xref: /freebsd/sys/geom/raid3/g_raid3.c (revision 35a04710d7286aa9538917fd7f8e417dbee95b82)
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/sysctl.h>
39 #include <sys/malloc.h>
40 #include <sys/eventhandler.h>
41 #include <vm/uma.h>
42 #include <geom/geom.h>
43 #include <sys/proc.h>
44 #include <sys/kthread.h>
45 #include <sys/sched.h>
46 #include <geom/raid3/g_raid3.h>
47 
48 
49 static MALLOC_DEFINE(M_RAID3, "raid3_data", "GEOM_RAID3 Data");
50 
51 SYSCTL_DECL(_kern_geom);
52 SYSCTL_NODE(_kern_geom, OID_AUTO, raid3, CTLFLAG_RW, 0, "GEOM_RAID3 stuff");
53 u_int g_raid3_debug = 0;
54 TUNABLE_INT("kern.geom.raid3.debug", &g_raid3_debug);
55 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, debug, CTLFLAG_RW, &g_raid3_debug, 0,
56     "Debug level");
57 static u_int g_raid3_timeout = 4;
58 TUNABLE_INT("kern.geom.raid3.timeout", &g_raid3_timeout);
59 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, timeout, CTLFLAG_RW, &g_raid3_timeout,
60     0, "Time to wait on all raid3 components");
61 static u_int g_raid3_idletime = 5;
62 TUNABLE_INT("kern.geom.raid3.idletime", &g_raid3_idletime);
63 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, idletime, CTLFLAG_RW,
64     &g_raid3_idletime, 0, "Mark components as clean when idling");
65 static u_int g_raid3_disconnect_on_failure = 1;
66 TUNABLE_INT("kern.geom.raid3.disconnect_on_failure",
67     &g_raid3_disconnect_on_failure);
68 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, disconnect_on_failure, CTLFLAG_RW,
69     &g_raid3_disconnect_on_failure, 0, "Disconnect component on I/O failure.");
70 static u_int g_raid3_syncreqs = 2;
71 TUNABLE_INT("kern.geom.raid3.sync_requests", &g_raid3_syncreqs);
72 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, sync_requests, CTLFLAG_RDTUN,
73     &g_raid3_syncreqs, 0, "Parallel synchronization I/O requests.");
74 static u_int g_raid3_use_malloc = 0;
75 TUNABLE_INT("kern.geom.raid3.use_malloc", &g_raid3_use_malloc);
76 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, use_malloc, CTLFLAG_RDTUN,
77     &g_raid3_use_malloc, 0, "Use malloc(9) instead of uma(9).");
78 
79 static u_int g_raid3_n64k = 50;
80 TUNABLE_INT("kern.geom.raid3.n64k", &g_raid3_n64k);
81 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, n64k, CTLFLAG_RD, &g_raid3_n64k, 0,
82     "Maximum number of 64kB allocations");
83 static u_int g_raid3_n16k = 200;
84 TUNABLE_INT("kern.geom.raid3.n16k", &g_raid3_n16k);
85 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, n16k, CTLFLAG_RD, &g_raid3_n16k, 0,
86     "Maximum number of 16kB allocations");
87 static u_int g_raid3_n4k = 1200;
88 TUNABLE_INT("kern.geom.raid3.n4k", &g_raid3_n4k);
89 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, n4k, CTLFLAG_RD, &g_raid3_n4k, 0,
90     "Maximum number of 4kB allocations");
91 
92 SYSCTL_NODE(_kern_geom_raid3, OID_AUTO, stat, CTLFLAG_RW, 0,
93     "GEOM_RAID3 statistics");
94 static u_int g_raid3_parity_mismatch = 0;
95 SYSCTL_UINT(_kern_geom_raid3_stat, OID_AUTO, parity_mismatch, CTLFLAG_RD,
96     &g_raid3_parity_mismatch, 0, "Number of failures in VERIFY mode");
97 
98 #define	MSLEEP(ident, mtx, priority, wmesg, timeout)	do {		\
99 	G_RAID3_DEBUG(4, "%s: Sleeping %p.", __func__, (ident));	\
100 	msleep((ident), (mtx), (priority), (wmesg), (timeout));		\
101 	G_RAID3_DEBUG(4, "%s: Woken up %p.", __func__, (ident));	\
102 } while (0)
103 
104 static eventhandler_tag g_raid3_pre_sync = NULL;
105 
106 static int g_raid3_destroy_geom(struct gctl_req *req, struct g_class *mp,
107     struct g_geom *gp);
108 static g_taste_t g_raid3_taste;
109 static void g_raid3_init(struct g_class *mp);
110 static void g_raid3_fini(struct g_class *mp);
111 
112 struct g_class g_raid3_class = {
113 	.name = G_RAID3_CLASS_NAME,
114 	.version = G_VERSION,
115 	.ctlreq = g_raid3_config,
116 	.taste = g_raid3_taste,
117 	.destroy_geom = g_raid3_destroy_geom,
118 	.init = g_raid3_init,
119 	.fini = g_raid3_fini
120 };
121 
122 
123 static void g_raid3_destroy_provider(struct g_raid3_softc *sc);
124 static int g_raid3_update_disk(struct g_raid3_disk *disk, u_int state);
125 static void g_raid3_update_device(struct g_raid3_softc *sc, boolean_t force);
126 static void g_raid3_dumpconf(struct sbuf *sb, const char *indent,
127     struct g_geom *gp, struct g_consumer *cp, struct g_provider *pp);
128 static void g_raid3_sync_stop(struct g_raid3_softc *sc, int type);
129 static int g_raid3_register_request(struct bio *pbp);
130 static void g_raid3_sync_release(struct g_raid3_softc *sc);
131 
132 
133 static const char *
134 g_raid3_disk_state2str(int state)
135 {
136 
137 	switch (state) {
138 	case G_RAID3_DISK_STATE_NODISK:
139 		return ("NODISK");
140 	case G_RAID3_DISK_STATE_NONE:
141 		return ("NONE");
142 	case G_RAID3_DISK_STATE_NEW:
143 		return ("NEW");
144 	case G_RAID3_DISK_STATE_ACTIVE:
145 		return ("ACTIVE");
146 	case G_RAID3_DISK_STATE_STALE:
147 		return ("STALE");
148 	case G_RAID3_DISK_STATE_SYNCHRONIZING:
149 		return ("SYNCHRONIZING");
150 	case G_RAID3_DISK_STATE_DISCONNECTED:
151 		return ("DISCONNECTED");
152 	default:
153 		return ("INVALID");
154 	}
155 }
156 
157 static const char *
158 g_raid3_device_state2str(int state)
159 {
160 
161 	switch (state) {
162 	case G_RAID3_DEVICE_STATE_STARTING:
163 		return ("STARTING");
164 	case G_RAID3_DEVICE_STATE_DEGRADED:
165 		return ("DEGRADED");
166 	case G_RAID3_DEVICE_STATE_COMPLETE:
167 		return ("COMPLETE");
168 	default:
169 		return ("INVALID");
170 	}
171 }
172 
173 const char *
174 g_raid3_get_diskname(struct g_raid3_disk *disk)
175 {
176 
177 	if (disk->d_consumer == NULL || disk->d_consumer->provider == NULL)
178 		return ("[unknown]");
179 	return (disk->d_name);
180 }
181 
182 static void *
183 g_raid3_alloc(struct g_raid3_softc *sc, size_t size, int flags)
184 {
185 	void *ptr;
186 
187 	if (g_raid3_use_malloc)
188 		ptr = malloc(size, M_RAID3, flags);
189 	else {
190 		ptr = uma_zalloc_arg(sc->sc_zones[g_raid3_zone(size)].sz_zone,
191 		   &sc->sc_zones[g_raid3_zone(size)], flags);
192 		sc->sc_zones[g_raid3_zone(size)].sz_requested++;
193 		if (ptr == NULL)
194 			sc->sc_zones[g_raid3_zone(size)].sz_failed++;
195 	}
196 	return (ptr);
197 }
198 
199 static void
200 g_raid3_free(struct g_raid3_softc *sc, void *ptr, size_t size)
201 {
202 
203 	if (g_raid3_use_malloc)
204 		free(ptr, M_RAID3);
205 	else {
206 		uma_zfree_arg(sc->sc_zones[g_raid3_zone(size)].sz_zone,
207 		    ptr, &sc->sc_zones[g_raid3_zone(size)]);
208 	}
209 }
210 
211 static int
212 g_raid3_uma_ctor(void *mem, int size, void *arg, int flags)
213 {
214 	struct g_raid3_zone *sz = arg;
215 
216 	if (sz->sz_max > 0 && sz->sz_inuse == sz->sz_max)
217 		return (ENOMEM);
218 	sz->sz_inuse++;
219 	return (0);
220 }
221 
222 static void
223 g_raid3_uma_dtor(void *mem, int size, void *arg)
224 {
225 	struct g_raid3_zone *sz = arg;
226 
227 	sz->sz_inuse--;
228 }
229 
230 #define	g_raid3_xor(src1, src2, dst, size)				\
231 	_g_raid3_xor((uint64_t *)(src1), (uint64_t *)(src2),		\
232 	    (uint64_t *)(dst), (size_t)size)
233 static void
234 _g_raid3_xor(uint64_t *src1, uint64_t *src2, uint64_t *dst, size_t size)
235 {
236 
237 	KASSERT((size % 128) == 0, ("Invalid size: %zu.", size));
238 	for (; size > 0; size -= 128) {
239 		*dst++ = (*src1++) ^ (*src2++);
240 		*dst++ = (*src1++) ^ (*src2++);
241 		*dst++ = (*src1++) ^ (*src2++);
242 		*dst++ = (*src1++) ^ (*src2++);
243 		*dst++ = (*src1++) ^ (*src2++);
244 		*dst++ = (*src1++) ^ (*src2++);
245 		*dst++ = (*src1++) ^ (*src2++);
246 		*dst++ = (*src1++) ^ (*src2++);
247 		*dst++ = (*src1++) ^ (*src2++);
248 		*dst++ = (*src1++) ^ (*src2++);
249 		*dst++ = (*src1++) ^ (*src2++);
250 		*dst++ = (*src1++) ^ (*src2++);
251 		*dst++ = (*src1++) ^ (*src2++);
252 		*dst++ = (*src1++) ^ (*src2++);
253 		*dst++ = (*src1++) ^ (*src2++);
254 		*dst++ = (*src1++) ^ (*src2++);
255 	}
256 }
257 
258 static int
259 g_raid3_is_zero(struct bio *bp)
260 {
261 	static const uint64_t zeros[] = {
262 	    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
263 	};
264 	u_char *addr;
265 	ssize_t size;
266 
267 	size = bp->bio_length;
268 	addr = (u_char *)bp->bio_data;
269 	for (; size > 0; size -= sizeof(zeros), addr += sizeof(zeros)) {
270 		if (bcmp(addr, zeros, sizeof(zeros)) != 0)
271 			return (0);
272 	}
273 	return (1);
274 }
275 
276 /*
277  * --- Events handling functions ---
278  * Events in geom_raid3 are used to maintain disks and device status
279  * from one thread to simplify locking.
280  */
281 static void
282 g_raid3_event_free(struct g_raid3_event *ep)
283 {
284 
285 	free(ep, M_RAID3);
286 }
287 
288 int
289 g_raid3_event_send(void *arg, int state, int flags)
290 {
291 	struct g_raid3_softc *sc;
292 	struct g_raid3_disk *disk;
293 	struct g_raid3_event *ep;
294 	int error;
295 
296 	ep = malloc(sizeof(*ep), M_RAID3, M_WAITOK);
297 	G_RAID3_DEBUG(4, "%s: Sending event %p.", __func__, ep);
298 	if ((flags & G_RAID3_EVENT_DEVICE) != 0) {
299 		disk = NULL;
300 		sc = arg;
301 	} else {
302 		disk = arg;
303 		sc = disk->d_softc;
304 	}
305 	ep->e_disk = disk;
306 	ep->e_state = state;
307 	ep->e_flags = flags;
308 	ep->e_error = 0;
309 	mtx_lock(&sc->sc_events_mtx);
310 	TAILQ_INSERT_TAIL(&sc->sc_events, ep, e_next);
311 	mtx_unlock(&sc->sc_events_mtx);
312 	G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, sc);
313 	mtx_lock(&sc->sc_queue_mtx);
314 	wakeup(sc);
315 	wakeup(&sc->sc_queue);
316 	mtx_unlock(&sc->sc_queue_mtx);
317 	if ((flags & G_RAID3_EVENT_DONTWAIT) != 0)
318 		return (0);
319 	sx_assert(&sc->sc_lock, SX_XLOCKED);
320 	G_RAID3_DEBUG(4, "%s: Sleeping %p.", __func__, ep);
321 	sx_xunlock(&sc->sc_lock);
322 	while ((ep->e_flags & G_RAID3_EVENT_DONE) == 0) {
323 		mtx_lock(&sc->sc_events_mtx);
324 		MSLEEP(ep, &sc->sc_events_mtx, PRIBIO | PDROP, "r3:event",
325 		    hz * 5);
326 	}
327 	error = ep->e_error;
328 	g_raid3_event_free(ep);
329 	sx_xlock(&sc->sc_lock);
330 	return (error);
331 }
332 
333 static struct g_raid3_event *
334 g_raid3_event_get(struct g_raid3_softc *sc)
335 {
336 	struct g_raid3_event *ep;
337 
338 	mtx_lock(&sc->sc_events_mtx);
339 	ep = TAILQ_FIRST(&sc->sc_events);
340 	mtx_unlock(&sc->sc_events_mtx);
341 	return (ep);
342 }
343 
344 static void
345 g_raid3_event_remove(struct g_raid3_softc *sc, struct g_raid3_event *ep)
346 {
347 
348 	mtx_lock(&sc->sc_events_mtx);
349 	TAILQ_REMOVE(&sc->sc_events, ep, e_next);
350 	mtx_unlock(&sc->sc_events_mtx);
351 }
352 
353 static void
354 g_raid3_event_cancel(struct g_raid3_disk *disk)
355 {
356 	struct g_raid3_softc *sc;
357 	struct g_raid3_event *ep, *tmpep;
358 
359 	sc = disk->d_softc;
360 	sx_assert(&sc->sc_lock, SX_XLOCKED);
361 
362 	mtx_lock(&sc->sc_events_mtx);
363 	TAILQ_FOREACH_SAFE(ep, &sc->sc_events, e_next, tmpep) {
364 		if ((ep->e_flags & G_RAID3_EVENT_DEVICE) != 0)
365 			continue;
366 		if (ep->e_disk != disk)
367 			continue;
368 		TAILQ_REMOVE(&sc->sc_events, ep, e_next);
369 		if ((ep->e_flags & G_RAID3_EVENT_DONTWAIT) != 0)
370 			g_raid3_event_free(ep);
371 		else {
372 			ep->e_error = ECANCELED;
373 			wakeup(ep);
374 		}
375 	}
376 	mtx_unlock(&sc->sc_events_mtx);
377 }
378 
379 /*
380  * Return the number of disks in the given state.
381  * If state is equal to -1, count all connected disks.
382  */
383 u_int
384 g_raid3_ndisks(struct g_raid3_softc *sc, int state)
385 {
386 	struct g_raid3_disk *disk;
387 	u_int n, ndisks;
388 
389 	sx_assert(&sc->sc_lock, SX_LOCKED);
390 
391 	for (n = ndisks = 0; n < sc->sc_ndisks; n++) {
392 		disk = &sc->sc_disks[n];
393 		if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
394 			continue;
395 		if (state == -1 || disk->d_state == state)
396 			ndisks++;
397 	}
398 	return (ndisks);
399 }
400 
401 static u_int
402 g_raid3_nrequests(struct g_raid3_softc *sc, struct g_consumer *cp)
403 {
404 	struct bio *bp;
405 	u_int nreqs = 0;
406 
407 	mtx_lock(&sc->sc_queue_mtx);
408 	TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
409 		if (bp->bio_from == cp)
410 			nreqs++;
411 	}
412 	mtx_unlock(&sc->sc_queue_mtx);
413 	return (nreqs);
414 }
415 
416 static int
417 g_raid3_is_busy(struct g_raid3_softc *sc, struct g_consumer *cp)
418 {
419 
420 	if (cp->index > 0) {
421 		G_RAID3_DEBUG(2,
422 		    "I/O requests for %s exist, can't destroy it now.",
423 		    cp->provider->name);
424 		return (1);
425 	}
426 	if (g_raid3_nrequests(sc, cp) > 0) {
427 		G_RAID3_DEBUG(2,
428 		    "I/O requests for %s in queue, can't destroy it now.",
429 		    cp->provider->name);
430 		return (1);
431 	}
432 	return (0);
433 }
434 
435 static void
436 g_raid3_destroy_consumer(void *arg, int flags __unused)
437 {
438 	struct g_consumer *cp;
439 
440 	g_topology_assert();
441 
442 	cp = arg;
443 	G_RAID3_DEBUG(1, "Consumer %s destroyed.", cp->provider->name);
444 	g_detach(cp);
445 	g_destroy_consumer(cp);
446 }
447 
448 static void
449 g_raid3_kill_consumer(struct g_raid3_softc *sc, struct g_consumer *cp)
450 {
451 	struct g_provider *pp;
452 	int retaste_wait;
453 
454 	g_topology_assert();
455 
456 	cp->private = NULL;
457 	if (g_raid3_is_busy(sc, cp))
458 		return;
459 	G_RAID3_DEBUG(2, "Consumer %s destroyed.", cp->provider->name);
460 	pp = cp->provider;
461 	retaste_wait = 0;
462 	if (cp->acw == 1) {
463 		if ((pp->geom->flags & G_GEOM_WITHER) == 0)
464 			retaste_wait = 1;
465 	}
466 	G_RAID3_DEBUG(2, "Access %s r%dw%de%d = %d", pp->name, -cp->acr,
467 	    -cp->acw, -cp->ace, 0);
468 	if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
469 		g_access(cp, -cp->acr, -cp->acw, -cp->ace);
470 	if (retaste_wait) {
471 		/*
472 		 * After retaste event was send (inside g_access()), we can send
473 		 * event to detach and destroy consumer.
474 		 * A class, which has consumer to the given provider connected
475 		 * will not receive retaste event for the provider.
