xref: /illumos-gate/usr/src/uts/common/fs/zfs/zio.c (revision 5d7b4d438c4a51eccc95e77a83a437b4d48380eb)
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, 2014 by Delphix. All rights reserved.
24  * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25  */
26 
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
29 #include <sys/spa.h>
30 #include <sys/txg.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio_impl.h>
34 #include <sys/zio_compress.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/arc.h>
38 #include <sys/ddt.h>
39 #include <sys/blkptr.h>
40 #include <sys/zfeature.h>
41 
42 /*
43  * ==========================================================================
44  * I/O type descriptions
45  * ==========================================================================
46  */
47 const char *zio_type_name[ZIO_TYPES] = {
48 	"zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
49 	"zio_ioctl"
50 };
51 
52 /*
53  * ==========================================================================
54  * I/O kmem caches
55  * ==========================================================================
56  */
57 kmem_cache_t *zio_cache;
58 kmem_cache_t *zio_link_cache;
59 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
60 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
61 
62 #ifdef _KERNEL
63 extern vmem_t *zio_alloc_arena;
64 #endif
65 
66 /*
67  * The following actions directly effect the spa's sync-to-convergence logic.
68  * The values below define the sync pass when we start performing the action.
69  * Care should be taken when changing these values as they directly impact
70  * spa_sync() performance. Tuning these values may introduce subtle performance
71  * pathologies and should only be done in the context of performance analysis.
72  * These tunables will eventually be removed and replaced with #defines once
73  * enough analysis has been done to determine optimal values.
74  *
75  * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
76  * regular blocks are not deferred.
77  */
78 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
79 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
80 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
81 
82 /*
83  * An allocating zio is one that either currently has the DVA allocate
84  * stage set or will have it later in its lifetime.
85  */
86 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
87 
88 boolean_t	zio_requeue_io_start_cut_in_line = B_TRUE;
89 
90 #ifdef ZFS_DEBUG
91 int zio_buf_debug_limit = 16384;
92 #else
93 int zio_buf_debug_limit = 0;
94 #endif
95 
96 void
97 zio_init(void)
98 {
99 	size_t c;
100 	vmem_t *data_alloc_arena = NULL;
101 
102 #ifdef _KERNEL
103 	data_alloc_arena = zio_alloc_arena;
104 #endif
105 	zio_cache = kmem_cache_create("zio_cache",
106 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
107 	zio_link_cache = kmem_cache_create("zio_link_cache",
108 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
109 
110 	/*
111 	 * For small buffers, we want a cache for each multiple of
112 	 * SPA_MINBLOCKSIZE.  For medium-size buffers, we want a cache
113 	 * for each quarter-power of 2.  For large buffers, we want
114 	 * a cache for each multiple of PAGESIZE.
115 	 */
116 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
117 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
118 		size_t p2 = size;
119 		size_t align = 0;
120 		size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
121 
122 		while (p2 & (p2 - 1))
123 			p2 &= p2 - 1;
124 
125 #ifndef _KERNEL
126 		/*
127 		 * If we are using watchpoints, put each buffer on its own page,
128 		 * to eliminate the performance overhead of trapping to the
129 		 * kernel when modifying a non-watched buffer that shares the
130 		 * page with a watched buffer.
131 		 */
132 		if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
133 			continue;
134 #endif
135 		if (size <= 4 * SPA_MINBLOCKSIZE) {
136 			align = SPA_MINBLOCKSIZE;
137 		} else if (IS_P2ALIGNED(size, PAGESIZE)) {
138 			align = PAGESIZE;
139 		} else if (IS_P2ALIGNED(size, p2 >> 2)) {
140 			align = p2 >> 2;
141 		}
142 
143 		if (align != 0) {
144 			char name[36];
145 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
146 			zio_buf_cache[c] = kmem_cache_create(name, size,
147 			    align, NULL, NULL, NULL, NULL, NULL, cflags);
148 
149 			/*
150 			 * Since zio_data bufs do not appear in crash dumps, we
151 			 * pass KMC_NOTOUCH so that no allocator metadata is
152 			 * stored with the buffers.
153 			 */
154 			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
155 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
156 			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
157 			    cflags | KMC_NOTOUCH);
158 		}
159 	}
160 
161 	while (--c != 0) {
162 		ASSERT(zio_buf_cache[c] != NULL);
163 		if (zio_buf_cache[c - 1] == NULL)
164 			zio_buf_cache[c - 1] = zio_buf_cache[c];
165 
166 		ASSERT(zio_data_buf_cache[c] != NULL);
167 		if (zio_data_buf_cache[c - 1] == NULL)
168 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
169 	}
170 
171 	zio_inject_init();
172 }
173 
174 void
175 zio_fini(void)
176 {
177 	size_t c;
178 	kmem_cache_t *last_cache = NULL;
179 	kmem_cache_t *last_data_cache = NULL;
180 
181 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
182 		if (zio_buf_cache[c] != last_cache) {
183 			last_cache = zio_buf_cache[c];
184 			kmem_cache_destroy(zio_buf_cache[c]);
185 		}
186 		zio_buf_cache[c] = NULL;
187 
188 		if (zio_data_buf_cache[c] != last_data_cache) {
189 			last_data_cache = zio_data_buf_cache[c];
190 			kmem_cache_destroy(zio_data_buf_cache[c]);
191 		}
192 		zio_data_buf_cache[c] = NULL;
193 	}
194 
195 	kmem_cache_destroy(zio_link_cache);
196 	kmem_cache_destroy(zio_cache);
197 
198 	zio_inject_fini();
199 }
200 
201 /*
202  * ==========================================================================
203  * Allocate and free I/O buffers
204  * ==========================================================================
205  */
206 
207 /*
208  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
209  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
210  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
211  * excess / transient data in-core during a crashdump.
212  */
213 void *
214 zio_buf_alloc(size_t size)
215 {
216 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
217 
218 	ASSERT3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
219 
220 	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
221 }
222 
223 /*
224  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
225  * crashdump if the kernel panics.  This exists so that we will limit the amount
226  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
227  * of kernel heap dumped to disk when the kernel panics)
228  */
229 void *
230 zio_data_buf_alloc(size_t size)
231 {
232 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
233 
234 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
235 
236 	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
237 }
238 
239 void
240 zio_buf_free(void *buf, size_t size)
241 {
242 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
243 
244 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
245 
246 	kmem_cache_free(zio_buf_cache[c], buf);
247 }
248 
249 void
250 zio_data_buf_free(void *buf, size_t size)
251 {
252 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
253 
254 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
255 
256 	kmem_cache_free(zio_data_buf_cache[c], buf);
257 }
258 
259 /*
260  * ==========================================================================
261  * Push and pop I/O transform buffers
262  * ==========================================================================
263  */
264 static void
265 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
266 	zio_transform_func_t *transform)
267 {
268 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
269 
270 	zt->zt_orig_data = zio->io_data;
271 	zt->zt_orig_size = zio->io_size;
272 	zt->zt_bufsize = bufsize;
273 	zt->zt_transform = transform;
274 
275 	zt->zt_next = zio->io_transform_stack;
276 	zio->io_transform_stack = zt;
277 
278 	zio->io_data = data;
279 	zio->io_size = size;
280 }
281 
282 static void
283 zio_pop_transforms(zio_t *zio)
284 {
285 	zio_transform_t *zt;
286 
287 	while ((zt = zio->io_transform_stack) != NULL) {
288 		if (zt->zt_transform != NULL)
289 			zt->zt_transform(zio,
290 			    zt->zt_orig_data, zt->zt_orig_size);
291 
292 		if (zt->zt_bufsize != 0)
293 			zio_buf_free(zio->io_data, zt->zt_bufsize);
294 
295 		zio->io_data = zt->zt_orig_data;
296 		zio->io_size = zt->zt_orig_size;
297 		zio->io_transform_stack = zt->zt_next;
298 
299 		kmem_free(zt, sizeof (zio_transform_t));
300 	}
301 }
302 
303 /*
304  * ==========================================================================
305  * I/O transform callbacks for subblocks and decompression
306  * ==========================================================================
307  */
308 static void
309 zio_subblock(zio_t *zio, void *data, uint64_t size)
310 {
311 	ASSERT(zio->io_size > size);
312 
313 	if (zio->io_type == ZIO_TYPE_READ)
314 		bcopy(zio->io_data, data, size);
315 }
316 
317 static void
318 zio_decompress(zio_t *zio, void *data, uint64_t size)
319 {
320 	if (zio->io_error == 0 &&
321 	    zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
322 	    zio->io_data, data, zio->io_size, size) != 0)
323 		zio->io_error = SET_ERROR(EIO);
324 }
325 
326 /*
327  * ==========================================================================
328  * I/O parent/child relationships and pipeline interlocks
329  * ==========================================================================
330  */
331 /*
332  * NOTE - Callers to zio_walk_parents() and zio_walk_children must
333  *        continue calling these functions until they return NULL.
334  *        Otherwise, the next caller will pick up the list walk in
335  *        some indeterminate state.  (Otherwise every caller would
336  *        have to pass in a cookie to keep the state represented by
337  *        io_walk_link, which gets annoying.)
338  */
339 zio_t *
340 zio_walk_parents(zio_t *cio)
341 {
342 	zio_link_t *zl = cio->io_walk_link;
343 	list_t *pl = &cio->io_parent_list;
344 
345 	zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
346 	cio->io_walk_link = zl;
347 
348 	if (zl == NULL)
349 		return (NULL);
350 
351 	ASSERT(zl->zl_child == cio);
352 	return (zl->zl_parent);
353 }
354 
355 zio_t *
356 zio_walk_children(zio_t *pio)
357 {
358 	zio_link_t *zl = pio->io_walk_link;
359 	list_t *cl = &pio->io_child_list;
360 
361 	zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
362 	pio->io_walk_link = zl;
363 
364 	if (zl == NULL)
365 		return (NULL);
366 
367 	ASSERT(zl->zl_parent == pio);
368 	return (zl->zl_child);
369 }
370 
371 zio_t *
372 zio_unique_parent(zio_t *cio)
373 {
374 	zio_t *pio = zio_walk_parents(cio);
375 
376 	VERIFY(zio_walk_parents(cio) == NULL);
377 	return (pio);
378 }
379 
380 void
381 zio_add_child(zio_t *pio, zio_t *cio)
382 {
383 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
384 
385 	/*
386 	 * Logical I/Os can have logical, gang, or vdev children.
387 	 * Gang I/Os can have gang or vdev children.
388 	 * Vdev I/Os can only have vdev children.
389 	 * The following ASSERT captures all of these constraints.
