xref: /freebsd/sys/geom/raid/md_nvidia.c (revision 7b2a8d7823218318e4a11183927cf085eee284ce)

/*-
 * Copyright (c) 2011 Alexander Motin <mav@FreeBSD.org>
 * Copyright (c) 2000 - 2008 Søren Schmidt <sos@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/bio.h>
#include <sys/endian.h>
#include <sys/kernel.h>
#include <sys/kobj.h>
#include <sys/limits.h>
#include <sys/lock.h>
#include <sys/malloc.h>
#include <sys/mutex.h>
#include <sys/systm.h>
#include <sys/taskqueue.h>
#include <geom/geom.h>
#include "geom/raid/g_raid.h"
#include "g_raid_md_if.h"

static MALLOC_DEFINE(M_MD_NVIDIA, "md_nvidia_data", "GEOM_RAID NVIDIA metadata");

struct nvidia_raid_conf {
	uint8_t		nvidia_id[8];
#define NVIDIA_MAGIC                "NVIDIA  "

	uint32_t	config_size;
	uint32_t	checksum;
	uint16_t	version;
	uint8_t		disk_number;
	uint8_t		dummy_0;
	uint32_t	total_sectors;
	uint32_t	sector_size;
	uint8_t		name[16];
	uint8_t		revision[4];
	uint32_t	disk_status;

	uint32_t	magic_0;
#define NVIDIA_MAGIC0		0x00640044

	uint64_t	volume_id[2];
	uint8_t		state;
#define NVIDIA_S_IDLE		0
#define NVIDIA_S_INIT		2
#define NVIDIA_S_REBUILD	3
#define NVIDIA_S_UPGRADE	4
#define NVIDIA_S_SYNC		5
	uint8_t		array_width;
	uint8_t		total_disks;
	uint8_t		orig_array_width;
	uint16_t	type;
#define NVIDIA_T_RAID0		0x0080
#define NVIDIA_T_RAID1		0x0081
#define NVIDIA_T_RAID3		0x0083
#define NVIDIA_T_RAID5		0x0085	/* RLQ = 00/02? */
#define NVIDIA_T_RAID5_SYM	0x0095	/* RLQ = 03 */
#define NVIDIA_T_RAID10		0x008a
#define NVIDIA_T_RAID01		0x8180
#define NVIDIA_T_CONCAT		0x00ff

	uint16_t	dummy_3;
	uint32_t	strip_sectors;
	uint32_t	strip_bytes;
	uint32_t	strip_shift;
	uint32_t	strip_mask;
	uint32_t	stripe_sectors;
	uint32_t	stripe_bytes;
	uint32_t	rebuild_lba;
	uint32_t	orig_type;
	uint32_t	orig_total_sectors;
	uint32_t	status;
#define NVIDIA_S_BOOTABLE	0x00000001
#define NVIDIA_S_DEGRADED	0x00000002

	uint32_t	filler[98];
} __packed;

struct g_raid_md_nvidia_perdisk {
	struct nvidia_raid_conf	*pd_meta;
	int			 pd_disk_pos;
	off_t			 pd_disk_size;
};

struct g_raid_md_nvidia_object {
	struct g_raid_md_object	 mdio_base;
	uint64_t		 mdio_volume_id[2];
	struct nvidia_raid_conf	*mdio_meta;
	struct callout		 mdio_start_co;	/* STARTING state timer. */
	int			 mdio_total_disks;
	int			 mdio_disks_present;
	int			 mdio_started;
	int			 mdio_incomplete;
	struct root_hold_token	*mdio_rootmount; /* Root mount delay token. */
};

static g_raid_md_create_t g_raid_md_create_nvidia;
static g_raid_md_taste_t g_raid_md_taste_nvidia;
static g_raid_md_event_t g_raid_md_event_nvidia;
static g_raid_md_ctl_t g_raid_md_ctl_nvidia;
static g_raid_md_write_t g_raid_md_write_nvidia;
static g_raid_md_fail_disk_t g_raid_md_fail_disk_nvidia;
static g_raid_md_free_disk_t g_raid_md_free_disk_nvidia;
static g_raid_md_free_t g_raid_md_free_nvidia;

static kobj_method_t g_raid_md_nvidia_methods[] = {
	KOBJMETHOD(g_raid_md_create,	g_raid_md_create_nvidia),
	KOBJMETHOD(g_raid_md_taste,	g_raid_md_taste_nvidia),
	KOBJMETHOD(g_raid_md_event,	g_raid_md_event_nvidia),
	KOBJMETHOD(g_raid_md_ctl,	g_raid_md_ctl_nvidia),
	KOBJMETHOD(g_raid_md_write,	g_raid_md_write_nvidia),
	KOBJMETHOD(g_raid_md_fail_disk,	g_raid_md_fail_disk_nvidia),
	KOBJMETHOD(g_raid_md_free_disk,	g_raid_md_free_disk_nvidia),
	KOBJMETHOD(g_raid_md_free,	g_raid_md_free_nvidia),
	{ 0, 0 }
};

static struct g_raid_md_class g_raid_md_nvidia_class = {
	"NVIDIA",
	g_raid_md_nvidia_methods,
	sizeof(struct g_raid_md_nvidia_object),
	.mdc_priority = 100
};

static int NVIDIANodeID = 1;

static void
g_raid_md_nvidia_print(struct nvidia_raid_conf *meta)
{

	if (g_raid_debug < 1)
		return;

	printf("********* ATA NVIDIA RAID Metadata *********\n");
	printf("nvidia_id           <%.8s>\n", meta->nvidia_id);
	printf("config_size         %u\n", meta->config_size);
	printf("checksum            0x%08x\n", meta->checksum);
	printf("version             0x%04x\n", meta->version);
	printf("disk_number         %d\n", meta->disk_number);
	printf("dummy_0             0x%02x\n", meta->dummy_0);
	printf("total_sectors       %u\n", meta->total_sectors);
	printf("sector_size         %u\n", meta->sector_size);
	printf("name                <%.16s>\n", meta->name);
	printf("revision            0x%02x%02x%02x%02x\n",
	    meta->revision[0], meta->revision[1],
	    meta->revision[2], meta->revision[3]);
	printf("disk_status         0x%08x\n", meta->disk_status);
	printf("magic_0             0x%08x\n", meta->magic_0);
	printf("volume_id           0x%016jx%016jx\n",
	    meta->volume_id[1], meta->volume_id[0]);
	printf("state               0x%02x\n", meta->state);
	printf("array_width         %u\n", meta->array_width);
	printf("total_disks         %u\n", meta->total_disks);
	printf("orig_array_width    %u\n", meta->orig_array_width);
	printf("type                0x%04x\n", meta->type);
	printf("dummy_3             0x%04x\n", meta->dummy_3);
	printf("strip_sectors       %u\n", meta->strip_sectors);
	printf("strip_bytes         %u\n", meta->strip_bytes);
	printf("strip_shift         %u\n", meta->strip_shift);
	printf("strip_mask          0x%08x\n", meta->strip_mask);
	printf("stripe_sectors      %u\n", meta->stripe_sectors);
	printf("stripe_bytes        %u\n", meta->stripe_bytes);
	printf("rebuild_lba         %u\n", meta->rebuild_lba);
	printf("orig_type           0x%04x\n", meta->orig_type);
	printf("orig_total_sectors  %u\n", meta->orig_total_sectors);
	printf("status              0x%08x\n", meta->status);
	printf("=================================================\n");
}

static struct nvidia_raid_conf *
nvidia_meta_copy(struct nvidia_raid_conf *meta)
{
	struct nvidia_raid_conf *nmeta;

	nmeta = malloc(sizeof(*meta), M_MD_NVIDIA, M_WAITOK);
	memcpy(nmeta, meta, sizeof(*meta));
	return (nmeta);
}

