xref: /freebsd/sys/geom/raid3/g_raid3.c (revision 595e514d0df2bac5b813d35f83e32875dbf16a83)
1 /*-
2  * Copyright (c) 2004-2006 Pawel Jakub Dawidek <pjd@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/kernel.h>
33 #include <sys/module.h>
34 #include <sys/limits.h>
35 #include <sys/lock.h>
36 #include <sys/mutex.h>
37 #include <sys/bio.h>
38 #include <sys/sbuf.h>
39 #include <sys/sysctl.h>
40 #include <sys/malloc.h>
41 #include <sys/eventhandler.h>
42 #include <vm/uma.h>
43 #include <geom/geom.h>
44 #include <sys/proc.h>
45 #include <sys/kthread.h>
46 #include <sys/sched.h>
47 #include <geom/raid3/g_raid3.h>
48 
49 FEATURE(geom_raid3, "GEOM RAID-3 functionality");
50 
51 static MALLOC_DEFINE(M_RAID3, "raid3_data", "GEOM_RAID3 Data");
52 
53 SYSCTL_DECL(_kern_geom);
54 static SYSCTL_NODE(_kern_geom, OID_AUTO, raid3, CTLFLAG_RW, 0,
55     "GEOM_RAID3 stuff");
56 u_int g_raid3_debug = 0;
57 TUNABLE_INT("kern.geom.raid3.debug", &g_raid3_debug);
58 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, debug, CTLFLAG_RW, &g_raid3_debug, 0,
59     "Debug level");
60 static u_int g_raid3_timeout = 4;
61 TUNABLE_INT("kern.geom.raid3.timeout", &g_raid3_timeout);
62 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, timeout, CTLFLAG_RW, &g_raid3_timeout,
63     0, "Time to wait on all raid3 components");
64 static u_int g_raid3_idletime = 5;
65 TUNABLE_INT("kern.geom.raid3.idletime", &g_raid3_idletime);
66 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, idletime, CTLFLAG_RW,
67     &g_raid3_idletime, 0, "Mark components as clean when idling");
68 static u_int g_raid3_disconnect_on_failure = 1;
69 TUNABLE_INT("kern.geom.raid3.disconnect_on_failure",
70     &g_raid3_disconnect_on_failure);
71 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, disconnect_on_failure, CTLFLAG_RW,
72     &g_raid3_disconnect_on_failure, 0, "Disconnect component on I/O failure.");
73 static u_int g_raid3_syncreqs = 2;
74 TUNABLE_INT("kern.geom.raid3.sync_requests", &g_raid3_syncreqs);
75 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, sync_requests, CTLFLAG_RDTUN,
76     &g_raid3_syncreqs, 0, "Parallel synchronization I/O requests.");
77 static u_int g_raid3_use_malloc = 0;
78 TUNABLE_INT("kern.geom.raid3.use_malloc", &g_raid3_use_malloc);
79 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, use_malloc, CTLFLAG_RDTUN,
80     &g_raid3_use_malloc, 0, "Use malloc(9) instead of uma(9).");
81 
82 static u_int g_raid3_n64k = 50;
83 TUNABLE_INT("kern.geom.raid3.n64k", &g_raid3_n64k);
84 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, n64k, CTLFLAG_RD, &g_raid3_n64k, 0,
85     "Maximum number of 64kB allocations");
86 static u_int g_raid3_n16k = 200;
87 TUNABLE_INT("kern.geom.raid3.n16k", &g_raid3_n16k);
88 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, n16k, CTLFLAG_RD, &g_raid3_n16k, 0,
89     "Maximum number of 16kB allocations");
90 static u_int g_raid3_n4k = 1200;
91 TUNABLE_INT("kern.geom.raid3.n4k", &g_raid3_n4k);
92 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, n4k, CTLFLAG_RD, &g_raid3_n4k, 0,
93     "Maximum number of 4kB allocations");
94 
95 static SYSCTL_NODE(_kern_geom_raid3, OID_AUTO, stat, CTLFLAG_RW, 0,
96     "GEOM_RAID3 statistics");
97 static u_int g_raid3_parity_mismatch = 0;
98 SYSCTL_UINT(_kern_geom_raid3_stat, OID_AUTO, parity_mismatch, CTLFLAG_RD,
99     &g_raid3_parity_mismatch, 0, "Number of failures in VERIFY mode");
100 
101 #define	MSLEEP(ident, mtx, priority, wmesg, timeout)	do {		\
102 	G_RAID3_DEBUG(4, "%s: Sleeping %p.", __func__, (ident));	\
103 	msleep((ident), (mtx), (priority), (wmesg), (timeout));		\
104 	G_RAID3_DEBUG(4, "%s: Woken up %p.", __func__, (ident));	\
105 } while (0)
106 
107 static eventhandler_tag g_raid3_post_sync = NULL;
108 static int g_raid3_shutdown = 0;
109 
110 static int g_raid3_destroy_geom(struct gctl_req *req, struct g_class *mp,
111     struct g_geom *gp);
112 static g_taste_t g_raid3_taste;
113 static void g_raid3_init(struct g_class *mp);
114 static void g_raid3_fini(struct g_class *mp);
115 
116 struct g_class g_raid3_class = {
117 	.name = G_RAID3_CLASS_NAME,
118 	.version = G_VERSION,
119 	.ctlreq = g_raid3_config,
120 	.taste = g_raid3_taste,
121 	.destroy_geom = g_raid3_destroy_geom,
122 	.init = g_raid3_init,
123 	.fini = g_raid3_fini
124 };
125 
126 
127 static void g_raid3_destroy_provider(struct g_raid3_softc *sc);
128 static int g_raid3_update_disk(struct g_raid3_disk *disk, u_int state);
129 static void g_raid3_update_device(struct g_raid3_softc *sc, boolean_t force);
130 static void g_raid3_dumpconf(struct sbuf *sb, const char *indent,
131     struct g_geom *gp, struct g_consumer *cp, struct g_provider *pp);
132 static void g_raid3_sync_stop(struct g_raid3_softc *sc, int type);
133 static int g_raid3_register_request(struct bio *pbp);
134 static void g_raid3_sync_release(struct g_raid3_softc *sc);
135 
136 
137 static const char *
138 g_raid3_disk_state2str(int state)
139 {
140 
141 	switch (state) {
142 	case G_RAID3_DISK_STATE_NODISK:
143 		return ("NODISK");
144 	case G_RAID3_DISK_STATE_NONE:
145 		return ("NONE");
146 	case G_RAID3_DISK_STATE_NEW:
147 		return ("NEW");
148 	case G_RAID3_DISK_STATE_ACTIVE:
149 		return ("ACTIVE");
150 	case G_RAID3_DISK_STATE_STALE:
151 		return ("STALE");
152 	case G_RAID3_DISK_STATE_SYNCHRONIZING:
153 		return ("SYNCHRONIZING");
154 	case G_RAID3_DISK_STATE_DISCONNECTED:
155 		return ("DISCONNECTED");
156 	default:
157 		return ("INVALID");
158 	}
159 }
160 
161 static const char *
162 g_raid3_device_state2str(int state)
163 {
164 
165 	switch (state) {
166 	case G_RAID3_DEVICE_STATE_STARTING:
167 		return ("STARTING");
168 	case G_RAID3_DEVICE_STATE_DEGRADED:
169 		return ("DEGRADED");
170 	case G_RAID3_DEVICE_STATE_COMPLETE:
171 		return ("COMPLETE");
172 	default:
173 		return ("INVALID");
174 	}
175 }
176 
177 const char *
178 g_raid3_get_diskname(struct g_raid3_disk *disk)
179 {
180 
181 	if (disk->d_consumer == NULL || disk->d_consumer->provider == NULL)
182 		return ("[unknown]");
183 	return (disk->d_name);
184 }
185 
186 static void *
187 g_raid3_alloc(struct g_raid3_softc *sc, size_t size, int flags)
188 {
189 	void *ptr;
190 	enum g_raid3_zones zone;
191 
192 	if (g_raid3_use_malloc ||
193 	    (zone = g_raid3_zone(size)) == G_RAID3_NUM_ZONES)
194 		ptr = malloc(size, M_RAID3, flags);
195 	else {
196 		ptr = uma_zalloc_arg(sc->sc_zones[zone].sz_zone,
197 		   &sc->sc_zones[zone], flags);
198 		sc->sc_zones[zone].sz_requested++;
199 		if (ptr == NULL)
200 			sc->sc_zones[zone].sz_failed++;
201 	}
202 	return (ptr);
203 }
204 
205 static void
206 g_raid3_free(struct g_raid3_softc *sc, void *ptr, size_t size)
207 {
208 	enum g_raid3_zones zone;
209 
210 	if (g_raid3_use_malloc ||
211 	    (zone = g_raid3_zone(size)) == G_RAID3_NUM_ZONES)
212 		free(ptr, M_RAID3);
213 	else {
214 		uma_zfree_arg(sc->sc_zones[zone].sz_zone,
215 		    ptr, &sc->sc_zones[zone]);
216 	}
217 }
218 
219 static int
220 g_raid3_uma_ctor(void *mem, int size, void *arg, int flags)
221 {
222 	struct g_raid3_zone *sz = arg;
223 
224 	if (sz->sz_max > 0 && sz->sz_inuse == sz->sz_max)
225 		return (ENOMEM);
226 	sz->sz_inuse++;
227 	return (0);
228 }
229 
230 static void
231 g_raid3_uma_dtor(void *mem, int size, void *arg)
232 {
233 	struct g_raid3_zone *sz = arg;
234 
235 	sz->sz_inuse--;
236 }
237 
238 #define	g_raid3_xor(src, dst, size)					\
239 	_g_raid3_xor((uint64_t *)(src),					\
240 	    (uint64_t *)(dst), (size_t)size)
241 static void
242 _g_raid3_xor(uint64_t *src, uint64_t *dst, size_t size)
243 {
244 
245 	KASSERT((size % 128) == 0, ("Invalid size: %zu.", size));
246 	for (; size > 0; size -= 128) {
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 		*dst++ ^= (*src++);
256 		*dst++ ^= (*src++);
257 		*dst++ ^= (*src++);
258 		*dst++ ^= (*src++);
259 		*dst++ ^= (*src++);
260 		*dst++ ^= (*src++);
261 		*dst++ ^= (*src++);
262 		*dst++ ^= (*src++);
263 	}
264 }
265 
266 static int
267 g_raid3_is_zero(struct bio *bp)
268 {
269 	static const uint64_t zeros[] = {
270 	    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
271 	};
272 	u_char *addr;
273 	ssize_t size;
274 
275 	size = bp->bio_length;
276 	addr = (u_char *)bp->bio_data;
277 	for (; size > 0; size -= sizeof(zeros), addr += sizeof(zeros)) {
278 		if (bcmp(addr, zeros, sizeof(zeros)) != 0)
279 			return (0);
280 	}
281 	return (1);
282 }
283 
284 /*
285  * --- Events handling functions ---
286  * Events in geom_raid3 are used to maintain disks and device status
287  * from one thread to simplify locking.
288  */
289 static void
290 g_raid3_event_free(struct g_raid3_event *ep)
291 {
292 
293 	free(ep, M_RAID3);
294 }
295 
296 int
297 g_raid3_event_send(void *arg, int state, int flags)
298 {
299 	struct g_raid3_softc *sc;
300 	struct g_raid3_disk *disk;
301 	struct g_raid3_event *ep;
302 	int error;
303 
304 	ep = malloc(sizeof(*ep), M_RAID3, M_WAITOK);
305 	G_RAID3_DEBUG(4, "%s: Sending event %p.", __func__, ep);
306 	if ((flags & G_RAID3_EVENT_DEVICE) != 0) {
307 		disk = NULL;
308 		sc = arg;
309 	} else {
310 		disk = arg;
311 		sc = disk->d_softc;
312 	}
313 	ep->e_disk = disk;
314 	ep->e_state = state;
315 	ep->e_flags = flags;
316 	ep->e_error = 0;
317 	mtx_lock(&sc->sc_events_mtx);
318 	TAILQ_INSERT_TAIL(&sc->sc_events, ep, e_next);
319 	mtx_unlock(&sc->sc_events_mtx);
320 	G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, sc);
321 	mtx_lock(&sc->sc_queue_mtx);
322 	wakeup(sc);
323 	wakeup(&sc->sc_queue);
324 	mtx_unlock(&sc->sc_queue_mtx);
325 	if ((flags & G_RAID3_EVENT_DONTWAIT) != 0)
326 		return (0);
327 	sx_assert(&sc->sc_lock, SX_XLOCKED);
328 	G_RAID3_DEBUG(4, "%s: Sleeping %p.", __func__, ep);
329 	sx_xunlock(&sc->sc_lock);
330 	while ((ep->e_flags & G_RAID3_EVENT_DONE) == 0) {
331 		mtx_lock(&sc->sc_events_mtx);
332 		MSLEEP(ep, &sc->sc_events_mtx, PRIBIO | PDROP, "r3:event",
333 		    hz * 5);
334 	}
335 	error = ep->e_error;
336 	g_raid3_event_free(ep);
337 	sx_xlock(&sc->sc_lock);
338 	return (error);
339 }
340 
341 static struct g_raid3_event *
342 g_raid3_event_get(struct g_raid3_softc *sc)
343 {
344 	struct g_raid3_event *ep;
345 
346 	mtx_lock(&sc->sc_events_mtx);
347 	ep = TAILQ_FIRST(&sc->sc_events);
348 	mtx_unlock(&sc->sc_events_mtx);
349 	return (ep);
350 }
351 
352 static void
353 g_raid3_event_remove(struct g_raid3_softc *sc, struct g_raid3_event *ep)
354 {
355 
356 	mtx_lock(&sc->sc_events_mtx);
357 	TAILQ_REMOVE(&sc->sc_events, ep, e_next);
358 	mtx_unlock(&sc->sc_events_mtx);
359 }
360 
361 static void
362 g_raid3_event_cancel(struct g_raid3_disk *disk)
363 {
364 	struct g_raid3_softc *sc;
365 	struct g_raid3_event *ep, *tmpep;
366 
367 	sc = disk->d_softc;
368 	sx_assert(&sc->sc_lock, SX_XLOCKED);
369 
370 	mtx_lock(&sc->sc_events_mtx);
371 	TAILQ_FOREACH_SAFE(ep, &sc->sc_events, e_next, tmpep) {
372 		if ((ep->e_flags & G_RAID3_EVENT_DEVICE) != 0)
373 			continue;
374 		if (ep->e_disk != disk)
375 			continue;
376 		TAILQ_REMOVE(&sc->sc_events, ep, e_next);
377 		if ((ep->e_flags & G_RAID3_EVENT_DONTWAIT) != 0)
378 			g_raid3_event_free(ep);
379 		else {
380 			ep->e_error = ECANCELED;
381 			wakeup(ep);
382 		}
383 	}
384 	mtx_unlock(&sc->sc_events_mtx);
385 }
386 
387 /*
388  * Return the number of disks in the given state.
389  * If state is equal to -1, count all connected disks.
390  */
391 u_int
392 g_raid3_ndisks(struct g_raid3_softc *sc, int state)
393 {
394 	struct g_raid3_disk *disk;
395 	u_int n, ndisks;
396 
397 	sx_assert(&sc->sc_lock, SX_LOCKED);
398 
399 	for (n = ndisks = 0; n < sc->sc_ndisks; n++) {
400 		disk = &sc->sc_disks[n];
401 		if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
402 			continue;
403 		if (state == -1 || disk->d_state == state)
404 			ndisks++;
405 	}
406 	return (ndisks);
407 }
408 
409 static u_int
410 g_raid3_nrequests(struct g_raid3_softc *sc, struct g_consumer *cp)
411 {
412 	struct bio *bp;
413 	u_int nreqs = 0;
414 
415 	mtx_lock(&sc->sc_queue_mtx);
416 	TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
417 		if (bp->bio_from == cp)
418 			nreqs++;
419 	}
420 	mtx_unlock(&sc->sc_queue_mtx);
421 	return (nreqs);
422 }
423 
424 static int
425 g_raid3_is_busy(struct g_raid3_softc *sc, struct g_consumer *cp)
426 {
427 
428 	if (cp->index > 0) {
429 		G_RAID3_DEBUG(2,
430 		    "I/O requests for %s exist, can't destroy it now.",
431 		    cp->provider->name);
432 		return (1);
433 	}
434 	if (g_raid3_nrequests(sc, cp) > 0) {
435 		G_RAID3_DEBUG(2,
436 		    "I/O requests for %s in queue, can't destroy it now.",
437 		    cp->provider->name);
438 		return (1);
439 	}
440 	return (0);
441 }
442 
443 static void
444 g_raid3_destroy_consumer(void *arg, int flags __unused)
445 {
446 	struct g_consumer *cp;
447 
448 	g_topology_assert();
449 
450 	cp = arg;
451 	G_RAID3_DEBUG(1, "Consumer %s destroyed.", cp->provider->name);
452 	g_detach(cp);
453 	g_destroy_consumer(cp);
454 }
455 
456 static void
457 g_raid3_kill_consumer(struct g_raid3_softc *sc, struct g_consumer *cp)
458 {
459 	struct g_provider *pp;
460 	int retaste_wait;
461 
462 	g_topology_assert();
463 
464 	cp->private = NULL;
465 	if (g_raid3_is_busy(sc, cp))
466 		return;
467 	G_RAID3_DEBUG(2, "Consumer %s destroyed.", cp->provider->name);
468 	pp = cp->provider;
469 	retaste_wait = 0;
470 	if (cp->acw == 1) {
471 		if ((pp->geom->flags & G_GEOM_WITHER) == 0)
472 			retaste_wait = 1;
473 	}
474 	G_RAID3_DEBUG(2, "Access %s r%dw%de%d = %d", pp->name, -cp->acr,
475 	    -cp->acw, -cp->ace, 0);
476 	if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
477 		g_access(cp, -cp->acr, -cp->acw, -cp->ace);
478 	if (retaste_wait) {
479 		/*
480 		 * After retaste event was send (inside g_access()), we can send
481 		 * event to detach and destroy consumer.
482 		 * A class, which has consumer to the given provider connected
483 		 * will not receive retaste event for the provider.
484 		 * This is the way how I ignore retaste events when I close
485 		 * consumers opened for write: I detach and destroy consumer
486 		 * after retaste event is sent.