390 	 */
391 	ASSERT(cio->io_child_type <= pio->io_child_type);
392 
393 	zl->zl_parent = pio;
394 	zl->zl_child = cio;
395 
396 	mutex_enter(&cio->io_lock);
397 	mutex_enter(&pio->io_lock);
398 
399 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
400 
401 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
402 		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
403 
404 	list_insert_head(&pio->io_child_list, zl);
405 	list_insert_head(&cio->io_parent_list, zl);
406 
407 	pio->io_child_count++;
408 	cio->io_parent_count++;
409 
410 	mutex_exit(&pio->io_lock);
411 	mutex_exit(&cio->io_lock);
412 }
413 
414 static void
415 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
416 {
417 	ASSERT(zl->zl_parent == pio);
418 	ASSERT(zl->zl_child == cio);
419 
420 	mutex_enter(&cio->io_lock);
421 	mutex_enter(&pio->io_lock);
422 
423 	list_remove(&pio->io_child_list, zl);
424 	list_remove(&cio->io_parent_list, zl);
425 
426 	pio->io_child_count--;
427 	cio->io_parent_count--;
428 
429 	mutex_exit(&pio->io_lock);
430 	mutex_exit(&cio->io_lock);
431 
432 	kmem_cache_free(zio_link_cache, zl);
433 }
434 
435 static boolean_t
436 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
437 {
438 	uint64_t *countp = &zio->io_children[child][wait];
439 	boolean_t waiting = B_FALSE;
440 
441 	mutex_enter(&zio->io_lock);
442 	ASSERT(zio->io_stall == NULL);
443 	if (*countp != 0) {
444 		zio->io_stage >>= 1;
445 		zio->io_stall = countp;
446 		waiting = B_TRUE;
447 	}
448 	mutex_exit(&zio->io_lock);
449 
450 	return (waiting);
451 }
452 
453 static void
454 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
455 {
456 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
457 	int *errorp = &pio->io_child_error[zio->io_child_type];
458 
459 	mutex_enter(&pio->io_lock);
460 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
461 		*errorp = zio_worst_error(*errorp, zio->io_error);
462 	pio->io_reexecute |= zio->io_reexecute;
463 	ASSERT3U(*countp, >, 0);
464 
465 	(*countp)--;
466 
467 	if (*countp == 0 && pio->io_stall == countp) {
468 		pio->io_stall = NULL;
469 		mutex_exit(&pio->io_lock);
470 		zio_execute(pio);
471 	} else {
472 		mutex_exit(&pio->io_lock);
473 	}
474 }
475 
476 static void
477 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
478 {
479 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
480 		zio->io_error = zio->io_child_error[c];
481 }
482 
483 /*
484  * ==========================================================================
485  * Create the various types of I/O (read, write, free, etc)
486  * ==========================================================================
487  */
488 static zio_t *
489 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
490     void *data, uint64_t size, zio_done_func_t *done, void *private,
491     zio_type_t type, zio_priority_t priority, enum zio_flag flags,
492     vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
493     enum zio_stage stage, enum zio_stage pipeline)
494 {
495 	zio_t *zio;
496 
497 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
498 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
499 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
500 
501 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
502 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
503 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
504 
505 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
506 	bzero(zio, sizeof (zio_t));
507 
508 	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
509 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
510 
511 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
512 	    offsetof(zio_link_t, zl_parent_node));
513 	list_create(&zio->io_child_list, sizeof (zio_link_t),
514 	    offsetof(zio_link_t, zl_child_node));
515 
516 	if (vd != NULL)
517 		zio->io_child_type = ZIO_CHILD_VDEV;
518 	else if (flags & ZIO_FLAG_GANG_CHILD)
519 		zio->io_child_type = ZIO_CHILD_GANG;
520 	else if (flags & ZIO_FLAG_DDT_CHILD)
521 		zio->io_child_type = ZIO_CHILD_DDT;
522 	else
523 		zio->io_child_type = ZIO_CHILD_LOGICAL;
524 
525 	if (bp != NULL) {
526 		zio->io_bp = (blkptr_t *)bp;
527 		zio->io_bp_copy = *bp;
528 		zio->io_bp_orig = *bp;
529 		if (type != ZIO_TYPE_WRITE ||
530 		    zio->io_child_type == ZIO_CHILD_DDT)
531 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
532 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
533 			zio->io_logical = zio;
534 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
535 			pipeline |= ZIO_GANG_STAGES;
536 	}
537 
538 	zio->io_spa = spa;
539 	zio->io_txg = txg;
540 	zio->io_done = done;
541 	zio->io_private = private;
542 	zio->io_type = type;
543 	zio->io_priority = priority;
544 	zio->io_vd = vd;
545 	zio->io_offset = offset;
546 	zio->io_orig_data = zio->io_data = data;
547 	zio->io_orig_size = zio->io_size = size;
548 	zio->io_orig_flags = zio->io_flags = flags;
549 	zio->io_orig_stage = zio->io_stage = stage;
550 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
551 
552 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
553 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
554 
555 	if (zb != NULL)
556 		zio->io_bookmark = *zb;
557 
558 	if (pio != NULL) {
559 		if (zio->io_logical == NULL)
560 			zio->io_logical = pio->io_logical;
561 		if (zio->io_child_type == ZIO_CHILD_GANG)
562 			zio->io_gang_leader = pio->io_gang_leader;
563 		zio_add_child(pio, zio);
564 	}
565 
566 	return (zio);
567 }
568 
569 static void
570 zio_destroy(zio_t *zio)
571 {
572 	list_destroy(&zio->io_parent_list);
573 	list_destroy(&zio->io_child_list);
574 	mutex_destroy(&zio->io_lock);
575 	cv_destroy(&zio->io_cv);
576 	kmem_cache_free(zio_cache, zio);
577 }
578 
579 zio_t *
580 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
581     void *private, enum zio_flag flags)
582 {
583 	zio_t *zio;
584 
585 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
586 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
587 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
588 
589 	return (zio);
590 }
591 
592 zio_t *
593 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
594 {
595 	return (zio_null(NULL, spa, NULL, done, private, flags));
596 }
597 
598 zio_t *
599 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
600     void *data, uint64_t size, zio_done_func_t *done, void *private,
601     zio_priority_t priority, enum zio_flag flags, const zbookmark_t *zb)
602 {
603 	zio_t *zio;
604 
605 	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
606 	    data, size, done, private,
607 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
608 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
609 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
610 
611 	return (zio);
612 }
613 
614 zio_t *
615 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
616     void *data, uint64_t size, const zio_prop_t *zp,
617     zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
618     void *private,
619     zio_priority_t priority, enum zio_flag flags, const zbookmark_t *zb)
620 {
621 	zio_t *zio;
622 
623 	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
624 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
625 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
626 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
627 	    DMU_OT_IS_VALID(zp->zp_type) &&
628 	    zp->zp_level < 32 &&
629 	    zp->zp_copies > 0 &&
630 	    zp->zp_copies <= spa_max_replication(spa));
631 
632 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
633 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
634 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
635 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
636 
637 	zio->io_ready = ready;
638 	zio->io_physdone = physdone;
639 	zio->io_prop = *zp;
640 
641 	/*
642 	 * Data can be NULL if we are going to call zio_write_override() to
643 	 * provide the already-allocated BP.  But we may need the data to
644 	 * verify a dedup hit (if requested).  In this case, don't try to
645 	 * dedup (just take the already-allocated BP verbatim).
646 	 */
647 	if (data == NULL && zio->io_prop.zp_dedup_verify) {
648 		zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
649 	}
650 
651 	return (zio);
652 }
653 
654 zio_t *
655 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
656     uint64_t size, zio_done_func_t *done, void *private,
657     zio_priority_t priority, enum zio_flag flags, zbookmark_t *zb)
658 {
659 	zio_t *zio;
660 
661 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
662 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
663 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
664 
665 	return (zio);
666 }
667 
668 void
669 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
670 {
671 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
672 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
673 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
674 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
675 
676 	/*
677 	 * We must reset the io_prop to match the values that existed
678 	 * when the bp was first written by dmu_sync() keeping in mind
679 	 * that nopwrite and dedup are mutually exclusive.
680 	 */
681 	zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
682 	zio->io_prop.zp_nopwrite = nopwrite;
683 	zio->io_prop.zp_copies = copies;
684 	zio->io_bp_override = bp;
685 }
686 
687 void
688 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
689 {
690 
691 	/*
692 	 * The check for EMBEDDED is a performance optimization.  We
693 	 * process the free here (by ignoring it) rather than
694 	 * putting it on the list and then processing it in zio_free_sync().
695 	 */
696 	if (BP_IS_EMBEDDED(bp))
697 		return;
698 	metaslab_check_free(spa, bp);
699 
700 	/*
701 	 * Frees that are for the currently-syncing txg, are not going to be
702 	 * deferred, and which will not need to do a read (i.e. not GANG or
703 	 * DEDUP), can be processed immediately.  Otherwise, put them on the
704 	 * in-memory list for later processing.
705 	 */
706 	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
707 	    txg != spa->spa_syncing_txg ||
708 	    spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
709 		bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
710 	} else {
711 		VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
712 	}
713 }
714 
715 zio_t *
716 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
717     enum zio_flag flags)
718 {
719 	zio_t *zio;
720 	enum zio_stage stage = ZIO_FREE_PIPELINE;
721 
722 	ASSERT(!BP_IS_HOLE(bp));
723 	ASSERT(spa_syncing_txg(spa) == txg);
724 	ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
725 
726 	if (BP_IS_EMBEDDED(bp))
727 		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
728 
729 	metaslab_check_free(spa, bp);
730 	arc_freed(spa, bp);
731 
732 	/*
733 	 * GANG and DEDUP blocks can induce a read (for the gang block header,
734 	 * or the DDT), so issue them asynchronously so that this thread is
735 	 * not tied up.
736 	 */
737 	if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
738 		stage |= ZIO_STAGE_ISSUE_ASYNC;
739 
740 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
741 	    NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
742 	    NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
743 
744 	return (zio);
745 }
746 
747 zio_t *
748 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
749     zio_done_func_t *done, void *private, enum zio_flag flags)
750 {
751 	zio_t *zio;
752 
753 	dprintf_bp(bp, "claiming in txg %llu", txg);
754 
755 	if (BP_IS_EMBEDDED(bp))
756 		return (zio_null(pio, spa, NULL, NULL, NULL, 0));
757 
758 	/*
759 	 * A claim is an allocation of a specific block.  Claims are needed
760 	 * to support immediate writes in the intent log.  The issue is that
761 	 * immediate writes contain committed data, but in a txg that was
762 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
763 	 * the intent log claims all blocks that contain immediate write data
764 	 * so that the SPA knows they're in use.
765 	 *
766 	 * All claims *must* be resolved in the first txg -- before the SPA
767 	 * starts allocating blocks -- so that nothing is allocated twice.
768 	 * If txg == 0 we just verify that the block is claimable.
769 	 */
770 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
771 	ASSERT(txg == spa_first_txg(spa) || txg == 0);
772 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
773 
774 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
775 	    done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
776 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
777 
778 	return (zio);
779 }
780 
781 zio_t *
782 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
783     zio_done_func_t *done, void *private, enum zio_flag flags)
784 {
785 	zio_t *zio;
786 	int c;
787 
788 	if (vd->vdev_children == 0) {
789 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
790 		    ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
791 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
792 
793 		zio->io_cmd = cmd;
794 	} else {
795 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
796 
797 		for (c = 0; c < vd->vdev_children; c++)
798 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
799 			    done, private, flags));
800 	}
801 
802 	return (zio);
803 }
804 
805 zio_t *
806 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
807     void *data, int checksum, zio_done_func_t *done, void *private,
808     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
809 {
810 	zio_t *zio;
811 
812 	ASSERT(vd->vdev_children == 0);
813 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
814 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
815 	ASSERT3U(offset + size, <=, vd->vdev_psize);
816 
817 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
818 	    ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
819 	    ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
820 
821 	zio->io_prop.zp_checksum = checksum;
822 
823 	return (zio);
824 }
825 
826 zio_t *
827 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
828     void *data, int checksum, zio_done_func_t *done, void *private,
829     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
830 {
831 	zio_t *zio;
832 
833 	ASSERT(vd->vdev_children == 0);
834 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
835 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
836 	ASSERT3U(offset + size, <=, vd->vdev_psize);
837 
838 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
839 	    ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
840 	    ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
841 
842 	zio->io_prop.zp_checksum = checksum;
843 
844 	if (zio_checksum_table[checksum].ci_eck) {
845 		/*
846 		 * zec checksums are necessarily destructive -- they modify
847 		 * the end of the write buffer to hold the verifier/checksum.
848 		 * Therefore, we must make a local copy in case the data is
849 		 * being written to multiple places in parallel.
850 		 */
851 		void *wbuf = zio_buf_alloc(size);
852 		bcopy(data, wbuf, size);
853 		zio_push_transform(zio, wbuf, size, size, NULL);
854 	}
855 
856 	return (zio);
857 }
858 
859 /*
860  * Create a child I/O to do some work for us.
861  */
862 zio_t *
863 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
864 	void *data, uint64_t size, int type, zio_priority_t priority,
865 	enum zio_flag flags, zio_done_func_t *done, void *private)
866 {
867 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
868 	zio_t *zio;
869 
870 	ASSERT(vd->vdev_parent ==
871 	    (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
872 
873 	if (type == ZIO_TYPE_READ && bp != NULL) {
874 		/*
875 		 * If we have the bp, then the child should perform the
876 		 * checksum and the parent need not.  This pushes error
877 		 * detection as close to the leaves as possible and
878 		 * eliminates redundant checksums in the interior nodes.
879 		 */
880 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
881 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
882 	}
883 
884 	if (vd->vdev_children == 0)
885 		offset += VDEV_LABEL_START_SIZE;
886 
887 	flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
888 
889 	/*
890 	 * If we've decided to do a repair, the write is not speculative --
891 	 * even if the original read was.
