xref: /titanic_52/usr/src/uts/common/fs/zfs/zio.c (revision 7aa76ffc594f84c1c092911a84f85a79ddb44c73)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2011 by Delphix. All rights reserved.
24  * 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 		if (size <= 4 * SPA_MINBLOCKSIZE) {
129 			align = SPA_MINBLOCKSIZE;
130 		} else if (P2PHASE(size, PAGESIZE) == 0) {
131 			align = PAGESIZE;
132 		} else if (P2PHASE(size, p2 >> 2) == 0) {
133 			align = p2 >> 2;
134 		}
135 
136 		if (align != 0) {
137 			char name[36];
138 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
139 			zio_buf_cache[c] = kmem_cache_create(name, size,
140 			    align, NULL, NULL, NULL, NULL, NULL, cflags);
141 
142 			/*
143 			 * Since zio_data bufs do not appear in crash dumps, we
144 			 * pass KMC_NOTOUCH so that no allocator metadata is
145 			 * stored with the buffers.
146 			 */
147 			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
148 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
149 			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
150 			    cflags | KMC_NOTOUCH);
151 		}
152 	}
153 
154 	while (--c != 0) {
155 		ASSERT(zio_buf_cache[c] != NULL);
156 		if (zio_buf_cache[c - 1] == NULL)
157 			zio_buf_cache[c - 1] = zio_buf_cache[c];
158 
159 		ASSERT(zio_data_buf_cache[c] != NULL);
160 		if (zio_data_buf_cache[c - 1] == NULL)
161 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
162 	}
163 
164 	/*
165 	 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
166 	 * to fail 3 times per txg or 8 failures, whichever is greater.
167 	 */
168 	zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
169 
170 	zio_inject_init();
171 }
172 
173 void
174 zio_fini(void)
175 {
176 	size_t c;
177 	kmem_cache_t *last_cache = NULL;
178 	kmem_cache_t *last_data_cache = NULL;
179 
180 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
181 		if (zio_buf_cache[c] != last_cache) {
182 			last_cache = zio_buf_cache[c];
183 			kmem_cache_destroy(zio_buf_cache[c]);
184 		}
185 		zio_buf_cache[c] = NULL;
186 
187 		if (zio_data_buf_cache[c] != last_data_cache) {
188 			last_data_cache = zio_data_buf_cache[c];
189 			kmem_cache_destroy(zio_data_buf_cache[c]);
190 		}
191 		zio_data_buf_cache[c] = NULL;
192 	}
193 
194 	kmem_cache_destroy(zio_link_cache);
195 	kmem_cache_destroy(zio_cache);
196 
197 	zio_inject_fini();
198 }
199 
200 /*
201  * ==========================================================================
202  * Allocate and free I/O buffers
203  * ==========================================================================
204  */
205 
206 /*
207  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
208  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
209  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
210  * excess / transient data in-core during a crashdump.
211  */
212 void *
213 zio_buf_alloc(size_t size)
214 {
215 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
216 
217 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
218 
219 	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
220 }
221 
222 /*
223  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
224  * crashdump if the kernel panics.  This exists so that we will limit the amount
225  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
226  * of kernel heap dumped to disk when the kernel panics)
227  */
228 void *
229 zio_data_buf_alloc(size_t size)
230 {
231 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
232 
233 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
234 
235 	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
236 }
237 
238 void
239 zio_buf_free(void *buf, size_t size)
240 {
241 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
242 
243 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
244 
245 	kmem_cache_free(zio_buf_cache[c], buf);
246 }
247 
248 void
249 zio_data_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_data_buf_cache[c], buf);
256 }
257 
258 /*
259  * ==========================================================================
260  * Push and pop I/O transform buffers
261  * ==========================================================================
262  */
263 static void
264 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
265 	zio_transform_func_t *transform)
266 {
267 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
268 
269 	zt->zt_orig_data = zio->io_data;
270 	zt->zt_orig_size = zio->io_size;
271 	zt->zt_bufsize = bufsize;
272 	zt->zt_transform = transform;
273 
274 	zt->zt_next = zio->io_transform_stack;
275 	zio->io_transform_stack = zt;
276 
277 	zio->io_data = data;
278 	zio->io_size = size;
279 }
280 
281 static void
282 zio_pop_transforms(zio_t *zio)
283 {
284 	zio_transform_t *zt;
285 
286 	while ((zt = zio->io_transform_stack) != NULL) {
287 		if (zt->zt_transform != NULL)
288 			zt->zt_transform(zio,
289 			    zt->zt_orig_data, zt->zt_orig_size);
290 
291 		if (zt->zt_bufsize != 0)
292 			zio_buf_free(zio->io_data, zt->zt_bufsize);
293 
294 		zio->io_data = zt->zt_orig_data;
295 		zio->io_size = zt->zt_orig_size;
296 		zio->io_transform_stack = zt->zt_next;
297 
298 		kmem_free(zt, sizeof (zio_transform_t));
299 	}
300 }
301 
302 /*
303  * ==========================================================================
304  * I/O transform callbacks for subblocks and decompression
305  * ==========================================================================
306  */
307 static void
308 zio_subblock(zio_t *zio, void *data, uint64_t size)
309 {
310 	ASSERT(zio->io_size > size);
311 
312 	if (zio->io_type == ZIO_TYPE_READ)
313 		bcopy(zio->io_data, data, size);
314 }
315 
316 static void
317 zio_decompress(zio_t *zio, void *data, uint64_t size)
318 {
319 	if (zio->io_error == 0 &&
320 	    zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
321 	    zio->io_data, data, zio->io_size, size) != 0)
322 		zio->io_error = EIO;
323 }
324 
325 /*
326  * ==========================================================================
327  * I/O parent/child relationships and pipeline interlocks
328  * ==========================================================================
329  */
330 /*
331  * NOTE - Callers to zio_walk_parents() and zio_walk_children must
332  *        continue calling these functions until they return NULL.
333  *        Otherwise, the next caller will pick up the list walk in
334  *        some indeterminate state.  (Otherwise every caller would
335  *        have to pass in a cookie to keep the state represented by
336  *        io_walk_link, which gets annoying.)
337  */
338 zio_t *
339 zio_walk_parents(zio_t *cio)
340 {
341 	zio_link_t *zl = cio->io_walk_link;
342 	list_t *pl = &cio->io_parent_list;
343 
344 	zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
345 	cio->io_walk_link = zl;
346 
347 	if (zl == NULL)
348 		return (NULL);
349 
350 	ASSERT(zl->zl_child == cio);
351 	return (zl->zl_parent);
352 }
353 
354 zio_t *
355 zio_walk_children(zio_t *pio)
356 {
357 	zio_link_t *zl = pio->io_walk_link;
358 	list_t *cl = &pio->io_child_list;
359 
360 	zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
361 	pio->io_walk_link = zl;
362 
363 	if (zl == NULL)
364 		return (NULL);
365 
366 	ASSERT(zl->zl_parent == pio);
367 	return (zl->zl_child);
368 }
369 
370 zio_t *
371 zio_unique_parent(zio_t *cio)
372 {
373 	zio_t *pio = zio_walk_parents(cio);
374 
375 	VERIFY(zio_walk_parents(cio) == NULL);
376 	return (pio);
377 }
378 
379 void
380 zio_add_child(zio_t *pio, zio_t *cio)
381 {
382 	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
383 
384 	/*
385 	 * Logical I/Os can have logical, gang, or vdev children.
386 	 * Gang I/Os can have gang or vdev children.
387 	 * Vdev I/Os can only have vdev children.
388 	 * The following ASSERT captures all of these constraints.
389 	 */
390 	ASSERT(cio->io_child_type <= pio->io_child_type);
391 
392 	zl->zl_parent = pio;
393 	zl->zl_child = cio;
394 
395 	mutex_enter(&cio->io_lock);
396 	mutex_enter(&pio->io_lock);
397 
398 	ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
399 
400 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
401 		pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
402 
403 	list_insert_head(&pio->io_child_list, zl);
404 	list_insert_head(&cio->io_parent_list, zl);
405 
406 	pio->io_child_count++;
407 	cio->io_parent_count++;
408 
409 	mutex_exit(&pio->io_lock);
410 	mutex_exit(&cio->io_lock);
411 }
412 
413 static void
414 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
415 {
416 	ASSERT(zl->zl_parent == pio);
417 	ASSERT(zl->zl_child == cio);
418 
419 	mutex_enter(&cio->io_lock);
420 	mutex_enter(&pio->io_lock);
421 
422 	list_remove(&pio->io_child_list, zl);
423 	list_remove(&cio->io_parent_list, zl);
424 
425 	pio->io_child_count--;
426 	cio->io_parent_count--;
427 
428 	mutex_exit(&pio->io_lock);
429 	mutex_exit(&cio->io_lock);
430 
431 	kmem_cache_free(zio_link_cache, zl);
432 }
433 
434 static boolean_t
435 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
436 {
437 	uint64_t *countp = &zio->io_children[child][wait];
438 	boolean_t waiting = B_FALSE;
439 
440 	mutex_enter(&zio->io_lock);
441 	ASSERT(zio->io_stall == NULL);
442 	if (*countp != 0) {
443 		zio->io_stage >>= 1;
444 		zio->io_stall = countp;
445 		waiting = B_TRUE;
446 	}
447 	mutex_exit(&zio->io_lock);
448 
449 	return (waiting);
450 }
451 
452 static void
453 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
454 {
455 	uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
456 	int *errorp = &pio->io_child_error[zio->io_child_type];
457 
458 	mutex_enter(&pio->io_lock);
459 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
460 		*errorp = zio_worst_error(*errorp, zio->io_error);
461 	pio->io_reexecute |= zio->io_reexecute;
462 	ASSERT3U(*countp, >, 0);
463 	if (--*countp == 0 && pio->io_stall == countp) {
464 		pio->io_stall = NULL;
465 		mutex_exit(&pio->io_lock);
466 		zio_execute(pio);
467 	} else {
468 		mutex_exit(&pio->io_lock);
469 	}
470 }
471 
472 static void
473 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
474 {
475 	if (zio->io_child_error[c] != 0 && zio->io_error == 0)
476 		zio->io_error = zio->io_child_error[c];
477 }
478 
479 /*
480  * ==========================================================================
481  * Create the various types of I/O (read, write, free, etc)
482  * ==========================================================================
483  */
484 static zio_t *
485 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
486     void *data, uint64_t size, zio_done_func_t *done, void *private,
487     zio_type_t type, int priority, enum zio_flag flags,
488     vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
489     enum zio_stage stage, enum zio_stage pipeline)
490 {
491 	zio_t *zio;
492 
493 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
494 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
495 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
496 
497 	ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
498 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
499 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
500 
501 	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
502 	bzero(zio, sizeof (zio_t));
503 
504 	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
505 	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
506 
507 	list_create(&zio->io_parent_list, sizeof (zio_link_t),
508 	    offsetof(zio_link_t, zl_parent_node));
509 	list_create(&zio->io_child_list, sizeof (zio_link_t),
510 	    offsetof(zio_link_t, zl_child_node));
511 
512 	if (vd != NULL)
513 		zio->io_child_type = ZIO_CHILD_VDEV;
514 	else if (flags & ZIO_FLAG_GANG_CHILD)
515 		zio->io_child_type = ZIO_CHILD_GANG;
516 	else if (flags & ZIO_FLAG_DDT_CHILD)
517 		zio->io_child_type = ZIO_CHILD_DDT;
518 	else
519 		zio->io_child_type = ZIO_CHILD_LOGICAL;
520 
521 	if (bp != NULL) {
522 		zio->io_bp = (blkptr_t *)bp;
523 		zio->io_bp_copy = *bp;
524 		zio->io_bp_orig = *bp;
525 		if (type != ZIO_TYPE_WRITE ||
526 		    zio->io_child_type == ZIO_CHILD_DDT)
527 			zio->io_bp = &zio->io_bp_copy;	/* so caller can free */
528 		if (zio->io_child_type == ZIO_CHILD_LOGICAL)
529 			zio->io_logical = zio;
530 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
531 			pipeline |= ZIO_GANG_STAGES;
532 	}
533 
534 	zio->io_spa = spa;
535 	zio->io_txg = txg;
536 	zio->io_done = done;
537 	zio->io_private = private;
538 	zio->io_type = type;
539 	zio->io_priority = priority;
540 	zio->io_vd = vd;
541 	zio->io_offset = offset;
542 	zio->io_orig_data = zio->io_data = data;