487 		 */
488 		g_post_event(g_raid3_destroy_consumer, cp, M_WAITOK, NULL);
489 		return;
490 	}
491 	G_RAID3_DEBUG(1, "Consumer %s destroyed.", pp->name);
492 	g_detach(cp);
493 	g_destroy_consumer(cp);
494 }
495 
496 static int
497 g_raid3_connect_disk(struct g_raid3_disk *disk, struct g_provider *pp)
498 {
499 	struct g_consumer *cp;
500 	int error;
501 
502 	g_topology_assert_not();
503 	KASSERT(disk->d_consumer == NULL,
504 	    ("Disk already connected (device %s).", disk->d_softc->sc_name));
505 
506 	g_topology_lock();
507 	cp = g_new_consumer(disk->d_softc->sc_geom);
508 	error = g_attach(cp, pp);
509 	if (error != 0) {
510 		g_destroy_consumer(cp);
511 		g_topology_unlock();
512 		return (error);
513 	}
514 	error = g_access(cp, 1, 1, 1);
515 		g_topology_unlock();
516 	if (error != 0) {
517 		g_detach(cp);
518 		g_destroy_consumer(cp);
519 		G_RAID3_DEBUG(0, "Cannot open consumer %s (error=%d).",
520 		    pp->name, error);
521 		return (error);
522 	}
523 	disk->d_consumer = cp;
524 	disk->d_consumer->private = disk;
525 	disk->d_consumer->index = 0;
526 	G_RAID3_DEBUG(2, "Disk %s connected.", g_raid3_get_diskname(disk));
527 	return (0);
528 }
529 
530 static void
531 g_raid3_disconnect_consumer(struct g_raid3_softc *sc, struct g_consumer *cp)
532 {
533 
534 	g_topology_assert();
535 
536 	if (cp == NULL)
537 		return;
538 	if (cp->provider != NULL)
539 		g_raid3_kill_consumer(sc, cp);
540 	else
541 		g_destroy_consumer(cp);
542 }
543 
544 /*
545  * Initialize disk. This means allocate memory, create consumer, attach it
546  * to the provider and open access (r1w1e1) to it.
547  */
548 static struct g_raid3_disk *
549 g_raid3_init_disk(struct g_raid3_softc *sc, struct g_provider *pp,
550     struct g_raid3_metadata *md, int *errorp)
551 {
552 	struct g_raid3_disk *disk;
553 	int error;
554 
555 	disk = &sc->sc_disks[md->md_no];
556 	error = g_raid3_connect_disk(disk, pp);
557 	if (error != 0) {
558 		if (errorp != NULL)
559 			*errorp = error;
560 		return (NULL);
561 	}
562 	disk->d_state = G_RAID3_DISK_STATE_NONE;
563 	disk->d_flags = md->md_dflags;
564 	if (md->md_provider[0] != '\0')
565 		disk->d_flags |= G_RAID3_DISK_FLAG_HARDCODED;
566 	disk->d_sync.ds_consumer = NULL;
567 	disk->d_sync.ds_offset = md->md_sync_offset;
568 	disk->d_sync.ds_offset_done = md->md_sync_offset;
569 	disk->d_genid = md->md_genid;
570 	disk->d_sync.ds_syncid = md->md_syncid;
571 	if (errorp != NULL)
572 		*errorp = 0;
573 	return (disk);
574 }
575 
576 static void
577 g_raid3_destroy_disk(struct g_raid3_disk *disk)
578 {
579 	struct g_raid3_softc *sc;
580 
581 	g_topology_assert_not();
582 	sc = disk->d_softc;
583 	sx_assert(&sc->sc_lock, SX_XLOCKED);
584 
585 	if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
586 		return;
587 	g_raid3_event_cancel(disk);
588 	switch (disk->d_state) {
589 	case G_RAID3_DISK_STATE_SYNCHRONIZING:
590 		if (sc->sc_syncdisk != NULL)
591 			g_raid3_sync_stop(sc, 1);
592 		/* FALLTHROUGH */
593 	case G_RAID3_DISK_STATE_NEW:
594 	case G_RAID3_DISK_STATE_STALE:
595 	case G_RAID3_DISK_STATE_ACTIVE:
596 		g_topology_lock();
597 		g_raid3_disconnect_consumer(sc, disk->d_consumer);
598 		g_topology_unlock();
599 		disk->d_consumer = NULL;
600 		break;
601 	default:
602 		KASSERT(0 == 1, ("Wrong disk state (%s, %s).",
603 		    g_raid3_get_diskname(disk),
604 		    g_raid3_disk_state2str(disk->d_state)));
605 	}
606 	disk->d_state = G_RAID3_DISK_STATE_NODISK;
607 }
608 
609 static void
610 g_raid3_destroy_device(struct g_raid3_softc *sc)
611 {
612 	struct g_raid3_event *ep;
613 	struct g_raid3_disk *disk;
614 	struct g_geom *gp;
615 	struct g_consumer *cp;
616 	u_int n;
617 
618 	g_topology_assert_not();
619 	sx_assert(&sc->sc_lock, SX_XLOCKED);
620 
621 	gp = sc->sc_geom;
622 	if (sc->sc_provider != NULL)
623 		g_raid3_destroy_provider(sc);
624 	for (n = 0; n < sc->sc_ndisks; n++) {
625 		disk = &sc->sc_disks[n];
626 		if (disk->d_state != G_RAID3_DISK_STATE_NODISK) {
627 			disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
628 			g_raid3_update_metadata(disk);
629 			g_raid3_destroy_disk(disk);
630 		}
631 	}
632 	while ((ep = g_raid3_event_get(sc)) != NULL) {
633 		g_raid3_event_remove(sc, ep);
634 		if ((ep->e_flags & G_RAID3_EVENT_DONTWAIT) != 0)
635 			g_raid3_event_free(ep);
636 		else {
637 			ep->e_error = ECANCELED;
638 			ep->e_flags |= G_RAID3_EVENT_DONE;
639 			G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, ep);
640 			mtx_lock(&sc->sc_events_mtx);
641 			wakeup(ep);
642 			mtx_unlock(&sc->sc_events_mtx);
643 		}
644 	}
645 	callout_drain(&sc->sc_callout);
646 	cp = LIST_FIRST(&sc->sc_sync.ds_geom->consumer);
647 	g_topology_lock();
648 	if (cp != NULL)
649 		g_raid3_disconnect_consumer(sc, cp);
650 	g_wither_geom(sc->sc_sync.ds_geom, ENXIO);
651 	G_RAID3_DEBUG(0, "Device %s destroyed.", gp->name);
652 	g_wither_geom(gp, ENXIO);
653 	g_topology_unlock();
654 	if (!g_raid3_use_malloc) {
655 		uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_64K].sz_zone);
656 		uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_16K].sz_zone);
657 		uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_4K].sz_zone);
658 	}
659 	mtx_destroy(&sc->sc_queue_mtx);
660 	mtx_destroy(&sc->sc_events_mtx);
661 	sx_xunlock(&sc->sc_lock);
662 	sx_destroy(&sc->sc_lock);
663 }
664 
665 static void
666 g_raid3_orphan(struct g_consumer *cp)
667 {
668 	struct g_raid3_disk *disk;
669 
670 	g_topology_assert();
671 
672 	disk = cp->private;
673 	if (disk == NULL)
674 		return;
675 	disk->d_softc->sc_bump_id = G_RAID3_BUMP_SYNCID;
676 	g_raid3_event_send(disk, G_RAID3_DISK_STATE_DISCONNECTED,
677 	    G_RAID3_EVENT_DONTWAIT);
678 }
679 
680 static int
681 g_raid3_write_metadata(struct g_raid3_disk *disk, struct g_raid3_metadata *md)
682 {
683 	struct g_raid3_softc *sc;
684 	struct g_consumer *cp;
685 	off_t offset, length;
686 	u_char *sector;
687 	int error = 0;
688 
689 	g_topology_assert_not();
690 	sc = disk->d_softc;
691 	sx_assert(&sc->sc_lock, SX_LOCKED);
692 
693 	cp = disk->d_consumer;
694 	KASSERT(cp != NULL, ("NULL consumer (%s).", sc->sc_name));
695 	KASSERT(cp->provider != NULL, ("NULL provider (%s).", sc->sc_name));
696 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
697 	    ("Consumer %s closed? (r%dw%de%d).", cp->provider->name, cp->acr,
698 	    cp->acw, cp->ace));
699 	length = cp->provider->sectorsize;
700 	offset = cp->provider->mediasize - length;
701 	sector = malloc((size_t)length, M_RAID3, M_WAITOK | M_ZERO);
702 	if (md != NULL)
703 		raid3_metadata_encode(md, sector);
704 	error = g_write_data(cp, offset, sector, length);
705 	free(sector, M_RAID3);
706 	if (error != 0) {
707 		if ((disk->d_flags & G_RAID3_DISK_FLAG_BROKEN) == 0) {
708 			G_RAID3_DEBUG(0, "Cannot write metadata on %s "
709 			    "(device=%s, error=%d).",
710 			    g_raid3_get_diskname(disk), sc->sc_name, error);
711 			disk->d_flags |= G_RAID3_DISK_FLAG_BROKEN;
712 		} else {
713 			G_RAID3_DEBUG(1, "Cannot write metadata on %s "
714 			    "(device=%s, error=%d).",
715 			    g_raid3_get_diskname(disk), sc->sc_name, error);
716 		}
717 		if (g_raid3_disconnect_on_failure &&
718 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
719 			sc->sc_bump_id |= G_RAID3_BUMP_GENID;
720 			g_raid3_event_send(disk,
721 			    G_RAID3_DISK_STATE_DISCONNECTED,
722 			    G_RAID3_EVENT_DONTWAIT);
723 		}
724 	}
725 	return (error);
726 }
727 
728 int
729 g_raid3_clear_metadata(struct g_raid3_disk *disk)
730 {
731 	int error;
732 
733 	g_topology_assert_not();
734 	sx_assert(&disk->d_softc->sc_lock, SX_LOCKED);
735 
736 	error = g_raid3_write_metadata(disk, NULL);
737 	if (error == 0) {
738 		G_RAID3_DEBUG(2, "Metadata on %s cleared.",
739 		    g_raid3_get_diskname(disk));
740 	} else {
741 		G_RAID3_DEBUG(0,
742 		    "Cannot clear metadata on disk %s (error=%d).",
743 		    g_raid3_get_diskname(disk), error);
744 	}
745 	return (error);
746 }
747 
748 void
749 g_raid3_fill_metadata(struct g_raid3_disk *disk, struct g_raid3_metadata *md)
750 {
751 	struct g_raid3_softc *sc;
752 	struct g_provider *pp;
753 
754 	sc = disk->d_softc;
755 	strlcpy(md->md_magic, G_RAID3_MAGIC, sizeof(md->md_magic));
756 	md->md_version = G_RAID3_VERSION;
757 	strlcpy(md->md_name, sc->sc_name, sizeof(md->md_name));
758 	md->md_id = sc->sc_id;
759 	md->md_all = sc->sc_ndisks;
760 	md->md_genid = sc->sc_genid;
761 	md->md_mediasize = sc->sc_mediasize;
762 	md->md_sectorsize = sc->sc_sectorsize;
763 	md->md_mflags = (sc->sc_flags & G_RAID3_DEVICE_FLAG_MASK);
764 	md->md_no = disk->d_no;
765 	md->md_syncid = disk->d_sync.ds_syncid;
766 	md->md_dflags = (disk->d_flags & G_RAID3_DISK_FLAG_MASK);
767 	if (disk->d_state != G_RAID3_DISK_STATE_SYNCHRONIZING)
768 		md->md_sync_offset = 0;
769 	else {
770 		md->md_sync_offset =
771 		    disk->d_sync.ds_offset_done / (sc->sc_ndisks - 1);
772 	}
773 	if (disk->d_consumer != NULL && disk->d_consumer->provider != NULL)
774 		pp = disk->d_consumer->provider;
775 	else
776 		pp = NULL;
777 	if ((disk->d_flags & G_RAID3_DISK_FLAG_HARDCODED) != 0 && pp != NULL)
778 		strlcpy(md->md_provider, pp->name, sizeof(md->md_provider));
779 	else
780 		bzero(md->md_provider, sizeof(md->md_provider));
781 	if (pp != NULL)
782 		md->md_provsize = pp->mediasize;
783 	else
784 		md->md_provsize = 0;
785 }
786 
787 void
788 g_raid3_update_metadata(struct g_raid3_disk *disk)
789 {
790 	struct g_raid3_softc *sc;
791 	struct g_raid3_metadata md;
792 	int error;
793 
794 	g_topology_assert_not();
795 	sc = disk->d_softc;
796 	sx_assert(&sc->sc_lock, SX_LOCKED);
797 
798 	g_raid3_fill_metadata(disk, &md);
799 	error = g_raid3_write_metadata(disk, &md);
800 	if (error == 0) {
801 		G_RAID3_DEBUG(2, "Metadata on %s updated.",
802 		    g_raid3_get_diskname(disk));
803 	} else {
804 		G_RAID3_DEBUG(0,
805 		    "Cannot update metadata on disk %s (error=%d).",
806 		    g_raid3_get_diskname(disk), error);
807 	}
808 }
809 
810 static void
811 g_raid3_bump_syncid(struct g_raid3_softc *sc)
812 {
813 	struct g_raid3_disk *disk;
814 	u_int n;
815 
816 	g_topology_assert_not();
817 	sx_assert(&sc->sc_lock, SX_XLOCKED);
818 	KASSERT(g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) > 0,
819 	    ("%s called with no active disks (device=%s).", __func__,
820 	    sc->sc_name));
821 
822 	sc->sc_syncid++;
823 	G_RAID3_DEBUG(1, "Device %s: syncid bumped to %u.", sc->sc_name,
824 	    sc->sc_syncid);
825 	for (n = 0; n < sc->sc_ndisks; n++) {
826 		disk = &sc->sc_disks[n];
827 		if (disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
828 		    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
829 			disk->d_sync.ds_syncid = sc->sc_syncid;
830 			g_raid3_update_metadata(disk);
831 		}
832 	}
833 }
834 
835 static void
836 g_raid3_bump_genid(struct g_raid3_softc *sc)
837 {
838 	struct g_raid3_disk *disk;
839 	u_int n;
840 
841 	g_topology_assert_not();
842 	sx_assert(&sc->sc_lock, SX_XLOCKED);
843 	KASSERT(g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) > 0,
844 	    ("%s called with no active disks (device=%s).", __func__,
845 	    sc->sc_name));
846 
847 	sc->sc_genid++;
848 	G_RAID3_DEBUG(1, "Device %s: genid bumped to %u.", sc->sc_name,
849 	    sc->sc_genid);
850 	for (n = 0; n < sc->sc_ndisks; n++) {
851 		disk = &sc->sc_disks[n];
852 		if (disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
853 		    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
854 			disk->d_genid = sc->sc_genid;
855 			g_raid3_update_metadata(disk);
856 		}
857 	}
858 }
859 
860 static int
861 g_raid3_idle(struct g_raid3_softc *sc, int acw)
862 {
863 	struct g_raid3_disk *disk;
864 	u_int i;
865 	int timeout;
866 
867 	g_topology_assert_not();
868 	sx_assert(&sc->sc_lock, SX_XLOCKED);
869 
870 	if (sc->sc_provider == NULL)
871 		return (0);
872 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) != 0)
873 		return (0);
874 	if (sc->sc_idle)
875 		return (0);
876 	if (sc->sc_writes > 0)
877 		return (0);
878 	if (acw > 0 || (acw == -1 && sc->sc_provider->acw > 0)) {
879 		timeout = g_raid3_idletime - (time_uptime - sc->sc_last_write);
880 		if (!g_raid3_shutdown && timeout > 0)
881 			return (timeout);
882 	}
883 	sc->sc_idle = 1;
884 	for (i = 0; i < sc->sc_ndisks; i++) {
885 		disk = &sc->sc_disks[i];
886 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE)
887 			continue;
888 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as clean.",
889 		    g_raid3_get_diskname(disk), sc->sc_name);
890 		disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
891 		g_raid3_update_metadata(disk);
892 	}
893 	return (0);
894 }
895 
896 static void
897 g_raid3_unidle(struct g_raid3_softc *sc)
898 {
899 	struct g_raid3_disk *disk;
900 	u_int i;
901 
902 	g_topology_assert_not();
903 	sx_assert(&sc->sc_lock, SX_XLOCKED);
904 
905 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) != 0)
906 		return;
907 	sc->sc_idle = 0;
908 	sc->sc_last_write = time_uptime;
909 	for (i = 0; i < sc->sc_ndisks; i++) {
910 		disk = &sc->sc_disks[i];
911 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE)
912 			continue;
913 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as dirty.",
914 		    g_raid3_get_diskname(disk), sc->sc_name);
915 		disk->d_flags |= G_RAID3_DISK_FLAG_DIRTY;
916 		g_raid3_update_metadata(disk);
917 	}
918 }
919 
920 /*
921  * Treat bio_driver1 field in parent bio as list head and field bio_caller1
922  * in child bio as pointer to the next element on the list.