476 		 * This is the way how I ignore retaste events when I close
477 		 * consumers opened for write: I detach and destroy consumer
478 		 * after retaste event is sent.
479 		 */
480 		g_post_event(g_raid3_destroy_consumer, cp, M_WAITOK, NULL);
481 		return;
482 	}
483 	G_RAID3_DEBUG(1, "Consumer %s destroyed.", pp->name);
484 	g_detach(cp);
485 	g_destroy_consumer(cp);
486 }
487 
488 static int
489 g_raid3_connect_disk(struct g_raid3_disk *disk, struct g_provider *pp)
490 {
491 	struct g_consumer *cp;
492 	int error;
493 
494 	g_topology_assert_not();
495 	KASSERT(disk->d_consumer == NULL,
496 	    ("Disk already connected (device %s).", disk->d_softc->sc_name));
497 
498 	g_topology_lock();
499 	cp = g_new_consumer(disk->d_softc->sc_geom);
500 	error = g_attach(cp, pp);
501 	if (error != 0) {
502 		g_destroy_consumer(cp);
503 		g_topology_unlock();
504 		return (error);
505 	}
506 	error = g_access(cp, 1, 1, 1);
507 		g_topology_unlock();
508 	if (error != 0) {
509 		g_detach(cp);
510 		g_destroy_consumer(cp);
511 		G_RAID3_DEBUG(0, "Cannot open consumer %s (error=%d).",
512 		    pp->name, error);
513 		return (error);
514 	}
515 	disk->d_consumer = cp;
516 	disk->d_consumer->private = disk;
517 	disk->d_consumer->index = 0;
518 	G_RAID3_DEBUG(2, "Disk %s connected.", g_raid3_get_diskname(disk));
519 	return (0);
520 }
521 
522 static void
523 g_raid3_disconnect_consumer(struct g_raid3_softc *sc, struct g_consumer *cp)
524 {
525 
526 	g_topology_assert();
527 
528 	if (cp == NULL)
529 		return;
530 	if (cp->provider != NULL)
531 		g_raid3_kill_consumer(sc, cp);
532 	else
533 		g_destroy_consumer(cp);
534 }
535 
536 /*
537  * Initialize disk. This means allocate memory, create consumer, attach it
538  * to the provider and open access (r1w1e1) to it.
539  */
540 static struct g_raid3_disk *
541 g_raid3_init_disk(struct g_raid3_softc *sc, struct g_provider *pp,
542     struct g_raid3_metadata *md, int *errorp)
543 {
544 	struct g_raid3_disk *disk;
545 	int error;
546 
547 	disk = &sc->sc_disks[md->md_no];
548 	error = g_raid3_connect_disk(disk, pp);
549 	if (error != 0) {
550 		if (errorp != NULL)
551 			*errorp = error;
552 		return (NULL);
553 	}
554 	disk->d_state = G_RAID3_DISK_STATE_NONE;
555 	disk->d_flags = md->md_dflags;
556 	if (md->md_provider[0] != '\0')
557 		disk->d_flags |= G_RAID3_DISK_FLAG_HARDCODED;
558 	disk->d_sync.ds_consumer = NULL;
559 	disk->d_sync.ds_offset = md->md_sync_offset;
560 	disk->d_sync.ds_offset_done = md->md_sync_offset;
561 	disk->d_genid = md->md_genid;
562 	disk->d_sync.ds_syncid = md->md_syncid;
563 	if (errorp != NULL)
564 		*errorp = 0;
565 	return (disk);
566 }
567 
568 static void
569 g_raid3_destroy_disk(struct g_raid3_disk *disk)
570 {
571 	struct g_raid3_softc *sc;
572 
573 	g_topology_assert_not();
574 	sc = disk->d_softc;
575 	sx_assert(&sc->sc_lock, SX_XLOCKED);
576 
577 	if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
578 		return;
579 	g_raid3_event_cancel(disk);
580 	switch (disk->d_state) {
581 	case G_RAID3_DISK_STATE_SYNCHRONIZING:
582 		if (sc->sc_syncdisk != NULL)
583 			g_raid3_sync_stop(sc, 1);
584 		/* FALLTHROUGH */
585 	case G_RAID3_DISK_STATE_NEW:
586 	case G_RAID3_DISK_STATE_STALE:
587 	case G_RAID3_DISK_STATE_ACTIVE:
588 		g_topology_lock();
589 		g_raid3_disconnect_consumer(sc, disk->d_consumer);
590 		g_topology_unlock();
591 		disk->d_consumer = NULL;
592 		break;
593 	default:
594 		KASSERT(0 == 1, ("Wrong disk state (%s, %s).",
595 		    g_raid3_get_diskname(disk),
596 		    g_raid3_disk_state2str(disk->d_state)));
597 	}
598 	disk->d_state = G_RAID3_DISK_STATE_NODISK;
599 }
600 
601 static void
602 g_raid3_destroy_device(struct g_raid3_softc *sc)
603 {
604 	struct g_raid3_event *ep;
605 	struct g_raid3_disk *disk;
606 	struct g_geom *gp;
607 	struct g_consumer *cp;
608 	u_int n;
609 
610 	g_topology_assert_not();
611 	sx_assert(&sc->sc_lock, SX_XLOCKED);
612 
613 	gp = sc->sc_geom;
614 	if (sc->sc_provider != NULL)
615 		g_raid3_destroy_provider(sc);
616 	for (n = 0; n < sc->sc_ndisks; n++) {
617 		disk = &sc->sc_disks[n];
618 		if (disk->d_state != G_RAID3_DISK_STATE_NODISK) {
619 			disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
620 			g_raid3_update_metadata(disk);
621 			g_raid3_destroy_disk(disk);
622 		}
623 	}
624 	while ((ep = g_raid3_event_get(sc)) != NULL) {
625 		g_raid3_event_remove(sc, ep);
626 		if ((ep->e_flags & G_RAID3_EVENT_DONTWAIT) != 0)
627 			g_raid3_event_free(ep);
628 		else {
629 			ep->e_error = ECANCELED;
630 			ep->e_flags |= G_RAID3_EVENT_DONE;
631 			G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, ep);
632 			mtx_lock(&sc->sc_events_mtx);
633 			wakeup(ep);
634 			mtx_unlock(&sc->sc_events_mtx);
635 		}
636 	}
637 	callout_drain(&sc->sc_callout);
638 	cp = LIST_FIRST(&sc->sc_sync.ds_geom->consumer);
639 	g_topology_lock();
640 	if (cp != NULL)
641 		g_raid3_disconnect_consumer(sc, cp);
642 	g_wither_geom(sc->sc_sync.ds_geom, ENXIO);
643 	G_RAID3_DEBUG(0, "Device %s destroyed.", gp->name);
644 	g_wither_geom(gp, ENXIO);
645 	g_topology_unlock();
646 	if (!g_raid3_use_malloc) {
647 		uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_64K].sz_zone);
648 		uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_16K].sz_zone);
649 		uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_4K].sz_zone);
650 	}
651 	mtx_destroy(&sc->sc_queue_mtx);
652 	mtx_destroy(&sc->sc_events_mtx);
653 	sx_xunlock(&sc->sc_lock);
654 	sx_destroy(&sc->sc_lock);
655 }
656 
657 static void
658 g_raid3_orphan(struct g_consumer *cp)
659 {
660 	struct g_raid3_disk *disk;
661 
662 	g_topology_assert();
663 
664 	disk = cp->private;
665 	if (disk == NULL)
666 		return;
667 	disk->d_softc->sc_bump_id = G_RAID3_BUMP_SYNCID;
668 	g_raid3_event_send(disk, G_RAID3_DISK_STATE_DISCONNECTED,
669 	    G_RAID3_EVENT_DONTWAIT);
670 }
671 
672 static int
673 g_raid3_write_metadata(struct g_raid3_disk *disk, struct g_raid3_metadata *md)
674 {
675 	struct g_raid3_softc *sc;
676 	struct g_consumer *cp;
677 	off_t offset, length;
678 	u_char *sector;
679 	int error = 0;
680 
681 	g_topology_assert_not();
682 	sc = disk->d_softc;
683 	sx_assert(&sc->sc_lock, SX_LOCKED);
684 
685 	cp = disk->d_consumer;
686 	KASSERT(cp != NULL, ("NULL consumer (%s).", sc->sc_name));
687 	KASSERT(cp->provider != NULL, ("NULL provider (%s).", sc->sc_name));
688 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
689 	    ("Consumer %s closed? (r%dw%de%d).", cp->provider->name, cp->acr,
690 	    cp->acw, cp->ace));
691 	length = cp->provider->sectorsize;
692 	offset = cp->provider->mediasize - length;
693 	sector = malloc((size_t)length, M_RAID3, M_WAITOK | M_ZERO);
694 	if (md != NULL)
695 		raid3_metadata_encode(md, sector);
696 	error = g_write_data(cp, offset, sector, length);
697 	free(sector, M_RAID3);
698 	if (error != 0) {
699 		if ((disk->d_flags & G_RAID3_DISK_FLAG_BROKEN) == 0) {
700 			G_RAID3_DEBUG(0, "Cannot write metadata on %s "
701 			    "(device=%s, error=%d).",
702 			    g_raid3_get_diskname(disk), sc->sc_name, error);
703 			disk->d_flags |= G_RAID3_DISK_FLAG_BROKEN;
704 		} else {
705 			G_RAID3_DEBUG(1, "Cannot write metadata on %s "
706 			    "(device=%s, error=%d).",
707 			    g_raid3_get_diskname(disk), sc->sc_name, error);
708 		}
709 		if (g_raid3_disconnect_on_failure &&
710 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
711 			sc->sc_bump_id |= G_RAID3_BUMP_GENID;
712 			g_raid3_event_send(disk,
713 			    G_RAID3_DISK_STATE_DISCONNECTED,
714 			    G_RAID3_EVENT_DONTWAIT);
715 		}
716 	}
717 	return (error);
718 }
719 
720 int
721 g_raid3_clear_metadata(struct g_raid3_disk *disk)
722 {
723 	int error;
724 
725 	g_topology_assert_not();
726 	sx_assert(&disk->d_softc->sc_lock, SX_LOCKED);
727 
728 	error = g_raid3_write_metadata(disk, NULL);
729 	if (error == 0) {
730 		G_RAID3_DEBUG(2, "Metadata on %s cleared.",
731 		    g_raid3_get_diskname(disk));
732 	} else {
733 		G_RAID3_DEBUG(0,
734 		    "Cannot clear metadata on disk %s (error=%d).",
735 		    g_raid3_get_diskname(disk), error);
736 	}
737 	return (error);
738 }
739 
740 void
741 g_raid3_fill_metadata(struct g_raid3_disk *disk, struct g_raid3_metadata *md)
742 {
743 	struct g_raid3_softc *sc;
744 	struct g_provider *pp;
745 
746 	sc = disk->d_softc;
747 	strlcpy(md->md_magic, G_RAID3_MAGIC, sizeof(md->md_magic));
748 	md->md_version = G_RAID3_VERSION;
749 	strlcpy(md->md_name, sc->sc_name, sizeof(md->md_name));
750 	md->md_id = sc->sc_id;
751 	md->md_all = sc->sc_ndisks;
752 	md->md_genid = sc->sc_genid;
753 	md->md_mediasize = sc->sc_mediasize;
754 	md->md_sectorsize = sc->sc_sectorsize;
755 	md->md_mflags = (sc->sc_flags & G_RAID3_DEVICE_FLAG_MASK);
756 	md->md_no = disk->d_no;
757 	md->md_syncid = disk->d_sync.ds_syncid;
758 	md->md_dflags = (disk->d_flags & G_RAID3_DISK_FLAG_MASK);
759 	if (disk->d_state != G_RAID3_DISK_STATE_SYNCHRONIZING)
760 		md->md_sync_offset = 0;
761 	else {
762 		md->md_sync_offset =
763 		    disk->d_sync.ds_offset_done / (sc->sc_ndisks - 1);
764 	}
765 	if (disk->d_consumer != NULL && disk->d_consumer->provider != NULL)
766 		pp = disk->d_consumer->provider;
767 	else
768 		pp = NULL;
769 	if ((disk->d_flags & G_RAID3_DISK_FLAG_HARDCODED) != 0 && pp != NULL)
770 		strlcpy(md->md_provider, pp->name, sizeof(md->md_provider));
771 	else
772 		bzero(md->md_provider, sizeof(md->md_provider));
773 	if (pp != NULL)
774 		md->md_provsize = pp->mediasize;
775 	else
776 		md->md_provsize = 0;
777 }
778 
779 void
780 g_raid3_update_metadata(struct g_raid3_disk *disk)
781 {
782 	struct g_raid3_softc *sc;
783 	struct g_raid3_metadata md;
784 	int error;
785 
786 	g_topology_assert_not();
787 	sc = disk->d_softc;
788 	sx_assert(&sc->sc_lock, SX_LOCKED);
789 
790 	g_raid3_fill_metadata(disk, &md);
791 	error = g_raid3_write_metadata(disk, &md);
792 	if (error == 0) {
793 		G_RAID3_DEBUG(2, "Metadata on %s updated.",
794 		    g_raid3_get_diskname(disk));
795 	} else {
796 		G_RAID3_DEBUG(0,
797 		    "Cannot update metadata on disk %s (error=%d).",
798 		    g_raid3_get_diskname(disk), error);
799 	}
800 }
801 
802 static void
803 g_raid3_bump_syncid(struct g_raid3_softc *sc)
804 {
805 	struct g_raid3_disk *disk;
806 	u_int n;
807 
808 	g_topology_assert_not();
809 	sx_assert(&sc->sc_lock, SX_XLOCKED);
810 	KASSERT(g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) > 0,
811 	    ("%s called with no active disks (device=%s).", __func__,
812 	    sc->sc_name));
813 
814 	sc->sc_syncid++;
815 	G_RAID3_DEBUG(1, "Device %s: syncid bumped to %u.", sc->sc_name,
816 	    sc->sc_syncid);
817 	for (n = 0; n < sc->sc_ndisks; n++) {
818 		disk = &sc->sc_disks[n];
819 		if (disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
820 		    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
821 			disk->d_sync.ds_syncid = sc->sc_syncid;
822 			g_raid3_update_metadata(disk);
823 		}
824 	}
825 }
826 
827 static void
828 g_raid3_bump_genid(struct g_raid3_softc *sc)
829 {
830 	struct g_raid3_disk *disk;
831 	u_int n;
832 
833 	g_topology_assert_not();
834 	sx_assert(&sc->sc_lock, SX_XLOCKED);
835 	KASSERT(g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) > 0,
836 	    ("%s called with no active disks (device=%s).", __func__,
837 	    sc->sc_name));
838 
839 	sc->sc_genid++;
840 	G_RAID3_DEBUG(1, "Device %s: genid bumped to %u.", sc->sc_name,
841 	    sc->sc_genid);
842 	for (n = 0; n < sc->sc_ndisks; n++) {
843 		disk = &sc->sc_disks[n];
844 		if (disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
845 		    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
846 			disk->d_genid = sc->sc_genid;
847 			g_raid3_update_metadata(disk);
848 		}
849 	}
850 }
851 
852 static int
853 g_raid3_idle(struct g_raid3_softc *sc, int acw)
854 {
855 	struct g_raid3_disk *disk;
856 	u_int i;
857 	int timeout;
858 
859 	g_topology_assert_not();
860 	sx_assert(&sc->sc_lock, SX_XLOCKED);
861 
862 	if (sc->sc_provider == NULL)
863 		return (0);
864 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) != 0)
865 		return (0);
866 	if (sc->sc_idle)
867 		return (0);
868 	if (sc->sc_writes > 0)
869 		return (0);
870 	if (acw > 0 || (acw == -1 && sc->sc_provider->acw > 0)) {
871 		timeout = g_raid3_idletime - (time_uptime - sc->sc_last_write);
872 		if (timeout > 0)
873 			return (timeout);
874 	}
875 	sc->sc_idle = 1;
876 	for (i = 0; i < sc->sc_ndisks; i++) {
877 		disk = &sc->sc_disks[i];
878 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE)
879 			continue;
880 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as clean.",
881 		    g_raid3_get_diskname(disk), sc->sc_name);
882 		disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
883 		g_raid3_update_metadata(disk);
884 	}
885 	return (0);
886 }
887 
888 static void
889 g_raid3_unidle(struct g_raid3_softc *sc)
890 {
891 	struct g_raid3_disk *disk;
892 	u_int i;
893 
894 	g_topology_assert_not();
895 	sx_assert(&sc->sc_lock, SX_XLOCKED);
896 
897 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) != 0)
898 		return;
899 	sc->sc_idle = 0;
900 	sc->sc_last_write = time_uptime;
901 	for (i = 0; i < sc->sc_ndisks; i++) {
902 		disk = &sc->sc_disks[i];
903 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE)
904 			continue;
905 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as dirty.",
906 		    g_raid3_get_diskname(disk), sc->sc_name);
907 		disk->d_flags |= G_RAID3_DISK_FLAG_DIRTY;
908 		g_raid3_update_metadata(disk);
909 	}
910 }
911 
912 /*
913  * Treat bio_driver1 field in parent bio as list head and field bio_caller1
914  * in child bio as pointer to the next element on the list.
