g_stripe.c revision 142727
1129473Spjd/*-
2142727Spjd * Copyright (c) 2004-2005 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 142727 2005-02-27 23:07:47Z 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;
72134528SpjdTUNABLE_INT("kern.geom.stripe.debug", &g_stripe_debug);
73131878SpjdSYSCTL_UINT(_kern_geom_stripe, OID_AUTO, debug, CTLFLAG_RW, &g_stripe_debug, 0,
74131878Spjd    "Debug level");
75138623Spjdstatic int g_stripe_fast = 0;
76131878SpjdTUNABLE_INT("kern.geom.stripe.fast", &g_stripe_fast);
77131878Spjdstatic int
78131878Spjdg_sysctl_stripe_fast(SYSCTL_HANDLER_ARGS)
79131878Spjd{
80131878Spjd	int error, fast;
81129473Spjd
82131878Spjd	fast = g_stripe_fast;
83131878Spjd	error = sysctl_handle_int(oidp, &fast, sizeof(fast), req);
84131878Spjd	if (error == 0 && req->newptr != NULL)
85131878Spjd		g_stripe_fast = fast;
86131878Spjd	return (error);
87131878Spjd}
88131878SpjdSYSCTL_PROC(_kern_geom_stripe, OID_AUTO, fast, CTLTYPE_INT | CTLFLAG_RW,
89132095Spjd    NULL, 0, g_sysctl_stripe_fast, "I", "Fast, but memory-consuming, mode");
90138623Spjdstatic u_int g_stripe_maxmem = MAX_IO_SIZE * 100;
91131878SpjdTUNABLE_INT("kern.geom.stripe.maxmem", &g_stripe_maxmem);
92131878SpjdSYSCTL_UINT(_kern_geom_stripe, OID_AUTO, maxmem, CTLFLAG_RD, &g_stripe_maxmem,
93132095Spjd    0, "Maximum memory that can be allocated in \"fast\" mode (in bytes)");
94133205Spjdstatic u_int g_stripe_fast_failed = 0;
95133205SpjdSYSCTL_UINT(_kern_geom_stripe, OID_AUTO, fast_failed, CTLFLAG_RD,
96133205Spjd    &g_stripe_fast_failed, 0, "How many times \"fast\" mode failed");
97131878Spjd
98129473Spjd/*
99129473Spjd * Greatest Common Divisor.
100129473Spjd */
101129473Spjdstatic u_int
102129473Spjdgcd(u_int a, u_int b)
103129473Spjd{
104129473Spjd	u_int c;
105129473Spjd
106129473Spjd	while (b != 0) {
107129473Spjd		c = a;
108129473Spjd		a = b;
109129473Spjd		b = (c % b);
110129473Spjd	}
111129473Spjd	return (a);
112129473Spjd}
113129473Spjd
114129473Spjd/*
115129473Spjd * Least Common Multiple.
116129473Spjd */
117129473Spjdstatic u_int
118129473Spjdlcm(u_int a, u_int b)
119129473Spjd{
120129473Spjd
121129473Spjd	return ((a * b) / gcd(a, b));
122129473Spjd}
123129473Spjd
124131878Spjdstatic void
125131878Spjdg_stripe_init(struct g_class *mp __unused)
126131878Spjd{
127131878Spjd
128131878Spjd	g_stripe_zone = uma_zcreate("g_stripe_zone", MAX_IO_SIZE, NULL, NULL,
129131878Spjd	    NULL, NULL, 0, 0);
130131878Spjd	g_stripe_maxmem -= g_stripe_maxmem % MAX_IO_SIZE;
131131878Spjd	uma_zone_set_max(g_stripe_zone, g_stripe_maxmem / MAX_IO_SIZE);
132131878Spjd}
133131878Spjd
134131878Spjdstatic void
135131878Spjdg_stripe_fini(struct g_class *mp __unused)
136131878Spjd{
137131878Spjd
138131878Spjd	uma_zdestroy(g_stripe_zone);
139131878Spjd}
140131878Spjd
141129473Spjd/*
142129473Spjd * Return the number of valid disks.
143129473Spjd */
144129473Spjdstatic u_int
145129473Spjdg_stripe_nvalid(struct g_stripe_softc *sc)
146129473Spjd{
147129473Spjd	u_int i, no;
148129473Spjd
149129473Spjd	no = 0;
150129473Spjd	for (i = 0; i < sc->sc_ndisks; i++) {
151129473Spjd		if (sc->sc_disks[i] != NULL)
152129473Spjd			no++;
153129473Spjd	}
154129473Spjd
155129473Spjd	return (no);
156129473Spjd}
157129473Spjd
158129473Spjdstatic void
159129473Spjdg_stripe_remove_disk(struct g_consumer *cp)
160129473Spjd{
161129473Spjd	struct g_stripe_softc *sc;
162129473Spjd	u_int no;
163129473Spjd
164129473Spjd	KASSERT(cp != NULL, ("Non-valid disk in %s.", __func__));
165129473Spjd	sc = (struct g_stripe_softc *)cp->private;
166129473Spjd	KASSERT(sc != NULL, ("NULL sc in %s.", __func__));
167129473Spjd	no = cp->index;
168129473Spjd
169129473Spjd	G_STRIPE_DEBUG(0, "Disk %s removed from %s.", cp->provider->name,
170132664Spjd	    sc->sc_name);
171129473Spjd
172129473Spjd	sc->sc_disks[no] = NULL;
173129473Spjd	if (sc->sc_provider != NULL) {
174129473Spjd		g_orphan_provider(sc->sc_provider, ENXIO);
175129473Spjd		sc->sc_provider = NULL;
176132664Spjd		G_STRIPE_DEBUG(0, "Device %s removed.", sc->sc_name);
177129473Spjd	}
178129473Spjd
179129473Spjd	if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
180129473Spjd		g_access(cp, -cp->acr, -cp->acw, -cp->ace);
181129473Spjd	g_detach(cp);
182129473Spjd	g_destroy_consumer(cp);
183129473Spjd}
184129473Spjd
185129473Spjdstatic void
186129473Spjdg_stripe_orphan(struct g_consumer *cp)
187129473Spjd{
188129473Spjd	struct g_stripe_softc *sc;
189129473Spjd	struct g_geom *gp;
190129473Spjd
191129473Spjd	g_topology_assert();
192129473Spjd	gp = cp->geom;
193129473Spjd	sc = gp->softc;
194129473Spjd	if (sc == NULL)
195129473Spjd		return;
196129473Spjd
197129473Spjd	g_stripe_remove_disk(cp);
198129473Spjd	/* If there are no valid disks anymore, remove device. */
199129473Spjd	if (g_stripe_nvalid(sc) == 0)
200129473Spjd		g_stripe_destroy(sc, 1);
201129473Spjd}
202129473Spjd
203129473Spjdstatic int
204129473Spjdg_stripe_access(struct g_provider *pp, int dr, int dw, int de)
205129473Spjd{
206129473Spjd	struct g_consumer *cp1, *cp2;
207129473Spjd	struct g_stripe_softc *sc;
208129473Spjd	struct g_geom *gp;
209129473Spjd	int error;
210129473Spjd
211129473Spjd	gp = pp->geom;
212129473Spjd	sc = gp->softc;
213129473Spjd
214129473Spjd	if (sc == NULL) {
215129473Spjd		/*
216129473Spjd		 * It looks like geom is being withered.