923  */
924 #define	G_RAID3_HEAD_BIO(pbp)	(pbp)->bio_driver1
925 
926 #define	G_RAID3_NEXT_BIO(cbp)	(cbp)->bio_caller1
927 
928 #define	G_RAID3_FOREACH_BIO(pbp, bp)					\
929 	for ((bp) = G_RAID3_HEAD_BIO(pbp); (bp) != NULL;		\
930 	    (bp) = G_RAID3_NEXT_BIO(bp))
931 
932 #define	G_RAID3_FOREACH_SAFE_BIO(pbp, bp, tmpbp)			\
933 	for ((bp) = G_RAID3_HEAD_BIO(pbp);				\
934 	    (bp) != NULL && ((tmpbp) = G_RAID3_NEXT_BIO(bp), 1);	\
935 	    (bp) = (tmpbp))
936 
937 static void
938 g_raid3_init_bio(struct bio *pbp)
939 {
940 
941 	G_RAID3_HEAD_BIO(pbp) = NULL;
942 }
943 
944 static void
945 g_raid3_remove_bio(struct bio *cbp)
946 {
947 	struct bio *pbp, *bp;
948 
949 	pbp = cbp->bio_parent;
950 	if (G_RAID3_HEAD_BIO(pbp) == cbp)
951 		G_RAID3_HEAD_BIO(pbp) = G_RAID3_NEXT_BIO(cbp);
952 	else {
953 		G_RAID3_FOREACH_BIO(pbp, bp) {
954 			if (G_RAID3_NEXT_BIO(bp) == cbp) {
955 				G_RAID3_NEXT_BIO(bp) = G_RAID3_NEXT_BIO(cbp);
956 				break;
957 			}
958 		}
959 	}
960 	G_RAID3_NEXT_BIO(cbp) = NULL;
961 }
962 
963 static void
964 g_raid3_replace_bio(struct bio *sbp, struct bio *dbp)
965 {
966 	struct bio *pbp, *bp;
967 
968 	g_raid3_remove_bio(sbp);
969 	pbp = dbp->bio_parent;
970 	G_RAID3_NEXT_BIO(sbp) = G_RAID3_NEXT_BIO(dbp);
971 	if (G_RAID3_HEAD_BIO(pbp) == dbp)
972 		G_RAID3_HEAD_BIO(pbp) = sbp;
973 	else {
974 		G_RAID3_FOREACH_BIO(pbp, bp) {
975 			if (G_RAID3_NEXT_BIO(bp) == dbp) {
976 				G_RAID3_NEXT_BIO(bp) = sbp;
977 				break;
978 			}
979 		}
980 	}
981 	G_RAID3_NEXT_BIO(dbp) = NULL;
982 }
983 
984 static void
985 g_raid3_destroy_bio(struct g_raid3_softc *sc, struct bio *cbp)
986 {
987 	struct bio *bp, *pbp;
988 	size_t size;
989 
990 	pbp = cbp->bio_parent;
991 	pbp->bio_children--;
992 	KASSERT(cbp->bio_data != NULL, ("NULL bio_data"));
993 	size = pbp->bio_length / (sc->sc_ndisks - 1);
994 	g_raid3_free(sc, cbp->bio_data, size);
995 	if (G_RAID3_HEAD_BIO(pbp) == cbp) {
996 		G_RAID3_HEAD_BIO(pbp) = G_RAID3_NEXT_BIO(cbp);
997 		G_RAID3_NEXT_BIO(cbp) = NULL;
998 		g_destroy_bio(cbp);
999 	} else {
1000 		G_RAID3_FOREACH_BIO(pbp, bp) {
1001 			if (G_RAID3_NEXT_BIO(bp) == cbp)
1002 				break;
1003 		}
1004 		if (bp != NULL) {
1005 			KASSERT(G_RAID3_NEXT_BIO(bp) != NULL,
1006 			    ("NULL bp->bio_driver1"));
1007 			G_RAID3_NEXT_BIO(bp) = G_RAID3_NEXT_BIO(cbp);
1008 			G_RAID3_NEXT_BIO(cbp) = NULL;
1009 		}
1010 		g_destroy_bio(cbp);
1011 	}
1012 }
1013 
1014 static struct bio *
1015 g_raid3_clone_bio(struct g_raid3_softc *sc, struct bio *pbp)
1016 {
1017 	struct bio *bp, *cbp;
1018 	size_t size;
1019 	int memflag;
1020 
1021 	cbp = g_clone_bio(pbp);
1022 	if (cbp == NULL)
1023 		return (NULL);
1024 	size = pbp->bio_length / (sc->sc_ndisks - 1);
1025 	if ((pbp->bio_cflags & G_RAID3_BIO_CFLAG_REGULAR) != 0)
1026 		memflag = M_WAITOK;
1027 	else
1028 		memflag = M_NOWAIT;
1029 	cbp->bio_data = g_raid3_alloc(sc, size, memflag);
1030 	if (cbp->bio_data == NULL) {
1031 		pbp->bio_children--;
1032 		g_destroy_bio(cbp);
1033 		return (NULL);
1034 	}
1035 	G_RAID3_NEXT_BIO(cbp) = NULL;
1036 	if (G_RAID3_HEAD_BIO(pbp) == NULL)
1037 		G_RAID3_HEAD_BIO(pbp) = cbp;
1038 	else {
1039 		G_RAID3_FOREACH_BIO(pbp, bp) {
1040 			if (G_RAID3_NEXT_BIO(bp) == NULL) {
1041 				G_RAID3_NEXT_BIO(bp) = cbp;
1042 				break;
1043 			}
1044 		}
1045 	}
1046 	return (cbp);
1047 }
1048 
1049 static void
1050 g_raid3_scatter(struct bio *pbp)
1051 {
1052 	struct g_raid3_softc *sc;
1053 	struct g_raid3_disk *disk;
1054 	struct bio *bp, *cbp, *tmpbp;
1055 	off_t atom, cadd, padd, left;
1056 	int first;
1057 
1058 	sc = pbp->bio_to->geom->softc;
1059 	bp = NULL;
1060 	if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_NOPARITY) == 0) {
1061 		/*
1062 		 * Find bio for which we should calculate data.
1063 		 */
1064 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1065 			if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_PARITY) != 0) {
1066 				bp = cbp;
1067 				break;
1068 			}
1069 		}
1070 		KASSERT(bp != NULL, ("NULL parity bio."));
1071 	}
1072 	atom = sc->sc_sectorsize / (sc->sc_ndisks - 1);
1073 	cadd = padd = 0;
1074 	for (left = pbp->bio_length; left > 0; left -= sc->sc_sectorsize) {
1075 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1076 			if (cbp == bp)
1077 				continue;
1078 			bcopy(pbp->bio_data + padd, cbp->bio_data + cadd, atom);
1079 			padd += atom;
1080 		}
1081 		cadd += atom;
1082 	}
1083 	if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_NOPARITY) == 0) {
1084 		/*
1085 		 * Calculate parity.
1086 		 */
1087 		first = 1;
1088 		G_RAID3_FOREACH_SAFE_BIO(pbp, cbp, tmpbp) {
1089 			if (cbp == bp)
1090 				continue;
1091 			if (first) {
1092 				bcopy(cbp->bio_data, bp->bio_data,
1093 				    bp->bio_length);
1094 				first = 0;
1095 			} else {
1096 				g_raid3_xor(cbp->bio_data, bp->bio_data,
1097 				    bp->bio_length);
1098 			}
1099 			if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_NODISK) != 0)
1100 				g_raid3_destroy_bio(sc, cbp);
1101 		}
1102 	}
1103 	G_RAID3_FOREACH_SAFE_BIO(pbp, cbp, tmpbp) {
1104 		struct g_consumer *cp;
1105 
1106 		disk = cbp->bio_caller2;
1107 		cp = disk->d_consumer;
1108 		cbp->bio_to = cp->provider;
1109 		G_RAID3_LOGREQ(3, cbp, "Sending request.");
1110 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1111 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1112 		    cp->acr, cp->acw, cp->ace));
1113 		cp->index++;
1114 		sc->sc_writes++;
1115 		g_io_request(cbp, cp);
1116 	}
1117 }
1118 
1119 static void
1120 g_raid3_gather(struct bio *pbp)
1121 {
1122 	struct g_raid3_softc *sc;
1123 	struct g_raid3_disk *disk;
1124 	struct bio *xbp, *fbp, *cbp;
1125 	off_t atom, cadd, padd, left;
1126 
1127 	sc = pbp->bio_to->geom->softc;
1128 	/*
1129 	 * Find bio for which we have to calculate data.
1130 	 * While going through this path, check if all requests
1131 	 * succeeded, if not, deny whole request.
1132 	 * If we're in COMPLETE mode, we allow one request to fail,
1133 	 * so if we find one, we're sending it to the parity consumer.
1134 	 * If there are more failed requests, we deny whole request.
1135 	 */
1136 	xbp = fbp = NULL;
1137 	G_RAID3_FOREACH_BIO(pbp, cbp) {
1138 		if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_PARITY) != 0) {
1139 			KASSERT(xbp == NULL, ("More than one parity bio."));
1140 			xbp = cbp;
1141 		}
1142 		if (cbp->bio_error == 0)
1143 			continue;
1144 		/*
1145 		 * Found failed request.
1146 		 */
1147 		if (fbp == NULL) {
1148 			if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_DEGRADED) != 0) {
1149 				/*
1150 				 * We are already in degraded mode, so we can't
1151 				 * accept any failures.
1152 				 */
1153 				if (pbp->bio_error == 0)
1154 					pbp->bio_error = cbp->bio_error;
1155 			} else {
1156 				fbp = cbp;
1157 			}
1158 		} else {
1159 			/*
1160 			 * Next failed request, that's too many.
1161 			 */
1162 			if (pbp->bio_error == 0)
1163 				pbp->bio_error = fbp->bio_error;
1164 		}
1165 		disk = cbp->bio_caller2;
1166 		if (disk == NULL)
1167 			continue;
1168 		if ((disk->d_flags & G_RAID3_DISK_FLAG_BROKEN) == 0) {
1169 			disk->d_flags |= G_RAID3_DISK_FLAG_BROKEN;
1170 			G_RAID3_LOGREQ(0, cbp, "Request failed (error=%d).",
1171 			    cbp->bio_error);
1172 		} else {
1173 			G_RAID3_LOGREQ(1, cbp, "Request failed (error=%d).",
1174 			    cbp->bio_error);
1175 		}
1176 		if (g_raid3_disconnect_on_failure &&
1177 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1178 			sc->sc_bump_id |= G_RAID3_BUMP_GENID;
1179 			g_raid3_event_send(disk,
1180 			    G_RAID3_DISK_STATE_DISCONNECTED,
1181 			    G_RAID3_EVENT_DONTWAIT);
1182 		}
1183 	}
1184 	if (pbp->bio_error != 0)
1185 		goto finish;
1186 	if (fbp != NULL && (pbp->bio_pflags & G_RAID3_BIO_PFLAG_VERIFY) != 0) {
1187 		pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_VERIFY;
1188 		if (xbp != fbp)
1189 			g_raid3_replace_bio(xbp, fbp);
1190 		g_raid3_destroy_bio(sc, fbp);
1191 	} else if (fbp != NULL) {
1192 		struct g_consumer *cp;
1193 
1194 		/*
1195 		 * One request failed, so send the same request to
1196 		 * the parity consumer.
1197 		 */
1198 		disk = pbp->bio_driver2;
1199 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE) {
1200 			pbp->bio_error = fbp->bio_error;
1201 			goto finish;
1202 		}
1203 		pbp->bio_pflags |= G_RAID3_BIO_PFLAG_DEGRADED;
1204 		pbp->bio_inbed--;
1205 		fbp->bio_flags &= ~(BIO_DONE | BIO_ERROR);
1206 		if (disk->d_no == sc->sc_ndisks - 1)
1207 			fbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1208 		fbp->bio_error = 0;
1209 		fbp->bio_completed = 0;
1210 		fbp->bio_children = 0;
1211 		fbp->bio_inbed = 0;
1212 		cp = disk->d_consumer;
1213 		fbp->bio_caller2 = disk;
1214 		fbp->bio_to = cp->provider;
1215 		G_RAID3_LOGREQ(3, fbp, "Sending request (recover).");
1216 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1217 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1218 		    cp->acr, cp->acw, cp->ace));
1219 		cp->index++;
1220 		g_io_request(fbp, cp);
1221 		return;
1222 	}
1223 	if (xbp != NULL) {
1224 		/*
1225 		 * Calculate parity.
1226 		 */
1227 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1228 			if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_PARITY) != 0)
1229 				continue;
1230 			g_raid3_xor(cbp->bio_data, xbp->bio_data,
1231 			    xbp->bio_length);
1232 		}
1233 		xbp->bio_cflags &= ~G_RAID3_BIO_CFLAG_PARITY;
1234 		if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_VERIFY) != 0) {
1235 			if (!g_raid3_is_zero(xbp)) {
1236 				g_raid3_parity_mismatch++;
1237 				pbp->bio_error = EIO;
1238 				goto finish;
1239 			}
1240 			g_raid3_destroy_bio(sc, xbp);
1241 		}
1242 	}
1243 	atom = sc->sc_sectorsize / (sc->sc_ndisks - 1);
1244 	cadd = padd = 0;
1245 	for (left = pbp->bio_length; left > 0; left -= sc->sc_sectorsize) {
1246 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1247 			bcopy(cbp->bio_data + cadd, pbp->bio_data + padd, atom);
1248 			pbp->bio_completed += atom;
1249 			padd += atom;
1250 		}
1251 		cadd += atom;
1252 	}
1253 finish:
1254 	if (pbp->bio_error == 0)
1255 		G_RAID3_LOGREQ(3, pbp, "Request finished.");
1256 	else {
1257 		if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_VERIFY) != 0)
1258 			G_RAID3_LOGREQ(1, pbp, "Verification error.");
1259 		else
1260 			G_RAID3_LOGREQ(0, pbp, "Request failed.");
1261 	}
1262 	pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_MASK;
1263 	while ((cbp = G_RAID3_HEAD_BIO(pbp)) != NULL)
1264 		g_raid3_destroy_bio(sc, cbp);
1265 	g_io_deliver(pbp, pbp->bio_error);
1266 }
1267 
1268 static void
1269 g_raid3_done(struct bio *bp)
1270 {
1271 	struct g_raid3_softc *sc;
1272 
1273 	sc = bp->bio_from->geom->softc;
1274 	bp->bio_cflags |= G_RAID3_BIO_CFLAG_REGULAR;
1275 	G_RAID3_LOGREQ(3, bp, "Regular request done (error=%d).", bp->bio_error);
1276 	mtx_lock(&sc->sc_queue_mtx);
1277 	bioq_insert_head(&sc->sc_queue, bp);
1278 	mtx_unlock(&sc->sc_queue_mtx);
1279 	wakeup(sc);
1280 	wakeup(&sc->sc_queue);
1281 }
1282 
1283 static void
1284 g_raid3_regular_request(struct bio *cbp)
1285 {
1286 	struct g_raid3_softc *sc;
1287 	struct g_raid3_disk *disk;
1288 	struct bio *pbp;
1289 
1290 	g_topology_assert_not();
1291 
1292 	pbp = cbp->bio_parent;
1293 	sc = pbp->bio_to->geom->softc;
1294 	cbp->bio_from->index--;
1295 	if (cbp->bio_cmd == BIO_WRITE)
1296 		sc->sc_writes--;
1297 	disk = cbp->bio_from->private;
1298 	if (disk == NULL) {
1299 		g_topology_lock();
1300 		g_raid3_kill_consumer(sc, cbp->bio_from);
1301 		g_topology_unlock();
1302 	}
1303 
1304 	G_RAID3_LOGREQ(3, cbp, "Request finished.");
1305 	pbp->bio_inbed++;
1306 	KASSERT(pbp->bio_inbed <= pbp->bio_children,
1307 	    ("bio_inbed (%u) is bigger than bio_children (%u).", pbp->bio_inbed,
1308 	    pbp->bio_children));
1309 	if (pbp->bio_inbed != pbp->bio_children)
1310 		return;
1311 	switch (pbp->bio_cmd) {
1312 	case BIO_READ:
1313 		g_raid3_gather(pbp);
1314 		break;
1315 	case BIO_WRITE:
1316 	case BIO_DELETE:
1317 	    {
1318 		int error = 0;
1319 
1320 		pbp->bio_completed = pbp->bio_length;
1321 		while ((cbp = G_RAID3_HEAD_BIO(pbp)) != NULL) {
1322 			if (cbp->bio_error == 0) {
1323 				g_raid3_destroy_bio(sc, cbp);
1324 				continue;
1325 			}
1326 
1327 			if (error == 0)
1328 				error = cbp->bio_error;
1329 			else if (pbp->bio_error == 0) {
1330 				/*
1331 				 * Next failed request, that's too many.
1332 				 */
1333 				pbp->bio_error = error;
1334 			}
1335 
1336 			disk = cbp->bio_caller2;
1337 			if (disk == NULL) {
1338 				g_raid3_destroy_bio(sc, cbp);
1339 				continue;
1340 			}
1341 
1342 			if ((disk->d_flags & G_RAID3_DISK_FLAG_BROKEN) == 0) {
1343 				disk->d_flags |= G_RAID3_DISK_FLAG_BROKEN;
1344 				G_RAID3_LOGREQ(0, cbp,
1345 				    "Request failed (error=%d).",
1346 				    cbp->bio_error);
1347 			} else {
1348 				G_RAID3_LOGREQ(1, cbp,
1349 				    "Request failed (error=%d).",
1350 				    cbp->bio_error);
1351 			}
1352 			if (g_raid3_disconnect_on_failure &&
1353 			    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1354 				sc->sc_bump_id |= G_RAID3_BUMP_GENID;
1355 				g_raid3_event_send(disk,
1356 				    G_RAID3_DISK_STATE_DISCONNECTED,
1357 				    G_RAID3_EVENT_DONTWAIT);
1358 			}
1359 			g_raid3_destroy_bio(sc, cbp);
1360 		}
1361 		if (pbp->bio_error == 0)
1362 			G_RAID3_LOGREQ(3, pbp, "Request finished.");
1363 		else
1364 			G_RAID3_LOGREQ(0, pbp, "Request failed.");
1365 		pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_DEGRADED;
1366 		pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_NOPARITY;
1367 		bioq_remove(&sc->sc_inflight, pbp);
1368 		/* Release delayed sync requests if possible. */
1369 		g_raid3_sync_release(sc);
1370 		g_io_deliver(pbp, pbp->bio_error);
1371 		break;
1372 	    }
1373 	}
1374 }
1375 
1376 static void
1377 g_raid3_sync_done(struct bio *bp)
1378 {
1379 	struct g_raid3_softc *sc;
1380 
1381 	G_RAID3_LOGREQ(3, bp, "Synchronization request delivered.");
1382 	sc = bp->bio_from->geom->softc;
1383 	bp->bio_cflags |= G_RAID3_BIO_CFLAG_SYNC;
1384 	mtx_lock(&sc->sc_queue_mtx);
1385 	bioq_insert_head(&sc->sc_queue, bp);
1386 	mtx_unlock(&sc->sc_queue_mtx);
1387 	wakeup(sc);
1388 	wakeup(&sc->sc_queue);
1389 }
1390 
1391 static void
1392 g_raid3_flush(struct g_raid3_softc *sc, struct bio *bp)
1393 {
1394 	struct bio_queue_head queue;
1395 	struct g_raid3_disk *disk;
1396 	struct g_consumer *cp;
1397 	struct bio *cbp;
1398 	u_int i;
1399 
1400 	bioq_init(&queue);
1401 	for (i = 0; i < sc->sc_ndisks; i++) {
1402 		disk = &sc->sc_disks[i];
1403 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE)
1404 			continue;
1405 		cbp = g_clone_bio(bp);
1406 		if (cbp == NULL) {
1407 			for (cbp = bioq_first(&queue); cbp != NULL;
1408 			    cbp = bioq_first(&queue)) {
1409 				bioq_remove(&queue, cbp);
1410 				g_destroy_bio(cbp);
1411 			}
1412 			if (bp->bio_error == 0)
1413 				bp->bio_error = ENOMEM;
1414 			g_io_deliver(bp, bp->bio_error);
1415 			return;
1416 		}
1417 		bioq_insert_tail(&queue, cbp);
1418 		cbp->bio_done = g_std_done;
1419 		cbp->bio_caller1 = disk;
1420 		cbp->bio_to = disk->d_consumer->provider;
1421 	}
1422 	for (cbp = bioq_first(&queue); cbp != NULL; cbp = bioq_first(&queue)) {
1423 		bioq_remove(&queue, cbp);
1424 		G_RAID3_LOGREQ(3, cbp, "Sending request.");
1425 		disk = cbp->bio_caller1;
1426 		cbp->bio_caller1 = NULL;
1427 		cp = disk->d_consumer;
1428 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1429 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1430 		    cp->acr, cp->acw, cp->ace));
1431 		g_io_request(cbp, disk->d_consumer);
1432 	}
1433 }
1434 
1435 static void
1436 g_raid3_start(struct bio *bp)
1437 {
1438 	struct g_raid3_softc *sc;
1439 
1440 	sc = bp->bio_to->geom->softc;
1441 	/*
1442 	 * If sc == NULL or there are no valid disks, provider's error
1443 	 * should be set and g_raid3_start() should not be called at all.
