/*-
 * Copyright (c) 2004 Pawel Jakub Dawidek <pjd@FreeBSD.org>
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 * 
 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
 * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 */

#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");

#include <sys/param.h>
#include <sys/systm.h>
#include <sys/kernel.h>
#include <sys/module.h>
#include <sys/lock.h>
#include <sys/mutex.h>
#include <sys/bio.h>
#include <sys/sysctl.h>
#include <sys/malloc.h>
#include <sys/bitstring.h>
#include <vm/uma.h>
#include <machine/atomic.h>
#include <geom/geom.h>
#include <sys/proc.h>
#include <sys/kthread.h>
#include <geom/mirror/g_mirror.h>


static MALLOC_DEFINE(M_MIRROR, "mirror data", "GEOM_MIRROR Data");

SYSCTL_DECL(_kern_geom);
SYSCTL_NODE(_kern_geom, OID_AUTO, mirror, CTLFLAG_RW, 0, "GEOM_MIRROR stuff");
static u_int g_mirror_debug = 4;
SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, debug, CTLFLAG_RW, &g_mirror_debug, 0,
    "Debug level");
static u_int g_mirror_sync_block_size = 131072;
SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, sync_block_size, CTLFLAG_RW,
    &g_mirror_sync_block_size, 0, "Synchronization block size");
static u_int g_mirror_timeout = 8;
SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, timeout, CTLFLAG_RW, &g_mirror_timeout,
    0, "Time to wait on all mirror components");

static int g_mirror_destroy_geom(struct gctl_req *req, struct g_class *mp,
    struct g_geom *gp);
static g_taste_t g_mirror_taste;

struct g_class g_mirror_class = {
	.name = G_MIRROR_CLASS_NAME,
	.ctlreq = g_mirror_config,
	.taste = g_mirror_taste,
	.destroy_geom = g_mirror_destroy_geom
};


static void g_mirror_destroy_disk(struct g_mirror_disk *disk);
static void g_mirror_destroy_provider(struct g_mirror_softc *sc);
static int g_mirror_update_disk(struct g_mirror_disk *disk, u_int state);
static void g_mirror_update_device(struct g_mirror_softc *sc, boolean_t force);
static void g_mirror_dumpconf(struct sbuf *sb, const char *indent,
    struct g_geom *gp, struct g_consumer *cp, struct g_provider *pp);
static void g_mirror_sync_stop(struct g_mirror_disk *disk);


static const char *
g_mirror_disk_state2str(int state)
{

	switch (state) {
	case G_MIRROR_DISK_STATE_NONE:
		return ("NONE");
	case G_MIRROR_DISK_STATE_NEW:
		return ("NEW");
	case G_MIRROR_DISK_STATE_ACTIVE:
		return ("ACTIVE");
	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
		return ("SYNCHRONIZING");
	case G_MIRROR_DISK_STATE_DISCONNECTED:
		return ("DISCONNECTED");
	case G_MIRROR_DISK_STATE_DESTROY:
		return ("DESTROY");
	default:
		return ("INVALID");
	}
}

static const char *
g_mirror_device_state2str(int state)
{

	switch (state) {
	case G_MIRROR_DEVICE_STATE_STARTING:
		return ("STARTING");
	case G_MIRROR_DEVICE_STATE_RUNNING:
		return ("RUNNING");
	default:
		return ("INVALID");
	}
}

static const char *
g_mirror_get_diskname(struct g_mirror_disk *disk)
{

	if (disk->d_consumer == NULL || disk->d_consumer->provider == NULL)
		return ("[unknown]");
	return (disk->d_name);
}

/*
 * --- Events handling functions ---
 * Events in geom_mirror are used to maintain disks and device status
 * from one thread to simplify locking.
 */
static void
g_mirror_event_free(struct g_mirror_event *ep)
{

	free(ep, M_MIRROR);
}

int
g_mirror_event_send(void *arg, int state, int flags)
{
	struct g_mirror_softc *sc;
	struct g_mirror_disk *disk;
	struct g_mirror_event *ep;
	int error;

	ep = malloc(sizeof(*ep), M_MIRROR, M_WAITOK);
	G_MIRROR_DEBUG(4, "%s: Sending event %p.", __func__, ep);
	if ((flags & G_MIRROR_EVENT_DEVICE) != 0) {
		disk = NULL;
		sc = arg;
	} else {
		disk = arg;
		sc = disk->d_softc;
	}
	ep->e_disk = disk;
	ep->e_state = state;
	ep->e_flags = flags;
	ep->e_error = 0;
	mtx_lock(&sc->sc_events_mtx);
	TAILQ_INSERT_TAIL(&sc->sc_events, ep, e_next);
	mtx_unlock(&sc->sc_events_mtx);
	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
	mtx_lock(&sc->sc_queue_mtx);
	wakeup(sc);
	mtx_unlock(&sc->sc_queue_mtx);
	if ((flags & G_MIRROR_EVENT_DONTWAIT) != 0)
		return (0);
	g_topology_assert();
	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, ep);
	g_topology_unlock();
	while ((ep->e_flags & G_MIRROR_EVENT_DONE) == 0) {
		mtx_lock(&sc->sc_events_mtx);
		msleep(ep, &sc->sc_events_mtx, PRIBIO | PDROP,
		    "mirror_event_wait", hz * 5);
	}
	g_topology_lock();
	G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, ep);
	error = ep->e_error;
	g_mirror_event_free(ep);
	return (error);
}

static struct g_mirror_event *
g_mirror_event_get(struct g_mirror_softc *sc)
{
	struct g_mirror_event *ep;

	mtx_lock(&sc->sc_events_mtx);
	ep = TAILQ_FIRST(&sc->sc_events);
	if (ep != NULL)
		TAILQ_REMOVE(&sc->sc_events, ep, e_next);
	mtx_unlock(&sc->sc_events_mtx);
	return (ep);
}

static void
g_mirror_event_cancel(struct g_mirror_disk *disk)
{
	struct g_mirror_softc *sc;
	struct g_mirror_event *ep, *tmpep;

	g_topology_assert();
	sc = disk->d_softc;
	mtx_lock(&sc->sc_events_mtx);
	TAILQ_FOREACH_SAFE(ep, &sc->sc_events, e_next, tmpep) {
		if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0)
			continue;
		if (ep->e_disk != disk)
			continue;
		TAILQ_REMOVE(&sc->sc_events, ep, e_next);
		if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
			g_mirror_event_free(ep);
		else {
			ep->e_error = ECANCELED;
			wakeup(ep);
		}
	}
	mtx_unlock(&sc->sc_events_mtx);
}

/*
 * Return the number of disks in given state.
 * If state is equal to -1, count all connected disks.
 */
u_int
g_mirror_ndisks(struct g_mirror_softc *sc, int state)
{
	struct g_mirror_disk *disk;
	u_int n = 0;

	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
		if (state == -1 || disk->d_state == state)
			n++;
	}
	return (n);
}

/*
 * Find a disk in mirror by its disk ID.
 */
static struct g_mirror_disk *
g_mirror_id2disk(struct g_mirror_softc *sc, uint32_t id)
{
	struct g_mirror_disk *disk;

