A GEOM cache can speed up read performance by sending fixed size

read requests to its consumer.  It has been developed to address
the problem of a horrible read performance of a 64k blocksize FS
residing on a RAID3 array with 8 data components, where a single
disk component would only get 8k read requests, thus effectively
killing disk performance under high load.  Documentation will be
provided later.  I'd like to thank Vsevolod Lobko for his bright
ideas, and Pawel Jakub Dawidek for helping me fix the nasty bug.

1 files changed, 1037 insertions, 0 deletions

diff --git a/sys/geom/cache/g_cache.c b/sys/geom/cache/g_cache.c
new file mode 100644
index 000000000000..60b0254e1a73
--- /dev/null
+++ b/sys/geom/cache/g_cache.c
@@ -0,0 +1,1037 @@
+/*-
+ * Copyright (c) 2006 Ruslan Ermilov <ru@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 AUTHOR 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 AUTHOR 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/queue.h>
+#include <sys/time.h>
+#include <vm/uma.h>
+#include <geom/geom.h>
+#include <geom/cache/g_cache.h>
+
+static MALLOC_DEFINE(M_GCACHE, "gcache_data", "GEOM_CACHE Data");
+
+SYSCTL_DECL(_kern_geom);
+SYSCTL_NODE(_kern_geom, OID_AUTO, cache, CTLFLAG_RW, 0, "GEOM_CACHE stuff");
+static u_int g_cache_debug = 0;
+SYSCTL_UINT(_kern_geom_cache, OID_AUTO, debug, CTLFLAG_RW, &g_cache_debug, 0,
+    "Debug level");
+static u_int g_cache_enable = 1;
+SYSCTL_UINT(_kern_geom_cache, OID_AUTO, enable, CTLFLAG_RW, &g_cache_enable, 0,
+    "");
+static u_int g_cache_timeout = 10;
+SYSCTL_UINT(_kern_geom_cache, OID_AUTO, timeout, CTLFLAG_RW, &g_cache_timeout,
+    0, "");
+static u_int g_cache_idletime = 5;
+SYSCTL_UINT(_kern_geom_cache, OID_AUTO, idletime, CTLFLAG_RW, &g_cache_idletime,
+    0, "");
+static u_int g_cache_used_lo = 5;
+static u_int g_cache_used_hi = 20;
+static int
+sysctl_handle_pct(SYSCTL_HANDLER_ARGS)
+{
+	u_int val = *(u_int *)arg1;
+	int error;
+
+	error = sysctl_handle_int(oidp, &val, sizeof(val), req);
+	if (error || !req->newptr)
+		return (error);
+	if (val < 0 || val > 100)
+		return (EINVAL);
+	if ((arg1 == &g_cache_used_lo && val > g_cache_used_hi) ||
+	    (arg1 == &g_cache_used_hi && g_cache_used_lo > val))
+		return (EINVAL);
+	*(u_int *)arg1 = val;
+	return (0);
+}
+SYSCTL_PROC(_kern_geom_cache, OID_AUTO, used_lo, CTLTYPE_UINT|CTLFLAG_RW,
+	&g_cache_used_lo, 0, sysctl_handle_pct, "IU", "");
+SYSCTL_PROC(_kern_geom_cache, OID_AUTO, used_hi, CTLTYPE_UINT|CTLFLAG_RW,
+	&g_cache_used_hi, 0, sysctl_handle_pct, "IU", "");
+
+
+static int g_cache_destroy(struct g_cache_softc *sc, boolean_t force);
+static g_ctl_destroy_geom_t g_cache_destroy_geom;
+
+static g_taste_t g_cache_taste;
+static g_ctl_req_t g_cache_config;
+static g_dumpconf_t g_cache_dumpconf;
+
+struct g_class g_cache_class = {
+	.name = G_CACHE_CLASS_NAME,
+	.version = G_VERSION,
+	.ctlreq = g_cache_config,
+	.taste = g_cache_taste,
+	.destroy_geom = g_cache_destroy_geom
+};
+
+#define	OFF2BNO(off, sc)	((off) >> (sc)->sc_bshift)
+#define	BNO2OFF(bno, sc)	((bno) << (sc)->sc_bshift)
+
+
+static struct g_cache_desc *
+g_cache_alloc(struct g_cache_softc *sc)
+{
+	struct g_cache_desc *dp;
+
+	mtx_assert(&sc->sc_mtx, MA_OWNED);
+
+	if (!TAILQ_EMPTY(&sc->sc_usedlist)) {
+		dp = TAILQ_FIRST(&sc->sc_usedlist);
