xref: /freebsd/sys/contrib/openzfs/module/zfs/zio.c (revision 3332f1b444d4a73238e9f59cca27bfc95fe936bd)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2011, 2020 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/ddt.h>
45 #include <sys/blkptr.h>
46 #include <sys/zfeature.h>
47 #include <sys/dsl_scan.h>
48 #include <sys/metaslab_impl.h>
49 #include <sys/time.h>
50 #include <sys/trace_zfs.h>
51 #include <sys/abd.h>
52 #include <sys/dsl_crypt.h>
53 #include <cityhash.h>
54 
55 /*
56  * ==========================================================================
57  * I/O type descriptions
58  * ==========================================================================
59  */
60 const char *zio_type_name[ZIO_TYPES] = {
61 	/*
62 	 * Note: Linux kernel thread name length is limited
63 	 * so these names will differ from upstream open zfs.
64 	 */
65 	"z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim"
66 };
67 
68 int zio_dva_throttle_enabled = B_TRUE;
69 int zio_deadman_log_all = B_FALSE;
70 
71 /*
72  * ==========================================================================
73  * I/O kmem caches
74  * ==========================================================================
75  */
76 kmem_cache_t *zio_cache;
77 kmem_cache_t *zio_link_cache;
78 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
79 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
80 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
81 uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
82 uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
83 #endif
84 
85 /* Mark IOs as "slow" if they take longer than 30 seconds */
86 int zio_slow_io_ms = (30 * MILLISEC);
87 
88 #define	BP_SPANB(indblkshift, level) \
89 	(((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
90 #define	COMPARE_META_LEVEL	0x80000000ul
91 /*
92  * The following actions directly effect the spa's sync-to-convergence logic.
93  * The values below define the sync pass when we start performing the action.
94  * Care should be taken when changing these values as they directly impact
95  * spa_sync() performance. Tuning these values may introduce subtle performance
96  * pathologies and should only be done in the context of performance analysis.
97  * These tunables will eventually be removed and replaced with #defines once
98  * enough analysis has been done to determine optimal values.
99  *
100  * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
101  * regular blocks are not deferred.
102  *
103  * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable
104  * compression (including of metadata).  In practice, we don't have this
105  * many sync passes, so this has no effect.
106  *
107  * The original intent was that disabling compression would help the sync
108  * passes to converge. However, in practice disabling compression increases
109  * the average number of sync passes, because when we turn compression off, a
110  * lot of block's size will change and thus we have to re-allocate (not
111  * overwrite) them. It also increases the number of 128KB allocations (e.g.
112  * for indirect blocks and spacemaps) because these will not be compressed.
113  * The 128K allocations are especially detrimental to performance on highly
114  * fragmented systems, which may have very few free segments of this size,
115  * and may need to load new metaslabs to satisfy 128K allocations.
116  */
117 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
118 int zfs_sync_pass_dont_compress = 8; /* don't compress starting in this pass */
119 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
120 
121 /*
122  * An allocating zio is one that either currently has the DVA allocate
123  * stage set or will have it later in its lifetime.
124  */
125 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
126 
127 /*
128  * Enable smaller cores by excluding metadata
129  * allocations as well.
130  */
131 int zio_exclude_metadata = 0;
132 int zio_requeue_io_start_cut_in_line = 1;
133 
134 #ifdef ZFS_DEBUG
135 int zio_buf_debug_limit = 16384;
136 #else
137 int zio_buf_debug_limit = 0;
138 #endif
139 
140 static inline void __zio_execute(zio_t *zio);
141 
142 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
143 
144 void
145 zio_init(void)
146 {
147 	size_t c;
148 
149 	zio_cache = kmem_cache_create("zio_cache",
150 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
151 	zio_link_cache = kmem_cache_create("zio_link_cache",
152 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
153 
154 	/*
155 	 * For small buffers, we want a cache for each multiple of
156 	 * SPA_MINBLOCKSIZE.  For larger buffers, we want a cache
157 	 * for each quarter-power of 2.
158 	 */
159 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
160 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
161 		size_t p2 = size;
162 		size_t align = 0;
163 		size_t data_cflags, cflags;
164 
165 		data_cflags = KMC_NODEBUG;
166 		cflags = (zio_exclude_metadata || size > zio_buf_debug_limit) ?
167 		    KMC_NODEBUG : 0;
168 
169 #if defined(_ILP32) && defined(_KERNEL)
170 		/*
171 		 * Cache size limited to 1M on 32-bit platforms until ARC
172 		 * buffers no longer require virtual address space.
173 		 */
174 		if (size > zfs_max_recordsize)
175 			break;
176 #endif
177 
178 		while (!ISP2(p2))
179 			p2 &= p2 - 1;
180 
181 #ifndef _KERNEL
182 		/*
183 		 * If we are using watchpoints, put each buffer on its own page,
184 		 * to eliminate the performance overhead of trapping to the
185 		 * kernel when modifying a non-watched buffer that shares the
186 		 * page with a watched buffer.
187 		 */
188 		if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
189 			continue;
190 		/*
191 		 * Here's the problem - on 4K native devices in userland on
192 		 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
193 		 * will fail with EINVAL, causing zdb (and others) to coredump.
194 		 * Since userland probably doesn't need optimized buffer caches,
195 		 * we just force 4K alignment on everything.
196 		 */
197 		align = 8 * SPA_MINBLOCKSIZE;
198 #else
199 		if (size < PAGESIZE) {
200 			align = SPA_MINBLOCKSIZE;
201 		} else if (IS_P2ALIGNED(size, p2 >> 2)) {
202 			align = PAGESIZE;
203 		}
204 #endif
205 
206 		if (align != 0) {
207 			char name[36];
208 			if (cflags == data_cflags) {
209 				/*
210 				 * Resulting kmem caches would be identical.
211 				 * Save memory by creating only one.
212 				 */
213 				(void) snprintf(name, sizeof (name),
214 				    "zio_buf_comb_%lu", (ulong_t)size);
215 				zio_buf_cache[c] = kmem_cache_create(name,
216 				    size, align, NULL, NULL, NULL, NULL, NULL,
217 				    cflags);
218 				zio_data_buf_cache[c] = zio_buf_cache[c];
219 				continue;
220 			}
221 			(void) snprintf(name, sizeof (name), "zio_buf_%lu",
222 			    (ulong_t)size);
223 			zio_buf_cache[c] = kmem_cache_create(name, size,
224 			    align, NULL, NULL, NULL, NULL, NULL, cflags);
225 
226 			(void) snprintf(name, sizeof (name), "zio_data_buf_%lu",
227 			    (ulong_t)size);
228 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
229 			    align, NULL, NULL, NULL, NULL, NULL, data_cflags);
230 		}
231 	}
232 
233 	while (--c != 0) {
234 		ASSERT(zio_buf_cache[c] != NULL);
235 		if (zio_buf_cache[c - 1] == NULL)
236 			zio_buf_cache[c - 1] = zio_buf_cache[c];
237 
238 		ASSERT(zio_data_buf_cache[c] != NULL);
239 		if (zio_data_buf_cache[c - 1] == NULL)
240 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
241 	}
242 
243 	zio_inject_init();
244 
245 	lz4_init();
246 }
247 
248 void
249 zio_fini(void)
250 {
251 	size_t n = SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT;
252 
253 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
254 	for (size_t i = 0; i < n; i++) {
255 		if (zio_buf_cache_allocs[i] != zio_buf_cache_frees[i])
256 			(void) printf("zio_fini: [%d] %llu != %llu\n",
257 			    (int)((i + 1) << SPA_MINBLOCKSHIFT),
258 			    (long long unsigned)zio_buf_cache_allocs[i],
259 			    (long long unsigned)zio_buf_cache_frees[i]);
260 	}
261 #endif
262 
263 	/*
264 	 * The same kmem cache can show up multiple times in both zio_buf_cache
265 	 * and zio_data_buf_cache. Do a wasteful but trivially correct scan to
266 	 * sort it out.
267 	 */
268 	for (size_t i = 0; i < n; i++) {
269 		kmem_cache_t *cache = zio_buf_cache[i];
270 		if (cache == NULL)
271 			continue;
272 		for (size_t j = i; j < n; j++) {
273 			if (cache == zio_buf_cache[j])
274 				zio_buf_cache[j] = NULL;
275 			if (cache == zio_data_buf_cache[j])
276 				zio_data_buf_cache[j] = NULL;
277 		}
278 		kmem_cache_destroy(cache);
279 	}
280 
281 	for (size_t i = 0; i < n; i++) {
282 		kmem_cache_t *cache = zio_data_buf_cache[i];
283 		if (cache == NULL)
284 			continue;
285 		for (size_t j = i; j < n; j++) {
286 			if (cache == zio_data_buf_cache[j])
287 				zio_data_buf_cache[j] = NULL;
288 		}
289 		kmem_cache_destroy(cache);
290 	}
291 
292 	for (size_t i = 0; i < n; i++) {
293 		VERIFY3P(zio_buf_cache[i], ==, NULL);
294 		VERIFY3P(zio_data_buf_cache[i], ==, NULL);
295 	}
296 
297 	kmem_cache_destroy(zio_link_cache);
298 	kmem_cache_destroy(zio_cache);
299 
300 	zio_inject_fini();
301 
302 	lz4_fini();
303 }
304 
305 /*
306  * ==========================================================================
307  * Allocate and free I/O buffers
308  * ==========================================================================
309  */
310 
311 /*
312  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
313  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
314  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
315  * excess / transient data in-core during a crashdump.
316  */
317 void *
318 zio_buf_alloc(size_t size)
319 {
320 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
321 
322 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
323 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
324 	atomic_add_64(&zio_buf_cache_allocs[c], 1);
325 #endif
326 
327 	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
328 }
329 
330 /*
331  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
332  * crashdump if the kernel panics.  This exists so that we will limit the amount
333  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
334  * of kernel heap dumped to disk when the kernel panics)
335  */
336 void *
337 zio_data_buf_alloc(size_t size)
338 {
339 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
340 
341 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
342 
343 	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
344 }
345 
346 void
347 zio_buf_free(void *buf, size_t size)
348 {
349 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
350 
351 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
352 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
353 	atomic_add_64(&zio_buf_cache_frees[c], 1);
354 #endif
355 
356 	kmem_cache_free(zio_buf_cache[c], buf);
357 }
358 
359 void
360 zio_data_buf_free(void *buf, size_t size)
361 {
362 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
363 
364 	VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
365 
366 	kmem_cache_free(zio_data_buf_cache[c], buf);
367 }
368 
369 static void
370 zio_abd_free(void *abd, size_t size)
371 {
372 	abd_free((abd_t *)abd);
373 }
374 
375 /*
376  * ==========================================================================
377  * Push and pop I/O transform buffers
378  * ==========================================================================
379  */
380 void
381 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
382     zio_transform_func_t *transform)
383 {
384 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
385 
386 	zt->zt_orig_abd = zio->io_abd;
387 	zt->zt_orig_size = zio->io_size;
388 	zt->zt_bufsize = bufsize;
389 	zt->zt_transform = transform;
390 
391 	zt->zt_next = zio->io_transform_stack;
392 	zio->io_transform_stack = zt;
393 
394 	zio->io_abd = data;
395 	zio->io_size = size;
396 }
397 
398 void
399 zio_pop_transforms(zio_t *zio)
400 {
401 	zio_transform_t *zt;
402 
403 	while ((zt = zio->io_transform_stack) != NULL) {
404 		if (zt->zt_transform != NULL)
405 			zt->zt_transform(zio,
406 			    zt->zt_orig_abd, zt->zt_orig_size);
407 
408 		if (zt->zt_bufsize != 0)
409 			abd_free(zio->io_abd);
410 
411 		zio->io_abd = zt->zt_orig_abd;
412 		zio->io_size = zt->zt_orig_size;
413 		zio->io_transform_stack = zt->zt_next;
414 
415 		kmem_free(zt, sizeof (zio_transform_t));
416 	}
417 }
418 
419 /*
420  * ==========================================================================
421  * I/O transform callbacks for subblocks, decompression, and decryption
422  * ==========================================================================
423  */
424 static void
425 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
426 {
427 	ASSERT(zio->io_size > size);
428 
429 	if (zio->io_type == ZIO_TYPE_READ)
430 		abd_copy(data, zio->io_abd, size);
431 }
432 
433 static void
434 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
435 {
436 	if (zio->io_error == 0) {
437 		void *tmp = abd_borrow_buf(data, size);
438 		int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
439 		    zio->io_abd, tmp, zio->io_size, size,
440 		    &zio->io_prop.zp_complevel);
441 		abd_return_buf_copy(data, tmp, size);
442 
443 		if (zio_injection_enabled && ret == 0)
444 			ret = zio_handle_fault_injection(zio, EINVAL);
445 
446 		if (ret != 0)
447 			zio->io_error = SET_ERROR(EIO);
448 	}
449 }
450 
451 static void
452 zio_decrypt(zio_t *zio, abd_t *data, uint64_t size)
453 {
454 	int ret;
455 	void *tmp;
456 	blkptr_t *bp = zio->io_bp;
457 	spa_t *spa = zio->io_spa;
458 	uint64_t dsobj = zio->io_bookmark.zb_objset;
459 	uint64_t lsize = BP_GET_LSIZE(bp);
460 	dmu_object_type_t ot = BP_GET_TYPE(bp);
461 	uint8_t salt[ZIO_DATA_SALT_LEN];
462 	uint8_t iv[ZIO_DATA_IV_LEN];
463 	uint8_t mac[ZIO_DATA_MAC_LEN];
464 	boolean_t no_crypt = B_FALSE;
465 
466 	ASSERT(BP_USES_CRYPT(bp));
467 	ASSERT3U(size, !=, 0);
468 
469 	if (zio->io_error != 0)
470 		return;
471 
472 	/*
473 	 * Verify the cksum of MACs stored in an indirect bp. It will always
474 	 * be possible to verify this since it does not require an encryption
475 	 * key.
476 	 */
477 	if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {
478 		zio_crypt_decode_mac_bp(bp, mac);
479 
480 		if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
481 			/*
482 			 * We haven't decompressed the data yet, but
483 			 * zio_crypt_do_indirect_mac_checksum() requires
484 			 * decompressed data to be able to parse out the MACs
485 			 * from the indirect block. We decompress it now and
486 			 * throw away the result after we are finished.
487 			 */
488 			tmp = zio_buf_alloc(lsize);
489 			ret = zio_decompress_data(BP_GET_COMPRESS(bp),
490 			    zio->io_abd, tmp, zio->io_size, lsize,
491 			    &zio->io_prop.zp_complevel);
492 			if (ret != 0) {
493 				ret = SET_ERROR(EIO);
494 				goto error;
495 			}
496 			ret = zio_crypt_do_indirect_mac_checksum(B_FALSE,
497 			    tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac);
498 			zio_buf_free(tmp, lsize);
499 		} else {
500 			ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
501 			    zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac);
502 		}
503 		abd_copy(data, zio->io_abd, size);
504 
505 		if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) {
506 			ret = zio_handle_decrypt_injection(spa,
507 			    &zio->io_bookmark, ot, ECKSUM);
508 		}
509 		if (ret != 0)
510 			goto error;
511 
512 		return;
513 	}
514 
515 	/*
516 	 * If this is an authenticated block, just check the MAC. It would be
517 	 * nice to separate this out into its own flag, but for the moment
518 	 * enum zio_flag is out of bits.
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 			(void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
575 			    spa, NULL, &zio->io_bookmark, zio, 0);
576 		}
577 	} else {
578 		zio->io_error = ret;
579 	}
580 }
581 
582 /*
583  * ==========================================================================
584  * I/O parent/child relationships and pipeline interlocks
585  * ==========================================================================
586  */
587 zio_t *
588 zio_walk_parents(zio_t *cio, zio_link_t **zl)
589 {
590 	list_t *pl = &cio->io_parent_list;
591 
592 	*zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
593 	if (*zl == NULL)
594 		return (NULL);
595 
596 	ASSERT((*zl)->zl_child == cio);
597 	return ((*zl)->zl_parent);
598 }
599 
600 zio_t *
601 zio_walk_children(zio_t *pio, zio_link_t **zl)
602 {
603 	list_t *cl = &pio->io_child_list;
604 
605 	ASSERT(MUTEX_HELD(&pio->io_lock));
606 
607 	*zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
608 	if (*zl == NULL)
609 		return (NULL);
610 
611 	ASSERT((*zl)->zl_parent == pio);
612 	return ((*zl)->zl_child);
613 }
614 
615 zio_t *
616 zio_unique_parent(zio_t *cio)
617 {
618 	zio_link_t *zl = NULL;
619 	zio_t *pio = zio_walk_parents(cio, &zl);
620 
621 	VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
622 	return (pio);
623 }
624 
625 void
626 zio_add_child(zio_t *pio, zio_t *cio)
627 {
628 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
629 
630 	/*
631 	 * Logical I/Os can have logical, gang, or vdev children.
632 	 * Gang I/Os can have gang or vdev children.
633 	 * Vdev I/Os can only have vdev children.
634 	 * The following ASSERT captures all of these constraints.
635 	 */
636 	ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
637 
638 	zl->zl_parent = pio;
639 	zl->zl_child = cio;
640 
641 	mutex_enter(&pio->io_lock);
642 	mutex_enter(&cio->io_lock);
643 
644 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
645 
646 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
647 		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
648 
649 	list_insert_head(&pio->io_child_list, zl);
650 	list_insert_head(&cio->io_parent_list, zl);
651 
652 	pio->io_child_count++;
653 	cio->io_parent_count++;
654 
655 	mutex_exit(&cio->io_lock);
656 	mutex_exit(&pio->io_lock);
657 }
658 
659 static void
660 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
661 {
662 	ASSERT(zl->zl_parent == pio);
663 	ASSERT(zl->zl_child == cio);
664 
665 	mutex_enter(&pio->io_lock);
666 	mutex_enter(&cio->io_lock);
667 
668 	list_remove(&pio->io_child_list, zl);
669 	list_remove(&cio->io_parent_list, zl);
670 
671 	pio->io_child_count--;
672 	cio->io_parent_count--;
673 
674 	mutex_exit(&cio->io_lock);
675 	mutex_exit(&pio->io_lock);
676 	kmem_cache_free(zio_link_cache, zl);
677 }
678 
679 static boolean_t
680 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
681 {
682 	boolean_t waiting = B_FALSE;
683 
684 	mutex_enter(&zio->io_lock);
685 	ASSERT(zio->io_stall == NULL);
686 	for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
687 		if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
688 			continue;
689 
690 		uint64_t *countp = &zio->io_children[c][wait];
691 		if (*countp != 0) {
692 			zio->io_stage >>= 1;
693 			ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
694 			zio->io_stall = countp;
695 			waiting = B_TRUE;
696 			break;
697 		}
698 	}
699 	mutex_exit(&zio->io_lock);
700 	return (waiting);
701 }
702 
703 __attribute__((always_inline))
704 static inline void
705 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait,
706     zio_t **next_to_executep)
707 {
708 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
709 	int *errorp = &pio->io_child_error[zio->io_child_type];
710 
711 	mutex_enter(&pio->io_lock);
712 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
713 		*errorp = zio_worst_error(*errorp, zio->io_error);
714 	pio->io_reexecute |= zio->io_reexecute;
715 	ASSERT3U(*countp, >, 0);
716 
717 	(*countp)--;
718 
719 	if (*countp == 0 && pio->io_stall == countp) {
720 		zio_taskq_type_t type =
721 		    pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
722 		    ZIO_TASKQ_INTERRUPT;
723 		pio->io_stall = NULL;
724 		mutex_exit(&pio->io_lock);
725 
726 		/*
727 		 * If we can tell the caller to execute this parent next, do
728 		 * so.  Otherwise dispatch the parent zio as its own task.
729 		 *
730 		 * Having the caller execute the parent when possible reduces
731 		 * locking on the zio taskq's, reduces context switch
732 		 * overhead, and has no recursion penalty.  Note that one
733 		 * read from disk typically causes at least 3 zio's: a
734 		 * zio_null(), the logical zio_read(), and then a physical
735 		 * zio.  When the physical ZIO completes, we are able to call
736 		 * zio_done() on all 3 of these zio's from one invocation of
737 		 * zio_execute() by returning the parent back to
738 		 * zio_execute().  Since the parent isn't executed until this
739 		 * thread returns back to zio_execute(), the caller should do
740 		 * so promptly.
741 		 *
742 		 * In other cases, dispatching the parent prevents
743 		 * overflowing the stack when we have deeply nested
744 		 * parent-child relationships, as we do with the "mega zio"
745 		 * of writes for spa_sync(), and the chain of ZIL blocks.