	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
		if (disk->d_id == id)
			return (disk);
	}
	return (NULL);
}

static int
g_mirror_connect_disk(struct g_mirror_disk *disk, struct g_provider *pp)
{
	int error;

	g_topology_assert();

	error = g_attach(disk->d_consumer, pp);
	if (error != 0)
		return (error);
	G_MIRROR_DEBUG(3, "Disk %s connected.", g_mirror_get_diskname(disk));
	/*
	 * Gain full disk access right now.
	 * XXX: We don't support read-only devices.
	 */
	error = g_access(disk->d_consumer, 1, 1, 1);
	if (error != 0)
		return (error);
	G_MIRROR_DEBUG(3, "Disk %s opened.", g_mirror_get_diskname(disk));
	return (0);
}

static void
g_mirror_disconnect_disk(struct g_mirror_disk *disk)
{
	struct g_consumer *cp;

	g_topology_assert();
	cp = disk->d_consumer;
	if (cp->acr != 0 || cp->acw != 0 || cp->ace != 0)
		g_access(cp, -cp->ace, -cp->acw, -cp->ace);
	if (cp->provider != NULL)
		g_detach(cp);
}

/*
 * Initialize disk. This means allocate memory, create consumer, attach it
 * to the provider and open access (r1w1e1) to it.
 */
static struct g_mirror_disk *
g_mirror_init_disk(struct g_mirror_softc *sc, struct g_provider *pp,
    struct g_mirror_metadata *md, int *errorp)
{
	struct g_mirror_disk *disk;
	int error;

	disk = malloc(sizeof(*disk), M_MIRROR, M_NOWAIT | M_ZERO);
	if (disk == NULL) {
		error = ENOMEM;
		goto fail;
	}
	disk->d_softc = sc;
	disk->d_consumer = g_new_consumer(sc->sc_geom);
	disk->d_consumer->private = disk;
	error = g_mirror_connect_disk(disk, pp);
	if (error != 0)
		goto fail;
	disk->d_id = md->md_did;
	disk->d_count = 0;
	disk->d_state = G_MIRROR_DISK_STATE_NONE;
	disk->d_flags = md->md_flags & G_MIRROR_DISK_FLAG_MASK;
	disk->d_sync.ds_consumer = NULL;
	disk->d_sync.ds_offset = md->md_sync_offset;
	disk->d_sync.ds_offset_done = md->md_sync_offset;
	disk->d_sync.ds_syncid = md->md_syncid;
	if (errorp != NULL)
		*errorp = 0;
	return (disk);
fail:
	if (errorp != NULL)
		*errorp = error;
	if (disk != NULL) {
		if (disk->d_consumer != NULL) {
			g_mirror_disconnect_disk(disk);
			g_destroy_consumer(disk->d_consumer);
		}
		free(disk, M_MIRROR);
	}
	return (NULL);
}

/*
 * Free the disk.
 */
static void
g_mirror_free_disk(struct g_mirror_disk *disk)
{

	g_topology_assert();
	g_mirror_disconnect_disk(disk);
	g_destroy_consumer(disk->d_consumer);
	free(disk, M_MIRROR);
}

static void
g_mirror_destroy_device(struct g_mirror_softc *sc)
{
	struct g_mirror_disk *disk;
	struct g_mirror_event *ep;
	struct g_geom *gp;

	g_topology_assert();
	gp = sc->sc_geom;
	if (sc->sc_provider != NULL)
		g_mirror_destroy_provider(sc);
	for (disk = LIST_FIRST(&sc->sc_disks); disk != NULL;
	    disk = LIST_FIRST(&sc->sc_disks)) {
		g_mirror_destroy_disk(disk);
	}
	while ((ep = g_mirror_event_get(sc)) != NULL) {
		if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
			g_mirror_event_free(ep);
		else {
			ep->e_error = ECANCELED;
			ep->e_flags |= G_MIRROR_EVENT_DONE;
			G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, ep);
			mtx_lock(&sc->sc_events_mtx);
			wakeup(ep);
			mtx_unlock(&sc->sc_events_mtx);
		}
	}
	callout_drain(&sc->sc_callout);
	gp->softc = NULL;
	uma_zdestroy(sc->sc_sync.ds_zone);
	g_wither_geom(sc->sc_sync.ds_geom, ENXIO);
	mtx_destroy(&sc->sc_queue_mtx);
	mtx_destroy(&sc->sc_events_mtx);
	G_MIRROR_DEBUG(0, "Device %s destroyed.", gp->name);
	g_wither_geom(gp, ENXIO);
}

static void
g_mirror_destroy_disk(struct g_mirror_disk *disk)
{
	struct g_mirror_softc *sc;

	g_topology_assert();
	LIST_REMOVE(disk, d_next);
	g_mirror_event_cancel(disk);
	sc = disk->d_softc;
	if (sc->sc_hint == disk)
		sc->sc_hint = NULL;
	switch (disk->d_state) {
	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
		g_mirror_sync_stop(disk);
		/* FALLTHROUGH */
	case G_MIRROR_DISK_STATE_NEW:
	case G_MIRROR_DISK_STATE_ACTIVE:
		g_mirror_free_disk(disk);
		break;
	default:
		KASSERT(0 == 1, ("Wrong disk state (%s, %s).",
		    g_mirror_get_diskname(disk),
		    g_mirror_disk_state2str(disk->d_state)));
	}
}

static void
g_mirror_orphan(struct g_consumer *cp)
{
	struct g_mirror_disk *disk;

	disk = cp->private;
	g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
	    G_MIRROR_EVENT_DONTWAIT);
}

/*
 * Function should return the next active disk on list.
 * It is possible that it will be the same disk as given.
 * If there are no active disks on list, NULL is returned.
 */
static __inline struct g_mirror_disk *
g_mirror_find_next(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
{
	struct g_mirror_disk *dp;

	for (dp = LIST_NEXT(disk, d_next); dp != disk;
	    dp = LIST_NEXT(dp, d_next)) {
		if (dp == NULL)
			dp = LIST_FIRST(&sc->sc_disks);
		if (dp->d_state == G_MIRROR_DISK_STATE_ACTIVE)
			break;
	}
	if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
		return (NULL);
	return (dp);
}

static struct g_mirror_disk *
g_mirror_get_disk(struct g_mirror_softc *sc)
{
	struct g_mirror_disk *disk;

	if (sc->sc_hint == NULL) {
		sc->sc_hint = LIST_FIRST(&sc->sc_disks);
		if (sc->sc_hint == NULL)
			return (NULL);
	}
	disk = sc->sc_hint;
	if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE) {
		disk = g_mirror_find_next(sc, disk);
		if (disk == NULL)
			return (NULL);
	}
	sc->sc_hint = g_mirror_find_next(sc, disk);
	return (disk);
}

static int
g_mirror_clear_metadata(struct g_mirror_disk *disk)
{
	struct g_mirror_softc *sc;
	off_t offset, length;
	u_char *sector;
	int error;

	sc = disk->d_softc;
	KASSERT(disk->d_consumer != NULL, ("NULL consumer (%s).",
	    sc->sc_geom->name));
	KASSERT(disk->d_consumer->provider != NULL, ("NULL provider (%s).",
	    sc->sc_geom->name));
	length = disk->d_consumer->provider->sectorsize;
	offset = disk->d_consumer->provider->mediasize - length;
	sector = malloc((size_t)length, M_MIRROR, M_NOWAIT | M_ZERO);
	if (sector == NULL)
		return (ENOMEM);
	error = g_write_data(disk->d_consumer, offset, sector, length);
	free(sector, M_MIRROR);
	if (error != 0) {
		G_MIRROR_DEBUG(0, "Cannot clear metadata on disk %s.",
		    g_mirror_get_diskname(disk));
		g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
		    G_MIRROR_EVENT_DONTWAIT);
		return (error);
	}
	G_MIRROR_DEBUG(4, "Metadata on %s cleared.",
	    g_mirror_get_diskname(disk));
	return (0);
}

