xref: /freebsd/sys/geom/mirror/g_mirror.c (revision 4543ef516683042d46f3bd3bb8a4f3f746e00499)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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/param.h>
30 #include <sys/systm.h>
31 #include <sys/bio.h>
32 #include <sys/eventhandler.h>
33 #include <sys/fail.h>
34 #include <sys/kernel.h>
35 #include <sys/kthread.h>
36 #include <sys/limits.h>
37 #include <sys/lock.h>
38 #include <sys/malloc.h>
39 #include <sys/mutex.h>
40 #include <sys/proc.h>
41 #include <sys/reboot.h>
42 #include <sys/sbuf.h>
43 #include <sys/sched.h>
44 #include <sys/sx.h>
45 #include <sys/sysctl.h>
46 
47 #include <geom/geom.h>
48 #include <geom/geom_dbg.h>
49 #include <geom/geom_disk.h>
50 #include <geom/mirror/g_mirror.h>
51 
52 FEATURE(geom_mirror, "GEOM mirroring support");
53 
54 static MALLOC_DEFINE(M_MIRROR, "mirror_data", "GEOM_MIRROR Data");
55 
56 SYSCTL_DECL(_kern_geom);
57 static SYSCTL_NODE(_kern_geom, OID_AUTO, mirror, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
58     "GEOM_MIRROR stuff");
59 int g_mirror_debug = 0;
60 SYSCTL_INT(_kern_geom_mirror, OID_AUTO, debug, CTLFLAG_RWTUN, &g_mirror_debug, 0,
61     "Debug level");
62 bool g_launch_mirror_before_timeout = true;
63 SYSCTL_BOOL(_kern_geom_mirror, OID_AUTO, launch_mirror_before_timeout,
64     CTLFLAG_RWTUN, &g_launch_mirror_before_timeout, 0,
65     "If false, force gmirror to wait out the full kern.geom.mirror.timeout "
66     "before launching mirrors");
67 static u_int g_mirror_timeout = 4;
68 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, timeout, CTLFLAG_RWTUN, &g_mirror_timeout,
69     0, "Time to wait on all mirror components");
70 static u_int g_mirror_idletime = 5;
71 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, idletime, CTLFLAG_RWTUN,
72     &g_mirror_idletime, 0, "Mark components as clean when idling");
73 static u_int g_mirror_disconnect_on_failure = 1;
74 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, disconnect_on_failure, CTLFLAG_RWTUN,
75     &g_mirror_disconnect_on_failure, 0, "Disconnect component on I/O failure.");
76 static u_int g_mirror_syncreqs = 2;
77 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, sync_requests, CTLFLAG_RDTUN,
78     &g_mirror_syncreqs, 0, "Parallel synchronization I/O requests.");
79 static u_int g_mirror_sync_period = 5;
80 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, sync_update_period, CTLFLAG_RWTUN,
81     &g_mirror_sync_period, 0,
82     "Metadata update period during synchronization, in seconds");
83 
84 #define	MSLEEP(ident, mtx, priority, wmesg, timeout)	do {		\
85 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, (ident));	\
86 	msleep((ident), (mtx), (priority), (wmesg), (timeout));		\
87 	G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, (ident));	\
88 } while (0)
89 
90 static eventhandler_tag g_mirror_post_sync = NULL;
91 static int g_mirror_shutdown = 0;
92 
93 static g_ctl_destroy_geom_t g_mirror_destroy_geom;
94 static g_taste_t g_mirror_taste;
95 static g_init_t g_mirror_init;
96 static g_fini_t g_mirror_fini;
97 static g_provgone_t g_mirror_providergone;
98 static g_resize_t g_mirror_resize;
99 
100 struct g_class g_mirror_class = {
101 	.name = G_MIRROR_CLASS_NAME,
102 	.version = G_VERSION,
103 	.ctlreq = g_mirror_config,
104 	.taste = g_mirror_taste,
105 	.destroy_geom = g_mirror_destroy_geom,
106 	.init = g_mirror_init,
107 	.fini = g_mirror_fini,
108 	.providergone = g_mirror_providergone,
109 	.resize = g_mirror_resize
110 };
111 
112 static void g_mirror_destroy_provider(struct g_mirror_softc *sc);
113 static int g_mirror_update_disk(struct g_mirror_disk *disk, u_int state);
114 static void g_mirror_update_device(struct g_mirror_softc *sc, bool force);
115 static void g_mirror_dumpconf(struct sbuf *sb, const char *indent,
116     struct g_geom *gp, struct g_consumer *cp, struct g_provider *pp);
117 static void g_mirror_timeout_drain(struct g_mirror_softc *sc);
118 static int g_mirror_refresh_device(struct g_mirror_softc *sc,
119     const struct g_provider *pp, const struct g_mirror_metadata *md);
120 static void g_mirror_sync_reinit(const struct g_mirror_disk *disk,
121     struct bio *bp, off_t offset);
122 static void g_mirror_sync_stop(struct g_mirror_disk *disk, int type);
123 static void g_mirror_register_request(struct g_mirror_softc *sc,
124     struct bio *bp);
125 static void g_mirror_sync_release(struct g_mirror_softc *sc);
126 
127 static const char *
128 g_mirror_disk_state2str(int state)
129 {
130 
131 	switch (state) {
132 	case G_MIRROR_DISK_STATE_NONE:
133 		return ("NONE");
134 	case G_MIRROR_DISK_STATE_NEW:
135 		return ("NEW");
136 	case G_MIRROR_DISK_STATE_ACTIVE:
137 		return ("ACTIVE");
138 	case G_MIRROR_DISK_STATE_STALE:
139 		return ("STALE");
140 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
141 		return ("SYNCHRONIZING");
142 	case G_MIRROR_DISK_STATE_DISCONNECTED:
143 		return ("DISCONNECTED");
144 	case G_MIRROR_DISK_STATE_DESTROY:
145 		return ("DESTROY");
146 	default:
147 		return ("INVALID");
148 	}
149 }
150 
151 static const char *
152 g_mirror_device_state2str(int state)
153 {
154 
155 	switch (state) {
156 	case G_MIRROR_DEVICE_STATE_STARTING:
157 		return ("STARTING");
158 	case G_MIRROR_DEVICE_STATE_RUNNING:
159 		return ("RUNNING");
160 	default:
161 		return ("INVALID");
162 	}
163 }
164 
165 static const char *
166 g_mirror_get_diskname(struct g_mirror_disk *disk)
167 {
168 
169 	if (disk->d_consumer == NULL || disk->d_consumer->provider == NULL)
170 		return ("[unknown]");
171 	return (disk->d_name);
172 }
173 
174 /*
175  * --- Events handling functions ---
176  * Events in geom_mirror are used to maintain disks and device status
177  * from one thread to simplify locking.
178  */
179 static void
180 g_mirror_event_free(struct g_mirror_event *ep)
181 {
182 
183 	free(ep, M_MIRROR);
184 }
185 
186 static int
187 g_mirror_event_dispatch(struct g_mirror_event *ep, void *arg, int state,
188     int flags)
189 {
190 	struct g_mirror_softc *sc;
191 	struct g_mirror_disk *disk;
192 	int error;
193 
194 	G_MIRROR_DEBUG(4, "%s: Sending event %p.", __func__, ep);
195 	if ((flags & G_MIRROR_EVENT_DEVICE) != 0) {
196 		disk = NULL;
197 		sc = arg;
198 	} else {
199 		disk = arg;
200 		sc = disk->d_softc;
201 	}
202 	ep->e_disk = disk;
203 	ep->e_state = state;
204 	ep->e_flags = flags;
205 	ep->e_error = 0;
206 	mtx_lock(&sc->sc_events_mtx);
207 	TAILQ_INSERT_TAIL(&sc->sc_events, ep, e_next);
208 	mtx_unlock(&sc->sc_events_mtx);
209 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
210 	mtx_lock(&sc->sc_queue_mtx);
211 	wakeup(sc);
212 	mtx_unlock(&sc->sc_queue_mtx);
213 	if ((flags & G_MIRROR_EVENT_DONTWAIT) != 0)
214 		return (0);
215 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, ep);
216 	sx_xunlock(&sc->sc_lock);
217 	while ((ep->e_flags & G_MIRROR_EVENT_DONE) == 0) {
218 		mtx_lock(&sc->sc_events_mtx);
219 		MSLEEP(ep, &sc->sc_events_mtx, PRIBIO | PDROP, "m:event",
220 		    hz * 5);
221 	}
222 	error = ep->e_error;
223 	g_mirror_event_free(ep);
224 	sx_xlock(&sc->sc_lock);
225 	return (error);
226 }
227 
228 int
229 g_mirror_event_send(void *arg, int state, int flags)
230 {
231 	struct g_mirror_event *ep;
232 
233 	ep = malloc(sizeof(*ep), M_MIRROR, M_WAITOK);
234 	return (g_mirror_event_dispatch(ep, arg, state, flags));
235 }
236 
237 static struct g_mirror_event *
238 g_mirror_event_first(struct g_mirror_softc *sc)
239 {
240 	struct g_mirror_event *ep;
241 
242 	mtx_lock(&sc->sc_events_mtx);
243 	ep = TAILQ_FIRST(&sc->sc_events);
244 	mtx_unlock(&sc->sc_events_mtx);
245 	return (ep);
246 }
247 
248 static void
249 g_mirror_event_remove(struct g_mirror_softc *sc, struct g_mirror_event *ep)
250 {
251 
252 	mtx_lock(&sc->sc_events_mtx);
253 	TAILQ_REMOVE(&sc->sc_events, ep, e_next);
254 	mtx_unlock(&sc->sc_events_mtx);
255 }
256 
257 static void
258 g_mirror_event_cancel(struct g_mirror_disk *disk)
259 {
260 	struct g_mirror_softc *sc;
261 	struct g_mirror_event *ep, *tmpep;
262 
263 	sc = disk->d_softc;
264 	sx_assert(&sc->sc_lock, SX_XLOCKED);
265 
266 	mtx_lock(&sc->sc_events_mtx);
267 	TAILQ_FOREACH_SAFE(ep, &sc->sc_events, e_next, tmpep) {
268 		if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0)
269 			continue;
270 		if (ep->e_disk != disk)
271 			continue;
272 		TAILQ_REMOVE(&sc->sc_events, ep, e_next);
273 		if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
274 			g_mirror_event_free(ep);
275 		else {
276 			ep->e_error = ECANCELED;
277 			wakeup(ep);
278 		}
279 	}
280 	mtx_unlock(&sc->sc_events_mtx);
281 }
282 
283 /*
284  * Return the number of disks in given state.
285  * If state is equal to -1, count all connected disks.
286  */
287 u_int
288 g_mirror_ndisks(struct g_mirror_softc *sc, int state)
289 {
290 	struct g_mirror_disk *disk;
291 	u_int n = 0;
292 
293 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
294 		if (state == -1 || disk->d_state == state)
295 			n++;
296 	}
297 	return (n);
298 }
299 
300 /*
301  * Find a disk in mirror by its disk ID.
302  */
303 static struct g_mirror_disk *
304 g_mirror_id2disk(struct g_mirror_softc *sc, uint32_t id)
305 {
306 	struct g_mirror_disk *disk;
307 
308 	sx_assert(&sc->sc_lock, SX_XLOCKED);
309 
310 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
311 		if (disk->d_id == id)
312 			return (disk);
313 	}
314 	return (NULL);
315 }
316 
317 static u_int
318 g_mirror_nrequests(struct g_mirror_softc *sc, struct g_consumer *cp)
319 {
320 	struct bio *bp;
321 	u_int nreqs = 0;
322 
323 	mtx_lock(&sc->sc_queue_mtx);
324 	TAILQ_FOREACH(bp, &sc->sc_queue, bio_queue) {
325 		if (bp->bio_from == cp)
326 			nreqs++;
327 	}
328 	mtx_unlock(&sc->sc_queue_mtx);
329 	return (nreqs);
330 }
331 
332 static int
333 g_mirror_is_busy(struct g_mirror_softc *sc, struct g_consumer *cp)
334 {
335 
336 	if (cp->index > 0) {
337 		G_MIRROR_DEBUG(2,
338 		    "I/O requests for %s exist, can't destroy it now.",
339 		    cp->provider->name);
340 		return (1);
341 	}
342 	if (g_mirror_nrequests(sc, cp) > 0) {
343 		G_MIRROR_DEBUG(2,
344 		    "I/O requests for %s in queue, can't destroy it now.",
345 		    cp->provider->name);
346 		return (1);
347 	}
348 	return (0);
349 }
350 
351 static void
352 g_mirror_destroy_consumer(void *arg, int flags __unused)
353 {
354 	struct g_consumer *cp;
355 
356 	g_topology_assert();
357 
358 	cp = arg;
359 	G_MIRROR_DEBUG(1, "Consumer %s destroyed.", cp->provider->name);
360 	g_detach(cp);
361 	g_destroy_consumer(cp);
362 }
363 
364 static void
365 g_mirror_kill_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
366 {
367 	struct g_provider *pp;
368 	int retaste_wait;
369 
370 	g_topology_assert();
371 
372 	cp->private = NULL;
373 	if (g_mirror_is_busy(sc, cp))
374 		return;
375 	pp = cp->provider;
376 	retaste_wait = 0;
377 	if (cp->acw == 1) {
378 		if ((pp->geom->flags & G_GEOM_WITHER) == 0)
379 			retaste_wait = 1;
380 	}
381 	G_MIRROR_DEBUG(2, "Access %s r%dw%de%d = %d", pp->name, -cp->acr,
382 	    -cp->acw, -cp->ace, 0);
383 	if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
384 		g_access(cp, -cp->acr, -cp->acw, -cp->ace);
385 	if (retaste_wait) {
386 		/*
387 		 * After retaste event was send (inside g_access()), we can send
388 		 * event to detach and destroy consumer.
389 		 * A class, which has consumer to the given provider connected
390 		 * will not receive retaste event for the provider.
391 		 * This is the way how I ignore retaste events when I close
392 		 * consumers opened for write: I detach and destroy consumer
393 		 * after retaste event is sent.
394 		 */
395 		g_post_event(g_mirror_destroy_consumer, cp, M_WAITOK, NULL);
396 		return;
397 	}
398 	G_MIRROR_DEBUG(1, "Consumer %s destroyed.", pp->name);
399 	g_detach(cp);
400 	g_destroy_consumer(cp);
401 }
402 
403 static int
404 g_mirror_connect_disk(struct g_mirror_disk *disk, struct g_provider *pp)
405 {
406 	struct g_consumer *cp;
407 	int error;
408 
409 	g_topology_assert_not();
410 	KASSERT(disk->d_consumer == NULL,
411 	    ("Disk already connected (device %s).", disk->d_softc->sc_name));
412 
413 	g_topology_lock();
414 	cp = g_new_consumer(disk->d_softc->sc_geom);
415 	cp->flags |= G_CF_DIRECT_RECEIVE;
416 	error = g_attach(cp, pp);
417 	if (error != 0) {
418 		g_destroy_consumer(cp);
419 		g_topology_unlock();
420 		return (error);
421 	}
422 	error = g_access(cp, 1, 1, 1);
423 	if (error != 0) {
424 		g_detach(cp);
425 		g_destroy_consumer(cp);
426 		g_topology_unlock();
427 		G_MIRROR_DEBUG(0, "Cannot open consumer %s (error=%d).",
428 		    pp->name, error);
429 		return (error);
430 	}
431 	g_topology_unlock();
432 	disk->d_consumer = cp;
433 	disk->d_consumer->private = disk;
434 	disk->d_consumer->index = 0;
435 
436 	G_MIRROR_DEBUG(2, "Disk %s connected.", g_mirror_get_diskname(disk));
437 	return (0);
438 }
439 
440 static void
441 g_mirror_disconnect_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
442 {
443 
444 	g_topology_assert();
445 
446 	if (cp == NULL)
447 		return;
448 	if (cp->provider != NULL)
449 		g_mirror_kill_consumer(sc, cp);
450 	else
451 		g_destroy_consumer(cp);
452 }
453 
454 /*
455  * Initialize disk. This means allocate memory, create consumer, attach it
456  * to the provider and open access (r1w1e1) to it.
