xref: /freebsd/sys/geom/mirror/g_mirror.c (revision cc68614da8232d8baaca0ae0d0dd8f890f06623e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2004-2006 Pawel Jakub Dawidek <pjd@FreeBSD.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/bio.h>
35 #include <sys/eventhandler.h>
36 #include <sys/fail.h>
37 #include <sys/kernel.h>
38 #include <sys/kthread.h>
39 #include <sys/limits.h>
40 #include <sys/lock.h>
41 #include <sys/malloc.h>
42 #include <sys/mutex.h>
43 #include <sys/proc.h>
44 #include <sys/sbuf.h>
45 #include <sys/sched.h>
46 #include <sys/sx.h>
47 #include <sys/sysctl.h>
48 
49 #include <geom/geom.h>
50 #include <geom/geom_dbg.h>
51 #include <geom/mirror/g_mirror.h>
52 
53 FEATURE(geom_mirror, "GEOM mirroring support");
54 
55 static MALLOC_DEFINE(M_MIRROR, "mirror_data", "GEOM_MIRROR Data");
56 
57 SYSCTL_DECL(_kern_geom);
58 static SYSCTL_NODE(_kern_geom, OID_AUTO, mirror, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
59     "GEOM_MIRROR stuff");
60 int g_mirror_debug = 0;
61 SYSCTL_INT(_kern_geom_mirror, OID_AUTO, debug, CTLFLAG_RWTUN, &g_mirror_debug, 0,
62     "Debug level");
63 bool g_launch_mirror_before_timeout = true;
64 SYSCTL_BOOL(_kern_geom_mirror, OID_AUTO, launch_mirror_before_timeout,
65     CTLFLAG_RWTUN, &g_launch_mirror_before_timeout, 0,
66     "If false, force gmirror to wait out the full kern.geom.mirror.timeout "
67     "before launching mirrors");
68 static u_int g_mirror_timeout = 4;
69 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, timeout, CTLFLAG_RWTUN, &g_mirror_timeout,
70     0, "Time to wait on all mirror components");
71 static u_int g_mirror_idletime = 5;
72 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, idletime, CTLFLAG_RWTUN,
73     &g_mirror_idletime, 0, "Mark components as clean when idling");
74 static u_int g_mirror_disconnect_on_failure = 1;
75 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, disconnect_on_failure, CTLFLAG_RWTUN,
76     &g_mirror_disconnect_on_failure, 0, "Disconnect component on I/O failure.");
77 static u_int g_mirror_syncreqs = 2;
78 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, sync_requests, CTLFLAG_RDTUN,
79     &g_mirror_syncreqs, 0, "Parallel synchronization I/O requests.");
80 static u_int g_mirror_sync_period = 5;
81 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, sync_update_period, CTLFLAG_RWTUN,
82     &g_mirror_sync_period, 0,
83     "Metadata update period during synchronization, in seconds");
84 
85 #define	MSLEEP(ident, mtx, priority, wmesg, timeout)	do {		\
86 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, (ident));	\
87 	msleep((ident), (mtx), (priority), (wmesg), (timeout));		\
88 	G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, (ident));	\
89 } while (0)
90 
91 static eventhandler_tag g_mirror_post_sync = NULL;
92 static int g_mirror_shutdown = 0;
93 
94 static g_ctl_destroy_geom_t g_mirror_destroy_geom;
95 static g_taste_t g_mirror_taste;
96 static g_init_t g_mirror_init;
97 static g_fini_t g_mirror_fini;
98 static g_provgone_t g_mirror_providergone;
99 static g_resize_t g_mirror_resize;
100 
101 struct g_class g_mirror_class = {
102 	.name = G_MIRROR_CLASS_NAME,
103 	.version = G_VERSION,
104 	.ctlreq = g_mirror_config,
105 	.taste = g_mirror_taste,
106 	.destroy_geom = g_mirror_destroy_geom,
107 	.init = g_mirror_init,
108 	.fini = g_mirror_fini,
109 	.providergone = g_mirror_providergone,
110 	.resize = g_mirror_resize
111 };
112 
113 static void g_mirror_destroy_provider(struct g_mirror_softc *sc);
114 static int g_mirror_update_disk(struct g_mirror_disk *disk, u_int state);
115 static void g_mirror_update_device(struct g_mirror_softc *sc, bool force);
116 static void g_mirror_dumpconf(struct sbuf *sb, const char *indent,
117     struct g_geom *gp, struct g_consumer *cp, struct g_provider *pp);
118 static void g_mirror_timeout_drain(struct g_mirror_softc *sc);
119 static int g_mirror_refresh_device(struct g_mirror_softc *sc,
120     const struct g_provider *pp, const struct g_mirror_metadata *md);
121 static void g_mirror_sync_reinit(const struct g_mirror_disk *disk,
122     struct bio *bp, off_t offset);
123 static void g_mirror_sync_stop(struct g_mirror_disk *disk, int type);
124 static void g_mirror_register_request(struct g_mirror_softc *sc,
125     struct bio *bp);
126 static void g_mirror_sync_release(struct g_mirror_softc *sc);
127 
128 static const char *
129 g_mirror_disk_state2str(int state)
130 {
131 
132 	switch (state) {
133 	case G_MIRROR_DISK_STATE_NONE:
134 		return ("NONE");
135 	case G_MIRROR_DISK_STATE_NEW:
136 		return ("NEW");
137 	case G_MIRROR_DISK_STATE_ACTIVE:
138 		return ("ACTIVE");
139 	case G_MIRROR_DISK_STATE_STALE:
140 		return ("STALE");
141 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
142 		return ("SYNCHRONIZING");
143 	case G_MIRROR_DISK_STATE_DISCONNECTED:
144 		return ("DISCONNECTED");
145 	case G_MIRROR_DISK_STATE_DESTROY:
146 		return ("DESTROY");
147 	default:
148 		return ("INVALID");
149 	}
150 }
151 
152 static const char *
153 g_mirror_device_state2str(int state)
154 {
155 
156 	switch (state) {
157 	case G_MIRROR_DEVICE_STATE_STARTING:
158 		return ("STARTING");
159 	case G_MIRROR_DEVICE_STATE_RUNNING:
160 		return ("RUNNING");
161 	default:
162 		return ("INVALID");
163 	}
164 }
165 
166 static const char *
167 g_mirror_get_diskname(struct g_mirror_disk *disk)
168 {
169 
170 	if (disk->d_consumer == NULL || disk->d_consumer->provider == NULL)
171 		return ("[unknown]");
172 	return (disk->d_name);
173 }
174 
175 /*
176  * --- Events handling functions ---
177  * Events in geom_mirror are used to maintain disks and device status
178  * from one thread to simplify locking.
179  */
180 static void
181 g_mirror_event_free(struct g_mirror_event *ep)
182 {
183 
184 	free(ep, M_MIRROR);
185 }
186 
187 static int
188 g_mirror_event_dispatch(struct g_mirror_event *ep, void *arg, int state,
189     int flags)
190 {
191 	struct g_mirror_softc *sc;
192 	struct g_mirror_disk *disk;
193 	int error;
194 
195 	G_MIRROR_DEBUG(4, "%s: Sending event %p.", __func__, ep);
196 	if ((flags & G_MIRROR_EVENT_DEVICE) != 0) {
197 		disk = NULL;
198 		sc = arg;
199 	} else {
200 		disk = arg;
201 		sc = disk->d_softc;
202 	}
203 	ep->e_disk = disk;
204 	ep->e_state = state;
205 	ep->e_flags = flags;
206 	ep->e_error = 0;
207 	mtx_lock(&sc->sc_events_mtx);
208 	TAILQ_INSERT_TAIL(&sc->sc_events, ep, e_next);
209 	mtx_unlock(&sc->sc_events_mtx);
210 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
211 	mtx_lock(&sc->sc_queue_mtx);
212 	wakeup(sc);
213 	mtx_unlock(&sc->sc_queue_mtx);
214 	if ((flags & G_MIRROR_EVENT_DONTWAIT) != 0)
215 		return (0);
216 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, ep);
217 	sx_xunlock(&sc->sc_lock);
218 	while ((ep->e_flags & G_MIRROR_EVENT_DONE) == 0) {
219 		mtx_lock(&sc->sc_events_mtx);
220 		MSLEEP(ep, &sc->sc_events_mtx, PRIBIO | PDROP, "m:event",
221 		    hz * 5);
222 	}
223 	error = ep->e_error;
224 	g_mirror_event_free(ep);
225 	sx_xlock(&sc->sc_lock);
226 	return (error);
227 }
228 
229 int
230 g_mirror_event_send(void *arg, int state, int flags)
231 {
232 	struct g_mirror_event *ep;
233 
234 	ep = malloc(sizeof(*ep), M_MIRROR, M_WAITOK);
235 	return (g_mirror_event_dispatch(ep, arg, state, flags));
236 }
237 
238 static struct g_mirror_event *
239 g_mirror_event_first(struct g_mirror_softc *sc)
240 {
241 	struct g_mirror_event *ep;
242 
243 	mtx_lock(&sc->sc_events_mtx);
244 	ep = TAILQ_FIRST(&sc->sc_events);
245 	mtx_unlock(&sc->sc_events_mtx);
246 	return (ep);
247 }
248 
249 static void
250 g_mirror_event_remove(struct g_mirror_softc *sc, struct g_mirror_event *ep)
251 {
252 
253 	mtx_lock(&sc->sc_events_mtx);
254 	TAILQ_REMOVE(&sc->sc_events, ep, e_next);
255 	mtx_unlock(&sc->sc_events_mtx);
256 }
257 
258 static void
259 g_mirror_event_cancel(struct g_mirror_disk *disk)
260 {
261 	struct g_mirror_softc *sc;
262 	struct g_mirror_event *ep, *tmpep;
263 
264 	sc = disk->d_softc;
265 	sx_assert(&sc->sc_lock, SX_XLOCKED);
266 
267 	mtx_lock(&sc->sc_events_mtx);
268 	TAILQ_FOREACH_SAFE(ep, &sc->sc_events, e_next, tmpep) {
269 		if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0)
270 			continue;
271 		if (ep->e_disk != disk)
272 			continue;
273 		TAILQ_REMOVE(&sc->sc_events, ep, e_next);
274 		if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
275 			g_mirror_event_free(ep);
276 		else {
277 			ep->e_error = ECANCELED;
278 			wakeup(ep);
279 		}
280 	}
281 	mtx_unlock(&sc->sc_events_mtx);
282 }
283 
284 /*
285  * Return the number of disks in given state.
286  * If state is equal to -1, count all connected disks.
287  */
288 u_int
289 g_mirror_ndisks(struct g_mirror_softc *sc, int state)
290 {
291 	struct g_mirror_disk *disk;
292 	u_int n = 0;
293 
294 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
295 		if (state == -1 || disk->d_state == state)
296 			n++;
297 	}
298 	return (n);
299 }
300 
301 /*
302  * Find a disk in mirror by its disk ID.
303  */
304 static struct g_mirror_disk *
305 g_mirror_id2disk(struct g_mirror_softc *sc, uint32_t id)
306 {
307 	struct g_mirror_disk *disk;
308 
309 	sx_assert(&sc->sc_lock, SX_XLOCKED);
310 
311 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
312 		if (disk->d_id == id)
313 			return (disk);
314 	}
315 	return (NULL);
316 }
317 
318 static u_int
319 g_mirror_nrequests(struct g_mirror_softc *sc, struct g_consumer *cp)
320 {
321 	struct bio *bp;
322 	u_int nreqs = 0;
323 
324 	mtx_lock(&sc->sc_queue_mtx);
325 	TAILQ_FOREACH(bp, &sc->sc_queue, bio_queue) {
326 		if (bp->bio_from == cp)
327 			nreqs++;
328 	}
329 	mtx_unlock(&sc->sc_queue_mtx);
330 	return (nreqs);
331 }
332 
333 static int
334 g_mirror_is_busy(struct g_mirror_softc *sc, struct g_consumer *cp)
335 {
336 
337 	if (cp->index > 0) {
338 		G_MIRROR_DEBUG(2,
339 		    "I/O requests for %s exist, can't destroy it now.",
340 		    cp->provider->name);
341 		return (1);
342 	}
343 	if (g_mirror_nrequests(sc, cp) > 0) {
344 		G_MIRROR_DEBUG(2,
345 		    "I/O requests for %s in queue, can't destroy it now.",
346 		    cp->provider->name);
347 		return (1);
348 	}
349 	return (0);
350 }
351 
352 static void
353 g_mirror_destroy_consumer(void *arg, int flags __unused)
354 {
355 	struct g_consumer *cp;
356 
357 	g_topology_assert();
358 
359 	cp = arg;
360 	G_MIRROR_DEBUG(1, "Consumer %s destroyed.", cp->provider->name);
361 	g_detach(cp);
362 	g_destroy_consumer(cp);
363 }
364 
365 static void
366 g_mirror_kill_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
367 {
368 	struct g_provider *pp;
369 	int retaste_wait;
370 
371 	g_topology_assert();
372 
373 	cp->private = NULL;
374 	if (g_mirror_is_busy(sc, cp))
375 		return;
376 	pp = cp->provider;
377 	retaste_wait = 0;
378 	if (cp->acw == 1) {
379 		if ((pp->geom->flags & G_GEOM_WITHER) == 0)
380 			retaste_wait = 1;
381 	}
382 	G_MIRROR_DEBUG(2, "Access %s r%dw%de%d = %d", pp->name, -cp->acr,
383 	    -cp->acw, -cp->ace, 0);
384 	if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
385 		g_access(cp, -cp->acr, -cp->acw, -cp->ace);
386 	if (retaste_wait) {
387 		/*
388 		 * After retaste event was send (inside g_access()), we can send
389 		 * event to detach and destroy consumer.
390 		 * A class, which has consumer to the given provider connected
391 		 * will not receive retaste event for the provider.
392 		 * This is the way how I ignore retaste events when I close
393 		 * consumers opened for write: I detach and destroy consumer
394 		 * after retaste event is sent.
395 		 */
396 		g_post_event(g_mirror_destroy_consumer, cp, M_WAITOK, NULL);
397 		return;
398 	}
399 	G_MIRROR_DEBUG(1, "Consumer %s destroyed.", pp->name);
400 	g_detach(cp);
401 	g_destroy_consumer(cp);
402 }
403 
404 static int
405 g_mirror_connect_disk(struct g_mirror_disk *disk, struct g_provider *pp)
406 {
407 	struct g_consumer *cp;
408 	int error;
409 
410 	g_topology_assert_not();
411 	KASSERT(disk->d_consumer == NULL,
412 	    ("Disk already connected (device %s).", disk->d_softc->sc_name));
413 
414 	g_topology_lock();
415 	cp = g_new_consumer(disk->d_softc->sc_geom);
416 	cp->flags |= G_CF_DIRECT_RECEIVE;
417 	error = g_attach(cp, pp);
418 	if (error != 0) {
419 		g_destroy_consumer(cp);
420 		g_topology_unlock();
421 		return (error);
422 	}
423 	error = g_access(cp, 1, 1, 1);
424 	if (error != 0) {
425 		g_detach(cp);
426 		g_destroy_consumer(cp);
427 		g_topology_unlock();
428 		G_MIRROR_DEBUG(0, "Cannot open consumer %s (error=%d).",
429 		    pp->name, error);
430 		return (error);
431 	}
432 	g_topology_unlock();
433 	disk->d_consumer = cp;
434 	disk->d_consumer->private = disk;
435 	disk->d_consumer->index = 0;
436 
437 	G_MIRROR_DEBUG(2, "Disk %s connected.", g_mirror_get_diskname(disk));
438 	return (0);
439 }
440 
441 static void
442 g_mirror_disconnect_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
443 {
444 
445 	g_topology_assert();
446 
447 	if (cp == NULL)
448 		return;
449 	if (cp->provider != NULL)
450 		g_mirror_kill_consumer(sc, cp);
451 	else
452 		g_destroy_consumer(cp);
453 }
454 
455 /*
456  * Initialize disk. This means allocate memory, create consumer, attach it
457  * to the provider and open access (r1w1e1) to it.