void
g_mirror_fill_metadata(struct g_mirror_softc *sc, struct g_mirror_disk *disk,
    struct g_mirror_metadata *md)
{

	strlcpy(md->md_magic, G_MIRROR_MAGIC, sizeof(md->md_magic));
	md->md_version = G_MIRROR_VERSION;
	strlcpy(md->md_name, sc->sc_name, sizeof(md->md_name));
	md->md_mid = sc->sc_id;
	md->md_all = sc->sc_ndisks;
	md->md_slice = sc->sc_slice;
	md->md_balance = sc->sc_balance;
	md->md_mediasize = sc->sc_mediasize;
	md->md_sectorsize = sc->sc_sectorsize;
	md->md_flags = sc->sc_flags;
	if (disk == NULL) {
		md->md_did = arc4random();
		md->md_syncid = 0;
		md->md_sync_offset = 0;
	} else {
		md->md_did = disk->d_id;
		md->md_syncid = disk->d_sync.ds_syncid;
		md->md_flags |= disk->d_flags;
		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
			md->md_sync_offset = disk->d_sync.ds_offset_done;
		else
			md->md_sync_offset = 0;
	}
}

void
g_mirror_update_metadata(struct g_mirror_disk *disk)
{
	struct g_mirror_softc *sc;
	struct g_mirror_metadata md;
	off_t offset, length;
	u_char *sector;
	int error;

	sc = disk->d_softc;
	KASSERT(disk->d_consumer != NULL, ("NULL consumer (%s).",
	    sc->sc_geom->name));
	KASSERT(disk->d_consumer->provider != NULL, ("NULL provider (%s).",
	    sc->sc_geom->name));
	length = disk->d_consumer->provider->sectorsize;
	offset = disk->d_consumer->provider->mediasize - length;
	sector = malloc((size_t)length, M_MIRROR, M_NOWAIT);
	if (sector == NULL)
		error = ENOMEM;
	else {
		g_mirror_fill_metadata(sc, disk, &md);
		mirror_metadata_encode(&md, sector);
		error = g_write_data(disk->d_consumer, offset, sector, length);
		free(sector, M_MIRROR);
	}
	if (error != 0) {
		G_MIRROR_DEBUG(0, "Cannot update metadata on disk %s.",
		    g_mirror_get_diskname(disk));
		g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
		    G_MIRROR_EVENT_DONTWAIT);
		return;
	}
	G_MIRROR_DEBUG(4, "Metadata on %s updated.",
	    g_mirror_get_diskname(disk));
}

static void
g_mirror_bump_syncid(struct g_mirror_softc *sc)
{
	struct g_mirror_disk *disk;

	KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
	    ("%s called with no active disks (device=%s).", __func__,
	    sc->sc_geom->name));
	sc->sc_syncid++;
	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
			disk->d_sync.ds_syncid = sc->sc_syncid;
			g_mirror_update_metadata(disk);
		}
	}
}

static void
g_mirror_done(struct bio *bp)
{
	struct g_mirror_softc *sc;
	struct g_mirror_disk *disk;
	struct bio *pbp;

	g_topology_assert_not();
	pbp = bp->bio_parent;
	sc = pbp->bio_to->geom->softc;

	pbp->bio_inbed++;
	if (bp->bio_error == 0) {
		G_MIRROR_LOGREQ(bp, "Request delivered.");
		g_destroy_bio(bp);
		if (pbp->bio_children == pbp->bio_inbed) {
			pbp->bio_completed = pbp->bio_length;
			G_MIRROR_LOGREQ(pbp, "Request delivered.");
			pbp->bio_error = 0;
			g_io_deliver(pbp, pbp->bio_error);
		}
		return;
	}

	/* An error occur, try to save situation. */

	G_MIRROR_LOGREQ(bp, "Request failed (error=%d).", bp->bio_error);
	disk = bp->bio_caller1;
	g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
	    G_MIRROR_EVENT_DONTWAIT);

	switch (pbp->bio_cmd) {
	case BIO_READ:
		g_destroy_bio(bp);
		mtx_lock(&sc->sc_queue_mtx);
		bioq_disksort(&sc->sc_queue, pbp);
		G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
		wakeup(sc);
		mtx_unlock(&sc->sc_queue_mtx);
		break;
	case BIO_DELETE:
	case BIO_WRITE:
		pbp->bio_inbed--;
		pbp->bio_children--;
		if (pbp->bio_error == 0)
			pbp->bio_error = bp->bio_error;
		g_destroy_bio(bp);
		if (pbp->bio_children == 0) {
			/*
			 * All requests failed.
			 */
		} else if (pbp->bio_children == pbp->bio_inbed) {
			/* At least one request succeeded. */
			pbp->bio_error = 0;
			pbp->bio_completed = pbp->bio_length;
		} else /* if (pbp->bio_inbed != pbp->bio_children) */ {
			KASSERT(pbp->bio_inbed < pbp->bio_children,
			    ("bio_inbed (%u) is not less then bio_children (%u).",
			    pbp->bio_inbed, pbp->bio_children));
			break;
		}
		g_io_deliver(pbp, pbp->bio_error);
		break;
	default:
		KASSERT(1 == 0, ("Invalid request: %u.", pbp->bio_cmd));
	}
}

static void
g_mirror_sync_done(struct bio *bp)
{
	struct g_mirror_softc *sc;

	G_MIRROR_LOGREQ(bp, "Synchronization request delivered.");
	sc = bp->bio_from->geom->softc;
	bp->bio_flags = BIO_FLAG1; 
	mtx_lock(&sc->sc_queue_mtx);
	bioq_disksort(&sc->sc_queue, bp);
	wakeup(sc);
	mtx_unlock(&sc->sc_queue_mtx);
}

static void
g_mirror_start(struct bio *bp)
{
	struct g_mirror_softc *sc;

	sc = bp->bio_to->geom->softc;
	/*
	 * If sc == NULL or there are no valid disks, provider's error
	 * should be set and g_mirror_start() should not be called at all.
	 */
	KASSERT(sc != NULL && sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
	    ("Provider's error should be set (error=%d)(mirror=%s).",
	    bp->bio_to->error, bp->bio_to->name));
	G_MIRROR_LOGREQ(bp, "Request received.");

	switch (bp->bio_cmd) {
	case BIO_READ:
	case BIO_WRITE:
	case BIO_DELETE:
		break;
	case BIO_GETATTR:
	default:
		g_io_deliver(bp, EOPNOTSUPP);
		return;
	}
	mtx_lock(&sc->sc_queue_mtx);
	bioq_disksort(&sc->sc_queue, bp);
	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
	wakeup(sc);
	mtx_unlock(&sc->sc_queue_mtx);
}