457  */
458 static struct g_mirror_disk *
459 g_mirror_init_disk(struct g_mirror_softc *sc, struct g_provider *pp,
460     struct g_mirror_metadata *md, int *errorp)
461 {
462 	struct g_mirror_disk *disk;
463 	int i, error;
464 
465 	disk = malloc(sizeof(*disk), M_MIRROR, M_NOWAIT | M_ZERO);
466 	if (disk == NULL) {
467 		error = ENOMEM;
468 		goto fail;
469 	}
470 	disk->d_softc = sc;
471 	error = g_mirror_connect_disk(disk, pp);
472 	if (error != 0)
473 		goto fail;
474 	disk->d_id = md->md_did;
475 	disk->d_state = G_MIRROR_DISK_STATE_NONE;
476 	disk->d_priority = md->md_priority;
477 	disk->d_flags = md->md_dflags;
478 	error = g_getattr("GEOM::candelete", disk->d_consumer, &i);
479 	if (error == 0 && i != 0)
480 		disk->d_flags |= G_MIRROR_DISK_FLAG_CANDELETE;
481 	error = g_getattr("GEOM::rotation_rate", disk->d_consumer,
482 		&disk->d_rotation_rate);
483 	if (error)
484 		disk->d_rotation_rate = DISK_RR_UNKNOWN;
485 	if (md->md_provider[0] != '\0')
486 		disk->d_flags |= G_MIRROR_DISK_FLAG_HARDCODED;
487 	disk->d_sync.ds_consumer = NULL;
488 	disk->d_sync.ds_offset = md->md_sync_offset;
489 	disk->d_sync.ds_offset_done = md->md_sync_offset;
490 	disk->d_sync.ds_update_ts = time_uptime;
491 	disk->d_genid = md->md_genid;
492 	disk->d_sync.ds_syncid = md->md_syncid;
493 	disk->d_init_ndisks = md->md_all;
494 	disk->d_init_slice = md->md_slice;
495 	disk->d_init_balance = md->md_balance;
496 	disk->d_init_mediasize = md->md_mediasize;
497 	if (errorp != NULL)
498 		*errorp = 0;
499 	return (disk);
500 fail:
501 	if (errorp != NULL)
502 		*errorp = error;
503 	if (disk != NULL)
504 		free(disk, M_MIRROR);
505 	return (NULL);
506 }
507 
508 static void
509 g_mirror_destroy_disk(struct g_mirror_disk *disk)
510 {
511 	struct g_mirror_softc *sc;
512 
513 	g_topology_assert_not();
514 	sc = disk->d_softc;
515 	sx_assert(&sc->sc_lock, SX_XLOCKED);
516 
517 	g_topology_lock();
518 	LIST_REMOVE(disk, d_next);
519 	g_topology_unlock();
520 	g_mirror_event_cancel(disk);
521 	if (sc->sc_hint == disk)
522 		sc->sc_hint = NULL;
523 	switch (disk->d_state) {
524 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
525 		g_mirror_sync_stop(disk, 1);
526 		/* FALLTHROUGH */
527 	case G_MIRROR_DISK_STATE_NEW:
528 	case G_MIRROR_DISK_STATE_STALE:
529 	case G_MIRROR_DISK_STATE_ACTIVE:
530 		g_topology_lock();
531 		g_mirror_disconnect_consumer(sc, disk->d_consumer);
532 		g_topology_unlock();
533 		free(disk, M_MIRROR);
534 		break;
535 	default:
536 		KASSERT(0 == 1, ("Wrong disk state (%s, %s).",
537 		    g_mirror_get_diskname(disk),
538 		    g_mirror_disk_state2str(disk->d_state)));
539 	}
540 }
541 
542 static void
543 g_mirror_free_device(struct g_mirror_softc *sc)
544 {
545 
546 	g_topology_assert();
547 
548 	mtx_destroy(&sc->sc_queue_mtx);
549 	mtx_destroy(&sc->sc_events_mtx);
550 	mtx_destroy(&sc->sc_done_mtx);
551 	sx_destroy(&sc->sc_lock);
552 	free(sc, M_MIRROR);
553 }
554 
555 static void
556 g_mirror_providergone(struct g_provider *pp)
557 {
558 	struct g_mirror_softc *sc = pp->private;
559 
560 	if ((--sc->sc_refcnt) == 0)
561 		g_mirror_free_device(sc);
562 }
563 
564 static void
565 g_mirror_destroy_device(struct g_mirror_softc *sc)
566 {
567 	struct g_mirror_disk *disk;
568 	struct g_mirror_event *ep;
569 	struct g_geom *gp;
570 	struct g_consumer *cp, *tmpcp;
571 
572 	g_topology_assert_not();
573 	sx_assert(&sc->sc_lock, SX_XLOCKED);
574 
575 	gp = sc->sc_geom;
576 	if (sc->sc_provider != NULL)
577 		g_mirror_destroy_provider(sc);
578 	for (disk = LIST_FIRST(&sc->sc_disks); disk != NULL;
579 	    disk = LIST_FIRST(&sc->sc_disks)) {
580 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
581 		g_mirror_update_metadata(disk);
582 		g_mirror_destroy_disk(disk);
583 	}
584 	while ((ep = g_mirror_event_first(sc)) != NULL) {
585 		g_mirror_event_remove(sc, ep);
586 		if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
587 			g_mirror_event_free(ep);
588 		else {
589 			ep->e_error = ECANCELED;
590 			ep->e_flags |= G_MIRROR_EVENT_DONE;
591 			G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, ep);
592 			mtx_lock(&sc->sc_events_mtx);
593 			wakeup(ep);
594 			mtx_unlock(&sc->sc_events_mtx);
595 		}
596 	}
597 	g_mirror_timeout_drain(sc);
598 
599 	g_topology_lock();
600 	LIST_FOREACH_SAFE(cp, &sc->sc_sync.ds_geom->consumer, consumer, tmpcp) {
601 		g_mirror_disconnect_consumer(sc, cp);
602 	}
603 	g_wither_geom(sc->sc_sync.ds_geom, ENXIO);
604 	G_MIRROR_DEBUG(0, "Device %s destroyed.", gp->name);
605 	g_wither_geom(gp, ENXIO);
606 	sx_xunlock(&sc->sc_lock);
607 	if ((--sc->sc_refcnt) == 0)
608 		g_mirror_free_device(sc);
609 	g_topology_unlock();
610 }
611 
612 static void
613 g_mirror_orphan(struct g_consumer *cp)
614 {
615 	struct g_mirror_disk *disk;
616 
617 	g_topology_assert();
618 
619 	disk = cp->private;
620 	if (disk == NULL)
621 		return;
622 	disk->d_softc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
623 	g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
624 	    G_MIRROR_EVENT_DONTWAIT);
625 }
626 
627 /*
628  * Function should return the next active disk on the list.
629  * It is possible that it will be the same disk as given.
630  * If there are no active disks on list, NULL is returned.
631  */
632 static __inline struct g_mirror_disk *
633 g_mirror_find_next(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
634 {
635 	struct g_mirror_disk *dp;
636 
637 	for (dp = LIST_NEXT(disk, d_next); dp != disk;
638 	    dp = LIST_NEXT(dp, d_next)) {
639 		if (dp == NULL)
640 			dp = LIST_FIRST(&sc->sc_disks);
641 		if (dp->d_state == G_MIRROR_DISK_STATE_ACTIVE)
642 			break;
643 	}
644 	if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
645 		return (NULL);
646 	return (dp);
647 }
648 
649 static struct g_mirror_disk *
650 g_mirror_get_disk(struct g_mirror_softc *sc)
651 {
652 	struct g_mirror_disk *disk;
653 
654 	if (sc->sc_hint == NULL) {
655 		sc->sc_hint = LIST_FIRST(&sc->sc_disks);
656 		if (sc->sc_hint == NULL)
657 			return (NULL);
658 	}
659 	disk = sc->sc_hint;
660 	if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE) {
661 		disk = g_mirror_find_next(sc, disk);
662 		if (disk == NULL)
663 			return (NULL);
664 	}
665 	sc->sc_hint = g_mirror_find_next(sc, disk);
666 	return (disk);
667 }
668 
669 static int
670 g_mirror_write_metadata(struct g_mirror_disk *disk,
671     struct g_mirror_metadata *md)
672 {
673 	struct g_mirror_softc *sc;
674 	struct g_consumer *cp;
675 	off_t offset, length;
676 	u_char *sector;
677 	int error = 0;
678 
679 	g_topology_assert_not();
680 	sc = disk->d_softc;
681 	sx_assert(&sc->sc_lock, SX_LOCKED);
682 
683 	cp = disk->d_consumer;
684 	KASSERT(cp != NULL, ("NULL consumer (%s).", sc->sc_name));
685 	KASSERT(cp->provider != NULL, ("NULL provider (%s).", sc->sc_name));
686 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
687 	    ("Consumer %s closed? (r%dw%de%d).", cp->provider->name, cp->acr,
688 	    cp->acw, cp->ace));
689 	length = cp->provider->sectorsize;
690 	offset = cp->provider->mediasize - length;
691 	sector = malloc((size_t)length, M_MIRROR, M_WAITOK | M_ZERO);
692 	if (md != NULL &&
693 	    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_WIPE) == 0) {
694 		/*
695 		 * Handle the case, when the size of parent provider reduced.
696 		 */
697 		if (offset < md->md_mediasize)
698 			error = ENOSPC;
699 		else
700 			mirror_metadata_encode(md, sector);
701 	}
702 	KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_metadata_write, error);
703 	if (error == 0)
704 		error = g_write_data(cp, offset, sector, length);
705 	free(sector, M_MIRROR);
706 	if (error != 0) {
707 		if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
708 			disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
709 			G_MIRROR_DEBUG(0, "Cannot write metadata on %s "
710 			    "(device=%s, error=%d).",
711 			    g_mirror_get_diskname(disk), sc->sc_name, error);
712 		} else {
713 			G_MIRROR_DEBUG(1, "Cannot write metadata on %s "
714 			    "(device=%s, error=%d).",
715 			    g_mirror_get_diskname(disk), sc->sc_name, error);
716 		}
717 		if (g_mirror_disconnect_on_failure &&
718 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1) {
719 			sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
720 			g_mirror_event_send(disk,
721 			    G_MIRROR_DISK_STATE_DISCONNECTED,
722 			    G_MIRROR_EVENT_DONTWAIT);
723 		}
724 	}
725 	return (error);
726 }
727 
728 static int
729 g_mirror_clear_metadata(struct g_mirror_disk *disk)
730 {
731 	int error;
732 
733 	g_topology_assert_not();
734 	sx_assert(&disk->d_softc->sc_lock, SX_LOCKED);
735 
736 	if (disk->d_softc->sc_type != G_MIRROR_TYPE_AUTOMATIC)
737 		return (0);
738 	error = g_mirror_write_metadata(disk, NULL);
739 	if (error == 0) {
740 		G_MIRROR_DEBUG(2, "Metadata on %s cleared.",
741 		    g_mirror_get_diskname(disk));
742 	} else {
743 		G_MIRROR_DEBUG(0,
744 		    "Cannot clear metadata on disk %s (error=%d).",
745 		    g_mirror_get_diskname(disk), error);
746 	}
747 	return (error);
748 }
749 
750 void
751 g_mirror_fill_metadata(struct g_mirror_softc *sc, struct g_mirror_disk *disk,
752     struct g_mirror_metadata *md)
753 {
754 
755 	bzero(md, sizeof(*md));
756 	strlcpy(md->md_magic, G_MIRROR_MAGIC, sizeof(md->md_magic));
757 	md->md_version = G_MIRROR_VERSION;
758 	strlcpy(md->md_name, sc->sc_name, sizeof(md->md_name));
759 	md->md_mid = sc->sc_id;
760 	md->md_all = sc->sc_ndisks;
761 	md->md_slice = sc->sc_slice;
762 	md->md_balance = sc->sc_balance;
763 	md->md_genid = sc->sc_genid;
764 	md->md_mediasize = sc->sc_mediasize;
765 	md->md_sectorsize = sc->sc_sectorsize;
766 	md->md_mflags = (sc->sc_flags & G_MIRROR_DEVICE_FLAG_MASK);
767 	if (disk == NULL) {
768 		md->md_did = arc4random();
769 	} else {
770 		md->md_did = disk->d_id;
771 		md->md_priority = disk->d_priority;
772 		md->md_syncid = disk->d_sync.ds_syncid;
773 		md->md_dflags = (disk->d_flags & G_MIRROR_DISK_FLAG_MASK);
774 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
775 			md->md_sync_offset = disk->d_sync.ds_offset_done;
776 		if ((disk->d_flags & G_MIRROR_DISK_FLAG_HARDCODED) != 0) {
777 			strlcpy(md->md_provider,
778 			    disk->d_consumer->provider->name,
779 			    sizeof(md->md_provider));
780 		}
781 		md->md_provsize = disk->d_consumer->provider->mediasize;
782 	}
783 }
784 
785 void
786 g_mirror_update_metadata(struct g_mirror_disk *disk)
787 {
788 	struct g_mirror_softc *sc;
789 	struct g_mirror_metadata md;
790 	int error;
791 
792 	g_topology_assert_not();
793 	sc = disk->d_softc;
794 	sx_assert(&sc->sc_lock, SX_LOCKED);
795 
796 	if (sc->sc_type != G_MIRROR_TYPE_AUTOMATIC)
797 		return;
798 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_WIPE) == 0)
799 		g_mirror_fill_metadata(sc, disk, &md);
800 	error = g_mirror_write_metadata(disk, &md);
801 	if (error == 0) {
802 		G_MIRROR_DEBUG(2, "Metadata on %s updated.",
803 		    g_mirror_get_diskname(disk));
804 	} else {
805 		G_MIRROR_DEBUG(0,
806 		    "Cannot update metadata on disk %s (error=%d).",
807 		    g_mirror_get_diskname(disk), error);
808 	}
809 }
810 
811 static void
812 g_mirror_bump_syncid(struct g_mirror_softc *sc)
813 {
814 	struct g_mirror_disk *disk;
815 
816 	g_topology_assert_not();
817 	sx_assert(&sc->sc_lock, SX_XLOCKED);
818 	KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
819 	    ("%s called with no active disks (device=%s).", __func__,
820 	    sc->sc_name));
821 
822 	sc->sc_syncid++;
823 	G_MIRROR_DEBUG(1, "Device %s: syncid bumped to %u.", sc->sc_name,
824 	    sc->sc_syncid);
825 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
826 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
827 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
828 			disk->d_sync.ds_syncid = sc->sc_syncid;
829 			g_mirror_update_metadata(disk);
830 		}
831 	}
832 }
833 
834 static void
835 g_mirror_bump_genid(struct g_mirror_softc *sc)
836 {
837 	struct g_mirror_disk *disk;
838 
839 	g_topology_assert_not();
840 	sx_assert(&sc->sc_lock, SX_XLOCKED);
841 	KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
842 	    ("%s called with no active disks (device=%s).", __func__,
843 	    sc->sc_name));
844 
845 	sc->sc_genid++;
846 	G_MIRROR_DEBUG(1, "Device %s: genid bumped to %u.", sc->sc_name,
847 	    sc->sc_genid);
848 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
849 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
850 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
851 			disk->d_genid = sc->sc_genid;
852 			g_mirror_update_metadata(disk);
853 		}
854 	}
855 }
856 
857 static int
858 g_mirror_idle(struct g_mirror_softc *sc, int acw)
859 {
860 	struct g_mirror_disk *disk;
861 	int timeout;
862 
863 	g_topology_assert_not();
864 	sx_assert(&sc->sc_lock, SX_XLOCKED);
865 
866 	if (sc->sc_provider == NULL)
867 		return (0);
868 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
869 		return (0);
870 	if (sc->sc_idle)
871 		return (0);
872 	if (sc->sc_writes > 0)
873 		return (0);
874 	if (acw > 0 || (acw == -1 && sc->sc_provider->acw > 0)) {
875 		timeout = g_mirror_idletime - (time_uptime - sc->sc_last_write);
876 		if (!g_mirror_shutdown && timeout > 0)
877 			return (timeout);
878 	}
879 	sc->sc_idle = 1;
880 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
881 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
882 			continue;
883 		G_MIRROR_DEBUG(2, "Disk %s (device %s) marked as clean.",
884 		    g_mirror_get_diskname(disk), sc->sc_name);
885 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
886 		g_mirror_update_metadata(disk);
887 	}
888 	return (0);
889 }
890 
891 static void
892 g_mirror_unidle(struct g_mirror_softc *sc)
893 {
894 	struct g_mirror_disk *disk;
895 
896 	g_topology_assert_not();
897 	sx_assert(&sc->sc_lock, SX_XLOCKED);
898 
899 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
900 		return;
901 	sc->sc_idle = 0;
902 	sc->sc_last_write = time_uptime;
903 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
904 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
905 			continue;
906 		G_MIRROR_DEBUG(2, "Disk %s (device %s) marked as dirty.",
907 		    g_mirror_get_diskname(disk), sc->sc_name);
908 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
909 		g_mirror_update_metadata(disk);
910 	}
911 }
912 
913 static void
914 g_mirror_done(struct bio *bp)
915 {
916 	struct g_mirror_softc *sc;
917 
918 	sc = bp->bio_from->geom->softc;
919 	bp->bio_cflags = G_MIRROR_BIO_FLAG_REGULAR;
920 	mtx_lock(&sc->sc_queue_mtx);
921 	TAILQ_INSERT_TAIL(&sc->sc_queue, bp, bio_queue);
922 	mtx_unlock(&sc->sc_queue_mtx);
923 	wakeup(sc);
924 }
925 
926 static void
927 g_mirror_regular_request_error(struct g_mirror_softc *sc,
928     struct g_mirror_disk *disk, struct bio *bp)
929 {
930 
931 	if ((bp->bio_cmd == BIO_FLUSH || bp->bio_cmd == BIO_SPEEDUP) &&
932 	    bp->bio_error == EOPNOTSUPP)
933 		return;
934 
935 	if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
936 		disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
937 		G_MIRROR_LOGREQ(0, bp, "Request failed (error=%d).",
938 		    bp->bio_error);
939 	} else {
940 		G_MIRROR_LOGREQ(1, bp, "Request failed (error=%d).",
941 		    bp->bio_error);
942 	}
943 	if (g_mirror_disconnect_on_failure &&
944 	    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1) {
945 		if (bp->bio_error == ENXIO &&
946 		    bp->bio_cmd == BIO_READ)
947 			sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
948 		else if (bp->bio_error == ENXIO)
949 			sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID_NOW;
950 		else
951 			sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
952 		g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
953 		    G_MIRROR_EVENT_DONTWAIT);
954 	}
955 }
956 
957 static void
958 g_mirror_regular_request(struct g_mirror_softc *sc, struct bio *bp)
959 {
960 	struct g_mirror_disk *disk;
961 	struct bio *pbp;
962 
963 	g_topology_assert_not();
964 	KASSERT(sc->sc_provider == bp->bio_parent->bio_to,
965 	    ("regular request %p with unexpected origin", bp));
966 
967 	pbp = bp->bio_parent;
968 	bp->bio_from->index--;
969 	if (bp->bio_cmd == BIO_WRITE || bp->bio_cmd == BIO_DELETE)
970 		sc->sc_writes--;
971 	disk = bp->bio_from->private;
972 	if (disk == NULL) {
973 		g_topology_lock();
974 		g_mirror_kill_consumer(sc, bp->bio_from);
975 		g_topology_unlock();
976 	}
977 
978 	switch (bp->bio_cmd) {
979 	case BIO_READ:
980 		KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_read,
981 		    bp->bio_error);
982 		break;
983 	case BIO_WRITE:
984 		KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_write,
985 		    bp->bio_error);
986 		break;
987 	case BIO_DELETE:
988 		KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_delete,
989 		    bp->bio_error);
990 		break;
991 	case BIO_FLUSH:
992 		KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_flush,
993 		    bp->bio_error);
994 		break;
995 	case BIO_SPEEDUP:
996 		KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_speedup,
997 		    bp->bio_error);
998 		break;
999 	}
1000 
1001 	pbp->bio_inbed++;
1002 	KASSERT(pbp->bio_inbed <= pbp->bio_children,
1003 	    ("bio_inbed (%u) is bigger than bio_children (%u).", pbp->bio_inbed,
1004 	    pbp->bio_children));
1005 	if (bp->bio_error == 0 && pbp->bio_error == 0) {
1006 		G_MIRROR_LOGREQ(3, bp, "Request delivered.");
1007 		g_destroy_bio(bp);
1008 		if (pbp->bio_children == pbp->bio_inbed) {
1009 			G_MIRROR_LOGREQ(3, pbp, "Request delivered.");
1010 			pbp->bio_completed = pbp->bio_length;
1011 			if (pbp->bio_cmd == BIO_WRITE ||
1012 			    pbp->bio_cmd == BIO_DELETE) {
1013 				TAILQ_REMOVE(&sc->sc_inflight, pbp, bio_queue);
1014 				/* Release delayed sync requests if possible. */
1015 				g_mirror_sync_release(sc);
1016 			}
1017 			g_io_deliver(pbp, pbp->bio_error);
1018 		}
1019 		return;
1020 	} else if (bp->bio_error != 0) {
1021 		if (pbp->bio_error == 0)
1022 			pbp->bio_error = bp->bio_error;
1023 		if (disk != NULL)
1024 			g_mirror_regular_request_error(sc, disk, bp);
1025 		switch (pbp->bio_cmd) {
1026 		case BIO_DELETE:
1027 		case BIO_WRITE:
1028 		case BIO_FLUSH:
1029 		case BIO_SPEEDUP:
1030 			pbp->bio_inbed--;
1031 			pbp->bio_children--;
1032 			break;
1033 		}
1034 	}
1035 	g_destroy_bio(bp);
1036 
1037 	switch (pbp->bio_cmd) {
1038 	case BIO_READ:
1039 		if (pbp->bio_inbed < pbp->bio_children)
1040 			break;
1041 
1042 		/*
1043 		 * If there is only one active disk we want to double-check that
1044 		 * it is, in fact, the disk that we already tried.  This is
1045 		 * necessary because we might have just lost a race with a
1046 		 * removal of the tried disk (likely because of the same error)
1047 		 * and the only remaining disk is still viable for a retry.
