xref: /titanic_52/usr/src/uts/common/fs/zfs/zio.c (revision fa9e4066f08beec538e775443c5be79dd423fcab)
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, 0);
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[300];
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 ? bp : &zio->io_bp_copy);
752 		dprintf("ZFS: %s (%s on %s off %llx: zio %p %s): error %d\n",
753 		    zio->io_error == ECKSUM ? "bad checksum" : "I/O failure",
754 		    zio_type_name[zio->io_type],
755 		    vdev_description(vd),
756 		    (u_longlong_t)zio->io_offset,
757 		    zio, blkbuf, zio->io_error);
758 	}
759 
760 	if (zio->io_numerrors != 0 && zio->io_type == ZIO_TYPE_WRITE) {
761 		sprintf_blkptr(blkbuf, bp ? bp : &zio->io_bp_copy);
762 		dprintf("ZFS: %s (%s on %s off %llx: zio %p %s): %d errors\n",
763 		    "partial write",
764 		    zio_type_name[zio->io_type],
765 		    vdev_description(vd),
766 		    (u_longlong_t)zio->io_offset,
767 		    zio, blkbuf, zio->io_numerrors);
768 	}
769 
770 	if (zio->io_error && !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
771 		sprintf_blkptr(blkbuf, bp ? bp : &zio->io_bp_copy);
772 		panic("ZFS: %s (%s on %s off %llx: zio %p %s): error %d",
773 		    zio->io_error == ECKSUM ? "bad checksum" : "I/O failure",
774 		    zio_type_name[zio->io_type],
775 		    vdev_description(vd),
776 		    (u_longlong_t)zio->io_offset,
777 		    zio, blkbuf, zio->io_error);
778 	}
779 
780 	zio_clear_transform_stack(zio);
781 
782 	if (zio->io_done)
783 		zio->io_done(zio);
784 
785 	ASSERT(zio->io_delegate_list == NULL);
786 	ASSERT(zio->io_delegate_next == NULL);
787 
788 	if (pio != NULL) {
789 		zio_t *next, *prev;
790 
791 		mutex_enter(&pio->io_lock);
792 		next = zio->io_sibling_next;
793 		prev = zio->io_sibling_prev;
794 		if (next != NULL)
795 			next->io_sibling_prev = prev;
796 		if (prev != NULL)
797 			prev->io_sibling_next = next;
798 		if (pio->io_child == zio)
799 			pio->io_child = next;
800 		mutex_exit(&pio->io_lock);
801 
802 		zio_notify_parent(zio, ZIO_STAGE_WAIT_CHILDREN_DONE,
803 		    &pio->io_children_notdone);
804 	}
805 
806 	if (pio == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_HELD))
807 		spa_config_exit(spa);
808 
809 	if (zio->io_waiter != NULL) {
810 		mutex_enter(&zio->io_lock);
811 		ASSERT(zio->io_stage == ZIO_STAGE_DONE);
812 		zio->io_stalled = zio->io_stage;
813 		cv_broadcast(&zio->io_cv);
814 		mutex_exit(&zio->io_lock);
815 	} else {
816 		kmem_free(zio, sizeof (zio_t));
817 	}
818 }
819 
820 /*
821  * ==========================================================================
822  * Compression support
823  * ==========================================================================
824  */
825 static void
826 zio_write_compress(zio_t *zio)
827 {
828 	int compress = zio->io_compress;
829 	blkptr_t *bp = zio->io_bp;
830 	void *cbuf;
831 	uint64_t lsize = zio->io_size;
832 	uint64_t csize = lsize;
833 	uint64_t cbufsize = 0;
834 	int pass;
835 
836 	if (bp->blk_birth == zio->io_txg) {
837 		/*
838 		 * We're rewriting an existing block, which means we're
839 		 * working on behalf of spa_sync().  For spa_sync() to
840 		 * converge, it must eventually be the case that we don't
841 		 * have to allocate new blocks.  But compression changes
842 		 * the blocksize, which forces a reallocate, and makes
843 		 * convergence take longer.  Therefore, after the first
844 		 * few passes, stop compressing to ensure convergence.
845 		 */
846 		pass = spa_sync_pass(zio->io_spa);
847 		if (pass > zio_sync_pass.zp_dontcompress)
848 			compress = ZIO_COMPRESS_OFF;
849 	} else {
850 		ASSERT(BP_IS_HOLE(bp));
851 		pass = 1;
852 	}
853 
854 	if (compress != ZIO_COMPRESS_OFF)
855 		if (!zio_compress_data(compress, zio->io_data, zio->io_size,
856 		    &cbuf, &csize, &cbufsize))
857 			compress = ZIO_COMPRESS_OFF;
858 
859 	if (compress != ZIO_COMPRESS_OFF && csize != 0)
860 		zio_push_transform(zio, cbuf, csize, cbufsize);
861 
862 	/*
863 	 * The final pass of spa_sync() must be all rewrites, but the first
864 	 * few passes offer a trade-off: allocating blocks defers convergence,
865 	 * but newly allocated blocks are sequential, so they can be written
866 	 * to disk faster.  Therefore, we allow the first few passes of
867 	 * spa_sync() to reallocate new blocks, but force rewrites after that.
