g_stripe.c revision 163886
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.
13155174Spjd *
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 163886 2006-11-01 22:16:53Z 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)
45151897Srwatsonstatic 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) {
174148092Spjd		sc->sc_provider->flags |= G_PF_WITHER;
175129473Spjd		g_orphan_provider(sc->sc_provider, ENXIO);
176129473Spjd		sc->sc_provider = NULL;
177132664Spjd		G_STRIPE_DEBUG(0, "Device %s removed.", sc->sc_name);
178129473Spjd	}
179129473Spjd
180129473Spjd	if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
181129473Spjd		g_access(cp, -cp->acr, -cp->acw, -cp->ace);
182129473Spjd	g_detach(cp);
183129473Spjd	g_destroy_consumer(cp);
184129473Spjd}
185129473Spjd
186129473Spjdstatic void
187129473Spjdg_stripe_orphan(struct g_consumer *cp)
188129473Spjd{
189129473Spjd	struct g_stripe_softc *sc;
190129473Spjd	struct g_geom *gp;
191129473Spjd
192129473Spjd	g_topology_assert();
193129473Spjd	gp = cp->geom;
194129473Spjd	sc = gp->softc;
195129473Spjd	if (sc == NULL)
196129473Spjd		return;
197129473Spjd
198129473Spjd	g_stripe_remove_disk(cp);
199129473Spjd	/* If there are no valid disks anymore, remove device. */
200129473Spjd	if (g_stripe_nvalid(sc) == 0)
201129473Spjd		g_stripe_destroy(sc, 1);
202129473Spjd}
203129473Spjd
204129473Spjdstatic int
205129473Spjdg_stripe_access(struct g_provider *pp, int dr, int dw, int de)
206129473Spjd{
207129473Spjd	struct g_consumer *cp1, *cp2;
208129473Spjd	struct g_stripe_softc *sc;
209129473Spjd	struct g_geom *gp;
210129473Spjd	int error;
211129473Spjd
212129473Spjd	gp = pp->geom;
213129473Spjd	sc = gp->softc;
214129473Spjd
215129473Spjd	if (sc == NULL) {
216129473Spjd		/*
217129473Spjd		 * It looks like geom is being withered.
218129473Spjd		 * In that case we allow only negative requests.
219129473Spjd		 */
220129473Spjd		KASSERT(dr <= 0 && dw <= 0 && de <= 0,
221129473Spjd		    ("Positive access request (device=%s).", pp->name));
222129473Spjd		if ((pp->acr + dr) == 0 && (pp->acw + dw) == 0 &&
223129473Spjd		    (pp->ace + de) == 0) {
224129473Spjd			G_STRIPE_DEBUG(0, "Device %s definitely destroyed.",
225129473Spjd			    gp->name);
226129473Spjd		}
227129473Spjd		return (0);
228129473Spjd	}
229129473Spjd
230129473Spjd	/* On first open, grab an extra "exclusive" bit */
231129473Spjd	if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)
232129473Spjd		de++;
233129473Spjd	/* ... and let go of it on last close */
234129473Spjd	if ((pp->acr + dr) == 0 && (pp->acw + dw) == 0 && (pp->ace + de) == 0)
235129473Spjd		de--;
236129473Spjd
237129473Spjd	error = ENXIO;
238129473Spjd	LIST_FOREACH(cp1, &gp->consumer, consumer) {
239129473Spjd		error = g_access(cp1, dr, dw, de);
240129473Spjd		if (error == 0)
241129473Spjd			continue;
242129473Spjd		/*
243129473Spjd		 * If we fail here, backout all previous changes.
244129473Spjd		 */
245129473Spjd		LIST_FOREACH(cp2, &gp->consumer, consumer) {
246129473Spjd			if (cp1 == cp2)
247129473Spjd				return (error);
248129473Spjd			g_access(cp2, -dr, -dw, -de);
249129473Spjd		}
250129473Spjd		/* NOTREACHED */
251129473Spjd	}
252129473Spjd
253129473Spjd	return (error);
254129473Spjd}
255129473Spjd
256129473Spjdstatic void
257131878Spjdg_stripe_copy(struct g_stripe_softc *sc, char *src, char *dst, off_t offset,
258131878Spjd    off_t length, int mode)
259129473Spjd{
260131878Spjd	u_int stripesize;
261131878Spjd	size_t len;
262131878Spjd
263131878Spjd	stripesize = sc->sc_stripesize;
264131878Spjd	len = (size_t)(stripesize - (offset & (stripesize - 1)));
265131878Spjd	do {
266131878Spjd		bcopy(src, dst, len);
267131878Spjd		if (mode) {
268131878Spjd			dst += len + stripesize * (sc->sc_ndisks - 1);
269131878Spjd			src += len;
270131878Spjd		} else {
271131878Spjd			dst += len;
272131878Spjd			src += len + stripesize * (sc->sc_ndisks - 1);
273131878Spjd		}
274131878Spjd		length -= len;
275131878Spjd		KASSERT(length >= 0,
276131878Spjd		    ("Length < 0 (stripesize=%zu, offset=%jd, length=%jd).",
277131878Spjd		    (size_t)stripesize, (intmax_t)offset, (intmax_t)length));
278131878Spjd		if (length > stripesize)
279131878Spjd			len = stripesize;
280131878Spjd		else
281131878Spjd			len = length;
282131878Spjd	} while (length > 0);
283131878Spjd}
284131878Spjd
285131878Spjdstatic void
286131878Spjdg_stripe_done(struct bio *bp)
287131878Spjd{
288129473Spjd	struct g_stripe_softc *sc;
289131878Spjd	struct bio *pbp;
290131878Spjd
291131878Spjd	pbp = bp->bio_parent;
292131878Spjd	sc = pbp->bio_to->geom->softc;
293131878Spjd	if (pbp->bio_error == 0)
294131878Spjd		pbp->bio_error = bp->bio_error;
295131878Spjd	pbp->bio_completed += bp->bio_completed;
296133204Spjd	if (bp->bio_cmd == BIO_READ && bp->bio_caller1 != NULL) {
297133204Spjd		g_stripe_copy(sc, bp->bio_data, bp->bio_caller1, bp->bio_offset,
298131878Spjd		    bp->bio_length, 1);
299133204Spjd		bp->bio_data = bp->bio_caller1;
300133204Spjd		bp->bio_caller1 = NULL;
301131878Spjd	}
302131878Spjd	g_destroy_bio(bp);
303131878Spjd	pbp->bio_inbed++;
304131878Spjd	if (pbp->bio_children == pbp->bio_inbed) {
305133204Spjd		if (pbp->bio_driver1 != NULL)
306133204Spjd			uma_zfree(g_stripe_zone, pbp->bio_driver1);
307131878Spjd		g_io_deliver(pbp, pbp->bio_error);
308131878Spjd	}
309131878Spjd}
310131878Spjd
311131878Spjdstatic int
312131878Spjdg_stripe_start_fast(struct bio *bp, u_int no, off_t offset, off_t length)
313131878Spjd{
314131878Spjd	TAILQ_HEAD(, bio) queue = TAILQ_HEAD_INITIALIZER(queue);
315131878Spjd	u_int nparts = 0, stripesize;
316131878Spjd	struct g_stripe_softc *sc;
317131878Spjd	char *addr, *data = NULL;
318129473Spjd	struct bio *cbp;
319131878Spjd	int error;
320131878Spjd
321131878Spjd	sc = bp->bio_to->geom->softc;
322131878Spjd
323131878Spjd	addr = bp->bio_data;
324131878Spjd	stripesize = sc->sc_stripesize;
325131878Spjd
326131878Spjd	cbp = g_clone_bio(bp);
327131878Spjd	if (cbp == NULL) {
328131878Spjd		error = ENOMEM;
329131878Spjd		goto failure;
330131878Spjd	}
331131878Spjd	TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
332131878Spjd	nparts++;
333131878Spjd	/*
334131878Spjd	 * Fill in the component buf structure.