1444 	 */
1445 	KASSERT(sc != NULL && (sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
1446 	    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE),
1447 	    ("Provider's error should be set (error=%d)(device=%s).",
1448 	    bp->bio_to->error, bp->bio_to->name));
1449 	G_RAID3_LOGREQ(3, bp, "Request received.");
1450 
1451 	switch (bp->bio_cmd) {
1452 	case BIO_READ:
1453 	case BIO_WRITE:
1454 	case BIO_DELETE:
1455 		break;
1456 	case BIO_FLUSH:
1457 		g_raid3_flush(sc, bp);
1458 		return;
1459 	case BIO_GETATTR:
1460 	default:
1461 		g_io_deliver(bp, EOPNOTSUPP);
1462 		return;
1463 	}
1464 	mtx_lock(&sc->sc_queue_mtx);
1465 	bioq_insert_tail(&sc->sc_queue, bp);
1466 	mtx_unlock(&sc->sc_queue_mtx);
1467 	G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1468 	wakeup(sc);
1469 }
1470 
1471 /*
1472  * Return TRUE if the given request is colliding with a in-progress
1473  * synchronization request.
1474  */
1475 static int
1476 g_raid3_sync_collision(struct g_raid3_softc *sc, struct bio *bp)
1477 {
1478 	struct g_raid3_disk *disk;
1479 	struct bio *sbp;
1480 	off_t rstart, rend, sstart, send;
1481 	int i;
1482 
1483 	disk = sc->sc_syncdisk;
1484 	if (disk == NULL)
1485 		return (0);
1486 	rstart = bp->bio_offset;
1487 	rend = bp->bio_offset + bp->bio_length;
1488 	for (i = 0; i < g_raid3_syncreqs; i++) {
1489 		sbp = disk->d_sync.ds_bios[i];
1490 		if (sbp == NULL)
1491 			continue;
1492 		sstart = sbp->bio_offset;
1493 		send = sbp->bio_length;
1494 		if (sbp->bio_cmd == BIO_WRITE) {
1495 			sstart *= sc->sc_ndisks - 1;
1496 			send *= sc->sc_ndisks - 1;
1497 		}
1498 		send += sstart;
1499 		if (rend > sstart && rstart < send)
1500 			return (1);
1501 	}
1502 	return (0);
1503 }
1504 
1505 /*
1506  * Return TRUE if the given sync request is colliding with a in-progress regular
1507  * request.
1508  */
1509 static int
1510 g_raid3_regular_collision(struct g_raid3_softc *sc, struct bio *sbp)
1511 {
1512 	off_t rstart, rend, sstart, send;
1513 	struct bio *bp;
1514 
1515 	if (sc->sc_syncdisk == NULL)
1516 		return (0);
1517 	sstart = sbp->bio_offset;
1518 	send = sstart + sbp->bio_length;
1519 	TAILQ_FOREACH(bp, &sc->sc_inflight.queue, bio_queue) {
1520 		rstart = bp->bio_offset;
1521 		rend = bp->bio_offset + bp->bio_length;
1522 		if (rend > sstart && rstart < send)
1523 			return (1);
1524 	}
1525 	return (0);
1526 }
1527 
1528 /*
1529  * Puts request onto delayed queue.
1530  */
1531 static void
1532 g_raid3_regular_delay(struct g_raid3_softc *sc, struct bio *bp)
1533 {
1534 
1535 	G_RAID3_LOGREQ(2, bp, "Delaying request.");
1536 	bioq_insert_head(&sc->sc_regular_delayed, bp);
1537 }
1538 
1539 /*
1540  * Puts synchronization request onto delayed queue.
1541  */
1542 static void
1543 g_raid3_sync_delay(struct g_raid3_softc *sc, struct bio *bp)
1544 {
1545 
1546 	G_RAID3_LOGREQ(2, bp, "Delaying synchronization request.");
1547 	bioq_insert_tail(&sc->sc_sync_delayed, bp);
1548 }
1549 
1550 /*
1551  * Releases delayed regular requests which don't collide anymore with sync
1552  * requests.
1553  */
1554 static void
1555 g_raid3_regular_release(struct g_raid3_softc *sc)
1556 {
1557 	struct bio *bp, *bp2;
1558 
1559 	TAILQ_FOREACH_SAFE(bp, &sc->sc_regular_delayed.queue, bio_queue, bp2) {
1560 		if (g_raid3_sync_collision(sc, bp))
1561 			continue;
1562 		bioq_remove(&sc->sc_regular_delayed, bp);
1563 		G_RAID3_LOGREQ(2, bp, "Releasing delayed request (%p).", bp);
1564 		mtx_lock(&sc->sc_queue_mtx);
1565 		bioq_insert_head(&sc->sc_queue, bp);
1566 #if 0
1567 		/*
1568 		 * wakeup() is not needed, because this function is called from
1569 		 * the worker thread.
1570 		 */
1571 		wakeup(&sc->sc_queue);
1572 #endif
1573 		mtx_unlock(&sc->sc_queue_mtx);
1574 	}
1575 }
1576 
1577 /*
1578  * Releases delayed sync requests which don't collide anymore with regular
1579  * requests.
1580  */
1581 static void
1582 g_raid3_sync_release(struct g_raid3_softc *sc)
1583 {
1584 	struct bio *bp, *bp2;
1585 
1586 	TAILQ_FOREACH_SAFE(bp, &sc->sc_sync_delayed.queue, bio_queue, bp2) {
1587 		if (g_raid3_regular_collision(sc, bp))
1588 			continue;
1589 		bioq_remove(&sc->sc_sync_delayed, bp);
1590 		G_RAID3_LOGREQ(2, bp,
1591 		    "Releasing delayed synchronization request.");
1592 		g_io_request(bp, bp->bio_from);
1593 	}
1594 }
1595 
1596 /*
1597  * Handle synchronization requests.
1598  * Every synchronization request is two-steps process: first, READ request is
1599  * send to active provider and then WRITE request (with read data) to the provider
1600  * beeing synchronized. When WRITE is finished, new synchronization request is
1601  * send.
1602  */
1603 static void
1604 g_raid3_sync_request(struct bio *bp)
1605 {
1606 	struct g_raid3_softc *sc;
1607 	struct g_raid3_disk *disk;
1608 
1609 	bp->bio_from->index--;
1610 	sc = bp->bio_from->geom->softc;
1611 	disk = bp->bio_from->private;
1612 	if (disk == NULL) {
1613 		sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1614 		g_topology_lock();
1615 		g_raid3_kill_consumer(sc, bp->bio_from);
1616 		g_topology_unlock();
1617 		free(bp->bio_data, M_RAID3);
1618 		g_destroy_bio(bp);
1619 		sx_xlock(&sc->sc_lock);
1620 		return;
1621 	}
1622 
1623 	/*
1624 	 * Synchronization request.
1625 	 */
1626 	switch (bp->bio_cmd) {
1627 	case BIO_READ:
1628 	    {
1629 		struct g_consumer *cp;
1630 		u_char *dst, *src;
1631 		off_t left;
1632 		u_int atom;
1633 
1634 		if (bp->bio_error != 0) {
1635 			G_RAID3_LOGREQ(0, bp,
1636 			    "Synchronization request failed (error=%d).",
1637 			    bp->bio_error);
1638 			g_destroy_bio(bp);
1639 			return;
1640 		}
1641 		G_RAID3_LOGREQ(3, bp, "Synchronization request finished.");
1642 		atom = sc->sc_sectorsize / (sc->sc_ndisks - 1);
1643 		dst = src = bp->bio_data;
1644 		if (disk->d_no == sc->sc_ndisks - 1) {
1645 			u_int n;
1646 
1647 			/* Parity component. */
1648 			for (left = bp->bio_length; left > 0;
1649 			    left -= sc->sc_sectorsize) {
1650 				bcopy(src, dst, atom);
1651 				src += atom;
1652 				for (n = 1; n < sc->sc_ndisks - 1; n++) {
1653 					g_raid3_xor(src, dst, atom);
1654 					src += atom;
1655 				}
1656 				dst += atom;
1657 			}
1658 		} else {
1659 			/* Regular component. */
1660 			src += atom * disk->d_no;
1661 			for (left = bp->bio_length; left > 0;
1662 			    left -= sc->sc_sectorsize) {
1663 				bcopy(src, dst, atom);
1664 				src += sc->sc_sectorsize;
1665 				dst += atom;
1666 			}
1667 		}
1668 		bp->bio_driver1 = bp->bio_driver2 = NULL;
1669 		bp->bio_pflags = 0;
1670 		bp->bio_offset /= sc->sc_ndisks - 1;
1671 		bp->bio_length /= sc->sc_ndisks - 1;
1672 		bp->bio_cmd = BIO_WRITE;
1673 		bp->bio_cflags = 0;
1674 		bp->bio_children = bp->bio_inbed = 0;
1675 		cp = disk->d_consumer;
1676 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1677 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1678 		    cp->acr, cp->acw, cp->ace));
1679 		cp->index++;
1680 		g_io_request(bp, cp);
1681 		return;
1682 	    }
1683 	case BIO_WRITE:
1684 	    {
1685 		struct g_raid3_disk_sync *sync;
1686 		off_t boffset, moffset;
1687 		void *data;
1688 		int i;
1689 
1690 		if (bp->bio_error != 0) {
1691 			G_RAID3_LOGREQ(0, bp,
1692 			    "Synchronization request failed (error=%d).",
1693 			    bp->bio_error);
1694 			g_destroy_bio(bp);
1695 			sc->sc_bump_id |= G_RAID3_BUMP_GENID;
1696 			g_raid3_event_send(disk,
1697 			    G_RAID3_DISK_STATE_DISCONNECTED,
1698 			    G_RAID3_EVENT_DONTWAIT);
1699 			return;
1700 		}
1701 		G_RAID3_LOGREQ(3, bp, "Synchronization request finished.");
1702 		sync = &disk->d_sync;
1703 		if (sync->ds_offset == sc->sc_mediasize / (sc->sc_ndisks - 1) ||
1704 		    sync->ds_consumer == NULL ||
1705 		    (sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
1706 			/* Don't send more synchronization requests. */
1707 			sync->ds_inflight--;
1708 			if (sync->ds_bios != NULL) {
1709 				i = (int)(uintptr_t)bp->bio_caller1;
1710 				sync->ds_bios[i] = NULL;
1711 			}
1712 			free(bp->bio_data, M_RAID3);
1713 			g_destroy_bio(bp);
1714 			if (sync->ds_inflight > 0)
1715 				return;
1716 			if (sync->ds_consumer == NULL ||
1717 			    (sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
1718 				return;
1719 			}
1720 			/*
1721 			 * Disk up-to-date, activate it.
1722 			 */
1723 			g_raid3_event_send(disk, G_RAID3_DISK_STATE_ACTIVE,
1724 			    G_RAID3_EVENT_DONTWAIT);
1725 			return;
1726 		}
1727 
1728 		/* Send next synchronization request. */
1729 		data = bp->bio_data;
1730 		bzero(bp, sizeof(*bp));
1731 		bp->bio_cmd = BIO_READ;
1732 		bp->bio_offset = sync->ds_offset * (sc->sc_ndisks - 1);
1733 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1734 		sync->ds_offset += bp->bio_length / (sc->sc_ndisks - 1);
1735 		bp->bio_done = g_raid3_sync_done;
1736 		bp->bio_data = data;
1737 		bp->bio_from = sync->ds_consumer;
1738 		bp->bio_to = sc->sc_provider;
1739 		G_RAID3_LOGREQ(3, bp, "Sending synchronization request.");
1740 		sync->ds_consumer->index++;
1741 		/*
1742 		 * Delay the request if it is colliding with a regular request.
1743 		 */
1744 		if (g_raid3_regular_collision(sc, bp))
1745 			g_raid3_sync_delay(sc, bp);
1746 		else
1747 			g_io_request(bp, sync->ds_consumer);
1748 
1749 		/* Release delayed requests if possible. */
1750 		g_raid3_regular_release(sc);
1751 
1752 		/* Find the smallest offset. */
1753 		moffset = sc->sc_mediasize;
1754 		for (i = 0; i < g_raid3_syncreqs; i++) {
1755 			bp = sync->ds_bios[i];
1756 			boffset = bp->bio_offset;
1757 			if (bp->bio_cmd == BIO_WRITE)
1758 				boffset *= sc->sc_ndisks - 1;
1759 			if (boffset < moffset)
1760 				moffset = boffset;
1761 		}
1762 		if (sync->ds_offset_done + (MAXPHYS * 100) < moffset) {
1763 			/* Update offset_done on every 100 blocks. */
1764 			sync->ds_offset_done = moffset;
1765 			g_raid3_update_metadata(disk);
1766 		}
1767 		return;
1768 	    }
1769 	default:
1770 		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1771 		    bp->bio_cmd, sc->sc_name));
1772 		break;
1773 	}
1774 }
1775 
1776 static int
1777 g_raid3_register_request(struct bio *pbp)
1778 {
1779 	struct g_raid3_softc *sc;
1780 	struct g_raid3_disk *disk;
1781 	struct g_consumer *cp;
1782 	struct bio *cbp, *tmpbp;
1783 	off_t offset, length;
1784 	u_int n, ndisks;
1785 	int round_robin, verify;
1786 
1787 	ndisks = 0;
1788 	sc = pbp->bio_to->geom->softc;
1789 	if ((pbp->bio_cflags & G_RAID3_BIO_CFLAG_REGSYNC) != 0 &&
1790 	    sc->sc_syncdisk == NULL) {
1791 		g_io_deliver(pbp, EIO);
1792 		return (0);
1793 	}
1794 	g_raid3_init_bio(pbp);
1795 	length = pbp->bio_length / (sc->sc_ndisks - 1);
1796 	offset = pbp->bio_offset / (sc->sc_ndisks - 1);
1797 	round_robin = verify = 0;
1798 	switch (pbp->bio_cmd) {
1799 	case BIO_READ:
1800 		if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_VERIFY) != 0 &&
1801 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1802 			pbp->bio_pflags |= G_RAID3_BIO_PFLAG_VERIFY;
1803 			verify = 1;
1804 			ndisks = sc->sc_ndisks;
1805 		} else {
1806 			verify = 0;
1807 			ndisks = sc->sc_ndisks - 1;
1808 		}
1809 		if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_ROUND_ROBIN) != 0 &&
1810 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1811 			round_robin = 1;
1812 		} else {
1813 			round_robin = 0;
1814 		}
1815 		KASSERT(!round_robin || !verify,
1816 		    ("ROUND-ROBIN and VERIFY are mutually exclusive."));
1817 		pbp->bio_driver2 = &sc->sc_disks[sc->sc_ndisks - 1];
1818 		break;
1819 	case BIO_WRITE:
1820 	case BIO_DELETE:
1821 		/*
1822 		 * Delay the request if it is colliding with a synchronization
1823 		 * request.
1824 		 */
1825 		if (g_raid3_sync_collision(sc, pbp)) {
1826 			g_raid3_regular_delay(sc, pbp);
1827 			return (0);
1828 		}
1829 
1830 		if (sc->sc_idle)
1831 			g_raid3_unidle(sc);
1832 		else
1833 			sc->sc_last_write = time_uptime;
1834 
1835 		ndisks = sc->sc_ndisks;
1836 		break;
1837 	}
1838 	for (n = 0; n < ndisks; n++) {
1839 		disk = &sc->sc_disks[n];
1840 		cbp = g_raid3_clone_bio(sc, pbp);
1841 		if (cbp == NULL) {
1842 			while ((cbp = G_RAID3_HEAD_BIO(pbp)) != NULL)
1843 				g_raid3_destroy_bio(sc, cbp);
1844 			/*
1845 			 * To prevent deadlock, we must run back up
1846 			 * with the ENOMEM for failed requests of any
1847 			 * of our consumers.  Our own sync requests
1848 			 * can stick around, as they are finite.
1849 			 */
1850 			if ((pbp->bio_cflags &
1851 			    G_RAID3_BIO_CFLAG_REGULAR) != 0) {
1852 				g_io_deliver(pbp, ENOMEM);
1853 				return (0);
1854 			}
1855 			return (ENOMEM);
1856 		}
1857 		cbp->bio_offset = offset;
1858 		cbp->bio_length = length;
1859 		cbp->bio_done = g_raid3_done;
1860 		switch (pbp->bio_cmd) {
1861 		case BIO_READ:
1862 			if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE) {
1863 				/*
1864 				 * Replace invalid component with the parity
1865 				 * component.
1866 				 */
1867 				disk = &sc->sc_disks[sc->sc_ndisks - 1];
1868 				cbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1869 				pbp->bio_pflags |= G_RAID3_BIO_PFLAG_DEGRADED;
1870 			} else if (round_robin &&
1871 			    disk->d_no == sc->sc_round_robin) {
1872 				/*
1873 				 * In round-robin mode skip one data component
1874 				 * and use parity component when reading.
