fs/zfs/zio.c

/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License (the "License").
 * You may not use this file except in compliance with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */
/*
 * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

#pragma ident	"@(#)zio.c	1.31	07/12/12 SMI"

#include <sys/zfs_context.h>
#include <sys/fm/fs/zfs.h>
#include <sys/spa.h>
#include <sys/txg.h>
#include <sys/spa_impl.h>
#include <sys/vdev_impl.h>
#include <sys/zio_impl.h>
#include <sys/zio_compress.h>
#include <sys/zio_checksum.h>

/*
 * ==========================================================================
 * I/O priority table
 * ==========================================================================
 */
uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
	0,	/* ZIO_PRIORITY_NOW		*/
	0,	/* ZIO_PRIORITY_SYNC_READ	*/
	0,	/* ZIO_PRIORITY_SYNC_WRITE	*/
	6,	/* ZIO_PRIORITY_ASYNC_READ	*/
	4,	/* ZIO_PRIORITY_ASYNC_WRITE	*/
	4,	/* ZIO_PRIORITY_FREE		*/
	0,	/* ZIO_PRIORITY_CACHE_FILL	*/
	0,	/* ZIO_PRIORITY_LOG_WRITE	*/
	10,	/* ZIO_PRIORITY_RESILVER	*/
	20,	/* ZIO_PRIORITY_SCRUB		*/
};

/*
 * ==========================================================================
 * I/O type descriptions
 * ==========================================================================
 */
char *zio_type_name[ZIO_TYPES] = {
	"null", "read", "write", "free", "claim", "ioctl" };

/* Force an allocation failure when non-zero */
uint16_t zio_zil_fail_shift = 0;
uint16_t zio_io_fail_shift = 0;

/* Enable/disable the write-retry logic */
int zio_write_retry = 1;

/* Taskq to handle reissuing of I/Os */
taskq_t *zio_taskq;
int zio_resume_threads = 4;

typedef struct zio_sync_pass {
	int	zp_defer_free;		/* defer frees after this pass */
	int	zp_dontcompress;	/* don't compress after this pass */
	int	zp_rewrite;		/* rewrite new bps after this pass */
} zio_sync_pass_t;

zio_sync_pass_t zio_sync_pass = {
	1,	/* zp_defer_free */
	4,	/* zp_dontcompress */
	1,	/* zp_rewrite */
};

static boolean_t zio_io_should_fail(uint16_t);

/*
 * ==========================================================================
 * I/O kmem caches
 * ==========================================================================
 */
kmem_cache_t *zio_cache;
kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];

#ifdef _KERNEL
extern vmem_t *zio_alloc_arena;
#endif

/*
 * Determine if we are allowed to issue the IO based on the
 * pool state. If we must wait then block until we are told
 * that we may continue.
 */
#define	ZIO_ENTER(spa) {						\
	if (spa->spa_state == POOL_STATE_IO_FAILURE) {			\
		mutex_enter(&spa->spa_zio_lock);			\
		while (spa->spa_state == POOL_STATE_IO_FAILURE)		\
			cv_wait(&spa->spa_zio_cv, &spa->spa_zio_lock);	\
		mutex_exit(&spa->spa_zio_lock);				\
	}								\
}

/*
 * An allocation zio is one that either currently has the DVA allocate
 * stage set or will have it later in it's lifetime.
 */
#define	IO_IS_ALLOCATING(zio) \
	((zio)->io_orig_pipeline & (1U << ZIO_STAGE_DVA_ALLOCATE))

void
zio_init(void)
{
	size_t c;
	vmem_t *data_alloc_arena = NULL;

#ifdef _KERNEL
	data_alloc_arena = zio_alloc_arena;
#endif

	zio_cache = kmem_cache_create("zio_cache", sizeof (zio_t), 0,
	    NULL, NULL, NULL, NULL, NULL, 0);