335131878Spjd	 */
336131878Spjd	cbp->bio_done = g_stripe_done;
337131878Spjd	cbp->bio_offset = offset;
338131878Spjd	cbp->bio_data = addr;
339133204Spjd	cbp->bio_caller1 = NULL;
340131878Spjd	cbp->bio_length = length;
341133204Spjd	cbp->bio_caller2 = sc->sc_disks[no];
342131878Spjd
343131878Spjd	/* offset -= offset % stripesize; */
344131878Spjd	offset -= offset & (stripesize - 1);
345131878Spjd	addr += length;
346131878Spjd	length = bp->bio_length - length;
347131878Spjd	for (no++; length > 0; no++, length -= stripesize, addr += stripesize) {
348131878Spjd		if (no > sc->sc_ndisks - 1) {
349131878Spjd			no = 0;
350131878Spjd			offset += stripesize;
351131878Spjd		}
352131878Spjd		if (nparts >= sc->sc_ndisks) {
353131878Spjd			cbp = TAILQ_NEXT(cbp, bio_queue);
354131878Spjd			if (cbp == NULL)
355131878Spjd				cbp = TAILQ_FIRST(&queue);
356131878Spjd			nparts++;
357131878Spjd			/*
358131878Spjd			 * Update bio structure.
359131878Spjd			 */
360131878Spjd			/*
361131878Spjd			 * MIN() is in case when
362131878Spjd			 * (bp->bio_length % sc->sc_stripesize) != 0.
363131878Spjd			 */
364131878Spjd			cbp->bio_length += MIN(stripesize, length);
365133204Spjd			if (cbp->bio_caller1 == NULL) {
366133204Spjd				cbp->bio_caller1 = cbp->bio_data;
367131878Spjd				cbp->bio_data = NULL;
368131878Spjd				if (data == NULL) {
369131878Spjd					data = uma_zalloc(g_stripe_zone,
370131878Spjd					    M_NOWAIT);
371131878Spjd					if (data == NULL) {
372131878Spjd						error = ENOMEM;
373131878Spjd						goto failure;
374131878Spjd					}
375131878Spjd				}
376131878Spjd			}
377131878Spjd		} else {
378131878Spjd			cbp = g_clone_bio(bp);
379131878Spjd			if (cbp == NULL) {
380131878Spjd				error = ENOMEM;
381131878Spjd				goto failure;
382131878Spjd			}
383131878Spjd			TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
384131878Spjd			nparts++;
385131878Spjd			/*
386131878Spjd			 * Fill in the component buf structure.
387131878Spjd			 */
388131878Spjd			cbp->bio_done = g_stripe_done;
389131878Spjd			cbp->bio_offset = offset;
390131878Spjd			cbp->bio_data = addr;
391133204Spjd			cbp->bio_caller1 = NULL;
392131878Spjd			/*
393131878Spjd			 * MIN() is in case when
394131878Spjd			 * (bp->bio_length % sc->sc_stripesize) != 0.
395131878Spjd			 */
396131878Spjd			cbp->bio_length = MIN(stripesize, length);
397133204Spjd			cbp->bio_caller2 = sc->sc_disks[no];
398131878Spjd		}
399131878Spjd	}
400131878Spjd	if (data != NULL)
401133444Spjd		bp->bio_driver1 = data;
402131878Spjd	/*
403131878Spjd	 * Fire off all allocated requests!
404131878Spjd	 */
405131878Spjd	while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
406131878Spjd		struct g_consumer *cp;
407131878Spjd
408131878Spjd		TAILQ_REMOVE(&queue, cbp, bio_queue);
409133204Spjd		cp = cbp->bio_caller2;
410133204Spjd		cbp->bio_caller2 = NULL;
411131878Spjd		cbp->bio_to = cp->provider;
412133204Spjd		if (cbp->bio_caller1 != NULL) {
413131878Spjd			cbp->bio_data = data;
414131878Spjd			if (bp->bio_cmd == BIO_WRITE) {
415133204Spjd				g_stripe_copy(sc, cbp->bio_caller1, data,
416131878Spjd				    cbp->bio_offset, cbp->bio_length, 0);
417131878Spjd			}
418131878Spjd			data += cbp->bio_length;
419131878Spjd		}
420131878Spjd		G_STRIPE_LOGREQ(cbp, "Sending request.");
421131878Spjd		g_io_request(cbp, cp);
422131878Spjd	}
423131878Spjd	return (0);
424131878Spjdfailure:
425131878Spjd	if (data != NULL)
426131878Spjd		uma_zfree(g_stripe_zone, data);
427131878Spjd	while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
428131878Spjd		TAILQ_REMOVE(&queue, cbp, bio_queue);
429133204Spjd		if (cbp->bio_caller1 != NULL) {
430133204Spjd			cbp->bio_data = cbp->bio_caller1;
431133204Spjd			cbp->bio_caller1 = NULL;
432131878Spjd		}
433133201Spjd		bp->bio_children--;
434131878Spjd		g_destroy_bio(cbp);
435131878Spjd	}
436131878Spjd	return (error);
437131878Spjd}
438131878Spjd
439131878Spjdstatic int
440131878Spjdg_stripe_start_economic(struct bio *bp, u_int no, off_t offset, off_t length)
441131878Spjd{
442131878Spjd	TAILQ_HEAD(, bio) queue = TAILQ_HEAD_INITIALIZER(queue);
443131878Spjd	struct g_stripe_softc *sc;
444129473Spjd	uint32_t stripesize;
445131878Spjd	struct bio *cbp;
446129473Spjd	char *addr;
447131878Spjd	int error;
448129473Spjd
449131878Spjd	sc = bp->bio_to->geom->softc;
450131878Spjd
451131878Spjd	addr = bp->bio_data;
452131878Spjd	stripesize = sc->sc_stripesize;
453131878Spjd
454131878Spjd	cbp = g_clone_bio(bp);
455131878Spjd	if (cbp == NULL) {
456131878Spjd		error = ENOMEM;
457131878Spjd		goto failure;
458131878Spjd	}
459131878Spjd	TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
460129473Spjd	/*
461131878Spjd	 * Fill in the component buf structure.