static int
nvidia_meta_translate_disk(struct nvidia_raid_conf *meta, int md_disk_pos)
{
	int disk_pos;

	if (md_disk_pos >= 0 && meta->type == NVIDIA_T_RAID01) {
		disk_pos = (md_disk_pos / meta->array_width) +
		    (md_disk_pos % meta->array_width) * meta->array_width;
	} else
		disk_pos = md_disk_pos;
	return (disk_pos);
}

static void
nvidia_meta_get_name(struct nvidia_raid_conf *meta, char *buf)
{
	int i;

	strncpy(buf, meta->name, 16);
	buf[16] = 0;
	for (i = 15; i >= 0; i--) {
		if (buf[i] > 0x20)
			break;
		buf[i] = 0;
	}
}

static void
nvidia_meta_put_name(struct nvidia_raid_conf *meta, char *buf)
{

	memset(meta->name, 0x20, 16);
	memcpy(meta->name, buf, MIN(strlen(buf), 16));
}

static struct nvidia_raid_conf *
nvidia_meta_read(struct g_consumer *cp)
{
	struct g_provider *pp;
	struct nvidia_raid_conf *meta;
	char *buf;
	int error, i;
	uint32_t checksum, *ptr;

	pp = cp->provider;

	/* Read the anchor sector. */
	buf = g_read_data(cp,
	    pp->mediasize - 2 * pp->sectorsize, pp->sectorsize, &error);
	if (buf == NULL) {
		G_RAID_DEBUG(1, "Cannot read metadata from %s (error=%d).",
		    pp->name, error);
		return (NULL);
	}
	meta = malloc(sizeof(*meta), M_MD_NVIDIA, M_WAITOK);
	memcpy(meta, buf, min(sizeof(*meta), pp->sectorsize));
	g_free(buf);

	/* Check if this is an NVIDIA RAID struct */
	if (strncmp(meta->nvidia_id, NVIDIA_MAGIC, strlen(NVIDIA_MAGIC))) {
		G_RAID_DEBUG(1, "NVIDIA signature check failed on %s", pp->name);
		free(meta, M_MD_NVIDIA);
		return (NULL);
	}
	if (meta->config_size > 128 ||
	    meta->config_size < 30) {
		G_RAID_DEBUG(1, "NVIDIA metadata size looks wrong: %d",
		    meta->config_size);
		free(meta, M_MD_NVIDIA);
		return (NULL);
	}

	/* Check metadata checksum. */
	for (checksum = 0, ptr = (uint32_t *)meta,
	    i = 0; i < meta->config_size; i++)
		checksum += *ptr++;
	if (checksum != 0) {
		G_RAID_DEBUG(1, "NVIDIA checksum check failed on %s", pp->name);
		free(meta, M_MD_NVIDIA);
		return (NULL);
	}

	/* Check volume state. */
	if (meta->state != NVIDIA_S_IDLE && meta->state != NVIDIA_S_INIT &&
	    meta->state != NVIDIA_S_REBUILD && meta->state != NVIDIA_S_SYNC) {
		G_RAID_DEBUG(1, "NVIDIA unknown state on %s (0x%02x)",
		    pp->name, meta->state);
		free(meta, M_MD_NVIDIA);
		return (NULL);
	}

	/* Check raid type. */
	if (meta->type != NVIDIA_T_RAID0 && meta->type != NVIDIA_T_RAID1 &&
	    meta->type != NVIDIA_T_RAID3 && meta->type != NVIDIA_T_RAID5 &&
	    meta->type != NVIDIA_T_RAID5_SYM &&
	    meta->type != NVIDIA_T_RAID01 && meta->type != NVIDIA_T_CONCAT) {
		G_RAID_DEBUG(1, "NVIDIA unknown RAID level on %s (0x%02x)",
		    pp->name, meta->type);
		free(meta, M_MD_NVIDIA);
		return (NULL);
	}

	return (meta);
}

static int
nvidia_meta_write(struct g_consumer *cp, struct nvidia_raid_conf *meta)
{
	struct g_provider *pp;
	char *buf;
	int error, i;
	uint32_t checksum, *ptr;

	pp = cp->provider;

	/* Recalculate checksum for case if metadata were changed. */
	meta->checksum = 0;
	for (checksum = 0, ptr = (uint32_t *)meta,
	    i = 0; i < meta->config_size; i++)
		checksum += *ptr++;
	meta->checksum -= checksum;

	/* Create and fill buffer. */
	buf = malloc(pp->sectorsize, M_MD_NVIDIA, M_WAITOK | M_ZERO);
	memcpy(buf, meta, sizeof(*meta));

	/* Write metadata. */
	error = g_write_data(cp,
	    pp->mediasize - 2 * pp->sectorsize, buf, pp->sectorsize);
	if (error != 0) {
		G_RAID_DEBUG(1, "Cannot write metadata to %s (error=%d).",
		    pp->name, error);
	}

	free(buf, M_MD_NVIDIA);
	return (error);
}

static int
nvidia_meta_erase(struct g_consumer *cp)
{
	struct g_provider *pp;
	char *buf;
	int error;

	pp = cp->provider;
	buf = malloc(pp->sectorsize, M_MD_NVIDIA, M_WAITOK | M_ZERO);
	error = g_write_data(cp,
	    pp->mediasize - 2 * pp->sectorsize, buf, pp->sectorsize);
	if (error != 0) {
		G_RAID_DEBUG(1, "Cannot erase metadata on %s (error=%d).",
		    pp->name, error);
	}
	free(buf, M_MD_NVIDIA);
	return (error);
}

static struct g_raid_disk *
g_raid_md_nvidia_get_disk(struct g_raid_softc *sc, int id)
{
	struct g_raid_disk	*disk;
	struct g_raid_md_nvidia_perdisk *pd;

	TAILQ_FOREACH(disk, &sc->sc_disks, d_next) {
		pd = (struct g_raid_md_nvidia_perdisk *)disk->d_md_data;
		if (pd->pd_disk_pos == id)
			break;
	}
	return (disk);
}

static int
g_raid_md_nvidia_supported(int level, int qual, int disks, int force)
{

	switch (level) {
	case G_RAID_VOLUME_RL_RAID0:
		if (disks < 1)
			return (0);
		if (!force && (disks < 2 || disks > 6))
			return (0);
		break;
	case G_RAID_VOLUME_RL_RAID1:
		if (disks < 1)
			return (0);
		if (!force && (disks != 2))
			return (0);
		break;
	case G_RAID_VOLUME_RL_RAID1E:
		if (disks < 2)
			return (0);
		if (disks % 2 != 0)
			return (0);
		if (!force && (disks < 4))
			return (0);
		break;
	case G_RAID_VOLUME_RL_SINGLE:
		if (disks != 1)
			return (0);
		break;
	case G_RAID_VOLUME_RL_CONCAT:
		if (disks < 2)
			return (0);
		break;
	case G_RAID_VOLUME_RL_RAID5:
		if (disks < 3)
			return (0);
		if (qual != G_RAID_VOLUME_RLQ_R5LA &&
		    qual != G_RAID_VOLUME_RLQ_R5LS)
			return (0);
		break;
	default:
		return (0);
	}
	if (level != G_RAID_VOLUME_RL_RAID5 && qual != G_RAID_VOLUME_RLQ_NONE)
		return (0);
	return (1);
}

static int
g_raid_md_nvidia_start_disk(struct g_raid_disk *disk)
{
	struct g_raid_softc *sc;
	struct g_raid_subdisk *sd, *tmpsd;
	struct g_raid_disk *olddisk, *tmpdisk;
	struct g_raid_md_object *md;
	struct g_raid_md_nvidia_object *mdi;
	struct g_raid_md_nvidia_perdisk *pd, *oldpd;
	struct nvidia_raid_conf *meta;
	int disk_pos, resurrection = 0;

	sc = disk->d_softc;
	md = sc->sc_md;
	mdi = (struct g_raid_md_nvidia_object *)md;
	meta = mdi->mdio_meta;
	pd = (struct g_raid_md_nvidia_perdisk *)disk->d_md_data;
	olddisk = NULL;

