xref: /freebsd/sys/geom/mirror/g_mirror.c (revision d056fa046c6a91b90cd98165face0e42a33a5173)
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/mirror/g_mirror.h>
47 
48 
49 static MALLOC_DEFINE(M_MIRROR, "mirror_data", "GEOM_MIRROR Data");
50 
51 SYSCTL_DECL(_kern_geom);
52 SYSCTL_NODE(_kern_geom, OID_AUTO, mirror, CTLFLAG_RW, 0, "GEOM_MIRROR stuff");
53 u_int g_mirror_debug = 0;
54 TUNABLE_INT("kern.geom.mirror.debug", &g_mirror_debug);
55 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, debug, CTLFLAG_RW, &g_mirror_debug, 0,
56     "Debug level");
57 static u_int g_mirror_timeout = 4;
58 TUNABLE_INT("kern.geom.mirror.timeout", &g_mirror_timeout);
59 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, timeout, CTLFLAG_RW, &g_mirror_timeout,
60     0, "Time to wait on all mirror components");
61 static u_int g_mirror_idletime = 5;
62 TUNABLE_INT("kern.geom.mirror.idletime", &g_mirror_idletime);
63 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, idletime, CTLFLAG_RW,
64     &g_mirror_idletime, 0, "Mark components as clean when idling");
65 static u_int g_mirror_disconnect_on_failure = 1;
66 TUNABLE_INT("kern.geom.mirror.disconnect_on_failure",
67     &g_mirror_disconnect_on_failure);
68 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, disconnect_on_failure, CTLFLAG_RW,
69     &g_mirror_disconnect_on_failure, 0, "Disconnect component on I/O failure.");
70 static u_int g_mirror_syncreqs = 2;
71 TUNABLE_INT("kern.geom.mirror.sync_requests", &g_mirror_syncreqs);
72 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, sync_requests, CTLFLAG_RDTUN,
73     &g_mirror_syncreqs, 0, "Parallel synchronization I/O requests.");
74 
75 #define	MSLEEP(ident, mtx, priority, wmesg, timeout)	do {		\
76 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, (ident));	\
77 	msleep((ident), (mtx), (priority), (wmesg), (timeout));		\
78 	G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, (ident));	\
79 } while (0)
80 
81 static eventhandler_tag g_mirror_pre_sync = NULL;
82 
83 static int g_mirror_destroy_geom(struct gctl_req *req, struct g_class *mp,
84     struct g_geom *gp);
85 static g_taste_t g_mirror_taste;
86 static void g_mirror_init(struct g_class *mp);
87 static void g_mirror_fini(struct g_class *mp);
88 
89 struct g_class g_mirror_class = {
90 	.name = G_MIRROR_CLASS_NAME,
91 	.version = G_VERSION,
92 	.ctlreq = g_mirror_config,
93 	.taste = g_mirror_taste,
94 	.destroy_geom = g_mirror_destroy_geom,
95 	.init = g_mirror_init,
96 	.fini = g_mirror_fini
97 };
98 
99 
100 static void g_mirror_destroy_provider(struct g_mirror_softc *sc);
101 static int g_mirror_update_disk(struct g_mirror_disk *disk, u_int state);
102 static void g_mirror_update_device(struct g_mirror_softc *sc, boolean_t force);
103 static void g_mirror_dumpconf(struct sbuf *sb, const char *indent,
104     struct g_geom *gp, struct g_consumer *cp, struct g_provider *pp);
105 static void g_mirror_sync_stop(struct g_mirror_disk *disk, int type);
106 static void g_mirror_register_request(struct bio *bp);
107 static void g_mirror_sync_release(struct g_mirror_softc *sc);
108 
109 
110 static const char *
111 g_mirror_disk_state2str(int state)
112 {
113 
114 	switch (state) {
115 	case G_MIRROR_DISK_STATE_NONE:
116 		return ("NONE");
117 	case G_MIRROR_DISK_STATE_NEW:
118 		return ("NEW");
119 	case G_MIRROR_DISK_STATE_ACTIVE:
120 		return ("ACTIVE");
121 	case G_MIRROR_DISK_STATE_STALE:
122 		return ("STALE");
123 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
124 		return ("SYNCHRONIZING");
125 	case G_MIRROR_DISK_STATE_DISCONNECTED:
126 		return ("DISCONNECTED");
127 	case G_MIRROR_DISK_STATE_DESTROY:
128 		return ("DESTROY");
129 	default:
130 		return ("INVALID");
131 	}
132 }
133 
134 static const char *
135 g_mirror_device_state2str(int state)
136 {
137 
138 	switch (state) {
139 	case G_MIRROR_DEVICE_STATE_STARTING:
140 		return ("STARTING");
141 	case G_MIRROR_DEVICE_STATE_RUNNING:
142 		return ("RUNNING");
143 	default:
144 		return ("INVALID");
145 	}
146 }
147 
148 static const char *
149 g_mirror_get_diskname(struct g_mirror_disk *disk)
150 {
151 
152 	if (disk->d_consumer == NULL || disk->d_consumer->provider == NULL)
153 		return ("[unknown]");
154 	return (disk->d_name);
155 }
156 
157 /*
158  * --- Events handling functions ---
159  * Events in geom_mirror are used to maintain disks and device status
160  * from one thread to simplify locking.
161  */
162 static void
163 g_mirror_event_free(struct g_mirror_event *ep)
164 {
165 
166 	free(ep, M_MIRROR);
167 }
168 
169 int
170 g_mirror_event_send(void *arg, int state, int flags)
171 {
172 	struct g_mirror_softc *sc;
173 	struct g_mirror_disk *disk;
174 	struct g_mirror_event *ep;
175 	int error;
176 
177 	ep = malloc(sizeof(*ep), M_MIRROR, M_WAITOK);
178 	G_MIRROR_DEBUG(4, "%s: Sending event %p.", __func__, ep);
179 	if ((flags & G_MIRROR_EVENT_DEVICE) != 0) {
180 		disk = NULL;
181 		sc = arg;
182 	} else {
183 		disk = arg;
184 		sc = disk->d_softc;
185 	}
186 	ep->e_disk = disk;
187 	ep->e_state = state;
188 	ep->e_flags = flags;
189 	ep->e_error = 0;
190 	mtx_lock(&sc->sc_events_mtx);
191 	TAILQ_INSERT_TAIL(&sc->sc_events, ep, e_next);
192 	mtx_unlock(&sc->sc_events_mtx);
193 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
194 	mtx_lock(&sc->sc_queue_mtx);
195 	wakeup(sc);
196 	mtx_unlock(&sc->sc_queue_mtx);
197 	if ((flags & G_MIRROR_EVENT_DONTWAIT) != 0)
198 		return (0);
199 	sx_assert(&sc->sc_lock, SX_XLOCKED);
200 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, ep);
201 	sx_xunlock(&sc->sc_lock);
202 	while ((ep->e_flags & G_MIRROR_EVENT_DONE) == 0) {
203 		mtx_lock(&sc->sc_events_mtx);
204 		MSLEEP(ep, &sc->sc_events_mtx, PRIBIO | PDROP, "m:event",
205 		    hz * 5);
206 	}
207 	error = ep->e_error;
208 	g_mirror_event_free(ep);
209 	sx_xlock(&sc->sc_lock);
210 	return (error);
211 }
212 
213 static struct g_mirror_event *
214 g_mirror_event_get(struct g_mirror_softc *sc)
215 {
216 	struct g_mirror_event *ep;
217 
218 	mtx_lock(&sc->sc_events_mtx);
219 	ep = TAILQ_FIRST(&sc->sc_events);
220 	mtx_unlock(&sc->sc_events_mtx);
221 	return (ep);
222 }
223 
224 static void
225 g_mirror_event_remove(struct g_mirror_softc *sc, struct g_mirror_event *ep)
226 {
227 
228 	mtx_lock(&sc->sc_events_mtx);
229 	TAILQ_REMOVE(&sc->sc_events, ep, e_next);
230 	mtx_unlock(&sc->sc_events_mtx);
231 }
232 
233 static void
234 g_mirror_event_cancel(struct g_mirror_disk *disk)
235 {
236 	struct g_mirror_softc *sc;
237 	struct g_mirror_event *ep, *tmpep;
238 
239 	sc = disk->d_softc;
240 	sx_assert(&sc->sc_lock, SX_XLOCKED);
241 
242 	mtx_lock(&sc->sc_events_mtx);
243 	TAILQ_FOREACH_SAFE(ep, &sc->sc_events, e_next, tmpep) {
244 		if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0)
245 			continue;
246 		if (ep->e_disk != disk)
247 			continue;
248 		TAILQ_REMOVE(&sc->sc_events, ep, e_next);
249 		if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
250 			g_mirror_event_free(ep);
251 		else {
252 			ep->e_error = ECANCELED;
253 			wakeup(ep);
254 		}
255 	}
256 	mtx_unlock(&sc->sc_events_mtx);
257 }
258 
259 /*
260  * Return the number of disks in given state.
261  * If state is equal to -1, count all connected disks.
262  */
263 u_int
264 g_mirror_ndisks(struct g_mirror_softc *sc, int state)
265 {
266 	struct g_mirror_disk *disk;
267 	u_int n = 0;
268 
269 	sx_assert(&sc->sc_lock, SX_LOCKED);
270 
271 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
272 		if (state == -1 || disk->d_state == state)
273 			n++;
274 	}
275 	return (n);
276 }
277 
278 /*
279  * Find a disk in mirror by its disk ID.
280  */
281 static struct g_mirror_disk *
282 g_mirror_id2disk(struct g_mirror_softc *sc, uint32_t id)
283 {
284 	struct g_mirror_disk *disk;
285 
286 	sx_assert(&sc->sc_lock, SX_XLOCKED);
287 
288 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
289 		if (disk->d_id == id)
290 			return (disk);
291 	}
292 	return (NULL);
293 }
294 
295 static u_int
296 g_mirror_nrequests(struct g_mirror_softc *sc, struct g_consumer *cp)
297 {
298 	struct bio *bp;
299 	u_int nreqs = 0;
300 
301 	mtx_lock(&sc->sc_queue_mtx);
302 	TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
303 		if (bp->bio_from == cp)
304 			nreqs++;
305 	}
306 	mtx_unlock(&sc->sc_queue_mtx);
307 	return (nreqs);
308 }
309 
310 static int
311 g_mirror_is_busy(struct g_mirror_softc *sc, struct g_consumer *cp)
312 {
313 
314 	if (cp->index > 0) {
315 		G_MIRROR_DEBUG(2,
316 		    "I/O requests for %s exist, can't destroy it now.",
317 		    cp->provider->name);
318 		return (1);
319 	}
320 	if (g_mirror_nrequests(sc, cp) > 0) {
321 		G_MIRROR_DEBUG(2,
322 		    "I/O requests for %s in queue, can't destroy it now.",
323 		    cp->provider->name);
324 		return (1);
325 	}
326 	return (0);
327 }
328 
329 static void
330 g_mirror_destroy_consumer(void *arg, int flags __unused)
331 {
332 	struct g_consumer *cp;
333 
334 	g_topology_assert();
335 
336 	cp = arg;
337 	G_MIRROR_DEBUG(1, "Consumer %s destroyed.", cp->provider->name);
338 	g_detach(cp);
339 	g_destroy_consumer(cp);
340 }
341 
342 static void
343 g_mirror_kill_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
344 {
345 	struct g_provider *pp;
346 	int retaste_wait;
347 
348 	g_topology_assert();
349 
350 	cp->private = NULL;
351 	if (g_mirror_is_busy(sc, cp))
352 		return;
353 	pp = cp->provider;
354 	retaste_wait = 0;
355 	if (cp->acw == 1) {
356 		if ((pp->geom->flags & G_GEOM_WITHER) == 0)
357 			retaste_wait = 1;
358 	}
359 	G_MIRROR_DEBUG(2, "Access %s r%dw%de%d = %d", pp->name, -cp->acr,
360 	    -cp->acw, -cp->ace, 0);
361 	if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
362 		g_access(cp, -cp->acr, -cp->acw, -cp->ace);
363 	if (retaste_wait) {
364 		/*
365 		 * After retaste event was send (inside g_access()), we can send
366 		 * event to detach and destroy consumer.
367 		 * A class, which has consumer to the given provider connected
368 		 * will not receive retaste event for the provider.
369 		 * This is the way how I ignore retaste events when I close
370 		 * consumers opened for write: I detach and destroy consumer
371 		 * after retaste event is sent.
372 		 */
373 		g_post_event(g_mirror_destroy_consumer, cp, M_WAITOK, NULL);
374 		return;
375 	}
376 	G_MIRROR_DEBUG(1, "Consumer %s destroyed.", pp->name);
377 	g_detach(cp);
378 	g_destroy_consumer(cp);
379 }
380 
381 static int
382 g_mirror_connect_disk(struct g_mirror_disk *disk, struct g_provider *pp)
383 {
384 	struct g_consumer *cp;
385 	int error;
386 
387 	g_topology_assert_not();
388 	KASSERT(disk->d_consumer == NULL,
389 	    ("Disk already connected (device %s).", disk->d_softc->sc_name));
390 
391 	g_topology_lock();
392 	cp = g_new_consumer(disk->d_softc->sc_geom);
393 	error = g_attach(cp, pp);
394 	if (error != 0) {
395 		g_destroy_consumer(cp);
396 		g_topology_unlock();
397 		return (error);
398 	}
399 	error = g_access(cp, 1, 1, 1);
400 	if (error != 0) {
401 		g_detach(cp);
402 		g_destroy_consumer(cp);
403 		g_topology_unlock();
404 		G_MIRROR_DEBUG(0, "Cannot open consumer %s (error=%d).",
405 		    pp->name, error);
406 		return (error);
407 	}
408 	g_topology_unlock();
409 	disk->d_consumer = cp;
410 	disk->d_consumer->private = disk;
411 	disk->d_consumer->index = 0;
412 
413 	G_MIRROR_DEBUG(2, "Disk %s connected.", g_mirror_get_diskname(disk));
414 	return (0);
415 }
416 
417 static void
418 g_mirror_disconnect_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
419 {
420 
421 	g_topology_assert();
422 
423 	if (cp == NULL)
424 		return;
425 	if (cp->provider != NULL)
426 		g_mirror_kill_consumer(sc, cp);
427 	else
428 		g_destroy_consumer(cp);
429 }
430 
431 /*
432  * Initialize disk. This means allocate memory, create consumer, attach it
433  * to the provider and open access (r1w1e1) to it.