868 	 * There should only be a handful of blocks after pass 1 in any case.
869 	 */
870 	if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == csize &&
871 	    pass > zio_sync_pass.zp_rewrite) {
872 		ASSERT(csize != 0);
873 		ASSERT3U(BP_GET_COMPRESS(bp), ==, compress);
874 		ASSERT3U(BP_GET_LSIZE(bp), ==, lsize);
875 
876 		zio->io_pipeline = ZIO_REWRITE_PIPELINE;
877 	} else {
878 		if (bp->blk_birth == zio->io_txg) {
879 			ASSERT3U(BP_GET_LSIZE(bp), ==, lsize);
880 			bzero(bp, sizeof (blkptr_t));
881 		}
882 		if (csize == 0) {
883 			BP_ZERO(bp);
884 			zio->io_pipeline = ZIO_WAIT_FOR_CHILDREN_PIPELINE;
885 		} else {
886 			BP_SET_LSIZE(bp, lsize);
887 			BP_SET_PSIZE(bp, csize);
888 			BP_SET_COMPRESS(bp, compress);
889 			zio->io_pipeline = ZIO_WRITE_ALLOCATE_PIPELINE;
890 		}
891 	}
892 
893 	zio_next_stage(zio);
894 }
895 
896 static void
897 zio_read_decompress(zio_t *zio)
898 {
899 	blkptr_t *bp = zio->io_bp;
900 	void *data;
901 	uint64_t size;
902 	uint64_t bufsize;
903 	int compress = BP_GET_COMPRESS(bp);
904 
905 	ASSERT(compress != ZIO_COMPRESS_OFF);
906 
907 	zio_pop_transform(zio, &data, &size, &bufsize);
908 
909 	if (zio_decompress_data(compress, data, size,
910 	    zio->io_data, zio->io_size))
911 		zio->io_error = EIO;
912 
913 	zio_buf_free(data, bufsize);
914 
915 	zio_next_stage(zio);
916 }
917 
918 /*
919  * ==========================================================================
920  * Gang block support
921  * ==========================================================================
922  */
923 static void
924 zio_gang_pipeline(zio_t *zio)
925 {
926 	/*
927 	 * By default, the pipeline assumes that we're dealing with a gang
928 	 * block.  If we're not, strip out any gang-specific stages.
929 	 */
930 	if (!DVA_GET_GANG(ZIO_GET_DVA(zio)))
931 		zio->io_pipeline &= ~ZIO_GANG_STAGES;
932 
933 	zio_next_stage(zio);
934 }
935 
936 static void
937 zio_gang_byteswap(zio_t *zio)
938 {
939 	ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
940 
941 	if (BP_SHOULD_BYTESWAP(zio->io_bp))
942 		byteswap_uint64_array(zio->io_data, zio->io_size);
943 }
944 
945 static void
946 zio_get_gang_header(zio_t *zio)
947 {
948 	blkptr_t *bp = zio->io_bp;
949 	uint64_t gsize = SPA_GANGBLOCKSIZE;
950 	void *gbuf = zio_buf_alloc(gsize);
951 
952 	ASSERT(DVA_GET_GANG(ZIO_GET_DVA(zio)));
953 
954 	zio_push_transform(zio, gbuf, gsize, gsize);
955 
956 	zio_nowait(zio_create(zio, zio->io_spa, bp->blk_birth, bp, gbuf, gsize,
957 	    NULL, NULL, ZIO_TYPE_READ, zio->io_priority,
958 	    zio->io_flags & ZIO_FLAG_GANG_INHERIT,
959 	    ZIO_STAGE_OPEN, ZIO_READ_PIPELINE));
960 
961 	zio_wait_children_done(zio);
962 }
963 
964 static void
965 zio_read_gang_members(zio_t *zio)
966 {
967 	zio_gbh_phys_t *gbh;
968 	uint64_t gsize, gbufsize, loff, lsize;
969 	int i;
970 
971 	ASSERT(DVA_GET_GANG(ZIO_GET_DVA(zio)));
972 
973 	zio_gang_byteswap(zio);
974 	zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize);
975 
976 	for (loff = 0, i = 0; loff != zio->io_size; loff += lsize, i++) {
977 		blkptr_t *gbp = &gbh->zg_blkptr[i];
978 		lsize = BP_GET_PSIZE(gbp);
979 