1048 		 */
1049 		if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 1 &&
1050 		    disk != NULL &&
1051 		    disk->d_state == G_MIRROR_DISK_STATE_ACTIVE) {
1052 			g_io_deliver(pbp, pbp->bio_error);
1053 		} else {
1054 			pbp->bio_error = 0;
1055 			mtx_lock(&sc->sc_queue_mtx);
1056 			TAILQ_INSERT_TAIL(&sc->sc_queue, pbp, bio_queue);
1057 			mtx_unlock(&sc->sc_queue_mtx);
1058 			G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1059 			wakeup(sc);
1060 		}
1061 		break;
1062 	case BIO_DELETE:
1063 	case BIO_WRITE:
1064 	case BIO_FLUSH:
1065 	case BIO_SPEEDUP:
1066 		if (pbp->bio_children == 0) {
1067 			/*
1068 			 * All requests failed.
1069 			 */
1070 		} else if (pbp->bio_inbed < pbp->bio_children) {
1071 			/* Do nothing. */
1072 			break;
1073 		} else if (pbp->bio_children == pbp->bio_inbed) {
1074 			/* Some requests succeeded. */
1075 			pbp->bio_error = 0;
1076 			pbp->bio_completed = pbp->bio_length;
1077 		}
1078 		if (pbp->bio_cmd == BIO_WRITE || pbp->bio_cmd == BIO_DELETE) {
1079 			TAILQ_REMOVE(&sc->sc_inflight, pbp, bio_queue);
1080 			/* Release delayed sync requests if possible. */
1081 			g_mirror_sync_release(sc);
1082 		}
1083 		g_io_deliver(pbp, pbp->bio_error);
1084 		break;
1085 	default:
1086 		KASSERT(1 == 0, ("Invalid request: %u.", pbp->bio_cmd));
1087 		break;
1088 	}
1089 }
1090 
1091 static void
1092 g_mirror_sync_done(struct bio *bp)
1093 {
1094 	struct g_mirror_softc *sc;
1095 
1096 	G_MIRROR_LOGREQ(3, bp, "Synchronization request delivered.");
1097 	sc = bp->bio_from->geom->softc;
1098 	bp->bio_cflags = G_MIRROR_BIO_FLAG_SYNC;
1099 	mtx_lock(&sc->sc_queue_mtx);
1100 	TAILQ_INSERT_TAIL(&sc->sc_queue, bp, bio_queue);
1101 	mtx_unlock(&sc->sc_queue_mtx);
1102 	wakeup(sc);
1103 }
1104 
1105 static void
1106 g_mirror_candelete(struct bio *bp)
1107 {
1108 	struct g_mirror_softc *sc;
1109 	struct g_mirror_disk *disk;
1110 	int val;
1111 
1112 	sc = bp->bio_to->private;
1113 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1114 		if (disk->d_flags & G_MIRROR_DISK_FLAG_CANDELETE)
1115 			break;
1116 	}
1117 	val = disk != NULL;
1118 	g_handleattr(bp, "GEOM::candelete", &val, sizeof(val));
1119 }
1120 
1121 static void
1122 g_mirror_kernel_dump(struct bio *bp)
1123 {
1124 	struct g_mirror_softc *sc;
1125 	struct g_mirror_disk *disk;
1126 	struct bio *cbp;
1127 	struct g_kerneldump *gkd;
1128 
1129 	/*
1130 	 * We configure dumping to the first component, because this component
1131 	 * will be used for reading with 'prefer' balance algorithm.
1132 	 * If the component with the highest priority is currently disconnected
1133 	 * we will not be able to read the dump after the reboot if it will be
1134 	 * connected and synchronized later. Can we do something better?
1135 	 */
1136 	sc = bp->bio_to->private;
1137 	disk = LIST_FIRST(&sc->sc_disks);
1138 
1139 	gkd = (struct g_kerneldump *)bp->bio_data;
1140 	if (gkd->length > bp->bio_to->mediasize)
1141 		gkd->length = bp->bio_to->mediasize;
1142 	cbp = g_clone_bio(bp);
1143 	if (cbp == NULL) {
1144 		g_io_deliver(bp, ENOMEM);
1145 		return;
1146 	}
1147 	cbp->bio_done = g_std_done;
1148 	g_io_request(cbp, disk->d_consumer);
1149 	G_MIRROR_DEBUG(1, "Kernel dump will go to %s.",
1150 	    g_mirror_get_diskname(disk));
1151 }
1152 
1153 static void
1154 g_mirror_rotation_rate(struct bio *bp)
1155 {
1156 	struct g_mirror_softc *sc;
1157 	struct g_mirror_disk *disk;
1158 	bool first = true;
1159 	uint16_t rr = DISK_RR_UNKNOWN;
1160 
1161 	sc = bp->bio_to->private;
1162 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1163 		if (first)
1164 			rr = disk->d_rotation_rate;
1165 		else if (rr != disk->d_rotation_rate) {
1166 			rr = DISK_RR_UNKNOWN;
1167 			break;
1168 		}
1169 		first = false;
1170 	}
1171 	g_handleattr(bp, "GEOM::rotation_rate", &rr, sizeof(rr));
1172 }
1173 
1174 static void
1175 g_mirror_start(struct bio *bp)
1176 {
1177 	struct g_mirror_softc *sc;
1178 
1179 	sc = bp->bio_to->private;
1180 	/*
1181 	 * If sc == NULL or there are no valid disks, provider's error
1182 	 * should be set and g_mirror_start() should not be called at all.
1183 	 */
1184 	KASSERT(sc != NULL && sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1185 	    ("Provider's error should be set (error=%d)(mirror=%s).",
1186 	    bp->bio_to->error, bp->bio_to->name));
1187 	G_MIRROR_LOGREQ(3, bp, "Request received.");
1188 
1189 	switch (bp->bio_cmd) {
1190 	case BIO_READ:
1191 	case BIO_WRITE:
1192 	case BIO_DELETE:
1193 	case BIO_SPEEDUP:
1194 	case BIO_FLUSH:
1195 		break;
1196 	case BIO_GETATTR:
1197 		if (!strcmp(bp->bio_attribute, "GEOM::candelete")) {
1198 			g_mirror_candelete(bp);
1199 			return;
1200 		} else if (strcmp("GEOM::kerneldump", bp->bio_attribute) == 0) {
1201 			g_mirror_kernel_dump(bp);
1202 			return;
1203 		} else if (!strcmp(bp->bio_attribute, "GEOM::rotation_rate")) {
1204 			g_mirror_rotation_rate(bp);
1205 			return;
1206 		}
1207 		/* FALLTHROUGH */
1208 	default:
1209 		g_io_deliver(bp, EOPNOTSUPP);
1210 		return;
1211 	}
1212 	mtx_lock(&sc->sc_queue_mtx);
1213 	if (bp->bio_to->error != 0) {
1214 		mtx_unlock(&sc->sc_queue_mtx);
1215 		g_io_deliver(bp, bp->bio_to->error);
1216 		return;
1217 	}
1218 	TAILQ_INSERT_TAIL(&sc->sc_queue, bp, bio_queue);
1219 	mtx_unlock(&sc->sc_queue_mtx);
1220 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1221 	wakeup(sc);
1222 }
1223 
1224 /*
1225  * Return TRUE if the given request is colliding with a in-progress
1226  * synchronization request.
1227  */
1228 static bool
1229 g_mirror_sync_collision(struct g_mirror_softc *sc, struct bio *bp)
1230 {
1231 	struct g_mirror_disk *disk;
1232 	struct bio *sbp;
1233 	off_t rstart, rend, sstart, send;
1234 	u_int i;
1235 
1236 	if (sc->sc_sync.ds_ndisks == 0)
1237 		return (false);
1238 	rstart = bp->bio_offset;
1239 	rend = bp->bio_offset + bp->bio_length;
1240 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1241 		if (disk->d_state != G_MIRROR_DISK_STATE_SYNCHRONIZING)
1242 			continue;
1243 		for (i = 0; i < g_mirror_syncreqs; i++) {
1244 			sbp = disk->d_sync.ds_bios[i];
1245 			if (sbp == NULL)
1246 				continue;
1247 			sstart = sbp->bio_offset;
1248 			send = sbp->bio_offset + sbp->bio_length;
1249 			if (rend > sstart && rstart < send)
1250 				return (true);
1251 		}
1252 	}
1253 	return (false);
1254 }
1255 
1256 /*
1257  * Return TRUE if the given sync request is colliding with a in-progress regular
1258  * request.
1259  */
1260 static bool
1261 g_mirror_regular_collision(struct g_mirror_softc *sc, struct bio *sbp)
1262 {
1263 	off_t rstart, rend, sstart, send;
1264 	struct bio *bp;
1265 
1266 	if (sc->sc_sync.ds_ndisks == 0)
1267 		return (false);
1268 	sstart = sbp->bio_offset;
1269 	send = sbp->bio_offset + sbp->bio_length;
1270 	TAILQ_FOREACH(bp, &sc->sc_inflight, bio_queue) {
1271 		rstart = bp->bio_offset;
1272 		rend = bp->bio_offset + bp->bio_length;
1273 		if (rend > sstart && rstart < send)
1274 			return (true);
1275 	}
1276 	return (false);
1277 }
1278 
1279 /*
1280  * Puts regular request onto delayed queue.
1281  */
1282 static void
1283 g_mirror_regular_delay(struct g_mirror_softc *sc, struct bio *bp)
1284 {
1285 
1286 	G_MIRROR_LOGREQ(2, bp, "Delaying request.");
1287 	TAILQ_INSERT_TAIL(&sc->sc_regular_delayed, bp, bio_queue);
1288 }
1289 
1290 /*
1291  * Puts synchronization request onto delayed queue.
1292  */
1293 static void
1294 g_mirror_sync_delay(struct g_mirror_softc *sc, struct bio *bp)
1295 {
1296 
1297 	G_MIRROR_LOGREQ(2, bp, "Delaying synchronization request.");
1298 	TAILQ_INSERT_TAIL(&sc->sc_sync_delayed, bp, bio_queue);
1299 }
1300 
1301 /*
1302  * Requeue delayed regular requests.
1303  */
1304 static void
1305 g_mirror_regular_release(struct g_mirror_softc *sc)
1306 {
1307 	struct bio *bp;
1308 
1309 	if ((bp = TAILQ_FIRST(&sc->sc_regular_delayed)) == NULL)
1310 		return;
1311 	if (g_mirror_sync_collision(sc, bp))
1312 		return;
1313 
1314 	G_MIRROR_DEBUG(2, "Requeuing regular requests after collision.");
1315 	mtx_lock(&sc->sc_queue_mtx);
1316 	TAILQ_CONCAT(&sc->sc_regular_delayed, &sc->sc_queue, bio_queue);
1317 	TAILQ_SWAP(&sc->sc_regular_delayed, &sc->sc_queue, bio, bio_queue);
1318 	mtx_unlock(&sc->sc_queue_mtx);
1319 }
1320 
1321 /*
1322  * Releases delayed sync requests which don't collide anymore with regular
1323  * requests.
1324  */
1325 static void
1326 g_mirror_sync_release(struct g_mirror_softc *sc)
1327 {
1328 	struct bio *bp, *bp2;
1329 
1330 	TAILQ_FOREACH_SAFE(bp, &sc->sc_sync_delayed, bio_queue, bp2) {
1331 		if (g_mirror_regular_collision(sc, bp))
1332 			continue;
1333 		TAILQ_REMOVE(&sc->sc_sync_delayed, bp, bio_queue);
1334 		G_MIRROR_LOGREQ(2, bp,
1335 		    "Releasing delayed synchronization request.");
1336 		g_io_request(bp, bp->bio_from);
1337 	}
1338 }
1339 
1340 /*
1341  * Free a synchronization request and clear its slot in the array.
1342  */
1343 static void
1344 g_mirror_sync_request_free(struct g_mirror_disk *disk, struct bio *bp)
1345 {
1346 	int idx;
1347 
1348 	if (disk != NULL && disk->d_sync.ds_bios != NULL) {
1349 		idx = (int)(uintptr_t)bp->bio_caller1;
1350 		KASSERT(disk->d_sync.ds_bios[idx] == bp,
1351 		    ("unexpected sync BIO at %p:%d", disk, idx));
1352 		disk->d_sync.ds_bios[idx] = NULL;
1353 	}
1354 	free(bp->bio_data, M_MIRROR);
1355 	g_destroy_bio(bp);
1356 }
1357 
1358 /*
1359  * Handle synchronization requests.
1360  * Every synchronization request is a two-step process: first, a read request is
1361  * sent to the mirror provider via the sync consumer. If that request completes
1362  * successfully, it is converted to a write and sent to the disk being
1363  * synchronized. If the write also completes successfully, the synchronization
1364  * offset is advanced and a new read request is submitted.
1365  */
1366 static void
1367 g_mirror_sync_request(struct g_mirror_softc *sc, struct bio *bp)
1368 {
1369 	struct g_mirror_disk *disk;
1370 	struct g_mirror_disk_sync *sync;
1371 
1372 	KASSERT((bp->bio_cmd == BIO_READ &&
1373 	    bp->bio_from->geom == sc->sc_sync.ds_geom) ||
1374 	    (bp->bio_cmd == BIO_WRITE && bp->bio_from->geom == sc->sc_geom),
1375 	    ("Sync BIO %p with unexpected origin", bp));
1376 
1377 	bp->bio_from->index--;
1378 	disk = bp->bio_from->private;
1379 	if (disk == NULL) {
1380 		sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1381 		g_topology_lock();
1382 		g_mirror_kill_consumer(sc, bp->bio_from);
1383 		g_topology_unlock();
1384 		g_mirror_sync_request_free(NULL, bp);
1385 		sx_xlock(&sc->sc_lock);
1386 		return;
1387 	}
1388 
1389 	sync = &disk->d_sync;
1390 
1391 	/*
1392 	 * Synchronization request.
1393 	 */
1394 	switch (bp->bio_cmd) {
1395 	case BIO_READ: {
1396 		struct g_consumer *cp;
1397 
1398 		KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_sync_request_read,
1399 		    bp->bio_error);
1400 
1401 		if (bp->bio_error != 0) {
1402 			G_MIRROR_LOGREQ(0, bp,
1403 			    "Synchronization request failed (error=%d).",
1404 			    bp->bio_error);
1405 
1406 			/*
1407 			 * The read error will trigger a syncid bump, so there's
1408 			 * no need to do that here.