+		TAILQ_REMOVE(&sc->sc_usedlist, dp, d_used);
+		sc->sc_nused--;
+		dp->d_flags = 0;
+		LIST_REMOVE(dp, d_next);
+		return (dp);
+	}
+	if (sc->sc_nent > sc->sc_maxent) {
+		sc->sc_cachefull++;
+		return (NULL);
+	}
+	dp = malloc(sizeof(*dp), M_GCACHE, M_NOWAIT | M_ZERO);
+	if (dp == NULL)
+		return (NULL);
+	dp->d_data = uma_zalloc(sc->sc_zone, M_NOWAIT);
+	if (dp->d_data == NULL) {
+		free(dp, M_GCACHE);
+		return (NULL);
+	}
+	sc->sc_nent++;
+	return (dp);
+}
+
+static void
+g_cache_free(struct g_cache_softc *sc, struct g_cache_desc *dp)
+{
+
+	mtx_assert(&sc->sc_mtx, MA_OWNED);
+
+	uma_zfree(sc->sc_zone, dp->d_data);
+	free(dp, M_GCACHE);
+	sc->sc_nent--;
+}
+
+static void
+g_cache_free_used(struct g_cache_softc *sc)
+{
+	struct g_cache_desc *dp;
+	u_int n;
+
+	mtx_assert(&sc->sc_mtx, MA_OWNED);
+
+	n = g_cache_used_lo * sc->sc_maxent / 100;
+	while (sc->sc_nused > n) {
+		KASSERT(!TAILQ_EMPTY(&sc->sc_usedlist), ("used list empty"));
+		dp = TAILQ_FIRST(&sc->sc_usedlist);
+		TAILQ_REMOVE(&sc->sc_usedlist, dp, d_used);
+		sc->sc_nused--;
+		LIST_REMOVE(dp, d_next);
+		g_cache_free(sc, dp);
+	}
+}
+
+static void
+g_cache_deliver(struct g_cache_softc *sc, struct bio *bp,
+    struct g_cache_desc *dp, int error)
+{
+	off_t off1, off, len;
+
+	mtx_assert(&sc->sc_mtx, MA_OWNED);
+	KASSERT(OFF2BNO(bp->bio_offset, sc) <= dp->d_bno, ("wrong entry"));
+	KASSERT(OFF2BNO(bp->bio_offset + bp->bio_length - 1, sc) >=
+	    dp->d_bno, ("wrong entry"));
+
+	off1 = BNO2OFF(dp->d_bno, sc);
+	off = MAX(bp->bio_offset, off1);
+	len = MIN(bp->bio_offset + bp->bio_length, off1 + sc->sc_bsize) - off;
+
+	if (bp->bio_error == 0)
+		bp->bio_error = error;
+	if (bp->bio_error == 0) {
+		bcopy(dp->d_data + (off - off1),
+		    bp->bio_data + (off - bp->bio_offset), len);
+	}
+	bp->bio_completed += len;
+	KASSERT(bp->bio_completed <= bp->bio_length, ("extra data"));
+	if (bp->bio_completed == bp->bio_length) {
+		if (bp->bio_error != 0)
+			bp->bio_completed = 0;
+		g_io_deliver(bp, bp->bio_error);
+	}
+
+	if (dp->d_flags & D_FLAG_USED) {
+		TAILQ_REMOVE(&sc->sc_usedlist, dp, d_used);
+		TAILQ_INSERT_TAIL(&sc->sc_usedlist, dp, d_used);
+	} else if (OFF2BNO(off + len, sc) > dp->d_bno) {
+		TAILQ_INSERT_TAIL(&sc->sc_usedlist, dp, d_used);
+		sc->sc_nused++;
+		dp->d_flags |= D_FLAG_USED;
+	}
+	dp->d_atime = time_uptime;
+}
+
+static void
+g_cache_done(struct bio *bp)
+{
+	struct g_cache_softc *sc;
+	struct g_cache_desc *dp;
+	struct bio *bp2, *tmpbp;
+
+	sc = bp->bio_from->geom->softc;
+	KASSERT(G_CACHE_DESC1(bp) == sc, ("corrupt bio_caller in g_cache_done()"));
+	dp = G_CACHE_DESC2(bp);
+	mtx_lock(&sc->sc_mtx);
+	bp2 = dp->d_biolist;
+	while (bp2 != NULL) {
+		KASSERT(G_CACHE_NEXT_BIO1(bp2) == sc, ("corrupt bio_driver in g_cache_done()"));
+		tmpbp = G_CACHE_NEXT_BIO2(bp2);
+		g_cache_deliver(sc, bp2, dp, bp->bio_error);
+		bp2 = tmpbp;
+	}
+	dp->d_biolist = NULL;
+	if (dp->d_flags & D_FLAG_INVALID) {
+		sc->sc_invalid--;
+		g_cache_free(sc, dp);
+	} else if (bp->bio_error) {
+		LIST_REMOVE(dp, d_next);
+		if (dp->d_flags & D_FLAG_USED) {
+			TAILQ_REMOVE(&sc->sc_usedlist, dp, d_used);
+			sc->sc_nused--;
+		}
+		g_cache_free(sc, dp);
+	}
+	mtx_unlock(&sc->sc_mtx);
+	g_destroy_bio(bp);
+}
+
+static struct g_cache_desc *
+g_cache_lookup(struct g_cache_softc *sc, off_t bno)