	/* Find disk position in metadata by it's serial. */
	if (pd->pd_meta != NULL) {
		disk_pos = pd->pd_meta->disk_number;
		if (disk_pos >= meta->total_disks || mdi->mdio_started)
			disk_pos = -3;
	} else
		disk_pos = -3;
	/* For RAID0+1 we need to translate order. */
	disk_pos = nvidia_meta_translate_disk(meta, disk_pos);
	if (disk_pos < 0) {
		G_RAID_DEBUG1(1, sc, "Unknown, probably new or stale disk");
		/* If we are in the start process, that's all for now. */
		if (!mdi->mdio_started)
			goto nofit;
		/*
		 * If we have already started - try to get use of the disk.
		 * Try to replace OFFLINE disks first, then FAILED.
		 */
		TAILQ_FOREACH(tmpdisk, &sc->sc_disks, d_next) {
			if (tmpdisk->d_state != G_RAID_DISK_S_OFFLINE &&
			    tmpdisk->d_state != G_RAID_DISK_S_FAILED)
				continue;
			/* Make sure this disk is big enough. */
			TAILQ_FOREACH(sd, &tmpdisk->d_subdisks, sd_next) {
				if (sd->sd_offset + sd->sd_size + 2 * 512 >
				    pd->pd_disk_size) {
					G_RAID_DEBUG1(1, sc,
					    "Disk too small (%ju < %ju)",
					    pd->pd_disk_size,
					    sd->sd_offset + sd->sd_size + 512);
					break;
				}
			}
			if (sd != NULL)
				continue;
			if (tmpdisk->d_state == G_RAID_DISK_S_OFFLINE) {
				olddisk = tmpdisk;
				break;
			} else if (olddisk == NULL)
				olddisk = tmpdisk;
		}
		if (olddisk == NULL) {
nofit:
			g_raid_change_disk_state(disk, G_RAID_DISK_S_SPARE);
			return (1);
		}
		oldpd = (struct g_raid_md_nvidia_perdisk *)olddisk->d_md_data;
		disk_pos = oldpd->pd_disk_pos;
		resurrection = 1;
	}

	if (olddisk == NULL) {
		/* Find placeholder by position. */
		olddisk = g_raid_md_nvidia_get_disk(sc, disk_pos);
		if (olddisk == NULL)
			panic("No disk at position %d!", disk_pos);
		if (olddisk->d_state != G_RAID_DISK_S_OFFLINE) {
			G_RAID_DEBUG1(1, sc, "More then one disk for pos %d",
			    disk_pos);
			g_raid_change_disk_state(disk, G_RAID_DISK_S_STALE);
			return (0);
		}
		oldpd = (struct g_raid_md_nvidia_perdisk *)olddisk->d_md_data;
	}

	/* Replace failed disk or placeholder with new disk. */
	TAILQ_FOREACH_SAFE(sd, &olddisk->d_subdisks, sd_next, tmpsd) {
		TAILQ_REMOVE(&olddisk->d_subdisks, sd, sd_next);
		TAILQ_INSERT_TAIL(&disk->d_subdisks, sd, sd_next);
		sd->sd_disk = disk;
	}
	oldpd->pd_disk_pos = -2;
	pd->pd_disk_pos = disk_pos;

	/* If it was placeholder -- destroy it. */
	if (olddisk->d_state == G_RAID_DISK_S_OFFLINE) {
		g_raid_destroy_disk(olddisk);
	} else {
		/* Otherwise, make it STALE_FAILED. */
		g_raid_change_disk_state(olddisk, G_RAID_DISK_S_STALE_FAILED);
	}

	/* Welcome the new disk. */
	if (resurrection)
		g_raid_change_disk_state(disk, G_RAID_DISK_S_ACTIVE);
	else// if (pd->pd_meta->disk_status == NVIDIA_S_CURRENT ||
	    //pd->pd_meta->disk_status == NVIDIA_S_REBUILD)
		g_raid_change_disk_state(disk, G_RAID_DISK_S_ACTIVE);
//	else
//		g_raid_change_disk_state(disk, G_RAID_DISK_S_FAILED);
	TAILQ_FOREACH(sd, &disk->d_subdisks, sd_next) {

		/*
		 * Different disks may have different sizes,
		 * in concat mode. Update from real disk size.
		 */
		if (meta->type == NVIDIA_T_CONCAT)
			sd->sd_size = pd->pd_disk_size - 0x800 * 512;

		if (resurrection) {
			/* New or ex-spare disk. */
			g_raid_change_subdisk_state(sd,
			    G_RAID_SUBDISK_S_NEW);
		} else if (meta->state == NVIDIA_S_REBUILD &&
		    (pd->pd_meta->disk_status & 0x100)) {
			/* Rebuilding disk. */
			g_raid_change_subdisk_state(sd,
			    G_RAID_SUBDISK_S_REBUILD);
			sd->sd_rebuild_pos = (off_t)pd->pd_meta->rebuild_lba /
			    meta->array_width * pd->pd_meta->sector_size;
		} else if (meta->state == NVIDIA_S_SYNC) {
			/* Resyncing/dirty disk. */
			g_raid_change_subdisk_state(sd,
			    G_RAID_SUBDISK_S_RESYNC);
			sd->sd_rebuild_pos = (off_t)pd->pd_meta->rebuild_lba /
			    meta->array_width * pd->pd_meta->sector_size;
		} else {
			/* Up to date disk. */
			g_raid_change_subdisk_state(sd,
			    G_RAID_SUBDISK_S_ACTIVE);
		}
		g_raid_event_send(sd, G_RAID_SUBDISK_E_NEW,
		    G_RAID_EVENT_SUBDISK);
	}

	/* Update status of our need for spare. */
	if (mdi->mdio_started) {
		mdi->mdio_incomplete =
		    (g_raid_ndisks(sc, G_RAID_DISK_S_ACTIVE) <
		     mdi->mdio_total_disks);
	}

	return (resurrection);
}

static void
g_disk_md_nvidia_retaste(void *arg, int pending)
{

	G_RAID_DEBUG(1, "Array is not complete, trying to retaste.");
	g_retaste(&g_raid_class);
	free(arg, M_MD_NVIDIA);
}

static void
g_raid_md_nvidia_refill(struct g_raid_softc *sc)
{
	struct g_raid_md_object *md;
	struct g_raid_md_nvidia_object *mdi;
	struct g_raid_disk *disk;
	struct task *task;
	int update, na;

	md = sc->sc_md;
	mdi = (struct g_raid_md_nvidia_object *)md;
	update = 0;
	do {
		/* Make sure we miss anything. */
		na = g_raid_ndisks(sc, G_RAID_DISK_S_ACTIVE);
		if (na == mdi->mdio_total_disks)
			break;

		G_RAID_DEBUG1(1, md->mdo_softc,
		    "Array is not complete (%d of %d), "
		    "trying to refill.", na, mdi->mdio_total_disks);

		/* Try to get use some of STALE disks. */
		TAILQ_FOREACH(disk, &sc->sc_disks, d_next) {
			if (disk->d_state == G_RAID_DISK_S_STALE) {
				update += g_raid_md_nvidia_start_disk(disk);
				if (disk->d_state == G_RAID_DISK_S_ACTIVE)
					break;
			}
		}
		if (disk != NULL)
			continue;

		/* Try to get use some of SPARE disks. */
		TAILQ_FOREACH(disk, &sc->sc_disks, d_next) {
			if (disk->d_state == G_RAID_DISK_S_SPARE) {
				update += g_raid_md_nvidia_start_disk(disk);
				if (disk->d_state == G_RAID_DISK_S_ACTIVE)
					break;
			}
		}
	} while (disk != NULL);