980 		ASSERT(BP_GET_COMPRESS(gbp) == ZIO_COMPRESS_OFF);
981 		ASSERT3U(lsize, ==, BP_GET_LSIZE(gbp));
982 		ASSERT3U(loff + lsize, <=, zio->io_size);
983 		ASSERT(i < SPA_GBH_NBLKPTRS);
984 		ASSERT(!BP_IS_HOLE(gbp));
985 
986 		zio_nowait(zio_read(zio, zio->io_spa, gbp,
987 		    (char *)zio->io_data + loff, lsize, NULL, NULL,
988 		    zio->io_priority, zio->io_flags & ZIO_FLAG_GANG_INHERIT));
989 	}
990 
991 	zio_buf_free(gbh, gbufsize);
992 	zio_wait_children_done(zio);
993 }
994 
995 static void
996 zio_rewrite_gang_members(zio_t *zio)
997 {
998 	zio_gbh_phys_t *gbh;
999 	uint64_t gsize, gbufsize, loff, lsize;
1000 	int i;
1001 
1002 	ASSERT(DVA_GET_GANG(ZIO_GET_DVA(zio)));
1003 	ASSERT3U(zio->io_size, ==, SPA_GANGBLOCKSIZE);
1004 
1005 	zio_gang_byteswap(zio);
1006 	zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize);
1007 
1008 	ASSERT(gsize == gbufsize);
1009 
1010 	for (loff = 0, i = 0; loff != zio->io_size; loff += lsize, i++) {
1011 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1012 		lsize = BP_GET_PSIZE(gbp);
1013 
1014 		ASSERT(BP_GET_COMPRESS(gbp) == ZIO_COMPRESS_OFF);
1015 		ASSERT3U(lsize, ==, BP_GET_LSIZE(gbp));
1016 		ASSERT3U(loff + lsize, <=, zio->io_size);
1017 		ASSERT(i < SPA_GBH_NBLKPTRS);
1018 		ASSERT(!BP_IS_HOLE(gbp));
1019 
1020 		zio_nowait(zio_rewrite(zio, zio->io_spa, zio->io_checksum,
1021 		    zio->io_txg, gbp, (char *)zio->io_data + loff, lsize,
1022 		    NULL, NULL, zio->io_priority, zio->io_flags));
1023 	}
1024 
1025 	zio_push_transform(zio, gbh, gsize, gbufsize);
1026 	zio_wait_children_ready(zio);
1027 }
1028 
1029 static void
1030 zio_free_gang_members(zio_t *zio)
1031 {
1032 	zio_gbh_phys_t *gbh;
1033 	uint64_t gsize, gbufsize;
1034 	int i;
1035 
1036 	ASSERT(DVA_GET_GANG(ZIO_GET_DVA(zio)));
1037 
1038 	zio_gang_byteswap(zio);
1039 	zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize);
1040 
1041 	for (i = 0; i < SPA_GBH_NBLKPTRS; i++) {
1042 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1043 
1044 		if (BP_IS_HOLE(gbp))
1045 			continue;
1046 		zio_nowait(zio_free(zio, zio->io_spa, zio->io_txg,
1047 		    gbp, NULL, NULL));
1048 	}
1049 
1050 	zio_buf_free(gbh, gbufsize);
1051 	zio_next_stage(zio);
1052 }
1053 
1054 static void
1055 zio_claim_gang_members(zio_t *zio)
1056 {
1057 	zio_gbh_phys_t *gbh;
1058 	uint64_t gsize, gbufsize;
1059 	int i;
1060 
1061 	ASSERT(DVA_GET_GANG(ZIO_GET_DVA(zio)));
1062 
1063 	zio_gang_byteswap(zio);
1064 	zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize);
1065 
1066 	for (i = 0; i < SPA_GBH_NBLKPTRS; i++) {
1067 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1068 		if (BP_IS_HOLE(gbp))
1069 			continue;
1070 		zio_nowait(zio_claim(zio, zio->io_spa, zio->io_txg,
1071 		    gbp, NULL, NULL));
1072 	}
1073 
1074 	zio_buf_free(gbh, gbufsize);
1075 	zio_next_stage(zio);
1076 }
1077 
1078 static void
1079 zio_write_allocate_gang_member_done(zio_t *zio)
1080 {
1081 	zio_t *pio = zio->io_parent;
1082 	dva_t *cdva = ZIO_GET_DVA(zio);
1083 	dva_t *pdva = ZIO_GET_DVA(pio);
1084 	uint64_t asize;
1085 
1086 	ASSERT(DVA_GET_GANG(pdva));
1087 
1088 	/* XXBP - Need to be careful here with multiple DVAs */
