xref: /freebsd/sys/geom/mirror/g_mirror.c (revision bb15ca603fa442c72dde3f3cb8b46db6970e3950)
1 /*-
2  * Copyright (c) 2004-2006 Pawel Jakub Dawidek <pjd@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/kernel.h>
33 #include <sys/module.h>
34 #include <sys/limits.h>
35 #include <sys/lock.h>
36 #include <sys/mutex.h>
37 #include <sys/bio.h>
38 #include <sys/sbuf.h>
39 #include <sys/sysctl.h>
40 #include <sys/malloc.h>
41 #include <sys/eventhandler.h>
42 #include <vm/uma.h>
43 #include <geom/geom.h>
44 #include <sys/proc.h>
45 #include <sys/kthread.h>
46 #include <sys/sched.h>
47 #include <geom/mirror/g_mirror.h>
48 
49 FEATURE(geom_mirror, "GEOM mirroring support");
50 
51 static MALLOC_DEFINE(M_MIRROR, "mirror_data", "GEOM_MIRROR Data");
52 
53 SYSCTL_DECL(_kern_geom);
54 static SYSCTL_NODE(_kern_geom, OID_AUTO, mirror, CTLFLAG_RW, 0,
55     "GEOM_MIRROR stuff");
56 u_int g_mirror_debug = 0;
57 TUNABLE_INT("kern.geom.mirror.debug", &g_mirror_debug);
58 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, debug, CTLFLAG_RW, &g_mirror_debug, 0,
59     "Debug level");
60 static u_int g_mirror_timeout = 4;
61 TUNABLE_INT("kern.geom.mirror.timeout", &g_mirror_timeout);
62 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, timeout, CTLFLAG_RW, &g_mirror_timeout,
63     0, "Time to wait on all mirror components");
64 static u_int g_mirror_idletime = 5;
65 TUNABLE_INT("kern.geom.mirror.idletime", &g_mirror_idletime);
66 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, idletime, CTLFLAG_RW,
67     &g_mirror_idletime, 0, "Mark components as clean when idling");
68 static u_int g_mirror_disconnect_on_failure = 1;
69 TUNABLE_INT("kern.geom.mirror.disconnect_on_failure",
70     &g_mirror_disconnect_on_failure);
71 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, disconnect_on_failure, CTLFLAG_RW,
72     &g_mirror_disconnect_on_failure, 0, "Disconnect component on I/O failure.");
73 static u_int g_mirror_syncreqs = 2;
74 TUNABLE_INT("kern.geom.mirror.sync_requests", &g_mirror_syncreqs);
75 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, sync_requests, CTLFLAG_RDTUN,
76     &g_mirror_syncreqs, 0, "Parallel synchronization I/O requests.");
77 
78 #define	MSLEEP(ident, mtx, priority, wmesg, timeout)	do {		\
79 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, (ident));	\
80 	msleep((ident), (mtx), (priority), (wmesg), (timeout));		\
81 	G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, (ident));	\
82 } while (0)
83 
84 static eventhandler_tag g_mirror_pre_sync = NULL;
85 
86 static int g_mirror_destroy_geom(struct gctl_req *req, struct g_class *mp,
87     struct g_geom *gp);
88 static g_taste_t g_mirror_taste;
89 static void g_mirror_init(struct g_class *mp);
90 static void g_mirror_fini(struct g_class *mp);
91 
92 struct g_class g_mirror_class = {
93 	.name = G_MIRROR_CLASS_NAME,
94 	.version = G_VERSION,
95 	.ctlreq = g_mirror_config,
96 	.taste = g_mirror_taste,
97 	.destroy_geom = g_mirror_destroy_geom,
98 	.init = g_mirror_init,
99 	.fini = g_mirror_fini
100 };
101 
102 
103 static void g_mirror_destroy_provider(struct g_mirror_softc *sc);
104 static int g_mirror_update_disk(struct g_mirror_disk *disk, u_int state);
105 static void g_mirror_update_device(struct g_mirror_softc *sc, boolean_t force);
106 static void g_mirror_dumpconf(struct sbuf *sb, const char *indent,
107     struct g_geom *gp, struct g_consumer *cp, struct g_provider *pp);
108 static void g_mirror_sync_stop(struct g_mirror_disk *disk, int type);
109 static void g_mirror_register_request(struct bio *bp);
110 static void g_mirror_sync_release(struct g_mirror_softc *sc);
111 
112 
113 static const char *
114 g_mirror_disk_state2str(int state)
115 {
116 
117 	switch (state) {
118 	case G_MIRROR_DISK_STATE_NONE:
119 		return ("NONE");
120 	case G_MIRROR_DISK_STATE_NEW:
121 		return ("NEW");
122 	case G_MIRROR_DISK_STATE_ACTIVE:
123 		return ("ACTIVE");
124 	case G_MIRROR_DISK_STATE_STALE:
125 		return ("STALE");
126 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
127 		return ("SYNCHRONIZING");
128 	case G_MIRROR_DISK_STATE_DISCONNECTED:
129 		return ("DISCONNECTED");
130 	case G_MIRROR_DISK_STATE_DESTROY:
131 		return ("DESTROY");
132 	default:
133 		return ("INVALID");
134 	}
135 }
136 
137 static const char *
138 g_mirror_device_state2str(int state)
139 {
140 
141 	switch (state) {
142 	case G_MIRROR_DEVICE_STATE_STARTING:
143 		return ("STARTING");
144 	case G_MIRROR_DEVICE_STATE_RUNNING:
145 		return ("RUNNING");
146 	default:
147 		return ("INVALID");
148 	}
149 }
150 
151 static const char *
152 g_mirror_get_diskname(struct g_mirror_disk *disk)
153 {
154 
155 	if (disk->d_consumer == NULL || disk->d_consumer->provider == NULL)
156 		return ("[unknown]");
157 	return (disk->d_name);
158 }
159 
160 /*
161  * --- Events handling functions ---
162  * Events in geom_mirror are used to maintain disks and device status
163  * from one thread to simplify locking.
164  */
165 static void
166 g_mirror_event_free(struct g_mirror_event *ep)
167 {
168 
169 	free(ep, M_MIRROR);
170 }
171 
172 int
173 g_mirror_event_send(void *arg, int state, int flags)
174 {
175 	struct g_mirror_softc *sc;
176 	struct g_mirror_disk *disk;
177 	struct g_mirror_event *ep;
178 	int error;
179 
180 	ep = malloc(sizeof(*ep), M_MIRROR, M_WAITOK);
181 	G_MIRROR_DEBUG(4, "%s: Sending event %p.", __func__, ep);
182 	if ((flags & G_MIRROR_EVENT_DEVICE) != 0) {
183 		disk = NULL;
184 		sc = arg;
185 	} else {
186 		disk = arg;
187 		sc = disk->d_softc;
188 	}
189 	ep->e_disk = disk;
190 	ep->e_state = state;
191 	ep->e_flags = flags;
192 	ep->e_error = 0;
193 	mtx_lock(&sc->sc_events_mtx);
194 	TAILQ_INSERT_TAIL(&sc->sc_events, ep, e_next);
195 	mtx_unlock(&sc->sc_events_mtx);
196 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
197 	mtx_lock(&sc->sc_queue_mtx);
198 	wakeup(sc);
199 	mtx_unlock(&sc->sc_queue_mtx);
200 	if ((flags & G_MIRROR_EVENT_DONTWAIT) != 0)
201 		return (0);
202 	sx_assert(&sc->sc_lock, SX_XLOCKED);
203 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, ep);
204 	sx_xunlock(&sc->sc_lock);
205 	while ((ep->e_flags & G_MIRROR_EVENT_DONE) == 0) {
206 		mtx_lock(&sc->sc_events_mtx);
207 		MSLEEP(ep, &sc->sc_events_mtx, PRIBIO | PDROP, "m:event",
208 		    hz * 5);
209 	}
210 	error = ep->e_error;
211 	g_mirror_event_free(ep);
212 	sx_xlock(&sc->sc_lock);
213 	return (error);
214 }
215 
216 static struct g_mirror_event *
217 g_mirror_event_get(struct g_mirror_softc *sc)
218 {
219 	struct g_mirror_event *ep;
220 
221 	mtx_lock(&sc->sc_events_mtx);
222 	ep = TAILQ_FIRST(&sc->sc_events);
223 	mtx_unlock(&sc->sc_events_mtx);
224 	return (ep);
225 }
226 
227 static void
228 g_mirror_event_remove(struct g_mirror_softc *sc, struct g_mirror_event *ep)
229 {
230 
231 	mtx_lock(&sc->sc_events_mtx);
232 	TAILQ_REMOVE(&sc->sc_events, ep, e_next);
233 	mtx_unlock(&sc->sc_events_mtx);
234 }
235 
236 static void
237 g_mirror_event_cancel(struct g_mirror_disk *disk)
238 {
239 	struct g_mirror_softc *sc;
240 	struct g_mirror_event *ep, *tmpep;
241 
242 	sc = disk->d_softc;
243 	sx_assert(&sc->sc_lock, SX_XLOCKED);
244 
245 	mtx_lock(&sc->sc_events_mtx);
246 	TAILQ_FOREACH_SAFE(ep, &sc->sc_events, e_next, tmpep) {
247 		if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0)
248 			continue;
249 		if (ep->e_disk != disk)
250 			continue;
251 		TAILQ_REMOVE(&sc->sc_events, ep, e_next);
252 		if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
253 			g_mirror_event_free(ep);
254 		else {
255 			ep->e_error = ECANCELED;
256 			wakeup(ep);
257 		}
258 	}
259 	mtx_unlock(&sc->sc_events_mtx);
260 }
261 
262 /*
263  * Return the number of disks in given state.
264  * If state is equal to -1, count all connected disks.
265  */
266 u_int
267 g_mirror_ndisks(struct g_mirror_softc *sc, int state)
268 {
269 	struct g_mirror_disk *disk;
270 	u_int n = 0;
271 
272 	sx_assert(&sc->sc_lock, SX_LOCKED);
273 
274 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
275 		if (state == -1 || disk->d_state == state)
276 			n++;
277 	}
278 	return (n);
279 }
280 
281 /*
282  * Find a disk in mirror by its disk ID.
283  */
284 static struct g_mirror_disk *
285 g_mirror_id2disk(struct g_mirror_softc *sc, uint32_t id)
286 {
287 	struct g_mirror_disk *disk;
288 
289 	sx_assert(&sc->sc_lock, SX_XLOCKED);
290 
291 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
292 		if (disk->d_id == id)
293 			return (disk);
294 	}
295 	return (NULL);
296 }
297 
298 static u_int
299 g_mirror_nrequests(struct g_mirror_softc *sc, struct g_consumer *cp)
300 {
301 	struct bio *bp;
302 	u_int nreqs = 0;
303 
304 	mtx_lock(&sc->sc_queue_mtx);
305 	TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
306 		if (bp->bio_from == cp)
307 			nreqs++;
308 	}
309 	mtx_unlock(&sc->sc_queue_mtx);
310 	return (nreqs);
311 }
312 
313 static int
314 g_mirror_is_busy(struct g_mirror_softc *sc, struct g_consumer *cp)
315 {
316 
317 	if (cp->index > 0) {
318 		G_MIRROR_DEBUG(2,
319 		    "I/O requests for %s exist, can't destroy it now.",
320 		    cp->provider->name);
321 		return (1);
322 	}
323 	if (g_mirror_nrequests(sc, cp) > 0) {
324 		G_MIRROR_DEBUG(2,
325 		    "I/O requests for %s in queue, can't destroy it now.",
326 		    cp->provider->name);
327 		return (1);
328 	}
329 	return (0);
330 }
331 
332 static void
333 g_mirror_destroy_consumer(void *arg, int flags __unused)
334 {
335 	struct g_consumer *cp;
336 
337 	g_topology_assert();
338 
339 	cp = arg;
340 	G_MIRROR_DEBUG(1, "Consumer %s destroyed.", cp->provider->name);
341 	g_detach(cp);
342 	g_destroy_consumer(cp);
343 }
344 
345 static void
346 g_mirror_kill_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
347 {
348 	struct g_provider *pp;
349 	int retaste_wait;
350 
351 	g_topology_assert();
352 
353 	cp->private = NULL;
354 	if (g_mirror_is_busy(sc, cp))
355 		return;
356 	pp = cp->provider;
357 	retaste_wait = 0;
358 	if (cp->acw == 1) {
359 		if ((pp->geom->flags & G_GEOM_WITHER) == 0)
360 			retaste_wait = 1;
361 	}
362 	G_MIRROR_DEBUG(2, "Access %s r%dw%de%d = %d", pp->name, -cp->acr,
363 	    -cp->acw, -cp->ace, 0);
364 	if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
365 		g_access(cp, -cp->acr, -cp->acw, -cp->ace);
366 	if (retaste_wait) {
367 		/*
368 		 * After retaste event was send (inside g_access()), we can send
369 		 * event to detach and destroy consumer.
370 		 * A class, which has consumer to the given provider connected
371 		 * will not receive retaste event for the provider.
372 		 * This is the way how I ignore retaste events when I close
373 		 * consumers opened for write: I detach and destroy consumer
374 		 * after retaste event is sent.
375 		 */
376 		g_post_event(g_mirror_destroy_consumer, cp, M_WAITOK, NULL);
377 		return;
378 	}
379 	G_MIRROR_DEBUG(1, "Consumer %s destroyed.", pp->name);
380 	g_detach(cp);
381 	g_destroy_consumer(cp);
382 }
383 
384 static int
385 g_mirror_connect_disk(struct g_mirror_disk *disk, struct g_provider *pp)
386 {
387 	struct g_consumer *cp;
388 	int error;
389 
390 	g_topology_assert_not();
391 	KASSERT(disk->d_consumer == NULL,
392 	    ("Disk already connected (device %s).", disk->d_softc->sc_name));
393 
394 	g_topology_lock();
395 	cp = g_new_consumer(disk->d_softc->sc_geom);
396 	error = g_attach(cp, pp);
397 	if (error != 0) {
398 		g_destroy_consumer(cp);
399 		g_topology_unlock();
400 		return (error);
401 	}
402 	error = g_access(cp, 1, 1, 1);
403 	if (error != 0) {
404 		g_detach(cp);
405 		g_destroy_consumer(cp);
406 		g_topology_unlock();
407 		G_MIRROR_DEBUG(0, "Cannot open consumer %s (error=%d).",
408 		    pp->name, error);
409 		return (error);
410 	}
411 	g_topology_unlock();
412 	disk->d_consumer = cp;
413 	disk->d_consumer->private = disk;
414 	disk->d_consumer->index = 0;
415 
416 	G_MIRROR_DEBUG(2, "Disk %s connected.", g_mirror_get_diskname(disk));
417 	return (0);
418 }
419 
420 static void
421 g_mirror_disconnect_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
422 {
423 
424 	g_topology_assert();
425 
426 	if (cp == NULL)
427 		return;
428 	if (cp->provider != NULL)
429 		g_mirror_kill_consumer(sc, cp);
430 	else
431 		g_destroy_consumer(cp);
432 }
433 
434 /*
435  * Initialize disk. This means allocate memory, create consumer, attach it
436  * to the provider and open access (r1w1e1) to it.