746 		 */
747 		if (next_to_executep != NULL && *next_to_executep == NULL) {
748 			*next_to_executep = pio;
749 		} else {
750 			zio_taskq_dispatch(pio, type, B_FALSE);
751 		}
752 	} else {
753 		mutex_exit(&pio->io_lock);
754 	}
755 }
756 
757 static void
758 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
759 {
760 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
761 		zio->io_error = zio->io_child_error[c];
762 }
763 
764 int
765 zio_bookmark_compare(const void *x1, const void *x2)
766 {
767 	const zio_t *z1 = x1;
768 	const zio_t *z2 = x2;
769 
770 	if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
771 		return (-1);
772 	if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
773 		return (1);
774 
775 	if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
776 		return (-1);
777 	if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
778 		return (1);
779 
780 	if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
781 		return (-1);
782 	if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
783 		return (1);
784 
785 	if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
786 		return (-1);
787 	if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
788 		return (1);
789 
790 	if (z1 < z2)
791 		return (-1);
792 	if (z1 > z2)
793 		return (1);
794 
795 	return (0);
796 }
797 
798 /*
799  * ==========================================================================
800  * Create the various types of I/O (read, write, free, etc)
801  * ==========================================================================
802  */
803 static zio_t *
804 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
805     abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
806     void *private, zio_type_t type, zio_priority_t priority,
807     enum zio_flag flags, vdev_t *vd, uint64_t offset,
808     const zbookmark_phys_t *zb, enum zio_stage stage,
809     enum zio_stage pipeline)
810 {
811 	zio_t *zio;
812 
813 	IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE);
814 	ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
815 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
816 
817 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
818 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
819 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
820 
821 	IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0);
822 
823 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
824 	bzero(zio, sizeof (zio_t));
825 
826 	mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL);
827 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
828 
829 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
830 	    offsetof(zio_link_t, zl_parent_node));
831 	list_create(&zio->io_child_list, sizeof (zio_link_t),
832 	    offsetof(zio_link_t, zl_child_node));
833 	metaslab_trace_init(&zio->io_alloc_list);
834 
835 	if (vd != NULL)
836 		zio->io_child_type = ZIO_CHILD_VDEV;
837 	else if (flags & ZIO_FLAG_GANG_CHILD)
838 		zio->io_child_type = ZIO_CHILD_GANG;
839 	else if (flags & ZIO_FLAG_DDT_CHILD)
840 		zio->io_child_type = ZIO_CHILD_DDT;
841 	else
842 		zio->io_child_type = ZIO_CHILD_LOGICAL;
843 
844 	if (bp != NULL) {
845 		zio->io_bp = (blkptr_t *)bp;
846 		zio->io_bp_copy = *bp;
847 		zio->io_bp_orig = *bp;
848 		if (type != ZIO_TYPE_WRITE ||
849 		    zio->io_child_type == ZIO_CHILD_DDT)
850 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
851 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
852 			zio->io_logical = zio;
853 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
854 			pipeline |= ZIO_GANG_STAGES;
855 	}
856 
857 	zio->io_spa = spa;
858 	zio->io_txg = txg;
859 	zio->io_done = done;
860 	zio->io_private = private;
861 	zio->io_type = type;
862 	zio->io_priority = priority;
863 	zio->io_vd = vd;
864 	zio->io_offset = offset;
865 	zio->io_orig_abd = zio->io_abd = data;
866 	zio->io_orig_size = zio->io_size = psize;
867 	zio->io_lsize = lsize;
868 	zio->io_orig_flags = zio->io_flags = flags;
869 	zio->io_orig_stage = zio->io_stage = stage;
870 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
871 	zio->io_pipeline_trace = ZIO_STAGE_OPEN;
872 
873 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
874 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
875 
876 	if (zb != NULL)
877 		zio->io_bookmark = *zb;
878 
879 	if (pio != NULL) {
880 		zio->io_metaslab_class = pio->io_metaslab_class;
881 		if (zio->io_logical == NULL)
882 			zio->io_logical = pio->io_logical;
883 		if (zio->io_child_type == ZIO_CHILD_GANG)
884 			zio->io_gang_leader = pio->io_gang_leader;
885 		zio_add_child(pio, zio);
886 	}
887 
888 	taskq_init_ent(&zio->io_tqent);
889 
890 	return (zio);
891 }
892 
893 static void
894 zio_destroy(zio_t *zio)
895 {
896 	metaslab_trace_fini(&zio->io_alloc_list);
897 	list_destroy(&zio->io_parent_list);
898 	list_destroy(&zio->io_child_list);
899 	mutex_destroy(&zio->io_lock);
900 	cv_destroy(&zio->io_cv);
901 	kmem_cache_free(zio_cache, zio);
902 }
903 
904 zio_t *
905 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
906     void *private, enum zio_flag flags)
907 {
908 	zio_t *zio;
909 
910 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
911 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
912 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
913 
914 	return (zio);
915 }
916 
917 zio_t *
918 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
919 {
920 	return (zio_null(NULL, spa, NULL, done, private, flags));
921 }
922 
923 static int
924 zfs_blkptr_verify_log(spa_t *spa, const blkptr_t *bp,
925     enum blk_verify_flag blk_verify, const char *fmt, ...)
926 {
927 	va_list adx;
928 	char buf[256];
929 
930 	va_start(adx, fmt);
931 	(void) vsnprintf(buf, sizeof (buf), fmt, adx);
932 	va_end(adx);
933 
934 	switch (blk_verify) {
935 	case BLK_VERIFY_HALT:
936 		dprintf_bp(bp, "blkptr at %p dprintf_bp():", bp);
937 		zfs_panic_recover("%s: %s", spa_name(spa), buf);
938 		break;
939 	case BLK_VERIFY_LOG:
940 		zfs_dbgmsg("%s: %s", spa_name(spa), buf);
941 		break;
942 	case BLK_VERIFY_ONLY:
943 		break;
944 	}
945 
946 	return (1);
947 }
948 
949 /*
950  * Verify the block pointer fields contain reasonable values.  This means
951  * it only contains known object types, checksum/compression identifiers,
952  * block sizes within the maximum allowed limits, valid DVAs, etc.
953  *
954  * If everything checks out B_TRUE is returned.  The zfs_blkptr_verify
955  * argument controls the behavior when an invalid field is detected.
956  *
957  * Modes for zfs_blkptr_verify:
958  *   1) BLK_VERIFY_ONLY (evaluate the block)
959  *   2) BLK_VERIFY_LOG (evaluate the block and log problems)
960  *   3) BLK_VERIFY_HALT (call zfs_panic_recover on error)
961  */
962 boolean_t
963 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp, boolean_t config_held,
964     enum blk_verify_flag blk_verify)
965 {
966 	int errors = 0;
967 
968 	if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
969 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
970 		    "blkptr at %p has invalid TYPE %llu",
971 		    bp, (longlong_t)BP_GET_TYPE(bp));
972 	}
973 	if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
974 	    BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
975 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
976 		    "blkptr at %p has invalid CHECKSUM %llu",
977 		    bp, (longlong_t)BP_GET_CHECKSUM(bp));
978 	}
979 	if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
980 	    BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
981 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
982 		    "blkptr at %p has invalid COMPRESS %llu",
983 		    bp, (longlong_t)BP_GET_COMPRESS(bp));
984 	}
985 	if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
986 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
987 		    "blkptr at %p has invalid LSIZE %llu",
988 		    bp, (longlong_t)BP_GET_LSIZE(bp));
989 	}
990 	if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
991 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
992 		    "blkptr at %p has invalid PSIZE %llu",
993 		    bp, (longlong_t)BP_GET_PSIZE(bp));
994 	}
995 
996 	if (BP_IS_EMBEDDED(bp)) {
997 		if (BPE_GET_ETYPE(bp) >= NUM_BP_EMBEDDED_TYPES) {
998 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
999 			    "blkptr at %p has invalid ETYPE %llu",
1000 			    bp, (longlong_t)BPE_GET_ETYPE(bp));
1001 		}
1002 	}
1003 
1004 	/*
1005 	 * Do not verify individual DVAs if the config is not trusted. This
1006 	 * will be done once the zio is executed in vdev_mirror_map_alloc.
1007 	 */
1008 	if (!spa->spa_trust_config)
1009 		return (errors == 0);
1010 
1011 	if (!config_held)
1012 		spa_config_enter(spa, SCL_VDEV, bp, RW_READER);
1013 	else
1014 		ASSERT(spa_config_held(spa, SCL_VDEV, RW_WRITER));
1015 	/*
1016 	 * Pool-specific checks.
1017 	 *
1018 	 * Note: it would be nice to verify that the blk_birth and
1019 	 * BP_PHYSICAL_BIRTH() are not too large.  However, spa_freeze()
1020 	 * allows the birth time of log blocks (and dmu_sync()-ed blocks
1021 	 * that are in the log) to be arbitrarily large.
1022 	 */
1023 	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
1024 		const dva_t *dva = &bp->blk_dva[i];
1025 		uint64_t vdevid = DVA_GET_VDEV(dva);
1026 
1027 		if (vdevid >= spa->spa_root_vdev->vdev_children) {
1028 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1029 			    "blkptr at %p DVA %u has invalid VDEV %llu",
1030 			    bp, i, (longlong_t)vdevid);
1031 			continue;
1032 		}
1033 		vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1034 		if (vd == NULL) {
1035 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1036 			    "blkptr at %p DVA %u has invalid VDEV %llu",
1037 			    bp, i, (longlong_t)vdevid);
1038 			continue;
1039 		}
1040 		if (vd->vdev_ops == &vdev_hole_ops) {
1041 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1042 			    "blkptr at %p DVA %u has hole VDEV %llu",
1043 			    bp, i, (longlong_t)vdevid);
1044 			continue;
1045 		}
1046 		if (vd->vdev_ops == &vdev_missing_ops) {
1047 			/*
1048 			 * "missing" vdevs are valid during import, but we
1049 			 * don't have their detailed info (e.g. asize), so
1050 			 * we can't perform any more checks on them.
1051 			 */
1052 			continue;
1053 		}
1054 		uint64_t offset = DVA_GET_OFFSET(dva);
1055 		uint64_t asize = DVA_GET_ASIZE(dva);
1056 		if (DVA_GET_GANG(dva))
1057 			asize = vdev_gang_header_asize(vd);
1058 		if (offset + asize > vd->vdev_asize) {
1059 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1060 			    "blkptr at %p DVA %u has invalid OFFSET %llu",
1061 			    bp, i, (longlong_t)offset);
1062 		}
1063 	}
1064 	if (errors > 0)
1065 		dprintf_bp(bp, "blkptr at %p dprintf_bp():", bp);
1066 	if (!config_held)
1067 		spa_config_exit(spa, SCL_VDEV, bp);
1068 
1069 	return (errors == 0);
1070 }
1071 
1072 boolean_t
1073 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
1074 {
1075 	uint64_t vdevid = DVA_GET_VDEV(dva);
1076 
1077 	if (vdevid >= spa->spa_root_vdev->vdev_children)
1078 		return (B_FALSE);
1079 
1080 	vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1081 	if (vd == NULL)
1082 		return (B_FALSE);
1083 
1084 	if (vd->vdev_ops == &vdev_hole_ops)
1085 		return (B_FALSE);
1086 
1087 	if (vd->vdev_ops == &vdev_missing_ops) {
1088 		return (B_FALSE);
1089 	}
1090 
1091 	uint64_t offset = DVA_GET_OFFSET(dva);
1092 	uint64_t asize = DVA_GET_ASIZE(dva);
1093 
1094 	if (DVA_GET_GANG(dva))
1095 		asize = vdev_gang_header_asize(vd);
1096 	if (offset + asize > vd->vdev_asize)
1097 		return (B_FALSE);
1098 
1099 	return (B_TRUE);
1100 }
1101 
1102 zio_t *
1103 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
1104     abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
1105     zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
1106 {
1107 	zio_t *zio;
1108 
1109 	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
1110 	    data, size, size, done, private,
1111 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
1112 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1113 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
1114 
1115 	return (zio);
1116 }
1117 
1118 zio_t *
1119 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
1120     abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
1121     zio_done_func_t *ready, zio_done_func_t *children_ready,
1122     zio_done_func_t *physdone, zio_done_func_t *done,
1123     void *private, zio_priority_t priority, enum zio_flag flags,
1124     const zbookmark_phys_t *zb)
1125 {
1126 	zio_t *zio;
1127 
1128 	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
1129 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
1130 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
1131 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
1132 	    DMU_OT_IS_VALID(zp->zp_type) &&
1133 	    zp->zp_level < 32 &&
1134 	    zp->zp_copies > 0 &&
1135 	    zp->zp_copies <= spa_max_replication(spa));
1136 
1137 	zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
1138 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
1139 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1140 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
1141 
1142 	zio->io_ready = ready;
1143 	zio->io_children_ready = children_ready;
1144 	zio->io_physdone = physdone;
1145 	zio->io_prop = *zp;
1146 
1147 	/*
1148 	 * Data can be NULL if we are going to call zio_write_override() to
1149 	 * provide the already-allocated BP.  But we may need the data to
1150 	 * verify a dedup hit (if requested).  In this case, don't try to
1151 	 * dedup (just take the already-allocated BP verbatim). Encrypted
1152 	 * dedup blocks need data as well so we also disable dedup in this
1153 	 * case.
1154 	 */
1155 	if (data == NULL &&
1156 	    (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) {
1157 		zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
1158 	}
1159 
1160 	return (zio);
1161 }
1162 
1163 zio_t *
1164 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
1165     uint64_t size, zio_done_func_t *done, void *private,
1166     zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
1167 {
1168 	zio_t *zio;
1169 
1170 	zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
1171 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
1172 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
1173 
1174 	return (zio);
1175 }
1176 
1177 void
1178 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
1179 {
1180 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1181 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1182 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1183 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
1184 
1185 	/*
1186 	 * We must reset the io_prop to match the values that existed
1187 	 * when the bp was first written by dmu_sync() keeping in mind
1188 	 * that nopwrite and dedup are mutually exclusive.
1189 	 */
1190 	zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
1191 	zio->io_prop.zp_nopwrite = nopwrite;
1192 	zio->io_prop.zp_copies = copies;
1193 	zio->io_bp_override = bp;
1194 }
1195 
1196 void
1197 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
1198 {
1199 
1200 	(void) zfs_blkptr_verify(spa, bp, B_FALSE, BLK_VERIFY_HALT);
1201 
1202 	/*
1203 	 * The check for EMBEDDED is a performance optimization.  We
1204 	 * process the free here (by ignoring it) rather than
1205 	 * putting it on the list and then processing it in zio_free_sync().
1206 	 */
1207 	if (BP_IS_EMBEDDED(bp))
1208 		return;
1209 	metaslab_check_free(spa, bp);
1210 
1211 	/*
1212 	 * Frees that are for the currently-syncing txg, are not going to be
1213 	 * deferred, and which will not need to do a read (i.e. not GANG or
1214 	 * DEDUP), can be processed immediately.  Otherwise, put them on the
1215 	 * in-memory list for later processing.
1216 	 *
1217 	 * Note that we only defer frees after zfs_sync_pass_deferred_free
1218 	 * when the log space map feature is disabled. [see relevant comment
1219 	 * in spa_sync_iterate_to_convergence()]
1220 	 */
1221 	if (BP_IS_GANG(bp) ||
1222 	    BP_GET_DEDUP(bp) ||
1223 	    txg != spa->spa_syncing_txg ||
1224 	    (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free &&
1225 	    !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))) {
1226 		bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1227 	} else {
1228 		VERIFY3P(zio_free_sync(NULL, spa, txg, bp, 0), ==, NULL);
1229 	}
1230 }
1231 
1232 /*
1233  * To improve performance, this function may return NULL if we were able
1234  * to do the free immediately.  This avoids the cost of creating a zio
1235  * (and linking it to the parent, etc).
1236  */
1237 zio_t *
1238 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1239     enum zio_flag flags)
1240 {
1241 	ASSERT(!BP_IS_HOLE(bp));
1242 	ASSERT(spa_syncing_txg(spa) == txg);
1243 
1244 	if (BP_IS_EMBEDDED(bp))
1245 		return (NULL);
1246 
1247 	metaslab_check_free(spa, bp);
1248 	arc_freed(spa, bp);
1249 	dsl_scan_freed(spa, bp);
1250 
1251 	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) {
1252 		/*
1253 		 * GANG and DEDUP blocks can induce a read (for the gang block
1254 		 * header, or the DDT), so issue them asynchronously so that
1255 		 * this thread is not tied up.
1256 		 */
1257 		enum zio_stage stage =
1258 		    ZIO_FREE_PIPELINE | ZIO_STAGE_ISSUE_ASYNC;
1259 
1260 		return (zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1261 		    BP_GET_PSIZE(bp), NULL, NULL,
1262 		    ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1263 		    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage));
1264 	} else {
1265 		metaslab_free(spa, bp, txg, B_FALSE);
1266 		return (NULL);
1267 	}
1268 }
1269 
1270 zio_t *
1271 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1272     zio_done_func_t *done, void *private, enum zio_flag flags)
1273 {
1274 	zio_t *zio;
1275 
1276 	(void) zfs_blkptr_verify(spa, bp, flags & ZIO_FLAG_CONFIG_WRITER,
1277 	    BLK_VERIFY_HALT);
1278 
1279 	if (BP_IS_EMBEDDED(bp))
1280 		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1281 
1282 	/*
1283 	 * A claim is an allocation of a specific block.  Claims are needed
1284 	 * to support immediate writes in the intent log.  The issue is that
1285 	 * immediate writes contain committed data, but in a txg that was
1286 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
1287 	 * the intent log claims all blocks that contain immediate write data
1288 	 * so that the SPA knows they're in use.
1289 	 *
1290 	 * All claims *must* be resolved in the first txg -- before the SPA
1291 	 * starts allocating blocks -- so that nothing is allocated twice.
1292 	 * If txg == 0 we just verify that the block is claimable.
1293 	 */
1294 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1295 	    spa_min_claim_txg(spa));
1296 	ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1297 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(8) */
1298 
1299 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1300 	    BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1301 	    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1302 	ASSERT0(zio->io_queued_timestamp);
1303 
1304 	return (zio);
1305 }
1306 
1307 zio_t *
1308 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
1309     zio_done_func_t *done, void *private, enum zio_flag flags)
1310 {
1311 	zio_t *zio;
1312 	int c;
1313 
1314 	if (vd->vdev_children == 0) {
1315 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1316 		    ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1317 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1318 
1319 		zio->io_cmd = cmd;
1320 	} else {
1321 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1322 
1323 		for (c = 0; c < vd->vdev_children; c++)
1324 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1325 			    done, private, flags));
1326 	}
1327 
1328 	return (zio);
1329 }
1330 
1331 zio_t *
1332 zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1333     zio_done_func_t *done, void *private, zio_priority_t priority,
1334     enum zio_flag flags, enum trim_flag trim_flags)
1335 {
1336 	zio_t *zio;
1337 
1338 	ASSERT0(vd->vdev_children);
1339 	ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
1340 	ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1341 	ASSERT3U(size, !=, 0);
1342 
1343 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done,
1344 	    private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL,
1345 	    vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE);
1346 	zio->io_trim_flags = trim_flags;
1347 
1348 	return (zio);
1349 }
1350 
1351 zio_t *
1352 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1353     abd_t *data, int checksum, zio_done_func_t *done, void *private,
1354     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1355 {
1356 	zio_t *zio;
1357 
1358 	ASSERT(vd->vdev_children == 0);
1359 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1360 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1361 	ASSERT3U(offset + size, <=, vd->vdev_psize);
1362 
1363 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1364 	    private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1365 	    offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1366 
1367 	zio->io_prop.zp_checksum = checksum;
1368 
1369 	return (zio);
1370 }
1371 
1372 zio_t *
1373 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1374     abd_t *data, int checksum, zio_done_func_t *done, void *private,
1375     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1376 {
1377 	zio_t *zio;
1378 
1379 	ASSERT(vd->vdev_children == 0);
1380 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1381 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1382 	ASSERT3U(offset + size, <=, vd->vdev_psize);
1383 
1384 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1385 	    private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1386 	    offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1387 
1388 	zio->io_prop.zp_checksum = checksum;
1389 
1390 	if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1391 		/*
1392 		 * zec checksums are necessarily destructive -- they modify
1393 		 * the end of the write buffer to hold the verifier/checksum.
1394 		 * Therefore, we must make a local copy in case the data is
1395 		 * being written to multiple places in parallel.
1396 		 */
1397 		abd_t *wbuf = abd_alloc_sametype(data, size);
1398 		abd_copy(wbuf, data, size);
1399 
1400 		zio_push_transform(zio, wbuf, size, size, NULL);
1401 	}
1402 
1403 	return (zio);
1404 }
1405 
1406 /*
1407  * Create a child I/O to do some work for us.
1408  */
1409 zio_t *
1410 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1411     abd_t *data, uint64_t size, int type, zio_priority_t priority,
1412     enum zio_flag flags, zio_done_func_t *done, void *private)
1413 {
1414 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1415 	zio_t *zio;
1416 
1417 	/*
1418 	 * vdev child I/Os do not propagate their error to the parent.