/*
 * Send one synchronization request.
 */
static void
g_mirror_sync_one(struct g_mirror_disk *disk)
{
	struct g_mirror_softc *sc;
	struct bio *bp;

	sc = disk->d_softc;
	KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
	    ("Disk %s is not marked for synchronization.",
	    g_mirror_get_diskname(disk)));

	bp = g_new_bio();
	if (bp == NULL)
		return;
	bp->bio_parent = NULL;
	bp->bio_cmd = BIO_READ;
	bp->bio_offset = disk->d_sync.ds_offset;
	bp->bio_length = MIN(sc->sc_sync.ds_block,
	    sc->sc_mediasize - bp->bio_offset);
	disk->d_sync.ds_offset += bp->bio_length;
	bp->bio_flags = 0;
	bp->bio_done = g_mirror_sync_done;
	bp->bio_data = uma_zalloc(sc->sc_sync.ds_zone, M_NOWAIT | M_ZERO);
	if (bp->bio_data == NULL) {
		g_destroy_bio(bp);
		return;
	}
	bp->bio_to = sc->sc_provider;
	bp->bio_caller1 = disk;
	G_MIRROR_LOGREQ(bp, "Sending synchronization request.");
	g_io_request(bp, disk->d_sync.ds_consumer);
}

static void
g_mirror_sync_request(struct bio *bp)
{
	struct g_mirror_softc *sc;
	struct g_mirror_disk *disk;

	disk = bp->bio_caller1;
	sc = disk->d_softc;

	/*
	 * Synchronization request.
	 */
	switch (bp->bio_cmd) {
	case BIO_READ:
		if (bp->bio_error != 0) {
			uma_zfree(sc->sc_sync.ds_zone, bp->bio_data);
			g_destroy_bio(bp);
			return;
		}
		bp->bio_cmd = BIO_WRITE;
		bp->bio_flags = 0;
		G_MIRROR_LOGREQ(bp, "READ finished.");
		g_io_request(bp, disk->d_consumer);
		return;
	case BIO_WRITE:
		uma_zfree(sc->sc_sync.ds_zone, bp->bio_data);
		if (bp->bio_error != 0) {
			g_destroy_bio(bp);
			g_mirror_event_send(disk,
			    G_MIRROR_DISK_STATE_DISCONNECTED,
			    G_MIRROR_EVENT_DONTWAIT);
			return;
		}
		G_MIRROR_LOGREQ(bp, "WRITE finished.");
		g_destroy_bio(bp);
		disk->d_sync.ds_offset_done = bp->bio_offset + bp->bio_length;
		if (bp->bio_offset + bp->bio_length ==
		    sc->sc_provider->mediasize) {
			/*
			 * Disk up-to-date, activate it.
			 */
			g_mirror_event_send(disk, G_MIRROR_DISK_STATE_ACTIVE,
			    G_MIRROR_EVENT_DONTWAIT);
			return;
		} else if ((disk->d_sync.ds_offset_done %
		    (sc->sc_sync.ds_block * 100)) == 0) {
			/*
			 * Update offset_done on every 100 blocks.
			 * XXX: This should be configurable.
			 */
			g_mirror_update_metadata(disk);
		}
		return;
	default:
		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
		    bp->bio_cmd, sc->sc_geom->name));
	}
}

static void
g_mirror_register_request(struct bio *bp)
{
	struct g_mirror_softc *sc;
	struct g_mirror_disk *disk;
	struct bio *cbp;

	sc = bp->bio_to->geom->softc;

	switch (bp->bio_cmd) {
	case BIO_READ:
		disk = g_mirror_get_disk(sc);
		if (disk == NULL) {
			if (bp->bio_error == 0)
				bp->bio_error = ENXIO;
			g_io_deliver(bp, bp->bio_error);
			return;
		}
		cbp = g_clone_bio(bp);
		if (cbp == NULL) {
			if (bp->bio_error == 0)
				bp->bio_error = ENOMEM;
			g_io_deliver(bp, bp->bio_error);
			return;
		}
		/*
		 * Fill in the component buf structure.
		 */
		cbp->bio_done = g_mirror_done;
		cbp->bio_caller1 = disk;
		cbp->bio_to = disk->d_consumer->provider;
		cbp->bio_caller1 = disk;
		G_MIRROR_LOGREQ(cbp, "Sending request.");
		g_io_request(cbp, disk->d_consumer);
		return;
	case BIO_WRITE:
	case BIO_DELETE:
	    {
		struct bio_queue_head queue;

		/*
		 * Allocate all bios before sending any request, so we can
		 * return ENOMEM in nice and clean way.
		 */
		bioq_init(&queue);
		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
			switch (disk->d_state) {
			case G_MIRROR_DISK_STATE_ACTIVE:
				break;
			case G_MIRROR_DISK_STATE_SYNCHRONIZING:
				if (bp->bio_offset >= disk->d_sync.ds_offset)
					continue;
				break;
			default:
				continue;
			}
			cbp = g_clone_bio(bp);
			if (cbp == NULL) {
				for (cbp = bioq_first(&queue); cbp != NULL;
				    cbp = bioq_first(&queue)) {
					bioq_remove(&queue, cbp);
					g_destroy_bio(cbp);
				}
				if (bp->bio_error == 0)
					bp->bio_error = ENOMEM;
				g_io_deliver(bp, bp->bio_error);
				return;
			}
			bioq_insert_tail(&queue, cbp);
		}
		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
			switch (disk->d_state) {
			case G_MIRROR_DISK_STATE_ACTIVE:
				break;
			case G_MIRROR_DISK_STATE_SYNCHRONIZING:
				if (bp->bio_offset >= disk->d_sync.ds_offset)
					continue;
				break;
			default:
				continue;
			}
			cbp = bioq_first(&queue);
			KASSERT(cbp != NULL, ("NULL cbp! (device %s).",
			    sc->sc_geom->name));
			bioq_remove(&queue, cbp);
			cbp->bio_done = g_mirror_done;
			cbp->bio_caller1 = disk;
			cbp->bio_to = disk->d_consumer->provider;
			G_MIRROR_LOGREQ(cbp, "Sending request.");
			g_io_request(cbp, disk->d_consumer);
		}
		/*
		 * Bump syncid on first write.
		 */
		if (sc->sc_bump_syncid) {
			sc->sc_bump_syncid = 0;
			g_mirror_bump_syncid(sc);
		}
		return;
	    }
	default:
		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
		    bp->bio_cmd, sc->sc_geom->name));
	}
}

/*
 * Worker thread.
 */
static void
g_mirror_worker(void *arg)
{
	struct g_mirror_softc *sc;
	struct g_mirror_disk *disk;
	struct g_mirror_event *ep;
	struct bio *bp;

	sc = arg;
	curthread->td_base_pri = PRIBIO;