437  */
438 static struct g_mirror_disk *
439 g_mirror_init_disk(struct g_mirror_softc *sc, struct g_provider *pp,
440     struct g_mirror_metadata *md, int *errorp)
441 {
442 	struct g_mirror_disk *disk;
443 	int error;
444 
445 	disk = malloc(sizeof(*disk), M_MIRROR, M_NOWAIT | M_ZERO);
446 	if (disk == NULL) {
447 		error = ENOMEM;
448 		goto fail;
449 	}
450 	disk->d_softc = sc;
451 	error = g_mirror_connect_disk(disk, pp);
452 	if (error != 0)
453 		goto fail;
454 	disk->d_id = md->md_did;
455 	disk->d_state = G_MIRROR_DISK_STATE_NONE;
456 	disk->d_priority = md->md_priority;
457 	disk->d_flags = md->md_dflags;
458 	if (md->md_provider[0] != '\0')
459 		disk->d_flags |= G_MIRROR_DISK_FLAG_HARDCODED;
460 	disk->d_sync.ds_consumer = NULL;
461 	disk->d_sync.ds_offset = md->md_sync_offset;
462 	disk->d_sync.ds_offset_done = md->md_sync_offset;
463 	disk->d_genid = md->md_genid;
464 	disk->d_sync.ds_syncid = md->md_syncid;
465 	if (errorp != NULL)
466 		*errorp = 0;
467 	return (disk);
468 fail:
469 	if (errorp != NULL)
470 		*errorp = error;
471 	if (disk != NULL)
472 		free(disk, M_MIRROR);
473 	return (NULL);
474 }
475 
476 static void
477 g_mirror_destroy_disk(struct g_mirror_disk *disk)
478 {
479 	struct g_mirror_softc *sc;
480 
481 	g_topology_assert_not();
482 	sc = disk->d_softc;
483 	sx_assert(&sc->sc_lock, SX_XLOCKED);
484 
485 	LIST_REMOVE(disk, d_next);
486 	g_mirror_event_cancel(disk);
487 	if (sc->sc_hint == disk)
488 		sc->sc_hint = NULL;
489 	switch (disk->d_state) {
490 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
491 		g_mirror_sync_stop(disk, 1);
492 		/* FALLTHROUGH */
493 	case G_MIRROR_DISK_STATE_NEW:
494 	case G_MIRROR_DISK_STATE_STALE:
495 	case G_MIRROR_DISK_STATE_ACTIVE:
496 		g_topology_lock();
497 		g_mirror_disconnect_consumer(sc, disk->d_consumer);
498 		g_topology_unlock();
499 		free(disk, M_MIRROR);
500 		break;
501 	default:
502 		KASSERT(0 == 1, ("Wrong disk state (%s, %s).",
503 		    g_mirror_get_diskname(disk),
504 		    g_mirror_disk_state2str(disk->d_state)));
505 	}
506 }
507 
508 static void
509 g_mirror_destroy_device(struct g_mirror_softc *sc)
510 {
511 	struct g_mirror_disk *disk;
512 	struct g_mirror_event *ep;
513 	struct g_geom *gp;
514 	struct g_consumer *cp, *tmpcp;
515 
516 	g_topology_assert_not();
517 	sx_assert(&sc->sc_lock, SX_XLOCKED);
518 
519 	gp = sc->sc_geom;
520 	if (sc->sc_provider != NULL)
521 		g_mirror_destroy_provider(sc);
522 	for (disk = LIST_FIRST(&sc->sc_disks); disk != NULL;
523 	    disk = LIST_FIRST(&sc->sc_disks)) {
524 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
525 		g_mirror_update_metadata(disk);
526 		g_mirror_destroy_disk(disk);
527 	}
528 	while ((ep = g_mirror_event_get(sc)) != NULL) {
529 		g_mirror_event_remove(sc, ep);
530 		if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
531 			g_mirror_event_free(ep);
532 		else {
533 			ep->e_error = ECANCELED;
534 			ep->e_flags |= G_MIRROR_EVENT_DONE;
535 			G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, ep);
536 			mtx_lock(&sc->sc_events_mtx);
537 			wakeup(ep);
538 			mtx_unlock(&sc->sc_events_mtx);
539 		}
540 	}
541 	callout_drain(&sc->sc_callout);
542 
543 	g_topology_lock();
544 	LIST_FOREACH_SAFE(cp, &sc->sc_sync.ds_geom->consumer, consumer, tmpcp) {
545 		g_mirror_disconnect_consumer(sc, cp);
546 	}
547 	g_wither_geom(sc->sc_sync.ds_geom, ENXIO);
548 	G_MIRROR_DEBUG(0, "Device %s destroyed.", gp->name);
549 	g_wither_geom(gp, ENXIO);
550 	g_topology_unlock();
551 	mtx_destroy(&sc->sc_queue_mtx);
552 	mtx_destroy(&sc->sc_events_mtx);
553 	sx_xunlock(&sc->sc_lock);
554 	sx_destroy(&sc->sc_lock);
555 }
556 
557 static void
558 g_mirror_orphan(struct g_consumer *cp)
559 {
560 	struct g_mirror_disk *disk;
561 
562 	g_topology_assert();
563 
564 	disk = cp->private;
565 	if (disk == NULL)
566 		return;
567 	disk->d_softc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
568 	g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
569 	    G_MIRROR_EVENT_DONTWAIT);
570 }
571 
572 /*
573  * Function should return the next active disk on the list.
574  * It is possible that it will be the same disk as given.
575  * If there are no active disks on list, NULL is returned.
576  */
577 static __inline struct g_mirror_disk *
578 g_mirror_find_next(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
579 {
580 	struct g_mirror_disk *dp;
581 
582 	for (dp = LIST_NEXT(disk, d_next); dp != disk;
583 	    dp = LIST_NEXT(dp, d_next)) {
584 		if (dp == NULL)
585 			dp = LIST_FIRST(&sc->sc_disks);
586 		if (dp->d_state == G_MIRROR_DISK_STATE_ACTIVE)
587 			break;
588 	}
589 	if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
590 		return (NULL);
591 	return (dp);
592 }
593 
594 static struct g_mirror_disk *
595 g_mirror_get_disk(struct g_mirror_softc *sc)
596 {
597 	struct g_mirror_disk *disk;
598 
599 	if (sc->sc_hint == NULL) {
600 		sc->sc_hint = LIST_FIRST(&sc->sc_disks);
601 		if (sc->sc_hint == NULL)
602 			return (NULL);
603 	}
604 	disk = sc->sc_hint;
605 	if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE) {
606 		disk = g_mirror_find_next(sc, disk);
607 		if (disk == NULL)
608 			return (NULL);
609 	}
610 	sc->sc_hint = g_mirror_find_next(sc, disk);
611 	return (disk);
612 }
613 
614 static int
615 g_mirror_write_metadata(struct g_mirror_disk *disk,
616     struct g_mirror_metadata *md)
617 {
618 	struct g_mirror_softc *sc;
619 	struct g_consumer *cp;
620 	off_t offset, length;
621 	u_char *sector;
622 	int error = 0;
623 
624 	g_topology_assert_not();
625 	sc = disk->d_softc;
626 	sx_assert(&sc->sc_lock, SX_LOCKED);
627 
628 	cp = disk->d_consumer;
629 	KASSERT(cp != NULL, ("NULL consumer (%s).", sc->sc_name));
630 	KASSERT(cp->provider != NULL, ("NULL provider (%s).", sc->sc_name));
631 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
632 	    ("Consumer %s closed? (r%dw%de%d).", cp->provider->name, cp->acr,
633 	    cp->acw, cp->ace));
634 	length = cp->provider->sectorsize;
635 	offset = cp->provider->mediasize - length;
636 	sector = malloc((size_t)length, M_MIRROR, M_WAITOK | M_ZERO);
637 	if (md != NULL)
638 		mirror_metadata_encode(md, sector);
639 	error = g_write_data(cp, offset, sector, length);
640 	free(sector, M_MIRROR);
641 	if (error != 0) {
642 		if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
643 			disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
644 			G_MIRROR_DEBUG(0, "Cannot write metadata on %s "
645 			    "(device=%s, error=%d).",
646 			    g_mirror_get_diskname(disk), sc->sc_name, error);
647 		} else {
648 			G_MIRROR_DEBUG(1, "Cannot write metadata on %s "
649 			    "(device=%s, error=%d).",
650 			    g_mirror_get_diskname(disk), sc->sc_name, error);
651 		}
652 		if (g_mirror_disconnect_on_failure &&
653 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1) {
654 			sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
655 			g_mirror_event_send(disk,
656 			    G_MIRROR_DISK_STATE_DISCONNECTED,
657 			    G_MIRROR_EVENT_DONTWAIT);
658 		}
659 	}
660 	return (error);
661 }
662 
663 static int
664 g_mirror_clear_metadata(struct g_mirror_disk *disk)
665 {
666 	int error;
667 
668 	g_topology_assert_not();
669 	sx_assert(&disk->d_softc->sc_lock, SX_LOCKED);
670 
671 	error = g_mirror_write_metadata(disk, NULL);
672 	if (error == 0) {
673 		G_MIRROR_DEBUG(2, "Metadata on %s cleared.",
674 		    g_mirror_get_diskname(disk));
675 	} else {
676 		G_MIRROR_DEBUG(0,
677 		    "Cannot clear metadata on disk %s (error=%d).",
678 		    g_mirror_get_diskname(disk), error);
679 	}
680 	return (error);
681 }
682 
683 void
684 g_mirror_fill_metadata(struct g_mirror_softc *sc, struct g_mirror_disk *disk,
685     struct g_mirror_metadata *md)
686 {
687 
688 	strlcpy(md->md_magic, G_MIRROR_MAGIC, sizeof(md->md_magic));
689 	md->md_version = G_MIRROR_VERSION;
690 	strlcpy(md->md_name, sc->sc_name, sizeof(md->md_name));
691 	md->md_mid = sc->sc_id;
692 	md->md_all = sc->sc_ndisks;
693 	md->md_slice = sc->sc_slice;
694 	md->md_balance = sc->sc_balance;
695 	md->md_genid = sc->sc_genid;
696 	md->md_mediasize = sc->sc_mediasize;
697 	md->md_sectorsize = sc->sc_sectorsize;
698 	md->md_mflags = (sc->sc_flags & G_MIRROR_DEVICE_FLAG_MASK);
699 	bzero(md->md_provider, sizeof(md->md_provider));
700 	if (disk == NULL) {
701 		md->md_did = arc4random();
702 		md->md_priority = 0;
703 		md->md_syncid = 0;
704 		md->md_dflags = 0;
705 		md->md_sync_offset = 0;
706 		md->md_provsize = 0;
707 	} else {
708 		md->md_did = disk->d_id;
709 		md->md_priority = disk->d_priority;
710 		md->md_syncid = disk->d_sync.ds_syncid;
711 		md->md_dflags = (disk->d_flags & G_MIRROR_DISK_FLAG_MASK);
712 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
713 			md->md_sync_offset = disk->d_sync.ds_offset_done;
714 		else
715 			md->md_sync_offset = 0;
716 		if ((disk->d_flags & G_MIRROR_DISK_FLAG_HARDCODED) != 0) {
717 			strlcpy(md->md_provider,
718 			    disk->d_consumer->provider->name,
719 			    sizeof(md->md_provider));
720 		}
721 		md->md_provsize = disk->d_consumer->provider->mediasize;
722 	}
723 }
724 
725 void
726 g_mirror_update_metadata(struct g_mirror_disk *disk)
727 {
728 	struct g_mirror_softc *sc;
729 	struct g_mirror_metadata md;
730 	int error;
731 
732 	g_topology_assert_not();
733 	sc = disk->d_softc;
734 	sx_assert(&sc->sc_lock, SX_LOCKED);
735 
736 	g_mirror_fill_metadata(sc, disk, &md);
737 	error = g_mirror_write_metadata(disk, &md);
738 	if (error == 0) {
739 		G_MIRROR_DEBUG(2, "Metadata on %s updated.",
740 		    g_mirror_get_diskname(disk));
741 	} else {
742 		G_MIRROR_DEBUG(0,
743 		    "Cannot update metadata on disk %s (error=%d).",
744 		    g_mirror_get_diskname(disk), error);
745 	}
746 }
747 
748 static void
749 g_mirror_bump_syncid(struct g_mirror_softc *sc)
750 {
751 	struct g_mirror_disk *disk;
752 
753 	g_topology_assert_not();
754 	sx_assert(&sc->sc_lock, SX_XLOCKED);
755 	KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
756 	    ("%s called with no active disks (device=%s).", __func__,
757 	    sc->sc_name));
758 
759 	sc->sc_syncid++;
760 	G_MIRROR_DEBUG(1, "Device %s: syncid bumped to %u.", sc->sc_name,
761 	    sc->sc_syncid);
762 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
763 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
764 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
765 			disk->d_sync.ds_syncid = sc->sc_syncid;
766 			g_mirror_update_metadata(disk);
767 		}
768 	}
769 }
770 
771 static void
772 g_mirror_bump_genid(struct g_mirror_softc *sc)
773 {
774 	struct g_mirror_disk *disk;
775 
776 	g_topology_assert_not();
777 	sx_assert(&sc->sc_lock, SX_XLOCKED);
778 	KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
779 	    ("%s called with no active disks (device=%s).", __func__,
780 	    sc->sc_name));
781 
782 	sc->sc_genid++;
783 	G_MIRROR_DEBUG(1, "Device %s: genid bumped to %u.", sc->sc_name,
784 	    sc->sc_genid);
785 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
786 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
787 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
788 			disk->d_genid = sc->sc_genid;
789 			g_mirror_update_metadata(disk);
790 		}
791 	}
792 }
793 
794 static int
795 g_mirror_idle(struct g_mirror_softc *sc, int acw)
796 {
797 	struct g_mirror_disk *disk;
798 	int timeout;
799 
800 	g_topology_assert_not();
801 	sx_assert(&sc->sc_lock, SX_XLOCKED);
802 
803 	if (sc->sc_provider == NULL)
804 		return (0);
805 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
806 		return (0);
807 	if (sc->sc_idle)
808 		return (0);
809 	if (sc->sc_writes > 0)
810 		return (0);
811 	if (acw > 0 || (acw == -1 && sc->sc_provider->acw > 0)) {
812 		timeout = g_mirror_idletime - (time_uptime - sc->sc_last_write);
813 		if (timeout > 0)
814 			return (timeout);
815 	}
816 	sc->sc_idle = 1;
817 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
818 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
819 			continue;
820 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as clean.",
821 		    g_mirror_get_diskname(disk), sc->sc_name);
822 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
823 		g_mirror_update_metadata(disk);
824 	}
825 	return (0);
826 }
827 
828 static void
829 g_mirror_unidle(struct g_mirror_softc *sc)
830 {
831 	struct g_mirror_disk *disk;
832 
833 	g_topology_assert_not();
834 	sx_assert(&sc->sc_lock, SX_XLOCKED);
835 
836 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
837 		return;
838 	sc->sc_idle = 0;
839 	sc->sc_last_write = time_uptime;
840 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
841 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
842 			continue;
843 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as dirty.",
844 		    g_mirror_get_diskname(disk), sc->sc_name);
845 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
846 		g_mirror_update_metadata(disk);
847 	}
848 }
849 
850 static void
851 g_mirror_done(struct bio *bp)
852 {
853 	struct g_mirror_softc *sc;
854 
855 	sc = bp->bio_from->geom->softc;
856 	bp->bio_cflags = G_MIRROR_BIO_FLAG_REGULAR;
857 	mtx_lock(&sc->sc_queue_mtx);
858 	bioq_disksort(&sc->sc_queue, bp);
859 	mtx_unlock(&sc->sc_queue_mtx);
860 	wakeup(sc);
861 }
862 
863 static void
864 g_mirror_regular_request(struct bio *bp)
865 {
866 	struct g_mirror_softc *sc;
867 	struct g_mirror_disk *disk;
868 	struct bio *pbp;
869 
870 	g_topology_assert_not();
871 
872 	pbp = bp->bio_parent;
873 	sc = pbp->bio_to->geom->softc;
874 	bp->bio_from->index--;
875 	if (bp->bio_cmd == BIO_WRITE)
876 		sc->sc_writes--;
877 	disk = bp->bio_from->private;
878 	if (disk == NULL) {
879 		g_topology_lock();
880 		g_mirror_kill_consumer(sc, bp->bio_from);
881 		g_topology_unlock();
882 	}
883 
884 	pbp->bio_inbed++;
885 	KASSERT(pbp->bio_inbed <= pbp->bio_children,
886 	    ("bio_inbed (%u) is bigger than bio_children (%u).", pbp->bio_inbed,
887 	    pbp->bio_children));
888 	if (bp->bio_error == 0 && pbp->bio_error == 0) {
889 		G_MIRROR_LOGREQ(3, bp, "Request delivered.");
890 		g_destroy_bio(bp);
891 		if (pbp->bio_children == pbp->bio_inbed) {
892 			G_MIRROR_LOGREQ(3, pbp, "Request delivered.");
893 			pbp->bio_completed = pbp->bio_length;
894 			if (pbp->bio_cmd == BIO_WRITE) {
895 				bioq_remove(&sc->sc_inflight, pbp);
896 				/* Release delayed sync requests if possible. */
897 				g_mirror_sync_release(sc);
898 			}
899 			g_io_deliver(pbp, pbp->bio_error);
900 		}
901 		return;
902 	} else if (bp->bio_error != 0) {
903 		if (pbp->bio_error == 0)
904 			pbp->bio_error = bp->bio_error;
905 		if (disk != NULL) {
906 			if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
907 				disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
908 				G_MIRROR_LOGREQ(0, bp,
909 				    "Request failed (error=%d).",
910 				    bp->bio_error);
911 			} else {
912 				G_MIRROR_LOGREQ(1, bp,
913 				    "Request failed (error=%d).",
914 				    bp->bio_error);
915 			}
916 			if (g_mirror_disconnect_on_failure &&
917 			    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1)
918 			{
919 				sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
920 				g_mirror_event_send(disk,
921 				    G_MIRROR_DISK_STATE_DISCONNECTED,
922 				    G_MIRROR_EVENT_DONTWAIT);
923 			}
924 		}
925 		switch (pbp->bio_cmd) {
926 		case BIO_DELETE:
927 		case BIO_WRITE:
928 			pbp->bio_inbed--;
929 			pbp->bio_children--;
930 			break;
931 		}
932 	}
933 	g_destroy_bio(bp);
934 
935 	switch (pbp->bio_cmd) {
936 	case BIO_READ:
937 		if (pbp->bio_inbed < pbp->bio_children)
938 			break;
939 		if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 1)
940 			g_io_deliver(pbp, pbp->bio_error);
941 		else {
942 			pbp->bio_error = 0;
943 			mtx_lock(&sc->sc_queue_mtx);
944 			bioq_disksort(&sc->sc_queue, pbp);
945 			mtx_unlock(&sc->sc_queue_mtx);
946 			G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
947 			wakeup(sc);
948 		}
949 		break;
950 	case BIO_DELETE:
951 	case BIO_WRITE:
952 		if (pbp->bio_children == 0) {
953 			/*
954 			 * All requests failed.
