xref: /freebsd/sys/geom/raid3/g_raid3.c (revision 06c3fb2749bda94cb5201f81ffdb8fa6c3161b2e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2004-2006 Pawel Jakub Dawidek <pjd@FreeBSD.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/param.h>
30 #include <sys/systm.h>
31 #include <sys/bio.h>
32 #include <sys/eventhandler.h>
33 #include <sys/kernel.h>
34 #include <sys/kthread.h>
35 #include <sys/limits.h>
36 #include <sys/lock.h>
37 #include <sys/malloc.h>
38 #include <sys/module.h>
39 #include <sys/mutex.h>
40 #include <sys/proc.h>
41 #include <sys/reboot.h>
42 #include <sys/sbuf.h>
43 #include <sys/sched.h>
44 #include <sys/sysctl.h>
45 
46 #include <vm/uma.h>
47 
48 #include <geom/geom.h>
49 #include <geom/geom_dbg.h>
50 #include <geom/raid3/g_raid3.h>
51 
52 FEATURE(geom_raid3, "GEOM RAID-3 functionality");
53 
54 static MALLOC_DEFINE(M_RAID3, "raid3_data", "GEOM_RAID3 Data");
55 
56 SYSCTL_DECL(_kern_geom);
57 static SYSCTL_NODE(_kern_geom, OID_AUTO, raid3, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
58     "GEOM_RAID3 stuff");
59 u_int g_raid3_debug = 0;
60 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, debug, CTLFLAG_RWTUN, &g_raid3_debug, 0,
61     "Debug level");
62 static u_int g_raid3_timeout = 4;
63 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, timeout, CTLFLAG_RWTUN, &g_raid3_timeout,
64     0, "Time to wait on all raid3 components");
65 static u_int g_raid3_idletime = 5;
66 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, idletime, CTLFLAG_RWTUN,
67     &g_raid3_idletime, 0, "Mark components as clean when idling");
68 static u_int g_raid3_disconnect_on_failure = 1;
69 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, disconnect_on_failure, CTLFLAG_RWTUN,
70     &g_raid3_disconnect_on_failure, 0, "Disconnect component on I/O failure.");
71 static u_int g_raid3_syncreqs = 2;
72 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, sync_requests, CTLFLAG_RDTUN,
73     &g_raid3_syncreqs, 0, "Parallel synchronization I/O requests.");
74 static u_int g_raid3_use_malloc = 0;
75 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, use_malloc, CTLFLAG_RDTUN,
76     &g_raid3_use_malloc, 0, "Use malloc(9) instead of uma(9).");
77 
78 static u_int g_raid3_n64k = 50;
79 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, n64k, CTLFLAG_RDTUN, &g_raid3_n64k, 0,
80     "Maximum number of 64kB allocations");
81 static u_int g_raid3_n16k = 200;
82 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, n16k, CTLFLAG_RDTUN, &g_raid3_n16k, 0,
83     "Maximum number of 16kB allocations");
84 static u_int g_raid3_n4k = 1200;
85 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, n4k, CTLFLAG_RDTUN, &g_raid3_n4k, 0,
86     "Maximum number of 4kB allocations");
87 
88 static SYSCTL_NODE(_kern_geom_raid3, OID_AUTO, stat,
89     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
90     "GEOM_RAID3 statistics");
91 static u_int g_raid3_parity_mismatch = 0;
92 SYSCTL_UINT(_kern_geom_raid3_stat, OID_AUTO, parity_mismatch, CTLFLAG_RD,
93     &g_raid3_parity_mismatch, 0, "Number of failures in VERIFY mode");
94 
95 #define	MSLEEP(ident, mtx, priority, wmesg, timeout)	do {		\
96 	G_RAID3_DEBUG(4, "%s: Sleeping %p.", __func__, (ident));	\
97 	msleep((ident), (mtx), (priority), (wmesg), (timeout));		\
98 	G_RAID3_DEBUG(4, "%s: Woken up %p.", __func__, (ident));	\
99 } while (0)
100 
101 static eventhandler_tag g_raid3_post_sync = NULL;
102 static int g_raid3_shutdown = 0;
103 
104 static int g_raid3_destroy_geom(struct gctl_req *req, struct g_class *mp,
105     struct g_geom *gp);
106 static g_taste_t g_raid3_taste;
107 static void g_raid3_init(struct g_class *mp);
108 static void g_raid3_fini(struct g_class *mp);
109 static void g_raid3_providergone(struct g_provider *pp);
110 
111 struct g_class g_raid3_class = {
112 	.name = G_RAID3_CLASS_NAME,
113 	.version = G_VERSION,
114 	.ctlreq = g_raid3_config,
115 	.taste = g_raid3_taste,
116 	.destroy_geom = g_raid3_destroy_geom,
117 	.init = g_raid3_init,
118 	.fini = g_raid3_fini,
119 	.providergone = g_raid3_providergone,
120 };
121 
122 static void g_raid3_destroy_provider(struct g_raid3_softc *sc);
123 static int g_raid3_update_disk(struct g_raid3_disk *disk, u_int state);
124 static void g_raid3_update_device(struct g_raid3_softc *sc, boolean_t force);
125 static void g_raid3_dumpconf(struct sbuf *sb, const char *indent,
126     struct g_geom *gp, struct g_consumer *cp, struct g_provider *pp);
127 static void g_raid3_sync_stop(struct g_raid3_softc *sc, int type);
128 static int g_raid3_register_request(struct bio *pbp);
129 static void g_raid3_sync_release(struct g_raid3_softc *sc);
130 static void g_raid3_timeout_drain(struct g_raid3_softc *sc);
131 
132 static const char *
133 g_raid3_disk_state2str(int state)
134 {
135 
136 	switch (state) {
137 	case G_RAID3_DISK_STATE_NODISK:
138 		return ("NODISK");
139 	case G_RAID3_DISK_STATE_NONE:
140 		return ("NONE");
141 	case G_RAID3_DISK_STATE_NEW:
142 		return ("NEW");
143 	case G_RAID3_DISK_STATE_ACTIVE:
144 		return ("ACTIVE");
145 	case G_RAID3_DISK_STATE_STALE:
146 		return ("STALE");
147 	case G_RAID3_DISK_STATE_SYNCHRONIZING:
148 		return ("SYNCHRONIZING");
149 	case G_RAID3_DISK_STATE_DISCONNECTED:
150 		return ("DISCONNECTED");
151 	default:
152 		return ("INVALID");
153 	}
154 }
155 
156 static const char *
157 g_raid3_device_state2str(int state)
158 {
159 
160 	switch (state) {
161 	case G_RAID3_DEVICE_STATE_STARTING:
162 		return ("STARTING");
163 	case G_RAID3_DEVICE_STATE_DEGRADED:
164 		return ("DEGRADED");
165 	case G_RAID3_DEVICE_STATE_COMPLETE:
166 		return ("COMPLETE");
167 	default:
168 		return ("INVALID");
169 	}
170 }
171 
172 const char *
173 g_raid3_get_diskname(struct g_raid3_disk *disk)
174 {
175 
176 	if (disk->d_consumer == NULL || disk->d_consumer->provider == NULL)
177 		return ("[unknown]");
178 	return (disk->d_name);
179 }
180 
181 static void *
182 g_raid3_alloc(struct g_raid3_softc *sc, size_t size, int flags)
183 {
184 	void *ptr;
185 	enum g_raid3_zones zone;
186 
187 	if (g_raid3_use_malloc ||
188 	    (zone = g_raid3_zone(size)) == G_RAID3_NUM_ZONES)
189 		ptr = malloc(size, M_RAID3, flags);
190 	else {
191 		ptr = uma_zalloc_arg(sc->sc_zones[zone].sz_zone,
192 		   &sc->sc_zones[zone], flags);
193 		sc->sc_zones[zone].sz_requested++;
194 		if (ptr == NULL)
195 			sc->sc_zones[zone].sz_failed++;
196 	}
197 	return (ptr);
198 }
199 
200 static void
201 g_raid3_free(struct g_raid3_softc *sc, void *ptr, size_t size)
202 {
203 	enum g_raid3_zones zone;
204 
205 	if (g_raid3_use_malloc ||
206 	    (zone = g_raid3_zone(size)) == G_RAID3_NUM_ZONES)
207 		free(ptr, M_RAID3);
208 	else {
209 		uma_zfree_arg(sc->sc_zones[zone].sz_zone,
210 		    ptr, &sc->sc_zones[zone]);
211 	}
212 }
213 
214 static int
215 g_raid3_uma_ctor(void *mem, int size, void *arg, int flags)
216 {
217 	struct g_raid3_zone *sz = arg;
218 
219 	if (sz->sz_max > 0 && sz->sz_inuse == sz->sz_max)
220 		return (ENOMEM);
221 	sz->sz_inuse++;
222 	return (0);
223 }
224 
225 static void
226 g_raid3_uma_dtor(void *mem, int size, void *arg)
227 {
228 	struct g_raid3_zone *sz = arg;
229 
230 	sz->sz_inuse--;
231 }
232 
233 #define	g_raid3_xor(src, dst, size)					\
234 	_g_raid3_xor((uint64_t *)(src),					\
235 	    (uint64_t *)(dst), (size_t)size)
236 static void
237 _g_raid3_xor(uint64_t *src, uint64_t *dst, size_t size)
238 {
239 
240 	KASSERT((size % 128) == 0, ("Invalid size: %zu.", size));
241 	for (; size > 0; size -= 128) {
242 		*dst++ ^= (*src++);
243 		*dst++ ^= (*src++);
244 		*dst++ ^= (*src++);
245 		*dst++ ^= (*src++);
246 		*dst++ ^= (*src++);
247 		*dst++ ^= (*src++);
248 		*dst++ ^= (*src++);
249 		*dst++ ^= (*src++);
250 		*dst++ ^= (*src++);
251 		*dst++ ^= (*src++);
252 		*dst++ ^= (*src++);
253 		*dst++ ^= (*src++);
254 		*dst++ ^= (*src++);
255 		*dst++ ^= (*src++);
256 		*dst++ ^= (*src++);
257 		*dst++ ^= (*src++);
258 	}
259 }
260 
261 static int
262 g_raid3_is_zero(struct bio *bp)
263 {
264 	static const uint64_t zeros[] = {
265 	    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
266 	};
267 	u_char *addr;
268 	ssize_t size;
269 
270 	size = bp->bio_length;
271 	addr = (u_char *)bp->bio_data;
272 	for (; size > 0; size -= sizeof(zeros), addr += sizeof(zeros)) {
273 		if (bcmp(addr, zeros, sizeof(zeros)) != 0)
274 			return (0);
275 	}
276 	return (1);
277 }
278 
279 /*
280  * --- Events handling functions ---
281  * Events in geom_raid3 are used to maintain disks and device status
282  * from one thread to simplify locking.
283  */
284 static void
285 g_raid3_event_free(struct g_raid3_event *ep)
286 {
287 
288 	free(ep, M_RAID3);
289 }
290 
291 static int
292 g_raid3_event_dispatch(struct g_raid3_event *ep, void *arg, int state,
293     int flags)
294 {
295 	struct g_raid3_softc *sc;
296 	struct g_raid3_disk *disk;
297 	int error;
298 
299 	G_RAID3_DEBUG(4, "%s: Sending event %p.", __func__, ep);
300 	if ((flags & G_RAID3_EVENT_DEVICE) != 0) {
301 		disk = NULL;
302 		sc = arg;
303 	} else {
304 		disk = arg;
305 		sc = disk->d_softc;
306 	}
307 	ep->e_disk = disk;
308 	ep->e_state = state;
309 	ep->e_flags = flags;
310 	ep->e_error = 0;
311 	mtx_lock(&sc->sc_events_mtx);
312 	TAILQ_INSERT_TAIL(&sc->sc_events, ep, e_next);
313 	mtx_unlock(&sc->sc_events_mtx);
314 	G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, sc);
315 	mtx_lock(&sc->sc_queue_mtx);
316 	wakeup(sc);
317 	wakeup(&sc->sc_queue);
318 	mtx_unlock(&sc->sc_queue_mtx);
319 	if ((flags & G_RAID3_EVENT_DONTWAIT) != 0)
320 		return (0);
321 	sx_assert(&sc->sc_lock, SX_XLOCKED);
322 	G_RAID3_DEBUG(4, "%s: Sleeping %p.", __func__, ep);
323 	sx_xunlock(&sc->sc_lock);
324 	while ((ep->e_flags & G_RAID3_EVENT_DONE) == 0) {
325 		mtx_lock(&sc->sc_events_mtx);
326 		MSLEEP(ep, &sc->sc_events_mtx, PRIBIO | PDROP, "r3:event",
327 		    hz * 5);
328 	}
329 	error = ep->e_error;
330 	g_raid3_event_free(ep);
331 	sx_xlock(&sc->sc_lock);
332 	return (error);
333 }
334 
335 int
336 g_raid3_event_send(void *arg, int state, int flags)
337 {
338 	struct g_raid3_event *ep;
339 
340 	ep = malloc(sizeof(*ep), M_RAID3, M_WAITOK);
341 	return (g_raid3_event_dispatch(ep, arg, state, flags));
342 }
343 
344 static struct g_raid3_event *
345 g_raid3_event_get(struct g_raid3_softc *sc)
346 {
347 	struct g_raid3_event *ep;
348 
349 	mtx_lock(&sc->sc_events_mtx);
350 	ep = TAILQ_FIRST(&sc->sc_events);
351 	mtx_unlock(&sc->sc_events_mtx);
352 	return (ep);
353 }
354 
355 static void
356 g_raid3_event_remove(struct g_raid3_softc *sc, struct g_raid3_event *ep)
357 {
358 
359 	mtx_lock(&sc->sc_events_mtx);
360 	TAILQ_REMOVE(&sc->sc_events, ep, e_next);
361 	mtx_unlock(&sc->sc_events_mtx);
362 }
363 
364 static void
365 g_raid3_event_cancel(struct g_raid3_disk *disk)
366 {
367 	struct g_raid3_softc *sc;
368 	struct g_raid3_event *ep, *tmpep;
369 
370 	sc = disk->d_softc;
371 	sx_assert(&sc->sc_lock, SX_XLOCKED);
372 
373 	mtx_lock(&sc->sc_events_mtx);
374 	TAILQ_FOREACH_SAFE(ep, &sc->sc_events, e_next, tmpep) {
375 		if ((ep->e_flags & G_RAID3_EVENT_DEVICE) != 0)
376 			continue;
377 		if (ep->e_disk != disk)
378 			continue;
379 		TAILQ_REMOVE(&sc->sc_events, ep, e_next);
380 		if ((ep->e_flags & G_RAID3_EVENT_DONTWAIT) != 0)
381 			g_raid3_event_free(ep);
382 		else {
383 			ep->e_error = ECANCELED;
384 			wakeup(ep);
385 		}
386 	}
387 	mtx_unlock(&sc->sc_events_mtx);
388 }
389 
390 /*
391  * Return the number of disks in the given state.
392  * If state is equal to -1, count all connected disks.
393  */
394 u_int
395 g_raid3_ndisks(struct g_raid3_softc *sc, int state)
396 {
397 	struct g_raid3_disk *disk;
398 	u_int n, ndisks;
399 
400 	sx_assert(&sc->sc_lock, SX_LOCKED);
401 
402 	for (n = ndisks = 0; n < sc->sc_ndisks; n++) {
403 		disk = &sc->sc_disks[n];
404 		if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
405 			continue;
406 		if (state == -1 || disk->d_state == state)
407 			ndisks++;
408 	}
409 	return (ndisks);
410 }
411 
412 static u_int
413 g_raid3_nrequests(struct g_raid3_softc *sc, struct g_consumer *cp)
414 {
415 	struct bio *bp;
416 	u_int nreqs = 0;
417 
418 	mtx_lock(&sc->sc_queue_mtx);
419 	TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
420 		if (bp->bio_from == cp)
421 			nreqs++;
422 	}
423 	mtx_unlock(&sc->sc_queue_mtx);
424 	return (nreqs);
425 }
426 
427 static int
428 g_raid3_is_busy(struct g_raid3_softc *sc, struct g_consumer *cp)
429 {
430 
431 	if (cp->index > 0) {
432 		G_RAID3_DEBUG(2,
433 		    "I/O requests for %s exist, can't destroy it now.",
434 		    cp->provider->name);
435 		return (1);
436 	}
437 	if (g_raid3_nrequests(sc, cp) > 0) {
438 		G_RAID3_DEBUG(2,
439 		    "I/O requests for %s in queue, can't destroy it now.",
440 		    cp->provider->name);
441 		return (1);
442 	}
443 	return (0);
444 }
445 
446 static void
447 g_raid3_destroy_consumer(void *arg, int flags __unused)
448 {
449 	struct g_consumer *cp;
450 
451 	g_topology_assert();
452 
453 	cp = arg;
454 	G_RAID3_DEBUG(1, "Consumer %s destroyed.", cp->provider->name);
455 	g_detach(cp);
456 	g_destroy_consumer(cp);
457 }
458 
459 static void
460 g_raid3_kill_consumer(struct g_raid3_softc *sc, struct g_consumer *cp)
461 {
462 	struct g_provider *pp;
463 	int retaste_wait;
464 
465 	g_topology_assert();
466 
467 	cp->private = NULL;
468 	if (g_raid3_is_busy(sc, cp))
469 		return;
470 	G_RAID3_DEBUG(2, "Consumer %s destroyed.", cp->provider->name);
471 	pp = cp->provider;
472 	retaste_wait = 0;
473 	if (cp->acw == 1) {
474 		if ((pp->geom->flags & G_GEOM_WITHER) == 0)
475 			retaste_wait = 1;
476 	}
477 	G_RAID3_DEBUG(2, "Access %s r%dw%de%d = %d", pp->name, -cp->acr,
478 	    -cp->acw, -cp->ace, 0);
479 	if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
480 		g_access(cp, -cp->acr, -cp->acw, -cp->ace);
481 	if (retaste_wait) {
482 		/*
483 		 * After retaste event was send (inside g_access()), we can send
484 		 * event to detach and destroy consumer.
485 		 * A class, which has consumer to the given provider connected
486 		 * will not receive retaste event for the provider.
487 		 * This is the way how I ignore retaste events when I close
488 		 * consumers opened for write: I detach and destroy consumer
489 		 * after retaste event is sent.