	/* Write new metadata if we changed something. */
	if (update)
		g_raid_md_write_nvidia(md, NULL, NULL, NULL);

	/* Update status of our need for spare. */
	mdi->mdio_incomplete = (g_raid_ndisks(sc, G_RAID_DISK_S_ACTIVE) <
	    mdi->mdio_total_disks);

	/* Request retaste hoping to find spare. */
	if (mdi->mdio_incomplete) {
		task = malloc(sizeof(struct task),
		    M_MD_NVIDIA, M_WAITOK | M_ZERO);
		TASK_INIT(task, 0, g_disk_md_nvidia_retaste, task);
		taskqueue_enqueue(taskqueue_swi, task);
	}
}

static void
g_raid_md_nvidia_start(struct g_raid_softc *sc)
{
	struct g_raid_md_object *md;
	struct g_raid_md_nvidia_object *mdi;
	struct g_raid_md_nvidia_perdisk *pd;
	struct nvidia_raid_conf *meta;
	struct g_raid_volume *vol;
	struct g_raid_subdisk *sd;
	struct g_raid_disk *disk;
	off_t size;
	int j, disk_pos;
	char buf[17];

	md = sc->sc_md;
	mdi = (struct g_raid_md_nvidia_object *)md;
	meta = mdi->mdio_meta;

	/* Create volumes and subdisks. */
	nvidia_meta_get_name(meta, buf);
	vol = g_raid_create_volume(sc, buf, -1);
	vol->v_mediasize = (off_t)meta->total_sectors * 512;
	vol->v_raid_level_qualifier = G_RAID_VOLUME_RLQ_NONE;
	if (meta->type == NVIDIA_T_RAID0) {
		vol->v_raid_level = G_RAID_VOLUME_RL_RAID0;
		size = vol->v_mediasize / mdi->mdio_total_disks;
	} else if (meta->type == NVIDIA_T_RAID1) {
		vol->v_raid_level = G_RAID_VOLUME_RL_RAID1;
		size = vol->v_mediasize;
	} else if (meta->type == NVIDIA_T_RAID01) {
		vol->v_raid_level = G_RAID_VOLUME_RL_RAID1E;
		size = vol->v_mediasize / (mdi->mdio_total_disks / 2);
	} else if (meta->type == NVIDIA_T_CONCAT) {
		if (mdi->mdio_total_disks == 1)
			vol->v_raid_level = G_RAID_VOLUME_RL_SINGLE;
		else
			vol->v_raid_level = G_RAID_VOLUME_RL_CONCAT;
		size = 0;
	} else if (meta->type == NVIDIA_T_RAID5) {
		vol->v_raid_level = G_RAID_VOLUME_RL_RAID5;
		vol->v_raid_level_qualifier = G_RAID_VOLUME_RLQ_R5LA;
		size = vol->v_mediasize / (mdi->mdio_total_disks - 1);
	} else if (meta->type == NVIDIA_T_RAID5_SYM) {
		vol->v_raid_level = G_RAID_VOLUME_RL_RAID5;
		vol->v_raid_level_qualifier = G_RAID_VOLUME_RLQ_R5LS;
		size = vol->v_mediasize / (mdi->mdio_total_disks - 1);
	} else {
		vol->v_raid_level = G_RAID_VOLUME_RL_UNKNOWN;
		size = 0;
	}
	vol->v_strip_size = meta->strip_sectors * 512; //ZZZ
	vol->v_disks_count = mdi->mdio_total_disks;
	vol->v_sectorsize = 512; //ZZZ
	for (j = 0; j < vol->v_disks_count; j++) {
		sd = &vol->v_subdisks[j];
		sd->sd_offset = 0;
		sd->sd_size = size;
	}
	g_raid_start_volume(vol);

	/* Create disk placeholders to store data for later writing. */
	for (disk_pos = 0; disk_pos < mdi->mdio_total_disks; disk_pos++) {
		pd = malloc(sizeof(*pd), M_MD_NVIDIA, M_WAITOK | M_ZERO);
		pd->pd_disk_pos = disk_pos;
		disk = g_raid_create_disk(sc);
		disk->d_md_data = (void *)pd;
		disk->d_state = G_RAID_DISK_S_OFFLINE;
		sd = &vol->v_subdisks[disk_pos];
		sd->sd_disk = disk;
		TAILQ_INSERT_TAIL(&disk->d_subdisks, sd, sd_next);
	}

	/* Make all disks found till the moment take their places. */
	do {
		TAILQ_FOREACH(disk, &sc->sc_disks, d_next) {
			if (disk->d_state == G_RAID_DISK_S_NONE) {
				g_raid_md_nvidia_start_disk(disk);
				break;
			}
		}
	} while (disk != NULL);

	mdi->mdio_started = 1;
	G_RAID_DEBUG1(0, sc, "Array started.");
	g_raid_md_write_nvidia(md, NULL, NULL, NULL);

	/* Pickup any STALE/SPARE disks to refill array if needed. */
	g_raid_md_nvidia_refill(sc);

	g_raid_event_send(vol, G_RAID_VOLUME_E_START, G_RAID_EVENT_VOLUME);

	callout_stop(&mdi->mdio_start_co);
	G_RAID_DEBUG1(1, sc, "root_mount_rel %p", mdi->mdio_rootmount);
	root_mount_rel(mdi->mdio_rootmount);
	mdi->mdio_rootmount = NULL;
}

static void
g_raid_md_nvidia_new_disk(struct g_raid_disk *disk)
{
	struct g_raid_softc *sc;
	struct g_raid_md_object *md;
	struct g_raid_md_nvidia_object *mdi;
	struct nvidia_raid_conf *pdmeta;
	struct g_raid_md_nvidia_perdisk *pd;

	sc = disk->d_softc;
	md = sc->sc_md;
	mdi = (struct g_raid_md_nvidia_object *)md;
	pd = (struct g_raid_md_nvidia_perdisk *)disk->d_md_data;
	pdmeta = pd->pd_meta;

	if (mdi->mdio_started) {
		if (g_raid_md_nvidia_start_disk(disk))
			g_raid_md_write_nvidia(md, NULL, NULL, NULL);
	} else {
		if (mdi->mdio_meta == NULL ||
		    mdi->mdio_meta->disk_number >= mdi->mdio_meta->total_disks) {
			G_RAID_DEBUG1(1, sc, "Newer disk");
			if (mdi->mdio_meta != NULL)
				free(mdi->mdio_meta, M_MD_NVIDIA);
			mdi->mdio_meta = nvidia_meta_copy(pdmeta);
			mdi->mdio_total_disks = pdmeta->total_disks;
			mdi->mdio_disks_present = 1;
		} else if (pdmeta->disk_number < mdi->mdio_meta->total_disks) {
			mdi->mdio_disks_present++;
			G_RAID_DEBUG1(1, sc, "Matching disk (%d of %d up)",
			    mdi->mdio_disks_present,
			    mdi->mdio_total_disks);
		} else
			G_RAID_DEBUG1(1, sc, "Spare disk");

