xref: /freebsd/sys/geom/raid3/g_raid3.c (revision 3c5ba95ad12285ad37c182a4bfc1b240ec6d18a7)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2004-2006 Pawel Jakub Dawidek <pjd@FreeBSD.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/kernel.h>
35 #include <sys/module.h>
36 #include <sys/limits.h>
37 #include <sys/lock.h>
38 #include <sys/mutex.h>
39 #include <sys/bio.h>
40 #include <sys/sbuf.h>
41 #include <sys/sysctl.h>
42 #include <sys/malloc.h>
43 #include <sys/eventhandler.h>
44 #include <vm/uma.h>
45 #include <geom/geom.h>
46 #include <sys/proc.h>
47 #include <sys/kthread.h>
48 #include <sys/sched.h>
49 #include <geom/raid3/g_raid3.h>
50 
51 FEATURE(geom_raid3, "GEOM RAID-3 functionality");
52 
53 static MALLOC_DEFINE(M_RAID3, "raid3_data", "GEOM_RAID3 Data");
54 
55 SYSCTL_DECL(_kern_geom);
56 static SYSCTL_NODE(_kern_geom, OID_AUTO, raid3, CTLFLAG_RW, 0,
57     "GEOM_RAID3 stuff");
58 u_int g_raid3_debug = 0;
59 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, debug, CTLFLAG_RWTUN, &g_raid3_debug, 0,
60     "Debug level");
61 static u_int g_raid3_timeout = 4;
62 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, timeout, CTLFLAG_RWTUN, &g_raid3_timeout,
63     0, "Time to wait on all raid3 components");
64 static u_int g_raid3_idletime = 5;
65 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, idletime, CTLFLAG_RWTUN,
66     &g_raid3_idletime, 0, "Mark components as clean when idling");
67 static u_int g_raid3_disconnect_on_failure = 1;
68 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, disconnect_on_failure, CTLFLAG_RWTUN,
69     &g_raid3_disconnect_on_failure, 0, "Disconnect component on I/O failure.");
70 static u_int g_raid3_syncreqs = 2;
71 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, sync_requests, CTLFLAG_RDTUN,
72     &g_raid3_syncreqs, 0, "Parallel synchronization I/O requests.");
73 static u_int g_raid3_use_malloc = 0;
74 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, use_malloc, CTLFLAG_RDTUN,
75     &g_raid3_use_malloc, 0, "Use malloc(9) instead of uma(9).");
76 
77 static u_int g_raid3_n64k = 50;
78 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, n64k, CTLFLAG_RDTUN, &g_raid3_n64k, 0,
79     "Maximum number of 64kB allocations");
80 static u_int g_raid3_n16k = 200;
81 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, n16k, CTLFLAG_RDTUN, &g_raid3_n16k, 0,
82     "Maximum number of 16kB allocations");
83 static u_int g_raid3_n4k = 1200;
84 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, n4k, CTLFLAG_RDTUN, &g_raid3_n4k, 0,
85     "Maximum number of 4kB allocations");
86 
87 static SYSCTL_NODE(_kern_geom_raid3, OID_AUTO, stat, CTLFLAG_RW, 0,
88     "GEOM_RAID3 statistics");
89 static u_int g_raid3_parity_mismatch = 0;
90 SYSCTL_UINT(_kern_geom_raid3_stat, OID_AUTO, parity_mismatch, CTLFLAG_RD,
91     &g_raid3_parity_mismatch, 0, "Number of failures in VERIFY mode");
92 
93 #define	MSLEEP(ident, mtx, priority, wmesg, timeout)	do {		\
94 	G_RAID3_DEBUG(4, "%s: Sleeping %p.", __func__, (ident));	\
95 	msleep((ident), (mtx), (priority), (wmesg), (timeout));		\
96 	G_RAID3_DEBUG(4, "%s: Woken up %p.", __func__, (ident));	\
97 } while (0)
98 
99 static eventhandler_tag g_raid3_post_sync = NULL;
100 static int g_raid3_shutdown = 0;
101 
102 static int g_raid3_destroy_geom(struct gctl_req *req, struct g_class *mp,
103     struct g_geom *gp);
104 static g_taste_t g_raid3_taste;
105 static void g_raid3_init(struct g_class *mp);
106 static void g_raid3_fini(struct g_class *mp);
107 
108 struct g_class g_raid3_class = {
109 	.name = G_RAID3_CLASS_NAME,
110 	.version = G_VERSION,
111 	.ctlreq = g_raid3_config,
112 	.taste = g_raid3_taste,
113 	.destroy_geom = g_raid3_destroy_geom,
114 	.init = g_raid3_init,
115 	.fini = g_raid3_fini
116 };
117 
118 
119 static void g_raid3_destroy_provider(struct g_raid3_softc *sc);
120 static int g_raid3_update_disk(struct g_raid3_disk *disk, u_int state);
121 static void g_raid3_update_device(struct g_raid3_softc *sc, boolean_t force);
122 static void g_raid3_dumpconf(struct sbuf *sb, const char *indent,
123     struct g_geom *gp, struct g_consumer *cp, struct g_provider *pp);
124 static void g_raid3_sync_stop(struct g_raid3_softc *sc, int type);
125 static int g_raid3_register_request(struct bio *pbp);
126 static void g_raid3_sync_release(struct g_raid3_softc *sc);
127 
128 
129 static const char *
130 g_raid3_disk_state2str(int state)
131 {
132 
133 	switch (state) {
134 	case G_RAID3_DISK_STATE_NODISK:
135 		return ("NODISK");
136 	case G_RAID3_DISK_STATE_NONE:
137 		return ("NONE");
138 	case G_RAID3_DISK_STATE_NEW:
139 		return ("NEW");
140 	case G_RAID3_DISK_STATE_ACTIVE:
141 		return ("ACTIVE");
142 	case G_RAID3_DISK_STATE_STALE:
143 		return ("STALE");
144 	case G_RAID3_DISK_STATE_SYNCHRONIZING:
145 		return ("SYNCHRONIZING");
146 	case G_RAID3_DISK_STATE_DISCONNECTED:
147 		return ("DISCONNECTED");
148 	default:
149 		return ("INVALID");
150 	}
151 }
152 
153 static const char *
154 g_raid3_device_state2str(int state)
155 {
156 
157 	switch (state) {
158 	case G_RAID3_DEVICE_STATE_STARTING:
159 		return ("STARTING");
160 	case G_RAID3_DEVICE_STATE_DEGRADED:
161 		return ("DEGRADED");
162 	case G_RAID3_DEVICE_STATE_COMPLETE:
163 		return ("COMPLETE");
164 	default:
165 		return ("INVALID");
166 	}
167 }
168 
169 const char *
170 g_raid3_get_diskname(struct g_raid3_disk *disk)
171 {
172 
173 	if (disk->d_consumer == NULL || disk->d_consumer->provider == NULL)
174 		return ("[unknown]");
175 	return (disk->d_name);
176 }
177 
178 static void *
179 g_raid3_alloc(struct g_raid3_softc *sc, size_t size, int flags)
180 {
181 	void *ptr;
182 	enum g_raid3_zones zone;
183 
184 	if (g_raid3_use_malloc ||
185 	    (zone = g_raid3_zone(size)) == G_RAID3_NUM_ZONES)
186 		ptr = malloc(size, M_RAID3, flags);
187 	else {
188 		ptr = uma_zalloc_arg(sc->sc_zones[zone].sz_zone,
189 		   &sc->sc_zones[zone], flags);
190 		sc->sc_zones[zone].sz_requested++;
191 		if (ptr == NULL)
192 			sc->sc_zones[zone].sz_failed++;
193 	}
194 	return (ptr);
195 }
196 
197 static void
198 g_raid3_free(struct g_raid3_softc *sc, void *ptr, size_t size)
199 {
200 	enum g_raid3_zones zone;
201 
202 	if (g_raid3_use_malloc ||
203 	    (zone = g_raid3_zone(size)) == G_RAID3_NUM_ZONES)
204 		free(ptr, M_RAID3);
205 	else {
206 		uma_zfree_arg(sc->sc_zones[zone].sz_zone,
207 		    ptr, &sc->sc_zones[zone]);
208 	}
209 }
210 
211 static int
212 g_raid3_uma_ctor(void *mem, int size, void *arg, int flags)
213 {
214 	struct g_raid3_zone *sz = arg;
215 
216 	if (sz->sz_max > 0 && sz->sz_inuse == sz->sz_max)
217 		return (ENOMEM);
218 	sz->sz_inuse++;
219 	return (0);
220 }
221 
222 static void
223 g_raid3_uma_dtor(void *mem, int size, void *arg)
224 {
225 	struct g_raid3_zone *sz = arg;
226 
227 	sz->sz_inuse--;
228 }
229 
230 #define	g_raid3_xor(src, dst, size)					\
231 	_g_raid3_xor((uint64_t *)(src),					\
232 	    (uint64_t *)(dst), (size_t)size)
233 static void
234 _g_raid3_xor(uint64_t *src, uint64_t *dst, size_t size)
235 {
236 
237 	KASSERT((size % 128) == 0, ("Invalid size: %zu.", size));
238 	for (; size > 0; size -= 128) {
239 		*dst++ ^= (*src++);
240 		*dst++ ^= (*src++);
241 		*dst++ ^= (*src++);
242 		*dst++ ^= (*src++);
243 		*dst++ ^= (*src++);
244 		*dst++ ^= (*src++);
245 		*dst++ ^= (*src++);
246 		*dst++ ^= (*src++);
247 		*dst++ ^= (*src++);
248 		*dst++ ^= (*src++);
249 		*dst++ ^= (*src++);
250 		*dst++ ^= (*src++);
251 		*dst++ ^= (*src++);
252 		*dst++ ^= (*src++);
253 		*dst++ ^= (*src++);
254 		*dst++ ^= (*src++);
255 	}
256 }
257 
258 static int
259 g_raid3_is_zero(struct bio *bp)
260 {
261 	static const uint64_t zeros[] = {
262 	    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
263 	};
264 	u_char *addr;
265 	ssize_t size;
266 
267 	size = bp->bio_length;
268 	addr = (u_char *)bp->bio_data;
269 	for (; size > 0; size -= sizeof(zeros), addr += sizeof(zeros)) {
270 		if (bcmp(addr, zeros, sizeof(zeros)) != 0)
271 			return (0);
272 	}
273 	return (1);
274 }
275 
276 /*
277  * --- Events handling functions ---
278  * Events in geom_raid3 are used to maintain disks and device status
279  * from one thread to simplify locking.
280  */
281 static void
282 g_raid3_event_free(struct g_raid3_event *ep)
283 {
284 
285 	free(ep, M_RAID3);
286 }
287 
288 int
289 g_raid3_event_send(void *arg, int state, int flags)
290 {
291 	struct g_raid3_softc *sc;
292 	struct g_raid3_disk *disk;
293 	struct g_raid3_event *ep;
294 	int error;
295 
296 	ep = malloc(sizeof(*ep), M_RAID3, M_WAITOK);
297 	G_RAID3_DEBUG(4, "%s: Sending event %p.", __func__, ep);
298 	if ((flags & G_RAID3_EVENT_DEVICE) != 0) {
299 		disk = NULL;
300 		sc = arg;
301 	} else {
302 		disk = arg;
303 		sc = disk->d_softc;
304 	}
305 	ep->e_disk = disk;
306 	ep->e_state = state;
307 	ep->e_flags = flags;
308 	ep->e_error = 0;
309 	mtx_lock(&sc->sc_events_mtx);
310 	TAILQ_INSERT_TAIL(&sc->sc_events, ep, e_next);
311 	mtx_unlock(&sc->sc_events_mtx);
312 	G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, sc);
313 	mtx_lock(&sc->sc_queue_mtx);
314 	wakeup(sc);
315 	wakeup(&sc->sc_queue);
316 	mtx_unlock(&sc->sc_queue_mtx);
317 	if ((flags & G_RAID3_EVENT_DONTWAIT) != 0)
318 		return (0);
319 	sx_assert(&sc->sc_lock, SX_XLOCKED);
320 	G_RAID3_DEBUG(4, "%s: Sleeping %p.", __func__, ep);
321 	sx_xunlock(&sc->sc_lock);
322 	while ((ep->e_flags & G_RAID3_EVENT_DONE) == 0) {
323 		mtx_lock(&sc->sc_events_mtx);
324 		MSLEEP(ep, &sc->sc_events_mtx, PRIBIO | PDROP, "r3:event",
325 		    hz * 5);
326 	}
327 	error = ep->e_error;
328 	g_raid3_event_free(ep);
329 	sx_xlock(&sc->sc_lock);
330 	return (error);
331 }
332 
333 static struct g_raid3_event *
334 g_raid3_event_get(struct g_raid3_softc *sc)
335 {
336 	struct g_raid3_event *ep;
337 
338 	mtx_lock(&sc->sc_events_mtx);
339 	ep = TAILQ_FIRST(&sc->sc_events);
340 	mtx_unlock(&sc->sc_events_mtx);
341 	return (ep);
342 }
343 
344 static void
345 g_raid3_event_remove(struct g_raid3_softc *sc, struct g_raid3_event *ep)
346 {
347 
348 	mtx_lock(&sc->sc_events_mtx);
349 	TAILQ_REMOVE(&sc->sc_events, ep, e_next);
350 	mtx_unlock(&sc->sc_events_mtx);
351 }
352 
353 static void
354 g_raid3_event_cancel(struct g_raid3_disk *disk)
355 {
356 	struct g_raid3_softc *sc;
357 	struct g_raid3_event *ep, *tmpep;
358 
359 	sc = disk->d_softc;
360 	sx_assert(&sc->sc_lock, SX_XLOCKED);
361 
362 	mtx_lock(&sc->sc_events_mtx);
363 	TAILQ_FOREACH_SAFE(ep, &sc->sc_events, e_next, tmpep) {
364 		if ((ep->e_flags & G_RAID3_EVENT_DEVICE) != 0)
365 			continue;
366 		if (ep->e_disk != disk)
367 			continue;
368 		TAILQ_REMOVE(&sc->sc_events, ep, e_next);
369 		if ((ep->e_flags & G_RAID3_EVENT_DONTWAIT) != 0)
370 			g_raid3_event_free(ep);
371 		else {
372 			ep->e_error = ECANCELED;
373 			wakeup(ep);
374 		}
375 	}
376 	mtx_unlock(&sc->sc_events_mtx);
377 }
378 
379 /*
380  * Return the number of disks in the given state.
381  * If state is equal to -1, count all connected disks.
382  */
383 u_int
384 g_raid3_ndisks(struct g_raid3_softc *sc, int state)
385 {
386 	struct g_raid3_disk *disk;
387 	u_int n, ndisks;
388 
389 	sx_assert(&sc->sc_lock, SX_LOCKED);
390 
391 	for (n = ndisks = 0; n < sc->sc_ndisks; n++) {
392 		disk = &sc->sc_disks[n];
393 		if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
394 			continue;
395 		if (state == -1 || disk->d_state == state)
396 			ndisks++;
397 	}
398 	return (ndisks);
399 }
400 
401 static u_int
402 g_raid3_nrequests(struct g_raid3_softc *sc, struct g_consumer *cp)
403 {
404 	struct bio *bp;
405 	u_int nreqs = 0;
406 
407 	mtx_lock(&sc->sc_queue_mtx);
408 	TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
409 		if (bp->bio_from == cp)
410 			nreqs++;
411 	}
412 	mtx_unlock(&sc->sc_queue_mtx);
413 	return (nreqs);
414 }
415 
416 static int
417 g_raid3_is_busy(struct g_raid3_softc *sc, struct g_consumer *cp)
418 {
419 
420 	if (cp->index > 0) {
421 		G_RAID3_DEBUG(2,
422 		    "I/O requests for %s exist, can't destroy it now.",
423 		    cp->provider->name);
424 		return (1);
425 	}
426 	if (g_raid3_nrequests(sc, cp) > 0) {
427 		G_RAID3_DEBUG(2,
428 		    "I/O requests for %s in queue, can't destroy it now.",
429 		    cp->provider->name);
430 		return (1);
431 	}
432 	return (0);
433 }
434 
435 static void
436 g_raid3_destroy_consumer(void *arg, int flags __unused)
437 {
438 	struct g_consumer *cp;
439 
440 	g_topology_assert();
441 
442 	cp = arg;
443 	G_RAID3_DEBUG(1, "Consumer %s destroyed.", cp->provider->name);
444 	g_detach(cp);
445 	g_destroy_consumer(cp);
446 }
447 
448 static void
449 g_raid3_kill_consumer(struct g_raid3_softc *sc, struct g_consumer *cp)
450 {
451 	struct g_provider *pp;
452 	int retaste_wait;
453 
454 	g_topology_assert();
455 
456 	cp->private = NULL;
457 	if (g_raid3_is_busy(sc, cp))
458 		return;
459 	G_RAID3_DEBUG(2, "Consumer %s destroyed.", cp->provider->name);
460 	pp = cp->provider;
461 	retaste_wait = 0;
462 	if (cp->acw == 1) {
463 		if ((pp->geom->flags & G_GEOM_WITHER) == 0)
464 			retaste_wait = 1;
465 	}
466 	G_RAID3_DEBUG(2, "Access %s r%dw%de%d = %d", pp->name, -cp->acr,
467 	    -cp->acw, -cp->ace, 0);
468 	if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
469 		g_access(cp, -cp->acr, -cp->acw, -cp->ace);
470 	if (retaste_wait) {
471 		/*
472 		 * After retaste event was send (inside g_access()), we can send
473 		 * event to detach and destroy consumer.
474 		 * A class, which has consumer to the given provider connected
475 		 * will not receive retaste event for the provider.
476 		 * This is the way how I ignore retaste events when I close
477 		 * consumers opened for write: I detach and destroy consumer
478 		 * after retaste event is sent.
