xref: /freebsd/sys/contrib/openzfs/module/zfs/zio.c (revision a57ca37dd1848cd42844d9082c4a74c2ed57f68a)
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, Klara Inc.
27  * Copyright (c) 2019, Allan Jude
28  * Copyright (c) 2021, Datto, Inc.
29  */
30 
31 #include <sys/sysmacros.h>
32 #include <sys/zfs_context.h>
33 #include <sys/fm/fs/zfs.h>
34 #include <sys/spa.h>
35 #include <sys/txg.h>
36 #include <sys/spa_impl.h>
37 #include <sys/vdev_impl.h>
38 #include <sys/vdev_trim.h>
39 #include <sys/zio_impl.h>
40 #include <sys/zio_compress.h>
41 #include <sys/zio_checksum.h>
42 #include <sys/dmu_objset.h>
43 #include <sys/arc.h>
44 #include <sys/brt.h>
45 #include <sys/ddt.h>
46 #include <sys/blkptr.h>
47 #include <sys/zfeature.h>
48 #include <sys/dsl_scan.h>
49 #include <sys/metaslab_impl.h>
50 #include <sys/time.h>
51 #include <sys/trace_zfs.h>
52 #include <sys/abd.h>
53 #include <sys/dsl_crypt.h>
54 #include <cityhash.h>
55 
56 /*
57  * ==========================================================================
58  * I/O type descriptions
59  * ==========================================================================
60  */
61 const char *const zio_type_name[ZIO_TYPES] = {
62 	/*
63 	 * Note: Linux kernel thread name length is limited
64 	 * so these names will differ from upstream open zfs.
65 	 */
66 	"z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim"
67 };
68 
69 int zio_dva_throttle_enabled = B_TRUE;
70 static int zio_deadman_log_all = B_FALSE;
71 
72 /*
73  * ==========================================================================
74  * I/O kmem caches
75  * ==========================================================================
76  */
77 static kmem_cache_t *zio_cache;
78 static kmem_cache_t *zio_link_cache;
79 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
80 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
81 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
82 static uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
83 static uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
84 #endif
85 
86 /* Mark IOs as "slow" if they take longer than 30 seconds */
87 static uint_t zio_slow_io_ms = (30 * MILLISEC);
88 
89 #define	BP_SPANB(indblkshift, level) \
90 	(((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
91 #define	COMPARE_META_LEVEL	0x80000000ul
92 /*
93  * The following actions directly effect the spa's sync-to-convergence logic.
94  * The values below define the sync pass when we start performing the action.
95  * Care should be taken when changing these values as they directly impact
96  * spa_sync() performance. Tuning these values may introduce subtle performance
97  * pathologies and should only be done in the context of performance analysis.
98  * These tunables will eventually be removed and replaced with #defines once
99  * enough analysis has been done to determine optimal values.
100  *
101  * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
102  * regular blocks are not deferred.
103  *
104  * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable
105  * compression (including of metadata).  In practice, we don't have this
106  * many sync passes, so this has no effect.
107  *
108  * The original intent was that disabling compression would help the sync
109  * passes to converge. However, in practice disabling compression increases
110  * the average number of sync passes, because when we turn compression off, a
111  * lot of block's size will change and thus we have to re-allocate (not
112  * overwrite) them. It also increases the number of 128KB allocations (e.g.
113  * for indirect blocks and spacemaps) because these will not be compressed.
114  * The 128K allocations are especially detrimental to performance on highly
115  * fragmented systems, which may have very few free segments of this size,
116  * and may need to load new metaslabs to satisfy 128K allocations.
117  */
118 
119 /* defer frees starting in this pass */
120 uint_t zfs_sync_pass_deferred_free = 2;
121 
122 /* don't compress starting in this pass */
123 static uint_t zfs_sync_pass_dont_compress = 8;
124 
125 /* rewrite new bps starting in this pass */
126 static uint_t zfs_sync_pass_rewrite = 2;
127 
128 /*
129  * An allocating zio is one that either currently has the DVA allocate
130  * stage set or will have it later in its lifetime.
131  */
132 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
133 
134 /*
135  * Enable smaller cores by excluding metadata
136  * allocations as well.
137  */
138 int zio_exclude_metadata = 0;
139 static int zio_requeue_io_start_cut_in_line = 1;
140 
141 #ifdef ZFS_DEBUG
142 static const int zio_buf_debug_limit = 16384;
143 #else
144 static const int zio_buf_debug_limit = 0;
145 #endif
146 
147 static inline void __zio_execute(zio_t *zio);
148 
149 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
150 
151 void
152 zio_init(void)
153 {
154 	size_t c;
155 
156 	zio_cache = kmem_cache_create("zio_cache",
157 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
158 	zio_link_cache = kmem_cache_create("zio_link_cache",
159 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
160 
161 	/*
162 	 * For small buffers, we want a cache for each multiple of
163 	 * SPA_MINBLOCKSIZE.  For larger buffers, we want a cache
164 	 * for each quarter-power of 2.
165 	 */
166 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
167 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
168 		size_t p2 = size;
169 		size_t align = 0;
170 		size_t data_cflags, cflags;
171 
172 		data_cflags = KMC_NODEBUG;
173 		cflags = (zio_exclude_metadata || size > zio_buf_debug_limit) ?
174 		    KMC_NODEBUG : 0;
175 
176 		while (!ISP2(p2))
177 			p2 &= p2 - 1;
178 
179 #ifndef _KERNEL
180 		/*
181 		 * If we are using watchpoints, put each buffer on its own page,
182 		 * to eliminate the performance overhead of trapping to the
183 		 * kernel when modifying a non-watched buffer that shares the
184 		 * page with a watched buffer.
185 		 */
186 		if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
187 			continue;
188 		/*
189 		 * Here's the problem - on 4K native devices in userland on
190 		 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
191 		 * will fail with EINVAL, causing zdb (and others) to coredump.
192 		 * Since userland probably doesn't need optimized buffer caches,
193 		 * we just force 4K alignment on everything.
194 		 */
195 		align = 8 * SPA_MINBLOCKSIZE;
196 #else
197 		if (size < PAGESIZE) {
198 			align = SPA_MINBLOCKSIZE;
199 		} else if (IS_P2ALIGNED(size, p2 >> 2)) {
200 			align = PAGESIZE;
201 		}
202 #endif
203 
204 		if (align != 0) {
205 			char name[36];
206 			if (cflags == data_cflags) {
207 				/*
208 				 * Resulting kmem caches would be identical.
209 				 * Save memory by creating only one.
210 				 */
211 				(void) snprintf(name, sizeof (name),
212 				    "zio_buf_comb_%lu", (ulong_t)size);
213 				zio_buf_cache[c] = kmem_cache_create(name,
214 				    size, align, NULL, NULL, NULL, NULL, NULL,
215 				    cflags);
216 				zio_data_buf_cache[c] = zio_buf_cache[c];
217 				continue;
218 			}
219 			(void) snprintf(name, sizeof (name), "zio_buf_%lu",
220 			    (ulong_t)size);
221 			zio_buf_cache[c] = kmem_cache_create(name, size,
222 			    align, NULL, NULL, NULL, NULL, NULL, cflags);
223 
224 			(void) snprintf(name, sizeof (name), "zio_data_buf_%lu",
225 			    (ulong_t)size);
226 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
227 			    align, NULL, NULL, NULL, NULL, NULL, data_cflags);
228 		}
229 	}
230 
231 	while (--c != 0) {
232 		ASSERT(zio_buf_cache[c] != NULL);
233 		if (zio_buf_cache[c - 1] == NULL)
234 			zio_buf_cache[c - 1] = zio_buf_cache[c];
235 
236 		ASSERT(zio_data_buf_cache[c] != NULL);
237 		if (zio_data_buf_cache[c - 1] == NULL)
238 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
239 	}
240 
241 	zio_inject_init();
242 
243 	lz4_init();
244 }
245 
246 void
247 zio_fini(void)
248 {
249 	size_t n = SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT;
250 
251 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
252 	for (size_t i = 0; i < n; i++) {
253 		if (zio_buf_cache_allocs[i] != zio_buf_cache_frees[i])
254 			(void) printf("zio_fini: [%d] %llu != %llu\n",
255 			    (int)((i + 1) << SPA_MINBLOCKSHIFT),
256 			    (long long unsigned)zio_buf_cache_allocs[i],
257 			    (long long unsigned)zio_buf_cache_frees[i]);
258 	}
259 #endif
260 
261 	/*
262 	 * The same kmem cache can show up multiple times in both zio_buf_cache
263 	 * and zio_data_buf_cache. Do a wasteful but trivially correct scan to
264 	 * sort it out.
265 	 */
266 	for (size_t i = 0; i < n; i++) {
267 		kmem_cache_t *cache = zio_buf_cache[i];
268 		if (cache == NULL)
269 			continue;
270 		for (size_t j = i; j < n; j++) {
271 			if (cache == zio_buf_cache[j])
272 				zio_buf_cache[j] = NULL;
273 			if (cache == zio_data_buf_cache[j])
274 				zio_data_buf_cache[j] = NULL;
275 		}
276 		kmem_cache_destroy(cache);
277 	}
278 
279 	for (size_t i = 0; i < n; i++) {
280 		kmem_cache_t *cache = zio_data_buf_cache[i];
281 		if (cache == NULL)
282 			continue;
283 		for (size_t j = i; j < n; j++) {
284 			if (cache == zio_data_buf_cache[j])
285 				zio_data_buf_cache[j] = NULL;
286 		}
287 		kmem_cache_destroy(cache);
288 	}
289 
290 	for (size_t i = 0; i < n; i++) {
291 		VERIFY3P(zio_buf_cache[i], ==, NULL);
292 		VERIFY3P(zio_data_buf_cache[i], ==, NULL);
293 	}
294 
295 	kmem_cache_destroy(zio_link_cache);
296 	kmem_cache_destroy(zio_cache);
297 
298 	zio_inject_fini();
299 
300 	lz4_fini();
301 }
302 
303 /*
304  * ==========================================================================
305  * Allocate and free I/O buffers
306  * ==========================================================================
307  */
308 
309 /*
310  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
311  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
312  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
313  * excess / transient data in-core during a crashdump.
314  */
315 void *
316 zio_buf_alloc(size_t size)
317 {
318 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
319 
320 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
321 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
322 	atomic_add_64(&zio_buf_cache_allocs[c], 1);
323 #endif
324 
325 	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
326 }
327 
328 /*
329  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
330  * crashdump if the kernel panics.  This exists so that we will limit the amount
331  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
332  * of kernel heap dumped to disk when the kernel panics)
333  */
334 void *
335 zio_data_buf_alloc(size_t size)
336 {
337 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
338 
339 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
340 
341 	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
342 }
343 
344 void
345 zio_buf_free(void *buf, size_t size)
346 {
347 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
348 
349 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
350 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
351 	atomic_add_64(&zio_buf_cache_frees[c], 1);
352 #endif
353 
354 	kmem_cache_free(zio_buf_cache[c], buf);
355 }
356 
357 void
358 zio_data_buf_free(void *buf, size_t size)
359 {
360 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
361 
362 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
363 
364 	kmem_cache_free(zio_data_buf_cache[c], buf);
365 }
366 
367 static void
368 zio_abd_free(void *abd, size_t size)
369 {
370 	(void) size;
371 	abd_free((abd_t *)abd);
372 }
373 
374 /*
375  * ==========================================================================
376  * Push and pop I/O transform buffers
377  * ==========================================================================
378  */
379 void
380 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
381     zio_transform_func_t *transform)
382 {
383 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
384 
385 	zt->zt_orig_abd = zio->io_abd;
386 	zt->zt_orig_size = zio->io_size;
387 	zt->zt_bufsize = bufsize;
388 	zt->zt_transform = transform;
389 
390 	zt->zt_next = zio->io_transform_stack;
391 	zio->io_transform_stack = zt;
392 
393 	zio->io_abd = data;
394 	zio->io_size = size;
395 }
396 
397 void
398 zio_pop_transforms(zio_t *zio)
399 {
400 	zio_transform_t *zt;
401 
402 	while ((zt = zio->io_transform_stack) != NULL) {
403 		if (zt->zt_transform != NULL)
404 			zt->zt_transform(zio,
405 			    zt->zt_orig_abd, zt->zt_orig_size);
406 
407 		if (zt->zt_bufsize != 0)
408 			abd_free(zio->io_abd);
409 
410 		zio->io_abd = zt->zt_orig_abd;
411 		zio->io_size = zt->zt_orig_size;
412 		zio->io_transform_stack = zt->zt_next;
413 
414 		kmem_free(zt, sizeof (zio_transform_t));
415 	}
416 }
417 
418 /*
419  * ==========================================================================
420  * I/O transform callbacks for subblocks, decompression, and decryption
421  * ==========================================================================
422  */
423 static void
424 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
425 {
426 	ASSERT(zio->io_size > size);
427 
428 	if (zio->io_type == ZIO_TYPE_READ)
429 		abd_copy(data, zio->io_abd, size);
430 }
431 
432 static void
433 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
434 {
435 	if (zio->io_error == 0) {
436 		void *tmp = abd_borrow_buf(data, size);
437 		int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
438 		    zio->io_abd, tmp, zio->io_size, size,
439 		    &zio->io_prop.zp_complevel);
440 		abd_return_buf_copy(data, tmp, size);
441 
442 		if (zio_injection_enabled && ret == 0)
443 			ret = zio_handle_fault_injection(zio, EINVAL);
444 
445 		if (ret != 0)
446 			zio->io_error = SET_ERROR(EIO);
447 	}
448 }
449 
450 static void
451 zio_decrypt(zio_t *zio, abd_t *data, uint64_t size)
452 {
453 	int ret;
454 	void *tmp;
455 	blkptr_t *bp = zio->io_bp;
456 	spa_t *spa = zio->io_spa;
457 	uint64_t dsobj = zio->io_bookmark.zb_objset;
458 	uint64_t lsize = BP_GET_LSIZE(bp);
459 	dmu_object_type_t ot = BP_GET_TYPE(bp);
460 	uint8_t salt[ZIO_DATA_SALT_LEN];
461 	uint8_t iv[ZIO_DATA_IV_LEN];
462 	uint8_t mac[ZIO_DATA_MAC_LEN];
463 	boolean_t no_crypt = B_FALSE;
464 
465 	ASSERT(BP_USES_CRYPT(bp));
466 	ASSERT3U(size, !=, 0);
467 
468 	if (zio->io_error != 0)
469 		return;
470 
471 	/*
472 	 * Verify the cksum of MACs stored in an indirect bp. It will always
473 	 * be possible to verify this since it does not require an encryption
474 	 * key.
475 	 */
476 	if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {
477 		zio_crypt_decode_mac_bp(bp, mac);
478 
479 		if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
480 			/*
481 			 * We haven't decompressed the data yet, but
482 			 * zio_crypt_do_indirect_mac_checksum() requires
483 			 * decompressed data to be able to parse out the MACs
484 			 * from the indirect block. We decompress it now and
485 			 * throw away the result after we are finished.
486 			 */
487 			tmp = zio_buf_alloc(lsize);
488 			ret = zio_decompress_data(BP_GET_COMPRESS(bp),
489 			    zio->io_abd, tmp, zio->io_size, lsize,
490 			    &zio->io_prop.zp_complevel);
491 			if (ret != 0) {
492 				ret = SET_ERROR(EIO);
493 				goto error;
494 			}
495 			ret = zio_crypt_do_indirect_mac_checksum(B_FALSE,
496 			    tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac);
497 			zio_buf_free(tmp, lsize);
498 		} else {
499 			ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
500 			    zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac);
501 		}
502 		abd_copy(data, zio->io_abd, size);
503 
504 		if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) {
505 			ret = zio_handle_decrypt_injection(spa,
506 			    &zio->io_bookmark, ot, ECKSUM);
507 		}
508 		if (ret != 0)
509 			goto error;
510 
511 		return;
512 	}
513 
514 	/*
515 	 * If this is an authenticated block, just check the MAC. It would be
516 	 * nice to separate this out into its own flag, but when this was done,
517 	 * we had run out of bits in what is now zio_flag_t. Future cleanup
518 	 * could make this a flag bit.
519 	 */
520 	if (BP_IS_AUTHENTICATED(bp)) {
521 		if (ot == DMU_OT_OBJSET) {
522 			ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa,
523 			    dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp));
524 		} else {
525 			zio_crypt_decode_mac_bp(bp, mac);
526 			ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj,
527 			    zio->io_abd, size, mac);
528 			if (zio_injection_enabled && ret == 0) {
529 				ret = zio_handle_decrypt_injection(spa,
530 				    &zio->io_bookmark, ot, ECKSUM);
531 			}
532 		}
533 		abd_copy(data, zio->io_abd, size);
534 
535 		if (ret != 0)
536 			goto error;
537 
538 		return;
539 	}
540 
541 	zio_crypt_decode_params_bp(bp, salt, iv);
542 
543 	if (ot == DMU_OT_INTENT_LOG) {
544 		tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t));
545 		zio_crypt_decode_mac_zil(tmp, mac);
546 		abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t));
547 	} else {
548 		zio_crypt_decode_mac_bp(bp, mac);
549 	}
550 
551 	ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp),
552 	    BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data,
553 	    zio->io_abd, &no_crypt);
554 	if (no_crypt)
555 		abd_copy(data, zio->io_abd, size);
556 
557 	if (ret != 0)
558 		goto error;
559 
560 	return;
561 
562 error:
563 	/* assert that the key was found unless this was speculative */
564 	ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE));
565 
566 	/*
567 	 * If there was a decryption / authentication error return EIO as
568 	 * the io_error. If this was not a speculative zio, create an ereport.
569 	 */
570 	if (ret == ECKSUM) {
571 		zio->io_error = SET_ERROR(EIO);
572 		if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
573 			spa_log_error(spa, &zio->io_bookmark,
574 			    &zio->io_bp->blk_birth);
575 			(void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
576 			    spa, NULL, &zio->io_bookmark, zio, 0);
577 		}
578 	} else {
579 		zio->io_error = ret;
580 	}
581 }
582 
583 /*
584  * ==========================================================================
585  * I/O parent/child relationships and pipeline interlocks
586  * ==========================================================================
587  */
588 zio_t *
589 zio_walk_parents(zio_t *cio, zio_link_t **zl)
590 {
591 	list_t *pl = &cio->io_parent_list;
592 
593 	*zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
594 	if (*zl == NULL)
595 		return (NULL);
596 
597 	ASSERT((*zl)->zl_child == cio);
598 	return ((*zl)->zl_parent);
599 }
600 
601 zio_t *
602 zio_walk_children(zio_t *pio, zio_link_t **zl)
603 {
604 	list_t *cl = &pio->io_child_list;
605 
606 	ASSERT(MUTEX_HELD(&pio->io_lock));
607 
608 	*zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
609 	if (*zl == NULL)
610 		return (NULL);
611 
612 	ASSERT((*zl)->zl_parent == pio);
613 	return ((*zl)->zl_child);
614 }
615 
616 zio_t *
617 zio_unique_parent(zio_t *cio)
618 {
619 	zio_link_t *zl = NULL;
620 	zio_t *pio = zio_walk_parents(cio, &zl);
621 
622 	VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
623 	return (pio);
624 }
625 
626 void
627 zio_add_child(zio_t *pio, zio_t *cio)
628 {
629 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
630 
631 	/*
632 	 * Logical I/Os can have logical, gang, or vdev children.
633 	 * Gang I/Os can have gang or vdev children.
634 	 * Vdev I/Os can only have vdev children.
635 	 * The following ASSERT captures all of these constraints.
636 	 */
637 	ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
638 
639 	zl->zl_parent = pio;
640 	zl->zl_child = cio;
641 
642 	mutex_enter(&pio->io_lock);
643 	mutex_enter(&cio->io_lock);
644 
645 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
646 
647 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
648 		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
649 
650 	list_insert_head(&pio->io_child_list, zl);
651 	list_insert_head(&cio->io_parent_list, zl);
652 
653 	pio->io_child_count++;
654 	cio->io_parent_count++;
655 
656 	mutex_exit(&cio->io_lock);
657 	mutex_exit(&pio->io_lock);
658 }
659 
660 static void
661 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
662 {
663 	ASSERT(zl->zl_parent == pio);
664 	ASSERT(zl->zl_child == cio);
665 
666 	mutex_enter(&pio->io_lock);
667 	mutex_enter(&cio->io_lock);
668 
669 	list_remove(&pio->io_child_list, zl);
670 	list_remove(&cio->io_parent_list, zl);
671 
672 	pio->io_child_count--;
673 	cio->io_parent_count--;
674 
675 	mutex_exit(&cio->io_lock);
676 	mutex_exit(&pio->io_lock);
677 	kmem_cache_free(zio_link_cache, zl);
678 }
679 
680 static boolean_t
681 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
682 {
683 	boolean_t waiting = B_FALSE;
684 
685 	mutex_enter(&zio->io_lock);
686 	ASSERT(zio->io_stall == NULL);
687 	for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
688 		if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
689 			continue;
690 
691 		uint64_t *countp = &zio->io_children[c][wait];
692 		if (*countp != 0) {
693 			zio->io_stage >>= 1;
694 			ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
695 			zio->io_stall = countp;
696 			waiting = B_TRUE;
697 			break;
698 		}
699 	}
700 	mutex_exit(&zio->io_lock);
701 	return (waiting);
702 }
703 
704 __attribute__((always_inline))
705 static inline void
706 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait,
707     zio_t **next_to_executep)
708 {
709 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
710 	int *errorp = &pio->io_child_error[zio->io_child_type];
711 
712 	mutex_enter(&pio->io_lock);
713 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
714 		*errorp = zio_worst_error(*errorp, zio->io_error);
715 	pio->io_reexecute |= zio->io_reexecute;
716 	ASSERT3U(*countp, >, 0);
717 
718 	(*countp)--;
719 
720 	if (*countp == 0 && pio->io_stall == countp) {
721 		zio_taskq_type_t type =
722 		    pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
723 		    ZIO_TASKQ_INTERRUPT;
724 		pio->io_stall = NULL;
725 		mutex_exit(&pio->io_lock);
726 
727 		/*
728 		 * If we can tell the caller to execute this parent next, do
729 		 * so. We only do this if the parent's zio type matches the
730 		 * child's type. Otherwise dispatch the parent zio in its
731 		 * own taskq.