+{
+	struct g_cache_desc *dp;
+
+	mtx_assert(&sc->sc_mtx, MA_OWNED);
+
+	LIST_FOREACH(dp, &sc->sc_desclist[G_CACHE_BUCKET(bno)], d_next)
+		if (dp->d_bno == bno)
+			return (dp);
+	return (NULL);
+}
+
+static int
+g_cache_read(struct g_cache_softc *sc, struct bio *bp)
+{
+	struct bio *cbp;
+	struct g_cache_desc *dp;
+
+	mtx_lock(&sc->sc_mtx);
+	dp = g_cache_lookup(sc,
+	    OFF2BNO(bp->bio_offset + bp->bio_completed, sc));
+	if (dp != NULL) {
+		/* Add to waiters list or deliver. */
+		sc->sc_cachehits++;
+		if (dp->d_biolist != NULL) {
+			G_CACHE_NEXT_BIO1(bp) = sc;
+			G_CACHE_NEXT_BIO2(bp) = dp->d_biolist;
+			dp->d_biolist = bp;
+		} else
+			g_cache_deliver(sc, bp, dp, 0);
+		mtx_unlock(&sc->sc_mtx);
+		return (0);
+	}
+
+	/* Cache miss.  Allocate entry and schedule bio.  */
+	sc->sc_cachemisses++;
+	dp = g_cache_alloc(sc);
+	if (dp == NULL) {
+		mtx_unlock(&sc->sc_mtx);
+		return (ENOMEM);
+	}
+	cbp = g_clone_bio(bp);
+	if (cbp == NULL) {
+		g_cache_free(sc, dp);
+		mtx_unlock(&sc->sc_mtx);
+		return (ENOMEM);
+	}
+
+	dp->d_bno = OFF2BNO(bp->bio_offset + bp->bio_completed, sc);
+	G_CACHE_NEXT_BIO1(bp) = sc;
+	G_CACHE_NEXT_BIO2(bp) = NULL;
+	dp->d_biolist = bp;
+	LIST_INSERT_HEAD(&sc->sc_desclist[G_CACHE_BUCKET(dp->d_bno)],
+	    dp, d_next);
+	mtx_unlock(&sc->sc_mtx);
+
+	G_CACHE_DESC1(cbp) = sc;
+	G_CACHE_DESC2(cbp) = dp;
+	cbp->bio_done = g_cache_done;
+	cbp->bio_offset = BNO2OFF(dp->d_bno, sc);
+	cbp->bio_data = dp->d_data;
+	cbp->bio_length = sc->sc_bsize;
+	g_io_request(cbp, LIST_FIRST(&bp->bio_to->geom->consumer));
+	return (0);
+}
+
+static void
+g_cache_invalidate(struct g_cache_softc *sc, struct bio *bp)
+{
+	struct g_cache_desc *dp;
+	off_t bno, lim;
+
+	mtx_lock(&sc->sc_mtx);
+	bno = OFF2BNO(bp->bio_offset, sc);
+	lim = OFF2BNO(bp->bio_offset + bp->bio_length - 1, sc);
+	do {
+		if ((dp = g_cache_lookup(sc, bno)) != NULL) {
+			LIST_REMOVE(dp, d_next);
+			if (dp->d_flags & D_FLAG_USED) {
+				TAILQ_REMOVE(&sc->sc_usedlist, dp, d_used);
+				sc->sc_nused--;
+			}
+			if (dp->d_biolist == NULL)
+				g_cache_free(sc, dp);
+			else {
+				dp->d_flags = D_FLAG_INVALID;
+				sc->sc_invalid++;
+			}
+		}
+		bno++;
+	} while (bno <= lim);
+	mtx_unlock(&sc->sc_mtx);
+}
+
+static void
+g_cache_start(struct bio *bp)
+{
+	struct g_cache_softc *sc;
+	struct g_geom *gp;
+	struct g_cache_desc *dp;
+	struct bio *cbp;
+
+	gp = bp->bio_to->geom;
+	sc = gp->softc;
+	G_CACHE_LOGREQ(bp, "Request received.");
+	switch (bp->bio_cmd) {
+	case BIO_READ:
+		sc->sc_reads++;
+		sc->sc_readbytes += bp->bio_length;
+		if (!g_cache_enable)
+			break;
+		if (bp->bio_offset + bp->bio_length > sc->sc_tail)
+			break;
+		if (OFF2BNO(bp->bio_offset, sc) ==
+		    OFF2BNO(bp->bio_offset + bp->bio_length - 1, sc)) {
+			sc->sc_cachereads++;
+			sc->sc_cachereadbytes += bp->bio_length;
+			if (g_cache_read(sc, bp) == 0)
+				return;
+			sc->sc_cachereads--;
+			sc->sc_cachereadbytes -= bp->bio_length;
+			break;
+		} else if (OFF2BNO(bp->bio_offset, sc) + 1 ==
+		    OFF2BNO(bp->bio_offset + bp->bio_length - 1, sc)) {
+			mtx_lock(&sc->sc_mtx);
+			dp = g_cache_lookup(sc, OFF2BNO(bp->bio_offset, sc));
+			if (dp == NULL || dp->d_biolist != NULL) {
+				mtx_unlock(&sc->sc_mtx);
+				break;
+			}
+			sc->sc_cachereads++;
+			sc->sc_cachereadbytes += bp->bio_length;
