xref: /illumos-gate/usr/src/uts/common/fs/zfs/zio.c (revision a98aba88)
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 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/zfs_context.h>
29 #include <sys/fm/fs/zfs.h>
30 #include <sys/spa.h>
31 #include <sys/txg.h>
32 #include <sys/spa_impl.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio_impl.h>
35 #include <sys/zio_compress.h>
36 #include <sys/zio_checksum.h>
37 
38 /*
39  * ==========================================================================
40  * I/O priority table
41  * ==========================================================================
42  */
43 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
44 	0,	/* ZIO_PRIORITY_NOW		*/
45 	0,	/* ZIO_PRIORITY_SYNC_READ	*/
46 	0,	/* ZIO_PRIORITY_SYNC_WRITE	*/
47 	6,	/* ZIO_PRIORITY_ASYNC_READ	*/
48 	4,	/* ZIO_PRIORITY_ASYNC_WRITE	*/
49 	4,	/* ZIO_PRIORITY_FREE		*/
50 	0,	/* ZIO_PRIORITY_CACHE_FILL	*/
51 	0,	/* ZIO_PRIORITY_LOG_WRITE	*/
52 	10,	/* ZIO_PRIORITY_RESILVER	*/
53 	20,	/* ZIO_PRIORITY_SCRUB		*/
54 };
55 
56 /*
57  * ==========================================================================
58  * I/O type descriptions
59  * ==========================================================================
60  */
61 char *zio_type_name[ZIO_TYPES] = {
62 	"null", "read", "write", "free", "claim", "ioctl" };
63 
64 /* At or above this size, force gang blocking - for testing */
65 uint64_t zio_gang_bang = SPA_MAXBLOCKSIZE + 1;
66 
67 typedef struct zio_sync_pass {
68 	int	zp_defer_free;		/* defer frees after this pass */
69 	int	zp_dontcompress;	/* don't compress after this pass */
70 	int	zp_rewrite;		/* rewrite new bps after this pass */
71 } zio_sync_pass_t;
72 
73 zio_sync_pass_t zio_sync_pass = {
74 	1,	/* zp_defer_free */
75 	4,	/* zp_dontcompress */
76 	1,	/* zp_rewrite */
77 };
78 
79 /*
80  * ==========================================================================
81  * I/O kmem caches
82  * ==========================================================================
83  */
84 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
85 
86 void
87 zio_init(void)
88 {
89 	size_t c;
90 
91 	/*
92 	 * For small buffers, we want a cache for each multiple of
93 	 * SPA_MINBLOCKSIZE.  For medium-size buffers, we want a cache
94 	 * for each quarter-power of 2.  For large buffers, we want
95 	 * a cache for each multiple of PAGESIZE.
96 	 */
97 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
98 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
99 		size_t p2 = size;
100 		size_t align = 0;
101 
102 		while (p2 & (p2 - 1))
103 			p2 &= p2 - 1;
104 
105 		if (size <= 4 * SPA_MINBLOCKSIZE) {
106 			align = SPA_MINBLOCKSIZE;
107 		} else if (P2PHASE(size, PAGESIZE) == 0) {
108 			align = PAGESIZE;
109 		} else if (P2PHASE(size, p2 >> 2) == 0) {
110 			align = p2 >> 2;
111 		}
112 
113 		if (align != 0) {
114 			char name[30];
115 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
116 			zio_buf_cache[c] = kmem_cache_create(name, size,
117 			    align, NULL, NULL, NULL, NULL, NULL, KMC_NODEBUG);
118 			dprintf("creating cache for size %5lx align %5lx\n",
119 			    size, align);
120 		}
121 	}
122 
123 	while (--c != 0) {
124 		ASSERT(zio_buf_cache[c] != NULL);
125 		if (zio_buf_cache[c - 1] == NULL)
126 			zio_buf_cache[c - 1] = zio_buf_cache[c];
127 	}
128 
129 	zio_inject_init();
130 }
131 
132 void
133 zio_fini(void)
134 {
135 	size_t c;
136 	kmem_cache_t *last_cache = NULL;
137 
138 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
139 		if (zio_buf_cache[c] != last_cache) {
140 			last_cache = zio_buf_cache[c];
141 			kmem_cache_destroy(zio_buf_cache[c]);
142 		}
143 		zio_buf_cache[c] = NULL;
144 	}
145 
146 	zio_inject_fini();
147 }
148 
149 /*
150  * ==========================================================================
151  * Allocate and free I/O buffers
152  * ==========================================================================
153  */
154 void *
155 zio_buf_alloc(size_t size)
156 {
157 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
158 
159 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
160 
161 	return (kmem_cache_alloc(zio_buf_cache[c], KM_SLEEP));
162 }
163 
164 void
165 zio_buf_free(void *buf, size_t size)
166 {
167 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
168 
169 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
170 
171 	kmem_cache_free(zio_buf_cache[c], buf);
172 }
173 
174 /*
175  * ==========================================================================
176  * Push and pop I/O transform buffers
177  * ==========================================================================
178  */
179 static void
180 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize)
181 {
182 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
183 
184 	zt->zt_data = data;
185 	zt->zt_size = size;
186 	zt->zt_bufsize = bufsize;
187 
188 	zt->zt_next = zio->io_transform_stack;
189 	zio->io_transform_stack = zt;
190 
191 	zio->io_data = data;
192 	zio->io_size = size;
193 }
194 
195 static void
196 zio_pop_transform(zio_t *zio, void **data, uint64_t *size, uint64_t *bufsize)
197 {
198 	zio_transform_t *zt = zio->io_transform_stack;
199 
200 	*data = zt->zt_data;
201 	*size = zt->zt_size;
202 	*bufsize = zt->zt_bufsize;
203 
204 	zio->io_transform_stack = zt->zt_next;
205 	kmem_free(zt, sizeof (zio_transform_t));
206 
207 	if ((zt = zio->io_transform_stack) != NULL) {
208 		zio->io_data = zt->zt_data;
209 		zio->io_size = zt->zt_size;
210 	}
211 }
212 
213 static void
214 zio_clear_transform_stack(zio_t *zio)
215 {
216 	void *data;
217 	uint64_t size, bufsize;
218 
219 	ASSERT(zio->io_transform_stack != NULL);
220 
221 	zio_pop_transform(zio, &data, &size, &bufsize);
222 	while (zio->io_transform_stack != NULL) {
223 		zio_buf_free(data, bufsize);
224 		zio_pop_transform(zio, &data, &size, &bufsize);
225 	}
226 }
227 
228 /*
229  * ==========================================================================
230  * Create the various types of I/O (read, write, free)
231  * ==========================================================================
232  */
233 static zio_t *
234 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
235     void *data, uint64_t size, zio_done_func_t *done, void *private,
236     zio_type_t type, int priority, int flags, uint8_t stage, uint32_t pipeline)
237 {
238 	zio_t *zio;
239 
240 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
241 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
242 
243 	zio = kmem_zalloc(sizeof (zio_t), KM_SLEEP);
244 	zio->io_parent = pio;
245 	zio->io_spa = spa;
246 	zio->io_txg = txg;
247 	if (bp != NULL) {
248 		zio->io_bp = bp;
249 		zio->io_bp_copy = *bp;
250 		zio->io_bp_orig = *bp;
251 	}
252 	zio->io_done = done;
253 	zio->io_private = private;
254 	zio->io_type = type;
255 	zio->io_priority = priority;
256 	zio->io_stage = stage;
257 	zio->io_pipeline = pipeline;
258 	zio->io_async_stages = ZIO_ASYNC_PIPELINE_STAGES;
259 	zio->io_timestamp = lbolt64;
260 	zio->io_flags = flags;
261 	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
262 	zio_push_transform(zio, data, size, size);
263 
264 	if (pio == NULL) {
265 		if (!(flags & ZIO_FLAG_CONFIG_HELD))
266 			spa_config_enter(zio->io_spa, RW_READER, zio);
267 		zio->io_root = zio;
268 	} else {
269 		zio->io_root = pio->io_root;
270 		if (!(flags & ZIO_FLAG_NOBOOKMARK))
271 			zio->io_logical = pio->io_logical;
272 		mutex_enter(&pio->io_lock);
273 		if (stage < ZIO_STAGE_READY)
274 			pio->io_children_notready++;
275 		pio->io_children_notdone++;
276 		zio->io_sibling_next = pio->io_child;
277 		zio->io_sibling_prev = NULL;
278 		if (pio->io_child != NULL)
279 			pio->io_child->io_sibling_prev = zio;
280 		pio->io_child = zio;
281 		zio->io_ndvas = pio->io_ndvas;
282 		mutex_exit(&pio->io_lock);
283 	}
284 
285 	return (zio);
286 }
287 
288 zio_t *
289 zio_null(zio_t *pio, spa_t *spa, zio_done_func_t *done, void *private,
290 	int flags)
291 {
292 	zio_t *zio;
293 
294 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
295 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, ZIO_STAGE_OPEN,
296 	    ZIO_WAIT_FOR_CHILDREN_PIPELINE);
297 
298 	return (zio);
299 }
300 
301 zio_t *
302 zio_root(spa_t *spa, zio_done_func_t *done, void *private, int flags)
303 {
304 	return (zio_null(NULL, spa, done, private, flags));
305 }
306 
307 zio_t *
308 zio_read(zio_t *pio, spa_t *spa, blkptr_t *bp, void *data,
309     uint64_t size, zio_done_func_t *done, void *private,
310     int priority, int flags, zbookmark_t *zb)
311 {
312 	zio_t *zio;
313 
314 	ASSERT3U(size, ==, BP_GET_LSIZE(bp));
315 
316 	zio = zio_create(pio, spa, bp->blk_birth, bp, data, size, done, private,
317 	    ZIO_TYPE_READ, priority, flags | ZIO_FLAG_USER,
318 	    ZIO_STAGE_OPEN, ZIO_READ_PIPELINE);
319 	zio->io_bookmark = *zb;
320 
321 	zio->io_logical = zio;
322 
323 	/*
324 	 * Work off our copy of the bp so the caller can free it.