955 			 */
956 		} else if (pbp->bio_inbed < pbp->bio_children) {
957 			/* Do nothing. */
958 			break;
959 		} else if (pbp->bio_children == pbp->bio_inbed) {
960 			/* Some requests succeeded. */
961 			pbp->bio_error = 0;
962 			pbp->bio_completed = pbp->bio_length;
963 		}
964 		bioq_remove(&sc->sc_inflight, pbp);
965 		/* Release delayed sync requests if possible. */
966 		g_mirror_sync_release(sc);
967 		g_io_deliver(pbp, pbp->bio_error);
968 		break;
969 	default:
970 		KASSERT(1 == 0, ("Invalid request: %u.", pbp->bio_cmd));
971 		break;
972 	}
973 }
974 
975 static void
976 g_mirror_sync_done(struct bio *bp)
977 {
978 	struct g_mirror_softc *sc;
979 
980 	G_MIRROR_LOGREQ(3, bp, "Synchronization request delivered.");
981 	sc = bp->bio_from->geom->softc;
982 	bp->bio_cflags = G_MIRROR_BIO_FLAG_SYNC;
983 	mtx_lock(&sc->sc_queue_mtx);
984 	bioq_disksort(&sc->sc_queue, bp);
985 	mtx_unlock(&sc->sc_queue_mtx);
986 	wakeup(sc);
987 }
988 
989 static void
990 g_mirror_kernel_dump(struct bio *bp)
991 {
992 	struct g_mirror_softc *sc;
993 	struct g_mirror_disk *disk;
994 	struct bio *cbp;
995 	struct g_kerneldump *gkd;
996 
997 	/*
998 	 * We configure dumping to the first component, because this component
999 	 * will be used for reading with 'prefer' balance algorithm.
1000 	 * If the component with the higest priority is currently disconnected
1001 	 * we will not be able to read the dump after the reboot if it will be
1002 	 * connected and synchronized later. Can we do something better?
1003 	 */
1004 	sc = bp->bio_to->geom->softc;
1005 	disk = LIST_FIRST(&sc->sc_disks);
1006 
1007 	gkd = (struct g_kerneldump *)bp->bio_data;
1008 	if (gkd->length > bp->bio_to->mediasize)
1009 		gkd->length = bp->bio_to->mediasize;
1010 	cbp = g_clone_bio(bp);
1011 	if (cbp == NULL) {
1012 		g_io_deliver(bp, ENOMEM);
1013 		return;
1014 	}
1015 	cbp->bio_done = g_std_done;
1016 	g_io_request(cbp, disk->d_consumer);
1017 	G_MIRROR_DEBUG(1, "Kernel dump will go to %s.",
1018 	    g_mirror_get_diskname(disk));
1019 }
1020 
1021 static void
1022 g_mirror_flush(struct g_mirror_softc *sc, struct bio *bp)
1023 {
1024 	struct bio_queue_head queue;
1025 	struct g_mirror_disk *disk;
1026 	struct g_consumer *cp;
1027 	struct bio *cbp;
1028 
1029 	bioq_init(&queue);
1030 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1031 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1032 			continue;
1033 		cbp = g_clone_bio(bp);
1034 		if (cbp == NULL) {
1035 			for (cbp = bioq_first(&queue); cbp != NULL;
1036 			    cbp = bioq_first(&queue)) {
1037 				bioq_remove(&queue, cbp);
1038 				g_destroy_bio(cbp);
1039 			}
1040 			if (bp->bio_error == 0)
1041 				bp->bio_error = ENOMEM;
1042 			g_io_deliver(bp, bp->bio_error);
1043 			return;
1044 		}
1045 		bioq_insert_tail(&queue, cbp);
1046 		cbp->bio_done = g_std_done;
1047 		cbp->bio_caller1 = disk;
1048 		cbp->bio_to = disk->d_consumer->provider;
1049 	}
1050 	for (cbp = bioq_first(&queue); cbp != NULL; cbp = bioq_first(&queue)) {
1051 		bioq_remove(&queue, cbp);
1052 		G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1053 		disk = cbp->bio_caller1;
1054 		cbp->bio_caller1 = NULL;
1055 		cp = disk->d_consumer;
1056 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1057 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1058 		    cp->acr, cp->acw, cp->ace));
1059 		g_io_request(cbp, disk->d_consumer);
1060 	}
1061 }
1062 
1063 static void
1064 g_mirror_start(struct bio *bp)
1065 {
1066 	struct g_mirror_softc *sc;
1067 
1068 	sc = bp->bio_to->geom->softc;
1069 	/*
1070 	 * If sc == NULL or there are no valid disks, provider's error
1071 	 * should be set and g_mirror_start() should not be called at all.
1072 	 */
1073 	KASSERT(sc != NULL && sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1074 	    ("Provider's error should be set (error=%d)(mirror=%s).",
1075 	    bp->bio_to->error, bp->bio_to->name));
1076 	G_MIRROR_LOGREQ(3, bp, "Request received.");
1077 
1078 	switch (bp->bio_cmd) {
1079 	case BIO_READ:
1080 	case BIO_WRITE:
1081 	case BIO_DELETE:
1082 		break;
1083 	case BIO_FLUSH:
1084 		g_mirror_flush(sc, bp);
1085 		return;
1086 	case BIO_GETATTR:
1087 		if (strcmp("GEOM::kerneldump", bp->bio_attribute) == 0) {
1088 			g_mirror_kernel_dump(bp);
1089 			return;
1090 		}
1091 		/* FALLTHROUGH */
1092 	default:
1093 		g_io_deliver(bp, EOPNOTSUPP);
1094 		return;
1095 	}
1096 	mtx_lock(&sc->sc_queue_mtx);
1097 	bioq_disksort(&sc->sc_queue, bp);
1098 	mtx_unlock(&sc->sc_queue_mtx);
1099 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1100 	wakeup(sc);
1101 }
1102 
1103 /*
1104  * Return TRUE if the given request is colliding with a in-progress
1105  * synchronization request.
1106  */
1107 static int
1108 g_mirror_sync_collision(struct g_mirror_softc *sc, struct bio *bp)
1109 {
1110 	struct g_mirror_disk *disk;
1111 	struct bio *sbp;
1112 	off_t rstart, rend, sstart, send;
1113 	int i;
1114 
1115 	if (sc->sc_sync.ds_ndisks == 0)
1116 		return (0);
1117 	rstart = bp->bio_offset;
1118 	rend = bp->bio_offset + bp->bio_length;
1119 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1120 		if (disk->d_state != G_MIRROR_DISK_STATE_SYNCHRONIZING)
1121 			continue;
1122 		for (i = 0; i < g_mirror_syncreqs; i++) {
1123 			sbp = disk->d_sync.ds_bios[i];
1124 			if (sbp == NULL)
1125 				continue;
1126 			sstart = sbp->bio_offset;
1127 			send = sbp->bio_offset + sbp->bio_length;
1128 			if (rend > sstart && rstart < send)
1129 				return (1);
1130 		}
1131 	}
1132 	return (0);
1133 }
1134 
1135 /*
1136  * Return TRUE if the given sync request is colliding with a in-progress regular
1137  * request.
1138  */
1139 static int
1140 g_mirror_regular_collision(struct g_mirror_softc *sc, struct bio *sbp)
1141 {
1142 	off_t rstart, rend, sstart, send;
1143 	struct bio *bp;
1144 
1145 	if (sc->sc_sync.ds_ndisks == 0)
1146 		return (0);
1147 	sstart = sbp->bio_offset;
1148 	send = sbp->bio_offset + sbp->bio_length;
1149 	TAILQ_FOREACH(bp, &sc->sc_inflight.queue, bio_queue) {
1150 		rstart = bp->bio_offset;
1151 		rend = bp->bio_offset + bp->bio_length;
1152 		if (rend > sstart && rstart < send)
1153 			return (1);
1154 	}
1155 	return (0);
1156 }
1157 
1158 /*
1159  * Puts request onto delayed queue.
1160  */
1161 static void
1162 g_mirror_regular_delay(struct g_mirror_softc *sc, struct bio *bp)
1163 {
1164 
1165 	G_MIRROR_LOGREQ(2, bp, "Delaying request.");
1166 	bioq_insert_head(&sc->sc_regular_delayed, bp);
1167 }
1168 
1169 /*
1170  * Puts synchronization request onto delayed queue.
1171  */
1172 static void
1173 g_mirror_sync_delay(struct g_mirror_softc *sc, struct bio *bp)
1174 {
1175 
1176 	G_MIRROR_LOGREQ(2, bp, "Delaying synchronization request.");
1177 	bioq_insert_tail(&sc->sc_sync_delayed, bp);
1178 }
1179 
1180 /*
1181  * Releases delayed regular requests which don't collide anymore with sync
1182  * requests.
1183  */
1184 static void
1185 g_mirror_regular_release(struct g_mirror_softc *sc)
1186 {
1187 	struct bio *bp, *bp2;
1188 
1189 	TAILQ_FOREACH_SAFE(bp, &sc->sc_regular_delayed.queue, bio_queue, bp2) {
1190 		if (g_mirror_sync_collision(sc, bp))
1191 			continue;
1192 		bioq_remove(&sc->sc_regular_delayed, bp);
1193 		G_MIRROR_LOGREQ(2, bp, "Releasing delayed request (%p).", bp);
1194 		mtx_lock(&sc->sc_queue_mtx);
1195 		bioq_insert_head(&sc->sc_queue, bp);
1196 #if 0
1197 		/*
1198 		 * wakeup() is not needed, because this function is called from
1199 		 * the worker thread.
1200 		 */
1201 		wakeup(&sc->sc_queue);
1202 #endif
1203 		mtx_unlock(&sc->sc_queue_mtx);
1204 	}
1205 }
1206 
1207 /*
1208  * Releases delayed sync requests which don't collide anymore with regular
1209  * requests.
1210  */
1211 static void
1212 g_mirror_sync_release(struct g_mirror_softc *sc)
1213 {
1214 	struct bio *bp, *bp2;
1215 
1216 	TAILQ_FOREACH_SAFE(bp, &sc->sc_sync_delayed.queue, bio_queue, bp2) {
1217 		if (g_mirror_regular_collision(sc, bp))
1218 			continue;
1219 		bioq_remove(&sc->sc_sync_delayed, bp);
1220 		G_MIRROR_LOGREQ(2, bp,
1221 		    "Releasing delayed synchronization request.");
1222 		g_io_request(bp, bp->bio_from);
1223 	}
1224 }
1225 
1226 /*
1227  * Handle synchronization requests.
1228  * Every synchronization request is two-steps process: first, READ request is
1229  * send to active provider and then WRITE request (with read data) to the provider
1230  * beeing synchronized. When WRITE is finished, new synchronization request is
1231  * send.
1232  */
1233 static void
1234 g_mirror_sync_request(struct bio *bp)
1235 {
1236 	struct g_mirror_softc *sc;
1237 	struct g_mirror_disk *disk;
1238 
1239 	bp->bio_from->index--;
1240 	sc = bp->bio_from->geom->softc;
1241 	disk = bp->bio_from->private;
1242 	if (disk == NULL) {
1243 		sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1244 		g_topology_lock();
1245 		g_mirror_kill_consumer(sc, bp->bio_from);
1246 		g_topology_unlock();
1247 		free(bp->bio_data, M_MIRROR);
1248 		g_destroy_bio(bp);
1249 		sx_xlock(&sc->sc_lock);
1250 		return;
1251 	}
1252 
1253 	/*
1254 	 * Synchronization request.