1089 	mutex_enter(&pio->io_lock);
1090 	asize = DVA_GET_ASIZE(pdva);
1091 	asize += DVA_GET_ASIZE(cdva);
1092 	DVA_SET_ASIZE(pdva, asize);
1093 	mutex_exit(&pio->io_lock);
1094 }
1095 
1096 static void
1097 zio_write_allocate_gang_members(zio_t *zio)
1098 {
1099 	blkptr_t *bp = zio->io_bp;
1100 	dva_t *dva = ZIO_GET_DVA(zio);
1101 	zio_gbh_phys_t *gbh;
1102 	uint64_t resid = zio->io_size;
1103 	uint64_t maxalloc = P2ROUNDUP(zio->io_size >> 1, SPA_MINBLOCKSIZE);
1104 	uint64_t gsize, loff, lsize;
1105 	uint32_t gbps_left;
1106 	int error;
1107 	int i;
1108 
1109 	gsize = SPA_GANGBLOCKSIZE;
1110 	gbps_left = SPA_GBH_NBLKPTRS;
1111 
1112 	error = metaslab_alloc(zio->io_spa, gsize, dva, zio->io_txg);
1113 	if (error == ENOSPC)
1114 		panic("can't allocate gang block header");
1115 	ASSERT(error == 0);
1116 
1117 	DVA_SET_GANG(dva, 1);
1118 
1119 	bp->blk_birth = zio->io_txg;
1120 
1121 	gbh = zio_buf_alloc(gsize);
1122 	bzero(gbh, gsize);
1123 
1124 	for (loff = 0, i = 0; loff != zio->io_size;
1125 	    loff += lsize, resid -= lsize, gbps_left--, i++) {
1126 		blkptr_t *gbp = &gbh->zg_blkptr[i];
1127 		dva = &gbp->blk_dva[0];
1128 
1129 		ASSERT(gbps_left != 0);
1130 		maxalloc = MIN(maxalloc, resid);
1131 
1132 		while (resid <= maxalloc * gbps_left) {
1133 			error = metaslab_alloc(zio->io_spa, maxalloc, dva,
1134 			    zio->io_txg);
1135 			if (error == 0)
1136 				break;
1137 			ASSERT3U(error, ==, ENOSPC);
1138 			if (maxalloc == SPA_MINBLOCKSIZE)
1139 				panic("really out of space");
1140 			maxalloc = P2ROUNDUP(maxalloc >> 1, SPA_MINBLOCKSIZE);
1141 		}
1142 
1143 		if (resid <= maxalloc * gbps_left) {
1144 			lsize = maxalloc;
1145 			BP_SET_LSIZE(gbp, lsize);
1146 			BP_SET_PSIZE(gbp, lsize);
1147 			BP_SET_COMPRESS(gbp, ZIO_COMPRESS_OFF);
1148 			gbp->blk_birth = zio->io_txg;
1149 			zio_nowait(zio_rewrite(zio, zio->io_spa,
1150 			    zio->io_checksum, zio->io_txg, gbp,
1151 			    (char *)zio->io_data + loff, lsize,
1152 			    zio_write_allocate_gang_member_done, NULL,
1153 			    zio->io_priority, zio->io_flags));
1154 		} else {
1155 			lsize = P2ROUNDUP(resid / gbps_left, SPA_MINBLOCKSIZE);
1156 			ASSERT(lsize != SPA_MINBLOCKSIZE);
1157 			zio_nowait(zio_write_allocate(zio, zio->io_spa,
1158 			    zio->io_checksum, zio->io_txg, gbp,
1159 			    (char *)zio->io_data + loff, lsize,
1160 			    zio_write_allocate_gang_member_done, NULL,
1161 			    zio->io_priority, zio->io_flags));
1162 		}
1163 	}
1164 
1165 	ASSERT(resid == 0 && loff == zio->io_size);
1166 
1167 	zio->io_pipeline |= 1U << ZIO_STAGE_GANG_CHECKSUM_GENERATE;
1168 
1169 	zio_push_transform(zio, gbh, gsize, gsize);
1170 	zio_wait_children_done(zio);
1171 }
1172 
1173 /*
1174  * ==========================================================================
1175  * Allocate and free blocks
1176  * ==========================================================================
1177  */
1178 static void
1179 zio_dva_allocate(zio_t *zio)
1180 {
1181 	blkptr_t *bp = zio->io_bp;
1182 	dva_t *dva = ZIO_GET_DVA(zio);
1183 	int error;
1184 
1185 	ASSERT(BP_IS_HOLE(bp));
1186 