325 	 */
326 	zio->io_bp = &zio->io_bp_copy;
327 
328 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
329 		uint64_t csize = BP_GET_PSIZE(bp);
330 		void *cbuf = zio_buf_alloc(csize);
331 
332 		zio_push_transform(zio, cbuf, csize, csize);
333 		zio->io_pipeline |= 1U << ZIO_STAGE_READ_DECOMPRESS;
334 	}
335 
336 	if (BP_IS_GANG(bp)) {
337 		uint64_t gsize = SPA_GANGBLOCKSIZE;
338 		void *gbuf = zio_buf_alloc(gsize);
339 
340 		zio_push_transform(zio, gbuf, gsize, gsize);
341 		zio->io_pipeline |= 1U << ZIO_STAGE_READ_GANG_MEMBERS;
342 	}
343 
344 	return (zio);
345 }
346 
347 zio_t *
348 zio_write(zio_t *pio, spa_t *spa, int checksum, int compress, int ncopies,
349     uint64_t txg, blkptr_t *bp, void *data, uint64_t size,
350     zio_done_func_t *done, void *private, int priority, int flags,
351     zbookmark_t *zb)
352 {
353 	zio_t *zio;
354 
355 	ASSERT(checksum >= ZIO_CHECKSUM_OFF &&
356 	    checksum < ZIO_CHECKSUM_FUNCTIONS);
357 
358 	ASSERT(compress >= ZIO_COMPRESS_OFF &&
359 	    compress < ZIO_COMPRESS_FUNCTIONS);
360 
361 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
362 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_USER,
363 	    ZIO_STAGE_OPEN, ZIO_WRITE_PIPELINE);
364 
365 	zio->io_bookmark = *zb;
366 
367 	zio->io_logical = zio;
368 
369 	zio->io_checksum = checksum;
370 	zio->io_compress = compress;
371 	zio->io_ndvas = ncopies;
372 
373 	if (compress != ZIO_COMPRESS_OFF)
374 		zio->io_async_stages |= 1U << ZIO_STAGE_WRITE_COMPRESS;
375 
376 	if (bp->blk_birth != txg) {
377 		/* XXX the bp usually (always?) gets re-zeroed later */
378 		BP_ZERO(bp);
379 		BP_SET_LSIZE(bp, size);
380 		BP_SET_PSIZE(bp, size);
381 	} else {
382 		/* Make sure someone doesn't change their mind on overwrites */
383 		ASSERT(MIN(zio->io_ndvas + BP_IS_GANG(bp),
384 		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
385 	}
386 
387 	return (zio);
388 }
389 
390 zio_t *
391 zio_rewrite(zio_t *pio, spa_t *spa, int checksum,
392     uint64_t txg, blkptr_t *bp, void *data, uint64_t size,
393     zio_done_func_t *done, void *private, int priority, int flags,
394     zbookmark_t *zb)
395 {
396 	zio_t *zio;
397 
398 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
399 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_USER,
400 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
401 
402 	zio->io_bookmark = *zb;
403 	zio->io_checksum = checksum;
404 	zio->io_compress = ZIO_COMPRESS_OFF;
405 
406 	if (pio != NULL)
407 		ASSERT3U(zio->io_ndvas, <=, BP_GET_NDVAS(bp));
408 
409 	return (zio);
410 }
411 
412 static zio_t *
413 zio_write_allocate(zio_t *pio, spa_t *spa, int checksum,
414     uint64_t txg, blkptr_t *bp, void *data, uint64_t size,
415     zio_done_func_t *done, void *private, int priority, int flags)
416 {
417 	zio_t *zio;
418 
419 	BP_ZERO(bp);
420 	BP_SET_LSIZE(bp, size);
421 	BP_SET_PSIZE(bp, size);
422 	BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF);
423 
424 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
425 	    ZIO_TYPE_WRITE, priority, flags,
426 	    ZIO_STAGE_OPEN, ZIO_WRITE_ALLOCATE_PIPELINE);
427 
428 	zio->io_checksum = checksum;
429 	zio->io_compress = ZIO_COMPRESS_OFF;
430 
431 	return (zio);
432 }
433 
434 zio_t *
435 zio_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
436     zio_done_func_t *done, void *private)
437 {
438 	zio_t *zio;
439 
440 	ASSERT(!BP_IS_HOLE(bp));
441 
442 	if (txg == spa->spa_syncing_txg &&
443 	    spa->spa_sync_pass > zio_sync_pass.zp_defer_free) {
444 		bplist_enqueue_deferred(&spa->spa_sync_bplist, bp);
445 		return (zio_null(pio, spa, NULL, NULL, 0));
446 	}
447 
448 	zio = zio_create(pio, spa, txg, bp, NULL, 0, done, private,
449 	    ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, ZIO_FLAG_USER,
450 	    ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
451 
452 	zio->io_bp = &zio->io_bp_copy;
453 
454 	return (zio);
455 }
456 
457 zio_t *
458 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
459     zio_done_func_t *done, void *private)
460 {
461 	zio_t *zio;
462 
463 	/*
464 	 * A claim is an allocation of a specific block.  Claims are needed
465 	 * to support immediate writes in the intent log.  The issue is that
466 	 * immediate writes contain committed data, but in a txg that was
467 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
468 	 * the intent log claims all blocks that contain immediate write data
469 	 * so that the SPA knows they're in use.
470 	 *
471 	 * All claims *must* be resolved in the first txg -- before the SPA
472 	 * starts allocating blocks -- so that nothing is allocated twice.
