xref: /titanic_41/usr/src/uts/common/fs/zfs/zio.c (revision 430b4c467020edf2445feb0c21db01c88b86243a)
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 by Delphix. All rights reserved.
24  */
25 
26 #include <sys/zfs_context.h>
27 #include <sys/fm/fs/zfs.h>
28 #include <sys/spa.h>
29 #include <sys/txg.h>
30 #include <sys/spa_impl.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/zio_impl.h>
33 #include <sys/zio_compress.h>
34 #include <sys/zio_checksum.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/arc.h>
37 #include <sys/ddt.h>
38 
39 /*
40  * ==========================================================================
41  * I/O priority table
42  * ==========================================================================
43  */
44 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
45 	0,	/* ZIO_PRIORITY_NOW		*/
46 	0,	/* ZIO_PRIORITY_SYNC_READ	*/
47 	0,	/* ZIO_PRIORITY_SYNC_WRITE	*/
48 	0,	/* ZIO_PRIORITY_LOG_WRITE	*/
49 	1,	/* ZIO_PRIORITY_CACHE_FILL	*/
50 	1,	/* ZIO_PRIORITY_AGG		*/
51 	4,	/* ZIO_PRIORITY_FREE		*/
52 	4,	/* ZIO_PRIORITY_ASYNC_WRITE	*/
53 	6,	/* ZIO_PRIORITY_ASYNC_READ	*/
54 	10,	/* ZIO_PRIORITY_RESILVER	*/
55 	20,	/* ZIO_PRIORITY_SCRUB		*/
56 	2,	/* ZIO_PRIORITY_DDT_PREFETCH	*/
57 };
58 
59 /*
60  * ==========================================================================
61  * I/O type descriptions
62  * ==========================================================================
63  */
64 char *zio_type_name[ZIO_TYPES] = {
65 	"zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
66 	"zio_ioctl"
67 };
68 
69 /*
70  * ==========================================================================
71  * I/O kmem caches
72  * ==========================================================================
73  */
74 kmem_cache_t *zio_cache;
75 kmem_cache_t *zio_link_cache;
76 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
77 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
78 
79 #ifdef _KERNEL
80 extern vmem_t *zio_alloc_arena;
81 #endif
82 extern int zfs_mg_alloc_failures;
83 
84 /*
85  * An allocating zio is one that either currently has the DVA allocate
86  * stage set or will have it later in its lifetime.
87  */
88 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
89 
90 boolean_t	zio_requeue_io_start_cut_in_line = B_TRUE;
91 
92 #ifdef ZFS_DEBUG
93 int zio_buf_debug_limit = 16384;
94 #else
95 int zio_buf_debug_limit = 0;
96 #endif
97 
98 void
99 zio_init(void)
100 {
101 	size_t c;
102 	vmem_t *data_alloc_arena = NULL;
103 
104 #ifdef _KERNEL
105 	data_alloc_arena = zio_alloc_arena;
106 #endif
107 	zio_cache = kmem_cache_create("zio_cache",
108 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
109 	zio_link_cache = kmem_cache_create("zio_link_cache",
110 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
111 
112 	/*
113 	 * For small buffers, we want a cache for each multiple of
114 	 * SPA_MINBLOCKSIZE.  For medium-size buffers, we want a cache
115 	 * for each quarter-power of 2.  For large buffers, we want
116 	 * a cache for each multiple of PAGESIZE.
117 	 */
118 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
119 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
120 		size_t p2 = size;
121 		size_t align = 0;
122 		size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
123 
124 		while (p2 & (p2 - 1))
125 			p2 &= p2 - 1;
126 
127 		if (size <= 4 * SPA_MINBLOCKSIZE) {
128 			align = SPA_MINBLOCKSIZE;
129 		} else if (P2PHASE(size, PAGESIZE) == 0) {
130 			align = PAGESIZE;
131 		} else if (P2PHASE(size, p2 >> 2) == 0) {
132 			align = p2 >> 2;
133 		}
134 
135 		if (align != 0) {
136 			char name[36];
137 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
138 			zio_buf_cache[c] = kmem_cache_create(name, size,
139 			    align, NULL, NULL, NULL, NULL, NULL, cflags);
140 
141 			/*
142 			 * Since zio_data bufs do not appear in crash dumps, we
143 			 * pass KMC_NOTOUCH so that no allocator metadata is
144 			 * stored with the buffers.
145 			 */
146 			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
147 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
148 			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
149 			    cflags | KMC_NOTOUCH);
150 		}
151 	}
152 
153 	while (--c != 0) {
154 		ASSERT(zio_buf_cache[c] != NULL);
155 		if (zio_buf_cache[c - 1] == NULL)
156 			zio_buf_cache[c - 1] = zio_buf_cache[c];
157 
158 		ASSERT(zio_data_buf_cache[c] != NULL);
159 		if (zio_data_buf_cache[c - 1] == NULL)
160 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
161 	}
162 
163 	/*
164 	 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
165 	 * to fail 3 times per txg or 8 failures, whichever is greater.
166 	 */
167 	zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
168 
169 	zio_inject_init();
170 }
171 
172 void
173 zio_fini(void)
174 {
175 	size_t c;
176 	kmem_cache_t *last_cache = NULL;
177 	kmem_cache_t *last_data_cache = NULL;
178 
179 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
180 		if (zio_buf_cache[c] != last_cache) {
181 			last_cache = zio_buf_cache[c];
182 			kmem_cache_destroy(zio_buf_cache[c]);
183 		}
184 		zio_buf_cache[c] = NULL;
185 
186 		if (zio_data_buf_cache[c] != last_data_cache) {
187 			last_data_cache = zio_data_buf_cache[c];
188 			kmem_cache_destroy(zio_data_buf_cache[c]);
189 		}
190 		zio_data_buf_cache[c] = NULL;
191 	}
192 
193 	kmem_cache_destroy(zio_link_cache);
194 	kmem_cache_destroy(zio_cache);
195 
196 	zio_inject_fini();
197 }
198 
199 /*
200  * ==========================================================================
201  * Allocate and free I/O buffers
202  * ==========================================================================
203  */
204 
205 /*
206  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
207  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
208  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
209  * excess / transient data in-core during a crashdump.
210  */
211 void *
212 zio_buf_alloc(size_t size)
213 {
214 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
215 
216 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
217 
218 	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
219 }
220 
221 /*
222  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
223  * crashdump if the kernel panics.  This exists so that we will limit the amount
224  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
225  * of kernel heap dumped to disk when the kernel panics)
226  */
227 void *
228 zio_data_buf_alloc(size_t size)
229 {
230 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
231 
232 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
233 
234 	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
235 }
236 
237 void
238 zio_buf_free(void *buf, size_t size)
239 {
240 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
241 
242 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
243 
244 	kmem_cache_free(zio_buf_cache[c], buf);
245 }
246 
247 void
248 zio_data_buf_free(void *buf, size_t size)
249 {
250 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
251 
252 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
253 
254 	kmem_cache_free(zio_data_buf_cache[c], buf);
255 }
256 
257 /*
258  * ==========================================================================
259  * Push and pop I/O transform buffers
260  * ==========================================================================
261  */
262 static void
263 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
264 	zio_transform_func_t *transform)
265 {
266 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
267 
268 	zt->zt_orig_data = zio->io_data;
269 	zt->zt_orig_size = zio->io_size;
270 	zt->zt_bufsize = bufsize;
271 	zt->zt_transform = transform;
272 
273 	zt->zt_next = zio->io_transform_stack;
274 	zio->io_transform_stack = zt;
275 
276 	zio->io_data = data;
277 	zio->io_size = size;
278 }
279 
280 static void
281 zio_pop_transforms(zio_t *zio)
282 {
283 	zio_transform_t *zt;
284 
285 	while ((zt = zio->io_transform_stack) != NULL) {
286 		if (zt->zt_transform != NULL)
287 			zt->zt_transform(zio,
288 			    zt->zt_orig_data, zt->zt_orig_size);
289 
290 		if (zt->zt_bufsize != 0)
291 			zio_buf_free(zio->io_data, zt->zt_bufsize);
292 
293 		zio->io_data = zt->zt_orig_data;
294 		zio->io_size = zt->zt_orig_size;
295 		zio->io_transform_stack = zt->zt_next;
296 
297 		kmem_free(zt, sizeof (zio_transform_t));
298 	}
299 }
300 
301 /*
302  * ==========================================================================
303  * I/O transform callbacks for subblocks and decompression
304  * ==========================================================================
305  */
306 static void
307 zio_subblock(zio_t *zio, void *data, uint64_t size)
308 {
309 	ASSERT(zio->io_size > size);
310 
311 	if (zio->io_type == ZIO_TYPE_READ)
312 		bcopy(zio->io_data, data, size);
313 }
314 
315 static void
316 zio_decompress(zio_t *zio, void *data, uint64_t size)
317 {
318 	if (zio->io_error == 0 &&
319 	    zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
320 	    zio->io_data, data, zio->io_size, size) != 0)
321 		zio->io_error = EIO;
322 }
323 
324 /*
325  * ==========================================================================
326  * I/O parent/child relationships and pipeline interlocks
327  * ==========================================================================
328  */
329 /*
330  * NOTE - Callers to zio_walk_parents() and zio_walk_children must
331  *        continue calling these functions until they return NULL.
332  *        Otherwise, the next caller will pick up the list walk in
333  *        some indeterminate state.  (Otherwise every caller would
334  *        have to pass in a cookie to keep the state represented by
335  *        io_walk_link, which gets annoying.)
336  */
337 zio_t *
338 zio_walk_parents(zio_t *cio)
339 {
340 	zio_link_t *zl = cio->io_walk_link;
341 	list_t *pl = &cio->io_parent_list;
342 
343 	zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
344 	cio->io_walk_link = zl;
345 
346 	if (zl == NULL)
347 		return (NULL);
348 
349 	ASSERT(zl->zl_child == cio);
350 	return (zl->zl_parent);
351 }
352 
353 zio_t *
354 zio_walk_children(zio_t *pio)
355 {
356 	zio_link_t *zl = pio->io_walk_link;
357 	list_t *cl = &pio->io_child_list;
358 
359 	zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
360 	pio->io_walk_link = zl;
361 
362 	if (zl == NULL)
363 		return (NULL);
364 
365 	ASSERT(zl->zl_parent == pio);
366 	return (zl->zl_child);
367 }
368 
369 zio_t *
370 zio_unique_parent(zio_t *cio)
371 {
372 	zio_t *pio = zio_walk_parents(cio);
373 
374 	VERIFY(zio_walk_parents(cio) == NULL);
375 	return (pio);
376 }
377 
378 void
379 zio_add_child(zio_t *pio, zio_t *cio)
380 {
381 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
382 
383 	/*
384 	 * Logical I/Os can have logical, gang, or vdev children.
385 	 * Gang I/Os can have gang or vdev children.
386 	 * Vdev I/Os can only have vdev children.
387 	 * The following ASSERT captures all of these constraints.