+			g_cache_deliver(sc, bp, dp, 0);
+			mtx_unlock(&sc->sc_mtx);
+			if (g_cache_read(sc, bp) == 0)
+				return;
+			sc->sc_cachereads--;
+			sc->sc_cachereadbytes -= bp->bio_length;
+			break;
+		}
+		break;
+	case BIO_WRITE:
+		sc->sc_writes++;
+		sc->sc_wrotebytes += bp->bio_length;
+		g_cache_invalidate(sc, bp);
+		break;
+	}
+	cbp = g_clone_bio(bp);
+	if (cbp == NULL) {
+		g_io_deliver(bp, ENOMEM);
+		return;
+	}
+	cbp->bio_done = g_std_done;
+	G_CACHE_LOGREQ(cbp, "Sending request.");
+	g_io_request(cbp, LIST_FIRST(&gp->consumer));
+}
+
+static void
+g_cache_go(void *arg)
+{
+	struct g_cache_softc *sc = arg;
+	struct g_cache_desc *dp;
+	int i;
+
+	mtx_assert(&sc->sc_mtx, MA_OWNED);
+
+	/* Forcibly mark idle ready entries as used. */
+	for (i = 0; i < G_CACHE_BUCKETS; i++) {
+		LIST_FOREACH(dp, &sc->sc_desclist[i], d_next) {
+			if (dp->d_flags & D_FLAG_USED ||
+			    dp->d_biolist != NULL ||
+			    time_uptime - dp->d_atime < g_cache_idletime)
+				continue;
+			TAILQ_INSERT_TAIL(&sc->sc_usedlist, dp, d_used);
+			sc->sc_nused++;
+			dp->d_flags |= D_FLAG_USED;
+		}
+	}
+
+	/* Keep the number of used entries low. */
+	if (sc->sc_nused > g_cache_used_hi * sc->sc_maxent / 100)
+		g_cache_free_used(sc);
+
+	callout_reset(&sc->sc_callout, g_cache_timeout * hz, g_cache_go, sc);
+}
+
+static int
+g_cache_access(struct g_provider *pp, int dr, int dw, int de)
+{
+	struct g_geom *gp;
+	struct g_consumer *cp;
+	int error;
+
+	gp = pp->geom;
+	cp = LIST_FIRST(&gp->consumer);
+	error = g_access(cp, dr, dw, de);
+
+	return (error);
+}
+
+static void
+g_cache_orphan(struct g_consumer *cp)
+{
+
+	g_topology_assert();
+	g_cache_destroy(cp->geom->softc, 1);
+}
+
+static struct g_cache_softc *
+g_cache_find_device(struct g_class *mp, const char *name)
+{
+	struct g_geom *gp;
+
+	LIST_FOREACH(gp, &mp->geom, geom) {
+		if (strcmp(gp->name, name) == 0)
+			return (gp->softc);
+	}
+	return (NULL);
+}
+
+static struct g_geom *
+g_cache_create(struct g_class *mp, struct g_provider *pp,
+    const struct g_cache_metadata *md, u_int type)
+{
+	struct g_cache_softc *sc;
+	struct g_geom *gp;
+	struct g_provider *newpp;
+	struct g_consumer *cp;
+	u_int bshift;
+	int i;
+
+	g_topology_assert();
+
+	gp = NULL;
+	newpp = NULL;
+	cp = NULL;
+
+	G_CACHE_DEBUG(1, "Creating device %s.", md->md_name);
+
+	/* Cache size is minimum 100. */
+	if (md->md_size < 100) {
+		G_CACHE_DEBUG(0, "Invalid size for device %s.", md->md_name);
+		return (NULL);
+	}
+
+	/* Block size restrictions. */
+	bshift = ffs(md->md_bsize) - 1;
+	if (md->md_bsize == 0 || md->md_bsize > MAXPHYS ||
+	    md->md_bsize != 1 << bshift ||
+	    (md->md_bsize % pp->sectorsize) != 0) {
+		G_CACHE_DEBUG(0, "Invalid blocksize for provider %s.", pp->name);
+		return (NULL);
+	}
+
+	/* Check for duplicate unit. */
+	if (g_cache_find_device(mp, (const char *)&md->md_name) != NULL) {
+		G_CACHE_DEBUG(0, "Provider %s already exists.", md->md_name);
+		return (NULL);
+	}
+
+	gp = g_new_geomf(mp, md->md_name);
+	if (gp == NULL) {
+		G_CACHE_DEBUG(0, "Cannot create geom %s.", md->md_name);
+		return (NULL);
+	}
+	gp->softc = NULL;	/* for a moment */
+
+	sc = g_malloc(sizeof(*sc), M_WAITOK | M_ZERO);
+	sc->sc_type = type;
+	sc->sc_bshift = bshift;
+	sc->sc_bsize = 1 << bshift;
