xref: /freebsd/sys/geom/mirror/g_mirror.c (revision 884a2a699669ec61e2366e3e358342dbc94be24a)
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 FEATURE(geom_mirror, "GEOM mirroring support");
49 
50 static MALLOC_DEFINE(M_MIRROR, "mirror_data", "GEOM_MIRROR Data");
51 
52 SYSCTL_DECL(_kern_geom);
53 SYSCTL_NODE(_kern_geom, OID_AUTO, mirror, CTLFLAG_RW, 0, "GEOM_MIRROR stuff");
54 u_int g_mirror_debug = 0;
55 TUNABLE_INT("kern.geom.mirror.debug", &g_mirror_debug);
56 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, debug, CTLFLAG_RW, &g_mirror_debug, 0,
57     "Debug level");
58 static u_int g_mirror_timeout = 4;
59 TUNABLE_INT("kern.geom.mirror.timeout", &g_mirror_timeout);
60 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, timeout, CTLFLAG_RW, &g_mirror_timeout,
61     0, "Time to wait on all mirror components");
62 static u_int g_mirror_idletime = 5;
63 TUNABLE_INT("kern.geom.mirror.idletime", &g_mirror_idletime);
64 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, idletime, CTLFLAG_RW,
65     &g_mirror_idletime, 0, "Mark components as clean when idling");
66 static u_int g_mirror_disconnect_on_failure = 1;
67 TUNABLE_INT("kern.geom.mirror.disconnect_on_failure",
68     &g_mirror_disconnect_on_failure);
69 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, disconnect_on_failure, CTLFLAG_RW,
70     &g_mirror_disconnect_on_failure, 0, "Disconnect component on I/O failure.");
71 static u_int g_mirror_syncreqs = 2;
72 TUNABLE_INT("kern.geom.mirror.sync_requests", &g_mirror_syncreqs);
73 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, sync_requests, CTLFLAG_RDTUN,
74     &g_mirror_syncreqs, 0, "Parallel synchronization I/O requests.");
75 
76 #define	MSLEEP(ident, mtx, priority, wmesg, timeout)	do {		\
77 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, (ident));	\
78 	msleep((ident), (mtx), (priority), (wmesg), (timeout));		\
79 	G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, (ident));	\
80 } while (0)
81 
82 static eventhandler_tag g_mirror_pre_sync = NULL;
83 
84 static int g_mirror_destroy_geom(struct gctl_req *req, struct g_class *mp,
85     struct g_geom *gp);
86 static g_taste_t g_mirror_taste;
87 static void g_mirror_init(struct g_class *mp);
88 static void g_mirror_fini(struct g_class *mp);
89 
90 struct g_class g_mirror_class = {
91 	.name = G_MIRROR_CLASS_NAME,
92 	.version = G_VERSION,
93 	.ctlreq = g_mirror_config,
94 	.taste = g_mirror_taste,
95 	.destroy_geom = g_mirror_destroy_geom,
96 	.init = g_mirror_init,
97 	.fini = g_mirror_fini
98 };
99 
100 
101 static void g_mirror_destroy_provider(struct g_mirror_softc *sc);
102 static int g_mirror_update_disk(struct g_mirror_disk *disk, u_int state);
103 static void g_mirror_update_device(struct g_mirror_softc *sc, boolean_t force);
104 static void g_mirror_dumpconf(struct sbuf *sb, const char *indent,
105     struct g_geom *gp, struct g_consumer *cp, struct g_provider *pp);
106 static void g_mirror_sync_stop(struct g_mirror_disk *disk, int type);
107 static void g_mirror_register_request(struct bio *bp);
108 static void g_mirror_sync_release(struct g_mirror_softc *sc);
109 
110 
111 static const char *
112 g_mirror_disk_state2str(int state)
113 {
114 
115 	switch (state) {
116 	case G_MIRROR_DISK_STATE_NONE:
117 		return ("NONE");
118 	case G_MIRROR_DISK_STATE_NEW:
119 		return ("NEW");
120 	case G_MIRROR_DISK_STATE_ACTIVE:
121 		return ("ACTIVE");
122 	case G_MIRROR_DISK_STATE_STALE:
123 		return ("STALE");
124 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
125 		return ("SYNCHRONIZING");
126 	case G_MIRROR_DISK_STATE_DISCONNECTED:
127 		return ("DISCONNECTED");
128 	case G_MIRROR_DISK_STATE_DESTROY:
129 		return ("DESTROY");
130 	default:
131 		return ("INVALID");
132 	}
133 }
134 
135 static const char *
136 g_mirror_device_state2str(int state)
137 {
138 
139 	switch (state) {
140 	case G_MIRROR_DEVICE_STATE_STARTING:
141 		return ("STARTING");
142 	case G_MIRROR_DEVICE_STATE_RUNNING:
143 		return ("RUNNING");
144 	default:
145 		return ("INVALID");
146 	}
147 }
148 
149 static const char *
150 g_mirror_get_diskname(struct g_mirror_disk *disk)
151 {
152 
153 	if (disk->d_consumer == NULL || disk->d_consumer->provider == NULL)
154 		return ("[unknown]");
155 	return (disk->d_name);
156 }
157 
158 /*
159  * --- Events handling functions ---
160  * Events in geom_mirror are used to maintain disks and device status
161  * from one thread to simplify locking.
162  */
163 static void
164 g_mirror_event_free(struct g_mirror_event *ep)
165 {
166 
167 	free(ep, M_MIRROR);
168 }
169 
170 int
171 g_mirror_event_send(void *arg, int state, int flags)
172 {
173 	struct g_mirror_softc *sc;
174 	struct g_mirror_disk *disk;
175 	struct g_mirror_event *ep;
176 	int error;
177 
178 	ep = malloc(sizeof(*ep), M_MIRROR, M_WAITOK);
179 	G_MIRROR_DEBUG(4, "%s: Sending event %p.", __func__, ep);
180 	if ((flags & G_MIRROR_EVENT_DEVICE) != 0) {
181 		disk = NULL;
182 		sc = arg;
183 	} else {
184 		disk = arg;
185 		sc = disk->d_softc;
186 	}
187 	ep->e_disk = disk;
188 	ep->e_state = state;
189 	ep->e_flags = flags;
190 	ep->e_error = 0;
191 	mtx_lock(&sc->sc_events_mtx);
192 	TAILQ_INSERT_TAIL(&sc->sc_events, ep, e_next);
193 	mtx_unlock(&sc->sc_events_mtx);
194 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
195 	mtx_lock(&sc->sc_queue_mtx);
196 	wakeup(sc);
197 	mtx_unlock(&sc->sc_queue_mtx);
198 	if ((flags & G_MIRROR_EVENT_DONTWAIT) != 0)
199 		return (0);
200 	sx_assert(&sc->sc_lock, SX_XLOCKED);
201 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, ep);
202 	sx_xunlock(&sc->sc_lock);
203 	while ((ep->e_flags & G_MIRROR_EVENT_DONE) == 0) {
204 		mtx_lock(&sc->sc_events_mtx);
205 		MSLEEP(ep, &sc->sc_events_mtx, PRIBIO | PDROP, "m:event",
206 		    hz * 5);
207 	}
208 	error = ep->e_error;
209 	g_mirror_event_free(ep);
210 	sx_xlock(&sc->sc_lock);
211 	return (error);
212 }
213 
214 static struct g_mirror_event *
215 g_mirror_event_get(struct g_mirror_softc *sc)
216 {
217 	struct g_mirror_event *ep;
218 
219 	mtx_lock(&sc->sc_events_mtx);
220 	ep = TAILQ_FIRST(&sc->sc_events);
221 	mtx_unlock(&sc->sc_events_mtx);
222 	return (ep);
223 }
224 
225 static void
226 g_mirror_event_remove(struct g_mirror_softc *sc, struct g_mirror_event *ep)
227 {
228 
229 	mtx_lock(&sc->sc_events_mtx);
230 	TAILQ_REMOVE(&sc->sc_events, ep, e_next);
231 	mtx_unlock(&sc->sc_events_mtx);
232 }
233 
234 static void
235 g_mirror_event_cancel(struct g_mirror_disk *disk)
236 {
237 	struct g_mirror_softc *sc;
238 	struct g_mirror_event *ep, *tmpep;
239 
240 	sc = disk->d_softc;
241 	sx_assert(&sc->sc_lock, SX_XLOCKED);
242 
243 	mtx_lock(&sc->sc_events_mtx);
244 	TAILQ_FOREACH_SAFE(ep, &sc->sc_events, e_next, tmpep) {
245 		if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0)
246 			continue;
247 		if (ep->e_disk != disk)
248 			continue;
249 		TAILQ_REMOVE(&sc->sc_events, ep, e_next);
250 		if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
251 			g_mirror_event_free(ep);
252 		else {
253 			ep->e_error = ECANCELED;
254 			wakeup(ep);
255 		}
256 	}
257 	mtx_unlock(&sc->sc_events_mtx);
258 }
259 
260 /*
261  * Return the number of disks in given state.
262  * If state is equal to -1, count all connected disks.
263  */
264 u_int
265 g_mirror_ndisks(struct g_mirror_softc *sc, int state)
266 {
267 	struct g_mirror_disk *disk;
268 	u_int n = 0;
269 
270 	sx_assert(&sc->sc_lock, SX_LOCKED);
271 
272 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
273 		if (state == -1 || disk->d_state == state)
274 			n++;
275 	}
276 	return (n);
277 }
278 
279 /*
280  * Find a disk in mirror by its disk ID.
281  */
282 static struct g_mirror_disk *
283 g_mirror_id2disk(struct g_mirror_softc *sc, uint32_t id)
284 {
285 	struct g_mirror_disk *disk;
286 
287 	sx_assert(&sc->sc_lock, SX_XLOCKED);
288 
289 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
290 		if (disk->d_id == id)
291 			return (disk);
292 	}
293 	return (NULL);
294 }
295 
296 static u_int
297 g_mirror_nrequests(struct g_mirror_softc *sc, struct g_consumer *cp)
298 {
299 	struct bio *bp;
300 	u_int nreqs = 0;
301 
302 	mtx_lock(&sc->sc_queue_mtx);
303 	TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
304 		if (bp->bio_from == cp)
305 			nreqs++;
306 	}
307 	mtx_unlock(&sc->sc_queue_mtx);
308 	return (nreqs);
309 }
310 
311 static int
312 g_mirror_is_busy(struct g_mirror_softc *sc, struct g_consumer *cp)
313 {
314 
315 	if (cp->index > 0) {
316 		G_MIRROR_DEBUG(2,
317 		    "I/O requests for %s exist, can't destroy it now.",
318 		    cp->provider->name);
319 		return (1);
320 	}
321 	if (g_mirror_nrequests(sc, cp) > 0) {
322 		G_MIRROR_DEBUG(2,
323 		    "I/O requests for %s in queue, can't destroy it now.",
324 		    cp->provider->name);
325 		return (1);
326 	}
327 	return (0);
328 }
329 
330 static void
331 g_mirror_destroy_consumer(void *arg, int flags __unused)
332 {
333 	struct g_consumer *cp;
334 
335 	g_topology_assert();
336 
337 	cp = arg;
338 	G_MIRROR_DEBUG(1, "Consumer %s destroyed.", cp->provider->name);
339 	g_detach(cp);
340 	g_destroy_consumer(cp);
341 }
342 
343 static void
344 g_mirror_kill_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
345 {
346 	struct g_provider *pp;
347 	int retaste_wait;
348 
349 	g_topology_assert();
350 
351 	cp->private = NULL;
352 	if (g_mirror_is_busy(sc, cp))
353 		return;
354 	pp = cp->provider;
355 	retaste_wait = 0;
356 	if (cp->acw == 1) {
357 		if ((pp->geom->flags & G_GEOM_WITHER) == 0)
358 			retaste_wait = 1;
359 	}
360 	G_MIRROR_DEBUG(2, "Access %s r%dw%de%d = %d", pp->name, -cp->acr,
361 	    -cp->acw, -cp->ace, 0);
362 	if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
363 		g_access(cp, -cp->acr, -cp->acw, -cp->ace);
364 	if (retaste_wait) {
365 		/*
366 		 * After retaste event was send (inside g_access()), we can send
367 		 * event to detach and destroy consumer.
368 		 * A class, which has consumer to the given provider connected
369 		 * will not receive retaste event for the provider.
370 		 * This is the way how I ignore retaste events when I close
371 		 * consumers opened for write: I detach and destroy consumer
372 		 * after retaste event is sent.
373 		 */
374 		g_post_event(g_mirror_destroy_consumer, cp, M_WAITOK, NULL);
375 		return;
376 	}
377 	G_MIRROR_DEBUG(1, "Consumer %s destroyed.", pp->name);
378 	g_detach(cp);
379 	g_destroy_consumer(cp);
380 }
381 
382 static int
383 g_mirror_connect_disk(struct g_mirror_disk *disk, struct g_provider *pp)
384 {
385 	struct g_consumer *cp;
386 	int error;
387 
388 	g_topology_assert_not();
389 	KASSERT(disk->d_consumer == NULL,
390 	    ("Disk already connected (device %s).", disk->d_softc->sc_name));
391 
392 	g_topology_lock();
393 	cp = g_new_consumer(disk->d_softc->sc_geom);
394 	error = g_attach(cp, pp);
395 	if (error != 0) {
396 		g_destroy_consumer(cp);
397 		g_topology_unlock();
398 		return (error);
399 	}
400 	error = g_access(cp, 1, 1, 1);
401 	if (error != 0) {
402 		g_detach(cp);
403 		g_destroy_consumer(cp);
404 		g_topology_unlock();
405 		G_MIRROR_DEBUG(0, "Cannot open consumer %s (error=%d).",
406 		    pp->name, error);
407 		return (error);
408 	}
409 	g_topology_unlock();
410 	disk->d_consumer = cp;
411 	disk->d_consumer->private = disk;
412 	disk->d_consumer->index = 0;
413 
414 	G_MIRROR_DEBUG(2, "Disk %s connected.", g_mirror_get_diskname(disk));
415 	return (0);
416 }
417 
418 static void
419 g_mirror_disconnect_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
420 {
421 
422 	g_topology_assert();
423 
424 	if (cp == NULL)
425 		return;
426 	if (cp->provider != NULL)
427 		g_mirror_kill_consumer(sc, cp);
428 	else
429 		g_destroy_consumer(cp);
430 }
431 
432 /*
433  * Initialize disk. This means allocate memory, create consumer, attach it
434  * to the provider and open access (r1w1e1) to it.
