g_stripe.c revision 133373
1129473Spjd/*-
2129473Spjd * Copyright (c) 2003 Pawel Jakub Dawidek <pjd@FreeBSD.org>
3129473Spjd * All rights reserved.
4129473Spjd *
5129473Spjd * Redistribution and use in source and binary forms, with or without
6129473Spjd * modification, are permitted provided that the following conditions
7129473Spjd * are met:
8129473Spjd * 1. Redistributions of source code must retain the above copyright
9129473Spjd *    notice, this list of conditions and the following disclaimer.
10129473Spjd * 2. Redistributions in binary form must reproduce the above copyright
11129473Spjd *    notice, this list of conditions and the following disclaimer in the
12129473Spjd *    documentation and/or other materials provided with the distribution.
13129473Spjd *
14129473Spjd * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
15129473Spjd * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16129473Spjd * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17129473Spjd * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
18129473Spjd * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19129473Spjd * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20129473Spjd * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21129473Spjd * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22129473Spjd * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23129473Spjd * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24129473Spjd * SUCH DAMAGE.
25129473Spjd */
26129473Spjd
27129473Spjd#include <sys/cdefs.h>
28129473Spjd__FBSDID("$FreeBSD: head/sys/geom/stripe/g_stripe.c 133373 2004-08-09 11:29:42Z pjd $");
29129473Spjd
30129473Spjd#include <sys/param.h>
31129473Spjd#include <sys/systm.h>
32129473Spjd#include <sys/kernel.h>
33129473Spjd#include <sys/module.h>
34129473Spjd#include <sys/lock.h>
35129473Spjd#include <sys/mutex.h>
36129473Spjd#include <sys/bio.h>
37129473Spjd#include <sys/sysctl.h>
38129473Spjd#include <sys/malloc.h>
39131878Spjd#include <vm/uma.h>
40129473Spjd#include <geom/geom.h>
41129473Spjd#include <geom/stripe/g_stripe.h>
42129473Spjd
43129473Spjd
44131878Spjd#define	MAX_IO_SIZE	(DFLTPHYS * 2)
45129473Spjdstatic MALLOC_DEFINE(M_STRIPE, "stripe data", "GEOM_STRIPE Data");
46129473Spjd
47131878Spjdstatic uma_zone_t g_stripe_zone;
48129473Spjd
49129473Spjdstatic int g_stripe_destroy(struct g_stripe_softc *sc, boolean_t force);
50129473Spjdstatic int g_stripe_destroy_geom(struct gctl_req *req, struct g_class *mp,
51129473Spjd    struct g_geom *gp);
52129473Spjd
53129473Spjdstatic g_taste_t g_stripe_taste;
54129473Spjdstatic g_ctl_req_t g_stripe_config;
55129473Spjdstatic g_dumpconf_t g_stripe_dumpconf;
56131878Spjdstatic g_init_t g_stripe_init;
57131878Spjdstatic g_fini_t g_stripe_fini;
58129473Spjd
59129473Spjdstruct g_class g_stripe_class = {
60129473Spjd	.name = G_STRIPE_CLASS_NAME,
61133318Sphk	.version = G_VERSION,
62129473Spjd	.ctlreq = g_stripe_config,
63129473Spjd	.taste = g_stripe_taste,
64131878Spjd	.destroy_geom = g_stripe_destroy_geom,
65131878Spjd	.init = g_stripe_init,
66131878Spjd	.fini = g_stripe_fini
67129473Spjd};
68129473Spjd
69131878SpjdSYSCTL_DECL(_kern_geom);
70131878SpjdSYSCTL_NODE(_kern_geom, OID_AUTO, stripe, CTLFLAG_RW, 0, "GEOM_STRIPE stuff");
71131878Spjdstatic u_int g_stripe_debug = 0;
72131878SpjdSYSCTL_UINT(_kern_geom_stripe, OID_AUTO, debug, CTLFLAG_RW, &g_stripe_debug, 0,
73131878Spjd    "Debug level");
74131878Spjdstatic int g_stripe_fast = 1;
75131878SpjdTUNABLE_INT("kern.geom.stripe.fast", &g_stripe_fast);
76131878Spjdstatic int
77131878Spjdg_sysctl_stripe_fast(SYSCTL_HANDLER_ARGS)
78131878Spjd{
79131878Spjd	int error, fast;
80129473Spjd
81131878Spjd	fast = g_stripe_fast;
82131878Spjd	error = sysctl_handle_int(oidp, &fast, sizeof(fast), req);
83131878Spjd	if (error == 0 && req->newptr != NULL)
84131878Spjd		g_stripe_fast = fast;
85131878Spjd	return (error);
86131878Spjd}
87131878SpjdSYSCTL_PROC(_kern_geom_stripe, OID_AUTO, fast, CTLTYPE_INT | CTLFLAG_RW,
88132095Spjd    NULL, 0, g_sysctl_stripe_fast, "I", "Fast, but memory-consuming, mode");
89131878Spjdstatic u_int g_stripe_maxmem = MAX_IO_SIZE * 10;
90131878SpjdTUNABLE_INT("kern.geom.stripe.maxmem", &g_stripe_maxmem);
91131878SpjdSYSCTL_UINT(_kern_geom_stripe, OID_AUTO, maxmem, CTLFLAG_RD, &g_stripe_maxmem,
92132095Spjd    0, "Maximum memory that can be allocated in \"fast\" mode (in bytes)");
93133205Spjdstatic u_int g_stripe_fast_failed = 0;
94133205SpjdSYSCTL_UINT(_kern_geom_stripe, OID_AUTO, fast_failed, CTLFLAG_RD,
95133205Spjd    &g_stripe_fast_failed, 0, "How many times \"fast\" mode failed");
96131878Spjd
97129473Spjd/*
98129473Spjd * Greatest Common Divisor.
99129473Spjd */
100129473Spjdstatic u_int
101129473Spjdgcd(u_int a, u_int b)
102129473Spjd{
103129473Spjd	u_int c;
104129473Spjd
105129473Spjd	while (b != 0) {
106129473Spjd		c = a;
107129473Spjd		a = b;
108129473Spjd		b = (c % b);
109129473Spjd	}
110129473Spjd	return (a);
111129473Spjd}
112129473Spjd
113129473Spjd/*
114129473Spjd * Least Common Multiple.
115129473Spjd */
116129473Spjdstatic u_int
117129473Spjdlcm(u_int a, u_int b)
118129473Spjd{
119129473Spjd
120129473Spjd	return ((a * b) / gcd(a, b));
121129473Spjd}
122129473Spjd
123131878Spjdstatic void
124131878Spjdg_stripe_init(struct g_class *mp __unused)
125131878Spjd{
126131878Spjd
127131878Spjd	g_stripe_zone = uma_zcreate("g_stripe_zone", MAX_IO_SIZE, NULL, NULL,
128131878Spjd	    NULL, NULL, 0, 0);
129131878Spjd	g_stripe_maxmem -= g_stripe_maxmem % MAX_IO_SIZE;
130131878Spjd	uma_zone_set_max(g_stripe_zone, g_stripe_maxmem / MAX_IO_SIZE);
131131878Spjd}
132131878Spjd
133131878Spjdstatic void
134131878Spjdg_stripe_fini(struct g_class *mp __unused)
135131878Spjd{
136131878Spjd
137131878Spjd	uma_zdestroy(g_stripe_zone);
138131878Spjd}
139131878Spjd
140129473Spjd/*
141129473Spjd * Return the number of valid disks.