	/*
	 * For small buffers, we want a cache for each multiple of
	 * SPA_MINBLOCKSIZE.  For medium-size buffers, we want a cache
	 * for each quarter-power of 2.  For large buffers, we want
	 * a cache for each multiple of PAGESIZE.
	 */
	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
		size_t p2 = size;
		size_t align = 0;

		while (p2 & (p2 - 1))
			p2 &= p2 - 1;

		if (size <= 4 * SPA_MINBLOCKSIZE) {
			align = SPA_MINBLOCKSIZE;
		} else if (P2PHASE(size, PAGESIZE) == 0) {
			align = PAGESIZE;
		} else if (P2PHASE(size, p2 >> 2) == 0) {
			align = p2 >> 2;
		}

		if (align != 0) {
			char name[36];
			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
			zio_buf_cache[c] = kmem_cache_create(name, size,
			    align, NULL, NULL, NULL, NULL, NULL, KMC_NODEBUG);

			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
			zio_data_buf_cache[c] = kmem_cache_create(name, size,
			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
			    KMC_NODEBUG);

		}
	}

	while (--c != 0) {
		ASSERT(zio_buf_cache[c] != NULL);
		if (zio_buf_cache[c - 1] == NULL)
			zio_buf_cache[c - 1] = zio_buf_cache[c];

		ASSERT(zio_data_buf_cache[c] != NULL);
		if (zio_data_buf_cache[c - 1] == NULL)
			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
	}

	zio_taskq = taskq_create("zio_taskq", zio_resume_threads,
	    maxclsyspri, 50, INT_MAX, TASKQ_PREPOPULATE);

	zio_inject_init();
}

void
zio_fini(void)
{
	size_t c;
	kmem_cache_t *last_cache = NULL;
	kmem_cache_t *last_data_cache = NULL;

	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
		if (zio_buf_cache[c] != last_cache) {
			last_cache = zio_buf_cache[c];
			kmem_cache_destroy(zio_buf_cache[c]);
		}
		zio_buf_cache[c] = NULL;

		if (zio_data_buf_cache[c] != last_data_cache) {
			last_data_cache = zio_data_buf_cache[c];
			kmem_cache_destroy(zio_data_buf_cache[c]);
		}
		zio_data_buf_cache[c] = NULL;
	}

	taskq_destroy(zio_taskq);

	kmem_cache_destroy(zio_cache);

	zio_inject_fini();
}

/*
 * ==========================================================================
 * Allocate and free I/O buffers
 * ==========================================================================
 */

/*
 * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
 * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
 * excess / transient data in-core during a crashdump.
 */
void *
zio_buf_alloc(size_t size)
{
	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;

	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);

	return (kmem_cache_alloc(zio_buf_cache[c], KM_SLEEP));
}

/*
 * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
 * crashdump if the kernel panics.  This exists so that we will limit the amount
 * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
 * of kernel heap dumped to disk when the kernel panics)
 */
void *
zio_data_buf_alloc(size_t size)
{
	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;

	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);

	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_SLEEP));
}

void
zio_buf_free(void *buf, size_t size)
{
	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;

	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);

	kmem_cache_free(zio_buf_cache[c], buf);
}

void
zio_data_buf_free(void *buf, size_t size)
{
	size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;

	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);

	kmem_cache_free(zio_data_buf_cache[c], buf);
}

/*
 * ==========================================================================
 * Push and pop I/O transform buffers
 * ==========================================================================
 */
static void
zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize)
{
	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);

	zt->zt_data = data;
	zt->zt_size = size;
	zt->zt_bufsize = bufsize;

	zt->zt_next = zio->io_transform_stack;
	zio->io_transform_stack = zt;

	zio->io_data = data;
	zio->io_size = size;
}

static void
zio_pop_transform(zio_t *zio, void **data, uint64_t *size, uint64_t *bufsize)
{
	zio_transform_t *zt = zio->io_transform_stack;

	*data = zt->zt_data;
	*size = zt->zt_size;
	*bufsize = zt->zt_bufsize;

	zio->io_transform_stack = zt->zt_next;
	kmem_free(zt, sizeof (zio_transform_t));

	if ((zt = zio->io_transform_stack) != NULL) {
		zio->io_data = zt->zt_data;
		zio->io_size = zt->zt_size;
	}
}

static void
zio_clear_transform_stack(zio_t *zio)
{
	void *data;
	uint64_t size, bufsize;

	ASSERT(zio->io_transform_stack != NULL);

	zio_pop_transform(zio, &data, &size, &bufsize);
	while (zio->io_transform_stack != NULL) {
		zio_buf_free(data, bufsize);
		zio_pop_transform(zio, &data, &size, &bufsize);
	}
}

