xref: /titanic_50/usr/src/uts/common/fs/zfs/zio.c (revision 15721462a45d32950a5b1f05d7b7a0fe34f81811)
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) 2012 by Delphix. All rights reserved.
24  * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25  */
26 
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
29 #include <sys/spa.h>
30 #include <sys/txg.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio_impl.h>
34 #include <sys/zio_compress.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/arc.h>
38 #include <sys/ddt.h>
39 
40 /*
41  * ==========================================================================
42  * I/O priority table
43  * ==========================================================================
44  */
45 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
46 	0,	/* ZIO_PRIORITY_NOW		*/
47 	0,	/* ZIO_PRIORITY_SYNC_READ	*/
48 	0,	/* ZIO_PRIORITY_SYNC_WRITE	*/
49 	0,	/* ZIO_PRIORITY_LOG_WRITE	*/
50 	1,	/* ZIO_PRIORITY_CACHE_FILL	*/
51 	1,	/* ZIO_PRIORITY_AGG		*/
52 	4,	/* ZIO_PRIORITY_FREE		*/
53 	4,	/* ZIO_PRIORITY_ASYNC_WRITE	*/
54 	6,	/* ZIO_PRIORITY_ASYNC_READ	*/
55 	10,	/* ZIO_PRIORITY_RESILVER	*/
56 	20,	/* ZIO_PRIORITY_SCRUB		*/
57 	2,	/* ZIO_PRIORITY_DDT_PREFETCH	*/
58 };
59 
60 /*
61  * ==========================================================================
62  * I/O type descriptions
63  * ==========================================================================
64  */
65 char *zio_type_name[ZIO_TYPES] = {
66 	"zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
67 	"zio_ioctl"
68 };
69 
70 /*
71  * ==========================================================================
72  * I/O kmem caches
73  * ==========================================================================
74  */
75 kmem_cache_t *zio_cache;
76 kmem_cache_t *zio_link_cache;
77 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
78 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
79 
80 #ifdef _KERNEL
81 extern vmem_t *zio_alloc_arena;
82 #endif
83 extern int zfs_mg_alloc_failures;
84 
85 /*
86  * An allocating zio is one that either currently has the DVA allocate
87  * stage set or will have it later in its lifetime.
88  */
89 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
90 
91 boolean_t	zio_requeue_io_start_cut_in_line = B_TRUE;
92 
93 #ifdef ZFS_DEBUG
94 int zio_buf_debug_limit = 16384;
95 #else
96 int zio_buf_debug_limit = 0;
97 #endif
98 
99 void
100 zio_init(void)
101 {
102 	size_t c;
103 	vmem_t *data_alloc_arena = NULL;
104 
105 #ifdef _KERNEL
106 	data_alloc_arena = zio_alloc_arena;
107 #endif
108 	zio_cache = kmem_cache_create("zio_cache",
109 	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
110 	zio_link_cache = kmem_cache_create("zio_link_cache",
111 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
112 
113 	/*
114 	 * For small buffers, we want a cache for each multiple of
115 	 * SPA_MINBLOCKSIZE.  For medium-size buffers, we want a cache
116 	 * for each quarter-power of 2.  For large buffers, we want
117 	 * a cache for each multiple of PAGESIZE.
118 	 */
119 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
120 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
121 		size_t p2 = size;
122 		size_t align = 0;
123 		size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
124 
125 		while (p2 & (p2 - 1))
126 			p2 &= p2 - 1;
127 
128 #ifndef _KERNEL
129 		/*
130 		 * If we are using watchpoints, put each buffer on its own page,
131 		 * to eliminate the performance overhead of trapping to the
132 		 * kernel when modifying a non-watched buffer that shares the
133 		 * page with a watched buffer.
134 		 */
135 		if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
136 			continue;
137 #endif
138 		if (size <= 4 * SPA_MINBLOCKSIZE) {
139 			align = SPA_MINBLOCKSIZE;
140 		} else if (IS_P2ALIGNED(size, PAGESIZE)) {
141 			align = PAGESIZE;
142 		} else if (IS_P2ALIGNED(size, p2 >> 2)) {
143 			align = p2 >> 2;
144 		}
145 
146 		if (align != 0) {
147 			char name[36];
148 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
149 			zio_buf_cache[c] = kmem_cache_create(name, size,
150 			    align, NULL, NULL, NULL, NULL, NULL, cflags);
151 
152 			/*
153 			 * Since zio_data bufs do not appear in crash dumps, we
154 			 * pass KMC_NOTOUCH so that no allocator metadata is
155 			 * stored with the buffers.
156 			 */
157 			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
158 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
159 			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
160 			    cflags | KMC_NOTOUCH);
161 		}
162 	}
163 
164 	while (--c != 0) {
165 		ASSERT(zio_buf_cache[c] != NULL);
166 		if (zio_buf_cache[c - 1] == NULL)
167 			zio_buf_cache[c - 1] = zio_buf_cache[c];
168 
169 		ASSERT(zio_data_buf_cache[c] != NULL);
170 		if (zio_data_buf_cache[c - 1] == NULL)
171 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
172 	}
173 
174 	/*
175 	 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
176 	 * to fail 3 times per txg or 8 failures, whichever is greater.
177 	 */
178 	zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
179 
180 	zio_inject_init();
181 }
182 
183 void
184 zio_fini(void)
185 {
186 	size_t c;
187 	kmem_cache_t *last_cache = NULL;
188 	kmem_cache_t *last_data_cache = NULL;
189 
190 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
191 		if (zio_buf_cache[c] != last_cache) {
192 			last_cache = zio_buf_cache[c];
193 			kmem_cache_destroy(zio_buf_cache[c]);
194 		}
195 		zio_buf_cache[c] = NULL;
196 
197 		if (zio_data_buf_cache[c] != last_data_cache) {
198 			last_data_cache = zio_data_buf_cache[c];
199 			kmem_cache_destroy(zio_data_buf_cache[c]);
200 		}
201 		zio_data_buf_cache[c] = NULL;
202 	}
203 
204 	kmem_cache_destroy(zio_link_cache);
205 	kmem_cache_destroy(zio_cache);
206 
207 	zio_inject_fini();
208 }
209 
210 /*
211  * ==========================================================================
212  * Allocate and free I/O buffers
213  * ==========================================================================
214  */
215 
216 /*
217  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
218  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
219  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
220  * excess / transient data in-core during a crashdump.
221  */
222 void *
223 zio_buf_alloc(size_t size)
224 {
225 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
226 
227 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
228 
229 	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
230 }
231 
232 /*
233  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
234  * crashdump if the kernel panics.  This exists so that we will limit the amount
235  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
236  * of kernel heap dumped to disk when the kernel panics)
237  */
238 void *
239 zio_data_buf_alloc(size_t size)
240 {
241 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
242 
243 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
244 
245 	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
246 }
247 
248 void
249 zio_buf_free(void *buf, size_t size)
250 {
251 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
252 
253 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
254 
255 	kmem_cache_free(zio_buf_cache[c], buf);
256 }
257 
258 void
259 zio_data_buf_free(void *buf, size_t size)
260 {
261 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
262 
263 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
264 
265 	kmem_cache_free(zio_data_buf_cache[c], buf);
266 }
267 
268 /*
269  * ==========================================================================
270  * Push and pop I/O transform buffers
271  * ==========================================================================
272  */
273 static void
274 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
275 	zio_transform_func_t *transform)
276 {
277 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
278 
279 	zt->zt_orig_data = zio->io_data;
280 	zt->zt_orig_size = zio->io_size;
281 	zt->zt_bufsize = bufsize;
282 	zt->zt_transform = transform;
283 
284 	zt->zt_next = zio->io_transform_stack;
285 	zio->io_transform_stack = zt;
286 
287 	zio->io_data = data;
288 	zio->io_size = size;
289 }
290 
291 static void
292 zio_pop_transforms(zio_t *zio)
293 {
294 	zio_transform_t *zt;
295 
296 	while ((zt = zio->io_transform_stack) != NULL) {
297 		if (zt->zt_transform != NULL)
298 			zt->zt_transform(zio,
299 			    zt->zt_orig_data, zt->zt_orig_size);
300 
301 		if (zt->zt_bufsize != 0)
302 			zio_buf_free(zio->io_data, zt->zt_bufsize);
303 
304 		zio->io_data = zt->zt_orig_data;
305 		zio->io_size = zt->zt_orig_size;
306 		zio->io_transform_stack = zt->zt_next;
307 
308 		kmem_free(zt, sizeof (zio_transform_t));
309 	}
310 }
311 
312 /*
313  * ==========================================================================
314  * I/O transform callbacks for subblocks and decompression
315  * ==========================================================================
316  */
317 static void
318 zio_subblock(zio_t *zio, void *data, uint64_t size)
319 {
320 	ASSERT(zio->io_size > size);
321 
322 	if (zio->io_type == ZIO_TYPE_READ)
323 		bcopy(zio->io_data, data, size);
324 }
325 
326 static void
327 zio_decompress(zio_t *zio, void *data, uint64_t size)
328 {
329 	if (zio->io_error == 0 &&
330 	    zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
331 	    zio->io_data, data, zio->io_size, size) != 0)
332 		zio->io_error = EIO;
333 }
334 
335 /*
336  * ==========================================================================
337  * I/O parent/child relationships and pipeline interlocks
338  * ==========================================================================
339  */
340 /*
341  * NOTE - Callers to zio_walk_parents() and zio_walk_children must
342  *        continue calling these functions until they return NULL.
343  *        Otherwise, the next caller will pick up the list walk in
344  *        some indeterminate state.  (Otherwise every caller would
345  *        have to pass in a cookie to keep the state represented by
346  *        io_walk_link, which gets annoying.)
347  */
348 zio_t *
349 zio_walk_parents(zio_t *cio)
350 {
351 	zio_link_t *zl = cio->io_walk_link;
352 	list_t *pl = &cio->io_parent_list;
353 
354 	zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
355 	cio->io_walk_link = zl;
356 
357 	if (zl == NULL)
358 		return (NULL);
359 
360 	ASSERT(zl->zl_child == cio);
361 	return (zl->zl_parent);
362 }
363 
364 zio_t *
365 zio_walk_children(zio_t *pio)
366 {
367 	zio_link_t *zl = pio->io_walk_link;
368 	list_t *cl = &pio->io_child_list;
369 
370 	zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
371 	pio->io_walk_link = zl;
372 
373 	if (zl == NULL)
374 		return (NULL);
375 
376 	ASSERT(zl->zl_parent == pio);
377 	return (zl->zl_child);
378 }
379 
380 zio_t *
381 zio_unique_parent(zio_t *cio)
382 {
383 	zio_t *pio = zio_walk_parents(cio);
384 
385 	VERIFY(zio_walk_parents(cio) == NULL);
386 	return (pio);
387 }
388 
389 void
390 zio_add_child(zio_t *pio, zio_t *cio)
391 {
392 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
393 
394 	/*
395 	 * Logical I/Os can have logical, gang, or vdev children.
396 	 * Gang I/Os can have gang or vdev children.
397 	 * Vdev I/Os can only have vdev children.
398 	 * The following ASSERT captures all of these constraints.
