xref: /titanic_41/usr/src/uts/common/fs/zfs/zio.c (revision 749f21d359d8fbd020c974a1a5227316221bfc9c)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <sys/zfs_context.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 static void zio_vdev_io_enter(zio_t *zio);
39 static void zio_vdev_io_exit(zio_t *zio);
40 
41 /*
42  * ==========================================================================
43  * I/O priority table
44  * ==========================================================================
45  */
46 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
47 	0,	/* ZIO_PRIORITY_NOW		*/
48 	0,	/* ZIO_PRIORITY_SYNC_READ	*/
49 	0,	/* ZIO_PRIORITY_SYNC_WRITE	*/
50 	6,	/* ZIO_PRIORITY_ASYNC_READ	*/
51 	4,	/* ZIO_PRIORITY_ASYNC_WRITE	*/
52 	4,	/* ZIO_PRIORITY_FREE		*/
53 	0,	/* ZIO_PRIORITY_CACHE_FILL	*/
54 	0,	/* ZIO_PRIORITY_LOG_WRITE	*/
55 	10,	/* ZIO_PRIORITY_RESILVER	*/
56 	20,	/* ZIO_PRIORITY_SCRUB		*/
57 };
58 
59 /*
60  * ==========================================================================
61  * I/O type descriptions
62  * ==========================================================================
63  */
64 char *zio_type_name[ZIO_TYPES] = {
65 	"null", "read", "write", "free", "claim", "ioctl" };
66 
67 /* At or above this size, force gang blocking - for testing */
68 uint64_t zio_gang_bang = SPA_MAXBLOCKSIZE + 1;
69 
70 typedef struct zio_sync_pass {
71 	int	zp_defer_free;		/* defer frees after this pass */
72 	int	zp_dontcompress;	/* don't compress after this pass */
73 	int	zp_rewrite;		/* rewrite new bps after this pass */
74 } zio_sync_pass_t;
75 
76 zio_sync_pass_t zio_sync_pass = {
77 	1,	/* zp_defer_free */
78 	4,	/* zp_dontcompress */
79 	1,	/* zp_rewrite */
80 };
81 
82 /*
83  * ==========================================================================
84  * I/O kmem caches
85  * ==========================================================================
86  */
87 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
88 
89 void
90 zio_init(void)
91 {
92 	size_t c;
93 
94 	/*
95 	 * For small buffers, we want a cache for each multiple of
96 	 * SPA_MINBLOCKSIZE.  For medium-size buffers, we want a cache
97 	 * for each quarter-power of 2.  For large buffers, we want
98 	 * a cache for each multiple of PAGESIZE.
99 	 */
100 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
101 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
102 		size_t p2 = size;
103 		size_t align = 0;
104 
105 		while (p2 & (p2 - 1))
106 			p2 &= p2 - 1;
107 
108 		if (size <= 4 * SPA_MINBLOCKSIZE) {
109 			align = SPA_MINBLOCKSIZE;
110 		} else if (P2PHASE(size, PAGESIZE) == 0) {
111 			align = PAGESIZE;
112 		} else if (P2PHASE(size, p2 >> 2) == 0) {
113 			align = p2 >> 2;
114 		}
115 
116 		if (align != 0) {
117 			char name[30];
118 			(void) sprintf(name, "zio_buf_%lu", size);
119 			zio_buf_cache[c] = kmem_cache_create(name, size,
120 			    align, NULL, NULL, NULL, NULL, NULL, KMC_NODEBUG);
121 			dprintf("creating cache for size %5lx align %5lx\n",
122 			    size, align);
123 		}
124 	}
125 
126 	while (--c != 0) {
127 		ASSERT(zio_buf_cache[c] != NULL);
128 		if (zio_buf_cache[c - 1] == NULL)
129 			zio_buf_cache[c - 1] = zio_buf_cache[c];
130 	}
131 }
132 
133 void
134 zio_fini(void)
135 {
136 	size_t c;
137 	kmem_cache_t *last_cache = NULL;
138 
139 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
140 		if (zio_buf_cache[c] != last_cache) {
141 			last_cache = zio_buf_cache[c];
142 			kmem_cache_destroy(zio_buf_cache[c]);
143 		}
144 		zio_buf_cache[c] = NULL;
145 	}
146 }
147 
148 /*
149  * ==========================================================================
150  * Allocate and free I/O buffers
151  * ==========================================================================
152  */
153 void *
154 zio_buf_alloc(size_t size)
155 {
156 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
157 
158 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
159 
160 	return (kmem_cache_alloc(zio_buf_cache[c], KM_SLEEP));
161 }
162 
163 void
164 zio_buf_free(void *buf, size_t size)
165 {
166 	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
167 
168 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
169 
170 	kmem_cache_free(zio_buf_cache[c], buf);
171 }
172 
173 /*
174  * ==========================================================================
175  * Push and pop I/O transform buffers
176  * ==========================================================================
177  */
178 static void
179 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize)
180 {
181 	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
182 
183 	zt->zt_data = data;
184 	zt->zt_size = size;
185 	zt->zt_bufsize = bufsize;
186 
187 	zt->zt_next = zio->io_transform_stack;
188 	zio->io_transform_stack = zt;
189 
190 	zio->io_data = data;
191 	zio->io_size = size;
192 }
193 
194 static void
195 zio_pop_transform(zio_t *zio, void **data, uint64_t *size, uint64_t *bufsize)
196 {
197 	zio_transform_t *zt = zio->io_transform_stack;
198 
199 	*data = zt->zt_data;
200 	*size = zt->zt_size;
201 	*bufsize = zt->zt_bufsize;
202 
203 	zio->io_transform_stack = zt->zt_next;
204 	kmem_free(zt, sizeof (zio_transform_t));
205 
206 	if ((zt = zio->io_transform_stack) != NULL) {
207 		zio->io_data = zt->zt_data;
208 		zio->io_size = zt->zt_size;
209 	}
210 }
211 
212 static void
213 zio_clear_transform_stack(zio_t *zio)
214 {
215 	void *data;
216 	uint64_t size, bufsize;
217 
218 	ASSERT(zio->io_transform_stack != NULL);
219 
220 	zio_pop_transform(zio, &data, &size, &bufsize);
221 	while (zio->io_transform_stack != NULL) {
222 		zio_buf_free(data, bufsize);
223 		zio_pop_transform(zio, &data, &size, &bufsize);
224 	}
225 }
226 
227 /*
228  * ==========================================================================
229  * Create the various types of I/O (read, write, free)
230  * ==========================================================================
231  */
232 static zio_t *
233 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
234     void *data, uint64_t size, zio_done_func_t *done, void *private,
235     zio_type_t type, int priority, int flags, uint8_t stage, uint32_t pipeline)
236 {
237 	zio_t *zio;
238 
239 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
240 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
241 
242 	zio = kmem_zalloc(sizeof (zio_t), KM_SLEEP);
243 	zio->io_parent = pio;
244 	zio->io_spa = spa;
245 	zio->io_txg = txg;
246 	if (bp != NULL) {
247 		zio->io_bp = bp;
248 		zio->io_bp_copy = *bp;
249 		zio->io_bp_orig = *bp;
250 		/* XXBP - Need to inherit this when it matters */
251 		zio->io_dva_index = 0;
252 	}
253 	zio->io_done = done;
254 	zio->io_private = private;
255 	zio->io_type = type;
256 	zio->io_priority = priority;
257 	zio->io_stage = stage;
258 	zio->io_pipeline = pipeline;
259 	zio->io_async_stages = ZIO_ASYNC_PIPELINE_STAGES;
260 	zio->io_timestamp = lbolt64;
261 	zio->io_flags = flags;
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);
267 		zio->io_root = zio;
268 	} else {
269 		zio->io_root = pio->io_root;
270 
271 		mutex_enter(&pio->io_lock);
272 		if (stage < ZIO_STAGE_READY)
273 			pio->io_children_notready++;
274 		pio->io_children_notdone++;
275 		zio->io_sibling_next = pio->io_child;
276 		zio->io_sibling_prev = NULL;
277 		if (pio->io_child != NULL)
278 			pio->io_child->io_sibling_prev = zio;
279 		pio->io_child = zio;
280 		mutex_exit(&pio->io_lock);
281 	}
282 
283 	return (zio);
284 }
285 
286 zio_t *
287 zio_null(zio_t *pio, spa_t *spa, zio_done_func_t *done, void *private,
288 	int flags)
289 {
290 	zio_t *zio;
291 
292 	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
293 	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, ZIO_STAGE_OPEN,
294 	    ZIO_WAIT_FOR_CHILDREN_PIPELINE);
295 
296 	return (zio);
297 }
298 
299 zio_t *
300 zio_root(spa_t *spa, zio_done_func_t *done, void *private, int flags)
301 {
302 	return (zio_null(NULL, spa, done, private, flags));
303 }
304 
305 zio_t *
306 zio_read(zio_t *pio, spa_t *spa, blkptr_t *bp, void *data,
307     uint64_t size, zio_done_func_t *done, void *private,
308     int priority, int flags)
309 {
310 	zio_t *zio;
311 	dva_t *dva;
312 
313 	ASSERT3U(size, ==, BP_GET_LSIZE(bp));
314 
315 	zio = zio_create(pio, spa, bp->blk_birth, bp, data, size, done, private,
316 	    ZIO_TYPE_READ, priority, flags, ZIO_STAGE_OPEN, ZIO_READ_PIPELINE);
317 
318 	/*
319 	 * Work off our copy of the bp so the caller can free it.