892 	 */
893 	if (flags & ZIO_FLAG_IO_REPAIR)
894 		flags &= ~ZIO_FLAG_SPECULATIVE;
895 
896 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
897 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
898 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
899 
900 	zio->io_physdone = pio->io_physdone;
901 	if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
902 		zio->io_logical->io_phys_children++;
903 
904 	return (zio);
905 }
906 
907 zio_t *
908 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
909 	int type, zio_priority_t priority, enum zio_flag flags,
910 	zio_done_func_t *done, void *private)
911 {
912 	zio_t *zio;
913 
914 	ASSERT(vd->vdev_ops->vdev_op_leaf);
915 
916 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
917 	    data, size, done, private, type, priority,
918 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
919 	    vd, offset, NULL,
920 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
921 
922 	return (zio);
923 }
924 
925 void
926 zio_flush(zio_t *zio, vdev_t *vd)
927 {
928 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
929 	    NULL, NULL,
930 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
931 }
932 
933 void
934 zio_shrink(zio_t *zio, uint64_t size)
935 {
936 	ASSERT(zio->io_executor == NULL);
937 	ASSERT(zio->io_orig_size == zio->io_size);
938 	ASSERT(size <= zio->io_size);
939 
940 	/*
941 	 * We don't shrink for raidz because of problems with the
942 	 * reconstruction when reading back less than the block size.
943 	 * Note, BP_IS_RAIDZ() assumes no compression.
944 	 */
945 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
946 	if (!BP_IS_RAIDZ(zio->io_bp))
947 		zio->io_orig_size = zio->io_size = size;
948 }
949 
950 /*
951  * ==========================================================================
952  * Prepare to read and write logical blocks
953  * ==========================================================================
954  */
955 
956 static int
957 zio_read_bp_init(zio_t *zio)
958 {
959 	blkptr_t *bp = zio->io_bp;
960 
961 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
962 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
963 	    !(zio->io_flags & ZIO_FLAG_RAW)) {
964 		uint64_t psize =
965 		    BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
966 		void *cbuf = zio_buf_alloc(psize);
967 
968 		zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
969 	}
970 
971 	if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
972 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
973 		decode_embedded_bp_compressed(bp, zio->io_data);
974 	} else {
975 		ASSERT(!BP_IS_EMBEDDED(bp));
976 	}
977 
978 	if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
979 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
980 
981 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
982 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
983 
984 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
985 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
986 
987 	return (ZIO_PIPELINE_CONTINUE);
988 }
989 
990 static int
991 zio_write_bp_init(zio_t *zio)
992 {
993 	spa_t *spa = zio->io_spa;
994 	zio_prop_t *zp = &zio->io_prop;
995 	enum zio_compress compress = zp->zp_compress;
996 	blkptr_t *bp = zio->io_bp;
997 	uint64_t lsize = zio->io_size;
998 	uint64_t psize = lsize;
999 	int pass = 1;
1000 
1001 	/*
1002 	 * If our children haven't all reached the ready stage,
1003 	 * wait for them and then repeat this pipeline stage.
1004 	 */
1005 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1006 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1007 		return (ZIO_PIPELINE_STOP);
1008 
1009 	if (!IO_IS_ALLOCATING(zio))
1010 		return (ZIO_PIPELINE_CONTINUE);
1011 
1012 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1013 
1014 	if (zio->io_bp_override) {
1015 		ASSERT(bp->blk_birth != zio->io_txg);
1016 		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1017 
1018 		*bp = *zio->io_bp_override;
1019 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1020 
1021 		if (BP_IS_EMBEDDED(bp))
1022 			return (ZIO_PIPELINE_CONTINUE);
1023 
1024 		/*
1025 		 * If we've been overridden and nopwrite is set then
1026 		 * set the flag accordingly to indicate that a nopwrite
1027 		 * has already occurred.
1028 		 */
1029 		if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1030 			ASSERT(!zp->zp_dedup);
1031 			zio->io_flags |= ZIO_FLAG_NOPWRITE;
1032 			return (ZIO_PIPELINE_CONTINUE);
1033 		}
1034 
1035 		ASSERT(!zp->zp_nopwrite);
1036 
1037 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1038 			return (ZIO_PIPELINE_CONTINUE);
1039 
1040 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1041 		    zp->zp_dedup_verify);
1042 
1043 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1044 			BP_SET_DEDUP(bp, 1);
1045 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1046 			return (ZIO_PIPELINE_CONTINUE);
1047 		}
1048 		zio->io_bp_override = NULL;
1049 		BP_ZERO(bp);
1050 	}
1051 
1052 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1053 		/*
1054 		 * We're rewriting an existing block, which means we're
1055 		 * working on behalf of spa_sync().  For spa_sync() to
1056 		 * converge, it must eventually be the case that we don't
1057 		 * have to allocate new blocks.  But compression changes
1058 		 * the blocksize, which forces a reallocate, and makes
1059 		 * convergence take longer.  Therefore, after the first
1060 		 * few passes, stop compressing to ensure convergence.
1061 		 */
1062 		pass = spa_sync_pass(spa);
1063 
1064 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
1065 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1066 		ASSERT(!BP_GET_DEDUP(bp));
1067 
1068 		if (pass >= zfs_sync_pass_dont_compress)
1069 			compress = ZIO_COMPRESS_OFF;
1070 
1071 		/* Make sure someone doesn't change their mind on overwrites */
1072 		ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1073 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1074 	}
1075 
1076 	if (compress != ZIO_COMPRESS_OFF) {
1077 		void *cbuf = zio_buf_alloc(lsize);
1078 		psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1079 		if (psize == 0 || psize == lsize) {
1080 			compress = ZIO_COMPRESS_OFF;
1081 			zio_buf_free(cbuf, lsize);
1082 		} else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1083 		    zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1084 		    spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1085 			encode_embedded_bp_compressed(bp,
1086 			    cbuf, compress, lsize, psize);
1087 			BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1088 			BP_SET_TYPE(bp, zio->io_prop.zp_type);
1089 			BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1090 			zio_buf_free(cbuf, lsize);
1091 			bp->blk_birth = zio->io_txg;
1092 			zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1093 			ASSERT(spa_feature_is_active(spa,
1094 			    SPA_FEATURE_EMBEDDED_DATA));
1095 			return (ZIO_PIPELINE_CONTINUE);
1096 		} else {
1097 			/*
1098 			 * Round up compressed size to MINBLOCKSIZE and
1099 			 * zero the tail.
1100 			 */
1101 			size_t rounded =
1102 			    P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE);
1103 			if (rounded > psize) {
1104 				bzero((char *)cbuf + psize, rounded - psize);
1105 				psize = rounded;
1106 			}
1107 			if (psize == lsize) {
1108 				compress = ZIO_COMPRESS_OFF;
1109 				zio_buf_free(cbuf, lsize);
1110 			} else {
1111 				zio_push_transform(zio, cbuf,
1112 				    psize, lsize, NULL);
1113 			}
1114 		}
1115 	}
1116 
1117 	/*
1118 	 * The final pass of spa_sync() must be all rewrites, but the first
1119 	 * few passes offer a trade-off: allocating blocks defers convergence,
1120 	 * but newly allocated blocks are sequential, so they can be written
1121 	 * to disk faster.  Therefore, we allow the first few passes of
1122 	 * spa_sync() to allocate new blocks, but force rewrites after that.
1123 	 * There should only be a handful of blocks after pass 1 in any case.
1124 	 */
1125 	if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1126 	    BP_GET_PSIZE(bp) == psize &&
1127 	    pass >= zfs_sync_pass_rewrite) {
1128 		ASSERT(psize != 0);
1129 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1130 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1131 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1132 	} else {
1133 		BP_ZERO(bp);
1134 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1135 	}
1136 
1137 	if (psize == 0) {
1138 		if (zio->io_bp_orig.blk_birth != 0 &&
1139 		    spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1140 			BP_SET_LSIZE(bp, lsize);
1141 			BP_SET_TYPE(bp, zp->zp_type);
1142 			BP_SET_LEVEL(bp, zp->zp_level);
1143 			BP_SET_BIRTH(bp, zio->io_txg, 0);
1144 		}
1145 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1146 	} else {
1147 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1148 		BP_SET_LSIZE(bp, lsize);
1149 		BP_SET_TYPE(bp, zp->zp_type);
1150 		BP_SET_LEVEL(bp, zp->zp_level);
1151 		BP_SET_PSIZE(bp, psize);
1152 		BP_SET_COMPRESS(bp, compress);
1153 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1154 		BP_SET_DEDUP(bp, zp->zp_dedup);
1155 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1156 		if (zp->zp_dedup) {
1157 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1158 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1159 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1160 		}
1161 		if (zp->zp_nopwrite) {
1162 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1163 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1164 			zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1165 		}
1166 	}
1167 
1168 	return (ZIO_PIPELINE_CONTINUE);
1169 }
1170 
1171 static int
1172 zio_free_bp_init(zio_t *zio)
1173 {
1174 	blkptr_t *bp = zio->io_bp;
1175 
1176 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1177 		if (BP_GET_DEDUP(bp))
1178 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1179 	}
1180 
1181 	return (ZIO_PIPELINE_CONTINUE);
1182 }
1183 
1184 /*
1185  * ==========================================================================
1186  * Execute the I/O pipeline
1187  * ==========================================================================
1188  */
1189 
1190 static void
1191 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1192 {
1193 	spa_t *spa = zio->io_spa;
1194 	zio_type_t t = zio->io_type;
1195 	int flags = (cutinline ? TQ_FRONT : 0);
1196 
1197 	/*
1198 	 * If we're a config writer or a probe, the normal issue and
1199 	 * interrupt threads may all be blocked waiting for the config lock.
1200 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1201 	 */
1202 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1203 		t = ZIO_TYPE_NULL;
1204 
1205 	/*
1206 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1207 	 */
1208 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1209 		t = ZIO_TYPE_NULL;
1210 
1211 	/*
1212 	 * If this is a high priority I/O, then use the high priority taskq if
1213 	 * available.
1214 	 */
1215 	if (zio->io_priority == ZIO_PRIORITY_NOW &&
1216 	    spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1217 		q++;
1218 
1219 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1220 
1221 	/*
1222 	 * NB: We are assuming that the zio can only be dispatched
1223 	 * to a single taskq at a time.  It would be a grievous error
1224 	 * to dispatch the zio to another taskq at the same time.
1225 	 */
1226 	ASSERT(zio->io_tqent.tqent_next == NULL);
1227 	spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1228 	    flags, &zio->io_tqent);
1229 }
1230 
1231 static boolean_t
1232 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1233 {
1234 	kthread_t *executor = zio->io_executor;
1235 	spa_t *spa = zio->io_spa;
1236 
1237 	for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1238 		spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1239 		uint_t i;
1240 		for (i = 0; i < tqs->stqs_count; i++) {
1241 			if (taskq_member(tqs->stqs_taskq[i], executor))
1242 				return (B_TRUE);
1243 		}
1244 	}
1245 
1246 	return (B_FALSE);
1247 }
1248 
1249 static int
1250 zio_issue_async(zio_t *zio)
1251 {
1252 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1253 
1254 	return (ZIO_PIPELINE_STOP);
1255 }
1256 
1257 void
1258 zio_interrupt(zio_t *zio)
1259 {
1260 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1261 }
1262 
1263 /*
1264  * Execute the I/O pipeline until one of the following occurs:
1265  *
1266  *	(1) the I/O completes
1267  *	(2) the pipeline stalls waiting for dependent child I/Os
1268  *	(3) the I/O issues, so we're waiting for an I/O completion interrupt
1269  *	(4) the I/O is delegated by vdev-level caching or aggregation
1270  *	(5) the I/O is deferred due to vdev-level queueing
1271  *	(6) the I/O is handed off to another thread.
1272  *
1273  * In all cases, the pipeline stops whenever there's no CPU work; it never
1274  * burns a thread in cv_wait().
1275  *
1276  * There's no locking on io_stage because there's no legitimate way
1277  * for multiple threads to be attempting to process the same I/O.
1278  */
1279 static zio_pipe_stage_t *zio_pipeline[];
1280 
1281 void
1282 zio_execute(zio_t *zio)
1283 {
1284 	zio->io_executor = curthread;
1285 
1286 	while (zio->io_stage < ZIO_STAGE_DONE) {
1287 		enum zio_stage pipeline = zio->io_pipeline;
1288 		enum zio_stage stage = zio->io_stage;
1289 		int rv;
1290 
1291 		ASSERT(!MUTEX_HELD(&zio->io_lock));
1292 		ASSERT(ISP2(stage));
1293 		ASSERT(zio->io_stall == NULL);
1294 
1295 		do {
1296 			stage <<= 1;
1297 		} while ((stage & pipeline) == 0);
1298 
1299 		ASSERT(stage <= ZIO_STAGE_DONE);
1300 
1301 		/*
1302 		 * If we are in interrupt context and this pipeline stage
1303 		 * will grab a config lock that is held across I/O,
1304 		 * or may wait for an I/O that needs an interrupt thread
1305 		 * to complete, issue async to avoid deadlock.