473 	 */
474 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
475 	ASSERT3U(spa_first_txg(spa), <=, txg);
476 
477 	zio = zio_create(pio, spa, txg, bp, NULL, 0, done, private,
478 	    ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, 0,
479 	    ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
480 
481 	zio->io_bp = &zio->io_bp_copy;
482 
483 	return (zio);
484 }
485 
486 zio_t *
487 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
488     zio_done_func_t *done, void *private, int priority, int flags)
489 {
490 	zio_t *zio;
491 	int c;
492 
493 	if (vd->vdev_children == 0) {
494 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
495 		    ZIO_TYPE_IOCTL, priority, flags,
496 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
497 
498 		zio->io_vd = vd;
499 		zio->io_cmd = cmd;
500 	} else {
501 		zio = zio_null(pio, spa, NULL, NULL, flags);
502 
503 		for (c = 0; c < vd->vdev_children; c++)
504 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
505 			    done, private, priority, flags));
506 	}
507 
508 	return (zio);
509 }
510 
511 static void
512 zio_phys_bp_init(vdev_t *vd, blkptr_t *bp, uint64_t offset, uint64_t size,
513     int checksum)
514 {
515 	ASSERT(vd->vdev_children == 0);
516 
517 	ASSERT(size <= SPA_MAXBLOCKSIZE);
518 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
519 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
520 
521 	ASSERT(offset + size <= VDEV_LABEL_START_SIZE ||
522 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
523 	ASSERT3U(offset + size, <=, vd->vdev_psize);
524 
525 	BP_ZERO(bp);
526 
527 	BP_SET_LSIZE(bp, size);
528 	BP_SET_PSIZE(bp, size);
529 
530 	BP_SET_CHECKSUM(bp, checksum);
531 	BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF);
532 	BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
533 
534 	if (checksum != ZIO_CHECKSUM_OFF)
535 		ZIO_SET_CHECKSUM(&bp->blk_cksum, offset, 0, 0, 0);
536 }
537 
538 zio_t *
539 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
540     void *data, int checksum, zio_done_func_t *done, void *private,
541     int priority, int flags)
542 {
543 	zio_t *zio;
544 	blkptr_t blk;
545 
546 	zio_phys_bp_init(vd, &blk, offset, size, checksum);
547 
548 	zio = zio_create(pio, vd->vdev_spa, 0, &blk, data, size, done, private,
549 	    ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL,
550 	    ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
551 
552 	zio->io_vd = vd;
553 	zio->io_offset = offset;
554 
555 	/*
556 	 * Work off our copy of the bp so the caller can free it.
557 	 */
558 	zio->io_bp = &zio->io_bp_copy;
559 
560 	return (zio);
561 }
562 
563 zio_t *
564 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
565     void *data, int checksum, zio_done_func_t *done, void *private,
566     int priority, int flags)
567 {
568 	zio_block_tail_t *zbt;
569 	void *wbuf;
570 	zio_t *zio;
571 	blkptr_t blk;
572 
573 	zio_phys_bp_init(vd, &blk, offset, size, checksum);
574 
575 	zio = zio_create(pio, vd->vdev_spa, 0, &blk, data, size, done, private,
576 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL,
577 	    ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
578 
579 	zio->io_vd = vd;
580 	zio->io_offset = offset;
581 
582 	zio->io_bp = &zio->io_bp_copy;
583 	zio->io_checksum = checksum;
584 
585 	if (zio_checksum_table[checksum].ci_zbt) {
586 		/*
587 		 * zbt checksums are necessarily destructive -- they modify
588 		 * one word of the write buffer to hold the verifier/checksum.
589 		 * Therefore, we must make a local copy in case the data is
590 		 * being written to multiple places.
591 		 */
592 		wbuf = zio_buf_alloc(size);
593 		bcopy(data, wbuf, size);
594 		zio_push_transform(zio, wbuf, size, size);
595 
596 		zbt = (zio_block_tail_t *)((char *)wbuf + size) - 1;
597 		zbt->zbt_cksum = blk.blk_cksum;
598 	}
599 
600 	return (zio);
601 }
602 
603 /*
604  * Create a child I/O to do some work for us.  It has no associated bp.
605  */
606 zio_t *
607 zio_vdev_child_io(zio_t *zio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
608 	void *data, uint64_t size, int type, int priority, int flags,
609 	zio_done_func_t *done, void *private)
610 {
611 	uint32_t pipeline = ZIO_VDEV_CHILD_PIPELINE;
612 	zio_t *cio;
613 
614 	if (type == ZIO_TYPE_READ && bp != NULL) {
615 		/*
616 		 * If we have the bp, then the child should perform the
617 		 * checksum and the parent need not.  This pushes error
618 		 * detection as close to the leaves as possible and
619 		 * eliminates redundant checksums in the interior nodes.
620 		 */
621 		pipeline |= 1U << ZIO_STAGE_CHECKSUM_VERIFY;
622 		zio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY);
623 	}
624 
625 	cio = zio_create(zio, zio->io_spa, zio->io_txg, bp, data, size,
626 	    done, private, type, priority,
627 	    (zio->io_flags & ZIO_FLAG_VDEV_INHERIT) | ZIO_FLAG_CANFAIL | flags,
628 	    ZIO_STAGE_VDEV_IO_START - 1, pipeline);
629 
630 	cio->io_vd = vd;
631 	cio->io_offset = offset;
632 
633 	return (cio);
634 }
635 
636 /*
637  * ==========================================================================
638  * Initiate I/O, either sync or async
639  * ==========================================================================
640  */
641 int
642 zio_wait(zio_t *zio)
643 {
644 	int error;
645 
646 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
647 
648 	zio->io_waiter = curthread;
649 
650 	zio_next_stage_async(zio);
651 
652 	mutex_enter(&zio->io_lock);
653 	while (zio->io_stalled != ZIO_STAGE_DONE)
654 		cv_wait(&zio->io_cv, &zio->io_lock);
655 	mutex_exit(&zio->io_lock);
656 
657 	error = zio->io_error;
658 	mutex_destroy(&zio->io_lock);
659 	kmem_free(zio, sizeof (zio_t));
660 
661 	return (error);
662 }
663 
664 void
665 zio_nowait(zio_t *zio)
666 {
667 	zio_next_stage_async(zio);
668 }
669 
670 /*
671  * ==========================================================================
672  * I/O pipeline interlocks: parent/child dependency scoreboarding
673  * ==========================================================================
674  */
675 static void
676 zio_wait_for_children(zio_t *zio, uint32_t stage, uint64_t *countp)
677 {
678 	mutex_enter(&zio->io_lock);
679 	if (*countp == 0) {
680 		ASSERT(zio->io_stalled == 0);
681 		mutex_exit(&zio->io_lock);
682 		zio_next_stage(zio);
683 	} else {
684 		zio->io_stalled = stage;
685 		mutex_exit(&zio->io_lock);
686 	}
687 }
688 
689 static void
690 zio_notify_parent(zio_t *zio, uint32_t stage, uint64_t *countp)
691 {
692 	zio_t *pio = zio->io_parent;
693 
694 	mutex_enter(&pio->io_lock);
695 	if (pio->io_error == 0 && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
696 		pio->io_error = zio->io_error;
697 	if (--*countp == 0 && pio->io_stalled == stage) {
698 		pio->io_stalled = 0;
699 		mutex_exit(&pio->io_lock);
700 		zio_next_stage_async(pio);
701 	} else {
702 		mutex_exit(&pio->io_lock);
703 	}
704 }
705 
706 static void
707 zio_wait_children_ready(zio_t *zio)
708 {
709 	zio_wait_for_children(zio, ZIO_STAGE_WAIT_CHILDREN_READY,
710 	    &zio->io_children_notready);
711 }
712 
713 void
714 zio_wait_children_done(zio_t *zio)
715 {
716 	zio_wait_for_children(zio, ZIO_STAGE_WAIT_CHILDREN_DONE,
717 	    &zio->io_children_notdone);
718 }
719 
720 static void
721 zio_ready(zio_t *zio)
722 {
723 	zio_t *pio = zio->io_parent;
724 
725 	if (pio != NULL)
726 		zio_notify_parent(zio, ZIO_STAGE_WAIT_CHILDREN_READY,
727 		    &pio->io_children_notready);
728 
729 	if (zio->io_bp)
730 		zio->io_bp_copy = *zio->io_bp;
731 
732 	zio_next_stage(zio);
733 }
734 
735 static void
736 zio_done(zio_t *zio)
737 {
738 	zio_t *pio = zio->io_parent;
739 	spa_t *spa = zio->io_spa;
740 	blkptr_t *bp = zio->io_bp;
741 	vdev_t *vd = zio->io_vd;
742 	char blkbuf[BP_SPRINTF_LEN];
743 
744 	ASSERT(zio->io_children_notready == 0);
745 	ASSERT(zio->io_children_notdone == 0);
746 
747 	if (bp != NULL) {
748 		ASSERT(bp->blk_pad[0] == 0);
749 		ASSERT(bp->blk_pad[1] == 0);
750 		ASSERT(bp->blk_pad[2] == 0);
751 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0);
752 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
753 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
754 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
755 			if (zio->io_ndvas != 0)
756 				ASSERT3U(zio->io_ndvas, <=, BP_GET_NDVAS(bp));
757 			ASSERT(BP_COUNT_GANG(bp) == 0 ||
758 			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
759 		}
760 	}
761 
762 	if (vd != NULL)
763 		vdev_stat_update(zio);
764 
765 	if (zio->io_error) {
766 		/*
767 		 * If this I/O is attached to a particular vdev,
768 		 * generate an error message describing the I/O failure
769 		 * at the block level.  We ignore these errors if the
770 		 * device is currently unavailable.