	for (;;) {
		G_MIRROR_DEBUG(5, "%s: Let's see...", __func__);
		ep = g_mirror_event_get(sc);
		if (ep != NULL) {
			g_topology_lock();
			if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0) {
				/* Update only device status. */
				G_MIRROR_DEBUG(4,
				    "Running event for device %s.",
				    sc->sc_geom->name);
				ep->e_error = 0;
				g_mirror_update_device(sc, 1);
			} else {
				/* Update disk status. */
				G_MIRROR_DEBUG(4, "Running event for disk %s.",
				     g_mirror_get_diskname(ep->e_disk));
				ep->e_error = g_mirror_update_disk(ep->e_disk,
				    ep->e_state);
				if (ep->e_error == 0)
					g_mirror_update_device(sc, 0);
			}
			g_topology_unlock();
			if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0) {
				KASSERT(ep->e_error == 0,
				    ("Error cannot be handled."));
				g_mirror_event_free(ep);
			} else {
				ep->e_flags |= G_MIRROR_EVENT_DONE;
				G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
				    ep);
				mtx_lock(&sc->sc_events_mtx);
				wakeup(ep);
				mtx_unlock(&sc->sc_events_mtx);
			}
			if ((sc->sc_flags &
			    G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
end:
				if ((sc->sc_flags &
				    G_MIRROR_DEVICE_FLAG_WAIT) != 0) {
					G_MIRROR_DEBUG(4, "%s: Waking up %p.",
					    __func__, &sc->sc_worker);
					wakeup(&sc->sc_worker);
					sc->sc_worker = NULL;
				} else {
					g_topology_lock();
					g_mirror_destroy_device(sc);
					g_topology_unlock();
					free(sc, M_MIRROR);
				}
				kthread_exit(0);
			}
			continue;
		}
		/* Get first request from the queue. */
		mtx_lock(&sc->sc_queue_mtx);
		bp = bioq_first(&sc->sc_queue);
		if (bp == NULL) {
			mtx_unlock(&sc->sc_queue_mtx);
			if ((sc->sc_flags &
			    G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
				goto end;
			}
			/*
			 * No pending I/O request, so it is time for
			 * synchronization.
			 * XXX: It is possible that due to heavy device
			 *      load we will not be able to sync disks
			 *      at all, in that case we need to consider
			 *	forcing synchronization requests.
			 */
			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
				if (disk->d_state !=
				    G_MIRROR_DISK_STATE_SYNCHRONIZING) {
					continue;
				}
				if (disk->d_sync.ds_offset >=
				    sc->sc_provider->mediasize) {
					continue;
				}
				if (disk->d_sync.ds_offset >
				    disk->d_sync.ds_offset_done) {
					continue;
				}
				g_mirror_sync_one(disk);
			}
			G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, sc);
			mtx_lock(&sc->sc_queue_mtx);
			msleep(sc, &sc->sc_queue_mtx, PRIBIO | PDROP,
			    "mirror:wait", 0);
			G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, sc);
			continue;
		}
		bioq_remove(&sc->sc_queue, bp);
		mtx_unlock(&sc->sc_queue_mtx);

		if ((bp->bio_flags & BIO_FLAG1) != 0)
			g_mirror_sync_request(bp);
		else
			g_mirror_register_request(bp);
	}
}

static void
g_mirror_sync_start(struct g_mirror_disk *disk)
{
	struct g_mirror_softc *sc;
	int error;

	g_topology_assert();

	sc = disk->d_softc;
	KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
	    ("Device not in RUNNING state (%s, %u).", sc->sc_geom->name,
	    sc->sc_state));
	KASSERT(disk->d_sync.ds_consumer == NULL,
	    ("Sync consumer already exists (device=%s, disk=%s).",
	    sc->sc_geom->name, g_mirror_get_diskname(disk)));
	disk->d_sync.ds_consumer = g_new_consumer(sc->sc_sync.ds_geom);
	error = g_attach(disk->d_sync.ds_consumer, disk->d_softc->sc_provider);
	KASSERT(error == 0, ("Cannot attach to %s.",
	    disk->d_softc->sc_geom->name));
	error = g_access(disk->d_sync.ds_consumer, 1, 0, 0);
	KASSERT(error == 0, ("Cannot access to %s.",
	    disk->d_softc->sc_geom->name));
}

static void
g_mirror_sync_stop(struct g_mirror_disk *disk)
{

	g_topology_assert();

	KASSERT(disk->d_sync.ds_consumer != NULL,
	    ("No sync consumer (device=%s, disk=%s).",
	    disk->d_softc->sc_geom->name, g_mirror_get_diskname(disk)));
	g_access(disk->d_sync.ds_consumer, -1, 0, 0);
	g_detach(disk->d_sync.ds_consumer);
	g_destroy_consumer(disk->d_sync.ds_consumer);
	disk->d_sync.ds_consumer = NULL;
}

static void
g_mirror_launch_provider(struct g_mirror_softc *sc)
{
	struct g_provider *pp;

	g_topology_assert();

	pp = g_new_providerf(sc->sc_geom, sc->sc_geom->name);
	pp->mediasize = sc->sc_mediasize;
	pp->sectorsize = sc->sc_sectorsize;
	sc->sc_provider = pp;
	g_error_provider(pp, 0);
}

static void
g_mirror_destroy_provider(struct g_mirror_softc *sc)
{
	struct g_mirror_disk *disk;
	struct bio *bp;

	g_topology_assert();
	KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
	    sc->sc_geom->name));
	g_error_provider(sc->sc_provider, ENXIO);
	mtx_lock(&sc->sc_queue_mtx);
	while ((bp = bioq_first(&sc->sc_queue)) != NULL) {
		bioq_remove(&sc->sc_queue, bp);
		g_io_deliver(bp, ENXIO);
	}
	mtx_unlock(&sc->sc_queue_mtx);
	sc->sc_provider->flags |= G_PF_WITHER;
	g_orphan_provider(sc->sc_provider, ENXIO);
	sc->sc_provider = NULL;
	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
			g_mirror_sync_stop(disk);
	}
}

static void
g_mirror_go(void *arg)
{
	struct g_mirror_softc *sc;

	sc = arg;
	G_MIRROR_DEBUG(0, "Force %s start due to timeout.", sc->sc_geom->name);
	g_mirror_event_send(sc, 0,
	    G_MIRROR_EVENT_DONTWAIT | G_MIRROR_EVENT_DEVICE);
}

static u_int
g_mirror_determine_state(struct g_mirror_disk *disk)
{
	struct g_mirror_softc *sc;
	u_int state;

	sc = disk->d_softc;
	if (sc->sc_syncid == disk->d_sync.ds_syncid) {
		if ((disk->d_flags &
		    G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
			/* Disk does not need synchronization. */
			state = G_MIRROR_DISK_STATE_ACTIVE;
		} else {
			/*
			 * We can start synchronization from
			 * the stored offset.
			 */
			state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
		}
	} else if (sc->sc_syncid > disk->d_sync.ds_syncid) {
		/*
		 * Reset all synchronization data for this disk,
		 * because if it even was synchronized, it was
		 * synchronized to disks with different syncid.
		 */
		disk->d_flags |= G_MIRROR_DISK_FLAG_SYNCHRONIZING;
		disk->d_sync.ds_offset = 0;
		disk->d_sync.ds_offset_done = 0;
		disk->d_sync.ds_syncid = sc->sc_syncid;
		state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
	} else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
		/*
		 * Not good, NOT GOOD!
		 * It means that mirror was started on stale disks
		 * and more fresh disk just arrive.
		 * If there were writes, mirror is fucked up, sorry.
		 * I think the best choice here is don't touch
		 * this disk and inform the user laudly.
		 */
		printf("Mirror %s was started before the freshest disk "
		    "(%s) arrives!! It will not be connected to the "
		    "running mirror.\n", sc->sc_geom->name,
		    g_mirror_get_diskname(disk));
		g_mirror_destroy_disk(disk);
		state = G_MIRROR_DISK_STATE_NONE;
	}
	G_MIRROR_DEBUG(1, "State for %s disk: %s.",
	    g_mirror_get_diskname(disk), g_mirror_disk_state2str(state));
	return (state);
}

/*
 * Update device state.
 */
static void
g_mirror_update_device(struct g_mirror_softc *sc, boolean_t force)
{
	struct g_mirror_disk *disk;
	u_int state;

	g_topology_assert();
	switch (sc->sc_state) {
	case G_MIRROR_DEVICE_STATE_STARTING:
	    {
		uint16_t syncid;