458  */
459 static struct g_mirror_disk *
460 g_mirror_init_disk(struct g_mirror_softc *sc, struct g_provider *pp,
461     struct g_mirror_metadata *md, int *errorp)
462 {
463 	struct g_mirror_disk *disk;
464 	int i, error;
465 
466 	disk = malloc(sizeof(*disk), M_MIRROR, M_NOWAIT | M_ZERO);
467 	if (disk == NULL) {
468 		error = ENOMEM;
469 		goto fail;
470 	}
471 	disk->d_softc = sc;
472 	error = g_mirror_connect_disk(disk, pp);
473 	if (error != 0)
474 		goto fail;
475 	disk->d_id = md->md_did;
476 	disk->d_state = G_MIRROR_DISK_STATE_NONE;
477 	disk->d_priority = md->md_priority;
478 	disk->d_flags = md->md_dflags;
479 	error = g_getattr("GEOM::candelete", disk->d_consumer, &i);
480 	if (error == 0 && i != 0)
481 		disk->d_flags |= G_MIRROR_DISK_FLAG_CANDELETE;
482 	if (md->md_provider[0] != '\0')
483 		disk->d_flags |= G_MIRROR_DISK_FLAG_HARDCODED;
484 	disk->d_sync.ds_consumer = NULL;
485 	disk->d_sync.ds_offset = md->md_sync_offset;
486 	disk->d_sync.ds_offset_done = md->md_sync_offset;
487 	disk->d_sync.ds_update_ts = time_uptime;
488 	disk->d_genid = md->md_genid;
489 	disk->d_sync.ds_syncid = md->md_syncid;
490 	disk->d_init_ndisks = md->md_all;
491 	disk->d_init_slice = md->md_slice;
492 	disk->d_init_balance = md->md_balance;
493 	disk->d_init_mediasize = md->md_mediasize;
494 	if (errorp != NULL)
495 		*errorp = 0;
496 	return (disk);
497 fail:
498 	if (errorp != NULL)
499 		*errorp = error;
500 	if (disk != NULL)
501 		free(disk, M_MIRROR);
502 	return (NULL);
503 }
504 
505 static void
506 g_mirror_destroy_disk(struct g_mirror_disk *disk)
507 {
508 	struct g_mirror_softc *sc;
509 
510 	g_topology_assert_not();
511 	sc = disk->d_softc;
512 	sx_assert(&sc->sc_lock, SX_XLOCKED);
513 
514 	g_topology_lock();
515 	LIST_REMOVE(disk, d_next);
516 	g_topology_unlock();
517 	g_mirror_event_cancel(disk);
518 	if (sc->sc_hint == disk)
519 		sc->sc_hint = NULL;
520 	switch (disk->d_state) {
521 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
522 		g_mirror_sync_stop(disk, 1);
523 		/* FALLTHROUGH */
524 	case G_MIRROR_DISK_STATE_NEW:
525 	case G_MIRROR_DISK_STATE_STALE:
526 	case G_MIRROR_DISK_STATE_ACTIVE:
527 		g_topology_lock();
528 		g_mirror_disconnect_consumer(sc, disk->d_consumer);
529 		g_topology_unlock();
530 		free(disk, M_MIRROR);
531 		break;
532 	default:
533 		KASSERT(0 == 1, ("Wrong disk state (%s, %s).",
534 		    g_mirror_get_diskname(disk),
535 		    g_mirror_disk_state2str(disk->d_state)));
536 	}
537 }
538 
539 static void
540 g_mirror_free_device(struct g_mirror_softc *sc)
541 {
542 
543 	g_topology_assert();
544 
545 	mtx_destroy(&sc->sc_queue_mtx);
546 	mtx_destroy(&sc->sc_events_mtx);
547 	mtx_destroy(&sc->sc_done_mtx);
548 	sx_destroy(&sc->sc_lock);
549 	free(sc, M_MIRROR);
550 }
551 
552 static void
553 g_mirror_providergone(struct g_provider *pp)
554 {
555 	struct g_mirror_softc *sc = pp->private;
556 
557 	if ((--sc->sc_refcnt) == 0)
558 		g_mirror_free_device(sc);
559 }
560 
561 static void
562 g_mirror_destroy_device(struct g_mirror_softc *sc)
563 {
564 	struct g_mirror_disk *disk;
565 	struct g_mirror_event *ep;
566 	struct g_geom *gp;
567 	struct g_consumer *cp, *tmpcp;
568 
569 	g_topology_assert_not();
570 	sx_assert(&sc->sc_lock, SX_XLOCKED);
571 
572 	gp = sc->sc_geom;
573 	if (sc->sc_provider != NULL)
574 		g_mirror_destroy_provider(sc);
575 	for (disk = LIST_FIRST(&sc->sc_disks); disk != NULL;
576 	    disk = LIST_FIRST(&sc->sc_disks)) {
577 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
578 		g_mirror_update_metadata(disk);
579 		g_mirror_destroy_disk(disk);
580 	}
581 	while ((ep = g_mirror_event_first(sc)) != NULL) {
582 		g_mirror_event_remove(sc, ep);
583 		if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
584 			g_mirror_event_free(ep);
585 		else {
586 			ep->e_error = ECANCELED;
587 			ep->e_flags |= G_MIRROR_EVENT_DONE;
588 			G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, ep);
589 			mtx_lock(&sc->sc_events_mtx);
590 			wakeup(ep);
591 			mtx_unlock(&sc->sc_events_mtx);
592 		}
593 	}
594 	g_mirror_timeout_drain(sc);
595 
596 	g_topology_lock();
597 	LIST_FOREACH_SAFE(cp, &sc->sc_sync.ds_geom->consumer, consumer, tmpcp) {
598 		g_mirror_disconnect_consumer(sc, cp);
599 	}
600 	g_wither_geom(sc->sc_sync.ds_geom, ENXIO);
601 	G_MIRROR_DEBUG(0, "Device %s destroyed.", gp->name);
602 	g_wither_geom(gp, ENXIO);
603 	sx_xunlock(&sc->sc_lock);
604 	if ((--sc->sc_refcnt) == 0)
605 		g_mirror_free_device(sc);
606 	g_topology_unlock();
607 }
608 
609 static void
610 g_mirror_orphan(struct g_consumer *cp)
611 {
612 	struct g_mirror_disk *disk;
613 
614 	g_topology_assert();
615 
616 	disk = cp->private;
617 	if (disk == NULL)
618 		return;
619 	disk->d_softc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
620 	g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
621 	    G_MIRROR_EVENT_DONTWAIT);
622 }
623 
624 /*
625  * Function should return the next active disk on the list.
626  * It is possible that it will be the same disk as given.
627  * If there are no active disks on list, NULL is returned.
628  */
629 static __inline struct g_mirror_disk *
630 g_mirror_find_next(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
631 {
632 	struct g_mirror_disk *dp;
633 
634 	for (dp = LIST_NEXT(disk, d_next); dp != disk;
635 	    dp = LIST_NEXT(dp, d_next)) {
636 		if (dp == NULL)
637 			dp = LIST_FIRST(&sc->sc_disks);
638 		if (dp->d_state == G_MIRROR_DISK_STATE_ACTIVE)
639 			break;
640 	}
641 	if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
642 		return (NULL);
643 	return (dp);
644 }
645 
646 static struct g_mirror_disk *
647 g_mirror_get_disk(struct g_mirror_softc *sc)
648 {
649 	struct g_mirror_disk *disk;
650 
651 	if (sc->sc_hint == NULL) {
652 		sc->sc_hint = LIST_FIRST(&sc->sc_disks);
653 		if (sc->sc_hint == NULL)
654 			return (NULL);
655 	}
656 	disk = sc->sc_hint;
657 	if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE) {
658 		disk = g_mirror_find_next(sc, disk);
659 		if (disk == NULL)
660 			return (NULL);
661 	}
662 	sc->sc_hint = g_mirror_find_next(sc, disk);
663 	return (disk);
664 }
665 
666 static int
667 g_mirror_write_metadata(struct g_mirror_disk *disk,
668     struct g_mirror_metadata *md)
669 {
670 	struct g_mirror_softc *sc;
671 	struct g_consumer *cp;
672 	off_t offset, length;
673 	u_char *sector;
674 	int error = 0;
675 
676 	g_topology_assert_not();
677 	sc = disk->d_softc;
678 	sx_assert(&sc->sc_lock, SX_LOCKED);
679 
680 	cp = disk->d_consumer;
681 	KASSERT(cp != NULL, ("NULL consumer (%s).", sc->sc_name));
682 	KASSERT(cp->provider != NULL, ("NULL provider (%s).", sc->sc_name));
683 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
684 	    ("Consumer %s closed? (r%dw%de%d).", cp->provider->name, cp->acr,
685 	    cp->acw, cp->ace));
686 	length = cp->provider->sectorsize;
687 	offset = cp->provider->mediasize - length;
688 	sector = malloc((size_t)length, M_MIRROR, M_WAITOK | M_ZERO);
689 	if (md != NULL &&
690 	    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_WIPE) == 0) {
691 		/*
692 		 * Handle the case, when the size of parent provider reduced.
693 		 */
694 		if (offset < md->md_mediasize)
695 			error = ENOSPC;
696 		else
697 			mirror_metadata_encode(md, sector);
698 	}
699 	KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_metadata_write, error);
700 	if (error == 0)
701 		error = g_write_data(cp, offset, sector, length);
702 	free(sector, M_MIRROR);
703 	if (error != 0) {
704 		if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
705 			disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
706 			G_MIRROR_DEBUG(0, "Cannot write metadata on %s "
707 			    "(device=%s, error=%d).",
708 			    g_mirror_get_diskname(disk), sc->sc_name, error);
709 		} else {
710 			G_MIRROR_DEBUG(1, "Cannot write metadata on %s "
711 			    "(device=%s, error=%d).",
712 			    g_mirror_get_diskname(disk), sc->sc_name, error);
713 		}
714 		if (g_mirror_disconnect_on_failure &&
715 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1) {
716 			sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
717 			g_mirror_event_send(disk,
718 			    G_MIRROR_DISK_STATE_DISCONNECTED,
719 			    G_MIRROR_EVENT_DONTWAIT);
720 		}
721 	}
722 	return (error);
723 }
724 
725 static int
726 g_mirror_clear_metadata(struct g_mirror_disk *disk)
727 {
728 	int error;
729 
730 	g_topology_assert_not();
731 	sx_assert(&disk->d_softc->sc_lock, SX_LOCKED);
732 
733 	if (disk->d_softc->sc_type != G_MIRROR_TYPE_AUTOMATIC)
734 		return (0);
735 	error = g_mirror_write_metadata(disk, NULL);
736 	if (error == 0) {
737 		G_MIRROR_DEBUG(2, "Metadata on %s cleared.",
738 		    g_mirror_get_diskname(disk));
739 	} else {
740 		G_MIRROR_DEBUG(0,
741 		    "Cannot clear metadata on disk %s (error=%d).",
742 		    g_mirror_get_diskname(disk), error);
743 	}
744 	return (error);
745 }
746 
747 void
748 g_mirror_fill_metadata(struct g_mirror_softc *sc, struct g_mirror_disk *disk,
749     struct g_mirror_metadata *md)
750 {
751 
752 	bzero(md, sizeof(*md));
753 	strlcpy(md->md_magic, G_MIRROR_MAGIC, sizeof(md->md_magic));
754 	md->md_version = G_MIRROR_VERSION;
755 	strlcpy(md->md_name, sc->sc_name, sizeof(md->md_name));
756 	md->md_mid = sc->sc_id;
757 	md->md_all = sc->sc_ndisks;
758 	md->md_slice = sc->sc_slice;
759 	md->md_balance = sc->sc_balance;
760 	md->md_genid = sc->sc_genid;
761 	md->md_mediasize = sc->sc_mediasize;
762 	md->md_sectorsize = sc->sc_sectorsize;
763 	md->md_mflags = (sc->sc_flags & G_MIRROR_DEVICE_FLAG_MASK);
764 	if (disk == NULL) {
765 		md->md_did = arc4random();
766 	} else {
767 		md->md_did = disk->d_id;
768 		md->md_priority = disk->d_priority;
769 		md->md_syncid = disk->d_sync.ds_syncid;
770 		md->md_dflags = (disk->d_flags & G_MIRROR_DISK_FLAG_MASK);
771 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
772 			md->md_sync_offset = disk->d_sync.ds_offset_done;
773 		if ((disk->d_flags & G_MIRROR_DISK_FLAG_HARDCODED) != 0) {
774 			strlcpy(md->md_provider,
775 			    disk->d_consumer->provider->name,
776 			    sizeof(md->md_provider));
777 		}
778 		md->md_provsize = disk->d_consumer->provider->mediasize;
779 	}
780 }
781 
782 void
783 g_mirror_update_metadata(struct g_mirror_disk *disk)
784 {
785 	struct g_mirror_softc *sc;
786 	struct g_mirror_metadata md;
787 	int error;
788 
789 	g_topology_assert_not();
790 	sc = disk->d_softc;
791 	sx_assert(&sc->sc_lock, SX_LOCKED);
792 
793 	if (sc->sc_type != G_MIRROR_TYPE_AUTOMATIC)
794 		return;
795 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_WIPE) == 0)
796 		g_mirror_fill_metadata(sc, disk, &md);
797 	error = g_mirror_write_metadata(disk, &md);
798 	if (error == 0) {
799 		G_MIRROR_DEBUG(2, "Metadata on %s updated.",
800 		    g_mirror_get_diskname(disk));
801 	} else {
802 		G_MIRROR_DEBUG(0,
803 		    "Cannot update metadata on disk %s (error=%d).",
804 		    g_mirror_get_diskname(disk), error);
805 	}
806 }
807 
808 static void
809 g_mirror_bump_syncid(struct g_mirror_softc *sc)
810 {
811 	struct g_mirror_disk *disk;
812 
813 	g_topology_assert_not();
814 	sx_assert(&sc->sc_lock, SX_XLOCKED);
815 	KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
816 	    ("%s called with no active disks (device=%s).", __func__,
817 	    sc->sc_name));
818 
819 	sc->sc_syncid++;
820 	G_MIRROR_DEBUG(1, "Device %s: syncid bumped to %u.", sc->sc_name,
821 	    sc->sc_syncid);
822 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
823 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
824 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
825 			disk->d_sync.ds_syncid = sc->sc_syncid;
826 			g_mirror_update_metadata(disk);
827 		}
828 	}
829 }
830 
831 static void
832 g_mirror_bump_genid(struct g_mirror_softc *sc)
833 {
834 	struct g_mirror_disk *disk;
835 
836 	g_topology_assert_not();
837 	sx_assert(&sc->sc_lock, SX_XLOCKED);
838 	KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
839 	    ("%s called with no active disks (device=%s).", __func__,
840 	    sc->sc_name));
841 
842 	sc->sc_genid++;
843 	G_MIRROR_DEBUG(1, "Device %s: genid bumped to %u.", sc->sc_name,
844 	    sc->sc_genid);
845 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
846 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
847 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
848 			disk->d_genid = sc->sc_genid;
849 			g_mirror_update_metadata(disk);
850 		}
851 	}
852 }
853 
854 static int
855 g_mirror_idle(struct g_mirror_softc *sc, int acw)
856 {
857 	struct g_mirror_disk *disk;
858 	int timeout;
859 
860 	g_topology_assert_not();
861 	sx_assert(&sc->sc_lock, SX_XLOCKED);
862 
863 	if (sc->sc_provider == NULL)
864 		return (0);
865 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
866 		return (0);
867 	if (sc->sc_idle)
868 		return (0);
869 	if (sc->sc_writes > 0)
870 		return (0);
871 	if (acw > 0 || (acw == -1 && sc->sc_provider->acw > 0)) {
872 		timeout = g_mirror_idletime - (time_uptime - sc->sc_last_write);
873 		if (!g_mirror_shutdown && timeout > 0)
874 			return (timeout);
875 	}
876 	sc->sc_idle = 1;
877 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
878 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
879 			continue;
880 		G_MIRROR_DEBUG(2, "Disk %s (device %s) marked as clean.",
881 		    g_mirror_get_diskname(disk), sc->sc_name);
882 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
883 		g_mirror_update_metadata(disk);
884 	}
885 	return (0);
886 }
887 
888 static void
889 g_mirror_unidle(struct g_mirror_softc *sc)
890 {
891 	struct g_mirror_disk *disk;
892 
893 	g_topology_assert_not();
894 	sx_assert(&sc->sc_lock, SX_XLOCKED);
895 
896 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
897 		return;
898 	sc->sc_idle = 0;
899 	sc->sc_last_write = time_uptime;
900 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
901 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
902 			continue;
903 		G_MIRROR_DEBUG(2, "Disk %s (device %s) marked as dirty.",
904 		    g_mirror_get_diskname(disk), sc->sc_name);
905 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
906 		g_mirror_update_metadata(disk);
907 	}
908 }
909 
910 static void
911 g_mirror_done(struct bio *bp)
912 {
913 	struct g_mirror_softc *sc;
914 
915 	sc = bp->bio_from->geom->softc;
916 	bp->bio_cflags = G_MIRROR_BIO_FLAG_REGULAR;
917 	mtx_lock(&sc->sc_queue_mtx);
918 	TAILQ_INSERT_TAIL(&sc->sc_queue, bp, bio_queue);
919 	mtx_unlock(&sc->sc_queue_mtx);
920 	wakeup(sc);
921 }
922 
923 static void
924 g_mirror_regular_request_error(struct g_mirror_softc *sc,
925     struct g_mirror_disk *disk, struct bio *bp)
926 {
927 
928 	if ((bp->bio_cmd == BIO_FLUSH || bp->bio_cmd == BIO_SPEEDUP) &&
929 	    bp->bio_error == EOPNOTSUPP)
930 		return;
931 
932 	if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
933 		disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
934 		G_MIRROR_LOGREQ(0, bp, "Request failed (error=%d).",
935 		    bp->bio_error);
936 	} else {
937 		G_MIRROR_LOGREQ(1, bp, "Request failed (error=%d).",
938 		    bp->bio_error);
939 	}
940 	if (g_mirror_disconnect_on_failure &&
941 	    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1) {
942 		if (bp->bio_error == ENXIO &&
943 		    bp->bio_cmd == BIO_READ)
944 			sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
945 		else if (bp->bio_error == ENXIO)
946 			sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID_NOW;
947 		else
948 			sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
949 		g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
950 		    G_MIRROR_EVENT_DONTWAIT);
951 	}
952 }
953 
954 static void
955 g_mirror_regular_request(struct g_mirror_softc *sc, struct bio *bp)
956 {
957 	struct g_mirror_disk *disk;
958 	struct bio *pbp;
959 
960 	g_topology_assert_not();
961 	KASSERT(sc->sc_provider == bp->bio_parent->bio_to,
962 	    ("regular request %p with unexpected origin", bp));
963 
964 	pbp = bp->bio_parent;
965 	bp->bio_from->index--;
966 	if (bp->bio_cmd == BIO_WRITE || bp->bio_cmd == BIO_DELETE)
967 		sc->sc_writes--;
968 	disk = bp->bio_from->private;
969 	if (disk == NULL) {
970 		g_topology_lock();
971 		g_mirror_kill_consumer(sc, bp->bio_from);
972 		g_topology_unlock();
973 	}
974 
975 	switch (bp->bio_cmd) {
976 	case BIO_READ:
977 		KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_read,
978 		    bp->bio_error);
979 		break;
980 	case BIO_WRITE:
981 		KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_write,
982 		    bp->bio_error);
983 		break;
984 	case BIO_DELETE:
985 		KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_delete,
986 		    bp->bio_error);
987 		break;
988 	case BIO_FLUSH:
989 		KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_flush,
990 		    bp->bio_error);
991 		break;
992 	case BIO_SPEEDUP:
993 		KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_speedup,
994 		    bp->bio_error);
995 		break;
996 	}
997 
998 	pbp->bio_inbed++;
999 	KASSERT(pbp->bio_inbed <= pbp->bio_children,
1000 	    ("bio_inbed (%u) is bigger than bio_children (%u).", pbp->bio_inbed,
1001 	    pbp->bio_children));
1002 	if (bp->bio_error == 0 && pbp->bio_error == 0) {
1003 		G_MIRROR_LOGREQ(3, bp, "Request delivered.");
1004 		g_destroy_bio(bp);
1005 		if (pbp->bio_children == pbp->bio_inbed) {
1006 			G_MIRROR_LOGREQ(3, pbp, "Request delivered.");
1007 			pbp->bio_completed = pbp->bio_length;
1008 			if (pbp->bio_cmd == BIO_WRITE ||
1009 			    pbp->bio_cmd == BIO_DELETE) {
1010 				TAILQ_REMOVE(&sc->sc_inflight, pbp, bio_queue);
1011 				/* Release delayed sync requests if possible. */
1012 				g_mirror_sync_release(sc);
1013 			}
1014 			g_io_deliver(pbp, pbp->bio_error);
1015 		}
1016 		return;
1017 	} else if (bp->bio_error != 0) {
1018 		if (pbp->bio_error == 0)
1019 			pbp->bio_error = bp->bio_error;
1020 		if (disk != NULL)
1021 			g_mirror_regular_request_error(sc, disk, bp);
1022 		switch (pbp->bio_cmd) {
1023 		case BIO_DELETE:
1024 		case BIO_WRITE:
1025 		case BIO_FLUSH:
1026 		case BIO_SPEEDUP:
1027 			pbp->bio_inbed--;
1028 			pbp->bio_children--;
1029 			break;
1030 		}
1031 	}
1032 	g_destroy_bio(bp);
1033 
1034 	switch (pbp->bio_cmd) {
1035 	case BIO_READ:
1036 		if (pbp->bio_inbed < pbp->bio_children)
1037 			break;
1038 		if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 1)
1039 			g_io_deliver(pbp, pbp->bio_error);
1040 		else {
1041 			pbp->bio_error = 0;
1042 			mtx_lock(&sc->sc_queue_mtx);
1043 			TAILQ_INSERT_TAIL(&sc->sc_queue, pbp, bio_queue);
1044 			mtx_unlock(&sc->sc_queue_mtx);
1045 			G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1046 			wakeup(sc);
1047 		}
1048 		break;
1049 	case BIO_DELETE:
1050 	case BIO_WRITE:
1051 	case BIO_FLUSH:
1052 	case BIO_SPEEDUP:
1053 		if (pbp->bio_children == 0) {
1054 			/*
1055 			 * All requests failed.