1875 				 */
1876 				pbp->bio_driver2 = disk;
1877 				disk = &sc->sc_disks[sc->sc_ndisks - 1];
1878 				cbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1879 				sc->sc_round_robin++;
1880 				round_robin = 0;
1881 			} else if (verify && disk->d_no == sc->sc_ndisks - 1) {
1882 				cbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1883 			}
1884 			break;
1885 		case BIO_WRITE:
1886 		case BIO_DELETE:
1887 			if (disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
1888 			    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
1889 				if (n == ndisks - 1) {
1890 					/*
1891 					 * Active parity component, mark it as such.
1892 					 */
1893 					cbp->bio_cflags |=
1894 					    G_RAID3_BIO_CFLAG_PARITY;
1895 				}
1896 			} else {
1897 				pbp->bio_pflags |= G_RAID3_BIO_PFLAG_DEGRADED;
1898 				if (n == ndisks - 1) {
1899 					/*
1900 					 * Parity component is not connected,
1901 					 * so destroy its request.
1902 					 */
1903 					pbp->bio_pflags |=
1904 					    G_RAID3_BIO_PFLAG_NOPARITY;
1905 					g_raid3_destroy_bio(sc, cbp);
1906 					cbp = NULL;
1907 				} else {
1908 					cbp->bio_cflags |=
1909 					    G_RAID3_BIO_CFLAG_NODISK;
1910 					disk = NULL;
1911 				}
1912 			}
1913 			break;
1914 		}
1915 		if (cbp != NULL)
1916 			cbp->bio_caller2 = disk;
1917 	}
1918 	switch (pbp->bio_cmd) {
1919 	case BIO_READ:
1920 		if (round_robin) {
1921 			/*
1922 			 * If we are in round-robin mode and 'round_robin' is
1923 			 * still 1, it means, that we skipped parity component
1924 			 * for this read and must reset sc_round_robin field.
1925 			 */
1926 			sc->sc_round_robin = 0;
1927 		}
1928 		G_RAID3_FOREACH_SAFE_BIO(pbp, cbp, tmpbp) {
1929 			disk = cbp->bio_caller2;
1930 			cp = disk->d_consumer;
1931 			cbp->bio_to = cp->provider;
1932 			G_RAID3_LOGREQ(3, cbp, "Sending request.");
1933 			KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1934 			    ("Consumer %s not opened (r%dw%de%d).",
1935 			    cp->provider->name, cp->acr, cp->acw, cp->ace));
1936 			cp->index++;
1937 			g_io_request(cbp, cp);
1938 		}
1939 		break;
1940 	case BIO_WRITE:
1941 	case BIO_DELETE:
1942 		/*
1943 		 * Put request onto inflight queue, so we can check if new
1944 		 * synchronization requests don't collide with it.
1945 		 */
1946 		bioq_insert_tail(&sc->sc_inflight, pbp);
1947 
1948 		/*
1949 		 * Bump syncid on first write.
1950 		 */
1951 		if ((sc->sc_bump_id & G_RAID3_BUMP_SYNCID) != 0) {
1952 			sc->sc_bump_id &= ~G_RAID3_BUMP_SYNCID;
1953 			g_raid3_bump_syncid(sc);
1954 		}
1955 		g_raid3_scatter(pbp);
1956 		break;
1957 	}
1958 	return (0);
1959 }
1960 
1961 static int
1962 g_raid3_can_destroy(struct g_raid3_softc *sc)
1963 {
1964 	struct g_geom *gp;
1965 	struct g_consumer *cp;
1966 
1967 	g_topology_assert();
1968 	gp = sc->sc_geom;
1969 	if (gp->softc == NULL)
1970 		return (1);
1971 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1972 		if (g_raid3_is_busy(sc, cp))
1973 			return (0);
1974 	}
1975 	gp = sc->sc_sync.ds_geom;
1976 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1977 		if (g_raid3_is_busy(sc, cp))
1978 			return (0);
1979 	}
1980 	G_RAID3_DEBUG(2, "No I/O requests for %s, it can be destroyed.",
1981 	    sc->sc_name);
1982 	return (1);
1983 }
1984 
1985 static int
1986 g_raid3_try_destroy(struct g_raid3_softc *sc)
1987 {
1988 
1989 	g_topology_assert_not();
1990 	sx_assert(&sc->sc_lock, SX_XLOCKED);
1991 
1992 	if (sc->sc_rootmount != NULL) {
1993 		G_RAID3_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
1994 		    sc->sc_rootmount);
1995 		root_mount_rel(sc->sc_rootmount);
1996 		sc->sc_rootmount = NULL;
1997 	}
1998 
1999 	g_topology_lock();
2000 	if (!g_raid3_can_destroy(sc)) {
2001 		g_topology_unlock();
2002 		return (0);
2003 	}
2004 	sc->sc_geom->softc = NULL;
2005 	sc->sc_sync.ds_geom->softc = NULL;
2006 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_WAIT) != 0) {
2007 		g_topology_unlock();
2008 		G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__,
2009 		    &sc->sc_worker);
2010 		/* Unlock sc_lock here, as it can be destroyed after wakeup. */
2011 		sx_xunlock(&sc->sc_lock);
2012 		wakeup(&sc->sc_worker);
2013 		sc->sc_worker = NULL;
2014 	} else {
2015 		g_topology_unlock();
2016 		g_raid3_destroy_device(sc);
2017 		free(sc->sc_disks, M_RAID3);
2018 		free(sc, M_RAID3);
2019 	}
2020 	return (1);
2021 }
2022 
2023 /*
2024  * Worker thread.
2025  */
2026 static void
2027 g_raid3_worker(void *arg)
2028 {
2029 	struct g_raid3_softc *sc;
2030 	struct g_raid3_event *ep;
2031 	struct bio *bp;
2032 	int timeout;
2033 
2034 	sc = arg;
2035 	thread_lock(curthread);
2036 	sched_prio(curthread, PRIBIO);
2037 	thread_unlock(curthread);
2038 
2039 	sx_xlock(&sc->sc_lock);
2040 	for (;;) {
2041 		G_RAID3_DEBUG(5, "%s: Let's see...", __func__);
2042 		/*
2043 		 * First take a look at events.
2044 		 * This is important to handle events before any I/O requests.
2045 		 */
2046 		ep = g_raid3_event_get(sc);
2047 		if (ep != NULL) {
2048 			g_raid3_event_remove(sc, ep);
2049 			if ((ep->e_flags & G_RAID3_EVENT_DEVICE) != 0) {
2050 				/* Update only device status. */
2051 				G_RAID3_DEBUG(3,
2052 				    "Running event for device %s.",
2053 				    sc->sc_name);
2054 				ep->e_error = 0;
2055 				g_raid3_update_device(sc, 1);
2056 			} else {
2057 				/* Update disk status. */
2058 				G_RAID3_DEBUG(3, "Running event for disk %s.",
2059 				     g_raid3_get_diskname(ep->e_disk));
2060 				ep->e_error = g_raid3_update_disk(ep->e_disk,
2061 				    ep->e_state);
2062 				if (ep->e_error == 0)
2063 					g_raid3_update_device(sc, 0);
2064 			}
2065 			if ((ep->e_flags & G_RAID3_EVENT_DONTWAIT) != 0) {
2066 				KASSERT(ep->e_error == 0,
2067 				    ("Error cannot be handled."));
2068 				g_raid3_event_free(ep);
2069 			} else {
2070 				ep->e_flags |= G_RAID3_EVENT_DONE;
2071 				G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__,
2072 				    ep);
2073 				mtx_lock(&sc->sc_events_mtx);
2074 				wakeup(ep);
2075 				mtx_unlock(&sc->sc_events_mtx);
2076 			}
2077 			if ((sc->sc_flags &
2078 			    G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
2079 				if (g_raid3_try_destroy(sc)) {
2080 					curthread->td_pflags &= ~TDP_GEOM;
2081 					G_RAID3_DEBUG(1, "Thread exiting.");
2082 					kproc_exit(0);
2083 				}
2084 			}
2085 			G_RAID3_DEBUG(5, "%s: I'm here 1.", __func__);
2086 			continue;
2087 		}
2088 		/*
2089 		 * Check if we can mark array as CLEAN and if we can't take
2090 		 * how much seconds should we wait.
2091 		 */
2092 		timeout = g_raid3_idle(sc, -1);
2093 		/*
2094 		 * Now I/O requests.
2095 		 */
2096 		/* Get first request from the queue. */
2097 		mtx_lock(&sc->sc_queue_mtx);
2098 		bp = bioq_first(&sc->sc_queue);
2099 		if (bp == NULL) {
2100 			if ((sc->sc_flags &
2101 			    G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
2102 				mtx_unlock(&sc->sc_queue_mtx);
2103 				if (g_raid3_try_destroy(sc)) {
2104 					curthread->td_pflags &= ~TDP_GEOM;
2105 					G_RAID3_DEBUG(1, "Thread exiting.");
2106 					kproc_exit(0);
2107 				}
2108 				mtx_lock(&sc->sc_queue_mtx);
2109 			}
2110 			sx_xunlock(&sc->sc_lock);
2111 			/*
2112 			 * XXX: We can miss an event here, because an event
2113 			 *      can be added without sx-device-lock and without
2114 			 *      mtx-queue-lock. Maybe I should just stop using
2115 			 *      dedicated mutex for events synchronization and
2116 			 *      stick with the queue lock?
2117 			 *      The event will hang here until next I/O request
2118 			 *      or next event is received.
2119 			 */
2120 			MSLEEP(sc, &sc->sc_queue_mtx, PRIBIO | PDROP, "r3:w1",
2121 			    timeout * hz);
2122 			sx_xlock(&sc->sc_lock);
2123 			G_RAID3_DEBUG(5, "%s: I'm here 4.", __func__);
2124 			continue;
2125 		}
2126 process:
2127 		bioq_remove(&sc->sc_queue, bp);
2128 		mtx_unlock(&sc->sc_queue_mtx);
2129 
2130 		if (bp->bio_from->geom == sc->sc_sync.ds_geom &&
2131 		    (bp->bio_cflags & G_RAID3_BIO_CFLAG_SYNC) != 0) {
2132 			g_raid3_sync_request(bp);	/* READ */
2133 		} else if (bp->bio_to != sc->sc_provider) {
2134 			if ((bp->bio_cflags & G_RAID3_BIO_CFLAG_REGULAR) != 0)
2135 				g_raid3_regular_request(bp);
2136 			else if ((bp->bio_cflags & G_RAID3_BIO_CFLAG_SYNC) != 0)
2137 				g_raid3_sync_request(bp);	/* WRITE */
2138 			else {
2139 				KASSERT(0,
2140 				    ("Invalid request cflags=0x%hhx to=%s.",
2141 				    bp->bio_cflags, bp->bio_to->name));
2142 			}
2143 		} else if (g_raid3_register_request(bp) != 0) {
2144 			mtx_lock(&sc->sc_queue_mtx);
2145 			bioq_insert_head(&sc->sc_queue, bp);
2146 			/*
2147 			 * We are short in memory, let see if there are finished
2148 			 * request we can free.
2149 			 */
2150 			TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
2151 				if (bp->bio_cflags & G_RAID3_BIO_CFLAG_REGULAR)
2152 					goto process;
2153 			}
2154 			/*
2155 			 * No finished regular request, so at least keep
2156 			 * synchronization running.
2157 			 */
2158 			TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
2159 				if (bp->bio_cflags & G_RAID3_BIO_CFLAG_SYNC)
2160 					goto process;
2161 			}
2162 			sx_xunlock(&sc->sc_lock);
2163 			MSLEEP(&sc->sc_queue, &sc->sc_queue_mtx, PRIBIO | PDROP,
2164 			    "r3:lowmem", hz / 10);
2165 			sx_xlock(&sc->sc_lock);
2166 		}
2167 		G_RAID3_DEBUG(5, "%s: I'm here 9.", __func__);
2168 	}
2169 }
2170 
2171 static void
2172 g_raid3_update_idle(struct g_raid3_softc *sc, struct g_raid3_disk *disk)
2173 {
2174 
2175 	sx_assert(&sc->sc_lock, SX_LOCKED);
2176 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) != 0)
2177 		return;
2178 	if (!sc->sc_idle && (disk->d_flags & G_RAID3_DISK_FLAG_DIRTY) == 0) {
2179 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as dirty.",
2180 		    g_raid3_get_diskname(disk), sc->sc_name);
2181 		disk->d_flags |= G_RAID3_DISK_FLAG_DIRTY;
2182 	} else if (sc->sc_idle &&
2183 	    (disk->d_flags & G_RAID3_DISK_FLAG_DIRTY) != 0) {
2184 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as clean.",
2185 		    g_raid3_get_diskname(disk), sc->sc_name);
2186 		disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2187 	}
2188 }
2189 
2190 static void
2191 g_raid3_sync_start(struct g_raid3_softc *sc)
2192 {
2193 	struct g_raid3_disk *disk;
2194 	struct g_consumer *cp;
2195 	struct bio *bp;
2196 	int error;
2197 	u_int n;
2198 
2199 	g_topology_assert_not();
2200 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2201 
2202 	KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED,
2203 	    ("Device not in DEGRADED state (%s, %u).", sc->sc_name,
2204 	    sc->sc_state));
2205 	KASSERT(sc->sc_syncdisk == NULL, ("Syncdisk is not NULL (%s, %u).",
2206 	    sc->sc_name, sc->sc_state));
2207 	disk = NULL;
2208 	for (n = 0; n < sc->sc_ndisks; n++) {
2209 		if (sc->sc_disks[n].d_state != G_RAID3_DISK_STATE_SYNCHRONIZING)
2210 			continue;
2211 		disk = &sc->sc_disks[n];
2212 		break;
2213 	}
2214 	if (disk == NULL)
2215 		return;
2216 
2217 	sx_xunlock(&sc->sc_lock);
2218 	g_topology_lock();
2219 	cp = g_new_consumer(sc->sc_sync.ds_geom);
2220 	error = g_attach(cp, sc->sc_provider);
2221 	KASSERT(error == 0,
2222 	    ("Cannot attach to %s (error=%d).", sc->sc_name, error));
2223 	error = g_access(cp, 1, 0, 0);
2224 	KASSERT(error == 0, ("Cannot open %s (error=%d).", sc->sc_name, error));
2225 	g_topology_unlock();
2226 	sx_xlock(&sc->sc_lock);
2227 
2228 	G_RAID3_DEBUG(0, "Device %s: rebuilding provider %s.", sc->sc_name,
2229 	    g_raid3_get_diskname(disk));
2230 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) == 0)
2231 		disk->d_flags |= G_RAID3_DISK_FLAG_DIRTY;
2232 	KASSERT(disk->d_sync.ds_consumer == NULL,
2233 	    ("Sync consumer already exists (device=%s, disk=%s).",
2234 	    sc->sc_name, g_raid3_get_diskname(disk)));
2235 
2236 	disk->d_sync.ds_consumer = cp;
2237 	disk->d_sync.ds_consumer->private = disk;
2238 	disk->d_sync.ds_consumer->index = 0;
2239 	sc->sc_syncdisk = disk;
2240 
2241 	/*
2242 	 * Allocate memory for synchronization bios and initialize them.
2243 	 */
2244 	disk->d_sync.ds_bios = malloc(sizeof(struct bio *) * g_raid3_syncreqs,
2245 	    M_RAID3, M_WAITOK);
2246 	for (n = 0; n < g_raid3_syncreqs; n++) {
2247 		bp = g_alloc_bio();
2248 		disk->d_sync.ds_bios[n] = bp;
2249 		bp->bio_parent = NULL;
2250 		bp->bio_cmd = BIO_READ;
2251 		bp->bio_data = malloc(MAXPHYS, M_RAID3, M_WAITOK);
2252 		bp->bio_cflags = 0;
2253 		bp->bio_offset = disk->d_sync.ds_offset * (sc->sc_ndisks - 1);
2254 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
2255 		disk->d_sync.ds_offset += bp->bio_length / (sc->sc_ndisks - 1);
2256 		bp->bio_done = g_raid3_sync_done;
2257 		bp->bio_from = disk->d_sync.ds_consumer;
2258 		bp->bio_to = sc->sc_provider;
2259 		bp->bio_caller1 = (void *)(uintptr_t)n;
2260 	}
2261 
2262 	/* Set the number of in-flight synchronization requests. */
2263 	disk->d_sync.ds_inflight = g_raid3_syncreqs;
2264 
2265 	/*
2266 	 * Fire off first synchronization requests.
2267 	 */
2268 	for (n = 0; n < g_raid3_syncreqs; n++) {
2269 		bp = disk->d_sync.ds_bios[n];
2270 		G_RAID3_LOGREQ(3, bp, "Sending synchronization request.");
2271 		disk->d_sync.ds_consumer->index++;
2272 		/*
2273 		 * Delay the request if it is colliding with a regular request.
2274 		 */
2275 		if (g_raid3_regular_collision(sc, bp))
2276 			g_raid3_sync_delay(sc, bp);
2277 		else
2278 			g_io_request(bp, disk->d_sync.ds_consumer);
2279 	}
2280 }
2281 
2282 /*
2283  * Stop synchronization process.