388 	 */
389 	ASSERT(cio->io_child_type <= pio->io_child_type);
390 
391 	zl->zl_parent = pio;
392 	zl->zl_child = cio;
393 
394 	mutex_enter(&cio->io_lock);
395 	mutex_enter(&pio->io_lock);
396 
397 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
398 
399 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
400 		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
401 
402 	list_insert_head(&pio->io_child_list, zl);
403 	list_insert_head(&cio->io_parent_list, zl);
404 
405 	pio->io_child_count++;
406 	cio->io_parent_count++;
407 
408 	mutex_exit(&pio->io_lock);
409 	mutex_exit(&cio->io_lock);
410 }
411 
412 static void
413 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
414 {
415 	ASSERT(zl->zl_parent == pio);
416 	ASSERT(zl->zl_child == cio);
417 
418 	mutex_enter(&cio->io_lock);
419 	mutex_enter(&pio->io_lock);
420 
421 	list_remove(&pio->io_child_list, zl);
422 	list_remove(&cio->io_parent_list, zl);
423 
424 	pio->io_child_count--;
425 	cio->io_parent_count--;
426 
427 	mutex_exit(&pio->io_lock);
428 	mutex_exit(&cio->io_lock);
429 
430 	kmem_cache_free(zio_link_cache, zl);
431 }
432 
433 static boolean_t
434 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
435 {
436 	uint64_t *countp = &zio->io_children[child][wait];
437 	boolean_t waiting = B_FALSE;
438 
439 	mutex_enter(&zio->io_lock);
440 	ASSERT(zio->io_stall == NULL);
441 	if (*countp != 0) {
442 		zio->io_stage >>= 1;
443 		zio->io_stall = countp;
444 		waiting = B_TRUE;
445 	}
446 	mutex_exit(&zio->io_lock);
447 
448 	return (waiting);
449 }
450 
451 static void
452 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
453 {
454 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
455 	int *errorp = &pio->io_child_error[zio->io_child_type];
456 
457 	mutex_enter(&pio->io_lock);
458 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
459 		*errorp = zio_worst_error(*errorp, zio->io_error);
460 	pio->io_reexecute |= zio->io_reexecute;
461 	ASSERT3U(*countp, >, 0);
462 	if (--*countp == 0 && pio->io_stall == countp) {
463 		pio->io_stall = NULL;
464 		mutex_exit(&pio->io_lock);
465 		zio_execute(pio);
466 	} else {
467 		mutex_exit(&pio->io_lock);
468 	}
469 }
470 
471 static void
472 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
473 {
474 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
475 		zio->io_error = zio->io_child_error[c];
476 }
477 
478 /*
479  * ==========================================================================
480  * Create the various types of I/O (read, write, free, etc)
481  * ==========================================================================
482  */
483 static zio_t *
484 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
485     void *data, uint64_t size, zio_done_func_t *done, void *private,
486     zio_type_t type, int priority, enum zio_flag flags,
487     vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
488     enum zio_stage stage, enum zio_stage pipeline)
489 {
490 	zio_t *zio;
491 
492 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
493 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
494 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
495 
496 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
497 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
498 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
499 
500 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
501 	bzero(zio, sizeof (zio_t));
502 
503 	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
504 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
505 
506 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
507 	    offsetof(zio_link_t, zl_parent_node));
508 	list_create(&zio->io_child_list, sizeof (zio_link_t),
509 	    offsetof(zio_link_t, zl_child_node));
510 
511 	if (vd != NULL)
512 		zio->io_child_type = ZIO_CHILD_VDEV;
513 	else if (flags & ZIO_FLAG_GANG_CHILD)
514 		zio->io_child_type = ZIO_CHILD_GANG;
515 	else if (flags & ZIO_FLAG_DDT_CHILD)
516 		zio->io_child_type = ZIO_CHILD_DDT;
517 	else
518 		zio->io_child_type = ZIO_CHILD_LOGICAL;
519 
520 	if (bp != NULL) {
521 		zio->io_bp = (blkptr_t *)bp;
522 		zio->io_bp_copy = *bp;
523 		zio->io_bp_orig = *bp;
524 		if (type != ZIO_TYPE_WRITE ||
525 		    zio->io_child_type == ZIO_CHILD_DDT)
526 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
527 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
528 			zio->io_logical = zio;
529 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
530 			pipeline |= ZIO_GANG_STAGES;
531 	}
532 
533 	zio->io_spa = spa;
534 	zio->io_txg = txg;
535 	zio->io_done = done;
536 	zio->io_private = private;
537 	zio->io_type = type;
538 	zio->io_priority = priority;
539 	zio->io_vd = vd;
540 	zio->io_offset = offset;
541 	zio->io_orig_data = zio->io_data = data;
542 	zio->io_orig_size = zio->io_size = size;
543 	zio->io_orig_flags = zio->io_flags = flags;
544 	zio->io_orig_stage = zio->io_stage = stage;
545 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
546 
547 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
548 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
549 
550 	if (zb != NULL)
551 		zio->io_bookmark = *zb;
552 
553 	if (pio != NULL) {
554 		if (zio->io_logical == NULL)
555 			zio->io_logical = pio->io_logical;
556 		if (zio->io_child_type == ZIO_CHILD_GANG)
557 			zio->io_gang_leader = pio->io_gang_leader;
558 		zio_add_child(pio, zio);
559 	}
560 
561 	return (zio);
562 }
563 
564 static void
565 zio_destroy(zio_t *zio)
566 {
567 	list_destroy(&zio->io_parent_list);
568 	list_destroy(&zio->io_child_list);
569 	mutex_destroy(&zio->io_lock);
570 	cv_destroy(&zio->io_cv);
571 	kmem_cache_free(zio_cache, zio);
572 }
573 
574 zio_t *
575 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
576     void *private, enum zio_flag flags)
577 {
578 	zio_t *zio;
579 
580 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
581 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
582 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
583 
584 	return (zio);
585 }
586 
587 zio_t *
588 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
589 {
590 	return (zio_null(NULL, spa, NULL, done, private, flags));
591 }
592 
593 zio_t *
594 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
595     void *data, uint64_t size, zio_done_func_t *done, void *private,
596     int priority, enum zio_flag flags, const zbookmark_t *zb)
597 {
598 	zio_t *zio;
599 
600 	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
601 	    data, size, done, private,
602 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
603 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
604 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
605 
606 	return (zio);
607 }
608 
609 zio_t *
610 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
611     void *data, uint64_t size, const zio_prop_t *zp,
612     zio_done_func_t *ready, zio_done_func_t *done, void *private,
613     int priority, enum zio_flag flags, const zbookmark_t *zb)
614 {
615 	zio_t *zio;
616 
617 	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
618 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
619 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
620 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
621 	    zp->zp_type < DMU_OT_NUMTYPES &&
622 	    zp->zp_level < 32 &&
623 	    zp->zp_copies > 0 &&
624 	    zp->zp_copies <= spa_max_replication(spa) &&
625 	    zp->zp_dedup <= 1 &&
626 	    zp->zp_dedup_verify <= 1);
627 
628 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
629 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
630 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
631 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
632 
633 	zio->io_ready = ready;
634 	zio->io_prop = *zp;
635 
636 	return (zio);
637 }
638 
639 zio_t *
640 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
641     uint64_t size, zio_done_func_t *done, void *private, int priority,
642     enum zio_flag flags, zbookmark_t *zb)
643 {
644 	zio_t *zio;
645 
646 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
647 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
648 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
649 
650 	return (zio);
651 }
652 
653 void
654 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
655 {
656 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
657 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
658 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
659 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
660 
661 	zio->io_prop.zp_copies = copies;
662 	zio->io_bp_override = bp;
663 }
664 
665 void
666 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
667 {
668 	bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
669 }
670 
671 zio_t *
672 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
673     enum zio_flag flags)
674 {
675 	zio_t *zio;
676 
677 	dprintf_bp(bp, "freeing in txg %llu, pass %u",
678 	    (longlong_t)txg, spa->spa_sync_pass);
679 
680 	ASSERT(!BP_IS_HOLE(bp));
681 	ASSERT(spa_syncing_txg(spa) == txg);
682 	ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
683 
684 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
685 	    NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
686 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
687 
688 	return (zio);
689 }
690 
691 zio_t *
692 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
693     zio_done_func_t *done, void *private, enum zio_flag flags)
694 {
695 	zio_t *zio;
696 
697 	/*
698 	 * A claim is an allocation of a specific block.  Claims are needed
699 	 * to support immediate writes in the intent log.  The issue is that
700 	 * immediate writes contain committed data, but in a txg that was
701 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
702 	 * the intent log claims all blocks that contain immediate write data
703 	 * so that the SPA knows they're in use.
704 	 *
705 	 * All claims *must* be resolved in the first txg -- before the SPA
706 	 * starts allocating blocks -- so that nothing is allocated twice.
707 	 * If txg == 0 we just verify that the block is claimable.
708 	 */
709 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
710 	ASSERT(txg == spa_first_txg(spa) || txg == 0);
711 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
712 
713 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
714 	    done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
715 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
716 
717 	return (zio);
718 }
719 
720 zio_t *
721 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
722     zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
723 {
724 	zio_t *zio;
725 	int c;
726 
727 	if (vd->vdev_children == 0) {
728 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
729 		    ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
730 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
731 
732 		zio->io_cmd = cmd;
733 	} else {
734 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
735 
736 		for (c = 0; c < vd->vdev_children; c++)
737 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
738 			    done, private, priority, flags));
739 	}
740 
741 	return (zio);
742 }
743 
744 zio_t *
745 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
746     void *data, int checksum, zio_done_func_t *done, void *private,
747     int priority, enum zio_flag flags, boolean_t labels)
748 {
749 	zio_t *zio;
750 
751 	ASSERT(vd->vdev_children == 0);
752 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
753 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
754 	ASSERT3U(offset + size, <=, vd->vdev_psize);
755 
756 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
757 	    ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
758 	    ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
759 
760 	zio->io_prop.zp_checksum = checksum;
761 
762 	return (zio);
763 }
764 
765 zio_t *
766 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
767     void *data, int checksum, zio_done_func_t *done, void *private,
768     int priority, enum zio_flag flags, boolean_t labels)
769 {
770 	zio_t *zio;
771 
772 	ASSERT(vd->vdev_children == 0);
773 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
774 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
775 	ASSERT3U(offset + size, <=, vd->vdev_psize);
776 
777 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
778 	    ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
779 	    ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
780 
781 	zio->io_prop.zp_checksum = checksum;
782 
783 	if (zio_checksum_table[checksum].ci_eck) {
784 		/*
785 		 * zec checksums are necessarily destructive -- they modify
786 		 * the end of the write buffer to hold the verifier/checksum.
787 		 * Therefore, we must make a local copy in case the data is
788 		 * being written to multiple places in parallel.
789 		 */
790 		void *wbuf = zio_buf_alloc(size);
791 		bcopy(data, wbuf, size);
792 		zio_push_transform(zio, wbuf, size, size, NULL);
793 	}
794 
795 	return (zio);
796 }
797 
798 /*
799  * Create a child I/O to do some work for us.
800  */
801 zio_t *
802 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
803 	void *data, uint64_t size, int type, int priority, enum zio_flag flags,
804 	zio_done_func_t *done, void *private)
805 {
806 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
807 	zio_t *zio;
808 
809 	ASSERT(vd->vdev_parent ==
810 	    (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
811 
812 	if (type == ZIO_TYPE_READ && bp != NULL) {
813 		/*
814 		 * If we have the bp, then the child should perform the
815 		 * checksum and the parent need not.  This pushes error
816 		 * detection as close to the leaves as possible and
817 		 * eliminates redundant checksums in the interior nodes.