1187 	/* For testing, make some blocks above a certain size be gang blocks */
1188 	if (zio->io_size >= zio_gang_bang && (lbolt & 0x3) == 0) {
1189 		zio_write_allocate_gang_members(zio);
1190 		return;
1191 	}
1192 
1193 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
1194 
1195 	error = metaslab_alloc(zio->io_spa, zio->io_size, dva, zio->io_txg);
1196 
1197 	if (error == 0) {
1198 		bp->blk_birth = zio->io_txg;
1199 	} else if (error == ENOSPC) {
1200 		if (zio->io_size == SPA_MINBLOCKSIZE)
1201 			panic("really, truly out of space");
1202 		zio_write_allocate_gang_members(zio);
1203 		return;
1204 	} else {
1205 		zio->io_error = error;
1206 	}
1207 	zio_next_stage(zio);
1208 }
1209 
1210 static void
1211 zio_dva_free(zio_t *zio)
1212 {
1213 	blkptr_t *bp = zio->io_bp;
1214 	dva_t *dva = ZIO_GET_DVA(zio);
1215 
1216 	ASSERT(!BP_IS_HOLE(bp));
1217 
1218 	metaslab_free(zio->io_spa, dva, zio->io_txg);
1219 
1220 	BP_ZERO(bp);
1221 
1222 	zio_next_stage(zio);
1223 }
1224 
1225 static void
1226 zio_dva_claim(zio_t *zio)
1227 {
1228 	blkptr_t *bp = zio->io_bp;
1229 	dva_t *dva = ZIO_GET_DVA(zio);
1230 
1231 	ASSERT(!BP_IS_HOLE(bp));
1232 
1233 	zio->io_error = metaslab_claim(zio->io_spa, dva, zio->io_txg);
1234 
1235 	zio_next_stage(zio);
1236 }
1237 
1238 static void
1239 zio_dva_translate(zio_t *zio)
1240 {
1241 	spa_t *spa = zio->io_spa;
1242 	dva_t *dva = ZIO_GET_DVA(zio);
1243 	uint64_t vdev = DVA_GET_VDEV(dva);
1244 	uint64_t offset = DVA_GET_OFFSET(dva);
1245 
1246 	ASSERT3U(zio->io_size, ==, ZIO_GET_IOSIZE(zio));
1247 
1248 	zio->io_offset = offset;
1249 
1250 	if ((zio->io_vd = vdev_lookup_top(spa, vdev)) == NULL)
1251 		zio->io_error = ENXIO;
1252 	else if (offset + zio->io_size > zio->io_vd->vdev_asize)
1253 		zio->io_error = EOVERFLOW;
1254 
1255 	zio_next_stage(zio);
1256 }
1257 
1258 /*
1259  * ==========================================================================
1260  * Read and write to physical devices
1261  * ==========================================================================
1262  */
1263 static void
1264 zio_vdev_io_enter(zio_t *zio)
1265 {
1266 	vdev_t *tvd = zio->io_vd->vdev_top;
1267 
1268 	mutex_enter(&tvd->vdev_io_lock);
1269 	ASSERT(zio->io_pending.list_next == NULL);
1270 	list_insert_tail(&tvd->vdev_io_pending, zio);
1271 	mutex_exit(&tvd->vdev_io_lock);
1272 }
1273 
1274 static void
1275 zio_vdev_io_exit(zio_t *zio)
1276 {
1277 	vdev_t *tvd = zio->io_vd->vdev_top;
1278 
1279 	mutex_enter(&tvd->vdev_io_lock);
1280 	ASSERT(zio->io_pending.list_next != NULL);
1281 	list_remove(&tvd->vdev_io_pending, zio);
1282 	if (list_head(&tvd->vdev_io_pending) == NULL)
1283 		cv_broadcast(&tvd->vdev_io_cv);
1284 	mutex_exit(&tvd->vdev_io_lock);
1285 }
1286 
1287 static void
1288 zio_vdev_io_retry(void *vdarg)
1289 {
1290 	vdev_t *vd = vdarg;
1291 	zio_t *zio, *zq;
1292 
1293 	ASSERT(vd == vd->vdev_top);
1294 
1295 	/* XXPOLICY */
1296 	delay(hz);
1297 
1298 	vdev_reopen(vd, &zq);
1299 
1300 	while ((zio = zq) != NULL) {
1301 		zq = zio->io_retry_next;
1302 		zio->io_retry_next = NULL;
1303 		dprintf("async retry #%d for I/O to %s offset %llx\n",