479 		 */
480 		g_post_event(g_raid3_destroy_consumer, cp, M_WAITOK, NULL);
481 		return;
482 	}
483 	G_RAID3_DEBUG(1, "Consumer %s destroyed.", pp->name);
484 	g_detach(cp);
485 	g_destroy_consumer(cp);
486 }
487 
488 static int
489 g_raid3_connect_disk(struct g_raid3_disk *disk, struct g_provider *pp)
490 {
491 	struct g_consumer *cp;
492 	int error;
493 
494 	g_topology_assert_not();
495 	KASSERT(disk->d_consumer == NULL,
496 	    ("Disk already connected (device %s).", disk->d_softc->sc_name));
497 
498 	g_topology_lock();
499 	cp = g_new_consumer(disk->d_softc->sc_geom);
500 	error = g_attach(cp, pp);
501 	if (error != 0) {
502 		g_destroy_consumer(cp);
503 		g_topology_unlock();
504 		return (error);
505 	}
506 	error = g_access(cp, 1, 1, 1);
507 		g_topology_unlock();
508 	if (error != 0) {
509 		g_detach(cp);
510 		g_destroy_consumer(cp);
511 		G_RAID3_DEBUG(0, "Cannot open consumer %s (error=%d).",
512 		    pp->name, error);
513 		return (error);
514 	}
515 	disk->d_consumer = cp;
516 	disk->d_consumer->private = disk;
517 	disk->d_consumer->index = 0;
518 	G_RAID3_DEBUG(2, "Disk %s connected.", g_raid3_get_diskname(disk));
519 	return (0);
520 }
521 
522 static void
523 g_raid3_disconnect_consumer(struct g_raid3_softc *sc, struct g_consumer *cp)
524 {
525 
526 	g_topology_assert();
527 
528 	if (cp == NULL)
529 		return;
530 	if (cp->provider != NULL)
531 		g_raid3_kill_consumer(sc, cp);
532 	else
533 		g_destroy_consumer(cp);
534 }
535 
536 /*
537  * Initialize disk. This means allocate memory, create consumer, attach it
538  * to the provider and open access (r1w1e1) to it.
539  */
540 static struct g_raid3_disk *
541 g_raid3_init_disk(struct g_raid3_softc *sc, struct g_provider *pp,
542     struct g_raid3_metadata *md, int *errorp)
543 {
544 	struct g_raid3_disk *disk;
545 	int error;
546 
547 	disk = &sc->sc_disks[md->md_no];
548 	error = g_raid3_connect_disk(disk, pp);
549 	if (error != 0) {
550 		if (errorp != NULL)
551 			*errorp = error;
552 		return (NULL);
553 	}
554 	disk->d_state = G_RAID3_DISK_STATE_NONE;
555 	disk->d_flags = md->md_dflags;
556 	if (md->md_provider[0] != '\0')
557 		disk->d_flags |= G_RAID3_DISK_FLAG_HARDCODED;
558 	disk->d_sync.ds_consumer = NULL;
559 	disk->d_sync.ds_offset = md->md_sync_offset;
560 	disk->d_sync.ds_offset_done = md->md_sync_offset;
561 	disk->d_genid = md->md_genid;
562 	disk->d_sync.ds_syncid = md->md_syncid;
563 	if (errorp != NULL)
564 		*errorp = 0;
565 	return (disk);
566 }
567 
568 static void
569 g_raid3_destroy_disk(struct g_raid3_disk *disk)
570 {
571 	struct g_raid3_softc *sc;
572 
573 	g_topology_assert_not();
574 	sc = disk->d_softc;
575 	sx_assert(&sc->sc_lock, SX_XLOCKED);
576 
577 	if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
578 		return;
579 	g_raid3_event_cancel(disk);
580 	switch (disk->d_state) {
581 	case G_RAID3_DISK_STATE_SYNCHRONIZING:
582 		if (sc->sc_syncdisk != NULL)
583 			g_raid3_sync_stop(sc, 1);
584 		/* FALLTHROUGH */
585 	case G_RAID3_DISK_STATE_NEW:
586 	case G_RAID3_DISK_STATE_STALE:
587 	case G_RAID3_DISK_STATE_ACTIVE:
588 		g_topology_lock();
589 		g_raid3_disconnect_consumer(sc, disk->d_consumer);
590 		g_topology_unlock();
591 		disk->d_consumer = NULL;
592 		break;
593 	default:
594 		KASSERT(0 == 1, ("Wrong disk state (%s, %s).",
595 		    g_raid3_get_diskname(disk),
596 		    g_raid3_disk_state2str(disk->d_state)));
597 	}
598 	disk->d_state = G_RAID3_DISK_STATE_NODISK;
599 }
600 
601 static void
602 g_raid3_destroy_device(struct g_raid3_softc *sc)
603 {
604 	struct g_raid3_event *ep;
605 	struct g_raid3_disk *disk;
606 	struct g_geom *gp;
607 	struct g_consumer *cp;
608 	u_int n;
609 
610 	g_topology_assert_not();
611 	sx_assert(&sc->sc_lock, SX_XLOCKED);
612 
613 	gp = sc->sc_geom;
614 	if (sc->sc_provider != NULL)
615 		g_raid3_destroy_provider(sc);
616 	for (n = 0; n < sc->sc_ndisks; n++) {
617 		disk = &sc->sc_disks[n];
618 		if (disk->d_state != G_RAID3_DISK_STATE_NODISK) {
619 			disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
620 			g_raid3_update_metadata(disk);
621 			g_raid3_destroy_disk(disk);
622 		}
623 	}
624 	while ((ep = g_raid3_event_get(sc)) != NULL) {
625 		g_raid3_event_remove(sc, ep);
626 		if ((ep->e_flags & G_RAID3_EVENT_DONTWAIT) != 0)
627 			g_raid3_event_free(ep);
628 		else {
629 			ep->e_error = ECANCELED;
630 			ep->e_flags |= G_RAID3_EVENT_DONE;
631 			G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, ep);
632 			mtx_lock(&sc->sc_events_mtx);
633 			wakeup(ep);
634 			mtx_unlock(&sc->sc_events_mtx);
635 		}
636 	}
637 	callout_drain(&sc->sc_callout);
638 	cp = LIST_FIRST(&sc->sc_sync.ds_geom->consumer);
639 	g_topology_lock();
640 	if (cp != NULL)
641 		g_raid3_disconnect_consumer(sc, cp);
642 	g_wither_geom(sc->sc_sync.ds_geom, ENXIO);
643 	G_RAID3_DEBUG(0, "Device %s destroyed.", gp->name);
644 	g_wither_geom(gp, ENXIO);
645 	g_topology_unlock();
646 	if (!g_raid3_use_malloc) {
647 		uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_64K].sz_zone);
648 		uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_16K].sz_zone);
649 		uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_4K].sz_zone);
650 	}
651 	mtx_destroy(&sc->sc_queue_mtx);
652 	mtx_destroy(&sc->sc_events_mtx);
653 	sx_xunlock(&sc->sc_lock);
654 	sx_destroy(&sc->sc_lock);
655 }
656 
657 static void
658 g_raid3_orphan(struct g_consumer *cp)
659 {
660 	struct g_raid3_disk *disk;
661 
662 	g_topology_assert();
663 
664 	disk = cp->private;
665 	if (disk == NULL)
666 		return;
667 	disk->d_softc->sc_bump_id = G_RAID3_BUMP_SYNCID;
668 	g_raid3_event_send(disk, G_RAID3_DISK_STATE_DISCONNECTED,
669 	    G_RAID3_EVENT_DONTWAIT);
670 }
671 
672 static int
673 g_raid3_write_metadata(struct g_raid3_disk *disk, struct g_raid3_metadata *md)
674 {
675 	struct g_raid3_softc *sc;
676 	struct g_consumer *cp;
677 	off_t offset, length;
678 	u_char *sector;
679 	int error = 0;
680 
681 	g_topology_assert_not();
682 	sc = disk->d_softc;
683 	sx_assert(&sc->sc_lock, SX_LOCKED);
684 
685 	cp = disk->d_consumer;
686 	KASSERT(cp != NULL, ("NULL consumer (%s).", sc->sc_name));
687 	KASSERT(cp->provider != NULL, ("NULL provider (%s).", sc->sc_name));
688 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
689 	    ("Consumer %s closed? (r%dw%de%d).", cp->provider->name, cp->acr,
690 	    cp->acw, cp->ace));
691 	length = cp->provider->sectorsize;
692 	offset = cp->provider->mediasize - length;
693 	sector = malloc((size_t)length, M_RAID3, M_WAITOK | M_ZERO);
694 	if (md != NULL)
695 		raid3_metadata_encode(md, sector);
696 	error = g_write_data(cp, offset, sector, length);
697 	free(sector, M_RAID3);
698 	if (error != 0) {
699 		if ((disk->d_flags & G_RAID3_DISK_FLAG_BROKEN) == 0) {
700 			G_RAID3_DEBUG(0, "Cannot write metadata on %s "
701 			    "(device=%s, error=%d).",
702 			    g_raid3_get_diskname(disk), sc->sc_name, error);
703 			disk->d_flags |= G_RAID3_DISK_FLAG_BROKEN;
704 		} else {
705 			G_RAID3_DEBUG(1, "Cannot write metadata on %s "
706 			    "(device=%s, error=%d).",
707 			    g_raid3_get_diskname(disk), sc->sc_name, error);
708 		}
709 		if (g_raid3_disconnect_on_failure &&
710 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
711 			sc->sc_bump_id |= G_RAID3_BUMP_GENID;
712 			g_raid3_event_send(disk,
713 			    G_RAID3_DISK_STATE_DISCONNECTED,
714 			    G_RAID3_EVENT_DONTWAIT);
715 		}
716 	}
717 	return (error);
718 }
719 
720 int
721 g_raid3_clear_metadata(struct g_raid3_disk *disk)
722 {
723 	int error;
724 
725 	g_topology_assert_not();
726 	sx_assert(&disk->d_softc->sc_lock, SX_LOCKED);
727 
728 	error = g_raid3_write_metadata(disk, NULL);
729 	if (error == 0) {
730 		G_RAID3_DEBUG(2, "Metadata on %s cleared.",
731 		    g_raid3_get_diskname(disk));
732 	} else {
733 		G_RAID3_DEBUG(0,
734 		    "Cannot clear metadata on disk %s (error=%d).",
735 		    g_raid3_get_diskname(disk), error);
736 	}
737 	return (error);
738 }
739 
740 void
741 g_raid3_fill_metadata(struct g_raid3_disk *disk, struct g_raid3_metadata *md)
742 {
743 	struct g_raid3_softc *sc;
744 	struct g_provider *pp;
745 
746 	sc = disk->d_softc;
747 	strlcpy(md->md_magic, G_RAID3_MAGIC, sizeof(md->md_magic));
748 	md->md_version = G_RAID3_VERSION;
749 	strlcpy(md->md_name, sc->sc_name, sizeof(md->md_name));
750 	md->md_id = sc->sc_id;
751 	md->md_all = sc->sc_ndisks;
752 	md->md_genid = sc->sc_genid;
753 	md->md_mediasize = sc->sc_mediasize;
754 	md->md_sectorsize = sc->sc_sectorsize;
755 	md->md_mflags = (sc->sc_flags & G_RAID3_DEVICE_FLAG_MASK);
756 	md->md_no = disk->d_no;
757 	md->md_syncid = disk->d_sync.ds_syncid;
758 	md->md_dflags = (disk->d_flags & G_RAID3_DISK_FLAG_MASK);
759 	if (disk->d_state != G_RAID3_DISK_STATE_SYNCHRONIZING)
760 		md->md_sync_offset = 0;
761 	else {
762 		md->md_sync_offset =
763 		    disk->d_sync.ds_offset_done / (sc->sc_ndisks - 1);
764 	}
765 	if (disk->d_consumer != NULL && disk->d_consumer->provider != NULL)
766 		pp = disk->d_consumer->provider;
767 	else
768 		pp = NULL;
769 	if ((disk->d_flags & G_RAID3_DISK_FLAG_HARDCODED) != 0 && pp != NULL)
770 		strlcpy(md->md_provider, pp->name, sizeof(md->md_provider));
771 	else
772 		bzero(md->md_provider, sizeof(md->md_provider));
773 	if (pp != NULL)
774 		md->md_provsize = pp->mediasize;
775 	else
776 		md->md_provsize = 0;
777 }
778 
779 void
780 g_raid3_update_metadata(struct g_raid3_disk *disk)
781 {
782 	struct g_raid3_softc *sc;
783 	struct g_raid3_metadata md;
784 	int error;
785 
786 	g_topology_assert_not();
787 	sc = disk->d_softc;
788 	sx_assert(&sc->sc_lock, SX_LOCKED);
789 
790 	g_raid3_fill_metadata(disk, &md);
791 	error = g_raid3_write_metadata(disk, &md);
792 	if (error == 0) {
793 		G_RAID3_DEBUG(2, "Metadata on %s updated.",
794 		    g_raid3_get_diskname(disk));
795 	} else {
796 		G_RAID3_DEBUG(0,
797 		    "Cannot update metadata on disk %s (error=%d).",
798 		    g_raid3_get_diskname(disk), error);
799 	}
800 }
801 
802 static void
803 g_raid3_bump_syncid(struct g_raid3_softc *sc)
804 {
805 	struct g_raid3_disk *disk;
806 	u_int n;
807 
808 	g_topology_assert_not();
809 	sx_assert(&sc->sc_lock, SX_XLOCKED);
810 	KASSERT(g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) > 0,
811 	    ("%s called with no active disks (device=%s).", __func__,
812 	    sc->sc_name));
813 
814 	sc->sc_syncid++;
815 	G_RAID3_DEBUG(1, "Device %s: syncid bumped to %u.", sc->sc_name,
816 	    sc->sc_syncid);
817 	for (n = 0; n < sc->sc_ndisks; n++) {
818 		disk = &sc->sc_disks[n];
819 		if (disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
820 		    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
821 			disk->d_sync.ds_syncid = sc->sc_syncid;
822 			g_raid3_update_metadata(disk);
823 		}
824 	}
825 }
826 
827 static void
828 g_raid3_bump_genid(struct g_raid3_softc *sc)
829 {
830 	struct g_raid3_disk *disk;
831 	u_int n;
832 
833 	g_topology_assert_not();
834 	sx_assert(&sc->sc_lock, SX_XLOCKED);
835 	KASSERT(g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) > 0,
836 	    ("%s called with no active disks (device=%s).", __func__,
837 	    sc->sc_name));
838 
839 	sc->sc_genid++;
840 	G_RAID3_DEBUG(1, "Device %s: genid bumped to %u.", sc->sc_name,
841 	    sc->sc_genid);
842 	for (n = 0; n < sc->sc_ndisks; n++) {
843 		disk = &sc->sc_disks[n];
844 		if (disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
845 		    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
846 			disk->d_genid = sc->sc_genid;
847 			g_raid3_update_metadata(disk);
848 		}
849 	}
850 }
851 
852 static int
853 g_raid3_idle(struct g_raid3_softc *sc, int acw)
854 {
855 	struct g_raid3_disk *disk;
856 	u_int i;
857 	int timeout;
858 
859 	g_topology_assert_not();
860 	sx_assert(&sc->sc_lock, SX_XLOCKED);
861 
862 	if (sc->sc_provider == NULL)
863 		return (0);
864 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) != 0)
865 		return (0);
866 	if (sc->sc_idle)
867 		return (0);
868 	if (sc->sc_writes > 0)
869 		return (0);
870 	if (acw > 0 || (acw == -1 && sc->sc_provider->acw > 0)) {
871 		timeout = g_raid3_idletime - (time_uptime - sc->sc_last_write);
872 		if (!g_raid3_shutdown && timeout > 0)
873 			return (timeout);
874 	}
875 	sc->sc_idle = 1;
876 	for (i = 0; i < sc->sc_ndisks; i++) {
877 		disk = &sc->sc_disks[i];
878 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE)
879 			continue;
880 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as clean.",
881 		    g_raid3_get_diskname(disk), sc->sc_name);
882 		disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
883 		g_raid3_update_metadata(disk);
884 	}
885 	return (0);
886 }
887 
888 static void
889 g_raid3_unidle(struct g_raid3_softc *sc)
890 {
891 	struct g_raid3_disk *disk;
892 	u_int i;
893 
894 	g_topology_assert_not();
895 	sx_assert(&sc->sc_lock, SX_XLOCKED);
896 
897 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) != 0)
898 		return;
899 	sc->sc_idle = 0;
900 	sc->sc_last_write = time_uptime;
901 	for (i = 0; i < sc->sc_ndisks; i++) {
902 		disk = &sc->sc_disks[i];
903 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE)
904 			continue;
905 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as dirty.",
906 		    g_raid3_get_diskname(disk), sc->sc_name);
907 		disk->d_flags |= G_RAID3_DISK_FLAG_DIRTY;
908 		g_raid3_update_metadata(disk);
909 	}
910 }
911 
912 /*
913  * Treat bio_driver1 field in parent bio as list head and field bio_caller1
914  * in child bio as pointer to the next element on the list.
