xref: /freebsd/sys/contrib/openzfs/module/zfs/zio.c (revision 13ec1e3155c7e9bf037b12af186351b7fa9b9450)
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 *const 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 static int zio_deadman_log_all = B_FALSE;
70 
71 /*
72  * ==========================================================================
73  * I/O kmem caches
74  * ==========================================================================
75  */
76 static kmem_cache_t *zio_cache;
77 static 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 static uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
82 static 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 static 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 static int zfs_sync_pass_dont_compress = 8; /* don't compress s. i. t. p. */
119 static int zfs_sync_pass_rewrite = 2; /* rewrite new bps s. i. t. p. */
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 static int zio_requeue_io_start_cut_in_line = 1;
133 
134 #ifdef ZFS_DEBUG
135 static const int zio_buf_debug_limit = 16384;
136 #else
137 static const 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 	(void) size;
373 	abd_free((abd_t *)abd);
374 }
375 
376 /*
377  * ==========================================================================
378  * Push and pop I/O transform buffers
379  * ==========================================================================
380  */
381 void
382 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
383     zio_transform_func_t *transform)
384 {
385 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
386 
387 	zt->zt_orig_abd = zio->io_abd;
388 	zt->zt_orig_size = zio->io_size;
389 	zt->zt_bufsize = bufsize;
390 	zt->zt_transform = transform;
391 
392 	zt->zt_next = zio->io_transform_stack;
393 	zio->io_transform_stack = zt;
394 
395 	zio->io_abd = data;
396 	zio->io_size = size;
397 }
398 
399 void
400 zio_pop_transforms(zio_t *zio)
401 {
402 	zio_transform_t *zt;
403 
404 	while ((zt = zio->io_transform_stack) != NULL) {
405 		if (zt->zt_transform != NULL)
406 			zt->zt_transform(zio,
407 			    zt->zt_orig_abd, zt->zt_orig_size);
408 
409 		if (zt->zt_bufsize != 0)
410 			abd_free(zio->io_abd);
411 
412 		zio->io_abd = zt->zt_orig_abd;
413 		zio->io_size = zt->zt_orig_size;
414 		zio->io_transform_stack = zt->zt_next;
415 
416 		kmem_free(zt, sizeof (zio_transform_t));
417 	}
418 }
419 
420 /*
421  * ==========================================================================
422  * I/O transform callbacks for subblocks, decompression, and decryption
423  * ==========================================================================
424  */
425 static void
426 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
427 {
428 	ASSERT(zio->io_size > size);
429 
430 	if (zio->io_type == ZIO_TYPE_READ)
431 		abd_copy(data, zio->io_abd, size);
432 }
433 
434 static void
435 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
436 {
437 	if (zio->io_error == 0) {
438 		void *tmp = abd_borrow_buf(data, size);
439 		int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
440 		    zio->io_abd, tmp, zio->io_size, size,
441 		    &zio->io_prop.zp_complevel);
442 		abd_return_buf_copy(data, tmp, size);
443 
444 		if (zio_injection_enabled && ret == 0)
445 			ret = zio_handle_fault_injection(zio, EINVAL);
446 
447 		if (ret != 0)
448 			zio->io_error = SET_ERROR(EIO);
449 	}
450 }
451 
452 static void
453 zio_decrypt(zio_t *zio, abd_t *data, uint64_t size)
454 {
455 	int ret;
456 	void *tmp;
457 	blkptr_t *bp = zio->io_bp;
458 	spa_t *spa = zio->io_spa;
459 	uint64_t dsobj = zio->io_bookmark.zb_objset;
460 	uint64_t lsize = BP_GET_LSIZE(bp);
461 	dmu_object_type_t ot = BP_GET_TYPE(bp);
462 	uint8_t salt[ZIO_DATA_SALT_LEN];
463 	uint8_t iv[ZIO_DATA_IV_LEN];
464 	uint8_t mac[ZIO_DATA_MAC_LEN];
465 	boolean_t no_crypt = B_FALSE;
466 
467 	ASSERT(BP_USES_CRYPT(bp));
468 	ASSERT3U(size, !=, 0);
469 
470 	if (zio->io_error != 0)
471 		return;
472 
473 	/*
474 	 * Verify the cksum of MACs stored in an indirect bp. It will always
475 	 * be possible to verify this since it does not require an encryption
476 	 * key.
477 	 */
478 	if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {
479 		zio_crypt_decode_mac_bp(bp, mac);
480 
481 		if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
482 			/*
483 			 * We haven't decompressed the data yet, but
484 			 * zio_crypt_do_indirect_mac_checksum() requires
485 			 * decompressed data to be able to parse out the MACs
486 			 * from the indirect block. We decompress it now and
487 			 * throw away the result after we are finished.
488 			 */
489 			tmp = zio_buf_alloc(lsize);
490 			ret = zio_decompress_data(BP_GET_COMPRESS(bp),
491 			    zio->io_abd, tmp, zio->io_size, lsize,
492 			    &zio->io_prop.zp_complevel);
493 			if (ret != 0) {
494 				ret = SET_ERROR(EIO);
495 				goto error;
496 			}
497 			ret = zio_crypt_do_indirect_mac_checksum(B_FALSE,
498 			    tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac);
499 			zio_buf_free(tmp, lsize);
500 		} else {
501 			ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
502 			    zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac);
503 		}
504 		abd_copy(data, zio->io_abd, size);
505 
506 		if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) {
507 			ret = zio_handle_decrypt_injection(spa,
508 			    &zio->io_bookmark, ot, ECKSUM);
509 		}
510 		if (ret != 0)
511 			goto error;
512 
513 		return;
514 	}
515 
516 	/*
517 	 * If this is an authenticated block, just check the MAC. It would be
518 	 * nice to separate this out into its own flag, but for the moment
519 	 * enum zio_flag is out of bits.
520 	 */
521 	if (BP_IS_AUTHENTICATED(bp)) {
522 		if (ot == DMU_OT_OBJSET) {
523 			ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa,
524 			    dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp));
525 		} else {
526 			zio_crypt_decode_mac_bp(bp, mac);
527 			ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj,
528 			    zio->io_abd, size, mac);
529 			if (zio_injection_enabled && ret == 0) {
530 				ret = zio_handle_decrypt_injection(spa,
531 				    &zio->io_bookmark, ot, ECKSUM);
532 			}
533 		}
534 		abd_copy(data, zio->io_abd, size);
535 
536 		if (ret != 0)
537 			goto error;
538 
539 		return;
540 	}
541 
542 	zio_crypt_decode_params_bp(bp, salt, iv);
543 
544 	if (ot == DMU_OT_INTENT_LOG) {
545 		tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t));
546 		zio_crypt_decode_mac_zil(tmp, mac);
547 		abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t));
548 	} else {
549 		zio_crypt_decode_mac_bp(bp, mac);
550 	}
551 
552 	ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp),
553 	    BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data,
554 	    zio->io_abd, &no_crypt);
555 	if (no_crypt)
556 		abd_copy(data, zio->io_abd, size);
557 
558 	if (ret != 0)
559 		goto error;
560 
561 	return;
562 
563 error:
564 	/* assert that the key was found unless this was speculative */
565 	ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE));
566 
567 	/*
568 	 * If there was a decryption / authentication error return EIO as
569 	 * the io_error. If this was not a speculative zio, create an ereport.
570 	 */
571 	if (ret == ECKSUM) {
572 		zio->io_error = SET_ERROR(EIO);
573 		if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
574 			spa_log_error(spa, &zio->io_bookmark);
575 			(void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
576 			    spa, NULL, &zio->io_bookmark, zio, 0);
577 		}
578 	} else {
579 		zio->io_error = ret;
580 	}
581 }
582 
583 /*
584  * ==========================================================================
585  * I/O parent/child relationships and pipeline interlocks
586  * ==========================================================================
587  */
588 zio_t *
589 zio_walk_parents(zio_t *cio, zio_link_t **zl)
590 {
591 	list_t *pl = &cio->io_parent_list;
592 
593 	*zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
594 	if (*zl == NULL)
595 		return (NULL);
596 
597 	ASSERT((*zl)->zl_child == cio);
598 	return ((*zl)->zl_parent);
599 }
600 
601 zio_t *
602 zio_walk_children(zio_t *pio, zio_link_t **zl)
603 {
604 	list_t *cl = &pio->io_child_list;
605 
606 	ASSERT(MUTEX_HELD(&pio->io_lock));
607 
608 	*zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
609 	if (*zl == NULL)
610 		return (NULL);
611 
612 	ASSERT((*zl)->zl_parent == pio);
613 	return ((*zl)->zl_child);
614 }
615 
616 zio_t *
617 zio_unique_parent(zio_t *cio)
618 {
619 	zio_link_t *zl = NULL;
620 	zio_t *pio = zio_walk_parents(cio, &zl);
621 
622 	VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
623 	return (pio);
624 }
625 
626 void
627 zio_add_child(zio_t *pio, zio_t *cio)
628 {
629 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
630 
631 	/*
632 	 * Logical I/Os can have logical, gang, or vdev children.
633 	 * Gang I/Os can have gang or vdev children.
634 	 * Vdev I/Os can only have vdev children.
635 	 * The following ASSERT captures all of these constraints.
636 	 */
637 	ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
638 
639 	zl->zl_parent = pio;
640 	zl->zl_child = cio;
641 
642 	mutex_enter(&pio->io_lock);
643 	mutex_enter(&cio->io_lock);
644 
645 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
646 
647 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
648 		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
649 
650 	list_insert_head(&pio->io_child_list, zl);
651 	list_insert_head(&cio->io_parent_list, zl);
652 
653 	pio->io_child_count++;
654 	cio->io_parent_count++;
655 
656 	mutex_exit(&cio->io_lock);
657 	mutex_exit(&pio->io_lock);
658 }
659 
660 static void
661 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
662 {
663 	ASSERT(zl->zl_parent == pio);
664 	ASSERT(zl->zl_child == cio);
665 
666 	mutex_enter(&pio->io_lock);
667 	mutex_enter(&cio->io_lock);
668 
669 	list_remove(&pio->io_child_list, zl);
670 	list_remove(&cio->io_parent_list, zl);
671 
672 	pio->io_child_count--;
673 	cio->io_parent_count--;
674 
675 	mutex_exit(&cio->io_lock);
676 	mutex_exit(&pio->io_lock);
677 	kmem_cache_free(zio_link_cache, zl);
678 }
679 
680 static boolean_t
681 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
682 {
683 	boolean_t waiting = B_FALSE;
684 
685 	mutex_enter(&zio->io_lock);
686 	ASSERT(zio->io_stall == NULL);
687 	for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
688 		if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
689 			continue;
690 
691 		uint64_t *countp = &zio->io_children[c][wait];
692 		if (*countp != 0) {
693 			zio->io_stage >>= 1;
694 			ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
695 			zio->io_stall = countp;
696 			waiting = B_TRUE;
697 			break;
698 		}
699 	}
700 	mutex_exit(&zio->io_lock);
701 	return (waiting);
702 }
703 
704 __attribute__((always_inline))
705 static inline void
706 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait,
707     zio_t **next_to_executep)
708 {
709 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
710 	int *errorp = &pio->io_child_error[zio->io_child_type];
711 
712 	mutex_enter(&pio->io_lock);
713 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
714 		*errorp = zio_worst_error(*errorp, zio->io_error);
715 	pio->io_reexecute |= zio->io_reexecute;
716 	ASSERT3U(*countp, >, 0);
717 
718 	(*countp)--;
719 
720 	if (*countp == 0 && pio->io_stall == countp) {
721 		zio_taskq_type_t type =
722 		    pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
723 		    ZIO_TASKQ_INTERRUPT;
724 		pio->io_stall = NULL;
725 		mutex_exit(&pio->io_lock);
726 
727 		/*
728 		 * If we can tell the caller to execute this parent next, do
729 		 * so.  Otherwise dispatch the parent zio as its own task.
730 		 *
731 		 * Having the caller execute the parent when possible reduces
732 		 * locking on the zio taskq's, reduces context switch
733 		 * overhead, and has no recursion penalty.  Note that one
734 		 * read from disk typically causes at least 3 zio's: a
735 		 * zio_null(), the logical zio_read(), and then a physical
736 		 * zio.  When the physical ZIO completes, we are able to call
737 		 * zio_done() on all 3 of these zio's from one invocation of
738 		 * zio_execute() by returning the parent back to
739 		 * zio_execute().  Since the parent isn't executed until this
740 		 * thread returns back to zio_execute(), the caller should do
741 		 * so promptly.
742 		 *
743 		 * In other cases, dispatching the parent prevents
744 		 * overflowing the stack when we have deeply nested
745 		 * parent-child relationships, as we do with the "mega zio"
746 		 * of writes for spa_sync(), and the chain of ZIL blocks.
747 		 */
748 		if (next_to_executep != NULL && *next_to_executep == NULL) {
749 			*next_to_executep = pio;
750 		} else {
751 			zio_taskq_dispatch(pio, type, B_FALSE);
752 		}
753 	} else {
754 		mutex_exit(&pio->io_lock);
755 	}
756 }
757 
758 static void
759 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
760 {
761 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
762 		zio->io_error = zio->io_child_error[c];
763 }
764 
765 int
766 zio_bookmark_compare(const void *x1, const void *x2)
767 {
768 	const zio_t *z1 = x1;
769 	const zio_t *z2 = x2;
770 
771 	if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
772 		return (-1);
773 	if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
774 		return (1);
775 
776 	if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
777 		return (-1);
778 	if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
779 		return (1);
780 
781 	if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
782 		return (-1);
783 	if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
784 		return (1);
785 
786 	if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
787 		return (-1);
788 	if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
789 		return (1);
790 
791 	if (z1 < z2)
792 		return (-1);
793 	if (z1 > z2)
794 		return (1);
795 
796 	return (0);
797 }
798 
799 /*
800  * ==========================================================================
801  * Create the various types of I/O (read, write, free, etc)
802  * ==========================================================================
803  */
804 static zio_t *
805 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
806     abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
807     void *private, zio_type_t type, zio_priority_t priority,
808     enum zio_flag flags, vdev_t *vd, uint64_t offset,
809     const zbookmark_phys_t *zb, enum zio_stage stage,
810     enum zio_stage pipeline)
811 {
812 	zio_t *zio;
813 
814 	IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE);
815 	ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
816 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
817 
818 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
819 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
820 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
821 
822 	IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0);
823 
824 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
825 	bzero(zio, sizeof (zio_t));
826 
827 	mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL);
828 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
829 
830 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
831 	    offsetof(zio_link_t, zl_parent_node));
832 	list_create(&zio->io_child_list, sizeof (zio_link_t),
833 	    offsetof(zio_link_t, zl_child_node));
834 	metaslab_trace_init(&zio->io_alloc_list);
835 
836 	if (vd != NULL)
837 		zio->io_child_type = ZIO_CHILD_VDEV;
838 	else if (flags & ZIO_FLAG_GANG_CHILD)
839 		zio->io_child_type = ZIO_CHILD_GANG;
840 	else if (flags & ZIO_FLAG_DDT_CHILD)
841 		zio->io_child_type = ZIO_CHILD_DDT;
842 	else
843 		zio->io_child_type = ZIO_CHILD_LOGICAL;
844 
845 	if (bp != NULL) {
846 		zio->io_bp = (blkptr_t *)bp;
847 		zio->io_bp_copy = *bp;
848 		zio->io_bp_orig = *bp;
849 		if (type != ZIO_TYPE_WRITE ||
850 		    zio->io_child_type == ZIO_CHILD_DDT)
851 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
852 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
853 			zio->io_logical = zio;
854 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
855 			pipeline |= ZIO_GANG_STAGES;
856 	}
857 
858 	zio->io_spa = spa;
859 	zio->io_txg = txg;
860 	zio->io_done = done;
861 	zio->io_private = private;
862 	zio->io_type = type;
863 	zio->io_priority = priority;
864 	zio->io_vd = vd;
865 	zio->io_offset = offset;
866 	zio->io_orig_abd = zio->io_abd = data;
867 	zio->io_orig_size = zio->io_size = psize;
868 	zio->io_lsize = lsize;
869 	zio->io_orig_flags = zio->io_flags = flags;
870 	zio->io_orig_stage = zio->io_stage = stage;
871 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
872 	zio->io_pipeline_trace = ZIO_STAGE_OPEN;
873 
874 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
875 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
876 
877 	if (zb != NULL)
878 		zio->io_bookmark = *zb;
879 
880 	if (pio != NULL) {
881 		zio->io_metaslab_class = pio->io_metaslab_class;
882 		if (zio->io_logical == NULL)
883 			zio->io_logical = pio->io_logical;
884 		if (zio->io_child_type == ZIO_CHILD_GANG)
885 			zio->io_gang_leader = pio->io_gang_leader;
886 		zio_add_child(pio, zio);
887 	}
888 
889 	taskq_init_ent(&zio->io_tqent);
890 
891 	return (zio);
892 }
893 
894 static void
895 zio_destroy(zio_t *zio)
896 {
897 	metaslab_trace_fini(&zio->io_alloc_list);
898 	list_destroy(&zio->io_parent_list);
899 	list_destroy(&zio->io_child_list);
900 	mutex_destroy(&zio->io_lock);
901 	cv_destroy(&zio->io_cv);
902 	kmem_cache_free(zio_cache, zio);
903 }
904 
905 zio_t *
906 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
907     void *private, enum zio_flag flags)
908 {
909 	zio_t *zio;
910 
911 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
912 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
913 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
914 
915 	return (zio);
916 }
917 
918 zio_t *
919 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
920 {
921 	return (zio_null(NULL, spa, NULL, done, private, flags));
922 }
923 
924 static int
925 zfs_blkptr_verify_log(spa_t *spa, const blkptr_t *bp,
926     enum blk_verify_flag blk_verify, const char *fmt, ...)
927 {
928 	va_list adx;
929 	char buf[256];
930 
931 	va_start(adx, fmt);
932 	(void) vsnprintf(buf, sizeof (buf), fmt, adx);
933 	va_end(adx);
934 
935 	switch (blk_verify) {
936 	case BLK_VERIFY_HALT:
937 		dprintf_bp(bp, "blkptr at %p dprintf_bp():", bp);
938 		zfs_panic_recover("%s: %s", spa_name(spa), buf);
939 		break;
940 	case BLK_VERIFY_LOG:
941 		zfs_dbgmsg("%s: %s", spa_name(spa), buf);
942 		break;
943 	case BLK_VERIFY_ONLY:
944 		break;
945 	}
946 
947 	return (1);
948 }
949 
950 /*
951  * Verify the block pointer fields contain reasonable values.  This means
952  * it only contains known object types, checksum/compression identifiers,
953  * block sizes within the maximum allowed limits, valid DVAs, etc.
954  *
955  * If everything checks out B_TRUE is returned.  The zfs_blkptr_verify
956  * argument controls the behavior when an invalid field is detected.
957  *
958  * Modes for zfs_blkptr_verify:
959  *   1) BLK_VERIFY_ONLY (evaluate the block)
960  *   2) BLK_VERIFY_LOG (evaluate the block and log problems)
961  *   3) BLK_VERIFY_HALT (call zfs_panic_recover on error)
962  */
963 boolean_t
964 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp, boolean_t config_held,
965     enum blk_verify_flag blk_verify)
966 {
967 	int errors = 0;
968 
969 	if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
970 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
971 		    "blkptr at %p has invalid TYPE %llu",
972 		    bp, (longlong_t)BP_GET_TYPE(bp));
973 	}
974 	if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
975 	    BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
976 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
977 		    "blkptr at %p has invalid CHECKSUM %llu",
978 		    bp, (longlong_t)BP_GET_CHECKSUM(bp));
979 	}
980 	if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
981 	    BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
982 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
983 		    "blkptr at %p has invalid COMPRESS %llu",
984 		    bp, (longlong_t)BP_GET_COMPRESS(bp));
985 	}
986 	if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
987 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
988 		    "blkptr at %p has invalid LSIZE %llu",
989 		    bp, (longlong_t)BP_GET_LSIZE(bp));
990 	}
991 	if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
992 		errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
993 		    "blkptr at %p has invalid PSIZE %llu",
994 		    bp, (longlong_t)BP_GET_PSIZE(bp));
995 	}
996 
997 	if (BP_IS_EMBEDDED(bp)) {
998 		if (BPE_GET_ETYPE(bp) >= NUM_BP_EMBEDDED_TYPES) {
999 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1000 			    "blkptr at %p has invalid ETYPE %llu",
1001 			    bp, (longlong_t)BPE_GET_ETYPE(bp));
1002 		}
1003 	}
1004 
1005 	/*
1006 	 * Do not verify individual DVAs if the config is not trusted. This
1007 	 * will be done once the zio is executed in vdev_mirror_map_alloc.