1419 	 * Therefore, for correct operation the caller *must* check for
1420 	 * and handle the error in the child i/o's done callback.
1421 	 * The only exceptions are i/os that we don't care about
1422 	 * (OPTIONAL or REPAIR).
1423 	 */
1424 	ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1425 	    done != NULL);
1426 
1427 	if (type == ZIO_TYPE_READ && bp != NULL) {
1428 		/*
1429 		 * If we have the bp, then the child should perform the
1430 		 * checksum and the parent need not.  This pushes error
1431 		 * detection as close to the leaves as possible and
1432 		 * eliminates redundant checksums in the interior nodes.
1433 		 */
1434 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1435 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1436 	}
1437 
1438 	if (vd->vdev_ops->vdev_op_leaf) {
1439 		ASSERT0(vd->vdev_children);
1440 		offset += VDEV_LABEL_START_SIZE;
1441 	}
1442 
1443 	flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1444 
1445 	/*
1446 	 * If we've decided to do a repair, the write is not speculative --
1447 	 * even if the original read was.
1448 	 */
1449 	if (flags & ZIO_FLAG_IO_REPAIR)
1450 		flags &= ~ZIO_FLAG_SPECULATIVE;
1451 
1452 	/*
1453 	 * If we're creating a child I/O that is not associated with a
1454 	 * top-level vdev, then the child zio is not an allocating I/O.
1455 	 * If this is a retried I/O then we ignore it since we will
1456 	 * have already processed the original allocating I/O.
1457 	 */
1458 	if (flags & ZIO_FLAG_IO_ALLOCATING &&
1459 	    (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1460 		ASSERT(pio->io_metaslab_class != NULL);
1461 		ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled);
1462 		ASSERT(type == ZIO_TYPE_WRITE);
1463 		ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1464 		ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1465 		ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1466 		    pio->io_child_type == ZIO_CHILD_GANG);
1467 
1468 		flags &= ~ZIO_FLAG_IO_ALLOCATING;
1469 	}
1470 
1471 
1472 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1473 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1474 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1475 	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1476 
1477 	zio->io_physdone = pio->io_physdone;
1478 	if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1479 		zio->io_logical->io_phys_children++;
1480 
1481 	return (zio);
1482 }
1483 
1484 zio_t *
1485 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1486     zio_type_t type, zio_priority_t priority, enum zio_flag flags,
1487     zio_done_func_t *done, void *private)
1488 {
1489 	zio_t *zio;
1490 
1491 	ASSERT(vd->vdev_ops->vdev_op_leaf);
1492 
1493 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1494 	    data, size, size, done, private, type, priority,
1495 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1496 	    vd, offset, NULL,
1497 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1498 
1499 	return (zio);
1500 }
1501 
1502 void
1503 zio_flush(zio_t *zio, vdev_t *vd)
1504 {
1505 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1506 	    NULL, NULL,
1507 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1508 }
1509 
1510 void
1511 zio_shrink(zio_t *zio, uint64_t size)
1512 {
1513 	ASSERT3P(zio->io_executor, ==, NULL);
1514 	ASSERT3U(zio->io_orig_size, ==, zio->io_size);
1515 	ASSERT3U(size, <=, zio->io_size);
1516 
1517 	/*
1518 	 * We don't shrink for raidz because of problems with the
1519 	 * reconstruction when reading back less than the block size.
1520 	 * Note, BP_IS_RAIDZ() assumes no compression.
1521 	 */
1522 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1523 	if (!BP_IS_RAIDZ(zio->io_bp)) {
1524 		/* we are not doing a raw write */
1525 		ASSERT3U(zio->io_size, ==, zio->io_lsize);
1526 		zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1527 	}
1528 }
1529 
1530 /*
1531  * ==========================================================================
1532  * Prepare to read and write logical blocks
1533  * ==========================================================================
1534  */
1535 
1536 static zio_t *
1537 zio_read_bp_init(zio_t *zio)
1538 {
1539 	blkptr_t *bp = zio->io_bp;
1540 	uint64_t psize =
1541 	    BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1542 
1543 	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1544 
1545 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1546 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
1547 	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1548 		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1549 		    psize, psize, zio_decompress);
1550 	}
1551 
1552 	if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) ||
1553 	    BP_HAS_INDIRECT_MAC_CKSUM(bp)) &&
1554 	    zio->io_child_type == ZIO_CHILD_LOGICAL) {
1555 		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1556 		    psize, psize, zio_decrypt);
1557 	}
1558 
1559 	if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1560 		int psize = BPE_GET_PSIZE(bp);
1561 		void *data = abd_borrow_buf(zio->io_abd, psize);
1562 
1563 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1564 		decode_embedded_bp_compressed(bp, data);
1565 		abd_return_buf_copy(zio->io_abd, data, psize);
1566 	} else {
1567 		ASSERT(!BP_IS_EMBEDDED(bp));
1568 		ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1569 	}
1570 
1571 	if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1572 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1573 
1574 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1575 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1576 
1577 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1578 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1579 
1580 	return (zio);
1581 }
1582 
1583 static zio_t *
1584 zio_write_bp_init(zio_t *zio)
1585 {
1586 	if (!IO_IS_ALLOCATING(zio))
1587 		return (zio);
1588 
1589 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1590 
1591 	if (zio->io_bp_override) {
1592 		blkptr_t *bp = zio->io_bp;
1593 		zio_prop_t *zp = &zio->io_prop;
1594 
1595 		ASSERT(bp->blk_birth != zio->io_txg);
1596 		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1597 
1598 		*bp = *zio->io_bp_override;
1599 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1600 
1601 		if (BP_IS_EMBEDDED(bp))
1602 			return (zio);
1603 
1604 		/*
1605 		 * If we've been overridden and nopwrite is set then
1606 		 * set the flag accordingly to indicate that a nopwrite
1607 		 * has already occurred.
1608 		 */
1609 		if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1610 			ASSERT(!zp->zp_dedup);
1611 			ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1612 			zio->io_flags |= ZIO_FLAG_NOPWRITE;
1613 			return (zio);
1614 		}
1615 
1616 		ASSERT(!zp->zp_nopwrite);
1617 
1618 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1619 			return (zio);
1620 
1621 		ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1622 		    ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1623 
1624 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum &&
1625 		    !zp->zp_encrypt) {
1626 			BP_SET_DEDUP(bp, 1);
1627 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1628 			return (zio);
1629 		}
1630 
1631 		/*
1632 		 * We were unable to handle this as an override bp, treat
1633 		 * it as a regular write I/O.
1634 		 */
1635 		zio->io_bp_override = NULL;
1636 		*bp = zio->io_bp_orig;
1637 		zio->io_pipeline = zio->io_orig_pipeline;
1638 	}
1639 
1640 	return (zio);
1641 }
1642 
1643 static zio_t *
1644 zio_write_compress(zio_t *zio)
1645 {
1646 	spa_t *spa = zio->io_spa;
1647 	zio_prop_t *zp = &zio->io_prop;
1648 	enum zio_compress compress = zp->zp_compress;
1649 	blkptr_t *bp = zio->io_bp;
1650 	uint64_t lsize = zio->io_lsize;
1651 	uint64_t psize = zio->io_size;
1652 	int pass = 1;
1653 
1654 	/*
1655 	 * If our children haven't all reached the ready stage,
1656 	 * wait for them and then repeat this pipeline stage.
1657 	 */
1658 	if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1659 	    ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1660 		return (NULL);
1661 	}
1662 
1663 	if (!IO_IS_ALLOCATING(zio))
1664 		return (zio);
1665 
1666 	if (zio->io_children_ready != NULL) {
1667 		/*
1668 		 * Now that all our children are ready, run the callback
1669 		 * associated with this zio in case it wants to modify the
1670 		 * data to be written.
1671 		 */
1672 		ASSERT3U(zp->zp_level, >, 0);
1673 		zio->io_children_ready(zio);
1674 	}
1675 
1676 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1677 	ASSERT(zio->io_bp_override == NULL);
1678 
1679 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1680 		/*
1681 		 * We're rewriting an existing block, which means we're
1682 		 * working on behalf of spa_sync().  For spa_sync() to
1683 		 * converge, it must eventually be the case that we don't
1684 		 * have to allocate new blocks.  But compression changes
1685 		 * the blocksize, which forces a reallocate, and makes
1686 		 * convergence take longer.  Therefore, after the first
1687 		 * few passes, stop compressing to ensure convergence.
1688 		 */
1689 		pass = spa_sync_pass(spa);
1690 
1691 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
1692 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1693 		ASSERT(!BP_GET_DEDUP(bp));
1694 
1695 		if (pass >= zfs_sync_pass_dont_compress)
1696 			compress = ZIO_COMPRESS_OFF;
1697 
1698 		/* Make sure someone doesn't change their mind on overwrites */
1699 		ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1700 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1701 	}
1702 
1703 	/* If it's a compressed write that is not raw, compress the buffer. */
1704 	if (compress != ZIO_COMPRESS_OFF &&
1705 	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1706 		void *cbuf = zio_buf_alloc(lsize);
1707 		psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize,
1708 		    zp->zp_complevel);
1709 		if (psize == 0 || psize >= lsize) {
1710 			compress = ZIO_COMPRESS_OFF;
1711 			zio_buf_free(cbuf, lsize);
1712 		} else if (!zp->zp_dedup && !zp->zp_encrypt &&
1713 		    psize <= BPE_PAYLOAD_SIZE &&
1714 		    zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1715 		    spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1716 			encode_embedded_bp_compressed(bp,
1717 			    cbuf, compress, lsize, psize);
1718 			BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1719 			BP_SET_TYPE(bp, zio->io_prop.zp_type);
1720 			BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1721 			zio_buf_free(cbuf, lsize);
1722 			bp->blk_birth = zio->io_txg;
1723 			zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1724 			ASSERT(spa_feature_is_active(spa,
1725 			    SPA_FEATURE_EMBEDDED_DATA));
1726 			return (zio);
1727 		} else {
1728 			/*
1729 			 * Round compressed size up to the minimum allocation
1730 			 * size of the smallest-ashift device, and zero the
1731 			 * tail. This ensures that the compressed size of the
1732 			 * BP (and thus compressratio property) are correct,
1733 			 * in that we charge for the padding used to fill out
1734 			 * the last sector.
1735 			 */
1736 			ASSERT3U(spa->spa_min_alloc, >=, SPA_MINBLOCKSHIFT);
1737 			size_t rounded = (size_t)roundup(psize,
1738 			    spa->spa_min_alloc);
1739 			if (rounded >= lsize) {
1740 				compress = ZIO_COMPRESS_OFF;
1741 				zio_buf_free(cbuf, lsize);
1742 				psize = lsize;
1743 			} else {
1744 				abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1745 				abd_take_ownership_of_buf(cdata, B_TRUE);
1746 				abd_zero_off(cdata, psize, rounded - psize);
1747 				psize = rounded;
1748 				zio_push_transform(zio, cdata,
1749 				    psize, lsize, NULL);
1750 			}
1751 		}
1752 
1753 		/*
1754 		 * We were unable to handle this as an override bp, treat
1755 		 * it as a regular write I/O.
1756 		 */
1757 		zio->io_bp_override = NULL;
1758 		*bp = zio->io_bp_orig;
1759 		zio->io_pipeline = zio->io_orig_pipeline;
1760 
1761 	} else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 &&
1762 	    zp->zp_type == DMU_OT_DNODE) {
1763 		/*
1764 		 * The DMU actually relies on the zio layer's compression
1765 		 * to free metadnode blocks that have had all contained
1766 		 * dnodes freed. As a result, even when doing a raw
1767 		 * receive, we must check whether the block can be compressed
1768 		 * to a hole.
1769 		 */
1770 		psize = zio_compress_data(ZIO_COMPRESS_EMPTY,
1771 		    zio->io_abd, NULL, lsize, zp->zp_complevel);
1772 		if (psize == 0 || psize >= lsize)
1773 			compress = ZIO_COMPRESS_OFF;
1774 	} else if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS) {
1775 		size_t rounded = MIN((size_t)roundup(psize,
1776 		    spa->spa_min_alloc), lsize);
1777 
1778 		if (rounded != psize) {
1779 			abd_t *cdata = abd_alloc_linear(rounded, B_TRUE);
1780 			abd_zero_off(cdata, psize, rounded - psize);
1781 			abd_copy_off(cdata, zio->io_abd, 0, 0, psize);
1782 			psize = rounded;
1783 			zio_push_transform(zio, cdata,
1784 			    psize, rounded, NULL);
1785 		}
1786 	} else {
1787 		ASSERT3U(psize, !=, 0);
1788 	}
1789 
1790 	/*
1791 	 * The final pass of spa_sync() must be all rewrites, but the first
1792 	 * few passes offer a trade-off: allocating blocks defers convergence,
1793 	 * but newly allocated blocks are sequential, so they can be written
1794 	 * to disk faster.  Therefore, we allow the first few passes of
1795 	 * spa_sync() to allocate new blocks, but force rewrites after that.
1796 	 * There should only be a handful of blocks after pass 1 in any case.
1797 	 */
1798 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1799 	    BP_GET_PSIZE(bp) == psize &&
1800 	    pass >= zfs_sync_pass_rewrite) {
1801 		VERIFY3U(psize, !=, 0);
1802 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1803 
1804 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1805 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1806 	} else {
1807 		BP_ZERO(bp);
1808 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1809 	}
1810 
1811 	if (psize == 0) {
1812 		if (zio->io_bp_orig.blk_birth != 0 &&
1813 		    spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1814 			BP_SET_LSIZE(bp, lsize);
1815 			BP_SET_TYPE(bp, zp->zp_type);
1816 			BP_SET_LEVEL(bp, zp->zp_level);
1817 			BP_SET_BIRTH(bp, zio->io_txg, 0);
1818 		}
1819 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1820 	} else {
1821 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1822 		BP_SET_LSIZE(bp, lsize);
1823 		BP_SET_TYPE(bp, zp->zp_type);
1824 		BP_SET_LEVEL(bp, zp->zp_level);
1825 		BP_SET_PSIZE(bp, psize);
1826 		BP_SET_COMPRESS(bp, compress);
1827 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1828 		BP_SET_DEDUP(bp, zp->zp_dedup);
1829 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1830 		if (zp->zp_dedup) {
1831 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1832 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1833 			ASSERT(!zp->zp_encrypt ||
1834 			    DMU_OT_IS_ENCRYPTED(zp->zp_type));
1835 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1836 		}
1837 		if (zp->zp_nopwrite) {
1838 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1839 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1840 			zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1841 		}
1842 	}
1843 	return (zio);
1844 }
1845 
1846 static zio_t *
1847 zio_free_bp_init(zio_t *zio)
1848 {
1849 	blkptr_t *bp = zio->io_bp;
1850 
1851 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1852 		if (BP_GET_DEDUP(bp))
1853 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1854 	}
1855 
1856 	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1857 
1858 	return (zio);
1859 }
1860 
1861 /*
1862  * ==========================================================================
1863  * Execute the I/O pipeline
1864  * ==========================================================================
1865  */
1866 
1867 static void
1868 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1869 {
1870 	spa_t *spa = zio->io_spa;
1871 	zio_type_t t = zio->io_type;
1872 	int flags = (cutinline ? TQ_FRONT : 0);
1873 
1874 	/*
1875 	 * If we're a config writer or a probe, the normal issue and
1876 	 * interrupt threads may all be blocked waiting for the config lock.
1877 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1878 	 */
1879 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1880 		t = ZIO_TYPE_NULL;
1881 
1882 	/*
1883 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1884 	 */
1885 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1886 		t = ZIO_TYPE_NULL;
1887 
1888 	/*
1889 	 * If this is a high priority I/O, then use the high priority taskq if
1890 	 * available.
1891 	 */
1892 	if ((zio->io_priority == ZIO_PRIORITY_NOW ||
1893 	    zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) &&
1894 	    spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1895 		q++;
1896 
1897 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1898 
1899 	/*
1900 	 * NB: We are assuming that the zio can only be dispatched
1901 	 * to a single taskq at a time.  It would be a grievous error
1902 	 * to dispatch the zio to another taskq at the same time.
1903 	 */
1904 	ASSERT(taskq_empty_ent(&zio->io_tqent));
1905 	spa_taskq_dispatch_ent(spa, t, q, zio_execute, zio, flags,
1906 	    &zio->io_tqent);
1907 }
1908 
1909 static boolean_t
1910 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1911 {
1912 	spa_t *spa = zio->io_spa;
1913 
1914 	taskq_t *tq = taskq_of_curthread();
1915 
1916 	for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1917 		spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1918 		uint_t i;
1919 		for (i = 0; i < tqs->stqs_count; i++) {
1920 			if (tqs->stqs_taskq[i] == tq)
1921 				return (B_TRUE);
1922 		}
1923 	}
1924 
1925 	return (B_FALSE);
1926 }
1927 
1928 static zio_t *
1929 zio_issue_async(zio_t *zio)
1930 {
1931 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1932 
1933 	return (NULL);
1934 }
1935 
1936 void
1937 zio_interrupt(void *zio)
1938 {
1939 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1940 }
1941 
1942 void
1943 zio_delay_interrupt(zio_t *zio)
1944 {
1945 	/*
1946 	 * The timeout_generic() function isn't defined in userspace, so
1947 	 * rather than trying to implement the function, the zio delay
1948 	 * functionality has been disabled for userspace builds.
1949 	 */
1950 
1951 #ifdef _KERNEL
1952 	/*
1953 	 * If io_target_timestamp is zero, then no delay has been registered
1954 	 * for this IO, thus jump to the end of this function and "skip" the
1955 	 * delay; issuing it directly to the zio layer.
1956 	 */
1957 	if (zio->io_target_timestamp != 0) {
1958 		hrtime_t now = gethrtime();
1959 
1960 		if (now >= zio->io_target_timestamp) {
1961 			/*
1962 			 * This IO has already taken longer than the target
1963 			 * delay to complete, so we don't want to delay it
1964 			 * any longer; we "miss" the delay and issue it
1965 			 * directly to the zio layer. This is likely due to
1966 			 * the target latency being set to a value less than
1967 			 * the underlying hardware can satisfy (e.g. delay
1968 			 * set to 1ms, but the disks take 10ms to complete an
1969 			 * IO request).
1970 			 */
1971 
1972 			DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1973 			    hrtime_t, now);
1974 
1975 			zio_interrupt(zio);
1976 		} else {
1977 			taskqid_t tid;
1978 			hrtime_t diff = zio->io_target_timestamp - now;
1979 			clock_t expire_at_tick = ddi_get_lbolt() +
1980 			    NSEC_TO_TICK(diff);
1981 
1982 			DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1983 			    hrtime_t, now, hrtime_t, diff);
1984 
1985 			if (NSEC_TO_TICK(diff) == 0) {
1986 				/* Our delay is less than a jiffy - just spin */
1987 				zfs_sleep_until(zio->io_target_timestamp);
1988 				zio_interrupt(zio);
1989 			} else {
1990 				/*
1991 				 * Use taskq_dispatch_delay() in the place of
1992 				 * OpenZFS's timeout_generic().
1993 				 */
1994 				tid = taskq_dispatch_delay(system_taskq,
1995 				    zio_interrupt, zio, TQ_NOSLEEP,
1996 				    expire_at_tick);
1997 				if (tid == TASKQID_INVALID) {
1998 					/*
1999 					 * Couldn't allocate a task.  Just
2000 					 * finish the zio without a delay.
2001 					 */
2002 					zio_interrupt(zio);
2003 				}
2004 			}
2005 		}
2006 		return;
2007 	}
2008 #endif
2009 	DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
2010 	zio_interrupt(zio);
2011 }
2012 
2013 static void
2014 zio_deadman_impl(zio_t *pio, int ziodepth)
2015 {
2016 	zio_t *cio, *cio_next;
2017 	zio_link_t *zl = NULL;
2018 	vdev_t *vd = pio->io_vd;
2019 
2020 	if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) {
2021 		vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL;
2022 		zbookmark_phys_t *zb = &pio->io_bookmark;
2023 		uint64_t delta = gethrtime() - pio->io_timestamp;
2024 		uint64_t failmode = spa_get_deadman_failmode(pio->io_spa);
2025 
2026 		zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu "
2027 		    "delta=%llu queued=%llu io=%llu "
2028 		    "path=%s "
2029 		    "last=%llu type=%d "
2030 		    "priority=%d flags=0x%x stage=0x%x "
2031 		    "pipeline=0x%x pipeline-trace=0x%x "
2032 		    "objset=%llu object=%llu "
2033 		    "level=%llu blkid=%llu "
2034 		    "offset=%llu size=%llu "
2035 		    "error=%d",
2036 		    ziodepth, pio, pio->io_timestamp,
2037 		    (u_longlong_t)delta, pio->io_delta, pio->io_delay,
2038 		    vd ? vd->vdev_path : "NULL",
2039 		    vq ? vq->vq_io_complete_ts : 0, pio->io_type,
2040 		    pio->io_priority, pio->io_flags, pio->io_stage,
2041 		    pio->io_pipeline, pio->io_pipeline_trace,
2042 		    (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object,
2043 		    (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid,
2044 		    (u_longlong_t)pio->io_offset, (u_longlong_t)pio->io_size,
2045 		    pio->io_error);
2046 		(void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN,
2047 		    pio->io_spa, vd, zb, pio, 0);
2048 
2049 		if (failmode == ZIO_FAILURE_MODE_CONTINUE &&
2050 		    taskq_empty_ent(&pio->io_tqent)) {
2051 			zio_interrupt(pio);
2052 		}
2053 	}
2054 
2055 	mutex_enter(&pio->io_lock);
2056 	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2057 		cio_next = zio_walk_children(pio, &zl);
2058 		zio_deadman_impl(cio, ziodepth + 1);
2059 	}
2060 	mutex_exit(&pio->io_lock);
2061 }
2062 
2063 /*
2064  * Log the critical information describing this zio and all of its children
2065  * using the zfs_dbgmsg() interface then post deadman event for the ZED.
