_prop_t *zp = &zio->io_prop;
1859 
1860 		ASSERT(BP_GET_LOGICAL_BIRTH(bp) != zio->io_txg);
1861 
1862 		*bp = *zio->io_bp_override;
1863 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1864 
1865 		if (zp->zp_brtwrite)
1866 			return (zio);
1867 
1868 		ASSERT(!BP_GET_DEDUP(zio->io_bp_override));
1869 
1870 		if (BP_IS_EMBEDDED(bp))
1871 			return (zio);
1872 
1873 		/*
1874 		 * If we've been overridden and nopwrite is set then
1875 		 * set the flag accordingly to indicate that a nopwrite
1876 		 * has already occurred.
1877 		 */
1878 		if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1879 			ASSERT(!zp->zp_dedup);
1880 			ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1881 			zio->io_flags |= ZIO_FLAG_NOPWRITE;
1882 			return (zio);
1883 		}
1884 
1885 		ASSERT(!zp->zp_nopwrite);
1886 
1887 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1888 			return (zio);
1889 
1890 		ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1891 		    ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1892 
1893 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum &&
1894 		    !zp->zp_encrypt) {
1895 			BP_SET_DEDUP(bp, 1);
1896 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1897 			return (zio);
1898 		}
1899 
1900 		/*
1901 		 * We were unable to handle this as an override bp, treat
1902 		 * it as a regular write I/O.
1903 		 */
1904 		zio->io_bp_override = NULL;
1905 		*bp = zio->io_bp_orig;
1906 		zio->io_pipeline = zio->io_orig_pipeline;
1907 	}
1908 
1909 	return (zio);
1910 }
1911 
1912 static zio_t *
1913 zio_write_compress(zio_t *zio)
1914 {
1915 	spa_t *spa = zio->io_spa;
1916 	zio_prop_t *zp = &zio->io_prop;
1917 	enum zio_compress compress = zp->zp_compress;
1918 	blkptr_t *bp = zio->io_bp;
1919 	uint64_t lsize = zio->io_lsize;
1920 	uint64_t psize = zio->io_size;
1921 	uint32_t pass = 1;
1922 
1923 	/*
1924 	 * If our children haven't all reached the ready stage,
1925 	 * wait for them and then repeat this pipeline stage.
1926 	 */
1927 	if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1928 	    ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1929 		return (NULL);
1930 	}
1931 
1932 	if (!IO_IS_ALLOCATING(zio))
1933 		return (zio);
1934 
1935 	if (zio->io_children_ready != NULL) {
1936 		/*
1937 		 * Now that all our children are ready, run the callback
1938 		 * associated with this zio in case it wants to modify the
1939 		 * data to be written.
1940 		 */
1941 		ASSERT3U(zp->zp_level, >, 0);
1942 		zio->io_children_ready(zio);
1943 	}
1944 
1945 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1946 	ASSERT(zio->io_bp_override == NULL);
1947 
1948 	if (!BP_IS_HOLE(bp) && BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg) {
1949 		/*
1950 		 * We're rewriting an existing block, which means we're
1951 		 * working on behalf of spa_sync().  For spa_sync() to
1952 		 * converge, it must eventually be the case that we don't
1953 		 * have to allocate new blocks.  But compression changes
1954 		 * the blocksize, which forces a reallocate, and makes
1955 		 * convergence take longer.  Therefore, after the first
1956 		 * few passes, stop compressing to ensure convergence.
1957 		 */
1958 		pass = spa_sync_pass(spa);
1959 
1960 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
1961 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1962 		ASSERT(!BP_GET_DEDUP(bp));
1963 
1964 		if (pass >= zfs_sync_pass_dont_compress)
1965 			compress = ZIO_COMPRESS_OFF;
1966 
1967 		/* Make sure someone doesn't change their mind on overwrites */
1968 		ASSERT(BP_IS_EMBEDDED(bp) || BP_IS_GANG(bp) ||
1969 		    MIN(zp->zp_copies, spa_max_replication(spa))
1970 		    == BP_GET_NDVAS(bp));
1971 	}
1972 
1973 	/* If it's a compressed write that is not raw, compress the buffer. */
1974 	if (compress != ZIO_COMPRESS_OFF &&
1975 	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1976 		abd_t *cabd = NULL;
1977 		if (abd_cmp_zero(zio->io_abd, lsize) == 0)
1978 			psize = 0;
1979 		else if (compress == ZIO_COMPRESS_EMPTY)
1980 			psize = lsize;
1981 		else
1982 			psize = zio_compress_data(compress, zio->io_abd, &cabd,
1983 			    lsize,
1984 			    zio_get_compression_max_size(compress,
1985 			    spa->spa_gcd_alloc, spa->spa_min_alloc, lsize),
1986 			    zp->zp_complevel);
1987 		if (psize == 0) {
1988 			compress = ZIO_COMPRESS_OFF;
1989 		} else if (psize >= lsize) {
1990 			compress = ZIO_COMPRESS_OFF;
1991 			if (cabd != NULL)
1992 				abd_free(cabd);
1993 		} else if (!zp->zp_dedup && !zp->zp_encrypt &&
1994 		    psize <= BPE_PAYLOAD_SIZE &&
1995 		    zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1996 		    spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1997 			void *cbuf = abd_borrow_buf_copy(cabd, lsize);
1998 			encode_embedded_bp_compressed(bp,
1999 			    cbuf, compress, lsize, psize);
2000 			BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
2001 			BP_SET_TYPE(bp, zio->io_prop.zp_type);
2002 			BP_SET_LEVEL(bp, zio->io_prop.zp_level);
2003 			abd_return_buf(cabd, cbuf, lsize);
2004 			abd_free(cabd);
2005 			BP_SET_LOGICAL_BIRTH(bp, zio->io_txg);
2006 			zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2007 			ASSERT(spa_feature_is_active(spa,
2008 			    SPA_FEATURE_EMBEDDED_DATA));
2009 			return (zio);
2010 		} else {
2011 			/*
2012 			 * Round compressed size up to the minimum allocation
2013 			 * size of the smallest-ashift device, and zero the
2014 			 * tail. This ensures that the compressed size of the
2015 			 * BP (and thus compressratio property) are correct,
2016 			 * in that we charge for the padding used to fill out
2017 			 * the last sector.
2018 			 */
2019 			size_t rounded = (size_t)zio_roundup_alloc_size(spa,
2020 			    psize);
2021 			if (rounded >= lsize) {
2022 				compress = ZIO_COMPRESS_OFF;
2023 				abd_free(cabd);
2024 				psize = lsize;
2025 			} else {
2026 				abd_zero_off(cabd, psize, rounded - psize);
2027 				psize = rounded;
2028 				zio_push_transform(zio, cabd,
2029 				    psize, lsize, NULL);
2030 			}
2031 		}
2032 
2033 		/*
2034 		 * We were unable to handle this as an override bp, treat
2035 		 * it as a regular write I/O.
2036 		 */
2037 		zio->io_bp_override = NULL;
2038 		*bp = zio->io_bp_orig;
2039 		zio->io_pipeline = zio->io_orig_pipeline;
2040 
2041 	} else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 &&
2042 	    zp->zp_type == DMU_OT_DNODE) {
2043 		/*
2044 		 * The DMU actually relies on the zio layer's compression
2045 		 * to free metadnode blocks that have had all contained
2046 		 * dnodes freed. As a result, even when doing a raw
2047 		 * receive, we must check whether the block can be compressed
2048 		 * to a hole.
2049 		 */
2050 		if (abd_cmp_zero(zio->io_abd, lsize) == 0) {
2051 			psize = 0;
2052 			compress = ZIO_COMPRESS_OFF;
2053 		} else {
2054 			psize = lsize;
2055 		}
2056 	} else if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS &&
2057 	    !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) {
2058 		/*
2059 		 * If we are raw receiving an encrypted dataset we should not
2060 		 * take this codepath because it will change the on-disk block
2061 		 * and decryption will fail.
2062 		 */
2063 		size_t rounded = MIN((size_t)zio_roundup_alloc_size(spa, psize),
2064 		    lsize);
2065 
2066 		if (rounded != psize) {
2067 			abd_t *cdata = abd_alloc_linear(rounded, B_TRUE);
2068 			abd_zero_off(cdata, psize, rounded - psize);
2069 			abd_copy_off(cdata, zio->io_abd, 0, 0, psize);
2070 			psize = rounded;
2071 			zio_push_transform(zio, cdata,
2072 			    psize, rounded, NULL);
2073 		}
2074 	} else {
2075 		ASSERT3U(psize, !=, 0);
2076 	}
2077 
2078 	/*
2079 	 * The final pass of spa_sync() must be all rewrites, but the first
2080 	 * few passes offer a trade-off: allocating blocks defers convergence,
2081 	 * but newly allocated blocks are sequential, so they can be written
2082 	 * to disk faster.  Therefore, we allow the first few passes of
2083 	 * spa_sync() to allocate new blocks, but force rewrites after that.
2084 	 * There should only be a handful of blocks after pass 1 in any case.
2085 	 */
2086 	if (!BP_IS_HOLE(bp) && BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg &&
2087 	    BP_GET_PSIZE(bp) == psize &&
2088 	    pass >= zfs_sync_pass_rewrite) {
2089 		VERIFY3U(psize, !=, 0);
2090 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
2091 
2092 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
2093 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
2094 	} else {
2095 		BP_ZERO(bp);
2096 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2097 	}
2098 
2099 	if (psize == 0) {
2100 		if (BP_GET_LOGICAL_BIRTH(&zio->io_bp_orig) != 0 &&
2101 		    spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
2102 			BP_SET_LSIZE(bp, lsize);
2103 			BP_SET_TYPE(bp, zp->zp_type);
2104 			BP_SET_LEVEL(bp, zp->zp_level);
2105 			BP_SET_BIRTH(bp, zio->io_txg, 0);
2106 		}
2107 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2108 	} else {
2109 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
2110 		BP_SET_LSIZE(bp, lsize);
2111 		BP_SET_TYPE(bp, zp->zp_type);
2112 		BP_SET_LEVEL(bp, zp->zp_level);
2113 		BP_SET_PSIZE(bp, psize);
2114 		BP_SET_COMPRESS(bp, compress);
2115 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
2116 		BP_SET_DEDUP(bp, zp->zp_dedup);
2117 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
2118 		if (zp->zp_dedup) {
2119 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2120 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2121 			ASSERT(!zp->zp_encrypt ||
2122 			    DMU_OT_IS_ENCRYPTED(zp->zp_type));
2123 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
2124 		}
2125 		if (zp->zp_nopwrite) {
2126 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2127 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2128 			zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
2129 		}
2130 	}
2131 	return (zio);
2132 }
2133 
2134 static zio_t *
2135 zio_free_bp_init(zio_t *zio)
2136 {
2137 	blkptr_t *bp = zio->io_bp;
2138 
2139 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
2140 		if (BP_GET_DEDUP(bp))
2141 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
2142 	}
2143 
2144 	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
2145 
2146 	return (zio);
2147 }
2148 
2149 /*
2150  * ==========================================================================
2151  * Execute the I/O pipeline
2152  * ==========================================================================
2153  */
2154 
2155 static void
2156 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
2157 {
2158 	spa_t *spa = zio->io_spa;
2159 	zio_type_t t = zio->io_type;
2160 
2161 	/*
2162 	 * If we're a config writer or a probe, the normal issue and
2163 	 * interrupt threads may all be blocked waiting for the config lock.
2164 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
2165 	 */
2166 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
2167 		t = ZIO_TYPE_NULL;
2168 
2169 	/*
2170 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
2171 	 */
2172 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
2173 		t = ZIO_TYPE_NULL;
2174 
2175 	/*
2176 	 * If this is a high priority I/O, then use the high priority taskq if
2177 	 * available or cut the line otherwise.
2178 	 */
2179 	if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) {
2180 		if (spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
2181 			q++;
2182 		else
2183 			cutinline = B_TRUE;
2184 	}
2185 
2186 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
2187 
2188 	spa_taskq_dispatch(spa, t, q, zio_execute, zio, cutinline);
2189 }
2190 
2191 static boolean_t
2192 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
2193 {
2194 	spa_t *spa = zio->io_spa;
2195 
2196 	taskq_t *tq = taskq_of_curthread();
2197 
2198 	for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
2199 		spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
2200 		uint_t i;
2201 		for (i = 0; i < tqs->stqs_count; i++) {
2202 			if (tqs->stqs_taskq[i] == tq)
2203 				return (B_TRUE);
2204 		}
2205 	}
2206 
2207 	return (B_FALSE);
2208 }
2209 
2210 static zio_t *
2211 zio_issue_async(zio_t *zio)
2212 {
2213 	ASSERT((zio->io_type != ZIO_TYPE_WRITE) || ZIO_HAS_ALLOCATOR(zio));
2214 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2215 	return (NULL);
2216 }
2217 
2218 void
2219 zio_interrupt(void *zio)
2220 {
2221 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
2222 }
2223 
2224 void
2225 zio_delay_interrupt(zio_t *zio)
2226 {
2227 	/*
2228 	 * The timeout_generic() function isn't defined in userspace, so
2229 	 * rather than trying to implement the function, the zio delay
2230 	 * functionality has been disabled for userspace builds.
2231 	 */
2232 
2233 #ifdef _KERNEL
2234 	/*
2235 	 * If io_target_timestamp is zero, then no delay has been registered
2236 	 * for this IO, thus jump to the end of this function and "skip" the
2237 	 * delay; issuing it directly to the zio layer.
2238 	 */
2239 	if (zio->io_target_timestamp != 0) {
2240 		hrtime_t now = gethrtime();
2241 
2242 		if (now >= zio->io_target_timestamp) {
2243 			/*
2244 			 * This IO has already taken longer than the target
2245 			 * delay to complete, so we don't want to delay it
2246 			 * any longer; we "miss" the delay and issue it
2247 			 * directly to the zio layer. This is likely due to
2248 			 * the target latency being set to a value less than
2249 			 * the underlying hardware can satisfy (e.g. delay
2250 			 * set to 1ms, but the disks take 10ms to complete an
2251 			 * IO request).
2252 			 */
2253 
2254 			DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
2255 			    hrtime_t, now);
2256 
2257 			zio_interrupt(zio);
2258 		} else {
2259 			taskqid_t tid;
2260 			hrtime_t diff = zio->io_target_timestamp - now;
2261 			int ticks = MAX(1, NSEC_TO_TICK(diff));
2262 			clock_t expire_at_tick = ddi_get_lbolt() + ticks;
2263 
2264 			DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
2265 			    hrtime_t, now, hrtime_t, diff);
2266 
2267 			tid = taskq_dispatch_delay(system_taskq, zio_interrupt,
2268 			    zio, TQ_NOSLEEP, expire_at_tick);
2269 			if (tid == TASKQID_INVALID) {
2270 				/*
2271 				 * Couldn't allocate a task.  Just finish the
2272 				 * zio without a delay.
2273 				 */
2274 				zio_interrupt(zio);
2275 			}
2276 		}
2277 		return;
2278 	}
2279 #endif
2280 	DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
2281 	zio_interrupt(zio);
2282 }
2283 
2284 static void
2285 zio_deadman_impl(zio_t *pio, int ziodepth)
2286 {
2287 	zio_t *cio, *cio_next;
2288 	zio_link_t *zl = NULL;
2289 	vdev_t *vd = pio->io_vd;
2290 
2291 	if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) {
2292 		vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL;
2293 		zbookmark_phys_t *zb = &pio->io_bookmark;
2294 		uint64_t delta = gethrtime() - pio->io_timestamp;
2295 		uint64_t failmode = spa_get_deadman_failmode(pio->io_spa);
2296 
2297 		zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu "
2298 		    "delta=%llu queued=%llu io=%llu "
2299 		    "path=%s "
2300 		    "last=%llu type=%d "
2301 		    "priority=%d flags=0x%llx stage=0x%x "
2302 		    "pipeline=0x%x pipeline-trace=0x%x "
2303 		    "objset=%llu object=%llu "
2304 		    "level=%llu blkid=%llu "
2305 		    "offset=%llu size=%llu "
2306 		    "error=%d",
2307 		    ziodepth, pio, pio->io_timestamp,
2308 		    (u_longlong_t)delta, pio->io_delta, pio->io_delay,
2309 		    vd ? vd->vdev_path : "NULL",
2310 		    vq ? vq->vq_io_complete_ts : 0, pio->io_type,
2311 		    pio->io_priority, (u_longlong_t)pio->io_flags,
2312 		    pio->io_stage, pio->io_pipeline, pio->io_pipeline_trace,
2313 		    (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object,
2314 		    (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid,
2315 		    (u_longlong_t)pio->io_offset, (u_longlong_t)pio->io_size,
2316 		    pio->io_error);
2317 		(void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN,
2318 		    pio->io_spa, vd, zb, pio, 0);
2319 
2320 		if (failmode == ZIO_FAILURE_MODE_CONTINUE &&
2321 		    taskq_empty_ent(&pio->io_tqent)) {
2322 			zio_interrupt(pio);
2323 		}
2324 	}
2325 
2326 	mutex_enter(&pio->io_lock);
2327 	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2328 		cio_next = zio_walk_children(pio, &zl);
2329 		zio_deadman_impl(cio, ziodepth + 1);
2330 	}
2331 	mutex_exit(&pio->io_lock);
2332 }
2333 
2334 /*
2335  * Log the critical information describing this zio and all of its children
2336  * using the zfs_dbgmsg() interface then post deadman event for the ZED.