434  */
435 static struct g_mirror_disk *
436 g_mirror_init_disk(struct g_mirror_softc *sc, struct g_provider *pp,
437     struct g_mirror_metadata *md, int *errorp)
438 {
439 	struct g_mirror_disk *disk;
440 	int error;
441 
442 	disk = malloc(sizeof(*disk), M_MIRROR, M_NOWAIT | M_ZERO);
443 	if (disk == NULL) {
444 		error = ENOMEM;
445 		goto fail;
446 	}
447 	disk->d_softc = sc;
448 	error = g_mirror_connect_disk(disk, pp);
449 	if (error != 0)
450 		goto fail;
451 	disk->d_id = md->md_did;
452 	disk->d_state = G_MIRROR_DISK_STATE_NONE;
453 	disk->d_priority = md->md_priority;
454 	disk->d_delay.sec = 0;
455 	disk->d_delay.frac = 0;
456 	binuptime(&disk->d_last_used);
457 	disk->d_flags = md->md_dflags;
458 	if (md->md_provider[0] != '\0')
459 		disk->d_flags |= G_MIRROR_DISK_FLAG_HARDCODED;
460 	disk->d_sync.ds_consumer = NULL;
461 	disk->d_sync.ds_offset = md->md_sync_offset;
462 	disk->d_sync.ds_offset_done = md->md_sync_offset;
463 	disk->d_genid = md->md_genid;
464 	disk->d_sync.ds_syncid = md->md_syncid;
465 	if (errorp != NULL)
466 		*errorp = 0;
467 	return (disk);
468 fail:
469 	if (errorp != NULL)
470 		*errorp = error;
471 	if (disk != NULL)
472 		free(disk, M_MIRROR);
473 	return (NULL);
474 }
475 
476 static void
477 g_mirror_destroy_disk(struct g_mirror_disk *disk)
478 {
479 	struct g_mirror_softc *sc;
480 
481 	g_topology_assert_not();
482 	sc = disk->d_softc;
483 	sx_assert(&sc->sc_lock, SX_XLOCKED);
484 
485 	LIST_REMOVE(disk, d_next);
486 	g_mirror_event_cancel(disk);
487 	if (sc->sc_hint == disk)
488 		sc->sc_hint = NULL;
489 	switch (disk->d_state) {
490 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
491 		g_mirror_sync_stop(disk, 1);
492 		/* FALLTHROUGH */
493 	case G_MIRROR_DISK_STATE_NEW:
494 	case G_MIRROR_DISK_STATE_STALE:
495 	case G_MIRROR_DISK_STATE_ACTIVE:
496 		g_topology_lock();
497 		g_mirror_disconnect_consumer(sc, disk->d_consumer);
498 		g_topology_unlock();
499 		free(disk, M_MIRROR);
500 		break;
501 	default:
502 		KASSERT(0 == 1, ("Wrong disk state (%s, %s).",
503 		    g_mirror_get_diskname(disk),
504 		    g_mirror_disk_state2str(disk->d_state)));
505 	}
506 }
507 
508 static void
509 g_mirror_destroy_device(struct g_mirror_softc *sc)
510 {
511 	struct g_mirror_disk *disk;
512 	struct g_mirror_event *ep;
513 	struct g_geom *gp;
514 	struct g_consumer *cp, *tmpcp;
515 
516 	g_topology_assert_not();
517 	sx_assert(&sc->sc_lock, SX_XLOCKED);
518 
519 	gp = sc->sc_geom;
520 	if (sc->sc_provider != NULL)
521 		g_mirror_destroy_provider(sc);
522 	for (disk = LIST_FIRST(&sc->sc_disks); disk != NULL;
523 	    disk = LIST_FIRST(&sc->sc_disks)) {
524 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
525 		g_mirror_update_metadata(disk);
526 		g_mirror_destroy_disk(disk);
527 	}
528 	while ((ep = g_mirror_event_get(sc)) != NULL) {
529 		g_mirror_event_remove(sc, ep);
530 		if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
531 			g_mirror_event_free(ep);
532 		else {
533 			ep->e_error = ECANCELED;
534 			ep->e_flags |= G_MIRROR_EVENT_DONE;
535 			G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, ep);
536 			mtx_lock(&sc->sc_events_mtx);
537 			wakeup(ep);
538 			mtx_unlock(&sc->sc_events_mtx);
539 		}
540 	}
541 	callout_drain(&sc->sc_callout);
542 
543 	g_topology_lock();
544 	LIST_FOREACH_SAFE(cp, &sc->sc_sync.ds_geom->consumer, consumer, tmpcp) {
545 		g_mirror_disconnect_consumer(sc, cp);
546 	}
547 	g_wither_geom(sc->sc_sync.ds_geom, ENXIO);
548 	G_MIRROR_DEBUG(0, "Device %s destroyed.", gp->name);
549 	g_wither_geom(gp, ENXIO);
550 	g_topology_unlock();
551 	mtx_destroy(&sc->sc_queue_mtx);
552 	mtx_destroy(&sc->sc_events_mtx);
553 	sx_xunlock(&sc->sc_lock);
554 	sx_destroy(&sc->sc_lock);
555 }
556 
557 static void
558 g_mirror_orphan(struct g_consumer *cp)
559 {
560 	struct g_mirror_disk *disk;
561 
562 	g_topology_assert();
563 
564 	disk = cp->private;
565 	if (disk == NULL)
566 		return;
567 	disk->d_softc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
568 	g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
569 	    G_MIRROR_EVENT_DONTWAIT);
570 }
571 
572 /*
573  * Function should return the next active disk on the list.
574  * It is possible that it will be the same disk as given.
575  * If there are no active disks on list, NULL is returned.
576  */
577 static __inline struct g_mirror_disk *
578 g_mirror_find_next(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
579 {
580 	struct g_mirror_disk *dp;
581 
582 	for (dp = LIST_NEXT(disk, d_next); dp != disk;
583 	    dp = LIST_NEXT(dp, d_next)) {
584 		if (dp == NULL)
585 			dp = LIST_FIRST(&sc->sc_disks);
586 		if (dp->d_state == G_MIRROR_DISK_STATE_ACTIVE)
587 			break;
588 	}
589 	if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
590 		return (NULL);
591 	return (dp);
592 }
593 
594 static struct g_mirror_disk *
595 g_mirror_get_disk(struct g_mirror_softc *sc)
596 {
597 	struct g_mirror_disk *disk;
598 
599 	if (sc->sc_hint == NULL) {
600 		sc->sc_hint = LIST_FIRST(&sc->sc_disks);
601 		if (sc->sc_hint == NULL)
602 			return (NULL);
603 	}
604 	disk = sc->sc_hint;
605 	if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE) {
606 		disk = g_mirror_find_next(sc, disk);
607 		if (disk == NULL)
608 			return (NULL);
609 	}
610 	sc->sc_hint = g_mirror_find_next(sc, disk);
611 	return (disk);
612 }
613 
614 static int
615 g_mirror_write_metadata(struct g_mirror_disk *disk,
616     struct g_mirror_metadata *md)
617 {
618 	struct g_mirror_softc *sc;
619 	struct g_consumer *cp;
620 	off_t offset, length;
621 	u_char *sector;
622 	int error = 0;
623 
624 	g_topology_assert_not();
625 	sc = disk->d_softc;
626 	sx_assert(&sc->sc_lock, SX_LOCKED);
627 
628 	cp = disk->d_consumer;
629 	KASSERT(cp != NULL, ("NULL consumer (%s).", sc->sc_name));
630 	KASSERT(cp->provider != NULL, ("NULL provider (%s).", sc->sc_name));
631 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
632 	    ("Consumer %s closed? (r%dw%de%d).", cp->provider->name, cp->acr,
633 	    cp->acw, cp->ace));
634 	length = cp->provider->sectorsize;
635 	offset = cp->provider->mediasize - length;
636 	sector = malloc((size_t)length, M_MIRROR, M_WAITOK | M_ZERO);
637 	if (md != NULL)
638 		mirror_metadata_encode(md, sector);
639 	error = g_write_data(cp, offset, sector, length);
640 	free(sector, M_MIRROR);
641 	if (error != 0) {
642 		if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
643 			disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
644 			G_MIRROR_DEBUG(0, "Cannot write metadata on %s "
645 			    "(device=%s, error=%d).",
646 			    g_mirror_get_diskname(disk), sc->sc_name, error);
647 		} else {
648 			G_MIRROR_DEBUG(1, "Cannot write metadata on %s "
649 			    "(device=%s, error=%d).",
650 			    g_mirror_get_diskname(disk), sc->sc_name, error);
651 		}
652 		if (g_mirror_disconnect_on_failure &&
653 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1) {
654 			sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
655 			g_mirror_event_send(disk,
656 			    G_MIRROR_DISK_STATE_DISCONNECTED,
657 			    G_MIRROR_EVENT_DONTWAIT);
658 		}
659 	}
660 	return (error);
661 }
662 
663 static int
664 g_mirror_clear_metadata(struct g_mirror_disk *disk)
665 {
666 	int error;
667 
668 	g_topology_assert_not();
669 	sx_assert(&disk->d_softc->sc_lock, SX_LOCKED);
670 
671 	error = g_mirror_write_metadata(disk, NULL);
672 	if (error == 0) {
673 		G_MIRROR_DEBUG(2, "Metadata on %s cleared.",
674 		    g_mirror_get_diskname(disk));
675 	} else {
676 		G_MIRROR_DEBUG(0,
677 		    "Cannot clear metadata on disk %s (error=%d).",
678 		    g_mirror_get_diskname(disk), error);
679 	}
680 	return (error);
681 }
682 
683 void
684 g_mirror_fill_metadata(struct g_mirror_softc *sc, struct g_mirror_disk *disk,
685     struct g_mirror_metadata *md)
686 {
687 
688 	strlcpy(md->md_magic, G_MIRROR_MAGIC, sizeof(md->md_magic));
689 	md->md_version = G_MIRROR_VERSION;
690 	strlcpy(md->md_name, sc->sc_name, sizeof(md->md_name));
691 	md->md_mid = sc->sc_id;
692 	md->md_all = sc->sc_ndisks;
693 	md->md_slice = sc->sc_slice;
694 	md->md_balance = sc->sc_balance;
695 	md->md_genid = sc->sc_genid;
696 	md->md_mediasize = sc->sc_mediasize;
697 	md->md_sectorsize = sc->sc_sectorsize;
698 	md->md_mflags = (sc->sc_flags & G_MIRROR_DEVICE_FLAG_MASK);
699 	bzero(md->md_provider, sizeof(md->md_provider));
700 	if (disk == NULL) {
701 		md->md_did = arc4random();
702 		md->md_priority = 0;
703 		md->md_syncid = 0;
704 		md->md_dflags = 0;
705 		md->md_sync_offset = 0;
706 		md->md_provsize = 0;
707 	} else {
708 		md->md_did = disk->d_id;
709 		md->md_priority = disk->d_priority;
710 		md->md_syncid = disk->d_sync.ds_syncid;
711 		md->md_dflags = (disk->d_flags & G_MIRROR_DISK_FLAG_MASK);
712 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
713 			md->md_sync_offset = disk->d_sync.ds_offset_done;
714 		else
715 			md->md_sync_offset = 0;
716 		if ((disk->d_flags & G_MIRROR_DISK_FLAG_HARDCODED) != 0) {
717 			strlcpy(md->md_provider,
718 			    disk->d_consumer->provider->name,
719 			    sizeof(md->md_provider));
720 		}
721 		md->md_provsize = disk->d_consumer->provider->mediasize;
722 	}
723 }
724 
725 void
726 g_mirror_update_metadata(struct g_mirror_disk *disk)
727 {
728 	struct g_mirror_softc *sc;
729 	struct g_mirror_metadata md;
730 	int error;
731 
732 	g_topology_assert_not();
733 	sc = disk->d_softc;
734 	sx_assert(&sc->sc_lock, SX_LOCKED);
735 
736 	g_mirror_fill_metadata(sc, disk, &md);
737 	error = g_mirror_write_metadata(disk, &md);
738 	if (error == 0) {
739 		G_MIRROR_DEBUG(2, "Metadata on %s updated.",
740 		    g_mirror_get_diskname(disk));
741 	} else {
742 		G_MIRROR_DEBUG(0,
743 		    "Cannot update metadata on disk %s (error=%d).",
744 		    g_mirror_get_diskname(disk), error);
745 	}
746 }
747 
748 static void
749 g_mirror_bump_syncid(struct g_mirror_softc *sc)
750 {
751 	struct g_mirror_disk *disk;
752 
753 	g_topology_assert_not();
754 	sx_assert(&sc->sc_lock, SX_XLOCKED);
755 	KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
756 	    ("%s called with no active disks (device=%s).", __func__,
757 	    sc->sc_name));
758 
759 	sc->sc_syncid++;
760 	G_MIRROR_DEBUG(1, "Device %s: syncid bumped to %u.", sc->sc_name,
761 	    sc->sc_syncid);
762 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
763 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
764 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
765 			disk->d_sync.ds_syncid = sc->sc_syncid;
766 			g_mirror_update_metadata(disk);
767 		}
768 	}
769 }
770 
771 static void
772 g_mirror_bump_genid(struct g_mirror_softc *sc)
773 {
774 	struct g_mirror_disk *disk;
775 
776 	g_topology_assert_not();
777 	sx_assert(&sc->sc_lock, SX_XLOCKED);
778 	KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
779 	    ("%s called with no active disks (device=%s).", __func__,
780 	    sc->sc_name));
781 
782 	sc->sc_genid++;
783 	G_MIRROR_DEBUG(1, "Device %s: genid bumped to %u.", sc->sc_name,
784 	    sc->sc_genid);
785 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
786 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
787 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
788 			disk->d_genid = sc->sc_genid;
789 			g_mirror_update_metadata(disk);
790 		}
791 	}
792 }
793 
794 static int
795 g_mirror_idle(struct g_mirror_softc *sc, int acw)
796 {
797 	struct g_mirror_disk *disk;
798 	int timeout;
799 
800 	g_topology_assert_not();
801 	sx_assert(&sc->sc_lock, SX_XLOCKED);
802 
803 	if (sc->sc_provider == NULL)
804 		return (0);
805 	if (sc->sc_idle)
806 		return (0);
807 	if (sc->sc_writes > 0)
808 		return (0);
809 	if (acw > 0 || (acw == -1 && sc->sc_provider->acw > 0)) {
810 		timeout = g_mirror_idletime - (time_uptime - sc->sc_last_write);
811 		if (timeout > 0)
812 			return (timeout);
813 	}
814 	sc->sc_idle = 1;
815 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
816 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
817 			continue;
818 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as clean.",
819 		    g_mirror_get_diskname(disk), sc->sc_name);
820 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