1008 	 */
1009 	if (!spa->spa_trust_config)
1010 		return (errors == 0);
1011 
1012 	if (!config_held)
1013 		spa_config_enter(spa, SCL_VDEV, bp, RW_READER);
1014 	else
1015 		ASSERT(spa_config_held(spa, SCL_VDEV, RW_WRITER));
1016 	/*
1017 	 * Pool-specific checks.
1018 	 *
1019 	 * Note: it would be nice to verify that the blk_birth and
1020 	 * BP_PHYSICAL_BIRTH() are not too large.  However, spa_freeze()
1021 	 * allows the birth time of log blocks (and dmu_sync()-ed blocks
1022 	 * that are in the log) to be arbitrarily large.
1023 	 */
1024 	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
1025 		const dva_t *dva = &bp->blk_dva[i];
1026 		uint64_t vdevid = DVA_GET_VDEV(dva);
1027 
1028 		if (vdevid >= spa->spa_root_vdev->vdev_children) {
1029 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1030 			    "blkptr at %p DVA %u has invalid VDEV %llu",
1031 			    bp, i, (longlong_t)vdevid);
1032 			continue;
1033 		}
1034 		vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1035 		if (vd == NULL) {
1036 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1037 			    "blkptr at %p DVA %u has invalid VDEV %llu",
1038 			    bp, i, (longlong_t)vdevid);
1039 			continue;
1040 		}
1041 		if (vd->vdev_ops == &vdev_hole_ops) {
1042 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1043 			    "blkptr at %p DVA %u has hole VDEV %llu",
1044 			    bp, i, (longlong_t)vdevid);
1045 			continue;
1046 		}
1047 		if (vd->vdev_ops == &vdev_missing_ops) {
1048 			/*
1049 			 * "missing" vdevs are valid during import, but we
1050 			 * don't have their detailed info (e.g. asize), so
1051 			 * we can't perform any more checks on them.
1052 			 */
1053 			continue;
1054 		}
1055 		uint64_t offset = DVA_GET_OFFSET(dva);
1056 		uint64_t asize = DVA_GET_ASIZE(dva);
1057 		if (DVA_GET_GANG(dva))
1058 			asize = vdev_gang_header_asize(vd);
1059 		if (offset + asize > vd->vdev_asize) {
1060 			errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1061 			    "blkptr at %p DVA %u has invalid OFFSET %llu",
1062 			    bp, i, (longlong_t)offset);
1063 		}
1064 	}
1065 	if (errors > 0)
1066 		dprintf_bp(bp, "blkptr at %p dprintf_bp():", bp);
1067 	if (!config_held)
1068 		spa_config_exit(spa, SCL_VDEV, bp);
1069 
1070 	return (errors == 0);
1071 }
1072 
1073 boolean_t
1074 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
1075 {
1076 	(void) bp;
1077 	uint64_t vdevid = DVA_GET_VDEV(dva);
1078 
1079 	if (vdevid >= spa->spa_root_vdev->vdev_children)
1080 		return (B_FALSE);
1081 
1082 	vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1083 	if (vd == NULL)
1084 		return (B_FALSE);
1085 
1086 	if (vd->vdev_ops == &vdev_hole_ops)
1087 		return (B_FALSE);
1088 
1089 	if (vd->vdev_ops == &vdev_missing_ops) {
1090 		return (B_FALSE);
1091 	}
1092 
1093 	uint64_t offset = DVA_GET_OFFSET(dva);
1094 	uint64_t asize = DVA_GET_ASIZE(dva);
1095 
1096 	if (DVA_GET_GANG(dva))
1097 		asize = vdev_gang_header_asize(vd);
1098 	if (offset + asize > vd->vdev_asize)
1099 		return (B_FALSE);
1100 
1101 	return (B_TRUE);
1102 }
1103 
1104 zio_t *
1105 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
1106     abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
1107     zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
1108 {
1109 	zio_t *zio;
1110 
1111 	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
1112 	    data, size, size, done, private,
1113 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
1114 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1115 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
1116 
1117 	return (zio);
1118 }
1119 
1120 zio_t *
1121 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
1122     abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
1123     zio_done_func_t *ready, zio_done_func_t *children_ready,
1124     zio_done_func_t *physdone, zio_done_func_t *done,
1125     void *private, zio_priority_t priority, enum zio_flag flags,
1126     const zbookmark_phys_t *zb)
1127 {
1128 	zio_t *zio;
1129 
1130 	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
1131 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
1132 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
1133 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
1134 	    DMU_OT_IS_VALID(zp->zp_type) &&
1135 	    zp->zp_level < 32 &&
1136 	    zp->zp_copies > 0 &&
1137 	    zp->zp_copies <= spa_max_replication(spa));
1138 
1139 	zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
1140 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
1141 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1142 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
1143 
1144 	zio->io_ready = ready;
1145 	zio->io_children_ready = children_ready;
1146 	zio->io_physdone = physdone;
1147 	zio->io_prop = *zp;
1148 
1149 	/*
1150 	 * Data can be NULL if we are going to call zio_write_override() to
1151 	 * provide the already-allocated BP.  But we may need the data to
1152 	 * verify a dedup hit (if requested).  In this case, don't try to
1153 	 * dedup (just take the already-allocated BP verbatim). Encrypted
1154 	 * dedup blocks need data as well so we also disable dedup in this
1155 	 * case.
1156 	 */
1157 	if (data == NULL &&
1158 	    (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) {
1159 		zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
1160 	}
1161 
1162 	return (zio);
1163 }
1164 
1165 zio_t *
1166 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
1167     uint64_t size, zio_done_func_t *done, void *private,
1168     zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
1169 {
1170 	zio_t *zio;
1171 
1172 	zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
1173 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
1174 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
1175 
1176 	return (zio);
1177 }
1178 
1179 void
1180 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
1181 {
1182 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1183 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1184 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1185 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
1186 
1187 	/*
1188 	 * We must reset the io_prop to match the values that existed
1189 	 * when the bp was first written by dmu_sync() keeping in mind
1190 	 * that nopwrite and dedup are mutually exclusive.
1191 	 */
1192 	zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
1193 	zio->io_prop.zp_nopwrite = nopwrite;
1194 	zio->io_prop.zp_copies = copies;
1195 	zio->io_bp_override = bp;
1196 }
1197 
1198 void
1199 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
1200 {
1201 
1202 	(void) zfs_blkptr_verify(spa, bp, B_FALSE, BLK_VERIFY_HALT);
1203 
1204 	/*
1205 	 * The check for EMBEDDED is a performance optimization.  We
1206 	 * process the free here (by ignoring it) rather than
1207 	 * putting it on the list and then processing it in zio_free_sync().
1208 	 */
1209 	if (BP_IS_EMBEDDED(bp))
1210 		return;
1211 	metaslab_check_free(spa, bp);
1212 
1213 	/*
1214 	 * Frees that are for the currently-syncing txg, are not going to be
1215 	 * deferred, and which will not need to do a read (i.e. not GANG or
1216 	 * DEDUP), can be processed immediately.  Otherwise, put them on the
1217 	 * in-memory list for later processing.
1218 	 *
1219 	 * Note that we only defer frees after zfs_sync_pass_deferred_free
1220 	 * when the log space map feature is disabled. [see relevant comment
1221 	 * in spa_sync_iterate_to_convergence()]
1222 	 */
1223 	if (BP_IS_GANG(bp) ||
1224 	    BP_GET_DEDUP(bp) ||
1225 	    txg != spa->spa_syncing_txg ||
1226 	    (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free &&
1227 	    !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))) {
1228 		bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1229 	} else {
1230 		VERIFY3P(zio_free_sync(NULL, spa, txg, bp, 0), ==, NULL);
1231 	}
1232 }
1233 
1234 /*
1235  * To improve performance, this function may return NULL if we were able
1236  * to do the free immediately.  This avoids the cost of creating a zio
1237  * (and linking it to the parent, etc).
1238  */
1239 zio_t *
1240 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1241     enum zio_flag flags)
1242 {
1243 	ASSERT(!BP_IS_HOLE(bp));
1244 	ASSERT(spa_syncing_txg(spa) == txg);
1245 
1246 	if (BP_IS_EMBEDDED(bp))
1247 		return (NULL);
1248 
1249 	metaslab_check_free(spa, bp);
1250 	arc_freed(spa, bp);
1251 	dsl_scan_freed(spa, bp);
1252 
1253 	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) {
1254 		/*
1255 		 * GANG and DEDUP blocks can induce a read (for the gang block
1256 		 * header, or the DDT), so issue them asynchronously so that
1257 		 * this thread is not tied up.
1258 		 */
1259 		enum zio_stage stage =
1260 		    ZIO_FREE_PIPELINE | ZIO_STAGE_ISSUE_ASYNC;
1261 
1262 		return (zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1263 		    BP_GET_PSIZE(bp), NULL, NULL,
1264 		    ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1265 		    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage));
1266 	} else {
1267 		metaslab_free(spa, bp, txg, B_FALSE);
1268 		return (NULL);
1269 	}
1270 }
1271 
1272 zio_t *
1273 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1274     zio_done_func_t *done, void *private, enum zio_flag flags)
1275 {
1276 	zio_t *zio;
1277 
1278 	(void) zfs_blkptr_verify(spa, bp, flags & ZIO_FLAG_CONFIG_WRITER,
1279 	    BLK_VERIFY_HALT);
1280 
1281 	if (BP_IS_EMBEDDED(bp))
1282 		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1283 
1284 	/*
1285 	 * A claim is an allocation of a specific block.  Claims are needed
1286 	 * to support immediate writes in the intent log.  The issue is that
1287 	 * immediate writes contain committed data, but in a txg that was
1288 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
1289 	 * the intent log claims all blocks that contain immediate write data
1290 	 * so that the SPA knows they're in use.
1291 	 *
1292 	 * All claims *must* be resolved in the first txg -- before the SPA
1293 	 * starts allocating blocks -- so that nothing is allocated twice.
1294 	 * If txg == 0 we just verify that the block is claimable.
1295 	 */
1296 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1297 	    spa_min_claim_txg(spa));
1298 	ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1299 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(8) */
1300 
1301 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1302 	    BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1303 	    flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1304 	ASSERT0(zio->io_queued_timestamp);
1305 
1306 	return (zio);
1307 }
1308 
1309 zio_t *
1310 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
1311     zio_done_func_t *done, void *private, enum zio_flag flags)
1312 {
1313 	zio_t *zio;
1314 	int c;
1315 
1316 	if (vd->vdev_children == 0) {
1317 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1318 		    ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1319 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1320 
1321 		zio->io_cmd = cmd;
1322 	} else {
1323 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1324 
1325 		for (c = 0; c < vd->vdev_children; c++)
1326 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1327 			    done, private, flags));
1328 	}
1329 
1330 	return (zio);
1331 }
1332 
1333 zio_t *
1334 zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1335     zio_done_func_t *done, void *private, zio_priority_t priority,
1336     enum zio_flag flags, enum trim_flag trim_flags)
1337 {
1338 	zio_t *zio;
1339 
1340 	ASSERT0(vd->vdev_children);
1341 	ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
1342 	ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1343 	ASSERT3U(size, !=, 0);
1344 
1345 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done,
1346 	    private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL,
1347 	    vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE);
1348 	zio->io_trim_flags = trim_flags;
1349 
1350 	return (zio);
1351 }
1352 
1353 zio_t *
1354 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1355     abd_t *data, int checksum, zio_done_func_t *done, void *private,
1356     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1357 {
1358 	zio_t *zio;
1359 
1360 	ASSERT(vd->vdev_children == 0);
1361 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1362 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1363 	ASSERT3U(offset + size, <=, vd->vdev_psize);
1364 
1365 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1366 	    private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1367 	    offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1368 
1369 	zio->io_prop.zp_checksum = checksum;
1370 
1371 	return (zio);
1372 }
1373 
1374 zio_t *
1375 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1376     abd_t *data, int checksum, zio_done_func_t *done, void *private,
1377     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1378 {
1379 	zio_t *zio;
1380 
1381 	ASSERT(vd->vdev_children == 0);
1382 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1383 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1384 	ASSERT3U(offset + size, <=, vd->vdev_psize);
1385 
1386 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1387 	    private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1388 	    offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1389 
1390 	zio->io_prop.zp_checksum = checksum;
1391 
1392 	if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1393 		/*
1394 		 * zec checksums are necessarily destructive -- they modify
1395 		 * the end of the write buffer to hold the verifier/checksum.
1396 		 * Therefore, we must make a local copy in case the data is
1397 		 * being written to multiple places in parallel.
1398 		 */
1399 		abd_t *wbuf = abd_alloc_sametype(data, size);
1400 		abd_copy(wbuf, data, size);
1401 
1402 		zio_push_transform(zio, wbuf, size, size, NULL);
1403 	}
1404 
1405 	return (zio);
1406 }
1407 
1408 /*
1409  * Create a child I/O to do some work for us.
1410  */
1411 zio_t *
1412 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1413     abd_t *data, uint64_t size, int type, zio_priority_t priority,
1414     enum zio_flag flags, zio_done_func_t *done, void *private)
1415 {
1416 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1417 	zio_t *zio;
1418 
1419 	/*
1420 	 * vdev child I/Os do not propagate their error to the parent.
1421 	 * Therefore, for correct operation the caller *must* check for
1422 	 * and handle the error in the child i/o's done callback.
1423 	 * The only exceptions are i/os that we don't care about
1424 	 * (OPTIONAL or REPAIR).
1425 	 */
1426 	ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1427 	    done != NULL);
1428 
1429 	if (type == ZIO_TYPE_READ && bp != NULL) {
1430 		/*
1431 		 * If we have the bp, then the child should perform the
1432 		 * checksum and the parent need not.  This pushes error
1433 		 * detection as close to the leaves as possible and
1434 		 * eliminates redundant checksums in the interior nodes.
1435 		 */
1436 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1437 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1438 	}
1439 
1440 	if (vd->vdev_ops->vdev_op_leaf) {
1441 		ASSERT0(vd->vdev_children);
1442 		offset += VDEV_LABEL_START_SIZE;
1443 	}
1444 
1445 	flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1446 
1447 	/*
1448 	 * If we've decided to do a repair, the write is not speculative --
1449 	 * even if the original read was.
1450 	 */
1451 	if (flags & ZIO_FLAG_IO_REPAIR)
1452 		flags &= ~ZIO_FLAG_SPECULATIVE;
1453 
1454 	/*
1455 	 * If we're creating a child I/O that is not associated with a
1456 	 * top-level vdev, then the child zio is not an allocating I/O.
1457 	 * If this is a retried I/O then we ignore it since we will
1458 	 * have already processed the original allocating I/O.
1459 	 */
1460 	if (flags & ZIO_FLAG_IO_ALLOCATING &&
1461 	    (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1462 		ASSERT(pio->io_metaslab_class != NULL);
1463 		ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled);
1464 		ASSERT(type == ZIO_TYPE_WRITE);
1465 		ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1466 		ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1467 		ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1468 		    pio->io_child_type == ZIO_CHILD_GANG);
1469 
1470 		flags &= ~ZIO_FLAG_IO_ALLOCATING;
1471 	}
1472 
1473 
1474 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1475 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1476 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1477 	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1478 
1479 	zio->io_physdone = pio->io_physdone;
1480 	if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1481 		zio->io_logical->io_phys_children++;
1482 
1483 	return (zio);
1484 }
1485 
1486 zio_t *
1487 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1488     zio_type_t type, zio_priority_t priority, enum zio_flag flags,
1489     zio_done_func_t *done, void *private)
1490 {
1491 	zio_t *zio;
1492 
1493 	ASSERT(vd->vdev_ops->vdev_op_leaf);
1494 
1495 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1496 	    data, size, size, done, private, type, priority,
1497 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1498 	    vd, offset, NULL,
1499 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1500 
1501 	return (zio);
1502 }
1503 
1504 void
1505 zio_flush(zio_t *zio, vdev_t *vd)
1506 {
1507 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1508 	    NULL, NULL,
1509 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1510 }
1511 
1512 void
1513 zio_shrink(zio_t *zio, uint64_t size)
1514 {
1515 	ASSERT3P(zio->io_executor, ==, NULL);
1516 	ASSERT3U(zio->io_orig_size, ==, zio->io_size);
1517 	ASSERT3U(size, <=, zio->io_size);
1518 
1519 	/*
1520 	 * We don't shrink for raidz because of problems with the
1521 	 * reconstruction when reading back less than the block size.
1522 	 * Note, BP_IS_RAIDZ() assumes no compression.
1523 	 */
1524 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1525 	if (!BP_IS_RAIDZ(zio->io_bp)) {
1526 		/* we are not doing a raw write */
1527 		ASSERT3U(zio->io_size, ==, zio->io_lsize);
1528 		zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1529 	}
1530 }
1531 
1532 /*
1533  * ==========================================================================
1534  * Prepare to read and write logical blocks
1535  * ==========================================================================
1536  */
1537 
1538 static zio_t *
1539 zio_read_bp_init(zio_t *zio)
1540 {
1541 	blkptr_t *bp = zio->io_bp;
1542 	uint64_t psize =
1543 	    BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1544 
1545 	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1546 
1547 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1548 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
1549 	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1550 		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1551 		    psize, psize, zio_decompress);
1552 	}
1553 
1554 	if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) ||
1555 	    BP_HAS_INDIRECT_MAC_CKSUM(bp)) &&
1556 	    zio->io_child_type == ZIO_CHILD_LOGICAL) {
1557 		zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1558 		    psize, psize, zio_decrypt);
1559 	}
1560 
1561 	if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1562 		int psize = BPE_GET_PSIZE(bp);
1563 		void *data = abd_borrow_buf(zio->io_abd, psize);
1564 
1565 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1566 		decode_embedded_bp_compressed(bp, data);
1567 		abd_return_buf_copy(zio->io_abd, data, psize);
1568 	} else {
1569 		ASSERT(!BP_IS_EMBEDDED(bp));
1570 		ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1571 	}
1572 
1573 	if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1574 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1575 
1576 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1577 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1578 
1579 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1580 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1581 
1582 	return (zio);
1583 }
1584 
1585 static zio_t *
1586 zio_write_bp_init(zio_t *zio)
1587 {
1588 	if (!IO_IS_ALLOCATING(zio))
1589 		return (zio);
1590 
1591 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1592 
1593 	if (zio->io_bp_override) {
1594 		blkptr_t *bp = zio->io_bp;
1595 		zio_prop_t *zp = &zio->io_prop;
1596 
1597 		ASSERT(bp->blk_birth != zio->io_txg);
1598 		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1599 
1600 		*bp = *zio->io_bp_override;
1601 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1602 
1603 		if (BP_IS_EMBEDDED(bp))
1604 			return (zio);
1605 
1606 		/*
1607 		 * If we've been overridden and nopwrite is set then
1608 		 * set the flag accordingly to indicate that a nopwrite
1609 		 * has already occurred.