1306 		 *
1307 		 * For VDEV_IO_START, we cut in line so that the io will
1308 		 * be sent to disk promptly.
1309 		 */
1310 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1311 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1312 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1313 			    zio_requeue_io_start_cut_in_line : B_FALSE;
1314 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1315 			return;
1316 		}
1317 
1318 		zio->io_stage = stage;
1319 		rv = zio_pipeline[highbit64(stage) - 1](zio);
1320 
1321 		if (rv == ZIO_PIPELINE_STOP)
1322 			return;
1323 
1324 		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1325 	}
1326 }
1327 
1328 /*
1329  * ==========================================================================
1330  * Initiate I/O, either sync or async
1331  * ==========================================================================
1332  */
1333 int
1334 zio_wait(zio_t *zio)
1335 {
1336 	int error;
1337 
1338 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1339 	ASSERT(zio->io_executor == NULL);
1340 
1341 	zio->io_waiter = curthread;
1342 
1343 	zio_execute(zio);
1344 
1345 	mutex_enter(&zio->io_lock);
1346 	while (zio->io_executor != NULL)
1347 		cv_wait(&zio->io_cv, &zio->io_lock);
1348 	mutex_exit(&zio->io_lock);
1349 
1350 	error = zio->io_error;
1351 	zio_destroy(zio);
1352 
1353 	return (error);
1354 }
1355 
1356 void
1357 zio_nowait(zio_t *zio)
1358 {
1359 	ASSERT(zio->io_executor == NULL);
1360 
1361 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1362 	    zio_unique_parent(zio) == NULL) {
1363 		/*
1364 		 * This is a logical async I/O with no parent to wait for it.
1365 		 * We add it to the spa_async_root_zio "Godfather" I/O which
1366 		 * will ensure they complete prior to unloading the pool.
1367 		 */
1368 		spa_t *spa = zio->io_spa;
1369 
1370 		zio_add_child(spa->spa_async_zio_root, zio);
1371 	}
1372 
1373 	zio_execute(zio);
1374 }
1375 
1376 /*
1377  * ==========================================================================
1378  * Reexecute or suspend/resume failed I/O
1379  * ==========================================================================
1380  */
1381 
1382 static void
1383 zio_reexecute(zio_t *pio)
1384 {
1385 	zio_t *cio, *cio_next;
1386 
1387 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1388 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1389 	ASSERT(pio->io_gang_leader == NULL);
1390 	ASSERT(pio->io_gang_tree == NULL);
1391 
1392 	pio->io_flags = pio->io_orig_flags;
1393 	pio->io_stage = pio->io_orig_stage;
1394 	pio->io_pipeline = pio->io_orig_pipeline;
1395 	pio->io_reexecute = 0;
1396 	pio->io_flags |= ZIO_FLAG_REEXECUTED;
1397 	pio->io_error = 0;
1398 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1399 		pio->io_state[w] = 0;
1400 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1401 		pio->io_child_error[c] = 0;
1402 
1403 	if (IO_IS_ALLOCATING(pio))
1404 		BP_ZERO(pio->io_bp);
1405 
1406 	/*
1407 	 * As we reexecute pio's children, new children could be created.
1408 	 * New children go to the head of pio's io_child_list, however,
1409 	 * so we will (correctly) not reexecute them.  The key is that
1410 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1411 	 * cannot be affected by any side effects of reexecuting 'cio'.
1412 	 */
1413 	for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1414 		cio_next = zio_walk_children(pio);
1415 		mutex_enter(&pio->io_lock);
1416 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1417 			pio->io_children[cio->io_child_type][w]++;
1418 		mutex_exit(&pio->io_lock);
1419 		zio_reexecute(cio);
1420 	}
1421 
1422 	/*
1423 	 * Now that all children have been reexecuted, execute the parent.
1424 	 * We don't reexecute "The Godfather" I/O here as it's the
1425 	 * responsibility of the caller to wait on him.
1426 	 */
1427 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1428 		zio_execute(pio);
1429 }
1430 
1431 void
1432 zio_suspend(spa_t *spa, zio_t *zio)
1433 {
1434 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1435 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1436 		    "failure and the failure mode property for this pool "
1437 		    "is set to panic.", spa_name(spa));
1438 
1439 	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1440 
1441 	mutex_enter(&spa->spa_suspend_lock);
1442 
1443 	if (spa->spa_suspend_zio_root == NULL)
1444 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1445 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1446 		    ZIO_FLAG_GODFATHER);
1447 
1448 	spa->spa_suspended = B_TRUE;
1449 
1450 	if (zio != NULL) {
1451 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1452 		ASSERT(zio != spa->spa_suspend_zio_root);
1453 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1454 		ASSERT(zio_unique_parent(zio) == NULL);
1455 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1456 		zio_add_child(spa->spa_suspend_zio_root, zio);
1457 	}
1458 
1459 	mutex_exit(&spa->spa_suspend_lock);
1460 }
1461 
1462 int
1463 zio_resume(spa_t *spa)
1464 {
1465 	zio_t *pio;
1466 
1467 	/*
1468 	 * Reexecute all previously suspended i/o.
1469 	 */
1470 	mutex_enter(&spa->spa_suspend_lock);
1471 	spa->spa_suspended = B_FALSE;
1472 	cv_broadcast(&spa->spa_suspend_cv);
1473 	pio = spa->spa_suspend_zio_root;
1474 	spa->spa_suspend_zio_root = NULL;
1475 	mutex_exit(&spa->spa_suspend_lock);
1476 
1477 	if (pio == NULL)
1478 		return (0);
1479 
1480 	zio_reexecute(pio);
1481 	return (zio_wait(pio));
1482 }
1483 
1484 void
1485 zio_resume_wait(spa_t *spa)
1486 {
1487 	mutex_enter(&spa->spa_suspend_lock);
1488 	while (spa_suspended(spa))
1489 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1490 	mutex_exit(&spa->spa_suspend_lock);
1491 }
1492 
1493 /*
1494  * ==========================================================================
1495  * Gang blocks.
1496  *
1497  * A gang block is a collection of small blocks that looks to the DMU
1498  * like one large block.  When zio_dva_allocate() cannot find a block
1499  * of the requested size, due to either severe fragmentation or the pool
1500  * being nearly full, it calls zio_write_gang_block() to construct the
1501  * block from smaller fragments.
1502  *
1503  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1504  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
1505  * an indirect block: it's an array of block pointers.  It consumes
1506  * only one sector and hence is allocatable regardless of fragmentation.
1507  * The gang header's bps point to its gang members, which hold the data.
1508  *
1509  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1510  * as the verifier to ensure uniqueness of the SHA256 checksum.
1511  * Critically, the gang block bp's blk_cksum is the checksum of the data,
1512  * not the gang header.  This ensures that data block signatures (needed for
1513  * deduplication) are independent of how the block is physically stored.
1514  *
1515  * Gang blocks can be nested: a gang member may itself be a gang block.
1516  * Thus every gang block is a tree in which root and all interior nodes are
1517  * gang headers, and the leaves are normal blocks that contain user data.
1518  * The root of the gang tree is called the gang leader.
1519  *
1520  * To perform any operation (read, rewrite, free, claim) on a gang block,
1521  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1522  * in the io_gang_tree field of the original logical i/o by recursively
1523  * reading the gang leader and all gang headers below it.  This yields
1524  * an in-core tree containing the contents of every gang header and the
1525  * bps for every constituent of the gang block.
1526  *
1527  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1528  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
1529  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1530  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1531  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1532  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
1533  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1534  * of the gang header plus zio_checksum_compute() of the data to update the
1535  * gang header's blk_cksum as described above.
1536  *
1537  * The two-phase assemble/issue model solves the problem of partial failure --
1538  * what if you'd freed part of a gang block but then couldn't read the
1539  * gang header for another part?  Assembling the entire gang tree first
1540  * ensures that all the necessary gang header I/O has succeeded before
1541  * starting the actual work of free, claim, or write.  Once the gang tree
1542  * is assembled, free and claim are in-memory operations that cannot fail.
1543  *
1544  * In the event that a gang write fails, zio_dva_unallocate() walks the
1545  * gang tree to immediately free (i.e. insert back into the space map)
1546  * everything we've allocated.  This ensures that we don't get ENOSPC
1547  * errors during repeated suspend/resume cycles due to a flaky device.
1548  *
1549  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
1550  * the gang tree, we won't modify the block, so we can safely defer the free
1551  * (knowing that the block is still intact).  If we *can* assemble the gang
1552  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1553  * each constituent bp and we can allocate a new block on the next sync pass.
1554  *
1555  * In all cases, the gang tree allows complete recovery from partial failure.
1556  * ==========================================================================
1557  */
1558 
1559 static zio_t *
1560 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1561 {
1562 	if (gn != NULL)
1563 		return (pio);
1564 
1565 	return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1566 	    NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1567 	    &pio->io_bookmark));
1568 }
1569 
1570 zio_t *
1571 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1572 {
1573 	zio_t *zio;
1574 
1575 	if (gn != NULL) {
1576 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1577 		    gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1578 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1579 		/*
1580 		 * As we rewrite each gang header, the pipeline will compute
1581 		 * a new gang block header checksum for it; but no one will
1582 		 * compute a new data checksum, so we do that here.  The one
1583 		 * exception is the gang leader: the pipeline already computed
1584 		 * its data checksum because that stage precedes gang assembly.
1585 		 * (Presently, nothing actually uses interior data checksums;
1586 		 * this is just good hygiene.)
1587 		 */
1588 		if (gn != pio->io_gang_leader->io_gang_tree) {
1589 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1590 			    data, BP_GET_PSIZE(bp));
1591 		}
1592 		/*
1593 		 * If we are here to damage data for testing purposes,
1594 		 * leave the GBH alone so that we can detect the damage.
1595 		 */
1596 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1597 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1598 	} else {
1599 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1600 		    data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1601 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1602 	}
1603 
1604 	return (zio);
1605 }
1606 
1607 /* ARGSUSED */
1608 zio_t *
1609 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1610 {
1611 	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1612 	    ZIO_GANG_CHILD_FLAGS(pio)));
1613 }
1614 
1615 /* ARGSUSED */
1616 zio_t *
1617 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1618 {
1619 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1620 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1621 }
1622 
1623 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1624 	NULL,
1625 	zio_read_gang,
1626 	zio_rewrite_gang,
1627 	zio_free_gang,
1628 	zio_claim_gang,
1629 	NULL
1630 };
1631 
1632 static void zio_gang_tree_assemble_done(zio_t *zio);
1633 
1634 static zio_gang_node_t *
1635 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1636 {
1637 	zio_gang_node_t *gn;
1638 
1639 	ASSERT(*gnpp == NULL);
1640 
1641 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1642 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1643 	*gnpp = gn;
1644 
1645 	return (gn);
1646 }
1647 
1648 static void
1649 zio_gang_node_free(zio_gang_node_t **gnpp)
1650 {
1651 	zio_gang_node_t *gn = *gnpp;
1652 
1653 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1654 		ASSERT(gn->gn_child[g] == NULL);
1655 
1656 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1657 	kmem_free(gn, sizeof (*gn));
1658 	*gnpp = NULL;
1659 }
1660 
1661 static void
1662 zio_gang_tree_free(zio_gang_node_t **gnpp)
1663 {
1664 	zio_gang_node_t *gn = *gnpp;
1665 
1666 	if (gn == NULL)
1667 		return;
1668 
1669 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1670 		zio_gang_tree_free(&gn->gn_child[g]);
1671 
1672 	zio_gang_node_free(gnpp);
1673 }
1674 
1675 static void
1676 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1677 {
1678 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1679 
1680 	ASSERT(gio->io_gang_leader == gio);
1681 	ASSERT(BP_IS_GANG(bp));
1682 
1683 	zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1684 	    SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1685 	    gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1686 }
1687 
1688 static void
1689 zio_gang_tree_assemble_done(zio_t *zio)
1690 {
1691 	zio_t *gio = zio->io_gang_leader;
1692 	zio_gang_node_t *gn = zio->io_private;
1693 	blkptr_t *bp = zio->io_bp;
1694 
1695 	ASSERT(gio == zio_unique_parent(zio));
1696 	ASSERT(zio->io_child_count == 0);
1697 
1698 	if (zio->io_error)
1699 		return;
1700 
1701 	if (BP_SHOULD_BYTESWAP(bp))
1702 		byteswap_uint64_array(zio->io_data, zio->io_size);
1703 
1704 	ASSERT(zio->io_data == gn->gn_gbh);
1705 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1706 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1707 
1708 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1709 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1710 		if (!BP_IS_GANG(gbp))
1711 			continue;
1712 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1713 	}
1714 }
1715 
1716 static void
1717 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1718 {
1719 	zio_t *gio = pio->io_gang_leader;
1720 	zio_t *zio;
1721 
1722 	ASSERT(BP_IS_GANG(bp) == !!gn);
1723 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1724 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1725 
1726 	/*
1727 	 * If you're a gang header, your data is in gn->gn_gbh.