		/* If we collected all needed disks - start array. */
		if (mdi->mdio_disks_present == mdi->mdio_total_disks)
			g_raid_md_nvidia_start(sc);
	}
}

static void
g_raid_nvidia_go(void *arg)
{
	struct g_raid_softc *sc;
	struct g_raid_md_object *md;
	struct g_raid_md_nvidia_object *mdi;

	sc = arg;
	md = sc->sc_md;
	mdi = (struct g_raid_md_nvidia_object *)md;
	if (!mdi->mdio_started) {
		G_RAID_DEBUG1(0, sc, "Force array start due to timeout.");
		g_raid_event_send(sc, G_RAID_NODE_E_START, 0);
	}
}

static int
g_raid_md_create_nvidia(struct g_raid_md_object *md, struct g_class *mp,
    struct g_geom **gp)
{
	struct g_raid_softc *sc;
	struct g_raid_md_nvidia_object *mdi;
	char name[32];

	mdi = (struct g_raid_md_nvidia_object *)md;
	arc4rand(&mdi->mdio_volume_id, 16, 0);
	snprintf(name, sizeof(name), "NVIDIA-%d",
	    atomic_fetchadd_int(&NVIDIANodeID, 1));
	sc = g_raid_create_node(mp, name, md);
	if (sc == NULL)
		return (G_RAID_MD_TASTE_FAIL);
	md->mdo_softc = sc;
	*gp = sc->sc_geom;
	return (G_RAID_MD_TASTE_NEW);
}

static int
g_raid_md_taste_nvidia(struct g_raid_md_object *md, struct g_class *mp,
                              struct g_consumer *cp, struct g_geom **gp)
{
	struct g_consumer *rcp;
	struct g_provider *pp;
	struct g_raid_md_nvidia_object *mdi, *mdi1;
	struct g_raid_softc *sc;
	struct g_raid_disk *disk;
	struct nvidia_raid_conf *meta;
	struct g_raid_md_nvidia_perdisk *pd;
	struct g_geom *geom;
	int error, result, spare, len;
	char name[32];
	uint16_t vendor;

	G_RAID_DEBUG(1, "Tasting NVIDIA on %s", cp->provider->name);
	mdi = (struct g_raid_md_nvidia_object *)md;
	pp = cp->provider;

	/* Read metadata from device. */
	meta = NULL;
	vendor = 0xffff;
	if (g_access(cp, 1, 0, 0) != 0)
		return (G_RAID_MD_TASTE_FAIL);
	g_topology_unlock();
	len = 2;
	if (pp->geom->rank == 1)
		g_io_getattr("GEOM::hba_vendor", cp, &len, &vendor);
	meta = nvidia_meta_read(cp);
	g_topology_lock();
	g_access(cp, -1, 0, 0);
	if (meta == NULL) {
		if (g_raid_aggressive_spare) {
			if (vendor == 0x10de) {
				G_RAID_DEBUG(1,
				    "No NVIDIA metadata, forcing spare.");
				spare = 2;
				goto search;
			} else {
				G_RAID_DEBUG(1,
				    "NVIDIA vendor mismatch 0x%04x != 0x10de",
				    vendor);
			}
		}
		return (G_RAID_MD_TASTE_FAIL);
	}

	/* Metadata valid. Print it. */
	g_raid_md_nvidia_print(meta);
	G_RAID_DEBUG(1, "NVIDIA disk position %d", meta->disk_number);
	spare = 0;//(meta->type == NVIDIA_T_SPARE) ? 1 : 0;

search:
	/* Search for matching node. */
	sc = NULL;
	mdi1 = NULL;
	LIST_FOREACH(geom, &mp->geom, geom) {
		sc = geom->softc;
		if (sc == NULL)
			continue;
		if (sc->sc_stopping != 0)
			continue;
		if (sc->sc_md->mdo_class != md->mdo_class)
			continue;
		mdi1 = (struct g_raid_md_nvidia_object *)sc->sc_md;
		if (spare) {
			if (mdi1->mdio_incomplete)
				break;
		} else {
			if (memcmp(&mdi1->mdio_volume_id,
			     &meta->volume_id, 16) == 0)
				break;
		}
	}

	/* Found matching node. */
	if (geom != NULL) {
		G_RAID_DEBUG(1, "Found matching array %s", sc->sc_name);
		result = G_RAID_MD_TASTE_EXISTING;

	} else if (spare) { /* Not found needy node -- left for later. */
		G_RAID_DEBUG(1, "Spare is not needed at this time");
		goto fail1;

	} else { /* Not found matching node -- create one. */
		result = G_RAID_MD_TASTE_NEW;
		memcpy(&mdi->mdio_volume_id, &meta->volume_id, 16);
		snprintf(name, sizeof(name), "NVIDIA-%d",
		    atomic_fetchadd_int(&NVIDIANodeID, 1));
		sc = g_raid_create_node(mp, name, md);
		md->mdo_softc = sc;
		geom = sc->sc_geom;
		callout_init(&mdi->mdio_start_co, 1);
		callout_reset(&mdi->mdio_start_co, g_raid_start_timeout * hz,
		    g_raid_nvidia_go, sc);
		mdi->mdio_rootmount = root_mount_hold("GRAID-NVIDIA");
		G_RAID_DEBUG1(1, sc, "root_mount_hold %p", mdi->mdio_rootmount);
	}

	rcp = g_new_consumer(geom);
	g_attach(rcp, pp);
	if (g_access(rcp, 1, 1, 1) != 0)
		; //goto fail1;

	g_topology_unlock();
	sx_xlock(&sc->sc_lock);

	pd = malloc(sizeof(*pd), M_MD_NVIDIA, M_WAITOK | M_ZERO);
	pd->pd_meta = meta;
	if (spare == 2) {
		pd->pd_disk_pos = -3;
	} else {
		pd->pd_disk_pos = -1;
	}
	pd->pd_disk_size = pp->mediasize;
	disk = g_raid_create_disk(sc);
	disk->d_md_data = (void *)pd;
	disk->d_consumer = rcp;
	rcp->private = disk;

	/* Read kernel dumping information. */
	disk->d_kd.offset = 0;
	disk->d_kd.length = OFF_MAX;
	len = sizeof(disk->d_kd);
	error = g_io_getattr("GEOM::kerneldump", rcp, &len, &disk->d_kd);
	if (disk->d_kd.di.dumper == NULL)
		G_RAID_DEBUG1(2, sc, "Dumping not supported by %s: %d.",
		    rcp->provider->name, error);

	g_raid_md_nvidia_new_disk(disk);

	sx_xunlock(&sc->sc_lock);
	g_topology_lock();
	*gp = geom;
	return (result);
fail1:
	free(meta, M_MD_NVIDIA);
	return (G_RAID_MD_TASTE_FAIL);
}

static int
g_raid_md_event_nvidia(struct g_raid_md_object *md,
    struct g_raid_disk *disk, u_int event)
{
	struct g_raid_softc *sc;
	struct g_raid_subdisk *sd;
	struct g_raid_md_nvidia_object *mdi;
	struct g_raid_md_nvidia_perdisk *pd;

	sc = md->mdo_softc;
	mdi = (struct g_raid_md_nvidia_object *)md;
	if (disk == NULL) {
		switch (event) {
		case G_RAID_NODE_E_START:
			if (!mdi->mdio_started) {
				/* Bump volume ID to drop missing disks. */
				arc4rand(&mdi->mdio_volume_id, 16, 0);
				g_raid_md_nvidia_start(sc);
			}
			return (0);
		}
		return (-1);
	}
	pd = (struct g_raid_md_nvidia_perdisk *)disk->d_md_data;
	switch (event) {
	case G_RAID_DISK_E_DISCONNECTED:
		/* If disk was assigned, just update statuses. */
		if (pd->pd_disk_pos >= 0) {
			g_raid_change_disk_state(disk, G_RAID_DISK_S_OFFLINE);
			if (disk->d_consumer) {
				g_raid_kill_consumer(sc, disk->d_consumer);
				disk->d_consumer = NULL;
			}
			TAILQ_FOREACH(sd, &disk->d_subdisks, sd_next) {
				g_raid_change_subdisk_state(sd,
				    G_RAID_SUBDISK_S_NONE);
				g_raid_event_send(sd, G_RAID_SUBDISK_E_DISCONNECTED,
				    G_RAID_EVENT_SUBDISK);
			}
		} else {
			/* Otherwise -- delete. */
			g_raid_change_disk_state(disk, G_RAID_DISK_S_NONE);
			g_raid_destroy_disk(disk);
		}

		if (mdi->mdio_started) {
			/* Bump volume ID to prevent disk resurrection. */
			if (pd->pd_disk_pos >= 0)
				arc4rand(&mdi->mdio_volume_id, 16, 0);