821 		g_mirror_update_metadata(disk);
822 	}
823 	return (0);
824 }
825 
826 static void
827 g_mirror_unidle(struct g_mirror_softc *sc)
828 {
829 	struct g_mirror_disk *disk;
830 
831 	g_topology_assert_not();
832 	sx_assert(&sc->sc_lock, SX_XLOCKED);
833 
834 	sc->sc_idle = 0;
835 	sc->sc_last_write = time_uptime;
836 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
837 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
838 			continue;
839 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as dirty.",
840 		    g_mirror_get_diskname(disk), sc->sc_name);
841 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
842 		g_mirror_update_metadata(disk);
843 	}
844 }
845 
846 static __inline int
847 bintime_cmp(struct bintime *bt1, struct bintime *bt2)
848 {
849 
850 	if (bt1->sec < bt2->sec)
851 		return (-1);
852 	else if (bt1->sec > bt2->sec)
853 		return (1);
854 	if (bt1->frac < bt2->frac)
855 		return (-1);
856 	else if (bt1->frac > bt2->frac)
857 		return (1);
858 	return (0);
859 }
860 
861 static void
862 g_mirror_update_delay(struct g_mirror_disk *disk, struct bio *bp)
863 {
864 
865 	if (disk->d_softc->sc_balance != G_MIRROR_BALANCE_LOAD)
866 		return;
867 	binuptime(&disk->d_delay);
868 	bintime_sub(&disk->d_delay, &bp->bio_t0);
869 }
870 
871 static void
872 g_mirror_done(struct bio *bp)
873 {
874 	struct g_mirror_softc *sc;
875 
876 	sc = bp->bio_from->geom->softc;
877 	bp->bio_cflags |= G_MIRROR_BIO_FLAG_REGULAR;
878 	mtx_lock(&sc->sc_queue_mtx);
879 	bioq_disksort(&sc->sc_queue, bp);
880 	wakeup(sc);
881 	mtx_unlock(&sc->sc_queue_mtx);
882 }
883 
884 static void
885 g_mirror_regular_request(struct bio *bp)
886 {
887 	struct g_mirror_softc *sc;
888 	struct g_mirror_disk *disk;
889 	struct bio *pbp;
890 
891 	g_topology_assert_not();
892 
893 	pbp = bp->bio_parent;
894 	sc = pbp->bio_to->geom->softc;
895 	bp->bio_from->index--;
896 	if (bp->bio_cmd == BIO_WRITE)
897 		sc->sc_writes--;
898 	disk = bp->bio_from->private;
899 	if (disk == NULL) {
900 		g_topology_lock();
901 		g_mirror_kill_consumer(sc, bp->bio_from);
902 		g_topology_unlock();
903 	} else {
904 		g_mirror_update_delay(disk, bp);
905 	}
906 
907 	pbp->bio_inbed++;
908 	KASSERT(pbp->bio_inbed <= pbp->bio_children,
909 	    ("bio_inbed (%u) is bigger than bio_children (%u).", pbp->bio_inbed,
910 	    pbp->bio_children));
911 	if (bp->bio_error == 0 && pbp->bio_error == 0) {
912 		G_MIRROR_LOGREQ(3, bp, "Request delivered.");
913 		g_destroy_bio(bp);
914 		if (pbp->bio_children == pbp->bio_inbed) {
915 			G_MIRROR_LOGREQ(3, pbp, "Request delivered.");
916 			pbp->bio_completed = pbp->bio_length;
917 			if (pbp->bio_cmd == BIO_WRITE) {
918 				bioq_remove(&sc->sc_inflight, pbp);
919 				/* Release delayed sync requests if possible. */
920 				g_mirror_sync_release(sc);
921 			}
922 			g_io_deliver(pbp, pbp->bio_error);
923 		}
924 		return;
925 	} else if (bp->bio_error != 0) {
926 		if (pbp->bio_error == 0)
927 			pbp->bio_error = bp->bio_error;
928 		if (disk != NULL) {
929 			if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
930 				disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
931 				G_MIRROR_LOGREQ(0, bp,
932 				    "Request failed (error=%d).",
933 				    bp->bio_error);
934 			} else {
935 				G_MIRROR_LOGREQ(1, bp,
936 				    "Request failed (error=%d).",
937 				    bp->bio_error);
938 			}
939 			if (g_mirror_disconnect_on_failure &&
940 			    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1)
941 			{
942 				sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
943 				g_mirror_event_send(disk,
944 				    G_MIRROR_DISK_STATE_DISCONNECTED,
945 				    G_MIRROR_EVENT_DONTWAIT);
946 			}
947 		}
948 		switch (pbp->bio_cmd) {
949 		case BIO_DELETE:
950 		case BIO_WRITE:
951 			pbp->bio_inbed--;
952 			pbp->bio_children--;
953 			break;
954 		}
955 	}
956 	g_destroy_bio(bp);
957 
958 	switch (pbp->bio_cmd) {
959 	case BIO_READ:
960 		if (pbp->bio_inbed < pbp->bio_children)
961 			break;
962 		if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 1)
963 			g_io_deliver(pbp, pbp->bio_error);
964 		else {
965 			pbp->bio_error = 0;
966 			mtx_lock(&sc->sc_queue_mtx);
967 			bioq_disksort(&sc->sc_queue, pbp);
968 			G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
969 			wakeup(sc);
970 			mtx_unlock(&sc->sc_queue_mtx);
971 		}
972 		break;
973 	case BIO_DELETE:
974 	case BIO_WRITE:
975 		if (pbp->bio_children == 0) {
976 			/*
977 			 * All requests failed.
978 			 */
979 		} else if (pbp->bio_inbed < pbp->bio_children) {
980 			/* Do nothing. */
981 			break;
982 		} else if (pbp->bio_children == pbp->bio_inbed) {
983 			/* Some requests succeeded. */
984 			pbp->bio_error = 0;
985 			pbp->bio_completed = pbp->bio_length;
986 		}
987 		bioq_remove(&sc->sc_inflight, pbp);
988 		/* Release delayed sync requests if possible. */
989 		g_mirror_sync_release(sc);
990 		g_io_deliver(pbp, pbp->bio_error);
991 		break;
992 	default:
993 		KASSERT(1 == 0, ("Invalid request: %u.", pbp->bio_cmd));
994 		break;
995 	}
996 }
997 
998 static void
999 g_mirror_sync_done(struct bio *bp)
1000 {
1001 	struct g_mirror_softc *sc;
1002 
1003 	G_MIRROR_LOGREQ(3, bp, "Synchronization request delivered.");
1004 	sc = bp->bio_from->geom->softc;
1005 	bp->bio_cflags |= G_MIRROR_BIO_FLAG_SYNC;
1006 	mtx_lock(&sc->sc_queue_mtx);
1007 	bioq_disksort(&sc->sc_queue, bp);
1008 	wakeup(sc);
1009 	mtx_unlock(&sc->sc_queue_mtx);
1010 }
1011 
1012 static void
1013 g_mirror_kernel_dump(struct bio *bp)
1014 {
1015 	struct g_mirror_softc *sc;
1016 	struct g_mirror_disk *disk;
1017 	struct bio *cbp;
1018 	struct g_kerneldump *gkd;
1019 
1020 	/*
1021 	 * We configure dumping to the first component, because this component
1022 	 * will be used for reading with 'prefer' balance algorithm.
1023 	 * If the component with the higest priority is currently disconnected
1024 	 * we will not be able to read the dump after the reboot if it will be
1025 	 * connected and synchronized later. Can we do something better?
1026 	 */
1027 	sc = bp->bio_to->geom->softc;
1028 	disk = LIST_FIRST(&sc->sc_disks);
1029 
1030 	gkd = (struct g_kerneldump *)bp->bio_data;
1031 	if (gkd->length > bp->bio_to->mediasize)
1032 		gkd->length = bp->bio_to->mediasize;
1033 	cbp = g_clone_bio(bp);
1034 	if (cbp == NULL) {
1035 		g_io_deliver(bp, ENOMEM);
1036 		return;
1037 	}
1038 	cbp->bio_done = g_std_done;
1039 	g_io_request(cbp, disk->d_consumer);
1040 	G_MIRROR_DEBUG(1, "Kernel dump will go to %s.",
1041 	    g_mirror_get_diskname(disk));
1042 }
1043 
1044 static void
1045 g_mirror_start(struct bio *bp)
1046 {
1047 	struct g_mirror_softc *sc;
1048 
1049 	sc = bp->bio_to->geom->softc;
1050 	/*
1051 	 * If sc == NULL or there are no valid disks, provider's error
1052 	 * should be set and g_mirror_start() should not be called at all.
1053 	 */
1054 	KASSERT(sc != NULL && sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1055 	    ("Provider's error should be set (error=%d)(mirror=%s).",
1056 	    bp->bio_to->error, bp->bio_to->name));
1057 	G_MIRROR_LOGREQ(3, bp, "Request received.");
1058 
1059 	switch (bp->bio_cmd) {
1060 	case BIO_READ:
1061 	case BIO_WRITE:
1062 	case BIO_DELETE:
1063 		break;
1064 	case BIO_GETATTR:
1065 		if (strcmp("GEOM::kerneldump", bp->bio_attribute) == 0) {
1066 			g_mirror_kernel_dump(bp);
1067 			return;
1068 		}
1069 		/* FALLTHROUGH */
1070 	default:
1071 		g_io_deliver(bp, EOPNOTSUPP);
1072 		return;
1073 	}
1074 	mtx_lock(&sc->sc_queue_mtx);
1075 	bioq_disksort(&sc->sc_queue, bp);
1076 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1077 	wakeup(sc);
1078 	mtx_unlock(&sc->sc_queue_mtx);
1079 }
1080 
1081 /*
1082  * Return TRUE if the given request is colliding with a in-progress
1083  * synchronization request.
1084  */
1085 static int
1086 g_mirror_sync_collision(struct g_mirror_softc *sc, struct bio *bp)
1087 {
1088 	struct g_mirror_disk *disk;
1089 	struct bio *sbp;
1090 	off_t rstart, rend, sstart, send;
1091 	int i;
1092 
1093 	if (sc->sc_sync.ds_ndisks == 0)
1094 		return (0);
1095 	rstart = bp->bio_offset;
1096 	rend = bp->bio_offset + bp->bio_length;
1097 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1098 		if (disk->d_state != G_MIRROR_DISK_STATE_SYNCHRONIZING)
1099 			continue;
1100 		for (i = 0; i < g_mirror_syncreqs; i++) {
1101 			sbp = disk->d_sync.ds_bios[i];
1102 			if (sbp == NULL)
1103 				continue;
1104 			sstart = sbp->bio_offset;
1105 			send = sbp->bio_offset + sbp->bio_length;
1106 			if (rend > sstart && rstart < send)
1107 				return (1);
1108 		}
1109 	}
1110 	return (0);
1111 }
1112 
1113 /*
1114  * Return TRUE if the given sync request is colliding with a in-progress regular
1115  * request.
1116  */
1117 static int
1118 g_mirror_regular_collision(struct g_mirror_softc *sc, struct bio *sbp)
1119 {
1120 	off_t rstart, rend, sstart, send;
1121 	struct bio *bp;
1122 
1123 	if (sc->sc_sync.ds_ndisks == 0)
1124 		return (0);
1125 	sstart = sbp->bio_offset;
1126 	send = sbp->bio_offset + sbp->bio_length;
1127 	TAILQ_FOREACH(bp, &sc->sc_inflight.queue, bio_queue) {
1128 		rstart = bp->bio_offset;
1129 		rend = bp->bio_offset + bp->bio_length;
1130 		if (rend > sstart && rstart < send)
1131 			return (1);
1132 	}
1133 	return (0);
1134 }
1135 
1136 /*
1137  * Puts request onto delayed queue.
1138  */
1139 static void
1140 g_mirror_regular_delay(struct g_mirror_softc *sc, struct bio *bp)
1141 {
1142 
1143 	G_MIRROR_LOGREQ(2, bp, "Delaying request.");
1144 	bioq_insert_head(&sc->sc_regular_delayed, bp);
1145 }
1146 
1147 /*
1148  * Puts synchronization request onto delayed queue.
1149  */
1150 static void
1151 g_mirror_sync_delay(struct g_mirror_softc *sc, struct bio *bp)
1152 {
1153 
1154 	G_MIRROR_LOGREQ(2, bp, "Delaying synchronization request.");
1155 	bioq_insert_tail(&sc->sc_sync_delayed, bp);
1156 }
1157 
1158 /*
1159  * Releases delayed regular requests which don't collide anymore with sync
1160  * requests.
1161  */
1162 static void
1163 g_mirror_regular_release(struct g_mirror_softc *sc)
1164 {
1165 	struct bio *bp, *bp2;
1166 
1167 	TAILQ_FOREACH_SAFE(bp, &sc->sc_regular_delayed.queue, bio_queue, bp2) {
1168 		if (g_mirror_sync_collision(sc, bp))
1169 			continue;
1170 		bioq_remove(&sc->sc_regular_delayed, bp);
1171 		G_MIRROR_LOGREQ(2, bp, "Releasing delayed request (%p).", bp);
1172 		mtx_lock(&sc->sc_queue_mtx);
1173 		bioq_insert_head(&sc->sc_queue, bp);
1174 #if 0
1175 		/*
1176 		 * wakeup() is not needed, because this function is called from
1177 		 * the worker thread.
1178 		 */
1179 		wakeup(&sc->sc_queue);
1180 #endif
1181 		mtx_unlock(&sc->sc_queue_mtx);
1182 	}
1183 }
1184 
1185 /*
1186  * Releases delayed sync requests which don't collide anymore with regular
1187  * requests.
1188  */
1189 static void
1190 g_mirror_sync_release(struct g_mirror_softc *sc)
1191 {
1192 	struct bio *bp, *bp2;
1193 
1194 	TAILQ_FOREACH_SAFE(bp, &sc->sc_sync_delayed.queue, bio_queue, bp2) {
1195 		if (g_mirror_regular_collision(sc, bp))
1196 			continue;
1197 		bioq_remove(&sc->sc_sync_delayed, bp);
1198 		G_MIRROR_LOGREQ(2, bp,
1199 		    "Releasing delayed synchronization request.");
1200 		g_io_request(bp, bp->bio_from);
1201 	}
1202 }
1203 
1204 /*
1205  * Handle synchronization requests.
1206  * Every synchronization request is two-steps process: first, READ request is
1207  * send to active provider and then WRITE request (with read data) to the provider
1208  * beeing synchronized. When WRITE is finished, new synchronization request is
1209  * send.
1210  */
1211 static void
1212 g_mirror_sync_request(struct bio *bp)
1213 {
1214 	struct g_mirror_softc *sc;
1215 	struct g_mirror_disk *disk;
1216 
1217 	bp->bio_from->index--;
1218 	sc = bp->bio_from->geom->softc;
1219 	disk = bp->bio_from->private;
1220 	if (disk == NULL) {
1221 		sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1222 		g_topology_lock();
1223 		g_mirror_kill_consumer(sc, bp->bio_from);
1224 		g_topology_unlock();
1225 		free(bp->bio_data, M_MIRROR);
1226 		g_destroy_bio(bp);
1227 		sx_xlock(&sc->sc_lock);
1228 		return;
1229 	}
1230 
1231 	/*
1232 	 * Synchronization request.