771 		 */
772 		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
773 			zfs_ereport_post(FM_EREPORT_ZFS_IO,
774 			    zio->io_spa, vd, zio, 0, 0);
775 
776 		if ((zio->io_error == EIO ||
777 		    !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) &&
778 		    zio->io_logical == zio) {
779 			/*
780 			 * For root I/O requests, tell the SPA to log the error
781 			 * appropriately.  Also, generate a logical data
782 			 * ereport.
783 			 */
784 			spa_log_error(zio->io_spa, zio);
785 
786 			zfs_ereport_post(FM_EREPORT_ZFS_DATA,
787 			    zio->io_spa, NULL, zio, 0, 0);
788 		}
789 
790 		/*
791 		 * For I/O requests that cannot fail, panic appropriately.
792 		 */
793 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL)) {
794 			sprintf_blkptr(blkbuf, BP_SPRINTF_LEN,
795 			    bp ? bp : &zio->io_bp_copy);
796 			panic("ZFS: %s (%s on %s off %llx: zio %p %s): error "
797 			    "%d", zio->io_error == ECKSUM ?
798 			    "bad checksum" : "I/O failure",
799 			    zio_type_name[zio->io_type],
800 			    vdev_description(vd),
801 			    (u_longlong_t)zio->io_offset,
802 			    zio, blkbuf, zio->io_error);
803 		}
804 	}
805 
806 	zio_clear_transform_stack(zio);
807 
808 	if (zio->io_done)
809 		zio->io_done(zio);
810 
811 	ASSERT(zio->io_delegate_list == NULL);
812 	ASSERT(zio->io_delegate_next == NULL);
813 
814 	if (pio != NULL) {
815 		zio_t *next, *prev;
816 
817 		mutex_enter(&pio->io_lock);
818 		next = zio->io_sibling_next;
819 		prev = zio->io_sibling_prev;
820 		if (next != NULL)
821 			next->io_sibling_prev = prev;
822 		if (prev != NULL)
823 			prev->io_sibling_next = next;
824 		if (pio->io_child == zio)
825 			pio->io_child = next;
826 		mutex_exit(&pio->io_lock);
827 
828 		zio_notify_parent(zio, ZIO_STAGE_WAIT_CHILDREN_DONE,
829 		    &pio->io_children_notdone);
830 	}
831 
832 	if (pio == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_HELD))
833 		spa_config_exit(spa, zio);
834 
835 	if (zio->io_waiter != NULL) {
836 		mutex_enter(&zio->io_lock);
837 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
838 		zio->io_stalled = zio->io_stage;
839 		cv_broadcast(&zio->io_cv);
840 		mutex_exit(&zio->io_lock);
841 	} else {
842 		kmem_free(zio, sizeof (zio_t));
843 	}
844 }
845 
846 /*
847  * ==========================================================================
848  * Compression support
849  * ==========================================================================
850  */
851 static void
852 zio_write_compress(zio_t *zio)
853 {
854 	int compress = zio->io_compress;
855 	blkptr_t *bp = zio->io_bp;
856 	void *cbuf;
857 	uint64_t lsize = zio->io_size;
858 	uint64_t csize = lsize;
859 	uint64_t cbufsize = 0;
860 	int pass;
861 
862 	if (bp->blk_birth == zio->io_txg) {
863 		/*
864 		 * We're rewriting an existing block, which means we're
865 		 * working on behalf of spa_sync().  For spa_sync() to
866 		 * converge, it must eventually be the case that we don't
867 		 * have to allocate new blocks.  But compression changes
868 		 * the blocksize, which forces a reallocate, and makes
869 		 * convergence take longer.  Therefore, after the first
870 		 * few passes, stop compressing to ensure convergence.
871 		 */
872 		pass = spa_sync_pass(zio->io_spa);
873 		if (pass > zio_sync_pass.zp_dontcompress)
874 			compress = ZIO_COMPRESS_OFF;
875 	} else {
876 		ASSERT(BP_IS_HOLE(bp));
877 		pass = 1;
878 	}
879 
880 	if (compress != ZIO_COMPRESS_OFF)
881 		if (!zio_compress_data(compress, zio->io_data, zio->io_size,
882 		    &cbuf, &csize, &cbufsize))
883 			compress = ZIO_COMPRESS_OFF;
884 
885 	if (compress != ZIO_COMPRESS_OFF && csize != 0)
886 		zio_push_transform(zio, cbuf, csize, cbufsize);
887 
888 	/*
889 	 * The final pass of spa_sync() must be all rewrites, but the first
890 	 * few passes offer a trade-off: allocating blocks defers convergence,
891 	 * but newly allocated blocks are sequential, so they can be written
892 	 * to disk faster.  Therefore, we allow the first few passes of
893 	 * spa_sync() to reallocate new blocks, but force rewrites after that.
894 	 * There should only be a handful of blocks after pass 1 in any case.
895 	 */
896 	if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == csize &&
897 	    pass > zio_sync_pass.zp_rewrite) {
898 		ASSERT(csize != 0);
899 		BP_SET_LSIZE(bp, lsize);
900 		BP_SET_COMPRESS(bp, compress);
901 		zio->io_pipeline = ZIO_REWRITE_PIPELINE;
902 	} else {
903 		if (bp->blk_birth == zio->io_txg) {
904 			ASSERT3U(BP_GET_LSIZE(bp), ==, lsize);
905 			bzero(bp, sizeof (blkptr_t));
906 		}
907 		if (csize == 0) {
908 			BP_ZERO(bp);
909 			zio->io_pipeline = ZIO_WAIT_FOR_CHILDREN_PIPELINE;
910 		} else {
911 			ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
912 			BP_SET_LSIZE(bp, lsize);
913 			BP_SET_PSIZE(bp, csize);
914 			BP_SET_COMPRESS(bp, compress);
915 			zio->io_pipeline = ZIO_WRITE_ALLOCATE_PIPELINE;
916 		}
917 	}
918 
919 	zio_next_stage(zio);
920 }
921 
922 static void
923 zio_read_decompress(zio_t *zio)
924 {
925 	blkptr_t *bp = zio->io_bp;
926 	void *data;
927 	uint64_t size;
928 	uint64_t bufsize;
929 	int compress = BP_GET_COMPRESS(bp);
930 
931 	ASSERT(compress != ZIO_COMPRESS_OFF);
932 
933 	zio_pop_transform(zio, &data, &size, &bufsize);
934 
935 	if (zio_decompress_data(compress, data, size,
936 	    zio->io_data, zio->io_size))
937 		zio->io_error = EIO;
938 
939 	zio_buf_free(data, bufsize);
940 
941 	zio_next_stage(zio);
942 }
943 
944 /*
945  * ==========================================================================
946  * Gang block support
947  * ==========================================================================
948  */
949 static void
950 zio_gang_pipeline(zio_t *zio)
951 {
952 	/*
953 	 * By default, the pipeline assumes that we're dealing with a gang
954 	 * block.  If we're not, strip out any gang-specific stages.