490 		 */
491 		g_post_event(g_raid3_destroy_consumer, cp, M_WAITOK, NULL);
492 		return;
493 	}
494 	G_RAID3_DEBUG(1, "Consumer %s destroyed.", pp->name);
495 	g_detach(cp);
496 	g_destroy_consumer(cp);
497 }
498 
499 static int
500 g_raid3_connect_disk(struct g_raid3_disk *disk, struct g_provider *pp)
501 {
502 	struct g_consumer *cp;
503 	int error;
504 
505 	g_topology_assert_not();
506 	KASSERT(disk->d_consumer == NULL,
507 	    ("Disk already connected (device %s).", disk->d_softc->sc_name));
508 
509 	g_topology_lock();
510 	cp = g_new_consumer(disk->d_softc->sc_geom);
511 	error = g_attach(cp, pp);
512 	if (error != 0) {
513 		g_destroy_consumer(cp);
514 		g_topology_unlock();
515 		return (error);
516 	}
517 	error = g_access(cp, 1, 1, 1);
518 		g_topology_unlock();
519 	if (error != 0) {
520 		g_detach(cp);
521 		g_destroy_consumer(cp);
522 		G_RAID3_DEBUG(0, "Cannot open consumer %s (error=%d).",
523 		    pp->name, error);
524 		return (error);
525 	}
526 	disk->d_consumer = cp;
527 	disk->d_consumer->private = disk;
528 	disk->d_consumer->index = 0;
529 	G_RAID3_DEBUG(2, "Disk %s connected.", g_raid3_get_diskname(disk));
530 	return (0);
531 }
532 
533 static void
534 g_raid3_disconnect_consumer(struct g_raid3_softc *sc, struct g_consumer *cp)
535 {
536 
537 	g_topology_assert();
538 
539 	if (cp == NULL)
540 		return;
541 	if (cp->provider != NULL)
542 		g_raid3_kill_consumer(sc, cp);
543 	else
544 		g_destroy_consumer(cp);
545 }
546 
547 /*
548  * Initialize disk. This means allocate memory, create consumer, attach it
549  * to the provider and open access (r1w1e1) to it.
550  */
551 static struct g_raid3_disk *
552 g_raid3_init_disk(struct g_raid3_softc *sc, struct g_provider *pp,
553     struct g_raid3_metadata *md, int *errorp)
554 {
555 	struct g_raid3_disk *disk;
556 	int error;
557 
558 	disk = &sc->sc_disks[md->md_no];
559 	error = g_raid3_connect_disk(disk, pp);
560 	if (error != 0) {
561 		if (errorp != NULL)
562 			*errorp = error;
563 		return (NULL);
564 	}
565 	disk->d_state = G_RAID3_DISK_STATE_NONE;
566 	disk->d_flags = md->md_dflags;
567 	if (md->md_provider[0] != '\0')
568 		disk->d_flags |= G_RAID3_DISK_FLAG_HARDCODED;
569 	disk->d_sync.ds_consumer = NULL;
570 	disk->d_sync.ds_offset = md->md_sync_offset;
571 	disk->d_sync.ds_offset_done = md->md_sync_offset;
572 	disk->d_genid = md->md_genid;
573 	disk->d_sync.ds_syncid = md->md_syncid;
574 	if (errorp != NULL)
575 		*errorp = 0;
576 	return (disk);
577 }
578 
579 static void
580 g_raid3_destroy_disk(struct g_raid3_disk *disk)
581 {
582 	struct g_raid3_softc *sc;
583 
584 	g_topology_assert_not();
585 	sc = disk->d_softc;
586 	sx_assert(&sc->sc_lock, SX_XLOCKED);
587 
588 	if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
589 		return;
590 	g_raid3_event_cancel(disk);
591 	switch (disk->d_state) {
592 	case G_RAID3_DISK_STATE_SYNCHRONIZING:
593 		if (sc->sc_syncdisk != NULL)
594 			g_raid3_sync_stop(sc, 1);
595 		/* FALLTHROUGH */
596 	case G_RAID3_DISK_STATE_NEW:
597 	case G_RAID3_DISK_STATE_STALE:
598 	case G_RAID3_DISK_STATE_ACTIVE:
599 		g_topology_lock();
600 		g_raid3_disconnect_consumer(sc, disk->d_consumer);
601 		g_topology_unlock();
602 		disk->d_consumer = NULL;
603 		break;
604 	default:
605 		KASSERT(0 == 1, ("Wrong disk state (%s, %s).",
606 		    g_raid3_get_diskname(disk),
607 		    g_raid3_disk_state2str(disk->d_state)));
608 	}
609 	disk->d_state = G_RAID3_DISK_STATE_NODISK;
610 }
611 
612 static void
613 g_raid3_free_device(struct g_raid3_softc *sc)
614 {
615 	KASSERT(sc->sc_refcnt == 0,
616 	    ("%s: non-zero refcount %u", __func__, sc->sc_refcnt));
617 
618 	if (!g_raid3_use_malloc) {
619 		uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_64K].sz_zone);
620 		uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_16K].sz_zone);
621 		uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_4K].sz_zone);
622 	}
623 	mtx_destroy(&sc->sc_queue_mtx);
624 	mtx_destroy(&sc->sc_events_mtx);
625 	sx_xunlock(&sc->sc_lock);
626 	sx_destroy(&sc->sc_lock);
627 	free(sc->sc_disks, M_RAID3);
628 	free(sc, M_RAID3);
629 }
630 
631 static void
632 g_raid3_providergone(struct g_provider *pp)
633 {
634 	struct g_raid3_softc *sc = pp->private;
635 
636 	if (--sc->sc_refcnt == 0)
637 		g_raid3_free_device(sc);
638 }
639 
640 static void
641 g_raid3_destroy_device(struct g_raid3_softc *sc)
642 {
643 	struct g_raid3_event *ep;
644 	struct g_raid3_disk *disk;
645 	struct g_geom *gp;
646 	struct g_consumer *cp;
647 	u_int n;
648 
649 	g_topology_assert_not();
650 	sx_assert(&sc->sc_lock, SX_XLOCKED);
651 
652 	gp = sc->sc_geom;
653 	if (sc->sc_provider != NULL)
654 		g_raid3_destroy_provider(sc);
655 	for (n = 0; n < sc->sc_ndisks; n++) {
656 		disk = &sc->sc_disks[n];
657 		if (disk->d_state != G_RAID3_DISK_STATE_NODISK) {
658 			disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
659 			g_raid3_update_metadata(disk);
660 			g_raid3_destroy_disk(disk);
661 		}
662 	}
663 	while ((ep = g_raid3_event_get(sc)) != NULL) {
664 		g_raid3_event_remove(sc, ep);
665 		if ((ep->e_flags & G_RAID3_EVENT_DONTWAIT) != 0)
666 			g_raid3_event_free(ep);
667 		else {
668 			ep->e_error = ECANCELED;
669 			ep->e_flags |= G_RAID3_EVENT_DONE;
670 			G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, ep);
671 			mtx_lock(&sc->sc_events_mtx);
672 			wakeup(ep);
673 			mtx_unlock(&sc->sc_events_mtx);
674 		}
675 	}
676 	g_raid3_timeout_drain(sc);
677 	cp = LIST_FIRST(&sc->sc_sync.ds_geom->consumer);
678 	g_topology_lock();
679 	if (cp != NULL)
680 		g_raid3_disconnect_consumer(sc, cp);
681 	g_wither_geom(sc->sc_sync.ds_geom, ENXIO);
682 	G_RAID3_DEBUG(0, "Device %s destroyed.", gp->name);
683 	g_wither_geom(gp, ENXIO);
684 	if (--sc->sc_refcnt == 0)
685 		g_raid3_free_device(sc);
686 	g_topology_unlock();
687 }
688 
689 static void
690 g_raid3_orphan(struct g_consumer *cp)
691 {
692 	struct g_raid3_disk *disk;
693 
694 	g_topology_assert();
695 
696 	disk = cp->private;
697 	if (disk == NULL)
698 		return;
699 	disk->d_softc->sc_bump_id = G_RAID3_BUMP_SYNCID;
700 	g_raid3_event_send(disk, G_RAID3_DISK_STATE_DISCONNECTED,
701 	    G_RAID3_EVENT_DONTWAIT);
702 }
703 
704 static int
705 g_raid3_write_metadata(struct g_raid3_disk *disk, struct g_raid3_metadata *md)
706 {
707 	struct g_raid3_softc *sc;
708 	struct g_consumer *cp;
709 	off_t offset, length;
710 	u_char *sector;
711 	int error = 0;
712 
713 	g_topology_assert_not();
714 	sc = disk->d_softc;
715 	sx_assert(&sc->sc_lock, SX_LOCKED);
716 
717 	cp = disk->d_consumer;
718 	KASSERT(cp != NULL, ("NULL consumer (%s).", sc->sc_name));
719 	KASSERT(cp->provider != NULL, ("NULL provider (%s).", sc->sc_name));
720 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
721 	    ("Consumer %s closed? (r%dw%de%d).", cp->provider->name, cp->acr,
722 	    cp->acw, cp->ace));
723 	length = cp->provider->sectorsize;
724 	offset = cp->provider->mediasize - length;
725 	sector = malloc((size_t)length, M_RAID3, M_WAITOK | M_ZERO);
726 	if (md != NULL)
727 		raid3_metadata_encode(md, sector);
728 	error = g_write_data(cp, offset, sector, length);
729 	free(sector, M_RAID3);
730 	if (error != 0) {
731 		if ((disk->d_flags & G_RAID3_DISK_FLAG_BROKEN) == 0) {
732 			G_RAID3_DEBUG(0, "Cannot write metadata on %s "
733 			    "(device=%s, error=%d).",
734 			    g_raid3_get_diskname(disk), sc->sc_name, error);
735 			disk->d_flags |= G_RAID3_DISK_FLAG_BROKEN;
736 		} else {
737 			G_RAID3_DEBUG(1, "Cannot write metadata on %s "
738 			    "(device=%s, error=%d).",
739 			    g_raid3_get_diskname(disk), sc->sc_name, error);
740 		}
741 		if (g_raid3_disconnect_on_failure &&
742 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
743 			sc->sc_bump_id |= G_RAID3_BUMP_GENID;
744 			g_raid3_event_send(disk,
745 			    G_RAID3_DISK_STATE_DISCONNECTED,
746 			    G_RAID3_EVENT_DONTWAIT);
747 		}
748 	}
749 	return (error);
750 }
751 
752 int
753 g_raid3_clear_metadata(struct g_raid3_disk *disk)
754 {
755 	int error;
756 
757 	g_topology_assert_not();
758 	sx_assert(&disk->d_softc->sc_lock, SX_LOCKED);
759 
760 	error = g_raid3_write_metadata(disk, NULL);
761 	if (error == 0) {
762 		G_RAID3_DEBUG(2, "Metadata on %s cleared.",
763 		    g_raid3_get_diskname(disk));
764 	} else {
765 		G_RAID3_DEBUG(0,
766 		    "Cannot clear metadata on disk %s (error=%d).",
767 		    g_raid3_get_diskname(disk), error);
768 	}
769 	return (error);
770 }
771 
772 void
773 g_raid3_fill_metadata(struct g_raid3_disk *disk, struct g_raid3_metadata *md)
774 {
775 	struct g_raid3_softc *sc;
776 	struct g_provider *pp;
777 
778 	bzero(md, sizeof(*md));
779 	sc = disk->d_softc;
780 	strlcpy(md->md_magic, G_RAID3_MAGIC, sizeof(md->md_magic));
781 	md->md_version = G_RAID3_VERSION;
782 	strlcpy(md->md_name, sc->sc_name, sizeof(md->md_name));
783 	md->md_id = sc->sc_id;
784 	md->md_all = sc->sc_ndisks;
785 	md->md_genid = sc->sc_genid;
786 	md->md_mediasize = sc->sc_mediasize;
787 	md->md_sectorsize = sc->sc_sectorsize;
788 	md->md_mflags = (sc->sc_flags & G_RAID3_DEVICE_FLAG_MASK);
789 	md->md_no = disk->d_no;
790 	md->md_syncid = disk->d_sync.ds_syncid;
791 	md->md_dflags = (disk->d_flags & G_RAID3_DISK_FLAG_MASK);
792 	if (disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
793 		md->md_sync_offset =
794 		    disk->d_sync.ds_offset_done / (sc->sc_ndisks - 1);
795 	}
796 	if (disk->d_consumer != NULL && disk->d_consumer->provider != NULL)
797 		pp = disk->d_consumer->provider;
798 	else
799 		pp = NULL;
800 	if ((disk->d_flags & G_RAID3_DISK_FLAG_HARDCODED) != 0 && pp != NULL)
801 		strlcpy(md->md_provider, pp->name, sizeof(md->md_provider));
802 	if (pp != NULL)
803 		md->md_provsize = pp->mediasize;
804 }
805 
806 void
807 g_raid3_update_metadata(struct g_raid3_disk *disk)
808 {
809 	struct g_raid3_softc *sc __diagused;
810 	struct g_raid3_metadata md;
811 	int error;
812 
813 	g_topology_assert_not();
814 	sc = disk->d_softc;
815 	sx_assert(&sc->sc_lock, SX_LOCKED);
816 
817 	g_raid3_fill_metadata(disk, &md);
818 	error = g_raid3_write_metadata(disk, &md);
819 	if (error == 0) {
820 		G_RAID3_DEBUG(2, "Metadata on %s updated.",
821 		    g_raid3_get_diskname(disk));
822 	} else {
823 		G_RAID3_DEBUG(0,
824 		    "Cannot update metadata on disk %s (error=%d).",
825 		    g_raid3_get_diskname(disk), error);
826 	}
827 }
828 
829 static void
830 g_raid3_bump_syncid(struct g_raid3_softc *sc)
831 {
832 	struct g_raid3_disk *disk;
833 	u_int n;
834 
835 	g_topology_assert_not();
836 	sx_assert(&sc->sc_lock, SX_XLOCKED);
837 	KASSERT(g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) > 0,
838 	    ("%s called with no active disks (device=%s).", __func__,
839 	    sc->sc_name));
840 
841 	sc->sc_syncid++;
842 	G_RAID3_DEBUG(1, "Device %s: syncid bumped to %u.", sc->sc_name,
843 	    sc->sc_syncid);
844 	for (n = 0; n < sc->sc_ndisks; n++) {
845 		disk = &sc->sc_disks[n];
846 		if (disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
847 		    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
848 			disk->d_sync.ds_syncid = sc->sc_syncid;
849 			g_raid3_update_metadata(disk);
850 		}
851 	}
852 }
853 
854 static void
855 g_raid3_bump_genid(struct g_raid3_softc *sc)
856 {
857 	struct g_raid3_disk *disk;
858 	u_int n;
859 
860 	g_topology_assert_not();
861 	sx_assert(&sc->sc_lock, SX_XLOCKED);
862 	KASSERT(g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) > 0,
863 	    ("%s called with no active disks (device=%s).", __func__,
864 	    sc->sc_name));
865 
866 	sc->sc_genid++;
867 	G_RAID3_DEBUG(1, "Device %s: genid bumped to %u.", sc->sc_name,
868 	    sc->sc_genid);
869 	for (n = 0; n < sc->sc_ndisks; n++) {
870 		disk = &sc->sc_disks[n];
871 		if (disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
872 		    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
873 			disk->d_genid = sc->sc_genid;
874 			g_raid3_update_metadata(disk);
875 		}
876 	}
877 }
878 
879 static int
880 g_raid3_idle(struct g_raid3_softc *sc, int acw)
881 {
882 	struct g_raid3_disk *disk;
883 	u_int i;
884 	int timeout;
885 
886 	g_topology_assert_not();
887 	sx_assert(&sc->sc_lock, SX_XLOCKED);
888 
889 	if (sc->sc_provider == NULL)
890 		return (0);
891 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) != 0)
892 		return (0);
893 	if (sc->sc_idle)
894 		return (0);
895 	if (sc->sc_writes > 0)
896 		return (0);
897 	if (acw > 0 || (acw == -1 && sc->sc_provider->acw > 0)) {
898 		timeout = g_raid3_idletime - (time_uptime - sc->sc_last_write);
899 		if (!g_raid3_shutdown && timeout > 0)
900 			return (timeout);
901 	}
902 	sc->sc_idle = 1;
903 	for (i = 0; i < sc->sc_ndisks; i++) {
904 		disk = &sc->sc_disks[i];
905 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE)
906 			continue;
907 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as clean.",
908 		    g_raid3_get_diskname(disk), sc->sc_name);
909 		disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
910 		g_raid3_update_metadata(disk);
911 	}
912 	return (0);
913 }
914 
915 static void
916 g_raid3_unidle(struct g_raid3_softc *sc)
917 {
918 	struct g_raid3_disk *disk;
919 	u_int i;
920 
921 	g_topology_assert_not();
922 	sx_assert(&sc->sc_lock, SX_XLOCKED);
923 
924 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) != 0)
925 		return;
926 	sc->sc_idle = 0;
927 	sc->sc_last_write = time_uptime;
928 	for (i = 0; i < sc->sc_ndisks; i++) {
929 		disk = &sc->sc_disks[i];
930 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE)
931 			continue;
932 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as dirty.",
933 		    g_raid3_get_diskname(disk), sc->sc_name);
934 		disk->d_flags |= G_RAID3_DISK_FLAG_DIRTY;
935 		g_raid3_update_metadata(disk);
936 	}
937 }
938 
939 /*
940  * Treat bio_driver1 field in parent bio as list head and field bio_caller1
941  * in child bio as pointer to the next element on the list.
