xref: /freebsd/sys/geom/mirror/g_mirror.c (revision eb6d21b4ca6d668cf89afd99eef7baeafa712197)
1 /*-
2  * Copyright (c) 2004-2006 Pawel Jakub Dawidek <pjd@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/kernel.h>
33 #include <sys/module.h>
34 #include <sys/limits.h>
35 #include <sys/lock.h>
36 #include <sys/mutex.h>
37 #include <sys/bio.h>
38 #include <sys/sysctl.h>
39 #include <sys/malloc.h>
40 #include <sys/eventhandler.h>
41 #include <vm/uma.h>
42 #include <geom/geom.h>
43 #include <sys/proc.h>
44 #include <sys/kthread.h>
45 #include <sys/sched.h>
46 #include <geom/mirror/g_mirror.h>
47 
48 
49 static MALLOC_DEFINE(M_MIRROR, "mirror_data", "GEOM_MIRROR Data");
50 
51 SYSCTL_DECL(_kern_geom);
52 SYSCTL_NODE(_kern_geom, OID_AUTO, mirror, CTLFLAG_RW, 0, "GEOM_MIRROR stuff");
53 u_int g_mirror_debug = 0;
54 TUNABLE_INT("kern.geom.mirror.debug", &g_mirror_debug);
55 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, debug, CTLFLAG_RW, &g_mirror_debug, 0,
56     "Debug level");
57 static u_int g_mirror_timeout = 4;
58 TUNABLE_INT("kern.geom.mirror.timeout", &g_mirror_timeout);
59 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, timeout, CTLFLAG_RW, &g_mirror_timeout,
60     0, "Time to wait on all mirror components");
61 static u_int g_mirror_idletime = 5;
62 TUNABLE_INT("kern.geom.mirror.idletime", &g_mirror_idletime);
63 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, idletime, CTLFLAG_RW,
64     &g_mirror_idletime, 0, "Mark components as clean when idling");
65 static u_int g_mirror_disconnect_on_failure = 1;
66 TUNABLE_INT("kern.geom.mirror.disconnect_on_failure",
67     &g_mirror_disconnect_on_failure);
68 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, disconnect_on_failure, CTLFLAG_RW,
69     &g_mirror_disconnect_on_failure, 0, "Disconnect component on I/O failure.");
70 static u_int g_mirror_syncreqs = 2;
71 TUNABLE_INT("kern.geom.mirror.sync_requests", &g_mirror_syncreqs);
72 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, sync_requests, CTLFLAG_RDTUN,
73     &g_mirror_syncreqs, 0, "Parallel synchronization I/O requests.");
74 
75 #define	MSLEEP(ident, mtx, priority, wmesg, timeout)	do {		\
76 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, (ident));	\
77 	msleep((ident), (mtx), (priority), (wmesg), (timeout));		\
78 	G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, (ident));	\
79 } while (0)
80 
81 static eventhandler_tag g_mirror_pre_sync = NULL;
82 
83 static int g_mirror_destroy_geom(struct gctl_req *req, struct g_class *mp,
84     struct g_geom *gp);
85 static g_taste_t g_mirror_taste;
86 static void g_mirror_init(struct g_class *mp);
87 static void g_mirror_fini(struct g_class *mp);
88 
89 struct g_class g_mirror_class = {
90 	.name = G_MIRROR_CLASS_NAME,
91 	.version = G_VERSION,
92 	.ctlreq = g_mirror_config,
93 	.taste = g_mirror_taste,
94 	.destroy_geom = g_mirror_destroy_geom,
95 	.init = g_mirror_init,
96 	.fini = g_mirror_fini
97 };
98 
99 
100 static void g_mirror_destroy_provider(struct g_mirror_softc *sc);
101 static int g_mirror_update_disk(struct g_mirror_disk *disk, u_int state);
102 static void g_mirror_update_device(struct g_mirror_softc *sc, boolean_t force);
103 static void g_mirror_dumpconf(struct sbuf *sb, const char *indent,
104     struct g_geom *gp, struct g_consumer *cp, struct g_provider *pp);
105 static void g_mirror_sync_stop(struct g_mirror_disk *disk, int type);
106 static void g_mirror_register_request(struct bio *bp);
107 static void g_mirror_sync_release(struct g_mirror_softc *sc);
108 
109 
110 static const char *
111 g_mirror_disk_state2str(int state)
112 {
113 
114 	switch (state) {
115 	case G_MIRROR_DISK_STATE_NONE:
116 		return ("NONE");
117 	case G_MIRROR_DISK_STATE_NEW:
118 		return ("NEW");
119 	case G_MIRROR_DISK_STATE_ACTIVE:
120 		return ("ACTIVE");
121 	case G_MIRROR_DISK_STATE_STALE:
122 		return ("STALE");
123 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
124 		return ("SYNCHRONIZING");
125 	case G_MIRROR_DISK_STATE_DISCONNECTED:
126 		return ("DISCONNECTED");
127 	case G_MIRROR_DISK_STATE_DESTROY:
128 		return ("DESTROY");
129 	default:
130 		return ("INVALID");
131 	}
132 }
133 
134 static const char *
135 g_mirror_device_state2str(int state)
136 {
137 
138 	switch (state) {
139 	case G_MIRROR_DEVICE_STATE_STARTING:
140 		return ("STARTING");
141 	case G_MIRROR_DEVICE_STATE_RUNNING:
142 		return ("RUNNING");
143 	default:
144 		return ("INVALID");
145 	}
146 }
147 
148 static const char *
149 g_mirror_get_diskname(struct g_mirror_disk *disk)
150 {
151 
152 	if (disk->d_consumer == NULL || disk->d_consumer->provider == NULL)
153 		return ("[unknown]");
154 	return (disk->d_name);
155 }
156 
157 /*
158  * --- Events handling functions ---
159  * Events in geom_mirror are used to maintain disks and device status
160  * from one thread to simplify locking.
161  */
162 static void
163 g_mirror_event_free(struct g_mirror_event *ep)
164 {
165 
166 	free(ep, M_MIRROR);
167 }
168 
169 int
170 g_mirror_event_send(void *arg, int state, int flags)
171 {
172 	struct g_mirror_softc *sc;
173 	struct g_mirror_disk *disk;
174 	struct g_mirror_event *ep;
175 	int error;
176 
177 	ep = malloc(sizeof(*ep), M_MIRROR, M_WAITOK);
178 	G_MIRROR_DEBUG(4, "%s: Sending event %p.", __func__, ep);
179 	if ((flags & G_MIRROR_EVENT_DEVICE) != 0) {
180 		disk = NULL;
181 		sc = arg;
182 	} else {
183 		disk = arg;
184 		sc = disk->d_softc;
185 	}
186 	ep->e_disk = disk;
187 	ep->e_state = state;
188 	ep->e_flags = flags;
189 	ep->e_error = 0;
190 	mtx_lock(&sc->sc_events_mtx);
191 	TAILQ_INSERT_TAIL(&sc->sc_events, ep, e_next);
192 	mtx_unlock(&sc->sc_events_mtx);
193 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
194 	mtx_lock(&sc->sc_queue_mtx);
195 	wakeup(sc);
196 	mtx_unlock(&sc->sc_queue_mtx);
197 	if ((flags & G_MIRROR_EVENT_DONTWAIT) != 0)
198 		return (0);
199 	sx_assert(&sc->sc_lock, SX_XLOCKED);
200 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, ep);
201 	sx_xunlock(&sc->sc_lock);
202 	while ((ep->e_flags & G_MIRROR_EVENT_DONE) == 0) {
203 		mtx_lock(&sc->sc_events_mtx);
204 		MSLEEP(ep, &sc->sc_events_mtx, PRIBIO | PDROP, "m:event",
205 		    hz * 5);
206 	}
207 	error = ep->e_error;
208 	g_mirror_event_free(ep);
209 	sx_xlock(&sc->sc_lock);
210 	return (error);
211 }
212 
213 static struct g_mirror_event *
214 g_mirror_event_get(struct g_mirror_softc *sc)
215 {
216 	struct g_mirror_event *ep;
217 
218 	mtx_lock(&sc->sc_events_mtx);
219 	ep = TAILQ_FIRST(&sc->sc_events);
220 	mtx_unlock(&sc->sc_events_mtx);
221 	return (ep);
222 }
223 
224 static void
225 g_mirror_event_remove(struct g_mirror_softc *sc, struct g_mirror_event *ep)
226 {
227 
228 	mtx_lock(&sc->sc_events_mtx);
229 	TAILQ_REMOVE(&sc->sc_events, ep, e_next);
230 	mtx_unlock(&sc->sc_events_mtx);
231 }
232 
233 static void
234 g_mirror_event_cancel(struct g_mirror_disk *disk)
235 {
236 	struct g_mirror_softc *sc;
237 	struct g_mirror_event *ep, *tmpep;
238 
239 	sc = disk->d_softc;
240 	sx_assert(&sc->sc_lock, SX_XLOCKED);
241 
242 	mtx_lock(&sc->sc_events_mtx);
243 	TAILQ_FOREACH_SAFE(ep, &sc->sc_events, e_next, tmpep) {
244 		if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0)
245 			continue;
246 		if (ep->e_disk != disk)
247 			continue;
248 		TAILQ_REMOVE(&sc->sc_events, ep, e_next);
249 		if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
250 			g_mirror_event_free(ep);
251 		else {
252 			ep->e_error = ECANCELED;
253 			wakeup(ep);
254 		}
255 	}
256 	mtx_unlock(&sc->sc_events_mtx);
257 }
258 
259 /*
260  * Return the number of disks in given state.
261  * If state is equal to -1, count all connected disks.
262  */
263 u_int
264 g_mirror_ndisks(struct g_mirror_softc *sc, int state)
265 {
266 	struct g_mirror_disk *disk;
267 	u_int n = 0;
268 
269 	sx_assert(&sc->sc_lock, SX_LOCKED);
270 
271 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
272 		if (state == -1 || disk->d_state == state)
273 			n++;
274 	}
275 	return (n);
276 }
277 
278 /*
279  * Find a disk in mirror by its disk ID.
280  */
281 static struct g_mirror_disk *
282 g_mirror_id2disk(struct g_mirror_softc *sc, uint32_t id)
283 {
284 	struct g_mirror_disk *disk;
285 
286 	sx_assert(&sc->sc_lock, SX_XLOCKED);
287 
288 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
289 		if (disk->d_id == id)
290 			return (disk);
291 	}
292 	return (NULL);
293 }
294 
295 static u_int
296 g_mirror_nrequests(struct g_mirror_softc *sc, struct g_consumer *cp)
297 {
298 	struct bio *bp;
299 	u_int nreqs = 0;
300 
301 	mtx_lock(&sc->sc_queue_mtx);
302 	TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
303 		if (bp->bio_from == cp)
304 			nreqs++;
305 	}
306 	mtx_unlock(&sc->sc_queue_mtx);
307 	return (nreqs);
308 }
309 
310 static int
311 g_mirror_is_busy(struct g_mirror_softc *sc, struct g_consumer *cp)
312 {
313 
314 	if (cp->index > 0) {
315 		G_MIRROR_DEBUG(2,
316 		    "I/O requests for %s exist, can't destroy it now.",
317 		    cp->provider->name);
318 		return (1);
319 	}
320 	if (g_mirror_nrequests(sc, cp) > 0) {
321 		G_MIRROR_DEBUG(2,
322 		    "I/O requests for %s in queue, can't destroy it now.",
323 		    cp->provider->name);
324 		return (1);
325 	}
326 	return (0);
327 }
328 
329 static void
330 g_mirror_destroy_consumer(void *arg, int flags __unused)
331 {
332 	struct g_consumer *cp;
333 
334 	g_topology_assert();
335 
336 	cp = arg;
337 	G_MIRROR_DEBUG(1, "Consumer %s destroyed.", cp->provider->name);
338 	g_detach(cp);
339 	g_destroy_consumer(cp);
340 }
341 
342 static void
343 g_mirror_kill_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
344 {
345 	struct g_provider *pp;
346 	int retaste_wait;
347 
348 	g_topology_assert();
349 
350 	cp->private = NULL;
351 	if (g_mirror_is_busy(sc, cp))
352 		return;
353 	pp = cp->provider;
354 	retaste_wait = 0;
355 	if (cp->acw == 1) {
356 		if ((pp->geom->flags & G_GEOM_WITHER) == 0)
357 			retaste_wait = 1;
358 	}
359 	G_MIRROR_DEBUG(2, "Access %s r%dw%de%d = %d", pp->name, -cp->acr,
360 	    -cp->acw, -cp->ace, 0);
361 	if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
362 		g_access(cp, -cp->acr, -cp->acw, -cp->ace);
363 	if (retaste_wait) {
364 		/*
365 		 * After retaste event was send (inside g_access()), we can send
366 		 * event to detach and destroy consumer.
367 		 * A class, which has consumer to the given provider connected
368 		 * will not receive retaste event for the provider.
369 		 * This is the way how I ignore retaste events when I close
370 		 * consumers opened for write: I detach and destroy consumer
371 		 * after retaste event is sent.
372 		 */
373 		g_post_event(g_mirror_destroy_consumer, cp, M_WAITOK, NULL);
374 		return;
375 	}
376 	G_MIRROR_DEBUG(1, "Consumer %s destroyed.", pp->name);
377 	g_detach(cp);
378 	g_destroy_consumer(cp);
379 }
380 
381 static int
382 g_mirror_connect_disk(struct g_mirror_disk *disk, struct g_provider *pp)
383 {
384 	struct g_consumer *cp;
385 	int error;
386 
387 	g_topology_assert_not();
388 	KASSERT(disk->d_consumer == NULL,
389 	    ("Disk already connected (device %s).", disk->d_softc->sc_name));
390 
391 	g_topology_lock();
392 	cp = g_new_consumer(disk->d_softc->sc_geom);
393 	error = g_attach(cp, pp);
394 	if (error != 0) {
395 		g_destroy_consumer(cp);
396 		g_topology_unlock();
397 		return (error);
398 	}
399 	error = g_access(cp, 1, 1, 1);
400 	if (error != 0) {
401 		g_detach(cp);
402 		g_destroy_consumer(cp);
403 		g_topology_unlock();
404 		G_MIRROR_DEBUG(0, "Cannot open consumer %s (error=%d).",
405 		    pp->name, error);
406 		return (error);
407 	}
408 	g_topology_unlock();
409 	disk->d_consumer = cp;
410 	disk->d_consumer->private = disk;
411 	disk->d_consumer->index = 0;
412 
413 	G_MIRROR_DEBUG(2, "Disk %s connected.", g_mirror_get_diskname(disk));
414 	return (0);
415 }
416 
417 static void
418 g_mirror_disconnect_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
419 {
420 
421 	g_topology_assert();
422 
423 	if (cp == NULL)
424 		return;
425 	if (cp->provider != NULL)
426 		g_mirror_kill_consumer(sc, cp);
427 	else
428 		g_destroy_consumer(cp);
429 }
430 
431 /*
432  * Initialize disk. This means allocate memory, create consumer, attach it
433  * to the provider and open access (r1w1e1) to it.