818 		 */
819 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
820 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
821 	}
822 
823 	if (vd->vdev_children == 0)
824 		offset += VDEV_LABEL_START_SIZE;
825 
826 	flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
827 
828 	/*
829 	 * If we've decided to do a repair, the write is not speculative --
830 	 * even if the original read was.
831 	 */
832 	if (flags & ZIO_FLAG_IO_REPAIR)
833 		flags &= ~ZIO_FLAG_SPECULATIVE;
834 
835 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
836 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
837 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
838 
839 	return (zio);
840 }
841 
842 zio_t *
843 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
844 	int type, int priority, enum zio_flag flags,
845 	zio_done_func_t *done, void *private)
846 {
847 	zio_t *zio;
848 
849 	ASSERT(vd->vdev_ops->vdev_op_leaf);
850 
851 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
852 	    data, size, done, private, type, priority,
853 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
854 	    vd, offset, NULL,
855 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
856 
857 	return (zio);
858 }
859 
860 void
861 zio_flush(zio_t *zio, vdev_t *vd)
862 {
863 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
864 	    NULL, NULL, ZIO_PRIORITY_NOW,
865 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
866 }
867 
868 void
869 zio_shrink(zio_t *zio, uint64_t size)
870 {
871 	ASSERT(zio->io_executor == NULL);
872 	ASSERT(zio->io_orig_size == zio->io_size);
873 	ASSERT(size <= zio->io_size);
874 
875 	/*
876 	 * We don't shrink for raidz because of problems with the
877 	 * reconstruction when reading back less than the block size.
878 	 * Note, BP_IS_RAIDZ() assumes no compression.
879 	 */
880 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
881 	if (!BP_IS_RAIDZ(zio->io_bp))
882 		zio->io_orig_size = zio->io_size = size;
883 }
884 
885 /*
886  * ==========================================================================
887  * Prepare to read and write logical blocks
888  * ==========================================================================
889  */
890 
891 static int
892 zio_read_bp_init(zio_t *zio)
893 {
894 	blkptr_t *bp = zio->io_bp;
895 
896 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
897 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
898 	    !(zio->io_flags & ZIO_FLAG_RAW)) {
899 		uint64_t psize = BP_GET_PSIZE(bp);
900 		void *cbuf = zio_buf_alloc(psize);
901 
902 		zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
903 	}
904 
905 	if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
906 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
907 
908 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
909 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
910 
911 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
912 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
913 
914 	return (ZIO_PIPELINE_CONTINUE);
915 }
916 
917 static int
918 zio_write_bp_init(zio_t *zio)
919 {
920 	spa_t *spa = zio->io_spa;
921 	zio_prop_t *zp = &zio->io_prop;
922 	enum zio_compress compress = zp->zp_compress;
923 	blkptr_t *bp = zio->io_bp;
924 	uint64_t lsize = zio->io_size;
925 	uint64_t psize = lsize;
926 	int pass = 1;
927 
928 	/*
929 	 * If our children haven't all reached the ready stage,
930 	 * wait for them and then repeat this pipeline stage.
931 	 */
932 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
933 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
934 		return (ZIO_PIPELINE_STOP);
935 
936 	if (!IO_IS_ALLOCATING(zio))
937 		return (ZIO_PIPELINE_CONTINUE);
938 
939 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
940 
941 	if (zio->io_bp_override) {
942 		ASSERT(bp->blk_birth != zio->io_txg);
943 		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
944 
945 		*bp = *zio->io_bp_override;
946 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
947 
948 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
949 			return (ZIO_PIPELINE_CONTINUE);
950 
951 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
952 		    zp->zp_dedup_verify);
953 
954 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
955 			BP_SET_DEDUP(bp, 1);
956 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
957 			return (ZIO_PIPELINE_CONTINUE);
958 		}
959 		zio->io_bp_override = NULL;
960 		BP_ZERO(bp);
961 	}
962 
963 	if (bp->blk_birth == zio->io_txg) {
964 		/*
965 		 * We're rewriting an existing block, which means we're
966 		 * working on behalf of spa_sync().  For spa_sync() to
967 		 * converge, it must eventually be the case that we don't
968 		 * have to allocate new blocks.  But compression changes
969 		 * the blocksize, which forces a reallocate, and makes
970 		 * convergence take longer.  Therefore, after the first
971 		 * few passes, stop compressing to ensure convergence.
972 		 */
973 		pass = spa_sync_pass(spa);
974 
975 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
976 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
977 		ASSERT(!BP_GET_DEDUP(bp));
978 
979 		if (pass > SYNC_PASS_DONT_COMPRESS)
980 			compress = ZIO_COMPRESS_OFF;
981 
982 		/* Make sure someone doesn't change their mind on overwrites */
983 		ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
984 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
985 	}
986 
987 	if (compress != ZIO_COMPRESS_OFF) {
988 		void *cbuf = zio_buf_alloc(lsize);
989 		psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
990 		if (psize == 0 || psize == lsize) {
991 			compress = ZIO_COMPRESS_OFF;
992 			zio_buf_free(cbuf, lsize);
993 		} else {
994 			ASSERT(psize < lsize);
995 			zio_push_transform(zio, cbuf, psize, lsize, NULL);
996 		}
997 	}
998 
999 	/*
1000 	 * The final pass of spa_sync() must be all rewrites, but the first
1001 	 * few passes offer a trade-off: allocating blocks defers convergence,
1002 	 * but newly allocated blocks are sequential, so they can be written
1003 	 * to disk faster.  Therefore, we allow the first few passes of
1004 	 * spa_sync() to allocate new blocks, but force rewrites after that.
1005 	 * There should only be a handful of blocks after pass 1 in any case.
1006 	 */
1007 	if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1008 	    pass > SYNC_PASS_REWRITE) {
1009 		ASSERT(psize != 0);
1010 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1011 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1012 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1013 	} else {
1014 		BP_ZERO(bp);
1015 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1016 	}
1017 
1018 	if (psize == 0) {
1019 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1020 	} else {
1021 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1022 		BP_SET_LSIZE(bp, lsize);
1023 		BP_SET_PSIZE(bp, psize);
1024 		BP_SET_COMPRESS(bp, compress);
1025 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1026 		BP_SET_TYPE(bp, zp->zp_type);
1027 		BP_SET_LEVEL(bp, zp->zp_level);
1028 		BP_SET_DEDUP(bp, zp->zp_dedup);
1029 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1030 		if (zp->zp_dedup) {
1031 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1032 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1033 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1034 		}
1035 	}
1036 
1037 	return (ZIO_PIPELINE_CONTINUE);
1038 }
1039 
1040 static int
1041 zio_free_bp_init(zio_t *zio)
1042 {
1043 	blkptr_t *bp = zio->io_bp;
1044 
1045 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1046 		if (BP_GET_DEDUP(bp))
1047 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1048 	}
1049 
1050 	return (ZIO_PIPELINE_CONTINUE);
1051 }
1052 
1053 /*
1054  * ==========================================================================
1055  * Execute the I/O pipeline
1056  * ==========================================================================
1057  */
1058 
1059 static void
1060 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1061 {
1062 	spa_t *spa = zio->io_spa;
1063 	zio_type_t t = zio->io_type;
1064 	int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0);
1065 
1066 	/*
1067 	 * If we're a config writer or a probe, the normal issue and
1068 	 * interrupt threads may all be blocked waiting for the config lock.
1069 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1070 	 */
1071 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1072 		t = ZIO_TYPE_NULL;
1073 
1074 	/*
1075 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1076 	 */
1077 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1078 		t = ZIO_TYPE_NULL;
1079 
1080 	/*
1081 	 * If this is a high priority I/O, then use the high priority taskq.
1082 	 */
1083 	if (zio->io_priority == ZIO_PRIORITY_NOW &&
1084 	    spa->spa_zio_taskq[t][q + 1] != NULL)
1085 		q++;
1086 
1087 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1088 	(void) taskq_dispatch(spa->spa_zio_taskq[t][q],
1089 	    (task_func_t *)zio_execute, zio, flags);
1090 }
1091 
1092 static boolean_t
1093 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1094 {
1095 	kthread_t *executor = zio->io_executor;
1096 	spa_t *spa = zio->io_spa;
1097 
1098 	for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1099 		if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1100 			return (B_TRUE);
1101 
1102 	return (B_FALSE);
1103 }
1104 
1105 static int
1106 zio_issue_async(zio_t *zio)
1107 {
1108 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1109 
1110 	return (ZIO_PIPELINE_STOP);
1111 }
1112 
1113 void
1114 zio_interrupt(zio_t *zio)
1115 {
1116 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1117 }
1118 
1119 /*
1120  * Execute the I/O pipeline until one of the following occurs:
1121  * (1) the I/O completes; (2) the pipeline stalls waiting for
1122  * dependent child I/Os; (3) the I/O issues, so we're waiting
1123  * for an I/O completion interrupt; (4) the I/O is delegated by
1124  * vdev-level caching or aggregation; (5) the I/O is deferred
1125  * due to vdev-level queueing; (6) the I/O is handed off to
1126  * another thread.  In all cases, the pipeline stops whenever
1127  * there's no CPU work; it never burns a thread in cv_wait().
1128  *
1129  * There's no locking on io_stage because there's no legitimate way
1130  * for multiple threads to be attempting to process the same I/O.
1131  */
1132 static zio_pipe_stage_t *zio_pipeline[];
1133 
1134 void
1135 zio_execute(zio_t *zio)
1136 {
1137 	zio->io_executor = curthread;
1138 
1139 	while (zio->io_stage < ZIO_STAGE_DONE) {
1140 		enum zio_stage pipeline = zio->io_pipeline;
1141 		enum zio_stage stage = zio->io_stage;
1142 		int rv;
1143 
1144 		ASSERT(!MUTEX_HELD(&zio->io_lock));
1145 		ASSERT(ISP2(stage));
1146 		ASSERT(zio->io_stall == NULL);
1147 
1148 		do {
1149 			stage <<= 1;
1150 		} while ((stage & pipeline) == 0);
1151 
1152 		ASSERT(stage <= ZIO_STAGE_DONE);
1153 
1154 		/*
1155 		 * If we are in interrupt context and this pipeline stage
1156 		 * will grab a config lock that is held across I/O,
1157 		 * or may wait for an I/O that needs an interrupt thread
1158 		 * to complete, issue async to avoid deadlock.
1159 		 *
1160 		 * For VDEV_IO_START, we cut in line so that the io will
1161 		 * be sent to disk promptly.