			/* Write updated metadata to all disks. */
			g_raid_md_write_nvidia(md, NULL, NULL, NULL);
		}

		/* Check if anything left except placeholders. */
		if (g_raid_ndisks(sc, -1) ==
		    g_raid_ndisks(sc, G_RAID_DISK_S_OFFLINE))
			g_raid_destroy_node(sc, 0);
		else
			g_raid_md_nvidia_refill(sc);
		return (0);
	}
	return (-2);
}

static int
g_raid_md_ctl_nvidia(struct g_raid_md_object *md,
    struct gctl_req *req)
{
	struct g_raid_softc *sc;
	struct g_raid_volume *vol;
	struct g_raid_subdisk *sd;
	struct g_raid_disk *disk;
	struct g_raid_md_nvidia_object *mdi;
	struct g_raid_md_nvidia_perdisk *pd;
	struct g_consumer *cp;
	struct g_provider *pp;
	char arg[16];
	const char *verb, *volname, *levelname, *diskname;
	int *nargs, *force;
	off_t size, sectorsize, strip, volsize;
	intmax_t *sizearg, *striparg;
	int numdisks, i, len, level, qual, update;
	int error;

	sc = md->mdo_softc;
	mdi = (struct g_raid_md_nvidia_object *)md;
	verb = gctl_get_param(req, "verb", NULL);
	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
	error = 0;
	if (strcmp(verb, "label") == 0) {

		if (*nargs < 4) {
			gctl_error(req, "Invalid number of arguments.");
			return (-1);
		}
		volname = gctl_get_asciiparam(req, "arg1");
		if (volname == NULL) {
			gctl_error(req, "No volume name.");
			return (-2);
		}
		levelname = gctl_get_asciiparam(req, "arg2");
		if (levelname == NULL) {
			gctl_error(req, "No RAID level.");
			return (-3);
		}
		if (strcasecmp(levelname, "RAID5") == 0)
			levelname = "RAID5-LS";
		if (g_raid_volume_str2level(levelname, &level, &qual)) {
			gctl_error(req, "Unknown RAID level '%s'.", levelname);
			return (-4);
		}
		numdisks = *nargs - 3;
		force = gctl_get_paraml(req, "force", sizeof(*force));
		if (!g_raid_md_nvidia_supported(level, qual, numdisks,
		    force ? *force : 0)) {
			gctl_error(req, "Unsupported RAID level "
			    "(0x%02x/0x%02x), or number of disks (%d).",
			    level, qual, numdisks);
			return (-5);
		}

		/* Search for disks, connect them and probe. */
		size = 0x7fffffffffffffffllu;
		sectorsize = 0;
		for (i = 0; i < numdisks; i++) {
			snprintf(arg, sizeof(arg), "arg%d", i + 3);
			diskname = gctl_get_asciiparam(req, arg);
			if (diskname == NULL) {
				gctl_error(req, "No disk name (%s).", arg);
				error = -6;
				break;
			}
			if (strcmp(diskname, "NONE") == 0) {
				cp = NULL;
				pp = NULL;
			} else {
				g_topology_lock();
				cp = g_raid_open_consumer(sc, diskname);
				if (cp == NULL) {
					gctl_error(req, "Can't open '%s'.",
					    diskname);
					g_topology_unlock();
					error = -7;
					break;
				}
				pp = cp->provider;
			}
			pd = malloc(sizeof(*pd), M_MD_NVIDIA, M_WAITOK | M_ZERO);
			pd->pd_disk_pos = i;
			disk = g_raid_create_disk(sc);
			disk->d_md_data = (void *)pd;
			disk->d_consumer = cp;
			if (cp == NULL)
				continue;
			cp->private = disk;
			g_topology_unlock();

			/* Read kernel dumping information. */
			disk->d_kd.offset = 0;
			disk->d_kd.length = OFF_MAX;
			len = sizeof(disk->d_kd);
			g_io_getattr("GEOM::kerneldump", cp, &len, &disk->d_kd);
			if (disk->d_kd.di.dumper == NULL)
				G_RAID_DEBUG1(2, sc,
				    "Dumping not supported by %s.",
				    cp->provider->name);

			pd->pd_disk_size = pp->mediasize;
			if (size > pp->mediasize)
				size = pp->mediasize;
			if (sectorsize < pp->sectorsize)
				sectorsize = pp->sectorsize;
		}
		if (error != 0)
			return (error);

		if (sectorsize <= 0) {
			gctl_error(req, "Can't get sector size.");
			return (-8);
		}

		/* Reserve space for metadata. */
		size -= 2 * sectorsize;

		/* Handle size argument. */
		len = sizeof(*sizearg);
		sizearg = gctl_get_param(req, "size", &len);
		if (sizearg != NULL && len == sizeof(*sizearg) &&
		    *sizearg > 0) {
			if (*sizearg > size) {
				gctl_error(req, "Size too big %lld > %lld.",
				    (long long)*sizearg, (long long)size);
				return (-9);
			}
			size = *sizearg;
		}

		/* Handle strip argument. */
		strip = 131072;
		len = sizeof(*striparg);
		striparg = gctl_get_param(req, "strip", &len);
		if (striparg != NULL && len == sizeof(*striparg) &&
		    *striparg > 0) {
			if (*striparg < sectorsize) {
				gctl_error(req, "Strip size too small.");
				return (-10);
			}
			if (*striparg % sectorsize != 0) {
				gctl_error(req, "Incorrect strip size.");
				return (-11);
			}
			if (strip > 65535 * sectorsize) {
				gctl_error(req, "Strip size too big.");
				return (-12);
			}
			strip = *striparg;
		}

		/* Round size down to strip or sector. */
		if (level == G_RAID_VOLUME_RL_RAID1)
			size -= (size % sectorsize);
		else if (level == G_RAID_VOLUME_RL_RAID1E &&
		    (numdisks & 1) != 0)
			size -= (size % (2 * strip));
		else
			size -= (size % strip);
		if (size <= 0) {
			gctl_error(req, "Size too small.");
			return (-13);
		}

		if (level == G_RAID_VOLUME_RL_RAID0 ||
		    level == G_RAID_VOLUME_RL_CONCAT ||
		    level == G_RAID_VOLUME_RL_SINGLE)
			volsize = size * numdisks;
		else if (level == G_RAID_VOLUME_RL_RAID1)
			volsize = size;
		else if (level == G_RAID_VOLUME_RL_RAID5)
			volsize = size * (numdisks - 1);
		else { /* RAID1E */
			volsize = ((size * numdisks) / strip / 2) *
			    strip;
		}
		if (volsize > 0xffffffffllu * sectorsize) {
			gctl_error(req, "Size too big.");
			return (-14);
		}

		/* We have all we need, create things: volume, ... */
		mdi->mdio_total_disks = numdisks;
		mdi->mdio_started = 1;
		vol = g_raid_create_volume(sc, volname, -1);
		vol->v_md_data = (void *)(intptr_t)0;
		vol->v_raid_level = level;
		vol->v_raid_level_qualifier = qual;
		vol->v_strip_size = strip;
		vol->v_disks_count = numdisks;
		vol->v_mediasize = volsize;
		vol->v_sectorsize = sectorsize;
		g_raid_start_volume(vol);