1255 	 */
1256 	switch (bp->bio_cmd) {
1257 	case BIO_READ:
1258 	    {
1259 		struct g_consumer *cp;
1260 
1261 		if (bp->bio_error != 0) {
1262 			G_MIRROR_LOGREQ(0, bp,
1263 			    "Synchronization request failed (error=%d).",
1264 			    bp->bio_error);
1265 			g_destroy_bio(bp);
1266 			return;
1267 		}
1268 		G_MIRROR_LOGREQ(3, bp,
1269 		    "Synchronization request half-finished.");
1270 		bp->bio_cmd = BIO_WRITE;
1271 		bp->bio_cflags = 0;
1272 		cp = disk->d_consumer;
1273 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1274 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1275 		    cp->acr, cp->acw, cp->ace));
1276 		cp->index++;
1277 		g_io_request(bp, cp);
1278 		return;
1279 	    }
1280 	case BIO_WRITE:
1281 	    {
1282 		struct g_mirror_disk_sync *sync;
1283 		off_t offset;
1284 		void *data;
1285 		int i;
1286 
1287 		if (bp->bio_error != 0) {
1288 			G_MIRROR_LOGREQ(0, bp,
1289 			    "Synchronization request failed (error=%d).",
1290 			    bp->bio_error);
1291 			g_destroy_bio(bp);
1292 			sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
1293 			g_mirror_event_send(disk,
1294 			    G_MIRROR_DISK_STATE_DISCONNECTED,
1295 			    G_MIRROR_EVENT_DONTWAIT);
1296 			return;
1297 		}
1298 		G_MIRROR_LOGREQ(3, bp, "Synchronization request finished.");
1299 		sync = &disk->d_sync;
1300 		if (sync->ds_offset == sc->sc_mediasize ||
1301 		    sync->ds_consumer == NULL ||
1302 		    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1303 			/* Don't send more synchronization requests. */
1304 			sync->ds_inflight--;
1305 			if (sync->ds_bios != NULL) {
1306 				i = (int)(uintptr_t)bp->bio_caller1;
1307 				sync->ds_bios[i] = NULL;
1308 			}
1309 			free(bp->bio_data, M_MIRROR);
1310 			g_destroy_bio(bp);
1311 			if (sync->ds_inflight > 0)
1312 				return;
1313 			if (sync->ds_consumer == NULL ||
1314 			    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1315 				return;
1316 			}
1317 			/* Disk up-to-date, activate it. */
1318 			g_mirror_event_send(disk, G_MIRROR_DISK_STATE_ACTIVE,
1319 			    G_MIRROR_EVENT_DONTWAIT);
1320 			return;
1321 		}
1322 
1323 		/* Send next synchronization request. */
1324 		data = bp->bio_data;
1325 		bzero(bp, sizeof(*bp));
1326 		bp->bio_cmd = BIO_READ;
1327 		bp->bio_offset = sync->ds_offset;
1328 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1329 		sync->ds_offset += bp->bio_length;
1330 		bp->bio_done = g_mirror_sync_done;
1331 		bp->bio_data = data;
1332 		bp->bio_from = sync->ds_consumer;
1333 		bp->bio_to = sc->sc_provider;
1334 		G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
1335 		sync->ds_consumer->index++;
1336 		/*
1337 		 * Delay the request if it is colliding with a regular request.
1338 		 */
1339 		if (g_mirror_regular_collision(sc, bp))
1340 			g_mirror_sync_delay(sc, bp);
1341 		else
1342 			g_io_request(bp, sync->ds_consumer);
1343 
1344 		/* Release delayed requests if possible. */
1345 		g_mirror_regular_release(sc);
1346 
1347 		/* Find the smallest offset */
1348 		offset = sc->sc_mediasize;
1349 		for (i = 0; i < g_mirror_syncreqs; i++) {
1350 			bp = sync->ds_bios[i];
1351 			if (bp->bio_offset < offset)
1352 				offset = bp->bio_offset;
1353 		}
1354 		if (sync->ds_offset_done + (MAXPHYS * 100) < offset) {
1355 			/* Update offset_done on every 100 blocks. */
1356 			sync->ds_offset_done = offset;
1357 			g_mirror_update_metadata(disk);
1358 		}
1359 		return;
1360 	    }
1361 	default:
1362 		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1363 		    bp->bio_cmd, sc->sc_name));
1364 		break;
1365 	}
1366 }
1367 
1368 static void
1369 g_mirror_request_prefer(struct g_mirror_softc *sc, struct bio *bp)
1370 {
1371 	struct g_mirror_disk *disk;
1372 	struct g_consumer *cp;
1373 	struct bio *cbp;
1374 
1375 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1376 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE)
1377 			break;
1378 	}
1379 	if (disk == NULL) {
1380 		if (bp->bio_error == 0)
1381 			bp->bio_error = ENXIO;
1382 		g_io_deliver(bp, bp->bio_error);
1383 		return;
1384 	}
1385 	cbp = g_clone_bio(bp);
1386 	if (cbp == NULL) {
1387 		if (bp->bio_error == 0)
1388 			bp->bio_error = ENOMEM;
1389 		g_io_deliver(bp, bp->bio_error);
1390 		return;
1391 	}
1392 	/*
1393 	 * Fill in the component buf structure.
1394 	 */
1395 	cp = disk->d_consumer;
1396 	cbp->bio_done = g_mirror_done;
1397 	cbp->bio_to = cp->provider;
1398 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1399 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1400 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1401 	    cp->acw, cp->ace));
1402 	cp->index++;
1403 	g_io_request(cbp, cp);
1404 }
1405 
1406 static void
1407 g_mirror_request_round_robin(struct g_mirror_softc *sc, struct bio *bp)
1408 {
1409 	struct g_mirror_disk *disk;
1410 	struct g_consumer *cp;
1411 	struct bio *cbp;
1412 
1413 	disk = g_mirror_get_disk(sc);
1414 	if (disk == NULL) {
1415 		if (bp->bio_error == 0)
1416 			bp->bio_error = ENXIO;
1417 		g_io_deliver(bp, bp->bio_error);
1418 		return;
1419 	}
1420 	cbp = g_clone_bio(bp);
1421 	if (cbp == NULL) {
1422 		if (bp->bio_error == 0)
1423 			bp->bio_error = ENOMEM;
1424 		g_io_deliver(bp, bp->bio_error);
1425 		return;
1426 	}
1427 	/*
1428 	 * Fill in the component buf structure.
1429 	 */
1430 	cp = disk->d_consumer;
1431 	cbp->bio_done = g_mirror_done;
1432 	cbp->bio_to = cp->provider;
1433 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1434 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1435 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1436 	    cp->acw, cp->ace));
1437 	cp->index++;
1438 	g_io_request(cbp, cp);
1439 }
1440 
1441 #define TRACK_SIZE  (1 * 1024 * 1024)
1442 #define LOAD_SCALE	256
1443 #define ABS(x)		(((x) >= 0) ? (x) : (-(x)))
1444 
1445 static void
1446 g_mirror_request_load(struct g_mirror_softc *sc, struct bio *bp)
1447 {
1448 	struct g_mirror_disk *disk, *dp;
1449 	struct g_consumer *cp;
1450 	struct bio *cbp;
1451 	int prio, best;
1452 
1453 	/* Find a disk with the smallest load. */
1454 	disk = NULL;
1455 	best = INT_MAX;
1456 	LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1457 		if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1458 			continue;
1459 		prio = dp->load;
1460 		/* If disk head is precisely in position - highly prefer it. */
1461 		if (dp->d_last_offset == bp->bio_offset)
1462 			prio -= 2 * LOAD_SCALE;
1463 		else
1464 		/* If disk head is close to position - prefer it. */
1465 		if (ABS(dp->d_last_offset - bp->bio_offset) < TRACK_SIZE)
1466 			prio -= 1 * LOAD_SCALE;
1467 		if (prio <= best) {
1468 			disk = dp;
1469 			best = prio;
1470 		}
1471 	}
1472 	KASSERT(disk != NULL, ("NULL disk for %s.", sc->sc_name));
1473 	cbp = g_clone_bio(bp);
1474 	if (cbp == NULL) {
1475 		if (bp->bio_error == 0)
1476 			bp->bio_error = ENOMEM;
1477 		g_io_deliver(bp, bp->bio_error);
1478 		return;
1479 	}
1480 	/*
1481 	 * Fill in the component buf structure.
1482 	 */
1483 	cp = disk->d_consumer;
1484 	cbp->bio_done = g_mirror_done;
1485 	cbp->bio_to = cp->provider;
1486 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1487 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1488 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1489 	    cp->acw, cp->ace));
1490 	cp->index++;
1491 	/* Remember last head position */
1492 	disk->d_last_offset = bp->bio_offset + bp->bio_length;
1493 	/* Update loads. */
1494 	LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1495 		dp->load = (dp->d_consumer->index * LOAD_SCALE +
1496 		    dp->load * 7) / 8;
1497 	}
1498 	g_io_request(cbp, cp);
1499 }
1500 
1501 static void
1502 g_mirror_request_split(struct g_mirror_softc *sc, struct bio *bp)
1503 {
1504 	struct bio_queue_head queue;
1505 	struct g_mirror_disk *disk;
1506 	struct g_consumer *cp;
1507 	struct bio *cbp;
1508 	off_t left, mod, offset, slice;
1509 	u_char *data;
1510 	u_int ndisks;
1511 
1512 	if (bp->bio_length <= sc->sc_slice) {
1513 		g_mirror_request_round_robin(sc, bp);
1514 		return;
1515 	}
1516 	ndisks = g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE);
1517 	slice = bp->bio_length / ndisks;
1518 	mod = slice % sc->sc_provider->sectorsize;
1519 	if (mod != 0)
1520 		slice += sc->sc_provider->sectorsize - mod;
1521 	/*
1522 	 * Allocate all bios before sending any request, so we can
1523 	 * return ENOMEM in nice and clean way.
1524 	 */
1525 	left = bp->bio_length;
1526 	offset = bp->bio_offset;
1527 	data = bp->bio_data;
1528 	bioq_init(&queue);
1529 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1530 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1531 			continue;
1532 		cbp = g_clone_bio(bp);
1533 		if (cbp == NULL) {
1534 			for (cbp = bioq_first(&queue); cbp != NULL;
1535 			    cbp = bioq_first(&queue)) {
1536 				bioq_remove(&queue, cbp);
1537 				g_destroy_bio(cbp);
1538 			}
1539 			if (bp->bio_error == 0)
1540 				bp->bio_error = ENOMEM;
1541 			g_io_deliver(bp, bp->bio_error);
1542 			return;
1543 		}
1544 		bioq_insert_tail(&queue, cbp);
1545 		cbp->bio_done = g_mirror_done;
1546 		cbp->bio_caller1 = disk;
1547 		cbp->bio_to = disk->d_consumer->provider;
1548 		cbp->bio_offset = offset;
1549 		cbp->bio_data = data;
1550 		cbp->bio_length = MIN(left, slice);
1551 		left -= cbp->bio_length;
1552 		if (left == 0)
1553 			break;
1554 		offset += cbp->bio_length;
1555 		data += cbp->bio_length;
1556 	}
1557 	for (cbp = bioq_first(&queue); cbp != NULL; cbp = bioq_first(&queue)) {
1558 		bioq_remove(&queue, cbp);
1559 		G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1560 		disk = cbp->bio_caller1;
1561 		cbp->bio_caller1 = NULL;
1562 		cp = disk->d_consumer;
1563 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1564 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1565 		    cp->acr, cp->acw, cp->ace));
1566 		disk->d_consumer->index++;
1567 		g_io_request(cbp, disk->d_consumer);
1568 	}
1569 }
1570 
1571 static void
1572 g_mirror_register_request(struct bio *bp)
1573 {
1574 	struct g_mirror_softc *sc;
1575 
1576 	sc = bp->bio_to->geom->softc;
1577 	switch (bp->bio_cmd) {
1578 	case BIO_READ:
1579 		switch (sc->sc_balance) {
1580 		case G_MIRROR_BALANCE_LOAD:
1581 			g_mirror_request_load(sc, bp);
1582 			break;
1583 		case G_MIRROR_BALANCE_PREFER:
1584 			g_mirror_request_prefer(sc, bp);
1585 			break;
1586 		case G_MIRROR_BALANCE_ROUND_ROBIN:
1587 			g_mirror_request_round_robin(sc, bp);
1588 			break;
1589 		case G_MIRROR_BALANCE_SPLIT:
1590 			g_mirror_request_split(sc, bp);
1591 			break;
1592 		}
1593 		return;
1594 	case BIO_WRITE:
1595 	case BIO_DELETE:
1596 	    {
1597 		struct g_mirror_disk *disk;
1598 		struct g_mirror_disk_sync *sync;
1599 		struct bio_queue_head queue;
1600 		struct g_consumer *cp;
1601 		struct bio *cbp;
1602 
1603 		/*
1604 		 * Delay the request if it is colliding with a synchronization
1605 		 * request.
1606 		 */
1607 		if (g_mirror_sync_collision(sc, bp)) {
1608 			g_mirror_regular_delay(sc, bp);
1609 			return;
1610 		}
1611 
1612 		if (sc->sc_idle)
1613 			g_mirror_unidle(sc);
1614 		else
1615 			sc->sc_last_write = time_uptime;
1616 
1617 		/*
1618 		 * Allocate all bios before sending any request, so we can
1619 		 * return ENOMEM in nice and clean way.
1620 		 */
1621 		bioq_init(&queue);
1622 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1623 			sync = &disk->d_sync;
1624 			switch (disk->d_state) {
1625 			case G_MIRROR_DISK_STATE_ACTIVE:
1626 				break;
1627 			case G_MIRROR_DISK_STATE_SYNCHRONIZING:
1628 				if (bp->bio_offset >= sync->ds_offset)
1629 					continue;
1630 				break;
1631 			default:
1632 				continue;
1633 			}
1634 			cbp = g_clone_bio(bp);
1635 			if (cbp == NULL) {
1636 				for (cbp = bioq_first(&queue); cbp != NULL;
1637 				    cbp = bioq_first(&queue)) {
1638 					bioq_remove(&queue, cbp);
1639 					g_destroy_bio(cbp);
1640 				}
1641 				if (bp->bio_error == 0)
1642 					bp->bio_error = ENOMEM;
1643 				g_io_deliver(bp, bp->bio_error);
1644 				return;
1645 			}
1646 			bioq_insert_tail(&queue, cbp);
1647 			cbp->bio_done = g_mirror_done;
1648 			cp = disk->d_consumer;
1649 			cbp->bio_caller1 = cp;
1650 			cbp->bio_to = cp->provider;
1651 			KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1652 			    ("Consumer %s not opened (r%dw%de%d).",
1653 			    cp->provider->name, cp->acr, cp->acw, cp->ace));
1654 		}
1655 		for (cbp = bioq_first(&queue); cbp != NULL;
1656 		    cbp = bioq_first(&queue)) {
1657 			bioq_remove(&queue, cbp);
1658 			G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1659 			cp = cbp->bio_caller1;
1660 			cbp->bio_caller1 = NULL;
1661 			cp->index++;
1662 			sc->sc_writes++;
1663 			g_io_request(cbp, cp);
1664 		}
1665 		/*
1666 		 * Put request onto inflight queue, so we can check if new
1667 		 * synchronization requests don't collide with it.
1668 		 */
1669 		bioq_insert_tail(&sc->sc_inflight, bp);
1670 		/*
1671 		 * Bump syncid on first write.