399 	 */
400 	ASSERT(cio->io_child_type <= pio->io_child_type);
401 
402 	zl->zl_parent = pio;
403 	zl->zl_child = cio;
404 
405 	mutex_enter(&cio->io_lock);
406 	mutex_enter(&pio->io_lock);
407 
408 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
409 
410 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
411 		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
412 
413 	list_insert_head(&pio->io_child_list, zl);
414 	list_insert_head(&cio->io_parent_list, zl);
415 
416 	pio->io_child_count++;
417 	cio->io_parent_count++;
418 
419 	mutex_exit(&pio->io_lock);
420 	mutex_exit(&cio->io_lock);
421 }
422 
423 static void
424 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
425 {
426 	ASSERT(zl->zl_parent == pio);
427 	ASSERT(zl->zl_child == cio);
428 
429 	mutex_enter(&cio->io_lock);
430 	mutex_enter(&pio->io_lock);
431 
432 	list_remove(&pio->io_child_list, zl);
433 	list_remove(&cio->io_parent_list, zl);
434 
435 	pio->io_child_count--;
436 	cio->io_parent_count--;
437 
438 	mutex_exit(&pio->io_lock);
439 	mutex_exit(&cio->io_lock);
440 
441 	kmem_cache_free(zio_link_cache, zl);
442 }
443 
444 static boolean_t
445 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
446 {
447 	uint64_t *countp = &zio->io_children[child][wait];
448 	boolean_t waiting = B_FALSE;
449 
450 	mutex_enter(&zio->io_lock);
451 	ASSERT(zio->io_stall == NULL);
452 	if (*countp != 0) {
453 		zio->io_stage >>= 1;
454 		zio->io_stall = countp;
455 		waiting = B_TRUE;
456 	}
457 	mutex_exit(&zio->io_lock);
458 
459 	return (waiting);
460 }
461 
462 static void
463 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
464 {
465 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
466 	int *errorp = &pio->io_child_error[zio->io_child_type];
467 
468 	mutex_enter(&pio->io_lock);
469 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
470 		*errorp = zio_worst_error(*errorp, zio->io_error);
471 	pio->io_reexecute |= zio->io_reexecute;
472 	ASSERT3U(*countp, >, 0);
473 	if (--*countp == 0 && pio->io_stall == countp) {
474 		pio->io_stall = NULL;
475 		mutex_exit(&pio->io_lock);
476 		zio_execute(pio);
477 	} else {
478 		mutex_exit(&pio->io_lock);
479 	}
480 }
481 
482 static void
483 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
484 {
485 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
486 		zio->io_error = zio->io_child_error[c];
487 }
488 
489 /*
490  * ==========================================================================
491  * Create the various types of I/O (read, write, free, etc)
492  * ==========================================================================
493  */
494 static zio_t *
495 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
496     void *data, uint64_t size, zio_done_func_t *done, void *private,
497     zio_type_t type, int priority, enum zio_flag flags,
498     vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
499     enum zio_stage stage, enum zio_stage pipeline)
500 {
501 	zio_t *zio;
502 
503 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
504 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
505 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
506 
507 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
508 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
509 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
510 
511 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
512 	bzero(zio, sizeof (zio_t));
513 
514 	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
515 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
516 
517 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
518 	    offsetof(zio_link_t, zl_parent_node));
519 	list_create(&zio->io_child_list, sizeof (zio_link_t),
520 	    offsetof(zio_link_t, zl_child_node));
521 
522 	if (vd != NULL)
523 		zio->io_child_type = ZIO_CHILD_VDEV;
524 	else if (flags & ZIO_FLAG_GANG_CHILD)
525 		zio->io_child_type = ZIO_CHILD_GANG;
526 	else if (flags & ZIO_FLAG_DDT_CHILD)
527 		zio->io_child_type = ZIO_CHILD_DDT;
528 	else
529 		zio->io_child_type = ZIO_CHILD_LOGICAL;
530 
531 	if (bp != NULL) {
532 		zio->io_bp = (blkptr_t *)bp;
533 		zio->io_bp_copy = *bp;
534 		zio->io_bp_orig = *bp;
535 		if (type != ZIO_TYPE_WRITE ||
536 		    zio->io_child_type == ZIO_CHILD_DDT)
537 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
538 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
539 			zio->io_logical = zio;
540 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
541 			pipeline |= ZIO_GANG_STAGES;
542 	}
543 
544 	zio->io_spa = spa;
545 	zio->io_txg = txg;
546 	zio->io_done = done;
547 	zio->io_private = private;
548 	zio->io_type = type;
549 	zio->io_priority = priority;
550 	zio->io_vd = vd;
551 	zio->io_offset = offset;
552 	zio->io_orig_data = zio->io_data = data;
553 	zio->io_orig_size = zio->io_size = size;
554 	zio->io_orig_flags = zio->io_flags = flags;
555 	zio->io_orig_stage = zio->io_stage = stage;
556 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
557 
558 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
559 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
560 
561 	if (zb != NULL)
562 		zio->io_bookmark = *zb;
563 
564 	if (pio != NULL) {
565 		if (zio->io_logical == NULL)
566 			zio->io_logical = pio->io_logical;
567 		if (zio->io_child_type == ZIO_CHILD_GANG)
568 			zio->io_gang_leader = pio->io_gang_leader;
569 		zio_add_child(pio, zio);
570 	}
571 
572 	return (zio);
573 }
574 
575 static void
576 zio_destroy(zio_t *zio)
577 {
578 	list_destroy(&zio->io_parent_list);
579 	list_destroy(&zio->io_child_list);
580 	mutex_destroy(&zio->io_lock);
581 	cv_destroy(&zio->io_cv);
582 	kmem_cache_free(zio_cache, zio);
583 }
584 
585 zio_t *
586 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
587     void *private, enum zio_flag flags)
588 {
589 	zio_t *zio;
590 
591 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
592 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
593 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
594 
595 	return (zio);
596 }
597 
598 zio_t *
599 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
600 {
601 	return (zio_null(NULL, spa, NULL, done, private, flags));
602 }
603 
604 zio_t *
605 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
606     void *data, uint64_t size, zio_done_func_t *done, void *private,
607     int priority, enum zio_flag flags, const zbookmark_t *zb)
608 {
609 	zio_t *zio;
610 
611 	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
612 	    data, size, done, private,
613 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
614 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
615 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
616 
617 	return (zio);
618 }
619 
620 zio_t *
621 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
622     void *data, uint64_t size, const zio_prop_t *zp,
623     zio_done_func_t *ready, zio_done_func_t *done, void *private,
624     int priority, enum zio_flag flags, const zbookmark_t *zb)
625 {
626 	zio_t *zio;
627 
628 	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
629 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
630 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
631 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
632 	    DMU_OT_IS_VALID(zp->zp_type) &&
633 	    zp->zp_level < 32 &&
634 	    zp->zp_copies > 0 &&
635 	    zp->zp_copies <= spa_max_replication(spa) &&
636 	    zp->zp_dedup <= 1 &&
637 	    zp->zp_dedup_verify <= 1);
638 
639 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
640 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
641 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
642 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
643 
644 	zio->io_ready = ready;
645 	zio->io_prop = *zp;
646 
647 	return (zio);
648 }
649 
650 zio_t *
651 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
652     uint64_t size, zio_done_func_t *done, void *private, int priority,
653     enum zio_flag flags, zbookmark_t *zb)
654 {
655 	zio_t *zio;
656 
657 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
658 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
659 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
660 
661 	return (zio);
662 }
663 
664 void
665 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
666 {
667 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
668 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
669 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
670 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
671 
672 	zio->io_prop.zp_copies = copies;
673 	zio->io_bp_override = bp;
674 }
675 
676 void
677 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
678 {
679 	bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
680 }
681 
682 zio_t *
683 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
684     enum zio_flag flags)
685 {
686 	zio_t *zio;
687 
688 	dprintf_bp(bp, "freeing in txg %llu, pass %u",
689 	    (longlong_t)txg, spa->spa_sync_pass);
690 
691 	ASSERT(!BP_IS_HOLE(bp));
692 	ASSERT(spa_syncing_txg(spa) == txg);
693 	ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
694 
695 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
696 	    NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
697 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
698 
699 	return (zio);
700 }
701 
702 zio_t *
703 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
704     zio_done_func_t *done, void *private, enum zio_flag flags)
705 {
706 	zio_t *zio;
707 
708 	/*
709 	 * A claim is an allocation of a specific block.  Claims are needed
710 	 * to support immediate writes in the intent log.  The issue is that
711 	 * immediate writes contain committed data, but in a txg that was
712 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
713 	 * the intent log claims all blocks that contain immediate write data
714 	 * so that the SPA knows they're in use.
715 	 *
716 	 * All claims *must* be resolved in the first txg -- before the SPA
717 	 * starts allocating blocks -- so that nothing is allocated twice.
718 	 * If txg == 0 we just verify that the block is claimable.
719 	 */
720 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
721 	ASSERT(txg == spa_first_txg(spa) || txg == 0);
722 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
723 
724 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
725 	    done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
726 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
727 
728 	return (zio);
729 }
730 
731 zio_t *
732 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
733     zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
734 {
735 	zio_t *zio;
736 	int c;
737 
738 	if (vd->vdev_children == 0) {
739 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
740 		    ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
741 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
742 
743 		zio->io_cmd = cmd;
744 	} else {
745 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
746 
747 		for (c = 0; c < vd->vdev_children; c++)
748 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
749 			    done, private, priority, flags));
750 	}
751 
752 	return (zio);
753 }
754 
755 zio_t *
756 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
757     void *data, int checksum, zio_done_func_t *done, void *private,
758     int priority, enum zio_flag flags, boolean_t labels)
759 {
760 	zio_t *zio;
761 
762 	ASSERT(vd->vdev_children == 0);
763 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
764 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
765 	ASSERT3U(offset + size, <=, vd->vdev_psize);
766 
767 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
768 	    ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
769 	    ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
770 
771 	zio->io_prop.zp_checksum = checksum;
772 
773 	return (zio);
774 }
775 
776 zio_t *
777 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
778     void *data, int checksum, zio_done_func_t *done, void *private,
779     int priority, enum zio_flag flags, boolean_t labels)
780 {
781 	zio_t *zio;
782 
783 	ASSERT(vd->vdev_children == 0);
784 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
785 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
786 	ASSERT3U(offset + size, <=, vd->vdev_psize);
787 
788 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
789 	    ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
790 	    ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
791 
792 	zio->io_prop.zp_checksum = checksum;
793 
794 	if (zio_checksum_table[checksum].ci_eck) {
795 		/*
796 		 * zec checksums are necessarily destructive -- they modify
797 		 * the end of the write buffer to hold the verifier/checksum.
798 		 * Therefore, we must make a local copy in case the data is
799 		 * being written to multiple places in parallel.
800 		 */
801 		void *wbuf = zio_buf_alloc(size);
802 		bcopy(data, wbuf, size);
803 		zio_push_transform(zio, wbuf, size, size, NULL);
804 	}
805 
806 	return (zio);
807 }
808 
809 /*
810  * Create a child I/O to do some work for us.
811  */
812 zio_t *
813 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
814 	void *data, uint64_t size, int type, int priority, enum zio_flag flags,
815 	zio_done_func_t *done, void *private)
816 {
817 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
818 	zio_t *zio;
819 
820 	ASSERT(vd->vdev_parent ==
821 	    (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
822 
823 	if (type == ZIO_TYPE_READ && bp != NULL) {
824 		/*
825 		 * If we have the bp, then the child should perform the
826 		 * checksum and the parent need not.  This pushes error
827 		 * detection as close to the leaves as possible and
828 		 * eliminates redundant checksums in the interior nodes.
829 		 */
830 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
831 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
832 	}
833 
834 	if (vd->vdev_children == 0)
835 		offset += VDEV_LABEL_START_SIZE;
836 
837 	flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
838 
839 	/*
840 	 * If we've decided to do a repair, the write is not speculative --
841 	 * even if the original read was.