942  */
943 #define	G_RAID3_HEAD_BIO(pbp)	(pbp)->bio_driver1
944 
945 #define	G_RAID3_NEXT_BIO(cbp)	(cbp)->bio_caller1
946 
947 #define	G_RAID3_FOREACH_BIO(pbp, bp)					\
948 	for ((bp) = G_RAID3_HEAD_BIO(pbp); (bp) != NULL;		\
949 	    (bp) = G_RAID3_NEXT_BIO(bp))
950 
951 #define	G_RAID3_FOREACH_SAFE_BIO(pbp, bp, tmpbp)			\
952 	for ((bp) = G_RAID3_HEAD_BIO(pbp);				\
953 	    (bp) != NULL && ((tmpbp) = G_RAID3_NEXT_BIO(bp), 1);	\
954 	    (bp) = (tmpbp))
955 
956 static void
957 g_raid3_init_bio(struct bio *pbp)
958 {
959 
960 	G_RAID3_HEAD_BIO(pbp) = NULL;
961 }
962 
963 static void
964 g_raid3_remove_bio(struct bio *cbp)
965 {
966 	struct bio *pbp, *bp;
967 
968 	pbp = cbp->bio_parent;
969 	if (G_RAID3_HEAD_BIO(pbp) == cbp)
970 		G_RAID3_HEAD_BIO(pbp) = G_RAID3_NEXT_BIO(cbp);
971 	else {
972 		G_RAID3_FOREACH_BIO(pbp, bp) {
973 			if (G_RAID3_NEXT_BIO(bp) == cbp) {
974 				G_RAID3_NEXT_BIO(bp) = G_RAID3_NEXT_BIO(cbp);
975 				break;
976 			}
977 		}
978 	}
979 	G_RAID3_NEXT_BIO(cbp) = NULL;
980 }
981 
982 static void
983 g_raid3_replace_bio(struct bio *sbp, struct bio *dbp)
984 {
985 	struct bio *pbp, *bp;
986 
987 	g_raid3_remove_bio(sbp);
988 	pbp = dbp->bio_parent;
989 	G_RAID3_NEXT_BIO(sbp) = G_RAID3_NEXT_BIO(dbp);
990 	if (G_RAID3_HEAD_BIO(pbp) == dbp)
991 		G_RAID3_HEAD_BIO(pbp) = sbp;
992 	else {
993 		G_RAID3_FOREACH_BIO(pbp, bp) {
994 			if (G_RAID3_NEXT_BIO(bp) == dbp) {
995 				G_RAID3_NEXT_BIO(bp) = sbp;
996 				break;
997 			}
998 		}
999 	}
1000 	G_RAID3_NEXT_BIO(dbp) = NULL;
1001 }
1002 
1003 static void
1004 g_raid3_destroy_bio(struct g_raid3_softc *sc, struct bio *cbp)
1005 {
1006 	struct bio *bp, *pbp;
1007 	size_t size;
1008 
1009 	pbp = cbp->bio_parent;
1010 	pbp->bio_children--;
1011 	KASSERT(cbp->bio_data != NULL, ("NULL bio_data"));
1012 	size = pbp->bio_length / (sc->sc_ndisks - 1);
1013 	g_raid3_free(sc, cbp->bio_data, size);
1014 	if (G_RAID3_HEAD_BIO(pbp) == cbp) {
1015 		G_RAID3_HEAD_BIO(pbp) = G_RAID3_NEXT_BIO(cbp);
1016 		G_RAID3_NEXT_BIO(cbp) = NULL;
1017 		g_destroy_bio(cbp);
1018 	} else {
1019 		G_RAID3_FOREACH_BIO(pbp, bp) {
1020 			if (G_RAID3_NEXT_BIO(bp) == cbp)
1021 				break;
1022 		}
1023 		if (bp != NULL) {
1024 			KASSERT(G_RAID3_NEXT_BIO(bp) != NULL,
1025 			    ("NULL bp->bio_driver1"));
1026 			G_RAID3_NEXT_BIO(bp) = G_RAID3_NEXT_BIO(cbp);
1027 			G_RAID3_NEXT_BIO(cbp) = NULL;
1028 		}
1029 		g_destroy_bio(cbp);
1030 	}
1031 }
1032 
1033 static struct bio *
1034 g_raid3_clone_bio(struct g_raid3_softc *sc, struct bio *pbp)
1035 {
1036 	struct bio *bp, *cbp;
1037 	size_t size;
1038 	int memflag;
1039 
1040 	cbp = g_clone_bio(pbp);
1041 	if (cbp == NULL)
1042 		return (NULL);
1043 	size = pbp->bio_length / (sc->sc_ndisks - 1);
1044 	if ((pbp->bio_cflags & G_RAID3_BIO_CFLAG_REGULAR) != 0)
1045 		memflag = M_WAITOK;
1046 	else
1047 		memflag = M_NOWAIT;
1048 	cbp->bio_data = g_raid3_alloc(sc, size, memflag);
1049 	if (cbp->bio_data == NULL) {
1050 		pbp->bio_children--;
1051 		g_destroy_bio(cbp);
1052 		return (NULL);
1053 	}
1054 	G_RAID3_NEXT_BIO(cbp) = NULL;
1055 	if (G_RAID3_HEAD_BIO(pbp) == NULL)
1056 		G_RAID3_HEAD_BIO(pbp) = cbp;
1057 	else {
1058 		G_RAID3_FOREACH_BIO(pbp, bp) {
1059 			if (G_RAID3_NEXT_BIO(bp) == NULL) {
1060 				G_RAID3_NEXT_BIO(bp) = cbp;
1061 				break;
1062 			}
1063 		}
1064 	}
1065 	return (cbp);
1066 }
1067 
1068 static void
1069 g_raid3_scatter(struct bio *pbp)
1070 {
1071 	struct g_raid3_softc *sc;
1072 	struct g_raid3_disk *disk;
1073 	struct bio *bp, *cbp, *tmpbp;
1074 	off_t atom, cadd, padd, left;
1075 	int first;
1076 
1077 	sc = pbp->bio_to->private;
1078 	bp = NULL;
1079 	if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_NOPARITY) == 0) {
1080 		/*
1081 		 * Find bio for which we should calculate data.
1082 		 */
1083 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1084 			if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_PARITY) != 0) {
1085 				bp = cbp;
1086 				break;
1087 			}
1088 		}
1089 		KASSERT(bp != NULL, ("NULL parity bio."));
1090 	}
1091 	atom = sc->sc_sectorsize / (sc->sc_ndisks - 1);
1092 	cadd = padd = 0;
1093 	for (left = pbp->bio_length; left > 0; left -= sc->sc_sectorsize) {
1094 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1095 			if (cbp == bp)
1096 				continue;
1097 			bcopy(pbp->bio_data + padd, cbp->bio_data + cadd, atom);
1098 			padd += atom;
1099 		}
1100 		cadd += atom;
1101 	}
1102 	if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_NOPARITY) == 0) {
1103 		/*
1104 		 * Calculate parity.
1105 		 */
1106 		first = 1;
1107 		G_RAID3_FOREACH_SAFE_BIO(pbp, cbp, tmpbp) {
1108 			if (cbp == bp)
1109 				continue;
1110 			if (first) {
1111 				bcopy(cbp->bio_data, bp->bio_data,
1112 				    bp->bio_length);
1113 				first = 0;
1114 			} else {
1115 				g_raid3_xor(cbp->bio_data, bp->bio_data,
1116 				    bp->bio_length);
1117 			}
1118 			if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_NODISK) != 0)
1119 				g_raid3_destroy_bio(sc, cbp);
1120 		}
1121 	}
1122 	G_RAID3_FOREACH_SAFE_BIO(pbp, cbp, tmpbp) {
1123 		struct g_consumer *cp;
1124 
1125 		disk = cbp->bio_caller2;
1126 		cp = disk->d_consumer;
1127 		cbp->bio_to = cp->provider;
1128 		G_RAID3_LOGREQ(3, cbp, "Sending request.");
1129 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1130 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1131 		    cp->acr, cp->acw, cp->ace));
1132 		cp->index++;
1133 		sc->sc_writes++;
1134 		g_io_request(cbp, cp);
1135 	}
1136 }
1137 
1138 static void
1139 g_raid3_gather(struct bio *pbp)
1140 {
1141 	struct g_raid3_softc *sc;
1142 	struct g_raid3_disk *disk;
1143 	struct bio *xbp, *fbp, *cbp;
1144 	off_t atom, cadd, padd, left;
1145 
1146 	sc = pbp->bio_to->private;
1147 	/*
1148 	 * Find bio for which we have to calculate data.
1149 	 * While going through this path, check if all requests
1150 	 * succeeded, if not, deny whole request.
1151 	 * If we're in COMPLETE mode, we allow one request to fail,
1152 	 * so if we find one, we're sending it to the parity consumer.
1153 	 * If there are more failed requests, we deny whole request.
1154 	 */
1155 	xbp = fbp = NULL;
1156 	G_RAID3_FOREACH_BIO(pbp, cbp) {
1157 		if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_PARITY) != 0) {
1158 			KASSERT(xbp == NULL, ("More than one parity bio."));
1159 			xbp = cbp;
1160 		}
1161 		if (cbp->bio_error == 0)
1162 			continue;
1163 		/*
1164 		 * Found failed request.
1165 		 */
1166 		if (fbp == NULL) {
1167 			if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_DEGRADED) != 0) {
1168 				/*
1169 				 * We are already in degraded mode, so we can't
1170 				 * accept any failures.
1171 				 */
1172 				if (pbp->bio_error == 0)
1173 					pbp->bio_error = cbp->bio_error;
1174 			} else {
1175 				fbp = cbp;
1176 			}
1177 		} else {
1178 			/*
1179 			 * Next failed request, that's too many.
1180 			 */
1181 			if (pbp->bio_error == 0)
1182 				pbp->bio_error = fbp->bio_error;
1183 		}
1184 		disk = cbp->bio_caller2;
1185 		if (disk == NULL)
1186 			continue;
1187 		if ((disk->d_flags & G_RAID3_DISK_FLAG_BROKEN) == 0) {
1188 			disk->d_flags |= G_RAID3_DISK_FLAG_BROKEN;
1189 			G_RAID3_LOGREQ(0, cbp, "Request failed (error=%d).",
1190 			    cbp->bio_error);
1191 		} else {
1192 			G_RAID3_LOGREQ(1, cbp, "Request failed (error=%d).",
1193 			    cbp->bio_error);
1194 		}
1195 		if (g_raid3_disconnect_on_failure &&
1196 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1197 			sc->sc_bump_id |= G_RAID3_BUMP_GENID;
1198 			g_raid3_event_send(disk,
1199 			    G_RAID3_DISK_STATE_DISCONNECTED,
1200 			    G_RAID3_EVENT_DONTWAIT);
1201 		}
1202 	}
1203 	if (pbp->bio_error != 0)
1204 		goto finish;
1205 	if (fbp != NULL && (pbp->bio_pflags & G_RAID3_BIO_PFLAG_VERIFY) != 0) {
1206 		pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_VERIFY;
1207 		if (xbp != fbp)
1208 			g_raid3_replace_bio(xbp, fbp);
1209 		g_raid3_destroy_bio(sc, fbp);
1210 	} else if (fbp != NULL) {
1211 		struct g_consumer *cp;
1212 
1213 		/*
1214 		 * One request failed, so send the same request to
1215 		 * the parity consumer.
1216 		 */
1217 		disk = pbp->bio_driver2;
1218 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE) {
1219 			pbp->bio_error = fbp->bio_error;
1220 			goto finish;
1221 		}
1222 		pbp->bio_pflags |= G_RAID3_BIO_PFLAG_DEGRADED;
1223 		pbp->bio_inbed--;
1224 		fbp->bio_flags &= ~(BIO_DONE | BIO_ERROR);
1225 		if (disk->d_no == sc->sc_ndisks - 1)
1226 			fbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1227 		fbp->bio_error = 0;
1228 		fbp->bio_completed = 0;
1229 		fbp->bio_children = 0;
1230 		fbp->bio_inbed = 0;
1231 		cp = disk->d_consumer;
1232 		fbp->bio_caller2 = disk;
1233 		fbp->bio_to = cp->provider;
1234 		G_RAID3_LOGREQ(3, fbp, "Sending request (recover).");
1235 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1236 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1237 		    cp->acr, cp->acw, cp->ace));
1238 		cp->index++;
1239 		g_io_request(fbp, cp);
1240 		return;
1241 	}
1242 	if (xbp != NULL) {
1243 		/*
1244 		 * Calculate parity.
1245 		 */
1246 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1247 			if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_PARITY) != 0)
1248 				continue;
1249 			g_raid3_xor(cbp->bio_data, xbp->bio_data,
1250 			    xbp->bio_length);
1251 		}
1252 		xbp->bio_cflags &= ~G_RAID3_BIO_CFLAG_PARITY;
1253 		if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_VERIFY) != 0) {
1254 			if (!g_raid3_is_zero(xbp)) {
1255 				g_raid3_parity_mismatch++;
1256 				pbp->bio_error = EIO;
1257 				goto finish;
1258 			}
1259 			g_raid3_destroy_bio(sc, xbp);
1260 		}
1261 	}
1262 	atom = sc->sc_sectorsize / (sc->sc_ndisks - 1);
1263 	cadd = padd = 0;
1264 	for (left = pbp->bio_length; left > 0; left -= sc->sc_sectorsize) {
1265 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1266 			bcopy(cbp->bio_data + cadd, pbp->bio_data + padd, atom);
1267 			pbp->bio_completed += atom;
1268 			padd += atom;
1269 		}
1270 		cadd += atom;
1271 	}
1272 finish:
1273 	if (pbp->bio_error == 0)
1274 		G_RAID3_LOGREQ(3, pbp, "Request finished.");
1275 	else {
1276 		if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_VERIFY) != 0)
1277 			G_RAID3_LOGREQ(1, pbp, "Verification error.");
1278 		else
1279 			G_RAID3_LOGREQ(0, pbp, "Request failed.");
1280 	}
1281 	pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_MASK;
1282 	while ((cbp = G_RAID3_HEAD_BIO(pbp)) != NULL)
1283 		g_raid3_destroy_bio(sc, cbp);
1284 	g_io_deliver(pbp, pbp->bio_error);
1285 }
1286 
1287 static void
1288 g_raid3_done(struct bio *bp)
1289 {
1290 	struct g_raid3_softc *sc;
1291 
1292 	sc = bp->bio_from->geom->softc;
1293 	bp->bio_cflags |= G_RAID3_BIO_CFLAG_REGULAR;
1294 	G_RAID3_LOGREQ(3, bp, "Regular request done (error=%d).", bp->bio_error);
1295 	mtx_lock(&sc->sc_queue_mtx);
1296 	bioq_insert_head(&sc->sc_queue, bp);
1297 	mtx_unlock(&sc->sc_queue_mtx);
1298 	wakeup(sc);
1299 	wakeup(&sc->sc_queue);
1300 }
1301 
1302 static void
1303 g_raid3_regular_request(struct bio *cbp)
1304 {
1305 	struct g_raid3_softc *sc;
1306 	struct g_raid3_disk *disk;
1307 	struct bio *pbp;
1308 
1309 	g_topology_assert_not();
1310 
1311 	pbp = cbp->bio_parent;
1312 	sc = pbp->bio_to->private;
1313 	cbp->bio_from->index--;
1314 	if (cbp->bio_cmd == BIO_WRITE)
1315 		sc->sc_writes--;
1316 	disk = cbp->bio_from->private;
1317 	if (disk == NULL) {
1318 		g_topology_lock();
1319 		g_raid3_kill_consumer(sc, cbp->bio_from);
1320 		g_topology_unlock();
1321 	}
1322 
1323 	G_RAID3_LOGREQ(3, cbp, "Request finished.");
1324 	pbp->bio_inbed++;
1325 	KASSERT(pbp->bio_inbed <= pbp->bio_children,
1326 	    ("bio_inbed (%u) is bigger than bio_children (%u).", pbp->bio_inbed,
1327 	    pbp->bio_children));
1328 	if (pbp->bio_inbed != pbp->bio_children)
1329 		return;
1330 	switch (pbp->bio_cmd) {
1331 	case BIO_READ:
1332 		g_raid3_gather(pbp);
1333 		break;
1334 	case BIO_WRITE:
1335 	case BIO_DELETE:
1336 	    {
1337 		int error = 0;
1338 
1339 		pbp->bio_completed = pbp->bio_length;
1340 		while ((cbp = G_RAID3_HEAD_BIO(pbp)) != NULL) {
1341 			if (cbp->bio_error == 0) {
1342 				g_raid3_destroy_bio(sc, cbp);
1343 				continue;
1344 			}
1345 
1346 			if (error == 0)
1347 				error = cbp->bio_error;
1348 			else if (pbp->bio_error == 0) {
1349 				/*
1350 				 * Next failed request, that's too many.
1351 				 */
1352 				pbp->bio_error = error;
1353 			}
1354 
1355 			disk = cbp->bio_caller2;
1356 			if (disk == NULL) {
1357 				g_raid3_destroy_bio(sc, cbp);
1358 				continue;
1359 			}
1360 
1361 			if ((disk->d_flags & G_RAID3_DISK_FLAG_BROKEN) == 0) {
1362 				disk->d_flags |= G_RAID3_DISK_FLAG_BROKEN;
1363 				G_RAID3_LOGREQ(0, cbp,
1364 				    "Request failed (error=%d).",
1365 				    cbp->bio_error);
1366 			} else {
1367 				G_RAID3_LOGREQ(1, cbp,
1368 				    "Request failed (error=%d).",
1369 				    cbp->bio_error);
1370 			}
1371 			if (g_raid3_disconnect_on_failure &&
1372 			    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1373 				sc->sc_bump_id |= G_RAID3_BUMP_GENID;
1374 				g_raid3_event_send(disk,
1375 				    G_RAID3_DISK_STATE_DISCONNECTED,
1376 				    G_RAID3_EVENT_DONTWAIT);
1377 			}
1378 			g_raid3_destroy_bio(sc, cbp);
1379 		}
1380 		if (pbp->bio_error == 0)
1381 			G_RAID3_LOGREQ(3, pbp, "Request finished.");
1382 		else
1383 			G_RAID3_LOGREQ(0, pbp, "Request failed.");
1384 		pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_DEGRADED;
1385 		pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_NOPARITY;
1386 		bioq_remove(&sc->sc_inflight, pbp);
1387 		/* Release delayed sync requests if possible. */
1388 		g_raid3_sync_release(sc);
1389 		g_io_deliver(pbp, pbp->bio_error);
1390 		break;
1391 	    }
1392 	}
1393 }
1394 
1395 static void
1396 g_raid3_sync_done(struct bio *bp)
1397 {
1398 	struct g_raid3_softc *sc;
1399 
1400 	G_RAID3_LOGREQ(3, bp, "Synchronization request delivered.");
1401 	sc = bp->bio_from->geom->softc;
1402 	bp->bio_cflags |= G_RAID3_BIO_CFLAG_SYNC;
1403 	mtx_lock(&sc->sc_queue_mtx);
1404 	bioq_insert_head(&sc->sc_queue, bp);
1405 	mtx_unlock(&sc->sc_queue_mtx);
1406 	wakeup(sc);
1407 	wakeup(&sc->sc_queue);
1408 }
1409 
1410 static void
1411 g_raid3_flush(struct g_raid3_softc *sc, struct bio *bp)
1412 {
1413 	struct bio_queue_head queue;
1414 	struct g_raid3_disk *disk;
1415 	struct g_consumer *cp __diagused;
1416 	struct bio *cbp;
1417 	u_int i;
1418 
1419 	bioq_init(&queue);
1420 	for (i = 0; i < sc->sc_ndisks; i++) {
1421 		disk = &sc->sc_disks[i];
1422 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE)
1423 			continue;
1424 		cbp = g_clone_bio(bp);
1425 		if (cbp == NULL) {
1426 			for (cbp = bioq_first(&queue); cbp != NULL;
1427 			    cbp = bioq_first(&queue)) {
1428 				bioq_remove(&queue, cbp);
1429 				g_destroy_bio(cbp);
1430 			}
1431 			if (bp->bio_error == 0)
1432 				bp->bio_error = ENOMEM;
1433 			g_io_deliver(bp, bp->bio_error);
1434 			return;
1435 		}
1436 		bioq_insert_tail(&queue, cbp);
1437 		cbp->bio_done = g_std_done;
1438 		cbp->bio_caller1 = disk;
1439 		cbp->bio_to = disk->d_consumer->provider;
1440 	}
1441 	for (cbp = bioq_first(&queue); cbp != NULL; cbp = bioq_first(&queue)) {
1442 		bioq_remove(&queue, cbp);
1443 		G_RAID3_LOGREQ(3, cbp, "Sending request.");
1444 		disk = cbp->bio_caller1;
1445 		cbp->bio_caller1 = NULL;
1446 		cp = disk->d_consumer;
1447 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1448 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1449 		    cp->acr, cp->acw, cp->ace));
1450 		g_io_request(cbp, disk->d_consumer);
1451 	}
1452 }
1453 
1454 static void
1455 g_raid3_start(struct bio *bp)
1456 {
1457 	struct g_raid3_softc *sc;
1458 
1459 	sc = bp->bio_to->private;
1460 	/*
1461 	 * If sc == NULL or there are no valid disks, provider's error
1462 	 * should be set and g_raid3_start() should not be called at all.