2337  */
2338 void
2339 zio_deadman(zio_t *pio, const char *tag)
2340 {
2341 	spa_t *spa = pio->io_spa;
2342 	char *name = spa_name(spa);
2343 
2344 	if (!zfs_deadman_enabled || spa_suspended(spa))
2345 		return;
2346 
2347 	zio_deadman_impl(pio, 0);
2348 
2349 	switch (spa_get_deadman_failmode(spa)) {
2350 	case ZIO_FAILURE_MODE_WAIT:
2351 		zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name);
2352 		break;
2353 
2354 	case ZIO_FAILURE_MODE_CONTINUE:
2355 		zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name);
2356 		break;
2357 
2358 	case ZIO_FAILURE_MODE_PANIC:
2359 		fm_panic("%s determined I/O to pool '%s' is hung.", tag, name);
2360 		break;
2361 	}
2362 }
2363 
2364 /*
2365  * Execute the I/O pipeline until one of the following occurs:
2366  * (1) the I/O completes; (2) the pipeline stalls waiting for
2367  * dependent child I/Os; (3) the I/O issues, so we're waiting
2368  * for an I/O completion interrupt; (4) the I/O is delegated by
2369  * vdev-level caching or aggregation; (5) the I/O is deferred
2370  * due to vdev-level queueing; (6) the I/O is handed off to
2371  * another thread.  In all cases, the pipeline stops whenever
2372  * there's no CPU work; it never burns a thread in cv_wait_io().
2373  *
2374  * There's no locking on io_stage because there's no legitimate way
2375  * for multiple threads to be attempting to process the same I/O.
2376  */
2377 static zio_pipe_stage_t *zio_pipeline[];
2378 
2379 /*
2380  * zio_execute() is a wrapper around the static function
2381  * __zio_execute() so that we can force  __zio_execute() to be
2382  * inlined.  This reduces stack overhead which is important
2383  * because __zio_execute() is called recursively in several zio
2384  * code paths.  zio_execute() itself cannot be inlined because
2385  * it is externally visible.
2386  */
2387 void
2388 zio_execute(void *zio)
2389 {
2390 	fstrans_cookie_t cookie;
2391 
2392 	cookie = spl_fstrans_mark();
2393 	__zio_execute(zio);
2394 	spl_fstrans_unmark(cookie);
2395 }
2396 
2397 /*
2398  * Used to determine if in the current context the stack is sized large
2399  * enough to allow zio_execute() to be called recursively.  A minimum
2400  * stack size of 16K is required to avoid needing to re-dispatch the zio.
2401  */
2402 static boolean_t
2403 zio_execute_stack_check(zio_t *zio)
2404 {
2405 #if !defined(HAVE_LARGE_STACKS)
2406 	dsl_pool_t *dp = spa_get_dsl(zio->io_spa);
2407 
2408 	/* Executing in txg_sync_thread() context. */
2409 	if (dp && curthread == dp->dp_tx.tx_sync_thread)
2410 		return (B_TRUE);
2411 
2412 	/* Pool initialization outside of zio_taskq context. */
2413 	if (dp && spa_is_initializing(dp->dp_spa) &&
2414 	    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
2415 	    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))
2416 		return (B_TRUE);
2417 #else
2418 	(void) zio;
2419 #endif /* HAVE_LARGE_STACKS */
2420 
2421 	return (B_FALSE);
2422 }
2423 
2424 __attribute__((always_inline))
2425 static inline void
2426 __zio_execute(zio_t *zio)
2427 {
2428 	ASSERT3U(zio->io_queued_timestamp, >, 0);
2429 
2430 	while (zio->io_stage < ZIO_STAGE_DONE) {
2431 		enum zio_stage pipeline = zio->io_pipeline;
2432 		enum zio_stage stage = zio->io_stage;
2433 
2434 		zio->io_executor = curthread;
2435 
2436 		ASSERT(!MUTEX_HELD(&zio->io_lock));
2437 		ASSERT(ISP2(stage));
2438 		ASSERT(zio->io_stall == NULL);
2439 
2440 		do {
2441 			stage <<= 1;
2442 		} while ((stage & pipeline) == 0);
2443 
2444 		ASSERT(stage <= ZIO_STAGE_DONE);
2445 
2446 		/*
2447 		 * If we are in interrupt context and this pipeline stage
2448 		 * will grab a config lock that is held across I/O,
2449 		 * or may wait for an I/O that needs an interrupt thread
2450 		 * to complete, issue async to avoid deadlock.
2451 		 *
2452 		 * For VDEV_IO_START, we cut in line so that the io will
2453 		 * be sent to disk promptly.
2454 		 */
2455 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
2456 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
2457 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2458 			    zio_requeue_io_start_cut_in_line : B_FALSE;
2459 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2460 			return;
2461 		}
2462 
2463 		/*
2464 		 * If the current context doesn't have large enough stacks
2465 		 * the zio must be issued asynchronously to prevent overflow.
2466 		 */
2467 		if (zio_execute_stack_check(zio)) {
2468 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2469 			    zio_requeue_io_start_cut_in_line : B_FALSE;
2470 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2471 			return;
2472 		}
2473 
2474 		zio->io_stage = stage;
2475 		zio->io_pipeline_trace |= zio->io_stage;
2476 
2477 		/*
2478 		 * The zio pipeline stage returns the next zio to execute
2479 		 * (typically the same as this one), or NULL if we should
2480 		 * stop.
2481 		 */
2482 		zio = zio_pipeline[highbit64(stage) - 1](zio);
2483 
2484 		if (zio == NULL)
2485 			return;
2486 	}
2487 }
2488 
2489 
2490 /*
2491  * ==========================================================================
2492  * Initiate I/O, either sync or async
2493  * ==========================================================================
2494  */
2495 int
2496 zio_wait(zio_t *zio)
2497 {
2498 	/*
2499 	 * Some routines, like zio_free_sync(), may return a NULL zio
2500 	 * to avoid the performance overhead of creating and then destroying
2501 	 * an unneeded zio.  For the callers' simplicity, we accept a NULL
2502 	 * zio and ignore it.
2503 	 */
2504 	if (zio == NULL)
2505 		return (0);
2506 
2507 	long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms);
2508 	int error;
2509 
2510 	ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN);
2511 	ASSERT3P(zio->io_executor, ==, NULL);
2512 
2513 	zio->io_waiter = curthread;
2514 	ASSERT0(zio->io_queued_timestamp);
2515 	zio->io_queued_timestamp = gethrtime();
2516 
2517 	if (zio->io_type == ZIO_TYPE_WRITE) {
2518 		spa_select_allocator(zio);
2519 	}
2520 	__zio_execute(zio);
2521 
2522 	mutex_enter(&zio->io_lock);
2523 	while (zio->io_executor != NULL) {
2524 		error = cv_timedwait_io(&zio->io_cv, &zio->io_lock,
2525 		    ddi_get_lbolt() + timeout);
2526 
2527 		if (zfs_deadman_enabled && error == -1 &&
2528 		    gethrtime() - zio->io_queued_timestamp >
2529 		    spa_deadman_ziotime(zio->io_spa)) {
2530 			mutex_exit(&zio->io_lock);
2531 			timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms);
2532 			zio_deadman(zio, FTAG);
2533 			mutex_enter(&zio->io_lock);
2534 		}
2535 	}
2536 	mutex_exit(&zio->io_lock);
2537 
2538 	error = zio->io_error;
2539 	zio_destroy(zio);
2540 
2541 	return (error);
2542 }
2543 
2544 void
2545 zio_nowait(zio_t *zio)
2546 {
2547 	/*
2548 	 * See comment in zio_wait().
2549 	 */
2550 	if (zio == NULL)
2551 		return;
2552 
2553 	ASSERT3P(zio->io_executor, ==, NULL);
2554 
2555 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
2556 	    list_is_empty(&zio->io_parent_list)) {
2557 		zio_t *pio;
2558 
2559 		/*
2560 		 * This is a logical async I/O with no parent to wait for it.
2561 		 * We add it to the spa_async_root_zio "Godfather" I/O which
2562 		 * will ensure they complete prior to unloading the pool.
2563 		 */
2564 		spa_t *spa = zio->io_spa;
2565 		pio = spa->spa_async_zio_root[CPU_SEQID_UNSTABLE];
2566 
2567 		zio_add_child(pio, zio);
2568 	}
2569 
2570 	ASSERT0(zio->io_queued_timestamp);
2571 	zio->io_queued_timestamp = gethrtime();
2572 	if (zio->io_type == ZIO_TYPE_WRITE) {
2573 		spa_select_allocator(zio);
2574 	}
2575 	__zio_execute(zio);
2576 }
2577 
2578 /*
2579  * ==========================================================================
2580  * Reexecute, cancel, or suspend/resume failed I/O
2581  * ==========================================================================
2582  */
2583 
2584 static void
2585 zio_reexecute(void *arg)
2586 {
2587 	zio_t *pio = arg;
2588 	zio_t *cio, *cio_next, *gio;
2589 
2590 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
2591 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
2592 	ASSERT(pio->io_gang_leader == NULL);
2593 	ASSERT(pio->io_gang_tree == NULL);
2594 
2595 	mutex_enter(&pio->io_lock);
2596 	pio->io_flags = pio->io_orig_flags;
2597 	pio->io_stage = pio->io_orig_stage;
2598 	pio->io_pipeline = pio->io_orig_pipeline;
2599 	pio->io_reexecute = 0;
2600 	pio->io_flags |= ZIO_FLAG_REEXECUTED;
2601 	pio->io_pipeline_trace = 0;
2602 	pio->io_error = 0;
2603 	pio->io_state[ZIO_WAIT_READY] = (pio->io_stage >= ZIO_STAGE_READY) ||
2604 	    (pio->io_pipeline & ZIO_STAGE_READY) == 0;
2605 	pio->io_state[ZIO_WAIT_DONE] = (pio->io_stage >= ZIO_STAGE_DONE);
2606 
2607 	/*
2608 	 * It's possible for a failed ZIO to be a descendant of more than one
2609 	 * ZIO tree. When reexecuting it, we have to be sure to add its wait
2610 	 * states to all parent wait counts.
2611 	 *
2612 	 * Those parents, in turn, may have other children that are currently
2613 	 * active, usually because they've already been reexecuted after
2614 	 * resuming. Those children may be executing and may call
2615 	 * zio_notify_parent() at the same time as we're updating our parent's
2616 	 * counts. To avoid races while updating the counts, we take
2617 	 * gio->io_lock before each update.
2618 	 */
2619 	zio_link_t *zl = NULL;
2620 	while ((gio = zio_walk_parents(pio, &zl)) != NULL) {
2621 		mutex_enter(&gio->io_lock);
2622 		for (int w = 0; w < ZIO_WAIT_TYPES; w++) {
2623 			gio->io_children[pio->io_child_type][w] +=
2624 			    !pio->io_state[w];
2625 		}
2626 		mutex_exit(&gio->io_lock);
2627 	}
2628 
2629 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2630 		pio->io_child_error[c] = 0;
2631 
2632 	if (IO_IS_ALLOCATING(pio))
2633 		BP_ZERO(pio->io_bp);
2634 
2635 	/*
2636 	 * As we reexecute pio's children, new children could be created.
2637 	 * New children go to the head of pio's io_child_list, however,
2638 	 * so we will (correctly) not reexecute them.  The key is that
2639 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
2640 	 * cannot be affected by any side effects of reexecuting 'cio'.
2641 	 */
2642 	zl = NULL;
2643 	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2644 		cio_next = zio_walk_children(pio, &zl);
2645 		mutex_exit(&pio->io_lock);
2646 		zio_reexecute(cio);
2647 		mutex_enter(&pio->io_lock);
2648 	}
2649 	mutex_exit(&pio->io_lock);
2650 
2651 	/*
2652 	 * Now that all children have been reexecuted, execute the parent.
2653 	 * We don't reexecute "The Godfather" I/O here as it's the
2654 	 * responsibility of the caller to wait on it.
2655 	 */
2656 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
2657 		pio->io_queued_timestamp = gethrtime();
2658 		__zio_execute(pio);
2659 	}
2660 }
2661 
2662 void
2663 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
2664 {
2665 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
2666 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2667 		    "failure and the failure mode property for this pool "
2668 		    "is set to panic.", spa_name(spa));
2669 
2670 	if (reason != ZIO_SUSPEND_MMP) {
2671 		cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable "
2672 		    "I/O failure and has been suspended.", spa_name(spa));
2673 	}
2674 
2675 	(void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
2676 	    NULL, NULL, 0);
2677 
2678 	mutex_enter(&spa->spa_suspend_lock);
2679 
2680 	if (spa->spa_suspend_zio_root == NULL)
2681 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
2682 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2683 		    ZIO_FLAG_GODFATHER);
2684 
2685 	spa->spa_suspended = reason;
2686 
2687 	if (zio != NULL) {
2688 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2689 		ASSERT(zio != spa->spa_suspend_zio_root);
2690 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2691 		ASSERT(zio_unique_parent(zio) == NULL);
2692 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
2693 		zio_add_child(spa->spa_suspend_zio_root, zio);
2694 	}
2695 
2696 	mutex_exit(&spa->spa_suspend_lock);
2697 }
2698 
2699 int
2700 zio_resume(spa_t *spa)
2701 {
2702 	zio_t *pio;
2703 
2704 	/*
2705 	 * Reexecute all previously suspended i/o.
2706 	 */
2707 	mutex_enter(&spa->spa_suspend_lock);
2708 	if (spa->spa_suspended != ZIO_SUSPEND_NONE)
2709 		cmn_err(CE_WARN, "Pool '%s' was suspended and is being "
2710 		    "resumed. Failed I/O will be retried.",
2711 		    spa_name(spa));
2712 	spa->spa_suspended = ZIO_SUSPEND_NONE;
2713 	cv_broadcast(&spa->spa_suspend_cv);
2714 	pio = spa->spa_suspend_zio_root;
2715 	spa->spa_suspend_zio_root = NULL;
2716 	mutex_exit(&spa->spa_suspend_lock);
2717 
2718 	if (pio == NULL)
2719 		return (0);
2720 
2721 	zio_reexecute(pio);
2722 	return (zio_wait(pio));
2723 }
2724 
2725 void
2726 zio_resume_wait(spa_t *spa)
2727 {
2728 	mutex_enter(&spa->spa_suspend_lock);
2729 	while (spa_suspended(spa))
2730 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
2731 	mutex_exit(&spa->spa_suspend_lock);
2732 }
2733 
2734 /*
2735  * ==========================================================================
2736  * Gang blocks.
2737  *
2738  * A gang block is a collection of small blocks that looks to the DMU
2739  * like one large block.  When zio_dva_allocate() cannot find a block
2740  * of the requested size, due to either severe fragmentation or the pool
2741  * being nearly full, it calls zio_write_gang_block() to construct the
2742  * block from smaller fragments.
2743  *
2744  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2745  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
2746  * an indirect block: it's an array of block pointers.  It consumes
2747  * only one sector and hence is allocatable regardless of fragmentation.
2748  * The gang header's bps point to its gang members, which hold the data.
2749  *
2750  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2751  * as the verifier to ensure uniqueness of the SHA256 checksum.
2752  * Critically, the gang block bp's blk_cksum is the checksum of the data,
2753  * not the gang header.  This ensures that data block signatures (needed for
2754  * deduplication) are independent of how the block is physically stored.
2755  *
2756  * Gang blocks can be nested: a gang member may itself be a gang block.
2757  * Thus every gang block is a tree in which root and all interior nodes are
2758  * gang headers, and the leaves are normal blocks that contain user data.
2759  * The root of the gang tree is called the gang leader.
2760  *
2761  * To perform any operation (read, rewrite, free, claim) on a gang block,
2762  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2763  * in the io_gang_tree field of the original logical i/o by recursively
2764  * reading the gang leader and all gang headers below it.  This yields
2765  * an in-core tree containing the contents of every gang header and the
2766  * bps for every constituent of the gang block.
2767  *
2768  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2769  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
2770  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2771  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2772  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2773  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
2774  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2775  * of the gang header plus zio_checksum_compute() of the data to update the
2776  * gang header's blk_cksum as described above.
2777  *
2778  * The two-phase assemble/issue model solves the problem of partial failure --
2779  * what if you'd freed part of a gang block but then couldn't read the
2780  * gang header for another part?  Assembling the entire gang tree first
2781  * ensures that all the necessary gang header I/O has succeeded before
2782  * starting the actual work of free, claim, or write.  Once the gang tree
2783  * is assembled, free and claim are in-memory operations that cannot fail.
2784  *
2785  * In the event that a gang write fails, zio_dva_unallocate() walks the
2786  * gang tree to immediately free (i.e. insert back into the space map)
2787  * everything we've allocated.  This ensures that we don't get ENOSPC
2788  * errors during repeated suspend/resume cycles due to a flaky device.
2789  *
2790  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
2791  * the gang tree, we won't modify the block, so we can safely defer the free
2792  * (knowing that the block is still intact).  If we *can* assemble the gang
2793  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2794  * each constituent bp and we can allocate a new block on the next sync pass.
2795  *
2796  * In all cases, the gang tree allows complete recovery from partial failure.
2797  * ==========================================================================
2798  */
2799 
2800 static void
2801 zio_gang_issue_func_done(zio_t *zio)
2802 {
2803 	abd_free(zio->io_abd);
2804 }
2805 
2806 static zio_t *
2807 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2808     uint64_t offset)
2809 {
2810 	if (gn != NULL)
2811 		return (pio);
2812 
2813 	return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2814 	    BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2815 	    NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2816 	    &pio->io_bookmark));
2817 }
2818 
2819 static zio_t *
2820 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2821     uint64_t offset)
2822 {
2823 	zio_t *zio;
2824 
2825 	if (gn != NULL) {
2826 		abd_t *gbh_abd =
2827 		    abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2828 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2829 		    gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2830 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2831 		    &pio->io_bookmark);
2832 		/*
2833 		 * As we rewrite each gang header, the pipeline will compute
2834 		 * a new gang block header checksum for it; but no one will
2835 		 * compute a new data checksum, so we do that here.  The one
2836 		 * exception is the gang leader: the pipeline already computed
2837 		 * its data checksum because that stage precedes gang assembly.
2838 		 * (Presently, nothing actually uses interior data checksums;
2839 		 * this is just good hygiene.)
2840 		 */
2841 		if (gn != pio->io_gang_leader->io_gang_tree) {
2842 			abd_t *buf = abd_get_offset(data, offset);
2843 
2844 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2845 			    buf, BP_GET_PSIZE(bp));
2846 
2847 			abd_free(buf);
2848 		}
2849 		/*
2850 		 * If we are here to damage data for testing purposes,
2851 		 * leave the GBH alone so that we can detect the damage.