955 	 */
956 	if (!BP_IS_GANG(zio->io_bp))
957 		zio->io_pipeline &= ~ZIO_GANG_STAGES;
958 
959 	zio_next_stage(zio);
960 }
961 
962 static void
963 zio_gang_byteswap(zio_t *zio)
964 {
965 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
966 
967 	if (BP_SHOULD_BYTESWAP(zio->io_bp))
968 		byteswap_uint64_array(zio->io_data, zio->io_size);
969 }
970 
971 static void
972 zio_get_gang_header(zio_t *zio)
973 {
974 	blkptr_t *bp = zio->io_bp;
975 	uint64_t gsize = SPA_GANGBLOCKSIZE;
976 	void *gbuf = zio_buf_alloc(gsize);
977 
978 	ASSERT(BP_IS_GANG(bp));
979 
980 	zio_push_transform(zio, gbuf, gsize, gsize);
981 
982 	zio_nowait(zio_create(zio, zio->io_spa, bp->blk_birth, bp, gbuf, gsize,
983 	    NULL, NULL, ZIO_TYPE_READ, zio->io_priority,
984 	    zio->io_flags & ZIO_FLAG_GANG_INHERIT,
985 	    ZIO_STAGE_OPEN, ZIO_READ_PIPELINE));
986 
987 	zio_wait_children_done(zio);
988 }
989 
990 static void
991 zio_read_gang_members(zio_t *zio)
992 {
993 	zio_gbh_phys_t *gbh;
994 	uint64_t gsize, gbufsize, loff, lsize;
995 	int i;
996 
997 	ASSERT(BP_IS_GANG(zio->io_bp));
998 
999 	zio_gang_byteswap(zio);
1000 	zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize);
1001 
1002 	for (loff = 0, i = 0; loff != zio->io_size; loff += lsize, i++) {
1003 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1004 		lsize = BP_GET_PSIZE(gbp);
1005 
1006 		ASSERT(BP_GET_COMPRESS(gbp) == ZIO_COMPRESS_OFF);
1007 		ASSERT3U(lsize, ==, BP_GET_LSIZE(gbp));
1008 		ASSERT3U(loff + lsize, <=, zio->io_size);
1009 		ASSERT(i < SPA_GBH_NBLKPTRS);
1010 		ASSERT(!BP_IS_HOLE(gbp));
1011 
1012 		zio_nowait(zio_read(zio, zio->io_spa, gbp,
1013 		    (char *)zio->io_data + loff, lsize, NULL, NULL,
1014 		    zio->io_priority, zio->io_flags & ZIO_FLAG_GANG_INHERIT,
1015 		    &zio->io_bookmark));
1016 	}
1017 
1018 	zio_buf_free(gbh, gbufsize);
1019 	zio_wait_children_done(zio);
1020 }
1021 
1022 static void
1023 zio_rewrite_gang_members(zio_t *zio)
1024 {
1025 	zio_gbh_phys_t *gbh;
1026 	uint64_t gsize, gbufsize, loff, lsize;
1027 	int i;
1028 
1029 	ASSERT(BP_IS_GANG(zio->io_bp));
1030 	ASSERT3U(zio->io_size, ==, SPA_GANGBLOCKSIZE);
1031 
1032 	zio_gang_byteswap(zio);
1033 	zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize);
1034 
1035 	ASSERT(gsize == gbufsize);
1036 
1037 	for (loff = 0, i = 0; loff != zio->io_size; loff += lsize, i++) {
1038 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1039 		lsize = BP_GET_PSIZE(gbp);
1040 
1041 		ASSERT(BP_GET_COMPRESS(gbp) == ZIO_COMPRESS_OFF);
1042 		ASSERT3U(lsize, ==, BP_GET_LSIZE(gbp));
1043 		ASSERT3U(loff + lsize, <=, zio->io_size);
1044 		ASSERT(i < SPA_GBH_NBLKPTRS);
1045 		ASSERT(!BP_IS_HOLE(gbp));
1046 
1047 		zio_nowait(zio_rewrite(zio, zio->io_spa, zio->io_checksum,
1048 		    zio->io_txg, gbp, (char *)zio->io_data + loff, lsize,
1049 		    NULL, NULL, zio->io_priority, zio->io_flags,
1050 		    &zio->io_bookmark));
1051 	}
1052 
1053 	zio_push_transform(zio, gbh, gsize, gbufsize);
1054 	zio_wait_children_ready(zio);
1055 }
1056 
1057 static void
1058 zio_free_gang_members(zio_t *zio)
1059 {
1060 	zio_gbh_phys_t *gbh;
1061 	uint64_t gsize, gbufsize;
1062 	int i;
1063 
1064 	ASSERT(BP_IS_GANG(zio->io_bp));
1065 
1066 	zio_gang_byteswap(zio);
1067 	zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize);
1068 
1069 	for (i = 0; i < SPA_GBH_NBLKPTRS; i++) {
1070 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1071 
1072 		if (BP_IS_HOLE(gbp))
1073 			continue;
1074 		zio_nowait(zio_free(zio, zio->io_spa, zio->io_txg,
1075 		    gbp, NULL, NULL));
1076 	}
1077 
1078 	zio_buf_free(gbh, gbufsize);
1079 	zio_next_stage(zio);
1080 }
1081 
1082 static void
1083 zio_claim_gang_members(zio_t *zio)
1084 {
1085 	zio_gbh_phys_t *gbh;
1086 	uint64_t gsize, gbufsize;
1087 	int i;
1088 
1089 	ASSERT(BP_IS_GANG(zio->io_bp));
1090 
1091 	zio_gang_byteswap(zio);
1092 	zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize);
1093 
1094 	for (i = 0; i < SPA_GBH_NBLKPTRS; i++) {
1095 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1096 		if (BP_IS_HOLE(gbp))
1097 			continue;
1098 		zio_nowait(zio_claim(zio, zio->io_spa, zio->io_txg,
1099 		    gbp, NULL, NULL));
1100 	}
1101 
1102 	zio_buf_free(gbh, gbufsize);
1103 	zio_next_stage(zio);
1104 }
1105 
1106 static void
1107 zio_write_allocate_gang_member_done(zio_t *zio)
1108 {
1109 	zio_t *pio = zio->io_parent;
1110 	dva_t *cdva = zio->io_bp->blk_dva;
1111 	dva_t *pdva = pio->io_bp->blk_dva;
1112 	uint64_t asize;
1113 	int d;
1114 
1115 	ASSERT3U(pio->io_ndvas, ==, zio->io_ndvas);
1116 	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1117 	ASSERT3U(zio->io_ndvas, <=, BP_GET_NDVAS(zio->io_bp));
1118 	ASSERT3U(pio->io_ndvas, <=, BP_GET_NDVAS(pio->io_bp));
1119 
1120 	mutex_enter(&pio->io_lock);
1121 	for (d = 0; d < BP_GET_NDVAS(pio->io_bp); d++) {
1122 		ASSERT(DVA_GET_GANG(&pdva[d]));
1123 		asize = DVA_GET_ASIZE(&pdva[d]);
1124 		asize += DVA_GET_ASIZE(&cdva[d]);
1125 		DVA_SET_ASIZE(&pdva[d], asize);
1126 	}
1127 	mutex_exit(&pio->io_lock);
1128 }
1129 
1130 static void
1131 zio_write_allocate_gang_members(zio_t *zio)
1132 {
1133 	blkptr_t *bp = zio->io_bp;
1134 	dva_t *dva = bp->blk_dva;
1135 	spa_t *spa = zio->io_spa;
1136 	zio_gbh_phys_t *gbh;
1137 	uint64_t txg = zio->io_txg;
1138 	uint64_t resid = zio->io_size;
1139 	uint64_t maxalloc = P2ROUNDUP(zio->io_size >> 1, SPA_MINBLOCKSIZE);
1140 	uint64_t gsize, loff, lsize;
1141 	uint32_t gbps_left;
1142 	int ndvas = zio->io_ndvas;
1143 	int gbh_ndvas = MIN(ndvas + 1, spa_max_replication(spa));
1144 	int error;
1145 	int i, d;
1146 