462131878Spjd	 */
463131878Spjd	cbp->bio_done = g_std_done;
464131878Spjd	cbp->bio_offset = offset;
465131878Spjd	cbp->bio_data = addr;
466131878Spjd	cbp->bio_length = length;
467133204Spjd	cbp->bio_caller2 = sc->sc_disks[no];
468131878Spjd
469131878Spjd	/* offset -= offset % stripesize; */
470131878Spjd	offset -= offset & (stripesize - 1);
471131878Spjd	addr += length;
472131878Spjd	length = bp->bio_length - length;
473131878Spjd	for (no++; length > 0; no++, length -= stripesize, addr += stripesize) {
474131878Spjd		if (no > sc->sc_ndisks - 1) {
475131878Spjd			no = 0;
476131878Spjd			offset += stripesize;
477131878Spjd		}
478131878Spjd		cbp = g_clone_bio(bp);
479131878Spjd		if (cbp == NULL) {
480131878Spjd			error = ENOMEM;
481131878Spjd			goto failure;
482131878Spjd		}
483131878Spjd		TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
484131878Spjd
485131878Spjd		/*
486131878Spjd		 * Fill in the component buf structure.
487131878Spjd		 */
488131878Spjd		cbp->bio_done = g_std_done;
489131878Spjd		cbp->bio_offset = offset;
490131878Spjd		cbp->bio_data = addr;
491131878Spjd		/*
492131878Spjd		 * MIN() is in case when
493131878Spjd		 * (bp->bio_length % sc->sc_stripesize) != 0.
494131878Spjd		 */
495131878Spjd		cbp->bio_length = MIN(stripesize, length);
496131878Spjd
497133204Spjd		cbp->bio_caller2 = sc->sc_disks[no];
498131878Spjd	}
499131878Spjd	/*
500131878Spjd	 * Fire off all allocated requests!
501131878Spjd	 */
502131878Spjd	while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
503131878Spjd		struct g_consumer *cp;
504131878Spjd
505131878Spjd		TAILQ_REMOVE(&queue, cbp, bio_queue);
506133204Spjd		cp = cbp->bio_caller2;
507133204Spjd		cbp->bio_caller2 = NULL;
508131878Spjd		cbp->bio_to = cp->provider;
509131878Spjd		G_STRIPE_LOGREQ(cbp, "Sending request.");
510131878Spjd		g_io_request(cbp, cp);
511131878Spjd	}
512131878Spjd	return (0);
513131878Spjdfailure:
514131878Spjd	while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
515131878Spjd		TAILQ_REMOVE(&queue, cbp, bio_queue);
516133201Spjd		bp->bio_children--;
517131878Spjd		g_destroy_bio(cbp);
518131878Spjd	}
519131878Spjd	return (error);
520131878Spjd}
521131878Spjd
522131878Spjdstatic void
523163836Spjdg_stripe_flush(struct g_stripe_softc *sc, struct bio *bp)
524163836Spjd{
525163836Spjd	struct bio_queue_head queue;
526163836Spjd	struct g_consumer *cp;
527163836Spjd	struct bio *cbp;
528163836Spjd	u_int no;
529163836Spjd
530163836Spjd	bioq_init(&queue);
531163836Spjd	for (no = 0; no < sc->sc_ndisks; no++) {
532163836Spjd		cbp = g_clone_bio(bp);
533163836Spjd		if (cbp == NULL) {
534163836Spjd			for (cbp = bioq_first(&queue); cbp != NULL;
535163836Spjd			    cbp = bioq_first(&queue)) {
536163836Spjd				bioq_remove(&queue, cbp);
537163836Spjd				g_destroy_bio(cbp);
538163836Spjd			}
539163836Spjd			if (bp->bio_error == 0)
540163836Spjd				bp->bio_error = ENOMEM;
541163836Spjd			g_io_deliver(bp, bp->bio_error);
542163836Spjd			return;
543163836Spjd		}
544163836Spjd		bioq_insert_tail(&queue, cbp);
545163836Spjd		cbp->bio_done = g_std_done;
546163836Spjd		cbp->bio_caller1 = sc->sc_disks[no];
547163836Spjd		cbp->bio_to = sc->sc_disks[no]->provider;
548163836Spjd	}
549163836Spjd	for (cbp = bioq_first(&queue); cbp != NULL; cbp = bioq_first(&queue)) {
550163836Spjd		bioq_remove(&queue, cbp);
551163836Spjd		G_STRIPE_LOGREQ(cbp, "Sending request.");
552163836Spjd		cp = cbp->bio_caller1;
553163836Spjd		cbp->bio_caller1 = NULL;
554163836Spjd		g_io_request(cbp, cp);
555163836Spjd	}
556163836Spjd}
557163836Spjd
558163836Spjdstatic void
559131878Spjdg_stripe_start(struct bio *bp)
560131878Spjd{
561131878Spjd	off_t offset, start, length, nstripe;
562131878Spjd	struct g_stripe_softc *sc;
563131878Spjd	u_int no, stripesize;
564131878Spjd	int error, fast = 0;
565131878Spjd
566131878Spjd	sc = bp->bio_to->geom->softc;
567131878Spjd	/*
568129473Spjd	 * If sc == NULL, provider's error should be set and g_stripe_start()
569129473Spjd	 * should not be called at all.
570129473Spjd	 */
571129473Spjd	KASSERT(sc != NULL,
572129473Spjd	    ("Provider's error should be set (error=%d)(device=%s).",
573129473Spjd	    bp->bio_to->error, bp->bio_to->name));
574129473Spjd
575129473Spjd	G_STRIPE_LOGREQ(bp, "Request received.");
576129473Spjd
577129473Spjd	switch (bp->bio_cmd) {
578129473Spjd	case BIO_READ:
579129473Spjd	case BIO_WRITE:
580129473Spjd	case BIO_DELETE:
581129473Spjd		break;
582163886Spjd	case BIO_FLUSH:
583163886Spjd		g_stripe_flush(sc, bp);
584163886Spjd		return;
585129473Spjd	case BIO_GETATTR:
586129473Spjd		/* To which provider it should be delivered? */
587129473Spjd	default:
588129473Spjd		g_io_deliver(bp, EOPNOTSUPP);
589129473Spjd		return;
590129473Spjd	}
591129473Spjd
592129473Spjd	stripesize = sc->sc_stripesize;
593129473Spjd
594129473Spjd	/*
595131878Spjd	 * Calculations are quite messy, but fast I hope.