915  */
916 #define	G_RAID3_HEAD_BIO(pbp)	(pbp)->bio_driver1
917 
918 #define	G_RAID3_NEXT_BIO(cbp)	(cbp)->bio_caller1
919 
920 #define	G_RAID3_FOREACH_BIO(pbp, bp)					\
921 	for ((bp) = G_RAID3_HEAD_BIO(pbp); (bp) != NULL;		\
922 	    (bp) = G_RAID3_NEXT_BIO(bp))
923 
924 #define	G_RAID3_FOREACH_SAFE_BIO(pbp, bp, tmpbp)			\
925 	for ((bp) = G_RAID3_HEAD_BIO(pbp);				\
926 	    (bp) != NULL && ((tmpbp) = G_RAID3_NEXT_BIO(bp), 1);	\
927 	    (bp) = (tmpbp))
928 
929 static void
930 g_raid3_init_bio(struct bio *pbp)
931 {
932 
933 	G_RAID3_HEAD_BIO(pbp) = NULL;
934 }
935 
936 static void
937 g_raid3_remove_bio(struct bio *cbp)
938 {
939 	struct bio *pbp, *bp;
940 
941 	pbp = cbp->bio_parent;
942 	if (G_RAID3_HEAD_BIO(pbp) == cbp)
943 		G_RAID3_HEAD_BIO(pbp) = G_RAID3_NEXT_BIO(cbp);
944 	else {
945 		G_RAID3_FOREACH_BIO(pbp, bp) {
946 			if (G_RAID3_NEXT_BIO(bp) == cbp) {
947 				G_RAID3_NEXT_BIO(bp) = G_RAID3_NEXT_BIO(cbp);
948 				break;
949 			}
950 		}
951 	}
952 	G_RAID3_NEXT_BIO(cbp) = NULL;
953 }
954 
955 static void
956 g_raid3_replace_bio(struct bio *sbp, struct bio *dbp)
957 {
958 	struct bio *pbp, *bp;
959 
960 	g_raid3_remove_bio(sbp);
961 	pbp = dbp->bio_parent;
962 	G_RAID3_NEXT_BIO(sbp) = G_RAID3_NEXT_BIO(dbp);
963 	if (G_RAID3_HEAD_BIO(pbp) == dbp)
964 		G_RAID3_HEAD_BIO(pbp) = sbp;
965 	else {
966 		G_RAID3_FOREACH_BIO(pbp, bp) {
967 			if (G_RAID3_NEXT_BIO(bp) == dbp) {
968 				G_RAID3_NEXT_BIO(bp) = sbp;
969 				break;
970 			}
971 		}
972 	}
973 	G_RAID3_NEXT_BIO(dbp) = NULL;
974 }
975 
976 static void
977 g_raid3_destroy_bio(struct g_raid3_softc *sc, struct bio *cbp)
978 {
979 	struct bio *bp, *pbp;
980 	size_t size;
981 
982 	pbp = cbp->bio_parent;
983 	pbp->bio_children--;
984 	KASSERT(cbp->bio_data != NULL, ("NULL bio_data"));
985 	size = pbp->bio_length / (sc->sc_ndisks - 1);
986 	g_raid3_free(sc, cbp->bio_data, size);
987 	if (G_RAID3_HEAD_BIO(pbp) == cbp) {
988 		G_RAID3_HEAD_BIO(pbp) = G_RAID3_NEXT_BIO(cbp);
989 		G_RAID3_NEXT_BIO(cbp) = NULL;
990 		g_destroy_bio(cbp);
991 	} else {
992 		G_RAID3_FOREACH_BIO(pbp, bp) {
993 			if (G_RAID3_NEXT_BIO(bp) == cbp)
994 				break;
995 		}
996 		if (bp != NULL) {
997 			KASSERT(G_RAID3_NEXT_BIO(bp) != NULL,
998 			    ("NULL bp->bio_driver1"));
999 			G_RAID3_NEXT_BIO(bp) = G_RAID3_NEXT_BIO(cbp);
1000 			G_RAID3_NEXT_BIO(cbp) = NULL;
1001 		}
1002 		g_destroy_bio(cbp);
1003 	}
1004 }
1005 
1006 static struct bio *
1007 g_raid3_clone_bio(struct g_raid3_softc *sc, struct bio *pbp)
1008 {
1009 	struct bio *bp, *cbp;
1010 	size_t size;
1011 	int memflag;
1012 
1013 	cbp = g_clone_bio(pbp);
1014 	if (cbp == NULL)
1015 		return (NULL);
1016 	size = pbp->bio_length / (sc->sc_ndisks - 1);
1017 	if ((pbp->bio_cflags & G_RAID3_BIO_CFLAG_REGULAR) != 0)
1018 		memflag = M_WAITOK;
1019 	else
1020 		memflag = M_NOWAIT;
1021 	cbp->bio_data = g_raid3_alloc(sc, size, memflag);
1022 	if (cbp->bio_data == NULL) {
1023 		pbp->bio_children--;
1024 		g_destroy_bio(cbp);
1025 		return (NULL);
1026 	}
1027 	G_RAID3_NEXT_BIO(cbp) = NULL;
1028 	if (G_RAID3_HEAD_BIO(pbp) == NULL)
1029 		G_RAID3_HEAD_BIO(pbp) = cbp;
1030 	else {
1031 		G_RAID3_FOREACH_BIO(pbp, bp) {
1032 			if (G_RAID3_NEXT_BIO(bp) == NULL) {
1033 				G_RAID3_NEXT_BIO(bp) = cbp;
1034 				break;
1035 			}
1036 		}
1037 	}
1038 	return (cbp);
1039 }
1040 
1041 static void
1042 g_raid3_scatter(struct bio *pbp)
1043 {
1044 	struct g_raid3_softc *sc;
1045 	struct g_raid3_disk *disk;
1046 	struct bio *bp, *cbp, *tmpbp;
1047 	off_t atom, cadd, padd, left;
1048 
1049 	sc = pbp->bio_to->geom->softc;
1050 	bp = NULL;
1051 	if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_NOPARITY) == 0) {
1052 		/*
1053 		 * Find bio for which we should calculate data.
1054 		 */
1055 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1056 			if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_PARITY) != 0) {
1057 				bp = cbp;
1058 				break;
1059 			}
1060 		}
1061 		KASSERT(bp != NULL, ("NULL parity bio."));
1062 	}
1063 	atom = sc->sc_sectorsize / (sc->sc_ndisks - 1);
1064 	cadd = padd = 0;
1065 	for (left = pbp->bio_length; left > 0; left -= sc->sc_sectorsize) {
1066 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1067 			if (cbp == bp)
1068 				continue;
1069 			bcopy(pbp->bio_data + padd, cbp->bio_data + cadd, atom);
1070 			padd += atom;
1071 		}
1072 		cadd += atom;
1073 	}
1074 	if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_NOPARITY) == 0) {
1075 		/*
1076 		 * Calculate parity.
1077 		 */
1078 		bzero(bp->bio_data, bp->bio_length);
1079 		G_RAID3_FOREACH_SAFE_BIO(pbp, cbp, tmpbp) {
1080 			if (cbp == bp)
1081 				continue;
1082 			g_raid3_xor(cbp->bio_data, bp->bio_data, bp->bio_data,
1083 			    bp->bio_length);
1084 			if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_NODISK) != 0)
1085 				g_raid3_destroy_bio(sc, cbp);
1086 		}
1087 	}
1088 	G_RAID3_FOREACH_SAFE_BIO(pbp, cbp, tmpbp) {
1089 		struct g_consumer *cp;
1090 
1091 		disk = cbp->bio_caller2;
1092 		cp = disk->d_consumer;
1093 		cbp->bio_to = cp->provider;
1094 		G_RAID3_LOGREQ(3, cbp, "Sending request.");
1095 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1096 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1097 		    cp->acr, cp->acw, cp->ace));
1098 		cp->index++;
1099 		sc->sc_writes++;
1100 		g_io_request(cbp, cp);
1101 	}
1102 }
1103 
1104 static void
1105 g_raid3_gather(struct bio *pbp)
1106 {
1107 	struct g_raid3_softc *sc;
1108 	struct g_raid3_disk *disk;
1109 	struct bio *xbp, *fbp, *cbp;
1110 	off_t atom, cadd, padd, left;
1111 
1112 	sc = pbp->bio_to->geom->softc;
1113 	/*
1114 	 * Find bio for which we have to calculate data.
1115 	 * While going through this path, check if all requests
1116 	 * succeeded, if not, deny whole request.
1117 	 * If we're in COMPLETE mode, we allow one request to fail,
1118 	 * so if we find one, we're sending it to the parity consumer.
1119 	 * If there are more failed requests, we deny whole request.
1120 	 */
1121 	xbp = fbp = NULL;
1122 	G_RAID3_FOREACH_BIO(pbp, cbp) {
1123 		if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_PARITY) != 0) {
1124 			KASSERT(xbp == NULL, ("More than one parity bio."));
1125 			xbp = cbp;
1126 		}
1127 		if (cbp->bio_error == 0)
1128 			continue;
1129 		/*
1130 		 * Found failed request.
1131 		 */
1132 		if (fbp == NULL) {
1133 			if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_DEGRADED) != 0) {
1134 				/*
1135 				 * We are already in degraded mode, so we can't
1136 				 * accept any failures.
1137 				 */
1138 				if (pbp->bio_error == 0)
1139 					pbp->bio_error = cbp->bio_error;
1140 			} else {
1141 				fbp = cbp;
1142 			}
1143 		} else {
1144 			/*
1145 			 * Next failed request, that's too many.
1146 			 */
1147 			if (pbp->bio_error == 0)
1148 				pbp->bio_error = fbp->bio_error;
1149 		}
1150 		disk = cbp->bio_caller2;
1151 		if (disk == NULL)
1152 			continue;
1153 		if ((disk->d_flags & G_RAID3_DISK_FLAG_BROKEN) == 0) {
1154 			disk->d_flags |= G_RAID3_DISK_FLAG_BROKEN;
1155 			G_RAID3_LOGREQ(0, cbp, "Request failed (error=%d).",
1156 			    cbp->bio_error);
1157 		} else {
1158 			G_RAID3_LOGREQ(1, cbp, "Request failed (error=%d).",
1159 			    cbp->bio_error);
1160 		}
1161 		if (g_raid3_disconnect_on_failure &&
1162 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1163 			sc->sc_bump_id |= G_RAID3_BUMP_GENID;
1164 			g_raid3_event_send(disk,
1165 			    G_RAID3_DISK_STATE_DISCONNECTED,
1166 			    G_RAID3_EVENT_DONTWAIT);
1167 		}
1168 	}
1169 	if (pbp->bio_error != 0)
1170 		goto finish;
1171 	if (fbp != NULL && (pbp->bio_pflags & G_RAID3_BIO_PFLAG_VERIFY) != 0) {
1172 		pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_VERIFY;
1173 		if (xbp != fbp)
1174 			g_raid3_replace_bio(xbp, fbp);
1175 		g_raid3_destroy_bio(sc, fbp);
1176 	} else if (fbp != NULL) {
1177 		struct g_consumer *cp;
1178 
1179 		/*
1180 		 * One request failed, so send the same request to
1181 		 * the parity consumer.
1182 		 */
1183 		disk = pbp->bio_driver2;
1184 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE) {
1185 			pbp->bio_error = fbp->bio_error;
1186 			goto finish;
1187 		}
1188 		pbp->bio_pflags |= G_RAID3_BIO_PFLAG_DEGRADED;
1189 		pbp->bio_inbed--;
1190 		fbp->bio_flags &= ~(BIO_DONE | BIO_ERROR);
1191 		if (disk->d_no == sc->sc_ndisks - 1)
1192 			fbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1193 		fbp->bio_error = 0;
1194 		fbp->bio_completed = 0;
1195 		fbp->bio_children = 0;
1196 		fbp->bio_inbed = 0;
1197 		cp = disk->d_consumer;
1198 		fbp->bio_caller2 = disk;
1199 		fbp->bio_to = cp->provider;
1200 		G_RAID3_LOGREQ(3, fbp, "Sending request (recover).");
1201 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1202 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1203 		    cp->acr, cp->acw, cp->ace));
1204 		cp->index++;
1205 		g_io_request(fbp, cp);
1206 		return;
1207 	}
1208 	if (xbp != NULL) {
1209 		/*
1210 		 * Calculate parity.
1211 		 */
1212 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1213 			if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_PARITY) != 0)
1214 				continue;
1215 			g_raid3_xor(cbp->bio_data, xbp->bio_data, xbp->bio_data,
1216 			    xbp->bio_length);
1217 		}
1218 		xbp->bio_cflags &= ~G_RAID3_BIO_CFLAG_PARITY;
1219 		if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_VERIFY) != 0) {
1220 			if (!g_raid3_is_zero(xbp)) {
1221 				g_raid3_parity_mismatch++;
1222 				pbp->bio_error = EIO;
1223 				goto finish;
1224 			}
1225 			g_raid3_destroy_bio(sc, xbp);
1226 		}
1227 	}
1228 	atom = sc->sc_sectorsize / (sc->sc_ndisks - 1);
1229 	cadd = padd = 0;
1230 	for (left = pbp->bio_length; left > 0; left -= sc->sc_sectorsize) {
1231 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1232 			bcopy(cbp->bio_data + cadd, pbp->bio_data + padd, atom);
1233 			pbp->bio_completed += atom;
1234 			padd += atom;
1235 		}
1236 		cadd += atom;
1237 	}
1238 finish:
1239 	if (pbp->bio_error == 0)
1240 		G_RAID3_LOGREQ(3, pbp, "Request finished.");
1241 	else {
1242 		if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_VERIFY) != 0)
1243 			G_RAID3_LOGREQ(1, pbp, "Verification error.");
1244 		else
1245 			G_RAID3_LOGREQ(0, pbp, "Request failed.");
1246 	}
1247 	pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_MASK;
1248 	while ((cbp = G_RAID3_HEAD_BIO(pbp)) != NULL)
1249 		g_raid3_destroy_bio(sc, cbp);
1250 	g_io_deliver(pbp, pbp->bio_error);
1251 }
1252 
1253 static void
1254 g_raid3_done(struct bio *bp)
1255 {
1256 	struct g_raid3_softc *sc;
1257 
1258 	sc = bp->bio_from->geom->softc;
1259 	bp->bio_cflags |= G_RAID3_BIO_CFLAG_REGULAR;
1260 	G_RAID3_LOGREQ(3, bp, "Regular request done (error=%d).", bp->bio_error);
1261 	mtx_lock(&sc->sc_queue_mtx);
1262 	bioq_insert_head(&sc->sc_queue, bp);
1263 	wakeup(sc);
1264 	wakeup(&sc->sc_queue);
1265 	mtx_unlock(&sc->sc_queue_mtx);
1266 }
1267 
1268 static void
1269 g_raid3_regular_request(struct bio *cbp)
1270 {
1271 	struct g_raid3_softc *sc;
1272 	struct g_raid3_disk *disk;
1273 	struct bio *pbp;
1274 
1275 	g_topology_assert_not();
1276 
1277 	pbp = cbp->bio_parent;
1278 	sc = pbp->bio_to->geom->softc;
1279 	cbp->bio_from->index--;
1280 	if (cbp->bio_cmd == BIO_WRITE)
1281 		sc->sc_writes--;
1282 	disk = cbp->bio_from->private;
1283 	if (disk == NULL) {
1284 		g_topology_lock();
1285 		g_raid3_kill_consumer(sc, cbp->bio_from);
1286 		g_topology_unlock();
1287 	}
1288 
1289 	G_RAID3_LOGREQ(3, cbp, "Request finished.");
1290 	pbp->bio_inbed++;
1291 	KASSERT(pbp->bio_inbed <= pbp->bio_children,
1292 	    ("bio_inbed (%u) is bigger than bio_children (%u).", pbp->bio_inbed,
1293 	    pbp->bio_children));
1294 	if (pbp->bio_inbed != pbp->bio_children)
1295 		return;
1296 	switch (pbp->bio_cmd) {
1297 	case BIO_READ:
1298 		g_raid3_gather(pbp);
1299 		break;
1300 	case BIO_WRITE:
1301 	case BIO_DELETE:
1302 	    {
1303 		int error = 0;
1304 
1305 		pbp->bio_completed = pbp->bio_length;
1306 		while ((cbp = G_RAID3_HEAD_BIO(pbp)) != NULL) {
1307 			if (cbp->bio_error == 0) {
1308 				g_raid3_destroy_bio(sc, cbp);
1309 				continue;
1310 			}
1311 
1312 			if (error == 0)
1313 				error = cbp->bio_error;
1314 			else if (pbp->bio_error == 0) {
1315 				/*
1316 				 * Next failed request, that's too many.
1317 				 */
1318 				pbp->bio_error = error;
1319 			}
1320 
1321 			disk = cbp->bio_caller2;
1322 			if (disk == NULL) {
1323 				g_raid3_destroy_bio(sc, cbp);
1324 				continue;
1325 			}
1326 
1327 			if ((disk->d_flags & G_RAID3_DISK_FLAG_BROKEN) == 0) {
1328 				disk->d_flags |= G_RAID3_DISK_FLAG_BROKEN;
1329 				G_RAID3_LOGREQ(0, cbp,
1330 				    "Request failed (error=%d).",
1331 				    cbp->bio_error);
1332 			} else {
1333 				G_RAID3_LOGREQ(1, cbp,
1334 				    "Request failed (error=%d).",
1335 				    cbp->bio_error);
1336 			}
1337 			if (g_raid3_disconnect_on_failure &&
1338 			    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1339 				sc->sc_bump_id |= G_RAID3_BUMP_GENID;
1340 				g_raid3_event_send(disk,
1341 				    G_RAID3_DISK_STATE_DISCONNECTED,
1342 				    G_RAID3_EVENT_DONTWAIT);
1343 			}
1344 			g_raid3_destroy_bio(sc, cbp);
1345 		}
1346 		if (pbp->bio_error == 0)
1347 			G_RAID3_LOGREQ(3, pbp, "Request finished.");
1348 		else
1349 			G_RAID3_LOGREQ(0, pbp, "Request failed.");
1350 		pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_DEGRADED;
1351 		pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_NOPARITY;
1352 		bioq_remove(&sc->sc_inflight, pbp);
1353 		/* Release delayed sync requests if possible. */
1354 		g_raid3_sync_release(sc);
1355 		g_io_deliver(pbp, pbp->bio_error);
1356 		break;
1357 	    }
1358 	}
1359 }
1360 
1361 static void
1362 g_raid3_sync_done(struct bio *bp)
1363 {
1364 	struct g_raid3_softc *sc;
1365 
1366 	G_RAID3_LOGREQ(3, bp, "Synchronization request delivered.");
1367 	sc = bp->bio_from->geom->softc;
1368 	bp->bio_cflags |= G_RAID3_BIO_CFLAG_SYNC;
1369 	mtx_lock(&sc->sc_queue_mtx);
1370 	bioq_insert_head(&sc->sc_queue, bp);
1371 	wakeup(sc);
1372 	wakeup(&sc->sc_queue);
1373 	mtx_unlock(&sc->sc_queue_mtx);
1374 }
1375 
1376 static void
1377 g_raid3_flush(struct g_raid3_softc *sc, struct bio *bp)
1378 {
1379 	struct bio_queue_head queue;
1380 	struct g_raid3_disk *disk;
1381 	struct g_consumer *cp;
1382 	struct bio *cbp;
1383 	u_int i;
1384 
1385 	bioq_init(&queue);
1386 	for (i = 0; i < sc->sc_ndisks; i++) {
1387 		disk = &sc->sc_disks[i];
1388 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE)
1389 			continue;
1390 		cbp = g_clone_bio(bp);
1391 		if (cbp == NULL) {
1392 			for (cbp = bioq_first(&queue); cbp != NULL;
1393 			    cbp = bioq_first(&queue)) {
1394 				bioq_remove(&queue, cbp);
1395 				g_destroy_bio(cbp);
1396 			}
1397 			if (bp->bio_error == 0)
1398 				bp->bio_error = ENOMEM;
1399 			g_io_deliver(bp, bp->bio_error);
1400 			return;
1401 		}
1402 		bioq_insert_tail(&queue, cbp);
1403 		cbp->bio_done = g_std_done;
1404 		cbp->bio_caller1 = disk;
1405 		cbp->bio_to = disk->d_consumer->provider;
1406 	}
1407 	for (cbp = bioq_first(&queue); cbp != NULL; cbp = bioq_first(&queue)) {
1408 		bioq_remove(&queue, cbp);
1409 		G_RAID3_LOGREQ(3, cbp, "Sending request.");
1410 		disk = cbp->bio_caller1;
1411 		cbp->bio_caller1 = NULL;
1412 		cp = disk->d_consumer;
1413 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1414 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1415 		    cp->acr, cp->acw, cp->ace));
1416 		g_io_request(cbp, disk->d_consumer);
1417 	}
1418 }
1419 
1420 static void
1421 g_raid3_start(struct bio *bp)
1422 {
1423 	struct g_raid3_softc *sc;
1424 
1425 	sc = bp->bio_to->geom->softc;
1426 	/*
1427 	 * If sc == NULL or there are no valid disks, provider's error
1428 	 * should be set and g_raid3_start() should not be called at all.