217129473Spjd		 * In that case we allow only negative requests.
218129473Spjd		 */
219129473Spjd		KASSERT(dr <= 0 && dw <= 0 && de <= 0,
220129473Spjd		    ("Positive access request (device=%s).", pp->name));
221129473Spjd		if ((pp->acr + dr) == 0 && (pp->acw + dw) == 0 &&
222129473Spjd		    (pp->ace + de) == 0) {
223129473Spjd			G_STRIPE_DEBUG(0, "Device %s definitely destroyed.",
224129473Spjd			    gp->name);
225129473Spjd		}
226129473Spjd		return (0);
227129473Spjd	}
228129473Spjd
229129473Spjd	/* On first open, grab an extra "exclusive" bit */
230129473Spjd	if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)
231129473Spjd		de++;
232129473Spjd	/* ... and let go of it on last close */
233129473Spjd	if ((pp->acr + dr) == 0 && (pp->acw + dw) == 0 && (pp->ace + de) == 0)
234129473Spjd		de--;
235129473Spjd
236129473Spjd	error = ENXIO;
237129473Spjd	LIST_FOREACH(cp1, &gp->consumer, consumer) {
238129473Spjd		error = g_access(cp1, dr, dw, de);
239129473Spjd		if (error == 0)
240129473Spjd			continue;
241129473Spjd		/*
242129473Spjd		 * If we fail here, backout all previous changes.
243129473Spjd		 */
244129473Spjd		LIST_FOREACH(cp2, &gp->consumer, consumer) {
245129473Spjd			if (cp1 == cp2)
246129473Spjd				return (error);
247129473Spjd			g_access(cp2, -dr, -dw, -de);
248129473Spjd		}
249129473Spjd		/* NOTREACHED */
250129473Spjd	}
251129473Spjd
252129473Spjd	return (error);
253129473Spjd}
254129473Spjd
255129473Spjdstatic void
256131878Spjdg_stripe_copy(struct g_stripe_softc *sc, char *src, char *dst, off_t offset,
257131878Spjd    off_t length, int mode)
258129473Spjd{
259131878Spjd	u_int stripesize;
260131878Spjd	size_t len;
261131878Spjd
262131878Spjd	stripesize = sc->sc_stripesize;
263131878Spjd	len = (size_t)(stripesize - (offset & (stripesize - 1)));
264131878Spjd	do {
265131878Spjd		bcopy(src, dst, len);
266131878Spjd		if (mode) {
267131878Spjd			dst += len + stripesize * (sc->sc_ndisks - 1);
268131878Spjd			src += len;
269131878Spjd		} else {
270131878Spjd			dst += len;
271131878Spjd			src += len + stripesize * (sc->sc_ndisks - 1);
272131878Spjd		}
273131878Spjd		length -= len;
274131878Spjd		KASSERT(length >= 0,
275131878Spjd		    ("Length < 0 (stripesize=%zu, offset=%jd, length=%jd).",
276131878Spjd		    (size_t)stripesize, (intmax_t)offset, (intmax_t)length));
277131878Spjd		if (length > stripesize)
278131878Spjd			len = stripesize;
279131878Spjd		else
280131878Spjd			len = length;
281131878Spjd	} while (length > 0);
282131878Spjd}
283131878Spjd
284131878Spjdstatic void
285131878Spjdg_stripe_done(struct bio *bp)
286131878Spjd{
287129473Spjd	struct g_stripe_softc *sc;
288131878Spjd	struct bio *pbp;
289131878Spjd
290131878Spjd	pbp = bp->bio_parent;
291131878Spjd	sc = pbp->bio_to->geom->softc;
292131878Spjd	if (pbp->bio_error == 0)
293131878Spjd		pbp->bio_error = bp->bio_error;
294131878Spjd	pbp->bio_completed += bp->bio_completed;
295133204Spjd	if (bp->bio_cmd == BIO_READ && bp->bio_caller1 != NULL) {
296133204Spjd		g_stripe_copy(sc, bp->bio_data, bp->bio_caller1, bp->bio_offset,
297131878Spjd		    bp->bio_length, 1);
298133204Spjd		bp->bio_data = bp->bio_caller1;
299133204Spjd		bp->bio_caller1 = NULL;
300131878Spjd	}
301131878Spjd	g_destroy_bio(bp);
302131878Spjd	pbp->bio_inbed++;
303131878Spjd	if (pbp->bio_children == pbp->bio_inbed) {
304133204Spjd		if (pbp->bio_driver1 != NULL)
305133204Spjd			uma_zfree(g_stripe_zone, pbp->bio_driver1);
306131878Spjd		g_io_deliver(pbp, pbp->bio_error);
307131878Spjd	}
308131878Spjd}
309131878Spjd
310131878Spjdstatic int
311131878Spjdg_stripe_start_fast(struct bio *bp, u_int no, off_t offset, off_t length)
312131878Spjd{
313131878Spjd	TAILQ_HEAD(, bio) queue = TAILQ_HEAD_INITIALIZER(queue);
314131878Spjd	u_int nparts = 0, stripesize;
315131878Spjd	struct g_stripe_softc *sc;
316131878Spjd	char *addr, *data = NULL;
317129473Spjd	struct bio *cbp;
318131878Spjd	int error;
319131878Spjd
320131878Spjd	sc = bp->bio_to->geom->softc;
321131878Spjd
322131878Spjd	addr = bp->bio_data;
323131878Spjd	stripesize = sc->sc_stripesize;
324131878Spjd
325131878Spjd	cbp = g_clone_bio(bp);
326131878Spjd	if (cbp == NULL) {
327131878Spjd		error = ENOMEM;
328131878Spjd		goto failure;
329131878Spjd	}
330131878Spjd	TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
331131878Spjd	nparts++;
332131878Spjd	/*
333131878Spjd	 * Fill in the component buf structure.
334131878Spjd	 */
335131878Spjd	cbp->bio_done = g_stripe_done;
336131878Spjd	cbp->bio_offset = offset;
337131878Spjd	cbp->bio_data = addr;
338133204Spjd	cbp->bio_caller1 = NULL;
339131878Spjd	cbp->bio_length = length;
340133204Spjd	cbp->bio_caller2 = sc->sc_disks[no];
341131878Spjd
342131878Spjd	/* offset -= offset % stripesize; */
343131878Spjd	offset -= offset & (stripesize - 1);
344131878Spjd	addr += length;
345131878Spjd	length = bp->bio_length - length;
346131878Spjd	for (no++; length > 0; no++, length -= stripesize, addr += stripesize) {
347131878Spjd		if (no > sc->sc_ndisks - 1) {
348131878Spjd			no = 0;
349131878Spjd			offset += stripesize;
350131878Spjd		}
351131878Spjd		if (nparts >= sc->sc_ndisks) {
352131878Spjd			cbp = TAILQ_NEXT(cbp, bio_queue);
353131878Spjd			if (cbp == NULL)
354131878Spjd				cbp = TAILQ_FIRST(&queue);
355131878Spjd			nparts++;
356131878Spjd			/*
357131878Spjd			 * Update bio structure.