1162 		 */
1163 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1164 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1165 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1166 			    zio_requeue_io_start_cut_in_line : B_FALSE;
1167 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1168 			return;
1169 		}
1170 
1171 		zio->io_stage = stage;
1172 		rv = zio_pipeline[highbit(stage) - 1](zio);
1173 
1174 		if (rv == ZIO_PIPELINE_STOP)
1175 			return;
1176 
1177 		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1178 	}
1179 }
1180 
1181 /*
1182  * ==========================================================================
1183  * Initiate I/O, either sync or async
1184  * ==========================================================================
1185  */
1186 int
1187 zio_wait(zio_t *zio)
1188 {
1189 	int error;
1190 
1191 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1192 	ASSERT(zio->io_executor == NULL);
1193 
1194 	zio->io_waiter = curthread;
1195 
1196 	zio_execute(zio);
1197 
1198 	mutex_enter(&zio->io_lock);
1199 	while (zio->io_executor != NULL)
1200 		cv_wait(&zio->io_cv, &zio->io_lock);
1201 	mutex_exit(&zio->io_lock);
1202 
1203 	error = zio->io_error;
1204 	zio_destroy(zio);
1205 
1206 	return (error);
1207 }
1208 
1209 void
1210 zio_nowait(zio_t *zio)
1211 {
1212 	ASSERT(zio->io_executor == NULL);
1213 
1214 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1215 	    zio_unique_parent(zio) == NULL) {
1216 		/*
1217 		 * This is a logical async I/O with no parent to wait for it.
1218 		 * We add it to the spa_async_root_zio "Godfather" I/O which
1219 		 * will ensure they complete prior to unloading the pool.
1220 		 */
1221 		spa_t *spa = zio->io_spa;
1222 
1223 		zio_add_child(spa->spa_async_zio_root, zio);
1224 	}
1225 
1226 	zio_execute(zio);
1227 }
1228 
1229 /*
1230  * ==========================================================================
1231  * Reexecute or suspend/resume failed I/O
1232  * ==========================================================================
1233  */
1234 
1235 static void
1236 zio_reexecute(zio_t *pio)
1237 {
1238 	zio_t *cio, *cio_next;
1239 
1240 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1241 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1242 	ASSERT(pio->io_gang_leader == NULL);
1243 	ASSERT(pio->io_gang_tree == NULL);
1244 
1245 	pio->io_flags = pio->io_orig_flags;
1246 	pio->io_stage = pio->io_orig_stage;
1247 	pio->io_pipeline = pio->io_orig_pipeline;
1248 	pio->io_reexecute = 0;
1249 	pio->io_error = 0;
1250 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1251 		pio->io_state[w] = 0;
1252 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1253 		pio->io_child_error[c] = 0;
1254 
1255 	if (IO_IS_ALLOCATING(pio))
1256 		BP_ZERO(pio->io_bp);
1257 
1258 	/*
1259 	 * As we reexecute pio's children, new children could be created.
1260 	 * New children go to the head of pio's io_child_list, however,
1261 	 * so we will (correctly) not reexecute them.  The key is that
1262 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1263 	 * cannot be affected by any side effects of reexecuting 'cio'.
1264 	 */
1265 	for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1266 		cio_next = zio_walk_children(pio);
1267 		mutex_enter(&pio->io_lock);
1268 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1269 			pio->io_children[cio->io_child_type][w]++;
1270 		mutex_exit(&pio->io_lock);
1271 		zio_reexecute(cio);
1272 	}
1273 
1274 	/*
1275 	 * Now that all children have been reexecuted, execute the parent.
1276 	 * We don't reexecute "The Godfather" I/O here as it's the
1277 	 * responsibility of the caller to wait on him.
1278 	 */
1279 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1280 		zio_execute(pio);
1281 }
1282 
1283 void
1284 zio_suspend(spa_t *spa, zio_t *zio)
1285 {
1286 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1287 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1288 		    "failure and the failure mode property for this pool "
1289 		    "is set to panic.", spa_name(spa));
1290 
1291 	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1292 
1293 	mutex_enter(&spa->spa_suspend_lock);
1294 
1295 	if (spa->spa_suspend_zio_root == NULL)
1296 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1297 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1298 		    ZIO_FLAG_GODFATHER);
1299 
1300 	spa->spa_suspended = B_TRUE;
1301 
1302 	if (zio != NULL) {
1303 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1304 		ASSERT(zio != spa->spa_suspend_zio_root);
1305 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1306 		ASSERT(zio_unique_parent(zio) == NULL);
1307 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1308 		zio_add_child(spa->spa_suspend_zio_root, zio);
1309 	}
1310 
1311 	mutex_exit(&spa->spa_suspend_lock);
1312 }
1313 
1314 int
1315 zio_resume(spa_t *spa)
1316 {
1317 	zio_t *pio;
1318 
1319 	/*
1320 	 * Reexecute all previously suspended i/o.
1321 	 */
1322 	mutex_enter(&spa->spa_suspend_lock);
1323 	spa->spa_suspended = B_FALSE;
1324 	cv_broadcast(&spa->spa_suspend_cv);
1325 	pio = spa->spa_suspend_zio_root;
1326 	spa->spa_suspend_zio_root = NULL;
1327 	mutex_exit(&spa->spa_suspend_lock);
1328 
1329 	if (pio == NULL)
1330 		return (0);
1331 
1332 	zio_reexecute(pio);
1333 	return (zio_wait(pio));
1334 }
1335 
1336 void
1337 zio_resume_wait(spa_t *spa)
1338 {
1339 	mutex_enter(&spa->spa_suspend_lock);
1340 	while (spa_suspended(spa))
1341 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1342 	mutex_exit(&spa->spa_suspend_lock);
1343 }
1344 
1345 /*
1346  * ==========================================================================
1347  * Gang blocks.
1348  *
1349  * A gang block is a collection of small blocks that looks to the DMU
1350  * like one large block.  When zio_dva_allocate() cannot find a block
1351  * of the requested size, due to either severe fragmentation or the pool
1352  * being nearly full, it calls zio_write_gang_block() to construct the
1353  * block from smaller fragments.
1354  *
1355  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1356  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
1357  * an indirect block: it's an array of block pointers.  It consumes
1358  * only one sector and hence is allocatable regardless of fragmentation.
1359  * The gang header's bps point to its gang members, which hold the data.
1360  *
1361  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1362  * as the verifier to ensure uniqueness of the SHA256 checksum.
1363  * Critically, the gang block bp's blk_cksum is the checksum of the data,
1364  * not the gang header.  This ensures that data block signatures (needed for
1365  * deduplication) are independent of how the block is physically stored.
1366  *
1367  * Gang blocks can be nested: a gang member may itself be a gang block.
1368  * Thus every gang block is a tree in which root and all interior nodes are
1369  * gang headers, and the leaves are normal blocks that contain user data.
1370  * The root of the gang tree is called the gang leader.
1371  *
1372  * To perform any operation (read, rewrite, free, claim) on a gang block,
1373  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1374  * in the io_gang_tree field of the original logical i/o by recursively
1375  * reading the gang leader and all gang headers below it.  This yields
1376  * an in-core tree containing the contents of every gang header and the
1377  * bps for every constituent of the gang block.
1378  *
1379  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1380  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
1381  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1382  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1383  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1384  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
1385  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1386  * of the gang header plus zio_checksum_compute() of the data to update the
1387  * gang header's blk_cksum as described above.
1388  *
1389  * The two-phase assemble/issue model solves the problem of partial failure --
1390  * what if you'd freed part of a gang block but then couldn't read the
1391  * gang header for another part?  Assembling the entire gang tree first
1392  * ensures that all the necessary gang header I/O has succeeded before
1393  * starting the actual work of free, claim, or write.  Once the gang tree
1394  * is assembled, free and claim are in-memory operations that cannot fail.
1395  *
1396  * In the event that a gang write fails, zio_dva_unallocate() walks the
1397  * gang tree to immediately free (i.e. insert back into the space map)
1398  * everything we've allocated.  This ensures that we don't get ENOSPC
1399  * errors during repeated suspend/resume cycles due to a flaky device.
1400  *
1401  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
1402  * the gang tree, we won't modify the block, so we can safely defer the free
1403  * (knowing that the block is still intact).  If we *can* assemble the gang
1404  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1405  * each constituent bp and we can allocate a new block on the next sync pass.
1406  *
1407  * In all cases, the gang tree allows complete recovery from partial failure.
1408  * ==========================================================================
1409  */
1410 
1411 static zio_t *
1412 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1413 {
1414 	if (gn != NULL)
1415 		return (pio);
1416 
1417 	return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1418 	    NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1419 	    &pio->io_bookmark));
1420 }
1421 
1422 zio_t *
1423 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1424 {
1425 	zio_t *zio;
1426 
1427 	if (gn != NULL) {
1428 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1429 		    gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1430 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1431 		/*
1432 		 * As we rewrite each gang header, the pipeline will compute
1433 		 * a new gang block header checksum for it; but no one will
1434 		 * compute a new data checksum, so we do that here.  The one
1435 		 * exception is the gang leader: the pipeline already computed
1436 		 * its data checksum because that stage precedes gang assembly.
1437 		 * (Presently, nothing actually uses interior data checksums;
1438 		 * this is just good hygiene.)
1439 		 */
1440 		if (gn != pio->io_gang_leader->io_gang_tree) {
1441 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1442 			    data, BP_GET_PSIZE(bp));
1443 		}
1444 		/*
1445 		 * If we are here to damage data for testing purposes,
1446 		 * leave the GBH alone so that we can detect the damage.
1447 		 */
1448 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1449 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1450 	} else {
1451 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1452 		    data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1453 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1454 	}
1455 
1456 	return (zio);
1457 }
1458 
1459 /* ARGSUSED */
1460 zio_t *
1461 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1462 {
1463 	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1464 	    ZIO_GANG_CHILD_FLAGS(pio)));
1465 }
1466 
1467 /* ARGSUSED */
1468 zio_t *
1469 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1470 {
1471 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1472 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1473 }
1474 
1475 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1476 	NULL,
1477 	zio_read_gang,
1478 	zio_rewrite_gang,
1479 	zio_free_gang,
1480 	zio_claim_gang,
1481 	NULL
1482 };
1483 
1484 static void zio_gang_tree_assemble_done(zio_t *zio);
1485 
1486 static zio_gang_node_t *
1487 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1488 {
1489 	zio_gang_node_t *gn;
1490 
1491 	ASSERT(*gnpp == NULL);
1492 
1493 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1494 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1495 	*gnpp = gn;
1496 
1497 	return (gn);
1498 }
1499 
1500 static void
1501 zio_gang_node_free(zio_gang_node_t **gnpp)
1502 {
1503 	zio_gang_node_t *gn = *gnpp;
1504 
1505 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1506 		ASSERT(gn->gn_child[g] == NULL);
1507 
1508 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1509 	kmem_free(gn, sizeof (*gn));
1510 	*gnpp = NULL;
1511 }
1512 
1513 static void
1514 zio_gang_tree_free(zio_gang_node_t **gnpp)
1515 {
1516 	zio_gang_node_t *gn = *gnpp;
1517 
1518 	if (gn == NULL)
1519 		return;
1520 
1521 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1522 		zio_gang_tree_free(&gn->gn_child[g]);
1523 
1524 	zio_gang_node_free(gnpp);
1525 }
1526 
1527 static void
1528 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1529 {
1530 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1531 
1532 	ASSERT(gio->io_gang_leader == gio);
1533 	ASSERT(BP_IS_GANG(bp));
1534 
1535 	zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1536 	    SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1537 	    gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1538 }
1539 
1540 static void
1541 zio_gang_tree_assemble_done(zio_t *zio)
1542 {
1543 	zio_t *gio = zio->io_gang_leader;
1544 	zio_gang_node_t *gn = zio->io_private;
1545 	blkptr_t *bp = zio->io_bp;
1546 
1547 	ASSERT(gio == zio_unique_parent(zio));
1548 	ASSERT(zio->io_child_count == 0);
1549 
1550 	if (zio->io_error)
1551 		return;
1552 
1553 	if (BP_SHOULD_BYTESWAP(bp))
1554 		byteswap_uint64_array(zio->io_data, zio->io_size);
1555 
1556 	ASSERT(zio->io_data == gn->gn_gbh);
1557 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1558 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1559 
1560 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1561 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1562 		if (!BP_IS_GANG(gbp))
1563 			continue;
1564 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1565 	}
1566 }
1567 
1568 static void
1569 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1570 {
1571 	zio_t *gio = pio->io_gang_leader;
1572 	zio_t *zio;
1573 
1574 	ASSERT(BP_IS_GANG(bp) == !!gn);
1575 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1576 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1577 
1578 	/*
1579 	 * If you're a gang header, your data is in gn->gn_gbh.