732 		 *
733 		 * Having the caller execute the parent when possible reduces
734 		 * locking on the zio taskq's, reduces context switch
735 		 * overhead, and has no recursion penalty.  Note that one
736 		 * read from disk typically causes at least 3 zio's: a
737 		 * zio_null(), the logical zio_read(), and then a physical
738 		 * zio.  When the physical ZIO completes, we are able to call
739 		 * zio_done() on all 3 of these zio's from one invocation of
740 		 * zio_execute() by returning the parent back to
741 		 * zio_execute().  Since the parent isn't executed until this
742 		 * thread returns back to zio_execute(), the caller should do
743 		 * so promptly.
744 		 *
745 		 * In other cases, dispatching the parent prevents
746 		 * overflowing the stack when we have deeply nested
747 		 * parent-child relationships, as we do with the "mega zio"
748 		 * of writes for spa_sync(), and the chain of ZIL blocks.
749 		 */
750 		if (next_to_executep != NULL && *next_to_executep == NULL &&
751 		    pio->io_type == zio->io_type) {
752 			*next_to_executep = pio;
753 		} else {
754 			zio_taskq_dispatch(pio, type, B_FALSE);
755 		}
756 	} else {
757 		mutex_exit(&pio->io_lock);
758 	}
759 }
760 
761 static void
762 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
763 {
764 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
765 		zio->io_error = zio->io_child_error[c];
766 }
767 
768 int
769 zio_bookmark_compare(const void *x1, const void *x2)
770 {
771 	const zio_t *z1 = x1;
772 	const zio_t *z2 = x2;
773 
774 	if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
775 		return (-1);
776 	if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
777 		return (1);
778 
779 	if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
780 		return (-1);
781 	if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
782 		return (1);
783 
784 	if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
785 		return (-1);
786 	if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
787 		return (1);
788 
789 	if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
790 		return (-1);
791 	if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
792 		return (1);
793 
794 	if (z1 < z2)
795 		return (-1);
796 	if (z1 > z2)
797 		return (1);
798 
799 	return (0);
800 }
801 
802 /*
803  * ==========================================================================
804  * Create the various types of I/O (read, write, free, etc)
805  * ==========================================================================
806  */
807 static zio_t *
808 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
809     abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
810     void *private, zio_type_t type, zio_priority_t priority,
811     zio_flag_t flags, vdev_t *vd, uint64_t offset,
812     const zbookmark_phys_t *zb, enum zio_stage stage,
813     enum zio_stage pipeline)
814 {
815 	zio_t *zio;
816 
817 	IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE);
818 	ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
819 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
820 
821 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
822 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
823 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
824 
825 	IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0);
826 
827 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
828 	memset(zio, 0, sizeof (zio_t));
829 
830 	mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL);
831 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
832 
833 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
834 	    offsetof(zio_link_t, zl_parent_node));
835 	list_create(&zio->io_child_list, sizeof (zio_link_t),
836 	    offsetof(zio_link_t, zl_child_node));
837 	metaslab_trace_init(&zio->io_alloc_list);
838 
839 	if (vd != NULL)
840 		zio->io_child_type = ZIO_CHILD_VDEV;
841 	else if (flags & ZIO_FLAG_GANG_CHILD)
842 		zio->io_child_type = ZIO_CHILD_GANG;
843 	else if (flags & ZIO_FLAG_DDT_CHILD)
844 		zio->io_child_type = ZIO_CHILD_DDT;
845 	else
846 		zio->io_child_type = ZIO_CHILD_LOGICAL;
847 
848 	if (bp != NULL) {
849 		zio->io_bp = (blkptr_t *)bp;
850 		zio->io_bp_copy = *bp;
851 		zio->io_bp_orig = *bp;
852 		if (type != ZIO_TYPE_WRITE ||
853 		    zio->io_child_type == ZIO_CHILD_DDT)
854 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
855 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
856 			zio->io_logical = zio;
857 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
858 			pipeline |= ZIO_GANG_STAGES;
859 	}
860 
861 	zio->io_spa = spa;
862 	zio->io_txg = txg;
863 	zio->io_done = done;
864 	zio->io_private = private;
865 	zio->io_type = type;
866 	zio->io_priority = priority;
867 	zio->io_vd = vd;
868 	zio->io_offset = offset;
869 	zio->io_orig_abd = zio->io_abd = data;
870 	zio->io_orig_size = zio->io_size = psize;
871 	zio->io_lsize = lsize;
872 	zio->io_orig_flags = zio->io_flags = flags;
873 	zio->io_orig_stage = zio->io_stage = stage;
874 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
875 	zio->io_pipeline_trace = ZIO_STAGE_OPEN;
876 
877 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
878 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
879 
880 	if (zb != NULL)
881 		zio->io_bookmark = *zb;
882 
883 	if (pio != NULL) {
884 		zio->io_metaslab_class = pio->io_metaslab_class;
885 		if (zio->io_logical == NULL)
886 			zio->io_logical = pio->io_logical;
887 		if (zio->io_child_type == ZIO_CHILD_GANG)
888 			zio->io_gang_leader = pio->io_gang_leader;
889 		zio_add_child(pio, zio);
890 	}
891 
892 	taskq_init_ent(&zio->io_tqent);
893 
894 	return (zio);
895 }
896 
897 void
898 zio_destroy(zio_t *zio)
899 {
900 	metaslab_trace_fini(&zio->io_alloc_list);
901 	list_destroy(&zio->io_parent_list);
902 	list_destroy(&zio->io_child_list);
903 	mutex_destroy(&zio->io_lock);
904 	cv_destroy(&zio->io_cv);
905 	kmem_cache_free(zio_cache, zio);
906 }
907 
908 zio_t *
909 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
910     void *private, zio_flag_t flags)
911 {
912 	zio_t *zio;
913 
914 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
915 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
916 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
917 
918 	return (zio);
919 }
920 
921 zio_t *
922 zio_root(spa_t *spa, zio_done_func_t *done, void *private, zio_flag_t flags)
923 {
924 	return (zio_null(NULL, spa, NULL, done, private, flags));
925 }
926 
927 static int
928 zfs_blkptr_verify_log(spa_t *spa, const blkptr_t *bp,
929     enum blk_verify_flag blk_verify, const char *fmt, ...)
930 {
931 	va_list adx;
932 	char buf[256];
933 
934 	va_start(adx, fmt);
935 	(void) vsnprintf(buf, sizeof (buf), fmt, adx);
936 	va_end(adx);
937 
938 	zfs_dbgmsg("bad blkptr at %px: "
939 	    "DVA[0]=%#llx/%#llx "
940 	    "DVA[1]=%#llx/%#llx "
941 	    "DVA[2]=%#llx/%#llx "
942 	    "prop=%#llx "
943 	    "pad=%#llx,%#llx "
944 	    "phys_birth=%#llx "
945 	    "birth=%#llx "
946 	    "fill=%#llx "
947 	    "cksum=%#llx/%#llx/%#llx/%#llx",
948 	    bp,
949 	    (long long)bp->blk_dva[0].dva_word[0],
950 	    (long long)bp->blk_dva[0].dva_word[1],
951 	    (long long)bp->blk_dva[1].dva_word[0],
952 	    (long long)bp->blk_dva[1].dva_word[1],
953 	    (long long)bp->blk_dva[2].dva_word[0],
954 	    (long long)bp->blk_dva[2].dva_word[1],
955 	    (long long)bp->blk_prop,
956 	    (long long)bp->blk_pad[0],
957 	    (long long)bp->blk_pad[1],
958 	    (long long)bp->blk_phys_birth,
959 	    (long long)bp->blk_birth,
960 	    (long long)bp->blk_fill,
961 	    (long long)bp->blk_cksum.zc_word[0],
962 	    (long long)bp->blk_cksum.zc_word[1],
963 	    (long long)bp->blk_cksum.zc_word[2],
964 	    (long long)bp->blk_cksum.zc_word[3]);
965 	switch (blk_verify) {
966 	case BLK_VERIFY_HALT:
967 		zfs_panic_recover("%s: %s", spa_name(spa), buf);
968 		break;
969 	case BLK_VERIFY_LOG:
970 		zfs_dbgmsg("%s: %s", spa_name(spa), buf);
971 		break;
972 	case BLK_VERIFY_ONLY:
973 		break;
974 	}
975 
976 	return (1);
977 }
978 
979 /*
980  * Verify the block pointer fields contain reasonable values.  This means
981  * it only contains known object types, checksum/compression identifiers,
982  * block sizes within the maximum allowed limits, valid DVAs, etc.
983  *
984  * If everything checks out B_TRUE is returned.  The zfs_blkptr_verify
985  * argument controls the behavior when an invalid field is detected.
986  *
987  * Values for blk_verify_flag:
988  *   BLK_VERIFY_ONLY: evaluate the block
989  *   BLK_VERIFY_LOG: evaluate the block and log problems
990  *   BLK_VERIFY_HALT: call zfs_panic_recover on error
991  *
992  * Values for blk_config_flag:
993  *   BLK_CONFIG_HELD: caller holds SCL_VDEV for writer
994  *   BLK_CONFIG_NEEDED: caller holds no config lock, SCL_VDEV will be
995  *   obtained for reader
996  *   BLK_CONFIG_SKIP: skip checks which require SCL_VDEV, for better
997  *   performance
998  */
999 boolean_t
1000 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp,
1001     enum blk_config_flag blk_config, enum blk_verify_flag blk_verify)
1002 {
1003 	int errors = 0;
1004 
1005 	if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
1006 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1007 		    "blkptr at %px has invalid TYPE %llu",
1008 		    bp, (longlong_t)BP_GET_TYPE(bp));
1009 	}
1010 	if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS) {
1011 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1012 		    "blkptr at %px has invalid CHECKSUM %llu",
1013 		    bp, (longlong_t)BP_GET_CHECKSUM(bp));
1014 	}
1015 	if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS) {
1016 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1017 		    "blkptr at %px has invalid COMPRESS %llu",
1018 		    bp, (longlong_t)BP_GET_COMPRESS(bp));
1019 	}
1020 	if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
1021 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1022 		    "blkptr at %px has invalid LSIZE %llu",
1023 		    bp, (longlong_t)BP_GET_LSIZE(bp));
1024 	}
1025 	if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
1026 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1027 		    "blkptr at %px has invalid PSIZE %llu",
1028 		    bp, (longlong_t)BP_GET_PSIZE(bp));
1029 	}
1030 
1031 	if (BP_IS_EMBEDDED(bp)) {
1032 		if (BPE_GET_ETYPE(bp) >= NUM_BP_EMBEDDED_TYPES) {
1033 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1034 			    "blkptr at %px has invalid ETYPE %llu",
1035 			    bp, (longlong_t)BPE_GET_ETYPE(bp));
1036 		}
1037 	}
1038 
1039 	/*
1040 	 * Do not verify individual DVAs if the config is not trusted. This
1041 	 * will be done once the zio is executed in vdev_mirror_map_alloc.
1042 	 */
1043 	if (!spa->spa_trust_config)
1044 		return (errors == 0);
1045 
1046 	switch (blk_config) {
1047 	case BLK_CONFIG_HELD:
1048 		ASSERT(spa_config_held(spa, SCL_VDEV, RW_WRITER));
1049 		break;
1050 	case BLK_CONFIG_NEEDED:
1051 		spa_config_enter(spa, SCL_VDEV, bp, RW_READER);
1052 		break;
1053 	case BLK_CONFIG_SKIP:
1054 		return (errors == 0);
1055 	default:
1056 		panic("invalid blk_config %u", blk_config);
1057 	}
1058 
1059 	/*
1060 	 * Pool-specific checks.
1061 	 *
1062 	 * Note: it would be nice to verify that the blk_birth and
1063 	 * BP_PHYSICAL_BIRTH() are not too large.  However, spa_freeze()
1064 	 * allows the birth time of log blocks (and dmu_sync()-ed blocks
1065 	 * that are in the log) to be arbitrarily large.
1066 	 */
1067 	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
1068 		const dva_t *dva = &bp->blk_dva[i];
1069 		uint64_t vdevid = DVA_GET_VDEV(dva);
1070 
1071 		if (vdevid >= spa->spa_root_vdev->vdev_children) {
1072 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1073 			    "blkptr at %px DVA %u has invalid VDEV %llu",
1074 			    bp, i, (longlong_t)vdevid);
1075 			continue;
1076 		}
1077 		vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1078 		if (vd == NULL) {
1079 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1080 			    "blkptr at %px DVA %u has invalid VDEV %llu",
1081 			    bp, i, (longlong_t)vdevid);
1082 			continue;
1083 		}
1084 		if (vd->vdev_ops == &vdev_hole_ops) {
1085 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1086 			    "blkptr at %px DVA %u has hole VDEV %llu",
1087 			    bp, i, (longlong_t)vdevid);
1088 			continue;
1089 		}
1090 		if (vd->vdev_ops == &vdev_missing_ops) {
1091 			/*
1092 			 * "missing" vdevs are valid during import, but we
1093 			 * don't have their detailed info (e.g. asize), so
1094 			 * we can't perform any more checks on them.
1095 			 */
1096 			continue;
1097 		}
1098 		uint64_t offset = DVA_GET_OFFSET(dva);
1099 		uint64_t asize = DVA_GET_ASIZE(dva);
1100 		if (DVA_GET_GANG(dva))
1101 			asize = vdev_gang_header_asize(vd);
1102 		if (offset + asize > vd->vdev_asize) {
1103 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1104 			    "blkptr at %px DVA %u has invalid OFFSET %llu",
1105 			    bp, i, (longlong_t)offset);
1106 		}
1107 	}
1108 	if (blk_config == BLK_CONFIG_NEEDED)
1109 		spa_config_exit(spa, SCL_VDEV, bp);
1110 
1111 	return (errors == 0);
1112 }
1113 
1114 boolean_t
1115 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
1116 {
1117 	(void) bp;
1118 	uint64_t vdevid = DVA_GET_VDEV(dva);
1119 
1120 	if (vdevid >= spa->spa_root_vdev->vdev_children)
1121 		return (B_FALSE);
1122 
1123 	vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1124 	if (vd == NULL)
1125 		return (B_FALSE);
1126 
1127 	if (vd->vdev_ops == &vdev_hole_ops)
1128 		return (B_FALSE);
1129 
1130 	if (vd->vdev_ops == &vdev_missing_ops) {
1131 		return (B_FALSE);
1132 	}
1133 
1134 	uint64_t offset = DVA_GET_OFFSET(dva);
1135 	uint64_t asize = DVA_GET_ASIZE(dva);
1136 
1137 	if (DVA_GET_GANG(dva))
1138 		asize = vdev_gang_header_asize(vd);
1139 	if (offset + asize > vd->vdev_asize)
1140 		return (B_FALSE);
1141 
1142 	return (B_TRUE);
1143 }
1144 
1145 zio_t *
1146 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
1147     abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
1148     zio_priority_t priority, zio_flag_t flags, const zbookmark_phys_t *zb)
1149 {
1150 	zio_t *zio;
1151 
1152 	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
1153 	    data, size, size, done, private,
1154 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
1155 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1156 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
1157 
1158 	return (zio);
1159 }
1160 
1161 zio_t *
1162 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
1163     abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
1164     zio_done_func_t *ready, zio_done_func_t *children_ready,
1165     zio_done_func_t *physdone, zio_done_func_t *done,
1166     void *private, zio_priority_t priority, zio_flag_t flags,
1167     const zbookmark_phys_t *zb)
1168 {
1169 	zio_t *zio;
1170 
1171 	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
1172 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
1173 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
1174 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
1175 	    DMU_OT_IS_VALID(zp->zp_type) &&
1176 	    zp->zp_level < 32 &&
1177 	    zp->zp_copies > 0 &&
1178 	    zp->zp_copies <= spa_max_replication(spa));
1179 
1180 	zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
1181 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
1182 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1183 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
1184 
1185 	zio->io_ready = ready;
1186 	zio->io_children_ready = children_ready;
1187 	zio->io_physdone = physdone;
1188 	zio->io_prop = *zp;
1189 
1190 	/*
1191 	 * Data can be NULL if we are going to call zio_write_override() to
1192 	 * provide the already-allocated BP.  But we may need the data to
1193 	 * verify a dedup hit (if requested).  In this case, don't try to
1194 	 * dedup (just take the already-allocated BP verbatim). Encrypted
1195 	 * dedup blocks need data as well so we also disable dedup in this
1196 	 * case.
1197 	 */
1198 	if (data == NULL &&
1199 	    (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) {
1200 		zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
1201 	}
1202 
1203 	return (zio);
1204 }
1205 
1206 zio_t *
1207 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
1208     uint64_t size, zio_done_func_t *done, void *private,
1209     zio_priority_t priority, zio_flag_t flags, zbookmark_phys_t *zb)
1210 {
1211 	zio_t *zio;
1212 
1213 	zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
1214 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
1215 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
1216 
1217 	return (zio);
1218 }
1219 
1220 void
1221 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite,
1222     boolean_t brtwrite)
1223 {
1224 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1225 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1226 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1227 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
1228 	ASSERT(!brtwrite || !nopwrite);
1229 
1230 	/*
1231 	 * We must reset the io_prop to match the values that existed
1232 	 * when the bp was first written by dmu_sync() keeping in mind
1233 	 * that nopwrite and dedup are mutually exclusive.
1234 	 */
1235 	zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
1236 	zio->io_prop.zp_nopwrite = nopwrite;
1237 	zio->io_prop.zp_brtwrite = brtwrite;
1238 	zio->io_prop.zp_copies = copies;
1239 	zio->io_bp_override = bp;
1240 }
1241 
1242 void
1243 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
1244 {
1245 
1246 	(void) zfs_blkptr_verify(spa, bp, BLK_CONFIG_NEEDED, BLK_VERIFY_HALT);
1247 
1248 	/*
1249 	 * The check for EMBEDDED is a performance optimization.  We
1250 	 * process the free here (by ignoring it) rather than
1251 	 * putting it on the list and then processing it in zio_free_sync().
1252 	 */
1253 	if (BP_IS_EMBEDDED(bp))
1254 		return;
1255 
1256 	/*
1257 	 * Frees that are for the currently-syncing txg, are not going to be
1258 	 * deferred, and which will not need to do a read (i.e. not GANG or
1259 	 * DEDUP), can be processed immediately.  Otherwise, put them on the
1260 	 * in-memory list for later processing.
1261 	 *
1262 	 * Note that we only defer frees after zfs_sync_pass_deferred_free
1263 	 * when the log space map feature is disabled. [see relevant comment
1264 	 * in spa_sync_iterate_to_convergence()]
1265 	 */
1266 	if (BP_IS_GANG(bp) ||
1267 	    BP_GET_DEDUP(bp) ||
1268 	    txg != spa->spa_syncing_txg ||
1269 	    (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free &&
1270 	    !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) ||
1271 	    brt_maybe_exists(spa, bp)) {
1272 		metaslab_check_free(spa, bp);
1273 		bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1274 	} else {
1275 		VERIFY3P(zio_free_sync(NULL, spa, txg, bp, 0), ==, NULL);
1276 	}
1277 }
1278 
1279 /*
1280  * To improve performance, this function may return NULL if we were able
1281  * to do the free immediately.  This avoids the cost of creating a zio
1282  * (and linking it to the parent, etc).
1283  */
1284 zio_t *
1285 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1286     zio_flag_t flags)
1287 {
1288 	ASSERT(!BP_IS_HOLE(bp));
1289 	ASSERT(spa_syncing_txg(spa) == txg);
1290 
1291 	if (BP_IS_EMBEDDED(bp))
1292 		return (NULL);
1293 
1294 	metaslab_check_free(spa, bp);
1295 	arc_freed(spa, bp);
1296 	dsl_scan_freed(spa, bp);
1297 
1298 	if (BP_IS_GANG(bp) ||
1299 	    BP_GET_DEDUP(bp) ||
1300 	    brt_maybe_exists(spa, bp)) {
1301 		/*
1302 		 * GANG, DEDUP and BRT blocks can induce a read (for the gang
1303 		 * block header, the DDT or the BRT), so issue them
1304 		 * asynchronously so that this thread is not tied up.
1305 		 */
1306 		enum zio_stage stage =
1307 		    ZIO_FREE_PIPELINE | ZIO_STAGE_ISSUE_ASYNC;
1308 
1309 		return (zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1310 		    BP_GET_PSIZE(bp), NULL, NULL,
1311 		    ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1312 		    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage));
1313 	} else {
1314 		metaslab_free(spa, bp, txg, B_FALSE);
1315 		return (NULL);
1316 	}
1317 }
1318 
1319 zio_t *
1320 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1321     zio_done_func_t *done, void *private, zio_flag_t flags)
1322 {
1323 	zio_t *zio;
1324 
1325 	(void) zfs_blkptr_verify(spa, bp, (flags & ZIO_FLAG_CONFIG_WRITER) ?
1326 	    BLK_CONFIG_HELD : BLK_CONFIG_NEEDED, BLK_VERIFY_HALT);
1327 
1328 	if (BP_IS_EMBEDDED(bp))
1329 		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1330 
1331 	/*
1332 	 * A claim is an allocation of a specific block.  Claims are needed
1333 	 * to support immediate writes in the intent log.  The issue is that
1334 	 * immediate writes contain committed data, but in a txg that was
1335 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
1336 	 * the intent log claims all blocks that contain immediate write data
1337 	 * so that the SPA knows they're in use.
1338 	 *
1339 	 * All claims *must* be resolved in the first txg -- before the SPA
1340 	 * starts allocating blocks -- so that nothing is allocated twice.