596129473Spjd	 */
597129473Spjd
598129473Spjd	/* Stripe number. */
599129473Spjd	/* nstripe = bp->bio_offset / stripesize; */
600129473Spjd	nstripe = bp->bio_offset >> (off_t)sc->sc_stripebits;
601129473Spjd	/* Disk number. */
602129473Spjd	no = nstripe % sc->sc_ndisks;
603129473Spjd	/* Start position in stripe. */
604129473Spjd	/* start = bp->bio_offset % stripesize; */
605129473Spjd	start = bp->bio_offset & (stripesize - 1);
606129473Spjd	/* Start position in disk. */
607131878Spjd	/* offset = (nstripe / sc->sc_ndisks) * stripesize + start; */
608131878Spjd	offset = ((nstripe / sc->sc_ndisks) << sc->sc_stripebits) + start;
609129473Spjd	/* Length of data to operate. */
610129473Spjd	length = MIN(bp->bio_length, stripesize - start);
611129473Spjd
612131878Spjd	/*
613131878Spjd	 * Do use "fast" mode when:
614131878Spjd	 * 1. "Fast" mode is ON.
615131878Spjd	 * and
616131878Spjd	 * 2. Request size is less than or equal to MAX_IO_SIZE (128kB),
617131878Spjd	 *    which should always be true.
618131878Spjd	 * and
619131878Spjd	 * 3. Request size is bigger than stripesize * ndisks. If it isn't,
620131878Spjd	 *    there will be no need to send more than one I/O request to
621131878Spjd	 *    a provider, so there is nothing to optmize.
622131878Spjd	 */
623131878Spjd	if (g_stripe_fast && bp->bio_length <= MAX_IO_SIZE &&
624131878Spjd	    bp->bio_length >= stripesize * sc->sc_ndisks) {
625131878Spjd		fast = 1;
626129473Spjd	}
627131878Spjd	error = 0;
628133205Spjd	if (fast) {
629131878Spjd		error = g_stripe_start_fast(bp, no, offset, length);
630133205Spjd		if (error != 0)
631133205Spjd			g_stripe_fast_failed++;
632133205Spjd	}
633129473Spjd	/*
634131878Spjd	 * Do use "economic" when:
635131878Spjd	 * 1. "Economic" mode is ON.
636131878Spjd	 * or
637131878Spjd	 * 2. "Fast" mode failed. It can only failed if there is no memory.
638129473Spjd	 */
639131878Spjd	if (!fast || error != 0)
640131878Spjd		error = g_stripe_start_economic(bp, no, offset, length);
641131878Spjd	if (error != 0) {
642131878Spjd		if (bp->bio_error == 0)
643131878Spjd			bp->bio_error = error;
644131878Spjd		g_io_deliver(bp, bp->bio_error);
645129473Spjd	}
646129473Spjd}
647129473Spjd
648129473Spjdstatic void
649129473Spjdg_stripe_check_and_run(struct g_stripe_softc *sc)
650129473Spjd{
651129473Spjd	off_t mediasize, ms;
652129473Spjd	u_int no, sectorsize = 0;
653129473Spjd
654129473Spjd	if (g_stripe_nvalid(sc) != sc->sc_ndisks)
655129473Spjd		return;
656129473Spjd
657132664Spjd	sc->sc_provider = g_new_providerf(sc->sc_geom, "stripe/%s",
658132664Spjd	    sc->sc_name);
659129473Spjd	/*
660129473Spjd	 * Find the smallest disk.
661129473Spjd	 */
662129473Spjd	mediasize = sc->sc_disks[0]->provider->mediasize;
663129473Spjd	if (sc->sc_type == G_STRIPE_TYPE_AUTOMATIC)
664129473Spjd		mediasize -= sc->sc_disks[0]->provider->sectorsize;
665129473Spjd	mediasize -= mediasize % sc->sc_stripesize;
666129473Spjd	sectorsize = sc->sc_disks[0]->provider->sectorsize;
667129473Spjd	for (no = 1; no < sc->sc_ndisks; no++) {
668129473Spjd		ms = sc->sc_disks[no]->provider->mediasize;
669129473Spjd		if (sc->sc_type == G_STRIPE_TYPE_AUTOMATIC)
670129473Spjd			ms -= sc->sc_disks[no]->provider->sectorsize;
671129473Spjd		ms -= ms % sc->sc_stripesize;
672129473Spjd		if (ms < mediasize)
673129473Spjd			mediasize = ms;
674129473Spjd		sectorsize = lcm(sectorsize,
675129473Spjd		    sc->sc_disks[no]->provider->sectorsize);
676129473Spjd	}
677129473Spjd	sc->sc_provider->sectorsize = sectorsize;
678129473Spjd	sc->sc_provider->mediasize = mediasize * sc->sc_ndisks;
679129473Spjd	g_error_provider(sc->sc_provider, 0);
680129473Spjd
681132664Spjd	G_STRIPE_DEBUG(0, "Device %s activated.", sc->sc_name);
682129473Spjd}
683129473Spjd
684129473Spjdstatic int
685129473Spjdg_stripe_read_metadata(struct g_consumer *cp, struct g_stripe_metadata *md)
686129473Spjd{
687129473Spjd	struct g_provider *pp;
688129473Spjd	u_char *buf;
689129473Spjd	int error;
690129473Spjd
691129473Spjd	g_topology_assert();
692129473Spjd
693129473Spjd	error = g_access(cp, 1, 0, 0);
694129473Spjd	if (error != 0)
695129473Spjd		return (error);
696129473Spjd	pp = cp->provider;
697129473Spjd	g_topology_unlock();
698129473Spjd	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
699129473Spjd	    &error);
700129473Spjd	g_topology_lock();
701129473Spjd	g_access(cp, -1, 0, 0);
702129473Spjd	if (buf == NULL)
703129473Spjd		return (error);
704129473Spjd
705129473Spjd	/* Decode metadata. */
706129473Spjd	stripe_metadata_decode(buf, md);
707129473Spjd	g_free(buf);
708129473Spjd
709129473Spjd	return (0);
710129473Spjd}
711129473Spjd
712129473Spjd/*
713129473Spjd * Add disk to given device.