2284  * type: 0 - synchronization finished
2285  *       1 - synchronization stopped
2286  */
2287 static void
2288 g_raid3_sync_stop(struct g_raid3_softc *sc, int type)
2289 {
2290 	struct g_raid3_disk *disk;
2291 	struct g_consumer *cp;
2292 
2293 	g_topology_assert_not();
2294 	sx_assert(&sc->sc_lock, SX_LOCKED);
2295 
2296 	KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED,
2297 	    ("Device not in DEGRADED state (%s, %u).", sc->sc_name,
2298 	    sc->sc_state));
2299 	disk = sc->sc_syncdisk;
2300 	sc->sc_syncdisk = NULL;
2301 	KASSERT(disk != NULL, ("No disk was synchronized (%s).", sc->sc_name));
2302 	KASSERT(disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING,
2303 	    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2304 	    g_raid3_disk_state2str(disk->d_state)));
2305 	if (disk->d_sync.ds_consumer == NULL)
2306 		return;
2307 
2308 	if (type == 0) {
2309 		G_RAID3_DEBUG(0, "Device %s: rebuilding provider %s finished.",
2310 		    sc->sc_name, g_raid3_get_diskname(disk));
2311 	} else /* if (type == 1) */ {
2312 		G_RAID3_DEBUG(0, "Device %s: rebuilding provider %s stopped.",
2313 		    sc->sc_name, g_raid3_get_diskname(disk));
2314 	}
2315 	free(disk->d_sync.ds_bios, M_RAID3);
2316 	disk->d_sync.ds_bios = NULL;
2317 	cp = disk->d_sync.ds_consumer;
2318 	disk->d_sync.ds_consumer = NULL;
2319 	disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2320 	sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
2321 	g_topology_lock();
2322 	g_raid3_kill_consumer(sc, cp);
2323 	g_topology_unlock();
2324 	sx_xlock(&sc->sc_lock);
2325 }
2326 
2327 static void
2328 g_raid3_launch_provider(struct g_raid3_softc *sc)
2329 {
2330 	struct g_provider *pp;
2331 	struct g_raid3_disk *disk;
2332 	int n;
2333 
2334 	sx_assert(&sc->sc_lock, SX_LOCKED);
2335 
2336 	g_topology_lock();
2337 	pp = g_new_providerf(sc->sc_geom, "raid3/%s", sc->sc_name);
2338 	pp->mediasize = sc->sc_mediasize;
2339 	pp->sectorsize = sc->sc_sectorsize;
2340 	pp->stripesize = 0;
2341 	pp->stripeoffset = 0;
2342 	for (n = 0; n < sc->sc_ndisks; n++) {
2343 		disk = &sc->sc_disks[n];
2344 		if (disk->d_consumer && disk->d_consumer->provider &&
2345 		    disk->d_consumer->provider->stripesize > pp->stripesize) {
2346 			pp->stripesize = disk->d_consumer->provider->stripesize;
2347 			pp->stripeoffset = disk->d_consumer->provider->stripeoffset;
2348 		}
2349 	}
2350 	pp->stripesize *= sc->sc_ndisks - 1;
2351 	pp->stripeoffset *= sc->sc_ndisks - 1;
2352 	sc->sc_provider = pp;
2353 	g_error_provider(pp, 0);
2354 	g_topology_unlock();
2355 	G_RAID3_DEBUG(0, "Device %s launched (%u/%u).", pp->name,
2356 	    g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE), sc->sc_ndisks);
2357 
2358 	if (sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED)
2359 		g_raid3_sync_start(sc);
2360 }
2361 
2362 static void
2363 g_raid3_destroy_provider(struct g_raid3_softc *sc)
2364 {
2365 	struct bio *bp;
2366 
2367 	g_topology_assert_not();
2368 	KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
2369 	    sc->sc_name));
2370 
2371 	g_topology_lock();
2372 	g_error_provider(sc->sc_provider, ENXIO);
2373 	mtx_lock(&sc->sc_queue_mtx);
2374 	while ((bp = bioq_first(&sc->sc_queue)) != NULL) {
2375 		bioq_remove(&sc->sc_queue, bp);
2376 		g_io_deliver(bp, ENXIO);
2377 	}
2378 	mtx_unlock(&sc->sc_queue_mtx);
2379 	G_RAID3_DEBUG(0, "Device %s: provider %s destroyed.", sc->sc_name,
2380 	    sc->sc_provider->name);
2381 	sc->sc_provider->flags |= G_PF_WITHER;
2382 	g_orphan_provider(sc->sc_provider, ENXIO);
2383 	g_topology_unlock();
2384 	sc->sc_provider = NULL;
2385 	if (sc->sc_syncdisk != NULL)
2386 		g_raid3_sync_stop(sc, 1);
2387 }
2388 
2389 static void
2390 g_raid3_go(void *arg)
2391 {
2392 	struct g_raid3_softc *sc;
2393 
2394 	sc = arg;
2395 	G_RAID3_DEBUG(0, "Force device %s start due to timeout.", sc->sc_name);
2396 	g_raid3_event_send(sc, 0,
2397 	    G_RAID3_EVENT_DONTWAIT | G_RAID3_EVENT_DEVICE);
2398 }
2399 
2400 static u_int
2401 g_raid3_determine_state(struct g_raid3_disk *disk)
2402 {
2403 	struct g_raid3_softc *sc;
2404 	u_int state;
2405 
2406 	sc = disk->d_softc;
2407 	if (sc->sc_syncid == disk->d_sync.ds_syncid) {
2408 		if ((disk->d_flags &
2409 		    G_RAID3_DISK_FLAG_SYNCHRONIZING) == 0) {
2410 			/* Disk does not need synchronization. */
2411 			state = G_RAID3_DISK_STATE_ACTIVE;
2412 		} else {
2413 			if ((sc->sc_flags &
2414 			     G_RAID3_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2415 			    (disk->d_flags &
2416 			     G_RAID3_DISK_FLAG_FORCE_SYNC) != 0) {
2417 				/*
2418 				 * We can start synchronization from
2419 				 * the stored offset.
2420 				 */
2421 				state = G_RAID3_DISK_STATE_SYNCHRONIZING;
2422 			} else {
2423 				state = G_RAID3_DISK_STATE_STALE;
2424 			}
2425 		}
2426 	} else if (disk->d_sync.ds_syncid < sc->sc_syncid) {
2427 		/*
2428 		 * Reset all synchronization data for this disk,
2429 		 * because if it even was synchronized, it was
2430 		 * synchronized to disks with different syncid.
2431 		 */
2432 		disk->d_flags |= G_RAID3_DISK_FLAG_SYNCHRONIZING;
2433 		disk->d_sync.ds_offset = 0;
2434 		disk->d_sync.ds_offset_done = 0;
2435 		disk->d_sync.ds_syncid = sc->sc_syncid;
2436 		if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2437 		    (disk->d_flags & G_RAID3_DISK_FLAG_FORCE_SYNC) != 0) {
2438 			state = G_RAID3_DISK_STATE_SYNCHRONIZING;
2439 		} else {
2440 			state = G_RAID3_DISK_STATE_STALE;
2441 		}
2442 	} else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
2443 		/*
2444 		 * Not good, NOT GOOD!
2445 		 * It means that device was started on stale disks
2446 		 * and more fresh disk just arrive.
2447 		 * If there were writes, device is broken, sorry.
2448 		 * I think the best choice here is don't touch
2449 		 * this disk and inform the user loudly.
2450 		 */
2451 		G_RAID3_DEBUG(0, "Device %s was started before the freshest "
2452 		    "disk (%s) arrives!! It will not be connected to the "
2453 		    "running device.", sc->sc_name,
2454 		    g_raid3_get_diskname(disk));
2455 		g_raid3_destroy_disk(disk);
2456 		state = G_RAID3_DISK_STATE_NONE;
2457 		/* Return immediately, because disk was destroyed. */
2458 		return (state);
2459 	}
2460 	G_RAID3_DEBUG(3, "State for %s disk: %s.",
2461 	    g_raid3_get_diskname(disk), g_raid3_disk_state2str(state));
2462 	return (state);
2463 }
2464 
2465 /*
2466  * Update device state.
2467  */
2468 static void
2469 g_raid3_update_device(struct g_raid3_softc *sc, boolean_t force)
2470 {
2471 	struct g_raid3_disk *disk;
2472 	u_int state;
2473 
2474 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2475 
2476 	switch (sc->sc_state) {
2477 	case G_RAID3_DEVICE_STATE_STARTING:
2478 	    {
2479 		u_int n, ndirty, ndisks, genid, syncid;
2480 
2481 		KASSERT(sc->sc_provider == NULL,
2482 		    ("Non-NULL provider in STARTING state (%s).", sc->sc_name));
2483 		/*
2484 		 * Are we ready? We are, if all disks are connected or
2485 		 * one disk is missing and 'force' is true.
2486 		 */
2487 		if (g_raid3_ndisks(sc, -1) + force == sc->sc_ndisks) {
2488 			if (!force)
2489 				callout_drain(&sc->sc_callout);
2490 		} else {
2491 			if (force) {
2492 				/*
2493 				 * Timeout expired, so destroy device.
2494 				 */
2495 				sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
2496 				G_RAID3_DEBUG(1, "root_mount_rel[%u] %p",
2497 				    __LINE__, sc->sc_rootmount);
2498 				root_mount_rel(sc->sc_rootmount);
2499 				sc->sc_rootmount = NULL;
2500 			}
2501 			return;
2502 		}
2503 
2504 		/*
2505 		 * Find the biggest genid.
2506 		 */
2507 		genid = 0;
2508 		for (n = 0; n < sc->sc_ndisks; n++) {
2509 			disk = &sc->sc_disks[n];
2510 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2511 				continue;
2512 			if (disk->d_genid > genid)
2513 				genid = disk->d_genid;
2514 		}
2515 		sc->sc_genid = genid;
2516 		/*
2517 		 * Remove all disks without the biggest genid.
2518 		 */
2519 		for (n = 0; n < sc->sc_ndisks; n++) {
2520 			disk = &sc->sc_disks[n];
2521 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2522 				continue;
2523 			if (disk->d_genid < genid) {
2524 				G_RAID3_DEBUG(0,
2525 				    "Component %s (device %s) broken, skipping.",
2526 				    g_raid3_get_diskname(disk), sc->sc_name);
2527 				g_raid3_destroy_disk(disk);
2528 			}
2529 		}
2530 
2531 		/*
2532 		 * There must be at least 'sc->sc_ndisks - 1' components
2533 		 * with the same syncid and without SYNCHRONIZING flag.
2534 		 */
2535 
2536 		/*
2537 		 * Find the biggest syncid, number of valid components and
2538 		 * number of dirty components.
2539 		 */
2540 		ndirty = ndisks = syncid = 0;
2541 		for (n = 0; n < sc->sc_ndisks; n++) {
2542 			disk = &sc->sc_disks[n];
2543 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2544 				continue;
2545 			if ((disk->d_flags & G_RAID3_DISK_FLAG_DIRTY) != 0)
2546 				ndirty++;
2547 			if (disk->d_sync.ds_syncid > syncid) {
2548 				syncid = disk->d_sync.ds_syncid;
2549 				ndisks = 0;
2550 			} else if (disk->d_sync.ds_syncid < syncid) {
2551 				continue;
2552 			}
2553 			if ((disk->d_flags &
2554 			    G_RAID3_DISK_FLAG_SYNCHRONIZING) != 0) {
2555 				continue;
2556 			}
2557 			ndisks++;
2558 		}
2559 		/*
2560 		 * Do we have enough valid components?
2561 		 */
2562 		if (ndisks + 1 < sc->sc_ndisks) {
2563 			G_RAID3_DEBUG(0,
2564 			    "Device %s is broken, too few valid components.",
2565 			    sc->sc_name);
2566 			sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
2567 			return;
2568 		}
2569 		/*
2570 		 * If there is one DIRTY component and all disks are present,
2571 		 * mark it for synchronization. If there is more than one DIRTY
2572 		 * component, mark parity component for synchronization.
2573 		 */
2574 		if (ndisks == sc->sc_ndisks && ndirty == 1) {
2575 			for (n = 0; n < sc->sc_ndisks; n++) {
2576 				disk = &sc->sc_disks[n];
2577 				if ((disk->d_flags &
2578 				    G_RAID3_DISK_FLAG_DIRTY) == 0) {
2579 					continue;
2580 				}
2581 				disk->d_flags |=
2582 				    G_RAID3_DISK_FLAG_SYNCHRONIZING;
2583 			}
2584 		} else if (ndisks == sc->sc_ndisks && ndirty > 1) {
2585 			disk = &sc->sc_disks[sc->sc_ndisks - 1];
2586 			disk->d_flags |= G_RAID3_DISK_FLAG_SYNCHRONIZING;
2587 		}
2588 
2589 		sc->sc_syncid = syncid;
2590 		if (force) {
2591 			/* Remember to bump syncid on first write. */
2592 			sc->sc_bump_id |= G_RAID3_BUMP_SYNCID;
2593 		}
2594 		if (ndisks == sc->sc_ndisks)
2595 			state = G_RAID3_DEVICE_STATE_COMPLETE;
2596 		else /* if (ndisks == sc->sc_ndisks - 1) */
2597 			state = G_RAID3_DEVICE_STATE_DEGRADED;
2598 		G_RAID3_DEBUG(1, "Device %s state changed from %s to %s.",
2599 		    sc->sc_name, g_raid3_device_state2str(sc->sc_state),
2600 		    g_raid3_device_state2str(state));
2601 		sc->sc_state = state;
2602 		for (n = 0; n < sc->sc_ndisks; n++) {
2603 			disk = &sc->sc_disks[n];
2604 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2605 				continue;
2606 			state = g_raid3_determine_state(disk);
2607 			g_raid3_event_send(disk, state, G_RAID3_EVENT_DONTWAIT);
2608 			if (state == G_RAID3_DISK_STATE_STALE)
2609 				sc->sc_bump_id |= G_RAID3_BUMP_SYNCID;
2610 		}
2611 		break;
2612 	    }
2613 	case G_RAID3_DEVICE_STATE_DEGRADED:
2614 		/*
2615 		 * Genid need to be bumped immediately, so do it here.
2616 		 */
2617 		if ((sc->sc_bump_id & G_RAID3_BUMP_GENID) != 0) {
2618 			sc->sc_bump_id &= ~G_RAID3_BUMP_GENID;
2619 			g_raid3_bump_genid(sc);
2620 		}
2621 
2622 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_NEW) > 0)
2623 			return;
2624 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) <
2625 		    sc->sc_ndisks - 1) {
2626 			if (sc->sc_provider != NULL)
2627 				g_raid3_destroy_provider(sc);
2628 			sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
2629 			return;
2630 		}
2631 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) ==
2632 		    sc->sc_ndisks) {
2633 			state = G_RAID3_DEVICE_STATE_COMPLETE;
2634 			G_RAID3_DEBUG(1,
2635 			    "Device %s state changed from %s to %s.",
2636 			    sc->sc_name, g_raid3_device_state2str(sc->sc_state),
2637 			    g_raid3_device_state2str(state));
2638 			sc->sc_state = state;
2639 		}
2640 		if (sc->sc_provider == NULL)
2641 			g_raid3_launch_provider(sc);
2642 		if (sc->sc_rootmount != NULL) {
2643 			G_RAID3_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2644 			    sc->sc_rootmount);
2645 			root_mount_rel(sc->sc_rootmount);
2646 			sc->sc_rootmount = NULL;
2647 		}
2648 		break;
2649 	case G_RAID3_DEVICE_STATE_COMPLETE:
2650 		/*
2651 		 * Genid need to be bumped immediately, so do it here.
2652 		 */
2653 		if ((sc->sc_bump_id & G_RAID3_BUMP_GENID) != 0) {
2654 			sc->sc_bump_id &= ~G_RAID3_BUMP_GENID;
2655 			g_raid3_bump_genid(sc);
2656 		}
2657 
2658 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_NEW) > 0)
2659 			return;
2660 		KASSERT(g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) >=
2661 		    sc->sc_ndisks - 1,
2662 		    ("Too few ACTIVE components in COMPLETE state (device %s).",
2663 		    sc->sc_name));
2664 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) ==
2665 		    sc->sc_ndisks - 1) {
2666 			state = G_RAID3_DEVICE_STATE_DEGRADED;
2667 			G_RAID3_DEBUG(1,
2668 			    "Device %s state changed from %s to %s.",
2669 			    sc->sc_name, g_raid3_device_state2str(sc->sc_state),
2670 			    g_raid3_device_state2str(state));
2671 			sc->sc_state = state;
2672 		}
2673 		if (sc->sc_provider == NULL)
2674 			g_raid3_launch_provider(sc);
2675 		if (sc->sc_rootmount != NULL) {
2676 			G_RAID3_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2677 			    sc->sc_rootmount);
2678 			root_mount_rel(sc->sc_rootmount);
2679 			sc->sc_rootmount = NULL;
2680 		}
2681 		break;
2682 	default:
2683 		KASSERT(1 == 0, ("Wrong device state (%s, %s).", sc->sc_name,
2684 		    g_raid3_device_state2str(sc->sc_state)));
2685 		break;
2686 	}
2687 }
2688 
2689 /*
2690  * Update disk state and device state if needed.
2691  */
2692 #define	DISK_STATE_CHANGED()	G_RAID3_DEBUG(1,			\
2693 	"Disk %s state changed from %s to %s (device %s).",		\
2694 	g_raid3_get_diskname(disk),					\
2695 	g_raid3_disk_state2str(disk->d_state),				\
2696 	g_raid3_disk_state2str(state), sc->sc_name)
2697 static int
2698 g_raid3_update_disk(struct g_raid3_disk *disk, u_int state)
2699 {
2700 	struct g_raid3_softc *sc;
2701 
2702 	sc = disk->d_softc;
2703 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2704 
2705 again:
2706 	G_RAID3_DEBUG(3, "Changing disk %s state from %s to %s.",
2707 	    g_raid3_get_diskname(disk), g_raid3_disk_state2str(disk->d_state),
2708 	    g_raid3_disk_state2str(state));
2709 	switch (state) {
2710 	case G_RAID3_DISK_STATE_NEW:
2711 		/*
2712 		 * Possible scenarios:
2713 		 * 1. New disk arrive.
2714 		 */
2715 		/* Previous state should be NONE. */
2716 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NONE,
2717 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2718 		    g_raid3_disk_state2str(disk->d_state)));
2719 		DISK_STATE_CHANGED();
2720 
2721 		disk->d_state = state;
2722 		G_RAID3_DEBUG(1, "Device %s: provider %s detected.",
2723 		    sc->sc_name, g_raid3_get_diskname(disk));
2724 		if (sc->sc_state == G_RAID3_DEVICE_STATE_STARTING)
2725 			break;
2726 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2727 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2728 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2729 		    g_raid3_device_state2str(sc->sc_state),
2730 		    g_raid3_get_diskname(disk),
2731 		    g_raid3_disk_state2str(disk->d_state)));
2732 		state = g_raid3_determine_state(disk);
2733 		if (state != G_RAID3_DISK_STATE_NONE)
2734 			goto again;
2735 		break;
2736 	case G_RAID3_DISK_STATE_ACTIVE:
2737 		/*
2738 		 * Possible scenarios:
2739 		 * 1. New disk does not need synchronization.
2740 		 * 2. Synchronization process finished successfully.