1429 	 */
1430 	KASSERT(sc != NULL && (sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
1431 	    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE),
1432 	    ("Provider's error should be set (error=%d)(device=%s).",
1433 	    bp->bio_to->error, bp->bio_to->name));
1434 	G_RAID3_LOGREQ(3, bp, "Request received.");
1435 
1436 	switch (bp->bio_cmd) {
1437 	case BIO_READ:
1438 	case BIO_WRITE:
1439 	case BIO_DELETE:
1440 		break;
1441 	case BIO_FLUSH:
1442 		g_raid3_flush(sc, bp);
1443 		return;
1444 	case BIO_GETATTR:
1445 	default:
1446 		g_io_deliver(bp, EOPNOTSUPP);
1447 		return;
1448 	}
1449 	mtx_lock(&sc->sc_queue_mtx);
1450 	bioq_insert_tail(&sc->sc_queue, bp);
1451 	G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1452 	wakeup(sc);
1453 	mtx_unlock(&sc->sc_queue_mtx);
1454 }
1455 
1456 /*
1457  * Return TRUE if the given request is colliding with a in-progress
1458  * synchronization request.
1459  */
1460 static int
1461 g_raid3_sync_collision(struct g_raid3_softc *sc, struct bio *bp)
1462 {
1463 	struct g_raid3_disk *disk;
1464 	struct bio *sbp;
1465 	off_t rstart, rend, sstart, send;
1466 	int i;
1467 
1468 	disk = sc->sc_syncdisk;
1469 	if (disk == NULL)
1470 		return (0);
1471 	rstart = bp->bio_offset;
1472 	rend = bp->bio_offset + bp->bio_length;
1473 	for (i = 0; i < g_raid3_syncreqs; i++) {
1474 		sbp = disk->d_sync.ds_bios[i];
1475 		if (sbp == NULL)
1476 			continue;
1477 		sstart = sbp->bio_offset;
1478 		send = sbp->bio_length;
1479 		if (sbp->bio_cmd == BIO_WRITE) {
1480 			sstart *= sc->sc_ndisks - 1;
1481 			send *= sc->sc_ndisks - 1;
1482 		}
1483 		send += sstart;
1484 		if (rend > sstart && rstart < send)
1485 			return (1);
1486 	}
1487 	return (0);
1488 }
1489 
1490 /*
1491  * Return TRUE if the given sync request is colliding with a in-progress regular
1492  * request.
1493  */
1494 static int
1495 g_raid3_regular_collision(struct g_raid3_softc *sc, struct bio *sbp)
1496 {
1497 	off_t rstart, rend, sstart, send;
1498 	struct bio *bp;
1499 
1500 	if (sc->sc_syncdisk == NULL)
1501 		return (0);
1502 	sstart = sbp->bio_offset;
1503 	send = sstart + sbp->bio_length;
1504 	TAILQ_FOREACH(bp, &sc->sc_inflight.queue, bio_queue) {
1505 		rstart = bp->bio_offset;
1506 		rend = bp->bio_offset + bp->bio_length;
1507 		if (rend > sstart && rstart < send)
1508 			return (1);
1509 	}
1510 	return (0);
1511 }
1512 
1513 /*
1514  * Puts request onto delayed queue.
1515  */
1516 static void
1517 g_raid3_regular_delay(struct g_raid3_softc *sc, struct bio *bp)
1518 {
1519 
1520 	G_RAID3_LOGREQ(2, bp, "Delaying request.");
1521 	bioq_insert_head(&sc->sc_regular_delayed, bp);
1522 }
1523 
1524 /*
1525  * Puts synchronization request onto delayed queue.
1526  */
1527 static void
1528 g_raid3_sync_delay(struct g_raid3_softc *sc, struct bio *bp)
1529 {
1530 
1531 	G_RAID3_LOGREQ(2, bp, "Delaying synchronization request.");
1532 	bioq_insert_tail(&sc->sc_sync_delayed, bp);
1533 }
1534 
1535 /*
1536  * Releases delayed regular requests which don't collide anymore with sync
1537  * requests.
1538  */
1539 static void
1540 g_raid3_regular_release(struct g_raid3_softc *sc)
1541 {
1542 	struct bio *bp, *bp2;
1543 
1544 	TAILQ_FOREACH_SAFE(bp, &sc->sc_regular_delayed.queue, bio_queue, bp2) {
1545 		if (g_raid3_sync_collision(sc, bp))
1546 			continue;
1547 		bioq_remove(&sc->sc_regular_delayed, bp);
1548 		G_RAID3_LOGREQ(2, bp, "Releasing delayed request (%p).", bp);
1549 		mtx_lock(&sc->sc_queue_mtx);
1550 		bioq_insert_head(&sc->sc_queue, bp);
1551 #if 0
1552 		/*
1553 		 * wakeup() is not needed, because this function is called from
1554 		 * the worker thread.
1555 		 */
1556 		wakeup(&sc->sc_queue);
1557 #endif
1558 		mtx_unlock(&sc->sc_queue_mtx);
1559 	}
1560 }
1561 
1562 /*
1563  * Releases delayed sync requests which don't collide anymore with regular
1564  * requests.
1565  */
1566 static void
1567 g_raid3_sync_release(struct g_raid3_softc *sc)
1568 {
1569 	struct bio *bp, *bp2;
1570 
1571 	TAILQ_FOREACH_SAFE(bp, &sc->sc_sync_delayed.queue, bio_queue, bp2) {
1572 		if (g_raid3_regular_collision(sc, bp))
1573 			continue;
1574 		bioq_remove(&sc->sc_sync_delayed, bp);
1575 		G_RAID3_LOGREQ(2, bp,
1576 		    "Releasing delayed synchronization request.");
1577 		g_io_request(bp, bp->bio_from);
1578 	}
1579 }
1580 
1581 /*
1582  * Handle synchronization requests.
1583  * Every synchronization request is two-steps process: first, READ request is
1584  * send to active provider and then WRITE request (with read data) to the provider
1585  * beeing synchronized. When WRITE is finished, new synchronization request is
1586  * send.
1587  */
1588 static void
1589 g_raid3_sync_request(struct bio *bp)
1590 {
1591 	struct g_raid3_softc *sc;
1592 	struct g_raid3_disk *disk;
1593 
1594 	bp->bio_from->index--;
1595 	sc = bp->bio_from->geom->softc;
1596 	disk = bp->bio_from->private;
1597 	if (disk == NULL) {
1598 		sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1599 		g_topology_lock();
1600 		g_raid3_kill_consumer(sc, bp->bio_from);
1601 		g_topology_unlock();
1602 		free(bp->bio_data, M_RAID3);
1603 		g_destroy_bio(bp);
1604 		sx_xlock(&sc->sc_lock);
1605 		return;
1606 	}
1607 
1608 	/*
1609 	 * Synchronization request.
1610 	 */
1611 	switch (bp->bio_cmd) {
1612 	case BIO_READ:
1613 	    {
1614 		struct g_consumer *cp;
1615 		u_char *dst, *src;
1616 		off_t left;
1617 		u_int atom;
1618 
1619 		if (bp->bio_error != 0) {
1620 			G_RAID3_LOGREQ(0, bp,
1621 			    "Synchronization request failed (error=%d).",
1622 			    bp->bio_error);
1623 			g_destroy_bio(bp);
1624 			return;
1625 		}
1626 		G_RAID3_LOGREQ(3, bp, "Synchronization request finished.");
1627 		atom = sc->sc_sectorsize / (sc->sc_ndisks - 1);
1628 		dst = src = bp->bio_data;
1629 		if (disk->d_no == sc->sc_ndisks - 1) {
1630 			u_int n;
1631 
1632 			/* Parity component. */
1633 			for (left = bp->bio_length; left > 0;
1634 			    left -= sc->sc_sectorsize) {
1635 				bcopy(src, dst, atom);
1636 				src += atom;
1637 				for (n = 1; n < sc->sc_ndisks - 1; n++) {
1638 					g_raid3_xor(src, dst, dst, atom);
1639 					src += atom;
1640 				}
1641 				dst += atom;
1642 			}
1643 		} else {
1644 			/* Regular component. */
1645 			src += atom * disk->d_no;
1646 			for (left = bp->bio_length; left > 0;
1647 			    left -= sc->sc_sectorsize) {
1648 				bcopy(src, dst, atom);
1649 				src += sc->sc_sectorsize;
1650 				dst += atom;
1651 			}
1652 		}
1653 		bp->bio_driver1 = bp->bio_driver2 = NULL;
1654 		bp->bio_pflags = 0;
1655 		bp->bio_offset /= sc->sc_ndisks - 1;
1656 		bp->bio_length /= sc->sc_ndisks - 1;
1657 		bp->bio_cmd = BIO_WRITE;
1658 		bp->bio_cflags = 0;
1659 		bp->bio_children = bp->bio_inbed = 0;
1660 		cp = disk->d_consumer;
1661 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1662 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1663 		    cp->acr, cp->acw, cp->ace));
1664 		cp->index++;
1665 		g_io_request(bp, cp);
1666 		return;
1667 	    }
1668 	case BIO_WRITE:
1669 	    {
1670 		struct g_raid3_disk_sync *sync;
1671 		off_t boffset, moffset;
1672 		void *data;
1673 		int i;
1674 
1675 		if (bp->bio_error != 0) {
1676 			G_RAID3_LOGREQ(0, bp,
1677 			    "Synchronization request failed (error=%d).",
1678 			    bp->bio_error);
1679 			g_destroy_bio(bp);
1680 			sc->sc_bump_id |= G_RAID3_BUMP_GENID;
1681 			g_raid3_event_send(disk,
1682 			    G_RAID3_DISK_STATE_DISCONNECTED,
1683 			    G_RAID3_EVENT_DONTWAIT);
1684 			return;
1685 		}
1686 		G_RAID3_LOGREQ(3, bp, "Synchronization request finished.");
1687 		sync = &disk->d_sync;
1688 		if (sync->ds_offset == sc->sc_mediasize / (sc->sc_ndisks - 1) ||
1689 		    sync->ds_consumer == NULL ||
1690 		    (sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
1691 			/* Don't send more synchronization requests. */
1692 			sync->ds_inflight--;
1693 			if (sync->ds_bios != NULL) {
1694 				i = (int)(uintptr_t)bp->bio_caller1;
1695 				sync->ds_bios[i] = NULL;
1696 			}
1697 			free(bp->bio_data, M_RAID3);
1698 			g_destroy_bio(bp);
1699 			if (sync->ds_inflight > 0)
1700 				return;
1701 			if (sync->ds_consumer == NULL ||
1702 			    (sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
1703 				return;
1704 			}
1705 			/*
1706 			 * Disk up-to-date, activate it.
1707 			 */
1708 			g_raid3_event_send(disk, G_RAID3_DISK_STATE_ACTIVE,
1709 			    G_RAID3_EVENT_DONTWAIT);
1710 			return;
1711 		}
1712 
1713 		/* Send next synchronization request. */
1714 		data = bp->bio_data;
1715 		bzero(bp, sizeof(*bp));
1716 		bp->bio_cmd = BIO_READ;
1717 		bp->bio_offset = sync->ds_offset * (sc->sc_ndisks - 1);
1718 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1719 		sync->ds_offset += bp->bio_length / (sc->sc_ndisks - 1);
1720 		bp->bio_done = g_raid3_sync_done;
1721 		bp->bio_data = data;
1722 		bp->bio_from = sync->ds_consumer;
1723 		bp->bio_to = sc->sc_provider;
1724 		G_RAID3_LOGREQ(3, bp, "Sending synchronization request.");
1725 		sync->ds_consumer->index++;
1726 		/*
1727 		 * Delay the request if it is colliding with a regular request.
1728 		 */
1729 		if (g_raid3_regular_collision(sc, bp))
1730 			g_raid3_sync_delay(sc, bp);
1731 		else
1732 			g_io_request(bp, sync->ds_consumer);
1733 
1734 		/* Release delayed requests if possible. */
1735 		g_raid3_regular_release(sc);
1736 
1737 		/* Find the smallest offset. */
1738 		moffset = sc->sc_mediasize;
1739 		for (i = 0; i < g_raid3_syncreqs; i++) {
1740 			bp = sync->ds_bios[i];
1741 			boffset = bp->bio_offset;
1742 			if (bp->bio_cmd == BIO_WRITE)
1743 				boffset *= sc->sc_ndisks - 1;
1744 			if (boffset < moffset)
1745 				moffset = boffset;
1746 		}
1747 		if (sync->ds_offset_done + (MAXPHYS * 100) < moffset) {
1748 			/* Update offset_done on every 100 blocks. */
1749 			sync->ds_offset_done = moffset;
1750 			g_raid3_update_metadata(disk);
1751 		}
1752 		return;
1753 	    }
1754 	default:
1755 		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1756 		    bp->bio_cmd, sc->sc_name));
1757 		break;
1758 	}
1759 }
1760 
1761 static int
1762 g_raid3_register_request(struct bio *pbp)
1763 {
1764 	struct g_raid3_softc *sc;
1765 	struct g_raid3_disk *disk;
1766 	struct g_consumer *cp;
1767 	struct bio *cbp, *tmpbp;
1768 	off_t offset, length;
1769 	u_int n, ndisks;
1770 	int round_robin, verify;
1771 
1772 	ndisks = 0;
1773 	sc = pbp->bio_to->geom->softc;
1774 	if ((pbp->bio_cflags & G_RAID3_BIO_CFLAG_REGSYNC) != 0 &&
1775 	    sc->sc_syncdisk == NULL) {
1776 		g_io_deliver(pbp, EIO);
1777 		return (0);
1778 	}
1779 	g_raid3_init_bio(pbp);
1780 	length = pbp->bio_length / (sc->sc_ndisks - 1);
1781 	offset = pbp->bio_offset / (sc->sc_ndisks - 1);
1782 	round_robin = verify = 0;
1783 	switch (pbp->bio_cmd) {
1784 	case BIO_READ:
1785 		if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_VERIFY) != 0 &&
1786 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1787 			pbp->bio_pflags |= G_RAID3_BIO_PFLAG_VERIFY;
1788 			verify = 1;
1789 			ndisks = sc->sc_ndisks;
1790 		} else {
1791 			verify = 0;
1792 			ndisks = sc->sc_ndisks - 1;
1793 		}
1794 		if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_ROUND_ROBIN) != 0 &&
1795 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1796 			round_robin = 1;
1797 		} else {
1798 			round_robin = 0;
1799 		}
1800 		KASSERT(!round_robin || !verify,
1801 		    ("ROUND-ROBIN and VERIFY are mutually exclusive."));
1802 		pbp->bio_driver2 = &sc->sc_disks[sc->sc_ndisks - 1];
1803 		break;
1804 	case BIO_WRITE:
1805 	case BIO_DELETE:
1806 		/*
1807 		 * Delay the request if it is colliding with a synchronization
1808 		 * request.
1809 		 */
1810 		if (g_raid3_sync_collision(sc, pbp)) {
1811 			g_raid3_regular_delay(sc, pbp);
1812 			return (0);
1813 		}
1814 
1815 		if (sc->sc_idle)
1816 			g_raid3_unidle(sc);
1817 		else
1818 			sc->sc_last_write = time_uptime;
1819 
1820 		ndisks = sc->sc_ndisks;
1821 		break;
1822 	}
1823 	for (n = 0; n < ndisks; n++) {
1824 		disk = &sc->sc_disks[n];
1825 		cbp = g_raid3_clone_bio(sc, pbp);
1826 		if (cbp == NULL) {
1827 			while ((cbp = G_RAID3_HEAD_BIO(pbp)) != NULL)
1828 				g_raid3_destroy_bio(sc, cbp);
1829 			/*
1830 			 * To prevent deadlock, we must run back up
1831 			 * with the ENOMEM for failed requests of any
1832 			 * of our consumers.  Our own sync requests
1833 			 * can stick around, as they are finite.
1834 			 */
1835 			if ((pbp->bio_cflags &
1836 			    G_RAID3_BIO_CFLAG_REGULAR) != 0) {
1837 				g_io_deliver(pbp, ENOMEM);
1838 				return (0);
1839 			}
1840 			return (ENOMEM);
1841 		}
1842 		cbp->bio_offset = offset;
1843 		cbp->bio_length = length;
1844 		cbp->bio_done = g_raid3_done;
1845 		switch (pbp->bio_cmd) {
1846 		case BIO_READ:
1847 			if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE) {
1848 				/*
1849 				 * Replace invalid component with the parity
1850 				 * component.