1056 			 */
1057 		} else if (pbp->bio_inbed < pbp->bio_children) {
1058 			/* Do nothing. */
1059 			break;
1060 		} else if (pbp->bio_children == pbp->bio_inbed) {
1061 			/* Some requests succeeded. */
1062 			pbp->bio_error = 0;
1063 			pbp->bio_completed = pbp->bio_length;
1064 		}
1065 		if (pbp->bio_cmd == BIO_WRITE || pbp->bio_cmd == BIO_DELETE) {
1066 			TAILQ_REMOVE(&sc->sc_inflight, pbp, bio_queue);
1067 			/* Release delayed sync requests if possible. */
1068 			g_mirror_sync_release(sc);
1069 		}
1070 		g_io_deliver(pbp, pbp->bio_error);
1071 		break;
1072 	default:
1073 		KASSERT(1 == 0, ("Invalid request: %u.", pbp->bio_cmd));
1074 		break;
1075 	}
1076 }
1077 
1078 static void
1079 g_mirror_sync_done(struct bio *bp)
1080 {
1081 	struct g_mirror_softc *sc;
1082 
1083 	G_MIRROR_LOGREQ(3, bp, "Synchronization request delivered.");
1084 	sc = bp->bio_from->geom->softc;
1085 	bp->bio_cflags = G_MIRROR_BIO_FLAG_SYNC;
1086 	mtx_lock(&sc->sc_queue_mtx);
1087 	TAILQ_INSERT_TAIL(&sc->sc_queue, bp, bio_queue);
1088 	mtx_unlock(&sc->sc_queue_mtx);
1089 	wakeup(sc);
1090 }
1091 
1092 static void
1093 g_mirror_candelete(struct bio *bp)
1094 {
1095 	struct g_mirror_softc *sc;
1096 	struct g_mirror_disk *disk;
1097 	int val;
1098 
1099 	sc = bp->bio_to->private;
1100 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1101 		if (disk->d_flags & G_MIRROR_DISK_FLAG_CANDELETE)
1102 			break;
1103 	}
1104 	val = disk != NULL;
1105 	g_handleattr(bp, "GEOM::candelete", &val, sizeof(val));
1106 }
1107 
1108 static void
1109 g_mirror_kernel_dump(struct bio *bp)
1110 {
1111 	struct g_mirror_softc *sc;
1112 	struct g_mirror_disk *disk;
1113 	struct bio *cbp;
1114 	struct g_kerneldump *gkd;
1115 
1116 	/*
1117 	 * We configure dumping to the first component, because this component
1118 	 * will be used for reading with 'prefer' balance algorithm.
1119 	 * If the component with the highest priority is currently disconnected
1120 	 * we will not be able to read the dump after the reboot if it will be
1121 	 * connected and synchronized later. Can we do something better?
1122 	 */
1123 	sc = bp->bio_to->private;
1124 	disk = LIST_FIRST(&sc->sc_disks);
1125 
1126 	gkd = (struct g_kerneldump *)bp->bio_data;
1127 	if (gkd->length > bp->bio_to->mediasize)
1128 		gkd->length = bp->bio_to->mediasize;
1129 	cbp = g_clone_bio(bp);
1130 	if (cbp == NULL) {
1131 		g_io_deliver(bp, ENOMEM);
1132 		return;
1133 	}
1134 	cbp->bio_done = g_std_done;
1135 	g_io_request(cbp, disk->d_consumer);
1136 	G_MIRROR_DEBUG(1, "Kernel dump will go to %s.",
1137 	    g_mirror_get_diskname(disk));
1138 }
1139 
1140 static void
1141 g_mirror_start(struct bio *bp)
1142 {
1143 	struct g_mirror_softc *sc;
1144 
1145 	sc = bp->bio_to->private;
1146 	/*
1147 	 * If sc == NULL or there are no valid disks, provider's error
1148 	 * should be set and g_mirror_start() should not be called at all.
1149 	 */
1150 	KASSERT(sc != NULL && sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1151 	    ("Provider's error should be set (error=%d)(mirror=%s).",
1152 	    bp->bio_to->error, bp->bio_to->name));
1153 	G_MIRROR_LOGREQ(3, bp, "Request received.");
1154 
1155 	switch (bp->bio_cmd) {
1156 	case BIO_READ:
1157 	case BIO_WRITE:
1158 	case BIO_DELETE:
1159 	case BIO_SPEEDUP:
1160 	case BIO_FLUSH:
1161 		break;
1162 	case BIO_GETATTR:
1163 		if (!strcmp(bp->bio_attribute, "GEOM::candelete")) {
1164 			g_mirror_candelete(bp);
1165 			return;
1166 		} else if (strcmp("GEOM::kerneldump", bp->bio_attribute) == 0) {
1167 			g_mirror_kernel_dump(bp);
1168 			return;
1169 		}
1170 		/* FALLTHROUGH */
1171 	default:
1172 		g_io_deliver(bp, EOPNOTSUPP);
1173 		return;
1174 	}
1175 	mtx_lock(&sc->sc_queue_mtx);
1176 	if (bp->bio_to->error != 0) {
1177 		mtx_unlock(&sc->sc_queue_mtx);
1178 		g_io_deliver(bp, bp->bio_to->error);
1179 		return;
1180 	}
1181 	TAILQ_INSERT_TAIL(&sc->sc_queue, bp, bio_queue);
1182 	mtx_unlock(&sc->sc_queue_mtx);
1183 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1184 	wakeup(sc);
1185 }
1186 
1187 /*
1188  * Return TRUE if the given request is colliding with a in-progress
1189  * synchronization request.
1190  */
1191 static bool
1192 g_mirror_sync_collision(struct g_mirror_softc *sc, struct bio *bp)
1193 {
1194 	struct g_mirror_disk *disk;
1195 	struct bio *sbp;
1196 	off_t rstart, rend, sstart, send;
1197 	u_int i;
1198 
1199 	if (sc->sc_sync.ds_ndisks == 0)
1200 		return (false);
1201 	rstart = bp->bio_offset;
1202 	rend = bp->bio_offset + bp->bio_length;
1203 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1204 		if (disk->d_state != G_MIRROR_DISK_STATE_SYNCHRONIZING)
1205 			continue;
1206 		for (i = 0; i < g_mirror_syncreqs; i++) {
1207 			sbp = disk->d_sync.ds_bios[i];
1208 			if (sbp == NULL)
1209 				continue;
1210 			sstart = sbp->bio_offset;
1211 			send = sbp->bio_offset + sbp->bio_length;
1212 			if (rend > sstart && rstart < send)
1213 				return (true);
1214 		}
1215 	}
1216 	return (false);
1217 }
1218 
1219 /*
1220  * Return TRUE if the given sync request is colliding with a in-progress regular
1221  * request.
1222  */
1223 static bool
1224 g_mirror_regular_collision(struct g_mirror_softc *sc, struct bio *sbp)
1225 {
1226 	off_t rstart, rend, sstart, send;
1227 	struct bio *bp;
1228 
1229 	if (sc->sc_sync.ds_ndisks == 0)
1230 		return (false);
1231 	sstart = sbp->bio_offset;
1232 	send = sbp->bio_offset + sbp->bio_length;
1233 	TAILQ_FOREACH(bp, &sc->sc_inflight, bio_queue) {
1234 		rstart = bp->bio_offset;
1235 		rend = bp->bio_offset + bp->bio_length;
1236 		if (rend > sstart && rstart < send)
1237 			return (true);
1238 	}
1239 	return (false);
1240 }
1241 
1242 /*
1243  * Puts regular request onto delayed queue.
1244  */
1245 static void
1246 g_mirror_regular_delay(struct g_mirror_softc *sc, struct bio *bp)
1247 {
1248 
1249 	G_MIRROR_LOGREQ(2, bp, "Delaying request.");
1250 	TAILQ_INSERT_TAIL(&sc->sc_regular_delayed, bp, bio_queue);
1251 }
1252 
1253 /*
1254  * Puts synchronization request onto delayed queue.
1255  */
1256 static void
1257 g_mirror_sync_delay(struct g_mirror_softc *sc, struct bio *bp)
1258 {
1259 
1260 	G_MIRROR_LOGREQ(2, bp, "Delaying synchronization request.");
1261 	TAILQ_INSERT_TAIL(&sc->sc_sync_delayed, bp, bio_queue);
1262 }
1263 
1264 /*
1265  * Requeue delayed regular requests.
1266  */
1267 static void
1268 g_mirror_regular_release(struct g_mirror_softc *sc)
1269 {
1270 	struct bio *bp;
1271 
1272 	if ((bp = TAILQ_FIRST(&sc->sc_regular_delayed)) == NULL)
1273 		return;
1274 	if (g_mirror_sync_collision(sc, bp))
1275 		return;
1276 
1277 	G_MIRROR_DEBUG(2, "Requeuing regular requests after collision.");
1278 	mtx_lock(&sc->sc_queue_mtx);
1279 	TAILQ_CONCAT(&sc->sc_regular_delayed, &sc->sc_queue, bio_queue);
1280 	TAILQ_SWAP(&sc->sc_regular_delayed, &sc->sc_queue, bio, bio_queue);
1281 	mtx_unlock(&sc->sc_queue_mtx);
1282 }
1283 
1284 /*
1285  * Releases delayed sync requests which don't collide anymore with regular
1286  * requests.
1287  */
1288 static void
1289 g_mirror_sync_release(struct g_mirror_softc *sc)
1290 {
1291 	struct bio *bp, *bp2;
1292 
1293 	TAILQ_FOREACH_SAFE(bp, &sc->sc_sync_delayed, bio_queue, bp2) {
1294 		if (g_mirror_regular_collision(sc, bp))
1295 			continue;
1296 		TAILQ_REMOVE(&sc->sc_sync_delayed, bp, bio_queue);
1297 		G_MIRROR_LOGREQ(2, bp,
1298 		    "Releasing delayed synchronization request.");
1299 		g_io_request(bp, bp->bio_from);
1300 	}
1301 }
1302 
1303 /*
1304  * Free a synchronization request and clear its slot in the array.
1305  */
1306 static void
1307 g_mirror_sync_request_free(struct g_mirror_disk *disk, struct bio *bp)
1308 {
1309 	int idx;
1310 
1311 	if (disk != NULL && disk->d_sync.ds_bios != NULL) {
1312 		idx = (int)(uintptr_t)bp->bio_caller1;
1313 		KASSERT(disk->d_sync.ds_bios[idx] == bp,
1314 		    ("unexpected sync BIO at %p:%d", disk, idx));
1315 		disk->d_sync.ds_bios[idx] = NULL;
1316 	}
1317 	free(bp->bio_data, M_MIRROR);
1318 	g_destroy_bio(bp);
1319 }
1320 
1321 /*
1322  * Handle synchronization requests.
1323  * Every synchronization request is a two-step process: first, a read request is
1324  * sent to the mirror provider via the sync consumer. If that request completes
1325  * successfully, it is converted to a write and sent to the disk being
1326  * synchronized. If the write also completes successfully, the synchronization
1327  * offset is advanced and a new read request is submitted.
1328  */
1329 static void
1330 g_mirror_sync_request(struct g_mirror_softc *sc, struct bio *bp)
1331 {
1332 	struct g_mirror_disk *disk;
1333 	struct g_mirror_disk_sync *sync;
1334 
1335 	KASSERT((bp->bio_cmd == BIO_READ &&
1336 	    bp->bio_from->geom == sc->sc_sync.ds_geom) ||
1337 	    (bp->bio_cmd == BIO_WRITE && bp->bio_from->geom == sc->sc_geom),
1338 	    ("Sync BIO %p with unexpected origin", bp));
1339 
1340 	bp->bio_from->index--;
1341 	disk = bp->bio_from->private;
1342 	if (disk == NULL) {
1343 		sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1344 		g_topology_lock();
1345 		g_mirror_kill_consumer(sc, bp->bio_from);
1346 		g_topology_unlock();
1347 		g_mirror_sync_request_free(NULL, bp);
1348 		sx_xlock(&sc->sc_lock);
1349 		return;
1350 	}
1351 
1352 	sync = &disk->d_sync;
1353 
1354 	/*
1355 	 * Synchronization request.
1356 	 */
1357 	switch (bp->bio_cmd) {
1358 	case BIO_READ: {
1359 		struct g_consumer *cp;
1360 
1361 		KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_sync_request_read,
1362 		    bp->bio_error);
1363 
1364 		if (bp->bio_error != 0) {
1365 			G_MIRROR_LOGREQ(0, bp,
1366 			    "Synchronization request failed (error=%d).",
1367 			    bp->bio_error);
1368 
1369 			/*
1370 			 * The read error will trigger a syncid bump, so there's
1371 			 * no need to do that here.
1372 			 *
1373 			 * The read error handling for regular requests will
1374 			 * retry the read from all active mirrors before passing
1375 			 * the error back up, so there's no need to retry here.