1409 			 *
1410 			 * The read error handling for regular requests will
1411 			 * retry the read from all active mirrors before passing
1412 			 * the error back up, so there's no need to retry here.
1413 			 */
1414 			g_mirror_sync_request_free(disk, bp);
1415 			g_mirror_event_send(disk,
1416 			    G_MIRROR_DISK_STATE_DISCONNECTED,
1417 			    G_MIRROR_EVENT_DONTWAIT);
1418 			return;
1419 		}
1420 		G_MIRROR_LOGREQ(3, bp,
1421 		    "Synchronization request half-finished.");
1422 		bp->bio_cmd = BIO_WRITE;
1423 		bp->bio_cflags = 0;
1424 		cp = disk->d_consumer;
1425 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1426 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1427 		    cp->acr, cp->acw, cp->ace));
1428 		cp->index++;
1429 		g_io_request(bp, cp);
1430 		return;
1431 	}
1432 	case BIO_WRITE: {
1433 		off_t offset;
1434 		int i;
1435 
1436 		KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_sync_request_write,
1437 		    bp->bio_error);
1438 
1439 		if (bp->bio_error != 0) {
1440 			G_MIRROR_LOGREQ(0, bp,
1441 			    "Synchronization request failed (error=%d).",
1442 			    bp->bio_error);
1443 			g_mirror_sync_request_free(disk, bp);
1444 			sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
1445 			g_mirror_event_send(disk,
1446 			    G_MIRROR_DISK_STATE_DISCONNECTED,
1447 			    G_MIRROR_EVENT_DONTWAIT);
1448 			return;
1449 		}
1450 		G_MIRROR_LOGREQ(3, bp, "Synchronization request finished.");
1451 		if (sync->ds_offset >= sc->sc_mediasize ||
1452 		    sync->ds_consumer == NULL ||
1453 		    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1454 			/* Don't send more synchronization requests. */
1455 			sync->ds_inflight--;
1456 			g_mirror_sync_request_free(disk, bp);
1457 			if (sync->ds_inflight > 0)
1458 				return;
1459 			if (sync->ds_consumer == NULL ||
1460 			    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1461 				return;
1462 			}
1463 			/* Disk up-to-date, activate it. */
1464 			g_mirror_event_send(disk, G_MIRROR_DISK_STATE_ACTIVE,
1465 			    G_MIRROR_EVENT_DONTWAIT);
1466 			return;
1467 		}
1468 
1469 		/* Send next synchronization request. */
1470 		g_mirror_sync_reinit(disk, bp, sync->ds_offset);
1471 		sync->ds_offset += bp->bio_length;
1472 
1473 		G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
1474 		sync->ds_consumer->index++;
1475 
1476 		/*
1477 		 * Delay the request if it is colliding with a regular request.
1478 		 */
1479 		if (g_mirror_regular_collision(sc, bp))
1480 			g_mirror_sync_delay(sc, bp);
1481 		else
1482 			g_io_request(bp, sync->ds_consumer);
1483 
1484 		/* Requeue delayed requests if possible. */
1485 		g_mirror_regular_release(sc);
1486 
1487 		/* Find the smallest offset */
1488 		offset = sc->sc_mediasize;
1489 		for (i = 0; i < g_mirror_syncreqs; i++) {
1490 			bp = sync->ds_bios[i];
1491 			if (bp != NULL && bp->bio_offset < offset)
1492 				offset = bp->bio_offset;
1493 		}
1494 		if (g_mirror_sync_period > 0 &&
1495 		    time_uptime - sync->ds_update_ts > g_mirror_sync_period) {
1496 			sync->ds_offset_done = offset;
1497 			g_mirror_update_metadata(disk);
1498 			sync->ds_update_ts = time_uptime;
1499 		}
1500 		return;
1501 	}
1502 	default:
1503 		panic("Invalid I/O request %p", bp);
1504 	}
1505 }
1506 
1507 static void
1508 g_mirror_request_prefer(struct g_mirror_softc *sc, struct bio *bp)
1509 {
1510 	struct g_mirror_disk *disk;
1511 	struct g_consumer *cp;
1512 	struct bio *cbp;
1513 
1514 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1515 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE)
1516 			break;
1517 	}
1518 	if (disk == NULL) {
1519 		if (bp->bio_error == 0)
1520 			bp->bio_error = ENXIO;
1521 		g_io_deliver(bp, bp->bio_error);
1522 		return;
1523 	}
1524 	cbp = g_clone_bio(bp);
1525 	if (cbp == NULL) {
1526 		if (bp->bio_error == 0)
1527 			bp->bio_error = ENOMEM;
1528 		g_io_deliver(bp, bp->bio_error);
1529 		return;
1530 	}
1531 	/*
1532 	 * Fill in the component buf structure.
1533 	 */
1534 	cp = disk->d_consumer;
1535 	cbp->bio_done = g_mirror_done;
1536 	cbp->bio_to = cp->provider;
1537 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1538 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1539 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1540 	    cp->acw, cp->ace));
1541 	cp->index++;
1542 	g_io_request(cbp, cp);
1543 }
1544 
1545 static void
1546 g_mirror_request_round_robin(struct g_mirror_softc *sc, struct bio *bp)
1547 {
1548 	struct g_mirror_disk *disk;
1549 	struct g_consumer *cp;
1550 	struct bio *cbp;
1551 
1552 	disk = g_mirror_get_disk(sc);
1553 	if (disk == NULL) {
1554 		if (bp->bio_error == 0)
1555 			bp->bio_error = ENXIO;
1556 		g_io_deliver(bp, bp->bio_error);
1557 		return;
1558 	}
1559 	cbp = g_clone_bio(bp);
1560 	if (cbp == NULL) {
1561 		if (bp->bio_error == 0)
1562 			bp->bio_error = ENOMEM;
1563 		g_io_deliver(bp, bp->bio_error);
1564 		return;
1565 	}
1566 	/*
1567 	 * Fill in the component buf structure.
1568 	 */
1569 	cp = disk->d_consumer;
1570 	cbp->bio_done = g_mirror_done;
1571 	cbp->bio_to = cp->provider;
1572 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1573 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1574 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1575 	    cp->acw, cp->ace));
1576 	cp->index++;
1577 	g_io_request(cbp, cp);
1578 }
1579 
1580 #define TRACK_SIZE  (1 * 1024 * 1024)
1581 #define LOAD_SCALE	256
1582 #define ABS(x)		(((x) >= 0) ? (x) : (-(x)))
1583 
1584 static void
1585 g_mirror_request_load(struct g_mirror_softc *sc, struct bio *bp)
1586 {
1587 	struct g_mirror_disk *disk, *dp;
1588 	struct g_consumer *cp;
1589 	struct bio *cbp;
1590 	int prio, best;
1591 
1592 	/* Find a disk with the smallest load. */
1593 	disk = NULL;
1594 	best = INT_MAX;
1595 	LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1596 		if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1597 			continue;
1598 		prio = dp->load;
1599 		/* If disk head is precisely in position - highly prefer it. */
1600 		if (dp->d_last_offset == bp->bio_offset)
1601 			prio -= 2 * LOAD_SCALE;
1602 		else
1603 		/* If disk head is close to position - prefer it. */
1604 		if (ABS(dp->d_last_offset - bp->bio_offset) < TRACK_SIZE)
1605 			prio -= 1 * LOAD_SCALE;
1606 		if (prio <= best) {
1607 			disk = dp;
1608 			best = prio;
1609 		}
1610 	}
1611 	KASSERT(disk != NULL, ("NULL disk for %s.", sc->sc_name));
1612 	cbp = g_clone_bio(bp);
1613 	if (cbp == NULL) {
1614 		if (bp->bio_error == 0)
1615 			bp->bio_error = ENOMEM;
1616 		g_io_deliver(bp, bp->bio_error);
1617 		return;
1618 	}
1619 	/*
1620 	 * Fill in the component buf structure.
1621 	 */
1622 	cp = disk->d_consumer;
1623 	cbp->bio_done = g_mirror_done;
1624 	cbp->bio_to = cp->provider;
1625 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1626 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1627 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1628 	    cp->acw, cp->ace));
1629 	cp->index++;
1630 	/* Remember last head position */
1631 	disk->d_last_offset = bp->bio_offset + bp->bio_length;
1632 	/* Update loads. */
1633 	LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1634 		dp->load = (dp->d_consumer->index * LOAD_SCALE +
1635 		    dp->load * 7) / 8;
1636 	}
1637 	g_io_request(cbp, cp);
1638 }
1639 
1640 static void
1641 g_mirror_request_split(struct g_mirror_softc *sc, struct bio *bp)
1642 {
1643 	struct bio_queue queue;
1644 	struct g_mirror_disk *disk;
1645 	struct g_consumer *cp __diagused;
1646 	struct bio *cbp;
1647 	off_t left, mod, offset, slice;
1648 	u_char *data;
1649 	u_int ndisks;
1650 
1651 	if (bp->bio_length <= sc->sc_slice) {
1652 		g_mirror_request_round_robin(sc, bp);
1653 		return;
1654 	}
1655 	ndisks = g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE);
1656 	slice = bp->bio_length / ndisks;
1657 	mod = slice % sc->sc_provider->sectorsize;
1658 	if (mod != 0)
1659 		slice += sc->sc_provider->sectorsize - mod;
1660 	/*
1661 	 * Allocate all bios before sending any request, so we can
1662 	 * return ENOMEM in nice and clean way.
1663 	 */
1664 	left = bp->bio_length;
1665 	offset = bp->bio_offset;
1666 	data = bp->bio_data;
1667 	TAILQ_INIT(&queue);
1668 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1669 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1670 			continue;
1671 		cbp = g_clone_bio(bp);
1672 		if (cbp == NULL) {
1673 			while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1674 				TAILQ_REMOVE(&queue, cbp, bio_queue);
1675 				g_destroy_bio(cbp);
1676 			}
1677 			if (bp->bio_error == 0)
1678 				bp->bio_error = ENOMEM;
1679 			g_io_deliver(bp, bp->bio_error);
1680 			return;
1681 		}
1682 		TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
1683 		cbp->bio_done = g_mirror_done;
1684 		cbp->bio_caller1 = disk;
1685 		cbp->bio_to = disk->d_consumer->provider;
1686 		cbp->bio_offset = offset;
1687 		cbp->bio_data = data;
1688 		cbp->bio_length = MIN(left, slice);
1689 		left -= cbp->bio_length;
1690 		if (left == 0)
1691 			break;
1692 		offset += cbp->bio_length;
1693 		data += cbp->bio_length;
1694 	}
1695 	while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1696 		TAILQ_REMOVE(&queue, cbp, bio_queue);
1697 		G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1698 		disk = cbp->bio_caller1;
1699 		cbp->bio_caller1 = NULL;
1700 		cp = disk->d_consumer;
1701 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1702 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1703 		    cp->acr, cp->acw, cp->ace));
1704 		disk->d_consumer->index++;
1705 		g_io_request(cbp, disk->d_consumer);
1706 	}
1707 }
1708 
1709 static void
1710 g_mirror_register_request(struct g_mirror_softc *sc, struct bio *bp)
1711 {
1712 	struct bio_queue queue;
1713 	struct bio *cbp;
1714 	struct g_consumer *cp;
1715 	struct g_mirror_disk *disk;
1716 
1717 	sx_assert(&sc->sc_lock, SA_XLOCKED);
1718 
1719 	/*
1720 	 * To avoid ordering issues, if a write is deferred because of a
1721 	 * collision with a sync request, all I/O is deferred until that
1722 	 * write is initiated.
1723 	 */
1724 	if (bp->bio_from->geom != sc->sc_sync.ds_geom &&
1725 	    !TAILQ_EMPTY(&sc->sc_regular_delayed)) {
1726 		g_mirror_regular_delay(sc, bp);
1727 		return;
1728 	}
1729 
1730 	switch (bp->bio_cmd) {
1731 	case BIO_READ:
1732 		switch (sc->sc_balance) {
1733 		case G_MIRROR_BALANCE_LOAD:
1734 			g_mirror_request_load(sc, bp);
1735 			break;
1736 		case G_MIRROR_BALANCE_PREFER:
1737 			g_mirror_request_prefer(sc, bp);
1738 			break;
1739 		case G_MIRROR_BALANCE_ROUND_ROBIN:
1740 			g_mirror_request_round_robin(sc, bp);
1741 			break;
1742 		case G_MIRROR_BALANCE_SPLIT:
1743 			g_mirror_request_split(sc, bp);
1744 			break;
1745 		}
1746 		return;
1747 	case BIO_WRITE:
1748 	case BIO_DELETE:
1749 		/*
1750 		 * Delay the request if it is colliding with a synchronization
1751 		 * request.
1752 		 */
1753 		if (g_mirror_sync_collision(sc, bp)) {
1754 			g_mirror_regular_delay(sc, bp);
1755 			return;
1756 		}
1757 
1758 		if (sc->sc_idle)
1759 			g_mirror_unidle(sc);
1760 		else
1761 			sc->sc_last_write = time_uptime;
1762 
1763 		/*
1764 		 * Bump syncid on first write.
1765 		 */
1766 		if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0) {
1767 			sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
1768 			g_mirror_bump_syncid(sc);
1769 		}
1770 
1771 		/*
1772 		 * Allocate all bios before sending any request, so we can
1773 		 * return ENOMEM in nice and clean way.
1774 		 */
1775 		TAILQ_INIT(&queue);
1776 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1777 			switch (disk->d_state) {
1778 			case G_MIRROR_DISK_STATE_ACTIVE:
1779 				break;
1780 			case G_MIRROR_DISK_STATE_SYNCHRONIZING:
1781 				if (bp->bio_offset >= disk->d_sync.ds_offset)
1782 					continue;
1783 				break;
1784 			default:
1785 				continue;
1786 			}
1787 			if (bp->bio_cmd == BIO_DELETE &&
1788 			    (disk->d_flags & G_MIRROR_DISK_FLAG_CANDELETE) == 0)
1789 				continue;
1790 			cbp = g_clone_bio(bp);
1791 			if (cbp == NULL) {
1792 				while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1793 					TAILQ_REMOVE(&queue, cbp, bio_queue);
1794 					g_destroy_bio(cbp);
1795 				}
1796 				if (bp->bio_error == 0)
1797 					bp->bio_error = ENOMEM;
1798 				g_io_deliver(bp, bp->bio_error);
1799 				return;
1800 			}
1801 			TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
1802 			cbp->bio_done = g_mirror_done;
1803 			cp = disk->d_consumer;
1804 			cbp->bio_caller1 = cp;
1805 			cbp->bio_to = cp->provider;
1806 			KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1807 			    ("Consumer %s not opened (r%dw%de%d).",
1808 			    cp->provider->name, cp->acr, cp->acw, cp->ace));
1809 		}
1810 		if (TAILQ_EMPTY(&queue)) {
1811 			KASSERT(bp->bio_cmd == BIO_DELETE,
1812 			    ("No consumers for regular request %p", bp));
1813 			g_io_deliver(bp, EOPNOTSUPP);
1814 			return;
1815 		}
1816 		while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1817 			G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1818 			TAILQ_REMOVE(&queue, cbp, bio_queue);
1819 			cp = cbp->bio_caller1;
1820 			cbp->bio_caller1 = NULL;
1821 			cp->index++;
1822 			sc->sc_writes++;
1823 			g_io_request(cbp, cp);
1824 		}
1825 		/*
1826 		 * Put request onto inflight queue, so we can check if new
1827 		 * synchronization requests don't collide with it.
1828 		 */
1829 		TAILQ_INSERT_TAIL(&sc->sc_inflight, bp, bio_queue);
1830 		return;
1831 	case BIO_SPEEDUP:
1832 	case BIO_FLUSH:
1833 		TAILQ_INIT(&queue);
1834 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1835 			if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1836 				continue;
1837 			cbp = g_clone_bio(bp);
1838 			if (cbp == NULL) {
1839 				while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1840 					TAILQ_REMOVE(&queue, cbp, bio_queue);
1841 					g_destroy_bio(cbp);
1842 				}
1843 				if (bp->bio_error == 0)
1844 					bp->bio_error = ENOMEM;
1845 				g_io_deliver(bp, bp->bio_error);
1846 				return;
1847 			}
1848 			TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
1849 			cbp->bio_done = g_mirror_done;
1850 			cbp->bio_caller1 = disk;
1851 			cbp->bio_to = disk->d_consumer->provider;
1852 		}
1853 		KASSERT(!TAILQ_EMPTY(&queue),
1854 		    ("No consumers for regular request %p", bp));
1855 		while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1856 			G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1857 			TAILQ_REMOVE(&queue, cbp, bio_queue);
1858 			disk = cbp->bio_caller1;
1859 			cbp->bio_caller1 = NULL;
1860 			cp = disk->d_consumer;
1861 			KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1862 			    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1863 			    cp->acr, cp->acw, cp->ace));
1864 			cp->index++;
1865 			g_io_request(cbp, cp);
1866 		}
1867 		break;
1868 	default:
1869 		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1870 		    bp->bio_cmd, sc->sc_name));
1871 		break;
1872 	}
1873 }
1874 
1875 static int
1876 g_mirror_can_destroy(struct g_mirror_softc *sc)
1877 {
1878 	struct g_geom *gp;
1879 	struct g_consumer *cp;
1880 
1881 	g_topology_assert();
1882 	gp = sc->sc_geom;
1883 	if (gp->softc == NULL)
1884 		return (1);
1885 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_TASTING) != 0)
1886 		return (0);
1887 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1888 		if (g_mirror_is_busy(sc, cp))
1889 			return (0);
1890 	}
1891 	gp = sc->sc_sync.ds_geom;
1892 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1893 		if (g_mirror_is_busy(sc, cp))
1894 			return (0);
1895 	}
1896 	G_MIRROR_DEBUG(2, "No I/O requests for %s, it can be destroyed.",
1897 	    sc->sc_name);
1898 	return (1);
1899 }
1900 
1901 static int
1902 g_mirror_try_destroy(struct g_mirror_softc *sc)
1903 {
1904 
1905 	if (sc->sc_rootmount != NULL) {
1906 		G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
1907 		    sc->sc_rootmount);
1908 		root_mount_rel(sc->sc_rootmount);
1909 		sc->sc_rootmount = NULL;
1910 	}
1911 	g_topology_lock();
1912 	if (!g_mirror_can_destroy(sc)) {
1913 		g_topology_unlock();
1914 		return (0);
1915 	}
1916 	sc->sc_geom->softc = NULL;
1917 	sc->sc_sync.ds_geom->softc = NULL;
1918 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DRAIN) != 0) {
1919 		g_topology_unlock();
1920 		G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1921 		    &sc->sc_worker);
1922 		/* Unlock sc_lock here, as it can be destroyed after wakeup. */
1923 		sx_xunlock(&sc->sc_lock);
1924 		wakeup(&sc->sc_worker);
1925 		sc->sc_worker = NULL;
1926 	} else {
1927 		g_topology_unlock();
1928 		g_mirror_destroy_device(sc);
1929 	}
1930 	return (1);
1931 }
1932 
1933 /*
1934  * Worker thread.