1304 		    zio->io_retries, vdev_description(vd), zio->io_offset);
1305 		zio_next_stage_async(zio);
1306 	}
1307 }
1308 
1309 static void
1310 zio_vdev_io_setup(zio_t *zio)
1311 {
1312 	vdev_t *vd = zio->io_vd;
1313 
1314 	/* XXPOLICY */
1315 	if (zio->io_retries == 0 && vd == vd->vdev_top)
1316 		zio->io_flags |= ZIO_FLAG_FAILFAST;
1317 
1318 	if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && vd->vdev_children == 0) {
1319 		zio->io_flags |= ZIO_FLAG_PHYSICAL;
1320 		zio->io_offset += VDEV_LABEL_START_SIZE;
1321 	}
1322 
1323 	zio_vdev_io_enter(zio);
1324 
1325 	zio_next_stage(zio);
1326 }
1327 
1328 static void
1329 zio_vdev_io_start(zio_t *zio)
1330 {
1331 	blkptr_t *bp = zio->io_bp;
1332 
1333 	ASSERT(P2PHASE(zio->io_offset, 1ULL << zio->io_vd->vdev_ashift) == 0);
1334 	ASSERT(P2PHASE(zio->io_size, 1ULL << zio->io_vd->vdev_ashift) == 0);
1335 	ASSERT(bp == NULL || ZIO_GET_IOSIZE(zio) == zio->io_size);
1336 	ASSERT(zio->io_type != ZIO_TYPE_WRITE || (spa_mode & FWRITE));
1337 
1338 	vdev_io_start(zio);
1339 
1340 	/* zio_next_stage_async() gets called from io completion interrupt */
1341 }
1342 
1343 static void
1344 zio_vdev_io_done(zio_t *zio)
1345 {
1346 	vdev_io_done(zio);
1347 }
1348 
1349 /* XXPOLICY */
1350 static boolean_t
1351 zio_should_retry(zio_t *zio)
1352 {
1353 	vdev_t *vd = zio->io_vd;
1354 
1355 	if (zio->io_error == 0)
1356 		return (B_FALSE);
1357 	if (zio->io_delegate_list != NULL)
1358 		return (B_FALSE);
1359 	if (vd != vd->vdev_top)
1360 		return (B_FALSE);
1361 	if (zio->io_flags & ZIO_FLAG_DONT_RETRY)
1362 		return (B_FALSE);
1363 	if (zio->io_retries > 300 &&
1364 	    (zio->io_flags & (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_CANFAIL)))
1365 		return (B_FALSE);
1366 	if (zio->io_retries > 1 &&
1367 	    (zio->io_error == ECKSUM || zio->io_error == ENXIO))
1368 		return (B_FALSE);
1369 
1370 	return (B_TRUE);
1371 }
1372 
1373 static void
1374 zio_vdev_io_assess(zio_t *zio)
1375 {
1376 	vdev_t *vd = zio->io_vd;
1377 	vdev_t *tvd = vd->vdev_top;
1378 
1379 	zio_vdev_io_exit(zio);
1380 
1381 	ASSERT(zio->io_vsd == NULL);
1382 
1383 	/*
1384 	 * If the I/O failed, determine whether we should attempt to retry it.
1385 	 */
1386 	/* XXPOLICY */
1387 	if (zio_should_retry(zio)) {
1388 		zio_t *zq;
1389 
1390 		ASSERT(tvd == vd);
1391 		ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE));
1392 
1393 		zio->io_retries++;
1394 		zio->io_error = 0;
1395 		zio->io_flags &= ZIO_FLAG_VDEV_INHERIT;
1396 		/* XXPOLICY */
1397 		zio->io_flags &= ~ZIO_FLAG_FAILFAST;
1398 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1399 		zio->io_stage = ZIO_STAGE_VDEV_IO_SETUP - 1;
1400 
1401 		dprintf("retry #%d for %s to %s offset %llx\n",
1402 		    zio->io_retries, zio_type_name[zio->io_type],
1403 		    vdev_description(vd), zio->io_offset);
1404 
1405 		/*
1406 		 * If this is the first retry, do it immediately.
1407 		 */
1408 		/* XXPOLICY */
1409 		if (zio->io_retries == 1) {
1410 			zio_next_stage_async(zio);
1411 			return;
1412 		}
1413 
1414 		/*
1415 		 * This was not the first retry, so go through the
1416 		 * longer enqueue/delay/vdev_reopen() process.