320 	 */
321 	zio->io_bp = &zio->io_bp_copy;
322 
323 	bp = zio->io_bp;
324 	dva = ZIO_GET_DVA(zio);
325 
326 	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
327 		uint64_t csize = BP_GET_PSIZE(bp);
328 		void *cbuf = zio_buf_alloc(csize);
329 
330 		zio_push_transform(zio, cbuf, csize, csize);
331 		zio->io_pipeline |= 1U << ZIO_STAGE_READ_DECOMPRESS;
332 	}
333 
334 	if (DVA_GET_GANG(dva)) {
335 		uint64_t gsize = SPA_GANGBLOCKSIZE;
336 		void *gbuf = zio_buf_alloc(gsize);
337 
338 		zio_push_transform(zio, gbuf, gsize, gsize);
339 		zio->io_pipeline |= 1U << ZIO_STAGE_READ_GANG_MEMBERS;
340 	}
341 
342 	return (zio);
343 }
344 
345 zio_t *
346 zio_write(zio_t *pio, spa_t *spa, int checksum, int compress,
347     uint64_t txg, blkptr_t *bp, void *data, uint64_t size,
348     zio_done_func_t *done, void *private, int priority, int flags)
349 {
350 	zio_t *zio;
351 
352 	ASSERT(checksum >= ZIO_CHECKSUM_OFF &&
353 	    checksum < ZIO_CHECKSUM_FUNCTIONS);
354 
355 	ASSERT(compress >= ZIO_COMPRESS_OFF &&
356 	    compress < ZIO_COMPRESS_FUNCTIONS);
357 
358 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
359 	    ZIO_TYPE_WRITE, priority, flags,
360 	    ZIO_STAGE_OPEN, ZIO_WRITE_PIPELINE);
361 
362 	zio->io_checksum = checksum;
363 	zio->io_compress = compress;
364 
365 	if (compress != ZIO_COMPRESS_OFF)
366 		zio->io_async_stages |= 1U << ZIO_STAGE_WRITE_COMPRESS;
367 
368 	if (bp->blk_birth != txg) {
369 		/* XXX the bp usually (always?) gets re-zeroed later */
370 		BP_ZERO(bp);
371 		BP_SET_LSIZE(bp, size);
372 		BP_SET_PSIZE(bp, size);
373 	}
374 
375 	return (zio);
376 }
377 
378 zio_t *
379 zio_rewrite(zio_t *pio, spa_t *spa, int checksum,
380     uint64_t txg, blkptr_t *bp, void *data, uint64_t size,
381     zio_done_func_t *done, void *private, int priority, int flags)
382 {
383 	zio_t *zio;
384 
385 	/* XXBP - We need to re-evaluate when to insert pipeline stages */
386 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
387 	    ZIO_TYPE_WRITE, priority, flags,
388 	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
389 
390 	zio->io_checksum = checksum;
391 	zio->io_compress = ZIO_COMPRESS_OFF;
392 
393 	return (zio);
394 }
395 
396 static zio_t *
397 zio_write_allocate(zio_t *pio, spa_t *spa, int checksum,
398     uint64_t txg, blkptr_t *bp, void *data, uint64_t size,
399     zio_done_func_t *done, void *private, int priority, int flags)
400 {
401 	zio_t *zio;
402 
403 	BP_ZERO(bp);
404 	BP_SET_LSIZE(bp, size);
405 	BP_SET_PSIZE(bp, size);
406 	BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF);
407 
408 	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
409 	    ZIO_TYPE_WRITE, priority, flags,
410 	    ZIO_STAGE_OPEN, ZIO_WRITE_ALLOCATE_PIPELINE);
411 
412 	zio->io_checksum = checksum;
413 	zio->io_compress = ZIO_COMPRESS_OFF;
414 
415 	return (zio);
416 }
417 
418 zio_t *
419 zio_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
420     zio_done_func_t *done, void *private)
421 {
422 	zio_t *zio;
423 
424 	ASSERT(!BP_IS_HOLE(bp));
425 
426 	if (txg == spa->spa_syncing_txg &&
427 	    spa->spa_sync_pass > zio_sync_pass.zp_defer_free) {
428 		bplist_enqueue_deferred(&spa->spa_sync_bplist, bp);
429 		return (zio_null(pio, spa, NULL, NULL, 0));
430 	}
431 
432 	/* XXBP - We need to re-evaluate when to insert pipeline stages */
433 	zio = zio_create(pio, spa, txg, bp, NULL, 0, done, private,
434 	    ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, 0,
435 	    ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
436 
437 	zio->io_bp = &zio->io_bp_copy;
438 
439 	return (zio);
440 }
441 
442 zio_t *
443 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
444     zio_done_func_t *done, void *private)
445 {
446 	zio_t *zio;
447 
448 	/*
449 	 * A claim is an allocation of a specific block.  Claims are needed
450 	 * to support immediate writes in the intent log.  The issue is that
451 	 * immediate writes contain committed data, but in a txg that was
452 	 * *not* committed.  Upon opening the pool after an unclean shutdown,
453 	 * the intent log claims all blocks that contain immediate write data
454 	 * so that the SPA knows they're in use.
455 	 *
456 	 * All claims *must* be resolved in the first txg -- before the SPA
457 	 * starts allocating blocks -- so that nothing is allocated twice.