434  */
435 static struct g_mirror_disk *
436 g_mirror_init_disk(struct g_mirror_softc *sc, struct g_provider *pp,
437     struct g_mirror_metadata *md, int *errorp)
438 {
439 	struct g_mirror_disk *disk;
440 	int error;
441 
442 	disk = malloc(sizeof(*disk), M_MIRROR, M_NOWAIT | M_ZERO);
443 	if (disk == NULL) {
444 		error = ENOMEM;
445 		goto fail;
446 	}
447 	disk->d_softc = sc;
448 	error = g_mirror_connect_disk(disk, pp);
449 	if (error != 0)
450 		goto fail;
451 	disk->d_id = md->md_did;
452 	disk->d_state = G_MIRROR_DISK_STATE_NONE;
453 	disk->d_priority = md->md_priority;
454 	disk->d_flags = md->md_dflags;
455 	if (md->md_provider[0] != '\0')
456 		disk->d_flags |= G_MIRROR_DISK_FLAG_HARDCODED;
457 	disk->d_sync.ds_consumer = NULL;
458 	disk->d_sync.ds_offset = md->md_sync_offset;
459 	disk->d_sync.ds_offset_done = md->md_sync_offset;
460 	disk->d_genid = md->md_genid;
461 	disk->d_sync.ds_syncid = md->md_syncid;
462 	if (errorp != NULL)
463 		*errorp = 0;
464 	return (disk);
465 fail:
466 	if (errorp != NULL)
467 		*errorp = error;
468 	if (disk != NULL)
469 		free(disk, M_MIRROR);
470 	return (NULL);
471 }
472 
473 static void
474 g_mirror_destroy_disk(struct g_mirror_disk *disk)
475 {
476 	struct g_mirror_softc *sc;
477 
478 	g_topology_assert_not();
479 	sc = disk->d_softc;
480 	sx_assert(&sc->sc_lock, SX_XLOCKED);
481 
482 	LIST_REMOVE(disk, d_next);
483 	g_mirror_event_cancel(disk);
484 	if (sc->sc_hint == disk)
485 		sc->sc_hint = NULL;
486 	switch (disk->d_state) {
487 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
488 		g_mirror_sync_stop(disk, 1);
489 		/* FALLTHROUGH */
490 	case G_MIRROR_DISK_STATE_NEW:
491 	case G_MIRROR_DISK_STATE_STALE:
492 	case G_MIRROR_DISK_STATE_ACTIVE:
493 		g_topology_lock();
494 		g_mirror_disconnect_consumer(sc, disk->d_consumer);
495 		g_topology_unlock();
496 		free(disk, M_MIRROR);
497 		break;
498 	default:
499 		KASSERT(0 == 1, ("Wrong disk state (%s, %s).",
500 		    g_mirror_get_diskname(disk),
501 		    g_mirror_disk_state2str(disk->d_state)));
502 	}
503 }
504 
505 static void
506 g_mirror_destroy_device(struct g_mirror_softc *sc)
507 {
508 	struct g_mirror_disk *disk;
509 	struct g_mirror_event *ep;
510 	struct g_geom *gp;
511 	struct g_consumer *cp, *tmpcp;
512 
513 	g_topology_assert_not();
514 	sx_assert(&sc->sc_lock, SX_XLOCKED);
515 
516 	gp = sc->sc_geom;
517 	if (sc->sc_provider != NULL)
518 		g_mirror_destroy_provider(sc);
519 	for (disk = LIST_FIRST(&sc->sc_disks); disk != NULL;
520 	    disk = LIST_FIRST(&sc->sc_disks)) {
521 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
522 		g_mirror_update_metadata(disk);
523 		g_mirror_destroy_disk(disk);
524 	}
525 	while ((ep = g_mirror_event_get(sc)) != NULL) {
526 		g_mirror_event_remove(sc, ep);
527 		if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
528 			g_mirror_event_free(ep);
529 		else {
530 			ep->e_error = ECANCELED;
531 			ep->e_flags |= G_MIRROR_EVENT_DONE;
532 			G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, ep);
533 			mtx_lock(&sc->sc_events_mtx);
534 			wakeup(ep);
535 			mtx_unlock(&sc->sc_events_mtx);
536 		}
537 	}
538 	callout_drain(&sc->sc_callout);
539 
540 	g_topology_lock();
541 	LIST_FOREACH_SAFE(cp, &sc->sc_sync.ds_geom->consumer, consumer, tmpcp) {
542 		g_mirror_disconnect_consumer(sc, cp);
543 	}
544 	g_wither_geom(sc->sc_sync.ds_geom, ENXIO);
545 	G_MIRROR_DEBUG(0, "Device %s destroyed.", gp->name);
546 	g_wither_geom(gp, ENXIO);
547 	g_topology_unlock();
548 	mtx_destroy(&sc->sc_queue_mtx);
549 	mtx_destroy(&sc->sc_events_mtx);
550 	sx_xunlock(&sc->sc_lock);
551 	sx_destroy(&sc->sc_lock);
552 }
553 
554 static void
555 g_mirror_orphan(struct g_consumer *cp)
556 {
557 	struct g_mirror_disk *disk;
558 
559 	g_topology_assert();
560 
561 	disk = cp->private;
562 	if (disk == NULL)
563 		return;
564 	disk->d_softc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
565 	g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
566 	    G_MIRROR_EVENT_DONTWAIT);
567 }
568 
569 /*
570  * Function should return the next active disk on the list.
571  * It is possible that it will be the same disk as given.
572  * If there are no active disks on list, NULL is returned.
573  */
574 static __inline struct g_mirror_disk *
575 g_mirror_find_next(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
576 {
577 	struct g_mirror_disk *dp;
578 
579 	for (dp = LIST_NEXT(disk, d_next); dp != disk;
580 	    dp = LIST_NEXT(dp, d_next)) {
581 		if (dp == NULL)
582 			dp = LIST_FIRST(&sc->sc_disks);
583 		if (dp->d_state == G_MIRROR_DISK_STATE_ACTIVE)
584 			break;
585 	}
586 	if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
587 		return (NULL);
588 	return (dp);
589 }
590 
591 static struct g_mirror_disk *
592 g_mirror_get_disk(struct g_mirror_softc *sc)
593 {
594 	struct g_mirror_disk *disk;
595 
596 	if (sc->sc_hint == NULL) {
597 		sc->sc_hint = LIST_FIRST(&sc->sc_disks);
598 		if (sc->sc_hint == NULL)
599 			return (NULL);
600 	}
601 	disk = sc->sc_hint;
602 	if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE) {
603 		disk = g_mirror_find_next(sc, disk);
604 		if (disk == NULL)
605 			return (NULL);
606 	}
607 	sc->sc_hint = g_mirror_find_next(sc, disk);
608 	return (disk);
609 }
610 
611 static int
612 g_mirror_write_metadata(struct g_mirror_disk *disk,
613     struct g_mirror_metadata *md)
614 {
615 	struct g_mirror_softc *sc;
616 	struct g_consumer *cp;
617 	off_t offset, length;
618 	u_char *sector;
619 	int error = 0;
620 
621 	g_topology_assert_not();
622 	sc = disk->d_softc;
623 	sx_assert(&sc->sc_lock, SX_LOCKED);
624 
625 	cp = disk->d_consumer;
626 	KASSERT(cp != NULL, ("NULL consumer (%s).", sc->sc_name));
627 	KASSERT(cp->provider != NULL, ("NULL provider (%s).", sc->sc_name));
628 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
629 	    ("Consumer %s closed? (r%dw%de%d).", cp->provider->name, cp->acr,
630 	    cp->acw, cp->ace));
631 	length = cp->provider->sectorsize;
632 	offset = cp->provider->mediasize - length;
633 	sector = malloc((size_t)length, M_MIRROR, M_WAITOK | M_ZERO);
634 	if (md != NULL)
635 		mirror_metadata_encode(md, sector);
636 	error = g_write_data(cp, offset, sector, length);
637 	free(sector, M_MIRROR);
638 	if (error != 0) {
639 		if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
640 			disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
641 			G_MIRROR_DEBUG(0, "Cannot write metadata on %s "
642 			    "(device=%s, error=%d).",
643 			    g_mirror_get_diskname(disk), sc->sc_name, error);
644 		} else {
645 			G_MIRROR_DEBUG(1, "Cannot write metadata on %s "
646 			    "(device=%s, error=%d).",
647 			    g_mirror_get_diskname(disk), sc->sc_name, error);
648 		}
649 		if (g_mirror_disconnect_on_failure &&
650 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1) {
651 			sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
652 			g_mirror_event_send(disk,
653 			    G_MIRROR_DISK_STATE_DISCONNECTED,
654 			    G_MIRROR_EVENT_DONTWAIT);
655 		}
656 	}
657 	return (error);
658 }
659 
660 static int
661 g_mirror_clear_metadata(struct g_mirror_disk *disk)
662 {
663 	int error;
664 
665 	g_topology_assert_not();
666 	sx_assert(&disk->d_softc->sc_lock, SX_LOCKED);
667 
668 	error = g_mirror_write_metadata(disk, NULL);
669 	if (error == 0) {
670 		G_MIRROR_DEBUG(2, "Metadata on %s cleared.",
671 		    g_mirror_get_diskname(disk));
672 	} else {
673 		G_MIRROR_DEBUG(0,
674 		    "Cannot clear metadata on disk %s (error=%d).",
675 		    g_mirror_get_diskname(disk), error);
676 	}
677 	return (error);
678 }
679 
680 void
681 g_mirror_fill_metadata(struct g_mirror_softc *sc, struct g_mirror_disk *disk,
682     struct g_mirror_metadata *md)
683 {
684 
685 	strlcpy(md->md_magic, G_MIRROR_MAGIC, sizeof(md->md_magic));
686 	md->md_version = G_MIRROR_VERSION;
687 	strlcpy(md->md_name, sc->sc_name, sizeof(md->md_name));
688 	md->md_mid = sc->sc_id;
689 	md->md_all = sc->sc_ndisks;
690 	md->md_slice = sc->sc_slice;
691 	md->md_balance = sc->sc_balance;
692 	md->md_genid = sc->sc_genid;
693 	md->md_mediasize = sc->sc_mediasize;
694 	md->md_sectorsize = sc->sc_sectorsize;
695 	md->md_mflags = (sc->sc_flags & G_MIRROR_DEVICE_FLAG_MASK);
696 	bzero(md->md_provider, sizeof(md->md_provider));
697 	if (disk == NULL) {
698 		md->md_did = arc4random();
699 		md->md_priority = 0;
700 		md->md_syncid = 0;
701 		md->md_dflags = 0;
702 		md->md_sync_offset = 0;
703 		md->md_provsize = 0;
704 	} else {
705 		md->md_did = disk->d_id;
706 		md->md_priority = disk->d_priority;
707 		md->md_syncid = disk->d_sync.ds_syncid;
708 		md->md_dflags = (disk->d_flags & G_MIRROR_DISK_FLAG_MASK);
709 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
710 			md->md_sync_offset = disk->d_sync.ds_offset_done;
711 		else
712 			md->md_sync_offset = 0;
713 		if ((disk->d_flags & G_MIRROR_DISK_FLAG_HARDCODED) != 0) {
714 			strlcpy(md->md_provider,
715 			    disk->d_consumer->provider->name,
716 			    sizeof(md->md_provider));
717 		}
718 		md->md_provsize = disk->d_consumer->provider->mediasize;
719 	}
720 }
721 
722 void
723 g_mirror_update_metadata(struct g_mirror_disk *disk)
724 {
725 	struct g_mirror_softc *sc;
726 	struct g_mirror_metadata md;
727 	int error;
728 
729 	g_topology_assert_not();
730 	sc = disk->d_softc;
731 	sx_assert(&sc->sc_lock, SX_LOCKED);
732 
733 	g_mirror_fill_metadata(sc, disk, &md);
734 	error = g_mirror_write_metadata(disk, &md);
735 	if (error == 0) {
736 		G_MIRROR_DEBUG(2, "Metadata on %s updated.",
737 		    g_mirror_get_diskname(disk));
738 	} else {
739 		G_MIRROR_DEBUG(0,
740 		    "Cannot update metadata on disk %s (error=%d).",
741 		    g_mirror_get_diskname(disk), error);
742 	}
743 }
744 
745 static void
746 g_mirror_bump_syncid(struct g_mirror_softc *sc)
747 {
748 	struct g_mirror_disk *disk;
749 
750 	g_topology_assert_not();
751 	sx_assert(&sc->sc_lock, SX_XLOCKED);
752 	KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
753 	    ("%s called with no active disks (device=%s).", __func__,
754 	    sc->sc_name));
755 
756 	sc->sc_syncid++;
757 	G_MIRROR_DEBUG(1, "Device %s: syncid bumped to %u.", sc->sc_name,
758 	    sc->sc_syncid);
759 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
760 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
761 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
762 			disk->d_sync.ds_syncid = sc->sc_syncid;
763 			g_mirror_update_metadata(disk);
764 		}
765 	}
766 }
767 
768 static void
769 g_mirror_bump_genid(struct g_mirror_softc *sc)
770 {
771 	struct g_mirror_disk *disk;
772 
773 	g_topology_assert_not();
774 	sx_assert(&sc->sc_lock, SX_XLOCKED);
775 	KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
776 	    ("%s called with no active disks (device=%s).", __func__,
777 	    sc->sc_name));
778 
779 	sc->sc_genid++;
780 	G_MIRROR_DEBUG(1, "Device %s: genid bumped to %u.", sc->sc_name,
781 	    sc->sc_genid);
782 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
783 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
784 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
785 			disk->d_genid = sc->sc_genid;
786 			g_mirror_update_metadata(disk);
787 		}
788 	}
789 }
790 
791 static int
792 g_mirror_idle(struct g_mirror_softc *sc, int acw)
793 {
794 	struct g_mirror_disk *disk;
795 	int timeout;
796 
797 	g_topology_assert_not();
798 	sx_assert(&sc->sc_lock, SX_XLOCKED);
799 
800 	if (sc->sc_provider == NULL)
801 		return (0);
802 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
803 		return (0);
804 	if (sc->sc_idle)
805 		return (0);
806 	if (sc->sc_writes > 0)
807 		return (0);
808 	if (acw > 0 || (acw == -1 && sc->sc_provider->acw > 0)) {
809 		timeout = g_mirror_idletime - (time_uptime - sc->sc_last_write);
810 		if (timeout > 0)
811 			return (timeout);
812 	}
813 	sc->sc_idle = 1;
814 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
815 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
816 			continue;
817 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as clean.",
818 		    g_mirror_get_diskname(disk), sc->sc_name);
819 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
820 		g_mirror_update_metadata(disk);
821 	}
822 	return (0);
823 }
824 
825 static void
826 g_mirror_unidle(struct g_mirror_softc *sc)
827 {
828 	struct g_mirror_disk *disk;
829 
830 	g_topology_assert_not();
831 	sx_assert(&sc->sc_lock, SX_XLOCKED);
832 
833 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
834 		return;
835 	sc->sc_idle = 0;
836 	sc->sc_last_write = time_uptime;
837 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
838 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
839 			continue;
840 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as dirty.",
841 		    g_mirror_get_diskname(disk), sc->sc_name);
842 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
843 		g_mirror_update_metadata(disk);
844 	}
845 }
846 
847 static __inline int
848 bintime_cmp(struct bintime *bt1, struct bintime *bt2)
849 {
850 
851 	if (bt1->sec < bt2->sec)
852 		return (-1);
853 	else if (bt1->sec > bt2->sec)
854 		return (1);
855 	if (bt1->frac < bt2->frac)
856 		return (-1);
857 	else if (bt1->frac > bt2->frac)
858 		return (1);
859 	return (0);
860 }
861 
862 static void
863 g_mirror_done(struct bio *bp)
864 {
865 	struct g_mirror_softc *sc;
866 
867 	sc = bp->bio_from->geom->softc;
868 	bp->bio_cflags = G_MIRROR_BIO_FLAG_REGULAR;
869 	mtx_lock(&sc->sc_queue_mtx);
870 	bioq_disksort(&sc->sc_queue, bp);
871 	wakeup(sc);
872 	mtx_unlock(&sc->sc_queue_mtx);
873 }
874 
875 static void
876 g_mirror_regular_request(struct bio *bp)
877 {
878 	struct g_mirror_softc *sc;
879 	struct g_mirror_disk *disk;
880 	struct bio *pbp;
881 
882 	g_topology_assert_not();
883 
884 	pbp = bp->bio_parent;
885 	sc = pbp->bio_to->geom->softc;
886 	bp->bio_from->index--;
887 	if (bp->bio_cmd == BIO_WRITE)
888 		sc->sc_writes--;
889 	disk = bp->bio_from->private;
890 	if (disk == NULL) {
891 		g_topology_lock();
892 		g_mirror_kill_consumer(sc, bp->bio_from);
893 		g_topology_unlock();
894 	}
895 
896 	pbp->bio_inbed++;
897 	KASSERT(pbp->bio_inbed <= pbp->bio_children,
898 	    ("bio_inbed (%u) is bigger than bio_children (%u).", pbp->bio_inbed,
899 	    pbp->bio_children));
900 	if (bp->bio_error == 0 && pbp->bio_error == 0) {
901 		G_MIRROR_LOGREQ(3, bp, "Request delivered.");
902 		g_destroy_bio(bp);
903 		if (pbp->bio_children == pbp->bio_inbed) {
904 			G_MIRROR_LOGREQ(3, pbp, "Request delivered.");
905 			pbp->bio_completed = pbp->bio_length;
906 			if (pbp->bio_cmd == BIO_WRITE) {
907 				bioq_remove(&sc->sc_inflight, pbp);
908 				/* Release delayed sync requests if possible. */
909 				g_mirror_sync_release(sc);
910 			}
911 			g_io_deliver(pbp, pbp->bio_error);
912 		}
913 		return;
914 	} else if (bp->bio_error != 0) {
915 		if (pbp->bio_error == 0)
916 			pbp->bio_error = bp->bio_error;
917 		if (disk != NULL) {
918 			if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
919 				disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
920 				G_MIRROR_LOGREQ(0, bp,
921 				    "Request failed (error=%d).",
922 				    bp->bio_error);
923 			} else {
924 				G_MIRROR_LOGREQ(1, bp,
925 				    "Request failed (error=%d).",
926 				    bp->bio_error);
927 			}
928 			if (g_mirror_disconnect_on_failure &&
929 			    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1)
930 			{
931 				sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
932 				g_mirror_event_send(disk,
933 				    G_MIRROR_DISK_STATE_DISCONNECTED,
934 				    G_MIRROR_EVENT_DONTWAIT);
935 			}
936 		}
937 		switch (pbp->bio_cmd) {
938 		case BIO_DELETE:
939 		case BIO_WRITE:
940 			pbp->bio_inbed--;
941 			pbp->bio_children--;
942 			break;
943 		}
944 	}
945 	g_destroy_bio(bp);
946 
947 	switch (pbp->bio_cmd) {
948 	case BIO_READ:
949 		if (pbp->bio_inbed < pbp->bio_children)
950 			break;
951 		if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 1)
952 			g_io_deliver(pbp, pbp->bio_error);
953 		else {
954 			pbp->bio_error = 0;
955 			mtx_lock(&sc->sc_queue_mtx);
956 			bioq_disksort(&sc->sc_queue, pbp);
957 			G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
958 			wakeup(sc);
959 			mtx_unlock(&sc->sc_queue_mtx);
960 		}
961 		break;
962 	case BIO_DELETE:
963 	case BIO_WRITE:
964 		if (pbp->bio_children == 0) {
965 			/*
966 			 * All requests failed.