1417 		 */
1418 		mutex_enter(&tvd->vdev_io_lock);
1419 		ASSERT(zio->io_retry_next == NULL);
1420 		zio->io_retry_next = zq = tvd->vdev_io_retry;
1421 		tvd->vdev_io_retry = zio;
1422 		mutex_exit(&tvd->vdev_io_lock);
1423 		if (zq == NULL)
1424 			(void) taskq_dispatch(
1425 			    tvd->vdev_spa->spa_vdev_retry_taskq,
1426 			    zio_vdev_io_retry, tvd, TQ_SLEEP);
1427 		return;
1428 	}
1429 
1430 	zio_next_stage(zio);
1431 }
1432 
1433 void
1434 zio_vdev_io_reissue(zio_t *zio)
1435 {
1436 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
1437 	ASSERT(zio->io_error == 0);
1438 
1439 	zio->io_stage--;
1440 }
1441 
1442 void
1443 zio_vdev_io_redone(zio_t *zio)
1444 {
1445 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
1446 
1447 	zio->io_stage--;
1448 }
1449 
1450 void
1451 zio_vdev_io_bypass(zio_t *zio)
1452 {
1453 	ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
1454 	ASSERT(zio->io_error == 0);
1455 
1456 	zio->io_flags |= ZIO_FLAG_IO_BYPASS;
1457 	zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS - 1;
1458 }
1459 
1460 /*
1461  * ==========================================================================
1462  * Generate and verify checksums
1463  * ==========================================================================
1464  */
1465 static void
1466 zio_checksum_generate(zio_t *zio)
1467 {
1468 	int checksum = zio->io_checksum;
1469 	blkptr_t *bp = zio->io_bp;
1470 
1471 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
1472 
1473 	BP_SET_CHECKSUM(bp, checksum);
1474 	BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1475 
1476 	zio_checksum(checksum, &bp->blk_cksum, zio->io_data, zio->io_size);
1477 
1478 	zio_next_stage(zio);
1479 }
1480 
1481 static void
1482 zio_gang_checksum_generate(zio_t *zio)
1483 {
1484 	zio_cksum_t zc;
1485 	zio_gbh_phys_t *gbh = zio->io_data;
1486 
1487 	ASSERT3U(zio->io_size, ==, SPA_GANGBLOCKSIZE);
1488 	ASSERT(DVA_GET_GANG(ZIO_GET_DVA(zio)));
1489 
1490 	zio_set_gang_verifier(zio, &gbh->zg_tail.zbt_cksum);
1491 
1492 	zio_checksum(ZIO_CHECKSUM_GANG_HEADER, &zc, zio->io_data, zio->io_size);
1493 
1494 	zio_next_stage(zio);
1495 }
1496 
1497 static void
1498 zio_checksum_verify(zio_t *zio)
1499 {
1500 	if (zio->io_bp != NULL) {
1501 		zio->io_error = zio_checksum_error(zio);
1502 		if (zio->io_error) {
1503 			dprintf("bad checksum on vdev %s\n",
1504 			    vdev_description(zio->io_vd));
1505 		}
1506 	}
1507 
1508 	zio_next_stage(zio);
1509 }
1510 
1511 /*
1512  * Called by RAID-Z to ensure we don't compute the checksum twice.
1513  */
1514 void
1515 zio_checksum_verified(zio_t *zio)
1516 {
1517 	zio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY);
1518 }
1519 
1520 /*
1521  * Set the external verifier for a gang block based on stuff in the bp
1522  */
1523 void
1524 zio_set_gang_verifier(zio_t *zio, zio_cksum_t *zcp)
1525 {
1526 	zcp->zc_word[0] = DVA_GET_VDEV(ZIO_GET_DVA(zio));
1527 	zcp->zc_word[1] = DVA_GET_OFFSET(ZIO_GET_DVA(zio));
1528 	zcp->zc_word[2] = zio->io_bp->blk_birth;
1529 	zcp->zc_word[3] = 0;
1530 }
1531 
1532 /*
1533  * ==========================================================================
1534  * Define the pipeline
1535  * ==========================================================================
1536  */
1537 typedef void zio_pipe_stage_t(zio_t *zio);
1538 
1539 static void
1540 zio_badop(zio_t *zio)
1541 {
1542 	panic("Invalid I/O pipeline stage %u for zio %p", zio->io_stage, zio);
1543 }
1544 
1545 zio_pipe_stage_t *zio_pipeline[ZIO_STAGE_DONE + 2] = {
1546 	zio_badop,
1547 	zio_wait_children_ready,
1548 	zio_write_compress,
1549 	zio_checksum_generate,
1550 	zio_gang_pipeline,
1551 	zio_get_gang_header,
1552 	zio_rewrite_gang_members,
1553 	zio_free_gang_members,
1554 	zio_claim_gang_members,
1555 	zio_dva_allocate,
1556 	zio_dva_free,
1557 	zio_dva_claim,
1558 	zio_gang_checksum_generate,
1559 	zio_ready,
1560 	zio_dva_translate,
1561 	zio_vdev_io_setup,
1562 	zio_vdev_io_start,
1563 	zio_vdev_io_done,
1564 	zio_vdev_io_assess,
1565 	zio_wait_children_done,
1566 	zio_checksum_verify,
1567 	zio_read_gang_members,
1568 	zio_read_decompress,
1569 	zio_done,
1570 	zio_badop
1571 };
1572 
1573 /*
1574  * Move an I/O to the next stage of the pipeline and execute that stage.