2741 		 */
2742 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2743 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2744 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2745 		    g_raid3_device_state2str(sc->sc_state),
2746 		    g_raid3_get_diskname(disk),
2747 		    g_raid3_disk_state2str(disk->d_state)));
2748 		/* Previous state should be NEW or SYNCHRONIZING. */
2749 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW ||
2750 		    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING,
2751 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2752 		    g_raid3_disk_state2str(disk->d_state)));
2753 		DISK_STATE_CHANGED();
2754 
2755 		if (disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
2756 			disk->d_flags &= ~G_RAID3_DISK_FLAG_SYNCHRONIZING;
2757 			disk->d_flags &= ~G_RAID3_DISK_FLAG_FORCE_SYNC;
2758 			g_raid3_sync_stop(sc, 0);
2759 		}
2760 		disk->d_state = state;
2761 		disk->d_sync.ds_offset = 0;
2762 		disk->d_sync.ds_offset_done = 0;
2763 		g_raid3_update_idle(sc, disk);
2764 		g_raid3_update_metadata(disk);
2765 		G_RAID3_DEBUG(1, "Device %s: provider %s activated.",
2766 		    sc->sc_name, g_raid3_get_diskname(disk));
2767 		break;
2768 	case G_RAID3_DISK_STATE_STALE:
2769 		/*
2770 		 * Possible scenarios:
2771 		 * 1. Stale disk was connected.
2772 		 */
2773 		/* Previous state should be NEW. */
2774 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW,
2775 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2776 		    g_raid3_disk_state2str(disk->d_state)));
2777 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2778 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
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 		/*
2784 		 * STALE state is only possible if device is marked
2785 		 * NOAUTOSYNC.
2786 		 */
2787 		KASSERT((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOAUTOSYNC) != 0,
2788 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2789 		    g_raid3_device_state2str(sc->sc_state),
2790 		    g_raid3_get_diskname(disk),
2791 		    g_raid3_disk_state2str(disk->d_state)));
2792 		DISK_STATE_CHANGED();
2793 
2794 		disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2795 		disk->d_state = state;
2796 		g_raid3_update_metadata(disk);
2797 		G_RAID3_DEBUG(0, "Device %s: provider %s is stale.",
2798 		    sc->sc_name, g_raid3_get_diskname(disk));
2799 		break;
2800 	case G_RAID3_DISK_STATE_SYNCHRONIZING:
2801 		/*
2802 		 * Possible scenarios:
2803 		 * 1. Disk which needs synchronization was connected.
2804 		 */
2805 		/* Previous state should be NEW. */
2806 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW,
2807 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2808 		    g_raid3_disk_state2str(disk->d_state)));
2809 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2810 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2811 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2812 		    g_raid3_device_state2str(sc->sc_state),
2813 		    g_raid3_get_diskname(disk),
2814 		    g_raid3_disk_state2str(disk->d_state)));
2815 		DISK_STATE_CHANGED();
2816 
2817 		if (disk->d_state == G_RAID3_DISK_STATE_NEW)
2818 			disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2819 		disk->d_state = state;
2820 		if (sc->sc_provider != NULL) {
2821 			g_raid3_sync_start(sc);
2822 			g_raid3_update_metadata(disk);
2823 		}
2824 		break;
2825 	case G_RAID3_DISK_STATE_DISCONNECTED:
2826 		/*
2827 		 * Possible scenarios:
2828 		 * 1. Device wasn't running yet, but disk disappear.
2829 		 * 2. Disk was active and disapppear.
2830 		 * 3. Disk disappear during synchronization process.
2831 		 */
2832 		if (sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2833 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
2834 			/*
2835 			 * Previous state should be ACTIVE, STALE or
2836 			 * SYNCHRONIZING.
2837 			 */
2838 			KASSERT(disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
2839 			    disk->d_state == G_RAID3_DISK_STATE_STALE ||
2840 			    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING,
2841 			    ("Wrong disk state (%s, %s).",
2842 			    g_raid3_get_diskname(disk),
2843 			    g_raid3_disk_state2str(disk->d_state)));
2844 		} else if (sc->sc_state == G_RAID3_DEVICE_STATE_STARTING) {
2845 			/* Previous state should be NEW. */
2846 			KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW,
2847 			    ("Wrong disk state (%s, %s).",
2848 			    g_raid3_get_diskname(disk),
2849 			    g_raid3_disk_state2str(disk->d_state)));
2850 			/*
2851 			 * Reset bumping syncid if disk disappeared in STARTING
2852 			 * state.
2853 			 */
2854 			if ((sc->sc_bump_id & G_RAID3_BUMP_SYNCID) != 0)
2855 				sc->sc_bump_id &= ~G_RAID3_BUMP_SYNCID;
2856 #ifdef	INVARIANTS
2857 		} else {
2858 			KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
2859 			    sc->sc_name,
2860 			    g_raid3_device_state2str(sc->sc_state),
2861 			    g_raid3_get_diskname(disk),
2862 			    g_raid3_disk_state2str(disk->d_state)));
2863 #endif
2864 		}
2865 		DISK_STATE_CHANGED();
2866 		G_RAID3_DEBUG(0, "Device %s: provider %s disconnected.",
2867 		    sc->sc_name, g_raid3_get_diskname(disk));
2868 
2869 		g_raid3_destroy_disk(disk);
2870 		break;
2871 	default:
2872 		KASSERT(1 == 0, ("Unknown state (%u).", state));
2873 		break;
2874 	}
2875 	return (0);
2876 }
2877 #undef	DISK_STATE_CHANGED
2878 
2879 int
2880 g_raid3_read_metadata(struct g_consumer *cp, struct g_raid3_metadata *md)
2881 {
2882 	struct g_provider *pp;
2883 	u_char *buf;
2884 	int error;
2885 
2886 	g_topology_assert();
2887 
2888 	error = g_access(cp, 1, 0, 0);
2889 	if (error != 0)
2890 		return (error);
2891 	pp = cp->provider;
2892 	g_topology_unlock();
2893 	/* Metadata are stored on last sector. */
2894 	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
2895 	    &error);
2896 	g_topology_lock();
2897 	g_access(cp, -1, 0, 0);
2898 	if (buf == NULL) {
2899 		G_RAID3_DEBUG(1, "Cannot read metadata from %s (error=%d).",
2900 		    cp->provider->name, error);
2901 		return (error);
2902 	}
2903 
2904 	/* Decode metadata. */
2905 	error = raid3_metadata_decode(buf, md);
2906 	g_free(buf);
2907 	if (strcmp(md->md_magic, G_RAID3_MAGIC) != 0)
2908 		return (EINVAL);
2909 	if (md->md_version > G_RAID3_VERSION) {
2910 		G_RAID3_DEBUG(0,
2911 		    "Kernel module is too old to handle metadata from %s.",
2912 		    cp->provider->name);
2913 		return (EINVAL);
2914 	}
2915 	if (error != 0) {
2916 		G_RAID3_DEBUG(1, "MD5 metadata hash mismatch for provider %s.",
2917 		    cp->provider->name);
2918 		return (error);
2919 	}
2920 	if (md->md_sectorsize > MAXPHYS) {
2921 		G_RAID3_DEBUG(0, "The blocksize is too big.");
2922 		return (EINVAL);
2923 	}
2924 
2925 	return (0);
2926 }
2927 
2928 static int
2929 g_raid3_check_metadata(struct g_raid3_softc *sc, struct g_provider *pp,
2930     struct g_raid3_metadata *md)
2931 {
2932 
2933 	if (md->md_no >= sc->sc_ndisks) {
2934 		G_RAID3_DEBUG(1, "Invalid disk %s number (no=%u), skipping.",
2935 		    pp->name, md->md_no);
2936 		return (EINVAL);
2937 	}
2938 	if (sc->sc_disks[md->md_no].d_state != G_RAID3_DISK_STATE_NODISK) {
2939 		G_RAID3_DEBUG(1, "Disk %s (no=%u) already exists, skipping.",
2940 		    pp->name, md->md_no);
2941 		return (EEXIST);
2942 	}
2943 	if (md->md_all != sc->sc_ndisks) {
2944 		G_RAID3_DEBUG(1,
2945 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2946 		    "md_all", pp->name, sc->sc_name);
2947 		return (EINVAL);
2948 	}
2949 	if ((md->md_mediasize % md->md_sectorsize) != 0) {
2950 		G_RAID3_DEBUG(1, "Invalid metadata (mediasize %% sectorsize != "
2951 		    "0) on disk %s (device %s), skipping.", pp->name,
2952 		    sc->sc_name);
2953 		return (EINVAL);
2954 	}
2955 	if (md->md_mediasize != sc->sc_mediasize) {
2956 		G_RAID3_DEBUG(1,
2957 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2958 		    "md_mediasize", pp->name, sc->sc_name);
2959 		return (EINVAL);
2960 	}
2961 	if ((md->md_mediasize % (sc->sc_ndisks - 1)) != 0) {
2962 		G_RAID3_DEBUG(1,
2963 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2964 		    "md_mediasize", pp->name, sc->sc_name);
2965 		return (EINVAL);
2966 	}
2967 	if ((sc->sc_mediasize / (sc->sc_ndisks - 1)) > pp->mediasize) {
2968 		G_RAID3_DEBUG(1,
2969 		    "Invalid size of disk %s (device %s), skipping.", pp->name,
2970 		    sc->sc_name);
2971 		return (EINVAL);
2972 	}
2973 	if ((md->md_sectorsize / pp->sectorsize) < sc->sc_ndisks - 1) {
2974 		G_RAID3_DEBUG(1,
2975 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2976 		    "md_sectorsize", pp->name, sc->sc_name);
2977 		return (EINVAL);
2978 	}
2979 	if (md->md_sectorsize != sc->sc_sectorsize) {
2980 		G_RAID3_DEBUG(1,
2981 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2982 		    "md_sectorsize", pp->name, sc->sc_name);
2983 		return (EINVAL);
2984 	}
2985 	if ((sc->sc_sectorsize % pp->sectorsize) != 0) {
2986 		G_RAID3_DEBUG(1,
2987 		    "Invalid sector size of disk %s (device %s), skipping.",
2988 		    pp->name, sc->sc_name);
2989 		return (EINVAL);
2990 	}
2991 	if ((md->md_mflags & ~G_RAID3_DEVICE_FLAG_MASK) != 0) {
2992 		G_RAID3_DEBUG(1,
2993 		    "Invalid device flags on disk %s (device %s), skipping.",
2994 		    pp->name, sc->sc_name);
2995 		return (EINVAL);
2996 	}
2997 	if ((md->md_mflags & G_RAID3_DEVICE_FLAG_VERIFY) != 0 &&
2998 	    (md->md_mflags & G_RAID3_DEVICE_FLAG_ROUND_ROBIN) != 0) {
2999 		/*
3000 		 * VERIFY and ROUND-ROBIN options are mutally exclusive.
3001 		 */
3002 		G_RAID3_DEBUG(1, "Both VERIFY and ROUND-ROBIN flags exist on "
3003 		    "disk %s (device %s), skipping.", pp->name, sc->sc_name);
3004 		return (EINVAL);
3005 	}
3006 	if ((md->md_dflags & ~G_RAID3_DISK_FLAG_MASK) != 0) {
3007 		G_RAID3_DEBUG(1,
3008 		    "Invalid disk flags on disk %s (device %s), skipping.",
3009 		    pp->name, sc->sc_name);
3010 		return (EINVAL);
3011 	}
3012 	return (0);
3013 }
3014 
3015 int
3016 g_raid3_add_disk(struct g_raid3_softc *sc, struct g_provider *pp,
3017     struct g_raid3_metadata *md)
3018 {
3019 	struct g_raid3_disk *disk;
3020 	int error;
3021 
3022 	g_topology_assert_not();
3023 	G_RAID3_DEBUG(2, "Adding disk %s.", pp->name);
3024 
3025 	error = g_raid3_check_metadata(sc, pp, md);
3026 	if (error != 0)
3027 		return (error);
3028 	if (sc->sc_state != G_RAID3_DEVICE_STATE_STARTING &&
3029 	    md->md_genid < sc->sc_genid) {
3030 		G_RAID3_DEBUG(0, "Component %s (device %s) broken, skipping.",
3031 		    pp->name, sc->sc_name);
3032 		return (EINVAL);
3033 	}
3034 	disk = g_raid3_init_disk(sc, pp, md, &error);
3035 	if (disk == NULL)
3036 		return (error);
3037 	error = g_raid3_event_send(disk, G_RAID3_DISK_STATE_NEW,
3038 	    G_RAID3_EVENT_WAIT);
3039 	if (error != 0)
3040 		return (error);
3041 	if (md->md_version < G_RAID3_VERSION) {
3042 		G_RAID3_DEBUG(0, "Upgrading metadata on %s (v%d->v%d).",
3043 		    pp->name, md->md_version, G_RAID3_VERSION);
3044 		g_raid3_update_metadata(disk);
3045 	}
3046 	return (0);
3047 }
3048 
3049 static void
3050 g_raid3_destroy_delayed(void *arg, int flag)
3051 {
3052 	struct g_raid3_softc *sc;
3053 	int error;
3054 
3055 	if (flag == EV_CANCEL) {
3056 		G_RAID3_DEBUG(1, "Destroying canceled.");
3057 		return;
3058 	}
3059 	sc = arg;
3060 	g_topology_unlock();
3061 	sx_xlock(&sc->sc_lock);
3062 	KASSERT((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) == 0,
3063 	    ("DESTROY flag set on %s.", sc->sc_name));
3064 	KASSERT((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROYING) != 0,
3065 	    ("DESTROYING flag not set on %s.", sc->sc_name));
3066 	G_RAID3_DEBUG(0, "Destroying %s (delayed).", sc->sc_name);
3067 	error = g_raid3_destroy(sc, G_RAID3_DESTROY_SOFT);
3068 	if (error != 0) {
3069 		G_RAID3_DEBUG(0, "Cannot destroy %s.", sc->sc_name);
3070 		sx_xunlock(&sc->sc_lock);
3071 	}
3072 	g_topology_lock();
3073 }
3074 
3075 static int
3076 g_raid3_access(struct g_provider *pp, int acr, int acw, int ace)
3077 {
3078 	struct g_raid3_softc *sc;
3079 	int dcr, dcw, dce, error = 0;
3080 
3081 	g_topology_assert();
3082 	G_RAID3_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, acr,
3083 	    acw, ace);
3084 
3085 	sc = pp->geom->softc;
3086 	if (sc == NULL && acr <= 0 && acw <= 0 && ace <= 0)
3087 		return (0);
3088 	KASSERT(sc != NULL, ("NULL softc (provider=%s).", pp->name));
3089 
3090 	dcr = pp->acr + acr;
3091 	dcw = pp->acw + acw;
3092 	dce = pp->ace + ace;
3093 
3094 	g_topology_unlock();
3095 	sx_xlock(&sc->sc_lock);
3096 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) != 0 ||
3097 	    g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) < sc->sc_ndisks - 1) {
3098 		if (acr > 0 || acw > 0 || ace > 0)
3099 			error = ENXIO;
3100 		goto end;
3101 	}
3102 	if (dcw == 0)
3103 		g_raid3_idle(sc, dcw);
3104 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROYING) != 0) {
3105 		if (acr > 0 || acw > 0 || ace > 0) {
3106 			error = ENXIO;
3107 			goto end;
3108 		}
3109 		if (dcr == 0 && dcw == 0 && dce == 0) {
3110 			g_post_event(g_raid3_destroy_delayed, sc, M_WAITOK,
3111 			    sc, NULL);
3112 		}
3113 	}
3114 end:
3115 	sx_xunlock(&sc->sc_lock);
3116 	g_topology_lock();
3117 	return (error);
3118 }
3119 
3120 static struct g_geom *
3121 g_raid3_create(struct g_class *mp, const struct g_raid3_metadata *md)
3122 {
3123 	struct g_raid3_softc *sc;
3124 	struct g_geom *gp;
3125 	int error, timeout;
3126 	u_int n;
3127 
3128 	g_topology_assert();
3129 	G_RAID3_DEBUG(1, "Creating device %s (id=%u).", md->md_name, md->md_id);
3130 
3131 	/* One disk is minimum. */
3132 	if (md->md_all < 1)
3133 		return (NULL);
3134 	/*
3135 	 * Action geom.
3136 	 */
3137 	gp = g_new_geomf(mp, "%s", md->md_name);
3138 	sc = malloc(sizeof(*sc), M_RAID3, M_WAITOK | M_ZERO);
3139 	sc->sc_disks = malloc(sizeof(struct g_raid3_disk) * md->md_all, M_RAID3,
3140 	    M_WAITOK | M_ZERO);
3141 	gp->start = g_raid3_start;
3142 	gp->orphan = g_raid3_orphan;
3143 	gp->access = g_raid3_access;
3144 	gp->dumpconf = g_raid3_dumpconf;
3145 
3146 	sc->sc_id = md->md_id;
3147 	sc->sc_mediasize = md->md_mediasize;
3148 	sc->sc_sectorsize = md->md_sectorsize;
3149 	sc->sc_ndisks = md->md_all;
3150 	sc->sc_round_robin = 0;
3151 	sc->sc_flags = md->md_mflags;
3152 	sc->sc_bump_id = 0;
3153 	sc->sc_idle = 1;
3154 	sc->sc_last_write = time_uptime;
3155 	sc->sc_writes = 0;
3156 	for (n = 0; n < sc->sc_ndisks; n++) {
3157 		sc->sc_disks[n].d_softc = sc;
3158 		sc->sc_disks[n].d_no = n;
3159 		sc->sc_disks[n].d_state = G_RAID3_DISK_STATE_NODISK;
3160 	}
3161 	sx_init(&sc->sc_lock, "graid3:lock");
3162 	bioq_init(&sc->sc_queue);
3163 	mtx_init(&sc->sc_queue_mtx, "graid3:queue", NULL, MTX_DEF);
3164 	bioq_init(&sc->sc_regular_delayed);
3165 	bioq_init(&sc->sc_inflight);
3166 	bioq_init(&sc->sc_sync_delayed);
3167 	TAILQ_INIT(&sc->sc_events);
3168 	mtx_init(&sc->sc_events_mtx, "graid3:events", NULL, MTX_DEF);
3169 	callout_init(&sc->sc_callout, CALLOUT_MPSAFE);
3170 	sc->sc_state = G_RAID3_DEVICE_STATE_STARTING;
3171 	gp->softc = sc;
3172 	sc->sc_geom = gp;
3173 	sc->sc_provider = NULL;
3174 	/*
3175 	 * Synchronization geom.