435  */
436 static struct g_mirror_disk *
437 g_mirror_init_disk(struct g_mirror_softc *sc, struct g_provider *pp,
438     struct g_mirror_metadata *md, int *errorp)
439 {
440 	struct g_mirror_disk *disk;
441 	int error;
442 
443 	disk = malloc(sizeof(*disk), M_MIRROR, M_NOWAIT | M_ZERO);
444 	if (disk == NULL) {
445 		error = ENOMEM;
446 		goto fail;
447 	}
448 	disk->d_softc = sc;
449 	error = g_mirror_connect_disk(disk, pp);
450 	if (error != 0)
451 		goto fail;
452 	disk->d_id = md->md_did;
453 	disk->d_state = G_MIRROR_DISK_STATE_NONE;
454 	disk->d_priority = md->md_priority;
455 	disk->d_flags = md->md_dflags;
456 	if (md->md_provider[0] != '\0')
457 		disk->d_flags |= G_MIRROR_DISK_FLAG_HARDCODED;
458 	disk->d_sync.ds_consumer = NULL;
459 	disk->d_sync.ds_offset = md->md_sync_offset;
460 	disk->d_sync.ds_offset_done = md->md_sync_offset;
461 	disk->d_genid = md->md_genid;
462 	disk->d_sync.ds_syncid = md->md_syncid;
463 	if (errorp != NULL)
464 		*errorp = 0;
465 	return (disk);
466 fail:
467 	if (errorp != NULL)
468 		*errorp = error;
469 	if (disk != NULL)
470 		free(disk, M_MIRROR);
471 	return (NULL);
472 }
473 
474 static void
475 g_mirror_destroy_disk(struct g_mirror_disk *disk)
476 {
477 	struct g_mirror_softc *sc;
478 
479 	g_topology_assert_not();
480 	sc = disk->d_softc;
481 	sx_assert(&sc->sc_lock, SX_XLOCKED);
482 
483 	LIST_REMOVE(disk, d_next);
484 	g_mirror_event_cancel(disk);
485 	if (sc->sc_hint == disk)
486 		sc->sc_hint = NULL;
487 	switch (disk->d_state) {
488 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
489 		g_mirror_sync_stop(disk, 1);
490 		/* FALLTHROUGH */
491 	case G_MIRROR_DISK_STATE_NEW:
492 	case G_MIRROR_DISK_STATE_STALE:
493 	case G_MIRROR_DISK_STATE_ACTIVE:
494 		g_topology_lock();
495 		g_mirror_disconnect_consumer(sc, disk->d_consumer);
496 		g_topology_unlock();
497 		free(disk, M_MIRROR);
498 		break;
499 	default:
500 		KASSERT(0 == 1, ("Wrong disk state (%s, %s).",
501 		    g_mirror_get_diskname(disk),
502 		    g_mirror_disk_state2str(disk->d_state)));
503 	}
504 }
505 
506 static void
507 g_mirror_destroy_device(struct g_mirror_softc *sc)
508 {
509 	struct g_mirror_disk *disk;
510 	struct g_mirror_event *ep;
511 	struct g_geom *gp;
512 	struct g_consumer *cp, *tmpcp;
513 
514 	g_topology_assert_not();
515 	sx_assert(&sc->sc_lock, SX_XLOCKED);
516 
517 	gp = sc->sc_geom;
518 	if (sc->sc_provider != NULL)
519 		g_mirror_destroy_provider(sc);
520 	for (disk = LIST_FIRST(&sc->sc_disks); disk != NULL;
521 	    disk = LIST_FIRST(&sc->sc_disks)) {
522 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
523 		g_mirror_update_metadata(disk);
524 		g_mirror_destroy_disk(disk);
525 	}
526 	while ((ep = g_mirror_event_get(sc)) != NULL) {
527 		g_mirror_event_remove(sc, ep);
528 		if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
529 			g_mirror_event_free(ep);
530 		else {
531 			ep->e_error = ECANCELED;
532 			ep->e_flags |= G_MIRROR_EVENT_DONE;
533 			G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, ep);
534 			mtx_lock(&sc->sc_events_mtx);
535 			wakeup(ep);
536 			mtx_unlock(&sc->sc_events_mtx);
537 		}
538 	}
539 	callout_drain(&sc->sc_callout);
540 
541 	g_topology_lock();
542 	LIST_FOREACH_SAFE(cp, &sc->sc_sync.ds_geom->consumer, consumer, tmpcp) {
543 		g_mirror_disconnect_consumer(sc, cp);
544 	}
545 	g_wither_geom(sc->sc_sync.ds_geom, ENXIO);
546 	G_MIRROR_DEBUG(0, "Device %s destroyed.", gp->name);
547 	g_wither_geom(gp, ENXIO);
548 	g_topology_unlock();
549 	mtx_destroy(&sc->sc_queue_mtx);
550 	mtx_destroy(&sc->sc_events_mtx);
551 	sx_xunlock(&sc->sc_lock);
552 	sx_destroy(&sc->sc_lock);
553 }
554 
555 static void
556 g_mirror_orphan(struct g_consumer *cp)
557 {
558 	struct g_mirror_disk *disk;
559 
560 	g_topology_assert();
561 
562 	disk = cp->private;
563 	if (disk == NULL)
564 		return;
565 	disk->d_softc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
566 	g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
567 	    G_MIRROR_EVENT_DONTWAIT);
568 }
569 
570 /*
571  * Function should return the next active disk on the list.
572  * It is possible that it will be the same disk as given.
573  * If there are no active disks on list, NULL is returned.
574  */
575 static __inline struct g_mirror_disk *
576 g_mirror_find_next(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
577 {
578 	struct g_mirror_disk *dp;
579 
580 	for (dp = LIST_NEXT(disk, d_next); dp != disk;
581 	    dp = LIST_NEXT(dp, d_next)) {
582 		if (dp == NULL)
583 			dp = LIST_FIRST(&sc->sc_disks);
584 		if (dp->d_state == G_MIRROR_DISK_STATE_ACTIVE)
585 			break;
586 	}
587 	if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
588 		return (NULL);
589 	return (dp);
590 }
591 
592 static struct g_mirror_disk *
593 g_mirror_get_disk(struct g_mirror_softc *sc)
594 {
595 	struct g_mirror_disk *disk;
596 
597 	if (sc->sc_hint == NULL) {
598 		sc->sc_hint = LIST_FIRST(&sc->sc_disks);
599 		if (sc->sc_hint == NULL)
600 			return (NULL);
601 	}
602 	disk = sc->sc_hint;
603 	if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE) {
604 		disk = g_mirror_find_next(sc, disk);
605 		if (disk == NULL)
606 			return (NULL);
607 	}
608 	sc->sc_hint = g_mirror_find_next(sc, disk);
609 	return (disk);
610 }
611 
612 static int
613 g_mirror_write_metadata(struct g_mirror_disk *disk,
614     struct g_mirror_metadata *md)
615 {
616 	struct g_mirror_softc *sc;
617 	struct g_consumer *cp;
618 	off_t offset, length;
619 	u_char *sector;
620 	int error = 0;
621 
622 	g_topology_assert_not();
623 	sc = disk->d_softc;
624 	sx_assert(&sc->sc_lock, SX_LOCKED);
625 
626 	cp = disk->d_consumer;
627 	KASSERT(cp != NULL, ("NULL consumer (%s).", sc->sc_name));
628 	KASSERT(cp->provider != NULL, ("NULL provider (%s).", sc->sc_name));
629 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
630 	    ("Consumer %s closed? (r%dw%de%d).", cp->provider->name, cp->acr,
631 	    cp->acw, cp->ace));
632 	length = cp->provider->sectorsize;
633 	offset = cp->provider->mediasize - length;
634 	sector = malloc((size_t)length, M_MIRROR, M_WAITOK | M_ZERO);
635 	if (md != NULL)
636 		mirror_metadata_encode(md, sector);
637 	error = g_write_data(cp, offset, sector, length);
638 	free(sector, M_MIRROR);
639 	if (error != 0) {
640 		if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
641 			disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
642 			G_MIRROR_DEBUG(0, "Cannot write metadata on %s "
643 			    "(device=%s, error=%d).",
644 			    g_mirror_get_diskname(disk), sc->sc_name, error);
645 		} else {
646 			G_MIRROR_DEBUG(1, "Cannot write metadata on %s "
647 			    "(device=%s, error=%d).",
648 			    g_mirror_get_diskname(disk), sc->sc_name, error);
649 		}
650 		if (g_mirror_disconnect_on_failure &&
651 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1) {
652 			sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
653 			g_mirror_event_send(disk,
654 			    G_MIRROR_DISK_STATE_DISCONNECTED,
655 			    G_MIRROR_EVENT_DONTWAIT);
656 		}
657 	}
658 	return (error);
659 }
660 
661 static int
662 g_mirror_clear_metadata(struct g_mirror_disk *disk)
663 {
664 	int error;
665 
666 	g_topology_assert_not();
667 	sx_assert(&disk->d_softc->sc_lock, SX_LOCKED);
668 
669 	error = g_mirror_write_metadata(disk, NULL);
670 	if (error == 0) {
671 		G_MIRROR_DEBUG(2, "Metadata on %s cleared.",
672 		    g_mirror_get_diskname(disk));
673 	} else {
674 		G_MIRROR_DEBUG(0,
675 		    "Cannot clear metadata on disk %s (error=%d).",
676 		    g_mirror_get_diskname(disk), error);
677 	}
678 	return (error);
679 }
680 
681 void
682 g_mirror_fill_metadata(struct g_mirror_softc *sc, struct g_mirror_disk *disk,
683     struct g_mirror_metadata *md)
684 {
685 
686 	strlcpy(md->md_magic, G_MIRROR_MAGIC, sizeof(md->md_magic));
687 	md->md_version = G_MIRROR_VERSION;
688 	strlcpy(md->md_name, sc->sc_name, sizeof(md->md_name));
689 	md->md_mid = sc->sc_id;
690 	md->md_all = sc->sc_ndisks;
691 	md->md_slice = sc->sc_slice;
692 	md->md_balance = sc->sc_balance;
693 	md->md_genid = sc->sc_genid;
694 	md->md_mediasize = sc->sc_mediasize;
695 	md->md_sectorsize = sc->sc_sectorsize;
696 	md->md_mflags = (sc->sc_flags & G_MIRROR_DEVICE_FLAG_MASK);
697 	bzero(md->md_provider, sizeof(md->md_provider));
698 	if (disk == NULL) {
699 		md->md_did = arc4random();
700 		md->md_priority = 0;
701 		md->md_syncid = 0;
702 		md->md_dflags = 0;
703 		md->md_sync_offset = 0;
704 		md->md_provsize = 0;
705 	} else {
706 		md->md_did = disk->d_id;
707 		md->md_priority = disk->d_priority;
708 		md->md_syncid = disk->d_sync.ds_syncid;
709 		md->md_dflags = (disk->d_flags & G_MIRROR_DISK_FLAG_MASK);
710 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
711 			md->md_sync_offset = disk->d_sync.ds_offset_done;
712 		else
713 			md->md_sync_offset = 0;
714 		if ((disk->d_flags & G_MIRROR_DISK_FLAG_HARDCODED) != 0) {
715 			strlcpy(md->md_provider,
716 			    disk->d_consumer->provider->name,
717 			    sizeof(md->md_provider));
718 		}
719 		md->md_provsize = disk->d_consumer->provider->mediasize;
720 	}
721 }
722 
723 void
724 g_mirror_update_metadata(struct g_mirror_disk *disk)
725 {
726 	struct g_mirror_softc *sc;
727 	struct g_mirror_metadata md;
728 	int error;
729 
730 	g_topology_assert_not();
731 	sc = disk->d_softc;
732 	sx_assert(&sc->sc_lock, SX_LOCKED);
733 
734 	g_mirror_fill_metadata(sc, disk, &md);
735 	error = g_mirror_write_metadata(disk, &md);
736 	if (error == 0) {
737 		G_MIRROR_DEBUG(2, "Metadata on %s updated.",
738 		    g_mirror_get_diskname(disk));
739 	} else {
740 		G_MIRROR_DEBUG(0,
741 		    "Cannot update metadata on disk %s (error=%d).",
742 		    g_mirror_get_diskname(disk), error);
743 	}
744 }
745 
746 static void
747 g_mirror_bump_syncid(struct g_mirror_softc *sc)
748 {
749 	struct g_mirror_disk *disk;
750 
751 	g_topology_assert_not();
752 	sx_assert(&sc->sc_lock, SX_XLOCKED);
753 	KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
754 	    ("%s called with no active disks (device=%s).", __func__,
755 	    sc->sc_name));
756 
757 	sc->sc_syncid++;
758 	G_MIRROR_DEBUG(1, "Device %s: syncid bumped to %u.", sc->sc_name,
759 	    sc->sc_syncid);
760 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
761 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
762 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
763 			disk->d_sync.ds_syncid = sc->sc_syncid;
764 			g_mirror_update_metadata(disk);
765 		}
766 	}
767 }
768 
769 static void
770 g_mirror_bump_genid(struct g_mirror_softc *sc)
771 {
772 	struct g_mirror_disk *disk;
773 
774 	g_topology_assert_not();
775 	sx_assert(&sc->sc_lock, SX_XLOCKED);
776 	KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
777 	    ("%s called with no active disks (device=%s).", __func__,
778 	    sc->sc_name));
779 
780 	sc->sc_genid++;
781 	G_MIRROR_DEBUG(1, "Device %s: genid bumped to %u.", sc->sc_name,
782 	    sc->sc_genid);
783 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
784 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
785 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
786 			disk->d_genid = sc->sc_genid;
787 			g_mirror_update_metadata(disk);
788 		}
789 	}
790 }
791 
792 static int
793 g_mirror_idle(struct g_mirror_softc *sc, int acw)
794 {
795 	struct g_mirror_disk *disk;
796 	int timeout;
797 
798 	g_topology_assert_not();
799 	sx_assert(&sc->sc_lock, SX_XLOCKED);
800 
801 	if (sc->sc_provider == NULL)
802 		return (0);
803 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
804 		return (0);
805 	if (sc->sc_idle)
806 		return (0);
807 	if (sc->sc_writes > 0)
808 		return (0);
809 	if (acw > 0 || (acw == -1 && sc->sc_provider->acw > 0)) {
810 		timeout = g_mirror_idletime - (time_uptime - sc->sc_last_write);
811 		if (timeout > 0)
812 			return (timeout);
813 	}
814 	sc->sc_idle = 1;
815 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
816 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
817 			continue;
818 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as clean.",
819 		    g_mirror_get_diskname(disk), sc->sc_name);
820 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
821 		g_mirror_update_metadata(disk);
822 	}
823 	return (0);
824 }
825 
826 static void
827 g_mirror_unidle(struct g_mirror_softc *sc)
828 {
829 	struct g_mirror_disk *disk;
830 
831 	g_topology_assert_not();
832 	sx_assert(&sc->sc_lock, SX_XLOCKED);
833 
834 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
835 		return;
836 	sc->sc_idle = 0;
837 	sc->sc_last_write = time_uptime;
838 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
839 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
840 			continue;
841 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as dirty.",
842 		    g_mirror_get_diskname(disk), sc->sc_name);
843 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
844 		g_mirror_update_metadata(disk);
845 	}
846 }
847 
848 static void
849 g_mirror_done(struct bio *bp)
850 {
851 	struct g_mirror_softc *sc;
852 
853 	sc = bp->bio_from->geom->softc;
854 	bp->bio_cflags = G_MIRROR_BIO_FLAG_REGULAR;
855 	mtx_lock(&sc->sc_queue_mtx);
856 	bioq_disksort(&sc->sc_queue, bp);
857 	mtx_unlock(&sc->sc_queue_mtx);
858 	wakeup(sc);
859 }
860 
861 static void
862 g_mirror_regular_request(struct bio *bp)
863 {
864 	struct g_mirror_softc *sc;
865 	struct g_mirror_disk *disk;
866 	struct bio *pbp;
867 
868 	g_topology_assert_not();
869 
870 	pbp = bp->bio_parent;
871 	sc = pbp->bio_to->geom->softc;
872 	bp->bio_from->index--;
873 	if (bp->bio_cmd == BIO_WRITE)
874 		sc->sc_writes--;
875 	disk = bp->bio_from->private;
876 	if (disk == NULL) {
877 		g_topology_lock();
878 		g_mirror_kill_consumer(sc, bp->bio_from);
879 		g_topology_unlock();
880 	}
881 
882 	pbp->bio_inbed++;
883 	KASSERT(pbp->bio_inbed <= pbp->bio_children,
884 	    ("bio_inbed (%u) is bigger than bio_children (%u).", pbp->bio_inbed,
885 	    pbp->bio_children));
886 	if (bp->bio_error == 0 && pbp->bio_error == 0) {
887 		G_MIRROR_LOGREQ(3, bp, "Request delivered.");
888 		g_destroy_bio(bp);
889 		if (pbp->bio_children == pbp->bio_inbed) {
890 			G_MIRROR_LOGREQ(3, pbp, "Request delivered.");
891 			pbp->bio_completed = pbp->bio_length;
892 			if (pbp->bio_cmd == BIO_WRITE) {
893 				bioq_remove(&sc->sc_inflight, pbp);
894 				/* Release delayed sync requests if possible. */
895 				g_mirror_sync_release(sc);
896 			}
897 			g_io_deliver(pbp, pbp->bio_error);
898 		}
899 		return;
900 	} else if (bp->bio_error != 0) {
901 		if (pbp->bio_error == 0)
902 			pbp->bio_error = bp->bio_error;
903 		if (disk != NULL) {
904 			if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
905 				disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
906 				G_MIRROR_LOGREQ(0, bp,
907 				    "Request failed (error=%d).",
908 				    bp->bio_error);
909 			} else {
910 				G_MIRROR_LOGREQ(1, bp,
911 				    "Request failed (error=%d).",
912 				    bp->bio_error);
913 			}
914 			if (g_mirror_disconnect_on_failure &&
915 			    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1)
916 			{
917 				sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
918 				g_mirror_event_send(disk,
919 				    G_MIRROR_DISK_STATE_DISCONNECTED,
920 				    G_MIRROR_EVENT_DONTWAIT);
921 			}
922 		}
923 		switch (pbp->bio_cmd) {
924 		case BIO_DELETE:
925 		case BIO_WRITE:
926 			pbp->bio_inbed--;
927 			pbp->bio_children--;
928 			break;
929 		}
930 	}
931 	g_destroy_bio(bp);
932 
933 	switch (pbp->bio_cmd) {
934 	case BIO_READ:
935 		if (pbp->bio_inbed < pbp->bio_children)
936 			break;
937 		if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 1)
938 			g_io_deliver(pbp, pbp->bio_error);
939 		else {
940 			pbp->bio_error = 0;
941 			mtx_lock(&sc->sc_queue_mtx);
942 			bioq_disksort(&sc->sc_queue, pbp);
943 			mtx_unlock(&sc->sc_queue_mtx);
944 			G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
945 			wakeup(sc);
946 		}
947 		break;
948 	case BIO_DELETE:
949 	case BIO_WRITE:
950 		if (pbp->bio_children == 0) {
951 			/*
952 			 * All requests failed.