1580 	 * If you're a gang member, your data is in 'data' and gn == NULL.
1581 	 */
1582 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1583 
1584 	if (gn != NULL) {
1585 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1586 
1587 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1588 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1589 			if (BP_IS_HOLE(gbp))
1590 				continue;
1591 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1592 			data = (char *)data + BP_GET_PSIZE(gbp);
1593 		}
1594 	}
1595 
1596 	if (gn == gio->io_gang_tree)
1597 		ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1598 
1599 	if (zio != pio)
1600 		zio_nowait(zio);
1601 }
1602 
1603 static int
1604 zio_gang_assemble(zio_t *zio)
1605 {
1606 	blkptr_t *bp = zio->io_bp;
1607 
1608 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1609 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1610 
1611 	zio->io_gang_leader = zio;
1612 
1613 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1614 
1615 	return (ZIO_PIPELINE_CONTINUE);
1616 }
1617 
1618 static int
1619 zio_gang_issue(zio_t *zio)
1620 {
1621 	blkptr_t *bp = zio->io_bp;
1622 
1623 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1624 		return (ZIO_PIPELINE_STOP);
1625 
1626 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1627 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1628 
1629 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1630 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1631 	else
1632 		zio_gang_tree_free(&zio->io_gang_tree);
1633 
1634 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1635 
1636 	return (ZIO_PIPELINE_CONTINUE);
1637 }
1638 
1639 static void
1640 zio_write_gang_member_ready(zio_t *zio)
1641 {
1642 	zio_t *pio = zio_unique_parent(zio);
1643 	zio_t *gio = zio->io_gang_leader;
1644 	dva_t *cdva = zio->io_bp->blk_dva;
1645 	dva_t *pdva = pio->io_bp->blk_dva;
1646 	uint64_t asize;
1647 
1648 	if (BP_IS_HOLE(zio->io_bp))
1649 		return;
1650 
1651 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1652 
1653 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1654 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1655 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1656 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1657 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1658 
1659 	mutex_enter(&pio->io_lock);
1660 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1661 		ASSERT(DVA_GET_GANG(&pdva[d]));
1662 		asize = DVA_GET_ASIZE(&pdva[d]);
1663 		asize += DVA_GET_ASIZE(&cdva[d]);
1664 		DVA_SET_ASIZE(&pdva[d], asize);
1665 	}
1666 	mutex_exit(&pio->io_lock);
1667 }
1668 
1669 static int
1670 zio_write_gang_block(zio_t *pio)
1671 {
1672 	spa_t *spa = pio->io_spa;
1673 	blkptr_t *bp = pio->io_bp;
1674 	zio_t *gio = pio->io_gang_leader;
1675 	zio_t *zio;
1676 	zio_gang_node_t *gn, **gnpp;
1677 	zio_gbh_phys_t *gbh;
1678 	uint64_t txg = pio->io_txg;
1679 	uint64_t resid = pio->io_size;
1680 	uint64_t lsize;
1681 	int copies = gio->io_prop.zp_copies;
1682 	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1683 	zio_prop_t zp;
1684 	int error;
1685 
1686 	error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1687 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1688 	    METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1689 	if (error) {
1690 		pio->io_error = error;
1691 		return (ZIO_PIPELINE_CONTINUE);
1692 	}
1693 
1694 	if (pio == gio) {
1695 		gnpp = &gio->io_gang_tree;
1696 	} else {
1697 		gnpp = pio->io_private;
1698 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
1699 	}
1700 
1701 	gn = zio_gang_node_alloc(gnpp);
1702 	gbh = gn->gn_gbh;
1703 	bzero(gbh, SPA_GANGBLOCKSIZE);
1704 
1705 	/*
1706 	 * Create the gang header.
1707 	 */
1708 	zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1709 	    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1710 
1711 	/*
1712 	 * Create and nowait the gang children.
1713 	 */
1714 	for (int g = 0; resid != 0; resid -= lsize, g++) {
1715 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1716 		    SPA_MINBLOCKSIZE);
1717 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1718 
1719 		zp.zp_checksum = gio->io_prop.zp_checksum;
1720 		zp.zp_compress = ZIO_COMPRESS_OFF;
1721 		zp.zp_type = DMU_OT_NONE;
1722 		zp.zp_level = 0;
1723 		zp.zp_copies = gio->io_prop.zp_copies;
1724 		zp.zp_dedup = 0;
1725 		zp.zp_dedup_verify = 0;
1726 
1727 		zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1728 		    (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1729 		    zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1730 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1731 		    &pio->io_bookmark));
1732 	}
1733 
1734 	/*
1735 	 * Set pio's pipeline to just wait for zio to finish.
1736 	 */
1737 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1738 
1739 	zio_nowait(zio);
1740 
1741 	return (ZIO_PIPELINE_CONTINUE);
1742 }
1743 
1744 /*
1745  * ==========================================================================
1746  * Dedup
1747  * ==========================================================================
1748  */
1749 static void
1750 zio_ddt_child_read_done(zio_t *zio)
1751 {
1752 	blkptr_t *bp = zio->io_bp;
1753 	ddt_entry_t *dde = zio->io_private;
1754 	ddt_phys_t *ddp;
1755 	zio_t *pio = zio_unique_parent(zio);
1756 
1757 	mutex_enter(&pio->io_lock);
1758 	ddp = ddt_phys_select(dde, bp);
1759 	if (zio->io_error == 0)
1760 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
1761 	if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1762 		dde->dde_repair_data = zio->io_data;
1763 	else
1764 		zio_buf_free(zio->io_data, zio->io_size);
1765 	mutex_exit(&pio->io_lock);
1766 }
1767 
1768 static int
1769 zio_ddt_read_start(zio_t *zio)
1770 {
1771 	blkptr_t *bp = zio->io_bp;
1772 
1773 	ASSERT(BP_GET_DEDUP(bp));
1774 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1775 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1776 
1777 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1778 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1779 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1780 		ddt_phys_t *ddp = dde->dde_phys;
1781 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1782 		blkptr_t blk;
1783 
1784 		ASSERT(zio->io_vsd == NULL);
1785 		zio->io_vsd = dde;
1786 
1787 		if (ddp_self == NULL)
1788 			return (ZIO_PIPELINE_CONTINUE);
1789 
1790 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1791 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1792 				continue;
1793 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1794 			    &blk);
1795 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
1796 			    zio_buf_alloc(zio->io_size), zio->io_size,
1797 			    zio_ddt_child_read_done, dde, zio->io_priority,
1798 			    ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1799 			    &zio->io_bookmark));
1800 		}
1801 		return (ZIO_PIPELINE_CONTINUE);
1802 	}
1803 
1804 	zio_nowait(zio_read(zio, zio->io_spa, bp,
1805 	    zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1806 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1807 
1808 	return (ZIO_PIPELINE_CONTINUE);
1809 }
1810 
1811 static int
1812 zio_ddt_read_done(zio_t *zio)
1813 {
1814 	blkptr_t *bp = zio->io_bp;
1815 
1816 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1817 		return (ZIO_PIPELINE_STOP);
1818 
1819 	ASSERT(BP_GET_DEDUP(bp));
1820 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1821 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1822 
1823 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1824 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1825 		ddt_entry_t *dde = zio->io_vsd;
1826 		if (ddt == NULL) {
1827 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1828 			return (ZIO_PIPELINE_CONTINUE);
1829 		}
1830 		if (dde == NULL) {
1831 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1832 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1833 			return (ZIO_PIPELINE_STOP);
1834 		}
1835 		if (dde->dde_repair_data != NULL) {
1836 			bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1837 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
1838 		}
1839 		ddt_repair_done(ddt, dde);
1840 		zio->io_vsd = NULL;
1841 	}
1842 
1843 	ASSERT(zio->io_vsd == NULL);
1844 
1845 	return (ZIO_PIPELINE_CONTINUE);
1846 }
1847 
1848 static boolean_t
1849 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1850 {
1851 	spa_t *spa = zio->io_spa;
1852 
1853 	/*
1854 	 * Note: we compare the original data, not the transformed data,
1855 	 * because when zio->io_bp is an override bp, we will not have
1856 	 * pushed the I/O transforms.  That's an important optimization
1857 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1858 	 */
1859 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1860 		zio_t *lio = dde->dde_lead_zio[p];
1861 
1862 		if (lio != NULL) {
1863 			return (lio->io_orig_size != zio->io_orig_size ||
1864 			    bcmp(zio->io_orig_data, lio->io_orig_data,
1865 			    zio->io_orig_size) != 0);
1866 		}
1867 	}
1868 
1869 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1870 		ddt_phys_t *ddp = &dde->dde_phys[p];
1871 
1872 		if (ddp->ddp_phys_birth != 0) {
1873 			arc_buf_t *abuf = NULL;
1874 			uint32_t aflags = ARC_WAIT;
1875 			blkptr_t blk = *zio->io_bp;
1876 			int error;
1877 
1878 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1879 
1880 			ddt_exit(ddt);
1881 
1882 			error = arc_read_nolock(NULL, spa, &blk,
1883 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1884 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1885 			    &aflags, &zio->io_bookmark);
1886 
1887 			if (error == 0) {
1888 				if (arc_buf_size(abuf) != zio->io_orig_size ||
1889 				    bcmp(abuf->b_data, zio->io_orig_data,
1890 				    zio->io_orig_size) != 0)
1891 					error = EEXIST;
1892 				VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1893 			}
1894 
1895 			ddt_enter(ddt);
1896 			return (error != 0);
1897 		}
1898 	}
1899 
1900 	return (B_FALSE);
1901 }
1902 
1903 static void
1904 zio_ddt_child_write_ready(zio_t *zio)
1905 {
1906 	int p = zio->io_prop.zp_copies;
1907 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1908 	ddt_entry_t *dde = zio->io_private;
1909 	ddt_phys_t *ddp = &dde->dde_phys[p];
1910 	zio_t *pio;
1911 
1912 	if (zio->io_error)
1913 		return;
1914 
1915 	ddt_enter(ddt);
1916 
1917 	ASSERT(dde->dde_lead_zio[p] == zio);
1918 
1919 	ddt_phys_fill(ddp, zio->io_bp);
1920 
1921 	while ((pio = zio_walk_parents(zio)) != NULL)
1922 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1923 
1924 	ddt_exit(ddt);
1925 }
1926 
1927 static void
1928 zio_ddt_child_write_done(zio_t *zio)
1929 {
1930 	int p = zio->io_prop.zp_copies;
1931 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1932 	ddt_entry_t *dde = zio->io_private;
1933 	ddt_phys_t *ddp = &dde->dde_phys[p];
1934 
1935 	ddt_enter(ddt);
1936 
1937 	ASSERT(ddp->ddp_refcnt == 0);
1938 	ASSERT(dde->dde_lead_zio[p] == zio);
1939 	dde->dde_lead_zio[p] = NULL;
1940 
1941 	if (zio->io_error == 0) {
1942 		while (zio_walk_parents(zio) != NULL)
1943 			ddt_phys_addref(ddp);
1944 	} else {
1945 		ddt_phys_clear(ddp);
1946 	}
1947 
1948 	ddt_exit(ddt);
1949 }
1950 
1951 static void
1952 zio_ddt_ditto_write_done(zio_t *zio)
1953 {
1954 	int p = DDT_PHYS_DITTO;
1955 	zio_prop_t *zp = &zio->io_prop;
1956 	blkptr_t *bp = zio->io_bp;
1957 	ddt_t *ddt = ddt_select(zio->io_spa, bp);
1958 	ddt_entry_t *dde = zio->io_private;
1959 	ddt_phys_t *ddp = &dde->dde_phys[p];
1960 	ddt_key_t *ddk = &dde->dde_key;
1961 
1962 	ddt_enter(ddt);
1963 
1964 	ASSERT(ddp->ddp_refcnt == 0);
1965 	ASSERT(dde->dde_lead_zio[p] == zio);
1966 	dde->dde_lead_zio[p] = NULL;
1967 
1968 	if (zio->io_error == 0) {
1969 		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1970 		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1971 		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1972 		if (ddp->ddp_phys_birth != 0)
1973 			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1974 		ddt_phys_fill(ddp, bp);
1975 	}
1976 
1977 	ddt_exit(ddt);
1978 }
1979 
1980 static int
1981 zio_ddt_write(zio_t *zio)
1982 {
1983 	spa_t *spa = zio->io_spa;
1984 	blkptr_t *bp = zio->io_bp;
1985 	uint64_t txg = zio->io_txg;
1986 	zio_prop_t *zp = &zio->io_prop;
1987 	int p = zp->zp_copies;
1988 	int ditto_copies;
1989 	zio_t *cio = NULL;
1990 	zio_t *dio = NULL;
1991 	ddt_t *ddt = ddt_select(spa, bp);
1992 	ddt_entry_t *dde;
1993 	ddt_phys_t *ddp;
1994 
1995 	ASSERT(BP_GET_DEDUP(bp));
1996 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
1997 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
1998 
1999 	ddt_enter(ddt);
2000 	dde = ddt_lookup(ddt, bp, B_TRUE);
2001 	ddp = &dde->dde_phys[p];
2002 
2003 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2004 		/*
2005 		 * If we're using a weak checksum, upgrade to a strong checksum
2006 		 * and try again.  If we're already using a strong checksum,
2007 		 * we can't resolve it, so just convert to an ordinary write.