915  */
916 #define	G_RAID3_HEAD_BIO(pbp)	(pbp)->bio_driver1
917 
918 #define	G_RAID3_NEXT_BIO(cbp)	(cbp)->bio_caller1
919 
920 #define	G_RAID3_FOREACH_BIO(pbp, bp)					\
921 	for ((bp) = G_RAID3_HEAD_BIO(pbp); (bp) != NULL;		\
922 	    (bp) = G_RAID3_NEXT_BIO(bp))
923 
924 #define	G_RAID3_FOREACH_SAFE_BIO(pbp, bp, tmpbp)			\
925 	for ((bp) = G_RAID3_HEAD_BIO(pbp);				\
926 	    (bp) != NULL && ((tmpbp) = G_RAID3_NEXT_BIO(bp), 1);	\
927 	    (bp) = (tmpbp))
928 
929 static void
930 g_raid3_init_bio(struct bio *pbp)
931 {
932 
933 	G_RAID3_HEAD_BIO(pbp) = NULL;
934 }
935 
936 static void
937 g_raid3_remove_bio(struct bio *cbp)
938 {
939 	struct bio *pbp, *bp;
940 
941 	pbp = cbp->bio_parent;
942 	if (G_RAID3_HEAD_BIO(pbp) == cbp)
943 		G_RAID3_HEAD_BIO(pbp) = G_RAID3_NEXT_BIO(cbp);
944 	else {
945 		G_RAID3_FOREACH_BIO(pbp, bp) {
946 			if (G_RAID3_NEXT_BIO(bp) == cbp) {
947 				G_RAID3_NEXT_BIO(bp) = G_RAID3_NEXT_BIO(cbp);
948 				break;
949 			}
950 		}
951 	}
952 	G_RAID3_NEXT_BIO(cbp) = NULL;
953 }
954 
955 static void
956 g_raid3_replace_bio(struct bio *sbp, struct bio *dbp)
957 {
958 	struct bio *pbp, *bp;
959 
960 	g_raid3_remove_bio(sbp);
961 	pbp = dbp->bio_parent;
962 	G_RAID3_NEXT_BIO(sbp) = G_RAID3_NEXT_BIO(dbp);
963 	if (G_RAID3_HEAD_BIO(pbp) == dbp)
964 		G_RAID3_HEAD_BIO(pbp) = sbp;
965 	else {
966 		G_RAID3_FOREACH_BIO(pbp, bp) {
967 			if (G_RAID3_NEXT_BIO(bp) == dbp) {
968 				G_RAID3_NEXT_BIO(bp) = sbp;
969 				break;
970 			}
971 		}
972 	}
973 	G_RAID3_NEXT_BIO(dbp) = NULL;
974 }
975 
976 static void
977 g_raid3_destroy_bio(struct g_raid3_softc *sc, struct bio *cbp)
978 {
979 	struct bio *bp, *pbp;
980 	size_t size;
981 
982 	pbp = cbp->bio_parent;
983 	pbp->bio_children--;
984 	KASSERT(cbp->bio_data != NULL, ("NULL bio_data"));
985 	size = pbp->bio_length / (sc->sc_ndisks - 1);
986 	g_raid3_free(sc, cbp->bio_data, size);
987 	if (G_RAID3_HEAD_BIO(pbp) == cbp) {
988 		G_RAID3_HEAD_BIO(pbp) = G_RAID3_NEXT_BIO(cbp);
989 		G_RAID3_NEXT_BIO(cbp) = NULL;
990 		g_destroy_bio(cbp);
991 	} else {
992 		G_RAID3_FOREACH_BIO(pbp, bp) {
993 			if (G_RAID3_NEXT_BIO(bp) == cbp)
994 				break;
995 		}
996 		if (bp != NULL) {
997 			KASSERT(G_RAID3_NEXT_BIO(bp) != NULL,
998 			    ("NULL bp->bio_driver1"));
999 			G_RAID3_NEXT_BIO(bp) = G_RAID3_NEXT_BIO(cbp);
1000 			G_RAID3_NEXT_BIO(cbp) = NULL;
1001 		}
1002 		g_destroy_bio(cbp);
1003 	}
1004 }
1005 
1006 static struct bio *
1007 g_raid3_clone_bio(struct g_raid3_softc *sc, struct bio *pbp)
1008 {
1009 	struct bio *bp, *cbp;
1010 	size_t size;
1011 	int memflag;
1012 
1013 	cbp = g_clone_bio(pbp);
1014 	if (cbp == NULL)
1015 		return (NULL);
1016 	size = pbp->bio_length / (sc->sc_ndisks - 1);
1017 	if ((pbp->bio_cflags & G_RAID3_BIO_CFLAG_REGULAR) != 0)
1018 		memflag = M_WAITOK;
1019 	else
1020 		memflag = M_NOWAIT;
1021 	cbp->bio_data = g_raid3_alloc(sc, size, memflag);
1022 	if (cbp->bio_data == NULL) {
1023 		pbp->bio_children--;
1024 		g_destroy_bio(cbp);
1025 		return (NULL);
1026 	}
1027 	G_RAID3_NEXT_BIO(cbp) = NULL;
1028 	if (G_RAID3_HEAD_BIO(pbp) == NULL)
1029 		G_RAID3_HEAD_BIO(pbp) = cbp;
1030 	else {
1031 		G_RAID3_FOREACH_BIO(pbp, bp) {
1032 			if (G_RAID3_NEXT_BIO(bp) == NULL) {
1033 				G_RAID3_NEXT_BIO(bp) = cbp;
1034 				break;
1035 			}
1036 		}
1037 	}
1038 	return (cbp);
1039 }
1040 
1041 static void
1042 g_raid3_scatter(struct bio *pbp)
1043 {
1044 	struct g_raid3_softc *sc;
1045 	struct g_raid3_disk *disk;
1046 	struct bio *bp, *cbp, *tmpbp;
1047 	off_t atom, cadd, padd, left;
1048 	int first;
1049 
1050 	sc = pbp->bio_to->geom->softc;
1051 	bp = NULL;
1052 	if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_NOPARITY) == 0) {
1053 		/*
1054 		 * Find bio for which we should calculate data.
1055 		 */
1056 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1057 			if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_PARITY) != 0) {
1058 				bp = cbp;
1059 				break;
1060 			}
1061 		}
1062 		KASSERT(bp != NULL, ("NULL parity bio."));
1063 	}
1064 	atom = sc->sc_sectorsize / (sc->sc_ndisks - 1);
1065 	cadd = padd = 0;
1066 	for (left = pbp->bio_length; left > 0; left -= sc->sc_sectorsize) {
1067 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1068 			if (cbp == bp)
1069 				continue;
1070 			bcopy(pbp->bio_data + padd, cbp->bio_data + cadd, atom);
1071 			padd += atom;
1072 		}
1073 		cadd += atom;
1074 	}
1075 	if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_NOPARITY) == 0) {
1076 		/*
1077 		 * Calculate parity.
1078 		 */
1079 		first = 1;
1080 		G_RAID3_FOREACH_SAFE_BIO(pbp, cbp, tmpbp) {
1081 			if (cbp == bp)
1082 				continue;
1083 			if (first) {
1084 				bcopy(cbp->bio_data, bp->bio_data,
1085 				    bp->bio_length);
1086 				first = 0;
1087 			} else {
1088 				g_raid3_xor(cbp->bio_data, bp->bio_data,
1089 				    bp->bio_length);
1090 			}
1091 			if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_NODISK) != 0)
1092 				g_raid3_destroy_bio(sc, cbp);
1093 		}
1094 	}
1095 	G_RAID3_FOREACH_SAFE_BIO(pbp, cbp, tmpbp) {
1096 		struct g_consumer *cp;
1097 
1098 		disk = cbp->bio_caller2;
1099 		cp = disk->d_consumer;
1100 		cbp->bio_to = cp->provider;
1101 		G_RAID3_LOGREQ(3, cbp, "Sending request.");
1102 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1103 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1104 		    cp->acr, cp->acw, cp->ace));
1105 		cp->index++;
1106 		sc->sc_writes++;
1107 		g_io_request(cbp, cp);
1108 	}
1109 }
1110 
1111 static void
1112 g_raid3_gather(struct bio *pbp)
1113 {
1114 	struct g_raid3_softc *sc;
1115 	struct g_raid3_disk *disk;
1116 	struct bio *xbp, *fbp, *cbp;
1117 	off_t atom, cadd, padd, left;
1118 
1119 	sc = pbp->bio_to->geom->softc;
1120 	/*
1121 	 * Find bio for which we have to calculate data.
1122 	 * While going through this path, check if all requests
1123 	 * succeeded, if not, deny whole request.
1124 	 * If we're in COMPLETE mode, we allow one request to fail,
1125 	 * so if we find one, we're sending it to the parity consumer.
1126 	 * If there are more failed requests, we deny whole request.
1127 	 */
1128 	xbp = fbp = NULL;
1129 	G_RAID3_FOREACH_BIO(pbp, cbp) {
1130 		if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_PARITY) != 0) {
1131 			KASSERT(xbp == NULL, ("More than one parity bio."));
1132 			xbp = cbp;
1133 		}
1134 		if (cbp->bio_error == 0)
1135 			continue;
1136 		/*
1137 		 * Found failed request.
1138 		 */
1139 		if (fbp == NULL) {
1140 			if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_DEGRADED) != 0) {
1141 				/*
1142 				 * We are already in degraded mode, so we can't
1143 				 * accept any failures.
1144 				 */
1145 				if (pbp->bio_error == 0)
1146 					pbp->bio_error = cbp->bio_error;
1147 			} else {
1148 				fbp = cbp;
1149 			}
1150 		} else {
1151 			/*
1152 			 * Next failed request, that's too many.
1153 			 */
1154 			if (pbp->bio_error == 0)
1155 				pbp->bio_error = fbp->bio_error;
1156 		}
1157 		disk = cbp->bio_caller2;
1158 		if (disk == NULL)
1159 			continue;
1160 		if ((disk->d_flags & G_RAID3_DISK_FLAG_BROKEN) == 0) {
1161 			disk->d_flags |= G_RAID3_DISK_FLAG_BROKEN;
1162 			G_RAID3_LOGREQ(0, cbp, "Request failed (error=%d).",
1163 			    cbp->bio_error);
1164 		} else {
1165 			G_RAID3_LOGREQ(1, cbp, "Request failed (error=%d).",
1166 			    cbp->bio_error);
1167 		}
1168 		if (g_raid3_disconnect_on_failure &&
1169 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1170 			sc->sc_bump_id |= G_RAID3_BUMP_GENID;
1171 			g_raid3_event_send(disk,
1172 			    G_RAID3_DISK_STATE_DISCONNECTED,
1173 			    G_RAID3_EVENT_DONTWAIT);
1174 		}
1175 	}
1176 	if (pbp->bio_error != 0)
1177 		goto finish;
1178 	if (fbp != NULL && (pbp->bio_pflags & G_RAID3_BIO_PFLAG_VERIFY) != 0) {
1179 		pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_VERIFY;
1180 		if (xbp != fbp)
1181 			g_raid3_replace_bio(xbp, fbp);
1182 		g_raid3_destroy_bio(sc, fbp);
1183 	} else if (fbp != NULL) {
1184 		struct g_consumer *cp;
1185 
1186 		/*
1187 		 * One request failed, so send the same request to
1188 		 * the parity consumer.
1189 		 */
1190 		disk = pbp->bio_driver2;
1191 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE) {
1192 			pbp->bio_error = fbp->bio_error;
1193 			goto finish;
1194 		}
1195 		pbp->bio_pflags |= G_RAID3_BIO_PFLAG_DEGRADED;
1196 		pbp->bio_inbed--;
1197 		fbp->bio_flags &= ~(BIO_DONE | BIO_ERROR);
1198 		if (disk->d_no == sc->sc_ndisks - 1)
1199 			fbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1200 		fbp->bio_error = 0;
1201 		fbp->bio_completed = 0;
1202 		fbp->bio_children = 0;
1203 		fbp->bio_inbed = 0;
1204 		cp = disk->d_consumer;
1205 		fbp->bio_caller2 = disk;
1206 		fbp->bio_to = cp->provider;
1207 		G_RAID3_LOGREQ(3, fbp, "Sending request (recover).");
1208 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1209 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1210 		    cp->acr, cp->acw, cp->ace));
1211 		cp->index++;
1212 		g_io_request(fbp, cp);
1213 		return;
1214 	}
1215 	if (xbp != NULL) {
1216 		/*
1217 		 * Calculate parity.
1218 		 */
1219 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1220 			if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_PARITY) != 0)
1221 				continue;
1222 			g_raid3_xor(cbp->bio_data, xbp->bio_data,
1223 			    xbp->bio_length);
1224 		}
1225 		xbp->bio_cflags &= ~G_RAID3_BIO_CFLAG_PARITY;
1226 		if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_VERIFY) != 0) {
1227 			if (!g_raid3_is_zero(xbp)) {
1228 				g_raid3_parity_mismatch++;
1229 				pbp->bio_error = EIO;
1230 				goto finish;
1231 			}
1232 			g_raid3_destroy_bio(sc, xbp);
1233 		}
1234 	}
1235 	atom = sc->sc_sectorsize / (sc->sc_ndisks - 1);
1236 	cadd = padd = 0;
1237 	for (left = pbp->bio_length; left > 0; left -= sc->sc_sectorsize) {
1238 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1239 			bcopy(cbp->bio_data + cadd, pbp->bio_data + padd, atom);
1240 			pbp->bio_completed += atom;
1241 			padd += atom;
1242 		}
1243 		cadd += atom;
1244 	}
1245 finish:
1246 	if (pbp->bio_error == 0)
1247 		G_RAID3_LOGREQ(3, pbp, "Request finished.");
1248 	else {
1249 		if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_VERIFY) != 0)
1250 			G_RAID3_LOGREQ(1, pbp, "Verification error.");
1251 		else
1252 			G_RAID3_LOGREQ(0, pbp, "Request failed.");
1253 	}
1254 	pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_MASK;
1255 	while ((cbp = G_RAID3_HEAD_BIO(pbp)) != NULL)
1256 		g_raid3_destroy_bio(sc, cbp);
1257 	g_io_deliver(pbp, pbp->bio_error);
1258 }
1259 
1260 static void
1261 g_raid3_done(struct bio *bp)
1262 {
1263 	struct g_raid3_softc *sc;
1264 
1265 	sc = bp->bio_from->geom->softc;
1266 	bp->bio_cflags |= G_RAID3_BIO_CFLAG_REGULAR;
1267 	G_RAID3_LOGREQ(3, bp, "Regular request done (error=%d).", bp->bio_error);
1268 	mtx_lock(&sc->sc_queue_mtx);
1269 	bioq_insert_head(&sc->sc_queue, bp);
1270 	mtx_unlock(&sc->sc_queue_mtx);
1271 	wakeup(sc);
1272 	wakeup(&sc->sc_queue);
1273 }
1274 
1275 static void
1276 g_raid3_regular_request(struct bio *cbp)
1277 {
1278 	struct g_raid3_softc *sc;
1279 	struct g_raid3_disk *disk;
1280 	struct bio *pbp;
1281 
1282 	g_topology_assert_not();
1283 
1284 	pbp = cbp->bio_parent;
1285 	sc = pbp->bio_to->geom->softc;
1286 	cbp->bio_from->index--;
1287 	if (cbp->bio_cmd == BIO_WRITE)
1288 		sc->sc_writes--;
1289 	disk = cbp->bio_from->private;
1290 	if (disk == NULL) {
1291 		g_topology_lock();
1292 		g_raid3_kill_consumer(sc, cbp->bio_from);
1293 		g_topology_unlock();
1294 	}
1295 
1296 	G_RAID3_LOGREQ(3, cbp, "Request finished.");
1297 	pbp->bio_inbed++;
1298 	KASSERT(pbp->bio_inbed <= pbp->bio_children,
1299 	    ("bio_inbed (%u) is bigger than bio_children (%u).", pbp->bio_inbed,
1300 	    pbp->bio_children));
1301 	if (pbp->bio_inbed != pbp->bio_children)
1302 		return;
1303 	switch (pbp->bio_cmd) {
1304 	case BIO_READ:
1305 		g_raid3_gather(pbp);
1306 		break;
1307 	case BIO_WRITE:
1308 	case BIO_DELETE:
1309 	    {
1310 		int error = 0;
1311 
1312 		pbp->bio_completed = pbp->bio_length;
1313 		while ((cbp = G_RAID3_HEAD_BIO(pbp)) != NULL) {
1314 			if (cbp->bio_error == 0) {
1315 				g_raid3_destroy_bio(sc, cbp);
1316 				continue;
1317 			}
1318 
1319 			if (error == 0)
1320 				error = cbp->bio_error;
1321 			else if (pbp->bio_error == 0) {
1322 				/*
1323 				 * Next failed request, that's too many.
1324 				 */
1325 				pbp->bio_error = error;
1326 			}
1327 
1328 			disk = cbp->bio_caller2;
1329 			if (disk == NULL) {
1330 				g_raid3_destroy_bio(sc, cbp);
1331 				continue;
1332 			}
1333 
1334 			if ((disk->d_flags & G_RAID3_DISK_FLAG_BROKEN) == 0) {
1335 				disk->d_flags |= G_RAID3_DISK_FLAG_BROKEN;
1336 				G_RAID3_LOGREQ(0, cbp,
1337 				    "Request failed (error=%d).",
1338 				    cbp->bio_error);
1339 			} else {
1340 				G_RAID3_LOGREQ(1, cbp,
1341 				    "Request failed (error=%d).",
1342 				    cbp->bio_error);
1343 			}
1344 			if (g_raid3_disconnect_on_failure &&
1345 			    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1346 				sc->sc_bump_id |= G_RAID3_BUMP_GENID;
1347 				g_raid3_event_send(disk,
1348 				    G_RAID3_DISK_STATE_DISCONNECTED,
1349 				    G_RAID3_EVENT_DONTWAIT);
1350 			}
1351 			g_raid3_destroy_bio(sc, cbp);
1352 		}
1353 		if (pbp->bio_error == 0)
1354 			G_RAID3_LOGREQ(3, pbp, "Request finished.");
1355 		else
1356 			G_RAID3_LOGREQ(0, pbp, "Request failed.");
1357 		pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_DEGRADED;
1358 		pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_NOPARITY;
1359 		bioq_remove(&sc->sc_inflight, pbp);
1360 		/* Release delayed sync requests if possible. */
1361 		g_raid3_sync_release(sc);
1362 		g_io_deliver(pbp, pbp->bio_error);
1363 		break;
1364 	    }
1365 	}
1366 }
1367 
1368 static void
1369 g_raid3_sync_done(struct bio *bp)
1370 {
1371 	struct g_raid3_softc *sc;
1372 
1373 	G_RAID3_LOGREQ(3, bp, "Synchronization request delivered.");
1374 	sc = bp->bio_from->geom->softc;
1375 	bp->bio_cflags |= G_RAID3_BIO_CFLAG_SYNC;
1376 	mtx_lock(&sc->sc_queue_mtx);
1377 	bioq_insert_head(&sc->sc_queue, bp);
1378 	mtx_unlock(&sc->sc_queue_mtx);
1379 	wakeup(sc);
1380 	wakeup(&sc->sc_queue);
1381 }
1382 
1383 static void
1384 g_raid3_flush(struct g_raid3_softc *sc, struct bio *bp)
1385 {
1386 	struct bio_queue_head queue;
1387 	struct g_raid3_disk *disk;
1388 	struct g_consumer *cp;
1389 	struct bio *cbp;
1390 	u_int i;
1391 
1392 	bioq_init(&queue);
1393 	for (i = 0; i < sc->sc_ndisks; i++) {
1394 		disk = &sc->sc_disks[i];
1395 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE)
1396 			continue;
1397 		cbp = g_clone_bio(bp);
1398 		if (cbp == NULL) {
1399 			for (cbp = bioq_first(&queue); cbp != NULL;
1400 			    cbp = bioq_first(&queue)) {
1401 				bioq_remove(&queue, cbp);
1402 				g_destroy_bio(cbp);
1403 			}
1404 			if (bp->bio_error == 0)
1405 				bp->bio_error = ENOMEM;
1406 			g_io_deliver(bp, bp->bio_error);
1407 			return;
1408 		}
1409 		bioq_insert_tail(&queue, cbp);
1410 		cbp->bio_done = g_std_done;
1411 		cbp->bio_caller1 = disk;
1412 		cbp->bio_to = disk->d_consumer->provider;
1413 	}
1414 	for (cbp = bioq_first(&queue); cbp != NULL; cbp = bioq_first(&queue)) {
1415 		bioq_remove(&queue, cbp);
1416 		G_RAID3_LOGREQ(3, cbp, "Sending request.");
1417 		disk = cbp->bio_caller1;
1418 		cbp->bio_caller1 = NULL;
1419 		cp = disk->d_consumer;
1420 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1421 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1422 		    cp->acr, cp->acw, cp->ace));
1423 		g_io_request(cbp, disk->d_consumer);
1424 	}
1425 }
1426 
1427 static void
1428 g_raid3_start(struct bio *bp)
1429 {
1430 	struct g_raid3_softc *sc;
1431 
1432 	sc = bp->bio_to->geom->softc;
1433 	/*
1434 	 * If sc == NULL or there are no valid disks, provider's error
1435 	 * should be set and g_raid3_start() should not be called at all.