1610 		 */
1611 		if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1612 			ASSERT(!zp->zp_dedup);
1613 			ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1614 			zio->io_flags |= ZIO_FLAG_NOPWRITE;
1615 			return (zio);
1616 		}
1617 
1618 		ASSERT(!zp->zp_nopwrite);
1619 
1620 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1621 			return (zio);
1622 
1623 		ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1624 		    ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1625 
1626 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum &&
1627 		    !zp->zp_encrypt) {
1628 			BP_SET_DEDUP(bp, 1);
1629 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1630 			return (zio);
1631 		}
1632 
1633 		/*
1634 		 * We were unable to handle this as an override bp, treat
1635 		 * it as a regular write I/O.
1636 		 */
1637 		zio->io_bp_override = NULL;
1638 		*bp = zio->io_bp_orig;
1639 		zio->io_pipeline = zio->io_orig_pipeline;
1640 	}
1641 
1642 	return (zio);
1643 }
1644 
1645 static zio_t *
1646 zio_write_compress(zio_t *zio)
1647 {
1648 	spa_t *spa = zio->io_spa;
1649 	zio_prop_t *zp = &zio->io_prop;
1650 	enum zio_compress compress = zp->zp_compress;
1651 	blkptr_t *bp = zio->io_bp;
1652 	uint64_t lsize = zio->io_lsize;
1653 	uint64_t psize = zio->io_size;
1654 	int pass = 1;
1655 
1656 	/*
1657 	 * If our children haven't all reached the ready stage,
1658 	 * wait for them and then repeat this pipeline stage.
1659 	 */
1660 	if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1661 	    ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1662 		return (NULL);
1663 	}
1664 
1665 	if (!IO_IS_ALLOCATING(zio))
1666 		return (zio);
1667 
1668 	if (zio->io_children_ready != NULL) {
1669 		/*
1670 		 * Now that all our children are ready, run the callback
1671 		 * associated with this zio in case it wants to modify the
1672 		 * data to be written.
1673 		 */
1674 		ASSERT3U(zp->zp_level, >, 0);
1675 		zio->io_children_ready(zio);
1676 	}
1677 
1678 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1679 	ASSERT(zio->io_bp_override == NULL);
1680 
1681 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1682 		/*
1683 		 * We're rewriting an existing block, which means we're
1684 		 * working on behalf of spa_sync().  For spa_sync() to
1685 		 * converge, it must eventually be the case that we don't
1686 		 * have to allocate new blocks.  But compression changes
1687 		 * the blocksize, which forces a reallocate, and makes
1688 		 * convergence take longer.  Therefore, after the first
1689 		 * few passes, stop compressing to ensure convergence.
1690 		 */
1691 		pass = spa_sync_pass(spa);
1692 
1693 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
1694 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1695 		ASSERT(!BP_GET_DEDUP(bp));
1696 
1697 		if (pass >= zfs_sync_pass_dont_compress)
1698 			compress = ZIO_COMPRESS_OFF;
1699 
1700 		/* Make sure someone doesn't change their mind on overwrites */
1701 		ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1702 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1703 	}
1704 
1705 	/* If it's a compressed write that is not raw, compress the buffer. */
1706 	if (compress != ZIO_COMPRESS_OFF &&
1707 	    !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1708 		void *cbuf = zio_buf_alloc(lsize);
1709 		psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize,
1710 		    zp->zp_complevel);
1711 		if (psize == 0 || psize >= lsize) {
1712 			compress = ZIO_COMPRESS_OFF;
1713 			zio_buf_free(cbuf, lsize);
1714 		} else if (!zp->zp_dedup && !zp->zp_encrypt &&
1715 		    psize <= BPE_PAYLOAD_SIZE &&
1716 		    zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1717 		    spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1718 			encode_embedded_bp_compressed(bp,
1719 			    cbuf, compress, lsize, psize);
1720 			BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1721 			BP_SET_TYPE(bp, zio->io_prop.zp_type);
1722 			BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1723 			zio_buf_free(cbuf, lsize);
1724 			bp->blk_birth = zio->io_txg;
1725 			zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1726 			ASSERT(spa_feature_is_active(spa,
1727 			    SPA_FEATURE_EMBEDDED_DATA));
1728 			return (zio);
1729 		} else {
1730 			/*
1731 			 * Round compressed size up to the minimum allocation
1732 			 * size of the smallest-ashift device, and zero the
1733 			 * tail. This ensures that the compressed size of the
1734 			 * BP (and thus compressratio property) are correct,
1735 			 * in that we charge for the padding used to fill out
1736 			 * the last sector.
1737 			 */
1738 			ASSERT3U(spa->spa_min_alloc, >=, SPA_MINBLOCKSHIFT);
1739 			size_t rounded = (size_t)roundup(psize,
1740 			    spa->spa_min_alloc);
1741 			if (rounded >= lsize) {
1742 				compress = ZIO_COMPRESS_OFF;
1743 				zio_buf_free(cbuf, lsize);
1744 				psize = lsize;
1745 			} else {
1746 				abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1747 				abd_take_ownership_of_buf(cdata, B_TRUE);
1748 				abd_zero_off(cdata, psize, rounded - psize);
1749 				psize = rounded;
1750 				zio_push_transform(zio, cdata,
1751 				    psize, lsize, NULL);
1752 			}
1753 		}
1754 
1755 		/*
1756 		 * We were unable to handle this as an override bp, treat
1757 		 * it as a regular write I/O.
1758 		 */
1759 		zio->io_bp_override = NULL;
1760 		*bp = zio->io_bp_orig;
1761 		zio->io_pipeline = zio->io_orig_pipeline;
1762 
1763 	} else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 &&
1764 	    zp->zp_type == DMU_OT_DNODE) {
1765 		/*
1766 		 * The DMU actually relies on the zio layer's compression
1767 		 * to free metadnode blocks that have had all contained
1768 		 * dnodes freed. As a result, even when doing a raw
1769 		 * receive, we must check whether the block can be compressed
1770 		 * to a hole.
1771 		 */
1772 		psize = zio_compress_data(ZIO_COMPRESS_EMPTY,
1773 		    zio->io_abd, NULL, lsize, zp->zp_complevel);
1774 		if (psize == 0 || psize >= lsize)
1775 			compress = ZIO_COMPRESS_OFF;
1776 	} else if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS) {
1777 		size_t rounded = MIN((size_t)roundup(psize,
1778 		    spa->spa_min_alloc), lsize);
1779 
1780 		if (rounded != psize) {
1781 			abd_t *cdata = abd_alloc_linear(rounded, B_TRUE);
1782 			abd_zero_off(cdata, psize, rounded - psize);
1783 			abd_copy_off(cdata, zio->io_abd, 0, 0, psize);
1784 			psize = rounded;
1785 			zio_push_transform(zio, cdata,
1786 			    psize, rounded, NULL);
1787 		}
1788 	} else {
1789 		ASSERT3U(psize, !=, 0);
1790 	}
1791 
1792 	/*
1793 	 * The final pass of spa_sync() must be all rewrites, but the first
1794 	 * few passes offer a trade-off: allocating blocks defers convergence,
1795 	 * but newly allocated blocks are sequential, so they can be written
1796 	 * to disk faster.  Therefore, we allow the first few passes of
1797 	 * spa_sync() to allocate new blocks, but force rewrites after that.
1798 	 * There should only be a handful of blocks after pass 1 in any case.
1799 	 */
1800 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1801 	    BP_GET_PSIZE(bp) == psize &&
1802 	    pass >= zfs_sync_pass_rewrite) {
1803 		VERIFY3U(psize, !=, 0);
1804 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1805 
1806 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1807 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1808 	} else {
1809 		BP_ZERO(bp);
1810 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1811 	}
1812 
1813 	if (psize == 0) {
1814 		if (zio->io_bp_orig.blk_birth != 0 &&
1815 		    spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1816 			BP_SET_LSIZE(bp, lsize);
1817 			BP_SET_TYPE(bp, zp->zp_type);
1818 			BP_SET_LEVEL(bp, zp->zp_level);
1819 			BP_SET_BIRTH(bp, zio->io_txg, 0);
1820 		}
1821 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1822 	} else {
1823 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1824 		BP_SET_LSIZE(bp, lsize);
1825 		BP_SET_TYPE(bp, zp->zp_type);
1826 		BP_SET_LEVEL(bp, zp->zp_level);
1827 		BP_SET_PSIZE(bp, psize);
1828 		BP_SET_COMPRESS(bp, compress);
1829 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1830 		BP_SET_DEDUP(bp, zp->zp_dedup);
1831 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1832 		if (zp->zp_dedup) {
1833 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1834 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1835 			ASSERT(!zp->zp_encrypt ||
1836 			    DMU_OT_IS_ENCRYPTED(zp->zp_type));
1837 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1838 		}
1839 		if (zp->zp_nopwrite) {
1840 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1841 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1842 			zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1843 		}
1844 	}
1845 	return (zio);
1846 }
1847 
1848 static zio_t *
1849 zio_free_bp_init(zio_t *zio)
1850 {
1851 	blkptr_t *bp = zio->io_bp;
1852 
1853 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1854 		if (BP_GET_DEDUP(bp))
1855 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1856 	}
1857 
1858 	ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1859 
1860 	return (zio);
1861 }
1862 
1863 /*
1864  * ==========================================================================
1865  * Execute the I/O pipeline
1866  * ==========================================================================
1867  */
1868 
1869 static void
1870 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1871 {
1872 	spa_t *spa = zio->io_spa;
1873 	zio_type_t t = zio->io_type;
1874 	int flags = (cutinline ? TQ_FRONT : 0);
1875 
1876 	/*
1877 	 * If we're a config writer or a probe, the normal issue and
1878 	 * interrupt threads may all be blocked waiting for the config lock.
1879 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1880 	 */
1881 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1882 		t = ZIO_TYPE_NULL;
1883 
1884 	/*
1885 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1886 	 */
1887 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1888 		t = ZIO_TYPE_NULL;
1889 
1890 	/*
1891 	 * If this is a high priority I/O, then use the high priority taskq if
1892 	 * available.
1893 	 */
1894 	if ((zio->io_priority == ZIO_PRIORITY_NOW ||
1895 	    zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) &&
1896 	    spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1897 		q++;
1898 
1899 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1900 
1901 	/*
1902 	 * NB: We are assuming that the zio can only be dispatched
1903 	 * to a single taskq at a time.  It would be a grievous error
1904 	 * to dispatch the zio to another taskq at the same time.
1905 	 */
1906 	ASSERT(taskq_empty_ent(&zio->io_tqent));
1907 	spa_taskq_dispatch_ent(spa, t, q, zio_execute, zio, flags,
1908 	    &zio->io_tqent);
1909 }
1910 
1911 static boolean_t
1912 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1913 {
1914 	spa_t *spa = zio->io_spa;
1915 
1916 	taskq_t *tq = taskq_of_curthread();
1917 
1918 	for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1919 		spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1920 		uint_t i;
1921 		for (i = 0; i < tqs->stqs_count; i++) {
1922 			if (tqs->stqs_taskq[i] == tq)
1923 				return (B_TRUE);
1924 		}
1925 	}
1926 
1927 	return (B_FALSE);
1928 }
1929 
1930 static zio_t *
1931 zio_issue_async(zio_t *zio)
1932 {
1933 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1934 
1935 	return (NULL);
1936 }
1937 
1938 void
1939 zio_interrupt(void *zio)
1940 {
1941 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1942 }
1943 
1944 void
1945 zio_delay_interrupt(zio_t *zio)
1946 {
1947 	/*
1948 	 * The timeout_generic() function isn't defined in userspace, so
1949 	 * rather than trying to implement the function, the zio delay
1950 	 * functionality has been disabled for userspace builds.
1951 	 */
1952 
1953 #ifdef _KERNEL
1954 	/*
1955 	 * If io_target_timestamp is zero, then no delay has been registered
1956 	 * for this IO, thus jump to the end of this function and "skip" the
1957 	 * delay; issuing it directly to the zio layer.
1958 	 */
1959 	if (zio->io_target_timestamp != 0) {
1960 		hrtime_t now = gethrtime();
1961 
1962 		if (now >= zio->io_target_timestamp) {
1963 			/*
1964 			 * This IO has already taken longer than the target
1965 			 * delay to complete, so we don't want to delay it
1966 			 * any longer; we "miss" the delay and issue it
1967 			 * directly to the zio layer. This is likely due to
1968 			 * the target latency being set to a value less than
1969 			 * the underlying hardware can satisfy (e.g. delay
1970 			 * set to 1ms, but the disks take 10ms to complete an
1971 			 * IO request).
1972 			 */
1973 
1974 			DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1975 			    hrtime_t, now);
1976 
1977 			zio_interrupt(zio);
1978 		} else {
1979 			taskqid_t tid;
1980 			hrtime_t diff = zio->io_target_timestamp - now;
1981 			clock_t expire_at_tick = ddi_get_lbolt() +
1982 			    NSEC_TO_TICK(diff);
1983 
1984 			DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1985 			    hrtime_t, now, hrtime_t, diff);
1986 
1987 			if (NSEC_TO_TICK(diff) == 0) {
1988 				/* Our delay is less than a jiffy - just spin */
1989 				zfs_sleep_until(zio->io_target_timestamp);
1990 				zio_interrupt(zio);
1991 			} else {
1992 				/*
1993 				 * Use taskq_dispatch_delay() in the place of
1994 				 * OpenZFS's timeout_generic().
1995 				 */
1996 				tid = taskq_dispatch_delay(system_taskq,
1997 				    zio_interrupt, zio, TQ_NOSLEEP,
1998 				    expire_at_tick);
1999 				if (tid == TASKQID_INVALID) {
2000 					/*
2001 					 * Couldn't allocate a task.  Just
2002 					 * finish the zio without a delay.
2003 					 */
2004 					zio_interrupt(zio);
2005 				}
2006 			}
2007 		}
2008 		return;
2009 	}
2010 #endif
2011 	DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
2012 	zio_interrupt(zio);
2013 }
2014 
2015 static void
2016 zio_deadman_impl(zio_t *pio, int ziodepth)
2017 {
2018 	zio_t *cio, *cio_next;
2019 	zio_link_t *zl = NULL;
2020 	vdev_t *vd = pio->io_vd;
2021 
2022 	if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) {
2023 		vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL;
2024 		zbookmark_phys_t *zb = &pio->io_bookmark;
2025 		uint64_t delta = gethrtime() - pio->io_timestamp;
2026 		uint64_t failmode = spa_get_deadman_failmode(pio->io_spa);
2027 
2028 		zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu "
2029 		    "delta=%llu queued=%llu io=%llu "
2030 		    "path=%s "
2031 		    "last=%llu type=%d "
2032 		    "priority=%d flags=0x%x stage=0x%x "
2033 		    "pipeline=0x%x pipeline-trace=0x%x "
2034 		    "objset=%llu object=%llu "
2035 		    "level=%llu blkid=%llu "
2036 		    "offset=%llu size=%llu "
2037 		    "error=%d",
2038 		    ziodepth, pio, pio->io_timestamp,
2039 		    (u_longlong_t)delta, pio->io_delta, pio->io_delay,
2040 		    vd ? vd->vdev_path : "NULL",
2041 		    vq ? vq->vq_io_complete_ts : 0, pio->io_type,
2042 		    pio->io_priority, pio->io_flags, pio->io_stage,
2043 		    pio->io_pipeline, pio->io_pipeline_trace,
2044 		    (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object,
2045 		    (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid,
2046 		    (u_longlong_t)pio->io_offset, (u_longlong_t)pio->io_size,
2047 		    pio->io_error);
2048 		(void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN,
2049 		    pio->io_spa, vd, zb, pio, 0);
2050 
2051 		if (failmode == ZIO_FAILURE_MODE_CONTINUE &&
2052 		    taskq_empty_ent(&pio->io_tqent)) {
2053 			zio_interrupt(pio);
2054 		}
2055 	}
2056 
2057 	mutex_enter(&pio->io_lock);
2058 	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2059 		cio_next = zio_walk_children(pio, &zl);
2060 		zio_deadman_impl(cio, ziodepth + 1);
2061 	}
2062 	mutex_exit(&pio->io_lock);
2063 }
2064 
2065 /*
2066  * Log the critical information describing this zio and all of its children
2067  * using the zfs_dbgmsg() interface then post deadman event for the ZED.
2068  */
2069 void
2070 zio_deadman(zio_t *pio, char *tag)
2071 {
2072 	spa_t *spa = pio->io_spa;
2073 	char *name = spa_name(spa);
2074 
2075 	if (!zfs_deadman_enabled || spa_suspended(spa))
2076 		return;
2077 
2078 	zio_deadman_impl(pio, 0);
2079 
2080 	switch (spa_get_deadman_failmode(spa)) {
2081 	case ZIO_FAILURE_MODE_WAIT:
2082 		zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name);
2083 		break;
2084 
2085 	case ZIO_FAILURE_MODE_CONTINUE:
2086 		zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name);
2087 		break;
2088 
2089 	case ZIO_FAILURE_MODE_PANIC:
2090 		fm_panic("%s determined I/O to pool '%s' is hung.", tag, name);
2091 		break;
2092 	}
2093 }
2094 
2095 /*
2096  * Execute the I/O pipeline until one of the following occurs:
2097  * (1) the I/O completes; (2) the pipeline stalls waiting for
2098  * dependent child I/Os; (3) the I/O issues, so we're waiting
2099  * for an I/O completion interrupt; (4) the I/O is delegated by
2100  * vdev-level caching or aggregation; (5) the I/O is deferred
2101  * due to vdev-level queueing; (6) the I/O is handed off to
2102  * another thread.  In all cases, the pipeline stops whenever
2103  * there's no CPU work; it never burns a thread in cv_wait_io().
2104  *
2105  * There's no locking on io_stage because there's no legitimate way
2106  * for multiple threads to be attempting to process the same I/O.
2107  */
2108 static zio_pipe_stage_t *zio_pipeline[];
2109 
2110 /*
2111  * zio_execute() is a wrapper around the static function
2112  * __zio_execute() so that we can force  __zio_execute() to be
2113  * inlined.  This reduces stack overhead which is important
2114  * because __zio_execute() is called recursively in several zio
2115  * code paths.  zio_execute() itself cannot be inlined because
2116  * it is externally visible.
2117  */
2118 void
2119 zio_execute(void *zio)
2120 {
2121 	fstrans_cookie_t cookie;
2122 
2123 	cookie = spl_fstrans_mark();
2124 	__zio_execute(zio);
2125 	spl_fstrans_unmark(cookie);
2126 }
2127 
2128 /*
2129  * Used to determine if in the current context the stack is sized large
2130  * enough to allow zio_execute() to be called recursively.  A minimum
2131  * stack size of 16K is required to avoid needing to re-dispatch the zio.
2132  */
2133 static boolean_t
2134 zio_execute_stack_check(zio_t *zio)
2135 {
2136 #if !defined(HAVE_LARGE_STACKS)
2137 	dsl_pool_t *dp = spa_get_dsl(zio->io_spa);
2138 
2139 	/* Executing in txg_sync_thread() context. */
2140 	if (dp && curthread == dp->dp_tx.tx_sync_thread)
2141 		return (B_TRUE);
2142 
2143 	/* Pool initialization outside of zio_taskq context. */
2144 	if (dp && spa_is_initializing(dp->dp_spa) &&
2145 	    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
2146 	    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))
2147 		return (B_TRUE);
2148 #else
2149 	(void) zio;
2150 #endif /* HAVE_LARGE_STACKS */
2151 
2152 	return (B_FALSE);
2153 }
2154 
2155 __attribute__((always_inline))
2156 static inline void
2157 __zio_execute(zio_t *zio)
2158 {
2159 	ASSERT3U(zio->io_queued_timestamp, >, 0);
2160 
2161 	while (zio->io_stage < ZIO_STAGE_DONE) {
2162 		enum zio_stage pipeline = zio->io_pipeline;
2163 		enum zio_stage stage = zio->io_stage;
2164 
2165 		zio->io_executor = curthread;
2166 
2167 		ASSERT(!MUTEX_HELD(&zio->io_lock));
2168 		ASSERT(ISP2(stage));
2169 		ASSERT(zio->io_stall == NULL);
2170 
2171 		do {
2172 			stage <<= 1;
2173 		} while ((stage & pipeline) == 0);
2174 
2175 		ASSERT(stage <= ZIO_STAGE_DONE);
2176 
2177 		/*
2178 		 * If we are in interrupt context and this pipeline stage
2179 		 * will grab a config lock that is held across I/O,
2180 		 * or may wait for an I/O that needs an interrupt thread
2181 		 * to complete, issue async to avoid deadlock.