967 			 */
968 		} else if (pbp->bio_inbed < pbp->bio_children) {
969 			/* Do nothing. */
970 			break;
971 		} else if (pbp->bio_children == pbp->bio_inbed) {
972 			/* Some requests succeeded. */
973 			pbp->bio_error = 0;
974 			pbp->bio_completed = pbp->bio_length;
975 		}
976 		bioq_remove(&sc->sc_inflight, pbp);
977 		/* Release delayed sync requests if possible. */
978 		g_mirror_sync_release(sc);
979 		g_io_deliver(pbp, pbp->bio_error);
980 		break;
981 	default:
982 		KASSERT(1 == 0, ("Invalid request: %u.", pbp->bio_cmd));
983 		break;
984 	}
985 }
986 
987 static void
988 g_mirror_sync_done(struct bio *bp)
989 {
990 	struct g_mirror_softc *sc;
991 
992 	G_MIRROR_LOGREQ(3, bp, "Synchronization request delivered.");
993 	sc = bp->bio_from->geom->softc;
994 	bp->bio_cflags = G_MIRROR_BIO_FLAG_SYNC;
995 	mtx_lock(&sc->sc_queue_mtx);
996 	bioq_disksort(&sc->sc_queue, bp);
997 	wakeup(sc);
998 	mtx_unlock(&sc->sc_queue_mtx);
999 }
1000 
1001 static void
1002 g_mirror_kernel_dump(struct bio *bp)
1003 {
1004 	struct g_mirror_softc *sc;
1005 	struct g_mirror_disk *disk;
1006 	struct bio *cbp;
1007 	struct g_kerneldump *gkd;
1008 
1009 	/*
1010 	 * We configure dumping to the first component, because this component
1011 	 * will be used for reading with 'prefer' balance algorithm.
1012 	 * If the component with the higest priority is currently disconnected
1013 	 * we will not be able to read the dump after the reboot if it will be
1014 	 * connected and synchronized later. Can we do something better?
1015 	 */
1016 	sc = bp->bio_to->geom->softc;
1017 	disk = LIST_FIRST(&sc->sc_disks);
1018 
1019 	gkd = (struct g_kerneldump *)bp->bio_data;
1020 	if (gkd->length > bp->bio_to->mediasize)
1021 		gkd->length = bp->bio_to->mediasize;
1022 	cbp = g_clone_bio(bp);
1023 	if (cbp == NULL) {
1024 		g_io_deliver(bp, ENOMEM);
1025 		return;
1026 	}
1027 	cbp->bio_done = g_std_done;
1028 	g_io_request(cbp, disk->d_consumer);
1029 	G_MIRROR_DEBUG(1, "Kernel dump will go to %s.",
1030 	    g_mirror_get_diskname(disk));
1031 }
1032 
1033 static void
1034 g_mirror_flush(struct g_mirror_softc *sc, struct bio *bp)
1035 {
1036 	struct bio_queue_head queue;
1037 	struct g_mirror_disk *disk;
1038 	struct g_consumer *cp;
1039 	struct bio *cbp;
1040 
1041 	bioq_init(&queue);
1042 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1043 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1044 			continue;
1045 		cbp = g_clone_bio(bp);
1046 		if (cbp == NULL) {
1047 			for (cbp = bioq_first(&queue); cbp != NULL;
1048 			    cbp = bioq_first(&queue)) {
1049 				bioq_remove(&queue, cbp);
1050 				g_destroy_bio(cbp);
1051 			}
1052 			if (bp->bio_error == 0)
1053 				bp->bio_error = ENOMEM;
1054 			g_io_deliver(bp, bp->bio_error);
1055 			return;
1056 		}
1057 		bioq_insert_tail(&queue, cbp);
1058 		cbp->bio_done = g_std_done;
1059 		cbp->bio_caller1 = disk;
1060 		cbp->bio_to = disk->d_consumer->provider;
1061 	}
1062 	for (cbp = bioq_first(&queue); cbp != NULL; cbp = bioq_first(&queue)) {
1063 		bioq_remove(&queue, cbp);
1064 		G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1065 		disk = cbp->bio_caller1;
1066 		cbp->bio_caller1 = NULL;
1067 		cp = disk->d_consumer;
1068 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1069 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1070 		    cp->acr, cp->acw, cp->ace));
1071 		g_io_request(cbp, disk->d_consumer);
1072 	}
1073 }
1074 
1075 static void
1076 g_mirror_start(struct bio *bp)
1077 {
1078 	struct g_mirror_softc *sc;
1079 
1080 	sc = bp->bio_to->geom->softc;
1081 	/*
1082 	 * If sc == NULL or there are no valid disks, provider's error
1083 	 * should be set and g_mirror_start() should not be called at all.
1084 	 */
1085 	KASSERT(sc != NULL && sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1086 	    ("Provider's error should be set (error=%d)(mirror=%s).",
1087 	    bp->bio_to->error, bp->bio_to->name));
1088 	G_MIRROR_LOGREQ(3, bp, "Request received.");
1089 
1090 	switch (bp->bio_cmd) {
1091 	case BIO_READ:
1092 	case BIO_WRITE:
1093 	case BIO_DELETE:
1094 		break;
1095 	case BIO_FLUSH:
1096 		g_mirror_flush(sc, bp);
1097 		return;
1098 	case BIO_GETATTR:
1099 		if (strcmp("GEOM::kerneldump", bp->bio_attribute) == 0) {
1100 			g_mirror_kernel_dump(bp);
1101 			return;
1102 		}
1103 		/* FALLTHROUGH */
1104 	default:
1105 		g_io_deliver(bp, EOPNOTSUPP);
1106 		return;
1107 	}
1108 	mtx_lock(&sc->sc_queue_mtx);
1109 	bioq_disksort(&sc->sc_queue, bp);
1110 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1111 	wakeup(sc);
1112 	mtx_unlock(&sc->sc_queue_mtx);
1113 }
1114 
1115 /*
1116  * Return TRUE if the given request is colliding with a in-progress
1117  * synchronization request.
1118  */
1119 static int
1120 g_mirror_sync_collision(struct g_mirror_softc *sc, struct bio *bp)
1121 {
1122 	struct g_mirror_disk *disk;
1123 	struct bio *sbp;
1124 	off_t rstart, rend, sstart, send;
1125 	int i;
1126 
1127 	if (sc->sc_sync.ds_ndisks == 0)
1128 		return (0);
1129 	rstart = bp->bio_offset;
1130 	rend = bp->bio_offset + bp->bio_length;
1131 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1132 		if (disk->d_state != G_MIRROR_DISK_STATE_SYNCHRONIZING)
1133 			continue;
1134 		for (i = 0; i < g_mirror_syncreqs; i++) {
1135 			sbp = disk->d_sync.ds_bios[i];
1136 			if (sbp == NULL)
1137 				continue;
1138 			sstart = sbp->bio_offset;
1139 			send = sbp->bio_offset + sbp->bio_length;
1140 			if (rend > sstart && rstart < send)
1141 				return (1);
1142 		}
1143 	}
1144 	return (0);
1145 }
1146 
1147 /*
1148  * Return TRUE if the given sync request is colliding with a in-progress regular
1149  * request.
1150  */
1151 static int
1152 g_mirror_regular_collision(struct g_mirror_softc *sc, struct bio *sbp)
1153 {
1154 	off_t rstart, rend, sstart, send;
1155 	struct bio *bp;
1156 
1157 	if (sc->sc_sync.ds_ndisks == 0)
1158 		return (0);
1159 	sstart = sbp->bio_offset;
1160 	send = sbp->bio_offset + sbp->bio_length;
1161 	TAILQ_FOREACH(bp, &sc->sc_inflight.queue, bio_queue) {
1162 		rstart = bp->bio_offset;
1163 		rend = bp->bio_offset + bp->bio_length;
1164 		if (rend > sstart && rstart < send)
1165 			return (1);
1166 	}
1167 	return (0);
1168 }
1169 
1170 /*
1171  * Puts request onto delayed queue.
1172  */
1173 static void
1174 g_mirror_regular_delay(struct g_mirror_softc *sc, struct bio *bp)
1175 {
1176 
1177 	G_MIRROR_LOGREQ(2, bp, "Delaying request.");
1178 	bioq_insert_head(&sc->sc_regular_delayed, bp);
1179 }
1180 
1181 /*
1182  * Puts synchronization request onto delayed queue.
1183  */
1184 static void
1185 g_mirror_sync_delay(struct g_mirror_softc *sc, struct bio *bp)
1186 {
1187 
1188 	G_MIRROR_LOGREQ(2, bp, "Delaying synchronization request.");
1189 	bioq_insert_tail(&sc->sc_sync_delayed, bp);
1190 }
1191 
1192 /*
1193  * Releases delayed regular requests which don't collide anymore with sync
1194  * requests.
1195  */
1196 static void
1197 g_mirror_regular_release(struct g_mirror_softc *sc)
1198 {
1199 	struct bio *bp, *bp2;
1200 
1201 	TAILQ_FOREACH_SAFE(bp, &sc->sc_regular_delayed.queue, bio_queue, bp2) {
1202 		if (g_mirror_sync_collision(sc, bp))
1203 			continue;
1204 		bioq_remove(&sc->sc_regular_delayed, bp);
1205 		G_MIRROR_LOGREQ(2, bp, "Releasing delayed request (%p).", bp);
1206 		mtx_lock(&sc->sc_queue_mtx);
1207 		bioq_insert_head(&sc->sc_queue, bp);
1208 #if 0
1209 		/*
1210 		 * wakeup() is not needed, because this function is called from
1211 		 * the worker thread.
1212 		 */
1213 		wakeup(&sc->sc_queue);
1214 #endif
1215 		mtx_unlock(&sc->sc_queue_mtx);
1216 	}
1217 }
1218 
1219 /*
1220  * Releases delayed sync requests which don't collide anymore with regular
1221  * requests.
1222  */
1223 static void
1224 g_mirror_sync_release(struct g_mirror_softc *sc)
1225 {
1226 	struct bio *bp, *bp2;
1227 
1228 	TAILQ_FOREACH_SAFE(bp, &sc->sc_sync_delayed.queue, bio_queue, bp2) {
1229 		if (g_mirror_regular_collision(sc, bp))
1230 			continue;
1231 		bioq_remove(&sc->sc_sync_delayed, bp);
1232 		G_MIRROR_LOGREQ(2, bp,
1233 		    "Releasing delayed synchronization request.");
1234 		g_io_request(bp, bp->bio_from);
1235 	}
1236 }
1237 
1238 /*
1239  * Handle synchronization requests.
1240  * Every synchronization request is two-steps process: first, READ request is
1241  * send to active provider and then WRITE request (with read data) to the provider
1242  * beeing synchronized. When WRITE is finished, new synchronization request is
1243  * send.
1244  */
1245 static void
1246 g_mirror_sync_request(struct bio *bp)
1247 {
1248 	struct g_mirror_softc *sc;
1249 	struct g_mirror_disk *disk;
1250 
1251 	bp->bio_from->index--;
1252 	sc = bp->bio_from->geom->softc;
1253 	disk = bp->bio_from->private;
1254 	if (disk == NULL) {
1255 		sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1256 		g_topology_lock();
1257 		g_mirror_kill_consumer(sc, bp->bio_from);
1258 		g_topology_unlock();
1259 		free(bp->bio_data, M_MIRROR);
1260 		g_destroy_bio(bp);
1261 		sx_xlock(&sc->sc_lock);
1262 		return;
1263 	}
1264 
1265 	/*
1266 	 * Synchronization request.