+	sc->sc_zone = uma_zcreate("gcache", sc->sc_bsize, NULL, NULL, NULL, NULL,
+	    UMA_ALIGN_PTR, 0);
+	mtx_init(&sc->sc_mtx, "GEOM CACHE mutex", NULL, MTX_DEF);
+	for (i = 0; i < G_CACHE_BUCKETS; i++)
+		LIST_INIT(&sc->sc_desclist[i]);
+	TAILQ_INIT(&sc->sc_usedlist);
+	sc->sc_maxent = md->md_size;
+	callout_init_mtx(&sc->sc_callout, &sc->sc_mtx, 0);
+	gp->softc = sc;
+	sc->sc_geom = gp;
+	gp->start = g_cache_start;
+	gp->orphan = g_cache_orphan;
+	gp->access = g_cache_access;
+	gp->dumpconf = g_cache_dumpconf;
+
+	newpp = g_new_providerf(gp, "cache/%s", gp->name);
+	if (newpp == NULL) {
+		G_CACHE_DEBUG(0, "Cannot create provider cache/%s.", gp->name);
+		goto fail;
+	}
+	newpp->sectorsize = pp->sectorsize;
+	newpp->mediasize = pp->mediasize;
+	if (type == G_CACHE_TYPE_AUTOMATIC)
+		newpp->mediasize -= pp->sectorsize;
+	sc->sc_tail = BNO2OFF(OFF2BNO(newpp->mediasize, sc), sc);
+
+	cp = g_new_consumer(gp);
+	if (cp == NULL) {
+		G_CACHE_DEBUG(0, "Cannot create consumer for %s.", gp->name);
+		goto fail;
+	}
+	if (g_attach(cp, pp) != 0) {
+		G_CACHE_DEBUG(0, "Cannot attach to provider %s.", pp->name);
+		goto fail;
+	}
+
+	g_error_provider(newpp, 0);
+	G_CACHE_DEBUG(0, "Device %s created.", gp->name);
+	callout_reset(&sc->sc_callout, g_cache_timeout * hz, g_cache_go, sc);
+	return (gp);
+fail:
+	if (cp != NULL) {
+		if (cp->provider != NULL)
+			g_detach(cp);
+		g_destroy_consumer(cp);
+	}
+	if (newpp != NULL)
+		g_destroy_provider(newpp);
+	if (gp != NULL) {
+		if (gp->softc != NULL) {
+			mtx_destroy(&sc->sc_mtx);
+			g_free(gp->softc);
+		}
+		g_destroy_geom(gp);
+	}
+	return (NULL);
+}
+
+static int
+g_cache_destroy(struct g_cache_softc *sc, boolean_t force)
+{
+	struct g_geom *gp;
+	struct g_provider *pp;
+	struct g_cache_desc *dp, *dp2;
+	int i;
+
+	g_topology_assert();
+	if (sc == NULL)
+		return (ENXIO);
+	gp = sc->sc_geom;
+	pp = LIST_FIRST(&gp->provider);
+	if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
+		if (force) {
+			G_CACHE_DEBUG(0, "Device %s is still open, so it "
+			    "can't be definitely removed.", pp->name);
+		} else {
+			G_CACHE_DEBUG(1, "Device %s is still open (r%dw%de%d).",
+			    pp->name, pp->acr, pp->acw, pp->ace);
+			return (EBUSY);
+		}
+	} else {
+		G_CACHE_DEBUG(0, "Device %s removed.", gp->name);
+	}
+	callout_drain(&sc->sc_callout);
+	mtx_lock(&sc->sc_mtx);
+	for (i = 0; i < G_CACHE_BUCKETS; i++) {
+		dp = LIST_FIRST(&sc->sc_desclist[i]);
+		while (dp != NULL) {
+			dp2 = LIST_NEXT(dp, d_next);
+			g_cache_free(sc, dp);
+			dp = dp2;
+		}
+	}
+	mtx_unlock(&sc->sc_mtx);
+	mtx_destroy(&sc->sc_mtx);
+	uma_zdestroy(sc->sc_zone);
+	g_free(sc);
+	gp->softc = NULL;
+	g_wither_geom(gp, ENXIO);
+
+	return (0);
+}
+
+static int
+g_cache_destroy_geom(struct gctl_req *req, struct g_class *mp, struct g_geom *gp)
+{
+
+	return (g_cache_destroy(gp->softc, 0));
+}
+
+static int
+g_cache_read_metadata(struct g_consumer *cp, struct g_cache_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();
+	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
+	    &error);
+	g_topology_lock();
+	g_access(cp, -1, 0, 0);
+	if (buf == NULL)
+		return (error);
+
+	/* Decode metadata. */
+	cache_metadata_decode(buf, md);
+	g_free(buf);
+
+	return (0);
+}
+
+static int
+g_cache_write_metadata(struct g_consumer *cp, struct g_cache_metadata *md)
+{
+	struct g_provider *pp;