1728 	 * If you're a gang member, your data is in 'data' and gn == NULL.
1729 	 */
1730 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1731 
1732 	if (gn != NULL) {
1733 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1734 
1735 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1736 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1737 			if (BP_IS_HOLE(gbp))
1738 				continue;
1739 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1740 			data = (char *)data + BP_GET_PSIZE(gbp);
1741 		}
1742 	}
1743 
1744 	if (gn == gio->io_gang_tree)
1745 		ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1746 
1747 	if (zio != pio)
1748 		zio_nowait(zio);
1749 }
1750 
1751 static int
1752 zio_gang_assemble(zio_t *zio)
1753 {
1754 	blkptr_t *bp = zio->io_bp;
1755 
1756 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1757 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1758 
1759 	zio->io_gang_leader = zio;
1760 
1761 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1762 
1763 	return (ZIO_PIPELINE_CONTINUE);
1764 }
1765 
1766 static int
1767 zio_gang_issue(zio_t *zio)
1768 {
1769 	blkptr_t *bp = zio->io_bp;
1770 
1771 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1772 		return (ZIO_PIPELINE_STOP);
1773 
1774 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1775 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1776 
1777 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1778 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1779 	else
1780 		zio_gang_tree_free(&zio->io_gang_tree);
1781 
1782 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1783 
1784 	return (ZIO_PIPELINE_CONTINUE);
1785 }
1786 
1787 static void
1788 zio_write_gang_member_ready(zio_t *zio)
1789 {
1790 	zio_t *pio = zio_unique_parent(zio);
1791 	zio_t *gio = zio->io_gang_leader;
1792 	dva_t *cdva = zio->io_bp->blk_dva;
1793 	dva_t *pdva = pio->io_bp->blk_dva;
1794 	uint64_t asize;
1795 
1796 	if (BP_IS_HOLE(zio->io_bp))
1797 		return;
1798 
1799 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1800 
1801 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1802 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1803 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1804 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1805 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1806 
1807 	mutex_enter(&pio->io_lock);
1808 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1809 		ASSERT(DVA_GET_GANG(&pdva[d]));
1810 		asize = DVA_GET_ASIZE(&pdva[d]);
1811 		asize += DVA_GET_ASIZE(&cdva[d]);
1812 		DVA_SET_ASIZE(&pdva[d], asize);
1813 	}
1814 	mutex_exit(&pio->io_lock);
1815 }
1816 
1817 static int
1818 zio_write_gang_block(zio_t *pio)
1819 {
1820 	spa_t *spa = pio->io_spa;
1821 	blkptr_t *bp = pio->io_bp;
1822 	zio_t *gio = pio->io_gang_leader;
1823 	zio_t *zio;
1824 	zio_gang_node_t *gn, **gnpp;
1825 	zio_gbh_phys_t *gbh;
1826 	uint64_t txg = pio->io_txg;
1827 	uint64_t resid = pio->io_size;
1828 	uint64_t lsize;
1829 	int copies = gio->io_prop.zp_copies;
1830 	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1831 	zio_prop_t zp;
1832 	int error;
1833 
1834 	error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1835 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1836 	    METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1837 	if (error) {
1838 		pio->io_error = error;
1839 		return (ZIO_PIPELINE_CONTINUE);
1840 	}
1841 
1842 	if (pio == gio) {
1843 		gnpp = &gio->io_gang_tree;
1844 	} else {
1845 		gnpp = pio->io_private;
1846 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
1847 	}
1848 
1849 	gn = zio_gang_node_alloc(gnpp);
1850 	gbh = gn->gn_gbh;
1851 	bzero(gbh, SPA_GANGBLOCKSIZE);
1852 
1853 	/*
1854 	 * Create the gang header.
1855 	 */
1856 	zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1857 	    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1858 
1859 	/*
1860 	 * Create and nowait the gang children.
1861 	 */
1862 	for (int g = 0; resid != 0; resid -= lsize, g++) {
1863 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1864 		    SPA_MINBLOCKSIZE);
1865 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1866 
1867 		zp.zp_checksum = gio->io_prop.zp_checksum;
1868 		zp.zp_compress = ZIO_COMPRESS_OFF;
1869 		zp.zp_type = DMU_OT_NONE;
1870 		zp.zp_level = 0;
1871 		zp.zp_copies = gio->io_prop.zp_copies;
1872 		zp.zp_dedup = B_FALSE;
1873 		zp.zp_dedup_verify = B_FALSE;
1874 		zp.zp_nopwrite = B_FALSE;
1875 
1876 		zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1877 		    (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1878 		    zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1879 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1880 		    &pio->io_bookmark));
1881 	}
1882 
1883 	/*
1884 	 * Set pio's pipeline to just wait for zio to finish.
1885 	 */
1886 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1887 
1888 	zio_nowait(zio);
1889 
1890 	return (ZIO_PIPELINE_CONTINUE);
1891 }
1892 
1893 /*
1894  * The zio_nop_write stage in the pipeline determines if allocating
1895  * a new bp is necessary.  By leveraging a cryptographically secure checksum,
1896  * such as SHA256, we can compare the checksums of the new data and the old
1897  * to determine if allocating a new block is required.  The nopwrite
1898  * feature can handle writes in either syncing or open context (i.e. zil
1899  * writes) and as a result is mutually exclusive with dedup.
1900  */
1901 static int
1902 zio_nop_write(zio_t *zio)
1903 {
1904 	blkptr_t *bp = zio->io_bp;
1905 	blkptr_t *bp_orig = &zio->io_bp_orig;
1906 	zio_prop_t *zp = &zio->io_prop;
1907 
1908 	ASSERT(BP_GET_LEVEL(bp) == 0);
1909 	ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1910 	ASSERT(zp->zp_nopwrite);
1911 	ASSERT(!zp->zp_dedup);
1912 	ASSERT(zio->io_bp_override == NULL);
1913 	ASSERT(IO_IS_ALLOCATING(zio));
1914 
1915 	/*
1916 	 * Check to see if the original bp and the new bp have matching
1917 	 * characteristics (i.e. same checksum, compression algorithms, etc).
1918 	 * If they don't then just continue with the pipeline which will
1919 	 * allocate a new bp.
1920 	 */
1921 	if (BP_IS_HOLE(bp_orig) ||
1922 	    !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1923 	    BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1924 	    BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1925 	    BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1926 	    zp->zp_copies != BP_GET_NDVAS(bp_orig))
1927 		return (ZIO_PIPELINE_CONTINUE);
1928 
1929 	/*
1930 	 * If the checksums match then reset the pipeline so that we
1931 	 * avoid allocating a new bp and issuing any I/O.
1932 	 */
1933 	if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1934 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1935 		ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1936 		ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1937 		ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1938 		ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1939 		    sizeof (uint64_t)) == 0);
1940 
1941 		*bp = *bp_orig;
1942 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1943 		zio->io_flags |= ZIO_FLAG_NOPWRITE;
1944 	}
1945 
1946 	return (ZIO_PIPELINE_CONTINUE);
1947 }
1948 
1949 /*
1950  * ==========================================================================
1951  * Dedup
1952  * ==========================================================================
1953  */
1954 static void
1955 zio_ddt_child_read_done(zio_t *zio)
1956 {
1957 	blkptr_t *bp = zio->io_bp;
1958 	ddt_entry_t *dde = zio->io_private;
1959 	ddt_phys_t *ddp;
1960 	zio_t *pio = zio_unique_parent(zio);
1961 
1962 	mutex_enter(&pio->io_lock);
1963 	ddp = ddt_phys_select(dde, bp);
1964 	if (zio->io_error == 0)
1965 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
1966 	if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1967 		dde->dde_repair_data = zio->io_data;
1968 	else
1969 		zio_buf_free(zio->io_data, zio->io_size);
1970 	mutex_exit(&pio->io_lock);
1971 }
1972 
1973 static int
1974 zio_ddt_read_start(zio_t *zio)
1975 {
1976 	blkptr_t *bp = zio->io_bp;
1977 
1978 	ASSERT(BP_GET_DEDUP(bp));
1979 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1980 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1981 
1982 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1983 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1984 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1985 		ddt_phys_t *ddp = dde->dde_phys;
1986 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1987 		blkptr_t blk;
1988 
1989 		ASSERT(zio->io_vsd == NULL);
1990 		zio->io_vsd = dde;
1991 
1992 		if (ddp_self == NULL)
1993 			return (ZIO_PIPELINE_CONTINUE);
1994 
1995 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1996 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1997 				continue;
1998 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1999 			    &blk);
2000 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
2001 			    zio_buf_alloc(zio->io_size), zio->io_size,
2002 			    zio_ddt_child_read_done, dde, zio->io_priority,
2003 			    ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2004 			    &zio->io_bookmark));
2005 		}
2006 		return (ZIO_PIPELINE_CONTINUE);
2007 	}
2008 
2009 	zio_nowait(zio_read(zio, zio->io_spa, bp,
2010 	    zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2011 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2012 
2013 	return (ZIO_PIPELINE_CONTINUE);
2014 }
2015 
2016 static int
2017 zio_ddt_read_done(zio_t *zio)
2018 {
2019 	blkptr_t *bp = zio->io_bp;
2020 
2021 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2022 		return (ZIO_PIPELINE_STOP);
2023 
2024 	ASSERT(BP_GET_DEDUP(bp));
2025 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2026 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2027 
2028 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
2029 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
2030 		ddt_entry_t *dde = zio->io_vsd;
2031 		if (ddt == NULL) {
2032 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2033 			return (ZIO_PIPELINE_CONTINUE);
2034 		}
2035 		if (dde == NULL) {
2036 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2037 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2038 			return (ZIO_PIPELINE_STOP);
2039 		}
2040 		if (dde->dde_repair_data != NULL) {
2041 			bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2042 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
2043 		}
2044 		ddt_repair_done(ddt, dde);
2045 		zio->io_vsd = NULL;
2046 	}
2047 
2048 	ASSERT(zio->io_vsd == NULL);
2049 
2050 	return (ZIO_PIPELINE_CONTINUE);
2051 }
2052 
2053 static boolean_t
2054 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2055 {
2056 	spa_t *spa = zio->io_spa;
2057 
2058 	/*
2059 	 * Note: we compare the original data, not the transformed data,
2060 	 * because when zio->io_bp is an override bp, we will not have
2061 	 * pushed the I/O transforms.  That's an important optimization
2062 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2063 	 */
2064 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2065 		zio_t *lio = dde->dde_lead_zio[p];
2066 
2067 		if (lio != NULL) {
2068 			return (lio->io_orig_size != zio->io_orig_size ||
2069 			    bcmp(zio->io_orig_data, lio->io_orig_data,
2070 			    zio->io_orig_size) != 0);
2071 		}
2072 	}
2073 
2074 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2075 		ddt_phys_t *ddp = &dde->dde_phys[p];
2076 
2077 		if (ddp->ddp_phys_birth != 0) {
2078 			arc_buf_t *abuf = NULL;
2079 			uint32_t aflags = ARC_WAIT;
2080 			blkptr_t blk = *zio->io_bp;
2081 			int error;
2082 
2083 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2084 
2085 			ddt_exit(ddt);
2086 
2087 			error = arc_read(NULL, spa, &blk,