1267 	 */
1268 	switch (bp->bio_cmd) {
1269 	case BIO_READ:
1270 	    {
1271 		struct g_consumer *cp;
1272 
1273 		if (bp->bio_error != 0) {
1274 			G_MIRROR_LOGREQ(0, bp,
1275 			    "Synchronization request failed (error=%d).",
1276 			    bp->bio_error);
1277 			g_destroy_bio(bp);
1278 			return;
1279 		}
1280 		G_MIRROR_LOGREQ(3, bp,
1281 		    "Synchronization request half-finished.");
1282 		bp->bio_cmd = BIO_WRITE;
1283 		bp->bio_cflags = 0;
1284 		cp = disk->d_consumer;
1285 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1286 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1287 		    cp->acr, cp->acw, cp->ace));
1288 		cp->index++;
1289 		g_io_request(bp, cp);
1290 		return;
1291 	    }
1292 	case BIO_WRITE:
1293 	    {
1294 		struct g_mirror_disk_sync *sync;
1295 		off_t offset;
1296 		void *data;
1297 		int i;
1298 
1299 		if (bp->bio_error != 0) {
1300 			G_MIRROR_LOGREQ(0, bp,
1301 			    "Synchronization request failed (error=%d).",
1302 			    bp->bio_error);
1303 			g_destroy_bio(bp);
1304 			sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
1305 			g_mirror_event_send(disk,
1306 			    G_MIRROR_DISK_STATE_DISCONNECTED,
1307 			    G_MIRROR_EVENT_DONTWAIT);
1308 			return;
1309 		}
1310 		G_MIRROR_LOGREQ(3, bp, "Synchronization request finished.");
1311 		sync = &disk->d_sync;
1312 		if (sync->ds_offset == sc->sc_mediasize ||
1313 		    sync->ds_consumer == NULL ||
1314 		    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1315 			/* Don't send more synchronization requests. */
1316 			sync->ds_inflight--;
1317 			if (sync->ds_bios != NULL) {
1318 				i = (int)(uintptr_t)bp->bio_caller1;
1319 				sync->ds_bios[i] = NULL;
1320 			}
1321 			free(bp->bio_data, M_MIRROR);
1322 			g_destroy_bio(bp);
1323 			if (sync->ds_inflight > 0)
1324 				return;
1325 			if (sync->ds_consumer == NULL ||
1326 			    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1327 				return;
1328 			}
1329 			/* Disk up-to-date, activate it. */
1330 			g_mirror_event_send(disk, G_MIRROR_DISK_STATE_ACTIVE,
1331 			    G_MIRROR_EVENT_DONTWAIT);
1332 			return;
1333 		}
1334 
1335 		/* Send next synchronization request. */
1336 		data = bp->bio_data;
1337 		bzero(bp, sizeof(*bp));
1338 		bp->bio_cmd = BIO_READ;
1339 		bp->bio_offset = sync->ds_offset;
1340 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1341 		sync->ds_offset += bp->bio_length;
1342 		bp->bio_done = g_mirror_sync_done;
1343 		bp->bio_data = data;
1344 		bp->bio_from = sync->ds_consumer;
1345 		bp->bio_to = sc->sc_provider;
1346 		G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
1347 		sync->ds_consumer->index++;
1348 		/*
1349 		 * Delay the request if it is colliding with a regular request.
1350 		 */
1351 		if (g_mirror_regular_collision(sc, bp))
1352 			g_mirror_sync_delay(sc, bp);
1353 		else
1354 			g_io_request(bp, sync->ds_consumer);
1355 
1356 		/* Release delayed requests if possible. */
1357 		g_mirror_regular_release(sc);
1358 
1359 		/* Find the smallest offset */
1360 		offset = sc->sc_mediasize;
1361 		for (i = 0; i < g_mirror_syncreqs; i++) {
1362 			bp = sync->ds_bios[i];
1363 			if (bp->bio_offset < offset)
1364 				offset = bp->bio_offset;
1365 		}
1366 		if (sync->ds_offset_done + (MAXPHYS * 100) < offset) {
1367 			/* Update offset_done on every 100 blocks. */
1368 			sync->ds_offset_done = offset;
1369 			g_mirror_update_metadata(disk);
1370 		}
1371 		return;
1372 	    }
1373 	default:
1374 		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1375 		    bp->bio_cmd, sc->sc_name));
1376 		break;
1377 	}
1378 }
1379 
1380 static void
1381 g_mirror_request_prefer(struct g_mirror_softc *sc, struct bio *bp)
1382 {
1383 	struct g_mirror_disk *disk;
1384 	struct g_consumer *cp;
1385 	struct bio *cbp;
1386 
1387 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1388 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE)
1389 			break;
1390 	}
1391 	if (disk == NULL) {
1392 		if (bp->bio_error == 0)
1393 			bp->bio_error = ENXIO;
1394 		g_io_deliver(bp, bp->bio_error);
1395 		return;
1396 	}
1397 	cbp = g_clone_bio(bp);
1398 	if (cbp == NULL) {
1399 		if (bp->bio_error == 0)
1400 			bp->bio_error = ENOMEM;
1401 		g_io_deliver(bp, bp->bio_error);
1402 		return;
1403 	}
1404 	/*
1405 	 * Fill in the component buf structure.
1406 	 */
1407 	cp = disk->d_consumer;
1408 	cbp->bio_done = g_mirror_done;
1409 	cbp->bio_to = cp->provider;
1410 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1411 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1412 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1413 	    cp->acw, cp->ace));
1414 	cp->index++;
1415 	g_io_request(cbp, cp);
1416 }
1417 
1418 static void
1419 g_mirror_request_round_robin(struct g_mirror_softc *sc, struct bio *bp)
1420 {
1421 	struct g_mirror_disk *disk;
1422 	struct g_consumer *cp;
1423 	struct bio *cbp;
1424 
1425 	disk = g_mirror_get_disk(sc);
1426 	if (disk == NULL) {
1427 		if (bp->bio_error == 0)
1428 			bp->bio_error = ENXIO;
1429 		g_io_deliver(bp, bp->bio_error);
1430 		return;
1431 	}
1432 	cbp = g_clone_bio(bp);
1433 	if (cbp == NULL) {
1434 		if (bp->bio_error == 0)
1435 			bp->bio_error = ENOMEM;
1436 		g_io_deliver(bp, bp->bio_error);
1437 		return;
1438 	}
1439 	/*
1440 	 * Fill in the component buf structure.
1441 	 */
1442 	cp = disk->d_consumer;
1443 	cbp->bio_done = g_mirror_done;
1444 	cbp->bio_to = cp->provider;
1445 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1446 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1447 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1448 	    cp->acw, cp->ace));
1449 	cp->index++;
1450 	g_io_request(cbp, cp);
1451 }
1452 
1453 #define TRACK_SIZE  (1 * 1024 * 1024)
1454 #define LOAD_SCALE	256
1455 #define ABS(x)		(((x) >= 0) ? (x) : (-(x)))
1456 
1457 static void
1458 g_mirror_request_load(struct g_mirror_softc *sc, struct bio *bp)
1459 {
1460 	struct g_mirror_disk *disk, *dp;
1461 	struct g_consumer *cp;
1462 	struct bio *cbp;
1463 	int prio, best;
1464 
1465 	/* Find a disk with the smallest load. */
1466 	disk = NULL;
1467 	best = INT_MAX;
1468 	LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1469 		if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1470 			continue;
1471 		prio = dp->load;
1472 		/* If disk head is precisely in position - highly prefer it. */
1473 		if (dp->d_last_offset == bp->bio_offset)
1474 			prio -= 2 * LOAD_SCALE;
1475 		else
1476 		/* If disk head is close to position - prefer it. */
1477 		if (ABS(dp->d_last_offset - bp->bio_offset) < TRACK_SIZE)
1478 			prio -= 1 * LOAD_SCALE;
1479 		if (prio <= best) {
1480 			disk = dp;
1481 			best = prio;
1482 		}
1483 	}
1484 	KASSERT(disk != NULL, ("NULL disk for %s.", sc->sc_name));
1485 	cbp = g_clone_bio(bp);
1486 	if (cbp == NULL) {
1487 		if (bp->bio_error == 0)
1488 			bp->bio_error = ENOMEM;
1489 		g_io_deliver(bp, bp->bio_error);
1490 		return;
1491 	}
1492 	/*
1493 	 * Fill in the component buf structure.
1494 	 */
1495 	cp = disk->d_consumer;
1496 	cbp->bio_done = g_mirror_done;
1497 	cbp->bio_to = cp->provider;
1498 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1499 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1500 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1501 	    cp->acw, cp->ace));
1502 	cp->index++;
1503 	/* Remember last head position */
1504 	disk->d_last_offset = bp->bio_offset + bp->bio_length;
1505 	/* Update loads. */
1506 	LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1507 		dp->load = (dp->d_consumer->index * LOAD_SCALE +
1508 		    dp->load * 7) / 8;
1509 	}
1510 	g_io_request(cbp, cp);
1511 }
1512 
1513 static void
1514 g_mirror_request_split(struct g_mirror_softc *sc, struct bio *bp)
1515 {
1516 	struct bio_queue_head queue;
1517 	struct g_mirror_disk *disk;
1518 	struct g_consumer *cp;
1519 	struct bio *cbp;
1520 	off_t left, mod, offset, slice;
1521 	u_char *data;
1522 	u_int ndisks;
1523 
1524 	if (bp->bio_length <= sc->sc_slice) {
1525 		g_mirror_request_round_robin(sc, bp);
1526 		return;
1527 	}
1528 	ndisks = g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE);
1529 	slice = bp->bio_length / ndisks;
1530 	mod = slice % sc->sc_provider->sectorsize;
1531 	if (mod != 0)
1532 		slice += sc->sc_provider->sectorsize - mod;
1533 	/*
1534 	 * Allocate all bios before sending any request, so we can
1535 	 * return ENOMEM in nice and clean way.
1536 	 */
1537 	left = bp->bio_length;
1538 	offset = bp->bio_offset;
1539 	data = bp->bio_data;
1540 	bioq_init(&queue);
1541 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1542 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1543 			continue;
1544 		cbp = g_clone_bio(bp);
1545 		if (cbp == NULL) {
1546 			for (cbp = bioq_first(&queue); cbp != NULL;
1547 			    cbp = bioq_first(&queue)) {
1548 				bioq_remove(&queue, cbp);
1549 				g_destroy_bio(cbp);
1550 			}
1551 			if (bp->bio_error == 0)
1552 				bp->bio_error = ENOMEM;
1553 			g_io_deliver(bp, bp->bio_error);
1554 			return;
1555 		}
1556 		bioq_insert_tail(&queue, cbp);
1557 		cbp->bio_done = g_mirror_done;
1558 		cbp->bio_caller1 = disk;
1559 		cbp->bio_to = disk->d_consumer->provider;
1560 		cbp->bio_offset = offset;
1561 		cbp->bio_data = data;
1562 		cbp->bio_length = MIN(left, slice);
1563 		left -= cbp->bio_length;
1564 		if (left == 0)
1565 			break;
1566 		offset += cbp->bio_length;
1567 		data += cbp->bio_length;
1568 	}
1569 	for (cbp = bioq_first(&queue); cbp != NULL; cbp = bioq_first(&queue)) {
1570 		bioq_remove(&queue, cbp);
1571 		G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1572 		disk = cbp->bio_caller1;
1573 		cbp->bio_caller1 = NULL;
1574 		cp = disk->d_consumer;
1575 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1576 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1577 		    cp->acr, cp->acw, cp->ace));
1578 		disk->d_consumer->index++;
1579 		g_io_request(cbp, disk->d_consumer);
1580 	}
1581 }
1582 
1583 static void
1584 g_mirror_register_request(struct bio *bp)
1585 {
1586 	struct g_mirror_softc *sc;
1587 
1588 	sc = bp->bio_to->geom->softc;
1589 	switch (bp->bio_cmd) {
1590 	case BIO_READ:
1591 		switch (sc->sc_balance) {
1592 		case G_MIRROR_BALANCE_LOAD:
1593 			g_mirror_request_load(sc, bp);
1594 			break;
1595 		case G_MIRROR_BALANCE_PREFER:
1596 			g_mirror_request_prefer(sc, bp);
1597 			break;
1598 		case G_MIRROR_BALANCE_ROUND_ROBIN:
1599 			g_mirror_request_round_robin(sc, bp);
1600 			break;
1601 		case G_MIRROR_BALANCE_SPLIT:
1602 			g_mirror_request_split(sc, bp);
1603 			break;
1604 		}
1605 		return;
1606 	case BIO_WRITE:
1607 	case BIO_DELETE:
1608 	    {
1609 		struct g_mirror_disk *disk;
1610 		struct g_mirror_disk_sync *sync;
1611 		struct bio_queue_head queue;
1612 		struct g_consumer *cp;
1613 		struct bio *cbp;
1614 
1615 		/*
1616 		 * Delay the request if it is colliding with a synchronization
1617 		 * request.
1618 		 */
1619 		if (g_mirror_sync_collision(sc, bp)) {
1620 			g_mirror_regular_delay(sc, bp);
1621 			return;
1622 		}
1623 
1624 		if (sc->sc_idle)
1625 			g_mirror_unidle(sc);
1626 		else
1627 			sc->sc_last_write = time_uptime;
1628 
1629 		/*
1630 		 * Allocate all bios before sending any request, so we can
1631 		 * return ENOMEM in nice and clean way.
1632 		 */
1633 		bioq_init(&queue);
1634 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1635 			sync = &disk->d_sync;
1636 			switch (disk->d_state) {
1637 			case G_MIRROR_DISK_STATE_ACTIVE:
1638 				break;
1639 			case G_MIRROR_DISK_STATE_SYNCHRONIZING:
1640 				if (bp->bio_offset >= sync->ds_offset)
1641 					continue;
1642 				break;
1643 			default:
1644 				continue;
1645 			}
1646 			cbp = g_clone_bio(bp);
1647 			if (cbp == NULL) {
1648 				for (cbp = bioq_first(&queue); cbp != NULL;
1649 				    cbp = bioq_first(&queue)) {
1650 					bioq_remove(&queue, cbp);
1651 					g_destroy_bio(cbp);
1652 				}
1653 				if (bp->bio_error == 0)
1654 					bp->bio_error = ENOMEM;
1655 				g_io_deliver(bp, bp->bio_error);
1656 				return;
1657 			}
1658 			bioq_insert_tail(&queue, cbp);
1659 			cbp->bio_done = g_mirror_done;
1660 			cp = disk->d_consumer;
1661 			cbp->bio_caller1 = cp;
1662 			cbp->bio_to = cp->provider;
1663 			KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1664 			    ("Consumer %s not opened (r%dw%de%d).",
1665 			    cp->provider->name, cp->acr, cp->acw, cp->ace));
1666 		}
1667 		for (cbp = bioq_first(&queue); cbp != NULL;
1668 		    cbp = bioq_first(&queue)) {
1669 			bioq_remove(&queue, cbp);
1670 			G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1671 			cp = cbp->bio_caller1;
1672 			cbp->bio_caller1 = NULL;
1673 			cp->index++;
1674 			sc->sc_writes++;
1675 			g_io_request(cbp, cp);
1676 		}
1677 		/*
1678 		 * Put request onto inflight queue, so we can check if new
1679 		 * synchronization requests don't collide with it.