2852 		 */
2853 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2854 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2855 	} else {
2856 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2857 		    abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2858 		    zio_gang_issue_func_done, NULL, pio->io_priority,
2859 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2860 	}
2861 
2862 	return (zio);
2863 }
2864 
2865 static zio_t *
2866 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2867     uint64_t offset)
2868 {
2869 	(void) gn, (void) data, (void) offset;
2870 
2871 	zio_t *zio = zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2872 	    ZIO_GANG_CHILD_FLAGS(pio));
2873 	if (zio == NULL) {
2874 		zio = zio_null(pio, pio->io_spa,
2875 		    NULL, NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio));
2876 	}
2877 	return (zio);
2878 }
2879 
2880 static zio_t *
2881 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2882     uint64_t offset)
2883 {
2884 	(void) gn, (void) data, (void) offset;
2885 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2886 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2887 }
2888 
2889 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2890 	NULL,
2891 	zio_read_gang,
2892 	zio_rewrite_gang,
2893 	zio_free_gang,
2894 	zio_claim_gang,
2895 	NULL
2896 };
2897 
2898 static void zio_gang_tree_assemble_done(zio_t *zio);
2899 
2900 static zio_gang_node_t *
2901 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2902 {
2903 	zio_gang_node_t *gn;
2904 
2905 	ASSERT(*gnpp == NULL);
2906 
2907 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2908 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2909 	*gnpp = gn;
2910 
2911 	return (gn);
2912 }
2913 
2914 static void
2915 zio_gang_node_free(zio_gang_node_t **gnpp)
2916 {
2917 	zio_gang_node_t *gn = *gnpp;
2918 
2919 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2920 		ASSERT(gn->gn_child[g] == NULL);
2921 
2922 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2923 	kmem_free(gn, sizeof (*gn));
2924 	*gnpp = NULL;
2925 }
2926 
2927 static void
2928 zio_gang_tree_free(zio_gang_node_t **gnpp)
2929 {
2930 	zio_gang_node_t *gn = *gnpp;
2931 
2932 	if (gn == NULL)
2933 		return;
2934 
2935 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2936 		zio_gang_tree_free(&gn->gn_child[g]);
2937 
2938 	zio_gang_node_free(gnpp);
2939 }
2940 
2941 static void
2942 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2943 {
2944 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2945 	abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2946 
2947 	ASSERT(gio->io_gang_leader == gio);
2948 	ASSERT(BP_IS_GANG(bp));
2949 
2950 	zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2951 	    zio_gang_tree_assemble_done, gn, gio->io_priority,
2952 	    ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2953 }
2954 
2955 static void
2956 zio_gang_tree_assemble_done(zio_t *zio)
2957 {
2958 	zio_t *gio = zio->io_gang_leader;
2959 	zio_gang_node_t *gn = zio->io_private;
2960 	blkptr_t *bp = zio->io_bp;
2961 
2962 	ASSERT(gio == zio_unique_parent(zio));
2963 	ASSERT(list_is_empty(&zio->io_child_list));
2964 
2965 	if (zio->io_error)
2966 		return;
2967 
2968 	/* this ABD was created from a linear buf in zio_gang_tree_assemble */
2969 	if (BP_SHOULD_BYTESWAP(bp))
2970 		byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2971 
2972 	ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2973 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2974 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2975 
2976 	abd_free(zio->io_abd);
2977 
2978 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2979 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2980 		if (!BP_IS_GANG(gbp))
2981 			continue;
2982 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2983 	}
2984 }
2985 
2986 static void
2987 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2988     uint64_t offset)
2989 {
2990 	zio_t *gio = pio->io_gang_leader;
2991 	zio_t *zio;
2992 
2993 	ASSERT(BP_IS_GANG(bp) == !!gn);
2994 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2995 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2996 
2997 	/*
2998 	 * If you're a gang header, your data is in gn->gn_gbh.
2999 	 * If you're a gang member, your data is in 'data' and gn == NULL.
3000 	 */
3001 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
3002 
3003 	if (gn != NULL) {
3004 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
3005 
3006 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3007 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
3008 			if (BP_IS_HOLE(gbp))
3009 				continue;
3010 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
3011 			    offset);
3012 			offset += BP_GET_PSIZE(gbp);
3013 		}
3014 	}
3015 
3016 	if (gn == gio->io_gang_tree)
3017 		ASSERT3U(gio->io_size, ==, offset);
3018 
3019 	if (zio != pio)
3020 		zio_nowait(zio);
3021 }
3022 
3023 static zio_t *
3024 zio_gang_assemble(zio_t *zio)
3025 {
3026 	blkptr_t *bp = zio->io_bp;
3027 
3028 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
3029 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3030 
3031 	zio->io_gang_leader = zio;
3032 
3033 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
3034 
3035 	return (zio);
3036 }
3037 
3038 static zio_t *
3039 zio_gang_issue(zio_t *zio)
3040 {
3041 	blkptr_t *bp = zio->io_bp;
3042 
3043 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
3044 		return (NULL);
3045 	}
3046 
3047 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
3048 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3049 
3050 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
3051 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
3052 		    0);
3053 	else
3054 		zio_gang_tree_free(&zio->io_gang_tree);
3055 
3056 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3057 
3058 	return (zio);
3059 }
3060 
3061 static void
3062 zio_gang_inherit_allocator(zio_t *pio, zio_t *cio)
3063 {
3064 	cio->io_allocator = pio->io_allocator;
3065 }
3066 
3067 static void
3068 zio_write_gang_member_ready(zio_t *zio)
3069 {
3070 	zio_t *pio = zio_unique_parent(zio);
3071 	dva_t *cdva = zio->io_bp->blk_dva;
3072 	dva_t *pdva = pio->io_bp->blk_dva;
3073 	uint64_t asize;
3074 	zio_t *gio __maybe_unused = zio->io_gang_leader;
3075 
3076 	if (BP_IS_HOLE(zio->io_bp))
3077 		return;
3078 
3079 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
3080 
3081 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
3082 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
3083 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
3084 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
3085 	VERIFY3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
3086 
3087 	mutex_enter(&pio->io_lock);
3088 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
3089 		ASSERT(DVA_GET_GANG(&pdva[d]));
3090 		asize = DVA_GET_ASIZE(&pdva[d]);
3091 		asize += DVA_GET_ASIZE(&cdva[d]);
3092 		DVA_SET_ASIZE(&pdva[d], asize);
3093 	}
3094 	mutex_exit(&pio->io_lock);
3095 }
3096 
3097 static void
3098 zio_write_gang_done(zio_t *zio)
3099 {
3100 	/*
3101 	 * The io_abd field will be NULL for a zio with no data.  The io_flags
3102 	 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
3103 	 * check for it here as it is cleared in zio_ready.
3104 	 */
3105 	if (zio->io_abd != NULL)
3106 		abd_free(zio->io_abd);
3107 }
3108 
3109 static zio_t *
3110 zio_write_gang_block(zio_t *pio, metaslab_class_t *mc)
3111 {
3112 	spa_t *spa = pio->io_spa;
3113 	blkptr_t *bp = pio->io_bp;
3114 	zio_t *gio = pio->io_gang_leader;
3115 	zio_t *zio;
3116 	zio_gang_node_t *gn, **gnpp;
3117 	zio_gbh_phys_t *gbh;
3118 	abd_t *gbh_abd;
3119 	uint64_t txg = pio->io_txg;
3120 	uint64_t resid = pio->io_size;
3121 	uint64_t lsize;
3122 	int copies = gio->io_prop.zp_copies;
3123 	zio_prop_t zp;
3124 	int error;
3125 	boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
3126 
3127 	/*
3128 	 * If one copy was requested, store 2 copies of the GBH, so that we
3129 	 * can still traverse all the data (e.g. to free or scrub) even if a
3130 	 * block is damaged.  Note that we can't store 3 copies of the GBH in
3131 	 * all cases, e.g. with encryption, which uses DVA[2] for the IV+salt.
3132 	 */
3133 	int gbh_copies = copies;
3134 	if (gbh_copies == 1) {
3135 		gbh_copies = MIN(2, spa_max_replication(spa));
3136 	}
3137 
3138 	ASSERT(ZIO_HAS_ALLOCATOR(pio));
3139 	int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
3140 	if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3141 		ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3142 		ASSERT(has_data);
3143 
3144 		flags |= METASLAB_ASYNC_ALLOC;
3145 		VERIFY(zfs_refcount_held(&mc->mc_allocator[pio->io_allocator].
3146 		    mca_alloc_slots, pio));
3147 
3148 		/*
3149 		 * The logical zio has already placed a reservation for
3150 		 * 'copies' allocation slots but gang blocks may require
3151 		 * additional copies. These additional copies
3152 		 * (i.e. gbh_copies - copies) are guaranteed to succeed
3153 		 * since metaslab_class_throttle_reserve() always allows
3154 		 * additional reservations for gang blocks.
3155 		 */
3156 		VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
3157 		    pio->io_allocator, pio, flags));
3158 	}
3159 
3160 	error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
3161 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
3162 	    &pio->io_alloc_list, pio, pio->io_allocator);
3163 	if (error) {
3164 		if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3165 			ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3166 			ASSERT(has_data);
3167 
3168 			/*
3169 			 * If we failed to allocate the gang block header then
3170 			 * we remove any additional allocation reservations that
3171 			 * we placed here. The original reservation will
3172 			 * be removed when the logical I/O goes to the ready
3173 			 * stage.
3174 			 */
3175 			metaslab_class_throttle_unreserve(mc,
3176 			    gbh_copies - copies, pio->io_allocator, pio);
3177 		}
3178 
3179 		pio->io_error = error;
3180 		return (pio);
3181 	}
3182 
3183 	if (pio == gio) {
3184 		gnpp = &gio->io_gang_tree;
3185 	} else {
3186 		gnpp = pio->io_private;
3187 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
3188 	}
3189 
3190 	gn = zio_gang_node_alloc(gnpp);
3191 	gbh = gn->gn_gbh;
3192 	memset(gbh, 0, SPA_GANGBLOCKSIZE);
3193 	gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
3194 
3195 	/*
3196 	 * Create the gang header.
3197 	 */
3198 	zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
3199 	    zio_write_gang_done, NULL, pio->io_priority,
3200 	    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
3201 
3202 	zio_gang_inherit_allocator(pio, zio);
3203 
3204 	/*
3205 	 * Create and nowait the gang children.
3206 	 */
3207 	for (int g = 0; resid != 0; resid -= lsize, g++) {
3208 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
3209 		    SPA_MINBLOCKSIZE);
3210 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
3211 
3212 		zp.zp_checksum = gio->io_prop.zp_checksum;
3213 		zp.zp_compress = ZIO_COMPRESS_OFF;
3214 		zp.zp_complevel = gio->io_prop.zp_complevel;
3215 		zp.zp_type = zp.zp_storage_type = DMU_OT_NONE;
3216 		zp.zp_level = 0;
3217 		zp.zp_copies = gio->io_prop.zp_copies;
3218 		zp.zp_dedup = B_FALSE;
3219 		zp.zp_dedup_verify = B_FALSE;
3220 		zp.zp_nopwrite = B_FALSE;
3221 		zp.zp_encrypt = gio->io_prop.zp_encrypt;
3222 		zp.zp_byteorder = gio->io_prop.zp_byteorder;
3223 		zp.zp_direct_write = B_FALSE;
3224 		memset(zp.zp_salt, 0, ZIO_DATA_SALT_LEN);
3225 		memset(zp.zp_iv, 0, ZIO_DATA_IV_LEN);
3226 		memset(zp.zp_mac, 0, ZIO_DATA_MAC_LEN);
3227 
3228 		zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
3229 		    has_data ? abd_get_offset(pio->io_abd, pio->io_size -
3230 		    resid) : NULL, lsize, lsize, &zp,
3231 		    zio_write_gang_member_ready, NULL,
3232 		    zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
3233 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
3234 
3235 		zio_gang_inherit_allocator(zio, cio);
3236 
3237 		if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3238 			ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3239 			ASSERT(has_data);
3240 
3241 			/*
3242 			 * Gang children won't throttle but we should
3243 			 * account for their work, so reserve an allocation
3244 			 * slot for them here.
3245 			 */
3246 			VERIFY(metaslab_class_throttle_reserve(mc,
3247 			    zp.zp_copies, cio->io_allocator, cio, flags));
3248 		}
3249 		zio_nowait(cio);
3250 	}
3251 
3252 	/*
3253 	 * Set pio's pipeline to just wait for zio to finish.
3254 	 */
3255 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3256 
3257 	zio_nowait(zio);
3258 
3259 	return (pio);
3260 }
3261 
3262 /*
3263  * The zio_nop_write stage in the pipeline determines if allocating a
3264  * new bp is necessary.  The nopwrite feature can handle writes in
3265  * either syncing or open context (i.e. zil writes) and as a result is
3266  * mutually exclusive with dedup.
3267  *
3268  * By leveraging a cryptographically secure checksum, such as SHA256, we
3269  * can compare the checksums of the new data and the old to determine if
3270  * allocating a new block is required.  Note that our requirements for
3271  * cryptographic strength are fairly weak: there can't be any accidental
3272  * hash collisions, but we don't need to be secure against intentional
3273  * (malicious) collisions.  To trigger a nopwrite, you have to be able
3274  * to write the file to begin with, and triggering an incorrect (hash
3275  * collision) nopwrite is no worse than simply writing to the file.
3276  * That said, there are no known attacks against the checksum algorithms
3277  * used for nopwrite, assuming that the salt and the checksums
3278  * themselves remain secret.
3279  */
3280 static zio_t *
3281 zio_nop_write(zio_t *zio)
3282 {
3283 	blkptr_t *bp = zio->io_bp;
3284 	blkptr_t *bp_orig = &zio->io_bp_orig;
3285 	zio_prop_t *zp = &zio->io_prop;
3286 
3287 	ASSERT(BP_IS_HOLE(bp));
3288 	ASSERT(BP_GET_LEVEL(bp) == 0);
3289 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
3290 	ASSERT(zp->zp_nopwrite);
3291 	ASSERT(!zp->zp_dedup);
3292 	ASSERT(zio->io_bp_override == NULL);
3293 	ASSERT(IO_IS_ALLOCATING(zio));
3294 
3295 	/*
3296 	 * Check to see if the original bp and the new bp have matching
3297 	 * characteristics (i.e. same checksum, compression algorithms, etc).
3298 	 * If they don't then just continue with the pipeline which will
3299 	 * allocate a new bp.
3300 	 */
3301 	if (BP_IS_HOLE(bp_orig) ||
3302 	    !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
3303 	    ZCHECKSUM_FLAG_NOPWRITE) ||
3304 	    BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) ||
3305 	    BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
3306 	    BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
3307 	    BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
3308 	    zp->zp_copies != BP_GET_NDVAS(bp_orig))
3309 		return (zio);
3310 
3311 	/*
3312 	 * If the checksums match then reset the pipeline so that we
3313 	 * avoid allocating a new bp and issuing any I/O.
3314 	 */
3315 	if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
3316 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
3317 		    ZCHECKSUM_FLAG_NOPWRITE);
3318 		ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
3319 		ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
3320 		ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
3321 		ASSERT3U(bp->blk_prop, ==, bp_orig->blk_prop);
3322 
3323 		/*
3324 		 * If we're overwriting a block that is currently on an
3325 		 * indirect vdev, then ignore the nopwrite request and
3326 		 * allow a new block to be allocated on a concrete vdev.
3327 		 */
3328 		spa_config_enter(zio->io_spa, SCL_VDEV, FTAG, RW_READER);
3329 		for (int d = 0; d < BP_GET_NDVAS(bp_orig); d++) {
3330 			vdev_t *tvd = vdev_lookup_top(zio->io_spa,
3331 			    DVA_GET_VDEV(&bp_orig->blk_dva[d]));
3332 			if (tvd->vdev_ops == &vdev_indirect_ops) {
3333 				spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3334 				return (zio);
3335 			}
3336 		}
3337 		spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3338 
3339 		*bp = *bp_orig;
3340 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3341 		zio->io_flags |= ZIO_FLAG_NOPWRITE;
3342 	}
3343 
3344 	return (zio);
3345 }
3346 
3347 /*
3348  * ==========================================================================
3349  * Block Reference Table
3350  * ==========================================================================
3351  */
3352 static zio_t *
3353 zio_brt_free(zio_t *zio)
3354 {
3355 	blkptr_t *bp;
3356 
3357 	bp = zio->io_bp;
3358 
3359 	if (BP_GET_LEVEL(bp) > 0 ||
3360 	    BP_IS_METADATA(bp) ||
3361 	    !brt_maybe_exists(zio->io_spa, bp)) {
3362 		return (zio);
3363 	}
3364 
3365 	if (!brt_entry_decref(zio->io_spa, bp)) {
3366 		/*
3367 		 * This isn't the last reference, so we cannot free
3368 		 * the data yet.