1851 				 */
1852 				disk = &sc->sc_disks[sc->sc_ndisks - 1];
1853 				cbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1854 				pbp->bio_pflags |= G_RAID3_BIO_PFLAG_DEGRADED;
1855 			} else if (round_robin &&
1856 			    disk->d_no == sc->sc_round_robin) {
1857 				/*
1858 				 * In round-robin mode skip one data component
1859 				 * and use parity component when reading.
1860 				 */
1861 				pbp->bio_driver2 = disk;
1862 				disk = &sc->sc_disks[sc->sc_ndisks - 1];
1863 				cbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1864 				sc->sc_round_robin++;
1865 				round_robin = 0;
1866 			} else if (verify && disk->d_no == sc->sc_ndisks - 1) {
1867 				cbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1868 			}
1869 			break;
1870 		case BIO_WRITE:
1871 		case BIO_DELETE:
1872 			if (disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
1873 			    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
1874 				if (n == ndisks - 1) {
1875 					/*
1876 					 * Active parity component, mark it as such.
1877 					 */
1878 					cbp->bio_cflags |=
1879 					    G_RAID3_BIO_CFLAG_PARITY;
1880 				}
1881 			} else {
1882 				pbp->bio_pflags |= G_RAID3_BIO_PFLAG_DEGRADED;
1883 				if (n == ndisks - 1) {
1884 					/*
1885 					 * Parity component is not connected,
1886 					 * so destroy its request.
1887 					 */
1888 					pbp->bio_pflags |=
1889 					    G_RAID3_BIO_PFLAG_NOPARITY;
1890 					g_raid3_destroy_bio(sc, cbp);
1891 					cbp = NULL;
1892 				} else {
1893 					cbp->bio_cflags |=
1894 					    G_RAID3_BIO_CFLAG_NODISK;
1895 					disk = NULL;
1896 				}
1897 			}
1898 			break;
1899 		}
1900 		if (cbp != NULL)
1901 			cbp->bio_caller2 = disk;
1902 	}
1903 	switch (pbp->bio_cmd) {
1904 	case BIO_READ:
1905 		if (round_robin) {
1906 			/*
1907 			 * If we are in round-robin mode and 'round_robin' is
1908 			 * still 1, it means, that we skipped parity component
1909 			 * for this read and must reset sc_round_robin field.
1910 			 */
1911 			sc->sc_round_robin = 0;
1912 		}
1913 		G_RAID3_FOREACH_SAFE_BIO(pbp, cbp, tmpbp) {
1914 			disk = cbp->bio_caller2;
1915 			cp = disk->d_consumer;
1916 			cbp->bio_to = cp->provider;
1917 			G_RAID3_LOGREQ(3, cbp, "Sending request.");
1918 			KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1919 			    ("Consumer %s not opened (r%dw%de%d).",
1920 			    cp->provider->name, cp->acr, cp->acw, cp->ace));
1921 			cp->index++;
1922 			g_io_request(cbp, cp);
1923 		}
1924 		break;
1925 	case BIO_WRITE:
1926 	case BIO_DELETE:
1927 		/*
1928 		 * Put request onto inflight queue, so we can check if new
1929 		 * synchronization requests don't collide with it.
1930 		 */
1931 		bioq_insert_tail(&sc->sc_inflight, pbp);
1932 
1933 		/*
1934 		 * Bump syncid on first write.
1935 		 */
1936 		if ((sc->sc_bump_id & G_RAID3_BUMP_SYNCID) != 0) {
1937 			sc->sc_bump_id &= ~G_RAID3_BUMP_SYNCID;
1938 			g_raid3_bump_syncid(sc);
1939 		}
1940 		g_raid3_scatter(pbp);
1941 		break;
1942 	}
1943 	return (0);
1944 }
1945 
1946 static int
1947 g_raid3_can_destroy(struct g_raid3_softc *sc)
1948 {
1949 	struct g_geom *gp;
1950 	struct g_consumer *cp;
1951 
1952 	g_topology_assert();
1953 	gp = sc->sc_geom;
1954 	if (gp->softc == NULL)
1955 		return (1);
1956 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1957 		if (g_raid3_is_busy(sc, cp))
1958 			return (0);
1959 	}
1960 	gp = sc->sc_sync.ds_geom;
1961 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1962 		if (g_raid3_is_busy(sc, cp))
1963 			return (0);
1964 	}
1965 	G_RAID3_DEBUG(2, "No I/O requests for %s, it can be destroyed.",
1966 	    sc->sc_name);
1967 	return (1);
1968 }
1969 
1970 static int
1971 g_raid3_try_destroy(struct g_raid3_softc *sc)
1972 {
1973 
1974 	g_topology_assert_not();
1975 	sx_assert(&sc->sc_lock, SX_XLOCKED);
1976 
1977 	if (sc->sc_rootmount != NULL) {
1978 		G_RAID3_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
1979 		    sc->sc_rootmount);
1980 		root_mount_rel(sc->sc_rootmount);
1981 		sc->sc_rootmount = NULL;
1982 	}
1983 
1984 	g_topology_lock();
1985 	if (!g_raid3_can_destroy(sc)) {
1986 		g_topology_unlock();
1987 		return (0);
1988 	}
1989 	sc->sc_geom->softc = NULL;
1990 	sc->sc_sync.ds_geom->softc = NULL;
1991 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_WAIT) != 0) {
1992 		g_topology_unlock();
1993 		G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__,
1994 		    &sc->sc_worker);
1995 		/* Unlock sc_lock here, as it can be destroyed after wakeup. */
1996 		sx_xunlock(&sc->sc_lock);
1997 		wakeup(&sc->sc_worker);
1998 		sc->sc_worker = NULL;
1999 	} else {
2000 		g_topology_unlock();
2001 		g_raid3_destroy_device(sc);
2002 		free(sc->sc_disks, M_RAID3);
2003 		free(sc, M_RAID3);
2004 	}
2005 	return (1);
2006 }
2007 
2008 /*
2009  * Worker thread.
2010  */
2011 static void
2012 g_raid3_worker(void *arg)
2013 {
2014 	struct g_raid3_softc *sc;
2015 	struct g_raid3_event *ep;
2016 	struct bio *bp;
2017 	int timeout;
2018 
2019 	sc = arg;
2020 	thread_lock(curthread);
2021 	sched_prio(curthread, PRIBIO);
2022 	thread_unlock(curthread);
2023 
2024 	sx_xlock(&sc->sc_lock);
2025 	for (;;) {
2026 		G_RAID3_DEBUG(5, "%s: Let's see...", __func__);
2027 		/*
2028 		 * First take a look at events.
2029 		 * This is important to handle events before any I/O requests.
2030 		 */
2031 		ep = g_raid3_event_get(sc);
2032 		if (ep != NULL) {
2033 			g_raid3_event_remove(sc, ep);
2034 			if ((ep->e_flags & G_RAID3_EVENT_DEVICE) != 0) {
2035 				/* Update only device status. */
2036 				G_RAID3_DEBUG(3,
2037 				    "Running event for device %s.",
2038 				    sc->sc_name);
2039 				ep->e_error = 0;
2040 				g_raid3_update_device(sc, 1);
2041 			} else {
2042 				/* Update disk status. */
2043 				G_RAID3_DEBUG(3, "Running event for disk %s.",
2044 				     g_raid3_get_diskname(ep->e_disk));
2045 				ep->e_error = g_raid3_update_disk(ep->e_disk,
2046 				    ep->e_state);
2047 				if (ep->e_error == 0)
2048 					g_raid3_update_device(sc, 0);
2049 			}
2050 			if ((ep->e_flags & G_RAID3_EVENT_DONTWAIT) != 0) {
2051 				KASSERT(ep->e_error == 0,
2052 				    ("Error cannot be handled."));
2053 				g_raid3_event_free(ep);
2054 			} else {
2055 				ep->e_flags |= G_RAID3_EVENT_DONE;
2056 				G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__,
2057 				    ep);
2058 				mtx_lock(&sc->sc_events_mtx);
2059 				wakeup(ep);
2060 				mtx_unlock(&sc->sc_events_mtx);
2061 			}
2062 			if ((sc->sc_flags &
2063 			    G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
2064 				if (g_raid3_try_destroy(sc)) {
2065 					curthread->td_pflags &= ~TDP_GEOM;
2066 					G_RAID3_DEBUG(1, "Thread exiting.");
2067 					kproc_exit(0);
2068 				}
2069 			}
2070 			G_RAID3_DEBUG(5, "%s: I'm here 1.", __func__);
2071 			continue;
2072 		}
2073 		/*
2074 		 * Check if we can mark array as CLEAN and if we can't take
2075 		 * how much seconds should we wait.
2076 		 */
2077 		timeout = g_raid3_idle(sc, -1);
2078 		/*
2079 		 * Now I/O requests.
2080 		 */
2081 		/* Get first request from the queue. */
2082 		mtx_lock(&sc->sc_queue_mtx);
2083 		bp = bioq_first(&sc->sc_queue);
2084 		if (bp == NULL) {
2085 			if ((sc->sc_flags &
2086 			    G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
2087 				mtx_unlock(&sc->sc_queue_mtx);
2088 				if (g_raid3_try_destroy(sc)) {
2089 					curthread->td_pflags &= ~TDP_GEOM;
2090 					G_RAID3_DEBUG(1, "Thread exiting.");
2091 					kproc_exit(0);
2092 				}
2093 				mtx_lock(&sc->sc_queue_mtx);
2094 			}
2095 			sx_xunlock(&sc->sc_lock);
2096 			/*
2097 			 * XXX: We can miss an event here, because an event
2098 			 *      can be added without sx-device-lock and without
2099 			 *      mtx-queue-lock. Maybe I should just stop using
2100 			 *      dedicated mutex for events synchronization and
2101 			 *      stick with the queue lock?
2102 			 *      The event will hang here until next I/O request
2103 			 *      or next event is received.
2104 			 */
2105 			MSLEEP(sc, &sc->sc_queue_mtx, PRIBIO | PDROP, "r3:w1",
2106 			    timeout * hz);
2107 			sx_xlock(&sc->sc_lock);
2108 			G_RAID3_DEBUG(5, "%s: I'm here 4.", __func__);
2109 			continue;
2110 		}
2111 process:
2112 		bioq_remove(&sc->sc_queue, bp);
2113 		mtx_unlock(&sc->sc_queue_mtx);
2114 
2115 		if (bp->bio_from->geom == sc->sc_sync.ds_geom &&
2116 		    (bp->bio_cflags & G_RAID3_BIO_CFLAG_SYNC) != 0) {
2117 			g_raid3_sync_request(bp);	/* READ */
2118 		} else if (bp->bio_to != sc->sc_provider) {
2119 			if ((bp->bio_cflags & G_RAID3_BIO_CFLAG_REGULAR) != 0)
2120 				g_raid3_regular_request(bp);
2121 			else if ((bp->bio_cflags & G_RAID3_BIO_CFLAG_SYNC) != 0)
2122 				g_raid3_sync_request(bp);	/* WRITE */
2123 			else {
2124 				KASSERT(0,
2125 				    ("Invalid request cflags=0x%hhx to=%s.",
2126 				    bp->bio_cflags, bp->bio_to->name));
2127 			}
2128 		} else if (g_raid3_register_request(bp) != 0) {
2129 			mtx_lock(&sc->sc_queue_mtx);
2130 			bioq_insert_head(&sc->sc_queue, bp);
2131 			/*
2132 			 * We are short in memory, let see if there are finished
2133 			 * request we can free.
2134 			 */
2135 			TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
2136 				if (bp->bio_cflags & G_RAID3_BIO_CFLAG_REGULAR)
2137 					goto process;
2138 			}
2139 			/*
2140 			 * No finished regular request, so at least keep
2141 			 * synchronization running.
2142 			 */
2143 			TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
2144 				if (bp->bio_cflags & G_RAID3_BIO_CFLAG_SYNC)
2145 					goto process;
2146 			}
2147 			sx_xunlock(&sc->sc_lock);
2148 			MSLEEP(&sc->sc_queue, &sc->sc_queue_mtx, PRIBIO | PDROP,
2149 			    "r3:lowmem", hz / 10);
2150 			sx_xlock(&sc->sc_lock);
2151 		}
2152 		G_RAID3_DEBUG(5, "%s: I'm here 9.", __func__);
2153 	}
2154 }
2155 
2156 static void
2157 g_raid3_update_idle(struct g_raid3_softc *sc, struct g_raid3_disk *disk)
2158 {
2159 
2160 	sx_assert(&sc->sc_lock, SX_LOCKED);
2161 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) != 0)
2162 		return;
2163 	if (!sc->sc_idle && (disk->d_flags & G_RAID3_DISK_FLAG_DIRTY) == 0) {
2164 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as dirty.",
2165 		    g_raid3_get_diskname(disk), sc->sc_name);
2166 		disk->d_flags |= G_RAID3_DISK_FLAG_DIRTY;
2167 	} else if (sc->sc_idle &&
2168 	    (disk->d_flags & G_RAID3_DISK_FLAG_DIRTY) != 0) {
2169 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as clean.",
2170 		    g_raid3_get_diskname(disk), sc->sc_name);
2171 		disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2172 	}
2173 }
2174 
2175 static void
2176 g_raid3_sync_start(struct g_raid3_softc *sc)
2177 {
2178 	struct g_raid3_disk *disk;
2179 	struct g_consumer *cp;
2180 	struct bio *bp;
2181 	int error;
2182 	u_int n;
2183 
2184 	g_topology_assert_not();
2185 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2186 
2187 	KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED,
2188 	    ("Device not in DEGRADED state (%s, %u).", sc->sc_name,
2189 	    sc->sc_state));
2190 	KASSERT(sc->sc_syncdisk == NULL, ("Syncdisk is not NULL (%s, %u).",
2191 	    sc->sc_name, sc->sc_state));
2192 	disk = NULL;
2193 	for (n = 0; n < sc->sc_ndisks; n++) {
2194 		if (sc->sc_disks[n].d_state != G_RAID3_DISK_STATE_SYNCHRONIZING)
2195 			continue;
2196 		disk = &sc->sc_disks[n];
2197 		break;
2198 	}
2199 	if (disk == NULL)
2200 		return;
2201 
2202 	sx_xunlock(&sc->sc_lock);
2203 	g_topology_lock();
2204 	cp = g_new_consumer(sc->sc_sync.ds_geom);
2205 	error = g_attach(cp, sc->sc_provider);
2206 	KASSERT(error == 0,
2207 	    ("Cannot attach to %s (error=%d).", sc->sc_name, error));
2208 	error = g_access(cp, 1, 0, 0);
2209 	KASSERT(error == 0, ("Cannot open %s (error=%d).", sc->sc_name, error));
2210 	g_topology_unlock();
2211 	sx_xlock(&sc->sc_lock);
2212 
2213 	G_RAID3_DEBUG(0, "Device %s: rebuilding provider %s.", sc->sc_name,
2214 	    g_raid3_get_diskname(disk));
2215 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) == 0)
2216 		disk->d_flags |= G_RAID3_DISK_FLAG_DIRTY;
2217 	KASSERT(disk->d_sync.ds_consumer == NULL,
2218 	    ("Sync consumer already exists (device=%s, disk=%s).",
2219 	    sc->sc_name, g_raid3_get_diskname(disk)));
2220 
2221 	disk->d_sync.ds_consumer = cp;
2222 	disk->d_sync.ds_consumer->private = disk;
2223 	disk->d_sync.ds_consumer->index = 0;
2224 	sc->sc_syncdisk = disk;
2225 
2226 	/*
2227 	 * Allocate memory for synchronization bios and initialize them.
2228 	 */
2229 	disk->d_sync.ds_bios = malloc(sizeof(struct bio *) * g_raid3_syncreqs,
2230 	    M_RAID3, M_WAITOK);
2231 	for (n = 0; n < g_raid3_syncreqs; n++) {
2232 		bp = g_alloc_bio();
2233 		disk->d_sync.ds_bios[n] = bp;
2234 		bp->bio_parent = NULL;
2235 		bp->bio_cmd = BIO_READ;
2236 		bp->bio_data = malloc(MAXPHYS, M_RAID3, M_WAITOK);
2237 		bp->bio_cflags = 0;
2238 		bp->bio_offset = disk->d_sync.ds_offset * (sc->sc_ndisks - 1);
2239 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
2240 		disk->d_sync.ds_offset += bp->bio_length / (sc->sc_ndisks - 1);
2241 		bp->bio_done = g_raid3_sync_done;
2242 		bp->bio_from = disk->d_sync.ds_consumer;
2243 		bp->bio_to = sc->sc_provider;
2244 		bp->bio_caller1 = (void *)(uintptr_t)n;
2245 	}
2246 
2247 	/* Set the number of in-flight synchronization requests. */
2248 	disk->d_sync.ds_inflight = g_raid3_syncreqs;
2249 
2250 	/*
2251 	 * Fire off first synchronization requests.
2252 	 */
2253 	for (n = 0; n < g_raid3_syncreqs; n++) {
2254 		bp = disk->d_sync.ds_bios[n];
2255 		G_RAID3_LOGREQ(3, bp, "Sending synchronization request.");
2256 		disk->d_sync.ds_consumer->index++;
2257 		/*
2258 		 * Delay the request if it is colliding with a regular request.
2259 		 */
2260 		if (g_raid3_regular_collision(sc, bp))
2261 			g_raid3_sync_delay(sc, bp);
2262 		else
2263 			g_io_request(bp, disk->d_sync.ds_consumer);
2264 	}
2265 }
2266 
2267 /*
2268  * Stop synchronization process.