+	u_char *buf;
+	int error;
+
+	g_topology_assert();
+
+	error = g_access(cp, 0, 1, 0);
+	if (error != 0)
+		return (error);
+	pp = cp->provider;
+	buf = malloc((size_t)pp->sectorsize, M_GCACHE, M_WAITOK | M_ZERO);
+	cache_metadata_encode(md, buf);
+	g_topology_unlock();
+	error = g_write_data(cp, pp->mediasize - pp->sectorsize, buf, pp->sectorsize);
+	g_topology_lock();
+	g_access(cp, 0, -1, 0);
+	free(buf, M_GCACHE);
+
+	return (error);
+}
+
+static struct g_geom *
+g_cache_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
+{
+	struct g_cache_metadata md;
+	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();
+
+	G_CACHE_DEBUG(3, "Tasting %s.", pp->name);
+
+	gp = g_new_geomf(mp, "cache:taste");
+	gp->start = g_cache_start;
+	gp->orphan = g_cache_orphan;
+	gp->access = g_cache_access;
+	cp = g_new_consumer(gp);
+	g_attach(cp, pp);
+	error = g_cache_read_metadata(cp, &md);
+	g_detach(cp);
+	g_destroy_consumer(cp);
+	g_destroy_geom(gp);
+	if (error != 0)
+		return (NULL);
+
+	if (strcmp(md.md_magic, G_CACHE_MAGIC) != 0)
+		return (NULL);
+	if (md.md_version > G_CACHE_VERSION) {
+		printf("geom_cache.ko module is too old to handle %s.\n",
+		    pp->name);
+		return (NULL);
+	}
+	if (md.md_provsize != pp->mediasize)
+		return (NULL);
+
+	gp = g_cache_create(mp, pp, &md, G_CACHE_TYPE_AUTOMATIC);
+	if (gp == NULL) {
+		G_CACHE_DEBUG(0, "Can't create %s.", md.md_name);
+		return (NULL);
+	}
+	return (gp);
+}
+
+static void
+g_cache_ctl_create(struct gctl_req *req, struct g_class *mp)
+{
+	struct g_cache_metadata md;
+	struct g_provider *pp;
+	struct g_geom *gp;
+	intmax_t *bsize, *size;
+	const char *name;
+	int *nargs;
+
+	g_topology_assert();
+
+	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
+	if (nargs == NULL) {
+		gctl_error(req, "No '%s' argument", "nargs");
+		return;
+	}
+	if (*nargs != 2) {
+		gctl_error(req, "Invalid number of arguments.");
+		return;
+	}
+
+	strlcpy(md.md_magic, G_CACHE_MAGIC, sizeof(md.md_magic));
+	md.md_version = G_CACHE_VERSION;
+	name = gctl_get_asciiparam(req, "arg0");
+	if (name == NULL) {
+		gctl_error(req, "No 'arg0' argument");
+		return;
+	}
+	strlcpy(md.md_name, name, sizeof(md.md_name));
+
+	size = gctl_get_paraml(req, "size", sizeof(*size));
+	if (size == NULL) {
+		gctl_error(req, "No '%s' argument", "size");
+		return;
+	}
+	if ((u_int)*size < 100) {
+		gctl_error(req, "Invalid '%s' argument", "size");
+		return;
+	}
+	md.md_size = (u_int)*size;
+
+	bsize = gctl_get_paraml(req, "blocksize", sizeof(*bsize));
+	if (bsize == NULL) {
+		gctl_error(req, "No '%s' argument", "blocksize");
+		return;
+	}
+	if (*bsize < 0) {
+		gctl_error(req, "Invalid '%s' argument", "blocksize");
+		return;
+	}
+	md.md_bsize = (u_int)*bsize;
+
+	/* This field is not important here. */
+	md.md_provsize = 0;
+
+	name = gctl_get_asciiparam(req, "arg1");
+	if (name == NULL) {
+		gctl_error(req, "No 'arg1' argument");
+		return;
+	}
+	if (strncmp(name, "/dev/", strlen("/dev/")) == 0)
+		name += strlen("/dev/");
+	pp = g_provider_by_name(name);
+	if (pp == NULL) {
+		G_CACHE_DEBUG(1, "Provider %s is invalid.", name);
+		gctl_error(req, "Provider %s is invalid.", name);
+		return;
+	}
+	gp = g_cache_create(mp, pp, &md, G_CACHE_TYPE_MANUAL);
+	if (gp == NULL) {
+		gctl_error(req, "Can't create %s.", md.md_name);