3369 		 */
3370 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3371 	}
3372 
3373 	return (zio);
3374 }
3375 
3376 /*
3377  * ==========================================================================
3378  * Dedup
3379  * ==========================================================================
3380  */
3381 static void
3382 zio_ddt_child_read_done(zio_t *zio)
3383 {
3384 	blkptr_t *bp = zio->io_bp;
3385 	ddt_t *ddt;
3386 	ddt_entry_t *dde = zio->io_private;
3387 	zio_t *pio = zio_unique_parent(zio);
3388 
3389 	mutex_enter(&pio->io_lock);
3390 	ddt = ddt_select(zio->io_spa, bp);
3391 
3392 	if (zio->io_error == 0) {
3393 		ddt_phys_variant_t v = ddt_phys_select(ddt, dde, bp);
3394 		/* this phys variant doesn't need repair */
3395 		ddt_phys_clear(dde->dde_phys, v);
3396 	}
3397 
3398 	if (zio->io_error == 0 && dde->dde_io->dde_repair_abd == NULL)
3399 		dde->dde_io->dde_repair_abd = zio->io_abd;
3400 	else
3401 		abd_free(zio->io_abd);
3402 	mutex_exit(&pio->io_lock);
3403 }
3404 
3405 static zio_t *
3406 zio_ddt_read_start(zio_t *zio)
3407 {
3408 	blkptr_t *bp = zio->io_bp;
3409 
3410 	ASSERT(BP_GET_DEDUP(bp));
3411 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3412 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3413 
3414 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
3415 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
3416 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
3417 		ddt_phys_variant_t v_self = ddt_phys_select(ddt, dde, bp);
3418 		ddt_univ_phys_t *ddp = dde->dde_phys;
3419 		blkptr_t blk;
3420 
3421 		ASSERT(zio->io_vsd == NULL);
3422 		zio->io_vsd = dde;
3423 
3424 		if (v_self == DDT_PHYS_NONE)
3425 			return (zio);
3426 
3427 		/* issue I/O for the other copies */
3428 		for (int p = 0; p < DDT_NPHYS(ddt); p++) {
3429 			ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p);
3430 
3431 			if (ddt_phys_birth(ddp, v) == 0 || v == v_self)
3432 				continue;
3433 
3434 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key,
3435 			    ddp, v, &blk);
3436 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
3437 			    abd_alloc_for_io(zio->io_size, B_TRUE),
3438 			    zio->io_size, zio_ddt_child_read_done, dde,
3439 			    zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
3440 			    ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
3441 		}
3442 		return (zio);
3443 	}
3444 
3445 	zio_nowait(zio_read(zio, zio->io_spa, bp,
3446 	    zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
3447 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
3448 
3449 	return (zio);
3450 }
3451 
3452 static zio_t *
3453 zio_ddt_read_done(zio_t *zio)
3454 {
3455 	blkptr_t *bp = zio->io_bp;
3456 
3457 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
3458 		return (NULL);
3459 	}
3460 
3461 	ASSERT(BP_GET_DEDUP(bp));
3462 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3463 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3464 
3465 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
3466 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
3467 		ddt_entry_t *dde = zio->io_vsd;
3468 		if (ddt == NULL) {
3469 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
3470 			return (zio);
3471 		}
3472 		if (dde == NULL) {
3473 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
3474 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
3475 			return (NULL);
3476 		}
3477 		if (dde->dde_io->dde_repair_abd != NULL) {
3478 			abd_copy(zio->io_abd, dde->dde_io->dde_repair_abd,
3479 			    zio->io_size);
3480 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
3481 		}
3482 		ddt_repair_done(ddt, dde);
3483 		zio->io_vsd = NULL;
3484 	}
3485 
3486 	ASSERT(zio->io_vsd == NULL);
3487 
3488 	return (zio);
3489 }
3490 
3491 static boolean_t
3492 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
3493 {
3494 	spa_t *spa = zio->io_spa;
3495 	boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW);
3496 
3497 	ASSERT(!(zio->io_bp_override && do_raw));
3498 
3499 	/*
3500 	 * Note: we compare the original data, not the transformed data,
3501 	 * because when zio->io_bp is an override bp, we will not have
3502 	 * pushed the I/O transforms.  That's an important optimization
3503 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
3504 	 * However, we should never get a raw, override zio so in these
3505 	 * cases we can compare the io_abd directly. This is useful because
3506 	 * it allows us to do dedup verification even if we don't have access
3507 	 * to the original data (for instance, if the encryption keys aren't
3508 	 * loaded).
3509 	 */
3510 
3511 	for (int p = 0; p < DDT_NPHYS(ddt); p++) {
3512 		if (DDT_PHYS_IS_DITTO(ddt, p))
3513 			continue;
3514 
3515 		if (dde->dde_io == NULL)
3516 			continue;
3517 
3518 		zio_t *lio = dde->dde_io->dde_lead_zio[p];
3519 		if (lio == NULL)
3520 			continue;
3521 
3522 		if (do_raw)
3523 			return (lio->io_size != zio->io_size ||
3524 			    abd_cmp(zio->io_abd, lio->io_abd) != 0);
3525 
3526 		return (lio->io_orig_size != zio->io_orig_size ||
3527 		    abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0);
3528 	}
3529 
3530 	for (int p = 0; p < DDT_NPHYS(ddt); p++) {
3531 		ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p);
3532 		uint64_t phys_birth = ddt_phys_birth(dde->dde_phys, v);
3533 
3534 		if (phys_birth != 0 && do_raw) {
3535 			blkptr_t blk = *zio->io_bp;
3536 			uint64_t psize;
3537 			abd_t *tmpabd;
3538 			int error;
3539 
3540 			ddt_bp_fill(dde->dde_phys, v, &blk, phys_birth);
3541 			psize = BP_GET_PSIZE(&blk);
3542 
3543 			if (psize != zio->io_size)
3544 				return (B_TRUE);
3545 
3546 			ddt_exit(ddt);
3547 
3548 			tmpabd = abd_alloc_for_io(psize, B_TRUE);
3549 
3550 			error = zio_wait(zio_read(NULL, spa, &blk, tmpabd,
3551 			    psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
3552 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
3553 			    ZIO_FLAG_RAW, &zio->io_bookmark));
3554 
3555 			if (error == 0) {
3556 				if (abd_cmp(tmpabd, zio->io_abd) != 0)
3557 					error = SET_ERROR(ENOENT);
3558 			}
3559 
3560 			abd_free(tmpabd);
3561 			ddt_enter(ddt);
3562 			return (error != 0);
3563 		} else if (phys_birth != 0) {
3564 			arc_buf_t *abuf = NULL;
3565 			arc_flags_t aflags = ARC_FLAG_WAIT;
3566 			blkptr_t blk = *zio->io_bp;
3567 			int error;
3568 
3569 			ddt_bp_fill(dde->dde_phys, v, &blk, phys_birth);
3570 
3571 			if (BP_GET_LSIZE(&blk) != zio->io_orig_size)
3572 				return (B_TRUE);
3573 
3574 			ddt_exit(ddt);
3575 
3576 			error = arc_read(NULL, spa, &blk,
3577 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
3578 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
3579 			    &aflags, &zio->io_bookmark);
3580 
3581 			if (error == 0) {
3582 				if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
3583 				    zio->io_orig_size) != 0)
3584 					error = SET_ERROR(ENOENT);
3585 				arc_buf_destroy(abuf, &abuf);
3586 			}
3587 
3588 			ddt_enter(ddt);
3589 			return (error != 0);
3590 		}
3591 	}
3592 
3593 	return (B_FALSE);
3594 }
3595 
3596 static void
3597 zio_ddt_child_write_done(zio_t *zio)
3598 {
3599 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3600 	ddt_entry_t *dde = zio->io_private;
3601 
3602 	zio_link_t *zl = NULL;
3603 	ASSERT3P(zio_walk_parents(zio, &zl), !=, NULL);
3604 
3605 	int p = DDT_PHYS_FOR_COPIES(ddt, zio->io_prop.zp_copies);
3606 	ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p);
3607 	ddt_univ_phys_t *ddp = dde->dde_phys;
3608 
3609 	ddt_enter(ddt);
3610 
3611 	/* we're the lead, so once we're done there's no one else outstanding */
3612 	if (dde->dde_io->dde_lead_zio[p] == zio)
3613 		dde->dde_io->dde_lead_zio[p] = NULL;
3614 
3615 	ddt_univ_phys_t *orig = &dde->dde_io->dde_orig_phys;
3616 
3617 	if (zio->io_error != 0) {
3618 		/*
3619 		 * The write failed, so we're about to abort the entire IO
3620 		 * chain. We need to revert the entry back to what it was at
3621 		 * the last time it was successfully extended.
3622 		 */
3623 		ddt_phys_copy(ddp, orig, v);
3624 		ddt_phys_clear(orig, v);
3625 
3626 		ddt_exit(ddt);
3627 		return;
3628 	}
3629 
3630 	/*
3631 	 * We've successfully added new DVAs to the entry. Clear the saved
3632 	 * state or, if there's still outstanding IO, remember it so we can
3633 	 * revert to a known good state if that IO fails.
3634 	 */
3635 	if (dde->dde_io->dde_lead_zio[p] == NULL)
3636 		ddt_phys_clear(orig, v);
3637 	else
3638 		ddt_phys_copy(orig, ddp, v);
3639 
3640 	/*
3641 	 * Add references for all dedup writes that were waiting on the
3642 	 * physical one, skipping any other physical writes that are waiting.
3643 	 */
3644 	zio_t *pio;
3645 	zl = NULL;
3646 	while ((pio = zio_walk_parents(zio, &zl)) != NULL) {
3647 		if (!(pio->io_flags & ZIO_FLAG_DDT_CHILD))
3648 			ddt_phys_addref(ddp, v);
3649 	}
3650 
3651 	ddt_exit(ddt);
3652 }
3653 
3654 static void
3655 zio_ddt_child_write_ready(zio_t *zio)
3656 {
3657 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3658 	ddt_entry_t *dde = zio->io_private;
3659 
3660 	zio_link_t *zl = NULL;
3661 	ASSERT3P(zio_walk_parents(zio, &zl), !=, NULL);
3662 
3663 	int p = DDT_PHYS_FOR_COPIES(ddt, zio->io_prop.zp_copies);
3664 	ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p);
3665 
3666 	if (zio->io_error != 0)
3667 		return;
3668 
3669 	ddt_enter(ddt);
3670 
3671 	ddt_phys_extend(dde->dde_phys, v, zio->io_bp);
3672 
3673 	zio_t *pio;
3674 	zl = NULL;
3675 	while ((pio = zio_walk_parents(zio, &zl)) != NULL) {
3676 		if (!(pio->io_flags & ZIO_FLAG_DDT_CHILD))
3677 			ddt_bp_fill(dde->dde_phys, v, pio->io_bp, zio->io_txg);
3678 	}
3679 
3680 	ddt_exit(ddt);
3681 }
3682 
3683 static zio_t *
3684 zio_ddt_write(zio_t *zio)
3685 {
3686 	spa_t *spa = zio->io_spa;
3687 	blkptr_t *bp = zio->io_bp;
3688 	uint64_t txg = zio->io_txg;
3689 	zio_prop_t *zp = &zio->io_prop;
3690 	ddt_t *ddt = ddt_select(spa, bp);
3691 	ddt_entry_t *dde;
3692 
3693 	ASSERT(BP_GET_DEDUP(bp));
3694 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
3695 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
3696 	ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
3697 	/*
3698 	 * Deduplication will not take place for Direct I/O writes. The
3699 	 * ddt_tree will be emptied in syncing context. Direct I/O writes take
3700 	 * place in the open-context. Direct I/O write can not attempt to
3701 	 * modify the ddt_tree while issuing out a write.
3702 	 */
3703 	ASSERT3B(zio->io_prop.zp_direct_write, ==, B_FALSE);
3704 
3705 	ddt_enter(ddt);
3706 	dde = ddt_lookup(ddt, bp);
3707 	if (dde == NULL) {
3708 		/* DDT size is over its quota so no new entries */
3709 		zp->zp_dedup = B_FALSE;
3710 		BP_SET_DEDUP(bp, B_FALSE);
3711 		if (zio->io_bp_override == NULL)
3712 			zio->io_pipeline = ZIO_WRITE_PIPELINE;
3713 		ddt_exit(ddt);
3714 		return (zio);
3715 	}
3716 
3717 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
3718 		/*
3719 		 * If we're using a weak checksum, upgrade to a strong checksum
3720 		 * and try again.  If we're already using a strong checksum,
3721 		 * we can't resolve it, so just convert to an ordinary write.
3722 		 * (And automatically e-mail a paper to Nature?)
3723 		 */
3724 		if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
3725 		    ZCHECKSUM_FLAG_DEDUP)) {
3726 			zp->zp_checksum = spa_dedup_checksum(spa);
3727 			zio_pop_transforms(zio);
3728 			zio->io_stage = ZIO_STAGE_OPEN;
3729 			BP_ZERO(bp);
3730 		} else {
3731 			zp->zp_dedup = B_FALSE;
3732 			BP_SET_DEDUP(bp, B_FALSE);
3733 		}
3734 		ASSERT(!BP_GET_DEDUP(bp));
3735 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
3736 		ddt_exit(ddt);
3737 		return (zio);
3738 	}
3739 
3740 	int p = DDT_PHYS_FOR_COPIES(ddt, zp->zp_copies);
3741 	ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p);
3742 	ddt_univ_phys_t *ddp = dde->dde_phys;
3743 
3744 	/*
3745 	 * In the common cases, at this point we have a regular BP with no
3746 	 * allocated DVAs, and the corresponding DDT entry for its checksum.
3747 	 * Our goal is to fill the BP with enough DVAs to satisfy its copies=
3748 	 * requirement.
3749 	 *
3750 	 * One of three things needs to happen to fulfill this:
3751 	 *
3752 	 * - if the DDT entry has enough DVAs to satisfy the BP, we just copy
3753 	 *   them out of the entry and return;
3754 	 *
3755 	 * - if the DDT entry has no DVAs (ie its brand new), then we have to
3756 	 *   issue the write as normal so that DVAs can be allocated and the
3757 	 *   data land on disk. We then copy the DVAs into the DDT entry on
3758 	 *   return.
3759 	 *
3760 	 * - if the DDT entry has some DVAs, but too few, we have to issue the
3761 	 *   write, adjusted to have allocate fewer copies. When it returns, we
3762 	 *   add the new DVAs to the DDT entry, and update the BP to have the
3763 	 *   full amount it originally requested.
3764 	 *
3765 	 * In all cases, if there's already a writing IO in flight, we need to
3766 	 * defer the action until after the write is done. If our action is to
3767 	 * write, we need to adjust our request for additional DVAs to match
3768 	 * what will be in the DDT entry after it completes. In this way every
3769 	 * IO can be guaranteed to recieve enough DVAs simply by joining the
3770 	 * end of the chain and letting the sequence play out.
3771 	 */
3772 
3773 	/*
3774 	 * Number of DVAs in the DDT entry. If the BP is encrypted we ignore
3775 	 * the third one as normal.
3776 	 */
3777 	int have_dvas = ddt_phys_dva_count(ddp, v, BP_IS_ENCRYPTED(bp));
3778 	IMPLY(have_dvas == 0, ddt_phys_birth(ddp, v) == 0);
3779 
3780 	/* Number of DVAs requested bya the IO. */
3781 	uint8_t need_dvas = zp->zp_copies;
3782 
3783 	/*
3784 	 * What we do next depends on whether or not there's IO outstanding that
3785 	 * will update this entry.
3786 	 */
3787 	if (dde->dde_io == NULL || dde->dde_io->dde_lead_zio[p] == NULL) {
3788 		/*
3789 		 * No IO outstanding, so we only need to worry about ourselves.
3790 		 */
3791 
3792 		/*
3793 		 * Override BPs bring their own DVAs and their own problems.
3794 		 */
3795 		if (zio->io_bp_override) {
3796 			/*
3797 			 * For a brand-new entry, all the work has been done
3798 			 * for us, and we can just fill it out from the provided
3799 			 * block and leave.
3800 			 */
3801 			if (have_dvas == 0) {
3802 				ASSERT(BP_GET_LOGICAL_BIRTH(bp) == txg);
3803 				ASSERT(BP_EQUAL(bp, zio->io_bp_override));
3804 				ddt_phys_extend(ddp, v, bp);
3805 				ddt_phys_addref(ddp, v);
3806 				ddt_exit(ddt);
3807 				return (zio);
3808 			}
3809 
3810 			/*
3811 			 * If we already have this entry, then we want to treat
3812 			 * it like a regular write. To do this we just wipe
3813 			 * them out and proceed like a regular write.
3814 			 *
3815 			 * Even if there are some DVAs in the entry, we still
3816 			 * have to clear them out. We can't use them to fill
3817 			 * out the dedup entry, as they are all referenced
3818 			 * together by a bp already on disk, and will be freed
3819 			 * as a group.
3820 			 */
3821 			BP_ZERO_DVAS(bp);
3822 			BP_SET_BIRTH(bp, 0, 0);
3823 		}
3824 
3825 		/*
3826 		 * If there are enough DVAs in the entry to service our request,
3827 		 * then we can just use them as-is.
3828 		 */
3829 		if (have_dvas >= need_dvas) {
3830 			ddt_bp_fill(ddp, v, bp, txg);
3831 			ddt_phys_addref(ddp, v);
3832 			ddt_exit(ddt);
3833 			return (zio);
3834 		}
3835 
3836 		/*
3837 		 * Otherwise, we have to issue IO to fill the entry up to the
3838 		 * amount we need.
3839 		 */
3840 		need_dvas -= have_dvas;
3841 	} else {
3842 		/*
3843 		 * There's a write in-flight. If there's already enough DVAs on
3844 		 * the entry, then either there were already enough to start
3845 		 * with, or the in-flight IO is between READY and DONE, and so
3846 		 * has extended the entry with new DVAs. Either way, we don't
3847 		 * need to do anything, we can just slot in behind it.
3848 		 */
3849 
3850 		if (zio->io_bp_override) {
3851 			/*
3852 			 * If there's a write out, then we're soon going to
3853 			 * have our own copies of this block, so clear out the
3854 			 * override block and treat it as a regular dedup
3855 			 * write. See comment above.
3856 			 */
3857 			BP_ZERO_DVAS(bp);
3858 			BP_SET_BIRTH(bp, 0, 0);
3859 		}
3860 
3861 		if (have_dvas >= need_dvas) {
3862 			/*
3863 			 * A minor point: there might already be enough
3864 			 * committed DVAs in the entry to service our request,
3865 			 * but we don't know which are completed and which are
3866 			 * allocated but not yet written. In this case, should
3867 			 * the IO for the new DVAs fail, we will be on the end
3868 			 * of the IO chain and will also recieve an error, even
3869 			 * though our request could have been serviced.
3870 			 *
3871 			 * This is an extremely rare case, as it requires the
3872 			 * original block to be copied with a request for a
3873 			 * larger number of DVAs, then copied again requesting
3874 			 * the same (or already fulfilled) number of DVAs while
3875 			 * the first request is active, and then that first
3876 			 * request errors. In return, the logic required to
3877 			 * catch and handle it is complex. For now, I'm just
3878 			 * not going to bother with it.