1935  */
1936 static void
1937 g_mirror_worker(void *arg)
1938 {
1939 	struct g_mirror_softc *sc;
1940 	struct g_mirror_event *ep;
1941 	struct bio *bp;
1942 	int timeout;
1943 
1944 	sc = arg;
1945 	thread_lock(curthread);
1946 	sched_prio(curthread, PRIBIO);
1947 	thread_unlock(curthread);
1948 
1949 	sx_xlock(&sc->sc_lock);
1950 	for (;;) {
1951 		G_MIRROR_DEBUG(5, "%s: Let's see...", __func__);
1952 		/*
1953 		 * First take a look at events.
1954 		 * This is important to handle events before any I/O requests.
1955 		 */
1956 		ep = g_mirror_event_first(sc);
1957 		if (ep != NULL) {
1958 			g_mirror_event_remove(sc, ep);
1959 			if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0) {
1960 				/* Update only device status. */
1961 				G_MIRROR_DEBUG(3,
1962 				    "Running event for device %s.",
1963 				    sc->sc_name);
1964 				ep->e_error = 0;
1965 				g_mirror_update_device(sc, true);
1966 			} else {
1967 				/* Update disk status. */
1968 				G_MIRROR_DEBUG(3, "Running event for disk %s.",
1969 				     g_mirror_get_diskname(ep->e_disk));
1970 				ep->e_error = g_mirror_update_disk(ep->e_disk,
1971 				    ep->e_state);
1972 				if (ep->e_error == 0)
1973 					g_mirror_update_device(sc, false);
1974 			}
1975 			if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0) {
1976 				KASSERT(ep->e_error == 0,
1977 				    ("Error cannot be handled."));
1978 				g_mirror_event_free(ep);
1979 			} else {
1980 				ep->e_flags |= G_MIRROR_EVENT_DONE;
1981 				G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1982 				    ep);
1983 				mtx_lock(&sc->sc_events_mtx);
1984 				wakeup(ep);
1985 				mtx_unlock(&sc->sc_events_mtx);
1986 			}
1987 			if ((sc->sc_flags &
1988 			    G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1989 				if (g_mirror_try_destroy(sc)) {
1990 					curthread->td_pflags &= ~TDP_GEOM;
1991 					G_MIRROR_DEBUG(1, "Thread exiting.");
1992 					kproc_exit(0);
1993 				}
1994 			}
1995 			G_MIRROR_DEBUG(5, "%s: I'm here 1.", __func__);
1996 			continue;
1997 		}
1998 
1999 		/*
2000 		 * Check if we can mark array as CLEAN and if we can't take
2001 		 * how much seconds should we wait.
2002 		 */
2003 		timeout = g_mirror_idle(sc, -1);
2004 
2005 		/*
2006 		 * Handle I/O requests.
2007 		 */
2008 		mtx_lock(&sc->sc_queue_mtx);
2009 		bp = TAILQ_FIRST(&sc->sc_queue);
2010 		if (bp != NULL)
2011 			TAILQ_REMOVE(&sc->sc_queue, bp, bio_queue);
2012 		else {
2013 			if ((sc->sc_flags &
2014 			    G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
2015 				mtx_unlock(&sc->sc_queue_mtx);
2016 				if (g_mirror_try_destroy(sc)) {
2017 					curthread->td_pflags &= ~TDP_GEOM;
2018 					G_MIRROR_DEBUG(1, "Thread exiting.");
2019 					kproc_exit(0);
2020 				}
2021 				mtx_lock(&sc->sc_queue_mtx);
2022 				if (!TAILQ_EMPTY(&sc->sc_queue)) {
2023 					mtx_unlock(&sc->sc_queue_mtx);
2024 					continue;
2025 				}
2026 			}
2027 			if (g_mirror_event_first(sc) != NULL) {
2028 				mtx_unlock(&sc->sc_queue_mtx);
2029 				continue;
2030 			}
2031 			sx_xunlock(&sc->sc_lock);
2032 			MSLEEP(sc, &sc->sc_queue_mtx, PRIBIO | PDROP, "m:w1",
2033 			    timeout * hz);
2034 			sx_xlock(&sc->sc_lock);
2035 			G_MIRROR_DEBUG(5, "%s: I'm here 4.", __func__);
2036 			continue;
2037 		}
2038 		mtx_unlock(&sc->sc_queue_mtx);
2039 
2040 		if (bp->bio_from->geom == sc->sc_sync.ds_geom &&
2041 		    (bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0) {
2042 			/*
2043 			 * Handle completion of the first half (the read) of a
2044 			 * block synchronization operation.
2045 			 */
2046 			g_mirror_sync_request(sc, bp);
2047 		} else if (bp->bio_to != sc->sc_provider) {
2048 			if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_REGULAR) != 0)
2049 				/*
2050 				 * Handle completion of a regular I/O request.
2051 				 */
2052 				g_mirror_regular_request(sc, bp);
2053 			else if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0)
2054 				/*
2055 				 * Handle completion of the second half (the
2056 				 * write) of a block synchronization operation.
2057 				 */
2058 				g_mirror_sync_request(sc, bp);
2059 			else {
2060 				KASSERT(0,
2061 				    ("Invalid request cflags=0x%hx to=%s.",
2062 				    bp->bio_cflags, bp->bio_to->name));
2063 			}
2064 		} else {
2065 			/*
2066 			 * Initiate an I/O request.
2067 			 */
2068 			g_mirror_register_request(sc, bp);
2069 		}
2070 		G_MIRROR_DEBUG(5, "%s: I'm here 9.", __func__);
2071 	}
2072 }
2073 
2074 static void
2075 g_mirror_update_idle(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
2076 {
2077 
2078 	sx_assert(&sc->sc_lock, SX_LOCKED);
2079 
2080 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
2081 		return;
2082 	if (!sc->sc_idle && (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) == 0) {
2083 		G_MIRROR_DEBUG(2, "Disk %s (device %s) marked as dirty.",
2084 		    g_mirror_get_diskname(disk), sc->sc_name);
2085 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
2086 	} else if (sc->sc_idle &&
2087 	    (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
2088 		G_MIRROR_DEBUG(2, "Disk %s (device %s) marked as clean.",
2089 		    g_mirror_get_diskname(disk), sc->sc_name);
2090 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2091 	}
2092 }
2093 
2094 static void
2095 g_mirror_sync_reinit(const struct g_mirror_disk *disk, struct bio *bp,
2096     off_t offset)
2097 {
2098 	void *data;
2099 	int idx;
2100 
2101 	data = bp->bio_data;
2102 	idx = (int)(uintptr_t)bp->bio_caller1;
2103 	g_reset_bio(bp);
2104 
2105 	bp->bio_cmd = BIO_READ;
2106 	bp->bio_data = data;
2107 	bp->bio_done = g_mirror_sync_done;
2108 	bp->bio_from = disk->d_sync.ds_consumer;
2109 	bp->bio_to = disk->d_softc->sc_provider;
2110 	bp->bio_caller1 = (void *)(uintptr_t)idx;
2111 	bp->bio_offset = offset;
2112 	bp->bio_length = MIN(maxphys,
2113 	    disk->d_softc->sc_mediasize - bp->bio_offset);
2114 }
2115 
2116 static void
2117 g_mirror_sync_start(struct g_mirror_disk *disk)
2118 {
2119 	struct g_mirror_softc *sc;
2120 	struct g_mirror_disk_sync *sync;
2121 	struct g_consumer *cp;
2122 	struct bio *bp;
2123 	int error __diagused, i;
2124 
2125 	g_topology_assert_not();
2126 	sc = disk->d_softc;
2127 	sync = &disk->d_sync;
2128 	sx_assert(&sc->sc_lock, SX_LOCKED);
2129 
2130 	KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2131 	    ("Disk %s is not marked for synchronization.",
2132 	    g_mirror_get_diskname(disk)));
2133 	KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2134 	    ("Device not in RUNNING state (%s, %u).", sc->sc_name,
2135 	    sc->sc_state));
2136 
2137 	sx_xunlock(&sc->sc_lock);
2138 	g_topology_lock();
2139 	cp = g_new_consumer(sc->sc_sync.ds_geom);
2140 	cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2141 	error = g_attach(cp, sc->sc_provider);
2142 	KASSERT(error == 0,
2143 	    ("Cannot attach to %s (error=%d).", sc->sc_name, error));
2144 	error = g_access(cp, 1, 0, 0);
2145 	KASSERT(error == 0, ("Cannot open %s (error=%d).", sc->sc_name, error));
2146 	g_topology_unlock();
2147 	sx_xlock(&sc->sc_lock);
2148 
2149 	G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s.", sc->sc_name,
2150 	    g_mirror_get_diskname(disk));
2151 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) == 0)
2152 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
2153 	KASSERT(sync->ds_consumer == NULL,
2154 	    ("Sync consumer already exists (device=%s, disk=%s).",
2155 	    sc->sc_name, g_mirror_get_diskname(disk)));
2156 
2157 	sync->ds_consumer = cp;
2158 	sync->ds_consumer->private = disk;
2159 	sync->ds_consumer->index = 0;
2160 
2161 	/*
2162 	 * Allocate memory for synchronization bios and initialize them.
2163 	 */
2164 	sync->ds_bios = malloc(sizeof(struct bio *) * g_mirror_syncreqs,
2165 	    M_MIRROR, M_WAITOK);
2166 	for (i = 0; i < g_mirror_syncreqs; i++) {
2167 		bp = g_alloc_bio();
2168 		sync->ds_bios[i] = bp;
2169 
2170 		bp->bio_data = malloc(maxphys, M_MIRROR, M_WAITOK);
2171 		bp->bio_caller1 = (void *)(uintptr_t)i;
2172 		g_mirror_sync_reinit(disk, bp, sync->ds_offset);
2173 		sync->ds_offset += bp->bio_length;
2174 	}
2175 
2176 	/* Increase the number of disks in SYNCHRONIZING state. */
2177 	sc->sc_sync.ds_ndisks++;
2178 	/* Set the number of in-flight synchronization requests. */
2179 	sync->ds_inflight = g_mirror_syncreqs;
2180 
2181 	/*
2182 	 * Fire off first synchronization requests.
2183 	 */
2184 	for (i = 0; i < g_mirror_syncreqs; i++) {
2185 		bp = sync->ds_bios[i];
2186 		G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
2187 		sync->ds_consumer->index++;
2188 		/*
2189 		 * Delay the request if it is colliding with a regular request.
2190 		 */
2191 		if (g_mirror_regular_collision(sc, bp))
2192 			g_mirror_sync_delay(sc, bp);
2193 		else
2194 			g_io_request(bp, sync->ds_consumer);
2195 	}
2196 }
2197 
2198 /*
2199  * Stop synchronization process.
2200  * type: 0 - synchronization finished
2201  *       1 - synchronization stopped
2202  */
2203 static void
2204 g_mirror_sync_stop(struct g_mirror_disk *disk, int type)
2205 {
2206 	struct g_mirror_softc *sc;
2207 	struct g_consumer *cp;
2208 
2209 	g_topology_assert_not();
2210 	sc = disk->d_softc;
2211 	sx_assert(&sc->sc_lock, SX_LOCKED);
2212 
2213 	KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2214 	    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2215 	    g_mirror_disk_state2str(disk->d_state)));
2216 	if (disk->d_sync.ds_consumer == NULL)
2217 		return;
2218 
2219 	if (type == 0) {
2220 		G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s finished.",
2221 		    sc->sc_name, g_mirror_get_diskname(disk));
2222 	} else /* if (type == 1) */ {
2223 		G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s stopped.",
2224 		    sc->sc_name, g_mirror_get_diskname(disk));
2225 	}
2226 	g_mirror_regular_release(sc);
2227 	free(disk->d_sync.ds_bios, M_MIRROR);
2228 	disk->d_sync.ds_bios = NULL;
2229 	cp = disk->d_sync.ds_consumer;
2230 	disk->d_sync.ds_consumer = NULL;
2231 	disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2232 	sc->sc_sync.ds_ndisks--;
2233 	sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
2234 	g_topology_lock();
2235 	g_mirror_kill_consumer(sc, cp);
2236 	g_topology_unlock();
2237 	sx_xlock(&sc->sc_lock);
2238 }
2239 
2240 static void
2241 g_mirror_launch_provider(struct g_mirror_softc *sc)
2242 {
2243 	struct g_mirror_disk *disk;
2244 	struct g_provider *pp, *dp;
2245 
2246 	sx_assert(&sc->sc_lock, SX_LOCKED);
2247 
2248 	g_topology_lock();
2249 	pp = g_new_providerf(sc->sc_geom, "mirror/%s", sc->sc_name);
2250 	pp->flags |= G_PF_DIRECT_RECEIVE;
2251 	pp->mediasize = sc->sc_mediasize;
2252 	pp->sectorsize = sc->sc_sectorsize;
2253 	pp->stripesize = 0;
2254 	pp->stripeoffset = 0;
2255 
2256 	/* Splitting of unmapped BIO's could work but isn't implemented now */
2257 	if (sc->sc_balance != G_MIRROR_BALANCE_SPLIT)
2258 		pp->flags |= G_PF_ACCEPT_UNMAPPED;
2259 
2260 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2261 		if (disk->d_consumer && disk->d_consumer->provider) {
2262 			dp = disk->d_consumer->provider;
2263 			if (dp->stripesize > pp->stripesize) {
2264 				pp->stripesize = dp->stripesize;
2265 				pp->stripeoffset = dp->stripeoffset;
2266 			}
2267 			/* A provider underneath us doesn't support unmapped */
2268 			if ((dp->flags & G_PF_ACCEPT_UNMAPPED) == 0) {
2269 				G_MIRROR_DEBUG(0, "Cancelling unmapped "
2270 				    "because of %s.", dp->name);
2271 				pp->flags &= ~G_PF_ACCEPT_UNMAPPED;
2272 			}
2273 		}
2274 	}
2275 	pp->private = sc;
2276 	sc->sc_refcnt++;
2277 	sc->sc_provider = pp;
2278 	g_error_provider(pp, 0);
2279 	g_topology_unlock();
2280 	G_MIRROR_DEBUG(0, "Device %s launched (%u/%u).", pp->name,
2281 	    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE), sc->sc_ndisks);
2282 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2283 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2284 			g_mirror_sync_start(disk);
2285 	}
2286 }
2287 
2288 static void
2289 g_mirror_destroy_provider(struct g_mirror_softc *sc)
2290 {
2291 	struct g_mirror_disk *disk;
2292 	struct bio *bp;
2293 
2294 	g_topology_assert_not();
2295 	KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
2296 	    sc->sc_name));
2297 
2298 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2299 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2300 			g_mirror_sync_stop(disk, 1);
2301 	}
2302 
2303 	g_topology_lock();
2304 	g_error_provider(sc->sc_provider, ENXIO);
2305 	mtx_lock(&sc->sc_queue_mtx);
2306 	while ((bp = TAILQ_FIRST(&sc->sc_queue)) != NULL) {
2307 		TAILQ_REMOVE(&sc->sc_queue, bp, bio_queue);
2308 		/*
2309 		 * Abort any pending I/O that wasn't generated by us.
2310 		 * Synchronization requests and requests destined for individual
2311 		 * mirror components can be destroyed immediately.