1680 		 */
1681 		bioq_insert_tail(&sc->sc_inflight, bp);
1682 		/*
1683 		 * Bump syncid on first write.
1684 		 */
1685 		if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0) {
1686 			sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
1687 			g_mirror_bump_syncid(sc);
1688 		}
1689 		return;
1690 	    }
1691 	default:
1692 		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1693 		    bp->bio_cmd, sc->sc_name));
1694 		break;
1695 	}
1696 }
1697 
1698 static int
1699 g_mirror_can_destroy(struct g_mirror_softc *sc)
1700 {
1701 	struct g_geom *gp;
1702 	struct g_consumer *cp;
1703 
1704 	g_topology_assert();
1705 	gp = sc->sc_geom;
1706 	if (gp->softc == NULL)
1707 		return (1);
1708 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1709 		if (g_mirror_is_busy(sc, cp))
1710 			return (0);
1711 	}
1712 	gp = sc->sc_sync.ds_geom;
1713 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1714 		if (g_mirror_is_busy(sc, cp))
1715 			return (0);
1716 	}
1717 	G_MIRROR_DEBUG(2, "No I/O requests for %s, it can be destroyed.",
1718 	    sc->sc_name);
1719 	return (1);
1720 }
1721 
1722 static int
1723 g_mirror_try_destroy(struct g_mirror_softc *sc)
1724 {
1725 
1726 	if (sc->sc_rootmount != NULL) {
1727 		G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
1728 		    sc->sc_rootmount);
1729 		root_mount_rel(sc->sc_rootmount);
1730 		sc->sc_rootmount = NULL;
1731 	}
1732 	g_topology_lock();
1733 	if (!g_mirror_can_destroy(sc)) {
1734 		g_topology_unlock();
1735 		return (0);
1736 	}
1737 	sc->sc_geom->softc = NULL;
1738 	sc->sc_sync.ds_geom->softc = NULL;
1739 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_WAIT) != 0) {
1740 		g_topology_unlock();
1741 		G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1742 		    &sc->sc_worker);
1743 		/* Unlock sc_lock here, as it can be destroyed after wakeup. */
1744 		sx_xunlock(&sc->sc_lock);
1745 		wakeup(&sc->sc_worker);
1746 		sc->sc_worker = NULL;
1747 	} else {
1748 		g_topology_unlock();
1749 		g_mirror_destroy_device(sc);
1750 		free(sc, M_MIRROR);
1751 	}
1752 	return (1);
1753 }
1754 
1755 /*
1756  * Worker thread.
1757  */
1758 static void
1759 g_mirror_worker(void *arg)
1760 {
1761 	struct g_mirror_softc *sc;
1762 	struct g_mirror_event *ep;
1763 	struct bio *bp;
1764 	int timeout;
1765 
1766 	sc = arg;
1767 	thread_lock(curthread);
1768 	sched_prio(curthread, PRIBIO);
1769 	thread_unlock(curthread);
1770 
1771 	sx_xlock(&sc->sc_lock);
1772 	for (;;) {
1773 		G_MIRROR_DEBUG(5, "%s: Let's see...", __func__);
1774 		/*
1775 		 * First take a look at events.
1776 		 * This is important to handle events before any I/O requests.
1777 		 */
1778 		ep = g_mirror_event_get(sc);
1779 		if (ep != NULL) {
1780 			g_mirror_event_remove(sc, ep);
1781 			if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0) {
1782 				/* Update only device status. */
1783 				G_MIRROR_DEBUG(3,
1784 				    "Running event for device %s.",
1785 				    sc->sc_name);
1786 				ep->e_error = 0;
1787 				g_mirror_update_device(sc, 1);
1788 			} else {
1789 				/* Update disk status. */
1790 				G_MIRROR_DEBUG(3, "Running event for disk %s.",
1791 				     g_mirror_get_diskname(ep->e_disk));
1792 				ep->e_error = g_mirror_update_disk(ep->e_disk,
1793 				    ep->e_state);
1794 				if (ep->e_error == 0)
1795 					g_mirror_update_device(sc, 0);
1796 			}
1797 			if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0) {
1798 				KASSERT(ep->e_error == 0,
1799 				    ("Error cannot be handled."));
1800 				g_mirror_event_free(ep);
1801 			} else {
1802 				ep->e_flags |= G_MIRROR_EVENT_DONE;
1803 				G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1804 				    ep);
1805 				mtx_lock(&sc->sc_events_mtx);
1806 				wakeup(ep);
1807 				mtx_unlock(&sc->sc_events_mtx);
1808 			}
1809 			if ((sc->sc_flags &
1810 			    G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1811 				if (g_mirror_try_destroy(sc)) {
1812 					curthread->td_pflags &= ~TDP_GEOM;
1813 					G_MIRROR_DEBUG(1, "Thread exiting.");
1814 					kproc_exit(0);
1815 				}
1816 			}
1817 			G_MIRROR_DEBUG(5, "%s: I'm here 1.", __func__);
1818 			continue;
1819 		}
1820 		/*
1821 		 * Check if we can mark array as CLEAN and if we can't take
1822 		 * how much seconds should we wait.
1823 		 */
1824 		timeout = g_mirror_idle(sc, -1);
1825 		/*
1826 		 * Now I/O requests.
1827 		 */
1828 		/* Get first request from the queue. */
1829 		mtx_lock(&sc->sc_queue_mtx);
1830 		bp = bioq_first(&sc->sc_queue);
1831 		if (bp == NULL) {
1832 			if ((sc->sc_flags &
1833 			    G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1834 				mtx_unlock(&sc->sc_queue_mtx);
1835 				if (g_mirror_try_destroy(sc)) {
1836 					curthread->td_pflags &= ~TDP_GEOM;
1837 					G_MIRROR_DEBUG(1, "Thread exiting.");
1838 					kproc_exit(0);
1839 				}
1840 				mtx_lock(&sc->sc_queue_mtx);
1841 			}
1842 			sx_xunlock(&sc->sc_lock);
1843 			/*
1844 			 * XXX: We can miss an event here, because an event
1845 			 *      can be added without sx-device-lock and without
1846 			 *      mtx-queue-lock. Maybe I should just stop using
1847 			 *      dedicated mutex for events synchronization and
1848 			 *      stick with the queue lock?
1849 			 *      The event will hang here until next I/O request
1850 			 *      or next event is received.
1851 			 */
1852 			MSLEEP(sc, &sc->sc_queue_mtx, PRIBIO | PDROP, "m:w1",
1853 			    timeout * hz);
1854 			sx_xlock(&sc->sc_lock);
1855 			G_MIRROR_DEBUG(5, "%s: I'm here 4.", __func__);
1856 			continue;
1857 		}
1858 		bioq_remove(&sc->sc_queue, bp);
1859 		mtx_unlock(&sc->sc_queue_mtx);
1860 
1861 		if (bp->bio_from->geom == sc->sc_sync.ds_geom &&
1862 		    (bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0) {
1863 			g_mirror_sync_request(bp);	/* READ */
1864 		} else if (bp->bio_to != sc->sc_provider) {
1865 			if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_REGULAR) != 0)
1866 				g_mirror_regular_request(bp);
1867 			else if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0)
1868 				g_mirror_sync_request(bp);	/* WRITE */
1869 			else {
1870 				KASSERT(0,
1871 				    ("Invalid request cflags=0x%hhx to=%s.",
1872 				    bp->bio_cflags, bp->bio_to->name));
1873 			}
1874 		} else {
1875 			g_mirror_register_request(bp);
1876 		}
1877 		G_MIRROR_DEBUG(5, "%s: I'm here 9.", __func__);
1878 	}
1879 }
1880 
1881 static void
1882 g_mirror_update_idle(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
1883 {
1884 
1885 	sx_assert(&sc->sc_lock, SX_LOCKED);
1886 
1887 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
1888 		return;
1889 	if (!sc->sc_idle && (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) == 0) {
1890 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as dirty.",
1891 		    g_mirror_get_diskname(disk), sc->sc_name);
1892 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
1893 	} else if (sc->sc_idle &&
1894 	    (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
1895 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as clean.",
1896 		    g_mirror_get_diskname(disk), sc->sc_name);
1897 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
1898 	}
1899 }
1900 
1901 static void
1902 g_mirror_sync_start(struct g_mirror_disk *disk)
1903 {
1904 	struct g_mirror_softc *sc;
1905 	struct g_consumer *cp;
1906 	struct bio *bp;
1907 	int error, i;
1908 
1909 	g_topology_assert_not();
1910 	sc = disk->d_softc;
1911 	sx_assert(&sc->sc_lock, SX_LOCKED);
1912 
1913 	KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
1914 	    ("Disk %s is not marked for synchronization.",
1915 	    g_mirror_get_diskname(disk)));
1916 	KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1917 	    ("Device not in RUNNING state (%s, %u).", sc->sc_name,
1918 	    sc->sc_state));
1919 
1920 	sx_xunlock(&sc->sc_lock);
1921 	g_topology_lock();
1922 	cp = g_new_consumer(sc->sc_sync.ds_geom);
1923 	error = g_attach(cp, sc->sc_provider);
1924 	KASSERT(error == 0,
1925 	    ("Cannot attach to %s (error=%d).", sc->sc_name, error));
1926 	error = g_access(cp, 1, 0, 0);
1927 	KASSERT(error == 0, ("Cannot open %s (error=%d).", sc->sc_name, error));
1928 	g_topology_unlock();
1929 	sx_xlock(&sc->sc_lock);
1930 
1931 	G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s.", sc->sc_name,
1932 	    g_mirror_get_diskname(disk));
1933 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) == 0)
1934 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
1935 	KASSERT(disk->d_sync.ds_consumer == NULL,
1936 	    ("Sync consumer already exists (device=%s, disk=%s).",
1937 	    sc->sc_name, g_mirror_get_diskname(disk)));
1938 
1939 	disk->d_sync.ds_consumer = cp;
1940 	disk->d_sync.ds_consumer->private = disk;
1941 	disk->d_sync.ds_consumer->index = 0;
1942 
1943 	/*
1944 	 * Allocate memory for synchronization bios and initialize them.
1945 	 */
1946 	disk->d_sync.ds_bios = malloc(sizeof(struct bio *) * g_mirror_syncreqs,
1947 	    M_MIRROR, M_WAITOK);
1948 	for (i = 0; i < g_mirror_syncreqs; i++) {
1949 		bp = g_alloc_bio();
1950 		disk->d_sync.ds_bios[i] = bp;
1951 		bp->bio_parent = NULL;
1952 		bp->bio_cmd = BIO_READ;
1953 		bp->bio_data = malloc(MAXPHYS, M_MIRROR, M_WAITOK);
1954 		bp->bio_cflags = 0;
1955 		bp->bio_offset = disk->d_sync.ds_offset;
1956 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1957 		disk->d_sync.ds_offset += bp->bio_length;
1958 		bp->bio_done = g_mirror_sync_done;
1959 		bp->bio_from = disk->d_sync.ds_consumer;
1960 		bp->bio_to = sc->sc_provider;
1961 		bp->bio_caller1 = (void *)(uintptr_t)i;
1962 	}
1963 
1964 	/* Increase the number of disks in SYNCHRONIZING state. */
1965 	sc->sc_sync.ds_ndisks++;
1966 	/* Set the number of in-flight synchronization requests. */
1967 	disk->d_sync.ds_inflight = g_mirror_syncreqs;
1968 
1969 	/*
1970 	 * Fire off first synchronization requests.
1971 	 */
1972 	for (i = 0; i < g_mirror_syncreqs; i++) {
1973 		bp = disk->d_sync.ds_bios[i];
1974 		G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
1975 		disk->d_sync.ds_consumer->index++;
1976 		/*
1977 		 * Delay the request if it is colliding with a regular request.
1978 		 */
1979 		if (g_mirror_regular_collision(sc, bp))
1980 			g_mirror_sync_delay(sc, bp);
1981 		else
1982 			g_io_request(bp, disk->d_sync.ds_consumer);
1983 	}
1984 }
1985 
1986 /*
1987  * Stop synchronization process.
1988  * type: 0 - synchronization finished
1989  *       1 - synchronization stopped
1990  */
1991 static void
1992 g_mirror_sync_stop(struct g_mirror_disk *disk, int type)
1993 {
1994 	struct g_mirror_softc *sc;
1995 	struct g_consumer *cp;
1996 
1997 	g_topology_assert_not();
1998 	sc = disk->d_softc;
1999 	sx_assert(&sc->sc_lock, SX_LOCKED);
2000 
2001 	KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2002 	    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2003 	    g_mirror_disk_state2str(disk->d_state)));
2004 	if (disk->d_sync.ds_consumer == NULL)
2005 		return;
2006 
2007 	if (type == 0) {
2008 		G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s finished.",
2009 		    sc->sc_name, g_mirror_get_diskname(disk));
2010 	} else /* if (type == 1) */ {
2011 		G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s stopped.",
2012 		    sc->sc_name, g_mirror_get_diskname(disk));
2013 	}
2014 	free(disk->d_sync.ds_bios, M_MIRROR);
2015 	disk->d_sync.ds_bios = NULL;
2016 	cp = disk->d_sync.ds_consumer;
2017 	disk->d_sync.ds_consumer = NULL;
2018 	disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2019 	sc->sc_sync.ds_ndisks--;
2020 	sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
2021 	g_topology_lock();
2022 	g_mirror_kill_consumer(sc, cp);
2023 	g_topology_unlock();
2024 	sx_xlock(&sc->sc_lock);
2025 }
2026 
2027 static void
2028 g_mirror_launch_provider(struct g_mirror_softc *sc)
2029 {
2030 	struct g_mirror_disk *disk;
2031 	struct g_provider *pp;
2032 
2033 	sx_assert(&sc->sc_lock, SX_LOCKED);
2034 
2035 	g_topology_lock();
2036 	pp = g_new_providerf(sc->sc_geom, "mirror/%s", sc->sc_name);
2037 	pp->mediasize = sc->sc_mediasize;
2038 	pp->sectorsize = sc->sc_sectorsize;
2039 	sc->sc_provider = pp;
2040 	g_error_provider(pp, 0);
2041 	g_topology_unlock();
2042 	G_MIRROR_DEBUG(0, "Device %s launched (%u/%u).", pp->name,
2043 	    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE), sc->sc_ndisks);
2044 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2045 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2046 			g_mirror_sync_start(disk);
2047 	}
2048 }
2049 
2050 static void
2051 g_mirror_destroy_provider(struct g_mirror_softc *sc)
2052 {
2053 	struct g_mirror_disk *disk;
2054 	struct bio *bp;
2055 
2056 	g_topology_assert_not();
2057 	KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
2058 	    sc->sc_name));
2059 
2060 	g_topology_lock();
2061 	g_error_provider(sc->sc_provider, ENXIO);
2062 	mtx_lock(&sc->sc_queue_mtx);
2063 	while ((bp = bioq_first(&sc->sc_queue)) != NULL) {
2064 		bioq_remove(&sc->sc_queue, bp);
2065 		g_io_deliver(bp, ENXIO);
2066 	}
2067 	mtx_unlock(&sc->sc_queue_mtx);
2068 	G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.", sc->sc_name,
2069 	    sc->sc_provider->name);
2070 	sc->sc_provider->flags |= G_PF_WITHER;
2071 	g_orphan_provider(sc->sc_provider, ENXIO);
2072 	g_topology_unlock();
2073 	sc->sc_provider = NULL;
2074 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2075 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2076 			g_mirror_sync_stop(disk, 1);
2077 	}
2078 }
2079 
2080 static void
2081 g_mirror_go(void *arg)
2082 {
2083 	struct g_mirror_softc *sc;
2084 
2085 	sc = arg;
2086 	G_MIRROR_DEBUG(0, "Force device %s start due to timeout.", sc->sc_name);
2087 	g_mirror_event_send(sc, 0,
2088 	    G_MIRROR_EVENT_DONTWAIT | G_MIRROR_EVENT_DEVICE);
2089 }
2090 
2091 static u_int
2092 g_mirror_determine_state(struct g_mirror_disk *disk)
2093 {
2094 	struct g_mirror_softc *sc;
2095 	u_int state;
2096 
2097 	sc = disk->d_softc;
2098 	if (sc->sc_syncid == disk->d_sync.ds_syncid) {
2099 		if ((disk->d_flags &
2100 		    G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2101 			/* Disk does not need synchronization. */
2102 			state = G_MIRROR_DISK_STATE_ACTIVE;
2103 		} else {
2104 			if ((sc->sc_flags &
2105 			     G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2106 			    (disk->d_flags &
2107 			     G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2108 				/*
2109 				 * We can start synchronization from
2110 				 * the stored offset.