1376 			 */
1377 			g_mirror_sync_request_free(disk, bp);
1378 			g_mirror_event_send(disk,
1379 			    G_MIRROR_DISK_STATE_DISCONNECTED,
1380 			    G_MIRROR_EVENT_DONTWAIT);
1381 			return;
1382 		}
1383 		G_MIRROR_LOGREQ(3, bp,
1384 		    "Synchronization request half-finished.");
1385 		bp->bio_cmd = BIO_WRITE;
1386 		bp->bio_cflags = 0;
1387 		cp = disk->d_consumer;
1388 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1389 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1390 		    cp->acr, cp->acw, cp->ace));
1391 		cp->index++;
1392 		g_io_request(bp, cp);
1393 		return;
1394 	}
1395 	case BIO_WRITE: {
1396 		off_t offset;
1397 		int i;
1398 
1399 		KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_sync_request_write,
1400 		    bp->bio_error);
1401 
1402 		if (bp->bio_error != 0) {
1403 			G_MIRROR_LOGREQ(0, bp,
1404 			    "Synchronization request failed (error=%d).",
1405 			    bp->bio_error);
1406 			g_mirror_sync_request_free(disk, bp);
1407 			sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
1408 			g_mirror_event_send(disk,
1409 			    G_MIRROR_DISK_STATE_DISCONNECTED,
1410 			    G_MIRROR_EVENT_DONTWAIT);
1411 			return;
1412 		}
1413 		G_MIRROR_LOGREQ(3, bp, "Synchronization request finished.");
1414 		if (sync->ds_offset >= sc->sc_mediasize ||
1415 		    sync->ds_consumer == NULL ||
1416 		    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1417 			/* Don't send more synchronization requests. */
1418 			sync->ds_inflight--;
1419 			g_mirror_sync_request_free(disk, bp);
1420 			if (sync->ds_inflight > 0)
1421 				return;
1422 			if (sync->ds_consumer == NULL ||
1423 			    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1424 				return;
1425 			}
1426 			/* Disk up-to-date, activate it. */
1427 			g_mirror_event_send(disk, G_MIRROR_DISK_STATE_ACTIVE,
1428 			    G_MIRROR_EVENT_DONTWAIT);
1429 			return;
1430 		}
1431 
1432 		/* Send next synchronization request. */
1433 		g_mirror_sync_reinit(disk, bp, sync->ds_offset);
1434 		sync->ds_offset += bp->bio_length;
1435 
1436 		G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
1437 		sync->ds_consumer->index++;
1438 
1439 		/*
1440 		 * Delay the request if it is colliding with a regular request.
1441 		 */
1442 		if (g_mirror_regular_collision(sc, bp))
1443 			g_mirror_sync_delay(sc, bp);
1444 		else
1445 			g_io_request(bp, sync->ds_consumer);
1446 
1447 		/* Requeue delayed requests if possible. */
1448 		g_mirror_regular_release(sc);
1449 
1450 		/* Find the smallest offset */
1451 		offset = sc->sc_mediasize;
1452 		for (i = 0; i < g_mirror_syncreqs; i++) {
1453 			bp = sync->ds_bios[i];
1454 			if (bp != NULL && bp->bio_offset < offset)
1455 				offset = bp->bio_offset;
1456 		}
1457 		if (g_mirror_sync_period > 0 &&
1458 		    time_uptime - sync->ds_update_ts > g_mirror_sync_period) {
1459 			sync->ds_offset_done = offset;
1460 			g_mirror_update_metadata(disk);
1461 			sync->ds_update_ts = time_uptime;
1462 		}
1463 		return;
1464 	}
1465 	default:
1466 		panic("Invalid I/O request %p", bp);
1467 	}
1468 }
1469 
1470 static void
1471 g_mirror_request_prefer(struct g_mirror_softc *sc, struct bio *bp)
1472 {
1473 	struct g_mirror_disk *disk;
1474 	struct g_consumer *cp;
1475 	struct bio *cbp;
1476 
1477 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1478 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE)
1479 			break;
1480 	}
1481 	if (disk == NULL) {
1482 		if (bp->bio_error == 0)
1483 			bp->bio_error = ENXIO;
1484 		g_io_deliver(bp, bp->bio_error);
1485 		return;
1486 	}
1487 	cbp = g_clone_bio(bp);
1488 	if (cbp == NULL) {
1489 		if (bp->bio_error == 0)
1490 			bp->bio_error = ENOMEM;
1491 		g_io_deliver(bp, bp->bio_error);
1492 		return;
1493 	}
1494 	/*
1495 	 * Fill in the component buf structure.
1496 	 */
1497 	cp = disk->d_consumer;
1498 	cbp->bio_done = g_mirror_done;
1499 	cbp->bio_to = cp->provider;
1500 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1501 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1502 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1503 	    cp->acw, cp->ace));
1504 	cp->index++;
1505 	g_io_request(cbp, cp);
1506 }
1507 
1508 static void
1509 g_mirror_request_round_robin(struct g_mirror_softc *sc, struct bio *bp)
1510 {
1511 	struct g_mirror_disk *disk;
1512 	struct g_consumer *cp;
1513 	struct bio *cbp;
1514 
1515 	disk = g_mirror_get_disk(sc);
1516 	if (disk == NULL) {
1517 		if (bp->bio_error == 0)
1518 			bp->bio_error = ENXIO;
1519 		g_io_deliver(bp, bp->bio_error);
1520 		return;
1521 	}
1522 	cbp = g_clone_bio(bp);
1523 	if (cbp == NULL) {
1524 		if (bp->bio_error == 0)
1525 			bp->bio_error = ENOMEM;
1526 		g_io_deliver(bp, bp->bio_error);
1527 		return;
1528 	}
1529 	/*
1530 	 * Fill in the component buf structure.
1531 	 */
1532 	cp = disk->d_consumer;
1533 	cbp->bio_done = g_mirror_done;
1534 	cbp->bio_to = cp->provider;
1535 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1536 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1537 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1538 	    cp->acw, cp->ace));
1539 	cp->index++;
1540 	g_io_request(cbp, cp);
1541 }
1542 
1543 #define TRACK_SIZE  (1 * 1024 * 1024)
1544 #define LOAD_SCALE	256
1545 #define ABS(x)		(((x) >= 0) ? (x) : (-(x)))
1546 
1547 static void
1548 g_mirror_request_load(struct g_mirror_softc *sc, struct bio *bp)
1549 {
1550 	struct g_mirror_disk *disk, *dp;
1551 	struct g_consumer *cp;
1552 	struct bio *cbp;
1553 	int prio, best;
1554 
1555 	/* Find a disk with the smallest load. */
1556 	disk = NULL;
1557 	best = INT_MAX;
1558 	LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1559 		if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1560 			continue;
1561 		prio = dp->load;
1562 		/* If disk head is precisely in position - highly prefer it. */
1563 		if (dp->d_last_offset == bp->bio_offset)
1564 			prio -= 2 * LOAD_SCALE;
1565 		else
1566 		/* If disk head is close to position - prefer it. */
1567 		if (ABS(dp->d_last_offset - bp->bio_offset) < TRACK_SIZE)
1568 			prio -= 1 * LOAD_SCALE;
1569 		if (prio <= best) {
1570 			disk = dp;
1571 			best = prio;
1572 		}
1573 	}
1574 	KASSERT(disk != NULL, ("NULL disk for %s.", sc->sc_name));
1575 	cbp = g_clone_bio(bp);
1576 	if (cbp == NULL) {
1577 		if (bp->bio_error == 0)
1578 			bp->bio_error = ENOMEM;
1579 		g_io_deliver(bp, bp->bio_error);
1580 		return;
1581 	}
1582 	/*
1583 	 * Fill in the component buf structure.
1584 	 */
1585 	cp = disk->d_consumer;
1586 	cbp->bio_done = g_mirror_done;
1587 	cbp->bio_to = cp->provider;
1588 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1589 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1590 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1591 	    cp->acw, cp->ace));
1592 	cp->index++;
1593 	/* Remember last head position */
1594 	disk->d_last_offset = bp->bio_offset + bp->bio_length;
1595 	/* Update loads. */
1596 	LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1597 		dp->load = (dp->d_consumer->index * LOAD_SCALE +
1598 		    dp->load * 7) / 8;
1599 	}
1600 	g_io_request(cbp, cp);
1601 }
1602 
1603 static void
1604 g_mirror_request_split(struct g_mirror_softc *sc, struct bio *bp)
1605 {
1606 	struct bio_queue queue;
1607 	struct g_mirror_disk *disk;
1608 	struct g_consumer *cp __diagused;
1609 	struct bio *cbp;
1610 	off_t left, mod, offset, slice;
1611 	u_char *data;
1612 	u_int ndisks;
1613 
1614 	if (bp->bio_length <= sc->sc_slice) {
1615 		g_mirror_request_round_robin(sc, bp);
1616 		return;
1617 	}
1618 	ndisks = g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE);
1619 	slice = bp->bio_length / ndisks;
1620 	mod = slice % sc->sc_provider->sectorsize;
1621 	if (mod != 0)
1622 		slice += sc->sc_provider->sectorsize - mod;
1623 	/*
1624 	 * Allocate all bios before sending any request, so we can
1625 	 * return ENOMEM in nice and clean way.
1626 	 */
1627 	left = bp->bio_length;
1628 	offset = bp->bio_offset;
1629 	data = bp->bio_data;
1630 	TAILQ_INIT(&queue);
1631 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1632 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1633 			continue;
1634 		cbp = g_clone_bio(bp);
1635 		if (cbp == NULL) {
1636 			while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1637 				TAILQ_REMOVE(&queue, cbp, bio_queue);
1638 				g_destroy_bio(cbp);
1639 			}
1640 			if (bp->bio_error == 0)
1641 				bp->bio_error = ENOMEM;
1642 			g_io_deliver(bp, bp->bio_error);
1643 			return;
1644 		}
1645 		TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
1646 		cbp->bio_done = g_mirror_done;
1647 		cbp->bio_caller1 = disk;
1648 		cbp->bio_to = disk->d_consumer->provider;
1649 		cbp->bio_offset = offset;
1650 		cbp->bio_data = data;
1651 		cbp->bio_length = MIN(left, slice);
1652 		left -= cbp->bio_length;
1653 		if (left == 0)
1654 			break;
1655 		offset += cbp->bio_length;
1656 		data += cbp->bio_length;
1657 	}
1658 	while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1659 		TAILQ_REMOVE(&queue, cbp, bio_queue);
1660 		G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1661 		disk = cbp->bio_caller1;
1662 		cbp->bio_caller1 = NULL;
1663 		cp = disk->d_consumer;
1664 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1665 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1666 		    cp->acr, cp->acw, cp->ace));
1667 		disk->d_consumer->index++;
1668 		g_io_request(cbp, disk->d_consumer);
1669 	}
1670 }
1671 
1672 static void
1673 g_mirror_register_request(struct g_mirror_softc *sc, struct bio *bp)
1674 {
1675 	struct bio_queue queue;
1676 	struct bio *cbp;
1677 	struct g_consumer *cp;
1678 	struct g_mirror_disk *disk;
1679 
1680 	sx_assert(&sc->sc_lock, SA_XLOCKED);
1681 
1682 	/*
1683 	 * To avoid ordering issues, if a write is deferred because of a
1684 	 * collision with a sync request, all I/O is deferred until that
1685 	 * write is initiated.
1686 	 */
1687 	if (bp->bio_from->geom != sc->sc_sync.ds_geom &&
1688 	    !TAILQ_EMPTY(&sc->sc_regular_delayed)) {
1689 		g_mirror_regular_delay(sc, bp);
1690 		return;
1691 	}
1692 
1693 	switch (bp->bio_cmd) {
1694 	case BIO_READ:
1695 		switch (sc->sc_balance) {
1696 		case G_MIRROR_BALANCE_LOAD:
1697 			g_mirror_request_load(sc, bp);
1698 			break;
1699 		case G_MIRROR_BALANCE_PREFER:
1700 			g_mirror_request_prefer(sc, bp);
1701 			break;
1702 		case G_MIRROR_BALANCE_ROUND_ROBIN:
1703 			g_mirror_request_round_robin(sc, bp);
1704 			break;
1705 		case G_MIRROR_BALANCE_SPLIT:
1706 			g_mirror_request_split(sc, bp);
1707 			break;
1708 		}
1709 		return;
1710 	case BIO_WRITE:
1711 	case BIO_DELETE:
1712 		/*
1713 		 * Delay the request if it is colliding with a synchronization
1714 		 * request.
1715 		 */
1716 		if (g_mirror_sync_collision(sc, bp)) {
1717 			g_mirror_regular_delay(sc, bp);
1718 			return;
1719 		}
1720 
1721 		if (sc->sc_idle)
1722 			g_mirror_unidle(sc);
1723 		else
1724 			sc->sc_last_write = time_uptime;
1725 
1726 		/*
1727 		 * Bump syncid on first write.
1728 		 */
1729 		if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0) {
1730 			sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
1731 			g_mirror_bump_syncid(sc);
1732 		}
1733 
1734 		/*
1735 		 * Allocate all bios before sending any request, so we can
1736 		 * return ENOMEM in nice and clean way.
1737 		 */
1738 		TAILQ_INIT(&queue);
1739 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1740 			switch (disk->d_state) {
1741 			case G_MIRROR_DISK_STATE_ACTIVE:
1742 				break;
1743 			case G_MIRROR_DISK_STATE_SYNCHRONIZING:
1744 				if (bp->bio_offset >= disk->d_sync.ds_offset)
1745 					continue;
1746 				break;
1747 			default:
1748 				continue;
1749 			}
1750 			if (bp->bio_cmd == BIO_DELETE &&
1751 			    (disk->d_flags & G_MIRROR_DISK_FLAG_CANDELETE) == 0)
1752 				continue;
1753 			cbp = g_clone_bio(bp);
1754 			if (cbp == NULL) {
1755 				while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1756 					TAILQ_REMOVE(&queue, cbp, bio_queue);
1757 					g_destroy_bio(cbp);
1758 				}
1759 				if (bp->bio_error == 0)
1760 					bp->bio_error = ENOMEM;
1761 				g_io_deliver(bp, bp->bio_error);
1762 				return;
1763 			}
1764 			TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
1765 			cbp->bio_done = g_mirror_done;
1766 			cp = disk->d_consumer;
1767 			cbp->bio_caller1 = cp;
1768 			cbp->bio_to = cp->provider;
1769 			KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1770 			    ("Consumer %s not opened (r%dw%de%d).",
1771 			    cp->provider->name, cp->acr, cp->acw, cp->ace));
1772 		}
1773 		if (TAILQ_EMPTY(&queue)) {
1774 			KASSERT(bp->bio_cmd == BIO_DELETE,
1775 			    ("No consumers for regular request %p", bp));
1776 			g_io_deliver(bp, EOPNOTSUPP);
1777 			return;
1778 		}
1779 		while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1780 			G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1781 			TAILQ_REMOVE(&queue, cbp, bio_queue);
1782 			cp = cbp->bio_caller1;
1783 			cbp->bio_caller1 = NULL;
1784 			cp->index++;
1785 			sc->sc_writes++;
1786 			g_io_request(cbp, cp);
1787 		}
1788 		/*
1789 		 * Put request onto inflight queue, so we can check if new
1790 		 * synchronization requests don't collide with it.
1791 		 */
1792 		TAILQ_INSERT_TAIL(&sc->sc_inflight, bp, bio_queue);
1793 		return;
1794 	case BIO_SPEEDUP:
1795 	case BIO_FLUSH:
1796 		TAILQ_INIT(&queue);
1797 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1798 			if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1799 				continue;
1800 			cbp = g_clone_bio(bp);
1801 			if (cbp == NULL) {
1802 				while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1803 					TAILQ_REMOVE(&queue, cbp, bio_queue);
1804 					g_destroy_bio(cbp);
1805 				}
1806 				if (bp->bio_error == 0)
1807 					bp->bio_error = ENOMEM;
1808 				g_io_deliver(bp, bp->bio_error);
1809 				return;
1810 			}
1811 			TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
1812 			cbp->bio_done = g_mirror_done;
1813 			cbp->bio_caller1 = disk;
1814 			cbp->bio_to = disk->d_consumer->provider;
1815 		}
1816 		KASSERT(!TAILQ_EMPTY(&queue),
1817 		    ("No consumers for regular request %p", bp));
1818 		while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1819 			G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1820 			TAILQ_REMOVE(&queue, cbp, bio_queue);
1821 			disk = cbp->bio_caller1;
1822 			cbp->bio_caller1 = NULL;
1823 			cp = disk->d_consumer;
1824 			KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1825 			    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1826 			    cp->acr, cp->acw, cp->ace));
1827 			cp->index++;
1828 			g_io_request(cbp, cp);
1829 		}
1830 		break;
1831 	default:
1832 		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1833 		    bp->bio_cmd, sc->sc_name));
1834 		break;
1835 	}
1836 }
1837 
1838 static int
1839 g_mirror_can_destroy(struct g_mirror_softc *sc)
1840 {
1841 	struct g_geom *gp;
1842 	struct g_consumer *cp;
1843 
1844 	g_topology_assert();
1845 	gp = sc->sc_geom;
1846 	if (gp->softc == NULL)
1847 		return (1);
1848 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_TASTING) != 0)
1849 		return (0);
1850 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1851 		if (g_mirror_is_busy(sc, cp))
1852 			return (0);
1853 	}
1854 	gp = sc->sc_sync.ds_geom;
1855 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1856 		if (g_mirror_is_busy(sc, cp))
1857 			return (0);
1858 	}
1859 	G_MIRROR_DEBUG(2, "No I/O requests for %s, it can be destroyed.",
1860 	    sc->sc_name);
1861 	return (1);
1862 }
1863 
1864 static int
1865 g_mirror_try_destroy(struct g_mirror_softc *sc)
1866 {
1867 
1868 	if (sc->sc_rootmount != NULL) {
1869 		G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
1870 		    sc->sc_rootmount);
1871 		root_mount_rel(sc->sc_rootmount);
1872 		sc->sc_rootmount = NULL;
1873 	}
1874 	g_topology_lock();
1875 	if (!g_mirror_can_destroy(sc)) {
1876 		g_topology_unlock();
1877 		return (0);
1878 	}
1879 	sc->sc_geom->softc = NULL;
1880 	sc->sc_sync.ds_geom->softc = NULL;
1881 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DRAIN) != 0) {
1882 		g_topology_unlock();
1883 		G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1884 		    &sc->sc_worker);
1885 		/* Unlock sc_lock here, as it can be destroyed after wakeup. */
1886 		sx_xunlock(&sc->sc_lock);
1887 		wakeup(&sc->sc_worker);
1888 		sc->sc_worker = NULL;
1889 	} else {
1890 		g_topology_unlock();
1891 		g_mirror_destroy_device(sc);
1892 	}
1893 	return (1);
1894 }
1895 
1896 /*
1897  * Worker thread.