1436 	 */
1437 	KASSERT(sc != NULL && (sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
1438 	    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE),
1439 	    ("Provider's error should be set (error=%d)(device=%s).",
1440 	    bp->bio_to->error, bp->bio_to->name));
1441 	G_RAID3_LOGREQ(3, bp, "Request received.");
1442 
1443 	switch (bp->bio_cmd) {
1444 	case BIO_READ:
1445 	case BIO_WRITE:
1446 	case BIO_DELETE:
1447 		break;
1448 	case BIO_FLUSH:
1449 		g_raid3_flush(sc, bp);
1450 		return;
1451 	case BIO_GETATTR:
1452 	default:
1453 		g_io_deliver(bp, EOPNOTSUPP);
1454 		return;
1455 	}
1456 	mtx_lock(&sc->sc_queue_mtx);
1457 	bioq_insert_tail(&sc->sc_queue, bp);
1458 	mtx_unlock(&sc->sc_queue_mtx);
1459 	G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1460 	wakeup(sc);
1461 }
1462 
1463 /*
1464  * Return TRUE if the given request is colliding with a in-progress
1465  * synchronization request.
1466  */
1467 static int
1468 g_raid3_sync_collision(struct g_raid3_softc *sc, struct bio *bp)
1469 {
1470 	struct g_raid3_disk *disk;
1471 	struct bio *sbp;
1472 	off_t rstart, rend, sstart, send;
1473 	int i;
1474 
1475 	disk = sc->sc_syncdisk;
1476 	if (disk == NULL)
1477 		return (0);
1478 	rstart = bp->bio_offset;
1479 	rend = bp->bio_offset + bp->bio_length;
1480 	for (i = 0; i < g_raid3_syncreqs; i++) {
1481 		sbp = disk->d_sync.ds_bios[i];
1482 		if (sbp == NULL)
1483 			continue;
1484 		sstart = sbp->bio_offset;
1485 		send = sbp->bio_length;
1486 		if (sbp->bio_cmd == BIO_WRITE) {
1487 			sstart *= sc->sc_ndisks - 1;
1488 			send *= sc->sc_ndisks - 1;
1489 		}
1490 		send += sstart;
1491 		if (rend > sstart && rstart < send)
1492 			return (1);
1493 	}
1494 	return (0);
1495 }
1496 
1497 /*
1498  * Return TRUE if the given sync request is colliding with a in-progress regular
1499  * request.
1500  */
1501 static int
1502 g_raid3_regular_collision(struct g_raid3_softc *sc, struct bio *sbp)
1503 {
1504 	off_t rstart, rend, sstart, send;
1505 	struct bio *bp;
1506 
1507 	if (sc->sc_syncdisk == NULL)
1508 		return (0);
1509 	sstart = sbp->bio_offset;
1510 	send = sstart + sbp->bio_length;
1511 	TAILQ_FOREACH(bp, &sc->sc_inflight.queue, bio_queue) {
1512 		rstart = bp->bio_offset;
1513 		rend = bp->bio_offset + bp->bio_length;
1514 		if (rend > sstart && rstart < send)
1515 			return (1);
1516 	}
1517 	return (0);
1518 }
1519 
1520 /*
1521  * Puts request onto delayed queue.
1522  */
1523 static void
1524 g_raid3_regular_delay(struct g_raid3_softc *sc, struct bio *bp)
1525 {
1526 
1527 	G_RAID3_LOGREQ(2, bp, "Delaying request.");
1528 	bioq_insert_head(&sc->sc_regular_delayed, bp);
1529 }
1530 
1531 /*
1532  * Puts synchronization request onto delayed queue.
1533  */
1534 static void
1535 g_raid3_sync_delay(struct g_raid3_softc *sc, struct bio *bp)
1536 {
1537 
1538 	G_RAID3_LOGREQ(2, bp, "Delaying synchronization request.");
1539 	bioq_insert_tail(&sc->sc_sync_delayed, bp);
1540 }
1541 
1542 /*
1543  * Releases delayed regular requests which don't collide anymore with sync
1544  * requests.
1545  */
1546 static void
1547 g_raid3_regular_release(struct g_raid3_softc *sc)
1548 {
1549 	struct bio *bp, *bp2;
1550 
1551 	TAILQ_FOREACH_SAFE(bp, &sc->sc_regular_delayed.queue, bio_queue, bp2) {
1552 		if (g_raid3_sync_collision(sc, bp))
1553 			continue;
1554 		bioq_remove(&sc->sc_regular_delayed, bp);
1555 		G_RAID3_LOGREQ(2, bp, "Releasing delayed request (%p).", bp);
1556 		mtx_lock(&sc->sc_queue_mtx);
1557 		bioq_insert_head(&sc->sc_queue, bp);
1558 #if 0
1559 		/*
1560 		 * wakeup() is not needed, because this function is called from
1561 		 * the worker thread.
1562 		 */
1563 		wakeup(&sc->sc_queue);
1564 #endif
1565 		mtx_unlock(&sc->sc_queue_mtx);
1566 	}
1567 }
1568 
1569 /*
1570  * Releases delayed sync requests which don't collide anymore with regular
1571  * requests.
1572  */
1573 static void
1574 g_raid3_sync_release(struct g_raid3_softc *sc)
1575 {
1576 	struct bio *bp, *bp2;
1577 
1578 	TAILQ_FOREACH_SAFE(bp, &sc->sc_sync_delayed.queue, bio_queue, bp2) {
1579 		if (g_raid3_regular_collision(sc, bp))
1580 			continue;
1581 		bioq_remove(&sc->sc_sync_delayed, bp);
1582 		G_RAID3_LOGREQ(2, bp,
1583 		    "Releasing delayed synchronization request.");
1584 		g_io_request(bp, bp->bio_from);
1585 	}
1586 }
1587 
1588 /*
1589  * Handle synchronization requests.
1590  * Every synchronization request is two-steps process: first, READ request is
1591  * send to active provider and then WRITE request (with read data) to the provider
1592  * being synchronized. When WRITE is finished, new synchronization request is
1593  * send.
1594  */
1595 static void
1596 g_raid3_sync_request(struct bio *bp)
1597 {
1598 	struct g_raid3_softc *sc;
1599 	struct g_raid3_disk *disk;
1600 
1601 	bp->bio_from->index--;
1602 	sc = bp->bio_from->geom->softc;
1603 	disk = bp->bio_from->private;
1604 	if (disk == NULL) {
1605 		sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1606 		g_topology_lock();
1607 		g_raid3_kill_consumer(sc, bp->bio_from);
1608 		g_topology_unlock();
1609 		free(bp->bio_data, M_RAID3);
1610 		g_destroy_bio(bp);
1611 		sx_xlock(&sc->sc_lock);
1612 		return;
1613 	}
1614 
1615 	/*
1616 	 * Synchronization request.
1617 	 */
1618 	switch (bp->bio_cmd) {
1619 	case BIO_READ:
1620 	    {
1621 		struct g_consumer *cp;
1622 		u_char *dst, *src;
1623 		off_t left;
1624 		u_int atom;
1625 
1626 		if (bp->bio_error != 0) {
1627 			G_RAID3_LOGREQ(0, bp,
1628 			    "Synchronization request failed (error=%d).",
1629 			    bp->bio_error);
1630 			g_destroy_bio(bp);
1631 			return;
1632 		}
1633 		G_RAID3_LOGREQ(3, bp, "Synchronization request finished.");
1634 		atom = sc->sc_sectorsize / (sc->sc_ndisks - 1);
1635 		dst = src = bp->bio_data;
1636 		if (disk->d_no == sc->sc_ndisks - 1) {
1637 			u_int n;
1638 
1639 			/* Parity component. */
1640 			for (left = bp->bio_length; left > 0;
1641 			    left -= sc->sc_sectorsize) {
1642 				bcopy(src, dst, atom);
1643 				src += atom;
1644 				for (n = 1; n < sc->sc_ndisks - 1; n++) {
1645 					g_raid3_xor(src, dst, atom);
1646 					src += atom;
1647 				}
1648 				dst += atom;
1649 			}
1650 		} else {
1651 			/* Regular component. */
1652 			src += atom * disk->d_no;
1653 			for (left = bp->bio_length; left > 0;
1654 			    left -= sc->sc_sectorsize) {
1655 				bcopy(src, dst, atom);
1656 				src += sc->sc_sectorsize;
1657 				dst += atom;
1658 			}
1659 		}
1660 		bp->bio_driver1 = bp->bio_driver2 = NULL;
1661 		bp->bio_pflags = 0;
1662 		bp->bio_offset /= sc->sc_ndisks - 1;
1663 		bp->bio_length /= sc->sc_ndisks - 1;
1664 		bp->bio_cmd = BIO_WRITE;
1665 		bp->bio_cflags = 0;
1666 		bp->bio_children = bp->bio_inbed = 0;
1667 		cp = disk->d_consumer;
1668 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1669 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1670 		    cp->acr, cp->acw, cp->ace));
1671 		cp->index++;
1672 		g_io_request(bp, cp);
1673 		return;
1674 	    }
1675 	case BIO_WRITE:
1676 	    {
1677 		struct g_raid3_disk_sync *sync;
1678 		off_t boffset, moffset;
1679 		void *data;
1680 		int i;
1681 
1682 		if (bp->bio_error != 0) {
1683 			G_RAID3_LOGREQ(0, bp,
1684 			    "Synchronization request failed (error=%d).",
1685 			    bp->bio_error);
1686 			g_destroy_bio(bp);
1687 			sc->sc_bump_id |= G_RAID3_BUMP_GENID;
1688 			g_raid3_event_send(disk,
1689 			    G_RAID3_DISK_STATE_DISCONNECTED,
1690 			    G_RAID3_EVENT_DONTWAIT);
1691 			return;
1692 		}
1693 		G_RAID3_LOGREQ(3, bp, "Synchronization request finished.");
1694 		sync = &disk->d_sync;
1695 		if (sync->ds_offset == sc->sc_mediasize / (sc->sc_ndisks - 1) ||
1696 		    sync->ds_consumer == NULL ||
1697 		    (sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
1698 			/* Don't send more synchronization requests. */
1699 			sync->ds_inflight--;
1700 			if (sync->ds_bios != NULL) {
1701 				i = (int)(uintptr_t)bp->bio_caller1;
1702 				sync->ds_bios[i] = NULL;
1703 			}
1704 			free(bp->bio_data, M_RAID3);
1705 			g_destroy_bio(bp);
1706 			if (sync->ds_inflight > 0)
1707 				return;
1708 			if (sync->ds_consumer == NULL ||
1709 			    (sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
1710 				return;
1711 			}
1712 			/*
1713 			 * Disk up-to-date, activate it.
1714 			 */
1715 			g_raid3_event_send(disk, G_RAID3_DISK_STATE_ACTIVE,
1716 			    G_RAID3_EVENT_DONTWAIT);
1717 			return;
1718 		}
1719 
1720 		/* Send next synchronization request. */
1721 		data = bp->bio_data;
1722 		g_reset_bio(bp);
1723 		bp->bio_cmd = BIO_READ;
1724 		bp->bio_offset = sync->ds_offset * (sc->sc_ndisks - 1);
1725 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1726 		sync->ds_offset += bp->bio_length / (sc->sc_ndisks - 1);
1727 		bp->bio_done = g_raid3_sync_done;
1728 		bp->bio_data = data;
1729 		bp->bio_from = sync->ds_consumer;
1730 		bp->bio_to = sc->sc_provider;
1731 		G_RAID3_LOGREQ(3, bp, "Sending synchronization request.");
1732 		sync->ds_consumer->index++;
1733 		/*
1734 		 * Delay the request if it is colliding with a regular request.
1735 		 */
1736 		if (g_raid3_regular_collision(sc, bp))
1737 			g_raid3_sync_delay(sc, bp);
1738 		else
1739 			g_io_request(bp, sync->ds_consumer);
1740 
1741 		/* Release delayed requests if possible. */
1742 		g_raid3_regular_release(sc);
1743 
1744 		/* Find the smallest offset. */
1745 		moffset = sc->sc_mediasize;
1746 		for (i = 0; i < g_raid3_syncreqs; i++) {
1747 			bp = sync->ds_bios[i];
1748 			boffset = bp->bio_offset;
1749 			if (bp->bio_cmd == BIO_WRITE)
1750 				boffset *= sc->sc_ndisks - 1;
1751 			if (boffset < moffset)
1752 				moffset = boffset;
1753 		}
1754 		if (sync->ds_offset_done + (MAXPHYS * 100) < moffset) {
1755 			/* Update offset_done on every 100 blocks. */
1756 			sync->ds_offset_done = moffset;
1757 			g_raid3_update_metadata(disk);
1758 		}
1759 		return;
1760 	    }
1761 	default:
1762 		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1763 		    bp->bio_cmd, sc->sc_name));
1764 		break;
1765 	}
1766 }
1767 
1768 static int
1769 g_raid3_register_request(struct bio *pbp)
1770 {
1771 	struct g_raid3_softc *sc;
1772 	struct g_raid3_disk *disk;
1773 	struct g_consumer *cp;
1774 	struct bio *cbp, *tmpbp;
1775 	off_t offset, length;
1776 	u_int n, ndisks;
1777 	int round_robin, verify;
1778 
1779 	ndisks = 0;
1780 	sc = pbp->bio_to->geom->softc;
1781 	if ((pbp->bio_cflags & G_RAID3_BIO_CFLAG_REGSYNC) != 0 &&
1782 	    sc->sc_syncdisk == NULL) {
1783 		g_io_deliver(pbp, EIO);
1784 		return (0);
1785 	}
1786 	g_raid3_init_bio(pbp);
1787 	length = pbp->bio_length / (sc->sc_ndisks - 1);
1788 	offset = pbp->bio_offset / (sc->sc_ndisks - 1);
1789 	round_robin = verify = 0;
1790 	switch (pbp->bio_cmd) {
1791 	case BIO_READ:
1792 		if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_VERIFY) != 0 &&
1793 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1794 			pbp->bio_pflags |= G_RAID3_BIO_PFLAG_VERIFY;
1795 			verify = 1;
1796 			ndisks = sc->sc_ndisks;
1797 		} else {
1798 			verify = 0;
1799 			ndisks = sc->sc_ndisks - 1;
1800 		}
1801 		if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_ROUND_ROBIN) != 0 &&
1802 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1803 			round_robin = 1;
1804 		} else {
1805 			round_robin = 0;
1806 		}
1807 		KASSERT(!round_robin || !verify,
1808 		    ("ROUND-ROBIN and VERIFY are mutually exclusive."));
1809 		pbp->bio_driver2 = &sc->sc_disks[sc->sc_ndisks - 1];
1810 		break;
1811 	case BIO_WRITE:
1812 	case BIO_DELETE:
1813 		/*
1814 		 * Delay the request if it is colliding with a synchronization
1815 		 * request.
1816 		 */
1817 		if (g_raid3_sync_collision(sc, pbp)) {
1818 			g_raid3_regular_delay(sc, pbp);
1819 			return (0);
1820 		}
1821 
1822 		if (sc->sc_idle)
1823 			g_raid3_unidle(sc);
1824 		else
1825 			sc->sc_last_write = time_uptime;
1826 
1827 		ndisks = sc->sc_ndisks;
1828 		break;
1829 	}
1830 	for (n = 0; n < ndisks; n++) {
1831 		disk = &sc->sc_disks[n];
1832 		cbp = g_raid3_clone_bio(sc, pbp);
1833 		if (cbp == NULL) {
1834 			while ((cbp = G_RAID3_HEAD_BIO(pbp)) != NULL)
1835 				g_raid3_destroy_bio(sc, cbp);
1836 			/*
1837 			 * To prevent deadlock, we must run back up
1838 			 * with the ENOMEM for failed requests of any
1839 			 * of our consumers.  Our own sync requests
1840 			 * can stick around, as they are finite.
1841 			 */
1842 			if ((pbp->bio_cflags &
1843 			    G_RAID3_BIO_CFLAG_REGULAR) != 0) {
1844 				g_io_deliver(pbp, ENOMEM);
1845 				return (0);
1846 			}
1847 			return (ENOMEM);
1848 		}
1849 		cbp->bio_offset = offset;
1850 		cbp->bio_length = length;
1851 		cbp->bio_done = g_raid3_done;
1852 		switch (pbp->bio_cmd) {
1853 		case BIO_READ:
1854 			if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE) {
1855 				/*
1856 				 * Replace invalid component with the parity
1857 				 * component.
1858 				 */
1859 				disk = &sc->sc_disks[sc->sc_ndisks - 1];
1860 				cbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1861 				pbp->bio_pflags |= G_RAID3_BIO_PFLAG_DEGRADED;
1862 			} else if (round_robin &&
1863 			    disk->d_no == sc->sc_round_robin) {
1864 				/*
1865 				 * In round-robin mode skip one data component
1866 				 * and use parity component when reading.
1867 				 */
1868 				pbp->bio_driver2 = disk;
1869 				disk = &sc->sc_disks[sc->sc_ndisks - 1];
1870 				cbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1871 				sc->sc_round_robin++;
1872 				round_robin = 0;
1873 			} else if (verify && disk->d_no == sc->sc_ndisks - 1) {
1874 				cbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1875 			}
1876 			break;
1877 		case BIO_WRITE:
1878 		case BIO_DELETE:
1879 			if (disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
1880 			    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
1881 				if (n == ndisks - 1) {
1882 					/*
1883 					 * Active parity component, mark it as such.