3879 			 */
3880 
3881 			/*
3882 			 * We always fill the bp here as we may have arrived
3883 			 * after the in-flight write has passed READY, and so
3884 			 * missed out.
3885 			 */
3886 			ddt_bp_fill(ddp, v, bp, txg);
3887 			zio_add_child(zio, dde->dde_io->dde_lead_zio[p]);
3888 			ddt_exit(ddt);
3889 			return (zio);
3890 		}
3891 
3892 		/*
3893 		 * There's not enough in the entry yet, so we need to look at
3894 		 * the write in-flight and see how many DVAs it will have once
3895 		 * it completes.
3896 		 *
3897 		 * The in-flight write has potentially had its copies request
3898 		 * reduced (if we're filling out an existing entry), so we need
3899 		 * to reach in and get the original write to find out what it is
3900 		 * expecting.
3901 		 *
3902 		 * Note that the parent of the lead zio will always have the
3903 		 * highest zp_copies of any zio in the chain, because ones that
3904 		 * can be serviced without additional IO are always added to
3905 		 * the back of the chain.
3906 		 */
3907 		zio_link_t *zl = NULL;
3908 		zio_t *pio =
3909 		    zio_walk_parents(dde->dde_io->dde_lead_zio[p], &zl);
3910 		ASSERT(pio);
3911 		uint8_t parent_dvas = pio->io_prop.zp_copies;
3912 
3913 		if (parent_dvas >= need_dvas) {
3914 			zio_add_child(zio, dde->dde_io->dde_lead_zio[p]);
3915 			ddt_exit(ddt);
3916 			return (zio);
3917 		}
3918 
3919 		/*
3920 		 * Still not enough, so we will need to issue to get the
3921 		 * shortfall.
3922 		 */
3923 		need_dvas -= parent_dvas;
3924 	}
3925 
3926 	/*
3927 	 * We need to write. We will create a new write with the copies
3928 	 * property adjusted to match the number of DVAs we need to need to
3929 	 * grow the DDT entry by to satisfy the request.
3930 	 */
3931 	zio_prop_t czp = *zp;
3932 	czp.zp_copies = need_dvas;
3933 	zio_t *cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3934 	    zio->io_orig_size, zio->io_orig_size, &czp,
3935 	    zio_ddt_child_write_ready, NULL,
3936 	    zio_ddt_child_write_done, dde, zio->io_priority,
3937 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3938 
3939 	zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
3940 
3941 	/*
3942 	 * We are the new lead zio, because our parent has the highest
3943 	 * zp_copies that has been requested for this entry so far.
3944 	 */
3945 	ddt_alloc_entry_io(dde);
3946 	if (dde->dde_io->dde_lead_zio[p] == NULL) {
3947 		/*
3948 		 * First time out, take a copy of the stable entry to revert
3949 		 * to if there's an error (see zio_ddt_child_write_done())
3950 		 */
3951 		ddt_phys_copy(&dde->dde_io->dde_orig_phys, dde->dde_phys, v);
3952 	} else {
3953 		/*
3954 		 * Make the existing chain our child, because it cannot
3955 		 * complete until we have.
3956 		 */
3957 		zio_add_child(cio, dde->dde_io->dde_lead_zio[p]);
3958 	}
3959 	dde->dde_io->dde_lead_zio[p] = cio;
3960 
3961 	ddt_exit(ddt);
3962 
3963 	zio_nowait(cio);
3964 
3965 	return (zio);
3966 }
3967 
3968 static ddt_entry_t *freedde; /* for debugging */
3969 
3970 static zio_t *
3971 zio_ddt_free(zio_t *zio)
3972 {
3973 	spa_t *spa = zio->io_spa;
3974 	blkptr_t *bp = zio->io_bp;
3975 	ddt_t *ddt = ddt_select(spa, bp);
3976 	ddt_entry_t *dde = NULL;
3977 
3978 	ASSERT(BP_GET_DEDUP(bp));
3979 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3980 
3981 	ddt_enter(ddt);
3982 	freedde = dde = ddt_lookup(ddt, bp);
3983 	if (dde) {
3984 		ddt_phys_variant_t v = ddt_phys_select(ddt, dde, bp);
3985 		if (v != DDT_PHYS_NONE)
3986 			ddt_phys_decref(dde->dde_phys, v);
3987 	}
3988 	ddt_exit(ddt);
3989 
3990 	/*
3991 	 * When no entry was found, it must have been pruned,
3992 	 * so we can free it now instead of decrementing the
3993 	 * refcount in the DDT.
3994 	 */
3995 	if (!dde) {
3996 		BP_SET_DEDUP(bp, 0);
3997 		zio->io_pipeline |= ZIO_STAGE_DVA_FREE;
3998 	}
3999 
4000 	return (zio);
4001 }
4002 
4003 /*
4004  * ==========================================================================
4005  * Allocate and free blocks
4006  * ==========================================================================
4007  */
4008 
4009 static zio_t *
4010 zio_io_to_allocate(spa_t *spa, int allocator)
4011 {
4012 	zio_t *zio;
4013 
4014 	ASSERT(MUTEX_HELD(&spa->spa_allocs[allocator].spaa_lock));
4015 
4016 	zio = avl_first(&spa->spa_allocs[allocator].spaa_tree);
4017 	if (zio == NULL)
4018 		return (NULL);
4019 
4020 	ASSERT(IO_IS_ALLOCATING(zio));
4021 	ASSERT(ZIO_HAS_ALLOCATOR(zio));
4022 
4023 	/*
4024 	 * Try to place a reservation for this zio. If we're unable to
4025 	 * reserve then we throttle.
4026 	 */
4027 	ASSERT3U(zio->io_allocator, ==, allocator);
4028 	if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
4029 	    zio->io_prop.zp_copies, allocator, zio, 0)) {
4030 		return (NULL);
4031 	}
4032 
4033 	avl_remove(&spa->spa_allocs[allocator].spaa_tree, zio);
4034 	ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
4035 
4036 	return (zio);
4037 }
4038 
4039 static zio_t *
4040 zio_dva_throttle(zio_t *zio)
4041 {
4042 	spa_t *spa = zio->io_spa;
4043 	zio_t *nio;
4044 	metaslab_class_t *mc;
4045 
4046 	/* locate an appropriate allocation class */
4047 	mc = spa_preferred_class(spa, zio);
4048 
4049 	if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
4050 	    !mc->mc_alloc_throttle_enabled ||
4051 	    zio->io_child_type == ZIO_CHILD_GANG ||
4052 	    zio->io_flags & ZIO_FLAG_NODATA) {
4053 		return (zio);
4054 	}
4055 
4056 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4057 	ASSERT(ZIO_HAS_ALLOCATOR(zio));
4058 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
4059 	ASSERT3U(zio->io_queued_timestamp, >, 0);
4060 	ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
4061 
4062 	int allocator = zio->io_allocator;
4063 	zio->io_metaslab_class = mc;
4064 	mutex_enter(&spa->spa_allocs[allocator].spaa_lock);
4065 	avl_add(&spa->spa_allocs[allocator].spaa_tree, zio);
4066 	nio = zio_io_to_allocate(spa, allocator);
4067 	mutex_exit(&spa->spa_allocs[allocator].spaa_lock);
4068 	return (nio);
4069 }
4070 
4071 static void
4072 zio_allocate_dispatch(spa_t *spa, int allocator)
4073 {
4074 	zio_t *zio;
4075 
4076 	mutex_enter(&spa->spa_allocs[allocator].spaa_lock);
4077 	zio = zio_io_to_allocate(spa, allocator);
4078 	mutex_exit(&spa->spa_allocs[allocator].spaa_lock);
4079 	if (zio == NULL)
4080 		return;
4081 
4082 	ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
4083 	ASSERT0(zio->io_error);
4084 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
4085 }
4086 
4087 static zio_t *
4088 zio_dva_allocate(zio_t *zio)
4089 {
4090 	spa_t *spa = zio->io_spa;
4091 	metaslab_class_t *mc;
4092 	blkptr_t *bp = zio->io_bp;
4093 	int error;
4094 	int flags = 0;
4095 
4096 	if (zio->io_gang_leader == NULL) {
4097 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
4098 		zio->io_gang_leader = zio;
4099 	}
4100 
4101 	ASSERT(BP_IS_HOLE(bp));
4102 	ASSERT0(BP_GET_NDVAS(bp));
4103 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
4104 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
4105 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
4106 
4107 	if (zio->io_flags & ZIO_FLAG_NODATA)
4108 		flags |= METASLAB_DONT_THROTTLE;
4109 	if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
4110 		flags |= METASLAB_GANG_CHILD;
4111 	if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
4112 		flags |= METASLAB_ASYNC_ALLOC;
4113 
4114 	/*
4115 	 * if not already chosen, locate an appropriate allocation class
4116 	 */
4117 	mc = zio->io_metaslab_class;
4118 	if (mc == NULL) {
4119 		mc = spa_preferred_class(spa, zio);
4120 		zio->io_metaslab_class = mc;
4121 	}
4122 	ZIOSTAT_BUMP(ziostat_total_allocations);
4123 
4124 	/*
4125 	 * Try allocating the block in the usual metaslab class.
4126 	 * If that's full, allocate it in the normal class.
4127 	 * If that's full, allocate as a gang block,
4128 	 * and if all are full, the allocation fails (which shouldn't happen).
4129 	 *
4130 	 * Note that we do not fall back on embedded slog (ZIL) space, to
4131 	 * preserve unfragmented slog space, which is critical for decent
4132 	 * sync write performance.  If a log allocation fails, we will fall
4133 	 * back to spa_sync() which is abysmal for performance.
4134 	 */
4135 	ASSERT(ZIO_HAS_ALLOCATOR(zio));
4136 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
4137 	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
4138 	    &zio->io_alloc_list, zio, zio->io_allocator);
4139 
4140 	/*
4141 	 * Fallback to normal class when an alloc class is full
4142 	 */
4143 	if (error == ENOSPC && mc != spa_normal_class(spa)) {
4144 		/*
4145 		 * When the dedup or special class is spilling into the  normal
4146 		 * class, there can still be significant space available due
4147 		 * to deferred frees that are in-flight.  We track the txg when
4148 		 * this occurred and back off adding new DDT entries for a few
4149 		 * txgs to allow the free blocks to be processed.
4150 		 */
4151 		if ((mc == spa_dedup_class(spa) || (spa_special_has_ddt(spa) &&
4152 		    mc == spa_special_class(spa))) &&
4153 		    spa->spa_dedup_class_full_txg != zio->io_txg) {
4154 			spa->spa_dedup_class_full_txg = zio->io_txg;
4155 			zfs_dbgmsg("%s[%d]: %s class spilling, req size %d, "
4156 			    "%llu allocated of %llu",
4157 			    spa_name(spa), (int)zio->io_txg,
4158 			    mc == spa_dedup_class(spa) ? "dedup" : "special",
4159 			    (int)zio->io_size,
4160 			    (u_longlong_t)metaslab_class_get_alloc(mc),
4161 			    (u_longlong_t)metaslab_class_get_space(mc));
4162 		}
4163 
4164 		/*
4165 		 * If throttling, transfer reservation over to normal class.
4166 		 * The io_allocator slot can remain the same even though we
4167 		 * are switching classes.
4168 		 */
4169 		if (mc->mc_alloc_throttle_enabled &&
4170 		    (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) {
4171 			metaslab_class_throttle_unreserve(mc,
4172 			    zio->io_prop.zp_copies, zio->io_allocator, zio);
4173 			zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING;
4174 
4175 			VERIFY(metaslab_class_throttle_reserve(
4176 			    spa_normal_class(spa),
4177 			    zio->io_prop.zp_copies, zio->io_allocator, zio,
4178 			    flags | METASLAB_MUST_RESERVE));
4179 		}
4180 		zio->io_metaslab_class = mc = spa_normal_class(spa);
4181 		if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) {
4182 			zfs_dbgmsg("%s: metaslab allocation failure, "
4183 			    "trying normal class: zio %px, size %llu, error %d",
4184 			    spa_name(spa), zio, (u_longlong_t)zio->io_size,
4185 			    error);
4186 		}
4187 
4188 		ZIOSTAT_BUMP(ziostat_alloc_class_fallbacks);
4189 		error = metaslab_alloc(spa, mc, zio->io_size, bp,
4190 		    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
4191 		    &zio->io_alloc_list, zio, zio->io_allocator);
4192 	}
4193 
4194 	if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) {
4195 		if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) {
4196 			zfs_dbgmsg("%s: metaslab allocation failure, "
4197 			    "trying ganging: zio %px, size %llu, error %d",
4198 			    spa_name(spa), zio, (u_longlong_t)zio->io_size,
4199 			    error);
4200 		}
4201 		ZIOSTAT_BUMP(ziostat_gang_writes);
4202 		if (flags & METASLAB_GANG_CHILD)
4203 			ZIOSTAT_BUMP(ziostat_gang_multilevel);
4204 		return (zio_write_gang_block(zio, mc));
4205 	}
4206 	if (error != 0) {
4207 		if (error != ENOSPC ||
4208 		    (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC)) {
4209 			zfs_dbgmsg("%s: metaslab allocation failure: zio %px, "
4210 			    "size %llu, error %d",
4211 			    spa_name(spa), zio, (u_longlong_t)zio->io_size,
4212 			    error);
4213 		}
4214 		zio->io_error = error;
4215 	}
4216 
4217 	return (zio);
4218 }
4219 
4220 static zio_t *
4221 zio_dva_free(zio_t *zio)
4222 {
4223 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
4224 
4225 	return (zio);
4226 }
4227 
4228 static zio_t *
4229 zio_dva_claim(zio_t *zio)
4230 {
4231 	int error;
4232 
4233 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
4234 	if (error)
4235 		zio->io_error = error;
4236 
4237 	return (zio);
4238 }
4239 
4240 /*
4241  * Undo an allocation.  This is used by zio_done() when an I/O fails
4242  * and we want to give back the block we just allocated.
4243  * This handles both normal blocks and gang blocks.
4244  */
4245 static void
4246 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
4247 {
4248 	ASSERT(BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg || BP_IS_HOLE(bp));
4249 	ASSERT(zio->io_bp_override == NULL);
4250 
4251 	if (!BP_IS_HOLE(bp)) {
4252 		metaslab_free(zio->io_spa, bp, BP_GET_LOGICAL_BIRTH(bp),
4253 		    B_TRUE);
4254 	}
4255 
4256 	if (gn != NULL) {
4257 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
4258 			zio_dva_unallocate(zio, gn->gn_child[g],
4259 			    &gn->gn_gbh->zg_blkptr[g]);
4260 		}
4261 	}
4262 }
4263 
4264 /*
4265  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
4266  */
4267 int
4268 zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
4269     uint64_t size, boolean_t *slog)
4270 {
4271 	int error = 1;
4272 	zio_alloc_list_t io_alloc_list;
4273 
4274 	ASSERT(txg > spa_syncing_txg(spa));
4275 
4276 	metaslab_trace_init(&io_alloc_list);
4277 
4278 	/*
4279 	 * Block pointer fields are useful to metaslabs for stats and debugging.
4280 	 * Fill in the obvious ones before calling into metaslab_alloc().
4281 	 */
4282 	BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
4283 	BP_SET_PSIZE(new_bp, size);
4284 	BP_SET_LEVEL(new_bp, 0);
4285 
4286 	/*
4287 	 * When allocating a zil block, we don't have information about
4288 	 * the final destination of the block except the objset it's part
4289 	 * of, so we just hash the objset ID to pick the allocator to get
4290 	 * some parallelism.
4291 	 */
4292 	int flags = METASLAB_ZIL;
4293 	int allocator = (uint_t)cityhash1(os->os_dsl_dataset->ds_object)
4294 	    % spa->spa_alloc_count;
4295 	ZIOSTAT_BUMP(ziostat_total_allocations);
4296 	error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
4297 	    txg, NULL, flags, &io_alloc_list, NULL, allocator);
4298 	*slog = (error == 0);
4299 	if (error != 0) {
4300 		error = metaslab_alloc(spa, spa_embedded_log_class(spa), size,
4301 		    new_bp, 1, txg, NULL, flags,
4302 		    &io_alloc_list, NULL, allocator);
4303 	}
4304 	if (error != 0) {
4305 		ZIOSTAT_BUMP(ziostat_alloc_class_fallbacks);
4306 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
4307 		    new_bp, 1, txg, NULL, flags,
4308 		    &io_alloc_list, NULL, allocator);
4309 	}
4310 	metaslab_trace_fini(&io_alloc_list);
4311 
4312 	if (error == 0) {
4313 		BP_SET_LSIZE(new_bp, size);
4314 		BP_SET_PSIZE(new_bp, size);
4315 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
4316 		BP_SET_CHECKSUM(new_bp,
4317 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
4318 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
4319 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
4320 		BP_SET_LEVEL(new_bp, 0);
4321 		BP_SET_DEDUP(new_bp, 0);
4322 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
4323 
4324 		/*
4325 		 * encrypted blocks will require an IV and salt. We generate
4326 		 * these now since we will not be rewriting the bp at
4327 		 * rewrite time.
4328 		 */
4329 		if (os->os_encrypted) {
4330 			uint8_t iv[ZIO_DATA_IV_LEN];
4331 			uint8_t salt[ZIO_DATA_SALT_LEN];
4332 
4333 			BP_SET_CRYPT(new_bp, B_TRUE);
4334 			VERIFY0(spa_crypt_get_salt(spa,
4335 			    dmu_objset_id(os), salt));
4336 			VERIFY0(zio_crypt_generate_iv(iv));
4337 
4338 			zio_crypt_encode_params_bp(new_bp, salt, iv);
4339 		}
4340 	} else {
4341 		zfs_dbgmsg("%s: zil block allocation failure: "
4342 		    "size %llu, error %d", spa_name(spa), (u_longlong_t)size,
4343 		    error);
4344 	}
4345 
4346 	return (error);
4347 }
4348 
4349 /*
4350  * ==========================================================================
4351  * Read and write to physical devices
4352  * ==========================================================================
4353  */
4354 
4355 /*
4356  * Issue an I/O to the underlying vdev. Typically the issue pipeline
4357  * stops after this stage and will resume upon I/O completion.