543 	zio->io_orig_size = zio->io_size = size;
544 	zio->io_orig_flags = zio->io_flags = flags;
545 	zio->io_orig_stage = zio->io_stage = stage;
546 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
547 
548 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
549 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
550 
551 	if (zb != NULL)
552 		zio->io_bookmark = *zb;
553 
554 	if (pio != NULL) {
555 		if (zio->io_logical == NULL)
556 			zio->io_logical = pio->io_logical;
557 		if (zio->io_child_type == ZIO_CHILD_GANG)
558 			zio->io_gang_leader = pio->io_gang_leader;
559 		zio_add_child(pio, zio);
560 	}
561 
562 	return (zio);
563 }
564 
565 static void
566 zio_destroy(zio_t *zio)
567 {
568 	list_destroy(&zio->io_parent_list);
569 	list_destroy(&zio->io_child_list);
570 	mutex_destroy(&zio->io_lock);
571 	cv_destroy(&zio->io_cv);
572 	kmem_cache_free(zio_cache, zio);
573 }
574 
575 zio_t *
576 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
577     void *private, enum zio_flag flags)
578 {
579 	zio_t *zio;
580 
581 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
582 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
583 	    ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
584 
585 	return (zio);
586 }
587 
588 zio_t *
589 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
590 {
591 	return (zio_null(NULL, spa, NULL, done, private, flags));
592 }
593 
594 zio_t *
595 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
596     void *data, uint64_t size, zio_done_func_t *done, void *private,
597     int priority, enum zio_flag flags, const zbookmark_t *zb)
598 {
599 	zio_t *zio;
600 
601 	zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
602 	    data, size, done, private,
603 	    ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
604 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
605 	    ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
606 
607 	return (zio);
608 }
609 
610 zio_t *
611 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
612     void *data, uint64_t size, const zio_prop_t *zp,
613     zio_done_func_t *ready, zio_done_func_t *done, void *private,
614     int priority, enum zio_flag flags, const zbookmark_t *zb)
615 {
616 	zio_t *zio;
617 
618 	ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
619 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
620 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
621 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
622 	    zp->zp_type < DMU_OT_NUMTYPES &&
623 	    zp->zp_level < 32 &&
624 	    zp->zp_copies > 0 &&
625 	    zp->zp_copies <= spa_max_replication(spa) &&
626 	    zp->zp_dedup <= 1 &&
627 	    zp->zp_dedup_verify <= 1);
628 
629 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
630 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
631 	    ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
632 	    ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
633 
634 	zio->io_ready = ready;
635 	zio->io_prop = *zp;
636 
637 	return (zio);
638 }
639 
640 zio_t *
641 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
642     uint64_t size, zio_done_func_t *done, void *private, int priority,
643     enum zio_flag flags, zbookmark_t *zb)
644 {
645 	zio_t *zio;
646 
647 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
648 	    ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
649 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
650 
651 	return (zio);
652 }
653 
654 void
655 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
656 {
657 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
658 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
659 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
660 	ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
661 
662 	zio->io_prop.zp_copies = copies;
663 	zio->io_bp_override = bp;
664 }
665 
666 void
667 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
668 {
669 	bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
670 }
671 
672 zio_t *
673 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
674     enum zio_flag flags)
675 {
676 	zio_t *zio;
677 
678 	dprintf_bp(bp, "freeing in txg %llu, pass %u",
679 	    (longlong_t)txg, spa->spa_sync_pass);
680 
681 	ASSERT(!BP_IS_HOLE(bp));
682 	ASSERT(spa_syncing_txg(spa) == txg);
683 	ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
684 
685 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
686 	    NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
687 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
688 
689 	return (zio);
690 }
691 
692 zio_t *
693 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
694     zio_done_func_t *done, void *private, enum zio_flag flags)
695 {
696 	zio_t *zio;
697 
698 	/*
699 	 * A claim is an allocation of a specific block.  Claims are needed
700 	 * to support immediate writes in the intent log.  The issue is that
701 	 * immediate writes contain committed data, but in a txg that was
702 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
703 	 * the intent log claims all blocks that contain immediate write data
704 	 * so that the SPA knows they're in use.
705 	 *
706 	 * All claims *must* be resolved in the first txg -- before the SPA
707 	 * starts allocating blocks -- so that nothing is allocated twice.
708 	 * If txg == 0 we just verify that the block is claimable.
709 	 */
710 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
711 	ASSERT(txg == spa_first_txg(spa) || txg == 0);
712 	ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));	/* zdb(1M) */
713 
714 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
715 	    done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
716 	    NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
717 
718 	return (zio);
719 }
720 
721 zio_t *
722 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
723     zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
724 {
725 	zio_t *zio;
726 	int c;
727 
728 	if (vd->vdev_children == 0) {
729 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
730 		    ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
731 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
732 
733 		zio->io_cmd = cmd;
734 	} else {
735 		zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
736 
737 		for (c = 0; c < vd->vdev_children; c++)
738 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
739 			    done, private, priority, flags));
740 	}
741 
742 	return (zio);
743 }
744 
745 zio_t *
746 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
747     void *data, int checksum, zio_done_func_t *done, void *private,
748     int priority, enum zio_flag flags, boolean_t labels)
749 {
750 	zio_t *zio;
751 
752 	ASSERT(vd->vdev_children == 0);
753 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
754 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
755 	ASSERT3U(offset + size, <=, vd->vdev_psize);
756 
757 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
758 	    ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
759 	    ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
760 
761 	zio->io_prop.zp_checksum = checksum;
762 
763 	return (zio);
764 }
765 
766 zio_t *
767 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
768     void *data, int checksum, zio_done_func_t *done, void *private,
769     int priority, enum zio_flag flags, boolean_t labels)
770 {
771 	zio_t *zio;
772 
773 	ASSERT(vd->vdev_children == 0);
774 	ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
775 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
776 	ASSERT3U(offset + size, <=, vd->vdev_psize);
777 
778 	zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
779 	    ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
780 	    ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
781 
782 	zio->io_prop.zp_checksum = checksum;
783 
784 	if (zio_checksum_table[checksum].ci_eck) {
785 		/*
786 		 * zec checksums are necessarily destructive -- they modify
787 		 * the end of the write buffer to hold the verifier/checksum.
788 		 * Therefore, we must make a local copy in case the data is
789 		 * being written to multiple places in parallel.
790 		 */
791 		void *wbuf = zio_buf_alloc(size);
792 		bcopy(data, wbuf, size);
793 		zio_push_transform(zio, wbuf, size, size, NULL);
794 	}
795 
796 	return (zio);
797 }
798 
799 /*
800  * Create a child I/O to do some work for us.
801  */
802 zio_t *
803 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
804 	void *data, uint64_t size, int type, int priority, enum zio_flag flags,
805 	zio_done_func_t *done, void *private)
806 {
807 	enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
808 	zio_t *zio;
809 
810 	ASSERT(vd->vdev_parent ==
811 	    (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
812 
813 	if (type == ZIO_TYPE_READ && bp != NULL) {
814 		/*
815 		 * If we have the bp, then the child should perform the
816 		 * checksum and the parent need not.  This pushes error
817 		 * detection as close to the leaves as possible and
818 		 * eliminates redundant checksums in the interior nodes.
819 		 */
820 		pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
821 		pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
822 	}
823 
824 	if (vd->vdev_children == 0)
825 		offset += VDEV_LABEL_START_SIZE;
826 
827 	flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
828 
829 	/*
830 	 * If we've decided to do a repair, the write is not speculative --
831 	 * even if the original read was.
832 	 */
833 	if (flags & ZIO_FLAG_IO_REPAIR)
834 		flags &= ~ZIO_FLAG_SPECULATIVE;
835 
836 	zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
837 	    done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
838 	    ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
839 
840 	return (zio);
841 }
842 
843 zio_t *
844 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
845 	int type, int priority, enum zio_flag flags,
846 	zio_done_func_t *done, void *private)
847 {
848 	zio_t *zio;
849 
850 	ASSERT(vd->vdev_ops->vdev_op_leaf);
851 
852 	zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
853 	    data, size, done, private, type, priority,
854 	    flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
855 	    vd, offset, NULL,
856 	    ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
857 
858 	return (zio);
859 }
860 
861 void
862 zio_flush(zio_t *zio, vdev_t *vd)
863 {
864 	zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
865 	    NULL, NULL, ZIO_PRIORITY_NOW,
866 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
867 }
868 
869 void
870 zio_shrink(zio_t *zio, uint64_t size)
871 {
872 	ASSERT(zio->io_executor == NULL);
873 	ASSERT(zio->io_orig_size == zio->io_size);
874 	ASSERT(size <= zio->io_size);
875 
876 	/*
877 	 * We don't shrink for raidz because of problems with the
878 	 * reconstruction when reading back less than the block size.
879 	 * Note, BP_IS_RAIDZ() assumes no compression.
880 	 */
881 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
882 	if (!BP_IS_RAIDZ(zio->io_bp))
883 		zio->io_orig_size = zio->io_size = size;
884 }
885 
886 /*
887  * ==========================================================================
888  * Prepare to read and write logical blocks
889  * ==========================================================================
890  */
891 
892 static int
893 zio_read_bp_init(zio_t *zio)
894 {
895 	blkptr_t *bp = zio->io_bp;
896 
897 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
898 	    zio->io_child_type == ZIO_CHILD_LOGICAL &&
899 	    !(zio->io_flags & ZIO_FLAG_RAW)) {
900 		uint64_t psize = BP_GET_PSIZE(bp);
901 		void *cbuf = zio_buf_alloc(psize);
902 
903 		zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
904 	}
905 
906 	if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
907 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
908 
909 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
910 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
911 
912 	if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
913 		zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
914 
915 	return (ZIO_PIPELINE_CONTINUE);
916 }
917 
918 static int
919 zio_write_bp_init(zio_t *zio)
920 {
921 	spa_t *spa = zio->io_spa;
922 	zio_prop_t *zp = &zio->io_prop;
923 	enum zio_compress compress = zp->zp_compress;
924 	blkptr_t *bp = zio->io_bp;
925 	uint64_t lsize = zio->io_size;
926 	uint64_t psize = lsize;
927 	int pass = 1;
928 
929 	/*
930 	 * If our children haven't all reached the ready stage,
931 	 * wait for them and then repeat this pipeline stage.