1884 					 */
1885 					cbp->bio_cflags |=
1886 					    G_RAID3_BIO_CFLAG_PARITY;
1887 				}
1888 			} else {
1889 				pbp->bio_pflags |= G_RAID3_BIO_PFLAG_DEGRADED;
1890 				if (n == ndisks - 1) {
1891 					/*
1892 					 * Parity component is not connected,
1893 					 * so destroy its request.
1894 					 */
1895 					pbp->bio_pflags |=
1896 					    G_RAID3_BIO_PFLAG_NOPARITY;
1897 					g_raid3_destroy_bio(sc, cbp);
1898 					cbp = NULL;
1899 				} else {
1900 					cbp->bio_cflags |=
1901 					    G_RAID3_BIO_CFLAG_NODISK;
1902 					disk = NULL;
1903 				}
1904 			}
1905 			break;
1906 		}
1907 		if (cbp != NULL)
1908 			cbp->bio_caller2 = disk;
1909 	}
1910 	switch (pbp->bio_cmd) {
1911 	case BIO_READ:
1912 		if (round_robin) {
1913 			/*
1914 			 * If we are in round-robin mode and 'round_robin' is
1915 			 * still 1, it means, that we skipped parity component
1916 			 * for this read and must reset sc_round_robin field.
1917 			 */
1918 			sc->sc_round_robin = 0;
1919 		}
1920 		G_RAID3_FOREACH_SAFE_BIO(pbp, cbp, tmpbp) {
1921 			disk = cbp->bio_caller2;
1922 			cp = disk->d_consumer;
1923 			cbp->bio_to = cp->provider;
1924 			G_RAID3_LOGREQ(3, cbp, "Sending request.");
1925 			KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1926 			    ("Consumer %s not opened (r%dw%de%d).",
1927 			    cp->provider->name, cp->acr, cp->acw, cp->ace));
1928 			cp->index++;
1929 			g_io_request(cbp, cp);
1930 		}
1931 		break;
1932 	case BIO_WRITE:
1933 	case BIO_DELETE:
1934 		/*
1935 		 * Put request onto inflight queue, so we can check if new
1936 		 * synchronization requests don't collide with it.
1937 		 */
1938 		bioq_insert_tail(&sc->sc_inflight, pbp);
1939 
1940 		/*
1941 		 * Bump syncid on first write.
1942 		 */
1943 		if ((sc->sc_bump_id & G_RAID3_BUMP_SYNCID) != 0) {
1944 			sc->sc_bump_id &= ~G_RAID3_BUMP_SYNCID;
1945 			g_raid3_bump_syncid(sc);
1946 		}
1947 		g_raid3_scatter(pbp);
1948 		break;
1949 	}
1950 	return (0);
1951 }
1952 
1953 static int
1954 g_raid3_can_destroy(struct g_raid3_softc *sc)
1955 {
1956 	struct g_geom *gp;
1957 	struct g_consumer *cp;
1958 
1959 	g_topology_assert();
1960 	gp = sc->sc_geom;
1961 	if (gp->softc == NULL)
1962 		return (1);
1963 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1964 		if (g_raid3_is_busy(sc, cp))
1965 			return (0);
1966 	}
1967 	gp = sc->sc_sync.ds_geom;
1968 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1969 		if (g_raid3_is_busy(sc, cp))
1970 			return (0);
1971 	}
1972 	G_RAID3_DEBUG(2, "No I/O requests for %s, it can be destroyed.",
1973 	    sc->sc_name);
1974 	return (1);
1975 }
1976 
1977 static int
1978 g_raid3_try_destroy(struct g_raid3_softc *sc)
1979 {
1980 
1981 	g_topology_assert_not();
1982 	sx_assert(&sc->sc_lock, SX_XLOCKED);
1983 
1984 	if (sc->sc_rootmount != NULL) {
1985 		G_RAID3_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
1986 		    sc->sc_rootmount);
1987 		root_mount_rel(sc->sc_rootmount);
1988 		sc->sc_rootmount = NULL;
1989 	}
1990 
1991 	g_topology_lock();
1992 	if (!g_raid3_can_destroy(sc)) {
1993 		g_topology_unlock();
1994 		return (0);
1995 	}
1996 	sc->sc_geom->softc = NULL;
1997 	sc->sc_sync.ds_geom->softc = NULL;
1998 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_WAIT) != 0) {
1999 		g_topology_unlock();
2000 		G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__,
2001 		    &sc->sc_worker);
2002 		/* Unlock sc_lock here, as it can be destroyed after wakeup. */
2003 		sx_xunlock(&sc->sc_lock);
2004 		wakeup(&sc->sc_worker);
2005 		sc->sc_worker = NULL;
2006 	} else {
2007 		g_topology_unlock();
2008 		g_raid3_destroy_device(sc);
2009 		free(sc->sc_disks, M_RAID3);
2010 		free(sc, M_RAID3);
2011 	}
2012 	return (1);
2013 }
2014 
2015 /*
2016  * Worker thread.
2017  */
2018 static void
2019 g_raid3_worker(void *arg)
2020 {
2021 	struct g_raid3_softc *sc;
2022 	struct g_raid3_event *ep;
2023 	struct bio *bp;
2024 	int timeout;
2025 
2026 	sc = arg;
2027 	thread_lock(curthread);
2028 	sched_prio(curthread, PRIBIO);
2029 	thread_unlock(curthread);
2030 
2031 	sx_xlock(&sc->sc_lock);
2032 	for (;;) {
2033 		G_RAID3_DEBUG(5, "%s: Let's see...", __func__);
2034 		/*
2035 		 * First take a look at events.
2036 		 * This is important to handle events before any I/O requests.
2037 		 */
2038 		ep = g_raid3_event_get(sc);
2039 		if (ep != NULL) {
2040 			g_raid3_event_remove(sc, ep);
2041 			if ((ep->e_flags & G_RAID3_EVENT_DEVICE) != 0) {
2042 				/* Update only device status. */
2043 				G_RAID3_DEBUG(3,
2044 				    "Running event for device %s.",
2045 				    sc->sc_name);
2046 				ep->e_error = 0;
2047 				g_raid3_update_device(sc, 1);
2048 			} else {
2049 				/* Update disk status. */
2050 				G_RAID3_DEBUG(3, "Running event for disk %s.",
2051 				     g_raid3_get_diskname(ep->e_disk));
2052 				ep->e_error = g_raid3_update_disk(ep->e_disk,
2053 				    ep->e_state);
2054 				if (ep->e_error == 0)
2055 					g_raid3_update_device(sc, 0);
2056 			}
2057 			if ((ep->e_flags & G_RAID3_EVENT_DONTWAIT) != 0) {
2058 				KASSERT(ep->e_error == 0,
2059 				    ("Error cannot be handled."));
2060 				g_raid3_event_free(ep);
2061 			} else {
2062 				ep->e_flags |= G_RAID3_EVENT_DONE;
2063 				G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__,
2064 				    ep);
2065 				mtx_lock(&sc->sc_events_mtx);
2066 				wakeup(ep);
2067 				mtx_unlock(&sc->sc_events_mtx);
2068 			}
2069 			if ((sc->sc_flags &
2070 			    G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
2071 				if (g_raid3_try_destroy(sc)) {
2072 					curthread->td_pflags &= ~TDP_GEOM;
2073 					G_RAID3_DEBUG(1, "Thread exiting.");
2074 					kproc_exit(0);
2075 				}
2076 			}
2077 			G_RAID3_DEBUG(5, "%s: I'm here 1.", __func__);
2078 			continue;
2079 		}
2080 		/*
2081 		 * Check if we can mark array as CLEAN and if we can't take
2082 		 * how much seconds should we wait.
2083 		 */
2084 		timeout = g_raid3_idle(sc, -1);
2085 		/*
2086 		 * Now I/O requests.
2087 		 */
2088 		/* Get first request from the queue. */
2089 		mtx_lock(&sc->sc_queue_mtx);
2090 		bp = bioq_first(&sc->sc_queue);
2091 		if (bp == NULL) {
2092 			if ((sc->sc_flags &
2093 			    G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
2094 				mtx_unlock(&sc->sc_queue_mtx);
2095 				if (g_raid3_try_destroy(sc)) {
2096 					curthread->td_pflags &= ~TDP_GEOM;
2097 					G_RAID3_DEBUG(1, "Thread exiting.");
2098 					kproc_exit(0);
2099 				}
2100 				mtx_lock(&sc->sc_queue_mtx);
2101 			}
2102 			sx_xunlock(&sc->sc_lock);
2103 			/*
2104 			 * XXX: We can miss an event here, because an event
2105 			 *      can be added without sx-device-lock and without
2106 			 *      mtx-queue-lock. Maybe I should just stop using
2107 			 *      dedicated mutex for events synchronization and
2108 			 *      stick with the queue lock?
2109 			 *      The event will hang here until next I/O request
2110 			 *      or next event is received.
2111 			 */
2112 			MSLEEP(sc, &sc->sc_queue_mtx, PRIBIO | PDROP, "r3:w1",
2113 			    timeout * hz);
2114 			sx_xlock(&sc->sc_lock);
2115 			G_RAID3_DEBUG(5, "%s: I'm here 4.", __func__);
2116 			continue;
2117 		}
2118 process:
2119 		bioq_remove(&sc->sc_queue, bp);
2120 		mtx_unlock(&sc->sc_queue_mtx);
2121 
2122 		if (bp->bio_from->geom == sc->sc_sync.ds_geom &&
2123 		    (bp->bio_cflags & G_RAID3_BIO_CFLAG_SYNC) != 0) {
2124 			g_raid3_sync_request(bp);	/* READ */
2125 		} else if (bp->bio_to != sc->sc_provider) {
2126 			if ((bp->bio_cflags & G_RAID3_BIO_CFLAG_REGULAR) != 0)
2127 				g_raid3_regular_request(bp);
2128 			else if ((bp->bio_cflags & G_RAID3_BIO_CFLAG_SYNC) != 0)
2129 				g_raid3_sync_request(bp);	/* WRITE */
2130 			else {
2131 				KASSERT(0,
2132 				    ("Invalid request cflags=0x%hx to=%s.",
2133 				    bp->bio_cflags, bp->bio_to->name));
2134 			}
2135 		} else if (g_raid3_register_request(bp) != 0) {
2136 			mtx_lock(&sc->sc_queue_mtx);
2137 			bioq_insert_head(&sc->sc_queue, bp);
2138 			/*
2139 			 * We are short in memory, let see if there are finished
2140 			 * request we can free.
2141 			 */
2142 			TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
2143 				if (bp->bio_cflags & G_RAID3_BIO_CFLAG_REGULAR)
2144 					goto process;
2145 			}
2146 			/*
2147 			 * No finished regular request, so at least keep
2148 			 * synchronization running.
2149 			 */
2150 			TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
2151 				if (bp->bio_cflags & G_RAID3_BIO_CFLAG_SYNC)
2152 					goto process;
2153 			}
2154 			sx_xunlock(&sc->sc_lock);
2155 			MSLEEP(&sc->sc_queue, &sc->sc_queue_mtx, PRIBIO | PDROP,
2156 			    "r3:lowmem", hz / 10);
2157 			sx_xlock(&sc->sc_lock);
2158 		}
2159 		G_RAID3_DEBUG(5, "%s: I'm here 9.", __func__);
2160 	}
2161 }
2162 
2163 static void
2164 g_raid3_update_idle(struct g_raid3_softc *sc, struct g_raid3_disk *disk)
2165 {
2166 
2167 	sx_assert(&sc->sc_lock, SX_LOCKED);
2168 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) != 0)
2169 		return;
2170 	if (!sc->sc_idle && (disk->d_flags & G_RAID3_DISK_FLAG_DIRTY) == 0) {
2171 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as dirty.",
2172 		    g_raid3_get_diskname(disk), sc->sc_name);
2173 		disk->d_flags |= G_RAID3_DISK_FLAG_DIRTY;
2174 	} else if (sc->sc_idle &&
2175 	    (disk->d_flags & G_RAID3_DISK_FLAG_DIRTY) != 0) {
2176 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as clean.",
2177 		    g_raid3_get_diskname(disk), sc->sc_name);
2178 		disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2179 	}
2180 }
2181 
2182 static void
2183 g_raid3_sync_start(struct g_raid3_softc *sc)
2184 {
2185 	struct g_raid3_disk *disk;
2186 	struct g_consumer *cp;
2187 	struct bio *bp;
2188 	int error;
2189 	u_int n;
2190 
2191 	g_topology_assert_not();
2192 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2193 
2194 	KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED,
2195 	    ("Device not in DEGRADED state (%s, %u).", sc->sc_name,
2196 	    sc->sc_state));
2197 	KASSERT(sc->sc_syncdisk == NULL, ("Syncdisk is not NULL (%s, %u).",
2198 	    sc->sc_name, sc->sc_state));
2199 	disk = NULL;
2200 	for (n = 0; n < sc->sc_ndisks; n++) {
2201 		if (sc->sc_disks[n].d_state != G_RAID3_DISK_STATE_SYNCHRONIZING)
2202 			continue;
2203 		disk = &sc->sc_disks[n];
2204 		break;
2205 	}
2206 	if (disk == NULL)
2207 		return;
2208 
2209 	sx_xunlock(&sc->sc_lock);
2210 	g_topology_lock();
2211 	cp = g_new_consumer(sc->sc_sync.ds_geom);
2212 	error = g_attach(cp, sc->sc_provider);
2213 	KASSERT(error == 0,
2214 	    ("Cannot attach to %s (error=%d).", sc->sc_name, error));
2215 	error = g_access(cp, 1, 0, 0);
2216 	KASSERT(error == 0, ("Cannot open %s (error=%d).", sc->sc_name, error));
2217 	g_topology_unlock();
2218 	sx_xlock(&sc->sc_lock);
2219 
2220 	G_RAID3_DEBUG(0, "Device %s: rebuilding provider %s.", sc->sc_name,
2221 	    g_raid3_get_diskname(disk));
2222 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) == 0)
2223 		disk->d_flags |= G_RAID3_DISK_FLAG_DIRTY;
2224 	KASSERT(disk->d_sync.ds_consumer == NULL,
2225 	    ("Sync consumer already exists (device=%s, disk=%s).",
2226 	    sc->sc_name, g_raid3_get_diskname(disk)));
2227 
2228 	disk->d_sync.ds_consumer = cp;
2229 	disk->d_sync.ds_consumer->private = disk;
2230 	disk->d_sync.ds_consumer->index = 0;
2231 	sc->sc_syncdisk = disk;
2232 
2233 	/*
2234 	 * Allocate memory for synchronization bios and initialize them.
2235 	 */
2236 	disk->d_sync.ds_bios = malloc(sizeof(struct bio *) * g_raid3_syncreqs,
2237 	    M_RAID3, M_WAITOK);
2238 	for (n = 0; n < g_raid3_syncreqs; n++) {
2239 		bp = g_alloc_bio();
2240 		disk->d_sync.ds_bios[n] = bp;
2241 		bp->bio_parent = NULL;
2242 		bp->bio_cmd = BIO_READ;
2243 		bp->bio_data = malloc(MAXPHYS, M_RAID3, M_WAITOK);
2244 		bp->bio_cflags = 0;
2245 		bp->bio_offset = disk->d_sync.ds_offset * (sc->sc_ndisks - 1);
2246 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
2247 		disk->d_sync.ds_offset += bp->bio_length / (sc->sc_ndisks - 1);
2248 		bp->bio_done = g_raid3_sync_done;
2249 		bp->bio_from = disk->d_sync.ds_consumer;
2250 		bp->bio_to = sc->sc_provider;
2251 		bp->bio_caller1 = (void *)(uintptr_t)n;
2252 	}
2253 
2254 	/* Set the number of in-flight synchronization requests. */
2255 	disk->d_sync.ds_inflight = g_raid3_syncreqs;
2256 
2257 	/*
2258 	 * Fire off first synchronization requests.
2259 	 */
2260 	for (n = 0; n < g_raid3_syncreqs; n++) {
2261 		bp = disk->d_sync.ds_bios[n];
2262 		G_RAID3_LOGREQ(3, bp, "Sending synchronization request.");
2263 		disk->d_sync.ds_consumer->index++;
2264 		/*
2265 		 * Delay the request if it is colliding with a regular request.
2266 		 */
2267 		if (g_raid3_regular_collision(sc, bp))
2268 			g_raid3_sync_delay(sc, bp);
2269 		else
2270 			g_io_request(bp, disk->d_sync.ds_consumer);
2271 	}
2272 }
2273 
2274 /*
2275  * Stop synchronization process.