953 			 */
954 		} else if (pbp->bio_inbed < pbp->bio_children) {
955 			/* Do nothing. */
956 			break;
957 		} else if (pbp->bio_children == pbp->bio_inbed) {
958 			/* Some requests succeeded. */
959 			pbp->bio_error = 0;
960 			pbp->bio_completed = pbp->bio_length;
961 		}
962 		bioq_remove(&sc->sc_inflight, pbp);
963 		/* Release delayed sync requests if possible. */
964 		g_mirror_sync_release(sc);
965 		g_io_deliver(pbp, pbp->bio_error);
966 		break;
967 	default:
968 		KASSERT(1 == 0, ("Invalid request: %u.", pbp->bio_cmd));
969 		break;
970 	}
971 }
972 
973 static void
974 g_mirror_sync_done(struct bio *bp)
975 {
976 	struct g_mirror_softc *sc;
977 
978 	G_MIRROR_LOGREQ(3, bp, "Synchronization request delivered.");
979 	sc = bp->bio_from->geom->softc;
980 	bp->bio_cflags = G_MIRROR_BIO_FLAG_SYNC;
981 	mtx_lock(&sc->sc_queue_mtx);
982 	bioq_disksort(&sc->sc_queue, bp);
983 	mtx_unlock(&sc->sc_queue_mtx);
984 	wakeup(sc);
985 }
986 
987 static void
988 g_mirror_kernel_dump(struct bio *bp)
989 {
990 	struct g_mirror_softc *sc;
991 	struct g_mirror_disk *disk;
992 	struct bio *cbp;
993 	struct g_kerneldump *gkd;
994 
995 	/*
996 	 * We configure dumping to the first component, because this component
997 	 * will be used for reading with 'prefer' balance algorithm.
998 	 * If the component with the higest priority is currently disconnected
999 	 * we will not be able to read the dump after the reboot if it will be
1000 	 * connected and synchronized later. Can we do something better?
1001 	 */
1002 	sc = bp->bio_to->geom->softc;
1003 	disk = LIST_FIRST(&sc->sc_disks);
1004 
1005 	gkd = (struct g_kerneldump *)bp->bio_data;
1006 	if (gkd->length > bp->bio_to->mediasize)
1007 		gkd->length = bp->bio_to->mediasize;
1008 	cbp = g_clone_bio(bp);
1009 	if (cbp == NULL) {
1010 		g_io_deliver(bp, ENOMEM);
1011 		return;
1012 	}
1013 	cbp->bio_done = g_std_done;
1014 	g_io_request(cbp, disk->d_consumer);
1015 	G_MIRROR_DEBUG(1, "Kernel dump will go to %s.",
1016 	    g_mirror_get_diskname(disk));
1017 }
1018 
1019 static void
1020 g_mirror_flush(struct g_mirror_softc *sc, struct bio *bp)
1021 {
1022 	struct bio_queue_head queue;
1023 	struct g_mirror_disk *disk;
1024 	struct g_consumer *cp;
1025 	struct bio *cbp;
1026 
1027 	bioq_init(&queue);
1028 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1029 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1030 			continue;
1031 		cbp = g_clone_bio(bp);
1032 		if (cbp == NULL) {
1033 			for (cbp = bioq_first(&queue); cbp != NULL;
1034 			    cbp = bioq_first(&queue)) {
1035 				bioq_remove(&queue, cbp);
1036 				g_destroy_bio(cbp);
1037 			}
1038 			if (bp->bio_error == 0)
1039 				bp->bio_error = ENOMEM;
1040 			g_io_deliver(bp, bp->bio_error);
1041 			return;
1042 		}
1043 		bioq_insert_tail(&queue, cbp);
1044 		cbp->bio_done = g_std_done;
1045 		cbp->bio_caller1 = disk;
1046 		cbp->bio_to = disk->d_consumer->provider;
1047 	}
1048 	for (cbp = bioq_first(&queue); cbp != NULL; cbp = bioq_first(&queue)) {
1049 		bioq_remove(&queue, cbp);
1050 		G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1051 		disk = cbp->bio_caller1;
1052 		cbp->bio_caller1 = NULL;
1053 		cp = disk->d_consumer;
1054 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1055 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1056 		    cp->acr, cp->acw, cp->ace));
1057 		g_io_request(cbp, disk->d_consumer);
1058 	}
1059 }
1060 
1061 static void
1062 g_mirror_start(struct bio *bp)
1063 {
1064 	struct g_mirror_softc *sc;
1065 
1066 	sc = bp->bio_to->geom->softc;
1067 	/*
1068 	 * If sc == NULL or there are no valid disks, provider's error
1069 	 * should be set and g_mirror_start() should not be called at all.
1070 	 */
1071 	KASSERT(sc != NULL && sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1072 	    ("Provider's error should be set (error=%d)(mirror=%s).",
1073 	    bp->bio_to->error, bp->bio_to->name));
1074 	G_MIRROR_LOGREQ(3, bp, "Request received.");
1075 
1076 	switch (bp->bio_cmd) {
1077 	case BIO_READ:
1078 	case BIO_WRITE:
1079 	case BIO_DELETE:
1080 		break;
1081 	case BIO_FLUSH:
1082 		g_mirror_flush(sc, bp);
1083 		return;
1084 	case BIO_GETATTR:
1085 		if (strcmp("GEOM::kerneldump", bp->bio_attribute) == 0) {
1086 			g_mirror_kernel_dump(bp);
1087 			return;
1088 		}
1089 		/* FALLTHROUGH */
1090 	default:
1091 		g_io_deliver(bp, EOPNOTSUPP);
1092 		return;
1093 	}
1094 	mtx_lock(&sc->sc_queue_mtx);
1095 	bioq_disksort(&sc->sc_queue, bp);
1096 	mtx_unlock(&sc->sc_queue_mtx);
1097 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1098 	wakeup(sc);
1099 }
1100 
1101 /*
1102  * Return TRUE if the given request is colliding with a in-progress
1103  * synchronization request.
1104  */
1105 static int
1106 g_mirror_sync_collision(struct g_mirror_softc *sc, struct bio *bp)
1107 {
1108 	struct g_mirror_disk *disk;
1109 	struct bio *sbp;
1110 	off_t rstart, rend, sstart, send;
1111 	int i;
1112 
1113 	if (sc->sc_sync.ds_ndisks == 0)
1114 		return (0);
1115 	rstart = bp->bio_offset;
1116 	rend = bp->bio_offset + bp->bio_length;
1117 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1118 		if (disk->d_state != G_MIRROR_DISK_STATE_SYNCHRONIZING)
1119 			continue;
1120 		for (i = 0; i < g_mirror_syncreqs; i++) {
1121 			sbp = disk->d_sync.ds_bios[i];
1122 			if (sbp == NULL)
1123 				continue;
1124 			sstart = sbp->bio_offset;
1125 			send = sbp->bio_offset + sbp->bio_length;
1126 			if (rend > sstart && rstart < send)
1127 				return (1);
1128 		}
1129 	}
1130 	return (0);
1131 }
1132 
1133 /*
1134  * Return TRUE if the given sync request is colliding with a in-progress regular
1135  * request.
1136  */
1137 static int
1138 g_mirror_regular_collision(struct g_mirror_softc *sc, struct bio *sbp)
1139 {
1140 	off_t rstart, rend, sstart, send;
1141 	struct bio *bp;
1142 
1143 	if (sc->sc_sync.ds_ndisks == 0)
1144 		return (0);
1145 	sstart = sbp->bio_offset;
1146 	send = sbp->bio_offset + sbp->bio_length;
1147 	TAILQ_FOREACH(bp, &sc->sc_inflight.queue, bio_queue) {
1148 		rstart = bp->bio_offset;
1149 		rend = bp->bio_offset + bp->bio_length;
1150 		if (rend > sstart && rstart < send)
1151 			return (1);
1152 	}
1153 	return (0);
1154 }
1155 
1156 /*
1157  * Puts request onto delayed queue.
1158  */
1159 static void
1160 g_mirror_regular_delay(struct g_mirror_softc *sc, struct bio *bp)
1161 {
1162 
1163 	G_MIRROR_LOGREQ(2, bp, "Delaying request.");
1164 	bioq_insert_head(&sc->sc_regular_delayed, bp);
1165 }
1166 
1167 /*
1168  * Puts synchronization request onto delayed queue.
1169  */
1170 static void
1171 g_mirror_sync_delay(struct g_mirror_softc *sc, struct bio *bp)
1172 {
1173 
1174 	G_MIRROR_LOGREQ(2, bp, "Delaying synchronization request.");
1175 	bioq_insert_tail(&sc->sc_sync_delayed, bp);
1176 }
1177 
1178 /*
1179  * Releases delayed regular requests which don't collide anymore with sync
1180  * requests.
1181  */
1182 static void
1183 g_mirror_regular_release(struct g_mirror_softc *sc)
1184 {
1185 	struct bio *bp, *bp2;
1186 
1187 	TAILQ_FOREACH_SAFE(bp, &sc->sc_regular_delayed.queue, bio_queue, bp2) {
1188 		if (g_mirror_sync_collision(sc, bp))
1189 			continue;
1190 		bioq_remove(&sc->sc_regular_delayed, bp);
1191 		G_MIRROR_LOGREQ(2, bp, "Releasing delayed request (%p).", bp);
1192 		mtx_lock(&sc->sc_queue_mtx);
1193 		bioq_insert_head(&sc->sc_queue, bp);
1194 #if 0
1195 		/*
1196 		 * wakeup() is not needed, because this function is called from
1197 		 * the worker thread.
1198 		 */
1199 		wakeup(&sc->sc_queue);
1200 #endif
1201 		mtx_unlock(&sc->sc_queue_mtx);
1202 	}
1203 }
1204 
1205 /*
1206  * Releases delayed sync requests which don't collide anymore with regular
1207  * requests.
1208  */
1209 static void
1210 g_mirror_sync_release(struct g_mirror_softc *sc)
1211 {
1212 	struct bio *bp, *bp2;
1213 
1214 	TAILQ_FOREACH_SAFE(bp, &sc->sc_sync_delayed.queue, bio_queue, bp2) {
1215 		if (g_mirror_regular_collision(sc, bp))
1216 			continue;
1217 		bioq_remove(&sc->sc_sync_delayed, bp);
1218 		G_MIRROR_LOGREQ(2, bp,
1219 		    "Releasing delayed synchronization request.");
1220 		g_io_request(bp, bp->bio_from);
1221 	}
1222 }
1223 
1224 /*
1225  * Handle synchronization requests.
1226  * Every synchronization request is two-steps process: first, READ request is
1227  * send to active provider and then WRITE request (with read data) to the provider
1228  * beeing synchronized. When WRITE is finished, new synchronization request is
1229  * send.
1230  */
1231 static void
1232 g_mirror_sync_request(struct bio *bp)
1233 {
1234 	struct g_mirror_softc *sc;
1235 	struct g_mirror_disk *disk;
1236 
1237 	bp->bio_from->index--;
1238 	sc = bp->bio_from->geom->softc;
1239 	disk = bp->bio_from->private;
1240 	if (disk == NULL) {
1241 		sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1242 		g_topology_lock();
1243 		g_mirror_kill_consumer(sc, bp->bio_from);
1244 		g_topology_unlock();
1245 		free(bp->bio_data, M_MIRROR);
1246 		g_destroy_bio(bp);
1247 		sx_xlock(&sc->sc_lock);
1248 		return;
1249 	}
1250 
1251 	/*
1252 	 * Synchronization request.