3176 	 */
3177 	gp = g_new_geomf(mp, "%s.sync", md->md_name);
3178 	gp->softc = sc;
3179 	gp->orphan = g_raid3_orphan;
3180 	sc->sc_sync.ds_geom = gp;
3181 
3182 	if (!g_raid3_use_malloc) {
3183 		sc->sc_zones[G_RAID3_ZONE_64K].sz_zone = uma_zcreate("gr3:64k",
3184 		    65536, g_raid3_uma_ctor, g_raid3_uma_dtor, NULL, NULL,
3185 		    UMA_ALIGN_PTR, 0);
3186 		sc->sc_zones[G_RAID3_ZONE_64K].sz_inuse = 0;
3187 		sc->sc_zones[G_RAID3_ZONE_64K].sz_max = g_raid3_n64k;
3188 		sc->sc_zones[G_RAID3_ZONE_64K].sz_requested =
3189 		    sc->sc_zones[G_RAID3_ZONE_64K].sz_failed = 0;
3190 		sc->sc_zones[G_RAID3_ZONE_16K].sz_zone = uma_zcreate("gr3:16k",
3191 		    16384, g_raid3_uma_ctor, g_raid3_uma_dtor, NULL, NULL,
3192 		    UMA_ALIGN_PTR, 0);
3193 		sc->sc_zones[G_RAID3_ZONE_16K].sz_inuse = 0;
3194 		sc->sc_zones[G_RAID3_ZONE_16K].sz_max = g_raid3_n16k;
3195 		sc->sc_zones[G_RAID3_ZONE_16K].sz_requested =
3196 		    sc->sc_zones[G_RAID3_ZONE_16K].sz_failed = 0;
3197 		sc->sc_zones[G_RAID3_ZONE_4K].sz_zone = uma_zcreate("gr3:4k",
3198 		    4096, g_raid3_uma_ctor, g_raid3_uma_dtor, NULL, NULL,
3199 		    UMA_ALIGN_PTR, 0);
3200 		sc->sc_zones[G_RAID3_ZONE_4K].sz_inuse = 0;
3201 		sc->sc_zones[G_RAID3_ZONE_4K].sz_max = g_raid3_n4k;
3202 		sc->sc_zones[G_RAID3_ZONE_4K].sz_requested =
3203 		    sc->sc_zones[G_RAID3_ZONE_4K].sz_failed = 0;
3204 	}
3205 
3206 	error = kproc_create(g_raid3_worker, sc, &sc->sc_worker, 0, 0,
3207 	    "g_raid3 %s", md->md_name);
3208 	if (error != 0) {
3209 		G_RAID3_DEBUG(1, "Cannot create kernel thread for %s.",
3210 		    sc->sc_name);
3211 		if (!g_raid3_use_malloc) {
3212 			uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_64K].sz_zone);
3213 			uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_16K].sz_zone);
3214 			uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_4K].sz_zone);
3215 		}
3216 		g_destroy_geom(sc->sc_sync.ds_geom);
3217 		mtx_destroy(&sc->sc_events_mtx);
3218 		mtx_destroy(&sc->sc_queue_mtx);
3219 		sx_destroy(&sc->sc_lock);
3220 		g_destroy_geom(sc->sc_geom);
3221 		free(sc->sc_disks, M_RAID3);
3222 		free(sc, M_RAID3);
3223 		return (NULL);
3224 	}
3225 
3226 	G_RAID3_DEBUG(1, "Device %s created (%u components, id=%u).",
3227 	    sc->sc_name, sc->sc_ndisks, sc->sc_id);
3228 
3229 	sc->sc_rootmount = root_mount_hold("GRAID3");
3230 	G_RAID3_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
3231 
3232 	/*
3233 	 * Run timeout.
3234 	 */
3235 	timeout = atomic_load_acq_int(&g_raid3_timeout);
3236 	callout_reset(&sc->sc_callout, timeout * hz, g_raid3_go, sc);
3237 	return (sc->sc_geom);
3238 }
3239 
3240 int
3241 g_raid3_destroy(struct g_raid3_softc *sc, int how)
3242 {
3243 	struct g_provider *pp;
3244 
3245 	g_topology_assert_not();
3246 	if (sc == NULL)
3247 		return (ENXIO);
3248 	sx_assert(&sc->sc_lock, SX_XLOCKED);
3249 
3250 	pp = sc->sc_provider;
3251 	if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
3252 		switch (how) {
3253 		case G_RAID3_DESTROY_SOFT:
3254 			G_RAID3_DEBUG(1,
3255 			    "Device %s is still open (r%dw%de%d).", pp->name,
3256 			    pp->acr, pp->acw, pp->ace);
3257 			return (EBUSY);
3258 		case G_RAID3_DESTROY_DELAYED:
3259 			G_RAID3_DEBUG(1,
3260 			    "Device %s will be destroyed on last close.",
3261 			    pp->name);
3262 			if (sc->sc_syncdisk != NULL)
3263 				g_raid3_sync_stop(sc, 1);
3264 			sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROYING;
3265 			return (EBUSY);
3266 		case G_RAID3_DESTROY_HARD:
3267 			G_RAID3_DEBUG(1, "Device %s is still open, so it "
3268 			    "can't be definitely removed.", pp->name);
3269 			break;
3270 		}
3271 	}
3272 
3273 	g_topology_lock();
3274 	if (sc->sc_geom->softc == NULL) {
3275 		g_topology_unlock();
3276 		return (0);
3277 	}
3278 	sc->sc_geom->softc = NULL;
3279 	sc->sc_sync.ds_geom->softc = NULL;
3280 	g_topology_unlock();
3281 
3282 	sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
3283 	sc->sc_flags |= G_RAID3_DEVICE_FLAG_WAIT;
3284 	G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, sc);
3285 	sx_xunlock(&sc->sc_lock);
3286 	mtx_lock(&sc->sc_queue_mtx);
3287 	wakeup(sc);
3288 	wakeup(&sc->sc_queue);
3289 	mtx_unlock(&sc->sc_queue_mtx);
3290 	G_RAID3_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
3291 	while (sc->sc_worker != NULL)
3292 		tsleep(&sc->sc_worker, PRIBIO, "r3:destroy", hz / 5);
3293 	G_RAID3_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
3294 	sx_xlock(&sc->sc_lock);
3295 	g_raid3_destroy_device(sc);
3296 	free(sc->sc_disks, M_RAID3);
3297 	free(sc, M_RAID3);
3298 	return (0);
3299 }
3300 
3301 static void
3302 g_raid3_taste_orphan(struct g_consumer *cp)
3303 {
3304 
3305 	KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
3306 	    cp->provider->name));
3307 }
3308 
3309 static struct g_geom *
3310 g_raid3_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
3311 {
3312 	struct g_raid3_metadata md;
3313 	struct g_raid3_softc *sc;
3314 	struct g_consumer *cp;
3315 	struct g_geom *gp;
3316 	int error;
3317 
3318 	g_topology_assert();
3319 	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
3320 	G_RAID3_DEBUG(2, "Tasting %s.", pp->name);
3321 
3322 	gp = g_new_geomf(mp, "raid3:taste");
3323 	/* This orphan function should be never called. */
3324 	gp->orphan = g_raid3_taste_orphan;
3325 	cp = g_new_consumer(gp);
3326 	g_attach(cp, pp);
3327 	error = g_raid3_read_metadata(cp, &md);
3328 	g_detach(cp);
3329 	g_destroy_consumer(cp);
3330 	g_destroy_geom(gp);
3331 	if (error != 0)
3332 		return (NULL);
3333 	gp = NULL;
3334 
3335 	if (md.md_provider[0] != '\0' &&
3336 	    !g_compare_names(md.md_provider, pp->name))
3337 		return (NULL);
3338 	if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
3339 		return (NULL);
3340 	if (g_raid3_debug >= 2)
3341 		raid3_metadata_dump(&md);
3342 
3343 	/*
3344 	 * Let's check if device already exists.
3345 	 */
3346 	sc = NULL;
3347 	LIST_FOREACH(gp, &mp->geom, geom) {
3348 		sc = gp->softc;
3349 		if (sc == NULL)
3350 			continue;
3351 		if (sc->sc_sync.ds_geom == gp)
3352 			continue;
3353 		if (strcmp(md.md_name, sc->sc_name) != 0)
3354 			continue;
3355 		if (md.md_id != sc->sc_id) {
3356 			G_RAID3_DEBUG(0, "Device %s already configured.",
3357 			    sc->sc_name);
3358 			return (NULL);
3359 		}
3360 		break;
3361 	}
3362 	if (gp == NULL) {
3363 		gp = g_raid3_create(mp, &md);
3364 		if (gp == NULL) {
3365 			G_RAID3_DEBUG(0, "Cannot create device %s.",
3366 			    md.md_name);
3367 			return (NULL);
3368 		}
3369 		sc = gp->softc;
3370 	}
3371 	G_RAID3_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
3372 	g_topology_unlock();
3373 	sx_xlock(&sc->sc_lock);
3374 	error = g_raid3_add_disk(sc, pp, &md);
3375 	if (error != 0) {
3376 		G_RAID3_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
3377 		    pp->name, gp->name, error);
3378 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_NODISK) ==
3379 		    sc->sc_ndisks) {
3380 			g_cancel_event(sc);
3381 			g_raid3_destroy(sc, G_RAID3_DESTROY_HARD);
3382 			g_topology_lock();
3383 			return (NULL);
3384 		}
3385 		gp = NULL;
3386 	}
3387 	sx_xunlock(&sc->sc_lock);
3388 	g_topology_lock();
3389 	return (gp);
3390 }
3391 
3392 static int
3393 g_raid3_destroy_geom(struct gctl_req *req __unused, struct g_class *mp __unused,
3394     struct g_geom *gp)
3395 {
3396 	struct g_raid3_softc *sc;
3397 	int error;
3398 
3399 	g_topology_unlock();
3400 	sc = gp->softc;
3401 	sx_xlock(&sc->sc_lock);
3402 	g_cancel_event(sc);
3403 	error = g_raid3_destroy(gp->softc, G_RAID3_DESTROY_SOFT);
3404 	if (error != 0)
3405 		sx_xunlock(&sc->sc_lock);
3406 	g_topology_lock();
3407 	return (error);
3408 }
3409 
3410 static void
3411 g_raid3_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
3412     struct g_consumer *cp, struct g_provider *pp)
3413 {
3414 	struct g_raid3_softc *sc;
3415 
3416 	g_topology_assert();
3417 
3418 	sc = gp->softc;
3419 	if (sc == NULL)
3420 		return;
3421 	/* Skip synchronization geom. */
3422 	if (gp == sc->sc_sync.ds_geom)
3423 		return;
3424 	if (pp != NULL) {
3425 		/* Nothing here. */
3426 	} else if (cp != NULL) {
3427 		struct g_raid3_disk *disk;
3428 
3429 		disk = cp->private;
3430 		if (disk == NULL)
3431 			return;
3432 		g_topology_unlock();
3433 		sx_xlock(&sc->sc_lock);
3434 		sbuf_printf(sb, "%s<Type>", indent);
3435 		if (disk->d_no == sc->sc_ndisks - 1)
3436 			sbuf_printf(sb, "PARITY");
3437 		else
3438 			sbuf_printf(sb, "DATA");
3439 		sbuf_printf(sb, "</Type>\n");
3440 		sbuf_printf(sb, "%s<Number>%u</Number>\n", indent,
3441 		    (u_int)disk->d_no);
3442 		if (disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
3443 			sbuf_printf(sb, "%s<Synchronized>", indent);
3444 			if (disk->d_sync.ds_offset == 0)
3445 				sbuf_printf(sb, "0%%");
3446 			else {
3447 				sbuf_printf(sb, "%u%%",
3448 				    (u_int)((disk->d_sync.ds_offset * 100) /
3449 				    (sc->sc_mediasize / (sc->sc_ndisks - 1))));
3450 			}
3451 			sbuf_printf(sb, "</Synchronized>\n");
3452 			if (disk->d_sync.ds_offset > 0) {
3453 				sbuf_printf(sb, "%s<BytesSynced>%jd"
3454 				    "</BytesSynced>\n", indent,
3455 				    (intmax_t)disk->d_sync.ds_offset);
3456 			}
3457 		}
3458 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent,
3459 		    disk->d_sync.ds_syncid);
3460 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, disk->d_genid);
3461 		sbuf_printf(sb, "%s<Flags>", indent);
3462 		if (disk->d_flags == 0)
3463 			sbuf_printf(sb, "NONE");
3464 		else {
3465 			int first = 1;
3466 
3467 #define	ADD_FLAG(flag, name)	do {					\
3468 	if ((disk->d_flags & (flag)) != 0) {				\
3469 		if (!first)						\
3470 			sbuf_printf(sb, ", ");				\
3471 		else							\
3472 			first = 0;					\
3473 		sbuf_printf(sb, name);					\
3474 	}								\
3475 } while (0)
3476 			ADD_FLAG(G_RAID3_DISK_FLAG_DIRTY, "DIRTY");
3477 			ADD_FLAG(G_RAID3_DISK_FLAG_HARDCODED, "HARDCODED");
3478 			ADD_FLAG(G_RAID3_DISK_FLAG_SYNCHRONIZING,
3479 			    "SYNCHRONIZING");
3480 			ADD_FLAG(G_RAID3_DISK_FLAG_FORCE_SYNC, "FORCE_SYNC");
3481 			ADD_FLAG(G_RAID3_DISK_FLAG_BROKEN, "BROKEN");
3482 #undef	ADD_FLAG
3483 		}
3484 		sbuf_printf(sb, "</Flags>\n");
3485 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3486 		    g_raid3_disk_state2str(disk->d_state));
3487 		sx_xunlock(&sc->sc_lock);
3488 		g_topology_lock();
3489 	} else {
3490 		g_topology_unlock();
3491 		sx_xlock(&sc->sc_lock);
3492 		if (!g_raid3_use_malloc) {
3493 			sbuf_printf(sb,
3494 			    "%s<Zone4kRequested>%u</Zone4kRequested>\n", indent,
3495 			    sc->sc_zones[G_RAID3_ZONE_4K].sz_requested);
3496 			sbuf_printf(sb,
3497 			    "%s<Zone4kFailed>%u</Zone4kFailed>\n", indent,
3498 			    sc->sc_zones[G_RAID3_ZONE_4K].sz_failed);
3499 			sbuf_printf(sb,
3500 			    "%s<Zone16kRequested>%u</Zone16kRequested>\n", indent,
3501 			    sc->sc_zones[G_RAID3_ZONE_16K].sz_requested);
3502 			sbuf_printf(sb,
3503 			    "%s<Zone16kFailed>%u</Zone16kFailed>\n", indent,
3504 			    sc->sc_zones[G_RAID3_ZONE_16K].sz_failed);
3505 			sbuf_printf(sb,
3506 			    "%s<Zone64kRequested>%u</Zone64kRequested>\n", indent,
3507 			    sc->sc_zones[G_RAID3_ZONE_64K].sz_requested);
3508 			sbuf_printf(sb,
3509 			    "%s<Zone64kFailed>%u</Zone64kFailed>\n", indent,
3510 			    sc->sc_zones[G_RAID3_ZONE_64K].sz_failed);
3511 		}
3512 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
3513 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent, sc->sc_syncid);
3514 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, sc->sc_genid);
3515 		sbuf_printf(sb, "%s<Flags>", indent);
3516 		if (sc->sc_flags == 0)
3517 			sbuf_printf(sb, "NONE");
3518 		else {
3519 			int first = 1;
3520 
3521 #define	ADD_FLAG(flag, name)	do {					\
3522 	if ((sc->sc_flags & (flag)) != 0) {				\
3523 		if (!first)						\
3524 			sbuf_printf(sb, ", ");				\
3525 		else							\
3526 			first = 0;					\
3527 		sbuf_printf(sb, name);					\
3528 	}								\
3529 } while (0)
3530 			ADD_FLAG(G_RAID3_DEVICE_FLAG_NOFAILSYNC, "NOFAILSYNC");
3531 			ADD_FLAG(G_RAID3_DEVICE_FLAG_NOAUTOSYNC, "NOAUTOSYNC");
3532 			ADD_FLAG(G_RAID3_DEVICE_FLAG_ROUND_ROBIN,
3533 			    "ROUND-ROBIN");
3534 			ADD_FLAG(G_RAID3_DEVICE_FLAG_VERIFY, "VERIFY");
3535 #undef	ADD_FLAG
3536 		}
3537 		sbuf_printf(sb, "</Flags>\n");
3538 		sbuf_printf(sb, "%s<Components>%u</Components>\n", indent,
3539 		    sc->sc_ndisks);
3540 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3541 		    g_raid3_device_state2str(sc->sc_state));
3542 		sx_xunlock(&sc->sc_lock);
3543 		g_topology_lock();
3544 	}
3545 }
3546 
3547 static void
3548 g_raid3_shutdown_post_sync(void *arg, int howto)
3549 {
3550 	struct g_class *mp;
3551 	struct g_geom *gp, *gp2;
3552 	struct g_raid3_softc *sc;
3553 	int error;
3554 
3555 	mp = arg;
3556 	DROP_GIANT();
3557 	g_topology_lock();
3558 	g_raid3_shutdown = 1;
3559 	LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
3560 		if ((sc = gp->softc) == NULL)
3561 			continue;
3562 		/* Skip synchronization geom. */
3563 		if (gp == sc->sc_sync.ds_geom)
3564 			continue;
3565 		g_topology_unlock();
3566 		sx_xlock(&sc->sc_lock);
3567 		g_raid3_idle(sc, -1);
3568 		g_cancel_event(sc);
3569 		error = g_raid3_destroy(sc, G_RAID3_DESTROY_DELAYED);
3570 		if (error != 0)
3571 			sx_xunlock(&sc->sc_lock);
3572 		g_topology_lock();
3573 	}
3574 	g_topology_unlock();
3575 	PICKUP_GIANT();
3576 }
3577 
3578 static void
3579 g_raid3_init(struct g_class *mp)
3580 {
3581 
3582 	g_raid3_post_sync = EVENTHANDLER_REGISTER(shutdown_post_sync,
3583 	    g_raid3_shutdown_post_sync, mp, SHUTDOWN_PRI_FIRST);
3584 	if (g_raid3_post_sync == NULL)
3585 		G_RAID3_DEBUG(0, "Warning! Cannot register shutdown event.");
3586 }
3587 
3588 static void
3589 g_raid3_fini(struct g_class *mp)
3590 {
3591 
3592 	if (g_raid3_post_sync != NULL)
3593 		EVENTHANDLER_DEREGISTER(shutdown_post_sync, g_raid3_post_sync);
3594 }
3595 
3596 DECLARE_GEOM_CLASS(g_raid3_class, g_raid3);
3597