714129473Spjd */
715129473Spjdstatic int
716129473Spjdg_stripe_add_disk(struct g_stripe_softc *sc, struct g_provider *pp, u_int no)
717129473Spjd{
718129473Spjd	struct g_consumer *cp, *fcp;
719129473Spjd	struct g_geom *gp;
720129473Spjd	int error;
721129473Spjd
722129473Spjd	/* Metadata corrupted? */
723129473Spjd	if (no >= sc->sc_ndisks)
724129473Spjd		return (EINVAL);
725129473Spjd
726129473Spjd	/* Check if disk is not already attached. */
727129473Spjd	if (sc->sc_disks[no] != NULL)
728129473Spjd		return (EEXIST);
729129473Spjd
730129473Spjd	gp = sc->sc_geom;
731129473Spjd	fcp = LIST_FIRST(&gp->consumer);
732129473Spjd
733129473Spjd	cp = g_new_consumer(gp);
734129473Spjd	error = g_attach(cp, pp);
735129473Spjd	if (error != 0) {
736129473Spjd		g_destroy_consumer(cp);
737129473Spjd		return (error);
738129473Spjd	}
739129473Spjd
740129473Spjd	if (fcp != NULL && (fcp->acr > 0 || fcp->acw > 0 || fcp->ace > 0)) {
741129473Spjd		error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
742129473Spjd		if (error != 0) {
743129473Spjd			g_detach(cp);
744129473Spjd			g_destroy_consumer(cp);
745129473Spjd			return (error);
746129473Spjd		}
747129473Spjd	}
748129473Spjd	if (sc->sc_type == G_STRIPE_TYPE_AUTOMATIC) {
749129473Spjd		struct g_stripe_metadata md;
750129473Spjd
751129473Spjd		/* Reread metadata. */
752129473Spjd		error = g_stripe_read_metadata(cp, &md);
753129473Spjd		if (error != 0)
754129473Spjd			goto fail;
755129473Spjd
756129473Spjd		if (strcmp(md.md_magic, G_STRIPE_MAGIC) != 0 ||
757129473Spjd		    strcmp(md.md_name, sc->sc_name) != 0 ||
758129473Spjd		    md.md_id != sc->sc_id) {
759129473Spjd			G_STRIPE_DEBUG(0, "Metadata on %s changed.", pp->name);
760129473Spjd			goto fail;
761129473Spjd		}
762129473Spjd	}
763129473Spjd
764129473Spjd	cp->private = sc;
765129473Spjd	cp->index = no;
766129473Spjd	sc->sc_disks[no] = cp;
767129473Spjd
768132664Spjd	G_STRIPE_DEBUG(0, "Disk %s attached to %s.", pp->name, sc->sc_name);
769129473Spjd
770129473Spjd	g_stripe_check_and_run(sc);
771129473Spjd
772129473Spjd	return (0);
773129473Spjdfail:
774129473Spjd	if (fcp != NULL && (fcp->acr > 0 || fcp->acw > 0 || fcp->ace > 0))
775129473Spjd		g_access(cp, -fcp->acr, -fcp->acw, -fcp->ace);
776129473Spjd	g_detach(cp);
777129473Spjd	g_destroy_consumer(cp);
778129473Spjd	return (error);
779129473Spjd}
780129473Spjd
781129473Spjdstatic struct g_geom *
782129473Spjdg_stripe_create(struct g_class *mp, const struct g_stripe_metadata *md,
783129473Spjd    u_int type)
784129473Spjd{
785129473Spjd	struct g_stripe_softc *sc;
786129473Spjd	struct g_geom *gp;
787129473Spjd	u_int no;
788129473Spjd
789132664Spjd	G_STRIPE_DEBUG(1, "Creating device %s (id=%u).", md->md_name,
790129473Spjd	    md->md_id);
791129473Spjd
792129473Spjd	/* Two disks is minimum. */
793132664Spjd	if (md->md_all < 2) {
794132664Spjd		G_STRIPE_DEBUG(0, "Too few disks defined for %s.", md->md_name);
795129473Spjd		return (NULL);
796129473Spjd	}
797129473Spjd#if 0
798129473Spjd	/* Stripe size have to be grater than or equal to sector size. */
799129473Spjd	if (md->md_stripesize < sectorsize) {
800132664Spjd		G_STRIPE_DEBUG(0, "Invalid stripe size for %s.", md->md_name);
801129473Spjd		return (NULL);
802129473Spjd	}
803129473Spjd#endif
804129473Spjd	/* Stripe size have to be power of 2. */
805129473Spjd	if (!powerof2(md->md_stripesize)) {
806132664Spjd		G_STRIPE_DEBUG(0, "Invalid stripe size for %s.", md->md_name);
807129473Spjd		return (NULL);
808129473Spjd	}
809129473Spjd
810129473Spjd	/* Check for duplicate unit */
811129473Spjd	LIST_FOREACH(gp, &mp->geom, geom) {
812129473Spjd		sc = gp->softc;
813129473Spjd		if (sc != NULL && strcmp(sc->sc_name, md->md_name) == 0) {
814129473Spjd			G_STRIPE_DEBUG(0, "Device %s already configured.",
815132664Spjd			    sc->sc_name);
816129473Spjd			return (NULL);
817129473Spjd		}
818129473Spjd	}
819132664Spjd	gp = g_new_geomf(mp, "%s", md->md_name);
820129473Spjd	gp->softc = NULL;	/* for a moment */
821129473Spjd
822132662Spjd	sc = malloc(sizeof(*sc), M_STRIPE, M_WAITOK | M_ZERO);
823129473Spjd	gp->start = g_stripe_start;
824129473Spjd	gp->spoiled = g_stripe_orphan;
825129473Spjd	gp->orphan = g_stripe_orphan;
826129473Spjd	gp->access = g_stripe_access;
827129473Spjd	gp->dumpconf = g_stripe_dumpconf;
828129473Spjd
829129473Spjd	sc->sc_id = md->md_id;
830129473Spjd	sc->sc_stripesize = md->md_stripesize;
831149300Spjd	sc->sc_stripebits = bitcount32(sc->sc_stripesize - 1);
832129473Spjd	sc->sc_ndisks = md->md_all;