1463 	 */
1464 	KASSERT(sc != NULL && (sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
1465 	    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE),
1466 	    ("Provider's error should be set (error=%d)(device=%s).",
1467 	    bp->bio_to->error, bp->bio_to->name));
1468 	G_RAID3_LOGREQ(3, bp, "Request received.");
1469 
1470 	switch (bp->bio_cmd) {
1471 	case BIO_READ:
1472 	case BIO_WRITE:
1473 	case BIO_DELETE:
1474 		break;
1475 	case BIO_SPEEDUP:
1476 	case BIO_FLUSH:
1477 		g_raid3_flush(sc, bp);
1478 		return;
1479 	case BIO_GETATTR:
1480 	default:
1481 		g_io_deliver(bp, EOPNOTSUPP);
1482 		return;
1483 	}
1484 	mtx_lock(&sc->sc_queue_mtx);
1485 	bioq_insert_tail(&sc->sc_queue, bp);
1486 	mtx_unlock(&sc->sc_queue_mtx);
1487 	G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1488 	wakeup(sc);
1489 }
1490 
1491 /*
1492  * Return TRUE if the given request is colliding with a in-progress
1493  * synchronization request.
1494  */
1495 static int
1496 g_raid3_sync_collision(struct g_raid3_softc *sc, struct bio *bp)
1497 {
1498 	struct g_raid3_disk *disk;
1499 	struct bio *sbp;
1500 	off_t rstart, rend, sstart, send;
1501 	int i;
1502 
1503 	disk = sc->sc_syncdisk;
1504 	if (disk == NULL)
1505 		return (0);
1506 	rstart = bp->bio_offset;
1507 	rend = bp->bio_offset + bp->bio_length;
1508 	for (i = 0; i < g_raid3_syncreqs; i++) {
1509 		sbp = disk->d_sync.ds_bios[i];
1510 		if (sbp == NULL)
1511 			continue;
1512 		sstart = sbp->bio_offset;
1513 		send = sbp->bio_length;
1514 		if (sbp->bio_cmd == BIO_WRITE) {
1515 			sstart *= sc->sc_ndisks - 1;
1516 			send *= sc->sc_ndisks - 1;
1517 		}
1518 		send += sstart;
1519 		if (rend > sstart && rstart < send)
1520 			return (1);
1521 	}
1522 	return (0);
1523 }
1524 
1525 /*
1526  * Return TRUE if the given sync request is colliding with a in-progress regular
1527  * request.
1528  */
1529 static int
1530 g_raid3_regular_collision(struct g_raid3_softc *sc, struct bio *sbp)
1531 {
1532 	off_t rstart, rend, sstart, send;
1533 	struct bio *bp;
1534 
1535 	if (sc->sc_syncdisk == NULL)
1536 		return (0);
1537 	sstart = sbp->bio_offset;
1538 	send = sstart + sbp->bio_length;
1539 	TAILQ_FOREACH(bp, &sc->sc_inflight.queue, bio_queue) {
1540 		rstart = bp->bio_offset;
1541 		rend = bp->bio_offset + bp->bio_length;
1542 		if (rend > sstart && rstart < send)
1543 			return (1);
1544 	}
1545 	return (0);
1546 }
1547 
1548 /*
1549  * Puts request onto delayed queue.
1550  */
1551 static void
1552 g_raid3_regular_delay(struct g_raid3_softc *sc, struct bio *bp)
1553 {
1554 
1555 	G_RAID3_LOGREQ(2, bp, "Delaying request.");
1556 	bioq_insert_head(&sc->sc_regular_delayed, bp);
1557 }
1558 
1559 /*
1560  * Puts synchronization request onto delayed queue.
1561  */
1562 static void
1563 g_raid3_sync_delay(struct g_raid3_softc *sc, struct bio *bp)
1564 {
1565 
1566 	G_RAID3_LOGREQ(2, bp, "Delaying synchronization request.");
1567 	bioq_insert_tail(&sc->sc_sync_delayed, bp);
1568 }
1569 
1570 /*
1571  * Releases delayed regular requests which don't collide anymore with sync
1572  * requests.
1573  */
1574 static void
1575 g_raid3_regular_release(struct g_raid3_softc *sc)
1576 {
1577 	struct bio *bp, *bp2;
1578 
1579 	TAILQ_FOREACH_SAFE(bp, &sc->sc_regular_delayed.queue, bio_queue, bp2) {
1580 		if (g_raid3_sync_collision(sc, bp))
1581 			continue;
1582 		bioq_remove(&sc->sc_regular_delayed, bp);
1583 		G_RAID3_LOGREQ(2, bp, "Releasing delayed request (%p).", bp);
1584 		mtx_lock(&sc->sc_queue_mtx);
1585 		bioq_insert_head(&sc->sc_queue, bp);
1586 #if 0
1587 		/*
1588 		 * wakeup() is not needed, because this function is called from
1589 		 * the worker thread.
1590 		 */
1591 		wakeup(&sc->sc_queue);
1592 #endif
1593 		mtx_unlock(&sc->sc_queue_mtx);
1594 	}
1595 }
1596 
1597 /*
1598  * Releases delayed sync requests which don't collide anymore with regular
1599  * requests.
1600  */
1601 static void
1602 g_raid3_sync_release(struct g_raid3_softc *sc)
1603 {
1604 	struct bio *bp, *bp2;
1605 
1606 	TAILQ_FOREACH_SAFE(bp, &sc->sc_sync_delayed.queue, bio_queue, bp2) {
1607 		if (g_raid3_regular_collision(sc, bp))
1608 			continue;
1609 		bioq_remove(&sc->sc_sync_delayed, bp);
1610 		G_RAID3_LOGREQ(2, bp,
1611 		    "Releasing delayed synchronization request.");
1612 		g_io_request(bp, bp->bio_from);
1613 	}
1614 }
1615 
1616 /*
1617  * Handle synchronization requests.
1618  * Every synchronization request is two-steps process: first, READ request is
1619  * send to active provider and then WRITE request (with read data) to the provider
1620  * being synchronized. When WRITE is finished, new synchronization request is
1621  * send.
1622  */
1623 static void
1624 g_raid3_sync_request(struct bio *bp)
1625 {
1626 	struct g_raid3_softc *sc;
1627 	struct g_raid3_disk *disk;
1628 
1629 	bp->bio_from->index--;
1630 	sc = bp->bio_from->geom->softc;
1631 	disk = bp->bio_from->private;
1632 	if (disk == NULL) {
1633 		sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1634 		g_topology_lock();
1635 		g_raid3_kill_consumer(sc, bp->bio_from);
1636 		g_topology_unlock();
1637 		free(bp->bio_data, M_RAID3);
1638 		g_destroy_bio(bp);
1639 		sx_xlock(&sc->sc_lock);
1640 		return;
1641 	}
1642 
1643 	/*
1644 	 * Synchronization request.
1645 	 */
1646 	switch (bp->bio_cmd) {
1647 	case BIO_READ:
1648 	    {
1649 		struct g_consumer *cp;
1650 		u_char *dst, *src;
1651 		off_t left;
1652 		u_int atom;
1653 
1654 		if (bp->bio_error != 0) {
1655 			G_RAID3_LOGREQ(0, bp,
1656 			    "Synchronization request failed (error=%d).",
1657 			    bp->bio_error);
1658 			g_destroy_bio(bp);
1659 			return;
1660 		}
1661 		G_RAID3_LOGREQ(3, bp, "Synchronization request finished.");
1662 		atom = sc->sc_sectorsize / (sc->sc_ndisks - 1);
1663 		dst = src = bp->bio_data;
1664 		if (disk->d_no == sc->sc_ndisks - 1) {
1665 			u_int n;
1666 
1667 			/* Parity component. */
1668 			for (left = bp->bio_length; left > 0;
1669 			    left -= sc->sc_sectorsize) {
1670 				bcopy(src, dst, atom);
1671 				src += atom;
1672 				for (n = 1; n < sc->sc_ndisks - 1; n++) {
1673 					g_raid3_xor(src, dst, atom);
1674 					src += atom;
1675 				}
1676 				dst += atom;
1677 			}
1678 		} else {
1679 			/* Regular component. */
1680 			src += atom * disk->d_no;
1681 			for (left = bp->bio_length; left > 0;
1682 			    left -= sc->sc_sectorsize) {
1683 				bcopy(src, dst, atom);
1684 				src += sc->sc_sectorsize;
1685 				dst += atom;
1686 			}
1687 		}
1688 		bp->bio_driver1 = bp->bio_driver2 = NULL;
1689 		bp->bio_pflags = 0;
1690 		bp->bio_offset /= sc->sc_ndisks - 1;
1691 		bp->bio_length /= sc->sc_ndisks - 1;
1692 		bp->bio_cmd = BIO_WRITE;
1693 		bp->bio_cflags = 0;
1694 		bp->bio_children = bp->bio_inbed = 0;
1695 		cp = disk->d_consumer;
1696 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1697 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1698 		    cp->acr, cp->acw, cp->ace));
1699 		cp->index++;
1700 		g_io_request(bp, cp);
1701 		return;
1702 	    }
1703 	case BIO_WRITE:
1704 	    {
1705 		struct g_raid3_disk_sync *sync;
1706 		off_t boffset, moffset;
1707 		void *data;
1708 		int i;
1709 
1710 		if (bp->bio_error != 0) {
1711 			G_RAID3_LOGREQ(0, bp,
1712 			    "Synchronization request failed (error=%d).",
1713 			    bp->bio_error);
1714 			g_destroy_bio(bp);
1715 			sc->sc_bump_id |= G_RAID3_BUMP_GENID;
1716 			g_raid3_event_send(disk,
1717 			    G_RAID3_DISK_STATE_DISCONNECTED,
1718 			    G_RAID3_EVENT_DONTWAIT);
1719 			return;
1720 		}
1721 		G_RAID3_LOGREQ(3, bp, "Synchronization request finished.");
1722 		sync = &disk->d_sync;
1723 		if (sync->ds_offset == sc->sc_mediasize / (sc->sc_ndisks - 1) ||
1724 		    sync->ds_consumer == NULL ||
1725 		    (sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
1726 			/* Don't send more synchronization requests. */
1727 			sync->ds_inflight--;
1728 			if (sync->ds_bios != NULL) {
1729 				i = (int)(uintptr_t)bp->bio_caller1;
1730 				sync->ds_bios[i] = NULL;
1731 			}
1732 			free(bp->bio_data, M_RAID3);
1733 			g_destroy_bio(bp);
1734 			if (sync->ds_inflight > 0)
1735 				return;
1736 			if (sync->ds_consumer == NULL ||
1737 			    (sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
1738 				return;
1739 			}
1740 			/*
1741 			 * Disk up-to-date, activate it.
1742 			 */
1743 			g_raid3_event_send(disk, G_RAID3_DISK_STATE_ACTIVE,
1744 			    G_RAID3_EVENT_DONTWAIT);
1745 			return;
1746 		}
1747 
1748 		/* Send next synchronization request. */
1749 		data = bp->bio_data;
1750 		g_reset_bio(bp);
1751 		bp->bio_cmd = BIO_READ;
1752 		bp->bio_offset = sync->ds_offset * (sc->sc_ndisks - 1);
1753 		bp->bio_length = MIN(maxphys, sc->sc_mediasize - bp->bio_offset);
1754 		sync->ds_offset += bp->bio_length / (sc->sc_ndisks - 1);
1755 		bp->bio_done = g_raid3_sync_done;
1756 		bp->bio_data = data;
1757 		bp->bio_from = sync->ds_consumer;
1758 		bp->bio_to = sc->sc_provider;
1759 		G_RAID3_LOGREQ(3, bp, "Sending synchronization request.");
1760 		sync->ds_consumer->index++;
1761 		/*
1762 		 * Delay the request if it is colliding with a regular request.
1763 		 */
1764 		if (g_raid3_regular_collision(sc, bp))
1765 			g_raid3_sync_delay(sc, bp);
1766 		else
1767 			g_io_request(bp, sync->ds_consumer);
1768 
1769 		/* Release delayed requests if possible. */
1770 		g_raid3_regular_release(sc);
1771 
1772 		/* Find the smallest offset. */
1773 		moffset = sc->sc_mediasize;
1774 		for (i = 0; i < g_raid3_syncreqs; i++) {
1775 			bp = sync->ds_bios[i];
1776 			boffset = bp->bio_offset;
1777 			if (bp->bio_cmd == BIO_WRITE)
1778 				boffset *= sc->sc_ndisks - 1;
1779 			if (boffset < moffset)
1780 				moffset = boffset;
1781 		}
1782 		if (sync->ds_offset_done + maxphys * 100 < moffset) {
1783 			/* Update offset_done on every 100 blocks. */
1784 			sync->ds_offset_done = moffset;
1785 			g_raid3_update_metadata(disk);
1786 		}
1787 		return;
1788 	    }
1789 	default:
1790 		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1791 		    bp->bio_cmd, sc->sc_name));
1792 		break;
1793 	}
1794 }
1795 
1796 static int
1797 g_raid3_register_request(struct bio *pbp)
1798 {
1799 	struct g_raid3_softc *sc;
1800 	struct g_raid3_disk *disk;
1801 	struct g_consumer *cp;
1802 	struct bio *cbp, *tmpbp;
1803 	off_t offset, length;
1804 	u_int n, ndisks;
1805 	int round_robin, verify;
1806 
1807 	ndisks = 0;
1808 	sc = pbp->bio_to->private;
1809 	if ((pbp->bio_cflags & G_RAID3_BIO_CFLAG_REGSYNC) != 0 &&
1810 	    sc->sc_syncdisk == NULL) {
1811 		g_io_deliver(pbp, EIO);
1812 		return (0);
1813 	}
1814 	g_raid3_init_bio(pbp);
1815 	length = pbp->bio_length / (sc->sc_ndisks - 1);
1816 	offset = pbp->bio_offset / (sc->sc_ndisks - 1);
1817 	round_robin = verify = 0;
1818 	switch (pbp->bio_cmd) {
1819 	case BIO_READ:
1820 		if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_VERIFY) != 0 &&
1821 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1822 			pbp->bio_pflags |= G_RAID3_BIO_PFLAG_VERIFY;
1823 			verify = 1;
1824 			ndisks = sc->sc_ndisks;
1825 		} else {
1826 			verify = 0;
1827 			ndisks = sc->sc_ndisks - 1;
1828 		}
1829 		if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_ROUND_ROBIN) != 0 &&
1830 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1831 			round_robin = 1;
1832 		} else {
1833 			round_robin = 0;
1834 		}
1835 		KASSERT(!round_robin || !verify,
1836 		    ("ROUND-ROBIN and VERIFY are mutually exclusive."));
1837 		pbp->bio_driver2 = &sc->sc_disks[sc->sc_ndisks - 1];
1838 		break;
1839 	case BIO_WRITE:
1840 	case BIO_DELETE:
1841 		/*
1842 		 * Delay the request if it is colliding with a synchronization
1843 		 * request.
1844 		 */
1845 		if (g_raid3_sync_collision(sc, pbp)) {
1846 			g_raid3_regular_delay(sc, pbp);
1847 			return (0);
1848 		}
1849 
1850 		if (sc->sc_idle)
1851 			g_raid3_unidle(sc);
1852 		else
1853 			sc->sc_last_write = time_uptime;
1854 
1855 		ndisks = sc->sc_ndisks;
1856 		break;
1857 	}
1858 	for (n = 0; n < ndisks; n++) {
1859 		disk = &sc->sc_disks[n];
1860 		cbp = g_raid3_clone_bio(sc, pbp);
1861 		if (cbp == NULL) {
1862 			while ((cbp = G_RAID3_HEAD_BIO(pbp)) != NULL)
1863 				g_raid3_destroy_bio(sc, cbp);
1864 			/*
1865 			 * To prevent deadlock, we must run back up
1866 			 * with the ENOMEM for failed requests of any
1867 			 * of our consumers.  Our own sync requests
1868 			 * can stick around, as they are finite.
1869 			 */
1870 			if ((pbp->bio_cflags &
1871 			    G_RAID3_BIO_CFLAG_REGULAR) != 0) {
1872 				g_io_deliver(pbp, ENOMEM);
1873 				return (0);
1874 			}
1875 			return (ENOMEM);
1876 		}
1877 		cbp->bio_offset = offset;
1878 		cbp->bio_length = length;
1879 		cbp->bio_done = g_raid3_done;
1880 		switch (pbp->bio_cmd) {
1881 		case BIO_READ:
1882 			if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE) {
1883 				/*
1884 				 * Replace invalid component with the parity
1885 				 * component.
1886 				 */
1887 				disk = &sc->sc_disks[sc->sc_ndisks - 1];
1888 				cbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1889 				pbp->bio_pflags |= G_RAID3_BIO_PFLAG_DEGRADED;
1890 			} else if (round_robin &&
1891 			    disk->d_no == sc->sc_round_robin) {
1892 				/*
1893 				 * In round-robin mode skip one data component
1894 				 * and use parity component when reading.