358131878Spjd			 */
359131878Spjd			/*
360131878Spjd			 * MIN() is in case when
361131878Spjd			 * (bp->bio_length % sc->sc_stripesize) != 0.
362131878Spjd			 */
363131878Spjd			cbp->bio_length += MIN(stripesize, length);
364133204Spjd			if (cbp->bio_caller1 == NULL) {
365133204Spjd				cbp->bio_caller1 = cbp->bio_data;
366131878Spjd				cbp->bio_data = NULL;
367131878Spjd				if (data == NULL) {
368131878Spjd					data = uma_zalloc(g_stripe_zone,
369131878Spjd					    M_NOWAIT);
370131878Spjd					if (data == NULL) {
371131878Spjd						error = ENOMEM;
372131878Spjd						goto failure;
373131878Spjd					}
374131878Spjd				}
375131878Spjd			}
376131878Spjd		} else {
377131878Spjd			cbp = g_clone_bio(bp);
378131878Spjd			if (cbp == NULL) {
379131878Spjd				error = ENOMEM;
380131878Spjd				goto failure;
381131878Spjd			}
382131878Spjd			TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
383131878Spjd			nparts++;
384131878Spjd			/*
385131878Spjd			 * Fill in the component buf structure.
386131878Spjd			 */
387131878Spjd			cbp->bio_done = g_stripe_done;
388131878Spjd			cbp->bio_offset = offset;
389131878Spjd			cbp->bio_data = addr;
390133204Spjd			cbp->bio_caller1 = NULL;
391131878Spjd			/*
392131878Spjd			 * MIN() is in case when
393131878Spjd			 * (bp->bio_length % sc->sc_stripesize) != 0.
394131878Spjd			 */
395131878Spjd			cbp->bio_length = MIN(stripesize, length);
396133204Spjd			cbp->bio_caller2 = sc->sc_disks[no];
397131878Spjd		}
398131878Spjd	}
399131878Spjd	if (data != NULL)
400133444Spjd		bp->bio_driver1 = data;
401131878Spjd	/*
402131878Spjd	 * Fire off all allocated requests!
403131878Spjd	 */
404131878Spjd	while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
405131878Spjd		struct g_consumer *cp;
406131878Spjd
407131878Spjd		TAILQ_REMOVE(&queue, cbp, bio_queue);
408133204Spjd		cp = cbp->bio_caller2;
409133204Spjd		cbp->bio_caller2 = NULL;
410131878Spjd		cbp->bio_to = cp->provider;
411133204Spjd		if (cbp->bio_caller1 != NULL) {
412131878Spjd			cbp->bio_data = data;
413131878Spjd			if (bp->bio_cmd == BIO_WRITE) {
414133204Spjd				g_stripe_copy(sc, cbp->bio_caller1, data,
415131878Spjd				    cbp->bio_offset, cbp->bio_length, 0);
416131878Spjd			}
417131878Spjd			data += cbp->bio_length;
418131878Spjd		}
419131878Spjd		G_STRIPE_LOGREQ(cbp, "Sending request.");
420131878Spjd		g_io_request(cbp, cp);
421131878Spjd	}
422131878Spjd	return (0);
423131878Spjdfailure:
424131878Spjd	if (data != NULL)
425131878Spjd		uma_zfree(g_stripe_zone, data);
426131878Spjd	while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
427131878Spjd		TAILQ_REMOVE(&queue, cbp, bio_queue);
428133204Spjd		if (cbp->bio_caller1 != NULL) {
429133204Spjd			cbp->bio_data = cbp->bio_caller1;
430133204Spjd			cbp->bio_caller1 = NULL;
431131878Spjd		}
432133201Spjd		bp->bio_children--;
433131878Spjd		g_destroy_bio(cbp);
434131878Spjd	}
435131878Spjd	return (error);
436131878Spjd}
437131878Spjd
438131878Spjdstatic int
439131878Spjdg_stripe_start_economic(struct bio *bp, u_int no, off_t offset, off_t length)
440131878Spjd{
441131878Spjd	TAILQ_HEAD(, bio) queue = TAILQ_HEAD_INITIALIZER(queue);
442131878Spjd	struct g_stripe_softc *sc;
443129473Spjd	uint32_t stripesize;
444131878Spjd	struct bio *cbp;
445129473Spjd	char *addr;
446131878Spjd	int error;
447129473Spjd
448131878Spjd	sc = bp->bio_to->geom->softc;
449131878Spjd
450131878Spjd	addr = bp->bio_data;
451131878Spjd	stripesize = sc->sc_stripesize;
452131878Spjd
453131878Spjd	cbp = g_clone_bio(bp);
454131878Spjd	if (cbp == NULL) {
455131878Spjd		error = ENOMEM;
456131878Spjd		goto failure;
457131878Spjd	}
458131878Spjd	TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
459129473Spjd	/*
460131878Spjd	 * Fill in the component buf structure.
461131878Spjd	 */
462131878Spjd	cbp->bio_done = g_std_done;
463131878Spjd	cbp->bio_offset = offset;
464131878Spjd	cbp->bio_data = addr;
465131878Spjd	cbp->bio_length = length;
466133204Spjd	cbp->bio_caller2 = sc->sc_disks[no];
467131878Spjd
468131878Spjd	/* offset -= offset % stripesize; */
469131878Spjd	offset -= offset & (stripesize - 1);
470131878Spjd	addr += length;
471131878Spjd	length = bp->bio_length - length;
472131878Spjd	for (no++; length > 0; no++, length -= stripesize, addr += stripesize) {
473131878Spjd		if (no > sc->sc_ndisks - 1) {
474131878Spjd			no = 0;
475131878Spjd			offset += stripesize;
476131878Spjd		}
477131878Spjd		cbp = g_clone_bio(bp);
478131878Spjd		if (cbp == NULL) {
479131878Spjd			error = ENOMEM;
480131878Spjd			goto failure;
481131878Spjd		}
482131878Spjd		TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
483131878Spjd
484131878Spjd		/*
485131878Spjd		 * Fill in the component buf structure.
486131878Spjd		 */
487131878Spjd		cbp->bio_done = g_std_done;
488131878Spjd		cbp->bio_offset = offset;
489131878Spjd		cbp->bio_data = addr;
490131878Spjd		/*
491131878Spjd		 * MIN() is in case when
492131878Spjd		 * (bp->bio_length % sc->sc_stripesize) != 0.
493131878Spjd		 */
494131878Spjd		cbp->bio_length = MIN(stripesize, length);
495131878Spjd
496133204Spjd		cbp->bio_caller2 = sc->sc_disks[no];
497131878Spjd	}
498131878Spjd	/*
499131878Spjd	 * Fire off all allocated requests!