2111 				 */
2112 				state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2113 			} else {
2114 				state = G_MIRROR_DISK_STATE_STALE;
2115 			}
2116 		}
2117 	} else if (disk->d_sync.ds_syncid < sc->sc_syncid) {
2118 		/*
2119 		 * Reset all synchronization data for this disk,
2120 		 * because if it even was synchronized, it was
2121 		 * synchronized to disks with different syncid.
2122 		 */
2123 		disk->d_flags |= G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2124 		disk->d_sync.ds_offset = 0;
2125 		disk->d_sync.ds_offset_done = 0;
2126 		disk->d_sync.ds_syncid = sc->sc_syncid;
2127 		if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2128 		    (disk->d_flags & G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2129 			state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2130 		} else {
2131 			state = G_MIRROR_DISK_STATE_STALE;
2132 		}
2133 	} else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
2134 		/*
2135 		 * Not good, NOT GOOD!
2136 		 * It means that mirror was started on stale disks
2137 		 * and more fresh disk just arrive.
2138 		 * If there were writes, mirror is broken, sorry.
2139 		 * I think the best choice here is don't touch
2140 		 * this disk and inform the user loudly.
2141 		 */
2142 		G_MIRROR_DEBUG(0, "Device %s was started before the freshest "
2143 		    "disk (%s) arrives!! It will not be connected to the "
2144 		    "running device.", sc->sc_name,
2145 		    g_mirror_get_diskname(disk));
2146 		g_mirror_destroy_disk(disk);
2147 		state = G_MIRROR_DISK_STATE_NONE;
2148 		/* Return immediately, because disk was destroyed. */
2149 		return (state);
2150 	}
2151 	G_MIRROR_DEBUG(3, "State for %s disk: %s.",
2152 	    g_mirror_get_diskname(disk), g_mirror_disk_state2str(state));
2153 	return (state);
2154 }
2155 
2156 /*
2157  * Update device state.
2158  */
2159 static void
2160 g_mirror_update_device(struct g_mirror_softc *sc, boolean_t force)
2161 {
2162 	struct g_mirror_disk *disk;
2163 	u_int state;
2164 
2165 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2166 
2167 	switch (sc->sc_state) {
2168 	case G_MIRROR_DEVICE_STATE_STARTING:
2169 	    {
2170 		struct g_mirror_disk *pdisk, *tdisk;
2171 		u_int dirty, ndisks, genid, syncid;
2172 
2173 		KASSERT(sc->sc_provider == NULL,
2174 		    ("Non-NULL provider in STARTING state (%s).", sc->sc_name));
2175 		/*
2176 		 * Are we ready? We are, if all disks are connected or
2177 		 * if we have any disks and 'force' is true.
2178 		 */
2179 		ndisks = g_mirror_ndisks(sc, -1);
2180 		if (sc->sc_ndisks == ndisks || (force && ndisks > 0)) {
2181 			;
2182 		} else if (ndisks == 0) {
2183 			/*
2184 			 * Disks went down in starting phase, so destroy
2185 			 * device.
2186 			 */
2187 			callout_drain(&sc->sc_callout);
2188 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2189 			G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2190 			    sc->sc_rootmount);
2191 			root_mount_rel(sc->sc_rootmount);
2192 			sc->sc_rootmount = NULL;
2193 			return;
2194 		} else {
2195 			return;
2196 		}
2197 
2198 		/*
2199 		 * Activate all disks with the biggest syncid.
2200 		 */
2201 		if (force) {
2202 			/*
2203 			 * If 'force' is true, we have been called due to
2204 			 * timeout, so don't bother canceling timeout.
2205 			 */
2206 			ndisks = 0;
2207 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2208 				if ((disk->d_flags &
2209 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2210 					ndisks++;
2211 				}
2212 			}
2213 			if (ndisks == 0) {
2214 				/* No valid disks found, destroy device. */
2215 				sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2216 				G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2217 				    __LINE__, sc->sc_rootmount);
2218 				root_mount_rel(sc->sc_rootmount);
2219 				sc->sc_rootmount = NULL;
2220 				return;
2221 			}
2222 		} else {
2223 			/* Cancel timeout. */
2224 			callout_drain(&sc->sc_callout);
2225 		}
2226 
2227 		/*
2228 		 * Find the biggest genid.
2229 		 */
2230 		genid = 0;
2231 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2232 			if (disk->d_genid > genid)
2233 				genid = disk->d_genid;
2234 		}
2235 		sc->sc_genid = genid;
2236 		/*
2237 		 * Remove all disks without the biggest genid.
2238 		 */
2239 		LIST_FOREACH_SAFE(disk, &sc->sc_disks, d_next, tdisk) {
2240 			if (disk->d_genid < genid) {
2241 				G_MIRROR_DEBUG(0,
2242 				    "Component %s (device %s) broken, skipping.",
2243 				    g_mirror_get_diskname(disk), sc->sc_name);
2244 				g_mirror_destroy_disk(disk);
2245 			}
2246 		}
2247 
2248 		/*
2249 		 * Find the biggest syncid.
2250 		 */
2251 		syncid = 0;
2252 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2253 			if (disk->d_sync.ds_syncid > syncid)
2254 				syncid = disk->d_sync.ds_syncid;
2255 		}
2256 
2257 		/*
2258 		 * Here we need to look for dirty disks and if all disks
2259 		 * with the biggest syncid are dirty, we have to choose
2260 		 * one with the biggest priority and rebuild the rest.
2261 		 */
2262 		/*
2263 		 * Find the number of dirty disks with the biggest syncid.
2264 		 * Find the number of disks with the biggest syncid.
2265 		 * While here, find a disk with the biggest priority.
2266 		 */
2267 		dirty = ndisks = 0;
2268 		pdisk = NULL;
2269 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2270 			if (disk->d_sync.ds_syncid != syncid)
2271 				continue;
2272 			if ((disk->d_flags &
2273 			    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2274 				continue;
2275 			}
2276 			ndisks++;
2277 			if ((disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
2278 				dirty++;
2279 				if (pdisk == NULL ||
2280 				    pdisk->d_priority < disk->d_priority) {
2281 					pdisk = disk;
2282 				}
2283 			}
2284 		}
2285 		if (dirty == 0) {
2286 			/* No dirty disks at all, great. */
2287 		} else if (dirty == ndisks) {
2288 			/*
2289 			 * Force synchronization for all dirty disks except one
2290 			 * with the biggest priority.
2291 			 */
2292 			KASSERT(pdisk != NULL, ("pdisk == NULL"));
2293 			G_MIRROR_DEBUG(1, "Using disk %s (device %s) as a "
2294 			    "master disk for synchronization.",
2295 			    g_mirror_get_diskname(pdisk), sc->sc_name);
2296 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2297 				if (disk->d_sync.ds_syncid != syncid)
2298 					continue;
2299 				if ((disk->d_flags &
2300 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2301 					continue;
2302 				}
2303 				KASSERT((disk->d_flags &
2304 				    G_MIRROR_DISK_FLAG_DIRTY) != 0,
2305 				    ("Disk %s isn't marked as dirty.",
2306 				    g_mirror_get_diskname(disk)));
2307 				/* Skip the disk with the biggest priority. */
2308 				if (disk == pdisk)
2309 					continue;
2310 				disk->d_sync.ds_syncid = 0;
2311 			}
2312 		} else if (dirty < ndisks) {
2313 			/*
2314 			 * Force synchronization for all dirty disks.
2315 			 * We have some non-dirty disks.
2316 			 */
2317 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2318 				if (disk->d_sync.ds_syncid != syncid)
2319 					continue;
2320 				if ((disk->d_flags &
2321 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2322 					continue;
2323 				}
2324 				if ((disk->d_flags &
2325 				    G_MIRROR_DISK_FLAG_DIRTY) == 0) {
2326 					continue;
2327 				}
2328 				disk->d_sync.ds_syncid = 0;
2329 			}
2330 		}
2331 
2332 		/* Reset hint. */
2333 		sc->sc_hint = NULL;
2334 		sc->sc_syncid = syncid;
2335 		if (force) {
2336 			/* Remember to bump syncid on first write. */
2337 			sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2338 		}
2339 		state = G_MIRROR_DEVICE_STATE_RUNNING;
2340 		G_MIRROR_DEBUG(1, "Device %s state changed from %s to %s.",
2341 		    sc->sc_name, g_mirror_device_state2str(sc->sc_state),
2342 		    g_mirror_device_state2str(state));
2343 		sc->sc_state = state;
2344 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2345 			state = g_mirror_determine_state(disk);
2346 			g_mirror_event_send(disk, state,
2347 			    G_MIRROR_EVENT_DONTWAIT);
2348 			if (state == G_MIRROR_DISK_STATE_STALE)
2349 				sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2350 		}
2351 		break;
2352 	    }
2353 	case G_MIRROR_DEVICE_STATE_RUNNING:
2354 		if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 0 &&
2355 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2356 			/*
2357 			 * No active disks or no disks at all,
2358 			 * so destroy device.
2359 			 */
2360 			if (sc->sc_provider != NULL)
2361 				g_mirror_destroy_provider(sc);
2362 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2363 			break;
2364 		} else if (g_mirror_ndisks(sc,
2365 		    G_MIRROR_DISK_STATE_ACTIVE) > 0 &&
2366 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2367 			/*
2368 			 * We have active disks, launch provider if it doesn't
2369 			 * exist.
2370 			 */
2371 			if (sc->sc_provider == NULL)
2372 				g_mirror_launch_provider(sc);
2373 			if (sc->sc_rootmount != NULL) {
2374 				G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2375 				    __LINE__, sc->sc_rootmount);
2376 				root_mount_rel(sc->sc_rootmount);
2377 				sc->sc_rootmount = NULL;
2378 			}
2379 		}
2380 		/*
2381 		 * Genid should be bumped immediately, so do it here.
2382 		 */
2383 		if ((sc->sc_bump_id & G_MIRROR_BUMP_GENID) != 0) {
2384 			sc->sc_bump_id &= ~G_MIRROR_BUMP_GENID;
2385 			g_mirror_bump_genid(sc);
2386 		}
2387 		break;
2388 	default:
2389 		KASSERT(1 == 0, ("Wrong device state (%s, %s).",
2390 		    sc->sc_name, g_mirror_device_state2str(sc->sc_state)));
2391 		break;
2392 	}
2393 }
2394 
2395 /*
2396  * Update disk state and device state if needed.
2397  */
2398 #define	DISK_STATE_CHANGED()	G_MIRROR_DEBUG(1,			\
2399 	"Disk %s state changed from %s to %s (device %s).",		\
2400 	g_mirror_get_diskname(disk),					\
2401 	g_mirror_disk_state2str(disk->d_state),				\
2402 	g_mirror_disk_state2str(state), sc->sc_name)
2403 static int
2404 g_mirror_update_disk(struct g_mirror_disk *disk, u_int state)
2405 {
2406 	struct g_mirror_softc *sc;
2407 
2408 	sc = disk->d_softc;
2409 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2410 
2411 again:
2412 	G_MIRROR_DEBUG(3, "Changing disk %s state from %s to %s.",
2413 	    g_mirror_get_diskname(disk), g_mirror_disk_state2str(disk->d_state),
2414 	    g_mirror_disk_state2str(state));
2415 	switch (state) {
2416 	case G_MIRROR_DISK_STATE_NEW:
2417 		/*
2418 		 * Possible scenarios:
2419 		 * 1. New disk arrive.
2420 		 */
2421 		/* Previous state should be NONE. */
2422 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NONE,
2423 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2424 		    g_mirror_disk_state2str(disk->d_state)));
2425 		DISK_STATE_CHANGED();
2426 
2427 		disk->d_state = state;
2428 		if (LIST_EMPTY(&sc->sc_disks))
2429 			LIST_INSERT_HEAD(&sc->sc_disks, disk, d_next);
2430 		else {
2431 			struct g_mirror_disk *dp;
2432 
2433 			LIST_FOREACH(dp, &sc->sc_disks, d_next) {
2434 				if (disk->d_priority >= dp->d_priority) {
2435 					LIST_INSERT_BEFORE(dp, disk, d_next);
2436 					dp = NULL;
2437 					break;
2438 				}
2439 				if (LIST_NEXT(dp, d_next) == NULL)
2440 					break;
2441 			}
2442 			if (dp != NULL)
2443 				LIST_INSERT_AFTER(dp, disk, d_next);
2444 		}
2445 		G_MIRROR_DEBUG(1, "Device %s: provider %s detected.",
2446 		    sc->sc_name, g_mirror_get_diskname(disk));
2447 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
2448 			break;
2449 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2450 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2451 		    g_mirror_device_state2str(sc->sc_state),
2452 		    g_mirror_get_diskname(disk),
2453 		    g_mirror_disk_state2str(disk->d_state)));
2454 		state = g_mirror_determine_state(disk);
2455 		if (state != G_MIRROR_DISK_STATE_NONE)
2456 			goto again;
2457 		break;
2458 	case G_MIRROR_DISK_STATE_ACTIVE:
2459 		/*
2460 		 * Possible scenarios:
2461 		 * 1. New disk does not need synchronization.
2462 		 * 2. Synchronization process finished successfully.
2463 		 */
2464 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2465 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2466 		    g_mirror_device_state2str(sc->sc_state),
2467 		    g_mirror_get_diskname(disk),
2468 		    g_mirror_disk_state2str(disk->d_state)));
2469 		/* Previous state should be NEW or SYNCHRONIZING. */
2470 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW ||
2471 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2472 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2473 		    g_mirror_disk_state2str(disk->d_state)));
2474 		DISK_STATE_CHANGED();
2475 
2476 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
2477 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2478 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_FORCE_SYNC;
2479 			g_mirror_sync_stop(disk, 0);
2480 		}
2481 		disk->d_state = state;
2482 		disk->d_sync.ds_offset = 0;
2483 		disk->d_sync.ds_offset_done = 0;
2484 		g_mirror_update_idle(sc, disk);
2485 		g_mirror_update_metadata(disk);
2486 		G_MIRROR_DEBUG(1, "Device %s: provider %s activated.",
2487 		    sc->sc_name, g_mirror_get_diskname(disk));
2488 		break;
2489 	case G_MIRROR_DISK_STATE_STALE:
2490 		/*
2491 		 * Possible scenarios:
2492 		 * 1. Stale disk was connected.