1233 	 */
1234 	switch (bp->bio_cmd) {
1235 	case BIO_READ:
1236 	    {
1237 		struct g_consumer *cp;
1238 
1239 		if (bp->bio_error != 0) {
1240 			G_MIRROR_LOGREQ(0, bp,
1241 			    "Synchronization request failed (error=%d).",
1242 			    bp->bio_error);
1243 			g_destroy_bio(bp);
1244 			return;
1245 		}
1246 		G_MIRROR_LOGREQ(3, bp,
1247 		    "Synchronization request half-finished.");
1248 		bp->bio_cmd = BIO_WRITE;
1249 		bp->bio_cflags = 0;
1250 		cp = disk->d_consumer;
1251 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1252 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1253 		    cp->acr, cp->acw, cp->ace));
1254 		cp->index++;
1255 		g_io_request(bp, cp);
1256 		return;
1257 	    }
1258 	case BIO_WRITE:
1259 	    {
1260 		struct g_mirror_disk_sync *sync;
1261 		off_t offset;
1262 		void *data;
1263 		int i;
1264 
1265 		if (bp->bio_error != 0) {
1266 			G_MIRROR_LOGREQ(0, bp,
1267 			    "Synchronization request failed (error=%d).",
1268 			    bp->bio_error);
1269 			g_destroy_bio(bp);
1270 			sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
1271 			g_mirror_event_send(disk,
1272 			    G_MIRROR_DISK_STATE_DISCONNECTED,
1273 			    G_MIRROR_EVENT_DONTWAIT);
1274 			return;
1275 		}
1276 		G_MIRROR_LOGREQ(3, bp, "Synchronization request finished.");
1277 		sync = &disk->d_sync;
1278 		if (sync->ds_offset == sc->sc_mediasize ||
1279 		    sync->ds_consumer == NULL ||
1280 		    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1281 			/* Don't send more synchronization requests. */
1282 			sync->ds_inflight--;
1283 			if (sync->ds_bios != NULL) {
1284 				i = (int)(uintptr_t)bp->bio_caller1;
1285 				sync->ds_bios[i] = NULL;
1286 			}
1287 			free(bp->bio_data, M_MIRROR);
1288 			g_destroy_bio(bp);
1289 			if (sync->ds_inflight > 0)
1290 				return;
1291 			if (sync->ds_consumer == NULL ||
1292 			    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1293 				return;
1294 			}
1295 			/* Disk up-to-date, activate it. */
1296 			g_mirror_event_send(disk, G_MIRROR_DISK_STATE_ACTIVE,
1297 			    G_MIRROR_EVENT_DONTWAIT);
1298 			return;
1299 		}
1300 
1301 		/* Send next synchronization request. */
1302 		data = bp->bio_data;
1303 		bzero(bp, sizeof(*bp));
1304 		bp->bio_cmd = BIO_READ;
1305 		bp->bio_offset = sync->ds_offset;
1306 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1307 		sync->ds_offset += bp->bio_length;
1308 		bp->bio_done = g_mirror_sync_done;
1309 		bp->bio_data = data;
1310 		bp->bio_from = sync->ds_consumer;
1311 		bp->bio_to = sc->sc_provider;
1312 		G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
1313 		sync->ds_consumer->index++;
1314 		/*
1315 		 * Delay the request if it is colliding with a regular request.
1316 		 */
1317 		if (g_mirror_regular_collision(sc, bp))
1318 			g_mirror_sync_delay(sc, bp);
1319 		else
1320 			g_io_request(bp, sync->ds_consumer);
1321 
1322 		/* Release delayed requests if possible. */
1323 		g_mirror_regular_release(sc);
1324 
1325 		/* Find the smallest offset */
1326 		offset = sc->sc_mediasize;
1327 		for (i = 0; i < g_mirror_syncreqs; i++) {
1328 			bp = sync->ds_bios[i];
1329 			if (bp->bio_offset < offset)
1330 				offset = bp->bio_offset;
1331 		}
1332 		if (sync->ds_offset_done + (MAXPHYS * 100) < offset) {
1333 			/* Update offset_done on every 100 blocks. */
1334 			sync->ds_offset_done = offset;
1335 			g_mirror_update_metadata(disk);
1336 		}
1337 		return;
1338 	    }
1339 	default:
1340 		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1341 		    bp->bio_cmd, sc->sc_name));
1342 		break;
1343 	}
1344 }
1345 
1346 static void
1347 g_mirror_request_prefer(struct g_mirror_softc *sc, struct bio *bp)
1348 {
1349 	struct g_mirror_disk *disk;
1350 	struct g_consumer *cp;
1351 	struct bio *cbp;
1352 
1353 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1354 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE)
1355 			break;
1356 	}
1357 	if (disk == NULL) {
1358 		if (bp->bio_error == 0)
1359 			bp->bio_error = ENXIO;
1360 		g_io_deliver(bp, bp->bio_error);
1361 		return;
1362 	}
1363 	cbp = g_clone_bio(bp);
1364 	if (cbp == NULL) {
1365 		if (bp->bio_error == 0)
1366 			bp->bio_error = ENOMEM;
1367 		g_io_deliver(bp, bp->bio_error);
1368 		return;
1369 	}
1370 	/*
1371 	 * Fill in the component buf structure.
1372 	 */
1373 	cp = disk->d_consumer;
1374 	cbp->bio_done = g_mirror_done;
1375 	cbp->bio_to = cp->provider;
1376 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1377 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1378 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1379 	    cp->acw, cp->ace));
1380 	cp->index++;
1381 	g_io_request(cbp, cp);
1382 }
1383 
1384 static void
1385 g_mirror_request_round_robin(struct g_mirror_softc *sc, struct bio *bp)
1386 {
1387 	struct g_mirror_disk *disk;
1388 	struct g_consumer *cp;
1389 	struct bio *cbp;
1390 
1391 	disk = g_mirror_get_disk(sc);
1392 	if (disk == NULL) {
1393 		if (bp->bio_error == 0)
1394 			bp->bio_error = ENXIO;
1395 		g_io_deliver(bp, bp->bio_error);
1396 		return;
1397 	}
1398 	cbp = g_clone_bio(bp);
1399 	if (cbp == NULL) {
1400 		if (bp->bio_error == 0)
1401 			bp->bio_error = ENOMEM;
1402 		g_io_deliver(bp, bp->bio_error);
1403 		return;
1404 	}
1405 	/*
1406 	 * Fill in the component buf structure.
1407 	 */
1408 	cp = disk->d_consumer;
1409 	cbp->bio_done = g_mirror_done;
1410 	cbp->bio_to = cp->provider;
1411 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1412 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1413 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1414 	    cp->acw, cp->ace));
1415 	cp->index++;
1416 	g_io_request(cbp, cp);
1417 }
1418 
1419 static void
1420 g_mirror_request_load(struct g_mirror_softc *sc, struct bio *bp)
1421 {
1422 	struct g_mirror_disk *disk, *dp;
1423 	struct g_consumer *cp;
1424 	struct bio *cbp;
1425 	struct bintime curtime;
1426 
1427 	binuptime(&curtime);
1428 	/*
1429 	 * Find a disk which the smallest load.
1430 	 */
1431 	disk = NULL;
1432 	LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1433 		if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1434 			continue;
1435 		/* If disk wasn't used for more than 2 sec, use it. */
1436 		if (curtime.sec - dp->d_last_used.sec >= 2) {
1437 			disk = dp;
1438 			break;
1439 		}
1440 		if (disk == NULL ||
1441 		    bintime_cmp(&dp->d_delay, &disk->d_delay) < 0) {
1442 			disk = dp;
1443 		}
1444 	}
1445 	KASSERT(disk != NULL, ("NULL disk for %s.", sc->sc_name));
1446 	cbp = g_clone_bio(bp);
1447 	if (cbp == NULL) {
1448 		if (bp->bio_error == 0)
1449 			bp->bio_error = ENOMEM;
1450 		g_io_deliver(bp, bp->bio_error);
1451 		return;
1452 	}
1453 	/*
1454 	 * Fill in the component buf structure.
1455 	 */
1456 	cp = disk->d_consumer;
1457 	cbp->bio_done = g_mirror_done;
1458 	cbp->bio_to = cp->provider;
1459 	binuptime(&disk->d_last_used);
1460 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1461 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1462 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1463 	    cp->acw, cp->ace));
1464 	cp->index++;
1465 	g_io_request(cbp, cp);
1466 }
1467 
1468 static void
1469 g_mirror_request_split(struct g_mirror_softc *sc, struct bio *bp)
1470 {
1471 	struct bio_queue_head queue;
1472 	struct g_mirror_disk *disk;
1473 	struct g_consumer *cp;
1474 	struct bio *cbp;
1475 	off_t left, mod, offset, slice;
1476 	u_char *data;
1477 	u_int ndisks;
1478 
1479 	if (bp->bio_length <= sc->sc_slice) {
1480 		g_mirror_request_round_robin(sc, bp);
1481 		return;
1482 	}
1483 	ndisks = g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE);
1484 	slice = bp->bio_length / ndisks;
1485 	mod = slice % sc->sc_provider->sectorsize;
1486 	if (mod != 0)
1487 		slice += sc->sc_provider->sectorsize - mod;
1488 	/*
1489 	 * Allocate all bios before sending any request, so we can
1490 	 * return ENOMEM in nice and clean way.
1491 	 */
1492 	left = bp->bio_length;
1493 	offset = bp->bio_offset;
1494 	data = bp->bio_data;
1495 	bioq_init(&queue);
1496 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1497 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1498 			continue;
1499 		cbp = g_clone_bio(bp);
1500 		if (cbp == NULL) {
1501 			for (cbp = bioq_first(&queue); cbp != NULL;
1502 			    cbp = bioq_first(&queue)) {
1503 				bioq_remove(&queue, cbp);
1504 				g_destroy_bio(cbp);
1505 			}
1506 			if (bp->bio_error == 0)
1507 				bp->bio_error = ENOMEM;
1508 			g_io_deliver(bp, bp->bio_error);
1509 			return;
1510 		}
1511 		bioq_insert_tail(&queue, cbp);
1512 		cbp->bio_done = g_mirror_done;
1513 		cbp->bio_caller1 = disk;
1514 		cbp->bio_to = disk->d_consumer->provider;
1515 		cbp->bio_offset = offset;
1516 		cbp->bio_data = data;
1517 		cbp->bio_length = MIN(left, slice);
1518 		left -= cbp->bio_length;
1519 		if (left == 0)
1520 			break;
1521 		offset += cbp->bio_length;
1522 		data += cbp->bio_length;
1523 	}
1524 	for (cbp = bioq_first(&queue); cbp != NULL; cbp = bioq_first(&queue)) {
1525 		bioq_remove(&queue, cbp);
1526 		G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1527 		disk = cbp->bio_caller1;
1528 		cbp->bio_caller1 = NULL;
1529 		cp = disk->d_consumer;
1530 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1531 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1532 		    cp->acr, cp->acw, cp->ace));
1533 		disk->d_consumer->index++;
1534 		g_io_request(cbp, disk->d_consumer);
1535 	}
1536 }
1537 
1538 static void
1539 g_mirror_register_request(struct bio *bp)
1540 {
1541 	struct g_mirror_softc *sc;
1542 
1543 	sc = bp->bio_to->geom->softc;
1544 	switch (bp->bio_cmd) {
1545 	case BIO_READ:
1546 		switch (sc->sc_balance) {
1547 		case G_MIRROR_BALANCE_LOAD:
1548 			g_mirror_request_load(sc, bp);
1549 			break;
1550 		case G_MIRROR_BALANCE_PREFER:
1551 			g_mirror_request_prefer(sc, bp);
1552 			break;
1553 		case G_MIRROR_BALANCE_ROUND_ROBIN:
1554 			g_mirror_request_round_robin(sc, bp);
1555 			break;
1556 		case G_MIRROR_BALANCE_SPLIT:
1557 			g_mirror_request_split(sc, bp);
1558 			break;
1559 		}
1560 		return;
1561 	case BIO_WRITE:
1562 	case BIO_DELETE:
1563 	    {
1564 		struct g_mirror_disk *disk;
1565 		struct g_mirror_disk_sync *sync;
1566 		struct bio_queue_head queue;
1567 		struct g_consumer *cp;
1568 		struct bio *cbp;
1569 
1570 		/*
1571 		 * Delay the request if it is colliding with a synchronization
1572 		 * request.
1573 		 */
1574 		if (g_mirror_sync_collision(sc, bp)) {
1575 			g_mirror_regular_delay(sc, bp);
1576 			return;
1577 		}
1578 
1579 		if (sc->sc_idle)
1580 			g_mirror_unidle(sc);
1581 		else
1582 			sc->sc_last_write = time_uptime;
1583 
1584 		/*
1585 		 * Allocate all bios before sending any request, so we can
1586 		 * return ENOMEM in nice and clean way.
1587 		 */
1588 		bioq_init(&queue);
1589 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1590 			sync = &disk->d_sync;
1591 			switch (disk->d_state) {
1592 			case G_MIRROR_DISK_STATE_ACTIVE:
1593 				break;
1594 			case G_MIRROR_DISK_STATE_SYNCHRONIZING:
1595 				if (bp->bio_offset >= sync->ds_offset)
1596 					continue;
1597 				break;
1598 			default:
1599 				continue;
1600 			}
1601 			cbp = g_clone_bio(bp);
1602 			if (cbp == NULL) {
1603 				for (cbp = bioq_first(&queue); cbp != NULL;
1604 				    cbp = bioq_first(&queue)) {
1605 					bioq_remove(&queue, cbp);
1606 					g_destroy_bio(cbp);
1607 				}
1608 				if (bp->bio_error == 0)
1609 					bp->bio_error = ENOMEM;
1610 				g_io_deliver(bp, bp->bio_error);
1611 				return;
1612 			}
1613 			bioq_insert_tail(&queue, cbp);
1614 			cbp->bio_done = g_mirror_done;
1615 			cp = disk->d_consumer;
1616 			cbp->bio_caller1 = cp;
1617 			cbp->bio_to = cp->provider;
1618 			KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1619 			    ("Consumer %s not opened (r%dw%de%d).",
1620 			    cp->provider->name, cp->acr, cp->acw, cp->ace));
1621 		}
1622 		for (cbp = bioq_first(&queue); cbp != NULL;
1623 		    cbp = bioq_first(&queue)) {
1624 			bioq_remove(&queue, cbp);
1625 			G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1626 			cp = cbp->bio_caller1;
1627 			cbp->bio_caller1 = NULL;
1628 			cp->index++;
1629 			sc->sc_writes++;
1630 			g_io_request(cbp, cp);
1631 		}
1632 		/*
1633 		 * Put request onto inflight queue, so we can check if new
1634 		 * synchronization requests don't collide with it.
1635 		 */
1636 		bioq_insert_tail(&sc->sc_inflight, bp);
1637 		/*
1638 		 * Bump syncid on first write.