2088 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2089 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2090 			    &aflags, &zio->io_bookmark);
2091 
2092 			if (error == 0) {
2093 				if (arc_buf_size(abuf) != zio->io_orig_size ||
2094 				    bcmp(abuf->b_data, zio->io_orig_data,
2095 				    zio->io_orig_size) != 0)
2096 					error = SET_ERROR(EEXIST);
2097 				VERIFY(arc_buf_remove_ref(abuf, &abuf));
2098 			}
2099 
2100 			ddt_enter(ddt);
2101 			return (error != 0);
2102 		}
2103 	}
2104 
2105 	return (B_FALSE);
2106 }
2107 
2108 static void
2109 zio_ddt_child_write_ready(zio_t *zio)
2110 {
2111 	int p = zio->io_prop.zp_copies;
2112 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2113 	ddt_entry_t *dde = zio->io_private;
2114 	ddt_phys_t *ddp = &dde->dde_phys[p];
2115 	zio_t *pio;
2116 
2117 	if (zio->io_error)
2118 		return;
2119 
2120 	ddt_enter(ddt);
2121 
2122 	ASSERT(dde->dde_lead_zio[p] == zio);
2123 
2124 	ddt_phys_fill(ddp, zio->io_bp);
2125 
2126 	while ((pio = zio_walk_parents(zio)) != NULL)
2127 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2128 
2129 	ddt_exit(ddt);
2130 }
2131 
2132 static void
2133 zio_ddt_child_write_done(zio_t *zio)
2134 {
2135 	int p = zio->io_prop.zp_copies;
2136 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2137 	ddt_entry_t *dde = zio->io_private;
2138 	ddt_phys_t *ddp = &dde->dde_phys[p];
2139 
2140 	ddt_enter(ddt);
2141 
2142 	ASSERT(ddp->ddp_refcnt == 0);
2143 	ASSERT(dde->dde_lead_zio[p] == zio);
2144 	dde->dde_lead_zio[p] = NULL;
2145 
2146 	if (zio->io_error == 0) {
2147 		while (zio_walk_parents(zio) != NULL)
2148 			ddt_phys_addref(ddp);
2149 	} else {
2150 		ddt_phys_clear(ddp);
2151 	}
2152 
2153 	ddt_exit(ddt);
2154 }
2155 
2156 static void
2157 zio_ddt_ditto_write_done(zio_t *zio)
2158 {
2159 	int p = DDT_PHYS_DITTO;
2160 	zio_prop_t *zp = &zio->io_prop;
2161 	blkptr_t *bp = zio->io_bp;
2162 	ddt_t *ddt = ddt_select(zio->io_spa, bp);
2163 	ddt_entry_t *dde = zio->io_private;
2164 	ddt_phys_t *ddp = &dde->dde_phys[p];
2165 	ddt_key_t *ddk = &dde->dde_key;
2166 
2167 	ddt_enter(ddt);
2168 
2169 	ASSERT(ddp->ddp_refcnt == 0);
2170 	ASSERT(dde->dde_lead_zio[p] == zio);
2171 	dde->dde_lead_zio[p] = NULL;
2172 
2173 	if (zio->io_error == 0) {
2174 		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2175 		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2176 		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2177 		if (ddp->ddp_phys_birth != 0)
2178 			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2179 		ddt_phys_fill(ddp, bp);
2180 	}
2181 
2182 	ddt_exit(ddt);
2183 }
2184 
2185 static int
2186 zio_ddt_write(zio_t *zio)
2187 {
2188 	spa_t *spa = zio->io_spa;
2189 	blkptr_t *bp = zio->io_bp;
2190 	uint64_t txg = zio->io_txg;
2191 	zio_prop_t *zp = &zio->io_prop;
2192 	int p = zp->zp_copies;
2193 	int ditto_copies;
2194 	zio_t *cio = NULL;
2195 	zio_t *dio = NULL;
2196 	ddt_t *ddt = ddt_select(spa, bp);
2197 	ddt_entry_t *dde;
2198 	ddt_phys_t *ddp;
2199 
2200 	ASSERT(BP_GET_DEDUP(bp));
2201 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2202 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2203 
2204 	ddt_enter(ddt);
2205 	dde = ddt_lookup(ddt, bp, B_TRUE);
2206 	ddp = &dde->dde_phys[p];
2207 
2208 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2209 		/*
2210 		 * If we're using a weak checksum, upgrade to a strong checksum
2211 		 * and try again.  If we're already using a strong checksum,
2212 		 * we can't resolve it, so just convert to an ordinary write.
2213 		 * (And automatically e-mail a paper to Nature?)
2214 		 */
2215 		if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2216 			zp->zp_checksum = spa_dedup_checksum(spa);
2217 			zio_pop_transforms(zio);
2218 			zio->io_stage = ZIO_STAGE_OPEN;
2219 			BP_ZERO(bp);
2220 		} else {
2221 			zp->zp_dedup = B_FALSE;
2222 		}
2223 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2224 		ddt_exit(ddt);
2225 		return (ZIO_PIPELINE_CONTINUE);
2226 	}
2227 
2228 	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2229 	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2230 
2231 	if (ditto_copies > ddt_ditto_copies_present(dde) &&
2232 	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2233 		zio_prop_t czp = *zp;
2234 
2235 		czp.zp_copies = ditto_copies;
2236 
2237 		/*
2238 		 * If we arrived here with an override bp, we won't have run
2239 		 * the transform stack, so we won't have the data we need to
2240 		 * generate a child i/o.  So, toss the override bp and restart.
2241 		 * This is safe, because using the override bp is just an
2242 		 * optimization; and it's rare, so the cost doesn't matter.
2243 		 */
2244 		if (zio->io_bp_override) {
2245 			zio_pop_transforms(zio);
2246 			zio->io_stage = ZIO_STAGE_OPEN;
2247 			zio->io_pipeline = ZIO_WRITE_PIPELINE;
2248 			zio->io_bp_override = NULL;
2249 			BP_ZERO(bp);
2250 			ddt_exit(ddt);
2251 			return (ZIO_PIPELINE_CONTINUE);
2252 		}
2253 
2254 		dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2255 		    zio->io_orig_size, &czp, NULL, NULL,
2256 		    zio_ddt_ditto_write_done, dde, zio->io_priority,
2257 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2258 
2259 		zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2260 		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2261 	}
2262 
2263 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2264 		if (ddp->ddp_phys_birth != 0)
2265 			ddt_bp_fill(ddp, bp, txg);
2266 		if (dde->dde_lead_zio[p] != NULL)
2267 			zio_add_child(zio, dde->dde_lead_zio[p]);
2268 		else
2269 			ddt_phys_addref(ddp);
2270 	} else if (zio->io_bp_override) {
2271 		ASSERT(bp->blk_birth == txg);
2272 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2273 		ddt_phys_fill(ddp, bp);
2274 		ddt_phys_addref(ddp);
2275 	} else {
2276 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2277 		    zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2278 		    zio_ddt_child_write_done, dde, zio->io_priority,
2279 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2280 
2281 		zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2282 		dde->dde_lead_zio[p] = cio;
2283 	}
2284 
2285 	ddt_exit(ddt);
2286 
2287 	if (cio)
2288 		zio_nowait(cio);
2289 	if (dio)
2290 		zio_nowait(dio);
2291 
2292 	return (ZIO_PIPELINE_CONTINUE);
2293 }
2294 
2295 ddt_entry_t *freedde; /* for debugging */
2296 
2297 static int
2298 zio_ddt_free(zio_t *zio)
2299 {
2300 	spa_t *spa = zio->io_spa;
2301 	blkptr_t *bp = zio->io_bp;
2302 	ddt_t *ddt = ddt_select(spa, bp);
2303 	ddt_entry_t *dde;
2304 	ddt_phys_t *ddp;
2305 
2306 	ASSERT(BP_GET_DEDUP(bp));
2307 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2308 
2309 	ddt_enter(ddt);
2310 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2311 	ddp = ddt_phys_select(dde, bp);
2312 	ddt_phys_decref(ddp);
2313 	ddt_exit(ddt);
2314 
2315 	return (ZIO_PIPELINE_CONTINUE);
2316 }
2317 
2318 /*
2319  * ==========================================================================
2320  * Allocate and free blocks
2321  * ==========================================================================
2322  */
2323 static int
2324 zio_dva_allocate(zio_t *zio)
2325 {
2326 	spa_t *spa = zio->io_spa;
2327 	metaslab_class_t *mc = spa_normal_class(spa);
2328 	blkptr_t *bp = zio->io_bp;
2329 	int error;
2330 	int flags = 0;
2331 
2332 	if (zio->io_gang_leader == NULL) {
2333 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2334 		zio->io_gang_leader = zio;
2335 	}
2336 
2337 	ASSERT(BP_IS_HOLE(bp));
2338 	ASSERT0(BP_GET_NDVAS(bp));
2339 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
2340 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2341 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2342 
2343 	/*
2344 	 * The dump device does not support gang blocks so allocation on
2345 	 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2346 	 * the "fast" gang feature.
2347 	 */
2348 	flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2349 	flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2350 	    METASLAB_GANG_CHILD : 0;
2351 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
2352 	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2353 
2354 	if (error) {
2355 		spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2356 		    "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2357 		    error);
2358 		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2359 			return (zio_write_gang_block(zio));
2360 		zio->io_error = error;
2361 	}
2362 
2363 	return (ZIO_PIPELINE_CONTINUE);
2364 }
2365 
2366 static int
2367 zio_dva_free(zio_t *zio)
2368 {
2369 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2370 
2371 	return (ZIO_PIPELINE_CONTINUE);
2372 }
2373 
2374 static int
2375 zio_dva_claim(zio_t *zio)
2376 {
2377 	int error;
2378 
2379 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2380 	if (error)
2381 		zio->io_error = error;
2382 
2383 	return (ZIO_PIPELINE_CONTINUE);
2384 }
2385 
2386 /*
2387  * Undo an allocation.  This is used by zio_done() when an I/O fails
2388  * and we want to give back the block we just allocated.
2389  * This handles both normal blocks and gang blocks.
2390  */
2391 static void
2392 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2393 {
2394 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2395 	ASSERT(zio->io_bp_override == NULL);
2396 
2397 	if (!BP_IS_HOLE(bp))
2398 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2399 
2400 	if (gn != NULL) {
2401 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2402 			zio_dva_unallocate(zio, gn->gn_child[g],
2403 			    &gn->gn_gbh->zg_blkptr[g]);
2404 		}
2405 	}
2406 }
2407 
2408 /*
2409  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
2410  */
2411 int
2412 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2413     uint64_t size, boolean_t use_slog)
2414 {
2415 	int error = 1;
2416 
2417 	ASSERT(txg > spa_syncing_txg(spa));
2418 
2419 	/*
2420 	 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2421 	 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2422 	 * when allocating them.
2423 	 */
2424 	if (use_slog) {
2425 		error = metaslab_alloc(spa, spa_log_class(spa), size,
2426 		    new_bp, 1, txg, old_bp,
2427 		    METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2428 	}
2429 
2430 	if (error) {
2431 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
2432 		    new_bp, 1, txg, old_bp,
2433 		    METASLAB_HINTBP_AVOID);
2434 	}
2435 
2436 	if (error == 0) {
2437 		BP_SET_LSIZE(new_bp, size);
2438 		BP_SET_PSIZE(new_bp, size);
2439 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2440 		BP_SET_CHECKSUM(new_bp,
2441 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2442 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2443 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2444 		BP_SET_LEVEL(new_bp, 0);
2445 		BP_SET_DEDUP(new_bp, 0);
2446 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2447 	}
2448 
2449 	return (error);
2450 }
2451 
2452 /*
2453  * Free an intent log block.
2454  */
2455 void
2456 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2457 {
2458 	ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2459 	ASSERT(!BP_IS_GANG(bp));
2460 
2461 	zio_free(spa, txg, bp);
2462 }
2463 
2464 /*
2465  * ==========================================================================
2466  * Read and write to physical devices
2467  * ==========================================================================
2468  */
2469 static int
2470 zio_vdev_io_start(zio_t *zio)
2471 {
2472 	vdev_t *vd = zio->io_vd;
2473 	uint64_t align;
2474 	spa_t *spa = zio->io_spa;
2475 
2476 	ASSERT(zio->io_error == 0);
2477 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2478 
2479 	if (vd == NULL) {
2480 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2481 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2482 
2483 		/*
2484 		 * The mirror_ops handle multiple DVAs in a single BP.