458 	 */
459 	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
460 	ASSERT3U(spa_first_txg(spa), <=, txg);
461 
462 	/* XXBP - We need to re-evaluate when to insert pipeline stages */
463 	zio = zio_create(pio, spa, txg, bp, NULL, 0, done, private,
464 	    ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, 0,
465 	    ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
466 
467 	zio->io_bp = &zio->io_bp_copy;
468 
469 	return (zio);
470 }
471 
472 zio_t *
473 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
474     zio_done_func_t *done, void *private, int priority, int flags)
475 {
476 	zio_t *zio;
477 	int c;
478 
479 	if (vd->vdev_children == 0) {
480 		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
481 		    ZIO_TYPE_IOCTL, priority, flags,
482 		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
483 
484 		zio->io_vd = vd;
485 		zio->io_cmd = cmd;
486 	} else {
487 		zio = zio_null(pio, spa, NULL, NULL, flags);
488 
489 		for (c = 0; c < vd->vdev_children; c++)
490 			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
491 			    done, private, priority, flags));
492 	}
493 
494 	return (zio);
495 }
496 
497 static void
498 zio_phys_bp_init(vdev_t *vd, blkptr_t *bp, uint64_t offset, uint64_t size,
499     int checksum)
500 {
501 	ASSERT(vd->vdev_children == 0);
502 
503 	ASSERT(size <= SPA_MAXBLOCKSIZE);
504 	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
505 	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
506 
507 	ASSERT(offset + size <= VDEV_LABEL_START_SIZE ||
508 	    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
509 	ASSERT3U(offset + size, <=, vd->vdev_psize);
510 
511 	BP_ZERO(bp);
512 
513 	BP_SET_LSIZE(bp, size);
514 	BP_SET_PSIZE(bp, size);
515 
516 	BP_SET_CHECKSUM(bp, checksum);
517 	BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF);
518 	BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
519 
520 	if (checksum != ZIO_CHECKSUM_OFF)
521 		ZIO_SET_CHECKSUM(&bp->blk_cksum, offset, 0, 0, 0);
522 }
523 
524 zio_t *
525 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
526     void *data, int checksum, zio_done_func_t *done, void *private,
527     int priority, int flags)
528 {
529 	zio_t *zio;
530 	blkptr_t blk;
531 
532 	zio_phys_bp_init(vd, &blk, offset, size, checksum);
533 
534 	zio = zio_create(pio, vd->vdev_spa, 0, &blk, data, size, done, private,
535 	    ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL,
536 	    ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
537 
538 	zio->io_vd = vd;
539 	zio->io_offset = offset;
540 
541 	/*
542 	 * Work off our copy of the bp so the caller can free it.
543 	 */
544 	zio->io_bp = &zio->io_bp_copy;
545 
546 	return (zio);
547 }
548 
549 zio_t *
550 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
551     void *data, int checksum, zio_done_func_t *done, void *private,
552     int priority, int flags)
553 {
554 	zio_block_tail_t *zbt;
555 	void *wbuf;
556 	zio_t *zio;
557 	blkptr_t blk;
558 
559 	zio_phys_bp_init(vd, &blk, offset, size, checksum);
560 
561 	zio = zio_create(pio, vd->vdev_spa, 0, &blk, data, size, done, private,
562 	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL,
563 	    ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
564 
565 	zio->io_vd = vd;
566 	zio->io_offset = offset;
567 
568 	zio->io_bp = &zio->io_bp_copy;
569 	zio->io_checksum = checksum;
570 
571 	if (zio_checksum_table[checksum].ci_zbt) {
572 		/*
573 		 * zbt checksums are necessarily destructive -- they modify
574 		 * one word of the write buffer to hold the verifier/checksum.
575 		 * Therefore, we must make a local copy in case the data is
576 		 * being written to multiple places.
577 		 */
578 		wbuf = zio_buf_alloc(size);
579 		bcopy(data, wbuf, size);
580 		zio_push_transform(zio, wbuf, size, size);
581 
582 		zbt = (zio_block_tail_t *)((char *)wbuf + size) - 1;
583 		zbt->zbt_cksum = blk.blk_cksum;
584 	}
585 
586 	return (zio);
587 }
588 
589 /*
590  * Create a child I/O to do some work for us.  It has no associated bp.
591  */
592 zio_t *
593 zio_vdev_child_io(zio_t *zio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
594 	void *data, uint64_t size, int type, int priority, int flags,
595 	zio_done_func_t *done, void *private)
596 {
597 	uint32_t pipeline = ZIO_VDEV_CHILD_PIPELINE;
598 	zio_t *cio;
599 
600 	if (type == ZIO_TYPE_READ && bp != NULL) {
601 		/*
602 		 * If we have the bp, then the child should perform the
603 		 * checksum and the parent need not.  This pushes error
604 		 * detection as close to the leaves as possible and
605 		 * eliminates redundant checksums in the interior nodes.
606 		 */
607 		pipeline |= 1U << ZIO_STAGE_CHECKSUM_VERIFY;
608 		zio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY);
609 	}
610 
611 	cio = zio_create(zio, zio->io_spa, zio->io_txg, bp, data, size,
612 	    done, private, type, priority,
613 	    (zio->io_flags & ZIO_FLAG_VDEV_INHERIT) | ZIO_FLAG_CANFAIL | flags,
614 	    ZIO_STAGE_VDEV_IO_SETUP - 1, pipeline);
615 
616 	cio->io_vd = vd;
617 	cio->io_offset = offset;
618 
619 	return (cio);
620 }
621 
622 /*
623  * ==========================================================================
624  * Initiate I/O, either sync or async
625  * ==========================================================================
626  */
627 int
628 zio_wait(zio_t *zio)
629 {
630 	int error;
631 
632 	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
633 
634 	zio->io_waiter = curthread;
635 
636 	zio_next_stage_async(zio);
637 
638 	mutex_enter(&zio->io_lock);
639 	while (zio->io_stalled != ZIO_STAGE_DONE)
640 		cv_wait(&zio->io_cv, &zio->io_lock);
641 	mutex_exit(&zio->io_lock);
642 
643 	error = zio->io_error;
644 
645 	kmem_free(zio, sizeof (zio_t));
646 
647 	return (error);
648 }
649 
650 void
651 zio_nowait(zio_t *zio)
652 {
653 	zio_next_stage_async(zio);
654 }
655 
656 /*
657  * ==========================================================================
658  * I/O pipeline interlocks: parent/child dependency scoreboarding
659  * ==========================================================================
660  */
661 static void
662 zio_wait_for_children(zio_t *zio, uint32_t stage, uint64_t *countp)
663 {
664 	mutex_enter(&zio->io_lock);
665 	if (*countp == 0) {
666 		ASSERT(zio->io_stalled == 0);
667 		mutex_exit(&zio->io_lock);
668 		zio_next_stage(zio);
669 	} else {