142129473Spjd */
143129473Spjdstatic u_int
144129473Spjdg_stripe_nvalid(struct g_stripe_softc *sc)
145129473Spjd{
146129473Spjd	u_int i, no;
147129473Spjd
148129473Spjd	no = 0;
149129473Spjd	for (i = 0; i < sc->sc_ndisks; i++) {
150129473Spjd		if (sc->sc_disks[i] != NULL)
151129473Spjd			no++;
152129473Spjd	}
153129473Spjd
154129473Spjd	return (no);
155129473Spjd}
156129473Spjd
157129473Spjdstatic void
158129473Spjdg_stripe_remove_disk(struct g_consumer *cp)
159129473Spjd{
160129473Spjd	struct g_stripe_softc *sc;
161129473Spjd	u_int no;
162129473Spjd
163129473Spjd	KASSERT(cp != NULL, ("Non-valid disk in %s.", __func__));
164129473Spjd	sc = (struct g_stripe_softc *)cp->private;
165129473Spjd	KASSERT(sc != NULL, ("NULL sc in %s.", __func__));
166129473Spjd	no = cp->index;
167129473Spjd
168129473Spjd	G_STRIPE_DEBUG(0, "Disk %s removed from %s.", cp->provider->name,
169132664Spjd	    sc->sc_name);
170129473Spjd
171129473Spjd	sc->sc_disks[no] = NULL;
172129473Spjd	if (sc->sc_provider != NULL) {
173129473Spjd		g_orphan_provider(sc->sc_provider, ENXIO);
174129473Spjd		sc->sc_provider = NULL;
175132664Spjd		G_STRIPE_DEBUG(0, "Device %s removed.", sc->sc_name);
176129473Spjd	}
177129473Spjd
178129473Spjd	if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
179129473Spjd		g_access(cp, -cp->acr, -cp->acw, -cp->ace);
180129473Spjd	g_detach(cp);
181129473Spjd	g_destroy_consumer(cp);
182129473Spjd}
183129473Spjd
184129473Spjdstatic void
185129473Spjdg_stripe_orphan(struct g_consumer *cp)
186129473Spjd{
187129473Spjd	struct g_stripe_softc *sc;
188129473Spjd	struct g_geom *gp;
189129473Spjd
190129473Spjd	g_topology_assert();
191129473Spjd	gp = cp->geom;
192129473Spjd	sc = gp->softc;
193129473Spjd	if (sc == NULL)
194129473Spjd		return;
195129473Spjd
196129473Spjd	g_stripe_remove_disk(cp);
197129473Spjd	/* If there are no valid disks anymore, remove device. */
198129473Spjd	if (g_stripe_nvalid(sc) == 0)
199129473Spjd		g_stripe_destroy(sc, 1);
200129473Spjd}
201129473Spjd
202129473Spjdstatic int
203129473Spjdg_stripe_access(struct g_provider *pp, int dr, int dw, int de)
204129473Spjd{
205129473Spjd	struct g_consumer *cp1, *cp2;
206129473Spjd	struct g_stripe_softc *sc;
207129473Spjd	struct g_geom *gp;
208129473Spjd	int error;
209129473Spjd
210129473Spjd	gp = pp->geom;
211129473Spjd	sc = gp->softc;
212129473Spjd
213129473Spjd	if (sc == NULL) {
214129473Spjd		/*
215129473Spjd		 * It looks like geom is being withered.
216129473Spjd		 * In that case we allow only negative requests.
217129473Spjd		 */
218129473Spjd		KASSERT(dr <= 0 && dw <= 0 && de <= 0,
219129473Spjd		    ("Positive access request (device=%s).", pp->name));
220129473Spjd		if ((pp->acr + dr) == 0 && (pp->acw + dw) == 0 &&
221129473Spjd		    (pp->ace + de) == 0) {
222129473Spjd			G_STRIPE_DEBUG(0, "Device %s definitely destroyed.",
223129473Spjd			    gp->name);
224129473Spjd		}
225129473Spjd		return (0);
226129473Spjd	}
227129473Spjd
228129473Spjd	/* On first open, grab an extra "exclusive" bit */
229129473Spjd	if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)
230129473Spjd		de++;
231129473Spjd	/* ... and let go of it on last close */
232129473Spjd	if ((pp->acr + dr) == 0 && (pp->acw + dw) == 0 && (pp->ace + de) == 0)
233129473Spjd		de--;
234129473Spjd
235129473Spjd	error = ENXIO;
236129473Spjd	LIST_FOREACH(cp1, &gp->consumer, consumer) {
237129473Spjd		error = g_access(cp1, dr, dw, de);
238129473Spjd		if (error == 0)
239129473Spjd			continue;
240129473Spjd		/*
241129473Spjd		 * If we fail here, backout all previous changes.
242129473Spjd		 */
243129473Spjd		LIST_FOREACH(cp2, &gp->consumer, consumer) {
244129473Spjd			if (cp1 == cp2)
245129473Spjd				return (error);
246129473Spjd			g_access(cp2, -dr, -dw, -de);
247129473Spjd		}
248129473Spjd		/* NOTREACHED */
249129473Spjd	}
250129473Spjd
251129473Spjd	return (error);
252129473Spjd}
253129473Spjd
254129473Spjdstatic void
255131878Spjdg_stripe_copy(struct g_stripe_softc *sc, char *src, char *dst, off_t offset,
256131878Spjd    off_t length, int mode)
257129473Spjd{
258131878Spjd	u_int stripesize;
259131878Spjd	size_t len;
260131878Spjd
261131878Spjd	stripesize = sc->sc_stripesize;
262131878Spjd	len = (size_t)(stripesize - (offset & (stripesize - 1)));
263131878Spjd	do {
264131878Spjd		bcopy(src, dst, len);
265131878Spjd		if (mode) {
266131878Spjd			dst += len + stripesize * (sc->sc_ndisks - 1);
267131878Spjd			src += len;
268131878Spjd		} else {
269131878Spjd			dst += len;
270131878Spjd			src += len + stripesize * (sc->sc_ndisks - 1);
271131878Spjd		}
272131878Spjd		length -= len;
273131878Spjd		KASSERT(length >= 0,
274131878Spjd		    ("Length < 0 (stripesize=%zu, offset=%jd, length=%jd).",
275131878Spjd		    (size_t)stripesize, (intmax_t)offset, (intmax_t)length));
276131878Spjd		if (length > stripesize)
277131878Spjd			len = stripesize;
278131878Spjd		else
279131878Spjd			len = length;
280131878Spjd	} while (length > 0);
281131878Spjd}
282131878Spjd
283131878Spjdstatic void
284131878Spjdg_stripe_done(struct bio *bp)
285131878Spjd{
286129473Spjd	struct g_stripe_softc *sc;
287131878Spjd	struct bio *pbp;
288131878Spjd
289131878Spjd	pbp = bp->bio_parent;
290131878Spjd	sc = pbp->bio_to->geom->softc;
291131878Spjd	if (pbp->bio_error == 0)
292131878Spjd		pbp->bio_error = bp->bio_error;
293131878Spjd	pbp->bio_completed += bp->bio_completed;
294133204Spjd	if (bp->bio_cmd == BIO_READ && bp->bio_caller1 != NULL) {
295133204Spjd		g_stripe_copy(sc, bp->bio_data, bp->bio_caller1, bp->bio_offset,
296131878Spjd		    bp->bio_length, 1);
297133204Spjd		bp->bio_data = bp->bio_caller1;
298133204Spjd		bp->bio_caller1 = NULL;
299131878Spjd	}
300131878Spjd	g_destroy_bio(bp);
301131878Spjd	pbp->bio_inbed++;
302131878Spjd	if (pbp->bio_children == pbp->bio_inbed) {
303133204Spjd		if (pbp->bio_driver1 != NULL)
304133204Spjd			uma_zfree(g_stripe_zone, pbp->bio_driver1);
305131878Spjd		g_io_deliver(pbp, pbp->bio_error);
306131878Spjd	}
307131878Spjd}
308131878Spjd
309131878Spjdstatic int
310131878Spjdg_stripe_start_fast(struct bio *bp, u_int no, off_t offset, off_t length)
311131878Spjd{
312131878Spjd	TAILQ_HEAD(, bio) queue = TAILQ_HEAD_INITIALIZER(queue);
313131878Spjd	u_int nparts = 0, stripesize;
314131878Spjd	struct g_stripe_softc *sc;
315131878Spjd	char *addr, *data = NULL;
316129473Spjd	struct bio *cbp;
317131878Spjd	int error;
318131878Spjd
319131878Spjd	sc = bp->bio_to->geom->softc;
320131878Spjd
321131878Spjd	addr = bp->bio_data;
322131878Spjd	stripesize = sc->sc_stripesize;
323131878Spjd
324131878Spjd	cbp = g_clone_bio(bp);
325131878Spjd	if (cbp == NULL) {
326131878Spjd		error = ENOMEM;
327131878Spjd		goto failure;
328131878Spjd	}
329131878Spjd	TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
330131878Spjd	nparts++;
331131878Spjd	/*
332131878Spjd	 * Fill in the component buf structure.