842 	 */
843 	if (flags & ZIO_FLAG_IO_REPAIR)
844 		flags &= ~ZIO_FLAG_SPECULATIVE;
845 
846 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
847 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
848 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
849 
850 	return (zio);
851 }
852 
853 zio_t *
854 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
855 	int type, int priority, enum zio_flag flags,
856 	zio_done_func_t *done, void *private)
857 {
858 	zio_t *zio;
859 
860 	ASSERT(vd->vdev_ops->vdev_op_leaf);
861 
862 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
863 	    data, size, done, private, type, priority,
864 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
865 	    vd, offset, NULL,
866 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
867 
868 	return (zio);
869 }
870 
871 void
872 zio_flush(zio_t *zio, vdev_t *vd)
873 {
874 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
875 	    NULL, NULL, ZIO_PRIORITY_NOW,
876 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
877 }
878 
879 void
880 zio_shrink(zio_t *zio, uint64_t size)
881 {
882 	ASSERT(zio->io_executor == NULL);
883 	ASSERT(zio->io_orig_size == zio->io_size);
884 	ASSERT(size <= zio->io_size);
885 
886 	/*
887 	 * We don't shrink for raidz because of problems with the
888 	 * reconstruction when reading back less than the block size.
889 	 * Note, BP_IS_RAIDZ() assumes no compression.
890 	 */
891 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
892 	if (!BP_IS_RAIDZ(zio->io_bp))
893 		zio->io_orig_size = zio->io_size = size;
894 }
895 
896 /*
897  * ==========================================================================
898  * Prepare to read and write logical blocks
899  * ==========================================================================
900  */
901 
902 static int
903 zio_read_bp_init(zio_t *zio)
904 {
905 	blkptr_t *bp = zio->io_bp;
906 
907 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
908 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
909 	    !(zio->io_flags & ZIO_FLAG_RAW)) {
910 		uint64_t psize = BP_GET_PSIZE(bp);
911 		void *cbuf = zio_buf_alloc(psize);
912 
913 		zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
914 	}
915 
916 	if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
917 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
918 
919 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
920 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
921 
922 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
923 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
924 
925 	return (ZIO_PIPELINE_CONTINUE);
926 }
927 
928 static int
929 zio_write_bp_init(zio_t *zio)
930 {
931 	spa_t *spa = zio->io_spa;
932 	zio_prop_t *zp = &zio->io_prop;
933 	enum zio_compress compress = zp->zp_compress;
934 	blkptr_t *bp = zio->io_bp;
935 	uint64_t lsize = zio->io_size;
936 	uint64_t psize = lsize;
937 	int pass = 1;
938 
939 	/*
940 	 * If our children haven't all reached the ready stage,
941 	 * wait for them and then repeat this pipeline stage.
942 	 */
943 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
944 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
945 		return (ZIO_PIPELINE_STOP);
946 
947 	if (!IO_IS_ALLOCATING(zio))
948 		return (ZIO_PIPELINE_CONTINUE);
949 
950 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
951 
952 	if (zio->io_bp_override) {
953 		ASSERT(bp->blk_birth != zio->io_txg);
954 		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
955 
956 		*bp = *zio->io_bp_override;
957 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
958 
959 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
960 			return (ZIO_PIPELINE_CONTINUE);
961 
962 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
963 		    zp->zp_dedup_verify);
964 
965 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
966 			BP_SET_DEDUP(bp, 1);
967 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
968 			return (ZIO_PIPELINE_CONTINUE);
969 		}
970 		zio->io_bp_override = NULL;
971 		BP_ZERO(bp);
972 	}
973 
974 	if (bp->blk_birth == zio->io_txg) {
975 		/*
976 		 * We're rewriting an existing block, which means we're
977 		 * working on behalf of spa_sync().  For spa_sync() to
978 		 * converge, it must eventually be the case that we don't
979 		 * have to allocate new blocks.  But compression changes
980 		 * the blocksize, which forces a reallocate, and makes
981 		 * convergence take longer.  Therefore, after the first
982 		 * few passes, stop compressing to ensure convergence.
983 		 */
984 		pass = spa_sync_pass(spa);
985 
986 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
987 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
988 		ASSERT(!BP_GET_DEDUP(bp));
989 
990 		if (pass > SYNC_PASS_DONT_COMPRESS)
991 			compress = ZIO_COMPRESS_OFF;
992 
993 		/* Make sure someone doesn't change their mind on overwrites */
994 		ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
995 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
996 	}
997 
998 	if (compress != ZIO_COMPRESS_OFF) {
999 		void *cbuf = zio_buf_alloc(lsize);
1000 		psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1001 		if (psize == 0 || psize == lsize) {
1002 			compress = ZIO_COMPRESS_OFF;
1003 			zio_buf_free(cbuf, lsize);
1004 		} else {
1005 			ASSERT(psize < lsize);
1006 			zio_push_transform(zio, cbuf, psize, lsize, NULL);
1007 		}
1008 	}
1009 
1010 	/*
1011 	 * The final pass of spa_sync() must be all rewrites, but the first
1012 	 * few passes offer a trade-off: allocating blocks defers convergence,
1013 	 * but newly allocated blocks are sequential, so they can be written
1014 	 * to disk faster.  Therefore, we allow the first few passes of
1015 	 * spa_sync() to allocate new blocks, but force rewrites after that.
1016 	 * There should only be a handful of blocks after pass 1 in any case.
1017 	 */
1018 	if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1019 	    pass > SYNC_PASS_REWRITE) {
1020 		ASSERT(psize != 0);
1021 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1022 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1023 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1024 	} else {
1025 		BP_ZERO(bp);
1026 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1027 	}
1028 
1029 	if (psize == 0) {
1030 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1031 	} else {
1032 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1033 		BP_SET_LSIZE(bp, lsize);
1034 		BP_SET_PSIZE(bp, psize);
1035 		BP_SET_COMPRESS(bp, compress);
1036 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1037 		BP_SET_TYPE(bp, zp->zp_type);
1038 		BP_SET_LEVEL(bp, zp->zp_level);
1039 		BP_SET_DEDUP(bp, zp->zp_dedup);
1040 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1041 		if (zp->zp_dedup) {
1042 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1043 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1044 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1045 		}
1046 	}
1047 
1048 	return (ZIO_PIPELINE_CONTINUE);
1049 }
1050 
1051 static int
1052 zio_free_bp_init(zio_t *zio)
1053 {
1054 	blkptr_t *bp = zio->io_bp;
1055 
1056 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1057 		if (BP_GET_DEDUP(bp))
1058 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1059 	}
1060 
1061 	return (ZIO_PIPELINE_CONTINUE);
1062 }
1063 
1064 /*
1065  * ==========================================================================
1066  * Execute the I/O pipeline
1067  * ==========================================================================
1068  */
1069 
1070 static void
1071 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1072 {
1073 	spa_t *spa = zio->io_spa;
1074 	zio_type_t t = zio->io_type;
1075 	int flags = (cutinline ? TQ_FRONT : 0);
1076 
1077 	/*
1078 	 * If we're a config writer or a probe, the normal issue and
1079 	 * interrupt threads may all be blocked waiting for the config lock.
1080 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1081 	 */
1082 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1083 		t = ZIO_TYPE_NULL;
1084 
1085 	/*
1086 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1087 	 */
1088 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1089 		t = ZIO_TYPE_NULL;
1090 
1091 	/*
1092 	 * If this is a high priority I/O, then use the high priority taskq.
1093 	 */
1094 	if (zio->io_priority == ZIO_PRIORITY_NOW &&
1095 	    spa->spa_zio_taskq[t][q + 1] != NULL)
1096 		q++;
1097 
1098 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1099 
1100 	/*
1101 	 * NB: We are assuming that the zio can only be dispatched
1102 	 * to a single taskq at a time.  It would be a grievous error
1103 	 * to dispatch the zio to another taskq at the same time.
1104 	 */
1105 	ASSERT(zio->io_tqent.tqent_next == NULL);
1106 	taskq_dispatch_ent(spa->spa_zio_taskq[t][q],
1107 	    (task_func_t *)zio_execute, zio, flags, &zio->io_tqent);
1108 }
1109 
1110 static boolean_t
1111 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1112 {
1113 	kthread_t *executor = zio->io_executor;
1114 	spa_t *spa = zio->io_spa;
1115 
1116 	for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1117 		if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1118 			return (B_TRUE);
1119 
1120 	return (B_FALSE);
1121 }
1122 
1123 static int
1124 zio_issue_async(zio_t *zio)
1125 {
1126 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1127 
1128 	return (ZIO_PIPELINE_STOP);
1129 }
1130 
1131 void
1132 zio_interrupt(zio_t *zio)
1133 {
1134 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1135 }
1136 
1137 /*
1138  * Execute the I/O pipeline until one of the following occurs:
1139  * (1) the I/O completes; (2) the pipeline stalls waiting for
1140  * dependent child I/Os; (3) the I/O issues, so we're waiting
1141  * for an I/O completion interrupt; (4) the I/O is delegated by
1142  * vdev-level caching or aggregation; (5) the I/O is deferred
1143  * due to vdev-level queueing; (6) the I/O is handed off to
1144  * another thread.  In all cases, the pipeline stops whenever
1145  * there's no CPU work; it never burns a thread in cv_wait().
1146  *
1147  * There's no locking on io_stage because there's no legitimate way
1148  * for multiple threads to be attempting to process the same I/O.
1149  */
1150 static zio_pipe_stage_t *zio_pipeline[];
1151 
1152 void
1153 zio_execute(zio_t *zio)
1154 {
1155 	zio->io_executor = curthread;
1156 
1157 	while (zio->io_stage < ZIO_STAGE_DONE) {
1158 		enum zio_stage pipeline = zio->io_pipeline;
1159 		enum zio_stage stage = zio->io_stage;
1160 		int rv;
1161 
1162 		ASSERT(!MUTEX_HELD(&zio->io_lock));
1163 		ASSERT(ISP2(stage));
1164 		ASSERT(zio->io_stall == NULL);
1165 
1166 		do {
1167 			stage <<= 1;
1168 		} while ((stage & pipeline) == 0);
1169 
1170 		ASSERT(stage <= ZIO_STAGE_DONE);
1171 
1172 		/*
1173 		 * If we are in interrupt context and this pipeline stage
1174 		 * will grab a config lock that is held across I/O,
1175 		 * or may wait for an I/O that needs an interrupt thread
1176 		 * to complete, issue async to avoid deadlock.
1177 		 *
1178 		 * For VDEV_IO_START, we cut in line so that the io will
1179 		 * be sent to disk promptly.
1180 		 */
1181 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1182 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1183 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1184 			    zio_requeue_io_start_cut_in_line : B_FALSE;
1185 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1186 			return;
1187 		}
1188 
1189 		zio->io_stage = stage;
1190 		rv = zio_pipeline[highbit(stage) - 1](zio);
1191 
1192 		if (rv == ZIO_PIPELINE_STOP)
1193 			return;
1194 
1195 		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1196 	}
1197 }
1198 
1199 /*
1200  * ==========================================================================
1201  * Initiate I/O, either sync or async
1202  * ==========================================================================
1203  */
1204 int
1205 zio_wait(zio_t *zio)
1206 {
1207 	int error;
1208 
1209 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1210 	ASSERT(zio->io_executor == NULL);
1211 
1212 	zio->io_waiter = curthread;
1213 
1214 	zio_execute(zio);
1215 
1216 	mutex_enter(&zio->io_lock);
1217 	while (zio->io_executor != NULL)
1218 		cv_wait(&zio->io_cv, &zio->io_lock);
1219 	mutex_exit(&zio->io_lock);
1220 
1221 	error = zio->io_error;
1222 	zio_destroy(zio);
1223 
1224 	return (error);
1225 }
1226 
1227 void
1228 zio_nowait(zio_t *zio)
1229 {
1230 	ASSERT(zio->io_executor == NULL);
1231 
1232 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1233 	    zio_unique_parent(zio) == NULL) {
1234 		/*
1235 		 * This is a logical async I/O with no parent to wait for it.