932 	 */
933 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
934 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
935 		return (ZIO_PIPELINE_STOP);
936 
937 	if (!IO_IS_ALLOCATING(zio))
938 		return (ZIO_PIPELINE_CONTINUE);
939 
940 	ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
941 
942 	if (zio->io_bp_override) {
943 		ASSERT(bp->blk_birth != zio->io_txg);
944 		ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
945 
946 		*bp = *zio->io_bp_override;
947 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
948 
949 		if (BP_IS_HOLE(bp) || !zp->zp_dedup)
950 			return (ZIO_PIPELINE_CONTINUE);
951 
952 		ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
953 		    zp->zp_dedup_verify);
954 
955 		if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
956 			BP_SET_DEDUP(bp, 1);
957 			zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
958 			return (ZIO_PIPELINE_CONTINUE);
959 		}
960 		zio->io_bp_override = NULL;
961 		BP_ZERO(bp);
962 	}
963 
964 	if (bp->blk_birth == zio->io_txg) {
965 		/*
966 		 * We're rewriting an existing block, which means we're
967 		 * working on behalf of spa_sync().  For spa_sync() to
968 		 * converge, it must eventually be the case that we don't
969 		 * have to allocate new blocks.  But compression changes
970 		 * the blocksize, which forces a reallocate, and makes
971 		 * convergence take longer.  Therefore, after the first
972 		 * few passes, stop compressing to ensure convergence.
973 		 */
974 		pass = spa_sync_pass(spa);
975 
976 		ASSERT(zio->io_txg == spa_syncing_txg(spa));
977 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
978 		ASSERT(!BP_GET_DEDUP(bp));
979 
980 		if (pass > SYNC_PASS_DONT_COMPRESS)
981 			compress = ZIO_COMPRESS_OFF;
982 
983 		/* Make sure someone doesn't change their mind on overwrites */
984 		ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
985 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
986 	}
987 
988 	if (compress != ZIO_COMPRESS_OFF) {
989 		void *cbuf = zio_buf_alloc(lsize);
990 		psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
991 		if (psize == 0 || psize == lsize) {
992 			compress = ZIO_COMPRESS_OFF;
993 			zio_buf_free(cbuf, lsize);
994 		} else {
995 			ASSERT(psize < lsize);
996 			zio_push_transform(zio, cbuf, psize, lsize, NULL);
997 		}
998 	}
999 
1000 	/*
1001 	 * The final pass of spa_sync() must be all rewrites, but the first
1002 	 * few passes offer a trade-off: allocating blocks defers convergence,
1003 	 * but newly allocated blocks are sequential, so they can be written
1004 	 * to disk faster.  Therefore, we allow the first few passes of
1005 	 * spa_sync() to allocate new blocks, but force rewrites after that.
1006 	 * There should only be a handful of blocks after pass 1 in any case.
1007 	 */
1008 	if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1009 	    pass > SYNC_PASS_REWRITE) {
1010 		ASSERT(psize != 0);
1011 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1012 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1013 		zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1014 	} else {
1015 		BP_ZERO(bp);
1016 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
1017 	}
1018 
1019 	if (psize == 0) {
1020 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1021 	} else {
1022 		ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1023 		BP_SET_LSIZE(bp, lsize);
1024 		BP_SET_PSIZE(bp, psize);
1025 		BP_SET_COMPRESS(bp, compress);
1026 		BP_SET_CHECKSUM(bp, zp->zp_checksum);
1027 		BP_SET_TYPE(bp, zp->zp_type);
1028 		BP_SET_LEVEL(bp, zp->zp_level);
1029 		BP_SET_DEDUP(bp, zp->zp_dedup);
1030 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1031 		if (zp->zp_dedup) {
1032 			ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1033 			ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1034 			zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1035 		}
1036 	}
1037 
1038 	return (ZIO_PIPELINE_CONTINUE);
1039 }
1040 
1041 static int
1042 zio_free_bp_init(zio_t *zio)
1043 {
1044 	blkptr_t *bp = zio->io_bp;
1045 
1046 	if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1047 		if (BP_GET_DEDUP(bp))
1048 			zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1049 	}
1050 
1051 	return (ZIO_PIPELINE_CONTINUE);
1052 }
1053 
1054 /*
1055  * ==========================================================================
1056  * Execute the I/O pipeline
1057  * ==========================================================================
1058  */
1059 
1060 static void
1061 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1062 {
1063 	spa_t *spa = zio->io_spa;
1064 	zio_type_t t = zio->io_type;
1065 	int flags = (cutinline ? TQ_FRONT : 0);
1066 
1067 	/*
1068 	 * If we're a config writer or a probe, the normal issue and
1069 	 * interrupt threads may all be blocked waiting for the config lock.
1070 	 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1071 	 */
1072 	if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1073 		t = ZIO_TYPE_NULL;
1074 
1075 	/*
1076 	 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1077 	 */
1078 	if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1079 		t = ZIO_TYPE_NULL;
1080 
1081 	/*
1082 	 * If this is a high priority I/O, then use the high priority taskq.
1083 	 */
1084 	if (zio->io_priority == ZIO_PRIORITY_NOW &&
1085 	    spa->spa_zio_taskq[t][q + 1] != NULL)
1086 		q++;
1087 
1088 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1089 
1090 	/*
1091 	 * NB: We are assuming that the zio can only be dispatched
1092 	 * to a single taskq at a time.  It would be a grievous error
1093 	 * to dispatch the zio to another taskq at the same time.
1094 	 */
1095 	ASSERT(zio->io_tqent.tqent_next == NULL);
1096 	taskq_dispatch_ent(spa->spa_zio_taskq[t][q],
1097 	    (task_func_t *)zio_execute, zio, flags, &zio->io_tqent);
1098 }
1099 
1100 static boolean_t
1101 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1102 {
1103 	kthread_t *executor = zio->io_executor;
1104 	spa_t *spa = zio->io_spa;
1105 
1106 	for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1107 		if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1108 			return (B_TRUE);
1109 
1110 	return (B_FALSE);
1111 }
1112 
1113 static int
1114 zio_issue_async(zio_t *zio)
1115 {
1116 	zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1117 
1118 	return (ZIO_PIPELINE_STOP);
1119 }
1120 
1121 void
1122 zio_interrupt(zio_t *zio)
1123 {
1124 	zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1125 }
1126 
1127 /*
1128  * Execute the I/O pipeline until one of the following occurs:
1129  * (1) the I/O completes; (2) the pipeline stalls waiting for
1130  * dependent child I/Os; (3) the I/O issues, so we're waiting
1131  * for an I/O completion interrupt; (4) the I/O is delegated by
1132  * vdev-level caching or aggregation; (5) the I/O is deferred
1133  * due to vdev-level queueing; (6) the I/O is handed off to
1134  * another thread.  In all cases, the pipeline stops whenever
1135  * there's no CPU work; it never burns a thread in cv_wait().
1136  *
1137  * There's no locking on io_stage because there's no legitimate way
1138  * for multiple threads to be attempting to process the same I/O.
1139  */
1140 static zio_pipe_stage_t *zio_pipeline[];
1141 
1142 void
1143 zio_execute(zio_t *zio)
1144 {
1145 	zio->io_executor = curthread;
1146 
1147 	while (zio->io_stage < ZIO_STAGE_DONE) {
1148 		enum zio_stage pipeline = zio->io_pipeline;
1149 		enum zio_stage stage = zio->io_stage;
1150 		int rv;
1151 
1152 		ASSERT(!MUTEX_HELD(&zio->io_lock));
1153 		ASSERT(ISP2(stage));
1154 		ASSERT(zio->io_stall == NULL);
1155 
1156 		do {
1157 			stage <<= 1;
1158 		} while ((stage & pipeline) == 0);
1159 
1160 		ASSERT(stage <= ZIO_STAGE_DONE);
1161 
1162 		/*
1163 		 * If we are in interrupt context and this pipeline stage
1164 		 * will grab a config lock that is held across I/O,
1165 		 * or may wait for an I/O that needs an interrupt thread
1166 		 * to complete, issue async to avoid deadlock.
1167 		 *
1168 		 * For VDEV_IO_START, we cut in line so that the io will
1169 		 * be sent to disk promptly.
1170 		 */
1171 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1172 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1173 			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1174 			    zio_requeue_io_start_cut_in_line : B_FALSE;
1175 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1176 			return;
1177 		}
1178 
1179 		zio->io_stage = stage;
1180 		rv = zio_pipeline[highbit(stage) - 1](zio);
1181 
1182 		if (rv == ZIO_PIPELINE_STOP)
1183 			return;
1184 
1185 		ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1186 	}
1187 }
1188 
1189 /*
1190  * ==========================================================================
1191  * Initiate I/O, either sync or async
1192  * ==========================================================================
1193  */
1194 int
1195 zio_wait(zio_t *zio)
1196 {
1197 	int error;
1198 
1199 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1200 	ASSERT(zio->io_executor == NULL);
1201 
1202 	zio->io_waiter = curthread;
1203 
1204 	zio_execute(zio);
1205 
1206 	mutex_enter(&zio->io_lock);
1207 	while (zio->io_executor != NULL)
1208 		cv_wait(&zio->io_cv, &zio->io_lock);
1209 	mutex_exit(&zio->io_lock);
1210 
1211 	error = zio->io_error;
1212 	zio_destroy(zio);
1213 
1214 	return (error);
1215 }
1216 
1217 void
1218 zio_nowait(zio_t *zio)
1219 {
1220 	ASSERT(zio->io_executor == NULL);
1221 
1222 	if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1223 	    zio_unique_parent(zio) == NULL) {
1224 		/*
1225 		 * This is a logical async I/O with no parent to wait for it.
1226 		 * We add it to the spa_async_root_zio "Godfather" I/O which
1227 		 * will ensure they complete prior to unloading the pool.
1228 		 */
1229 		spa_t *spa = zio->io_spa;
1230 
1231 		zio_add_child(spa->spa_async_zio_root, zio);
1232 	}
1233 
1234 	zio_execute(zio);
1235 }
1236 
1237 /*
1238  * ==========================================================================
1239  * Reexecute or suspend/resume failed I/O
1240  * ==========================================================================
1241  */
1242 
1243 static void
1244 zio_reexecute(zio_t *pio)
1245 {
1246 	zio_t *cio, *cio_next;
1247 
1248 	ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1249 	ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1250 	ASSERT(pio->io_gang_leader == NULL);
1251 	ASSERT(pio->io_gang_tree == NULL);
1252 
1253 	pio->io_flags = pio->io_orig_flags;
1254 	pio->io_stage = pio->io_orig_stage;
1255 	pio->io_pipeline = pio->io_orig_pipeline;
1256 	pio->io_reexecute = 0;
1257 	pio->io_error = 0;
1258 	for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1259 		pio->io_state[w] = 0;
1260 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1261 		pio->io_child_error[c] = 0;
1262 
1263 	if (IO_IS_ALLOCATING(pio))
1264 		BP_ZERO(pio->io_bp);
1265 
1266 	/*
1267 	 * As we reexecute pio's children, new children could be created.