333131878Spjd	 */
334131878Spjd	cbp->bio_done = g_stripe_done;
335131878Spjd	cbp->bio_offset = offset;
336131878Spjd	cbp->bio_data = addr;
337133204Spjd	cbp->bio_caller1 = NULL;
338131878Spjd	cbp->bio_length = length;
339133204Spjd	cbp->bio_caller2 = sc->sc_disks[no];
340131878Spjd
341131878Spjd	/* offset -= offset % stripesize; */
342131878Spjd	offset -= offset & (stripesize - 1);
343131878Spjd	addr += length;
344131878Spjd	length = bp->bio_length - length;
345131878Spjd	for (no++; length > 0; no++, length -= stripesize, addr += stripesize) {
346131878Spjd		if (no > sc->sc_ndisks - 1) {
347131878Spjd			no = 0;
348131878Spjd			offset += stripesize;
349131878Spjd		}
350131878Spjd		if (nparts >= sc->sc_ndisks) {
351131878Spjd			cbp = TAILQ_NEXT(cbp, bio_queue);
352131878Spjd			if (cbp == NULL)
353131878Spjd				cbp = TAILQ_FIRST(&queue);
354131878Spjd			nparts++;
355131878Spjd			/*
356131878Spjd			 * Update bio structure.
357131878Spjd			 */
358131878Spjd			/*
359131878Spjd			 * MIN() is in case when
360131878Spjd			 * (bp->bio_length % sc->sc_stripesize) != 0.
361131878Spjd			 */
362131878Spjd			cbp->bio_length += MIN(stripesize, length);
363133204Spjd			if (cbp->bio_caller1 == NULL) {
364133204Spjd				cbp->bio_caller1 = cbp->bio_data;
365131878Spjd				cbp->bio_data = NULL;
366131878Spjd				if (data == NULL) {
367131878Spjd					data = uma_zalloc(g_stripe_zone,
368131878Spjd					    M_NOWAIT);
369131878Spjd					if (data == NULL) {
370131878Spjd						error = ENOMEM;
371131878Spjd						goto failure;
372131878Spjd					}
373131878Spjd				}
374131878Spjd			}
375131878Spjd		} else {
376131878Spjd			cbp = g_clone_bio(bp);
377131878Spjd			if (cbp == NULL) {
378131878Spjd				error = ENOMEM;
379131878Spjd				goto failure;
380131878Spjd			}
381131878Spjd			TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
382131878Spjd			nparts++;
383131878Spjd			/*
384131878Spjd			 * Fill in the component buf structure.
385131878Spjd			 */
386131878Spjd			cbp->bio_done = g_stripe_done;
387131878Spjd			cbp->bio_offset = offset;
388131878Spjd			cbp->bio_data = addr;
389133204Spjd			cbp->bio_caller1 = NULL;
390131878Spjd			/*
391131878Spjd			 * MIN() is in case when
392131878Spjd			 * (bp->bio_length % sc->sc_stripesize) != 0.
393131878Spjd			 */
394131878Spjd			cbp->bio_length = MIN(stripesize, length);
395133204Spjd			cbp->bio_caller2 = sc->sc_disks[no];
396131878Spjd		}
397131878Spjd	}
398131878Spjd	if (data != NULL)
399131878Spjd		bp->bio_caller1 = data;
400131878Spjd	/*
401131878Spjd	 * Fire off all allocated requests!
402131878Spjd	 */
403131878Spjd	while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
404131878Spjd		struct g_consumer *cp;
405131878Spjd
406131878Spjd		TAILQ_REMOVE(&queue, cbp, bio_queue);
407133204Spjd		cp = cbp->bio_caller2;
408133204Spjd		cbp->bio_caller2 = NULL;
409131878Spjd		cbp->bio_to = cp->provider;
410133204Spjd		if (cbp->bio_caller1 != NULL) {
411131878Spjd			cbp->bio_data = data;
412131878Spjd			if (bp->bio_cmd == BIO_WRITE) {
413133204Spjd				g_stripe_copy(sc, cbp->bio_caller1, data,
414131878Spjd				    cbp->bio_offset, cbp->bio_length, 0);
415131878Spjd			}
416131878Spjd			data += cbp->bio_length;
417131878Spjd		}
418131878Spjd		G_STRIPE_LOGREQ(cbp, "Sending request.");
419131878Spjd		g_io_request(cbp, cp);
420131878Spjd	}
421131878Spjd	return (0);
422131878Spjdfailure:
423131878Spjd	if (data != NULL)
424131878Spjd		uma_zfree(g_stripe_zone, data);
425131878Spjd	while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
426131878Spjd		TAILQ_REMOVE(&queue, cbp, bio_queue);
427133204Spjd		if (cbp->bio_caller1 != NULL) {
428133204Spjd			cbp->bio_data = cbp->bio_caller1;
429133204Spjd			cbp->bio_caller1 = NULL;
430131878Spjd		}
431133201Spjd		bp->bio_children--;
432131878Spjd		g_destroy_bio(cbp);
433131878Spjd	}
434131878Spjd	return (error);
435131878Spjd}
436131878Spjd
437131878Spjdstatic int
438131878Spjdg_stripe_start_economic(struct bio *bp, u_int no, off_t offset, off_t length)
439131878Spjd{
440131878Spjd	TAILQ_HEAD(, bio) queue = TAILQ_HEAD_INITIALIZER(queue);
441131878Spjd	struct g_stripe_softc *sc;
442129473Spjd	uint32_t stripesize;
443131878Spjd	struct bio *cbp;
444129473Spjd	char *addr;
445131878Spjd	int error;
446129473Spjd
447131878Spjd	sc = bp->bio_to->geom->softc;
448131878Spjd
449131878Spjd	addr = bp->bio_data;
450131878Spjd	stripesize = sc->sc_stripesize;
451131878Spjd
452131878Spjd	cbp = g_clone_bio(bp);
453131878Spjd	if (cbp == NULL) {
454131878Spjd		error = ENOMEM;
455131878Spjd		goto failure;
456131878Spjd	}
457131878Spjd	TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
458129473Spjd	/*
459131878Spjd	 * Fill in the component buf structure.