1898  */
1899 static void
1900 g_mirror_worker(void *arg)
1901 {
1902 	struct g_mirror_softc *sc;
1903 	struct g_mirror_event *ep;
1904 	struct bio *bp;
1905 	int timeout;
1906 
1907 	sc = arg;
1908 	thread_lock(curthread);
1909 	sched_prio(curthread, PRIBIO);
1910 	thread_unlock(curthread);
1911 
1912 	sx_xlock(&sc->sc_lock);
1913 	for (;;) {
1914 		G_MIRROR_DEBUG(5, "%s: Let's see...", __func__);
1915 		/*
1916 		 * First take a look at events.
1917 		 * This is important to handle events before any I/O requests.
1918 		 */
1919 		ep = g_mirror_event_first(sc);
1920 		if (ep != NULL) {
1921 			g_mirror_event_remove(sc, ep);
1922 			if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0) {
1923 				/* Update only device status. */
1924 				G_MIRROR_DEBUG(3,
1925 				    "Running event for device %s.",
1926 				    sc->sc_name);
1927 				ep->e_error = 0;
1928 				g_mirror_update_device(sc, true);
1929 			} else {
1930 				/* Update disk status. */
1931 				G_MIRROR_DEBUG(3, "Running event for disk %s.",
1932 				     g_mirror_get_diskname(ep->e_disk));
1933 				ep->e_error = g_mirror_update_disk(ep->e_disk,
1934 				    ep->e_state);
1935 				if (ep->e_error == 0)
1936 					g_mirror_update_device(sc, false);
1937 			}
1938 			if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0) {
1939 				KASSERT(ep->e_error == 0,
1940 				    ("Error cannot be handled."));
1941 				g_mirror_event_free(ep);
1942 			} else {
1943 				ep->e_flags |= G_MIRROR_EVENT_DONE;
1944 				G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1945 				    ep);
1946 				mtx_lock(&sc->sc_events_mtx);
1947 				wakeup(ep);
1948 				mtx_unlock(&sc->sc_events_mtx);
1949 			}
1950 			if ((sc->sc_flags &
1951 			    G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1952 				if (g_mirror_try_destroy(sc)) {
1953 					curthread->td_pflags &= ~TDP_GEOM;
1954 					G_MIRROR_DEBUG(1, "Thread exiting.");
1955 					kproc_exit(0);
1956 				}
1957 			}
1958 			G_MIRROR_DEBUG(5, "%s: I'm here 1.", __func__);
1959 			continue;
1960 		}
1961 
1962 		/*
1963 		 * Check if we can mark array as CLEAN and if we can't take
1964 		 * how much seconds should we wait.
1965 		 */
1966 		timeout = g_mirror_idle(sc, -1);
1967 
1968 		/*
1969 		 * Handle I/O requests.
1970 		 */
1971 		mtx_lock(&sc->sc_queue_mtx);
1972 		bp = TAILQ_FIRST(&sc->sc_queue);
1973 		if (bp != NULL)
1974 			TAILQ_REMOVE(&sc->sc_queue, bp, bio_queue);
1975 		else {
1976 			if ((sc->sc_flags &
1977 			    G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1978 				mtx_unlock(&sc->sc_queue_mtx);
1979 				if (g_mirror_try_destroy(sc)) {
1980 					curthread->td_pflags &= ~TDP_GEOM;
1981 					G_MIRROR_DEBUG(1, "Thread exiting.");
1982 					kproc_exit(0);
1983 				}
1984 				mtx_lock(&sc->sc_queue_mtx);
1985 				if (!TAILQ_EMPTY(&sc->sc_queue)) {
1986 					mtx_unlock(&sc->sc_queue_mtx);
1987 					continue;
1988 				}
1989 			}
1990 			if (g_mirror_event_first(sc) != NULL) {
1991 				mtx_unlock(&sc->sc_queue_mtx);
1992 				continue;
1993 			}
1994 			sx_xunlock(&sc->sc_lock);
1995 			MSLEEP(sc, &sc->sc_queue_mtx, PRIBIO | PDROP, "m:w1",
1996 			    timeout * hz);
1997 			sx_xlock(&sc->sc_lock);
1998 			G_MIRROR_DEBUG(5, "%s: I'm here 4.", __func__);
1999 			continue;
2000 		}
2001 		mtx_unlock(&sc->sc_queue_mtx);
2002 
2003 		if (bp->bio_from->geom == sc->sc_sync.ds_geom &&
2004 		    (bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0) {
2005 			/*
2006 			 * Handle completion of the first half (the read) of a
2007 			 * block synchronization operation.
2008 			 */
2009 			g_mirror_sync_request(sc, bp);
2010 		} else if (bp->bio_to != sc->sc_provider) {
2011 			if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_REGULAR) != 0)
2012 				/*
2013 				 * Handle completion of a regular I/O request.
2014 				 */
2015 				g_mirror_regular_request(sc, bp);
2016 			else if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0)
2017 				/*
2018 				 * Handle completion of the second half (the
2019 				 * write) of a block synchronization operation.
2020 				 */
2021 				g_mirror_sync_request(sc, bp);
2022 			else {
2023 				KASSERT(0,
2024 				    ("Invalid request cflags=0x%hx to=%s.",
2025 				    bp->bio_cflags, bp->bio_to->name));
2026 			}
2027 		} else {
2028 			/*
2029 			 * Initiate an I/O request.
2030 			 */
2031 			g_mirror_register_request(sc, bp);
2032 		}
2033 		G_MIRROR_DEBUG(5, "%s: I'm here 9.", __func__);
2034 	}
2035 }
2036 
2037 static void
2038 g_mirror_update_idle(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
2039 {
2040 
2041 	sx_assert(&sc->sc_lock, SX_LOCKED);
2042 
2043 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
2044 		return;
2045 	if (!sc->sc_idle && (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) == 0) {
2046 		G_MIRROR_DEBUG(2, "Disk %s (device %s) marked as dirty.",
2047 		    g_mirror_get_diskname(disk), sc->sc_name);
2048 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
2049 	} else if (sc->sc_idle &&
2050 	    (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
2051 		G_MIRROR_DEBUG(2, "Disk %s (device %s) marked as clean.",
2052 		    g_mirror_get_diskname(disk), sc->sc_name);
2053 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2054 	}
2055 }
2056 
2057 static void
2058 g_mirror_sync_reinit(const struct g_mirror_disk *disk, struct bio *bp,
2059     off_t offset)
2060 {
2061 	void *data;
2062 	int idx;
2063 
2064 	data = bp->bio_data;
2065 	idx = (int)(uintptr_t)bp->bio_caller1;
2066 	g_reset_bio(bp);
2067 
2068 	bp->bio_cmd = BIO_READ;
2069 	bp->bio_data = data;
2070 	bp->bio_done = g_mirror_sync_done;
2071 	bp->bio_from = disk->d_sync.ds_consumer;
2072 	bp->bio_to = disk->d_softc->sc_provider;
2073 	bp->bio_caller1 = (void *)(uintptr_t)idx;
2074 	bp->bio_offset = offset;
2075 	bp->bio_length = MIN(maxphys,
2076 	    disk->d_softc->sc_mediasize - bp->bio_offset);
2077 }
2078 
2079 static void
2080 g_mirror_sync_start(struct g_mirror_disk *disk)
2081 {
2082 	struct g_mirror_softc *sc;
2083 	struct g_mirror_disk_sync *sync;
2084 	struct g_consumer *cp;
2085 	struct bio *bp;
2086 	int error __diagused, i;
2087 
2088 	g_topology_assert_not();
2089 	sc = disk->d_softc;
2090 	sync = &disk->d_sync;
2091 	sx_assert(&sc->sc_lock, SX_LOCKED);
2092 
2093 	KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2094 	    ("Disk %s is not marked for synchronization.",
2095 	    g_mirror_get_diskname(disk)));
2096 	KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2097 	    ("Device not in RUNNING state (%s, %u).", sc->sc_name,
2098 	    sc->sc_state));
2099 
2100 	sx_xunlock(&sc->sc_lock);
2101 	g_topology_lock();
2102 	cp = g_new_consumer(sc->sc_sync.ds_geom);
2103 	cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2104 	error = g_attach(cp, sc->sc_provider);
2105 	KASSERT(error == 0,
2106 	    ("Cannot attach to %s (error=%d).", sc->sc_name, error));
2107 	error = g_access(cp, 1, 0, 0);
2108 	KASSERT(error == 0, ("Cannot open %s (error=%d).", sc->sc_name, error));
2109 	g_topology_unlock();
2110 	sx_xlock(&sc->sc_lock);
2111 
2112 	G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s.", sc->sc_name,
2113 	    g_mirror_get_diskname(disk));
2114 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) == 0)
2115 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
2116 	KASSERT(sync->ds_consumer == NULL,
2117 	    ("Sync consumer already exists (device=%s, disk=%s).",
2118 	    sc->sc_name, g_mirror_get_diskname(disk)));
2119 
2120 	sync->ds_consumer = cp;
2121 	sync->ds_consumer->private = disk;
2122 	sync->ds_consumer->index = 0;
2123 
2124 	/*
2125 	 * Allocate memory for synchronization bios and initialize them.
2126 	 */
2127 	sync->ds_bios = malloc(sizeof(struct bio *) * g_mirror_syncreqs,
2128 	    M_MIRROR, M_WAITOK);
2129 	for (i = 0; i < g_mirror_syncreqs; i++) {
2130 		bp = g_alloc_bio();
2131 		sync->ds_bios[i] = bp;
2132 
2133 		bp->bio_data = malloc(maxphys, M_MIRROR, M_WAITOK);
2134 		bp->bio_caller1 = (void *)(uintptr_t)i;
2135 		g_mirror_sync_reinit(disk, bp, sync->ds_offset);
2136 		sync->ds_offset += bp->bio_length;
2137 	}
2138 
2139 	/* Increase the number of disks in SYNCHRONIZING state. */
2140 	sc->sc_sync.ds_ndisks++;
2141 	/* Set the number of in-flight synchronization requests. */
2142 	sync->ds_inflight = g_mirror_syncreqs;
2143 
2144 	/*
2145 	 * Fire off first synchronization requests.
2146 	 */
2147 	for (i = 0; i < g_mirror_syncreqs; i++) {
2148 		bp = sync->ds_bios[i];
2149 		G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
2150 		sync->ds_consumer->index++;
2151 		/*
2152 		 * Delay the request if it is colliding with a regular request.
2153 		 */
2154 		if (g_mirror_regular_collision(sc, bp))
2155 			g_mirror_sync_delay(sc, bp);
2156 		else
2157 			g_io_request(bp, sync->ds_consumer);
2158 	}
2159 }
2160 
2161 /*
2162  * Stop synchronization process.
2163  * type: 0 - synchronization finished
2164  *       1 - synchronization stopped
2165  */
2166 static void
2167 g_mirror_sync_stop(struct g_mirror_disk *disk, int type)
2168 {
2169 	struct g_mirror_softc *sc;
2170 	struct g_consumer *cp;
2171 
2172 	g_topology_assert_not();
2173 	sc = disk->d_softc;
2174 	sx_assert(&sc->sc_lock, SX_LOCKED);
2175 
2176 	KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2177 	    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2178 	    g_mirror_disk_state2str(disk->d_state)));
2179 	if (disk->d_sync.ds_consumer == NULL)
2180 		return;
2181 
2182 	if (type == 0) {
2183 		G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s finished.",
2184 		    sc->sc_name, g_mirror_get_diskname(disk));
2185 	} else /* if (type == 1) */ {
2186 		G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s stopped.",
2187 		    sc->sc_name, g_mirror_get_diskname(disk));
2188 	}
2189 	g_mirror_regular_release(sc);
2190 	free(disk->d_sync.ds_bios, M_MIRROR);
2191 	disk->d_sync.ds_bios = NULL;
2192 	cp = disk->d_sync.ds_consumer;
2193 	disk->d_sync.ds_consumer = NULL;
2194 	disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2195 	sc->sc_sync.ds_ndisks--;
2196 	sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
2197 	g_topology_lock();
2198 	g_mirror_kill_consumer(sc, cp);
2199 	g_topology_unlock();
2200 	sx_xlock(&sc->sc_lock);
2201 }
2202 
2203 static void
2204 g_mirror_launch_provider(struct g_mirror_softc *sc)
2205 {
2206 	struct g_mirror_disk *disk;
2207 	struct g_provider *pp, *dp;
2208 
2209 	sx_assert(&sc->sc_lock, SX_LOCKED);
2210 
2211 	g_topology_lock();
2212 	pp = g_new_providerf(sc->sc_geom, "mirror/%s", sc->sc_name);
2213 	pp->flags |= G_PF_DIRECT_RECEIVE;
2214 	pp->mediasize = sc->sc_mediasize;
2215 	pp->sectorsize = sc->sc_sectorsize;
2216 	pp->stripesize = 0;
2217 	pp->stripeoffset = 0;
2218 
2219 	/* Splitting of unmapped BIO's could work but isn't implemented now */
2220 	if (sc->sc_balance != G_MIRROR_BALANCE_SPLIT)
2221 		pp->flags |= G_PF_ACCEPT_UNMAPPED;
2222 
2223 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2224 		if (disk->d_consumer && disk->d_consumer->provider) {
2225 			dp = disk->d_consumer->provider;
2226 			if (dp->stripesize > pp->stripesize) {
2227 				pp->stripesize = dp->stripesize;
2228 				pp->stripeoffset = dp->stripeoffset;
2229 			}
2230 			/* A provider underneath us doesn't support unmapped */
2231 			if ((dp->flags & G_PF_ACCEPT_UNMAPPED) == 0) {
2232 				G_MIRROR_DEBUG(0, "Cancelling unmapped "
2233 				    "because of %s.", dp->name);
2234 				pp->flags &= ~G_PF_ACCEPT_UNMAPPED;
2235 			}
2236 		}
2237 	}
2238 	pp->private = sc;
2239 	sc->sc_refcnt++;
2240 	sc->sc_provider = pp;
2241 	g_error_provider(pp, 0);
2242 	g_topology_unlock();
2243 	G_MIRROR_DEBUG(0, "Device %s launched (%u/%u).", pp->name,
2244 	    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE), sc->sc_ndisks);
2245 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2246 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2247 			g_mirror_sync_start(disk);
2248 	}
2249 }
2250 
2251 static void
2252 g_mirror_destroy_provider(struct g_mirror_softc *sc)
2253 {
2254 	struct g_mirror_disk *disk;
2255 	struct bio *bp;
2256 
2257 	g_topology_assert_not();
2258 	KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
2259 	    sc->sc_name));
2260 
2261 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2262 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2263 			g_mirror_sync_stop(disk, 1);
2264 	}
2265 
2266 	g_topology_lock();
2267 	g_error_provider(sc->sc_provider, ENXIO);
2268 	mtx_lock(&sc->sc_queue_mtx);
2269 	while ((bp = TAILQ_FIRST(&sc->sc_queue)) != NULL) {
2270 		TAILQ_REMOVE(&sc->sc_queue, bp, bio_queue);
2271 		/*
2272 		 * Abort any pending I/O that wasn't generated by us.
2273 		 * Synchronization requests and requests destined for individual
2274 		 * mirror components can be destroyed immediately.
2275 		 */
2276 		if (bp->bio_to == sc->sc_provider &&
2277 		    bp->bio_from->geom != sc->sc_sync.ds_geom) {
2278 			g_io_deliver(bp, ENXIO);
2279 		} else {
2280 			if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0)
2281 				free(bp->bio_data, M_MIRROR);
2282 			g_destroy_bio(bp);
2283 		}
2284 	}
2285 	mtx_unlock(&sc->sc_queue_mtx);
2286 	g_wither_provider(sc->sc_provider, ENXIO);
2287 	sc->sc_provider = NULL;
2288 	G_MIRROR_DEBUG(0, "Device %s: provider destroyed.", sc->sc_name);
2289 	g_topology_unlock();
2290 }
2291 
2292 static void
2293 g_mirror_go(void *arg)
2294 {
2295 	struct g_mirror_softc *sc;
2296 	struct g_mirror_event *ep;
2297 
2298 	sc = arg;
2299 	G_MIRROR_DEBUG(0, "Force device %s start due to timeout.", sc->sc_name);
2300 	ep = sc->sc_timeout_event;
2301 	sc->sc_timeout_event = NULL;
2302 	g_mirror_event_dispatch(ep, sc, 0,
2303 	    G_MIRROR_EVENT_DONTWAIT | G_MIRROR_EVENT_DEVICE);
2304 }
2305 
2306 static void
2307 g_mirror_timeout_drain(struct g_mirror_softc *sc)
2308 {
2309 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2310 
2311 	callout_drain(&sc->sc_callout);
2312 	g_mirror_event_free(sc->sc_timeout_event);
2313 	sc->sc_timeout_event = NULL;
2314 }
2315 
2316 static u_int
2317 g_mirror_determine_state(struct g_mirror_disk *disk)
2318 {
2319 	struct g_mirror_softc *sc;
2320 	u_int state;
2321 
2322 	sc = disk->d_softc;
2323 	if (sc->sc_syncid == disk->d_sync.ds_syncid) {
2324 		if ((disk->d_flags &
2325 		    G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0 &&
2326 		    (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 0 ||
2327 		     (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) == 0)) {
2328 			/* Disk does not need synchronization. */
2329 			state = G_MIRROR_DISK_STATE_ACTIVE;
2330 		} else {
2331 			if ((sc->sc_flags &
2332 			     G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2333 			    (disk->d_flags &
2334 			     G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2335 				/*
2336 				 * We can start synchronization from
2337 				 * the stored offset.