		KASSERT(sc->sc_provider == NULL,
		    ("Non-NULL provider in STARTING state (%s).",
		    sc->sc_geom->name));
		/*
		 * Are we ready? We are, if all disks are connected or
		 * if we have any disks and 'force' is true.
		 */
		if ((force && g_mirror_ndisks(sc, -1) > 0) ||
		    sc->sc_ndisks == g_mirror_ndisks(sc, -1)) {
			;
		} else if (g_mirror_ndisks(sc, -1) == 0) {
			/*
			 * Disks went down in starting phase, so destroy
			 * device.
			 */
			callout_drain(&sc->sc_callout);
			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
			return;
		} else {
			return;
		}

		/*
		 * Activate all disks with the biggest syncid.
		 */
		if (force) {
			u_int ndisks;

			/*
			 * If called with 'force' true, we're called from
			 * timeout * procedure, so don't bother canceling
			 * timeout.
			 */
			ndisks = 0;
			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
				if ((disk->d_flags &
				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
					ndisks++;
				}
			}
			if (ndisks == 0) {
				int timeout;

				/* No valid disks still, wait some more. */
				timeout =
				    atomic_load_acq_int(&g_mirror_timeout);
				callout_reset(&sc->sc_callout, timeout * hz,
				    g_mirror_go, sc);
				return;
			}
		} else {
			/* Cancel timeout. */
			callout_drain(&sc->sc_callout);
			/*
			 * If timeout procedure was in progress, check if
			 * device was turned on already.
			 */
			if (sc->sc_state != G_MIRROR_DEVICE_STATE_STARTING)
				return;
		}
		/* Find disk with the biggest syncid. */
		syncid = 0;
		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
			if (disk->d_sync.ds_syncid > syncid)
				syncid = disk->d_sync.ds_syncid;
		}
		/* Reset hint. */
		sc->sc_hint = NULL;
		sc->sc_syncid = syncid;
		if (force) {
			/* Remember to bump syncid on first write. */
			sc->sc_bump_syncid = 1;
		}
		/* Create and run provider. */
		g_mirror_launch_provider(sc);
		state = G_MIRROR_DEVICE_STATE_RUNNING;
		G_MIRROR_DEBUG(1, "Device %s state changed from %s to %s.",
		    sc->sc_geom->name, g_mirror_device_state2str(sc->sc_state),
		    g_mirror_device_state2str(state));
		sc->sc_state = state;
		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
			state = g_mirror_determine_state(disk);
			g_mirror_event_send(disk, state,
			    G_MIRROR_EVENT_DONTWAIT);
		}
		break;
	    }
	case G_MIRROR_DEVICE_STATE_RUNNING:
		KASSERT(sc->sc_provider != NULL,
		    ("NULL provider in RUNNING state (%s).",
		    sc->sc_geom->name));
		/*
		 * Basically, we should check if device may still run.
		 * It may run if we have at least one active disk.
		 */
		if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0)
			break;
		/*
		 * No active disks, we need to destroy device.
		 */
		sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
		break;
	default:
		KASSERT(1 == 0, ("Wrong device state (%s, %s).",
		    sc->sc_geom->name,
		    g_mirror_device_state2str(sc->sc_state)));
	}
}

/*
 * Update disk state and device state if needed.
 */
static int
g_mirror_update_disk(struct g_mirror_disk *disk, u_int state)
{
	struct g_mirror_softc *sc;

	g_topology_assert();
	sc = disk->d_softc;
again:
	G_MIRROR_DEBUG(3, "Changing disk %s state from %s to %s.",
	    g_mirror_get_diskname(disk), g_mirror_disk_state2str(disk->d_state),
	    g_mirror_disk_state2str(state));
	switch (state) {
	case G_MIRROR_DISK_STATE_NEW:
	    {
		struct g_mirror_disk *dp;

		/*
		 * Possible scenarios:
		 * 1. New disk arrive.
		 */
		/* Previous state should be NONE. */
		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NONE,
		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
		    g_mirror_disk_state2str(disk->d_state)));

		/*
		 * Check for duplication.
		 */
		dp = g_mirror_id2disk(sc, disk->d_id);
		if (dp != NULL) {
			struct g_provider *pp;

			/*
			 * We have a duplication.
			 */
			pp = disk->d_consumer->provider;
			g_mirror_free_disk(disk);
			G_MIRROR_DEBUG(1,
			    "Disk %s have the same ID as disk %s!", pp->name,
			    dp->d_consumer->provider->name);
			return (EEXIST);
		} 
		G_MIRROR_DEBUG(1,
		    "Disk %s state changed from %s to %s (device %s).",
		    g_mirror_get_diskname(disk),
		    g_mirror_disk_state2str(disk->d_state),
		    g_mirror_disk_state2str(state),
		    sc->sc_geom->name);
		disk->d_state = state;
		LIST_INSERT_HEAD(&sc->sc_disks, disk, d_next);
		if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
			break;
		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_geom->name,
		    g_mirror_device_state2str(sc->sc_state),
		    g_mirror_get_diskname(disk),
		    g_mirror_disk_state2str(disk->d_state)));
		state = g_mirror_determine_state(disk);
		if (state == G_MIRROR_DISK_STATE_NONE)
			return (0);
		goto again;
	    }
	case G_MIRROR_DISK_STATE_ACTIVE:
		/*
		 * Possible scenarios:
		 * 1. New disk does not need synchronization.
		 * 2. Synchronization process finished successfully.
		 */
		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_geom->name,
		    g_mirror_device_state2str(sc->sc_state),
		    g_mirror_get_diskname(disk),
		    g_mirror_disk_state2str(disk->d_state)));
		/* Previous state should be NEW or SYNCHRONIZING. */
		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW ||
		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
		    g_mirror_disk_state2str(disk->d_state)));
		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
			disk->d_flags &= ~G_MIRROR_DISK_FLAG_SYNCHRONIZING;
			g_mirror_sync_stop(disk);
		}
		G_MIRROR_DEBUG(1,
		    "Disk %s state changed from %s to %s (device %s).",
		    g_mirror_get_diskname(disk),
		    g_mirror_disk_state2str(disk->d_state),
		    g_mirror_disk_state2str(state),
		    sc->sc_geom->name);
		disk->d_state = state;
		disk->d_sync.ds_offset = 0;
		disk->d_sync.ds_offset_done = 0;
		disk->d_flags &= ~G_MIRROR_DISK_FLAG_SYNCHRONIZING;
		g_mirror_update_metadata(disk);
		break;
	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
		/*
		 * Possible scenarios:
		 * 1. Disk which needs synchronization was connected.
		 */
		/* Previous state should be NEW. */
		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
		    g_mirror_disk_state2str(disk->d_state)));
		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_geom->name,
		    g_mirror_device_state2str(sc->sc_state),
		    g_mirror_get_diskname(disk),
		    g_mirror_disk_state2str(disk->d_state)));
		G_MIRROR_DEBUG(1,
		    "Disk %s state changed from %s to %s (device %s).",
		    g_mirror_get_diskname(disk),
		    g_mirror_disk_state2str(disk->d_state),
		    g_mirror_disk_state2str(state), sc->sc_geom->name);
		disk->d_state = state;
		g_mirror_update_metadata(disk);
		if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING)
			g_mirror_sync_start(disk);
		break;
	case G_MIRROR_DISK_STATE_DISCONNECTED:
		/*
		 * Possible scenarios:
		 * 1. Device wasn't running yet, but disk disappear.
		 * 2. Disk was active and disapppear.
		 * 3. Disk disappear during synchronization process.
		 */
		if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING) {
			/* Previous state should be ACTIVE or SYNCHRONIZING. */
			KASSERT(disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
			    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
			    ("Wrong disk state (%s, %s).",
			    g_mirror_get_diskname(disk),
			    g_mirror_disk_state2str(disk->d_state)));
			/*
			 * Remember to bump syncid on first write, but only if
			 * we're in RUNNING state.
			 */
			sc->sc_bump_syncid = 1;
#ifdef	INVARIANTS
		} else if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING) {
			/* Previous state should be NEW. */
			KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
			    ("Wrong disk state (%s, %s).",
			    g_mirror_get_diskname(disk),
			    g_mirror_disk_state2str(disk->d_state)));
		} else {
			KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
			    sc->sc_geom->name,
			    g_mirror_device_state2str(sc->sc_state),
			    g_mirror_get_diskname(disk),
			    g_mirror_disk_state2str(disk->d_state)));
#endif
		}
		G_MIRROR_DEBUG(1,
		    "Disk %s state changed from %s to %s (device %s).",
		    g_mirror_get_diskname(disk),
		    g_mirror_disk_state2str(disk->d_state),
		    g_mirror_disk_state2str(state), sc->sc_geom->name);
		g_mirror_destroy_disk(disk);
		break;
	case G_MIRROR_DISK_STATE_DESTROY:
	    {
		int error;