1268 	 * New children go to the head of pio's io_child_list, however,
1269 	 * so we will (correctly) not reexecute them.  The key is that
1270 	 * the remainder of pio's io_child_list, from 'cio_next' onward,
1271 	 * cannot be affected by any side effects of reexecuting 'cio'.
1272 	 */
1273 	for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1274 		cio_next = zio_walk_children(pio);
1275 		mutex_enter(&pio->io_lock);
1276 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1277 			pio->io_children[cio->io_child_type][w]++;
1278 		mutex_exit(&pio->io_lock);
1279 		zio_reexecute(cio);
1280 	}
1281 
1282 	/*
1283 	 * Now that all children have been reexecuted, execute the parent.
1284 	 * We don't reexecute "The Godfather" I/O here as it's the
1285 	 * responsibility of the caller to wait on him.
1286 	 */
1287 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1288 		zio_execute(pio);
1289 }
1290 
1291 void
1292 zio_suspend(spa_t *spa, zio_t *zio)
1293 {
1294 	if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1295 		fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1296 		    "failure and the failure mode property for this pool "
1297 		    "is set to panic.", spa_name(spa));
1298 
1299 	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1300 
1301 	mutex_enter(&spa->spa_suspend_lock);
1302 
1303 	if (spa->spa_suspend_zio_root == NULL)
1304 		spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1305 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1306 		    ZIO_FLAG_GODFATHER);
1307 
1308 	spa->spa_suspended = B_TRUE;
1309 
1310 	if (zio != NULL) {
1311 		ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1312 		ASSERT(zio != spa->spa_suspend_zio_root);
1313 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1314 		ASSERT(zio_unique_parent(zio) == NULL);
1315 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1316 		zio_add_child(spa->spa_suspend_zio_root, zio);
1317 	}
1318 
1319 	mutex_exit(&spa->spa_suspend_lock);
1320 }
1321 
1322 int
1323 zio_resume(spa_t *spa)
1324 {
1325 	zio_t *pio;
1326 
1327 	/*
1328 	 * Reexecute all previously suspended i/o.
1329 	 */
1330 	mutex_enter(&spa->spa_suspend_lock);
1331 	spa->spa_suspended = B_FALSE;
1332 	cv_broadcast(&spa->spa_suspend_cv);
1333 	pio = spa->spa_suspend_zio_root;
1334 	spa->spa_suspend_zio_root = NULL;
1335 	mutex_exit(&spa->spa_suspend_lock);
1336 
1337 	if (pio == NULL)
1338 		return (0);
1339 
1340 	zio_reexecute(pio);
1341 	return (zio_wait(pio));
1342 }
1343 
1344 void
1345 zio_resume_wait(spa_t *spa)
1346 {
1347 	mutex_enter(&spa->spa_suspend_lock);
1348 	while (spa_suspended(spa))
1349 		cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1350 	mutex_exit(&spa->spa_suspend_lock);
1351 }
1352 
1353 /*
1354  * ==========================================================================
1355  * Gang blocks.
1356  *
1357  * A gang block is a collection of small blocks that looks to the DMU
1358  * like one large block.  When zio_dva_allocate() cannot find a block
1359  * of the requested size, due to either severe fragmentation or the pool
1360  * being nearly full, it calls zio_write_gang_block() to construct the
1361  * block from smaller fragments.
1362  *
1363  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1364  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
1365  * an indirect block: it's an array of block pointers.  It consumes
1366  * only one sector and hence is allocatable regardless of fragmentation.
1367  * The gang header's bps point to its gang members, which hold the data.
1368  *
1369  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1370  * as the verifier to ensure uniqueness of the SHA256 checksum.
1371  * Critically, the gang block bp's blk_cksum is the checksum of the data,
1372  * not the gang header.  This ensures that data block signatures (needed for
1373  * deduplication) are independent of how the block is physically stored.
1374  *
1375  * Gang blocks can be nested: a gang member may itself be a gang block.
1376  * Thus every gang block is a tree in which root and all interior nodes are
1377  * gang headers, and the leaves are normal blocks that contain user data.
1378  * The root of the gang tree is called the gang leader.
1379  *
1380  * To perform any operation (read, rewrite, free, claim) on a gang block,
1381  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1382  * in the io_gang_tree field of the original logical i/o by recursively
1383  * reading the gang leader and all gang headers below it.  This yields
1384  * an in-core tree containing the contents of every gang header and the
1385  * bps for every constituent of the gang block.
1386  *
1387  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1388  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
1389  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1390  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1391  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1392  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
1393  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1394  * of the gang header plus zio_checksum_compute() of the data to update the
1395  * gang header's blk_cksum as described above.
1396  *
1397  * The two-phase assemble/issue model solves the problem of partial failure --
1398  * what if you'd freed part of a gang block but then couldn't read the
1399  * gang header for another part?  Assembling the entire gang tree first
1400  * ensures that all the necessary gang header I/O has succeeded before
1401  * starting the actual work of free, claim, or write.  Once the gang tree
1402  * is assembled, free and claim are in-memory operations that cannot fail.
1403  *
1404  * In the event that a gang write fails, zio_dva_unallocate() walks the
1405  * gang tree to immediately free (i.e. insert back into the space map)
1406  * everything we've allocated.  This ensures that we don't get ENOSPC
1407  * errors during repeated suspend/resume cycles due to a flaky device.
1408  *
1409  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
1410  * the gang tree, we won't modify the block, so we can safely defer the free
1411  * (knowing that the block is still intact).  If we *can* assemble the gang
1412  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1413  * each constituent bp and we can allocate a new block on the next sync pass.
1414  *
1415  * In all cases, the gang tree allows complete recovery from partial failure.
1416  * ==========================================================================
1417  */
1418 
1419 static zio_t *
1420 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1421 {
1422 	if (gn != NULL)
1423 		return (pio);
1424 
1425 	return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1426 	    NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1427 	    &pio->io_bookmark));
1428 }
1429 
1430 zio_t *
1431 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1432 {
1433 	zio_t *zio;
1434 
1435 	if (gn != NULL) {
1436 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1437 		    gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1438 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1439 		/*
1440 		 * As we rewrite each gang header, the pipeline will compute
1441 		 * a new gang block header checksum for it; but no one will
1442 		 * compute a new data checksum, so we do that here.  The one
1443 		 * exception is the gang leader: the pipeline already computed
1444 		 * its data checksum because that stage precedes gang assembly.
1445 		 * (Presently, nothing actually uses interior data checksums;
1446 		 * this is just good hygiene.)
1447 		 */
1448 		if (gn != pio->io_gang_leader->io_gang_tree) {
1449 			zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1450 			    data, BP_GET_PSIZE(bp));
1451 		}
1452 		/*
1453 		 * If we are here to damage data for testing purposes,
1454 		 * leave the GBH alone so that we can detect the damage.
1455 		 */
1456 		if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1457 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1458 	} else {
1459 		zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1460 		    data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1461 		    ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1462 	}
1463 
1464 	return (zio);
1465 }
1466 
1467 /* ARGSUSED */
1468 zio_t *
1469 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1470 {
1471 	return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1472 	    ZIO_GANG_CHILD_FLAGS(pio)));
1473 }
1474 
1475 /* ARGSUSED */
1476 zio_t *
1477 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1478 {
1479 	return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1480 	    NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1481 }
1482 
1483 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1484 	NULL,
1485 	zio_read_gang,
1486 	zio_rewrite_gang,
1487 	zio_free_gang,
1488 	zio_claim_gang,
1489 	NULL
1490 };
1491 
1492 static void zio_gang_tree_assemble_done(zio_t *zio);
1493 
1494 static zio_gang_node_t *
1495 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1496 {
1497 	zio_gang_node_t *gn;
1498 
1499 	ASSERT(*gnpp == NULL);
1500 
1501 	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1502 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1503 	*gnpp = gn;
1504 
1505 	return (gn);
1506 }
1507 
1508 static void
1509 zio_gang_node_free(zio_gang_node_t **gnpp)
1510 {
1511 	zio_gang_node_t *gn = *gnpp;
1512 
1513 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1514 		ASSERT(gn->gn_child[g] == NULL);
1515 
1516 	zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1517 	kmem_free(gn, sizeof (*gn));
1518 	*gnpp = NULL;
1519 }
1520 
1521 static void
1522 zio_gang_tree_free(zio_gang_node_t **gnpp)
1523 {
1524 	zio_gang_node_t *gn = *gnpp;
1525 
1526 	if (gn == NULL)
1527 		return;
1528 
1529 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1530 		zio_gang_tree_free(&gn->gn_child[g]);
1531 
1532 	zio_gang_node_free(gnpp);
1533 }
1534 
1535 static void
1536 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1537 {
1538 	zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1539 
1540 	ASSERT(gio->io_gang_leader == gio);
1541 	ASSERT(BP_IS_GANG(bp));
1542 
1543 	zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1544 	    SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1545 	    gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1546 }
1547 
1548 static void
1549 zio_gang_tree_assemble_done(zio_t *zio)
1550 {
1551 	zio_t *gio = zio->io_gang_leader;
1552 	zio_gang_node_t *gn = zio->io_private;
1553 	blkptr_t *bp = zio->io_bp;
1554 
1555 	ASSERT(gio == zio_unique_parent(zio));
1556 	ASSERT(zio->io_child_count == 0);
1557 
1558 	if (zio->io_error)
1559 		return;
1560 
1561 	if (BP_SHOULD_BYTESWAP(bp))
1562 		byteswap_uint64_array(zio->io_data, zio->io_size);
1563 
1564 	ASSERT(zio->io_data == gn->gn_gbh);
1565 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1566 	ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1567 
1568 	for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1569 		blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1570 		if (!BP_IS_GANG(gbp))
1571 			continue;
1572 		zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1573 	}
1574 }
1575 
1576 static void
1577 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1578 {
1579 	zio_t *gio = pio->io_gang_leader;
1580 	zio_t *zio;
1581 
1582 	ASSERT(BP_IS_GANG(bp) == !!gn);
1583 	ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1584 	ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1585 
1586 	/*
1587 	 * If you're a gang header, your data is in gn->gn_gbh.
1588 	 * If you're a gang member, your data is in 'data' and gn == NULL.