2008 		 * (And automatically e-mail a paper to Nature?)
2009 		 */
2010 		if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2011 			zp->zp_checksum = spa_dedup_checksum(spa);
2012 			zio_pop_transforms(zio);
2013 			zio->io_stage = ZIO_STAGE_OPEN;
2014 			BP_ZERO(bp);
2015 		} else {
2016 			zp->zp_dedup = 0;
2017 		}
2018 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2019 		ddt_exit(ddt);
2020 		return (ZIO_PIPELINE_CONTINUE);
2021 	}
2022 
2023 	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2024 	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2025 
2026 	if (ditto_copies > ddt_ditto_copies_present(dde) &&
2027 	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2028 		zio_prop_t czp = *zp;
2029 
2030 		czp.zp_copies = ditto_copies;
2031 
2032 		/*
2033 		 * If we arrived here with an override bp, we won't have run
2034 		 * the transform stack, so we won't have the data we need to
2035 		 * generate a child i/o.  So, toss the override bp and restart.
2036 		 * This is safe, because using the override bp is just an
2037 		 * optimization; and it's rare, so the cost doesn't matter.
2038 		 */
2039 		if (zio->io_bp_override) {
2040 			zio_pop_transforms(zio);
2041 			zio->io_stage = ZIO_STAGE_OPEN;
2042 			zio->io_pipeline = ZIO_WRITE_PIPELINE;
2043 			zio->io_bp_override = NULL;
2044 			BP_ZERO(bp);
2045 			ddt_exit(ddt);
2046 			return (ZIO_PIPELINE_CONTINUE);
2047 		}
2048 
2049 		dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2050 		    zio->io_orig_size, &czp, NULL,
2051 		    zio_ddt_ditto_write_done, dde, zio->io_priority,
2052 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2053 
2054 		zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2055 		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2056 	}
2057 
2058 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2059 		if (ddp->ddp_phys_birth != 0)
2060 			ddt_bp_fill(ddp, bp, txg);
2061 		if (dde->dde_lead_zio[p] != NULL)
2062 			zio_add_child(zio, dde->dde_lead_zio[p]);
2063 		else
2064 			ddt_phys_addref(ddp);
2065 	} else if (zio->io_bp_override) {
2066 		ASSERT(bp->blk_birth == txg);
2067 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2068 		ddt_phys_fill(ddp, bp);
2069 		ddt_phys_addref(ddp);
2070 	} else {
2071 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2072 		    zio->io_orig_size, zp, zio_ddt_child_write_ready,
2073 		    zio_ddt_child_write_done, dde, zio->io_priority,
2074 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2075 
2076 		zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2077 		dde->dde_lead_zio[p] = cio;
2078 	}
2079 
2080 	ddt_exit(ddt);
2081 
2082 	if (cio)
2083 		zio_nowait(cio);
2084 	if (dio)
2085 		zio_nowait(dio);
2086 
2087 	return (ZIO_PIPELINE_CONTINUE);
2088 }
2089 
2090 ddt_entry_t *freedde; /* for debugging */
2091 
2092 static int
2093 zio_ddt_free(zio_t *zio)
2094 {
2095 	spa_t *spa = zio->io_spa;
2096 	blkptr_t *bp = zio->io_bp;
2097 	ddt_t *ddt = ddt_select(spa, bp);
2098 	ddt_entry_t *dde;
2099 	ddt_phys_t *ddp;
2100 
2101 	ASSERT(BP_GET_DEDUP(bp));
2102 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2103 
2104 	ddt_enter(ddt);
2105 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2106 	ddp = ddt_phys_select(dde, bp);
2107 	ddt_phys_decref(ddp);
2108 	ddt_exit(ddt);
2109 
2110 	return (ZIO_PIPELINE_CONTINUE);
2111 }
2112 
2113 /*
2114  * ==========================================================================
2115  * Allocate and free blocks
2116  * ==========================================================================
2117  */
2118 static int
2119 zio_dva_allocate(zio_t *zio)
2120 {
2121 	spa_t *spa = zio->io_spa;
2122 	metaslab_class_t *mc = spa_normal_class(spa);
2123 	blkptr_t *bp = zio->io_bp;
2124 	int error;
2125 	int flags = 0;
2126 
2127 	if (zio->io_gang_leader == NULL) {
2128 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2129 		zio->io_gang_leader = zio;
2130 	}
2131 
2132 	ASSERT(BP_IS_HOLE(bp));
2133 	ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2134 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
2135 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2136 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2137 
2138 	/*
2139 	 * The dump device does not support gang blocks so allocation on
2140 	 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2141 	 * the "fast" gang feature.
2142 	 */
2143 	flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2144 	flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2145 	    METASLAB_GANG_CHILD : 0;
2146 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
2147 	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2148 
2149 	if (error) {
2150 		spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2151 		    "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2152 		    error);
2153 		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2154 			return (zio_write_gang_block(zio));
2155 		zio->io_error = error;
2156 	}
2157 
2158 	return (ZIO_PIPELINE_CONTINUE);
2159 }
2160 
2161 static int
2162 zio_dva_free(zio_t *zio)
2163 {
2164 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2165 
2166 	return (ZIO_PIPELINE_CONTINUE);
2167 }
2168 
2169 static int
2170 zio_dva_claim(zio_t *zio)
2171 {
2172 	int error;
2173 
2174 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2175 	if (error)
2176 		zio->io_error = error;
2177 
2178 	return (ZIO_PIPELINE_CONTINUE);
2179 }
2180 
2181 /*
2182  * Undo an allocation.  This is used by zio_done() when an I/O fails
2183  * and we want to give back the block we just allocated.
2184  * This handles both normal blocks and gang blocks.
2185  */
2186 static void
2187 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2188 {
2189 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2190 	ASSERT(zio->io_bp_override == NULL);
2191 
2192 	if (!BP_IS_HOLE(bp))
2193 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2194 
2195 	if (gn != NULL) {
2196 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2197 			zio_dva_unallocate(zio, gn->gn_child[g],
2198 			    &gn->gn_gbh->zg_blkptr[g]);
2199 		}
2200 	}
2201 }
2202 
2203 /*
2204  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
2205  */
2206 int
2207 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2208     uint64_t size, boolean_t use_slog)
2209 {
2210 	int error = 1;
2211 
2212 	ASSERT(txg > spa_syncing_txg(spa));
2213 
2214 	if (use_slog)
2215 		error = metaslab_alloc(spa, spa_log_class(spa), size,
2216 		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2217 
2218 	if (error)
2219 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
2220 		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2221 
2222 	if (error == 0) {
2223 		BP_SET_LSIZE(new_bp, size);
2224 		BP_SET_PSIZE(new_bp, size);
2225 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2226 		BP_SET_CHECKSUM(new_bp,
2227 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2228 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2229 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2230 		BP_SET_LEVEL(new_bp, 0);
2231 		BP_SET_DEDUP(new_bp, 0);
2232 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2233 	}
2234 
2235 	return (error);
2236 }
2237 
2238 /*
2239  * Free an intent log block.
2240  */
2241 void
2242 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2243 {
2244 	ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2245 	ASSERT(!BP_IS_GANG(bp));
2246 
2247 	zio_free(spa, txg, bp);
2248 }
2249 
2250 /*
2251  * ==========================================================================
2252  * Read and write to physical devices
2253  * ==========================================================================
2254  */
2255 static int
2256 zio_vdev_io_start(zio_t *zio)
2257 {
2258 	vdev_t *vd = zio->io_vd;
2259 	uint64_t align;
2260 	spa_t *spa = zio->io_spa;
2261 
2262 	ASSERT(zio->io_error == 0);
2263 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2264 
2265 	if (vd == NULL) {
2266 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2267 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2268 
2269 		/*
2270 		 * The mirror_ops handle multiple DVAs in a single BP.
2271 		 */
2272 		return (vdev_mirror_ops.vdev_op_io_start(zio));
2273 	}
2274 
2275 	/*
2276 	 * We keep track of time-sensitive I/Os so that the scan thread
2277 	 * can quickly react to certain workloads.  In particular, we care
2278 	 * about non-scrubbing, top-level reads and writes with the following
2279 	 * characteristics:
2280 	 * 	- synchronous writes of user data to non-slog devices
2281 	 *	- any reads of user data
2282 	 * When these conditions are met, adjust the timestamp of spa_last_io
2283 	 * which allows the scan thread to adjust its workload accordingly.