2485 		 */
2486 		return (vdev_mirror_ops.vdev_op_io_start(zio));
2487 	}
2488 
2489 	/*
2490 	 * We keep track of time-sensitive I/Os so that the scan thread
2491 	 * can quickly react to certain workloads.  In particular, we care
2492 	 * about non-scrubbing, top-level reads and writes with the following
2493 	 * characteristics:
2494 	 * 	- synchronous writes of user data to non-slog devices
2495 	 *	- any reads of user data
2496 	 * When these conditions are met, adjust the timestamp of spa_last_io
2497 	 * which allows the scan thread to adjust its workload accordingly.
2498 	 */
2499 	if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2500 	    vd == vd->vdev_top && !vd->vdev_islog &&
2501 	    zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2502 	    zio->io_txg != spa_syncing_txg(spa)) {
2503 		uint64_t old = spa->spa_last_io;
2504 		uint64_t new = ddi_get_lbolt64();
2505 		if (old != new)
2506 			(void) atomic_cas_64(&spa->spa_last_io, old, new);
2507 	}
2508 
2509 	align = 1ULL << vd->vdev_top->vdev_ashift;
2510 
2511 	if (P2PHASE(zio->io_size, align) != 0) {
2512 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
2513 		char *abuf = zio_buf_alloc(asize);
2514 		ASSERT(vd == vd->vdev_top);
2515 		if (zio->io_type == ZIO_TYPE_WRITE) {
2516 			bcopy(zio->io_data, abuf, zio->io_size);
2517 			bzero(abuf + zio->io_size, asize - zio->io_size);
2518 		}
2519 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2520 	}
2521 
2522 	ASSERT(P2PHASE(zio->io_offset, align) == 0);
2523 	ASSERT(P2PHASE(zio->io_size, align) == 0);
2524 	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2525 
2526 	/*
2527 	 * If this is a repair I/O, and there's no self-healing involved --
2528 	 * that is, we're just resilvering what we expect to resilver --
2529 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2530 	 * This prevents spurious resilvering with nested replication.
2531 	 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2532 	 * A is out of date, we'll read from C+D, then use the data to
2533 	 * resilver A+B -- but we don't actually want to resilver B, just A.
2534 	 * The top-level mirror has no way to know this, so instead we just
2535 	 * discard unnecessary repairs as we work our way down the vdev tree.
2536 	 * The same logic applies to any form of nested replication:
2537 	 * ditto + mirror, RAID-Z + replacing, etc.  This covers them all.
2538 	 */
2539 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2540 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2541 	    zio->io_txg != 0 &&	/* not a delegated i/o */
2542 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2543 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2544 		zio_vdev_io_bypass(zio);
2545 		return (ZIO_PIPELINE_CONTINUE);
2546 	}
2547 
2548 	if (vd->vdev_ops->vdev_op_leaf &&
2549 	    (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2550 
2551 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
2552 			return (ZIO_PIPELINE_CONTINUE);
2553 
2554 		if ((zio = vdev_queue_io(zio)) == NULL)
2555 			return (ZIO_PIPELINE_STOP);
2556 
2557 		if (!vdev_accessible(vd, zio)) {
2558 			zio->io_error = SET_ERROR(ENXIO);
2559 			zio_interrupt(zio);
2560 			return (ZIO_PIPELINE_STOP);
2561 		}
2562 	}
2563 
2564 	return (vd->vdev_ops->vdev_op_io_start(zio));
2565 }
2566 
2567 static int
2568 zio_vdev_io_done(zio_t *zio)
2569 {
2570 	vdev_t *vd = zio->io_vd;
2571 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2572 	boolean_t unexpected_error = B_FALSE;
2573 
2574 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2575 		return (ZIO_PIPELINE_STOP);
2576 
2577 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2578 
2579 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2580 
2581 		vdev_queue_io_done(zio);
2582 
2583 		if (zio->io_type == ZIO_TYPE_WRITE)
2584 			vdev_cache_write(zio);
2585 
2586 		if (zio_injection_enabled && zio->io_error == 0)
2587 			zio->io_error = zio_handle_device_injection(vd,
2588 			    zio, EIO);
2589 
2590 		if (zio_injection_enabled && zio->io_error == 0)
2591 			zio->io_error = zio_handle_label_injection(zio, EIO);
2592 
2593 		if (zio->io_error) {
2594 			if (!vdev_accessible(vd, zio)) {
2595 				zio->io_error = SET_ERROR(ENXIO);
2596 			} else {
2597 				unexpected_error = B_TRUE;
2598 			}
2599 		}
2600 	}
2601 
2602 	ops->vdev_op_io_done(zio);
2603 
2604 	if (unexpected_error)
2605 		VERIFY(vdev_probe(vd, zio) == NULL);
2606 
2607 	return (ZIO_PIPELINE_CONTINUE);
2608 }
2609 
2610 /*
2611  * For non-raidz ZIOs, we can just copy aside the bad data read from the
2612  * disk, and use that to finish the checksum ereport later.
2613  */
2614 static void
2615 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2616     const void *good_buf)
2617 {
2618 	/* no processing needed */
2619 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2620 }
2621 
2622 /*ARGSUSED*/
2623 void
2624 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2625 {
2626 	void *buf = zio_buf_alloc(zio->io_size);
2627 
2628 	bcopy(zio->io_data, buf, zio->io_size);
2629 
2630 	zcr->zcr_cbinfo = zio->io_size;
2631 	zcr->zcr_cbdata = buf;
2632 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
2633 	zcr->zcr_free = zio_buf_free;
2634 }
2635 
2636 static int
2637 zio_vdev_io_assess(zio_t *zio)
2638 {
2639 	vdev_t *vd = zio->io_vd;
2640 
2641 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2642 		return (ZIO_PIPELINE_STOP);
2643 
2644 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2645 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2646 
2647 	if (zio->io_vsd != NULL) {
2648 		zio->io_vsd_ops->vsd_free(zio);
2649 		zio->io_vsd = NULL;
2650 	}
2651 
2652 	if (zio_injection_enabled && zio->io_error == 0)
2653 		zio->io_error = zio_handle_fault_injection(zio, EIO);
2654 
2655 	/*
2656 	 * If the I/O failed, determine whether we should attempt to retry it.
2657 	 *
2658 	 * On retry, we cut in line in the issue queue, since we don't want
2659 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2660 	 */
2661 	if (zio->io_error && vd == NULL &&
2662 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2663 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
2664 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
2665 		zio->io_error = 0;
2666 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
2667 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2668 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2669 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2670 		    zio_requeue_io_start_cut_in_line);
2671 		return (ZIO_PIPELINE_STOP);
2672 	}
2673 
2674 	/*
2675 	 * If we got an error on a leaf device, convert it to ENXIO
2676 	 * if the device is not accessible at all.
2677 	 */
2678 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2679 	    !vdev_accessible(vd, zio))
2680 		zio->io_error = SET_ERROR(ENXIO);
2681 
2682 	/*
2683 	 * If we can't write to an interior vdev (mirror or RAID-Z),
2684 	 * set vdev_cant_write so that we stop trying to allocate from it.
2685 	 */
2686 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2687 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2688 		vd->vdev_cant_write = B_TRUE;
2689 	}
2690 
2691 	if (zio->io_error)
2692 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2693 
2694 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2695 	    zio->io_physdone != NULL) {
2696 		ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2697 		ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2698 		zio->io_physdone(zio->io_logical);
2699 	}
2700 
2701 	return (ZIO_PIPELINE_CONTINUE);
2702 }
2703 
2704 void
2705 zio_vdev_io_reissue(zio_t *zio)
2706 {
2707 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2708 	ASSERT(zio->io_error == 0);
2709 
2710 	zio->io_stage >>= 1;
2711 }
2712 
2713 void
2714 zio_vdev_io_redone(zio_t *zio)
2715 {
2716 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2717 
2718 	zio->io_stage >>= 1;
2719 }
2720 
2721 void
2722 zio_vdev_io_bypass(zio_t *zio)
2723 {
2724 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2725 	ASSERT(zio->io_error == 0);
2726 
2727 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2728 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2729 }
2730 
2731 /*
2732  * ==========================================================================
2733  * Generate and verify checksums
2734  * ==========================================================================
2735  */
2736 static int
2737 zio_checksum_generate(zio_t *zio)
2738 {
2739 	blkptr_t *bp = zio->io_bp;
2740 	enum zio_checksum checksum;
2741 
2742 	if (bp == NULL) {
2743 		/*
2744 		 * This is zio_write_phys().
2745 		 * We're either generating a label checksum, or none at all.
2746 		 */
2747 		checksum = zio->io_prop.zp_checksum;
2748 
2749 		if (checksum == ZIO_CHECKSUM_OFF)
2750 			return (ZIO_PIPELINE_CONTINUE);
2751 
2752 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2753 	} else {
2754 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2755 			ASSERT(!IO_IS_ALLOCATING(zio));
2756 			checksum = ZIO_CHECKSUM_GANG_HEADER;
2757 		} else {
2758 			checksum = BP_GET_CHECKSUM(bp);
2759 		}
2760 	}
2761 
2762 	zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2763 
2764 	return (ZIO_PIPELINE_CONTINUE);
2765 }
2766 
2767 static int
2768 zio_checksum_verify(zio_t *zio)
2769 {
2770 	zio_bad_cksum_t info;
2771 	blkptr_t *bp = zio->io_bp;
2772 	int error;
2773 
2774 	ASSERT(zio->io_vd != NULL);
2775 
2776 	if (bp == NULL) {
2777 		/*
2778 		 * This is zio_read_phys().
2779 		 * We're either verifying a label checksum, or nothing at all.
2780 		 */
2781 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2782 			return (ZIO_PIPELINE_CONTINUE);
2783 
2784 		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2785 	}
2786 
2787 	if ((error = zio_checksum_error(zio, &info)) != 0) {
2788 		zio->io_error = error;
2789 		if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2790 			zfs_ereport_start_checksum(zio->io_spa,
2791 			    zio->io_vd, zio, zio->io_offset,
2792 			    zio->io_size, NULL, &info);
2793 		}
2794 	}
2795 
2796 	return (ZIO_PIPELINE_CONTINUE);
2797 }
2798 
2799 /*
2800  * Called by RAID-Z to ensure we don't compute the checksum twice.
2801  */
2802 void
2803 zio_checksum_verified(zio_t *zio)
2804 {
2805 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2806 }
2807 
2808 /*
2809  * ==========================================================================
2810  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2811  * An error of 0 indicates success.  ENXIO indicates whole-device failure,
2812  * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
2813  * indicate errors that are specific to one I/O, and most likely permanent.
2814  * Any other error is presumed to be worse because we weren't expecting it.
2815  * ==========================================================================
2816  */
2817 int
2818 zio_worst_error(int e1, int e2)
2819 {
2820 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2821 	int r1, r2;
2822 
2823 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2824 		if (e1 == zio_error_rank[r1])
2825 			break;
2826 
2827 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2828 		if (e2 == zio_error_rank[r2])
2829 			break;
2830 
2831 	return (r1 > r2 ? e1 : e2);
2832 }
2833 
2834 /*
2835  * ==========================================================================
2836  * I/O completion
2837  * ==========================================================================
2838  */
2839 static int
2840 zio_ready(zio_t *zio)
2841 {
2842 	blkptr_t *bp = zio->io_bp;
2843 	zio_t *pio, *pio_next;
2844 
2845 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2846 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2847 		return (ZIO_PIPELINE_STOP);
2848 
2849 	if (zio->io_ready) {
2850 		ASSERT(IO_IS_ALLOCATING(zio));
2851 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2852 		    (zio->io_flags & ZIO_FLAG_NOPWRITE));
2853 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2854 
2855 		zio->io_ready(zio);
2856 	}
2857 
2858 	if (bp != NULL && bp != &zio->io_bp_copy)
2859 		zio->io_bp_copy = *bp;
2860 
2861 	if (zio->io_error)
2862 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2863 
2864 	mutex_enter(&zio->io_lock);
2865 	zio->io_state[ZIO_WAIT_READY] = 1;
2866 	pio = zio_walk_parents(zio);
2867 	mutex_exit(&zio->io_lock);
2868 
2869 	/*
2870 	 * As we notify zio's parents, new parents could be added.