+		return;
+	}
+}
+
+static void
+g_cache_ctl_configure(struct gctl_req *req, struct g_class *mp)
+{
+	struct g_cache_metadata md;
+	struct g_cache_softc *sc;
+	struct g_consumer *cp;
+	intmax_t *bsize, *size;
+	const char *name;
+	int error, *nargs;
+
+	g_topology_assert();
+
+	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
+	if (nargs == NULL) {
+		gctl_error(req, "No '%s' argument", "nargs");
+		return;
+	}
+	if (*nargs != 1) {
+		gctl_error(req, "Missing device.");
+		return;
+	}
+
+	name = gctl_get_asciiparam(req, "arg0");
+	if (name == NULL) {
+		gctl_error(req, "No 'arg0' argument");
+		return;
+	}
+	sc = g_cache_find_device(mp, name);
+	if (sc == NULL) {
+		G_CACHE_DEBUG(1, "Device %s is invalid.", name);
+		gctl_error(req, "Device %s is invalid.", name);
+		return;
+	}
+
+	size = gctl_get_paraml(req, "size", sizeof(*size));
+	if (size == NULL) {
+		gctl_error(req, "No '%s' argument", "size");
+		return;
+	}
+	if ((u_int)*size != 0 && (u_int)*size < 100) {
+		gctl_error(req, "Invalid '%s' argument", "size");
+		return;
+	}
+	if ((u_int)*size != 0)
+		sc->sc_maxent = (u_int)*size;
+
+	bsize = gctl_get_paraml(req, "blocksize", sizeof(*bsize));
+	if (bsize == NULL) {
+		gctl_error(req, "No '%s' argument", "blocksize");
+		return;
+	}
+	if (*bsize < 0) {
+		gctl_error(req, "Invalid '%s' argument", "blocksize");
+		return;
+	}
+
+	if (sc->sc_type != G_CACHE_TYPE_AUTOMATIC)
+		return;
+
+	strlcpy(md.md_name, name, sizeof(md.md_name));
+	strlcpy(md.md_magic, G_CACHE_MAGIC, sizeof(md.md_magic));
+	md.md_version = G_CACHE_VERSION;
+	if ((u_int)*size != 0)
+		md.md_size = (u_int)*size;
+	else
+		md.md_size = sc->sc_maxent;
+	if ((u_int)*bsize != 0)
+		md.md_bsize = (u_int)*bsize;
+	else
+		md.md_bsize = sc->sc_bsize;
+	cp = LIST_FIRST(&sc->sc_geom->consumer);
+	md.md_provsize = cp->provider->mediasize;
+	error = g_cache_write_metadata(cp, &md);
+	if (error == 0)
+		G_CACHE_DEBUG(2, "Metadata on %s updated.", cp->provider->name);
+	else
+		G_CACHE_DEBUG(0, "Cannot update metadata on %s (error=%d).",
+		    cp->provider->name, error);
+}
+
+static void
+g_cache_ctl_destroy(struct gctl_req *req, struct g_class *mp)
+{
+	int *nargs, *force, error, i;
+	struct g_cache_softc *sc;
+	const char *name;
+	char param[16];
+
+	g_topology_assert();
+
+	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
+	if (nargs == NULL) {
+		gctl_error(req, "No '%s' argument", "nargs");
+		return;
+	}
+	if (*nargs <= 0) {
+		gctl_error(req, "Missing device(s).");
+		return;
+	}
+	force = gctl_get_paraml(req, "force", sizeof(*force));
+	if (force == NULL) {
+		gctl_error(req, "No 'force' argument");
+		return;
+	}
+
+	for (i = 0; i < *nargs; i++) {
+		snprintf(param, sizeof(param), "arg%d", i);
+		name = gctl_get_asciiparam(req, param);
+		if (name == NULL) {
+			gctl_error(req, "No 'arg%d' argument", i);
+			return;
+		}
+		sc = g_cache_find_device(mp, name);
+		if (sc == NULL) {
+			G_CACHE_DEBUG(1, "Device %s is invalid.", name);
+			gctl_error(req, "Device %s is invalid.", name);
+			return;
+		}
+		error = g_cache_destroy(sc, *force);
+		if (error != 0) {
+			gctl_error(req, "Cannot destroy device %s (error=%d).",
+			    sc->sc_name, error);
+			return;
+		}
+	}
+}
+
+static void
+g_cache_ctl_reset(struct gctl_req *req, struct g_class *mp)
+{