1253 	 */
1254 	switch (bp->bio_cmd) {
1255 	case BIO_READ:
1256 	    {
1257 		struct g_consumer *cp;
1258 
1259 		if (bp->bio_error != 0) {
1260 			G_MIRROR_LOGREQ(0, bp,
1261 			    "Synchronization request failed (error=%d).",
1262 			    bp->bio_error);
1263 			g_destroy_bio(bp);
1264 			return;
1265 		}
1266 		G_MIRROR_LOGREQ(3, bp,
1267 		    "Synchronization request half-finished.");
1268 		bp->bio_cmd = BIO_WRITE;
1269 		bp->bio_cflags = 0;
1270 		cp = disk->d_consumer;
1271 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1272 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1273 		    cp->acr, cp->acw, cp->ace));
1274 		cp->index++;
1275 		g_io_request(bp, cp);
1276 		return;
1277 	    }
1278 	case BIO_WRITE:
1279 	    {
1280 		struct g_mirror_disk_sync *sync;
1281 		off_t offset;
1282 		void *data;
1283 		int i;
1284 
1285 		if (bp->bio_error != 0) {
1286 			G_MIRROR_LOGREQ(0, bp,
1287 			    "Synchronization request failed (error=%d).",
1288 			    bp->bio_error);
1289 			g_destroy_bio(bp);
1290 			sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
1291 			g_mirror_event_send(disk,
1292 			    G_MIRROR_DISK_STATE_DISCONNECTED,
1293 			    G_MIRROR_EVENT_DONTWAIT);
1294 			return;
1295 		}
1296 		G_MIRROR_LOGREQ(3, bp, "Synchronization request finished.");
1297 		sync = &disk->d_sync;
1298 		if (sync->ds_offset == sc->sc_mediasize ||
1299 		    sync->ds_consumer == NULL ||
1300 		    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1301 			/* Don't send more synchronization requests. */
1302 			sync->ds_inflight--;
1303 			if (sync->ds_bios != NULL) {
1304 				i = (int)(uintptr_t)bp->bio_caller1;
1305 				sync->ds_bios[i] = NULL;
1306 			}
1307 			free(bp->bio_data, M_MIRROR);
1308 			g_destroy_bio(bp);
1309 			if (sync->ds_inflight > 0)
1310 				return;
1311 			if (sync->ds_consumer == NULL ||
1312 			    (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1313 				return;
1314 			}
1315 			/* Disk up-to-date, activate it. */
1316 			g_mirror_event_send(disk, G_MIRROR_DISK_STATE_ACTIVE,
1317 			    G_MIRROR_EVENT_DONTWAIT);
1318 			return;
1319 		}
1320 
1321 		/* Send next synchronization request. */
1322 		data = bp->bio_data;
1323 		bzero(bp, sizeof(*bp));
1324 		bp->bio_cmd = BIO_READ;
1325 		bp->bio_offset = sync->ds_offset;
1326 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1327 		sync->ds_offset += bp->bio_length;
1328 		bp->bio_done = g_mirror_sync_done;
1329 		bp->bio_data = data;
1330 		bp->bio_from = sync->ds_consumer;
1331 		bp->bio_to = sc->sc_provider;
1332 		G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
1333 		sync->ds_consumer->index++;
1334 		/*
1335 		 * Delay the request if it is colliding with a regular request.
1336 		 */
1337 		if (g_mirror_regular_collision(sc, bp))
1338 			g_mirror_sync_delay(sc, bp);
1339 		else
1340 			g_io_request(bp, sync->ds_consumer);
1341 
1342 		/* Release delayed requests if possible. */
1343 		g_mirror_regular_release(sc);
1344 
1345 		/* Find the smallest offset */
1346 		offset = sc->sc_mediasize;
1347 		for (i = 0; i < g_mirror_syncreqs; i++) {
1348 			bp = sync->ds_bios[i];
1349 			if (bp->bio_offset < offset)
1350 				offset = bp->bio_offset;
1351 		}
1352 		if (sync->ds_offset_done + (MAXPHYS * 100) < offset) {
1353 			/* Update offset_done on every 100 blocks. */
1354 			sync->ds_offset_done = offset;
1355 			g_mirror_update_metadata(disk);
1356 		}
1357 		return;
1358 	    }
1359 	default:
1360 		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1361 		    bp->bio_cmd, sc->sc_name));
1362 		break;
1363 	}
1364 }
1365 
1366 static void
1367 g_mirror_request_prefer(struct g_mirror_softc *sc, struct bio *bp)
1368 {
1369 	struct g_mirror_disk *disk;
1370 	struct g_consumer *cp;
1371 	struct bio *cbp;
1372 
1373 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1374 		if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE)
1375 			break;
1376 	}
1377 	if (disk == NULL) {
1378 		if (bp->bio_error == 0)
1379 			bp->bio_error = ENXIO;
1380 		g_io_deliver(bp, bp->bio_error);
1381 		return;
1382 	}
1383 	cbp = g_clone_bio(bp);
1384 	if (cbp == NULL) {
1385 		if (bp->bio_error == 0)
1386 			bp->bio_error = ENOMEM;
1387 		g_io_deliver(bp, bp->bio_error);
1388 		return;
1389 	}
1390 	/*
1391 	 * Fill in the component buf structure.
1392 	 */
1393 	cp = disk->d_consumer;
1394 	cbp->bio_done = g_mirror_done;
1395 	cbp->bio_to = cp->provider;
1396 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1397 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1398 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1399 	    cp->acw, cp->ace));
1400 	cp->index++;
1401 	g_io_request(cbp, cp);
1402 }
1403 
1404 static void
1405 g_mirror_request_round_robin(struct g_mirror_softc *sc, struct bio *bp)
1406 {
1407 	struct g_mirror_disk *disk;
1408 	struct g_consumer *cp;
1409 	struct bio *cbp;
1410 
1411 	disk = g_mirror_get_disk(sc);
1412 	if (disk == NULL) {
1413 		if (bp->bio_error == 0)
1414 			bp->bio_error = ENXIO;
1415 		g_io_deliver(bp, bp->bio_error);
1416 		return;
1417 	}
1418 	cbp = g_clone_bio(bp);
1419 	if (cbp == NULL) {
1420 		if (bp->bio_error == 0)
1421 			bp->bio_error = ENOMEM;
1422 		g_io_deliver(bp, bp->bio_error);
1423 		return;
1424 	}
1425 	/*
1426 	 * Fill in the component buf structure.
1427 	 */
1428 	cp = disk->d_consumer;
1429 	cbp->bio_done = g_mirror_done;
1430 	cbp->bio_to = cp->provider;
1431 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1432 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1433 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1434 	    cp->acw, cp->ace));
1435 	cp->index++;
1436 	g_io_request(cbp, cp);
1437 }
1438 
1439 #define TRACK_SIZE  (1 * 1024 * 1024)
1440 #define LOAD_SCALE	256
1441 #define ABS(x)		(((x) >= 0) ? (x) : (-(x)))
1442 
1443 static void
1444 g_mirror_request_load(struct g_mirror_softc *sc, struct bio *bp)
1445 {
1446 	struct g_mirror_disk *disk, *dp;
1447 	struct g_consumer *cp;
1448 	struct bio *cbp;
1449 	int prio, best;
1450 
1451 	/* Find a disk with the smallest load. */
1452 	disk = NULL;
1453 	best = INT_MAX;
1454 	LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1455 		if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1456 			continue;
1457 		prio = dp->load;
1458 		/* If disk head is precisely in position - highly prefer it. */
1459 		if (dp->d_last_offset == bp->bio_offset)
1460 			prio -= 2 * LOAD_SCALE;
1461 		else
1462 		/* If disk head is close to position - prefer it. */
1463 		if (ABS(dp->d_last_offset - bp->bio_offset) < TRACK_SIZE)
1464 			prio -= 1 * LOAD_SCALE;
1465 		if (prio <= best) {
1466 			disk = dp;
1467 			best = prio;
1468 		}
1469 	}
1470 	KASSERT(disk != NULL, ("NULL disk for %s.", sc->sc_name));
1471 	cbp = g_clone_bio(bp);
1472 	if (cbp == NULL) {
1473 		if (bp->bio_error == 0)
1474 			bp->bio_error = ENOMEM;
1475 		g_io_deliver(bp, bp->bio_error);
1476 		return;
1477 	}
1478 	/*
1479 	 * Fill in the component buf structure.
1480 	 */
1481 	cp = disk->d_consumer;
1482 	cbp->bio_done = g_mirror_done;
1483 	cbp->bio_to = cp->provider;
1484 	G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1485 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1486 	    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1487 	    cp->acw, cp->ace));
1488 	cp->index++;
1489 	/* Remember last head position */
1490 	disk->d_last_offset = bp->bio_offset + bp->bio_length;
1491 	/* Update loads. */
1492 	LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1493 		dp->load = (dp->d_consumer->index * LOAD_SCALE +
1494 		    dp->load * 7) / 8;
1495 	}
1496 	g_io_request(cbp, cp);
1497 }
1498 
1499 static void
1500 g_mirror_request_split(struct g_mirror_softc *sc, struct bio *bp)
1501 {
1502 	struct bio_queue_head queue;
1503 	struct g_mirror_disk *disk;
1504 	struct g_consumer *cp;
1505 	struct bio *cbp;
1506 	off_t left, mod, offset, slice;
1507 	u_char *data;
1508 	u_int ndisks;
1509 
1510 	if (bp->bio_length <= sc->sc_slice) {
1511 		g_mirror_request_round_robin(sc, bp);
1512 		return;
1513 	}
1514 	ndisks = g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE);
1515 	slice = bp->bio_length / ndisks;
1516 	mod = slice % sc->sc_provider->sectorsize;
1517 	if (mod != 0)
1518 		slice += sc->sc_provider->sectorsize - mod;
1519 	/*
1520 	 * Allocate all bios before sending any request, so we can
1521 	 * return ENOMEM in nice and clean way.
1522 	 */
1523 	left = bp->bio_length;
1524 	offset = bp->bio_offset;
1525 	data = bp->bio_data;
1526 	bioq_init(&queue);
1527 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1528 		if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1529 			continue;
1530 		cbp = g_clone_bio(bp);
1531 		if (cbp == NULL) {
1532 			for (cbp = bioq_first(&queue); cbp != NULL;
1533 			    cbp = bioq_first(&queue)) {
1534 				bioq_remove(&queue, cbp);
1535 				g_destroy_bio(cbp);
1536 			}
1537 			if (bp->bio_error == 0)
1538 				bp->bio_error = ENOMEM;
1539 			g_io_deliver(bp, bp->bio_error);
1540 			return;
1541 		}
1542 		bioq_insert_tail(&queue, cbp);
1543 		cbp->bio_done = g_mirror_done;
1544 		cbp->bio_caller1 = disk;
1545 		cbp->bio_to = disk->d_consumer->provider;
1546 		cbp->bio_offset = offset;
1547 		cbp->bio_data = data;
1548 		cbp->bio_length = MIN(left, slice);
1549 		left -= cbp->bio_length;
1550 		if (left == 0)
1551 			break;
1552 		offset += cbp->bio_length;
1553 		data += cbp->bio_length;
1554 	}
1555 	for (cbp = bioq_first(&queue); cbp != NULL; cbp = bioq_first(&queue)) {
1556 		bioq_remove(&queue, cbp);
1557 		G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1558 		disk = cbp->bio_caller1;
1559 		cbp->bio_caller1 = NULL;
1560 		cp = disk->d_consumer;
1561 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1562 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1563 		    cp->acr, cp->acw, cp->ace));
1564 		disk->d_consumer->index++;
1565 		g_io_request(cbp, disk->d_consumer);
1566 	}
1567 }
1568 
1569 static void
1570 g_mirror_register_request(struct bio *bp)
1571 {
1572 	struct g_mirror_softc *sc;
1573 
1574 	sc = bp->bio_to->geom->softc;
1575 	switch (bp->bio_cmd) {
1576 	case BIO_READ:
1577 		switch (sc->sc_balance) {
1578 		case G_MIRROR_BALANCE_LOAD:
1579 			g_mirror_request_load(sc, bp);
1580 			break;
1581 		case G_MIRROR_BALANCE_PREFER:
1582 			g_mirror_request_prefer(sc, bp);
1583 			break;
1584 		case G_MIRROR_BALANCE_ROUND_ROBIN:
1585 			g_mirror_request_round_robin(sc, bp);
1586 			break;
1587 		case G_MIRROR_BALANCE_SPLIT:
1588 			g_mirror_request_split(sc, bp);
1589 			break;
1590 		}
1591 		return;
1592 	case BIO_WRITE:
1593 	case BIO_DELETE:
1594 	    {
1595 		struct g_mirror_disk *disk;
1596 		struct g_mirror_disk_sync *sync;
1597 		struct bio_queue_head queue;
1598 		struct g_consumer *cp;
1599 		struct bio *cbp;
1600 
1601 		/*
1602 		 * Delay the request if it is colliding with a synchronization
1603 		 * request.
1604 		 */
1605 		if (g_mirror_sync_collision(sc, bp)) {
1606 			g_mirror_regular_delay(sc, bp);
1607 			return;
1608 		}
1609 
1610 		if (sc->sc_idle)
1611 			g_mirror_unidle(sc);
1612 		else
1613 			sc->sc_last_write = time_uptime;
1614 
1615 		/*
1616 		 * Allocate all bios before sending any request, so we can
1617 		 * return ENOMEM in nice and clean way.
1618 		 */
1619 		bioq_init(&queue);
1620 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1621 			sync = &disk->d_sync;
1622 			switch (disk->d_state) {
1623 			case G_MIRROR_DISK_STATE_ACTIVE:
1624 				break;
1625 			case G_MIRROR_DISK_STATE_SYNCHRONIZING:
1626 				if (bp->bio_offset >= sync->ds_offset)
1627 					continue;
1628 				break;
1629 			default:
1630 				continue;
1631 			}
1632 			cbp = g_clone_bio(bp);
1633 			if (cbp == NULL) {
1634 				for (cbp = bioq_first(&queue); cbp != NULL;
1635 				    cbp = bioq_first(&queue)) {
1636 					bioq_remove(&queue, cbp);
1637 					g_destroy_bio(cbp);
1638 				}
1639 				if (bp->bio_error == 0)
1640 					bp->bio_error = ENOMEM;
1641 				g_io_deliver(bp, bp->bio_error);
1642 				return;
1643 			}
1644 			bioq_insert_tail(&queue, cbp);
1645 			cbp->bio_done = g_mirror_done;
1646 			cp = disk->d_consumer;
1647 			cbp->bio_caller1 = cp;
1648 			cbp->bio_to = cp->provider;
1649 			KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1650 			    ("Consumer %s not opened (r%dw%de%d).",
1651 			    cp->provider->name, cp->acr, cp->acw, cp->ace));
1652 		}
1653 		for (cbp = bioq_first(&queue); cbp != NULL;
1654 		    cbp = bioq_first(&queue)) {
1655 			bioq_remove(&queue, cbp);
1656 			G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1657 			cp = cbp->bio_caller1;
1658 			cbp->bio_caller1 = NULL;
1659 			cp->index++;
1660 			sc->sc_writes++;
1661 			g_io_request(cbp, cp);
1662 		}
1663 		/*
1664 		 * Put request onto inflight queue, so we can check if new
1665 		 * synchronization requests don't collide with it.
1666 		 */
1667 		bioq_insert_tail(&sc->sc_inflight, bp);
1668 		/*
1669 		 * Bump syncid on first write.