2182 		 *
2183 		 * For VDEV_IO_START, we cut in line so that the io will
2184 		 * be sent to disk promptly.
2185 		 */
2186 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
2187 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
2188 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2189 			    zio_requeue_io_start_cut_in_line : B_FALSE;
2190 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2191 			return;
2192 		}
2193 
2194 		/*
2195 		 * If the current context doesn't have large enough stacks
2196 		 * the zio must be issued asynchronously to prevent overflow.
2197 		 */
2198 		if (zio_execute_stack_check(zio)) {
2199 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2200 			    zio_requeue_io_start_cut_in_line : B_FALSE;
2201 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2202 			return;
2203 		}
2204 
2205 		zio->io_stage = stage;
2206 		zio->io_pipeline_trace |= zio->io_stage;
2207 
2208 		/*
2209 		 * The zio pipeline stage returns the next zio to execute
2210 		 * (typically the same as this one), or NULL if we should
2211 		 * stop.
2212 		 */
2213 		zio = zio_pipeline[highbit64(stage) - 1](zio);
2214 
2215 		if (zio == NULL)
2216 			return;
2217 	}
2218 }
2219 
2220 
2221 /*
2222  * ==========================================================================
2223  * Initiate I/O, either sync or async
2224  * ==========================================================================
2225  */
2226 int
2227 zio_wait(zio_t *zio)
2228 {
2229 	/*
2230 	 * Some routines, like zio_free_sync(), may return a NULL zio
2231 	 * to avoid the performance overhead of creating and then destroying
2232 	 * an unneeded zio.  For the callers' simplicity, we accept a NULL
2233 	 * zio and ignore it.
2234 	 */
2235 	if (zio == NULL)
2236 		return (0);
2237 
2238 	long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms);
2239 	int error;
2240 
2241 	ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN);
2242 	ASSERT3P(zio->io_executor, ==, NULL);
2243 
2244 	zio->io_waiter = curthread;
2245 	ASSERT0(zio->io_queued_timestamp);
2246 	zio->io_queued_timestamp = gethrtime();
2247 
2248 	__zio_execute(zio);
2249 
2250 	mutex_enter(&zio->io_lock);
2251 	while (zio->io_executor != NULL) {
2252 		error = cv_timedwait_io(&zio->io_cv, &zio->io_lock,
2253 		    ddi_get_lbolt() + timeout);
2254 
2255 		if (zfs_deadman_enabled && error == -1 &&
2256 		    gethrtime() - zio->io_queued_timestamp >
2257 		    spa_deadman_ziotime(zio->io_spa)) {
2258 			mutex_exit(&zio->io_lock);
2259 			timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms);
2260 			zio_deadman(zio, FTAG);
2261 			mutex_enter(&zio->io_lock);
2262 		}
2263 	}
2264 	mutex_exit(&zio->io_lock);
2265 
2266 	error = zio->io_error;
2267 	zio_destroy(zio);
2268 
2269 	return (error);
2270 }
2271 
2272 void
2273 zio_nowait(zio_t *zio)
2274 {
2275 	/*
2276 	 * See comment in zio_wait().
2277 	 */
2278 	if (zio == NULL)
2279 		return;
2280 
2281 	ASSERT3P(zio->io_executor, ==, NULL);
2282 
2283 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
2284 	    zio_unique_parent(zio) == NULL) {
2285 		zio_t *pio;
2286 
2287 		/*
2288 		 * This is a logical async I/O with no parent to wait for it.
2289 		 * We add it to the spa_async_root_zio "Godfather" I/O which
2290 		 * will ensure they complete prior to unloading the pool.
2291 		 */
2292 		spa_t *spa = zio->io_spa;
2293 		pio = spa->spa_async_zio_root[CPU_SEQID_UNSTABLE];
2294 
2295 		zio_add_child(pio, zio);
2296 	}
2297 
2298 	ASSERT0(zio->io_queued_timestamp);
2299 	zio->io_queued_timestamp = gethrtime();
2300 	__zio_execute(zio);
2301 }
2302 
2303 /*
2304  * ==========================================================================
2305  * Reexecute, cancel, or suspend/resume failed I/O
2306  * ==========================================================================
2307  */
2308 
2309 static void
2310 zio_reexecute(void *arg)
2311 {
2312 	zio_t *pio = arg;
2313 	zio_t *cio, *cio_next;
2314 
2315 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
2316 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
2317 	ASSERT(pio->io_gang_leader == NULL);
2318 	ASSERT(pio->io_gang_tree == NULL);
2319 
2320 	pio->io_flags = pio->io_orig_flags;
2321 	pio->io_stage = pio->io_orig_stage;
2322 	pio->io_pipeline = pio->io_orig_pipeline;
2323 	pio->io_reexecute = 0;
2324 	pio->io_flags |= ZIO_FLAG_REEXECUTED;
2325 	pio->io_pipeline_trace = 0;
2326 	pio->io_error = 0;
2327 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2328 		pio->io_state[w] = 0;
2329 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2330 		pio->io_child_error[c] = 0;
2331 
2332 	if (IO_IS_ALLOCATING(pio))
2333 		BP_ZERO(pio->io_bp);
2334 
2335 	/*
2336 	 * As we reexecute pio's children, new children could be created.
2337 	 * New children go to the head of pio's io_child_list, however,
2338 	 * so we will (correctly) not reexecute them.  The key is that
2339 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
2340 	 * cannot be affected by any side effects of reexecuting 'cio'.
2341 	 */
2342 	zio_link_t *zl = NULL;
2343 	mutex_enter(&pio->io_lock);
2344 	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2345 		cio_next = zio_walk_children(pio, &zl);
2346 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2347 			pio->io_children[cio->io_child_type][w]++;
2348 		mutex_exit(&pio->io_lock);
2349 		zio_reexecute(cio);
2350 		mutex_enter(&pio->io_lock);
2351 	}
2352 	mutex_exit(&pio->io_lock);
2353 
2354 	/*
2355 	 * Now that all children have been reexecuted, execute the parent.
2356 	 * We don't reexecute "The Godfather" I/O here as it's the
2357 	 * responsibility of the caller to wait on it.
2358 	 */
2359 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
2360 		pio->io_queued_timestamp = gethrtime();
2361 		__zio_execute(pio);
2362 	}
2363 }
2364 
2365 void
2366 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
2367 {
2368 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
2369 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2370 		    "failure and the failure mode property for this pool "
2371 		    "is set to panic.", spa_name(spa));
2372 
2373 	cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
2374 	    "failure and has been suspended.\n", spa_name(spa));
2375 
2376 	(void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
2377 	    NULL, NULL, 0);
2378 
2379 	mutex_enter(&spa->spa_suspend_lock);
2380 
2381 	if (spa->spa_suspend_zio_root == NULL)
2382 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
2383 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2384 		    ZIO_FLAG_GODFATHER);
2385 
2386 	spa->spa_suspended = reason;
2387 
2388 	if (zio != NULL) {
2389 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2390 		ASSERT(zio != spa->spa_suspend_zio_root);
2391 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2392 		ASSERT(zio_unique_parent(zio) == NULL);
2393 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
2394 		zio_add_child(spa->spa_suspend_zio_root, zio);
2395 	}
2396 
2397 	mutex_exit(&spa->spa_suspend_lock);
2398 }
2399 
2400 int
2401 zio_resume(spa_t *spa)
2402 {
2403 	zio_t *pio;
2404 
2405 	/*
2406 	 * Reexecute all previously suspended i/o.
2407 	 */
2408 	mutex_enter(&spa->spa_suspend_lock);
2409 	spa->spa_suspended = ZIO_SUSPEND_NONE;
2410 	cv_broadcast(&spa->spa_suspend_cv);
2411 	pio = spa->spa_suspend_zio_root;
2412 	spa->spa_suspend_zio_root = NULL;
2413 	mutex_exit(&spa->spa_suspend_lock);
2414 
2415 	if (pio == NULL)
2416 		return (0);
2417 
2418 	zio_reexecute(pio);
2419 	return (zio_wait(pio));
2420 }
2421 
2422 void
2423 zio_resume_wait(spa_t *spa)
2424 {
2425 	mutex_enter(&spa->spa_suspend_lock);
2426 	while (spa_suspended(spa))
2427 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
2428 	mutex_exit(&spa->spa_suspend_lock);
2429 }
2430 
2431 /*
2432  * ==========================================================================
2433  * Gang blocks.
2434  *
2435  * A gang block is a collection of small blocks that looks to the DMU
2436  * like one large block.  When zio_dva_allocate() cannot find a block
2437  * of the requested size, due to either severe fragmentation or the pool
2438  * being nearly full, it calls zio_write_gang_block() to construct the
2439  * block from smaller fragments.
2440  *
2441  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2442  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
2443  * an indirect block: it's an array of block pointers.  It consumes
2444  * only one sector and hence is allocatable regardless of fragmentation.
2445  * The gang header's bps point to its gang members, which hold the data.
2446  *
2447  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2448  * as the verifier to ensure uniqueness of the SHA256 checksum.
2449  * Critically, the gang block bp's blk_cksum is the checksum of the data,
2450  * not the gang header.  This ensures that data block signatures (needed for
2451  * deduplication) are independent of how the block is physically stored.
2452  *
2453  * Gang blocks can be nested: a gang member may itself be a gang block.
2454  * Thus every gang block is a tree in which root and all interior nodes are
2455  * gang headers, and the leaves are normal blocks that contain user data.
2456  * The root of the gang tree is called the gang leader.
2457  *
2458  * To perform any operation (read, rewrite, free, claim) on a gang block,
2459  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2460  * in the io_gang_tree field of the original logical i/o by recursively
2461  * reading the gang leader and all gang headers below it.  This yields
2462  * an in-core tree containing the contents of every gang header and the
2463  * bps for every constituent of the gang block.
2464  *
2465  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2466  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
2467  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2468  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2469  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2470  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
2471  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2472  * of the gang header plus zio_checksum_compute() of the data to update the
2473  * gang header's blk_cksum as described above.
2474  *
2475  * The two-phase assemble/issue model solves the problem of partial failure --
2476  * what if you'd freed part of a gang block but then couldn't read the
2477  * gang header for another part?  Assembling the entire gang tree first
2478  * ensures that all the necessary gang header I/O has succeeded before
2479  * starting the actual work of free, claim, or write.  Once the gang tree
2480  * is assembled, free and claim are in-memory operations that cannot fail.
2481  *
2482  * In the event that a gang write fails, zio_dva_unallocate() walks the
2483  * gang tree to immediately free (i.e. insert back into the space map)
2484  * everything we've allocated.  This ensures that we don't get ENOSPC
2485  * errors during repeated suspend/resume cycles due to a flaky device.
2486  *
2487  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
2488  * the gang tree, we won't modify the block, so we can safely defer the free
2489  * (knowing that the block is still intact).  If we *can* assemble the gang
2490  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2491  * each constituent bp and we can allocate a new block on the next sync pass.
2492  *
2493  * In all cases, the gang tree allows complete recovery from partial failure.
2494  * ==========================================================================
2495  */
2496 
2497 static void
2498 zio_gang_issue_func_done(zio_t *zio)
2499 {
2500 	abd_free(zio->io_abd);
2501 }
2502 
2503 static zio_t *
2504 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2505     uint64_t offset)
2506 {
2507 	if (gn != NULL)
2508 		return (pio);
2509 
2510 	return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2511 	    BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2512 	    NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2513 	    &pio->io_bookmark));
2514 }
2515 
2516 static zio_t *
2517 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2518     uint64_t offset)
2519 {
2520 	zio_t *zio;
2521 
2522 	if (gn != NULL) {
2523 		abd_t *gbh_abd =
2524 		    abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2525 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2526 		    gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2527 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2528 		    &pio->io_bookmark);
2529 		/*
2530 		 * As we rewrite each gang header, the pipeline will compute
2531 		 * a new gang block header checksum for it; but no one will
2532 		 * compute a new data checksum, so we do that here.  The one
2533 		 * exception is the gang leader: the pipeline already computed
2534 		 * its data checksum because that stage precedes gang assembly.
2535 		 * (Presently, nothing actually uses interior data checksums;
2536 		 * this is just good hygiene.)
2537 		 */
2538 		if (gn != pio->io_gang_leader->io_gang_tree) {
2539 			abd_t *buf = abd_get_offset(data, offset);
2540 
2541 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2542 			    buf, BP_GET_PSIZE(bp));
2543 
2544 			abd_free(buf);
2545 		}
2546 		/*
2547 		 * If we are here to damage data for testing purposes,
2548 		 * leave the GBH alone so that we can detect the damage.
2549 		 */
2550 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2551 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2552 	} else {
2553 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2554 		    abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2555 		    zio_gang_issue_func_done, NULL, pio->io_priority,
2556 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2557 	}
2558 
2559 	return (zio);
2560 }
2561 
2562 static zio_t *
2563 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2564     uint64_t offset)
2565 {
2566 	(void) gn, (void) data, (void) offset;
2567 
2568 	zio_t *zio = zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2569 	    ZIO_GANG_CHILD_FLAGS(pio));
2570 	if (zio == NULL) {
2571 		zio = zio_null(pio, pio->io_spa,
2572 		    NULL, NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio));
2573 	}
2574 	return (zio);
2575 }
2576 
2577 static zio_t *
2578 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2579     uint64_t offset)
2580 {
2581 	(void) gn, (void) data, (void) offset;
2582 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2583 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2584 }
2585 
2586 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2587 	NULL,
2588 	zio_read_gang,
2589 	zio_rewrite_gang,
2590 	zio_free_gang,
2591 	zio_claim_gang,
2592 	NULL
2593 };
2594 
2595 static void zio_gang_tree_assemble_done(zio_t *zio);
2596 
2597 static zio_gang_node_t *
2598 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2599 {
2600 	zio_gang_node_t *gn;
2601 
2602 	ASSERT(*gnpp == NULL);
2603 
2604 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2605 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2606 	*gnpp = gn;
2607 
2608 	return (gn);
2609 }
2610 
2611 static void
2612 zio_gang_node_free(zio_gang_node_t **gnpp)
2613 {
2614 	zio_gang_node_t *gn = *gnpp;
2615 
2616 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2617 		ASSERT(gn->gn_child[g] == NULL);
2618 
2619 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2620 	kmem_free(gn, sizeof (*gn));
2621 	*gnpp = NULL;
2622 }
2623 
2624 static void
2625 zio_gang_tree_free(zio_gang_node_t **gnpp)
2626 {
2627 	zio_gang_node_t *gn = *gnpp;
2628 
2629 	if (gn == NULL)
2630 		return;
2631 
2632 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2633 		zio_gang_tree_free(&gn->gn_child[g]);
2634 
2635 	zio_gang_node_free(gnpp);
2636 }
2637 
2638 static void
2639 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2640 {
2641 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2642 	abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2643 
2644 	ASSERT(gio->io_gang_leader == gio);
2645 	ASSERT(BP_IS_GANG(bp));
2646 
2647 	zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2648 	    zio_gang_tree_assemble_done, gn, gio->io_priority,
2649 	    ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2650 }
2651 
2652 static void
2653 zio_gang_tree_assemble_done(zio_t *zio)
2654 {
2655 	zio_t *gio = zio->io_gang_leader;
2656 	zio_gang_node_t *gn = zio->io_private;
2657 	blkptr_t *bp = zio->io_bp;
2658 
2659 	ASSERT(gio == zio_unique_parent(zio));
2660 	ASSERT(zio->io_child_count == 0);
2661 
2662 	if (zio->io_error)
2663 		return;
2664 
2665 	/* this ABD was created from a linear buf in zio_gang_tree_assemble */
2666 	if (BP_SHOULD_BYTESWAP(bp))
2667 		byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2668 
2669 	ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2670 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2671 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2672 
2673 	abd_free(zio->io_abd);
2674 
2675 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2676 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2677 		if (!BP_IS_GANG(gbp))
2678 			continue;
2679 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2680 	}
2681 }
2682 
2683 static void
2684 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2685     uint64_t offset)
2686 {
2687 	zio_t *gio = pio->io_gang_leader;
2688 	zio_t *zio;
2689 
2690 	ASSERT(BP_IS_GANG(bp) == !!gn);
2691 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2692 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2693 
2694 	/*
2695 	 * If you're a gang header, your data is in gn->gn_gbh.
2696 	 * If you're a gang member, your data is in 'data' and gn == NULL.
2697 	 */
2698 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2699 
2700 	if (gn != NULL) {
2701 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2702 
2703 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2704 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2705 			if (BP_IS_HOLE(gbp))
2706 				continue;
2707 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2708 			    offset);
2709 			offset += BP_GET_PSIZE(gbp);
2710 		}
2711 	}
2712 
2713 	if (gn == gio->io_gang_tree)
2714 		ASSERT3U(gio->io_size, ==, offset);
2715 
2716 	if (zio != pio)
2717 		zio_nowait(zio);
2718 }
2719 
2720 static zio_t *
2721 zio_gang_assemble(zio_t *zio)
2722 {
2723 	blkptr_t *bp = zio->io_bp;
2724 
2725 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2726 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2727 
2728 	zio->io_gang_leader = zio;
2729 
2730 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2731 
2732 	return (zio);
2733 }
2734 
2735 static zio_t *
2736 zio_gang_issue(zio_t *zio)
2737 {
2738 	blkptr_t *bp = zio->io_bp;
2739 
2740 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2741 		return (NULL);
2742 	}
2743 
2744 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2745 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2746 
2747 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2748 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2749 		    0);
2750 	else
2751 		zio_gang_tree_free(&zio->io_gang_tree);
2752 
2753 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2754 
2755 	return (zio);
2756 }
2757 
2758 static void
2759 zio_write_gang_member_ready(zio_t *zio)
2760 {
2761 	zio_t *pio = zio_unique_parent(zio);
2762 	dva_t *cdva = zio->io_bp->blk_dva;
2763 	dva_t *pdva = pio->io_bp->blk_dva;
2764 	uint64_t asize;
2765 	zio_t *gio __maybe_unused = zio->io_gang_leader;
2766 
2767 	if (BP_IS_HOLE(zio->io_bp))
2768 		return;
2769 
2770 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2771 
2772 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2773 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2774 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2775 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2776 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2777 
2778 	mutex_enter(&pio->io_lock);
2779 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2780 		ASSERT(DVA_GET_GANG(&pdva[d]));
2781 		asize = DVA_GET_ASIZE(&pdva[d]);
2782 		asize += DVA_GET_ASIZE(&cdva[d]);
2783 		DVA_SET_ASIZE(&pdva[d], asize);
2784 	}
2785 	mutex_exit(&pio->io_lock);
2786 }
2787 
2788 static void
2789 zio_write_gang_done(zio_t *zio)
2790 {
2791 	/*
2792 	 * The io_abd field will be NULL for a zio with no data.  The io_flags
2793 	 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2794 	 * check for it here as it is cleared in zio_ready.