460131878Spjd	 */
461131878Spjd	cbp->bio_done = g_std_done;
462131878Spjd	cbp->bio_offset = offset;
463131878Spjd	cbp->bio_data = addr;
464131878Spjd	cbp->bio_length = length;
465133204Spjd	cbp->bio_caller2 = sc->sc_disks[no];
466131878Spjd
467131878Spjd	/* offset -= offset % stripesize; */
468131878Spjd	offset -= offset & (stripesize - 1);
469131878Spjd	addr += length;
470131878Spjd	length = bp->bio_length - length;
471131878Spjd	for (no++; length > 0; no++, length -= stripesize, addr += stripesize) {
472131878Spjd		if (no > sc->sc_ndisks - 1) {
473131878Spjd			no = 0;
474131878Spjd			offset += stripesize;
475131878Spjd		}
476131878Spjd		cbp = g_clone_bio(bp);
477131878Spjd		if (cbp == NULL) {
478131878Spjd			error = ENOMEM;
479131878Spjd			goto failure;
480131878Spjd		}
481131878Spjd		TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
482131878Spjd
483131878Spjd		/*
484131878Spjd		 * Fill in the component buf structure.
485131878Spjd		 */
486131878Spjd		cbp->bio_done = g_std_done;
487131878Spjd		cbp->bio_offset = offset;
488131878Spjd		cbp->bio_data = addr;
489131878Spjd		/*
490131878Spjd		 * MIN() is in case when
491131878Spjd		 * (bp->bio_length % sc->sc_stripesize) != 0.
492131878Spjd		 */
493131878Spjd		cbp->bio_length = MIN(stripesize, length);
494131878Spjd
495133204Spjd		cbp->bio_caller2 = sc->sc_disks[no];
496131878Spjd	}
497131878Spjd	/*
498131878Spjd	 * Fire off all allocated requests!
499131878Spjd	 */
500131878Spjd	while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
501131878Spjd		struct g_consumer *cp;
502131878Spjd
503131878Spjd		TAILQ_REMOVE(&queue, cbp, bio_queue);
504133204Spjd		cp = cbp->bio_caller2;
505133204Spjd		cbp->bio_caller2 = NULL;
506131878Spjd		cbp->bio_to = cp->provider;
507131878Spjd		G_STRIPE_LOGREQ(cbp, "Sending request.");
508131878Spjd		g_io_request(cbp, cp);
509131878Spjd	}
510131878Spjd	return (0);
511131878Spjdfailure:
512131878Spjd	while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
513131878Spjd		TAILQ_REMOVE(&queue, cbp, bio_queue);
514133201Spjd		bp->bio_children--;
515131878Spjd		g_destroy_bio(cbp);
516131878Spjd	}
517131878Spjd	return (error);
518131878Spjd}
519131878Spjd
520131878Spjdstatic void
521131878Spjdg_stripe_start(struct bio *bp)
522131878Spjd{
523131878Spjd	off_t offset, start, length, nstripe;
524131878Spjd	struct g_stripe_softc *sc;
525131878Spjd	u_int no, stripesize;
526131878Spjd	int error, fast = 0;
527131878Spjd
528131878Spjd	sc = bp->bio_to->geom->softc;
529131878Spjd	/*
530129473Spjd	 * If sc == NULL, provider's error should be set and g_stripe_start()
531129473Spjd	 * should not be called at all.
532129473Spjd	 */
533129473Spjd	KASSERT(sc != NULL,
534129473Spjd	    ("Provider's error should be set (error=%d)(device=%s).",
535129473Spjd	    bp->bio_to->error, bp->bio_to->name));
536129473Spjd
537129473Spjd	G_STRIPE_LOGREQ(bp, "Request received.");
538129473Spjd
539129473Spjd	switch (bp->bio_cmd) {
540129473Spjd	case BIO_READ:
541129473Spjd	case BIO_WRITE:
542129473Spjd	case BIO_DELETE:
543129473Spjd		/*
544129473Spjd		 * Only those requests are supported.
545129473Spjd		 */
546129473Spjd		break;
547129473Spjd	case BIO_GETATTR:
548129473Spjd		/* To which provider it should be delivered? */
549129473Spjd	default:
550129473Spjd		g_io_deliver(bp, EOPNOTSUPP);
551129473Spjd		return;
552129473Spjd	}
553129473Spjd
554129473Spjd	stripesize = sc->sc_stripesize;
555129473Spjd
556129473Spjd	/*
557131878Spjd	 * Calculations are quite messy, but fast I hope.
558129473Spjd	 */
559129473Spjd
560129473Spjd	/* Stripe number. */
561129473Spjd	/* nstripe = bp->bio_offset / stripesize; */
562129473Spjd	nstripe = bp->bio_offset >> (off_t)sc->sc_stripebits;
563129473Spjd	/* Disk number. */
564129473Spjd	no = nstripe % sc->sc_ndisks;
565129473Spjd	/* Start position in stripe. */
566129473Spjd	/* start = bp->bio_offset % stripesize; */
567129473Spjd	start = bp->bio_offset & (stripesize - 1);
568129473Spjd	/* Start position in disk. */
569131878Spjd	/* offset = (nstripe / sc->sc_ndisks) * stripesize + start; */
570131878Spjd	offset = ((nstripe / sc->sc_ndisks) << sc->sc_stripebits) + start;
571129473Spjd	/* Length of data to operate. */
572129473Spjd	length = MIN(bp->bio_length, stripesize - start);
573129473Spjd
574131878Spjd	/*
575131878Spjd	 * Do use "fast" mode when:
576131878Spjd	 * 1. "Fast" mode is ON.
577131878Spjd	 * and
578131878Spjd	 * 2. Request size is less than or equal to MAX_IO_SIZE (128kB),
579131878Spjd	 *    which should always be true.
580131878Spjd	 * and
581131878Spjd	 * 3. Request size is bigger than stripesize * ndisks. If it isn't,
582131878Spjd	 *    there will be no need to send more than one I/O request to
583131878Spjd	 *    a provider, so there is nothing to optmize.
584131878Spjd	 */
585131878Spjd	if (g_stripe_fast && bp->bio_length <= MAX_IO_SIZE &&
586131878Spjd	    bp->bio_length >= stripesize * sc->sc_ndisks) {
587131878Spjd		fast = 1;
588129473Spjd	}
589131878Spjd	error = 0;
590133205Spjd	if (fast) {
591131878Spjd		error = g_stripe_start_fast(bp, no, offset, length);
592133205Spjd		if (error != 0)
593133205Spjd			g_stripe_fast_failed++;
594133205Spjd	}
595129473Spjd	/*
596131878Spjd	 * Do use "economic" when:
597131878Spjd	 * 1. "Economic" mode is ON.
598131878Spjd	 * or
599131878Spjd	 * 2. "Fast" mode failed. It can only failed if there is no memory.