1147 	gsize = SPA_GANGBLOCKSIZE;
1148 	gbps_left = SPA_GBH_NBLKPTRS;
1149 
1150 	error = metaslab_alloc(spa, gsize, bp, gbh_ndvas, txg, NULL);
1151 	if (error == ENOSPC)
1152 		panic("can't allocate gang block header");
1153 	ASSERT(error == 0);
1154 
1155 	for (d = 0; d < gbh_ndvas; d++)
1156 		DVA_SET_GANG(&dva[d], 1);
1157 
1158 	bp->blk_birth = txg;
1159 
1160 	gbh = zio_buf_alloc(gsize);
1161 	bzero(gbh, gsize);
1162 
1163 	/* We need to test multi-level gang blocks */
1164 	if (maxalloc >= zio_gang_bang && (lbolt & 0x1) == 0)
1165 		maxalloc = MAX(maxalloc >> 2, SPA_MINBLOCKSIZE);
1166 
1167 	for (loff = 0, i = 0; loff != zio->io_size;
1168 	    loff += lsize, resid -= lsize, gbps_left--, i++) {
1169 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1170 		dva = gbp->blk_dva;
1171 
1172 		ASSERT(gbps_left != 0);
1173 		maxalloc = MIN(maxalloc, resid);
1174 
1175 		while (resid <= maxalloc * gbps_left) {
1176 			error = metaslab_alloc(spa, maxalloc, gbp, ndvas,
1177 			    txg, bp);
1178 			if (error == 0)
1179 				break;
1180 			ASSERT3U(error, ==, ENOSPC);
1181 			if (maxalloc == SPA_MINBLOCKSIZE)
1182 				panic("really out of space");
1183 			maxalloc = P2ROUNDUP(maxalloc >> 1, SPA_MINBLOCKSIZE);
1184 		}
1185 
1186 		if (resid <= maxalloc * gbps_left) {
1187 			lsize = maxalloc;
1188 			BP_SET_LSIZE(gbp, lsize);
1189 			BP_SET_PSIZE(gbp, lsize);
1190 			BP_SET_COMPRESS(gbp, ZIO_COMPRESS_OFF);
1191 			gbp->blk_birth = txg;
1192 			zio_nowait(zio_rewrite(zio, spa,
1193 			    zio->io_checksum, txg, gbp,
1194 			    (char *)zio->io_data + loff, lsize,
1195 			    zio_write_allocate_gang_member_done, NULL,
1196 			    zio->io_priority, zio->io_flags,
1197 			    &zio->io_bookmark));
1198 		} else {
1199 			lsize = P2ROUNDUP(resid / gbps_left, SPA_MINBLOCKSIZE);
1200 			ASSERT(lsize != SPA_MINBLOCKSIZE);
1201 			zio_nowait(zio_write_allocate(zio, spa,
1202 			    zio->io_checksum, txg, gbp,
1203 			    (char *)zio->io_data + loff, lsize,
1204 			    zio_write_allocate_gang_member_done, NULL,
1205 			    zio->io_priority, zio->io_flags));
1206 		}
1207 	}
1208 
1209 	ASSERT(resid == 0 && loff == zio->io_size);
1210 
1211 	zio->io_pipeline |= 1U << ZIO_STAGE_GANG_CHECKSUM_GENERATE;
1212 
1213 	zio_push_transform(zio, gbh, gsize, gsize);
1214 	/*
1215 	 * As much as we'd like this to be zio_wait_children_ready(),
1216 	 * updating our ASIZE doesn't happen until the io_done callback,
1217 	 * so we have to wait for that to finish in order for our BP
1218 	 * to be stable.
1219 	 */
1220 	zio_wait_children_done(zio);
1221 }
1222 
1223 /*
1224  * ==========================================================================
1225  * Allocate and free blocks
1226  * ==========================================================================
1227  */
1228 static void
1229 zio_dva_allocate(zio_t *zio)
1230 {
1231 	blkptr_t *bp = zio->io_bp;
1232 	int error;
1233 
1234 	ASSERT(BP_IS_HOLE(bp));
1235 	ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
1236 	ASSERT3U(zio->io_ndvas, >, 0);
1237 	ASSERT3U(zio->io_ndvas, <=, spa_max_replication(zio->io_spa));
1238 
1239 	/* For testing, make some blocks above a certain size be gang blocks */
1240 	if (zio->io_size >= zio_gang_bang && (lbolt & 0x3) == 0) {
1241 		zio_write_allocate_gang_members(zio);
1242 		return;
1243 	}
1244 
1245 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
1246 
1247 	error = metaslab_alloc(zio->io_spa, zio->io_size, bp, zio->io_ndvas,
1248 	    zio->io_txg, NULL);
1249 
1250 	if (error == 0) {
1251 		bp->blk_birth = zio->io_txg;
1252 	} else if (error == ENOSPC) {
1253 		if (zio->io_size == SPA_MINBLOCKSIZE)
1254 			panic("really, truly out of space");
1255 		zio_write_allocate_gang_members(zio);
1256 		return;
1257 	} else {
1258 		zio->io_error = error;
1259 	}
1260 	zio_next_stage(zio);
1261 }
1262 
1263 static void
1264 zio_dva_free(zio_t *zio)
1265 {
1266 	blkptr_t *bp = zio->io_bp;
1267 
1268 	metaslab_free(zio->io_spa, bp, zio->io_txg, B_FALSE);
1269 
1270 	BP_ZERO(bp);
1271 
1272 	zio_next_stage(zio);
1273 }
1274 
1275 static void
1276 zio_dva_claim(zio_t *zio)
1277 {
1278 	zio->io_error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
1279 
1280 	zio_next_stage(zio);
1281 }
1282 
1283 /*
1284  * ==========================================================================
1285  * Read and write to physical devices
1286  * ==========================================================================
1287  */
1288 
1289 static void
1290 zio_vdev_io_start(zio_t *zio)
1291 {
1292 	vdev_t *vd = zio->io_vd;
1293 	vdev_t *tvd = vd ? vd->vdev_top : NULL;
1294 	blkptr_t *bp = zio->io_bp;
1295 	uint64_t align;
1296 
1297 	if (vd == NULL) {
1298 		/* The mirror_ops handle multiple DVAs in a single BP */
1299 		vdev_mirror_ops.vdev_op_io_start(zio);
1300 		return;
1301 	}
1302 
1303 	align = 1ULL << tvd->vdev_ashift;
1304 
1305 	if (zio->io_retries == 0 && vd == tvd)
1306 		zio->io_flags |= ZIO_FLAG_FAILFAST;
1307 
1308 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
1309 	    vd->vdev_children == 0) {
1310 		zio->io_flags |= ZIO_FLAG_PHYSICAL;
1311 		zio->io_offset += VDEV_LABEL_START_SIZE;
1312 	}
1313 
1314 	if (P2PHASE(zio->io_size, align) != 0) {
1315 		uint64_t asize = P2ROUNDUP(zio->io_size, align);
1316 		char *abuf = zio_buf_alloc(asize);
1317 		ASSERT(vd == tvd);
1318 		if (zio->io_type == ZIO_TYPE_WRITE) {
1319 			bcopy(zio->io_data, abuf, zio->io_size);
1320 			bzero(abuf + zio->io_size, asize - zio->io_size);
1321 		}
1322 		zio_push_transform(zio, abuf, asize, asize);