500131878Spjd	 */
501131878Spjd	while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
502131878Spjd		struct g_consumer *cp;
503131878Spjd
504131878Spjd		TAILQ_REMOVE(&queue, cbp, bio_queue);
505133204Spjd		cp = cbp->bio_caller2;
506133204Spjd		cbp->bio_caller2 = NULL;
507131878Spjd		cbp->bio_to = cp->provider;
508131878Spjd		G_STRIPE_LOGREQ(cbp, "Sending request.");
509131878Spjd		g_io_request(cbp, cp);
510131878Spjd	}
511131878Spjd	return (0);
512131878Spjdfailure:
513131878Spjd	while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
514131878Spjd		TAILQ_REMOVE(&queue, cbp, bio_queue);
515133201Spjd		bp->bio_children--;
516131878Spjd		g_destroy_bio(cbp);
517131878Spjd	}
518131878Spjd	return (error);
519131878Spjd}
520131878Spjd
521131878Spjdstatic void
522131878Spjdg_stripe_start(struct bio *bp)
523131878Spjd{
524131878Spjd	off_t offset, start, length, nstripe;
525131878Spjd	struct g_stripe_softc *sc;
526131878Spjd	u_int no, stripesize;
527131878Spjd	int error, fast = 0;
528131878Spjd
529131878Spjd	sc = bp->bio_to->geom->softc;
530131878Spjd	/*
531129473Spjd	 * If sc == NULL, provider's error should be set and g_stripe_start()
532129473Spjd	 * should not be called at all.
533129473Spjd	 */
534129473Spjd	KASSERT(sc != NULL,
535129473Spjd	    ("Provider's error should be set (error=%d)(device=%s).",
536129473Spjd	    bp->bio_to->error, bp->bio_to->name));
537129473Spjd
538129473Spjd	G_STRIPE_LOGREQ(bp, "Request received.");
539129473Spjd
540129473Spjd	switch (bp->bio_cmd) {
541129473Spjd	case BIO_READ:
542129473Spjd	case BIO_WRITE:
543129473Spjd	case BIO_DELETE:
544129473Spjd		/*
545129473Spjd		 * Only those requests are supported.
546129473Spjd		 */
547129473Spjd		break;
548129473Spjd	case BIO_GETATTR:
549129473Spjd		/* To which provider it should be delivered? */
550129473Spjd	default:
551129473Spjd		g_io_deliver(bp, EOPNOTSUPP);
552129473Spjd		return;
553129473Spjd	}
554129473Spjd
555129473Spjd	stripesize = sc->sc_stripesize;
556129473Spjd
557129473Spjd	/*
558131878Spjd	 * Calculations are quite messy, but fast I hope.
559129473Spjd	 */
560129473Spjd
561129473Spjd	/* Stripe number. */
562129473Spjd	/* nstripe = bp->bio_offset / stripesize; */
563129473Spjd	nstripe = bp->bio_offset >> (off_t)sc->sc_stripebits;
564129473Spjd	/* Disk number. */
565129473Spjd	no = nstripe % sc->sc_ndisks;
566129473Spjd	/* Start position in stripe. */
567129473Spjd	/* start = bp->bio_offset % stripesize; */
568129473Spjd	start = bp->bio_offset & (stripesize - 1);
569129473Spjd	/* Start position in disk. */
570131878Spjd	/* offset = (nstripe / sc->sc_ndisks) * stripesize + start; */
571131878Spjd	offset = ((nstripe / sc->sc_ndisks) << sc->sc_stripebits) + start;
572129473Spjd	/* Length of data to operate. */
573129473Spjd	length = MIN(bp->bio_length, stripesize - start);
574129473Spjd
575131878Spjd	/*
576131878Spjd	 * Do use "fast" mode when:
577131878Spjd	 * 1. "Fast" mode is ON.
578131878Spjd	 * and
579131878Spjd	 * 2. Request size is less than or equal to MAX_IO_SIZE (128kB),
580131878Spjd	 *    which should always be true.
581131878Spjd	 * and
582131878Spjd	 * 3. Request size is bigger than stripesize * ndisks. If it isn't,
583131878Spjd	 *    there will be no need to send more than one I/O request to
584131878Spjd	 *    a provider, so there is nothing to optmize.
585131878Spjd	 */
586131878Spjd	if (g_stripe_fast && bp->bio_length <= MAX_IO_SIZE &&
587131878Spjd	    bp->bio_length >= stripesize * sc->sc_ndisks) {
588131878Spjd		fast = 1;
589129473Spjd	}
590131878Spjd	error = 0;
591133205Spjd	if (fast) {
592131878Spjd		error = g_stripe_start_fast(bp, no, offset, length);
593133205Spjd		if (error != 0)
594133205Spjd			g_stripe_fast_failed++;
595133205Spjd	}
596129473Spjd	/*
597131878Spjd	 * Do use "economic" when:
598131878Spjd	 * 1. "Economic" mode is ON.
599131878Spjd	 * or
600131878Spjd	 * 2. "Fast" mode failed. It can only failed if there is no memory.
601129473Spjd	 */
602131878Spjd	if (!fast || error != 0)
603131878Spjd		error = g_stripe_start_economic(bp, no, offset, length);
604131878Spjd	if (error != 0) {
605131878Spjd		if (bp->bio_error == 0)
606131878Spjd			bp->bio_error = error;
607131878Spjd		g_io_deliver(bp, bp->bio_error);
608129473Spjd	}
609129473Spjd}
610129473Spjd
611129473Spjdstatic void
612129473Spjdg_stripe_check_and_run(struct g_stripe_softc *sc)
613129473Spjd{
614129473Spjd	off_t mediasize, ms;
615129473Spjd	u_int no, sectorsize = 0;
616129473Spjd
617129473Spjd	if (g_stripe_nvalid(sc) != sc->sc_ndisks)
618129473Spjd		return;
619129473Spjd
620132664Spjd	sc->sc_provider = g_new_providerf(sc->sc_geom, "stripe/%s",
621132664Spjd	    sc->sc_name);
622129473Spjd	/*
623129473Spjd	 * Find the smallest disk.