670 		if (zio->io_stage == ZIO_STAGE_VDEV_IO_START)
671 			zio_vdev_io_exit(zio);
672 		zio->io_stalled = stage;
673 		mutex_exit(&zio->io_lock);
674 	}
675 }
676 
677 static void
678 zio_notify_parent(zio_t *zio, uint32_t stage, uint64_t *countp)
679 {
680 	zio_t *pio = zio->io_parent;
681 
682 	mutex_enter(&pio->io_lock);
683 	if (pio->io_error == 0 && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
684 		pio->io_error = zio->io_error;
685 	if (--*countp == 0 && pio->io_stalled == stage) {
686 		if (pio->io_stage == ZIO_STAGE_VDEV_IO_START)
687 			zio_vdev_io_enter(pio);
688 		pio->io_stalled = 0;
689 		mutex_exit(&pio->io_lock);
690 		zio_next_stage_async(pio);
691 	} else {
692 		mutex_exit(&pio->io_lock);
693 	}
694 }
695 
696 static void
697 zio_wait_children_ready(zio_t *zio)
698 {
699 	zio_wait_for_children(zio, ZIO_STAGE_WAIT_CHILDREN_READY,
700 	    &zio->io_children_notready);
701 }
702 
703 void
704 zio_wait_children_done(zio_t *zio)
705 {
706 	zio_wait_for_children(zio, ZIO_STAGE_WAIT_CHILDREN_DONE,
707 	    &zio->io_children_notdone);
708 }
709 
710 static void
711 zio_ready(zio_t *zio)
712 {
713 	zio_t *pio = zio->io_parent;
714 
715 	if (pio != NULL)
716 		zio_notify_parent(zio, ZIO_STAGE_WAIT_CHILDREN_READY,
717 		    &pio->io_children_notready);
718 
719 	if (zio->io_bp)
720 		zio->io_bp_copy = *zio->io_bp;
721 
722 	zio_next_stage(zio);
723 }
724 
725 static void
726 zio_done(zio_t *zio)
727 {
728 	zio_t *pio = zio->io_parent;
729 	spa_t *spa = zio->io_spa;
730 	blkptr_t *bp = zio->io_bp;
731 	vdev_t *vd = zio->io_vd;
732 	char blkbuf[BP_SPRINTF_LEN];
733 
734 	ASSERT(zio->io_children_notready == 0);
735 	ASSERT(zio->io_children_notdone == 0);
736 
737 	if (bp != NULL) {
738 		ASSERT(bp->blk_pad[0] == 0);
739 		ASSERT(bp->blk_pad[1] == 0);
740 		ASSERT(bp->blk_pad[2] == 0);
741 		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0);
742 		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
743 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
744 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
745 	}
746 
747 	if (vd != NULL)
748 		vdev_stat_update(zio);
749 
750 	if (zio->io_error) {
751 		sprintf_blkptr(blkbuf, BP_SPRINTF_LEN,
752 		    bp ? bp : &zio->io_bp_copy);
753 		dprintf("ZFS: %s (%s on %s off %llx: zio %p %s): error %d\n",
754 		    zio->io_error == ECKSUM ? "bad checksum" : "I/O failure",
755 		    zio_type_name[zio->io_type],
756 		    vdev_description(vd),
757 		    (u_longlong_t)zio->io_offset,
758 		    zio, blkbuf, zio->io_error);
759 	}
760 
761 	if (zio->io_numerrors != 0 && zio->io_type == ZIO_TYPE_WRITE) {
762 		sprintf_blkptr(blkbuf, BP_SPRINTF_LEN,
763 		    bp ? bp : &zio->io_bp_copy);
764 		dprintf("ZFS: %s (%s on %s off %llx: zio %p %s): %d errors\n",
765 		    "partial write",
766 		    zio_type_name[zio->io_type],
767 		    vdev_description(vd),
768 		    (u_longlong_t)zio->io_offset,
769 		    zio, blkbuf, zio->io_numerrors);
770 	}
771 
772 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
773 		sprintf_blkptr(blkbuf, BP_SPRINTF_LEN,
774 		    bp ? bp : &zio->io_bp_copy);
775 		panic("ZFS: %s (%s on %s off %llx: zio %p %s): error %d",
776 		    zio->io_error == ECKSUM ? "bad checksum" : "I/O failure",
777 		    zio_type_name[zio->io_type],
778 		    vdev_description(vd),
779 		    (u_longlong_t)zio->io_offset,
780 		    zio, blkbuf, zio->io_error);
781 	}
782 
783 	zio_clear_transform_stack(zio);
784 
785 	if (zio->io_done)
786 		zio->io_done(zio);
787 
788 	ASSERT(zio->io_delegate_list == NULL);
789 	ASSERT(zio->io_delegate_next == NULL);
790 
791 	if (pio != NULL) {
792 		zio_t *next, *prev;
793 
794 		mutex_enter(&pio->io_lock);
795 		next = zio->io_sibling_next;
796 		prev = zio->io_sibling_prev;
797 		if (next != NULL)
798 			next->io_sibling_prev = prev;
799 		if (prev != NULL)
800 			prev->io_sibling_next = next;
801 		if (pio->io_child == zio)
802 			pio->io_child = next;
803 		mutex_exit(&pio->io_lock);
804 
805 		zio_notify_parent(zio, ZIO_STAGE_WAIT_CHILDREN_DONE,
806 		    &pio->io_children_notdone);
807 	}
808 
809 	if (pio == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_HELD))
810 		spa_config_exit(spa);
811 
812 	if (zio->io_waiter != NULL) {
813 		mutex_enter(&zio->io_lock);
814 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
815 		zio->io_stalled = zio->io_stage;
816 		cv_broadcast(&zio->io_cv);
817 		mutex_exit(&zio->io_lock);
818 	} else {
819 		kmem_free(zio, sizeof (zio_t));
820 	}
821 }
822 
823 /*
824  * ==========================================================================
825  * Compression support
826  * ==========================================================================
827  */
828 static void
829 zio_write_compress(zio_t *zio)
830 {
831 	int compress = zio->io_compress;
832 	blkptr_t *bp = zio->io_bp;
833 	void *cbuf;
834 	uint64_t lsize = zio->io_size;
835 	uint64_t csize = lsize;
836 	uint64_t cbufsize = 0;
837 	int pass;
838 
839 	if (bp->blk_birth == zio->io_txg) {
840 		/*
841 		 * We're rewriting an existing block, which means we're
842 		 * working on behalf of spa_sync().  For spa_sync() to
843 		 * converge, it must eventually be the case that we don't
844 		 * have to allocate new blocks.  But compression changes
845 		 * the blocksize, which forces a reallocate, and makes
846 		 * convergence take longer.  Therefore, after the first
847 		 * few passes, stop compressing to ensure convergence.
848 		 */
849 		pass = spa_sync_pass(zio->io_spa);
850 		if (pass > zio_sync_pass.zp_dontcompress)
851 			compress = ZIO_COMPRESS_OFF;
852 	} else {
853 		ASSERT(BP_IS_HOLE(bp));
854 		pass = 1;
855 	}
856 
857 	if (compress != ZIO_COMPRESS_OFF)
858 		if (!zio_compress_data(compress, zio->io_data, zio->io_size,
859 		    &cbuf, &csize, &cbufsize))
860 			compress = ZIO_COMPRESS_OFF;
861 
862 	if (compress != ZIO_COMPRESS_OFF && csize != 0)
863 		zio_push_transform(zio, cbuf, csize, cbufsize);
864 
865 	/*
866 	 * The final pass of spa_sync() must be all rewrites, but the first
867 	 * few passes offer a trade-off: allocating blocks defers convergence,
868 	 * but newly allocated blocks are sequential, so they can be written
869 	 * to disk faster.  Therefore, we allow the first few passes of
870 	 * spa_sync() to reallocate new blocks, but force rewrites after that.
871 	 * There should only be a handful of blocks after pass 1 in any case.