2795 	 */
2796 	if (zio->io_abd != NULL)
2797 		abd_free(zio->io_abd);
2798 }
2799 
2800 static zio_t *
2801 zio_write_gang_block(zio_t *pio, metaslab_class_t *mc)
2802 {
2803 	spa_t *spa = pio->io_spa;
2804 	blkptr_t *bp = pio->io_bp;
2805 	zio_t *gio = pio->io_gang_leader;
2806 	zio_t *zio;
2807 	zio_gang_node_t *gn, **gnpp;
2808 	zio_gbh_phys_t *gbh;
2809 	abd_t *gbh_abd;
2810 	uint64_t txg = pio->io_txg;
2811 	uint64_t resid = pio->io_size;
2812 	uint64_t lsize;
2813 	int copies = gio->io_prop.zp_copies;
2814 	int gbh_copies;
2815 	zio_prop_t zp;
2816 	int error;
2817 	boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
2818 
2819 	/*
2820 	 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2821 	 * have a third copy.
2822 	 */
2823 	gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2824 	if (gio->io_prop.zp_encrypt && gbh_copies >= SPA_DVAS_PER_BP)
2825 		gbh_copies = SPA_DVAS_PER_BP - 1;
2826 
2827 	int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2828 	if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2829 		ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2830 		ASSERT(has_data);
2831 
2832 		flags |= METASLAB_ASYNC_ALLOC;
2833 		VERIFY(zfs_refcount_held(&mc->mc_allocator[pio->io_allocator].
2834 		    mca_alloc_slots, pio));
2835 
2836 		/*
2837 		 * The logical zio has already placed a reservation for
2838 		 * 'copies' allocation slots but gang blocks may require
2839 		 * additional copies. These additional copies
2840 		 * (i.e. gbh_copies - copies) are guaranteed to succeed
2841 		 * since metaslab_class_throttle_reserve() always allows
2842 		 * additional reservations for gang blocks.
2843 		 */
2844 		VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2845 		    pio->io_allocator, pio, flags));
2846 	}
2847 
2848 	error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2849 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2850 	    &pio->io_alloc_list, pio, pio->io_allocator);
2851 	if (error) {
2852 		if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2853 			ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2854 			ASSERT(has_data);
2855 
2856 			/*
2857 			 * If we failed to allocate the gang block header then
2858 			 * we remove any additional allocation reservations that
2859 			 * we placed here. The original reservation will
2860 			 * be removed when the logical I/O goes to the ready
2861 			 * stage.
2862 			 */
2863 			metaslab_class_throttle_unreserve(mc,
2864 			    gbh_copies - copies, pio->io_allocator, pio);
2865 		}
2866 
2867 		pio->io_error = error;
2868 		return (pio);
2869 	}
2870 
2871 	if (pio == gio) {
2872 		gnpp = &gio->io_gang_tree;
2873 	} else {
2874 		gnpp = pio->io_private;
2875 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
2876 	}
2877 
2878 	gn = zio_gang_node_alloc(gnpp);
2879 	gbh = gn->gn_gbh;
2880 	bzero(gbh, SPA_GANGBLOCKSIZE);
2881 	gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2882 
2883 	/*
2884 	 * Create the gang header.
2885 	 */
2886 	zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2887 	    zio_write_gang_done, NULL, pio->io_priority,
2888 	    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2889 
2890 	/*
2891 	 * Create and nowait the gang children.
2892 	 */
2893 	for (int g = 0; resid != 0; resid -= lsize, g++) {
2894 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2895 		    SPA_MINBLOCKSIZE);
2896 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2897 
2898 		zp.zp_checksum = gio->io_prop.zp_checksum;
2899 		zp.zp_compress = ZIO_COMPRESS_OFF;
2900 		zp.zp_complevel = gio->io_prop.zp_complevel;
2901 		zp.zp_type = DMU_OT_NONE;
2902 		zp.zp_level = 0;
2903 		zp.zp_copies = gio->io_prop.zp_copies;
2904 		zp.zp_dedup = B_FALSE;
2905 		zp.zp_dedup_verify = B_FALSE;
2906 		zp.zp_nopwrite = B_FALSE;
2907 		zp.zp_encrypt = gio->io_prop.zp_encrypt;
2908 		zp.zp_byteorder = gio->io_prop.zp_byteorder;
2909 		bzero(zp.zp_salt, ZIO_DATA_SALT_LEN);
2910 		bzero(zp.zp_iv, ZIO_DATA_IV_LEN);
2911 		bzero(zp.zp_mac, ZIO_DATA_MAC_LEN);
2912 
2913 		zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2914 		    has_data ? abd_get_offset(pio->io_abd, pio->io_size -
2915 		    resid) : NULL, lsize, lsize, &zp,
2916 		    zio_write_gang_member_ready, NULL, NULL,
2917 		    zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2918 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2919 
2920 		if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2921 			ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2922 			ASSERT(has_data);
2923 
2924 			/*
2925 			 * Gang children won't throttle but we should
2926 			 * account for their work, so reserve an allocation
2927 			 * slot for them here.
2928 			 */
2929 			VERIFY(metaslab_class_throttle_reserve(mc,
2930 			    zp.zp_copies, cio->io_allocator, cio, flags));
2931 		}
2932 		zio_nowait(cio);
2933 	}
2934 
2935 	/*
2936 	 * Set pio's pipeline to just wait for zio to finish.
2937 	 */
2938 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2939 
2940 	/*
2941 	 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2942 	 */
2943 	pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
2944 
2945 	zio_nowait(zio);
2946 
2947 	return (pio);
2948 }
2949 
2950 /*
2951  * The zio_nop_write stage in the pipeline determines if allocating a
2952  * new bp is necessary.  The nopwrite feature can handle writes in
2953  * either syncing or open context (i.e. zil writes) and as a result is
2954  * mutually exclusive with dedup.
2955  *
2956  * By leveraging a cryptographically secure checksum, such as SHA256, we
2957  * can compare the checksums of the new data and the old to determine if
2958  * allocating a new block is required.  Note that our requirements for
2959  * cryptographic strength are fairly weak: there can't be any accidental
2960  * hash collisions, but we don't need to be secure against intentional
2961  * (malicious) collisions.  To trigger a nopwrite, you have to be able
2962  * to write the file to begin with, and triggering an incorrect (hash
2963  * collision) nopwrite is no worse than simply writing to the file.
2964  * That said, there are no known attacks against the checksum algorithms
2965  * used for nopwrite, assuming that the salt and the checksums
2966  * themselves remain secret.
2967  */
2968 static zio_t *
2969 zio_nop_write(zio_t *zio)
2970 {
2971 	blkptr_t *bp = zio->io_bp;
2972 	blkptr_t *bp_orig = &zio->io_bp_orig;
2973 	zio_prop_t *zp = &zio->io_prop;
2974 
2975 	ASSERT(BP_GET_LEVEL(bp) == 0);
2976 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2977 	ASSERT(zp->zp_nopwrite);
2978 	ASSERT(!zp->zp_dedup);
2979 	ASSERT(zio->io_bp_override == NULL);
2980 	ASSERT(IO_IS_ALLOCATING(zio));
2981 
2982 	/*
2983 	 * Check to see if the original bp and the new bp have matching
2984 	 * characteristics (i.e. same checksum, compression algorithms, etc).
2985 	 * If they don't then just continue with the pipeline which will
2986 	 * allocate a new bp.
2987 	 */
2988 	if (BP_IS_HOLE(bp_orig) ||
2989 	    !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2990 	    ZCHECKSUM_FLAG_NOPWRITE) ||
2991 	    BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) ||
2992 	    BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2993 	    BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2994 	    BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2995 	    zp->zp_copies != BP_GET_NDVAS(bp_orig))
2996 		return (zio);
2997 
2998 	/*
2999 	 * If the checksums match then reset the pipeline so that we
3000 	 * avoid allocating a new bp and issuing any I/O.
3001 	 */
3002 	if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
3003 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
3004 		    ZCHECKSUM_FLAG_NOPWRITE);
3005 		ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
3006 		ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
3007 		ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
3008 		ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
3009 		    sizeof (uint64_t)) == 0);
3010 
3011 		/*
3012 		 * If we're overwriting a block that is currently on an
3013 		 * indirect vdev, then ignore the nopwrite request and
3014 		 * allow a new block to be allocated on a concrete vdev.
3015 		 */
3016 		spa_config_enter(zio->io_spa, SCL_VDEV, FTAG, RW_READER);
3017 		vdev_t *tvd = vdev_lookup_top(zio->io_spa,
3018 		    DVA_GET_VDEV(&bp->blk_dva[0]));
3019 		if (tvd->vdev_ops == &vdev_indirect_ops) {
3020 			spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3021 			return (zio);
3022 		}
3023 		spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3024 
3025 		*bp = *bp_orig;
3026 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3027 		zio->io_flags |= ZIO_FLAG_NOPWRITE;
3028 	}
3029 
3030 	return (zio);
3031 }
3032 
3033 /*
3034  * ==========================================================================
3035  * Dedup
3036  * ==========================================================================
3037  */
3038 static void
3039 zio_ddt_child_read_done(zio_t *zio)
3040 {
3041 	blkptr_t *bp = zio->io_bp;
3042 	ddt_entry_t *dde = zio->io_private;
3043 	ddt_phys_t *ddp;
3044 	zio_t *pio = zio_unique_parent(zio);
3045 
3046 	mutex_enter(&pio->io_lock);
3047 	ddp = ddt_phys_select(dde, bp);
3048 	if (zio->io_error == 0)
3049 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
3050 
3051 	if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
3052 		dde->dde_repair_abd = zio->io_abd;
3053 	else
3054 		abd_free(zio->io_abd);
3055 	mutex_exit(&pio->io_lock);
3056 }
3057 
3058 static zio_t *
3059 zio_ddt_read_start(zio_t *zio)
3060 {
3061 	blkptr_t *bp = zio->io_bp;
3062 
3063 	ASSERT(BP_GET_DEDUP(bp));
3064 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3065 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3066 
3067 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
3068 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
3069 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
3070 		ddt_phys_t *ddp = dde->dde_phys;
3071 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
3072 		blkptr_t blk;
3073 
3074 		ASSERT(zio->io_vsd == NULL);
3075 		zio->io_vsd = dde;
3076 
3077 		if (ddp_self == NULL)
3078 			return (zio);
3079 
3080 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
3081 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
3082 				continue;
3083 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
3084 			    &blk);
3085 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
3086 			    abd_alloc_for_io(zio->io_size, B_TRUE),
3087 			    zio->io_size, zio_ddt_child_read_done, dde,
3088 			    zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
3089 			    ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
3090 		}
3091 		return (zio);
3092 	}
3093 
3094 	zio_nowait(zio_read(zio, zio->io_spa, bp,
3095 	    zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
3096 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
3097 
3098 	return (zio);
3099 }
3100 
3101 static zio_t *
3102 zio_ddt_read_done(zio_t *zio)
3103 {
3104 	blkptr_t *bp = zio->io_bp;
3105 
3106 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
3107 		return (NULL);
3108 	}
3109 
3110 	ASSERT(BP_GET_DEDUP(bp));
3111 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3112 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3113 
3114 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
3115 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
3116 		ddt_entry_t *dde = zio->io_vsd;
3117 		if (ddt == NULL) {
3118 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
3119 			return (zio);
3120 		}
3121 		if (dde == NULL) {
3122 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
3123 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
3124 			return (NULL);
3125 		}
3126 		if (dde->dde_repair_abd != NULL) {
3127 			abd_copy(zio->io_abd, dde->dde_repair_abd,
3128 			    zio->io_size);
3129 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
3130 		}
3131 		ddt_repair_done(ddt, dde);
3132 		zio->io_vsd = NULL;
3133 	}
3134 
3135 	ASSERT(zio->io_vsd == NULL);
3136 
3137 	return (zio);
3138 }
3139 
3140 static boolean_t
3141 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
3142 {
3143 	spa_t *spa = zio->io_spa;
3144 	boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW);
3145 
3146 	ASSERT(!(zio->io_bp_override && do_raw));
3147 
3148 	/*
3149 	 * Note: we compare the original data, not the transformed data,
3150 	 * because when zio->io_bp is an override bp, we will not have
3151 	 * pushed the I/O transforms.  That's an important optimization
3152 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
3153 	 * However, we should never get a raw, override zio so in these
3154 	 * cases we can compare the io_abd directly. This is useful because
3155 	 * it allows us to do dedup verification even if we don't have access
3156 	 * to the original data (for instance, if the encryption keys aren't
3157 	 * loaded).
3158 	 */
3159 
3160 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
3161 		zio_t *lio = dde->dde_lead_zio[p];
3162 
3163 		if (lio != NULL && do_raw) {
3164 			return (lio->io_size != zio->io_size ||
3165 			    abd_cmp(zio->io_abd, lio->io_abd) != 0);
3166 		} else if (lio != NULL) {
3167 			return (lio->io_orig_size != zio->io_orig_size ||
3168 			    abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0);
3169 		}
3170 	}
3171 
3172 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
3173 		ddt_phys_t *ddp = &dde->dde_phys[p];
3174 
3175 		if (ddp->ddp_phys_birth != 0 && do_raw) {
3176 			blkptr_t blk = *zio->io_bp;
3177 			uint64_t psize;
3178 			abd_t *tmpabd;
3179 			int error;
3180 
3181 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
3182 			psize = BP_GET_PSIZE(&blk);
3183 
3184 			if (psize != zio->io_size)
3185 				return (B_TRUE);
3186 
3187 			ddt_exit(ddt);
3188 
3189 			tmpabd = abd_alloc_for_io(psize, B_TRUE);
3190 
3191 			error = zio_wait(zio_read(NULL, spa, &blk, tmpabd,
3192 			    psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
3193 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
3194 			    ZIO_FLAG_RAW, &zio->io_bookmark));
3195 
3196 			if (error == 0) {
3197 				if (abd_cmp(tmpabd, zio->io_abd) != 0)
3198 					error = SET_ERROR(ENOENT);
3199 			}
3200 
3201 			abd_free(tmpabd);
3202 			ddt_enter(ddt);
3203 			return (error != 0);
3204 		} else if (ddp->ddp_phys_birth != 0) {
3205 			arc_buf_t *abuf = NULL;
3206 			arc_flags_t aflags = ARC_FLAG_WAIT;
3207 			blkptr_t blk = *zio->io_bp;
3208 			int error;
3209 
3210 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
3211 
3212 			if (BP_GET_LSIZE(&blk) != zio->io_orig_size)
3213 				return (B_TRUE);
3214 
3215 			ddt_exit(ddt);
3216 
3217 			error = arc_read(NULL, spa, &blk,
3218 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
3219 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
3220 			    &aflags, &zio->io_bookmark);
3221 
3222 			if (error == 0) {
3223 				if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
3224 				    zio->io_orig_size) != 0)
3225 					error = SET_ERROR(ENOENT);
3226 				arc_buf_destroy(abuf, &abuf);
3227 			}
3228 
3229 			ddt_enter(ddt);
3230 			return (error != 0);
3231 		}
3232 	}
3233 
3234 	return (B_FALSE);
3235 }
3236 
3237 static void
3238 zio_ddt_child_write_ready(zio_t *zio)
3239 {
3240 	int p = zio->io_prop.zp_copies;
3241 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3242 	ddt_entry_t *dde = zio->io_private;
3243 	ddt_phys_t *ddp = &dde->dde_phys[p];
3244 	zio_t *pio;
3245 
3246 	if (zio->io_error)
3247 		return;
3248 
3249 	ddt_enter(ddt);
3250 
3251 	ASSERT(dde->dde_lead_zio[p] == zio);
3252 
3253 	ddt_phys_fill(ddp, zio->io_bp);
3254 
3255 	zio_link_t *zl = NULL;
3256 	while ((pio = zio_walk_parents(zio, &zl)) != NULL)
3257 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
3258 
3259 	ddt_exit(ddt);
3260 }
3261 
3262 static void
3263 zio_ddt_child_write_done(zio_t *zio)
3264 {
3265 	int p = zio->io_prop.zp_copies;
3266 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3267 	ddt_entry_t *dde = zio->io_private;
3268 	ddt_phys_t *ddp = &dde->dde_phys[p];
3269 
3270 	ddt_enter(ddt);
3271 
3272 	ASSERT(ddp->ddp_refcnt == 0);
3273 	ASSERT(dde->dde_lead_zio[p] == zio);
3274 	dde->dde_lead_zio[p] = NULL;
3275 
3276 	if (zio->io_error == 0) {
3277 		zio_link_t *zl = NULL;
3278 		while (zio_walk_parents(zio, &zl) != NULL)
3279 			ddt_phys_addref(ddp);
3280 	} else {
3281 		ddt_phys_clear(ddp);
3282 	}
3283 
3284 	ddt_exit(ddt);
3285 }
3286 
3287 static zio_t *
3288 zio_ddt_write(zio_t *zio)
3289 {
3290 	spa_t *spa = zio->io_spa;
3291 	blkptr_t *bp = zio->io_bp;
3292 	uint64_t txg = zio->io_txg;
3293 	zio_prop_t *zp = &zio->io_prop;
3294 	int p = zp->zp_copies;
3295 	zio_t *cio = NULL;
3296 	ddt_t *ddt = ddt_select(spa, bp);
3297 	ddt_entry_t *dde;
3298 	ddt_phys_t *ddp;
3299 
3300 	ASSERT(BP_GET_DEDUP(bp));
3301 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
3302 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
3303 	ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
3304 
3305 	ddt_enter(ddt);
3306 	dde = ddt_lookup(ddt, bp, B_TRUE);
3307 	ddp = &dde->dde_phys[p];
3308 
3309 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
3310 		/*
3311 		 * If we're using a weak checksum, upgrade to a strong checksum
3312 		 * and try again.  If we're already using a strong checksum,
3313 		 * we can't resolve it, so just convert to an ordinary write.
3314 		 * (And automatically e-mail a paper to Nature?)
3315 		 */
3316 		if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
3317 		    ZCHECKSUM_FLAG_DEDUP)) {
3318 			zp->zp_checksum = spa_dedup_checksum(spa);
3319 			zio_pop_transforms(zio);
3320 			zio->io_stage = ZIO_STAGE_OPEN;
3321 			BP_ZERO(bp);
3322 		} else {
3323 			zp->zp_dedup = B_FALSE;
3324 			BP_SET_DEDUP(bp, B_FALSE);
3325 		}
3326 		ASSERT(!BP_GET_DEDUP(bp));
3327 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
3328 		ddt_exit(ddt);
3329 		return (zio);
3330 	}
3331 
3332 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
3333 		if (ddp->ddp_phys_birth != 0)
3334 			ddt_bp_fill(ddp, bp, txg);
3335 		if (dde->dde_lead_zio[p] != NULL)
3336 			zio_add_child(zio, dde->dde_lead_zio[p]);
3337 		else
3338 			ddt_phys_addref(ddp);
3339 	} else if (zio->io_bp_override) {
3340 		ASSERT(bp->blk_birth == txg);
3341 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
3342 		ddt_phys_fill(ddp, bp);
3343 		ddt_phys_addref(ddp);
3344 	} else {
3345 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3346 		    zio->io_orig_size, zio->io_orig_size, zp,
3347 		    zio_ddt_child_write_ready, NULL, NULL,
3348 		    zio_ddt_child_write_done, dde, zio->io_priority,
3349 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3350 
3351 		zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
3352 		dde->dde_lead_zio[p] = cio;
3353 	}
3354 
3355 	ddt_exit(ddt);
3356 
3357 	zio_nowait(cio);
3358 
3359 	return (zio);
3360 }
3361 
3362 ddt_entry_t *freedde; /* for debugging */
3363 
3364 static zio_t *
3365 zio_ddt_free(zio_t *zio)
3366 {
3367 	spa_t *spa = zio->io_spa;
3368 	blkptr_t *bp = zio->io_bp;
3369 	ddt_t *ddt = ddt_select(spa, bp);
3370 	ddt_entry_t *dde;
3371 	ddt_phys_t *ddp;
3372 
3373 	ASSERT(BP_GET_DEDUP(bp));
3374 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3375 
3376 	ddt_enter(ddt);
3377 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
3378 	if (dde) {
3379 		ddp = ddt_phys_select(dde, bp);
3380 		if (ddp)
3381 			ddt_phys_decref(ddp);
3382 	}
3383 	ddt_exit(ddt);
3384 
3385 	return (zio);
3386 }
3387 
3388 /*
3389  * ==========================================================================
3390  * Allocate and free blocks
3391  * ==========================================================================
3392  */
3393 
3394 static zio_t *
3395 zio_io_to_allocate(spa_t *spa, int allocator)
3396 {
3397 	zio_t *zio;
3398 
3399 	ASSERT(MUTEX_HELD(&spa->spa_allocs[allocator].spaa_lock));
3400 
3401 	zio = avl_first(&spa->spa_allocs[allocator].spaa_tree);
3402 	if (zio == NULL)
3403 		return (NULL);
3404 
3405 	ASSERT(IO_IS_ALLOCATING(zio));
3406 
3407 	/*
3408 	 * Try to place a reservation for this zio. If we're unable to
3409 	 * reserve then we throttle.