2269  * type: 0 - synchronization finished
2270  *       1 - synchronization stopped
2271  */
2272 static void
2273 g_raid3_sync_stop(struct g_raid3_softc *sc, int type)
2274 {
2275 	struct g_raid3_disk *disk;
2276 	struct g_consumer *cp;
2277 
2278 	g_topology_assert_not();
2279 	sx_assert(&sc->sc_lock, SX_LOCKED);
2280 
2281 	KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED,
2282 	    ("Device not in DEGRADED state (%s, %u).", sc->sc_name,
2283 	    sc->sc_state));
2284 	disk = sc->sc_syncdisk;
2285 	sc->sc_syncdisk = NULL;
2286 	KASSERT(disk != NULL, ("No disk was synchronized (%s).", sc->sc_name));
2287 	KASSERT(disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING,
2288 	    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2289 	    g_raid3_disk_state2str(disk->d_state)));
2290 	if (disk->d_sync.ds_consumer == NULL)
2291 		return;
2292 
2293 	if (type == 0) {
2294 		G_RAID3_DEBUG(0, "Device %s: rebuilding provider %s finished.",
2295 		    sc->sc_name, g_raid3_get_diskname(disk));
2296 	} else /* if (type == 1) */ {
2297 		G_RAID3_DEBUG(0, "Device %s: rebuilding provider %s stopped.",
2298 		    sc->sc_name, g_raid3_get_diskname(disk));
2299 	}
2300 	free(disk->d_sync.ds_bios, M_RAID3);
2301 	disk->d_sync.ds_bios = NULL;
2302 	cp = disk->d_sync.ds_consumer;
2303 	disk->d_sync.ds_consumer = NULL;
2304 	disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2305 	sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
2306 	g_topology_lock();
2307 	g_raid3_kill_consumer(sc, cp);
2308 	g_topology_unlock();
2309 	sx_xlock(&sc->sc_lock);
2310 }
2311 
2312 static void
2313 g_raid3_launch_provider(struct g_raid3_softc *sc)
2314 {
2315 	struct g_provider *pp;
2316 
2317 	sx_assert(&sc->sc_lock, SX_LOCKED);
2318 
2319 	g_topology_lock();
2320 	pp = g_new_providerf(sc->sc_geom, "raid3/%s", sc->sc_name);
2321 	pp->mediasize = sc->sc_mediasize;
2322 	pp->sectorsize = sc->sc_sectorsize;
2323 	sc->sc_provider = pp;
2324 	g_error_provider(pp, 0);
2325 	g_topology_unlock();
2326 	G_RAID3_DEBUG(0, "Device %s launched (%u/%u).", pp->name,
2327 	    g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE), sc->sc_ndisks);
2328 
2329 	if (sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED)
2330 		g_raid3_sync_start(sc);
2331 }
2332 
2333 static void
2334 g_raid3_destroy_provider(struct g_raid3_softc *sc)
2335 {
2336 	struct bio *bp;
2337 
2338 	g_topology_assert_not();
2339 	KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
2340 	    sc->sc_name));
2341 
2342 	g_topology_lock();
2343 	g_error_provider(sc->sc_provider, ENXIO);
2344 	mtx_lock(&sc->sc_queue_mtx);
2345 	while ((bp = bioq_first(&sc->sc_queue)) != NULL) {
2346 		bioq_remove(&sc->sc_queue, bp);
2347 		g_io_deliver(bp, ENXIO);
2348 	}
2349 	mtx_unlock(&sc->sc_queue_mtx);
2350 	G_RAID3_DEBUG(0, "Device %s: provider %s destroyed.", sc->sc_name,
2351 	    sc->sc_provider->name);
2352 	sc->sc_provider->flags |= G_PF_WITHER;
2353 	g_orphan_provider(sc->sc_provider, ENXIO);
2354 	g_topology_unlock();
2355 	sc->sc_provider = NULL;
2356 	if (sc->sc_syncdisk != NULL)
2357 		g_raid3_sync_stop(sc, 1);
2358 }
2359 
2360 static void
2361 g_raid3_go(void *arg)
2362 {
2363 	struct g_raid3_softc *sc;
2364 
2365 	sc = arg;
2366 	G_RAID3_DEBUG(0, "Force device %s start due to timeout.", sc->sc_name);
2367 	g_raid3_event_send(sc, 0,
2368 	    G_RAID3_EVENT_DONTWAIT | G_RAID3_EVENT_DEVICE);
2369 }
2370 
2371 static u_int
2372 g_raid3_determine_state(struct g_raid3_disk *disk)
2373 {
2374 	struct g_raid3_softc *sc;
2375 	u_int state;
2376 
2377 	sc = disk->d_softc;
2378 	if (sc->sc_syncid == disk->d_sync.ds_syncid) {
2379 		if ((disk->d_flags &
2380 		    G_RAID3_DISK_FLAG_SYNCHRONIZING) == 0) {
2381 			/* Disk does not need synchronization. */
2382 			state = G_RAID3_DISK_STATE_ACTIVE;
2383 		} else {
2384 			if ((sc->sc_flags &
2385 			     G_RAID3_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2386 			    (disk->d_flags &
2387 			     G_RAID3_DISK_FLAG_FORCE_SYNC) != 0) {
2388 				/*
2389 				 * We can start synchronization from
2390 				 * the stored offset.
2391 				 */
2392 				state = G_RAID3_DISK_STATE_SYNCHRONIZING;
2393 			} else {
2394 				state = G_RAID3_DISK_STATE_STALE;
2395 			}
2396 		}
2397 	} else if (disk->d_sync.ds_syncid < sc->sc_syncid) {
2398 		/*
2399 		 * Reset all synchronization data for this disk,
2400 		 * because if it even was synchronized, it was
2401 		 * synchronized to disks with different syncid.
2402 		 */
2403 		disk->d_flags |= G_RAID3_DISK_FLAG_SYNCHRONIZING;
2404 		disk->d_sync.ds_offset = 0;
2405 		disk->d_sync.ds_offset_done = 0;
2406 		disk->d_sync.ds_syncid = sc->sc_syncid;
2407 		if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2408 		    (disk->d_flags & G_RAID3_DISK_FLAG_FORCE_SYNC) != 0) {
2409 			state = G_RAID3_DISK_STATE_SYNCHRONIZING;
2410 		} else {
2411 			state = G_RAID3_DISK_STATE_STALE;
2412 		}
2413 	} else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
2414 		/*
2415 		 * Not good, NOT GOOD!
2416 		 * It means that device was started on stale disks
2417 		 * and more fresh disk just arrive.
2418 		 * If there were writes, device is broken, sorry.
2419 		 * I think the best choice here is don't touch
2420 		 * this disk and inform the user loudly.
2421 		 */
2422 		G_RAID3_DEBUG(0, "Device %s was started before the freshest "
2423 		    "disk (%s) arrives!! It will not be connected to the "
2424 		    "running device.", sc->sc_name,
2425 		    g_raid3_get_diskname(disk));
2426 		g_raid3_destroy_disk(disk);
2427 		state = G_RAID3_DISK_STATE_NONE;
2428 		/* Return immediately, because disk was destroyed. */
2429 		return (state);
2430 	}
2431 	G_RAID3_DEBUG(3, "State for %s disk: %s.",
2432 	    g_raid3_get_diskname(disk), g_raid3_disk_state2str(state));
2433 	return (state);
2434 }
2435 
2436 /*
2437  * Update device state.
2438  */
2439 static void
2440 g_raid3_update_device(struct g_raid3_softc *sc, boolean_t force)
2441 {
2442 	struct g_raid3_disk *disk;
2443 	u_int state;
2444 
2445 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2446 
2447 	switch (sc->sc_state) {
2448 	case G_RAID3_DEVICE_STATE_STARTING:
2449 	    {
2450 		u_int n, ndirty, ndisks, genid, syncid;
2451 
2452 		KASSERT(sc->sc_provider == NULL,
2453 		    ("Non-NULL provider in STARTING state (%s).", sc->sc_name));
2454 		/*
2455 		 * Are we ready? We are, if all disks are connected or
2456 		 * one disk is missing and 'force' is true.
2457 		 */
2458 		if (g_raid3_ndisks(sc, -1) + force == sc->sc_ndisks) {
2459 			if (!force)
2460 				callout_drain(&sc->sc_callout);
2461 		} else {
2462 			if (force) {
2463 				/*
2464 				 * Timeout expired, so destroy device.
2465 				 */
2466 				sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
2467 				G_RAID3_DEBUG(1, "root_mount_rel[%u] %p",
2468 				    __LINE__, sc->sc_rootmount);
2469 				root_mount_rel(sc->sc_rootmount);
2470 				sc->sc_rootmount = NULL;
2471 			}
2472 			return;
2473 		}
2474 
2475 		/*
2476 		 * Find the biggest genid.
2477 		 */
2478 		genid = 0;
2479 		for (n = 0; n < sc->sc_ndisks; n++) {
2480 			disk = &sc->sc_disks[n];
2481 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2482 				continue;
2483 			if (disk->d_genid > genid)
2484 				genid = disk->d_genid;
2485 		}
2486 		sc->sc_genid = genid;
2487 		/*
2488 		 * Remove all disks without the biggest genid.
2489 		 */
2490 		for (n = 0; n < sc->sc_ndisks; n++) {
2491 			disk = &sc->sc_disks[n];
2492 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2493 				continue;
2494 			if (disk->d_genid < genid) {
2495 				G_RAID3_DEBUG(0,
2496 				    "Component %s (device %s) broken, skipping.",
2497 				    g_raid3_get_diskname(disk), sc->sc_name);
2498 				g_raid3_destroy_disk(disk);
2499 			}
2500 		}
2501 
2502 		/*
2503 		 * There must be at least 'sc->sc_ndisks - 1' components
2504 		 * with the same syncid and without SYNCHRONIZING flag.
2505 		 */
2506 
2507 		/*
2508 		 * Find the biggest syncid, number of valid components and
2509 		 * number of dirty components.
2510 		 */
2511 		ndirty = ndisks = syncid = 0;
2512 		for (n = 0; n < sc->sc_ndisks; n++) {
2513 			disk = &sc->sc_disks[n];
2514 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2515 				continue;
2516 			if ((disk->d_flags & G_RAID3_DISK_FLAG_DIRTY) != 0)
2517 				ndirty++;
2518 			if (disk->d_sync.ds_syncid > syncid) {
2519 				syncid = disk->d_sync.ds_syncid;
2520 				ndisks = 0;
2521 			} else if (disk->d_sync.ds_syncid < syncid) {
2522 				continue;
2523 			}
2524 			if ((disk->d_flags &
2525 			    G_RAID3_DISK_FLAG_SYNCHRONIZING) != 0) {
2526 				continue;
2527 			}
2528 			ndisks++;
2529 		}
2530 		/*
2531 		 * Do we have enough valid components?
2532 		 */
2533 		if (ndisks + 1 < sc->sc_ndisks) {
2534 			G_RAID3_DEBUG(0,
2535 			    "Device %s is broken, too few valid components.",
2536 			    sc->sc_name);
2537 			sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
2538 			return;
2539 		}
2540 		/*
2541 		 * If there is one DIRTY component and all disks are present,
2542 		 * mark it for synchronization. If there is more than one DIRTY
2543 		 * component, mark parity component for synchronization.
2544 		 */
2545 		if (ndisks == sc->sc_ndisks && ndirty == 1) {
2546 			for (n = 0; n < sc->sc_ndisks; n++) {
2547 				disk = &sc->sc_disks[n];
2548 				if ((disk->d_flags &
2549 				    G_RAID3_DISK_FLAG_DIRTY) == 0) {
2550 					continue;
2551 				}
2552 				disk->d_flags |=
2553 				    G_RAID3_DISK_FLAG_SYNCHRONIZING;
2554 			}
2555 		} else if (ndisks == sc->sc_ndisks && ndirty > 1) {
2556 			disk = &sc->sc_disks[sc->sc_ndisks - 1];
2557 			disk->d_flags |= G_RAID3_DISK_FLAG_SYNCHRONIZING;
2558 		}
2559 
2560 		sc->sc_syncid = syncid;
2561 		if (force) {
2562 			/* Remember to bump syncid on first write. */
2563 			sc->sc_bump_id |= G_RAID3_BUMP_SYNCID;
2564 		}
2565 		if (ndisks == sc->sc_ndisks)
2566 			state = G_RAID3_DEVICE_STATE_COMPLETE;
2567 		else /* if (ndisks == sc->sc_ndisks - 1) */
2568 			state = G_RAID3_DEVICE_STATE_DEGRADED;
2569 		G_RAID3_DEBUG(1, "Device %s state changed from %s to %s.",
2570 		    sc->sc_name, g_raid3_device_state2str(sc->sc_state),
2571 		    g_raid3_device_state2str(state));
2572 		sc->sc_state = state;
2573 		for (n = 0; n < sc->sc_ndisks; n++) {
2574 			disk = &sc->sc_disks[n];
2575 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2576 				continue;
2577 			state = g_raid3_determine_state(disk);
2578 			g_raid3_event_send(disk, state, G_RAID3_EVENT_DONTWAIT);
2579 			if (state == G_RAID3_DISK_STATE_STALE)
2580 				sc->sc_bump_id |= G_RAID3_BUMP_SYNCID;
2581 		}
2582 		break;
2583 	    }
2584 	case G_RAID3_DEVICE_STATE_DEGRADED:
2585 		/*
2586 		 * Genid need to be bumped immediately, so do it here.
2587 		 */
2588 		if ((sc->sc_bump_id & G_RAID3_BUMP_GENID) != 0) {
2589 			sc->sc_bump_id &= ~G_RAID3_BUMP_GENID;
2590 			g_raid3_bump_genid(sc);
2591 		}
2592 
2593 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_NEW) > 0)
2594 			return;
2595 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) <
2596 		    sc->sc_ndisks - 1) {
2597 			if (sc->sc_provider != NULL)
2598 				g_raid3_destroy_provider(sc);
2599 			sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
2600 			return;
2601 		}
2602 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) ==
2603 		    sc->sc_ndisks) {
2604 			state = G_RAID3_DEVICE_STATE_COMPLETE;
2605 			G_RAID3_DEBUG(1,
2606 			    "Device %s state changed from %s to %s.",
2607 			    sc->sc_name, g_raid3_device_state2str(sc->sc_state),
2608 			    g_raid3_device_state2str(state));
2609 			sc->sc_state = state;
2610 		}
2611 		if (sc->sc_provider == NULL)
2612 			g_raid3_launch_provider(sc);
2613 		if (sc->sc_rootmount != NULL) {
2614 			G_RAID3_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2615 			    sc->sc_rootmount);
2616 			root_mount_rel(sc->sc_rootmount);
2617 			sc->sc_rootmount = NULL;
2618 		}
2619 		break;
2620 	case G_RAID3_DEVICE_STATE_COMPLETE:
2621 		/*
2622 		 * Genid need to be bumped immediately, so do it here.
2623 		 */
2624 		if ((sc->sc_bump_id & G_RAID3_BUMP_GENID) != 0) {
2625 			sc->sc_bump_id &= ~G_RAID3_BUMP_GENID;
2626 			g_raid3_bump_genid(sc);
2627 		}
2628 
2629 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_NEW) > 0)
2630 			return;
2631 		KASSERT(g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) >=
2632 		    sc->sc_ndisks - 1,
2633 		    ("Too few ACTIVE components in COMPLETE state (device %s).",
2634 		    sc->sc_name));
2635 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) ==
2636 		    sc->sc_ndisks - 1) {
2637 			state = G_RAID3_DEVICE_STATE_DEGRADED;
2638 			G_RAID3_DEBUG(1,
2639 			    "Device %s state changed from %s to %s.",
2640 			    sc->sc_name, g_raid3_device_state2str(sc->sc_state),
2641 			    g_raid3_device_state2str(state));
2642 			sc->sc_state = state;
2643 		}
2644 		if (sc->sc_provider == NULL)
2645 			g_raid3_launch_provider(sc);
2646 		if (sc->sc_rootmount != NULL) {
2647 			G_RAID3_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2648 			    sc->sc_rootmount);
2649 			root_mount_rel(sc->sc_rootmount);
2650 			sc->sc_rootmount = NULL;
2651 		}
2652 		break;
2653 	default:
2654 		KASSERT(1 == 0, ("Wrong device state (%s, %s).", sc->sc_name,
2655 		    g_raid3_device_state2str(sc->sc_state)));
2656 		break;
2657 	}
2658 }
2659 
2660 /*
2661  * Update disk state and device state if needed.
2662  */
2663 #define	DISK_STATE_CHANGED()	G_RAID3_DEBUG(1,			\
2664 	"Disk %s state changed from %s to %s (device %s).",		\
2665 	g_raid3_get_diskname(disk),					\
2666 	g_raid3_disk_state2str(disk->d_state),				\
2667 	g_raid3_disk_state2str(state), sc->sc_name)
2668 static int
2669 g_raid3_update_disk(struct g_raid3_disk *disk, u_int state)
2670 {
2671 	struct g_raid3_softc *sc;
2672 
2673 	sc = disk->d_softc;
2674 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2675 
2676 again:
2677 	G_RAID3_DEBUG(3, "Changing disk %s state from %s to %s.",
2678 	    g_raid3_get_diskname(disk), g_raid3_disk_state2str(disk->d_state),
2679 	    g_raid3_disk_state2str(state));
2680 	switch (state) {
2681 	case G_RAID3_DISK_STATE_NEW:
2682 		/*
2683 		 * Possible scenarios:
2684 		 * 1. New disk arrive.
2685 		 */
2686 		/* Previous state should be NONE. */
2687 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NONE,
2688 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2689 		    g_raid3_disk_state2str(disk->d_state)));
2690 		DISK_STATE_CHANGED();
2691 
2692 		disk->d_state = state;
2693 		G_RAID3_DEBUG(1, "Device %s: provider %s detected.",
2694 		    sc->sc_name, g_raid3_get_diskname(disk));
2695 		if (sc->sc_state == G_RAID3_DEVICE_STATE_STARTING)
2696 			break;
2697 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2698 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2699 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2700 		    g_raid3_device_state2str(sc->sc_state),
2701 		    g_raid3_get_diskname(disk),
2702 		    g_raid3_disk_state2str(disk->d_state)));
2703 		state = g_raid3_determine_state(disk);
2704 		if (state != G_RAID3_DISK_STATE_NONE)
2705 			goto again;
2706 		break;
2707 	case G_RAID3_DISK_STATE_ACTIVE:
2708 		/*
2709 		 * Possible scenarios:
2710 		 * 1. New disk does not need synchronization.
2711 		 * 2. Synchronization process finished successfully.