600129473Spjd	 */
601131878Spjd	if (!fast || error != 0)
602131878Spjd		error = g_stripe_start_economic(bp, no, offset, length);
603131878Spjd	if (error != 0) {
604131878Spjd		if (bp->bio_error == 0)
605131878Spjd			bp->bio_error = error;
606131878Spjd		g_io_deliver(bp, bp->bio_error);
607129473Spjd	}
608129473Spjd}
609129473Spjd
610129473Spjdstatic void
611129473Spjdg_stripe_check_and_run(struct g_stripe_softc *sc)
612129473Spjd{
613129473Spjd	off_t mediasize, ms;
614129473Spjd	u_int no, sectorsize = 0;
615129473Spjd
616129473Spjd	if (g_stripe_nvalid(sc) != sc->sc_ndisks)
617129473Spjd		return;
618129473Spjd
619132664Spjd	sc->sc_provider = g_new_providerf(sc->sc_geom, "stripe/%s",
620132664Spjd	    sc->sc_name);
621129473Spjd	/*
622129473Spjd	 * Find the smallest disk.
623129473Spjd	 */
624129473Spjd	mediasize = sc->sc_disks[0]->provider->mediasize;
625129473Spjd	if (sc->sc_type == G_STRIPE_TYPE_AUTOMATIC)
626129473Spjd		mediasize -= sc->sc_disks[0]->provider->sectorsize;
627129473Spjd	mediasize -= mediasize % sc->sc_stripesize;
628129473Spjd	sectorsize = sc->sc_disks[0]->provider->sectorsize;
629129473Spjd	for (no = 1; no < sc->sc_ndisks; no++) {
630129473Spjd		ms = sc->sc_disks[no]->provider->mediasize;
631129473Spjd		if (sc->sc_type == G_STRIPE_TYPE_AUTOMATIC)
632129473Spjd			ms -= sc->sc_disks[no]->provider->sectorsize;
633129473Spjd		ms -= ms % sc->sc_stripesize;
634129473Spjd		if (ms < mediasize)
635129473Spjd			mediasize = ms;
636129473Spjd		sectorsize = lcm(sectorsize,
637129473Spjd		    sc->sc_disks[no]->provider->sectorsize);
638129473Spjd	}
639129473Spjd	sc->sc_provider->sectorsize = sectorsize;
640129473Spjd	sc->sc_provider->mediasize = mediasize * sc->sc_ndisks;
641129473Spjd	g_error_provider(sc->sc_provider, 0);
642129473Spjd
643132664Spjd	G_STRIPE_DEBUG(0, "Device %s activated.", sc->sc_name);
644129473Spjd}
645129473Spjd
646129473Spjdstatic int
647129473Spjdg_stripe_read_metadata(struct g_consumer *cp, struct g_stripe_metadata *md)
648129473Spjd{
649129473Spjd	struct g_provider *pp;
650129473Spjd	u_char *buf;
651129473Spjd	int error;
652129473Spjd
653129473Spjd	g_topology_assert();
654129473Spjd
655129473Spjd	error = g_access(cp, 1, 0, 0);
656129473Spjd	if (error != 0)
657129473Spjd		return (error);
658129473Spjd	pp = cp->provider;
659129473Spjd	g_topology_unlock();
660129473Spjd	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
661129473Spjd	    &error);
662129473Spjd	g_topology_lock();
663129473Spjd	g_access(cp, -1, 0, 0);
664129473Spjd	if (buf == NULL)
665129473Spjd		return (error);
666129473Spjd
667129473Spjd	/* Decode metadata. */
668129473Spjd	stripe_metadata_decode(buf, md);
669129473Spjd	g_free(buf);
670129473Spjd
671129473Spjd	return (0);
672129473Spjd}
673129473Spjd
674129473Spjd/*
675129473Spjd * Add disk to given device.
676129473Spjd */
677129473Spjdstatic int
678129473Spjdg_stripe_add_disk(struct g_stripe_softc *sc, struct g_provider *pp, u_int no)
679129473Spjd{
680129473Spjd	struct g_consumer *cp, *fcp;
681129473Spjd	struct g_geom *gp;
682129473Spjd	int error;
683129473Spjd
684129473Spjd	/* Metadata corrupted? */
685129473Spjd	if (no >= sc->sc_ndisks)
686129473Spjd		return (EINVAL);
687129473Spjd
688129473Spjd	/* Check if disk is not already attached. */
689129473Spjd	if (sc->sc_disks[no] != NULL)
690129473Spjd		return (EEXIST);
691129473Spjd
692129473Spjd	gp = sc->sc_geom;
693129473Spjd	fcp = LIST_FIRST(&gp->consumer);
694129473Spjd
695129473Spjd	cp = g_new_consumer(gp);
696129473Spjd	error = g_attach(cp, pp);
697129473Spjd	if (error != 0) {
698129473Spjd		g_destroy_consumer(cp);
699129473Spjd		return (error);
700129473Spjd	}
701129473Spjd
702129473Spjd	if (fcp != NULL && (fcp->acr > 0 || fcp->acw > 0 || fcp->ace > 0)) {
703129473Spjd		error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
704129473Spjd		if (error != 0) {
705129473Spjd			g_detach(cp);
706129473Spjd			g_destroy_consumer(cp);
707129473Spjd			return (error);
708129473Spjd		}
709129473Spjd	}
710129473Spjd	if (sc->sc_type == G_STRIPE_TYPE_AUTOMATIC) {
711129473Spjd		struct g_stripe_metadata md;
712129473Spjd
713129473Spjd		/* Reread metadata. */
714129473Spjd		error = g_stripe_read_metadata(cp, &md);
715129473Spjd		if (error != 0)
716129473Spjd			goto fail;
717129473Spjd
718129473Spjd		if (strcmp(md.md_magic, G_STRIPE_MAGIC) != 0 ||
719129473Spjd		    strcmp(md.md_name, sc->sc_name) != 0 ||
720129473Spjd		    md.md_id != sc->sc_id) {
721129473Spjd			G_STRIPE_DEBUG(0, "Metadata on %s changed.", pp->name);
722129473Spjd			goto fail;
723129473Spjd		}
724129473Spjd	}
725129473Spjd
726129473Spjd	cp->private = sc;
727129473Spjd	cp->index = no;
728129473Spjd	sc->sc_disks[no] = cp;
729129473Spjd
730132664Spjd	G_STRIPE_DEBUG(0, "Disk %s attached to %s.", pp->name, sc->sc_name);
731129473Spjd
732129473Spjd	g_stripe_check_and_run(sc);
733129473Spjd
734129473Spjd	return (0);
735129473Spjdfail:
736129473Spjd	if (fcp != NULL && (fcp->acr > 0 || fcp->acw > 0 || fcp->ace > 0))
737129473Spjd		g_access(cp, -fcp->acr, -fcp->acw, -fcp->ace);
738129473Spjd	g_detach(cp);
739129473Spjd	g_destroy_consumer(cp);
740129473Spjd	return (error);
741129473Spjd}
742129473Spjd
743129473Spjdstatic struct g_geom *
744129473Spjdg_stripe_create(struct g_class *mp, const struct g_stripe_metadata *md,
745129473Spjd    u_int type)
746129473Spjd{
747129473Spjd	struct g_stripe_softc *sc;
748129473Spjd	struct g_geom *gp;
749129473Spjd	u_int no;
750129473Spjd
751132664Spjd	G_STRIPE_DEBUG(1, "Creating device %s (id=%u).", md->md_name,
752129473Spjd	    md->md_id);
753129473Spjd
754129473Spjd	/* Two disks is minimum. */
755132664Spjd	if (md->md_all < 2) {