1236 		 * We add it to the spa_async_root_zio "Godfather" I/O which
1237 		 * will ensure they complete prior to unloading the pool.
1238 		 */
1239 		spa_t *spa = zio->io_spa;
1240 
1241 		zio_add_child(spa->spa_async_zio_root, zio);
1242 	}
1243 
1244 	zio_execute(zio);
1245 }
1246 
1247 /*
1248  * ==========================================================================
1249  * Reexecute or suspend/resume failed I/O
1250  * ==========================================================================
1251  */
1252 
1253 static void
1254 zio_reexecute(zio_t *pio)
1255 {
1256 	zio_t *cio, *cio_next;
1257 
1258 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1259 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1260 	ASSERT(pio->io_gang_leader == NULL);
1261 	ASSERT(pio->io_gang_tree == NULL);
1262 
1263 	pio->io_flags = pio->io_orig_flags;
1264 	pio->io_stage = pio->io_orig_stage;
1265 	pio->io_pipeline = pio->io_orig_pipeline;
1266 	pio->io_reexecute = 0;
1267 	pio->io_error = 0;
1268 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1269 		pio->io_state[w] = 0;
1270 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1271 		pio->io_child_error[c] = 0;
1272 
1273 	if (IO_IS_ALLOCATING(pio))
1274 		BP_ZERO(pio->io_bp);
1275 
1276 	/*
1277 	 * As we reexecute pio's children, new children could be created.
1278 	 * New children go to the head of pio's io_child_list, however,
1279 	 * so we will (correctly) not reexecute them.  The key is that
1280 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1281 	 * cannot be affected by any side effects of reexecuting 'cio'.
1282 	 */
1283 	for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1284 		cio_next = zio_walk_children(pio);
1285 		mutex_enter(&pio->io_lock);
1286 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1287 			pio->io_children[cio->io_child_type][w]++;
1288 		mutex_exit(&pio->io_lock);
1289 		zio_reexecute(cio);
1290 	}
1291 
1292 	/*
1293 	 * Now that all children have been reexecuted, execute the parent.
1294 	 * We don't reexecute "The Godfather" I/O here as it's the
1295 	 * responsibility of the caller to wait on him.
1296 	 */
1297 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1298 		zio_execute(pio);
1299 }
1300 
1301 void
1302 zio_suspend(spa_t *spa, zio_t *zio)
1303 {
1304 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1305 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1306 		    "failure and the failure mode property for this pool "
1307 		    "is set to panic.", spa_name(spa));
1308 
1309 	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1310 
1311 	mutex_enter(&spa->spa_suspend_lock);
1312 
1313 	if (spa->spa_suspend_zio_root == NULL)
1314 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1315 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1316 		    ZIO_FLAG_GODFATHER);
1317 
1318 	spa->spa_suspended = B_TRUE;
1319 
1320 	if (zio != NULL) {
1321 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1322 		ASSERT(zio != spa->spa_suspend_zio_root);
1323 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1324 		ASSERT(zio_unique_parent(zio) == NULL);
1325 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1326 		zio_add_child(spa->spa_suspend_zio_root, zio);
1327 	}
1328 
1329 	mutex_exit(&spa->spa_suspend_lock);
1330 }
1331 
1332 int
1333 zio_resume(spa_t *spa)
1334 {
1335 	zio_t *pio;
1336 
1337 	/*
1338 	 * Reexecute all previously suspended i/o.
1339 	 */
1340 	mutex_enter(&spa->spa_suspend_lock);
1341 	spa->spa_suspended = B_FALSE;
1342 	cv_broadcast(&spa->spa_suspend_cv);
1343 	pio = spa->spa_suspend_zio_root;
1344 	spa->spa_suspend_zio_root = NULL;
1345 	mutex_exit(&spa->spa_suspend_lock);
1346 
1347 	if (pio == NULL)
1348 		return (0);
1349 
1350 	zio_reexecute(pio);
1351 	return (zio_wait(pio));
1352 }
1353 
1354 void
1355 zio_resume_wait(spa_t *spa)
1356 {
1357 	mutex_enter(&spa->spa_suspend_lock);
1358 	while (spa_suspended(spa))
1359 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1360 	mutex_exit(&spa->spa_suspend_lock);
1361 }
1362 
1363 /*
1364  * ==========================================================================
1365  * Gang blocks.
1366  *
1367  * A gang block is a collection of small blocks that looks to the DMU
1368  * like one large block.  When zio_dva_allocate() cannot find a block
1369  * of the requested size, due to either severe fragmentation or the pool
1370  * being nearly full, it calls zio_write_gang_block() to construct the
1371  * block from smaller fragments.
1372  *
1373  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1374  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
1375  * an indirect block: it's an array of block pointers.  It consumes
1376  * only one sector and hence is allocatable regardless of fragmentation.
1377  * The gang header's bps point to its gang members, which hold the data.
1378  *
1379  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1380  * as the verifier to ensure uniqueness of the SHA256 checksum.
1381  * Critically, the gang block bp's blk_cksum is the checksum of the data,
1382  * not the gang header.  This ensures that data block signatures (needed for
1383  * deduplication) are independent of how the block is physically stored.
1384  *
1385  * Gang blocks can be nested: a gang member may itself be a gang block.
1386  * Thus every gang block is a tree in which root and all interior nodes are
1387  * gang headers, and the leaves are normal blocks that contain user data.
1388  * The root of the gang tree is called the gang leader.
1389  *
1390  * To perform any operation (read, rewrite, free, claim) on a gang block,
1391  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1392  * in the io_gang_tree field of the original logical i/o by recursively
1393  * reading the gang leader and all gang headers below it.  This yields
1394  * an in-core tree containing the contents of every gang header and the
1395  * bps for every constituent of the gang block.
1396  *
1397  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1398  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
1399  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1400  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1401  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1402  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
1403  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1404  * of the gang header plus zio_checksum_compute() of the data to update the
1405  * gang header's blk_cksum as described above.
1406  *
1407  * The two-phase assemble/issue model solves the problem of partial failure --
1408  * what if you'd freed part of a gang block but then couldn't read the
1409  * gang header for another part?  Assembling the entire gang tree first
1410  * ensures that all the necessary gang header I/O has succeeded before
1411  * starting the actual work of free, claim, or write.  Once the gang tree
1412  * is assembled, free and claim are in-memory operations that cannot fail.
1413  *
1414  * In the event that a gang write fails, zio_dva_unallocate() walks the
1415  * gang tree to immediately free (i.e. insert back into the space map)
1416  * everything we've allocated.  This ensures that we don't get ENOSPC
1417  * errors during repeated suspend/resume cycles due to a flaky device.
1418  *
1419  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
1420  * the gang tree, we won't modify the block, so we can safely defer the free
1421  * (knowing that the block is still intact).  If we *can* assemble the gang
1422  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1423  * each constituent bp and we can allocate a new block on the next sync pass.
1424  *
1425  * In all cases, the gang tree allows complete recovery from partial failure.
1426  * ==========================================================================
1427  */
1428 
1429 static zio_t *
1430 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1431 {
1432 	if (gn != NULL)
1433 		return (pio);
1434 
1435 	return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1436 	    NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1437 	    &pio->io_bookmark));
1438 }
1439 
1440 zio_t *
1441 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1442 {
1443 	zio_t *zio;
1444 
1445 	if (gn != NULL) {
1446 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1447 		    gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1448 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1449 		/*
1450 		 * As we rewrite each gang header, the pipeline will compute
1451 		 * a new gang block header checksum for it; but no one will
1452 		 * compute a new data checksum, so we do that here.  The one
1453 		 * exception is the gang leader: the pipeline already computed
1454 		 * its data checksum because that stage precedes gang assembly.
1455 		 * (Presently, nothing actually uses interior data checksums;
1456 		 * this is just good hygiene.)
1457 		 */
1458 		if (gn != pio->io_gang_leader->io_gang_tree) {
1459 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1460 			    data, BP_GET_PSIZE(bp));
1461 		}
1462 		/*
1463 		 * If we are here to damage data for testing purposes,
1464 		 * leave the GBH alone so that we can detect the damage.
1465 		 */
1466 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1467 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1468 	} else {
1469 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1470 		    data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1471 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1472 	}
1473 
1474 	return (zio);
1475 }
1476 
1477 /* ARGSUSED */
1478 zio_t *
1479 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1480 {
1481 	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1482 	    ZIO_GANG_CHILD_FLAGS(pio)));
1483 }
1484 
1485 /* ARGSUSED */
1486 zio_t *
1487 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1488 {
1489 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1490 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1491 }
1492 
1493 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1494 	NULL,
1495 	zio_read_gang,
1496 	zio_rewrite_gang,
1497 	zio_free_gang,
1498 	zio_claim_gang,
1499 	NULL
1500 };
1501 
1502 static void zio_gang_tree_assemble_done(zio_t *zio);
1503 
1504 static zio_gang_node_t *
1505 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1506 {
1507 	zio_gang_node_t *gn;
1508 
1509 	ASSERT(*gnpp == NULL);
1510 
1511 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1512 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1513 	*gnpp = gn;
1514 
1515 	return (gn);
1516 }
1517 
1518 static void
1519 zio_gang_node_free(zio_gang_node_t **gnpp)
1520 {
1521 	zio_gang_node_t *gn = *gnpp;
1522 
1523 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1524 		ASSERT(gn->gn_child[g] == NULL);
1525 
1526 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1527 	kmem_free(gn, sizeof (*gn));
1528 	*gnpp = NULL;
1529 }
1530 
1531 static void
1532 zio_gang_tree_free(zio_gang_node_t **gnpp)
1533 {
1534 	zio_gang_node_t *gn = *gnpp;
1535 
1536 	if (gn == NULL)
1537 		return;
1538 
1539 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1540 		zio_gang_tree_free(&gn->gn_child[g]);
1541 
1542 	zio_gang_node_free(gnpp);
1543 }
1544 
1545 static void
1546 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1547 {
1548 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1549 
1550 	ASSERT(gio->io_gang_leader == gio);
1551 	ASSERT(BP_IS_GANG(bp));
1552 
1553 	zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1554 	    SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1555 	    gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1556 }
1557 
1558 static void
1559 zio_gang_tree_assemble_done(zio_t *zio)
1560 {
1561 	zio_t *gio = zio->io_gang_leader;
1562 	zio_gang_node_t *gn = zio->io_private;
1563 	blkptr_t *bp = zio->io_bp;
1564 
1565 	ASSERT(gio == zio_unique_parent(zio));
1566 	ASSERT(zio->io_child_count == 0);
1567 
1568 	if (zio->io_error)
1569 		return;
1570 
1571 	if (BP_SHOULD_BYTESWAP(bp))
1572 		byteswap_uint64_array(zio->io_data, zio->io_size);
1573 
1574 	ASSERT(zio->io_data == gn->gn_gbh);
1575 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1576 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1577 
1578 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1579 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1580 		if (!BP_IS_GANG(gbp))
1581 			continue;
1582 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1583 	}
1584 }
1585 
1586 static void
1587 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1588 {
1589 	zio_t *gio = pio->io_gang_leader;
1590 	zio_t *zio;
1591 
1592 	ASSERT(BP_IS_GANG(bp) == !!gn);
1593 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1594 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1595 
1596 	/*
1597 	 * If you're a gang header, your data is in gn->gn_gbh.