2338 				 */
2339 				state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2340 			} else {
2341 				state = G_MIRROR_DISK_STATE_STALE;
2342 			}
2343 		}
2344 	} else if (disk->d_sync.ds_syncid < sc->sc_syncid) {
2345 		/*
2346 		 * Reset all synchronization data for this disk,
2347 		 * because if it even was synchronized, it was
2348 		 * synchronized to disks with different syncid.
2349 		 */
2350 		disk->d_flags |= G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2351 		disk->d_sync.ds_offset = 0;
2352 		disk->d_sync.ds_offset_done = 0;
2353 		disk->d_sync.ds_syncid = sc->sc_syncid;
2354 		if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2355 		    (disk->d_flags & G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2356 			state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2357 		} else {
2358 			state = G_MIRROR_DISK_STATE_STALE;
2359 		}
2360 	} else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
2361 		/*
2362 		 * Not good, NOT GOOD!
2363 		 * It means that mirror was started on stale disks
2364 		 * and more fresh disk just arrive.
2365 		 * If there were writes, mirror is broken, sorry.
2366 		 * I think the best choice here is don't touch
2367 		 * this disk and inform the user loudly.
2368 		 */
2369 		G_MIRROR_DEBUG(0, "Device %s was started before the freshest "
2370 		    "disk (%s) arrives!! It will not be connected to the "
2371 		    "running device.", sc->sc_name,
2372 		    g_mirror_get_diskname(disk));
2373 		g_mirror_destroy_disk(disk);
2374 		state = G_MIRROR_DISK_STATE_NONE;
2375 		/* Return immediately, because disk was destroyed. */
2376 		return (state);
2377 	}
2378 	G_MIRROR_DEBUG(3, "State for %s disk: %s.",
2379 	    g_mirror_get_diskname(disk), g_mirror_disk_state2str(state));
2380 	return (state);
2381 }
2382 
2383 /*
2384  * Update device state.
2385  */
2386 static void
2387 g_mirror_update_device(struct g_mirror_softc *sc, bool force)
2388 {
2389 	struct g_mirror_disk *disk;
2390 	u_int state;
2391 
2392 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2393 
2394 	switch (sc->sc_state) {
2395 	case G_MIRROR_DEVICE_STATE_STARTING:
2396 	    {
2397 		struct g_mirror_disk *pdisk, *tdisk;
2398 		const char *mismatch;
2399 		uintmax_t found, newest;
2400 		u_int dirty, ndisks;
2401 
2402 		/* Pre-flight checks */
2403 		LIST_FOREACH_SAFE(disk, &sc->sc_disks, d_next, tdisk) {
2404 			/*
2405 			 * Confirm we already detected the newest genid.
2406 			 */
2407 			KASSERT(sc->sc_genid >= disk->d_genid,
2408 			    ("%s: found newer genid %u (sc:%p had %u).", __func__,
2409 			    disk->d_genid, sc, sc->sc_genid));
2410 
2411 			/* Kick out any previously tasted stale components. */
2412 			if (disk->d_genid < sc->sc_genid) {
2413 				G_MIRROR_DEBUG(0, "Stale 'genid' field on %s "
2414 				    "(device %s) (component=%u latest=%u), skipping.",
2415 				    g_mirror_get_diskname(disk), sc->sc_name,
2416 				    disk->d_genid, sc->sc_genid);
2417 				g_mirror_destroy_disk(disk);
2418 				sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2419 				continue;
2420 			}
2421 
2422 			/*
2423 			 * Confirm we already detected the newest syncid.
2424 			 */
2425 			KASSERT(sc->sc_syncid >= disk->d_sync.ds_syncid,
2426 			    ("%s: found newer syncid %u (sc:%p had %u).",
2427 			     __func__, disk->d_sync.ds_syncid, sc,
2428 			     sc->sc_syncid));
2429 
2430 #define DETECT_MISMATCH(field, name) \
2431 			if (mismatch == NULL &&					\
2432 			    disk->d_init_ ## field != sc->sc_ ## field) {	\
2433 				mismatch = name;				\
2434 				found = (intmax_t)disk->d_init_ ## field;	\
2435 				newest = (intmax_t)sc->sc_ ## field;		\
2436 			}
2437 			mismatch = NULL;
2438 			DETECT_MISMATCH(ndisks, "md_all");
2439 			DETECT_MISMATCH(balance, "md_balance");
2440 			DETECT_MISMATCH(slice, "md_slice");
2441 			DETECT_MISMATCH(mediasize, "md_mediasize");
2442 #undef DETECT_MISMATCH
2443 			if (mismatch != NULL) {
2444 				G_MIRROR_DEBUG(0, "Found a mismatching '%s' "
2445 				    "field on %s (device %s) (found=%ju "
2446 				    "newest=%ju).", mismatch,
2447 				    g_mirror_get_diskname(disk), sc->sc_name,
2448 				    found, newest);
2449 				g_mirror_destroy_disk(disk);
2450 				sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2451 				continue;
2452 			}
2453 		}
2454 
2455 		KASSERT(sc->sc_provider == NULL,
2456 		    ("Non-NULL provider in STARTING state (%s).", sc->sc_name));
2457 		/*
2458 		 * Are we ready? If the timeout (force is true) has expired, and
2459 		 * any disks are present, then yes. If we're permitted to launch
2460 		 * before the timeout has expired and the expected number of
2461 		 * current-generation mirror disks have been tasted, then yes.
2462 		 */
2463 		ndisks = g_mirror_ndisks(sc, -1);
2464 		if ((force && ndisks > 0) ||
2465 		    (g_launch_mirror_before_timeout && ndisks == sc->sc_ndisks)) {
2466 			;
2467 		} else if (ndisks == 0) {
2468 			/*
2469 			 * Disks went down in starting phase, so destroy
2470 			 * device.
2471 			 */
2472 			g_mirror_timeout_drain(sc);
2473 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2474 			G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2475 			    sc->sc_rootmount);
2476 			root_mount_rel(sc->sc_rootmount);
2477 			sc->sc_rootmount = NULL;
2478 			return;
2479 		} else {
2480 			return;
2481 		}
2482 
2483 		/*
2484 		 * Activate all disks with the biggest syncid.
2485 		 */
2486 		if (force) {
2487 			/*
2488 			 * If 'force' is true, we have been called due to
2489 			 * timeout, so don't bother canceling timeout.
2490 			 */
2491 			ndisks = 0;
2492 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2493 				if ((disk->d_flags &
2494 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2495 					ndisks++;
2496 				}
2497 			}
2498 			if (ndisks == 0) {
2499 				/* No valid disks found, destroy device. */
2500 				sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2501 				G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2502 				    __LINE__, sc->sc_rootmount);
2503 				root_mount_rel(sc->sc_rootmount);
2504 				sc->sc_rootmount = NULL;
2505 				return;
2506 			}
2507 		} else {
2508 			/* Cancel timeout. */
2509 			g_mirror_timeout_drain(sc);
2510 		}
2511 
2512 		/*
2513 		 * Here we need to look for dirty disks and if all disks
2514 		 * with the biggest syncid are dirty, we have to choose
2515 		 * one with the biggest priority and rebuild the rest.
2516 		 */
2517 		/*
2518 		 * Find the number of dirty disks with the biggest syncid.
2519 		 * Find the number of disks with the biggest syncid.
2520 		 * While here, find a disk with the biggest priority.
2521 		 */
2522 		dirty = ndisks = 0;
2523 		pdisk = NULL;
2524 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2525 			if (disk->d_sync.ds_syncid != sc->sc_syncid)
2526 				continue;
2527 			if ((disk->d_flags &
2528 			    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2529 				continue;
2530 			}
2531 			ndisks++;
2532 			if ((disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
2533 				dirty++;
2534 				if (pdisk == NULL ||
2535 				    pdisk->d_priority < disk->d_priority) {
2536 					pdisk = disk;
2537 				}
2538 			}
2539 		}
2540 		if (dirty == 0) {
2541 			/* No dirty disks at all, great. */
2542 		} else if (dirty == ndisks) {
2543 			/*
2544 			 * Force synchronization for all dirty disks except one
2545 			 * with the biggest priority.
2546 			 */
2547 			KASSERT(pdisk != NULL, ("pdisk == NULL"));
2548 			G_MIRROR_DEBUG(1, "Using disk %s (device %s) as a "
2549 			    "master disk for synchronization.",
2550 			    g_mirror_get_diskname(pdisk), sc->sc_name);
2551 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2552 				if (disk->d_sync.ds_syncid != sc->sc_syncid)
2553 					continue;
2554 				if ((disk->d_flags &
2555 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2556 					continue;
2557 				}
2558 				KASSERT((disk->d_flags &
2559 				    G_MIRROR_DISK_FLAG_DIRTY) != 0,
2560 				    ("Disk %s isn't marked as dirty.",
2561 				    g_mirror_get_diskname(disk)));
2562 				/* Skip the disk with the biggest priority. */
2563 				if (disk == pdisk)
2564 					continue;
2565 				disk->d_sync.ds_syncid = 0;
2566 			}
2567 		} else if (dirty < ndisks) {
2568 			/*
2569 			 * Force synchronization for all dirty disks.
2570 			 * We have some non-dirty disks.
2571 			 */
2572 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2573 				if (disk->d_sync.ds_syncid != sc->sc_syncid)
2574 					continue;
2575 				if ((disk->d_flags &
2576 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2577 					continue;
2578 				}
2579 				if ((disk->d_flags &
2580 				    G_MIRROR_DISK_FLAG_DIRTY) == 0) {
2581 					continue;
2582 				}
2583 				disk->d_sync.ds_syncid = 0;
2584 			}
2585 		}
2586 
2587 		/* Reset hint. */
2588 		sc->sc_hint = NULL;
2589 		if (force) {
2590 			/* Remember to bump syncid on first write. */
2591 			sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2592 		}
2593 		state = G_MIRROR_DEVICE_STATE_RUNNING;
2594 		G_MIRROR_DEBUG(1, "Device %s state changed from %s to %s.",
2595 		    sc->sc_name, g_mirror_device_state2str(sc->sc_state),
2596 		    g_mirror_device_state2str(state));
2597 		sc->sc_state = state;
2598 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2599 			state = g_mirror_determine_state(disk);
2600 			g_mirror_event_send(disk, state,
2601 			    G_MIRROR_EVENT_DONTWAIT);
2602 			if (state == G_MIRROR_DISK_STATE_STALE)
2603 				sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2604 		}
2605 		break;
2606 	    }
2607 	case G_MIRROR_DEVICE_STATE_RUNNING:
2608 		if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 0 &&
2609 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2610 			/*
2611 			 * No usable disks, so destroy the device.
2612 			 */
2613 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2614 			break;
2615 		} else if (g_mirror_ndisks(sc,
2616 		    G_MIRROR_DISK_STATE_ACTIVE) > 0 &&
2617 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2618 			/*
2619 			 * We have active disks, launch provider if it doesn't
2620 			 * exist.
2621 			 */
2622 			if (sc->sc_provider == NULL)
2623 				g_mirror_launch_provider(sc);
2624 			if (sc->sc_rootmount != NULL) {
2625 				G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2626 				    __LINE__, sc->sc_rootmount);
2627 				root_mount_rel(sc->sc_rootmount);
2628 				sc->sc_rootmount = NULL;
2629 			}
2630 		}
2631 		/*
2632 		 * Genid should be bumped immediately, so do it here.
2633 		 */
2634 		if ((sc->sc_bump_id & G_MIRROR_BUMP_GENID) != 0) {
2635 			sc->sc_bump_id &= ~G_MIRROR_BUMP_GENID;
2636 			g_mirror_bump_genid(sc);
2637 		}
2638 		if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID_NOW) != 0) {
2639 			sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID_NOW;
2640 			g_mirror_bump_syncid(sc);
2641 		}
2642 		break;
2643 	default:
2644 		KASSERT(1 == 0, ("Wrong device state (%s, %s).",
2645 		    sc->sc_name, g_mirror_device_state2str(sc->sc_state)));
2646 		break;
2647 	}
2648 }
2649 
2650 /*
2651  * Update disk state and device state if needed.
2652  */
2653 #define	DISK_STATE_CHANGED()	G_MIRROR_DEBUG(1,			\
2654 	"Disk %s state changed from %s to %s (device %s).",		\
2655 	g_mirror_get_diskname(disk),					\
2656 	g_mirror_disk_state2str(disk->d_state),				\
2657 	g_mirror_disk_state2str(state), sc->sc_name)
2658 static int
2659 g_mirror_update_disk(struct g_mirror_disk *disk, u_int state)
2660 {
2661 	struct g_mirror_softc *sc;
2662 
2663 	sc = disk->d_softc;
2664 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2665 
2666 again:
2667 	G_MIRROR_DEBUG(3, "Changing disk %s state from %s to %s.",
2668 	    g_mirror_get_diskname(disk), g_mirror_disk_state2str(disk->d_state),
2669 	    g_mirror_disk_state2str(state));
2670 	switch (state) {
2671 	case G_MIRROR_DISK_STATE_NEW:
2672 		/*
2673 		 * Possible scenarios:
2674 		 * 1. New disk arrive.
2675 		 */
2676 		/* Previous state should be NONE. */
2677 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NONE,
2678 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2679 		    g_mirror_disk_state2str(disk->d_state)));
2680 		DISK_STATE_CHANGED();
2681 
2682 		disk->d_state = state;
2683 		g_topology_lock();
2684 		if (LIST_EMPTY(&sc->sc_disks))
2685 			LIST_INSERT_HEAD(&sc->sc_disks, disk, d_next);
2686 		else {
2687 			struct g_mirror_disk *dp;
2688 
2689 			LIST_FOREACH(dp, &sc->sc_disks, d_next) {
2690 				if (disk->d_priority >= dp->d_priority) {
2691 					LIST_INSERT_BEFORE(dp, disk, d_next);
2692 					dp = NULL;
2693 					break;
2694 				}
2695 				if (LIST_NEXT(dp, d_next) == NULL)
2696 					break;
2697 			}
2698 			if (dp != NULL)
2699 				LIST_INSERT_AFTER(dp, disk, d_next);
2700 		}
2701 		g_topology_unlock();
2702 		G_MIRROR_DEBUG(1, "Device %s: provider %s detected.",
2703 		    sc->sc_name, g_mirror_get_diskname(disk));
2704 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
2705 			break;
2706 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2707 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2708 		    g_mirror_device_state2str(sc->sc_state),
2709 		    g_mirror_get_diskname(disk),
2710 		    g_mirror_disk_state2str(disk->d_state)));
2711 		state = g_mirror_determine_state(disk);
2712 		if (state != G_MIRROR_DISK_STATE_NONE)
2713 			goto again;
2714 		break;
2715 	case G_MIRROR_DISK_STATE_ACTIVE:
2716 		/*
2717 		 * Possible scenarios:
2718 		 * 1. New disk does not need synchronization.
2719 		 * 2. Synchronization process finished successfully.
2720 		 */
2721 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2722 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2723 		    g_mirror_device_state2str(sc->sc_state),
2724 		    g_mirror_get_diskname(disk),
2725 		    g_mirror_disk_state2str(disk->d_state)));
2726 		/* Previous state should be NEW or SYNCHRONIZING. */
2727 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW ||
2728 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2729 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2730 		    g_mirror_disk_state2str(disk->d_state)));
2731 		DISK_STATE_CHANGED();
2732 
2733 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
2734 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2735 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_FORCE_SYNC;
2736 			g_mirror_sync_stop(disk, 0);
2737 		}
2738 		disk->d_state = state;
2739 		disk->d_sync.ds_offset = 0;
2740 		disk->d_sync.ds_offset_done = 0;
2741 		g_mirror_update_idle(sc, disk);
2742 		g_mirror_update_metadata(disk);
2743 		G_MIRROR_DEBUG(1, "Device %s: provider %s activated.",
2744 		    sc->sc_name, g_mirror_get_diskname(disk));
2745 		break;
2746 	case G_MIRROR_DISK_STATE_STALE:
2747 		/*
2748 		 * Possible scenarios:
2749 		 * 1. Stale disk was connected.
2750 		 */
2751 		/* Previous state should be NEW. */
2752 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2753 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2754 		    g_mirror_disk_state2str(disk->d_state)));
2755 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2756 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2757 		    g_mirror_device_state2str(sc->sc_state),
2758 		    g_mirror_get_diskname(disk),
2759 		    g_mirror_disk_state2str(disk->d_state)));
2760 		/*
2761 		 * STALE state is only possible if device is marked
2762 		 * NOAUTOSYNC.