+	struct g_cache_softc *sc;
+	const char *name;
+	char param[16];
+	int i, *nargs;
+
+	g_topology_assert();
+
+	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
+	if (nargs == NULL) {
+		gctl_error(req, "No '%s' argument", "nargs");
+		return;
+	}
+	if (*nargs <= 0) {
+		gctl_error(req, "Missing device(s).");
+		return;
+	}
+
+	for (i = 0; i < *nargs; i++) {
+		snprintf(param, sizeof(param), "arg%d", i);
+		name = gctl_get_asciiparam(req, param);
+		if (name == NULL) {
+			gctl_error(req, "No 'arg%d' argument", i);
+			return;
+		}
+		sc = g_cache_find_device(mp, name);
+		if (sc == NULL) {
+			G_CACHE_DEBUG(1, "Device %s is invalid.", name);
+			gctl_error(req, "Device %s is invalid.", name);
+			return;
+		}
+		sc->sc_reads = 0;
+		sc->sc_readbytes = 0;
+		sc->sc_cachereads = 0;
+		sc->sc_cachereadbytes = 0;
+		sc->sc_cachehits = 0;
+		sc->sc_cachemisses = 0;
+		sc->sc_cachefull = 0;
+		sc->sc_writes = 0;
+		sc->sc_wrotebytes = 0;
+	}
+}
+
+static void
+g_cache_config(struct gctl_req *req, struct g_class *mp, const char *verb)
+{
+	uint32_t *version;
+
+	g_topology_assert();
+
+	version = gctl_get_paraml(req, "version", sizeof(*version));
+	if (version == NULL) {
+		gctl_error(req, "No '%s' argument.", "version");
+		return;
+	}
+	if (*version != G_CACHE_VERSION) {
+		gctl_error(req, "Userland and kernel parts are out of sync.");
+		return;
+	}
+
+	if (strcmp(verb, "create") == 0) {
+		g_cache_ctl_create(req, mp);
+		return;
+	} else if (strcmp(verb, "configure") == 0) {
+		g_cache_ctl_configure(req, mp);
+		return;
+	} else if (strcmp(verb, "destroy") == 0 ||
+	    strcmp(verb, "stop") == 0) {
+		g_cache_ctl_destroy(req, mp);
+		return;
+	} else if (strcmp(verb, "reset") == 0) {
+		g_cache_ctl_reset(req, mp);
+		return;
+	}
+
+	gctl_error(req, "Unknown verb.");
+}
+
+static void
+g_cache_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
+    struct g_consumer *cp, struct g_provider *pp)
+{
+	struct g_cache_softc *sc;
+
+	if (pp != NULL || cp != NULL)
+		return;
+	sc = gp->softc;
+	sbuf_printf(sb, "%s<Size>%u</Size>\n", indent, sc->sc_maxent);
+	sbuf_printf(sb, "%s<BlockSize>%u</BlockSize>\n", indent, sc->sc_bsize);
+	sbuf_printf(sb, "%s<TailOffset>%ju</TailOffset>\n", indent,
+	    (uintmax_t)sc->sc_tail);
+	sbuf_printf(sb, "%s<Entries>%u</Entries>\n", indent, sc->sc_nent);
+	sbuf_printf(sb, "%s<UsedEntries>%u</UsedEntries>\n", indent,
+	    sc->sc_nused);
+	sbuf_printf(sb, "%s<InvalidEntries>%u</InvalidEntries>\n", indent,
+	    sc->sc_invalid);
+	sbuf_printf(sb, "%s<Reads>%ju</Reads>\n", indent, sc->sc_reads);
+	sbuf_printf(sb, "%s<ReadBytes>%ju</ReadBytes>\n", indent,
+	    sc->sc_readbytes);
+	sbuf_printf(sb, "%s<CacheReads>%ju</CacheReads>\n", indent,
+	    sc->sc_cachereads);
+	sbuf_printf(sb, "%s<CacheReadBytes>%ju</CacheReadBytes>\n", indent,
+	    sc->sc_cachereadbytes);
+	sbuf_printf(sb, "%s<CacheHits>%ju</CacheHits>\n", indent,
+	    sc->sc_cachehits);
+	sbuf_printf(sb, "%s<CacheMisses>%ju</CacheMisses>\n", indent,
+	    sc->sc_cachemisses);
+	sbuf_printf(sb, "%s<CacheFull>%ju</CacheFull>\n", indent,
+	    sc->sc_cachefull);
+	sbuf_printf(sb, "%s<Writes>%ju</Writes>\n", indent, sc->sc_writes);
+	sbuf_printf(sb, "%s<WroteBytes>%ju</WroteBytes>\n", indent,
+	    sc->sc_wrotebytes);
+}
+
+DECLARE_GEOM_CLASS(g_cache_class, g_cache);