2066  */
2067 void
2068 zio_deadman(zio_t *pio, char *tag)
2069 {
2070 	spa_t *spa = pio->io_spa;
2071 	char *name = spa_name(spa);
2072 
2073 	if (!zfs_deadman_enabled || spa_suspended(spa))
2074 		return;
2075 
2076 	zio_deadman_impl(pio, 0);
2077 
2078 	switch (spa_get_deadman_failmode(spa)) {
2079 	case ZIO_FAILURE_MODE_WAIT:
2080 		zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name);
2081 		break;
2082 
2083 	case ZIO_FAILURE_MODE_CONTINUE:
2084 		zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name);
2085 		break;
2086 
2087 	case ZIO_FAILURE_MODE_PANIC:
2088 		fm_panic("%s determined I/O to pool '%s' is hung.", tag, name);
2089 		break;
2090 	}
2091 }
2092 
2093 /*
2094  * Execute the I/O pipeline until one of the following occurs:
2095  * (1) the I/O completes; (2) the pipeline stalls waiting for
2096  * dependent child I/Os; (3) the I/O issues, so we're waiting
2097  * for an I/O completion interrupt; (4) the I/O is delegated by
2098  * vdev-level caching or aggregation; (5) the I/O is deferred
2099  * due to vdev-level queueing; (6) the I/O is handed off to
2100  * another thread.  In all cases, the pipeline stops whenever
2101  * there's no CPU work; it never burns a thread in cv_wait_io().
2102  *
2103  * There's no locking on io_stage because there's no legitimate way
2104  * for multiple threads to be attempting to process the same I/O.
2105  */
2106 static zio_pipe_stage_t *zio_pipeline[];
2107 
2108 /*
2109  * zio_execute() is a wrapper around the static function
2110  * __zio_execute() so that we can force  __zio_execute() to be
2111  * inlined.  This reduces stack overhead which is important
2112  * because __zio_execute() is called recursively in several zio
2113  * code paths.  zio_execute() itself cannot be inlined because
2114  * it is externally visible.
2115  */
2116 void
2117 zio_execute(void *zio)
2118 {
2119 	fstrans_cookie_t cookie;
2120 
2121 	cookie = spl_fstrans_mark();
2122 	__zio_execute(zio);
2123 	spl_fstrans_unmark(cookie);
2124 }
2125 
2126 /*
2127  * Used to determine if in the current context the stack is sized large
2128  * enough to allow zio_execute() to be called recursively.  A minimum
2129  * stack size of 16K is required to avoid needing to re-dispatch the zio.
2130  */
2131 static boolean_t
2132 zio_execute_stack_check(zio_t *zio)
2133 {
2134 #if !defined(HAVE_LARGE_STACKS)
2135 	dsl_pool_t *dp = spa_get_dsl(zio->io_spa);
2136 
2137 	/* Executing in txg_sync_thread() context. */
2138 	if (dp && curthread == dp->dp_tx.tx_sync_thread)
2139 		return (B_TRUE);
2140 
2141 	/* Pool initialization outside of zio_taskq context. */
2142 	if (dp && spa_is_initializing(dp->dp_spa) &&
2143 	    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
2144 	    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))
2145 		return (B_TRUE);
2146 #endif /* HAVE_LARGE_STACKS */
2147 
2148 	return (B_FALSE);
2149 }
2150 
2151 __attribute__((always_inline))
2152 static inline void
2153 __zio_execute(zio_t *zio)
2154 {
2155 	ASSERT3U(zio->io_queued_timestamp, >, 0);
2156 
2157 	while (zio->io_stage < ZIO_STAGE_DONE) {
2158 		enum zio_stage pipeline = zio->io_pipeline;
2159 		enum zio_stage stage = zio->io_stage;
2160 
2161 		zio->io_executor = curthread;
2162 
2163 		ASSERT(!MUTEX_HELD(&zio->io_lock));
2164 		ASSERT(ISP2(stage));
2165 		ASSERT(zio->io_stall == NULL);
2166 
2167 		do {
2168 			stage <<= 1;
2169 		} while ((stage & pipeline) == 0);
2170 
2171 		ASSERT(stage <= ZIO_STAGE_DONE);
2172 
2173 		/*
2174 		 * If we are in interrupt context and this pipeline stage
2175 		 * will grab a config lock that is held across I/O,
2176 		 * or may wait for an I/O that needs an interrupt thread
2177 		 * to complete, issue async to avoid deadlock.
2178 		 *
2179 		 * For VDEV_IO_START, we cut in line so that the io will
2180 		 * be sent to disk promptly.
2181 		 */
2182 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
2183 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
2184 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2185 			    zio_requeue_io_start_cut_in_line : B_FALSE;
2186 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2187 			return;
2188 		}
2189 
2190 		/*
2191 		 * If the current context doesn't have large enough stacks
2192 		 * the zio must be issued asynchronously to prevent overflow.
2193 		 */
2194 		if (zio_execute_stack_check(zio)) {
2195 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2196 			    zio_requeue_io_start_cut_in_line : B_FALSE;
2197 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2198 			return;
2199 		}
2200 
2201 		zio->io_stage = stage;
2202 		zio->io_pipeline_trace |= zio->io_stage;
2203 
2204 		/*
2205 		 * The zio pipeline stage returns the next zio to execute
2206 		 * (typically the same as this one), or NULL if we should
2207 		 * stop.
2208 		 */
2209 		zio = zio_pipeline[highbit64(stage) - 1](zio);
2210 
2211 		if (zio == NULL)
2212 			return;
2213 	}
2214 }
2215 
2216 
2217 /*
2218  * ==========================================================================
2219  * Initiate I/O, either sync or async
2220  * ==========================================================================
2221  */
2222 int
2223 zio_wait(zio_t *zio)
2224 {
2225 	/*
2226 	 * Some routines, like zio_free_sync(), may return a NULL zio
2227 	 * to avoid the performance overhead of creating and then destroying
2228 	 * an unneeded zio.  For the callers' simplicity, we accept a NULL
2229 	 * zio and ignore it.
2230 	 */
2231 	if (zio == NULL)
2232 		return (0);
2233 
2234 	long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms);
2235 	int error;
2236 
2237 	ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN);
2238 	ASSERT3P(zio->io_executor, ==, NULL);
2239 
2240 	zio->io_waiter = curthread;
2241 	ASSERT0(zio->io_queued_timestamp);
2242 	zio->io_queued_timestamp = gethrtime();
2243 
2244 	__zio_execute(zio);
2245 
2246 	mutex_enter(&zio->io_lock);
2247 	while (zio->io_executor != NULL) {
2248 		error = cv_timedwait_io(&zio->io_cv, &zio->io_lock,
2249 		    ddi_get_lbolt() + timeout);
2250 
2251 		if (zfs_deadman_enabled && error == -1 &&
2252 		    gethrtime() - zio->io_queued_timestamp >
2253 		    spa_deadman_ziotime(zio->io_spa)) {
2254 			mutex_exit(&zio->io_lock);
2255 			timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms);
2256 			zio_deadman(zio, FTAG);
2257 			mutex_enter(&zio->io_lock);
2258 		}
2259 	}
2260 	mutex_exit(&zio->io_lock);
2261 
2262 	error = zio->io_error;
2263 	zio_destroy(zio);
2264 
2265 	return (error);
2266 }
2267 
2268 void
2269 zio_nowait(zio_t *zio)
2270 {
2271 	/*
2272 	 * See comment in zio_wait().
2273 	 */
2274 	if (zio == NULL)
2275 		return;
2276 
2277 	ASSERT3P(zio->io_executor, ==, NULL);
2278 
2279 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
2280 	    zio_unique_parent(zio) == NULL) {
2281 		zio_t *pio;
2282 
2283 		/*
2284 		 * This is a logical async I/O with no parent to wait for it.
2285 		 * We add it to the spa_async_root_zio "Godfather" I/O which
2286 		 * will ensure they complete prior to unloading the pool.
2287 		 */
2288 		spa_t *spa = zio->io_spa;
2289 		pio = spa->spa_async_zio_root[CPU_SEQID_UNSTABLE];
2290 
2291 		zio_add_child(pio, zio);
2292 	}
2293 
2294 	ASSERT0(zio->io_queued_timestamp);
2295 	zio->io_queued_timestamp = gethrtime();
2296 	__zio_execute(zio);
2297 }
2298 
2299 /*
2300  * ==========================================================================
2301  * Reexecute, cancel, or suspend/resume failed I/O
2302  * ==========================================================================
2303  */
2304 
2305 static void
2306 zio_reexecute(void *arg)
2307 {
2308 	zio_t *pio = arg;
2309 	zio_t *cio, *cio_next;
2310 
2311 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
2312 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
2313 	ASSERT(pio->io_gang_leader == NULL);
2314 	ASSERT(pio->io_gang_tree == NULL);
2315 
2316 	pio->io_flags = pio->io_orig_flags;
2317 	pio->io_stage = pio->io_orig_stage;
2318 	pio->io_pipeline = pio->io_orig_pipeline;
2319 	pio->io_reexecute = 0;
2320 	pio->io_flags |= ZIO_FLAG_REEXECUTED;
2321 	pio->io_pipeline_trace = 0;
2322 	pio->io_error = 0;
2323 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2324 		pio->io_state[w] = 0;
2325 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2326 		pio->io_child_error[c] = 0;
2327 
2328 	if (IO_IS_ALLOCATING(pio))
2329 		BP_ZERO(pio->io_bp);
2330 
2331 	/*
2332 	 * As we reexecute pio's children, new children could be created.
2333 	 * New children go to the head of pio's io_child_list, however,
2334 	 * so we will (correctly) not reexecute them.  The key is that
2335 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
2336 	 * cannot be affected by any side effects of reexecuting 'cio'.
2337 	 */
2338 	zio_link_t *zl = NULL;
2339 	mutex_enter(&pio->io_lock);
2340 	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2341 		cio_next = zio_walk_children(pio, &zl);
2342 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2343 			pio->io_children[cio->io_child_type][w]++;
2344 		mutex_exit(&pio->io_lock);
2345 		zio_reexecute(cio);
2346 		mutex_enter(&pio->io_lock);
2347 	}
2348 	mutex_exit(&pio->io_lock);
2349 
2350 	/*
2351 	 * Now that all children have been reexecuted, execute the parent.
2352 	 * We don't reexecute "The Godfather" I/O here as it's the
2353 	 * responsibility of the caller to wait on it.
2354 	 */
2355 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
2356 		pio->io_queued_timestamp = gethrtime();
2357 		__zio_execute(pio);
2358 	}
2359 }
2360 
2361 void
2362 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
2363 {
2364 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
2365 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2366 		    "failure and the failure mode property for this pool "
2367 		    "is set to panic.", spa_name(spa));
2368 
2369 	cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
2370 	    "failure and has been suspended.\n", spa_name(spa));
2371 
2372 	(void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
2373 	    NULL, NULL, 0);
2374 
2375 	mutex_enter(&spa->spa_suspend_lock);
2376 
2377 	if (spa->spa_suspend_zio_root == NULL)
2378 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
2379 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2380 		    ZIO_FLAG_GODFATHER);
2381 
2382 	spa->spa_suspended = reason;
2383 
2384 	if (zio != NULL) {
2385 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2386 		ASSERT(zio != spa->spa_suspend_zio_root);
2387 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2388 		ASSERT(zio_unique_parent(zio) == NULL);
2389 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
2390 		zio_add_child(spa->spa_suspend_zio_root, zio);
2391 	}
2392 
2393 	mutex_exit(&spa->spa_suspend_lock);
2394 }
2395 
2396 int
2397 zio_resume(spa_t *spa)
2398 {
2399 	zio_t *pio;
2400 
2401 	/*
2402 	 * Reexecute all previously suspended i/o.
2403 	 */
2404 	mutex_enter(&spa->spa_suspend_lock);
2405 	spa->spa_suspended = ZIO_SUSPEND_NONE;
2406 	cv_broadcast(&spa->spa_suspend_cv);
2407 	pio = spa->spa_suspend_zio_root;
2408 	spa->spa_suspend_zio_root = NULL;
2409 	mutex_exit(&spa->spa_suspend_lock);
2410 
2411 	if (pio == NULL)
2412 		return (0);
2413 
2414 	zio_reexecute(pio);
2415 	return (zio_wait(pio));
2416 }
2417 
2418 void
2419 zio_resume_wait(spa_t *spa)
2420 {
2421 	mutex_enter(&spa->spa_suspend_lock);
2422 	while (spa_suspended(spa))
2423 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
2424 	mutex_exit(&spa->spa_suspend_lock);
2425 }
2426 
2427 /*
2428  * ==========================================================================
2429  * Gang blocks.
2430  *
2431  * A gang block is a collection of small blocks that looks to the DMU
2432  * like one large block.  When zio_dva_allocate() cannot find a block
2433  * of the requested size, due to either severe fragmentation or the pool
2434  * being nearly full, it calls zio_write_gang_block() to construct the
2435  * block from smaller fragments.
2436  *
2437  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2438  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
2439  * an indirect block: it's an array of block pointers.  It consumes
2440  * only one sector and hence is allocatable regardless of fragmentation.
2441  * The gang header's bps point to its gang members, which hold the data.
2442  *
2443  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2444  * as the verifier to ensure uniqueness of the SHA256 checksum.
2445  * Critically, the gang block bp's blk_cksum is the checksum of the data,
2446  * not the gang header.  This ensures that data block signatures (needed for
2447  * deduplication) are independent of how the block is physically stored.
2448  *
2449  * Gang blocks can be nested: a gang member may itself be a gang block.
2450  * Thus every gang block is a tree in which root and all interior nodes are
2451  * gang headers, and the leaves are normal blocks that contain user data.
2452  * The root of the gang tree is called the gang leader.
2453  *
2454  * To perform any operation (read, rewrite, free, claim) on a gang block,
2455  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2456  * in the io_gang_tree field of the original logical i/o by recursively
2457  * reading the gang leader and all gang headers below it.  This yields
2458  * an in-core tree containing the contents of every gang header and the
2459  * bps for every constituent of the gang block.
2460  *
2461  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2462  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
2463  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2464  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2465  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2466  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
2467  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2468  * of the gang header plus zio_checksum_compute() of the data to update the
2469  * gang header's blk_cksum as described above.
2470  *
2471  * The two-phase assemble/issue model solves the problem of partial failure --
2472  * what if you'd freed part of a gang block but then couldn't read the
2473  * gang header for another part?  Assembling the entire gang tree first
2474  * ensures that all the necessary gang header I/O has succeeded before
2475  * starting the actual work of free, claim, or write.  Once the gang tree
2476  * is assembled, free and claim are in-memory operations that cannot fail.
2477  *
2478  * In the event that a gang write fails, zio_dva_unallocate() walks the
2479  * gang tree to immediately free (i.e. insert back into the space map)
2480  * everything we've allocated.  This ensures that we don't get ENOSPC
2481  * errors during repeated suspend/resume cycles due to a flaky device.
2482  *
2483  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
2484  * the gang tree, we won't modify the block, so we can safely defer the free
2485  * (knowing that the block is still intact).  If we *can* assemble the gang
2486  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2487  * each constituent bp and we can allocate a new block on the next sync pass.
2488  *
2489  * In all cases, the gang tree allows complete recovery from partial failure.
2490  * ==========================================================================
2491  */
2492 
2493 static void
2494 zio_gang_issue_func_done(zio_t *zio)
2495 {
2496 	abd_free(zio->io_abd);
2497 }
2498 
2499 static zio_t *
2500 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2501     uint64_t offset)
2502 {
2503 	if (gn != NULL)
2504 		return (pio);
2505 
2506 	return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2507 	    BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2508 	    NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2509 	    &pio->io_bookmark));
2510 }
2511 
2512 static zio_t *
2513 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2514     uint64_t offset)
2515 {
2516 	zio_t *zio;
2517 
2518 	if (gn != NULL) {
2519 		abd_t *gbh_abd =
2520 		    abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2521 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2522 		    gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2523 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2524 		    &pio->io_bookmark);
2525 		/*
2526 		 * As we rewrite each gang header, the pipeline will compute
2527 		 * a new gang block header checksum for it; but no one will
2528 		 * compute a new data checksum, so we do that here.  The one
2529 		 * exception is the gang leader: the pipeline already computed
2530 		 * its data checksum because that stage precedes gang assembly.
2531 		 * (Presently, nothing actually uses interior data checksums;
2532 		 * this is just good hygiene.)
2533 		 */
2534 		if (gn != pio->io_gang_leader->io_gang_tree) {
2535 			abd_t *buf = abd_get_offset(data, offset);
2536 
2537 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2538 			    buf, BP_GET_PSIZE(bp));
2539 
2540 			abd_free(buf);
2541 		}
2542 		/*
2543 		 * If we are here to damage data for testing purposes,
2544 		 * leave the GBH alone so that we can detect the damage.
2545 		 */
2546 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2547 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2548 	} else {
2549 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2550 		    abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2551 		    zio_gang_issue_func_done, NULL, pio->io_priority,
2552 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2553 	}
2554 
2555 	return (zio);
2556 }
2557 
2558 /* ARGSUSED */
2559 static zio_t *
2560 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2561     uint64_t offset)
2562 {
2563 	zio_t *zio = zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2564 	    ZIO_GANG_CHILD_FLAGS(pio));
2565 	if (zio == NULL) {
2566 		zio = zio_null(pio, pio->io_spa,
2567 		    NULL, NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio));
2568 	}
2569 	return (zio);
2570 }
2571 
2572 /* ARGSUSED */
2573 static zio_t *
2574 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2575     uint64_t offset)
2576 {
2577 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2578 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2579 }
2580 
2581 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2582 	NULL,
2583 	zio_read_gang,
2584 	zio_rewrite_gang,
2585 	zio_free_gang,
2586 	zio_claim_gang,
2587 	NULL
2588 };
2589 
2590 static void zio_gang_tree_assemble_done(zio_t *zio);
2591 
2592 static zio_gang_node_t *
2593 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2594 {
2595 	zio_gang_node_t *gn;
2596 
2597 	ASSERT(*gnpp == NULL);
2598 
2599 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2600 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2601 	*gnpp = gn;
2602 
2603 	return (gn);
2604 }
2605 
2606 static void
2607 zio_gang_node_free(zio_gang_node_t **gnpp)
2608 {
2609 	zio_gang_node_t *gn = *gnpp;
2610 
2611 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2612 		ASSERT(gn->gn_child[g] == NULL);
2613 
2614 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2615 	kmem_free(gn, sizeof (*gn));
2616 	*gnpp = NULL;
2617 }
2618 
2619 static void
2620 zio_gang_tree_free(zio_gang_node_t **gnpp)
2621 {
2622 	zio_gang_node_t *gn = *gnpp;
2623 
2624 	if (gn == NULL)
2625 		return;
2626 
2627 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2628 		zio_gang_tree_free(&gn->gn_child[g]);
2629 
2630 	zio_gang_node_free(gnpp);
2631 }
2632 
2633 static void
2634 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2635 {
2636 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2637 	abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2638 
2639 	ASSERT(gio->io_gang_leader == gio);
2640 	ASSERT(BP_IS_GANG(bp));
2641 
2642 	zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2643 	    zio_gang_tree_assemble_done, gn, gio->io_priority,
2644 	    ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2645 }
2646 
2647 static void
2648 zio_gang_tree_assemble_done(zio_t *zio)
2649 {
2650 	zio_t *gio = zio->io_gang_leader;
2651 	zio_gang_node_t *gn = zio->io_private;
2652 	blkptr_t *bp = zio->io_bp;
2653 
2654 	ASSERT(gio == zio_unique_parent(zio));
2655 	ASSERT(zio->io_child_count == 0);
2656 
2657 	if (zio->io_error)
2658 		return;
2659 
2660 	/* this ABD was created from a linear buf in zio_gang_tree_assemble */
2661 	if (BP_SHOULD_BYTESWAP(bp))
2662 		byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2663 
2664 	ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2665 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2666 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2667 
2668 	abd_free(zio->io_abd);
2669 
2670 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2671 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2672 		if (!BP_IS_GANG(gbp))
2673 			continue;
2674 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2675 	}
2676 }
2677 
2678 static void
2679 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2680     uint64_t offset)
2681 {
2682 	zio_t *gio = pio->io_gang_leader;
2683 	zio_t *zio;
2684 
2685 	ASSERT(BP_IS_GANG(bp) == !!gn);
2686 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2687 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2688 
2689 	/*
2690 	 * If you're a gang header, your data is in gn->gn_gbh.