3410 	 */
3411 	ASSERT3U(zio->io_allocator, ==, allocator);
3412 	if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
3413 	    zio->io_prop.zp_copies, allocator, zio, 0)) {
3414 		return (NULL);
3415 	}
3416 
3417 	avl_remove(&spa->spa_allocs[allocator].spaa_tree, zio);
3418 	ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
3419 
3420 	return (zio);
3421 }
3422 
3423 static zio_t *
3424 zio_dva_throttle(zio_t *zio)
3425 {
3426 	spa_t *spa = zio->io_spa;
3427 	zio_t *nio;
3428 	metaslab_class_t *mc;
3429 
3430 	/* locate an appropriate allocation class */
3431 	mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type,
3432 	    zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk);
3433 
3434 	if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
3435 	    !mc->mc_alloc_throttle_enabled ||
3436 	    zio->io_child_type == ZIO_CHILD_GANG ||
3437 	    zio->io_flags & ZIO_FLAG_NODATA) {
3438 		return (zio);
3439 	}
3440 
3441 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3442 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3443 	ASSERT3U(zio->io_queued_timestamp, >, 0);
3444 	ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3445 
3446 	zbookmark_phys_t *bm = &zio->io_bookmark;
3447 	/*
3448 	 * We want to try to use as many allocators as possible to help improve
3449 	 * performance, but we also want logically adjacent IOs to be physically
3450 	 * adjacent to improve sequential read performance. We chunk each object
3451 	 * into 2^20 block regions, and then hash based on the objset, object,
3452 	 * level, and region to accomplish both of these goals.
3453 	 */
3454 	int allocator = (uint_t)cityhash4(bm->zb_objset, bm->zb_object,
3455 	    bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
3456 	zio->io_allocator = allocator;
3457 	zio->io_metaslab_class = mc;
3458 	mutex_enter(&spa->spa_allocs[allocator].spaa_lock);
3459 	avl_add(&spa->spa_allocs[allocator].spaa_tree, zio);
3460 	nio = zio_io_to_allocate(spa, allocator);
3461 	mutex_exit(&spa->spa_allocs[allocator].spaa_lock);
3462 	return (nio);
3463 }
3464 
3465 static void
3466 zio_allocate_dispatch(spa_t *spa, int allocator)
3467 {
3468 	zio_t *zio;
3469 
3470 	mutex_enter(&spa->spa_allocs[allocator].spaa_lock);
3471 	zio = zio_io_to_allocate(spa, allocator);
3472 	mutex_exit(&spa->spa_allocs[allocator].spaa_lock);
3473 	if (zio == NULL)
3474 		return;
3475 
3476 	ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3477 	ASSERT0(zio->io_error);
3478 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3479 }
3480 
3481 static zio_t *
3482 zio_dva_allocate(zio_t *zio)
3483 {
3484 	spa_t *spa = zio->io_spa;
3485 	metaslab_class_t *mc;
3486 	blkptr_t *bp = zio->io_bp;
3487 	int error;
3488 	int flags = 0;
3489 
3490 	if (zio->io_gang_leader == NULL) {
3491 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3492 		zio->io_gang_leader = zio;
3493 	}
3494 
3495 	ASSERT(BP_IS_HOLE(bp));
3496 	ASSERT0(BP_GET_NDVAS(bp));
3497 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
3498 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3499 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3500 
3501 	flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
3502 	if (zio->io_flags & ZIO_FLAG_NODATA)
3503 		flags |= METASLAB_DONT_THROTTLE;
3504 	if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
3505 		flags |= METASLAB_GANG_CHILD;
3506 	if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
3507 		flags |= METASLAB_ASYNC_ALLOC;
3508 
3509 	/*
3510 	 * if not already chosen, locate an appropriate allocation class
3511 	 */
3512 	mc = zio->io_metaslab_class;
3513 	if (mc == NULL) {
3514 		mc = spa_preferred_class(spa, zio->io_size,
3515 		    zio->io_prop.zp_type, zio->io_prop.zp_level,
3516 		    zio->io_prop.zp_zpl_smallblk);
3517 		zio->io_metaslab_class = mc;
3518 	}
3519 
3520 	/*
3521 	 * Try allocating the block in the usual metaslab class.
3522 	 * If that's full, allocate it in the normal class.
3523 	 * If that's full, allocate as a gang block,
3524 	 * and if all are full, the allocation fails (which shouldn't happen).
3525 	 *
3526 	 * Note that we do not fall back on embedded slog (ZIL) space, to
3527 	 * preserve unfragmented slog space, which is critical for decent
3528 	 * sync write performance.  If a log allocation fails, we will fall
3529 	 * back to spa_sync() which is abysmal for performance.
3530 	 */
3531 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
3532 	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3533 	    &zio->io_alloc_list, zio, zio->io_allocator);
3534 
3535 	/*
3536 	 * Fallback to normal class when an alloc class is full
3537 	 */
3538 	if (error == ENOSPC && mc != spa_normal_class(spa)) {
3539 		/*
3540 		 * If throttling, transfer reservation over to normal class.
3541 		 * The io_allocator slot can remain the same even though we
3542 		 * are switching classes.
3543 		 */
3544 		if (mc->mc_alloc_throttle_enabled &&
3545 		    (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) {
3546 			metaslab_class_throttle_unreserve(mc,
3547 			    zio->io_prop.zp_copies, zio->io_allocator, zio);
3548 			zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING;
3549 
3550 			VERIFY(metaslab_class_throttle_reserve(
3551 			    spa_normal_class(spa),
3552 			    zio->io_prop.zp_copies, zio->io_allocator, zio,
3553 			    flags | METASLAB_MUST_RESERVE));
3554 		}
3555 		zio->io_metaslab_class = mc = spa_normal_class(spa);
3556 		if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) {
3557 			zfs_dbgmsg("%s: metaslab allocation failure, "
3558 			    "trying normal class: zio %px, size %llu, error %d",
3559 			    spa_name(spa), zio, (u_longlong_t)zio->io_size,
3560 			    error);
3561 		}
3562 
3563 		error = metaslab_alloc(spa, mc, zio->io_size, bp,
3564 		    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3565 		    &zio->io_alloc_list, zio, zio->io_allocator);
3566 	}
3567 
3568 	if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) {
3569 		if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) {
3570 			zfs_dbgmsg("%s: metaslab allocation failure, "
3571 			    "trying ganging: zio %px, size %llu, error %d",
3572 			    spa_name(spa), zio, (u_longlong_t)zio->io_size,
3573 			    error);
3574 		}
3575 		return (zio_write_gang_block(zio, mc));
3576 	}
3577 	if (error != 0) {
3578 		if (error != ENOSPC ||
3579 		    (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC)) {
3580 			zfs_dbgmsg("%s: metaslab allocation failure: zio %px, "
3581 			    "size %llu, error %d",
3582 			    spa_name(spa), zio, (u_longlong_t)zio->io_size,
3583 			    error);
3584 		}
3585 		zio->io_error = error;
3586 	}
3587 
3588 	return (zio);
3589 }
3590 
3591 static zio_t *
3592 zio_dva_free(zio_t *zio)
3593 {
3594 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3595 
3596 	return (zio);
3597 }
3598 
3599 static zio_t *
3600 zio_dva_claim(zio_t *zio)
3601 {
3602 	int error;
3603 
3604 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3605 	if (error)
3606 		zio->io_error = error;
3607 
3608 	return (zio);
3609 }
3610 
3611 /*
3612  * Undo an allocation.  This is used by zio_done() when an I/O fails
3613  * and we want to give back the block we just allocated.
3614  * This handles both normal blocks and gang blocks.
3615  */
3616 static void
3617 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3618 {
3619 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3620 	ASSERT(zio->io_bp_override == NULL);
3621 
3622 	if (!BP_IS_HOLE(bp))
3623 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3624 
3625 	if (gn != NULL) {
3626 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3627 			zio_dva_unallocate(zio, gn->gn_child[g],
3628 			    &gn->gn_gbh->zg_blkptr[g]);
3629 		}
3630 	}
3631 }
3632 
3633 /*
3634  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
3635  */
3636 int
3637 zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
3638     uint64_t size, boolean_t *slog)
3639 {
3640 	int error = 1;
3641 	zio_alloc_list_t io_alloc_list;
3642 
3643 	ASSERT(txg > spa_syncing_txg(spa));
3644 
3645 	metaslab_trace_init(&io_alloc_list);
3646 
3647 	/*
3648 	 * Block pointer fields are useful to metaslabs for stats and debugging.
3649 	 * Fill in the obvious ones before calling into metaslab_alloc().
3650 	 */
3651 	BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3652 	BP_SET_PSIZE(new_bp, size);
3653 	BP_SET_LEVEL(new_bp, 0);
3654 
3655 	/*
3656 	 * When allocating a zil block, we don't have information about
3657 	 * the final destination of the block except the objset it's part
3658 	 * of, so we just hash the objset ID to pick the allocator to get
3659 	 * some parallelism.
3660 	 */
3661 	int flags = METASLAB_FASTWRITE | METASLAB_ZIL;
3662 	int allocator = (uint_t)cityhash4(0, 0, 0,
3663 	    os->os_dsl_dataset->ds_object) % spa->spa_alloc_count;
3664 	error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3665 	    txg, NULL, flags, &io_alloc_list, NULL, allocator);
3666 	*slog = (error == 0);
3667 	if (error != 0) {
3668 		error = metaslab_alloc(spa, spa_embedded_log_class(spa), size,
3669 		    new_bp, 1, txg, NULL, flags,
3670 		    &io_alloc_list, NULL, allocator);
3671 	}
3672 	if (error != 0) {
3673 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
3674 		    new_bp, 1, txg, NULL, flags,
3675 		    &io_alloc_list, NULL, allocator);
3676 	}
3677 	metaslab_trace_fini(&io_alloc_list);
3678 
3679 	if (error == 0) {
3680 		BP_SET_LSIZE(new_bp, size);
3681 		BP_SET_PSIZE(new_bp, size);
3682 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3683 		BP_SET_CHECKSUM(new_bp,
3684 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3685 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3686 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3687 		BP_SET_LEVEL(new_bp, 0);
3688 		BP_SET_DEDUP(new_bp, 0);
3689 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3690 
3691 		/*
3692 		 * encrypted blocks will require an IV and salt. We generate
3693 		 * these now since we will not be rewriting the bp at
3694 		 * rewrite time.
3695 		 */
3696 		if (os->os_encrypted) {
3697 			uint8_t iv[ZIO_DATA_IV_LEN];
3698 			uint8_t salt[ZIO_DATA_SALT_LEN];
3699 
3700 			BP_SET_CRYPT(new_bp, B_TRUE);
3701 			VERIFY0(spa_crypt_get_salt(spa,
3702 			    dmu_objset_id(os), salt));
3703 			VERIFY0(zio_crypt_generate_iv(iv));
3704 
3705 			zio_crypt_encode_params_bp(new_bp, salt, iv);
3706 		}
3707 	} else {
3708 		zfs_dbgmsg("%s: zil block allocation failure: "
3709 		    "size %llu, error %d", spa_name(spa), (u_longlong_t)size,
3710 		    error);
3711 	}
3712 
3713 	return (error);
3714 }
3715 
3716 /*
3717  * ==========================================================================
3718  * Read and write to physical devices
3719  * ==========================================================================
3720  */
3721 
3722 /*
3723  * Issue an I/O to the underlying vdev. Typically the issue pipeline
3724  * stops after this stage and will resume upon I/O completion.
3725  * However, there are instances where the vdev layer may need to
3726  * continue the pipeline when an I/O was not issued. Since the I/O
3727  * that was sent to the vdev layer might be different than the one
3728  * currently active in the pipeline (see vdev_queue_io()), we explicitly
3729  * force the underlying vdev layers to call either zio_execute() or
3730  * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3731  */
3732 static zio_t *
3733 zio_vdev_io_start(zio_t *zio)
3734 {
3735 	vdev_t *vd = zio->io_vd;
3736 	uint64_t align;
3737 	spa_t *spa = zio->io_spa;
3738 
3739 	zio->io_delay = 0;
3740 
3741 	ASSERT(zio->io_error == 0);
3742 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3743 
3744 	if (vd == NULL) {
3745 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3746 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3747 
3748 		/*
3749 		 * The mirror_ops handle multiple DVAs in a single BP.
3750 		 */
3751 		vdev_mirror_ops.vdev_op_io_start(zio);
3752 		return (NULL);
3753 	}
3754 
3755 	ASSERT3P(zio->io_logical, !=, zio);
3756 	if (zio->io_type == ZIO_TYPE_WRITE) {
3757 		ASSERT(spa->spa_trust_config);
3758 
3759 		/*
3760 		 * Note: the code can handle other kinds of writes,
3761 		 * but we don't expect them.
3762 		 */
3763 		if (zio->io_vd->vdev_noalloc) {
3764 			ASSERT(zio->io_flags &
3765 			    (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3766 			    ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3767 		}
3768 	}
3769 
3770 	align = 1ULL << vd->vdev_top->vdev_ashift;
3771 
3772 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3773 	    P2PHASE(zio->io_size, align) != 0) {
3774 		/* Transform logical writes to be a full physical block size. */
3775 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
3776 		abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3777 		ASSERT(vd == vd->vdev_top);
3778 		if (zio->io_type == ZIO_TYPE_WRITE) {
3779 			abd_copy(abuf, zio->io_abd, zio->io_size);
3780 			abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3781 		}
3782 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3783 	}
3784 
3785 	/*
3786 	 * If this is not a physical io, make sure that it is properly aligned
3787 	 * before proceeding.
3788 	 */
3789 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3790 		ASSERT0(P2PHASE(zio->io_offset, align));
3791 		ASSERT0(P2PHASE(zio->io_size, align));
3792 	} else {
3793 		/*
3794 		 * For physical writes, we allow 512b aligned writes and assume
3795 		 * the device will perform a read-modify-write as necessary.
3796 		 */
3797 		ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3798 		ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3799 	}
3800 
3801 	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
3802 
3803 	/*
3804 	 * If this is a repair I/O, and there's no self-healing involved --
3805 	 * that is, we're just resilvering what we expect to resilver --
3806 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3807 	 * This prevents spurious resilvering.
3808 	 *
3809 	 * There are a few ways that we can end up creating these spurious
3810 	 * resilver i/os:
3811 	 *
3812 	 * 1. A resilver i/o will be issued if any DVA in the BP has a
3813 	 * dirty DTL.  The mirror code will issue resilver writes to
3814 	 * each DVA, including the one(s) that are not on vdevs with dirty
3815 	 * DTLs.
3816 	 *
3817 	 * 2. With nested replication, which happens when we have a
3818 	 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3819 	 * For example, given mirror(replacing(A+B), C), it's likely that
3820 	 * only A is out of date (it's the new device). In this case, we'll
3821 	 * read from C, then use the data to resilver A+B -- but we don't
3822 	 * actually want to resilver B, just A. The top-level mirror has no
3823 	 * way to know this, so instead we just discard unnecessary repairs
3824 	 * as we work our way down the vdev tree.
3825 	 *
3826 	 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3827 	 * The same logic applies to any form of nested replication: ditto
3828 	 * + mirror, RAID-Z + replacing, etc.
3829 	 *
3830 	 * However, indirect vdevs point off to other vdevs which may have
3831 	 * DTL's, so we never bypass them.  The child i/os on concrete vdevs
3832 	 * will be properly bypassed instead.
3833 	 *
3834 	 * Leaf DTL_PARTIAL can be empty when a legitimate write comes from
3835 	 * a dRAID spare vdev. For example, when a dRAID spare is first
3836 	 * used, its spare blocks need to be written to but the leaf vdev's
3837 	 * of such blocks can have empty DTL_PARTIAL.
3838 	 *
3839 	 * There seemed no clean way to allow such writes while bypassing
3840 	 * spurious ones. At this point, just avoid all bypassing for dRAID
3841 	 * for correctness.
3842 	 */
3843 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3844 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3845 	    zio->io_txg != 0 &&	/* not a delegated i/o */
3846 	    vd->vdev_ops != &vdev_indirect_ops &&
3847 	    vd->vdev_top->vdev_ops != &vdev_draid_ops &&
3848 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3849 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3850 		zio_vdev_io_bypass(zio);
3851 		return (zio);
3852 	}
3853 
3854 	/*
3855 	 * Select the next best leaf I/O to process.  Distributed spares are
3856 	 * excluded since they dispatch the I/O directly to a leaf vdev after
3857 	 * applying the dRAID mapping.
3858 	 */
3859 	if (vd->vdev_ops->vdev_op_leaf &&
3860 	    vd->vdev_ops != &vdev_draid_spare_ops &&
3861 	    (zio->io_type == ZIO_TYPE_READ ||
3862 	    zio->io_type == ZIO_TYPE_WRITE ||
3863 	    zio->io_type == ZIO_TYPE_TRIM)) {
3864 
3865 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
3866 			return (zio);
3867 
3868 		if ((zio = vdev_queue_io(zio)) == NULL)
3869 			return (NULL);
3870 
3871 		if (!vdev_accessible(vd, zio)) {
3872 			zio->io_error = SET_ERROR(ENXIO);
3873 			zio_interrupt(zio);
3874 			return (NULL);
3875 		}
3876 		zio->io_delay = gethrtime();
3877 	}
3878 
3879 	vd->vdev_ops->vdev_op_io_start(zio);
3880 	return (NULL);
3881 }
3882 
3883 static zio_t *
3884 zio_vdev_io_done(zio_t *zio)
3885 {
3886 	vdev_t *vd = zio->io_vd;
3887 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3888 	boolean_t unexpected_error = B_FALSE;
3889 
3890 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3891 		return (NULL);
3892 	}
3893 
3894 	ASSERT(zio->io_type == ZIO_TYPE_READ ||
3895 	    zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM);
3896 
3897 	if (zio->io_delay)
3898 		zio->io_delay = gethrtime() - zio->io_delay;
3899 
3900 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3901 	    vd->vdev_ops != &vdev_draid_spare_ops) {
3902 		vdev_queue_io_done(zio);
3903 
3904 		if (zio->io_type == ZIO_TYPE_WRITE)
3905 			vdev_cache_write(zio);
3906 
3907 		if (zio_injection_enabled && zio->io_error == 0)
3908 			zio->io_error = zio_handle_device_injections(vd, zio,
3909 			    EIO, EILSEQ);
3910 
3911 		if (zio_injection_enabled && zio->io_error == 0)
3912 			zio->io_error = zio_handle_label_injection(zio, EIO);
3913 
3914 		if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) {
3915 			if (!vdev_accessible(vd, zio)) {
3916 				zio->io_error = SET_ERROR(ENXIO);
3917 			} else {
3918 				unexpected_error = B_TRUE;
3919 			}
3920 		}
3921 	}
3922 
3923 	ops->vdev_op_io_done(zio);
3924 
3925 	if (unexpected_error)
3926 		VERIFY(vdev_probe(vd, zio) == NULL);
3927 
3928 	return (zio);
3929 }
3930 
3931 /*
3932  * This function is used to change the priority of an existing zio that is
3933  * currently in-flight. This is used by the arc to upgrade priority in the
3934  * event that a demand read is made for a block that is currently queued
3935  * as a scrub or async read IO. Otherwise, the high priority read request
3936  * would end up having to wait for the lower priority IO.