1341 	 * If txg == 0 we just verify that the block is claimable.
1342 	 */
1343 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1344 	    spa_min_claim_txg(spa));
1345 	ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1346 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(8) */
1347 
1348 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1349 	    BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1350 	    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1351 	ASSERT0(zio->io_queued_timestamp);
1352 
1353 	return (zio);
1354 }
1355 
1356 zio_t *
1357 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
1358     zio_done_func_t *done, void *private, zio_flag_t flags)
1359 {
1360 	zio_t *zio;
1361 	int c;
1362 
1363 	if (vd->vdev_children == 0) {
1364 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1365 		    ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1366 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1367 
1368 		zio->io_cmd = cmd;
1369 	} else {
1370 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1371 
1372 		for (c = 0; c < vd->vdev_children; c++)
1373 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1374 			    done, private, flags));
1375 	}
1376 
1377 	return (zio);
1378 }
1379 
1380 zio_t *
1381 zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1382     zio_done_func_t *done, void *private, zio_priority_t priority,
1383     zio_flag_t flags, enum trim_flag trim_flags)
1384 {
1385 	zio_t *zio;
1386 
1387 	ASSERT0(vd->vdev_children);
1388 	ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
1389 	ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1390 	ASSERT3U(size, !=, 0);
1391 
1392 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done,
1393 	    private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL,
1394 	    vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE);
1395 	zio->io_trim_flags = trim_flags;
1396 
1397 	return (zio);
1398 }
1399 
1400 zio_t *
1401 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1402     abd_t *data, int checksum, zio_done_func_t *done, void *private,
1403     zio_priority_t priority, zio_flag_t flags, boolean_t labels)
1404 {
1405 	zio_t *zio;
1406 
1407 	ASSERT(vd->vdev_children == 0);
1408 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1409 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1410 	ASSERT3U(offset + size, <=, vd->vdev_psize);
1411 
1412 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1413 	    private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1414 	    offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1415 
1416 	zio->io_prop.zp_checksum = checksum;
1417 
1418 	return (zio);
1419 }
1420 
1421 zio_t *
1422 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1423     abd_t *data, int checksum, zio_done_func_t *done, void *private,
1424     zio_priority_t priority, zio_flag_t flags, boolean_t labels)
1425 {
1426 	zio_t *zio;
1427 
1428 	ASSERT(vd->vdev_children == 0);
1429 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1430 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1431 	ASSERT3U(offset + size, <=, vd->vdev_psize);
1432 
1433 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1434 	    private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1435 	    offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1436 
1437 	zio->io_prop.zp_checksum = checksum;
1438 
1439 	if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1440 		/*
1441 		 * zec checksums are necessarily destructive -- they modify
1442 		 * the end of the write buffer to hold the verifier/checksum.
1443 		 * Therefore, we must make a local copy in case the data is
1444 		 * being written to multiple places in parallel.
1445 		 */
1446 		abd_t *wbuf = abd_alloc_sametype(data, size);
1447 		abd_copy(wbuf, data, size);
1448 
1449 		zio_push_transform(zio, wbuf, size, size, NULL);
1450 	}
1451 
1452 	return (zio);
1453 }
1454 
1455 /*
1456  * Create a child I/O to do some work for us.
1457  */
1458 zio_t *
1459 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1460     abd_t *data, uint64_t size, int type, zio_priority_t priority,
1461     zio_flag_t flags, zio_done_func_t *done, void *private)
1462 {
1463 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1464 	zio_t *zio;
1465 
1466 	/*
1467 	 * vdev child I/Os do not propagate their error to the parent.
1468 	 * Therefore, for correct operation the caller *must* check for
1469 	 * and handle the error in the child i/o's done callback.
1470 	 * The only exceptions are i/os that we don't care about
1471 	 * (OPTIONAL or REPAIR).
1472 	 */
1473 	ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1474 	    done != NULL);
1475 
1476 	if (type == ZIO_TYPE_READ && bp != NULL) {
1477 		/*
1478 		 * If we have the bp, then the child should perform the
1479 		 * checksum and the parent need not.  This pushes error
1480 		 * detection as close to the leaves as possible and
1481 		 * eliminates redundant checksums in the interior nodes.
1482 		 */
1483 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1484 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1485 	}
1486 
1487 	if (vd->vdev_ops->vdev_op_leaf) {
1488 		ASSERT0(vd->vdev_children);
1489 		offset += VDEV_LABEL_START_SIZE;
1490 	}
1491 
1492 	flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1493 
1494 	/*
1495 	 * If we've decided to do a repair, the write is not speculative --
1496 	 * even if the original read was.
1497 	 */
1498 	if (flags & ZIO_FLAG_IO_REPAIR)
1499 		flags &= ~ZIO_FLAG_SPECULATIVE;
1500 
1501 	/*
1502 	 * If we're creating a child I/O that is not associated with a
1503 	 * top-level vdev, then the child zio is not an allocating I/O.
1504 	 * If this is a retried I/O then we ignore it since we will
1505 	 * have already processed the original allocating I/O.
1506 	 */
1507 	if (flags & ZIO_FLAG_IO_ALLOCATING &&
1508 	    (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1509 		ASSERT(pio->io_metaslab_class != NULL);
1510 		ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled);
1511 		ASSERT(type == ZIO_TYPE_WRITE);
1512 		ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1513 		ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1514 		ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1515 		    pio->io_child_type == ZIO_CHILD_GANG);
1516 
1517 		flags &= ~ZIO_FLAG_IO_ALLOCATING;
1518 	}
1519 
1520 
1521 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1522 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1523 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1524 	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1525 
1526 	zio->io_physdone = pio->io_physdone;
1527 	if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1528 		zio->io_logical->io_phys_children++;
1529 
1530 	return (zio);
1531 }
1532 
1533 zio_t *
1534 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1535     zio_type_t type, zio_priority_t priority, zio_flag_t flags,
1536     zio_done_func_t *done, void *private)
1537 {
1538 	zio_t *zio;
1539 
1540 	ASSERT(vd->vdev_ops->vdev_op_leaf);
1541 
1542 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1543 	    data, size, size, done, private, type, priority,
1544 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1545 	    vd, offset, NULL,
1546 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1547 
1548 	return (zio);
1549 }
1550 
1551 void
1552 zio_flush(zio_t *zio, vdev_t *vd)
1553 {
1554 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1555 	    NULL, NULL,
1556 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1557 }
1558 
1559 void
1560 zio_shrink(zio_t *zio, uint64_t size)
1561 {
1562 	ASSERT3P(zio->io_executor, ==, NULL);
1563 	ASSERT3U(zio->io_orig_size, ==, zio->io_size);
1564 	ASSERT3U(size, <=, zio->io_size);
1565 
1566 	/*
1567 	 * We don't shrink for raidz because of problems with the
1568 	 * reconstruction when reading back less than the block size.
1569 	 * Note, BP_IS_RAIDZ() assumes no compression.
1570 	 */
1571 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1572 	if (!BP_IS_RAIDZ(zio->io_bp)) {
1573 		/* we are not doing a raw write */
1574 		ASSERT3U(zio->io_size, ==, zio->io_lsize);
1575 		zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1576 	}
1577 }
1578 
1579 /*
1580  * ==========================================================================
1581  * Prepare to read and write logical blocks
1582  * ==========================================================================
1583  */
1584 
1585 static zio_t *
1586 zio_read_bp_init(zio_t *zio)
1587 {
1588 	blkptr_t *bp = zio->io_bp;
1589 	uint64_t psize =
1590 	    BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1591 
1592 	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1593 
1594 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1595 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
1596 	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1597 		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1598 		    psize, psize, zio_decompress);
1599 	}
1600 
1601 	if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) ||
1602 	    BP_HAS_INDIRECT_MAC_CKSUM(bp)) &&
1603 	    zio->io_child_type == ZIO_CHILD_LOGICAL) {
1604 		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1605 		    psize, psize, zio_decrypt);
1606 	}
1607 
1608 	if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1609 		int psize = BPE_GET_PSIZE(bp);
1610 		void *data = abd_borrow_buf(zio->io_abd, psize);
1611 
1612 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1613 		decode_embedded_bp_compressed(bp, data);
1614 		abd_return_buf_copy(zio->io_abd, data, psize);
1615 	} else {
1616 		ASSERT(!BP_IS_EMBEDDED(bp));
1617 		ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1618 	}
1619 
1620 	if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1621 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1622 
1623 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1624 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1625 
1626 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1627 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1628 
1629 	return (zio);
1630 }
1631 
1632 static zio_t *
1633 zio_write_bp_init(zio_t *zio)
1634 {
1635 	if (!IO_IS_ALLOCATING(zio))
1636 		return (zio);
1637 
1638 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1639 
1640 	if (zio->io_bp_override) {
1641 		blkptr_t *bp = zio->io_bp;
1642 		zio_prop_t *zp = &zio->io_prop;
1643 
1644 		ASSERT(bp->blk_birth != zio->io_txg);
1645 
1646 		*bp = *zio->io_bp_override;
1647 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1648 
1649 		if (zp->zp_brtwrite)
1650 			return (zio);
1651 
1652 		ASSERT(!BP_GET_DEDUP(zio->io_bp_override));
1653 
1654 		if (BP_IS_EMBEDDED(bp))
1655 			return (zio);
1656 
1657 		/*
1658 		 * If we've been overridden and nopwrite is set then
1659 		 * set the flag accordingly to indicate that a nopwrite
1660 		 * has already occurred.
1661 		 */
1662 		if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1663 			ASSERT(!zp->zp_dedup);
1664 			ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1665 			zio->io_flags |= ZIO_FLAG_NOPWRITE;
1666 			return (zio);
1667 		}
1668 
1669 		ASSERT(!zp->zp_nopwrite);
1670 
1671 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1672 			return (zio);
1673 
1674 		ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1675 		    ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1676 
1677 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum &&
1678 		    !zp->zp_encrypt) {
1679 			BP_SET_DEDUP(bp, 1);
1680 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1681 			return (zio);
1682 		}
1683 
1684 		/*
1685 		 * We were unable to handle this as an override bp, treat
1686 		 * it as a regular write I/O.
1687 		 */
1688 		zio->io_bp_override = NULL;
1689 		*bp = zio->io_bp_orig;
1690 		zio->io_pipeline = zio->io_orig_pipeline;
1691 	}
1692 
1693 	return (zio);
1694 }
1695 
1696 static zio_t *
1697 zio_write_compress(zio_t *zio)
1698 {
1699 	spa_t *spa = zio->io_spa;
1700 	zio_prop_t *zp = &zio->io_prop;
1701 	enum zio_compress compress = zp->zp_compress;
1702 	blkptr_t *bp = zio->io_bp;
1703 	uint64_t lsize = zio->io_lsize;
1704 	uint64_t psize = zio->io_size;
1705 	uint32_t pass = 1;
1706 
1707 	/*
1708 	 * If our children haven't all reached the ready stage,
1709 	 * wait for them and then repeat this pipeline stage.
1710 	 */
1711 	if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1712 	    ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1713 		return (NULL);
1714 	}
1715 
1716 	if (!IO_IS_ALLOCATING(zio))
1717 		return (zio);
1718 
1719 	if (zio->io_children_ready != NULL) {
1720 		/*
1721 		 * Now that all our children are ready, run the callback
1722 		 * associated with this zio in case it wants to modify the
1723 		 * data to be written.
1724 		 */
1725 		ASSERT3U(zp->zp_level, >, 0);
1726 		zio->io_children_ready(zio);
1727 	}
1728 
1729 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1730 	ASSERT(zio->io_bp_override == NULL);
1731 
1732 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1733 		/*
1734 		 * We're rewriting an existing block, which means we're
1735 		 * working on behalf of spa_sync().  For spa_sync() to
1736 		 * converge, it must eventually be the case that we don't
1737 		 * have to allocate new blocks.  But compression changes
1738 		 * the blocksize, which forces a reallocate, and makes
1739 		 * convergence take longer.  Therefore, after the first
1740 		 * few passes, stop compressing to ensure convergence.
1741 		 */
1742 		pass = spa_sync_pass(spa);
1743 
1744 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
1745 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1746 		ASSERT(!BP_GET_DEDUP(bp));
1747 
1748 		if (pass >= zfs_sync_pass_dont_compress)
1749 			compress = ZIO_COMPRESS_OFF;
1750 
1751 		/* Make sure someone doesn't change their mind on overwrites */
1752 		ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1753 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1754 	}
1755 
1756 	/* If it's a compressed write that is not raw, compress the buffer. */
1757 	if (compress != ZIO_COMPRESS_OFF &&
1758 	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1759 		void *cbuf = NULL;
1760 		psize = zio_compress_data(compress, zio->io_abd, &cbuf, lsize,
1761 		    zp->zp_complevel);
1762 		if (psize == 0) {
1763 			compress = ZIO_COMPRESS_OFF;
1764 		} else if (psize >= lsize) {
1765 			compress = ZIO_COMPRESS_OFF;
1766 			if (cbuf != NULL)
1767 				zio_buf_free(cbuf, lsize);
1768 		} else if (!zp->zp_dedup && !zp->zp_encrypt &&
1769 		    psize <= BPE_PAYLOAD_SIZE &&
1770 		    zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1771 		    spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1772 			encode_embedded_bp_compressed(bp,
1773 			    cbuf, compress, lsize, psize);
1774 			BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1775 			BP_SET_TYPE(bp, zio->io_prop.zp_type);
1776 			BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1777 			zio_buf_free(cbuf, lsize);
1778 			bp->blk_birth = zio->io_txg;
1779 			zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1780 			ASSERT(spa_feature_is_active(spa,
1781 			    SPA_FEATURE_EMBEDDED_DATA));
1782 			return (zio);
1783 		} else {
1784 			/*
1785 			 * Round compressed size up to the minimum allocation
1786 			 * size of the smallest-ashift device, and zero the
1787 			 * tail. This ensures that the compressed size of the
1788 			 * BP (and thus compressratio property) are correct,
1789 			 * in that we charge for the padding used to fill out
1790 			 * the last sector.
1791 			 */
1792 			ASSERT3U(spa->spa_min_alloc, >=, SPA_MINBLOCKSHIFT);
1793 			size_t rounded = (size_t)roundup(psize,
1794 			    spa->spa_min_alloc);
1795 			if (rounded >= lsize) {
1796 				compress = ZIO_COMPRESS_OFF;
1797 				zio_buf_free(cbuf, lsize);
1798 				psize = lsize;
1799 			} else {
1800 				abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1801 				abd_take_ownership_of_buf(cdata, B_TRUE);
1802 				abd_zero_off(cdata, psize, rounded - psize);
1803 				psize = rounded;
1804 				zio_push_transform(zio, cdata,
1805 				    psize, lsize, NULL);
1806 			}
1807 		}
1808 
1809 		/*
1810 		 * We were unable to handle this as an override bp, treat
1811 		 * it as a regular write I/O.
1812 		 */
1813 		zio->io_bp_override = NULL;
1814 		*bp = zio->io_bp_orig;
1815 		zio->io_pipeline = zio->io_orig_pipeline;
1816 
1817 	} else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 &&
1818 	    zp->zp_type == DMU_OT_DNODE) {
1819 		/*
1820 		 * The DMU actually relies on the zio layer's compression
1821 		 * to free metadnode blocks that have had all contained
1822 		 * dnodes freed. As a result, even when doing a raw
1823 		 * receive, we must check whether the block can be compressed
1824 		 * to a hole.
1825 		 */
1826 		psize = zio_compress_data(ZIO_COMPRESS_EMPTY,
1827 		    zio->io_abd, NULL, lsize, zp->zp_complevel);
1828 		if (psize == 0 || psize >= lsize)
1829 			compress = ZIO_COMPRESS_OFF;
1830 	} else if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS &&
1831 	    !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) {
1832 		/*
1833 		 * If we are raw receiving an encrypted dataset we should not
1834 		 * take this codepath because it will change the on-disk block
1835 		 * and decryption will fail.
1836 		 */
1837 		size_t rounded = MIN((size_t)roundup(psize,
1838 		    spa->spa_min_alloc), lsize);
1839 
1840 		if (rounded != psize) {
1841 			abd_t *cdata = abd_alloc_linear(rounded, B_TRUE);
1842 			abd_zero_off(cdata, psize, rounded - psize);
1843 			abd_copy_off(cdata, zio->io_abd, 0, 0, psize);
1844 			psize = rounded;
1845 			zio_push_transform(zio, cdata,
1846 			    psize, rounded, NULL);
1847 		}
1848 	} else {
1849 		ASSERT3U(psize, !=, 0);
1850 	}
1851 
1852 	/*
1853 	 * The final pass of spa_sync() must be all rewrites, but the first
1854 	 * few passes offer a trade-off: allocating blocks defers convergence,
1855 	 * but newly allocated blocks are sequential, so they can be written
1856 	 * to disk faster.  Therefore, we allow the first few passes of
1857 	 * spa_sync() to allocate new blocks, but force rewrites after that.
1858 	 * There should only be a handful of blocks after pass 1 in any case.
1859 	 */
1860 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1861 	    BP_GET_PSIZE(bp) == psize &&
1862 	    pass >= zfs_sync_pass_rewrite) {
1863 		VERIFY3U(psize, !=, 0);
1864 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1865 
1866 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1867 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1868 	} else {
1869 		BP_ZERO(bp);
1870 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1871 	}
1872 
1873 	if (psize == 0) {
1874 		if (zio->io_bp_orig.blk_birth != 0 &&
1875 		    spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1876 			BP_SET_LSIZE(bp, lsize);
1877 			BP_SET_TYPE(bp, zp->zp_type);
1878 			BP_SET_LEVEL(bp, zp->zp_level);
1879 			BP_SET_BIRTH(bp, zio->io_txg, 0);
1880 		}
1881 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1882 	} else {
1883 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1884 		BP_SET_LSIZE(bp, lsize);
1885 		BP_SET_TYPE(bp, zp->zp_type);
1886 		BP_SET_LEVEL(bp, zp->zp_level);
1887 		BP_SET_PSIZE(bp, psize);
1888 		BP_SET_COMPRESS(bp, compress);
1889 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1890 		BP_SET_DEDUP(bp, zp->zp_dedup);
1891 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1892 		if (zp->zp_dedup) {
1893 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1894 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1895 			ASSERT(!zp->zp_encrypt ||
1896 			    DMU_OT_IS_ENCRYPTED(zp->zp_type));
1897 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1898 		}
1899 		if (zp->zp_nopwrite) {
1900 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1901 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1902 			zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1903 		}
1904 	}
1905 	return (zio);
1906 }
1907 
1908 static zio_t *
1909 zio_free_bp_init(zio_t *zio)
1910 {
1911 	blkptr_t *bp = zio->io_bp;
1912 
1913 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1914 		if (BP_GET_DEDUP(bp))
1915 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1916 	}
1917 
1918 	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1919 
1920 	return (zio);
1921 }
1922 
1923 /*
1924  * ==========================================================================
1925  * Execute the I/O pipeline
1926  * ==========================================================================
1927  */
1928 
1929 static void
1930 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1931 {
1932 	spa_t *spa = zio->io_spa;
1933 	zio_type_t t = zio->io_type;
1934 	int flags = (cutinline ? TQ_FRONT : 0);
1935 
1936 	/*
1937 	 * If we're a config writer or a probe, the normal issue and
1938 	 * interrupt threads may all be blocked waiting for the config lock.
1939 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1940 	 */
1941 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1942 		t = ZIO_TYPE_NULL;
1943 
1944 	/*
1945 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1946 	 */
1947 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1948 		t = ZIO_TYPE_NULL;
1949 
1950 	/*
1951 	 * If this is a high priority I/O, then use the high priority taskq if
1952 	 * available.
1953 	 */
1954 	if ((zio->io_priority == ZIO_PRIORITY_NOW ||
1955 	    zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) &&
1956 	    spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1957 		q++;
1958 
1959 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1960 
1961 	/*
1962 	 * NB: We are assuming that the zio can only be dispatched
1963 	 * to a single taskq at a time.  It would be a grievous error
1964 	 * to dispatch the zio to another taskq at the same time.
1965 	 */
1966 	ASSERT(taskq_empty_ent(&zio->io_tqent));
1967 	spa_taskq_dispatch_ent(spa, t, q, zio_execute, zio, flags,
1968 	    &zio->io_tqent);
1969 }
1970 
1971 static boolean_t
1972 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1973 {
1974 	spa_t *spa = zio->io_spa;
1975 
1976 	taskq_t *tq = taskq_of_curthread();
1977 
1978 	for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1979 		spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1980 		uint_t i;
1981 		for (i = 0; i < tqs->stqs_count; i++) {
1982 			if (tqs->stqs_taskq[i] == tq)
1983 				return (B_TRUE);
1984 		}
1985 	}
1986 
1987 	return (B_FALSE);
1988 }
1989 
1990 static zio_t *
1991 zio_issue_async(zio_t *zio)
1992 {
1993 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1994 
1995 	return (NULL);
1996 }
1997 
1998 void
1999 zio_interrupt(void *zio)
2000 {
2001 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
2002 }
2003 
2004 void
2005 zio_delay_interrupt(zio_t *zio)
2006 {
2007 	/*
2008 	 * The timeout_generic() function isn't defined in userspace, so
2009 	 * rather than trying to implement the function, the zio delay
2010 	 * functionality has been disabled for userspace builds.
2011 	 */
2012 
2013 #ifdef _KERNEL
2014 	/*
2015 	 * If io_target_timestamp is zero, then no delay has been registered
2016 	 * for this IO, thus jump to the end of this function and "skip" the
2017 	 * delay; issuing it directly to the zio layer.