624129473Spjd	 */
625129473Spjd	mediasize = sc->sc_disks[0]->provider->mediasize;
626129473Spjd	if (sc->sc_type == G_STRIPE_TYPE_AUTOMATIC)
627129473Spjd		mediasize -= sc->sc_disks[0]->provider->sectorsize;
628129473Spjd	mediasize -= mediasize % sc->sc_stripesize;
629129473Spjd	sectorsize = sc->sc_disks[0]->provider->sectorsize;
630129473Spjd	for (no = 1; no < sc->sc_ndisks; no++) {
631129473Spjd		ms = sc->sc_disks[no]->provider->mediasize;
632129473Spjd		if (sc->sc_type == G_STRIPE_TYPE_AUTOMATIC)
633129473Spjd			ms -= sc->sc_disks[no]->provider->sectorsize;
634129473Spjd		ms -= ms % sc->sc_stripesize;
635129473Spjd		if (ms < mediasize)
636129473Spjd			mediasize = ms;
637129473Spjd		sectorsize = lcm(sectorsize,
638129473Spjd		    sc->sc_disks[no]->provider->sectorsize);
639129473Spjd	}
640129473Spjd	sc->sc_provider->sectorsize = sectorsize;
641129473Spjd	sc->sc_provider->mediasize = mediasize * sc->sc_ndisks;
642129473Spjd	g_error_provider(sc->sc_provider, 0);
643129473Spjd
644132664Spjd	G_STRIPE_DEBUG(0, "Device %s activated.", sc->sc_name);
645129473Spjd}
646129473Spjd
647129473Spjdstatic int
648129473Spjdg_stripe_read_metadata(struct g_consumer *cp, struct g_stripe_metadata *md)
649129473Spjd{
650129473Spjd	struct g_provider *pp;
651129473Spjd	u_char *buf;
652129473Spjd	int error;
653129473Spjd
654129473Spjd	g_topology_assert();
655129473Spjd
656129473Spjd	error = g_access(cp, 1, 0, 0);
657129473Spjd	if (error != 0)
658129473Spjd		return (error);
659129473Spjd	pp = cp->provider;
660129473Spjd	g_topology_unlock();
661129473Spjd	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
662129473Spjd	    &error);
663129473Spjd	g_topology_lock();
664129473Spjd	g_access(cp, -1, 0, 0);
665129473Spjd	if (buf == NULL)
666129473Spjd		return (error);
667129473Spjd
668129473Spjd	/* Decode metadata. */
669129473Spjd	stripe_metadata_decode(buf, md);
670129473Spjd	g_free(buf);
671129473Spjd
672129473Spjd	return (0);
673129473Spjd}
674129473Spjd
675129473Spjd/*
676129473Spjd * Add disk to given device.
677129473Spjd */
678129473Spjdstatic int
679129473Spjdg_stripe_add_disk(struct g_stripe_softc *sc, struct g_provider *pp, u_int no)
680129473Spjd{
681129473Spjd	struct g_consumer *cp, *fcp;
682129473Spjd	struct g_geom *gp;
683129473Spjd	int error;
684129473Spjd
685129473Spjd	/* Metadata corrupted? */
686129473Spjd	if (no >= sc->sc_ndisks)
687129473Spjd		return (EINVAL);
688129473Spjd
689129473Spjd	/* Check if disk is not already attached. */
690129473Spjd	if (sc->sc_disks[no] != NULL)
691129473Spjd		return (EEXIST);
692129473Spjd
693129473Spjd	gp = sc->sc_geom;
694129473Spjd	fcp = LIST_FIRST(&gp->consumer);
695129473Spjd
696129473Spjd	cp = g_new_consumer(gp);
697129473Spjd	error = g_attach(cp, pp);
698129473Spjd	if (error != 0) {
699129473Spjd		g_destroy_consumer(cp);
700129473Spjd		return (error);
701129473Spjd	}
702129473Spjd
703129473Spjd	if (fcp != NULL && (fcp->acr > 0 || fcp->acw > 0 || fcp->ace > 0)) {
704129473Spjd		error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
705129473Spjd		if (error != 0) {
706129473Spjd			g_detach(cp);
707129473Spjd			g_destroy_consumer(cp);
708129473Spjd			return (error);
709129473Spjd		}
710129473Spjd	}
711129473Spjd	if (sc->sc_type == G_STRIPE_TYPE_AUTOMATIC) {
712129473Spjd		struct g_stripe_metadata md;
713129473Spjd
714129473Spjd		/* Reread metadata. */
715129473Spjd		error = g_stripe_read_metadata(cp, &md);
716129473Spjd		if (error != 0)
717129473Spjd			goto fail;
718129473Spjd
719129473Spjd		if (strcmp(md.md_magic, G_STRIPE_MAGIC) != 0 ||
720129473Spjd		    strcmp(md.md_name, sc->sc_name) != 0 ||
721129473Spjd		    md.md_id != sc->sc_id) {
722129473Spjd			G_STRIPE_DEBUG(0, "Metadata on %s changed.", pp->name);
723129473Spjd			goto fail;
724129473Spjd		}
725129473Spjd	}
726129473Spjd
727129473Spjd	cp->private = sc;
728129473Spjd	cp->index = no;
729129473Spjd	sc->sc_disks[no] = cp;
730129473Spjd
731132664Spjd	G_STRIPE_DEBUG(0, "Disk %s attached to %s.", pp->name, sc->sc_name);
732129473Spjd
733129473Spjd	g_stripe_check_and_run(sc);
734129473Spjd
735129473Spjd	return (0);
736129473Spjdfail:
737129473Spjd	if (fcp != NULL && (fcp->acr > 0 || fcp->acw > 0 || fcp->ace > 0))
738129473Spjd		g_access(cp, -fcp->acr, -fcp->acw, -fcp->ace);
739129473Spjd	g_detach(cp);
740129473Spjd	g_destroy_consumer(cp);
741129473Spjd	return (error);
742129473Spjd}
743129473Spjd
744129473Spjdstatic struct g_geom *
745129473Spjdg_stripe_create(struct g_class *mp, const struct g_stripe_metadata *md,
746129473Spjd    u_int type)
747129473Spjd{
748129473Spjd	struct g_stripe_softc *sc;
749129473Spjd	struct g_geom *gp;
750129473Spjd	u_int no;
751129473Spjd
752132664Spjd	G_STRIPE_DEBUG(1, "Creating device %s (id=%u).", md->md_name,
753129473Spjd	    md->md_id);
754129473Spjd
755129473Spjd	/* Two disks is minimum. */
756132664Spjd	if (md->md_all < 2) {
757132664Spjd		G_STRIPE_DEBUG(0, "Too few disks defined for %s.", md->md_name);
758129473Spjd		return (NULL);
759129473Spjd	}
760129473Spjd#if 0
761129473Spjd	/* Stripe size have to be grater than or equal to sector size. */
762129473Spjd	if (md->md_stripesize < sectorsize) {
763132664Spjd		G_STRIPE_DEBUG(0, "Invalid stripe size for %s.", md->md_name);
764129473Spjd		return (NULL);
765129473Spjd	}
766129473Spjd#endif
767129473Spjd	/* Stripe size have to be power of 2. */
768129473Spjd	if (!powerof2(md->md_stripesize)) {
769132664Spjd		G_STRIPE_DEBUG(0, "Invalid stripe size for %s.", md->md_name);
770129473Spjd		return (NULL);
771129473Spjd	}
772129473Spjd