2691 	 * If you're a gang member, your data is in 'data' and gn == NULL.
2692 	 */
2693 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2694 
2695 	if (gn != NULL) {
2696 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2697 
2698 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2699 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2700 			if (BP_IS_HOLE(gbp))
2701 				continue;
2702 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2703 			    offset);
2704 			offset += BP_GET_PSIZE(gbp);
2705 		}
2706 	}
2707 
2708 	if (gn == gio->io_gang_tree)
2709 		ASSERT3U(gio->io_size, ==, offset);
2710 
2711 	if (zio != pio)
2712 		zio_nowait(zio);
2713 }
2714 
2715 static zio_t *
2716 zio_gang_assemble(zio_t *zio)
2717 {
2718 	blkptr_t *bp = zio->io_bp;
2719 
2720 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2721 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2722 
2723 	zio->io_gang_leader = zio;
2724 
2725 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2726 
2727 	return (zio);
2728 }
2729 
2730 static zio_t *
2731 zio_gang_issue(zio_t *zio)
2732 {
2733 	blkptr_t *bp = zio->io_bp;
2734 
2735 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2736 		return (NULL);
2737 	}
2738 
2739 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2740 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2741 
2742 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2743 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2744 		    0);
2745 	else
2746 		zio_gang_tree_free(&zio->io_gang_tree);
2747 
2748 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2749 
2750 	return (zio);
2751 }
2752 
2753 static void
2754 zio_write_gang_member_ready(zio_t *zio)
2755 {
2756 	zio_t *pio = zio_unique_parent(zio);
2757 	dva_t *cdva = zio->io_bp->blk_dva;
2758 	dva_t *pdva = pio->io_bp->blk_dva;
2759 	uint64_t asize;
2760 	zio_t *gio __maybe_unused = zio->io_gang_leader;
2761 
2762 	if (BP_IS_HOLE(zio->io_bp))
2763 		return;
2764 
2765 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2766 
2767 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2768 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2769 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2770 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2771 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2772 
2773 	mutex_enter(&pio->io_lock);
2774 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2775 		ASSERT(DVA_GET_GANG(&pdva[d]));
2776 		asize = DVA_GET_ASIZE(&pdva[d]);
2777 		asize += DVA_GET_ASIZE(&cdva[d]);
2778 		DVA_SET_ASIZE(&pdva[d], asize);
2779 	}
2780 	mutex_exit(&pio->io_lock);
2781 }
2782 
2783 static void
2784 zio_write_gang_done(zio_t *zio)
2785 {
2786 	/*
2787 	 * The io_abd field will be NULL for a zio with no data.  The io_flags
2788 	 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2789 	 * check for it here as it is cleared in zio_ready.
2790 	 */
2791 	if (zio->io_abd != NULL)
2792 		abd_free(zio->io_abd);
2793 }
2794 
2795 static zio_t *
2796 zio_write_gang_block(zio_t *pio, metaslab_class_t *mc)
2797 {
2798 	spa_t *spa = pio->io_spa;
2799 	blkptr_t *bp = pio->io_bp;
2800 	zio_t *gio = pio->io_gang_leader;
2801 	zio_t *zio;
2802 	zio_gang_node_t *gn, **gnpp;
2803 	zio_gbh_phys_t *gbh;
2804 	abd_t *gbh_abd;
2805 	uint64_t txg = pio->io_txg;
2806 	uint64_t resid = pio->io_size;
2807 	uint64_t lsize;
2808 	int copies = gio->io_prop.zp_copies;
2809 	int gbh_copies;
2810 	zio_prop_t zp;
2811 	int error;
2812 	boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
2813 
2814 	/*
2815 	 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2816 	 * have a third copy.
2817 	 */
2818 	gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2819 	if (gio->io_prop.zp_encrypt && gbh_copies >= SPA_DVAS_PER_BP)
2820 		gbh_copies = SPA_DVAS_PER_BP - 1;
2821 
2822 	int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2823 	if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2824 		ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2825 		ASSERT(has_data);
2826 
2827 		flags |= METASLAB_ASYNC_ALLOC;
2828 		VERIFY(zfs_refcount_held(&mc->mc_allocator[pio->io_allocator].
2829 		    mca_alloc_slots, pio));
2830 
2831 		/*
2832 		 * The logical zio has already placed a reservation for
2833 		 * 'copies' allocation slots but gang blocks may require
2834 		 * additional copies. These additional copies
2835 		 * (i.e. gbh_copies - copies) are guaranteed to succeed
2836 		 * since metaslab_class_throttle_reserve() always allows
2837 		 * additional reservations for gang blocks.
2838 		 */
2839 		VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2840 		    pio->io_allocator, pio, flags));
2841 	}
2842 
2843 	error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2844 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2845 	    &pio->io_alloc_list, pio, pio->io_allocator);
2846 	if (error) {
2847 		if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2848 			ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2849 			ASSERT(has_data);
2850 
2851 			/*
2852 			 * If we failed to allocate the gang block header then
2853 			 * we remove any additional allocation reservations that
2854 			 * we placed here. The original reservation will
2855 			 * be removed when the logical I/O goes to the ready
2856 			 * stage.
2857 			 */
2858 			metaslab_class_throttle_unreserve(mc,
2859 			    gbh_copies - copies, pio->io_allocator, pio);
2860 		}
2861 
2862 		pio->io_error = error;
2863 		return (pio);
2864 	}
2865 
2866 	if (pio == gio) {
2867 		gnpp = &gio->io_gang_tree;
2868 	} else {
2869 		gnpp = pio->io_private;
2870 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
2871 	}
2872 
2873 	gn = zio_gang_node_alloc(gnpp);
2874 	gbh = gn->gn_gbh;
2875 	bzero(gbh, SPA_GANGBLOCKSIZE);
2876 	gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2877 
2878 	/*
2879 	 * Create the gang header.
2880 	 */
2881 	zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2882 	    zio_write_gang_done, NULL, pio->io_priority,
2883 	    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2884 
2885 	/*
2886 	 * Create and nowait the gang children.
2887 	 */
2888 	for (int g = 0; resid != 0; resid -= lsize, g++) {
2889 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2890 		    SPA_MINBLOCKSIZE);
2891 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2892 
2893 		zp.zp_checksum = gio->io_prop.zp_checksum;
2894 		zp.zp_compress = ZIO_COMPRESS_OFF;
2895 		zp.zp_complevel = gio->io_prop.zp_complevel;
2896 		zp.zp_type = DMU_OT_NONE;
2897 		zp.zp_level = 0;
2898 		zp.zp_copies = gio->io_prop.zp_copies;
2899 		zp.zp_dedup = B_FALSE;
2900 		zp.zp_dedup_verify = B_FALSE;
2901 		zp.zp_nopwrite = B_FALSE;
2902 		zp.zp_encrypt = gio->io_prop.zp_encrypt;
2903 		zp.zp_byteorder = gio->io_prop.zp_byteorder;
2904 		bzero(zp.zp_salt, ZIO_DATA_SALT_LEN);
2905 		bzero(zp.zp_iv, ZIO_DATA_IV_LEN);
2906 		bzero(zp.zp_mac, ZIO_DATA_MAC_LEN);
2907 
2908 		zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2909 		    has_data ? abd_get_offset(pio->io_abd, pio->io_size -
2910 		    resid) : NULL, lsize, lsize, &zp,
2911 		    zio_write_gang_member_ready, NULL, NULL,
2912 		    zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2913 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2914 
2915 		if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2916 			ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2917 			ASSERT(has_data);
2918 
2919 			/*
2920 			 * Gang children won't throttle but we should
2921 			 * account for their work, so reserve an allocation
2922 			 * slot for them here.
2923 			 */
2924 			VERIFY(metaslab_class_throttle_reserve(mc,
2925 			    zp.zp_copies, cio->io_allocator, cio, flags));
2926 		}
2927 		zio_nowait(cio);
2928 	}
2929 
2930 	/*
2931 	 * Set pio's pipeline to just wait for zio to finish.
2932 	 */
2933 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2934 
2935 	/*
2936 	 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2937 	 */
2938 	pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
2939 
2940 	zio_nowait(zio);
2941 
2942 	return (pio);
2943 }
2944 
2945 /*
2946  * The zio_nop_write stage in the pipeline determines if allocating a
2947  * new bp is necessary.  The nopwrite feature can handle writes in
2948  * either syncing or open context (i.e. zil writes) and as a result is
2949  * mutually exclusive with dedup.
2950  *
2951  * By leveraging a cryptographically secure checksum, such as SHA256, we
2952  * can compare the checksums of the new data and the old to determine if
2953  * allocating a new block is required.  Note that our requirements for
2954  * cryptographic strength are fairly weak: there can't be any accidental
2955  * hash collisions, but we don't need to be secure against intentional
2956  * (malicious) collisions.  To trigger a nopwrite, you have to be able
2957  * to write the file to begin with, and triggering an incorrect (hash
2958  * collision) nopwrite is no worse than simply writing to the file.
2959  * That said, there are no known attacks against the checksum algorithms
2960  * used for nopwrite, assuming that the salt and the checksums
2961  * themselves remain secret.
2962  */
2963 static zio_t *
2964 zio_nop_write(zio_t *zio)
2965 {
2966 	blkptr_t *bp = zio->io_bp;
2967 	blkptr_t *bp_orig = &zio->io_bp_orig;
2968 	zio_prop_t *zp = &zio->io_prop;
2969 
2970 	ASSERT(BP_GET_LEVEL(bp) == 0);
2971 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2972 	ASSERT(zp->zp_nopwrite);
2973 	ASSERT(!zp->zp_dedup);
2974 	ASSERT(zio->io_bp_override == NULL);
2975 	ASSERT(IO_IS_ALLOCATING(zio));
2976 
2977 	/*
2978 	 * Check to see if the original bp and the new bp have matching
2979 	 * characteristics (i.e. same checksum, compression algorithms, etc).
2980 	 * If they don't then just continue with the pipeline which will
2981 	 * allocate a new bp.
2982 	 */
2983 	if (BP_IS_HOLE(bp_orig) ||
2984 	    !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2985 	    ZCHECKSUM_FLAG_NOPWRITE) ||
2986 	    BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) ||
2987 	    BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2988 	    BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2989 	    BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2990 	    zp->zp_copies != BP_GET_NDVAS(bp_orig))
2991 		return (zio);
2992 
2993 	/*
2994 	 * If the checksums match then reset the pipeline so that we
2995 	 * avoid allocating a new bp and issuing any I/O.
2996 	 */
2997 	if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2998 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2999 		    ZCHECKSUM_FLAG_NOPWRITE);
3000 		ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
3001 		ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
3002 		ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
3003 		ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
3004 		    sizeof (uint64_t)) == 0);
3005 
3006 		/*
3007 		 * If we're overwriting a block that is currently on an
3008 		 * indirect vdev, then ignore the nopwrite request and
3009 		 * allow a new block to be allocated on a concrete vdev.
3010 		 */
3011 		spa_config_enter(zio->io_spa, SCL_VDEV, FTAG, RW_READER);
3012 		vdev_t *tvd = vdev_lookup_top(zio->io_spa,
3013 		    DVA_GET_VDEV(&bp->blk_dva[0]));
3014 		if (tvd->vdev_ops == &vdev_indirect_ops) {
3015 			spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3016 			return (zio);
3017 		}
3018 		spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3019 
3020 		*bp = *bp_orig;
3021 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3022 		zio->io_flags |= ZIO_FLAG_NOPWRITE;
3023 	}
3024 
3025 	return (zio);
3026 }
3027 
3028 /*
3029  * ==========================================================================
3030  * Dedup
3031  * ==========================================================================
3032  */
3033 static void
3034 zio_ddt_child_read_done(zio_t *zio)
3035 {
3036 	blkptr_t *bp = zio->io_bp;
3037 	ddt_entry_t *dde = zio->io_private;
3038 	ddt_phys_t *ddp;
3039 	zio_t *pio = zio_unique_parent(zio);
3040 
3041 	mutex_enter(&pio->io_lock);
3042 	ddp = ddt_phys_select(dde, bp);
3043 	if (zio->io_error == 0)
3044 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
3045 
3046 	if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
3047 		dde->dde_repair_abd = zio->io_abd;
3048 	else
3049 		abd_free(zio->io_abd);
3050 	mutex_exit(&pio->io_lock);
3051 }
3052 
3053 static zio_t *
3054 zio_ddt_read_start(zio_t *zio)
3055 {
3056 	blkptr_t *bp = zio->io_bp;
3057 
3058 	ASSERT(BP_GET_DEDUP(bp));
3059 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3060 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3061 
3062 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
3063 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
3064 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
3065 		ddt_phys_t *ddp = dde->dde_phys;
3066 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
3067 		blkptr_t blk;
3068 
3069 		ASSERT(zio->io_vsd == NULL);
3070 		zio->io_vsd = dde;
3071 
3072 		if (ddp_self == NULL)
3073 			return (zio);
3074 
3075 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
3076 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
3077 				continue;
3078 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
3079 			    &blk);
3080 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
3081 			    abd_alloc_for_io(zio->io_size, B_TRUE),
3082 			    zio->io_size, zio_ddt_child_read_done, dde,
3083 			    zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
3084 			    ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
3085 		}
3086 		return (zio);
3087 	}
3088 
3089 	zio_nowait(zio_read(zio, zio->io_spa, bp,
3090 	    zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
3091 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
3092 
3093 	return (zio);
3094 }
3095 
3096 static zio_t *
3097 zio_ddt_read_done(zio_t *zio)
3098 {
3099 	blkptr_t *bp = zio->io_bp;
3100 
3101 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
3102 		return (NULL);
3103 	}
3104 
3105 	ASSERT(BP_GET_DEDUP(bp));
3106 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3107 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3108 
3109 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
3110 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
3111 		ddt_entry_t *dde = zio->io_vsd;
3112 		if (ddt == NULL) {
3113 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
3114 			return (zio);
3115 		}
3116 		if (dde == NULL) {
3117 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
3118 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
3119 			return (NULL);
3120 		}
3121 		if (dde->dde_repair_abd != NULL) {
3122 			abd_copy(zio->io_abd, dde->dde_repair_abd,
3123 			    zio->io_size);
3124 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
3125 		}
3126 		ddt_repair_done(ddt, dde);
3127 		zio->io_vsd = NULL;
3128 	}
3129 
3130 	ASSERT(zio->io_vsd == NULL);
3131 
3132 	return (zio);
3133 }
3134 
3135 static boolean_t
3136 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
3137 {
3138 	spa_t *spa = zio->io_spa;
3139 	boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW);
3140 
3141 	ASSERT(!(zio->io_bp_override && do_raw));
3142 
3143 	/*
3144 	 * Note: we compare the original data, not the transformed data,
3145 	 * because when zio->io_bp is an override bp, we will not have
3146 	 * pushed the I/O transforms.  That's an important optimization
3147 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
3148 	 * However, we should never get a raw, override zio so in these
3149 	 * cases we can compare the io_abd directly. This is useful because
3150 	 * it allows us to do dedup verification even if we don't have access
3151 	 * to the original data (for instance, if the encryption keys aren't
3152 	 * loaded).
3153 	 */
3154 
3155 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
3156 		zio_t *lio = dde->dde_lead_zio[p];
3157 
3158 		if (lio != NULL && do_raw) {
3159 			return (lio->io_size != zio->io_size ||
3160 			    abd_cmp(zio->io_abd, lio->io_abd) != 0);
3161 		} else if (lio != NULL) {
3162 			return (lio->io_orig_size != zio->io_orig_size ||
3163 			    abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0);
3164 		}
3165 	}
3166 
3167 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
3168 		ddt_phys_t *ddp = &dde->dde_phys[p];
3169 
3170 		if (ddp->ddp_phys_birth != 0 && do_raw) {
3171 			blkptr_t blk = *zio->io_bp;
3172 			uint64_t psize;
3173 			abd_t *tmpabd;
3174 			int error;
3175 
3176 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
3177 			psize = BP_GET_PSIZE(&blk);
3178 
3179 			if (psize != zio->io_size)
3180 				return (B_TRUE);
3181 
3182 			ddt_exit(ddt);
3183 
3184 			tmpabd = abd_alloc_for_io(psize, B_TRUE);
3185 
3186 			error = zio_wait(zio_read(NULL, spa, &blk, tmpabd,
3187 			    psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
3188 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
3189 			    ZIO_FLAG_RAW, &zio->io_bookmark));
3190 
3191 			if (error == 0) {
3192 				if (abd_cmp(tmpabd, zio->io_abd) != 0)
3193 					error = SET_ERROR(ENOENT);
3194 			}
3195 
3196 			abd_free(tmpabd);
3197 			ddt_enter(ddt);
3198 			return (error != 0);
3199 		} else if (ddp->ddp_phys_birth != 0) {
3200 			arc_buf_t *abuf = NULL;
3201 			arc_flags_t aflags = ARC_FLAG_WAIT;
3202 			blkptr_t blk = *zio->io_bp;
3203 			int error;
3204 
3205 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
3206 
3207 			if (BP_GET_LSIZE(&blk) != zio->io_orig_size)
3208 				return (B_TRUE);
3209 
3210 			ddt_exit(ddt);
3211 
3212 			error = arc_read(NULL, spa, &blk,
3213 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
3214 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
3215 			    &aflags, &zio->io_bookmark);
3216 
3217 			if (error == 0) {
3218 				if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
3219 				    zio->io_orig_size) != 0)
3220 					error = SET_ERROR(ENOENT);
3221 				arc_buf_destroy(abuf, &abuf);
3222 			}
3223 
3224 			ddt_enter(ddt);
3225 			return (error != 0);
3226 		}
3227 	}
3228 
3229 	return (B_FALSE);
3230 }
3231 
3232 static void
3233 zio_ddt_child_write_ready(zio_t *zio)
3234 {
3235 	int p = zio->io_prop.zp_copies;
3236 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3237 	ddt_entry_t *dde = zio->io_private;
3238 	ddt_phys_t *ddp = &dde->dde_phys[p];
3239 	zio_t *pio;
3240 
3241 	if (zio->io_error)
3242 		return;
3243 
3244 	ddt_enter(ddt);
3245 
3246 	ASSERT(dde->dde_lead_zio[p] == zio);
3247 
3248 	ddt_phys_fill(ddp, zio->io_bp);
3249 
3250 	zio_link_t *zl = NULL;
3251 	while ((pio = zio_walk_parents(zio, &zl)) != NULL)
3252 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
3253 
3254 	ddt_exit(ddt);
3255 }
3256 
3257 static void
3258 zio_ddt_child_write_done(zio_t *zio)
3259 {
3260 	int p = zio->io_prop.zp_copies;
3261 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3262 	ddt_entry_t *dde = zio->io_private;
3263 	ddt_phys_t *ddp = &dde->dde_phys[p];
3264 
3265 	ddt_enter(ddt);
3266 
3267 	ASSERT(ddp->ddp_refcnt == 0);
3268 	ASSERT(dde->dde_lead_zio[p] == zio);
3269 	dde->dde_lead_zio[p] = NULL;
3270 
3271 	if (zio->io_error == 0) {
3272 		zio_link_t *zl = NULL;
3273 		while (zio_walk_parents(zio, &zl) != NULL)
3274 			ddt_phys_addref(ddp);
3275 	} else {
3276 		ddt_phys_clear(ddp);
3277 	}
3278 
3279 	ddt_exit(ddt);
3280 }
3281 
3282 static zio_t *
3283 zio_ddt_write(zio_t *zio)
3284 {
3285 	spa_t *spa = zio->io_spa;
3286 	blkptr_t *bp = zio->io_bp;
3287 	uint64_t txg = zio->io_txg;
3288 	zio_prop_t *zp = &zio->io_prop;
3289 	int p = zp->zp_copies;
3290 	zio_t *cio = NULL;
3291 	ddt_t *ddt = ddt_select(spa, bp);
3292 	ddt_entry_t *dde;
3293 	ddt_phys_t *ddp;
3294 
3295 	ASSERT(BP_GET_DEDUP(bp));
3296 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
3297 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
3298 	ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
3299 
3300 	ddt_enter(ddt);
3301 	dde = ddt_lookup(ddt, bp, B_TRUE);
3302 	ddp = &dde->dde_phys[p];
3303 
3304 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
3305 		/*
3306 		 * If we're using a weak checksum, upgrade to a strong checksum
3307 		 * and try again.  If we're already using a strong checksum,
3308 		 * we can't resolve it, so just convert to an ordinary write.