2312 		 */
2313 		if (bp->bio_to == sc->sc_provider &&
2314 		    bp->bio_from->geom != sc->sc_sync.ds_geom) {
2315 			g_io_deliver(bp, ENXIO);
2316 		} else {
2317 			if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0)
2318 				free(bp->bio_data, M_MIRROR);
2319 			g_destroy_bio(bp);
2320 		}
2321 	}
2322 	mtx_unlock(&sc->sc_queue_mtx);
2323 	g_wither_provider(sc->sc_provider, ENXIO);
2324 	sc->sc_provider = NULL;
2325 	G_MIRROR_DEBUG(0, "Device %s: provider destroyed.", sc->sc_name);
2326 	g_topology_unlock();
2327 }
2328 
2329 static void
2330 g_mirror_go(void *arg)
2331 {
2332 	struct g_mirror_softc *sc;
2333 	struct g_mirror_event *ep;
2334 
2335 	sc = arg;
2336 	G_MIRROR_DEBUG(0, "Force device %s start due to timeout.", sc->sc_name);
2337 	ep = sc->sc_timeout_event;
2338 	sc->sc_timeout_event = NULL;
2339 	g_mirror_event_dispatch(ep, sc, 0,
2340 	    G_MIRROR_EVENT_DONTWAIT | G_MIRROR_EVENT_DEVICE);
2341 }
2342 
2343 static void
2344 g_mirror_timeout_drain(struct g_mirror_softc *sc)
2345 {
2346 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2347 
2348 	callout_drain(&sc->sc_callout);
2349 	g_mirror_event_free(sc->sc_timeout_event);
2350 	sc->sc_timeout_event = NULL;
2351 }
2352 
2353 static u_int
2354 g_mirror_determine_state(struct g_mirror_disk *disk)
2355 {
2356 	struct g_mirror_softc *sc;
2357 	u_int state;
2358 
2359 	sc = disk->d_softc;
2360 	if (sc->sc_syncid == disk->d_sync.ds_syncid) {
2361 		if ((disk->d_flags &
2362 		    G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0 &&
2363 		    (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 0 ||
2364 		     (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) == 0)) {
2365 			/* Disk does not need synchronization. */
2366 			state = G_MIRROR_DISK_STATE_ACTIVE;
2367 		} else {
2368 			if ((sc->sc_flags &
2369 			     G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2370 			    (disk->d_flags &
2371 			     G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2372 				/*
2373 				 * We can start synchronization from
2374 				 * the stored offset.
2375 				 */
2376 				state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2377 			} else {
2378 				state = G_MIRROR_DISK_STATE_STALE;
2379 			}
2380 		}
2381 	} else if (disk->d_sync.ds_syncid < sc->sc_syncid) {
2382 		/*
2383 		 * Reset all synchronization data for this disk,
2384 		 * because if it even was synchronized, it was
2385 		 * synchronized to disks with different syncid.
2386 		 */
2387 		disk->d_flags |= G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2388 		disk->d_sync.ds_offset = 0;
2389 		disk->d_sync.ds_offset_done = 0;
2390 		disk->d_sync.ds_syncid = sc->sc_syncid;
2391 		if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2392 		    (disk->d_flags & G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2393 			state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2394 		} else {
2395 			state = G_MIRROR_DISK_STATE_STALE;
2396 		}
2397 	} else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
2398 		/*
2399 		 * Not good, NOT GOOD!
2400 		 * It means that mirror was started on stale disks
2401 		 * and more fresh disk just arrive.
2402 		 * If there were writes, mirror is broken, sorry.
2403 		 * I think the best choice here is don't touch
2404 		 * this disk and inform the user loudly.
2405 		 */
2406 		G_MIRROR_DEBUG(0, "Device %s was started before the freshest "
2407 		    "disk (%s) arrives!! It will not be connected to the "
2408 		    "running device.", sc->sc_name,
2409 		    g_mirror_get_diskname(disk));
2410 		g_mirror_destroy_disk(disk);
2411 		state = G_MIRROR_DISK_STATE_NONE;
2412 		/* Return immediately, because disk was destroyed. */
2413 		return (state);
2414 	}
2415 	G_MIRROR_DEBUG(3, "State for %s disk: %s.",
2416 	    g_mirror_get_diskname(disk), g_mirror_disk_state2str(state));
2417 	return (state);
2418 }
2419 
2420 /*
2421  * Update device state.
2422  */
2423 static void
2424 g_mirror_update_device(struct g_mirror_softc *sc, bool force)
2425 {
2426 	struct g_mirror_disk *disk;
2427 	u_int state;
2428 
2429 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2430 
2431 	switch (sc->sc_state) {
2432 	case G_MIRROR_DEVICE_STATE_STARTING:
2433 	    {
2434 		struct g_mirror_disk *pdisk, *tdisk;
2435 		const char *mismatch;
2436 		uintmax_t found, newest;
2437 		u_int dirty, ndisks;
2438 
2439 		/* Pre-flight checks */
2440 		LIST_FOREACH_SAFE(disk, &sc->sc_disks, d_next, tdisk) {
2441 			/*
2442 			 * Confirm we already detected the newest genid.
2443 			 */
2444 			KASSERT(sc->sc_genid >= disk->d_genid,
2445 			    ("%s: found newer genid %u (sc:%p had %u).", __func__,
2446 			    disk->d_genid, sc, sc->sc_genid));
2447 
2448 			/* Kick out any previously tasted stale components. */
2449 			if (disk->d_genid < sc->sc_genid) {
2450 				G_MIRROR_DEBUG(0, "Stale 'genid' field on %s "
2451 				    "(device %s) (component=%u latest=%u), skipping.",
2452 				    g_mirror_get_diskname(disk), sc->sc_name,
2453 				    disk->d_genid, sc->sc_genid);
2454 				g_mirror_destroy_disk(disk);
2455 				sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2456 				continue;
2457 			}
2458 
2459 			/*
2460 			 * Confirm we already detected the newest syncid.
2461 			 */
2462 			KASSERT(sc->sc_syncid >= disk->d_sync.ds_syncid,
2463 			    ("%s: found newer syncid %u (sc:%p had %u).",
2464 			     __func__, disk->d_sync.ds_syncid, sc,
2465 			     sc->sc_syncid));
2466 
2467 #define DETECT_MISMATCH(field, name) \
2468 			if (mismatch == NULL &&					\
2469 			    disk->d_init_ ## field != sc->sc_ ## field) {	\
2470 				mismatch = name;				\
2471 				found = (intmax_t)disk->d_init_ ## field;	\
2472 				newest = (intmax_t)sc->sc_ ## field;		\
2473 			}
2474 			mismatch = NULL;
2475 			DETECT_MISMATCH(ndisks, "md_all");
2476 			DETECT_MISMATCH(balance, "md_balance");
2477 			DETECT_MISMATCH(slice, "md_slice");
2478 			DETECT_MISMATCH(mediasize, "md_mediasize");
2479 #undef DETECT_MISMATCH
2480 			if (mismatch != NULL) {
2481 				G_MIRROR_DEBUG(0, "Found a mismatching '%s' "
2482 				    "field on %s (device %s) (found=%ju "
2483 				    "newest=%ju).", mismatch,
2484 				    g_mirror_get_diskname(disk), sc->sc_name,
2485 				    found, newest);
2486 				g_mirror_destroy_disk(disk);
2487 				sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2488 				continue;
2489 			}
2490 		}
2491 
2492 		KASSERT(sc->sc_provider == NULL,
2493 		    ("Non-NULL provider in STARTING state (%s).", sc->sc_name));
2494 		/*
2495 		 * Are we ready? If the timeout (force is true) has expired, and
2496 		 * any disks are present, then yes. If we're permitted to launch
2497 		 * before the timeout has expired and the expected number of
2498 		 * current-generation mirror disks have been tasted, then yes.
2499 		 */
2500 		ndisks = g_mirror_ndisks(sc, -1);
2501 		if ((force && ndisks > 0) ||
2502 		    (g_launch_mirror_before_timeout && ndisks == sc->sc_ndisks)) {
2503 			;
2504 		} else if (ndisks == 0) {
2505 			/*
2506 			 * Disks went down in starting phase, so destroy
2507 			 * device.
2508 			 */
2509 			g_mirror_timeout_drain(sc);
2510 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2511 			G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2512 			    sc->sc_rootmount);
2513 			root_mount_rel(sc->sc_rootmount);
2514 			sc->sc_rootmount = NULL;
2515 			return;
2516 		} else {
2517 			return;
2518 		}
2519 
2520 		/*
2521 		 * Activate all disks with the biggest syncid.
2522 		 */
2523 		if (force) {
2524 			/*
2525 			 * If 'force' is true, we have been called due to
2526 			 * timeout, so don't bother canceling timeout.
2527 			 */
2528 			ndisks = 0;
2529 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2530 				if ((disk->d_flags &
2531 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2532 					ndisks++;
2533 				}
2534 			}
2535 			if (ndisks == 0) {
2536 				/* No valid disks found, destroy device. */
2537 				sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2538 				G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2539 				    __LINE__, sc->sc_rootmount);
2540 				root_mount_rel(sc->sc_rootmount);
2541 				sc->sc_rootmount = NULL;
2542 				return;
2543 			}
2544 		} else {
2545 			/* Cancel timeout. */
2546 			g_mirror_timeout_drain(sc);
2547 		}
2548 
2549 		/*
2550 		 * Here we need to look for dirty disks and if all disks
2551 		 * with the biggest syncid are dirty, we have to choose
2552 		 * one with the biggest priority and rebuild the rest.
2553 		 */
2554 		/*
2555 		 * Find the number of dirty disks with the biggest syncid.
2556 		 * Find the number of disks with the biggest syncid.
2557 		 * While here, find a disk with the biggest priority.
2558 		 */
2559 		dirty = ndisks = 0;
2560 		pdisk = NULL;
2561 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2562 			if (disk->d_sync.ds_syncid != sc->sc_syncid)
2563 				continue;
2564 			if ((disk->d_flags &
2565 			    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2566 				continue;
2567 			}
2568 			ndisks++;
2569 			if ((disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
2570 				dirty++;
2571 				if (pdisk == NULL ||
2572 				    pdisk->d_priority < disk->d_priority) {
2573 					pdisk = disk;
2574 				}
2575 			}
2576 		}
2577 		if (dirty == 0) {
2578 			/* No dirty disks at all, great. */
2579 		} else if (dirty == ndisks) {
2580 			/*
2581 			 * Force synchronization for all dirty disks except one
2582 			 * with the biggest priority.
2583 			 */
2584 			KASSERT(pdisk != NULL, ("pdisk == NULL"));
2585 			G_MIRROR_DEBUG(1, "Using disk %s (device %s) as a "
2586 			    "master disk for synchronization.",
2587 			    g_mirror_get_diskname(pdisk), sc->sc_name);
2588 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2589 				if (disk->d_sync.ds_syncid != sc->sc_syncid)
2590 					continue;
2591 				if ((disk->d_flags &
2592 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2593 					continue;
2594 				}
2595 				KASSERT((disk->d_flags &
2596 				    G_MIRROR_DISK_FLAG_DIRTY) != 0,
2597 				    ("Disk %s isn't marked as dirty.",
2598 				    g_mirror_get_diskname(disk)));
2599 				/* Skip the disk with the biggest priority. */
2600 				if (disk == pdisk)
2601 					continue;
2602 				disk->d_sync.ds_syncid = 0;
2603 			}
2604 		} else if (dirty < ndisks) {
2605 			/*
2606 			 * Force synchronization for all dirty disks.
2607 			 * We have some non-dirty disks.
2608 			 */
2609 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2610 				if (disk->d_sync.ds_syncid != sc->sc_syncid)
2611 					continue;
2612 				if ((disk->d_flags &
2613 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2614 					continue;
2615 				}
2616 				if ((disk->d_flags &
2617 				    G_MIRROR_DISK_FLAG_DIRTY) == 0) {
2618 					continue;
2619 				}
2620 				disk->d_sync.ds_syncid = 0;
2621 			}
2622 		}
2623 
2624 		/* Reset hint. */
2625 		sc->sc_hint = NULL;
2626 		if (force) {
2627 			/* Remember to bump syncid on first write. */
2628 			sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2629 		}
2630 		state = G_MIRROR_DEVICE_STATE_RUNNING;
2631 		G_MIRROR_DEBUG(1, "Device %s state changed from %s to %s.",
2632 		    sc->sc_name, g_mirror_device_state2str(sc->sc_state),
2633 		    g_mirror_device_state2str(state));
2634 		sc->sc_state = state;
2635 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2636 			state = g_mirror_determine_state(disk);
2637 			g_mirror_event_send(disk, state,
2638 			    G_MIRROR_EVENT_DONTWAIT);
2639 			if (state == G_MIRROR_DISK_STATE_STALE)
2640 				sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2641 		}
2642 		break;
2643 	    }
2644 	case G_MIRROR_DEVICE_STATE_RUNNING:
2645 		if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 0 &&
2646 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2647 			/*
2648 			 * No usable disks, so destroy the device.
2649 			 */
2650 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2651 			break;
2652 		} else if (g_mirror_ndisks(sc,
2653 		    G_MIRROR_DISK_STATE_ACTIVE) > 0 &&
2654 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2655 			/*
2656 			 * We have active disks, launch provider if it doesn't
2657 			 * exist.
2658 			 */
2659 			if (sc->sc_provider == NULL)
2660 				g_mirror_launch_provider(sc);
2661 			if (sc->sc_rootmount != NULL) {
2662 				G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2663 				    __LINE__, sc->sc_rootmount);
2664 				root_mount_rel(sc->sc_rootmount);
2665 				sc->sc_rootmount = NULL;
2666 			}
2667 		}
2668 		/*
2669 		 * Genid should be bumped immediately, so do it here.
2670 		 */
2671 		if ((sc->sc_bump_id & G_MIRROR_BUMP_GENID) != 0) {
2672 			sc->sc_bump_id &= ~G_MIRROR_BUMP_GENID;
2673 			g_mirror_bump_genid(sc);
2674 		}
2675 		if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID_NOW) != 0) {
2676 			sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID_NOW;
2677 			g_mirror_bump_syncid(sc);
2678 		}
2679 		break;
2680 	default:
2681 		KASSERT(1 == 0, ("Wrong device state (%s, %s).",
2682 		    sc->sc_name, g_mirror_device_state2str(sc->sc_state)));
2683 		break;
2684 	}
2685 }
2686 
2687 /*
2688  * Update disk state and device state if needed.
2689  */
2690 #define	DISK_STATE_CHANGED()	G_MIRROR_DEBUG(1,			\
2691 	"Disk %s state changed from %s to %s (device %s).",		\
2692 	g_mirror_get_diskname(disk),					\
2693 	g_mirror_disk_state2str(disk->d_state),				\
2694 	g_mirror_disk_state2str(state), sc->sc_name)
2695 static int
2696 g_mirror_update_disk(struct g_mirror_disk *disk, u_int state)
2697 {
2698 	struct g_mirror_softc *sc;
2699 
2700 	sc = disk->d_softc;
2701 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2702 
2703 again:
2704 	G_MIRROR_DEBUG(3, "Changing disk %s state from %s to %s.",
2705 	    g_mirror_get_diskname(disk), g_mirror_disk_state2str(disk->d_state),
2706 	    g_mirror_disk_state2str(state));
2707 	switch (state) {
2708 	case G_MIRROR_DISK_STATE_NEW:
2709 		/*
2710 		 * Possible scenarios:
2711 		 * 1. New disk arrive.
2712 		 */
2713 		/* Previous state should be NONE. */
2714 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NONE,
2715 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2716 		    g_mirror_disk_state2str(disk->d_state)));
2717 		DISK_STATE_CHANGED();
2718 
2719 		disk->d_state = state;
2720 		g_topology_lock();
2721 		if (LIST_EMPTY(&sc->sc_disks))
2722 			LIST_INSERT_HEAD(&sc->sc_disks, disk, d_next);
2723 		else {
2724 			struct g_mirror_disk *dp;
2725 
2726 			LIST_FOREACH(dp, &sc->sc_disks, d_next) {
2727 				if (disk->d_priority >= dp->d_priority) {
2728 					LIST_INSERT_BEFORE(dp, disk, d_next);
2729 					dp = NULL;
2730 					break;
2731 				}
2732 				if (LIST_NEXT(dp, d_next) == NULL)
2733 					break;
2734 			}
2735 			if (dp != NULL)
2736 				LIST_INSERT_AFTER(dp, disk, d_next);
2737 		}
2738 		g_topology_unlock();
2739 		G_MIRROR_DEBUG(1, "Device %s: provider %s detected.",
2740 		    sc->sc_name, g_mirror_get_diskname(disk));
2741 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
2742 			break;
2743 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2744 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2745 		    g_mirror_device_state2str(sc->sc_state),
2746 		    g_mirror_get_diskname(disk),
2747 		    g_mirror_disk_state2str(disk->d_state)));
2748 		state = g_mirror_determine_state(disk);
2749 		if (state != G_MIRROR_DISK_STATE_NONE)
2750 			goto again;
2751 		break;
2752 	case G_MIRROR_DISK_STATE_ACTIVE:
2753 		/*
2754 		 * Possible scenarios:
2755 		 * 1. New disk does not need synchronization.
2756 		 * 2. Synchronization process finished successfully.
2757 		 */
2758 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2759 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2760 		    g_mirror_device_state2str(sc->sc_state),
2761 		    g_mirror_get_diskname(disk),
2762 		    g_mirror_disk_state2str(disk->d_state)));
2763 		/* Previous state should be NEW or SYNCHRONIZING. */
2764 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW ||
2765 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2766 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2767 		    g_mirror_disk_state2str(disk->d_state)));
2768 		DISK_STATE_CHANGED();
2769 
2770 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
2771 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2772 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_FORCE_SYNC;
2773 			g_mirror_sync_stop(disk, 0);
2774 		}
2775 		disk->d_state = state;
2776 		disk->d_sync.ds_offset = 0;
2777 		disk->d_sync.ds_offset_done = 0;
2778 		g_mirror_update_idle(sc, disk);
2779 		g_mirror_update_metadata(disk);
2780 		G_MIRROR_DEBUG(1, "Device %s: provider %s activated.",
2781 		    sc->sc_name, g_mirror_get_diskname(disk));
2782 		break;
2783 	case G_MIRROR_DISK_STATE_STALE:
2784 		/*
2785 		 * Possible scenarios:
2786 		 * 1. Stale disk was connected.