756132664Spjd		G_STRIPE_DEBUG(0, "Too few disks defined for %s.", md->md_name);
757129473Spjd		return (NULL);
758129473Spjd	}
759129473Spjd#if 0
760129473Spjd	/* Stripe size have to be grater than or equal to sector size. */
761129473Spjd	if (md->md_stripesize < sectorsize) {
762132664Spjd		G_STRIPE_DEBUG(0, "Invalid stripe size for %s.", md->md_name);
763129473Spjd		return (NULL);
764129473Spjd	}
765129473Spjd#endif
766129473Spjd	/* Stripe size have to be power of 2. */
767129473Spjd	if (!powerof2(md->md_stripesize)) {
768132664Spjd		G_STRIPE_DEBUG(0, "Invalid stripe size for %s.", md->md_name);
769129473Spjd		return (NULL);
770129473Spjd	}
771129473Spjd
772129473Spjd	/* Check for duplicate unit */
773129473Spjd	LIST_FOREACH(gp, &mp->geom, geom) {
774129473Spjd		sc = gp->softc;
775129473Spjd		if (sc != NULL && strcmp(sc->sc_name, md->md_name) == 0) {
776129473Spjd			G_STRIPE_DEBUG(0, "Device %s already configured.",
777132664Spjd			    sc->sc_name);
778129473Spjd			return (NULL);
779129473Spjd		}
780129473Spjd	}
781132664Spjd	gp = g_new_geomf(mp, "%s", md->md_name);
782129473Spjd	gp->softc = NULL;	/* for a moment */
783129473Spjd
784132662Spjd	sc = malloc(sizeof(*sc), M_STRIPE, M_WAITOK | M_ZERO);
785129473Spjd	gp->start = g_stripe_start;
786129473Spjd	gp->spoiled = g_stripe_orphan;
787129473Spjd	gp->orphan = g_stripe_orphan;
788129473Spjd	gp->access = g_stripe_access;
789129473Spjd	gp->dumpconf = g_stripe_dumpconf;
790129473Spjd
791129473Spjd	sc->sc_id = md->md_id;
792129473Spjd	sc->sc_stripesize = md->md_stripesize;
793129473Spjd	sc->sc_stripebits = BITCOUNT(sc->sc_stripesize - 1);
794129473Spjd	sc->sc_ndisks = md->md_all;
795129473Spjd	sc->sc_disks = malloc(sizeof(struct g_consumer *) * sc->sc_ndisks,
796129473Spjd	    M_STRIPE, M_WAITOK | M_ZERO);
797129473Spjd	for (no = 0; no < sc->sc_ndisks; no++)
798129473Spjd		sc->sc_disks[no] = NULL;
799129473Spjd	sc->sc_type = type;
800129473Spjd
801129473Spjd	gp->softc = sc;
802129473Spjd	sc->sc_geom = gp;
803129473Spjd	sc->sc_provider = NULL;
804129473Spjd
805132664Spjd	G_STRIPE_DEBUG(0, "Device %s created (id=%u).", sc->sc_name, sc->sc_id);
806129473Spjd
807129473Spjd	return (gp);
808129473Spjd}
809129473Spjd
810129473Spjdstatic int
811129473Spjdg_stripe_destroy(struct g_stripe_softc *sc, boolean_t force)
812129473Spjd{
813129473Spjd	struct g_provider *pp;
814129473Spjd	struct g_geom *gp;
815129473Spjd	u_int no;
816129473Spjd
817129473Spjd	g_topology_assert();
818129473Spjd
819129473Spjd	if (sc == NULL)
820129473Spjd		return (ENXIO);
821129473Spjd
822129473Spjd	pp = sc->sc_provider;
823129473Spjd	if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
824129473Spjd		if (force) {
825129473Spjd			G_STRIPE_DEBUG(0, "Device %s is still open, so it "
826129473Spjd			    "can't be definitely removed.", pp->name);
827129473Spjd		} else {
828129473Spjd			G_STRIPE_DEBUG(1,
829129473Spjd			    "Device %s is still open (r%dw%de%d).", pp->name,
830129473Spjd			    pp->acr, pp->acw, pp->ace);
831129473Spjd			return (EBUSY);
832129473Spjd		}
833129473Spjd	}
834129473Spjd
835129473Spjd	for (no = 0; no < sc->sc_ndisks; no++) {
836129473Spjd		if (sc->sc_disks[no] != NULL)
837129473Spjd			g_stripe_remove_disk(sc->sc_disks[no]);
838129473Spjd	}
839129473Spjd
840129473Spjd	gp = sc->sc_geom;
841129473Spjd	gp->softc = NULL;
842129473Spjd	KASSERT(sc->sc_provider == NULL, ("Provider still exists? (device=%s)",
843129473Spjd	    gp->name));
844129473Spjd	free(sc->sc_disks, M_STRIPE);
845129473Spjd	free(sc, M_STRIPE);
846129473Spjd
847129473Spjd	pp = LIST_FIRST(&gp->provider);
848129473Spjd	if (pp == NULL || (pp->acr == 0 && pp->acw == 0 && pp->ace == 0))
849129473Spjd		G_STRIPE_DEBUG(0, "Device %s destroyed.", gp->name);
850129473Spjd
851129473Spjd	g_wither_geom(gp, ENXIO);
852129473Spjd
853129473Spjd	return (0);
854129473Spjd}
855129473Spjd
856129473Spjdstatic int
857129473Spjdg_stripe_destroy_geom(struct gctl_req *req __unused,
858129473Spjd    struct g_class *mp __unused, struct g_geom *gp)
859129473Spjd{
860129473Spjd	struct g_stripe_softc *sc;
861129473Spjd
862129473Spjd	sc = gp->softc;
863129473Spjd	return (g_stripe_destroy(sc, 0));
864129473Spjd}
865129473Spjd
866129473Spjdstatic struct g_geom *
867129473Spjdg_stripe_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
868129473Spjd{
869129473Spjd	struct g_stripe_metadata md;
870129473Spjd	struct g_stripe_softc *sc;
871129473Spjd	struct g_consumer *cp;
872129473Spjd	struct g_geom *gp;
873129473Spjd	int error;
874129473Spjd
875129473Spjd	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
876129473Spjd	g_topology_assert();
877129473Spjd
878129473Spjd	G_STRIPE_DEBUG(3, "Tasting %s.", pp->name);
879129473Spjd
880129473Spjd	gp = g_new_geomf(mp, "stripe:taste");
881129473Spjd	gp->start = g_stripe_start;
882129473Spjd	gp->access = g_stripe_access;
883129473Spjd	gp->orphan = g_stripe_orphan;
884129473Spjd	cp = g_new_consumer(gp);
885129473Spjd	g_attach(cp, pp);
886129473Spjd	error = g_stripe_read_metadata(cp, &md);
887133371Spjd	g_detach(cp);
888133371Spjd	g_destroy_consumer(cp);
889133371Spjd	g_destroy_geom(gp);
890129473Spjd	if (error != 0)
891129473Spjd		return (NULL);
892129473Spjd	gp = NULL;
893129473Spjd
894129473Spjd	if (strcmp(md.md_magic, G_STRIPE_MAGIC) != 0)
895129473Spjd		return (NULL);
896129473Spjd	if (md.md_version > G_STRIPE_VERSION) {
897129473Spjd		printf("geom_stripe.ko module is too old to handle %s.\n",
898129473Spjd		    pp->name);
899129473Spjd		return (NULL);
900129473Spjd	}
901133373Spjd	/*
902133373Spjd	 * Backward compatibility:
903133373Spjd	 * There was no md_provider field in earlier versions of metadata.