1589 	 */
1590 	zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1591 
1592 	if (gn != NULL) {
1593 		ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1594 
1595 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1596 			blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1597 			if (BP_IS_HOLE(gbp))
1598 				continue;
1599 			zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1600 			data = (char *)data + BP_GET_PSIZE(gbp);
1601 		}
1602 	}
1603 
1604 	if (gn == gio->io_gang_tree)
1605 		ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1606 
1607 	if (zio != pio)
1608 		zio_nowait(zio);
1609 }
1610 
1611 static int
1612 zio_gang_assemble(zio_t *zio)
1613 {
1614 	blkptr_t *bp = zio->io_bp;
1615 
1616 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1617 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1618 
1619 	zio->io_gang_leader = zio;
1620 
1621 	zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1622 
1623 	return (ZIO_PIPELINE_CONTINUE);
1624 }
1625 
1626 static int
1627 zio_gang_issue(zio_t *zio)
1628 {
1629 	blkptr_t *bp = zio->io_bp;
1630 
1631 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1632 		return (ZIO_PIPELINE_STOP);
1633 
1634 	ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1635 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1636 
1637 	if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1638 		zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1639 	else
1640 		zio_gang_tree_free(&zio->io_gang_tree);
1641 
1642 	zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1643 
1644 	return (ZIO_PIPELINE_CONTINUE);
1645 }
1646 
1647 static void
1648 zio_write_gang_member_ready(zio_t *zio)
1649 {
1650 	zio_t *pio = zio_unique_parent(zio);
1651 	zio_t *gio = zio->io_gang_leader;
1652 	dva_t *cdva = zio->io_bp->blk_dva;
1653 	dva_t *pdva = pio->io_bp->blk_dva;
1654 	uint64_t asize;
1655 
1656 	if (BP_IS_HOLE(zio->io_bp))
1657 		return;
1658 
1659 	ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1660 
1661 	ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1662 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1663 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1664 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1665 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1666 
1667 	mutex_enter(&pio->io_lock);
1668 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1669 		ASSERT(DVA_GET_GANG(&pdva[d]));
1670 		asize = DVA_GET_ASIZE(&pdva[d]);
1671 		asize += DVA_GET_ASIZE(&cdva[d]);
1672 		DVA_SET_ASIZE(&pdva[d], asize);
1673 	}
1674 	mutex_exit(&pio->io_lock);
1675 }
1676 
1677 static int
1678 zio_write_gang_block(zio_t *pio)
1679 {
1680 	spa_t *spa = pio->io_spa;
1681 	blkptr_t *bp = pio->io_bp;
1682 	zio_t *gio = pio->io_gang_leader;
1683 	zio_t *zio;
1684 	zio_gang_node_t *gn, **gnpp;
1685 	zio_gbh_phys_t *gbh;
1686 	uint64_t txg = pio->io_txg;
1687 	uint64_t resid = pio->io_size;
1688 	uint64_t lsize;
1689 	int copies = gio->io_prop.zp_copies;
1690 	int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1691 	zio_prop_t zp;
1692 	int error;
1693 
1694 	error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1695 	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1696 	    METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1697 	if (error) {
1698 		pio->io_error = error;
1699 		return (ZIO_PIPELINE_CONTINUE);
1700 	}
1701 
1702 	if (pio == gio) {
1703 		gnpp = &gio->io_gang_tree;
1704 	} else {
1705 		gnpp = pio->io_private;
1706 		ASSERT(pio->io_ready == zio_write_gang_member_ready);
1707 	}
1708 
1709 	gn = zio_gang_node_alloc(gnpp);
1710 	gbh = gn->gn_gbh;
1711 	bzero(gbh, SPA_GANGBLOCKSIZE);
1712 
1713 	/*
1714 	 * Create the gang header.
1715 	 */
1716 	zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1717 	    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1718 
1719 	/*
1720 	 * Create and nowait the gang children.
1721 	 */
1722 	for (int g = 0; resid != 0; resid -= lsize, g++) {
1723 		lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1724 		    SPA_MINBLOCKSIZE);
1725 		ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1726 
1727 		zp.zp_checksum = gio->io_prop.zp_checksum;
1728 		zp.zp_compress = ZIO_COMPRESS_OFF;
1729 		zp.zp_type = DMU_OT_NONE;
1730 		zp.zp_level = 0;
1731 		zp.zp_copies = gio->io_prop.zp_copies;
1732 		zp.zp_dedup = 0;
1733 		zp.zp_dedup_verify = 0;
1734 
1735 		zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1736 		    (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1737 		    zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1738 		    pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1739 		    &pio->io_bookmark));
1740 	}
1741 
1742 	/*
1743 	 * Set pio's pipeline to just wait for zio to finish.
1744 	 */
1745 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1746 
1747 	zio_nowait(zio);
1748 
1749 	return (ZIO_PIPELINE_CONTINUE);
1750 }
1751 
1752 /*
1753  * ==========================================================================
1754  * Dedup
1755  * ==========================================================================
1756  */
1757 static void
1758 zio_ddt_child_read_done(zio_t *zio)
1759 {
1760 	blkptr_t *bp = zio->io_bp;
1761 	ddt_entry_t *dde = zio->io_private;
1762 	ddt_phys_t *ddp;
1763 	zio_t *pio = zio_unique_parent(zio);
1764 
1765 	mutex_enter(&pio->io_lock);
1766 	ddp = ddt_phys_select(dde, bp);
1767 	if (zio->io_error == 0)
1768 		ddt_phys_clear(ddp);	/* this ddp doesn't need repair */
1769 	if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1770 		dde->dde_repair_data = zio->io_data;
1771 	else
1772 		zio_buf_free(zio->io_data, zio->io_size);
1773 	mutex_exit(&pio->io_lock);
1774 }
1775 
1776 static int
1777 zio_ddt_read_start(zio_t *zio)
1778 {
1779 	blkptr_t *bp = zio->io_bp;
1780 
1781 	ASSERT(BP_GET_DEDUP(bp));
1782 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1783 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1784 
1785 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1786 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1787 		ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1788 		ddt_phys_t *ddp = dde->dde_phys;
1789 		ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1790 		blkptr_t blk;
1791 
1792 		ASSERT(zio->io_vsd == NULL);
1793 		zio->io_vsd = dde;
1794 
1795 		if (ddp_self == NULL)
1796 			return (ZIO_PIPELINE_CONTINUE);
1797 
1798 		for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1799 			if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1800 				continue;
1801 			ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1802 			    &blk);
1803 			zio_nowait(zio_read(zio, zio->io_spa, &blk,
1804 			    zio_buf_alloc(zio->io_size), zio->io_size,
1805 			    zio_ddt_child_read_done, dde, zio->io_priority,
1806 			    ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1807 			    &zio->io_bookmark));
1808 		}
1809 		return (ZIO_PIPELINE_CONTINUE);
1810 	}
1811 
1812 	zio_nowait(zio_read(zio, zio->io_spa, bp,
1813 	    zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1814 	    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1815 
1816 	return (ZIO_PIPELINE_CONTINUE);
1817 }
1818 
1819 static int
1820 zio_ddt_read_done(zio_t *zio)
1821 {
1822 	blkptr_t *bp = zio->io_bp;
1823 
1824 	if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1825 		return (ZIO_PIPELINE_STOP);
1826 
1827 	ASSERT(BP_GET_DEDUP(bp));
1828 	ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1829 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1830 
1831 	if (zio->io_child_error[ZIO_CHILD_DDT]) {
1832 		ddt_t *ddt = ddt_select(zio->io_spa, bp);
1833 		ddt_entry_t *dde = zio->io_vsd;
1834 		if (ddt == NULL) {
1835 			ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1836 			return (ZIO_PIPELINE_CONTINUE);
1837 		}
1838 		if (dde == NULL) {
1839 			zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1840 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1841 			return (ZIO_PIPELINE_STOP);
1842 		}
1843 		if (dde->dde_repair_data != NULL) {
1844 			bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1845 			zio->io_child_error[ZIO_CHILD_DDT] = 0;
1846 		}
1847 		ddt_repair_done(ddt, dde);
1848 		zio->io_vsd = NULL;
1849 	}
1850 
1851 	ASSERT(zio->io_vsd == NULL);
1852 
1853 	return (ZIO_PIPELINE_CONTINUE);
1854 }
1855 
1856 static boolean_t
1857 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1858 {
1859 	spa_t *spa = zio->io_spa;
1860 
1861 	/*
1862 	 * Note: we compare the original data, not the transformed data,
1863 	 * because when zio->io_bp is an override bp, we will not have
1864 	 * pushed the I/O transforms.  That's an important optimization
1865 	 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1866 	 */
1867 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1868 		zio_t *lio = dde->dde_lead_zio[p];
1869 
1870 		if (lio != NULL) {
1871 			return (lio->io_orig_size != zio->io_orig_size ||
1872 			    bcmp(zio->io_orig_data, lio->io_orig_data,
1873 			    zio->io_orig_size) != 0);
1874 		}
1875 	}
1876 
1877 	for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1878 		ddt_phys_t *ddp = &dde->dde_phys[p];
1879 
1880 		if (ddp->ddp_phys_birth != 0) {
1881 			arc_buf_t *abuf = NULL;
1882 			uint32_t aflags = ARC_WAIT;
1883 			blkptr_t blk = *zio->io_bp;
1884 			int error;
1885 
1886 			ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1887 
1888 			ddt_exit(ddt);
1889 
1890 			error = arc_read_nolock(NULL, spa, &blk,
1891 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1892 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1893 			    &aflags, &zio->io_bookmark);
1894 
1895 			if (error == 0) {
1896 				if (arc_buf_size(abuf) != zio->io_orig_size ||
1897 				    bcmp(abuf->b_data, zio->io_orig_data,
1898 				    zio->io_orig_size) != 0)
1899 					error = EEXIST;
1900 				VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1901 			}
1902 
1903 			ddt_enter(ddt);
1904 			return (error != 0);
1905 		}
1906 	}
1907 
1908 	return (B_FALSE);
1909 }
1910 
1911 static void
1912 zio_ddt_child_write_ready(zio_t *zio)
1913 {
1914 	int p = zio->io_prop.zp_copies;
1915 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1916 	ddt_entry_t *dde = zio->io_private;
1917 	ddt_phys_t *ddp = &dde->dde_phys[p];
1918 	zio_t *pio;
1919 
1920 	if (zio->io_error)
1921 		return;
1922 
1923 	ddt_enter(ddt);
1924 
1925 	ASSERT(dde->dde_lead_zio[p] == zio);
1926 
1927 	ddt_phys_fill(ddp, zio->io_bp);
1928 
1929 	while ((pio = zio_walk_parents(zio)) != NULL)
1930 		ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1931 
1932 	ddt_exit(ddt);
1933 }
1934 
1935 static void
1936 zio_ddt_child_write_done(zio_t *zio)
1937 {
1938 	int p = zio->io_prop.zp_copies;
1939 	ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1940 	ddt_entry_t *dde = zio->io_private;
1941 	ddt_phys_t *ddp = &dde->dde_phys[p];
1942 
1943 	ddt_enter(ddt);
1944 
1945 	ASSERT(ddp->ddp_refcnt == 0);
1946 	ASSERT(dde->dde_lead_zio[p] == zio);
1947 	dde->dde_lead_zio[p] = NULL;
1948 
1949 	if (zio->io_error == 0) {
1950 		while (zio_walk_parents(zio) != NULL)
1951 			ddt_phys_addref(ddp);
1952 	} else {
1953 		ddt_phys_clear(ddp);
1954 	}
1955 
1956 	ddt_exit(ddt);
1957 }
1958 
1959 static void
1960 zio_ddt_ditto_write_done(zio_t *zio)
1961 {
1962 	int p = DDT_PHYS_DITTO;
1963 	zio_prop_t *zp = &zio->io_prop;
1964 	blkptr_t *bp = zio->io_bp;
1965 	ddt_t *ddt = ddt_select(zio->io_spa, bp);
1966 	ddt_entry_t *dde = zio->io_private;
1967 	ddt_phys_t *ddp = &dde->dde_phys[p];
1968 	ddt_key_t *ddk = &dde->dde_key;
1969 
1970 	ddt_enter(ddt);
1971 
1972 	ASSERT(ddp->ddp_refcnt == 0);
1973 	ASSERT(dde->dde_lead_zio[p] == zio);
1974 	dde->dde_lead_zio[p] = NULL;
1975 
1976 	if (zio->io_error == 0) {
1977 		ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1978 		ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1979 		ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1980 		if (ddp->ddp_phys_birth != 0)
1981 			ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1982 		ddt_phys_fill(ddp, bp);
1983 	}
1984 
1985 	ddt_exit(ddt);
1986 }
1987 
1988 static int
1989 zio_ddt_write(zio_t *zio)
1990 {
1991 	spa_t *spa = zio->io_spa;
1992 	blkptr_t *bp = zio->io_bp;
1993 	uint64_t txg = zio->io_txg;
1994 	zio_prop_t *zp = &zio->io_prop;
1995 	int p = zp->zp_copies;
1996 	int ditto_copies;
1997 	zio_t *cio = NULL;
1998 	zio_t *dio = NULL;
1999 	ddt_t *ddt = ddt_select(spa, bp);
2000 	ddt_entry_t *dde;
2001 	ddt_phys_t *ddp;
2002 
2003 	ASSERT(BP_GET_DEDUP(bp));
2004 	ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2005 	ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2006 
2007 	ddt_enter(ddt);
2008 	dde = ddt_lookup(ddt, bp, B_TRUE);
2009 	ddp = &dde->dde_phys[p];
2010 
2011 	if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2012 		/*
2013 		 * If we're using a weak checksum, upgrade to a strong checksum
2014 		 * and try again.  If we're already using a strong checksum,
2015 		 * we can't resolve it, so just convert to an ordinary write.