1639 		 */
1640 		if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0) {
1641 			sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
1642 			g_mirror_bump_syncid(sc);
1643 		}
1644 		return;
1645 	    }
1646 	default:
1647 		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1648 		    bp->bio_cmd, sc->sc_name));
1649 		break;
1650 	}
1651 }
1652 
1653 static int
1654 g_mirror_can_destroy(struct g_mirror_softc *sc)
1655 {
1656 	struct g_geom *gp;
1657 	struct g_consumer *cp;
1658 
1659 	g_topology_assert();
1660 	gp = sc->sc_geom;
1661 	if (gp->softc == NULL)
1662 		return (1);
1663 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1664 		if (g_mirror_is_busy(sc, cp))
1665 			return (0);
1666 	}
1667 	gp = sc->sc_sync.ds_geom;
1668 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1669 		if (g_mirror_is_busy(sc, cp))
1670 			return (0);
1671 	}
1672 	G_MIRROR_DEBUG(2, "No I/O requests for %s, it can be destroyed.",
1673 	    sc->sc_name);
1674 	return (1);
1675 }
1676 
1677 static int
1678 g_mirror_try_destroy(struct g_mirror_softc *sc)
1679 {
1680 
1681 	if (sc->sc_rootmount != NULL) {
1682 		G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
1683 		    sc->sc_rootmount);
1684 		root_mount_rel(sc->sc_rootmount);
1685 		sc->sc_rootmount = NULL;
1686 	}
1687 	g_topology_lock();
1688 	if (!g_mirror_can_destroy(sc)) {
1689 		g_topology_unlock();
1690 		return (0);
1691 	}
1692 	sc->sc_geom->softc = NULL;
1693 	sc->sc_sync.ds_geom->softc = NULL;
1694 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_WAIT) != 0) {
1695 		g_topology_unlock();
1696 		G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1697 		    &sc->sc_worker);
1698 		/* Unlock sc_lock here, as it can be destroyed after wakeup. */
1699 		sx_xunlock(&sc->sc_lock);
1700 		wakeup(&sc->sc_worker);
1701 		sc->sc_worker = NULL;
1702 	} else {
1703 		g_topology_unlock();
1704 		g_mirror_destroy_device(sc);
1705 		free(sc, M_MIRROR);
1706 	}
1707 	return (1);
1708 }
1709 
1710 /*
1711  * Worker thread.
1712  */
1713 static void
1714 g_mirror_worker(void *arg)
1715 {
1716 	struct g_mirror_softc *sc;
1717 	struct g_mirror_event *ep;
1718 	struct bio *bp;
1719 	int timeout;
1720 
1721 	sc = arg;
1722 	mtx_lock_spin(&sched_lock);
1723 	sched_prio(curthread, PRIBIO);
1724 	mtx_unlock_spin(&sched_lock);
1725 
1726 	sx_xlock(&sc->sc_lock);
1727 	for (;;) {
1728 		G_MIRROR_DEBUG(5, "%s: Let's see...", __func__);
1729 		/*
1730 		 * First take a look at events.
1731 		 * This is important to handle events before any I/O requests.
1732 		 */
1733 		ep = g_mirror_event_get(sc);
1734 		if (ep != NULL) {
1735 			g_mirror_event_remove(sc, ep);
1736 			if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0) {
1737 				/* Update only device status. */
1738 				G_MIRROR_DEBUG(3,
1739 				    "Running event for device %s.",
1740 				    sc->sc_name);
1741 				ep->e_error = 0;
1742 				g_mirror_update_device(sc, 1);
1743 			} else {
1744 				/* Update disk status. */
1745 				G_MIRROR_DEBUG(3, "Running event for disk %s.",
1746 				     g_mirror_get_diskname(ep->e_disk));
1747 				ep->e_error = g_mirror_update_disk(ep->e_disk,
1748 				    ep->e_state);
1749 				if (ep->e_error == 0)
1750 					g_mirror_update_device(sc, 0);
1751 			}
1752 			if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0) {
1753 				KASSERT(ep->e_error == 0,
1754 				    ("Error cannot be handled."));
1755 				g_mirror_event_free(ep);
1756 			} else {
1757 				ep->e_flags |= G_MIRROR_EVENT_DONE;
1758 				G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1759 				    ep);
1760 				mtx_lock(&sc->sc_events_mtx);
1761 				wakeup(ep);
1762 				mtx_unlock(&sc->sc_events_mtx);
1763 			}
1764 			if ((sc->sc_flags &
1765 			    G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1766 				if (g_mirror_try_destroy(sc)) {
1767 					curthread->td_pflags &= ~TDP_GEOM;
1768 					G_MIRROR_DEBUG(1, "Thread exiting.");
1769 					kthread_exit(0);
1770 				}
1771 			}
1772 			G_MIRROR_DEBUG(5, "%s: I'm here 1.", __func__);
1773 			continue;
1774 		}
1775 		/*
1776 		 * Check if we can mark array as CLEAN and if we can't take
1777 		 * how much seconds should we wait.
1778 		 */
1779 		timeout = g_mirror_idle(sc, -1);
1780 		/*
1781 		 * Now I/O requests.
1782 		 */
1783 		/* Get first request from the queue. */
1784 		mtx_lock(&sc->sc_queue_mtx);
1785 		bp = bioq_first(&sc->sc_queue);
1786 		if (bp == NULL) {
1787 			if ((sc->sc_flags &
1788 			    G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1789 				mtx_unlock(&sc->sc_queue_mtx);
1790 				if (g_mirror_try_destroy(sc)) {
1791 					curthread->td_pflags &= ~TDP_GEOM;
1792 					G_MIRROR_DEBUG(1, "Thread exiting.");
1793 					kthread_exit(0);
1794 				}
1795 				mtx_lock(&sc->sc_queue_mtx);
1796 			}
1797 			sx_xunlock(&sc->sc_lock);
1798 			/*
1799 			 * XXX: We can miss an event here, because an event
1800 			 *      can be added without sx-device-lock and without
1801 			 *      mtx-queue-lock. Maybe I should just stop using
1802 			 *      dedicated mutex for events synchronization and
1803 			 *      stick with the queue lock?
1804 			 *      The event will hang here until next I/O request
1805 			 *      or next event is received.
1806 			 */
1807 			MSLEEP(sc, &sc->sc_queue_mtx, PRIBIO | PDROP, "m:w1",
1808 			    timeout * hz);
1809 			sx_xlock(&sc->sc_lock);
1810 			G_MIRROR_DEBUG(5, "%s: I'm here 4.", __func__);
1811 			continue;
1812 		}
1813 		bioq_remove(&sc->sc_queue, bp);
1814 		mtx_unlock(&sc->sc_queue_mtx);
1815 
1816 		if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_REGULAR) != 0)
1817 			g_mirror_regular_request(bp);
1818 		else if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0)
1819 			g_mirror_sync_request(bp);
1820 		else
1821 			g_mirror_register_request(bp);
1822 		G_MIRROR_DEBUG(5, "%s: I'm here 9.", __func__);
1823 	}
1824 }
1825 
1826 static void
1827 g_mirror_update_idle(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
1828 {
1829 
1830 	sx_assert(&sc->sc_lock, SX_LOCKED);
1831 
1832 	if (!sc->sc_idle && (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) == 0) {
1833 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as dirty.",
1834 		    g_mirror_get_diskname(disk), sc->sc_name);
1835 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
1836 	} else if (sc->sc_idle &&
1837 	    (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
1838 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as clean.",
1839 		    g_mirror_get_diskname(disk), sc->sc_name);
1840 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
1841 	}
1842 }
1843 
1844 static void
1845 g_mirror_sync_start(struct g_mirror_disk *disk)
1846 {
1847 	struct g_mirror_softc *sc;
1848 	struct g_consumer *cp;
1849 	struct bio *bp;
1850 	int error, i;
1851 
1852 	g_topology_assert_not();
1853 	sc = disk->d_softc;
1854 	sx_assert(&sc->sc_lock, SX_LOCKED);
1855 
1856 	KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
1857 	    ("Disk %s is not marked for synchronization.",
1858 	    g_mirror_get_diskname(disk)));
1859 	KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1860 	    ("Device not in RUNNING state (%s, %u).", sc->sc_name,
1861 	    sc->sc_state));
1862 
1863 	sx_xunlock(&sc->sc_lock);
1864 	g_topology_lock();
1865 	cp = g_new_consumer(sc->sc_sync.ds_geom);
1866 	error = g_attach(cp, sc->sc_provider);
1867 	KASSERT(error == 0,
1868 	    ("Cannot attach to %s (error=%d).", sc->sc_name, error));
1869 	error = g_access(cp, 1, 0, 0);
1870 	KASSERT(error == 0, ("Cannot open %s (error=%d).", sc->sc_name, error));
1871 	g_topology_unlock();
1872 	sx_xlock(&sc->sc_lock);
1873 
1874 	G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s.", sc->sc_name,
1875 	    g_mirror_get_diskname(disk));
1876 	disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
1877 	KASSERT(disk->d_sync.ds_consumer == NULL,
1878 	    ("Sync consumer already exists (device=%s, disk=%s).",
1879 	    sc->sc_name, g_mirror_get_diskname(disk)));
1880 
1881 	disk->d_sync.ds_consumer = cp;
1882 	disk->d_sync.ds_consumer->private = disk;
1883 	disk->d_sync.ds_consumer->index = 0;
1884 
1885 	/*
1886 	 * Allocate memory for synchronization bios and initialize them.
1887 	 */
1888 	disk->d_sync.ds_bios = malloc(sizeof(struct bio *) * g_mirror_syncreqs,
1889 	    M_MIRROR, M_WAITOK);
1890 	for (i = 0; i < g_mirror_syncreqs; i++) {
1891 		bp = g_alloc_bio();
1892 		disk->d_sync.ds_bios[i] = bp;
1893 		bp->bio_parent = NULL;
1894 		bp->bio_cmd = BIO_READ;
1895 		bp->bio_data = malloc(MAXPHYS, M_MIRROR, M_WAITOK);
1896 		bp->bio_cflags = 0;
1897 		bp->bio_offset = disk->d_sync.ds_offset;
1898 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1899 		disk->d_sync.ds_offset += bp->bio_length;
1900 		bp->bio_done = g_mirror_sync_done;
1901 		bp->bio_from = disk->d_sync.ds_consumer;
1902 		bp->bio_to = sc->sc_provider;
1903 		bp->bio_caller1 = (void *)(uintptr_t)i;
1904 	}
1905 
1906 	/* Increase the number of disks in SYNCHRONIZING state. */
1907 	sc->sc_sync.ds_ndisks++;
1908 	/* Set the number of in-flight synchronization requests. */
1909 	disk->d_sync.ds_inflight = g_mirror_syncreqs;
1910 
1911 	/*
1912 	 * Fire off first synchronization requests.
1913 	 */
1914 	for (i = 0; i < g_mirror_syncreqs; i++) {
1915 		bp = disk->d_sync.ds_bios[i];
1916 		G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
1917 		disk->d_sync.ds_consumer->index++;
1918 		/*
1919 		 * Delay the request if it is colliding with a regular request.
1920 		 */
1921 		if (g_mirror_regular_collision(sc, bp))
1922 			g_mirror_sync_delay(sc, bp);
1923 		else
1924 			g_io_request(bp, disk->d_sync.ds_consumer);
1925 	}
1926 }
1927 
1928 /*
1929  * Stop synchronization process.
1930  * type: 0 - synchronization finished
1931  *       1 - synchronization stopped
1932  */
1933 static void
1934 g_mirror_sync_stop(struct g_mirror_disk *disk, int type)
1935 {
1936 	struct g_mirror_softc *sc;
1937 	struct g_consumer *cp;
1938 
1939 	g_topology_assert_not();
1940 	sc = disk->d_softc;
1941 	sx_assert(&sc->sc_lock, SX_LOCKED);
1942 
1943 	KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
1944 	    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
1945 	    g_mirror_disk_state2str(disk->d_state)));
1946 	if (disk->d_sync.ds_consumer == NULL)
1947 		return;
1948 
1949 	if (type == 0) {
1950 		G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s finished.",
1951 		    sc->sc_name, g_mirror_get_diskname(disk));
1952 	} else /* if (type == 1) */ {
1953 		G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s stopped.",
1954 		    sc->sc_name, g_mirror_get_diskname(disk));
1955 	}
1956 	free(disk->d_sync.ds_bios, M_MIRROR);
1957 	disk->d_sync.ds_bios = NULL;
1958 	cp = disk->d_sync.ds_consumer;
1959 	disk->d_sync.ds_consumer = NULL;
1960 	disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
1961 	sc->sc_sync.ds_ndisks--;
1962 	sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1963 	g_topology_lock();
1964 	g_mirror_kill_consumer(sc, cp);
1965 	g_topology_unlock();
1966 	sx_xlock(&sc->sc_lock);
1967 }
1968 
1969 static void
1970 g_mirror_launch_provider(struct g_mirror_softc *sc)
1971 {
1972 	struct g_mirror_disk *disk;
1973 	struct g_provider *pp;
1974 
1975 	sx_assert(&sc->sc_lock, SX_LOCKED);
1976 
1977 	g_topology_lock();
1978 	pp = g_new_providerf(sc->sc_geom, "mirror/%s", sc->sc_name);
1979 	pp->mediasize = sc->sc_mediasize;
1980 	pp->sectorsize = sc->sc_sectorsize;
1981 	sc->sc_provider = pp;
1982 	g_error_provider(pp, 0);
1983 	g_topology_unlock();
1984 	G_MIRROR_DEBUG(0, "Device %s: provider %s launched.", sc->sc_name,
1985 	    pp->name);
1986 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1987 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
1988 			g_mirror_sync_start(disk);
1989 	}
1990 }
1991 
1992 static void
1993 g_mirror_destroy_provider(struct g_mirror_softc *sc)
1994 {
1995 	struct g_mirror_disk *disk;
1996 	struct bio *bp;
1997 
1998 	g_topology_assert_not();
1999 	KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
2000 	    sc->sc_name));
2001 
2002 	g_topology_lock();
2003 	g_error_provider(sc->sc_provider, ENXIO);
2004 	mtx_lock(&sc->sc_queue_mtx);
2005 	while ((bp = bioq_first(&sc->sc_queue)) != NULL) {
2006 		bioq_remove(&sc->sc_queue, bp);
2007 		g_io_deliver(bp, ENXIO);
2008 	}
2009 	mtx_unlock(&sc->sc_queue_mtx);
2010 	G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.", sc->sc_name,
2011 	    sc->sc_provider->name);
2012 	sc->sc_provider->flags |= G_PF_WITHER;
2013 	g_orphan_provider(sc->sc_provider, ENXIO);
2014 	g_topology_unlock();
2015 	sc->sc_provider = NULL;
2016 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2017 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2018 			g_mirror_sync_stop(disk, 1);
2019 	}
2020 }
2021 
2022 static void
2023 g_mirror_go(void *arg)
2024 {
2025 	struct g_mirror_softc *sc;
2026 
2027 	sc = arg;
2028 	G_MIRROR_DEBUG(0, "Force device %s start due to timeout.", sc->sc_name);
2029 	g_mirror_event_send(sc, 0,
2030 	    G_MIRROR_EVENT_DONTWAIT | G_MIRROR_EVENT_DEVICE);
2031 }
2032 
2033 static u_int
2034 g_mirror_determine_state(struct g_mirror_disk *disk)
2035 {
2036 	struct g_mirror_softc *sc;
2037 	u_int state;
2038 
2039 	sc = disk->d_softc;
2040 	if (sc->sc_syncid == disk->d_sync.ds_syncid) {
2041 		if ((disk->d_flags &
2042 		    G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2043 			/* Disk does not need synchronization. */
2044 			state = G_MIRROR_DISK_STATE_ACTIVE;
2045 		} else {
2046 			if ((sc->sc_flags &
2047 			     G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2048 			    (disk->d_flags &
2049 			     G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2050 				/*
2051 				 * We can start synchronization from
2052 				 * the stored offset.