2018 	 */
2019 	if (zio->io_target_timestamp != 0) {
2020 		hrtime_t now = gethrtime();
2021 
2022 		if (now >= zio->io_target_timestamp) {
2023 			/*
2024 			 * This IO has already taken longer than the target
2025 			 * delay to complete, so we don't want to delay it
2026 			 * any longer; we "miss" the delay and issue it
2027 			 * directly to the zio layer. This is likely due to
2028 			 * the target latency being set to a value less than
2029 			 * the underlying hardware can satisfy (e.g. delay
2030 			 * set to 1ms, but the disks take 10ms to complete an
2031 			 * IO request).
2032 			 */
2033 
2034 			DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
2035 			    hrtime_t, now);
2036 
2037 			zio_interrupt(zio);
2038 		} else {
2039 			taskqid_t tid;
2040 			hrtime_t diff = zio->io_target_timestamp - now;
2041 			clock_t expire_at_tick = ddi_get_lbolt() +
2042 			    NSEC_TO_TICK(diff);
2043 
2044 			DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
2045 			    hrtime_t, now, hrtime_t, diff);
2046 
2047 			if (NSEC_TO_TICK(diff) == 0) {
2048 				/* Our delay is less than a jiffy - just spin */
2049 				zfs_sleep_until(zio->io_target_timestamp);
2050 				zio_interrupt(zio);
2051 			} else {
2052 				/*
2053 				 * Use taskq_dispatch_delay() in the place of
2054 				 * OpenZFS's timeout_generic().
2055 				 */
2056 				tid = taskq_dispatch_delay(system_taskq,
2057 				    zio_interrupt, zio, TQ_NOSLEEP,
2058 				    expire_at_tick);
2059 				if (tid == TASKQID_INVALID) {
2060 					/*
2061 					 * Couldn't allocate a task.  Just
2062 					 * finish the zio without a delay.
2063 					 */
2064 					zio_interrupt(zio);
2065 				}
2066 			}
2067 		}
2068 		return;
2069 	}
2070 #endif
2071 	DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
2072 	zio_interrupt(zio);
2073 }
2074 
2075 static void
2076 zio_deadman_impl(zio_t *pio, int ziodepth)
2077 {
2078 	zio_t *cio, *cio_next;
2079 	zio_link_t *zl = NULL;
2080 	vdev_t *vd = pio->io_vd;
2081 
2082 	if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) {
2083 		vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL;
2084 		zbookmark_phys_t *zb = &pio->io_bookmark;
2085 		uint64_t delta = gethrtime() - pio->io_timestamp;
2086 		uint64_t failmode = spa_get_deadman_failmode(pio->io_spa);
2087 
2088 		zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu "
2089 		    "delta=%llu queued=%llu io=%llu "
2090 		    "path=%s "
2091 		    "last=%llu type=%d "
2092 		    "priority=%d flags=0x%llx stage=0x%x "
2093 		    "pipeline=0x%x pipeline-trace=0x%x "
2094 		    "objset=%llu object=%llu "
2095 		    "level=%llu blkid=%llu "
2096 		    "offset=%llu size=%llu "
2097 		    "error=%d",
2098 		    ziodepth, pio, pio->io_timestamp,
2099 		    (u_longlong_t)delta, pio->io_delta, pio->io_delay,
2100 		    vd ? vd->vdev_path : "NULL",
2101 		    vq ? vq->vq_io_complete_ts : 0, pio->io_type,
2102 		    pio->io_priority, (u_longlong_t)pio->io_flags,
2103 		    pio->io_stage, pio->io_pipeline, pio->io_pipeline_trace,
2104 		    (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object,
2105 		    (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid,
2106 		    (u_longlong_t)pio->io_offset, (u_longlong_t)pio->io_size,
2107 		    pio->io_error);
2108 		(void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN,
2109 		    pio->io_spa, vd, zb, pio, 0);
2110 
2111 		if (failmode == ZIO_FAILURE_MODE_CONTINUE &&
2112 		    taskq_empty_ent(&pio->io_tqent)) {
2113 			zio_interrupt(pio);
2114 		}
2115 	}
2116 
2117 	mutex_enter(&pio->io_lock);
2118 	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2119 		cio_next = zio_walk_children(pio, &zl);
2120 		zio_deadman_impl(cio, ziodepth + 1);
2121 	}
2122 	mutex_exit(&pio->io_lock);
2123 }
2124 
2125 /*
2126  * Log the critical information describing this zio and all of its children
2127  * using the zfs_dbgmsg() interface then post deadman event for the ZED.
2128  */
2129 void
2130 zio_deadman(zio_t *pio, const char *tag)
2131 {
2132 	spa_t *spa = pio->io_spa;
2133 	char *name = spa_name(spa);
2134 
2135 	if (!zfs_deadman_enabled || spa_suspended(spa))
2136 		return;
2137 
2138 	zio_deadman_impl(pio, 0);
2139 
2140 	switch (spa_get_deadman_failmode(spa)) {
2141 	case ZIO_FAILURE_MODE_WAIT:
2142 		zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name);
2143 		break;
2144 
2145 	case ZIO_FAILURE_MODE_CONTINUE:
2146 		zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name);
2147 		break;
2148 
2149 	case ZIO_FAILURE_MODE_PANIC:
2150 		fm_panic("%s determined I/O to pool '%s' is hung.", tag, name);
2151 		break;
2152 	}
2153 }
2154 
2155 /*
2156  * Execute the I/O pipeline until one of the following occurs:
2157  * (1) the I/O completes; (2) the pipeline stalls waiting for
2158  * dependent child I/Os; (3) the I/O issues, so we're waiting
2159  * for an I/O completion interrupt; (4) the I/O is delegated by
2160  * vdev-level caching or aggregation; (5) the I/O is deferred
2161  * due to vdev-level queueing; (6) the I/O is handed off to
2162  * another thread.  In all cases, the pipeline stops whenever
2163  * there's no CPU work; it never burns a thread in cv_wait_io().
2164  *
2165  * There's no locking on io_stage because there's no legitimate way
2166  * for multiple threads to be attempting to process the same I/O.
2167  */
2168 static zio_pipe_stage_t *zio_pipeline[];
2169 
2170 /*
2171  * zio_execute() is a wrapper around the static function
2172  * __zio_execute() so that we can force  __zio_execute() to be
2173  * inlined.  This reduces stack overhead which is important
2174  * because __zio_execute() is called recursively in several zio
2175  * code paths.  zio_execute() itself cannot be inlined because
2176  * it is externally visible.
2177  */
2178 void
2179 zio_execute(void *zio)
2180 {
2181 	fstrans_cookie_t cookie;
2182 
2183 	cookie = spl_fstrans_mark();
2184 	__zio_execute(zio);
2185 	spl_fstrans_unmark(cookie);
2186 }
2187 
2188 /*
2189  * Used to determine if in the current context the stack is sized large
2190  * enough to allow zio_execute() to be called recursively.  A minimum
2191  * stack size of 16K is required to avoid needing to re-dispatch the zio.
2192  */
2193 static boolean_t
2194 zio_execute_stack_check(zio_t *zio)
2195 {
2196 #if !defined(HAVE_LARGE_STACKS)
2197 	dsl_pool_t *dp = spa_get_dsl(zio->io_spa);
2198 
2199 	/* Executing in txg_sync_thread() context. */
2200 	if (dp && curthread == dp->dp_tx.tx_sync_thread)
2201 		return (B_TRUE);
2202 
2203 	/* Pool initialization outside of zio_taskq context. */
2204 	if (dp && spa_is_initializing(dp->dp_spa) &&
2205 	    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
2206 	    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))
2207 		return (B_TRUE);
2208 #else
2209 	(void) zio;
2210 #endif /* HAVE_LARGE_STACKS */
2211 
2212 	return (B_FALSE);
2213 }
2214 
2215 __attribute__((always_inline))
2216 static inline void
2217 __zio_execute(zio_t *zio)
2218 {
2219 	ASSERT3U(zio->io_queued_timestamp, >, 0);
2220 
2221 	while (zio->io_stage < ZIO_STAGE_DONE) {
2222 		enum zio_stage pipeline = zio->io_pipeline;
2223 		enum zio_stage stage = zio->io_stage;
2224 
2225 		zio->io_executor = curthread;
2226 
2227 		ASSERT(!MUTEX_HELD(&zio->io_lock));
2228 		ASSERT(ISP2(stage));
2229 		ASSERT(zio->io_stall == NULL);
2230 
2231 		do {
2232 			stage <<= 1;
2233 		} while ((stage & pipeline) == 0);
2234 
2235 		ASSERT(stage <= ZIO_STAGE_DONE);
2236 
2237 		/*
2238 		 * If we are in interrupt context and this pipeline stage
2239 		 * will grab a config lock that is held across I/O,
2240 		 * or may wait for an I/O that needs an interrupt thread
2241 		 * to complete, issue async to avoid deadlock.
2242 		 *
2243 		 * For VDEV_IO_START, we cut in line so that the io will
2244 		 * be sent to disk promptly.
2245 		 */
2246 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
2247 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
2248 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2249 			    zio_requeue_io_start_cut_in_line : B_FALSE;
2250 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2251 			return;
2252 		}
2253 
2254 		/*
2255 		 * If the current context doesn't have large enough stacks
2256 		 * the zio must be issued asynchronously to prevent overflow.
2257 		 */
2258 		if (zio_execute_stack_check(zio)) {
2259 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2260 			    zio_requeue_io_start_cut_in_line : B_FALSE;
2261 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2262 			return;
2263 		}
2264 
2265 		zio->io_stage = stage;
2266 		zio->io_pipeline_trace |= zio->io_stage;
2267 
2268 		/*
2269 		 * The zio pipeline stage returns the next zio to execute
2270 		 * (typically the same as this one), or NULL if we should
2271 		 * stop.
2272 		 */
2273 		zio = zio_pipeline[highbit64(stage) - 1](zio);
2274 
2275 		if (zio == NULL)
2276 			return;
2277 	}
2278 }
2279 
2280 
2281 /*
2282  * ==========================================================================
2283  * Initiate I/O, either sync or async
2284  * ==========================================================================
2285  */
2286 int
2287 zio_wait(zio_t *zio)
2288 {
2289 	/*
2290 	 * Some routines, like zio_free_sync(), may return a NULL zio
2291 	 * to avoid the performance overhead of creating and then destroying
2292 	 * an unneeded zio.  For the callers' simplicity, we accept a NULL
2293 	 * zio and ignore it.
2294 	 */
2295 	if (zio == NULL)
2296 		return (0);
2297 
2298 	long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms);
2299 	int error;
2300 
2301 	ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN);
2302 	ASSERT3P(zio->io_executor, ==, NULL);
2303 
2304 	zio->io_waiter = curthread;
2305 	ASSERT0(zio->io_queued_timestamp);
2306 	zio->io_queued_timestamp = gethrtime();
2307 
2308 	__zio_execute(zio);
2309 
2310 	mutex_enter(&zio->io_lock);
2311 	while (zio->io_executor != NULL) {
2312 		error = cv_timedwait_io(&zio->io_cv, &zio->io_lock,
2313 		    ddi_get_lbolt() + timeout);
2314 
2315 		if (zfs_deadman_enabled && error == -1 &&
2316 		    gethrtime() - zio->io_queued_timestamp >
2317 		    spa_deadman_ziotime(zio->io_spa)) {
2318 			mutex_exit(&zio->io_lock);
2319 			timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms);
2320 			zio_deadman(zio, FTAG);
2321 			mutex_enter(&zio->io_lock);
2322 		}
2323 	}
2324 	mutex_exit(&zio->io_lock);
2325 
2326 	error = zio->io_error;
2327 	zio_destroy(zio);
2328 
2329 	return (error);
2330 }
2331 
2332 void
2333 zio_nowait(zio_t *zio)
2334 {
2335 	/*
2336 	 * See comment in zio_wait().
2337 	 */
2338 	if (zio == NULL)
2339 		return;
2340 
2341 	ASSERT3P(zio->io_executor, ==, NULL);
2342 
2343 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
2344 	    list_is_empty(&zio->io_parent_list)) {
2345 		zio_t *pio;
2346 
2347 		/*
2348 		 * This is a logical async I/O with no parent to wait for it.
2349 		 * We add it to the spa_async_root_zio "Godfather" I/O which
2350 		 * will ensure they complete prior to unloading the pool.
2351 		 */
2352 		spa_t *spa = zio->io_spa;
2353 		pio = spa->spa_async_zio_root[CPU_SEQID_UNSTABLE];
2354 
2355 		zio_add_child(pio, zio);
2356 	}
2357 
2358 	ASSERT0(zio->io_queued_timestamp);
2359 	zio->io_queued_timestamp = gethrtime();
2360 	__zio_execute(zio);
2361 }
2362 
2363 /*
2364  * ==========================================================================
2365  * Reexecute, cancel, or suspend/resume failed I/O
2366  * ==========================================================================
2367  */
2368 
2369 static void
2370 zio_reexecute(void *arg)
2371 {
2372 	zio_t *pio = arg;
2373 	zio_t *cio, *cio_next;
2374 
2375 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
2376 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
2377 	ASSERT(pio->io_gang_leader == NULL);
2378 	ASSERT(pio->io_gang_tree == NULL);
2379 
2380 	pio->io_flags = pio->io_orig_flags;
2381 	pio->io_stage = pio->io_orig_stage;
2382 	pio->io_pipeline = pio->io_orig_pipeline;
2383 	pio->io_reexecute = 0;
2384 	pio->io_flags |= ZIO_FLAG_REEXECUTED;
2385 	pio->io_pipeline_trace = 0;
2386 	pio->io_error = 0;
2387 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2388 		pio->io_state[w] = 0;
2389 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2390 		pio->io_child_error[c] = 0;
2391 
2392 	if (IO_IS_ALLOCATING(pio))
2393 		BP_ZERO(pio->io_bp);
2394 
2395 	/*
2396 	 * As we reexecute pio's children, new children could be created.
2397 	 * New children go to the head of pio's io_child_list, however,
2398 	 * so we will (correctly) not reexecute them.  The key is that
2399 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
2400 	 * cannot be affected by any side effects of reexecuting 'cio'.
2401 	 */
2402 	zio_link_t *zl = NULL;
2403 	mutex_enter(&pio->io_lock);
2404 	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2405 		cio_next = zio_walk_children(pio, &zl);
2406 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2407 			pio->io_children[cio->io_child_type][w]++;
2408 		mutex_exit(&pio->io_lock);
2409 		zio_reexecute(cio);
2410 		mutex_enter(&pio->io_lock);
2411 	}
2412 	mutex_exit(&pio->io_lock);
2413 
2414 	/*
2415 	 * Now that all children have been reexecuted, execute the parent.
2416 	 * We don't reexecute "The Godfather" I/O here as it's the
2417 	 * responsibility of the caller to wait on it.
2418 	 */
2419 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
2420 		pio->io_queued_timestamp = gethrtime();
2421 		__zio_execute(pio);
2422 	}
2423 }
2424 
2425 void
2426 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
2427 {
2428 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
2429 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2430 		    "failure and the failure mode property for this pool "
2431 		    "is set to panic.", spa_name(spa));
2432 
2433 	cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
2434 	    "failure and has been suspended.\n", spa_name(spa));
2435 
2436 	(void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
2437 	    NULL, NULL, 0);
2438 
2439 	mutex_enter(&spa->spa_suspend_lock);
2440 
2441 	if (spa->spa_suspend_zio_root == NULL)
2442 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
2443 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2444 		    ZIO_FLAG_GODFATHER);
2445 
2446 	spa->spa_suspended = reason;
2447 
2448 	if (zio != NULL) {
2449 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2450 		ASSERT(zio != spa->spa_suspend_zio_root);
2451 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2452 		ASSERT(zio_unique_parent(zio) == NULL);
2453 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
2454 		zio_add_child(spa->spa_suspend_zio_root, zio);
2455 	}
2456 
2457 	mutex_exit(&spa->spa_suspend_lock);
2458 }
2459 
2460 int
2461 zio_resume(spa_t *spa)
2462 {
2463 	zio_t *pio;
2464 
2465 	/*
2466 	 * Reexecute all previously suspended i/o.
2467 	 */
2468 	mutex_enter(&spa->spa_suspend_lock);
2469 	spa->spa_suspended = ZIO_SUSPEND_NONE;
2470 	cv_broadcast(&spa->spa_suspend_cv);
2471 	pio = spa->spa_suspend_zio_root;
2472 	spa->spa_suspend_zio_root = NULL;
2473 	mutex_exit(&spa->spa_suspend_lock);
2474 
2475 	if (pio == NULL)
2476 		return (0);
2477 
2478 	zio_reexecute(pio);
2479 	return (zio_wait(pio));
2480 }
2481 
2482 void
2483 zio_resume_wait(spa_t *spa)
2484 {
2485 	mutex_enter(&spa->spa_suspend_lock);
2486 	while (spa_suspended(spa))
2487 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
2488 	mutex_exit(&spa->spa_suspend_lock);
2489 }
2490 
2491 /*
2492  * ==========================================================================
2493  * Gang blocks.
2494  *
2495  * A gang block is a collection of small blocks that looks to the DMU
2496  * like one large block.  When zio_dva_allocate() cannot find a block
2497  * of the requested size, due to either severe fragmentation or the pool
2498  * being nearly full, it calls zio_write_gang_block() to construct the
2499  * block from smaller fragments.
2500  *
2501  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2502  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
2503  * an indirect block: it's an array of block pointers.  It consumes
2504  * only one sector and hence is allocatable regardless of fragmentation.
2505  * The gang header's bps point to its gang members, which hold the data.
2506  *
2507  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2508  * as the verifier to ensure uniqueness of the SHA256 checksum.
2509  * Critically, the gang block bp's blk_cksum is the checksum of the data,
2510  * not the gang header.  This ensures that data block signatures (needed for
2511  * deduplication) are independent of how the block is physically stored.
2512  *
2513  * Gang blocks can be nested: a gang member may itself be a gang block.
2514  * Thus every gang block is a tree in which root and all interior nodes are
2515  * gang headers, and the leaves are normal blocks that contain user data.
2516  * The root of the gang tree is called the gang leader.
2517  *
2518  * To perform any operation (read, rewrite, free, claim) on a gang block,
2519  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2520  * in the io_gang_tree field of the original logical i/o by recursively
2521  * reading the gang leader and all gang headers below it.  This yields
2522  * an in-core tree containing the contents of every gang header and the
2523  * bps for every constituent of the gang block.
2524  *
2525  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2526  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
2527  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2528  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2529  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2530  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
2531  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2532  * of the gang header plus zio_checksum_compute() of the data to update the
2533  * gang header's blk_cksum as described above.
2534  *
2535  * The two-phase assemble/issue model solves the problem of partial failure --
2536  * what if you'd freed part of a gang block but then couldn't read the
2537  * gang header for another part?  Assembling the entire gang tree first
2538  * ensures that all the necessary gang header I/O has succeeded before
2539  * starting the actual work of free, claim, or write.  Once the gang tree
2540  * is assembled, free and claim are in-memory operations that cannot fail.
2541  *
2542  * In the event that a gang write fails, zio_dva_unallocate() walks the
2543  * gang tree to immediately free (i.e. insert back into the space map)
2544  * everything we've allocated.  This ensures that we don't get ENOSPC
2545  * errors during repeated suspend/resume cycles due to a flaky device.
2546  *
2547  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
2548  * the gang tree, we won't modify the block, so we can safely defer the free
2549  * (knowing that the block is still intact).  If we *can* assemble the gang
2550  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2551  * each constituent bp and we can allocate a new block on the next sync pass.
2552  *
2553  * In all cases, the gang tree allows complete recovery from partial failure.
2554  * ==========================================================================
2555  */
2556 
2557 static void
2558 zio_gang_issue_func_done(zio_t *zio)
2559 {
2560 	abd_free(zio->io_abd);
2561 }
2562 
2563 static zio_t *
2564 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2565     uint64_t offset)
2566 {
2567 	if (gn != NULL)
2568 		return (pio);
2569 
2570 	return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2571 	    BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2572 	    NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2573 	    &pio->io_bookmark));
2574 }
2575 
2576 static zio_t *
2577 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2578     uint64_t offset)
2579 {
2580 	zio_t *zio;
2581 
2582 	if (gn != NULL) {
2583 		abd_t *gbh_abd =
2584 		    abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2585 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2586 		    gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2587 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2588 		    &pio->io_bookmark);
2589 		/*
2590 		 * As we rewrite each gang header, the pipeline will compute
2591 		 * a new gang block header checksum for it; but no one will
2592 		 * compute a new data checksum, so we do that here.  The one
2593 		 * exception is the gang leader: the pipeline already computed
2594 		 * its data checksum because that stage precedes gang assembly.
2595 		 * (Presently, nothing actually uses interior data checksums;
2596 		 * this is just good hygiene.)
2597 		 */
2598 		if (gn != pio->io_gang_leader->io_gang_tree) {
2599 			abd_t *buf = abd_get_offset(data, offset);
2600 
2601 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2602 			    buf, BP_GET_PSIZE(bp));
2603 
2604 			abd_free(buf);
2605 		}
2606 		/*
2607 		 * If we are here to damage data for testing purposes,
2608 		 * leave the GBH alone so that we can detect the damage.