1670 		 */
1671 		if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0) {
1672 			sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
1673 			g_mirror_bump_syncid(sc);
1674 		}
1675 		return;
1676 	    }
1677 	default:
1678 		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1679 		    bp->bio_cmd, sc->sc_name));
1680 		break;
1681 	}
1682 }
1683 
1684 static int
1685 g_mirror_can_destroy(struct g_mirror_softc *sc)
1686 {
1687 	struct g_geom *gp;
1688 	struct g_consumer *cp;
1689 
1690 	g_topology_assert();
1691 	gp = sc->sc_geom;
1692 	if (gp->softc == NULL)
1693 		return (1);
1694 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1695 		if (g_mirror_is_busy(sc, cp))
1696 			return (0);
1697 	}
1698 	gp = sc->sc_sync.ds_geom;
1699 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1700 		if (g_mirror_is_busy(sc, cp))
1701 			return (0);
1702 	}
1703 	G_MIRROR_DEBUG(2, "No I/O requests for %s, it can be destroyed.",
1704 	    sc->sc_name);
1705 	return (1);
1706 }
1707 
1708 static int
1709 g_mirror_try_destroy(struct g_mirror_softc *sc)
1710 {
1711 
1712 	if (sc->sc_rootmount != NULL) {
1713 		G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
1714 		    sc->sc_rootmount);
1715 		root_mount_rel(sc->sc_rootmount);
1716 		sc->sc_rootmount = NULL;
1717 	}
1718 	g_topology_lock();
1719 	if (!g_mirror_can_destroy(sc)) {
1720 		g_topology_unlock();
1721 		return (0);
1722 	}
1723 	sc->sc_geom->softc = NULL;
1724 	sc->sc_sync.ds_geom->softc = NULL;
1725 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_WAIT) != 0) {
1726 		g_topology_unlock();
1727 		G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1728 		    &sc->sc_worker);
1729 		/* Unlock sc_lock here, as it can be destroyed after wakeup. */
1730 		sx_xunlock(&sc->sc_lock);
1731 		wakeup(&sc->sc_worker);
1732 		sc->sc_worker = NULL;
1733 	} else {
1734 		g_topology_unlock();
1735 		g_mirror_destroy_device(sc);
1736 		free(sc, M_MIRROR);
1737 	}
1738 	return (1);
1739 }
1740 
1741 /*
1742  * Worker thread.
1743  */
1744 static void
1745 g_mirror_worker(void *arg)
1746 {
1747 	struct g_mirror_softc *sc;
1748 	struct g_mirror_event *ep;
1749 	struct bio *bp;
1750 	int timeout;
1751 
1752 	sc = arg;
1753 	thread_lock(curthread);
1754 	sched_prio(curthread, PRIBIO);
1755 	thread_unlock(curthread);
1756 
1757 	sx_xlock(&sc->sc_lock);
1758 	for (;;) {
1759 		G_MIRROR_DEBUG(5, "%s: Let's see...", __func__);
1760 		/*
1761 		 * First take a look at events.
1762 		 * This is important to handle events before any I/O requests.
1763 		 */
1764 		ep = g_mirror_event_get(sc);
1765 		if (ep != NULL) {
1766 			g_mirror_event_remove(sc, ep);
1767 			if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0) {
1768 				/* Update only device status. */
1769 				G_MIRROR_DEBUG(3,
1770 				    "Running event for device %s.",
1771 				    sc->sc_name);
1772 				ep->e_error = 0;
1773 				g_mirror_update_device(sc, 1);
1774 			} else {
1775 				/* Update disk status. */
1776 				G_MIRROR_DEBUG(3, "Running event for disk %s.",
1777 				     g_mirror_get_diskname(ep->e_disk));
1778 				ep->e_error = g_mirror_update_disk(ep->e_disk,
1779 				    ep->e_state);
1780 				if (ep->e_error == 0)
1781 					g_mirror_update_device(sc, 0);
1782 			}
1783 			if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0) {
1784 				KASSERT(ep->e_error == 0,
1785 				    ("Error cannot be handled."));
1786 				g_mirror_event_free(ep);
1787 			} else {
1788 				ep->e_flags |= G_MIRROR_EVENT_DONE;
1789 				G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1790 				    ep);
1791 				mtx_lock(&sc->sc_events_mtx);
1792 				wakeup(ep);
1793 				mtx_unlock(&sc->sc_events_mtx);
1794 			}
1795 			if ((sc->sc_flags &
1796 			    G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1797 				if (g_mirror_try_destroy(sc)) {
1798 					curthread->td_pflags &= ~TDP_GEOM;
1799 					G_MIRROR_DEBUG(1, "Thread exiting.");
1800 					kproc_exit(0);
1801 				}
1802 			}
1803 			G_MIRROR_DEBUG(5, "%s: I'm here 1.", __func__);
1804 			continue;
1805 		}
1806 		/*
1807 		 * Check if we can mark array as CLEAN and if we can't take
1808 		 * how much seconds should we wait.
1809 		 */
1810 		timeout = g_mirror_idle(sc, -1);
1811 		/*
1812 		 * Now I/O requests.
1813 		 */
1814 		/* Get first request from the queue. */
1815 		mtx_lock(&sc->sc_queue_mtx);
1816 		bp = bioq_first(&sc->sc_queue);
1817 		if (bp == NULL) {
1818 			if ((sc->sc_flags &
1819 			    G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1820 				mtx_unlock(&sc->sc_queue_mtx);
1821 				if (g_mirror_try_destroy(sc)) {
1822 					curthread->td_pflags &= ~TDP_GEOM;
1823 					G_MIRROR_DEBUG(1, "Thread exiting.");
1824 					kproc_exit(0);
1825 				}
1826 				mtx_lock(&sc->sc_queue_mtx);
1827 			}
1828 			sx_xunlock(&sc->sc_lock);
1829 			/*
1830 			 * XXX: We can miss an event here, because an event
1831 			 *      can be added without sx-device-lock and without
1832 			 *      mtx-queue-lock. Maybe I should just stop using
1833 			 *      dedicated mutex for events synchronization and
1834 			 *      stick with the queue lock?
1835 			 *      The event will hang here until next I/O request
1836 			 *      or next event is received.
1837 			 */
1838 			MSLEEP(sc, &sc->sc_queue_mtx, PRIBIO | PDROP, "m:w1",
1839 			    timeout * hz);
1840 			sx_xlock(&sc->sc_lock);
1841 			G_MIRROR_DEBUG(5, "%s: I'm here 4.", __func__);
1842 			continue;
1843 		}
1844 		bioq_remove(&sc->sc_queue, bp);
1845 		mtx_unlock(&sc->sc_queue_mtx);
1846 
1847 		if (bp->bio_from->geom == sc->sc_sync.ds_geom &&
1848 		    (bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0) {
1849 			g_mirror_sync_request(bp);	/* READ */
1850 		} else if (bp->bio_to != sc->sc_provider) {
1851 			if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_REGULAR) != 0)
1852 				g_mirror_regular_request(bp);
1853 			else if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0)
1854 				g_mirror_sync_request(bp);	/* WRITE */
1855 			else {
1856 				KASSERT(0,
1857 				    ("Invalid request cflags=0x%hhx to=%s.",
1858 				    bp->bio_cflags, bp->bio_to->name));
1859 			}
1860 		} else {
1861 			g_mirror_register_request(bp);
1862 		}
1863 		G_MIRROR_DEBUG(5, "%s: I'm here 9.", __func__);
1864 	}
1865 }
1866 
1867 static void
1868 g_mirror_update_idle(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
1869 {
1870 
1871 	sx_assert(&sc->sc_lock, SX_LOCKED);
1872 
1873 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
1874 		return;
1875 	if (!sc->sc_idle && (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) == 0) {
1876 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as dirty.",
1877 		    g_mirror_get_diskname(disk), sc->sc_name);
1878 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
1879 	} else if (sc->sc_idle &&
1880 	    (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
1881 		G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as clean.",
1882 		    g_mirror_get_diskname(disk), sc->sc_name);
1883 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
1884 	}
1885 }
1886 
1887 static void
1888 g_mirror_sync_start(struct g_mirror_disk *disk)
1889 {
1890 	struct g_mirror_softc *sc;
1891 	struct g_consumer *cp;
1892 	struct bio *bp;
1893 	int error, i;
1894 
1895 	g_topology_assert_not();
1896 	sc = disk->d_softc;
1897 	sx_assert(&sc->sc_lock, SX_LOCKED);
1898 
1899 	KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
1900 	    ("Disk %s is not marked for synchronization.",
1901 	    g_mirror_get_diskname(disk)));
1902 	KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1903 	    ("Device not in RUNNING state (%s, %u).", sc->sc_name,
1904 	    sc->sc_state));
1905 
1906 	sx_xunlock(&sc->sc_lock);
1907 	g_topology_lock();
1908 	cp = g_new_consumer(sc->sc_sync.ds_geom);
1909 	error = g_attach(cp, sc->sc_provider);
1910 	KASSERT(error == 0,
1911 	    ("Cannot attach to %s (error=%d).", sc->sc_name, error));
1912 	error = g_access(cp, 1, 0, 0);
1913 	KASSERT(error == 0, ("Cannot open %s (error=%d).", sc->sc_name, error));
1914 	g_topology_unlock();
1915 	sx_xlock(&sc->sc_lock);
1916 
1917 	G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s.", sc->sc_name,
1918 	    g_mirror_get_diskname(disk));
1919 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) == 0)
1920 		disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
1921 	KASSERT(disk->d_sync.ds_consumer == NULL,
1922 	    ("Sync consumer already exists (device=%s, disk=%s).",
1923 	    sc->sc_name, g_mirror_get_diskname(disk)));
1924 
1925 	disk->d_sync.ds_consumer = cp;
1926 	disk->d_sync.ds_consumer->private = disk;
1927 	disk->d_sync.ds_consumer->index = 0;
1928 
1929 	/*
1930 	 * Allocate memory for synchronization bios and initialize them.
1931 	 */
1932 	disk->d_sync.ds_bios = malloc(sizeof(struct bio *) * g_mirror_syncreqs,
1933 	    M_MIRROR, M_WAITOK);
1934 	for (i = 0; i < g_mirror_syncreqs; i++) {
1935 		bp = g_alloc_bio();
1936 		disk->d_sync.ds_bios[i] = bp;
1937 		bp->bio_parent = NULL;
1938 		bp->bio_cmd = BIO_READ;
1939 		bp->bio_data = malloc(MAXPHYS, M_MIRROR, M_WAITOK);
1940 		bp->bio_cflags = 0;
1941 		bp->bio_offset = disk->d_sync.ds_offset;
1942 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1943 		disk->d_sync.ds_offset += bp->bio_length;
1944 		bp->bio_done = g_mirror_sync_done;
1945 		bp->bio_from = disk->d_sync.ds_consumer;
1946 		bp->bio_to = sc->sc_provider;
1947 		bp->bio_caller1 = (void *)(uintptr_t)i;
1948 	}
1949 
1950 	/* Increase the number of disks in SYNCHRONIZING state. */
1951 	sc->sc_sync.ds_ndisks++;
1952 	/* Set the number of in-flight synchronization requests. */
1953 	disk->d_sync.ds_inflight = g_mirror_syncreqs;
1954 
1955 	/*
1956 	 * Fire off first synchronization requests.
1957 	 */
1958 	for (i = 0; i < g_mirror_syncreqs; i++) {
1959 		bp = disk->d_sync.ds_bios[i];
1960 		G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
1961 		disk->d_sync.ds_consumer->index++;
1962 		/*
1963 		 * Delay the request if it is colliding with a regular request.
1964 		 */
1965 		if (g_mirror_regular_collision(sc, bp))
1966 			g_mirror_sync_delay(sc, bp);
1967 		else
1968 			g_io_request(bp, disk->d_sync.ds_consumer);
1969 	}
1970 }
1971 
1972 /*
1973  * Stop synchronization process.
1974  * type: 0 - synchronization finished
1975  *       1 - synchronization stopped
1976  */
1977 static void
1978 g_mirror_sync_stop(struct g_mirror_disk *disk, int type)
1979 {
1980 	struct g_mirror_softc *sc;
1981 	struct g_consumer *cp;
1982 
1983 	g_topology_assert_not();
1984 	sc = disk->d_softc;
1985 	sx_assert(&sc->sc_lock, SX_LOCKED);
1986 
1987 	KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
1988 	    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
1989 	    g_mirror_disk_state2str(disk->d_state)));
1990 	if (disk->d_sync.ds_consumer == NULL)
1991 		return;
1992 
1993 	if (type == 0) {
1994 		G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s finished.",
1995 		    sc->sc_name, g_mirror_get_diskname(disk));
1996 	} else /* if (type == 1) */ {
1997 		G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s stopped.",
1998 		    sc->sc_name, g_mirror_get_diskname(disk));
1999 	}
2000 	free(disk->d_sync.ds_bios, M_MIRROR);
2001 	disk->d_sync.ds_bios = NULL;
2002 	cp = disk->d_sync.ds_consumer;
2003 	disk->d_sync.ds_consumer = NULL;
2004 	disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2005 	sc->sc_sync.ds_ndisks--;
2006 	sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
2007 	g_topology_lock();
2008 	g_mirror_kill_consumer(sc, cp);
2009 	g_topology_unlock();
2010 	sx_xlock(&sc->sc_lock);
2011 }
2012 
2013 static void
2014 g_mirror_launch_provider(struct g_mirror_softc *sc)
2015 {
2016 	struct g_mirror_disk *disk;
2017 	struct g_provider *pp;
2018 
2019 	sx_assert(&sc->sc_lock, SX_LOCKED);
2020 
2021 	g_topology_lock();
2022 	pp = g_new_providerf(sc->sc_geom, "mirror/%s", sc->sc_name);
2023 	pp->mediasize = sc->sc_mediasize;
2024 	pp->sectorsize = sc->sc_sectorsize;
2025 	pp->stripesize = 0;
2026 	pp->stripeoffset = 0;
2027 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2028 		if (disk->d_consumer && disk->d_consumer->provider &&
2029 		    disk->d_consumer->provider->stripesize > pp->stripesize) {
2030 			pp->stripesize = disk->d_consumer->provider->stripesize;
2031 			pp->stripeoffset = disk->d_consumer->provider->stripeoffset;
2032 		}
2033 	}
2034 	sc->sc_provider = pp;
2035 	g_error_provider(pp, 0);
2036 	g_topology_unlock();
2037 	G_MIRROR_DEBUG(0, "Device %s launched (%u/%u).", pp->name,
2038 	    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE), sc->sc_ndisks);
2039 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2040 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2041 			g_mirror_sync_start(disk);
2042 	}
2043 }
2044 
2045 static void
2046 g_mirror_destroy_provider(struct g_mirror_softc *sc)
2047 {
2048 	struct g_mirror_disk *disk;
2049 	struct bio *bp;
2050 
2051 	g_topology_assert_not();
2052 	KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
2053 	    sc->sc_name));
2054 
2055 	g_topology_lock();
2056 	g_error_provider(sc->sc_provider, ENXIO);
2057 	mtx_lock(&sc->sc_queue_mtx);
2058 	while ((bp = bioq_first(&sc->sc_queue)) != NULL) {
2059 		bioq_remove(&sc->sc_queue, bp);
2060 		g_io_deliver(bp, ENXIO);
2061 	}
2062 	mtx_unlock(&sc->sc_queue_mtx);
2063 	G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.", sc->sc_name,
2064 	    sc->sc_provider->name);
2065 	sc->sc_provider->flags |= G_PF_WITHER;
2066 	g_orphan_provider(sc->sc_provider, ENXIO);
2067 	g_topology_unlock();
2068 	sc->sc_provider = NULL;
2069 	LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2070 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2071 			g_mirror_sync_stop(disk, 1);
2072 	}
2073 }
2074 
2075 static void
2076 g_mirror_go(void *arg)
2077 {
2078 	struct g_mirror_softc *sc;
2079 
2080 	sc = arg;
2081 	G_MIRROR_DEBUG(0, "Force device %s start due to timeout.", sc->sc_name);
2082 	g_mirror_event_send(sc, 0,
2083 	    G_MIRROR_EVENT_DONTWAIT | G_MIRROR_EVENT_DEVICE);
2084 }
2085 
2086 static u_int
2087 g_mirror_determine_state(struct g_mirror_disk *disk)
2088 {
2089 	struct g_mirror_softc *sc;
2090 	u_int state;
2091 
2092 	sc = disk->d_softc;
2093 	if (sc->sc_syncid == disk->d_sync.ds_syncid) {
2094 		if ((disk->d_flags &
2095 		    G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2096 			/* Disk does not need synchronization. */
2097 			state = G_MIRROR_DISK_STATE_ACTIVE;
2098 		} else {
2099 			if ((sc->sc_flags &
2100 			     G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2101 			    (disk->d_flags &
2102 			     G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2103 				/*
2104 				 * We can start synchronization from
2105 				 * the stored offset.