1598 	 * If you're a gang member, your data is in 'data' and gn == NULL.
1599 	 */
1600 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1601 
1602 	if (gn != NULL) {
1603 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1604 
1605 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1606 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1607 			if (BP_IS_HOLE(gbp))
1608 				continue;
1609 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1610 			data = (char *)data + BP_GET_PSIZE(gbp);
1611 		}
1612 	}
1613 
1614 	if (gn == gio->io_gang_tree)
1615 		ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1616 
1617 	if (zio != pio)
1618 		zio_nowait(zio);
1619 }
1620 
1621 static int
1622 zio_gang_assemble(zio_t *zio)
1623 {
1624 	blkptr_t *bp = zio->io_bp;
1625 
1626 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1627 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1628 
1629 	zio->io_gang_leader = zio;
1630 
1631 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1632 
1633 	return (ZIO_PIPELINE_CONTINUE);
1634 }
1635 
1636 static int
1637 zio_gang_issue(zio_t *zio)
1638 {
1639 	blkptr_t *bp = zio->io_bp;
1640 
1641 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1642 		return (ZIO_PIPELINE_STOP);
1643 
1644 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1645 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1646 
1647 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1648 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1649 	else
1650 		zio_gang_tree_free(&zio->io_gang_tree);
1651 
1652 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1653 
1654 	return (ZIO_PIPELINE_CONTINUE);
1655 }
1656 
1657 static void
1658 zio_write_gang_member_ready(zio_t *zio)
1659 {
1660 	zio_t *pio = zio_unique_parent(zio);
1661 	zio_t *gio = zio->io_gang_leader;
1662 	dva_t *cdva = zio->io_bp->blk_dva;
1663 	dva_t *pdva = pio->io_bp->blk_dva;
1664 	uint64_t asize;
1665 
1666 	if (BP_IS_HOLE(zio->io_bp))
1667 		return;
1668 
1669 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1670 
1671 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1672 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1673 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1674 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1675 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1676 
1677 	mutex_enter(&pio->io_lock);
1678 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1679 		ASSERT(DVA_GET_GANG(&pdva[d]));
1680 		asize = DVA_GET_ASIZE(&pdva[d]);
1681 		asize += DVA_GET_ASIZE(&cdva[d]);
1682 		DVA_SET_ASIZE(&pdva[d], asize);
1683 	}
1684 	mutex_exit(&pio->io_lock);
1685 }
1686 
1687 static int
1688 zio_write_gang_block(zio_t *pio)
1689 {
1690 	spa_t *spa = pio->io_spa;
1691 	blkptr_t *bp = pio->io_bp;
1692 	zio_t *gio = pio->io_gang_leader;
1693 	zio_t *zio;
1694 	zio_gang_node_t *gn, **gnpp;
1695 	zio_gbh_phys_t *gbh;
1696 	uint64_t txg = pio->io_txg;
1697 	uint64_t resid = pio->io_size;
1698 	uint64_t lsize;
1699 	int copies = gio->io_prop.zp_copies;
1700 	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1701 	zio_prop_t zp;
1702 	int error;
1703 
1704 	error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1705 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1706 	    METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1707 	if (error) {
1708 		pio->io_error = error;
1709 		return (ZIO_PIPELINE_CONTINUE);
1710 	}
1711 
1712 	if (pio == gio) {
1713 		gnpp = &gio->io_gang_tree;
1714 	} else {
1715 		gnpp = pio->io_private;
1716 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
1717 	}
1718 
1719 	gn = zio_gang_node_alloc(gnpp);
1720 	gbh = gn->gn_gbh;
1721 	bzero(gbh, SPA_GANGBLOCKSIZE);
1722 
1723 	/*
1724 	 * Create the gang header.
1725 	 */
1726 	zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1727 	    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1728 
1729 	/*
1730 	 * Create and nowait the gang children.
1731 	 */
1732 	for (int g = 0; resid != 0; resid -= lsize, g++) {
1733 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1734 		    SPA_MINBLOCKSIZE);
1735 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1736 
1737 		zp.zp_checksum = gio->io_prop.zp_checksum;
1738 		zp.zp_compress = ZIO_COMPRESS_OFF;
1739 		zp.zp_type = DMU_OT_NONE;
1740 		zp.zp_level = 0;
1741 		zp.zp_copies = gio->io_prop.zp_copies;
1742 		zp.zp_dedup = 0;
1743 		zp.zp_dedup_verify = 0;
1744 
1745 		zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1746 		    (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1747 		    zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1748 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1749 		    &pio->io_bookmark));
1750 	}
1751 
1752 	/*
1753 	 * Set pio's pipeline to just wait for zio to finish.
1754 	 */
1755 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1756 
1757 	zio_nowait(zio);
1758 
1759 	return (ZIO_PIPELINE_CONTINUE);
1760 }
1761 
1762 /*
1763  * ==========================================================================
1764  * Dedup
1765  * ==========================================================================
1766  */
1767 static void
1768 zio_ddt_child_read_done(zio_t *zio)
1769 {
1770 	blkptr_t *bp = zio->io_bp;
1771 	ddt_entry_t *dde = zio->io_private;
1772 	ddt_phys_t *ddp;
1773 	zio_t *pio = zio_unique_parent(zio);
1774 
1775 	mutex_enter(&pio->io_lock);
1776 	ddp = ddt_phys_select(dde, bp);
1777 	if (zio->io_error == 0)
1778 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
1779 	if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1780 		dde->dde_repair_data = zio->io_data;
1781 	else
1782 		zio_buf_free(zio->io_data, zio->io_size);
1783 	mutex_exit(&pio->io_lock);
1784 }
1785 
1786 static int
1787 zio_ddt_read_start(zio_t *zio)
1788 {
1789 	blkptr_t *bp = zio->io_bp;
1790 
1791 	ASSERT(BP_GET_DEDUP(bp));
1792 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1793 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1794 
1795 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1796 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1797 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1798 		ddt_phys_t *ddp = dde->dde_phys;
1799 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1800 		blkptr_t blk;
1801 
1802 		ASSERT(zio->io_vsd == NULL);
1803 		zio->io_vsd = dde;
1804 
1805 		if (ddp_self == NULL)
1806 			return (ZIO_PIPELINE_CONTINUE);
1807 
1808 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1809 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1810 				continue;
1811 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1812 			    &blk);
1813 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
1814 			    zio_buf_alloc(zio->io_size), zio->io_size,
1815 			    zio_ddt_child_read_done, dde, zio->io_priority,
1816 			    ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1817 			    &zio->io_bookmark));
1818 		}
1819 		return (ZIO_PIPELINE_CONTINUE);
1820 	}
1821 
1822 	zio_nowait(zio_read(zio, zio->io_spa, bp,
1823 	    zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1824 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1825 
1826 	return (ZIO_PIPELINE_CONTINUE);
1827 }
1828 
1829 static int
1830 zio_ddt_read_done(zio_t *zio)
1831 {
1832 	blkptr_t *bp = zio->io_bp;
1833 
1834 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1835 		return (ZIO_PIPELINE_STOP);
1836 
1837 	ASSERT(BP_GET_DEDUP(bp));
1838 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1839 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1840 
1841 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1842 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1843 		ddt_entry_t *dde = zio->io_vsd;
1844 		if (ddt == NULL) {
1845 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1846 			return (ZIO_PIPELINE_CONTINUE);
1847 		}
1848 		if (dde == NULL) {
1849 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1850 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1851 			return (ZIO_PIPELINE_STOP);
1852 		}
1853 		if (dde->dde_repair_data != NULL) {
1854 			bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1855 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
1856 		}
1857 		ddt_repair_done(ddt, dde);
1858 		zio->io_vsd = NULL;
1859 	}
1860 
1861 	ASSERT(zio->io_vsd == NULL);
1862 
1863 	return (ZIO_PIPELINE_CONTINUE);
1864 }
1865 
1866 static boolean_t
1867 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1868 {
1869 	spa_t *spa = zio->io_spa;
1870 
1871 	/*
1872 	 * Note: we compare the original data, not the transformed data,
1873 	 * because when zio->io_bp is an override bp, we will not have
1874 	 * pushed the I/O transforms.  That's an important optimization
1875 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1876 	 */
1877 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1878 		zio_t *lio = dde->dde_lead_zio[p];
1879 
1880 		if (lio != NULL) {
1881 			return (lio->io_orig_size != zio->io_orig_size ||
1882 			    bcmp(zio->io_orig_data, lio->io_orig_data,
1883 			    zio->io_orig_size) != 0);
1884 		}
1885 	}
1886 
1887 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1888 		ddt_phys_t *ddp = &dde->dde_phys[p];
1889 
1890 		if (ddp->ddp_phys_birth != 0) {
1891 			arc_buf_t *abuf = NULL;
1892 			uint32_t aflags = ARC_WAIT;
1893 			blkptr_t blk = *zio->io_bp;
1894 			int error;
1895 
1896 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1897 
1898 			ddt_exit(ddt);
1899 
1900 			error = arc_read_nolock(NULL, spa, &blk,
1901 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1902 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1903 			    &aflags, &zio->io_bookmark);
1904 
1905 			if (error == 0) {
1906 				if (arc_buf_size(abuf) != zio->io_orig_size ||
1907 				    bcmp(abuf->b_data, zio->io_orig_data,
1908 				    zio->io_orig_size) != 0)
1909 					error = EEXIST;
1910 				VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1911 			}
1912 
1913 			ddt_enter(ddt);
1914 			return (error != 0);
1915 		}
1916 	}
1917 
1918 	return (B_FALSE);
1919 }
1920 
1921 static void
1922 zio_ddt_child_write_ready(zio_t *zio)
1923 {
1924 	int p = zio->io_prop.zp_copies;
1925 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1926 	ddt_entry_t *dde = zio->io_private;
1927 	ddt_phys_t *ddp = &dde->dde_phys[p];
1928 	zio_t *pio;
1929 
1930 	if (zio->io_error)
1931 		return;
1932 
1933 	ddt_enter(ddt);
1934 
1935 	ASSERT(dde->dde_lead_zio[p] == zio);
1936 
1937 	ddt_phys_fill(ddp, zio->io_bp);
1938 
1939 	while ((pio = zio_walk_parents(zio)) != NULL)
1940 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1941 
1942 	ddt_exit(ddt);
1943 }
1944 
1945 static void
1946 zio_ddt_child_write_done(zio_t *zio)
1947 {
1948 	int p = zio->io_prop.zp_copies;
1949 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1950 	ddt_entry_t *dde = zio->io_private;
1951 	ddt_phys_t *ddp = &dde->dde_phys[p];
1952 
1953 	ddt_enter(ddt);
1954 
1955 	ASSERT(ddp->ddp_refcnt == 0);
1956 	ASSERT(dde->dde_lead_zio[p] == zio);
1957 	dde->dde_lead_zio[p] = NULL;
1958 
1959 	if (zio->io_error == 0) {
1960 		while (zio_walk_parents(zio) != NULL)
1961 			ddt_phys_addref(ddp);
1962 	} else {
1963 		ddt_phys_clear(ddp);
1964 	}
1965 
1966 	ddt_exit(ddt);
1967 }
1968 
1969 static void
1970 zio_ddt_ditto_write_done(zio_t *zio)
1971 {
1972 	int p = DDT_PHYS_DITTO;
1973 	zio_prop_t *zp = &zio->io_prop;
1974 	blkptr_t *bp = zio->io_bp;
1975 	ddt_t *ddt = ddt_select(zio->io_spa, bp);
1976 	ddt_entry_t *dde = zio->io_private;
1977 	ddt_phys_t *ddp = &dde->dde_phys[p];
1978 	ddt_key_t *ddk = &dde->dde_key;
1979 
1980 	ddt_enter(ddt);
1981 
1982 	ASSERT(ddp->ddp_refcnt == 0);
1983 	ASSERT(dde->dde_lead_zio[p] == zio);
1984 	dde->dde_lead_zio[p] = NULL;
1985 
1986 	if (zio->io_error == 0) {
1987 		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1988 		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1989 		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1990 		if (ddp->ddp_phys_birth != 0)
1991 			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1992 		ddt_phys_fill(ddp, bp);
1993 	}
1994 
1995 	ddt_exit(ddt);
1996 }
1997 
1998 static int
1999 zio_ddt_write(zio_t *zio)
2000 {
2001 	spa_t *spa = zio->io_spa;
2002 	blkptr_t *bp = zio->io_bp;
2003 	uint64_t txg = zio->io_txg;
2004 	zio_prop_t *zp = &zio->io_prop;
2005 	int p = zp->zp_copies;
2006 	int ditto_copies;
2007 	zio_t *cio = NULL;
2008 	zio_t *dio = NULL;
2009 	ddt_t *ddt = ddt_select(spa, bp);
2010 	ddt_entry_t *dde;
2011 	ddt_phys_t *ddp;
2012 
2013 	ASSERT(BP_GET_DEDUP(bp));
2014 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2015 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2016 
2017 	ddt_enter(ddt);
2018 	dde = ddt_lookup(ddt, bp, B_TRUE);
2019 	ddp = &dde->dde_phys[p];
2020 
2021 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2022 		/*
2023 		 * If we're using a weak checksum, upgrade to a strong checksum
2024 		 * and try again.  If we're already using a strong checksum,
2025 		 * we can't resolve it, so just convert to an ordinary write.