773129473Spjd	/* Check for duplicate unit */
774129473Spjd	LIST_FOREACH(gp, &mp->geom, geom) {
775129473Spjd		sc = gp->softc;
776129473Spjd		if (sc != NULL && strcmp(sc->sc_name, md->md_name) == 0) {
777129473Spjd			G_STRIPE_DEBUG(0, "Device %s already configured.",
778132664Spjd			    sc->sc_name);
779129473Spjd			return (NULL);
780129473Spjd		}
781129473Spjd	}
782132664Spjd	gp = g_new_geomf(mp, "%s", md->md_name);
783129473Spjd	gp->softc = NULL;	/* for a moment */
784129473Spjd
785132662Spjd	sc = malloc(sizeof(*sc), M_STRIPE, M_WAITOK | M_ZERO);
786129473Spjd	gp->start = g_stripe_start;
787129473Spjd	gp->spoiled = g_stripe_orphan;
788129473Spjd	gp->orphan = g_stripe_orphan;
789129473Spjd	gp->access = g_stripe_access;
790129473Spjd	gp->dumpconf = g_stripe_dumpconf;
791129473Spjd
792129473Spjd	sc->sc_id = md->md_id;
793129473Spjd	sc->sc_stripesize = md->md_stripesize;
794129473Spjd	sc->sc_stripebits = BITCOUNT(sc->sc_stripesize - 1);
795129473Spjd	sc->sc_ndisks = md->md_all;
796129473Spjd	sc->sc_disks = malloc(sizeof(struct g_consumer *) * sc->sc_ndisks,
797129473Spjd	    M_STRIPE, M_WAITOK | M_ZERO);
798129473Spjd	for (no = 0; no < sc->sc_ndisks; no++)
799129473Spjd		sc->sc_disks[no] = NULL;
800129473Spjd	sc->sc_type = type;
801129473Spjd
802129473Spjd	gp->softc = sc;
803129473Spjd	sc->sc_geom = gp;
804129473Spjd	sc->sc_provider = NULL;
805129473Spjd
806132664Spjd	G_STRIPE_DEBUG(0, "Device %s created (id=%u).", sc->sc_name, sc->sc_id);
807129473Spjd
808129473Spjd	return (gp);
809129473Spjd}
810129473Spjd
811129473Spjdstatic int
812129473Spjdg_stripe_destroy(struct g_stripe_softc *sc, boolean_t force)
813129473Spjd{
814129473Spjd	struct g_provider *pp;
815129473Spjd	struct g_geom *gp;
816129473Spjd	u_int no;
817129473Spjd
818129473Spjd	g_topology_assert();
819129473Spjd
820129473Spjd	if (sc == NULL)
821129473Spjd		return (ENXIO);
822129473Spjd
823129473Spjd	pp = sc->sc_provider;
824129473Spjd	if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
825129473Spjd		if (force) {
826129473Spjd			G_STRIPE_DEBUG(0, "Device %s is still open, so it "
827129473Spjd			    "can't be definitely removed.", pp->name);
828129473Spjd		} else {
829129473Spjd			G_STRIPE_DEBUG(1,
830129473Spjd			    "Device %s is still open (r%dw%de%d).", pp->name,
831129473Spjd			    pp->acr, pp->acw, pp->ace);
832129473Spjd			return (EBUSY);
833129473Spjd		}
834129473Spjd	}
835129473Spjd
836129473Spjd	for (no = 0; no < sc->sc_ndisks; no++) {
837129473Spjd		if (sc->sc_disks[no] != NULL)
838129473Spjd			g_stripe_remove_disk(sc->sc_disks[no]);
839129473Spjd	}
840129473Spjd
841129473Spjd	gp = sc->sc_geom;
842129473Spjd	gp->softc = NULL;
843129473Spjd	KASSERT(sc->sc_provider == NULL, ("Provider still exists? (device=%s)",
844129473Spjd	    gp->name));
845129473Spjd	free(sc->sc_disks, M_STRIPE);
846129473Spjd	free(sc, M_STRIPE);
847129473Spjd
848129473Spjd	pp = LIST_FIRST(&gp->provider);
849129473Spjd	if (pp == NULL || (pp->acr == 0 && pp->acw == 0 && pp->ace == 0))
850129473Spjd		G_STRIPE_DEBUG(0, "Device %s destroyed.", gp->name);
851129473Spjd
852129473Spjd	g_wither_geom(gp, ENXIO);
853129473Spjd
854129473Spjd	return (0);
855129473Spjd}
856129473Spjd
857129473Spjdstatic int
858129473Spjdg_stripe_destroy_geom(struct gctl_req *req __unused,
859129473Spjd    struct g_class *mp __unused, struct g_geom *gp)
860129473Spjd{
861129473Spjd	struct g_stripe_softc *sc;
862129473Spjd
863129473Spjd	sc = gp->softc;
864129473Spjd	return (g_stripe_destroy(sc, 0));
865129473Spjd}
866129473Spjd
867129473Spjdstatic struct g_geom *
868129473Spjdg_stripe_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
869129473Spjd{
870129473Spjd	struct g_stripe_metadata md;
871129473Spjd	struct g_stripe_softc *sc;
872129473Spjd	struct g_consumer *cp;
873129473Spjd	struct g_geom *gp;
874129473Spjd	int error;
875129473Spjd
876129473Spjd	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
877129473Spjd	g_topology_assert();
878129473Spjd
879129473Spjd	G_STRIPE_DEBUG(3, "Tasting %s.", pp->name);
880129473Spjd
881129473Spjd	gp = g_new_geomf(mp, "stripe:taste");
882129473Spjd	gp->start = g_stripe_start;
883129473Spjd	gp->access = g_stripe_access;
884129473Spjd	gp->orphan = g_stripe_orphan;
885129473Spjd	cp = g_new_consumer(gp);
886129473Spjd	g_attach(cp, pp);
887129473Spjd	error = g_stripe_read_metadata(cp, &md);
888133371Spjd	g_detach(cp);
889133371Spjd	g_destroy_consumer(cp);
890133371Spjd	g_destroy_geom(gp);
891129473Spjd	if (error != 0)
892129473Spjd		return (NULL);
893129473Spjd	gp = NULL;
894129473Spjd
895129473Spjd	if (strcmp(md.md_magic, G_STRIPE_MAGIC) != 0)
896129473Spjd		return (NULL);
897129473Spjd	if (md.md_version > G_STRIPE_VERSION) {
898129473Spjd		printf("geom_stripe.ko module is too old to handle %s.\n",
899129473Spjd		    pp->name);
900129473Spjd		return (NULL);
901129473Spjd	}
902133373Spjd	/*
903133373Spjd	 * Backward compatibility:
904133373Spjd	 */
905142727Spjd	/* There was no md_provider field in earlier versions of metadata. */
906133373Spjd	if (md.md_version < 2)
907133373Spjd		bzero(md.md_provider, sizeof(md.md_provider));
908142727Spjd	/* There was no md_provsize field in earlier versions of metadata. */
909142727Spjd	if (md.md_version < 3)
910142727Spjd		md.md_provsize = pp->mediasize;
911129473Spjd
912133373Spjd	if (md.md_provider[0] != '\0' && strcmp(md.md_provider, pp->name) != 0)
913133373Spjd		return (NULL);
914142727Spjd	if (md.md_provsize != pp->mediasize)
915142727Spjd		return (NULL);
916133373Spjd
917129473Spjd	/*
918129473Spjd	 * Let's check if device already exists.