1672 		 */
1673 		if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0) {
1674 			sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
1675 			g_mirror_bump_syncid(sc);
1676 		}
1677 		return;
1678 	    }
1679 	default:
1680 		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1681 		    bp->bio_cmd, sc->sc_name));
1682 		break;
1683 	}
1684 }
1685 
1686 static int
1687 g_mirror_can_destroy(struct g_mirror_softc *sc)
1688 {
1689 	struct g_geom *gp;
1690 	struct g_consumer *cp;
1691 
1692 	g_topology_assert();
1693 	gp = sc->sc_geom;
1694 	if (gp->softc == NULL)
1695 		return (1);
1696 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1697 		if (g_mirror_is_busy(sc, cp))
1698 			return (0);
1699 	}
1700 	gp = sc->sc_sync.ds_geom;
1701 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1702 		if (g_mirror_is_busy(sc, cp))
1703 			return (0);
1704 	}
1705 	G_MIRROR_DEBUG(2, "No I/O requests for %s, it can be destroyed.",
1706 	    sc->sc_name);
1707 	return (1);
1708 }
1709 
1710 static int
1711 g_mirror_try_destroy(struct g_mirror_softc *sc)
1712 {
1713 
1714 	if (sc->sc_rootmount != NULL) {
1715 		G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
1716 		    sc->sc_rootmount);
1717 		root_mount_rel(sc->sc_rootmount);
1718 		sc->sc_rootmount = NULL;
1719 	}
1720 	g_topology_lock();
1721 	if (!g_mirror_can_destroy(sc)) {
1722 		g_topology_unlock();
1723 		return (0);
1724 	}
1725 	sc->sc_geom->softc = NULL;
1726 	sc->sc_sync.ds_geom->softc = NULL;
1727 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_WAIT) != 0) {
1728 		g_topology_unlock();
1729 		G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1730 		    &sc->sc_worker);
1731 		/* Unlock sc_lock here, as it can be destroyed after wakeup. */
1732 		sx_xunlock(&sc->sc_lock);
1733 		wakeup(&sc->sc_worker);
1734 		sc->sc_worker = NULL;
1735 	} else {
1736 		g_topology_unlock();
1737 		g_mirror_destroy_device(sc);
1738 		free(sc, M_MIRROR);
1739 	}
1740 	return (1);
1741 }
1742 
1743 /*
1744  * Worker thread.
1745  */
1746 static void
1747 g_mirror_worker(void *arg)
1748 {
1749 	struct g_mirror_softc *sc;
1750 	struct g_mirror_event *ep;
1751 	struct bio *bp;
1752 	int timeout;
1753 
1754 	sc = arg;
1755 	thread_lock(curthread);
1756 	sched_prio(curthread, PRIBIO);
1757 	thread_unlock(curthread);
1758 
1759 	sx_xlock(&sc->sc_lock);
1760 	for (;;) {
1761 		G_MIRROR_DEBUG(5, "%s: Let's see...", __func__);
1762 		/*
1763 		 * First take a look at events.
1764 		 * This is important to handle events before any I/O requests.
1765 		 */
1766 		ep = g_mirror_event_get(sc);
1767 		if (ep != NULL) {
1768 			g_mirror_event_remove(sc, ep);
1769 			if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0) {
1770 				/* Update only device status. */
1771 				G_MIRROR_DEBUG(3,
1772 				    "Running event for device %s.",
1773 				    sc->sc_name);
1774 				ep->e_error = 0;
1775 				g_mirror_update_device(sc, 1);
1776 			} else {
1777 				/* Update disk status. */
1778 				G_MIRROR_DEBUG(3, "Running event for disk %s.",
1779 				     g_mirror_get_diskname(ep->e_disk));
1780 				ep->e_error = g_mirror_update_disk(ep->e_disk,
1781 				    ep->e_state);
1782 				if (ep->e_error == 0)
1783 					g_mirror_update_device(sc, 0);
1784 			}
1785 			if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0) {
1786 				KASSERT(ep->e_error == 0,
1787 				    ("Error cannot be handled."));
1788 				g_mirror_event_free(ep);
1789 			} else {
1790 				ep->e_flags |= G_MIRROR_EVENT_DONE;
1791 				G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1792 				    ep);
1793 				mtx_lock(&sc->sc_events_mtx);
1794 				wakeup(ep);
1795 				mtx_unlock(&sc->sc_events_mtx);
1796 			}
1797 			if ((sc->sc_flags &
1798 			    G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1799 				if (g_mirror_try_destroy(sc)) {
1800 					curthread->td_pflags &= ~TDP_GEOM;
1801 					G_MIRROR_DEBUG(1, "Thread exiting.");
1802 					kproc_exit(0);
1803 				}
1804 			}
1805 			G_MIRROR_DEBUG(5, "%s: I'm here 1.", __func__);
1806 			continue;
1807 		}
1808 		/*
1809 		 * Check if we can mark array as CLEAN and if we can't take
1810 		 * how much seconds should we wait.
1811 		 */
1812 		timeout = g_mirror_idle(sc, -1);
1813 		/*
1814 		 * Now I/O requests.
1815 		 */
1816 		/* Get first request from the queue. */
1817 		mtx_lock(&sc->sc_queue_mtx);
1818 		bp = bioq_first(&sc->sc_queue);
1819 		if (bp == NULL) {
1820 			if ((sc->sc_flags &
1821 			    G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1822 				mtx_unlock(&sc->sc_queue_mtx);
1823 				if (g_mirror_try_destroy(sc)) {
1824 					curthread->td_pflags &= ~TDP_GEOM;
1825 					G_MIRROR_DEBUG(1, "Thread exiting.");
1826 					kproc_exit(0);
1827 				}
1828 				mtx_lock(&sc->sc_queue_mtx);
1829 			}
1830 			sx_xunlock(&sc->sc_lock);
1831 			/*
1832 			 * XXX: We can miss an event here, because an event
1833 			 *      can be added without sx-device-lock and without
1834 			 *      mtx-queue-lock. Maybe I should just stop using
1835 			 *      dedicated mutex for events synchronization and
1836 			 *      stick with the queue lock?
1837 			 *      The event will hang here until next I/O request
1838 			 *      or next event is received.
1839 			 */
1840 			MSLEEP(sc, &sc->sc_queue_mtx, PRIBIO | PDROP, "m:w1",
1841 			    timeout * hz);
1842 			sx_xlock(&sc->sc_lock);
1843 			G_MIRROR_DEBUG(5, "%s: I'm here 4.", __func__);
1844 			continue;
1845 		}
1846 		bioq_remove(&sc->sc_queue, bp);
1847 		mtx_unlock(&sc->sc_queue_mtx);
1848 
1849 		if (bp->bio_from->geom == sc->sc_sync.ds_geom &&
1850 		    (bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0) {
1851 			g_mirror_sync_request(bp);	/* READ */
1852 		} else if (bp->bio_to != sc->sc_provider) {
1853 			if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_REGULAR) != 0)
1854 				g_mirror_regular_request(bp);
1855 			else if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0)
1856 				g_mirror_sync_request(bp);	/* WRITE */
1857 			else {
1858 				KASSERT(0,
1859 				    ("Invalid request cflags=0x%hhx to=%s.",
1860 				    bp->bio_cflags, bp->bio_to->name));
1861 			}
1862 		} else {
1863 			g_mirror_register_request(bp);
1864 		}
1865 		G_MIRROR_DEBUG(5, "%s: I'm here 9.", __func__);
1866 	}
1867 }
1868 
1869 static void
1870 g_mirror_update_idle(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
1871 {
1872 
1873 	sx_assert(&sc->sc_lock, SX_LOCKED);
1874 
1875 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
1876 		return;
1877 	if (!sc->sc_idle && (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) == 0) {
1878 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as dirty.",
1879 		    g_mirror_get_diskname(disk), sc->sc_name);
1880 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
1881 	} else if (sc->sc_idle &&
1882 	    (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
1883 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as clean.",
1884 		    g_mirror_get_diskname(disk), sc->sc_name);
1885 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
1886 	}
1887 }
1888 
1889 static void
1890 g_mirror_sync_start(struct g_mirror_disk *disk)
1891 {
1892 	struct g_mirror_softc *sc;
1893 	struct g_consumer *cp;
1894 	struct bio *bp;
1895 	int error, i;
1896 
1897 	g_topology_assert_not();
1898 	sc = disk->d_softc;
1899 	sx_assert(&sc->sc_lock, SX_LOCKED);
1900 
1901 	KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
1902 	    ("Disk %s is not marked for synchronization.",
1903 	    g_mirror_get_diskname(disk)));
1904 	KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1905 	    ("Device not in RUNNING state (%s, %u).", sc->sc_name,
1906 	    sc->sc_state));
1907 
1908 	sx_xunlock(&sc->sc_lock);
1909 	g_topology_lock();
1910 	cp = g_new_consumer(sc->sc_sync.ds_geom);
1911 	error = g_attach(cp, sc->sc_provider);
1912 	KASSERT(error == 0,
1913 	    ("Cannot attach to %s (error=%d).", sc->sc_name, error));
1914 	error = g_access(cp, 1, 0, 0);
1915 	KASSERT(error == 0, ("Cannot open %s (error=%d).", sc->sc_name, error));
1916 	g_topology_unlock();
1917 	sx_xlock(&sc->sc_lock);
1918 
1919 	G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s.", sc->sc_name,
1920 	    g_mirror_get_diskname(disk));
1921 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) == 0)
1922 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
1923 	KASSERT(disk->d_sync.ds_consumer == NULL,
1924 	    ("Sync consumer already exists (device=%s, disk=%s).",
1925 	    sc->sc_name, g_mirror_get_diskname(disk)));
1926 
1927 	disk->d_sync.ds_consumer = cp;
1928 	disk->d_sync.ds_consumer->private = disk;
1929 	disk->d_sync.ds_consumer->index = 0;
1930 
1931 	/*
1932 	 * Allocate memory for synchronization bios and initialize them.
1933 	 */
1934 	disk->d_sync.ds_bios = malloc(sizeof(struct bio *) * g_mirror_syncreqs,
1935 	    M_MIRROR, M_WAITOK);
1936 	for (i = 0; i < g_mirror_syncreqs; i++) {
1937 		bp = g_alloc_bio();
1938 		disk->d_sync.ds_bios[i] = bp;
1939 		bp->bio_parent = NULL;
1940 		bp->bio_cmd = BIO_READ;
1941 		bp->bio_data = malloc(MAXPHYS, M_MIRROR, M_WAITOK);
1942 		bp->bio_cflags = 0;
1943 		bp->bio_offset = disk->d_sync.ds_offset;
1944 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1945 		disk->d_sync.ds_offset += bp->bio_length;
1946 		bp->bio_done = g_mirror_sync_done;
1947 		bp->bio_from = disk->d_sync.ds_consumer;
1948 		bp->bio_to = sc->sc_provider;
1949 		bp->bio_caller1 = (void *)(uintptr_t)i;
1950 	}
1951 
1952 	/* Increase the number of disks in SYNCHRONIZING state. */
1953 	sc->sc_sync.ds_ndisks++;
1954 	/* Set the number of in-flight synchronization requests. */
1955 	disk->d_sync.ds_inflight = g_mirror_syncreqs;
1956 
1957 	/*
1958 	 * Fire off first synchronization requests.
1959 	 */
1960 	for (i = 0; i < g_mirror_syncreqs; i++) {
1961 		bp = disk->d_sync.ds_bios[i];
1962 		G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
1963 		disk->d_sync.ds_consumer->index++;
1964 		/*
1965 		 * Delay the request if it is colliding with a regular request.
1966 		 */
1967 		if (g_mirror_regular_collision(sc, bp))
1968 			g_mirror_sync_delay(sc, bp);
1969 		else
1970 			g_io_request(bp, disk->d_sync.ds_consumer);
1971 	}
1972 }
1973 
1974 /*
1975  * Stop synchronization process.
1976  * type: 0 - synchronization finished
1977  *       1 - synchronization stopped
1978  */
1979 static void
1980 g_mirror_sync_stop(struct g_mirror_disk *disk, int type)
1981 {
1982 	struct g_mirror_softc *sc;
1983 	struct g_consumer *cp;
1984 
1985 	g_topology_assert_not();
1986 	sc = disk->d_softc;
1987 	sx_assert(&sc->sc_lock, SX_LOCKED);
1988 
1989 	KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
1990 	    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
1991 	    g_mirror_disk_state2str(disk->d_state)));
1992 	if (disk->d_sync.ds_consumer == NULL)
1993 		return;
1994 
1995 	if (type == 0) {
1996 		G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s finished.",
1997 		    sc->sc_name, g_mirror_get_diskname(disk));
1998 	} else /* if (type == 1) */ {
1999 		G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s stopped.",
2000 		    sc->sc_name, g_mirror_get_diskname(disk));
2001 	}
2002 	free(disk->d_sync.ds_bios, M_MIRROR);
2003 	disk->d_sync.ds_bios = NULL;
2004 	cp = disk->d_sync.ds_consumer;
2005 	disk->d_sync.ds_consumer = NULL;
2006 	disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2007 	sc->sc_sync.ds_ndisks--;
2008 	sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
2009 	g_topology_lock();
2010 	g_mirror_kill_consumer(sc, cp);
2011 	g_topology_unlock();
2012 	sx_xlock(&sc->sc_lock);
2013 }
2014 
2015 static void
2016 g_mirror_launch_provider(struct g_mirror_softc *sc)
2017 {
2018 	struct g_mirror_disk *disk;
2019 	struct g_provider *pp;
2020 
2021 	sx_assert(&sc->sc_lock, SX_LOCKED);
2022 
2023 	g_topology_lock();
2024 	pp = g_new_providerf(sc->sc_geom, "mirror/%s", sc->sc_name);
2025 	pp->mediasize = sc->sc_mediasize;
2026 	pp->sectorsize = sc->sc_sectorsize;
2027 	pp->stripesize = 0;
2028 	pp->stripeoffset = 0;
2029 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2030 		if (disk->d_consumer && disk->d_consumer->provider &&
2031 		    disk->d_consumer->provider->stripesize > pp->stripesize) {
2032 			pp->stripesize = disk->d_consumer->provider->stripesize;
2033 			pp->stripeoffset = disk->d_consumer->provider->stripeoffset;
2034 		}
2035 	}
2036 	sc->sc_provider = pp;
2037 	g_error_provider(pp, 0);
2038 	g_topology_unlock();
2039 	G_MIRROR_DEBUG(0, "Device %s launched (%u/%u).", pp->name,
2040 	    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE), sc->sc_ndisks);
2041 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2042 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2043 			g_mirror_sync_start(disk);
2044 	}
2045 }
2046 
2047 static void
2048 g_mirror_destroy_provider(struct g_mirror_softc *sc)
2049 {
2050 	struct g_mirror_disk *disk;
2051 	struct bio *bp;
2052 
2053 	g_topology_assert_not();
2054 	KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
2055 	    sc->sc_name));
2056 
2057 	g_topology_lock();
2058 	g_error_provider(sc->sc_provider, ENXIO);
2059 	mtx_lock(&sc->sc_queue_mtx);
2060 	while ((bp = bioq_first(&sc->sc_queue)) != NULL) {
2061 		bioq_remove(&sc->sc_queue, bp);
2062 		g_io_deliver(bp, ENXIO);
2063 	}
2064 	mtx_unlock(&sc->sc_queue_mtx);
2065 	G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.", sc->sc_name,
2066 	    sc->sc_provider->name);
2067 	sc->sc_provider->flags |= G_PF_WITHER;
2068 	g_orphan_provider(sc->sc_provider, ENXIO);
2069 	g_topology_unlock();
2070 	sc->sc_provider = NULL;
2071 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2072 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2073 			g_mirror_sync_stop(disk, 1);
2074 	}
2075 }
2076 
2077 static void
2078 g_mirror_go(void *arg)
2079 {
2080 	struct g_mirror_softc *sc;
2081 
2082 	sc = arg;
2083 	G_MIRROR_DEBUG(0, "Force device %s start due to timeout.", sc->sc_name);
2084 	g_mirror_event_send(sc, 0,
2085 	    G_MIRROR_EVENT_DONTWAIT | G_MIRROR_EVENT_DEVICE);
2086 }
2087 
2088 static u_int
2089 g_mirror_determine_state(struct g_mirror_disk *disk)
2090 {
2091 	struct g_mirror_softc *sc;
2092 	u_int state;
2093 
2094 	sc = disk->d_softc;
2095 	if (sc->sc_syncid == disk->d_sync.ds_syncid) {
2096 		if ((disk->d_flags &
2097 		    G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2098 			/* Disk does not need synchronization. */
2099 			state = G_MIRROR_DISK_STATE_ACTIVE;
2100 		} else {
2101 			if ((sc->sc_flags &
2102 			     G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2103 			    (disk->d_flags &
2104 			     G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2105 				/*
2106 				 * We can start synchronization from
2107 				 * the stored offset.