2053 				 */
2054 				state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2055 			} else {
2056 				state = G_MIRROR_DISK_STATE_STALE;
2057 			}
2058 		}
2059 	} else if (disk->d_sync.ds_syncid < sc->sc_syncid) {
2060 		/*
2061 		 * Reset all synchronization data for this disk,
2062 		 * because if it even was synchronized, it was
2063 		 * synchronized to disks with different syncid.
2064 		 */
2065 		disk->d_flags |= G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2066 		disk->d_sync.ds_offset = 0;
2067 		disk->d_sync.ds_offset_done = 0;
2068 		disk->d_sync.ds_syncid = sc->sc_syncid;
2069 		if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2070 		    (disk->d_flags & G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2071 			state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2072 		} else {
2073 			state = G_MIRROR_DISK_STATE_STALE;
2074 		}
2075 	} else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
2076 		/*
2077 		 * Not good, NOT GOOD!
2078 		 * It means that mirror was started on stale disks
2079 		 * and more fresh disk just arrive.
2080 		 * If there were writes, mirror is fucked up, sorry.
2081 		 * I think the best choice here is don't touch
2082 		 * this disk and inform the user laudly.
2083 		 */
2084 		G_MIRROR_DEBUG(0, "Device %s was started before the freshest "
2085 		    "disk (%s) arrives!! It will not be connected to the "
2086 		    "running device.", sc->sc_name,
2087 		    g_mirror_get_diskname(disk));
2088 		g_mirror_destroy_disk(disk);
2089 		state = G_MIRROR_DISK_STATE_NONE;
2090 		/* Return immediately, because disk was destroyed. */
2091 		return (state);
2092 	}
2093 	G_MIRROR_DEBUG(3, "State for %s disk: %s.",
2094 	    g_mirror_get_diskname(disk), g_mirror_disk_state2str(state));
2095 	return (state);
2096 }
2097 
2098 /*
2099  * Update device state.
2100  */
2101 static void
2102 g_mirror_update_device(struct g_mirror_softc *sc, boolean_t force)
2103 {
2104 	struct g_mirror_disk *disk;
2105 	u_int state;
2106 
2107 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2108 
2109 	switch (sc->sc_state) {
2110 	case G_MIRROR_DEVICE_STATE_STARTING:
2111 	    {
2112 		struct g_mirror_disk *pdisk, *tdisk;
2113 		u_int dirty, ndisks, genid, syncid;
2114 
2115 		KASSERT(sc->sc_provider == NULL,
2116 		    ("Non-NULL provider in STARTING state (%s).", sc->sc_name));
2117 		/*
2118 		 * Are we ready? We are, if all disks are connected or
2119 		 * if we have any disks and 'force' is true.
2120 		 */
2121 		ndisks = g_mirror_ndisks(sc, -1);
2122 		if (sc->sc_ndisks == ndisks || (force && ndisks > 0)) {
2123 			;
2124 		} else if (ndisks == 0) {
2125 			/*
2126 			 * Disks went down in starting phase, so destroy
2127 			 * device.
2128 			 */
2129 			callout_drain(&sc->sc_callout);
2130 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2131 			G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2132 			    sc->sc_rootmount);
2133 			root_mount_rel(sc->sc_rootmount);
2134 			sc->sc_rootmount = NULL;
2135 			return;
2136 		} else {
2137 			return;
2138 		}
2139 
2140 		/*
2141 		 * Activate all disks with the biggest syncid.
2142 		 */
2143 		if (force) {
2144 			/*
2145 			 * If 'force' is true, we have been called due to
2146 			 * timeout, so don't bother canceling timeout.
2147 			 */
2148 			ndisks = 0;
2149 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2150 				if ((disk->d_flags &
2151 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2152 					ndisks++;
2153 				}
2154 			}
2155 			if (ndisks == 0) {
2156 				/* No valid disks found, destroy device. */
2157 				sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2158 				G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2159 				    __LINE__, sc->sc_rootmount);
2160 				root_mount_rel(sc->sc_rootmount);
2161 				sc->sc_rootmount = NULL;
2162 				return;
2163 			}
2164 		} else {
2165 			/* Cancel timeout. */
2166 			callout_drain(&sc->sc_callout);
2167 		}
2168 
2169 		/*
2170 		 * Find the biggest genid.
2171 		 */
2172 		genid = 0;
2173 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2174 			if (disk->d_genid > genid)
2175 				genid = disk->d_genid;
2176 		}
2177 		sc->sc_genid = genid;
2178 		/*
2179 		 * Remove all disks without the biggest genid.
2180 		 */
2181 		LIST_FOREACH_SAFE(disk, &sc->sc_disks, d_next, tdisk) {
2182 			if (disk->d_genid < genid) {
2183 				G_MIRROR_DEBUG(0,
2184 				    "Component %s (device %s) broken, skipping.",
2185 				    g_mirror_get_diskname(disk), sc->sc_name);
2186 				g_mirror_destroy_disk(disk);
2187 			}
2188 		}
2189 
2190 		/*
2191 		 * Find the biggest syncid.
2192 		 */
2193 		syncid = 0;
2194 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2195 			if (disk->d_sync.ds_syncid > syncid)
2196 				syncid = disk->d_sync.ds_syncid;
2197 		}
2198 
2199 		/*
2200 		 * Here we need to look for dirty disks and if all disks
2201 		 * with the biggest syncid are dirty, we have to choose
2202 		 * one with the biggest priority and rebuild the rest.
2203 		 */
2204 		/*
2205 		 * Find the number of dirty disks with the biggest syncid.
2206 		 * Find the number of disks with the biggest syncid.
2207 		 * While here, find a disk with the biggest priority.
2208 		 */
2209 		dirty = ndisks = 0;
2210 		pdisk = NULL;
2211 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2212 			if (disk->d_sync.ds_syncid != syncid)
2213 				continue;
2214 			if ((disk->d_flags &
2215 			    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2216 				continue;
2217 			}
2218 			ndisks++;
2219 			if ((disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
2220 				dirty++;
2221 				if (pdisk == NULL ||
2222 				    pdisk->d_priority < disk->d_priority) {
2223 					pdisk = disk;
2224 				}
2225 			}
2226 		}
2227 		if (dirty == 0) {
2228 			/* No dirty disks at all, great. */
2229 		} else if (dirty == ndisks) {
2230 			/*
2231 			 * Force synchronization for all dirty disks except one
2232 			 * with the biggest priority.
2233 			 */
2234 			KASSERT(pdisk != NULL, ("pdisk == NULL"));
2235 			G_MIRROR_DEBUG(1, "Using disk %s (device %s) as a "
2236 			    "master disk for synchronization.",
2237 			    g_mirror_get_diskname(pdisk), sc->sc_name);
2238 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2239 				if (disk->d_sync.ds_syncid != syncid)
2240 					continue;
2241 				if ((disk->d_flags &
2242 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2243 					continue;
2244 				}
2245 				KASSERT((disk->d_flags &
2246 				    G_MIRROR_DISK_FLAG_DIRTY) != 0,
2247 				    ("Disk %s isn't marked as dirty.",
2248 				    g_mirror_get_diskname(disk)));
2249 				/* Skip the disk with the biggest priority. */
2250 				if (disk == pdisk)
2251 					continue;
2252 				disk->d_sync.ds_syncid = 0;
2253 			}
2254 		} else if (dirty < ndisks) {
2255 			/*
2256 			 * Force synchronization for all dirty disks.
2257 			 * We have some non-dirty disks.
2258 			 */
2259 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2260 				if (disk->d_sync.ds_syncid != syncid)
2261 					continue;
2262 				if ((disk->d_flags &
2263 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2264 					continue;
2265 				}
2266 				if ((disk->d_flags &
2267 				    G_MIRROR_DISK_FLAG_DIRTY) == 0) {
2268 					continue;
2269 				}
2270 				disk->d_sync.ds_syncid = 0;
2271 			}
2272 		}
2273 
2274 		/* Reset hint. */
2275 		sc->sc_hint = NULL;
2276 		sc->sc_syncid = syncid;
2277 		if (force) {
2278 			/* Remember to bump syncid on first write. */
2279 			sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2280 		}
2281 		state = G_MIRROR_DEVICE_STATE_RUNNING;
2282 		G_MIRROR_DEBUG(1, "Device %s state changed from %s to %s.",
2283 		    sc->sc_name, g_mirror_device_state2str(sc->sc_state),
2284 		    g_mirror_device_state2str(state));
2285 		sc->sc_state = state;
2286 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2287 			state = g_mirror_determine_state(disk);
2288 			g_mirror_event_send(disk, state,
2289 			    G_MIRROR_EVENT_DONTWAIT);
2290 			if (state == G_MIRROR_DISK_STATE_STALE)
2291 				sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2292 		}
2293 		break;
2294 	    }
2295 	case G_MIRROR_DEVICE_STATE_RUNNING:
2296 		if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 0 &&
2297 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2298 			/*
2299 			 * No active disks or no disks at all,
2300 			 * so destroy device.
2301 			 */
2302 			if (sc->sc_provider != NULL)
2303 				g_mirror_destroy_provider(sc);
2304 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2305 			break;
2306 		} else if (g_mirror_ndisks(sc,
2307 		    G_MIRROR_DISK_STATE_ACTIVE) > 0 &&
2308 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2309 			/*
2310 			 * We have active disks, launch provider if it doesn't
2311 			 * exist.
2312 			 */
2313 			if (sc->sc_provider == NULL)
2314 				g_mirror_launch_provider(sc);
2315 			if (sc->sc_rootmount != NULL) {
2316 				G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2317 				    __LINE__, sc->sc_rootmount);
2318 				root_mount_rel(sc->sc_rootmount);
2319 				sc->sc_rootmount = NULL;
2320 			}
2321 		}
2322 		/*
2323 		 * Genid should be bumped immediately, so do it here.
2324 		 */
2325 		if ((sc->sc_bump_id & G_MIRROR_BUMP_GENID) != 0) {
2326 			sc->sc_bump_id &= ~G_MIRROR_BUMP_GENID;
2327 			g_mirror_bump_genid(sc);
2328 		}
2329 		break;
2330 	default:
2331 		KASSERT(1 == 0, ("Wrong device state (%s, %s).",
2332 		    sc->sc_name, g_mirror_device_state2str(sc->sc_state)));
2333 		break;
2334 	}
2335 }
2336 
2337 /*
2338  * Update disk state and device state if needed.
2339  */
2340 #define	DISK_STATE_CHANGED()	G_MIRROR_DEBUG(1,			\
2341 	"Disk %s state changed from %s to %s (device %s).",		\
2342 	g_mirror_get_diskname(disk),					\
2343 	g_mirror_disk_state2str(disk->d_state),				\
2344 	g_mirror_disk_state2str(state), sc->sc_name)
2345 static int
2346 g_mirror_update_disk(struct g_mirror_disk *disk, u_int state)
2347 {
2348 	struct g_mirror_softc *sc;
2349 
2350 	sc = disk->d_softc;
2351 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2352 
2353 again:
2354 	G_MIRROR_DEBUG(3, "Changing disk %s state from %s to %s.",
2355 	    g_mirror_get_diskname(disk), g_mirror_disk_state2str(disk->d_state),
2356 	    g_mirror_disk_state2str(state));
2357 	switch (state) {
2358 	case G_MIRROR_DISK_STATE_NEW:
2359 		/*
2360 		 * Possible scenarios:
2361 		 * 1. New disk arrive.
2362 		 */
2363 		/* Previous state should be NONE. */
2364 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NONE,
2365 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2366 		    g_mirror_disk_state2str(disk->d_state)));
2367 		DISK_STATE_CHANGED();
2368 
2369 		disk->d_state = state;
2370 		if (LIST_EMPTY(&sc->sc_disks))
2371 			LIST_INSERT_HEAD(&sc->sc_disks, disk, d_next);
2372 		else {
2373 			struct g_mirror_disk *dp;
2374 
2375 			LIST_FOREACH(dp, &sc->sc_disks, d_next) {
2376 				if (disk->d_priority >= dp->d_priority) {
2377 					LIST_INSERT_BEFORE(dp, disk, d_next);
2378 					dp = NULL;
2379 					break;
2380 				}
2381 				if (LIST_NEXT(dp, d_next) == NULL)
2382 					break;
2383 			}
2384 			if (dp != NULL)
2385 				LIST_INSERT_AFTER(dp, disk, d_next);
2386 		}
2387 		G_MIRROR_DEBUG(0, "Device %s: provider %s detected.",
2388 		    sc->sc_name, g_mirror_get_diskname(disk));
2389 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
2390 			break;
2391 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2392 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2393 		    g_mirror_device_state2str(sc->sc_state),
2394 		    g_mirror_get_diskname(disk),
2395 		    g_mirror_disk_state2str(disk->d_state)));
2396 		state = g_mirror_determine_state(disk);
2397 		if (state != G_MIRROR_DISK_STATE_NONE)
2398 			goto again;
2399 		break;
2400 	case G_MIRROR_DISK_STATE_ACTIVE:
2401 		/*
2402 		 * Possible scenarios:
2403 		 * 1. New disk does not need synchronization.
2404 		 * 2. Synchronization process finished successfully.
2405 		 */
2406 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2407 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2408 		    g_mirror_device_state2str(sc->sc_state),
2409 		    g_mirror_get_diskname(disk),
2410 		    g_mirror_disk_state2str(disk->d_state)));
2411 		/* Previous state should be NEW or SYNCHRONIZING. */
2412 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW ||
2413 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2414 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2415 		    g_mirror_disk_state2str(disk->d_state)));
2416 		DISK_STATE_CHANGED();
2417 
2418 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
2419 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2420 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_FORCE_SYNC;
2421 			g_mirror_sync_stop(disk, 0);
2422 		}
2423 		disk->d_state = state;
2424 		disk->d_sync.ds_offset = 0;
2425 		disk->d_sync.ds_offset_done = 0;
2426 		g_mirror_update_idle(sc, disk);
2427 		g_mirror_update_metadata(disk);
2428 		G_MIRROR_DEBUG(0, "Device %s: provider %s activated.",
2429 		    sc->sc_name, g_mirror_get_diskname(disk));
2430 		break;
2431 	case G_MIRROR_DISK_STATE_STALE:
2432 		/*
2433 		 * Possible scenarios:
2434 		 * 1. Stale disk was connected.
2435 		 */
2436 		/* Previous state should be NEW. */
2437 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2438 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2439 		    g_mirror_disk_state2str(disk->d_state)));
2440 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2441 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2442 		    g_mirror_device_state2str(sc->sc_state),
2443 		    g_mirror_get_diskname(disk),
2444 		    g_mirror_disk_state2str(disk->d_state)));
2445 		/*
2446 		 * STALE state is only possible if device is marked
2447 		 * NOAUTOSYNC.