4358  * However, there are instances where the vdev layer may need to
4359  * continue the pipeline when an I/O was not issued. Since the I/O
4360  * that was sent to the vdev layer might be different than the one
4361  * currently active in the pipeline (see vdev_queue_io()), we explicitly
4362  * force the underlying vdev layers to call either zio_execute() or
4363  * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
4364  */
4365 static zio_t *
4366 zio_vdev_io_start(zio_t *zio)
4367 {
4368 	vdev_t *vd = zio->io_vd;
4369 	uint64_t align;
4370 	spa_t *spa = zio->io_spa;
4371 
4372 	zio->io_delay = 0;
4373 
4374 	ASSERT(zio->io_error == 0);
4375 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
4376 
4377 	if (vd == NULL) {
4378 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
4379 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
4380 
4381 		/*
4382 		 * The mirror_ops handle multiple DVAs in a single BP.
4383 		 */
4384 		vdev_mirror_ops.vdev_op_io_start(zio);
4385 		return (NULL);
4386 	}
4387 
4388 	ASSERT3P(zio->io_logical, !=, zio);
4389 	if (zio->io_type == ZIO_TYPE_WRITE) {
4390 		ASSERT(spa->spa_trust_config);
4391 
4392 		/*
4393 		 * Note: the code can handle other kinds of writes,
4394 		 * but we don't expect them.
4395 		 */
4396 		if (zio->io_vd->vdev_noalloc) {
4397 			ASSERT(zio->io_flags &
4398 			    (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
4399 			    ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
4400 		}
4401 	}
4402 
4403 	align = 1ULL << vd->vdev_top->vdev_ashift;
4404 
4405 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
4406 	    P2PHASE(zio->io_size, align) != 0) {
4407 		/* Transform logical writes to be a full physical block size. */
4408 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
4409 		abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
4410 		ASSERT(vd == vd->vdev_top);
4411 		if (zio->io_type == ZIO_TYPE_WRITE) {
4412 			abd_copy(abuf, zio->io_abd, zio->io_size);
4413 			abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
4414 		}
4415 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
4416 	}
4417 
4418 	/*
4419 	 * If this is not a physical io, make sure that it is properly aligned
4420 	 * before proceeding.
4421 	 */
4422 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
4423 		ASSERT0(P2PHASE(zio->io_offset, align));
4424 		ASSERT0(P2PHASE(zio->io_size, align));
4425 	} else {
4426 		/*
4427 		 * For physical writes, we allow 512b aligned writes and assume
4428 		 * the device will perform a read-modify-write as necessary.
4429 		 */
4430 		ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
4431 		ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
4432 	}
4433 
4434 	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
4435 
4436 	/*
4437 	 * If this is a repair I/O, and there's no self-healing involved --
4438 	 * that is, we're just resilvering what we expect to resilver --
4439 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
4440 	 * This prevents spurious resilvering.
4441 	 *
4442 	 * There are a few ways that we can end up creating these spurious
4443 	 * resilver i/os:
4444 	 *
4445 	 * 1. A resilver i/o will be issued if any DVA in the BP has a
4446 	 * dirty DTL.  The mirror code will issue resilver writes to
4447 	 * each DVA, including the one(s) that are not on vdevs with dirty
4448 	 * DTLs.
4449 	 *
4450 	 * 2. With nested replication, which happens when we have a
4451 	 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
4452 	 * For example, given mirror(replacing(A+B), C), it's likely that
4453 	 * only A is out of date (it's the new device). In this case, we'll
4454 	 * read from C, then use the data to resilver A+B -- but we don't
4455 	 * actually want to resilver B, just A. The top-level mirror has no
4456 	 * way to know this, so instead we just discard unnecessary repairs
4457 	 * as we work our way down the vdev tree.
4458 	 *
4459 	 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
4460 	 * The same logic applies to any form of nested replication: ditto
4461 	 * + mirror, RAID-Z + replacing, etc.
4462 	 *
4463 	 * However, indirect vdevs point off to other vdevs which may have
4464 	 * DTL's, so we never bypass them.  The child i/os on concrete vdevs
4465 	 * will be properly bypassed instead.
4466 	 *
4467 	 * Leaf DTL_PARTIAL can be empty when a legitimate write comes from
4468 	 * a dRAID spare vdev. For example, when a dRAID spare is first
4469 	 * used, its spare blocks need to be written to but the leaf vdev's
4470 	 * of such blocks can have empty DTL_PARTIAL.
4471 	 *
4472 	 * There seemed no clean way to allow such writes while bypassing
4473 	 * spurious ones. At this point, just avoid all bypassing for dRAID
4474 	 * for correctness.
4475 	 */
4476 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
4477 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
4478 	    zio->io_txg != 0 &&	/* not a delegated i/o */
4479 	    vd->vdev_ops != &vdev_indirect_ops &&
4480 	    vd->vdev_top->vdev_ops != &vdev_draid_ops &&
4481 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
4482 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4483 		zio_vdev_io_bypass(zio);
4484 		return (zio);
4485 	}
4486 
4487 	/*
4488 	 * Select the next best leaf I/O to process.  Distributed spares are
4489 	 * excluded since they dispatch the I/O directly to a leaf vdev after
4490 	 * applying the dRAID mapping.
4491 	 */
4492 	if (vd->vdev_ops->vdev_op_leaf &&
4493 	    vd->vdev_ops != &vdev_draid_spare_ops &&
4494 	    (zio->io_type == ZIO_TYPE_READ ||
4495 	    zio->io_type == ZIO_TYPE_WRITE ||
4496 	    zio->io_type == ZIO_TYPE_TRIM)) {
4497 
4498 		if ((zio = vdev_queue_io(zio)) == NULL)
4499 			return (NULL);
4500 
4501 		if (!vdev_accessible(vd, zio)) {
4502 			zio->io_error = SET_ERROR(ENXIO);
4503 			zio_interrupt(zio);
4504 			return (NULL);
4505 		}
4506 		zio->io_delay = gethrtime();
4507 
4508 		if (zio_handle_device_injection(vd, zio, ENOSYS) != 0) {
4509 			/*
4510 			 * "no-op" injections return success, but do no actual
4511 			 * work. Just return it.
4512 			 */
4513 			zio_delay_interrupt(zio);
4514 			return (NULL);
4515 		}
4516 	}
4517 
4518 	vd->vdev_ops->vdev_op_io_start(zio);
4519 	return (NULL);
4520 }
4521 
4522 static zio_t *
4523 zio_vdev_io_done(zio_t *zio)
4524 {
4525 	vdev_t *vd = zio->io_vd;
4526 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
4527 	boolean_t unexpected_error = B_FALSE;
4528 
4529 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
4530 		return (NULL);
4531 	}
4532 
4533 	ASSERT(zio->io_type == ZIO_TYPE_READ ||
4534 	    zio->io_type == ZIO_TYPE_WRITE ||
4535 	    zio->io_type == ZIO_TYPE_FLUSH ||
4536 	    zio->io_type == ZIO_TYPE_TRIM);
4537 
4538 	if (zio->io_delay)
4539 		zio->io_delay = gethrtime() - zio->io_delay;
4540 
4541 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
4542 	    vd->vdev_ops != &vdev_draid_spare_ops) {
4543 		if (zio->io_type != ZIO_TYPE_FLUSH)
4544 			vdev_queue_io_done(zio);
4545 
4546 		if (zio_injection_enabled && zio->io_error == 0)
4547 			zio->io_error = zio_handle_device_injections(vd, zio,
4548 			    EIO, EILSEQ);
4549 
4550 		if (zio_injection_enabled && zio->io_error == 0)
4551 			zio->io_error = zio_handle_label_injection(zio, EIO);
4552 
4553 		if (zio->io_error && zio->io_type != ZIO_TYPE_FLUSH &&
4554 		    zio->io_type != ZIO_TYPE_TRIM) {
4555 			if (!vdev_accessible(vd, zio)) {
4556 				zio->io_error = SET_ERROR(ENXIO);
4557 			} else {
4558 				unexpected_error = B_TRUE;
4559 			}
4560 		}
4561 	}
4562 
4563 	ops->vdev_op_io_done(zio);
4564 
4565 	if (unexpected_error && vd->vdev_remove_wanted == B_FALSE)
4566 		VERIFY(vdev_probe(vd, zio) == NULL);
4567 
4568 	return (zio);
4569 }
4570 
4571 /*
4572  * This function is used to change the priority of an existing zio that is
4573  * currently in-flight. This is used by the arc to upgrade priority in the
4574  * event that a demand read is made for a block that is currently queued
4575  * as a scrub or async read IO. Otherwise, the high priority read request
4576  * would end up having to wait for the lower priority IO.
4577  */
4578 void
4579 zio_change_priority(zio_t *pio, zio_priority_t priority)
4580 {
4581 	zio_t *cio, *cio_next;
4582 	zio_link_t *zl = NULL;
4583 
4584 	ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
4585 
4586 	if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
4587 		vdev_queue_change_io_priority(pio, priority);
4588 	} else {
4589 		pio->io_priority = priority;
4590 	}
4591 
4592 	mutex_enter(&pio->io_lock);
4593 	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
4594 		cio_next = zio_walk_children(pio, &zl);
4595 		zio_change_priority(cio, priority);
4596 	}
4597 	mutex_exit(&pio->io_lock);
4598 }
4599 
4600 /*
4601  * For non-raidz ZIOs, we can just copy aside the bad data read from the
4602  * disk, and use that to finish the checksum ereport later.
4603  */
4604 static void
4605 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
4606     const abd_t *good_buf)
4607 {
4608 	/* no processing needed */
4609 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
4610 }
4611 
4612 void
4613 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr)
4614 {
4615 	void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size);
4616 
4617 	abd_copy(abd, zio->io_abd, zio->io_size);
4618 
4619 	zcr->zcr_cbinfo = zio->io_size;
4620 	zcr->zcr_cbdata = abd;
4621 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
4622 	zcr->zcr_free = zio_abd_free;
4623 }
4624 
4625 static zio_t *
4626 zio_vdev_io_assess(zio_t *zio)
4627 {
4628 	vdev_t *vd = zio->io_vd;
4629 
4630 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
4631 		return (NULL);
4632 	}
4633 
4634 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
4635 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
4636 
4637 	if (zio->io_vsd != NULL) {
4638 		zio->io_vsd_ops->vsd_free(zio);
4639 		zio->io_vsd = NULL;
4640 	}
4641 
4642 	/*
4643 	 * If a Direct I/O operation has a checksum verify error then this I/O
4644 	 * should not attempt to be issued again.
4645 	 */
4646 	if (zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR) {
4647 		if (zio->io_type == ZIO_TYPE_WRITE) {
4648 			ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_LOGICAL);
4649 			ASSERT3U(zio->io_error, ==, EIO);
4650 		}
4651 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4652 		return (zio);
4653 	}
4654 
4655 	if (zio_injection_enabled && zio->io_error == 0)
4656 		zio->io_error = zio_handle_fault_injection(zio, EIO);
4657 
4658 	/*
4659 	 * If the I/O failed, determine whether we should attempt to retry it.
4660 	 *
4661 	 * On retry, we cut in line in the issue queue, since we don't want
4662 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
4663 	 */
4664 	if (zio->io_error && vd == NULL &&
4665 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
4666 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
4667 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
4668 		zio->io_error = 0;
4669 		zio->io_flags |= ZIO_FLAG_IO_RETRY | ZIO_FLAG_DONT_AGGREGATE;
4670 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
4671 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
4672 		    zio_requeue_io_start_cut_in_line);
4673 		return (NULL);
4674 	}
4675 
4676 	/*
4677 	 * If we got an error on a leaf device, convert it to ENXIO
4678 	 * if the device is not accessible at all.
4679 	 */
4680 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
4681 	    !vdev_accessible(vd, zio))
4682 		zio->io_error = SET_ERROR(ENXIO);
4683 
4684 	/*
4685 	 * If we can't write to an interior vdev (mirror or RAID-Z),
4686 	 * set vdev_cant_write so that we stop trying to allocate from it.
4687 	 */
4688 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
4689 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
4690 		vdev_dbgmsg(vd, "zio_vdev_io_assess(zio=%px) setting "
4691 		    "cant_write=TRUE due to write failure with ENXIO",
4692 		    zio);
4693 		vd->vdev_cant_write = B_TRUE;
4694 	}
4695 
4696 	/*
4697 	 * If a cache flush returns ENOTSUP we know that no future
4698 	 * attempts will ever succeed. In this case we set a persistent
4699 	 * boolean flag so that we don't bother with it in the future, and
4700 	 * then we act like the flush succeeded.
4701 	 */
4702 	if (zio->io_error == ENOTSUP && zio->io_type == ZIO_TYPE_FLUSH &&
4703 	    vd != NULL) {
4704 		vd->vdev_nowritecache = B_TRUE;
4705 		zio->io_error = 0;
4706 	}
4707 
4708 	if (zio->io_error)
4709 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4710 
4711 	return (zio);
4712 }
4713 
4714 void
4715 zio_vdev_io_reissue(zio_t *zio)
4716 {
4717 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4718 	ASSERT(zio->io_error == 0);
4719 
4720 	zio->io_stage >>= 1;
4721 }
4722 
4723 void
4724 zio_vdev_io_redone(zio_t *zio)
4725 {
4726 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
4727 
4728 	zio->io_stage >>= 1;
4729 }
4730 
4731 void
4732 zio_vdev_io_bypass(zio_t *zio)
4733 {
4734 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4735 	ASSERT(zio->io_error == 0);
4736 
4737 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
4738 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
4739 }
4740 
4741 /*
4742  * ==========================================================================
4743  * Encrypt and store encryption parameters
4744  * ==========================================================================
4745  */
4746 
4747 
4748 /*
4749  * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
4750  * managing the storage of encryption parameters and passing them to the
4751  * lower-level encryption functions.
4752  */
4753 static zio_t *
4754 zio_encrypt(zio_t *zio)
4755 {
4756 	zio_prop_t *zp = &zio->io_prop;
4757 	spa_t *spa = zio->io_spa;
4758 	blkptr_t *bp = zio->io_bp;
4759 	uint64_t psize = BP_GET_PSIZE(bp);
4760 	uint64_t dsobj = zio->io_bookmark.zb_objset;
4761 	dmu_object_type_t ot = BP_GET_TYPE(bp);
4762 	void *enc_buf = NULL;
4763 	abd_t *eabd = NULL;
4764 	uint8_t salt[ZIO_DATA_SALT_LEN];
4765 	uint8_t iv[ZIO_DATA_IV_LEN];
4766 	uint8_t mac[ZIO_DATA_MAC_LEN];
4767 	boolean_t no_crypt = B_FALSE;
4768 
4769 	/* the root zio already encrypted the data */
4770 	if (zio->io_child_type == ZIO_CHILD_GANG)
4771 		return (zio);
4772 
4773 	/* only ZIL blocks are re-encrypted on rewrite */
4774 	if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG)
4775 		return (zio);
4776 
4777 	if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) {
4778 		BP_SET_CRYPT(bp, B_FALSE);
4779 		return (zio);
4780 	}
4781 
4782 	/* if we are doing raw encryption set the provided encryption params */
4783 	if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) {
4784 		ASSERT0(BP_GET_LEVEL(bp));
4785 		BP_SET_CRYPT(bp, B_TRUE);
4786 		BP_SET_BYTEORDER(bp, zp->zp_byteorder);
4787 		if (ot != DMU_OT_OBJSET)
4788 			zio_crypt_encode_mac_bp(bp, zp->zp_mac);
4789 
4790 		/* dnode blocks must be written out in the provided byteorder */
4791 		if (zp->zp_byteorder != ZFS_HOST_BYTEORDER &&
4792 		    ot == DMU_OT_DNODE) {
4793 			void *bswap_buf = zio_buf_alloc(psize);
4794 			abd_t *babd = abd_get_from_buf(bswap_buf, psize);
4795 
4796 			ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4797 			abd_copy_to_buf(bswap_buf, zio->io_abd, psize);
4798 			dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf,
4799 			    psize);
4800 
4801 			abd_take_ownership_of_buf(babd, B_TRUE);
4802 			zio_push_transform(zio, babd, psize, psize, NULL);
4803 		}
4804 
4805 		if (DMU_OT_IS_ENCRYPTED(ot))
4806 			zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv);
4807 		return (zio);
4808 	}
4809 
4810 	/* indirect blocks only maintain a cksum of the lower level MACs */
4811 	if (BP_GET_LEVEL(bp) > 0) {
4812 		BP_SET_CRYPT(bp, B_TRUE);
4813 		VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE,
4814 		    zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp),
4815 		    mac));
4816 		zio_crypt_encode_mac_bp(bp, mac);
4817 		return (zio);
4818 	}
4819 
4820 	/*
4821 	 * Objset blocks are a special case since they have 2 256-bit MACs
4822 	 * embedded within them.
4823 	 */
4824 	if (ot == DMU_OT_OBJSET) {
4825 		ASSERT0(DMU_OT_IS_ENCRYPTED(ot));
4826 		ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4827 		BP_SET_CRYPT(bp, B_TRUE);
4828 		VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj,
4829 		    zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp)));
4830 		return (zio);
4831 	}
4832 
4833 	/* unencrypted object types are only authenticated with a MAC */
4834 	if (!DMU_OT_IS_ENCRYPTED(ot)) {
4835 		BP_SET_CRYPT(bp, B_TRUE);
4836 		VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj,
4837 		    zio->io_abd, psize, mac));
4838 		zio_crypt_encode_mac_bp(bp, mac);
4839 		return (zio);
4840 	}
4841 
4842 	/*
4843 	 * Later passes of sync-to-convergence may decide to rewrite data
4844 	 * in place to avoid more disk reallocations. This presents a problem
4845 	 * for encryption because this constitutes rewriting the new data with
4846 	 * the same encryption key and IV. However, this only applies to blocks
4847 	 * in the MOS (particularly the spacemaps) and we do not encrypt the
4848 	 * MOS. We assert that the zio is allocating or an intent log write
4849 	 * to enforce this.