2763 		 */
2764 		KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) != 0,
2765 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2766 		    g_mirror_device_state2str(sc->sc_state),
2767 		    g_mirror_get_diskname(disk),
2768 		    g_mirror_disk_state2str(disk->d_state)));
2769 		DISK_STATE_CHANGED();
2770 
2771 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2772 		disk->d_state = state;
2773 		g_mirror_update_metadata(disk);
2774 		G_MIRROR_DEBUG(0, "Device %s: provider %s is stale.",
2775 		    sc->sc_name, g_mirror_get_diskname(disk));
2776 		break;
2777 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
2778 		/*
2779 		 * Possible scenarios:
2780 		 * 1. Disk which needs synchronization was connected.
2781 		 */
2782 		/* Previous state should be NEW. */
2783 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2784 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2785 		    g_mirror_disk_state2str(disk->d_state)));
2786 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2787 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2788 		    g_mirror_device_state2str(sc->sc_state),
2789 		    g_mirror_get_diskname(disk),
2790 		    g_mirror_disk_state2str(disk->d_state)));
2791 		DISK_STATE_CHANGED();
2792 
2793 		if (disk->d_state == G_MIRROR_DISK_STATE_NEW)
2794 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2795 		disk->d_state = state;
2796 		if (sc->sc_provider != NULL) {
2797 			g_mirror_sync_start(disk);
2798 			g_mirror_update_metadata(disk);
2799 		}
2800 		break;
2801 	case G_MIRROR_DISK_STATE_DISCONNECTED:
2802 		/*
2803 		 * Possible scenarios:
2804 		 * 1. Device wasn't running yet, but disk disappear.
2805 		 * 2. Disk was active and disapppear.
2806 		 * 3. Disk disappear during synchronization process.
2807 		 */
2808 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING) {
2809 			/*
2810 			 * Previous state should be ACTIVE, STALE or
2811 			 * SYNCHRONIZING.
2812 			 */
2813 			KASSERT(disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
2814 			    disk->d_state == G_MIRROR_DISK_STATE_STALE ||
2815 			    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2816 			    ("Wrong disk state (%s, %s).",
2817 			    g_mirror_get_diskname(disk),
2818 			    g_mirror_disk_state2str(disk->d_state)));
2819 		} else if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING) {
2820 			/* Previous state should be NEW. */
2821 			KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2822 			    ("Wrong disk state (%s, %s).",
2823 			    g_mirror_get_diskname(disk),
2824 			    g_mirror_disk_state2str(disk->d_state)));
2825 			/*
2826 			 * Reset bumping syncid if disk disappeared in STARTING
2827 			 * state.
2828 			 */
2829 			if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0)
2830 				sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
2831 #ifdef	INVARIANTS
2832 		} else {
2833 			KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
2834 			    sc->sc_name,
2835 			    g_mirror_device_state2str(sc->sc_state),
2836 			    g_mirror_get_diskname(disk),
2837 			    g_mirror_disk_state2str(disk->d_state)));
2838 #endif
2839 		}
2840 		DISK_STATE_CHANGED();
2841 		G_MIRROR_DEBUG(0, "Device %s: provider %s disconnected.",
2842 		    sc->sc_name, g_mirror_get_diskname(disk));
2843 
2844 		g_mirror_destroy_disk(disk);
2845 		break;
2846 	case G_MIRROR_DISK_STATE_DESTROY:
2847 	    {
2848 		int error;
2849 
2850 		error = g_mirror_clear_metadata(disk);
2851 		if (error != 0) {
2852 			G_MIRROR_DEBUG(0,
2853 			    "Device %s: failed to clear metadata on %s: %d.",
2854 			    sc->sc_name, g_mirror_get_diskname(disk), error);
2855 			break;
2856 		}
2857 		DISK_STATE_CHANGED();
2858 		G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.",
2859 		    sc->sc_name, g_mirror_get_diskname(disk));
2860 
2861 		g_mirror_destroy_disk(disk);
2862 		sc->sc_ndisks--;
2863 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2864 			g_mirror_update_metadata(disk);
2865 		}
2866 		break;
2867 	    }
2868 	default:
2869 		KASSERT(1 == 0, ("Unknown state (%u).", state));
2870 		break;
2871 	}
2872 	return (0);
2873 }
2874 #undef	DISK_STATE_CHANGED
2875 
2876 int
2877 g_mirror_read_metadata(struct g_consumer *cp, struct g_mirror_metadata *md)
2878 {
2879 	struct g_provider *pp;
2880 	u_char *buf;
2881 	int error;
2882 
2883 	g_topology_assert();
2884 
2885 	error = g_access(cp, 1, 0, 0);
2886 	if (error != 0)
2887 		return (error);
2888 	pp = cp->provider;
2889 	g_topology_unlock();
2890 	/* Metadata are stored on last sector. */
2891 	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
2892 	    &error);
2893 	g_topology_lock();
2894 	g_access(cp, -1, 0, 0);
2895 	if (buf == NULL) {
2896 		G_MIRROR_DEBUG(1, "Cannot read metadata from %s (error=%d).",
2897 		    cp->provider->name, error);
2898 		return (error);
2899 	}
2900 
2901 	/* Decode metadata. */
2902 	error = mirror_metadata_decode(buf, md);
2903 	g_free(buf);
2904 	if (strcmp(md->md_magic, G_MIRROR_MAGIC) != 0)
2905 		return (EINVAL);
2906 	if (md->md_version > G_MIRROR_VERSION) {
2907 		G_MIRROR_DEBUG(0,
2908 		    "Kernel module is too old to handle metadata from %s.",
2909 		    cp->provider->name);
2910 		return (EINVAL);
2911 	}
2912 	if (error != 0) {
2913 		G_MIRROR_DEBUG(1, "MD5 metadata hash mismatch for provider %s.",
2914 		    cp->provider->name);
2915 		return (error);
2916 	}
2917 
2918 	return (0);
2919 }
2920 
2921 static int
2922 g_mirror_check_metadata(struct g_mirror_softc *sc, struct g_provider *pp,
2923     struct g_mirror_metadata *md)
2924 {
2925 
2926 	G_MIRROR_DEBUG(2, "%s: md_did 0x%u disk %s device %s md_all 0x%x "
2927 	    "sc_ndisks 0x%x md_slice 0x%x sc_slice 0x%x md_balance 0x%x "
2928 	    "sc_balance 0x%x sc_mediasize 0x%jx pp_mediasize 0x%jx "
2929 	    "md_sectorsize 0x%x sc_sectorsize 0x%x md_mflags 0x%jx "
2930 	    "md_dflags 0x%jx md_syncid 0x%x md_genid 0x%x md_priority 0x%x "
2931 	    "sc_state 0x%x.",
2932 	    __func__, md->md_did, pp->name, sc->sc_name, md->md_all,
2933 	    sc->sc_ndisks, md->md_slice, sc->sc_slice, md->md_balance,
2934 	    sc->sc_balance, (uintmax_t)sc->sc_mediasize,
2935 	    (uintmax_t)pp->mediasize, md->md_sectorsize, sc->sc_sectorsize,
2936 	    (uintmax_t)md->md_mflags, (uintmax_t)md->md_dflags, md->md_syncid,
2937 	    md->md_genid, md->md_priority, sc->sc_state);
2938 
2939 	if (g_mirror_id2disk(sc, md->md_did) != NULL) {
2940 		G_MIRROR_DEBUG(1, "Disk %s (id=%u) already exists, skipping.",
2941 		    pp->name, md->md_did);
2942 		return (EEXIST);
2943 	}
2944 	if (sc->sc_mediasize > pp->mediasize) {
2945 		G_MIRROR_DEBUG(1,
2946 		    "Invalid size of disk %s (device %s), skipping.", pp->name,
2947 		    sc->sc_name);
2948 		return (EINVAL);
2949 	}
2950 	if (md->md_sectorsize != sc->sc_sectorsize) {
2951 		G_MIRROR_DEBUG(1,
2952 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2953 		    "md_sectorsize", pp->name, sc->sc_name);
2954 		return (EINVAL);
2955 	}
2956 	if ((sc->sc_sectorsize % pp->sectorsize) != 0) {
2957 		G_MIRROR_DEBUG(1,
2958 		    "Invalid sector size of disk %s (device %s), skipping.",
2959 		    pp->name, sc->sc_name);
2960 		return (EINVAL);
2961 	}
2962 	if ((md->md_mflags & ~G_MIRROR_DEVICE_FLAG_MASK) != 0) {
2963 		G_MIRROR_DEBUG(1,
2964 		    "Invalid device flags on disk %s (device %s), skipping.",
2965 		    pp->name, sc->sc_name);
2966 		return (EINVAL);
2967 	}
2968 	if ((md->md_dflags & ~G_MIRROR_DISK_FLAG_MASK) != 0) {
2969 		G_MIRROR_DEBUG(1,
2970 		    "Invalid disk flags on disk %s (device %s), skipping.",
2971 		    pp->name, sc->sc_name);
2972 		return (EINVAL);
2973 	}
2974 	return (0);
2975 }
2976 
2977 int
2978 g_mirror_add_disk(struct g_mirror_softc *sc, struct g_provider *pp,
2979     struct g_mirror_metadata *md)
2980 {
2981 	struct g_mirror_disk *disk;
2982 	int error;
2983 
2984 	g_topology_assert_not();
2985 	G_MIRROR_DEBUG(2, "Adding disk %s.", pp->name);
2986 
2987 	error = g_mirror_check_metadata(sc, pp, md);
2988 	if (error != 0)
2989 		return (error);
2990 
2991 	if (md->md_genid < sc->sc_genid) {
2992 		G_MIRROR_DEBUG(0, "Component %s (device %s) broken, skipping.",
2993 		    pp->name, sc->sc_name);
2994 		return (EINVAL);
2995 	}
2996 
2997 	/*
2998 	 * If the component disk we're tasting has newer metadata than the
2999 	 * STARTING gmirror device, refresh the device from the component.
3000 	 */
3001 	error = g_mirror_refresh_device(sc, pp, md);
3002 	if (error != 0)
3003 		return (error);
3004 
3005 	disk = g_mirror_init_disk(sc, pp, md, &error);
3006 	if (disk == NULL)
3007 		return (error);
3008 	error = g_mirror_event_send(disk, G_MIRROR_DISK_STATE_NEW,
3009 	    G_MIRROR_EVENT_WAIT);
3010 	if (error != 0)
3011 		return (error);
3012 	if (md->md_version < G_MIRROR_VERSION) {
3013 		G_MIRROR_DEBUG(0, "Upgrading metadata on %s (v%d->v%d).",
3014 		    pp->name, md->md_version, G_MIRROR_VERSION);
3015 		g_mirror_update_metadata(disk);
3016 	}
3017 	return (0);
3018 }
3019 
3020 static void
3021 g_mirror_destroy_delayed(void *arg, int flag)
3022 {
3023 	struct g_mirror_softc *sc;
3024 	int error;
3025 
3026 	if (flag == EV_CANCEL) {
3027 		G_MIRROR_DEBUG(1, "Destroying canceled.");
3028 		return;
3029 	}
3030 	sc = arg;
3031 	g_topology_unlock();
3032 	sx_xlock(&sc->sc_lock);
3033 	KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) == 0,
3034 	    ("DESTROY flag set on %s.", sc->sc_name));
3035 	KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_CLOSEWAIT) != 0,
3036 	    ("CLOSEWAIT flag not set on %s.", sc->sc_name));
3037 	G_MIRROR_DEBUG(1, "Destroying %s (delayed).", sc->sc_name);
3038 	error = g_mirror_destroy(sc, G_MIRROR_DESTROY_SOFT);
3039 	if (error != 0) {
3040 		G_MIRROR_DEBUG(0, "Cannot destroy %s (error=%d).",
3041 		    sc->sc_name, error);
3042 		sx_xunlock(&sc->sc_lock);
3043 	}
3044 	g_topology_lock();
3045 }
3046 
3047 static int
3048 g_mirror_access(struct g_provider *pp, int acr, int acw, int ace)
3049 {
3050 	struct g_mirror_softc *sc;
3051 	int error = 0;
3052 
3053 	g_topology_assert();
3054 	G_MIRROR_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, acr,
3055 	    acw, ace);
3056 
3057 	sc = pp->private;
3058 	KASSERT(sc != NULL, ("NULL softc (provider=%s).", pp->name));
3059 
3060 	g_topology_unlock();
3061 	sx_xlock(&sc->sc_lock);
3062 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0 ||
3063 	    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_CLOSEWAIT) != 0 ||
3064 	    LIST_EMPTY(&sc->sc_disks)) {
3065 		if (acr > 0 || acw > 0 || ace > 0)
3066 			error = ENXIO;
3067 		goto end;
3068 	}
3069 	sc->sc_provider_open += acr + acw + ace;
3070 	if (pp->acw + acw == 0)
3071 		g_mirror_idle(sc, 0);
3072 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_CLOSEWAIT) != 0 &&
3073 	    sc->sc_provider_open == 0)
3074 		g_post_event(g_mirror_destroy_delayed, sc, M_WAITOK, sc, NULL);
3075 end:
3076 	sx_xunlock(&sc->sc_lock);
3077 	g_topology_lock();
3078 	return (error);
3079 }
3080 
3081 static void
3082 g_mirror_reinit_from_metadata(struct g_mirror_softc *sc,
3083     const struct g_mirror_metadata *md)
3084 {
3085 
3086 	sc->sc_genid = md->md_genid;
3087 	sc->sc_syncid = md->md_syncid;
3088 
3089 	sc->sc_slice = md->md_slice;
3090 	sc->sc_balance = md->md_balance;
3091 	sc->sc_mediasize = md->md_mediasize;
3092 	sc->sc_ndisks = md->md_all;
3093 	sc->sc_flags &= ~G_MIRROR_DEVICE_FLAG_MASK;
3094 	sc->sc_flags |= (md->md_mflags & G_MIRROR_DEVICE_FLAG_MASK);
3095 }
3096 
3097 struct g_geom *
3098 g_mirror_create(struct g_class *mp, const struct g_mirror_metadata *md,
3099     u_int type)
3100 {
3101 	struct g_mirror_softc *sc;
3102 	struct g_geom *gp;
3103 	int error, timeout;
3104 
3105 	g_topology_assert();
3106 	G_MIRROR_DEBUG(1, "Creating device %s (id=%u).", md->md_name,
3107 	    md->md_mid);
3108 
3109 	/* One disk is minimum. */
3110 	if (md->md_all < 1)
3111 		return (NULL);
3112 	/*
3113 	 * Action geom.
3114 	 */
3115 	gp = g_new_geomf(mp, "%s", md->md_name);
3116 	sc = malloc(sizeof(*sc), M_MIRROR, M_WAITOK | M_ZERO);
3117 	gp->start = g_mirror_start;
3118 	gp->orphan = g_mirror_orphan;
3119 	gp->access = g_mirror_access;
3120 	gp->dumpconf = g_mirror_dumpconf;
3121 
3122 	sc->sc_type = type;
3123 	sc->sc_id = md->md_mid;
3124 	g_mirror_reinit_from_metadata(sc, md);
3125 	sc->sc_sectorsize = md->md_sectorsize;
3126 	sc->sc_bump_id = 0;
3127 	sc->sc_idle = 1;
3128 	sc->sc_last_write = time_uptime;
3129 	sc->sc_writes = 0;
3130 	sc->sc_refcnt = 1;
3131 	sx_init(&sc->sc_lock, "gmirror:lock");
3132 	TAILQ_INIT(&sc->sc_queue);
3133 	mtx_init(&sc->sc_queue_mtx, "gmirror:queue", NULL, MTX_DEF);
3134 	TAILQ_INIT(&sc->sc_regular_delayed);
3135 	TAILQ_INIT(&sc->sc_inflight);
3136 	TAILQ_INIT(&sc->sc_sync_delayed);
3137 	LIST_INIT(&sc->sc_disks);
3138 	TAILQ_INIT(&sc->sc_events);
3139 	mtx_init(&sc->sc_events_mtx, "gmirror:events", NULL, MTX_DEF);
3140 	callout_init(&sc->sc_callout, 1);
3141 	mtx_init(&sc->sc_done_mtx, "gmirror:done", NULL, MTX_DEF);
3142 	sc->sc_state = G_MIRROR_DEVICE_STATE_STARTING;
3143 	gp->softc = sc;
3144 	sc->sc_geom = gp;
3145 	sc->sc_provider = NULL;
3146 	sc->sc_provider_open = 0;
3147 	/*
3148 	 * Synchronization geom.
3149 	 */
3150 	gp = g_new_geomf(mp, "%s.sync", md->md_name);
3151 	gp->softc = sc;
3152 	gp->orphan = g_mirror_orphan;
3153 	sc->sc_sync.ds_geom = gp;
3154 	sc->sc_sync.ds_ndisks = 0;
3155 	error = kproc_create(g_mirror_worker, sc, &sc->sc_worker, 0, 0,
3156 	    "g_mirror %s", md->md_name);
3157 	if (error != 0) {
3158 		G_MIRROR_DEBUG(1, "Cannot create kernel thread for %s.",
3159 		    sc->sc_name);
3160 		g_destroy_geom(sc->sc_sync.ds_geom);
3161 		g_destroy_geom(sc->sc_geom);
3162 		g_mirror_free_device(sc);
3163 		return (NULL);
3164 	}
3165 
3166 	G_MIRROR_DEBUG(1, "Device %s created (%u components, id=%u).",
3167 	    sc->sc_name, sc->sc_ndisks, sc->sc_id);
3168 
3169 	sc->sc_rootmount = root_mount_hold("GMIRROR");
3170 	G_MIRROR_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
3171 
3172 	/*
3173 	 * Schedule startup timeout.