1895 				 */
1896 				pbp->bio_driver2 = disk;
1897 				disk = &sc->sc_disks[sc->sc_ndisks - 1];
1898 				cbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1899 				sc->sc_round_robin++;
1900 				round_robin = 0;
1901 			} else if (verify && disk->d_no == sc->sc_ndisks - 1) {
1902 				cbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1903 			}
1904 			break;
1905 		case BIO_WRITE:
1906 		case BIO_DELETE:
1907 			if (disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
1908 			    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
1909 				if (n == ndisks - 1) {
1910 					/*
1911 					 * Active parity component, mark it as such.
1912 					 */
1913 					cbp->bio_cflags |=
1914 					    G_RAID3_BIO_CFLAG_PARITY;
1915 				}
1916 			} else {
1917 				pbp->bio_pflags |= G_RAID3_BIO_PFLAG_DEGRADED;
1918 				if (n == ndisks - 1) {
1919 					/*
1920 					 * Parity component is not connected,
1921 					 * so destroy its request.
1922 					 */
1923 					pbp->bio_pflags |=
1924 					    G_RAID3_BIO_PFLAG_NOPARITY;
1925 					g_raid3_destroy_bio(sc, cbp);
1926 					cbp = NULL;
1927 				} else {
1928 					cbp->bio_cflags |=
1929 					    G_RAID3_BIO_CFLAG_NODISK;
1930 					disk = NULL;
1931 				}
1932 			}
1933 			break;
1934 		}
1935 		if (cbp != NULL)
1936 			cbp->bio_caller2 = disk;
1937 	}
1938 	switch (pbp->bio_cmd) {
1939 	case BIO_READ:
1940 		if (round_robin) {
1941 			/*
1942 			 * If we are in round-robin mode and 'round_robin' is
1943 			 * still 1, it means, that we skipped parity component
1944 			 * for this read and must reset sc_round_robin field.
1945 			 */
1946 			sc->sc_round_robin = 0;
1947 		}
1948 		G_RAID3_FOREACH_SAFE_BIO(pbp, cbp, tmpbp) {
1949 			disk = cbp->bio_caller2;
1950 			cp = disk->d_consumer;
1951 			cbp->bio_to = cp->provider;
1952 			G_RAID3_LOGREQ(3, cbp, "Sending request.");
1953 			KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1954 			    ("Consumer %s not opened (r%dw%de%d).",
1955 			    cp->provider->name, cp->acr, cp->acw, cp->ace));
1956 			cp->index++;
1957 			g_io_request(cbp, cp);
1958 		}
1959 		break;
1960 	case BIO_WRITE:
1961 	case BIO_DELETE:
1962 		/*
1963 		 * Put request onto inflight queue, so we can check if new
1964 		 * synchronization requests don't collide with it.
1965 		 */
1966 		bioq_insert_tail(&sc->sc_inflight, pbp);
1967 
1968 		/*
1969 		 * Bump syncid on first write.
1970 		 */
1971 		if ((sc->sc_bump_id & G_RAID3_BUMP_SYNCID) != 0) {
1972 			sc->sc_bump_id &= ~G_RAID3_BUMP_SYNCID;
1973 			g_raid3_bump_syncid(sc);
1974 		}
1975 		g_raid3_scatter(pbp);
1976 		break;
1977 	}
1978 	return (0);
1979 }
1980 
1981 static int
1982 g_raid3_can_destroy(struct g_raid3_softc *sc)
1983 {
1984 	struct g_geom *gp;
1985 	struct g_consumer *cp;
1986 
1987 	g_topology_assert();
1988 	gp = sc->sc_geom;
1989 	if (gp->softc == NULL)
1990 		return (1);
1991 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1992 		if (g_raid3_is_busy(sc, cp))
1993 			return (0);
1994 	}
1995 	gp = sc->sc_sync.ds_geom;
1996 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1997 		if (g_raid3_is_busy(sc, cp))
1998 			return (0);
1999 	}
2000 	G_RAID3_DEBUG(2, "No I/O requests for %s, it can be destroyed.",
2001 	    sc->sc_name);
2002 	return (1);
2003 }
2004 
2005 static int
2006 g_raid3_try_destroy(struct g_raid3_softc *sc)
2007 {
2008 
2009 	g_topology_assert_not();
2010 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2011 
2012 	if (sc->sc_rootmount != NULL) {
2013 		G_RAID3_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2014 		    sc->sc_rootmount);
2015 		root_mount_rel(sc->sc_rootmount);
2016 		sc->sc_rootmount = NULL;
2017 	}
2018 
2019 	g_topology_lock();
2020 	if (!g_raid3_can_destroy(sc)) {
2021 		g_topology_unlock();
2022 		return (0);
2023 	}
2024 	sc->sc_geom->softc = NULL;
2025 	sc->sc_sync.ds_geom->softc = NULL;
2026 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_WAIT) != 0) {
2027 		g_topology_unlock();
2028 		G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__,
2029 		    &sc->sc_worker);
2030 		/* Unlock sc_lock here, as it can be destroyed after wakeup. */
2031 		sx_xunlock(&sc->sc_lock);
2032 		wakeup(&sc->sc_worker);
2033 		sc->sc_worker = NULL;
2034 	} else {
2035 		g_topology_unlock();
2036 		g_raid3_free_device(sc);
2037 	}
2038 	return (1);
2039 }
2040 
2041 /*
2042  * Worker thread.
2043  */
2044 static void
2045 g_raid3_worker(void *arg)
2046 {
2047 	struct g_raid3_softc *sc;
2048 	struct g_raid3_event *ep;
2049 	struct bio *bp;
2050 	int timeout;
2051 
2052 	sc = arg;
2053 	thread_lock(curthread);
2054 	sched_prio(curthread, PRIBIO);
2055 	thread_unlock(curthread);
2056 
2057 	sx_xlock(&sc->sc_lock);
2058 	for (;;) {
2059 		G_RAID3_DEBUG(5, "%s: Let's see...", __func__);
2060 		/*
2061 		 * First take a look at events.
2062 		 * This is important to handle events before any I/O requests.
2063 		 */
2064 		ep = g_raid3_event_get(sc);
2065 		if (ep != NULL) {
2066 			g_raid3_event_remove(sc, ep);
2067 			if ((ep->e_flags & G_RAID3_EVENT_DEVICE) != 0) {
2068 				/* Update only device status. */
2069 				G_RAID3_DEBUG(3,
2070 				    "Running event for device %s.",
2071 				    sc->sc_name);
2072 				ep->e_error = 0;
2073 				g_raid3_update_device(sc, 1);
2074 			} else {
2075 				/* Update disk status. */
2076 				G_RAID3_DEBUG(3, "Running event for disk %s.",
2077 				     g_raid3_get_diskname(ep->e_disk));
2078 				ep->e_error = g_raid3_update_disk(ep->e_disk,
2079 				    ep->e_state);
2080 				if (ep->e_error == 0)
2081 					g_raid3_update_device(sc, 0);
2082 			}
2083 			if ((ep->e_flags & G_RAID3_EVENT_DONTWAIT) != 0) {
2084 				KASSERT(ep->e_error == 0,
2085 				    ("Error cannot be handled."));
2086 				g_raid3_event_free(ep);
2087 			} else {
2088 				ep->e_flags |= G_RAID3_EVENT_DONE;
2089 				G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__,
2090 				    ep);
2091 				mtx_lock(&sc->sc_events_mtx);
2092 				wakeup(ep);
2093 				mtx_unlock(&sc->sc_events_mtx);
2094 			}
2095 			if ((sc->sc_flags &
2096 			    G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
2097 				if (g_raid3_try_destroy(sc)) {
2098 					curthread->td_pflags &= ~TDP_GEOM;
2099 					G_RAID3_DEBUG(1, "Thread exiting.");
2100 					kproc_exit(0);
2101 				}
2102 			}
2103 			G_RAID3_DEBUG(5, "%s: I'm here 1.", __func__);
2104 			continue;
2105 		}
2106 		/*
2107 		 * Check if we can mark array as CLEAN and if we can't take
2108 		 * how much seconds should we wait.
2109 		 */
2110 		timeout = g_raid3_idle(sc, -1);
2111 		/*
2112 		 * Now I/O requests.
2113 		 */
2114 		/* Get first request from the queue. */
2115 		mtx_lock(&sc->sc_queue_mtx);
2116 		bp = bioq_first(&sc->sc_queue);
2117 		if (bp == NULL) {
2118 			if ((sc->sc_flags &
2119 			    G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
2120 				mtx_unlock(&sc->sc_queue_mtx);
2121 				if (g_raid3_try_destroy(sc)) {
2122 					curthread->td_pflags &= ~TDP_GEOM;
2123 					G_RAID3_DEBUG(1, "Thread exiting.");
2124 					kproc_exit(0);
2125 				}
2126 				mtx_lock(&sc->sc_queue_mtx);
2127 			}
2128 			sx_xunlock(&sc->sc_lock);
2129 			/*
2130 			 * XXX: We can miss an event here, because an event
2131 			 *      can be added without sx-device-lock and without
2132 			 *      mtx-queue-lock. Maybe I should just stop using
2133 			 *      dedicated mutex for events synchronization and
2134 			 *      stick with the queue lock?
2135 			 *      The event will hang here until next I/O request
2136 			 *      or next event is received.
2137 			 */
2138 			MSLEEP(sc, &sc->sc_queue_mtx, PRIBIO | PDROP, "r3:w1",
2139 			    timeout * hz);
2140 			sx_xlock(&sc->sc_lock);
2141 			G_RAID3_DEBUG(5, "%s: I'm here 4.", __func__);
2142 			continue;
2143 		}
2144 process:
2145 		bioq_remove(&sc->sc_queue, bp);
2146 		mtx_unlock(&sc->sc_queue_mtx);
2147 
2148 		if (bp->bio_from->geom == sc->sc_sync.ds_geom &&
2149 		    (bp->bio_cflags & G_RAID3_BIO_CFLAG_SYNC) != 0) {
2150 			g_raid3_sync_request(bp);	/* READ */
2151 		} else if (bp->bio_to != sc->sc_provider) {
2152 			if ((bp->bio_cflags & G_RAID3_BIO_CFLAG_REGULAR) != 0)
2153 				g_raid3_regular_request(bp);
2154 			else if ((bp->bio_cflags & G_RAID3_BIO_CFLAG_SYNC) != 0)
2155 				g_raid3_sync_request(bp);	/* WRITE */
2156 			else {
2157 				KASSERT(0,
2158 				    ("Invalid request cflags=0x%hx to=%s.",
2159 				    bp->bio_cflags, bp->bio_to->name));
2160 			}
2161 		} else if (g_raid3_register_request(bp) != 0) {
2162 			mtx_lock(&sc->sc_queue_mtx);
2163 			bioq_insert_head(&sc->sc_queue, bp);
2164 			/*
2165 			 * We are short in memory, let see if there are finished
2166 			 * request we can free.
2167 			 */
2168 			TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
2169 				if (bp->bio_cflags & G_RAID3_BIO_CFLAG_REGULAR)
2170 					goto process;
2171 			}
2172 			/*
2173 			 * No finished regular request, so at least keep
2174 			 * synchronization running.
2175 			 */
2176 			TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
2177 				if (bp->bio_cflags & G_RAID3_BIO_CFLAG_SYNC)
2178 					goto process;
2179 			}
2180 			sx_xunlock(&sc->sc_lock);
2181 			MSLEEP(&sc->sc_queue, &sc->sc_queue_mtx, PRIBIO | PDROP,
2182 			    "r3:lowmem", hz / 10);
2183 			sx_xlock(&sc->sc_lock);
2184 		}
2185 		G_RAID3_DEBUG(5, "%s: I'm here 9.", __func__);
2186 	}
2187 }
2188 
2189 static void
2190 g_raid3_update_idle(struct g_raid3_softc *sc, struct g_raid3_disk *disk)
2191 {
2192 
2193 	sx_assert(&sc->sc_lock, SX_LOCKED);
2194 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) != 0)
2195 		return;
2196 	if (!sc->sc_idle && (disk->d_flags & G_RAID3_DISK_FLAG_DIRTY) == 0) {
2197 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as dirty.",
2198 		    g_raid3_get_diskname(disk), sc->sc_name);
2199 		disk->d_flags |= G_RAID3_DISK_FLAG_DIRTY;
2200 	} else if (sc->sc_idle &&
2201 	    (disk->d_flags & G_RAID3_DISK_FLAG_DIRTY) != 0) {
2202 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as clean.",
2203 		    g_raid3_get_diskname(disk), sc->sc_name);
2204 		disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2205 	}
2206 }
2207 
2208 static void
2209 g_raid3_sync_start(struct g_raid3_softc *sc)
2210 {
2211 	struct g_raid3_disk *disk;
2212 	struct g_consumer *cp;
2213 	struct bio *bp;
2214 	int error __diagused;
2215 	u_int n;
2216 
2217 	g_topology_assert_not();
2218 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2219 
2220 	KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED,
2221 	    ("Device not in DEGRADED state (%s, %u).", sc->sc_name,
2222 	    sc->sc_state));
2223 	KASSERT(sc->sc_syncdisk == NULL, ("Syncdisk is not NULL (%s, %u).",
2224 	    sc->sc_name, sc->sc_state));
2225 	disk = NULL;
2226 	for (n = 0; n < sc->sc_ndisks; n++) {
2227 		if (sc->sc_disks[n].d_state != G_RAID3_DISK_STATE_SYNCHRONIZING)
2228 			continue;
2229 		disk = &sc->sc_disks[n];
2230 		break;
2231 	}
2232 	if (disk == NULL)
2233 		return;
2234 
2235 	sx_xunlock(&sc->sc_lock);
2236 	g_topology_lock();
2237 	cp = g_new_consumer(sc->sc_sync.ds_geom);
2238 	error = g_attach(cp, sc->sc_provider);
2239 	KASSERT(error == 0,
2240 	    ("Cannot attach to %s (error=%d).", sc->sc_name, error));
2241 	error = g_access(cp, 1, 0, 0);
2242 	KASSERT(error == 0, ("Cannot open %s (error=%d).", sc->sc_name, error));
2243 	g_topology_unlock();
2244 	sx_xlock(&sc->sc_lock);
2245 
2246 	G_RAID3_DEBUG(0, "Device %s: rebuilding provider %s.", sc->sc_name,
2247 	    g_raid3_get_diskname(disk));
2248 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOFAILSYNC) == 0)
2249 		disk->d_flags |= G_RAID3_DISK_FLAG_DIRTY;
2250 	KASSERT(disk->d_sync.ds_consumer == NULL,
2251 	    ("Sync consumer already exists (device=%s, disk=%s).",
2252 	    sc->sc_name, g_raid3_get_diskname(disk)));
2253 
2254 	disk->d_sync.ds_consumer = cp;
2255 	disk->d_sync.ds_consumer->private = disk;
2256 	disk->d_sync.ds_consumer->index = 0;
2257 	sc->sc_syncdisk = disk;
2258 
2259 	/*
2260 	 * Allocate memory for synchronization bios and initialize them.
2261 	 */
2262 	disk->d_sync.ds_bios = malloc(sizeof(struct bio *) * g_raid3_syncreqs,
2263 	    M_RAID3, M_WAITOK);
2264 	for (n = 0; n < g_raid3_syncreqs; n++) {
2265 		bp = g_alloc_bio();
2266 		disk->d_sync.ds_bios[n] = bp;
2267 		bp->bio_parent = NULL;
2268 		bp->bio_cmd = BIO_READ;
2269 		bp->bio_data = malloc(maxphys, M_RAID3, M_WAITOK);
2270 		bp->bio_cflags = 0;
2271 		bp->bio_offset = disk->d_sync.ds_offset * (sc->sc_ndisks - 1);
2272 		bp->bio_length = MIN(maxphys, sc->sc_mediasize - bp->bio_offset);
2273 		disk->d_sync.ds_offset += bp->bio_length / (sc->sc_ndisks - 1);
2274 		bp->bio_done = g_raid3_sync_done;
2275 		bp->bio_from = disk->d_sync.ds_consumer;
2276 		bp->bio_to = sc->sc_provider;
2277 		bp->bio_caller1 = (void *)(uintptr_t)n;
2278 	}
2279 
2280 	/* Set the number of in-flight synchronization requests. */
2281 	disk->d_sync.ds_inflight = g_raid3_syncreqs;
2282 
2283 	/*
2284 	 * Fire off first synchronization requests.
2285 	 */
2286 	for (n = 0; n < g_raid3_syncreqs; n++) {
2287 		bp = disk->d_sync.ds_bios[n];
2288 		G_RAID3_LOGREQ(3, bp, "Sending synchronization request.");
2289 		disk->d_sync.ds_consumer->index++;
2290 		/*
2291 		 * Delay the request if it is colliding with a regular request.
2292 		 */
2293 		if (g_raid3_regular_collision(sc, bp))
2294 			g_raid3_sync_delay(sc, bp);
2295 		else
2296 			g_io_request(bp, disk->d_sync.ds_consumer);
2297 	}
2298 }
2299 
2300 /*
2301  * Stop synchronization process.