2712 		 */
2713 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2714 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2715 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2716 		    g_raid3_device_state2str(sc->sc_state),
2717 		    g_raid3_get_diskname(disk),
2718 		    g_raid3_disk_state2str(disk->d_state)));
2719 		/* Previous state should be NEW or SYNCHRONIZING. */
2720 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW ||
2721 		    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING,
2722 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2723 		    g_raid3_disk_state2str(disk->d_state)));
2724 		DISK_STATE_CHANGED();
2725 
2726 		if (disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
2727 			disk->d_flags &= ~G_RAID3_DISK_FLAG_SYNCHRONIZING;
2728 			disk->d_flags &= ~G_RAID3_DISK_FLAG_FORCE_SYNC;
2729 			g_raid3_sync_stop(sc, 0);
2730 		}
2731 		disk->d_state = state;
2732 		disk->d_sync.ds_offset = 0;
2733 		disk->d_sync.ds_offset_done = 0;
2734 		g_raid3_update_idle(sc, disk);
2735 		g_raid3_update_metadata(disk);
2736 		G_RAID3_DEBUG(1, "Device %s: provider %s activated.",
2737 		    sc->sc_name, g_raid3_get_diskname(disk));
2738 		break;
2739 	case G_RAID3_DISK_STATE_STALE:
2740 		/*
2741 		 * Possible scenarios:
2742 		 * 1. Stale disk was connected.
2743 		 */
2744 		/* Previous state should be NEW. */
2745 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW,
2746 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2747 		    g_raid3_disk_state2str(disk->d_state)));
2748 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2749 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2750 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2751 		    g_raid3_device_state2str(sc->sc_state),
2752 		    g_raid3_get_diskname(disk),
2753 		    g_raid3_disk_state2str(disk->d_state)));
2754 		/*
2755 		 * STALE state is only possible if device is marked
2756 		 * NOAUTOSYNC.
2757 		 */
2758 		KASSERT((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOAUTOSYNC) != 0,
2759 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2760 		    g_raid3_device_state2str(sc->sc_state),
2761 		    g_raid3_get_diskname(disk),
2762 		    g_raid3_disk_state2str(disk->d_state)));
2763 		DISK_STATE_CHANGED();
2764 
2765 		disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2766 		disk->d_state = state;
2767 		g_raid3_update_metadata(disk);
2768 		G_RAID3_DEBUG(0, "Device %s: provider %s is stale.",
2769 		    sc->sc_name, g_raid3_get_diskname(disk));
2770 		break;
2771 	case G_RAID3_DISK_STATE_SYNCHRONIZING:
2772 		/*
2773 		 * Possible scenarios:
2774 		 * 1. Disk which needs synchronization was connected.
2775 		 */
2776 		/* Previous state should be NEW. */
2777 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW,
2778 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2779 		    g_raid3_disk_state2str(disk->d_state)));
2780 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2781 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2782 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2783 		    g_raid3_device_state2str(sc->sc_state),
2784 		    g_raid3_get_diskname(disk),
2785 		    g_raid3_disk_state2str(disk->d_state)));
2786 		DISK_STATE_CHANGED();
2787 
2788 		if (disk->d_state == G_RAID3_DISK_STATE_NEW)
2789 			disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2790 		disk->d_state = state;
2791 		if (sc->sc_provider != NULL) {
2792 			g_raid3_sync_start(sc);
2793 			g_raid3_update_metadata(disk);
2794 		}
2795 		break;
2796 	case G_RAID3_DISK_STATE_DISCONNECTED:
2797 		/*
2798 		 * Possible scenarios:
2799 		 * 1. Device wasn't running yet, but disk disappear.
2800 		 * 2. Disk was active and disapppear.
2801 		 * 3. Disk disappear during synchronization process.
2802 		 */
2803 		if (sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2804 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
2805 			/*
2806 			 * Previous state should be ACTIVE, STALE or
2807 			 * SYNCHRONIZING.
2808 			 */
2809 			KASSERT(disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
2810 			    disk->d_state == G_RAID3_DISK_STATE_STALE ||
2811 			    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING,
2812 			    ("Wrong disk state (%s, %s).",
2813 			    g_raid3_get_diskname(disk),
2814 			    g_raid3_disk_state2str(disk->d_state)));
2815 		} else if (sc->sc_state == G_RAID3_DEVICE_STATE_STARTING) {
2816 			/* Previous state should be NEW. */
2817 			KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW,
2818 			    ("Wrong disk state (%s, %s).",
2819 			    g_raid3_get_diskname(disk),
2820 			    g_raid3_disk_state2str(disk->d_state)));
2821 			/*
2822 			 * Reset bumping syncid if disk disappeared in STARTING
2823 			 * state.
2824 			 */
2825 			if ((sc->sc_bump_id & G_RAID3_BUMP_SYNCID) != 0)
2826 				sc->sc_bump_id &= ~G_RAID3_BUMP_SYNCID;
2827 #ifdef	INVARIANTS
2828 		} else {
2829 			KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
2830 			    sc->sc_name,
2831 			    g_raid3_device_state2str(sc->sc_state),
2832 			    g_raid3_get_diskname(disk),
2833 			    g_raid3_disk_state2str(disk->d_state)));
2834 #endif
2835 		}
2836 		DISK_STATE_CHANGED();
2837 		G_RAID3_DEBUG(0, "Device %s: provider %s disconnected.",
2838 		    sc->sc_name, g_raid3_get_diskname(disk));
2839 
2840 		g_raid3_destroy_disk(disk);
2841 		break;
2842 	default:
2843 		KASSERT(1 == 0, ("Unknown state (%u).", state));
2844 		break;
2845 	}
2846 	return (0);
2847 }
2848 #undef	DISK_STATE_CHANGED
2849 
2850 int
2851 g_raid3_read_metadata(struct g_consumer *cp, struct g_raid3_metadata *md)
2852 {
2853 	struct g_provider *pp;
2854 	u_char *buf;
2855 	int error;
2856 
2857 	g_topology_assert();
2858 
2859 	error = g_access(cp, 1, 0, 0);
2860 	if (error != 0)
2861 		return (error);
2862 	pp = cp->provider;
2863 	g_topology_unlock();
2864 	/* Metadata are stored on last sector. */
2865 	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
2866 	    &error);
2867 	g_topology_lock();
2868 	g_access(cp, -1, 0, 0);
2869 	if (buf == NULL) {
2870 		G_RAID3_DEBUG(1, "Cannot read metadata from %s (error=%d).",
2871 		    cp->provider->name, error);
2872 		return (error);
2873 	}
2874 
2875 	/* Decode metadata. */
2876 	error = raid3_metadata_decode(buf, md);
2877 	g_free(buf);
2878 	if (strcmp(md->md_magic, G_RAID3_MAGIC) != 0)
2879 		return (EINVAL);
2880 	if (md->md_version > G_RAID3_VERSION) {
2881 		G_RAID3_DEBUG(0,
2882 		    "Kernel module is too old to handle metadata from %s.",
2883 		    cp->provider->name);
2884 		return (EINVAL);
2885 	}
2886 	if (error != 0) {
2887 		G_RAID3_DEBUG(1, "MD5 metadata hash mismatch for provider %s.",
2888 		    cp->provider->name);
2889 		return (error);
2890 	}
2891 
2892 	return (0);
2893 }
2894 
2895 static int
2896 g_raid3_check_metadata(struct g_raid3_softc *sc, struct g_provider *pp,
2897     struct g_raid3_metadata *md)
2898 {
2899 
2900 	if (md->md_no >= sc->sc_ndisks) {
2901 		G_RAID3_DEBUG(1, "Invalid disk %s number (no=%u), skipping.",
2902 		    pp->name, md->md_no);
2903 		return (EINVAL);
2904 	}
2905 	if (sc->sc_disks[md->md_no].d_state != G_RAID3_DISK_STATE_NODISK) {
2906 		G_RAID3_DEBUG(1, "Disk %s (no=%u) already exists, skipping.",
2907 		    pp->name, md->md_no);
2908 		return (EEXIST);
2909 	}
2910 	if (md->md_all != sc->sc_ndisks) {
2911 		G_RAID3_DEBUG(1,
2912 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2913 		    "md_all", pp->name, sc->sc_name);
2914 		return (EINVAL);
2915 	}
2916 	if ((md->md_mediasize % md->md_sectorsize) != 0) {
2917 		G_RAID3_DEBUG(1, "Invalid metadata (mediasize %% sectorsize != "
2918 		    "0) on disk %s (device %s), skipping.", pp->name,
2919 		    sc->sc_name);
2920 		return (EINVAL);
2921 	}
2922 	if (md->md_mediasize != sc->sc_mediasize) {
2923 		G_RAID3_DEBUG(1,
2924 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2925 		    "md_mediasize", pp->name, sc->sc_name);
2926 		return (EINVAL);
2927 	}
2928 	if ((md->md_mediasize % (sc->sc_ndisks - 1)) != 0) {
2929 		G_RAID3_DEBUG(1,
2930 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2931 		    "md_mediasize", pp->name, sc->sc_name);
2932 		return (EINVAL);
2933 	}
2934 	if ((sc->sc_mediasize / (sc->sc_ndisks - 1)) > pp->mediasize) {
2935 		G_RAID3_DEBUG(1,
2936 		    "Invalid size of disk %s (device %s), skipping.", pp->name,
2937 		    sc->sc_name);
2938 		return (EINVAL);
2939 	}
2940 	if ((md->md_sectorsize / pp->sectorsize) < sc->sc_ndisks - 1) {
2941 		G_RAID3_DEBUG(1,
2942 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2943 		    "md_sectorsize", pp->name, sc->sc_name);
2944 		return (EINVAL);
2945 	}
2946 	if (md->md_sectorsize != sc->sc_sectorsize) {
2947 		G_RAID3_DEBUG(1,
2948 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2949 		    "md_sectorsize", pp->name, sc->sc_name);
2950 		return (EINVAL);
2951 	}
2952 	if ((sc->sc_sectorsize % pp->sectorsize) != 0) {
2953 		G_RAID3_DEBUG(1,
2954 		    "Invalid sector size of disk %s (device %s), skipping.",
2955 		    pp->name, sc->sc_name);
2956 		return (EINVAL);
2957 	}
2958 	if ((md->md_mflags & ~G_RAID3_DEVICE_FLAG_MASK) != 0) {
2959 		G_RAID3_DEBUG(1,
2960 		    "Invalid device flags on disk %s (device %s), skipping.",
2961 		    pp->name, sc->sc_name);
2962 		return (EINVAL);
2963 	}
2964 	if ((md->md_mflags & G_RAID3_DEVICE_FLAG_VERIFY) != 0 &&
2965 	    (md->md_mflags & G_RAID3_DEVICE_FLAG_ROUND_ROBIN) != 0) {
2966 		/*
2967 		 * VERIFY and ROUND-ROBIN options are mutally exclusive.
2968 		 */
2969 		G_RAID3_DEBUG(1, "Both VERIFY and ROUND-ROBIN flags exist on "
2970 		    "disk %s (device %s), skipping.", pp->name, sc->sc_name);
2971 		return (EINVAL);
2972 	}
2973 	if ((md->md_dflags & ~G_RAID3_DISK_FLAG_MASK) != 0) {
2974 		G_RAID3_DEBUG(1,
2975 		    "Invalid disk flags on disk %s (device %s), skipping.",
2976 		    pp->name, sc->sc_name);
2977 		return (EINVAL);
2978 	}
2979 	return (0);
2980 }
2981 
2982 int
2983 g_raid3_add_disk(struct g_raid3_softc *sc, struct g_provider *pp,
2984     struct g_raid3_metadata *md)
2985 {
2986 	struct g_raid3_disk *disk;
2987 	int error;
2988 
2989 	g_topology_assert_not();
2990 	G_RAID3_DEBUG(2, "Adding disk %s.", pp->name);
2991 
2992 	error = g_raid3_check_metadata(sc, pp, md);
2993 	if (error != 0)
2994 		return (error);
2995 	if (sc->sc_state != G_RAID3_DEVICE_STATE_STARTING &&
2996 	    md->md_genid < sc->sc_genid) {
2997 		G_RAID3_DEBUG(0, "Component %s (device %s) broken, skipping.",
2998 		    pp->name, sc->sc_name);
2999 		return (EINVAL);
3000 	}
3001 	disk = g_raid3_init_disk(sc, pp, md, &error);
3002 	if (disk == NULL)
3003 		return (error);
3004 	error = g_raid3_event_send(disk, G_RAID3_DISK_STATE_NEW,
3005 	    G_RAID3_EVENT_WAIT);
3006 	if (error != 0)
3007 		return (error);
3008 	if (md->md_version < G_RAID3_VERSION) {
3009 		G_RAID3_DEBUG(0, "Upgrading metadata on %s (v%d->v%d).",
3010 		    pp->name, md->md_version, G_RAID3_VERSION);
3011 		g_raid3_update_metadata(disk);
3012 	}
3013 	return (0);
3014 }
3015 
3016 static void
3017 g_raid3_destroy_delayed(void *arg, int flag)
3018 {
3019 	struct g_raid3_softc *sc;
3020 	int error;
3021 
3022 	if (flag == EV_CANCEL) {
3023 		G_RAID3_DEBUG(1, "Destroying canceled.");
3024 		return;
3025 	}
3026 	sc = arg;
3027 	g_topology_unlock();
3028 	sx_xlock(&sc->sc_lock);
3029 	KASSERT((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) == 0,
3030 	    ("DESTROY flag set on %s.", sc->sc_name));
3031 	KASSERT((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROYING) != 0,
3032 	    ("DESTROYING flag not set on %s.", sc->sc_name));
3033 	G_RAID3_DEBUG(0, "Destroying %s (delayed).", sc->sc_name);
3034 	error = g_raid3_destroy(sc, G_RAID3_DESTROY_SOFT);
3035 	if (error != 0) {
3036 		G_RAID3_DEBUG(0, "Cannot destroy %s.", sc->sc_name);
3037 		sx_xunlock(&sc->sc_lock);
3038 	}
3039 	g_topology_lock();
3040 }
3041 
3042 static int
3043 g_raid3_access(struct g_provider *pp, int acr, int acw, int ace)
3044 {
3045 	struct g_raid3_softc *sc;
3046 	int dcr, dcw, dce, error = 0;
3047 
3048 	g_topology_assert();
3049 	G_RAID3_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, acr,
3050 	    acw, ace);
3051 
3052 	sc = pp->geom->softc;
3053 	if (sc == NULL && acr <= 0 && acw <= 0 && ace <= 0)
3054 		return (0);
3055 	KASSERT(sc != NULL, ("NULL softc (provider=%s).", pp->name));
3056 
3057 	dcr = pp->acr + acr;
3058 	dcw = pp->acw + acw;
3059 	dce = pp->ace + ace;
3060 
3061 	g_topology_unlock();
3062 	sx_xlock(&sc->sc_lock);
3063 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) != 0 ||
3064 	    g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) < sc->sc_ndisks - 1) {
3065 		if (acr > 0 || acw > 0 || ace > 0)
3066 			error = ENXIO;
3067 		goto end;
3068 	}
3069 	if (dcw == 0 && !sc->sc_idle)
3070 		g_raid3_idle(sc, dcw);
3071 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROYING) != 0) {
3072 		if (acr > 0 || acw > 0 || ace > 0) {
3073 			error = ENXIO;
3074 			goto end;
3075 		}
3076 		if (dcr == 0 && dcw == 0 && dce == 0) {
3077 			g_post_event(g_raid3_destroy_delayed, sc, M_WAITOK,
3078 			    sc, NULL);
3079 		}
3080 	}
3081 end:
3082 	sx_xunlock(&sc->sc_lock);
3083 	g_topology_lock();
3084 	return (error);
3085 }
3086 
3087 static struct g_geom *
3088 g_raid3_create(struct g_class *mp, const struct g_raid3_metadata *md)
3089 {
3090 	struct g_raid3_softc *sc;
3091 	struct g_geom *gp;
3092 	int error, timeout;
3093 	u_int n;
3094 
3095 	g_topology_assert();
3096 	G_RAID3_DEBUG(1, "Creating device %s (id=%u).", md->md_name, md->md_id);
3097 
3098 	/* One disk is minimum. */
3099 	if (md->md_all < 1)
3100 		return (NULL);
3101 	/*
3102 	 * Action geom.
3103 	 */
3104 	gp = g_new_geomf(mp, "%s", md->md_name);
3105 	sc = malloc(sizeof(*sc), M_RAID3, M_WAITOK | M_ZERO);
3106 	sc->sc_disks = malloc(sizeof(struct g_raid3_disk) * md->md_all, M_RAID3,
3107 	    M_WAITOK | M_ZERO);
3108 	gp->start = g_raid3_start;
3109 	gp->orphan = g_raid3_orphan;
3110 	gp->access = g_raid3_access;
3111 	gp->dumpconf = g_raid3_dumpconf;
3112 
3113 	sc->sc_id = md->md_id;
3114 	sc->sc_mediasize = md->md_mediasize;
3115 	sc->sc_sectorsize = md->md_sectorsize;
3116 	sc->sc_ndisks = md->md_all;
3117 	sc->sc_round_robin = 0;
3118 	sc->sc_flags = md->md_mflags;
3119 	sc->sc_bump_id = 0;
3120 	sc->sc_idle = 1;
3121 	sc->sc_last_write = time_uptime;
3122 	sc->sc_writes = 0;
3123 	for (n = 0; n < sc->sc_ndisks; n++) {
3124 		sc->sc_disks[n].d_softc = sc;
3125 		sc->sc_disks[n].d_no = n;
3126 		sc->sc_disks[n].d_state = G_RAID3_DISK_STATE_NODISK;
3127 	}
3128 	sx_init(&sc->sc_lock, "graid3:lock");
3129 	bioq_init(&sc->sc_queue);
3130 	mtx_init(&sc->sc_queue_mtx, "graid3:queue", NULL, MTX_DEF);
3131 	bioq_init(&sc->sc_regular_delayed);
3132 	bioq_init(&sc->sc_inflight);
3133 	bioq_init(&sc->sc_sync_delayed);
3134 	TAILQ_INIT(&sc->sc_events);
3135 	mtx_init(&sc->sc_events_mtx, "graid3:events", NULL, MTX_DEF);
3136 	callout_init(&sc->sc_callout, CALLOUT_MPSAFE);
3137 	sc->sc_state = G_RAID3_DEVICE_STATE_STARTING;
3138 	gp->softc = sc;
3139 	sc->sc_geom = gp;
3140 	sc->sc_provider = NULL;
3141 	/*
3142 	 * Synchronization geom.