2106 				 */
2107 				state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2108 			} else {
2109 				state = G_MIRROR_DISK_STATE_STALE;
2110 			}
2111 		}
2112 	} else if (disk->d_sync.ds_syncid < sc->sc_syncid) {
2113 		/*
2114 		 * Reset all synchronization data for this disk,
2115 		 * because if it even was synchronized, it was
2116 		 * synchronized to disks with different syncid.
2117 		 */
2118 		disk->d_flags |= G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2119 		disk->d_sync.ds_offset = 0;
2120 		disk->d_sync.ds_offset_done = 0;
2121 		disk->d_sync.ds_syncid = sc->sc_syncid;
2122 		if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2123 		    (disk->d_flags & G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2124 			state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2125 		} else {
2126 			state = G_MIRROR_DISK_STATE_STALE;
2127 		}
2128 	} else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
2129 		/*
2130 		 * Not good, NOT GOOD!
2131 		 * It means that mirror was started on stale disks
2132 		 * and more fresh disk just arrive.
2133 		 * If there were writes, mirror is broken, sorry.
2134 		 * I think the best choice here is don't touch
2135 		 * this disk and inform the user loudly.
2136 		 */
2137 		G_MIRROR_DEBUG(0, "Device %s was started before the freshest "
2138 		    "disk (%s) arrives!! It will not be connected to the "
2139 		    "running device.", sc->sc_name,
2140 		    g_mirror_get_diskname(disk));
2141 		g_mirror_destroy_disk(disk);
2142 		state = G_MIRROR_DISK_STATE_NONE;
2143 		/* Return immediately, because disk was destroyed. */
2144 		return (state);
2145 	}
2146 	G_MIRROR_DEBUG(3, "State for %s disk: %s.",
2147 	    g_mirror_get_diskname(disk), g_mirror_disk_state2str(state));
2148 	return (state);
2149 }
2150 
2151 /*
2152  * Update device state.
2153  */
2154 static void
2155 g_mirror_update_device(struct g_mirror_softc *sc, boolean_t force)
2156 {
2157 	struct g_mirror_disk *disk;
2158 	u_int state;
2159 
2160 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2161 
2162 	switch (sc->sc_state) {
2163 	case G_MIRROR_DEVICE_STATE_STARTING:
2164 	    {
2165 		struct g_mirror_disk *pdisk, *tdisk;
2166 		u_int dirty, ndisks, genid, syncid;
2167 
2168 		KASSERT(sc->sc_provider == NULL,
2169 		    ("Non-NULL provider in STARTING state (%s).", sc->sc_name));
2170 		/*
2171 		 * Are we ready? We are, if all disks are connected or
2172 		 * if we have any disks and 'force' is true.
2173 		 */
2174 		ndisks = g_mirror_ndisks(sc, -1);
2175 		if (sc->sc_ndisks == ndisks || (force && ndisks > 0)) {
2176 			;
2177 		} else if (ndisks == 0) {
2178 			/*
2179 			 * Disks went down in starting phase, so destroy
2180 			 * device.
2181 			 */
2182 			callout_drain(&sc->sc_callout);
2183 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2184 			G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2185 			    sc->sc_rootmount);
2186 			root_mount_rel(sc->sc_rootmount);
2187 			sc->sc_rootmount = NULL;
2188 			return;
2189 		} else {
2190 			return;
2191 		}
2192 
2193 		/*
2194 		 * Activate all disks with the biggest syncid.
2195 		 */
2196 		if (force) {
2197 			/*
2198 			 * If 'force' is true, we have been called due to
2199 			 * timeout, so don't bother canceling timeout.
2200 			 */
2201 			ndisks = 0;
2202 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2203 				if ((disk->d_flags &
2204 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2205 					ndisks++;
2206 				}
2207 			}
2208 			if (ndisks == 0) {
2209 				/* No valid disks found, destroy device. */
2210 				sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2211 				G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2212 				    __LINE__, sc->sc_rootmount);
2213 				root_mount_rel(sc->sc_rootmount);
2214 				sc->sc_rootmount = NULL;
2215 				return;
2216 			}
2217 		} else {
2218 			/* Cancel timeout. */
2219 			callout_drain(&sc->sc_callout);
2220 		}
2221 
2222 		/*
2223 		 * Find the biggest genid.
2224 		 */
2225 		genid = 0;
2226 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2227 			if (disk->d_genid > genid)
2228 				genid = disk->d_genid;
2229 		}
2230 		sc->sc_genid = genid;
2231 		/*
2232 		 * Remove all disks without the biggest genid.
2233 		 */
2234 		LIST_FOREACH_SAFE(disk, &sc->sc_disks, d_next, tdisk) {
2235 			if (disk->d_genid < genid) {
2236 				G_MIRROR_DEBUG(0,
2237 				    "Component %s (device %s) broken, skipping.",
2238 				    g_mirror_get_diskname(disk), sc->sc_name);
2239 				g_mirror_destroy_disk(disk);
2240 			}
2241 		}
2242 
2243 		/*
2244 		 * Find the biggest syncid.
2245 		 */
2246 		syncid = 0;
2247 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2248 			if (disk->d_sync.ds_syncid > syncid)
2249 				syncid = disk->d_sync.ds_syncid;
2250 		}
2251 
2252 		/*
2253 		 * Here we need to look for dirty disks and if all disks
2254 		 * with the biggest syncid are dirty, we have to choose
2255 		 * one with the biggest priority and rebuild the rest.
2256 		 */
2257 		/*
2258 		 * Find the number of dirty disks with the biggest syncid.
2259 		 * Find the number of disks with the biggest syncid.
2260 		 * While here, find a disk with the biggest priority.
2261 		 */
2262 		dirty = ndisks = 0;
2263 		pdisk = NULL;
2264 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2265 			if (disk->d_sync.ds_syncid != syncid)
2266 				continue;
2267 			if ((disk->d_flags &
2268 			    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2269 				continue;
2270 			}
2271 			ndisks++;
2272 			if ((disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
2273 				dirty++;
2274 				if (pdisk == NULL ||
2275 				    pdisk->d_priority < disk->d_priority) {
2276 					pdisk = disk;
2277 				}
2278 			}
2279 		}
2280 		if (dirty == 0) {
2281 			/* No dirty disks at all, great. */
2282 		} else if (dirty == ndisks) {
2283 			/*
2284 			 * Force synchronization for all dirty disks except one
2285 			 * with the biggest priority.
2286 			 */
2287 			KASSERT(pdisk != NULL, ("pdisk == NULL"));
2288 			G_MIRROR_DEBUG(1, "Using disk %s (device %s) as a "
2289 			    "master disk for synchronization.",
2290 			    g_mirror_get_diskname(pdisk), sc->sc_name);
2291 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2292 				if (disk->d_sync.ds_syncid != syncid)
2293 					continue;
2294 				if ((disk->d_flags &
2295 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2296 					continue;
2297 				}
2298 				KASSERT((disk->d_flags &
2299 				    G_MIRROR_DISK_FLAG_DIRTY) != 0,
2300 				    ("Disk %s isn't marked as dirty.",
2301 				    g_mirror_get_diskname(disk)));
2302 				/* Skip the disk with the biggest priority. */
2303 				if (disk == pdisk)
2304 					continue;
2305 				disk->d_sync.ds_syncid = 0;
2306 			}
2307 		} else if (dirty < ndisks) {
2308 			/*
2309 			 * Force synchronization for all dirty disks.
2310 			 * We have some non-dirty disks.
2311 			 */
2312 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2313 				if (disk->d_sync.ds_syncid != syncid)
2314 					continue;
2315 				if ((disk->d_flags &
2316 				    G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2317 					continue;
2318 				}
2319 				if ((disk->d_flags &
2320 				    G_MIRROR_DISK_FLAG_DIRTY) == 0) {
2321 					continue;
2322 				}
2323 				disk->d_sync.ds_syncid = 0;
2324 			}
2325 		}
2326 
2327 		/* Reset hint. */
2328 		sc->sc_hint = NULL;
2329 		sc->sc_syncid = syncid;
2330 		if (force) {
2331 			/* Remember to bump syncid on first write. */
2332 			sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2333 		}
2334 		state = G_MIRROR_DEVICE_STATE_RUNNING;
2335 		G_MIRROR_DEBUG(1, "Device %s state changed from %s to %s.",
2336 		    sc->sc_name, g_mirror_device_state2str(sc->sc_state),
2337 		    g_mirror_device_state2str(state));
2338 		sc->sc_state = state;
2339 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2340 			state = g_mirror_determine_state(disk);
2341 			g_mirror_event_send(disk, state,
2342 			    G_MIRROR_EVENT_DONTWAIT);
2343 			if (state == G_MIRROR_DISK_STATE_STALE)
2344 				sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2345 		}
2346 		break;
2347 	    }
2348 	case G_MIRROR_DEVICE_STATE_RUNNING:
2349 		if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 0 &&
2350 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2351 			/*
2352 			 * No active disks or no disks at all,
2353 			 * so destroy device.
2354 			 */
2355 			if (sc->sc_provider != NULL)
2356 				g_mirror_destroy_provider(sc);
2357 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2358 			break;
2359 		} else if (g_mirror_ndisks(sc,
2360 		    G_MIRROR_DISK_STATE_ACTIVE) > 0 &&
2361 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2362 			/*
2363 			 * We have active disks, launch provider if it doesn't
2364 			 * exist.
2365 			 */
2366 			if (sc->sc_provider == NULL)
2367 				g_mirror_launch_provider(sc);
2368 			if (sc->sc_rootmount != NULL) {
2369 				G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2370 				    __LINE__, sc->sc_rootmount);
2371 				root_mount_rel(sc->sc_rootmount);
2372 				sc->sc_rootmount = NULL;
2373 			}
2374 		}
2375 		/*
2376 		 * Genid should be bumped immediately, so do it here.
2377 		 */
2378 		if ((sc->sc_bump_id & G_MIRROR_BUMP_GENID) != 0) {
2379 			sc->sc_bump_id &= ~G_MIRROR_BUMP_GENID;
2380 			g_mirror_bump_genid(sc);
2381 		}
2382 		break;
2383 	default:
2384 		KASSERT(1 == 0, ("Wrong device state (%s, %s).",
2385 		    sc->sc_name, g_mirror_device_state2str(sc->sc_state)));
2386 		break;
2387 	}
2388 }
2389 
2390 /*
2391  * Update disk state and device state if needed.
2392  */
2393 #define	DISK_STATE_CHANGED()	G_MIRROR_DEBUG(1,			\
2394 	"Disk %s state changed from %s to %s (device %s).",		\
2395 	g_mirror_get_diskname(disk),					\
2396 	g_mirror_disk_state2str(disk->d_state),				\
2397 	g_mirror_disk_state2str(state), sc->sc_name)
2398 static int
2399 g_mirror_update_disk(struct g_mirror_disk *disk, u_int state)
2400 {
2401 	struct g_mirror_softc *sc;
2402 
2403 	sc = disk->d_softc;
2404 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2405 
2406 again:
2407 	G_MIRROR_DEBUG(3, "Changing disk %s state from %s to %s.",
2408 	    g_mirror_get_diskname(disk), g_mirror_disk_state2str(disk->d_state),
2409 	    g_mirror_disk_state2str(state));
2410 	switch (state) {
2411 	case G_MIRROR_DISK_STATE_NEW:
2412 		/*
2413 		 * Possible scenarios:
2414 		 * 1. New disk arrive.
2415 		 */
2416 		/* Previous state should be NONE. */
2417 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NONE,
2418 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2419 		    g_mirror_disk_state2str(disk->d_state)));
2420 		DISK_STATE_CHANGED();
2421 
2422 		disk->d_state = state;
2423 		if (LIST_EMPTY(&sc->sc_disks))
2424 			LIST_INSERT_HEAD(&sc->sc_disks, disk, d_next);
2425 		else {
2426 			struct g_mirror_disk *dp;
2427 
2428 			LIST_FOREACH(dp, &sc->sc_disks, d_next) {
2429 				if (disk->d_priority >= dp->d_priority) {
2430 					LIST_INSERT_BEFORE(dp, disk, d_next);
2431 					dp = NULL;
2432 					break;
2433 				}
2434 				if (LIST_NEXT(dp, d_next) == NULL)
2435 					break;
2436 			}
2437 			if (dp != NULL)
2438 				LIST_INSERT_AFTER(dp, disk, d_next);
2439 		}
2440 		G_MIRROR_DEBUG(1, "Device %s: provider %s detected.",
2441 		    sc->sc_name, g_mirror_get_diskname(disk));
2442 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
2443 			break;
2444 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2445 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2446 		    g_mirror_device_state2str(sc->sc_state),
2447 		    g_mirror_get_diskname(disk),
2448 		    g_mirror_disk_state2str(disk->d_state)));
2449 		state = g_mirror_determine_state(disk);
2450 		if (state != G_MIRROR_DISK_STATE_NONE)
2451 			goto again;
2452 		break;
2453 	case G_MIRROR_DISK_STATE_ACTIVE:
2454 		/*
2455 		 * Possible scenarios:
2456 		 * 1. New disk does not need synchronization.
2457 		 * 2. Synchronization process finished successfully.
2458 		 */
2459 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2460 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2461 		    g_mirror_device_state2str(sc->sc_state),
2462 		    g_mirror_get_diskname(disk),
2463 		    g_mirror_disk_state2str(disk->d_state)));
2464 		/* Previous state should be NEW or SYNCHRONIZING. */
2465 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW ||
2466 		    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2467 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2468 		    g_mirror_disk_state2str(disk->d_state)));
2469 		DISK_STATE_CHANGED();
2470 
2471 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
2472 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2473 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_FORCE_SYNC;
2474 			g_mirror_sync_stop(disk, 0);
2475 		}
2476 		disk->d_state = state;
2477 		disk->d_sync.ds_offset = 0;
2478 		disk->d_sync.ds_offset_done = 0;
2479 		g_mirror_update_idle(sc, disk);
2480 		g_mirror_update_metadata(disk);
2481 		G_MIRROR_DEBUG(1, "Device %s: provider %s activated.",
2482 		    sc->sc_name, g_mirror_get_diskname(disk));
2483 		break;
2484 	case G_MIRROR_DISK_STATE_STALE:
2485 		/*
2486 		 * Possible scenarios:
2487 		 * 1. Stale disk was connected.