/*
 * ==========================================================================
 * Create the various types of I/O (read, write, free)
 * ==========================================================================
 */
static zio_t *
zio_create(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
    void *data, uint64_t size, zio_done_func_t *done, void *private,
    zio_type_t type, int priority, int flags, uint8_t stage, uint32_t pipeline)
{
	zio_t *zio;

	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);

	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
	bzero(zio, sizeof (zio_t));
	zio->io_parent = pio;
	zio->io_spa = spa;
	zio->io_txg = txg;
	zio->io_flags = flags;
	if (bp != NULL) {
		zio->io_bp = bp;
		zio->io_bp_copy = *bp;
		zio->io_bp_orig = *bp;
	}
	zio->io_done = done;
	zio->io_private = private;
	zio->io_type = type;
	zio->io_priority = priority;
	zio->io_stage = stage;
	zio->io_pipeline = pipeline;
	zio->io_timestamp = lbolt64;
	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
	zio_push_transform(zio, data, size, size);

	/*
	 * Note on config lock:
	 *
	 * If CONFIG_HELD is set, then the caller already has the config
	 * lock, so we don't need it for this io.
	 *
	 * We set CONFIG_GRABBED to indicate that we have grabbed the
	 * config lock on behalf of this io, so it should be released
	 * in zio_done.
	 *
	 * Unless CONFIG_HELD is set, we will grab the config lock for
	 * any top-level (parent-less) io, *except* NULL top-level ios.
	 * The NULL top-level ios rarely have any children, so we delay
	 * grabbing the lock until the first child is added (but it is
	 * still grabbed on behalf of the top-level i/o, so additional
	 * children don't need to also grab it).  This greatly reduces
	 * contention on the config lock.
	 */
	if (pio == NULL) {
		if (type != ZIO_TYPE_NULL &&
		    !(flags & ZIO_FLAG_CONFIG_HELD)) {
			spa_config_enter(spa, RW_READER, zio);
			zio->io_flags |= ZIO_FLAG_CONFIG_GRABBED;
		}
		zio->io_root = zio;
	} else {
		zio->io_root = pio->io_root;
		if (!(flags & ZIO_FLAG_NOBOOKMARK))
			zio->io_logical = pio->io_logical;
		mutex_enter(&pio->io_lock);
		if (pio->io_parent == NULL &&
		    pio->io_type == ZIO_TYPE_NULL &&
		    !(pio->io_flags & ZIO_FLAG_CONFIG_GRABBED) &&
		    !(pio->io_flags & ZIO_FLAG_CONFIG_HELD)) {
			pio->io_flags |= ZIO_FLAG_CONFIG_GRABBED;
			spa_config_enter(spa, RW_READER, pio);
		}
		if (stage < ZIO_STAGE_READY)
			pio->io_children_notready++;
		pio->io_children_notdone++;
		zio->io_sibling_next = pio->io_child;
		zio->io_sibling_prev = NULL;
		if (pio->io_child != NULL)
			pio->io_child->io_sibling_prev = zio;
		pio->io_child = zio;
		zio->io_ndvas = pio->io_ndvas;
		mutex_exit(&pio->io_lock);
	}

	/*
	 * Save off the original state incase we need to retry later.
	 */
	zio->io_orig_stage = zio->io_stage;
	zio->io_orig_pipeline = zio->io_pipeline;
	zio->io_orig_flags = zio->io_flags;

	return (zio);
}

static void
zio_reset(zio_t *zio)
{
	zio_clear_transform_stack(zio);

	zio->io_flags = zio->io_orig_flags;
	zio->io_stage = zio->io_orig_stage;
	zio->io_pipeline = zio->io_orig_pipeline;
	zio_push_transform(zio, zio->io_data, zio->io_size, zio->io_size);
}

zio_t *
zio_null(zio_t *pio, spa_t *spa, zio_done_func_t *done, void *private,
	int flags)
{
	zio_t *zio;

	zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
	    ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, ZIO_STAGE_OPEN,
	    ZIO_WAIT_FOR_CHILDREN_PIPELINE);

	return (zio);
}

zio_t *
zio_root(spa_t *spa, zio_done_func_t *done, void *private, int flags)
{
	return (zio_null(NULL, spa, done, private, flags));
}

zio_t *
zio_read(zio_t *pio, spa_t *spa, blkptr_t *bp, void *data,
    uint64_t size, zio_done_func_t *done, void *private,
    int priority, int flags, zbookmark_t *zb)
{
	zio_t *zio;