919129473Spjd	 */
920129473Spjd	sc = NULL;
921129473Spjd	LIST_FOREACH(gp, &mp->geom, geom) {
922129473Spjd		sc = gp->softc;
923129473Spjd		if (sc == NULL)
924129473Spjd			continue;
925129473Spjd		if (sc->sc_type != G_STRIPE_TYPE_AUTOMATIC)
926129473Spjd			continue;
927129473Spjd		if (strcmp(md.md_name, sc->sc_name) != 0)
928129473Spjd			continue;
929129473Spjd		if (md.md_id != sc->sc_id)
930129473Spjd			continue;
931129473Spjd		break;
932129473Spjd	}
933129473Spjd	if (gp != NULL) {
934129473Spjd		G_STRIPE_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
935129473Spjd		error = g_stripe_add_disk(sc, pp, md.md_no);
936129473Spjd		if (error != 0) {
937129473Spjd			G_STRIPE_DEBUG(0,
938129473Spjd			    "Cannot add disk %s to %s (error=%d).", pp->name,
939129473Spjd			    gp->name, error);
940129473Spjd			return (NULL);
941129473Spjd		}
942129473Spjd	} else {
943129473Spjd		gp = g_stripe_create(mp, &md, G_STRIPE_TYPE_AUTOMATIC);
944129473Spjd		if (gp == NULL) {
945132664Spjd			G_STRIPE_DEBUG(0, "Cannot create device %s.",
946129473Spjd			    md.md_name);
947129473Spjd			return (NULL);
948129473Spjd		}
949129473Spjd		sc = gp->softc;
950129473Spjd		G_STRIPE_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
951129473Spjd		error = g_stripe_add_disk(sc, pp, md.md_no);
952129473Spjd		if (error != 0) {
953129473Spjd			G_STRIPE_DEBUG(0,
954129473Spjd			    "Cannot add disk %s to %s (error=%d).", pp->name,
955129473Spjd			    gp->name, error);
956129473Spjd			g_stripe_destroy(sc, 1);
957129473Spjd			return (NULL);
958129473Spjd		}
959129473Spjd	}
960129473Spjd
961129473Spjd	return (gp);
962129473Spjd}
963129473Spjd
964129473Spjdstatic void
965129473Spjdg_stripe_ctl_create(struct gctl_req *req, struct g_class *mp)
966129473Spjd{
967129473Spjd	u_int attached, no;
968129473Spjd	struct g_stripe_metadata md;
969129473Spjd	struct g_provider *pp;
970129473Spjd	struct g_stripe_softc *sc;
971129473Spjd	struct g_geom *gp;
972129473Spjd	struct sbuf *sb;
973129473Spjd	intmax_t *stripesize;
974129473Spjd	const char *name;
975129473Spjd	char param[16];
976129473Spjd	int *nargs;
977129473Spjd
978129473Spjd	g_topology_assert();
979129473Spjd	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
980129473Spjd	if (nargs == NULL) {
981129473Spjd		gctl_error(req, "No '%s' argument.", "nargs");
982129473Spjd		return;
983129473Spjd	}
984129473Spjd	if (*nargs <= 2) {
985129473Spjd		gctl_error(req, "Too few arguments.");
986129473Spjd		return;
987129473Spjd	}
988129473Spjd
989129473Spjd	strlcpy(md.md_magic, G_STRIPE_MAGIC, sizeof(md.md_magic));
990129473Spjd	md.md_version = G_STRIPE_VERSION;
991129473Spjd	name = gctl_get_asciiparam(req, "arg0");
992129473Spjd	if (name == NULL) {
993129473Spjd		gctl_error(req, "No 'arg%u' argument.", 0);
994129473Spjd		return;
995129473Spjd	}
996129473Spjd	strlcpy(md.md_name, name, sizeof(md.md_name));
997129473Spjd	md.md_id = arc4random();
998129473Spjd	md.md_no = 0;
999129473Spjd	md.md_all = *nargs - 1;
1000129473Spjd	stripesize = gctl_get_paraml(req, "stripesize", sizeof(*stripesize));
1001129473Spjd	if (stripesize == NULL) {
1002129473Spjd		gctl_error(req, "No '%s' argument.", "stripesize");
1003129473Spjd		return;
1004129473Spjd	}
1005129473Spjd	md.md_stripesize = *stripesize;
1006133373Spjd	bzero(md.md_provider, sizeof(md.md_provider));
1007142727Spjd	/* This field is not important here. */
1008142727Spjd	md.md_provsize = 0;
1009129473Spjd
1010129473Spjd	/* Check all providers are valid */
1011129473Spjd	for (no = 1; no < *nargs; no++) {
1012129473Spjd		snprintf(param, sizeof(param), "arg%u", no);
1013129473Spjd		name = gctl_get_asciiparam(req, param);
1014129473Spjd		if (name == NULL) {
1015129473Spjd			gctl_error(req, "No 'arg%u' argument.", no);
1016129473Spjd			return;
1017129473Spjd		}
1018129473Spjd		if (strncmp(name, "/dev/", strlen("/dev/")) == 0)
1019129473Spjd			name += strlen("/dev/");
1020129473Spjd		pp = g_provider_by_name(name);
1021129473Spjd		if (pp == NULL) {
1022129473Spjd			G_STRIPE_DEBUG(1, "Disk %s is invalid.", name);
1023129473Spjd			gctl_error(req, "Disk %s is invalid.", name);
1024129473Spjd			return;
1025129473Spjd		}
1026129473Spjd	}
1027129473Spjd
1028129473Spjd	gp = g_stripe_create(mp, &md, G_STRIPE_TYPE_MANUAL);
1029129473Spjd	if (gp == NULL) {
1030132664Spjd		gctl_error(req, "Can't configure %s.", md.md_name);
1031129473Spjd		return;
1032129473Spjd	}
1033129473Spjd
1034129473Spjd	sc = gp->softc;
1035129473Spjd	sb = sbuf_new(NULL, NULL, 0, SBUF_AUTOEXTEND);
1036129473Spjd	sbuf_printf(sb, "Can't attach disk(s) to %s:", gp->name);
1037129473Spjd	for (attached = 0, no = 1; no < *nargs; no++) {
1038129473Spjd		snprintf(param, sizeof(param), "arg%u", no);
1039129473Spjd		name = gctl_get_asciiparam(req, param);
1040129473Spjd		if (strncmp(name, "/dev/", strlen("/dev/")) == 0)
1041129473Spjd			name += strlen("/dev/");
1042129473Spjd		pp = g_provider_by_name(name);
1043129473Spjd		KASSERT(pp != NULL, ("Provider %s disappear?!", name));
1044129473Spjd		if (g_stripe_add_disk(sc, pp, no - 1) != 0) {
1045129473Spjd			G_STRIPE_DEBUG(1, "Disk %u (%s) not attached to %s.",
1046129473Spjd			    no, pp->name, gp->name);
1047129473Spjd			sbuf_printf(sb, " %s", pp->name);
1048129473Spjd			continue;
1049129473Spjd		}
1050129473Spjd		attached++;
1051129473Spjd	}
1052129473Spjd	sbuf_finish(sb);