2276  * type: 0 - synchronization finished
2277  *       1 - synchronization stopped
2278  */
2279 static void
2280 g_raid3_sync_stop(struct g_raid3_softc *sc, int type)
2281 {
2282 	struct g_raid3_disk *disk;
2283 	struct g_consumer *cp;
2284 
2285 	g_topology_assert_not();
2286 	sx_assert(&sc->sc_lock, SX_LOCKED);
2287 
2288 	KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED,
2289 	    ("Device not in DEGRADED state (%s, %u).", sc->sc_name,
2290 	    sc->sc_state));
2291 	disk = sc->sc_syncdisk;
2292 	sc->sc_syncdisk = NULL;
2293 	KASSERT(disk != NULL, ("No disk was synchronized (%s).", sc->sc_name));
2294 	KASSERT(disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING,
2295 	    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2296 	    g_raid3_disk_state2str(disk->d_state)));
2297 	if (disk->d_sync.ds_consumer == NULL)
2298 		return;
2299 
2300 	if (type == 0) {
2301 		G_RAID3_DEBUG(0, "Device %s: rebuilding provider %s finished.",
2302 		    sc->sc_name, g_raid3_get_diskname(disk));
2303 	} else /* if (type == 1) */ {
2304 		G_RAID3_DEBUG(0, "Device %s: rebuilding provider %s stopped.",
2305 		    sc->sc_name, g_raid3_get_diskname(disk));
2306 	}
2307 	free(disk->d_sync.ds_bios, M_RAID3);
2308 	disk->d_sync.ds_bios = NULL;
2309 	cp = disk->d_sync.ds_consumer;
2310 	disk->d_sync.ds_consumer = NULL;
2311 	disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2312 	sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
2313 	g_topology_lock();
2314 	g_raid3_kill_consumer(sc, cp);
2315 	g_topology_unlock();
2316 	sx_xlock(&sc->sc_lock);
2317 }
2318 
2319 static void
2320 g_raid3_launch_provider(struct g_raid3_softc *sc)
2321 {
2322 	struct g_provider *pp;
2323 	struct g_raid3_disk *disk;
2324 	int n;
2325 
2326 	sx_assert(&sc->sc_lock, SX_LOCKED);
2327 
2328 	g_topology_lock();
2329 	pp = g_new_providerf(sc->sc_geom, "raid3/%s", sc->sc_name);
2330 	pp->mediasize = sc->sc_mediasize;
2331 	pp->sectorsize = sc->sc_sectorsize;
2332 	pp->stripesize = 0;
2333 	pp->stripeoffset = 0;
2334 	for (n = 0; n < sc->sc_ndisks; n++) {
2335 		disk = &sc->sc_disks[n];
2336 		if (disk->d_consumer && disk->d_consumer->provider &&
2337 		    disk->d_consumer->provider->stripesize > pp->stripesize) {
2338 			pp->stripesize = disk->d_consumer->provider->stripesize;
2339 			pp->stripeoffset = disk->d_consumer->provider->stripeoffset;
2340 		}
2341 	}
2342 	pp->stripesize *= sc->sc_ndisks - 1;
2343 	pp->stripeoffset *= sc->sc_ndisks - 1;
2344 	sc->sc_provider = pp;
2345 	g_error_provider(pp, 0);
2346 	g_topology_unlock();
2347 	G_RAID3_DEBUG(0, "Device %s launched (%u/%u).", pp->name,
2348 	    g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE), sc->sc_ndisks);
2349 
2350 	if (sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED)
2351 		g_raid3_sync_start(sc);
2352 }
2353 
2354 static void
2355 g_raid3_destroy_provider(struct g_raid3_softc *sc)
2356 {
2357 	struct bio *bp;
2358 
2359 	g_topology_assert_not();
2360 	KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
2361 	    sc->sc_name));
2362 
2363 	g_topology_lock();
2364 	g_error_provider(sc->sc_provider, ENXIO);
2365 	mtx_lock(&sc->sc_queue_mtx);
2366 	while ((bp = bioq_first(&sc->sc_queue)) != NULL) {
2367 		bioq_remove(&sc->sc_queue, bp);
2368 		g_io_deliver(bp, ENXIO);
2369 	}
2370 	mtx_unlock(&sc->sc_queue_mtx);
2371 	G_RAID3_DEBUG(0, "Device %s: provider %s destroyed.", sc->sc_name,
2372 	    sc->sc_provider->name);
2373 	g_wither_provider(sc->sc_provider, ENXIO);
2374 	g_topology_unlock();
2375 	sc->sc_provider = NULL;
2376 	if (sc->sc_syncdisk != NULL)
2377 		g_raid3_sync_stop(sc, 1);
2378 }
2379 
2380 static void
2381 g_raid3_go(void *arg)
2382 {
2383 	struct g_raid3_softc *sc;
2384 
2385 	sc = arg;
2386 	G_RAID3_DEBUG(0, "Force device %s start due to timeout.", sc->sc_name);
2387 	g_raid3_event_send(sc, 0,
2388 	    G_RAID3_EVENT_DONTWAIT | G_RAID3_EVENT_DEVICE);
2389 }
2390 
2391 static u_int
2392 g_raid3_determine_state(struct g_raid3_disk *disk)
2393 {
2394 	struct g_raid3_softc *sc;
2395 	u_int state;
2396 
2397 	sc = disk->d_softc;
2398 	if (sc->sc_syncid == disk->d_sync.ds_syncid) {
2399 		if ((disk->d_flags &
2400 		    G_RAID3_DISK_FLAG_SYNCHRONIZING) == 0) {
2401 			/* Disk does not need synchronization. */
2402 			state = G_RAID3_DISK_STATE_ACTIVE;
2403 		} else {
2404 			if ((sc->sc_flags &
2405 			     G_RAID3_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2406 			    (disk->d_flags &
2407 			     G_RAID3_DISK_FLAG_FORCE_SYNC) != 0) {
2408 				/*
2409 				 * We can start synchronization from
2410 				 * the stored offset.
2411 				 */
2412 				state = G_RAID3_DISK_STATE_SYNCHRONIZING;
2413 			} else {
2414 				state = G_RAID3_DISK_STATE_STALE;
2415 			}
2416 		}
2417 	} else if (disk->d_sync.ds_syncid < sc->sc_syncid) {
2418 		/*
2419 		 * Reset all synchronization data for this disk,
2420 		 * because if it even was synchronized, it was
2421 		 * synchronized to disks with different syncid.
2422 		 */
2423 		disk->d_flags |= G_RAID3_DISK_FLAG_SYNCHRONIZING;
2424 		disk->d_sync.ds_offset = 0;
2425 		disk->d_sync.ds_offset_done = 0;
2426 		disk->d_sync.ds_syncid = sc->sc_syncid;
2427 		if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2428 		    (disk->d_flags & G_RAID3_DISK_FLAG_FORCE_SYNC) != 0) {
2429 			state = G_RAID3_DISK_STATE_SYNCHRONIZING;
2430 		} else {
2431 			state = G_RAID3_DISK_STATE_STALE;
2432 		}
2433 	} else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
2434 		/*
2435 		 * Not good, NOT GOOD!
2436 		 * It means that device was started on stale disks
2437 		 * and more fresh disk just arrive.
2438 		 * If there were writes, device is broken, sorry.
2439 		 * I think the best choice here is don't touch
2440 		 * this disk and inform the user loudly.
2441 		 */
2442 		G_RAID3_DEBUG(0, "Device %s was started before the freshest "
2443 		    "disk (%s) arrives!! It will not be connected to the "
2444 		    "running device.", sc->sc_name,
2445 		    g_raid3_get_diskname(disk));
2446 		g_raid3_destroy_disk(disk);
2447 		state = G_RAID3_DISK_STATE_NONE;
2448 		/* Return immediately, because disk was destroyed. */
2449 		return (state);
2450 	}
2451 	G_RAID3_DEBUG(3, "State for %s disk: %s.",
2452 	    g_raid3_get_diskname(disk), g_raid3_disk_state2str(state));
2453 	return (state);
2454 }
2455 
2456 /*
2457  * Update device state.
2458  */
2459 static void
2460 g_raid3_update_device(struct g_raid3_softc *sc, boolean_t force)
2461 {
2462 	struct g_raid3_disk *disk;
2463 	u_int state;
2464 
2465 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2466 
2467 	switch (sc->sc_state) {
2468 	case G_RAID3_DEVICE_STATE_STARTING:
2469 	    {
2470 		u_int n, ndirty, ndisks, genid, syncid;
2471 
2472 		KASSERT(sc->sc_provider == NULL,
2473 		    ("Non-NULL provider in STARTING state (%s).", sc->sc_name));
2474 		/*
2475 		 * Are we ready? We are, if all disks are connected or
2476 		 * one disk is missing and 'force' is true.
2477 		 */
2478 		if (g_raid3_ndisks(sc, -1) + force == sc->sc_ndisks) {
2479 			if (!force)
2480 				callout_drain(&sc->sc_callout);
2481 		} else {
2482 			if (force) {
2483 				/*
2484 				 * Timeout expired, so destroy device.
2485 				 */
2486 				sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
2487 				G_RAID3_DEBUG(1, "root_mount_rel[%u] %p",
2488 				    __LINE__, sc->sc_rootmount);
2489 				root_mount_rel(sc->sc_rootmount);
2490 				sc->sc_rootmount = NULL;
2491 			}
2492 			return;
2493 		}
2494 
2495 		/*
2496 		 * Find the biggest genid.
2497 		 */
2498 		genid = 0;
2499 		for (n = 0; n < sc->sc_ndisks; n++) {
2500 			disk = &sc->sc_disks[n];
2501 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2502 				continue;
2503 			if (disk->d_genid > genid)
2504 				genid = disk->d_genid;
2505 		}
2506 		sc->sc_genid = genid;
2507 		/*
2508 		 * Remove all disks without the biggest genid.
2509 		 */
2510 		for (n = 0; n < sc->sc_ndisks; n++) {
2511 			disk = &sc->sc_disks[n];
2512 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2513 				continue;
2514 			if (disk->d_genid < genid) {
2515 				G_RAID3_DEBUG(0,
2516 				    "Component %s (device %s) broken, skipping.",
2517 				    g_raid3_get_diskname(disk), sc->sc_name);
2518 				g_raid3_destroy_disk(disk);
2519 			}
2520 		}
2521 
2522 		/*
2523 		 * There must be at least 'sc->sc_ndisks - 1' components
2524 		 * with the same syncid and without SYNCHRONIZING flag.
2525 		 */
2526 
2527 		/*
2528 		 * Find the biggest syncid, number of valid components and
2529 		 * number of dirty components.
2530 		 */
2531 		ndirty = ndisks = syncid = 0;
2532 		for (n = 0; n < sc->sc_ndisks; n++) {
2533 			disk = &sc->sc_disks[n];
2534 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2535 				continue;
2536 			if ((disk->d_flags & G_RAID3_DISK_FLAG_DIRTY) != 0)
2537 				ndirty++;
2538 			if (disk->d_sync.ds_syncid > syncid) {
2539 				syncid = disk->d_sync.ds_syncid;
2540 				ndisks = 0;
2541 			} else if (disk->d_sync.ds_syncid < syncid) {
2542 				continue;
2543 			}
2544 			if ((disk->d_flags &
2545 			    G_RAID3_DISK_FLAG_SYNCHRONIZING) != 0) {
2546 				continue;
2547 			}
2548 			ndisks++;
2549 		}
2550 		/*
2551 		 * Do we have enough valid components?
2552 		 */
2553 		if (ndisks + 1 < sc->sc_ndisks) {
2554 			G_RAID3_DEBUG(0,
2555 			    "Device %s is broken, too few valid components.",
2556 			    sc->sc_name);
2557 			sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
2558 			return;
2559 		}
2560 		/*
2561 		 * If there is one DIRTY component and all disks are present,
2562 		 * mark it for synchronization. If there is more than one DIRTY
2563 		 * component, mark parity component for synchronization.
2564 		 */
2565 		if (ndisks == sc->sc_ndisks && ndirty == 1) {
2566 			for (n = 0; n < sc->sc_ndisks; n++) {
2567 				disk = &sc->sc_disks[n];
2568 				if ((disk->d_flags &
2569 				    G_RAID3_DISK_FLAG_DIRTY) == 0) {
2570 					continue;
2571 				}
2572 				disk->d_flags |=
2573 				    G_RAID3_DISK_FLAG_SYNCHRONIZING;
2574 			}
2575 		} else if (ndisks == sc->sc_ndisks && ndirty > 1) {
2576 			disk = &sc->sc_disks[sc->sc_ndisks - 1];
2577 			disk->d_flags |= G_RAID3_DISK_FLAG_SYNCHRONIZING;
2578 		}
2579 
2580 		sc->sc_syncid = syncid;
2581 		if (force) {
2582 			/* Remember to bump syncid on first write. */
2583 			sc->sc_bump_id |= G_RAID3_BUMP_SYNCID;
2584 		}
2585 		if (ndisks == sc->sc_ndisks)
2586 			state = G_RAID3_DEVICE_STATE_COMPLETE;
2587 		else /* if (ndisks == sc->sc_ndisks - 1) */
2588 			state = G_RAID3_DEVICE_STATE_DEGRADED;
2589 		G_RAID3_DEBUG(1, "Device %s state changed from %s to %s.",
2590 		    sc->sc_name, g_raid3_device_state2str(sc->sc_state),
2591 		    g_raid3_device_state2str(state));
2592 		sc->sc_state = state;
2593 		for (n = 0; n < sc->sc_ndisks; n++) {
2594 			disk = &sc->sc_disks[n];
2595 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2596 				continue;
2597 			state = g_raid3_determine_state(disk);
2598 			g_raid3_event_send(disk, state, G_RAID3_EVENT_DONTWAIT);
2599 			if (state == G_RAID3_DISK_STATE_STALE)
2600 				sc->sc_bump_id |= G_RAID3_BUMP_SYNCID;
2601 		}
2602 		break;
2603 	    }
2604 	case G_RAID3_DEVICE_STATE_DEGRADED:
2605 		/*
2606 		 * Genid need to be bumped immediately, so do it here.
2607 		 */
2608 		if ((sc->sc_bump_id & G_RAID3_BUMP_GENID) != 0) {
2609 			sc->sc_bump_id &= ~G_RAID3_BUMP_GENID;
2610 			g_raid3_bump_genid(sc);
2611 		}
2612 
2613 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_NEW) > 0)
2614 			return;
2615 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) <
2616 		    sc->sc_ndisks - 1) {
2617 			if (sc->sc_provider != NULL)
2618 				g_raid3_destroy_provider(sc);
2619 			sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
2620 			return;
2621 		}
2622 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) ==
2623 		    sc->sc_ndisks) {
2624 			state = G_RAID3_DEVICE_STATE_COMPLETE;
2625 			G_RAID3_DEBUG(1,
2626 			    "Device %s state changed from %s to %s.",
2627 			    sc->sc_name, g_raid3_device_state2str(sc->sc_state),
2628 			    g_raid3_device_state2str(state));
2629 			sc->sc_state = state;
2630 		}
2631 		if (sc->sc_provider == NULL)
2632 			g_raid3_launch_provider(sc);
2633 		if (sc->sc_rootmount != NULL) {
2634 			G_RAID3_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2635 			    sc->sc_rootmount);
2636 			root_mount_rel(sc->sc_rootmount);
2637 			sc->sc_rootmount = NULL;
2638 		}
2639 		break;
2640 	case G_RAID3_DEVICE_STATE_COMPLETE:
2641 		/*
2642 		 * Genid need to be bumped immediately, so do it here.
2643 		 */
2644 		if ((sc->sc_bump_id & G_RAID3_BUMP_GENID) != 0) {
2645 			sc->sc_bump_id &= ~G_RAID3_BUMP_GENID;
2646 			g_raid3_bump_genid(sc);
2647 		}
2648 
2649 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_NEW) > 0)
2650 			return;
2651 		KASSERT(g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) >=
2652 		    sc->sc_ndisks - 1,
2653 		    ("Too few ACTIVE components in COMPLETE state (device %s).",
2654 		    sc->sc_name));
2655 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) ==
2656 		    sc->sc_ndisks - 1) {
2657 			state = G_RAID3_DEVICE_STATE_DEGRADED;
2658 			G_RAID3_DEBUG(1,
2659 			    "Device %s state changed from %s to %s.",
2660 			    sc->sc_name, g_raid3_device_state2str(sc->sc_state),
2661 			    g_raid3_device_state2str(state));
2662 			sc->sc_state = state;
2663 		}
2664 		if (sc->sc_provider == NULL)
2665 			g_raid3_launch_provider(sc);
2666 		if (sc->sc_rootmount != NULL) {
2667 			G_RAID3_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2668 			    sc->sc_rootmount);
2669 			root_mount_rel(sc->sc_rootmount);
2670 			sc->sc_rootmount = NULL;
2671 		}
2672 		break;
2673 	default:
2674 		KASSERT(1 == 0, ("Wrong device state (%s, %s).", sc->sc_name,
2675 		    g_raid3_device_state2str(sc->sc_state)));
2676 		break;
2677 	}
2678 }
2679 
2680 /*
2681  * Update disk state and device state if needed.
2682  */
2683 #define	DISK_STATE_CHANGED()	G_RAID3_DEBUG(1,			\
2684 	"Disk %s state changed from %s to %s (device %s).",		\
2685 	g_raid3_get_diskname(disk),					\
2686 	g_raid3_disk_state2str(disk->d_state),				\
2687 	g_raid3_disk_state2str(state), sc->sc_name)
2688 static int
2689 g_raid3_update_disk(struct g_raid3_disk *disk, u_int state)
2690 {
2691 	struct g_raid3_softc *sc;
2692 
2693 	sc = disk->d_softc;
2694 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2695 
2696 again:
2697 	G_RAID3_DEBUG(3, "Changing disk %s state from %s to %s.",
2698 	    g_raid3_get_diskname(disk), g_raid3_disk_state2str(disk->d_state),
2699 	    g_raid3_disk_state2str(state));
2700 	switch (state) {
2701 	case G_RAID3_DISK_STATE_NEW:
2702 		/*
2703 		 * Possible scenarios:
2704 		 * 1. New disk arrive.
2705 		 */
2706 		/* Previous state should be NONE. */
2707 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NONE,
2708 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2709 		    g_raid3_disk_state2str(disk->d_state)));
2710 		DISK_STATE_CHANGED();
2711 
2712 		disk->d_state = state;
2713 		G_RAID3_DEBUG(1, "Device %s: provider %s detected.",
2714 		    sc->sc_name, g_raid3_get_diskname(disk));
2715 		if (sc->sc_state == G_RAID3_DEVICE_STATE_STARTING)
2716 			break;
2717 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2718 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2719 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2720 		    g_raid3_device_state2str(sc->sc_state),
2721 		    g_raid3_get_diskname(disk),
2722 		    g_raid3_disk_state2str(disk->d_state)));
2723 		state = g_raid3_determine_state(disk);
2724 		if (state != G_RAID3_DISK_STATE_NONE)
2725 			goto again;
2726 		break;
2727 	case G_RAID3_DISK_STATE_ACTIVE:
2728 		/*
2729 		 * Possible scenarios:
2730 		 * 1. New disk does not need synchronization.
2731 		 * 2. Synchronization process finished successfully.
2732 		 */
2733 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2734 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2735 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2736 		    g_raid3_device_state2str(sc->sc_state),
2737 		    g_raid3_get_diskname(disk),
2738 		    g_raid3_disk_state2str(disk->d_state)));
2739 		/* Previous state should be NEW or SYNCHRONIZING. */
2740 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW ||
2741 		    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING,
2742 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2743 		    g_raid3_disk_state2str(disk->d_state)));
2744 		DISK_STATE_CHANGED();
2745 
2746 		if (disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
2747 			disk->d_flags &= ~G_RAID3_DISK_FLAG_SYNCHRONIZING;
2748 			disk->d_flags &= ~G_RAID3_DISK_FLAG_FORCE_SYNC;
2749 			g_raid3_sync_stop(sc, 0);
2750 		}
2751 		disk->d_state = state;
2752 		disk->d_sync.ds_offset = 0;
2753 		disk->d_sync.ds_offset_done = 0;
2754 		g_raid3_update_idle(sc, disk);
2755 		g_raid3_update_metadata(disk);
2756 		G_RAID3_DEBUG(1, "Device %s: provider %s activated.",
2757 		    sc->sc_name, g_raid3_get_diskname(disk));
2758 		break;
2759 	case G_RAID3_DISK_STATE_STALE:
2760 		/*
2761 		 * Possible scenarios:
2762 		 * 1. Stale disk was connected.