4850 	 */
4851 	ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG);
4852 	ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG);
4853 	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION));
4854 	ASSERT3U(psize, !=, 0);
4855 
4856 	enc_buf = zio_buf_alloc(psize);
4857 	eabd = abd_get_from_buf(enc_buf, psize);
4858 	abd_take_ownership_of_buf(eabd, B_TRUE);
4859 
4860 	/*
4861 	 * For an explanation of what encryption parameters are stored
4862 	 * where, see the block comment in zio_crypt.c.
4863 	 */
4864 	if (ot == DMU_OT_INTENT_LOG) {
4865 		zio_crypt_decode_params_bp(bp, salt, iv);
4866 	} else {
4867 		BP_SET_CRYPT(bp, B_TRUE);
4868 	}
4869 
4870 	/* Perform the encryption. This should not fail */
4871 	VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark,
4872 	    BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
4873 	    salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt));
4874 
4875 	/* encode encryption metadata into the bp */
4876 	if (ot == DMU_OT_INTENT_LOG) {
4877 		/*
4878 		 * ZIL blocks store the MAC in the embedded checksum, so the
4879 		 * transform must always be applied.
4880 		 */
4881 		zio_crypt_encode_mac_zil(enc_buf, mac);
4882 		zio_push_transform(zio, eabd, psize, psize, NULL);
4883 	} else {
4884 		BP_SET_CRYPT(bp, B_TRUE);
4885 		zio_crypt_encode_params_bp(bp, salt, iv);
4886 		zio_crypt_encode_mac_bp(bp, mac);
4887 
4888 		if (no_crypt) {
4889 			ASSERT3U(ot, ==, DMU_OT_DNODE);
4890 			abd_free(eabd);
4891 		} else {
4892 			zio_push_transform(zio, eabd, psize, psize, NULL);
4893 		}
4894 	}
4895 
4896 	return (zio);
4897 }
4898 
4899 /*
4900  * ==========================================================================
4901  * Generate and verify checksums
4902  * ==========================================================================
4903  */
4904 static zio_t *
4905 zio_checksum_generate(zio_t *zio)
4906 {
4907 	blkptr_t *bp = zio->io_bp;
4908 	enum zio_checksum checksum;
4909 
4910 	if (bp == NULL) {
4911 		/*
4912 		 * This is zio_write_phys().
4913 		 * We're either generating a label checksum, or none at all.
4914 		 */
4915 		checksum = zio->io_prop.zp_checksum;
4916 
4917 		if (checksum == ZIO_CHECKSUM_OFF)
4918 			return (zio);
4919 
4920 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
4921 	} else {
4922 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
4923 			ASSERT(!IO_IS_ALLOCATING(zio));
4924 			checksum = ZIO_CHECKSUM_GANG_HEADER;
4925 		} else {
4926 			checksum = BP_GET_CHECKSUM(bp);
4927 		}
4928 	}
4929 
4930 	zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
4931 
4932 	return (zio);
4933 }
4934 
4935 static zio_t *
4936 zio_checksum_verify(zio_t *zio)
4937 {
4938 	zio_bad_cksum_t info;
4939 	blkptr_t *bp = zio->io_bp;
4940 	int error;
4941 
4942 	ASSERT(zio->io_vd != NULL);
4943 
4944 	if (bp == NULL) {
4945 		/*
4946 		 * This is zio_read_phys().
4947 		 * We're either verifying a label checksum, or nothing at all.
4948 		 */
4949 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
4950 			return (zio);
4951 
4952 		ASSERT3U(zio->io_prop.zp_checksum, ==, ZIO_CHECKSUM_LABEL);
4953 	}
4954 
4955 	ASSERT0(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR);
4956 	IMPLY(zio->io_flags & ZIO_FLAG_DIO_READ,
4957 	    !(zio->io_flags & ZIO_FLAG_SPECULATIVE));
4958 
4959 	if ((error = zio_checksum_error(zio, &info)) != 0) {
4960 		zio->io_error = error;
4961 		if (error == ECKSUM &&
4962 		    !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
4963 			if (zio->io_flags & ZIO_FLAG_DIO_READ) {
4964 				zio->io_flags |= ZIO_FLAG_DIO_CHKSUM_ERR;
4965 				zio_t *pio = zio_unique_parent(zio);
4966 				/*
4967 				 * Any Direct I/O read that has a checksum
4968 				 * error must be treated as suspicous as the
4969 				 * contents of the buffer could be getting
4970 				 * manipulated while the I/O is taking place.
4971 				 *
4972 				 * The checksum verify error will only be
4973 				 * reported here for disk and file VDEV's and
4974 				 * will be reported on those that the failure
4975 				 * occurred on. Other types of VDEV's report the
4976 				 * verify failure in their own code paths.
4977 				 */
4978 				if (pio->io_child_type == ZIO_CHILD_LOGICAL) {
4979 					zio_dio_chksum_verify_error_report(zio);
4980 				}
4981 			} else {
4982 				mutex_enter(&zio->io_vd->vdev_stat_lock);
4983 				zio->io_vd->vdev_stat.vs_checksum_errors++;
4984 				mutex_exit(&zio->io_vd->vdev_stat_lock);
4985 				(void) zfs_ereport_start_checksum(zio->io_spa,
4986 				    zio->io_vd, &zio->io_bookmark, zio,
4987 				    zio->io_offset, zio->io_size, &info);
4988 			}
4989 		}
4990 	}
4991 
4992 	return (zio);
4993 }
4994 
4995 static zio_t *
4996 zio_dio_checksum_verify(zio_t *zio)
4997 {
4998 	zio_t *pio = zio_unique_parent(zio);
4999 	int error;
5000 
5001 	ASSERT3P(zio->io_vd, !=, NULL);
5002 	ASSERT3P(zio->io_bp, !=, NULL);
5003 	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
5004 	ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
5005 	ASSERT3B(pio->io_prop.zp_direct_write, ==, B_TRUE);
5006 	ASSERT3U(pio->io_child_type, ==, ZIO_CHILD_LOGICAL);
5007 
5008 	if (zfs_vdev_direct_write_verify == 0 || zio->io_error != 0)
5009 		goto out;
5010 
5011 	if ((error = zio_checksum_error(zio, NULL)) != 0) {
5012 		zio->io_error = error;
5013 		if (error == ECKSUM) {
5014 			zio->io_flags |= ZIO_FLAG_DIO_CHKSUM_ERR;
5015 			zio_dio_chksum_verify_error_report(zio);
5016 		}
5017 	}
5018 
5019 out:
5020 	return (zio);
5021 }
5022 
5023 
5024 /*
5025  * Called by RAID-Z to ensure we don't compute the checksum twice.
5026  */
5027 void
5028 zio_checksum_verified(zio_t *zio)
5029 {
5030 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
5031 }
5032 
5033 /*
5034  * Report Direct I/O checksum verify error and create ZED event.
5035  */
5036 void
5037 zio_dio_chksum_verify_error_report(zio_t *zio)
5038 {
5039 	ASSERT(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR);
5040 
5041 	if (zio->io_child_type == ZIO_CHILD_LOGICAL)
5042 		return;
5043 
5044 	mutex_enter(&zio->io_vd->vdev_stat_lock);
5045 	zio->io_vd->vdev_stat.vs_dio_verify_errors++;
5046 	mutex_exit(&zio->io_vd->vdev_stat_lock);
5047 	if (zio->io_type == ZIO_TYPE_WRITE) {
5048 		/*
5049 		 * Convert checksum error for writes into EIO.
5050 		 */
5051 		zio->io_error = SET_ERROR(EIO);
5052 		/*
5053 		 * Report dio_verify_wr ZED event.
5054 		 */
5055 		(void) zfs_ereport_post(FM_EREPORT_ZFS_DIO_VERIFY_WR,
5056 		    zio->io_spa,  zio->io_vd, &zio->io_bookmark, zio, 0);
5057 	} else {
5058 		/*
5059 		 * Report dio_verify_rd ZED event.
5060 		 */
5061 		(void) zfs_ereport_post(FM_EREPORT_ZFS_DIO_VERIFY_RD,
5062 		    zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0);
5063 	}
5064 }
5065 
5066 /*
5067  * ==========================================================================
5068  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
5069  * An error of 0 indicates success.  ENXIO indicates whole-device failure,
5070  * which may be transient (e.g. unplugged) or permanent.  ECKSUM and EIO
5071  * indicate errors that are specific to one I/O, and most likely permanent.
5072  * Any other error is presumed to be worse because we weren't expecting it.
5073  * ==========================================================================
5074  */
5075 int
5076 zio_worst_error(int e1, int e2)
5077 {
5078 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
5079 	int r1, r2;
5080 
5081 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
5082 		if (e1 == zio_error_rank[r1])
5083 			break;
5084 
5085 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
5086 		if (e2 == zio_error_rank[r2])
5087 			break;
5088 
5089 	return (r1 > r2 ? e1 : e2);
5090 }
5091 
5092 /*
5093  * ==========================================================================
5094  * I/O completion
5095  * ==========================================================================
5096  */
5097 static zio_t *
5098 zio_ready(zio_t *zio)
5099 {
5100 	blkptr_t *bp = zio->io_bp;
5101 	zio_t *pio, *pio_next;
5102 	zio_link_t *zl = NULL;
5103 
5104 	if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
5105 	    ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT, ZIO_WAIT_READY)) {
5106 		return (NULL);
5107 	}
5108 
5109 	if (zio->io_ready) {
5110 		ASSERT(IO_IS_ALLOCATING(zio));
5111 		ASSERT(BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg ||
5112 		    BP_IS_HOLE(bp) || (zio->io_flags & ZIO_FLAG_NOPWRITE));
5113 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
5114 
5115 		zio->io_ready(zio);
5116 	}
5117 
5118 #ifdef ZFS_DEBUG
5119 	if (bp != NULL && bp != &zio->io_bp_copy)
5120 		zio->io_bp_copy = *bp;
5121 #endif
5122 
5123 	if (zio->io_error != 0) {
5124 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
5125 
5126 		if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
5127 			ASSERT(IO_IS_ALLOCATING(zio));
5128 			ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
5129 			ASSERT(zio->io_metaslab_class != NULL);
5130 			ASSERT(ZIO_HAS_ALLOCATOR(zio));
5131 
5132 			/*
5133 			 * We were unable to allocate anything, unreserve and
5134 			 * issue the next I/O to allocate.
5135 			 */
5136 			metaslab_class_throttle_unreserve(
5137 			    zio->io_metaslab_class, zio->io_prop.zp_copies,
5138 			    zio->io_allocator, zio);
5139 			zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
5140 		}
5141 	}
5142 
5143 	mutex_enter(&zio->io_lock);
5144 	zio->io_state[ZIO_WAIT_READY] = 1;
5145 	pio = zio_walk_parents(zio, &zl);
5146 	mutex_exit(&zio->io_lock);
5147 
5148 	/*
5149 	 * As we notify zio's parents, new parents could be added.
5150 	 * New parents go to the head of zio's io_parent_list, however,
5151 	 * so we will (correctly) not notify them.  The remainder of zio's
5152 	 * io_parent_list, from 'pio_next' onward, cannot change because
5153 	 * all parents must wait for us to be done before they can be done.
5154 	 */
5155 	for (; pio != NULL; pio = pio_next) {
5156 		pio_next = zio_walk_parents(zio, &zl);
5157 		zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL);
5158 	}
5159 
5160 	if (zio->io_flags & ZIO_FLAG_NODATA) {
5161 		if (bp != NULL && BP_IS_GANG(bp)) {
5162 			zio->io_flags &= ~ZIO_FLAG_NODATA;
5163 		} else {
5164 			ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
5165 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
5166 		}
5167 	}
5168 
5169 	if (zio_injection_enabled &&
5170 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
5171 		zio_handle_ignored_writes(zio);
5172 
5173 	return (zio);
5174 }
5175 
5176 /*
5177  * Update the allocation throttle accounting.
5178  */
5179 static void
5180 zio_dva_throttle_done(zio_t *zio)
5181 {
5182 	zio_t *lio __maybe_unused = zio->io_logical;
5183 	zio_t *pio = zio_unique_parent(zio);
5184 	vdev_t *vd = zio->io_vd;
5185 	int flags = METASLAB_ASYNC_ALLOC;
5186 
5187 	ASSERT3P(zio->io_bp, !=, NULL);
5188 	ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
5189 	ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
5190 	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
5191 	ASSERT(vd != NULL);
5192 	ASSERT3P(vd, ==, vd->vdev_top);
5193 	ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY));
5194 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
5195 	ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
5196 	ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
5197 	ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
5198 
5199 	/*
5200 	 * Parents of gang children can have two flavors -- ones that
5201 	 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
5202 	 * and ones that allocated the constituent blocks. The allocation
5203 	 * throttle needs to know the allocating parent zio so we must find
5204 	 * it here.
5205 	 */
5206 	if (pio->io_child_type == ZIO_CHILD_GANG) {
5207 		/*
5208 		 * If our parent is a rewrite gang child then our grandparent
5209 		 * would have been the one that performed the allocation.
5210 		 */
5211 		if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
5212 			pio = zio_unique_parent(pio);
5213 		flags |= METASLAB_GANG_CHILD;
5214 	}
5215 
5216 	ASSERT(IO_IS_ALLOCATING(pio));
5217 	ASSERT(ZIO_HAS_ALLOCATOR(pio));
5218 	ASSERT3P(zio, !=, zio->io_logical);
5219 	ASSERT(zio->io_logical != NULL);
5220 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
5221 	ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
5222 	ASSERT(zio->io_metaslab_class != NULL);
5223 
5224 	mutex_enter(&pio->io_lock);
5225 	metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
5226 	    pio->io_allocator, B_TRUE);
5227 	mutex_exit(&pio->io_lock);
5228 
5229 	metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1,
5230 	    pio->io_allocator, pio);
5231 
5232 	/*
5233 	 * Call into the pipeline to see if there is more work that
5234 	 * needs to be done. If there is work to be done it will be
5235 	 * dispatched to another taskq thread.
5236 	 */
5237 	zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
5238 }
5239 
5240 static zio_t *
5241 zio_done(zio_t *zio)
5242 {
5243 	/*
5244 	 * Always attempt to keep stack usage minimal here since
5245 	 * we can be called recursively up to 19 levels deep.
5246 	 */
5247 	const uint64_t psize = zio->io_size;
5248 	zio_t *pio, *pio_next;
5249 	zio_link_t *zl = NULL;
5250 
5251 	/*
5252 	 * If our children haven't all completed,
5253 	 * wait for them and then repeat this pipeline stage.
5254 	 */
5255 	if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
5256 		return (NULL);
5257 	}
5258 
5259 	/*
5260 	 * If the allocation throttle is enabled, then update the accounting.
5261 	 * We only track child I/Os that are part of an allocating async
5262 	 * write. We must do this since the allocation is performed
5263 	 * by the logical I/O but the actual write is done by child I/Os.
5264 	 */
5265 	if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
5266 	    zio->io_child_type == ZIO_CHILD_VDEV) {
5267 		ASSERT(zio->io_metaslab_class != NULL);
5268 		ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled);
5269 		zio_dva_throttle_done(zio);
5270 	}
5271 
5272 	/*
5273 	 * If the allocation throttle is enabled, verify that
5274 	 * we have decremented the refcounts for every I/O that was throttled.
5275 	 */
5276 	if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
5277 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
5278 		ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
5279 		ASSERT(zio->io_bp != NULL);
5280 		ASSERT(ZIO_HAS_ALLOCATOR(zio));
5281 
5282 		metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio,
5283 		    zio->io_allocator);
5284 		VERIFY(zfs_refcount_not_held(&zio->io_metaslab_class->
5285 		    mc_allocator[zio->io_allocator].mca_alloc_slots, zio));
5286 	}
5287 
5288 
5289 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
5290 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
5291 			ASSERT(zio->io_children[c][w] == 0);
5292 
5293 	if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) {
5294 		ASSERT(zio->io_bp->blk_pad[0] == 0);
5295 		ASSERT(zio->io_bp->blk_pad[1] == 0);
5296 		ASSERT(memcmp(zio->io_bp, &zio->io_bp_copy,
5297 		    sizeof (blkptr_t)) == 0 ||
5298 		    (zio->io_bp == zio_unique_parent(zio)->io_bp));
5299 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
5300 		    zio->io_bp_override == NULL &&
5301 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
5302 			ASSERT3U(zio->io_prop.zp_copies, <=,
5303 			    BP_GET_NDVAS(zio->io_bp));
5304 			ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
5305 			    (BP_COUNT_GANG(zio->io_bp) ==
5306 			    BP_GET_NDVAS(zio->io_bp)));
5307 		}
5308 		if (zio->io_flags & ZIO_FLAG_NOPWRITE)
5309 			VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig));
5310 	}
5311 
5312 	/*
5313 	 * If there were child vdev/gang/ddt errors, they apply to us now.
5314 	 */
5315 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
5316 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
5317 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
5318 
5319 	/*
5320 	 * If the I/O on the transformed data was successful, generate any
5321 	 * checksum reports now while we still have the transformed data.
5322 	 */
5323 	if (zio->io_error == 0) {
5324 		while (zio->io_cksum_report != NULL) {
5325 			zio_cksum_report_t *zcr = zio->io_cksum_report;
5326 			uint64_t align = zcr->zcr_align;
5327 			uint64_t asize = P2ROUNDUP(psize, align);
5328 			abd_t *adata = zio->io_abd;
5329 
5330 			if (adata != NULL && asize != psize) {
5331 				adata = abd_alloc(asize, B_TRUE);
5332 				abd_copy(adata, zio->io_abd, psize);
5333 				abd_zero_off(adata, psize, asize - psize);
5334 			}
5335 
5336 			zio->io_cksum_report = zcr->zcr_next;
5337 			zcr->zcr_next = NULL;
5338 			zcr->zcr_finish(zcr, adata);
5339 			zfs_ereport_free_checksum(zcr);
5340 
5341 			if (adata != NULL && asize != psize)
5342 				abd_free(adata);
5343 		}
5344 	}
5345 
5346 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
5347 
5348 	vdev_stat_update(zio, psize);
5349 
5350 	/*
5351 	 * If this I/O is attached to a particular vdev is slow, exceeding
5352 	 * 30 seconds to complete, post an error described the I/O delay.