833129473Spjd	sc->sc_disks = malloc(sizeof(struct g_consumer *) * sc->sc_ndisks,
834129473Spjd	    M_STRIPE, M_WAITOK | M_ZERO);
835129473Spjd	for (no = 0; no < sc->sc_ndisks; no++)
836129473Spjd		sc->sc_disks[no] = NULL;
837129473Spjd	sc->sc_type = type;
838129473Spjd
839129473Spjd	gp->softc = sc;
840129473Spjd	sc->sc_geom = gp;
841129473Spjd	sc->sc_provider = NULL;
842129473Spjd
843132664Spjd	G_STRIPE_DEBUG(0, "Device %s created (id=%u).", sc->sc_name, sc->sc_id);
844129473Spjd
845129473Spjd	return (gp);
846129473Spjd}
847129473Spjd
848129473Spjdstatic int
849129473Spjdg_stripe_destroy(struct g_stripe_softc *sc, boolean_t force)
850129473Spjd{
851129473Spjd	struct g_provider *pp;
852129473Spjd	struct g_geom *gp;
853129473Spjd	u_int no;
854129473Spjd
855129473Spjd	g_topology_assert();
856129473Spjd
857129473Spjd	if (sc == NULL)
858129473Spjd		return (ENXIO);
859129473Spjd
860129473Spjd	pp = sc->sc_provider;
861129473Spjd	if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
862129473Spjd		if (force) {
863129473Spjd			G_STRIPE_DEBUG(0, "Device %s is still open, so it "
864129473Spjd			    "can't be definitely removed.", pp->name);
865129473Spjd		} else {
866129473Spjd			G_STRIPE_DEBUG(1,
867129473Spjd			    "Device %s is still open (r%dw%de%d).", pp->name,
868129473Spjd			    pp->acr, pp->acw, pp->ace);
869129473Spjd			return (EBUSY);
870129473Spjd		}
871129473Spjd	}
872129473Spjd
873129473Spjd	for (no = 0; no < sc->sc_ndisks; no++) {
874129473Spjd		if (sc->sc_disks[no] != NULL)
875129473Spjd			g_stripe_remove_disk(sc->sc_disks[no]);
876129473Spjd	}
877129473Spjd
878129473Spjd	gp = sc->sc_geom;
879129473Spjd	gp->softc = NULL;
880129473Spjd	KASSERT(sc->sc_provider == NULL, ("Provider still exists? (device=%s)",
881129473Spjd	    gp->name));
882129473Spjd	free(sc->sc_disks, M_STRIPE);
883129473Spjd	free(sc, M_STRIPE);
884129473Spjd
885155174Spjd	pp = LIST_FIRST(&gp->provider);
886129473Spjd	if (pp == NULL || (pp->acr == 0 && pp->acw == 0 && pp->ace == 0))
887129473Spjd		G_STRIPE_DEBUG(0, "Device %s destroyed.", gp->name);
888129473Spjd
889129473Spjd	g_wither_geom(gp, ENXIO);
890129473Spjd
891129473Spjd	return (0);
892129473Spjd}
893129473Spjd
894129473Spjdstatic int
895129473Spjdg_stripe_destroy_geom(struct gctl_req *req __unused,
896129473Spjd    struct g_class *mp __unused, struct g_geom *gp)
897129473Spjd{
898129473Spjd	struct g_stripe_softc *sc;
899129473Spjd
900129473Spjd	sc = gp->softc;
901129473Spjd	return (g_stripe_destroy(sc, 0));
902129473Spjd}
903129473Spjd
904129473Spjdstatic struct g_geom *
905129473Spjdg_stripe_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
906129473Spjd{
907129473Spjd	struct g_stripe_metadata md;
908129473Spjd	struct g_stripe_softc *sc;
909129473Spjd	struct g_consumer *cp;
910129473Spjd	struct g_geom *gp;
911129473Spjd	int error;
912129473Spjd
913129473Spjd	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
914129473Spjd	g_topology_assert();
915129473Spjd
916129473Spjd	G_STRIPE_DEBUG(3, "Tasting %s.", pp->name);
917129473Spjd
918129473Spjd	gp = g_new_geomf(mp, "stripe:taste");
919129473Spjd	gp->start = g_stripe_start;
920129473Spjd	gp->access = g_stripe_access;
921129473Spjd	gp->orphan = g_stripe_orphan;
922129473Spjd	cp = g_new_consumer(gp);
923129473Spjd	g_attach(cp, pp);
924129473Spjd	error = g_stripe_read_metadata(cp, &md);
925133371Spjd	g_detach(cp);
926133371Spjd	g_destroy_consumer(cp);
927133371Spjd	g_destroy_geom(gp);
928129473Spjd	if (error != 0)
929129473Spjd		return (NULL);
930129473Spjd	gp = NULL;
931129473Spjd
932129473Spjd	if (strcmp(md.md_magic, G_STRIPE_MAGIC) != 0)
933129473Spjd		return (NULL);
934129473Spjd	if (md.md_version > G_STRIPE_VERSION) {
935129473Spjd		printf("geom_stripe.ko module is too old to handle %s.\n",
936129473Spjd		    pp->name);
937129473Spjd		return (NULL);
938129473Spjd	}
939133373Spjd	/*
940133373Spjd	 * Backward compatibility:
941133373Spjd	 */
942142727Spjd	/* There was no md_provider field in earlier versions of metadata. */
943133373Spjd	if (md.md_version < 2)
944133373Spjd		bzero(md.md_provider, sizeof(md.md_provider));
945142727Spjd	/* There was no md_provsize field in earlier versions of metadata. */
946142727Spjd	if (md.md_version < 3)
947142727Spjd		md.md_provsize = pp->mediasize;
948129473Spjd
949133373Spjd	if (md.md_provider[0] != '\0' && strcmp(md.md_provider, pp->name) != 0)
950133373Spjd		return (NULL);
951142727Spjd	if (md.md_provsize != pp->mediasize)
952142727Spjd		return (NULL);
953133373Spjd
954129473Spjd	/*
955129473Spjd	 * Let's check if device already exists.