2016 		 * (And automatically e-mail a paper to Nature?)
2017 		 */
2018 		if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2019 			zp->zp_checksum = spa_dedup_checksum(spa);
2020 			zio_pop_transforms(zio);
2021 			zio->io_stage = ZIO_STAGE_OPEN;
2022 			BP_ZERO(bp);
2023 		} else {
2024 			zp->zp_dedup = 0;
2025 		}
2026 		zio->io_pipeline = ZIO_WRITE_PIPELINE;
2027 		ddt_exit(ddt);
2028 		return (ZIO_PIPELINE_CONTINUE);
2029 	}
2030 
2031 	ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2032 	ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2033 
2034 	if (ditto_copies > ddt_ditto_copies_present(dde) &&
2035 	    dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2036 		zio_prop_t czp = *zp;
2037 
2038 		czp.zp_copies = ditto_copies;
2039 
2040 		/*
2041 		 * If we arrived here with an override bp, we won't have run
2042 		 * the transform stack, so we won't have the data we need to
2043 		 * generate a child i/o.  So, toss the override bp and restart.
2044 		 * This is safe, because using the override bp is just an
2045 		 * optimization; and it's rare, so the cost doesn't matter.
2046 		 */
2047 		if (zio->io_bp_override) {
2048 			zio_pop_transforms(zio);
2049 			zio->io_stage = ZIO_STAGE_OPEN;
2050 			zio->io_pipeline = ZIO_WRITE_PIPELINE;
2051 			zio->io_bp_override = NULL;
2052 			BP_ZERO(bp);
2053 			ddt_exit(ddt);
2054 			return (ZIO_PIPELINE_CONTINUE);
2055 		}
2056 
2057 		dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2058 		    zio->io_orig_size, &czp, NULL,
2059 		    zio_ddt_ditto_write_done, dde, zio->io_priority,
2060 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2061 
2062 		zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2063 		dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2064 	}
2065 
2066 	if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2067 		if (ddp->ddp_phys_birth != 0)
2068 			ddt_bp_fill(ddp, bp, txg);
2069 		if (dde->dde_lead_zio[p] != NULL)
2070 			zio_add_child(zio, dde->dde_lead_zio[p]);
2071 		else
2072 			ddt_phys_addref(ddp);
2073 	} else if (zio->io_bp_override) {
2074 		ASSERT(bp->blk_birth == txg);
2075 		ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2076 		ddt_phys_fill(ddp, bp);
2077 		ddt_phys_addref(ddp);
2078 	} else {
2079 		cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2080 		    zio->io_orig_size, zp, zio_ddt_child_write_ready,
2081 		    zio_ddt_child_write_done, dde, zio->io_priority,
2082 		    ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2083 
2084 		zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2085 		dde->dde_lead_zio[p] = cio;
2086 	}
2087 
2088 	ddt_exit(ddt);
2089 
2090 	if (cio)
2091 		zio_nowait(cio);
2092 	if (dio)
2093 		zio_nowait(dio);
2094 
2095 	return (ZIO_PIPELINE_CONTINUE);
2096 }
2097 
2098 ddt_entry_t *freedde; /* for debugging */
2099 
2100 static int
2101 zio_ddt_free(zio_t *zio)
2102 {
2103 	spa_t *spa = zio->io_spa;
2104 	blkptr_t *bp = zio->io_bp;
2105 	ddt_t *ddt = ddt_select(spa, bp);
2106 	ddt_entry_t *dde;
2107 	ddt_phys_t *ddp;
2108 
2109 	ASSERT(BP_GET_DEDUP(bp));
2110 	ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2111 
2112 	ddt_enter(ddt);
2113 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2114 	ddp = ddt_phys_select(dde, bp);
2115 	ddt_phys_decref(ddp);
2116 	ddt_exit(ddt);
2117 
2118 	return (ZIO_PIPELINE_CONTINUE);
2119 }
2120 
2121 /*
2122  * ==========================================================================
2123  * Allocate and free blocks
2124  * ==========================================================================
2125  */
2126 static int
2127 zio_dva_allocate(zio_t *zio)
2128 {
2129 	spa_t *spa = zio->io_spa;
2130 	metaslab_class_t *mc = spa_normal_class(spa);
2131 	blkptr_t *bp = zio->io_bp;
2132 	int error;
2133 	int flags = 0;
2134 
2135 	if (zio->io_gang_leader == NULL) {
2136 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2137 		zio->io_gang_leader = zio;
2138 	}
2139 
2140 	ASSERT(BP_IS_HOLE(bp));
2141 	ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2142 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
2143 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2144 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2145 
2146 	/*
2147 	 * The dump device does not support gang blocks so allocation on
2148 	 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2149 	 * the "fast" gang feature.
2150 	 */
2151 	flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2152 	flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2153 	    METASLAB_GANG_CHILD : 0;
2154 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
2155 	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2156 
2157 	if (error) {
2158 		spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2159 		    "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2160 		    error);
2161 		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2162 			return (zio_write_gang_block(zio));
2163 		zio->io_error = error;
2164 	}
2165 
2166 	return (ZIO_PIPELINE_CONTINUE);
2167 }
2168 
2169 static int
2170 zio_dva_free(zio_t *zio)
2171 {
2172 	metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2173 
2174 	return (ZIO_PIPELINE_CONTINUE);
2175 }
2176 
2177 static int
2178 zio_dva_claim(zio_t *zio)
2179 {
2180 	int error;
2181 
2182 	error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2183 	if (error)
2184 		zio->io_error = error;
2185 
2186 	return (ZIO_PIPELINE_CONTINUE);
2187 }
2188 
2189 /*
2190  * Undo an allocation.  This is used by zio_done() when an I/O fails
2191  * and we want to give back the block we just allocated.
2192  * This handles both normal blocks and gang blocks.
2193  */
2194 static void
2195 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2196 {
2197 	ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2198 	ASSERT(zio->io_bp_override == NULL);
2199 
2200 	if (!BP_IS_HOLE(bp))
2201 		metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2202 
2203 	if (gn != NULL) {
2204 		for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2205 			zio_dva_unallocate(zio, gn->gn_child[g],
2206 			    &gn->gn_gbh->zg_blkptr[g]);
2207 		}
2208 	}
2209 }
2210 
2211 /*
2212  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
2213  */
2214 int
2215 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2216     uint64_t size, boolean_t use_slog)
2217 {
2218 	int error = 1;
2219 
2220 	ASSERT(txg > spa_syncing_txg(spa));
2221 
2222 	if (use_slog)
2223 		error = metaslab_alloc(spa, spa_log_class(spa), size,
2224 		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2225 
2226 	if (error)
2227 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
2228 		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2229 
2230 	if (error == 0) {
2231 		BP_SET_LSIZE(new_bp, size);
2232 		BP_SET_PSIZE(new_bp, size);
2233 		BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2234 		BP_SET_CHECKSUM(new_bp,
2235 		    spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2236 		    ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2237 		BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2238 		BP_SET_LEVEL(new_bp, 0);
2239 		BP_SET_DEDUP(new_bp, 0);
2240 		BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2241 	}
2242 
2243 	return (error);
2244 }
2245 
2246 /*
2247  * Free an intent log block.
2248  */
2249 void
2250 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2251 {
2252 	ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2253 	ASSERT(!BP_IS_GANG(bp));
2254 
2255 	zio_free(spa, txg, bp);
2256 }
2257 
2258 /*
2259  * ==========================================================================
2260  * Read and write to physical devices
2261  * ==========================================================================
2262  */
2263 static int
2264 zio_vdev_io_start(zio_t *zio)
2265 {
2266 	vdev_t *vd = zio->io_vd;
2267 	uint64_t align;
2268 	spa_t *spa = zio->io_spa;
2269 
2270 	ASSERT(zio->io_error == 0);
2271 	ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2272 
2273 	if (vd == NULL) {
2274 		if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2275 			spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2276 
2277 		/*
2278 		 * The mirror_ops handle multiple DVAs in a single BP.
2279 		 */
2280 		return (vdev_mirror_ops.vdev_op_io_start(zio));
2281 	}
2282 
2283 	/*
2284 	 * We keep track of time-sensitive I/Os so that the scan thread
2285 	 * can quickly react to certain workloads.  In particular, we care
2286 	 * about non-scrubbing, top-level reads and writes with the following
2287 	 * characteristics:
2288 	 * 	- synchronous writes of user data to non-slog devices
2289 	 *	- any reads of user data
2290 	 * When these conditions are met, adjust the timestamp of spa_last_io
2291 	 * which allows the scan thread to adjust its workload accordingly.