2302  * type: 0 - synchronization finished
2303  *       1 - synchronization stopped
2304  */
2305 static void
2306 g_raid3_sync_stop(struct g_raid3_softc *sc, int type)
2307 {
2308 	struct g_raid3_disk *disk;
2309 	struct g_consumer *cp;
2310 
2311 	g_topology_assert_not();
2312 	sx_assert(&sc->sc_lock, SX_LOCKED);
2313 
2314 	KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED,
2315 	    ("Device not in DEGRADED state (%s, %u).", sc->sc_name,
2316 	    sc->sc_state));
2317 	disk = sc->sc_syncdisk;
2318 	sc->sc_syncdisk = NULL;
2319 	KASSERT(disk != NULL, ("No disk was synchronized (%s).", sc->sc_name));
2320 	KASSERT(disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING,
2321 	    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2322 	    g_raid3_disk_state2str(disk->d_state)));
2323 	if (disk->d_sync.ds_consumer == NULL)
2324 		return;
2325 
2326 	if (type == 0) {
2327 		G_RAID3_DEBUG(0, "Device %s: rebuilding provider %s finished.",
2328 		    sc->sc_name, g_raid3_get_diskname(disk));
2329 	} else /* if (type == 1) */ {
2330 		G_RAID3_DEBUG(0, "Device %s: rebuilding provider %s stopped.",
2331 		    sc->sc_name, g_raid3_get_diskname(disk));
2332 	}
2333 	free(disk->d_sync.ds_bios, M_RAID3);
2334 	disk->d_sync.ds_bios = NULL;
2335 	cp = disk->d_sync.ds_consumer;
2336 	disk->d_sync.ds_consumer = NULL;
2337 	disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2338 	sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
2339 	g_topology_lock();
2340 	g_raid3_kill_consumer(sc, cp);
2341 	g_topology_unlock();
2342 	sx_xlock(&sc->sc_lock);
2343 }
2344 
2345 static void
2346 g_raid3_launch_provider(struct g_raid3_softc *sc)
2347 {
2348 	struct g_provider *pp;
2349 	struct g_raid3_disk *disk;
2350 	int n;
2351 
2352 	sx_assert(&sc->sc_lock, SX_LOCKED);
2353 
2354 	g_topology_lock();
2355 	pp = g_new_providerf(sc->sc_geom, "raid3/%s", sc->sc_name);
2356 	pp->mediasize = sc->sc_mediasize;
2357 	pp->sectorsize = sc->sc_sectorsize;
2358 	pp->stripesize = 0;
2359 	pp->stripeoffset = 0;
2360 	for (n = 0; n < sc->sc_ndisks; n++) {
2361 		disk = &sc->sc_disks[n];
2362 		if (disk->d_consumer && disk->d_consumer->provider &&
2363 		    disk->d_consumer->provider->stripesize > pp->stripesize) {
2364 			pp->stripesize = disk->d_consumer->provider->stripesize;
2365 			pp->stripeoffset = disk->d_consumer->provider->stripeoffset;
2366 		}
2367 	}
2368 	pp->stripesize *= sc->sc_ndisks - 1;
2369 	pp->stripeoffset *= sc->sc_ndisks - 1;
2370 	pp->private = sc;
2371 	sc->sc_refcnt++;
2372 	sc->sc_provider = pp;
2373 	g_error_provider(pp, 0);
2374 	g_topology_unlock();
2375 	G_RAID3_DEBUG(0, "Device %s launched (%u/%u).", pp->name,
2376 	    g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE), sc->sc_ndisks);
2377 
2378 	if (sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED)
2379 		g_raid3_sync_start(sc);
2380 }
2381 
2382 static void
2383 g_raid3_destroy_provider(struct g_raid3_softc *sc)
2384 {
2385 	struct bio *bp;
2386 
2387 	g_topology_assert_not();
2388 	KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
2389 	    sc->sc_name));
2390 
2391 	g_topology_lock();
2392 	g_error_provider(sc->sc_provider, ENXIO);
2393 	mtx_lock(&sc->sc_queue_mtx);
2394 	while ((bp = bioq_first(&sc->sc_queue)) != NULL) {
2395 		bioq_remove(&sc->sc_queue, bp);
2396 		g_io_deliver(bp, ENXIO);
2397 	}
2398 	mtx_unlock(&sc->sc_queue_mtx);
2399 	G_RAID3_DEBUG(0, "Device %s: provider %s destroyed.", sc->sc_name,
2400 	    sc->sc_provider->name);
2401 	g_wither_provider(sc->sc_provider, ENXIO);
2402 	g_topology_unlock();
2403 	sc->sc_provider = NULL;
2404 	if (sc->sc_syncdisk != NULL)
2405 		g_raid3_sync_stop(sc, 1);
2406 }
2407 
2408 static void
2409 g_raid3_go(void *arg)
2410 {
2411 	struct g_raid3_softc *sc;
2412 	struct g_raid3_event *ep;
2413 
2414 	sc = arg;
2415 	G_RAID3_DEBUG(0, "Force device %s start due to timeout.", sc->sc_name);
2416 	ep = sc->sc_timeout_event;
2417 	sc->sc_timeout_event = NULL;
2418 	g_raid3_event_dispatch(ep, sc, 0,
2419 	    G_RAID3_EVENT_DONTWAIT | G_RAID3_EVENT_DEVICE);
2420 }
2421 
2422 static void
2423 g_raid3_timeout_drain(struct g_raid3_softc *sc)
2424 {
2425 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2426 
2427 	callout_drain(&sc->sc_callout);
2428 	g_raid3_event_free(sc->sc_timeout_event);
2429 	sc->sc_timeout_event = NULL;
2430 }
2431 
2432 static u_int
2433 g_raid3_determine_state(struct g_raid3_disk *disk)
2434 {
2435 	struct g_raid3_softc *sc;
2436 	u_int state;
2437 
2438 	sc = disk->d_softc;
2439 	if (sc->sc_syncid == disk->d_sync.ds_syncid) {
2440 		if ((disk->d_flags &
2441 		    G_RAID3_DISK_FLAG_SYNCHRONIZING) == 0) {
2442 			/* Disk does not need synchronization. */
2443 			state = G_RAID3_DISK_STATE_ACTIVE;
2444 		} else {
2445 			if ((sc->sc_flags &
2446 			     G_RAID3_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2447 			    (disk->d_flags &
2448 			     G_RAID3_DISK_FLAG_FORCE_SYNC) != 0) {
2449 				/*
2450 				 * We can start synchronization from
2451 				 * the stored offset.
2452 				 */
2453 				state = G_RAID3_DISK_STATE_SYNCHRONIZING;
2454 			} else {
2455 				state = G_RAID3_DISK_STATE_STALE;
2456 			}
2457 		}
2458 	} else if (disk->d_sync.ds_syncid < sc->sc_syncid) {
2459 		/*
2460 		 * Reset all synchronization data for this disk,
2461 		 * because if it even was synchronized, it was
2462 		 * synchronized to disks with different syncid.
2463 		 */
2464 		disk->d_flags |= G_RAID3_DISK_FLAG_SYNCHRONIZING;
2465 		disk->d_sync.ds_offset = 0;
2466 		disk->d_sync.ds_offset_done = 0;
2467 		disk->d_sync.ds_syncid = sc->sc_syncid;
2468 		if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2469 		    (disk->d_flags & G_RAID3_DISK_FLAG_FORCE_SYNC) != 0) {
2470 			state = G_RAID3_DISK_STATE_SYNCHRONIZING;
2471 		} else {
2472 			state = G_RAID3_DISK_STATE_STALE;
2473 		}
2474 	} else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
2475 		/*
2476 		 * Not good, NOT GOOD!
2477 		 * It means that device was started on stale disks
2478 		 * and more fresh disk just arrive.
2479 		 * If there were writes, device is broken, sorry.
2480 		 * I think the best choice here is don't touch
2481 		 * this disk and inform the user loudly.
2482 		 */
2483 		G_RAID3_DEBUG(0, "Device %s was started before the freshest "
2484 		    "disk (%s) arrives!! It will not be connected to the "
2485 		    "running device.", sc->sc_name,
2486 		    g_raid3_get_diskname(disk));
2487 		g_raid3_destroy_disk(disk);
2488 		state = G_RAID3_DISK_STATE_NONE;
2489 		/* Return immediately, because disk was destroyed. */
2490 		return (state);
2491 	}
2492 	G_RAID3_DEBUG(3, "State for %s disk: %s.",
2493 	    g_raid3_get_diskname(disk), g_raid3_disk_state2str(state));
2494 	return (state);
2495 }
2496 
2497 /*
2498  * Update device state.
2499  */
2500 static void
2501 g_raid3_update_device(struct g_raid3_softc *sc, boolean_t force)
2502 {
2503 	struct g_raid3_disk *disk;
2504 	u_int state;
2505 
2506 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2507 
2508 	switch (sc->sc_state) {
2509 	case G_RAID3_DEVICE_STATE_STARTING:
2510 	    {
2511 		u_int n, ndirty, ndisks, genid, syncid;
2512 
2513 		KASSERT(sc->sc_provider == NULL,
2514 		    ("Non-NULL provider in STARTING state (%s).", sc->sc_name));
2515 		/*
2516 		 * Are we ready? We are, if all disks are connected or
2517 		 * one disk is missing and 'force' is true.
2518 		 */
2519 		if (g_raid3_ndisks(sc, -1) + force == sc->sc_ndisks) {
2520 			if (!force)
2521 				g_raid3_timeout_drain(sc);
2522 		} else {
2523 			if (force) {
2524 				/*
2525 				 * Timeout expired, so destroy device.
2526 				 */
2527 				sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
2528 				G_RAID3_DEBUG(1, "root_mount_rel[%u] %p",
2529 				    __LINE__, sc->sc_rootmount);
2530 				root_mount_rel(sc->sc_rootmount);
2531 				sc->sc_rootmount = NULL;
2532 			}
2533 			return;
2534 		}
2535 
2536 		/*
2537 		 * Find the biggest genid.
2538 		 */
2539 		genid = 0;
2540 		for (n = 0; n < sc->sc_ndisks; n++) {
2541 			disk = &sc->sc_disks[n];
2542 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2543 				continue;
2544 			if (disk->d_genid > genid)
2545 				genid = disk->d_genid;
2546 		}
2547 		sc->sc_genid = genid;
2548 		/*
2549 		 * Remove all disks without the biggest genid.
2550 		 */
2551 		for (n = 0; n < sc->sc_ndisks; n++) {
2552 			disk = &sc->sc_disks[n];
2553 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2554 				continue;
2555 			if (disk->d_genid < genid) {
2556 				G_RAID3_DEBUG(0,
2557 				    "Component %s (device %s) broken, skipping.",
2558 				    g_raid3_get_diskname(disk), sc->sc_name);
2559 				g_raid3_destroy_disk(disk);
2560 			}
2561 		}
2562 
2563 		/*
2564 		 * There must be at least 'sc->sc_ndisks - 1' components
2565 		 * with the same syncid and without SYNCHRONIZING flag.
2566 		 */
2567 
2568 		/*
2569 		 * Find the biggest syncid, number of valid components and
2570 		 * number of dirty components.
2571 		 */
2572 		ndirty = ndisks = syncid = 0;
2573 		for (n = 0; n < sc->sc_ndisks; n++) {
2574 			disk = &sc->sc_disks[n];
2575 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2576 				continue;
2577 			if ((disk->d_flags & G_RAID3_DISK_FLAG_DIRTY) != 0)
2578 				ndirty++;
2579 			if (disk->d_sync.ds_syncid > syncid) {
2580 				syncid = disk->d_sync.ds_syncid;
2581 				ndisks = 0;
2582 			} else if (disk->d_sync.ds_syncid < syncid) {
2583 				continue;
2584 			}
2585 			if ((disk->d_flags &
2586 			    G_RAID3_DISK_FLAG_SYNCHRONIZING) != 0) {
2587 				continue;
2588 			}
2589 			ndisks++;
2590 		}
2591 		/*
2592 		 * Do we have enough valid components?
2593 		 */
2594 		if (ndisks + 1 < sc->sc_ndisks) {
2595 			G_RAID3_DEBUG(0,
2596 			    "Device %s is broken, too few valid components.",
2597 			    sc->sc_name);
2598 			sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
2599 			return;
2600 		}
2601 		/*
2602 		 * If there is one DIRTY component and all disks are present,
2603 		 * mark it for synchronization. If there is more than one DIRTY
2604 		 * component, mark parity component for synchronization.
2605 		 */
2606 		if (ndisks == sc->sc_ndisks && ndirty == 1) {
2607 			for (n = 0; n < sc->sc_ndisks; n++) {
2608 				disk = &sc->sc_disks[n];
2609 				if ((disk->d_flags &
2610 				    G_RAID3_DISK_FLAG_DIRTY) == 0) {
2611 					continue;
2612 				}
2613 				disk->d_flags |=
2614 				    G_RAID3_DISK_FLAG_SYNCHRONIZING;
2615 			}
2616 		} else if (ndisks == sc->sc_ndisks && ndirty > 1) {
2617 			disk = &sc->sc_disks[sc->sc_ndisks - 1];
2618 			disk->d_flags |= G_RAID3_DISK_FLAG_SYNCHRONIZING;
2619 		}
2620 
2621 		sc->sc_syncid = syncid;
2622 		if (force) {
2623 			/* Remember to bump syncid on first write. */
2624 			sc->sc_bump_id |= G_RAID3_BUMP_SYNCID;
2625 		}
2626 		if (ndisks == sc->sc_ndisks)
2627 			state = G_RAID3_DEVICE_STATE_COMPLETE;
2628 		else /* if (ndisks == sc->sc_ndisks - 1) */
2629 			state = G_RAID3_DEVICE_STATE_DEGRADED;
2630 		G_RAID3_DEBUG(1, "Device %s state changed from %s to %s.",
2631 		    sc->sc_name, g_raid3_device_state2str(sc->sc_state),
2632 		    g_raid3_device_state2str(state));
2633 		sc->sc_state = state;
2634 		for (n = 0; n < sc->sc_ndisks; n++) {
2635 			disk = &sc->sc_disks[n];
2636 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2637 				continue;
2638 			state = g_raid3_determine_state(disk);
2639 			g_raid3_event_send(disk, state, G_RAID3_EVENT_DONTWAIT);
2640 			if (state == G_RAID3_DISK_STATE_STALE)
2641 				sc->sc_bump_id |= G_RAID3_BUMP_SYNCID;
2642 		}
2643 		break;
2644 	    }
2645 	case G_RAID3_DEVICE_STATE_DEGRADED:
2646 		/*
2647 		 * Genid need to be bumped immediately, so do it here.
2648 		 */
2649 		if ((sc->sc_bump_id & G_RAID3_BUMP_GENID) != 0) {
2650 			sc->sc_bump_id &= ~G_RAID3_BUMP_GENID;
2651 			g_raid3_bump_genid(sc);
2652 		}
2653 
2654 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_NEW) > 0)
2655 			return;
2656 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) <
2657 		    sc->sc_ndisks - 1) {
2658 			if (sc->sc_provider != NULL)
2659 				g_raid3_destroy_provider(sc);
2660 			sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
2661 			return;
2662 		}
2663 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) ==
2664 		    sc->sc_ndisks) {
2665 			state = G_RAID3_DEVICE_STATE_COMPLETE;
2666 			G_RAID3_DEBUG(1,
2667 			    "Device %s state changed from %s to %s.",
2668 			    sc->sc_name, g_raid3_device_state2str(sc->sc_state),
2669 			    g_raid3_device_state2str(state));
2670 			sc->sc_state = state;
2671 		}
2672 		if (sc->sc_provider == NULL)
2673 			g_raid3_launch_provider(sc);
2674 		if (sc->sc_rootmount != NULL) {
2675 			G_RAID3_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2676 			    sc->sc_rootmount);
2677 			root_mount_rel(sc->sc_rootmount);
2678 			sc->sc_rootmount = NULL;
2679 		}
2680 		break;
2681 	case G_RAID3_DEVICE_STATE_COMPLETE:
2682 		/*
2683 		 * Genid need to be bumped immediately, so do it here.
2684 		 */
2685 		if ((sc->sc_bump_id & G_RAID3_BUMP_GENID) != 0) {
2686 			sc->sc_bump_id &= ~G_RAID3_BUMP_GENID;
2687 			g_raid3_bump_genid(sc);
2688 		}
2689 
2690 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_NEW) > 0)
2691 			return;
2692 		KASSERT(g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) >=
2693 		    sc->sc_ndisks - 1,
2694 		    ("Too few ACTIVE components in COMPLETE state (device %s).",
2695 		    sc->sc_name));
2696 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) ==
2697 		    sc->sc_ndisks - 1) {
2698 			state = G_RAID3_DEVICE_STATE_DEGRADED;
2699 			G_RAID3_DEBUG(1,
2700 			    "Device %s state changed from %s to %s.",
2701 			    sc->sc_name, g_raid3_device_state2str(sc->sc_state),
2702 			    g_raid3_device_state2str(state));
2703 			sc->sc_state = state;
2704 		}
2705 		if (sc->sc_provider == NULL)
2706 			g_raid3_launch_provider(sc);
2707 		if (sc->sc_rootmount != NULL) {
2708 			G_RAID3_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2709 			    sc->sc_rootmount);
2710 			root_mount_rel(sc->sc_rootmount);
2711 			sc->sc_rootmount = NULL;
2712 		}
2713 		break;
2714 	default:
2715 		KASSERT(1 == 0, ("Wrong device state (%s, %s).", sc->sc_name,
2716 		    g_raid3_device_state2str(sc->sc_state)));
2717 		break;
2718 	}
2719 }
2720 
2721 /*
2722  * Update disk state and device state if needed.
2723  */
2724 #define	DISK_STATE_CHANGED()	G_RAID3_DEBUG(1,			\
2725 	"Disk %s state changed from %s to %s (device %s).",		\
2726 	g_raid3_get_diskname(disk),					\
2727 	g_raid3_disk_state2str(disk->d_state),				\
2728 	g_raid3_disk_state2str(state), sc->sc_name)
2729 static int
2730 g_raid3_update_disk(struct g_raid3_disk *disk, u_int state)
2731 {
2732 	struct g_raid3_softc *sc;
2733 
2734 	sc = disk->d_softc;
2735 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2736 
2737 again:
2738 	G_RAID3_DEBUG(3, "Changing disk %s state from %s to %s.",
2739 	    g_raid3_get_diskname(disk), g_raid3_disk_state2str(disk->d_state),
2740 	    g_raid3_disk_state2str(state));
2741 	switch (state) {
2742 	case G_RAID3_DISK_STATE_NEW:
2743 		/*
2744 		 * Possible scenarios:
2745 		 * 1. New disk arrive.
2746 		 */
2747 		/* Previous state should be NONE. */
2748 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NONE,
2749 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2750 		    g_raid3_disk_state2str(disk->d_state)));
2751 		DISK_STATE_CHANGED();
2752 
2753 		disk->d_state = state;
2754 		G_RAID3_DEBUG(1, "Device %s: provider %s detected.",
2755 		    sc->sc_name, g_raid3_get_diskname(disk));
2756 		if (sc->sc_state == G_RAID3_DEVICE_STATE_STARTING)
2757 			break;
2758 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2759 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2760 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2761 		    g_raid3_device_state2str(sc->sc_state),
2762 		    g_raid3_get_diskname(disk),
2763 		    g_raid3_disk_state2str(disk->d_state)));
2764 		state = g_raid3_determine_state(disk);
2765 		if (state != G_RAID3_DISK_STATE_NONE)
2766 			goto again;
2767 		break;
2768 	case G_RAID3_DISK_STATE_ACTIVE:
2769 		/*
2770 		 * Possible scenarios:
2771 		 * 1. New disk does not need synchronization.
2772 		 * 2. Synchronization process finished successfully.