	ASSERT3U(size, ==, BP_GET_LSIZE(bp));

	/*
	 * If the user has specified that we allow I/Os to continue
	 * then attempt to satisfy the read.
	 */
	if (spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
		ZIO_ENTER(spa);

	zio = zio_create(pio, spa, bp->blk_birth, bp, data, size, done, private,
	    ZIO_TYPE_READ, priority, flags | ZIO_FLAG_USER,
	    ZIO_STAGE_OPEN, ZIO_READ_PIPELINE);
	zio->io_bookmark = *zb;

	zio->io_logical = zio;

	/*
	 * Work off our copy of the bp so the caller can free it.
	 */
	zio->io_bp = &zio->io_bp_copy;

	return (zio);
}

zio_t *
zio_write(zio_t *pio, spa_t *spa, int checksum, int compress, int ncopies,
    uint64_t txg, blkptr_t *bp, void *data, uint64_t size,
    zio_done_func_t *ready, zio_done_func_t *done, void *private, int priority,
    int flags, zbookmark_t *zb)
{
	zio_t *zio;

	ASSERT(checksum >= ZIO_CHECKSUM_OFF &&
	    checksum < ZIO_CHECKSUM_FUNCTIONS);

	ASSERT(compress >= ZIO_COMPRESS_OFF &&
	    compress < ZIO_COMPRESS_FUNCTIONS);

	ZIO_ENTER(spa);

	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_USER,
	    ZIO_STAGE_OPEN, ZIO_WRITE_PIPELINE);

	zio->io_ready = ready;

	zio->io_bookmark = *zb;

	zio->io_logical = zio;

	zio->io_checksum = checksum;
	zio->io_compress = compress;
	zio->io_ndvas = ncopies;

	if (bp->blk_birth != txg) {
		/* XXX the bp usually (always?) gets re-zeroed later */
		BP_ZERO(bp);
		BP_SET_LSIZE(bp, size);
		BP_SET_PSIZE(bp, size);
	} else {
		/* Make sure someone doesn't change their mind on overwrites */
		ASSERT(MIN(zio->io_ndvas + BP_IS_GANG(bp),
		    spa_max_replication(spa)) == BP_GET_NDVAS(bp));
	}

	return (zio);
}

zio_t *
zio_rewrite(zio_t *pio, spa_t *spa, int checksum,
    uint64_t txg, blkptr_t *bp, void *data, uint64_t size,
    zio_done_func_t *done, void *private, int priority, int flags,
    zbookmark_t *zb)
{
	zio_t *zio;

	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_USER,
	    ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE(bp));

	zio->io_bookmark = *zb;
	zio->io_checksum = checksum;
	zio->io_compress = ZIO_COMPRESS_OFF;

	if (pio != NULL)
		ASSERT3U(zio->io_ndvas, <=, BP_GET_NDVAS(bp));

	return (zio);
}

static void
zio_write_allocate_ready(zio_t *zio)
{
	/* Free up the previous block */
	if (!BP_IS_HOLE(&zio->io_bp_orig)) {
		zio_nowait(zio_free(zio, zio->io_spa, zio->io_txg,
		    &zio->io_bp_orig, NULL, NULL));
	}
}

static zio_t *
zio_write_allocate(zio_t *pio, spa_t *spa, int checksum,
    uint64_t txg, blkptr_t *bp, void *data, uint64_t size,
    zio_done_func_t *done, void *private, int priority, int flags)
{
	zio_t *zio;

	BP_ZERO(bp);
	BP_SET_LSIZE(bp, size);
	BP_SET_PSIZE(bp, size);
	BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF);

	zio = zio_create(pio, spa, txg, bp, data, size, done, private,
	    ZIO_TYPE_WRITE, priority, flags,
	    ZIO_STAGE_OPEN, ZIO_WRITE_ALLOCATE_PIPELINE);

	zio->io_checksum = checksum;
	zio->io_compress = ZIO_COMPRESS_OFF;
	zio->io_ready = zio_write_allocate_ready;

	return (zio);
}

zio_t *
zio_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
    zio_done_func_t *done, void *private)
{
	zio_t *zio;