1323 		ASSERT(!(zio->io_flags & ZIO_FLAG_SUBBLOCK));
1324 		zio->io_flags |= ZIO_FLAG_SUBBLOCK;
1325 	}
1326 
1327 	ASSERT(P2PHASE(zio->io_offset, align) == 0);
1328 	ASSERT(P2PHASE(zio->io_size, align) == 0);
1329 	ASSERT(bp == NULL ||
1330 	    P2ROUNDUP(ZIO_GET_IOSIZE(zio), align) == zio->io_size);
1331 	ASSERT(zio->io_type != ZIO_TYPE_WRITE || (spa_mode & FWRITE));
1332 
1333 	vdev_io_start(zio);
1334 
1335 	/* zio_next_stage_async() gets called from io completion interrupt */
1336 }
1337 
1338 static void
1339 zio_vdev_io_done(zio_t *zio)
1340 {
1341 	if (zio->io_vd == NULL)
1342 		/* The mirror_ops handle multiple DVAs in a single BP */
1343 		vdev_mirror_ops.vdev_op_io_done(zio);
1344 	else
1345 		vdev_io_done(zio);
1346 }
1347 
1348 /* XXPOLICY */
1349 boolean_t
1350 zio_should_retry(zio_t *zio)
1351 {
1352 	vdev_t *vd = zio->io_vd;
1353 
1354 	if (zio->io_error == 0)
1355 		return (B_FALSE);
1356 	if (zio->io_delegate_list != NULL)
1357 		return (B_FALSE);
1358 	if (vd && vd != vd->vdev_top)
1359 		return (B_FALSE);
1360 	if (zio->io_flags & ZIO_FLAG_DONT_RETRY)
1361 		return (B_FALSE);
1362 	if (zio->io_retries > 0)
1363 		return (B_FALSE);
1364 
1365 	return (B_TRUE);
1366 }
1367 
1368 static void
1369 zio_vdev_io_assess(zio_t *zio)
1370 {
1371 	vdev_t *vd = zio->io_vd;
1372 	vdev_t *tvd = vd ? vd->vdev_top : NULL;
1373 
1374 	ASSERT(zio->io_vsd == NULL);
1375 
1376 	if (zio->io_flags & ZIO_FLAG_SUBBLOCK) {
1377 		void *abuf;
1378 		uint64_t asize;
1379 		ASSERT(vd == tvd);
1380 		zio_pop_transform(zio, &abuf, &asize, &asize);
1381 		if (zio->io_type == ZIO_TYPE_READ)
1382 			bcopy(abuf, zio->io_data, zio->io_size);
1383 		zio_buf_free(abuf, asize);
1384 		zio->io_flags &= ~ZIO_FLAG_SUBBLOCK;
1385 	}
1386 
1387 	if (zio_injection_enabled && !zio->io_error)
1388 		zio->io_error = zio_handle_fault_injection(zio, EIO);
1389 
1390 	/*
1391 	 * If the I/O failed, determine whether we should attempt to retry it.
1392 	 */
1393 	/* XXPOLICY */
1394 	if (zio_should_retry(zio)) {
1395 		ASSERT(tvd == vd);
1396 
1397 		zio->io_retries++;
1398 		zio->io_error = 0;
1399 		zio->io_flags &= ZIO_FLAG_VDEV_INHERIT;
1400 		/* XXPOLICY */
1401 		zio->io_flags &= ~ZIO_FLAG_FAILFAST;
1402 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1403 		zio->io_stage = ZIO_STAGE_VDEV_IO_START - 1;
1404 
1405 		dprintf("retry #%d for %s to %s offset %llx\n",
1406 		    zio->io_retries, zio_type_name[zio->io_type],
1407 		    vdev_description(vd), zio->io_offset);
1408 
1409 		zio_next_stage_async(zio);
1410 		return;
1411 	}
1412 
1413 	if (zio->io_error != 0 && zio->io_error != ECKSUM &&
1414 	    !(zio->io_flags & ZIO_FLAG_SPECULATIVE) && vd) {
1415 		/*
1416 		 * Poor man's hotplug support.  Even if we're done retrying this
1417 		 * I/O, try to reopen the vdev to see if it's still attached.
1418 		 * To avoid excessive thrashing, we only try it once a minute.
1419 		 * This also has the effect of detecting when missing devices
1420 		 * have come back, by polling the device once a minute.
1421 		 *
1422 		 * We need to do this asynchronously because we can't grab
1423 		 * all the necessary locks way down here.
1424 		 */
1425 		if (gethrtime() - vd->vdev_last_try > 60ULL * NANOSEC) {
1426 			vd->vdev_last_try = gethrtime();
1427 			tvd->vdev_reopen_wanted = 1;
1428 			spa_async_request(vd->vdev_spa, SPA_ASYNC_REOPEN);
1429 		}
1430 	}
1431 
1432 	zio_next_stage(zio);
1433 }
1434 
1435 void
1436 zio_vdev_io_reissue(zio_t *zio)
1437 {
1438 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
1439 	ASSERT(zio->io_error == 0);
1440 
1441 	zio->io_stage--;
1442 }
1443 
1444 void
1445 zio_vdev_io_redone(zio_t *zio)
1446 {
1447 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
1448 
1449 	zio->io_stage--;
1450 }
1451 
1452 void
1453 zio_vdev_io_bypass(zio_t *zio)
1454 {
1455 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
1456 	ASSERT(zio->io_error == 0);
1457 
1458 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
1459 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS - 1;
1460 }
1461 
1462 /*
1463  * ==========================================================================
1464  * Generate and verify checksums
1465  * ==========================================================================
1466  */
1467 static void
1468 zio_checksum_generate(zio_t *zio)
1469 {
1470 	int checksum = zio->io_checksum;
1471 	blkptr_t *bp = zio->io_bp;
1472 
1473 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
1474 
1475 	BP_SET_CHECKSUM(bp, checksum);
1476 	BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1477 
1478 	zio_checksum(checksum, &bp->blk_cksum, zio->io_data, zio->io_size);
1479 
1480 	zio_next_stage(zio);
1481 }
1482 
1483 static void
1484 zio_gang_checksum_generate(zio_t *zio)
1485 {
1486 	zio_cksum_t zc;
1487 	zio_gbh_phys_t *gbh = zio->io_data;
1488 
1489 	ASSERT(BP_IS_GANG(zio->io_bp));
1490 	ASSERT3U(zio->io_size, ==, SPA_GANGBLOCKSIZE);
1491 
1492 	zio_set_gang_verifier(zio, &gbh->zg_tail.zbt_cksum);
1493 
1494 	zio_checksum(ZIO_CHECKSUM_GANG_HEADER, &zc, zio->io_data, zio->io_size);
1495 
1496 	zio_next_stage(zio);
1497 }
1498 
1499 static void
1500 zio_checksum_verify(zio_t *zio)
1501 {
1502 	if (zio->io_bp != NULL) {
1503 		zio->io_error = zio_checksum_error(zio);
1504 		if (zio->io_error && !(zio->io_flags & ZIO_FLAG_SPECULATIVE))
1505 			zfs_ereport_post(FM_EREPORT_ZFS_CHECKSUM,
1506 			    zio->io_spa, zio->io_vd, zio, 0, 0);
1507 	}
1508 
1509 	zio_next_stage(zio);
1510 }
1511 
1512 /*
1513  * Called by RAID-Z to ensure we don't compute the checksum twice.