2292 	 */
2293 	if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2294 	    vd == vd->vdev_top && !vd->vdev_islog &&
2295 	    zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2296 	    zio->io_txg != spa_syncing_txg(spa)) {
2297 		uint64_t old = spa->spa_last_io;
2298 		uint64_t new = ddi_get_lbolt64();
2299 		if (old != new)
2300 			(void) atomic_cas_64(&spa->spa_last_io, old, new);
2301 	}
2302 
2303 	align = 1ULL << vd->vdev_top->vdev_ashift;
2304 
2305 	if (P2PHASE(zio->io_size, align) != 0) {
2306 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
2307 		char *abuf = zio_buf_alloc(asize);
2308 		ASSERT(vd == vd->vdev_top);
2309 		if (zio->io_type == ZIO_TYPE_WRITE) {
2310 			bcopy(zio->io_data, abuf, zio->io_size);
2311 			bzero(abuf + zio->io_size, asize - zio->io_size);
2312 		}
2313 		zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2314 	}
2315 
2316 	ASSERT(P2PHASE(zio->io_offset, align) == 0);
2317 	ASSERT(P2PHASE(zio->io_size, align) == 0);
2318 	VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2319 
2320 	/*
2321 	 * If this is a repair I/O, and there's no self-healing involved --
2322 	 * that is, we're just resilvering what we expect to resilver --
2323 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2324 	 * This prevents spurious resilvering with nested replication.
2325 	 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2326 	 * A is out of date, we'll read from C+D, then use the data to
2327 	 * resilver A+B -- but we don't actually want to resilver B, just A.
2328 	 * The top-level mirror has no way to know this, so instead we just
2329 	 * discard unnecessary repairs as we work our way down the vdev tree.
2330 	 * The same logic applies to any form of nested replication:
2331 	 * ditto + mirror, RAID-Z + replacing, etc.  This covers them all.
2332 	 */
2333 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2334 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2335 	    zio->io_txg != 0 &&	/* not a delegated i/o */
2336 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2337 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2338 		zio_vdev_io_bypass(zio);
2339 		return (ZIO_PIPELINE_CONTINUE);
2340 	}
2341 
2342 	if (vd->vdev_ops->vdev_op_leaf &&
2343 	    (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2344 
2345 		if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2346 			return (ZIO_PIPELINE_CONTINUE);
2347 
2348 		if ((zio = vdev_queue_io(zio)) == NULL)
2349 			return (ZIO_PIPELINE_STOP);
2350 
2351 		if (!vdev_accessible(vd, zio)) {
2352 			zio->io_error = ENXIO;
2353 			zio_interrupt(zio);
2354 			return (ZIO_PIPELINE_STOP);
2355 		}
2356 	}
2357 
2358 	return (vd->vdev_ops->vdev_op_io_start(zio));
2359 }
2360 
2361 static int
2362 zio_vdev_io_done(zio_t *zio)
2363 {
2364 	vdev_t *vd = zio->io_vd;
2365 	vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2366 	boolean_t unexpected_error = B_FALSE;
2367 
2368 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2369 		return (ZIO_PIPELINE_STOP);
2370 
2371 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2372 
2373 	if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2374 
2375 		vdev_queue_io_done(zio);
2376 
2377 		if (zio->io_type == ZIO_TYPE_WRITE)
2378 			vdev_cache_write(zio);
2379 
2380 		if (zio_injection_enabled && zio->io_error == 0)
2381 			zio->io_error = zio_handle_device_injection(vd,
2382 			    zio, EIO);
2383 
2384 		if (zio_injection_enabled && zio->io_error == 0)
2385 			zio->io_error = zio_handle_label_injection(zio, EIO);
2386 
2387 		if (zio->io_error) {
2388 			if (!vdev_accessible(vd, zio)) {
2389 				zio->io_error = ENXIO;
2390 			} else {
2391 				unexpected_error = B_TRUE;
2392 			}
2393 		}
2394 	}
2395 
2396 	ops->vdev_op_io_done(zio);
2397 
2398 	if (unexpected_error)
2399 		VERIFY(vdev_probe(vd, zio) == NULL);
2400 
2401 	return (ZIO_PIPELINE_CONTINUE);
2402 }
2403 
2404 /*
2405  * For non-raidz ZIOs, we can just copy aside the bad data read from the
2406  * disk, and use that to finish the checksum ereport later.
2407  */
2408 static void
2409 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2410     const void *good_buf)
2411 {
2412 	/* no processing needed */
2413 	zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2414 }
2415 
2416 /*ARGSUSED*/
2417 void
2418 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2419 {
2420 	void *buf = zio_buf_alloc(zio->io_size);
2421 
2422 	bcopy(zio->io_data, buf, zio->io_size);
2423 
2424 	zcr->zcr_cbinfo = zio->io_size;
2425 	zcr->zcr_cbdata = buf;
2426 	zcr->zcr_finish = zio_vsd_default_cksum_finish;
2427 	zcr->zcr_free = zio_buf_free;
2428 }
2429 
2430 static int
2431 zio_vdev_io_assess(zio_t *zio)
2432 {
2433 	vdev_t *vd = zio->io_vd;
2434 
2435 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2436 		return (ZIO_PIPELINE_STOP);
2437 
2438 	if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2439 		spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2440 
2441 	if (zio->io_vsd != NULL) {
2442 		zio->io_vsd_ops->vsd_free(zio);
2443 		zio->io_vsd = NULL;
2444 	}
2445 
2446 	if (zio_injection_enabled && zio->io_error == 0)
2447 		zio->io_error = zio_handle_fault_injection(zio, EIO);
2448 
2449 	/*
2450 	 * If the I/O failed, determine whether we should attempt to retry it.
2451 	 *
2452 	 * On retry, we cut in line in the issue queue, since we don't want
2453 	 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2454 	 */
2455 	if (zio->io_error && vd == NULL &&
2456 	    !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2457 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));	/* not a leaf */
2458 		ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));	/* not a leaf */
2459 		zio->io_error = 0;
2460 		zio->io_flags |= ZIO_FLAG_IO_RETRY |
2461 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2462 		zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2463 		zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2464 		    zio_requeue_io_start_cut_in_line);
2465 		return (ZIO_PIPELINE_STOP);
2466 	}
2467 
2468 	/*
2469 	 * If we got an error on a leaf device, convert it to ENXIO
2470 	 * if the device is not accessible at all.
2471 	 */
2472 	if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2473 	    !vdev_accessible(vd, zio))
2474 		zio->io_error = ENXIO;
2475 
2476 	/*
2477 	 * If we can't write to an interior vdev (mirror or RAID-Z),
2478 	 * set vdev_cant_write so that we stop trying to allocate from it.
2479 	 */
2480 	if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2481 	    vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2482 		vd->vdev_cant_write = B_TRUE;
2483 
2484 	if (zio->io_error)
2485 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2486 
2487 	return (ZIO_PIPELINE_CONTINUE);
2488 }
2489 
2490 void
2491 zio_vdev_io_reissue(zio_t *zio)
2492 {
2493 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2494 	ASSERT(zio->io_error == 0);
2495 
2496 	zio->io_stage >>= 1;
2497 }
2498 
2499 void
2500 zio_vdev_io_redone(zio_t *zio)
2501 {
2502 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2503 
2504 	zio->io_stage >>= 1;
2505 }
2506 
2507 void
2508 zio_vdev_io_bypass(zio_t *zio)
2509 {
2510 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2511 	ASSERT(zio->io_error == 0);
2512 
2513 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2514 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2515 }
2516 
2517 /*
2518  * ==========================================================================
2519  * Generate and verify checksums
2520  * ==========================================================================
2521  */
2522 static int
2523 zio_checksum_generate(zio_t *zio)
2524 {
2525 	blkptr_t *bp = zio->io_bp;
2526 	enum zio_checksum checksum;
2527 
2528 	if (bp == NULL) {
2529 		/*
2530 		 * This is zio_write_phys().
2531 		 * We're either generating a label checksum, or none at all.
2532 		 */
2533 		checksum = zio->io_prop.zp_checksum;
2534 
2535 		if (checksum == ZIO_CHECKSUM_OFF)
2536 			return (ZIO_PIPELINE_CONTINUE);
2537 
2538 		ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2539 	} else {
2540 		if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2541 			ASSERT(!IO_IS_ALLOCATING(zio));
2542 			checksum = ZIO_CHECKSUM_GANG_HEADER;
2543 		} else {
2544 			checksum = BP_GET_CHECKSUM(bp);
2545 		}
2546 	}
2547 
2548 	zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2549 
2550 	return (ZIO_PIPELINE_CONTINUE);
2551 }
2552 
2553 static int
2554 zio_checksum_verify(zio_t *zio)
2555 {
2556 	zio_bad_cksum_t info;
2557 	blkptr_t *bp = zio->io_bp;
2558 	int error;
2559 
2560 	ASSERT(zio->io_vd != NULL);
2561 
2562 	if (bp == NULL) {
2563 		/*
2564 		 * This is zio_read_phys().
2565 		 * We're either verifying a label checksum, or nothing at all.
2566 		 */
2567 		if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2568 			return (ZIO_PIPELINE_CONTINUE);
2569 
2570 		ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2571 	}
2572 
2573 	if ((error = zio_checksum_error(zio, &info)) != 0) {
2574 		zio->io_error = error;
2575 		if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2576 			zfs_ereport_start_checksum(zio->io_spa,
2577 			    zio->io_vd, zio, zio->io_offset,
2578 			    zio->io_size, NULL, &info);
2579 		}
2580 	}
2581 
2582 	return (ZIO_PIPELINE_CONTINUE);
2583 }
2584 
2585 /*
2586  * Called by RAID-Z to ensure we don't compute the checksum twice.
2587  */
2588 void
2589 zio_checksum_verified(zio_t *zio)
2590 {
2591 	zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2592 }
2593 
2594 /*
2595  * ==========================================================================
2596  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2597  * An error of 0 indictes success.  ENXIO indicates whole-device failure,
2598  * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
2599  * indicate errors that are specific to one I/O, and most likely permanent.
2600  * Any other error is presumed to be worse because we weren't expecting it.
2601  * ==========================================================================
2602  */
2603 int
2604 zio_worst_error(int e1, int e2)
2605 {
2606 	static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2607 	int r1, r2;
2608 
2609 	for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2610 		if (e1 == zio_error_rank[r1])
2611 			break;
2612 
2613 	for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2614 		if (e2 == zio_error_rank[r2])
2615 			break;
2616 
2617 	return (r1 > r2 ? e1 : e2);
2618 }
2619 
2620 /*
2621  * ==========================================================================
2622  * I/O completion
2623  * ==========================================================================
2624  */
2625 static int
2626 zio_ready(zio_t *zio)
2627 {
2628 	blkptr_t *bp = zio->io_bp;
2629 	zio_t *pio, *pio_next;
2630 
2631 	if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2632 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2633 		return (ZIO_PIPELINE_STOP);
2634 
2635 	if (zio->io_ready) {
2636 		ASSERT(IO_IS_ALLOCATING(zio));
2637 		ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2638 		ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2639 
2640 		zio->io_ready(zio);
2641 	}
2642 
2643 	if (bp != NULL && bp != &zio->io_bp_copy)
2644 		zio->io_bp_copy = *bp;
2645 
2646 	if (zio->io_error)
2647 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2648 
2649 	mutex_enter(&zio->io_lock);
2650 	zio->io_state[ZIO_WAIT_READY] = 1;
2651 	pio = zio_walk_parents(zio);
2652 	mutex_exit(&zio->io_lock);
2653 
2654 	/*
2655 	 * As we notify zio's parents, new parents could be added.
2656 	 * New parents go to the head of zio's io_parent_list, however,
2657 	 * so we will (correctly) not notify them.  The remainder of zio's
2658 	 * io_parent_list, from 'pio_next' onward, cannot change because
2659 	 * all parents must wait for us to be done before they can be done.