2448 		 */
2449 		KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) != 0,
2450 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2451 		    g_mirror_device_state2str(sc->sc_state),
2452 		    g_mirror_get_diskname(disk),
2453 		    g_mirror_disk_state2str(disk->d_state)));
2454 		DISK_STATE_CHANGED();
2455 
2456 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2457 		disk->d_state = state;
2458 		g_mirror_update_metadata(disk);
2459 		G_MIRROR_DEBUG(0, "Device %s: provider %s is stale.",
2460 		    sc->sc_name, g_mirror_get_diskname(disk));
2461 		break;
2462 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
2463 		/*
2464 		 * Possible scenarios:
2465 		 * 1. Disk which needs synchronization was connected.
2466 		 */
2467 		/* Previous state should be NEW. */
2468 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2469 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2470 		    g_mirror_disk_state2str(disk->d_state)));
2471 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2472 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2473 		    g_mirror_device_state2str(sc->sc_state),
2474 		    g_mirror_get_diskname(disk),
2475 		    g_mirror_disk_state2str(disk->d_state)));
2476 		DISK_STATE_CHANGED();
2477 
2478 		if (disk->d_state == G_MIRROR_DISK_STATE_NEW)
2479 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2480 		disk->d_state = state;
2481 		if (sc->sc_provider != NULL) {
2482 			g_mirror_sync_start(disk);
2483 			g_mirror_update_metadata(disk);
2484 		}
2485 		break;
2486 	case G_MIRROR_DISK_STATE_DISCONNECTED:
2487 		/*
2488 		 * Possible scenarios:
2489 		 * 1. Device wasn't running yet, but disk disappear.
2490 		 * 2. Disk was active and disapppear.
2491 		 * 3. Disk disappear during synchronization process.
2492 		 */
2493 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING) {
2494 			/*
2495 			 * Previous state should be ACTIVE, STALE or
2496 			 * SYNCHRONIZING.
2497 			 */
2498 			KASSERT(disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
2499 			    disk->d_state == G_MIRROR_DISK_STATE_STALE ||
2500 			    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2501 			    ("Wrong disk state (%s, %s).",
2502 			    g_mirror_get_diskname(disk),
2503 			    g_mirror_disk_state2str(disk->d_state)));
2504 		} else if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING) {
2505 			/* Previous state should be NEW. */
2506 			KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2507 			    ("Wrong disk state (%s, %s).",
2508 			    g_mirror_get_diskname(disk),
2509 			    g_mirror_disk_state2str(disk->d_state)));
2510 			/*
2511 			 * Reset bumping syncid if disk disappeared in STARTING
2512 			 * state.
2513 			 */
2514 			if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0)
2515 				sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
2516 #ifdef	INVARIANTS
2517 		} else {
2518 			KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
2519 			    sc->sc_name,
2520 			    g_mirror_device_state2str(sc->sc_state),
2521 			    g_mirror_get_diskname(disk),
2522 			    g_mirror_disk_state2str(disk->d_state)));
2523 #endif
2524 		}
2525 		DISK_STATE_CHANGED();
2526 		G_MIRROR_DEBUG(0, "Device %s: provider %s disconnected.",
2527 		    sc->sc_name, g_mirror_get_diskname(disk));
2528 
2529 		g_mirror_destroy_disk(disk);
2530 		break;
2531 	case G_MIRROR_DISK_STATE_DESTROY:
2532 	    {
2533 		int error;
2534 
2535 		error = g_mirror_clear_metadata(disk);
2536 		if (error != 0)
2537 			return (error);
2538 		DISK_STATE_CHANGED();
2539 		G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.",
2540 		    sc->sc_name, g_mirror_get_diskname(disk));
2541 
2542 		g_mirror_destroy_disk(disk);
2543 		sc->sc_ndisks--;
2544 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2545 			g_mirror_update_metadata(disk);
2546 		}
2547 		break;
2548 	    }
2549 	default:
2550 		KASSERT(1 == 0, ("Unknown state (%u).", state));
2551 		break;
2552 	}
2553 	return (0);
2554 }
2555 #undef	DISK_STATE_CHANGED
2556 
2557 int
2558 g_mirror_read_metadata(struct g_consumer *cp, struct g_mirror_metadata *md)
2559 {
2560 	struct g_provider *pp;
2561 	u_char *buf;
2562 	int error;
2563 
2564 	g_topology_assert();
2565 
2566 	error = g_access(cp, 1, 0, 0);
2567 	if (error != 0)
2568 		return (error);
2569 	pp = cp->provider;
2570 	g_topology_unlock();
2571 	/* Metadata are stored on last sector. */
2572 	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
2573 	    &error);
2574 	g_topology_lock();
2575 	g_access(cp, -1, 0, 0);
2576 	if (buf == NULL) {
2577 		G_MIRROR_DEBUG(1, "Cannot read metadata from %s (error=%d).",
2578 		    cp->provider->name, error);
2579 		return (error);
2580 	}
2581 
2582 	/* Decode metadata. */
2583 	error = mirror_metadata_decode(buf, md);
2584 	g_free(buf);
2585 	if (strcmp(md->md_magic, G_MIRROR_MAGIC) != 0)
2586 		return (EINVAL);
2587 	if (md->md_version > G_MIRROR_VERSION) {
2588 		G_MIRROR_DEBUG(0,
2589 		    "Kernel module is too old to handle metadata from %s.",
2590 		    cp->provider->name);
2591 		return (EINVAL);
2592 	}
2593 	if (error != 0) {
2594 		G_MIRROR_DEBUG(1, "MD5 metadata hash mismatch for provider %s.",
2595 		    cp->provider->name);
2596 		return (error);
2597 	}
2598 
2599 	return (0);
2600 }
2601 
2602 static int
2603 g_mirror_check_metadata(struct g_mirror_softc *sc, struct g_provider *pp,
2604     struct g_mirror_metadata *md)
2605 {
2606 
2607 	if (g_mirror_id2disk(sc, md->md_did) != NULL) {
2608 		G_MIRROR_DEBUG(1, "Disk %s (id=%u) already exists, skipping.",
2609 		    pp->name, md->md_did);
2610 		return (EEXIST);
2611 	}
2612 	if (md->md_all != sc->sc_ndisks) {
2613 		G_MIRROR_DEBUG(1,
2614 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2615 		    "md_all", pp->name, sc->sc_name);
2616 		return (EINVAL);
2617 	}
2618 	if (md->md_slice != sc->sc_slice) {
2619 		G_MIRROR_DEBUG(1,
2620 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2621 		    "md_slice", pp->name, sc->sc_name);
2622 		return (EINVAL);
2623 	}
2624 	if (md->md_balance != sc->sc_balance) {
2625 		G_MIRROR_DEBUG(1,
2626 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2627 		    "md_balance", pp->name, sc->sc_name);
2628 		return (EINVAL);
2629 	}
2630 	if (md->md_mediasize != sc->sc_mediasize) {
2631 		G_MIRROR_DEBUG(1,
2632 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2633 		    "md_mediasize", pp->name, sc->sc_name);
2634 		return (EINVAL);
2635 	}
2636 	if (sc->sc_mediasize > pp->mediasize) {
2637 		G_MIRROR_DEBUG(1,
2638 		    "Invalid size of disk %s (device %s), skipping.", pp->name,
2639 		    sc->sc_name);
2640 		return (EINVAL);
2641 	}
2642 	if (md->md_sectorsize != sc->sc_sectorsize) {
2643 		G_MIRROR_DEBUG(1,
2644 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2645 		    "md_sectorsize", pp->name, sc->sc_name);
2646 		return (EINVAL);
2647 	}
2648 	if ((sc->sc_sectorsize % pp->sectorsize) != 0) {
2649 		G_MIRROR_DEBUG(1,
2650 		    "Invalid sector size of disk %s (device %s), skipping.",
2651 		    pp->name, sc->sc_name);
2652 		return (EINVAL);
2653 	}
2654 	if ((md->md_mflags & ~G_MIRROR_DEVICE_FLAG_MASK) != 0) {
2655 		G_MIRROR_DEBUG(1,
2656 		    "Invalid device flags on disk %s (device %s), skipping.",
2657 		    pp->name, sc->sc_name);
2658 		return (EINVAL);
2659 	}
2660 	if ((md->md_dflags & ~G_MIRROR_DISK_FLAG_MASK) != 0) {
2661 		G_MIRROR_DEBUG(1,
2662 		    "Invalid disk flags on disk %s (device %s), skipping.",
2663 		    pp->name, sc->sc_name);
2664 		return (EINVAL);
2665 	}
2666 	return (0);
2667 }
2668 
2669 int
2670 g_mirror_add_disk(struct g_mirror_softc *sc, struct g_provider *pp,
2671     struct g_mirror_metadata *md)
2672 {
2673 	struct g_mirror_disk *disk;
2674 	int error;
2675 
2676 	g_topology_assert_not();
2677 	G_MIRROR_DEBUG(2, "Adding disk %s.", pp->name);
2678 
2679 	error = g_mirror_check_metadata(sc, pp, md);
2680 	if (error != 0)
2681 		return (error);
2682 	if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING &&
2683 	    md->md_genid < sc->sc_genid) {
2684 		G_MIRROR_DEBUG(0, "Component %s (device %s) broken, skipping.",
2685 		    pp->name, sc->sc_name);
2686 		return (EINVAL);
2687 	}
2688 	disk = g_mirror_init_disk(sc, pp, md, &error);
2689 	if (disk == NULL)
2690 		return (error);
2691 	error = g_mirror_event_send(disk, G_MIRROR_DISK_STATE_NEW,
2692 	    G_MIRROR_EVENT_WAIT);
2693 	if (error != 0)
2694 		return (error);
2695 	if (md->md_version < G_MIRROR_VERSION) {
2696 		G_MIRROR_DEBUG(0, "Upgrading metadata on %s (v%d->v%d).",
2697 		    pp->name, md->md_version, G_MIRROR_VERSION);
2698 		g_mirror_update_metadata(disk);
2699 	}
2700 	return (0);
2701 }
2702 
2703 static void
2704 g_mirror_destroy_delayed(void *arg, int flag)
2705 {
2706 	struct g_mirror_softc *sc;
2707 	int error;
2708 
2709 	if (flag == EV_CANCEL) {
2710 		G_MIRROR_DEBUG(1, "Destroying canceled.");
2711 		return;
2712 	}
2713 	sc = arg;
2714 	g_topology_unlock();
2715 	sx_xlock(&sc->sc_lock);
2716 	KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) == 0,
2717 	    ("DESTROY flag set on %s.", sc->sc_name));
2718 	KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROYING) != 0,
2719 	    ("DESTROYING flag not set on %s.", sc->sc_name));
2720 	G_MIRROR_DEBUG(1, "Destroying %s (delayed).", sc->sc_name);
2721 	error = g_mirror_destroy(sc, G_MIRROR_DESTROY_SOFT);
2722 	if (error != 0) {
2723 		G_MIRROR_DEBUG(0, "Cannot destroy %s.", sc->sc_name);
2724 		sx_xunlock(&sc->sc_lock);
2725 	}
2726 	g_topology_lock();
2727 }
2728 
2729 static int
2730 g_mirror_access(struct g_provider *pp, int acr, int acw, int ace)
2731 {
2732 	struct g_mirror_softc *sc;
2733 	int dcr, dcw, dce, error = 0;
2734 
2735 	g_topology_assert();
2736 	G_MIRROR_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, acr,
2737 	    acw, ace);
2738 
2739 	sc = pp->geom->softc;
2740 	if (sc == NULL && acr <= 0 && acw <= 0 && ace <= 0)
2741 		return (0);
2742 	KASSERT(sc != NULL, ("NULL softc (provider=%s).", pp->name));
2743 
2744 	dcr = pp->acr + acr;
2745 	dcw = pp->acw + acw;
2746 	dce = pp->ace + ace;
2747 
2748 	g_topology_unlock();
2749 	sx_xlock(&sc->sc_lock);
2750 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0 ||
2751 	    LIST_EMPTY(&sc->sc_disks)) {
2752 		if (acr > 0 || acw > 0 || ace > 0)
2753 			error = ENXIO;
2754 		goto end;
2755 	}
2756 	if (dcw == 0 && !sc->sc_idle)
2757 		g_mirror_idle(sc, dcw);
2758 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROYING) != 0) {
2759 		if (acr > 0 || acw > 0 || ace > 0) {
2760 			error = ENXIO;
2761 			goto end;
2762 		}
2763 		if (dcr == 0 && dcw == 0 && dce == 0) {
2764 			g_post_event(g_mirror_destroy_delayed, sc, M_WAITOK,
2765 			    sc, NULL);
2766 		}
2767 	}
2768 end:
2769 	sx_xunlock(&sc->sc_lock);
2770 	g_topology_lock();
2771 	return (error);
2772 }
2773 
2774 static struct g_geom *
2775 g_mirror_create(struct g_class *mp, const struct g_mirror_metadata *md)
2776 {
2777 	struct g_mirror_softc *sc;
2778 	struct g_geom *gp;
2779 	int error, timeout;
2780 
2781 	g_topology_assert();
2782 	G_MIRROR_DEBUG(1, "Creating device %s (id=%u).", md->md_name,
2783 	    md->md_mid);
2784 
2785 	/* One disk is minimum. */
2786 	if (md->md_all < 1)
2787 		return (NULL);
2788 	/*
2789 	 * Action geom.
2790 	 */
2791 	gp = g_new_geomf(mp, "%s", md->md_name);
2792 	sc = malloc(sizeof(*sc), M_MIRROR, M_WAITOK | M_ZERO);
2793 	gp->start = g_mirror_start;
2794 	gp->orphan = g_mirror_orphan;
2795 	gp->access = g_mirror_access;
2796 	gp->dumpconf = g_mirror_dumpconf;
2797 
2798 	sc->sc_id = md->md_mid;
2799 	sc->sc_slice = md->md_slice;
2800 	sc->sc_balance = md->md_balance;
2801 	sc->sc_mediasize = md->md_mediasize;
2802 	sc->sc_sectorsize = md->md_sectorsize;
2803 	sc->sc_ndisks = md->md_all;
2804 	sc->sc_flags = md->md_mflags;
2805 	sc->sc_bump_id = 0;
2806 	sc->sc_idle = 1;
2807 	sc->sc_last_write = time_uptime;
2808 	sc->sc_writes = 0;
2809 	sx_init(&sc->sc_lock, "gmirror:lock");
2810 	bioq_init(&sc->sc_queue);
2811 	mtx_init(&sc->sc_queue_mtx, "gmirror:queue", NULL, MTX_DEF);
2812 	bioq_init(&sc->sc_regular_delayed);
2813 	bioq_init(&sc->sc_inflight);
2814 	bioq_init(&sc->sc_sync_delayed);
2815 	LIST_INIT(&sc->sc_disks);
2816 	TAILQ_INIT(&sc->sc_events);
2817 	mtx_init(&sc->sc_events_mtx, "gmirror:events", NULL, MTX_DEF);
2818 	callout_init(&sc->sc_callout, CALLOUT_MPSAFE);
2819 	sc->sc_state = G_MIRROR_DEVICE_STATE_STARTING;
2820 	gp->softc = sc;
2821 	sc->sc_geom = gp;
2822 	sc->sc_provider = NULL;
2823 	/*
2824 	 * Synchronization geom.