1514  */
1515 void
1516 zio_checksum_verified(zio_t *zio)
1517 {
1518 	zio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY);
1519 }
1520 
1521 /*
1522  * Set the external verifier for a gang block based on stuff in the bp
1523  */
1524 void
1525 zio_set_gang_verifier(zio_t *zio, zio_cksum_t *zcp)
1526 {
1527 	blkptr_t *bp = zio->io_bp;
1528 
1529 	zcp->zc_word[0] = DVA_GET_VDEV(BP_IDENTITY(bp));
1530 	zcp->zc_word[1] = DVA_GET_OFFSET(BP_IDENTITY(bp));
1531 	zcp->zc_word[2] = bp->blk_birth;
1532 	zcp->zc_word[3] = 0;
1533 }
1534 
1535 /*
1536  * ==========================================================================
1537  * Define the pipeline
1538  * ==========================================================================
1539  */
1540 typedef void zio_pipe_stage_t(zio_t *zio);
1541 
1542 static void
1543 zio_badop(zio_t *zio)
1544 {
1545 	panic("Invalid I/O pipeline stage %u for zio %p", zio->io_stage, zio);
1546 }
1547 
1548 zio_pipe_stage_t *zio_pipeline[ZIO_STAGE_DONE + 2] = {
1549 	zio_badop,
1550 	zio_wait_children_ready,
1551 	zio_write_compress,
1552 	zio_checksum_generate,
1553 	zio_gang_pipeline,
1554 	zio_get_gang_header,
1555 	zio_rewrite_gang_members,
1556 	zio_free_gang_members,
1557 	zio_claim_gang_members,
1558 	zio_dva_allocate,
1559 	zio_dva_free,
1560 	zio_dva_claim,
1561 	zio_gang_checksum_generate,
1562 	zio_ready,
1563 	zio_vdev_io_start,
1564 	zio_vdev_io_done,
1565 	zio_vdev_io_assess,
1566 	zio_wait_children_done,
1567 	zio_checksum_verify,
1568 	zio_read_gang_members,
1569 	zio_read_decompress,
1570 	zio_done,
1571 	zio_badop
1572 };
1573 
1574 /*
1575  * Move an I/O to the next stage of the pipeline and execute that stage.
1576  * There's no locking on io_stage because there's no legitimate way for
1577  * multiple threads to be attempting to process the same I/O.
1578  */
1579 void
1580 zio_next_stage(zio_t *zio)
1581 {
1582 	uint32_t pipeline = zio->io_pipeline;
1583 
1584 	ASSERT(!MUTEX_HELD(&zio->io_lock));
1585 
1586 	if (zio->io_error) {
1587 		dprintf("zio %p vdev %s offset %llx stage %d error %d\n",
1588 		    zio, vdev_description(zio->io_vd),
1589 		    zio->io_offset, zio->io_stage, zio->io_error);
1590 		if (((1U << zio->io_stage) & ZIO_VDEV_IO_PIPELINE) == 0)
1591 			pipeline &= ZIO_ERROR_PIPELINE_MASK;
1592 	}
1593 
1594 	while (((1U << ++zio->io_stage) & pipeline) == 0)
1595 		continue;
1596 
1597 	ASSERT(zio->io_stage <= ZIO_STAGE_DONE);
1598 	ASSERT(zio->io_stalled == 0);
1599 
1600 	zio_pipeline[zio->io_stage](zio);
1601 }
1602 
1603 void
1604 zio_next_stage_async(zio_t *zio)
1605 {
1606 	taskq_t *tq;
1607 	uint32_t pipeline = zio->io_pipeline;
1608 
1609 	ASSERT(!MUTEX_HELD(&zio->io_lock));
1610 
1611 	if (zio->io_error) {
1612 		dprintf("zio %p vdev %s offset %llx stage %d error %d\n",
1613 		    zio, vdev_description(zio->io_vd),
1614 		    zio->io_offset, zio->io_stage, zio->io_error);
1615 		if (((1U << zio->io_stage) & ZIO_VDEV_IO_PIPELINE) == 0)
1616 			pipeline &= ZIO_ERROR_PIPELINE_MASK;
1617 	}
1618 
1619 	while (((1U << ++zio->io_stage) & pipeline) == 0)
1620 		continue;
1621 
1622 	ASSERT(zio->io_stage <= ZIO_STAGE_DONE);
1623 	ASSERT(zio->io_stalled == 0);
1624 
1625 	/*
1626 	 * For performance, we'll probably want two sets of task queues:
1627 	 * per-CPU issue taskqs and per-CPU completion taskqs.  The per-CPU
1628 	 * part is for read performance: since we have to make a pass over
1629 	 * the data to checksum it anyway, we want to do this on the same CPU
1630 	 * that issued the read, because (assuming CPU scheduling affinity)
1631 	 * that thread is probably still there.  Getting this optimization
1632 	 * right avoids performance-hostile cache-to-cache transfers.
1633 	 *
1634 	 * Note that having two sets of task queues is also necessary for
1635 	 * correctness: if all of the issue threads get bogged down waiting
1636 	 * for dependent reads (e.g. metaslab freelist) to complete, then
1637 	 * there won't be any threads available to service I/O completion
1638 	 * interrupts.
1639 	 */
1640 	if ((1U << zio->io_stage) & zio->io_async_stages) {
1641 		if (zio->io_stage < ZIO_STAGE_VDEV_IO_DONE)
1642 			tq = zio->io_spa->spa_zio_issue_taskq[zio->io_type];
1643 		else
1644 			tq = zio->io_spa->spa_zio_intr_taskq[zio->io_type];
1645 		(void) taskq_dispatch(tq,
1646 		    (task_func_t *)zio_pipeline[zio->io_stage], zio, TQ_SLEEP);
1647 	} else {
1648 		zio_pipeline[zio->io_stage](zio);
1649 	}
1650 }
1651 
1652 /*
1653  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
1654  */
1655 int
1656 zio_alloc_blk(spa_t *spa, uint64_t size, blkptr_t *bp, uint64_t txg)
1657 {
1658 	int error;
1659 
1660 	spa_config_enter(spa, RW_READER, FTAG);
1661 
1662 	BP_ZERO(bp);
1663 
1664 	error = metaslab_alloc(spa, size, bp, 1, txg, NULL);
1665 
1666 	if (error == 0) {
1667 		BP_SET_LSIZE(bp, size);
1668 		BP_SET_PSIZE(bp, size);
1669 		BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF);
1670 		BP_SET_CHECKSUM(bp, ZIO_CHECKSUM_ZILOG);
1671 		BP_SET_TYPE(bp, DMU_OT_INTENT_LOG);
1672 		BP_SET_LEVEL(bp, 0);
1673 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1674 		bp->blk_birth = txg;
1675 	}
1676 
1677 	spa_config_exit(spa, FTAG);
1678 
1679 	return (error);
1680 }
1681 
1682 /*
1683  * Free an intent log block.  We know it can't be a gang block, so there's
1684  * nothing to do except metaslab_free() it.
1685  */
1686 void
1687 zio_free_blk(spa_t *spa, blkptr_t *bp, uint64_t txg)
1688 {
1689 	ASSERT(!BP_IS_GANG(bp));
1690 
1691 	spa_config_enter(spa, RW_READER, FTAG);
1692 
1693 	metaslab_free(spa, bp, txg, B_FALSE);
1694 
1695 	spa_config_exit(spa, FTAG);
1696 }
1697