		error = g_mirror_clear_metadata(disk);
		if (error != 0)
			return (error);
		G_MIRROR_DEBUG(1,
		    "Disk %s state changed from %s to %s (device %s).",
		    g_mirror_get_diskname(disk),
		    g_mirror_disk_state2str(disk->d_state),
		    g_mirror_disk_state2str(state), sc->sc_geom->name);
		g_mirror_destroy_disk(disk);
		sc->sc_ndisks--;
		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
			g_mirror_update_metadata(disk);
		}
		break;
	    }
	default:
		KASSERT(1 == 0, ("Unknown state (%u).", state));
	}
	return (0);
}

static int
g_mirror_read_metadata(struct g_consumer *cp, struct g_mirror_metadata *md)
{
	struct g_provider *pp;
	u_char *buf;
	int error;

	g_topology_assert();

	error = g_access(cp, 1, 0, 0);
	if (error != 0)
		return (error);
	pp = cp->provider;
	g_topology_unlock();
	/* Metadata are stored on last sector. */
	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
	    &error);
	g_topology_lock();
	if (buf == NULL) {
		g_access(cp, -1, 0, 0);
		return (error);
	}
	if (error != 0) {
		g_access(cp, -1, 0, 0);
		g_free(buf);
		return (error);
	}
	error = g_access(cp, -1, 0, 0);
	KASSERT(error == 0, ("Cannot decrease access count for %s.", pp->name));

	/* Decode metadata. */
	mirror_metadata_decode(buf, md);
	g_free(buf);

	return (0);
}

/*
 * Add disk to the mirror.
 *
 * Algorithm:
 *
 * If bitmap for disk exists, start synchronization process.
 * NOTE: If on-disk synchronization ID and mirror synchronization ID are equal,
 *       but bitmap exists, it is bogus. It means, that disk was disconnected
 *       and its synchronization ID was set manually. We could assert this, but
 *       don't be so cruel, life is cruel already, and just ignore this and
 *       start synchronization process depending on bitmap.
 *
 * If there is no bitmap and synchronization IDs are different, whole disk
 * have to be synchronized.
 * NOTE: If synchronization ID on connected disk is bigger that mirror's
 *       synchronization ID, it is bogus. It means, that someone change it
 *       for this disk manually. Again we could assert this, but again we
 *       have to remember that world is cruel enough and people are evil.
 *
 * The easiest situation to handle: synchronization IDs are equal and there is
 * no bitmap. It means that there were no writes while disk was disconnected.
 * Quite great, because we don't need any synchronization at all here.
 * But don't be so happy, remember that nobody likes you and this is the most
 * uncommon situation.
 *
 * ...I really should assert those "NOTE:" cases someday, I see no reason to be
 * better than cruel world, no reason at all.
 */
static int
g_mirror_add_disk(struct g_mirror_softc *sc, struct g_provider *pp,
    struct g_mirror_metadata *md)
{
	struct g_mirror_disk *disk;
	int error = 0;

	g_topology_assert();
	G_MIRROR_DEBUG(2, "Adding disk %s.", pp->name);
	disk = g_mirror_init_disk(sc, pp, md, &error);
	if (disk == NULL)
		return (error);
	error = g_mirror_event_send(disk, G_MIRROR_DISK_STATE_NEW,
	    G_MIRROR_EVENT_WAIT);
	return (error);
}

static struct g_geom *
g_mirror_create(struct g_class *mp, const struct g_mirror_metadata *md)
{
	struct g_mirror_softc *sc;
	struct g_geom *gp;
	int error, timeout;

	G_MIRROR_DEBUG(1, "Creating device %s.mirror (id=%u).", md->md_name,
	    md->md_mid);

	/* Two disks is minimum. */
	if (md->md_all <= 1)
		return (NULL);
	/*
	 * Action geom.
	 */
	gp = g_new_geomf(mp, "%s.mirror", md->md_name);
	sc = malloc(sizeof(*sc), M_MIRROR, M_NOWAIT | M_ZERO);
	if (sc == NULL) {
		G_MIRROR_DEBUG(0, "Can't allocate memory for device %s.",
		    gp->name);
		g_destroy_geom(gp);
		return (NULL);
	}
	gp->start = g_mirror_start;
	gp->spoiled = g_mirror_orphan;
	gp->orphan = g_mirror_orphan;
	gp->access = g_std_access;
	gp->dumpconf = g_mirror_dumpconf;

	strlcpy(sc->sc_name, md->md_name, sizeof(sc->sc_name));
	sc->sc_id = md->md_mid;
	sc->sc_slice = md->md_slice;
	sc->sc_balance = md->md_balance;
	sc->sc_mediasize = md->md_mediasize;
	sc->sc_sectorsize = md->md_sectorsize;
	sc->sc_ndisks = md->md_all;
	sc->sc_bump_syncid = 0;
	bioq_init(&sc->sc_queue);
	mtx_init(&sc->sc_queue_mtx, "gmirror:queue", NULL, MTX_DEF);
	LIST_INIT(&sc->sc_disks);
	TAILQ_INIT(&sc->sc_events);
	mtx_init(&sc->sc_events_mtx, "gmirror:events", NULL, MTX_DEF);
	callout_init(&sc->sc_callout, CALLOUT_MPSAFE);
	sc->sc_state = G_MIRROR_DEVICE_STATE_STARTING;
	gp->softc = sc;
	sc->sc_geom = gp;
	sc->sc_provider = NULL;
	/*
	 * Synchronization geom.
	 */
	gp = g_new_geomf(mp, "%s.mirror.sync", md->md_name);
	gp->softc = sc;
	gp->spoiled = g_mirror_orphan;
	gp->orphan = g_mirror_orphan;
	sc->sc_sync.ds_geom = gp;
	sc->sc_sync.ds_block = atomic_load_acq_int(&g_mirror_sync_block_size);
	sc->sc_sync.ds_zone = uma_zcreate("gmirror:sync", sc->sc_sync.ds_block,
	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
	error = kthread_create(g_mirror_worker, sc, &sc->sc_worker, 0, 0,
	    "g_mirror %s", md->md_name);
	if (error != 0) {
		G_MIRROR_DEBUG(1, "Cannot create kernel thread for %s.",
		    sc->sc_geom->name);
		uma_zdestroy(sc->sc_sync.ds_zone);
		g_destroy_geom(sc->sc_sync.ds_geom);
		mtx_destroy(&sc->sc_events_mtx);
		mtx_destroy(&sc->sc_queue_mtx);
		g_destroy_geom(sc->sc_geom);
		free(sc, M_MIRROR);
		return (NULL);
	}