2108 				 */
2109 				state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2110 			} else {
2111 				state = G_MIRROR_DISK_STATE_STALE;
2112 			}
2113 		}
2114 	} else if (disk->d_sync.ds_syncid < sc->sc_syncid) {
2115 		/*
2116 		 * Reset all synchronization data for this disk,
2117 		 * because if it even was synchronized, it was
2118 		 * synchronized to disks with different syncid.
2119 		 */
2120 		disk->d_flags |= G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2121 		disk->d_sync.ds_offset = 0;
2122 		disk->d_sync.ds_offset_done = 0;
2123 		disk->d_sync.ds_syncid = sc->sc_syncid;
2124 		if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2125 		    (disk->d_flags & G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2126 			state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2127 		} else {
2128 			state = G_MIRROR_DISK_STATE_STALE;
2129 		}
2130 	} else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
2131 		/*
2132 		 * Not good, NOT GOOD!
2133 		 * It means that mirror was started on stale disks
2134 		 * and more fresh disk just arrive.
2135 		 * If there were writes, mirror is broken, sorry.
2136 		 * I think the best choice here is don't touch
2137 		 * this disk and inform the user loudly.
2138 		 */
2139 		G_MIRROR_DEBUG(0, "Device %s was started before the freshest "
2140 		    "disk (%s) arrives!! It will not be connected to the "
2141 		    "running device.", sc->sc_name,
2142 		    g_mirror_get_diskname(disk));
2143 		g_mirror_destroy_disk(disk);
2144 		state = G_MIRROR_DISK_STATE_NONE;
2145 		/* Return immediately, because disk was destroyed. */
2146 		return (state);
2147 	}
2148 	G_MIRROR_DEBUG(3, "State for %s disk: %s.",
2149 	    g_mirror_get_diskname(disk), g_mirror_disk_state2str(state));
2150 	return (state);
2151 }
2152 
2153 /*
2154  * Update device state.
2155  */
2156 static void
2157 g_mirror_update_device(struct g_mirror_softc *sc, boolean_t force)
2158 {
2159 	struct g_mirror_disk *disk;
2160 	u_int state;
2161 
2162 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2163 
2164 	switch (sc->sc_state) {
2165 	case G_MIRROR_DEVICE_STATE_STARTING:
2166 	    {
2167 		struct g_mirror_disk *pdisk, *tdisk;
2168 		u_int dirty, ndisks, genid, syncid;
2169 
2170 		KASSERT(sc->sc_provider == NULL,
2171 		    ("Non-NULL provider in STARTING state (%s).", sc->sc_name));
2172 		/*
2173 		 * Are we ready? We are, if all disks are connected or
2174 		 * if we have any disks and 'force' is true.
2175 		 */
2176 		ndisks = g_mirror_ndisks(sc, -1);
2177 		if (sc->sc_ndisks == ndisks || (force && ndisks > 0)) {
2178 			;
2179 		} else if (ndisks == 0) {
2180 			/*
2181 			 * Disks went down in starting phase, so destroy
2182 			 * device.
2183 			 */
2184 			callout_drain(&sc->sc_callout);
2185 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2186 			G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2187 			    sc->sc_rootmount);
2188 			root_mount_rel(sc->sc_rootmount);
2189 			sc->sc_rootmount = NULL;
2190 			return;
2191 		} else {
2192 			return;
2193 		}
2194 
2195 		/*
2196 		 * Activate all disks with the biggest syncid.
2197 		 */
2198 		if (force) {
2199 			/*
2200 			 * If 'force' is true, we have been called due to
2201 			 * timeout, so don't bother canceling timeout.
2202 			 */
2203 			ndisks = 0;
2204 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2205 				if ((disk->d_flags &
2206 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2207 					ndisks++;
2208 				}
2209 			}
2210 			if (ndisks == 0) {
2211 				/* No valid disks found, destroy device. */
2212 				sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2213 				G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2214 				    __LINE__, sc->sc_rootmount);
2215 				root_mount_rel(sc->sc_rootmount);
2216 				sc->sc_rootmount = NULL;
2217 				return;
2218 			}
2219 		} else {
2220 			/* Cancel timeout. */
2221 			callout_drain(&sc->sc_callout);
2222 		}
2223 
2224 		/*
2225 		 * Find the biggest genid.
2226 		 */
2227 		genid = 0;
2228 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2229 			if (disk->d_genid > genid)
2230 				genid = disk->d_genid;
2231 		}
2232 		sc->sc_genid = genid;
2233 		/*
2234 		 * Remove all disks without the biggest genid.
2235 		 */
2236 		LIST_FOREACH_SAFE(disk, &sc->sc_disks, d_next, tdisk) {
2237 			if (disk->d_genid < genid) {
2238 				G_MIRROR_DEBUG(0,
2239 				    "Component %s (device %s) broken, skipping.",
2240 				    g_mirror_get_diskname(disk), sc->sc_name);
2241 				g_mirror_destroy_disk(disk);
2242 			}
2243 		}
2244 
2245 		/*
2246 		 * Find the biggest syncid.
2247 		 */
2248 		syncid = 0;
2249 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2250 			if (disk->d_sync.ds_syncid > syncid)
2251 				syncid = disk->d_sync.ds_syncid;
2252 		}
2253 
2254 		/*
2255 		 * Here we need to look for dirty disks and if all disks
2256 		 * with the biggest syncid are dirty, we have to choose
2257 		 * one with the biggest priority and rebuild the rest.
2258 		 */
2259 		/*
2260 		 * Find the number of dirty disks with the biggest syncid.
2261 		 * Find the number of disks with the biggest syncid.
2262 		 * While here, find a disk with the biggest priority.
2263 		 */
2264 		dirty = ndisks = 0;
2265 		pdisk = NULL;
2266 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2267 			if (disk->d_sync.ds_syncid != syncid)
2268 				continue;
2269 			if ((disk->d_flags &
2270 			    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2271 				continue;
2272 			}
2273 			ndisks++;
2274 			if ((disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
2275 				dirty++;
2276 				if (pdisk == NULL ||
2277 				    pdisk->d_priority < disk->d_priority) {
2278 					pdisk = disk;
2279 				}
2280 			}
2281 		}
2282 		if (dirty == 0) {
2283 			/* No dirty disks at all, great. */
2284 		} else if (dirty == ndisks) {
2285 			/*
2286 			 * Force synchronization for all dirty disks except one
2287 			 * with the biggest priority.
2288 			 */
2289 			KASSERT(pdisk != NULL, ("pdisk == NULL"));
2290 			G_MIRROR_DEBUG(1, "Using disk %s (device %s) as a "
2291 			    "master disk for synchronization.",
2292 			    g_mirror_get_diskname(pdisk), sc->sc_name);
2293 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2294 				if (disk->d_sync.ds_syncid != syncid)
2295 					continue;
2296 				if ((disk->d_flags &
2297 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2298 					continue;
2299 				}
2300 				KASSERT((disk->d_flags &
2301 				    G_MIRROR_DISK_FLAG_DIRTY) != 0,
2302 				    ("Disk %s isn't marked as dirty.",
2303 				    g_mirror_get_diskname(disk)));
2304 				/* Skip the disk with the biggest priority. */
2305 				if (disk == pdisk)
2306 					continue;
2307 				disk->d_sync.ds_syncid = 0;
2308 			}
2309 		} else if (dirty < ndisks) {
2310 			/*
2311 			 * Force synchronization for all dirty disks.
2312 			 * We have some non-dirty disks.
2313 			 */
2314 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2315 				if (disk->d_sync.ds_syncid != syncid)
2316 					continue;
2317 				if ((disk->d_flags &
2318 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2319 					continue;
2320 				}
2321 				if ((disk->d_flags &
2322 				    G_MIRROR_DISK_FLAG_DIRTY) == 0) {
2323 					continue;
2324 				}
2325 				disk->d_sync.ds_syncid = 0;
2326 			}
2327 		}
2328 
2329 		/* Reset hint. */
2330 		sc->sc_hint = NULL;
2331 		sc->sc_syncid = syncid;
2332 		if (force) {
2333 			/* Remember to bump syncid on first write. */
2334 			sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2335 		}
2336 		state = G_MIRROR_DEVICE_STATE_RUNNING;
2337 		G_MIRROR_DEBUG(1, "Device %s state changed from %s to %s.",
2338 		    sc->sc_name, g_mirror_device_state2str(sc->sc_state),
2339 		    g_mirror_device_state2str(state));
2340 		sc->sc_state = state;
2341 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2342 			state = g_mirror_determine_state(disk);
2343 			g_mirror_event_send(disk, state,
2344 			    G_MIRROR_EVENT_DONTWAIT);
2345 			if (state == G_MIRROR_DISK_STATE_STALE)
2346 				sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2347 		}
2348 		break;
2349 	    }
2350 	case G_MIRROR_DEVICE_STATE_RUNNING:
2351 		if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 0 &&
2352 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2353 			/*
2354 			 * No active disks or no disks at all,
2355 			 * so destroy device.
2356 			 */
2357 			if (sc->sc_provider != NULL)
2358 				g_mirror_destroy_provider(sc);
2359 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2360 			break;
2361 		} else if (g_mirror_ndisks(sc,
2362 		    G_MIRROR_DISK_STATE_ACTIVE) > 0 &&
2363 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2364 			/*
2365 			 * We have active disks, launch provider if it doesn't
2366 			 * exist.
2367 			 */
2368 			if (sc->sc_provider == NULL)
2369 				g_mirror_launch_provider(sc);
2370 			if (sc->sc_rootmount != NULL) {
2371 				G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2372 				    __LINE__, sc->sc_rootmount);
2373 				root_mount_rel(sc->sc_rootmount);
2374 				sc->sc_rootmount = NULL;
2375 			}
2376 		}
2377 		/*
2378 		 * Genid should be bumped immediately, so do it here.
2379 		 */
2380 		if ((sc->sc_bump_id & G_MIRROR_BUMP_GENID) != 0) {
2381 			sc->sc_bump_id &= ~G_MIRROR_BUMP_GENID;
2382 			g_mirror_bump_genid(sc);
2383 		}
2384 		break;
2385 	default:
2386 		KASSERT(1 == 0, ("Wrong device state (%s, %s).",
2387 		    sc->sc_name, g_mirror_device_state2str(sc->sc_state)));
2388 		break;
2389 	}
2390 }
2391 
2392 /*
2393  * Update disk state and device state if needed.
2394  */
2395 #define	DISK_STATE_CHANGED()	G_MIRROR_DEBUG(1,			\
2396 	"Disk %s state changed from %s to %s (device %s).",		\
2397 	g_mirror_get_diskname(disk),					\
2398 	g_mirror_disk_state2str(disk->d_state),				\
2399 	g_mirror_disk_state2str(state), sc->sc_name)
2400 static int
2401 g_mirror_update_disk(struct g_mirror_disk *disk, u_int state)
2402 {
2403 	struct g_mirror_softc *sc;
2404 
2405 	sc = disk->d_softc;
2406 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2407 
2408 again:
2409 	G_MIRROR_DEBUG(3, "Changing disk %s state from %s to %s.",
2410 	    g_mirror_get_diskname(disk), g_mirror_disk_state2str(disk->d_state),
2411 	    g_mirror_disk_state2str(state));
2412 	switch (state) {
2413 	case G_MIRROR_DISK_STATE_NEW:
2414 		/*
2415 		 * Possible scenarios:
2416 		 * 1. New disk arrive.
2417 		 */
2418 		/* Previous state should be NONE. */
2419 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NONE,
2420 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2421 		    g_mirror_disk_state2str(disk->d_state)));
2422 		DISK_STATE_CHANGED();
2423 
2424 		disk->d_state = state;
2425 		if (LIST_EMPTY(&sc->sc_disks))
2426 			LIST_INSERT_HEAD(&sc->sc_disks, disk, d_next);
2427 		else {
2428 			struct g_mirror_disk *dp;
2429 
2430 			LIST_FOREACH(dp, &sc->sc_disks, d_next) {
2431 				if (disk->d_priority >= dp->d_priority) {
2432 					LIST_INSERT_BEFORE(dp, disk, d_next);
2433 					dp = NULL;
2434 					break;
2435 				}
2436 				if (LIST_NEXT(dp, d_next) == NULL)
2437 					break;
2438 			}
2439 			if (dp != NULL)
2440 				LIST_INSERT_AFTER(dp, disk, d_next);
2441 		}
2442 		G_MIRROR_DEBUG(1, "Device %s: provider %s detected.",
2443 		    sc->sc_name, g_mirror_get_diskname(disk));
2444 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
2445 			break;
2446 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2447 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2448 		    g_mirror_device_state2str(sc->sc_state),
2449 		    g_mirror_get_diskname(disk),
2450 		    g_mirror_disk_state2str(disk->d_state)));
2451 		state = g_mirror_determine_state(disk);
2452 		if (state != G_MIRROR_DISK_STATE_NONE)
2453 			goto again;
2454 		break;
2455 	case G_MIRROR_DISK_STATE_ACTIVE:
2456 		/*
2457 		 * Possible scenarios:
2458 		 * 1. New disk does not need synchronization.
2459 		 * 2. Synchronization process finished successfully.
2460 		 */
2461 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2462 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2463 		    g_mirror_device_state2str(sc->sc_state),
2464 		    g_mirror_get_diskname(disk),
2465 		    g_mirror_disk_state2str(disk->d_state)));
2466 		/* Previous state should be NEW or SYNCHRONIZING. */
2467 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW ||
2468 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2469 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2470 		    g_mirror_disk_state2str(disk->d_state)));
2471 		DISK_STATE_CHANGED();
2472 
2473 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
2474 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2475 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_FORCE_SYNC;
2476 			g_mirror_sync_stop(disk, 0);
2477 		}
2478 		disk->d_state = state;
2479 		disk->d_sync.ds_offset = 0;
2480 		disk->d_sync.ds_offset_done = 0;
2481 		g_mirror_update_idle(sc, disk);
2482 		g_mirror_update_metadata(disk);
2483 		G_MIRROR_DEBUG(1, "Device %s: provider %s activated.",
2484 		    sc->sc_name, g_mirror_get_diskname(disk));
2485 		break;
2486 	case G_MIRROR_DISK_STATE_STALE:
2487 		/*
2488 		 * Possible scenarios:
2489 		 * 1. Stale disk was connected.