3309 		 * (And automatically e-mail a paper to Nature?)
3310 		 */
3311 		if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
3312 		    ZCHECKSUM_FLAG_DEDUP)) {
3313 			zp->zp_checksum = spa_dedup_checksum(spa);
3314 			zio_pop_transforms(zio);
3315 			zio->io_stage = ZIO_STAGE_OPEN;
3316 			BP_ZERO(bp);
3317 		} else {
3318 			zp->zp_dedup = B_FALSE;
3319 			BP_SET_DEDUP(bp, B_FALSE);
3320 		}
3321 		ASSERT(!BP_GET_DEDUP(bp));
3322 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
3323 		ddt_exit(ddt);
3324 		return (zio);
3325 	}
3326 
3327 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
3328 		if (ddp->ddp_phys_birth != 0)
3329 			ddt_bp_fill(ddp, bp, txg);
3330 		if (dde->dde_lead_zio[p] != NULL)
3331 			zio_add_child(zio, dde->dde_lead_zio[p]);
3332 		else
3333 			ddt_phys_addref(ddp);
3334 	} else if (zio->io_bp_override) {
3335 		ASSERT(bp->blk_birth == txg);
3336 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
3337 		ddt_phys_fill(ddp, bp);
3338 		ddt_phys_addref(ddp);
3339 	} else {
3340 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3341 		    zio->io_orig_size, zio->io_orig_size, zp,
3342 		    zio_ddt_child_write_ready, NULL, NULL,
3343 		    zio_ddt_child_write_done, dde, zio->io_priority,
3344 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3345 
3346 		zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
3347 		dde->dde_lead_zio[p] = cio;
3348 	}
3349 
3350 	ddt_exit(ddt);
3351 
3352 	zio_nowait(cio);
3353 
3354 	return (zio);
3355 }
3356 
3357 ddt_entry_t *freedde; /* for debugging */
3358 
3359 static zio_t *
3360 zio_ddt_free(zio_t *zio)
3361 {
3362 	spa_t *spa = zio->io_spa;
3363 	blkptr_t *bp = zio->io_bp;
3364 	ddt_t *ddt = ddt_select(spa, bp);
3365 	ddt_entry_t *dde;
3366 	ddt_phys_t *ddp;
3367 
3368 	ASSERT(BP_GET_DEDUP(bp));
3369 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3370 
3371 	ddt_enter(ddt);
3372 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
3373 	if (dde) {
3374 		ddp = ddt_phys_select(dde, bp);
3375 		if (ddp)
3376 			ddt_phys_decref(ddp);
3377 	}
3378 	ddt_exit(ddt);
3379 
3380 	return (zio);
3381 }
3382 
3383 /*
3384  * ==========================================================================
3385  * Allocate and free blocks
3386  * ==========================================================================
3387  */
3388 
3389 static zio_t *
3390 zio_io_to_allocate(spa_t *spa, int allocator)
3391 {
3392 	zio_t *zio;
3393 
3394 	ASSERT(MUTEX_HELD(&spa->spa_allocs[allocator].spaa_lock));
3395 
3396 	zio = avl_first(&spa->spa_allocs[allocator].spaa_tree);
3397 	if (zio == NULL)
3398 		return (NULL);
3399 
3400 	ASSERT(IO_IS_ALLOCATING(zio));
3401 
3402 	/*
3403 	 * Try to place a reservation for this zio. If we're unable to
3404 	 * reserve then we throttle.
3405 	 */
3406 	ASSERT3U(zio->io_allocator, ==, allocator);
3407 	if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
3408 	    zio->io_prop.zp_copies, allocator, zio, 0)) {
3409 		return (NULL);
3410 	}
3411 
3412 	avl_remove(&spa->spa_allocs[allocator].spaa_tree, zio);
3413 	ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
3414 
3415 	return (zio);
3416 }
3417 
3418 static zio_t *
3419 zio_dva_throttle(zio_t *zio)
3420 {
3421 	spa_t *spa = zio->io_spa;
3422 	zio_t *nio;
3423 	metaslab_class_t *mc;
3424 
3425 	/* locate an appropriate allocation class */
3426 	mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type,
3427 	    zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk);
3428 
3429 	if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
3430 	    !mc->mc_alloc_throttle_enabled ||
3431 	    zio->io_child_type == ZIO_CHILD_GANG ||
3432 	    zio->io_flags & ZIO_FLAG_NODATA) {
3433 		return (zio);
3434 	}
3435 
3436 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3437 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3438 	ASSERT3U(zio->io_queued_timestamp, >, 0);
3439 	ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3440 
3441 	zbookmark_phys_t *bm = &zio->io_bookmark;
3442 	/*
3443 	 * We want to try to use as many allocators as possible to help improve
3444 	 * performance, but we also want logically adjacent IOs to be physically
3445 	 * adjacent to improve sequential read performance. We chunk each object
3446 	 * into 2^20 block regions, and then hash based on the objset, object,
3447 	 * level, and region to accomplish both of these goals.
3448 	 */
3449 	int allocator = (uint_t)cityhash4(bm->zb_objset, bm->zb_object,
3450 	    bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
3451 	zio->io_allocator = allocator;
3452 	zio->io_metaslab_class = mc;
3453 	mutex_enter(&spa->spa_allocs[allocator].spaa_lock);
3454 	avl_add(&spa->spa_allocs[allocator].spaa_tree, zio);
3455 	nio = zio_io_to_allocate(spa, allocator);
3456 	mutex_exit(&spa->spa_allocs[allocator].spaa_lock);
3457 	return (nio);
3458 }
3459 
3460 static void
3461 zio_allocate_dispatch(spa_t *spa, int allocator)
3462 {
3463 	zio_t *zio;
3464 
3465 	mutex_enter(&spa->spa_allocs[allocator].spaa_lock);
3466 	zio = zio_io_to_allocate(spa, allocator);
3467 	mutex_exit(&spa->spa_allocs[allocator].spaa_lock);
3468 	if (zio == NULL)
3469 		return;
3470 
3471 	ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3472 	ASSERT0(zio->io_error);
3473 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3474 }
3475 
3476 static zio_t *
3477 zio_dva_allocate(zio_t *zio)
3478 {
3479 	spa_t *spa = zio->io_spa;
3480 	metaslab_class_t *mc;
3481 	blkptr_t *bp = zio->io_bp;
3482 	int error;
3483 	int flags = 0;
3484 
3485 	if (zio->io_gang_leader == NULL) {
3486 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3487 		zio->io_gang_leader = zio;
3488 	}
3489 
3490 	ASSERT(BP_IS_HOLE(bp));
3491 	ASSERT0(BP_GET_NDVAS(bp));
3492 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
3493 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3494 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3495 
3496 	flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
3497 	if (zio->io_flags & ZIO_FLAG_NODATA)
3498 		flags |= METASLAB_DONT_THROTTLE;
3499 	if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
3500 		flags |= METASLAB_GANG_CHILD;
3501 	if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
3502 		flags |= METASLAB_ASYNC_ALLOC;
3503 
3504 	/*
3505 	 * if not already chosen, locate an appropriate allocation class
3506 	 */
3507 	mc = zio->io_metaslab_class;
3508 	if (mc == NULL) {
3509 		mc = spa_preferred_class(spa, zio->io_size,
3510 		    zio->io_prop.zp_type, zio->io_prop.zp_level,
3511 		    zio->io_prop.zp_zpl_smallblk);
3512 		zio->io_metaslab_class = mc;
3513 	}
3514 
3515 	/*
3516 	 * Try allocating the block in the usual metaslab class.
3517 	 * If that's full, allocate it in the normal class.
3518 	 * If that's full, allocate as a gang block,
3519 	 * and if all are full, the allocation fails (which shouldn't happen).
3520 	 *
3521 	 * Note that we do not fall back on embedded slog (ZIL) space, to
3522 	 * preserve unfragmented slog space, which is critical for decent
3523 	 * sync write performance.  If a log allocation fails, we will fall
3524 	 * back to spa_sync() which is abysmal for performance.
3525 	 */
3526 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
3527 	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3528 	    &zio->io_alloc_list, zio, zio->io_allocator);
3529 
3530 	/*
3531 	 * Fallback to normal class when an alloc class is full
3532 	 */
3533 	if (error == ENOSPC && mc != spa_normal_class(spa)) {
3534 		/*
3535 		 * If throttling, transfer reservation over to normal class.
3536 		 * The io_allocator slot can remain the same even though we
3537 		 * are switching classes.
3538 		 */
3539 		if (mc->mc_alloc_throttle_enabled &&
3540 		    (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) {
3541 			metaslab_class_throttle_unreserve(mc,
3542 			    zio->io_prop.zp_copies, zio->io_allocator, zio);
3543 			zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING;
3544 
3545 			VERIFY(metaslab_class_throttle_reserve(
3546 			    spa_normal_class(spa),
3547 			    zio->io_prop.zp_copies, zio->io_allocator, zio,
3548 			    flags | METASLAB_MUST_RESERVE));
3549 		}
3550 		zio->io_metaslab_class = mc = spa_normal_class(spa);
3551 		if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) {
3552 			zfs_dbgmsg("%s: metaslab allocation failure, "
3553 			    "trying normal class: zio %px, size %llu, error %d",
3554 			    spa_name(spa), zio, (u_longlong_t)zio->io_size,
3555 			    error);
3556 		}
3557 
3558 		error = metaslab_alloc(spa, mc, zio->io_size, bp,
3559 		    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3560 		    &zio->io_alloc_list, zio, zio->io_allocator);
3561 	}
3562 
3563 	if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) {
3564 		if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) {
3565 			zfs_dbgmsg("%s: metaslab allocation failure, "
3566 			    "trying ganging: zio %px, size %llu, error %d",
3567 			    spa_name(spa), zio, (u_longlong_t)zio->io_size,
3568 			    error);
3569 		}
3570 		return (zio_write_gang_block(zio, mc));
3571 	}
3572 	if (error != 0) {
3573 		if (error != ENOSPC ||
3574 		    (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC)) {
3575 			zfs_dbgmsg("%s: metaslab allocation failure: zio %px, "
3576 			    "size %llu, error %d",
3577 			    spa_name(spa), zio, (u_longlong_t)zio->io_size,
3578 			    error);
3579 		}
3580 		zio->io_error = error;
3581 	}
3582 
3583 	return (zio);
3584 }
3585 
3586 static zio_t *
3587 zio_dva_free(zio_t *zio)
3588 {
3589 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3590 
3591 	return (zio);
3592 }
3593 
3594 static zio_t *
3595 zio_dva_claim(zio_t *zio)
3596 {
3597 	int error;
3598 
3599 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3600 	if (error)
3601 		zio->io_error = error;
3602 
3603 	return (zio);
3604 }
3605 
3606 /*
3607  * Undo an allocation.  This is used by zio_done() when an I/O fails
3608  * and we want to give back the block we just allocated.
3609  * This handles both normal blocks and gang blocks.
3610  */
3611 static void
3612 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3613 {
3614 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3615 	ASSERT(zio->io_bp_override == NULL);
3616 
3617 	if (!BP_IS_HOLE(bp))
3618 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3619 
3620 	if (gn != NULL) {
3621 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3622 			zio_dva_unallocate(zio, gn->gn_child[g],
3623 			    &gn->gn_gbh->zg_blkptr[g]);
3624 		}
3625 	}
3626 }
3627 
3628 /*
3629  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
3630  */
3631 int
3632 zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
3633     uint64_t size, boolean_t *slog)
3634 {
3635 	int error = 1;
3636 	zio_alloc_list_t io_alloc_list;
3637 
3638 	ASSERT(txg > spa_syncing_txg(spa));
3639 
3640 	metaslab_trace_init(&io_alloc_list);
3641 
3642 	/*
3643 	 * Block pointer fields are useful to metaslabs for stats and debugging.
3644 	 * Fill in the obvious ones before calling into metaslab_alloc().
3645 	 */
3646 	BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3647 	BP_SET_PSIZE(new_bp, size);
3648 	BP_SET_LEVEL(new_bp, 0);
3649 
3650 	/*
3651 	 * When allocating a zil block, we don't have information about
3652 	 * the final destination of the block except the objset it's part
3653 	 * of, so we just hash the objset ID to pick the allocator to get
3654 	 * some parallelism.
3655 	 */
3656 	int flags = METASLAB_FASTWRITE | METASLAB_ZIL;
3657 	int allocator = (uint_t)cityhash4(0, 0, 0,
3658 	    os->os_dsl_dataset->ds_object) % spa->spa_alloc_count;
3659 	error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3660 	    txg, NULL, flags, &io_alloc_list, NULL, allocator);
3661 	*slog = (error == 0);
3662 	if (error != 0) {
3663 		error = metaslab_alloc(spa, spa_embedded_log_class(spa), size,
3664 		    new_bp, 1, txg, NULL, flags,
3665 		    &io_alloc_list, NULL, allocator);
3666 	}
3667 	if (error != 0) {
3668 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
3669 		    new_bp, 1, txg, NULL, flags,
3670 		    &io_alloc_list, NULL, allocator);
3671 	}
3672 	metaslab_trace_fini(&io_alloc_list);
3673 
3674 	if (error == 0) {
3675 		BP_SET_LSIZE(new_bp, size);
3676 		BP_SET_PSIZE(new_bp, size);
3677 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3678 		BP_SET_CHECKSUM(new_bp,
3679 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3680 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3681 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3682 		BP_SET_LEVEL(new_bp, 0);
3683 		BP_SET_DEDUP(new_bp, 0);
3684 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3685 
3686 		/*
3687 		 * encrypted blocks will require an IV and salt. We generate
3688 		 * these now since we will not be rewriting the bp at
3689 		 * rewrite time.
3690 		 */
3691 		if (os->os_encrypted) {
3692 			uint8_t iv[ZIO_DATA_IV_LEN];
3693 			uint8_t salt[ZIO_DATA_SALT_LEN];
3694 
3695 			BP_SET_CRYPT(new_bp, B_TRUE);
3696 			VERIFY0(spa_crypt_get_salt(spa,
3697 			    dmu_objset_id(os), salt));
3698 			VERIFY0(zio_crypt_generate_iv(iv));
3699 
3700 			zio_crypt_encode_params_bp(new_bp, salt, iv);
3701 		}
3702 	} else {
3703 		zfs_dbgmsg("%s: zil block allocation failure: "
3704 		    "size %llu, error %d", spa_name(spa), (u_longlong_t)size,
3705 		    error);
3706 	}
3707 
3708 	return (error);
3709 }
3710 
3711 /*
3712  * ==========================================================================
3713  * Read and write to physical devices
3714  * ==========================================================================
3715  */
3716 
3717 /*
3718  * Issue an I/O to the underlying vdev. Typically the issue pipeline
3719  * stops after this stage and will resume upon I/O completion.
3720  * However, there are instances where the vdev layer may need to
3721  * continue the pipeline when an I/O was not issued. Since the I/O
3722  * that was sent to the vdev layer might be different than the one
3723  * currently active in the pipeline (see vdev_queue_io()), we explicitly
3724  * force the underlying vdev layers to call either zio_execute() or
3725  * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3726  */
3727 static zio_t *
3728 zio_vdev_io_start(zio_t *zio)
3729 {
3730 	vdev_t *vd = zio->io_vd;
3731 	uint64_t align;
3732 	spa_t *spa = zio->io_spa;
3733 
3734 	zio->io_delay = 0;
3735 
3736 	ASSERT(zio->io_error == 0);
3737 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3738 
3739 	if (vd == NULL) {
3740 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3741 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3742 
3743 		/*
3744 		 * The mirror_ops handle multiple DVAs in a single BP.
3745 		 */
3746 		vdev_mirror_ops.vdev_op_io_start(zio);
3747 		return (NULL);
3748 	}
3749 
3750 	ASSERT3P(zio->io_logical, !=, zio);
3751 	if (zio->io_type == ZIO_TYPE_WRITE) {
3752 		ASSERT(spa->spa_trust_config);
3753 
3754 		/*
3755 		 * Note: the code can handle other kinds of writes,
3756 		 * but we don't expect them.
3757 		 */
3758 		if (zio->io_vd->vdev_removing) {
3759 			ASSERT(zio->io_flags &
3760 			    (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3761 			    ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3762 		}
3763 	}
3764 
3765 	align = 1ULL << vd->vdev_top->vdev_ashift;
3766 
3767 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3768 	    P2PHASE(zio->io_size, align) != 0) {
3769 		/* Transform logical writes to be a full physical block size. */
3770 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
3771 		abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3772 		ASSERT(vd == vd->vdev_top);
3773 		if (zio->io_type == ZIO_TYPE_WRITE) {
3774 			abd_copy(abuf, zio->io_abd, zio->io_size);
3775 			abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3776 		}
3777 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3778 	}
3779 
3780 	/*
3781 	 * If this is not a physical io, make sure that it is properly aligned
3782 	 * before proceeding.
3783 	 */
3784 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3785 		ASSERT0(P2PHASE(zio->io_offset, align));
3786 		ASSERT0(P2PHASE(zio->io_size, align));
3787 	} else {
3788 		/*
3789 		 * For physical writes, we allow 512b aligned writes and assume
3790 		 * the device will perform a read-modify-write as necessary.
3791 		 */
3792 		ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3793 		ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3794 	}
3795 
3796 	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
3797 
3798 	/*
3799 	 * If this is a repair I/O, and there's no self-healing involved --
3800 	 * that is, we're just resilvering what we expect to resilver --
3801 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3802 	 * This prevents spurious resilvering.
3803 	 *
3804 	 * There are a few ways that we can end up creating these spurious
3805 	 * resilver i/os:
3806 	 *
3807 	 * 1. A resilver i/o will be issued if any DVA in the BP has a
3808 	 * dirty DTL.  The mirror code will issue resilver writes to
3809 	 * each DVA, including the one(s) that are not on vdevs with dirty
3810 	 * DTLs.
3811 	 *
3812 	 * 2. With nested replication, which happens when we have a
3813 	 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3814 	 * For example, given mirror(replacing(A+B), C), it's likely that
3815 	 * only A is out of date (it's the new device). In this case, we'll
3816 	 * read from C, then use the data to resilver A+B -- but we don't
3817 	 * actually want to resilver B, just A. The top-level mirror has no
3818 	 * way to know this, so instead we just discard unnecessary repairs
3819 	 * as we work our way down the vdev tree.
3820 	 *
3821 	 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3822 	 * The same logic applies to any form of nested replication: ditto
3823 	 * + mirror, RAID-Z + replacing, etc.
3824 	 *
3825 	 * However, indirect vdevs point off to other vdevs which may have
3826 	 * DTL's, so we never bypass them.  The child i/os on concrete vdevs
3827 	 * will be properly bypassed instead.
3828 	 *
3829 	 * Leaf DTL_PARTIAL can be empty when a legitimate write comes from
3830 	 * a dRAID spare vdev. For example, when a dRAID spare is first
3831 	 * used, its spare blocks need to be written to but the leaf vdev's
3832 	 * of such blocks can have empty DTL_PARTIAL.
3833 	 *
3834 	 * There seemed no clean way to allow such writes while bypassing
3835 	 * spurious ones. At this point, just avoid all bypassing for dRAID
3836 	 * for correctness.
3837 	 */
3838 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3839 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3840 	    zio->io_txg != 0 &&	/* not a delegated i/o */
3841 	    vd->vdev_ops != &vdev_indirect_ops &&
3842 	    vd->vdev_top->vdev_ops != &vdev_draid_ops &&
3843 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3844 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3845 		zio_vdev_io_bypass(zio);
3846 		return (zio);
3847 	}
3848 
3849 	/*
3850 	 * Select the next best leaf I/O to process.  Distributed spares are
3851 	 * excluded since they dispatch the I/O directly to a leaf vdev after
3852 	 * applying the dRAID mapping.