2609 		 */
2610 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2611 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2612 	} else {
2613 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2614 		    abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2615 		    zio_gang_issue_func_done, NULL, pio->io_priority,
2616 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2617 	}
2618 
2619 	return (zio);
2620 }
2621 
2622 static zio_t *
2623 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2624     uint64_t offset)
2625 {
2626 	(void) gn, (void) data, (void) offset;
2627 
2628 	zio_t *zio = zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2629 	    ZIO_GANG_CHILD_FLAGS(pio));
2630 	if (zio == NULL) {
2631 		zio = zio_null(pio, pio->io_spa,
2632 		    NULL, NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio));
2633 	}
2634 	return (zio);
2635 }
2636 
2637 static zio_t *
2638 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2639     uint64_t offset)
2640 {
2641 	(void) gn, (void) data, (void) offset;
2642 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2643 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2644 }
2645 
2646 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2647 	NULL,
2648 	zio_read_gang,
2649 	zio_rewrite_gang,
2650 	zio_free_gang,
2651 	zio_claim_gang,
2652 	NULL
2653 };
2654 
2655 static void zio_gang_tree_assemble_done(zio_t *zio);
2656 
2657 static zio_gang_node_t *
2658 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2659 {
2660 	zio_gang_node_t *gn;
2661 
2662 	ASSERT(*gnpp == NULL);
2663 
2664 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2665 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2666 	*gnpp = gn;
2667 
2668 	return (gn);
2669 }
2670 
2671 static void
2672 zio_gang_node_free(zio_gang_node_t **gnpp)
2673 {
2674 	zio_gang_node_t *gn = *gnpp;
2675 
2676 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2677 		ASSERT(gn->gn_child[g] == NULL);
2678 
2679 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2680 	kmem_free(gn, sizeof (*gn));
2681 	*gnpp = NULL;
2682 }
2683 
2684 static void
2685 zio_gang_tree_free(zio_gang_node_t **gnpp)
2686 {
2687 	zio_gang_node_t *gn = *gnpp;
2688 
2689 	if (gn == NULL)
2690 		return;
2691 
2692 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2693 		zio_gang_tree_free(&gn->gn_child[g]);
2694 
2695 	zio_gang_node_free(gnpp);
2696 }
2697 
2698 static void
2699 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2700 {
2701 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2702 	abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2703 
2704 	ASSERT(gio->io_gang_leader == gio);
2705 	ASSERT(BP_IS_GANG(bp));
2706 
2707 	zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2708 	    zio_gang_tree_assemble_done, gn, gio->io_priority,
2709 	    ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2710 }
2711 
2712 static void
2713 zio_gang_tree_assemble_done(zio_t *zio)
2714 {
2715 	zio_t *gio = zio->io_gang_leader;
2716 	zio_gang_node_t *gn = zio->io_private;
2717 	blkptr_t *bp = zio->io_bp;
2718 
2719 	ASSERT(gio == zio_unique_parent(zio));
2720 	ASSERT(zio->io_child_count == 0);
2721 
2722 	if (zio->io_error)
2723 		return;
2724 
2725 	/* this ABD was created from a linear buf in zio_gang_tree_assemble */
2726 	if (BP_SHOULD_BYTESWAP(bp))
2727 		byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2728 
2729 	ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2730 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2731 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2732 
2733 	abd_free(zio->io_abd);
2734 
2735 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2736 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2737 		if (!BP_IS_GANG(gbp))
2738 			continue;
2739 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2740 	}
2741 }
2742 
2743 static void
2744 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2745     uint64_t offset)
2746 {
2747 	zio_t *gio = pio->io_gang_leader;
2748 	zio_t *zio;
2749 
2750 	ASSERT(BP_IS_GANG(bp) == !!gn);
2751 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2752 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2753 
2754 	/*
2755 	 * If you're a gang header, your data is in gn->gn_gbh.
2756 	 * If you're a gang member, your data is in 'data' and gn == NULL.
2757 	 */
2758 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2759 
2760 	if (gn != NULL) {
2761 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2762 
2763 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2764 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2765 			if (BP_IS_HOLE(gbp))
2766 				continue;
2767 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2768 			    offset);
2769 			offset += BP_GET_PSIZE(gbp);
2770 		}
2771 	}
2772 
2773 	if (gn == gio->io_gang_tree)
2774 		ASSERT3U(gio->io_size, ==, offset);
2775 
2776 	if (zio != pio)
2777 		zio_nowait(zio);
2778 }
2779 
2780 static zio_t *
2781 zio_gang_assemble(zio_t *zio)
2782 {
2783 	blkptr_t *bp = zio->io_bp;
2784 
2785 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2786 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2787 
2788 	zio->io_gang_leader = zio;
2789 
2790 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2791 
2792 	return (zio);
2793 }
2794 
2795 static zio_t *
2796 zio_gang_issue(zio_t *zio)
2797 {
2798 	blkptr_t *bp = zio->io_bp;
2799 
2800 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2801 		return (NULL);
2802 	}
2803 
2804 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2805 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2806 
2807 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2808 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2809 		    0);
2810 	else
2811 		zio_gang_tree_free(&zio->io_gang_tree);
2812 
2813 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2814 
2815 	return (zio);
2816 }
2817 
2818 static void
2819 zio_write_gang_member_ready(zio_t *zio)
2820 {
2821 	zio_t *pio = zio_unique_parent(zio);
2822 	dva_t *cdva = zio->io_bp->blk_dva;
2823 	dva_t *pdva = pio->io_bp->blk_dva;
2824 	uint64_t asize;
2825 	zio_t *gio __maybe_unused = zio->io_gang_leader;
2826 
2827 	if (BP_IS_HOLE(zio->io_bp))
2828 		return;
2829 
2830 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2831 
2832 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2833 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2834 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2835 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2836 	VERIFY3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2837 
2838 	mutex_enter(&pio->io_lock);
2839 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2840 		ASSERT(DVA_GET_GANG(&pdva[d]));
2841 		asize = DVA_GET_ASIZE(&pdva[d]);
2842 		asize += DVA_GET_ASIZE(&cdva[d]);
2843 		DVA_SET_ASIZE(&pdva[d], asize);
2844 	}
2845 	mutex_exit(&pio->io_lock);
2846 }
2847 
2848 static void
2849 zio_write_gang_done(zio_t *zio)
2850 {
2851 	/*
2852 	 * The io_abd field will be NULL for a zio with no data.  The io_flags
2853 	 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2854 	 * check for it here as it is cleared in zio_ready.
2855 	 */
2856 	if (zio->io_abd != NULL)
2857 		abd_free(zio->io_abd);
2858 }
2859 
2860 static zio_t *
2861 zio_write_gang_block(zio_t *pio, metaslab_class_t *mc)
2862 {
2863 	spa_t *spa = pio->io_spa;
2864 	blkptr_t *bp = pio->io_bp;
2865 	zio_t *gio = pio->io_gang_leader;
2866 	zio_t *zio;
2867 	zio_gang_node_t *gn, **gnpp;
2868 	zio_gbh_phys_t *gbh;
2869 	abd_t *gbh_abd;
2870 	uint64_t txg = pio->io_txg;
2871 	uint64_t resid = pio->io_size;
2872 	uint64_t lsize;
2873 	int copies = gio->io_prop.zp_copies;
2874 	zio_prop_t zp;
2875 	int error;
2876 	boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
2877 
2878 	/*
2879 	 * If one copy was requested, store 2 copies of the GBH, so that we
2880 	 * can still traverse all the data (e.g. to free or scrub) even if a
2881 	 * block is damaged.  Note that we can't store 3 copies of the GBH in
2882 	 * all cases, e.g. with encryption, which uses DVA[2] for the IV+salt.
2883 	 */
2884 	int gbh_copies = copies;
2885 	if (gbh_copies == 1) {
2886 		gbh_copies = MIN(2, spa_max_replication(spa));
2887 	}
2888 
2889 	int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2890 	if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2891 		ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2892 		ASSERT(has_data);
2893 
2894 		flags |= METASLAB_ASYNC_ALLOC;
2895 		VERIFY(zfs_refcount_held(&mc->mc_allocator[pio->io_allocator].
2896 		    mca_alloc_slots, pio));
2897 
2898 		/*
2899 		 * The logical zio has already placed a reservation for
2900 		 * 'copies' allocation slots but gang blocks may require
2901 		 * additional copies. These additional copies
2902 		 * (i.e. gbh_copies - copies) are guaranteed to succeed
2903 		 * since metaslab_class_throttle_reserve() always allows
2904 		 * additional reservations for gang blocks.
2905 		 */
2906 		VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2907 		    pio->io_allocator, pio, flags));
2908 	}
2909 
2910 	error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2911 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2912 	    &pio->io_alloc_list, pio, pio->io_allocator);
2913 	if (error) {
2914 		if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2915 			ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2916 			ASSERT(has_data);
2917 
2918 			/*
2919 			 * If we failed to allocate the gang block header then
2920 			 * we remove any additional allocation reservations that
2921 			 * we placed here. The original reservation will
2922 			 * be removed when the logical I/O goes to the ready
2923 			 * stage.
2924 			 */
2925 			metaslab_class_throttle_unreserve(mc,
2926 			    gbh_copies - copies, pio->io_allocator, pio);
2927 		}
2928 
2929 		pio->io_error = error;
2930 		return (pio);
2931 	}
2932 
2933 	if (pio == gio) {
2934 		gnpp = &gio->io_gang_tree;
2935 	} else {
2936 		gnpp = pio->io_private;
2937 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
2938 	}
2939 
2940 	gn = zio_gang_node_alloc(gnpp);
2941 	gbh = gn->gn_gbh;
2942 	memset(gbh, 0, SPA_GANGBLOCKSIZE);
2943 	gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2944 
2945 	/*
2946 	 * Create the gang header.
2947 	 */
2948 	zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2949 	    zio_write_gang_done, NULL, pio->io_priority,
2950 	    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2951 
2952 	/*
2953 	 * Create and nowait the gang children.
2954 	 */
2955 	for (int g = 0; resid != 0; resid -= lsize, g++) {
2956 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2957 		    SPA_MINBLOCKSIZE);
2958 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2959 
2960 		zp.zp_checksum = gio->io_prop.zp_checksum;
2961 		zp.zp_compress = ZIO_COMPRESS_OFF;
2962 		zp.zp_complevel = gio->io_prop.zp_complevel;
2963 		zp.zp_type = DMU_OT_NONE;
2964 		zp.zp_level = 0;
2965 		zp.zp_copies = gio->io_prop.zp_copies;
2966 		zp.zp_dedup = B_FALSE;
2967 		zp.zp_dedup_verify = B_FALSE;
2968 		zp.zp_nopwrite = B_FALSE;
2969 		zp.zp_encrypt = gio->io_prop.zp_encrypt;
2970 		zp.zp_byteorder = gio->io_prop.zp_byteorder;
2971 		memset(zp.zp_salt, 0, ZIO_DATA_SALT_LEN);
2972 		memset(zp.zp_iv, 0, ZIO_DATA_IV_LEN);
2973 		memset(zp.zp_mac, 0, ZIO_DATA_MAC_LEN);
2974 
2975 		zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2976 		    has_data ? abd_get_offset(pio->io_abd, pio->io_size -
2977 		    resid) : NULL, lsize, lsize, &zp,
2978 		    zio_write_gang_member_ready, NULL, NULL,
2979 		    zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2980 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2981 
2982 		if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2983 			ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2984 			ASSERT(has_data);
2985 
2986 			/*
2987 			 * Gang children won't throttle but we should
2988 			 * account for their work, so reserve an allocation
2989 			 * slot for them here.
2990 			 */
2991 			VERIFY(metaslab_class_throttle_reserve(mc,
2992 			    zp.zp_copies, cio->io_allocator, cio, flags));
2993 		}
2994 		zio_nowait(cio);
2995 	}
2996 
2997 	/*
2998 	 * Set pio's pipeline to just wait for zio to finish.
2999 	 */
3000 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3001 
3002 	/*
3003 	 * We didn't allocate this bp, so make sure it doesn't get unmarked.
3004 	 */
3005 	pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
3006 
3007 	zio_nowait(zio);
3008 
3009 	return (pio);
3010 }
3011 
3012 /*
3013  * The zio_nop_write stage in the pipeline determines if allocating a
3014  * new bp is necessary.  The nopwrite feature can handle writes in
3015  * either syncing or open context (i.e. zil writes) and as a result is
3016  * mutually exclusive with dedup.
3017  *
3018  * By leveraging a cryptographically secure checksum, such as SHA256, we
3019  * can compare the checksums of the new data and the old to determine if
3020  * allocating a new block is required.  Note that our requirements for
3021  * cryptographic strength are fairly weak: there can't be any accidental
3022  * hash collisions, but we don't need to be secure against intentional
3023  * (malicious) collisions.  To trigger a nopwrite, you have to be able
3024  * to write the file to begin with, and triggering an incorrect (hash
3025  * collision) nopwrite is no worse than simply writing to the file.
3026  * That said, there are no known attacks against the checksum algorithms
3027  * used for nopwrite, assuming that the salt and the checksums
3028  * themselves remain secret.
3029  */
3030 static zio_t *
3031 zio_nop_write(zio_t *zio)
3032 {
3033 	blkptr_t *bp = zio->io_bp;
3034 	blkptr_t *bp_orig = &zio->io_bp_orig;
3035 	zio_prop_t *zp = &zio->io_prop;
3036 
3037 	ASSERT(BP_IS_HOLE(bp));
3038 	ASSERT(BP_GET_LEVEL(bp) == 0);
3039 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
3040 	ASSERT(zp->zp_nopwrite);
3041 	ASSERT(!zp->zp_dedup);
3042 	ASSERT(zio->io_bp_override == NULL);
3043 	ASSERT(IO_IS_ALLOCATING(zio));
3044 
3045 	/*
3046 	 * Check to see if the original bp and the new bp have matching
3047 	 * characteristics (i.e. same checksum, compression algorithms, etc).
3048 	 * If they don't then just continue with the pipeline which will
3049 	 * allocate a new bp.
3050 	 */
3051 	if (BP_IS_HOLE(bp_orig) ||
3052 	    !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
3053 	    ZCHECKSUM_FLAG_NOPWRITE) ||
3054 	    BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) ||
3055 	    BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
3056 	    BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
3057 	    BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
3058 	    zp->zp_copies != BP_GET_NDVAS(bp_orig))
3059 		return (zio);
3060 
3061 	/*
3062 	 * If the checksums match then reset the pipeline so that we
3063 	 * avoid allocating a new bp and issuing any I/O.
3064 	 */
3065 	if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
3066 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
3067 		    ZCHECKSUM_FLAG_NOPWRITE);
3068 		ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
3069 		ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
3070 		ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
3071 		ASSERT3U(bp->blk_prop, ==, bp_orig->blk_prop);
3072 
3073 		/*
3074 		 * If we're overwriting a block that is currently on an
3075 		 * indirect vdev, then ignore the nopwrite request and
3076 		 * allow a new block to be allocated on a concrete vdev.
3077 		 */
3078 		spa_config_enter(zio->io_spa, SCL_VDEV, FTAG, RW_READER);
3079 		for (int d = 0; d < BP_GET_NDVAS(bp_orig); d++) {
3080 			vdev_t *tvd = vdev_lookup_top(zio->io_spa,
3081 			    DVA_GET_VDEV(&bp_orig->blk_dva[d]));
3082 			if (tvd->vdev_ops == &vdev_indirect_ops) {
3083 				spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3084 				return (zio);
3085 			}
3086 		}
3087 		spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3088 
3089 		*bp = *bp_orig;
3090 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3091 		zio->io_flags |= ZIO_FLAG_NOPWRITE;
3092 	}
3093 
3094 	return (zio);
3095 }
3096 
3097 /*
3098  * ==========================================================================
3099  * Block Reference Table
3100  * ==========================================================================
3101  */
3102 static zio_t *
3103 zio_brt_free(zio_t *zio)
3104 {
3105 	blkptr_t *bp;
3106 
3107 	bp = zio->io_bp;
3108 
3109 	if (BP_GET_LEVEL(bp) > 0 ||
3110 	    BP_IS_METADATA(bp) ||
3111 	    !brt_maybe_exists(zio->io_spa, bp)) {
3112 		return (zio);
3113 	}
3114 
3115 	if (!brt_entry_decref(zio->io_spa, bp)) {
3116 		/*
3117 		 * This isn't the last reference, so we cannot free
3118 		 * the data yet.
3119 		 */
3120 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3121 	}
3122 
3123 	return (zio);
3124 }
3125 
3126 /*
3127  * ==========================================================================
3128  * Dedup
3129  * ==========================================================================
3130  */
3131 static void
3132 zio_ddt_child_read_done(zio_t *zio)
3133 {
3134 	blkptr_t *bp = zio->io_bp;
3135 	ddt_entry_t *dde = zio->io_private;
3136 	ddt_phys_t *ddp;
3137 	zio_t *pio = zio_unique_parent(zio);
3138 
3139 	mutex_enter(&pio->io_lock);
3140 	ddp = ddt_phys_select(dde, bp);
3141 	if (zio->io_error == 0)
3142 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
3143 
3144 	if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
3145 		dde->dde_repair_abd = zio->io_abd;
3146 	else
3147 		abd_free(zio->io_abd);
3148 	mutex_exit(&pio->io_lock);
3149 }
3150 
3151 static zio_t *
3152 zio_ddt_read_start(zio_t *zio)
3153 {
3154 	blkptr_t *bp = zio->io_bp;
3155 
3156 	ASSERT(BP_GET_DEDUP(bp));
3157 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3158 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3159 
3160 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
3161 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
3162 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
3163 		ddt_phys_t *ddp = dde->dde_phys;
3164 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
3165 		blkptr_t blk;
3166 
3167 		ASSERT(zio->io_vsd == NULL);
3168 		zio->io_vsd = dde;
3169 
3170 		if (ddp_self == NULL)
3171 			return (zio);
3172 
3173 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
3174 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
3175 				continue;
3176 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
3177 			    &blk);
3178 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
3179 			    abd_alloc_for_io(zio->io_size, B_TRUE),
3180 			    zio->io_size, zio_ddt_child_read_done, dde,
3181 			    zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
3182 			    ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
3183 		}
3184 		return (zio);
3185 	}
3186 
3187 	zio_nowait(zio_read(zio, zio->io_spa, bp,
3188 	    zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
3189 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
3190 
3191 	return (zio);
3192 }
3193 
3194 static zio_t *
3195 zio_ddt_read_done(zio_t *zio)
3196 {
3197 	blkptr_t *bp = zio->io_bp;
3198 
3199 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
3200 		return (NULL);
3201 	}
3202 
3203 	ASSERT(BP_GET_DEDUP(bp));
3204 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3205 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3206 
3207 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
3208 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
3209 		ddt_entry_t *dde = zio->io_vsd;
3210 		if (ddt == NULL) {
3211 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
3212 			return (zio);
3213 		}
3214 		if (dde == NULL) {
3215 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
3216 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
3217 			return (NULL);
3218 		}
3219 		if (dde->dde_repair_abd != NULL) {
3220 			abd_copy(zio->io_abd, dde->dde_repair_abd,
3221 			    zio->io_size);
3222 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
3223 		}
3224 		ddt_repair_done(ddt, dde);
3225 		zio->io_vsd = NULL;
3226 	}
3227 
3228 	ASSERT(zio->io_vsd == NULL);
3229 
3230 	return (zio);
3231 }
3232 
3233 static boolean_t
3234 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
3235 {
3236 	spa_t *spa = zio->io_spa;
3237 	boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW);
3238 
3239 	ASSERT(!(zio->io_bp_override && do_raw));
3240 
3241 	/*
3242 	 * Note: we compare the original data, not the transformed data,
3243 	 * because when zio->io_bp is an override bp, we will not have
3244 	 * pushed the I/O transforms.  That's an important optimization
3245 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
3246 	 * However, we should never get a raw, override zio so in these
3247 	 * cases we can compare the io_abd directly. This is useful because
3248 	 * it allows us to do dedup verification even if we don't have access
3249 	 * to the original data (for instance, if the encryption keys aren't
3250 	 * loaded).
3251 	 */
3252 
3253 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
3254 		zio_t *lio = dde->dde_lead_zio[p];
3255 
3256 		if (lio != NULL && do_raw) {
3257 			return (lio->io_size != zio->io_size ||
3258 			    abd_cmp(zio->io_abd, lio->io_abd) != 0);
3259 		} else if (lio != NULL) {
3260 			return (lio->io_orig_size != zio->io_orig_size ||
3261 			    abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0);
3262 		}
3263 	}
3264 
3265 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
3266 		ddt_phys_t *ddp = &dde->dde_phys[p];
3267 
3268 		if (ddp->ddp_phys_birth != 0 && do_raw) {
3269 			blkptr_t blk = *zio->io_bp;
3270 			uint64_t psize;
3271 			abd_t *tmpabd;
3272 			int error;
3273 
3274 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
3275 			psize = BP_GET_PSIZE(&blk);
3276 
3277 			if (psize != zio->io_size)
3278 				return (B_TRUE);
3279 
3280 			ddt_exit(ddt);
3281 
3282 			tmpabd = abd_alloc_for_io(psize, B_TRUE);
3283 
3284 			error = zio_wait(zio_read(NULL, spa, &blk, tmpabd,
3285 			    psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
3286 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
3287 			    ZIO_FLAG_RAW, &zio->io_bookmark));
3288 
3289 			if (error == 0) {
3290 				if (abd_cmp(tmpabd, zio->io_abd) != 0)
3291 					error = SET_ERROR(ENOENT);
3292 			}
3293 
3294 			abd_free(tmpabd);
3295 			ddt_enter(ddt);
3296 			return (error != 0);
3297 		} else if (ddp->ddp_phys_birth != 0) {
3298 			arc_buf_t *abuf = NULL;
3299 			arc_flags_t aflags = ARC_FLAG_WAIT;
3300 			blkptr_t blk = *zio->io_bp;
3301 			int error;
3302 
3303 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
3304 
3305 			if (BP_GET_LSIZE(&blk) != zio->io_orig_size)
3306 				return (B_TRUE);
3307 
3308 			ddt_exit(ddt);
3309 
3310 			error = arc_read(NULL, spa, &blk,
3311 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
3312 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
3313 			    &aflags, &zio->io_bookmark);
3314 
3315 			if (error == 0) {
3316 				if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
3317 				    zio->io_orig_size) != 0)
3318 					error = SET_ERROR(ENOENT);
3319 				arc_buf_destroy(abuf, &abuf);
3320 			}
3321 
3322 			ddt_enter(ddt);
3323 			return (error != 0);
3324 		}
3325 	}
3326 
3327 	return (B_FALSE);
3328 }
3329 
3330 static void
3331 zio_ddt_child_write_ready(zio_t *zio)
3332 {
3333 	int p = zio->io_prop.zp_copies;
3334 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3335 	ddt_entry_t *dde = zio->io_private;
3336 	ddt_phys_t *ddp = &dde->dde_phys[p];
3337 	zio_t *pio;
3338 
3339 	if (zio->io_error)
3340 		return;
3341 
3342 	ddt_enter(ddt);
3343 
3344 	ASSERT(dde->dde_lead_zio[p] == zio);
3345 
3346 	ddt_phys_fill(ddp, zio->io_bp);
3347 
3348 	zio_link_t *zl = NULL;
3349 	while ((pio = zio_walk_parents(zio, &zl)) != NULL)
3350 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
3351 
3352 	ddt_exit(ddt);
3353 }
3354 
3355 static void
3356 zio_ddt_child_write_done(zio_t *zio)
3357 {
3358 	int p = zio->io_prop.zp_copies;
3359 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3360 	ddt_entry_t *dde = zio->io_private;
3361 	ddt_phys_t *ddp = &dde->dde_phys[p];
3362 
3363 	ddt_enter(ddt);
3364 
3365 	ASSERT(ddp->ddp_refcnt == 0);
3366 	ASSERT(dde->dde_lead_zio[p] == zio);
3367 	dde->dde_lead_zio[p] = NULL;
3368 
3369 	if (zio->io_error == 0) {
3370 		zio_link_t *zl = NULL;
3371 		while (zio_walk_parents(zio, &zl) != NULL)
3372 			ddt_phys_addref(ddp);
3373 	} else {
3374 		ddt_phys_clear(ddp);
3375 	}
3376 
3377 	ddt_exit(ddt);
3378 }
3379 
3380 static zio_t *
3381 zio_ddt_write(zio_t *zio)
3382 {
3383 	spa_t *spa = zio->io_spa;
3384 	blkptr_t *bp = zio->io_bp;
3385 	uint64_t txg = zio->io_txg;
3386 	zio_prop_t *zp = &zio->io_prop;
3387 	int p = zp->zp_copies;
3388 	zio_t *cio = NULL;
3389 	ddt_t *ddt = ddt_select(spa, bp);
3390 	ddt_entry_t *dde;
3391 	ddt_phys_t *ddp;
3392 
3393 	ASSERT(BP_GET_DEDUP(bp));
3394 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
3395 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
3396 	ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
3397 
3398 	ddt_enter(ddt);
3399 	dde = ddt_lookup(ddt, bp, B_TRUE);
3400 	ddp = &dde->dde_phys[p];
3401 
3402 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
3403 		/*
3404 		 * If we're using a weak checksum, upgrade to a strong checksum
3405 		 * and try again.  If we're already using a strong checksum,
3406 		 * we can't resolve it, so just convert to an ordinary write.