2488 		 */
2489 		/* Previous state should be NEW. */
2490 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2491 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2492 		    g_mirror_disk_state2str(disk->d_state)));
2493 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2494 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2495 		    g_mirror_device_state2str(sc->sc_state),
2496 		    g_mirror_get_diskname(disk),
2497 		    g_mirror_disk_state2str(disk->d_state)));
2498 		/*
2499 		 * STALE state is only possible if device is marked
2500 		 * NOAUTOSYNC.
2501 		 */
2502 		KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) != 0,
2503 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2504 		    g_mirror_device_state2str(sc->sc_state),
2505 		    g_mirror_get_diskname(disk),
2506 		    g_mirror_disk_state2str(disk->d_state)));
2507 		DISK_STATE_CHANGED();
2508 
2509 		disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2510 		disk->d_state = state;
2511 		g_mirror_update_metadata(disk);
2512 		G_MIRROR_DEBUG(0, "Device %s: provider %s is stale.",
2513 		    sc->sc_name, g_mirror_get_diskname(disk));
2514 		break;
2515 	case G_MIRROR_DISK_STATE_SYNCHRONIZING:
2516 		/*
2517 		 * Possible scenarios:
2518 		 * 1. Disk which needs synchronization was connected.
2519 		 */
2520 		/* Previous state should be NEW. */
2521 		KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2522 		    ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2523 		    g_mirror_disk_state2str(disk->d_state)));
2524 		KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2525 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2526 		    g_mirror_device_state2str(sc->sc_state),
2527 		    g_mirror_get_diskname(disk),
2528 		    g_mirror_disk_state2str(disk->d_state)));
2529 		DISK_STATE_CHANGED();
2530 
2531 		if (disk->d_state == G_MIRROR_DISK_STATE_NEW)
2532 			disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2533 		disk->d_state = state;
2534 		if (sc->sc_provider != NULL) {
2535 			g_mirror_sync_start(disk);
2536 			g_mirror_update_metadata(disk);
2537 		}
2538 		break;
2539 	case G_MIRROR_DISK_STATE_DISCONNECTED:
2540 		/*
2541 		 * Possible scenarios:
2542 		 * 1. Device wasn't running yet, but disk disappear.
2543 		 * 2. Disk was active and disapppear.
2544 		 * 3. Disk disappear during synchronization process.
2545 		 */
2546 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING) {
2547 			/*
2548 			 * Previous state should be ACTIVE, STALE or
2549 			 * SYNCHRONIZING.
2550 			 */
2551 			KASSERT(disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
2552 			    disk->d_state == G_MIRROR_DISK_STATE_STALE ||
2553 			    disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2554 			    ("Wrong disk state (%s, %s).",
2555 			    g_mirror_get_diskname(disk),
2556 			    g_mirror_disk_state2str(disk->d_state)));
2557 		} else if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING) {
2558 			/* Previous state should be NEW. */
2559 			KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2560 			    ("Wrong disk state (%s, %s).",
2561 			    g_mirror_get_diskname(disk),
2562 			    g_mirror_disk_state2str(disk->d_state)));
2563 			/*
2564 			 * Reset bumping syncid if disk disappeared in STARTING
2565 			 * state.
2566 			 */
2567 			if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0)
2568 				sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
2569 #ifdef	INVARIANTS
2570 		} else {
2571 			KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
2572 			    sc->sc_name,
2573 			    g_mirror_device_state2str(sc->sc_state),
2574 			    g_mirror_get_diskname(disk),
2575 			    g_mirror_disk_state2str(disk->d_state)));
2576 #endif
2577 		}
2578 		DISK_STATE_CHANGED();
2579 		G_MIRROR_DEBUG(0, "Device %s: provider %s disconnected.",
2580 		    sc->sc_name, g_mirror_get_diskname(disk));
2581 
2582 		g_mirror_destroy_disk(disk);
2583 		break;
2584 	case G_MIRROR_DISK_STATE_DESTROY:
2585 	    {
2586 		int error;
2587 
2588 		error = g_mirror_clear_metadata(disk);
2589 		if (error != 0)
2590 			return (error);
2591 		DISK_STATE_CHANGED();
2592 		G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.",
2593 		    sc->sc_name, g_mirror_get_diskname(disk));
2594 
2595 		g_mirror_destroy_disk(disk);
2596 		sc->sc_ndisks--;
2597 		LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2598 			g_mirror_update_metadata(disk);
2599 		}
2600 		break;
2601 	    }
2602 	default:
2603 		KASSERT(1 == 0, ("Unknown state (%u).", state));
2604 		break;
2605 	}
2606 	return (0);
2607 }
2608 #undef	DISK_STATE_CHANGED
2609 
2610 int
2611 g_mirror_read_metadata(struct g_consumer *cp, struct g_mirror_metadata *md)
2612 {
2613 	struct g_provider *pp;
2614 	u_char *buf;
2615 	int error;
2616 
2617 	g_topology_assert();
2618 
2619 	error = g_access(cp, 1, 0, 0);
2620 	if (error != 0)
2621 		return (error);
2622 	pp = cp->provider;
2623 	g_topology_unlock();
2624 	/* Metadata are stored on last sector. */
2625 	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
2626 	    &error);
2627 	g_topology_lock();
2628 	g_access(cp, -1, 0, 0);
2629 	if (buf == NULL) {
2630 		G_MIRROR_DEBUG(1, "Cannot read metadata from %s (error=%d).",
2631 		    cp->provider->name, error);
2632 		return (error);
2633 	}
2634 
2635 	/* Decode metadata. */
2636 	error = mirror_metadata_decode(buf, md);
2637 	g_free(buf);
2638 	if (strcmp(md->md_magic, G_MIRROR_MAGIC) != 0)
2639 		return (EINVAL);
2640 	if (md->md_version > G_MIRROR_VERSION) {
2641 		G_MIRROR_DEBUG(0,
2642 		    "Kernel module is too old to handle metadata from %s.",
2643 		    cp->provider->name);
2644 		return (EINVAL);
2645 	}
2646 	if (error != 0) {
2647 		G_MIRROR_DEBUG(1, "MD5 metadata hash mismatch for provider %s.",
2648 		    cp->provider->name);
2649 		return (error);
2650 	}
2651 
2652 	return (0);
2653 }
2654 
2655 static int
2656 g_mirror_check_metadata(struct g_mirror_softc *sc, struct g_provider *pp,
2657     struct g_mirror_metadata *md)
2658 {
2659 
2660 	if (g_mirror_id2disk(sc, md->md_did) != NULL) {
2661 		G_MIRROR_DEBUG(1, "Disk %s (id=%u) already exists, skipping.",
2662 		    pp->name, md->md_did);
2663 		return (EEXIST);
2664 	}
2665 	if (md->md_all != sc->sc_ndisks) {
2666 		G_MIRROR_DEBUG(1,
2667 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2668 		    "md_all", pp->name, sc->sc_name);
2669 		return (EINVAL);
2670 	}
2671 	if (md->md_slice != sc->sc_slice) {
2672 		G_MIRROR_DEBUG(1,
2673 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2674 		    "md_slice", pp->name, sc->sc_name);
2675 		return (EINVAL);
2676 	}
2677 	if (md->md_balance != sc->sc_balance) {
2678 		G_MIRROR_DEBUG(1,
2679 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2680 		    "md_balance", pp->name, sc->sc_name);
2681 		return (EINVAL);
2682 	}
2683 	if (md->md_mediasize != sc->sc_mediasize) {
2684 		G_MIRROR_DEBUG(1,
2685 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2686 		    "md_mediasize", pp->name, sc->sc_name);
2687 		return (EINVAL);
2688 	}
2689 	if (sc->sc_mediasize > pp->mediasize) {
2690 		G_MIRROR_DEBUG(1,
2691 		    "Invalid size of disk %s (device %s), skipping.", pp->name,
2692 		    sc->sc_name);
2693 		return (EINVAL);
2694 	}
2695 	if (md->md_sectorsize != sc->sc_sectorsize) {
2696 		G_MIRROR_DEBUG(1,
2697 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2698 		    "md_sectorsize", pp->name, sc->sc_name);
2699 		return (EINVAL);
2700 	}
2701 	if ((sc->sc_sectorsize % pp->sectorsize) != 0) {
2702 		G_MIRROR_DEBUG(1,
2703 		    "Invalid sector size of disk %s (device %s), skipping.",
2704 		    pp->name, sc->sc_name);
2705 		return (EINVAL);
2706 	}
2707 	if ((md->md_mflags & ~G_MIRROR_DEVICE_FLAG_MASK) != 0) {
2708 		G_MIRROR_DEBUG(1,
2709 		    "Invalid device flags on disk %s (device %s), skipping.",
2710 		    pp->name, sc->sc_name);
2711 		return (EINVAL);
2712 	}
2713 	if ((md->md_dflags & ~G_MIRROR_DISK_FLAG_MASK) != 0) {
2714 		G_MIRROR_DEBUG(1,
2715 		    "Invalid disk flags on disk %s (device %s), skipping.",
2716 		    pp->name, sc->sc_name);
2717 		return (EINVAL);
2718 	}
2719 	return (0);
2720 }
2721 
2722 int
2723 g_mirror_add_disk(struct g_mirror_softc *sc, struct g_provider *pp,
2724     struct g_mirror_metadata *md)
2725 {
2726 	struct g_mirror_disk *disk;
2727 	int error;
2728 
2729 	g_topology_assert_not();
2730 	G_MIRROR_DEBUG(2, "Adding disk %s.", pp->name);
2731 
2732 	error = g_mirror_check_metadata(sc, pp, md);
2733 	if (error != 0)
2734 		return (error);
2735 	if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING &&
2736 	    md->md_genid < sc->sc_genid) {
2737 		G_MIRROR_DEBUG(0, "Component %s (device %s) broken, skipping.",
2738 		    pp->name, sc->sc_name);
2739 		return (EINVAL);
2740 	}
2741 	disk = g_mirror_init_disk(sc, pp, md, &error);
2742 	if (disk == NULL)
2743 		return (error);
2744 	error = g_mirror_event_send(disk, G_MIRROR_DISK_STATE_NEW,
2745 	    G_MIRROR_EVENT_WAIT);
2746 	if (error != 0)
2747 		return (error);
2748 	if (md->md_version < G_MIRROR_VERSION) {
2749 		G_MIRROR_DEBUG(0, "Upgrading metadata on %s (v%d->v%d).",
2750 		    pp->name, md->md_version, G_MIRROR_VERSION);
2751 		g_mirror_update_metadata(disk);
2752 	}
2753 	return (0);
2754 }
2755 
2756 static void
2757 g_mirror_destroy_delayed(void *arg, int flag)
2758 {
2759 	struct g_mirror_softc *sc;
2760 	int error;
2761 
2762 	if (flag == EV_CANCEL) {
2763 		G_MIRROR_DEBUG(1, "Destroying canceled.");
2764 		return;
2765 	}
2766 	sc = arg;
2767 	g_topology_unlock();
2768 	sx_xlock(&sc->sc_lock);
2769 	KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) == 0,
2770 	    ("DESTROY flag set on %s.", sc->sc_name));
2771 	KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROYING) != 0,
2772 	    ("DESTROYING flag not set on %s.", sc->sc_name));
2773 	G_MIRROR_DEBUG(1, "Destroying %s (delayed).", sc->sc_name);
2774 	error = g_mirror_destroy(sc, G_MIRROR_DESTROY_SOFT);
2775 	if (error != 0) {
2776 		G_MIRROR_DEBUG(0, "Cannot destroy %s.", sc->sc_name);
2777 		sx_xunlock(&sc->sc_lock);
2778 	}
2779 	g_topology_lock();
2780 }
2781 
2782 static int
2783 g_mirror_access(struct g_provider *pp, int acr, int acw, int ace)
2784 {
2785 	struct g_mirror_softc *sc;
2786 	int dcr, dcw, dce, error = 0;
2787 
2788 	g_topology_assert();
2789 	G_MIRROR_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, acr,
2790 	    acw, ace);
2791 
2792 	sc = pp->geom->softc;
2793 	if (sc == NULL && acr <= 0 && acw <= 0 && ace <= 0)
2794 		return (0);
2795 	KASSERT(sc != NULL, ("NULL softc (provider=%s).", pp->name));
2796 
2797 	dcr = pp->acr + acr;
2798 	dcw = pp->acw + acw;
2799 	dce = pp->ace + ace;
2800 
2801 	g_topology_unlock();
2802 	sx_xlock(&sc->sc_lock);
2803 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0 ||
2804 	    LIST_EMPTY(&sc->sc_disks)) {
2805 		if (acr > 0 || acw > 0 || ace > 0)
2806 			error = ENXIO;
2807 		goto end;
2808 	}
2809 	if (dcw == 0 && !sc->sc_idle)
2810 		g_mirror_idle(sc, dcw);
2811 	if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROYING) != 0) {
2812 		if (acr > 0 || acw > 0 || ace > 0) {
2813 			error = ENXIO;
2814 			goto end;
2815 		}
2816 		if (dcr == 0 && dcw == 0 && dce == 0) {
2817 			g_post_event(g_mirror_destroy_delayed, sc, M_WAITOK,
2818 			    sc, NULL);
2819 		}
2820 	}
2821 end:
2822 	sx_xunlock(&sc->sc_lock);
2823 	g_topology_lock();
2824 	return (error);
2825 }
2826 
2827 static struct g_geom *
2828 g_mirror_create(struct g_class *mp, const struct g_mirror_metadata *md)
2829 {
2830 	struct g_mirror_softc *sc;
2831 	struct g_geom *gp;
2832 	int error, timeout;
2833 
2834 	g_topology_assert();
2835 	G_MIRROR_DEBUG(1, "Creating device %s (id=%u).", md->md_name,
2836 	    md->md_mid);
2837 
2838 	/* One disk is minimum. */
2839 	if (md->md_all < 1)
2840 		return (NULL);
2841 	/*
2842 	 * Action geom.
2843 	 */
2844 	gp = g_new_geomf(mp, "%s", md->md_name);
2845 	sc = malloc(sizeof(*sc), M_MIRROR, M_WAITOK | M_ZERO);
2846 	gp->start = g_mirror_start;
2847 	gp->orphan = g_mirror_orphan;
2848 	gp->access = g_mirror_access;
2849 	gp->dumpconf = g_mirror_dumpconf;
2850 
2851 	sc->sc_id = md->md_mid;
2852 	sc->sc_slice = md->md_slice;
2853 	sc->sc_balance = md->md_balance;
2854 	sc->sc_mediasize = md->md_mediasize;
2855 	sc->sc_sectorsize = md->md_sectorsize;
2856 	sc->sc_ndisks = md->md_all;
2857 	sc->sc_flags = md->md_mflags;
2858 	sc->sc_bump_id = 0;
2859 	sc->sc_idle = 1;
2860 	sc->sc_last_write = time_uptime;
2861 	sc->sc_writes = 0;
2862 	sx_init(&sc->sc_lock, "gmirror:lock");
2863 	bioq_init(&sc->sc_queue);
2864 	mtx_init(&sc->sc_queue_mtx, "gmirror:queue", NULL, MTX_DEF);
2865 	bioq_init(&sc->sc_regular_delayed);
2866 	bioq_init(&sc->sc_inflight);
2867 	bioq_init(&sc->sc_sync_delayed);
2868 	LIST_INIT(&sc->sc_disks);
2869 	TAILQ_INIT(&sc->sc_events);
2870 	mtx_init(&sc->sc_events_mtx, "gmirror:events", NULL, MTX_DEF);
2871 	callout_init(&sc->sc_callout, CALLOUT_MPSAFE);
2872 	sc->sc_state = G_MIRROR_DEVICE_STATE_STARTING;
2873 	gp->softc = sc;
2874 	sc->sc_geom = gp;
2875 	sc->sc_provider = NULL;
2876 	/*
2877 	 * Synchronization geom.