		/* , and subdisks. */
		TAILQ_FOREACH(disk, &sc->sc_disks, d_next) {
			pd = (struct g_raid_md_nvidia_perdisk *)disk->d_md_data;
			sd = &vol->v_subdisks[pd->pd_disk_pos];
			sd->sd_disk = disk;
			sd->sd_offset = 0;
			sd->sd_size = size;
			TAILQ_INSERT_TAIL(&disk->d_subdisks, sd, sd_next);
			if (sd->sd_disk->d_consumer != NULL) {
				g_raid_change_disk_state(disk,
				    G_RAID_DISK_S_ACTIVE);
				g_raid_change_subdisk_state(sd,
				    G_RAID_SUBDISK_S_ACTIVE);
				g_raid_event_send(sd, G_RAID_SUBDISK_E_NEW,
				    G_RAID_EVENT_SUBDISK);
			} else {
				g_raid_change_disk_state(disk, G_RAID_DISK_S_OFFLINE);
			}
		}

		/* Write metadata based on created entities. */
		G_RAID_DEBUG1(0, sc, "Array started.");
		g_raid_md_write_nvidia(md, NULL, NULL, NULL);

		/* Pickup any STALE/SPARE disks to refill array if needed. */
		g_raid_md_nvidia_refill(sc);

		g_raid_event_send(vol, G_RAID_VOLUME_E_START,
		    G_RAID_EVENT_VOLUME);
		return (0);
	}
	if (strcmp(verb, "delete") == 0) {

		/* Check if some volume is still open. */
		force = gctl_get_paraml(req, "force", sizeof(*force));
		if (force != NULL && *force == 0 &&
		    g_raid_nopens(sc) != 0) {
			gctl_error(req, "Some volume is still open.");
			return (-4);
		}

		TAILQ_FOREACH(disk, &sc->sc_disks, d_next) {
			if (disk->d_consumer)
				nvidia_meta_erase(disk->d_consumer);
		}
		g_raid_destroy_node(sc, 0);
		return (0);
	}
	if (strcmp(verb, "remove") == 0 ||
	    strcmp(verb, "fail") == 0) {
		if (*nargs < 2) {
			gctl_error(req, "Invalid number of arguments.");
			return (-1);
		}
		for (i = 1; i < *nargs; i++) {
			snprintf(arg, sizeof(arg), "arg%d", i);
			diskname = gctl_get_asciiparam(req, arg);
			if (diskname == NULL) {
				gctl_error(req, "No disk name (%s).", arg);
				error = -2;
				break;
			}
			if (strncmp(diskname, "/dev/", 5) == 0)
				diskname += 5;

			TAILQ_FOREACH(disk, &sc->sc_disks, d_next) {
				if (disk->d_consumer != NULL &&
				    disk->d_consumer->provider != NULL &&
				    strcmp(disk->d_consumer->provider->name,
				     diskname) == 0)
					break;
			}
			if (disk == NULL) {
				gctl_error(req, "Disk '%s' not found.",
				    diskname);
				error = -3;
				break;
			}

			if (strcmp(verb, "fail") == 0) {
				g_raid_md_fail_disk_nvidia(md, NULL, disk);
				continue;
			}

			pd = (struct g_raid_md_nvidia_perdisk *)disk->d_md_data;

			/* Erase metadata on deleting disk. */
			nvidia_meta_erase(disk->d_consumer);

			/* If disk was assigned, just update statuses. */
			if (pd->pd_disk_pos >= 0) {
				g_raid_change_disk_state(disk, G_RAID_DISK_S_OFFLINE);
				g_raid_kill_consumer(sc, disk->d_consumer);
				disk->d_consumer = NULL;
				TAILQ_FOREACH(sd, &disk->d_subdisks, sd_next) {
					g_raid_change_subdisk_state(sd,
					    G_RAID_SUBDISK_S_NONE);
					g_raid_event_send(sd, G_RAID_SUBDISK_E_DISCONNECTED,
					    G_RAID_EVENT_SUBDISK);
				}
			} else {
				/* Otherwise -- delete. */
				g_raid_change_disk_state(disk, G_RAID_DISK_S_NONE);
				g_raid_destroy_disk(disk);
			}
		}

		/* Write updated metadata to remaining disks. */
		g_raid_md_write_nvidia(md, NULL, NULL, NULL);

		/* Check if anything left except placeholders. */
		if (g_raid_ndisks(sc, -1) ==
		    g_raid_ndisks(sc, G_RAID_DISK_S_OFFLINE))
			g_raid_destroy_node(sc, 0);
		else
			g_raid_md_nvidia_refill(sc);
		return (error);
	}
	if (strcmp(verb, "insert") == 0) {
		if (*nargs < 2) {
			gctl_error(req, "Invalid number of arguments.");
			return (-1);
		}
		update = 0;
		for (i = 1; i < *nargs; i++) {
			/* Get disk name. */
			snprintf(arg, sizeof(arg), "arg%d", i);
			diskname = gctl_get_asciiparam(req, arg);
			if (diskname == NULL) {
				gctl_error(req, "No disk name (%s).", arg);
				error = -3;
				break;
			}

			/* Try to find provider with specified name. */
			g_topology_lock();
			cp = g_raid_open_consumer(sc, diskname);
			if (cp == NULL) {
				gctl_error(req, "Can't open disk '%s'.",
				    diskname);
				g_topology_unlock();
				error = -4;
				break;
			}
			pp = cp->provider;

			pd = malloc(sizeof(*pd), M_MD_NVIDIA, M_WAITOK | M_ZERO);
			pd->pd_disk_pos = -3;
			pd->pd_disk_size = pp->mediasize;

			disk = g_raid_create_disk(sc);
			disk->d_consumer = cp;
			disk->d_md_data = (void *)pd;
			cp->private = disk;
			g_topology_unlock();

			/* Read kernel dumping information. */
			disk->d_kd.offset = 0;
			disk->d_kd.length = OFF_MAX;
			len = sizeof(disk->d_kd);
			g_io_getattr("GEOM::kerneldump", cp, &len, &disk->d_kd);
			if (disk->d_kd.di.dumper == NULL)
				G_RAID_DEBUG1(2, sc,
				    "Dumping not supported by %s.",
				    cp->provider->name);

			/* Welcome the "new" disk. */
			update += g_raid_md_nvidia_start_disk(disk);
			if (disk->d_state != G_RAID_DISK_S_SPARE &&
			    disk->d_state != G_RAID_DISK_S_ACTIVE) {
				gctl_error(req, "Disk '%s' doesn't fit.",
				    diskname);
				g_raid_destroy_disk(disk);
				error = -8;
				break;
			}
		}

		/* Write new metadata if we changed something. */
		if (update)
			g_raid_md_write_nvidia(md, NULL, NULL, NULL);
		return (error);
	}
	gctl_error(req, "Command '%s' is not supported.", verb);
	return (-100);
}

static int
g_raid_md_write_nvidia(struct g_raid_md_object *md, struct g_raid_volume *tvol,
    struct g_raid_subdisk *tsd, struct g_raid_disk *tdisk)
{
	struct g_raid_softc *sc;
	struct g_raid_volume *vol;
	struct g_raid_subdisk *sd;
	struct g_raid_disk *disk;
	struct g_raid_md_nvidia_object *mdi;
	struct g_raid_md_nvidia_perdisk *pd;
	struct nvidia_raid_conf *meta;
	int i, spares;

	sc = md->mdo_softc;
	mdi = (struct g_raid_md_nvidia_object *)md;

	if (sc->sc_stopping == G_RAID_DESTROY_HARD)
		return (0);

	/* There is only one volume. */
	vol = TAILQ_FIRST(&sc->sc_volumes);