2026 		 * (And automatically e-mail a paper to Nature?)
2027 		 */
2028 		if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2029 			zp->zp_checksum = spa_dedup_checksum(spa);
2030 			zio_pop_transforms(zio);
2031 			zio->io_stage = ZIO_STAGE_OPEN;
2032 			BP_ZERO(bp);
2033 		} else {
2034 			zp->zp_dedup = 0;
2035 		}
2036 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2037 		ddt_exit(ddt);
2038 		return (ZIO_PIPELINE_CONTINUE);
2039 	}
2040 
2041 	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2042 	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2043 
2044 	if (ditto_copies > ddt_ditto_copies_present(dde) &&
2045 	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2046 		zio_prop_t czp = *zp;
2047 
2048 		czp.zp_copies = ditto_copies;
2049 
2050 		/*
2051 		 * If we arrived here with an override bp, we won't have run
2052 		 * the transform stack, so we won't have the data we need to
2053 		 * generate a child i/o.  So, toss the override bp and restart.
2054 		 * This is safe, because using the override bp is just an
2055 		 * optimization; and it's rare, so the cost doesn't matter.
2056 		 */
2057 		if (zio->io_bp_override) {
2058 			zio_pop_transforms(zio);
2059 			zio->io_stage = ZIO_STAGE_OPEN;
2060 			zio->io_pipeline = ZIO_WRITE_PIPELINE;
2061 			zio->io_bp_override = NULL;
2062 			BP_ZERO(bp);
2063 			ddt_exit(ddt);
2064 			return (ZIO_PIPELINE_CONTINUE);
2065 		}
2066 
2067 		dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2068 		    zio->io_orig_size, &czp, NULL,
2069 		    zio_ddt_ditto_write_done, dde, zio->io_priority,
2070 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2071 
2072 		zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2073 		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2074 	}
2075 
2076 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2077 		if (ddp->ddp_phys_birth != 0)
2078 			ddt_bp_fill(ddp, bp, txg);
2079 		if (dde->dde_lead_zio[p] != NULL)
2080 			zio_add_child(zio, dde->dde_lead_zio[p]);
2081 		else
2082 			ddt_phys_addref(ddp);
2083 	} else if (zio->io_bp_override) {
2084 		ASSERT(bp->blk_birth == txg);
2085 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2086 		ddt_phys_fill(ddp, bp);
2087 		ddt_phys_addref(ddp);
2088 	} else {
2089 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2090 		    zio->io_orig_size, zp, zio_ddt_child_write_ready,
2091 		    zio_ddt_child_write_done, dde, zio->io_priority,
2092 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2093 
2094 		zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2095 		dde->dde_lead_zio[p] = cio;
2096 	}
2097 
2098 	ddt_exit(ddt);
2099 
2100 	if (cio)
2101 		zio_nowait(cio);
2102 	if (dio)
2103 		zio_nowait(dio);
2104 
2105 	return (ZIO_PIPELINE_CONTINUE);
2106 }
2107 
2108 ddt_entry_t *freedde; /* for debugging */
2109 
2110 static int
2111 zio_ddt_free(zio_t *zio)
2112 {
2113 	spa_t *spa = zio->io_spa;
2114 	blkptr_t *bp = zio->io_bp;
2115 	ddt_t *ddt = ddt_select(spa, bp);
2116 	ddt_entry_t *dde;
2117 	ddt_phys_t *ddp;
2118 
2119 	ASSERT(BP_GET_DEDUP(bp));
2120 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2121 
2122 	ddt_enter(ddt);
2123 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2124 	ddp = ddt_phys_select(dde, bp);
2125 	ddt_phys_decref(ddp);
2126 	ddt_exit(ddt);
2127 
2128 	return (ZIO_PIPELINE_CONTINUE);
2129 }
2130 
2131 /*
2132  * ==========================================================================
2133  * Allocate and free blocks
2134  * ==========================================================================
2135  */
2136 static int
2137 zio_dva_allocate(zio_t *zio)
2138 {
2139 	spa_t *spa = zio->io_spa;
2140 	metaslab_class_t *mc = spa_normal_class(spa);
2141 	blkptr_t *bp = zio->io_bp;
2142 	int error;
2143 	int flags = 0;
2144 
2145 	if (zio->io_gang_leader == NULL) {
2146 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2147 		zio->io_gang_leader = zio;
2148 	}
2149 
2150 	ASSERT(BP_IS_HOLE(bp));
2151 	ASSERT0(BP_GET_NDVAS(bp));
2152 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
2153 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2154 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2155 
2156 	/*
2157 	 * The dump device does not support gang blocks so allocation on
2158 	 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2159 	 * the "fast" gang feature.
2160 	 */
2161 	flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2162 	flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2163 	    METASLAB_GANG_CHILD : 0;
2164 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
2165 	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2166 
2167 	if (error) {
2168 		spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2169 		    "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2170 		    error);
2171 		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2172 			return (zio_write_gang_block(zio));
2173 		zio->io_error = error;
2174 	}
2175 
2176 	return (ZIO_PIPELINE_CONTINUE);
2177 }
2178 
2179 static int
2180 zio_dva_free(zio_t *zio)
2181 {
2182 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2183 
2184 	return (ZIO_PIPELINE_CONTINUE);
2185 }
2186 
2187 static int
2188 zio_dva_claim(zio_t *zio)
2189 {
2190 	int error;
2191 
2192 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2193 	if (error)
2194 		zio->io_error = error;
2195 
2196 	return (ZIO_PIPELINE_CONTINUE);
2197 }
2198 
2199 /*
2200  * Undo an allocation.  This is used by zio_done() when an I/O fails
2201  * and we want to give back the block we just allocated.
2202  * This handles both normal blocks and gang blocks.
2203  */
2204 static void
2205 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2206 {
2207 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2208 	ASSERT(zio->io_bp_override == NULL);
2209 
2210 	if (!BP_IS_HOLE(bp))
2211 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2212 
2213 	if (gn != NULL) {
2214 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2215 			zio_dva_unallocate(zio, gn->gn_child[g],
2216 			    &gn->gn_gbh->zg_blkptr[g]);
2217 		}
2218 	}
2219 }
2220 
2221 /*
2222  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
2223  */
2224 int
2225 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2226     uint64_t size, boolean_t use_slog)
2227 {
2228 	int error = 1;
2229 
2230 	ASSERT(txg > spa_syncing_txg(spa));
2231 
2232 	/*
2233 	 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2234 	 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2235 	 * when allocating them.
2236 	 */
2237 	if (use_slog) {
2238 		error = metaslab_alloc(spa, spa_log_class(spa), size,
2239 		    new_bp, 1, txg, old_bp,
2240 		    METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2241 	}
2242 
2243 	if (error) {
2244 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
2245 		    new_bp, 1, txg, old_bp,
2246 		    METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2247 	}
2248 
2249 	if (error == 0) {
2250 		BP_SET_LSIZE(new_bp, size);
2251 		BP_SET_PSIZE(new_bp, size);
2252 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2253 		BP_SET_CHECKSUM(new_bp,
2254 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2255 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2256 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2257 		BP_SET_LEVEL(new_bp, 0);
2258 		BP_SET_DEDUP(new_bp, 0);
2259 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2260 	}
2261 
2262 	return (error);
2263 }
2264 
2265 /*
2266  * Free an intent log block.
2267  */
2268 void
2269 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2270 {
2271 	ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2272 	ASSERT(!BP_IS_GANG(bp));
2273 
2274 	zio_free(spa, txg, bp);
2275 }
2276 
2277 /*
2278  * ==========================================================================
2279  * Read and write to physical devices
2280  * ==========================================================================
2281  */
2282 static int
2283 zio_vdev_io_start(zio_t *zio)
2284 {
2285 	vdev_t *vd = zio->io_vd;
2286 	uint64_t align;
2287 	spa_t *spa = zio->io_spa;
2288 
2289 	ASSERT(zio->io_error == 0);
2290 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2291 
2292 	if (vd == NULL) {
2293 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2294 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2295 
2296 		/*
2297 		 * The mirror_ops handle multiple DVAs in a single BP.
2298 		 */
2299 		return (vdev_mirror_ops.vdev_op_io_start(zio));
2300 	}
2301 
2302 	/*
2303 	 * We keep track of time-sensitive I/Os so that the scan thread
2304 	 * can quickly react to certain workloads.  In particular, we care
2305 	 * about non-scrubbing, top-level reads and writes with the following
2306 	 * characteristics:
2307 	 * 	- synchronous writes of user data to non-slog devices
2308 	 *	- any reads of user data
2309 	 * When these conditions are met, adjust the timestamp of spa_last_io
2310 	 * which allows the scan thread to adjust its workload accordingly.
2311 	 */
2312 	if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2313 	    vd == vd->vdev_top && !vd->vdev_islog &&
2314 	    zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2315 	    zio->io_txg != spa_syncing_txg(spa)) {
2316 		uint64_t old = spa->spa_last_io;
2317 		uint64_t new = ddi_get_lbolt64();
2318 		if (old != new)
2319 			(void) atomic_cas_64(&spa->spa_last_io, old, new);
2320 	}
2321 
2322 	align = 1ULL << vd->vdev_top->vdev_ashift;
2323 
2324 	if (P2PHASE(zio->io_size, align) != 0) {
2325 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
2326 		char *abuf = zio_buf_alloc(asize);
2327 		ASSERT(vd == vd->vdev_top);
2328 		if (zio->io_type == ZIO_TYPE_WRITE) {
2329 			bcopy(zio->io_data, abuf, zio->io_size);
2330 			bzero(abuf + zio->io_size, asize - zio->io_size);
2331 		}
2332 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2333 	}
2334 
2335 	ASSERT(P2PHASE(zio->io_offset, align) == 0);
2336 	ASSERT(P2PHASE(zio->io_size, align) == 0);
2337 	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2338 
2339 	/*
2340 	 * If this is a repair I/O, and there's no self-healing involved --
2341 	 * that is, we're just resilvering what we expect to resilver --
2342 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2343 	 * This prevents spurious resilvering with nested replication.