	ASSERT(!BP_IS_HOLE(bp));

	if (txg == spa->spa_syncing_txg &&
	    spa->spa_sync_pass > zio_sync_pass.zp_defer_free) {
		bplist_enqueue_deferred(&spa->spa_sync_bplist, bp);
		return (zio_null(pio, spa, NULL, NULL, 0));
	}

	zio = zio_create(pio, spa, txg, bp, NULL, 0, done, private,
	    ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, ZIO_FLAG_USER,
	    ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE(bp));

	zio->io_bp = &zio->io_bp_copy;

	return (zio);
}

zio_t *
zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
    zio_done_func_t *done, void *private)
{
	zio_t *zio;

	/*
	 * A claim is an allocation of a specific block.  Claims are needed
	 * to support immediate writes in the intent log.  The issue is that
	 * immediate writes contain committed data, but in a txg that was
	 * *not* committed.  Upon opening the pool after an unclean shutdown,
	 * the intent log claims all blocks that contain immediate write data
	 * so that the SPA knows they're in use.
	 *
	 * All claims *must* be resolved in the first txg -- before the SPA
	 * starts allocating blocks -- so that nothing is allocated twice.
	 */
	ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
	ASSERT3U(spa_first_txg(spa), <=, txg);

	zio = zio_create(pio, spa, txg, bp, NULL, 0, done, private,
	    ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, 0,
	    ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE(bp));

	zio->io_bp = &zio->io_bp_copy;

	return (zio);
}

zio_t *
zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
    zio_done_func_t *done, void *private, int priority, int flags)
{
	zio_t *zio;
	int c;

	if (vd->vdev_children == 0) {
		zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
		    ZIO_TYPE_IOCTL, priority, flags,
		    ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);

		zio->io_vd = vd;
		zio->io_cmd = cmd;
	} else {
		zio = zio_null(pio, spa, NULL, NULL, flags);

		for (c = 0; c < vd->vdev_children; c++)
			zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
			    done, private, priority, flags));
	}

	return (zio);
}

static void
zio_phys_bp_init(vdev_t *vd, blkptr_t *bp, uint64_t offset, uint64_t size,
    int checksum, boolean_t labels)
{
	ASSERT(vd->vdev_children == 0);

	ASSERT(size <= SPA_MAXBLOCKSIZE);
	ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
	ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);

#ifdef ZFS_DEBUG
	if (labels) {
		ASSERT(offset + size <= VDEV_LABEL_START_SIZE ||
		    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
	}
#endif
	ASSERT3U(offset + size, <=, vd->vdev_psize);

	BP_ZERO(bp);

	BP_SET_LSIZE(bp, size);
	BP_SET_PSIZE(bp, size);

	BP_SET_CHECKSUM(bp, checksum);
	BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF);
	BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);

	if (checksum != ZIO_CHECKSUM_OFF)
		ZIO_SET_CHECKSUM(&bp->blk_cksum, offset, 0, 0, 0);
}

zio_t *
zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
    void *data, int checksum, zio_done_func_t *done, void *private,
    int priority, int flags, boolean_t labels)
{
	zio_t *zio;
	blkptr_t blk;

	ZIO_ENTER(vd->vdev_spa);

	zio_phys_bp_init(vd, &blk, offset, size, checksum, labels);

	zio = zio_create(pio, vd->vdev_spa, 0, &blk, data, size, done, private,
	    ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL,
	    ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);

	zio->io_vd = vd;
	zio->io_offset = offset;

	/*
	 * Work off our copy of the bp so the caller can free it.
	 */
	zio->io_bp = &zio->io_bp_copy;

	return (zio);
}

zio_t *
zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
    void *data, int checksum, zio_done_func_t *done, void *private,
    int priority, int flags, boolean_t labels)
{
	zio_block_tail_t *zbt;
	void *wbuf;
	zio_t *zio;
	blkptr_t blk;