2787 		 */
2788 		/* Previous state should be NEW. */
2789 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2790 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2791 		    g_mirror_disk_state2str(disk->d_state)));
2792 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2793 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2794 		    g_mirror_device_state2str(sc->sc_state),
2795 		    g_mirror_get_diskname(disk),
2796 		    g_mirror_disk_state2str(disk->d_state)));
2797 		/*
2798 		 * STALE state is only possible if device is marked
2799 		 * NOAUTOSYNC.
2800 		 */
2801 		KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) != 0,
2802 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2803 		    g_mirror_device_state2str(sc->sc_state),
2804 		    g_mirror_get_diskname(disk),
2805 		    g_mirror_disk_state2str(disk->d_state)));
2806 		DISK_STATE_CHANGED();
2807 
2808 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2809 		disk->d_state = state;
2810 		g_mirror_update_metadata(disk);
2811 		G_MIRROR_DEBUG(0, "Device %s: provider %s is stale.",
2812 		    sc->sc_name, g_mirror_get_diskname(disk));
2813 		break;
2814 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
2815 		/*
2816 		 * Possible scenarios:
2817 		 * 1. Disk which needs synchronization was connected.
2818 		 */
2819 		/* Previous state should be NEW. */
2820 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2821 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2822 		    g_mirror_disk_state2str(disk->d_state)));
2823 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2824 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2825 		    g_mirror_device_state2str(sc->sc_state),
2826 		    g_mirror_get_diskname(disk),
2827 		    g_mirror_disk_state2str(disk->d_state)));
2828 		DISK_STATE_CHANGED();
2829 
2830 		if (disk->d_state == G_MIRROR_DISK_STATE_NEW)
2831 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2832 		disk->d_state = state;
2833 		if (sc->sc_provider != NULL) {
2834 			g_mirror_sync_start(disk);
2835 			g_mirror_update_metadata(disk);
2836 		}
2837 		break;
2838 	case G_MIRROR_DISK_STATE_DISCONNECTED:
2839 		/*
2840 		 * Possible scenarios:
2841 		 * 1. Device wasn't running yet, but disk disappear.
2842 		 * 2. Disk was active and disapppear.
2843 		 * 3. Disk disappear during synchronization process.
2844 		 */
2845 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING) {
2846 			/*
2847 			 * Previous state should be ACTIVE, STALE or
2848 			 * SYNCHRONIZING.
2849 			 */
2850 			KASSERT(disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
2851 			    disk->d_state == G_MIRROR_DISK_STATE_STALE ||
2852 			    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2853 			    ("Wrong disk state (%s, %s).",
2854 			    g_mirror_get_diskname(disk),
2855 			    g_mirror_disk_state2str(disk->d_state)));
2856 		} else if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING) {
2857 			/* Previous state should be NEW. */
2858 			KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2859 			    ("Wrong disk state (%s, %s).",
2860 			    g_mirror_get_diskname(disk),
2861 			    g_mirror_disk_state2str(disk->d_state)));
2862 			/*
2863 			 * Reset bumping syncid if disk disappeared in STARTING
2864 			 * state.
2865 			 */
2866 			if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0)
2867 				sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
2868 #ifdef	INVARIANTS
2869 		} else {
2870 			KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
2871 			    sc->sc_name,
2872 			    g_mirror_device_state2str(sc->sc_state),
2873 			    g_mirror_get_diskname(disk),
2874 			    g_mirror_disk_state2str(disk->d_state)));
2875 #endif
2876 		}
2877 		DISK_STATE_CHANGED();
2878 		G_MIRROR_DEBUG(0, "Device %s: provider %s disconnected.",
2879 		    sc->sc_name, g_mirror_get_diskname(disk));
2880 
2881 		g_mirror_destroy_disk(disk);
2882 		break;
2883 	case G_MIRROR_DISK_STATE_DESTROY:
2884 	    {
2885 		int error;
2886 
2887 		error = g_mirror_clear_metadata(disk);
2888 		if (error != 0) {
2889 			G_MIRROR_DEBUG(0,
2890 			    "Device %s: failed to clear metadata on %s: %d.",
2891 			    sc->sc_name, g_mirror_get_diskname(disk), error);
2892 			break;
2893 		}
2894 		DISK_STATE_CHANGED();
2895 		G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.",
2896 		    sc->sc_name, g_mirror_get_diskname(disk));
2897 
2898 		g_mirror_destroy_disk(disk);
2899 		sc->sc_ndisks--;
2900 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2901 			g_mirror_update_metadata(disk);
2902 		}
2903 		break;
2904 	    }
2905 	default:
2906 		KASSERT(1 == 0, ("Unknown state (%u).", state));
2907 		break;
2908 	}
2909 	return (0);
2910 }
2911 #undef	DISK_STATE_CHANGED
2912 
2913 int
2914 g_mirror_read_metadata(struct g_consumer *cp, struct g_mirror_metadata *md)
2915 {
2916 	struct g_provider *pp;
2917 	u_char *buf;
2918 	int error;
2919 
2920 	g_topology_assert();
2921 
2922 	error = g_access(cp, 1, 0, 0);
2923 	if (error != 0)
2924 		return (error);
2925 	pp = cp->provider;
2926 	g_topology_unlock();
2927 	/* Metadata are stored on last sector. */
2928 	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
2929 	    &error);
2930 	g_topology_lock();
2931 	g_access(cp, -1, 0, 0);
2932 	if (buf == NULL) {
2933 		G_MIRROR_DEBUG(1, "Cannot read metadata from %s (error=%d).",
2934 		    cp->provider->name, error);
2935 		return (error);
2936 	}
2937 
2938 	/* Decode metadata. */
2939 	error = mirror_metadata_decode(buf, md);
2940 	g_free(buf);
2941 	if (strcmp(md->md_magic, G_MIRROR_MAGIC) != 0)
2942 		return (EINVAL);
2943 	if (md->md_version > G_MIRROR_VERSION) {
2944 		G_MIRROR_DEBUG(0,
2945 		    "Kernel module is too old to handle metadata from %s.",
2946 		    cp->provider->name);
2947 		return (EINVAL);
2948 	}
2949 	if (error != 0) {
2950 		G_MIRROR_DEBUG(1, "MD5 metadata hash mismatch for provider %s.",
2951 		    cp->provider->name);
2952 		return (error);
2953 	}
2954 
2955 	return (0);
2956 }
2957 
2958 static int
2959 g_mirror_check_metadata(struct g_mirror_softc *sc, struct g_provider *pp,
2960     struct g_mirror_metadata *md)
2961 {
2962 
2963 	G_MIRROR_DEBUG(2, "%s: md_did 0x%u disk %s device %s md_all 0x%x "
2964 	    "sc_ndisks 0x%x md_slice 0x%x sc_slice 0x%x md_balance 0x%x "
2965 	    "sc_balance 0x%x sc_mediasize 0x%jx pp_mediasize 0x%jx "
2966 	    "md_sectorsize 0x%x sc_sectorsize 0x%x md_mflags 0x%jx "
2967 	    "md_dflags 0x%jx md_syncid 0x%x md_genid 0x%x md_priority 0x%x "
2968 	    "sc_state 0x%x.",
2969 	    __func__, md->md_did, pp->name, sc->sc_name, md->md_all,
2970 	    sc->sc_ndisks, md->md_slice, sc->sc_slice, md->md_balance,
2971 	    sc->sc_balance, (uintmax_t)sc->sc_mediasize,
2972 	    (uintmax_t)pp->mediasize, md->md_sectorsize, sc->sc_sectorsize,
2973 	    (uintmax_t)md->md_mflags, (uintmax_t)md->md_dflags, md->md_syncid,
2974 	    md->md_genid, md->md_priority, sc->sc_state);
2975 
2976 	if (g_mirror_id2disk(sc, md->md_did) != NULL) {
2977 		G_MIRROR_DEBUG(1, "Disk %s (id=%u) already exists, skipping.",
2978 		    pp->name, md->md_did);
2979 		return (EEXIST);
2980 	}
2981 	if (sc->sc_mediasize > pp->mediasize) {
2982 		G_MIRROR_DEBUG(1,
2983 		    "Invalid size of disk %s (device %s), skipping.", pp->name,
2984 		    sc->sc_name);
2985 		return (EINVAL);
2986 	}
2987 	if (md->md_sectorsize != sc->sc_sectorsize) {
2988 		G_MIRROR_DEBUG(1,
2989 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2990 		    "md_sectorsize", pp->name, sc->sc_name);
2991 		return (EINVAL);
2992 	}
2993 	if ((sc->sc_sectorsize % pp->sectorsize) != 0) {
2994 		G_MIRROR_DEBUG(1,
2995 		    "Invalid sector size of disk %s (device %s), skipping.",
2996 		    pp->name, sc->sc_name);
2997 		return (EINVAL);
2998 	}
2999 	if ((md->md_mflags & ~G_MIRROR_DEVICE_FLAG_MASK) != 0) {
3000 		G_MIRROR_DEBUG(1,
3001 		    "Invalid device flags on disk %s (device %s), skipping.",
3002 		    pp->name, sc->sc_name);
3003 		return (EINVAL);
3004 	}
3005 	if ((md->md_dflags & ~G_MIRROR_DISK_FLAG_MASK) != 0) {
3006 		G_MIRROR_DEBUG(1,
3007 		    "Invalid disk flags on disk %s (device %s), skipping.",
3008 		    pp->name, sc->sc_name);
3009 		return (EINVAL);
3010 	}
3011 	return (0);
3012 }
3013 
3014 int
3015 g_mirror_add_disk(struct g_mirror_softc *sc, struct g_provider *pp,
3016     struct g_mirror_metadata *md)
3017 {
3018 	struct g_mirror_disk *disk;
3019 	int error;
3020 
3021 	g_topology_assert_not();
3022 	G_MIRROR_DEBUG(2, "Adding disk %s.", pp->name);
3023 
3024 	error = g_mirror_check_metadata(sc, pp, md);
3025 	if (error != 0)
3026 		return (error);
3027 
3028 	if (md->md_genid < sc->sc_genid) {
3029 		G_MIRROR_DEBUG(0, "Component %s (device %s) broken, skipping.",
3030 		    pp->name, sc->sc_name);
3031 		return (EINVAL);
3032 	}
3033 
3034 	/*
3035 	 * If the component disk we're tasting has newer metadata than the
3036 	 * STARTING gmirror device, refresh the device from the component.
3037 	 */
3038 	error = g_mirror_refresh_device(sc, pp, md);
3039 	if (error != 0)
3040 		return (error);
3041 
3042 	disk = g_mirror_init_disk(sc, pp, md, &error);
3043 	if (disk == NULL)
3044 		return (error);
3045 	error = g_mirror_event_send(disk, G_MIRROR_DISK_STATE_NEW,
3046 	    G_MIRROR_EVENT_WAIT);
3047 	if (error != 0)
3048 		return (error);
3049 	if (md->md_version < G_MIRROR_VERSION) {
3050 		G_MIRROR_DEBUG(0, "Upgrading metadata on %s (v%d->v%d).",
3051 		    pp->name, md->md_version, G_MIRROR_VERSION);
3052 		g_mirror_update_metadata(disk);
3053 	}
3054 	return (0);
3055 }
3056 
3057 static void
3058 g_mirror_destroy_delayed(void *arg, int flag)
3059 {
3060 	struct g_mirror_softc *sc;
3061 	int error;
3062 
3063 	if (flag == EV_CANCEL) {
3064 		G_MIRROR_DEBUG(1, "Destroying canceled.");
3065 		return;
3066 	}
3067 	sc = arg;
3068 	g_topology_unlock();
3069 	sx_xlock(&sc->sc_lock);
3070 	KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) == 0,
3071 	    ("DESTROY flag set on %s.", sc->sc_name));
3072 	KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_CLOSEWAIT) != 0,
3073 	    ("CLOSEWAIT flag not set on %s.", sc->sc_name));
3074 	G_MIRROR_DEBUG(1, "Destroying %s (delayed).", sc->sc_name);
3075 	error = g_mirror_destroy(sc, G_MIRROR_DESTROY_SOFT);
3076 	if (error != 0) {
3077 		G_MIRROR_DEBUG(0, "Cannot destroy %s (error=%d).",
3078 		    sc->sc_name, error);
3079 		sx_xunlock(&sc->sc_lock);
3080 	}
3081 	g_topology_lock();
3082 }
3083 
3084 static int
3085 g_mirror_access(struct g_provider *pp, int acr, int acw, int ace)
3086 {
3087 	struct g_mirror_softc *sc;
3088 	int error = 0;
3089 
3090 	g_topology_assert();
3091 	G_MIRROR_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, acr,
3092 	    acw, ace);
3093 
3094 	sc = pp->private;
3095 	KASSERT(sc != NULL, ("NULL softc (provider=%s).", pp->name));
3096 
3097 	g_topology_unlock();
3098 	sx_xlock(&sc->sc_lock);
3099 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0 ||
3100 	    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_CLOSEWAIT) != 0 ||
3101 	    LIST_EMPTY(&sc->sc_disks)) {
3102 		if (acr > 0 || acw > 0 || ace > 0)
3103 			error = ENXIO;
3104 		goto end;
3105 	}
3106 	sc->sc_provider_open += acr + acw + ace;
3107 	if (pp->acw + acw == 0)
3108 		g_mirror_idle(sc, 0);
3109 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_CLOSEWAIT) != 0 &&
3110 	    sc->sc_provider_open == 0)
3111 		g_post_event(g_mirror_destroy_delayed, sc, M_WAITOK, sc, NULL);
3112 end:
3113 	sx_xunlock(&sc->sc_lock);
3114 	g_topology_lock();
3115 	return (error);
3116 }
3117 
3118 static void
3119 g_mirror_reinit_from_metadata(struct g_mirror_softc *sc,
3120     const struct g_mirror_metadata *md)
3121 {
3122 
3123 	sc->sc_genid = md->md_genid;
3124 	sc->sc_syncid = md->md_syncid;
3125 
3126 	sc->sc_slice = md->md_slice;
3127 	sc->sc_balance = md->md_balance;
3128 	sc->sc_mediasize = md->md_mediasize;
3129 	sc->sc_ndisks = md->md_all;
3130 	sc->sc_flags &= ~G_MIRROR_DEVICE_FLAG_MASK;
3131 	sc->sc_flags |= (md->md_mflags & G_MIRROR_DEVICE_FLAG_MASK);
3132 }
3133 
3134 struct g_geom *
3135 g_mirror_create(struct g_class *mp, const struct g_mirror_metadata *md,
3136     u_int type)
3137 {
3138 	struct g_mirror_softc *sc;
3139 	struct g_geom *gp;
3140 	int error, timeout;
3141 
3142 	g_topology_assert();
3143 	G_MIRROR_DEBUG(1, "Creating device %s (id=%u).", md->md_name,
3144 	    md->md_mid);
3145 
3146 	/* One disk is minimum. */
3147 	if (md->md_all < 1)
3148 		return (NULL);
3149 	/*
3150 	 * Action geom.
3151 	 */
3152 	gp = g_new_geomf(mp, "%s", md->md_name);
3153 	sc = malloc(sizeof(*sc), M_MIRROR, M_WAITOK | M_ZERO);
3154 	gp->start = g_mirror_start;
3155 	gp->orphan = g_mirror_orphan;
3156 	gp->access = g_mirror_access;
3157 	gp->dumpconf = g_mirror_dumpconf;
3158 
3159 	sc->sc_type = type;
3160 	sc->sc_id = md->md_mid;
3161 	g_mirror_reinit_from_metadata(sc, md);
3162 	sc->sc_sectorsize = md->md_sectorsize;
3163 	sc->sc_bump_id = 0;
3164 	sc->sc_idle = 1;
3165 	sc->sc_last_write = time_uptime;
3166 	sc->sc_writes = 0;
3167 	sc->sc_refcnt = 1;
3168 	sx_init(&sc->sc_lock, "gmirror:lock");
3169 	TAILQ_INIT(&sc->sc_queue);
3170 	mtx_init(&sc->sc_queue_mtx, "gmirror:queue", NULL, MTX_DEF);
3171 	TAILQ_INIT(&sc->sc_regular_delayed);
3172 	TAILQ_INIT(&sc->sc_inflight);
3173 	TAILQ_INIT(&sc->sc_sync_delayed);
3174 	LIST_INIT(&sc->sc_disks);
3175 	TAILQ_INIT(&sc->sc_events);
3176 	mtx_init(&sc->sc_events_mtx, "gmirror:events", NULL, MTX_DEF);
3177 	callout_init(&sc->sc_callout, 1);
3178 	mtx_init(&sc->sc_done_mtx, "gmirror:done", NULL, MTX_DEF);
3179 	sc->sc_state = G_MIRROR_DEVICE_STATE_STARTING;
3180 	gp->softc = sc;
3181 	sc->sc_geom = gp;
3182 	sc->sc_provider = NULL;
3183 	sc->sc_provider_open = 0;
3184 	/*
3185 	 * Synchronization geom.
3186 	 */
3187 	gp = g_new_geomf(mp, "%s.sync", md->md_name);
3188 	gp->softc = sc;
3189 	gp->orphan = g_mirror_orphan;
3190 	sc->sc_sync.ds_geom = gp;
3191 	sc->sc_sync.ds_ndisks = 0;
3192 	error = kproc_create(g_mirror_worker, sc, &sc->sc_worker, 0, 0,
3193 	    "g_mirror %s", md->md_name);
3194 	if (error != 0) {
3195 		G_MIRROR_DEBUG(1, "Cannot create kernel thread for %s.",
3196 		    sc->sc_name);
3197 		g_destroy_geom(sc->sc_sync.ds_geom);
3198 		g_destroy_geom(sc->sc_geom);
3199 		g_mirror_free_device(sc);
3200 		return (NULL);
3201 	}
3202 
3203 	G_MIRROR_DEBUG(1, "Device %s created (%u components, id=%u).",
3204 	    sc->sc_name, sc->sc_ndisks, sc->sc_id);
3205 
3206 	sc->sc_rootmount = root_mount_hold("GMIRROR");
3207 	G_MIRROR_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
3208 
3209 	/*
3210 	 * Schedule startup timeout.