904133373Spjd	 */
905133373Spjd	if (md.md_version < 2)
906133373Spjd		bzero(md.md_provider, sizeof(md.md_provider));
907129473Spjd
908133373Spjd	if (md.md_provider[0] != '\0' && strcmp(md.md_provider, pp->name) != 0)
909133373Spjd		return (NULL);
910133373Spjd
911129473Spjd	/*
912129473Spjd	 * Let's check if device already exists.
913129473Spjd	 */
914129473Spjd	sc = NULL;
915129473Spjd	LIST_FOREACH(gp, &mp->geom, geom) {
916129473Spjd		sc = gp->softc;
917129473Spjd		if (sc == NULL)
918129473Spjd			continue;
919129473Spjd		if (sc->sc_type != G_STRIPE_TYPE_AUTOMATIC)
920129473Spjd			continue;
921129473Spjd		if (strcmp(md.md_name, sc->sc_name) != 0)
922129473Spjd			continue;
923129473Spjd		if (md.md_id != sc->sc_id)
924129473Spjd			continue;
925129473Spjd		break;
926129473Spjd	}
927129473Spjd	if (gp != NULL) {
928129473Spjd		G_STRIPE_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
929129473Spjd		error = g_stripe_add_disk(sc, pp, md.md_no);
930129473Spjd		if (error != 0) {
931129473Spjd			G_STRIPE_DEBUG(0,
932129473Spjd			    "Cannot add disk %s to %s (error=%d).", pp->name,
933129473Spjd			    gp->name, error);
934129473Spjd			return (NULL);
935129473Spjd		}
936129473Spjd	} else {
937129473Spjd		gp = g_stripe_create(mp, &md, G_STRIPE_TYPE_AUTOMATIC);
938129473Spjd		if (gp == NULL) {
939132664Spjd			G_STRIPE_DEBUG(0, "Cannot create device %s.",
940129473Spjd			    md.md_name);
941129473Spjd			return (NULL);
942129473Spjd		}
943129473Spjd		sc = gp->softc;
944129473Spjd		G_STRIPE_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
945129473Spjd		error = g_stripe_add_disk(sc, pp, md.md_no);
946129473Spjd		if (error != 0) {
947129473Spjd			G_STRIPE_DEBUG(0,
948129473Spjd			    "Cannot add disk %s to %s (error=%d).", pp->name,
949129473Spjd			    gp->name, error);
950129473Spjd			g_stripe_destroy(sc, 1);
951129473Spjd			return (NULL);
952129473Spjd		}
953129473Spjd	}
954129473Spjd
955129473Spjd	return (gp);
956129473Spjd}
957129473Spjd
958129473Spjdstatic void
959129473Spjdg_stripe_ctl_create(struct gctl_req *req, struct g_class *mp)
960129473Spjd{
961129473Spjd	u_int attached, no;
962129473Spjd	struct g_stripe_metadata md;
963129473Spjd	struct g_provider *pp;
964129473Spjd	struct g_stripe_softc *sc;
965129473Spjd	struct g_geom *gp;
966129473Spjd	struct sbuf *sb;
967129473Spjd	intmax_t *stripesize;
968129473Spjd	const char *name;
969129473Spjd	char param[16];
970129473Spjd	int *nargs;
971129473Spjd
972129473Spjd	g_topology_assert();
973129473Spjd	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
974129473Spjd	if (nargs == NULL) {
975129473Spjd		gctl_error(req, "No '%s' argument.", "nargs");
976129473Spjd		return;
977129473Spjd	}
978129473Spjd	if (*nargs <= 2) {
979129473Spjd		gctl_error(req, "Too few arguments.");
980129473Spjd		return;
981129473Spjd	}
982129473Spjd
983129473Spjd	strlcpy(md.md_magic, G_STRIPE_MAGIC, sizeof(md.md_magic));
984129473Spjd	md.md_version = G_STRIPE_VERSION;
985129473Spjd	name = gctl_get_asciiparam(req, "arg0");
986129473Spjd	if (name == NULL) {
987129473Spjd		gctl_error(req, "No 'arg%u' argument.", 0);
988129473Spjd		return;
989129473Spjd	}
990129473Spjd	strlcpy(md.md_name, name, sizeof(md.md_name));
991129473Spjd	md.md_id = arc4random();
992129473Spjd	md.md_no = 0;
993129473Spjd	md.md_all = *nargs - 1;
994129473Spjd	stripesize = gctl_get_paraml(req, "stripesize", sizeof(*stripesize));
995129473Spjd	if (stripesize == NULL) {
996129473Spjd		gctl_error(req, "No '%s' argument.", "stripesize");
997129473Spjd		return;
998129473Spjd	}
999129473Spjd	md.md_stripesize = *stripesize;
1000133373Spjd	bzero(md.md_provider, sizeof(md.md_provider));
1001129473Spjd
1002129473Spjd	/* Check all providers are valid */
1003129473Spjd	for (no = 1; no < *nargs; no++) {
1004129473Spjd		snprintf(param, sizeof(param), "arg%u", no);
1005129473Spjd		name = gctl_get_asciiparam(req, param);
1006129473Spjd		if (name == NULL) {
1007129473Spjd			gctl_error(req, "No 'arg%u' argument.", no);
1008129473Spjd			return;
1009129473Spjd		}
1010129473Spjd		if (strncmp(name, "/dev/", strlen("/dev/")) == 0)
1011129473Spjd			name += strlen("/dev/");
1012129473Spjd		pp = g_provider_by_name(name);
1013129473Spjd		if (pp == NULL) {
1014129473Spjd			G_STRIPE_DEBUG(1, "Disk %s is invalid.", name);
1015129473Spjd			gctl_error(req, "Disk %s is invalid.", name);
1016129473Spjd			return;
1017129473Spjd		}
1018129473Spjd	}
1019129473Spjd
1020129473Spjd	gp = g_stripe_create(mp, &md, G_STRIPE_TYPE_MANUAL);
1021129473Spjd	if (gp == NULL) {
1022132664Spjd		gctl_error(req, "Can't configure %s.", md.md_name);
1023129473Spjd		return;
1024129473Spjd	}
1025129473Spjd
1026129473Spjd	sc = gp->softc;
1027129473Spjd	sb = sbuf_new(NULL, NULL, 0, SBUF_AUTOEXTEND);
1028129473Spjd	sbuf_printf(sb, "Can't attach disk(s) to %s:", gp->name);
1029129473Spjd	for (attached = 0, no = 1; no < *nargs; no++) {
1030129473Spjd		snprintf(param, sizeof(param), "arg%u", no);
1031129473Spjd		name = gctl_get_asciiparam(req, param);
1032129473Spjd		if (strncmp(name, "/dev/", strlen("/dev/")) == 0)
1033129473Spjd			name += strlen("/dev/");
1034129473Spjd		pp = g_provider_by_name(name);
1035129473Spjd		KASSERT(pp != NULL, ("Provider %s disappear?!", name));
1036129473Spjd		if (g_stripe_add_disk(sc, pp, no - 1) != 0) {
1037129473Spjd			G_STRIPE_DEBUG(1, "Disk %u (%s) not attached to %s.",
1038129473Spjd			    no, pp->name, gp->name);