3174 	 */
3175 	timeout = g_mirror_timeout * hz;
3176 	sc->sc_timeout_event = malloc(sizeof(struct g_mirror_event), M_MIRROR,
3177 	    M_WAITOK);
3178 	callout_reset(&sc->sc_callout, timeout, g_mirror_go, sc);
3179 	return (sc->sc_geom);
3180 }
3181 
3182 int
3183 g_mirror_destroy(struct g_mirror_softc *sc, int how)
3184 {
3185 	struct g_mirror_disk *disk;
3186 
3187 	g_topology_assert_not();
3188 	sx_assert(&sc->sc_lock, SX_XLOCKED);
3189 
3190 	if (sc->sc_provider_open != 0) {
3191 		switch (how) {
3192 		case G_MIRROR_DESTROY_SOFT:
3193 			G_MIRROR_DEBUG(1,
3194 			    "Device %s is still open (%d).", sc->sc_name,
3195 			    sc->sc_provider_open);
3196 			return (EBUSY);
3197 		case G_MIRROR_DESTROY_DELAYED:
3198 			G_MIRROR_DEBUG(1,
3199 			    "Device %s will be destroyed on last close.",
3200 			    sc->sc_name);
3201 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
3202 				if (disk->d_state ==
3203 				    G_MIRROR_DISK_STATE_SYNCHRONIZING) {
3204 					g_mirror_sync_stop(disk, 1);
3205 				}
3206 			}
3207 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_CLOSEWAIT;
3208 			return (EBUSY);
3209 		case G_MIRROR_DESTROY_HARD:
3210 			G_MIRROR_DEBUG(1, "Device %s is still open, so it "
3211 			    "can't be definitely removed.", sc->sc_name);
3212 		}
3213 	}
3214 
3215 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
3216 		sx_xunlock(&sc->sc_lock);
3217 		return (0);
3218 	}
3219 	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
3220 	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DRAIN;
3221 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
3222 	sx_xunlock(&sc->sc_lock);
3223 	mtx_lock(&sc->sc_queue_mtx);
3224 	wakeup(sc);
3225 	mtx_unlock(&sc->sc_queue_mtx);
3226 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
3227 	while (sc->sc_worker != NULL)
3228 		tsleep(&sc->sc_worker, PRIBIO, "m:destroy", hz / 5);
3229 	G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
3230 	sx_xlock(&sc->sc_lock);
3231 	g_mirror_destroy_device(sc);
3232 	return (0);
3233 }
3234 
3235 static void
3236 g_mirror_taste_orphan(struct g_consumer *cp)
3237 {
3238 
3239 	KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
3240 	    cp->provider->name));
3241 }
3242 
3243 static struct g_geom *
3244 g_mirror_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
3245 {
3246 	struct g_mirror_metadata md;
3247 	struct g_mirror_softc *sc;
3248 	struct g_consumer *cp;
3249 	struct g_geom *gp;
3250 	int error;
3251 
3252 	g_topology_assert();
3253 	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
3254 	G_MIRROR_DEBUG(2, "Tasting %s.", pp->name);
3255 
3256 	gp = g_new_geomf(mp, "mirror:taste");
3257 	/*
3258 	 * This orphan function should be never called.
3259 	 */
3260 	gp->orphan = g_mirror_taste_orphan;
3261 	cp = g_new_consumer(gp);
3262 	error = g_attach(cp, pp);
3263 	if (error == 0) {
3264 		error = g_mirror_read_metadata(cp, &md);
3265 		g_detach(cp);
3266 	}
3267 	g_destroy_consumer(cp);
3268 	g_destroy_geom(gp);
3269 	if (error != 0)
3270 		return (NULL);
3271 	gp = NULL;
3272 
3273 	if (md.md_provider[0] != '\0' &&
3274 	    !g_compare_names(md.md_provider, pp->name))
3275 		return (NULL);
3276 	if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
3277 		return (NULL);
3278 	if ((md.md_dflags & G_MIRROR_DISK_FLAG_INACTIVE) != 0) {
3279 		G_MIRROR_DEBUG(0,
3280 		    "Device %s: provider %s marked as inactive, skipping.",
3281 		    md.md_name, pp->name);
3282 		return (NULL);
3283 	}
3284 	if (g_mirror_debug >= 2)
3285 		mirror_metadata_dump(&md);
3286 
3287 	/*
3288 	 * Let's check if device already exists.
3289 	 */
3290 	sc = NULL;
3291 	LIST_FOREACH(gp, &mp->geom, geom) {
3292 		sc = gp->softc;
3293 		if (sc == NULL)
3294 			continue;
3295 		if (sc->sc_type != G_MIRROR_TYPE_AUTOMATIC)
3296 			continue;
3297 		if (sc->sc_sync.ds_geom == gp)
3298 			continue;
3299 		if (strcmp(md.md_name, sc->sc_name) != 0)
3300 			continue;
3301 		if (md.md_mid != sc->sc_id) {
3302 			G_MIRROR_DEBUG(0, "Device %s already configured.",
3303 			    sc->sc_name);
3304 			return (NULL);
3305 		}
3306 		break;
3307 	}
3308 	if (gp == NULL) {
3309 		gp = g_mirror_create(mp, &md, G_MIRROR_TYPE_AUTOMATIC);
3310 		if (gp == NULL) {
3311 			G_MIRROR_DEBUG(0, "Cannot create device %s.",
3312 			    md.md_name);
3313 			return (NULL);
3314 		}
3315 		sc = gp->softc;
3316 	}
3317 	G_MIRROR_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
3318 	g_topology_unlock();
3319 	sx_xlock(&sc->sc_lock);
3320 	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_TASTING;
3321 	error = g_mirror_add_disk(sc, pp, &md);
3322 	sc->sc_flags &= ~G_MIRROR_DEVICE_FLAG_TASTING;
3323 	if (error != 0) {
3324 		G_MIRROR_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
3325 		    pp->name, gp->name, error);
3326 		if (LIST_EMPTY(&sc->sc_disks)) {
3327 			g_cancel_event(sc);
3328 			g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3329 			g_topology_lock();
3330 			return (NULL);
3331 		}
3332 		gp = NULL;
3333 	}
3334 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
3335 		g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3336 		g_topology_lock();
3337 		return (NULL);
3338 	}
3339 	sx_xunlock(&sc->sc_lock);
3340 	g_topology_lock();
3341 	return (gp);
3342 }
3343 
3344 static void
3345 g_mirror_resize(struct g_consumer *cp)
3346 {
3347 	struct g_mirror_disk *disk;
3348 
3349 	g_topology_assert();
3350 	g_trace(G_T_TOPOLOGY, "%s(%s)", __func__, cp->provider->name);
3351 
3352 	disk = cp->private;
3353 	if (disk == NULL)
3354 		return;
3355 	g_topology_unlock();
3356 	g_mirror_update_metadata(disk);
3357 	g_topology_lock();
3358 }
3359 
3360 static int
3361 g_mirror_destroy_geom(struct gctl_req *req __unused,
3362     struct g_class *mp __unused, struct g_geom *gp)
3363 {
3364 	struct g_mirror_softc *sc;
3365 	int error;
3366 
3367 	g_topology_unlock();
3368 	sc = gp->softc;
3369 	sx_xlock(&sc->sc_lock);
3370 	g_cancel_event(sc);
3371 	error = g_mirror_destroy(gp->softc, G_MIRROR_DESTROY_SOFT);
3372 	if (error != 0)
3373 		sx_xunlock(&sc->sc_lock);
3374 	g_topology_lock();
3375 	return (error);
3376 }
3377 
3378 static void
3379 g_mirror_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
3380     struct g_consumer *cp, struct g_provider *pp)
3381 {
3382 	struct g_mirror_softc *sc;
3383 
3384 	g_topology_assert();
3385 
3386 	sc = gp->softc;
3387 	if (sc == NULL)
3388 		return;
3389 	/* Skip synchronization geom. */
3390 	if (gp == sc->sc_sync.ds_geom)
3391 		return;
3392 	if (pp != NULL) {
3393 		/* Nothing here. */
3394 	} else if (cp != NULL) {
3395 		struct g_mirror_disk *disk;
3396 
3397 		disk = cp->private;
3398 		if (disk == NULL)
3399 			return;
3400 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)disk->d_id);
3401 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
3402 			sbuf_printf(sb, "%s<Synchronized>", indent);
3403 			if (disk->d_sync.ds_offset == 0)
3404 				sbuf_cat(sb, "0%");
3405 			else
3406 				sbuf_printf(sb, "%u%%",
3407 				    (u_int)((disk->d_sync.ds_offset * 100) /
3408 				    sc->sc_mediasize));
3409 			sbuf_cat(sb, "</Synchronized>\n");
3410 			if (disk->d_sync.ds_offset > 0)
3411 				sbuf_printf(sb, "%s<BytesSynced>%jd"
3412 				    "</BytesSynced>\n", indent,
3413 				    (intmax_t)disk->d_sync.ds_offset);
3414 		}
3415 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent,
3416 		    disk->d_sync.ds_syncid);
3417 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent,
3418 		    disk->d_genid);
3419 		sbuf_printf(sb, "%s<Flags>", indent);
3420 		if (disk->d_flags == 0)
3421 			sbuf_cat(sb, "NONE");
3422 		else {
3423 			int first = 1;
3424 
3425 #define	ADD_FLAG(flag, name)	do {					\
3426 	if ((disk->d_flags & (flag)) != 0) {				\
3427 		if (!first)						\
3428 			sbuf_cat(sb, ", ");				\
3429 		else							\
3430 			first = 0;					\
3431 		sbuf_cat(sb, name);					\
3432 	}								\
3433 } while (0)
3434 			ADD_FLAG(G_MIRROR_DISK_FLAG_DIRTY, "DIRTY");
3435 			ADD_FLAG(G_MIRROR_DISK_FLAG_HARDCODED, "HARDCODED");
3436 			ADD_FLAG(G_MIRROR_DISK_FLAG_INACTIVE, "INACTIVE");
3437 			ADD_FLAG(G_MIRROR_DISK_FLAG_SYNCHRONIZING,
3438 			    "SYNCHRONIZING");
3439 			ADD_FLAG(G_MIRROR_DISK_FLAG_FORCE_SYNC, "FORCE_SYNC");
3440 			ADD_FLAG(G_MIRROR_DISK_FLAG_BROKEN, "BROKEN");
3441 #undef	ADD_FLAG
3442 		}
3443 		sbuf_cat(sb, "</Flags>\n");
3444 		sbuf_printf(sb, "%s<Priority>%u</Priority>\n", indent,
3445 		    disk->d_priority);
3446 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3447 		    g_mirror_disk_state2str(disk->d_state));
3448 	} else {
3449 		sbuf_printf(sb, "%s<Type>", indent);
3450 		switch (sc->sc_type) {
3451 		case G_MIRROR_TYPE_AUTOMATIC:
3452 			sbuf_cat(sb, "AUTOMATIC");
3453 			break;
3454 		case G_MIRROR_TYPE_MANUAL:
3455 			sbuf_cat(sb, "MANUAL");
3456 			break;
3457 		default:
3458 			sbuf_cat(sb, "UNKNOWN");
3459 			break;
3460 		}
3461 		sbuf_cat(sb, "</Type>\n");
3462 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
3463 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent, sc->sc_syncid);
3464 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, sc->sc_genid);
3465 		sbuf_printf(sb, "%s<Flags>", indent);
3466 		if (sc->sc_flags == 0)
3467 			sbuf_cat(sb, "NONE");
3468 		else {
3469 			int first = 1;
3470 
3471 #define	ADD_FLAG(flag, name)	do {					\
3472 	if ((sc->sc_flags & (flag)) != 0) {				\
3473 		if (!first)						\
3474 			sbuf_cat(sb, ", ");				\
3475 		else							\
3476 			first = 0;					\
3477 		sbuf_cat(sb, name);					\
3478 	}								\
3479 } while (0)
3480 			ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOFAILSYNC, "NOFAILSYNC");
3481 			ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOAUTOSYNC, "NOAUTOSYNC");
3482 #undef	ADD_FLAG
3483 		}
3484 		sbuf_cat(sb, "</Flags>\n");
3485 		sbuf_printf(sb, "%s<Slice>%u</Slice>\n", indent,
3486 		    (u_int)sc->sc_slice);
3487 		sbuf_printf(sb, "%s<Balance>%s</Balance>\n", indent,
3488 		    balance_name(sc->sc_balance));
3489 		sbuf_printf(sb, "%s<Components>%u</Components>\n", indent,
3490 		    sc->sc_ndisks);
3491 		sbuf_printf(sb, "%s<State>", indent);
3492 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
3493 			sbuf_printf(sb, "%s", "STARTING");
3494 		else if (sc->sc_ndisks ==
3495 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE))
3496 			sbuf_printf(sb, "%s", "COMPLETE");
3497 		else
3498 			sbuf_printf(sb, "%s", "DEGRADED");
3499 		sbuf_cat(sb, "</State>\n");
3500 	}
3501 }
3502 
3503 static void
3504 g_mirror_shutdown_post_sync(void *arg, int howto)
3505 {
3506 	struct g_class *mp;
3507 	struct g_geom *gp, *gp2;
3508 	struct g_mirror_softc *sc;
3509 	int error;
3510 
3511 	if (KERNEL_PANICKED())
3512 		return;
3513 
3514 	mp = arg;
3515 	g_topology_lock();
3516 	g_mirror_shutdown = 1;
3517 	LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
3518 		if ((sc = gp->softc) == NULL)
3519 			continue;
3520 		/* Skip synchronization geom. */
3521 		if (gp == sc->sc_sync.ds_geom)
3522 			continue;
3523 		g_topology_unlock();
3524 		sx_xlock(&sc->sc_lock);
3525 		g_mirror_idle(sc, -1);
3526 		g_cancel_event(sc);
3527 		error = g_mirror_destroy(sc, G_MIRROR_DESTROY_DELAYED);
3528 		if (error != 0)
3529 			sx_xunlock(&sc->sc_lock);
3530 		g_topology_lock();
3531 	}
3532 	g_topology_unlock();
3533 }
3534 
3535 static void
3536 g_mirror_init(struct g_class *mp)
3537 {
3538 
3539 	g_mirror_post_sync = EVENTHANDLER_REGISTER(shutdown_post_sync,
3540 	    g_mirror_shutdown_post_sync, mp, SHUTDOWN_PRI_FIRST);
3541 	if (g_mirror_post_sync == NULL)
3542 		G_MIRROR_DEBUG(0, "Warning! Cannot register shutdown event.");
3543 }
3544 
3545 static void
3546 g_mirror_fini(struct g_class *mp)
3547 {
3548 
3549 	if (g_mirror_post_sync != NULL)
3550 		EVENTHANDLER_DEREGISTER(shutdown_post_sync, g_mirror_post_sync);
3551 }
3552 
3553 /*
3554  * Refresh the mirror device's metadata when gmirror encounters a newer
3555  * generation as the individual components are being added to the mirror set.
3556  */
3557 static int
3558 g_mirror_refresh_device(struct g_mirror_softc *sc, const struct g_provider *pp,
3559     const struct g_mirror_metadata *md)
3560 {
3561 
3562 	g_topology_assert_not();
3563 	sx_assert(&sc->sc_lock, SX_XLOCKED);
3564 
3565 	KASSERT(sc->sc_genid <= md->md_genid,
3566 	    ("%s: attempted to refresh from stale component %s (device %s) "
3567 	    "(%u < %u).", __func__, pp->name, sc->sc_name, md->md_genid,
3568 	    sc->sc_genid));
3569 
3570 	if (sc->sc_genid > md->md_genid || (sc->sc_genid == md->md_genid &&
3571 	    sc->sc_syncid >= md->md_syncid))
3572 		return (0);
3573 
3574 	G_MIRROR_DEBUG(0, "Found newer version for device %s (genid: curr=%u "
3575 	    "new=%u; syncid: curr=%u new=%u; ndisks: curr=%u new=%u; "
3576 	    "provider=%s).", sc->sc_name, sc->sc_genid, md->md_genid,
3577 	    sc->sc_syncid, md->md_syncid, sc->sc_ndisks, md->md_all, pp->name);
3578 
3579 	if (sc->sc_state != G_MIRROR_DEVICE_STATE_STARTING) {
3580 		/* Probable data corruption detected */
3581 		G_MIRROR_DEBUG(0, "Cannot refresh metadata in %s state "
3582 		    "(device=%s genid=%u). A stale mirror device was launched.",
3583 		    g_mirror_device_state2str(sc->sc_state), sc->sc_name,
3584 		    sc->sc_genid);
3585 		return (EINVAL);
3586 	}
3587 
3588 	/* Update softc */
3589 	g_mirror_reinit_from_metadata(sc, md);
3590 
3591 	G_MIRROR_DEBUG(1, "Refresh device %s (id=%u, state=%s) from disk %s "
3592 	    "(genid=%u syncid=%u md_all=%u).", sc->sc_name, md->md_mid,
3593 	    g_mirror_device_state2str(sc->sc_state), pp->name, md->md_genid,
3594 	    md->md_syncid, (unsigned)md->md_all);
3595 
3596 	return (0);
3597 }
3598 
3599 DECLARE_GEOM_CLASS(g_mirror_class, g_mirror);
3600 MODULE_VERSION(geom_mirror, 0);
3601