2493 		 */
2494 		/* Previous state should be NEW. */
2495 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2496 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2497 		    g_mirror_disk_state2str(disk->d_state)));
2498 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2499 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2500 		    g_mirror_device_state2str(sc->sc_state),
2501 		    g_mirror_get_diskname(disk),
2502 		    g_mirror_disk_state2str(disk->d_state)));
2503 		/*
2504 		 * STALE state is only possible if device is marked
2505 		 * NOAUTOSYNC.
2506 		 */
2507 		KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) != 0,
2508 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2509 		    g_mirror_device_state2str(sc->sc_state),
2510 		    g_mirror_get_diskname(disk),
2511 		    g_mirror_disk_state2str(disk->d_state)));
2512 		DISK_STATE_CHANGED();
2513 
2514 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2515 		disk->d_state = state;
2516 		g_mirror_update_metadata(disk);
2517 		G_MIRROR_DEBUG(0, "Device %s: provider %s is stale.",
2518 		    sc->sc_name, g_mirror_get_diskname(disk));
2519 		break;
2520 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
2521 		/*
2522 		 * Possible scenarios:
2523 		 * 1. Disk which needs synchronization was connected.
2524 		 */
2525 		/* Previous state should be NEW. */
2526 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2527 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2528 		    g_mirror_disk_state2str(disk->d_state)));
2529 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2530 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2531 		    g_mirror_device_state2str(sc->sc_state),
2532 		    g_mirror_get_diskname(disk),
2533 		    g_mirror_disk_state2str(disk->d_state)));
2534 		DISK_STATE_CHANGED();
2535 
2536 		if (disk->d_state == G_MIRROR_DISK_STATE_NEW)
2537 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2538 		disk->d_state = state;
2539 		if (sc->sc_provider != NULL) {
2540 			g_mirror_sync_start(disk);
2541 			g_mirror_update_metadata(disk);
2542 		}
2543 		break;
2544 	case G_MIRROR_DISK_STATE_DISCONNECTED:
2545 		/*
2546 		 * Possible scenarios:
2547 		 * 1. Device wasn't running yet, but disk disappear.
2548 		 * 2. Disk was active and disapppear.
2549 		 * 3. Disk disappear during synchronization process.
2550 		 */
2551 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING) {
2552 			/*
2553 			 * Previous state should be ACTIVE, STALE or
2554 			 * SYNCHRONIZING.
2555 			 */
2556 			KASSERT(disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
2557 			    disk->d_state == G_MIRROR_DISK_STATE_STALE ||
2558 			    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2559 			    ("Wrong disk state (%s, %s).",
2560 			    g_mirror_get_diskname(disk),
2561 			    g_mirror_disk_state2str(disk->d_state)));
2562 		} else if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING) {
2563 			/* Previous state should be NEW. */
2564 			KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2565 			    ("Wrong disk state (%s, %s).",
2566 			    g_mirror_get_diskname(disk),
2567 			    g_mirror_disk_state2str(disk->d_state)));
2568 			/*
2569 			 * Reset bumping syncid if disk disappeared in STARTING
2570 			 * state.
2571 			 */
2572 			if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0)
2573 				sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
2574 #ifdef	INVARIANTS
2575 		} else {
2576 			KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
2577 			    sc->sc_name,
2578 			    g_mirror_device_state2str(sc->sc_state),
2579 			    g_mirror_get_diskname(disk),
2580 			    g_mirror_disk_state2str(disk->d_state)));
2581 #endif
2582 		}
2583 		DISK_STATE_CHANGED();
2584 		G_MIRROR_DEBUG(0, "Device %s: provider %s disconnected.",
2585 		    sc->sc_name, g_mirror_get_diskname(disk));
2586 
2587 		g_mirror_destroy_disk(disk);
2588 		break;
2589 	case G_MIRROR_DISK_STATE_DESTROY:
2590 	    {
2591 		int error;
2592 
2593 		error = g_mirror_clear_metadata(disk);
2594 		if (error != 0)
2595 			return (error);
2596 		DISK_STATE_CHANGED();
2597 		G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.",
2598 		    sc->sc_name, g_mirror_get_diskname(disk));
2599 
2600 		g_mirror_destroy_disk(disk);
2601 		sc->sc_ndisks--;
2602 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2603 			g_mirror_update_metadata(disk);
2604 		}
2605 		break;
2606 	    }
2607 	default:
2608 		KASSERT(1 == 0, ("Unknown state (%u).", state));
2609 		break;
2610 	}
2611 	return (0);
2612 }
2613 #undef	DISK_STATE_CHANGED
2614 
2615 int
2616 g_mirror_read_metadata(struct g_consumer *cp, struct g_mirror_metadata *md)
2617 {
2618 	struct g_provider *pp;
2619 	u_char *buf;
2620 	int error;
2621 
2622 	g_topology_assert();
2623 
2624 	error = g_access(cp, 1, 0, 0);
2625 	if (error != 0)
2626 		return (error);
2627 	pp = cp->provider;
2628 	g_topology_unlock();
2629 	/* Metadata are stored on last sector. */
2630 	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
2631 	    &error);
2632 	g_topology_lock();
2633 	g_access(cp, -1, 0, 0);
2634 	if (buf == NULL) {
2635 		G_MIRROR_DEBUG(1, "Cannot read metadata from %s (error=%d).",
2636 		    cp->provider->name, error);
2637 		return (error);
2638 	}
2639 
2640 	/* Decode metadata. */
2641 	error = mirror_metadata_decode(buf, md);
2642 	g_free(buf);
2643 	if (strcmp(md->md_magic, G_MIRROR_MAGIC) != 0)
2644 		return (EINVAL);
2645 	if (md->md_version > G_MIRROR_VERSION) {
2646 		G_MIRROR_DEBUG(0,
2647 		    "Kernel module is too old to handle metadata from %s.",
2648 		    cp->provider->name);
2649 		return (EINVAL);
2650 	}
2651 	if (error != 0) {
2652 		G_MIRROR_DEBUG(1, "MD5 metadata hash mismatch for provider %s.",
2653 		    cp->provider->name);
2654 		return (error);
2655 	}
2656 
2657 	return (0);
2658 }
2659 
2660 static int
2661 g_mirror_check_metadata(struct g_mirror_softc *sc, struct g_provider *pp,
2662     struct g_mirror_metadata *md)
2663 {
2664 
2665 	if (g_mirror_id2disk(sc, md->md_did) != NULL) {
2666 		G_MIRROR_DEBUG(1, "Disk %s (id=%u) already exists, skipping.",
2667 		    pp->name, md->md_did);
2668 		return (EEXIST);
2669 	}
2670 	if (md->md_all != sc->sc_ndisks) {
2671 		G_MIRROR_DEBUG(1,
2672 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2673 		    "md_all", pp->name, sc->sc_name);
2674 		return (EINVAL);
2675 	}
2676 	if (md->md_slice != sc->sc_slice) {
2677 		G_MIRROR_DEBUG(1,
2678 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2679 		    "md_slice", pp->name, sc->sc_name);
2680 		return (EINVAL);
2681 	}
2682 	if (md->md_balance != sc->sc_balance) {
2683 		G_MIRROR_DEBUG(1,
2684 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2685 		    "md_balance", pp->name, sc->sc_name);
2686 		return (EINVAL);
2687 	}
2688 	if (md->md_mediasize != sc->sc_mediasize) {
2689 		G_MIRROR_DEBUG(1,
2690 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2691 		    "md_mediasize", pp->name, sc->sc_name);
2692 		return (EINVAL);
2693 	}
2694 	if (sc->sc_mediasize > pp->mediasize) {
2695 		G_MIRROR_DEBUG(1,
2696 		    "Invalid size of disk %s (device %s), skipping.", pp->name,
2697 		    sc->sc_name);
2698 		return (EINVAL);
2699 	}
2700 	if (md->md_sectorsize != sc->sc_sectorsize) {
2701 		G_MIRROR_DEBUG(1,
2702 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2703 		    "md_sectorsize", pp->name, sc->sc_name);
2704 		return (EINVAL);
2705 	}
2706 	if ((sc->sc_sectorsize % pp->sectorsize) != 0) {
2707 		G_MIRROR_DEBUG(1,
2708 		    "Invalid sector size of disk %s (device %s), skipping.",
2709 		    pp->name, sc->sc_name);
2710 		return (EINVAL);
2711 	}
2712 	if ((md->md_mflags & ~G_MIRROR_DEVICE_FLAG_MASK) != 0) {
2713 		G_MIRROR_DEBUG(1,
2714 		    "Invalid device flags on disk %s (device %s), skipping.",
2715 		    pp->name, sc->sc_name);
2716 		return (EINVAL);
2717 	}
2718 	if ((md->md_dflags & ~G_MIRROR_DISK_FLAG_MASK) != 0) {
2719 		G_MIRROR_DEBUG(1,
2720 		    "Invalid disk flags on disk %s (device %s), skipping.",
2721 		    pp->name, sc->sc_name);
2722 		return (EINVAL);
2723 	}
2724 	return (0);
2725 }
2726 
2727 int
2728 g_mirror_add_disk(struct g_mirror_softc *sc, struct g_provider *pp,
2729     struct g_mirror_metadata *md)
2730 {
2731 	struct g_mirror_disk *disk;
2732 	int error;
2733 
2734 	g_topology_assert_not();
2735 	G_MIRROR_DEBUG(2, "Adding disk %s.", pp->name);
2736 
2737 	error = g_mirror_check_metadata(sc, pp, md);
2738 	if (error != 0)
2739 		return (error);
2740 	if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING &&
2741 	    md->md_genid < sc->sc_genid) {
2742 		G_MIRROR_DEBUG(0, "Component %s (device %s) broken, skipping.",
2743 		    pp->name, sc->sc_name);
2744 		return (EINVAL);
2745 	}
2746 	disk = g_mirror_init_disk(sc, pp, md, &error);
2747 	if (disk == NULL)
2748 		return (error);
2749 	error = g_mirror_event_send(disk, G_MIRROR_DISK_STATE_NEW,
2750 	    G_MIRROR_EVENT_WAIT);
2751 	if (error != 0)
2752 		return (error);
2753 	if (md->md_version < G_MIRROR_VERSION) {
2754 		G_MIRROR_DEBUG(0, "Upgrading metadata on %s (v%d->v%d).",
2755 		    pp->name, md->md_version, G_MIRROR_VERSION);
2756 		g_mirror_update_metadata(disk);
2757 	}
2758 	return (0);
2759 }
2760 
2761 static void
2762 g_mirror_destroy_delayed(void *arg, int flag)
2763 {
2764 	struct g_mirror_softc *sc;
2765 	int error;
2766 
2767 	if (flag == EV_CANCEL) {
2768 		G_MIRROR_DEBUG(1, "Destroying canceled.");
2769 		return;
2770 	}
2771 	sc = arg;
2772 	g_topology_unlock();
2773 	sx_xlock(&sc->sc_lock);
2774 	KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) == 0,
2775 	    ("DESTROY flag set on %s.", sc->sc_name));
2776 	KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROYING) != 0,
2777 	    ("DESTROYING flag not set on %s.", sc->sc_name));
2778 	G_MIRROR_DEBUG(1, "Destroying %s (delayed).", sc->sc_name);
2779 	error = g_mirror_destroy(sc, G_MIRROR_DESTROY_SOFT);
2780 	if (error != 0) {
2781 		G_MIRROR_DEBUG(0, "Cannot destroy %s.", sc->sc_name);
2782 		sx_xunlock(&sc->sc_lock);
2783 	}
2784 	g_topology_lock();
2785 }
2786 
2787 static int
2788 g_mirror_access(struct g_provider *pp, int acr, int acw, int ace)
2789 {
2790 	struct g_mirror_softc *sc;
2791 	int dcr, dcw, dce, error = 0;
2792 
2793 	g_topology_assert();
2794 	G_MIRROR_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, acr,
2795 	    acw, ace);
2796 
2797 	sc = pp->geom->softc;
2798 	if (sc == NULL && acr <= 0 && acw <= 0 && ace <= 0)
2799 		return (0);
2800 	KASSERT(sc != NULL, ("NULL softc (provider=%s).", pp->name));
2801 
2802 	dcr = pp->acr + acr;
2803 	dcw = pp->acw + acw;
2804 	dce = pp->ace + ace;
2805 
2806 	g_topology_unlock();
2807 	sx_xlock(&sc->sc_lock);
2808 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0 ||
2809 	    LIST_EMPTY(&sc->sc_disks)) {
2810 		if (acr > 0 || acw > 0 || ace > 0)
2811 			error = ENXIO;
2812 		goto end;
2813 	}
2814 	if (dcw == 0 && !sc->sc_idle)
2815 		g_mirror_idle(sc, dcw);
2816 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROYING) != 0) {
2817 		if (acr > 0 || acw > 0 || ace > 0) {
2818 			error = ENXIO;
2819 			goto end;
2820 		}
2821 		if (dcr == 0 && dcw == 0 && dce == 0) {
2822 			g_post_event(g_mirror_destroy_delayed, sc, M_WAITOK,
2823 			    sc, NULL);
2824 		}
2825 	}
2826 end:
2827 	sx_xunlock(&sc->sc_lock);
2828 	g_topology_lock();
2829 	return (error);
2830 }
2831 
2832 static struct g_geom *
2833 g_mirror_create(struct g_class *mp, const struct g_mirror_metadata *md)
2834 {
2835 	struct g_mirror_softc *sc;
2836 	struct g_geom *gp;
2837 	int error, timeout;
2838 
2839 	g_topology_assert();
2840 	G_MIRROR_DEBUG(1, "Creating device %s (id=%u).", md->md_name,
2841 	    md->md_mid);
2842 
2843 	/* One disk is minimum. */
2844 	if (md->md_all < 1)
2845 		return (NULL);
2846 	/*
2847 	 * Action geom.
2848 	 */
2849 	gp = g_new_geomf(mp, "%s", md->md_name);
2850 	sc = malloc(sizeof(*sc), M_MIRROR, M_WAITOK | M_ZERO);
2851 	gp->start = g_mirror_start;
2852 	gp->orphan = g_mirror_orphan;
2853 	gp->access = g_mirror_access;
2854 	gp->dumpconf = g_mirror_dumpconf;
2855 
2856 	sc->sc_id = md->md_mid;
2857 	sc->sc_slice = md->md_slice;
2858 	sc->sc_balance = md->md_balance;
2859 	sc->sc_mediasize = md->md_mediasize;
2860 	sc->sc_sectorsize = md->md_sectorsize;
2861 	sc->sc_ndisks = md->md_all;
2862 	sc->sc_flags = md->md_mflags;
2863 	sc->sc_bump_id = 0;
2864 	sc->sc_idle = 1;
2865 	sc->sc_last_write = time_uptime;
2866 	sc->sc_writes = 0;
2867 	sx_init(&sc->sc_lock, "gmirror:lock");
2868 	bioq_init(&sc->sc_queue);
2869 	mtx_init(&sc->sc_queue_mtx, "gmirror:queue", NULL, MTX_DEF);
2870 	bioq_init(&sc->sc_regular_delayed);
2871 	bioq_init(&sc->sc_inflight);
2872 	bioq_init(&sc->sc_sync_delayed);
2873 	LIST_INIT(&sc->sc_disks);
2874 	TAILQ_INIT(&sc->sc_events);
2875 	mtx_init(&sc->sc_events_mtx, "gmirror:events", NULL, MTX_DEF);
2876 	callout_init(&sc->sc_callout, CALLOUT_MPSAFE);
2877 	sc->sc_state = G_MIRROR_DEVICE_STATE_STARTING;
2878 	gp->softc = sc;
2879 	sc->sc_geom = gp;
2880 	sc->sc_provider = NULL;
2881 	/*
2882 	 * Synchronization geom.