3407 		 * (And automatically e-mail a paper to Nature?)
3408 		 */
3409 		if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
3410 		    ZCHECKSUM_FLAG_DEDUP)) {
3411 			zp->zp_checksum = spa_dedup_checksum(spa);
3412 			zio_pop_transforms(zio);
3413 			zio->io_stage = ZIO_STAGE_OPEN;
3414 			BP_ZERO(bp);
3415 		} else {
3416 			zp->zp_dedup = B_FALSE;
3417 			BP_SET_DEDUP(bp, B_FALSE);
3418 		}
3419 		ASSERT(!BP_GET_DEDUP(bp));
3420 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
3421 		ddt_exit(ddt);
3422 		return (zio);
3423 	}
3424 
3425 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
3426 		if (ddp->ddp_phys_birth != 0)
3427 			ddt_bp_fill(ddp, bp, txg);
3428 		if (dde->dde_lead_zio[p] != NULL)
3429 			zio_add_child(zio, dde->dde_lead_zio[p]);
3430 		else
3431 			ddt_phys_addref(ddp);
3432 	} else if (zio->io_bp_override) {
3433 		ASSERT(bp->blk_birth == txg);
3434 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
3435 		ddt_phys_fill(ddp, bp);
3436 		ddt_phys_addref(ddp);
3437 	} else {
3438 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3439 		    zio->io_orig_size, zio->io_orig_size, zp,
3440 		    zio_ddt_child_write_ready, NULL, NULL,
3441 		    zio_ddt_child_write_done, dde, zio->io_priority,
3442 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3443 
3444 		zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
3445 		dde->dde_lead_zio[p] = cio;
3446 	}
3447 
3448 	ddt_exit(ddt);
3449 
3450 	zio_nowait(cio);
3451 
3452 	return (zio);
3453 }
3454 
3455 static ddt_entry_t *freedde; /* for debugging */
3456 
3457 static zio_t *
3458 zio_ddt_free(zio_t *zio)
3459 {
3460 	spa_t *spa = zio->io_spa;
3461 	blkptr_t *bp = zio->io_bp;
3462 	ddt_t *ddt = ddt_select(spa, bp);
3463 	ddt_entry_t *dde;
3464 	ddt_phys_t *ddp;
3465 
3466 	ASSERT(BP_GET_DEDUP(bp));
3467 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3468 
3469 	ddt_enter(ddt);
3470 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
3471 	if (dde) {
3472 		ddp = ddt_phys_select(dde, bp);
3473 		if (ddp)
3474 			ddt_phys_decref(ddp);
3475 	}
3476 	ddt_exit(ddt);
3477 
3478 	return (zio);
3479 }
3480 
3481 /*
3482  * ==========================================================================
3483  * Allocate and free blocks
3484  * ==========================================================================
3485  */
3486 
3487 static zio_t *
3488 zio_io_to_allocate(spa_t *spa, int allocator)
3489 {
3490 	zio_t *zio;
3491 
3492 	ASSERT(MUTEX_HELD(&spa->spa_allocs[allocator].spaa_lock));
3493 
3494 	zio = avl_first(&spa->spa_allocs[allocator].spaa_tree);
3495 	if (zio == NULL)
3496 		return (NULL);
3497 
3498 	ASSERT(IO_IS_ALLOCATING(zio));
3499 
3500 	/*
3501 	 * Try to place a reservation for this zio. If we're unable to
3502 	 * reserve then we throttle.
3503 	 */
3504 	ASSERT3U(zio->io_allocator, ==, allocator);
3505 	if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
3506 	    zio->io_prop.zp_copies, allocator, zio, 0)) {
3507 		return (NULL);
3508 	}
3509 
3510 	avl_remove(&spa->spa_allocs[allocator].spaa_tree, zio);
3511 	ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
3512 
3513 	return (zio);
3514 }
3515 
3516 static zio_t *
3517 zio_dva_throttle(zio_t *zio)
3518 {
3519 	spa_t *spa = zio->io_spa;
3520 	zio_t *nio;
3521 	metaslab_class_t *mc;
3522 
3523 	/* locate an appropriate allocation class */
3524 	mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type,
3525 	    zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk);
3526 
3527 	if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
3528 	    !mc->mc_alloc_throttle_enabled ||
3529 	    zio->io_child_type == ZIO_CHILD_GANG ||
3530 	    zio->io_flags & ZIO_FLAG_NODATA) {
3531 		return (zio);
3532 	}
3533 
3534 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3535 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3536 	ASSERT3U(zio->io_queued_timestamp, >, 0);
3537 	ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3538 
3539 	zbookmark_phys_t *bm = &zio->io_bookmark;
3540 	/*
3541 	 * We want to try to use as many allocators as possible to help improve
3542 	 * performance, but we also want logically adjacent IOs to be physically
3543 	 * adjacent to improve sequential read performance. We chunk each object
3544 	 * into 2^20 block regions, and then hash based on the objset, object,
3545 	 * level, and region to accomplish both of these goals.
3546 	 */
3547 	int allocator = (uint_t)cityhash4(bm->zb_objset, bm->zb_object,
3548 	    bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
3549 	zio->io_allocator = allocator;
3550 	zio->io_metaslab_class = mc;
3551 	mutex_enter(&spa->spa_allocs[allocator].spaa_lock);
3552 	avl_add(&spa->spa_allocs[allocator].spaa_tree, zio);
3553 	nio = zio_io_to_allocate(spa, allocator);
3554 	mutex_exit(&spa->spa_allocs[allocator].spaa_lock);
3555 	return (nio);
3556 }
3557 
3558 static void
3559 zio_allocate_dispatch(spa_t *spa, int allocator)
3560 {
3561 	zio_t *zio;
3562 
3563 	mutex_enter(&spa->spa_allocs[allocator].spaa_lock);
3564 	zio = zio_io_to_allocate(spa, allocator);
3565 	mutex_exit(&spa->spa_allocs[allocator].spaa_lock);
3566 	if (zio == NULL)
3567 		return;
3568 
3569 	ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3570 	ASSERT0(zio->io_error);
3571 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3572 }
3573 
3574 static zio_t *
3575 zio_dva_allocate(zio_t *zio)
3576 {
3577 	spa_t *spa = zio->io_spa;
3578 	metaslab_class_t *mc;
3579 	blkptr_t *bp = zio->io_bp;
3580 	int error;
3581 	int flags = 0;
3582 
3583 	if (zio->io_gang_leader == NULL) {
3584 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3585 		zio->io_gang_leader = zio;
3586 	}
3587 
3588 	ASSERT(BP_IS_HOLE(bp));
3589 	ASSERT0(BP_GET_NDVAS(bp));
3590 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
3591 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3592 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3593 
3594 	flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
3595 	if (zio->io_flags & ZIO_FLAG_NODATA)
3596 		flags |= METASLAB_DONT_THROTTLE;
3597 	if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
3598 		flags |= METASLAB_GANG_CHILD;
3599 	if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
3600 		flags |= METASLAB_ASYNC_ALLOC;
3601 
3602 	/*
3603 	 * if not already chosen, locate an appropriate allocation class
3604 	 */
3605 	mc = zio->io_metaslab_class;
3606 	if (mc == NULL) {
3607 		mc = spa_preferred_class(spa, zio->io_size,
3608 		    zio->io_prop.zp_type, zio->io_prop.zp_level,
3609 		    zio->io_prop.zp_zpl_smallblk);
3610 		zio->io_metaslab_class = mc;
3611 	}
3612 
3613 	/*
3614 	 * Try allocating the block in the usual metaslab class.
3615 	 * If that's full, allocate it in the normal class.
3616 	 * If that's full, allocate as a gang block,
3617 	 * and if all are full, the allocation fails (which shouldn't happen).
3618 	 *
3619 	 * Note that we do not fall back on embedded slog (ZIL) space, to
3620 	 * preserve unfragmented slog space, which is critical for decent
3621 	 * sync write performance.  If a log allocation fails, we will fall
3622 	 * back to spa_sync() which is abysmal for performance.
3623 	 */
3624 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
3625 	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3626 	    &zio->io_alloc_list, zio, zio->io_allocator);
3627 
3628 	/*
3629 	 * Fallback to normal class when an alloc class is full
3630 	 */
3631 	if (error == ENOSPC && mc != spa_normal_class(spa)) {
3632 		/*
3633 		 * If throttling, transfer reservation over to normal class.
3634 		 * The io_allocator slot can remain the same even though we
3635 		 * are switching classes.
3636 		 */
3637 		if (mc->mc_alloc_throttle_enabled &&
3638 		    (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) {
3639 			metaslab_class_throttle_unreserve(mc,
3640 			    zio->io_prop.zp_copies, zio->io_allocator, zio);
3641 			zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING;
3642 
3643 			VERIFY(metaslab_class_throttle_reserve(
3644 			    spa_normal_class(spa),
3645 			    zio->io_prop.zp_copies, zio->io_allocator, zio,
3646 			    flags | METASLAB_MUST_RESERVE));
3647 		}
3648 		zio->io_metaslab_class = mc = spa_normal_class(spa);
3649 		if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) {
3650 			zfs_dbgmsg("%s: metaslab allocation failure, "
3651 			    "trying normal class: zio %px, size %llu, error %d",
3652 			    spa_name(spa), zio, (u_longlong_t)zio->io_size,
3653 			    error);
3654 		}
3655 
3656 		error = metaslab_alloc(spa, mc, zio->io_size, bp,
3657 		    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3658 		    &zio->io_alloc_list, zio, zio->io_allocator);
3659 	}
3660 
3661 	if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) {
3662 		if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) {
3663 			zfs_dbgmsg("%s: metaslab allocation failure, "
3664 			    "trying ganging: zio %px, size %llu, error %d",
3665 			    spa_name(spa), zio, (u_longlong_t)zio->io_size,
3666 			    error);
3667 		}
3668 		return (zio_write_gang_block(zio, mc));
3669 	}
3670 	if (error != 0) {
3671 		if (error != ENOSPC ||
3672 		    (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC)) {
3673 			zfs_dbgmsg("%s: metaslab allocation failure: zio %px, "
3674 			    "size %llu, error %d",
3675 			    spa_name(spa), zio, (u_longlong_t)zio->io_size,
3676 			    error);
3677 		}
3678 		zio->io_error = error;
3679 	}
3680 
3681 	return (zio);
3682 }
3683 
3684 static zio_t *
3685 zio_dva_free(zio_t *zio)
3686 {
3687 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3688 
3689 	return (zio);
3690 }
3691 
3692 static zio_t *
3693 zio_dva_claim(zio_t *zio)
3694 {
3695 	int error;
3696 
3697 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3698 	if (error)
3699 		zio->io_error = error;
3700 
3701 	return (zio);
3702 }
3703 
3704 /*
3705  * Undo an allocation.  This is used by zio_done() when an I/O fails
3706  * and we want to give back the block we just allocated.
3707  * This handles both normal blocks and gang blocks.
3708  */
3709 static void
3710 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3711 {
3712 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3713 	ASSERT(zio->io_bp_override == NULL);
3714 
3715 	if (!BP_IS_HOLE(bp))
3716 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3717 
3718 	if (gn != NULL) {
3719 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3720 			zio_dva_unallocate(zio, gn->gn_child[g],
3721 			    &gn->gn_gbh->zg_blkptr[g]);
3722 		}
3723 	}
3724 }
3725 
3726 /*
3727  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
3728  */
3729 int
3730 zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
3731     uint64_t size, boolean_t *slog)
3732 {
3733 	int error = 1;
3734 	zio_alloc_list_t io_alloc_list;
3735 
3736 	ASSERT(txg > spa_syncing_txg(spa));
3737 
3738 	metaslab_trace_init(&io_alloc_list);
3739 
3740 	/*
3741 	 * Block pointer fields are useful to metaslabs for stats and debugging.
3742 	 * Fill in the obvious ones before calling into metaslab_alloc().
3743 	 */
3744 	BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3745 	BP_SET_PSIZE(new_bp, size);
3746 	BP_SET_LEVEL(new_bp, 0);
3747 
3748 	/*
3749 	 * When allocating a zil block, we don't have information about
3750 	 * the final destination of the block except the objset it's part
3751 	 * of, so we just hash the objset ID to pick the allocator to get
3752 	 * some parallelism.
3753 	 */
3754 	int flags = METASLAB_FASTWRITE | METASLAB_ZIL;
3755 	int allocator = (uint_t)cityhash4(0, 0, 0,
3756 	    os->os_dsl_dataset->ds_object) % spa->spa_alloc_count;
3757 	error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3758 	    txg, NULL, flags, &io_alloc_list, NULL, allocator);
3759 	*slog = (error == 0);
3760 	if (error != 0) {
3761 		error = metaslab_alloc(spa, spa_embedded_log_class(spa), size,
3762 		    new_bp, 1, txg, NULL, flags,
3763 		    &io_alloc_list, NULL, allocator);
3764 	}
3765 	if (error != 0) {
3766 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
3767 		    new_bp, 1, txg, NULL, flags,
3768 		    &io_alloc_list, NULL, allocator);
3769 	}
3770 	metaslab_trace_fini(&io_alloc_list);
3771 
3772 	if (error == 0) {
3773 		BP_SET_LSIZE(new_bp, size);
3774 		BP_SET_PSIZE(new_bp, size);
3775 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3776 		BP_SET_CHECKSUM(new_bp,
3777 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3778 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3779 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3780 		BP_SET_LEVEL(new_bp, 0);
3781 		BP_SET_DEDUP(new_bp, 0);
3782 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3783 
3784 		/*
3785 		 * encrypted blocks will require an IV and salt. We generate
3786 		 * these now since we will not be rewriting the bp at
3787 		 * rewrite time.
3788 		 */
3789 		if (os->os_encrypted) {
3790 			uint8_t iv[ZIO_DATA_IV_LEN];
3791 			uint8_t salt[ZIO_DATA_SALT_LEN];
3792 
3793 			BP_SET_CRYPT(new_bp, B_TRUE);
3794 			VERIFY0(spa_crypt_get_salt(spa,
3795 			    dmu_objset_id(os), salt));
3796 			VERIFY0(zio_crypt_generate_iv(iv));
3797 
3798 			zio_crypt_encode_params_bp(new_bp, salt, iv);
3799 		}
3800 	} else {
3801 		zfs_dbgmsg("%s: zil block allocation failure: "
3802 		    "size %llu, error %d", spa_name(spa), (u_longlong_t)size,
3803 		    error);
3804 	}
3805 
3806 	return (error);
3807 }
3808 
3809 /*
3810  * ==========================================================================
3811  * Read and write to physical devices
3812  * ==========================================================================
3813  */
3814 
3815 /*
3816  * Issue an I/O to the underlying vdev. Typically the issue pipeline
3817  * stops after this stage and will resume upon I/O completion.
3818  * However, there are instances where the vdev layer may need to
3819  * continue the pipeline when an I/O was not issued. Since the I/O
3820  * that was sent to the vdev layer might be different than the one
3821  * currently active in the pipeline (see vdev_queue_io()), we explicitly
3822  * force the underlying vdev layers to call either zio_execute() or
3823  * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3824  */
3825 static zio_t *
3826 zio_vdev_io_start(zio_t *zio)
3827 {
3828 	vdev_t *vd = zio->io_vd;
3829 	uint64_t align;
3830 	spa_t *spa = zio->io_spa;
3831 
3832 	zio->io_delay = 0;
3833 
3834 	ASSERT(zio->io_error == 0);
3835 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3836 
3837 	if (vd == NULL) {
3838 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3839 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3840 
3841 		/*
3842 		 * The mirror_ops handle multiple DVAs in a single BP.
3843 		 */
3844 		vdev_mirror_ops.vdev_op_io_start(zio);
3845 		return (NULL);
3846 	}
3847 
3848 	ASSERT3P(zio->io_logical, !=, zio);
3849 	if (zio->io_type == ZIO_TYPE_WRITE) {
3850 		ASSERT(spa->spa_trust_config);
3851 
3852 		/*
3853 		 * Note: the code can handle other kinds of writes,
3854 		 * but we don't expect them.
3855 		 */
3856 		if (zio->io_vd->vdev_noalloc) {
3857 			ASSERT(zio->io_flags &
3858 			    (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3859 			    ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3860 		}
3861 	}
3862 
3863 	align = 1ULL << vd->vdev_top->vdev_ashift;
3864 
3865 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3866 	    P2PHASE(zio->io_size, align) != 0) {
3867 		/* Transform logical writes to be a full physical block size. */
3868 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
3869 		abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3870 		ASSERT(vd == vd->vdev_top);
3871 		if (zio->io_type == ZIO_TYPE_WRITE) {
3872 			abd_copy(abuf, zio->io_abd, zio->io_size);
3873 			abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3874 		}
3875 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3876 	}
3877 
3878 	/*
3879 	 * If this is not a physical io, make sure that it is properly aligned
3880 	 * before proceeding.
3881 	 */
3882 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3883 		ASSERT0(P2PHASE(zio->io_offset, align));
3884 		ASSERT0(P2PHASE(zio->io_size, align));
3885 	} else {
3886 		/*
3887 		 * For physical writes, we allow 512b aligned writes and assume
3888 		 * the device will perform a read-modify-write as necessary.
3889 		 */
3890 		ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3891 		ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3892 	}
3893 
3894 	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
3895 
3896 	/*
3897 	 * If this is a repair I/O, and there's no self-healing involved --
3898 	 * that is, we're just resilvering what we expect to resilver --
3899 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3900 	 * This prevents spurious resilvering.
3901 	 *
3902 	 * There are a few ways that we can end up creating these spurious
3903 	 * resilver i/os:
3904 	 *
3905 	 * 1. A resilver i/o will be issued if any DVA in the BP has a
3906 	 * dirty DTL.  The mirror code will issue resilver writes to
3907 	 * each DVA, including the one(s) that are not on vdevs with dirty
3908 	 * DTLs.
3909 	 *
3910 	 * 2. With nested replication, which happens when we have a
3911 	 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3912 	 * For example, given mirror(replacing(A+B), C), it's likely that
3913 	 * only A is out of date (it's the new device). In this case, we'll
3914 	 * read from C, then use the data to resilver A+B -- but we don't
3915 	 * actually want to resilver B, just A. The top-level mirror has no
3916 	 * way to know this, so instead we just discard unnecessary repairs
3917 	 * as we work our way down the vdev tree.
3918 	 *
3919 	 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3920 	 * The same logic applies to any form of nested replication: ditto
3921 	 * + mirror, RAID-Z + replacing, etc.
3922 	 *
3923 	 * However, indirect vdevs point off to other vdevs which may have
3924 	 * DTL's, so we never bypass them.  The child i/os on concrete vdevs
3925 	 * will be properly bypassed instead.
3926 	 *
3927 	 * Leaf DTL_PARTIAL can be empty when a legitimate write comes from
3928 	 * a dRAID spare vdev. For example, when a dRAID spare is first
3929 	 * used, its spare blocks need to be written to but the leaf vdev's
3930 	 * of such blocks can have empty DTL_PARTIAL.
3931 	 *
3932 	 * There seemed no clean way to allow such writes while bypassing
3933 	 * spurious ones. At this point, just avoid all bypassing for dRAID
3934 	 * for correctness.