3211 	 */
3212 	timeout = g_mirror_timeout * hz;
3213 	sc->sc_timeout_event = malloc(sizeof(struct g_mirror_event), M_MIRROR,
3214 	    M_WAITOK);
3215 	callout_reset(&sc->sc_callout, timeout, g_mirror_go, sc);
3216 	return (sc->sc_geom);
3217 }
3218 
3219 int
3220 g_mirror_destroy(struct g_mirror_softc *sc, int how)
3221 {
3222 	struct g_mirror_disk *disk;
3223 
3224 	g_topology_assert_not();
3225 	sx_assert(&sc->sc_lock, SX_XLOCKED);
3226 
3227 	if (sc->sc_provider_open != 0) {
3228 		switch (how) {
3229 		case G_MIRROR_DESTROY_SOFT:
3230 			G_MIRROR_DEBUG(1,
3231 			    "Device %s is still open (%d).", sc->sc_name,
3232 			    sc->sc_provider_open);
3233 			return (EBUSY);
3234 		case G_MIRROR_DESTROY_DELAYED:
3235 			G_MIRROR_DEBUG(1,
3236 			    "Device %s will be destroyed on last close.",
3237 			    sc->sc_name);
3238 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
3239 				if (disk->d_state ==
3240 				    G_MIRROR_DISK_STATE_SYNCHRONIZING) {
3241 					g_mirror_sync_stop(disk, 1);
3242 				}
3243 			}
3244 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_CLOSEWAIT;
3245 			return (EBUSY);
3246 		case G_MIRROR_DESTROY_HARD:
3247 			G_MIRROR_DEBUG(1, "Device %s is still open, so it "
3248 			    "can't be definitely removed.", sc->sc_name);
3249 		}
3250 	}
3251 
3252 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
3253 		sx_xunlock(&sc->sc_lock);
3254 		return (0);
3255 	}
3256 	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
3257 	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DRAIN;
3258 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
3259 	sx_xunlock(&sc->sc_lock);
3260 	mtx_lock(&sc->sc_queue_mtx);
3261 	wakeup(sc);
3262 	mtx_unlock(&sc->sc_queue_mtx);
3263 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
3264 	while (sc->sc_worker != NULL)
3265 		tsleep(&sc->sc_worker, PRIBIO, "m:destroy", hz / 5);
3266 	G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
3267 	sx_xlock(&sc->sc_lock);
3268 	g_mirror_destroy_device(sc);
3269 	return (0);
3270 }
3271 
3272 static void
3273 g_mirror_taste_orphan(struct g_consumer *cp)
3274 {
3275 
3276 	KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
3277 	    cp->provider->name));
3278 }
3279 
3280 static struct g_geom *
3281 g_mirror_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
3282 {
3283 	struct g_mirror_metadata md;
3284 	struct g_mirror_softc *sc;
3285 	struct g_consumer *cp;
3286 	struct g_geom *gp;
3287 	int error;
3288 
3289 	g_topology_assert();
3290 	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
3291 	G_MIRROR_DEBUG(2, "Tasting %s.", pp->name);
3292 
3293 	gp = g_new_geomf(mp, "mirror:taste");
3294 	/*
3295 	 * This orphan function should be never called.
3296 	 */
3297 	gp->orphan = g_mirror_taste_orphan;
3298 	cp = g_new_consumer(gp);
3299 	cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
3300 	error = g_attach(cp, pp);
3301 	if (error == 0) {
3302 		error = g_mirror_read_metadata(cp, &md);
3303 		g_detach(cp);
3304 	}
3305 	g_destroy_consumer(cp);
3306 	g_destroy_geom(gp);
3307 	if (error != 0)
3308 		return (NULL);
3309 	gp = NULL;
3310 
3311 	if (md.md_provider[0] != '\0' &&
3312 	    !g_compare_names(md.md_provider, pp->name))
3313 		return (NULL);
3314 	if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
3315 		return (NULL);
3316 	if ((md.md_dflags & G_MIRROR_DISK_FLAG_INACTIVE) != 0) {
3317 		G_MIRROR_DEBUG(0,
3318 		    "Device %s: provider %s marked as inactive, skipping.",
3319 		    md.md_name, pp->name);
3320 		return (NULL);
3321 	}
3322 	if (g_mirror_debug >= 2)
3323 		mirror_metadata_dump(&md);
3324 
3325 	/*
3326 	 * Let's check if device already exists.
3327 	 */
3328 	sc = NULL;
3329 	LIST_FOREACH(gp, &mp->geom, geom) {
3330 		sc = gp->softc;
3331 		if (sc == NULL)
3332 			continue;
3333 		if (sc->sc_type != G_MIRROR_TYPE_AUTOMATIC)
3334 			continue;
3335 		if (sc->sc_sync.ds_geom == gp)
3336 			continue;
3337 		if (strcmp(md.md_name, sc->sc_name) != 0)
3338 			continue;
3339 		if (md.md_mid != sc->sc_id) {
3340 			G_MIRROR_DEBUG(0, "Device %s already configured.",
3341 			    sc->sc_name);
3342 			return (NULL);
3343 		}
3344 		break;
3345 	}
3346 	if (gp == NULL) {
3347 		gp = g_mirror_create(mp, &md, G_MIRROR_TYPE_AUTOMATIC);
3348 		if (gp == NULL) {
3349 			G_MIRROR_DEBUG(0, "Cannot create device %s.",
3350 			    md.md_name);
3351 			return (NULL);
3352 		}
3353 		sc = gp->softc;
3354 	}
3355 	G_MIRROR_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
3356 	g_topology_unlock();
3357 	sx_xlock(&sc->sc_lock);
3358 	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_TASTING;
3359 	error = g_mirror_add_disk(sc, pp, &md);
3360 	sc->sc_flags &= ~G_MIRROR_DEVICE_FLAG_TASTING;
3361 	if (error != 0) {
3362 		G_MIRROR_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
3363 		    pp->name, gp->name, error);
3364 		if (LIST_EMPTY(&sc->sc_disks)) {
3365 			g_cancel_event(sc);
3366 			g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3367 			g_topology_lock();
3368 			return (NULL);
3369 		}
3370 		gp = NULL;
3371 	}
3372 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
3373 		g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3374 		g_topology_lock();
3375 		return (NULL);
3376 	}
3377 	sx_xunlock(&sc->sc_lock);
3378 	g_topology_lock();
3379 	return (gp);
3380 }
3381 
3382 static void
3383 g_mirror_resize(struct g_consumer *cp)
3384 {
3385 	struct g_mirror_disk *disk;
3386 
3387 	g_topology_assert();
3388 	g_trace(G_T_TOPOLOGY, "%s(%s)", __func__, cp->provider->name);
3389 
3390 	disk = cp->private;
3391 	if (disk == NULL)
3392 		return;
3393 	g_topology_unlock();
3394 	g_mirror_update_metadata(disk);
3395 	g_topology_lock();
3396 }
3397 
3398 static int
3399 g_mirror_destroy_geom(struct gctl_req *req __unused,
3400     struct g_class *mp __unused, struct g_geom *gp)
3401 {
3402 	struct g_mirror_softc *sc;
3403 	int error;
3404 
3405 	g_topology_unlock();
3406 	sc = gp->softc;
3407 	sx_xlock(&sc->sc_lock);
3408 	g_cancel_event(sc);
3409 	error = g_mirror_destroy(gp->softc, G_MIRROR_DESTROY_SOFT);
3410 	if (error != 0)
3411 		sx_xunlock(&sc->sc_lock);
3412 	g_topology_lock();
3413 	return (error);
3414 }
3415 
3416 static void
3417 g_mirror_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
3418     struct g_consumer *cp, struct g_provider *pp)
3419 {
3420 	struct g_mirror_softc *sc;
3421 
3422 	g_topology_assert();
3423 
3424 	sc = gp->softc;
3425 	if (sc == NULL)
3426 		return;
3427 	/* Skip synchronization geom. */
3428 	if (gp == sc->sc_sync.ds_geom)
3429 		return;
3430 	if (pp != NULL) {
3431 		/* Nothing here. */
3432 	} else if (cp != NULL) {
3433 		struct g_mirror_disk *disk;
3434 
3435 		disk = cp->private;
3436 		if (disk == NULL)
3437 			return;
3438 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)disk->d_id);
3439 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
3440 			sbuf_printf(sb, "%s<Synchronized>", indent);
3441 			if (disk->d_sync.ds_offset == 0)
3442 				sbuf_cat(sb, "0%");
3443 			else
3444 				sbuf_printf(sb, "%u%%",
3445 				    (u_int)((disk->d_sync.ds_offset * 100) /
3446 				    sc->sc_mediasize));
3447 			sbuf_cat(sb, "</Synchronized>\n");
3448 			if (disk->d_sync.ds_offset > 0)
3449 				sbuf_printf(sb, "%s<BytesSynced>%jd"
3450 				    "</BytesSynced>\n", indent,
3451 				    (intmax_t)disk->d_sync.ds_offset);
3452 		}
3453 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent,
3454 		    disk->d_sync.ds_syncid);
3455 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent,
3456 		    disk->d_genid);
3457 		sbuf_printf(sb, "%s<Flags>", indent);
3458 		if (disk->d_flags == 0)
3459 			sbuf_cat(sb, "NONE");
3460 		else {
3461 			int first = 1;
3462 
3463 #define	ADD_FLAG(flag, name)	do {					\
3464 	if ((disk->d_flags & (flag)) != 0) {				\
3465 		if (!first)						\
3466 			sbuf_cat(sb, ", ");				\
3467 		else							\
3468 			first = 0;					\
3469 		sbuf_cat(sb, name);					\
3470 	}								\
3471 } while (0)
3472 			ADD_FLAG(G_MIRROR_DISK_FLAG_DIRTY, "DIRTY");
3473 			ADD_FLAG(G_MIRROR_DISK_FLAG_HARDCODED, "HARDCODED");
3474 			ADD_FLAG(G_MIRROR_DISK_FLAG_INACTIVE, "INACTIVE");
3475 			ADD_FLAG(G_MIRROR_DISK_FLAG_SYNCHRONIZING,
3476 			    "SYNCHRONIZING");
3477 			ADD_FLAG(G_MIRROR_DISK_FLAG_FORCE_SYNC, "FORCE_SYNC");
3478 			ADD_FLAG(G_MIRROR_DISK_FLAG_BROKEN, "BROKEN");
3479 #undef	ADD_FLAG
3480 		}
3481 		sbuf_cat(sb, "</Flags>\n");
3482 		sbuf_printf(sb, "%s<Priority>%u</Priority>\n", indent,
3483 		    disk->d_priority);
3484 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3485 		    g_mirror_disk_state2str(disk->d_state));
3486 	} else {
3487 		sbuf_printf(sb, "%s<Type>", indent);
3488 		switch (sc->sc_type) {
3489 		case G_MIRROR_TYPE_AUTOMATIC:
3490 			sbuf_cat(sb, "AUTOMATIC");
3491 			break;
3492 		case G_MIRROR_TYPE_MANUAL:
3493 			sbuf_cat(sb, "MANUAL");
3494 			break;
3495 		default:
3496 			sbuf_cat(sb, "UNKNOWN");
3497 			break;
3498 		}
3499 		sbuf_cat(sb, "</Type>\n");
3500 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
3501 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent, sc->sc_syncid);
3502 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, sc->sc_genid);
3503 		sbuf_printf(sb, "%s<Flags>", indent);
3504 		if (sc->sc_flags == 0)
3505 			sbuf_cat(sb, "NONE");
3506 		else {
3507 			int first = 1;
3508 
3509 #define	ADD_FLAG(flag, name)	do {					\
3510 	if ((sc->sc_flags & (flag)) != 0) {				\
3511 		if (!first)						\
3512 			sbuf_cat(sb, ", ");				\
3513 		else							\
3514 			first = 0;					\
3515 		sbuf_cat(sb, name);					\
3516 	}								\
3517 } while (0)
3518 			ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOFAILSYNC, "NOFAILSYNC");
3519 			ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOAUTOSYNC, "NOAUTOSYNC");
3520 #undef	ADD_FLAG
3521 		}
3522 		sbuf_cat(sb, "</Flags>\n");
3523 		sbuf_printf(sb, "%s<Slice>%u</Slice>\n", indent,
3524 		    (u_int)sc->sc_slice);
3525 		sbuf_printf(sb, "%s<Balance>%s</Balance>\n", indent,
3526 		    balance_name(sc->sc_balance));
3527 		sbuf_printf(sb, "%s<Components>%u</Components>\n", indent,
3528 		    sc->sc_ndisks);
3529 		sbuf_printf(sb, "%s<State>", indent);
3530 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
3531 			sbuf_printf(sb, "%s", "STARTING");
3532 		else if (sc->sc_ndisks ==
3533 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE))
3534 			sbuf_printf(sb, "%s", "COMPLETE");
3535 		else
3536 			sbuf_printf(sb, "%s", "DEGRADED");
3537 		sbuf_cat(sb, "</State>\n");
3538 	}
3539 }
3540 
3541 static void
3542 g_mirror_shutdown_post_sync(void *arg, int howto)
3543 {
3544 	struct g_class *mp;
3545 	struct g_geom *gp, *gp2;
3546 	struct g_mirror_softc *sc;
3547 	int error;
3548 
3549 	if ((howto & RB_NOSYNC) != 0)
3550 		return;
3551 
3552 	mp = arg;
3553 	g_topology_lock();
3554 	g_mirror_shutdown = 1;
3555 	LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
3556 		if ((sc = gp->softc) == NULL)
3557 			continue;
3558 		/* Skip synchronization geom. */
3559 		if (gp == sc->sc_sync.ds_geom)
3560 			continue;
3561 		g_topology_unlock();
3562 		sx_xlock(&sc->sc_lock);
3563 		g_mirror_idle(sc, -1);
3564 		g_cancel_event(sc);
3565 		error = g_mirror_destroy(sc, G_MIRROR_DESTROY_DELAYED);
3566 		if (error != 0)
3567 			sx_xunlock(&sc->sc_lock);
3568 		g_topology_lock();
3569 	}
3570 	g_topology_unlock();
3571 }
3572 
3573 static void
3574 g_mirror_init(struct g_class *mp)
3575 {
3576 
3577 	g_mirror_post_sync = EVENTHANDLER_REGISTER(shutdown_post_sync,
3578 	    g_mirror_shutdown_post_sync, mp, SHUTDOWN_PRI_FIRST);
3579 	if (g_mirror_post_sync == NULL)
3580 		G_MIRROR_DEBUG(0, "Warning! Cannot register shutdown event.");
3581 }
3582 
3583 static void
3584 g_mirror_fini(struct g_class *mp)
3585 {
3586 
3587 	if (g_mirror_post_sync != NULL)
3588 		EVENTHANDLER_DEREGISTER(shutdown_post_sync, g_mirror_post_sync);
3589 }
3590 
3591 /*
3592  * Refresh the mirror device's metadata when gmirror encounters a newer
3593  * generation as the individual components are being added to the mirror set.
3594  */
3595 static int
3596 g_mirror_refresh_device(struct g_mirror_softc *sc, const struct g_provider *pp,
3597     const struct g_mirror_metadata *md)
3598 {
3599 
3600 	g_topology_assert_not();
3601 	sx_assert(&sc->sc_lock, SX_XLOCKED);
3602 
3603 	KASSERT(sc->sc_genid <= md->md_genid,
3604 	    ("%s: attempted to refresh from stale component %s (device %s) "
3605 	    "(%u < %u).", __func__, pp->name, sc->sc_name, md->md_genid,
3606 	    sc->sc_genid));
3607 
3608 	if (sc->sc_genid > md->md_genid || (sc->sc_genid == md->md_genid &&
3609 	    sc->sc_syncid >= md->md_syncid))
3610 		return (0);
3611 
3612 	G_MIRROR_DEBUG(0, "Found newer version for device %s (genid: curr=%u "
3613 	    "new=%u; syncid: curr=%u new=%u; ndisks: curr=%u new=%u; "
3614 	    "provider=%s).", sc->sc_name, sc->sc_genid, md->md_genid,
3615 	    sc->sc_syncid, md->md_syncid, sc->sc_ndisks, md->md_all, pp->name);
3616 
3617 	if (sc->sc_state != G_MIRROR_DEVICE_STATE_STARTING) {
3618 		/* Probable data corruption detected */
3619 		G_MIRROR_DEBUG(0, "Cannot refresh metadata in %s state "
3620 		    "(device=%s genid=%u). A stale mirror device was launched.",
3621 		    g_mirror_device_state2str(sc->sc_state), sc->sc_name,
3622 		    sc->sc_genid);
3623 		return (EINVAL);
3624 	}
3625 
3626 	/* Update softc */
3627 	g_mirror_reinit_from_metadata(sc, md);
3628 
3629 	G_MIRROR_DEBUG(1, "Refresh device %s (id=%u, state=%s) from disk %s "
3630 	    "(genid=%u syncid=%u md_all=%u).", sc->sc_name, md->md_mid,
3631 	    g_mirror_device_state2str(sc->sc_state), pp->name, md->md_genid,
3632 	    md->md_syncid, (unsigned)md->md_all);
3633 
3634 	return (0);
3635 }
3636 
3637 DECLARE_GEOM_CLASS(g_mirror_class, g_mirror);
3638 MODULE_VERSION(geom_mirror, 0);
3639