1039129473Spjd			sbuf_printf(sb, " %s", pp->name);
1040129473Spjd			continue;
1041129473Spjd		}
1042129473Spjd		attached++;
1043129473Spjd	}
1044129473Spjd	sbuf_finish(sb);
1045129473Spjd	if (md.md_all != attached) {
1046129473Spjd		g_stripe_destroy(gp->softc, 1);
1047129473Spjd		gctl_error(req, "%s", sbuf_data(sb));
1048129473Spjd	}
1049129473Spjd	sbuf_delete(sb);
1050129473Spjd}
1051129473Spjd
1052129473Spjdstatic struct g_stripe_softc *
1053129473Spjdg_stripe_find_device(struct g_class *mp, const char *name)
1054129473Spjd{
1055129473Spjd	struct g_stripe_softc *sc;
1056129473Spjd	struct g_geom *gp;
1057129473Spjd
1058129473Spjd	LIST_FOREACH(gp, &mp->geom, geom) {
1059129473Spjd		sc = gp->softc;
1060129473Spjd		if (sc == NULL)
1061129473Spjd			continue;
1062132664Spjd		if (strcmp(sc->sc_name, name) == 0)
1063129473Spjd			return (sc);
1064129473Spjd	}
1065129473Spjd	return (NULL);
1066129473Spjd}
1067129473Spjd
1068129473Spjdstatic void
1069129473Spjdg_stripe_ctl_destroy(struct gctl_req *req, struct g_class *mp)
1070129473Spjd{
1071129473Spjd	struct g_stripe_softc *sc;
1072129473Spjd	int *force, *nargs, error;
1073129473Spjd	const char *name;
1074129473Spjd	char param[16];
1075129473Spjd	u_int i;
1076129473Spjd
1077129473Spjd	g_topology_assert();
1078129473Spjd
1079129473Spjd	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
1080129473Spjd	if (nargs == NULL) {
1081129473Spjd		gctl_error(req, "No '%s' argument.", "nargs");
1082129473Spjd		return;
1083129473Spjd	}
1084129473Spjd	if (*nargs <= 0) {
1085129473Spjd		gctl_error(req, "Missing device(s).");
1086129473Spjd		return;
1087129473Spjd	}
1088129473Spjd	force = gctl_get_paraml(req, "force", sizeof(*force));
1089129473Spjd	if (force == NULL) {
1090129473Spjd		gctl_error(req, "No '%s' argument.", "force");
1091129473Spjd		return;
1092129473Spjd	}
1093129473Spjd
1094129473Spjd	for (i = 0; i < (u_int)*nargs; i++) {
1095129473Spjd		snprintf(param, sizeof(param), "arg%u", i);
1096129473Spjd		name = gctl_get_asciiparam(req, param);
1097129473Spjd		if (name == NULL) {
1098129473Spjd			gctl_error(req, "No 'arg%u' argument.", i);
1099129473Spjd			return;
1100129473Spjd		}
1101129473Spjd		sc = g_stripe_find_device(mp, name);
1102129473Spjd		if (sc == NULL) {
1103129473Spjd			gctl_error(req, "No such device: %s.", name);
1104129473Spjd			return;
1105129473Spjd		}
1106129473Spjd		error = g_stripe_destroy(sc, *force);
1107129473Spjd		if (error != 0) {
1108129473Spjd			gctl_error(req, "Cannot destroy device %s (error=%d).",
1109132664Spjd			    sc->sc_name, error);
1110129473Spjd			return;
1111129473Spjd		}
1112129473Spjd	}
1113129473Spjd}
1114129473Spjd
1115129473Spjdstatic void
1116129473Spjdg_stripe_config(struct gctl_req *req, struct g_class *mp, const char *verb)
1117129473Spjd{
1118129473Spjd	uint32_t *version;
1119129473Spjd
1120129473Spjd	g_topology_assert();
1121129473Spjd
1122129473Spjd	version = gctl_get_paraml(req, "version", sizeof(*version));
1123129473Spjd	if (version == NULL) {
1124129473Spjd		gctl_error(req, "No '%s' argument.", "version");
1125129473Spjd		return;
1126129473Spjd	}
1127129473Spjd	if (*version != G_STRIPE_VERSION) {
1128129473Spjd		gctl_error(req, "Userland and kernel parts are out of sync.");
1129129473Spjd		return;
1130129473Spjd	}
1131129473Spjd
1132129473Spjd	if (strcmp(verb, "create") == 0) {
1133129473Spjd		g_stripe_ctl_create(req, mp);
1134129473Spjd		return;
1135131649Spjd	} else if (strcmp(verb, "destroy") == 0 ||
1136131649Spjd	    strcmp(verb, "stop") == 0) {
1137129473Spjd		g_stripe_ctl_destroy(req, mp);
1138129473Spjd		return;
1139129473Spjd	}
1140129473Spjd
1141129473Spjd	gctl_error(req, "Unknown verb.");
1142129473Spjd}
1143129473Spjd
1144129473Spjdstatic void
1145129473Spjdg_stripe_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1146129473Spjd    struct g_consumer *cp, struct g_provider *pp)
1147129473Spjd{
1148129473Spjd	struct g_stripe_softc *sc;
1149129473Spjd
1150129473Spjd	sc = gp->softc;
1151132665Spjd	if (sc == NULL)
1152129473Spjd		return;
1153132665Spjd	if (pp != NULL) {
1154132665Spjd		/* Nothing here. */
1155132665Spjd	} else if (cp != NULL) {
1156132665Spjd		/* Nothing here. */
1157132665Spjd	} else {
1158132665Spjd		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
1159132665Spjd		sbuf_printf(sb, "%s<Stripesize>%u</Stripesize>\n", indent,
1160132665Spjd		    (u_int)sc->sc_stripesize);
1161132665Spjd		sbuf_printf(sb, "%s<Type>", indent);
1162132665Spjd		switch (sc->sc_type) {
1163132665Spjd		case G_STRIPE_TYPE_AUTOMATIC:
1164132665Spjd			sbuf_printf(sb, "AUTOMATIC");
1165132665Spjd			break;
1166132665Spjd		case G_STRIPE_TYPE_MANUAL:
1167132665Spjd			sbuf_printf(sb, "MANUAL");
1168132665Spjd			break;
1169132665Spjd		default:
1170132665Spjd			sbuf_printf(sb, "UNKNOWN");
1171132665Spjd			break;
1172132665Spjd		}
1173132665Spjd		sbuf_printf(sb, "</Type>\n");
1174132665Spjd		sbuf_printf(sb, "%s<Status>Total=%u, Online=%u</Status>\n",
1175132665Spjd		    indent, sc->sc_ndisks, g_stripe_nvalid(sc));
1176132665Spjd		sbuf_printf(sb, "%s<State>", indent);
1177132665Spjd		if (sc->sc_provider != NULL && sc->sc_provider->error == 0)
1178132665Spjd			sbuf_printf(sb, "UP");
1179132665Spjd		else
1180132665Spjd			sbuf_printf(sb, "DOWN");
1181132665Spjd		sbuf_printf(sb, "</State>\n");
1182129473Spjd	}
1183129473Spjd}
1184129473Spjd
1185129473SpjdDECLARE_GEOM_CLASS(g_stripe_class, g_stripe);
1186