2878 	 */
2879 	gp = g_new_geomf(mp, "%s.sync", md->md_name);
2880 	gp->softc = sc;
2881 	gp->orphan = g_mirror_orphan;
2882 	sc->sc_sync.ds_geom = gp;
2883 	sc->sc_sync.ds_ndisks = 0;
2884 	error = kproc_create(g_mirror_worker, sc, &sc->sc_worker, 0, 0,
2885 	    "g_mirror %s", md->md_name);
2886 	if (error != 0) {
2887 		G_MIRROR_DEBUG(1, "Cannot create kernel thread for %s.",
2888 		    sc->sc_name);
2889 		g_destroy_geom(sc->sc_sync.ds_geom);
2890 		mtx_destroy(&sc->sc_events_mtx);
2891 		mtx_destroy(&sc->sc_queue_mtx);
2892 		sx_destroy(&sc->sc_lock);
2893 		g_destroy_geom(sc->sc_geom);
2894 		free(sc, M_MIRROR);
2895 		return (NULL);
2896 	}
2897 
2898 	G_MIRROR_DEBUG(1, "Device %s created (%u components, id=%u).",
2899 	    sc->sc_name, sc->sc_ndisks, sc->sc_id);
2900 
2901 	sc->sc_rootmount = root_mount_hold("GMIRROR");
2902 	G_MIRROR_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
2903 	/*
2904 	 * Run timeout.
2905 	 */
2906 	timeout = g_mirror_timeout * hz;
2907 	callout_reset(&sc->sc_callout, timeout, g_mirror_go, sc);
2908 	return (sc->sc_geom);
2909 }
2910 
2911 int
2912 g_mirror_destroy(struct g_mirror_softc *sc, int how)
2913 {
2914 	struct g_mirror_disk *disk;
2915 	struct g_provider *pp;
2916 
2917 	g_topology_assert_not();
2918 	if (sc == NULL)
2919 		return (ENXIO);
2920 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2921 
2922 	pp = sc->sc_provider;
2923 	if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
2924 		switch (how) {
2925 		case G_MIRROR_DESTROY_SOFT:
2926 			G_MIRROR_DEBUG(1,
2927 			    "Device %s is still open (r%dw%de%d).", pp->name,
2928 			    pp->acr, pp->acw, pp->ace);
2929 			return (EBUSY);
2930 		case G_MIRROR_DESTROY_DELAYED:
2931 			G_MIRROR_DEBUG(1,
2932 			    "Device %s will be destroyed on last close.",
2933 			    pp->name);
2934 			LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2935 				if (disk->d_state ==
2936 				    G_MIRROR_DISK_STATE_SYNCHRONIZING) {
2937 					g_mirror_sync_stop(disk, 1);
2938 				}
2939 			}
2940 			sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROYING;
2941 			return (EBUSY);
2942 		case G_MIRROR_DESTROY_HARD:
2943 			G_MIRROR_DEBUG(1, "Device %s is still open, so it "
2944 			    "can't be definitely removed.", pp->name);
2945 		}
2946 	}
2947 
2948 	g_topology_lock();
2949 	if (sc->sc_geom->softc == NULL) {
2950 		g_topology_unlock();
2951 		return (0);
2952 	}
2953 	sc->sc_geom->softc = NULL;
2954 	sc->sc_sync.ds_geom->softc = NULL;
2955 	g_topology_unlock();
2956 
2957 	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2958 	sc->sc_flags |= G_MIRROR_DEVICE_FLAG_WAIT;
2959 	G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
2960 	sx_xunlock(&sc->sc_lock);
2961 	mtx_lock(&sc->sc_queue_mtx);
2962 	wakeup(sc);
2963 	mtx_unlock(&sc->sc_queue_mtx);
2964 	G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
2965 	while (sc->sc_worker != NULL)
2966 		tsleep(&sc->sc_worker, PRIBIO, "m:destroy", hz / 5);
2967 	G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
2968 	sx_xlock(&sc->sc_lock);
2969 	g_mirror_destroy_device(sc);
2970 	free(sc, M_MIRROR);
2971 	return (0);
2972 }
2973 
2974 static void
2975 g_mirror_taste_orphan(struct g_consumer *cp)
2976 {
2977 
2978 	KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
2979 	    cp->provider->name));
2980 }
2981 
2982 static struct g_geom *
2983 g_mirror_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
2984 {
2985 	struct g_mirror_metadata md;
2986 	struct g_mirror_softc *sc;
2987 	struct g_consumer *cp;
2988 	struct g_geom *gp;
2989 	int error;
2990 
2991 	g_topology_assert();
2992 	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
2993 	G_MIRROR_DEBUG(2, "Tasting %s.", pp->name);
2994 
2995 	gp = g_new_geomf(mp, "mirror:taste");
2996 	/*
2997 	 * This orphan function should be never called.
2998 	 */
2999 	gp->orphan = g_mirror_taste_orphan;
3000 	cp = g_new_consumer(gp);
3001 	g_attach(cp, pp);
3002 	error = g_mirror_read_metadata(cp, &md);
3003 	g_detach(cp);
3004 	g_destroy_consumer(cp);
3005 	g_destroy_geom(gp);
3006 	if (error != 0)
3007 		return (NULL);
3008 	gp = NULL;
3009 
3010 	if (md.md_provider[0] != '\0' &&
3011 	    !g_compare_names(md.md_provider, pp->name))
3012 		return (NULL);
3013 	if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
3014 		return (NULL);
3015 	if ((md.md_dflags & G_MIRROR_DISK_FLAG_INACTIVE) != 0) {
3016 		G_MIRROR_DEBUG(0,
3017 		    "Device %s: provider %s marked as inactive, skipping.",
3018 		    md.md_name, pp->name);
3019 		return (NULL);
3020 	}
3021 	if (g_mirror_debug >= 2)
3022 		mirror_metadata_dump(&md);
3023 
3024 	/*
3025 	 * Let's check if device already exists.
3026 	 */
3027 	sc = NULL;
3028 	LIST_FOREACH(gp, &mp->geom, geom) {
3029 		sc = gp->softc;
3030 		if (sc == NULL)
3031 			continue;
3032 		if (sc->sc_sync.ds_geom == gp)
3033 			continue;
3034 		if (strcmp(md.md_name, sc->sc_name) != 0)
3035 			continue;
3036 		if (md.md_mid != sc->sc_id) {
3037 			G_MIRROR_DEBUG(0, "Device %s already configured.",
3038 			    sc->sc_name);
3039 			return (NULL);
3040 		}
3041 		break;
3042 	}
3043 	if (gp == NULL) {
3044 		gp = g_mirror_create(mp, &md);
3045 		if (gp == NULL) {
3046 			G_MIRROR_DEBUG(0, "Cannot create device %s.",
3047 			    md.md_name);
3048 			return (NULL);
3049 		}
3050 		sc = gp->softc;
3051 	}
3052 	G_MIRROR_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
3053 	g_topology_unlock();
3054 	sx_xlock(&sc->sc_lock);
3055 	error = g_mirror_add_disk(sc, pp, &md);
3056 	if (error != 0) {
3057 		G_MIRROR_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
3058 		    pp->name, gp->name, error);
3059 		if (LIST_EMPTY(&sc->sc_disks)) {
3060 			g_cancel_event(sc);
3061 			g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3062 			g_topology_lock();
3063 			return (NULL);
3064 		}
3065 		gp = NULL;
3066 	}
3067 	sx_xunlock(&sc->sc_lock);
3068 	g_topology_lock();
3069 	return (gp);
3070 }
3071 
3072 static int
3073 g_mirror_destroy_geom(struct gctl_req *req __unused,
3074     struct g_class *mp __unused, struct g_geom *gp)
3075 {
3076 	struct g_mirror_softc *sc;
3077 	int error;
3078 
3079 	g_topology_unlock();
3080 	sc = gp->softc;
3081 	sx_xlock(&sc->sc_lock);
3082 	g_cancel_event(sc);
3083 	error = g_mirror_destroy(gp->softc, G_MIRROR_DESTROY_SOFT);
3084 	if (error != 0)
3085 		sx_xunlock(&sc->sc_lock);
3086 	g_topology_lock();
3087 	return (error);
3088 }
3089 
3090 static void
3091 g_mirror_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
3092     struct g_consumer *cp, struct g_provider *pp)
3093 {
3094 	struct g_mirror_softc *sc;
3095 
3096 	g_topology_assert();
3097 
3098 	sc = gp->softc;
3099 	if (sc == NULL)
3100 		return;
3101 	/* Skip synchronization geom. */
3102 	if (gp == sc->sc_sync.ds_geom)
3103 		return;
3104 	if (pp != NULL) {
3105 		/* Nothing here. */
3106 	} else if (cp != NULL) {
3107 		struct g_mirror_disk *disk;
3108 
3109 		disk = cp->private;
3110 		if (disk == NULL)
3111 			return;
3112 		g_topology_unlock();
3113 		sx_xlock(&sc->sc_lock);
3114 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)disk->d_id);
3115 		if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
3116 			sbuf_printf(sb, "%s<Synchronized>", indent);
3117 			if (disk->d_sync.ds_offset == 0)
3118 				sbuf_printf(sb, "0%%");
3119 			else {
3120 				sbuf_printf(sb, "%u%%",
3121 				    (u_int)((disk->d_sync.ds_offset * 100) /
3122 				    sc->sc_provider->mediasize));
3123 			}
3124 			sbuf_printf(sb, "</Synchronized>\n");
3125 		}
3126 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent,
3127 		    disk->d_sync.ds_syncid);
3128 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent,
3129 		    disk->d_genid);
3130 		sbuf_printf(sb, "%s<Flags>", indent);
3131 		if (disk->d_flags == 0)
3132 			sbuf_printf(sb, "NONE");
3133 		else {
3134 			int first = 1;
3135 
3136 #define	ADD_FLAG(flag, name)	do {					\
3137 	if ((disk->d_flags & (flag)) != 0) {				\
3138 		if (!first)						\
3139 			sbuf_printf(sb, ", ");				\
3140 		else							\
3141 			first = 0;					\
3142 		sbuf_printf(sb, name);					\
3143 	}								\
3144 } while (0)
3145 			ADD_FLAG(G_MIRROR_DISK_FLAG_DIRTY, "DIRTY");
3146 			ADD_FLAG(G_MIRROR_DISK_FLAG_HARDCODED, "HARDCODED");
3147 			ADD_FLAG(G_MIRROR_DISK_FLAG_INACTIVE, "INACTIVE");
3148 			ADD_FLAG(G_MIRROR_DISK_FLAG_SYNCHRONIZING,
3149 			    "SYNCHRONIZING");
3150 			ADD_FLAG(G_MIRROR_DISK_FLAG_FORCE_SYNC, "FORCE_SYNC");
3151 			ADD_FLAG(G_MIRROR_DISK_FLAG_BROKEN, "BROKEN");
3152 #undef	ADD_FLAG
3153 		}
3154 		sbuf_printf(sb, "</Flags>\n");
3155 		sbuf_printf(sb, "%s<Priority>%u</Priority>\n", indent,
3156 		    disk->d_priority);
3157 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3158 		    g_mirror_disk_state2str(disk->d_state));
3159 		sx_xunlock(&sc->sc_lock);
3160 		g_topology_lock();
3161 	} else {
3162 		g_topology_unlock();
3163 		sx_xlock(&sc->sc_lock);
3164 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
3165 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent, sc->sc_syncid);
3166 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, sc->sc_genid);
3167 		sbuf_printf(sb, "%s<Flags>", indent);
3168 		if (sc->sc_flags == 0)
3169 			sbuf_printf(sb, "NONE");
3170 		else {
3171 			int first = 1;
3172 
3173 #define	ADD_FLAG(flag, name)	do {					\
3174 	if ((sc->sc_flags & (flag)) != 0) {				\
3175 		if (!first)						\
3176 			sbuf_printf(sb, ", ");				\
3177 		else							\
3178 			first = 0;					\
3179 		sbuf_printf(sb, name);					\
3180 	}								\
3181 } while (0)
3182 			ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOFAILSYNC, "NOFAILSYNC");
3183 			ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOAUTOSYNC, "NOAUTOSYNC");
3184 #undef	ADD_FLAG
3185 		}
3186 		sbuf_printf(sb, "</Flags>\n");
3187 		sbuf_printf(sb, "%s<Slice>%u</Slice>\n", indent,
3188 		    (u_int)sc->sc_slice);
3189 		sbuf_printf(sb, "%s<Balance>%s</Balance>\n", indent,
3190 		    balance_name(sc->sc_balance));
3191 		sbuf_printf(sb, "%s<Components>%u</Components>\n", indent,
3192 		    sc->sc_ndisks);
3193 		sbuf_printf(sb, "%s<State>", indent);
3194 		if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
3195 			sbuf_printf(sb, "%s", "STARTING");
3196 		else if (sc->sc_ndisks ==
3197 		    g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE))
3198 			sbuf_printf(sb, "%s", "COMPLETE");
3199 		else
3200 			sbuf_printf(sb, "%s", "DEGRADED");
3201 		sbuf_printf(sb, "</State>\n");
3202 		sx_xunlock(&sc->sc_lock);
3203 		g_topology_lock();
3204 	}
3205 }
3206 
3207 static void
3208 g_mirror_shutdown_pre_sync(void *arg, int howto)
3209 {
3210 	struct g_class *mp;
3211 	struct g_geom *gp, *gp2;
3212 	struct g_mirror_softc *sc;
3213 	int error;
3214 
3215 	mp = arg;
3216 	DROP_GIANT();
3217 	g_topology_lock();
3218 	LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
3219 		if ((sc = gp->softc) == NULL)
3220 			continue;
3221 		/* Skip synchronization geom. */
3222 		if (gp == sc->sc_sync.ds_geom)
3223 			continue;
3224 		g_topology_unlock();
3225 		sx_xlock(&sc->sc_lock);
3226 		g_cancel_event(sc);
3227 		error = g_mirror_destroy(sc, G_MIRROR_DESTROY_DELAYED);
3228 		if (error != 0)
3229 			sx_xunlock(&sc->sc_lock);
3230 		g_topology_lock();
3231 	}
3232 	g_topology_unlock();
3233 	PICKUP_GIANT();
3234 }
3235 
3236 static void
3237 g_mirror_init(struct g_class *mp)
3238 {
3239 
3240 	g_mirror_pre_sync = EVENTHANDLER_REGISTER(shutdown_pre_sync,
3241 	    g_mirror_shutdown_pre_sync, mp, SHUTDOWN_PRI_FIRST);
3242 	if (g_mirror_pre_sync == NULL)
3243 		G_MIRROR_DEBUG(0, "Warning! Cannot register shutdown event.");
3244 }
3245 
3246 static void
3247 g_mirror_fini(struct g_class *mp)
3248 {
3249 
3250 	if (g_mirror_pre_sync != NULL)
3251 		EVENTHANDLER_DEREGISTER(shutdown_pre_sync, g_mirror_pre_sync);
3252 }
3253 
3254 DECLARE_GEOM_CLASS(g_mirror_class, g_mirror);
3255