2763 		 */
2764 		/* Previous state should be NEW. */
2765 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW,
2766 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2767 		    g_raid3_disk_state2str(disk->d_state)));
2768 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2769 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2770 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2771 		    g_raid3_device_state2str(sc->sc_state),
2772 		    g_raid3_get_diskname(disk),
2773 		    g_raid3_disk_state2str(disk->d_state)));
2774 		/*
2775 		 * STALE state is only possible if device is marked
2776 		 * NOAUTOSYNC.
2777 		 */
2778 		KASSERT((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOAUTOSYNC) != 0,
2779 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2780 		    g_raid3_device_state2str(sc->sc_state),
2781 		    g_raid3_get_diskname(disk),
2782 		    g_raid3_disk_state2str(disk->d_state)));
2783 		DISK_STATE_CHANGED();
2784 
2785 		disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2786 		disk->d_state = state;
2787 		g_raid3_update_metadata(disk);
2788 		G_RAID3_DEBUG(0, "Device %s: provider %s is stale.",
2789 		    sc->sc_name, g_raid3_get_diskname(disk));
2790 		break;
2791 	case G_RAID3_DISK_STATE_SYNCHRONIZING:
2792 		/*
2793 		 * Possible scenarios:
2794 		 * 1. Disk which needs synchronization was connected.
2795 		 */
2796 		/* Previous state should be NEW. */
2797 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW,
2798 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2799 		    g_raid3_disk_state2str(disk->d_state)));
2800 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2801 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2802 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2803 		    g_raid3_device_state2str(sc->sc_state),
2804 		    g_raid3_get_diskname(disk),
2805 		    g_raid3_disk_state2str(disk->d_state)));
2806 		DISK_STATE_CHANGED();
2807 
2808 		if (disk->d_state == G_RAID3_DISK_STATE_NEW)
2809 			disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2810 		disk->d_state = state;
2811 		if (sc->sc_provider != NULL) {
2812 			g_raid3_sync_start(sc);
2813 			g_raid3_update_metadata(disk);
2814 		}
2815 		break;
2816 	case G_RAID3_DISK_STATE_DISCONNECTED:
2817 		/*
2818 		 * Possible scenarios:
2819 		 * 1. Device wasn't running yet, but disk disappear.
2820 		 * 2. Disk was active and disapppear.
2821 		 * 3. Disk disappear during synchronization process.
2822 		 */
2823 		if (sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2824 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
2825 			/*
2826 			 * Previous state should be ACTIVE, STALE or
2827 			 * SYNCHRONIZING.
2828 			 */
2829 			KASSERT(disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
2830 			    disk->d_state == G_RAID3_DISK_STATE_STALE ||
2831 			    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING,
2832 			    ("Wrong disk state (%s, %s).",
2833 			    g_raid3_get_diskname(disk),
2834 			    g_raid3_disk_state2str(disk->d_state)));
2835 		} else if (sc->sc_state == G_RAID3_DEVICE_STATE_STARTING) {
2836 			/* Previous state should be NEW. */
2837 			KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW,
2838 			    ("Wrong disk state (%s, %s).",
2839 			    g_raid3_get_diskname(disk),
2840 			    g_raid3_disk_state2str(disk->d_state)));
2841 			/*
2842 			 * Reset bumping syncid if disk disappeared in STARTING
2843 			 * state.
2844 			 */
2845 			if ((sc->sc_bump_id & G_RAID3_BUMP_SYNCID) != 0)
2846 				sc->sc_bump_id &= ~G_RAID3_BUMP_SYNCID;
2847 #ifdef	INVARIANTS
2848 		} else {
2849 			KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
2850 			    sc->sc_name,
2851 			    g_raid3_device_state2str(sc->sc_state),
2852 			    g_raid3_get_diskname(disk),
2853 			    g_raid3_disk_state2str(disk->d_state)));
2854 #endif
2855 		}
2856 		DISK_STATE_CHANGED();
2857 		G_RAID3_DEBUG(0, "Device %s: provider %s disconnected.",
2858 		    sc->sc_name, g_raid3_get_diskname(disk));
2859 
2860 		g_raid3_destroy_disk(disk);
2861 		break;
2862 	default:
2863 		KASSERT(1 == 0, ("Unknown state (%u).", state));
2864 		break;
2865 	}
2866 	return (0);
2867 }
2868 #undef	DISK_STATE_CHANGED
2869 
2870 int
2871 g_raid3_read_metadata(struct g_consumer *cp, struct g_raid3_metadata *md)
2872 {
2873 	struct g_provider *pp;
2874 	u_char *buf;
2875 	int error;
2876 
2877 	g_topology_assert();
2878 
2879 	error = g_access(cp, 1, 0, 0);
2880 	if (error != 0)
2881 		return (error);
2882 	pp = cp->provider;
2883 	g_topology_unlock();
2884 	/* Metadata are stored on last sector. */
2885 	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
2886 	    &error);
2887 	g_topology_lock();
2888 	g_access(cp, -1, 0, 0);
2889 	if (buf == NULL) {
2890 		G_RAID3_DEBUG(1, "Cannot read metadata from %s (error=%d).",
2891 		    cp->provider->name, error);
2892 		return (error);
2893 	}
2894 
2895 	/* Decode metadata. */
2896 	error = raid3_metadata_decode(buf, md);
2897 	g_free(buf);
2898 	if (strcmp(md->md_magic, G_RAID3_MAGIC) != 0)
2899 		return (EINVAL);
2900 	if (md->md_version > G_RAID3_VERSION) {
2901 		G_RAID3_DEBUG(0,
2902 		    "Kernel module is too old to handle metadata from %s.",
2903 		    cp->provider->name);
2904 		return (EINVAL);
2905 	}
2906 	if (error != 0) {
2907 		G_RAID3_DEBUG(1, "MD5 metadata hash mismatch for provider %s.",
2908 		    cp->provider->name);
2909 		return (error);
2910 	}
2911 	if (md->md_sectorsize > MAXPHYS) {
2912 		G_RAID3_DEBUG(0, "The blocksize is too big.");
2913 		return (EINVAL);
2914 	}
2915 
2916 	return (0);
2917 }
2918 
2919 static int
2920 g_raid3_check_metadata(struct g_raid3_softc *sc, struct g_provider *pp,
2921     struct g_raid3_metadata *md)
2922 {
2923 
2924 	if (md->md_no >= sc->sc_ndisks) {
2925 		G_RAID3_DEBUG(1, "Invalid disk %s number (no=%u), skipping.",
2926 		    pp->name, md->md_no);
2927 		return (EINVAL);
2928 	}
2929 	if (sc->sc_disks[md->md_no].d_state != G_RAID3_DISK_STATE_NODISK) {
2930 		G_RAID3_DEBUG(1, "Disk %s (no=%u) already exists, skipping.",
2931 		    pp->name, md->md_no);
2932 		return (EEXIST);
2933 	}
2934 	if (md->md_all != sc->sc_ndisks) {
2935 		G_RAID3_DEBUG(1,
2936 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2937 		    "md_all", pp->name, sc->sc_name);
2938 		return (EINVAL);
2939 	}
2940 	if ((md->md_mediasize % md->md_sectorsize) != 0) {
2941 		G_RAID3_DEBUG(1, "Invalid metadata (mediasize %% sectorsize != "
2942 		    "0) on disk %s (device %s), skipping.", pp->name,
2943 		    sc->sc_name);
2944 		return (EINVAL);
2945 	}
2946 	if (md->md_mediasize != sc->sc_mediasize) {
2947 		G_RAID3_DEBUG(1,
2948 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2949 		    "md_mediasize", pp->name, sc->sc_name);
2950 		return (EINVAL);
2951 	}
2952 	if ((md->md_mediasize % (sc->sc_ndisks - 1)) != 0) {
2953 		G_RAID3_DEBUG(1,
2954 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2955 		    "md_mediasize", pp->name, sc->sc_name);
2956 		return (EINVAL);
2957 	}
2958 	if ((sc->sc_mediasize / (sc->sc_ndisks - 1)) > pp->mediasize) {
2959 		G_RAID3_DEBUG(1,
2960 		    "Invalid size of disk %s (device %s), skipping.", pp->name,
2961 		    sc->sc_name);
2962 		return (EINVAL);
2963 	}
2964 	if ((md->md_sectorsize / pp->sectorsize) < sc->sc_ndisks - 1) {
2965 		G_RAID3_DEBUG(1,
2966 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2967 		    "md_sectorsize", pp->name, sc->sc_name);
2968 		return (EINVAL);
2969 	}
2970 	if (md->md_sectorsize != sc->sc_sectorsize) {
2971 		G_RAID3_DEBUG(1,
2972 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2973 		    "md_sectorsize", pp->name, sc->sc_name);
2974 		return (EINVAL);
2975 	}
2976 	if ((sc->sc_sectorsize % pp->sectorsize) != 0) {
2977 		G_RAID3_DEBUG(1,
2978 		    "Invalid sector size of disk %s (device %s), skipping.",
2979 		    pp->name, sc->sc_name);
2980 		return (EINVAL);
2981 	}
2982 	if ((md->md_mflags & ~G_RAID3_DEVICE_FLAG_MASK) != 0) {
2983 		G_RAID3_DEBUG(1,
2984 		    "Invalid device flags on disk %s (device %s), skipping.",
2985 		    pp->name, sc->sc_name);
2986 		return (EINVAL);
2987 	}
2988 	if ((md->md_mflags & G_RAID3_DEVICE_FLAG_VERIFY) != 0 &&
2989 	    (md->md_mflags & G_RAID3_DEVICE_FLAG_ROUND_ROBIN) != 0) {
2990 		/*
2991 		 * VERIFY and ROUND-ROBIN options are mutally exclusive.
2992 		 */
2993 		G_RAID3_DEBUG(1, "Both VERIFY and ROUND-ROBIN flags exist on "
2994 		    "disk %s (device %s), skipping.", pp->name, sc->sc_name);
2995 		return (EINVAL);
2996 	}
2997 	if ((md->md_dflags & ~G_RAID3_DISK_FLAG_MASK) != 0) {
2998 		G_RAID3_DEBUG(1,
2999 		    "Invalid disk flags on disk %s (device %s), skipping.",
3000 		    pp->name, sc->sc_name);
3001 		return (EINVAL);
3002 	}
3003 	return (0);
3004 }
3005 
3006 int
3007 g_raid3_add_disk(struct g_raid3_softc *sc, struct g_provider *pp,
3008     struct g_raid3_metadata *md)
3009 {
3010 	struct g_raid3_disk *disk;
3011 	int error;
3012 
3013 	g_topology_assert_not();
3014 	G_RAID3_DEBUG(2, "Adding disk %s.", pp->name);
3015 
3016 	error = g_raid3_check_metadata(sc, pp, md);
3017 	if (error != 0)
3018 		return (error);
3019 	if (sc->sc_state != G_RAID3_DEVICE_STATE_STARTING &&
3020 	    md->md_genid < sc->sc_genid) {
3021 		G_RAID3_DEBUG(0, "Component %s (device %s) broken, skipping.",
3022 		    pp->name, sc->sc_name);
3023 		return (EINVAL);
3024 	}
3025 	disk = g_raid3_init_disk(sc, pp, md, &error);
3026 	if (disk == NULL)
3027 		return (error);
3028 	error = g_raid3_event_send(disk, G_RAID3_DISK_STATE_NEW,
3029 	    G_RAID3_EVENT_WAIT);
3030 	if (error != 0)
3031 		return (error);
3032 	if (md->md_version < G_RAID3_VERSION) {
3033 		G_RAID3_DEBUG(0, "Upgrading metadata on %s (v%d->v%d).",
3034 		    pp->name, md->md_version, G_RAID3_VERSION);
3035 		g_raid3_update_metadata(disk);
3036 	}
3037 	return (0);
3038 }
3039 
3040 static void
3041 g_raid3_destroy_delayed(void *arg, int flag)
3042 {
3043 	struct g_raid3_softc *sc;
3044 	int error;
3045 
3046 	if (flag == EV_CANCEL) {
3047 		G_RAID3_DEBUG(1, "Destroying canceled.");
3048 		return;
3049 	}
3050 	sc = arg;
3051 	g_topology_unlock();
3052 	sx_xlock(&sc->sc_lock);
3053 	KASSERT((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) == 0,
3054 	    ("DESTROY flag set on %s.", sc->sc_name));
3055 	KASSERT((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROYING) != 0,
3056 	    ("DESTROYING flag not set on %s.", sc->sc_name));
3057 	G_RAID3_DEBUG(0, "Destroying %s (delayed).", sc->sc_name);
3058 	error = g_raid3_destroy(sc, G_RAID3_DESTROY_SOFT);
3059 	if (error != 0) {
3060 		G_RAID3_DEBUG(0, "Cannot destroy %s.", sc->sc_name);
3061 		sx_xunlock(&sc->sc_lock);
3062 	}
3063 	g_topology_lock();
3064 }
3065 
3066 static int
3067 g_raid3_access(struct g_provider *pp, int acr, int acw, int ace)
3068 {
3069 	struct g_raid3_softc *sc;
3070 	int dcr, dcw, dce, error = 0;
3071 
3072 	g_topology_assert();
3073 	G_RAID3_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, acr,
3074 	    acw, ace);
3075 
3076 	sc = pp->geom->softc;
3077 	if (sc == NULL && acr <= 0 && acw <= 0 && ace <= 0)
3078 		return (0);
3079 	KASSERT(sc != NULL, ("NULL softc (provider=%s).", pp->name));
3080 
3081 	dcr = pp->acr + acr;
3082 	dcw = pp->acw + acw;
3083 	dce = pp->ace + ace;
3084 
3085 	g_topology_unlock();
3086 	sx_xlock(&sc->sc_lock);
3087 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) != 0 ||
3088 	    g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) < sc->sc_ndisks - 1) {
3089 		if (acr > 0 || acw > 0 || ace > 0)
3090 			error = ENXIO;
3091 		goto end;
3092 	}
3093 	if (dcw == 0)
3094 		g_raid3_idle(sc, dcw);
3095 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROYING) != 0) {
3096 		if (acr > 0 || acw > 0 || ace > 0) {
3097 			error = ENXIO;
3098 			goto end;
3099 		}
3100 		if (dcr == 0 && dcw == 0 && dce == 0) {
3101 			g_post_event(g_raid3_destroy_delayed, sc, M_WAITOK,
3102 			    sc, NULL);
3103 		}
3104 	}
3105 end:
3106 	sx_xunlock(&sc->sc_lock);
3107 	g_topology_lock();
3108 	return (error);
3109 }
3110 
3111 static struct g_geom *
3112 g_raid3_create(struct g_class *mp, const struct g_raid3_metadata *md)
3113 {
3114 	struct g_raid3_softc *sc;
3115 	struct g_geom *gp;
3116 	int error, timeout;
3117 	u_int n;
3118 
3119 	g_topology_assert();
3120 	G_RAID3_DEBUG(1, "Creating device %s (id=%u).", md->md_name, md->md_id);
3121 
3122 	/* One disk is minimum. */
3123 	if (md->md_all < 1)
3124 		return (NULL);
3125 	/*
3126 	 * Action geom.
3127 	 */
3128 	gp = g_new_geomf(mp, "%s", md->md_name);
3129 	sc = malloc(sizeof(*sc), M_RAID3, M_WAITOK | M_ZERO);
3130 	sc->sc_disks = malloc(sizeof(struct g_raid3_disk) * md->md_all, M_RAID3,
3131 	    M_WAITOK | M_ZERO);
3132 	gp->start = g_raid3_start;
3133 	gp->orphan = g_raid3_orphan;
3134 	gp->access = g_raid3_access;
3135 	gp->dumpconf = g_raid3_dumpconf;
3136 
3137 	sc->sc_id = md->md_id;
3138 	sc->sc_mediasize = md->md_mediasize;
3139 	sc->sc_sectorsize = md->md_sectorsize;
3140 	sc->sc_ndisks = md->md_all;
3141 	sc->sc_round_robin = 0;
3142 	sc->sc_flags = md->md_mflags;
3143 	sc->sc_bump_id = 0;
3144 	sc->sc_idle = 1;
3145 	sc->sc_last_write = time_uptime;
3146 	sc->sc_writes = 0;
3147 	for (n = 0; n < sc->sc_ndisks; n++) {
3148 		sc->sc_disks[n].d_softc = sc;
3149 		sc->sc_disks[n].d_no = n;
3150 		sc->sc_disks[n].d_state = G_RAID3_DISK_STATE_NODISK;
3151 	}
3152 	sx_init(&sc->sc_lock, "graid3:lock");
3153 	bioq_init(&sc->sc_queue);
3154 	mtx_init(&sc->sc_queue_mtx, "graid3:queue", NULL, MTX_DEF);
3155 	bioq_init(&sc->sc_regular_delayed);
3156 	bioq_init(&sc->sc_inflight);
3157 	bioq_init(&sc->sc_sync_delayed);
3158 	TAILQ_INIT(&sc->sc_events);
3159 	mtx_init(&sc->sc_events_mtx, "graid3:events", NULL, MTX_DEF);
3160 	callout_init(&sc->sc_callout, 1);
3161 	sc->sc_state = G_RAID3_DEVICE_STATE_STARTING;
3162 	gp->softc = sc;
3163 	sc->sc_geom = gp;
3164 	sc->sc_provider = NULL;
3165 	/*
3166 	 * Synchronization geom.