2773 		 */
2774 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2775 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2776 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2777 		    g_raid3_device_state2str(sc->sc_state),
2778 		    g_raid3_get_diskname(disk),
2779 		    g_raid3_disk_state2str(disk->d_state)));
2780 		/* Previous state should be NEW or SYNCHRONIZING. */
2781 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW ||
2782 		    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING,
2783 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2784 		    g_raid3_disk_state2str(disk->d_state)));
2785 		DISK_STATE_CHANGED();
2786 
2787 		if (disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
2788 			disk->d_flags &= ~G_RAID3_DISK_FLAG_SYNCHRONIZING;
2789 			disk->d_flags &= ~G_RAID3_DISK_FLAG_FORCE_SYNC;
2790 			g_raid3_sync_stop(sc, 0);
2791 		}
2792 		disk->d_state = state;
2793 		disk->d_sync.ds_offset = 0;
2794 		disk->d_sync.ds_offset_done = 0;
2795 		g_raid3_update_idle(sc, disk);
2796 		g_raid3_update_metadata(disk);
2797 		G_RAID3_DEBUG(1, "Device %s: provider %s activated.",
2798 		    sc->sc_name, g_raid3_get_diskname(disk));
2799 		break;
2800 	case G_RAID3_DISK_STATE_STALE:
2801 		/*
2802 		 * Possible scenarios:
2803 		 * 1. Stale disk was connected.
2804 		 */
2805 		/* Previous state should be NEW. */
2806 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW,
2807 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2808 		    g_raid3_disk_state2str(disk->d_state)));
2809 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2810 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2811 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2812 		    g_raid3_device_state2str(sc->sc_state),
2813 		    g_raid3_get_diskname(disk),
2814 		    g_raid3_disk_state2str(disk->d_state)));
2815 		/*
2816 		 * STALE state is only possible if device is marked
2817 		 * NOAUTOSYNC.
2818 		 */
2819 		KASSERT((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOAUTOSYNC) != 0,
2820 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2821 		    g_raid3_device_state2str(sc->sc_state),
2822 		    g_raid3_get_diskname(disk),
2823 		    g_raid3_disk_state2str(disk->d_state)));
2824 		DISK_STATE_CHANGED();
2825 
2826 		disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2827 		disk->d_state = state;
2828 		g_raid3_update_metadata(disk);
2829 		G_RAID3_DEBUG(0, "Device %s: provider %s is stale.",
2830 		    sc->sc_name, g_raid3_get_diskname(disk));
2831 		break;
2832 	case G_RAID3_DISK_STATE_SYNCHRONIZING:
2833 		/*
2834 		 * Possible scenarios:
2835 		 * 1. Disk which needs synchronization was connected.
2836 		 */
2837 		/* Previous state should be NEW. */
2838 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW,
2839 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2840 		    g_raid3_disk_state2str(disk->d_state)));
2841 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2842 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2843 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2844 		    g_raid3_device_state2str(sc->sc_state),
2845 		    g_raid3_get_diskname(disk),
2846 		    g_raid3_disk_state2str(disk->d_state)));
2847 		DISK_STATE_CHANGED();
2848 
2849 		if (disk->d_state == G_RAID3_DISK_STATE_NEW)
2850 			disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2851 		disk->d_state = state;
2852 		if (sc->sc_provider != NULL) {
2853 			g_raid3_sync_start(sc);
2854 			g_raid3_update_metadata(disk);
2855 		}
2856 		break;
2857 	case G_RAID3_DISK_STATE_DISCONNECTED:
2858 		/*
2859 		 * Possible scenarios:
2860 		 * 1. Device wasn't running yet, but disk disappear.
2861 		 * 2. Disk was active and disapppear.
2862 		 * 3. Disk disappear during synchronization process.
2863 		 */
2864 		if (sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2865 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
2866 			/*
2867 			 * Previous state should be ACTIVE, STALE or
2868 			 * SYNCHRONIZING.
2869 			 */
2870 			KASSERT(disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
2871 			    disk->d_state == G_RAID3_DISK_STATE_STALE ||
2872 			    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING,
2873 			    ("Wrong disk state (%s, %s).",
2874 			    g_raid3_get_diskname(disk),
2875 			    g_raid3_disk_state2str(disk->d_state)));
2876 		} else if (sc->sc_state == G_RAID3_DEVICE_STATE_STARTING) {
2877 			/* Previous state should be NEW. */
2878 			KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW,
2879 			    ("Wrong disk state (%s, %s).",
2880 			    g_raid3_get_diskname(disk),
2881 			    g_raid3_disk_state2str(disk->d_state)));
2882 			/*
2883 			 * Reset bumping syncid if disk disappeared in STARTING
2884 			 * state.
2885 			 */
2886 			if ((sc->sc_bump_id & G_RAID3_BUMP_SYNCID) != 0)
2887 				sc->sc_bump_id &= ~G_RAID3_BUMP_SYNCID;
2888 #ifdef	INVARIANTS
2889 		} else {
2890 			KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
2891 			    sc->sc_name,
2892 			    g_raid3_device_state2str(sc->sc_state),
2893 			    g_raid3_get_diskname(disk),
2894 			    g_raid3_disk_state2str(disk->d_state)));
2895 #endif
2896 		}
2897 		DISK_STATE_CHANGED();
2898 		G_RAID3_DEBUG(0, "Device %s: provider %s disconnected.",
2899 		    sc->sc_name, g_raid3_get_diskname(disk));
2900 
2901 		g_raid3_destroy_disk(disk);
2902 		break;
2903 	default:
2904 		KASSERT(1 == 0, ("Unknown state (%u).", state));
2905 		break;
2906 	}
2907 	return (0);
2908 }
2909 #undef	DISK_STATE_CHANGED
2910 
2911 int
2912 g_raid3_read_metadata(struct g_consumer *cp, struct g_raid3_metadata *md)
2913 {
2914 	struct g_provider *pp;
2915 	u_char *buf;
2916 	int error;
2917 
2918 	g_topology_assert();
2919 
2920 	error = g_access(cp, 1, 0, 0);
2921 	if (error != 0)
2922 		return (error);
2923 	pp = cp->provider;
2924 	g_topology_unlock();
2925 	/* Metadata are stored on last sector. */
2926 	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
2927 	    &error);
2928 	g_topology_lock();
2929 	g_access(cp, -1, 0, 0);
2930 	if (buf == NULL) {
2931 		G_RAID3_DEBUG(1, "Cannot read metadata from %s (error=%d).",
2932 		    cp->provider->name, error);
2933 		return (error);
2934 	}
2935 
2936 	/* Decode metadata. */
2937 	error = raid3_metadata_decode(buf, md);
2938 	g_free(buf);
2939 	if (strcmp(md->md_magic, G_RAID3_MAGIC) != 0)
2940 		return (EINVAL);
2941 	if (md->md_version > G_RAID3_VERSION) {
2942 		G_RAID3_DEBUG(0,
2943 		    "Kernel module is too old to handle metadata from %s.",
2944 		    cp->provider->name);
2945 		return (EINVAL);
2946 	}
2947 	if (error != 0) {
2948 		G_RAID3_DEBUG(1, "MD5 metadata hash mismatch for provider %s.",
2949 		    cp->provider->name);
2950 		return (error);
2951 	}
2952 	if (md->md_sectorsize > maxphys) {
2953 		G_RAID3_DEBUG(0, "The blocksize is too big.");
2954 		return (EINVAL);
2955 	}
2956 
2957 	return (0);
2958 }
2959 
2960 static int
2961 g_raid3_check_metadata(struct g_raid3_softc *sc, struct g_provider *pp,
2962     struct g_raid3_metadata *md)
2963 {
2964 
2965 	if (md->md_no >= sc->sc_ndisks) {
2966 		G_RAID3_DEBUG(1, "Invalid disk %s number (no=%u), skipping.",
2967 		    pp->name, md->md_no);
2968 		return (EINVAL);
2969 	}
2970 	if (sc->sc_disks[md->md_no].d_state != G_RAID3_DISK_STATE_NODISK) {
2971 		G_RAID3_DEBUG(1, "Disk %s (no=%u) already exists, skipping.",
2972 		    pp->name, md->md_no);
2973 		return (EEXIST);
2974 	}
2975 	if (md->md_all != sc->sc_ndisks) {
2976 		G_RAID3_DEBUG(1,
2977 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2978 		    "md_all", pp->name, sc->sc_name);
2979 		return (EINVAL);
2980 	}
2981 	if ((md->md_mediasize % md->md_sectorsize) != 0) {
2982 		G_RAID3_DEBUG(1, "Invalid metadata (mediasize %% sectorsize != "
2983 		    "0) on disk %s (device %s), skipping.", pp->name,
2984 		    sc->sc_name);
2985 		return (EINVAL);
2986 	}
2987 	if (md->md_mediasize != sc->sc_mediasize) {
2988 		G_RAID3_DEBUG(1,
2989 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2990 		    "md_mediasize", pp->name, sc->sc_name);
2991 		return (EINVAL);
2992 	}
2993 	if ((md->md_mediasize % (sc->sc_ndisks - 1)) != 0) {
2994 		G_RAID3_DEBUG(1,
2995 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2996 		    "md_mediasize", pp->name, sc->sc_name);
2997 		return (EINVAL);
2998 	}
2999 	if ((sc->sc_mediasize / (sc->sc_ndisks - 1)) > pp->mediasize) {
3000 		G_RAID3_DEBUG(1,
3001 		    "Invalid size of disk %s (device %s), skipping.", pp->name,
3002 		    sc->sc_name);
3003 		return (EINVAL);
3004 	}
3005 	if ((md->md_sectorsize / pp->sectorsize) < sc->sc_ndisks - 1) {
3006 		G_RAID3_DEBUG(1,
3007 		    "Invalid '%s' field on disk %s (device %s), skipping.",
3008 		    "md_sectorsize", pp->name, sc->sc_name);
3009 		return (EINVAL);
3010 	}
3011 	if (md->md_sectorsize != sc->sc_sectorsize) {
3012 		G_RAID3_DEBUG(1,
3013 		    "Invalid '%s' field on disk %s (device %s), skipping.",
3014 		    "md_sectorsize", pp->name, sc->sc_name);
3015 		return (EINVAL);
3016 	}
3017 	if ((sc->sc_sectorsize % pp->sectorsize) != 0) {
3018 		G_RAID3_DEBUG(1,
3019 		    "Invalid sector size of disk %s (device %s), skipping.",
3020 		    pp->name, sc->sc_name);
3021 		return (EINVAL);
3022 	}
3023 	if ((md->md_mflags & ~G_RAID3_DEVICE_FLAG_MASK) != 0) {
3024 		G_RAID3_DEBUG(1,
3025 		    "Invalid device flags on disk %s (device %s), skipping.",
3026 		    pp->name, sc->sc_name);
3027 		return (EINVAL);
3028 	}
3029 	if ((md->md_mflags & G_RAID3_DEVICE_FLAG_VERIFY) != 0 &&
3030 	    (md->md_mflags & G_RAID3_DEVICE_FLAG_ROUND_ROBIN) != 0) {
3031 		/*
3032 		 * VERIFY and ROUND-ROBIN options are mutally exclusive.
3033 		 */
3034 		G_RAID3_DEBUG(1, "Both VERIFY and ROUND-ROBIN flags exist on "
3035 		    "disk %s (device %s), skipping.", pp->name, sc->sc_name);
3036 		return (EINVAL);
3037 	}
3038 	if ((md->md_dflags & ~G_RAID3_DISK_FLAG_MASK) != 0) {
3039 		G_RAID3_DEBUG(1,
3040 		    "Invalid disk flags on disk %s (device %s), skipping.",
3041 		    pp->name, sc->sc_name);
3042 		return (EINVAL);
3043 	}
3044 	return (0);
3045 }
3046 
3047 int
3048 g_raid3_add_disk(struct g_raid3_softc *sc, struct g_provider *pp,
3049     struct g_raid3_metadata *md)
3050 {
3051 	struct g_raid3_disk *disk;
3052 	int error;
3053 
3054 	g_topology_assert_not();
3055 	G_RAID3_DEBUG(2, "Adding disk %s.", pp->name);
3056 
3057 	error = g_raid3_check_metadata(sc, pp, md);
3058 	if (error != 0)
3059 		return (error);
3060 	if (sc->sc_state != G_RAID3_DEVICE_STATE_STARTING &&
3061 	    md->md_genid < sc->sc_genid) {
3062 		G_RAID3_DEBUG(0, "Component %s (device %s) broken, skipping.",
3063 		    pp->name, sc->sc_name);
3064 		return (EINVAL);
3065 	}
3066 	disk = g_raid3_init_disk(sc, pp, md, &error);
3067 	if (disk == NULL)
3068 		return (error);
3069 	error = g_raid3_event_send(disk, G_RAID3_DISK_STATE_NEW,
3070 	    G_RAID3_EVENT_WAIT);
3071 	if (error != 0)
3072 		return (error);
3073 	if (md->md_version < G_RAID3_VERSION) {
3074 		G_RAID3_DEBUG(0, "Upgrading metadata on %s (v%d->v%d).",
3075 		    pp->name, md->md_version, G_RAID3_VERSION);
3076 		g_raid3_update_metadata(disk);
3077 	}
3078 	return (0);
3079 }
3080 
3081 static void
3082 g_raid3_destroy_delayed(void *arg, int flag)
3083 {
3084 	struct g_raid3_softc *sc;
3085 	int error;
3086 
3087 	if (flag == EV_CANCEL) {
3088 		G_RAID3_DEBUG(1, "Destroying canceled.");
3089 		return;
3090 	}
3091 	sc = arg;
3092 	g_topology_unlock();
3093 	sx_xlock(&sc->sc_lock);
3094 	KASSERT((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) == 0,
3095 	    ("DESTROY flag set on %s.", sc->sc_name));
3096 	KASSERT((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROYING) != 0,
3097 	    ("DESTROYING flag not set on %s.", sc->sc_name));
3098 	G_RAID3_DEBUG(0, "Destroying %s (delayed).", sc->sc_name);
3099 	error = g_raid3_destroy(sc, G_RAID3_DESTROY_SOFT);
3100 	if (error != 0) {
3101 		G_RAID3_DEBUG(0, "Cannot destroy %s.", sc->sc_name);
3102 		sx_xunlock(&sc->sc_lock);
3103 	}
3104 	g_topology_lock();
3105 }
3106 
3107 static int
3108 g_raid3_access(struct g_provider *pp, int acr, int acw, int ace)
3109 {
3110 	struct g_raid3_softc *sc;
3111 	int dcr, dcw, dce, error = 0;
3112 
3113 	g_topology_assert();
3114 	G_RAID3_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, acr,
3115 	    acw, ace);
3116 
3117 	sc = pp->private;
3118 	KASSERT(sc != NULL, ("NULL softc (provider=%s).", pp->name));
3119 
3120 	dcr = pp->acr + acr;
3121 	dcw = pp->acw + acw;
3122 	dce = pp->ace + ace;
3123 
3124 	g_topology_unlock();
3125 	sx_xlock(&sc->sc_lock);
3126 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) != 0 ||
3127 	    g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) < sc->sc_ndisks - 1) {
3128 		if (acr > 0 || acw > 0 || ace > 0)
3129 			error = ENXIO;
3130 		goto end;
3131 	}
3132 	if (dcw == 0)
3133 		g_raid3_idle(sc, dcw);
3134 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROYING) != 0) {
3135 		if (acr > 0 || acw > 0 || ace > 0) {
3136 			error = ENXIO;
3137 			goto end;
3138 		}
3139 		if (dcr == 0 && dcw == 0 && dce == 0) {
3140 			g_post_event(g_raid3_destroy_delayed, sc, M_WAITOK,
3141 			    sc, NULL);
3142 		}
3143 	}
3144 end:
3145 	sx_xunlock(&sc->sc_lock);
3146 	g_topology_lock();
3147 	return (error);
3148 }
3149 
3150 static struct g_geom *
3151 g_raid3_create(struct g_class *mp, const struct g_raid3_metadata *md)
3152 {
3153 	struct g_raid3_softc *sc;
3154 	struct g_geom *gp;
3155 	int error, timeout;
3156 	u_int n;
3157 
3158 	g_topology_assert();
3159 	G_RAID3_DEBUG(1, "Creating device %s (id=%u).", md->md_name, md->md_id);
3160 
3161 	/* One disk is minimum. */
3162 	if (md->md_all < 1)
3163 		return (NULL);
3164 	/*
3165 	 * Action geom.
3166 	 */
3167 	gp = g_new_geomf(mp, "%s", md->md_name);
3168 	sc = malloc(sizeof(*sc), M_RAID3, M_WAITOK | M_ZERO);
3169 	sc->sc_disks = malloc(sizeof(struct g_raid3_disk) * md->md_all, M_RAID3,
3170 	    M_WAITOK | M_ZERO);
3171 	gp->start = g_raid3_start;
3172 	gp->orphan = g_raid3_orphan;
3173 	gp->access = g_raid3_access;
3174 	gp->dumpconf = g_raid3_dumpconf;
3175 
3176 	sc->sc_id = md->md_id;
3177 	sc->sc_mediasize = md->md_mediasize;
3178 	sc->sc_sectorsize = md->md_sectorsize;
3179 	sc->sc_ndisks = md->md_all;
3180 	sc->sc_round_robin = 0;
3181 	sc->sc_flags = md->md_mflags;
3182 	sc->sc_bump_id = 0;
3183 	sc->sc_idle = 1;
3184 	sc->sc_last_write = time_uptime;
3185 	sc->sc_writes = 0;
3186 	sc->sc_refcnt = 1;
3187 	for (n = 0; n < sc->sc_ndisks; n++) {
3188 		sc->sc_disks[n].d_softc = sc;
3189 		sc->sc_disks[n].d_no = n;
3190 		sc->sc_disks[n].d_state = G_RAID3_DISK_STATE_NODISK;
3191 	}
3192 	sx_init(&sc->sc_lock, "graid3:lock");
3193 	bioq_init(&sc->sc_queue);
3194 	mtx_init(&sc->sc_queue_mtx, "graid3:queue", NULL, MTX_DEF);
3195 	bioq_init(&sc->sc_regular_delayed);
3196 	bioq_init(&sc->sc_inflight);
3197 	bioq_init(&sc->sc_sync_delayed);
3198 	TAILQ_INIT(&sc->sc_events);
3199 	mtx_init(&sc->sc_events_mtx, "graid3:events", NULL, MTX_DEF);
3200 	callout_init(&sc->sc_callout, 1);
3201 	sc->sc_state = G_RAID3_DEVICE_STATE_STARTING;
3202 	gp->softc = sc;
3203 	sc->sc_geom = gp;
3204 	sc->sc_provider = NULL;
3205 	/*
3206 	 * Synchronization geom.