3937  */
3938 void
3939 zio_change_priority(zio_t *pio, zio_priority_t priority)
3940 {
3941 	zio_t *cio, *cio_next;
3942 	zio_link_t *zl = NULL;
3943 
3944 	ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3945 
3946 	if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3947 		vdev_queue_change_io_priority(pio, priority);
3948 	} else {
3949 		pio->io_priority = priority;
3950 	}
3951 
3952 	mutex_enter(&pio->io_lock);
3953 	for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3954 		cio_next = zio_walk_children(pio, &zl);
3955 		zio_change_priority(cio, priority);
3956 	}
3957 	mutex_exit(&pio->io_lock);
3958 }
3959 
3960 /*
3961  * For non-raidz ZIOs, we can just copy aside the bad data read from the
3962  * disk, and use that to finish the checksum ereport later.
3963  */
3964 static void
3965 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3966     const abd_t *good_buf)
3967 {
3968 	/* no processing needed */
3969 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3970 }
3971 
3972 void
3973 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr)
3974 {
3975 	void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size);
3976 
3977 	abd_copy(abd, zio->io_abd, zio->io_size);
3978 
3979 	zcr->zcr_cbinfo = zio->io_size;
3980 	zcr->zcr_cbdata = abd;
3981 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
3982 	zcr->zcr_free = zio_abd_free;
3983 }
3984 
3985 static zio_t *
3986 zio_vdev_io_assess(zio_t *zio)
3987 {
3988 	vdev_t *vd = zio->io_vd;
3989 
3990 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3991 		return (NULL);
3992 	}
3993 
3994 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3995 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3996 
3997 	if (zio->io_vsd != NULL) {
3998 		zio->io_vsd_ops->vsd_free(zio);
3999 		zio->io_vsd = NULL;
4000 	}
4001 
4002 	if (zio_injection_enabled && zio->io_error == 0)
4003 		zio->io_error = zio_handle_fault_injection(zio, EIO);
4004 
4005 	/*
4006 	 * If the I/O failed, determine whether we should attempt to retry it.
4007 	 *
4008 	 * On retry, we cut in line in the issue queue, since we don't want
4009 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
4010 	 */
4011 	if (zio->io_error && vd == NULL &&
4012 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
4013 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
4014 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
4015 		zio->io_error = 0;
4016 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
4017 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
4018 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
4019 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
4020 		    zio_requeue_io_start_cut_in_line);
4021 		return (NULL);
4022 	}
4023 
4024 	/*
4025 	 * If we got an error on a leaf device, convert it to ENXIO
4026 	 * if the device is not accessible at all.
4027 	 */
4028 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
4029 	    !vdev_accessible(vd, zio))
4030 		zio->io_error = SET_ERROR(ENXIO);
4031 
4032 	/*
4033 	 * If we can't write to an interior vdev (mirror or RAID-Z),
4034 	 * set vdev_cant_write so that we stop trying to allocate from it.
4035 	 */
4036 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
4037 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
4038 		vdev_dbgmsg(vd, "zio_vdev_io_assess(zio=%px) setting "
4039 		    "cant_write=TRUE due to write failure with ENXIO",
4040 		    zio);
4041 		vd->vdev_cant_write = B_TRUE;
4042 	}
4043 
4044 	/*
4045 	 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
4046 	 * attempts will ever succeed. In this case we set a persistent
4047 	 * boolean flag so that we don't bother with it in the future.
4048 	 */
4049 	if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
4050 	    zio->io_type == ZIO_TYPE_IOCTL &&
4051 	    zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
4052 		vd->vdev_nowritecache = B_TRUE;
4053 
4054 	if (zio->io_error)
4055 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4056 
4057 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
4058 	    zio->io_physdone != NULL) {
4059 		ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
4060 		ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
4061 		zio->io_physdone(zio->io_logical);
4062 	}
4063 
4064 	return (zio);
4065 }
4066 
4067 void
4068 zio_vdev_io_reissue(zio_t *zio)
4069 {
4070 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4071 	ASSERT(zio->io_error == 0);
4072 
4073 	zio->io_stage >>= 1;
4074 }
4075 
4076 void
4077 zio_vdev_io_redone(zio_t *zio)
4078 {
4079 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
4080 
4081 	zio->io_stage >>= 1;
4082 }
4083 
4084 void
4085 zio_vdev_io_bypass(zio_t *zio)
4086 {
4087 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4088 	ASSERT(zio->io_error == 0);
4089 
4090 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
4091 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
4092 }
4093 
4094 /*
4095  * ==========================================================================
4096  * Encrypt and store encryption parameters
4097  * ==========================================================================
4098  */
4099 
4100 
4101 /*
4102  * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
4103  * managing the storage of encryption parameters and passing them to the
4104  * lower-level encryption functions.
4105  */
4106 static zio_t *
4107 zio_encrypt(zio_t *zio)
4108 {
4109 	zio_prop_t *zp = &zio->io_prop;
4110 	spa_t *spa = zio->io_spa;
4111 	blkptr_t *bp = zio->io_bp;
4112 	uint64_t psize = BP_GET_PSIZE(bp);
4113 	uint64_t dsobj = zio->io_bookmark.zb_objset;
4114 	dmu_object_type_t ot = BP_GET_TYPE(bp);
4115 	void *enc_buf = NULL;
4116 	abd_t *eabd = NULL;
4117 	uint8_t salt[ZIO_DATA_SALT_LEN];
4118 	uint8_t iv[ZIO_DATA_IV_LEN];
4119 	uint8_t mac[ZIO_DATA_MAC_LEN];
4120 	boolean_t no_crypt = B_FALSE;
4121 
4122 	/* the root zio already encrypted the data */
4123 	if (zio->io_child_type == ZIO_CHILD_GANG)
4124 		return (zio);
4125 
4126 	/* only ZIL blocks are re-encrypted on rewrite */
4127 	if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG)
4128 		return (zio);
4129 
4130 	if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) {
4131 		BP_SET_CRYPT(bp, B_FALSE);
4132 		return (zio);
4133 	}
4134 
4135 	/* if we are doing raw encryption set the provided encryption params */
4136 	if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) {
4137 		ASSERT0(BP_GET_LEVEL(bp));
4138 		BP_SET_CRYPT(bp, B_TRUE);
4139 		BP_SET_BYTEORDER(bp, zp->zp_byteorder);
4140 		if (ot != DMU_OT_OBJSET)
4141 			zio_crypt_encode_mac_bp(bp, zp->zp_mac);
4142 
4143 		/* dnode blocks must be written out in the provided byteorder */
4144 		if (zp->zp_byteorder != ZFS_HOST_BYTEORDER &&
4145 		    ot == DMU_OT_DNODE) {
4146 			void *bswap_buf = zio_buf_alloc(psize);
4147 			abd_t *babd = abd_get_from_buf(bswap_buf, psize);
4148 
4149 			ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4150 			abd_copy_to_buf(bswap_buf, zio->io_abd, psize);
4151 			dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf,
4152 			    psize);
4153 
4154 			abd_take_ownership_of_buf(babd, B_TRUE);
4155 			zio_push_transform(zio, babd, psize, psize, NULL);
4156 		}
4157 
4158 		if (DMU_OT_IS_ENCRYPTED(ot))
4159 			zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv);
4160 		return (zio);
4161 	}
4162 
4163 	/* indirect blocks only maintain a cksum of the lower level MACs */
4164 	if (BP_GET_LEVEL(bp) > 0) {
4165 		BP_SET_CRYPT(bp, B_TRUE);
4166 		VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE,
4167 		    zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp),
4168 		    mac));
4169 		zio_crypt_encode_mac_bp(bp, mac);
4170 		return (zio);
4171 	}
4172 
4173 	/*
4174 	 * Objset blocks are a special case since they have 2 256-bit MACs
4175 	 * embedded within them.
4176 	 */
4177 	if (ot == DMU_OT_OBJSET) {
4178 		ASSERT0(DMU_OT_IS_ENCRYPTED(ot));
4179 		ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4180 		BP_SET_CRYPT(bp, B_TRUE);
4181 		VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj,
4182 		    zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp)));
4183 		return (zio);
4184 	}
4185 
4186 	/* unencrypted object types are only authenticated with a MAC */
4187 	if (!DMU_OT_IS_ENCRYPTED(ot)) {
4188 		BP_SET_CRYPT(bp, B_TRUE);
4189 		VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj,
4190 		    zio->io_abd, psize, mac));
4191 		zio_crypt_encode_mac_bp(bp, mac);
4192 		return (zio);
4193 	}
4194 
4195 	/*
4196 	 * Later passes of sync-to-convergence may decide to rewrite data
4197 	 * in place to avoid more disk reallocations. This presents a problem
4198 	 * for encryption because this constitutes rewriting the new data with
4199 	 * the same encryption key and IV. However, this only applies to blocks
4200 	 * in the MOS (particularly the spacemaps) and we do not encrypt the
4201 	 * MOS. We assert that the zio is allocating or an intent log write
4202 	 * to enforce this.
4203 	 */
4204 	ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG);
4205 	ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG);
4206 	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION));
4207 	ASSERT3U(psize, !=, 0);
4208 
4209 	enc_buf = zio_buf_alloc(psize);
4210 	eabd = abd_get_from_buf(enc_buf, psize);
4211 	abd_take_ownership_of_buf(eabd, B_TRUE);
4212 
4213 	/*
4214 	 * For an explanation of what encryption parameters are stored
4215 	 * where, see the block comment in zio_crypt.c.
4216 	 */
4217 	if (ot == DMU_OT_INTENT_LOG) {
4218 		zio_crypt_decode_params_bp(bp, salt, iv);
4219 	} else {
4220 		BP_SET_CRYPT(bp, B_TRUE);
4221 	}
4222 
4223 	/* Perform the encryption. This should not fail */
4224 	VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark,
4225 	    BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
4226 	    salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt));
4227 
4228 	/* encode encryption metadata into the bp */
4229 	if (ot == DMU_OT_INTENT_LOG) {
4230 		/*
4231 		 * ZIL blocks store the MAC in the embedded checksum, so the
4232 		 * transform must always be applied.
4233 		 */
4234 		zio_crypt_encode_mac_zil(enc_buf, mac);
4235 		zio_push_transform(zio, eabd, psize, psize, NULL);
4236 	} else {
4237 		BP_SET_CRYPT(bp, B_TRUE);
4238 		zio_crypt_encode_params_bp(bp, salt, iv);
4239 		zio_crypt_encode_mac_bp(bp, mac);
4240 
4241 		if (no_crypt) {
4242 			ASSERT3U(ot, ==, DMU_OT_DNODE);
4243 			abd_free(eabd);
4244 		} else {
4245 			zio_push_transform(zio, eabd, psize, psize, NULL);
4246 		}
4247 	}
4248 
4249 	return (zio);
4250 }
4251 
4252 /*
4253  * ==========================================================================
4254  * Generate and verify checksums
4255  * ==========================================================================
4256  */
4257 static zio_t *
4258 zio_checksum_generate(zio_t *zio)
4259 {
4260 	blkptr_t *bp = zio->io_bp;
4261 	enum zio_checksum checksum;
4262 
4263 	if (bp == NULL) {
4264 		/*
4265 		 * This is zio_write_phys().
4266 		 * We're either generating a label checksum, or none at all.
4267 		 */
4268 		checksum = zio->io_prop.zp_checksum;
4269 
4270 		if (checksum == ZIO_CHECKSUM_OFF)
4271 			return (zio);
4272 
4273 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
4274 	} else {
4275 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
4276 			ASSERT(!IO_IS_ALLOCATING(zio));
4277 			checksum = ZIO_CHECKSUM_GANG_HEADER;
4278 		} else {
4279 			checksum = BP_GET_CHECKSUM(bp);
4280 		}
4281 	}
4282 
4283 	zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
4284 
4285 	return (zio);
4286 }
4287 
4288 static zio_t *
4289 zio_checksum_verify(zio_t *zio)
4290 {
4291 	zio_bad_cksum_t info;
4292 	blkptr_t *bp = zio->io_bp;
4293 	int error;
4294 
4295 	ASSERT(zio->io_vd != NULL);
4296 
4297 	if (bp == NULL) {
4298 		/*
4299 		 * This is zio_read_phys().
4300 		 * We're either verifying a label checksum, or nothing at all.
4301 		 */
4302 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
4303 			return (zio);
4304 
4305 		ASSERT3U(zio->io_prop.zp_checksum, ==, ZIO_CHECKSUM_LABEL);
4306 	}
4307 
4308 	if ((error = zio_checksum_error(zio, &info)) != 0) {
4309 		zio->io_error = error;
4310 		if (error == ECKSUM &&
4311 		    !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
4312 			(void) zfs_ereport_start_checksum(zio->io_spa,
4313 			    zio->io_vd, &zio->io_bookmark, zio,
4314 			    zio->io_offset, zio->io_size, &info);
4315 			mutex_enter(&zio->io_vd->vdev_stat_lock);
4316 			zio->io_vd->vdev_stat.vs_checksum_errors++;
4317 			mutex_exit(&zio->io_vd->vdev_stat_lock);
4318 		}
4319 	}
4320 
4321 	return (zio);
4322 }
4323 
4324 /*
4325  * Called by RAID-Z to ensure we don't compute the checksum twice.
4326  */
4327 void
4328 zio_checksum_verified(zio_t *zio)
4329 {
4330 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
4331 }
4332 
4333 /*
4334  * ==========================================================================
4335  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4336  * An error of 0 indicates success.  ENXIO indicates whole-device failure,
4337  * which may be transient (e.g. unplugged) or permanent.  ECKSUM and EIO
4338  * indicate errors that are specific to one I/O, and most likely permanent.
4339  * Any other error is presumed to be worse because we weren't expecting it.
4340  * ==========================================================================
4341  */
4342 int
4343 zio_worst_error(int e1, int e2)
4344 {
4345 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
4346 	int r1, r2;
4347 
4348 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
4349 		if (e1 == zio_error_rank[r1])
4350 			break;
4351 
4352 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
4353 		if (e2 == zio_error_rank[r2])
4354 			break;
4355 
4356 	return (r1 > r2 ? e1 : e2);
4357 }
4358 
4359 /*
4360  * ==========================================================================
4361  * I/O completion
4362  * ==========================================================================
4363  */
4364 static zio_t *
4365 zio_ready(zio_t *zio)
4366 {
4367 	blkptr_t *bp = zio->io_bp;
4368 	zio_t *pio, *pio_next;
4369 	zio_link_t *zl = NULL;
4370 
4371 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
4372 	    ZIO_WAIT_READY)) {
4373 		return (NULL);
4374 	}
4375 
4376 	if (zio->io_ready) {
4377 		ASSERT(IO_IS_ALLOCATING(zio));
4378 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
4379 		    (zio->io_flags & ZIO_FLAG_NOPWRITE));
4380 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
4381 
4382 		zio->io_ready(zio);
4383 	}
4384 
4385 	if (bp != NULL && bp != &zio->io_bp_copy)
4386 		zio->io_bp_copy = *bp;
4387 
4388 	if (zio->io_error != 0) {
4389 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4390 
4391 		if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4392 			ASSERT(IO_IS_ALLOCATING(zio));
4393 			ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4394 			ASSERT(zio->io_metaslab_class != NULL);
4395 
4396 			/*
4397 			 * We were unable to allocate anything, unreserve and
4398 			 * issue the next I/O to allocate.
4399 			 */
4400 			metaslab_class_throttle_unreserve(
4401 			    zio->io_metaslab_class, zio->io_prop.zp_copies,
4402 			    zio->io_allocator, zio);
4403 			zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
4404 		}
4405 	}
4406 
4407 	mutex_enter(&zio->io_lock);
4408 	zio->io_state[ZIO_WAIT_READY] = 1;
4409 	pio = zio_walk_parents(zio, &zl);
4410 	mutex_exit(&zio->io_lock);
4411 
4412 	/*
4413 	 * As we notify zio's parents, new parents could be added.
4414 	 * New parents go to the head of zio's io_parent_list, however,
4415 	 * so we will (correctly) not notify them.  The remainder of zio's
4416 	 * io_parent_list, from 'pio_next' onward, cannot change because
4417 	 * all parents must wait for us to be done before they can be done.
4418 	 */
4419 	for (; pio != NULL; pio = pio_next) {
4420 		pio_next = zio_walk_parents(zio, &zl);
4421 		zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL);
4422 	}
4423 
4424 	if (zio->io_flags & ZIO_FLAG_NODATA) {
4425 		if (BP_IS_GANG(bp)) {
4426 			zio->io_flags &= ~ZIO_FLAG_NODATA;
4427 		} else {
4428 			ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
4429 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
4430 		}
4431 	}
4432 
4433 	if (zio_injection_enabled &&
4434 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
4435 		zio_handle_ignored_writes(zio);
4436 
4437 	return (zio);
4438 }
4439 
4440 /*
4441  * Update the allocation throttle accounting.
4442  */
4443 static void
4444 zio_dva_throttle_done(zio_t *zio)
4445 {
4446 	zio_t *lio __maybe_unused = zio->io_logical;
4447 	zio_t *pio = zio_unique_parent(zio);
4448 	vdev_t *vd = zio->io_vd;
4449 	int flags = METASLAB_ASYNC_ALLOC;
4450 
4451 	ASSERT3P(zio->io_bp, !=, NULL);
4452 	ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
4453 	ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
4454 	ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
4455 	ASSERT(vd != NULL);
4456 	ASSERT3P(vd, ==, vd->vdev_top);
4457 	ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY));
4458 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4459 	ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
4460 	ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
4461 	ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
4462 
4463 	/*
4464 	 * Parents of gang children can have two flavors -- ones that
4465 	 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4466 	 * and ones that allocated the constituent blocks. The allocation
4467 	 * throttle needs to know the allocating parent zio so we must find
4468 	 * it here.
4469 	 */
4470 	if (pio->io_child_type == ZIO_CHILD_GANG) {
4471 		/*
4472 		 * If our parent is a rewrite gang child then our grandparent
4473 		 * would have been the one that performed the allocation.
4474 		 */
4475 		if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
4476 			pio = zio_unique_parent(pio);
4477 		flags |= METASLAB_GANG_CHILD;
4478 	}
4479 
4480 	ASSERT(IO_IS_ALLOCATING(pio));
4481 	ASSERT3P(zio, !=, zio->io_logical);
4482 	ASSERT(zio->io_logical != NULL);
4483 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4484 	ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
4485 	ASSERT(zio->io_metaslab_class != NULL);
4486 
4487 	mutex_enter(&pio->io_lock);
4488 	metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
4489 	    pio->io_allocator, B_TRUE);
4490 	mutex_exit(&pio->io_lock);
4491 
4492 	metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1,
4493 	    pio->io_allocator, pio);
4494 
4495 	/*
4496 	 * Call into the pipeline to see if there is more work that
4497 	 * needs to be done. If there is work to be done it will be
4498 	 * dispatched to another taskq thread.
4499 	 */
4500 	zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
4501 }
4502 
4503 static zio_t *
4504 zio_done(zio_t *zio)
4505 {
4506 	/*
4507 	 * Always attempt to keep stack usage minimal here since
4508 	 * we can be called recursively up to 19 levels deep.
4509 	 */
4510 	const uint64_t psize = zio->io_size;
4511 	zio_t *pio, *pio_next;
4512 	zio_link_t *zl = NULL;
4513 
4514 	/*
4515 	 * If our children haven't all completed,
4516 	 * wait for them and then repeat this pipeline stage.
4517 	 */
4518 	if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
4519 		return (NULL);
4520 	}
4521 
4522 	/*
4523 	 * If the allocation throttle is enabled, then update the accounting.
4524 	 * We only track child I/Os that are part of an allocating async
4525 	 * write. We must do this since the allocation is performed
4526 	 * by the logical I/O but the actual write is done by child I/Os.