2284 	 */
2285 	if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2286 	    vd == vd->vdev_top && !vd->vdev_islog &&
2287 	    zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2288 	    zio->io_txg != spa_syncing_txg(spa)) {
2289 		uint64_t old = spa->spa_last_io;
2290 		uint64_t new = ddi_get_lbolt64();
2291 		if (old != new)
2292 			(void) atomic_cas_64(&spa->spa_last_io, old, new);
2293 	}
2294 
2295 	align = 1ULL << vd->vdev_top->vdev_ashift;
2296 
2297 	if (P2PHASE(zio->io_size, align) != 0) {
2298 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
2299 		char *abuf = zio_buf_alloc(asize);
2300 		ASSERT(vd == vd->vdev_top);
2301 		if (zio->io_type == ZIO_TYPE_WRITE) {
2302 			bcopy(zio->io_data, abuf, zio->io_size);
2303 			bzero(abuf + zio->io_size, asize - zio->io_size);
2304 		}
2305 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2306 	}
2307 
2308 	ASSERT(P2PHASE(zio->io_offset, align) == 0);
2309 	ASSERT(P2PHASE(zio->io_size, align) == 0);
2310 	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2311 
2312 	/*
2313 	 * If this is a repair I/O, and there's no self-healing involved --
2314 	 * that is, we're just resilvering what we expect to resilver --
2315 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2316 	 * This prevents spurious resilvering with nested replication.
2317 	 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2318 	 * A is out of date, we'll read from C+D, then use the data to
2319 	 * resilver A+B -- but we don't actually want to resilver B, just A.
2320 	 * The top-level mirror has no way to know this, so instead we just
2321 	 * discard unnecessary repairs as we work our way down the vdev tree.
2322 	 * The same logic applies to any form of nested replication:
2323 	 * ditto + mirror, RAID-Z + replacing, etc.  This covers them all.
2324 	 */
2325 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2326 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2327 	    zio->io_txg != 0 &&	/* not a delegated i/o */
2328 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2329 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2330 		zio_vdev_io_bypass(zio);
2331 		return (ZIO_PIPELINE_CONTINUE);
2332 	}
2333 
2334 	if (vd->vdev_ops->vdev_op_leaf &&
2335 	    (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2336 
2337 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2338 			return (ZIO_PIPELINE_CONTINUE);
2339 
2340 		if ((zio = vdev_queue_io(zio)) == NULL)
2341 			return (ZIO_PIPELINE_STOP);
2342 
2343 		if (!vdev_accessible(vd, zio)) {
2344 			zio->io_error = ENXIO;
2345 			zio_interrupt(zio);
2346 			return (ZIO_PIPELINE_STOP);
2347 		}
2348 	}
2349 
2350 	return (vd->vdev_ops->vdev_op_io_start(zio));
2351 }
2352 
2353 static int
2354 zio_vdev_io_done(zio_t *zio)
2355 {
2356 	vdev_t *vd = zio->io_vd;
2357 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2358 	boolean_t unexpected_error = B_FALSE;
2359 
2360 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2361 		return (ZIO_PIPELINE_STOP);
2362 
2363 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2364 
2365 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2366 
2367 		vdev_queue_io_done(zio);
2368 
2369 		if (zio->io_type == ZIO_TYPE_WRITE)
2370 			vdev_cache_write(zio);
2371 
2372 		if (zio_injection_enabled && zio->io_error == 0)
2373 			zio->io_error = zio_handle_device_injection(vd,
2374 			    zio, EIO);
2375 
2376 		if (zio_injection_enabled && zio->io_error == 0)
2377 			zio->io_error = zio_handle_label_injection(zio, EIO);
2378 
2379 		if (zio->io_error) {
2380 			if (!vdev_accessible(vd, zio)) {
2381 				zio->io_error = ENXIO;
2382 			} else {
2383 				unexpected_error = B_TRUE;
2384 			}
2385 		}
2386 	}
2387 
2388 	ops->vdev_op_io_done(zio);
2389 
2390 	if (unexpected_error)
2391 		VERIFY(vdev_probe(vd, zio) == NULL);
2392 
2393 	return (ZIO_PIPELINE_CONTINUE);
2394 }
2395 
2396 /*
2397  * For non-raidz ZIOs, we can just copy aside the bad data read from the
2398  * disk, and use that to finish the checksum ereport later.
2399  */
2400 static void
2401 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2402     const void *good_buf)
2403 {
2404 	/* no processing needed */
2405 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2406 }
2407 
2408 /*ARGSUSED*/
2409 void
2410 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2411 {
2412 	void *buf = zio_buf_alloc(zio->io_size);
2413 
2414 	bcopy(zio->io_data, buf, zio->io_size);
2415 
2416 	zcr->zcr_cbinfo = zio->io_size;
2417 	zcr->zcr_cbdata = buf;
2418 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
2419 	zcr->zcr_free = zio_buf_free;
2420 }
2421 
2422 static int
2423 zio_vdev_io_assess(zio_t *zio)
2424 {
2425 	vdev_t *vd = zio->io_vd;
2426 
2427 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2428 		return (ZIO_PIPELINE_STOP);
2429 
2430 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2431 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2432 
2433 	if (zio->io_vsd != NULL) {
2434 		zio->io_vsd_ops->vsd_free(zio);
2435 		zio->io_vsd = NULL;
2436 	}
2437 
2438 	if (zio_injection_enabled && zio->io_error == 0)
2439 		zio->io_error = zio_handle_fault_injection(zio, EIO);
2440 
2441 	/*
2442 	 * If the I/O failed, determine whether we should attempt to retry it.
2443 	 *
2444 	 * On retry, we cut in line in the issue queue, since we don't want
2445 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2446 	 */
2447 	if (zio->io_error && vd == NULL &&
2448 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2449 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
2450 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
2451 		zio->io_error = 0;
2452 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
2453 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2454 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2455 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2456 		    zio_requeue_io_start_cut_in_line);
2457 		return (ZIO_PIPELINE_STOP);
2458 	}
2459 
2460 	/*
2461 	 * If we got an error on a leaf device, convert it to ENXIO
2462 	 * if the device is not accessible at all.
2463 	 */
2464 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2465 	    !vdev_accessible(vd, zio))
2466 		zio->io_error = ENXIO;
2467 
2468 	/*
2469 	 * If we can't write to an interior vdev (mirror or RAID-Z),
2470 	 * set vdev_cant_write so that we stop trying to allocate from it.
2471 	 */
2472 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2473 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2474 		vd->vdev_cant_write = B_TRUE;
2475 
2476 	if (zio->io_error)
2477 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2478 
2479 	return (ZIO_PIPELINE_CONTINUE);
2480 }
2481 
2482 void
2483 zio_vdev_io_reissue(zio_t *zio)
2484 {
2485 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2486 	ASSERT(zio->io_error == 0);
2487 
2488 	zio->io_stage >>= 1;
2489 }
2490 
2491 void
2492 zio_vdev_io_redone(zio_t *zio)
2493 {
2494 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2495 
2496 	zio->io_stage >>= 1;
2497 }
2498 
2499 void
2500 zio_vdev_io_bypass(zio_t *zio)
2501 {
2502 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2503 	ASSERT(zio->io_error == 0);
2504 
2505 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2506 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2507 }
2508 
2509 /*
2510  * ==========================================================================
2511  * Generate and verify checksums
2512  * ==========================================================================
2513  */
2514 static int
2515 zio_checksum_generate(zio_t *zio)
2516 {
2517 	blkptr_t *bp = zio->io_bp;
2518 	enum zio_checksum checksum;
2519 
2520 	if (bp == NULL) {
2521 		/*
2522 		 * This is zio_write_phys().
2523 		 * We're either generating a label checksum, or none at all.
2524 		 */
2525 		checksum = zio->io_prop.zp_checksum;
2526 
2527 		if (checksum == ZIO_CHECKSUM_OFF)
2528 			return (ZIO_PIPELINE_CONTINUE);
2529 
2530 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2531 	} else {
2532 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2533 			ASSERT(!IO_IS_ALLOCATING(zio));
2534 			checksum = ZIO_CHECKSUM_GANG_HEADER;
2535 		} else {
2536 			checksum = BP_GET_CHECKSUM(bp);
2537 		}
2538 	}
2539 
2540 	zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2541 
2542 	return (ZIO_PIPELINE_CONTINUE);
2543 }
2544 
2545 static int
2546 zio_checksum_verify(zio_t *zio)
2547 {
2548 	zio_bad_cksum_t info;
2549 	blkptr_t *bp = zio->io_bp;
2550 	int error;
2551 
2552 	ASSERT(zio->io_vd != NULL);
2553 
2554 	if (bp == NULL) {
2555 		/*
2556 		 * This is zio_read_phys().
2557 		 * We're either verifying a label checksum, or nothing at all.
2558 		 */
2559 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2560 			return (ZIO_PIPELINE_CONTINUE);
2561 
2562 		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2563 	}
2564 
2565 	if ((error = zio_checksum_error(zio, &info)) != 0) {
2566 		zio->io_error = error;
2567 		if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2568 			zfs_ereport_start_checksum(zio->io_spa,
2569 			    zio->io_vd, zio, zio->io_offset,
2570 			    zio->io_size, NULL, &info);
2571 		}
2572 	}
2573 
2574 	return (ZIO_PIPELINE_CONTINUE);
2575 }
2576 
2577 /*
2578  * Called by RAID-Z to ensure we don't compute the checksum twice.
2579  */
2580 void
2581 zio_checksum_verified(zio_t *zio)
2582 {
2583 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2584 }
2585 
2586 /*
2587  * ==========================================================================
2588  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2589  * An error of 0 indictes success.  ENXIO indicates whole-device failure,
2590  * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
2591  * indicate errors that are specific to one I/O, and most likely permanent.
2592  * Any other error is presumed to be worse because we weren't expecting it.
2593  * ==========================================================================
2594  */
2595 int
2596 zio_worst_error(int e1, int e2)
2597 {
2598 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2599 	int r1, r2;
2600 
2601 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2602 		if (e1 == zio_error_rank[r1])
2603 			break;
2604 
2605 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2606 		if (e2 == zio_error_rank[r2])
2607 			break;
2608 
2609 	return (r1 > r2 ? e1 : e2);
2610 }
2611 
2612 /*
2613  * ==========================================================================
2614  * I/O completion
2615  * ==========================================================================
2616  */
2617 static int
2618 zio_ready(zio_t *zio)
2619 {
2620 	blkptr_t *bp = zio->io_bp;
2621 	zio_t *pio, *pio_next;
2622 
2623 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2624 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2625 		return (ZIO_PIPELINE_STOP);
2626 
2627 	if (zio->io_ready) {
2628 		ASSERT(IO_IS_ALLOCATING(zio));
2629 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2630 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2631 
2632 		zio->io_ready(zio);
2633 	}
2634 
2635 	if (bp != NULL && bp != &zio->io_bp_copy)
2636 		zio->io_bp_copy = *bp;
2637 
2638 	if (zio->io_error)
2639 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2640 
2641 	mutex_enter(&zio->io_lock);
2642 	zio->io_state[ZIO_WAIT_READY] = 1;
2643 	pio = zio_walk_parents(zio);
2644 	mutex_exit(&zio->io_lock);
2645 
2646 	/*
2647 	 * As we notify zio's parents, new parents could be added.