	ZIO_ENTER(vd->vdev_spa);

	zio_phys_bp_init(vd, &blk, offset, size, checksum, labels);

	zio = zio_create(pio, vd->vdev_spa, 0, &blk, data, size, done, private,
	    ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL,
	    ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);

	zio->io_vd = vd;
	zio->io_offset = offset;

	zio->io_bp = &zio->io_bp_copy;
	zio->io_checksum = checksum;

	if (zio_checksum_table[checksum].ci_zbt) {
		/*
		 * zbt checksums are necessarily destructive -- they modify
		 * one word of the write buffer to hold the verifier/checksum.
		 * Therefore, we must make a local copy in case the data is
		 * being written to multiple places.
		 */
		wbuf = zio_buf_alloc(size);
		bcopy(data, wbuf, size);
		zio_push_transform(zio, wbuf, size, size);

		zbt = (zio_block_tail_t *)((char *)wbuf + size) - 1;
		zbt->zbt_cksum = blk.blk_cksum;
	}

	return (zio);
}

/*
 * Create a child I/O to do some work for us.  It has no associated bp.
 */
zio_t *
zio_vdev_child_io(zio_t *zio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
	void *data, uint64_t size, int type, int priority, int flags,
	zio_done_func_t *done, void *private)
{
	uint32_t pipeline = ZIO_VDEV_CHILD_PIPELINE;
	zio_t *cio;

	if (type == ZIO_TYPE_READ && bp != NULL) {
		/*
		 * If we have the bp, then the child should perform the
		 * checksum and the parent need not.  This pushes error
		 * detection as close to the leaves as possible and
		 * eliminates redundant checksums in the interior nodes.
		 */
		pipeline |= 1U << ZIO_STAGE_CHECKSUM_VERIFY;
		zio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY);
	}

	cio = zio_create(zio, zio->io_spa, zio->io_txg, bp, data, size,
	    done, private, type, priority,
	    (zio->io_flags & ZIO_FLAG_VDEV_INHERIT) | ZIO_FLAG_CANFAIL | flags,
	    ZIO_STAGE_VDEV_IO_START - 1, pipeline);

	cio->io_vd = vd;
	cio->io_offset = offset;

	return (cio);
}

/*
 * ==========================================================================
 * Initiate I/O, either sync or async
 * ==========================================================================
 */
int
zio_wait(zio_t *zio)
{
	int error;

	ASSERT(zio->io_stage == ZIO_STAGE_OPEN);

	zio->io_waiter = curthread;

	zio_execute(zio);

	mutex_enter(&zio->io_lock);
	while (zio->io_stalled != ZIO_STAGE_DONE)
		cv_wait(&zio->io_cv, &zio->io_lock);
	mutex_exit(&zio->io_lock);

	error = zio->io_error;
	mutex_destroy(&zio->io_lock);
	cv_destroy(&zio->io_cv);
	kmem_cache_free(zio_cache, zio);

	return (error);
}

void
zio_nowait(zio_t *zio)
{
	zio_execute(zio);
}

void
zio_interrupt(zio_t *zio)
{
	(void) taskq_dispatch(zio->io_spa->spa_zio_intr_taskq[zio->io_type],
	    (task_func_t *)zio_execute, zio, TQ_SLEEP);
}

static int
zio_issue_async(zio_t *zio)
{
	(void) taskq_dispatch(zio->io_spa->spa_zio_issue_taskq[zio->io_type],
	    (task_func_t *)zio_execute, zio, TQ_SLEEP);

	return (ZIO_PIPELINE_STOP);
}

/*
 * ==========================================================================
 * I/O pipeline interlocks: parent/child dependency scoreboarding
 * ==========================================================================
 */
static int
zio_wait_for_children(zio_t *zio, uint32_t stage, uint64_t *countp)
{
	int rv = ZIO_PIPELINE_CONTINUE;

	mutex_enter(&zio->io_lock);
	ASSERT(zio->io_stalled == 0);
	if (*countp != 0) {
		zio->io_stalled = stage;
		rv = ZIO_PIPELINE_STOP;
	}
	mutex_exit(&zio->io_lock);

	return (rv);
}

static void
zio_notify_parent(zio_t *zio, uint32_t stage, uint64_t *countp)
{
	zio_t *pio = zio->io_parent;