5353 	 * We ignore these errors if the device is currently unavailable.
5354 	 */
5355 	if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) {
5356 		if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) {
5357 			/*
5358 			 * We want to only increment our slow IO counters if
5359 			 * the IO is valid (i.e. not if the drive is removed).
5360 			 *
5361 			 * zfs_ereport_post() will also do these checks, but
5362 			 * it can also ratelimit and have other failures, so we
5363 			 * need to increment the slow_io counters independent
5364 			 * of it.
5365 			 */
5366 			if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY,
5367 			    zio->io_spa, zio->io_vd, zio)) {
5368 				mutex_enter(&zio->io_vd->vdev_stat_lock);
5369 				zio->io_vd->vdev_stat.vs_slow_ios++;
5370 				mutex_exit(&zio->io_vd->vdev_stat_lock);
5371 
5372 				(void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY,
5373 				    zio->io_spa, zio->io_vd, &zio->io_bookmark,
5374 				    zio, 0);
5375 			}
5376 		}
5377 	}
5378 
5379 	if (zio->io_error) {
5380 		/*
5381 		 * If this I/O is attached to a particular vdev,
5382 		 * generate an error message describing the I/O failure
5383 		 * at the block level.  We ignore these errors if the
5384 		 * device is currently unavailable.
5385 		 */
5386 		if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
5387 		    !vdev_is_dead(zio->io_vd) &&
5388 		    !(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR)) {
5389 			int ret = zfs_ereport_post(FM_EREPORT_ZFS_IO,
5390 			    zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0);
5391 			if (ret != EALREADY) {
5392 				mutex_enter(&zio->io_vd->vdev_stat_lock);
5393 				if (zio->io_type == ZIO_TYPE_READ)
5394 					zio->io_vd->vdev_stat.vs_read_errors++;
5395 				else if (zio->io_type == ZIO_TYPE_WRITE)
5396 					zio->io_vd->vdev_stat.vs_write_errors++;
5397 				mutex_exit(&zio->io_vd->vdev_stat_lock);
5398 			}
5399 		}
5400 
5401 		if ((zio->io_error == EIO || !(zio->io_flags &
5402 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
5403 		    !(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR) &&
5404 		    zio == zio->io_logical) {
5405 			/*
5406 			 * For logical I/O requests, tell the SPA to log the
5407 			 * error and generate a logical data ereport.
5408 			 */
5409 			spa_log_error(zio->io_spa, &zio->io_bookmark,
5410 			    BP_GET_LOGICAL_BIRTH(zio->io_bp));
5411 			(void) zfs_ereport_post(FM_EREPORT_ZFS_DATA,
5412 			    zio->io_spa, NULL, &zio->io_bookmark, zio, 0);
5413 		}
5414 	}
5415 
5416 	if (zio->io_error && zio == zio->io_logical) {
5417 		/*
5418 		 * Determine whether zio should be reexecuted.  This will
5419 		 * propagate all the way to the root via zio_notify_parent().
5420 		 */
5421 		ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
5422 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
5423 
5424 		if (IO_IS_ALLOCATING(zio) &&
5425 		    !(zio->io_flags & ZIO_FLAG_CANFAIL) &&
5426 		    !(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR)) {
5427 			if (zio->io_error != ENOSPC)
5428 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
5429 			else
5430 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
5431 		}
5432 
5433 		if ((zio->io_type == ZIO_TYPE_READ ||
5434 		    zio->io_type == ZIO_TYPE_FREE) &&
5435 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
5436 		    zio->io_error == ENXIO &&
5437 		    spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
5438 		    spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
5439 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
5440 
5441 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
5442 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
5443 
5444 		/*
5445 		 * Here is a possibly good place to attempt to do
5446 		 * either combinatorial reconstruction or error correction
5447 		 * based on checksums.  It also might be a good place
5448 		 * to send out preliminary ereports before we suspend
5449 		 * processing.
5450 		 */
5451 	}
5452 
5453 	/*
5454 	 * If there were logical child errors, they apply to us now.
5455 	 * We defer this until now to avoid conflating logical child
5456 	 * errors with errors that happened to the zio itself when
5457 	 * updating vdev stats and reporting FMA events above.
5458 	 */
5459 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
5460 
5461 	if ((zio->io_error || zio->io_reexecute) &&
5462 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
5463 	    !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
5464 		zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
5465 
5466 	zio_gang_tree_free(&zio->io_gang_tree);
5467 
5468 	/*
5469 	 * Godfather I/Os should never suspend.
5470 	 */
5471 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
5472 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
5473 		zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND;
5474 
5475 	if (zio->io_reexecute) {
5476 		/*
5477 		 * A Direct I/O operation that has a checksum verify error
5478 		 * should not attempt to reexecute. Instead, the error should
5479 		 * just be propagated back.
5480 		 */
5481 		ASSERT(!(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR));
5482 
5483 		/*
5484 		 * This is a logical I/O that wants to reexecute.
5485 		 *
5486 		 * Reexecute is top-down.  When an i/o fails, if it's not
5487 		 * the root, it simply notifies its parent and sticks around.
5488 		 * The parent, seeing that it still has children in zio_done(),
5489 		 * does the same.  This percolates all the way up to the root.
5490 		 * The root i/o will reexecute or suspend the entire tree.
5491 		 *
5492 		 * This approach ensures that zio_reexecute() honors
5493 		 * all the original i/o dependency relationships, e.g.
5494 		 * parents not executing until children are ready.
5495 		 */
5496 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
5497 
5498 		zio->io_gang_leader = NULL;
5499 
5500 		mutex_enter(&zio->io_lock);
5501 		zio->io_state[ZIO_WAIT_DONE] = 1;
5502 		mutex_exit(&zio->io_lock);
5503 
5504 		/*
5505 		 * "The Godfather" I/O monitors its children but is
5506 		 * not a true parent to them. It will track them through
5507 		 * the pipeline but severs its ties whenever they get into
5508 		 * trouble (e.g. suspended). This allows "The Godfather"
5509 		 * I/O to return status without blocking.
5510 		 */
5511 		zl = NULL;
5512 		for (pio = zio_walk_parents(zio, &zl); pio != NULL;
5513 		    pio = pio_next) {
5514 			zio_link_t *remove_zl = zl;
5515 			pio_next = zio_walk_parents(zio, &zl);
5516 
5517 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
5518 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
5519 				zio_remove_child(pio, zio, remove_zl);
5520 				/*
5521 				 * This is a rare code path, so we don't
5522 				 * bother with "next_to_execute".
5523 				 */
5524 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE,
5525 				    NULL);
5526 			}
5527 		}
5528 
5529 		if ((pio = zio_unique_parent(zio)) != NULL) {
5530 			/*
5531 			 * We're not a root i/o, so there's nothing to do
5532 			 * but notify our parent.  Don't propagate errors
5533 			 * upward since we haven't permanently failed yet.
5534 			 */
5535 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
5536 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
5537 			/*
5538 			 * This is a rare code path, so we don't bother with
5539 			 * "next_to_execute".
5540 			 */
5541 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL);
5542 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
5543 			/*
5544 			 * We'd fail again if we reexecuted now, so suspend
5545 			 * until conditions improve (e.g. device comes online).
5546 			 */
5547 			zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
5548 		} else {
5549 			/*
5550 			 * Reexecution is potentially a huge amount of work.
5551 			 * Hand it off to the otherwise-unused claim taskq.
5552 			 */
5553 			spa_taskq_dispatch(zio->io_spa,
5554 			    ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
5555 			    zio_reexecute, zio, B_FALSE);
5556 		}
5557 		return (NULL);
5558 	}
5559 
5560 	ASSERT(list_is_empty(&zio->io_child_list));
5561 	ASSERT(zio->io_reexecute == 0);
5562 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
5563 
5564 	/*
5565 	 * Report any checksum errors, since the I/O is complete.
5566 	 */
5567 	while (zio->io_cksum_report != NULL) {
5568 		zio_cksum_report_t *zcr = zio->io_cksum_report;
5569 		zio->io_cksum_report = zcr->zcr_next;
5570 		zcr->zcr_next = NULL;
5571 		zcr->zcr_finish(zcr, NULL);
5572 		zfs_ereport_free_checksum(zcr);
5573 	}
5574 
5575 	/*
5576 	 * It is the responsibility of the done callback to ensure that this
5577 	 * particular zio is no longer discoverable for adoption, and as
5578 	 * such, cannot acquire any new parents.
5579 	 */
5580 	if (zio->io_done)
5581 		zio->io_done(zio);
5582 
5583 	mutex_enter(&zio->io_lock);
5584 	zio->io_state[ZIO_WAIT_DONE] = 1;
5585 	mutex_exit(&zio->io_lock);
5586 
5587 	/*
5588 	 * We are done executing this zio.  We may want to execute a parent
5589 	 * next.  See the comment in zio_notify_parent().
5590 	 */
5591 	zio_t *next_to_execute = NULL;
5592 	zl = NULL;
5593 	for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
5594 		zio_link_t *remove_zl = zl;
5595 		pio_next = zio_walk_parents(zio, &zl);
5596 		zio_remove_child(pio, zio, remove_zl);
5597 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute);
5598 	}
5599 
5600 	if (zio->io_waiter != NULL) {
5601 		mutex_enter(&zio->io_lock);
5602 		zio->io_executor = NULL;
5603 		cv_broadcast(&zio->io_cv);
5604 		mutex_exit(&zio->io_lock);
5605 	} else {
5606 		zio_destroy(zio);
5607 	}
5608 
5609 	return (next_to_execute);
5610 }
5611 
5612 /*
5613  * ==========================================================================
5614  * I/O pipeline definition
5615  * ==========================================================================
5616  */
5617 static zio_pipe_stage_t *zio_pipeline[] = {
5618 	NULL,
5619 	zio_read_bp_init,
5620 	zio_write_bp_init,
5621 	zio_free_bp_init,
5622 	zio_issue_async,
5623 	zio_write_compress,
5624 	zio_encrypt,
5625 	zio_checksum_generate,
5626 	zio_nop_write,
5627 	zio_brt_free,
5628 	zio_ddt_read_start,
5629 	zio_ddt_read_done,
5630 	zio_ddt_write,
5631 	zio_ddt_free,
5632 	zio_gang_assemble,
5633 	zio_gang_issue,
5634 	zio_dva_throttle,
5635 	zio_dva_allocate,
5636 	zio_dva_free,
5637 	zio_dva_claim,
5638 	zio_ready,
5639 	zio_vdev_io_start,
5640 	zio_vdev_io_done,
5641 	zio_vdev_io_assess,
5642 	zio_checksum_verify,
5643 	zio_dio_checksum_verify,
5644 	zio_done
5645 };
5646 
5647 
5648 
5649 
5650 /*
5651  * Compare two zbookmark_phys_t's to see which we would reach first in a
5652  * pre-order traversal of the object tree.
5653  *
5654  * This is simple in every case aside from the meta-dnode object. For all other
5655  * objects, we traverse them in order (object 1 before object 2, and so on).
5656  * However, all of these objects are traversed while traversing object 0, since
5657  * the data it points to is the list of objects.  Thus, we need to convert to a
5658  * canonical representation so we can compare meta-dnode bookmarks to
5659  * non-meta-dnode bookmarks.
5660  *
5661  * We do this by calculating "equivalents" for each field of the zbookmark.
5662  * zbookmarks outside of the meta-dnode use their own object and level, and
5663  * calculate the level 0 equivalent (the first L0 blkid that is contained in the
5664  * blocks this bookmark refers to) by multiplying their blkid by their span
5665  * (the number of L0 blocks contained within one block at their level).
5666  * zbookmarks inside the meta-dnode calculate their object equivalent
5667  * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
5668  * level + 1<<31 (any value larger than a level could ever be) for their level.
5669  * This causes them to always compare before a bookmark in their object
5670  * equivalent, compare appropriately to bookmarks in other objects, and to
5671  * compare appropriately to other bookmarks in the meta-dnode.
5672  */
5673 int
5674 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
5675     const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
5676 {
5677 	/*
5678 	 * These variables represent the "equivalent" values for the zbookmark,
5679 	 * after converting zbookmarks inside the meta dnode to their
5680 	 * normal-object equivalents.
5681 	 */
5682 	uint64_t zb1obj, zb2obj;
5683 	uint64_t zb1L0, zb2L0;
5684 	uint64_t zb1level, zb2level;
5685 
5686 	if (zb1->zb_object == zb2->zb_object &&
5687 	    zb1->zb_level == zb2->zb_level &&
5688 	    zb1->zb_blkid == zb2->zb_blkid)
5689 		return (0);
5690 
5691 	IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT);
5692 	IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT);
5693 
5694 	/*
5695 	 * BP_SPANB calculates the span in blocks.
5696 	 */
5697 	zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
5698 	zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
5699 
5700 	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
5701 		zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
5702 		zb1L0 = 0;
5703 		zb1level = zb1->zb_level + COMPARE_META_LEVEL;
5704 	} else {
5705 		zb1obj = zb1->zb_object;
5706 		zb1level = zb1->zb_level;
5707 	}
5708 
5709 	if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
5710 		zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
5711 		zb2L0 = 0;
5712 		zb2level = zb2->zb_level + COMPARE_META_LEVEL;
5713 	} else {
5714 		zb2obj = zb2->zb_object;
5715 		zb2level = zb2->zb_level;
5716 	}
5717 
5718 	/* Now that we have a canonical representation, do the comparison. */
5719 	if (zb1obj != zb2obj)
5720 		return (zb1obj < zb2obj ? -1 : 1);
5721 	else if (zb1L0 != zb2L0)
5722 		return (zb1L0 < zb2L0 ? -1 : 1);
5723 	else if (zb1level != zb2level)
5724 		return (zb1level > zb2level ? -1 : 1);
5725 	/*
5726 	 * This can (theoretically) happen if the bookmarks have the same object
5727 	 * and level, but different blkids, if the block sizes are not the same.
5728 	 * There is presently no way to change the indirect block sizes
5729 	 */
5730 	return (0);
5731 }
5732 
5733 /*
5734  *  This function checks the following: given that last_block is the place that
5735  *  our traversal stopped last time, does that guarantee that we've visited
5736  *  every node under subtree_root?  Therefore, we can't just use the raw output
5737  *  of zbookmark_compare.  We have to pass in a modified version of
5738  *  subtree_root; by incrementing the block id, and then checking whether
5739  *  last_block is before or equal to that, we can tell whether or not having
5740  *  visited last_block implies that all of subtree_root's children have been
5741  *  visited.
5742  */
5743 boolean_t
5744 zbookmark_subtree_completed(const dnode_phys_t *dnp,
5745     const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
5746 {
5747 	zbookmark_phys_t mod_zb = *subtree_root;
5748 	mod_zb.zb_blkid++;
5749 	ASSERT0(last_block->zb_level);
5750 
5751 	/* The objset_phys_t isn't before anything. */
5752 	if (dnp == NULL)
5753 		return (B_FALSE);
5754 
5755 	/*
5756 	 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
5757 	 * data block size in sectors, because that variable is only used if
5758 	 * the bookmark refers to a block in the meta-dnode.  Since we don't
5759 	 * know without examining it what object it refers to, and there's no
5760 	 * harm in passing in this value in other cases, we always pass it in.
5761 	 *
5762 	 * We pass in 0 for the indirect block size shift because zb2 must be
5763 	 * level 0.  The indirect block size is only used to calculate the span
5764 	 * of the bookmark, but since the bookmark must be level 0, the span is
5765 	 * always 1, so the math works out.
5766 	 *
5767 	 * If you make changes to how the zbookmark_compare code works, be sure
5768 	 * to make sure that this code still works afterwards.
5769 	 */
5770 	return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
5771 	    1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
5772 	    last_block) <= 0);
5773 }
5774 
5775 /*
5776  * This function is similar to zbookmark_subtree_completed(), but returns true
5777  * if subtree_root is equal or ahead of last_block, i.e. still to be done.
5778  */
5779 boolean_t
5780 zbookmark_subtree_tbd(const dnode_phys_t *dnp,
5781     const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
5782 {
5783 	ASSERT0(last_block->zb_level);
5784 	if (dnp == NULL)
5785 		return (B_FALSE);
5786 	return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
5787 	    1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, subtree_root,
5788 	    last_block) >= 0);
5789 }
5790 
5791 EXPORT_SYMBOL(zio_type_name);
5792 EXPORT_SYMBOL(zio_buf_alloc);
5793 EXPORT_SYMBOL(zio_data_buf_alloc);
5794 EXPORT_SYMBOL(zio_buf_free);
5795 EXPORT_SYMBOL(zio_data_buf_free);
5796 
5797 ZFS_MODULE_PARAM(zfs_zio, zio_, slow_io_ms, INT, ZMOD_RW,
5798 	"Max I/O completion time (milliseconds) before marking it as slow");
5799 
5800 ZFS_MODULE_PARAM(zfs_zio, zio_, requeue_io_start_cut_in_line, INT, ZMOD_RW,
5801 	"Prioritize requeued I/O");
5802 
5803 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_deferred_free,  UINT, ZMOD_RW,
5804 	"Defer frees starting in this pass");
5805 
5806 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_dont_compress, UINT, ZMOD_RW,
5807 	"Don't compress starting in this pass");
5808 
5809 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_rewrite, UINT, ZMOD_RW,
5810 	"Rewrite new bps starting in this pass");
5811 
5812 ZFS_MODULE_PARAM(zfs_zio, zio_, dva_throttle_enabled, INT, ZMOD_RW,
5813 	"Throttle block allocations in the ZIO pipeline");
5814 
5815 ZFS_MODULE_PARAM(zfs_zio, zio_, deadman_log_all, INT, ZMOD_RW,
5816 	"Log all slow ZIOs, not just those with vdevs");
5817