3935 	 */
3936 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3937 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3938 	    zio->io_txg != 0 &&	/* not a delegated i/o */
3939 	    vd->vdev_ops != &vdev_indirect_ops &&
3940 	    vd->vdev_top->vdev_ops != &vdev_draid_ops &&
3941 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3942 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3943 		zio_vdev_io_bypass(zio);
3944 		return (zio);
3945 	}
3946 
3947 	/*
3948 	 * Select the next best leaf I/O to process.  Distributed spares are
3949 	 * excluded since they dispatch the I/O directly to a leaf vdev after
3950 	 * applying the dRAID mapping.
3951 	 */
3952 	if (vd->vdev_ops->vdev_op_leaf &&
3953 	    vd->vdev_ops != &vdev_draid_spare_ops &&
3954 	    (zio->io_type == ZIO_TYPE_READ ||
3955 	    zio->io_type == ZIO_TYPE_WRITE ||
3956 	    zio->io_type == ZIO_TYPE_TRIM)) {
3957 
3958 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
3959 			return (zio);
3960 
3961 		if ((zio = vdev_queue_io(zio)) == NULL)
3962 			return (NULL);
3963 
3964 		if (!vdev_accessible(vd, zio)) {
3965 			zio->io_error = SET_ERROR(ENXIO);
3966 			zio_interrupt(zio);
3967 			return (NULL);
3968 		}
3969 		zio->io_delay = gethrtime();
3970 	}
3971 
3972 	vd->vdev_ops->vdev_op_io_start(zio);
3973 	return (NULL);
3974 }
3975 
3976 static zio_t *
3977 zio_vdev_io_done(zio_t *zio)
3978 {
3979 	vdev_t *vd = zio->io_vd;
3980 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3981 	boolean_t unexpected_error = B_FALSE;
3982 
3983 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3984 		return (NULL);
3985 	}
3986 
3987 	ASSERT(zio->io_type == ZIO_TYPE_READ ||
3988 	    zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM);
3989 
3990 	if (zio->io_delay)
3991 		zio->io_delay = gethrtime() - zio->io_delay;
3992 
3993 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3994 	    vd->vdev_ops != &vdev_draid_spare_ops) {
3995 		vdev_queue_io_done(zio);
3996 
3997 		if (zio->io_type == ZIO_TYPE_WRITE)
3998 			vdev_cache_write(zio);
3999 
4000 		if (zio_injection_enabled && zio->io_error == 0)
4001 			zio->io_error = zio_handle_device_injections(vd, zio,
4002 			    EIO, EILSEQ);
4003 
4004 		if (zio_injection_enabled && zio->io_error == 0)
4005 			zio->io_error = zio_handle_label_injection(zio, EIO);
4006 
4007 		if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) {
4008 			if (!vdev_accessible(vd, zio)) {
4009 				zio->io_error = SET_ERROR(ENXIO);
4010 			} else {
4011 				unexpected_error = B_TRUE;
4012 			}
4013 		}
4014 	}
4015 
4016 	ops->vdev_op_io_done(zio);
4017 
4018 	if (unexpected_error && vd->vdev_remove_wanted == B_FALSE)
4019 		VERIFY(vdev_probe(vd, zio) == NULL);
4020 
4021 	return (zio);
4022 }
4023 
4024 /*
4025  * This function is used to change the priority of an existing zio that is
4026  * currently in-flight. This is used by the arc to upgrade priority in the
4027  * event that a demand read is made for a block that is currently queued
4028  * as a scrub or async read IO. Otherwise, the high priority read request
4029  * would end up having to wait for the lower priority IO.
4030  */
4031 void
4032 zio_change_priority(zio_t *pio, zio_priority_t priority)
4033 {
4034 	zio_t *cio, *cio_next;
4035 	zio_link_t *zl = NULL;
4036 
4037 	ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
4038 
4039 	if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
4040 		vdev_queue_change_io_priority(pio, priority);
4041 	} else {
4042 		pio->io_priority = priority;
4043 	}
4044 
4045 	mutex_enter(&pio->io_lock);
4046 	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
4047 		cio_next = zio_walk_children(pio, &zl);
4048 		zio_change_priority(cio, priority);
4049 	}
4050 	mutex_exit(&pio->io_lock);
4051 }
4052 
4053 /*
4054  * For non-raidz ZIOs, we can just copy aside the bad data read from the
4055  * disk, and use that to finish the checksum ereport later.
4056  */
4057 static void
4058 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
4059     const abd_t *good_buf)
4060 {
4061 	/* no processing needed */
4062 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
4063 }
4064 
4065 void
4066 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr)
4067 {
4068 	void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size);
4069 
4070 	abd_copy(abd, zio->io_abd, zio->io_size);
4071 
4072 	zcr->zcr_cbinfo = zio->io_size;
4073 	zcr->zcr_cbdata = abd;
4074 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
4075 	zcr->zcr_free = zio_abd_free;
4076 }
4077 
4078 static zio_t *
4079 zio_vdev_io_assess(zio_t *zio)
4080 {
4081 	vdev_t *vd = zio->io_vd;
4082 
4083 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
4084 		return (NULL);
4085 	}
4086 
4087 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
4088 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
4089 
4090 	if (zio->io_vsd != NULL) {
4091 		zio->io_vsd_ops->vsd_free(zio);
4092 		zio->io_vsd = NULL;
4093 	}
4094 
4095 	if (zio_injection_enabled && zio->io_error == 0)
4096 		zio->io_error = zio_handle_fault_injection(zio, EIO);
4097 
4098 	/*
4099 	 * If the I/O failed, determine whether we should attempt to retry it.
4100 	 *
4101 	 * On retry, we cut in line in the issue queue, since we don't want
4102 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
4103 	 */
4104 	if (zio->io_error && vd == NULL &&
4105 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
4106 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
4107 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
4108 		zio->io_error = 0;
4109 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
4110 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
4111 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
4112 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
4113 		    zio_requeue_io_start_cut_in_line);
4114 		return (NULL);
4115 	}
4116 
4117 	/*
4118 	 * If we got an error on a leaf device, convert it to ENXIO
4119 	 * if the device is not accessible at all.
4120 	 */
4121 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
4122 	    !vdev_accessible(vd, zio))
4123 		zio->io_error = SET_ERROR(ENXIO);
4124 
4125 	/*
4126 	 * If we can't write to an interior vdev (mirror or RAID-Z),
4127 	 * set vdev_cant_write so that we stop trying to allocate from it.
4128 	 */
4129 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
4130 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
4131 		vdev_dbgmsg(vd, "zio_vdev_io_assess(zio=%px) setting "
4132 		    "cant_write=TRUE due to write failure with ENXIO",
4133 		    zio);
4134 		vd->vdev_cant_write = B_TRUE;
4135 	}
4136 
4137 	/*
4138 	 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
4139 	 * attempts will ever succeed. In this case we set a persistent
4140 	 * boolean flag so that we don't bother with it in the future.
4141 	 */
4142 	if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
4143 	    zio->io_type == ZIO_TYPE_IOCTL &&
4144 	    zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
4145 		vd->vdev_nowritecache = B_TRUE;
4146 
4147 	if (zio->io_error)
4148 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4149 
4150 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
4151 	    zio->io_physdone != NULL) {
4152 		ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
4153 		ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
4154 		zio->io_physdone(zio->io_logical);
4155 	}
4156 
4157 	return (zio);
4158 }
4159 
4160 void
4161 zio_vdev_io_reissue(zio_t *zio)
4162 {
4163 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4164 	ASSERT(zio->io_error == 0);
4165 
4166 	zio->io_stage >>= 1;
4167 }
4168 
4169 void
4170 zio_vdev_io_redone(zio_t *zio)
4171 {
4172 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
4173 
4174 	zio->io_stage >>= 1;
4175 }
4176 
4177 void
4178 zio_vdev_io_bypass(zio_t *zio)
4179 {
4180 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4181 	ASSERT(zio->io_error == 0);
4182 
4183 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
4184 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
4185 }
4186 
4187 /*
4188  * ==========================================================================
4189  * Encrypt and store encryption parameters
4190  * ==========================================================================
4191  */
4192 
4193 
4194 /*
4195  * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
4196  * managing the storage of encryption parameters and passing them to the
4197  * lower-level encryption functions.
4198  */
4199 static zio_t *
4200 zio_encrypt(zio_t *zio)
4201 {
4202 	zio_prop_t *zp = &zio->io_prop;
4203 	spa_t *spa = zio->io_spa;
4204 	blkptr_t *bp = zio->io_bp;
4205 	uint64_t psize = BP_GET_PSIZE(bp);
4206 	uint64_t dsobj = zio->io_bookmark.zb_objset;
4207 	dmu_object_type_t ot = BP_GET_TYPE(bp);
4208 	void *enc_buf = NULL;
4209 	abd_t *eabd = NULL;
4210 	uint8_t salt[ZIO_DATA_SALT_LEN];
4211 	uint8_t iv[ZIO_DATA_IV_LEN];
4212 	uint8_t mac[ZIO_DATA_MAC_LEN];
4213 	boolean_t no_crypt = B_FALSE;
4214 
4215 	/* the root zio already encrypted the data */
4216 	if (zio->io_child_type == ZIO_CHILD_GANG)
4217 		return (zio);
4218 
4219 	/* only ZIL blocks are re-encrypted on rewrite */
4220 	if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG)
4221 		return (zio);
4222 
4223 	if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) {
4224 		BP_SET_CRYPT(bp, B_FALSE);
4225 		return (zio);
4226 	}
4227 
4228 	/* if we are doing raw encryption set the provided encryption params */
4229 	if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) {
4230 		ASSERT0(BP_GET_LEVEL(bp));
4231 		BP_SET_CRYPT(bp, B_TRUE);
4232 		BP_SET_BYTEORDER(bp, zp->zp_byteorder);
4233 		if (ot != DMU_OT_OBJSET)
4234 			zio_crypt_encode_mac_bp(bp, zp->zp_mac);
4235 
4236 		/* dnode blocks must be written out in the provided byteorder */
4237 		if (zp->zp_byteorder != ZFS_HOST_BYTEORDER &&
4238 		    ot == DMU_OT_DNODE) {
4239 			void *bswap_buf = zio_buf_alloc(psize);
4240 			abd_t *babd = abd_get_from_buf(bswap_buf, psize);
4241 
4242 			ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4243 			abd_copy_to_buf(bswap_buf, zio->io_abd, psize);
4244 			dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf,
4245 			    psize);
4246 
4247 			abd_take_ownership_of_buf(babd, B_TRUE);
4248 			zio_push_transform(zio, babd, psize, psize, NULL);
4249 		}
4250 
4251 		if (DMU_OT_IS_ENCRYPTED(ot))
4252 			zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv);
4253 		return (zio);
4254 	}
4255 
4256 	/* indirect blocks only maintain a cksum of the lower level MACs */
4257 	if (BP_GET_LEVEL(bp) > 0) {
4258 		BP_SET_CRYPT(bp, B_TRUE);
4259 		VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE,
4260 		    zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp),
4261 		    mac));
4262 		zio_crypt_encode_mac_bp(bp, mac);
4263 		return (zio);
4264 	}
4265 
4266 	/*
4267 	 * Objset blocks are a special case since they have 2 256-bit MACs
4268 	 * embedded within them.
4269 	 */
4270 	if (ot == DMU_OT_OBJSET) {
4271 		ASSERT0(DMU_OT_IS_ENCRYPTED(ot));
4272 		ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4273 		BP_SET_CRYPT(bp, B_TRUE);
4274 		VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj,
4275 		    zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp)));
4276 		return (zio);
4277 	}
4278 
4279 	/* unencrypted object types are only authenticated with a MAC */
4280 	if (!DMU_OT_IS_ENCRYPTED(ot)) {
4281 		BP_SET_CRYPT(bp, B_TRUE);
4282 		VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj,
4283 		    zio->io_abd, psize, mac));
4284 		zio_crypt_encode_mac_bp(bp, mac);
4285 		return (zio);
4286 	}
4287 
4288 	/*
4289 	 * Later passes of sync-to-convergence may decide to rewrite data
4290 	 * in place to avoid more disk reallocations. This presents a problem
4291 	 * for encryption because this constitutes rewriting the new data with
4292 	 * the same encryption key and IV. However, this only applies to blocks
4293 	 * in the MOS (particularly the spacemaps) and we do not encrypt the
4294 	 * MOS. We assert that the zio is allocating or an intent log write
4295 	 * to enforce this.
4296 	 */
4297 	ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG);
4298 	ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG);
4299 	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION));
4300 	ASSERT3U(psize, !=, 0);
4301 
4302 	enc_buf = zio_buf_alloc(psize);
4303 	eabd = abd_get_from_buf(enc_buf, psize);
4304 	abd_take_ownership_of_buf(eabd, B_TRUE);
4305 
4306 	/*
4307 	 * For an explanation of what encryption parameters are stored
4308 	 * where, see the block comment in zio_crypt.c.
4309 	 */
4310 	if (ot == DMU_OT_INTENT_LOG) {
4311 		zio_crypt_decode_params_bp(bp, salt, iv);
4312 	} else {
4313 		BP_SET_CRYPT(bp, B_TRUE);
4314 	}
4315 
4316 	/* Perform the encryption. This should not fail */
4317 	VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark,
4318 	    BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
4319 	    salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt));
4320 
4321 	/* encode encryption metadata into the bp */
4322 	if (ot == DMU_OT_INTENT_LOG) {
4323 		/*
4324 		 * ZIL blocks store the MAC in the embedded checksum, so the
4325 		 * transform must always be applied.
4326 		 */
4327 		zio_crypt_encode_mac_zil(enc_buf, mac);
4328 		zio_push_transform(zio, eabd, psize, psize, NULL);
4329 	} else {
4330 		BP_SET_CRYPT(bp, B_TRUE);
4331 		zio_crypt_encode_params_bp(bp, salt, iv);
4332 		zio_crypt_encode_mac_bp(bp, mac);
4333 
4334 		if (no_crypt) {
4335 			ASSERT3U(ot, ==, DMU_OT_DNODE);
4336 			abd_free(eabd);
4337 		} else {
4338 			zio_push_transform(zio, eabd, psize, psize, NULL);
4339 		}
4340 	}
4341 
4342 	return (zio);
4343 }
4344 
4345 /*
4346  * ==========================================================================
4347  * Generate and verify checksums
4348  * ==========================================================================
4349  */
4350 static zio_t *
4351 zio_checksum_generate(zio_t *zio)
4352 {
4353 	blkptr_t *bp = zio->io_bp;
4354 	enum zio_checksum checksum;
4355 
4356 	if (bp == NULL) {
4357 		/*
4358 		 * This is zio_write_phys().
4359 		 * We're either generating a label checksum, or none at all.
4360 		 */
4361 		checksum = zio->io_prop.zp_checksum;
4362 
4363 		if (checksum == ZIO_CHECKSUM_OFF)
4364 			return (zio);
4365 
4366 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
4367 	} else {
4368 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
4369 			ASSERT(!IO_IS_ALLOCATING(zio));
4370 			checksum = ZIO_CHECKSUM_GANG_HEADER;
4371 		} else {
4372 			checksum = BP_GET_CHECKSUM(bp);
4373 		}
4374 	}
4375 
4376 	zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
4377 
4378 	return (zio);
4379 }
4380 
4381 static zio_t *
4382 zio_checksum_verify(zio_t *zio)
4383 {
4384 	zio_bad_cksum_t info;
4385 	blkptr_t *bp = zio->io_bp;
4386 	int error;
4387 
4388 	ASSERT(zio->io_vd != NULL);
4389 
4390 	if (bp == NULL) {
4391 		/*
4392 		 * This is zio_read_phys().
4393 		 * We're either verifying a label checksum, or nothing at all.
4394 		 */
4395 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
4396 			return (zio);
4397 
4398 		ASSERT3U(zio->io_prop.zp_checksum, ==, ZIO_CHECKSUM_LABEL);
4399 	}
4400 
4401 	if ((error = zio_checksum_error(zio, &info)) != 0) {
4402 		zio->io_error = error;
4403 		if (error == ECKSUM &&
4404 		    !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
4405 			mutex_enter(&zio->io_vd->vdev_stat_lock);
4406 			zio->io_vd->vdev_stat.vs_checksum_errors++;
4407 			mutex_exit(&zio->io_vd->vdev_stat_lock);
4408 			(void) zfs_ereport_start_checksum(zio->io_spa,
4409 			    zio->io_vd, &zio->io_bookmark, zio,
4410 			    zio->io_offset, zio->io_size, &info);
4411 		}
4412 	}
4413 
4414 	return (zio);
4415 }
4416 
4417 /*
4418  * Called by RAID-Z to ensure we don't compute the checksum twice.
4419  */
4420 void
4421 zio_checksum_verified(zio_t *zio)
4422 {
4423 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
4424 }
4425 
4426 /*
4427  * ==========================================================================
4428  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4429  * An error of 0 indicates success.  ENXIO indicates whole-device failure,
4430  * which may be transient (e.g. unplugged) or permanent.  ECKSUM and EIO
4431  * indicate errors that are specific to one I/O, and most likely permanent.
4432  * Any other error is presumed to be worse because we weren't expecting it.
4433  * ==========================================================================
4434  */
4435 int
4436 zio_worst_error(int e1, int e2)
4437 {
4438 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
4439 	int r1, r2;
4440 
4441 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
4442 		if (e1 == zio_error_rank[r1])
4443 			break;
4444 
4445 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
4446 		if (e2 == zio_error_rank[r2])
4447 			break;
4448 
4449 	return (r1 > r2 ? e1 : e2);
4450 }
4451 
4452 /*
4453  * ==========================================================================
4454  * I/O completion
4455  * ==========================================================================
4456  */
4457 static zio_t *
4458 zio_ready(zio_t *zio)
4459 {
4460 	blkptr_t *bp = zio->io_bp;
4461 	zio_t *pio, *pio_next;
4462 	zio_link_t *zl = NULL;
4463 
4464 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
4465 	    ZIO_WAIT_READY)) {
4466 		return (NULL);
4467 	}
4468 
4469 	if (zio->io_ready) {
4470 		ASSERT(IO_IS_ALLOCATING(zio));
4471 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
4472 		    (zio->io_flags & ZIO_FLAG_NOPWRITE));
4473 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
4474 
4475 		zio->io_ready(zio);
4476 	}
4477 
4478 	if (bp != NULL && bp != &zio->io_bp_copy)
4479 		zio->io_bp_copy = *bp;
4480 
4481 	if (zio->io_error != 0) {
4482 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4483 
4484 		if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4485 			ASSERT(IO_IS_ALLOCATING(zio));
4486 			ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4487 			ASSERT(zio->io_metaslab_class != NULL);
4488 
4489 			/*
4490 			 * We were unable to allocate anything, unreserve and
4491 			 * issue the next I/O to allocate.
4492 			 */
4493 			metaslab_class_throttle_unreserve(
4494 			    zio->io_metaslab_class, zio->io_prop.zp_copies,
4495 			    zio->io_allocator, zio);
4496 			zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
4497 		}
4498 	}
4499 
4500 	mutex_enter(&zio->io_lock);
4501 	zio->io_state[ZIO_WAIT_READY] = 1;
4502 	pio = zio_walk_parents(zio, &zl);
4503 	mutex_exit(&zio->io_lock);
4504 
4505 	/*
4506 	 * As we notify zio's parents, new parents could be added.
4507 	 * New parents go to the head of zio's io_parent_list, however,
4508 	 * so we will (correctly) not notify them.  The remainder of zio's
4509 	 * io_parent_list, from 'pio_next' onward, cannot change because
4510 	 * all parents must wait for us to be done before they can be done.
4511 	 */
4512 	for (; pio != NULL; pio = pio_next) {
4513 		pio_next = zio_walk_parents(zio, &zl);
4514 		zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL);
4515 	}
4516 
4517 	if (zio->io_flags & ZIO_FLAG_NODATA) {
4518 		if (bp != NULL && BP_IS_GANG(bp)) {
4519 			zio->io_flags &= ~ZIO_FLAG_NODATA;
4520 		} else {
4521 			ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
4522 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
4523 		}
4524 	}
4525 
4526 	if (zio_injection_enabled &&
4527 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
4528 		zio_handle_ignored_writes(zio);
4529 
4530 	return (zio);
4531 }
4532 
4533 /*
4534  * Update the allocation throttle accounting.
4535  */
4536 static void
4537 zio_dva_throttle_done(zio_t *zio)
4538 {
4539 	zio_t *lio __maybe_unused = zio->io_logical;
4540 	zio_t *pio = zio_unique_parent(zio);
4541 	vdev_t *vd = zio->io_vd;
4542 	int flags = METASLAB_ASYNC_ALLOC;
4543 
4544 	ASSERT3P(zio->io_bp, !=, NULL);
4545 	ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
4546 	ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
4547 	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
4548 	ASSERT(vd != NULL);
4549 	ASSERT3P(vd, ==, vd->vdev_top);
4550 	ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY));
4551 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4552 	ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
4553 	ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
4554 	ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
4555 
4556 	/*
4557 	 * Parents of gang children can have two flavors -- ones that
4558 	 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4559 	 * and ones that allocated the constituent blocks. The allocation
4560 	 * throttle needs to know the allocating parent zio so we must find
4561 	 * it here.
4562 	 */
4563 	if (pio->io_child_type == ZIO_CHILD_GANG) {
4564 		/*
4565 		 * If our parent is a rewrite gang child then our grandparent
4566 		 * would have been the one that performed the allocation.
4567 		 */
4568 		if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
4569 			pio = zio_unique_parent(pio);
4570 		flags |= METASLAB_GANG_CHILD;
4571 	}
4572 
4573 	ASSERT(IO_IS_ALLOCATING(pio));
4574 	ASSERT3P(zio, !=, zio->io_logical);
4575 	ASSERT(zio->io_logical != NULL);
4576 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4577 	ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
4578 	ASSERT(zio->io_metaslab_class != NULL);
4579 
4580 	mutex_enter(&pio->io_lock);
4581 	metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
4582 	    pio->io_allocator, B_TRUE);
4583 	mutex_exit(&pio->io_lock);
4584 
4585 	metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1,
4586 	    pio->io_allocator, pio);
4587 
4588 	/*
4589 	 * Call into the pipeline to see if there is more work that
4590 	 * needs to be done. If there is work to be done it will be
4591 	 * dispatched to another taskq thread.