2883 	 */
2884 	gp = g_new_geomf(mp, "%s.sync", md->md_name);
2885 	gp->softc = sc;
2886 	gp->orphan = g_mirror_orphan;
2887 	sc->sc_sync.ds_geom = gp;
2888 	sc->sc_sync.ds_ndisks = 0;
2889 	error = kproc_create(g_mirror_worker, sc, &sc->sc_worker, 0, 0,
2890 	    "g_mirror %s", md->md_name);
2891 	if (error != 0) {
2892 		G_MIRROR_DEBUG(1, "Cannot create kernel thread for %s.",
2893 		    sc->sc_name);
2894 		g_destroy_geom(sc->sc_sync.ds_geom);
2895 		mtx_destroy(&sc->sc_events_mtx);
2896 		mtx_destroy(&sc->sc_queue_mtx);
2897 		sx_destroy(&sc->sc_lock);
2898 		g_destroy_geom(sc->sc_geom);
2899 		free(sc, M_MIRROR);
2900 		return (NULL);
2901 	}
2902 
2903 	G_MIRROR_DEBUG(1, "Device %s created (%u components, id=%u).",
2904 	    sc->sc_name, sc->sc_ndisks, sc->sc_id);
2905 
2906 	sc->sc_rootmount = root_mount_hold("GMIRROR");
2907 	G_MIRROR_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
2908 	/*
2909 	 * Run timeout.
2910 	 */
2911 	timeout = g_mirror_timeout * hz;
2912 	callout_reset(&sc->sc_callout, timeout, g_mirror_go, sc);
2913 	return (sc->sc_geom);
2914 }
2915 
2916 int
2917 g_mirror_destroy(struct g_mirror_softc *sc, int how)
2918 {
2919 	struct g_mirror_disk *disk;
2920 	struct g_provider *pp;
2921 
2922 	g_topology_assert_not();
2923 	if (sc == NULL)
2924 		return (ENXIO);
2925 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2926 
2927 	pp = sc->sc_provider;
2928 	if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
2929 		switch (how) {
2930 		case G_MIRROR_DESTROY_SOFT:
2931 			G_MIRROR_DEBUG(1,
2932 			    "Device %s is still open (r%dw%de%d).", pp->name,
2933 			    pp->acr, pp->acw, pp->ace);
2934 			return (EBUSY);
2935 		case G_MIRROR_DESTROY_DELAYED:
2936 			G_MIRROR_DEBUG(1,
2937 			    "Device %s will be destroyed on last close.",
2938 			    pp->name);
2939 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2940 				if (disk->d_state ==
2941 				    G_MIRROR_DISK_STATE_SYNCHRONIZING) {
2942 					g_mirror_sync_stop(disk, 1);
2943 				}
2944 			}
2945 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROYING;
2946 			return (EBUSY);
2947 		case G_MIRROR_DESTROY_HARD:
2948 			G_MIRROR_DEBUG(1, "Device %s is still open, so it "
2949 			    "can't be definitely removed.", pp->name);
2950 		}
2951 	}
2952 
2953 	g_topology_lock();
2954 	if (sc->sc_geom->softc == NULL) {
2955 		g_topology_unlock();
2956 		return (0);
2957 	}
2958 	sc->sc_geom->softc = NULL;
2959 	sc->sc_sync.ds_geom->softc = NULL;
2960 	g_topology_unlock();
2961 
2962 	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2963 	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_WAIT;
2964 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
2965 	sx_xunlock(&sc->sc_lock);
2966 	mtx_lock(&sc->sc_queue_mtx);
2967 	wakeup(sc);
2968 	mtx_unlock(&sc->sc_queue_mtx);
2969 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
2970 	while (sc->sc_worker != NULL)
2971 		tsleep(&sc->sc_worker, PRIBIO, "m:destroy", hz / 5);
2972 	G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
2973 	sx_xlock(&sc->sc_lock);
2974 	g_mirror_destroy_device(sc);
2975 	free(sc, M_MIRROR);
2976 	return (0);
2977 }
2978 
2979 static void
2980 g_mirror_taste_orphan(struct g_consumer *cp)
2981 {
2982 
2983 	KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
2984 	    cp->provider->name));
2985 }
2986 
2987 static struct g_geom *
2988 g_mirror_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
2989 {
2990 	struct g_mirror_metadata md;
2991 	struct g_mirror_softc *sc;
2992 	struct g_consumer *cp;
2993 	struct g_geom *gp;
2994 	int error;
2995 
2996 	g_topology_assert();
2997 	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
2998 	G_MIRROR_DEBUG(2, "Tasting %s.", pp->name);
2999 
3000 	gp = g_new_geomf(mp, "mirror:taste");
3001 	/*
3002 	 * This orphan function should be never called.
3003 	 */
3004 	gp->orphan = g_mirror_taste_orphan;
3005 	cp = g_new_consumer(gp);
3006 	g_attach(cp, pp);
3007 	error = g_mirror_read_metadata(cp, &md);
3008 	g_detach(cp);
3009 	g_destroy_consumer(cp);
3010 	g_destroy_geom(gp);
3011 	if (error != 0)
3012 		return (NULL);
3013 	gp = NULL;
3014 
3015 	if (md.md_provider[0] != '\0' && strcmp(md.md_provider, pp->name) != 0)
3016 		return (NULL);
3017 	if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
3018 		return (NULL);
3019 	if ((md.md_dflags & G_MIRROR_DISK_FLAG_INACTIVE) != 0) {
3020 		G_MIRROR_DEBUG(0,
3021 		    "Device %s: provider %s marked as inactive, skipping.",
3022 		    md.md_name, pp->name);
3023 		return (NULL);
3024 	}
3025 	if (g_mirror_debug >= 2)
3026 		mirror_metadata_dump(&md);
3027 
3028 	/*
3029 	 * Let's check if device already exists.
3030 	 */
3031 	sc = NULL;
3032 	LIST_FOREACH(gp, &mp->geom, geom) {
3033 		sc = gp->softc;
3034 		if (sc == NULL)
3035 			continue;
3036 		if (sc->sc_sync.ds_geom == gp)
3037 			continue;
3038 		if (strcmp(md.md_name, sc->sc_name) != 0)
3039 			continue;
3040 		if (md.md_mid != sc->sc_id) {
3041 			G_MIRROR_DEBUG(0, "Device %s already configured.",
3042 			    sc->sc_name);
3043 			return (NULL);
3044 		}
3045 		break;
3046 	}
3047 	if (gp == NULL) {
3048 		gp = g_mirror_create(mp, &md);
3049 		if (gp == NULL) {
3050 			G_MIRROR_DEBUG(0, "Cannot create device %s.",
3051 			    md.md_name);
3052 			return (NULL);
3053 		}
3054 		sc = gp->softc;
3055 	}
3056 	G_MIRROR_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
3057 	g_topology_unlock();
3058 	sx_xlock(&sc->sc_lock);
3059 	error = g_mirror_add_disk(sc, pp, &md);
3060 	if (error != 0) {
3061 		G_MIRROR_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
3062 		    pp->name, gp->name, error);
3063 		if (LIST_EMPTY(&sc->sc_disks)) {
3064 			g_cancel_event(sc);
3065 			g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3066 			g_topology_lock();
3067 			return (NULL);
3068 		}
3069 		gp = NULL;
3070 	}
3071 	sx_xunlock(&sc->sc_lock);
3072 	g_topology_lock();
3073 	return (gp);
3074 }
3075 
3076 static int
3077 g_mirror_destroy_geom(struct gctl_req *req __unused,
3078     struct g_class *mp __unused, struct g_geom *gp)
3079 {
3080 	struct g_mirror_softc *sc;
3081 	int error;
3082 
3083 	g_topology_unlock();
3084 	sc = gp->softc;
3085 	sx_xlock(&sc->sc_lock);
3086 	g_cancel_event(sc);
3087 	error = g_mirror_destroy(gp->softc, G_MIRROR_DESTROY_SOFT);
3088 	if (error != 0)
3089 		sx_xunlock(&sc->sc_lock);
3090 	g_topology_lock();
3091 	return (error);
3092 }
3093 
3094 static void
3095 g_mirror_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
3096     struct g_consumer *cp, struct g_provider *pp)
3097 {
3098 	struct g_mirror_softc *sc;
3099 
3100 	g_topology_assert();
3101 
3102 	sc = gp->softc;
3103 	if (sc == NULL)
3104 		return;
3105 	/* Skip synchronization geom. */
3106 	if (gp == sc->sc_sync.ds_geom)
3107 		return;
3108 	if (pp != NULL) {
3109 		/* Nothing here. */
3110 	} else if (cp != NULL) {
3111 		struct g_mirror_disk *disk;
3112 
3113 		disk = cp->private;
3114 		if (disk == NULL)
3115 			return;
3116 		g_topology_unlock();
3117 		sx_xlock(&sc->sc_lock);
3118 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)disk->d_id);
3119 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
3120 			sbuf_printf(sb, "%s<Synchronized>", indent);
3121 			if (disk->d_sync.ds_offset == 0)
3122 				sbuf_printf(sb, "0%%");
3123 			else {
3124 				sbuf_printf(sb, "%u%%",
3125 				    (u_int)((disk->d_sync.ds_offset * 100) /
3126 				    sc->sc_provider->mediasize));
3127 			}
3128 			sbuf_printf(sb, "</Synchronized>\n");
3129 		}
3130 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent,
3131 		    disk->d_sync.ds_syncid);
3132 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent,
3133 		    disk->d_genid);
3134 		sbuf_printf(sb, "%s<Flags>", indent);
3135 		if (disk->d_flags == 0)
3136 			sbuf_printf(sb, "NONE");
3137 		else {
3138 			int first = 1;
3139 
3140 #define	ADD_FLAG(flag, name)	do {					\
3141 	if ((disk->d_flags & (flag)) != 0) {				\
3142 		if (!first)						\
3143 			sbuf_printf(sb, ", ");				\
3144 		else							\
3145 			first = 0;					\
3146 		sbuf_printf(sb, name);					\
3147 	}								\
3148 } while (0)
3149 			ADD_FLAG(G_MIRROR_DISK_FLAG_DIRTY, "DIRTY");
3150 			ADD_FLAG(G_MIRROR_DISK_FLAG_HARDCODED, "HARDCODED");
3151 			ADD_FLAG(G_MIRROR_DISK_FLAG_INACTIVE, "INACTIVE");
3152 			ADD_FLAG(G_MIRROR_DISK_FLAG_SYNCHRONIZING,
3153 			    "SYNCHRONIZING");
3154 			ADD_FLAG(G_MIRROR_DISK_FLAG_FORCE_SYNC, "FORCE_SYNC");
3155 			ADD_FLAG(G_MIRROR_DISK_FLAG_BROKEN, "BROKEN");
3156 #undef	ADD_FLAG
3157 		}
3158 		sbuf_printf(sb, "</Flags>\n");
3159 		sbuf_printf(sb, "%s<Priority>%u</Priority>\n", indent,
3160 		    disk->d_priority);
3161 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3162 		    g_mirror_disk_state2str(disk->d_state));
3163 		sx_xunlock(&sc->sc_lock);
3164 		g_topology_lock();
3165 	} else {
3166 		g_topology_unlock();
3167 		sx_xlock(&sc->sc_lock);
3168 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
3169 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent, sc->sc_syncid);
3170 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, sc->sc_genid);
3171 		sbuf_printf(sb, "%s<Flags>", indent);
3172 		if (sc->sc_flags == 0)
3173 			sbuf_printf(sb, "NONE");
3174 		else {
3175 			int first = 1;
3176 
3177 #define	ADD_FLAG(flag, name)	do {					\
3178 	if ((sc->sc_flags & (flag)) != 0) {				\
3179 		if (!first)						\
3180 			sbuf_printf(sb, ", ");				\
3181 		else							\
3182 			first = 0;					\
3183 		sbuf_printf(sb, name);					\
3184 	}								\
3185 } while (0)
3186 			ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOFAILSYNC, "NOFAILSYNC");
3187 			ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOAUTOSYNC, "NOAUTOSYNC");
3188 #undef	ADD_FLAG
3189 		}
3190 		sbuf_printf(sb, "</Flags>\n");
3191 		sbuf_printf(sb, "%s<Slice>%u</Slice>\n", indent,
3192 		    (u_int)sc->sc_slice);
3193 		sbuf_printf(sb, "%s<Balance>%s</Balance>\n", indent,
3194 		    balance_name(sc->sc_balance));
3195 		sbuf_printf(sb, "%s<Components>%u</Components>\n", indent,
3196 		    sc->sc_ndisks);
3197 		sbuf_printf(sb, "%s<State>", indent);
3198 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
3199 			sbuf_printf(sb, "%s", "STARTING");
3200 		else if (sc->sc_ndisks ==
3201 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE))
3202 			sbuf_printf(sb, "%s", "COMPLETE");
3203 		else
3204 			sbuf_printf(sb, "%s", "DEGRADED");
3205 		sbuf_printf(sb, "</State>\n");
3206 		sx_xunlock(&sc->sc_lock);
3207 		g_topology_lock();
3208 	}
3209 }
3210 
3211 static void
3212 g_mirror_shutdown_pre_sync(void *arg, int howto)
3213 {
3214 	struct g_class *mp;
3215 	struct g_geom *gp, *gp2;
3216 	struct g_mirror_softc *sc;
3217 	int error;
3218 
3219 	mp = arg;
3220 	DROP_GIANT();
3221 	g_topology_lock();
3222 	LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
3223 		if ((sc = gp->softc) == NULL)
3224 			continue;
3225 		/* Skip synchronization geom. */
3226 		if (gp == sc->sc_sync.ds_geom)
3227 			continue;
3228 		g_topology_unlock();
3229 		sx_xlock(&sc->sc_lock);
3230 		g_cancel_event(sc);
3231 		error = g_mirror_destroy(sc, G_MIRROR_DESTROY_DELAYED);
3232 		if (error != 0)
3233 			sx_xunlock(&sc->sc_lock);
3234 		g_topology_lock();
3235 	}
3236 	g_topology_unlock();
3237 	PICKUP_GIANT();
3238 }
3239 
3240 static void
3241 g_mirror_init(struct g_class *mp)
3242 {
3243 
3244 	g_mirror_pre_sync = EVENTHANDLER_REGISTER(shutdown_pre_sync,
3245 	    g_mirror_shutdown_pre_sync, mp, SHUTDOWN_PRI_FIRST);
3246 	if (g_mirror_pre_sync == NULL)
3247 		G_MIRROR_DEBUG(0, "Warning! Cannot register shutdown event.");
3248 }
3249 
3250 static void
3251 g_mirror_fini(struct g_class *mp)
3252 {
3253 
3254 	if (g_mirror_pre_sync != NULL)
3255 		EVENTHANDLER_DEREGISTER(shutdown_pre_sync, g_mirror_pre_sync);
3256 }
3257 
3258 DECLARE_GEOM_CLASS(g_mirror_class, g_mirror);
3259