2648 	 * New parents go to the head of zio's io_parent_list, however,
2649 	 * so we will (correctly) not notify them.  The remainder of zio's
2650 	 * io_parent_list, from 'pio_next' onward, cannot change because
2651 	 * all parents must wait for us to be done before they can be done.
2652 	 */
2653 	for (; pio != NULL; pio = pio_next) {
2654 		pio_next = zio_walk_parents(zio);
2655 		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2656 	}
2657 
2658 	if (zio->io_flags & ZIO_FLAG_NODATA) {
2659 		if (BP_IS_GANG(bp)) {
2660 			zio->io_flags &= ~ZIO_FLAG_NODATA;
2661 		} else {
2662 			ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2663 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2664 		}
2665 	}
2666 
2667 	if (zio_injection_enabled &&
2668 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
2669 		zio_handle_ignored_writes(zio);
2670 
2671 	return (ZIO_PIPELINE_CONTINUE);
2672 }
2673 
2674 static int
2675 zio_done(zio_t *zio)
2676 {
2677 	spa_t *spa = zio->io_spa;
2678 	zio_t *lio = zio->io_logical;
2679 	blkptr_t *bp = zio->io_bp;
2680 	vdev_t *vd = zio->io_vd;
2681 	uint64_t psize = zio->io_size;
2682 	zio_t *pio, *pio_next;
2683 
2684 	/*
2685 	 * If our children haven't all completed,
2686 	 * wait for them and then repeat this pipeline stage.
2687 	 */
2688 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2689 	    zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2690 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2691 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2692 		return (ZIO_PIPELINE_STOP);
2693 
2694 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2695 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2696 			ASSERT(zio->io_children[c][w] == 0);
2697 
2698 	if (bp != NULL) {
2699 		ASSERT(bp->blk_pad[0] == 0);
2700 		ASSERT(bp->blk_pad[1] == 0);
2701 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2702 		    (bp == zio_unique_parent(zio)->io_bp));
2703 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2704 		    zio->io_bp_override == NULL &&
2705 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2706 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
2707 			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2708 			ASSERT(BP_COUNT_GANG(bp) == 0 ||
2709 			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2710 		}
2711 	}
2712 
2713 	/*
2714 	 * If there were child vdev/gang/ddt errors, they apply to us now.
2715 	 */
2716 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2717 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2718 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2719 
2720 	/*
2721 	 * If the I/O on the transformed data was successful, generate any
2722 	 * checksum reports now while we still have the transformed data.
2723 	 */
2724 	if (zio->io_error == 0) {
2725 		while (zio->io_cksum_report != NULL) {
2726 			zio_cksum_report_t *zcr = zio->io_cksum_report;
2727 			uint64_t align = zcr->zcr_align;
2728 			uint64_t asize = P2ROUNDUP(psize, align);
2729 			char *abuf = zio->io_data;
2730 
2731 			if (asize != psize) {
2732 				abuf = zio_buf_alloc(asize);
2733 				bcopy(zio->io_data, abuf, psize);
2734 				bzero(abuf + psize, asize - psize);
2735 			}
2736 
2737 			zio->io_cksum_report = zcr->zcr_next;
2738 			zcr->zcr_next = NULL;
2739 			zcr->zcr_finish(zcr, abuf);
2740 			zfs_ereport_free_checksum(zcr);
2741 
2742 			if (asize != psize)
2743 				zio_buf_free(abuf, asize);
2744 		}
2745 	}
2746 
2747 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
2748 
2749 	vdev_stat_update(zio, psize);
2750 
2751 	if (zio->io_error) {
2752 		/*
2753 		 * If this I/O is attached to a particular vdev,
2754 		 * generate an error message describing the I/O failure
2755 		 * at the block level.  We ignore these errors if the
2756 		 * device is currently unavailable.
2757 		 */
2758 		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2759 			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2760 
2761 		if ((zio->io_error == EIO || !(zio->io_flags &
2762 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2763 		    zio == lio) {
2764 			/*
2765 			 * For logical I/O requests, tell the SPA to log the
2766 			 * error and generate a logical data ereport.
2767 			 */
2768 			spa_log_error(spa, zio);
2769 			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2770 			    0, 0);
2771 		}
2772 	}
2773 
2774 	if (zio->io_error && zio == lio) {
2775 		/*
2776 		 * Determine whether zio should be reexecuted.  This will
2777 		 * propagate all the way to the root via zio_notify_parent().
2778 		 */
2779 		ASSERT(vd == NULL && bp != NULL);
2780 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2781 
2782 		if (IO_IS_ALLOCATING(zio) &&
2783 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2784 			if (zio->io_error != ENOSPC)
2785 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2786 			else
2787 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2788 		}
2789 
2790 		if ((zio->io_type == ZIO_TYPE_READ ||
2791 		    zio->io_type == ZIO_TYPE_FREE) &&
2792 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2793 		    zio->io_error == ENXIO &&
2794 		    spa_load_state(spa) == SPA_LOAD_NONE &&
2795 		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2796 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2797 
2798 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2799 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2800 
2801 		/*
2802 		 * Here is a possibly good place to attempt to do
2803 		 * either combinatorial reconstruction or error correction
2804 		 * based on checksums.  It also might be a good place
2805 		 * to send out preliminary ereports before we suspend
2806 		 * processing.
2807 		 */
2808 	}
2809 
2810 	/*
2811 	 * If there were logical child errors, they apply to us now.
2812 	 * We defer this until now to avoid conflating logical child
2813 	 * errors with errors that happened to the zio itself when
2814 	 * updating vdev stats and reporting FMA events above.
2815 	 */
2816 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2817 
2818 	if ((zio->io_error || zio->io_reexecute) &&
2819 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2820 	    !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2821 		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2822 
2823 	zio_gang_tree_free(&zio->io_gang_tree);
2824 
2825 	/*
2826 	 * Godfather I/Os should never suspend.
2827 	 */
2828 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2829 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2830 		zio->io_reexecute = 0;
2831 
2832 	if (zio->io_reexecute) {
2833 		/*
2834 		 * This is a logical I/O that wants to reexecute.
2835 		 *
2836 		 * Reexecute is top-down.  When an i/o fails, if it's not
2837 		 * the root, it simply notifies its parent and sticks around.
2838 		 * The parent, seeing that it still has children in zio_done(),
2839 		 * does the same.  This percolates all the way up to the root.
2840 		 * The root i/o will reexecute or suspend the entire tree.
2841 		 *
2842 		 * This approach ensures that zio_reexecute() honors
2843 		 * all the original i/o dependency relationships, e.g.
2844 		 * parents not executing until children are ready.
2845 		 */
2846 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2847 
2848 		zio->io_gang_leader = NULL;
2849 
2850 		mutex_enter(&zio->io_lock);
2851 		zio->io_state[ZIO_WAIT_DONE] = 1;
2852 		mutex_exit(&zio->io_lock);
2853 
2854 		/*
2855 		 * "The Godfather" I/O monitors its children but is
2856 		 * not a true parent to them. It will track them through
2857 		 * the pipeline but severs its ties whenever they get into
2858 		 * trouble (e.g. suspended). This allows "The Godfather"
2859 		 * I/O to return status without blocking.
2860 		 */
2861 		for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2862 			zio_link_t *zl = zio->io_walk_link;
2863 			pio_next = zio_walk_parents(zio);
2864 
2865 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2866 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2867 				zio_remove_child(pio, zio, zl);
2868 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2869 			}
2870 		}
2871 
2872 		if ((pio = zio_unique_parent(zio)) != NULL) {
2873 			/*
2874 			 * We're not a root i/o, so there's nothing to do
2875 			 * but notify our parent.  Don't propagate errors
2876 			 * upward since we haven't permanently failed yet.
2877 			 */
2878 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2879 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2880 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2881 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2882 			/*
2883 			 * We'd fail again if we reexecuted now, so suspend
2884 			 * until conditions improve (e.g. device comes online).
2885 			 */
2886 			zio_suspend(spa, zio);
2887 		} else {
2888 			/*
2889 			 * Reexecution is potentially a huge amount of work.
2890 			 * Hand it off to the otherwise-unused claim taskq.
2891 			 */
2892 			(void) taskq_dispatch(
2893 			    spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2894 			    (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2895 		}
2896 		return (ZIO_PIPELINE_STOP);
2897 	}
2898 
2899 	ASSERT(zio->io_child_count == 0);
2900 	ASSERT(zio->io_reexecute == 0);
2901 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2902 
2903 	/*
2904 	 * Report any checksum errors, since the I/O is complete.
2905 	 */
2906 	while (zio->io_cksum_report != NULL) {
2907 		zio_cksum_report_t *zcr = zio->io_cksum_report;
2908 		zio->io_cksum_report = zcr->zcr_next;
2909 		zcr->zcr_next = NULL;
2910 		zcr->zcr_finish(zcr, NULL);
2911 		zfs_ereport_free_checksum(zcr);
2912 	}
2913 
2914 	/*
2915 	 * It is the responsibility of the done callback to ensure that this
2916 	 * particular zio is no longer discoverable for adoption, and as
2917 	 * such, cannot acquire any new parents.
2918 	 */
2919 	if (zio->io_done)
2920 		zio->io_done(zio);
2921 
2922 	mutex_enter(&zio->io_lock);
2923 	zio->io_state[ZIO_WAIT_DONE] = 1;
2924 	mutex_exit(&zio->io_lock);
2925 
2926 	for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2927 		zio_link_t *zl = zio->io_walk_link;
2928 		pio_next = zio_walk_parents(zio);
2929 		zio_remove_child(pio, zio, zl);
2930 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2931 	}
2932 
2933 	if (zio->io_waiter != NULL) {
2934 		mutex_enter(&zio->io_lock);
2935 		zio->io_executor = NULL;
2936 		cv_broadcast(&zio->io_cv);
2937 		mutex_exit(&zio->io_lock);
2938 	} else {
2939 		zio_destroy(zio);
2940 	}
2941 
2942 	return (ZIO_PIPELINE_STOP);
2943 }
2944 
2945 /*
2946  * ==========================================================================
2947  * I/O pipeline definition
2948  * ==========================================================================
2949  */
2950 static zio_pipe_stage_t *zio_pipeline[] = {
2951 	NULL,
2952 	zio_read_bp_init,
2953 	zio_free_bp_init,
2954 	zio_issue_async,
2955 	zio_write_bp_init,
2956 	zio_checksum_generate,
2957 	zio_ddt_read_start,
2958 	zio_ddt_read_done,
2959 	zio_ddt_write,
2960 	zio_ddt_free,
2961 	zio_gang_assemble,
2962 	zio_gang_issue,
2963 	zio_dva_allocate,
2964 	zio_dva_free,
2965 	zio_dva_claim,
2966 	zio_ready,
2967 	zio_vdev_io_start,
2968 	zio_vdev_io_done,
2969 	zio_vdev_io_assess,
2970 	zio_checksum_verify,
2971 	zio_done
2972 };
2973