	mutex_enter(&pio->io_lock);
	if (pio->io_error == 0 && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
		pio->io_error = zio->io_error;
	ASSERT3U(*countp, >, 0);
	if (--*countp == 0 && pio->io_stalled == stage) {
		pio->io_stalled = 0;
		mutex_exit(&pio->io_lock);
		zio_execute(pio);
	} else {
		mutex_exit(&pio->io_lock);
	}
}

int
zio_wait_for_children_ready(zio_t *zio)
{
	return (zio_wait_for_children(zio, ZIO_STAGE_WAIT_FOR_CHILDREN_READY,
	    &zio->io_children_notready));
}

int
zio_wait_for_children_done(zio_t *zio)
{
	return (zio_wait_for_children(zio, ZIO_STAGE_WAIT_FOR_CHILDREN_DONE,
	    &zio->io_children_notdone));
}

static int
zio_read_init(zio_t *zio)
{
	blkptr_t *bp = zio->io_bp;

	if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
		uint64_t csize = BP_GET_PSIZE(bp);
		void *cbuf = zio_buf_alloc(csize);

		zio_push_transform(zio, cbuf, csize, csize);
		zio->io_pipeline |= 1U << ZIO_STAGE_READ_DECOMPRESS;
	}

	if (BP_IS_GANG(bp)) {
		uint64_t gsize = SPA_GANGBLOCKSIZE;
		void *gbuf = zio_buf_alloc(gsize);

		zio_push_transform(zio, gbuf, gsize, gsize);
		zio->io_pipeline |= 1U << ZIO_STAGE_READ_GANG_MEMBERS;
	}

	if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
		zio->io_flags |= ZIO_FLAG_DONT_CACHE;

	return (ZIO_PIPELINE_CONTINUE);
}

static int
zio_ready(zio_t *zio)
{
	zio_t *pio = zio->io_parent;

	if (zio->io_ready)
		zio->io_ready(zio);

	if (pio != NULL)
		zio_notify_parent(zio, ZIO_STAGE_WAIT_FOR_CHILDREN_READY,
		    &pio->io_children_notready);

	if (zio->io_bp)
		zio->io_bp_copy = *zio->io_bp;

	return (ZIO_PIPELINE_CONTINUE);
}

static int
zio_vdev_retry_io(zio_t *zio)
{
	zio_t *pio = zio->io_parent;

	/*
	 * Preserve the failed bp so that the io_ready() callback can
	 * update the accounting accordingly. The callback will also be
	 * responsible for freeing the previously allocated block, if one
	 * exists.
	 */
	zio->io_bp_orig = *zio->io_bp;

	/*
	 * We must zero out the old DVA and blk_birth before reallocating
	 * the bp.
	 */
	BP_ZERO_DVAS(zio->io_bp);
	zio_reset(zio);

	if (pio) {
		/*
		 * Let the parent know that we will
		 * re-alloc the write (=> new bp info).
		 */
		mutex_enter(&pio->io_lock);
		pio->io_children_notready++;

		/*
		 * If the parent I/O is still in the open stage, then
		 * don't bother telling it to retry since it hasn't
		 * progressed far enough for it to care.
		 */
		if (pio->io_stage > ZIO_STAGE_OPEN && IO_IS_ALLOCATING(pio))
			pio->io_flags |= ZIO_FLAG_WRITE_RETRY;

		ASSERT(pio->io_stage <= ZIO_STAGE_WAIT_FOR_CHILDREN_DONE);
		mutex_exit(&pio->io_lock);
	}

	/*
	 * We are getting ready to process the retry request so clear
	 * the flag and the zio's current error status.
	 */
	zio->io_flags &= ~ZIO_FLAG_WRITE_RETRY;
	zio->io_error = 0;

	return (ZIO_PIPELINE_CONTINUE);
}

int
zio_vdev_resume_io(spa_t *spa)
{
	zio_t *zio;

	mutex_enter(&spa->spa_zio_lock);

	/*
	 * Probe all of vdevs that have experienced an I/O error.
	 * If we are still unable to verify the integrity of the vdev
	 * then we prevent the resume from proceeeding.
	 */
	for (zio = list_head(&spa->spa_zio_list); zio != NULL;
	    zio = list_next(&spa->spa_zio_list, zio)) {
		int error = 0;