1575  * There's no locking on io_stage because there's no legitimate way for
1576  * multiple threads to be attempting to process the same I/O.
1577  */
1578 void
1579 zio_next_stage(zio_t *zio)
1580 {
1581 	uint32_t pipeline = zio->io_pipeline;
1582 
1583 	ASSERT(!MUTEX_HELD(&zio->io_lock));
1584 
1585 	if (zio->io_error) {
1586 		dprintf("zio %p vdev %s offset %llx stage %d error %d\n",
1587 		    zio, vdev_description(zio->io_vd),
1588 		    zio->io_offset, zio->io_stage, zio->io_error);
1589 		if (((1U << zio->io_stage) & ZIO_VDEV_IO_PIPELINE) == 0)
1590 			pipeline &= ZIO_ERROR_PIPELINE_MASK;
1591 	}
1592 
1593 	while (((1U << ++zio->io_stage) & pipeline) == 0)
1594 		continue;
1595 
1596 	ASSERT(zio->io_stage <= ZIO_STAGE_DONE);
1597 	ASSERT(zio->io_stalled == 0);
1598 
1599 	zio_pipeline[zio->io_stage](zio);
1600 }
1601 
1602 void
1603 zio_next_stage_async(zio_t *zio)
1604 {
1605 	taskq_t *tq;
1606 	uint32_t pipeline = zio->io_pipeline;
1607 
1608 	ASSERT(!MUTEX_HELD(&zio->io_lock));
1609 
1610 	if (zio->io_error) {
1611 		dprintf("zio %p vdev %s offset %llx stage %d error %d\n",
1612 		    zio, vdev_description(zio->io_vd),
1613 		    zio->io_offset, zio->io_stage, zio->io_error);
1614 		if (((1U << zio->io_stage) & ZIO_VDEV_IO_PIPELINE) == 0)
1615 			pipeline &= ZIO_ERROR_PIPELINE_MASK;
1616 	}
1617 
1618 	while (((1U << ++zio->io_stage) & pipeline) == 0)
1619 		continue;
1620 
1621 	ASSERT(zio->io_stage <= ZIO_STAGE_DONE);
1622 	ASSERT(zio->io_stalled == 0);
1623 
1624 	/*
1625 	 * For performance, we'll probably want two sets of task queues:
1626 	 * per-CPU issue taskqs and per-CPU completion taskqs.  The per-CPU
1627 	 * part is for read performance: since we have to make a pass over
1628 	 * the data to checksum it anyway, we want to do this on the same CPU
1629 	 * that issued the read, because (assuming CPU scheduling affinity)
1630 	 * that thread is probably still there.  Getting this optimization
1631 	 * right avoids performance-hostile cache-to-cache transfers.
1632 	 *
1633 	 * Note that having two sets of task queues is also necessary for
1634 	 * correctness: if all of the issue threads get bogged down waiting
1635 	 * for dependent reads (e.g. metaslab freelist) to complete, then
1636 	 * there won't be any threads available to service I/O completion
1637 	 * interrupts.
1638 	 */
1639 	if ((1U << zio->io_stage) & zio->io_async_stages) {
1640 		if (zio->io_stage < ZIO_STAGE_VDEV_IO_DONE)
1641 			tq = zio->io_spa->spa_zio_issue_taskq[zio->io_type];
1642 		else
1643 			tq = zio->io_spa->spa_zio_intr_taskq[zio->io_type];
1644 		(void) taskq_dispatch(tq,
1645 		    (task_func_t *)zio_pipeline[zio->io_stage], zio, TQ_SLEEP);
1646 	} else {
1647 		zio_pipeline[zio->io_stage](zio);
1648 	}
1649 }
1650 
1651 /*
1652  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
1653  */
1654 int
1655 zio_alloc_blk(spa_t *spa, int checksum, uint64_t size, blkptr_t *bp,
1656     uint64_t txg)
1657 {
1658 	int error;
1659 
1660 	spa_config_enter(spa, RW_READER);
1661 
1662 	BP_ZERO(bp);
1663 
1664 	error = metaslab_alloc(spa, size, BP_IDENTITY(bp), txg);
1665 
1666 	if (error == 0) {
1667 		BP_SET_CHECKSUM(bp, checksum);
1668 		BP_SET_LSIZE(bp, size);
1669 		BP_SET_PSIZE(bp, size);
1670 		BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF);
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);
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(DVA_GET_GANG(BP_IDENTITY(bp)) == 0);
1690 
1691 	dprintf_bp(bp, "txg %llu: ", txg);
1692 
1693 	spa_config_enter(spa, RW_READER);
1694 
1695 	metaslab_free(spa, BP_IDENTITY(bp), txg);
1696 
1697 	spa_config_exit(spa);
1698 }
1699