3143 	 */
3144 	gp = g_new_geomf(mp, "%s.sync", md->md_name);
3145 	gp->softc = sc;
3146 	gp->orphan = g_raid3_orphan;
3147 	sc->sc_sync.ds_geom = gp;
3148 
3149 	if (!g_raid3_use_malloc) {
3150 		sc->sc_zones[G_RAID3_ZONE_64K].sz_zone = uma_zcreate("gr3:64k",
3151 		    65536, g_raid3_uma_ctor, g_raid3_uma_dtor, NULL, NULL,
3152 		    UMA_ALIGN_PTR, 0);
3153 		sc->sc_zones[G_RAID3_ZONE_64K].sz_inuse = 0;
3154 		sc->sc_zones[G_RAID3_ZONE_64K].sz_max = g_raid3_n64k;
3155 		sc->sc_zones[G_RAID3_ZONE_64K].sz_requested =
3156 		    sc->sc_zones[G_RAID3_ZONE_64K].sz_failed = 0;
3157 		sc->sc_zones[G_RAID3_ZONE_16K].sz_zone = uma_zcreate("gr3:16k",
3158 		    16384, g_raid3_uma_ctor, g_raid3_uma_dtor, NULL, NULL,
3159 		    UMA_ALIGN_PTR, 0);
3160 		sc->sc_zones[G_RAID3_ZONE_16K].sz_inuse = 0;
3161 		sc->sc_zones[G_RAID3_ZONE_16K].sz_max = g_raid3_n16k;
3162 		sc->sc_zones[G_RAID3_ZONE_16K].sz_requested =
3163 		    sc->sc_zones[G_RAID3_ZONE_16K].sz_failed = 0;
3164 		sc->sc_zones[G_RAID3_ZONE_4K].sz_zone = uma_zcreate("gr3:4k",
3165 		    4096, g_raid3_uma_ctor, g_raid3_uma_dtor, NULL, NULL,
3166 		    UMA_ALIGN_PTR, 0);
3167 		sc->sc_zones[G_RAID3_ZONE_4K].sz_inuse = 0;
3168 		sc->sc_zones[G_RAID3_ZONE_4K].sz_max = g_raid3_n4k;
3169 		sc->sc_zones[G_RAID3_ZONE_4K].sz_requested =
3170 		    sc->sc_zones[G_RAID3_ZONE_4K].sz_failed = 0;
3171 	}
3172 
3173 	error = kproc_create(g_raid3_worker, sc, &sc->sc_worker, 0, 0,
3174 	    "g_raid3 %s", md->md_name);
3175 	if (error != 0) {
3176 		G_RAID3_DEBUG(1, "Cannot create kernel thread for %s.",
3177 		    sc->sc_name);
3178 		if (!g_raid3_use_malloc) {
3179 			uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_64K].sz_zone);
3180 			uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_16K].sz_zone);
3181 			uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_4K].sz_zone);
3182 		}
3183 		g_destroy_geom(sc->sc_sync.ds_geom);
3184 		mtx_destroy(&sc->sc_events_mtx);
3185 		mtx_destroy(&sc->sc_queue_mtx);
3186 		sx_destroy(&sc->sc_lock);
3187 		g_destroy_geom(sc->sc_geom);
3188 		free(sc->sc_disks, M_RAID3);
3189 		free(sc, M_RAID3);
3190 		return (NULL);
3191 	}
3192 
3193 	G_RAID3_DEBUG(1, "Device %s created (%u components, id=%u).",
3194 	    sc->sc_name, sc->sc_ndisks, sc->sc_id);
3195 
3196 	sc->sc_rootmount = root_mount_hold("GRAID3");
3197 	G_RAID3_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
3198 
3199 	/*
3200 	 * Run timeout.
3201 	 */
3202 	timeout = atomic_load_acq_int(&g_raid3_timeout);
3203 	callout_reset(&sc->sc_callout, timeout * hz, g_raid3_go, sc);
3204 	return (sc->sc_geom);
3205 }
3206 
3207 int
3208 g_raid3_destroy(struct g_raid3_softc *sc, int how)
3209 {
3210 	struct g_provider *pp;
3211 
3212 	g_topology_assert_not();
3213 	if (sc == NULL)
3214 		return (ENXIO);
3215 	sx_assert(&sc->sc_lock, SX_XLOCKED);
3216 
3217 	pp = sc->sc_provider;
3218 	if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
3219 		switch (how) {
3220 		case G_RAID3_DESTROY_SOFT:
3221 			G_RAID3_DEBUG(1,
3222 			    "Device %s is still open (r%dw%de%d).", pp->name,
3223 			    pp->acr, pp->acw, pp->ace);
3224 			return (EBUSY);
3225 		case G_RAID3_DESTROY_DELAYED:
3226 			G_RAID3_DEBUG(1,
3227 			    "Device %s will be destroyed on last close.",
3228 			    pp->name);
3229 			if (sc->sc_syncdisk != NULL)
3230 				g_raid3_sync_stop(sc, 1);
3231 			sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROYING;
3232 			return (EBUSY);
3233 		case G_RAID3_DESTROY_HARD:
3234 			G_RAID3_DEBUG(1, "Device %s is still open, so it "
3235 			    "can't be definitely removed.", pp->name);
3236 			break;
3237 		}
3238 	}
3239 
3240 	g_topology_lock();
3241 	if (sc->sc_geom->softc == NULL) {
3242 		g_topology_unlock();
3243 		return (0);
3244 	}
3245 	sc->sc_geom->softc = NULL;
3246 	sc->sc_sync.ds_geom->softc = NULL;
3247 	g_topology_unlock();
3248 
3249 	sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
3250 	sc->sc_flags |= G_RAID3_DEVICE_FLAG_WAIT;
3251 	G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, sc);
3252 	sx_xunlock(&sc->sc_lock);
3253 	mtx_lock(&sc->sc_queue_mtx);
3254 	wakeup(sc);
3255 	wakeup(&sc->sc_queue);
3256 	mtx_unlock(&sc->sc_queue_mtx);
3257 	G_RAID3_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
3258 	while (sc->sc_worker != NULL)
3259 		tsleep(&sc->sc_worker, PRIBIO, "r3:destroy", hz / 5);
3260 	G_RAID3_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
3261 	sx_xlock(&sc->sc_lock);
3262 	g_raid3_destroy_device(sc);
3263 	free(sc->sc_disks, M_RAID3);
3264 	free(sc, M_RAID3);
3265 	return (0);
3266 }
3267 
3268 static void
3269 g_raid3_taste_orphan(struct g_consumer *cp)
3270 {
3271 
3272 	KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
3273 	    cp->provider->name));
3274 }
3275 
3276 static struct g_geom *
3277 g_raid3_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
3278 {
3279 	struct g_raid3_metadata md;
3280 	struct g_raid3_softc *sc;
3281 	struct g_consumer *cp;
3282 	struct g_geom *gp;
3283 	int error;
3284 
3285 	g_topology_assert();
3286 	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
3287 	G_RAID3_DEBUG(2, "Tasting %s.", pp->name);
3288 
3289 	gp = g_new_geomf(mp, "raid3:taste");
3290 	/* This orphan function should be never called. */
3291 	gp->orphan = g_raid3_taste_orphan;
3292 	cp = g_new_consumer(gp);
3293 	g_attach(cp, pp);
3294 	error = g_raid3_read_metadata(cp, &md);
3295 	g_detach(cp);
3296 	g_destroy_consumer(cp);
3297 	g_destroy_geom(gp);
3298 	if (error != 0)
3299 		return (NULL);
3300 	gp = NULL;
3301 
3302 	if (md.md_provider[0] != '\0' && strcmp(md.md_provider, pp->name) != 0)
3303 		return (NULL);
3304 	if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
3305 		return (NULL);
3306 	if (g_raid3_debug >= 2)
3307 		raid3_metadata_dump(&md);
3308 
3309 	/*
3310 	 * Let's check if device already exists.
3311 	 */
3312 	sc = NULL;
3313 	LIST_FOREACH(gp, &mp->geom, geom) {
3314 		sc = gp->softc;
3315 		if (sc == NULL)
3316 			continue;
3317 		if (sc->sc_sync.ds_geom == gp)
3318 			continue;
3319 		if (strcmp(md.md_name, sc->sc_name) != 0)
3320 			continue;
3321 		if (md.md_id != sc->sc_id) {
3322 			G_RAID3_DEBUG(0, "Device %s already configured.",
3323 			    sc->sc_name);
3324 			return (NULL);
3325 		}
3326 		break;
3327 	}
3328 	if (gp == NULL) {
3329 		gp = g_raid3_create(mp, &md);
3330 		if (gp == NULL) {
3331 			G_RAID3_DEBUG(0, "Cannot create device %s.",
3332 			    md.md_name);
3333 			return (NULL);
3334 		}
3335 		sc = gp->softc;
3336 	}
3337 	G_RAID3_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
3338 	g_topology_unlock();
3339 	sx_xlock(&sc->sc_lock);
3340 	error = g_raid3_add_disk(sc, pp, &md);
3341 	if (error != 0) {
3342 		G_RAID3_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
3343 		    pp->name, gp->name, error);
3344 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_NODISK) ==
3345 		    sc->sc_ndisks) {
3346 			g_cancel_event(sc);
3347 			g_raid3_destroy(sc, G_RAID3_DESTROY_HARD);
3348 			g_topology_lock();
3349 			return (NULL);
3350 		}
3351 		gp = NULL;
3352 	}
3353 	sx_xunlock(&sc->sc_lock);
3354 	g_topology_lock();
3355 	return (gp);
3356 }
3357 
3358 static int
3359 g_raid3_destroy_geom(struct gctl_req *req __unused, struct g_class *mp __unused,
3360     struct g_geom *gp)
3361 {
3362 	struct g_raid3_softc *sc;
3363 	int error;
3364 
3365 	g_topology_unlock();
3366 	sc = gp->softc;
3367 	sx_xlock(&sc->sc_lock);
3368 	g_cancel_event(sc);
3369 	error = g_raid3_destroy(gp->softc, G_RAID3_DESTROY_SOFT);
3370 	if (error != 0)
3371 		sx_xunlock(&sc->sc_lock);
3372 	g_topology_lock();
3373 	return (error);
3374 }
3375 
3376 static void
3377 g_raid3_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
3378     struct g_consumer *cp, struct g_provider *pp)
3379 {
3380 	struct g_raid3_softc *sc;
3381 
3382 	g_topology_assert();
3383 
3384 	sc = gp->softc;
3385 	if (sc == NULL)
3386 		return;
3387 	/* Skip synchronization geom. */
3388 	if (gp == sc->sc_sync.ds_geom)
3389 		return;
3390 	if (pp != NULL) {
3391 		/* Nothing here. */
3392 	} else if (cp != NULL) {
3393 		struct g_raid3_disk *disk;
3394 
3395 		disk = cp->private;
3396 		if (disk == NULL)
3397 			return;
3398 		g_topology_unlock();
3399 		sx_xlock(&sc->sc_lock);
3400 		sbuf_printf(sb, "%s<Type>", indent);
3401 		if (disk->d_no == sc->sc_ndisks - 1)
3402 			sbuf_printf(sb, "PARITY");
3403 		else
3404 			sbuf_printf(sb, "DATA");
3405 		sbuf_printf(sb, "</Type>\n");
3406 		sbuf_printf(sb, "%s<Number>%u</Number>\n", indent,
3407 		    (u_int)disk->d_no);
3408 		if (disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
3409 			sbuf_printf(sb, "%s<Synchronized>", indent);
3410 			if (disk->d_sync.ds_offset == 0)
3411 				sbuf_printf(sb, "0%%");
3412 			else {
3413 				sbuf_printf(sb, "%u%%",
3414 				    (u_int)((disk->d_sync.ds_offset * 100) /
3415 				    (sc->sc_mediasize / (sc->sc_ndisks - 1))));
3416 			}
3417 			sbuf_printf(sb, "</Synchronized>\n");
3418 		}
3419 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent,
3420 		    disk->d_sync.ds_syncid);
3421 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, disk->d_genid);
3422 		sbuf_printf(sb, "%s<Flags>", indent);
3423 		if (disk->d_flags == 0)
3424 			sbuf_printf(sb, "NONE");
3425 		else {
3426 			int first = 1;
3427 
3428 #define	ADD_FLAG(flag, name)	do {					\
3429 	if ((disk->d_flags & (flag)) != 0) {				\
3430 		if (!first)						\
3431 			sbuf_printf(sb, ", ");				\
3432 		else							\
3433 			first = 0;					\
3434 		sbuf_printf(sb, name);					\
3435 	}								\
3436 } while (0)
3437 			ADD_FLAG(G_RAID3_DISK_FLAG_DIRTY, "DIRTY");
3438 			ADD_FLAG(G_RAID3_DISK_FLAG_HARDCODED, "HARDCODED");
3439 			ADD_FLAG(G_RAID3_DISK_FLAG_SYNCHRONIZING,
3440 			    "SYNCHRONIZING");
3441 			ADD_FLAG(G_RAID3_DISK_FLAG_FORCE_SYNC, "FORCE_SYNC");
3442 			ADD_FLAG(G_RAID3_DISK_FLAG_BROKEN, "BROKEN");
3443 #undef	ADD_FLAG
3444 		}
3445 		sbuf_printf(sb, "</Flags>\n");
3446 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3447 		    g_raid3_disk_state2str(disk->d_state));
3448 		sx_xunlock(&sc->sc_lock);
3449 		g_topology_lock();
3450 	} else {
3451 		g_topology_unlock();
3452 		sx_xlock(&sc->sc_lock);
3453 		if (!g_raid3_use_malloc) {
3454 			sbuf_printf(sb,
3455 			    "%s<Zone4kRequested>%u</Zone4kRequested>\n", indent,
3456 			    sc->sc_zones[G_RAID3_ZONE_4K].sz_requested);
3457 			sbuf_printf(sb,
3458 			    "%s<Zone4kFailed>%u</Zone4kFailed>\n", indent,
3459 			    sc->sc_zones[G_RAID3_ZONE_4K].sz_failed);
3460 			sbuf_printf(sb,
3461 			    "%s<Zone16kRequested>%u</Zone16kRequested>\n", indent,
3462 			    sc->sc_zones[G_RAID3_ZONE_16K].sz_requested);
3463 			sbuf_printf(sb,
3464 			    "%s<Zone16kFailed>%u</Zone16kFailed>\n", indent,
3465 			    sc->sc_zones[G_RAID3_ZONE_16K].sz_failed);
3466 			sbuf_printf(sb,
3467 			    "%s<Zone64kRequested>%u</Zone64kRequested>\n", indent,
3468 			    sc->sc_zones[G_RAID3_ZONE_64K].sz_requested);
3469 			sbuf_printf(sb,
3470 			    "%s<Zone64kFailed>%u</Zone64kFailed>\n", indent,
3471 			    sc->sc_zones[G_RAID3_ZONE_64K].sz_failed);
3472 		}
3473 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
3474 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent, sc->sc_syncid);
3475 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, sc->sc_genid);
3476 		sbuf_printf(sb, "%s<Flags>", indent);
3477 		if (sc->sc_flags == 0)
3478 			sbuf_printf(sb, "NONE");
3479 		else {
3480 			int first = 1;
3481 
3482 #define	ADD_FLAG(flag, name)	do {					\
3483 	if ((sc->sc_flags & (flag)) != 0) {				\
3484 		if (!first)						\
3485 			sbuf_printf(sb, ", ");				\
3486 		else							\
3487 			first = 0;					\
3488 		sbuf_printf(sb, name);					\
3489 	}								\
3490 } while (0)
3491 			ADD_FLAG(G_RAID3_DEVICE_FLAG_NOFAILSYNC, "NOFAILSYNC");
3492 			ADD_FLAG(G_RAID3_DEVICE_FLAG_NOAUTOSYNC, "NOAUTOSYNC");
3493 			ADD_FLAG(G_RAID3_DEVICE_FLAG_ROUND_ROBIN,
3494 			    "ROUND-ROBIN");
3495 			ADD_FLAG(G_RAID3_DEVICE_FLAG_VERIFY, "VERIFY");
3496 #undef	ADD_FLAG
3497 		}
3498 		sbuf_printf(sb, "</Flags>\n");
3499 		sbuf_printf(sb, "%s<Components>%u</Components>\n", indent,
3500 		    sc->sc_ndisks);
3501 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3502 		    g_raid3_device_state2str(sc->sc_state));
3503 		sx_xunlock(&sc->sc_lock);
3504 		g_topology_lock();
3505 	}
3506 }
3507 
3508 static void
3509 g_raid3_shutdown_pre_sync(void *arg, int howto)
3510 {
3511 	struct g_class *mp;
3512 	struct g_geom *gp, *gp2;
3513 	struct g_raid3_softc *sc;
3514 	int error;
3515 
3516 	mp = arg;
3517 	DROP_GIANT();
3518 	g_topology_lock();
3519 	LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
3520 		if ((sc = gp->softc) == NULL)
3521 			continue;
3522 		/* Skip synchronization geom. */
3523 		if (gp == sc->sc_sync.ds_geom)
3524 			continue;
3525 		g_topology_unlock();
3526 		sx_xlock(&sc->sc_lock);
3527 		g_cancel_event(sc);
3528 		error = g_raid3_destroy(sc, G_RAID3_DESTROY_DELAYED);
3529 		if (error != 0)
3530 			sx_xunlock(&sc->sc_lock);
3531 		g_topology_lock();
3532 	}
3533 	g_topology_unlock();
3534 	PICKUP_GIANT();
3535 }
3536 
3537 static void
3538 g_raid3_init(struct g_class *mp)
3539 {
3540 
3541 	g_raid3_pre_sync = EVENTHANDLER_REGISTER(shutdown_pre_sync,
3542 	    g_raid3_shutdown_pre_sync, mp, SHUTDOWN_PRI_FIRST);
3543 	if (g_raid3_pre_sync == NULL)
3544 		G_RAID3_DEBUG(0, "Warning! Cannot register shutdown event.");
3545 }
3546 
3547 static void
3548 g_raid3_fini(struct g_class *mp)
3549 {
3550 
3551 	if (g_raid3_pre_sync != NULL)
3552 		EVENTHANDLER_DEREGISTER(shutdown_pre_sync, g_raid3_pre_sync);
3553 }
3554 
3555 DECLARE_GEOM_CLASS(g_raid3_class, g_raid3);
3556