956129473Spjd	 */
957129473Spjd	sc = NULL;
958129473Spjd	LIST_FOREACH(gp, &mp->geom, geom) {
959129473Spjd		sc = gp->softc;
960129473Spjd		if (sc == NULL)
961129473Spjd			continue;
962129473Spjd		if (sc->sc_type != G_STRIPE_TYPE_AUTOMATIC)
963129473Spjd			continue;
964129473Spjd		if (strcmp(md.md_name, sc->sc_name) != 0)
965129473Spjd			continue;
966129473Spjd		if (md.md_id != sc->sc_id)
967129473Spjd			continue;
968129473Spjd		break;
969129473Spjd	}
970129473Spjd	if (gp != NULL) {
971129473Spjd		G_STRIPE_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
972129473Spjd		error = g_stripe_add_disk(sc, pp, md.md_no);
973129473Spjd		if (error != 0) {
974129473Spjd			G_STRIPE_DEBUG(0,
975129473Spjd			    "Cannot add disk %s to %s (error=%d).", pp->name,
976129473Spjd			    gp->name, error);
977129473Spjd			return (NULL);
978129473Spjd		}
979129473Spjd	} else {
980129473Spjd		gp = g_stripe_create(mp, &md, G_STRIPE_TYPE_AUTOMATIC);
981129473Spjd		if (gp == NULL) {
982132664Spjd			G_STRIPE_DEBUG(0, "Cannot create device %s.",
983129473Spjd			    md.md_name);
984129473Spjd			return (NULL);
985129473Spjd		}
986129473Spjd		sc = gp->softc;
987129473Spjd		G_STRIPE_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
988129473Spjd		error = g_stripe_add_disk(sc, pp, md.md_no);
989129473Spjd		if (error != 0) {
990129473Spjd			G_STRIPE_DEBUG(0,
991129473Spjd			    "Cannot add disk %s to %s (error=%d).", pp->name,
992129473Spjd			    gp->name, error);
993129473Spjd			g_stripe_destroy(sc, 1);
994129473Spjd			return (NULL);
995129473Spjd		}
996129473Spjd	}
997129473Spjd
998129473Spjd	return (gp);
999129473Spjd}
1000129473Spjd
1001129473Spjdstatic void
1002129473Spjdg_stripe_ctl_create(struct gctl_req *req, struct g_class *mp)
1003129473Spjd{
1004129473Spjd	u_int attached, no;
1005129473Spjd	struct g_stripe_metadata md;
1006129473Spjd	struct g_provider *pp;
1007129473Spjd	struct g_stripe_softc *sc;
1008129473Spjd	struct g_geom *gp;
1009129473Spjd	struct sbuf *sb;
1010129473Spjd	intmax_t *stripesize;
1011129473Spjd	const char *name;
1012129473Spjd	char param[16];
1013129473Spjd	int *nargs;
1014129473Spjd
1015129473Spjd	g_topology_assert();
1016129473Spjd	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
1017129473Spjd	if (nargs == NULL) {
1018129473Spjd		gctl_error(req, "No '%s' argument.", "nargs");
1019129473Spjd		return;
1020129473Spjd	}
1021129473Spjd	if (*nargs <= 2) {
1022129473Spjd		gctl_error(req, "Too few arguments.");
1023129473Spjd		return;
1024129473Spjd	}
1025129473Spjd
1026129473Spjd	strlcpy(md.md_magic, G_STRIPE_MAGIC, sizeof(md.md_magic));
1027129473Spjd	md.md_version = G_STRIPE_VERSION;
1028129473Spjd	name = gctl_get_asciiparam(req, "arg0");
1029129473Spjd	if (name == NULL) {
1030129473Spjd		gctl_error(req, "No 'arg%u' argument.", 0);
1031129473Spjd		return;
1032129473Spjd	}
1033129473Spjd	strlcpy(md.md_name, name, sizeof(md.md_name));
1034129473Spjd	md.md_id = arc4random();
1035129473Spjd	md.md_no = 0;
1036129473Spjd	md.md_all = *nargs - 1;
1037129473Spjd	stripesize = gctl_get_paraml(req, "stripesize", sizeof(*stripesize));
1038129473Spjd	if (stripesize == NULL) {
1039129473Spjd		gctl_error(req, "No '%s' argument.", "stripesize");
1040129473Spjd		return;
1041129473Spjd	}
1042129473Spjd	md.md_stripesize = *stripesize;
1043133373Spjd	bzero(md.md_provider, sizeof(md.md_provider));
1044142727Spjd	/* This field is not important here. */
1045142727Spjd	md.md_provsize = 0;
1046129473Spjd
1047129473Spjd	/* Check all providers are valid */
1048129473Spjd	for (no = 1; no < *nargs; no++) {
1049129473Spjd		snprintf(param, sizeof(param), "arg%u", no);
1050129473Spjd		name = gctl_get_asciiparam(req, param);
1051129473Spjd		if (name == NULL) {
1052129473Spjd			gctl_error(req, "No 'arg%u' argument.", no);
1053129473Spjd			return;
1054129473Spjd		}
1055129473Spjd		if (strncmp(name, "/dev/", strlen("/dev/")) == 0)
1056129473Spjd			name += strlen("/dev/");
1057129473Spjd		pp = g_provider_by_name(name);
1058129473Spjd		if (pp == NULL) {
1059129473Spjd			G_STRIPE_DEBUG(1, "Disk %s is invalid.", name);
1060129473Spjd			gctl_error(req, "Disk %s is invalid.", name);
1061129473Spjd			return;
1062129473Spjd		}
1063129473Spjd	}
1064129473Spjd
1065129473Spjd	gp = g_stripe_create(mp, &md, G_STRIPE_TYPE_MANUAL);
1066129473Spjd	if (gp == NULL) {
1067132664Spjd		gctl_error(req, "Can't configure %s.", md.md_name);
1068129473Spjd		return;
1069129473Spjd	}
1070129473Spjd
1071129473Spjd	sc = gp->softc;
1072129473Spjd	sb = sbuf_new(NULL, NULL, 0, SBUF_AUTOEXTEND);
1073129473Spjd	sbuf_printf(sb, "Can't attach disk(s) to %s:", gp->name);
1074129473Spjd	for (attached = 0, no = 1; no < *nargs; no++) {
1075129473Spjd		snprintf(param, sizeof(param), "arg%u", no);
1076129473Spjd		name = gctl_get_asciiparam(req, param);
1077146109Spjd		if (name == NULL) {
1078146109Spjd			gctl_error(req, "No 'arg%u' argument.", no);
1079146109Spjd			continue;
1080146109Spjd		}
1081129473Spjd		if (strncmp(name, "/dev/", strlen("/dev/")) == 0)
1082129473Spjd			name += strlen("/dev/");
1083129473Spjd		pp = g_provider_by_name(name);
1084129473Spjd		KASSERT(pp != NULL, ("Provider %s disappear?!", name));
1085129473Spjd		if (g_stripe_add_disk(sc, pp, no - 1) != 0) {
1086129473Spjd			G_STRIPE_DEBUG(1, "Disk %u (%s) not attached to %s.",
1087129473Spjd			    no, pp->name, gp->name);
1088129473Spjd			sbuf_printf(sb, " %s", pp->name);
1089129473Spjd			continue;