1053129473Spjd	if (md.md_all != attached) {
1054129473Spjd		g_stripe_destroy(gp->softc, 1);
1055129473Spjd		gctl_error(req, "%s", sbuf_data(sb));
1056129473Spjd	}
1057129473Spjd	sbuf_delete(sb);
1058129473Spjd}
1059129473Spjd
1060129473Spjdstatic struct g_stripe_softc *
1061129473Spjdg_stripe_find_device(struct g_class *mp, const char *name)
1062129473Spjd{
1063129473Spjd	struct g_stripe_softc *sc;
1064129473Spjd	struct g_geom *gp;
1065129473Spjd
1066129473Spjd	LIST_FOREACH(gp, &mp->geom, geom) {
1067129473Spjd		sc = gp->softc;
1068129473Spjd		if (sc == NULL)
1069129473Spjd			continue;
1070132664Spjd		if (strcmp(sc->sc_name, name) == 0)
1071129473Spjd			return (sc);
1072129473Spjd	}
1073129473Spjd	return (NULL);
1074129473Spjd}
1075129473Spjd
1076129473Spjdstatic void
1077129473Spjdg_stripe_ctl_destroy(struct gctl_req *req, struct g_class *mp)
1078129473Spjd{
1079129473Spjd	struct g_stripe_softc *sc;
1080129473Spjd	int *force, *nargs, error;
1081129473Spjd	const char *name;
1082129473Spjd	char param[16];
1083129473Spjd	u_int i;
1084129473Spjd
1085129473Spjd	g_topology_assert();
1086129473Spjd
1087129473Spjd	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
1088129473Spjd	if (nargs == NULL) {
1089129473Spjd		gctl_error(req, "No '%s' argument.", "nargs");
1090129473Spjd		return;
1091129473Spjd	}
1092129473Spjd	if (*nargs <= 0) {
1093129473Spjd		gctl_error(req, "Missing device(s).");
1094129473Spjd		return;
1095129473Spjd	}
1096129473Spjd	force = gctl_get_paraml(req, "force", sizeof(*force));
1097129473Spjd	if (force == NULL) {
1098129473Spjd		gctl_error(req, "No '%s' argument.", "force");
1099129473Spjd		return;
1100129473Spjd	}
1101129473Spjd
1102129473Spjd	for (i = 0; i < (u_int)*nargs; i++) {
1103129473Spjd		snprintf(param, sizeof(param), "arg%u", i);
1104129473Spjd		name = gctl_get_asciiparam(req, param);
1105129473Spjd		if (name == NULL) {
1106129473Spjd			gctl_error(req, "No 'arg%u' argument.", i);
1107129473Spjd			return;
1108129473Spjd		}
1109129473Spjd		sc = g_stripe_find_device(mp, name);
1110129473Spjd		if (sc == NULL) {
1111129473Spjd			gctl_error(req, "No such device: %s.", name);
1112129473Spjd			return;
1113129473Spjd		}
1114129473Spjd		error = g_stripe_destroy(sc, *force);
1115129473Spjd		if (error != 0) {
1116129473Spjd			gctl_error(req, "Cannot destroy device %s (error=%d).",
1117132664Spjd			    sc->sc_name, error);
1118129473Spjd			return;
1119129473Spjd		}
1120129473Spjd	}
1121129473Spjd}
1122129473Spjd
1123129473Spjdstatic void
1124129473Spjdg_stripe_config(struct gctl_req *req, struct g_class *mp, const char *verb)
1125129473Spjd{
1126129473Spjd	uint32_t *version;
1127129473Spjd
1128129473Spjd	g_topology_assert();
1129129473Spjd
1130129473Spjd	version = gctl_get_paraml(req, "version", sizeof(*version));
1131129473Spjd	if (version == NULL) {
1132129473Spjd		gctl_error(req, "No '%s' argument.", "version");
1133129473Spjd		return;
1134129473Spjd	}
1135129473Spjd	if (*version != G_STRIPE_VERSION) {
1136129473Spjd		gctl_error(req, "Userland and kernel parts are out of sync.");
1137129473Spjd		return;
1138129473Spjd	}
1139129473Spjd
1140129473Spjd	if (strcmp(verb, "create") == 0) {
1141129473Spjd		g_stripe_ctl_create(req, mp);
1142129473Spjd		return;
1143131649Spjd	} else if (strcmp(verb, "destroy") == 0 ||
1144131649Spjd	    strcmp(verb, "stop") == 0) {
1145129473Spjd		g_stripe_ctl_destroy(req, mp);
1146129473Spjd		return;
1147129473Spjd	}
1148129473Spjd
1149129473Spjd	gctl_error(req, "Unknown verb.");
1150129473Spjd}
1151129473Spjd
1152129473Spjdstatic void
1153129473Spjdg_stripe_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1154129473Spjd    struct g_consumer *cp, struct g_provider *pp)
1155129473Spjd{
1156129473Spjd	struct g_stripe_softc *sc;
1157129473Spjd
1158129473Spjd	sc = gp->softc;
1159132665Spjd	if (sc == NULL)
1160129473Spjd		return;
1161132665Spjd	if (pp != NULL) {
1162132665Spjd		/* Nothing here. */
1163132665Spjd	} else if (cp != NULL) {
1164134292Spjd		sbuf_printf(sb, "%s<Number>%u</Number>\n", indent,
1165134292Spjd		    (u_int)cp->index);
1166132665Spjd	} else {
1167132665Spjd		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
1168132665Spjd		sbuf_printf(sb, "%s<Stripesize>%u</Stripesize>\n", indent,
1169132665Spjd		    (u_int)sc->sc_stripesize);
1170132665Spjd		sbuf_printf(sb, "%s<Type>", indent);
1171132665Spjd		switch (sc->sc_type) {
1172132665Spjd		case G_STRIPE_TYPE_AUTOMATIC:
1173132665Spjd			sbuf_printf(sb, "AUTOMATIC");
1174132665Spjd			break;
1175132665Spjd		case G_STRIPE_TYPE_MANUAL:
1176132665Spjd			sbuf_printf(sb, "MANUAL");
1177132665Spjd			break;
1178132665Spjd		default:
1179132665Spjd			sbuf_printf(sb, "UNKNOWN");
1180132665Spjd			break;
1181132665Spjd		}
1182132665Spjd		sbuf_printf(sb, "</Type>\n");
1183132665Spjd		sbuf_printf(sb, "%s<Status>Total=%u, Online=%u</Status>\n",
1184132665Spjd		    indent, sc->sc_ndisks, g_stripe_nvalid(sc));
1185132665Spjd		sbuf_printf(sb, "%s<State>", indent);
1186132665Spjd		if (sc->sc_provider != NULL && sc->sc_provider->error == 0)
1187132665Spjd			sbuf_printf(sb, "UP");
1188132665Spjd		else
1189132665Spjd			sbuf_printf(sb, "DOWN");
1190132665Spjd		sbuf_printf(sb, "</State>\n");
1191129473Spjd	}
1192129473Spjd}
1193129473Spjd
1194129473SpjdDECLARE_GEOM_CLASS(g_stripe_class, g_stripe);
1195