2490 		 */
2491 		/* Previous state should be NEW. */
2492 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2493 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2494 		    g_mirror_disk_state2str(disk->d_state)));
2495 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2496 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2497 		    g_mirror_device_state2str(sc->sc_state),
2498 		    g_mirror_get_diskname(disk),
2499 		    g_mirror_disk_state2str(disk->d_state)));
2500 		/*
2501 		 * STALE state is only possible if device is marked
2502 		 * NOAUTOSYNC.
2503 		 */
2504 		KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) != 0,
2505 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2506 		    g_mirror_device_state2str(sc->sc_state),
2507 		    g_mirror_get_diskname(disk),
2508 		    g_mirror_disk_state2str(disk->d_state)));
2509 		DISK_STATE_CHANGED();
2510 
2511 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2512 		disk->d_state = state;
2513 		g_mirror_update_metadata(disk);
2514 		G_MIRROR_DEBUG(0, "Device %s: provider %s is stale.",
2515 		    sc->sc_name, g_mirror_get_diskname(disk));
2516 		break;
2517 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
2518 		/*
2519 		 * Possible scenarios:
2520 		 * 1. Disk which needs synchronization was connected.
2521 		 */
2522 		/* Previous state should be NEW. */
2523 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2524 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2525 		    g_mirror_disk_state2str(disk->d_state)));
2526 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2527 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2528 		    g_mirror_device_state2str(sc->sc_state),
2529 		    g_mirror_get_diskname(disk),
2530 		    g_mirror_disk_state2str(disk->d_state)));
2531 		DISK_STATE_CHANGED();
2532 
2533 		if (disk->d_state == G_MIRROR_DISK_STATE_NEW)
2534 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2535 		disk->d_state = state;
2536 		if (sc->sc_provider != NULL) {
2537 			g_mirror_sync_start(disk);
2538 			g_mirror_update_metadata(disk);
2539 		}
2540 		break;
2541 	case G_MIRROR_DISK_STATE_DISCONNECTED:
2542 		/*
2543 		 * Possible scenarios:
2544 		 * 1. Device wasn't running yet, but disk disappear.
2545 		 * 2. Disk was active and disapppear.
2546 		 * 3. Disk disappear during synchronization process.
2547 		 */
2548 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING) {
2549 			/*
2550 			 * Previous state should be ACTIVE, STALE or
2551 			 * SYNCHRONIZING.
2552 			 */
2553 			KASSERT(disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
2554 			    disk->d_state == G_MIRROR_DISK_STATE_STALE ||
2555 			    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2556 			    ("Wrong disk state (%s, %s).",
2557 			    g_mirror_get_diskname(disk),
2558 			    g_mirror_disk_state2str(disk->d_state)));
2559 		} else if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING) {
2560 			/* Previous state should be NEW. */
2561 			KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2562 			    ("Wrong disk state (%s, %s).",
2563 			    g_mirror_get_diskname(disk),
2564 			    g_mirror_disk_state2str(disk->d_state)));
2565 			/*
2566 			 * Reset bumping syncid if disk disappeared in STARTING
2567 			 * state.
2568 			 */
2569 			if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0)
2570 				sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
2571 #ifdef	INVARIANTS
2572 		} else {
2573 			KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
2574 			    sc->sc_name,
2575 			    g_mirror_device_state2str(sc->sc_state),
2576 			    g_mirror_get_diskname(disk),
2577 			    g_mirror_disk_state2str(disk->d_state)));
2578 #endif
2579 		}
2580 		DISK_STATE_CHANGED();
2581 		G_MIRROR_DEBUG(0, "Device %s: provider %s disconnected.",
2582 		    sc->sc_name, g_mirror_get_diskname(disk));
2583 
2584 		g_mirror_destroy_disk(disk);
2585 		break;
2586 	case G_MIRROR_DISK_STATE_DESTROY:
2587 	    {
2588 		int error;
2589 
2590 		error = g_mirror_clear_metadata(disk);
2591 		if (error != 0)
2592 			return (error);
2593 		DISK_STATE_CHANGED();
2594 		G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.",
2595 		    sc->sc_name, g_mirror_get_diskname(disk));
2596 
2597 		g_mirror_destroy_disk(disk);
2598 		sc->sc_ndisks--;
2599 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2600 			g_mirror_update_metadata(disk);
2601 		}
2602 		break;
2603 	    }
2604 	default:
2605 		KASSERT(1 == 0, ("Unknown state (%u).", state));
2606 		break;
2607 	}
2608 	return (0);
2609 }
2610 #undef	DISK_STATE_CHANGED
2611 
2612 int
2613 g_mirror_read_metadata(struct g_consumer *cp, struct g_mirror_metadata *md)
2614 {
2615 	struct g_provider *pp;
2616 	u_char *buf;
2617 	int error;
2618 
2619 	g_topology_assert();
2620 
2621 	error = g_access(cp, 1, 0, 0);
2622 	if (error != 0)
2623 		return (error);
2624 	pp = cp->provider;
2625 	g_topology_unlock();
2626 	/* Metadata are stored on last sector. */
2627 	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
2628 	    &error);
2629 	g_topology_lock();
2630 	g_access(cp, -1, 0, 0);
2631 	if (buf == NULL) {
2632 		G_MIRROR_DEBUG(1, "Cannot read metadata from %s (error=%d).",
2633 		    cp->provider->name, error);
2634 		return (error);
2635 	}
2636 
2637 	/* Decode metadata. */
2638 	error = mirror_metadata_decode(buf, md);
2639 	g_free(buf);
2640 	if (strcmp(md->md_magic, G_MIRROR_MAGIC) != 0)
2641 		return (EINVAL);
2642 	if (md->md_version > G_MIRROR_VERSION) {
2643 		G_MIRROR_DEBUG(0,
2644 		    "Kernel module is too old to handle metadata from %s.",
2645 		    cp->provider->name);
2646 		return (EINVAL);
2647 	}
2648 	if (error != 0) {
2649 		G_MIRROR_DEBUG(1, "MD5 metadata hash mismatch for provider %s.",
2650 		    cp->provider->name);
2651 		return (error);
2652 	}
2653 
2654 	return (0);
2655 }
2656 
2657 static int
2658 g_mirror_check_metadata(struct g_mirror_softc *sc, struct g_provider *pp,
2659     struct g_mirror_metadata *md)
2660 {
2661 
2662 	if (g_mirror_id2disk(sc, md->md_did) != NULL) {
2663 		G_MIRROR_DEBUG(1, "Disk %s (id=%u) already exists, skipping.",
2664 		    pp->name, md->md_did);
2665 		return (EEXIST);
2666 	}
2667 	if (md->md_all != sc->sc_ndisks) {
2668 		G_MIRROR_DEBUG(1,
2669 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2670 		    "md_all", pp->name, sc->sc_name);
2671 		return (EINVAL);
2672 	}
2673 	if (md->md_slice != sc->sc_slice) {
2674 		G_MIRROR_DEBUG(1,
2675 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2676 		    "md_slice", pp->name, sc->sc_name);
2677 		return (EINVAL);
2678 	}
2679 	if (md->md_balance != sc->sc_balance) {
2680 		G_MIRROR_DEBUG(1,
2681 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2682 		    "md_balance", pp->name, sc->sc_name);
2683 		return (EINVAL);
2684 	}
2685 	if (md->md_mediasize != sc->sc_mediasize) {
2686 		G_MIRROR_DEBUG(1,
2687 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2688 		    "md_mediasize", pp->name, sc->sc_name);
2689 		return (EINVAL);
2690 	}
2691 	if (sc->sc_mediasize > pp->mediasize) {
2692 		G_MIRROR_DEBUG(1,
2693 		    "Invalid size of disk %s (device %s), skipping.", pp->name,
2694 		    sc->sc_name);
2695 		return (EINVAL);
2696 	}
2697 	if (md->md_sectorsize != sc->sc_sectorsize) {
2698 		G_MIRROR_DEBUG(1,
2699 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2700 		    "md_sectorsize", pp->name, sc->sc_name);
2701 		return (EINVAL);
2702 	}
2703 	if ((sc->sc_sectorsize % pp->sectorsize) != 0) {
2704 		G_MIRROR_DEBUG(1,
2705 		    "Invalid sector size of disk %s (device %s), skipping.",
2706 		    pp->name, sc->sc_name);
2707 		return (EINVAL);
2708 	}
2709 	if ((md->md_mflags & ~G_MIRROR_DEVICE_FLAG_MASK) != 0) {
2710 		G_MIRROR_DEBUG(1,
2711 		    "Invalid device flags on disk %s (device %s), skipping.",
2712 		    pp->name, sc->sc_name);
2713 		return (EINVAL);
2714 	}
2715 	if ((md->md_dflags & ~G_MIRROR_DISK_FLAG_MASK) != 0) {
2716 		G_MIRROR_DEBUG(1,
2717 		    "Invalid disk flags on disk %s (device %s), skipping.",
2718 		    pp->name, sc->sc_name);
2719 		return (EINVAL);
2720 	}
2721 	return (0);
2722 }
2723 
2724 int
2725 g_mirror_add_disk(struct g_mirror_softc *sc, struct g_provider *pp,
2726     struct g_mirror_metadata *md)
2727 {
2728 	struct g_mirror_disk *disk;
2729 	int error;
2730 
2731 	g_topology_assert_not();
2732 	G_MIRROR_DEBUG(2, "Adding disk %s.", pp->name);
2733 
2734 	error = g_mirror_check_metadata(sc, pp, md);
2735 	if (error != 0)
2736 		return (error);
2737 	if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING &&
2738 	    md->md_genid < sc->sc_genid) {
2739 		G_MIRROR_DEBUG(0, "Component %s (device %s) broken, skipping.",
2740 		    pp->name, sc->sc_name);
2741 		return (EINVAL);
2742 	}
2743 	disk = g_mirror_init_disk(sc, pp, md, &error);
2744 	if (disk == NULL)
2745 		return (error);
2746 	error = g_mirror_event_send(disk, G_MIRROR_DISK_STATE_NEW,
2747 	    G_MIRROR_EVENT_WAIT);
2748 	if (error != 0)
2749 		return (error);
2750 	if (md->md_version < G_MIRROR_VERSION) {
2751 		G_MIRROR_DEBUG(0, "Upgrading metadata on %s (v%d->v%d).",
2752 		    pp->name, md->md_version, G_MIRROR_VERSION);
2753 		g_mirror_update_metadata(disk);
2754 	}
2755 	return (0);
2756 }
2757 
2758 static void
2759 g_mirror_destroy_delayed(void *arg, int flag)
2760 {
2761 	struct g_mirror_softc *sc;
2762 	int error;
2763 
2764 	if (flag == EV_CANCEL) {
2765 		G_MIRROR_DEBUG(1, "Destroying canceled.");
2766 		return;
2767 	}
2768 	sc = arg;
2769 	g_topology_unlock();
2770 	sx_xlock(&sc->sc_lock);
2771 	KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) == 0,
2772 	    ("DESTROY flag set on %s.", sc->sc_name));
2773 	KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROYING) != 0,
2774 	    ("DESTROYING flag not set on %s.", sc->sc_name));
2775 	G_MIRROR_DEBUG(1, "Destroying %s (delayed).", sc->sc_name);
2776 	error = g_mirror_destroy(sc, G_MIRROR_DESTROY_SOFT);
2777 	if (error != 0) {
2778 		G_MIRROR_DEBUG(0, "Cannot destroy %s.", sc->sc_name);
2779 		sx_xunlock(&sc->sc_lock);
2780 	}
2781 	g_topology_lock();
2782 }
2783 
2784 static int
2785 g_mirror_access(struct g_provider *pp, int acr, int acw, int ace)
2786 {
2787 	struct g_mirror_softc *sc;
2788 	int dcr, dcw, dce, error = 0;
2789 
2790 	g_topology_assert();
2791 	G_MIRROR_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, acr,
2792 	    acw, ace);
2793 
2794 	sc = pp->geom->softc;
2795 	if (sc == NULL && acr <= 0 && acw <= 0 && ace <= 0)
2796 		return (0);
2797 	KASSERT(sc != NULL, ("NULL softc (provider=%s).", pp->name));
2798 
2799 	dcr = pp->acr + acr;
2800 	dcw = pp->acw + acw;
2801 	dce = pp->ace + ace;
2802 
2803 	g_topology_unlock();
2804 	sx_xlock(&sc->sc_lock);
2805 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0 ||
2806 	    LIST_EMPTY(&sc->sc_disks)) {
2807 		if (acr > 0 || acw > 0 || ace > 0)
2808 			error = ENXIO;
2809 		goto end;
2810 	}
2811 	if (dcw == 0 && !sc->sc_idle)
2812 		g_mirror_idle(sc, dcw);
2813 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROYING) != 0) {
2814 		if (acr > 0 || acw > 0 || ace > 0) {
2815 			error = ENXIO;
2816 			goto end;
2817 		}
2818 		if (dcr == 0 && dcw == 0 && dce == 0) {
2819 			g_post_event(g_mirror_destroy_delayed, sc, M_WAITOK,
2820 			    sc, NULL);
2821 		}
2822 	}
2823 end:
2824 	sx_xunlock(&sc->sc_lock);
2825 	g_topology_lock();
2826 	return (error);
2827 }
2828 
2829 static struct g_geom *
2830 g_mirror_create(struct g_class *mp, const struct g_mirror_metadata *md)
2831 {
2832 	struct g_mirror_softc *sc;
2833 	struct g_geom *gp;
2834 	int error, timeout;
2835 
2836 	g_topology_assert();
2837 	G_MIRROR_DEBUG(1, "Creating device %s (id=%u).", md->md_name,
2838 	    md->md_mid);
2839 
2840 	/* One disk is minimum. */
2841 	if (md->md_all < 1)
2842 		return (NULL);
2843 	/*
2844 	 * Action geom.
2845 	 */
2846 	gp = g_new_geomf(mp, "%s", md->md_name);
2847 	sc = malloc(sizeof(*sc), M_MIRROR, M_WAITOK | M_ZERO);
2848 	gp->start = g_mirror_start;
2849 	gp->orphan = g_mirror_orphan;
2850 	gp->access = g_mirror_access;
2851 	gp->dumpconf = g_mirror_dumpconf;
2852 
2853 	sc->sc_id = md->md_mid;
2854 	sc->sc_slice = md->md_slice;
2855 	sc->sc_balance = md->md_balance;
2856 	sc->sc_mediasize = md->md_mediasize;
2857 	sc->sc_sectorsize = md->md_sectorsize;
2858 	sc->sc_ndisks = md->md_all;
2859 	sc->sc_flags = md->md_mflags;
2860 	sc->sc_bump_id = 0;
2861 	sc->sc_idle = 1;
2862 	sc->sc_last_write = time_uptime;
2863 	sc->sc_writes = 0;
2864 	sx_init(&sc->sc_lock, "gmirror:lock");
2865 	bioq_init(&sc->sc_queue);
2866 	mtx_init(&sc->sc_queue_mtx, "gmirror:queue", NULL, MTX_DEF);
2867 	bioq_init(&sc->sc_regular_delayed);
2868 	bioq_init(&sc->sc_inflight);
2869 	bioq_init(&sc->sc_sync_delayed);
2870 	LIST_INIT(&sc->sc_disks);
2871 	TAILQ_INIT(&sc->sc_events);
2872 	mtx_init(&sc->sc_events_mtx, "gmirror:events", NULL, MTX_DEF);
2873 	callout_init(&sc->sc_callout, CALLOUT_MPSAFE);
2874 	sc->sc_state = G_MIRROR_DEVICE_STATE_STARTING;
2875 	gp->softc = sc;
2876 	sc->sc_geom = gp;
2877 	sc->sc_provider = NULL;
2878 	/*
2879 	 * Synchronization geom.