872 	 */
873 	if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == csize &&
874 	    pass > zio_sync_pass.zp_rewrite) {
875 		ASSERT(csize != 0);
876 		ASSERT3U(BP_GET_COMPRESS(bp), ==, compress);
877 		ASSERT3U(BP_GET_LSIZE(bp), ==, lsize);
878 
879 		zio->io_pipeline = ZIO_REWRITE_PIPELINE;
880 	} else {
881 		if (bp->blk_birth == zio->io_txg) {
882 			ASSERT3U(BP_GET_LSIZE(bp), ==, lsize);
883 			bzero(bp, sizeof (blkptr_t));
884 		}
885 		if (csize == 0) {
886 			BP_ZERO(bp);
887 			zio->io_pipeline = ZIO_WAIT_FOR_CHILDREN_PIPELINE;
888 		} else {
889 			BP_SET_LSIZE(bp, lsize);
890 			BP_SET_PSIZE(bp, csize);
891 			BP_SET_COMPRESS(bp, compress);
892 			zio->io_pipeline = ZIO_WRITE_ALLOCATE_PIPELINE;
893 		}
894 	}
895 
896 	zio_next_stage(zio);
897 }
898 
899 static void
900 zio_read_decompress(zio_t *zio)
901 {
902 	blkptr_t *bp = zio->io_bp;
903 	void *data;
904 	uint64_t size;
905 	uint64_t bufsize;
906 	int compress = BP_GET_COMPRESS(bp);
907 
908 	ASSERT(compress != ZIO_COMPRESS_OFF);
909 
910 	zio_pop_transform(zio, &data, &size, &bufsize);
911 
912 	if (zio_decompress_data(compress, data, size,
913 	    zio->io_data, zio->io_size))
914 		zio->io_error = EIO;
915 
916 	zio_buf_free(data, bufsize);
917 
918 	zio_next_stage(zio);
919 }
920 
921 /*
922  * ==========================================================================
923  * Gang block support
924  * ==========================================================================
925  */
926 static void
927 zio_gang_pipeline(zio_t *zio)
928 {
929 	/*
930 	 * By default, the pipeline assumes that we're dealing with a gang
931 	 * block.  If we're not, strip out any gang-specific stages.
932 	 */
933 	if (!DVA_GET_GANG(ZIO_GET_DVA(zio)))
934 		zio->io_pipeline &= ~ZIO_GANG_STAGES;
935 
936 	zio_next_stage(zio);
937 }
938 
939 static void
940 zio_gang_byteswap(zio_t *zio)
941 {
942 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
943 
944 	if (BP_SHOULD_BYTESWAP(zio->io_bp))
945 		byteswap_uint64_array(zio->io_data, zio->io_size);
946 }
947 
948 static void
949 zio_get_gang_header(zio_t *zio)
950 {
951 	blkptr_t *bp = zio->io_bp;
952 	uint64_t gsize = SPA_GANGBLOCKSIZE;
953 	void *gbuf = zio_buf_alloc(gsize);
954 
955 	ASSERT(DVA_GET_GANG(ZIO_GET_DVA(zio)));
956 
957 	zio_push_transform(zio, gbuf, gsize, gsize);
958 
959 	zio_nowait(zio_create(zio, zio->io_spa, bp->blk_birth, bp, gbuf, gsize,
960 	    NULL, NULL, ZIO_TYPE_READ, zio->io_priority,
961 	    zio->io_flags & ZIO_FLAG_GANG_INHERIT,
962 	    ZIO_STAGE_OPEN, ZIO_READ_PIPELINE));
963 
964 	zio_wait_children_done(zio);
965 }
966 
967 static void
968 zio_read_gang_members(zio_t *zio)
969 {
970 	zio_gbh_phys_t *gbh;
971 	uint64_t gsize, gbufsize, loff, lsize;
972 	int i;
973 
974 	ASSERT(DVA_GET_GANG(ZIO_GET_DVA(zio)));
975 
976 	zio_gang_byteswap(zio);
977 	zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize);
978 
979 	for (loff = 0, i = 0; loff != zio->io_size; loff += lsize, i++) {
980 		blkptr_t *gbp = &gbh->zg_blkptr[i];
981 		lsize = BP_GET_PSIZE(gbp);
982 
983 		ASSERT(BP_GET_COMPRESS(gbp) == ZIO_COMPRESS_OFF);
984 		ASSERT3U(lsize, ==, BP_GET_LSIZE(gbp));
985 		ASSERT3U(loff + lsize, <=, zio->io_size);
986 		ASSERT(i < SPA_GBH_NBLKPTRS);
987 		ASSERT(!BP_IS_HOLE(gbp));
988 
989 		zio_nowait(zio_read(zio, zio->io_spa, gbp,
990 		    (char *)zio->io_data + loff, lsize, NULL, NULL,
991 		    zio->io_priority, zio->io_flags & ZIO_FLAG_GANG_INHERIT));
992 	}
993 
994 	zio_buf_free(gbh, gbufsize);
995 	zio_wait_children_done(zio);
996 }
997 
998 static void
999 zio_rewrite_gang_members(zio_t *zio)
1000 {
1001 	zio_gbh_phys_t *gbh;
1002 	uint64_t gsize, gbufsize, loff, lsize;
1003 	int i;
1004 
1005 	ASSERT(DVA_GET_GANG(ZIO_GET_DVA(zio)));
1006 	ASSERT3U(zio->io_size, ==, SPA_GANGBLOCKSIZE);
1007 
1008 	zio_gang_byteswap(zio);
1009 	zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize);
1010 
1011 	ASSERT(gsize == gbufsize);
1012 
1013 	for (loff = 0, i = 0; loff != zio->io_size; loff += lsize, i++) {
1014 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1015 		lsize = BP_GET_PSIZE(gbp);
1016 
1017 		ASSERT(BP_GET_COMPRESS(gbp) == ZIO_COMPRESS_OFF);
1018 		ASSERT3U(lsize, ==, BP_GET_LSIZE(gbp));
1019 		ASSERT3U(loff + lsize, <=, zio->io_size);
1020 		ASSERT(i < SPA_GBH_NBLKPTRS);
1021 		ASSERT(!BP_IS_HOLE(gbp));
1022 
1023 		zio_nowait(zio_rewrite(zio, zio->io_spa, zio->io_checksum,
1024 		    zio->io_txg, gbp, (char *)zio->io_data + loff, lsize,
1025 		    NULL, NULL, zio->io_priority, zio->io_flags));
1026 	}
1027 
1028 	zio_push_transform(zio, gbh, gsize, gbufsize);
1029 	zio_wait_children_ready(zio);
1030 }
1031 
1032 static void
1033 zio_free_gang_members(zio_t *zio)
1034 {
1035 	zio_gbh_phys_t *gbh;
1036 	uint64_t gsize, gbufsize;
1037 	int i;
1038 
1039 	ASSERT(DVA_GET_GANG(ZIO_GET_DVA(zio)));
1040 
1041 	zio_gang_byteswap(zio);
1042 	zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize);
1043 
1044 	for (i = 0; i < SPA_GBH_NBLKPTRS; i++) {
1045 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1046 
1047 		if (BP_IS_HOLE(gbp))
1048 			continue;
1049 		zio_nowait(zio_free(zio, zio->io_spa, zio->io_txg,
1050 		    gbp, NULL, NULL));
1051 	}
1052 
1053 	zio_buf_free(gbh, gbufsize);
1054 	zio_next_stage(zio);
1055 }
1056 
1057 static void
1058 zio_claim_gang_members(zio_t *zio)
1059 {
1060 	zio_gbh_phys_t *gbh;
1061 	uint64_t gsize, gbufsize;
1062 	int i;
1063 
1064 	ASSERT(DVA_GET_GANG(ZIO_GET_DVA(zio)));
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 		if (BP_IS_HOLE(gbp))
1072 			continue;
1073 		zio_nowait(zio_claim(zio, zio->io_spa, zio->io_txg,
1074 		    gbp, NULL, NULL));
1075 	}
1076 
1077 	zio_buf_free(gbh, gbufsize);
1078 	zio_next_stage(zio);
1079 }
1080 
1081 static void
1082 zio_write_allocate_gang_member_done(zio_t *zio)
1083 {
1084 	zio_t *pio = zio->io_parent;
1085 	dva_t *cdva = ZIO_GET_DVA(zio);
1086 	dva_t *pdva = ZIO_GET_DVA(pio);
1087 	uint64_t asize;
1088 
1089 	ASSERT(DVA_GET_GANG(pdva));
1090 