	G_MIRROR_DEBUG(0, "Device %s created (id=%u).", sc->sc_geom->name,
	    sc->sc_id);

	/*
	 * Run timeout.
	 */
	timeout = atomic_load_acq_int(&g_mirror_timeout);
	callout_reset(&sc->sc_callout, timeout * hz, g_mirror_go, sc);
	return (sc->sc_geom);
}

int
g_mirror_destroy(struct g_mirror_softc *sc, boolean_t force)
{
	struct g_provider *pp;

	g_topology_assert();
	if (sc == NULL)
		return (ENXIO);
	pp = sc->sc_provider;
	if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
		if (force) {
			G_MIRROR_DEBUG(0, "Device %s is still open, so it "
			    "can't be definitely removed.", pp->name);
		} else {
			G_MIRROR_DEBUG(1,
			    "Device %s is still open (r%dw%de%d).", pp->name,
			    pp->acr, pp->acw, pp->ace);
			return (EBUSY);
		}
	}

	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_WAIT;
	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
	mtx_lock(&sc->sc_queue_mtx);
	wakeup(sc);
	mtx_unlock(&sc->sc_queue_mtx);
	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
	while (sc->sc_worker != NULL)
		tsleep(&sc->sc_worker, PRIBIO, "mirror:destroy", hz / 5);
	G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
	g_mirror_destroy_device(sc);
	free(sc, M_MIRROR);
	return (0);
}

static void
g_mirror_dump_metadata(struct g_mirror_metadata *md)
{

	printf("     magic: %s\n", md->md_magic);
	printf("   version: %u\n", (u_int)md->md_version);
	printf("      name: %s\n", md->md_name);
	printf("       mid: %u\n", (u_int)md->md_mid);
	printf("       did: %u\n", (u_int)md->md_did);
	printf("       all: %u\n", (u_int)md->md_all);
	printf("    syncid: %u\n", (u_int)md->md_syncid);
	printf("     slice: %u\n", (u_int)md->md_slice);
	printf("   balance: %u\n", (u_int)md->md_balance);
	printf(" mediasize: %jd\n", (intmax_t)md->md_mediasize);
	printf("sectorsize: %u\n", (u_int)md->md_sectorsize);
	printf("syncoffset: %jd\n", (intmax_t)md->md_sync_offset);
	printf("     flags: %ju\n", (uintmax_t)md->md_flags);
}

static void
g_mirror_taste_orphan(struct g_consumer *cp)
{

	KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
	    cp->provider->name));
}

/* XXX: Temporary. */
int g_mirror_stop = 0;

static struct g_geom *
g_mirror_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
{
	struct g_mirror_metadata md;
	struct g_mirror_softc *sc;
	struct g_consumer *cp;
	struct g_geom *gp;
	int error;

	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
	g_topology_assert();

	if (g_mirror_stop)
		return (NULL);

	/*
	 * XXX: Hack to force mirror to use NOP providers only for now.
	 */
	if (strcmp(pp->geom->class->name, "NOP") != 0)
		return (NULL);
	G_MIRROR_DEBUG(3, "Tasting %s.", pp->name);

	gp = g_new_geomf(mp, "mirror:taste");
	/*
	 * This orphan function should be never called.
	 */
	gp->orphan = g_mirror_taste_orphan;
	cp = g_new_consumer(gp);
	g_attach(cp, pp);

	error = g_mirror_read_metadata(cp, &md);
	g_wither_geom(gp, ENXIO);
	if (error != 0)
		return (NULL);
	gp = NULL;

	if (strcmp(md.md_magic, G_MIRROR_MAGIC) != 0)
		return (NULL);
	if (md.md_version > G_MIRROR_VERSION) {
		printf("geom_mirror.ko module is too old to handle %s.\n",
		    pp->name);
		return (NULL);
	}
	if ((md.md_flags & G_MIRROR_DISK_FLAG_INACTIVE) != 0) {
		G_MIRROR_DEBUG(1, "Disk %s marked as inactive, skipping.",
		    pp->name);
		return (NULL);
	}
	g_mirror_dump_metadata(&md);

	/*
	 * Let's check if device already exists.
	 */
	LIST_FOREACH(gp, &mp->geom, geom) {
		sc = gp->softc;
		if (sc == NULL)
			continue;
		if (sc->sc_sync.ds_geom == gp)
			continue;
		if (strcmp(md.md_name, sc->sc_name) != 0)
			continue;
		if (md.md_mid != sc->sc_id) {
			G_MIRROR_DEBUG(0, "Device %s already configured.",
			    gp->name);
			return (NULL);
		}
		break;
	}
	if (gp == NULL) {
		gp = g_mirror_create(mp, &md);
		if (gp == NULL) {
			G_MIRROR_DEBUG(0, "Cannot create device %s.mirror",
			    md.md_name);
			return (NULL);
		}
		sc = gp->softc;
	}
	G_MIRROR_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
	error = g_mirror_add_disk(sc, pp, &md);
	if (error != 0) {
		G_MIRROR_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
		    pp->name, gp->name, error);
		if (LIST_EMPTY(&sc->sc_disks))
			g_mirror_destroy(sc, 1);
		return (NULL);
	}
	return (gp);
}

static int
g_mirror_destroy_geom(struct gctl_req *req __unused,
    struct g_class *mp __unused, struct g_geom *gp)
{

	return (g_mirror_destroy(gp->softc, 0));
}

static void
g_mirror_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
    struct g_consumer *cp, struct g_provider *pp)
{
	struct g_mirror_softc *sc;
	struct g_mirror_disk *disk;

	sc = gp->softc;
	if (sc == NULL || pp == NULL)
		return;
	sbuf_printf(sb, "%s<mid>%zu</mid>\n", indent, sc->sc_id);
	sbuf_printf(sb, "%s<syncid>%u</syncid>\n", indent, sc->sc_syncid);
	sbuf_printf(sb, "%s<slice>%zu</slice>\n", indent, sc->sc_slice);
	sbuf_printf(sb, "%s<balance>%s</balance>\n", indent,
	    balance_name(sc->sc_balance));
	sbuf_printf(sb, "%s<flags>%ju</flags>\n", indent,
	    (uintmax_t)sc->sc_flags);
	sbuf_printf(sb, "%s<nproviders>%u</nproviders>\n", indent,
	    sc->sc_ndisks);
	sbuf_printf(sb, "%s<providers>", indent);
	/* XXX: Disks list can change during our operation. */
	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
		sbuf_printf(sb, "%s(%s", g_mirror_get_diskname(disk),
		    g_mirror_disk_state2str(disk->d_state));
		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
			sbuf_printf(sb, " %u%%",
			    (u_int)((sc->sc_provider->mediasize * 100) /
			    disk->d_sync.ds_offset_done));
		}
		sbuf_printf(sb, ")");
		if (LIST_NEXT(disk, d_next) != NULL)
			sbuf_printf(sb, " ");
	}
	sbuf_printf(sb, "</providers>\n");
}

DECLARE_GEOM_CLASS(g_mirror_class, g_mirror);