2871 	 * New parents go to the head of zio's io_parent_list, however,
2872 	 * so we will (correctly) not notify them.  The remainder of zio's
2873 	 * io_parent_list, from 'pio_next' onward, cannot change because
2874 	 * all parents must wait for us to be done before they can be done.
2875 	 */
2876 	for (; pio != NULL; pio = pio_next) {
2877 		pio_next = zio_walk_parents(zio);
2878 		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2879 	}
2880 
2881 	if (zio->io_flags & ZIO_FLAG_NODATA) {
2882 		if (BP_IS_GANG(bp)) {
2883 			zio->io_flags &= ~ZIO_FLAG_NODATA;
2884 		} else {
2885 			ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2886 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2887 		}
2888 	}
2889 
2890 	if (zio_injection_enabled &&
2891 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
2892 		zio_handle_ignored_writes(zio);
2893 
2894 	return (ZIO_PIPELINE_CONTINUE);
2895 }
2896 
2897 static int
2898 zio_done(zio_t *zio)
2899 {
2900 	spa_t *spa = zio->io_spa;
2901 	zio_t *lio = zio->io_logical;
2902 	blkptr_t *bp = zio->io_bp;
2903 	vdev_t *vd = zio->io_vd;
2904 	uint64_t psize = zio->io_size;
2905 	zio_t *pio, *pio_next;
2906 
2907 	/*
2908 	 * If our children haven't all completed,
2909 	 * wait for them and then repeat this pipeline stage.
2910 	 */
2911 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2912 	    zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2913 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2914 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2915 		return (ZIO_PIPELINE_STOP);
2916 
2917 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2918 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2919 			ASSERT(zio->io_children[c][w] == 0);
2920 
2921 	if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
2922 		ASSERT(bp->blk_pad[0] == 0);
2923 		ASSERT(bp->blk_pad[1] == 0);
2924 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2925 		    (bp == zio_unique_parent(zio)->io_bp));
2926 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2927 		    zio->io_bp_override == NULL &&
2928 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2929 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
2930 			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2931 			ASSERT(BP_COUNT_GANG(bp) == 0 ||
2932 			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2933 		}
2934 		if (zio->io_flags & ZIO_FLAG_NOPWRITE)
2935 			VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
2936 	}
2937 
2938 	/*
2939 	 * If there were child vdev/gang/ddt errors, they apply to us now.
2940 	 */
2941 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2942 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2943 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2944 
2945 	/*
2946 	 * If the I/O on the transformed data was successful, generate any
2947 	 * checksum reports now while we still have the transformed data.
2948 	 */
2949 	if (zio->io_error == 0) {
2950 		while (zio->io_cksum_report != NULL) {
2951 			zio_cksum_report_t *zcr = zio->io_cksum_report;
2952 			uint64_t align = zcr->zcr_align;
2953 			uint64_t asize = P2ROUNDUP(psize, align);
2954 			char *abuf = zio->io_data;
2955 
2956 			if (asize != psize) {
2957 				abuf = zio_buf_alloc(asize);
2958 				bcopy(zio->io_data, abuf, psize);
2959 				bzero(abuf + psize, asize - psize);
2960 			}
2961 
2962 			zio->io_cksum_report = zcr->zcr_next;
2963 			zcr->zcr_next = NULL;
2964 			zcr->zcr_finish(zcr, abuf);
2965 			zfs_ereport_free_checksum(zcr);
2966 
2967 			if (asize != psize)
2968 				zio_buf_free(abuf, asize);
2969 		}
2970 	}
2971 
2972 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
2973 
2974 	vdev_stat_update(zio, psize);
2975 
2976 	if (zio->io_error) {
2977 		/*
2978 		 * If this I/O is attached to a particular vdev,
2979 		 * generate an error message describing the I/O failure
2980 		 * at the block level.  We ignore these errors if the
2981 		 * device is currently unavailable.
2982 		 */
2983 		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2984 			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2985 
2986 		if ((zio->io_error == EIO || !(zio->io_flags &
2987 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2988 		    zio == lio) {
2989 			/*
2990 			 * For logical I/O requests, tell the SPA to log the
2991 			 * error and generate a logical data ereport.
2992 			 */
2993 			spa_log_error(spa, zio);
2994 			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2995 			    0, 0);
2996 		}
2997 	}
2998 
2999 	if (zio->io_error && zio == lio) {
3000 		/*
3001 		 * Determine whether zio should be reexecuted.  This will
3002 		 * propagate all the way to the root via zio_notify_parent().
3003 		 */
3004 		ASSERT(vd == NULL && bp != NULL);
3005 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3006 
3007 		if (IO_IS_ALLOCATING(zio) &&
3008 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3009 			if (zio->io_error != ENOSPC)
3010 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3011 			else
3012 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3013 		}
3014 
3015 		if ((zio->io_type == ZIO_TYPE_READ ||
3016 		    zio->io_type == ZIO_TYPE_FREE) &&
3017 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3018 		    zio->io_error == ENXIO &&
3019 		    spa_load_state(spa) == SPA_LOAD_NONE &&
3020 		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3021 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3022 
3023 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3024 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3025 
3026 		/*
3027 		 * Here is a possibly good place to attempt to do
3028 		 * either combinatorial reconstruction or error correction
3029 		 * based on checksums.  It also might be a good place
3030 		 * to send out preliminary ereports before we suspend
3031 		 * processing.
3032 		 */
3033 	}
3034 
3035 	/*
3036 	 * If there were logical child errors, they apply to us now.
3037 	 * We defer this until now to avoid conflating logical child
3038 	 * errors with errors that happened to the zio itself when
3039 	 * updating vdev stats and reporting FMA events above.
3040 	 */
3041 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3042 
3043 	if ((zio->io_error || zio->io_reexecute) &&
3044 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3045 	    !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3046 		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3047 
3048 	zio_gang_tree_free(&zio->io_gang_tree);
3049 
3050 	/*
3051 	 * Godfather I/Os should never suspend.
3052 	 */
3053 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3054 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3055 		zio->io_reexecute = 0;
3056 
3057 	if (zio->io_reexecute) {
3058 		/*
3059 		 * This is a logical I/O that wants to reexecute.
3060 		 *
3061 		 * Reexecute is top-down.  When an i/o fails, if it's not
3062 		 * the root, it simply notifies its parent and sticks around.
3063 		 * The parent, seeing that it still has children in zio_done(),
3064 		 * does the same.  This percolates all the way up to the root.
3065 		 * The root i/o will reexecute or suspend the entire tree.
3066 		 *
3067 		 * This approach ensures that zio_reexecute() honors
3068 		 * all the original i/o dependency relationships, e.g.
3069 		 * parents not executing until children are ready.
3070 		 */
3071 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3072 
3073 		zio->io_gang_leader = NULL;
3074 
3075 		mutex_enter(&zio->io_lock);
3076 		zio->io_state[ZIO_WAIT_DONE] = 1;
3077 		mutex_exit(&zio->io_lock);
3078 
3079 		/*
3080 		 * "The Godfather" I/O monitors its children but is
3081 		 * not a true parent to them. It will track them through
3082 		 * the pipeline but severs its ties whenever they get into
3083 		 * trouble (e.g. suspended). This allows "The Godfather"
3084 		 * I/O to return status without blocking.
3085 		 */
3086 		for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3087 			zio_link_t *zl = zio->io_walk_link;
3088 			pio_next = zio_walk_parents(zio);
3089 
3090 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3091 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3092 				zio_remove_child(pio, zio, zl);
3093 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3094 			}
3095 		}
3096 
3097 		if ((pio = zio_unique_parent(zio)) != NULL) {
3098 			/*
3099 			 * We're not a root i/o, so there's nothing to do
3100 			 * but notify our parent.  Don't propagate errors
3101 			 * upward since we haven't permanently failed yet.
3102 			 */
3103 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3104 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3105 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3106 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3107 			/*
3108 			 * We'd fail again if we reexecuted now, so suspend
3109 			 * until conditions improve (e.g. device comes online).
3110 			 */
3111 			zio_suspend(spa, zio);
3112 		} else {
3113 			/*
3114 			 * Reexecution is potentially a huge amount of work.
3115 			 * Hand it off to the otherwise-unused claim taskq.
3116 			 */
3117 			ASSERT(zio->io_tqent.tqent_next == NULL);
3118 			spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3119 			    ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3120 			    0, &zio->io_tqent);
3121 		}
3122 		return (ZIO_PIPELINE_STOP);
3123 	}
3124 
3125 	ASSERT(zio->io_child_count == 0);
3126 	ASSERT(zio->io_reexecute == 0);
3127 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3128 
3129 	/*
3130 	 * Report any checksum errors, since the I/O is complete.
3131 	 */
3132 	while (zio->io_cksum_report != NULL) {
3133 		zio_cksum_report_t *zcr = zio->io_cksum_report;
3134 		zio->io_cksum_report = zcr->zcr_next;
3135 		zcr->zcr_next = NULL;
3136 		zcr->zcr_finish(zcr, NULL);
3137 		zfs_ereport_free_checksum(zcr);
3138 	}
3139 
3140 	/*
3141 	 * It is the responsibility of the done callback to ensure that this
3142 	 * particular zio is no longer discoverable for adoption, and as
3143 	 * such, cannot acquire any new parents.
3144 	 */
3145 	if (zio->io_done)
3146 		zio->io_done(zio);
3147 
3148 	mutex_enter(&zio->io_lock);
3149 	zio->io_state[ZIO_WAIT_DONE] = 1;
3150 	mutex_exit(&zio->io_lock);
3151 
3152 	for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3153 		zio_link_t *zl = zio->io_walk_link;
3154 		pio_next = zio_walk_parents(zio);
3155 		zio_remove_child(pio, zio, zl);
3156 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3157 	}
3158 
3159 	if (zio->io_waiter != NULL) {
3160 		mutex_enter(&zio->io_lock);
3161 		zio->io_executor = NULL;
3162 		cv_broadcast(&zio->io_cv);
3163 		mutex_exit(&zio->io_lock);
3164 	} else {
3165 		zio_destroy(zio);
3166 	}
3167 
3168 	return (ZIO_PIPELINE_STOP);
3169 }
3170 
3171 /*
3172  * ==========================================================================
3173  * I/O pipeline definition
3174  * ==========================================================================
3175  */
3176 static zio_pipe_stage_t *zio_pipeline[] = {
3177 	NULL,
3178 	zio_read_bp_init,
3179 	zio_free_bp_init,
3180 	zio_issue_async,
3181 	zio_write_bp_init,
3182 	zio_checksum_generate,
3183 	zio_nop_write,
3184 	zio_ddt_read_start,
3185 	zio_ddt_read_done,
3186 	zio_ddt_write,
3187 	zio_ddt_free,
3188 	zio_gang_assemble,
3189 	zio_gang_issue,
3190 	zio_dva_allocate,
3191 	zio_dva_free,
3192 	zio_dva_claim,
3193 	zio_ready,
3194 	zio_vdev_io_start,
3195 	zio_vdev_io_done,
3196 	zio_vdev_io_assess,
3197 	zio_checksum_verify,
3198 	zio_done
3199 };
3200 
3201 /* dnp is the dnode for zb1->zb_object */
3202 boolean_t
3203 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3204     const zbookmark_t *zb2)
3205 {
3206 	uint64_t zb1nextL0, zb2thisobj;
3207 
3208 	ASSERT(zb1->zb_objset == zb2->zb_objset);
3209 	ASSERT(zb2->zb_level == 0);
3210 
3211 	/*
3212 	 * A bookmark in the deadlist is considered to be after
3213 	 * everything else.
3214 	 */
3215 	if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3216 		return (B_TRUE);
3217 
3218 	/* The objset_phys_t isn't before anything. */
3219 	if (dnp == NULL)
3220 		return (B_FALSE);
3221 
3222 	zb1nextL0 = (zb1->zb_blkid + 1) <<
3223 	    ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3224 
3225 	zb2thisobj = zb2->zb_object ? zb2->zb_object :
3226 	    zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3227 
3228 	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3229 		uint64_t nextobj = zb1nextL0 *
3230 		    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3231 		return (nextobj <= zb2thisobj);
3232 	}
3233 
3234 	if (zb1->zb_object < zb2thisobj)
3235 		return (B_TRUE);
3236 	if (zb1->zb_object > zb2thisobj)
3237 		return (B_FALSE);
3238 	if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3239 		return (B_FALSE);
3240 	return (zb1nextL0 <= zb2->zb_blkid);
3241 }
3242