3207 	 */
3208 	gp = g_new_geomf(mp, "%s.sync", md->md_name);
3209 	gp->softc = sc;
3210 	gp->orphan = g_raid3_orphan;
3211 	sc->sc_sync.ds_geom = gp;
3212 
3213 	if (!g_raid3_use_malloc) {
3214 		sc->sc_zones[G_RAID3_ZONE_64K].sz_zone = uma_zcreate("gr3:64k",
3215 		    65536, g_raid3_uma_ctor, g_raid3_uma_dtor, NULL, NULL,
3216 		    UMA_ALIGN_PTR, 0);
3217 		sc->sc_zones[G_RAID3_ZONE_64K].sz_inuse = 0;
3218 		sc->sc_zones[G_RAID3_ZONE_64K].sz_max = g_raid3_n64k;
3219 		sc->sc_zones[G_RAID3_ZONE_64K].sz_requested =
3220 		    sc->sc_zones[G_RAID3_ZONE_64K].sz_failed = 0;
3221 		sc->sc_zones[G_RAID3_ZONE_16K].sz_zone = uma_zcreate("gr3:16k",
3222 		    16384, g_raid3_uma_ctor, g_raid3_uma_dtor, NULL, NULL,
3223 		    UMA_ALIGN_PTR, 0);
3224 		sc->sc_zones[G_RAID3_ZONE_16K].sz_inuse = 0;
3225 		sc->sc_zones[G_RAID3_ZONE_16K].sz_max = g_raid3_n16k;
3226 		sc->sc_zones[G_RAID3_ZONE_16K].sz_requested =
3227 		    sc->sc_zones[G_RAID3_ZONE_16K].sz_failed = 0;
3228 		sc->sc_zones[G_RAID3_ZONE_4K].sz_zone = uma_zcreate("gr3:4k",
3229 		    4096, g_raid3_uma_ctor, g_raid3_uma_dtor, NULL, NULL,
3230 		    UMA_ALIGN_PTR, 0);
3231 		sc->sc_zones[G_RAID3_ZONE_4K].sz_inuse = 0;
3232 		sc->sc_zones[G_RAID3_ZONE_4K].sz_max = g_raid3_n4k;
3233 		sc->sc_zones[G_RAID3_ZONE_4K].sz_requested =
3234 		    sc->sc_zones[G_RAID3_ZONE_4K].sz_failed = 0;
3235 	}
3236 
3237 	error = kproc_create(g_raid3_worker, sc, &sc->sc_worker, 0, 0,
3238 	    "g_raid3 %s", md->md_name);
3239 	if (error != 0) {
3240 		G_RAID3_DEBUG(1, "Cannot create kernel thread for %s.",
3241 		    sc->sc_name);
3242 		g_destroy_geom(sc->sc_geom);
3243 		g_raid3_free_device(sc);
3244 		return (NULL);
3245 	}
3246 
3247 	G_RAID3_DEBUG(1, "Device %s created (%u components, id=%u).",
3248 	    sc->sc_name, sc->sc_ndisks, sc->sc_id);
3249 
3250 	sc->sc_rootmount = root_mount_hold("GRAID3");
3251 	G_RAID3_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
3252 
3253 	/*
3254 	 * Schedule startup timeout.
3255 	 */
3256 	timeout = atomic_load_acq_int(&g_raid3_timeout);
3257 	sc->sc_timeout_event = malloc(sizeof(struct g_raid3_event), M_RAID3,
3258 	    M_WAITOK);
3259 	callout_reset(&sc->sc_callout, timeout * hz, g_raid3_go, sc);
3260 	return (sc->sc_geom);
3261 }
3262 
3263 int
3264 g_raid3_destroy(struct g_raid3_softc *sc, int how)
3265 {
3266 	struct g_provider *pp;
3267 
3268 	g_topology_assert_not();
3269 	sx_assert(&sc->sc_lock, SX_XLOCKED);
3270 
3271 	pp = sc->sc_provider;
3272 	if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
3273 		switch (how) {
3274 		case G_RAID3_DESTROY_SOFT:
3275 			G_RAID3_DEBUG(1,
3276 			    "Device %s is still open (r%dw%de%d).", pp->name,
3277 			    pp->acr, pp->acw, pp->ace);
3278 			return (EBUSY);
3279 		case G_RAID3_DESTROY_DELAYED:
3280 			G_RAID3_DEBUG(1,
3281 			    "Device %s will be destroyed on last close.",
3282 			    pp->name);
3283 			if (sc->sc_syncdisk != NULL)
3284 				g_raid3_sync_stop(sc, 1);
3285 			sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROYING;
3286 			return (EBUSY);
3287 		case G_RAID3_DESTROY_HARD:
3288 			G_RAID3_DEBUG(1, "Device %s is still open, so it "
3289 			    "can't be definitely removed.", pp->name);
3290 			break;
3291 		}
3292 	}
3293 
3294 	g_topology_lock();
3295 	if (sc->sc_geom->softc == NULL) {
3296 		g_topology_unlock();
3297 		return (0);
3298 	}
3299 	sc->sc_geom->softc = NULL;
3300 	sc->sc_sync.ds_geom->softc = NULL;
3301 	g_topology_unlock();
3302 
3303 	sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
3304 	sc->sc_flags |= G_RAID3_DEVICE_FLAG_WAIT;
3305 	G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, sc);
3306 	sx_xunlock(&sc->sc_lock);
3307 	mtx_lock(&sc->sc_queue_mtx);
3308 	wakeup(sc);
3309 	wakeup(&sc->sc_queue);
3310 	mtx_unlock(&sc->sc_queue_mtx);
3311 	G_RAID3_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
3312 	while (sc->sc_worker != NULL)
3313 		tsleep(&sc->sc_worker, PRIBIO, "r3:destroy", hz / 5);
3314 	G_RAID3_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
3315 	sx_xlock(&sc->sc_lock);
3316 	g_raid3_destroy_device(sc);
3317 	return (0);
3318 }
3319 
3320 static void
3321 g_raid3_taste_orphan(struct g_consumer *cp)
3322 {
3323 
3324 	KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
3325 	    cp->provider->name));
3326 }
3327 
3328 static struct g_geom *
3329 g_raid3_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
3330 {
3331 	struct g_raid3_metadata md;
3332 	struct g_raid3_softc *sc;
3333 	struct g_consumer *cp;
3334 	struct g_geom *gp;
3335 	int error;
3336 
3337 	g_topology_assert();
3338 	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
3339 	G_RAID3_DEBUG(2, "Tasting %s.", pp->name);
3340 
3341 	gp = g_new_geomf(mp, "raid3:taste");
3342 	/* This orphan function should be never called. */
3343 	gp->orphan = g_raid3_taste_orphan;
3344 	cp = g_new_consumer(gp);
3345 	cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
3346 	error = g_attach(cp, pp);
3347 	if (error == 0) {
3348 		error = g_raid3_read_metadata(cp, &md);
3349 		g_detach(cp);
3350 	}
3351 	g_destroy_consumer(cp);
3352 	g_destroy_geom(gp);
3353 	if (error != 0)
3354 		return (NULL);
3355 	gp = NULL;
3356 
3357 	if (md.md_provider[0] != '\0' &&
3358 	    !g_compare_names(md.md_provider, pp->name))
3359 		return (NULL);
3360 	if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
3361 		return (NULL);
3362 	if (g_raid3_debug >= 2)
3363 		raid3_metadata_dump(&md);
3364 
3365 	/*
3366 	 * Let's check if device already exists.
3367 	 */
3368 	sc = NULL;
3369 	LIST_FOREACH(gp, &mp->geom, geom) {
3370 		sc = gp->softc;
3371 		if (sc == NULL)
3372 			continue;
3373 		if (sc->sc_sync.ds_geom == gp)
3374 			continue;
3375 		if (strcmp(md.md_name, sc->sc_name) != 0)
3376 			continue;
3377 		if (md.md_id != sc->sc_id) {
3378 			G_RAID3_DEBUG(0, "Device %s already configured.",
3379 			    sc->sc_name);
3380 			return (NULL);
3381 		}
3382 		break;
3383 	}
3384 	if (gp == NULL) {
3385 		gp = g_raid3_create(mp, &md);
3386 		if (gp == NULL) {
3387 			G_RAID3_DEBUG(0, "Cannot create device %s.",
3388 			    md.md_name);
3389 			return (NULL);
3390 		}
3391 		sc = gp->softc;
3392 	}
3393 	G_RAID3_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
3394 	g_topology_unlock();
3395 	sx_xlock(&sc->sc_lock);
3396 	error = g_raid3_add_disk(sc, pp, &md);
3397 	if (error != 0) {
3398 		G_RAID3_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
3399 		    pp->name, gp->name, error);
3400 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_NODISK) ==
3401 		    sc->sc_ndisks) {
3402 			g_cancel_event(sc);
3403 			g_raid3_destroy(sc, G_RAID3_DESTROY_HARD);
3404 			g_topology_lock();
3405 			return (NULL);
3406 		}
3407 		gp = NULL;
3408 	}
3409 	sx_xunlock(&sc->sc_lock);
3410 	g_topology_lock();
3411 	return (gp);
3412 }
3413 
3414 static int
3415 g_raid3_destroy_geom(struct gctl_req *req __unused, struct g_class *mp __unused,
3416     struct g_geom *gp)
3417 {
3418 	struct g_raid3_softc *sc;
3419 	int error;
3420 
3421 	g_topology_unlock();
3422 	sc = gp->softc;
3423 	sx_xlock(&sc->sc_lock);
3424 	g_cancel_event(sc);
3425 	error = g_raid3_destroy(gp->softc, G_RAID3_DESTROY_SOFT);
3426 	if (error != 0)
3427 		sx_xunlock(&sc->sc_lock);
3428 	g_topology_lock();
3429 	return (error);
3430 }
3431 
3432 static void
3433 g_raid3_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
3434     struct g_consumer *cp, struct g_provider *pp)
3435 {
3436 	struct g_raid3_softc *sc;
3437 
3438 	g_topology_assert();
3439 
3440 	sc = gp->softc;
3441 	if (sc == NULL)
3442 		return;
3443 	/* Skip synchronization geom. */
3444 	if (gp == sc->sc_sync.ds_geom)
3445 		return;
3446 	if (pp != NULL) {
3447 		/* Nothing here. */
3448 	} else if (cp != NULL) {
3449 		struct g_raid3_disk *disk;
3450 
3451 		disk = cp->private;
3452 		if (disk == NULL)
3453 			return;
3454 		g_topology_unlock();
3455 		sx_xlock(&sc->sc_lock);
3456 		sbuf_printf(sb, "%s<Type>", indent);
3457 		if (disk->d_no == sc->sc_ndisks - 1)
3458 			sbuf_cat(sb, "PARITY");
3459 		else
3460 			sbuf_cat(sb, "DATA");
3461 		sbuf_cat(sb, "</Type>\n");
3462 		sbuf_printf(sb, "%s<Number>%u</Number>\n", indent,
3463 		    (u_int)disk->d_no);
3464 		if (disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
3465 			sbuf_printf(sb, "%s<Synchronized>", indent);
3466 			if (disk->d_sync.ds_offset == 0)
3467 				sbuf_cat(sb, "0%");
3468 			else {
3469 				sbuf_printf(sb, "%u%%",
3470 				    (u_int)((disk->d_sync.ds_offset * 100) /
3471 				    (sc->sc_mediasize / (sc->sc_ndisks - 1))));
3472 			}
3473 			sbuf_cat(sb, "</Synchronized>\n");
3474 			if (disk->d_sync.ds_offset > 0) {
3475 				sbuf_printf(sb, "%s<BytesSynced>%jd"
3476 				    "</BytesSynced>\n", indent,
3477 				    (intmax_t)disk->d_sync.ds_offset);
3478 			}
3479 		}
3480 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent,
3481 		    disk->d_sync.ds_syncid);
3482 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, disk->d_genid);
3483 		sbuf_printf(sb, "%s<Flags>", indent);
3484 		if (disk->d_flags == 0)
3485 			sbuf_cat(sb, "NONE");
3486 		else {
3487 			int first = 1;
3488 
3489 #define	ADD_FLAG(flag, name)	do {					\
3490 	if ((disk->d_flags & (flag)) != 0) {				\
3491 		if (!first)						\
3492 			sbuf_cat(sb, ", ");				\
3493 		else							\
3494 			first = 0;					\
3495 		sbuf_cat(sb, name);					\
3496 	}								\
3497 } while (0)
3498 			ADD_FLAG(G_RAID3_DISK_FLAG_DIRTY, "DIRTY");
3499 			ADD_FLAG(G_RAID3_DISK_FLAG_HARDCODED, "HARDCODED");
3500 			ADD_FLAG(G_RAID3_DISK_FLAG_SYNCHRONIZING,
3501 			    "SYNCHRONIZING");
3502 			ADD_FLAG(G_RAID3_DISK_FLAG_FORCE_SYNC, "FORCE_SYNC");
3503 			ADD_FLAG(G_RAID3_DISK_FLAG_BROKEN, "BROKEN");
3504 #undef	ADD_FLAG
3505 		}
3506 		sbuf_cat(sb, "</Flags>\n");
3507 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3508 		    g_raid3_disk_state2str(disk->d_state));
3509 		sx_xunlock(&sc->sc_lock);
3510 		g_topology_lock();
3511 	} else {
3512 		g_topology_unlock();
3513 		sx_xlock(&sc->sc_lock);
3514 		if (!g_raid3_use_malloc) {
3515 			sbuf_printf(sb,
3516 			    "%s<Zone4kRequested>%u</Zone4kRequested>\n", indent,
3517 			    sc->sc_zones[G_RAID3_ZONE_4K].sz_requested);
3518 			sbuf_printf(sb,
3519 			    "%s<Zone4kFailed>%u</Zone4kFailed>\n", indent,
3520 			    sc->sc_zones[G_RAID3_ZONE_4K].sz_failed);
3521 			sbuf_printf(sb,
3522 			    "%s<Zone16kRequested>%u</Zone16kRequested>\n", indent,
3523 			    sc->sc_zones[G_RAID3_ZONE_16K].sz_requested);
3524 			sbuf_printf(sb,
3525 			    "%s<Zone16kFailed>%u</Zone16kFailed>\n", indent,
3526 			    sc->sc_zones[G_RAID3_ZONE_16K].sz_failed);
3527 			sbuf_printf(sb,
3528 			    "%s<Zone64kRequested>%u</Zone64kRequested>\n", indent,
3529 			    sc->sc_zones[G_RAID3_ZONE_64K].sz_requested);
3530 			sbuf_printf(sb,
3531 			    "%s<Zone64kFailed>%u</Zone64kFailed>\n", indent,
3532 			    sc->sc_zones[G_RAID3_ZONE_64K].sz_failed);
3533 		}
3534 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
3535 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent, sc->sc_syncid);
3536 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, sc->sc_genid);
3537 		sbuf_printf(sb, "%s<Flags>", indent);
3538 		if (sc->sc_flags == 0)
3539 			sbuf_cat(sb, "NONE");
3540 		else {
3541 			int first = 1;
3542 
3543 #define	ADD_FLAG(flag, name)	do {					\
3544 	if ((sc->sc_flags & (flag)) != 0) {				\
3545 		if (!first)						\
3546 			sbuf_cat(sb, ", ");				\
3547 		else							\
3548 			first = 0;					\
3549 		sbuf_cat(sb, name);					\
3550 	}								\
3551 } while (0)
3552 			ADD_FLAG(G_RAID3_DEVICE_FLAG_NOFAILSYNC, "NOFAILSYNC");
3553 			ADD_FLAG(G_RAID3_DEVICE_FLAG_NOAUTOSYNC, "NOAUTOSYNC");
3554 			ADD_FLAG(G_RAID3_DEVICE_FLAG_ROUND_ROBIN,
3555 			    "ROUND-ROBIN");
3556 			ADD_FLAG(G_RAID3_DEVICE_FLAG_VERIFY, "VERIFY");
3557 #undef	ADD_FLAG
3558 		}
3559 		sbuf_cat(sb, "</Flags>\n");
3560 		sbuf_printf(sb, "%s<Components>%u</Components>\n", indent,
3561 		    sc->sc_ndisks);
3562 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3563 		    g_raid3_device_state2str(sc->sc_state));
3564 		sx_xunlock(&sc->sc_lock);
3565 		g_topology_lock();
3566 	}
3567 }
3568 
3569 static void
3570 g_raid3_shutdown_post_sync(void *arg, int howto)
3571 {
3572 	struct g_class *mp;
3573 	struct g_geom *gp, *gp2;
3574 	struct g_raid3_softc *sc;
3575 	int error;
3576 
3577 	if ((howto & RB_NOSYNC) != 0)
3578 		return;
3579 
3580 	mp = arg;
3581 	g_topology_lock();
3582 	g_raid3_shutdown = 1;
3583 	LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
3584 		if ((sc = gp->softc) == NULL)
3585 			continue;
3586 		/* Skip synchronization geom. */
3587 		if (gp == sc->sc_sync.ds_geom)
3588 			continue;
3589 		g_topology_unlock();
3590 		sx_xlock(&sc->sc_lock);
3591 		g_raid3_idle(sc, -1);
3592 		g_cancel_event(sc);
3593 		error = g_raid3_destroy(sc, G_RAID3_DESTROY_DELAYED);
3594 		if (error != 0)
3595 			sx_xunlock(&sc->sc_lock);
3596 		g_topology_lock();
3597 	}
3598 	g_topology_unlock();
3599 }
3600 
3601 static void
3602 g_raid3_init(struct g_class *mp)
3603 {
3604 
3605 	g_raid3_post_sync = EVENTHANDLER_REGISTER(shutdown_post_sync,
3606 	    g_raid3_shutdown_post_sync, mp, SHUTDOWN_PRI_FIRST);
3607 	if (g_raid3_post_sync == NULL)
3608 		G_RAID3_DEBUG(0, "Warning! Cannot register shutdown event.");
3609 }
3610 
3611 static void
3612 g_raid3_fini(struct g_class *mp)
3613 {
3614 
3615 	if (g_raid3_post_sync != NULL)
3616 		EVENTHANDLER_DEREGISTER(shutdown_post_sync, g_raid3_post_sync);
3617 }
3618 
3619 DECLARE_GEOM_CLASS(g_raid3_class, g_raid3);
3620 MODULE_VERSION(geom_raid3, 0);
3621