1091 	/* XXBP - Need to be careful here with multiple DVAs */
1092 	mutex_enter(&pio->io_lock);
1093 	asize = DVA_GET_ASIZE(pdva);
1094 	asize += DVA_GET_ASIZE(cdva);
1095 	DVA_SET_ASIZE(pdva, asize);
1096 	mutex_exit(&pio->io_lock);
1097 }
1098 
1099 static void
1100 zio_write_allocate_gang_members(zio_t *zio)
1101 {
1102 	blkptr_t *bp = zio->io_bp;
1103 	dva_t *dva = ZIO_GET_DVA(zio);
1104 	zio_gbh_phys_t *gbh;
1105 	uint64_t resid = zio->io_size;
1106 	uint64_t maxalloc = P2ROUNDUP(zio->io_size >> 1, SPA_MINBLOCKSIZE);
1107 	uint64_t gsize, loff, lsize;
1108 	uint32_t gbps_left;
1109 	int error;
1110 	int i;
1111 
1112 	gsize = SPA_GANGBLOCKSIZE;
1113 	gbps_left = SPA_GBH_NBLKPTRS;
1114 
1115 	error = metaslab_alloc(zio->io_spa, gsize, dva, zio->io_txg);
1116 	if (error == ENOSPC)
1117 		panic("can't allocate gang block header");
1118 	ASSERT(error == 0);
1119 
1120 	DVA_SET_GANG(dva, 1);
1121 
1122 	bp->blk_birth = zio->io_txg;
1123 
1124 	gbh = zio_buf_alloc(gsize);
1125 	bzero(gbh, gsize);
1126 
1127 	for (loff = 0, i = 0; loff != zio->io_size;
1128 	    loff += lsize, resid -= lsize, gbps_left--, i++) {
1129 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1130 		dva = &gbp->blk_dva[0];
1131 
1132 		ASSERT(gbps_left != 0);
1133 		maxalloc = MIN(maxalloc, resid);
1134 
1135 		while (resid <= maxalloc * gbps_left) {
1136 			error = metaslab_alloc(zio->io_spa, maxalloc, dva,
1137 			    zio->io_txg);
1138 			if (error == 0)
1139 				break;
1140 			ASSERT3U(error, ==, ENOSPC);
1141 			if (maxalloc == SPA_MINBLOCKSIZE)
1142 				panic("really out of space");
1143 			maxalloc = P2ROUNDUP(maxalloc >> 1, SPA_MINBLOCKSIZE);
1144 		}
1145 
1146 		if (resid <= maxalloc * gbps_left) {
1147 			lsize = maxalloc;
1148 			BP_SET_LSIZE(gbp, lsize);
1149 			BP_SET_PSIZE(gbp, lsize);
1150 			BP_SET_COMPRESS(gbp, ZIO_COMPRESS_OFF);
1151 			gbp->blk_birth = zio->io_txg;
1152 			zio_nowait(zio_rewrite(zio, zio->io_spa,
1153 			    zio->io_checksum, zio->io_txg, gbp,
1154 			    (char *)zio->io_data + loff, lsize,
1155 			    zio_write_allocate_gang_member_done, NULL,
1156 			    zio->io_priority, zio->io_flags));
1157 		} else {
1158 			lsize = P2ROUNDUP(resid / gbps_left, SPA_MINBLOCKSIZE);
1159 			ASSERT(lsize != SPA_MINBLOCKSIZE);
1160 			zio_nowait(zio_write_allocate(zio, zio->io_spa,
1161 			    zio->io_checksum, zio->io_txg, gbp,
1162 			    (char *)zio->io_data + loff, lsize,
1163 			    zio_write_allocate_gang_member_done, NULL,
1164 			    zio->io_priority, zio->io_flags));
1165 		}
1166 	}
1167 
1168 	ASSERT(resid == 0 && loff == zio->io_size);
1169 
1170 	zio->io_pipeline |= 1U << ZIO_STAGE_GANG_CHECKSUM_GENERATE;
1171 
1172 	zio_push_transform(zio, gbh, gsize, gsize);
1173 	zio_wait_children_done(zio);
1174 }
1175 
1176 /*
1177  * ==========================================================================
1178  * Allocate and free blocks
1179  * ==========================================================================
1180  */
1181 static void
1182 zio_dva_allocate(zio_t *zio)
1183 {
1184 	blkptr_t *bp = zio->io_bp;
1185 	dva_t *dva = ZIO_GET_DVA(zio);
1186 	int error;
1187 
1188 	ASSERT(BP_IS_HOLE(bp));
1189 
1190 	/* For testing, make some blocks above a certain size be gang blocks */
1191 	if (zio->io_size >= zio_gang_bang && (lbolt & 0x3) == 0) {
1192 		zio_write_allocate_gang_members(zio);
1193 		return;
1194 	}
1195 
1196 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
1197 
1198 	error = metaslab_alloc(zio->io_spa, zio->io_size, dva, zio->io_txg);
1199 
1200 	if (error == 0) {
1201 		bp->blk_birth = zio->io_txg;
1202 	} else if (error == ENOSPC) {
1203 		if (zio->io_size == SPA_MINBLOCKSIZE)
1204 			panic("really, truly out of space");
1205 		zio_write_allocate_gang_members(zio);
1206 		return;
1207 	} else {
1208 		zio->io_error = error;
1209 	}
1210 	zio_next_stage(zio);
1211 }
1212 
1213 static void
1214 zio_dva_free(zio_t *zio)
1215 {
1216 	blkptr_t *bp = zio->io_bp;
1217 	dva_t *dva = ZIO_GET_DVA(zio);
1218 
1219 	ASSERT(!BP_IS_HOLE(bp));
1220 
1221 	metaslab_free(zio->io_spa, dva, zio->io_txg);
1222 
1223 	BP_ZERO(bp);
1224 
1225 	zio_next_stage(zio);
1226 }
1227 
1228 static void
1229 zio_dva_claim(zio_t *zio)
1230 {
1231 	blkptr_t *bp = zio->io_bp;
1232 	dva_t *dva = ZIO_GET_DVA(zio);
1233 
1234 	ASSERT(!BP_IS_HOLE(bp));
1235 
1236 	zio->io_error = metaslab_claim(zio->io_spa, dva, zio->io_txg);
1237 
1238 	zio_next_stage(zio);
1239 }
1240 
1241 static void
1242 zio_dva_translate(zio_t *zio)
1243 {
1244 	spa_t *spa = zio->io_spa;
1245 	dva_t *dva = ZIO_GET_DVA(zio);
1246 	uint64_t vdev = DVA_GET_VDEV(dva);
1247 	uint64_t offset = DVA_GET_OFFSET(dva);
1248 
1249 	ASSERT3U(zio->io_size, ==, ZIO_GET_IOSIZE(zio));
1250 
1251 	zio->io_offset = offset;
1252 
1253 	if ((zio->io_vd = vdev_lookup_top(spa, vdev)) == NULL)
1254 		zio->io_error = ENXIO;
1255 	else if (offset + zio->io_size > zio->io_vd->vdev_asize)
1256 		zio->io_error = EOVERFLOW;
1257 
1258 	zio_next_stage(zio);
1259 }
1260 
1261 /*
1262  * ==========================================================================
1263  * Read and write to physical devices
1264  * ==========================================================================
1265  */
1266 static void
1267 zio_vdev_io_enter(zio_t *zio)
1268 {
1269 	vdev_t *tvd = zio->io_vd->vdev_top;
1270 
1271 	mutex_enter(&tvd->vdev_io_lock);
1272 	ASSERT(zio->io_pending.list_next == NULL);
1273 	list_insert_tail(&tvd->vdev_io_pending, zio);
1274 	mutex_exit(&tvd->vdev_io_lock);
1275 }
1276 
1277 static void
1278 zio_vdev_io_exit(zio_t *zio)
1279 {
1280 	vdev_t *tvd = zio->io_vd->vdev_top;
1281 
1282 	mutex_enter(&tvd->vdev_io_lock);
1283 	ASSERT(zio->io_pending.list_next != NULL);
1284 	list_remove(&tvd->vdev_io_pending, zio);
1285 	if (list_head(&tvd->vdev_io_pending) == NULL)
1286 		cv_broadcast(&tvd->vdev_io_cv);
1287 	mutex_exit(&tvd->vdev_io_lock);
1288 }
1289 
1290 static void
1291 zio_vdev_io_retry(void *vdarg)
1292 {
1293 	vdev_t *vd = vdarg;
1294 	zio_t *zio, *zq;