	/* Fill global fields. */
	meta = malloc(sizeof(*meta), M_MD_NVIDIA, M_WAITOK | M_ZERO);
	if (mdi->mdio_meta)
		memcpy(meta, mdi->mdio_meta, sizeof(*meta));
	memcpy(meta->nvidia_id, NVIDIA_MAGIC, sizeof(NVIDIA_MAGIC) - 1);
	meta->config_size = 30;
	meta->version = 0x0064;
	meta->total_sectors = vol->v_mediasize / vol->v_sectorsize;
	meta->sector_size = vol->v_sectorsize;
	nvidia_meta_put_name(meta, vol->v_name);
	meta->magic_0 = NVIDIA_MAGIC0;
	memcpy(&meta->volume_id, &mdi->mdio_volume_id, 16);
	meta->state = NVIDIA_S_IDLE;
	if (vol->v_raid_level == G_RAID_VOLUME_RL_RAID1)
		meta->array_width = 1;
	else if (vol->v_raid_level == G_RAID_VOLUME_RL_RAID1E)
		meta->array_width = vol->v_disks_count / 2;
	else if (vol->v_raid_level == G_RAID_VOLUME_RL_RAID5)
		meta->array_width = vol->v_disks_count - 1;
	else
		meta->array_width = vol->v_disks_count;
	meta->total_disks = vol->v_disks_count;
	meta->orig_array_width = meta->array_width;
	if (vol->v_raid_level == G_RAID_VOLUME_RL_RAID0)
		meta->type = NVIDIA_T_RAID0;
	else if (vol->v_raid_level == G_RAID_VOLUME_RL_RAID1)
		meta->type = NVIDIA_T_RAID1;
	else if (vol->v_raid_level == G_RAID_VOLUME_RL_RAID1E)
		meta->type = NVIDIA_T_RAID01;
	else if (vol->v_raid_level == G_RAID_VOLUME_RL_CONCAT ||
	    vol->v_raid_level == G_RAID_VOLUME_RL_SINGLE)
		meta->type = NVIDIA_T_CONCAT;
	else if (vol->v_raid_level_qualifier == G_RAID_VOLUME_RLQ_R5LA)
		meta->type = NVIDIA_T_RAID5;
	else
		meta->type = NVIDIA_T_RAID5_SYM;
	meta->strip_sectors = vol->v_strip_size / vol->v_sectorsize;
	meta->strip_bytes = vol->v_strip_size;
	meta->strip_shift = ffs(meta->strip_sectors) - 1;
	meta->strip_mask = meta->strip_sectors - 1;
	meta->stripe_sectors = meta->strip_sectors * meta->orig_array_width;
	meta->stripe_bytes = meta->stripe_sectors * vol->v_sectorsize;
	meta->rebuild_lba = 0;
	meta->orig_type = meta->type;
	meta->orig_total_sectors = meta->total_sectors;
	meta->status = 0;

	for (i = 0; i < vol->v_disks_count; i++) {
		sd = &vol->v_subdisks[i];
		if ((sd->sd_state == G_RAID_SUBDISK_S_STALE ||
		     sd->sd_state == G_RAID_SUBDISK_S_RESYNC ||
		     vol->v_dirty) &&
		     meta->state != NVIDIA_S_REBUILD)
			meta->state = NVIDIA_S_SYNC;
		else if (sd->sd_state == G_RAID_SUBDISK_S_NEW ||
		     sd->sd_state == G_RAID_SUBDISK_S_REBUILD)
			meta->state = NVIDIA_S_REBUILD;
	}

	/* We are done. Print meta data and store them to disks. */
	if (mdi->mdio_meta != NULL)
		free(mdi->mdio_meta, M_MD_NVIDIA);
	mdi->mdio_meta = meta;
	spares = 0;
	TAILQ_FOREACH(disk, &sc->sc_disks, d_next) {
		pd = (struct g_raid_md_nvidia_perdisk *)disk->d_md_data;
		if (disk->d_state != G_RAID_DISK_S_ACTIVE &&
		    disk->d_state != G_RAID_DISK_S_SPARE)
			continue;
		if (pd->pd_meta != NULL) {
			free(pd->pd_meta, M_MD_NVIDIA);
			pd->pd_meta = NULL;
		}
		pd->pd_meta = nvidia_meta_copy(meta);
		if ((sd = TAILQ_FIRST(&disk->d_subdisks)) != NULL) {
			/* For RAID0+1 we need to translate order. */
			pd->pd_meta->disk_number =
			    nvidia_meta_translate_disk(meta, sd->sd_pos);
			if (sd->sd_state != G_RAID_SUBDISK_S_ACTIVE) {
				pd->pd_meta->disk_status = 0x100;
				pd->pd_meta->rebuild_lba =
				    sd->sd_rebuild_pos / vol->v_sectorsize *
				    meta->array_width;
			}
		} else
			pd->pd_meta->disk_number = meta->total_disks + spares++;
		G_RAID_DEBUG(1, "Writing NVIDIA metadata to %s",
		    g_raid_get_diskname(disk));
		g_raid_md_nvidia_print(pd->pd_meta);
		nvidia_meta_write(disk->d_consumer, pd->pd_meta);
	}
	return (0);
}

static int
g_raid_md_fail_disk_nvidia(struct g_raid_md_object *md,
    struct g_raid_subdisk *tsd, struct g_raid_disk *tdisk)
{
	struct g_raid_softc *sc;
	struct g_raid_md_nvidia_perdisk *pd;
	struct g_raid_subdisk *sd;

	sc = md->mdo_softc;
	pd = (struct g_raid_md_nvidia_perdisk *)tdisk->d_md_data;

	/* We can't fail disk that is not a part of array now. */
	if (pd->pd_disk_pos < 0)
		return (-1);

	/* Erase metadata to prevent disks's later resurrection. */
	if (tdisk->d_consumer)
		nvidia_meta_erase(tdisk->d_consumer);

	/* Change states. */
	g_raid_change_disk_state(tdisk, G_RAID_DISK_S_FAILED);
	TAILQ_FOREACH(sd, &tdisk->d_subdisks, sd_next) {
		g_raid_change_subdisk_state(sd,
		    G_RAID_SUBDISK_S_FAILED);
		g_raid_event_send(sd, G_RAID_SUBDISK_E_FAILED,
		    G_RAID_EVENT_SUBDISK);
	}

	/* Write updated metadata to remaining disks. */
	g_raid_md_write_nvidia(md, NULL, NULL, tdisk);

	/* Check if anything left except placeholders. */
	if (g_raid_ndisks(sc, -1) ==
	    g_raid_ndisks(sc, G_RAID_DISK_S_OFFLINE))
		g_raid_destroy_node(sc, 0);
	else
		g_raid_md_nvidia_refill(sc);
	return (0);
}

static int
g_raid_md_free_disk_nvidia(struct g_raid_md_object *md,
    struct g_raid_disk *disk)
{
	struct g_raid_md_nvidia_perdisk *pd;

	pd = (struct g_raid_md_nvidia_perdisk *)disk->d_md_data;
	if (pd->pd_meta != NULL) {
		free(pd->pd_meta, M_MD_NVIDIA);
		pd->pd_meta = NULL;
	}
	free(pd, M_MD_NVIDIA);
	disk->d_md_data = NULL;
	return (0);
}

static int
g_raid_md_free_nvidia(struct g_raid_md_object *md)
{
	struct g_raid_md_nvidia_object *mdi;

	mdi = (struct g_raid_md_nvidia_object *)md;
	if (!mdi->mdio_started) {
		mdi->mdio_started = 0;
		callout_stop(&mdi->mdio_start_co);
		G_RAID_DEBUG1(1, md->mdo_softc,
		    "root_mount_rel %p", mdi->mdio_rootmount);
		root_mount_rel(mdi->mdio_rootmount);
		mdi->mdio_rootmount = NULL;
	}
	if (mdi->mdio_meta != NULL) {
		free(mdi->mdio_meta, M_MD_NVIDIA);
		mdi->mdio_meta = NULL;
	}
	return (0);
}

G_RAID_MD_DECLARE(g_raid_md_nvidia);