3167 	 */
3168 	gp = g_new_geomf(mp, "%s.sync", md->md_name);
3169 	gp->softc = sc;
3170 	gp->orphan = g_raid3_orphan;
3171 	sc->sc_sync.ds_geom = gp;
3172 
3173 	if (!g_raid3_use_malloc) {
3174 		sc->sc_zones[G_RAID3_ZONE_64K].sz_zone = uma_zcreate("gr3:64k",
3175 		    65536, g_raid3_uma_ctor, g_raid3_uma_dtor, NULL, NULL,
3176 		    UMA_ALIGN_PTR, 0);
3177 		sc->sc_zones[G_RAID3_ZONE_64K].sz_inuse = 0;
3178 		sc->sc_zones[G_RAID3_ZONE_64K].sz_max = g_raid3_n64k;
3179 		sc->sc_zones[G_RAID3_ZONE_64K].sz_requested =
3180 		    sc->sc_zones[G_RAID3_ZONE_64K].sz_failed = 0;
3181 		sc->sc_zones[G_RAID3_ZONE_16K].sz_zone = uma_zcreate("gr3:16k",
3182 		    16384, g_raid3_uma_ctor, g_raid3_uma_dtor, NULL, NULL,
3183 		    UMA_ALIGN_PTR, 0);
3184 		sc->sc_zones[G_RAID3_ZONE_16K].sz_inuse = 0;
3185 		sc->sc_zones[G_RAID3_ZONE_16K].sz_max = g_raid3_n16k;
3186 		sc->sc_zones[G_RAID3_ZONE_16K].sz_requested =
3187 		    sc->sc_zones[G_RAID3_ZONE_16K].sz_failed = 0;
3188 		sc->sc_zones[G_RAID3_ZONE_4K].sz_zone = uma_zcreate("gr3:4k",
3189 		    4096, g_raid3_uma_ctor, g_raid3_uma_dtor, NULL, NULL,
3190 		    UMA_ALIGN_PTR, 0);
3191 		sc->sc_zones[G_RAID3_ZONE_4K].sz_inuse = 0;
3192 		sc->sc_zones[G_RAID3_ZONE_4K].sz_max = g_raid3_n4k;
3193 		sc->sc_zones[G_RAID3_ZONE_4K].sz_requested =
3194 		    sc->sc_zones[G_RAID3_ZONE_4K].sz_failed = 0;
3195 	}
3196 
3197 	error = kproc_create(g_raid3_worker, sc, &sc->sc_worker, 0, 0,
3198 	    "g_raid3 %s", md->md_name);
3199 	if (error != 0) {
3200 		G_RAID3_DEBUG(1, "Cannot create kernel thread for %s.",
3201 		    sc->sc_name);
3202 		if (!g_raid3_use_malloc) {
3203 			uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_64K].sz_zone);
3204 			uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_16K].sz_zone);
3205 			uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_4K].sz_zone);
3206 		}
3207 		g_destroy_geom(sc->sc_sync.ds_geom);
3208 		mtx_destroy(&sc->sc_events_mtx);
3209 		mtx_destroy(&sc->sc_queue_mtx);
3210 		sx_destroy(&sc->sc_lock);
3211 		g_destroy_geom(sc->sc_geom);
3212 		free(sc->sc_disks, M_RAID3);
3213 		free(sc, M_RAID3);
3214 		return (NULL);
3215 	}
3216 
3217 	G_RAID3_DEBUG(1, "Device %s created (%u components, id=%u).",
3218 	    sc->sc_name, sc->sc_ndisks, sc->sc_id);
3219 
3220 	sc->sc_rootmount = root_mount_hold("GRAID3");
3221 	G_RAID3_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
3222 
3223 	/*
3224 	 * Run timeout.
3225 	 */
3226 	timeout = atomic_load_acq_int(&g_raid3_timeout);
3227 	callout_reset(&sc->sc_callout, timeout * hz, g_raid3_go, sc);
3228 	return (sc->sc_geom);
3229 }
3230 
3231 int
3232 g_raid3_destroy(struct g_raid3_softc *sc, int how)
3233 {
3234 	struct g_provider *pp;
3235 
3236 	g_topology_assert_not();
3237 	if (sc == NULL)
3238 		return (ENXIO);
3239 	sx_assert(&sc->sc_lock, SX_XLOCKED);
3240 
3241 	pp = sc->sc_provider;
3242 	if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
3243 		switch (how) {
3244 		case G_RAID3_DESTROY_SOFT:
3245 			G_RAID3_DEBUG(1,
3246 			    "Device %s is still open (r%dw%de%d).", pp->name,
3247 			    pp->acr, pp->acw, pp->ace);
3248 			return (EBUSY);
3249 		case G_RAID3_DESTROY_DELAYED:
3250 			G_RAID3_DEBUG(1,
3251 			    "Device %s will be destroyed on last close.",
3252 			    pp->name);
3253 			if (sc->sc_syncdisk != NULL)
3254 				g_raid3_sync_stop(sc, 1);
3255 			sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROYING;
3256 			return (EBUSY);
3257 		case G_RAID3_DESTROY_HARD:
3258 			G_RAID3_DEBUG(1, "Device %s is still open, so it "
3259 			    "can't be definitely removed.", pp->name);
3260 			break;
3261 		}
3262 	}
3263 
3264 	g_topology_lock();
3265 	if (sc->sc_geom->softc == NULL) {
3266 		g_topology_unlock();
3267 		return (0);
3268 	}
3269 	sc->sc_geom->softc = NULL;
3270 	sc->sc_sync.ds_geom->softc = NULL;
3271 	g_topology_unlock();
3272 
3273 	sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
3274 	sc->sc_flags |= G_RAID3_DEVICE_FLAG_WAIT;
3275 	G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, sc);
3276 	sx_xunlock(&sc->sc_lock);
3277 	mtx_lock(&sc->sc_queue_mtx);
3278 	wakeup(sc);
3279 	wakeup(&sc->sc_queue);
3280 	mtx_unlock(&sc->sc_queue_mtx);
3281 	G_RAID3_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
3282 	while (sc->sc_worker != NULL)
3283 		tsleep(&sc->sc_worker, PRIBIO, "r3:destroy", hz / 5);
3284 	G_RAID3_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
3285 	sx_xlock(&sc->sc_lock);
3286 	g_raid3_destroy_device(sc);
3287 	free(sc->sc_disks, M_RAID3);
3288 	free(sc, M_RAID3);
3289 	return (0);
3290 }
3291 
3292 static void
3293 g_raid3_taste_orphan(struct g_consumer *cp)
3294 {
3295 
3296 	KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
3297 	    cp->provider->name));
3298 }
3299 
3300 static struct g_geom *
3301 g_raid3_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
3302 {
3303 	struct g_raid3_metadata md;
3304 	struct g_raid3_softc *sc;
3305 	struct g_consumer *cp;
3306 	struct g_geom *gp;
3307 	int error;
3308 
3309 	g_topology_assert();
3310 	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
3311 	G_RAID3_DEBUG(2, "Tasting %s.", pp->name);
3312 
3313 	gp = g_new_geomf(mp, "raid3:taste");
3314 	/* This orphan function should be never called. */
3315 	gp->orphan = g_raid3_taste_orphan;
3316 	cp = g_new_consumer(gp);
3317 	g_attach(cp, pp);
3318 	error = g_raid3_read_metadata(cp, &md);
3319 	g_detach(cp);
3320 	g_destroy_consumer(cp);
3321 	g_destroy_geom(gp);
3322 	if (error != 0)
3323 		return (NULL);
3324 	gp = NULL;
3325 
3326 	if (md.md_provider[0] != '\0' &&
3327 	    !g_compare_names(md.md_provider, pp->name))
3328 		return (NULL);
3329 	if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
3330 		return (NULL);
3331 	if (g_raid3_debug >= 2)
3332 		raid3_metadata_dump(&md);
3333 
3334 	/*
3335 	 * Let's check if device already exists.
3336 	 */
3337 	sc = NULL;
3338 	LIST_FOREACH(gp, &mp->geom, geom) {
3339 		sc = gp->softc;
3340 		if (sc == NULL)
3341 			continue;
3342 		if (sc->sc_sync.ds_geom == gp)
3343 			continue;
3344 		if (strcmp(md.md_name, sc->sc_name) != 0)
3345 			continue;
3346 		if (md.md_id != sc->sc_id) {
3347 			G_RAID3_DEBUG(0, "Device %s already configured.",
3348 			    sc->sc_name);
3349 			return (NULL);
3350 		}
3351 		break;
3352 	}
3353 	if (gp == NULL) {
3354 		gp = g_raid3_create(mp, &md);
3355 		if (gp == NULL) {
3356 			G_RAID3_DEBUG(0, "Cannot create device %s.",
3357 			    md.md_name);
3358 			return (NULL);
3359 		}
3360 		sc = gp->softc;
3361 	}
3362 	G_RAID3_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
3363 	g_topology_unlock();
3364 	sx_xlock(&sc->sc_lock);
3365 	error = g_raid3_add_disk(sc, pp, &md);
3366 	if (error != 0) {
3367 		G_RAID3_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
3368 		    pp->name, gp->name, error);
3369 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_NODISK) ==
3370 		    sc->sc_ndisks) {
3371 			g_cancel_event(sc);
3372 			g_raid3_destroy(sc, G_RAID3_DESTROY_HARD);
3373 			g_topology_lock();
3374 			return (NULL);
3375 		}
3376 		gp = NULL;
3377 	}
3378 	sx_xunlock(&sc->sc_lock);
3379 	g_topology_lock();
3380 	return (gp);
3381 }
3382 
3383 static int
3384 g_raid3_destroy_geom(struct gctl_req *req __unused, struct g_class *mp __unused,
3385     struct g_geom *gp)
3386 {
3387 	struct g_raid3_softc *sc;
3388 	int error;
3389 
3390 	g_topology_unlock();
3391 	sc = gp->softc;
3392 	sx_xlock(&sc->sc_lock);
3393 	g_cancel_event(sc);
3394 	error = g_raid3_destroy(gp->softc, G_RAID3_DESTROY_SOFT);
3395 	if (error != 0)
3396 		sx_xunlock(&sc->sc_lock);
3397 	g_topology_lock();
3398 	return (error);
3399 }
3400 
3401 static void
3402 g_raid3_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
3403     struct g_consumer *cp, struct g_provider *pp)
3404 {
3405 	struct g_raid3_softc *sc;
3406 
3407 	g_topology_assert();
3408 
3409 	sc = gp->softc;
3410 	if (sc == NULL)
3411 		return;
3412 	/* Skip synchronization geom. */
3413 	if (gp == sc->sc_sync.ds_geom)
3414 		return;
3415 	if (pp != NULL) {
3416 		/* Nothing here. */
3417 	} else if (cp != NULL) {
3418 		struct g_raid3_disk *disk;
3419 
3420 		disk = cp->private;
3421 		if (disk == NULL)
3422 			return;
3423 		g_topology_unlock();
3424 		sx_xlock(&sc->sc_lock);
3425 		sbuf_printf(sb, "%s<Type>", indent);
3426 		if (disk->d_no == sc->sc_ndisks - 1)
3427 			sbuf_printf(sb, "PARITY");
3428 		else
3429 			sbuf_printf(sb, "DATA");
3430 		sbuf_printf(sb, "</Type>\n");
3431 		sbuf_printf(sb, "%s<Number>%u</Number>\n", indent,
3432 		    (u_int)disk->d_no);
3433 		if (disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
3434 			sbuf_printf(sb, "%s<Synchronized>", indent);
3435 			if (disk->d_sync.ds_offset == 0)
3436 				sbuf_printf(sb, "0%%");
3437 			else {
3438 				sbuf_printf(sb, "%u%%",
3439 				    (u_int)((disk->d_sync.ds_offset * 100) /
3440 				    (sc->sc_mediasize / (sc->sc_ndisks - 1))));
3441 			}
3442 			sbuf_printf(sb, "</Synchronized>\n");
3443 			if (disk->d_sync.ds_offset > 0) {
3444 				sbuf_printf(sb, "%s<BytesSynced>%jd"
3445 				    "</BytesSynced>\n", indent,
3446 				    (intmax_t)disk->d_sync.ds_offset);
3447 			}
3448 		}
3449 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent,
3450 		    disk->d_sync.ds_syncid);
3451 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, disk->d_genid);
3452 		sbuf_printf(sb, "%s<Flags>", indent);
3453 		if (disk->d_flags == 0)
3454 			sbuf_printf(sb, "NONE");
3455 		else {
3456 			int first = 1;
3457 
3458 #define	ADD_FLAG(flag, name)	do {					\
3459 	if ((disk->d_flags & (flag)) != 0) {				\
3460 		if (!first)						\
3461 			sbuf_printf(sb, ", ");				\
3462 		else							\
3463 			first = 0;					\
3464 		sbuf_printf(sb, name);					\
3465 	}								\
3466 } while (0)
3467 			ADD_FLAG(G_RAID3_DISK_FLAG_DIRTY, "DIRTY");
3468 			ADD_FLAG(G_RAID3_DISK_FLAG_HARDCODED, "HARDCODED");
3469 			ADD_FLAG(G_RAID3_DISK_FLAG_SYNCHRONIZING,
3470 			    "SYNCHRONIZING");
3471 			ADD_FLAG(G_RAID3_DISK_FLAG_FORCE_SYNC, "FORCE_SYNC");
3472 			ADD_FLAG(G_RAID3_DISK_FLAG_BROKEN, "BROKEN");
3473 #undef	ADD_FLAG
3474 		}
3475 		sbuf_printf(sb, "</Flags>\n");
3476 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3477 		    g_raid3_disk_state2str(disk->d_state));
3478 		sx_xunlock(&sc->sc_lock);
3479 		g_topology_lock();
3480 	} else {
3481 		g_topology_unlock();
3482 		sx_xlock(&sc->sc_lock);
3483 		if (!g_raid3_use_malloc) {
3484 			sbuf_printf(sb,
3485 			    "%s<Zone4kRequested>%u</Zone4kRequested>\n", indent,
3486 			    sc->sc_zones[G_RAID3_ZONE_4K].sz_requested);
3487 			sbuf_printf(sb,
3488 			    "%s<Zone4kFailed>%u</Zone4kFailed>\n", indent,
3489 			    sc->sc_zones[G_RAID3_ZONE_4K].sz_failed);
3490 			sbuf_printf(sb,
3491 			    "%s<Zone16kRequested>%u</Zone16kRequested>\n", indent,
3492 			    sc->sc_zones[G_RAID3_ZONE_16K].sz_requested);
3493 			sbuf_printf(sb,
3494 			    "%s<Zone16kFailed>%u</Zone16kFailed>\n", indent,
3495 			    sc->sc_zones[G_RAID3_ZONE_16K].sz_failed);
3496 			sbuf_printf(sb,
3497 			    "%s<Zone64kRequested>%u</Zone64kRequested>\n", indent,
3498 			    sc->sc_zones[G_RAID3_ZONE_64K].sz_requested);
3499 			sbuf_printf(sb,
3500 			    "%s<Zone64kFailed>%u</Zone64kFailed>\n", indent,
3501 			    sc->sc_zones[G_RAID3_ZONE_64K].sz_failed);
3502 		}
3503 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
3504 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent, sc->sc_syncid);
3505 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, sc->sc_genid);
3506 		sbuf_printf(sb, "%s<Flags>", indent);
3507 		if (sc->sc_flags == 0)
3508 			sbuf_printf(sb, "NONE");
3509 		else {
3510 			int first = 1;
3511 
3512 #define	ADD_FLAG(flag, name)	do {					\
3513 	if ((sc->sc_flags & (flag)) != 0) {				\
3514 		if (!first)						\
3515 			sbuf_printf(sb, ", ");				\
3516 		else							\
3517 			first = 0;					\
3518 		sbuf_printf(sb, name);					\
3519 	}								\
3520 } while (0)
3521 			ADD_FLAG(G_RAID3_DEVICE_FLAG_NOFAILSYNC, "NOFAILSYNC");
3522 			ADD_FLAG(G_RAID3_DEVICE_FLAG_NOAUTOSYNC, "NOAUTOSYNC");
3523 			ADD_FLAG(G_RAID3_DEVICE_FLAG_ROUND_ROBIN,
3524 			    "ROUND-ROBIN");
3525 			ADD_FLAG(G_RAID3_DEVICE_FLAG_VERIFY, "VERIFY");
3526 #undef	ADD_FLAG
3527 		}
3528 		sbuf_printf(sb, "</Flags>\n");
3529 		sbuf_printf(sb, "%s<Components>%u</Components>\n", indent,
3530 		    sc->sc_ndisks);
3531 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3532 		    g_raid3_device_state2str(sc->sc_state));
3533 		sx_xunlock(&sc->sc_lock);
3534 		g_topology_lock();
3535 	}
3536 }
3537 
3538 static void
3539 g_raid3_shutdown_post_sync(void *arg, int howto)
3540 {
3541 	struct g_class *mp;
3542 	struct g_geom *gp, *gp2;
3543 	struct g_raid3_softc *sc;
3544 	int error;
3545 
3546 	mp = arg;
3547 	g_topology_lock();
3548 	g_raid3_shutdown = 1;
3549 	LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
3550 		if ((sc = gp->softc) == NULL)
3551 			continue;
3552 		/* Skip synchronization geom. */
3553 		if (gp == sc->sc_sync.ds_geom)
3554 			continue;
3555 		g_topology_unlock();
3556 		sx_xlock(&sc->sc_lock);
3557 		g_raid3_idle(sc, -1);
3558 		g_cancel_event(sc);
3559 		error = g_raid3_destroy(sc, G_RAID3_DESTROY_DELAYED);
3560 		if (error != 0)
3561 			sx_xunlock(&sc->sc_lock);
3562 		g_topology_lock();
3563 	}
3564 	g_topology_unlock();
3565 }
3566 
3567 static void
3568 g_raid3_init(struct g_class *mp)
3569 {
3570 
3571 	g_raid3_post_sync = EVENTHANDLER_REGISTER(shutdown_post_sync,
3572 	    g_raid3_shutdown_post_sync, mp, SHUTDOWN_PRI_FIRST);
3573 	if (g_raid3_post_sync == NULL)
3574 		G_RAID3_DEBUG(0, "Warning! Cannot register shutdown event.");
3575 }
3576 
3577 static void
3578 g_raid3_fini(struct g_class *mp)
3579 {
3580 
3581 	if (g_raid3_post_sync != NULL)
3582 		EVENTHANDLER_DEREGISTER(shutdown_post_sync, g_raid3_post_sync);
3583 }
3584 
3585 DECLARE_GEOM_CLASS(g_raid3_class, g_raid3);
3586 MODULE_VERSION(geom_raid3, 0);
3587