2880 	 */
2881 	gp = g_new_geomf(mp, "%s.sync", md->md_name);
2882 	gp->softc = sc;
2883 	gp->orphan = g_mirror_orphan;
2884 	sc->sc_sync.ds_geom = gp;
2885 	sc->sc_sync.ds_ndisks = 0;
2886 	error = kproc_create(g_mirror_worker, sc, &sc->sc_worker, 0, 0,
2887 	    "g_mirror %s", md->md_name);
2888 	if (error != 0) {
2889 		G_MIRROR_DEBUG(1, "Cannot create kernel thread for %s.",
2890 		    sc->sc_name);
2891 		g_destroy_geom(sc->sc_sync.ds_geom);
2892 		mtx_destroy(&sc->sc_events_mtx);
2893 		mtx_destroy(&sc->sc_queue_mtx);
2894 		sx_destroy(&sc->sc_lock);
2895 		g_destroy_geom(sc->sc_geom);
2896 		free(sc, M_MIRROR);
2897 		return (NULL);
2898 	}
2899 
2900 	G_MIRROR_DEBUG(1, "Device %s created (%u components, id=%u).",
2901 	    sc->sc_name, sc->sc_ndisks, sc->sc_id);
2902 
2903 	sc->sc_rootmount = root_mount_hold("GMIRROR");
2904 	G_MIRROR_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
2905 	/*
2906 	 * Run timeout.
2907 	 */
2908 	timeout = g_mirror_timeout * hz;
2909 	callout_reset(&sc->sc_callout, timeout, g_mirror_go, sc);
2910 	return (sc->sc_geom);
2911 }
2912 
2913 int
2914 g_mirror_destroy(struct g_mirror_softc *sc, int how)
2915 {
2916 	struct g_mirror_disk *disk;
2917 	struct g_provider *pp;
2918 
2919 	g_topology_assert_not();
2920 	if (sc == NULL)
2921 		return (ENXIO);
2922 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2923 
2924 	pp = sc->sc_provider;
2925 	if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
2926 		switch (how) {
2927 		case G_MIRROR_DESTROY_SOFT:
2928 			G_MIRROR_DEBUG(1,
2929 			    "Device %s is still open (r%dw%de%d).", pp->name,
2930 			    pp->acr, pp->acw, pp->ace);
2931 			return (EBUSY);
2932 		case G_MIRROR_DESTROY_DELAYED:
2933 			G_MIRROR_DEBUG(1,
2934 			    "Device %s will be destroyed on last close.",
2935 			    pp->name);
2936 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2937 				if (disk->d_state ==
2938 				    G_MIRROR_DISK_STATE_SYNCHRONIZING) {
2939 					g_mirror_sync_stop(disk, 1);
2940 				}
2941 			}
2942 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROYING;
2943 			return (EBUSY);
2944 		case G_MIRROR_DESTROY_HARD:
2945 			G_MIRROR_DEBUG(1, "Device %s is still open, so it "
2946 			    "can't be definitely removed.", pp->name);
2947 		}
2948 	}
2949 
2950 	g_topology_lock();
2951 	if (sc->sc_geom->softc == NULL) {
2952 		g_topology_unlock();
2953 		return (0);
2954 	}
2955 	sc->sc_geom->softc = NULL;
2956 	sc->sc_sync.ds_geom->softc = NULL;
2957 	g_topology_unlock();
2958 
2959 	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2960 	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_WAIT;
2961 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
2962 	sx_xunlock(&sc->sc_lock);
2963 	mtx_lock(&sc->sc_queue_mtx);
2964 	wakeup(sc);
2965 	mtx_unlock(&sc->sc_queue_mtx);
2966 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
2967 	while (sc->sc_worker != NULL)
2968 		tsleep(&sc->sc_worker, PRIBIO, "m:destroy", hz / 5);
2969 	G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
2970 	sx_xlock(&sc->sc_lock);
2971 	g_mirror_destroy_device(sc);
2972 	free(sc, M_MIRROR);
2973 	return (0);
2974 }
2975 
2976 static void
2977 g_mirror_taste_orphan(struct g_consumer *cp)
2978 {
2979 
2980 	KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
2981 	    cp->provider->name));
2982 }
2983 
2984 static struct g_geom *
2985 g_mirror_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
2986 {
2987 	struct g_mirror_metadata md;
2988 	struct g_mirror_softc *sc;
2989 	struct g_consumer *cp;
2990 	struct g_geom *gp;
2991 	int error;
2992 
2993 	g_topology_assert();
2994 	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
2995 	G_MIRROR_DEBUG(2, "Tasting %s.", pp->name);
2996 
2997 	gp = g_new_geomf(mp, "mirror:taste");
2998 	/*
2999 	 * This orphan function should be never called.
3000 	 */
3001 	gp->orphan = g_mirror_taste_orphan;
3002 	cp = g_new_consumer(gp);
3003 	g_attach(cp, pp);
3004 	error = g_mirror_read_metadata(cp, &md);
3005 	g_detach(cp);
3006 	g_destroy_consumer(cp);
3007 	g_destroy_geom(gp);
3008 	if (error != 0)
3009 		return (NULL);
3010 	gp = NULL;
3011 
3012 	if (md.md_provider[0] != '\0' &&
3013 	    !g_compare_names(md.md_provider, pp->name))
3014 		return (NULL);
3015 	if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
3016 		return (NULL);
3017 	if ((md.md_dflags & G_MIRROR_DISK_FLAG_INACTIVE) != 0) {
3018 		G_MIRROR_DEBUG(0,
3019 		    "Device %s: provider %s marked as inactive, skipping.",
3020 		    md.md_name, pp->name);
3021 		return (NULL);
3022 	}
3023 	if (g_mirror_debug >= 2)
3024 		mirror_metadata_dump(&md);
3025 
3026 	/*
3027 	 * Let's check if device already exists.
3028 	 */
3029 	sc = NULL;
3030 	LIST_FOREACH(gp, &mp->geom, geom) {
3031 		sc = gp->softc;
3032 		if (sc == NULL)
3033 			continue;
3034 		if (sc->sc_sync.ds_geom == gp)
3035 			continue;
3036 		if (strcmp(md.md_name, sc->sc_name) != 0)
3037 			continue;
3038 		if (md.md_mid != sc->sc_id) {
3039 			G_MIRROR_DEBUG(0, "Device %s already configured.",
3040 			    sc->sc_name);
3041 			return (NULL);
3042 		}
3043 		break;
3044 	}
3045 	if (gp == NULL) {
3046 		gp = g_mirror_create(mp, &md);
3047 		if (gp == NULL) {
3048 			G_MIRROR_DEBUG(0, "Cannot create device %s.",
3049 			    md.md_name);
3050 			return (NULL);
3051 		}
3052 		sc = gp->softc;
3053 	}
3054 	G_MIRROR_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
3055 	g_topology_unlock();
3056 	sx_xlock(&sc->sc_lock);
3057 	error = g_mirror_add_disk(sc, pp, &md);
3058 	if (error != 0) {
3059 		G_MIRROR_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
3060 		    pp->name, gp->name, error);
3061 		if (LIST_EMPTY(&sc->sc_disks)) {
3062 			g_cancel_event(sc);
3063 			g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3064 			g_topology_lock();
3065 			return (NULL);
3066 		}
3067 		gp = NULL;
3068 	}
3069 	sx_xunlock(&sc->sc_lock);
3070 	g_topology_lock();
3071 	return (gp);
3072 }
3073 
3074 static int
3075 g_mirror_destroy_geom(struct gctl_req *req __unused,
3076     struct g_class *mp __unused, struct g_geom *gp)
3077 {
3078 	struct g_mirror_softc *sc;
3079 	int error;
3080 
3081 	g_topology_unlock();
3082 	sc = gp->softc;
3083 	sx_xlock(&sc->sc_lock);
3084 	g_cancel_event(sc);
3085 	error = g_mirror_destroy(gp->softc, G_MIRROR_DESTROY_SOFT);
3086 	if (error != 0)
3087 		sx_xunlock(&sc->sc_lock);
3088 	g_topology_lock();
3089 	return (error);
3090 }
3091 
3092 static void
3093 g_mirror_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
3094     struct g_consumer *cp, struct g_provider *pp)
3095 {
3096 	struct g_mirror_softc *sc;
3097 
3098 	g_topology_assert();
3099 
3100 	sc = gp->softc;
3101 	if (sc == NULL)
3102 		return;
3103 	/* Skip synchronization geom. */
3104 	if (gp == sc->sc_sync.ds_geom)
3105 		return;
3106 	if (pp != NULL) {
3107 		/* Nothing here. */
3108 	} else if (cp != NULL) {
3109 		struct g_mirror_disk *disk;
3110 
3111 		disk = cp->private;
3112 		if (disk == NULL)
3113 			return;
3114 		g_topology_unlock();
3115 		sx_xlock(&sc->sc_lock);
3116 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)disk->d_id);
3117 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
3118 			sbuf_printf(sb, "%s<Synchronized>", indent);
3119 			if (disk->d_sync.ds_offset == 0)
3120 				sbuf_printf(sb, "0%%");
3121 			else {
3122 				sbuf_printf(sb, "%u%%",
3123 				    (u_int)((disk->d_sync.ds_offset * 100) /
3124 				    sc->sc_provider->mediasize));
3125 			}
3126 			sbuf_printf(sb, "</Synchronized>\n");
3127 		}
3128 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent,
3129 		    disk->d_sync.ds_syncid);
3130 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent,
3131 		    disk->d_genid);
3132 		sbuf_printf(sb, "%s<Flags>", indent);
3133 		if (disk->d_flags == 0)
3134 			sbuf_printf(sb, "NONE");
3135 		else {
3136 			int first = 1;
3137 
3138 #define	ADD_FLAG(flag, name)	do {					\
3139 	if ((disk->d_flags & (flag)) != 0) {				\
3140 		if (!first)						\
3141 			sbuf_printf(sb, ", ");				\
3142 		else							\
3143 			first = 0;					\
3144 		sbuf_printf(sb, name);					\
3145 	}								\
3146 } while (0)
3147 			ADD_FLAG(G_MIRROR_DISK_FLAG_DIRTY, "DIRTY");
3148 			ADD_FLAG(G_MIRROR_DISK_FLAG_HARDCODED, "HARDCODED");
3149 			ADD_FLAG(G_MIRROR_DISK_FLAG_INACTIVE, "INACTIVE");
3150 			ADD_FLAG(G_MIRROR_DISK_FLAG_SYNCHRONIZING,
3151 			    "SYNCHRONIZING");
3152 			ADD_FLAG(G_MIRROR_DISK_FLAG_FORCE_SYNC, "FORCE_SYNC");
3153 			ADD_FLAG(G_MIRROR_DISK_FLAG_BROKEN, "BROKEN");
3154 #undef	ADD_FLAG
3155 		}
3156 		sbuf_printf(sb, "</Flags>\n");
3157 		sbuf_printf(sb, "%s<Priority>%u</Priority>\n", indent,
3158 		    disk->d_priority);
3159 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3160 		    g_mirror_disk_state2str(disk->d_state));
3161 		sx_xunlock(&sc->sc_lock);
3162 		g_topology_lock();
3163 	} else {
3164 		g_topology_unlock();
3165 		sx_xlock(&sc->sc_lock);
3166 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
3167 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent, sc->sc_syncid);
3168 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, sc->sc_genid);
3169 		sbuf_printf(sb, "%s<Flags>", indent);
3170 		if (sc->sc_flags == 0)
3171 			sbuf_printf(sb, "NONE");
3172 		else {
3173 			int first = 1;
3174 
3175 #define	ADD_FLAG(flag, name)	do {					\
3176 	if ((sc->sc_flags & (flag)) != 0) {				\
3177 		if (!first)						\
3178 			sbuf_printf(sb, ", ");				\
3179 		else							\
3180 			first = 0;					\
3181 		sbuf_printf(sb, name);					\
3182 	}								\
3183 } while (0)
3184 			ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOFAILSYNC, "NOFAILSYNC");
3185 			ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOAUTOSYNC, "NOAUTOSYNC");
3186 #undef	ADD_FLAG
3187 		}
3188 		sbuf_printf(sb, "</Flags>\n");
3189 		sbuf_printf(sb, "%s<Slice>%u</Slice>\n", indent,
3190 		    (u_int)sc->sc_slice);
3191 		sbuf_printf(sb, "%s<Balance>%s</Balance>\n", indent,
3192 		    balance_name(sc->sc_balance));
3193 		sbuf_printf(sb, "%s<Components>%u</Components>\n", indent,
3194 		    sc->sc_ndisks);
3195 		sbuf_printf(sb, "%s<State>", indent);
3196 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
3197 			sbuf_printf(sb, "%s", "STARTING");
3198 		else if (sc->sc_ndisks ==
3199 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE))
3200 			sbuf_printf(sb, "%s", "COMPLETE");
3201 		else
3202 			sbuf_printf(sb, "%s", "DEGRADED");
3203 		sbuf_printf(sb, "</State>\n");
3204 		sx_xunlock(&sc->sc_lock);
3205 		g_topology_lock();
3206 	}
3207 }
3208 
3209 static void
3210 g_mirror_shutdown_pre_sync(void *arg, int howto)
3211 {
3212 	struct g_class *mp;
3213 	struct g_geom *gp, *gp2;
3214 	struct g_mirror_softc *sc;
3215 	int error;
3216 
3217 	mp = arg;
3218 	DROP_GIANT();
3219 	g_topology_lock();
3220 	LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
3221 		if ((sc = gp->softc) == NULL)
3222 			continue;
3223 		/* Skip synchronization geom. */
3224 		if (gp == sc->sc_sync.ds_geom)
3225 			continue;
3226 		g_topology_unlock();
3227 		sx_xlock(&sc->sc_lock);
3228 		g_cancel_event(sc);
3229 		error = g_mirror_destroy(sc, G_MIRROR_DESTROY_DELAYED);
3230 		if (error != 0)
3231 			sx_xunlock(&sc->sc_lock);
3232 		g_topology_lock();
3233 	}
3234 	g_topology_unlock();
3235 	PICKUP_GIANT();
3236 }
3237 
3238 static void
3239 g_mirror_init(struct g_class *mp)
3240 {
3241 
3242 	g_mirror_pre_sync = EVENTHANDLER_REGISTER(shutdown_pre_sync,
3243 	    g_mirror_shutdown_pre_sync, mp, SHUTDOWN_PRI_FIRST);
3244 	if (g_mirror_pre_sync == NULL)
3245 		G_MIRROR_DEBUG(0, "Warning! Cannot register shutdown event.");
3246 }
3247 
3248 static void
3249 g_mirror_fini(struct g_class *mp)
3250 {
3251 
3252 	if (g_mirror_pre_sync != NULL)
3253 		EVENTHANDLER_DEREGISTER(shutdown_pre_sync, g_mirror_pre_sync);
3254 }
3255 
3256 DECLARE_GEOM_CLASS(g_mirror_class, g_mirror);
3257