1090129473Spjd		}
1091129473Spjd		attached++;
1092129473Spjd	}
1093129473Spjd	sbuf_finish(sb);
1094129473Spjd	if (md.md_all != attached) {
1095129473Spjd		g_stripe_destroy(gp->softc, 1);
1096129473Spjd		gctl_error(req, "%s", sbuf_data(sb));
1097129473Spjd	}
1098129473Spjd	sbuf_delete(sb);
1099129473Spjd}
1100129473Spjd
1101129473Spjdstatic struct g_stripe_softc *
1102129473Spjdg_stripe_find_device(struct g_class *mp, const char *name)
1103129473Spjd{
1104129473Spjd	struct g_stripe_softc *sc;
1105129473Spjd	struct g_geom *gp;
1106129473Spjd
1107129473Spjd	LIST_FOREACH(gp, &mp->geom, geom) {
1108129473Spjd		sc = gp->softc;
1109129473Spjd		if (sc == NULL)
1110129473Spjd			continue;
1111132664Spjd		if (strcmp(sc->sc_name, name) == 0)
1112129473Spjd			return (sc);
1113129473Spjd	}
1114129473Spjd	return (NULL);
1115129473Spjd}
1116129473Spjd
1117129473Spjdstatic void
1118129473Spjdg_stripe_ctl_destroy(struct gctl_req *req, struct g_class *mp)
1119129473Spjd{
1120129473Spjd	struct g_stripe_softc *sc;
1121129473Spjd	int *force, *nargs, error;
1122129473Spjd	const char *name;
1123129473Spjd	char param[16];
1124129473Spjd	u_int i;
1125129473Spjd
1126129473Spjd	g_topology_assert();
1127129473Spjd
1128129473Spjd	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
1129129473Spjd	if (nargs == NULL) {
1130129473Spjd		gctl_error(req, "No '%s' argument.", "nargs");
1131129473Spjd		return;
1132129473Spjd	}
1133129473Spjd	if (*nargs <= 0) {
1134129473Spjd		gctl_error(req, "Missing device(s).");
1135129473Spjd		return;
1136129473Spjd	}
1137129473Spjd	force = gctl_get_paraml(req, "force", sizeof(*force));
1138129473Spjd	if (force == NULL) {
1139129473Spjd		gctl_error(req, "No '%s' argument.", "force");
1140129473Spjd		return;
1141129473Spjd	}
1142129473Spjd
1143129473Spjd	for (i = 0; i < (u_int)*nargs; i++) {
1144129473Spjd		snprintf(param, sizeof(param), "arg%u", i);
1145129473Spjd		name = gctl_get_asciiparam(req, param);
1146129473Spjd		if (name == NULL) {
1147129473Spjd			gctl_error(req, "No 'arg%u' argument.", i);
1148129473Spjd			return;
1149129473Spjd		}
1150129473Spjd		sc = g_stripe_find_device(mp, name);
1151129473Spjd		if (sc == NULL) {
1152129473Spjd			gctl_error(req, "No such device: %s.", name);
1153129473Spjd			return;
1154129473Spjd		}
1155129473Spjd		error = g_stripe_destroy(sc, *force);
1156129473Spjd		if (error != 0) {
1157129473Spjd			gctl_error(req, "Cannot destroy device %s (error=%d).",
1158132664Spjd			    sc->sc_name, error);
1159129473Spjd			return;
1160129473Spjd		}
1161129473Spjd	}
1162129473Spjd}
1163129473Spjd
1164129473Spjdstatic void
1165129473Spjdg_stripe_config(struct gctl_req *req, struct g_class *mp, const char *verb)
1166129473Spjd{
1167129473Spjd	uint32_t *version;
1168129473Spjd
1169129473Spjd	g_topology_assert();
1170129473Spjd
1171129473Spjd	version = gctl_get_paraml(req, "version", sizeof(*version));
1172129473Spjd	if (version == NULL) {
1173129473Spjd		gctl_error(req, "No '%s' argument.", "version");
1174129473Spjd		return;
1175129473Spjd	}
1176129473Spjd	if (*version != G_STRIPE_VERSION) {
1177129473Spjd		gctl_error(req, "Userland and kernel parts are out of sync.");
1178129473Spjd		return;
1179129473Spjd	}
1180129473Spjd
1181129473Spjd	if (strcmp(verb, "create") == 0) {
1182129473Spjd		g_stripe_ctl_create(req, mp);
1183129473Spjd		return;
1184131649Spjd	} else if (strcmp(verb, "destroy") == 0 ||
1185131649Spjd	    strcmp(verb, "stop") == 0) {
1186129473Spjd		g_stripe_ctl_destroy(req, mp);
1187129473Spjd		return;
1188129473Spjd	}
1189129473Spjd
1190129473Spjd	gctl_error(req, "Unknown verb.");
1191129473Spjd}
1192129473Spjd
1193129473Spjdstatic void
1194129473Spjdg_stripe_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1195129473Spjd    struct g_consumer *cp, struct g_provider *pp)
1196129473Spjd{
1197129473Spjd	struct g_stripe_softc *sc;
1198129473Spjd
1199129473Spjd	sc = gp->softc;
1200132665Spjd	if (sc == NULL)
1201129473Spjd		return;
1202132665Spjd	if (pp != NULL) {
1203132665Spjd		/* Nothing here. */
1204132665Spjd	} else if (cp != NULL) {
1205134292Spjd		sbuf_printf(sb, "%s<Number>%u</Number>\n", indent,
1206134292Spjd		    (u_int)cp->index);
1207132665Spjd	} else {
1208132665Spjd		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
1209132665Spjd		sbuf_printf(sb, "%s<Stripesize>%u</Stripesize>\n", indent,
1210132665Spjd		    (u_int)sc->sc_stripesize);
1211132665Spjd		sbuf_printf(sb, "%s<Type>", indent);
1212132665Spjd		switch (sc->sc_type) {
1213132665Spjd		case G_STRIPE_TYPE_AUTOMATIC:
1214132665Spjd			sbuf_printf(sb, "AUTOMATIC");
1215132665Spjd			break;
1216132665Spjd		case G_STRIPE_TYPE_MANUAL:
1217132665Spjd			sbuf_printf(sb, "MANUAL");
1218132665Spjd			break;
1219132665Spjd		default:
1220132665Spjd			sbuf_printf(sb, "UNKNOWN");
1221132665Spjd			break;
1222132665Spjd		}
1223132665Spjd		sbuf_printf(sb, "</Type>\n");
1224132665Spjd		sbuf_printf(sb, "%s<Status>Total=%u, Online=%u</Status>\n",
1225132665Spjd		    indent, sc->sc_ndisks, g_stripe_nvalid(sc));
1226132665Spjd		sbuf_printf(sb, "%s<State>", indent);
1227132665Spjd		if (sc->sc_provider != NULL && sc->sc_provider->error == 0)
1228132665Spjd			sbuf_printf(sb, "UP");
1229132665Spjd		else
1230132665Spjd			sbuf_printf(sb, "DOWN");
1231132665Spjd		sbuf_printf(sb, "</State>\n");
1232129473Spjd	}
1233129473Spjd}
1234129473Spjd
1235129473SpjdDECLARE_GEOM_CLASS(g_stripe_class, g_stripe);
1236