4527 	 */
4528 	if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
4529 	    zio->io_child_type == ZIO_CHILD_VDEV) {
4530 		ASSERT(zio->io_metaslab_class != NULL);
4531 		ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled);
4532 		zio_dva_throttle_done(zio);
4533 	}
4534 
4535 	/*
4536 	 * If the allocation throttle is enabled, verify that
4537 	 * we have decremented the refcounts for every I/O that was throttled.
4538 	 */
4539 	if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4540 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4541 		ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4542 		ASSERT(zio->io_bp != NULL);
4543 
4544 		metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio,
4545 		    zio->io_allocator);
4546 		VERIFY(zfs_refcount_not_held(&zio->io_metaslab_class->
4547 		    mc_allocator[zio->io_allocator].mca_alloc_slots, zio));
4548 	}
4549 
4550 
4551 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
4552 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
4553 			ASSERT(zio->io_children[c][w] == 0);
4554 
4555 	if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) {
4556 		ASSERT(zio->io_bp->blk_pad[0] == 0);
4557 		ASSERT(zio->io_bp->blk_pad[1] == 0);
4558 		ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy,
4559 		    sizeof (blkptr_t)) == 0 ||
4560 		    (zio->io_bp == zio_unique_parent(zio)->io_bp));
4561 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
4562 		    zio->io_bp_override == NULL &&
4563 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
4564 			ASSERT3U(zio->io_prop.zp_copies, <=,
4565 			    BP_GET_NDVAS(zio->io_bp));
4566 			ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
4567 			    (BP_COUNT_GANG(zio->io_bp) ==
4568 			    BP_GET_NDVAS(zio->io_bp)));
4569 		}
4570 		if (zio->io_flags & ZIO_FLAG_NOPWRITE)
4571 			VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig));
4572 	}
4573 
4574 	/*
4575 	 * If there were child vdev/gang/ddt errors, they apply to us now.
4576 	 */
4577 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
4578 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
4579 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
4580 
4581 	/*
4582 	 * If the I/O on the transformed data was successful, generate any
4583 	 * checksum reports now while we still have the transformed data.
4584 	 */
4585 	if (zio->io_error == 0) {
4586 		while (zio->io_cksum_report != NULL) {
4587 			zio_cksum_report_t *zcr = zio->io_cksum_report;
4588 			uint64_t align = zcr->zcr_align;
4589 			uint64_t asize = P2ROUNDUP(psize, align);
4590 			abd_t *adata = zio->io_abd;
4591 
4592 			if (adata != NULL && asize != psize) {
4593 				adata = abd_alloc(asize, B_TRUE);
4594 				abd_copy(adata, zio->io_abd, psize);
4595 				abd_zero_off(adata, psize, asize - psize);
4596 			}
4597 
4598 			zio->io_cksum_report = zcr->zcr_next;
4599 			zcr->zcr_next = NULL;
4600 			zcr->zcr_finish(zcr, adata);
4601 			zfs_ereport_free_checksum(zcr);
4602 
4603 			if (adata != NULL && asize != psize)
4604 				abd_free(adata);
4605 		}
4606 	}
4607 
4608 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
4609 
4610 	vdev_stat_update(zio, psize);
4611 
4612 	/*
4613 	 * If this I/O is attached to a particular vdev is slow, exceeding
4614 	 * 30 seconds to complete, post an error described the I/O delay.
4615 	 * We ignore these errors if the device is currently unavailable.
4616 	 */
4617 	if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) {
4618 		if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) {
4619 			/*
4620 			 * We want to only increment our slow IO counters if
4621 			 * the IO is valid (i.e. not if the drive is removed).
4622 			 *
4623 			 * zfs_ereport_post() will also do these checks, but
4624 			 * it can also ratelimit and have other failures, so we
4625 			 * need to increment the slow_io counters independent
4626 			 * of it.
4627 			 */
4628 			if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY,
4629 			    zio->io_spa, zio->io_vd, zio)) {
4630 				mutex_enter(&zio->io_vd->vdev_stat_lock);
4631 				zio->io_vd->vdev_stat.vs_slow_ios++;
4632 				mutex_exit(&zio->io_vd->vdev_stat_lock);
4633 
4634 				(void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY,
4635 				    zio->io_spa, zio->io_vd, &zio->io_bookmark,
4636 				    zio, 0);
4637 			}
4638 		}
4639 	}
4640 
4641 	if (zio->io_error) {
4642 		/*
4643 		 * If this I/O is attached to a particular vdev,
4644 		 * generate an error message describing the I/O failure
4645 		 * at the block level.  We ignore these errors if the
4646 		 * device is currently unavailable.
4647 		 */
4648 		if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
4649 		    !vdev_is_dead(zio->io_vd)) {
4650 			int ret = zfs_ereport_post(FM_EREPORT_ZFS_IO,
4651 			    zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0);
4652 			if (ret != EALREADY) {
4653 				mutex_enter(&zio->io_vd->vdev_stat_lock);
4654 				if (zio->io_type == ZIO_TYPE_READ)
4655 					zio->io_vd->vdev_stat.vs_read_errors++;
4656 				else if (zio->io_type == ZIO_TYPE_WRITE)
4657 					zio->io_vd->vdev_stat.vs_write_errors++;
4658 				mutex_exit(&zio->io_vd->vdev_stat_lock);
4659 			}
4660 		}
4661 
4662 		if ((zio->io_error == EIO || !(zio->io_flags &
4663 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
4664 		    zio == zio->io_logical) {
4665 			/*
4666 			 * For logical I/O requests, tell the SPA to log the
4667 			 * error and generate a logical data ereport.
4668 			 */
4669 			spa_log_error(zio->io_spa, &zio->io_bookmark);
4670 			(void) zfs_ereport_post(FM_EREPORT_ZFS_DATA,
4671 			    zio->io_spa, NULL, &zio->io_bookmark, zio, 0);
4672 		}
4673 	}
4674 
4675 	if (zio->io_error && zio == zio->io_logical) {
4676 		/*
4677 		 * Determine whether zio should be reexecuted.  This will
4678 		 * propagate all the way to the root via zio_notify_parent().
4679 		 */
4680 		ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
4681 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4682 
4683 		if (IO_IS_ALLOCATING(zio) &&
4684 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
4685 			if (zio->io_error != ENOSPC)
4686 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
4687 			else
4688 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4689 		}
4690 
4691 		if ((zio->io_type == ZIO_TYPE_READ ||
4692 		    zio->io_type == ZIO_TYPE_FREE) &&
4693 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
4694 		    zio->io_error == ENXIO &&
4695 		    spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
4696 		    spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
4697 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4698 
4699 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
4700 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4701 
4702 		/*
4703 		 * Here is a possibly good place to attempt to do
4704 		 * either combinatorial reconstruction or error correction
4705 		 * based on checksums.  It also might be a good place
4706 		 * to send out preliminary ereports before we suspend
4707 		 * processing.
4708 		 */
4709 	}
4710 
4711 	/*
4712 	 * If there were logical child errors, they apply to us now.
4713 	 * We defer this until now to avoid conflating logical child
4714 	 * errors with errors that happened to the zio itself when
4715 	 * updating vdev stats and reporting FMA events above.
4716 	 */
4717 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
4718 
4719 	if ((zio->io_error || zio->io_reexecute) &&
4720 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
4721 	    !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
4722 		zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
4723 
4724 	zio_gang_tree_free(&zio->io_gang_tree);
4725 
4726 	/*
4727 	 * Godfather I/Os should never suspend.
4728 	 */
4729 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4730 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4731 		zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND;
4732 
4733 	if (zio->io_reexecute) {
4734 		/*
4735 		 * This is a logical I/O that wants to reexecute.
4736 		 *
4737 		 * Reexecute is top-down.  When an i/o fails, if it's not
4738 		 * the root, it simply notifies its parent and sticks around.
4739 		 * The parent, seeing that it still has children in zio_done(),
4740 		 * does the same.  This percolates all the way up to the root.
4741 		 * The root i/o will reexecute or suspend the entire tree.
4742 		 *
4743 		 * This approach ensures that zio_reexecute() honors
4744 		 * all the original i/o dependency relationships, e.g.
4745 		 * parents not executing until children are ready.
4746 		 */
4747 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4748 
4749 		zio->io_gang_leader = NULL;
4750 
4751 		mutex_enter(&zio->io_lock);
4752 		zio->io_state[ZIO_WAIT_DONE] = 1;
4753 		mutex_exit(&zio->io_lock);
4754 
4755 		/*
4756 		 * "The Godfather" I/O monitors its children but is
4757 		 * not a true parent to them. It will track them through
4758 		 * the pipeline but severs its ties whenever they get into
4759 		 * trouble (e.g. suspended). This allows "The Godfather"
4760 		 * I/O to return status without blocking.
4761 		 */
4762 		zl = NULL;
4763 		for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4764 		    pio = pio_next) {
4765 			zio_link_t *remove_zl = zl;
4766 			pio_next = zio_walk_parents(zio, &zl);
4767 
4768 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4769 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4770 				zio_remove_child(pio, zio, remove_zl);
4771 				/*
4772 				 * This is a rare code path, so we don't
4773 				 * bother with "next_to_execute".
4774 				 */
4775 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE,
4776 				    NULL);
4777 			}
4778 		}
4779 
4780 		if ((pio = zio_unique_parent(zio)) != NULL) {
4781 			/*
4782 			 * We're not a root i/o, so there's nothing to do
4783 			 * but notify our parent.  Don't propagate errors
4784 			 * upward since we haven't permanently failed yet.
4785 			 */
4786 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4787 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4788 			/*
4789 			 * This is a rare code path, so we don't bother with
4790 			 * "next_to_execute".
4791 			 */
4792 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL);
4793 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4794 			/*
4795 			 * We'd fail again if we reexecuted now, so suspend
4796 			 * until conditions improve (e.g. device comes online).
4797 			 */
4798 			zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
4799 		} else {
4800 			/*
4801 			 * Reexecution is potentially a huge amount of work.
4802 			 * Hand it off to the otherwise-unused claim taskq.
4803 			 */
4804 			ASSERT(taskq_empty_ent(&zio->io_tqent));
4805 			spa_taskq_dispatch_ent(zio->io_spa,
4806 			    ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
4807 			    zio_reexecute, zio, 0, &zio->io_tqent);
4808 		}
4809 		return (NULL);
4810 	}
4811 
4812 	ASSERT(zio->io_child_count == 0);
4813 	ASSERT(zio->io_reexecute == 0);
4814 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4815 
4816 	/*
4817 	 * Report any checksum errors, since the I/O is complete.
4818 	 */
4819 	while (zio->io_cksum_report != NULL) {
4820 		zio_cksum_report_t *zcr = zio->io_cksum_report;
4821 		zio->io_cksum_report = zcr->zcr_next;
4822 		zcr->zcr_next = NULL;
4823 		zcr->zcr_finish(zcr, NULL);
4824 		zfs_ereport_free_checksum(zcr);
4825 	}
4826 
4827 	if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
4828 	    !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) &&
4829 	    !(zio->io_flags & ZIO_FLAG_NOPWRITE)) {
4830 		metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
4831 	}
4832 
4833 	/*
4834 	 * It is the responsibility of the done callback to ensure that this
4835 	 * particular zio is no longer discoverable for adoption, and as
4836 	 * such, cannot acquire any new parents.
4837 	 */
4838 	if (zio->io_done)
4839 		zio->io_done(zio);
4840 
4841 	mutex_enter(&zio->io_lock);
4842 	zio->io_state[ZIO_WAIT_DONE] = 1;
4843 	mutex_exit(&zio->io_lock);
4844 
4845 	/*
4846 	 * We are done executing this zio.  We may want to execute a parent
4847 	 * next.  See the comment in zio_notify_parent().
4848 	 */
4849 	zio_t *next_to_execute = NULL;
4850 	zl = NULL;
4851 	for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4852 		zio_link_t *remove_zl = zl;
4853 		pio_next = zio_walk_parents(zio, &zl);
4854 		zio_remove_child(pio, zio, remove_zl);
4855 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute);
4856 	}
4857 
4858 	if (zio->io_waiter != NULL) {
4859 		mutex_enter(&zio->io_lock);
4860 		zio->io_executor = NULL;
4861 		cv_broadcast(&zio->io_cv);
4862 		mutex_exit(&zio->io_lock);
4863 	} else {
4864 		zio_destroy(zio);
4865 	}
4866 
4867 	return (next_to_execute);
4868 }
4869 
4870 /*
4871  * ==========================================================================
4872  * I/O pipeline definition
4873  * ==========================================================================
4874  */
4875 static zio_pipe_stage_t *zio_pipeline[] = {
4876 	NULL,
4877 	zio_read_bp_init,
4878 	zio_write_bp_init,
4879 	zio_free_bp_init,
4880 	zio_issue_async,
4881 	zio_write_compress,
4882 	zio_encrypt,
4883 	zio_checksum_generate,
4884 	zio_nop_write,
4885 	zio_ddt_read_start,
4886 	zio_ddt_read_done,
4887 	zio_ddt_write,
4888 	zio_ddt_free,
4889 	zio_gang_assemble,
4890 	zio_gang_issue,
4891 	zio_dva_throttle,
4892 	zio_dva_allocate,
4893 	zio_dva_free,
4894 	zio_dva_claim,
4895 	zio_ready,
4896 	zio_vdev_io_start,
4897 	zio_vdev_io_done,
4898 	zio_vdev_io_assess,
4899 	zio_checksum_verify,
4900 	zio_done
4901 };
4902 
4903 
4904 
4905 
4906 /*
4907  * Compare two zbookmark_phys_t's to see which we would reach first in a
4908  * pre-order traversal of the object tree.
4909  *
4910  * This is simple in every case aside from the meta-dnode object. For all other
4911  * objects, we traverse them in order (object 1 before object 2, and so on).
4912  * However, all of these objects are traversed while traversing object 0, since
4913  * the data it points to is the list of objects.  Thus, we need to convert to a
4914  * canonical representation so we can compare meta-dnode bookmarks to
4915  * non-meta-dnode bookmarks.
4916  *
4917  * We do this by calculating "equivalents" for each field of the zbookmark.
4918  * zbookmarks outside of the meta-dnode use their own object and level, and
4919  * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4920  * blocks this bookmark refers to) by multiplying their blkid by their span
4921  * (the number of L0 blocks contained within one block at their level).
4922  * zbookmarks inside the meta-dnode calculate their object equivalent
4923  * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4924  * level + 1<<31 (any value larger than a level could ever be) for their level.
4925  * This causes them to always compare before a bookmark in their object
4926  * equivalent, compare appropriately to bookmarks in other objects, and to
4927  * compare appropriately to other bookmarks in the meta-dnode.
4928  */
4929 int
4930 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4931     const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4932 {
4933 	/*
4934 	 * These variables represent the "equivalent" values for the zbookmark,
4935 	 * after converting zbookmarks inside the meta dnode to their
4936 	 * normal-object equivalents.
4937 	 */
4938 	uint64_t zb1obj, zb2obj;
4939 	uint64_t zb1L0, zb2L0;
4940 	uint64_t zb1level, zb2level;
4941 
4942 	if (zb1->zb_object == zb2->zb_object &&
4943 	    zb1->zb_level == zb2->zb_level &&
4944 	    zb1->zb_blkid == zb2->zb_blkid)
4945 		return (0);
4946 
4947 	IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT);
4948 	IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT);
4949 
4950 	/*
4951 	 * BP_SPANB calculates the span in blocks.
4952 	 */
4953 	zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4954 	zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4955 
4956 	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4957 		zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4958 		zb1L0 = 0;
4959 		zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4960 	} else {
4961 		zb1obj = zb1->zb_object;
4962 		zb1level = zb1->zb_level;
4963 	}
4964 
4965 	if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4966 		zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4967 		zb2L0 = 0;
4968 		zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4969 	} else {
4970 		zb2obj = zb2->zb_object;
4971 		zb2level = zb2->zb_level;
4972 	}
4973 
4974 	/* Now that we have a canonical representation, do the comparison. */
4975 	if (zb1obj != zb2obj)
4976 		return (zb1obj < zb2obj ? -1 : 1);
4977 	else if (zb1L0 != zb2L0)
4978 		return (zb1L0 < zb2L0 ? -1 : 1);
4979 	else if (zb1level != zb2level)
4980 		return (zb1level > zb2level ? -1 : 1);
4981 	/*
4982 	 * This can (theoretically) happen if the bookmarks have the same object
4983 	 * and level, but different blkids, if the block sizes are not the same.
4984 	 * There is presently no way to change the indirect block sizes
4985 	 */
4986 	return (0);
4987 }
4988 
4989 /*
4990  *  This function checks the following: given that last_block is the place that
4991  *  our traversal stopped last time, does that guarantee that we've visited
4992  *  every node under subtree_root?  Therefore, we can't just use the raw output
4993  *  of zbookmark_compare.  We have to pass in a modified version of
4994  *  subtree_root; by incrementing the block id, and then checking whether
4995  *  last_block is before or equal to that, we can tell whether or not having
4996  *  visited last_block implies that all of subtree_root's children have been
4997  *  visited.
4998  */
4999 boolean_t
5000 zbookmark_subtree_completed(const dnode_phys_t *dnp,
5001     const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
5002 {
5003 	zbookmark_phys_t mod_zb = *subtree_root;
5004 	mod_zb.zb_blkid++;
5005 	ASSERT(last_block->zb_level == 0);
5006 
5007 	/* The objset_phys_t isn't before anything. */
5008 	if (dnp == NULL)
5009 		return (B_FALSE);
5010 
5011 	/*
5012 	 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
5013 	 * data block size in sectors, because that variable is only used if
5014 	 * the bookmark refers to a block in the meta-dnode.  Since we don't
5015 	 * know without examining it what object it refers to, and there's no
5016 	 * harm in passing in this value in other cases, we always pass it in.
5017 	 *
5018 	 * We pass in 0 for the indirect block size shift because zb2 must be
5019 	 * level 0.  The indirect block size is only used to calculate the span
5020 	 * of the bookmark, but since the bookmark must be level 0, the span is
5021 	 * always 1, so the math works out.
5022 	 *
5023 	 * If you make changes to how the zbookmark_compare code works, be sure
5024 	 * to make sure that this code still works afterwards.
5025 	 */
5026 	return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
5027 	    1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
5028 	    last_block) <= 0);
5029 }
5030 
5031 EXPORT_SYMBOL(zio_type_name);
5032 EXPORT_SYMBOL(zio_buf_alloc);
5033 EXPORT_SYMBOL(zio_data_buf_alloc);
5034 EXPORT_SYMBOL(zio_buf_free);
5035 EXPORT_SYMBOL(zio_data_buf_free);
5036 
5037 /* BEGIN CSTYLED */
5038 ZFS_MODULE_PARAM(zfs_zio, zio_, slow_io_ms, INT, ZMOD_RW,
5039 	"Max I/O completion time (milliseconds) before marking it as slow");
5040 
5041 ZFS_MODULE_PARAM(zfs_zio, zio_, requeue_io_start_cut_in_line, INT, ZMOD_RW,
5042 	"Prioritize requeued I/O");
5043 
5044 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_deferred_free,  INT, ZMOD_RW,
5045 	"Defer frees starting in this pass");
5046 
5047 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_dont_compress, INT, ZMOD_RW,
5048 	"Don't compress starting in this pass");
5049 
5050 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_rewrite, INT, ZMOD_RW,
5051 	"Rewrite new bps starting in this pass");
5052 
5053 ZFS_MODULE_PARAM(zfs_zio, zio_, dva_throttle_enabled, INT, ZMOD_RW,
5054 	"Throttle block allocations in the ZIO pipeline");
5055 
5056 ZFS_MODULE_PARAM(zfs_zio, zio_, deadman_log_all, INT, ZMOD_RW,
5057 	"Log all slow ZIOs, not just those with vdevs");
5058 /* END CSTYLED */
5059