4592 	 */
4593 	zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
4594 }
4595 
4596 static zio_t *
4597 zio_done(zio_t *zio)
4598 {
4599 	/*
4600 	 * Always attempt to keep stack usage minimal here since
4601 	 * we can be called recursively up to 19 levels deep.
4602 	 */
4603 	const uint64_t psize = zio->io_size;
4604 	zio_t *pio, *pio_next;
4605 	zio_link_t *zl = NULL;
4606 
4607 	/*
4608 	 * If our children haven't all completed,
4609 	 * wait for them and then repeat this pipeline stage.
4610 	 */
4611 	if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
4612 		return (NULL);
4613 	}
4614 
4615 	/*
4616 	 * If the allocation throttle is enabled, then update the accounting.
4617 	 * We only track child I/Os that are part of an allocating async
4618 	 * write. We must do this since the allocation is performed
4619 	 * by the logical I/O but the actual write is done by child I/Os.
4620 	 */
4621 	if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
4622 	    zio->io_child_type == ZIO_CHILD_VDEV) {
4623 		ASSERT(zio->io_metaslab_class != NULL);
4624 		ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled);
4625 		zio_dva_throttle_done(zio);
4626 	}
4627 
4628 	/*
4629 	 * If the allocation throttle is enabled, verify that
4630 	 * we have decremented the refcounts for every I/O that was throttled.
4631 	 */
4632 	if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4633 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4634 		ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4635 		ASSERT(zio->io_bp != NULL);
4636 
4637 		metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio,
4638 		    zio->io_allocator);
4639 		VERIFY(zfs_refcount_not_held(&zio->io_metaslab_class->
4640 		    mc_allocator[zio->io_allocator].mca_alloc_slots, zio));
4641 	}
4642 
4643 
4644 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
4645 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
4646 			ASSERT(zio->io_children[c][w] == 0);
4647 
4648 	if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) {
4649 		ASSERT(zio->io_bp->blk_pad[0] == 0);
4650 		ASSERT(zio->io_bp->blk_pad[1] == 0);
4651 		ASSERT(memcmp(zio->io_bp, &zio->io_bp_copy,
4652 		    sizeof (blkptr_t)) == 0 ||
4653 		    (zio->io_bp == zio_unique_parent(zio)->io_bp));
4654 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
4655 		    zio->io_bp_override == NULL &&
4656 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
4657 			ASSERT3U(zio->io_prop.zp_copies, <=,
4658 			    BP_GET_NDVAS(zio->io_bp));
4659 			ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
4660 			    (BP_COUNT_GANG(zio->io_bp) ==
4661 			    BP_GET_NDVAS(zio->io_bp)));
4662 		}
4663 		if (zio->io_flags & ZIO_FLAG_NOPWRITE)
4664 			VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig));
4665 	}
4666 
4667 	/*
4668 	 * If there were child vdev/gang/ddt errors, they apply to us now.
4669 	 */
4670 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
4671 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
4672 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
4673 
4674 	/*
4675 	 * If the I/O on the transformed data was successful, generate any
4676 	 * checksum reports now while we still have the transformed data.
4677 	 */
4678 	if (zio->io_error == 0) {
4679 		while (zio->io_cksum_report != NULL) {
4680 			zio_cksum_report_t *zcr = zio->io_cksum_report;
4681 			uint64_t align = zcr->zcr_align;
4682 			uint64_t asize = P2ROUNDUP(psize, align);
4683 			abd_t *adata = zio->io_abd;
4684 
4685 			if (adata != NULL && asize != psize) {
4686 				adata = abd_alloc(asize, B_TRUE);
4687 				abd_copy(adata, zio->io_abd, psize);
4688 				abd_zero_off(adata, psize, asize - psize);
4689 			}
4690 
4691 			zio->io_cksum_report = zcr->zcr_next;
4692 			zcr->zcr_next = NULL;
4693 			zcr->zcr_finish(zcr, adata);
4694 			zfs_ereport_free_checksum(zcr);
4695 
4696 			if (adata != NULL && asize != psize)
4697 				abd_free(adata);
4698 		}
4699 	}
4700 
4701 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
4702 
4703 	vdev_stat_update(zio, psize);
4704 
4705 	/*
4706 	 * If this I/O is attached to a particular vdev is slow, exceeding
4707 	 * 30 seconds to complete, post an error described the I/O delay.
4708 	 * We ignore these errors if the device is currently unavailable.
4709 	 */
4710 	if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) {
4711 		if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) {
4712 			/*
4713 			 * We want to only increment our slow IO counters if
4714 			 * the IO is valid (i.e. not if the drive is removed).
4715 			 *
4716 			 * zfs_ereport_post() will also do these checks, but
4717 			 * it can also ratelimit and have other failures, so we
4718 			 * need to increment the slow_io counters independent
4719 			 * of it.
4720 			 */
4721 			if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY,
4722 			    zio->io_spa, zio->io_vd, zio)) {
4723 				mutex_enter(&zio->io_vd->vdev_stat_lock);
4724 				zio->io_vd->vdev_stat.vs_slow_ios++;
4725 				mutex_exit(&zio->io_vd->vdev_stat_lock);
4726 
4727 				(void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY,
4728 				    zio->io_spa, zio->io_vd, &zio->io_bookmark,
4729 				    zio, 0);
4730 			}
4731 		}
4732 	}
4733 
4734 	if (zio->io_error) {
4735 		/*
4736 		 * If this I/O is attached to a particular vdev,
4737 		 * generate an error message describing the I/O failure
4738 		 * at the block level.  We ignore these errors if the
4739 		 * device is currently unavailable.
4740 		 */
4741 		if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
4742 		    !vdev_is_dead(zio->io_vd)) {
4743 			int ret = zfs_ereport_post(FM_EREPORT_ZFS_IO,
4744 			    zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0);
4745 			if (ret != EALREADY) {
4746 				mutex_enter(&zio->io_vd->vdev_stat_lock);
4747 				if (zio->io_type == ZIO_TYPE_READ)
4748 					zio->io_vd->vdev_stat.vs_read_errors++;
4749 				else if (zio->io_type == ZIO_TYPE_WRITE)
4750 					zio->io_vd->vdev_stat.vs_write_errors++;
4751 				mutex_exit(&zio->io_vd->vdev_stat_lock);
4752 			}
4753 		}
4754 
4755 		if ((zio->io_error == EIO || !(zio->io_flags &
4756 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
4757 		    zio == zio->io_logical) {
4758 			/*
4759 			 * For logical I/O requests, tell the SPA to log the
4760 			 * error and generate a logical data ereport.
4761 			 */
4762 			spa_log_error(zio->io_spa, &zio->io_bookmark,
4763 			    &zio->io_bp->blk_birth);
4764 			(void) zfs_ereport_post(FM_EREPORT_ZFS_DATA,
4765 			    zio->io_spa, NULL, &zio->io_bookmark, zio, 0);
4766 		}
4767 	}
4768 
4769 	if (zio->io_error && zio == zio->io_logical) {
4770 		/*
4771 		 * Determine whether zio should be reexecuted.  This will
4772 		 * propagate all the way to the root via zio_notify_parent().
4773 		 */
4774 		ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
4775 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4776 
4777 		if (IO_IS_ALLOCATING(zio) &&
4778 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
4779 			if (zio->io_error != ENOSPC)
4780 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
4781 			else
4782 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4783 		}
4784 
4785 		if ((zio->io_type == ZIO_TYPE_READ ||
4786 		    zio->io_type == ZIO_TYPE_FREE) &&
4787 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
4788 		    zio->io_error == ENXIO &&
4789 		    spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
4790 		    spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
4791 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4792 
4793 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
4794 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4795 
4796 		/*
4797 		 * Here is a possibly good place to attempt to do
4798 		 * either combinatorial reconstruction or error correction
4799 		 * based on checksums.  It also might be a good place
4800 		 * to send out preliminary ereports before we suspend
4801 		 * processing.
4802 		 */
4803 	}
4804 
4805 	/*
4806 	 * If there were logical child errors, they apply to us now.
4807 	 * We defer this until now to avoid conflating logical child
4808 	 * errors with errors that happened to the zio itself when
4809 	 * updating vdev stats and reporting FMA events above.
4810 	 */
4811 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
4812 
4813 	if ((zio->io_error || zio->io_reexecute) &&
4814 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
4815 	    !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
4816 		zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
4817 
4818 	zio_gang_tree_free(&zio->io_gang_tree);
4819 
4820 	/*
4821 	 * Godfather I/Os should never suspend.
4822 	 */
4823 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4824 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4825 		zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND;
4826 
4827 	if (zio->io_reexecute) {
4828 		/*
4829 		 * This is a logical I/O that wants to reexecute.
4830 		 *
4831 		 * Reexecute is top-down.  When an i/o fails, if it's not
4832 		 * the root, it simply notifies its parent and sticks around.
4833 		 * The parent, seeing that it still has children in zio_done(),
4834 		 * does the same.  This percolates all the way up to the root.
4835 		 * The root i/o will reexecute or suspend the entire tree.
4836 		 *
4837 		 * This approach ensures that zio_reexecute() honors
4838 		 * all the original i/o dependency relationships, e.g.
4839 		 * parents not executing until children are ready.
4840 		 */
4841 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4842 
4843 		zio->io_gang_leader = NULL;
4844 
4845 		mutex_enter(&zio->io_lock);
4846 		zio->io_state[ZIO_WAIT_DONE] = 1;
4847 		mutex_exit(&zio->io_lock);
4848 
4849 		/*
4850 		 * "The Godfather" I/O monitors its children but is
4851 		 * not a true parent to them. It will track them through
4852 		 * the pipeline but severs its ties whenever they get into
4853 		 * trouble (e.g. suspended). This allows "The Godfather"
4854 		 * I/O to return status without blocking.
4855 		 */
4856 		zl = NULL;
4857 		for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4858 		    pio = pio_next) {
4859 			zio_link_t *remove_zl = zl;
4860 			pio_next = zio_walk_parents(zio, &zl);
4861 
4862 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4863 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4864 				zio_remove_child(pio, zio, remove_zl);
4865 				/*
4866 				 * This is a rare code path, so we don't
4867 				 * bother with "next_to_execute".
4868 				 */
4869 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE,
4870 				    NULL);
4871 			}
4872 		}
4873 
4874 		if ((pio = zio_unique_parent(zio)) != NULL) {
4875 			/*
4876 			 * We're not a root i/o, so there's nothing to do
4877 			 * but notify our parent.  Don't propagate errors
4878 			 * upward since we haven't permanently failed yet.
4879 			 */
4880 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4881 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4882 			/*
4883 			 * This is a rare code path, so we don't bother with
4884 			 * "next_to_execute".
4885 			 */
4886 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL);
4887 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4888 			/*
4889 			 * We'd fail again if we reexecuted now, so suspend
4890 			 * until conditions improve (e.g. device comes online).
4891 			 */
4892 			zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
4893 		} else {
4894 			/*
4895 			 * Reexecution is potentially a huge amount of work.
4896 			 * Hand it off to the otherwise-unused claim taskq.
4897 			 */
4898 			ASSERT(taskq_empty_ent(&zio->io_tqent));
4899 			spa_taskq_dispatch_ent(zio->io_spa,
4900 			    ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
4901 			    zio_reexecute, zio, 0, &zio->io_tqent);
4902 		}
4903 		return (NULL);
4904 	}
4905 
4906 	ASSERT(zio->io_child_count == 0);
4907 	ASSERT(zio->io_reexecute == 0);
4908 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4909 
4910 	/*
4911 	 * Report any checksum errors, since the I/O is complete.
4912 	 */
4913 	while (zio->io_cksum_report != NULL) {
4914 		zio_cksum_report_t *zcr = zio->io_cksum_report;
4915 		zio->io_cksum_report = zcr->zcr_next;
4916 		zcr->zcr_next = NULL;
4917 		zcr->zcr_finish(zcr, NULL);
4918 		zfs_ereport_free_checksum(zcr);
4919 	}
4920 
4921 	if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
4922 	    !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) &&
4923 	    !(zio->io_flags & ZIO_FLAG_NOPWRITE)) {
4924 		metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
4925 	}
4926 
4927 	/*
4928 	 * It is the responsibility of the done callback to ensure that this
4929 	 * particular zio is no longer discoverable for adoption, and as
4930 	 * such, cannot acquire any new parents.
4931 	 */
4932 	if (zio->io_done)
4933 		zio->io_done(zio);
4934 
4935 	mutex_enter(&zio->io_lock);
4936 	zio->io_state[ZIO_WAIT_DONE] = 1;
4937 	mutex_exit(&zio->io_lock);
4938 
4939 	/*
4940 	 * We are done executing this zio.  We may want to execute a parent
4941 	 * next.  See the comment in zio_notify_parent().
4942 	 */
4943 	zio_t *next_to_execute = NULL;
4944 	zl = NULL;
4945 	for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4946 		zio_link_t *remove_zl = zl;
4947 		pio_next = zio_walk_parents(zio, &zl);
4948 		zio_remove_child(pio, zio, remove_zl);
4949 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute);
4950 	}
4951 
4952 	if (zio->io_waiter != NULL) {
4953 		mutex_enter(&zio->io_lock);
4954 		zio->io_executor = NULL;
4955 		cv_broadcast(&zio->io_cv);
4956 		mutex_exit(&zio->io_lock);
4957 	} else {
4958 		zio_destroy(zio);
4959 	}
4960 
4961 	return (next_to_execute);
4962 }
4963 
4964 /*
4965  * ==========================================================================
4966  * I/O pipeline definition
4967  * ==========================================================================
4968  */
4969 static zio_pipe_stage_t *zio_pipeline[] = {
4970 	NULL,
4971 	zio_read_bp_init,
4972 	zio_write_bp_init,
4973 	zio_free_bp_init,
4974 	zio_issue_async,
4975 	zio_write_compress,
4976 	zio_encrypt,
4977 	zio_checksum_generate,
4978 	zio_nop_write,
4979 	zio_brt_free,
4980 	zio_ddt_read_start,
4981 	zio_ddt_read_done,
4982 	zio_ddt_write,
4983 	zio_ddt_free,
4984 	zio_gang_assemble,
4985 	zio_gang_issue,
4986 	zio_dva_throttle,
4987 	zio_dva_allocate,
4988 	zio_dva_free,
4989 	zio_dva_claim,
4990 	zio_ready,
4991 	zio_vdev_io_start,
4992 	zio_vdev_io_done,
4993 	zio_vdev_io_assess,
4994 	zio_checksum_verify,
4995 	zio_done
4996 };
4997 
4998 
4999 
5000 
5001 /*
5002  * Compare two zbookmark_phys_t's to see which we would reach first in a
5003  * pre-order traversal of the object tree.
5004  *
5005  * This is simple in every case aside from the meta-dnode object. For all other
5006  * objects, we traverse them in order (object 1 before object 2, and so on).
5007  * However, all of these objects are traversed while traversing object 0, since
5008  * the data it points to is the list of objects.  Thus, we need to convert to a
5009  * canonical representation so we can compare meta-dnode bookmarks to
5010  * non-meta-dnode bookmarks.
5011  *
5012  * We do this by calculating "equivalents" for each field of the zbookmark.
5013  * zbookmarks outside of the meta-dnode use their own object and level, and
5014  * calculate the level 0 equivalent (the first L0 blkid that is contained in the
5015  * blocks this bookmark refers to) by multiplying their blkid by their span
5016  * (the number of L0 blocks contained within one block at their level).
5017  * zbookmarks inside the meta-dnode calculate their object equivalent
5018  * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
5019  * level + 1<<31 (any value larger than a level could ever be) for their level.
5020  * This causes them to always compare before a bookmark in their object
5021  * equivalent, compare appropriately to bookmarks in other objects, and to
5022  * compare appropriately to other bookmarks in the meta-dnode.
5023  */
5024 int
5025 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
5026     const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
5027 {
5028 	/*
5029 	 * These variables represent the "equivalent" values for the zbookmark,
5030 	 * after converting zbookmarks inside the meta dnode to their
5031 	 * normal-object equivalents.
5032 	 */
5033 	uint64_t zb1obj, zb2obj;
5034 	uint64_t zb1L0, zb2L0;
5035 	uint64_t zb1level, zb2level;
5036 
5037 	if (zb1->zb_object == zb2->zb_object &&
5038 	    zb1->zb_level == zb2->zb_level &&
5039 	    zb1->zb_blkid == zb2->zb_blkid)
5040 		return (0);
5041 
5042 	IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT);
5043 	IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT);
5044 
5045 	/*
5046 	 * BP_SPANB calculates the span in blocks.
5047 	 */
5048 	zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
5049 	zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
5050 
5051 	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
5052 		zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
5053 		zb1L0 = 0;
5054 		zb1level = zb1->zb_level + COMPARE_META_LEVEL;
5055 	} else {
5056 		zb1obj = zb1->zb_object;
5057 		zb1level = zb1->zb_level;
5058 	}
5059 
5060 	if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
5061 		zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
5062 		zb2L0 = 0;
5063 		zb2level = zb2->zb_level + COMPARE_META_LEVEL;
5064 	} else {
5065 		zb2obj = zb2->zb_object;
5066 		zb2level = zb2->zb_level;
5067 	}
5068 
5069 	/* Now that we have a canonical representation, do the comparison. */
5070 	if (zb1obj != zb2obj)
5071 		return (zb1obj < zb2obj ? -1 : 1);
5072 	else if (zb1L0 != zb2L0)
5073 		return (zb1L0 < zb2L0 ? -1 : 1);
5074 	else if (zb1level != zb2level)
5075 		return (zb1level > zb2level ? -1 : 1);
5076 	/*
5077 	 * This can (theoretically) happen if the bookmarks have the same object
5078 	 * and level, but different blkids, if the block sizes are not the same.
5079 	 * There is presently no way to change the indirect block sizes
5080 	 */
5081 	return (0);
5082 }
5083 
5084 /*
5085  *  This function checks the following: given that last_block is the place that
5086  *  our traversal stopped last time, does that guarantee that we've visited
5087  *  every node under subtree_root?  Therefore, we can't just use the raw output
5088  *  of zbookmark_compare.  We have to pass in a modified version of
5089  *  subtree_root; by incrementing the block id, and then checking whether
5090  *  last_block is before or equal to that, we can tell whether or not having
5091  *  visited last_block implies that all of subtree_root's children have been
5092  *  visited.
5093  */
5094 boolean_t
5095 zbookmark_subtree_completed(const dnode_phys_t *dnp,
5096     const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
5097 {
5098 	zbookmark_phys_t mod_zb = *subtree_root;
5099 	mod_zb.zb_blkid++;
5100 	ASSERT0(last_block->zb_level);
5101 
5102 	/* The objset_phys_t isn't before anything. */
5103 	if (dnp == NULL)
5104 		return (B_FALSE);
5105 
5106 	/*
5107 	 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
5108 	 * data block size in sectors, because that variable is only used if
5109 	 * the bookmark refers to a block in the meta-dnode.  Since we don't
5110 	 * know without examining it what object it refers to, and there's no
5111 	 * harm in passing in this value in other cases, we always pass it in.
5112 	 *
5113 	 * We pass in 0 for the indirect block size shift because zb2 must be
5114 	 * level 0.  The indirect block size is only used to calculate the span
5115 	 * of the bookmark, but since the bookmark must be level 0, the span is
5116 	 * always 1, so the math works out.
5117 	 *
5118 	 * If you make changes to how the zbookmark_compare code works, be sure
5119 	 * to make sure that this code still works afterwards.
5120 	 */
5121 	return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
5122 	    1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
5123 	    last_block) <= 0);
5124 }
5125 
5126 /*
5127  * This function is similar to zbookmark_subtree_completed(), but returns true
5128  * if subtree_root is equal or ahead of last_block, i.e. still to be done.
5129  */
5130 boolean_t
5131 zbookmark_subtree_tbd(const dnode_phys_t *dnp,
5132     const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
5133 {
5134 	ASSERT0(last_block->zb_level);
5135 	if (dnp == NULL)
5136 		return (B_FALSE);
5137 	return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
5138 	    1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, subtree_root,
5139 	    last_block) >= 0);
5140 }
5141 
5142 EXPORT_SYMBOL(zio_type_name);
5143 EXPORT_SYMBOL(zio_buf_alloc);
5144 EXPORT_SYMBOL(zio_data_buf_alloc);
5145 EXPORT_SYMBOL(zio_buf_free);
5146 EXPORT_SYMBOL(zio_data_buf_free);
5147 
5148 ZFS_MODULE_PARAM(zfs_zio, zio_, slow_io_ms, INT, ZMOD_RW,
5149 	"Max I/O completion time (milliseconds) before marking it as slow");
5150 
5151 ZFS_MODULE_PARAM(zfs_zio, zio_, requeue_io_start_cut_in_line, INT, ZMOD_RW,
5152 	"Prioritize requeued I/O");
5153 
5154 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_deferred_free,  UINT, ZMOD_RW,
5155 	"Defer frees starting in this pass");
5156 
5157 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_dont_compress, UINT, ZMOD_RW,
5158 	"Don't compress starting in this pass");
5159 
5160 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_rewrite, UINT, ZMOD_RW,
5161 	"Rewrite new bps starting in this pass");
5162 
5163 ZFS_MODULE_PARAM(zfs_zio, zio_, dva_throttle_enabled, INT, ZMOD_RW,
5164 	"Throttle block allocations in the ZIO pipeline");
5165 
5166 ZFS_MODULE_PARAM(zfs_zio, zio_, deadman_log_all, INT, ZMOD_RW,
5167 	"Log all slow ZIOs, not just those with vdevs");
5168