2825 	 */
2826 	gp = g_new_geomf(mp, "%s.sync", md->md_name);
2827 	gp->softc = sc;
2828 	gp->orphan = g_mirror_orphan;
2829 	sc->sc_sync.ds_geom = gp;
2830 	sc->sc_sync.ds_ndisks = 0;
2831 	error = kthread_create(g_mirror_worker, sc, &sc->sc_worker, 0, 0,
2832 	    "g_mirror %s", md->md_name);
2833 	if (error != 0) {
2834 		G_MIRROR_DEBUG(1, "Cannot create kernel thread for %s.",
2835 		    sc->sc_name);
2836 		g_destroy_geom(sc->sc_sync.ds_geom);
2837 		mtx_destroy(&sc->sc_events_mtx);
2838 		mtx_destroy(&sc->sc_queue_mtx);
2839 		sx_destroy(&sc->sc_lock);
2840 		g_destroy_geom(sc->sc_geom);
2841 		free(sc, M_MIRROR);
2842 		return (NULL);
2843 	}
2844 
2845 	G_MIRROR_DEBUG(0, "Device %s created (id=%u).", sc->sc_name, sc->sc_id);
2846 
2847 	sc->sc_rootmount = root_mount_hold("GMIRROR");
2848 	G_MIRROR_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
2849 	/*
2850 	 * Run timeout.
2851 	 */
2852 	timeout = g_mirror_timeout * hz;
2853 	callout_reset(&sc->sc_callout, timeout, g_mirror_go, sc);
2854 	return (sc->sc_geom);
2855 }
2856 
2857 int
2858 g_mirror_destroy(struct g_mirror_softc *sc, int how)
2859 {
2860 	struct g_mirror_disk *disk;
2861 	struct g_provider *pp;
2862 
2863 	g_topology_assert_not();
2864 	if (sc == NULL)
2865 		return (ENXIO);
2866 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2867 
2868 	pp = sc->sc_provider;
2869 	if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
2870 		switch (how) {
2871 		case G_MIRROR_DESTROY_SOFT:
2872 			G_MIRROR_DEBUG(1,
2873 			    "Device %s is still open (r%dw%de%d).", pp->name,
2874 			    pp->acr, pp->acw, pp->ace);
2875 			return (EBUSY);
2876 		case G_MIRROR_DESTROY_DELAYED:
2877 			G_MIRROR_DEBUG(1,
2878 			    "Device %s will be destroyed on last close.",
2879 			    pp->name);
2880 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2881 				if (disk->d_state ==
2882 				    G_MIRROR_DISK_STATE_SYNCHRONIZING) {
2883 					g_mirror_sync_stop(disk, 1);
2884 				}
2885 			}
2886 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROYING;
2887 			return (EBUSY);
2888 		case G_MIRROR_DESTROY_HARD:
2889 			G_MIRROR_DEBUG(1, "Device %s is still open, so it "
2890 			    "can't be definitely removed.", pp->name);
2891 		}
2892 	}
2893 
2894 	g_topology_lock();
2895 	if (sc->sc_geom->softc == NULL) {
2896 		g_topology_unlock();
2897 		return (0);
2898 	}
2899 	sc->sc_geom->softc = NULL;
2900 	sc->sc_sync.ds_geom->softc = NULL;
2901 	g_topology_unlock();
2902 
2903 	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2904 	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_WAIT;
2905 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
2906 	sx_xunlock(&sc->sc_lock);
2907 	mtx_lock(&sc->sc_queue_mtx);
2908 	wakeup(sc);
2909 	mtx_unlock(&sc->sc_queue_mtx);
2910 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
2911 	while (sc->sc_worker != NULL)
2912 		tsleep(&sc->sc_worker, PRIBIO, "m:destroy", hz / 5);
2913 	G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
2914 	sx_xlock(&sc->sc_lock);
2915 	g_mirror_destroy_device(sc);
2916 	free(sc, M_MIRROR);
2917 	return (0);
2918 }
2919 
2920 static void
2921 g_mirror_taste_orphan(struct g_consumer *cp)
2922 {
2923 
2924 	KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
2925 	    cp->provider->name));
2926 }
2927 
2928 static struct g_geom *
2929 g_mirror_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
2930 {
2931 	struct g_mirror_metadata md;
2932 	struct g_mirror_softc *sc;
2933 	struct g_consumer *cp;
2934 	struct g_geom *gp;
2935 	int error;
2936 
2937 	g_topology_assert();
2938 	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
2939 	G_MIRROR_DEBUG(2, "Tasting %s.", pp->name);
2940 
2941 	gp = g_new_geomf(mp, "mirror:taste");
2942 	/*
2943 	 * This orphan function should be never called.
2944 	 */
2945 	gp->orphan = g_mirror_taste_orphan;
2946 	cp = g_new_consumer(gp);
2947 	g_attach(cp, pp);
2948 	error = g_mirror_read_metadata(cp, &md);
2949 	g_detach(cp);
2950 	g_destroy_consumer(cp);
2951 	g_destroy_geom(gp);
2952 	if (error != 0)
2953 		return (NULL);
2954 	gp = NULL;
2955 
2956 	if (md.md_provider[0] != '\0' && strcmp(md.md_provider, pp->name) != 0)
2957 		return (NULL);
2958 	if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
2959 		return (NULL);
2960 	if ((md.md_dflags & G_MIRROR_DISK_FLAG_INACTIVE) != 0) {
2961 		G_MIRROR_DEBUG(0,
2962 		    "Device %s: provider %s marked as inactive, skipping.",
2963 		    md.md_name, pp->name);
2964 		return (NULL);
2965 	}
2966 	if (g_mirror_debug >= 2)
2967 		mirror_metadata_dump(&md);
2968 
2969 	/*
2970 	 * Let's check if device already exists.
2971 	 */
2972 	sc = NULL;
2973 	LIST_FOREACH(gp, &mp->geom, geom) {
2974 		sc = gp->softc;
2975 		if (sc == NULL)
2976 			continue;
2977 		if (sc->sc_sync.ds_geom == gp)
2978 			continue;
2979 		if (strcmp(md.md_name, sc->sc_name) != 0)
2980 			continue;
2981 		if (md.md_mid != sc->sc_id) {
2982 			G_MIRROR_DEBUG(0, "Device %s already configured.",
2983 			    sc->sc_name);
2984 			return (NULL);
2985 		}
2986 		break;
2987 	}
2988 	if (gp == NULL) {
2989 		gp = g_mirror_create(mp, &md);
2990 		if (gp == NULL) {
2991 			G_MIRROR_DEBUG(0, "Cannot create device %s.",
2992 			    md.md_name);
2993 			return (NULL);
2994 		}
2995 		sc = gp->softc;
2996 	}
2997 	G_MIRROR_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
2998 	g_topology_unlock();
2999 	sx_xlock(&sc->sc_lock);
3000 	error = g_mirror_add_disk(sc, pp, &md);
3001 	if (error != 0) {
3002 		G_MIRROR_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
3003 		    pp->name, gp->name, error);
3004 		if (LIST_EMPTY(&sc->sc_disks)) {
3005 			g_cancel_event(sc);
3006 			g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3007 			g_topology_lock();
3008 			return (NULL);
3009 		}
3010 		gp = NULL;
3011 	}
3012 	sx_xunlock(&sc->sc_lock);
3013 	g_topology_lock();
3014 	return (gp);
3015 }
3016 
3017 static int
3018 g_mirror_destroy_geom(struct gctl_req *req __unused,
3019     struct g_class *mp __unused, struct g_geom *gp)
3020 {
3021 	struct g_mirror_softc *sc;
3022 	int error;
3023 
3024 	g_topology_unlock();
3025 	sc = gp->softc;
3026 	sx_xlock(&sc->sc_lock);
3027 	g_cancel_event(sc);
3028 	error = g_mirror_destroy(gp->softc, G_MIRROR_DESTROY_SOFT);
3029 	if (error != 0)
3030 		sx_xunlock(&sc->sc_lock);
3031 	g_topology_lock();
3032 	return (error);
3033 }
3034 
3035 static void
3036 g_mirror_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
3037     struct g_consumer *cp, struct g_provider *pp)
3038 {
3039 	struct g_mirror_softc *sc;
3040 
3041 	g_topology_assert();
3042 
3043 	sc = gp->softc;
3044 	if (sc == NULL)
3045 		return;
3046 	/* Skip synchronization geom. */
3047 	if (gp == sc->sc_sync.ds_geom)
3048 		return;
3049 	if (pp != NULL) {
3050 		/* Nothing here. */
3051 	} else if (cp != NULL) {
3052 		struct g_mirror_disk *disk;
3053 
3054 		disk = cp->private;
3055 		if (disk == NULL)
3056 			return;
3057 		g_topology_unlock();
3058 		sx_xlock(&sc->sc_lock);
3059 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)disk->d_id);
3060 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
3061 			sbuf_printf(sb, "%s<Synchronized>", indent);
3062 			if (disk->d_sync.ds_offset == 0)
3063 				sbuf_printf(sb, "0%%");
3064 			else {
3065 				sbuf_printf(sb, "%u%%",
3066 				    (u_int)((disk->d_sync.ds_offset * 100) /
3067 				    sc->sc_provider->mediasize));
3068 			}
3069 			sbuf_printf(sb, "</Synchronized>\n");
3070 		}
3071 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent,
3072 		    disk->d_sync.ds_syncid);
3073 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent,
3074 		    disk->d_genid);
3075 		sbuf_printf(sb, "%s<Flags>", indent);
3076 		if (disk->d_flags == 0)
3077 			sbuf_printf(sb, "NONE");
3078 		else {
3079 			int first = 1;
3080 
3081 #define	ADD_FLAG(flag, name)	do {					\
3082 	if ((disk->d_flags & (flag)) != 0) {				\
3083 		if (!first)						\
3084 			sbuf_printf(sb, ", ");				\
3085 		else							\
3086 			first = 0;					\
3087 		sbuf_printf(sb, name);					\
3088 	}								\
3089 } while (0)
3090 			ADD_FLAG(G_MIRROR_DISK_FLAG_DIRTY, "DIRTY");
3091 			ADD_FLAG(G_MIRROR_DISK_FLAG_HARDCODED, "HARDCODED");
3092 			ADD_FLAG(G_MIRROR_DISK_FLAG_INACTIVE, "INACTIVE");
3093 			ADD_FLAG(G_MIRROR_DISK_FLAG_SYNCHRONIZING,
3094 			    "SYNCHRONIZING");
3095 			ADD_FLAG(G_MIRROR_DISK_FLAG_FORCE_SYNC, "FORCE_SYNC");
3096 			ADD_FLAG(G_MIRROR_DISK_FLAG_BROKEN, "BROKEN");
3097 #undef	ADD_FLAG
3098 		}
3099 		sbuf_printf(sb, "</Flags>\n");
3100 		sbuf_printf(sb, "%s<Priority>%u</Priority>\n", indent,
3101 		    disk->d_priority);
3102 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3103 		    g_mirror_disk_state2str(disk->d_state));
3104 		sx_xunlock(&sc->sc_lock);
3105 		g_topology_lock();
3106 	} else {
3107 		g_topology_unlock();
3108 		sx_xlock(&sc->sc_lock);
3109 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
3110 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent, sc->sc_syncid);
3111 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, sc->sc_genid);
3112 		sbuf_printf(sb, "%s<Flags>", indent);
3113 		if (sc->sc_flags == 0)
3114 			sbuf_printf(sb, "NONE");
3115 		else {
3116 			int first = 1;
3117 
3118 #define	ADD_FLAG(flag, name)	do {					\
3119 	if ((sc->sc_flags & (flag)) != 0) {				\
3120 		if (!first)						\
3121 			sbuf_printf(sb, ", ");				\
3122 		else							\
3123 			first = 0;					\
3124 		sbuf_printf(sb, name);					\
3125 	}								\
3126 } while (0)
3127 			ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOAUTOSYNC, "NOAUTOSYNC");
3128 #undef	ADD_FLAG
3129 		}
3130 		sbuf_printf(sb, "</Flags>\n");
3131 		sbuf_printf(sb, "%s<Slice>%u</Slice>\n", indent,
3132 		    (u_int)sc->sc_slice);
3133 		sbuf_printf(sb, "%s<Balance>%s</Balance>\n", indent,
3134 		    balance_name(sc->sc_balance));
3135 		sbuf_printf(sb, "%s<Components>%u</Components>\n", indent,
3136 		    sc->sc_ndisks);
3137 		sbuf_printf(sb, "%s<State>", indent);
3138 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
3139 			sbuf_printf(sb, "%s", "STARTING");
3140 		else if (sc->sc_ndisks ==
3141 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE))
3142 			sbuf_printf(sb, "%s", "COMPLETE");
3143 		else
3144 			sbuf_printf(sb, "%s", "DEGRADED");
3145 		sbuf_printf(sb, "</State>\n");
3146 		sx_xunlock(&sc->sc_lock);
3147 		g_topology_lock();
3148 	}
3149 }
3150 
3151 static void
3152 g_mirror_shutdown_pre_sync(void *arg, int howto)
3153 {
3154 	struct g_class *mp;
3155 	struct g_geom *gp, *gp2;
3156 	struct g_mirror_softc *sc;
3157 	int error;
3158 
3159 	mp = arg;
3160 	DROP_GIANT();
3161 	g_topology_lock();
3162 	LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
3163 		if ((sc = gp->softc) == NULL)
3164 			continue;
3165 		/* Skip synchronization geom. */
3166 		if (gp == sc->sc_sync.ds_geom)
3167 			continue;
3168 		g_topology_unlock();
3169 		sx_xlock(&sc->sc_lock);
3170 		g_cancel_event(sc);
3171 		error = g_mirror_destroy(sc, G_MIRROR_DESTROY_DELAYED);
3172 		if (error != 0)
3173 			sx_xunlock(&sc->sc_lock);
3174 		g_topology_lock();
3175 	}
3176 	g_topology_unlock();
3177 	PICKUP_GIANT();
3178 }
3179 
3180 static void
3181 g_mirror_init(struct g_class *mp)
3182 {
3183 
3184 	g_mirror_pre_sync = EVENTHANDLER_REGISTER(shutdown_pre_sync,
3185 	    g_mirror_shutdown_pre_sync, mp, SHUTDOWN_PRI_FIRST);
3186 	if (g_mirror_pre_sync == NULL)
3187 		G_MIRROR_DEBUG(0, "Warning! Cannot register shutdown event.");
3188 }
3189 
3190 static void
3191 g_mirror_fini(struct g_class *mp)
3192 {
3193 
3194 	if (g_mirror_pre_sync != NULL)
3195 		EVENTHANDLER_DEREGISTER(shutdown_pre_sync, g_mirror_pre_sync);
3196 }
3197 
3198 DECLARE_GEOM_CLASS(g_mirror_class, g_mirror);
3199