3853 	 */
3854 	if (vd->vdev_ops->vdev_op_leaf &&
3855 	    vd->vdev_ops != &vdev_draid_spare_ops &&
3856 	    (zio->io_type == ZIO_TYPE_READ ||
3857 	    zio->io_type == ZIO_TYPE_WRITE ||
3858 	    zio->io_type == ZIO_TYPE_TRIM)) {
3859 
3860 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
3861 			return (zio);
3862 
3863 		if ((zio = vdev_queue_io(zio)) == NULL)
3864 			return (NULL);
3865 
3866 		if (!vdev_accessible(vd, zio)) {
3867 			zio->io_error = SET_ERROR(ENXIO);
3868 			zio_interrupt(zio);
3869 			return (NULL);
3870 		}
3871 		zio->io_delay = gethrtime();
3872 	}
3873 
3874 	vd->vdev_ops->vdev_op_io_start(zio);
3875 	return (NULL);
3876 }
3877 
3878 static zio_t *
3879 zio_vdev_io_done(zio_t *zio)
3880 {
3881 	vdev_t *vd = zio->io_vd;
3882 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3883 	boolean_t unexpected_error = B_FALSE;
3884 
3885 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3886 		return (NULL);
3887 	}
3888 
3889 	ASSERT(zio->io_type == ZIO_TYPE_READ ||
3890 	    zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM);
3891 
3892 	if (zio->io_delay)
3893 		zio->io_delay = gethrtime() - zio->io_delay;
3894 
3895 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3896 	    vd->vdev_ops != &vdev_draid_spare_ops) {
3897 		vdev_queue_io_done(zio);
3898 
3899 		if (zio->io_type == ZIO_TYPE_WRITE)
3900 			vdev_cache_write(zio);
3901 
3902 		if (zio_injection_enabled && zio->io_error == 0)
3903 			zio->io_error = zio_handle_device_injections(vd, zio,
3904 			    EIO, EILSEQ);
3905 
3906 		if (zio_injection_enabled && zio->io_error == 0)
3907 			zio->io_error = zio_handle_label_injection(zio, EIO);
3908 
3909 		if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) {
3910 			if (!vdev_accessible(vd, zio)) {
3911 				zio->io_error = SET_ERROR(ENXIO);
3912 			} else {
3913 				unexpected_error = B_TRUE;
3914 			}
3915 		}
3916 	}
3917 
3918 	ops->vdev_op_io_done(zio);
3919 
3920 	if (unexpected_error)
3921 		VERIFY(vdev_probe(vd, zio) == NULL);
3922 
3923 	return (zio);
3924 }
3925 
3926 /*
3927  * This function is used to change the priority of an existing zio that is
3928  * currently in-flight. This is used by the arc to upgrade priority in the
3929  * event that a demand read is made for a block that is currently queued
3930  * as a scrub or async read IO. Otherwise, the high priority read request
3931  * would end up having to wait for the lower priority IO.
3932  */
3933 void
3934 zio_change_priority(zio_t *pio, zio_priority_t priority)
3935 {
3936 	zio_t *cio, *cio_next;
3937 	zio_link_t *zl = NULL;
3938 
3939 	ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3940 
3941 	if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3942 		vdev_queue_change_io_priority(pio, priority);
3943 	} else {
3944 		pio->io_priority = priority;
3945 	}
3946 
3947 	mutex_enter(&pio->io_lock);
3948 	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3949 		cio_next = zio_walk_children(pio, &zl);
3950 		zio_change_priority(cio, priority);
3951 	}
3952 	mutex_exit(&pio->io_lock);
3953 }
3954 
3955 /*
3956  * For non-raidz ZIOs, we can just copy aside the bad data read from the
3957  * disk, and use that to finish the checksum ereport later.
3958  */
3959 static void
3960 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3961     const abd_t *good_buf)
3962 {
3963 	/* no processing needed */
3964 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3965 }
3966 
3967 /*ARGSUSED*/
3968 void
3969 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr)
3970 {
3971 	void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size);
3972 
3973 	abd_copy(abd, zio->io_abd, zio->io_size);
3974 
3975 	zcr->zcr_cbinfo = zio->io_size;
3976 	zcr->zcr_cbdata = abd;
3977 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
3978 	zcr->zcr_free = zio_abd_free;
3979 }
3980 
3981 static zio_t *
3982 zio_vdev_io_assess(zio_t *zio)
3983 {
3984 	vdev_t *vd = zio->io_vd;
3985 
3986 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3987 		return (NULL);
3988 	}
3989 
3990 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3991 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3992 
3993 	if (zio->io_vsd != NULL) {
3994 		zio->io_vsd_ops->vsd_free(zio);
3995 		zio->io_vsd = NULL;
3996 	}
3997 
3998 	if (zio_injection_enabled && zio->io_error == 0)
3999 		zio->io_error = zio_handle_fault_injection(zio, EIO);
4000 
4001 	/*
4002 	 * If the I/O failed, determine whether we should attempt to retry it.
4003 	 *
4004 	 * On retry, we cut in line in the issue queue, since we don't want
4005 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
4006 	 */
4007 	if (zio->io_error && vd == NULL &&
4008 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
4009 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
4010 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
4011 		zio->io_error = 0;
4012 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
4013 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
4014 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
4015 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
4016 		    zio_requeue_io_start_cut_in_line);
4017 		return (NULL);
4018 	}
4019 
4020 	/*
4021 	 * If we got an error on a leaf device, convert it to ENXIO
4022 	 * if the device is not accessible at all.
4023 	 */
4024 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
4025 	    !vdev_accessible(vd, zio))
4026 		zio->io_error = SET_ERROR(ENXIO);
4027 
4028 	/*
4029 	 * If we can't write to an interior vdev (mirror or RAID-Z),
4030 	 * set vdev_cant_write so that we stop trying to allocate from it.
4031 	 */
4032 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
4033 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
4034 		vdev_dbgmsg(vd, "zio_vdev_io_assess(zio=%px) setting "
4035 		    "cant_write=TRUE due to write failure with ENXIO",
4036 		    zio);
4037 		vd->vdev_cant_write = B_TRUE;
4038 	}
4039 
4040 	/*
4041 	 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
4042 	 * attempts will ever succeed. In this case we set a persistent
4043 	 * boolean flag so that we don't bother with it in the future.
4044 	 */
4045 	if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
4046 	    zio->io_type == ZIO_TYPE_IOCTL &&
4047 	    zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
4048 		vd->vdev_nowritecache = B_TRUE;
4049 
4050 	if (zio->io_error)
4051 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4052 
4053 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
4054 	    zio->io_physdone != NULL) {
4055 		ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
4056 		ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
4057 		zio->io_physdone(zio->io_logical);
4058 	}
4059 
4060 	return (zio);
4061 }
4062 
4063 void
4064 zio_vdev_io_reissue(zio_t *zio)
4065 {
4066 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4067 	ASSERT(zio->io_error == 0);
4068 
4069 	zio->io_stage >>= 1;
4070 }
4071 
4072 void
4073 zio_vdev_io_redone(zio_t *zio)
4074 {
4075 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
4076 
4077 	zio->io_stage >>= 1;
4078 }
4079 
4080 void
4081 zio_vdev_io_bypass(zio_t *zio)
4082 {
4083 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4084 	ASSERT(zio->io_error == 0);
4085 
4086 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
4087 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
4088 }
4089 
4090 /*
4091  * ==========================================================================
4092  * Encrypt and store encryption parameters
4093  * ==========================================================================
4094  */
4095 
4096 
4097 /*
4098  * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
4099  * managing the storage of encryption parameters and passing them to the
4100  * lower-level encryption functions.
4101  */
4102 static zio_t *
4103 zio_encrypt(zio_t *zio)
4104 {
4105 	zio_prop_t *zp = &zio->io_prop;
4106 	spa_t *spa = zio->io_spa;
4107 	blkptr_t *bp = zio->io_bp;
4108 	uint64_t psize = BP_GET_PSIZE(bp);
4109 	uint64_t dsobj = zio->io_bookmark.zb_objset;
4110 	dmu_object_type_t ot = BP_GET_TYPE(bp);
4111 	void *enc_buf = NULL;
4112 	abd_t *eabd = NULL;
4113 	uint8_t salt[ZIO_DATA_SALT_LEN];
4114 	uint8_t iv[ZIO_DATA_IV_LEN];
4115 	uint8_t mac[ZIO_DATA_MAC_LEN];
4116 	boolean_t no_crypt = B_FALSE;
4117 
4118 	/* the root zio already encrypted the data */
4119 	if (zio->io_child_type == ZIO_CHILD_GANG)
4120 		return (zio);
4121 
4122 	/* only ZIL blocks are re-encrypted on rewrite */
4123 	if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG)
4124 		return (zio);
4125 
4126 	if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) {
4127 		BP_SET_CRYPT(bp, B_FALSE);
4128 		return (zio);
4129 	}
4130 
4131 	/* if we are doing raw encryption set the provided encryption params */
4132 	if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) {
4133 		ASSERT0(BP_GET_LEVEL(bp));
4134 		BP_SET_CRYPT(bp, B_TRUE);
4135 		BP_SET_BYTEORDER(bp, zp->zp_byteorder);
4136 		if (ot != DMU_OT_OBJSET)
4137 			zio_crypt_encode_mac_bp(bp, zp->zp_mac);
4138 
4139 		/* dnode blocks must be written out in the provided byteorder */
4140 		if (zp->zp_byteorder != ZFS_HOST_BYTEORDER &&
4141 		    ot == DMU_OT_DNODE) {
4142 			void *bswap_buf = zio_buf_alloc(psize);
4143 			abd_t *babd = abd_get_from_buf(bswap_buf, psize);
4144 
4145 			ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4146 			abd_copy_to_buf(bswap_buf, zio->io_abd, psize);
4147 			dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf,
4148 			    psize);
4149 
4150 			abd_take_ownership_of_buf(babd, B_TRUE);
4151 			zio_push_transform(zio, babd, psize, psize, NULL);
4152 		}
4153 
4154 		if (DMU_OT_IS_ENCRYPTED(ot))
4155 			zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv);
4156 		return (zio);
4157 	}
4158 
4159 	/* indirect blocks only maintain a cksum of the lower level MACs */
4160 	if (BP_GET_LEVEL(bp) > 0) {
4161 		BP_SET_CRYPT(bp, B_TRUE);
4162 		VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE,
4163 		    zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp),
4164 		    mac));
4165 		zio_crypt_encode_mac_bp(bp, mac);
4166 		return (zio);
4167 	}
4168 
4169 	/*
4170 	 * Objset blocks are a special case since they have 2 256-bit MACs
4171 	 * embedded within them.
4172 	 */
4173 	if (ot == DMU_OT_OBJSET) {
4174 		ASSERT0(DMU_OT_IS_ENCRYPTED(ot));
4175 		ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4176 		BP_SET_CRYPT(bp, B_TRUE);
4177 		VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj,
4178 		    zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp)));
4179 		return (zio);
4180 	}
4181 
4182 	/* unencrypted object types are only authenticated with a MAC */
4183 	if (!DMU_OT_IS_ENCRYPTED(ot)) {
4184 		BP_SET_CRYPT(bp, B_TRUE);
4185 		VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj,
4186 		    zio->io_abd, psize, mac));
4187 		zio_crypt_encode_mac_bp(bp, mac);
4188 		return (zio);
4189 	}
4190 
4191 	/*
4192 	 * Later passes of sync-to-convergence may decide to rewrite data
4193 	 * in place to avoid more disk reallocations. This presents a problem
4194 	 * for encryption because this constitutes rewriting the new data with
4195 	 * the same encryption key and IV. However, this only applies to blocks
4196 	 * in the MOS (particularly the spacemaps) and we do not encrypt the
4197 	 * MOS. We assert that the zio is allocating or an intent log write
4198 	 * to enforce this.
4199 	 */
4200 	ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG);
4201 	ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG);
4202 	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION));
4203 	ASSERT3U(psize, !=, 0);
4204 
4205 	enc_buf = zio_buf_alloc(psize);
4206 	eabd = abd_get_from_buf(enc_buf, psize);
4207 	abd_take_ownership_of_buf(eabd, B_TRUE);
4208 
4209 	/*
4210 	 * For an explanation of what encryption parameters are stored
4211 	 * where, see the block comment in zio_crypt.c.
4212 	 */
4213 	if (ot == DMU_OT_INTENT_LOG) {
4214 		zio_crypt_decode_params_bp(bp, salt, iv);
4215 	} else {
4216 		BP_SET_CRYPT(bp, B_TRUE);
4217 	}
4218 
4219 	/* Perform the encryption. This should not fail */
4220 	VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark,
4221 	    BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
4222 	    salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt));
4223 
4224 	/* encode encryption metadata into the bp */
4225 	if (ot == DMU_OT_INTENT_LOG) {
4226 		/*
4227 		 * ZIL blocks store the MAC in the embedded checksum, so the
4228 		 * transform must always be applied.
4229 		 */
4230 		zio_crypt_encode_mac_zil(enc_buf, mac);
4231 		zio_push_transform(zio, eabd, psize, psize, NULL);
4232 	} else {
4233 		BP_SET_CRYPT(bp, B_TRUE);
4234 		zio_crypt_encode_params_bp(bp, salt, iv);
4235 		zio_crypt_encode_mac_bp(bp, mac);
4236 
4237 		if (no_crypt) {
4238 			ASSERT3U(ot, ==, DMU_OT_DNODE);
4239 			abd_free(eabd);
4240 		} else {
4241 			zio_push_transform(zio, eabd, psize, psize, NULL);
4242 		}
4243 	}
4244 
4245 	return (zio);
4246 }
4247 
4248 /*
4249  * ==========================================================================
4250  * Generate and verify checksums
4251  * ==========================================================================
4252  */
4253 static zio_t *
4254 zio_checksum_generate(zio_t *zio)
4255 {
4256 	blkptr_t *bp = zio->io_bp;
4257 	enum zio_checksum checksum;
4258 
4259 	if (bp == NULL) {
4260 		/*
4261 		 * This is zio_write_phys().
4262 		 * We're either generating a label checksum, or none at all.
4263 		 */
4264 		checksum = zio->io_prop.zp_checksum;
4265 
4266 		if (checksum == ZIO_CHECKSUM_OFF)
4267 			return (zio);
4268 
4269 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
4270 	} else {
4271 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
4272 			ASSERT(!IO_IS_ALLOCATING(zio));
4273 			checksum = ZIO_CHECKSUM_GANG_HEADER;
4274 		} else {
4275 			checksum = BP_GET_CHECKSUM(bp);
4276 		}
4277 	}
4278 
4279 	zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
4280 
4281 	return (zio);
4282 }
4283 
4284 static zio_t *
4285 zio_checksum_verify(zio_t *zio)
4286 {
4287 	zio_bad_cksum_t info;
4288 	blkptr_t *bp = zio->io_bp;
4289 	int error;
4290 
4291 	ASSERT(zio->io_vd != NULL);
4292 
4293 	if (bp == NULL) {
4294 		/*
4295 		 * This is zio_read_phys().
4296 		 * We're either verifying a label checksum, or nothing at all.
4297 		 */
4298 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
4299 			return (zio);
4300 
4301 		ASSERT3U(zio->io_prop.zp_checksum, ==, ZIO_CHECKSUM_LABEL);
4302 	}
4303 
4304 	if ((error = zio_checksum_error(zio, &info)) != 0) {
4305 		zio->io_error = error;
4306 		if (error == ECKSUM &&
4307 		    !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
4308 			(void) zfs_ereport_start_checksum(zio->io_spa,
4309 			    zio->io_vd, &zio->io_bookmark, zio,
4310 			    zio->io_offset, zio->io_size, &info);
4311 			mutex_enter(&zio->io_vd->vdev_stat_lock);
4312 			zio->io_vd->vdev_stat.vs_checksum_errors++;
4313 			mutex_exit(&zio->io_vd->vdev_stat_lock);
4314 		}
4315 	}
4316 
4317 	return (zio);
4318 }
4319 
4320 /*
4321  * Called by RAID-Z to ensure we don't compute the checksum twice.
4322  */
4323 void
4324 zio_checksum_verified(zio_t *zio)
4325 {
4326 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
4327 }
4328 
4329 /*
4330  * ==========================================================================
4331  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4332  * An error of 0 indicates success.  ENXIO indicates whole-device failure,
4333  * which may be transient (e.g. unplugged) or permanent.  ECKSUM and EIO
4334  * indicate errors that are specific to one I/O, and most likely permanent.
4335  * Any other error is presumed to be worse because we weren't expecting it.
4336  * ==========================================================================
4337  */
4338 int
4339 zio_worst_error(int e1, int e2)
4340 {
4341 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
4342 	int r1, r2;
4343 
4344 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
4345 		if (e1 == zio_error_rank[r1])
4346 			break;
4347 
4348 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
4349 		if (e2 == zio_error_rank[r2])
4350 			break;
4351 
4352 	return (r1 > r2 ? e1 : e2);
4353 }
4354 
4355 /*
4356  * ==========================================================================
4357  * I/O completion
4358  * ==========================================================================
4359  */
4360 static zio_t *
4361 zio_ready(zio_t *zio)
4362 {
4363 	blkptr_t *bp = zio->io_bp;
4364 	zio_t *pio, *pio_next;
4365 	zio_link_t *zl = NULL;
4366 
4367 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
4368 	    ZIO_WAIT_READY)) {
4369 		return (NULL);
4370 	}
4371 
4372 	if (zio->io_ready) {
4373 		ASSERT(IO_IS_ALLOCATING(zio));
4374 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
4375 		    (zio->io_flags & ZIO_FLAG_NOPWRITE));
4376 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
4377 
4378 		zio->io_ready(zio);
4379 	}
4380 
4381 	if (bp != NULL && bp != &zio->io_bp_copy)
4382 		zio->io_bp_copy = *bp;
4383 
4384 	if (zio->io_error != 0) {
4385 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4386 
4387 		if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4388 			ASSERT(IO_IS_ALLOCATING(zio));
4389 			ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4390 			ASSERT(zio->io_metaslab_class != NULL);
4391 
4392 			/*
4393 			 * We were unable to allocate anything, unreserve and
4394 			 * issue the next I/O to allocate.
4395 			 */
4396 			metaslab_class_throttle_unreserve(
4397 			    zio->io_metaslab_class, zio->io_prop.zp_copies,
4398 			    zio->io_allocator, zio);
4399 			zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
4400 		}
4401 	}
4402 
4403 	mutex_enter(&zio->io_lock);
4404 	zio->io_state[ZIO_WAIT_READY] = 1;
4405 	pio = zio_walk_parents(zio, &zl);
4406 	mutex_exit(&zio->io_lock);
4407 
4408 	/*
4409 	 * As we notify zio's parents, new parents could be added.
4410 	 * New parents go to the head of zio's io_parent_list, however,
4411 	 * so we will (correctly) not notify them.  The remainder of zio's
4412 	 * io_parent_list, from 'pio_next' onward, cannot change because
4413 	 * all parents must wait for us to be done before they can be done.
4414 	 */
4415 	for (; pio != NULL; pio = pio_next) {
4416 		pio_next = zio_walk_parents(zio, &zl);
4417 		zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL);
4418 	}
4419 
4420 	if (zio->io_flags & ZIO_FLAG_NODATA) {
4421 		if (BP_IS_GANG(bp)) {
4422 			zio->io_flags &= ~ZIO_FLAG_NODATA;
4423 		} else {
4424 			ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
4425 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
4426 		}
4427 	}
4428 
4429 	if (zio_injection_enabled &&
4430 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
4431 		zio_handle_ignored_writes(zio);
4432 
4433 	return (zio);
4434 }
4435 
4436 /*
4437  * Update the allocation throttle accounting.
4438  */
4439 static void
4440 zio_dva_throttle_done(zio_t *zio)
4441 {
4442 	zio_t *lio __maybe_unused = zio->io_logical;
4443 	zio_t *pio = zio_unique_parent(zio);
4444 	vdev_t *vd = zio->io_vd;
4445 	int flags = METASLAB_ASYNC_ALLOC;
4446 
4447 	ASSERT3P(zio->io_bp, !=, NULL);
4448 	ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
4449 	ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
4450 	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
4451 	ASSERT(vd != NULL);
4452 	ASSERT3P(vd, ==, vd->vdev_top);
4453 	ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY));
4454 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4455 	ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
4456 	ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
4457 	ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
4458 
4459 	/*
4460 	 * Parents of gang children can have two flavors -- ones that
4461 	 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4462 	 * and ones that allocated the constituent blocks. The allocation
4463 	 * throttle needs to know the allocating parent zio so we must find
4464 	 * it here.
4465 	 */
4466 	if (pio->io_child_type == ZIO_CHILD_GANG) {
4467 		/*
4468 		 * If our parent is a rewrite gang child then our grandparent
4469 		 * would have been the one that performed the allocation.
4470 		 */
4471 		if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
4472 			pio = zio_unique_parent(pio);
4473 		flags |= METASLAB_GANG_CHILD;
4474 	}
4475 
4476 	ASSERT(IO_IS_ALLOCATING(pio));
4477 	ASSERT3P(zio, !=, zio->io_logical);
4478 	ASSERT(zio->io_logical != NULL);
4479 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4480 	ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
4481 	ASSERT(zio->io_metaslab_class != NULL);
4482 
4483 	mutex_enter(&pio->io_lock);
4484 	metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
4485 	    pio->io_allocator, B_TRUE);
4486 	mutex_exit(&pio->io_lock);
4487 
4488 	metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1,
4489 	    pio->io_allocator, pio);
4490 
4491 	/*
4492 	 * Call into the pipeline to see if there is more work that
4493 	 * needs to be done. If there is work to be done it will be
4494 	 * dispatched to another taskq thread.