2344 	 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2345 	 * A is out of date, we'll read from C+D, then use the data to
2346 	 * resilver A+B -- but we don't actually want to resilver B, just A.
2347 	 * The top-level mirror has no way to know this, so instead we just
2348 	 * discard unnecessary repairs as we work our way down the vdev tree.
2349 	 * The same logic applies to any form of nested replication:
2350 	 * ditto + mirror, RAID-Z + replacing, etc.  This covers them all.
2351 	 */
2352 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2353 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2354 	    zio->io_txg != 0 &&	/* not a delegated i/o */
2355 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2356 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2357 		zio_vdev_io_bypass(zio);
2358 		return (ZIO_PIPELINE_CONTINUE);
2359 	}
2360 
2361 	if (vd->vdev_ops->vdev_op_leaf &&
2362 	    (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2363 
2364 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2365 			return (ZIO_PIPELINE_CONTINUE);
2366 
2367 		if ((zio = vdev_queue_io(zio)) == NULL)
2368 			return (ZIO_PIPELINE_STOP);
2369 
2370 		if (!vdev_accessible(vd, zio)) {
2371 			zio->io_error = ENXIO;
2372 			zio_interrupt(zio);
2373 			return (ZIO_PIPELINE_STOP);
2374 		}
2375 	}
2376 
2377 	return (vd->vdev_ops->vdev_op_io_start(zio));
2378 }
2379 
2380 static int
2381 zio_vdev_io_done(zio_t *zio)
2382 {
2383 	vdev_t *vd = zio->io_vd;
2384 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2385 	boolean_t unexpected_error = B_FALSE;
2386 
2387 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2388 		return (ZIO_PIPELINE_STOP);
2389 
2390 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2391 
2392 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2393 
2394 		vdev_queue_io_done(zio);
2395 
2396 		if (zio->io_type == ZIO_TYPE_WRITE)
2397 			vdev_cache_write(zio);
2398 
2399 		if (zio_injection_enabled && zio->io_error == 0)
2400 			zio->io_error = zio_handle_device_injection(vd,
2401 			    zio, EIO);
2402 
2403 		if (zio_injection_enabled && zio->io_error == 0)
2404 			zio->io_error = zio_handle_label_injection(zio, EIO);
2405 
2406 		if (zio->io_error) {
2407 			if (!vdev_accessible(vd, zio)) {
2408 				zio->io_error = ENXIO;
2409 			} else {
2410 				unexpected_error = B_TRUE;
2411 			}
2412 		}
2413 	}
2414 
2415 	ops->vdev_op_io_done(zio);
2416 
2417 	if (unexpected_error)
2418 		VERIFY(vdev_probe(vd, zio) == NULL);
2419 
2420 	return (ZIO_PIPELINE_CONTINUE);
2421 }
2422 
2423 /*
2424  * For non-raidz ZIOs, we can just copy aside the bad data read from the
2425  * disk, and use that to finish the checksum ereport later.
2426  */
2427 static void
2428 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2429     const void *good_buf)
2430 {
2431 	/* no processing needed */
2432 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2433 }
2434 
2435 /*ARGSUSED*/
2436 void
2437 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2438 {
2439 	void *buf = zio_buf_alloc(zio->io_size);
2440 
2441 	bcopy(zio->io_data, buf, zio->io_size);
2442 
2443 	zcr->zcr_cbinfo = zio->io_size;
2444 	zcr->zcr_cbdata = buf;
2445 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
2446 	zcr->zcr_free = zio_buf_free;
2447 }
2448 
2449 static int
2450 zio_vdev_io_assess(zio_t *zio)
2451 {
2452 	vdev_t *vd = zio->io_vd;
2453 
2454 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2455 		return (ZIO_PIPELINE_STOP);
2456 
2457 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2458 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2459 
2460 	if (zio->io_vsd != NULL) {
2461 		zio->io_vsd_ops->vsd_free(zio);
2462 		zio->io_vsd = NULL;
2463 	}
2464 
2465 	if (zio_injection_enabled && zio->io_error == 0)
2466 		zio->io_error = zio_handle_fault_injection(zio, EIO);
2467 
2468 	/*
2469 	 * If the I/O failed, determine whether we should attempt to retry it.
2470 	 *
2471 	 * On retry, we cut in line in the issue queue, since we don't want
2472 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2473 	 */
2474 	if (zio->io_error && vd == NULL &&
2475 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2476 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
2477 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
2478 		zio->io_error = 0;
2479 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
2480 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2481 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2482 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2483 		    zio_requeue_io_start_cut_in_line);
2484 		return (ZIO_PIPELINE_STOP);
2485 	}
2486 
2487 	/*
2488 	 * If we got an error on a leaf device, convert it to ENXIO
2489 	 * if the device is not accessible at all.
2490 	 */
2491 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2492 	    !vdev_accessible(vd, zio))
2493 		zio->io_error = ENXIO;
2494 
2495 	/*
2496 	 * If we can't write to an interior vdev (mirror or RAID-Z),
2497 	 * set vdev_cant_write so that we stop trying to allocate from it.
2498 	 */
2499 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2500 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2501 		vd->vdev_cant_write = B_TRUE;
2502 
2503 	if (zio->io_error)
2504 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2505 
2506 	return (ZIO_PIPELINE_CONTINUE);
2507 }
2508 
2509 void
2510 zio_vdev_io_reissue(zio_t *zio)
2511 {
2512 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2513 	ASSERT(zio->io_error == 0);
2514 
2515 	zio->io_stage >>= 1;
2516 }
2517 
2518 void
2519 zio_vdev_io_redone(zio_t *zio)
2520 {
2521 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2522 
2523 	zio->io_stage >>= 1;
2524 }
2525 
2526 void
2527 zio_vdev_io_bypass(zio_t *zio)
2528 {
2529 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2530 	ASSERT(zio->io_error == 0);
2531 
2532 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2533 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2534 }
2535 
2536 /*
2537  * ==========================================================================
2538  * Generate and verify checksums
2539  * ==========================================================================
2540  */
2541 static int
2542 zio_checksum_generate(zio_t *zio)
2543 {
2544 	blkptr_t *bp = zio->io_bp;
2545 	enum zio_checksum checksum;
2546 
2547 	if (bp == NULL) {
2548 		/*
2549 		 * This is zio_write_phys().
2550 		 * We're either generating a label checksum, or none at all.
2551 		 */
2552 		checksum = zio->io_prop.zp_checksum;
2553 
2554 		if (checksum == ZIO_CHECKSUM_OFF)
2555 			return (ZIO_PIPELINE_CONTINUE);
2556 
2557 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2558 	} else {
2559 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2560 			ASSERT(!IO_IS_ALLOCATING(zio));
2561 			checksum = ZIO_CHECKSUM_GANG_HEADER;
2562 		} else {
2563 			checksum = BP_GET_CHECKSUM(bp);
2564 		}
2565 	}
2566 
2567 	zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2568 
2569 	return (ZIO_PIPELINE_CONTINUE);
2570 }
2571 
2572 static int
2573 zio_checksum_verify(zio_t *zio)
2574 {
2575 	zio_bad_cksum_t info;
2576 	blkptr_t *bp = zio->io_bp;
2577 	int error;
2578 
2579 	ASSERT(zio->io_vd != NULL);
2580 
2581 	if (bp == NULL) {
2582 		/*
2583 		 * This is zio_read_phys().
2584 		 * We're either verifying a label checksum, or nothing at all.
2585 		 */
2586 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2587 			return (ZIO_PIPELINE_CONTINUE);
2588 
2589 		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2590 	}
2591 
2592 	if ((error = zio_checksum_error(zio, &info)) != 0) {
2593 		zio->io_error = error;
2594 		if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2595 			zfs_ereport_start_checksum(zio->io_spa,
2596 			    zio->io_vd, zio, zio->io_offset,
2597 			    zio->io_size, NULL, &info);
2598 		}
2599 	}
2600 
2601 	return (ZIO_PIPELINE_CONTINUE);
2602 }
2603 
2604 /*
2605  * Called by RAID-Z to ensure we don't compute the checksum twice.
2606  */
2607 void
2608 zio_checksum_verified(zio_t *zio)
2609 {
2610 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2611 }
2612 
2613 /*
2614  * ==========================================================================
2615  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2616  * An error of 0 indictes success.  ENXIO indicates whole-device failure,
2617  * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
2618  * indicate errors that are specific to one I/O, and most likely permanent.
2619  * Any other error is presumed to be worse because we weren't expecting it.
2620  * ==========================================================================
2621  */
2622 int
2623 zio_worst_error(int e1, int e2)
2624 {
2625 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2626 	int r1, r2;
2627 
2628 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2629 		if (e1 == zio_error_rank[r1])
2630 			break;
2631 
2632 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2633 		if (e2 == zio_error_rank[r2])
2634 			break;
2635 
2636 	return (r1 > r2 ? e1 : e2);
2637 }
2638 
2639 /*
2640  * ==========================================================================
2641  * I/O completion
2642  * ==========================================================================
2643  */
2644 static int
2645 zio_ready(zio_t *zio)
2646 {
2647 	blkptr_t *bp = zio->io_bp;
2648 	zio_t *pio, *pio_next;
2649 
2650 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2651 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2652 		return (ZIO_PIPELINE_STOP);
2653 
2654 	if (zio->io_ready) {
2655 		ASSERT(IO_IS_ALLOCATING(zio));
2656 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2657 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2658 
2659 		zio->io_ready(zio);
2660 	}
2661 
2662 	if (bp != NULL && bp != &zio->io_bp_copy)
2663 		zio->io_bp_copy = *bp;
2664 
2665 	if (zio->io_error)
2666 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2667 
2668 	mutex_enter(&zio->io_lock);
2669 	zio->io_state[ZIO_WAIT_READY] = 1;
2670 	pio = zio_walk_parents(zio);
2671 	mutex_exit(&zio->io_lock);
2672 
2673 	/*
2674 	 * As we notify zio's parents, new parents could be added.
2675 	 * New parents go to the head of zio's io_parent_list, however,
2676 	 * so we will (correctly) not notify them.  The remainder of zio's
2677 	 * io_parent_list, from 'pio_next' onward, cannot change because
2678 	 * all parents must wait for us to be done before they can be done.
2679 	 */
2680 	for (; pio != NULL; pio = pio_next) {
2681 		pio_next = zio_walk_parents(zio);
2682 		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2683 	}
2684 
2685 	if (zio->io_flags & ZIO_FLAG_NODATA) {
2686 		if (BP_IS_GANG(bp)) {
2687 			zio->io_flags &= ~ZIO_FLAG_NODATA;
2688 		} else {
2689 			ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2690 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2691 		}
2692 	}
2693 
2694 	if (zio_injection_enabled &&
2695 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
2696 		zio_handle_ignored_writes(zio);
2697 
2698 	return (ZIO_PIPELINE_CONTINUE);
2699 }
2700 
2701 static int
2702 zio_done(zio_t *zio)
2703 {
2704 	spa_t *spa = zio->io_spa;
2705 	zio_t *lio = zio->io_logical;
2706 	blkptr_t *bp = zio->io_bp;
2707 	vdev_t *vd = zio->io_vd;
2708 	uint64_t psize = zio->io_size;
2709 	zio_t *pio, *pio_next;
2710 
2711 	/*
2712 	 * If our children haven't all completed,
2713 	 * wait for them and then repeat this pipeline stage.