1295 
1296 	ASSERT(vd == vd->vdev_top);
1297 
1298 	/* XXPOLICY */
1299 	delay(hz);
1300 
1301 	vdev_reopen(vd, &zq);
1302 
1303 	while ((zio = zq) != NULL) {
1304 		zq = zio->io_retry_next;
1305 		zio->io_retry_next = NULL;
1306 		dprintf("async retry #%d for I/O to %s offset %llx\n",
1307 		    zio->io_retries, vdev_description(vd), zio->io_offset);
1308 		zio_next_stage_async(zio);
1309 	}
1310 }
1311 
1312 static void
1313 zio_vdev_io_setup(zio_t *zio)
1314 {
1315 	vdev_t *vd = zio->io_vd;
1316 
1317 	/* XXPOLICY */
1318 	if (zio->io_retries == 0 && vd == vd->vdev_top)
1319 		zio->io_flags |= ZIO_FLAG_FAILFAST;
1320 
1321 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && vd->vdev_children == 0) {
1322 		zio->io_flags |= ZIO_FLAG_PHYSICAL;
1323 		zio->io_offset += VDEV_LABEL_START_SIZE;
1324 	}
1325 
1326 	zio_vdev_io_enter(zio);
1327 
1328 	zio_next_stage(zio);
1329 }
1330 
1331 static void
1332 zio_vdev_io_start(zio_t *zio)
1333 {
1334 	blkptr_t *bp = zio->io_bp;
1335 
1336 	ASSERT(P2PHASE(zio->io_offset, 1ULL << zio->io_vd->vdev_ashift) == 0);
1337 	ASSERT(P2PHASE(zio->io_size, 1ULL << zio->io_vd->vdev_ashift) == 0);
1338 	ASSERT(bp == NULL || ZIO_GET_IOSIZE(zio) == zio->io_size);
1339 	ASSERT(zio->io_type != ZIO_TYPE_WRITE || (spa_mode & FWRITE));
1340 
1341 	vdev_io_start(zio);
1342 
1343 	/* zio_next_stage_async() gets called from io completion interrupt */
1344 }
1345 
1346 static void
1347 zio_vdev_io_done(zio_t *zio)
1348 {
1349 	vdev_io_done(zio);
1350 }
1351 
1352 /* XXPOLICY */
1353 static boolean_t
1354 zio_should_retry(zio_t *zio)
1355 {
1356 	vdev_t *vd = zio->io_vd;
1357 
1358 	if (zio->io_error == 0)
1359 		return (B_FALSE);
1360 	if (zio->io_delegate_list != NULL)
1361 		return (B_FALSE);
1362 	if (vd != vd->vdev_top)
1363 		return (B_FALSE);
1364 	if (zio->io_flags & ZIO_FLAG_DONT_RETRY)
1365 		return (B_FALSE);
1366 	if (zio->io_retries > 300 &&
1367 	    (zio->io_flags & (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_CANFAIL)))
1368 		return (B_FALSE);
1369 	if (zio->io_retries > 1 &&
1370 	    (zio->io_error == ECKSUM || zio->io_error == ENXIO))
1371 		return (B_FALSE);
1372 
1373 	return (B_TRUE);
1374 }
1375 
1376 static void
1377 zio_vdev_io_assess(zio_t *zio)
1378 {
1379 	vdev_t *vd = zio->io_vd;
1380 	vdev_t *tvd = vd->vdev_top;
1381 
1382 	zio_vdev_io_exit(zio);
1383 
1384 	ASSERT(zio->io_vsd == NULL);
1385 
1386 	/*
1387 	 * If the I/O failed, determine whether we should attempt to retry it.
1388 	 */
1389 	/* XXPOLICY */
1390 	if (zio_should_retry(zio)) {
1391 		zio_t *zq;
1392 
1393 		ASSERT(tvd == vd);
1394 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE));
1395 
1396 		zio->io_retries++;
1397 		zio->io_error = 0;
1398 		zio->io_flags &= ZIO_FLAG_VDEV_INHERIT;
1399 		/* XXPOLICY */
1400 		zio->io_flags &= ~ZIO_FLAG_FAILFAST;
1401 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1402 		zio->io_stage = ZIO_STAGE_VDEV_IO_SETUP - 1;
1403 
1404 		dprintf("retry #%d for %s to %s offset %llx\n",
1405 		    zio->io_retries, zio_type_name[zio->io_type],
1406 		    vdev_description(vd), zio->io_offset);
1407 
1408 		/*
1409 		 * If this is the first retry, do it immediately.
1410 		 */
1411 		/* XXPOLICY */
1412 		if (zio->io_retries == 1) {
1413 			zio_next_stage_async(zio);
1414 			return;
1415 		}
1416 
1417 		/*
1418 		 * This was not the first retry, so go through the
1419 		 * longer enqueue/delay/vdev_reopen() process.
1420 		 */
1421 		mutex_enter(&tvd->vdev_io_lock);
1422 		ASSERT(zio->io_retry_next == NULL);
1423 		zio->io_retry_next = zq = tvd->vdev_io_retry;
1424 		tvd->vdev_io_retry = zio;
1425 		mutex_exit(&tvd->vdev_io_lock);
1426 		if (zq == NULL)
1427 			(void) taskq_dispatch(
1428 			    tvd->vdev_spa->spa_vdev_retry_taskq,
1429 			    zio_vdev_io_retry, tvd, TQ_SLEEP);
1430 		return;
1431 	}
1432 
1433 	zio_next_stage(zio);
1434 }
1435 
1436 void
1437 zio_vdev_io_reissue(zio_t *zio)
1438 {
1439 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
1440 	ASSERT(zio->io_error == 0);
1441 
1442 	zio->io_stage--;
1443 }
1444 
1445 void
1446 zio_vdev_io_redone(zio_t *zio)
1447 {
1448 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
1449 
1450 	zio->io_stage--;
1451 }
1452 
1453 void
1454 zio_vdev_io_bypass(zio_t *zio)
1455 {
1456 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
1457 	ASSERT(zio->io_error == 0);
1458 
1459 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
1460 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS - 1;
1461 }
1462 
1463 /*
1464  * ==========================================================================
1465  * Generate and verify checksums
1466  * ==========================================================================
1467  */
1468 static void
1469 zio_checksum_generate(zio_t *zio)
1470 {
1471 	int checksum = zio->io_checksum;
1472 	blkptr_t *bp = zio->io_bp;
1473 
1474 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
1475 
1476 	BP_SET_CHECKSUM(bp, checksum);
1477 	BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1478 
1479 	zio_checksum(checksum, &bp->blk_cksum, zio->io_data, zio->io_size);
1480 
1481 	zio_next_stage(zio);
1482 }
1483 
1484 static void
1485 zio_gang_checksum_generate(zio_t *zio)
1486 {
1487 	zio_cksum_t zc;
1488 	zio_gbh_phys_t *gbh = zio->io_data;
1489 
1490 	ASSERT3U(zio->io_size, ==, SPA_GANGBLOCKSIZE);
1491 	ASSERT(DVA_GET_GANG(ZIO_GET_DVA(zio)));
1492 
1493 	zio_set_gang_verifier(zio, &gbh->zg_tail.zbt_cksum);
1494 
1495 	zio_checksum(ZIO_CHECKSUM_GANG_HEADER, &zc, zio->io_data, zio->io_size);
1496 
1497 	zio_next_stage(zio);
1498 }
1499 
1500 static void
1501 zio_checksum_verify(zio_t *zio)
1502 {
1503 	if (zio->io_bp != NULL) {
1504 		zio->io_error = zio_checksum_error(zio);
1505 		if (zio->io_error) {
1506 			dprintf("bad checksum on vdev %s\n",
1507 			    vdev_description(zio->io_vd));
1508 		}
1509 	}
1510 
1511 	zio_next_stage(zio);
1512 }
1513 
1514 /*
1515  * Called by RAID-Z to ensure we don't compute the checksum twice.