2660 	 */
2661 	for (; pio != NULL; pio = pio_next) {
2662 		pio_next = zio_walk_parents(zio);
2663 		zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2664 	}
2665 
2666 	if (zio->io_flags & ZIO_FLAG_NODATA) {
2667 		if (BP_IS_GANG(bp)) {
2668 			zio->io_flags &= ~ZIO_FLAG_NODATA;
2669 		} else {
2670 			ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2671 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2672 		}
2673 	}
2674 
2675 	if (zio_injection_enabled &&
2676 	    zio->io_spa->spa_syncing_txg == zio->io_txg)
2677 		zio_handle_ignored_writes(zio);
2678 
2679 	return (ZIO_PIPELINE_CONTINUE);
2680 }
2681 
2682 static int
2683 zio_done(zio_t *zio)
2684 {
2685 	spa_t *spa = zio->io_spa;
2686 	zio_t *lio = zio->io_logical;
2687 	blkptr_t *bp = zio->io_bp;
2688 	vdev_t *vd = zio->io_vd;
2689 	uint64_t psize = zio->io_size;
2690 	zio_t *pio, *pio_next;
2691 
2692 	/*
2693 	 * If our children haven't all completed,
2694 	 * wait for them and then repeat this pipeline stage.
2695 	 */
2696 	if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2697 	    zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2698 	    zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2699 	    zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2700 		return (ZIO_PIPELINE_STOP);
2701 
2702 	for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2703 		for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2704 			ASSERT(zio->io_children[c][w] == 0);
2705 
2706 	if (bp != NULL) {
2707 		ASSERT(bp->blk_pad[0] == 0);
2708 		ASSERT(bp->blk_pad[1] == 0);
2709 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2710 		    (bp == zio_unique_parent(zio)->io_bp));
2711 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2712 		    zio->io_bp_override == NULL &&
2713 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2714 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
2715 			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2716 			ASSERT(BP_COUNT_GANG(bp) == 0 ||
2717 			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2718 		}
2719 	}
2720 
2721 	/*
2722 	 * If there were child vdev/gang/ddt errors, they apply to us now.
2723 	 */
2724 	zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2725 	zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2726 	zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2727 
2728 	/*
2729 	 * If the I/O on the transformed data was successful, generate any
2730 	 * checksum reports now while we still have the transformed data.
2731 	 */
2732 	if (zio->io_error == 0) {
2733 		while (zio->io_cksum_report != NULL) {
2734 			zio_cksum_report_t *zcr = zio->io_cksum_report;
2735 			uint64_t align = zcr->zcr_align;
2736 			uint64_t asize = P2ROUNDUP(psize, align);
2737 			char *abuf = zio->io_data;
2738 
2739 			if (asize != psize) {
2740 				abuf = zio_buf_alloc(asize);
2741 				bcopy(zio->io_data, abuf, psize);
2742 				bzero(abuf + psize, asize - psize);
2743 			}
2744 
2745 			zio->io_cksum_report = zcr->zcr_next;
2746 			zcr->zcr_next = NULL;
2747 			zcr->zcr_finish(zcr, abuf);
2748 			zfs_ereport_free_checksum(zcr);
2749 
2750 			if (asize != psize)
2751 				zio_buf_free(abuf, asize);
2752 		}
2753 	}
2754 
2755 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
2756 
2757 	vdev_stat_update(zio, psize);
2758 
2759 	if (zio->io_error) {
2760 		/*
2761 		 * If this I/O is attached to a particular vdev,
2762 		 * generate an error message describing the I/O failure
2763 		 * at the block level.  We ignore these errors if the
2764 		 * device is currently unavailable.
2765 		 */
2766 		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2767 			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2768 
2769 		if ((zio->io_error == EIO || !(zio->io_flags &
2770 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2771 		    zio == lio) {
2772 			/*
2773 			 * For logical I/O requests, tell the SPA to log the
2774 			 * error and generate a logical data ereport.
2775 			 */
2776 			spa_log_error(spa, zio);
2777 			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2778 			    0, 0);
2779 		}
2780 	}
2781 
2782 	if (zio->io_error && zio == lio) {
2783 		/*
2784 		 * Determine whether zio should be reexecuted.  This will
2785 		 * propagate all the way to the root via zio_notify_parent().
2786 		 */
2787 		ASSERT(vd == NULL && bp != NULL);
2788 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2789 
2790 		if (IO_IS_ALLOCATING(zio) &&
2791 		    !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2792 			if (zio->io_error != ENOSPC)
2793 				zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2794 			else
2795 				zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2796 		}
2797 
2798 		if ((zio->io_type == ZIO_TYPE_READ ||
2799 		    zio->io_type == ZIO_TYPE_FREE) &&
2800 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2801 		    zio->io_error == ENXIO &&
2802 		    spa_load_state(spa) == SPA_LOAD_NONE &&
2803 		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2804 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2805 
2806 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2807 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2808 
2809 		/*
2810 		 * Here is a possibly good place to attempt to do
2811 		 * either combinatorial reconstruction or error correction
2812 		 * based on checksums.  It also might be a good place
2813 		 * to send out preliminary ereports before we suspend
2814 		 * processing.
2815 		 */
2816 	}
2817 
2818 	/*
2819 	 * If there were logical child errors, they apply to us now.
2820 	 * We defer this until now to avoid conflating logical child
2821 	 * errors with errors that happened to the zio itself when
2822 	 * updating vdev stats and reporting FMA events above.
2823 	 */
2824 	zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2825 
2826 	if ((zio->io_error || zio->io_reexecute) &&
2827 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2828 	    !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2829 		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2830 
2831 	zio_gang_tree_free(&zio->io_gang_tree);
2832 
2833 	/*
2834 	 * Godfather I/Os should never suspend.
2835 	 */
2836 	if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2837 	    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2838 		zio->io_reexecute = 0;
2839 
2840 	if (zio->io_reexecute) {
2841 		/*
2842 		 * This is a logical I/O that wants to reexecute.
2843 		 *
2844 		 * Reexecute is top-down.  When an i/o fails, if it's not
2845 		 * the root, it simply notifies its parent and sticks around.
2846 		 * The parent, seeing that it still has children in zio_done(),
2847 		 * does the same.  This percolates all the way up to the root.
2848 		 * The root i/o will reexecute or suspend the entire tree.
2849 		 *
2850 		 * This approach ensures that zio_reexecute() honors
2851 		 * all the original i/o dependency relationships, e.g.
2852 		 * parents not executing until children are ready.
2853 		 */
2854 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2855 
2856 		zio->io_gang_leader = NULL;
2857 
2858 		mutex_enter(&zio->io_lock);
2859 		zio->io_state[ZIO_WAIT_DONE] = 1;
2860 		mutex_exit(&zio->io_lock);
2861 
2862 		/*
2863 		 * "The Godfather" I/O monitors its children but is
2864 		 * not a true parent to them. It will track them through
2865 		 * the pipeline but severs its ties whenever they get into
2866 		 * trouble (e.g. suspended). This allows "The Godfather"
2867 		 * I/O to return status without blocking.
2868 		 */
2869 		for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2870 			zio_link_t *zl = zio->io_walk_link;
2871 			pio_next = zio_walk_parents(zio);
2872 
2873 			if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2874 			    (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2875 				zio_remove_child(pio, zio, zl);
2876 				zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2877 			}
2878 		}
2879 
2880 		if ((pio = zio_unique_parent(zio)) != NULL) {
2881 			/*
2882 			 * We're not a root i/o, so there's nothing to do
2883 			 * but notify our parent.  Don't propagate errors
2884 			 * upward since we haven't permanently failed yet.
2885 			 */
2886 			ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2887 			zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2888 			zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2889 		} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2890 			/*
2891 			 * We'd fail again if we reexecuted now, so suspend
2892 			 * until conditions improve (e.g. device comes online).
2893 			 */
2894 			zio_suspend(spa, zio);
2895 		} else {
2896 			/*
2897 			 * Reexecution is potentially a huge amount of work.
2898 			 * Hand it off to the otherwise-unused claim taskq.
2899 			 */
2900 			ASSERT(zio->io_tqent.tqent_next == NULL);
2901 			(void) taskq_dispatch_ent(
2902 			    spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2903 			    (task_func_t *)zio_reexecute, zio, 0,
2904 			    &zio->io_tqent);
2905 		}
2906 		return (ZIO_PIPELINE_STOP);
2907 	}
2908 
2909 	ASSERT(zio->io_child_count == 0);
2910 	ASSERT(zio->io_reexecute == 0);
2911 	ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2912 
2913 	/*
2914 	 * Report any checksum errors, since the I/O is complete.
2915 	 */
2916 	while (zio->io_cksum_report != NULL) {
2917 		zio_cksum_report_t *zcr = zio->io_cksum_report;
2918 		zio->io_cksum_report = zcr->zcr_next;
2919 		zcr->zcr_next = NULL;
2920 		zcr->zcr_finish(zcr, NULL);
2921 		zfs_ereport_free_checksum(zcr);
2922 	}
2923 
2924 	/*
2925 	 * It is the responsibility of the done callback to ensure that this
2926 	 * particular zio is no longer discoverable for adoption, and as
2927 	 * such, cannot acquire any new parents.
2928 	 */
2929 	if (zio->io_done)
2930 		zio->io_done(zio);
2931 
2932 	mutex_enter(&zio->io_lock);
2933 	zio->io_state[ZIO_WAIT_DONE] = 1;
2934 	mutex_exit(&zio->io_lock);
2935 
2936 	for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2937 		zio_link_t *zl = zio->io_walk_link;
2938 		pio_next = zio_walk_parents(zio);
2939 		zio_remove_child(pio, zio, zl);
2940 		zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2941 	}
2942 
2943 	if (zio->io_waiter != NULL) {
2944 		mutex_enter(&zio->io_lock);
2945 		zio->io_executor = NULL;
2946 		cv_broadcast(&zio->io_cv);
2947 		mutex_exit(&zio->io_lock);
2948 	} else {
2949 		zio_destroy(zio);
2950 	}
2951 
2952 	return (ZIO_PIPELINE_STOP);
2953 }
2954 
2955 /*
2956  * ==========================================================================
2957  * I/O pipeline definition
2958  * ==========================================================================
2959  */
2960 static zio_pipe_stage_t *zio_pipeline[] = {
2961 	NULL,
2962 	zio_read_bp_init,
2963 	zio_free_bp_init,
2964 	zio_issue_async,
2965 	zio_write_bp_init,
2966 	zio_checksum_generate,
2967 	zio_ddt_read_start,
2968 	zio_ddt_read_done,
2969 	zio_ddt_write,
2970 	zio_ddt_free,
2971 	zio_gang_assemble,
2972 	zio_gang_issue,
2973 	zio_dva_allocate,
2974 	zio_dva_free,
2975 	zio_dva_claim,
2976 	zio_ready,
2977 	zio_vdev_io_start,
2978 	zio_vdev_io_done,
2979 	zio_vdev_io_assess,
2980 	zio_checksum_verify,
2981 	zio_done
2982 };
2983