2714 	 */
2715 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2716 	    zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2717 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2718 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2719 		return (ZIO_PIPELINE_STOP);
2720 
2721 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2722 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2723 			ASSERT(zio->io_children[c][w] == 0);
2724 
2725 	if (bp != NULL) {
2726 		ASSERT(bp->blk_pad[0] == 0);
2727 		ASSERT(bp->blk_pad[1] == 0);
2728 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2729 		    (bp == zio_unique_parent(zio)->io_bp));
2730 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2731 		    zio->io_bp_override == NULL &&
2732 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2733 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
2734 			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2735 			ASSERT(BP_COUNT_GANG(bp) == 0 ||
2736 			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2737 		}
2738 	}
2739 
2740 	/*
2741 	 * If there were child vdev/gang/ddt errors, they apply to us now.
2742 	 */
2743 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2744 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2745 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2746 
2747 	/*
2748 	 * If the I/O on the transformed data was successful, generate any
2749 	 * checksum reports now while we still have the transformed data.
2750 	 */
2751 	if (zio->io_error == 0) {
2752 		while (zio->io_cksum_report != NULL) {
2753 			zio_cksum_report_t *zcr = zio->io_cksum_report;
2754 			uint64_t align = zcr->zcr_align;
2755 			uint64_t asize = P2ROUNDUP(psize, align);
2756 			char *abuf = zio->io_data;
2757 
2758 			if (asize != psize) {
2759 				abuf = zio_buf_alloc(asize);
2760 				bcopy(zio->io_data, abuf, psize);
2761 				bzero(abuf + psize, asize - psize);
2762 			}
2763 
2764 			zio->io_cksum_report = zcr->zcr_next;
2765 			zcr->zcr_next = NULL;
2766 			zcr->zcr_finish(zcr, abuf);
2767 			zfs_ereport_free_checksum(zcr);
2768 
2769 			if (asize != psize)
2770 				zio_buf_free(abuf, asize);
2771 		}
2772 	}
2773 
2774 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
2775 
2776 	vdev_stat_update(zio, psize);
2777 
2778 	if (zio->io_error) {
2779 		/*
2780 		 * If this I/O is attached to a particular vdev,
2781 		 * generate an error message describing the I/O failure
2782 		 * at the block level.  We ignore these errors if the
2783 		 * device is currently unavailable.
2784 		 */
2785 		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2786 			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2787 
2788 		if ((zio->io_error == EIO || !(zio->io_flags &
2789 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2790 		    zio == lio) {
2791 			/*
2792 			 * For logical I/O requests, tell the SPA to log the
2793 			 * error and generate a logical data ereport.
2794 			 */
2795 			spa_log_error(spa, zio);
2796 			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2797 			    0, 0);
2798 		}
2799 	}
2800 
2801 	if (zio->io_error && zio == lio) {
2802 		/*
2803 		 * Determine whether zio should be reexecuted.  This will
2804 		 * propagate all the way to the root via zio_notify_parent().
2805 		 */
2806 		ASSERT(vd == NULL && bp != NULL);
2807 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2808 
2809 		if (IO_IS_ALLOCATING(zio) &&
2810 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2811 			if (zio->io_error != ENOSPC)
2812 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2813 			else
2814 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2815 		}
2816 
2817 		if ((zio->io_type == ZIO_TYPE_READ ||
2818 		    zio->io_type == ZIO_TYPE_FREE) &&
2819 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2820 		    zio->io_error == ENXIO &&
2821 		    spa_load_state(spa) == SPA_LOAD_NONE &&
2822 		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2823 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2824 
2825 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2826 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2827 
2828 		/*
2829 		 * Here is a possibly good place to attempt to do
2830 		 * either combinatorial reconstruction or error correction
2831 		 * based on checksums.  It also might be a good place
2832 		 * to send out preliminary ereports before we suspend
2833 		 * processing.
2834 		 */
2835 	}
2836 
2837 	/*
2838 	 * If there were logical child errors, they apply to us now.
2839 	 * We defer this until now to avoid conflating logical child
2840 	 * errors with errors that happened to the zio itself when
2841 	 * updating vdev stats and reporting FMA events above.
2842 	 */
2843 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2844 
2845 	if ((zio->io_error || zio->io_reexecute) &&
2846 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2847 	    !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2848 		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2849 
2850 	zio_gang_tree_free(&zio->io_gang_tree);
2851 
2852 	/*
2853 	 * Godfather I/Os should never suspend.
2854 	 */
2855 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2856 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2857 		zio->io_reexecute = 0;
2858 
2859 	if (zio->io_reexecute) {
2860 		/*
2861 		 * This is a logical I/O that wants to reexecute.
2862 		 *
2863 		 * Reexecute is top-down.  When an i/o fails, if it's not
2864 		 * the root, it simply notifies its parent and sticks around.
2865 		 * The parent, seeing that it still has children in zio_done(),
2866 		 * does the same.  This percolates all the way up to the root.
2867 		 * The root i/o will reexecute or suspend the entire tree.
2868 		 *
2869 		 * This approach ensures that zio_reexecute() honors
2870 		 * all the original i/o dependency relationships, e.g.
2871 		 * parents not executing until children are ready.
2872 		 */
2873 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2874 
2875 		zio->io_gang_leader = NULL;
2876 
2877 		mutex_enter(&zio->io_lock);
2878 		zio->io_state[ZIO_WAIT_DONE] = 1;
2879 		mutex_exit(&zio->io_lock);
2880 
2881 		/*
2882 		 * "The Godfather" I/O monitors its children but is
2883 		 * not a true parent to them. It will track them through
2884 		 * the pipeline but severs its ties whenever they get into
2885 		 * trouble (e.g. suspended). This allows "The Godfather"
2886 		 * I/O to return status without blocking.
2887 		 */
2888 		for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2889 			zio_link_t *zl = zio->io_walk_link;
2890 			pio_next = zio_walk_parents(zio);
2891 
2892 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2893 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2894 				zio_remove_child(pio, zio, zl);
2895 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2896 			}
2897 		}
2898 
2899 		if ((pio = zio_unique_parent(zio)) != NULL) {
2900 			/*
2901 			 * We're not a root i/o, so there's nothing to do
2902 			 * but notify our parent.  Don't propagate errors
2903 			 * upward since we haven't permanently failed yet.
2904 			 */
2905 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2906 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2907 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2908 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2909 			/*
2910 			 * We'd fail again if we reexecuted now, so suspend
2911 			 * until conditions improve (e.g. device comes online).
2912 			 */
2913 			zio_suspend(spa, zio);
2914 		} else {
2915 			/*
2916 			 * Reexecution is potentially a huge amount of work.
2917 			 * Hand it off to the otherwise-unused claim taskq.
2918 			 */
2919 			ASSERT(zio->io_tqent.tqent_next == NULL);
2920 			(void) taskq_dispatch_ent(
2921 			    spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2922 			    (task_func_t *)zio_reexecute, zio, 0,
2923 			    &zio->io_tqent);
2924 		}
2925 		return (ZIO_PIPELINE_STOP);
2926 	}
2927 
2928 	ASSERT(zio->io_child_count == 0);
2929 	ASSERT(zio->io_reexecute == 0);
2930 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2931 
2932 	/*
2933 	 * Report any checksum errors, since the I/O is complete.
2934 	 */
2935 	while (zio->io_cksum_report != NULL) {
2936 		zio_cksum_report_t *zcr = zio->io_cksum_report;
2937 		zio->io_cksum_report = zcr->zcr_next;
2938 		zcr->zcr_next = NULL;
2939 		zcr->zcr_finish(zcr, NULL);
2940 		zfs_ereport_free_checksum(zcr);
2941 	}
2942 
2943 	/*
2944 	 * It is the responsibility of the done callback to ensure that this
2945 	 * particular zio is no longer discoverable for adoption, and as
2946 	 * such, cannot acquire any new parents.
2947 	 */
2948 	if (zio->io_done)
2949 		zio->io_done(zio);
2950 
2951 	mutex_enter(&zio->io_lock);
2952 	zio->io_state[ZIO_WAIT_DONE] = 1;
2953 	mutex_exit(&zio->io_lock);
2954 
2955 	for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2956 		zio_link_t *zl = zio->io_walk_link;
2957 		pio_next = zio_walk_parents(zio);
2958 		zio_remove_child(pio, zio, zl);
2959 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2960 	}
2961 
2962 	if (zio->io_waiter != NULL) {
2963 		mutex_enter(&zio->io_lock);
2964 		zio->io_executor = NULL;
2965 		cv_broadcast(&zio->io_cv);
2966 		mutex_exit(&zio->io_lock);
2967 	} else {
2968 		zio_destroy(zio);
2969 	}
2970 
2971 	return (ZIO_PIPELINE_STOP);
2972 }
2973 
2974 /*
2975  * ==========================================================================
2976  * I/O pipeline definition
2977  * ==========================================================================
2978  */
2979 static zio_pipe_stage_t *zio_pipeline[] = {
2980 	NULL,
2981 	zio_read_bp_init,
2982 	zio_free_bp_init,
2983 	zio_issue_async,
2984 	zio_write_bp_init,
2985 	zio_checksum_generate,
2986 	zio_ddt_read_start,
2987 	zio_ddt_read_done,
2988 	zio_ddt_write,
2989 	zio_ddt_free,
2990 	zio_gang_assemble,
2991 	zio_gang_issue,
2992 	zio_dva_allocate,
2993 	zio_dva_free,
2994 	zio_dva_claim,
2995 	zio_ready,
2996 	zio_vdev_io_start,
2997 	zio_vdev_io_done,
2998 	zio_vdev_io_assess,
2999 	zio_checksum_verify,
3000 	zio_done
3001 };
3002 
3003 /* dnp is the dnode for zb1->zb_object */
3004 boolean_t
3005 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3006     const zbookmark_t *zb2)
3007 {
3008 	uint64_t zb1nextL0, zb2thisobj;
3009 
3010 	ASSERT(zb1->zb_objset == zb2->zb_objset);
3011 	ASSERT(zb2->zb_level == 0);
3012 
3013 	/*
3014 	 * A bookmark in the deadlist is considered to be after
3015 	 * everything else.
3016 	 */
3017 	if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3018 		return (B_TRUE);
3019 
3020 	/* The objset_phys_t isn't before anything. */
3021 	if (dnp == NULL)
3022 		return (B_FALSE);
3023 
3024 	zb1nextL0 = (zb1->zb_blkid + 1) <<
3025 	    ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3026 
3027 	zb2thisobj = zb2->zb_object ? zb2->zb_object :
3028 	    zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3029 
3030 	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3031 		uint64_t nextobj = zb1nextL0 *
3032 		    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3033 		return (nextobj <= zb2thisobj);
3034 	}
3035 
3036 	if (zb1->zb_object < zb2thisobj)
3037 		return (B_TRUE);
3038 	if (zb1->zb_object > zb2thisobj)
3039 		return (B_FALSE);
3040 	if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3041 		return (B_FALSE);
3042 	return (zb1nextL0 <= zb2->zb_blkid);
3043 }
3044