1516  */
1517 void
1518 zio_checksum_verified(zio_t *zio)
1519 {
1520 	zio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY);
1521 }
1522 
1523 /*
1524  * Set the external verifier for a gang block based on stuff in the bp
1525  */
1526 void
1527 zio_set_gang_verifier(zio_t *zio, zio_cksum_t *zcp)
1528 {
1529 	zcp->zc_word[0] = DVA_GET_VDEV(ZIO_GET_DVA(zio));
1530 	zcp->zc_word[1] = DVA_GET_OFFSET(ZIO_GET_DVA(zio));
1531 	zcp->zc_word[2] = zio->io_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_dva_translate,
1564 	zio_vdev_io_setup,
1565 	zio_vdev_io_start,
1566 	zio_vdev_io_done,
1567 	zio_vdev_io_assess,
1568 	zio_wait_children_done,
1569 	zio_checksum_verify,
1570 	zio_read_gang_members,
1571 	zio_read_decompress,
1572 	zio_done,
1573 	zio_badop
1574 };
1575 
1576 /*
1577  * Move an I/O to the next stage of the pipeline and execute that stage.
1578  * There's no locking on io_stage because there's no legitimate way for
1579  * multiple threads to be attempting to process the same I/O.
1580  */
1581 void
1582 zio_next_stage(zio_t *zio)
1583 {
1584 	uint32_t pipeline = zio->io_pipeline;
1585 
1586 	ASSERT(!MUTEX_HELD(&zio->io_lock));
1587 
1588 	if (zio->io_error) {
1589 		dprintf("zio %p vdev %s offset %llx stage %d error %d\n",
1590 		    zio, vdev_description(zio->io_vd),
1591 		    zio->io_offset, zio->io_stage, zio->io_error);
1592 		if (((1U << zio->io_stage) & ZIO_VDEV_IO_PIPELINE) == 0)
1593 			pipeline &= ZIO_ERROR_PIPELINE_MASK;
1594 	}
1595 
1596 	while (((1U << ++zio->io_stage) & pipeline) == 0)
1597 		continue;
1598 
1599 	ASSERT(zio->io_stage <= ZIO_STAGE_DONE);
1600 	ASSERT(zio->io_stalled == 0);
1601 
1602 	zio_pipeline[zio->io_stage](zio);
1603 }
1604 
1605 void
1606 zio_next_stage_async(zio_t *zio)
1607 {
1608 	taskq_t *tq;
1609 	uint32_t pipeline = zio->io_pipeline;
1610 
1611 	ASSERT(!MUTEX_HELD(&zio->io_lock));
1612 
1613 	if (zio->io_error) {
1614 		dprintf("zio %p vdev %s offset %llx stage %d error %d\n",
1615 		    zio, vdev_description(zio->io_vd),
1616 		    zio->io_offset, zio->io_stage, zio->io_error);
1617 		if (((1U << zio->io_stage) & ZIO_VDEV_IO_PIPELINE) == 0)
1618 			pipeline &= ZIO_ERROR_PIPELINE_MASK;
1619 	}
1620 
1621 	while (((1U << ++zio->io_stage) & pipeline) == 0)
1622 		continue;
1623 
1624 	ASSERT(zio->io_stage <= ZIO_STAGE_DONE);
1625 	ASSERT(zio->io_stalled == 0);
1626 
1627 	/*
1628 	 * For performance, we'll probably want two sets of task queues:
1629 	 * per-CPU issue taskqs and per-CPU completion taskqs.  The per-CPU
1630 	 * part is for read performance: since we have to make a pass over
1631 	 * the data to checksum it anyway, we want to do this on the same CPU
1632 	 * that issued the read, because (assuming CPU scheduling affinity)
1633 	 * that thread is probably still there.  Getting this optimization
1634 	 * right avoids performance-hostile cache-to-cache transfers.
1635 	 *
1636 	 * Note that having two sets of task queues is also necessary for
1637 	 * correctness: if all of the issue threads get bogged down waiting
1638 	 * for dependent reads (e.g. metaslab freelist) to complete, then
1639 	 * there won't be any threads available to service I/O completion
1640 	 * interrupts.
1641 	 */
1642 	if ((1U << zio->io_stage) & zio->io_async_stages) {
1643 		if (zio->io_stage < ZIO_STAGE_VDEV_IO_DONE)
1644 			tq = zio->io_spa->spa_zio_issue_taskq[zio->io_type];
1645 		else
1646 			tq = zio->io_spa->spa_zio_intr_taskq[zio->io_type];
1647 		(void) taskq_dispatch(tq,
1648 		    (task_func_t *)zio_pipeline[zio->io_stage], zio, TQ_SLEEP);
1649 	} else {
1650 		zio_pipeline[zio->io_stage](zio);
1651 	}
1652 }
1653 
1654 /*
1655  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
1656  */
1657 int
1658 zio_alloc_blk(spa_t *spa, int checksum, uint64_t size, blkptr_t *bp,
1659     uint64_t txg)
1660 {
1661 	int error;
1662 
1663 	spa_config_enter(spa, RW_READER);
1664 
1665 	BP_ZERO(bp);
1666 
1667 	error = metaslab_alloc(spa, size, BP_IDENTITY(bp), txg);
1668 
1669 	if (error == 0) {
1670 		BP_SET_CHECKSUM(bp, checksum);
1671 		BP_SET_LSIZE(bp, size);
1672 		BP_SET_PSIZE(bp, size);
1673 		BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF);
1674 		BP_SET_TYPE(bp, DMU_OT_INTENT_LOG);
1675 		BP_SET_LEVEL(bp, 0);
1676 		BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1677 		bp->blk_birth = txg;
1678 	}
1679 
1680 	spa_config_exit(spa);
1681 
1682 	return (error);
1683 }
1684 
1685 /*
1686  * Free an intent log block.  We know it can't be a gang block, so there's
1687  * nothing to do except metaslab_free() it.
1688  */
1689 void
1690 zio_free_blk(spa_t *spa, blkptr_t *bp, uint64_t txg)
1691 {
1692 	ASSERT(DVA_GET_GANG(BP_IDENTITY(bp)) == 0);
1693 
1694 	dprintf_bp(bp, "txg %llu: ", txg);
1695 
1696 	spa_config_enter(spa, RW_READER);
1697 
1698 	metaslab_free(spa, BP_IDENTITY(bp), txg);
1699 
1700 	spa_config_exit(spa);
1701 }
1702