4495 	 */
4496 	zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
4497 }
4498 
4499 static zio_t *
4500 zio_done(zio_t *zio)
4501 {
4502 	/*
4503 	 * Always attempt to keep stack usage minimal here since
4504 	 * we can be called recursively up to 19 levels deep.
4505 	 */
4506 	const uint64_t psize = zio->io_size;
4507 	zio_t *pio, *pio_next;
4508 	zio_link_t *zl = NULL;
4509 
4510 	/*
4511 	 * If our children haven't all completed,
4512 	 * wait for them and then repeat this pipeline stage.
4513 	 */
4514 	if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
4515 		return (NULL);
4516 	}
4517 
4518 	/*
4519 	 * If the allocation throttle is enabled, then update the accounting.
4520 	 * We only track child I/Os that are part of an allocating async
4521 	 * write. We must do this since the allocation is performed
4522 	 * by the logical I/O but the actual write is done by child I/Os.
4523 	 */
4524 	if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
4525 	    zio->io_child_type == ZIO_CHILD_VDEV) {
4526 		ASSERT(zio->io_metaslab_class != NULL);
4527 		ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled);
4528 		zio_dva_throttle_done(zio);
4529 	}
4530 
4531 	/*
4532 	 * If the allocation throttle is enabled, verify that
4533 	 * we have decremented the refcounts for every I/O that was throttled.
4534 	 */
4535 	if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4536 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4537 		ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4538 		ASSERT(zio->io_bp != NULL);
4539 
4540 		metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio,
4541 		    zio->io_allocator);
4542 		VERIFY(zfs_refcount_not_held(&zio->io_metaslab_class->
4543 		    mc_allocator[zio->io_allocator].mca_alloc_slots, zio));
4544 	}
4545 
4546 
4547 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
4548 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
4549 			ASSERT(zio->io_children[c][w] == 0);
4550 
4551 	if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) {
4552 		ASSERT(zio->io_bp->blk_pad[0] == 0);
4553 		ASSERT(zio->io_bp->blk_pad[1] == 0);
4554 		ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy,
4555 		    sizeof (blkptr_t)) == 0 ||
4556 		    (zio->io_bp == zio_unique_parent(zio)->io_bp));
4557 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
4558 		    zio->io_bp_override == NULL &&
4559 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
4560 			ASSERT3U(zio->io_prop.zp_copies, <=,
4561 			    BP_GET_NDVAS(zio->io_bp));
4562 			ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
4563 			    (BP_COUNT_GANG(zio->io_bp) ==
4564 			    BP_GET_NDVAS(zio->io_bp)));
4565 		}
4566 		if (zio->io_flags & ZIO_FLAG_NOPWRITE)
4567 			VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig));
4568 	}
4569 
4570 	/*
4571 	 * If there were child vdev/gang/ddt errors, they apply to us now.
4572 	 */
4573 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
4574 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
4575 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
4576 
4577 	/*
4578 	 * If the I/O on the transformed data was successful, generate any
4579 	 * checksum reports now while we still have the transformed data.
4580 	 */
4581 	if (zio->io_error == 0) {
4582 		while (zio->io_cksum_report != NULL) {
4583 			zio_cksum_report_t *zcr = zio->io_cksum_report;
4584 			uint64_t align = zcr->zcr_align;
4585 			uint64_t asize = P2ROUNDUP(psize, align);
4586 			abd_t *adata = zio->io_abd;
4587 
4588 			if (adata != NULL && asize != psize) {
4589 				adata = abd_alloc(asize, B_TRUE);
4590 				abd_copy(adata, zio->io_abd, psize);
4591 				abd_zero_off(adata, psize, asize - psize);
4592 			}
4593 
4594 			zio->io_cksum_report = zcr->zcr_next;
4595 			zcr->zcr_next = NULL;
4596 			zcr->zcr_finish(zcr, adata);
4597 			zfs_ereport_free_checksum(zcr);
4598 
4599 			if (adata != NULL && asize != psize)
4600 				abd_free(adata);
4601 		}
4602 	}
4603 
4604 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
4605 
4606 	vdev_stat_update(zio, psize);
4607 
4608 	/*
4609 	 * If this I/O is attached to a particular vdev is slow, exceeding
4610 	 * 30 seconds to complete, post an error described the I/O delay.
4611 	 * We ignore these errors if the device is currently unavailable.
4612 	 */
4613 	if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) {
4614 		if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) {
4615 			/*
4616 			 * We want to only increment our slow IO counters if
4617 			 * the IO is valid (i.e. not if the drive is removed).
4618 			 *
4619 			 * zfs_ereport_post() will also do these checks, but
4620 			 * it can also ratelimit and have other failures, so we
4621 			 * need to increment the slow_io counters independent
4622 			 * of it.
4623 			 */
4624 			if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY,
4625 			    zio->io_spa, zio->io_vd, zio)) {
4626 				mutex_enter(&zio->io_vd->vdev_stat_lock);
4627 				zio->io_vd->vdev_stat.vs_slow_ios++;
4628 				mutex_exit(&zio->io_vd->vdev_stat_lock);
4629 
4630 				(void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY,
4631 				    zio->io_spa, zio->io_vd, &zio->io_bookmark,
4632 				    zio, 0);
4633 			}
4634 		}
4635 	}
4636 
4637 	if (zio->io_error) {
4638 		/*
4639 		 * If this I/O is attached to a particular vdev,
4640 		 * generate an error message describing the I/O failure
4641 		 * at the block level.  We ignore these errors if the
4642 		 * device is currently unavailable.
4643 		 */
4644 		if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
4645 		    !vdev_is_dead(zio->io_vd)) {
4646 			int ret = zfs_ereport_post(FM_EREPORT_ZFS_IO,
4647 			    zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0);
4648 			if (ret != EALREADY) {
4649 				mutex_enter(&zio->io_vd->vdev_stat_lock);
4650 				if (zio->io_type == ZIO_TYPE_READ)
4651 					zio->io_vd->vdev_stat.vs_read_errors++;
4652 				else if (zio->io_type == ZIO_TYPE_WRITE)
4653 					zio->io_vd->vdev_stat.vs_write_errors++;
4654 				mutex_exit(&zio->io_vd->vdev_stat_lock);
4655 			}
4656 		}
4657 
4658 		if ((zio->io_error == EIO || !(zio->io_flags &
4659 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
4660 		    zio == zio->io_logical) {
4661 			/*
4662 			 * For logical I/O requests, tell the SPA to log the
4663 			 * error and generate a logical data ereport.
4664 			 */
4665 			spa_log_error(zio->io_spa, &zio->io_bookmark);
4666 			(void) zfs_ereport_post(FM_EREPORT_ZFS_DATA,
4667 			    zio->io_spa, NULL, &zio->io_bookmark, zio, 0);
4668 		}
4669 	}
4670 
4671 	if (zio->io_error && zio == zio->io_logical) {
4672 		/*
4673 		 * Determine whether zio should be reexecuted.  This will
4674 		 * propagate all the way to the root via zio_notify_parent().
4675 		 */
4676 		ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
4677 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4678 
4679 		if (IO_IS_ALLOCATING(zio) &&
4680 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
4681 			if (zio->io_error != ENOSPC)
4682 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
4683 			else
4684 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4685 		}
4686 
4687 		if ((zio->io_type == ZIO_TYPE_READ ||
4688 		    zio->io_type == ZIO_TYPE_FREE) &&
4689 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
4690 		    zio->io_error == ENXIO &&
4691 		    spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
4692 		    spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
4693 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4694 
4695 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
4696 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4697 
4698 		/*
4699 		 * Here is a possibly good place to attempt to do
4700 		 * either combinatorial reconstruction or error correction
4701 		 * based on checksums.  It also might be a good place
4702 		 * to send out preliminary ereports before we suspend
4703 		 * processing.
4704 		 */
4705 	}
4706 
4707 	/*
4708 	 * If there were logical child errors, they apply to us now.
4709 	 * We defer this until now to avoid conflating logical child
4710 	 * errors with errors that happened to the zio itself when
4711 	 * updating vdev stats and reporting FMA events above.
4712 	 */
4713 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
4714 
4715 	if ((zio->io_error || zio->io_reexecute) &&
4716 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
4717 	    !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
4718 		zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
4719 
4720 	zio_gang_tree_free(&zio->io_gang_tree);
4721 
4722 	/*
4723 	 * Godfather I/Os should never suspend.
4724 	 */
4725 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4726 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4727 		zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND;
4728 
4729 	if (zio->io_reexecute) {
4730 		/*
4731 		 * This is a logical I/O that wants to reexecute.
4732 		 *
4733 		 * Reexecute is top-down.  When an i/o fails, if it's not
4734 		 * the root, it simply notifies its parent and sticks around.
4735 		 * The parent, seeing that it still has children in zio_done(),
4736 		 * does the same.  This percolates all the way up to the root.
4737 		 * The root i/o will reexecute or suspend the entire tree.
4738 		 *
4739 		 * This approach ensures that zio_reexecute() honors
4740 		 * all the original i/o dependency relationships, e.g.
4741 		 * parents not executing until children are ready.
4742 		 */
4743 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4744 
4745 		zio->io_gang_leader = NULL;
4746 
4747 		mutex_enter(&zio->io_lock);
4748 		zio->io_state[ZIO_WAIT_DONE] = 1;
4749 		mutex_exit(&zio->io_lock);
4750 
4751 		/*
4752 		 * "The Godfather" I/O monitors its children but is
4753 		 * not a true parent to them. It will track them through
4754 		 * the pipeline but severs its ties whenever they get into
4755 		 * trouble (e.g. suspended). This allows "The Godfather"
4756 		 * I/O to return status without blocking.
4757 		 */
4758 		zl = NULL;
4759 		for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4760 		    pio = pio_next) {
4761 			zio_link_t *remove_zl = zl;
4762 			pio_next = zio_walk_parents(zio, &zl);
4763 
4764 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4765 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4766 				zio_remove_child(pio, zio, remove_zl);
4767 				/*
4768 				 * This is a rare code path, so we don't
4769 				 * bother with "next_to_execute".
4770 				 */
4771 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE,
4772 				    NULL);
4773 			}
4774 		}
4775 
4776 		if ((pio = zio_unique_parent(zio)) != NULL) {
4777 			/*
4778 			 * We're not a root i/o, so there's nothing to do
4779 			 * but notify our parent.  Don't propagate errors
4780 			 * upward since we haven't permanently failed yet.
4781 			 */
4782 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4783 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4784 			/*
4785 			 * This is a rare code path, so we don't bother with
4786 			 * "next_to_execute".
4787 			 */
4788 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL);
4789 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4790 			/*
4791 			 * We'd fail again if we reexecuted now, so suspend
4792 			 * until conditions improve (e.g. device comes online).
4793 			 */
4794 			zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
4795 		} else {
4796 			/*
4797 			 * Reexecution is potentially a huge amount of work.
4798 			 * Hand it off to the otherwise-unused claim taskq.
4799 			 */
4800 			ASSERT(taskq_empty_ent(&zio->io_tqent));
4801 			spa_taskq_dispatch_ent(zio->io_spa,
4802 			    ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
4803 			    zio_reexecute, zio, 0, &zio->io_tqent);
4804 		}
4805 		return (NULL);
4806 	}
4807 
4808 	ASSERT(zio->io_child_count == 0);
4809 	ASSERT(zio->io_reexecute == 0);
4810 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4811 
4812 	/*
4813 	 * Report any checksum errors, since the I/O is complete.
4814 	 */
4815 	while (zio->io_cksum_report != NULL) {
4816 		zio_cksum_report_t *zcr = zio->io_cksum_report;
4817 		zio->io_cksum_report = zcr->zcr_next;
4818 		zcr->zcr_next = NULL;
4819 		zcr->zcr_finish(zcr, NULL);
4820 		zfs_ereport_free_checksum(zcr);
4821 	}
4822 
4823 	if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
4824 	    !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) &&
4825 	    !(zio->io_flags & ZIO_FLAG_NOPWRITE)) {
4826 		metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
4827 	}
4828 
4829 	/*
4830 	 * It is the responsibility of the done callback to ensure that this
4831 	 * particular zio is no longer discoverable for adoption, and as
4832 	 * such, cannot acquire any new parents.
4833 	 */
4834 	if (zio->io_done)
4835 		zio->io_done(zio);
4836 
4837 	mutex_enter(&zio->io_lock);
4838 	zio->io_state[ZIO_WAIT_DONE] = 1;
4839 	mutex_exit(&zio->io_lock);
4840 
4841 	/*
4842 	 * We are done executing this zio.  We may want to execute a parent
4843 	 * next.  See the comment in zio_notify_parent().
4844 	 */
4845 	zio_t *next_to_execute = NULL;
4846 	zl = NULL;
4847 	for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4848 		zio_link_t *remove_zl = zl;
4849 		pio_next = zio_walk_parents(zio, &zl);
4850 		zio_remove_child(pio, zio, remove_zl);
4851 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute);
4852 	}
4853 
4854 	if (zio->io_waiter != NULL) {
4855 		mutex_enter(&zio->io_lock);
4856 		zio->io_executor = NULL;
4857 		cv_broadcast(&zio->io_cv);
4858 		mutex_exit(&zio->io_lock);
4859 	} else {
4860 		zio_destroy(zio);
4861 	}
4862 
4863 	return (next_to_execute);
4864 }
4865 
4866 /*
4867  * ==========================================================================
4868  * I/O pipeline definition
4869  * ==========================================================================
4870  */
4871 static zio_pipe_stage_t *zio_pipeline[] = {
4872 	NULL,
4873 	zio_read_bp_init,
4874 	zio_write_bp_init,
4875 	zio_free_bp_init,
4876 	zio_issue_async,
4877 	zio_write_compress,
4878 	zio_encrypt,
4879 	zio_checksum_generate,
4880 	zio_nop_write,
4881 	zio_ddt_read_start,
4882 	zio_ddt_read_done,
4883 	zio_ddt_write,
4884 	zio_ddt_free,
4885 	zio_gang_assemble,
4886 	zio_gang_issue,
4887 	zio_dva_throttle,
4888 	zio_dva_allocate,
4889 	zio_dva_free,
4890 	zio_dva_claim,
4891 	zio_ready,
4892 	zio_vdev_io_start,
4893 	zio_vdev_io_done,
4894 	zio_vdev_io_assess,
4895 	zio_checksum_verify,
4896 	zio_done
4897 };
4898 
4899 
4900 
4901 
4902 /*
4903  * Compare two zbookmark_phys_t's to see which we would reach first in a
4904  * pre-order traversal of the object tree.
4905  *
4906  * This is simple in every case aside from the meta-dnode object. For all other
4907  * objects, we traverse them in order (object 1 before object 2, and so on).
4908  * However, all of these objects are traversed while traversing object 0, since
4909  * the data it points to is the list of objects.  Thus, we need to convert to a
4910  * canonical representation so we can compare meta-dnode bookmarks to
4911  * non-meta-dnode bookmarks.
4912  *
4913  * We do this by calculating "equivalents" for each field of the zbookmark.
4914  * zbookmarks outside of the meta-dnode use their own object and level, and
4915  * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4916  * blocks this bookmark refers to) by multiplying their blkid by their span
4917  * (the number of L0 blocks contained within one block at their level).
4918  * zbookmarks inside the meta-dnode calculate their object equivalent
4919  * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4920  * level + 1<<31 (any value larger than a level could ever be) for their level.
4921  * This causes them to always compare before a bookmark in their object
4922  * equivalent, compare appropriately to bookmarks in other objects, and to
4923  * compare appropriately to other bookmarks in the meta-dnode.
4924  */
4925 int
4926 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4927     const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4928 {
4929 	/*
4930 	 * These variables represent the "equivalent" values for the zbookmark,
4931 	 * after converting zbookmarks inside the meta dnode to their
4932 	 * normal-object equivalents.
4933 	 */
4934 	uint64_t zb1obj, zb2obj;
4935 	uint64_t zb1L0, zb2L0;
4936 	uint64_t zb1level, zb2level;
4937 
4938 	if (zb1->zb_object == zb2->zb_object &&
4939 	    zb1->zb_level == zb2->zb_level &&
4940 	    zb1->zb_blkid == zb2->zb_blkid)
4941 		return (0);
4942 
4943 	IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT);
4944 	IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT);
4945 
4946 	/*
4947 	 * BP_SPANB calculates the span in blocks.
4948 	 */
4949 	zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4950 	zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4951 
4952 	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4953 		zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4954 		zb1L0 = 0;
4955 		zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4956 	} else {
4957 		zb1obj = zb1->zb_object;
4958 		zb1level = zb1->zb_level;
4959 	}
4960 
4961 	if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4962 		zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4963 		zb2L0 = 0;
4964 		zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4965 	} else {
4966 		zb2obj = zb2->zb_object;
4967 		zb2level = zb2->zb_level;
4968 	}
4969 
4970 	/* Now that we have a canonical representation, do the comparison. */
4971 	if (zb1obj != zb2obj)
4972 		return (zb1obj < zb2obj ? -1 : 1);
4973 	else if (zb1L0 != zb2L0)
4974 		return (zb1L0 < zb2L0 ? -1 : 1);
4975 	else if (zb1level != zb2level)
4976 		return (zb1level > zb2level ? -1 : 1);
4977 	/*
4978 	 * This can (theoretically) happen if the bookmarks have the same object
4979 	 * and level, but different blkids, if the block sizes are not the same.
4980 	 * There is presently no way to change the indirect block sizes
4981 	 */
4982 	return (0);
4983 }
4984 
4985 /*
4986  *  This function checks the following: given that last_block is the place that
4987  *  our traversal stopped last time, does that guarantee that we've visited
4988  *  every node under subtree_root?  Therefore, we can't just use the raw output
4989  *  of zbookmark_compare.  We have to pass in a modified version of
4990  *  subtree_root; by incrementing the block id, and then checking whether
4991  *  last_block is before or equal to that, we can tell whether or not having
4992  *  visited last_block implies that all of subtree_root's children have been
4993  *  visited.
4994  */
4995 boolean_t
4996 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4997     const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4998 {
4999 	zbookmark_phys_t mod_zb = *subtree_root;
5000 	mod_zb.zb_blkid++;
5001 	ASSERT(last_block->zb_level == 0);
5002 
5003 	/* The objset_phys_t isn't before anything. */
5004 	if (dnp == NULL)
5005 		return (B_FALSE);
5006 
5007 	/*
5008 	 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
5009 	 * data block size in sectors, because that variable is only used if
5010 	 * the bookmark refers to a block in the meta-dnode.  Since we don't
5011 	 * know without examining it what object it refers to, and there's no
5012 	 * harm in passing in this value in other cases, we always pass it in.
5013 	 *
5014 	 * We pass in 0 for the indirect block size shift because zb2 must be
5015 	 * level 0.  The indirect block size is only used to calculate the span
5016 	 * of the bookmark, but since the bookmark must be level 0, the span is
5017 	 * always 1, so the math works out.
5018 	 *
5019 	 * If you make changes to how the zbookmark_compare code works, be sure
5020 	 * to make sure that this code still works afterwards.
5021 	 */
5022 	return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
5023 	    1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
5024 	    last_block) <= 0);
5025 }
5026 
5027 EXPORT_SYMBOL(zio_type_name);
5028 EXPORT_SYMBOL(zio_buf_alloc);
5029 EXPORT_SYMBOL(zio_data_buf_alloc);
5030 EXPORT_SYMBOL(zio_buf_free);
5031 EXPORT_SYMBOL(zio_data_buf_free);
5032 
5033 /* BEGIN CSTYLED */
5034 ZFS_MODULE_PARAM(zfs_zio, zio_, slow_io_ms, INT, ZMOD_RW,
5035 	"Max I/O completion time (milliseconds) before marking it as slow");
5036 
5037 ZFS_MODULE_PARAM(zfs_zio, zio_, requeue_io_start_cut_in_line, INT, ZMOD_RW,
5038 	"Prioritize requeued I/O");
5039 
5040 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_deferred_free,  INT, ZMOD_RW,
5041 	"Defer frees starting in this pass");
5042 
5043 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_dont_compress, INT, ZMOD_RW,
5044 	"Don't compress starting in this pass");
5045 
5046 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_rewrite, INT, ZMOD_RW,
5047 	"Rewrite new bps starting in this pass");
5048 
5049 ZFS_MODULE_PARAM(zfs_zio, zio_, dva_throttle_enabled, INT, ZMOD_RW,
5050 	"Throttle block allocations in the ZIO pipeline");
5051 
5052 ZFS_MODULE_PARAM(zfs_zio, zio_, deadman_log_all, INT, ZMOD_RW,
5053 	"Log all slow ZIOs, not just those with vdevs");
5054 /* END CSTYLED */
5055